text
stringlengths
1
1.19M
meta
dict
\section{Introduction}\label{sintro} Entanglement turned out to be a crucial resource for quantum computation. It plays a central r\^ole in quantum communication and quantum computation. A considerable effort is being put into quantifying quantum entanglement. It seems natural to focus the efforts on quantifying entanglement itself, that is, describing the \emph{impossibility} to prepare a state by means of LOCC (local operations and classical communications). One may, although, go another way around and try to quantify \emph{separability} rather than entanglement: this turned out to be applicable for building combinatorial entanglement patterns for multipartite quantum systems \cite{myjmo}. In this paper I dwell on the case of bipartite quantum systems. A state of such system is called {\sc separable} if it can be prepared by LOCC. In terms of density matrices that means that $\mathbf{p}$, its density matrix, can be represented as a mixture of pure product states. According to Carath\'eodory theorem, the number of this states can be reduced to $n^4$ where $n$ is the dimension of the state of a single particle. \medskip The idea to replace finite sums of projectors by continuous distributions on the set of unit vectors is put forward making it possible to provide a geometrical characterization of separable mixed states of a bipartite quantum system. To consistently describe the result presented in this paper recall some necessary definitions. \paragraph{Basics.} A density matrix $\mathbf{p}$ in the product space $\mathfrak{B}$ is called {\sc factorizable} if it is a tensor product of density matrices, $\mathbf{p}=\rho\otimes\rho'$. If $\mathbf{p}$ is a convex combination of factorizable operators, it is said to be {\sc separable} \begin{equation}\label{edefsepar} \mathbf{p} \;=\; \sum_\alpha\limits \,p_\alpha\, \rho_\alpha\otimes\rho'_\alpha \end{equation} \noindent A crucial feature of quantum mechanics, the phenomenon of quantum entanglement, stems from the fact that there exist density operators in the product space which are NOT separable, they are called {\sc entangled}. A density operator $\mathbf{p}$ is called {\sc robustly separable} if it has a neighborhood $U$ in $\mathcal{L}$ such that all operators $\rho'\in{}U$ are separable. \paragraph{A brief account.} In the Euclidean space $\mathfrak{L}$ of self-adjoint operators acting in the tensor product space $\mathfrak{B}=\mathcal{H}\otimes\mathcal{H}'$ we define a real-valued, positive functional ${\mathcal{K}}:\mathfrak{L}\to\mathbb{R}_{+}$ as follows \[ {\mathcal{K}}(X) \;=\; \iint e^{\bracket{\ppp{\phi}}{X}{\ppp{\phi}}} \ppp{\,d\mathbf{S}_{\lth}} \] \noindent where the integration is taken over the torus---the Cartesian product of unit spheres in $\mathcal{H},\mathcal{H}'$, respectively, and consider the hypersurface $\mathcal{K}\subset\mathfrak{L}$ \[ \mathcal{K} \;=\; \{X\in\mathfrak{L} \,\mid\, {\mathcal{K}}(X)=1 \} \] \noindent Then \begin{itemize} \item all robustly separable density operators in $\mathcal{H}$ are in 1--1 correspondence with the points of $\mathcal{K}$ \item the density matrix associated with a point $X\in\mathcal{K}$ is the normal vector to $\mathcal{K}$ at point $X$. \end{itemize} \section{Continuous optimal ensembles}\label{scontensemb} To make the account self-consistent, begin with necessary definitions. A {\sc density operator} is a non-negative self-adjoint operator whose trace equals to 1. In particular, for any unit vector $\ket{\phi}$ the one-dimensional projector $\raypr{\phi}$ is a density matrix. Note that for any set of density operators $\rho_\alpha$ the convex combination $\sum_\alpha{}\rho_\alpha$ is always a density operator. The set of all self-adjoint operators in $\mathcal{H}=\mathbb{C}^n$ has a natural structure of a real space $\mathbb{R}^{2n}$, in which the set of all density matrices is a hypersurface, which is the zero surface $T=0$ of the affine functional $T=\trc{}X-1$. In this paper a geometrical characterization of separable bipartite density operators is provided. It is based on the notion of continuous ensembles. Generalizing the fact that any convex combination of density operators is again a density operator, we represent density operators as probability distributions on the unit sphere in the state space $\mathcal{H}$ of the system. Let us pass to a more detailed account of this issue beginning with the case of a single quantum system. Let $\mathcal{H}=\mathbb{C}^n$ be a $n$-dimensional Hermitian space, let $\rho$ be a density matrix in $\mathcal{H}$. We would like to represent the state whose density operator is $\rho$ by an ensemble of pure states. We would like this ensemble to be continuous with the probability density expressed by a function $\mu(\phi)$ where $\phi$ ranges over all unit vectors in $\mathcal{H}$. \paragraph{Technical remark.} Pure states form a projective space rather than the unit sphere in $\mathcal{H}$. On the other hand, one may integrate over any probabilistic space. Usually distributions of pure states over the spectrum of observables are studied, sometimes probability distributions on the projective spaces are considered \cite{sqprop}. In this paper for technical reasons I prefer to represent ensembles of pure states by measures on unit vectors in $\mathcal{H}$. I use the Umegaki measure on $\cfield{B}_\lth$--- the uniform measure with respect to the action of $U(n)$ normalized so that $\int_{\cfield{B}_\lth}\,d\mathbf{S}_{\lth}=1$. \subsection{Effective definition}\label{scontens} The density operator of a continuous ensemble associated with the measure $\mu(\phi)$ on the set $\cfield{B}_\lth$ of unit vectors in $\mathcal{H}$ is calculated as the following (matrix) integral \begin{equation}\label{e01integral} \rho \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; \mu(\phi)\, \raypr{\phi} \,\,d\mathbf{S}_{\lth} \end{equation} \noindent where $\raypr{\phi}$ is the projector onto the vector $\bra{\phi}$ and $\,d\mathbf{S}_{\lth}$ is the above mentioned normalized measure on $\cfield{B}_\lth$: \begin{equation}\label{einvarmes} \int_{\phi\in\cfield{B}_\lth}\limits\; \,\,d\mathbf{S}_{\lth} \;=\; 1 \end{equation} \noindent Effectively, the operator integral $\rho$ in \eqref{e01integral} can be calculated by its matrix elements. In any fixed basis $\{\ket{\mathbf{e}_i}\}$ in $\mathcal{H}$, each its matrix element $\rho_{ij}=\bracket{\mathbf{e}_i}{\rho}{\mathbf{e}_j}$ is the following numerical integral: \begin{equation}\label{e01basis} \rho_{ij} \;=\; \bracket{\mathbf{e}_i}{\rho}{\mathbf{e}_j} \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; \mu(\phi)\, \braket{\mathbf{e}_i}{\phi} \braket{\phi}{\mathbf{e}_j} \,\,d\mathbf{S}_{\lth} \end{equation} \subsection{Optimal ensembles}\label{soptens} We need to solve the following variational problem. Given a functional $Q$ on $L^1(\cfield{B}_\lth)$ and given a density matrix $\rho$ in $\mathcal{H}$, find the distribution $\mu$ on the set $\cfield{B}_\lth$ of unit vectors in $\mathcal{H}$ such that \begin{equation}\label{e03} \left\lbrace \begin{array}{l} \int_{\phi\in\cfield{B}_\lth}\limits\; \mu(\phi)\,\raypr{\phi}\,d\mathbf{S}_{\lth} \;=\;\rho \\ \qquad \\ Q(\mu)\;\to\; \mbox{extr} \end{array} \right. \end{equation} \noindent We shall consider functionals $Q$ of the form \begin{equation}\label{e03q} Q(\mu) \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; q(\mu(\phi))\,d\mathbf{S}_{\lth} \end{equation} \noindent then, according to \eqref{e01basis}, the variational problem \eqref{e03} reads \[ \left\lbrace \begin{array}{l} \int_{\phi\in\cfield{B}_\lth}\limits\; \mu(\phi) \braket{\mathbf{e}_i}{\phi} \braket{\phi}{\mathbf{e}_j} \,d\mathbf{S}_{\lth}\;=\;\rho_{ij} \\ \int_{\phi\in\cfield{B}_\lth}\limits\; q(\mu(\phi))\,d\mathbf{S}_{\lth} \;\to\; \mbox{extr} \end{array}\right. \] \noindent Solving this variational problem by introducing Lagrangian multiples $X_{ij}$ we get \begin{equation}\label{e03a} q'(\mu(\phi)) \,-\, \sum_{ij} X_{ij} \braket{\mathbf{e}_i}{\phi} \braket{\phi}{\mathbf{e}_j} \;=\; 0 \end{equation} \noindent Combining the Lagrange multiples into the operator $X=\sum_{ij} X_{ij}\ketbra{\mathbf{e}_j}{\mathbf{e}_i}$ turns the equation \eqref{e03a} to \(q'(\mu(\phi)) \,=\, \bracket{\phi}{X}{\phi} \). Then, denoting by $f$ the inverse of $q'$ we write \eqref{e03a} as \begin{equation}\label{e01a} \mu(\phi) = f\left( \bracket{\phi}{X}{\phi} \right) \end{equation} \noindent and the problem reduces to finding $\mu$ from the condition \begin{equation}\label{e04ini} \int_{\phi\in\cfield{B}_\lth}\limits\; \mu(\phi) \raypr{\phi}\,d\mathbf{S}_{\lth} \;=\;\rho \end{equation} \noindent which according to \eqref{e01a} and \eqref{e01basis} can be written as \begin{equation}\label{e04} \bracket{\mathbf{e}_i}{\rho}{\mathbf{e}_j} \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; f\left( \bracket{\phi}{X}{\phi} \right) \raypr{\phi}\,d\mathbf{S}_{\lth} \end{equation} \noindent It follows from \eqref{e03a} that the coefficients $X_{ik}$ can be chosen so that $X_{ik}=\bar{X}_{ki}$. That means that the problem of finding the optimal ensemble reduces to that of finding the coefficients of a self-adjoint operator, that is, to finding $n^2$ numbers from $n^2$ equations. \subsection{Geometrical interpretation}\label{sgeominterpr} The equation \eqref{e04} can be given a direct geometrical meaning. Let $\mathcal{L}\simeq\mathbb{R}^{n^2}$ be the space of all self-adjoint operators in $\mathcal{H}$. Let $f:\mathbb{R}\to\mathbb{R}$ be a differentiable function. Consider the real valued functional $F:\mathcal{L}\to\mathbb{R}$ defined as \begin{equation}\label{edefderiv} F(X) \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; f\left( \bracket{\phi}{X}{\phi} \right) \,d\mathbf{S}_{\lth} \end{equation} \noindent which is well-defined as the set $\cfield{B}_\lth$ is compact. Fix a basis $\{\mathbf{e}_k\}$ in $\mathcal{H}$, then any $X\in\mathcal{L}$ is defined by its matrix elements $X_{ik}=\bracket{\mathbf{e}_i}{X}{\mathbf{e}_k}$, so $\bracket{\phi}{X}{\phi}=\sum_{ik}X_{ik}\braket{\phi}{\mathbf{e}_i}\braket{\mathbf{e}_k}{\phi}$. Then the expression \eqref{edefderiv} can be treated as an integral depending on the set of parameters $\{X_{ik}\}$. We may consider the derivatives of $F(X)$ with respect to these variables, calculate them \[ \frac{\partial}{\partial X_{ik}} \,F(X) \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; \frac{\partial}{\partial X_{ik}} \left( \vphantom{\frac{\partial}{\partial X_{ik}}} \, f\left( \bracket{\phi}{X}{\phi} \right) \right) \,d\mathbf{S}_{\lth} \;=\; \] \begin{equation}\label{ederivgen} \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; \, f'\left( \bracket{\phi}{X}{\phi} \right) \braket{\phi}{\mathbf{e}_i} \, \braket{\mathbf{e}_k}{\phi} \,d\mathbf{S}_{\lth} \;=\; \end{equation} \[ \;=\; \bracket{\mathbf{e}_k}{ \int_{\phi\in\cfield{B}_\lth}\limits\; f'\left( \bracket{\phi}{X}{\phi} \right) \raypr{\phi}\,d\mathbf{S}_{\lth} }{\mathbf{e}_i} \] \medskip \noindent So, the gradient of the functional $F$ is the operator which can be symbolically written as \begin{equation}\label{ederivsymb} \nabla F \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; f'\left( \bracket{\phi}{X}{\phi} \right) \raypr{\phi}\,d\mathbf{S}_{\lth} \end{equation} \noindent and effectively calculated using \eqref{ederivgen}. \subsection{Optimal entropy ensembles}\label{soptentropens} Let us specify the form of the optimality functional in \eqref{e03q} assuming it to be the differential entropy of the appropriate distribution: \begin{equation}\label{edefq} q(\mu) \;=\; -\mu\,\ln\mu \end{equation} \noindent then $q'=-(1+\ln\mu)$ and we have the following $f$ for \eqref{e04} \[ f(x) \;=\; e^{-(1+x)} \] \noindent Introduce, as in \eqref{edefderiv}, the functional ${\mathcal{K}}:\mathcal{L}\to\mathbb{R}$ on the set of all self-adjoint operators in $\mathcal{H}$ (the minus sign and the unit summand are omitted here being a matter of renormalization): \begin{equation}\label{edefk} {\mathcal{K}}(X) \;=\; \int_{\phi\in\cfield{B}_\lth}\limits\; \, e^{\bracket{\phi}{X}{\phi} } \,d\mathbf{S}_{\lth} \end{equation} \noindent Note that $\rho(X)=\int_{\phi\in\cfield{B}_\lth}\limits\; \, e^{\bracket{\phi}{X}{\phi} }\raypr{\phi}\,d\mathbf{S}_{\lth}$ is always a positive operator, then \[ {\mathcal{K}}(X) \;=\; \trc\int_{\phi\in\cfield{B}_\lth}\limits\; \, e^{\bracket{\phi}{X}{\phi} }\raypr{\phi}\,d\mathbf{S}_{\lth} \;=\;1 \] \noindent is a condition which defines a full-range density matrix $\rho(X)$ in $\mathcal{H}$. On the other hand, the condition ${\mathcal{K}}(X)=1$ defines a hypersurface in the Euclidean space $\mathcal{L}$. Together with the fact that $\left(e^x\right)'=e^x$ and \eqref{ederivsymb} we come to the following \paragraph{Statement.} Any full-range density matrix in $\rho$ is associated with a point on the hypersurface ${\mathcal{K}}(X)=1$ and the entries of $\rho$ are calculated as the components of the gradient: \begin{equation}\label{edefrrhgrad} \rho \;=\; \nabla{}{\mathcal{K}} \end{equation} \subsection{The existence}\label{sexist} Why optimal entropy ensembles do exist for all full-range density matrices? First note that for any full-range density matrix $\rho=\sum{}p_k\raypr{\mathbf{e}_k}$ there are infinitely many continuous ensembles (=probability measures on $\cfield{B}_\lth$ in our setting) associated with it. An example of such distribution is $\rho=\sum\,p_k\raypr{\mathbf{e}_k}=\int\mu(\phi)\raypr{\phi}\,d\mathbf{S}_{\lth}$ with \begin{equation}\label{esmeared} \mu(\phi) \;=\; \frac{ \bigl((L+1)n\bigr)! }{ L\,n!(L\,n)! } \; \sum_{k=1}^{n}\limits \left( p_k- \frac{1}{L(n+1)} \right) \,|\braket{\mathbf{e}_k}{\phi}|^{2Ln} \end{equation} \noindent as it follows from \cite{mygibbs}. Here $L$ is a parameter, such that $L>\frac{1}{p_0(n+1)}$ where $p_0>0$ is the smallest eigenvalue of $\rho$. Any probabilistic density $\mu$ whose support is $\cfield{B}_\lth$ is a point in the interior of the simplex of all probabilistic measures on $\cfield{B}_\lth$. For each probabilistic measure on $\cfield{B}_\lth$ its differential entropy can be calculated. The differential entropy is, in turn, a concave function in the affine space of probability distributions. Therefore if we have an affine subset of of probability measure on $\cfield{B}_\lth$, the differential entropy takes its maximal value in the interior of the simplex of probability measures. Now return to the condition in \eqref{e03}---we see that it is affine. Therefore, if we know that there exist at least one continuous ensemble representing $\rho$ (but we know that as mentioned above), that means that there exist a maximal entropy ensemble representing $X$, hence it has the representation \eqref{edefrrhgrad}. \section{Bipartite systems}\label{sbipart} Consider two finite-dimensional quantum systems whose state spaces are $\mathcal{H},\mathcal{H}'$. The state space of the composite system is the tensor product $\mathfrak{B}=\mathcal{H}\otimes\mathcal{H}'$. Denote by $\mathfrak{L}=\mathcal{L}\otimes\mathcal{L}$ the space of all self-adjoint operators in $\mathfrak{B}$. \subsection{Continuous ensembles in bipartite case}\label{scontbi} Let $\mathbf{p}$ be a robustly separable density matrix in the product space $\mathcal{H}\otimes\mathcal{H}'$. Then it can be represented (in infinitely many ways) as a continuous ensemble of pure product states. Carrying out exactly the same reasoning as in section \ref{sexist} we conclude that among those continuous ensembles there exists one having the least differential entropy, this will be the ensemble we are interested in. Like in section \ref{soptentropens}, formulate the variational problem. Let $\mathbf{p}$ be a density operator in a tensor product space $\mathfrak{B}=\mathcal{H}\otimes\mathcal{H}'$. The task is to find a probability density $\mu(\ppp\phi)$ defined on the Cartesian product $\mathfrak{T}=\cfield{B}_\lth\times\cfield{B}_\lth$ of the unit spheres in $\mathcal{H},\mathcal{H}'$, respectively. \begin{equation}\label{e03bi} \left\lbrace \begin{array}{l} \int_{\ppp\phi\in\mathfrak{T}}\limits\; \mu(\ppp\phi)\,\raypr{\ppp\phi}\ppp\,d\mathbf{S}_{\lth} \;=\;\mathbf{p} \\ \qquad \\ Q(\mu)\;\to\; \mbox{extr} \end{array} \right. \end{equation} Proceeding exactly in the same way as with single particle, we get the following representation: \begin{equation}\label{erepbi} \mathbf{p} \;=\; \int_{\ppp\phi\in\mathfrak{T}}\limits\; e^{\bracket{\ppp\phi}{X}{\ppp\phi}} \,\raypr{\ppp\phi}\ppp\,d\mathbf{S}_{\lth} \end{equation} \noindent for some self-adjoint operator $X$ in $\mathcal{L}$ whose existence is guaranteed by the same reasons as in section \ref{sexist}. Why such $X$ does not exist for entangled density operators? The reason is that the set of probability distributions among which $e^{\bracket{\ppp\phi}{X}{\ppp\phi}}$ is optimal is simply void in the entangled case. \subsection{Geometrical characterization of robustly separable quantum states}\label{sgeombi} Now we pass to the main result of this paper. Suppose we deal with a tensor product of two Hilbert spaces $\mathcal{H},\mathcal{H}'$, each of dimension $n$. Consider the space $\mathcal{L}$ of all self-adjoint linear operators in the tensor product $\mathfrak{B}=\mathcal{H}\otimes\mathcal{H}'$, being a Euclidean space of dimension $n^4$. For any $X\in\mathcal{L}$ we can always calculate the integral \begin{equation}\label{edefkbi} {\mathcal{K}}(X) \;=\; \int_{\ppp\phi\in\mathfrak{T}}\limits\; e^{\bracket{\ppp\phi}{X}{\ppp\phi}} \ppp\,d\mathbf{S}_{\lth} \end{equation} \noindent which is always well-defined (as an integral of a bounded function over a compact set), positive (as the exponent is always positive) functional from $\mathcal{L}$ to $\mathbb{R}_+$. Consider the hypersurface $\mathcal{K}$ in $\mathcal{L}$ defined by the equation \[ \mathcal{K} \;=\; \{X\in\mathcal{L} \,|\,{\mathcal{K}}(X)=1 \} \] \noindent In any point of $\mathcal{L}$ the gradient $\nabla {\mathcal{K}}$ can be calculated. In particular, at any point $X$ of $\mathcal{K}$ the gradient $\nabla {\mathcal{K}}$ will be a normal vector to $\mathcal{K}$. The surface $\mathcal{K}$ is something given once and forever, it depends only on the dimensionality of the state space. For any $X$ such that ${\mathcal{K}}(X)=1$, we can calculate the gradient $\mathbf{p}(X)=\nabla {\mathcal{K}}\left|{}_{X}\right.$ at point $X$ Fix bases $\{\mathbf{e}_i\}$, $\{\mathbf{e}'_{i'}\}$, then $X=\sum_{\ppp{i}\,\ppp{k}}\,X_{\ppp{i}\,\ppp{k}}\ketbra{\ppp{i}}{\ppp{k}}$ and the expression \eqref{erepbi} for the operator $\mathbf{p}$ has the following form: \begin{equation}\label{edefgradbi} \mathbf{p}_{\ppp{i}\,\ppp{k}} \;=\; \nabla {\mathcal{K}} \;=\; \frac{\partial{{\mathcal{K}}}}{\partial{X_{\ppp{i}\,\ppp{k}}}} \end{equation} Conversely, given a robustly separable bipartite density matrix $\mathbf{p}$, we know that it can be represented as a convex combination of product states: $\mathbf{p}=\sum p_{\alpha}\rho_{\alpha}\otimes\rho'_{\alpha}$. Each $\rho_{\alpha}$ can be, in turn, represented as a non-vanishing probability distribution \eqref{esmeared}. Then exactly the same reasoning as in section \ref{sexist} can be carried out and there is a point $X$ on the surface $\mathcal{K}$ associated with $\mathbf{p}$. So, together with \eqref{edefgradbi}, we have the main result: \begin{equation}\label{emainresbi} \bigl\{ \mbox{robustly separable states} \bigr\} \quad\leftrightarrow\quad \bigl\{ \mbox{the points of $\mathcal{K}$} \bigr\} \end{equation} \section*{Summary} A geometrical interpretation of robustly separable density operators of a bipartite quantum system with the state space $\mathfrak{B}=\mathcal{H}\otimes\mathcal{H}'$ is provided. They are represented as normal vectors to the hypersurface $\mathcal{K}$ in the (Euclidean) space $\mathfrak{L}$ of self-adjoint operators in $\mathfrak{B}$ defined by the following equation: \begin{equation}\label{edefkbiconc} \mathcal{K} \quad=\quad \left\{ \vphantom{\int_{\ppp\phi\in\mathfrak{T}}\limits} \,X\; \right| \left. \; \int_{\ppp\phi\in\mathfrak{T}}\limits\; e^{\bracket{\ppp\phi}{X}{\ppp\phi}} \ppp\,d\mathbf{S}_{\lth} \;=\;1\; \right\} \end{equation} \noindent where the integration is performed over the set of all unit product vectors $\bra{\ppp\phi}\in\mathfrak{B}$. Each point $X\in\mathcal{K}$ is a self-adjoint operator, the parameter of the probability distribution on the set of unit vectors which gives a density operator $\mathbf{p}$. Furthermore, the normal vector to $\mathcal{K}$ at point $X$ is $\mathbf{p}$ itself: \begin{equation}\label{erepbiconcl} \mathbf{p} \;=\; \nabla {\mathcal{K}}\left|{}_{X}\right. \;=\; \int_{\ppp\phi\in\mathfrak{T}}\limits\; e^{\bracket{\ppp\phi}{X}{\ppp\phi}} \,\raypr{\ppp\phi}\ppp\,d\mathbf{S}_{\lth} \end{equation} \paragraph{The final remark.} Given a density matrix $\mathbf{p}$ in $\mathfrak{B}$, a question arises if it is separable or not. When the dimension of at least one of spaces $\mathcal{H},\mathcal{H}'$ is 2, this question was given an effective answer---the positive partial transpose (PPT) criterion due to Peres-Horodecki was suggested \cite{perehorod}. The criterion states that $\mathbf{p}$ is separable if and only if its partial transpose $\mathbf{p}^{T_2}$ remains non-negative matrix. In higher dimensions PPT is only a necessary condition for a state to be factorizable as there exist entangled density matrices whose partial transpose if positive. Although a geometrical characterization of robustly separable density matrices is provided, it does not solve (directly, at least) the `inverse problem'. Nevertheless, the continuous ensemble method presented in this paper seems to be helpful for tackling the inverse problem as well. This issue is addressed in the next paper on continuous ensembles. \paragraph{Acknowledgments.} The idea to consider continuous ensemble was inspired by the paper \cite{vidaltarrach}, where the notion of robustness for entangled states was introduced, I am grateful to its authors for the inspiration. Much helpful advice from Serguei Krasnikov is highly appreciated. The financial support for this research was provided by the research grant No. 04-06-80215a from RFFI (Russian Basic Research Foundation). Several crucial issues related to this research were intensively duscussed during the meeting Glafka-2004 `Iconoclastic Approaches to Quantum Gravity' (15--18 June, 2004, Athens, Greece) supported by QUALCO Technologies (special thanks to its organizers---Ioannis Raptis and Orestis Tsakalotos).
{ "timestamp": "2005-03-21T11:39:50", "yymm": "0503", "arxiv_id": "quant-ph/0503173", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503173" }
\section{Introduction} The Bloch vector, or vector of coherence \cite{Alicki}, provides a geometric description of the density matrix of a spin-1/2 particle which is commonly used in nuclear magnetic resonance. Mathematically, the Bloch vector may be viewed as the adjoint representation of an $su(2)$ object in an $so(3)$ basis \cite{alta3}. Extension of the notion of vector of coherence to two-spin systems \cite{fano,quan}, and more generally to quantum spin systems of higher dimensions \cite{byrd}, has drawn attention in the contexts of quantum information theory and quantum computation. Specific motivations include the prospects of a useful quantification of entanglement for composite systems \cite{mahl,byrd,alta1} and the quest for equations describing observables in quantum networks \cite{quan}. In the present work, the extension of the Bloch formalism to two spins is used to obtain a geometric representation of the orbits of the vector of coherence for each spin system in the case that a nonlocal interaction of the form $\sigma_i\otimes\sigma_j$ is introduced. We propose that this geometric picture will be useful in devising schemes for control of a quantum state via quantum interfaces \cite{lloyd}, i.e., through the mediation of an ancillary system. In this vein, we investigate the limits of control of a quantum state $S$, mixed or pure, given a nonlocal interaction and an ancilla $Q$. The simple geometric picture developed below also applies to another special case of nonlocal interaction, namely the Heisenberg exchange Hamiltonian. As a second application of our formal results, we investigate the entanglement power of the Heisenberg interaction. \section{Product of operator basis for a density matrix} \subsection{One qubit} The density matrix $\rho$ of a two-state system is a positive semi-definite Hermitian $2 \times 2$ matrix having unit trace. It can always be given expression in terms of the three trace-free Pauli matrices $\sigma_i,~i=1,2,3 $, which are generators of $su(2)$, and $I/{\sqrt{2}}$ ($I$ being the unit matrix): \begin{equation} \rho=\frac{1}{2} I +{\bf v}{\bf \sigma}\,. \label{onequbit} \end{equation} Here $\bf v$ is the vector of coherence, whose magnitude is bounded by $0 \leq\parallel{\bf v}\parallel\leq 1/2$ because $1/2\leq {\rm Tr}(\rho^2)\leq1$. The two limiting values of the norm correspond to maximally mixed and pure states, respectively. The magnitude of the Bloch vector differs by a factor of $1/2$ from that of the vector of coherence, as a matter of convention. Unitary operations rotate the Bloch vector without changing its magnitude: $ SU(2)$ operations on the qubit correspond to $SO(3)$ operations on the Bloch vector. The dynamical evolution of the Bloch vector under non-local operations is considered in the next section. \subsection{Two qubits and the correlation tensor} In analogy to the representation (\ref{onequbit}), we adopt the generators of ${\mit G}=SU(4)$, i.e., the elements of the algebra ${\mit g}=su(4)$ (together with the unit matrix), as an orthonormal basis for the $4\times4$ density matrix of the two-qubit system. We employ this basis as it appears in Ref.~\onlinecite{alta1}, noting that it differs from the basis used in \cite{byrd,mahl} only in the coefficients. The dynamical evolution of the system becomes more transparent if we choose basis elements of the algebra ${\mit g}= su(4)$ in accordance with its Cartan Decomposition ${\mit g}={\mit p}\oplus{\mit e}$ \cite{Bro,Zhang}. The algebras ${\mit p}$ and ${\mit e}$ satisfy the commutations relations \begin{equation} [{\mit e},{\mit e}] \subset {\mit e}\,, \quad [{\mit p},{\mit e}] \subset {\mit p}\,, \quad [{\mit p},{\mit p}] \subset {\mit e}\,. \end{equation} The basis elements, $W_j,~j=1,\ldots,15$ of the orthogonal algebra pair $(e,p)$ are \begin{equation} {\mit e}= {\rm span} \frac{i}{2} \{\sigma_x\otimes1,\sigma_y\otimes1, \sigma_z\otimes1, 1\otimes\sigma_x,1\otimes\sigma_y,1\otimes\sigma_z\}\,, \label{cart1} \end{equation} \begin{equation} {\mit p}= {\rm span} \frac{i}{2} \{\sigma_x\otimes\sigma_x, \sigma_x\otimes \sigma_y,\sigma_x\otimes\sigma_z, \\ \sigma_y\otimes\sigma_x, \sigma_y\otimes \sigma_y,\sigma_y\otimes\sigma_z, \\ \sigma_z\otimes\sigma_x, \sigma_z\otimes \sigma_y,\sigma_z\otimes\sigma_z \}\,. \label{cart2} \end{equation} The basis defined by Eqs.~(\ref{cart1}) and (\ref{cart2}) is used to expand the density matrix as \begin{equation} \rho=\sum_{j=0}^{15} {\rm Tr}(\rho X_j) X_j =\sum_{j=0}^{15}\rho_j X_j\,, \end{equation} where $X_0=I/\sqrt{4}$, $\rho_0=1/\sqrt{4}$, and $X_j= -i W_j$ ($j=1, \ldots, 15$). In this representation, the density matrix is specified by three objects, namely the vectors of coherence ${\bf r}_1$ and ${\bf r}_2$ for the two subsystems along with the spin-spin correlation tensor $T_j^i$. \begin{equation} {\bf r}_1=\left( \begin{array}{c} \rho_1 \\ \rho_2 \\ \rho_3 \end{array}\right)\,, \qquad {\bf r}_2=\left( \begin{array}{c} \rho_4 \\ \rho_5 \\ \rho_6 \end{array}\right)\,, \qquad T_j^i=\left( \begin{array}{ccc} \rho_7 & \rho_8 & \rho_9 \\ \rho_{10} & \rho_{11} & \rho_{12} \\ \rho_{13} & \rho_{14} & \rho_{15} \end{array}\right)\,. \end{equation} We note that the object $ T_j^i$ has other names: Stokes tensor \cite{Boston}, entanglement tensor \cite{mahl}, and tensor of coherence (when combined with the coherence vectors in one object). Details of the properties of $T_j^i$ can be found in Ref.~\onlinecite{alta1}, where many prior studies are cited. This tensor contains information on the correlations between the two subsystems, of both classical and quantum nature. Necessary and sufficient conditions for separability of a pure state can be stated in terms of its properties, whereas in the case of a mixed state, only necessary conditions for separability can be given \cite{alta1}. \section{Evolution Under Local and non Local Operations} As we have seen, the Lie algebra ${\mit g}=su(4)$ possesses a Cartan decomposition $ {\mit g}={\mit e}\oplus {\mit p}$, which informs us that there exists within the Lie group ${\mit G}=SU(4)$ a subgroup of local operations ${\mit G}_L=SU(2)\otimes SU(2)$ generated by ${\mit e}$. All the other operations are nonlocal and members of the coset space $SU(4)/SU(2)\otimes SU(2)$, which does not form a subgroup of $SU(4)$. It is known (see {\it Proposition 1} of Ref.~\onlinecite{Zhang}) that any $U\in SU(4)$ can be written as \begin{equation} U=k_1Ak_2 \label{decom1} \end{equation} with \begin{equation} A = \exp\left[\frac{i}{2}(c_1\sigma_{x}^{1}\sigma_{x}^{2} +c_2\sigma_{y}^{1}\sigma_{y}^{2}+c_3\sigma_{z}^{1}\sigma_{z}^{2})\right]\,, \label{decom2} \end{equation} where $k_1,~k_2 \in SU(2)\otimes SU(2)$ and $c_1,~c_2,~c_3~\in R$. In the following, we focus on the effect of nonlocal operations generated by a single operator among the possibilities for $\sigma_i\otimes\sigma_j$, where $i,j \in\{x,y,z\}$. (This consideration includes the special case in which two of the parameters $c_1$, $c_2$, and $c_3$ in the decomposition (\ref{decom1})-(\ref{decom2}) are zero.) Such nonlocal operations will be called one-dimensional. \subsection{Local operations} Local operations are operations $g\in SU(2)\otimes SU(2)$ generated by the elements of ${\mit e}$. From the commutation relations $[{\mit e},{\mit e}] \subset {\mit e}$ and $[{\mit p},{\mit e}] \subset {\mit p}$ it is clear that the elements of the vectors ${\bf r}_i$ and tensor $(T_i^j)$ do not mix and do not affect one another. Under local operations, the vectors behave just like ordinary Cartesian vectors. In particular, a vector of coherence is rotated about some vector $\hat n$ as illustrated in Fig.~1, i.e., \begin{equation} (r')_{1}^{i}=R^{i}_{j}r_1^j\,, \qquad (r')_{2}^{i}=R^{i}_{j}r_2^j\,. \end{equation} On the other hand, the correlation tensor transforms like a mixed Cartesian tensor, \begin{equation} (T')^{i}_{j}=R^{i}_{m}(R')^{n}_{j} T^{m}_{n}\,. \end{equation} The magnitude of each object remains invariant under local operations. In addition, there exist fifteen more invariants which can be constructed from the vectors and the tensor \cite{makhl}. \begin{figure} \includegraphics[width=12cm]{fig1.eps} \caption{Local operations on the two spin subsystems produce a rotation of the corresponding vectors of coherence around some direction $\hat{{\bf n}}$. The effect is the same for both pure states (a) and mixed states (b).} \end{figure} \subsection{One-dimensional nonlocal operations } The nonlocal operations in the coset space $SU(4)/SU(2)\otimes SU(2)$ require, in their construction, exponentiation of at least one of the elements of ${\mit p}$. Hence, under these operations the elements of the tensor and vectors of coherence are mixed, due to the commutation relations $[{\mit p},{\mit e}] \subset {\mit p}$ and $[{\mit p},{\mit p}] \subset {\mit e}$. We shall establish that the one-dimensional nonlocal operations generated by the chosen interaction $\sigma_i\otimes\sigma_j$ give rise to elliptical orbits for the vectors of coherence of the subsystems. The characteristics of these elliptic paths depend on the indices $i$ and $j$, on the initial states of the subsystems, and on the degree of correlations between them. These orbits can be described by non-unitary transformations on each of the individual subsystems when one traces over the other's degrees of freedom. Accordingly, we take the interaction Hamiltonian between the two spins to be $H_I=\sigma_i\otimes\sigma_j/2$, and, for reasons of simplicity, we suppose that the internal Hamiltonians for the two spins may be ignored. Assuming that the duration of the interaction is $\phi$, and appealing to (i) the commutation relations as summarized in Ref.~\cite{Zhang} and (ii) the identity \begin{equation} \exp\left[-i(\phi/2)\sigma_i\otimes\sigma_j\right] = \cos(\phi/2)I-i\sin(\phi/2)\sigma_i\otimes\sigma_j\,, \end{equation} we can make the following observations: \begin{enumerate} \item[(i)] The components $r^{i}_1$ and $r^{j}_2$ of the vectors of coherence remain unaffected; hence the vectors are confined to planes perpendicular to the $i$-axis and $j$-axis respectively. \item[(ii)] Of the nine elements of the correlation tensor $T^{k}_{l}$, only four experience changes. The five that are unchanged under the action of $\sigma_i\otimes \sigma_j$ are $T^i_j$ and $T^k_l$ with $k\neq i$ and $l\neq j$. \item[(iii)] The vectors $r_1^m+T_j^m$ and $r_2^m+T_m^i$ are rotated, without change of magnitude, through an angle $\phi$ about the $i$ and $j$ axes, respectively. (Here $m$ ranges freely over $\{x,y,z\}$). \item[(iv)] More explicitly, the components of the vectors transform according to \begin{equation} \begin{array}{l} r_{1}^{i}\rightarrow (r')_{1}^{i}=r_{1}^{i}\,, \nonumber \\ r_{1}^{k}\rightarrow (r')_{1}^{k}=r_1^k \cos \phi- T^{l}_{j}\sin \phi \,,\nonumber \\ r_{1}^{l}\rightarrow (r')_{1}^{l}= T^{k}_{j}\sin \phi + r_{1}^{l}\cos \phi \,, \end{array} \begin{array}{l} r_{2}^{j}\rightarrow (r')_{2}^{j}=r_{2}^{j}\,,\nonumber \\ r_{2}^{m}\rightarrow (r')_{2}^{m}=r_{2}^{m}\cos \phi- T^{i}_{n}\sin \phi\,,\nonumber\\ r_{2}^{n}\rightarrow (r')_{2}^{n}= T^{i}_{m}\sin \phi +r_{2}^{n}\cos \phi \,, \end{array} \end{equation} and the components of the tensor of coherence, according to \begin{equation} \begin{array}{l} T_{j}^{i}\rightarrow (T')_{j}^{i}=T_{j}^{i}\,,\nonumber \\ T_{j}^{k}\rightarrow (T')_{j}^{k}=T^{k}_{j}\cos \phi-r^{l}_{1}\sin \phi \,,\nonumber\\ T_{j}^{l}\rightarrow (T')_{j}^{l} =r^{k}_{1}\sin \phi + T_{j}^{l} \cos \phi \,, \end{array},~~ \begin{array}{l} T_{j}^{i}\rightarrow (T')_{j}^{i}=T_{j}^{i}\,,\nonumber \\ T_{m}^{i}\rightarrow (T')_{m}^{i}=T^{i}_{m}\cos \phi -r^{n}_{2}\sin \phi \,,\nonumber\\ T_{n}^{i}\rightarrow (T')_{n}^{i} =r^{m}_{2}\sin \phi + T_{n}^{i}\cos \phi \,, \end{array} \end{equation} with no change in the tensor's other elements. The ordered sets of indices $(i,l,k)$ and $(j,m,n)$ belong to $\{(x,y,z),(y,z,x),(z,x,y)\} $. \end{enumerate} Given this behavior, it is not difficult to show that {\it ${\bf r}_1(\phi)$ and ${\bf r}_2(\phi)_2$ follow elliptical orbits}. Since the 1,2 labeling is arbitrary, it suffices to demonstrate this property for the the vector ${\bf r}_1(\phi)$. \smallskip \noindent {\it Proof.} Referring to Fig.~2(a), the coordinates for a vector $\bf s$ tracing an ellipse in the $x-y$ plane, with principal axes $a$ and $b$ rotated by an angle $\psi$, are \begin{equation} \begin{array}{l} s_x(\phi)= a ~{\rm cos}\phi~{\rm cos}\psi+ b~{\rm sin}\phi~{\rm sin}\psi\,,\nonumber\\ s_y(\phi)= - a~ {\rm cos}\phi~{\rm sin}\psi+ b~{\rm sin}\phi~{\rm cos}\psi \,. \end{array} \end{equation} The angle $\phi$ is zero when the vector $\bf s$ is aligned with the principal axis $a$. The coordinates of the vector of coherence ${\bf r}_1$ moving in the $k-l$ plane are given by \begin{equation} \begin{array}{l} r_1^k(\phi')=r_1^k(0)\cos \phi' - T_j^l(0) \sin \phi'\,,\nonumber \\ r_1^l(\phi')=T_j^k(0)\sin \phi'+r_1^l(0) \cos \phi' \,. \end{array} \end{equation} Of course, for the vector of coherence, $\phi' = 0 $ does not in general correspond to the principal axis $a$ (see Fig.~2(a)). In fact, $\phi'=\phi+\chi$, and the coordinates of ${\bf r}_1 $ can be rewritten as follows in terms of the phase difference $\chi$: \begin{equation} \begin{array}{l} r_1^k(\phi)=(r_1^k(0)\cos\chi- T_j^l(0)\sin\chi)\cos\phi+ (-T_j^l(0)\cos\chi-r_1^k(0)\sin\chi)\sin\phi \,, \nonumber\\ r_1^l(\phi)=(T_j^k(0)\sin\chi+r_1^l(0)\cos\chi) \cos\phi +(-r_1^l(0)\sin\chi+T_j^k(0)\cos\chi)\sin\phi \,. \end{array} \end{equation} Comparison of the two sets of coordinates $\{s_x(\phi),s_y(\phi)\}$ and $\{r_1^k(\phi),r_1^l(\phi) \}$ shows that a match can always be made, such that the parameters $a$, $b$, $\psi$, and $\chi$ can be determined by solving the system of equations \begin{equation} \begin{array}{c} a\cos\psi= r_1^k(0)\cos\chi- T_j^l(0)\sin\chi \,, \nonumber \\ b\sin\psi=-T_j^l(0)\cos\chi-r_1^k(0)\sin\chi \,, \nonumber \\ a\sin\psi=-T_j^k(0)\sin\chi-r_1^l(0)\cos\chi \,, \nonumber \\ b\cos\psi=-r_1^l(0)\sin\chi+T_j^k(0)\cos\chi\,. \label{system} \end{array} \end{equation} This completes the proof. It is important to note that the shape of the ellipse depends explicitly on the spin-spin correlation tensor. \smallskip \begin{figure} \includegraphics[width=12cm]{fig2.eps} \caption{(a) The vector $\bf s$ describes an ellipse with principal axes $a$ and $b$ rotated by angle $\psi$ with respect to the $y$-axis. The angle $\phi$, interpreted as the duration of a group operation, is measured relative to the $a$ principal axis. (b) The vector of coherence $\bf r$ corresponding to one of the spins of the two-spin system moves on an ellipse on the $k-l$ plane, with the angle $\phi'$ measured relative to the original direction of $\bf r$.} \end{figure} Solving Eqs.~(\ref{system}) for the angle $\chi$, we find \begin{equation} \tan(2\chi)= \frac{2\left[r^{k}_{1}(0)T^{l}_{j}(0)-r^{l}_{1}(0)T^{k}_{j}(0))\right]} {-(r^{l}_{1}(0))^2+(T^{k}_{j}(0))^2 -(r^{k}_{1}(0))^2+(T^{l}_{j}(0))^2}\,, \end{equation} which specifies the initial orientation of the coherence vector ${\bf r}_1$ with respect to the principal axis $a$. Suppose now the two-spin system is initially in a {\it product state}. For this case it is easy to prove these corollaries to our principal result: \begin{itemize} \item[(1)] The phase difference $\chi$ is zero. This means that the initial positions of both coherence vectors lie on the $a$ principal axis (as in Fig.~3(a)). It follows that the linear entropy of the state of each of the subsystems (defined by $1 - {\rm Tr}\,\rho^2$) can only decrease, showing it is possible to increase the entanglement of the system with this interaction. (This will depend on initial conditions. See Section \ref{sec:ent-heis}.) \item[(2)] The length of the semi-minor axis of the ellipse followed by subsystem 1 is given by $ b_1=|r_2^j(0)|[(r_1^l(0))^2+(r_1^k(0))^2]^{1/2}$ (and likewise for subsystem 2 with $1 \rightarrow 2$ and $\{j,k,l\} \rightarrow \{i,n,m\}$). It follows that for an initially pure state and the assumed single interaction $\sigma_i \otimes \sigma_j$, the maximum attainable entanglement is achieved at $\phi$ values of $\pi/2$ and $3 \pi/2$. \end{itemize} For the case of a initial state that is not pure but still separable, the phase difference $\chi$ does not vanish, in general (see figure~3(b)). Accordingly, the implied dynamical behavior of a classically correlated system distinguishes it from an uncorrelated system, but not from a system experiencing quantum entanglement. Moreover, the linear entropy of each subsystem can either increase or decrease, showing it is possible to increase or decrease the amount of entanglement in the system. \begin{figure} \includegraphics[width=12cm]{fig3.eps} \caption{The initial position of the vector of coherence of subsystem 1 or 2 is shown, together with its time evolution under a one-dimensional nonlocal interaction (dashed line). If the initial state of the two-spin system is a product state, then the initial position is on the $a$ principal axis of the elliptical path, as in (a). In general this agreement no longer occurs if the subsystems are initially correlated, either classically or quantum mechanically, as in (b).} \end{figure} \subsection{General nonlocal operations} From {\it Proposition 1} of Ref.~\onlinecite{Zhang}, any nonlocal operation can be decomposed as a product of two local operations and an operation of the form \begin{equation} A= \exp\left[ \frac{i}{2}(c_1\sigma_{x}^{1}\sigma_{x}^{2} +c_2\sigma_{y}^{1}\sigma_{y}^{2}+c_3\sigma_{z}^{1}\sigma_{z}^{2})\right] \,. \label{decom3} \end{equation} The operators $\{Y_i\}=\{i\sigma_{x}^{1}\sigma_{x}^{2}/2, i\sigma_{y}^{1}\sigma_{y}^{2}/2,i\sigma_{z}^{1}\sigma_{z}^{2}/2 \}$ span a maximal Abelian subalgebra of ${\mit P}$, and the relations \begin{equation} [Y_i,Y_j]=0\,, \qquad [Y_i,Y_j]_+ = -i|\epsilon_{ijk}|Y_k -\frac{1}{2}\delta_{ij} \end{equation} hold, where $[\cdot,\cdot]_+$ denotes the anticommutator. Consequently, $A$ of Eq.~(\ref{decom3}) can be written in product form, \begin{equation} A =\exp\left[\frac{i}{2}(c_1\sigma_{x}^{1}\sigma_{x}^{2})\right] \exp\left[\frac{i}{2}(c_2\sigma_{y}^{1}\sigma_{y}^{2})\right] \exp\left[\frac{i}{2}(c_3\sigma_{z}^{1}\sigma_{z}^{2})\right]\,. \label{prodform} \end{equation} The property (\ref{prodform}) tells us that {\it any nonlocal operation can be decomposed into a sequence of operations effecting a succession of circular and elliptic paths in Bloch space}. This result facilitates the calculation of the final states of the subsystems, but gives only limited insight into the geometric characteristics of the coherence vectors' time orbits. For all $c_i$ distinct, two general observations can be made: \begin{itemize} \item[(1)] The motion of the vectors of coherence is no longer restricted to a plane, since there is no linear combination of $\{\sigma_x\otimes 1, \sigma_y\otimes 1, \sigma_z\otimes 1\}$ or $\{1\otimes\sigma_x,1\otimes\sigma_y,1\otimes\sigma_z\}$ that is invariant under the action of $A$. \item[(2)] Characteristics of the trajectories such as periodicity depend in detail on the parameters $c_1$, $c_2$, and $c_3$. A trajectory is periodic only if $c_2/c_1$ and $c_3/c_1$ are both rational numbers. We note also that the set of parameters $\{c_1,c_2,c_3\}$ has been used to determine the equivalence classes of nonlocal interactions \cite{Zhang} as well as the invariants of the nonlocal interactions \cite{makhl}. \end{itemize} \subsection{Special case of the Heisenberg Hamiltonian} The Heisenberg exchange Hamiltonian, corresponding to $c_1=c_2=c_3=-c/2$, is not included in our general observations on nonlocal interactions (made for all $c_i$ distinct), but like the one-dimensional Hamiltonians, it admits a simple geometric picture. This interaction is the primary two-qubit interaction in several experimental proposals for quantum-dot qubits \cite{Loss:98, Kane:98, Vrijen:00}. It can also be used for universal quantum computing on encoded qubits of several types \cite{Bacon:00, Kempe:01, DiVincenzo:00a, Lidar:02, Wu:02, Byrd:02}. For these reasons, it warrants special attention. Introducing the time parameter $\phi$, the operator $A$ of Eq.~(\ref{prodform}) now takes the form \begin{eqnarray*} A(\phi) &=& \exp[-i(c\phi/2)(\sigma_x\otimes \sigma_x + \sigma_y\otimes \sigma_y + \sigma_z\otimes \sigma_z)] \nonumber \\ &=& \left[\cos^3(c\phi/2) -i\sin^3(c\phi/2)\right]I\otimes I \nonumber \\ && -(i/2)e^{ic\phi/2}\sin(c\phi)(\sigma_x\otimes \sigma_x + \sigma_y\otimes \sigma_y + \sigma_z\otimes \sigma_z ). \end{eqnarray*} The time development of the density matrix under the operator $A$ is given $\rho(\phi)= A(\phi)\rho(0)A^{\dagger}(\phi)$ and the corresponding coherence vectors change according to \begin{equation} \label{eq:ipart} r_1^i(\phi) = \frac{1}{2}[r^i_1(0)+r^i_2(0)+(r_1^i(0)-r_2^i(0))\cos(2c\phi) +(T_{k}^{j}(0)-T_{j}^{k}(0))\sin(2c\phi)] \,, \end{equation} where $i,j,k =1,2,3$ and cyclic permutations are implied. Similarly, for the coherence tensor we have \begin{equation} \label{eq:tpart} T_j^i(\phi) = \frac{1}{2}[T^{i}_{j}(0)+T^{j}_{i}(0) +(T^{i}_{j}(0)-T^{j}_{i}(0)) \cos(2c\phi) +(r_1^k(0)-r_2^k(0)) \sin(2c\phi)] \,. \end{equation} The elements of ${\bf r}_2(\phi)$ are found by symmetry $1\leftrightarrow 2$. The quantities $r_1^i+r_2^i$, $T^{i}_{j}+T^{j}_{i}$, and $T^{i}_{i}$ are unchanged by the operation, and the form of the one-parameter set that describes the time-evolving coherence vector is \begin{equation} {\bf r}_1(\phi) = {\bf R} + {\bf S}\cos(2c\phi) + {\bf V}\sin(2c\phi)\,, \end{equation} where ${\bf R}={\bf r}_1(0)+{\bf r}_2(0)$, ${\bf S} = {\bf r_1}(0)-{\bf r}_2(0)$, and \begin{equation} {\bf V} = \left(\begin{array}{c} T^{3}_{2}(0)-T^{2}_{3}(0) \\ T^{1}_{3}(0)-T^{3}_{1}(0) \\ T^{2}_{1}(0)-T^{1}_{2}(0) \end{array}\right)\,. \end{equation} Clearly the vector traces out an ellipse lying in the plane spanned by ${\bf S}$ and ${\bf V}$, defined by ${\bf S}\times{\bf V}$, and passing through the point ${\bf R}$. \section{Applications} We shall now illustrate some of the results of Section III with two examples. The first provides a controllability result for nonlocal unitary interactions and the second demonstrates how the orbit of the coherence vector can be used to describe the entangling power of the Heisenberg exchange interaction. \subsection{Quantum control via quantum controllers and one-dimensional nonlocal interactions} Let us now consider the implications of the findings of the preceding sections for the problem of quantum control. To this end, we adopt the nomenclature of Lloyd \cite{lloyd} and identify spin 1 with the system $S$ whose quantum state we wish to control, and spin 2 with the quantum controller or interface $Q$. It is assumed that (i) only one interaction Hamiltonian $H_I$ is in play between $S$ and $Q$ and (i) system $Q$ is completely controllable via control Hamiltonians $\{H_Q^{m}\}=\{1\otimes \sigma_x, 1\otimes \sigma_y,1\otimes \sigma_z\} $ that span the $su(2)$ algebra. The initial state of the bipartite system is taken to be a product state in the ensuing analysis. Suppose that the interaction Hamiltonian is nonlocal, but takes the one-dimensional form $H_I=\sigma_i\otimes \sigma_j $. Then the set $\{\{H_Q^{m}\},H_I\}$ $=\{1\otimes \sigma_x, 1\otimes \sigma_y, 1\otimes \sigma_z, \sigma_i\otimes \sigma_x, \sigma_i\otimes \sigma_y,\sigma_i\otimes \sigma_z\} $ comprises a closed six-element subalgebra ${\mit G}_6$ of ${\mit G} $. Given this set of operations, the vector of coherence ${\bf r}_S$ of system $S$ remains in the plane perpendicular to the $i$-axis. It has been established in Sec.~III that when $H_I=\sigma_i\otimes \sigma_j $ is the only element of the algebra $su(4)$ affecting the two-spin system, the vectors of coherence ${\bf r}_1$ and ${\bf r}_2$ are constrained to move in elliptical orbits. Now, with the six-element subalgebra ${\mit G}_6$ available to the two-spin system $S+Q$, the reachable set of the system $S$ is enlarged to an {\it elliptical disk} (see Fig.~4). The principle axis of the disk coincides with the initial coherence vector ${\bf r}_S(0)$ of the $S$ system, while the length of its semiminor axis is given by $b=[(r_S^k(0))^2+(r_S^l(0))^2]^{1/2}|{\bf r}_Q(0)|$, where ${\bf r}_Q(0)$ is the initial coherence vector of system $Q$. \smallskip \noindent {\it Proof.} First, if one implements the two-step sequence of a local operation $\in 1\otimes su(2)$ on system $Q$ followed by the nonlocal operation $H_I=\sigma_i\otimes \sigma_j$ on $S+Q$, the orbit of ${\bf r}_S$ is necessarily an ellipse whose $a$ principle axis lies along the initial coherence vector ${\bf r}_S(0)$ and whose semimajor axis $b$ is restricted by $0\le b\le [(r_S^k(0))^2+(r_S^j(0))^2]^{1/2}|{\bf r}_Q(0)|$. Hence the set reachable by this two-step procedure is the elliptic disk in question. Second, using the Baker-Campbell-Hausdorff formula one can show that all the elements of the six-element subalgebra ${\mit G}_6$ can be constructed by this two step sequence, so their reachable sets are the same. From this result we infer that {\it the entropy of system $S$ cannot be decreased by intervention of the quantum interface $Q$ if the interaction Hamiltonian is limited to the form $H_I=\sigma_i\otimes \sigma_j$}. Noting that $|{\bf r}_Q|\le 1/2$, it follows that $a\ge b$, where $a$ and $b$ are respectively the magnitudes of the semimajor and semiminor axes of the elliptical reachable set. Furthermore, it is seen that the systems $S$ and $Q$ become maximally entangled if the initial state of the system $S$ is situated on the equatorial plane perpendicular to $i$-axis. \begin{figure} \includegraphics[width=12cm]{fig4.eps} \caption{The gray area is the set of reachable states for the system $S$ if one has full control of the controller $Q$ and the interaction $\sigma_i\otimes\sigma_j$ is available. This elliptical disk is characterized by a semimajor axis coincident with the initial vector of coherence for $S$ and a semiminor axis with $b= [(r_S^k(0))^2+(r_S^j(0))^2]^{1/2}|{\bf r}_Q(0)|$.} \end{figure} \subsection{Entanglement power of Heisenberg interaction} \label{sec:ent-heis} Upon examining Eq.~(\ref{eq:ipart}), we see that the maximum entanglement, realized in a maximally entangled pure state, can be achieved if ${\bf {r}}_1(0) = - {\bf {r}}_2(0)$, $|{\bf r}_1|=1/2$, and $c\phi = \pm\pi/4$. Otherwise, the state is not perfectly entangled since the linear entropy 1-Tr($\rho^2$) is not minimized. This conclusion agrees with the result of Zhang {\it et al.} \cite{Zhang} that the only perfect entanglers that can be achieved with the Heisenberg Hamiltonian are the square-root of swap and its inverse. However, suppose that the initial state of the two-spin system is represented by $$ \rho(0) = \frac{1}{4}(I + \sigma_z)\otimes (I + \sigma_z) \,, $$ which is a pure-state density matrix for which $r_1^z = 1/2 = r_2^z$ and $T^{z}_{z} = 1/2$, all other elements of the coherence vectors and coherence tensor being zero. Then $$ r_1^{x}(\phi) = r_1^z(0)\cos^2(c\phi)+r_2^z(0)\sin^2(c\phi)\,, $$ while all other components of ${\bf r}_1$ and ${\bf r}_2$ vanish at time $\phi$, and all other $r_1^\alpha(0) =0$. In this case the ellipse collapses to a line and the coherence vector simply oscillates between two values along that line. The only element of the correlation tensor that changes is $$ T^{1}_{2} = \frac{1}{2}\sin(2c\phi)(r_1^z(0)-r_2^z(0))\,, $$ which vanishes for an initial tensor product of pure states for which the subsystems are polarized in the $+z$ direction. Therefore one cannot create maximally entangled states with these initial conditions. \section{Conclusions} In this paper we have developed a geometric representation for the orbits of the coherence vectors of a two-qubit system. In various circumstances we have shown that their evolution is described by elliptical orbits lying within the surface of the Bloch sphere. Importantly, every two-qubit unitary operation can be expressed as a combination of one of the evolutions we have considered, together with ``pre'' and ``post'' local single-qubit rotations. We anticipate that this geometric picture will be helpful in devising schemes for control of a quantum state via quantum interfaces, and we have obtained a controllability result appropriate for such applications. Given the utility of the coherence-vector picture for modeling quantum systems and describing their entanglement, further studies along similar lines may be fruitful. Such work could include analysis of the orbits of higher-dimensional quantum states, as well as consideration of the effects of measurement operations on controllability. \section*{Acknowledgments} This research was supported by the U.~S.\ National Science Foundation under Grant No.~PHY-0140316 (JWC and AM) and by the Nipher Fund.
{ "timestamp": "2005-08-19T20:34:04", "yymm": "0503", "arxiv_id": "quant-ph/0503208", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503208" }
\section*{The effective polarization density matrix} In these notes, our first aim is to re-derive the formulas for the effective density matrix $\rho_\mathrm{eff}$ we introduced in \cite{Aiello04c}, without using the concept of Stokes parameters and without using a circular polarization basis. To begin with, let us establish our notation. In order to make the results comparable with the ones in Refs. \cite{Peres_et_al}, we adopt a relativistic notation. All formulas are given in natural units $(c = \hbar = 1)$ and all quantized fields are transverse. In this context, single-photon plane-wave states are denoted by $|\mathbf{k}, \lambda \rangle $ where $\mathbf{k}$ is the spatial part of the four momentum $k = (k^0, \mathbf{k}), \; k^0 = |\mathbf{k}| \equiv \omega$, and $\lambda = 1,2$ is the {\em linear} polarization. They are created from the vacuum state $| 0 \rangle$ by the corresponding creation operator $\hat{ a} _{\lambda}^\dagger(\mathbf{k} )$ which, together with the annihilation operator $\hat{ a} _{\lambda}(\mathbf{k} )$, satisfies the commutation relation \begin{equation}\label{10} \bigl[ \hat{ a} _{\lambda}(\mathbf{k} ), \hat{ a} ^\dagger_{\lambda'}(\mathbf{k}' ) \bigr] = (2 \pi)^3 2 k^0 \delta_{\lambda \lambda'} \delta^{(3)}(\mathbf{k} - \mathbf{k}'). \end{equation} With this convention, the normalization condition between two plane-wave states reads \begin{equation}\label{20} \langle \mathbf{k}, \lambda | \mathbf{k}', \lambda' \rangle = (2 \pi)^3 2 k^0 \delta_{\lambda \lambda'} \delta^{(3)}(\mathbf{k} - \mathbf{k}'), \end{equation} which can be obtained directly by calculating the expectation value with respect to the vacuum state of the left side of Eq. (\ref{10}). Then the resolution of the identity can be written as \begin{equation}\label{30} 1 = |0\rangle \langle 0| + \int \tilde{d} \, \mathbf{k} \sum_{\lambda = 1}^2 |\mathbf{k}, \lambda \rangle \langle \mathbf{k}, \lambda |+ \sum \left\{ \mathrm{multiparticle} \; \mathrm{states} \right\}, \end{equation} where $\tilde{d} \, \mathbf{k}$ denotes the Lorentz-invariant measure \begin{equation}\label{40} \tilde{d} \, \mathbf{k} = \frac{d^3 \mathbf{k}}{(2 \pi)^3} \frac{1}{2 k^0}. \end{equation} The plane-wave states $| \mathbf{k}, \lambda \rangle$ are quite special since they are eigenstates of the linear momentum of the field. More generally, a single-photon state can be described by the density operator \begin{equation}\label{45} \hat{\rho} = \sum_{\lambda, \lambda'}^{1,2} \int \tilde{d} \, \bk \, \tilde{d} \, \bk' \rho_{\lambda \lambda'}(\mathbf{k}, \mathbf{k}') | \mathbf{k}, \lambda \rangle \langle \mathbf{k}', \lambda' |, \end{equation} where $\rho_{\lambda \lambda'}(\mathbf{k}, \mathbf{k}')$ is a Hermitian positive semidefinite $2 \times 2$ matrix. While the linear momentum and the polarization degrees of freedom are decoupled for photons in the plane-wave states $|\mathbf{k}, \lambda \rangle $ \cite{Aiello04c}, this is not the case for photons described by $\hat{\rho}$ which spans an $\infty^2$-dimensional vector space ($2$ polarization directions for each wave vector $\mathbf{k}$). Therefore, if one want to deal with polarization degrees of freedom only, it becomes necessary to { reduce} $\hat{\rho}$ to an {\em effective} finite-dimensional representation $\rho_{\mathrm{eff}}$ ($2 \times 2$ or $3 \times 3 $) \cite{Peres_et_al}. Now we present a derivation of $\rho_\mathrm{eff}$ which enlightens the geometrical aspects of the problem. To this end, we have to do some manipulations on the fields. In the Coulomb gauge, which is assumed to hold throughout this notes, the free field vector potential operator $\hat{\mathbf{A}}(\mathbf{r},t)$ is written \begin{equation}\label{50} \hat{\mathbf{A}}(\mathbf{r},t) = \int \tilde{d} \, \mathbf{k} \sum_{\lambda = 1}^2\bigl[ \bm \epsilon^{(\lambda)}(\mathbf{k})\hat{ a} _{\lambda}(\mathbf{k} ) e^{-i (k^0 t - \mathbf{k} \cdot \mathbf{r})} + \mathrm{h. c.} \bigr], \end{equation} where the real unit vectors $\bm \epsilon^{(\lambda)} (\mathbf{k})$ ({\em linear} polarization) satisfy the following orthogonality conditions: \begin{equation}\label{60} \bm \epsilon^{(1)}(\mathbf{k}) \cdot \bm \epsilon^{(2)}(\mathbf{k}) = 0, \qquad \bm \epsilon^{(1)}(\mathbf{k}) \times \bm \epsilon^{(2)}(\mathbf{k}) = \frac{\mathbf{k}}{k^0}. \end{equation} Since we are working with the {\em transverse} field ($A^0(\mathbf{r}, t) = 0$), we can assume an Euclidean metric in the spatial 3-dimensional space and make no distinction between high and low indices. It is useful to define $\bm \epsilon^{(3)}(\mathbf{k}) \equiv {\mathbf{k}}/{k^0}$ and to build the $\mathbf{k}$-dependent complete basis $\mathcal{E}(\mathbf{k})= \{ \bm \epsilon^{(1)}(\mathbf{k}), \bm \epsilon^{(2)}(\mathbf{k}), \bm \epsilon^{(3)}(\mathbf{k}) \}$. Let us consider now three unit vectors $\bm x, \bm y, \bm z$ of a Cartesian coordinate system or, more generally, a set $\mathcal{R}$ of three real orthogonal unit vectors $\mathcal{R} = \{ \bm e^{(1)}, \bm e^{(2)}, \bm e^{(3)} \}$: \begin{equation}\label{70} \bm e^{(a)}\cdot \bm e^{(b)} = \delta^{ab}, \qquad (a,b =1, \ldots, 3), \end{equation} which form a complete basis in the ordinary Euclidean 3-dimensional space: \begin{equation}\label{80} \sum_{a =1}^3 \bm e^{(a)} : \bm e^{(a)} = \mathbf{1} \quad \Leftrightarrow \quad\sum_{a =1}^3 e_i^{(a)} e_j^{(a)} = \delta_{ij}, \qquad (i,j =1, \ldots, 3), \end{equation} where the unit dyadic $\mathbf{1}$ has been written as a sum of dyadic products $\bm e^{(a)}: \bm e^{(a)}$ ($a = 1,2,3$). The vectors $\bm e^{(a)}$ define three orthogonal {\em spatial} orientations and they are independent from the {\em momentum} direction $\mathbf{k}/k^0$. However, for a given $\mathbf{k}$, one can write the orthogonal transformation $\Lambda(\mathbf{k})$ between the two basis $\mathcal{E}(\mathbf{k})$ and $\mathcal{R}$ as \begin{equation}\label{90} \begin{array}{rcl} \Lambda_{ab}(\mathbf{k}) & \equiv & \bm e^{(a)} \cdot \bm \epsilon^{(b)}(\mathbf{k})\\ & = & \displaystyle{\sum_{i = 1}^3 e^{(a)}_i \epsilon^{(b)}_i(\mathbf{k})}. \end{array} \end{equation} Then we can write \begin{equation}\label{100} \bm \epsilon^{(\lambda)} (\mathbf{k}) = \sum_{b =1}^3 \bm e^{(b)} \Lambda_{b \lambda} (\mathbf{k}), \qquad (\lambda = 1,2,3), \end{equation} and insert this formula in Eq. (\ref{50}) in order to obtain: \begin{equation}\label{110} \hat{\mathbf{A}}(\mathbf{r},t)= \sum_{b =1}^3 \bm e^{(b)} \hat{A}_b(\mathbf{r},t), \end{equation} where we have defined the $b$-th component of the field $\hat{\mathbf{A}}(\mathbf{r},t)$ as \begin{equation}\label{120} \begin{array}{rcl} \hat{A}_b(\mathbf{r},t) & = & \displaystyle{ \int \tilde{d} \, \mathbf{k} \sum_{\lambda = 1}^2\bigl[ \Lambda_{b \lambda} (\mathbf{k}) \hat{ a} _{\lambda}(\mathbf{k} ) e^{-i (k^0 t - \mathbf{k} \cdot \mathbf{r})} + \mathrm{h. c.} \bigr]}\\\\ & \equiv & \displaystyle{\int \tilde{d} \, \mathbf{k} \, \hat{\mathcal{A}}_b (\mathbf{k}) e^{-i (k^0 t - \mathbf{k} \cdot \mathbf{r})} + \mathrm{h. c.}} , \end{array} \end{equation} where we have defined the transformed annihilation operators $\hat{\mathcal{A}}_b(\mathbf{k} )$ ($b = 1,2,3$), as: \begin{equation}\label{130} \hat{\mathcal{A}}_b(\mathbf{k} ) \equiv \sum_{\lambda = 1}^2 \Lambda_b \/_\lambda (\mathbf{k}) \hat{ a} _{\lambda}(\mathbf{k} ), \qquad (b=1,2,3). \end{equation} It is easy to check that these operators satisfy the following commutation relations \begin{equation}\label{140} \bigl[\hat{\mathcal{A}}_a(\mathbf{k}), \hat{\mathcal{A}}_{a'}^\dagger(\mathbf{k}') \bigr] = (2 \pi)^3 2 k^0 \Delta_{a a'} \delta^{(3)} (\mathbf{k} - \mathbf{k}'), \end{equation} where we have defined the transverse Kronecker symbol $\Delta_{a a'}$ as \begin{equation}\label{150} \Delta_{aa'} \equiv \delta_{aa'} - \frac{k_a k_{a'}}{|\mathbf{k}|^2}. \end{equation} As expected, the longitudinal part $- \frac{k_a k_{a'}}{|\mathbf{k}|^2}$ of $\Delta_{aa'}$, spoils the {\em canonical} commutation relation. In the quantum theory of photo-detection it is a standard practice \cite{MandelBook} to define the positive frequency operators $\hat{V}_b(\mathbf{r},t)$ ($b = 1,2,3$) as \begin{equation}\label{160} \hat{V}_b(\mathbf{r},t) \equiv \int \tilde{d} \, \bk \sqrt{2 k^0} \hat{\mathcal{A}}_b (\mathbf{k}) e^{-i (k^0t - \mathbf{k} \cdot \mathbf{r})}, \end{equation} and such that $\sum_{b=1}^3 \hat{V}_b^\dagger(\mathbf{r},t)\hat{V}_b (\mathbf{r},t)$ represent the photon density in $(\mathbf{r},t)$. These operators can be used to build the {\em polarization correlation} operators \begin{equation}\label{170} \begin{array}{rcl} \hat{J}_{ab} & \equiv & \displaystyle{\int d^3 \mathbf{r} \, \hat{V}_a^{ \dagger} (\mathbf{r},t) \hat{V}_b}(\mathbf{r},t) \\\\ & = & \displaystyle{ \int \tilde{d} \, \bk \, \hat{\mathcal{A}}_a^{ \dagger}(\mathbf{k}) \hat{\mathcal{A}}_b(\mathbf{k})}, \end{array} \end{equation} where the last result follows immediately from Eq. (\ref{160}). The meaning of the matrix operator $\hat{J} \equiv ||\hat{J}_{ab}||$ becomes clear when we calculate its trace: \begin{equation}\label{180} \begin{array}{rcl} \mathrm{Tr} \hat{J} & = & \displaystyle{\sum_{a = 1}^3 \hat{J}_{aa} } \\\\ & = & \displaystyle{} \int \tilde{d} \, \bk \sum_{\lambda =1}^2 \hat{ a} _\lambda^\dagger (\mathbf{k}) \hat{ a} _\lambda(\mathbf{k}) \equiv \hat{N}, \end{array} \end{equation} where $\hat{N}$ denotes the photon-number operator. Now, by inserting the identity resolution Eq. (\ref{30}) in Eq. (\ref{170}), we obtain \begin{equation}\label{190} \begin{array}{rcl} \hat{J}_{ab} & = & \displaystyle{ \int \tilde{d} \, \bk \, \hat{\mathcal{A}}_a^{ \dagger}| 0 \rangle \langle 0 | \hat{\mathcal{A}}_b} + \sum \left\{ \mathrm{multiparticle} \; \mathrm{states} \right\} \\\\ & \equiv & \displaystyle{ \int_\mathcal{D} \tilde{d} \, \mathbf{k} \, | \mathbf{k}, a \rangle \langle \mathbf{k} , b | \qquad (a,b=1,2,3)}, \end{array} \end{equation} where the last equality holds {\em only} in the Hilbert spaces spanned by the one-photon states, $\mathcal{D}$ represents the set of \emph{detected} modes, and we have defined the single-photon states $ | \mathbf{k}, a \rangle$ as: \begin{equation}\label{200} \hat{\mathcal{A}}_a^{ \dagger}| 0 \rangle = | \mathbf{k}, a \rangle, \qquad (a=1,2,3). \end{equation} From Eqs. (\ref{180}-\ref{190}) it is clear that $\hat{J}_{11}, \hat{J}_{22}$ and $\hat{J}_{33}$ form a POVM (positive operator valued measure \cite{PeresBook}) in the space spanned by the one-photon states. As we saw previously, the operators $\hat{J}_{ab}$'s allow us to introduce a $3 \times 3$ correlation matrix operator \begin{equation}\label{192} \hat{J} \equiv \left( \begin{array}{ccc} \displaystyle{ \hat{J}_{11} } & \displaystyle{ \hat{J}_{21} } & \displaystyle{ \hat{J}_{31} } \\ \displaystyle{ \hat{J}_{12} } & \displaystyle{ \hat{J}_{22} } & \displaystyle{ \hat{J}_{32} } \\ \displaystyle{ \hat{J}_{13} } & \displaystyle{ \hat{J}_{23} } & \displaystyle{ \hat{J}_{33} } \end{array} \right). \end{equation} Now, for a light beam properly collimated around the direction $\bm e^{(3)} = \bm z$, the $2 \times 2$ matrix obtained by extracting the first two rows and two columns from $\mathbb{J} \equiv \langle \hat{J} \rangle $, coincides with the well known {\em coherency matrix} of the beam \cite{BornWolf}, where the bracket average $\bigl\langle \cdot \bigr\rangle$ is understood with respect to the state of the field. More generally, three independent $2 \times 2$ matrices can be extracted from $\hat{J}$: % \begin{equation}\label{193} \hat{J}^{(a)} = \left( \begin{array}{cc} \displaystyle{ \hat{J}_{bb} } & \displaystyle{ \hat{J}_{cb} } \\ \displaystyle{ \hat{J}_{bc} } & \displaystyle{ \hat{J}_{cc} } \end{array} \right), \quad \begin{array}{rcl} a,b,c &\in & \{1,2,3 \},\\ \qquad c & > & b, \\ c,b & \neq & a. \end{array} \end{equation} In principle, each $\mathbb{J}^{(a)} \equiv \langle \hat{J}^{(a)} \rangle$ can be determined by measuring the Stokes parameters of the beam (either classical or quantum), with a polarization analyzer whose axis is parallel to $\bm e^{(a)}$. For example, for a beam propagating along the axis $\bm z$, a set of {\em generalized} Hermitian Stokes operators \cite{JauchBook,Aiello04c} can be defined as \begin{equation}\label{194} \hat{S}_\mu \equiv \mathrm{Tr}\{ \sigma_{(\mu)} \hat{J}^{(3)} \}, \qquad (\mu = 0, \dots,3), \end{equation} where the $\sigma_{(\mu)}$ $(\mu=0,1,2,3)$ are the normalized Pauli's matrices \cite{Aiello04d}. Then from Eqs. (\ref{193}-\ref{194}) it readily follows \begin{equation}\label{195} \mathbb{J}^{(3)} = \sum_{\mu = 0 }^3 s_\mu \sigma_{(\mu)}, \end{equation} where $s_\mu \equiv \langle \hat{S}_\mu \rangle$. Now we are ready to accomplish our initial task by introducing the $2 \times 2$ effective reduced density matrix \begin{equation}\label{210} \rho_\mathrm{eff} \equiv \frac{\mathbb{J}^{(3)}}{\mathrm{Tr} \mathbb{J}^{(3)}}. \end{equation} It is easy to check that when the set $\mathcal{D}$ of the detected modes reduces to a single mode $\mathbf{k} \parallel \bm e^{(3)}$, the definition of $\rho_\mathrm{eff}$ above coincides with the well known polarization density matrix of a photon \cite{DauFieldRel}. More generally, by using Eq. (\ref{50}), it is possible to introduce an effective $3 \times 3$ reduced density matrix as $\rho \equiv \mathbb{J} / \mathrm{Tr} [ \mathbb{J}]$, which coincides with the one given by Peres {\em et al. }\cite{Peres_et_al}. \section*{Single-photon scattering} Let us consider now a generic scattering process which transform the initial single-photon density operator $\hat{\rho}^{\mathrm{in}}$ in the output density operator $\hat{\rho}^{\mathrm{out}}$. The most general linear transformation between $\hat{\rho}^{\mathrm{in}}$ and $\hat{\rho}^{\mathrm{out}}$, which leaves the density operator Hermitian and positive semidefinite, can be written \begin{equation}\label{n1} \hat{\rho}^{\mathrm{out}} = \sum_{A \in \mathcal{S}} p_A \mathcal{T}^{A } \hat{\rho}^{\mathrm{in}} \mathcal{T}^{A \dagger}, \end{equation} where the scattering system has been represented by the ensemble $\mathcal{S}$ of scattering matrices $\{ \mathcal{T}^{A} \}$, each of them occurring with probability $p_A \geq 0$. If we insert Eq. (\ref{45}) into Eq. (\ref{n1}) we obtain \begin{equation}\label{n2} \begin{array}{ccl} \hat{\rho}^{\mathrm{out}} & = & \displaystyle{ \sum_{\lambda, \lambda'}^{1,2} \int \tilde{d} \, \bk \, \tilde{d} \, \bk' \rho_{\lambda \lambda'}^\mathrm{in}(\mathbf{k}, \mathbf{k}') \sum_{A \in \mathcal{S}} p_A \mathcal{T}^{A} | \mathbf{k}, \lambda \rangle \langle \mathbf{k}', \lambda' |\mathcal{T}^{A \dagger} }, \end{array} \end{equation} where \begin{equation}\label{n4} \begin{array}{ccl} \displaystyle{ \sum_{A \in \mathcal{S}} p_A \mathcal{T}^{A } | \mathbf{k}, \lambda \rangle \langle \mathbf{k}', \lambda' |\mathcal{T}^{A \dagger} } & = & \displaystyle{ \sum_{\theta, \theta'}^{1,2} \int \tilde{d} \, \bq \, \tilde{d} \, \bq' \sum_{A \in \mathcal{S}} p_A |\mathbf{q} \theta \rangle \langle \mathbf{q} \theta | \mathcal{T}^{A} | \mathbf{k}, \lambda \rangle \langle \mathbf{k}', \lambda' |\mathcal{T}^{A \dagger} | \mathbf{q}' \theta' \rangle \langle \mathbf{q}' \theta' |}\\\\ & = & \displaystyle{ \sum_{\theta, \theta'}^{1,2} \int \tilde{d} \, \bq \, \tilde{d} \, \bq' \sum_{A \in \mathcal{S}} p_A \mathcal{T}^{A }_{\theta \lambda}(\mathbf{q},\mathbf{k}) \mathcal{T}^{A \dagger }_{ \lambda' \theta'}(\mathbf{k}', \mathbf{q}') |\mathbf{q} \theta \rangle \langle \mathbf{q}' \theta' |}. \end{array} \end{equation} From the equation above, it is straightforward to see that we can write \begin{equation}\label{n3} \hat{\rho}^\mathrm{out} = \displaystyle{ \sum_{\theta, \theta'}^{1,2} \int \tilde{d} \, \bq \, \tilde{d} \, \bq' \rho^\mathrm{out}_{\theta \theta'}(\mathbf{q}, \mathbf{q}') |\mathbf{q} \theta \rangle \langle \mathbf{q}' \theta' |} \end{equation} where \begin{equation}\label{n5} \rho^\mathrm{out}_{\theta \theta'}(\mathbf{q}, \mathbf{q}') = \sum_{A \in \mathcal{S}} p_A \sum_{\lambda, \lambda'}^{1,2} \int \tilde{d} \, \bk \, \tilde{d} \, \bk' \mathcal{T}^{A}_{\theta \lambda}(\mathbf{q},\mathbf{k}) \rho_{\lambda \lambda'}^\mathrm{in}(\mathbf{k}, \mathbf{k}') \mathcal{T}^{A \dagger}_{ \lambda' \theta'}(\mathbf{k}', \mathbf{q}'). \end{equation} Since $\rho^{\mathrm{t}}(\mathbf{k}, \mathbf{k}') = || \rho_{\lambda \lambda'}^{\mathrm{t}}(\mathbf{k}, \mathbf{k}') ||$ (t = in, out) are $2 \times 2$ matrices, it is always possible to express them in the complete Pauli basis as \cite{Aiello04d} \begin{equation}\label{n7} \rho^{\mathrm{t}}(\mathbf{k}, \mathbf{k}') = \sum_{\mu = 0} ^3 S_\mu^\mathrm{t}(\mathbf{k}, \mathbf{k}') \sigma_{(\mu)}, \end{equation} where we have introduced the \emph{two-mode Stokes parameters} $S_\mu^\mathrm{t}(\mathbf{k}, \mathbf{k}') = \mathrm{Tr} \{ \sigma_{(\mu)} \rho^{\mathrm{t}}(\mathbf{k}, \mathbf{k}') \}$. Since Eq. (\ref{n5}) can be written in matrix form as \begin{equation}\label{n7b} \rho^\mathrm{out}(\mathbf{q}, \mathbf{q}') = \sum_{A \in \mathcal{S}} p_A \int \tilde{d} \, \bk \, \tilde{d} \, \bk' \mathcal{T}^{A}(\mathbf{q},\mathbf{k}) \rho^\mathrm{in}(\mathbf{k}, \mathbf{k}') \mathcal{T}^{A \dagger}(\mathbf{k}', \mathbf{q}'), \end{equation} it is easy to write \begin{equation}\label{n8} \begin{array}{rcl} \displaystyle{ S^\mathrm{out}_\mu (\mathbf{q}, \mathbf{q}')} & = & \displaystyle{\sum_{\nu = 0} ^3 \int \tilde{d} \, \bk \, \tilde{d} \, \bk' S^\mathrm{in}_\nu (\mathbf{k}, \mathbf{k}') \sum_{A \in \mathcal{S}} p_A \mathrm{Tr} \left\{ \sigma_{(\mu)} \mathcal{T}^{A }(\mathbf{q},\mathbf{k}) \sigma_{(\nu)} \mathcal{T}^{A \dagger}(\mathbf{k}', \mathbf{q}')\right\}}\\\\ & \equiv & \displaystyle{ \sum_{\nu = 0} ^3 \int \tilde{d} \, \bk \, \tilde{d} \, \bk' S^\mathrm{in}_\nu (\mathbf{k}, \mathbf{k}') \sum_{A \in \mathcal{S}} p_A \, m_{\mu \nu}^A}(\mathbf{q}, \mathbf{q}'; \mathbf{k}, \mathbf{k}')\\\\ & \equiv & \displaystyle{ \sum_{\nu = 0} ^3 \int \tilde{d} \, \bk \, \tilde{d} \, \bk' M_{\mu \nu}}(\mathbf{q}, \mathbf{q}'; \mathbf{k}, \mathbf{k}') S^\mathrm{in}_\nu (\mathbf{k}, \mathbf{k}'), \end{array} \end{equation} where the four-mode density Mueller matrix \begin{equation}\label{n9} m_{\mu \nu}^A (\mathbf{q}, \mathbf{q}'; \mathbf{k}, \mathbf{k}') = \mathrm{Tr} \left\{ \sigma_{(\mu)} \mathcal{T}^{A}(\mathbf{q},\mathbf{k}) \sigma_{(\nu)} \mathcal{T}^{A \dagger} (\mathbf{k}', \mathbf{q}') \right\}, \end{equation} is defined for a single ensemble realization $A$, while \begin{equation}\label{n10} M_{\mu \nu} (\mathbf{q}, \mathbf{q}'; \mathbf{k}, \mathbf{k}') = \sum_{A \in \mathcal{S}} p_A \, m_{\mu \nu}^A (\mathbf{q}, \mathbf{q}'; \mathbf{k}, \mathbf{k}'), \end{equation} represents the ensemble-averaged four-mode density Mueller matrix. It is easy to see that when the input state is a single-mode $\mathbf{k}_0$ state: \begin{equation}\label{n20} {\rho}^\mathrm{in}_{\lambda \lambda'} (\mathbf{k} , \mathbf{k}') = \rho_{\lambda \lambda'} \delta^{(3)} (\mathbf{k} - \mathbf{k}_0)\delta^{(3)} (\mathbf{k}' - \mathbf{k}_0), \end{equation} then \begin{equation}\label{n30} S^\mathrm{in}_{\mu} (\mathbf{k} , \mathbf{k}') = s^\mathrm{in}_\mu \delta^{(3)} (\mathbf{k} - \mathbf{k}_0)\delta^{(3)} (\mathbf{k}' - \mathbf{k}_0), \end{equation} where $\rho = || \rho_{\lambda \lambda'} ||$ and $s^\mathrm{in}_\mu \equiv \mathrm{Tr}\{\sigma_{(\mu)} \rho \}$. In this case, the single-mode input and output Stokes parameters have the same functional relation as their classical counterparts: \begin{equation}\label{n40} \begin{array}{rcl} \displaystyle{ S^\mathrm{out}_\mu (\mathbf{q}_0 , \mathbf{q}_0)} & = & \displaystyle{ \sum_{\nu = 0} ^3 \int \tilde{d} \, \bk \, \tilde{d} \, \bk' M_{\mu \nu}(\mathbf{q}_0 , \mathbf{q}_0; \mathbf{k}, \mathbf{k}') s^\mathrm{in}_\nu \delta^{(3)} (\mathbf{k} - \mathbf{k}_0)\delta^{(3)} (\mathbf{k}' - \mathbf{k}_0)}\\\\ & = & \displaystyle{ \sum_{\nu = 0} ^3 M_{\mu \nu}(\mathbf{q}_0 , \mathbf{q}_0; \mathbf{k}_0, \mathbf{k}_0) s^\mathrm{in}_\nu }\\\\ & \equiv & \displaystyle{ \sum_{\nu = 0} ^3 M_{\mu \nu} s^\mathrm{in}_\nu }. \end{array} \end{equation} \begin{acknowledgments} We acknowledge support from the EU under the IST-ATESIT contract. This project is also supported by FOM. \end{acknowledgments}
{ "timestamp": "2005-03-14T13:39:17", "yymm": "0503", "arxiv_id": "quant-ph/0503124", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503124" }
\section{Introduction} Since Helmholtz \cite{helmholtz77}, it has become natural to describe a self-sustained\footnote{Self-sustained is a term indicating oscillation driven by a constant energy input.} musical instrument as an exciter coupled to a resonator. More recently, McIntyre et al.\ \cite{mcintyre83} have highlighted that simple models are able to describe the main functioning of most self-sustained musical instruments. These models rely on few equations whose implementation is not \textsc{cpu}-demanding, mainly because the nonlinearity is spatially localized in an area small compared to the wavelength. This makes them well adapted for real-time computation (including both transient and steady states). These models are particularly popular in the framework of sound synthesis. On the other hand, calculation in the frequency domain is suitable for determining periodic solutions of the model (the values of the harmonics as well as the playing frequency) for a given set of parameters. Such information can be provided by an iterative method named the harmonic balance method (HBM). Though the name ``harmonic balance'' seems to date back to 1936 \cite{krylov36}, the method was popularized nearly forty years ago for electrical and mechanical engineering purposes, first for forced vibrations \cite{urabe65}, later for auto-oscillating systems \cite{stokes72}. The modern version was presented rather shortly after by Nakhla and Vlach \cite{nakhla76}. In 1978, Schumacher was the first one to use the HBM for musical acoustics purposes with a focus on the clarinet \cite{schumacher78}. However in this paper, the playing frequency is not determined by the HBM. This shortcoming is the major improvement brought by Gilbert et al.\ \cite{gilbert89} eleven years later, who proposed a full study of the clarinet including the playing frequency as an unknown of the problem. The fact that the HBM can only calculate periodic solutions, may seem as a drawback. Certainly, transients such as the attack are impossible to calculate, and the periodic result is boring to listen to and does not represent the musicality of the instrument. Therefore the HBM is definitely not intended for sound synthesis. Nevertheless, self-sustained musical instruments are usually used to generate harmonic sounds, which are periodic by definition. The HBM is thus very useful to investigate the behavior of a physical model of an instrument, depending on its parameter values. This is possible for both stable and unstable solutions, without care of precise initial conditions. Moreover, HBM results can be compared to approximate analytical calculations (like the variable truncation method (VTM) \cite{kergomard00}), in order to check the validity of the approximate model considered. The present paper is based on the work of Gilbert et al \cite{gilbert89}. Our main contributions are: extension of the diversity of equations managed, improved convergence of the method, introduction of basic continuation facilities, and from a practical point of view, faster calculations. While the main idea is already described by Gilbert et al.\ \cite{gilbert89}, Section~\ref{s:nummeth} details the principle of the HBM, in particular the discretization of the problem, both in time and frequency. Section~\ref{s:harmbal} is devoted to the various contributions of the current work, which are applied in a computer program called Harmbal \cite{harmbal}. The framework is defined to include models with three equations: two linear differential equations, written in the frequency domain, and a nonlinear coupling equation in the time domain (see Sec.~\ref{s:self-sustained}). As usual in the HBM, this system of three equations is solved iteratively. The solving method chosen (Newton-Raphson, Sec.~\ref{s:harmdet}) has been investigated and its convergence has been improved through a backtracking scheme (Secs.~\ref{s:holes} and~\ref{s:backtracking}). To illustrate the advantages of the HBM and the improvements, a few case studies were performed and are presented in Section~\ref{s:case}. They are based on a classical model of single reed instruments which is presented in Section~\ref{s:clarinet}. In Sections~\ref{s:verif} and further, simplifications to each of the three equations are introduced so that the results could be compared to analytical calculations, both for cylindrical and stepped-cones bores. Finally the full model is compared to time-domain simulations. This also shows the modularity of Harmbal. The comparison is achieved through the investigation of bifurcation diagrams as the dimensionless blowing pressure is altered. The derivation of a branch of solution is obtained thanks to basic continuation with an auto-adaptative parameter step. Finally, various questions are tackled through practical experience from using Harmbal. Section~\ref{s:practexp} discusses multiplicity of solutions and poor robustness in the frequency estimation. \section{Numerical method}\label{s:nummeth} \subsection{The harmonic balance method}\label{s:HBM} The harmonic balance method is a numerical method to calculate the steady-state spectrum of periodic solutions of a nonlinear dynamical system. In this paper we are only concerned with periodic solutions. The following provides a detailed and general description of the method for a nonlinearly coupled exciter-resonator system. Let $X(\omega_k)$, $k = 0, \dots, N_t-1$ be the Discrete Fourier transform (DFT) of one period $x(t)$, $0\le t<T$, of a $T$-periodic solution of a mathematical system to be defined. $X(\omega_k)$ will have a number of complex components $N_t$, which depends on the sampling frequency $f_s=1/T_s$ with which we discretize $x(t)$ into $N_t=T/T_s$ equidistant samples. Furthermore, $\omega_k{=}2\pi f_p T_s k$ is the angular frequency of each harmonic of the fundamental frequency $f_p$ of the oscillation, referred to as the {\em playing frequency}. Note that the sampling frequency $f_s=N_tf_p$ is automatically adjusted to the current playing frequency so that we always consider one period of the oscillation while keeping $N_t$ constant. Note also that $N_t$ should be sufficiently large to avoid aliasing. Moreover, if it is chosen a power of two, the Fast Fourier transform (FFT) may be used. Assuming that $N_p<N_t/2$ harmonics is sufficient to describe the solution, we define $\vec X\in\mathbb{R}^{2N_p+2}$ as the $N_p+1$ first real components (denoted by $\Re$) of $X(\omega_k)$ followed by their imaginary components ($\Im$): \begin{align} \vec X=&\left[\Re\left(X(\omega_0)\right), \dots,\Re\left(X(\omega_{N_p})\right),\right.\\ &\qquad\left.\Im\left(X(\omega_0)\right), \dots, \Im\left( X(\omega_{N_p})\right)\right].\nonumber \end{align} Note that the components $X_0$ and $X_{N_p+1}$ are the real and imaginary DC components respectively (and that $X_{N_p+1}$ is always zero). Our mathematical system can thus be defined by the nonlinear function $F: \mathbb{R}^{2N_p+3} \to \mathbb{R}^{2N_p+2}$: \begin{equation} \vec X = \vec F(\vec X,f_p). \label{e:gensyst} \end{equation} Until now, the playing frequency has silently been assumed to be a known quantity. In autonomous systems, however, the frequency is an additional unknown, so that the $N_p$-harmonic solution seeked is defined by $2N_p{+}3$ unknowns linked through the $2N_p{+}2$ equations~(\ref{e:gensyst}). However, it is well known that as $\vec X$ is a periodic solution of a dynamical system, any $\vec X^{\prime}$ deduced from $\vec X$ by a phase rotation (i.e.\ a shift in the time domain) is also a solution. Thus an additional constraint has to be added in order to select a single periodic solution among the infinity of phase-rotated solutions. A common choice (see Ref.\ \onlinecite{gilbert89}) is to consider the solution for which the first harmonic is real (i.e.\ its imaginary part, $X_{N_p+2}$, is zero). This additional constraint decreases the number of unknowns to $2N_p{+}2$ for an $N_p$-harmonic periodic solution. Thus we get $\vec F: \mathbb{R}^{2N+2} \to \mathbb{R}^{2N+2}$, and it is now possible to find periodic solutions, if they exist. Finally, a simple way of avoiding trivial solutions to equation~(\ref{e:gensyst}) is to look for roots of the function $\vec G: \mathbb{R}^{2N_p+2} \to \mathbb{R}^{2N_p+2}$, defined by \begin{equation} \vec G(\vec X,f_p)=\frac{\vec X-\vec F(\vec X,f_p)}{X_1}, \label{e:G} \end{equation} i.e.\ $\vec G(\vec X,f_p)=0$. This equation is usually solved numerically through an iteration process, for instance by the Newton-Raphson method as in our case. How to handle the playing frequency $f_p$ will be discussed in the following section. \subsection{Iteration by Newton-Raphson}\label{s:newton} The equation $\vec G(\vec X,f) = 0$, $\vec G$ being defined by equation~\eqref{e:G}, is nonlinear and has usually no analytical solution. (For readability we leave out the index $p$ on the playing frequency until end of Sec.~\ref{s:harmbal}.) This section describes the common, iterative Newton-Raphson method. This is the method used in the program Harmbal (see Section~\ref{s:harmbal}) although it had to be refined with a backtracking procedure to improve its convergence, as discussed in Section~\ref{s:backtracking}. For the sake of later reference, it is useful to re\-col\-lect the principles of Newton's method for a one-dimensional problem $g(x)=0$. Starting with an estimate $x^0$ of the solution, the next estimate $x^1$ is defined as the intersection point between the tangent to $g$ at $x_0$ and the $x$-axis. The method can be summarized as \begin{equation} x^{i+1}=x^i-\frac{g(x^i)}{g'(x^i)}. \end{equation} This is repeated, as shown in Figure~\ref{f:newton}, while increasing the iteration index $i$ until $g(x^i)<\varepsilon$, where $\varepsilon$ is a user-defined threshold value. \begin{figure} \ifgalleyfig \includegraphics[width=1.95in]{eps/newton_g.eps \else \ifoutputfig \includegraphics[width=5.5in]{eps/newton_g.eps \fi \fi \caption [The iteration process of Newton's method] {\label{f:newton} \ifgalleyfig {The iteration process of Newton's method} \fi} \end{figure} In our $2N_p{+}2$-dimensional case, we have a vector problem: we search $(\vec X,f)$ for which $\vec G(\vec X,f)=0$. Newton's method is generalized to the Newton-Raphson method, which may be written \cite{numrec}: \begin{equation} (\vec X^{i+1}, f^{i+1})=(\vec X^i , f^i ) -\left(\mathbf{J}_G^i\right)^{-1}\!\!\cdot \vec G(\vec X^i,f^i), \label{e:fullstep} \end{equation} where $\mathbf{J}^i_G{\triangleq}\nabla G(\vec X^i,f^i)$ is the {\em Jacobian\/} matrix of $\vec G$ at $(\vec X^i,f^i)$. Note that all derivatives by $X_{N_p+2}$, which was chosen to be zero, are ignored. The column $N_p{+}2$ in the Jacobian is thus replaced by the derivatives with respect to the playing frequency $f$. $\mathbf{J}^i_G$ is thus a $(2N_p{+}2)$-square matrix. This means that line number $N_p{+}2$ in equation~\eqref{e:fullstep} gives the new frequency $f$ instead of $X_{N_p+2}$. We define the {\em Newton step\/} $\Delta\vec X{=}\vec X^{i+1}{-}\vec X^i$ (where $\Delta f{=}f^{i+1}{-}f^i$ replaces $\Delta X_{N_p+2}$), which follows the local steepest descent direction. The Jacobian may be found analytically if $\vec G$ is given analytically, but it is usually sufficient to use the first-order approximation \begin{equation} J_{jk}=\frac{\partial G_j}{\partial X_k} \simeq\frac{G_j(\vec X+\delta\vec X_k,f)-G_j(\vec X,f)}{\delta X}, \label{e:jacobij} \end{equation} except for $k=N_p+2$, in which case we use \begin{equation} J_{j,N+2}=\frac{\partial G_j}{\partial f} \simeq\frac{G_j(\vec X,f+\delta f)-G_j(\vec X,f)}{\delta f}. \label{e:jacobim} \end{equation} The components of $\delta\vec X_k$ are zero except for the $k$th one, which is the tiny perturbation $\delta X$. The iteration has converged when $|\vec G^i|{\triangleq}|\vec G(\vec X^i,f^i)|<\varepsilon$. We found $\varepsilon=10^{-5}$ to be a good compromize between computation time and solution accuracy. \section{Implementation and Harmbal}\label{s:harmbal} \subsection{Equations for self-sustained musical instruments \label{s:self-sustained} } Though, to the authors' knowledge, the harmonic balance method in the context of musical acoustics with unknown playing frequency has only been applied to study models of clarinet-like instruments, it should be possible to consider many different classes of self-sustained instruments. It is well accepted that sound production by a musical instrument results from the interaction between an exciter and a resonator through a nonlinear coupling. Moreover, in most playing conditions, linear modelling of both the exciter and the resonator is a good approximation. Therefore, within these hypotheses, any musical instrument could be modelled by the following three equations: \begin{equation} \qquad\quad\left\{ \begin{array}{l@{\qquad}c} Z_e(\omega) X_e(\omega) = X_c(\omega)& \mathrm{(a)}\\ \phantom{Z_e(\omega)}X_c(\omega) = Z_r(\omega) X_r(\omega) & \mathrm{(b)}\\ \mathcal{F}(x_c(t),x_e(t),x_r(t)) =0 & \mathrm{(c)} \end{array} \right. \label{e:any_instr} \end{equation} where $Z_e$ is the dynamic stiffness and $Z_r$ is the input impedance of the exciter and the resonator, respectively, and $X_e$ and $X_r$ are the spectra describing the dynamics of the exciter and the resonator during the steady state (periodicity assumption). $X_c$ is the spectrum of the coupling variable. All these quantities, and thus equations~(\ref{e:any_instr}a--b), are defined in the Fourier domain. Equation~(\ref{e:any_instr}c) is written in the time domain, where $\mathcal{F}$ is a nonlinear functional of $x_c$, $x_e$, and $x_r$, which are the inverse Fourier transforms of $X_c$, $X_e$, and $X_r$, respectively. We apply the discretization as described in Section~\ref{s:HBM}, implying that equations~(\ref{e:any_instr}a--b) become vector equations where the impedances must be written as real $(2N_p{+}2){\times}(2N_p{+}2)$-matrices to accommodate the rules of complex multiplication: \begin{equation} Z(f)=\left( \begin{array}{cc} \Re(\tilde{Z}(f))&-\Im(\tilde{Z}(f))\\ \Im(\tilde{Z}(f))&\Re(\tilde{Z}(f))\\ \end{array} \right) \label{e:impmat1} \end{equation} where \begin{equation} \tilde{Z}(f)=\left( \begin{array}{cccc} Z(0) &0&\cdots&0\\ 0& Z(\omega_1)&&0\\ \vdots&&\ddots&\vdots\\ 0& 0 &\cdots&Z(\omega_{N_p}) \end{array} \right) \label{e:impmat2} \end{equation} is complex, and $\Re(\tilde{Z})$ and $\Im(\tilde{Z})$ are the real and imaginary components of $\tilde{Z}$. The system~(\ref{e:any_instr}) is solved iteratively by Harmbal according to the scheme illustrated in Figure~\ref{f:any_instr_solve}. \begin{figure} \centerline% {\setlength{\unitlength}{1em}% \def$\vec X_r,f$}\def\Xc{$\vec X_c,f$}\def\Xe{$\vec X_e,f${$\vec X_r,f$}\def\Xc{$\vec X_c,f$}\def\Xe{$\vec X_e,f$} \def$x_r(t)$}\def\xc{$x_c(t)$}\def\xe{$x_e(t)${$x_r(t)$}\def\xc{$x_c(t)$}\def\xe{$x_e(t)$} \begin{picture}(22.5,12)(-1.5,-0.5) \put(0,10){\Xc} \put(1.8,9.3){\vector(1,-1){2}} \put(3,9){\footnotesize Eq.(\ref{e:any_instr}a)} \put(3,6.2){\Xe} \put(6,6.5){\vector(1,0){2.5}} \put(6,6.9){\footnotesize DFT$^{-1}$} \put(9,6.2){\xe} \put(3,10.3){\vector(1,0){5.5}} \put(4.7,10.7){\footnotesize DFT$^{-1}$} \put(9,10){\xc} \put(11.3,8){$\left.\hbox{\vbox to2.8\unitlength{}}\right\}$} \put(12.8,8.3){\vector(1,0){2.5}} \put(12.6,8.7){\footnotesize Eq.(\ref{e:any_instr}c)} \put(16,8){$x_r(t)$}\def\xc{$x_c(t)$}\def\xe{$x_e(t)$} \put(17,7.5){\vector(0,-1){1.5}} \put(17.5,6.5){\footnotesize DFT} \put(15.6,5){$\vec X_r,f$}\def\Xc{$\vec X_c,f$}\def\Xe{$\vec X_e,f$} \put(17,4.5){\vector(0,-1){1.5}} \put(17.5,3.5){\footnotesize Eq.(\ref{e:any_instr}b)} \put(15,2){$\vec F$(\Xc)} \put(14.7,2.3){\vector(-1,0){3}} \put(1,9.3){\line(0,-1){7}} \put(1,2.3){\vector(1,0){8.7}} \put(10.7,2.3){\circle{2}} \put(10.1,2){=?} \put(10.7,1.3){\line(0,-1){1}} \put(10.7,0.3){\vector(-1,0){5.5}} \put(1,0){$\Delta\vec X_c,\Delta f$} \put(7,.6){\small N-R} \put(0.7,0.3){\line(-1,0){1.7}} \put(-1,0.3){\line(0,1){8.5}} \put(-1,8.8){\vector(1,1){1}} \end{picture}}} \caption{The iteration loop of the harmonic balance method for a musical instrument (notations defined in the text)} \label{f:any_instr_solve} \end{figure} In Harmbal, these equations are easily defined by writing new C functions. Only superficial knowledge of the C language is necessary to do this. Three cases related to models of single reed instruments with cylindrical or stepped-conical bores are studied in particular in Section~\ref{s:case} in order to validate the code and to illustrate the modularity of Harmbal. \subsection{Practical characteristics of Harmbal \label{s:harmdet}} Both fast calculation, good portability, and independence of commercial software are easily achieved by programming in C, whose compiler is freely available for most computer platforms. It is, however, somewhat difficult to combine portability with easy usage, because an intuitive usage normally means a graphical and interactive user interface, while the handling of graphics varies a lot between the different platforms. We have chosen to write Harmbal with a nongraphical and non-interactive\footnote{The term {\em non-interactive\/} means that the user has no influence on the program while it is running.} user interface. The major advantage of this is that independent user interfaces may be further developed depending on need. Our concept is to save both the parameters and the solution in a single file. This file also serves as input to Harmbal while individual parameters can be changed through start-up arguments. The solution provided by the file works as the initial condition for the harmonic balance method. Thus the lack of a simple user interface is compensated by a simple way of re-using an existing solution to solve the system for a slightly different set of parameters. Solutions for a range of a parameter values may thereby be calculated by changing the parameter stepwise and providing the previous solution as an initial condition for the next run. The Perl script {\em hbmap\/} provides such zeroth-order continuation facilities. This procedure may also be used when searching for a solution where it is difficult to provide a sufficiently good initial condition, for instance by successively increasing $N_p$ when wanting many harmonics. \subsection{Convergence of Newton-Raphson \label{s:holes}} When merely employing the Newton-Raphson method to determine the solution of the system at a given set of parameters, we have found that it is impossible to find a solution at particular combinations of the parameters. Indeed, for the clarinet model of Section~\ref{s:helcyl}, no convergence was obtained for particular values of the parameter $\gamma$ (the dimensionless blowing pressure) and its neighborhood. This is seen as discontinuities, or \emph{holes}, in the curves in Figure~\ref{f:holes} (see Section~\ref{s:case} for the underlying equations and parameters). Note that the solutions seem to go continuously through this hole and that the positions of the holes and their extent vary with the number of harmonics $N_p$ taken into account. \begin{figure \ifgalleyfig% \includegraphics[width=3.25in]{eps/compmap.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/compmap.eps}% \fi \fi \caption [Solution holes: first pressure harmonic $P_1$ versus blowing pressure $\gamma$ for different $N_p$ with $N_t=128$, $\zeta=0.5$, and $\eta=10^{-3}$. (Even $N_p$ give the same as $N_p{-}1$.) Equations and parameters are defined in section \ref{s:case}.] {\label{f:holes} \ifgalleyfig {Solution holes: first pressure harmonic $P_1$ versus blowing pressure $\gamma$ for different $N_p$ with $N_t=128$, $\zeta=0.5$, and $\eta=10^{-3}$. (Even $N_p$ give the same as $N_p{-}1$.) Equations and parameters are defined in section \ref{s:case}.} \fi} \end{figure} The curves were calculated by the program {\em hbmap}. In this case we have decreased $\gamma$ from 0.5 downward in steps of $10^{-4}$ and drawn a line between them except across $\gamma$ values where solution failed. In the holes, the Newton-Raphson method did not converge, either by alternating between two values of $\vec P$ (i.e.\ $\vec X_c$) or by starting to diverge. To study the problem, we simplified the system to a one-dimensional problem by setting $N_p=1$, thus leaving $P_1$ as the only nonzero value. $G_1$ thus became the only contributor to $|\vec G|$, and a simple graph of $G_1$ around the solution $G_1=0$ could illustrate the problem, as shown in Figure~\ref{f:GvsP}. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/jumphole128.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/jumphole128.eps}% \fi \fi \caption [$G_1$ as $P_1$ varies around the solution $G_1=0$ for various $\gamma$ around a hole at $\gamma\simeq0.4196$. $N_t=128$ and $N_p=1$.] {\label{f:GvsP} \ifgalleyfig {$G_1$ as $P_1$ varies around the solution $G_1=0$ for various $\gamma$ around a hole at $\gamma\simeq0.4196$. $N_t=128$ and $N_p=1$.} \fi} \end{figure} We see that the curve of $G_1(P_1)$ has inflection points (visible as ``soft steps'' on the curve) at rather regular distances. At the centre of a convergence hole, i.e.\ for $\gamma\simeq0.4196$, an inflection point is located at the intersection with the horizontal axis. This is a school example of a situation where Newton's method does not converge because the Newton step $\Delta P_1$ brings us alternatingly from one side of the solution to the other, but not closer. In fact, the existence of inflection points is linked with the digital sampling of the continuous signal. If the sampling rate is increased, i.e.\ if $N_t$ is increased, the steps become smaller but occur more frequently, as shown for $N_t=32$, 128, and 1024 in Figures~\ref{f:sampling}a--c. The derivative $dG_1/dP_1$ is included in the figures to quantify the importance of the steps. According to the Figures~\ref{f:sampling}a--c it seems reasonable to increase $N_t$ to avoid convergence problems. However, this would significantly increase the computational cost. Another solution is therefore suggested in the following. \begin{figure*} \ifgalleyfig% \includegraphics[height=.595\width]{eps/jumps32.eps}\hspace{-.21\width}% \includegraphics[height=.595\width]{eps/jumps128.eps}\hspace{-.21\width}% \includegraphics[height=.595\width]{eps/jumps1024.eps}\\%}% \else \ifoutputfig% \includegraphics[height=.285\width]{eps/jumps32.eps}\hspace{-.1\width}% \includegraphics[height=.285\width]{eps/jumps128.eps}\hspace{-.1\width}% \includegraphics[height=.285\width]{eps/jumps1024.eps}% \fi \fi \caption [The effect of sampling rate on the ``smoothness'' of $G_1(P_1)$: (a) $N_t=32$, (b) 128, and (c) 1024. The derivative $dG_1/dP_1$ exhibits the ``roughness''.] {\label{f:sampling} \ifgalleyfig {The effect of sampling rate on the ``smoothness'' of $G_1(P_1)$: (a) $N_t=32$, (b) 128, and (c) 1024. The derivative $dG_1/dP_1$ exhibits the ``roughness''.} \fi} \end{figure*} \subsection{Backtracking}\label{s:backtracking} When the Newton-Raphson scheme fails to converge, it often happens because the Newton step $\Delta\vec X$ leads to a point where $|\vec G(\vec X,f)|$ is larger than in the previous step. However, acknowledging that the Newton step points in the direction of the steepest descent, there must be a point along $\Delta\vec X$ where $|\vec G(\vec X,f)|$ is smaller than in the previous iteration of the HBM. A backtracking algorithm described in Numerical Recipes \cite[Sec.9.7]{numrec} solves the problem elegantly by shortening the Newton step as described here. The principle is illustrated in the simple one-dimensional case in Figure~\ref{f:backtracking}, where $g(x)$ replaces $|\vec G(\vec X,f)|$, although we use the multidimensional notation in the following. \begin{figure} \ifgalleyfig% \includegraphics[width=2.4in]{eps/backtrack3.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/backtrack3.eps}% \fi \fi \caption [The principle of backtracking in one dimension. Objective is to estimate the root of $g(x)$ (solid curve). Broken lines with arrows show how the Newton step $\Delta x$ from $x$ leads to divergence. $h(\lambda)$ (dot-dashed curve) is a 2nd order expansion of $g(x)$ along the Newton step, i.e. the $\lambda$ axis. Minimum of $h(\lambda)$ should be closer to the root of $g(x)$ than $g(x+\Delta x)$.] {\label{f:backtracking} \ifgalleyfig {The principle of backtracking in one dimension. Objective is to estimate the root of $g(x)$ (solid curve). Broken lines with arrows show how the Newton step $\Delta x$ from $x$ leads to divergence. $h(\lambda)$ (dot-dashed curve) is a 2nd order expansion of $g(x)$ along the Newton step, i.e. the $\lambda$ axis. Minimum of $h(\lambda)$ should be closer to the root of $g(x)$ than $g(x+\Delta x)$.} \fi} \end{figure} Defining the $\lambda$ axis along the Newton step, we simply take a step $\lambda\Delta\vec X$ in the same direction, where $0<\lambda<1$. The optimal value for $\lambda$ is the one that minimizes the function $h(\lambda)$: \begin{equation} h(\lambda)=\textstyle\frac12|\vec G(\vec X^i+\lambda\Delta\vec X)|^2 \end{equation} with derivative \begin{equation} h'(\lambda)=\left(\mathbf{J}_G \cdot \vec G\right)\big| _{\vec X^i+\lambda\Delta\vec X} \cdot\Delta\vec X. \end{equation} During the calculation of the failing Newton step, we computed $\vec G(\vec X^i)$ and $\vec G(\vec X^{i+1})$, so now it is possible to calculate with nearly no additional computational effort $h(0) = \frac12|\vec G(\vec X^i)|^2$, $h'(0) = -|\vec G(\vec X^i)|^2$, and $h(1) = \frac12|\vec G(\vec X^i+\Delta\vec X)|^2 = \frac12|\vec G(\vec X^{i+1})|^2$. This allows to propose a quadratic approximation of $h$ for $\lambda$ between $0$ and $1$, for which the minimum is located at \begin{equation} \lambda_1=-\frac{\frac12h'(0)}{h(1)-h(0)-h'(0)}. \end{equation} It can be shown that $\lambda_1$ should not exceed 0.5, and in practice $\lambda_1\ge0.1$ is required to avoid a too short step at this stage. If $|\vec G(\vec X^i+\lambda_1\Delta\vec X)|$ still is larger than $|\vec G(\vec X^i)|$, $h(\lambda)$ is then modelled as a cubic function (using $h(\lambda_1)$ which has just been calculated). The minimum of this cubic function gives a new value $\lambda_2$, again restricted to $0.1\lambda_1<\lambda_2<0.5\lambda_1$. This calculation requires solving a system of two equations, so if also $\lambda_2$ is not accepted because $|\vec G(\vec X^i+\lambda_2\Delta\vec X)|$ is still too large, we do not enhance to a fourth-order model of $h$, which would increase the computational cost much more. Instead, subsequent cubic modellings are performed using the most two recent values of $\lambda$. In practice, however, not many repetitions should be necessary before finding a better solution, if possible. \section{Case studies}\label{s:case} \subsection{The equations for the clarinet \label{s:clarinet}} The three equations~(\ref{e:any_instr}a--c) may be constructed by physical modelling. In the case of the clarinet, a common simple model is described below. We limit the description in the following to a brief presentation based on dimensionless quantities, {\em dimensional\/} variables being denoted by a hat ($\hat{\ }$) hereafter (see Fritz et al.\ \cite{fritz04} for further details). \medskip The exciter is an oscillating reed which may be modelled as a spring with mass and damping: \begin{equation} \ddot{\hat{y}}+g_e\dot{\hat{y}}+\omega_e^2\hat{y} =\frac1{\mu_e}(\hat{p}-p_m), \label{e:lindiff_dim} \end{equation} where $\hat{y}$ is the dynamic reed displacement, and $\hat{p}$ and $p_m$ are the dynamic pressure in the mouthpiece, i.e.\ the {\em internal\/} pressure, and the static blowing pressure in the player's mouth, respectively. The constants $\mu_e$, $g_e$, and $\omega_e$ represent the mass per area, the damping factor, and the angular resonance frequency of the exciter (the reed). The dots over $\hat{y}$ denote the time derivative. In dimensionless form, equation~\eqref{e:lindiff_dim} becomes \begin{equation} M\ddot{x}+R\dot{x}+Kx=p, \label{e:lindiff} \end{equation} where $p=\hat p/p_M$ and $x=\hat y/H+\gamma/K$ with $\gamma=p_m/p_M$. The equilibrium reed opening is $H$ as shown in Figure~\ref{f:mouthpiece}. In the static regime, when blowing harder than a maximum pressure $p_M$, i.e.\ $p_m\ge p_M$ ($\gamma\ge1$), the reed blocks the opening, i.e.\ $\hat y=-H$, so we get $\hat p=0$ and can conclude that $K=1$ for the current reed model. \begin{figure} \ifgalleyfig% \includegraphics[width=1.9in]{eps/mouthpiece.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/mouthpiece.eps \fi \fi \caption [Illustration of the mouthpiece] {\label{f:mouthpiece} \ifgalleyfig {Illustration of the mouthpiece} \fi} \end{figure} Like Fritz et al.\ \cite{fritz04} we relate the dimensionless time to the resonance angular frequency $\omega_r$ of the resonator (the bore), i.e.\ $t=\omega_r\hat t$, so that the values of the dimensionless mass $M$, damping $R$, and spring constant $K$ become \begin{align} K&=\mu_e H\omega_e^2/p_M=1,\\ R\,&=Kg_e\omega_r/\omega_e^2=g_e\omega_r/\omega_e^2,\\ M&=K\omega_r^2/\omega_e^2=\omega_r^2/\omega_e^2. \label{e:MRK} \end{align} In the Fourier domain, Equation~\eqref{e:lindiff} thus takes the form of equation~(\ref{e:any_instr}a), $Z_e(\omega)X(\omega)=P(\omega)$, where \begin{equation} Z_e(\omega)=1-M\left(\!\frac\omega{2\pi}\!\right)^{\!2} +iR\left(\!\frac\omega{2\pi}\!\right), \label{e:reedimp} \end{equation} for $i=\sqrt{-1}$ and $\omega =2\pi\hat\omega/\omega_r =\hat\omega/f_r$ is the dimensionless angular frequency in the Fourier domain. A common minimum model for the clarinet assumes a simple reed with no mass or damping, thus $M=R=0$. Equation~\eqref{e:lindiff} reduces to $x=p$. \medskip The resonator (i.e.\ the air column in the bore of the instrument) is commonly described by its frequency response $\hat Z_r(\hat\omega)$. For a simple cylindrical bore of length $l$ with a closed and an ideal open end, the resonance frequencies are odd multiples of $f_r=c/4l$, $c$ being the sound speed in the air column \cite{fletcher91}. The input impedance of the bore may thus be expressed in dimensionless quantities as \begin{equation} Z_r(\omega)=\frac{\hat Z_r(\hat\omega)}{Z_0} =i\tan\!\left(\frac\omega4 + (1-i)\alpha(\omega)\!\right), \label{e:freqresponse} \end{equation} where $\alpha(\omega)\triangleq\psi\eta\sqrt{\omega/2\pi}$ with $\psi\simeq1.3$ for common conditions in air and $\eta$ being the dimensionless loss parameter, which depends on the tube length, typically 0.02 for a normal clarinet with all holes closed. $Z_0\triangleq \rho c/S$ is the characteristic impedance of the cylidrical resonator, $S$ being its cross section, and $\rho$ the density of air. The last term in the argument of equation~\eqref{e:freqresponse} includes the dispersion as the real part and viscous losses as the imaginary part. Equation~(\ref{e:any_instr}b) becomes \begin{equation} P(\omega)=Z_r(\omega)U(\omega), \label{e:imped} \end{equation} where $P(\omega)$ and $U(\omega)\triangleq\hat U(\omega)Z_0/p_M$ are the dimensionless internal pressure and volume flow of air through the mouthpiece in the Fourier domain. \medskip The coupling equation~(\ref{e:any_instr}c), is given by the Bernoulli theorem with some supplementary hypotheses applied between the mouth and the outlet of the reed channel. The coupling equation is nonlinear and must be calculated in the time domain. This leads to the following expression for the dimensionless airflow through the mouthpiece \cite{kergomard95}: \begin{equation} u(p,x)=\zeta\left(1+x-\gamma\right) \sqrt{|\gamma-p|}\,\mathrm{sign}(\gamma-p) \label{e:nonlin} \end{equation} as long as $x>\gamma-1$, and $u=0$ otherwise. $\zeta=Z_0wH\sqrt{2/\rho p_M}$ is a dimensionless embouchure parameter roughly describing the mouthpiece and the position of the player's mouth, $w$ being the width of the opening and $\rho$ the density of the air. $\zeta$ is also related to the maximum volume velocity entering the tube \cite{ollivier04a}. If the reed dynamics were not taken into account, we had $x=p$ and thus \begin{equation} u(p)=\zeta\left(1+p-\gamma\right) \sqrt{|\gamma-p|}\,\mathrm{sign}(\gamma-p) \label{e:simplenonlin} \end{equation} for $p>\gamma-1$, and, as before, $u=0$ otherwise. \subsection{Verification of method and models \label{s:verif}} In the following we want to verify that the HBM (and its implementation in Harmbal) gives correct results. By using very low losses in the resonator (small $\eta$) we can compare the results of the HBM with analytical results. Rising the attenuation in the resonator and including mass and damping for the exciter, we compare with numerical results from real-time synthesis of the same system. This also gives us the opportunity to illustrate the modularity of Harmbal as we change the models of the resonator and the nonlinear coupling. \subsubsection{Helmholtz oscillation for cylindrical tubes}\label{s:helcyl} To compare the HBM results with analytical results, we assume a nondissipative, nondispersive air column, i.e.\ setting $\eta=0$ and thus $\alpha=0$ in equation~\eqref{e:freqresponse}. Furthermore, we assume that the reed has neither mass nor damping and thus use equation~\eqref{e:simplenonlin}. The resulting square-wave amplitude (the Helmholtz motion \cite{helmholtz77}) may be found by solving $u(p)=u(-p)$, which results from the fact that the internal pressure $p(t)$ and the power $p(t)u(t)$ averaged over a period are zero according to the lossless hypothesis \cite{kergomard95}. This leads to the square oscillation with amplitude \begin{equation} p(\gamma)=\sqrt{-3\gamma^2+4\gamma-1}. \label{e:helmotion} \end{equation} This result is compared with the results calculated by Harmbal (for the same set of equations, but $\eta=10^{-5}$ instead of $\eta=0$ to avoid infinite impedance peaks) for 3, 9, 49, and 299 harmonics close to the oscillation threshold in Figure~\ref{f:nearthres-nl}, and at $\gamma=0.4$ in Figure~\ref{f:largeosc-nl}, which is far from the threshold. \begin{figure}[t] \ifgalleyfig% \includegraphics[width=3.25in]{eps/HBM-hh-freq-g0.334336.eps}% \\\includegraphics[width=3.25in]{eps/HBM-hh-time-g0.334336.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/HBM-hh-freq-g0.334336.eps}% \\\includegraphics[width=5.5in]{eps/HBM-hh-time-g0.334336.eps}% \fi \fi \caption [The Helmholtz solution, eq.\ \eqref{e:helmotion} compared with the HBM truncated to 3, 9, 49, and 299 harmonics close to the oscillation threshold ($\gamma=0.334$, $\zeta=0.5$, $\eta=10^{-5}$). (a) frequency domain. (b) time domain.] {\label{f:nearthres-nl} \ifgalleyfig {The Helmholtz solution, eq.\ \eqref{e:helmotion} compared with the HBM truncated to 3, 9, 49, and 299 harmonics close to the oscillation threshold ($\gamma=0.334$, $\zeta=0.5$, $\eta=10^{-5}$). (a) frequency domain. (b) time domain.} \fi} \end{figure} \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/HBM-hh-freq-g0.40.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/HBM-hh-freq-g0.40.eps}% \fi \fi \caption [The Helmholtz solution, eq.\ \eqref{e:helmotion} compared with the HBM truncated to 3, 9, 49, and 299 harmonics far from the oscillation threshold ($\gamma=0.40$, $\zeta=0.5$, $\eta=10^{-5}$) in the frequency domain.] {\label{f:largeosc-nl} \ifgalleyfig {The Helmholtz solution, eq.\ \eqref{e:helmotion} compared with the HBM truncated to 3, 9, 49, and 299 harmonics far from the oscillation threshold ($\gamma=0.40$, $\zeta=0.5$, $\eta=10^{-5}$) in the frequency domain.} \fi} \end{figure} As expected, the solution using the HBM shows good convergence towards the Helmholtz motion as the number of harmonics increases. Note the deviation for higher harmonics close to the threshold, even for 299 harmonics. Dissipation in the resonator ($\eta=10^{-5} \not= 0$) causes higher harmonics to be damped more in this area of $\gamma$ than for higher blowing pressures (as explained e.g.\ in Ref.\ \onlinecite{kergomard00}). The deviation from a square-wave signal is thus more noticeable close to the threshold, and as the HBM calculations imposed a nonzero dissipation, this is probably the reason for the small deviation in Figure~\ref{f:nearthres-nl}. The deviation is not visible in the time domain. A popular simplification of the nonlinear function~\eqref{e:simplenonlin} is a cubic expansion for small oscillations (e.g.\ Ref.\ \cite{mcintyre83,grand97,worman71}): \begin{equation} \tilde{u}(p) = u_{00}+Ap+Bp^2+Cp^3, \label{e:cubic} \end{equation} where $u_{00}$, $A$, $B$, and $C$ are easily found by expanding equation~\eqref{e:simplenonlin}. Its Helmholtz solution is easily calculated like above, yielding \begin{equation} p(\gamma)=\sqrt{-\frac AC}=\sqrt{\frac{8\gamma^2(3\gamma-1)}{\gamma+1}}. \label{e:pcubic} \end{equation} The influence of the difference between the two versions of the nonlinear function is investigated in Figure~\ref{f:compcub} for the lossless case. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/compcub-freq-g0.334336.eps}% \\\includegraphics[width=3.25in]{eps/compcub-freq-g0.40.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/compcub-freq-g0.334336.eps}% \\\includegraphics[width=5.5in]{eps/compcub-freq-g0.40.eps}% \fi \fi \caption [The (lossless) Helmholtz motion and the (almost lossless) HBM for 299 harmonics using the full nonlinearity~\eqref{e:simplenonlin} and the cubic expansion~\eqref{e:cubic} (a) close to the oscillation threshold ($\gamma=0.334$) and (b) far from it ($\gamma=0.40$) for $\zeta=0.5$, $\eta=10^{-5}$. Above the 11th harmonic only every 10th harmonic is shown.] {\label{f:compcub} \ifgalleyfig {The (lossless) Helmholtz motion and the (almost lossless) HBM for 299 harmonics using the full nonlinearity~\eqref{e:simplenonlin} and the cubic expansion~\eqref{e:cubic} (a) close to the oscillation threshold ($\gamma=0.334$) and (b) far from it ($\gamma=0.40$) for $\zeta=0.5$, $\eta=10^{-5}$. Above the 11th harmonic only every 10th harmonic is shown.} \fi} \end{figure} Close to the oscillation threshold, Figure~\ref{f:compcub}a, we see that there is no significant difference between the two versions of the nonlinear equation, as expected. The fact that the HBM is lower for higher harmonics is as before due to the small attenuation we had to include to perform the numerical calculations. Far from the threshold, however, Figure~\ref{f:compcub}b, we see that the cubic expansion fails to approximate the nonlinear equation. For lower harmonics this error is larger than the attenuation effect in the HBM calculations. This is further discussed by Fritz et al \cite{fritz04}. In Figure~\ref{f:P1-gamma} we have completed some of the curves that we failed to make in Figure~\ref{f:holes}, and even increased the number of harmonics, owing to the backtracking mechanism. Admittedly, at $N_p=49$, a few holes can still be seen, but the convergence is significantly improved. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/HBM-hh-gamma.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/HBM-hh-gamma.eps}% \fi \fi \caption [Amplitude of first harmonic as the blowing pressure increases for the Helmholtz solution \eqref{e:helmotion} and the HBM truncated to 1, 3, 9, and 49 harmonics, the last coinciding with Helmholtz ($\zeta=0.5$, $\eta=10^{-5}$)] {\label{f:P1-gamma} \ifgalleyfig {Amplitude of first harmonic as the blowing pressure increases for the Helmholtz solution \eqref{e:helmotion} and the HBM truncated to 1, 3, 9, and 49 harmonics, the last coinciding with Helmholtz ($\zeta=0.5$, $\eta=10^{-5}$)} \fi} \end{figure} Here the amplitude of the first harmonic is plotted for different numbers of harmonics as a function of the blowing pressure $\gamma$ together with first harmonic of the Helmholtz solution, deduced from equation~\eqref{e:helmotion}. In practice, the solution at $\gamma = 0.4$ was found and then \emph{hbmap} was used to make Harmbal calculate solution for each of a large number of subsequent values of $\gamma$ down to the oscillation threshold by using the previous solution as initial value. The procedure was repeated from $\gamma=0.4$ up to the point where the reed started to beat, i.e.\ for $p<\gamma-1$ in equation~\eqref{e:simplenonlin}. Without losses (Helmholtz solution) the beating threshold does not arrive before $\gamma=0.5$, and this should be expected for the nearly lossless case studied with the HBM also. However, the number of harmonics $N_p$ taken into account in the HBM calculations is too small to follow the sharp edges of the square signal. The resulting overshoots in $p(t)$, as seen in Figure~\ref{f:nearthres-nl}b, cause $p$ to prematurely exceed the criterion for beating. The beating threshold converges to 0.5 as $N_p$ increases (see also Ref.\ \onlinecite{fritz04}). Note that, for the chosen value of $\zeta$, it can be calculated following Hirschberg \cite[eq.(45)]{hirschberg95} that above $\gamma \simeq 0.45$, the Helmholtz solution loses its stability through a subharmonic bifurcation (a period-doubling occuring). By Figure~\ref{f:P1-gamma} we can also verify that the model experiences a direct Hopf bifurcation (which is known since the work of Grand et al.\ \cite{grand97}). Thus, a single harmonic is enough to study the solution around the threshold. Far from the threshold, more harmonics have to be taken into account for $P_1$ to converge toward the Helmholtz solution. This is not obvious and for example contradictory with the hypothesis made for the VTM \cite{kergomard00}. Thus Harmbal appears as an interesting tool to evaluate the relevance of approximate methods according to the parameter values. \subsubsection{Helmholtz oscillation for a stepped conical tube} The saxophone works similarly to the clarinet, but the bore has a conic form. In this section we compare the HBM calculations with analytical results, and in order to calculate the Helmholtz motion when losses are ignored, we need to simplify the cone by assuming that it consists of a sequence of $N$ sylinders of length $l$ and cross section $S_i=\frac12i(i+1)S_1$, $S_1=S$ being the cross section of the smallest cylinder, and $i=1,\dots,N$ (see Ref.\ \onlinecite{dalmont00}). The total length of the instrument is thus $L=Nl$. The input impedance of such a {\em stepped cone\/} may be written as \begin{equation} Z_r(\omega)=\frac{2i} {\cot\!\left(\!\frac{\omega'}{4} - i\alpha(\omega')\right) + \cot\!\left(\!\frac{\omega'}{4N}- i\alpha(\frac{\omega'}N)\right)}, \label{e:coneimp} \end{equation} where $\omega'\triangleq2\omega/(N+1)$ when $\omega=2\pi\hat f/f_r$, where $f_r$ is the first eigenfrequency of this resonator. We have ignored the dispersion term here. Equation~\eqref{e:coneimp} is used instead of equation~\eqref{e:freqresponse}, and the damping $\alpha(\omega)=\psi\eta\sqrt{\omega/2\pi}$ is zero in the analytic Helmholtz case and very small ($\eta=2\cdot10^{-5}$ below which convergence became difficult) for the calculations with the HBM. As before, the pressure amplitude of the ideal lossless case is calculated by solving $u(p)=u(-Np)$, and two solutions are possible:\footnote{This result corrects equation~(14) in ref.~\onlinecite{dalmont00}} \begin{equation} \begin{array}{rl \llap{$p^{\pm}$}(\gamma)\!\!\!\!&=\displaystyle\frac{(N{-}1)(2{-}3\gamma)}{2(N^2-N+1)}\\ &\displaystyle\pm\, \frac{ \sqrt{(N{-}1)^2+(N{+}1)^2(-3\gamma^2{+}4\gamma{-}1)}}{2(N^2-N+1)} \end{array} \label{e:coneN} \end{equation} as long as $\gamma<1/(N+1)$ for the standard Helmholtz motion ($p^+$) and $\gamma<N/(N+1)$ for the inverted one ($p^-$), which is unstable. Above these limits $p^+=\gamma$ and $p^-=-\gamma/N$. The magnitude of the first harmonic of a square or rectangular wave is then given by \begin{equation} P_1^{\pm}(\gamma)=\frac{\sin\frac\pi{N+1}}{\frac\pi{N+1}}p^{\pm}(\gamma). \label{e:P1p} \end{equation} For $N=1$, equation~\eqref{e:coneN} reduces to equation~\eqref{e:helmotion}. For higher $N$, the pressure oscillation becomes asymmetric. We take the case $N=2$ and get \begin{equation} p^{\pm}(\gamma)=\frac16\left(2-3\gamma\pm\sqrt{-27\gamma^2+36\gamma-8}\right). \label{e:coneN2} \end{equation} This result is compared with HBM calculations in Figure~\ref{f:coneN2A} for $\gamma=0.31$. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/freq-n2e-5g0.31A.eps}% \\\includegraphics[width=3.25in]{eps/time-n2e-5g0.31A.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/freq-n2e-5g0.31A.eps}% \\\includegraphics[width=5.5in]{eps/time-n2e-5g0.31A.eps}% \fi \fi \caption [Comparison between the standartd Helmholtz motion of a stepped cone ($N=2$) and the HBM for various $N_p$ at $\gamma=0.31$, $\zeta=0.2$, and $\eta=2\cdot10^{-5}$. (a) The magnitude of the harmonics and (b) one oscillation period. $N_t$ varies from 128 for $N_p{=}5$ to 1024 for $N_p{=}180$.] {\label{f:coneN2A} \ifgalleyfig {Comparison between the standartd Helmholtz motion of a stepped cone ($N=2$) and the HBM for various $N_p$ at $\gamma=0.31$, $\zeta=0.2$, and $\eta=2\cdot10^{-5}$. (a) The magnitude of the harmonics and (b) one oscillation period. $N_t$ varies from 128 for $N_p{=}5$ to 1024 for $N_p{=}180$.} \fi} \end{figure} Theoretically, the spectrum of the Helmholtz solution, Figure~\ref{f:coneN2A}a, shows that every third component is missing (actually zero) while the remaining components decrease in magnitude thus forming the asymmetric pressure oscillation as shown in Figure~\ref{f:coneN2A}a. The HBM, on the other hand, suggests that the first component in each pair be smaller than the second component. This results in a {\em dip\/} at the middle of the long, positive part of the period (i.e. on both extremities $t=0$ and $t=1024$ of the curve in Figure \ref{f:coneN2A}). The same was observed for $N=3$ and $N=4$, where the long part of the period was divided by similar dips into three and four parts, respectively (not shown). The number of time samples, $N_t$ did not change this fact, but as Figure~\ref{f:coneN2A} indicates, the dips gradually become narrower as the number of harmonics $N_p$ increases. This indicates that the HBM approaches the Helmholtz solution as $N_p$ approaches infinity. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/map-n2e-5P1.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/map-n2e-5P1.eps}% \fi \fi \caption [Amplitude of first harmonic $P_1$ as a function of the blowing pressure $\gamma$ for the Helmholtz solution \eqref{e:coneN2} for 2-stepped cone and the HBM truncated to 2, 5,\dots, and 63 harmonics, the last coinciding with Helmholtz ($\zeta=0.2$, $\eta=2\cdot10^{-5}$). Only nonbeating regimes are shown.] {\label{f:P1-gammaN2} \ifgalleyfig {Amplitude of first harmonic $P_1$ as a function of the blowing pressure $\gamma$ for the Helmholtz solution \eqref{e:coneN2} for 2-stepped cone and the HBM truncated to 2, 5,\dots, and 63 harmonics, the last coinciding with Helmholtz ($\zeta=0.2$, $\eta=2\cdot10^{-5}$). Only nonbeating regimes are shown.} \fi} \end{figure} A bifurcation diagram is plotted in Figure~\ref{f:P1-gammaN2}. Similarly to Figure~\ref{f:P1-gamma} for the cylindrical bore, the amplitude of the first harmonic is plotted for different number of harmonics as a function of the blowing pressure $\gamma$. The Helmholtz solution (equation~\eqref{e:P1p} with $N=2$) is also plotted. As shown by Ollivier et al.\ \cite{ollivier04b}, the lower part of the upper branch and the branch of the inverted Helmholtz motion are unstable. In practice, these curves are more difficult to obtain with {\em hbmap\/} than for the cylindrical bore, especially close to the subcritical oscillation threshold around $\gamma=0.28$, where computation was not possible at this low losses. More sophisticated continuation schemes should be considered to obtain complete curves. However, it is obvious from the diagram that the model experiences a sub-critical Hopf bifurcation, which agrees with the conclusion of Grand et al.\ \cite{grand97}. This means that a single-harmonic approximation is not enough to study the solution around this threshold, since the small-amplitude hypothesis does not hold. Further from the threshold, convergence toward the Helmholtz motion is ensured as the number of harmonics $N_p$ is increased. Only the nonbeating reed regime is considered in the figure and, similarly to Figure~\ref{f:P1-gamma}, it can be noted that the beating threshold for the model with $N_p$ harmonics depends on $N_p$ but converges toward the Helmholtz threshold $\gamma=1/3$ (corresponding to the lossless, continuous system) as $N_p$ is increased. \subsubsection{Validation with time-domain model} When adding a mass and damping to the reed or viscous losses and dispersion to the pipe, it is more difficult to compare Harmbal results with analytic solutions. This has been done by Fritz et al.\ \cite{fritz04} as far as the playing frequency is concerned, by comparison with approximate analytical formula. Here, we propose to confront both the playing frequency and the amplitude of the first partial with numerical results obtained with a time-domain method. We use a newly developed (real-time) time-domain method (here called TDM) by Guillemain et al.\ \cite{guillemain03a}. It is based on the same set of equations as presented in Section~\ref{s:clarinet} except that the impedance of the bore is slightly modified to be expressed as an infinite impulse response. In the Fourier domain, it can be expressed as \begin{equation} Z_r(\tilde\omega)=\frac{1-a_1e^{-i\tilde\omega}-b_0e^{-i\tilde\omega D}} {1 - a_1e^{-i\tilde\omega} + b_0 e^{-i\tilde\omega D}}. \label{e:philimp} \end{equation} where $\tilde\omega=\hat\omega/f_s$, $f_s$ being the sampling frequency, and the integer $D=\mathrm{round}(f_s/2f_r)$ the time delay in samples for the sound wave to propagate to the end of the bore and back. The constants $a_1$ and $b_0$ are to be adjusted so that the two first peaks of resonance have the same amplitude as the two first peaks of equation~\eqref{e:freqresponse}. To express equation~\eqref{e:philimp} using our terminology, we remember that $\omega=2\pi\hat f/f_r$ and obtain \begin{equation} Z_r(\omega)=\frac{1 - a_1e^{-i\omega\frac{f_r}{f_s}} - b_0 e^{-i\omega/2}} {1 - a_1e^{-i\omega\frac{f_r}{f_s}} + b_0 e^{-i\omega/2}}. \label{e:philimp2} \end{equation} In this section, we also include the mass and damping of the reed, so $M$ and $R$ are no longer zero. The TDM does not work for $M=R=0$, or even for values close to this, so we have used a reed with weak interaction with the pipe resonance as well as one with close to normal reed impedance. The corresponding values for $\omega_e$ and $q_e\triangleq g_e/\omega_e$ are shown in Table~\ref{t:phil}. \begin{table} \caption{The values of $M$ and $R$ for three strengths of reed interaction. The bore parameters are $D=247$ ($f_r=103.4$\,Hz), $a_1=0.899$, and $b_0=0.0946$ for sampling frequency $f_s=51100$\,Hz.} \smallskip \centerline{% \begin{tabular}{l|cccc}\hline Reed &$\omega_e$/Hz&$q_e$ &$M$&$R$\\\hline Weak &10000 & 0.1 &$1.070\e{-4}$ &$1.034\e{-3\vphantom{^1}}$\\ Normal &\ph02500 & 0.2 &$1.712\e{-3}$ &$\ph08.28\e{-3}$\\ \hline \end{tabular}} \label{t:phil} \end{table} Figure~\ref{f:phil-gamma}a shows the bifurcation diagram for two values of $\zeta$ and for weak and normal reed impedance, while Figure~\ref{f:phil-gamma}b shows the corresponding variation in the dimensionless playing frequency $f_p/f_r$. The lines represent the continuous solutions of the HBM, and the symbols show a set of results derived from the steady-state part of the TDM signal. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/phil-gamma-p15.eps}% \\\includegraphics[width=3.25in]{eps/freqcomp.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/phil-gamma-p15.eps}% \\\includegraphics[width=5.5in]{eps/freqcomp.eps}% \fi \fi \caption [Comparison between HBM and TDM of the amplitude of (a) the first harmonic $P_1$ and (b) the dimensionless playing frequency $f_p/f_r$ as the blowing pressure $\gamma$ increases for a clarinet-like system with viscous losses and weak and normal reed interaction. TDM values for $\zeta=0.50$ and $\gamma>0.48$ are omitted due to period doubling. So are the beating regimes of HBM calculations. ($f_s\,{=}\,51100$Hz, $N_t\,{=}\,512$, $f_r\,=\,103.4$\,Hz, $N_p\,{=}\,15$)] {\label{f:phil-gamma} \ifgalleyfig {Comparison between HBM and TDM of the amplitude of (a) the first harmonic $P_1$ and (b) the dimensionless playing frequency $f_p/f_r$ as the blowing pressure $\gamma$ increases for a clarinet-like system with viscous losses and weak and normal reed interaction. TDM values for $\zeta=0.50$ and $\gamma>0.48$ are omitted due to period doubling. So are the beating regimes of HBM calculations. ($f_s\,{=}\,51100$Hz, $N_t\,{=}\,512$, $f_r\,=\,103.4$\,Hz, $N_p\,{=}\,15$)} \fi} \end{figure} The TDM symbols fall well on the lines of the HBM, except for $\zeta=0.50$ when $\gamma$ approaches 0.5. Then the TDM experiences period doubling, i.e.\ two subsequent periods of the signal differ. At the same time, not being able to show subharmonics, the HBM shows signs of a beating reed, possibly a solution that is unstable and thus not attainable by time-domain methods. Note that three points have to be verified before comparing results from the HBM and the TDM: The numerical scheme used in the TDM to approximate the time derivatives in the reed equation~\eqref{e:lindiff} requires discretization. Depending on the sampling frequency $f_s$, the peak of resonance of the reed deviates more or less from the one given by the continuous equation. For normal reed interaction ($f_e$=2500 Hz), the deviation is negligible, but it may become significant in the case of weak reed interaction, where the peak is at 10000\,Hz. However, the fact that the reed and the bore interact weakly in the latter case, implies that the exact position of the peak has little importance. Therefore, at the used sampling frequency, the discretization in the TDM is not compensated for in the HBM calculations. Then there should be agreement between the sampling frequency $f_s$ in the TDM and the number of samples $N_t$ per period in the HBM. Their relation is given by \begin{equation} N_t=\frac{f_s}{f_p}. \end{equation} In order to have a sufficiently high sampling rate, we have chosen $N_t=512$. The playing frequency $f_p$ is plotted in Figure~\ref{f:phil-gamma}b, and we used an average $f_s=51100$\,Hz for both the HBM and the TDM. Finally, it seems also necessary that $N_p$ and $N_t$ are chosen so that \begin{equation} N_p+1 = \frac{N_t}{2}. \end{equation} In practice, however, when comparing bifurcation diagrams of the first harmonic $P_1$, as in Figure \ref{f:phil-gamma}, rather low values of $N_p$ give good results. Nevertheless, more harmonics are obviously needed to compare waveforms in the time domain, especially far from the oscillation threshold. \section{Practical experiences}\label{s:practexp}\label{s:disc} \subsection{Multiple solutions} As we consider a nonlinear problem, we cannot anticipate the number of solutions. Therefore, it should not be surprising that it is possible to obtain multiple solutions for a given set of parameter values. When searching for a particular solution, this may be a practical problem. Fritz et al.\ \cite{fritz04} have discovered that some solutions seem to disappear when increasing the number of harmonics $N_p$, implying that solutions may arise from the truncation to a finite $N_p$. We have now discovered alternative solutions that persist even at very high $N_p$. Let us illustrate this with the simple model of the clarinet used in Section~\ref{s:helcyl}, where the reed is a spring without mass or damping, the nonlinearity is given by equation~\eqref{e:simplenonlin}, and the bore is an ideal cylinder with nearly lossless propagation and no dispersion. Figure~\ref{f:bizsol} shows a three-level sister solution together with the related Helmholtz solution for a large number of harmonics, $N_p=2000$. \begin{figure \ifgalleyfig% \includegraphics[width=3.25in]{eps/comp_p_helmbiz2.eps}% \\\includegraphics[width=3.25in]{eps/comp_u_helmbiz2.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/comp_p_helmbiz2.eps \\\includegraphics[width=5.5in]{eps/comp_u_helmbiz2.eps \fi \fi \caption [The pressure (a) and volume-flow (b) wave form of the Helmholtz solution and a 3-level sister solution calculated by the HBM employing the simple clarinet model with $\zeta=0.5$, $\gamma=0.4$, $N_p=2000$, $\eta=10^{-5}$.] {\label{f:bizsol} \ifgalleyfig {The pressure (a) and volume-flow (b) wave form of the Helmholtz solution and a 3-level sister solution calculated by the HBM employing the simple clarinet model with $\zeta=0.5$, $\gamma=0.4$, $N_p=99$, $\eta=10^{-5}$.} \fi} \end{figure} A solution of the lossless problem should satisfy the criteria \cite{kergomard95} \begin{equation} \left\{ \begin{array}{l} p(t+\pi)=-p(t)\\ u(t+\pi)=u(t)\\ \end{array} \right. \label{e:pu_temp} \end{equation} (the dimensionless period being $2\pi$), as well as the conditions stated before equation~\eqref{e:helmotion}, noting that $p(t)\,{=}\,u(t)\,{=}\,0$ for all $t$ is the static solution. It is easily verified graphically that both of the solutions in Figure~\ref{f:bizsol} satisfy these conditions. Moreover, since they also satisfy equation~\eqref{e:simplenonlin}, the three-level solution is a solution of the lossless model. Whereas the system of time-domain equations~\eqref{e:pu_temp} has an infinity of solutions, truncation in frequency-domain limits the number of solutions. The unique solution of the HBM with only one harmonic is obviously a sine. Let us analyse the situation in the simplest nontrivial case of the lossless problem with two odd harmonics and a cubic expansion for nonlinear coupling. Ignoring even harmonics, the HBM gives a system of two equations (see Kergomard et al.\ \cite{kergomard00}): \begin{equation} \left\{ \begin{array}{rclr} \alpha & = & 3P_12(1+x+2|x|^2) & \mathrm{(a)}\\ \alpha x & = & P_12(1+3x|x|^2+6x), & \mathrm{(b)}\\ \end{array} \right. \label{e:3h_P} \end{equation} where $\alpha=-A/C$ and $x=P_3/P_1$. As equation~(\ref{e:3h_P}a) imposes $P_3$ to be real, solving this system amounts to solving \begin{equation} x^3+x^2-x=1/3. \label{e:3h_x} \end{equation} This equation has three real solutions $x \simeq -1.5151$, $-0.2776$ and $0.7926$. All of them are found by Harmbal for negligible losses ($\eta=10^{-5}$), and the corresponding waveforms are presented in Figure~\ref{f:helm_biz_3h}. We note that the second solution leads to the Helmholtz motion when increasing the number of harmonics (with the theoretical value known to be $x=-1/3$) whereas the third one corresponds to the three-level solution in Figure \ref{f:bizsol}. We can also easily imagine that these three solutions of the truncated problem are three-harmonic approximations of square waves that are distributed on three levels: $p^\pm \simeq \pm 0.5$ and $p=0$. Respectively, they have two, zero, and one steps at the zero-level. It should be noted that the conditions (\ref{e:pu_temp}) for the continuous problem do not constrain the duration of each step. Figure \ref{f:sol_np100} shows two such twin solutions for $N_p=99$ corresponding to the three-level solution in Figure \ref{f:helm_biz_3h}. This has to be kept in mind when increasing $N_p$ using the HBM. \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/helm_biz_3h_bis.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/helm_biz_3h_bis.eps}% \fi \fi \caption [The pressure waveform of the three solutions found by the HBM with $N_p=3$ employing the simple clarinet model with $\zeta=0.5$, $\gamma=0.4$, $\eta=10^{-5}$.] {\label{f:helm_biz_3h} \ifgalleyfig {The pressure waveform of the three solutions found by the HBM with $N_p=3$ employing the simple clarinet model with $\zeta=0.5$, $\gamma=0.4$, $\eta=10^{-5}$.} \fi} \end{figure} \begin{figure} \ifgalleyfig% \includegraphics[width=3.25in]{eps/deux_sol_np100.eps}% \else \ifoutputfig% \includegraphics[width=5.5in]{eps/deux_sol_np100.eps}% \fi \fi \caption [The pressure waveform of two solutions that differ by the duration of their steps, found by the HBM employing the simple clarinet model with $\zeta=0.5$, $\gamma=0.4$, $\eta=10^{-5}$, $N_p=99$.] {\label{f:sol_np100} \ifgalleyfig {The pressure waveform of two solutions that differ by the duration of their steps, found by the HBM employing the simple clarinet model with $\zeta=0.5$, $\gamma=0.4$, $\eta=10^{-5}$, $N_p=99$.} \fi} \end{figure} While the Helmholtz motion is known to be stable \cite{kergomard95}, the two three-level solutions can be considered as a combination of the static solution (the zero level) and the square wave (two levels with opposite values). Since we know from Kergomard \cite{kergomard95} that in the case of ideal propagation (neither losses nor dispersion), the stability domain of these two solutions are mutually exclusive, it can be concluded that the three-level solutions are unstable. Taking into account losses in the propagation does not make the three-level solutions vanish. But a simple reasoning to determine the stability of this solution is not possible in this case. To the authors knowledge, however, such a solution has never been observed experimentally at low level of excitation. \subsection{Initial value of the playing frequency} A practical difficulty encountered is the convergence of the playing frequency $f_p$. If its initial value is not close enough to the solution, divergence is almost inevitable. This occurs because the resonator impedance $Z_r$ tends to vanish outside the immediate surroundings of the resonance peaks of the resonator, rendering $\vec F(\vec P,f_p)$ very small and thereby $\vec G\simeq \vec P/P_1$ nearly constant with respect to $f_p$. The slope $\partial\vec G/\partial f_p$ thus becomes close to zero, the Newton step leads far away from the solution, and convergence fails. Dissipation widens the resonance peaks and thus also the convergence range. For a simple system where the playing frequency is known to correspond to a resonance peak of the tube, initializing $f_p$ is easy. However, with dispersion or other inharmonic effects, choosing an initial value for $f_p$ may be difficult. In Harmbal the problem may to some extent be avoided by the possibility of gradually adding the dispersion (or other inharmonic effects), so that the playing frequency can be followed quasi-continuously from a known solution without dispersion, for instance by using {\em hbmap}. \section{Conclusions}\label{s:concl} The harmonic balance method (HBM) is suited for studies of self-sustained oscillations of musical instruments, and the computer program {\em Harmbal\/} has been developed for this application. It is available with its source code \cite{harmbal}, has a free licence, and is already in use by several researchers. It is programmed in C, runs fast, and is easily used by other application, such as for continuation purposes. Some difficulties are related to the digital sampling of the signal and can be solved by introducing a backtracking mechanism. When using a large number of harmonics, the extreme case of the (lossless) Helmholtz motion can be solved for different shapes of resonators. Nevertheless, the value of the first harmonic $P_1$ seems to be well predicted by lower values of $N_p$, in particular close to the threshold of a direct bifurcation. For the saxophone we used a stepped-cone bore and observed one or more dips during the longest part of the period, depending on the number of steps. These dips approach pure impulses as $N_p$ increases. The number of samples $N_t$ in a period showed to be insignificant for these dips. The HBM can lead to some alternative solutions for a unique set of parameters. The nondissipative versions of these solutions satisfy the continuous model equations, but they are not stable and thus cannot be attained by ab initio time-domain calculations. Another problem is the great sensitivity to the guessed playing frequency. As a consequence, a certain expertise is needed in order to use the method, but, thanks to an automatic continuation procedure, the calculation is easy. We note that also experimental results can be used for the impedance of the resonator. \begin{acknowledgments} The Europeen Union through the MOSART project is acknowledged for financial support. We would also like to thank Claudia Fritz at IRCAM in Paris for thorough testing and valuable feedback, Jo\"el Gilbert at Laboratoire d'Acoustique de l'Universit\'e du Maine (LAUM) in Le Mans, and Philippe Guillemain at the Laboratoire de M\'ecanique et d'Acoustique at CNRS in Marseille for fruitful discussions during the work, and the latter also for kindly providing some Matlab code for the time-domain model. \end{acknowledgments} \bibliographystyle{jasasty}
{ "timestamp": "2005-03-07T11:23:46", "yymm": "0503", "arxiv_id": "physics/0503047", "language": "en", "url": "https://arxiv.org/abs/physics/0503047" }
\section{\@startsection{section}{1}% \z@{.7\linespacing\@plus\linespacing}{.5\linespacing}% {\normalfont\bfseries} \def\subsection{\@startsection{subsection}{2}% \z@{.5\linespacing\@plus.7\linespacing}{-.5em}% {\normalfont\scshape}} \catcode`\@=12 \catcode`\@=11 \def\sideset#1#2#3{% \@mathmeasure\z@\displaystyle{#3}% \global\setbox\@ne\vbox to\ht\z@{}\dp\@ne\dp\z@ \setbox\tw@\box\@ne \@mathmeasure4\displaystyle{\copy\tw@#1}% \@mathmeasure6\displaystyle{#3{#2}}% \dimen@-\wd6 \advance\dimen@\wd4 \advance\dimen@\wd\z@ \hbox to\dimen@{}\mathop{\kern-\dimen@\box4\box6}% } \catcode`\@=12 {\obeylines \gdef\MATH{\begingroup\parindent0pt\parskip0pt plus 0pt\obeylines% \def^^M{\vskip4pt}% \obeyspaces\tt\small}% } \def\goodbreakpoint{\par\penalty-5000% \vrule height10pt depth2pt width0pt\leavevmode} \def\endgroup{\endgroup} \def\MATHphi{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\phi$\hss}\hss}} \def\MATHGamma{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\Gamma$\hss}\hss}} \def\MATHpi{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\pi$\hss}\hss}} \def\MATHinfty{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\infty$\hss}\hss}} \def\MATHhStrich{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\vrule height4.5pt depth-3.5pt width5.24995pt}\hss}} \def\MATHluEck{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hskip2.12497pt \vrule height4.5pt depth1pt width1pt \vrule height4.5pt depth-3.5pt width2.12498pt}\hss}} \def\MATHruEck{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{% \vrule height4.5pt depth-3.5pt width2.12497pt \vrule height4.5pt depth1pt width1pt \hskip2.12498pt}\hss}} \def\MATHloEck{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hskip2.12497pt \vrule height9pt depth-3.5pt width1pt \vrule height4.5pt depth-3.5pt width2.12498pt}\hss}} \def\MATHroEck{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{% \vrule height4.5pt depth-3.5pt width2.12497pt \vrule height9pt depth-3.5pt width1pt \hskip2.12498pt}\hss}} \def\MATHvStrich{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hskip2.12497pt \vtop to 0pt{\hsize1pt\vss% \vrule height17pt depth6pt width1pt\vskip8pt\vss\par}% \hskip2.12498pt}\hss}} \def\MATHtStueck{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{% \vrule height4.5pt depth-3.5pt width2.12497pt \vrule height4.5pt depth2pt width1pt \vrule height4.5pt depth-3.5pt width2.12498pt}\hss}} \def\MATHbackslash{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\backslash$\hss}\hss}} \def\MATHlbrace{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\{$\hss}\hss}} \def\MATHrbrace{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\}$\hss}\hss}} \def<{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\langle$\hss}\hss}} \def<{<} \def>{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$\rangle$\hss}\hss}} \def>{>} \def\MATHhoch{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss$^\land$\hss}\hss}} \def\MATHtief{\leavevmode \hbox to 0pt{\hbox to 5.24995pt{\hss\vrule height0pt depth.8pt width3pt\hss}\hss}} \def\ringerl(#1 #2){\move(#1 #2)\fcir f:0 r:.1} \def\hdSchritt{\bsegment \lpatt(.05 .13) \rlvec(0.866025403784439 -.5) \savepos(0.866025403784439 -.5)(*ex *ey) \esegment \move(*ex *ey) } \def\vdSchritt{\bsegment \lpatt(.05 .13) \rlvec(0 -1) \savepos(0 -1)(*ex *ey) \esegment \move(*ex *ey) } \def\ldreieck{\bsegment \rlvec(0.866025403784439 .5) \rlvec(0 -1) \rlvec(-0.866025403784439 .5) \savepos(0.866025403784439 -.5)(*ex *ey) \esegment \move(*ex *ey) } \def\rdreieck{\bsegment \rlvec(0.866025403784439 -.5) \rlvec(-0.866025403784439 -.5) \rlvec(0 1) \savepos(0 -1)(*ex *ey) \esegment \move(*ex *ey) } \def\rhombus{\bsegment \rlvec(0.866025403784439 .5) \rlvec(0.866025403784439 -.5) \rlvec(-0.866025403784439 -.5) \rlvec(0 1) \rmove(0 -1) \rlvec(-0.866025403784439 .5) \savepos(0.866025403784439 -.5)(*ex *ey) \esegment \move(*ex *ey) } \def\RhombusA{\bsegment \rlvec(0.866025403784439 .5) \rlvec(0.866025403784439 -.5) \rlvec(-0.866025403784439 -.5) \rlvec(-0.866025403784439 .5) \savepos(0.866025403784439 -.5)(*ex *ey) \esegment \move(*ex *ey) } \def\RhombusB{\bsegment \rlvec(0.866025403784439 .5) \rlvec(0 -1) \rlvec(-0.866025403784439 -.5) \rlvec(0 1) \savepos(0 -1)(*ex *ey) \esegment \move(*ex *ey) } \def\RhombusC{\bsegment \rlvec(0.866025403784439 -.5) \rlvec(0 -1) \rlvec(-0.866025403784439 .5) \rlvec(0 1) \savepos(0.866025403784439 -.5)(*ex *ey) \esegment \move(*ex *ey) } \def\alpha{\alpha} \def\beta{\beta} \def\gamma{\gamma} \def\delta{\delta} \def\varepsilon{\varepsilon} \def\zeta{\zeta} \def\eta{\eta} \def\theta{\theta} \def\vartheta{\vartheta} \def\iota{\iota} \def\kappa{\kappa} \def\lambda{\lambda} \def\rho{\rho} \def\sigma{\sigma} \def\tau{\tau} \def\varphi{\varphi} \def\chi{\chi} \def\psi{\psi} \def\omega{\omega} \def\Gamma{\Gamma} \def\Delta{\Delta} \def\Theta{\Theta} \def\Lambda{\Lambda} \def\Sigma{\Sigma} \def\Phi{\Phi} \def\Psi{\Psi} \def\Omega{\Omega} \def\row#1#2#3{#1_{#2},\ldots,#1_{#3}} \def\rowup#1#2#3{#1^{#2},\ldots,#1^{#3}} \def\times{\times} \def\crf{} \def\rf{} \def\rfnew{} \def{\mathbb P}{{\mathbb P}} \def{\mathbb R}{{\mathbb R}} \def{\mathcal X}{{\mathcal X}} \def{\mathbb C}{{\mathbb C}} \def{\mathcal Mf}{{\mathcal Mf}} \def{\mathcal F\mathcal M}{{\mathcal F\mathcal M}} \def{\mathcal F}{{\mathcal F}} \def{\mathcal G}{{\mathcal G}} \def{\mathcal V}{{\mathcal V}} \def{\mathcal T}{{\mathcal T}} \def{\mathcal A}{{\mathcal A}} \def{\mathbb N}{{\mathbb N}} \def{\mathbb Z}{{\mathbb Z}} \def{\mathbb Q}{{\mathbb Q}} \def\left.\tfrac \partial{\partial t}\right\vert_0{\left.\tfrac \partial{\partial t}\right\vert_0} \def\dd#1{\tfrac \partial{\partial #1}} \def\today{\ifcase\month\or January\or February\or March\or April\or May\or June\or July\or August\or September\or October\or November\or December\fi \space\number\day, \number\year} \def\nmb#1#2{#2} \def\iprod#1#2{\langle#1,#2\rangle} \def\pder#1#2{\frac{\partial #1}{\partial #2}} \def\int\!\!\int{\int\!\!\int} \def\({\left(} \def\){\right)} \def\[{\left[} \def\]{\right]} \def\operatorname{supp}{\operatorname{supp}} \def\operatorname{Comp}{\operatorname{Comp}} \def\operatorname{Part}{\operatorname{Part}} \def\operatorname{Df}{\operatorname{Df}} \def\operatorname{dom}{\operatorname{dom}} \def\operatorname{Ker}{\operatorname{Ker}} \def\operatorname{Per}{\operatorname{Per}} \def\operatorname{Tr}{\operatorname{Tr}} \def\operatorname{Res}{\operatorname{Res}} \def\operatorname{Aut}{\operatorname{Aut}} \def\operatorname{kgV}{\operatorname{kgV}} \def\operatorname{ggT}{\operatorname{ggT}} \def\operatorname{diam}{\operatorname{diam}} \def\operatorname{Im}{\operatorname{Im}} \def\operatorname{Re}{\operatorname{Re}} \def\operatorname{ord}{\operatorname{ord}} \def\operatorname{rang}{\operatorname{rang}} \def\operatorname{rng}{\operatorname{rng}} \def\operatorname{grd}{\operatorname{grd}} \def\operatorname{inv}{\operatorname{inv}} \def\operatorname{maj}{\operatorname{maj}} \def\operatorname{fmaj}{\operatorname{fmaj}} \def\operatorname{nmaj}{\operatorname{nmaj}} \def\operatorname{neg}{\operatorname{neg}} \def\operatorname{sneg}{\operatorname{sneg}} \def\operatorname{des}{\operatorname{des}} \def\operatorname{\overline{maj}}{\operatorname{\overline{maj}}} \def\operatorname{\overline{des}}{\operatorname{\overline{des}}} \def\operatorname{\overline{maj}'}{\operatorname{\overline{maj}'}} \def\operatorname{maj'}{\operatorname{maj'}} \def\operatorname{zbk}{\operatorname{zbk}} \def\operatorname{nzbk}{\operatorname{nzbk}} \def\operatorname{NC}{\operatorname{NC}} \def\operatorname{NCmatch}{\operatorname{NCmatch}} \def\operatorname{ln}{\operatorname{ln}} \def\operatorname{der}{\operatorname{der}} \def\operatorname{Hom}{\operatorname{Hom}} \def\operatorname{tr}{\operatorname{tr}} \def\operatorname{Span}{\operatorname{Span}} \def\operatorname{grad}{\operatorname{grad}} \def\operatorname{div}{\operatorname{div}} \def\operatorname{rot}{\operatorname{rot}} \def\operatorname{Sp}{\operatorname{Sp}} \def\operatorname{sgn}{\operatorname{sgn}} \def\lim\limits{\lim\limits} \def\sup\limits{\sup\limits} \def\bigcup\limits{\bigcup\limits} \def\bigcap\limits{\bigcap\limits} \def\limsup\limits{\limsup\limits} \def\liminf\limits{\liminf\limits} \def\int\limits{\int\limits} \def\sum\limits{\sum\limits} \def\max\limits{\max\limits} \def\min\limits{\min\limits} \def\prod\limits{\prod\limits} \def\operatorname{tan}{\operatorname{tan}} \def\operatorname{cot}{\operatorname{cot}} \def\operatorname{arctan}{\operatorname{arctan}} \def\operatorname{arccot}{\operatorname{arccot}} \def\operatorname{arccot}{\operatorname{arccot}} \def\operatorname{tanh}{\operatorname{tanh}} \def\operatorname{coth}{\operatorname{coth}} \def\operatorname{arcsinh}{\operatorname{arcsinh}} \def\operatorname{arccosh}{\operatorname{arccosh}} \def\operatorname{arctanh}{\operatorname{arctanh}} \def\operatorname{arccoth}{\operatorname{arccoth}} \def\ss{\ss} \let\vv\v \def\operatorname{Tr}{\operatorname{Tr}} \def\po#1#2{(#1)_{#2}} \def\fl#1{\left\lfloor#1\right\rfloor} \def\cl#1{\left\lceil#1\right\rceil} \def\coef#1{\left\langle#1\right\rangle} \def{\bar X}{{\bar X}} \def{\sqrt{-1}}{{\sqrt{-1}}} \def{(\sqrt{-1})}{{(\sqrt{-1})}} \def\operatorname{bk}{\operatorname{bk}} \def\operatorname{CT}{\operatorname{CT}} \def\operatorname{NC}{\operatorname{NC}} \def\operatorname{rank}{\operatorname{rank}} \def\operatorname{stat}{\operatorname{stat}} \def\operatorname{NCmatch}{\operatorname{NCmatch}} \def\operatorname{Pf}{\operatorname{Pf}} \DeclareMathOperator{\h}{H} \def\raise-15pt\hbox{{\Huge$\square$}}{\raise-15pt\hbox{{\Huge$\square$}}} \def\qbinom#1#2{\left[\begin{smallmatrix} #1\\#2\end{smallmatrix}\right]_q} \begin{document} \newbox\Adr \setbox\Adr\vbox{ \centerline{\sc C.~Krattenthaler$^\dagger$} \vskip18pt \centerline{Institut Camille Jordan, Universit\'e Claude Bernard Lyon-I,} \centerline{21, avenue Claude Bernard, F-69622 Villeurbanne Cedex, France.} \centerline{E-mail: {\tt\footnotesize kratt@euler.univ-lyon1.fr}} \centerline{WWW: \footnotesize\tt http://igd.univ-lyon1.fr/\~{}kratt} } \title{Advanced Determinant Calculus: A Complement} \author[C.~Krattenthaler]{\box\Adr} \address{Institut Girard Desargues, Universit\'e Claude Bernard Lyon-I, 21, avenue Claude Bernard, F-69622 Villeurbanne Cedex, France.} \email{kratt@euler.univ-lyon1.fr} \thanks{$^\dagger$ Research partially supported by EC's IHRP Programme, grant HPRN-CT-2001-00272, ``Algebraic Combinatorics in Europe", and by the ``Algebraic Combinatorics" Programme during Spring 2005 of the Institut Mittag--Leffler of the Royal Swedish Academy of Sciences} \subjclass[2000]{Primary 05A19; Secondary 05A10 05A15 05A17 05A18 05A30 05E10 05E15 11B68 11B73 11C20 11Y60 15A15 33C45 33D45 33E05} \keywords{Determinants, Vandermonde determinant, Cauchy's double alternant, skew circulant matrix, confluent alternant, confluent Cauchy determinant, Pfaffian, Hankel determinants, orthogonal polynomials, Chebyshev polynomials, Meixner polynomials, Laguerre polynomials, continued fractions, binomial coefficient, Catalan numbers, Fibonacci numbers, Bernoulli numbers, Stirling numbers, non-intersecting lattice paths, plane partitions, tableaux, rhombus tilings, lozenge tilings, alternating sign matrices, non-crossing partitions, perfect matchings, permutations, signed permutations, inversion number, major index, compositions, integer partitions, descent algebra, non-commutative symmetric functions, elliptic functions, the number $\pi$, LLL-algorithm} \begin{abstract} This is a complement to my previous article {``Advanced Determinant Calculus"} ({\it S\'eminaire Lotharingien Combin.}\ {\bf 42} (1999), Article~B42q, 67~pp.). In the present article, I share with the reader my experience of applying the methods described in the previous article in order to solve a particular problem from number theory (G.~Almkvist, J.~Petersson and the author, {\it Experiment.\ Math.}\ {\bf 12} (2003), 441--456). Moreover, I add a list of determinant evaluations which I consider as interesting, which have been found since the appearance of the previous article, or which I failed to mention there, including several conjectures and open problems. \end{abstract} \maketitle \section{Introduction} In the previous article \machSeite{KratBN}\cite{KratBN}, I described several methods to evaluate determinants, and I provided a long list of known determinant evaluations. The present article is meant as a complement to \machSeite{KratBN}\cite{KratBN}. Its purpose is three-fold: first, I want to shed light on the problem of evaluating determinants from a slightly different angle, by sharing with the reader my experience of applying the methods from \machSeite{KratBN}\cite{KratBN} in order to solve a particular problem from number theory (see Sections~\ref{sec:det} and \ref{sec:eval}); second, I shall address the question why it is apparently in the first case combinatorialists (such as myself) who are so interested in determinant evaluations and get so easily excited about them (see Section~\ref{sec:comb}); and, finally third, I add a list of determinant evaluations, which I consider as interesting, which have been found since the appearance of \machSeite{KratBN}\cite{KratBN}, or which I failed to mention in the list given in Section~3 of \machSeite{KratBN}\cite{KratBN} (see Section~\ref{sec:detlist}), including several conjectures and open problems. \section{Enumerative combinatorics, nice formulae, and determinants} \label{sec:comb} Why are combinatorialists so fascinated by determinant evaluations? A simplistic answer to this question goes as follows. Clearly, binomial coefficients $\binom nk$ or Stirling numbers (of the second kind) $S(n,k)$ are basic objects in (enumerative) combinatorics; after all they count the subsets of cardinality $k$ of a set with $n$ elements, respectively the ways of partitioning such a set of $n$ elements into $k$ pairwise disjoint non-empty subsets. Thus, if one sees an identity such as\footnote{For more information on this determinant see Theorems~\ref{thm:MM} and \ref{thm:MM2} in this section and \machSeite{KratBN}\cite[Sections~2.2, 2.3 and 2.5]{KratBN}.} \begin{equation} \label{eq:M1} \det_{1\le i,j\le n}\(\binom {a+b}{a-i+j} \)= \prod _{i=1} ^{n}\frac {(a+b+i-1)!\,(i-1)!} {(a+i-1)!\,(b+i-1)!}, \end{equation} or\footnote{This determinant evaluation follows easily from the matrix factorisation $$\(S(i+j,i) \)_{1\le i,j\le n}=\((-1)^kk^i/(k!\,(i-k)!)\)_{1\le i,k\le n}\cdot\(k^j\)_{1\le k,j\le n},$$ application of \machSeite{KratBN}\cite[Theorem~26, (3.14)]{KratBN} to the first determinant, and application of the Vandermonde determinant evaluation to the second.} \begin{equation} \label{eq:S2} \det_{1\le i,j\le n}\(S(i+j,i) \)=\prod _{i=1} ^{n}i^i \end{equation} (and there are many more of that kind; see \machSeite{KratBN}\cite{KratBN} and Section~\ref{sec:detlist}), there is an obvious excitement that one cannot escape. Although this is indeed an explanation which applies in many cases, there is also an answer on a more substantial level, which brings us to the reason why {\it I\/} like (and need) determinant evaluations. The {\it favourite question} for an enumerative combinatorialist (such as myself) is $$\text {\it How many $\langle\dots\rangle$ are there?}$$ Here, $\langle\dots\rangle$ can be permutations with certain properties, certain partitions, certain paths, certain trees, etc. The {\it favourite theorem} then is: \begin{Theorem} \label{thm:fav} The number of $\langle\dots\rangle$ of size $n$ is equal to $$NICE(n).$$ \end{Theorem} I have already explained the meaning of $\langle\dots\rangle$. What does $NICE(n)$ stand for? Typical examples for $NICE(n)$ are formulae such as \begin{equation} \label{eq:Cat} \frac {1} {n+1}\binom {2n}n \end{equation} (\!{\it Catalan numbers}; cf.\ \machSeite{StanBI}\cite[Ex.~6.19]{StanBI}) or \begin{equation} \label{eq:alt} \prod _{i=0} ^{n-1}\frac {(3i+1)!}{(n+i)!} \end{equation} (the number of $n\times n$ alternating sign matrices and several other combinatorial objects; cf.\ \machSeite{BresAO}\cite{BresAO}). Let us be more precise. \begin{DefinitionA} \label{def:1} The symbol $NICE(n)$ is a formula of the type \begin{equation} \label{eq:NICE} \xi^n\cdot\text{\em Rat}(n)\cdot \prod _{i=1} ^{k}\frac {(a_in+b_i)!} {(c_in+d_i)!}, \end{equation} where $\text{\em Rat}(n)$ is a rational function in $n$, and where $a_i,c_i\in {\mathbb Z}$ for $i=1,2,\dots,k$, ${\mathbb Z}$ denoting the set of integers. The parameters $b_i,c_i,\xi$ can be arbitrary real or complex numbers. {\em(}If necessary, $(a_in+b_i)!$ has to be interpreted as $\Gamma(a_in+b_i+1)$, where $\Gamma(x)$ is the Euler gamma function, and similarly for $(c_in+d_i)!$.{\em)} \end{DefinitionA} Clearly, the formulae \eqref{eq:Cat} and \eqref{eq:alt} fit this ``Definition".\footnote{The writing $NICE(n)$ is borrowed from Doron Zeilberger \machSeite{ZeilAP}\cite[Recitation~III]{ZeilAP}. The technical term for a formula of the type \eqref{eq:NICE} is {\it``hypergeometric term"}, see \machSeite{PeWZAA}\cite[Sec.~3.2]{PeWZAA}, whereas, most often, the colloquial terms {\it``closed form"} or {\it``nice formula"} are used for it, see \machSeite{ZeilAP}\cite[Recitation~II]{ZeilAP}. More recently, some authors call sequences given by formulae of that type sequences of ``round" numbers, see \machSeite{KupeAH}\cite[Sec.~6]{KupeAH}.} If one is working on a particular problem, how can one recognise that one is looking at a sequence of numbers given by $NICE(n)$? The key observation is that, if we factorise $(an+b)!$ into its prime factors, where $a$ and $b$ are integers, then, as $n$ runs through the positive integers, the numbers $(an+b)!$ explode quickly, whereas the prime factors occurring in the factorisation will grow only moderately, more precisely, they will grow roughly linearly. Thus, if we encounter a sequence the prime factorisation of which has this property, we can be sure that there is a formula $NICE(n)$ for this sequence. Even better, as I explain in Appendix~A of \machSeite{KratBN}\cite{KratBN}, the program {\tt Rate}\footnote{\label{foot:Rate}{\tt Rate} is available from {\tt http://igd.univ-lyon1.fr/\~{}kratt}. It is based on a rather simple algorithm which involves rational interpolation. In contrast to what I read, with great surprise, in \machSeite{CoBWAA}\cite{CoBWAA}, the explanations of how {\tt Rate} works in Appendix~A of \machSeite{KratBN}\cite{KratBN} can be read and understood without any knowledge about determinants and, in particular, without any knowledge of the fifty or so pages that precede Appendix~A in \machSeite{KratBN}\cite{KratBN}.} will (normally\footnote{\label{foot:normally}{\tt Rate} will {\it always} be able to guess a formula of the type \eqref{eq:NICE} if there are enough initial terms of the sequence available. However, there is a larger class of sequences which have the property that the size of the primes in the prime factorisation of the terms of the sequence grows only slowly with $n$. These are sequences given by formulae containing ``Abelian" factors, such as $n^n$. Unfortunately, {\tt Rate} does not know how to handle such factors. Recently, Rubey \machSeite{RubeAD}\cite{RubeAD} proposed an algorithm for covering Abelian factors as well. His implementation {\tt Guess} is written in {\sl Axiom} and is available at {\tt http://www.mat.univie.ac.at/\~{}rubey/martin.html}.}) be able to guess the formula. To illustrate this, let us look at a particular example. Let us suppose that the first few values of our sequence are the following: \begin{multline*} 1, 2, 5, 14, 42, 132, 429, 1430, 4862, 16796, 58786, 208012, 742900, 2674440, \\ 9694845, 35357670, 129644790, 477638700, 1767263190, 6564120420. \end{multline*} The prime factorisation of the second-to-last number is (we are using {\sl Mathematica} here) \MATH \goodbreakpoint% In[1]:= FactorInteger[477638700] \goodbreakpoint% Out[1]= % \MATHlbrace % \MATHlbrace 2, 2% \MATHrbrace , % \MATHlbrace 3, 1% \MATHrbrace , % \MATHlbrace 5, 2% \MATHrbrace , % \MATHlbrace 7, 1% \MATHrbrace , % \MATHlbrace 11, 1% \MATHrbrace , % \MATHlbrace 23, 1% \MATHrbrace , % \MATHlbrace 29, 1% \MATHrbrace , % \MATHlbrace 31, 1% \MATHrbrace % \MATHrbrace \goodbreakpoint% \endgroup whereas the prime factorisations of the next-to-last and the last number in this sequence are \MATH \goodbreakpoint% In[2]:= FactorInteger[1767263190] \goodbreakpoint% Out[2]= % \MATHlbrace % \MATHlbrace 2, 1% \MATHrbrace , % \MATHlbrace 3, 1% \MATHrbrace , % \MATHlbrace 5, 1% \MATHrbrace , % \MATHlbrace 7, 1% \MATHrbrace , % \MATHlbrace 11, 1% \MATHrbrace , % \MATHlbrace 23, 1% \MATHrbrace , % \MATHlbrace 29, 1% \MATHrbrace , % \MATHlbrace 31, 1% \MATHrbrace , > % \MATHlbrace 37, 1% \MATHrbrace % \MATHrbrace \goodbreakpoint% In[3]:= FactorInteger[6564120420] \goodbreakpoint% Out[3]= % \MATHlbrace % \MATHlbrace 2, 2% \MATHrbrace , % \MATHlbrace 3, 1% \MATHrbrace , % \MATHlbrace 5, 1% \MATHrbrace , % \MATHlbrace 11, 1% \MATHrbrace , % \MATHlbrace 13, 1% \MATHrbrace , % \MATHlbrace 23, 1% \MATHrbrace , % \MATHlbrace 29, 1% \MATHrbrace , % \MATHlbrace 31, 1% \MATHrbrace , > % \MATHlbrace 37, 1% \MATHrbrace % \MATHrbrace \goodbreakpoint% \endgroup \noindent (To decipher this for the reader unfamiliar with {\sl Mathematica}: the prime factorisation of the last number is $ 2^ 2 3^ 1 5^ 1 11^ 1 13^ 1 23^ 1 29^ 1 31^ 1 37^ 1$.) One observes, first of all, that the occurring prime factors are rather small in comparison to the numbers of which they are factors, and, second, that the size of the prime factors grows only very slowly (from 31 to 37). Thus, we {\it can be sure} that there is a ``nice" formula $NICE(n)$ for this sequence. Indeed, {\tt Rate} needs only the first five members of the sequence to come up with a guess for $NICE(n)$: \MATH \goodbreakpoint% In[4]:= <{<}rate.m \goodbreakpoint% In[5]:= Rate[1,2,5,14,42] \goodbreakpoint% \leavevmode% i0 1 \leavevmode% 4 Gamma[- + i0] \leavevmode% 2 Out[5]= ---------------------- \leavevmode% Sqrt[Pi] Gamma[2 + i0] \goodbreakpoint% \endgroup As the reader will have guessed, {\tt Rate} uses the parameter $i_0$ instead of $n$. In fact, the formula is a fancy way to write $\frac {1} {i_0+1}\binom {2i_0}{i_0}$, that is, we were looking at the sequence of Catalan numbers \eqref{eq:Cat}. To see the sharp contrast, here are the first few terms of another sequence: $$1,2,9,272,589185.$$ (Also these are combinatorial numbers. They count the perfect matchings of the $n$-dimensional hypercube; cf.\ \machSeite{PropAH}\cite[Problem~19]{PropAH}.) Let us factorise the last two numbers: \MATH \goodbreakpoint% In[6]:= FactorInteger[272] \goodbreakpoint% Out[6]= % \MATHlbrace % \MATHlbrace 2, 4% \MATHrbrace , % \MATHlbrace 17, 1% \MATHrbrace % \MATHrbrace \goodbreakpoint% In[7]:= FactorInteger[589185] \goodbreakpoint% Out[7]= % \MATHlbrace % \MATHlbrace 3, 2% \MATHrbrace , % \MATHlbrace 5, 1% \MATHrbrace , % \MATHlbrace 13093, 1% \MATHrbrace % \MATHrbrace \goodbreakpoint% \endgroup The presence of the big prime factor 13093 in the last factorisation is a sure sign that we cannot expect a formula $NICE(n)$ as described in the ``Definition" for this sequence of numbers. (There may well be a simple formula of a different kind. It is not very likely, though. In any case, such a formula has not been found up to this date.) \medskip Now, that I have sufficiently explained all the ingredients in the ``prototype theorem'' Theorem~\ref{thm:fav}, I can explain why theorems of this form are so attractive (at least to me): the objects (i.e., the permutations, partitions, paths, trees, etc.) that it deals with are usually very simple to explain, the statement is very simple and can be understood by anybody, the result $NICE(n)$ has a very elegant form, and yet, very often it is not easy at all to give a proof (not to mention a true {\it explanation} why such an elegant result occurs.) Here are two examples. They concern {\it rhombus tilings}, by which I mean tilings of a region by rhombi with side lengths 1 and angles of $60^\circ$ and $120^\circ$. The first one is a one century old theorem due to MacMahon {\machSeite{MacMAA}\cite[Sec.~429, $q\rightarrow 1$; proof in Sec.~494]{MacMAA}}.\footnote{To be correct, MacMahon did not know anything about rhombus tilings, they did not exist in enumerative combinatorics at the time. The objects that he considered were {\it plane partitions}. However, there is a very simple bijection between plane partitions contained in an $a\times b\times c$ box and rhombus tilings of a hexagon with side lengths $a,b,c,a,b,c$, as explained for example in \machSeite{DT}\cite{DT}.} \begin{Theorem} \label{thm:MM} The number of rhombus tilings of a hexagon with side lengths $a,b,c,a,b,c$ whose angles are $120^\circ$ {\em(}see Figure~\ref{fig:1}.a for an example of such a hexagon, and Figure~\ref{fig:1}.b for an example of a rhombus tiling{\em)} is equal to \begin{equation} \label{eq:M2} \prod _{i=1} ^{c}\frac {(a+b+i-1)!\,(i-1)!} {(a+i-1)!\,(b+i-1)!}. \end{equation} \quad \quad \qed \end{Theorem} \begin{figure}[h] \centertexdraw{ \drawdim truecm \linewd.02 \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-0.866025403784439 -.5) \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -1) \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -1) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -2) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -3) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \move (-1.732050807568877 -4) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \move (-1.732050807568877 -5) \rdreieck \rhombus \rhombus \rhombus \rhombus \move(8 0) \bsegment \drawdim truecm \linewd.02 \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-0.866025403784439 -.5) \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -1) \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -1) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -2) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move (-1.732050807568877 -3) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \move (-1.732050807568877 -4) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \move (-1.732050807568877 -5) \rdreieck \rhombus \rhombus \rhombus \rhombus \linewd.12 \move(0 0) \RhombusA \RhombusB \RhombusB \RhombusA \RhombusA \RhombusB \RhombusA \RhombusB \RhombusB \move (-0.866025403784439 -.5) \RhombusA \RhombusB \RhombusB \RhombusB \RhombusB \RhombusA \RhombusA \RhombusB \RhombusA \move (-1.732050807568877 -1) \RhombusB \RhombusB \RhombusA \RhombusB \RhombusB \RhombusA \RhombusB \RhombusA \RhombusA \move (1.732050807568877 0) \RhombusC \RhombusC \RhombusC \move (1.732050807568877 -1) \RhombusC \RhombusC \RhombusC \move (3.464101615137755 -3) \RhombusC \move (-0.866025403784439 -.5) \RhombusC \move (-0.866025403784439 -1.5) \RhombusC \move (0.866025403784439 -2.5) \RhombusC \RhombusC \move (0.866025403784439 -3.5) \RhombusC \RhombusC \RhombusC \move (2.598076211353316 -5.5) \RhombusC \move (0.866025403784439 -5.5) \RhombusC \move (-1.732050807568877 -3) \RhombusC \move (-1.732050807568877 -4) \RhombusC \move (-1.732050807568877 -5) \RhombusC \RhombusC \esegment \htext (-1.5 -9){\small a. A hexagon with sides $a,b,c,a,b,c$,} \htext (-1.5 -9.5){\small \hphantom{a. }where $a=3$, $b=4$, $c=5$} \htext (6.8 -9){\small b. A rhombus tiling of a hexagon} \htext (6.8 -9.5){\small \hphantom{b. }with sides $a,b,c,a,b,c$} \rtext td:0 (4.3 -4.1){$\sideset {} c {\left.\vbox{\vskip2.6cm}\right\}}$} \rtext td:60 (2.6 -.55){$\sideset {} {} {\left.\vbox{\vskip2cm}\right\}}$} \rtext td:120 (-.34 -.2){$\sideset {} {} {\left.\vbox{\vskip1.7cm}\right\}}$} \rtext td:0 (-2.4 -3.6){$\sideset {c} {} {\left\{\vbox{\vskip2.6cm}\right.}$} \rtext td:240 (-0.1 -6.9){$\sideset {} {} {\left.\vbox{\vskip2cm}\right\}}$} \rtext td:300 (2.9 -7.3){$\sideset {} {} {\left.\vbox{\vskip1.7cm}\right\}}$} \htext (-.9 0.2){$a$} \htext (2.8 -.1){$b$} \htext (3.2 -7.9){$a$} \htext (-0.4 -7.65){$b$} } \caption{} \label{fig:1} \end{figure} The second one is more recent, and is due to Ciucu, Eisenk\"olbl, Zare and the author {\machSeite{CiEKAA}\cite[Theorem~1]{CiEKAA}}. \begin{Theorem} \label{enum} If $a,b,c$ have the same parity, then the number of lozenge tilings of a hexagon with side lengths $a,b+m,c,a+m,b,c+m$, with an equilateral triangle of side length $m$ removed from its centre {\em(}see Figure~\ref{hex} for an example{\em)} is given by \begin{multline} \label{eq:enum} \frac {\h(a + m)\h(b + m)\h(c + m)\h(a + b + c + m) } {\h(a + b + m)\h(a + c + m)\h(b + c + m) } \frac {\h(m + \left \lceil {\frac{a + b + c}{2}} \right \rceil) \h(m + \left \lfloor {\frac{a + b + c}{2}} \right \rfloor) } {\h({\frac{a + b}{2}} + m) \h({\frac{a + c}{2}} + m)\h({\frac{b + c}{2}} + m) } \\ \times\frac {\h(\left \lceil {\frac{a}{2}} \right \rceil) \h(\left \lceil {\frac{b}{2}} \right \rceil) \h(\left \lceil {\frac{c}{2}} \right \rceil) \h(\left \lfloor {\frac{a}{2}} \right \rfloor)\, \h(\left \lfloor {\frac{b}{2}} \right \rfloor)\, \h(\left \lfloor {\frac{c}{2}} \right \rfloor)\, } {\h({\frac{m}{2}} + \left \lceil {\frac{a}{2}} \right \rceil)\, \h({\frac{m}{2}} + \left \lceil {\frac{b}{2}} \right \rceil)\, \h({\frac{m}{2}} + \left \lceil {\frac{c}{2}} \right \rceil)\, \h({\frac{m}{2}} + \left \lfloor {\frac{a}{2}} \right \rfloor)\, \h({\frac{m}{2}} + \left \lfloor {\frac{b}{2}} \right \rfloor)\, \h({\frac{m}{2}} + \left \lfloor {\frac{c}{2}} \right \rfloor)\, }\\ \times \frac {\h(\frac{m}{2})^2 \h({\frac{a + b + m}{2}})^2 \h({\frac{a + c + m}{2}})^2 \h({\frac{b + c + m}{2}})^2 } {\h({\frac{m}{2}} + \left \lceil {\frac{a + b + c}{2}} \right \rceil) \h({\frac{m}{2}} + \left \lfloor {\frac{a + b + c}{2}} \right \rfloor) \h({\frac{a + b}{2}})\h({\frac{a + c}{2}})\h({\frac{b + c}{2}}) }, \end{multline} where \begin{equation} \label{eq:hyperfac} \h(n):=\begin{cases} \prod _{k=0} ^{n-1}{\Gamma(k+1)}\quad &\text {for $n$ an integer,}\\ \prod _{k=0} ^{n-\frac {1} {2}}{\Gamma(k+\frac {1} {2})} \quad &\text {for $n$ a half-integer}. \end{cases} \end{equation} \quad \quad \qed \end{Theorem} (There is a similar theorem if the parities of $a,b,c$ should not be the same, see \machSeite{CiEKAA}\cite[Theorem~2]{CiEKAA}. Together, the two theorems generalise MacMahon's Theorem~\ref{thm:MM}.\footnote{Bijective proofs of Theorem~\ref{thm:MM} which ``explain" the ``nice" formula are known \machSeite{KratAY}% \machSeite{KratBK}% \cite{KratAY,KratBK}. I do not ask for a bijective proof of Theorem~\ref{enum} because I consider the task of finding one as daunting.}) \begin{figure} \centertexdraw{ \drawdim truecm \linewd.02 \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move(-.866025 -.5) \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \move(-1.7305 -1) \rhombus \rhombus \rhombus \ldreieck \rmove(.866025 -.5) \rhombus \rhombus \rhombus \move(-1.7305 -1) \rdreieck \rhombus \rhombus \rhombus \ldreieck \rhombus \rhombus \rhombus \move(-1.7305 -2) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \move(-1.7305 -3) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \htext(-.8 0){$a$} \htext(4 -.7){$b+m$} \htext(6.9 -3.95){$\left. \vbox{\vskip.6cm} \right\} c$} \htext(-3.2 -4){$c+m \left\{ \vbox{\vskip1.6cm} \right.$} \htext(0 -6){$b$} \htext(4.9 -6){$a+m$} \htext(1.7 -3.95){$\left. \vbox{\vskip1cm} \right\}m$} \rtext td:60 (4 -1.3) {$\left. \vbox{\vskip3.6cm} \right\} $} \rtext td:-60 (-.8 0){$\left\{ \vbox{\vskip1.6cm} \right. $} \rtext td:-60 (4.6 -5.3) {$\left. \vbox{\vskip2.5cm} \right\} $} \rtext td:60 (0.3 -5.6){$\left\{ \vbox{\vskip2.7cm}\right. $} } \caption{\protect\small A hexagon with triangular hole} \label{hex} \end{figure} The reader should notice that the right-hand side of \eqref{eq:M2} is indeed of the form $NICE(a)$, while the right-hand side of \eqref{eq:enum} is of the form $NICE(m/2)$. \medskip Where is the connexion to determinants? As it turns out, these two theorems are in fact {\it determinant evaluation theorems}. More precisely, Theorem~\ref{thm:MM} is equivalent to the following theorem. \begin{Theorem} \label{thm:MM2} \begin{equation} \label{eq:M3} \det_{1\le i,j\le c}\(\binom {a+b}{a-i+j} \)= \prod _{i=1} ^{c}\frac {(a+b+i-1)!\,(i-1)!} {(a+i-1)!\,(b+i-1)!}. \end{equation} \quad \quad \qed \end{Theorem} (The reader should notice that this is exactly \eqref{eq:M1} with $n$ replaced by $c$.) On the other hand, Theorem~\ref{enum} is equivalent to the theorem below.\footnote{To be correct, this is a little bit oversimplified. The truth is that equivalence holds only if $m$ is even. An additional argument is necessary for proving the result for the case that $m$ is odd. We refer the reader who is interested in these details to \machSeite{CiEKAA}\cite[Sec.~2]{CiEKAA}.} \begin{Theorem} \label{enum2} If $m$ is even, the determinant \begin{equation} \label{mat1} \det_{1\le i,j\le a+m} \begin{pmatrix} \dbinom{b+c+m}{b-i+j}& \text {\scriptsize $1\le i\le a$}\\ \dbinom{\frac {b+c} {2}}{\frac {b+a} {2}-i+j}& \text {\scriptsize $a+1\le i \le a+m$} \end{pmatrix} \end{equation} is equal to \eqref{eq:enum}.\quad \quad \qed \end{Theorem} \begin{figure} \centertexdraw{ \drawdim cm \setunitscale.7 \linewd.01 \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \ldreieck \move(-.866025 -.5) \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \move(-1.73205 -1) \rhombus \rhombus \rhombus \ldreieck \rmove(.866025 -.5) \rhombus \rhombus \rhombus \move(-1.73205 -1) \rdreieck \rhombus \rhombus \rhombus \ldreieck \rhombus \rhombus \rhombus \move(-1.73205 -2) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \rhombus \move(-1.73205 -3) \rdreieck \rhombus \rhombus \rhombus \rhombus \rhombus \linewd.08 \move(0 0) \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusB \move(-.866025 -.5) \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusB \RhombusA \move(-1.73205 -1) \RhombusA \RhombusA \RhombusB \RhombusB \RhombusA \RhombusB \RhombusA \RhombusA \move(2.598 -3.5) \RhombusA \RhombusB \RhombusA \move(1.73205 -4) \RhombusB \RhombusA \RhombusA \move(-1.73205 -1) \RhombusC \RhombusC \move(-1.73205 -2) \RhombusC \move(-1.73205 -3)\RhombusC \RhombusC \RhombusC \move(.866025 -1.5) \RhombusC \move(.866025 -2.5) \RhombusC \htext(-4 -8){\small a. A lozenge tiling of the cored hexagon in Figure~\ref{hex}} \htext(-1 0.2){$a$} \htext(4.4 -1){$b+m$} \htext(6.9 -3.9){$\left. \vbox{\vskip.35cm} \right\} c$} \htext(-3.8 -4){$c+m \left\{ \vbox{\vskip1.2cm} \right.$} \htext(0 -6.2){$b$} \htext(5 -6){$a+m$} \htext(1.7 -4){$\left. \vbox{\vskip0.77cm} \right\}m$} \rtext td:60 (4 -1.3) {$\left. \vbox{\vskip2.52cm} \right\} $} \rtext td:-60 (-.8 0.2){$\left\{ \vbox{\vskip1.02cm} \right. $} \rtext td:-60 (4.6 -5.3) {$\left. \vbox{\vskip1.75cm} \right\} $} \rtext td:60 (0.3 -5.8){$\left\{ \vbox{\vskip1.89cm}\right. $} \move(11 0) \bsegment \linewd.05 \move(0 0) \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusB \move(-.866025 -.5) \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusA \RhombusB \RhombusA \move(-1.73205 -1) \RhombusA \RhombusA \RhombusB \RhombusB \RhombusA \RhombusB \RhombusA \RhombusA \move(2.598 -3.5) \RhombusA \RhombusB \RhombusA \move(1.73205 -4) \RhombusB \RhombusA \RhombusA \move(-1.73205 -1) \RhombusC \RhombusC \move(-1.73205 -2) \RhombusC \move(-1.73205 -3)\RhombusC \RhombusC \RhombusC \move(.866025 -1.5) \RhombusC \move(.866025 -2.5) \RhombusC \ringerl(.433012 .25) \hdSchritt \hdSchritt \hdSchritt \hdSchritt \hdSchritt \hdSchritt \hdSchritt \vdSchritt \ringerl(-.433012 -.25) \hdSchritt \hdSchritt \hdSchritt \hdSchritt \hdSchritt \hdSchritt \vdSchritt \hdSchritt \ringerl(-1.299037 -.75) \hdSchritt \hdSchritt \vdSchritt \vdSchritt \hdSchritt \vdSchritt \hdSchritt \hdSchritt \ringerl(3.031 -3.25) \hdSchritt \vdSchritt \hdSchritt \ringerl(2.165 -3.75) \vdSchritt \hdSchritt \hdSchritt \ringerl(6.4952 -4.25) \ringerl(5.6292 -4.75) \ringerl(4.7632 -5.25) \ringerl(3.8971 -5.75) \ringerl(3.031 -6.25) \htext(-2 -8){\small b. The corresponding path family} \esegment \htext(4 -18){\small c. The path family made orthogonal}} \vskip-7cm $$ \Einheit=.7cm \Gitter(11,7)(0,0) \Koordinatenachsen(11,7)(0,0) \Pfad(0,3),11\endPfad \Pfad(2,1),22\endPfad \Pfad(2,1),1\endPfad \Pfad(3,0),2\endPfad \Pfad(3,0),11\endPfad \Pfad(1,4),111111\endPfad \Pfad(7,3),2\endPfad \Pfad(7,3),1\endPfad \Pfad(2,5),1111111\endPfad \Pfad(9,4),2\endPfad \Pfad(4,1),2\endPfad \Pfad(4,1),11\endPfad \Pfad(5,3),1\endPfad \Pfad(6,2),2\endPfad \Pfad(6,2),1\endPfad \Kreis(0,3.02) \Kreis(5,0) \Label\lo{A_1}(0,3) \Label\ru{E_1}(5,0) \Kreis(1,4.02) \Kreis(8,3) \Label\lo{A_2}(1,4) \Label\ru{E_4}(8,3) \Kreis(2,5.02) \Kreis(9,4) \Label\lo{A_3}(2,5) \Label\ru{E_5}(9,4) \Kreis(4,2.02) \Kreis(6,1) \Label\lo{A_4}(4,2) \Label\ru{E_2}(6,1) \Kreis(5,3.02) \Kreis(7,2) \Label\lo{A_5}(5,3) \Label\ru{E_3}(7,2) \hskip6cm $$ \vskip1cm \caption{} \label{tiling} \end{figure} The link between rhombus tilings (and equivalent objects such as plane partitions, semistandard tableaux, etc.) and determinants which explains the above two equivalence statements is {\it non-intersecting lattice paths}.\footnote{There exists in fact a second link between rhombus tilings and determinants which is not less interesting or less important. It is a well-known fact that rhombus tilings are in bijection with perfect matchings of certain hexagonal graphs. (See for example \machSeite{KupeAG}\cite[Figures~13 and 14]{KupeAG}.) In view of this fact, this second link is given by Kasteleyn's theorem \machSeite{KastAA}\cite{KastAA} saying that the number of perfect matchings of a planar graph is given by the Pfaffian of a slight perturbation of the adjacency matrix of the graph. See \machSeite{KupeAG}\cite{KupeAG} for an exposition of Kasteleyn's result, including historical notes, and for adaptations taking symmetries of the graph into account.} The latter are families of paths in a lattice with the property that no two paths in the family have a point in common. Indeed, rhombus tilings are (usually) in bijection with families of non-intersecting paths in the integer lattice ${\mathbb Z}^2$ which consist of unit horizontal and vertical steps. (Figure~\ref{tiling} illustrates the bijection for the rhombus tilings which appear in Theorem~\ref{enum} in an example. In that bijection, all horizontal steps of the paths are in the positive direction, and all vertical steps are in the negative direction. See the explanations that accompany \machSeite{CiEKAA}\cite[Figure~8]{CiEKAA} for a detailed description. Since, as I explained, Theorem~\ref{enum} essentially is a generalisation of Theorem~\ref{thm:MM}, this gives also an idea for the bijection for the rhombus tilings which appear in the latter theorem. For other instances of bijections between rhombus tilings and non-intersecting lattice paths see \machSeite{CiKrAA}% \machSeite{CiKrAC}% \machSeite{CiKrAD}% \machSeite{EisTAA}% \machSeite{EisTAB}% \machSeite{EisTAF}% \machSeite{FiscAA}% \machSeite{KratBY}% \machSeite{OkKrAA}% \cite{CiKrAA,CiKrAC,CiKrAD,EisTAA,EisTAB,EisTAF,FiscAA,KratBY,OkKrAA}.). In the case that the starting points and the end points of the lattice paths are fixed, the following many-author-theorem applies.\footnote{% This result was discovered and rediscovered several times. In a probabilistic form, it occurs for the first time in work by Karlin and McGregor \machSeite{KaMGAB}% \machSeite{KaMGAC}% \cite{KaMGAB,KaMGAC}. In matroid theory, it is discovered in its discrete form by Lindstr\"om \machSeite{LindAA}\cite[Lemma~1]{LindAA}. Then, in the 1980s the theorem is rediscovered at about the same time in three different communities, not knowing from each other at the time: in statistical physics by Fisher \machSeite{FishAA}\cite[Sec.~5.3]{FishAA} in order to apply it to the analysis of vicious walkers as a model of wetting and melting, in combinatorial chemistry by John and Sachs \machSeite{JoSaAB}\cite{JoSaAB} and Gro\-nau, Just, Schade, Scheffler and Wojciechowski \machSeite{GrJSAA}\cite{GrJSAA} in order to compute Pauling's bond order in benzenoid hydrocarbon molecules, and in enumerative combinatorics by Gessel and Viennot \machSeite{GeViAA}% \machSeite{GeViAB}% \cite{GeViAA,GeViAB} in order to count tableaux and plane partitions. Since only Lindstr\"om, and then Gessel and Viennot state the result in its most general form (not reproduced here), I call this theorem most often the ``Lindstr\"om--Gessel--Viennot theorem." It must be also mentioned that the so-called ``Slater determinant" in quantum mechanics (cf.\ \machSeite{SlatZY}\cite{SlatZY} and \machSeite{SlatZZ}\cite[Ch.~11]{SlatZZ}) may qualify as an ``ancestor" of the Lindstr\"om--Gessel--Viennot determinant.} \begin{Theorem}[\sc Karlin--McGregor, Lindstr\"om, Gessel--Viennot, Fisher,\break John--Sachs, Gronau--Just--Schade--Scheffler--Wojciechowski] \label{thm:nonint} Let $A_1,A_2,\break \dots,A_n$ and $E_1,E_2,\dots,E_n$ be lattice points such that for $i<j$ and $k<l$ any lattice path between $A_i$ and $E_l$ has a common point with any lattice path between $A_j$ and $E_k$. Then the number of all families $(P_1,P_2,\dots,P_n)$ of non-intersecting lattice paths, $P_i$ running from $A_i$ to $E_i$, $i=1,2,\dots,n$, is given by $$\det_{1\le i,j\le n}\big(P(A_j\to E_i)\big),$$ where $P(A\to E)$ denotes the number of all lattice paths from $A$ to $E$.\quad \quad \qed \end{Theorem} It goes beyond the scope of this article to include the proof of this theorem here. However, I cannot help telling that it is an extremely beautiful and simple proof that {\it every} mathematician should have seen once, even if (s)he does not have any use for it in her/his own research. I refer the reader to \machSeite{GeViAA}% \machSeite{GeViAB}% \machSeite{StemAE}% \cite{GeViAA,GeViAB,StemAE}. \medskip Now the origin of the determinants becomes evident. In particular, since, for rhombus tilings, we have to deal with lattice paths in the integer lattice consisting of unit horizontal and vertical steps, and since the number of such lattice paths which connect two lattice points is given by a binomial coefficient, we see that the enumeration of rhombus tilings must be a rich source for binomial determinants. This is indeed the case, and there are several instances in which such determinants can be evaluated in the form $NICE(.)$ (see \machSeite{CiucAH}% \machSeite{CiEKAA}% \machSeite{CiKrAA}% \machSeite{CiKrA}% \machSeite{CiKrAD}% \machSeite{EisTAA}% \machSeite{EisTAB}% \machSeite{EisTAF}% \machSeite{FiscAA}% \machSeite{FuKrAC}% \machSeite{KratBN}% \cite{CiucAH,CiEKAA,CiKrAA,CiKrAC,CiKrAD,EisTAA,EisTAB,EisTAF,FiscAA,FuKrAC,KratBN} and Section~\ref{sec:detlist}). Often the evaluation part is highly non-trivial. The evaluation of the determinant \eqref{eq:M3} is not very difficult (see \machSeite{KratBN}\cite[Sections~2.2, 2.3, 2.5]{KratBN} for 3 different ways to evaluate it). On the other hand, the evaluation of the determinant \eqref{mat1} requires some effort (see \machSeite{CiEKAA}\cite[Sec.~7]{CiEKAA}). \medskip To conclude this section, I state another determinant evaluation, to which I shall come back later. Its origin lies as well in the enumeration of rhombus tilings and plane partitions (see \machSeite{KratBD}\cite[Theorem~10]{KratBD} and \machSeite{CiKrAB}\cite[Theorem~2.1]{CiKrAB}). \begin{Theorem} \label{thm:xy} For any complex numbers $x$ and $y$ there holds \begin{multline} \label{eq:Krat} \det_{0\le i,j\le n-1}\(\frac {(x+y+i+j-1)!} {(x+2i-j)!\,(y+2j-i)!}\)\\ =\prod _{i=0} ^{n-1}\frac {i!\,(x+y+i-1)!\,(2x+y+2i)_i\,(x+2y+2i)_i} {(x+2i)!\,(y+2i)!}, \end{multline} where the {\em shifted factorials} or {\em Pochhammer symbols} $(a)_k$ are defined by $(a)_k:=a(a+1)\cdots(a+k-1)$, $k\ge1$, and $(a)_0:=1$. {\em(}In this formula, a factorial $m!$ has to be interpreted as $\Gamma(m+1)$ if $m$ is not a non-negative integer.{\em)} \end{Theorem} \section{A determinant from number theory}\label{sec:det} However, determinants do not only arise in combinatorics, they also arise in other fields. In this section, I want to present a determinant which arose in number theory, explain in some detail its origin, and then outline the steps which led to its evaluation, thereby giving the reader an opportunity to look ``behind the scenes" while one tries to make the determinant evaluation methods described in \machSeite{KratBN}\cite{KratBN} work. The story begins with the following two series expansions for $\pi$. The first one is due to Bill Gosper \machSeite{Gosper}\cite{Gosper}, \begin{equation} \label{eq:Gosper} \pi =\sum_{n=0}^\infty \frac{50n-6}{\binom{3n}n2^n} , \end{equation} and was used by Fabrice Bellard \machSeite{Bellard}\cite[file {\tt pi1.c}]{Bellard} to find an algorithm for computing the $n$-th decimal of $\pi $ without computing the earlier ones, thus improving an earlier algorithm due to Simon Plouffe \machSeite{Plouffe}\cite{Plouffe}. The second one, \begin{equation} \label{eq:Bellard} \pi=\frac {1} {740025}\(\sum _{n=1} ^{\infty}\frac {3P(n)} {\binom {7n}{2n}2^{n-1}}-20379280\), \end{equation} where \begin{multline*} P(n)=-885673181n^5+3125347237n^4-2942969225n^3\\ +1031962795n^2-196882274n+10996648, \end{multline*} is due to Fabrice Bellard \machSeite{Bellard}\cite{Bellard}, and was used by him in his world record setting computation of the 1000 billionth {\it binary} digit of $\pi$, being based on the algorithm in \machSeite{BaBoPl}\cite{BaBoPl}. Going beyond that, my co-authors from \machSeite{AlKPAA}\cite{AlKPAA}, Gert Almkvist and Joakim Petersson, asked themselves the following question: {\em Are there more expansions of the type $$\pi=\sum_{n=0}^\infty \frac {S(n)}{\binom{mn}{pn}a^n},$$ where $S(n)$ is some polynomial in $n$ {\em(}depending on $m,p,a${\em)}?} How can one go about to get some intuition about this question? One chooses some specific $m,p,a$, goes to the computer, computes $$p(k)=\sum _{n=0} ^{\infty}\frac {n^k} {\binom {mn}{pn}a^n}$$ to many, many digits for $k=0,1,2,\dots$, puts $$\pi,p(0),p(1),p(2),\dots$$ into the LLL-algorithm (which comes, for example, with the {\sl Maple} computer algebra package), and one waits whether the algorithm comes up with an integral linear combination of $\pi,p(0),p(1),p(2),\dots$.\footnote{For readers unfamiliar with the LLL-algorithm: in this particular application, it takes as an input rational numbers $r_1,r_2,\dots,r_m$ (which, in our case, will be the numbers $1$ and the rational approximations of $\pi$, $p(0)$, $p(1)$, \dots \ which we computed), and, if successful, outputs {\it small\/} integers $c_1,c_2,\dots,c_m$ such that $c_1r_1+c_2r_2+\dots+c_mr_m$ is {\it very small}. Thus, if $r_i$ was a good approximation for the real number $x_i$, $i=1,2,\dots,m$, one can expect that actually $c_1x_1+c_2x_2+\dots+c_mx_m=0$. See \machSeite{LeLLAA}\cite[Sec.~1, in particular the last paragraph]{LeLLAA} and \machSeite{CohHAA}\cite[Ch.~2]{CohHAA} for the description of and more information on this important algorithm. In particular, also here, the output of the algorithm (if there is) is just a (very guided) {\it guess}. Thus, a proof is still needed, although the probability that the guess is wrong is infinitesimal. As a matter of fact, it is very likely that Bellard had no proof of his \hbox{formula \eqref{eq:Bellard} \dots}} Indeed, Table~\ref{tab:2} shows the parameter values, where the LLL-algorithm gave a result. \begin{table}[h] \begin{tabular}{c|c|l|cl} $m$ & $p$ & $\hphantom{-}a$ & $\deg(S)$ & \\ \cline{1-4} \hphantom{1}3 & \hphantom{1}1 & $\hphantom{-}2$ & \hphantom{1}1 & (Gosper) \\ \hphantom{1}7 & \hphantom{1}2 & $\hphantom{-}2$ & \hphantom{1}5 & (Bellard) \\ \hphantom{1}8 & \hphantom{1}4 & $-4$ & \hphantom{1}4 & \\ 10 & \hphantom{1}4 & $\hphantom{-}4$ & \hphantom{1}8 & \\ 12 & \hphantom{1}4 & $-4$ & \hphantom{1}8 & \\ 16 & \hphantom{1}8 & $\hphantom{-}16$ & \hphantom{1}8 & \\ 24 & 12 & $-64$ & 12 & \\ 32 & 16 & $\hphantom{-}256$ & 16 & \\ 40 & 20 & $-4^5$ & 20 & \\ 48 & 24 & $\hphantom{-}4^6$ & 24 & \\ 56 & 28 & $-4^7$ & 28 & \\ 64 & 32 & $\hphantom{-}4^8$ & 32 & \\ 72 & 36 & $-4^9$ & 36 & \\ 80 & 40 & $\hphantom{-}4^{10}$ & 40 & \end{tabular} \vskip10pt \caption{} \label{tab:2} \end{table} For example, it found $$ \pi =\frac 1r\sum_{n=0}^\infty \frac{S(n)}{\binom{16n}{8n}16^n}, $$ where $$ r=3^65^37^211^213^2 $$ and \begin{multline*} S(n)=-869897157255-3524219363487888n+112466777263118189n^2 \\ -1242789726208374386n^3+6693196178751930680n^4-19768094496651298112n^5 \\ +32808347163463348736n^6-28892659596072587264n^7+10530503748472012800n^8, \end{multline*} and $$ \pi =\frac 1r\sum_{n=0}^\infty \frac{S(n)}{\binom{32n}{16n}256^n} , $$ where $$ r=2^33^{10}5^67^311^1 13^217^219^223^229^231^2 $$ and {\allowdisplaybreaks \begin{align*} S(n)=&-2062111884756347479085709280875 \\ &+1505491740302839023753569717261882091900n \\ &-112401149404087658213839386716211975291975n^2 \\ &+3257881651942682891818557726225840674110002n^3 \\ &-51677309510890630500607898599463036267961280n^4 \\ &+517337977987354819322786909541179043148522720n^5 \\ &-3526396494329560718758086392841258152390245120n^6 \\ &+171145766235995166227501216110074805943799363584n^7 \\ &-60739416613228219940886539658145904402068029440n^8 \\ &+159935882563435860391195903248596461569183580160n^9 \\ &-313951952615028230229958218839819183812205608960n^{10} \\ &+457341091673257198565533286493831205566468325376n^{11} \\ &-486846784774707448105420279985074159657397780480n^{12} \\ &+367314505118245777241612044490633887668208926720n^{13} \\ &-185647326591648164598342857319777582801297080320n^{14} \\ &+56224688035707015687999128994324690418467340288n^{15} \\ &-7687255778816557786073977795149360408612044800n^{16} . \end{align*}}% Of course, there could be many more. If one looks more closely at Table~\ref{tab:2}, then, if one disregards the first, second and fourth line, one cannot escape to notice a pattern: {\it apparently, for each $k=1,2,\dots$, there is a formula $$ \pi =\sum_{n=0}^\infty \frac{S_k(n)}{\binom{8kn}{4kn}(-4)^{kn}} , $$ where $S_k(n)$ is some polynomial in $n$ of degree $4k$.} In order to make progress on this observation, we have to first see how one can prove such an identity, once it is found. In fact, this is not difficult at all. To illustrate the idea, let us go through a proof of Gosper's identity \eqref{eq:Gosper}. The beta integral evaluation (cf.\ \machSeite{AAR}\cite[Theorem~1.1.4]{AAR}) gives $$ \frac 1{\binom{3n}n}=(3n+1)\int_0^1x^{2n}(1-x)^ndx . $$ Hence the right hand side of the formula will be $$ \int_0^1\sum_{n=0}^\infty (50n-6)(3n+1)\left(\frac {x^2(1-x)}{2}\right)^ndx . $$ We have \begin{equation} \label{eq:rational} \sum_{n=0}^\infty (50n-6)(3n+1)y^n=\frac{2(56y^2+97y-3)}{(1-y)^3} . \end{equation} Thus, if substituted, we obtain \begin{align}\notag RHS&=8\int_0^1\frac{28x^6-56x^5+28x^4-97x^3+97x^2-6}{(x^3-x^2+2)^3}dx\\ &= \left[ \frac{4x(x-1)(x^3-28x^2+9x+8)}{(x^3-x^2+2)^2}+4\arctan (x-1)\right] _0^1=\pi . \label{eq:RHS} \end{align} (Clearly, both \eqref{eq:rational} and \eqref{eq:RHS} are routine calculations, and therefore we did not do it by hand, but let them be worked out by {\sl Maple}.) Now let us fix $k\ge1$. We apply the same procedure to $ \sum_{n=0}^\infty {S_k(n)}\big/{\binom{8kn}{4kn}(-4)^{kn}} , $ where $S_k(n)$ is (hopefully) some (unknown) polynomial in $n$. The beta integral evaluation gives $$ \frac 1{\binom{8kn}{4kn}}=(8kn+1)\int_0^1x^{4kn}(1-x)^{4kn}dx . $$ Hence, if $S_k(n)$ should have degree $d$ in $n$, \begin{align}\notag \sum _{n=0} ^{\infty}\frac {S_k(n)}{\binom{8kn}{4kn}(-4)^{kn}} &=\int_0^1\sum_{n=0}^\infty (8kn+1)S_k(n)\left(\frac {x^{4k}(1-x)^{4k}}{(-4)^k}\right)^n\,dx \notag\\ &=\int_0^1 \frac {P_k(x)} {\(x^{4k}(1-x)^{4k}-(-4)^k\)^{d+2}}\,dx, \label{eq:sumint} \end{align} where $P_k(x)$ is some polynomial in $x$. For convenience, let us write $P$ as a short-hand for $P_k$. Let $Q(x):=x^{4k}(1-x)^{4k}-(-4)^k$. Now we make the wild assumption that $$ \int \frac{P(x)}{Q(x)^{d+2}}\,dx=\frac{R(x)}{Q(x)^{d+1}}+2\arctan (x)+2\arctan (x-1) , $$ for some polynomial $R(x)$ with $R(0)=R(1)=0$. Then the original sum would indeed be equal to $\pi$. The last equality is equivalent to $$ \frac P{Q^{d+2}}=\frac{R^{\prime }}{Q^{d+1}}-(d+1)\frac{Q^{\prime }R}{Q^{d+2}% }+2\left(\frac 1{x^2+1}+\frac 1{x^2-2x+2}\right), $$ or $$QR^{\prime }-(d+1)Q^{\prime }R=P-2Q^{d+2}\left(\frac 1{x^2+1}+ \frac 1{x^2-2x+2}\right) . $$ In our examples, we observed that $$ R(x)=(2x-1)\check{R}\big(x(1-x)\big) $$ for a polynomial $\check R$. So, let us make the substitution $$t=x(1-x).$$ Then, after some simplification, the above differential equation becomes \begin{equation} \label{eq:diff} -(1-4t)Q\frac{d\check{R}}{dt}+(2Q+4k(4k+1)(1-4t)t^{4k-1})\check{R}-P+2(3-2t)% \frac{Q^{4k+2}}{t^2-2t+2}=0 , \end{equation} where $Q(t)=t^{4k}-(-4)^{k}$. Now, writing $N(k)=4k(4k+1)$, we make the Ansatz \begin{align*} \check R(t)&=\sum _{j=1} ^{N(k)-1}a(j)t^j,\\ S_k(n)&=\sum _{j=0} ^{4k}a(N(k)+j)n^j. \end{align*} (The reader should recall that $S_k(n)$ defines $P_k(t)=P(t)$ through \eqref{eq:sumint}.) Comparing coefficients of powers of $t$ on both sides of \eqref{eq:diff}, we get a system of $N(k)+4k$ linear equations for the unknowns $a(1),a(2),\dots,a(N(k)+4k)$. Hence: {\it If the determinant of this system of linear equations is non-zero, then there does indeed exist a representation $$\pi=\sum_{n=0}^\infty \frac {S_k(n)}{\binom{8kn}{4kn}(-4)^n}.$$ } To see whether we could indeed hope for the determinant to be non-zero, we went again to the computer and looked at the values of the determinant in some small instances. (Obviously, we do not want to do this by hand, since for $k=1$ the matrix is already a $24\times 24$ matrix!) So, let us program the matrix. (We shall see the mathematical definition of the matrix in just a moment, see \eqref{eq:Det}.\footnote{To tell the truth, this is the form of the matrix after some simplifications have already been carried out. (In particular, we are looking at a matrix which is slightly smaller than the original one.) See \machSeite{AlKPAA}\cite[beginning of Section~4]{AlKPAA} for these details. There, the matrix in \eqref{eq:Det} is called $M'''$.}) \MATH \goodbreakpoint% In[8]:= a[k\MATHtief ,j\MATHtief ]:=Module[% \MATHlbrace Var=j/(4k)% \MATHrbrace , \leavevmode% (-1)\MATHhoch (Var-1)*8k(4k+1)(-4)\MATHhoch (k*(Var+1))* \leavevmode% Product[4k*l-1,% \MATHlbrace l,1,4k-Var% \MATHrbrace ]*Product[4k*l+1,% \MATHlbrace l,1,Var-1% \MATHrbrace ] \leavevmode% ] In[9]:= A[k\MATHtief , i\MATHtief , j\MATHtief ] := Module[% \MATHlbrace Var% \MATHrbrace , \leavevmode% Var = % \MATHlbrace Floor[(i - 2)/(4*k - 1)], \leavevmode% Floor[(j - 1)/(4*k)], Mod[i - 2, 4*k - 1], \leavevmode% Mod[j - 1, 4*k]% \MATHrbrace ; \leavevmode% If[i == 1, \leavevmode% If[Mod[j, 4*k] === 0, a[k, j], 0], \leavevmode% If[Var[[1]] - Var[[2]] == 0, \leavevmode% Switch[Var[[3]] - Var[[4]], 0, f1[k, Var[[3]]+1, j], -1, \leavevmode% f0[k, Var[[3]]+1, j], \MATHtief , 0], \leavevmode% If[Var[[1]] - Var[[2]] == 1, \leavevmode% Switch[Var[[3]] - Var[[4]], 0, g1[k, Var[[3]]+1, j], -1, \leavevmode% g0[k, Var[[3]]+1, j], \MATHtief , 0], 0]]]] \goodbreakpoint% In[10]:= A[k\MATHtief ] := Table[A[k, i, j], % \MATHlbrace i, 1, 16*k\MATHhoch 2% \MATHrbrace , % \MATHlbrace j, 1, 16*k\MATHhoch 2% \MATHrbrace ] \goodbreakpoint% In[11]:= f0[k\MATHtief , t\MATHtief , j\MATHtief ] := j*(-4)\MATHhoch k; \leavevmode% f1[k\MATHtief , t\MATHtief , j\MATHtief ] := -(2 + 4*j)*(-4)\MATHhoch k; \leavevmode% g0[k\MATHtief , t\MATHtief , j\MATHtief ] := (4*k*(4*k + 1) - j); \leavevmode% g1[k\MATHtief , t\MATHtief , j\MATHtief ] := (-4*4*k*(4*k + 1) + 2 + 4*j) \goodbreakpoint% \endgroup We shall not try to digest this at this point. Let us accept the program as a black box, and let us compute the determinant for $k=2$. \MATH \goodbreakpoint% In[12]:= Det[A[2]] \goodbreakpoint% Out[12]= -601576375580370166777074138698518196031142518971568946712\MATHbackslash > 2204136674781038302774231725971306459064075121023092662279814\MATHbackslash > 015195545600000000000 \goodbreakpoint% \endgroup Magnificent! This is certainly {\it not\/} zero. However, what are we going to do with this gigantic number? Remembering our discussion about ``nice" numbers and ``nice" formulae in the preceding section, let us factorise it in its prime factors. \MATH \goodbreakpoint% In[13]:= FactorInteger[\%] \goodbreakpoint% Out[13]= % \MATHlbrace % \MATHlbrace -1, 1% \MATHrbrace , % \MATHlbrace 2, 325% \MATHrbrace , % \MATHlbrace 3, 39% \MATHrbrace , % \MATHlbrace 5, 11% \MATHrbrace , % \MATHlbrace 7, 11% \MATHrbrace , % \MATHlbrace 11, 3% \MATHrbrace , % \MATHlbrace 13, 2% \MATHrbrace % \MATHrbrace \goodbreakpoint% \endgroup I would say that this is sensational: a number with 139 digits, and the biggest prime factor is 13! As a matter of fact, this is not just a rare exception. Table~\ref{tab:1} shows the factorisations of the first five determinants. (We could not go further because of the exploding size of the matrix of which the determinant is taken.) \begin{table}[h] \vskip10pt \begin{tabular}{l|l} \hphantom{1}$k$ & $\hphantom{-}\det(A(k))$ \\ \hline\\[-8pt] \hphantom{1}1 & $\hphantom{-}2^{59}3^55^67^1$ \\ \hphantom{1}2 & $-2^{325}3^{39}5^{11}7^{11}11^313^2$ \\ \hphantom{1}3 & $\hphantom{-}2^{772}3^{146}5^{28}7^{17}11^{17}13^{18}17^419^323^1$ \\ \hphantom{1}4 & $-2^{1913}3^{111}5^{58}7^{38}11^{21}13^{22}17^{24}19^723^529^231^1$ \\ \hphantom{1}5 & $\hphantom{-} 2^{2932}3^{202}5^{306}7^{69}11^{29}13^{27}17^{28}19^{29}23^{9}29^631^537^2$% \end{tabular} \vskip10pt \caption{} \label{tab:1} \end{table} Thus, these experimental results {\it make us sure} that there must be a ``nice" formula for the determinant. Indeed, we prove in \machSeite{AlKPAA}\cite{AlKPAA} that\footnote{Strictly speaking, this is not a formula $NICE(k)$ according to my ``Definition" in the preceding section, because of the presence of the ``Abelian" factors $k^{8k^2+2k}$ and $(4k+1)!^{4k}$, see Footnote~\ref{foot:normally}. Nevertheless, the reader will certainly admit that this is a {\it nice} and {\it closed\/} formula.} \begin{equation} \label{eq:det(A(k))} \det (A(k))=(-1)^{k-1} 2^{16k^3+20k^2+6k}k^{8k^2+2k}(4k+1)!^{4k} \prod_{j=1}^{4k}\frac{(2j)!}{j!^2}. \end{equation} Hence the desired theorem follows. \begin{Theorem} \label{T1} For all $k\geq 1$ there is a formula $$ \pi =\sum_{n=0}^\infty \frac{S_k(n)}{\binom{8kn}{4kn}(-4)^{kn}}, $$ where $S_k(n)$ is a polynomial in $n$ of degree $4k$ with rational coefficients. The polynomial $S_k(n)$ can be found by solving the previously described system of linear equations. \end{Theorem} I must admit that we were extremely lucky that it was indeed possible to {\it evaluate} the determinant {\it explicitly}. To recall, ``all" we needed to prove our theorem (Theorem~\ref{T1}) was to show that the determinant was {\it non-zero}. To be honest, I would not have the slightest idea how to do this here without finding the exact value of the determinant. \medskip Now, after all this somewhat ``dry" discussion, let me present the determinant. We had to determine the determinant of the $16k^2\times 16k^2$ matrix \begin{equation} \label{eq:Det} \begin{pmatrix} 0\dots0\,*& 0\dots0\,*& 0\dots0\,*& \dots& \dots& \dots& 0\dots0\,*\\ \hbox{\Large$F_1$}&\hbox{\Large$0$}&\hbox{\Large$0$}& \dots&\dots&\dots&\hbox{\Large$0$}\\ \hbox{\Large$G_1$}&\hbox{\Large$F_2$}& \hbox{\Large$0$}& \dots&\dots&\dots&\hbox{\Large$0$}\\ \hbox{\Large$0$}&\hbox{\Large$G_2$}&\hbox{\Large$F_3$}& &&&\vdots\\ \hbox{\Large$0$}&\hbox{\Large$0$}&\hbox{\Large$G_3$}& & \ddots& &\vdots\\ \vdots&\ddots&\ddots&\ddots&\ddots&\ddots&\vdots\\ \vdots&&\ddots&\ddots&\ddots&\hbox{\Large$F_{4k-1}$}&\hbox{\Large$0$}\\ \vdots&&&\hbox{\Large$0$}&\hbox{\Large$0$}& \hbox{\Large$G_{4k-1}$}&\hbox{\Large$F_{4k}$}\\ \hbox{\Large$0$}&\dots&\dots&\dots&\hbox{\Large$0$}&\hbox{\Large$0$}&\hbox{\Large$G_{4k}$} \end{pmatrix}, \end{equation} where the $\ell$-th non-zero entry in the first row (these are marked by $*$) is $$ (-1)^{\ell -1}(-4)^{(\ell +1)k}8k(4k+1)\left(\prod _{i=1} ^{4k-\ell }(4ik-1)\right) \left(\prod _{i=1} ^{\ell -1}(4ik+1)\right), $$ and where each block $F_t$ and $G_t$ is a $(4k-1)\times(4k)$ matrix (that is, these are {\it rectangular} blocks!) with non-zero entries only on its (two) main diagonals, $$ F_t =\left(\smallmatrix f_1(4(t -1)k+1)&f_0(4(t -1)k+2)&0&\dots\\ 0&f_1(4(t -1)k+2)&f_0(4(t -1)k+3)&0&\dots\\ &\ddots&\ddots&\\ &&\ddots&\ddots&\\ &&0&f_1(4t k-2)&f_0(4t k-1)&0\\ &&&0&f_1(4t k-1)&f_0(4t k)\\ \endsmallmatrix\right) $$ and $$ G_t =\left(\smallmatrix g_1(4(t -1)k+1)&g_0(4(t -1)k+2)&0&\dots\\ 0&g_1(4(t -1)k+2)&g_0(4(t -1)k+3)&0&\dots\\ &\ddots&\ddots&\\ &&\ddots&\ddots&\\ &&0&g_1(4t k-2)&g_0(4t k-1)&0\\ &&&0&g_1(4t k-1)&g_0(4t k)\\ \endsmallmatrix\right). $$ We have almost worked our way through the definition of the determinant. The only missing piece is the definition of the functions $f_0,f_1,g_0,g_1$ in the blocks $F_t$ and $G_t$. Here it is: \begin{align} \notag f_0(j)&=j(-4)^k,\\ \notag f_1(j)&=-(4j+2)(-4)^k,\\ \notag g_0(j)&=(N(k)-j),\\ \label{eq:fg} g_1(j)&=-(4N(k)-4j-2), \end{align} where, as before, we write $N(k)=4k(4k+1)$ for short. \section{The evaluation of the determinant} \label{sec:eval} I now describe how the determinant of \eqref{eq:Det} was evaluated by applying to it the methods described in \machSeite{KratBN}\cite{KratBN}. To make this section as self-contained as possible, for each of them I briefly recall how it works before putting it into action. \medskip {\it ``Method" 0}: {\it Do row and column operations until the determinant reduces to something manageable.} In fact, at a first glance, this does not look too bad. Our matrix \eqref{eq:Det}, of which we want to compute the determinant and show that it is non-zero, is a very sparse matrix. Moreover, it looks almost like a two-diagonal matrix. It seems that one should be able to do a few row and column manipulations and thus reduce the matrix to a matrix of a simpler form of which we can evaluate the determinant. Well, we tried that. Unfortunately, the above impression is deceiving. First of all, the diagonals of the blocks do not really fit together to form diagonals which run from one end of the matrix to the other. Second, there remains still the first row which does not fit the pattern of the rest of the matrix. So, whatever we did, we ended up nowhere. Maybe we should try something more \hbox{sophisticated \dots} \medskip {\it Method 1} \machSeite{KratBN}\cite[Sec.~2.6]{KratBN}: {\it LU-factorisation}. Suppose we are given a family of matrices $A(1), A(2),A(3),\dots$ of which we want to compute the determinants. Suppose further that we can write $$A(k)\cdot U(k)=L(k),$$ where $U(k)$ is an upper triangular matrix with 1s on the diagonal, and where $L(k)$ is a lower triangular matrix. Then, clearly, $$\det(A(k))=\text {product of the diagonal entries of $L(k)$}.$$ But how do we find $U(k)$ and $L(k)$? We go to the computer, crank out $U(k)$ and $L(k)$ for $k=1,2,3,\dots$, until we are able to make a guess. Afterwards we prove the guess by proving the corresponding identities. Well, we programmed that, we stared at the output on the computer screen, but we could not make any sense of it. \medskip {\it Method 2} \machSeite{KratBN}\cite[Sec.~2.3]{KratBN}: {\it Condensation}. This is based on a determinant formula due to Jacobi (see \machSeite{BresAO}\cite[Ch.~4]{BresAO} and \machSeite{KnutAF}\cite[Sec.~3]{KnutAF}). Let $A$ be an $n\times n$ matrix. Let $A_{i_1,i_2,\dots,i_\ell}^{j_1,j_2,\dots,j_\ell}$ denote the submatrix of $A$ in which rows $i_1,i_2,\dots,i_\ell$ and columns $j_1,j_2,\dots,j_\ell$ are omitted. Then there holds \begin{equation} \label{eq:cond} \hfill \det A\cdot \det A_{1,n}^{1,n}=\det A_{1}^{1}\cdot \det A_n^n- \det A_1^n\cdot \det A_n^1. \end{equation} If we consider a family of matrices $A(1),A(2),\dots$, and if all the consecutive minors of $A(n)$ belong to the same family, then this allows one to give an inductive proof of a conjectured determinant evaluation for $A(n)$. Let me illustrate this by reproducing Amdeberhan's condensation proof \machSeite{AmdeAD}\cite{AmdeAD} of \eqref{eq:Krat}. Let $M_n(x,y)$ denote the determinant in \eqref{eq:Krat}. Then we have \begin{align} \notag \big(M_n(x,y)\big)_n^n&=M_{n-1}(x,y), \\ \notag \big(M_n(x,y)\big)_1^1&=M_{n-1}(x+1,y+1), \\ \notag \big(M_n(x,y)\big)_n^1&=M_{n-1}(x-1,y+2), \\ \notag \big(M_n(x,y)\big)_1^n&=M_{n-1}(x+2,y-1), \\ \label{eq:minors} \big(M_n(x,y)\big)_{1,n}^{1,n}&=M_{n-2}(x+1,y+1). \end{align} Thus, we know that Equation \eqref{eq:cond} is satisfied with $A$ replaced by $M_n(x,y)$, where the minors appearing in \eqref{eq:cond} are given by \eqref{eq:minors}. This can be interpreted as a recurrence for the sequence $\big(M_n(x,y)\big)_{n\ge0}$. Indeed, given $M_0(x,y)$ and $M_1(x,y)$, the equation \eqref{eq:cond} determines $M_n(x,y)$ uniquely for all $n\ge0$ (given that $M_n(x,y)$ never vanishes). Thus, since the right-hand side of \eqref{eq:Krat} is indeed never zero, for the proof of \eqref{eq:Krat} it suffices to check \eqref{eq:Krat} for $n=0$ and $n=1$, and that the right-hand side of \eqref{eq:Krat} also satisfies \eqref{eq:cond}, all of which is a routine task. \medskip Now, a short glance at the definition of our matrix \eqref{eq:Det} will convince us quickly that application of this method to it is rather hopeless. For example, omission of the first row already brings us outside of our family of matrices. So, also this method is not much help to solve our problem, which is really a pity, because it is the most painless of \hbox{all \dots} \medskip {\it Method 3} \machSeite{KratBN}\cite[Sec.~2.4]{KratBN}: {\it Identification of factors}. In order to sketch the idea, let us quickly go through a (standard) proof of the {\it Vandermonde determinant evaluation}, \begin{equation} \label{eq:Vandermonde} \det_{1\le i,j\le n}\(X_i^{j-1}\)=\prod _{1\le i<j\le n} ^{}(X_j-X_i). \end{equation} \begin{proof}If $X_{i_1}=X_{i_2}$ with $i_1\ne i_2$, then the Vandermonde determinant \eqref{eq:Vandermonde} certainly vanishes because in that case two rows of the determinant are identical. Hence, $(X_{i_1}-X_{i_2})$ divides the determinant as a polynomial in the $X_i$'s. But that means that the complete product $\prod _{1\le i<j\le n} (X_j-X_i)$ (which is exactly the right-hand side of \eqref{eq:Vandermonde}) must divide the determinant. On the other hand, the determinant is a polynomial in the $X_i$'s of degree at most $\binom n2$. Combined with the previous observation, this implies that the determinant equals the right-hand side product times, possibly, some constant. To compute the constant, compare coefficients of $X_1^0X_2^1\cdots X_n^{n-1}$ on both sides of \eqref{eq:Vandermonde}. This completes the proof of \eqref{eq:Vandermonde}. \end{proof} At this point, let us extract the essence of this proof. The basic steps are: {\em \begin{enumerate} \item[(S1)] Identification of factors \item[(S2)] Determination of degree bound \item[(S3)] Computation of the multiplicative constant. \end{enumerate} } As I report in \machSeite{KratBN}\cite{KratBN}, this turns out to be an extremely powerful method which has numerous applications. To given an idea of the flavour of the method, I show a few steps when it is applied to the determinant in \eqref{eq:Krat} (ignoring the fact that we have already found a very simple proof of its evaluation; see \machSeite{KratBD}\cite[proof of Theorem~10]{KratBD} for the complete proof using the ``identification of factors" method). \medskip To get started, we have to transform the assertion \eqref{eq:Krat} into an assertion about polynomials. This is easily done, we just have to factor $$(x+y+i-1)!/(x+2i)!/(y+2n-i-2)!$$ out of the $i$-th row of the determinant. If we subsequently cancel common factors on both sides of \eqref{eq:Krat}, we arrive at the equivalent assertion \begin{multline} \label{eq:Krat1} \det_{0\le i,j\le n-1}\big((x+y+i)_{j}\,(x+2i-j+1)_j\,(y+2j-i+1)_{2n-2j-2}\big)\\ =\prod _{i=0} ^{n-1}\big(i!\,(y+2i+1)_{n-i-1}\,(2x+y+2i)_i\,(x+2y+2i)_i\big), \end{multline} where, as before, $(\alpha)_k$ is the standard notation for shifted factorials (Pochhammer symbols) explained in the statement of Theorem~\ref{thm:xy}. In order to apply the same idea as in the above evaluation of the Vandermonde determinant, as a first step we have to show that the right-hand side of \eqref{eq:Krat1} divides the determinant on the left-hand side as a polynomial in $x$ and $y$. For example, we would have to prove that $(x+2y+2n-2)$ (actually, $(x+2y+2n-2)^{\fl{(n+1)/3}}$, we will come to that in a moment) divides the determinant. Equivalently, if we set $x=-2y-2n+2$ in the determinant, then it should vanish. How could we prove that? Well, if it vanishes then there must be a linear combination of the columns, or of the rows, that vanishes. Equivalently, for $x=-2y-2n+2$ we find a vector in the kernel of the matrix in \eqref{eq:Krat1}, respectively of its transpose. More generally (and this addresses the fact that we actually want to prove that $(x+2y+2n-2)^{\fl{(n+1)/3}}$ divides the determinant): \medskip {\em \leftskip1cm \rightskip1cm \noindent For proving that $(x+2y+2n-2)^E$ divides the determinant, we find $E$ linear independent vectors in the kernel. \par} \medskip \noindent (For a formal justification that this does indeed suffice, see Section~2 of \machSeite{KratBI}\cite{KratBI}, and in particular the Lemma in that section.) Okay, how is this done in practice? You go to your computer, crank out these vectors in the kernel, for $n=1,2,3,\dots$, and try to make a guess what they are in general. To see how this works, let us do it in our example. First of all, we program the kernel of the matrix in \eqref{eq:Krat1} with $x=-2y-2n+2$ (again, we are using {\sl Mathematica} here).\footnote{In the program, {\tt V[n]} represents the kernel, which is clearly a vector space. In the computer output, it is given in parametric form, the parameters being the {\tt c[i]}'s.} \MATH \goodbreakpoint% In[14]:= p=Pochhammer; \leavevmode% m[i\MATHtief ,j\MATHtief ,n\MATHtief ]:=p[x+y+i, j]*p[y+2*j+1-i, 2*n-2*j-2]*p[x-j+1+2i,j]; \leavevmode% V[n\MATHtief ]:=(x=-2y-2n+2; \leavevmode% Var=Sum[c[j]*Table[m[i,j,n],\MATHlbrace i,0,n-1\MATHrbrace ],\MATHlbrace j,0,n-1\MATHrbrace ]; \leavevmode% Var=Solve[Var==Table[0,\MATHlbrace n\MATHrbrace ],Table[c[i],\MATHlbrace i,0,n-1\MATHrbrace ]]; \leavevmode% Factor[Table[c[i],\MATHlbrace i,0,n-1\MATHrbrace ]/.Var]) \goodbreakpoin \endgroup What the computer gives is the following: \MATH \goodbreakpoint% In[15]:= V[2] Out[15]= \MATHlbrace \MATHlbrace -2 c[1], c[1]\MATHrbrace \MATHrbrace \goodbreakpoin In[16]:= V[3] Out[16]= \MATHlbrace \MATHlbrace -2 c[2], -c[2], c[2]\MATHrbrace \MATHrbrace \goodbreakpoin In[17]:= V[4] Out[17]= \MATHlbrace \MATHlbrace -2 c[3], -3 c[3], 0, c[3]\MATHrbrace \MATHrbrace \goodbreakpoin In[18]:= V[5] Out[18]= \MATHlbrace \MATHlbrace -2 c[4], -5 c[4], -2 c[3] - c[4], c[3], c[4]\MATHrbrace \MATHrbrace \goodbreakpoin In[19]:= V[6] Out[19]= \MATHlbrace \MATHlbrace -2 c[5], -7 c[5], -2 (c[4] + 2 c[5]), -c[4], c[4], c[5]\MATHrbrace \MATHrbrace \goodbreakpoin In[20]:= V[7] Out[20]= \MATHlbrace \MATHlbrace -2 c[6], -9 c[6], -2 c[5] - 9 c[6], % -3 c[5] - c[6], 0, % > c[5], c[6]\MATHrbrace \MATHrbrace \goodbreakpoin \endgroup At this point, the computations become somewhat slow. So we should help our computer. Indeed, on the basis of what we have obtained so far, it is ``obvious" that, somewhat unexpectedly, $y$ does not appear in the result. Therefore we simply set $y$ equal to some random number, and then the computer can go much further without any effort. \MATH \goodbreakpoin In[21]:= y=101 \goodbreakpoin In[22]:= V[8] Out[22]= \MATHlbrace \MATHlbrace -2 c[7], -11 c[7], -2 (c[6] + 8 c[7]), -5 (c[6] + c[7]), > -2 c[5] - c[6], c[5], c[6], c[7]\MATHrbrace \MATHrbrace \goodbreakpoin In[23]:= V[9] Out[23]= \MATHlbrace \MATHlbrace -2 c[8], -13 c[8], -2 c[7] - 25 c[8], -7 (c[7] + 2 c[8]), > -2 c[6] - 4 c[7] - c[8], -c[6], c[6], c[7], c[8]\MATHrbrace \MATHrbrace \goodbreakpoin In[24]:= V[10] Out[24]= \MATHlbrace \MATHlbrace -2 c[9], -15 c[9], -2 (c[8] + 18 c[9]), -3 (3 c[8] + 10 c[9]), > -2 c[7] - 9 c[8] - 6 c[9], -3 c[7] - c[8], 0, c[7], c[8], c[9]\MATHrbrace \MATHrbrace \goodbreakpoin In[25]:= V[11] Out[25]= \MATHlbrace \MATHlbrace -2 c[10], -17 c[10], -2 c[9] - 49 c[10], -11 (c[9] + 5 c[10]), > -2 (c[8] + 8 c[9] + 10 c[10]), -5 c[8] - 5 c[9] - c[10], > -2 c[7] - c[8], c[7], c[8], c[9], c[10]\MATHrbrace \MATHrbrace \goodbreakpoin \endgroup Let us extract some information out of these data. For convenience, we write $M_n$ for the matrix in \eqref{eq:Krat1} in the sequel. For example, by just looking at the coefficients of $c[n-1]$ appearing in $V[n]$, we extract that \begin{enumerate} \item [] the vector $(-2,1)$ is in the kernel of $M_2$, \item [] the vector $(-2,-1,1)$ is in the kernel of $M_3$, \item [] the vector $(-2,-3,0,1)$ is in the kernel of $M_4$, \item [] the vector $(-2,-5,-1,0,1)$ is in the kernel of $M_5$, \item [] the vector $(-2,-7,-4,0,0,1)$ is in the kernel of $M_6$, \item [] the vector $(-2,-9,-9,-1,0,0,1)$ is in the kernel of $M_7$, \item [] the vector $(-2,-11,-16,-5,0,0,0,1)$ is in the kernel of $M_8$, \item [] the vector $(-2,-13,-25,-14,-1,0,0,0,1)$ is in the kernel of $M_9$, \item [] the vector $(-2,-15,-36,-30,-6,0,0,0,0,1)$ is in the kernel of $M_{10}$, \item [] the vector $(-2,-17,-49,-55,-20,-1,0,0,0,0,1)$ is in the kernel of $M_{11}$. \end{enumerate} Okay, now we have to make sense out of this. Our vectors in the kernel have the following structure: first, there are some negative numbers, then follow a few zeroes, and finally there is a trailing 1. I believe that we do not have any problem to guess what the zeroeth\footnote{The indexing convention in the matrix in \eqref{eq:Krat1} of which the determinant is taken is that rows and columns are indexed by $0,1,\dots,n-1$. We keep this convention here.} or the first coordinate of our vector is. Since the second coordinates are always negatives of squares, there is also no problem there. What about the third coordinates? Starting with the vector for $M_7$, these are $-1,-5,-14,-30,-55,\dots$. I guess, rather than thinking hard, we should consult {\tt Rate} (see Footnote~\ref{foot:Rate}): \MATH \goodbreakpoint% In[26]:= Rate[-1,-5,-14,-30,-55] \leavevmode% -(i0 (1 + i0) (1 + 2 i0)) Out[26]= \MATHlbrace -------------------------\MATHrbrace \leavevmode% 6 \goodbreakpoint% \endgroup After replacing {\tt i0} by $n-6$ (as we should), this becomes $-(n-6)(n-5)(2n-11)/6$. An interesting feature of this formula is that it also works well for $n=6$ and $n=5$. Equipped with this experience, we let {\tt Rate} work out the fourth coordinate: \MATH \goodbreakpoint% In[27]:= Rate[0,0,0,-1,-6,-20] \leavevmode% 2 \leavevmode% -((-3 + i0) (-2 + i0) (-1 + i0)) Out[27]= \MATHlbrace ---------------------------------\MATHrbrace \leavevmode% 12 \goodbreakpoint% \endgroup After replacement of {\tt i0} by $n-5$, this is $-(n-8)(n-7)^2(n-6)/12$. Let us summarise our results so far: the first five coordinates of our vector in the kernel of $M_n$ are \begin{multline*} -2,\ -(2n-5),\ -\frac {(n-4)(2n-8)} {2}, \ -\frac {(n-6)(n-5)(2n-11)} {6}, \\ -\frac {(n-8)(n-7)(n-6)(2n-14)} {12}. \end{multline*} I would say, there is a clear pattern emerging: the $s$-th coordinate is equal to $$-\frac {(n-2s)_{s-1}\,(2n-3s-2)} {s!}= -\frac {(2n-3s-2)} {(n-s-1)}\frac {(n-2s)_{s}} {s!}.$$ Denoting the above expression by $f(n,s)$, the vector $$(f(n,0),f(n,1),\dots,f(n,n-2),1)$$ is apparently in the kernel of $M_n$ for $n\ge2$. To prove this, we have to show that \begin{multline*} \sum _{s=0} ^{n-2}\frac {(2n-3s-2)} {(n-s-1)}\frac {(n-2s)_{s}} {(s)!}\\ \cdot(-y-2n+i+2)_{s}\,(-2y-2n+2i-s+3)_{s}\, (y+2s-i+1)_{2n-2s-2}\\ =(-y-2n+i+2)_{n-1}\,(-2y-3n+2i+4)_{n-1}. \end{multline*} In \machSeite{KratBD}\cite{KratBD} it was argued that this identity follows from a certain hypergeometric identity due to Singh \machSeite{SinVAA}\cite{SinVAA}. However, for just having {\it some} proof of this identity, this careful literature search was not necessary. In fact, nowadays, {\it once you write down a binomial or hypergeometric identity, it is already proved!} One simply puts the binomial/hypergeometric sum into the {\it Gosper--Zeilberger algorithm} (see \machSeite{PeWZAA}% \machSeite{ZeilAP}% \machSeite{ZeilAM}% \machSeite{ZeilAV}% \cite{PeWZAA,ZeilAP,ZeilAM,ZeilAV}), which outputs a recurrence for it, and then the only task is to verify that the (conjectured) right-hand side also satisfies the same recurrence, and to check the identity for sufficiently many initial values (which one has already done anyway while producing the conjecture).\footnote{As you may have suspected, this is again a little bit oversimplified. But not much. The Gosper--Zeilberger algorithm applies {\it always} to hypergeometric sums, and there are only very few binomial sums where it does not apply. (For the sake of completeness, I mention that there are also several algorithms available to deal with multi-sums, see \machSeite{ChSaAA}% \machSeite{WiZeAC}% \cite{ChSaAA,WiZeAC}. These do, however, rather quickly challenge the resources of today's computers.) {\em Maple} implementations written by Doron Zeilberger are available from {\tt http://www.math.temple.edu/\~{}zeilberg}, those written by Fr\'ed\'eric Chyzak are available from {\tt http://algo.inria.fr/chyzak/mgfun.html}, {\sl Mathematica} implementations written by Peter Paule, Axel Riese, Markus Schorn, Kurt Wegschaider are available from {\tt http://www.risc.uni-linz.ac.at/research/combinat/risc/software}.} As I mentioned earlier, actually we need more vectors in the kernel. However, this is not difficult. Take a closer look, and you will see that the pattern persists (set $c[n-1]=0$ in the vector for $V[n]$, etc.). It will take you no time to work out a full-fledged conjecture for $\fl{(n+1)/3}$ linear independent vectors in the kernel of $M_n$. I do not want to go through Steps (S2) and (S3), that is, the degree calculation and the computation of the constant. As it turns out, to do this conveniently you need to introduce more variables in the determinant in \eqref{eq:Krat1}. Once you do this, everything works out very smoothly. I refer the reader to \machSeite{KratBD}\cite{KratBD} for these details. \medskip Now, let us come back to our determinant, the determinant of \eqref{eq:Det}, and apply ``identification of factors" to it.% \footnote{What I describe in the sequel is, except for very few excursions that ended up in a dead end, and which are therefore omitted here, the way how the determinant evaluation was found.} To begin with, here is bad news: ``identification of factors" crucially requires the existence of indeterminates. But, where are they in \eqref{eq:Det}? If we look at the definition of the matrix \eqref{eq:Det}, which, in the end, depends on the auxiliary functions $f_0,f_1,g_0,g_1$ defined in \eqref{eq:fg}, then we see that there are no indeterminates at all. Everything is (integral) numbers. So, to get even started, we need to introduce indeterminates in a way such that the more general determinant would still factor ``nicely." We do not have much guidance. Maybe, since we already made the abbreviation $N(k)=4k(4k+1)$, we should replace $N(k)$ by $X$? Okay, let us try this, that is, let us put \begin{align} \notag f_0(j)&=j(-4)^k,\\ \notag f_1(j)&=-(4j+2)(-4)^k,\\ \notag g_0(j)&=(X-j),\\ \label{eq:fg1} g_1(j)&=-(4X-4j-2) \end{align} instead of \eqref{eq:fg}. Let us compute the new determinant for $k=2$. We program the new functions $f_0,f_1,g_0,g_1$, \MATH \goodbreakpoint% In[28]:= f0[k\MATHtief , t\MATHtief , j\MATHtief ] := j*(-4)\MATHhoch k; \leavevmode% f1[k\MATHtief , t\MATHtief , j\MATHtief ] := -(2 + 4*j)*(-4)\MATHhoch k; \leavevmode% g0[k\MATHtief , t\MATHtief , j\MATHtief ] := (X - j); \leavevmode% g1[k\MATHtief , t\MATHtief , j\MATHtief ] := (-4*X + 2 + 4*j) \goodbreakpoint% \endgroup we enter the new determinant for $k=2$, \MATH \goodbreakpoint% In[29]:= Factor[Det[A[2]]] \goodbreakpoint% \endgroup and, after a waiting time of more than 15 minutes,\footnote{\label{foot:kompl}which I use to explain why our computer needs so long to calculate this determinant of a very sparse matrix of size $16\cdot 2^2=64$: isn't it true that, nowadays, determinants of matrices with several hundreds of rows and columns can be calculated without the slightest difficulty (particularly if they are very sparse)? Well, we should not forget that this is true for determinants of matrices with {\it numerical\/} entries. However, our matrix \eqref{eq:Det} with the modified definitions \eqref{eq:fg1} of $f_0,f_1,g_0,g_1$ has now entries which are {\it polynomials} in $X$. Hence, when our computer algebra program applies (internally) some elimination algorithm to compute the determinant, huge rational expressions will slowly build up and will slow down the calculations (and, at times, will make our computer crash \dots). As I learn from Dave Saunders, {\sl Maple} and {\sl Mathematica} do currently in fact not use the best known algorithms for dealing with determinants of matrices with polynomial entries. (This may have to do with the fact that the developers try to optimise the algorithms for numerical determinants in the first case.) It is known how to avoid the expression swell and compute polynomial matrix determinants in time about $mn^3$, where $n$ is the dimension of the matrix and $m$ is the bit length of the determinant (roughly, in univariate case, $m$ is degree times maximum coefficient length).} we obtain \MATH \goodbreakpoint% Out[29]= -1406399608474882323154910525986578515918369681041517636\MATHbackslash \leavevmode% 2 > 11783762359972003840000000 (-64 + X) (-48 + X) (-40 + X) \leavevmode% 3 4 5 6 7 > (-32 + X) (-24 + X) (-16 + X) (-8 + X) X \leavevmode% 2 > (9653078694297600 - 916000657637376 X + 36130368757760 X \leavevmode% 3 4 5 6 > - 758218948608 X + 8928558848 X - 55938432 X + 145673 X ) \goodbreakpoint% \endgroup Not bad. There are many factors which are linear in $X$. (This is what we were after.) However, the irreducible polynomial of degree 6 gives us some headache. (The degrees of the irreducible part of the polynomial will grow quickly with $k$.) How are we going to guess what this factor could be, and, even more daunting, even if we should be able to come up with a guess, how would we go about to prove it? So, maybe we should modify our choice of how to introduce indeterminates into the matrix. In fact, we overlooked something: maybe, in a hidden manner, the variable $X$ is also there at other places in \eqref{eq:fg}, that is, when $X$ is specialised to $N(k)=4k(4k+1)$ at these places it becomes invisible. More specifically, maybe we should insert the difference $X-4k(4k+1)$ in the definitions of $f_0$ and $f_1$ (which would disappear for $X=4k(4k+1)$). So, maybe we should try: \begin{align*} \notag f_0(j)&=(4k(4k + 1) - X + j)(-4)^k,\\ \notag f_1(j)&=-(16k(4k + 1) - 4X + 2 + 4j)(-4)^k,\\ \notag g_0(j)&=(X-j),\\ g_1(j)&=-(4X-4j-2), \end{align*} Okay, let us modify our computer program accordingly, \MATH \goodbreakpoint% In[30]:= f0[k\MATHtief , t\MATHtief , j\MATHtief ] := (4*k*(4*k + 1) - X + j)*(-4)\MATHhoch k; \leavevmode% f1[k\MATHtief , t\MATHtief , j\MATHtief ] := -(4*4*k*(4*k + 1) - 4*X + 2 + 4*j)*(-4)\MATHhoch k; \leavevmode% g0[k\MATHtief , t\MATHtief , j\MATHtief ] := (X - j); \leavevmode% g1[k\MATHtief , t\MATHtief , j\MATHtief ] := (-4*X + 2 + 4*j) \goodbreakpoint% \endgroup and let us compute the new determinant for $k=2$: \MATH \goodbreakpoint% In[31]:= Factor[Det[A[2]]] \goodbreakpoint% \endgroup This makes us wait for another 15 minutes, after which we are rewarded with: \MATH \goodbreakpoint% Out[31]= -296777975397624679901369809794412104454134763494070841\MATHbackslash > 1155365196124754770317472271790417634937439881166252558632\MATHbackslash > 616674197504000000000 (-141 + 2 X) (-139 + 2 X) (-137 + 2 X) > (-135 + 2 X) (-133 + 2 X) (-131 + 2 X) (-129 + 2 X) \goodbreakpoint% \endgroup Excellent! There is no big irreducible polynomial anymore. Everything is linear factors in $X$. But, wait, there is still a problem: in the end (recall Step~(S2)!) we will have to compare the degrees of the determinant and of the right-hand side as polynomials in $X$. If we expand the determinant according to its definition, then the conclusion is that the degree of the determinant is bounded above by $16k^2-1$, which, for $k=2$ is equal to $31$. The right-hand side polynomial however which we computed above has degree 7. This is a big gap! I skip some other things (ending up in dead ends \dots) that we tried at this point. Altogether they pointed to the fact that, apparently, {\it one} indeterminate is not sufficient. Perhaps it is a good idea to ``diversify" the variable $X$, that is, to make two variables, $X_1$ and $X_2$, out of $X$: \begin{align*} \notag f_0(j)&=(4k(4k + 1) - X_2 + j)(-4)^k,\\ \notag f_1(j)&=-(16k(4k + 1) - 4X_1 + 2 + 4j)(-4)^k,\\ \notag g_0(j)&=(X_2-j),\\ g_1(j)&=-(4X_1-4j-2). \end{align*} We program this, \MATH \goodbreakpoint% In[32]:= f0[k\MATHtief , t\MATHtief , j\MATHtief ] := (4*k*(4*k + 1) - X[2] + j)*(-4)\MATHhoch k; \leavevmode% f1[k\MATHtief , t\MATHtief , j\MATHtief ] := -(4*4*k*(4*k + 1) - 4*X[1] + 2 + 4*j)*(-4)\MATHhoch k; \leavevmode% g0[k\MATHtief , t\MATHtief , j\MATHtief ] := (X[2] - j); \leavevmode% g1[k\MATHtief , t\MATHtief , j\MATHtief ] := (-4*X[1] + 2 + 4*j) \goodbreakpoint% \endgroup and, in order to avoid overstraining our computer, compute this time the new determinant for $k=1$: \MATH \goodbreakpoint% In[33]:= Factor[Det[A[1]]] \goodbreakpoint% \endgroup After some minutes there appears \MATH \goodbreakpoint% Out[33]= 3242591731706757120000 (-37 + 2 X[1]) (-35 + 2 X[1]) \leavevmode% 3 2 > (-33 + 2 X[1]) (1 + 2 X[1] - 2 X[2]) (3 + 2 X[1] - 2 X[2]) > (5 + 2 X[1] - 2 X[2]) \goodbreakpoint% \endgroup on the computer screen. On the positive side: the determinant still factors completely into linear factors, something which we could not expect a priori. Moreover, the degree (in $X_1$ and $X_2$) has increased, it is now equal to 9 although we were only computing the determinant for $k=1$. However, a gap remains, the degree should be $16k^2-1=15$ if $k=1$. Thus, it may be wise to introduce another genuine variable, $Y$. For example, we may think of simply homogenising the definitions of $f_0,f_1,g_0,g_1$: \begin{align*} \notag f_0(j)&=(4k(4k + 1)Y - X_2 + jY)(-4)^k,\\ \notag f_1(j)&=-(16k(4k + 1)Y - 4X_1 + (2 + 4j)Y)(-4)^k,\\ \notag g_0(j)&=(X_2-jY),\\ g_1(j)&=-(4X_1-(4j+2)Y). \end{align*} We program this, \MATH \goodbreakpoint% In[34]:= f0[k\MATHtief , t\MATHtief , j\MATHtief ] := (4*k*(4*k + 1)*Y - X[2] + j*Y)*(-4)\MATHhoch k; \leavevmode% f1[k\MATHtief , t\MATHtief , j\MATHtief ] := -(4*4*k*(4*k + 1)*Y - 4*X[1] + \leavevmode% (2 + 4*j)*Y)*(-4)\MATHhoch k; \leavevmode% g0[k\MATHtief , t\MATHtief , j\MATHtief ] := (X[2] - j*Y); \leavevmode% g1[k\MATHtief , t\MATHtief , j\MATHtief ] := (-4*X[1] + (2 + 4*j)*Y) \goodbreakpoint% In[35]:= Factor[Det[A[1]]] \goodbreakpoint% \endgroup we wait for some more minutes, and we obtain \MATH \leavevmode% 6 Out[35]= -3242591731706757120000 Y (33 Y - 2 X[1]) (35 Y - 2 X[1]) \leavevmode% 3 2 > (37 Y - 2 X[1]) (Y + 2 X[1] - 2 X[2]) (3 Y + 2 X[1] - 2 X[2]) > (5 Y + 2 X[1] - 2 X[2]) \goodbreakpoint% \endgroup Great! The degree in $X_1,X_2,Y$ is 15, as it should be! At this point, one becomes greedy. The more variables we have, the easier will be the proof. We ``diversify" the variables $X_1,X_2,Y$, that is, we make them $X_{1,t},X_{2,t},Y_t$ if they appear in the blocks $F_t$ or $G_t$, respectively, $t=1,2,\dots,4k$ (cf.\ \eqref{eq:Det} and the {\sl Mathematica} code for the precise meaning of this definition): \begin{align} \notag f_0(j)&=(4k(4k + 1)Y_t - X_{2,t} + jY_t)(-4)^k,\\ \notag f_1(j)&=-(16k(4k + 1)Y_t - 4X_{1,t} + (2 + 4j)Y_t)(-4)^k,\\ \notag g_0(j)&=(X_{2,t}-jY_t),\\ \label{eq:fg5} g_1(j)&=-(4X_{1,t}-(4j+2)Y_t). \end{align} Now there are so many variables so that there is no way to do the factorisation of the new determinant for $k=1$ on the computer unless one plays tricks (which we\break did).% \footnote{\label{foot:tricks}See Footnote~\ref{foot:kompl} for the explanation of the complexity problem. ``Playing tricks" would mean to compute the determinant for various special choices of the variables $X_{1,t},X_{2,t},Y_t$, and then reconstruct the general result by interpolation. This is possible because we know an a priori degree bound (namely $15$) for the polynomial. However, this would become infeasible for $k=3$, for example. ``Playing tricks" then would mean to be content with an ``almost sure" guess, the latter being based on features of the (unknown) general result that are already visible in the earlier results, and on calculations done for special values of the variables. For example, if we encounter determinants $\det M(k)$, where the $M(k)$'s are some square matrices, $k=1,2,\dots$, and the results for $k=1,2,\dots,k_0-1$ show that $x-y$ must be a factor of $\det M(k)$ to some power, then one would specialise $y$ to some value that would make $x-y$ distinct from any other linear factors containing $x$, and, supposing that $y=17$ is such a choice, compute $\det M(k_0)$ with $y=17$. The exact power of $x-y$ in the unspecialised determinant $\det M(k_0)$ can then be read off from the exponent of $x-17$ in the specialised one. If it should happen that it is also infeasible to calculate $\det M(k_0)$ with $x$ still unspecialised, then there is still a way out. In that case, one specialises $y$ {\it and\/} $x$, in such a way that $x-y$ would be a prime $p$ that one expects not to occur as a prime factor in any other factor of the determinant $\det M(k_0)$. The exact power of $x-y$ in the unspecialised determinant $\det M(k_0)$ can then be read off from the exponent of $p$ in the prime factorisation of the specialised determinant. See Subsections~\ref{sec:signed} and \ref{sec:poset}, and in particular Footnote~\ref{foot:maj} for further instances where this trick was applied.} But let us pretend that we are able to do it: \MATH \goodbreakpoint% In[36]:= f0[k\MATHtief , t\MATHtief , j\MATHtief ] := (4*k*(4*k + 1)*Y[t] - X[2, t] \leavevmode% + j*Y[t])*(-4)\MATHhoch k; \leavevmode% f1[k\MATHtief , t\MATHtief , j\MATHtief ] := -(4*4*k (4*k + 1)*Y[t] - 4*X[1, t] + \leavevmode% (2 + 4*j)*Y[t])*(-4)\MATHhoch k; \leavevmode% g0[k\MATHtief , t\MATHtief , j\MATHtief ] := (X[2, t] - j*Y[t]); \leavevmode% g1[k\MATHtief , t\MATHtief , j\MATHtief ] := (-4*X[1, t] + (2 + 4*j)*Y[t]) \goodbreakpoint% In[37]:= Factor[Det[A[1]]] \goodbreakpoint% Out[37]= 3242591731706757120000 (2 X[1, 1] - 33 Y[1]) Y[1] > (2 X[1, 1] - 2 X[2, 1] + Y[1]) (2 X[1, 2] - 35 Y[2]) Y[2] > (2 X[1, 2] - 2 X[2, 2] + Y[2]) (-2 X[2, 2] Y[1] + 2 X[1, 1] > Y[2] + 3 Y[1] Y[2]) (2 X[1, 3] - 37 Y[3]) Y[3] (2 X[1, 3] - > 2 X[2, 3] + Y[3]) (-2 X[2, 3] Y[1] + 2 X[1, 1] Y[3] + > 5 Y[1] Y[3]) (-2 X[2, 3] Y[2] + 2 X[1, 2] Y[3] + 3 Y[2] Y[3]) \goodbreakpoint% \endgroup By staring a little bit at this result (and the one that we computed for $k=2$), we extracted that, apparently, we have \begin{multline} \label{eq:CK4} \det A^X=(-1)^{k-1}4^{2k(4k^2+7k+2)}k^{2k(4k+1)} \prod _{i=1} ^{4k}(i+1)_{4k-i+1}\\ \times \prod _{a=1} ^{4k-1}\left(2X_{1,a}-(32k^2+2a-1)Y_a\right)\\ \times \prod _{1\le a\le b\le 4k-1} ^{} (2X_{2,b}Y_a-2X_{1,a}Y_b-(2b-2a+1)Y_aY_b), \end{multline} where $A^X$ denotes the new general matrix given through \eqref{eq:Det} and \eqref{eq:fg5}, and where, as before, $(\alpha)_k$ is the standard notation for shifted factorials (Pochhammer symbols) explained in the statement of Theorem~\ref{thm:xy}. The special case that we need in the end to prove our Theorem~\ref{T1} is $X_{1,t}=X_{2,t}=N(k)$ and $Y_t=1$. \medskip Now we are in business. Here is the {\it Sketch of the proof of \eqref{eq:CK4}}: \medskip Re (S1): For each factor of the (conjectured) result \eqref{eq:CK4}, we find a linear combination of the rows which vanishes if the factor vanishes. (In other terms: if the indeterminates in the matrix are specialised so that a particular factor vanishes, we find a vector in the kernel of the transpose of the specialised matrix.) For example, to explain the factor $(2X_{1,1}-(32k^2+1)Y_1)$, we found: If $X_{1,1}=\frac {32k^2+1} {2}Y_1$, then \begin{multline} \label{eq:combin} \frac {2(X_{2,4k-1}-(N(k)-1)Y_{4k-1})} {(-4)^{k(4k+1)+1}(16k^2+1)\prod _{\ell=1} ^{4k-1}(4\ell k+1)}\cdot(\text {row 0 of $A^X$}) \\+ \sum _{s =0} ^{4k}\sum _{t =0} ^{4k-2}\Bigg(\frac {(-1)^{s (k-1)}2^t} {4^{s k}} \prod _{\ell=0} ^{s -1}\frac {4k-1+4\ell k} {16k^2+1-4\ell k} \prod _{\ell=4k-t } ^{4k-1}\frac {2X_{1,\ell}-(32k^2+2\ell-1)Y_\ell} {X_{2,\ell-1}-(16k^2+\ell-1)Y_{\ell-1}}\Bigg)\\ \cdot(\text {row $(16k^2-(4k-1)s -t -1)$ of $A^X$})=0, \end{multline} as is easy to verify. (Since the coefficients of the various rows in \eqref{eq:combin} are rational functions in the indeterminates $X_{1,t},X_{2,t},Y_t$, they are rather easy to work out from computer data. One does not even need {\tt Rate} \dots) \medskip Re (S2): The total degree in the $X_{1,t}$'s, $X_{2,t}$'s, $Y_{t}$'s of the product on the right-hand side of \eqref{eq:CK4} is $16k^2-1$. As we already remarked earlier, the degree of the determinant is at most $16k^2-1$. Hence, the determinant is equal to the product times, possibly, a constant. \medskip Re (S3): For the evaluation of the constant, we compare coefficients of $$ X_{1,1}^{4k}X_{1,2}^{4k-1}\cdots X_{1,4k-1}^2Y_1^1Y_2^2\cdots Y_{4k-1}^{4k-1}. $$ After some reflection, it turns out that the constant is equal to a determinant of the same form, that is, of the form \eqref{eq:Det}, but with auxiliary functions \begin{align} \notag f_0(j)&=(N(k)+j)(-4)^k,\\ \notag f_1(j)&=4(-4)^k,\\ \notag g_0(j)&=-j,\\ \label{eq:fg6} g_1(j)&=-4. \end{align} \medskip What a set-back! It seems that we are in the same situation as at the very beginning. We started with the determinant of the matrix \eqref{eq:Det} with auxiliary functions \eqref{eq:fg}, and we ended up with the same type of determinant, with auxiliary functions \eqref{eq:fg6}. There is little hope though: the functions in \eqref{eq:fg6} are somewhat simpler as those in \eqref{eq:fg}. Nevertheless, we have to play the same game again; that is, if we want to apply the method of identification of factors, then we have to introduce indeterminates. Skipping the experimental part, we came up with \begin{align*} f_0(j)&=(Z_t+j)(-4)^k,\\ f_1(j)&=4(-4)^kX_t,\\ g_0(j)&=-j,\\ g_1(j)&=-4X_t, \end{align*} where $t$ has the same meaning as before in \eqref{eq:fg5}. Denoting the new matrix by $A^Z$, computer calculations suggested that apparently \begin{equation} \label{eq:CK9} \det A^Z=(-1)^{k-1}2^{16k^3+20k^2+14k-1}k^{4k}(4k+1)! \prod _{a=1} ^{4k-1}\Bigg(X_a^{4k+1-a} \prod _{b=0} ^{a-1}(Z_a-4bk)\Bigg). \end{equation} The special case that we need is $Z_t=N(k)$ and $X_t=1$. So, we apply again the method of identification of factors. Everything runs smoothly (except that the details of the verification of the factors are somewhat more unpleasant here). When we come finally to the point that we want to determine the constant, it turns out that the constant is equal to --- no surprise anymore --- the determinant of a matrix of the form \eqref{eq:Det} with auxiliary functions \begin{align*} f_0(j)&=(-4)^k,\\ f_1(j)&=4(-4)^k,\\ g_0(j)&=0,\\ g_1(j)&=-4. \end{align*} Now, is this good or bad news? In other words, while painfully working through the steps of ``identification of factors," will we forever continue producing new determinants of the form \eqref{eq:Det}, which we must again handle by the same method? To give it away: this is indeed {\it very good} news. The function $g_0(j)$ vanishes identically! It makes it possible that now Method~0 (= do some row and column manipulations) works. (See \machSeite{AlKPAA}\cite{AlKPAA} for the details.) We are --- finally --- done with the proof of \eqref{eq:CK4}, and, since the right-hand side {\it does not\/} vanish for $X_{1,t}=X_{2,t}=N(k)$ and $Y_t=1$, with the proof of Theorem~\ref{T1}! \raise-15pt\hbox{{\Huge$\square$}} \section{More determinant evaluations}\label{sec:detlist} This section complements the list of known determinant evaluations given in Section~3 of \machSeite{KratBN}\cite{KratBN}. I list here several determinant evaluations which I believe are interesting or attractive (and, in the ideal case, both), that have appeared since \machSeite{KratBN}\cite{KratBN}, or that I failed to mention in \machSeite{KratBN}\cite{KratBN}. I also include several conjectures and open problems, some of them old, some of them new. As in \machSeite{KratBN}\cite{KratBN}, each evaluation is accompanied by some remarks providing information on the context in which it arose. Again, the selection of determinant evaluations presented reflects totally my taste, which must be blamed in the case of any shortcomings. The order of presentation follows loosely the order of presentation of determinants in \machSeite{KratBN}\cite{KratBN}. \subsection{More basic determinant evaluations} I begin with two determinant evaluations belonging to the category ``standard determinants" (see Section~2.1 in \machSeite{KratBN}\cite{KratBN}). They are among those which I missed to state in \machSeite{KratBN}\cite{KratBN}. The reminder for inclusion here is the paper \machSeite{AmZeAB}\cite{AmZeAB}. There, Amdeberhan and Zeilberger propose an {\it automated approach} towards determinant evaluations via the condensation method (see ``Method~2" in Section~\ref{sec:eval}). They provide a list of examples which can be obtained in that way. As they remark at the end of the paper, all of these are special cases of Lemma~5 in \machSeite{KratBN}\cite{KratBN}, with the exception of three, namely Eqs.~(8)--(10) in \machSeite{AmZeAB}\cite{AmZeAB}. In their turn, two of them, namely (8) and (9), are special cases of the following evaluation. (For (10), see Lemma~\ref{prop:AmZe} below.) \begin{Lemma} \label{lem:AmZe} Let $P(Z)$ be a polynomial in $Z$ of degree $n-1$ with leading coefficient $L$. Then \begin{equation} \label{eq:D1} \det_{1\le i,j\le n}\(P(X_i+Y_j)\)=L^n \prod _{i=1} ^{n}\binom {n-1}i\prod _{1\le i<j\le n} ^{}(X_i-X_j)(Y_j-Y_i). \end{equation} \quad \quad \qed \end{Lemma} This lemma is easily proved along the lines of the standard proof of the Vandermonde determinant evaluation which we recalled in Section~\ref{sec:eval} (see the proof of \eqref{eq:Vandermonde}) or by condensation. A multiplicative version of Lemma~\ref{lem:AmZe} is the following. \begin{Lemma} \label{lem:AmZe2} Let $P(Z)=p_{n-1}Z^{n-1}+p_{n-2}Z^{n-2}+\dots+p_0$. Then \begin{equation} \label{eq:D2} \det_{1\le i,j\le n}\(P(X_iY_j)\)= \prod _{i=0} ^{n-1}p_i\prod _{1\le i<j\le n} ^{}(X_i-X_j)(Y_i-Y_j). \end{equation} \quad \quad \qed \end{Lemma} On the other hand, identity~(10) from \machSeite{AmZeAB}\cite{AmZeAB} can be generalised to the following Cauchy-type determinant evaluation. As all the identities from \machSeite{AmZeAB}\cite{AmZeAB}, it can also be proved by the condensation method. \begin{Lemma} \label{prop:AmZe} Let $a_0,a_1,\dots,a_{n-1}$, $c_0,c_1,\dots,c_{n-1}$, $b$, $x$ and $y$ be indeterminates. Then, for any positive integer $n$, there holds \begin{multline} \label{eq:AmZe} \det_{0\le i,j\le n-1}\(\frac{(x+a_i+c_j)(y+bi+c_j)} { (x+a_i+bi+c_j)}\)\\=b^{n-1}\,(n-1)! \(\binom n2b+(n-1)x+y+\sum_{i=1}^{n-1}a_i + \sum_{i=0}^{n-1}c_i\)\\ \times \frac{\displaystyle\prod_{0\le i<j\le n-1}^{}(c_j-c_i)\ \prod_{i=1}^{n-1}(y-x-a_i) \prod_{1\le i<j\le n-1}^{}((j-i)b-a_i+a_j)} {\displaystyle\prod_{i=1}^{n-1}\prod_{j=0}^{n-1} {(x+a_i+bi+c_j)}}. \end{multline} \quad \quad \qed \end{Lemma} Speaking of Cauchy-type determinant evaluations, this brings us to a whole family of such evaluations which were instrumental in Kuperberg's recent advance \machSeite{KupeAH}\cite{KupeAH} on the enumeration of (symmetry classes of) {\it alternating sign matrices}. The reason that determinants, and also Pfaffians, play an important role in this context is due to Propp's discovery (described for the first time in \machSeite{ElKLAB}\cite[Sec.~7]{ElKLAB} and exploited in \machSeite{KupeAD}% \machSeite{KupeAH}\cite{KupeAD,KupeAH}) that alternating sign matrices are in bijection with {\it configurations in the six vertex model}, and due to determinant and Pfaffian formulae due to Izergin \machSeite{IzerAA}\cite{IzerAA} and Kuperberg \machSeite{KupeAH}\cite{KupeAH} for certain multivariable partition functions of the six vertex model under various boundary conditions. In many cases, this leads to determinants which are, or are similar to, {\it Cauchy's evaluation of the double alternant\/} (see \machSeite{MuirAB}% \cite[vol.~III, p.~311]{MuirAB} and \eqref{eq:Cauchy} below) or {\it Schur's Pfaffian version} \machSeite{SchuAA}\cite[pp.~226/227]{SchuAA} of it (see \eqref{eq:Schur} below). Let me recall that the {\it Pfaffian} $\operatorname{Pf}(A)$ of a skew-symmetric $(2n)\times(2n)$ matrix $A$ is defined by \begin{equation} \label{eq:Pfaff} \operatorname{Pf}(A)=\sum _{\pi} ^{}(-1)^{c(\pi)}\prod _{(ij)\in \pi} ^{}A_{ij}, \end{equation} where the sum is over all perfect matchings $\pi$ of the complete graph on $2n$ vertices, where $c(\pi)$ is the {\em crossing number} of $\pi$, and where the product is over all edges $(ij)$, $i<j$, in the matching $\pi$ (see e.g.\ \machSeite{StemAE}\cite[Sec.~2]{StemAE}). What links Pfaffians so closely to determinants is (aside from similarity of definitions) the fact that the Pfaffian of a skew-symmetric matrix is, up to sign, the square root of its determinant. That is, $\det(A)=\operatorname{Pf}(A)^2$ for any skew-symmetric $(2n)\times(2n)$ matrix $A$ (cf.\ \machSeite{StemAE}\cite[Prop.~2.2]{StemAE}). See the corresponding remarks and additional references in \machSeite{KratBN}\cite[Sec.~2.8]{KratBN}. The following three theorems present the relevant evaluations. They are Theorems~15--17 from \machSeite{KupeAH}\cite{KupeAH}. All of them are proved using identification of factors (see ``Method~3" in Section~\ref{sec:eval}). The results in Theorem~\ref{thm:Kup1} contain whole sets of indeterminates, whereas the results in Theorems~\ref{thm:Kup2} and \ref{thm:Kup3} only have two indeterminates $p$ and $q$, respectively three indeterminates $p$, $q$ and $r$, in them. Identity \eqref{eq:Kup4} is originally due to Laksov, Lascoux and Thorup \machSeite{LaLTAA}\cite{LaLTAA} and Stembridge \machSeite{StemAE}\cite{StemAE}, independently. The reader must be warned that the statements in \machSeite{KupeAH}\cite[Theorems~15--17]{KupeAH} are often blurred by typos. \begin{Theorem} \label{thm:Kup1} Let $x_1,x_2,\dots$ and $y_1,y_2,\dots$ be indeterminates. Then, for any positive integer $n$, there hold \begin{equation} \label{eq:Cauchy} \det_{1\leq i,j\leq n}\left( \frac{1}{x_i+y_j} \right) = \frac{\displaystyle \prod_{1\leq i<j\leq n}(x_i-x_j)(y_i-y_j) }{\displaystyle \prod_{1\leq i, j\leq n}(x_i+y_j) }, \end{equation} \begin{multline} \label{eq:Kup2} \det_{1\le i,j\le n}\(\frac1{x_i+y_j} - \frac1{1+x_iy_j}\) = \frac{\displaystyle\prod_{1\le i<j\le n} (1-x_ix_j)(1-y_iy_j)(x_j-x_i)(y_j-y_i)} {\displaystyle\prod_{1\le i,j\le n} (x_i+y_j)(1+x_iy_j)}\\ \times \prod_{i=1}^n (1-x_i)(1-y_i) , \end{multline} \begin{equation} \label{eq:Schur} \underset{1\le i,j\le 2n}\operatorname{Pf}\(\frac {x_i-x_j} {x_i+x_j}\)= \prod _{1\le i<j\le 2n} ^{}\frac {x_i-x_j} {x_i+x_j}. \end{equation} \begin{equation} \label{eq:Kup4} \underset{1\le i,j\le 2n}\operatorname{Pf}\(\frac{x_i-x_j}{1-x_ix_j}\) = \prod_{1\le i<j\le 2n} \frac{x_i-x_j}{1-x_ix_j}. \end{equation} \quad \quad \qed \end{Theorem} \begin{Theorem} \label{thm:Kup2} Let $p$ and $q$ be indeterminates. Then, for any positive integer $n$, there hold \begin{equation} \label{eq:Kup5} \det_{1\le i,j\le n}\(\frac{q^{n+j-i}-q^{-(n+j-i)}} {p^{n+j-i}-p^{-(n+j-i)}}\) = \frac{\displaystyle\prod_{1\le i \ne j\le n} (p^{j-i}-p^{-(j-i)}) \prod_{1\le i,j\le n} (qp^{j-i}-q^{-1}p^{-(j-i)})} {\displaystyle\prod_{1\le i,j\le n} (p^{n+j-i}-p^{-(n+j-i)})}, \end{equation} \begin{equation} \label{eq:Kup6} \det_{1\le i,j\le n}\(\frac{q^{j-i}+q^{-(j-i)}}{p^{j-i}+p^{-(j-i)}}\) = (-1)^{\binom{n}2} \frac{\displaystyle 2^n\prod_{\substack{1\le i \ne j\le n \\ 2\mid j-i}}(p^{j-i}-p^{-j+i}) \prod_{\substack{1\le i,j\le n \\ 2\nmid j-i}}(qp^{j-i}-q^{-1}p^{-j+i})} {\displaystyle\prod_{1\le i,j\le n} (p^{j-i}+p^{-j+i})} , \end{equation} \begin{multline} \label{eq:Kup7} \det_{1\le i,i\le n}\(\frac{q^{n+j+i}-q^{-(n+j+i)}}{p^{n+j+i}-p^{-(n+j+i)}} - \frac{q^{n+j-i}-q^{-(n+j-i)}}{p^{n+j-i}-p^{-(n+j-i)}}\) \\= \frac{\prod_{1\le i<j\le 2n}(p^{j-i}-p^{-(j-i)}) \prod_{\displaystyle\substack{1\le i,j \le 2n+1 \\ 2|j}} (qp^{j-i}-q^{-1}p^{-(j-i)})} {\displaystyle\prod_{1\le i,j\le n} (p^{n+j-i}-p^{-(n+j-i)}) (p^{n+j+i}-p^{-(n+j+i)})}, \end{multline} \begin{multline} \label{eq:Kup8} \det_{1\le i,j\le n}\(\frac{q^{j+i}+q^{-(j+i)}}{p^{j+i}+p^{-(j+i)}} - \frac{\displaystyle q^{j-i}+q^{-j+i}}{p^{j-i}+p^{-(j-i)}}\)\\ =(-1)^{\binom n2} \frac{\displaystyle 2^n\prod_{1\le i<j \le n} (p^{2(j-i)}-p^{-2(j-i)})^2 \prod_{\substack{1\le i,j \le 2n+1 \\ 2\nmid i,\,2|j}} (qp^{j-i}-q^{-1}p^{-(j-i)})} {\displaystyle\prod_{1\le i,j\le n} (p^{j-i}+p^{-(j-i)})(p^{j+i}+p^{-(j+i)})}. \end{multline} \quad \quad \qed \end{Theorem} \begin{Theorem} \label{thm:Kup3} Let $p$, $q$, and $r$ be indeterminates. Then, for any positive integer $n$, there hold \begin{multline} \label{eq:Kup9} \underset{1\le i,j\le 2n}\operatorname{Pf} \(\frac{(q^{j-i}-q^{-(j-i)})(r^{j-i}-r^{-(j-i)})} {(p^{j-i}-p^{-(j-i)})}\) \\ = \frac{\displaystyle\prod_{1\le i<j\le n} (p^{j-i}-p^{-(j-i)})^2 \prod_{1\le i,j\le n} (qp^{j-i}-q^{-1}p^{-(j-i)})(rp^{j-i}-r^{-1}p^{-(j-i)})} {\displaystyle\prod_{1\le i,j \le n} (p^{n+j-i}-p^{-(n+j-i)})} \end{multline} \begin{multline} \label{eq:Kup10} \underset{1\le i,j\le 2n}\operatorname{Pf} \Bigg((p^{j+i}-p^{-(j+i)})(p^{j-i}-p^{-(j-i)}) \biggl(\frac{q^{j+i}-q^{-(j+i)}}{p^{j+i}-p^{-(j+i)}} - \frac{q^{j-i}-q^{-(j-i)}}{p^{j-i}-p^{-(j-i)}}\biggr)\\ \cdot \biggl(\frac{r^{j+i}-r^{-(j+i)}} {p^{j+i}-p^{-(j+i)}} - \frac{r^{j-i}-r^{-(j-i)}}{p^{j-i}-p^{-(j-i)}}\biggr)\Bigg) \\ = \frac{\displaystyle\prod_{1\le i<j \le 2n}(p^{j-i}-p^{-(j-i)}) \prod_{\substack{1\le i,j \le 2n+1 \\ 2|j}} (qp^{j-i}-q^{-1}p^{-(j-i)})(rp^{j-i}-r^{-1}p^{-(j-i)})} {\displaystyle\prod_{1\le i<j \le 2n}(p^{j+i}-p^{-(j+i)})}. \end{multline} \quad \quad \qed \end{Theorem} Subsequent to Kuperberg's work, Okada \machSeite{OkadAJ}\cite{OkadAJ} related Kuperberg's determinants and Pfaffians to characters of classical groups, by coming up with rather complex, but still beautiful determinant identities. In particular, this allowed him to settle one more of the conjectured enumeration formulae on symmetry classes of alternating sign matrices. Generalising even further, Ishikawa, Okada, Tagawa and Zeng \machSeite{IsOTAA}\cite{IsOTAA} have found more such determinant identities. Putting them into the framework of certain special representations of the symmetric group, Lascoux \machSeite{LascAT}\cite{LascAT} has clarified the mechanism which gives rise to these identities. \medskip The next six determinant lemmas are corollaries of {\it elliptic determinant evaluations} due to Rosengren and Schlosser \machSeite{RoScAC}\cite{RoScAC}. (The latter will be addressed later in Subsection~\ref{sec:ell}.) They partly extend the fundamental determinant lemmas in \machSeite{KratBN}\cite[Sec.~2.2]{KratBN}. For the statements of the lemmas, we need the notion of a {\it norm} of a polynomial $a_0+a_1z+\dots+a_kz^k$, which we define to be the reciprocal of the product of its roots, or, more explicitly, as $(-1)^ka_k/a_0$. If we specialise $p=0$ in Lemma~\ref{wp}, \eqref{awpi}, then we obtain a determinant identity which generalises at the same time the Vandermonde determinant evaluation, Lemma~\ref{lem:AmZe} and Lemma~\ref{lem:AmZe2}. \begin{Lemma} \label{lem:Van1} Let $P_1,P_2,\dots,P_n$ be polynomials of degree $n$ and norm $t$, given by $$P_j(x)=(-1)^nta_{j,0}x^n+ \sum _{k=0} ^{n-1}a_{j,k}x^k.$$ Then \begin{equation} \label{eq:Van1} \det_{1\leq i,j\leq n}\left(P_j(x_i)\right)=(1-tx_1\dotsm x_n) \bigg(\prod_{1\leq i<j\leq n}(x_j-x_i)\bigg) \underset{0\le j\le n-1}{\det_{1\le i\le n}}(a_{i,j}). \end{equation} \quad \quad \qed \end{Lemma} Further determinant identities which generalise other {\it Weyl denominator formulae} (cf.\ \machSeite{KratBN}\cite[Lemma~2]{KratBN}) could be obtained from the special case $p=0$ of the other determinant evaluations in Lemma~\ref{wp}. A generalisation of Lemma~6 from \machSeite{KratBN}\cite{KratBN} in the same spirit can be obtained by setting $p=0$ in Theorem~\ref{adet}. It is given as Corollary~5.1 in \machSeite{RoScAC}\cite{RoScAC}. \begin{Lemma} \label{lem:RS1} Let $x_1,\dots,x_n$, $a_1,\dots,a_n$, and $t$ be indeterminates. For each $j=1,\dots,n$, let $P_j$ be a polynomial of degree $j$ and norm $ta_1\dotsm a_j$. Then there holds \begin{multline} \label{eq:RS1} \det_{1\le i,j\le n}\left(P_j(x_i) \prod_{k=j+1}^n(1-a_kx_i)\right)\\ =\frac{1-ta_1\dotsm a_nx_1\dotsm x_n}{1-t} \prod_{i=1}^nP_i(1/a_i) \prod_{1\le i<j\le n}a_j(x_j-x_i). \end{multline} \quad \quad \qed \end{Lemma} We continue with a consequence of Theorem~\ref{adetcor} (see Corollary~5.3 in \machSeite{RoScAC}\cite{RoScAC}). The special case $P_{j-1}(x)=1$, $j=1,\dots,n$, is Lemma~A.1 of \machSeite{SchlAB}\cite{SchlAB}, which was needed in order to obtain an {\it $A_n$ matrix inversion} that played a crucial role in the derivation of {\it multiple basic hypergeometric series identities}. A slight generalisation was given in \machSeite{SchlAF}\cite[Lemma~A.1]{SchlAF}. \begin{Lemma}\label{adetcorr} Let $x_1,\dots,x_n$ and $b$ be indeterminates. For each $j=1,\dots,n$, let $P_{j-1}(x)$ be a polynomial in $x$ of degree at most $j-1$ with constant term $1$, and let $Q(x)=(1-y_1x)\dotsm (1-y_{n+1}x)$. Then there holds \begin{multline}\label{adetcorrid} Q(b)\;\det_{1\le i,j\le n}\left(x_i^{n+1-j}P_{j-1}(x_i) -b^{n+1-j}P_{j-1}(b)\frac{Q(x_i)}{Q(b)}\right)\\ =(1-bx_1\cdots x_ny_1\dotsm y_{n+1}) \prod_{i=1}^n(x_i-b)\prod_{1\le i<j\le n}(x_i-x_j). \end{multline} \quad \quad \qed \end{Lemma} Pairing the $(i,j)$-entry in the determinant in \eqref{eq:RS1} with itself, but with $x_i$ replaced by $1/x_i$, one can construct another determinant which evaluates in closed form. The result given below is Corollary~5.5 in \machSeite{RoScAC}\cite{RoScAC}. It is the special case $p=0$ of Theorem~\ref{cdet}. \begin{Lemma}\label{cdetr} Let $x_1,\dots,x_n$, $a_1,\dots,a_n$, and $c_1,\dots,c_{n+2}$ be indeterminates. For each $j=1,\dots,n$, let $P_j$ be a polynomial of degree $j$ with norm $(c_1\dotsm c_{n+2}a_{j+1}\dotsm a_n)^{-1}$. Then there holds \begin{multline}\label{cdetrid} \det_{1\leq i,j\leq n}\left(x_i^{-n-1} \prod_{k=1}^{n+2}(1-c_kx_i)\, P_j(x_i)\prod_{k=j+1}^n(1-a_kx_i)\right.\\ \left.-x_i^{n+1}\prod_{k=1}^{n+2}(1-c_kx_i^{-1})\, P_j(x_i^{-1})\prod_{k=j+1}^n(1-a_kx_i^{-1})\right)\\ =\frac{a_1\dotsm a_n} {x_1\dotsm x_n\,(1-c_1\dotsm c_{n+2}a_1\dotsm a_n)} \prod_{i=1}^nP_i(1/a_i)\\\times \prod_{1\leq i<j\leq n+2}(1-c_ic_j)\prod_{i=1}^n(1-x_i^2) \prod_{1\le i<j\le n}a_j(x_i-x_j)(1-1/x_ix_j). \end{multline} \quad \quad \qed \end{Lemma} It is worthwhile to state the limit case $c_{n+2}\to\infty$ of this lemma separately, in which case the norm constraint on the polynomials $P_j$ drops out, but, in return, the degree of $P_j$ gets lowered by one (see \machSeite{RoScAC}\cite[Cor.~5.8]{RoScAC}). \begin{Lemma}\label{cdetr1} Let $x_1,\dots,x_n$, $a_2,\dots,a_n$, and $c_1,\dots,c_{n+1}$ be indeterminates. For each $j=1,\dots,n$, let $P_{j-1}$ be a polynomial of degree at most $j-1$. Then there holds \begin{multline}\label{cdetr1id} \det_{1\leq i,j\leq n}\left(x_i^{-n} \prod_{k=1}^{n+1}(1-c_kx_i)\, P_{j-1}(x_i)\prod_{k=j+1}^n(1-a_kx_i)\right.\\ \left.-x_i^n\prod_{k=1}^{n+1}(1-c_kx_i^{-1})\, P_{j-1}(x_i^{-1})\prod_{k=j+1}^n(1-a_kx_i^{-1})\right)\\ =\prod_{i=1}^nP_{i-1}(1/a_i)\prod_{1\leq i<j\leq n+1}(1-c_ic_j)\\\times \prod_{i=1}^nx_i^{-1}(1-x_i^2) \prod_{1\le i<j\le n}a_j(x_i-x_j)(1-1/x_ix_j). \end{multline} \quad \quad \qed \end{Lemma} Dividing both sides of \eqref{cdetr1id} by $\prod_{i=2}^n a_i^{i-1}$ and then letting $a_i$ tend to $\infty$, $i=2,3,\dots,n$, we arrive at the determinant evaluation below (see \machSeite{RoScAC}\cite[Cor.~5.11]{RoScAC}). Its special case $P_{j-1}(x)=1$, $j=1,\dots,n$, is Lemma~A.11 of \machSeite{SchlAB}\cite{SchlAB}, needed there in order to obtain a {\it $C_n$ matrix inversion}, which was later applied in \machSeite{SchlAG}\cite{SchlAG} to derive {\it multiple q-Abel and q-Rothe summations}. \begin{Lemma}\label{cdetr1cor} Let $x_1,\dots,x_n$, and $c_1,\dots,c_{n+1}$ be indeterminates. For each $j=1,\dots,n$, let $P_{j-1}$ be a polynomial of degree at most $j-1$. Then there holds \begin{multline}\label{cdetr1corid} \det_{1\leq i,j\leq n}\left(x_i^{-j} \prod_{k=1}^{n+1}(1-c_kx_i)\,P_{j-1}(x_i) -x_i^j\prod_{k=1}^{n+1}(1-c_kx_i^{-1})\,P_{j-1}(x_i^{-1})\right)\\ =\prod_{i=1}^n P_{i-1}(0) \prod_{1\leq i<j\leq n+1}(1-c_ic_j) \prod_{i=1}^nx_i^{-1}(1-x_i^2) \prod_{1\le i<j\le n}(x_j-x_i)(1-1/x_ix_j). \end{multline} \quad \quad \qed \end{Lemma} It is an attractive feature of this determinant identity that it contains, at the same time, the {\it Weyl denominator formulae} for the classical root systems $B_n$, $C_n$ and $D_n$ as special cases (cf.\ \machSeite{KratBN}\cite[Lemma~2]{KratBN}). This is seen by setting $P_j(x)=1$ for all $j$, $c_1=c_2=\dots=c_{n-1}=0$, and then $c_n=0$, $c_{n+1}=-1$ for the type $B_n$ case, $c_n=c_{n+1}=0$ for the type $C_n$ case, and $c_n=1$, $c_{n+1}=-1$ for the type $D_n$ case, respectively. \medskip A determinant which is of completely different type, but which also belongs to the category of basic determinant evaluations, is the determinant of a matrix where only two (circular) diagonals are filled with non-zero elements. It was applied with advantage in \machSeite{HaKrAA}\cite{HaKrAA} to evaluate {\it Scott-type permanents}. \begin{Lemma} \label{prop:2diag} Let $n$ and $r$ be positive integers, $r\le n$, and $x_1,x_2,\dots,x_n$, $y_1,y_2,\dots,y_n$ be indeterminates. Then, with $d=\gcd(r,n)$, we have \begin{multline} \det\begin{pmatrix} x_1&0&\dots&0&y_{n-r+1}&0&\\ 0&x_2&0&&0&y_{n-r+2}&0\\ &&\ddots&&&&\ddots&0\\ 0&&&&&&0&y_n\\ y_1&0&\\ 0&y_2&0\\ &0&\ddots&0&&&\ddots&0\\ &&0&y_{n-r}&0&&0&x_n\end{pmatrix}\\ \hskip3.6cm=\prod _{i=1} ^{d}\bigg(\prod _{j=1} ^{n/d}x_{i+(j-1)d}- (-1)^{n/d}\prod _{j=1} ^{n/d}y_{i+(j-1)d}\bigg). \end{multline} {\em(}I.e., in the matrix there are only nonzero entries along two diagonals, one of which is a broken diagonal.{\em)} \quad \quad \qed \end{Lemma} A further basic determinant evaluation which I missed to state in \machSeite{KratBN}\cite{KratBN} is the evaluation of the determinant of a {\it skew circulant matrix} attributed to Scott \machSeite{ScotAB}\cite{ScotAB} in \machSeite{MuirAB}\cite[p.~356]{MuirAB}. It was in fact recently used by Fulmek in \machSeite{FulmAF}\cite{FulmAF} to find a closed form formula for the number of {\it non-intersecting lattice paths with equally spaced starting and end points living on a cylinder}, improving on earlier results by Forrester \machSeite{ForrAC}\cite{ForrAC} on the {\it vicious walker model\/} in {\it statistical mechanics}, see \machSeite{FulmAF}\cite[Lemma~9]{FulmAF}. \begin{Theorem} \label{thm:circulant1} Let $n$ by a fixed positive integer, and let $a_0,a_1,\dots,a_{n-1}$ be indeterminates. Then \begin{multline} \label{eq:circulant1} \det\begin{pmatrix} a_0&a_1&a_2&\dots&a_{n-2}&a_{n-1}\\ -a_{n-1}&a_0&a_1&\dots&a_{n-3}&a_{n-2}\\ -a_{n-2}&-a_{n-1}&a_0&\dots&a_{n-4}&a_{n-3}\\ \hdotsfor6\\ -a_{1}&-a_2&-a_3&\dots&-a_{n-1}&a_{0} \end{pmatrix}\\=\prod _{i=0} ^{n-1}(a_0+\omega^{2i+1}a_1+\omega^{2(2i+1)}a_2+\dots+\omega^{(n-1)(2i+1)}a_{n-1}), \end{multline} where $\omega$ is a primitive $(2n)$-th root of unity.\quad \quad \qed \end{Theorem} \subsection{More confluent determinants} Here I continue the discussion from the beginning of Section~3 in \machSeite{KratBN}\cite[Theorems~20--24]{KratBN}. There I presented determinant evaluations of matrices which, essentially, consist of several vertical strips, each of which is formed by taking a certain column vector and gluing it together with its derivative, its second derivative, etc., respectively by a similar construction where the derivative is replaced by a difference or $q$-difference operator. Since most of this subsection will be under the influence of the so-called {\it ``$q$-disease'',}\footnote{\label{foot:q}The distinctive symptom of this disease is to invariably raise the question ``Is there also a $q$-analogue?" My epidemiological research on {\textsf MathSciNet} revealed that, while basically non-existent during the 1970s, this disease slowly spread during the 1980s, and then had a sharp increase around 1990, when it jumped from about 20 papers per year published with the word ``$q$-analog$*$" in it to over 80 in 1995, and since then it has been roughly stable at 60--70 papers per year. In its simplest form, somebody who is infected by this disease takes a combinatorial identity, and replaces every occurrence of a positive integer $n$ by its {\it ``$q$-analogue"} $1+q+q^2+\dots+q^{n-1}$, inserts some powers of $q$ at the appropriate places, and hopes that the result of these manipulations would be again an identity, thus constituting a ``$q$-analogue" of the original equation. I refer the reader to the bible \machSeite{GaRaAA}\cite{GaRaAA} for a rich source of $q$-identities, and for the right way to look at (most) combinatorial $q$-identities. In another form, given a certain set of objects of which one knows the exact number, one defines a {\it statistics} stat on these objects and now tries to evaluate $\sum _{O\text{ an object}} ^{}q^{\operatorname{stat}(O)}$. For a very instructive text following these lines see \machSeite{FoHaAL}\cite{FoHaAL}, with emphasis on the objects being permutations. There is also an important third form of the disease in which one works in the ring of polynomials in variables $x,y,\dots$ with coefficients being rational functions in $q$, and in which some pairs of variables satisfy commutation relations of the type $xy=qxy$. The study of such polynomial rings and algebras is often motivated by {\it quantum groups} and {\it quantum algebras}. The reader may want to consult \machSeite{KoorAG}\cite{KoorAG} to learn more about this direction. While my description did not make this clear, the three described forms of the $q$-disease are indeed strongly inter-related.} we shall need the standard $q$-notations $(a;q)_k$, denoting the {\em $q$-shifted factorial\/} and being given by $(a;q)_0:=1$ and $$(a;q)_k:=(1-a)(1-aq)\cdots(1-aq^{k-1})$$ if $k$ is a positive integer, as well as $\left[\begin{smallmatrix}\alpha\\k\end{smallmatrix}\right]_q$, denoting the {\em $q$-binomial coefficient\/} and being defined by $\left[\begin{smallmatrix}\alpha\\k\end{smallmatrix}\right]_q=0$ if $k<0$, $\left[\begin{smallmatrix}\alpha\\0\end{smallmatrix}\right]_q=1$, and $$\begin{bmatrix} \alpha\\k\end{bmatrix}_q:= \frac {(1-q^\alpha)(1-q^{\alpha-1})\cdots(1-q^{\alpha-k+1})} {(1-q^k)(1-q^{k-1})\cdots(1-q)}$$ if $k$ is a positive integer. Clearly we have $\lim_{q\to1}\[\smallmatrix \alpha\\k\endsmallmatrix\]_q= \binom \alpha k$. The first result that I present is a $q$-extension of the evaluation of the {\it confluent alternant\/} due to Schendel \machSeite{ScheAA}\cite{ScheAA} (cf.\ \machSeite{KratBN}\cite[paragraph before Theorem~20]{KratBN}). In fact, Theorem~23 of \machSeite{KratBN}\cite{KratBN} already provided a $q$-extension of (a generalisation of) Schendel's formula. However, in \machSeite{JohWAF}\cite[Theorem~1]{JohWAF}, Johnson found a different $q$-extension. The theorem below is a slight generalisation of it. (The theorem below reduces to Johnson's theorem if one puts $C=0$. For $q=1$, the theorem below and \machSeite{KratBN}\cite[Theorem~23]{KratBN} become equivalent. To go from one determinant to the other in this special case, one would have to take a certain factor out of each column.) \begin{Theorem} \label{thm:Johnson1} Let $n$ be a non-negative integer, and let $A_m(X)$ denote the $n\times m$ matrix $$\(\begin{bmatrix} C+i\\i-j\end{bmatrix}_q (X;q)_{i-j}\)_{0\le i\le n-1,\,0\le j\le m-1}.$$ Given a composition of $n$, $n=m_1+\dots+m_\ell$, there holds \begin{multline} \label{eq:Johnson1} \det_{n\times n}\big(A_{m_1}(X_1)\,A_{m_2}(X_2)\dots A_{m_\ell}(X_\ell)\big)\\= q^{\sum_{1\le i<j<k\le \ell}m_im_jm_k} \prod _{1\le i<j\le \ell} ^{} \prod _{g=1} ^{m_i} \prod _{h=1} ^{m_j} \dfrac{(q^{h-1}X_i-q^{g-1}X_j) (1-q^{C+g+h-1+\sum_{r=1}^{i-1}m_r)}} {(1-q^{g+h-1+\sum_{r=1}^{i-1}m_r)}}. \end{multline} \quad \quad \qed \end{Theorem} In \machSeite{JohWAF}\cite[Theorem~2]{JohWAF}, Johnson provides as well a confluent $q$-extension of the evaluation of Cauchy's double alternant \eqref{eq:Cauchy}. Already the case $q=1$ seems to not have appeared in the literature earlier. Here, I was not able to introduce an additional parameter (as, for example, the $C$ in Theorem~\ref{thm:Johnson1}). \begin{Theorem} \label{thm:Johnson2} Let $n$ be a non-negative integer, and let $A_m(X)$ denote the $n\times m$ matrix $$\(\frac {1} {(Y_i-X)(Y_i-qX)(Y_i-q^2X)\cdots(Y_i-q^{j-1}X)}\) _{1\le i\le n,\,1\le j\le m}.$$ Given a composition of $n$, $n=m_1+\dots+m_\ell$, there holds \begin{multline} \label{eq:Johnson2} \det_{n\times n}\big(A_{m_1}(X_1)\,A_{m_2}(X_2)\dots A_{m_\ell}(X_\ell)\big)\\= \frac{\displaystyle \(\prod _{1\le i<j\le n} ^{}(Y_i-Y_j)\) \(\prod _{1\le i<j\le \ell} ^{} \prod _{g=1} ^{m_i} \prod _{h=1} ^{m_j} (q^{h-1}X_j-q^{g-1}X_i) \)} {\displaystyle \prod _{i=1} ^{n} \prod _{j=1} ^{\ell}(Y_i-X_j)(Y_i-qX_j)(Y_i-q^2X_j)\cdots(Y_i-q^{m_j-1}X_j)}. \end{multline} \quad \quad \qed \end{Theorem} A surprising mixture between the confluent alternant and a confluent double alternant appears in \machSeite{CiucAL}\cite[Theorem~A.1]{CiucAL}. Ciucu used it there in order to prove a {\it Coulomb gas law} and a {\it superposition principle} for the joint correlation of certain collections of holes for the {\it rhombus tiling model on the triangular lattice}. (His main result is in fact based on an even more general, and more complex, determinant evaluation, see \machSeite{CiucAL}\cite[Theorem~8.1]{CiucAL}.) \begin{Theorem} \label{thm:Ciucu} Let $s_1,s_2, \dots, s_m\geq1$ and $t_1, t_2,\dots, t_n\geq1$ be integers. Write $S=\sum_{i=1}^m s_i$, $T=\sum_{j=1}^n t_j$, and assume $S\geq T$. Let $x_1,x_2,\dots,x_m$ and $y_1,y_2,\dots,y_n$ be indeterminates. Define $N$ to be the $S\times S$ matrix \begin{equation} \label{eq:Ciuc} N=\left[\begin{matrix} A&B \end{matrix}\right] \end{equation} whose blocks are given by \begin{multline} A= \\ \left(\! \begin{matrix} {\scriptscriptstyle \frac{{\binom 0 0}}{y_1-x_1}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 1 0}}{(y_1-x_1)^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_1-1} 0}}{(y_1-x_1)^{t_1}}}& \ & {\scriptscriptstyle \frac{{\binom 0 0}}{y_n-x_1}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 1 0}}{(y_n-x_1)^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_n-1} 0}}{(y_n-x_1)^{t_n}}} \\ {\scriptscriptstyle \frac{{\binom 1 1}}{(y_1-x_1)^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 2 1}}{(y_1-x_1)^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_1} 1}}{(y_1-x_1)^{t_1+1}}}& \ & {\scriptscriptstyle \frac{{\binom 1 1}}{(y_n-x_1)^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 2 1}}{(y_n-x_1)^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_n} 1}}{(y_n-x_1)^{t_n+1}}} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \!\!\!\!\cdots\!\!\! & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptscriptstyle \frac{{\binom {s_1-1} {s_1-1}}}{(y_1-x_1)^{s_1}}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_1} {s_1-1}}}{(y_1-x_1)^{s_1+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_1+t_1-2} {s_1-1}}}{(y_1-x_1)^{s_1+t_1-1}}}& \ & {\scriptscriptstyle \frac{{\binom {s_1-1} {s_1-1}}}{(y_n-x_1)^{s_1}}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_1} {s_1-1}}}{(y_n-x_1)^{s_1+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_1+t_n-2} {s_1-1}}}{(y_n-x_1)^{s_1+t_n-1}}} \\ \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ & \ & \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ \\ \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ & \ & \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ \\ \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ & \ & \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ \\ {\scriptscriptstyle \frac{{\binom 0 0}}{y_1-x_m}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 1 0}}{(y_1-x_m)^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_1-1} 0}}{(y_1-x_m)^{t_1}}}& \ & {\scriptscriptstyle \frac{{\binom 0 0}}{y_n-x_m}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 1 0}}{(y_n-x_m)^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_n-1} 0}}{(y_n-x_m)^{t_n}}} \\ {\scriptscriptstyle \frac{{\binom 1 1}}{(y_1-x_m)^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 2 1}}{(y_1-x_m)^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_1} 1}}{(y_1-x_m)^{t_1+1}}}& \ & {\scriptscriptstyle \frac{{\binom 1 1}}{(y_n-x_m)^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom 2 1}}{(y_n-x_m)^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {t_n} 1}}{(y_n-x_m)^{t_n+1}}} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \!\!\!\!\cdots\!\!\! & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptscriptstyle \frac{{\binom {s_m-1} {s_m-1}}}{(y_1-x_m)^{s_m}}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_m} {s_m-1}}}{(y_1-x_m)^{s_m+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_m+t_1-2} {s_m-1}}}{(y_1-x_m)^{s_m+t_1-1}}}& \ & {\scriptscriptstyle \frac{{\binom {s_m-1} {s_m-1}}}{(y_n-x_m)^{s_m}}}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_m} {s_m-1}}}{(y_n-x_m)^{s_m+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\binom {s_m+t_n-2} {s_m-1}}}{(y_n-x_m)^{s_m+t_n-1}}} \end{matrix} \!\right) \end{multline} and \begin{equation} B= \left(\begin{matrix} {\scriptstyle {\binom 0 0}x_1^0}\!\!\!& {\scriptstyle {\binom 1 0}x_1}\!\!\!&{\scriptstyle \cdots}\!\!\!& {\scriptstyle {\binom {S-T-1} {0}}x_1^{S-T-1}} \\ {\scriptstyle {\binom {0} {1}}x_1^{-1}}\!\!\!& {\scriptstyle {\binom {1} {1}}x_1^0}\!\!\!&{\scriptstyle \cdots}\!\!\!& {\scriptstyle {\binom {S-T-1} {1}}x_1^{S-T-2}} \\ {\scriptstyle \cdot}\!\!\!& {\scriptstyle \cdot}\!\!\!&{\scriptstyle \ }\!\!\!& {\scriptstyle \cdot} \\ {\scriptstyle \cdot}\!\!\!& {\scriptstyle \cdot}\!\!\!&{\scriptstyle \ }\!\!\!& {\scriptstyle \cdot} \\ {\scriptstyle \cdot}\!\!\!& {\scriptstyle \cdot}\!\!\!&{\scriptstyle \ }\!\!\!& {\scriptstyle \cdot} \\ {\scriptstyle {\binom {0} {s_1-1}}x_1^{1-s_1}}\!\!\!& {\scriptstyle {\binom {1} {s_1-1}}x_1^{2-s_1}}\!\!\!&{\scriptstyle \cdots}\!\!\!& {\scriptstyle {\binom {S-T-1} {s_1-1}}x_1^{S-T-s_1}} \\ {\scriptstyle \ }\!\!\!& {\scriptstyle \ }\!\!\!&\cdot \!\!\!& {\scriptstyle \ } \\ {\scriptstyle \ }\!\!\!& {\scriptstyle \ }\!\!\!&\cdot \!\!\!& {\scriptstyle \ } \\ {\scriptstyle \ }\!\!\!& {\scriptstyle \ }\!\!\!&\cdot \!\!\!& {\scriptstyle \ } \\ {\scriptstyle {\binom {0} {0}}x_m^0}\!\!\!& {\scriptstyle {\binom {1} {0}}x_m}\!\!\!&{\scriptstyle \cdots}\!\!\!& {\scriptstyle {\binom {S-T-1} {0}}x_m^{S-T-1}} \\ {\scriptstyle {\binom {0} {1}}x_m^{-1}}\!\!\!& {\scriptstyle {\binom {1} {1}}x_m^0}\!\!\!&{\scriptstyle \cdots}\!\!\!& {\scriptstyle {\binom {S-T-1} {1}}x_m^{S-T-2}} \\ {\scriptstyle \cdot}\!\!\!& {\scriptstyle \cdot}\!\!\!&{\scriptstyle \ }\!\!\!& {\scriptstyle \cdot} \\ {\scriptstyle \cdot}\!\!\!& {\scriptstyle \cdot}\!\!\!&{\scriptstyle \ }\!\!\!& {\scriptstyle \cdot} \\ {\scriptstyle \cdot}\!\!\!& {\scriptstyle \cdot}\!\!\!&{\scriptstyle \ }\!\!\!& {\scriptstyle \cdot} \\ {\scriptstyle {\binom {0} {s_m-1}}x_m^{1-s_m}}\!\!\!& {\scriptstyle {\binom {1} {s_m-1}}x_m^{2-s_m}}\!\!\!&{\scriptstyle \cdots}\!\!\!& {\scriptstyle {\binom {S-T-1} {s_m-1}}x_m^{S-T-s_m}} \end{matrix}\right). \end{equation} Then we have \vbox{\noindent \begin{equation} \det N =\frac{\prod_{1\leq i<j\leq m}(x_j-x_i)^{s_is_j}\prod_{1\leq i<j\leq n}(y_i-y_j)^{t_it_j}} {\prod_{i=1}^m\prod_{j=1}^n(y_j-x_i)^{s_it_j}}. \end{equation} \quad \quad \qed} \end{Theorem} This theorem generalises at the same time numerous previously obtained determinant evaluations. It reduces of course to Cauchy's double alternant when $m=n$ and $s_1=s_2=\dots=s_m=t_1=t_2=\dots=t_n=1$. (In that case, the submatrix $B$ is empty.) It reduces to the confluent alternant for $t_1=t_2=\dots=t_n=0$ (i.e., in the case where the submatrix $A$ is empty). The case $m=rn$, $s_1=s_2=\dots=s_m=1$, $t_1=t_2=\dots=t_n=r$ is stated as an exercise in \machSeite{MuirAD}\cite[Ex.~42, p.~360]{MuirAD}. Finally, a mixture of the double alternant and the Vandermonde determinant appeared already in \machSeite{HaKrAA}\cite[Theorem~(Cauchy+)]{HaKrAA} where it was used to evaluate {\it Scott-type permanents}. This mixture turns out to be the special case $s_1=s_2=\dots=s_m=t_1=t_2=\dots=t_n=1$ (but not necessarily $m=n$) of Theorem~\ref{thm:Ciucu}. If $S=T$ (i.e., in the case where the submatrix $B$ is empty), Theorem~\ref{thm:Ciucu} provides the evaluation of a confluent double alternant which is different from the one in Theorem~\ref{thm:Johnson2} for $C=0$ and $q=1$. While, for the general form of Theorem~\ref{thm:Ciucu}, I was not able to find a $q$-analogue, I was able to find one for this special case, that is, for the case where $B$ is empty. In view of the fact that there are also $q$-analogues for the other extreme case where the submatrix $A$ is empty (namely Theorem~\ref{thm:Johnson1} and \machSeite{KratBN}\cite[Theorem~23]{KratBN}), I still suspect that a $q$-analogue of the general form of Theorem~\ref{thm:Ciucu} should exist. \begin{Theorem} \label{thm:Ciucu1} Let $s_1, s_2,\dots, s_m\geq1$ and $t_1, t_2,\dots, t_n\geq1$ be integers such that $s_1+s_2+\dots+s_m=t_1+t_2+\dots+t_n$. Let $x_1,x_2,\dots,x_m$ and $y_1,y_2,\dots,y_n$ be indeterminates. Let $A$ be the matrix \begin{multline} A= \\ \left(\! \begin{matrix} {\scriptscriptstyle \frac{{\qbinom 0 0}}{\coef{x_1,y_1}^1}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 1 0}}{\coef{x_1,y_1}^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_1-1} 0}}{\coef{x_1,y_1}^{t_1}}}& \ & {\scriptscriptstyle \frac{{\qbinom 0 0}}{\coef{x_1,y_n}^1}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 1 0}}{\coef{x_1,y_n}^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_n-1} 0}}{\coef{x_1,y_n}^{t_n}}} \\ {\scriptscriptstyle \frac{{\qbinom 1 1}}{\coef{x_1,y_1}^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 2 1}}{\coef{x_1,y_1}^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_1} 1}}{\coef{x_1,y_1}^{t_1+1}}}& \ & {\scriptscriptstyle \frac{{\qbinom 1 1}}{\coef{x_1,y_n}^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 2 1}}{\coef{x_1,y_n}^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_n} 1}}{\coef{x_1,y_n}^{t_n+1}}} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \!\!\!\!\cdots\!\!\! & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptscriptstyle \frac{{\qbinom {s_1-1} {s_1-1}}}{\coef{x_1,y_1}^{s_1}}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_1} {s_1-1}}}{\coef{x_1,y_1}^{s_1+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_1+t_1-2} {s_1-1}}}{\coef{x_1,y_1}^{s_1+t_1-1}}}& \ & {\scriptscriptstyle \frac{{\qbinom {s_1-1} {s_1-1}}}{\coef{x_1,y_n}^{s_1}}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_1} {s_1-1}}}{\coef{x_1,y_n}^{s_1+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_1+t_n-2} {s_1-1}}}{\coef{x_1,y_n}^{s_1+t_n-1}}} \\ \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ & \ & \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ \\ \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ & \ & \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ \\ \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ & \ & \ \!\!\!\!& \ \!\!\!\!&\cdot\!\!\!\!& \ \\ {\scriptscriptstyle \frac{{\qbinom 0 0}}{\coef{x_m,y_1}^1}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 1 0}}{\coef{x_m,y_1}^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_1-1} 0}}{\coef{x_m,y_1}^{t_1}}}& \ & {\scriptscriptstyle \frac{{\qbinom 0 0}}{\coef{x_m,y_n}^1}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 1 0}}{\coef{x_m,y_n}^2}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_n-1} 0}}{\coef{x_m,y_n}^{t_n}}} \\ {\scriptscriptstyle \frac{{\qbinom 1 1}}{\coef{x_m,y_1}^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 2 1}}{\coef{x_m,y_1}^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_1} 1}}{\coef{x_m,y_1}^{t_1+1}}}& \ & {\scriptscriptstyle \frac{{\qbinom 1 1}}{\coef{x_m,y_n}^2}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom 2 1}}{\coef{x_m,y_n}^3}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {t_n} 1}}{\coef{x_m,y_n}^{t_n+1}}} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \!\!\!\!\cdots\!\!\! & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot}& \ & {\scriptstyle\cdot}\!\!\!\!& {\scriptstyle\cdot}\!\!\!\!&\ \!\!\!\!& {\scriptstyle\cdot} \\ {\scriptscriptstyle \frac{{\qbinom {s_m-1} {s_m-1}}}{\coef{x_m,y_1}^{s_m}}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_m} {s_m-1}}}{\coef{x_m,y_1}^{s_m+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_m+t_1-2} {s_m-1}}}{\coef{x_m,y_1}^{s_m+t_1-1}}}& \ & {\scriptscriptstyle \frac{{\qbinom {s_m-1} {s_m-1}}}{\coef{x_m,y_n}^{s_m}}}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_m} {s_m-1}}}{\coef{x_m,y_n}^{s_m+1}}}\!\!\!\!&{\scriptscriptstyle \cdots}\!\!\!\!& {\scriptscriptstyle \frac{{\qbinom {s_m+t_n-2} {s_m-1}}}{\coef{x_m,y_n}^{s_m+t_n-1}}} \end{matrix} \!\right), \end{multline} where $\coef{x,y}^e:=(y-x)(qy-x)(q^2y-x)\cdots(q^{e-1}y-x)$. Then we have \vbox{ \begin{multline} \det A = q^{\frac {1} {6} \sum _{i=1} ^{m}(s_i-1)s_i(2s_i-1)} \(\prod _{1\le i<j\le m} \prod _{g=1} ^{s_i} \prod _{h=1} ^{s_j} (q^{g-1}x_j-q^{h-1}x_i)\)\\ \times \(\prod _{1\le i<j\le n} \prod _{g=1} ^{t_i} \prod _{h=1} ^{t_j} (q^{g-1}y_i-q^{h-1}y_j)\) \(\prod _{j=1} ^{m} \prod _{i=1} ^{n} \prod _{g=1} ^{t_i} \prod _{h=1} ^{s_j} \frac 1{(q^{g+h-2}y_i-x_j)}\). \end{multline} \quad \quad \qed} \end{Theorem} \subsection{More determinants containing derivatives and compositions of series} Inspired by formulae of Mina \machSeite{MinaAA}\cite{MinaAA}, Kedlaya \machSeite{KedlAA}\cite{KedlAA} and Strehl and Wilf \machSeite{StWiAA}\cite{StWiAA} for determinants of matrices the entries of which being given by (coefficients of) {\it multiple derivatives} and {\it compositions of formal power series} (see also \machSeite{KratBN}\cite[Lemma~16]{KratBN}), Chu embedded all these in a larger context in the remarkable systematic study \machSeite{ChuWBG}\cite{ChuWBG}. He shows that, at the heart of these formulae, there is the {\it Fa\`a di Bruno formula}\footnote{As one can read in \machSeite{JohWAE}\cite{JohWAE}, ``Fa\`a di Bruno was neither the first to state the formula that bears his name nor the first to prove it." In Section~4 of that article, the author tries to trace back the roots of the formula. It is apparently impossible to find the author of the formula with certainty. In his book \machSeite{ArboAA}\cite[p.~312]{ArboAA}, Arbogast describes a recursive rule how, from the top term, to generate all other terms in the formula. However, the explicit formula is never written down. (I am not able to verify the conclusions in \machSeite{CraiAA}\cite{CraiAA}. It seems to me that the author mixes the knowledge that we have today with what is really written in \machSeite{ArboAA}\cite{ArboAA}.) The formula appears explicitly in Lacroix's book \machSeite{LacrAA}\cite[p.~629]{LacrAA}, but Lacroix's precise sources remain unknown. I refer the reader to \machSeite{JohWAE}\cite[Sec.~4]{JohWAE} and \machSeite{CraiAA}\cite{CraiAA} for more detailed remarks on the history of the formula.} for multiple derivatives of a composition of two formal power series. Using it, he derives the following determinant reduction formulae \machSeite{ChuWBG}\cite[Theorems~4.1 and 4.2]{ChuWBG} for determinants of matrices containing multiple derivatives of compositions of formal power series. \begin{Theorem} \label{thm:Chu1} Let $f(x)$ and $\phi_k(x)$ and $w_k(x)$, $k=0,1,\dots,n$, be formal power series in $x$ with coefficients in a commutative ring. Then \begin{equation} \label{eq:Chu1} \det_{0\le i,j,\le n}\(\frac {d^j} {dx^j}\bigg(w_j(x)\phi_i\big(f(x)\big)\bigg)\)= \big(f'(x)\big)^{\binom {n+1}2} \(\prod _{k=0} ^{n}w_k(x)\)\det_{0\le i,j,\le n}\(\phi_i^{(j)}\big(f(x)\big)\), \end{equation} where $\phi^{(j)}(x)$ is short for $\frac {d^j} {dx^j}\phi(x)$. If, in addition, $w_k(x)$ is a polynomial of degree at most $k$, $k=1,2,\dots,n$, then \begin{equation} \label{eq:Chu2} \det_{1\le i,j,\le n}\(\frac {d^j} {dx^j}\bigg(w_j(x)\phi_i\big(f(x)\big)\bigg)\)= \big(f'(x)\big)^{\binom {n+1}2} \(\prod _{k=1} ^{n}w_k(x)\)\det_{1\le i,j,\le n}\(\phi_i^{(j)}\big(f(x)\big)\). \end{equation} \quad \quad \qed \end{Theorem} Specialising the series $\phi_i(x)$ so that the determinants on the right-hand sides of \eqref{eq:Chu1} or \eqref{eq:Chu2} can be evaluated, he obtains numerous nice corollaries. Possible choices are $\phi_i(x)=\exp(y_ix)$, $\phi_i(x)=\log(1+y_ix)$, $\phi_i(x)=x^{y_i}$, or $\phi_i(x)=(a_i+b_ix)/(c_i+d_ix)$. See \machSeite{ChuWBG}\cite[Cor.~4.3 and 4.4]{ChuWBG} for the corresponding results. Further reduction formulae and determinant evaluations from \machSeite{ChuWBG}\cite{ChuWBG} address determinants of matrices formed out of coefficients of iterated compositions of formal power series. In order to have a convenient notation, let us write $f^{\coef{n}}(x)$ for the $n$-fold composition of $f$ with itself, $$f^{\coef{n}}(x)=f(f(\cdots(f(x)))),$$ with $n$ occurrences of $f$ on the right-hand side. Chu shows \machSeite{ChuWBG}\cite[Sec.~1.4]{ChuWBG} that it is possible to extend this $n$-fold composition to values of $n$ other than non-negative integers. This given, Theorems~4.6 and 4.7 from \machSeite{ChuWBG}\cite{ChuWBG} read as follows. \begin{Theorem} \label{thm:Chu3} Let $f(x)=x+ \sum _{m=2} ^{\infty}f_mx^m$, $g(x)= \sum _{m=1} ^{\infty}g_mx^m$ and $w_k(x)$, $k=1,2,\dots,n$, be formal power series with coefficients in some commutative ring. Then \begin{equation} \label{eq:Chu3} \det_{1\le i,j\le n}\Big([x^j]w_j(x)f^{\coef{y_i}}\big(g(x)\big)\Big)= f_2^{\binom n2}g_1^{\binom {n+1}2} \( \prod _{k=1} ^{n}w_k(0)\)\( \prod _{1\le i<j\le n} ^{}(y_j-y_i)\), \end{equation} where $[x^j]h(x)$ denotes the coefficient of $x^j$ in the series $h(x)$. If, in addition, $w_n(0)=0$, then \begin{multline} \label{eq:Chu4} \det_{1\le i,j\le n}\Big([x^{j+1}]w_j(x)f^{\coef{y_i}}\big(g(x)\big)\Big) \\= f_2^{\binom n2}g_1^{\binom {n+1}2} \(\prod _{1\le i<j\le n} ^{}(y_j-y_i)\) \det_{1\le i,j\le n}\Big([x^{1+j-i}]w_j(x)\Big). \end{multline} Furthermore, we have \begin{equation} \label{eq:Chu5} \det_{1\le i,j\le n}\Big([x^{j+1}]f^{\coef{y_i}}(x)\Big)= f_2^{\binom {n+1}2}\(\prod _{k=1} ^{n}y_k\)\( \prod _{1\le i<j\le n} ^{}(y_j-y_i)\). \end{equation} \quad \quad \qed \end{Theorem} \subsection{More on Hankel determinant evaluations} Section~2.7 of \machSeite{KratBN}\cite{KratBN} was devoted to {\it Hankel determinants}. There, I tried to convince the reader that, whenever you think that a certain Hankel determinant evaluates nicely, then the explanation will be (sometimes more sometimes less) hidden in the theory of {\em continued fractions} and {\em orthogonal polynomials}. In retrospect, it seems that the success of this try was mixed. Since readers are always right, this has to be blamed entirely on myself, and, indeed, the purpose of the present subsection is to rectify some shortcomings from then. Roughly speaking, I explained in \machSeite{KratBN}\cite{KratBN} that, given a Hankel determinant \begin{equation} \label{eq:Hankel1} \det_{0\le i,j\le n-1}(\mu_{i+j}), \end{equation} to find its evaluation one should expand the generating function of the sequence of coefficients $(\mu_k)_{k\ge0}$ in terms of a continued fraction, respectively find the sequence of orthogonal polynomials $(p_n(x))_{n\ge0}$ with moments $\mu_k$, $k=0,1,\dots$, and then the value of the Hankel determinant \eqref{eq:Hankel1} can be read off the coefficients of the continued fraction, respectively from the recursion coefficients of the orthogonal polynomials. What I missed to state is that the knowledge of the orthogonal polynomials makes it also possible to find the value of the Hankel determinants which start with $\mu_1$ and $\mu_2$, respectively (instead of $\mu_0$). In the theorem below I summarise the results that were already discussed in \machSeite{KratBN}\cite{KratBN} (for which classical references are \machSeite{WallCF}% \cite[Theorem 51.1]{WallCF} or \machSeite{VienAE}% \cite[Cor.~6, (19), on p.~IV-17; Proposition~1, (7), on p.~V-5]{VienAE}), and I add the two missing ones. \begin{Theorem} \label{cor:cfracHankel} Let $(\mu_k)_{k\ge0}$ be a sequence with generating function $\sum_{k=0}^\infty{\mu_k}x^k$ written in the form \begin{equation} \label{eq:momentgf} \sum_{k=0}^\infty{\mu_k}x^k=\cfrac{ \mu_0} {1+a_0x-\cfrac{ b_1x^2} {1+a_1x-\cfrac{ b_2x^2} {1+a_2x-\cdots}}}\quad . \end{equation} Then \begin{equation} \label{eq:Hankel2} \det_{0\le i,j\le n-1}(\mu_{i+j})=\mu_0^nb_1^{n-1}b_2^{n-2}\cdots b_{n-2}^2b_{n-1}. \end{equation} Let $(p_n(x))_{n\ge0}$ be a sequence of monic polynomials, the polynomial $p_n(x)$ having degree $n$, which is orthogonal with respect to some functional $L$, that is, $L(p_m(x)p_n(x))=\delta_{m,n}c_{n}$, where the $c_n$'s are some non-zero constants and $\delta_{m,n}$ is the Kronecker delta. Let \begin{equation} p_{n+1}(x)=(a_{n}+x)p_{n}(x)-b_{n}p_{n-1}(x) \label{eq:three-term2} \end{equation} be the corresponding three-term recurrence which is guaranteed by Favard's theorem. Then the generating function $\sum _{k=0} ^{\infty}\mu_kx^k$ for the moments $\mu_k=L(x^k)$ satisfies \eqref{eq:momentgf} with the $a_i$'s and $b_i$'s being the coefficients in the three-term recurrence \eqref{eq:three-term2}. In particular, the Hankel determinant evaluation \eqref{eq:Hankel2} holds, with the $b_i$'s from the three-term recurrence \eqref{eq:three-term2}. If $(q_n)_{n\ge0}$ is the sequence recursively defined by $q_0=1$, $q_1=-a_0$, and $$q_{n+1}=-a_n q_n-b_nq_{n-1},$$ then in the situation above we also have \begin{equation} \label{eq:Hankel3} \det_{0\le i,j\le n-1}(\mu_{i+j+1})=\mu_0^nb_1^{n-1}b_2^{n-2}\cdots b_{n-2}^2b_{n-1}q_n \end{equation} and \begin{equation} \label{eq:Hankel4} \det_{0\le i,j\le n-1}(\mu_{i+j+2})=\mu_0^nb_1^{n-1}b_2^{n-2}\cdots b_{n-2}^2b_{n-1} \sum _{k=0} ^{n}q_k^2b_{k+1}\cdots b_{n-1}b_n. \end{equation} \quad \quad \qed \end{Theorem} I did not find a reference for \eqref{eq:Hankel3} and \eqref{eq:Hankel4}. These two identities follow however easily from Viennot's combinatorial model \machSeite{VienAE}\cite{VienAE} for orthogonal polynomials and Hankel determinants of moments. More precisely, in this theory the moments $\mu_k$ are certain generating functions for {\it Motzkin paths}, and, due to Theorem~\ref{thm:nonint}, the Hankel determinants $\det_{0\le i,j\le n-1}(\mu_{i+j+m})$ are generating functions for families $(P_1,P_2,\dots,P_n)$ of non-intersecting Motzkin paths, $P_i$ running from $(-i,0)$ to $(j+m,0)$. In the case $m=0$, it is explained in \machSeite{VienAE}\cite[Ch.~IV]{VienAE} how to find the corresponding Hankel determinant evaluation \eqref{eq:Hankel2} using this combinatorial model. The idea is that in that case there is a unique family of non-intersecting Motzkin paths, and its weight gives the right-hand side of \eqref{eq:Hankel2}. If $m=1$ or $m=2$ one can argue similarly. The paths are uniquely determined with the exception of their portions in the strip $0\le x\le m$. The various possibilities that one has there then yield the right-hand sides of \eqref{eq:Hankel3} and \eqref{eq:Hankel4}. Since there are so many explicit families of orthogonal polynomials, and, hence, so many ways to apply the above theorem, I listed only a few standard Hankel determinant evaluations explicitly in \machSeite{KratBN}\cite{KratBN}. I did append a long list of references and sketched in which ways these give rise to more Hankel determinant evaluations. Apparently, these remarks were at times too cryptic, in particular concerning the theme {\it ``orthogonal polynomials as moments}.'' This is treated systematically in the two papers \machSeite{IsStAB}\machSeite{IsStAC}\cite{IsStAB,IsStAC} by Ismail and Stanton. There it is shown that certain classical polynomials $(r_n(x))_{n\ge0}$, such as, for example, the {\em Laguerre polynomials}, the {\em Meixner polynomials}, or the {\em Al-Salam--Chihara polynomials} (but there are others as well, see \machSeite{IsStAB}\machSeite{IsStAC}\cite{IsStAB,IsStAC}), are {\it moments} of other families of classical orthogonal polynomials. Thus, application of Theorem~\ref{cor:cfracHankel} with $\mu_n=r_n(x)$ immediately tells that the evaluations of the corresponding Hankel determinants \begin{equation} \label{eq:pn} \det_{0\le i,j\le n-1}\big(r_{i+j}(x)\big) \end{equation} (and also the higher ones in \eqref{eq:Hankel3} and \eqref{eq:Hankel4}) are known. In particular, the explicit forms can be extracted from the coefficients of the three-term recursions for these other families of orthogonal polynomials. Thus, whenever you encounter a determinant of the form \eqref{eq:pn}, you must check whether $(r_n(x))_{n\ge0}$ is a family of orthogonal polynomials (which, as I explained in \machSeite{KratBN}\cite{KratBN}, one does by consulting the compendium \machSeite{KoSwAA}\cite{KoSwAA} of hypergeometric orthogonal polynomials compiled by Koekoek and Swarttouw), and if the answer is ``yes", you will find the solution of your determinant evaluation through the results in \machSeite{IsStAB}\machSeite{IsStAC}\cite{IsStAB,IsStAC} by applying Theorem~\ref{cor:cfracHankel}. \medskip While Theorem~\ref{cor:cfracHankel} describes in detail the connexion between Hankel determinants and the continued fractions of the type \eqref{eq:momentgf}, which are often called {\it $J$-fractions} (which is short for {\it Jacobi continued fractions}), I missed to tell in \machSeite{KratBN}\cite{KratBN} that there is also a close relation between Hankel determinants and so-called {\it $S$-fractions} (which is short for {\it Stieltjes continued fractions}). I try to remedy this by the theorem below (cf.\ for example \machSeite{JoThAA}\cite[Theorem~7.2]{JoThAA}, where $S$-fractions are called {\it regular $C$-fractions}). In principle, since $S$-fractions are special cases of $J$-fractions \eqref{eq:momentgf} in which the coefficients $a_i$ are all zero, the corresponding result for the Hankel determinants is in fact implied by Theorem~\ref{cor:cfracHankel}. Nevertheless, it is useful to state it separately. I am not able to give a reference for \eqref{eq:Hankel7}, but, again, it is not too difficult to derive it from Viennot's combinatorial model \machSeite{VienAE}\cite{VienAE} for orthogonal polynomials and moments that was mentioned above. \begin{Theorem} \label{thm:cfrac2} Let $(\mu_k)_{k\ge0}$ be a sequence with generating function $\sum_{k=0}^\infty{\mu_k}x^k$ written in the form \begin{equation} \label{eq:momentgfS} \sum_{k=0}^\infty{\mu_k}x^k=\cfrac{ \mu_0} {1+\cfrac{ a_1x} {1+\cfrac{ a_2x} {1+\cdots}}}\quad . \end{equation} Then \begin{align} \label{eq:Hankel5} \det_{0\le i,j\le n-1}(\mu_{i+j})&=\mu_0^n(a_1a_2)^{n-1}(a_3a_4)^{n-2}\cdots (a_{2n-5}a_{2n-4})^2(a_{2n-3}a_{2n-2}),\\ \label{eq:Hankel6} \det_{0\le i,j\le n-1}(\mu_{i+j+1})&=(-1)^n\mu_0^n a_1^n(a_2a_3)^{n-1}(a_4a_5)^{n-2}\cdots (a_{2n-4}a_{2n-3})^2(a_{2n-2}a_{2n-1}), \end{align} and \vbox{ \begin{multline} \label{eq:Hankel7} \det_{0\le i,j\le n-1}(\mu_{i+j+2})=\mu_0^n a_1^n(a_2a_3)^{n-1}(a_4a_5)^{n-2}\cdots (a_{2n-4}a_{2n-3})^2(a_{2n-2}a_{2n-1})\\ \times \sum _{0\le i_1-1<i_2-2<\dots<i_n-n\le n} ^{}a_{i_1}a_{i_2}\cdots a_{i_n}. \end{multline} \quad \quad \qed} \end{Theorem} Using this theorem, Tamm \machSeite{TammAA}\cite[Theorem~3.1]{TammAA} observed that from {\it Gau\ss' continued fraction for the ratio of two contiguous $_2F_1$-series} one can deduce several interesting binomial Hankel determinant evaluations, some of them had already been found earlier by E\u gecio\u glu, Redmond and Ryavec \machSeite{EgRiAA}\cite[Theorem~4]{EgRiAA} while working on {\it polynomial Riemann hypotheses}. Gessel and Xin \machSeite{GeXiAB}\cite{GeXiAB} undertook a systematic analysis of this approach, and they arrived at a set of 18 Hankel determinant evaluations, which I list as \eqref{eq:3nA}--\eqref{eq:3nR} in the theorem below. They are preceded by the Hankel determinant evaluation \eqref{eq:3nAAAA}, which appears only in \machSeite{EgRiAA}\cite[Theorem~4]{EgRiAA}. \begin{Theorem} \label{thm:Tamm} For any positive integer $n$, there hold \begin{equation} \label{eq:3nAAAA} \det_{0\le i,j\le n-1}\(\binom {3i+3j+2}{i+j}\) = \prod _{i=0} ^{n-1}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nA} \det_{0\le i,j\le n-1}\(\frac {1} {3i+3j+1}\binom {3i+3j+1}{i+j}\) = \prod _{i=0} ^{n-1}\frac {(\frac {2} {3})_i\,(\frac {1} {6})_i\, (\frac {4} {3})_i\,(\frac {5} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nB} \det_{0\le i,j\le n-1}\(\frac {1} {3i+3j+4}\binom {3i+3j+4}{i+j+1}\) = \prod _{i=0} ^{n-1}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nC} \det_{0\le i,j\le n-1}\(\frac {1} {3i+3j+2}\binom {3i+3j+2}{i+j+1}\) = \prod _{i=0} ^{n-1}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nD} \det_{0\le i,j\le n-1}\(\frac {1} {3i+3j+5}\binom {3i+3j+5}{i+j+2}\) = \prod _{i=0} ^{n}\frac {(\frac {2} {3})_i\,(\frac {1} {6})_i\, (\frac {4} {3})_i\,(\frac {5} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nE} \det_{0\le i,j\le n-1}\(\frac {2} {3i+3j+1}\binom {3i+3j+1}{i+j+1}\) = \prod _{i=0} ^{n}\frac {(\frac {2} {3})_i\,(\frac {1} {6})_i\, (\frac {4} {3})_i\,(\frac {5} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nF} \det_{0\le i,j\le n-1}\(\frac {2} {3i+3j+4}\binom {3i+3j+4}{i+j+2}\) = \prod _{i=0} ^{n}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nG} \det_{0\le i,j\le n-1}\(\frac {2} {(3i+3j+1)(3i+3j+2)} \binom {3i+3j+2}{i+j+1}\) = \prod _{i=0} ^{n-1}2\frac {(\frac {5} {3})_i\,(\frac {1} {6})_i\, (\frac {7} {3})_i\,(\frac {5} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{multline} \label{eq:3nH} \det_{0\le i,j\le n-1}\(\frac {2} {(3i+3j+4)(3i+3j+5)} \binom {3i+3j+5}{i+j+2}\) \\ =(-1)^n \prod _{i=1} ^{n}\frac {(\frac {5} {3})_i\,(\frac {1} {6})_i\, (\frac {4} {3})_i\,(-\frac {1} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{multline} \begin{equation} \label{eq:3nI} \det_{0\le i,j\le n-1}\(\frac {(9i+9j+5)} {(3i+3j+1)(3i+3j+2)} \binom {3i+3j+2}{i+j+1}\) = \prod _{i=0} ^{n-1}5\frac {(\frac {2} {3})_i\,(\frac {7} {6})_i\, (\frac {4} {3})_i\,(\frac {11} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}, \end{equation} \begin{equation} \label{eq:3nJ} \det_{0\le i,j\le n-1}\(\frac {(9i+9j+14)} {(3i+3j+4)(3i+3j+5)} \binom {3i+3j+5}{i+j+2}\) = \prod _{i=1} ^{n}2\frac {(\frac {2} {3})_i\,(\frac {7} {6})_i\, (\frac {1} {3})_i\,(\frac {5} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Let $a_0=-2$ and $a_m=\frac {1} {3m+1}\binom {3m+1}m$ for $m\ge1$. Then \begin{equation} \label{eq:3nK} \det_{0\le i,j\le n-1}(a_{i+j})= \prod _{i=0} ^{n-1}(-2)\frac {(\frac {1} {3})_i\,(-\frac {1} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Let $b_0=10$ and $b_m=\frac {2} {3m+2}\binom {3m+2}m$ for $m\ge1$. Then \begin{equation} \label{eq:3nL} \det_{0\le i,j\le n-1}(b_{i+j})= \prod _{i=0} ^{n-1}10\frac {(\frac {2} {3})_i\,(\frac {1} {6})_i\, (\frac {7} {3})_i\,(\frac {11} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Furthermore, \begin{equation} \label{eq:3nM} \det_{0\le i,j\le n-1}\(\frac {2} {3i+3j+5}\binom {3i+3j+5}{i+j+1}\) = \prod _{i=0} ^{n}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Let $c_0=\frac {7} {2}$ and $c_m=\frac {2} {3m+1}\binom {3m+1}{m+1}$ for $m\ge1$. Then \begin{equation} \label{eq:3nN} \det_{0\le i,j\le n-1}(c_{i+j})= \prod _{i=0} ^{n}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Let $d_0=-5$ and $d_m=\frac {8} {(3m+1)(3m+2)}\binom {3m+2}m$ for $m\ge1$. Then \begin{equation} \label{eq:3nO} \det_{0\le i,j\le n-1}(d_{i+j})= \prod _{i=0} ^{n-1}(-5)\frac {(\frac {4} {3})_i\,(-\frac {1} {6})_i\, (\frac {8} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Furthermore, \begin{equation} \label{eq:3nP} \det_{0\le i,j\le n-1}\(\frac {8} {(3i+3j+4)(3i+3j+5)} \binom {3i+3j+5}{i+j+1}\) = \prod _{i=0} ^{n-1}2\frac {(\frac {7} {3})_i\,(\frac {5} {6})_i\, (\frac {8} {3})_i\,(\frac {7} {6})_i} {(\frac {5} {2})_{2i}\,(\frac {7} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Let $e_0=14$ and $e_m=\frac {2(9m+5)} {(3m+1)(3m+2)}\binom {3m+2}m$ for $m\ge1$. Then \begin{equation} \label{eq:3nQ} \det_{0\le i,j\le n-1}(e_{i+j})= \prod _{i=0} ^{n-1}14\frac {(\frac {1} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {13} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} Furthermore, \begin{equation} \label{eq:3nR} \det_{0\le i,j\le n-1}\(\frac {2(9i+9j+14)} {(3i+3j+4)(3i+3j+5)} \binom {3i+3j+5}{i+j+1}\) = \prod _{i=1} ^{n}2\frac {(\frac {2} {3})_i\,(\frac {7} {6})_i\, (\frac {1} {3})_i\,(\frac {5} {6})_i} {(\frac {1} {2})_{2i}\,(\frac {3} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} \quad \quad \qed \end{Theorem} Some of the numbers appearing on the right-hand sides of the formulae in this theorem have combinatorial significance, although no intrinsic explanation is known why this is the case. More precisely, the numbers on the right-hand sides of \eqref{eq:3nA}, \eqref{eq:3nD} and \eqref{eq:3nE} count {\it cyclically symmetric transpose-complementary plane partitions} (cf.\ \machSeite{MiRRAD}\cite{MiRRAD} and \machSeite{BresAO}\cite{BresAO}), whereas those on the right-hand sides of \eqref{eq:3nAAAA}, \eqref{eq:3nB}, \eqref{eq:3nC} and \eqref{eq:3nF} count {\em vertically symmetric alternating sign matrices} (cf.\ \machSeite{KupeAH}\cite{KupeAH} and \machSeite{BresAO}\cite{BresAO}). In \machSeite{EgRiAA}\cite[Theorem~4]{EgRiAA}, E\u gecio\u glu, Redmond and Ryavec prove also the following common generalisation of \eqref{eq:3nB} and \eqref{eq:3nC}. (The first identity is the special case $x=0$, while the second is the special case $x=1$ of the following theorem.) \begin{Theorem} \label{thm:EgRR} For $m\ge0$, let $s_m(x)= \sum _{k=0} ^{m}\frac {k+1} {m+1}\binom {3m-k+1}{m-k}x^k$. Then, for any positive integer $n$, there holds \begin{equation} \label{eq:EgRR} \det_{0\le i,j\le n-1}(s_{i+j}(x))= \prod _{i=0} ^{n-1}\frac {(\frac {4} {3})_i\,(\frac {5} {6})_i\, (\frac {5} {3})_i\,(\frac {7} {6})_i} {(\frac {3} {2})_{2i}\,(\frac {5} {2})_{2i}}\(\frac {27} {4}\)^{2i}. \end{equation} \quad \quad \qed \end{Theorem} As I mentioned above, in \machSeite{KratBN}\cite{KratBN} I only stated a few special Hankel determinant evaluations explicitly, because there are too many ways to apply Theorems~\ref{cor:cfracHankel} and \ref{thm:cfrac2}. I realise, however, that I should have stated the evaluation of the {\it Hankel determinant of Catalan numbers} there. I make this up now by doing this in the theorem below. I did not do it then because orthogonal polynomials are not needed for its evaluation (the orthogonal polynomials which are tied to Catalan numbers as moments are {\it Chebyshev polynomials}, but, via Theorems~\ref{cor:cfracHankel} and \ref{thm:cfrac2}, one would only cover the cases $m=0,1,2$ in the theorem below). In fact, the Catalan number $C_n=\frac {1} {n+1}\binom {2n}n$ can be alternatively written as $C_n=(-1)^n2^{2n+1}\binom {1/2}{n+1}$, and therefore the Hankel determinant evaluation below follows from \machSeite{KratBN}\cite[Theorem~26, (3.12)]{KratBN}. This latter observation shows that even a more general determinant, namely $\det_{0\le i,j\le n-1}(C_{\lambda_i+j})$, can be evaluated in closed form. For historical remarks on this ubiquitous determinant see \machSeite{GhKrAA}\cite[paragraph before the Appendix]{GhKrAA}. \begin{Theorem} \label{thm:Catalan} \begin{equation} \label{eq:Catalan} \det_{0\le i,j\le n-1}(C_{m+i+j})= \prod _{1\le i\le j\le m-1} ^{}\frac {2n+i+j} {i+j}. \end{equation} \quad \quad \qed \end{Theorem} As in \machSeite{KratBN}\cite{KratBN}, let me conclude the part on Hankel determinants by pointing the reader to further papers containing interesting results on them, high-lighting sometimes the point of view of orthogonal polynomials that I explained above, sometimes a combinatorial point of view. The first point of view is put forward in \machSeite{WimpAB}\cite{WimpAB} (see \machSeite{KratZZ}\cite{KratZZ} for the solution of the conjectures in that paper) in order to present Hankel determinant evaluations of matrices with {\it hypergeometric $_2F_1$-series} as entries. The orthogonal polynomials approach is also used in \machSeite{CvRIAA}\cite{CvRIAA} to show that a certain Hankel determinant defined by {\it Catalan numbers} evaluates to {\it Fibonacci numbers}. In \machSeite{AnWiAA}\cite{AnWiAA}, one finds Hankel determinant evaluations involving generalisations of the {\it Bernoulli numbers}. The combinatorial point of view dominates in \machSeite{AignAA}% \machSeite{CiglAM}% \machSeite{CiglAO}% \machSeite{CiglAV}% \machSeite{EhreAB}\cite{AignAA,CiglAM,CiglAO,CiglAV,EhreAB}, where Hankel determinants involving {\it $q$-Catalan numbers, $q$-Stirling numbers}, and {\it $q$-Fibonacci numbers} are considered. A very interesting new direction, which seems to have much potential, is opened up by Luque and Thibon in \machSeite{LuThAB}\cite{LuThAB}. They show that {\it Selberg-type integrals} can be evaluated by means of {\it Hankel hyperdeterminants}, and they prove many hyperdeterminant generalisations of classical Hankel determinant evaluations. At last, (but certainly not least!), I want to draw the reader's attention to Lascoux's ``unorthodox"\footnote{It could easily be that it is the ``modern" treatment of the theory which must be labelled with the attribute ``unorthodox." As Lascoux documents in \machSeite{LascAZ}\cite{LascAZ}, in his treatment {\it he} follows the tradition of great masters such as Cauchy, Jacobi or Wro\'nski \dots} approach to Hankel determinants and orthogonal polynomials through symmetric functions which he presents in detail in \machSeite{LascAZ}\cite[Ch.~4, 5, 8]{LascAZ}. In particular, Theorem~\ref{cor:cfracHankel}, Eq.~\eqref{eq:Hankel2} are the contents of Theorem~8.3.1 in \machSeite{LascAZ}\cite{LascAZ} (see also the end of Section~5.3 there), and Theorem~\ref{thm:cfrac2}, Eqs.~\eqref{eq:Hankel5} and \eqref{eq:Hankel6} are the contents of Theorem~4.2.1 in \machSeite{LascAZ}\cite{LascAZ}. The usefulness of this symmetric function approach is, for example, demonstrated in \machSeite{HoLMAA}% \machSeite{HoLMAB}\cite{HoLMAA,HoLMAB} in order to evaluate Hankel determinants of matrices the entries of which are {\it Rogers--Szeg\H o}, respectively {\it Meixner polynomials}. \subsection{More binomial determinants} A vast part of Section~3 in \machSeite{KratBN}\cite{KratBN} is occupied by binomial determinants. As I mentioned in Section~\ref{sec:comb} of the present article, an extremely rich source for binomial determinants is rhombus tiling enumeration. I want to present here some which did not already appear in \machSeite{KratBN}\cite{KratBN}. To begin with, I want to remind the reader of an old problem posed by Andrews in \machSeite{AndrAO}\cite[p.~105]{AndrAO}. The determinant in this problem is a variation of a determinant which enumerates {\it cyclically symmetric plane partitions} and {\it descending plane partitions}, which was evaluated by Andrews in \machSeite{AndrAN}\cite{AndrAN} (see also \machSeite{KratBN}\cite[Theorem~32]{KratBN}; the latter determinant arises from the one in \eqref{eq:desc-var} by replacing $j+1$ by $j$ in the bottom of the binomial coefficient). \begin{Problem} \label{prob:7} Evaluate the determinant \begin{equation} \label{eq:desc-var} D_1(n):=\det_{0\le i,j\le n-1}\left(\delta_{ij}+\binom {\mu+i+j}{j+1}\right), \end{equation} where $\delta_{ij}$ is the Kronecker delta. In particular, show that \begin{equation} \label{eq:fmu} \frac {D_1(2n)} {D_1(2n-1)}=(-1)^{\binom {n-1}2} \frac {2^n\,\(\frac\mu2+n\)_{\cl{n/2}}\,\(\frac {\mu} {2}+2n+\frac {1} {2}\)_{n-1}} {(n)_n\,\(-\frac {\mu} {2}-2n+\frac {3} {2}\)_{\cl{(n-2)/2}}}. \end{equation} \end{Problem} The determinants $D_1(n)$ are rather intriguing. Here are the first few values: \begin{align*} D_1(1)&=\mu+1,\\ D_1(2)&=(\mu+1)(\mu+2),\\ D_1(3)&= \frac {1} {12}{( \mu + 1 ) ( \mu + 2 ) ( \mu + 3 ) ( \mu + 14 ) },\\ D_1(4)&=\frac {1} {72} {( \mu + 1 ) ( \mu + 2 ) ( \mu + 3 ) ( \mu + 4 ) ( \mu + 9 ) ( \mu + 14 ) },\\ D_1(5)&= \frac {1} {8640} ( \mu + 1 ) ( \mu + 2 ) ( \mu + 3 ) ( \mu + 4 ) ( \mu + 5 ) ( \mu + 9 ) \\ &\kern3cm \times ( 3432 + 722 \mu + 45 \mu^2 + \mu^3 ) ,\\ D_1(6)&=\frac {1} {518400} ( \mu + 1 ) ( \mu + 2 ) ( \mu + 3 ) ( \mu + 4 ) ( \mu + 5 ) ( \mu + 6 ) ( \mu + 8 ) ( \mu + 13 ) \\ &\kern3cm \times ( \mu + 15 ) ( 3432 + 722 \mu + 45 \mu^2 + \mu^3 ) ,\\ D_1(7)&=\frac {1} {870912000} ( \mu + 1 ) ( \mu + 2 ) ( \mu + 3 ) ( \mu + 4 ) ( \mu + 5 ) ( \mu + 6 ) ( \mu + 7 ) ( \mu + 8 ) ^2\\ &\kern3cm \times ( \mu + 13 ) ( \mu + 15 ) ^2 ( \mu + 34 ) ( \mu^3 + 47 \mu^2 + 954 \mu + 5928) ,\\ D_1(8)&=\frac {1} {731566080000} ( \mu + 1 ) ( \mu + 2 ) ( \mu + 3 ) ( \mu + 4 ) ( \mu + 5 ) ( \mu + 6 ) ( \mu + 7 ) ( \mu + 8 ) ^3 \\ &\kern3cm\times ( \mu + 10 ) ( \mu + 15 ) ^2 ( \mu + 17 ) ( \mu + 19 ) \\ &\kern3cm\times ( \mu + 21 ) ( \mu + 34 ) ( \mu^3 + 47 \mu^2 + 954 \mu + 5928 ) . \end{align*} ``So," these determinants factor almost completely, there is only a relatively small (in degree) irreducible factor which is not linear. (For example, this factor is of degree $6$ for $D_1(9)$ and $D_1(10)$, and of degree $7$ for $D_1(11)$ and $D_1(12)$.) Moreover, this ``bigger" factor is always the same for $D_1(2n-1)$ and $D_1(2n)$. Not only that, the quotient which is predicted in \eqref{eq:fmu} is at the same time a building block in the result of the evaluation of the determinant which enumerates the cyclically symmetric and descending plane partitions (see \machSeite{AndrAO}\cite{AndrAO}). All this begs for an explanation in terms of a factorisation of the matrix of which the determinant is taken from. In fact, for the plane partition matrix there is such a factorisation, due to Mills, Robbins and Rumsey \machSeite{MiRRAD}\cite[Theorem~5]{MiRRAD} (see also \machSeite{KratBN}\cite[Theorem~36]{KratBN}). The question is whether there is a similar one for the matrix in \eqref{eq:desc-var}. Inspired by this conjecture and by the variations in \machSeite{CiEKAA}\cite[Theorems~11--13]{CiEKAA} (see \machSeite{KratBN}\cite[Theorem~35]{KratBN}) on Andrews' original determinant evaluation in \machSeite{AndrAN}\cite{AndrAN}, Guoce Xin (private communication) observed that, if we change the sign in front of the Kronecker delta in \eqref{eq:desc-var}, then the resulting determinant factors completely into linear factors. \begin{Conjecture} \label{conj:Xin1} Let $\mu$ be an indeterminate and $n$ be a non-negative integer. The determinant \begin{equation} \label{eq:Xin-det1} \det_{0\le i,j\le n-1}\left(-\delta_{ij}+\binom {\mu+i+j}{j+1}\right) \end{equation} is equal to \begin{multline} \label{eq:Xin-erg1a} (-1)^{n/2}2^{n(n+2)/4} \frac {\(\frac \mu2\)_{n/2}} {\(\frac n2\)!} \( \prod_{i=0}^{(n-2)/2}\frac {i!^2}{(2i)!^2} \)\\ \times \( \prod_{i=0}^{\fl{(n-4)/4}} \(\frac\mu2+3i+\frac52\)_{(n-4i-2)/2}^2 \(-\frac\mu2-\frac{3n}2+3i+3\)_{(n-4i-4)/2}^2\) \end{multline} if $n$ is even, and it is equal to \begin{multline} \label{eq:Xin-erg1b} (-1)^{(n-1)/2}2^{(n+3)(n+1)/4}\,\(\frac {\mu-1} {2}\)_{(n+1)/2} \( \prod_{i=0}^{(n-1)/2}\frac{i!\,(i+1)!} {(2i)!\,(2i+2)!}\)\\ \times \(\prod_{i=0}^{\fl{(n-3)/4}}\(\frac\mu2+3i+\frac52\)_{(n-4i-3)/2}^2 \(-\frac\mu2-\frac{3n}2+\frac32+3i\)_{(n-4i-1)/2}^2\) \end{multline} if $n$ is odd. \end{Conjecture} In fact, it seems that also the ``next" determinant, the determinant where one replaces $j+1$ at the bottom of the binomial coefficient in \eqref{eq:Xin-det1} by $j+2$ factors completely when $n$ is odd. (It does not when $n$ is even, though.) \begin{Conjecture} \label{conj:Xin2} Let $\mu$ be an indeterminate. For any odd non-negative integer $n$ there holds \begin{multline} \label{eq:Xin-det2} \det_{0\le i,j\le n-1}\left(-\delta_{ij}+\binom {\mu+i+j}{j+2}\right)\\= (-1)^{(n-1)/2}2^{(n-1)(n+5)/4}(\mu+1) \frac {\(\frac \mu2-1\)_{(n+1)/2}} {\(\frac{n+1}2\)!} \(\prod_{i=0}^{(n-1)/2}\frac {i!^2} {(2i)!^2} \(\frac\mu2+3i+\frac32\)_{(n-4i-1)/2}^2\)\\ \times \( \prod_{i=0}^{\fl{(n-3)/4}} \(-\frac\mu2-\frac{3n}2+3i+\frac52\)_{(n-4i-3)/2}^2\) . \end{multline} \end{Conjecture} For the combinatorialist I add that all the determinants in Problem~\ref{prob:7} and Conjectures~\ref{conj:Xin1} and \ref{conj:Xin2} count certain rhombus tilings, as do the original determinants in \machSeite{AndrAN}% \machSeite{AndrAO}% \machSeite{CiEKAA}\cite{AndrAN,AndrAO,CiEKAA}. Alain Lascoux (private communication) did not understand why we should stop here, and he hinted at a parametric family of determinant evaluations into which the case of odd $n$ of Conjecture~\ref{conj:Xin1} is embedded as a special case. \begin{Conjecture} \label{conj:Xin3} Let $\mu$ be an indeterminate. For any odd non-negative integers $n$ and $r$ there holds \begin{multline} \label{eq:Xin-det3} \det_{0\le i,j\le n-1}\left(-\delta_{i,j+r-1}+\binom {\mu+i+j}{j+r}\right)\\\kern-1pt = (-1)^{(n-r)/2}2^{(n^2+6n-2nr+r^2-4r+2)/4} \( \prod _{i=0} ^{r-2}i!\) \( \prod _{i=0} ^{(r-3)/2}\frac {(n-2i-2)!^2} {\(\frac {n-2i-3} {2}\)!^2\, (n+2i)!\, (n+2i+2)!}\)\kern-1pt\\ \times (\mu-r)\(\frac {\mu+1} {2}\)_{(n-r)/2}\( \prod _{i=1} ^{r-1}(m-r+i)_{n+r-2i+1}\) \( \prod_{i=0}^{(n-1)/2}\frac{i!\,(i+1)!} {(2i)!\,(2i+2)!}\)\\ \times \(\prod_{i=0}^{\fl{(n-r-2)/4}}\(\frac\mu2+3i+r+\frac32\)_{(n-4i-r-2)/2}^2 \(-\frac\mu2-\frac{3n}2+\frac {r} {2}+3i+1\)_{(n-4i-r)/2}^2\). \end{multline} \end{Conjecture} \medskip The next binomial determinant that I want to mention is, strictly speaking, not a determinant but a Pfaffian (see \eqref{eq:Pfaff} for the definition). While doing {\it $(-1)$-enumeration of sef-complementary plane partitions}, Eisenk\"olbl \machSeite{EisTAF}\cite{EisTAF} encountered an, I admit, complicated looking Pfaffian, \begin{equation} \label{eq:mn} \underset{1\le i,j\le a}\operatorname{Pf}\big(M(m_1,m_2,n_1,n_2,a,b)\big), \end{equation} where $a$ is even and $b$ is odd, and where \begin{multline*} M_{ij}(m_1, m_2, n_1, n_2, a, b) \\ = \sum_{l=1} ^{\fl{(a + b - 1)/4)}} (-1)^{i + j} \(\binom {n_1}{ (b - 1)/2 + \fl{(i - 1)/2} - l + 1} \binom {m_1} { -a/2 + \fl{(j - 1)/2} + l} \right.\\ - \left. \binom {n_1} {(b - 1)/2 + \fl{(j - 1)/2} - l + 1} \binom {m_1} {-a/2 + \fl{(i - 1)/2} + l}\) \\ + \sum _{l= 1} ^{\cl{(a + b - 1)/4}} \( \binom {n_2} { (b - 1)/2 + \fl{i/2} - l +1} \binom {m_2} {-a/2 + \fl{j/2} + l - 1} \right. \\ - \left. \binom {n_2} {(b - 1)/2 + \fl{j/2} - l+1} \binom {m_2} {-a/2 + \fl{i/2} + l-1}\). \end{multline*} Remarkably however, experimentally this Pfaffian, first of all, factors completely into factors which are linear in the variables $m_1,m_2,n_1,n_2$, but not only that, there seems to be complete separation, that is, each linear factor contains only one of $m_1,m_2,n_1,n_2$. One has the impression that this phenomenon should have an explanation in a factorisation of the matrix in \eqref{eq:mn}. However, the task of finding one does not seem to be an easy one in view of the ``entangledness" of the parameters in the sums of the matrix entries. \begin{Problem} \label{prob:6} \leavevmode \kern-5pt\footnote{Theresia Eisenk\"olbl has recently solved this problem in \machSeite{EisTAG}\cite{EisTAG}.} Find and prove the closed form evaluation of the Pfaffian in \eqref{eq:mn}. \end{Problem} Our next determinants can be considered as {\it shuffles} of two binomial determinants. Let us first consider \begin{equation} \label{det-ep} \det_{1\le i,j\le a+m} \begin{pmatrix} \dbinom{b+c+m}{b-i+j}& \text {\scriptsize $1\le i\le a$}\\ \dbinom{\frac {b+c} {2}}{\frac {b+a} {2}-i+j+\varepsilon}& \text {\scriptsize $a+1\le i \le a+m$} \end{pmatrix}. \end{equation} In fact, if $\varepsilon=0$, and if $a,b,c$ all have the same parity, then this is exactly the determinant in \eqref{mat1}, the evaluation of which proves Theorem~\ref{enum}, as we explained in Section~\ref{sec:comb}. If $\varepsilon=1/2$ and $a$ has parity different from that of $b$ and $c$, then the corresponding determinant was also evaluated in \machSeite{CiEKAA}\cite{CiEKAA}, and this evaluation implied the companion result to Theorem~\ref{enum} that we mentioned immediately after the statement of the theorem. In the last section of \machSeite{CiEKAA}\cite{CiEKAA}, it is reported that, apparently, there are also nice closed forms for the determinant in \eqref{det-ep} for $\varepsilon=1$ and $\varepsilon=3/2$, both of which imply as well enumeration theorems for {\it rhombus tilings of a hexagon with an equilateral triangle removed from its interior} (see Conjectures~1 and 2 in \machSeite{CiEKAA}\cite{CiEKAA}). We reproduce the conjecture for $\varepsilon=1$ here, the one for $\varepsilon=3/2$ is very similar in form. \begin{Conjecture} \label{conj:loch1} Let $a,b,c,m$ be non-negative integers, $a,b,c$ having the same parity. Then for $\varepsilon=1$ the determinant in \eqref{det-ep} is equal to \begin{multline} \label{eq:1-step} \frac {1} {4}\frac {\h(a + m)\h(b + m)\h(c + m)\h(a + b + c + m) } {\h(a + b + m)\h(a + c + m)\h(b + c + m) }\\ \times \frac {\h(m + \left \lceil {\frac{a + b + c}{2}} \right \rceil) \h(m + \left \lfloor {\frac{a + b + c}{2}} \right \rfloor) } {\h({\frac{a + b}{2}} + m+1) \h({\frac{a + c}{2}} + m-1)\h({\frac{b + c}{2}} + m) } \\ \times\frac {\h(\left \lceil {\frac{a}{2}} \right \rceil) \h(\left \lceil {\frac{b}{2}} \right \rceil) \h(\left \lceil {\frac{c}{2}} \right \rceil) \h(\left \lfloor {\frac{a}{2}} \right \rfloor)\, \h(\left \lfloor {\frac{b}{2}} \right \rfloor)\, \h(\left \lfloor {\frac{c}{2}} \right \rfloor)\, } {\h({\frac{m}{2}} + \left \lceil {\frac{a}{2}} \right \rceil)\, \h({\frac{m}{2}} + \left \lceil {\frac{b}{2}} \right \rceil)\, \h({\frac{m}{2}} + \left \lceil {\frac{c}{2}} \right \rceil)\, \h({\frac{m}{2}} + \left \lfloor {\frac{a}{2}} \right \rfloor)\, \h({\frac{m}{2}} + \left \lfloor {\frac{b}{2}} \right \rfloor)\, \h({\frac{m}{2}} + \left \lfloor {\frac{c}{2}} \right \rfloor)\, }\\ \times \frac {\h(\frac{m}{2})^2 \h({\frac{a + b + m}{2}})^2 \h({\frac{a + c + m}{2}})^2 \h({\frac{b + c + m}{2}})^2 } {\h({\frac{m}{2}} + \left \lceil {\frac{a + b + c}{2}} \right \rceil) \h({\frac{m}{2}} + \left \lfloor {\frac{a + b + c}{2}} \right \rfloor) \h({\frac{a + b}{2}}-1)\h({\frac{a + c}{2}}+1)\h({\frac{b + c}{2}}) }P_1(a,b,c,m), \end{multline} where $P_1(a,b,c,m)$ is the polynomial given by $$P_1(a,b,c,m)=\begin{cases} (a+b)(a+c)+2am&\text {if $a$ is even,}\\ (a+b)(a+c)+2(a+b+c+m)m&\text {if $a$ is odd,}\end{cases}$$ and where $\h(n)$ is the hyperfactorial defined in \eqref{eq:hyperfac}. \end{Conjecture} Two other examples of determinants in which the upper part is given by one binomial matrix, while the lower part is given by a different one, arose in \machSeite{CiKrAC}\cite[Conjectures~A.1 and A.2]{CiKrAC}. Again, both of them seem to factor completely into linear factors, and both of them imply enumeration results for {\it rhombus tilings of a certain V-shaped region}. The right-hand sides of the (conjectured) results are the weirdest ``closed" forms in enumeration that I am aware of.\footnote{No non-trivial simplifications seem to be possible.} We state just the first of the two conjectures, the other is very similar. \begin{Conjecture} Let $x,y,m$ be non-negative integers. Then the determinant \begin{equation} \label{eq:A.2} \det_{1\le i,j\le m+y}\left(\left\{\begin{matrix} \binom {x+i}{x-i+j}&i=1,\dots,m\hfill\\ \binom{x+2m-i+1}{m+y-2i+j+1}&i=m+1,\dots,m+y \end{matrix}\right\}\right). \end{equation} is equal to \begin{multline} \label{eq:A.1} \prod _{i=1} ^{m}\frac {(x+i)!} {(x-i+m+y+1)!\,(2i-1)!} \prod _{i=m+1} ^{m+y}\frac {(x+2m-i+1)!} {(2m+2y-2i+1)!\,(m+x-y+i-1)!}\\ \times {2^{\binom {m}2 + \binom y2 }} \prod_{i = 1}^{m-1}i! \prod_{i = 1}^{y-1}i! \prod_{i \ge 0}^{} ({ \textstyle x+i+{\frac{3}{2}} }) _{m-2i-1} \prod_{i \ge 0}^{} ({ \textstyle x - y+{\frac{5}{2}} + 3 i}) _{ \left \lfloor {\frac{3 y}{2}} -{\frac{9 i}{2}} \right \rfloor-2}\\ \times \prod_{i \ge 0}^{} ({ \textstyle x + {\frac{3 m}{2}} - y + \left \lceil {\frac{3 i}{2}} \right \rceil+\frac {3} {2}}) _{ 3 \left \lceil {\frac{y}{2}} \right \rceil - \left \lceil {\frac{9 i}{2}} \right \rceil -2} \prod_{i \ge 0} ^{} ({ \textstyle { x+ {\frac{3 m}{2}} - y + \left \lfloor {\frac{3 i}{2}} \right \rfloor+2}}) _{ 3 \left \lfloor {\frac{y}{2}} \right \rfloor - \left \lfloor {\frac{9 i}{2}} \right \rfloor-1}\\ \times \prod_{i \ge 0}^{} ({ \textstyle x+m - \left \lfloor {\frac{y}{2}} \right \rfloor}+i+1) _{ 2 \left \lfloor {\frac{y}{2}} \right \rfloor-m - 2 i } \prod_{i \ge 0} ^{} ({ \textstyle x + \left \lfloor {\frac{y}{2}} \right \rfloor+i+2}) _{m - 2 \left \lfloor {\frac{y}{2}} \right \rfloor-2i-2} \\\times {\frac{ \displaystyle \prod_{i = 0}^{y} ({ \textstyle x - y+3i+1}) _{m + 2 y-4i} \prod_{i = 0}^{ \left \lceil {\frac{y}{2}} \right \rceil-1} ({ \textstyle x+m - y+i+1}) _{3 y-m-4i} } {\displaystyle \prod_{i \ge 0}^{} ({ \textstyle x+ {\frac{m}{2}} - {\frac{y}{2}}+i+1}) _{y-2i}\, ({ \textstyle x + {\frac{m}{2}}- {\frac{y}{2}}+i+{\frac{3}{2}}}) _{y-2i-1} }}\\ \times\frac {\displaystyle \prod_{i = 0}^{y} ({ \textstyle x+i+2 }) _{2m - 2 i - 1} } { ({ \textstyle x + y+2}) _{ m - y-1} \,(m+x-y+1)_{m+y} }. \end{multline} Here, shifted factorials occur with positive as well as with negative indices. The convention with respect to which these have to be interpreted is $$(\alpha)_k:=\begin{cases} \alpha(\alpha+1)\cdots(\alpha+k-1)&\text {if }k>0,\\ 1&\text {if }k=0,\\ 1/(\alpha-1)(\alpha-2)\cdots(\alpha+k)&\text {if }k<0. \end{cases}$$ All products $\prod _{i\ge0} ^{}(f(i))_{g(i)}$ in \eqref{eq:A.1} have to interpreted as the products over all $i\ge0$ for which $g(i)\ge0$. \end{Conjecture} For further conjectures of determinants of shuffles of two binomial matrices I refer the reader to Conjectures~1--3 in Section~4 of \machSeite{FuKrAC}\cite{FuKrAC}. All of them imply also enumeration results for rhombus tilings of hexagons. This time, these would be results about the number of {\it rhombus tilings of a symmetric hexagon with some fixed rhombi on the symmetry axis}. \subsection{Determinants of matrices with recursive entries} Binomial coefficients $\binom {i+j}i$ satisfy the basic recurrence of the Pascal triangle, \begin{equation} \label{eq:Pascal} p_{i,j}=p_{i,j-1}+p_{i-1,j}. \end{equation} We have seen many determinants of matrices with entries containing binomial coefficients in the preceding subsection and in \machSeite{KratBN}\cite[Sec.~3]{KratBN}. In \machSeite{BacRAA}\cite{BacRAA}, Bacher reports an experimental study of determinants of matrices $(p_{i,j})_{0\le i,j\le n-1}$, where the coefficients $p_{i,j}$ satisfy the recurrence \eqref{eq:Pascal} (and sometimes more general recurrences), but where the initial conditions for $p_{i,0}$ and $p_{0,i}$, $i\ge 0$, are different from the ones for binomial coefficients. He makes many interesting observations. The most intriguing one says that these determinants satisfy also a linear recurrence (albeit a much longer one). It is intriguing because it points towards the possibility of {\it automatising determinant evaluations}\footnote{The reader should recall that the successful automatisation \machSeite{PeWZAA}% \machSeite{WegsAA}% \machSeite{WiZeAC}% \machSeite{ZeilAM}% \machSeite{ZeilAV}% \cite{PeWZAA,WegsAA,WiZeAC,ZeilAM,ZeilAV} of the evaluation of binomial and hypergeometric sums is fundamentally based on producing recurrences by the computer.}, something that several authors (cf.\ e.g.\ \machSeite{AmZeAB}% \machSeite{KratBN}% \machSeite{PeWiAA}% \cite{AmZeAB,KratBN,PeWiAA}) have been aiming at (albeit, with only limited success up to now). The conjecture (and, in fact, a generalisation thereof) has been proved by Petkov\v sek and Zakraj\v sek in \machSeite{ZaPeAA}\cite{ZaPeAA}. Still, there remains a large gap to fill until computers will replace humans doing determinant evaluations. The paper \machSeite{BacRAA}\cite{BacRAA} contained as well several pretty conjectures on closed form evaluations of special cases of such determinants. These were subsequently proved in \machSeite{KratBU}\cite{KratBU}. We state three of them in the following three theorems. The first two are proved in \machSeite{KratBU}\cite{KratBU} by working out the LU-factorisation (see ``Method~1" in Section~\ref{sec:eval}) for the matrices of which the determinant is computed. The third one is derived by simple row and column operations. \begin{Theorem} \label{thm:Pasc1} Let $(a_{i,j})_{i,j\ge0}$ be the sequence given by the recurrence $$a_{i,j}=a_{i-1,j}+a_{i,j-1}+x\,a_{i-1,j-1},\quad \quad i,j\ge1, $$ and the initial conditions $a_{i,0}=\rho^i$ and $a_{0,i}=\sigma^i$, $i\ge0$. Then \begin{equation} \label{eq:Pasc1} \det_{0\le i,j\le n-1}(a_{i,j})=(1+x)^{\binom {n-1}2}(x+\rho+\sigma-\rho\sigma)^{n-1}. \end{equation} \quad \quad \qed \end{Theorem} \begin{Theorem} \label{thm:Pasc2} Let $(a_{i,j})_{i,j\ge0}$ be the sequence given by the recurrence $$a_{i,j}=a_{i-1,j}+a_{i,j-1}+x\,a_{i-1,j-1},\quad \quad i,j\ge1, $$ and the initial conditions $a_{i,i}=0$, $i\ge0$, $a_{i,0}=\rho^{i-1}$ and $a_{0,i}=-\rho^{i-1}$, $i\ge1$. Then \begin{equation} \label{eq:Pasc2} \det_{0\le i,j\le 2n-1}(a_{i,j})=(1+x)^{2(n-1)^2} (x+\rho)^{2n-2}. \end{equation} \quad \quad \qed \end{Theorem} \begin{Theorem}\label{thm:Pasc3} Let $(a_{i,j})_{i,j\ge0}$ be the sequence given by the recurrence $$a_{i,j}=a_{i-1,j}+a_{i,j-1}+x\,a_{i-1,j-1},\quad \quad i,j\ge1, $$ and the initial conditions $a_{i,0}=i$ and $a_{0,i}=-i$, $i\ge0$. Then \begin{equation} \label{eq:Pasc3} \det_{0\le i,j\le 2n-1}(a_{i,j})=(1+x)^{2n(n-1)}. \end{equation} \quad \quad \qed \end{Theorem} Certainly, the proofs in \machSeite{KratBU}\cite{KratBU} are not very illuminating. Neuwirth \machSeite{NeuwAE}\cite{NeuwAE} has looked more carefully into the structure of recursive sequences of the type as those in Theorems~\ref{thm:Pasc1}--\ref{thm:Pasc3}. Even more generally, he looks at sequences $(f_{i,j})_{i,j\ge0}$ satisfying the recurrence relation \begin{equation} \label{eq:rec} f_{i,j}=c_{j}f_{i-1,j}+d_jf_{i,j-1}+e_jf_{i-1,j-1},\quad i,j\ge1, \end{equation} for some given sequences $(c_j)_{j\ge1}$, $(d_j)_{j\ge1}$, $(e_j)_{j\ge1}$. He approaches the problem by finding appropriate {\it matrix decompositions} for the (infinite) matrix $(f_{i,j})_{i,j\ge0}$. In two special cases, he is able to apply his decomposition results to work out the LU-factorisation of the matrix $(f_{i,j})_{i,j\ge0}$ explicitly, which then yields an elegant determinant evaluation in both of these cases. Neuwirth's first result \machSeite{NeuwAE}\cite[Theorem~5]{NeuwAE} addresses the case where the initial values $f_{0,j}$ satisfy a first order recurrence determined by the coefficients $d_j$ from \eqref{eq:rec}. It generalises Theorem~\ref{thm:Pasc1}. There is no restriction on the initial values $f_{i,0}$ for $i\ge1$. \begin{Theorem} \label{thm:Neuw1} Let $(c_j)_{j\ge1}$, $(d_j)_{j\ge1}$ and $(e_j)_{j\ge1}$ be given sequences, and let $(f_{i,j})_{i,j\ge0}$ be the doubly indexed sequence given by the recurrence \eqref{eq:rec} and the initial conditions $f_{0,0}=1$ and $f_{0,j}=d_jf_{0,j-1}$, $j\ge1$. Then \begin{equation} \label{eq:Neuw1} \det_{0\le i,j\le n-1}(f_{i,j})= \prod _{0\le i<j\le n-1} ^{}(e_{i+1}+c_jd_{i+1}). \end{equation} \quad \quad \qed \end{Theorem} Neuwirth's second result \machSeite{NeuwAE}\cite[Theorem~6]{NeuwAE} also generalises Theorem~\ref{thm:Pasc1}, but in a different way. This time, the initial values $f_{0,j}$, $j\ge1$, are free, whereas the initial values $f_{i,0}$ satisfy a first order recurrence determined by the coefficients $c_j$ from the recurrence \eqref{eq:rec}. Below, we state its most attractive special case, in which all the $c_j$'s are identical. \begin{Theorem} \label{thm:Neuw2} Let $(d_j)_{j\ge1}$ and $(e_j)_{j\ge1}$ be given sequences, and let $(f_{i,j})_{i,j\ge0}$ be the doubly indexed sequence given by the recurrence \eqref{eq:rec} and the initial conditions $f_{0,0}=1$ and $f_{i,0}=cf_{i-1,0}$, $i\ge1$. Then \begin{equation} \label{eq:Neuw2} \det_{0\le i,j\le n-1}(f_{i,j})= \prod _{i=1} ^{n-1}(cd_{i}+e_{i})^{n-i}. \end{equation} \quad \quad \qed \end{Theorem} \subsection{Determinants for signed permutations} \label{sec:signed} The next class of determinants that we consider are determinants of matrices in which rows and columns are indexed by {\it elements of reflection groups} (the latter being groups generated by reflections of hyperplanes in real $n$-dimensional space; see \machSeite{HumpAC}\cite{HumpAC} for more information on these groups, and, more generally, on Coxeter groups). The prototypical example of a reflection group is the {\it symmetric group} $\mathfrak S_n$ of permutations of an $n$-element set. In \machSeite{KratBN}\cite{KratBN}, there appeared two determinant evaluations associated to the symmetric group, see Theorems~55 and 56 in \machSeite{KratBN}\cite{KratBN}. They concerned evaluations of determinants of the type \begin{equation} \label{eq:stat} \det_{\sigma,\pi\in \mathfrak S_n}\(q^{\operatorname{stat}(\sigma\pi^{-1})}\), \end{equation} due to Varchenko, Zagier, and Thibon, respectively, in which stat is the statistic {\it ``number of inversions,"} respectively {\it``major index."} We know that in many fields of mathematics there exist certain diseases which are typical for that field. Algebraic combinatorics is no exception. Here, I am {\it not\/} talking of the earlier mentioned ``$q$-disease" (see Footnote~\ref{foot:q}; although, due to the presence of $q$ we might also count it as a case of $q$-disease), but of the disease which manifests itself by the question ``And what about the other types?"\footnote{\label{foot:root}% In order to give a reader who is not acquainted with the language and theory of reflection groups an idea what this question is referring to, I mention that all finite reflection groups have been classified, each having been assigned a certain ``type." So, usually one proves something for the symmetric group $\mathfrak S_n$, which, according to this classification, has type $A_{n-1}$, and then somebody (which could be oneself) will ask the question ``Can you also do this for the other types?", meaning whether or not there exists an analogous result for the other finite reflection groups.} So let us ask this question, that is, are there theorems similar to the two theorems which we mentioned above for other reflection groups? So, first of all we need analogues of the statistics ``number of inversions" and ``major index" for other reflection groups. Indeed, these are available in the literature. The analogue of ``number of inversions" is the so-called {\it length} of an element in a Coxeter group (see \machSeite{HumpAC}\cite{HumpAC} for the definition). As a matter of fact, a closed form evaluation of the determinant \eqref{eq:stat}, where $\mathfrak S_n$ is replaced by any finite or affine reflection group, and where stat is the length, is known (and was already implicitly mentioned in \machSeite{KratBN}\cite{KratBN}). This result is due to Varchenko \machSeite{VarcAC}\cite[Theorem~(1.1), where $a(H)$ is specialised to $q$]{VarcAC}. His result is actually much more general, as it is valid for real hyperplane arrangements in which each hyperplane is assigned a different weight. I will not state it here explicitly because I do not want to go through the definitions and notations which would be necessary for doing that. So, what about analogues of the ``major index" for other reflection groups? These are also available, and there are in fact several of them. The first person to introduce a major index for reflection groups other than the symmetric groups was Reiner in \machSeite{ReivAC}\cite{ReivAC}. He proposed a major index for the {\it hyperoctahedral group} $B_n$, which arose naturally in his study of {\it $P$-partitions for signed posets}. The elements of $B_n$ are often called {\it signed permutations}, and they are all elements of the form $\pi_1\pi_2\dots\pi_n$, where $\pi_i\in\{\pm1,\pm2,\dots,\pm n\}$, $i=1,2,\dots,n$, and where $\vert\pi_1\vert\vert\pi_2\vert\dots\vert\pi_n\vert$ is a permutation in $\mathfrak S_n$. To define their multiplicative structure, it is most convenient to view $\pi=\pi_1\pi_2\dots\pi_n$ as a linear operator on ${\mathbb R}^n$ acting by permutation and sign changes of the co-ordinates. To be precise, the action is given by $\pi(e_i)=(\operatorname{sgn}\pi_i)e_{\vert\pi_i\vert}$, where $e_i$ is the $i$-th standard basis vector in ${\mathbb R}^n$, $i=1,2,\dots,n$. The multiplication of two signed permutations is then simply the composition of the corresponding linear operators. The major index $\operatorname{maj}_B\pi$ of an element $\pi\in B_n$ which Reiner defined is, as in the symmetric group case, the sum of all positions of {\it descents} in $\pi$. (There is a natural notion of ``descent" for any Coxeter group.) Concretely, it is $$\operatorname{maj}_B\pi:=\chi(\pi_n<0)+ \sum _{i=1} ^{n-1}i\cdot\chi(\pi_i>_B\pi_{i+1}),$$ where we impose the order $1<_B2<_B\cdots<_Bn<_B-n<_B\cdots<_B-2<_B-1$ on our ground set $\{\pm1,\pm2,\dots,\pm n\}$, and where $\chi(\mathcal A)=1$ if $\mathcal A$ is true and $\chi(\mathcal A)=0$ otherwise. There is overwhelming computational evidence\footnote{\label{foot:maj}The reader may wonder what this computational evidence could actually be. After all, we are talking about a determinant of a matrix of size $2^nn!$. More concretely, for $n=1,2,3,4,5$ these are matrices of size $2$, $8$, $48$, $384$, $3840$, respectively. While {\sl Maple} or {\sl Mathematica} have no problem to compute these determinants for $n=1$ and $n=2$, it takes already considerable time to do the computation for $n=3$, and it is, of course, completely hopeless to let them compute the one for $n=4$, a determinant of a matrix of size $384$ which has polynomial entries (cf.\ Footnote~\ref{foot:kompl}). However, the results for $n=1,2,3$ already ``show" that the determinant will factor completely into factors of the form $1-q^i$, $i=1,2,\dots,2n$. One starts to expect the same to be true for higher $n$. To get a formula for $n=4$, one would then apply the tricks explained in Footnote~\ref{foot:tricks}. That is, one specialises $q$ to $4$, at which value the first $8$ cyclotomic polynomials (in fact, even more) are clearly distinguishable by their prime factorisations, and one computes the determinant. The exponents of the various factors $1-q^i$ can then be extracted from the exponents of the prime factors in the prime factorisation of the determinant with $q=4$. Unfortunately, the data collected for $n=1,2,3,4$ do not suffice to come up with a guess, and, on the other hand, {\sl Maple} and {\sl Mathematica} will certainly be incapable to compute a determinant of a matrix of size $3840$ (which, just to store it on the disk, occupies already 10 megabytes \dots). So then, what did I mean when I said that the conjecture is based on data including $n=5$? This turned out to become a ``test case" for {\sl LinBox}, a C++ template library for exact high-performance linear algebra \machSeite{DGGGAA}\cite{DGGGAA}, which is freely available under {\tt http://linalg.org}. To be honest, I was helped by Dave Saunders and Zhendong Wan (two of the developers) who applied {\sl LinBox} to do rank and Smith normal form computations for the specialised matrix with respect to various prime powers (each of which taking several hours). The specific computational approach that worked here is quite recent (thus, it came just in time for our purpose), and is documented in \machSeite{SaWaAA}\cite{SaWaAA}. The results of the computations made it possible to come up with a ``sure" prediction for the exponents with which the various prime factors occur in the prime factorisation of the specialised determinant. As in the case $n=4$, the exponents of the various factors $1-q^i$, $i=1,2,\dots,2n$ can then easily be extracted. (The guesses were subsequently also tested with special values of $q$ other than $q=4$.)} that the ``major-de\-ter\-min\-ant" for $B_n$, i.e., the determinant \eqref{eq:stat} with $\text{stat}=\operatorname{maj}_B$ and with $\mathfrak S_n$ replaced by $B_n$, factors completely into cyclotomic polynomials. \begin{Conjecture} \label{prob:1} For any positive integer $n$, we have \begin{equation} \label{eq:maj-Bn} \det_{\sigma,\pi\in B_n}\(q^{\operatorname{maj}_B(\sigma\pi^{-1})}\)= \prod_{i=1}^n (1-q^{2i})^{2^{n-1}n!/i} \prod_{i=2}^n (1-q^i)^{2^n n!(i-1)/i}. \end{equation} \end{Conjecture} A different major index for $B_n$ was proposed by Adin and Roichman in \machSeite{AdRoAC}\cite{AdRoAC}. It arises there naturally in a combinatorial study of {\it polynomial algebras which are diagonally invariant under $B_n$}. (In fact, more generally, {\it wreath products} of the form $C_m\wr \mathfrak S_n$, where $C_m$ is the cyclic group of order $m$, and their diagonal actions on polynomial algebras are studied in \machSeite{AdRoAC}\cite{AdRoAC}. These groups are also sometimes called {\it generalised reflection groups}. In this context, $B_n$ is the special case $C_2\wr\mathfrak S_n$.) If we write $\operatorname{neg}\pi$ for the number of $i$ for which $\pi_i$ is negative, then the {\it flag-major index} fmaj of Adin and Roichman is defined by \begin{equation} \label{eq:fmaj-def} \operatorname{fmaj}\pi:=2\operatorname{maj}_A\pi+\operatorname{neg}\pi, \end{equation} where $\operatorname{maj}_A$ is the ``ordinary" major index due to MacMahon, $$\operatorname{maj}_A\pi:=\sum _{i=1} ^{n-1}i\cdot\chi(\pi_i>\pi_{i+1}).$$ If one now goes to the computer and calculates the determinant on the left-hand side of \eqref{eq:maj-Bn} with $\operatorname{maj}_B$ replaced by $\operatorname{fmaj}$ for $n=1,2,3,4,5$ (see Footnote~\ref{foot:maj} for the precise meaning of ``calculating the determinant for $n=1,2,3,4,5$"), then again the results factor completely into cyclotomic factors. Even more generally, it seems that one can treat the two parts on the right-hand side of \eqref{eq:fmaj-def}, that is ``major index" and ``number of negative letters," separately. \begin{Conjecture} \label{prob:2} For any positive integer $n$, we have \begin{equation} \label{eq:fmaj-Bn} \det_{\sigma,\pi\in B_n}\(q^{\operatorname{maj}_A(\sigma\pi^{-1})}p^{\operatorname{neg}(\sigma\pi^{-1})}\)= \prod _{i=1} ^{n}(1-p^{2i})^{2^{n-1}n!/i} \prod _{i=2} ^{n}(1-q^i)^{2^{n}n!(i-1)/i}. \end{equation} \end{Conjecture} I should remark that Adin and Roichman have shown in \machSeite{AdRoAC}\cite{AdRoAC} that the statistics fmaj is equidistributed with the statistics length on $B_n$. However, even in the case where we just look at the flag-major determinant (that is, the case where $q=p^2$ in Conjecture~\ref{prob:2}), this does not seem to help. (Neither length nor flag-major index satisfy a simple law with respect to multiplication of signed permutations.) In fact, from the data one sees that the flag-major determinants are different from the length determinants (that is, the determinants \eqref{eq:stat}, where $\mathfrak S_n$ is replaced by $B_n$ and stat is flag-major, respectively length). Initially, I had my program wrong, and, instead of taking the (ordinary) major index $\operatorname{maj}_A$ of the signed permutation $\pi=\pi_1\pi_2\dots\pi_n$ in \eqref{eq:fmaj-def}, I computed taking the major index of the {\it absolute value} of $\pi$. This absolute value is obtained by forgetting all signs of the letters of $\pi$, that is, writing $\vert\pi\vert$ for the absolute value of $\pi$, $\vert\pi\vert=\vert\pi_1\vert\,\vert\pi_2\vert\,\dots\,\vert\pi_n\vert$. Curiously, it seems that also this ``wrong" determinant factors nicely. (Again, the evidence for this conjecture is based on data which were obtained in the way described in Footnote~\ref{foot:maj}.) \begin{Conjecture} \label{prob:3} For any positive integer $n$, we have \begin{equation} \label{eq:fabsmaj-Bn} \det_{\sigma,\pi\in B_n}\(q^{\operatorname{maj}\vert\sigma\pi^{-1}\vert}p^{\operatorname{neg}(\sigma\pi^{-1})}\)= (1-p^{2})^{2^{n-1}n\cdot n!} \prod _{i=2} ^{n}(1-q^i)^{2^{n}n!(i-1)/i}. \end{equation} \end{Conjecture} Since, as I indicated earlier, Adin and Roichman actually define a flag-major index for wreath products $C_m\wr\mathfrak S_n$, a question that suggests itself is whether or not we can expect closed product formulae for the corresponding determinants. Clearly, since we are now dealing with determinants of the size $m^nn!$, computer computations will exhaust our computer's resources even faster if $m>2$. The calculations that I was able to do suggest strongly that there is indeed an extension of the statement in Conjecture~\ref{prob:2} to the case of arbitrary $m$, if one uses the definition of major index and the ``negative" statistics for $C_m\wr\mathfrak S_n$ as in \machSeite{AdRoAC}\cite{AdRoAC}. (See \machSeite{AdRoAC}\cite[Section~3]{AdRoAC} for the definition of the major index. The sum on the right-hand side of \machSeite{AdRoAC}\cite[(3.1)]{AdRoAC} must be taken as the extension of the ``negative" statistics neg to $C_m\wr\mathfrak S_n$.) \begin{Problem} \label{prob:4} Find and prove the closed form evaluation of \begin{equation} \label{eq:maj-wreath} \det_{\sigma,\pi\in C_m\wr \mathfrak S_n}\(q^{\operatorname{maj}(\sigma\pi^{-1})} p^{\operatorname{neg}(\sigma\pi^{-1})}\), \end{equation} where $\operatorname{maj}$ and $\operatorname{neg}$ are the extensions to $C_m\wr \mathfrak S_n$ of the statistics $\operatorname{maj}_A$ and $\operatorname{neg}$ in Conjecture~{\em\ref{prob:2}}, as described in the paragaph above. \end{Problem} Together with Brenti, Adin and Roichman proposed another major statistics for signed permutations in \machSeite{AdBRAA}\cite{AdBRAA}. They call it the {\it negative major index}, denoted nmaj, and it is defined as the sum of the ordinary major index and the sum of the absolute values of the negative letters, that is, $$\operatorname{nmaj}\pi:=\operatorname{maj}_A\pi+\operatorname{sneg}\pi,$$ where $\operatorname{sneg}\pi:=-\sum _{i=1} ^{n}\chi(\pi_i<0)\pi_i$. Also for this statistics, the corresponding determinant seems to factor nicely. In fact, it seems that one can again treat the two components of the definition of the statistics, that is, ``major index" and ``sum of negative letters," separately. (Once more, the evidence for this conjecture is based on data which were obtained in the way described in Footnote~\ref{foot:maj}.) \begin{Conjecture} \label{prob:8} For any positive integer $n$, we have \begin{equation} \label{eq:nmaj-Bn} \det_{\sigma,\pi\in B_n}\(q^{\operatorname{maj}_A(\sigma\pi^{-1})}p^{\operatorname{sneg}(\sigma\pi^{-1})}\)= \prod _{i=1} ^{n}(1-p^{2i^2})^{2^{n-1}n!/i} \prod _{i=2} ^{n}(1-q^i)^{2^{n}n!(i-1)/i}. \end{equation} \end{Conjecture} If one compares the (conjectured) result with the (conjectured) one for the ``flag-major determinant" in Conjecture~\ref{prob:2}, then one notices the somewhat mind-boggling fact that one obtains the right-hand side of \eqref{eq:nmaj-Bn} from the one of \eqref{eq:fmaj-Bn} by simply replacing (in the factored form of the latter) $1-p^{2i}$ by $1-p^{2i^2}$, everything else, the exponents, the ``$q$-part", is identical. It is difficult to imagine an intrinsic explanation why this should be the case. Since Thibon's proof of the evaluation of the determinant \eqref{eq:stat} with stat being the (ordinary) major index for permutations (see \machSeite{KratBN}\cite[Appendix~C]{KratBN}) involved the {\em descent algebra} of the symmetric group, viewed in terms of {\em non-commutative symmetric functions}, one might speculate that to prove Conjectures~\ref{prob:1}--\ref{prob:3} and \ref{prob:8} it may be necessary to work with $B_n$ versions of descent algebras (which exist, see \machSeite{SoloAA}\cite{SoloAA}) and non-commutative symmetric functions (which also exist, see \machSeite{ChoCAA}\cite{ChoCAA}). Adriano Garsia points out that all the determinants in \eqref{eq:stat}, \eqref{eq:maj-Bn}, \eqref{eq:fmaj-Bn}--\eqref{eq:nmaj-Bn} are special instances of {\it group determinants}. (See the excellent survey article \machSeite{LamTAA}\cite{LamTAA} for information on group determinants.) The main theorem on group determinants, due to Frobenius, says that a general group determinant factorises into irreducible factors, each of which corresponding to an irreducible representation of the group, and the exact exponent to which the irreducible factor is raised is the degree of the corresponding irreducible representation. In view of this, Garsia poses the following problem, a solution of which would refine the above conjectures, Problem~\ref{prob:4}, and the earlier mentioned results of Varchenko, Zagier, and Thibon. \begin{Problem} \label{prob:Frob} For each of the above special group determinants, determine the closed formula for the value of the irreducible factor corresponding to a fixed irreducible representation of the involved group {\em(}$\mathfrak S_n$, $B_n$, $C_m\wr \mathfrak S_n$, respectively{\em)}. \end{Problem} It seems that a solution to this problem has not even been worked out for the determinant which is the subject of the results of Varchenko and Zagier, that is, for the determinant \eqref{eq:stat} with stat being the number of inversions. For further work on statistics for (generalised) reflection groups (thus providing further prospective candidates for forming interesting determinants), I refer the reader to \machSeite{AdBRAA}% \machSeite{AdBRAB}% \machSeite{BagnAA}% \machSeite{BernAA}% \machSeite{BernAB}% \machSeite{BiagAA}% \machSeite{BiCaAA}% \machSeite{FoHaAM}% \machSeite{FoHaAN}% \machSeite{FoHaAO}% \machSeite{HaLRAA}% \machSeite{ReRoAA}% \cite{AdBRAA,AdBRAB,BagnAA,BernAA,BernAB,BiagAA,BiCaAA,FoHaAM,FoHaAN,FoHaAO,HaLRAA,ReRoAA}. I must report that, somewhat disappointingly, it seems that the various major indices proposed for the group $D_n$ of even signed permutations (see \machSeite{BiagAA}% \machSeite{BiCaAA}% \machSeite{ReivAC}% \cite{BiagAA,BiCaAA,ReivAC}) apparently do not give rise to determinants in the same way as above that have nice product formulae. This remark seems to also apply to determinants formed in an anlogous way by using the various statistics proposed for the alternating group in \machSeite{ReRoAB}\cite{ReRoAB}. \subsection{More poset and lattice determinants} \label{sec:poset} Continuing the discussion of determinants which arise under the influence of the above-mentioned ``reflection group disease," we turn our attention to two miraculous determinants which were among the last things Rodica Simion was able to look at. Some of her considerations in this direction are reported in \machSeite{SchFAA}\cite{SchFAA}. The first of the two is a determinant of a matrix the rows and columns of which are indexed by {\it type $B$ non-crossing partitions}. This determinant is inspired by the evaluation of an analogous one for {\it ordinary} non-crossing partitions (that is, in ``reflection group language," {\it type $A$ non-crossing partitions}), due to Dahab \machSeite{DahaAA}\cite{DahaAA} and Tutte \machSeite{TuttAC}\cite{TuttAC} (see \machSeite{KratBN}\cite[Theorem~57, (3.69)]{KratBN}). Recall (see \machSeite{StanAP}\cite[Ch.~1 and 3]{StanAP} for more information) that a {\it {\rm(}set{\rm)} partition} of a set $S$ is a collection $\{B_1,B_2,\dots,B_k\}$ of pairwise disjoint non-empty subsets of $S$ such that their union is equal to $S$. The subsets $B_i$ are also called {\it blocks} of the partition. One partially orders partitions by refinement. With respect to this partial order, the partitions form a lattice. We write $\pi\lor_{A}\gamma$ (the $A$ stands for the fact that, in ``reflection group language", we are looking at ``type $A$ partitions") for the join of $\pi$ and $\gamma$ in this lattice. Roughly speaking, the join of $\pi$ and $\gamma$ is formed by considering altogether all the blocks of $\pi$ and $\gamma$. Subsequently, whenever we find two blocks which have a non-empty intersection, we merge them into a bigger block, and we keep doing this until all the (merged) blocks are pairwise disjoint. If $S=\{1,2,\dots,n\}$, we call a partition {\it non-crossing} if for any $i<j<k<l$ the elements $i$ and $k$ are in the same block at the same time as the elements $j$ and $l$ are in the same block only if these two blocks are the same. (I refer the reader to \machSeite{SimiAD}\cite{SimiAD} for a survey on non-crossing partitions.) Reiner \machSeite{ReivAD}\cite{ReivAD} introduced non-crossing partitions in type $B$. Partitions of type $B_n$ are (ordinary) partitions of $\{1,2,\dots,n,-1,-2,\dots,-n\}$ with the property that whenever $B$ is a block then so is $-B:=\{-b:b\in B\}$, and that there is at most one block $B$ with $B=-B$. A block $B$ with $B=-B$, if present, is called the {\it zero-block} of the partition. We denote the set of all type $B_n$ partitions by $\Pi_n^B$, the number of zero blocks of a partition $\pi$by $\operatorname{zbk} \pi$, and we write $\operatorname{nzbk}\pi$ for half of the number of the non-zero blocks. Type $B_n$ non-crossing partitions are a subset of type $B_n$ partitions. Imposing the order $1<2<\dots<n<-1<-2<\dots<-n$ on our ground-set, the definition of type $B_n$ non-crossing partitions is identical with the one for type $A$ non-crossing partitions, that is, given this order on the ground-set, a $B_n$ partition is called {\it non-crossing} if for any $i<j<k<l$ the elements $i$ and $k$ are in the same block at the same time as the elements $j$ and $l$ are in the same block only if these two blocks are the same. We write $\operatorname{NC}_n^B$ for the set of all $B_n$ non-crossing partitions. The determinant defined by type $B_n$ non-crossing partitions that Simion tried to evaluate was the one in \eqref{eq:Simion1} below.\footnote{In fact, instead of $\lor_A$, the ``ordinary" join, she used the join in the type $B_n$ partition lattice $\Pi_n^B$. However, the number of {\it non-zero} blocks will be the same regardless of whether we take the join of two type $B_n$ non-crossing partitions with respect to ``ordinary" join or with respect to ``type $B_n$" join. This is in contrast to the numbers of zero blocks, which can differ largely. (To be more precise, one way to form the ``type $B_n$" join is to first form the ``ordinary" join, and then merge all zero blocks into one big block.) The reason that I insist on using $\lor_A$ is that this is crucial for the more general Conjecture~\ref{conj:Simion2}. To tell the truth, the discovery of the latter conjecture is due to a programming error on my behalf (that is, originally I aimed to program the ``type $B$" join, but it happened to be the ``ordinary" join \dots).} The use of the ``type $A$" join $\lor_A$ for two type $B$ non-crossing partitions in \eqref{eq:Simion1} may seem strange. However, this is certainly a well-defined operation. The result may neither be a type $B$ partition nor a non-crossing one, it will just be an ordinary partition of the ground-set $\{1,2,\dots,n,-1,-2,\dots,-n\}$. We extend the notion of ``zero block" and ``non-zero block" to these objects in the obvious way. Simion observed that, as in the case of the analogous type $A_n$ partition determinant due to Dahab and Tutte, it factors apparently completely into factors which are Chebyshev polynomials. Based on some additional numerical calculations,\footnote{\label{foot:Simion}% Evidently, more than five years later, thanks to technical progress since then, one can go much farther when doing computer calculations. The evidence for Conjecture~\ref{conj:Simion1} which I have is based on, similar to the conjectures and calculations on determinants for signed permutations in Subsection~\ref{sec:signed} (see Footnote~\ref{foot:maj}), the exact form of the determinants for $n=1,2,3,4$, which were already computed by Simion, and, essentially, the exact form of the determinants for $n=5$ and $6$. By ``essentially" I mean, as earlier, that I computed the determinant for many special values of $q$, which then let me make a guess on the basis of comparison of the prime factors in the factorised results with the prime factors of the candidate factors, that is the irreducible factors of the Chebyshev polynomials. Finally, for guessing the general form of the exponents, the available data were not sufficient for {\tt Rate} (see Footnote~\ref{foot:Rate}). So I consulted the fabulous {\it On-Line Encyclopedia of Integer Sequences} ({\tt http://www.research.att.com/\~{}njas/sequences/Seis.html}), originally created by Neil Sloane and Simon Plouffe \machSeite{SlPlAA}\cite{SlPlAA}, and since many years continuously further developed by Sloane and his team \machSeite{SloaAA}\cite{SloaAA}. An appropriate selection from the results turned up by the Encyclopedia then led to the exponents on the right-hand side of \eqref{eq:Simion1}.} I propose the following conjecture. \begin{Conjecture} \label{conj:Simion1} For any positive integer $n$, we have \begin{equation} \label{eq:Simion1} \det_{\pi,\gamma\in\operatorname{NC}_n^B}\(q^{\operatorname{nzbk} (\pi\lor_{A}\gamma)}\)= \prod _{i=1} ^{n}\left(\dfrac {U_{3i-1}(\sqrt q/2)} {U_{i-1}(\sqrt q/2)}\)^{\binom {2n}{n-i}}, \end{equation} where $U_{m}(x):=\sum _{j\ge0} ^{}(-1)^j\binom {m-j}j(2x)^{m-2j}$ is the $m$-th {\em Chebyshev polynomial of the second kind}. \end{Conjecture} If proved, this would solve Problem~1 in \machSeite{SchFAA}\cite{SchFAA}. It would also solve Problem~2 from \machSeite{SchFAA}\cite{SchFAA}, because $U_{m-1}(\sqrt q/2)$ is, up to multiplication by a power of $q$, equal to the product $\prod _{j\mid m} ^{}f_j(q)$, where the polynomials $f_j(q)$ are the ones of \machSeite{SchFAA}\cite{SchFAA}. A simple computation then shows that, when the right-hand side product of \eqref{eq:Simion1} is expressed in terms of the $f_j(q)$'s, one obtains \begin{equation} \label{eq:fprod} \prod _{k=1} ^{n}f_{3k}(q)^{e_{n,k}}, \end{equation} where $$e_{n,k}=\underset {\ell\not\equiv 0\,(\text{mod }3)} {\sum _{\ell=1} ^{\fl{n/k}}}\binom {2n}{n-\ell k}.$$ This agrees with the data in \machSeite{SchFAA}\cite{SchFAA} and with the further ones I have computed (see Footnote~\ref{foot:Simion}). Even more seems to be true. The following conjecture predicts the evaluation of the more general determinant where we also keep track of the zero blocks. \begin{Conjecture} \label{conj:Simion2} For any positive integer $n$, we have \begin{equation} \label{eq:Simion2} \det_{\pi,\gamma\in\operatorname{NC}_n^B}\(q^{\operatorname{nzbk} (\pi\lor_{A}\gamma)}z^{\operatorname{zbk} (\pi\lor_{A}\gamma)}\)=z^{\frac {1} {2}\binom {2n}{n}} \prod _{i=1} ^{n}\Big(2T_{2i}(\sqrt q/2)+2-z\Big)^{\binom {2n}{n-i}}, \end{equation} where $T_{m}(x):=\frac {1} {2}\sum _{j\ge0} ^{}(-1)^j\frac {m} {m-j}\binom {m-j}j(2x)^{m-2j}$ is the $m$-th {\em Chebyshev polynomial of the first kind}. \end{Conjecture} Again, the conjecture is supported by extensive numerical calculations. It is not too difficult to show, by using some identities for Chebyshev polynomials, that Conjecture~\ref{conj:Simion2} implies Conjecture~\ref{conj:Simion1}. The other determinant which Simion looked at (cf.\ \machSeite{SchFAA}\cite[Problem~9ff]{SchFAA}), was the $B_n$ analogue of a determinant of a matrix the rows and columns of which are indexed by {\it non-crossing matchings}, due to Lickorish \machSeite{LickAA}\cite{LickAA}, and evaluated by Ko and Smolinsky \machSeite{KoSmAA}\cite{KoSmAA} and independently by Di~Francesco \machSeite{DiFrAA}\cite{DiFrAA} (see \machSeite{KratBN}\cite[Theorem~58]{KratBN}). As we may regard (ordinary) non-crossing matchings as partitions all the blocks of which consist of two elements, we {\it define} a {\it $B_n$ non-crossing matching} to be a $B_n$ non-crossing partition all the blocks of which consist of two elements. We shall be concerned with $B_{2n}$ non-crossing matchings, which we denote by $\operatorname{NCmatch}(2n)$. With this notation, the following seems to be true. \begin{Conjecture} \label{conj:Simion3} For any positive integer $n$, we have \begin{equation} \label{eq:Simion3} \det_{\pi,\gamma\in\operatorname{NCmatch}(2n)}\(q^{\operatorname{nzbk} (\pi\lor_{A}\gamma)}z^{\operatorname{zbk} (\pi\lor_{A}\gamma)}\)= \prod _{i=1} ^{n}\Big(2T_{2i}(q/2)+2-z^2\Big)^{\binom {2n}{n-i}}. \end{equation} \end{Conjecture} The reader should notice the remarkable fact that, in the case that Conjectures~\ref{conj:Simion2} and \ref{conj:Simion3} are true, the right-hand side of \eqref{eq:Simion3} is, up to a power of $z$, equal to the right-hand side of \eqref{eq:Simion2} with $q$ replaced by $q^2$ and $z$ replaced by $z^2$. An intrinsic explanation why this should be the case is not known. An analogous relation between the determinants of Tutte and of Lickorish, respectively, was observed, and proved, in \machSeite{CoSSAA}\cite{CoSSAA}. Also here, no intrinsic explanation is known. The reader is referred to \machSeite{SchFAA}\cite{SchFAA} for further open problems related to the determinants in Conjectures~\ref{conj:Simion1}--\ref{conj:Simion3}. Finally, it may also be worthwhile to look at determinants defined using $D_n$ non-crossing partitions and non-crossing matchings, see \machSeite{AtReAA}\cite{AtReAA} and \machSeite{ReivAD}\cite{ReivAD} for two possible definitions of those. \medskip The reader may have wondered why in Conjectures~\ref{conj:Simion1} and \ref{conj:Simion2} we considered determinants defined by type $B_n$ non-crossing partitions, which form in fact a lattice, but used the extraneous type $A$ join\footnote{and not even the one in the type $A$ {\it non-crossing} partition lattice!} in the definition of the determinant, instead of the join which is intrinsic to the lattice of type $B_n$ non-crossing partitions. In particular, what would happen if we would make the latter choice? As it turns out, for that situation there exists an elegant general theorem due to Lindstr\"om \machSeite{LindAB}\cite{LindAB}, which I missed to state in \machSeite{KratBN}\cite{KratBN}. I refer to \machSeite{StanAP}\cite[Ch.~3]{StanAP} for the explanation of the poset terminology used in the statement. \begin{Theorem} \label{thm:Lind} Let $L$ be a finite meet semilattice, $R$ be a commutative ring, and $f:L\times L\to R$ be an incidence function, that is, $f(x,y)=0$ unless $x\land y=x$. Then \begin{equation} \label{eq:Lind} \det_{x,y\in L}\big(f(x\land y,x)\big)= \prod _{y\in L} ^{} \(\sum _{x\in L} ^{}\mu(x,y)f(x,y) \), \end{equation} where $\mu$ is the M\"obius function of $L$.\quad \quad \qed \end{Theorem} Clearly, this does indeed answer our question, we just have to specialise $f(x,y)=h(x)$ for $x\land y=x$, where $h$ is some function from $L$ to $R$. The fact that the above theorem talks about meets instead of joins is of course no problem because this is just a matter of convention. Having an answer in such a great generality, one is tempted to pose the problem of finding a general theorem that would encompass the above-mentioned determinant evaluations due to Tutte, Dahab, Ko and Smolinsky, Di~Francesco, as well as Conjectures~\ref{conj:Simion1} and \ref{conj:Simion2}. This problem is essentially Problem~6 in \machSeite{SchFAA}\cite{SchFAA}. \begin{Problem} \label{prob:9} Let $L$ and $L'$ be two lattices {\em(}semilattices?{\em)} with $L'\subseteq L$. Furthermore, let $R$ be a commutative ring, and let $f$ be a function from $L$ to $R$. Under which conditions is there a compact formula for the determinant \begin{equation} \label{eq:det9} \det_{x,y\in L'}\big(f(x\land_L y)\big) , \end{equation} where $\land_L$ is the meet operation in $L$? \end{Problem} By specialisation in Theorem~\ref{thm:Lind}, one can derive numerous corollaries. For example, a very attractive one is the evaluation of the ``GCD determinant" due to Smith \machSeite{SmitAA}\cite{SmitAA}. (In fact, Smith's result is a more general one for {\it factor closed subsets} of the positive integers.) \begin{Theorem} \label{thm:Smith} For any positive integer $n$, we have $$\det_{1\le i,j\le n}\big(\gcd(i,j)\big)= \prod _{i=1} ^{n}\phi(i),$$ where $\phi$ denotes the Euler totient function.\quad \quad \qed \end{Theorem} An interesting generalisation of Theorem~\ref{thm:Lind} to posets was given by Altini\c sik, Sagan and Tu\u glu \machSeite{AlSTAA}\cite{AlSTAA}. Again, all undefined terminology can be found in \machSeite{StanAP}\cite[Ch.~3]{StanAP}. \begin{Theorem} \label{thm:AlSaTu} Let $P$ be a finite poset, $R$ be a commutative ring, and $f,g:P\times P\to R$ be two incidence functions, that is, $f(x,y)=0$ unless $x\le y$ in $P$, the same being true for $g$. Then $$\det_{x,y\in P}\( \sum _{z\in P} ^{}f(z,x)g(z,y)\)= \prod _{x\in P} ^{}f(x,x)g(x,x).$$ \quad \quad \qed \end{Theorem} The reader is referred to Section~3 of \machSeite{AlSTAA}\cite{AlSTAA} for the explanation why this theorem implies Lindstr\"om's. \subsection{Determinants for compositions} Our next family of determinants consists of determinants of matrices the rows and columns of which are indexed by {\it compositions}. Recall that a composition of a non-negative integer $n$ is a vector $(\alpha_1,\alpha_2,\dots,\alpha_k)$ of non-negative integers such that $\alpha_1+\alpha_2+\dots+\alpha_k=n$, for some $k$. For a fixed $k$, let $\mathcal C(n,k)$ denote the corresponding set of compositions of $n$. While working on a problem in global optimisation, Brunat and Montes \machSeite{BrMoAA}\cite{BrMoAA} discovered the following surprising determinant evaluation. It allowed them to show how to explicitly express a multivariable polynomial as a {\it difference of convex functions}. In the statement, we use standard multi-index notation: if $\boldsymbol\alpha=(\alpha_1,\alpha_2,\dots,\alpha_k)$ and $\boldsymbol \beta=(\beta_1,\beta_2,\dots,\beta_k)$ are two compositions, we let $$\boldsymbol\alpha^{\boldsymbol \beta}:=\alpha_1^{\beta_1}\alpha_2^{\beta_2}\cdots \alpha_k^{\beta_k}.$$ \begin{Theorem} \label{thm:Brunat} For any positive integers $n$ and $k$, we have \begin{equation} \label{eq:Brunat} \det_{\boldsymbol\alpha,\boldsymbol\beta\in\mathcal C(n,k)}\left(\boldsymbol\alpha^{\boldsymbol \beta}\right)= n^{{\binom{ n + k-1} k}+k-1}\, \prod_{i = 1}^{ n-1} i^{( n -i+1) {\binom {n + k-i-1} { k-2}}}. \end{equation} \quad \quad \qed \end{Theorem} In recent joint work \machSeite{BrKMAA}\machSeite{BrMoAB}\cite{BrKMAA,BrMoAB}, Brunat, Montes and the author showed that there is in fact a polynomial generalisation of this determinant evaluation. \begin{Theorem} \label{conj:Brunat1} Let $\mathbf x=(x_1,x_2,\dots,x_k)$ be a vector of indeterminates, and let $\lambda$ be an indeterminate. Then, for any non-negative integers $n$ and $k$, we have \begin{equation} \label{eq:Brunat1} \det_{\boldsymbol\alpha,\boldsymbol\beta\in\mathcal C(n,k)}\left((\mathbf x+\lambda\boldsymbol\alpha)^ {\boldsymbol \beta}\right)=\lambda^{(k-1)\binom {n+k-1}k} \left({ \vert \mathbf x\vert+\lambda n }\right) ^{\binom{n+k-1} {k}} \prod _{i=1} ^{n}i^{(k-1)\binom {n+k-i-1}{k-1}}, \end{equation} where $\mathbf x+\lambda\boldsymbol\alpha$ is short for $(x_1+\lambda\alpha_1,x_2+\lambda\alpha_2,\dots,x_k+\lambda\alpha_k)$, and where $\vert \mathbf x\vert=x_1+x_2+\dots+x_k$. \end{Theorem} As a matter of fact, there is actually a binomial variant which implies the above theorem. Extending our multi-index notation, let $$\binom{\boldsymbol\alpha}{\boldsymbol \beta}:= \binom{\alpha_1}{\beta_1}\binom{\alpha_2}{\beta_2}\cdots \binom{\alpha_k}{\beta_k}.$$ \begin{Theorem} \label{conj:Brunat2} Let $\mathbf x=(x_1,x_2,\dots,x_k)$ be a vector of indeterminates, and let $\lambda$ be an indeterminate. Then, using the notation from Theorem~{\em\ref{conj:Brunat1}}, for any non-negative integers $n$ and $k$, we have \begin{equation} \label{eq:Brunat2} \det_{\boldsymbol\alpha,\boldsymbol\beta\in\mathcal C(n,k)}\left(\binom{\mathbf x+\lambda\boldsymbol\alpha} {\boldsymbol \beta}\right)=\lambda^{(k-1)\binom {n+k-1}k} \prod _{i=1} ^{n} \left(\frac{{ \vert \mathbf x\vert+(\lambda-1)n+i }} i\right) ^{\binom{n+k-i-1} {k-1}}. \end{equation} \end{Theorem} The above theorem is proved in \machSeite{BrKMAA}\cite{BrKMAA} by identification of factors (see ``Method~3" in Section~\ref{sec:eval}). Theorem~\ref{conj:Brunat2} follows by extracting the highest homogeneous component in \eqref{eq:Brunat1}. I report that, if one naively replaces ``compositions" by ``integer partitions" in the above considerations, then the arising determinants do not have nice product formulae. \medskip Another interesting determinant of a matrix with rows indexed by compositions appears in the work of Bergeron, Reutenauer, Rosas and Zabrocki \machSeite{BeRRAA}\cite[Theorem~4.8]{BeRRAA} on {\it Hopf algebras of non-commutative symmetric functions}. It was used there to show that a certain set of generators of non-commutative symmetric functions were algebraically independent. To state their determinant evaluation, we need to introduce some notation. Given a composition $\boldsymbol\alpha=(\alpha_1,\alpha_2,\dots,\alpha_k)$ of $n$ with all summands $\alpha_i$ positive, we let $D(\boldsymbol\alpha)=\{\alpha_1,\alpha_1+\alpha_2,\dots,\alpha_1+\alpha_2+\dots+\alpha_{k-1}\}$. Furthermore, for two compositions $\boldsymbol\alpha$ and $\boldsymbol\beta$ of $n$, we write $\boldsymbol\alpha\cup\boldsymbol\beta$ for the (unique) composition $\boldsymbol\gamma$ of $n$ with $D(\boldsymbol\gamma)=D(\boldsymbol\alpha)\cup D(\boldsymbol\beta)$. Finally, $\boldsymbol\alpha!$ is short for $\alpha_1!\,\alpha_2!\cdots\alpha_k!$. \begin{Theorem} \label{thm:BRRZ} Let $\operatorname{Comp}(n)$ denote the set of all compositions of $n$ all summands of which are positive. Then \begin{equation} \label{eq:BRRZ} \det_{\boldsymbol\alpha,\boldsymbol\beta\in\operatorname{Comp}(n)} \big((\boldsymbol\alpha\cup\boldsymbol\beta)!\big)= \prod _{\boldsymbol\gamma\in\operatorname{Comp}(n)} ^{} \prod _{i=1} ^{\ell(\boldsymbol\gamma)}a_{\gamma_{i}}, \end{equation} where $\ell(\boldsymbol\gamma)$ is the number of summands {\em(}components{\em)} of the composition $\boldsymbol\gamma$, and where $a_m$ denotes the number of {\em indecomposable permutations} of $m$ {\em(}cf.\ \machSeite{StanBI}\cite[Ex.~5.13(b)]{StanBI}{\em)}. These numbers can be computed recursively by $a_1=1$ and $$a_n=n!-\sum _{i=1} ^{n-1}a_i(n-1)!, \quad \quad n>1.$$ \quad \quad \qed \end{Theorem} As explained in \machSeite{BeRRAA}\cite{BeRRAA}, one proves the theorem by factoring the matrix in \eqref{eq:BRRZ} in the form $CDC^t$, where $C$ is the ``incidence matrix" of ``refinement of compositions," and where $D$ is a diagonal matrix. Thus, in particular, the LU-factorisation of the matrix is determined. \subsection{Two partition determinants} On the surface, {\it integer partitions} (see below for their definition) seem to be very closely related to compositions, as they can be considered as ``compositions where the order of the summands is without importance." However, experience shows that integer partitions are much more complex combinatorial objects than compositions. This may be the reason that the ``composition determinants" from the preceding subsection do not seem to have analogues for integer partitions. Leaving aside this disappointment, here {\it is} a determinant of a matrix in which rows and columns are indexed by integer partitions. This determinant arose in work on {\it linear forms of values of the Riemann zeta function evaluated at positive integers}, although the traces of it have now been completely erased in the final version of the article \machSeite{KrRiAA}\cite{KrRiAA}. (The symmetric function calculus in Section~12 of the earlier version \machSeite{KrRiAAA}\cite{KrRiAAA} gives a vague idea where it may have come from.) Recall that the {\it power symmetric function} of degree $d$ in $x_1,x_2,\dots,x_k$ is given by $x_1^d+x_2^d+\dots+x_k^d$, and is denoted by $p_d(x_1,x_2,\dots,x_k)$. (See \machSeite{LascAZ}\cite[Ch.~1 and 2]{LascAZ}, \machSeite{MacdAC}\cite[Ch.~I]{MacdAC} and \machSeite{StanBI}\cite[Ch.~7]{StanBI} for in-depth expositions of the theory of {\it symmetric functions}.) Then, while working on \machSeite{KrRiAA}\cite{KrRiAA}, Rivoal and the author needed to evaluate the determinant \begin{equation} \label{eq:Part} \det_{\lambda,\mu\in\operatorname{Part}(n,k)}\big(p_{\lambda}(\mu_1,\mu_2,\dots,\mu_k)\big), \end{equation} where $\operatorname{Part}(n,k)$ is the set of integer partitions of $n$ with at most $k$ parts, that is, the set of all possibilities to write $n$ as a sum of non-negative integers, $n=\lambda_1+\lambda_2+\dots+\lambda_k$, with $\lambda_1\ge\lambda_2\ge\dots\ge\lambda_k\ge0$ (the non-zero $\lambda_i$'s being called the {\it parts of $\lambda$}), and where $$p_\lambda(x_1,x_2,\dots,x_k)=p_{\lambda_1}(x_1,\dots,x_k) p_{\lambda_1}(x_1,\dots,x_k)\cdots p_{\lambda_k}(x_1,\dots,x_k). $$ Following the advice given in Section~\ref{sec:comb}, we went to the computer and let it calculate the prime factorisations of the values of this determinant for small values of $n$ and $k$. Indeed, the prime factors turned out be always small so that we were sure that a ``nice" formula exists for the determinant. However, even more seemed to be true. Recall that, in order to facilitate a proof of a (still unknown) formula, it is (almost) always a good idea to try to introduce more parameters (see \machSeite{KratBN}\cite[Sec.~2]{KratBN}). This is what we did. It led us consider the following determinant, \begin{equation} \label{eq:Part-gen} \det_{\lambda,\mu\in\operatorname{Part}(n,k)}\big(p_{\lambda}(\mu_1+X_1,\mu_2+X_2,\dots,\mu_k+X_k) \big), \end{equation} where $X_1,X_2,\dots,X_k$ are indeterminates. Here are some values of the determinant \eqref{eq:Part-gen} for special values of $n$ and $k$. For $n=k=3$ we obtain $$ 6 \,( X_1 - X_2+2 ) ( X_1 - X_3+1 ) ( X_2 - X_3+1 ) ( X_1 + X_2 + X_3+3) ^4, $$ for $n=4$ and $k=3$ we obtain \begin{multline*} 8 \,( X_1 - X_2 +1) ( X_1 - X_2+2 ) ( X_1 - X_2+3 ) ( X_1 - X_3+2 ) \\ \times ( X_2 - X_3+1 ) ( X_1 + X_2 + X_3+4) ^7, \end{multline*} for $n=5$ and $k=3$ we get \begin{multline*} 8 \,( X_1 - X_2+1) ( X_1 - X_2+2) ( X_1 - X_2+3) ( X_1 - X_2+4) ( X_1 - X_3+2)\\ \times ( X_1 - X_3+3) ( X_2 - X_3+1) ( X_2 - X_3+2) ( X_1 + X_2 + X_3+5)^{11}, \end{multline*} for $n=6$ and $k=3$ we get \begin{multline*} 576 \,( X_1 - X_2+1 ) ( X_1 - X_2+2 ) ^2 ( X_1 - X_2+3 ) ^2 ( X_1 - X_2+4 ) ( X_1 - X_2+5 )\\ \times ( X_1 - X_3+1 ) ( X_1 - X_3+2 ) ( X_1 - X_3+3 ) ( X_1 - X_3+4 ) ( X_2 - X_3+1 ) ^2 \\ \times ( X_2 - X_3+2 ) ^2 ( X_1 + X_2 + X_3+6 ) ^{16}, \end{multline*} for $n=4$ and $k=4$ we get \begin{multline*} 192\,( X_1 - X_2 +1) ( X_1 - X_2+2 ) ( X_1 - X_2+3 ) ( X_1 - X_3+2 ) ( X_2 - X_3+1 ) \\ \times ( X_1 - X_4+1 ) ( X_2 - X_4+1 ) ( X_3 - X_4+1 ) ( X_1 + X_2 + X_3 + X_4+4 ) ^7 \end{multline*} while for $n=5$ and $k=4$ we get \begin{multline*} 48 \,( X_1 - X_2+1 ) ( X_1 - X_2+2 ) ( X_1 - X_2+3 ) ( X_1 - X_2+4 ) ( X_1 - X_3+2 ) \\ \times ( X_1 - X_3+3 ) ( X_2 - X_3+1 ) ( X_2 - X_3+2 ) ( X_1 - X_4+2 ) \\ \times ( X_2 - X_4+1) ( X_3 - X_4+1 ) ( X_1 + X_2 + X_3 + X_4+5 ) ^{12}. \end{multline*} It is ``therefore" evident that there will be one factor which is a power of $n+\sum _{i=1} ^{k}X_i$, whereas the other factors will be of the form $X_i-X_j+c_{i,j}$, again raised to some power. In fact, the determinant is easy to compute for $k=1$ because, in that case, it reduces to a special case of the Vandermonde determinant evaluation. However, we were not able to come even close to computing \eqref{eq:Part-gen} in general. (As I already indicated, finally we managed to avoid the determinant evaluation in our work in \machSeite{KrRiAA}\cite{KrRiAA}.) To proceed further, we remark that evaluating \eqref{eq:Part-gen} is equivalent to evaluating the same determinant, but with the power symmetric functions $p_\lambda$ replaced by the {\it Schur functions} $s_\lambda$, because the transition matrix between these two bases of symmetric functions is the {\it character table of the symmetric group} of the corresponding order (cf.\ \machSeite{MacdAC}\cite[Ch.~I, Sec.~7]{MacdAC}), the determinant of which is known (see Theorem~\ref{thm:chi} below; since in our determinants \eqref{eq:Part} and \eqref{eq:Part-gen} the indices range over all partitions of $n$ {\it with at most $k$ parts}, it is in fact the refinement given in Theorem~\ref{thm:chik} which we have to apply). The determinant with Schur functions has some advantages over the one with power symmetric functions since the former decomposes into a finer block structure. Alain Lascoux observed that, in fact, there is a generalisation of the Schur function determinant to {\it Gra\3mannian Schubert polynomials}, which contains another set of variables, $Y_1,Y_2,\dots,Y_{n+k-1}$. More precisely, given $\lambda=(\lambda_1,\lambda_2,\dots,\lambda_k)$, let $\mathbb Y_\lambda(X_1,X_2,\dots,X_k;Y_1,Y_2,\dots)$ denote the polynomial in the variables $X_1,X_2,\dots,X_k$ and $Y_1,Y_2,\dots$, defined by (see \machSeite{LascAZ}\cite[Sections~1.4 and 9.7; the order of the $B_{k_i}$ should be reversed in (9.7.2) and analogous places]{LascAZ}) $$\mathbb Y_\lambda(X_1,X_2,\dots,X_k;Y_1,Y_2,\dots):=\det_{1\le i,j\le k}(S_{\lambda_i-i+j} (X_1,\dots,X_k;Y_1,\dots,Y_{\lambda_{i}+k-i})),$$ where the entries of the determinant are given by \begin{equation} \label{eq:XY} \sum _{m=0} ^{\infty}S_m (X_1,\dots,X_k;Y_1,\dots,Y_l)x^m= \frac {\prod _{i=1} ^{l}(1-Y_ix)} {\prod _{i=1} ^{k}(1-X_ix)}. \end{equation} The Gra\3mannian Schubert polynomial $\mathbb Y_\lambda$ reduces to the Schur function $s_\lambda$ when all the variables $Y_i$, $i=1,2,\dots$, are set equal to $0$. Given these definitions, Alain Lascoux (private communication) established the following result. \begin{Theorem} \label{thm:Lascoux} Let $X_1,X_2,\dots,X_k,Y_1,Y_2,\dots,Y_{n+k-1}$ be indeterminates. Then, \begin{multline} \label{eq:Part-gena} \det_{\lambda,\mu\in\operatorname{Part}(n,k)} \big(\mathbb Y_{\lambda}(\mu_1+X_1,\mu_2+X_2,\dots,\mu_k+X_k; Y_1,Y_2,\dots) \big)\\=\Pi(n,k) \prod _{\sigma\in G(n,k)} ^{}\(n+\sum _{i=1} ^{k}X_i-\sum _{i=1} ^{k}Y_{\sigma(i)}\), \end{multline} where $G(n,k)$ denotes the set of all {\it Gra\3mannian permutations},\footnote{A permutation $\sigma$ in $\mathfrak S_\infty$ (the set of all permutations of the natural numbers $\mathbb N$ which fix all but a finite number of elements of $\mathbb N$) is called {\it Gra\3mannian} if $\sigma(i)<\sigma(i+1)$ for all $i$ except possibly for one, the latter being called the {\it descent} of $\sigma$ (see \machSeite{MacdAE}\cite[p.~13]{MacdAE}). An inversion of $\sigma$ is a pair $(i,j)$ such that $i<j$ but $\sigma(i)>\sigma(j)$. We remark that the number of Gra\3mannian permutations with descent (if existent) at $k$ and at most $n-1$ inversions is equal to the number of partitions of at most $n-1$ with at most $k$ parts (including the empty partition). More concretely, if we denote this number by $g_{n,k}$, the generating function of the numbers $g_{n,k}$ is given by $$ \sum _{n=1} ^{\infty}g_{n,k}x^{n-1}= \frac {1} {1-q}\prod _{i=1} ^{k}\frac {1} {1-q^i}.$$ } the descent of which {\em(}if existent{\em)} is at $k$, and which contain at most $n-1$ inversions, and where $\Pi(n,k)$ is given recursively by \begin{equation} \label{eq:Pi} \Pi(n,k)=\Pi(n,k-1)\Pi(n-k,k) \prod _{\mu\in\operatorname{Part}^+(n,k)} ^{} \prod _{j=1} ^{k}(X_j+\mu_j-X_k), \end{equation} $\operatorname{Part}^+(n,k)$ denoting the set of partitions of $n$ into {\em exactly} $k$ {\em(}positive{\em)} parts, with initial conditions $\Pi(n,k)=1$ if $k=1$ or $n\le1$. Explicitly, \begin{equation} \label{eq:Piexpl} \Pi(n,k)=\underset{\vert w\vert_B<k-1} {\prod _{w} ^{}} \prod _{\mu\in\operatorname{Part}(n-\vert w\vert_A k+\operatorname{inv} w,k-\vert w\vert_B)} ^{} \prod _{j=1} ^{k-\vert w\vert_B}(X_j+\mu_j-X_{k-\vert w\vert_B}), \end{equation} where the product over $w$ runs over all finite-length words $w$ with letters from $\{A,B\}$, including the empty word. The notation $\vert w\vert_B$ means the number of occurrences of $B$ in $w$, with the analogous meaning for $\vert w\vert_A$. The quantity $\operatorname{inv} w$ denotes the number of inversions of $w=w_1w_2\dots$, which is the number of pairs of letters $(w_i,w_j)$, $i<j$, such that $w_i=B$ and $w_j=A$. \end{Theorem} We obtain the evaluation of the determinant \eqref{eq:Part} if we set $X_i=Y_i=0$ for all $i$ in the above theorem and multiply by \eqref{eq:chik}. Likewise, we obtain the evaluation of the determinant \eqref{eq:Part-gen} if we set $Y_i=0$ for all $i$ in the above theorem and multiply the result by \eqref{eq:chik}. For example, here is the determinant \eqref{eq:Part-gena} for $n=3$ and $k=2$, \begin{multline*} (X_1-X_2+2)(X_1+X_2-Y_1-Y_2+3)(X_1+X_2-Y_1-Y_3+3)\\ \times (X_1+X_2-Y_2-Y_3+3)(X_1+X_2-Y_1-Y_4+3), \end{multline*} and the following is the one for $n=k=3$, \begin{multline*} (X_1-X_2+2)(X_1-X_3+1)(X_2-X_3+1)\\ \times (X_1+X_2+X_3-Y_1-Y_2-Y_3+3) (X_1+X_2+X_3-Y_1-Y_2-Y_4+3)\\ \times (X_1+X_2+X_3-Y_1-Y_3-Y_4+3) (X_1+X_2+X_3-Y_1-Y_2-Y_5+3). \end{multline*} In the sequel, I sketch Lascoux's proof of Theorem~\ref{thm:Lascoux}. The first step consists in applying the {\it Monk formula for double Schubert polynomials} in the case of Gra\3mannian Schubert polynomials (see \machSeite{KoVeAA}\cite{KoVeAA}), $$ \(\sum _{i=1} ^{k}X_i-\sum _{i=1} ^{k}Y_{\lambda_{k-i+1}+i}\) \mathbb Y_\lambda = \mathbb Y_{\lambda+(1,0,0,\dots)} + \sum_\mu \mathbb Y_\mu, $$ where $\mathbb Y_\lambda$ is short for $\mathbb Y_\lambda(X_1,X_2,\dots,X_k;Y_1,Y_2,\dots)$, and where the sum on the right-hand side is over all partitions $\mu$ of the same size as $\lambda+(1,0,0,\dots)$ but lexicographically smaller. Clearly, by using this identity, appropriate row operations in the determinant \eqref{eq:Part-gena} show that \begin{equation} \label{eq:factor} n+\sum _{i=1} ^{k}X_i-\sum _{i=1} ^{k}Y_{\sigma(i)}, \end{equation} is one of its factors, the relation between $\sigma$ and $\lambda$ being $\sigma(i)=\lambda_{k-i+1}+i$, $i=1,2,\dots,k$. In particular, $\sigma$ can be extended (in a unique way) to a Gra\3mannian permutation. Moreover, doing these row operations, and taking out the factors of the form \eqref{eq:factor}, we collect on the one hand the product in \eqref{eq:Part-gena}, and we may on the other hand reduce the determinant \eqref{eq:Part-gena} to a determinant of the same form, in which, however, the partitions $\lambda=(\lambda_1,\lambda_2,\dots,\lambda_k)$ run over all partitions of size {\it at most\/} $n$ with the {\it additional property that\/} $\lambda_1=\lambda_2$ (instead of over all partitions from $\operatorname{Part}(n,k)$). As it turns out, the determinant thus obtained is independent of the variables $Y_1,Y_2,\break \dots,Y_{n+k-1}$. Indeed, if we expand each Schubert polynomial $\mathbb Y_\lambda(\mu_1+X_1,\mu_2+X_2,\dots,\mu_k+X_k;Y_1,Y_2,\dots)$ in the determinant as a linear combination of Schur functions in $X_1,X_2,\break \dots,X_k$ with coefficients being polynomials in the $Y_1,Y_2,\dots,Y_{n+k-1}$, then, by also using that $$S_1(\mu_1+X_1,\mu_2+X_2,\dots,\mu_k+X_k)= n+\sum _{i=1} ^{k}X_i$$ is independent of $\mu$, it is not difficult to see that one can eliminate all the $Y_i$'s by appropriate row operations. To summarise the current state of the discussion: we have already explained the occurrence of the product on the right-hand side of \eqref{eq:Part-gena} as a factor of the determinant. Moreover, the remaining factor is given by a determinant of the same form as in \eqref{eq:Part-gena}, \begin{equation} \label{eq:det-Y} \det_{\lambda,\mu} \big(\mathbb Y_{\lambda}(\mu_1+X_1,\mu_2+X_2,\dots,\mu_k+X_k; Y_1,Y_2,\dots) \big), \end{equation} where, as before, $\mu$ runs over all partitions in $\operatorname{Part}(n,k)$, but where $\lambda$ runs over partitions $(\lambda_1,\lambda_2,\dots,\lambda_k)$ of size at most $n$, with the additional restriction that $\lambda_1=\lambda_2$. The final observation was that this latter determinant is in fact independent of the $Y_i$'s. This allows us to specify $Y_1$ arbitrarily, say $Y_1=X_k$. Now another property of double Schubert polynomials, namely that (this follows from the definition of double Schubert polynomials by means of divided differences, see \machSeite{LascAZ}\cite[(10.2.3)]{LascAZ}, and standard properties of divided differences) \begin{equation} \label{eq:Y1} \mathbb Y_{\rho+(1,1,\dots,1)}(X_1,X_2,\dots,X_k;Y_1,Y_2,\dots)= \mathbb Y_{\rho}(X_1,X_2,\dots,X_k;Y_2,\dots)\prod _{j=1} ^{k}(X_j-Y_1) \end{equation} comes in handy. (In the vector $(1,1,\dots,1)$ on the left-hand side of \eqref{eq:Y1} there are $k$ occurrences of $1$.) Namely, if $Y_1=X_k$, the matrix in \eqref{eq:det-Y} (of which the determinant is taken), $M(n,k;Y_1,Y_2,\dots)$ say, decomposes in block form. If $\lambda$ is a partition with $\lambda_k>0$ and $\mu$ is a partition with $\mu_k=0$, then, because of \eqref{eq:Y1}, the corresponding entry in \eqref{eq:det-Y} vanishes. Furthermore, in the block where $\lambda$ and $\mu$ are partitions with $\lambda_k=\mu_k=0$, because of the definition \eqref{eq:XY} of the quantities $S_m(.)$ the corresponding entry reduces to \begin{multline*} \mathbb Y_{\lambda}(\mu_1+X_1,\mu_2+X_2,\dots,\mu_{k-1}+X_{k-1},X_k; X_k,Y_2,\dots)\\ =\mathbb Y_{\lambda}(\mu_1+X_1,\mu_2+X_2,\dots,\mu_{k-1}+X_{k-1}; Y_2,\dots). \end{multline*} In other words, this block is identical with $M(n,k-1;Y_2,\dots)$. Finally, in the block indexed by partitions $\lambda$ and $\mu$ with $\lambda_k>0$ and $\mu_k>0$, we may use \eqref{eq:Y1} to factor $ \prod _{j=1} ^{k}(X_j+\mu_j-X_k)$ out of the column indexed by $\mu$, for all such $\mu$. What remains is identical with $M(n-k,k;Y_2,\dots)$. Taking determinants, we obtain the recurrence \eqref{eq:Pi}. (Here we use again that the determinants in \eqref{eq:det-Y}, that is, in particular, the determinants of $M(n,k;Y_1,\dots)$, $M(n,k-1;Y_2,\dots)$, and of $M(n-k,k;Y_2,\dots)$, are all independent of the $Y_i$'s.) The explicit form \eqref{eq:Piexpl} for $\Pi(n,k)$ can be easily derived by induction on $n$ and $k$. \medskip A few paragraphs above, we mentioned in passing another interesting determinant of a matrix the rows and columns of which are indexed by (integer) partitions: the {\it determinant of the character table of the symmetric group $\mathfrak S_n$} (cf.\ \machSeite{JameAA}\cite[Cor.~6.5]{JameAA}). Since this is a classical and beautiful determinant evaluation which I missed to state in \machSeite{KratBN}\cite{KratBN}, I present it now in the theorem below. There, the notation $\lambda\vdash n$ stands for ``$\lambda$ is a partition of $n$." For all undefined notation, I refer the reader to standard texts on the representation theory of symmetric groups, as for example \machSeite{JameAA}% \machSeite{JaKeAA}% \machSeite{SagaAQ}% \cite{JameAA,JaKeAA,SagaAQ}. \begin{Theorem} \label{thm:chi} For partitions $\lambda$ and $\rho$ of $n$, let $\chi^\lambda(\rho)$ denote the value of the irreducible character $\chi^\lambda$ evaluated at a permutation of cycle type $\rho$. Then \begin{equation} \label{eq:chi} \det_{\lambda,\rho\,\vdash n}\(\chi^\lambda(\rho)\)= \prod _{\mu\,\vdash n} ^{} \prod _{i\ge1} ^{}\mu_i. \end{equation} In words: the determinant of the character table of the symmetric group $\mathfrak S_n$ is equal to the products of all the parts of all the partitions of $n$.\quad \quad \qed \end{Theorem} A refinement of this statement, where we restrict to partitions of $n$ with at most $k$ parts, is the following. \begin{Theorem} \label{thm:chik} With the notation of Theorem~{\em\ref{thm:chi}}, for all positive integers $n$ and $k$, $n\ge k$, we have \begin{equation} \label{eq:chik} \det_{\lambda,\rho\in\operatorname{Part}(n,k)}\(\chi^\lambda(\rho)\)= \prod _{\mu\in\operatorname{Part}(n,k)} ^{} \prod _{i\ge1} ^{}m_i(\mu)!, \end{equation} where $m_i(\mu)$ is the number of times $i$ occurs as a part in the partition $\mu$.\quad \quad \qed \end{Theorem} This determinant evaluation follows from the decomposition of the {\it full\/} character table of $\mathfrak S_n$ in the form \begin{equation} \label{eq:LK} \(\chi^\lambda(\rho)\)_{\lambda,\rho\,\vdash n}=L\cdot K^{-1}, \end{equation} where $L=(L_{\lambda,\mu})_{\lambda,\mu\,\vdash n}$ is the transition matrix from power symmetric functions to monomial symmetric functions, and where $K=(K_{\lambda,\mu})_{\lambda,\mu\,\vdash n}$ is the {\it Kostka matrix}, the transition matrix from Schur functions to monomial symmetric functions (see \machSeite{MacdAC}\cite[Ch.~I, (6.12)]{MacdAC}). For, if we order the partitions of $n$ so that the partitions in $\operatorname{Part}(n,k)$ come before the partitions in $\operatorname{Part}(n,k+1)$, $k=1,2,\dots,n-1$, and within $\operatorname{Part}(n,k)$ lexicographically, then, with respect to this order, the matrix $L$ is lower triangular and the matrix $K$, and hence also $K^{-1}$, is upper triangular. Furthermore, the matrix $K$, and hence also $K^{-1}$, is even block diagonal, the blocks along the diagonal being the ones which are formed by the rows and columns indexed by the partitions in $\operatorname{Part}(n,k)$, $k=1,2,\dots,n$. These facts together imply that the decomposition \eqref{eq:LK} restricts to the submatrices indexed by partitions in $\operatorname{Part}(n,k)$, \begin{equation* \(\chi^\lambda(\rho)\)_{\lambda,\rho\in \operatorname{Part}(n,k)}= (L_{\lambda,\mu})_{\lambda,\mu\in\operatorname{Part}(n,k)}\cdot (K_{\lambda,\mu})_{\lambda,\mu\in\operatorname{Part}(n,k)}^{-1}. \end{equation*} If we now take determinants on both sides, then, in view of $K_{\mu,\mu}=1$ and of $$L_{\mu,\mu}=\prod _{i\ge1} ^{}m_i(\mu)!$$ for all $\mu$, the theorem follows. \medskip Further examples of nice determinant evaluations of tables of {\it characters of representations of symmetric groups and their double covers} can be found in \machSeite{BeOSAA}% \machSeite{OlssAB}% \cite{BeOSAA,OlssAB}. Determinants of tables of {\it characters of the alternating group} can be found in \machSeite{BeOlAE}\cite{BeOlAE}. \subsection{Elliptic determinant evaluations}\label{sec:ell} In special functions theory there is currently a disease rapidly spreading, generalising the earlier mentioned $q$-disease (see Footnote~\ref{foot:q}). It could be called the {\it ``elliptic disease}." Recall that, during the $q$-disease, we replaced every positive integer $n$ by $1+q+q^2+\dots+q^{n-1}=(1-q^n) /(1-q)$, and, more generally, shifted factorials $a(a+1)\cdots(a+k-1)$ by $q$-shifted factorials $(1-\alpha)(1-q\alpha)\cdots(1-\alpha q^{k-1})$. (Here, $\alpha$ takes the role of $q^a$, and one drops the powers of $1-q$ in order to ease notation.) Doing this with some ``ordinary" identity, we arrived (hopefully) at its {\it $q$-analogue}. Now, once infected by the elliptic disease, we would replace every occurrence of a term $1-x$ (and, looking at the definition of $q$-shifted factorials, we can see that there will be many) by its {\it elliptic analogue} $\theta(x;p)$: $$\theta(x)=\theta(x;p)=\prod_{j=0}^\infty(1-p^jx)(1-p^{j+1}/x). $$ Here, $p$ is a complex number with $|p|<1$, which will be fixed throughout. Up to a trivial factor, $\theta(e^{2\pi ix};e^{2\pi i \tau})$ equals the {\it Jacobi theta function} $\theta_1(x|\tau)$ (cf.\ \machSeite{WW}\cite{WW}). Clearly, $\theta(x)$ reduces to $1-x$ if $p=0$. At first sight, one will be sceptical if this is a fruitful thing to do. After all, for working with the functions $\theta(x)$, the only identities which are available are the (trivial) {\it inversion formula} \begin{equation}\label{ti}\theta(1/x)=-\frac1 x\,\theta(x), \end{equation} the (trivial) {\it quasi-periodicity} \begin{equation}\label{tp}\theta(px)=-\frac1 x\,\theta(x),\end{equation} and {\it Riemann's} (highly non-trivial) {\it addition formula} (cf.\ \machSeite{WW}\cite[p.~451, Example~5]{WW}) \begin{equation}\label{tadd} \theta(xy)\,\theta(x/y)\,\theta(uv)\,\theta(u/v)- \theta(xv)\,\theta(x/v)\,\theta(uy)\,\theta(u/y) =\frac uy\,\theta(yv)\,\theta(y/v)\,\theta(xu)\,\theta(x/u). \end{equation} Nevertheless, it has turned out recently that a surprising number of identities from the ``ordinary" and from the ``$q$-world" can be lifted to the elliptic level. This is particularly true for series of hypergeometric nature. We refer the reader to Chapter~11 of \machSeite{GaRaAA}\cite{GaRaAA} for an account of the current state of the art in the theory of, as they are called now, {\it elliptic hypergeometric series}. On the following pages, I give elliptic determinant evaluations a rather extensive coverage because, first of all, they were non-existent in \machSeite{KratBN}\cite{KratBN} (with the exception of the mention of the papers \machSeite{MilnAO}% \machSeite{MilnAP}\cite{MilnAP,MilnAO} by Milne), and, second, because I believe that the ``elliptic research" is a research direction that will further prosper in the next future and will have numerous applications in many fields, also outside of just special functions theory and number theory. I further believe that the determinant evaluations presented in this subsection will turn out to be as fundamental as the determinant evaluations in Sections~2.1 and 2.2 in \machSeite{KratBN}\cite{KratBN}. For some of them this belief is already a fact. For example, determinant evaluations involving elliptic functions have come into the picture in the theory of {\it multiple elliptic hypergeometric series}, see \machSeite{KN}% \machSeite{RainAA}% \machSeite{RoseAA}% \machSeite{Ro}% \machSeite{RoScAB}% \machSeite{Sp}% \machSeite{WarnAG}% \cite{KN,RainAA,RoseAA,Ro,RoScAB,Sp,WarnAG}. They have also an important role in the study of \emph{Ruijsenaars operators} and related {\it integrable systems} \machSeite{H}% \machSeite{Ru}% \cite{H,Ru}. Furthermore, they have recently found applications in number theory to the problem of counting the number of {\it representations of an integer as a sum of triangular numbers} \machSeite{RoseAB}\cite{RoseAB}. Probably the first elliptic determinant evaluation is due to Frobenius \machSeite{Fr}\cite[(12)]{Fr}. This identity has found applications to {\it Ruijsenaars operators} \machSeite{Ru}\cite{Ru}, to {\it multidimensional elliptic hypergeometric series} and {\it integrals} \machSeite{KN}\cite{KN}, \machSeite{RainAA}\cite{RainAA} and to {\it number theory} \machSeite{RoseAB}\cite{RoseAB}. For a generalisation to {\it higher genus Riemann surfaces}, see \machSeite{F}\cite[Cor.~2.19]{F}. Amdeberhan \machSeite{AmdeAC}\cite{AmdeAC} observed that it can be easily proved using the condensation method (see ``Method~2" in Section~\ref{sec:eval}). \begin{Theorem} \label{froa} Let $x_1,x_2,\dots,x_n$, $a_1,a_2,\dots,a_n$ and $t$ be indeterminates. Then there holds \begin{equation}\label{froaid} \det_{1\leq i,j\leq n}\left(\frac{\theta(ta_jx_i)}{\theta(t)\, \theta(a_jx_i)}\right) =\frac{\theta(ta_1\dotsm a_nx_1\dotsm x_n)}{\theta(t)} \frac{\displaystyle\prod_{1\leq i<j\leq n} a_jx_j\,\theta(a_i/a_j)\,\theta(x_i/x_j)} {\displaystyle\prod_{i,j=1}^n\theta(a_jx_i)}. \end{equation} \quad \quad \qed \end{Theorem} For $p=0$ and $t\to\infty$, this determinant identity reduces to Cauchy's evaluation \eqref{eq:Cauchy} of the double alternant, and, thus, may be regarded as an ``elliptic analogue" of the latter. Okada \machSeite{OkadAK}\cite[Theorem~1.1]{OkadAK} has recently found an elliptic extension of Schur's evaluation \eqref{eq:Schur} of a Cauchy-type Pfaffian. His proof works by the Pfaffian version of the condensation method. \begin{Theorem} \label{thm:Okada} Let $x_1,x_2,\dots,x_n$, $t$ and $w$ be indeterminates. Then there holds \begin{multline}\label{eq:Okada} \underset{1\leq i,j\leq 2n}\operatorname{Pf}\left(\frac{\theta(x_j/x_i)}{\theta(x_ix_j)} \frac{\theta(tx_ix_j)}{\theta(t)} \frac{\theta(wx_ix_j)}{\theta(w)} \right) \\= \frac{\theta(tx_1\dotsm x_{2n})}{\theta(t)} \frac{\theta(wx_1\dotsm x_{2n})}{\theta(w)} \prod _{1\le i<j\le 2n} ^{} \frac{\theta(x_j/x_i)} {x_j\,\theta(x_ix_j)}. \end{multline} \quad \quad \qed \end{Theorem} The next group of determinant evaluations is from \machSeite{RoScAC}\cite[Sec.~3]{RoScAC}. As the Vandermonde determinant evaluation, or the other Weyl denominator formulae (cf.\ \machSeite{KratBN}\cite[Lemma~2]{KratBN}), are fundamental {\it polynomial} determinant evaluations, the evaluations in Lemma~\ref{wp} below are equally fundamental in the elliptic domain as they can be considered as the elliptic analogues of the former. Indeed, Rosengren and Schlosser show that they {\it imply} the {\it Macdonald identities associated to affine root systems} \machSeite{MacdAA}\cite{MacdAA}, which are the affine analogues of the Weyl denominator formulae. In particular, in this way they obtain new proofs of the Macdonald identities. In order to conveniently formulate Rosengren and Schlosser's determinant evaluations, we shall adopt the following terminology from \machSeite{RoScAC}\cite{RoScAC}. For $0<|p|<1$ and $t\neq 0$, an {\it $A_{n-1}$ theta function $f$ of norm $t$} is a holomorphic function for $x\neq 0$ such that \begin{equation}\label{ade}f(px)=\frac{(-1)^n}{tx^n}\,f(x).\end{equation} Moreover, if $R$ denotes either of the root systems $B_n$, $B^\vee_n$, $C_n$, $C^\vee_n$, $BC_n$ or $D_n$ (see Footnote~\ref{foot:root} and \machSeite{HumpAC}\cite{HumpAC} for information on root systems), we call $f$ an {\it$R$ theta function} if {\allowdisplaybreaks \begin{align*} f(px)&=-\frac{1}{p^{n-1}x^{2n-1}}\,f(x),& f(1/x)&=-\frac 1x\,f(x),& R&=B_n,\\ f(px)&=-\frac{1}{p^{n}x^{2n}}\,f(x),& f(1/x)&=-f(x),& R&=B_n^\vee,\\ f(px)&=\frac{1}{p^{n+1}x^{2n+2}}\,f(x),& f(1/x)&=-f(x),& R&=C_n,\\ f(px)&=\frac{1}{p^{n-\frac12}x^{2n}}\,f(x),& f(1/x)&=-\frac 1x\,f(x),& R&=C_n^\vee,\\ f(px)&=\frac{1}{p^{n}x^{2n+1}}\,f(x),& f(1/x)&=-\frac 1x\,f(x),& R&=BC_n,\\ f(px)&=\frac{1}{p^{n-1}x^{2n-2}}\, f(x),& f(1/x)&=f(x),& R&=D_n. \end{align*} } Given this definition, Rosengren and Schlosser \machSeite{RoScAC}\cite[Lemma~3.2]{RoScAC} show that a function $f$ is an $A_{n-1}$ theta function of norm $t$ if and only if there exist constants $C$, $b_1,\dots,b_{n}$ such that $b_1\dotsm b_n=t$ and $$f(x)=C\,\theta(b_1x)\cdots\theta(b_nx), $$ and for the other six cases, they show that $f$ is an $R$ theta function if and only if there exist constants $C$, $b_1,\dots,b_{n-1}$ such that \begin{align*}f(x)&=C\,\theta(x)\,\theta(b_1x)\,\theta(b_1/x) \cdots\theta(b_{n-1}x)\,\theta(b_{n-1}/x),& R&=B_n,\\ f(x)&=C\,x^{-1}\theta(x^2;p^2)\,\theta(b_1x)\,\theta(b_1/x) \cdots\theta(b_{n-1}x)\,\theta(b_{n-1}/x),& R&=B_n^\vee,\\ f(x)&=C\,x^{-1}\theta(x^2)\,\theta(b_1x)\,\theta(b_1/x) \cdots\theta(b_{n-1}x)\,\theta(b_{n-1}/x),& R&=C_n,\\ f(x)&=C\,\theta(x;p^{\frac12})\,\theta(b_1x)\,\theta(b_1/x) \cdots\theta(b_{n-1}x)\,\theta(b_{n-1}/x),& R&=C_n^\vee,\\ f(x)&=C\,\theta(x)\,\theta(px^2;p^2)\,\theta(b_1x)\,\theta(b_1/x) \cdots\theta(b_{n-1}x)\,\theta(b_{n-1}/x),& R&=BC_n,\\ f(x)&=C\,\theta(b_1x)\,\theta(b_1/x) \cdots\theta(b_{n-1}x)\,\theta(b_{n-1}/x),& R&=D_n, \end{align*} where $\theta(x)=\theta(x;p)$. If one puts $p=0$, then an $A_{n-1}$ theta function of norm $t$ becomes a polynomial of degree $n$ such that the reciprocal of the product of its roots is equal to $t$. Similarly, if one puts $p=0$, then a $D_n$ theta function becomes a polynomial in $(x+1/x)$ of degree $n$. This is the specialisation of some of the following results which is relevant for obtaining the earlier Lemmas~\ref{lem:RS1}--\ref{cdetr1cor}. The elliptic extension of the Weyl denominator formulae is the following formula. (See \machSeite{RoScAC}\cite[Prop.~3.4]{RoScAC}.) \begin{Lemma}\label{wp} Let $f_1,\dots,f_n$ be $A_{n-1}$ theta functions of norm $t$. Then, \begin{equation}\label{awpi}\det_{1\leq i,j\leq n}\left(f_j(x_i)\right)= C\,W_{A_{n-1}}(x), \end{equation} for some constant $C$, where $$W_{A_{n-1}}(x)=\theta(tx_1\dotsm x_n) \,\prod_{1\leq i<j\leq n}x_j\theta(x_i/x_j).$$ Moreover, if $R$ denotes either $B_n$, $B^\vee_n$, $C_n$, $C^\vee_n$, $BC_n$ or $D_n$ and $f_1,\dots,f_n$ are $R$ theta functions, we have \begin{equation}\label{wpi} \det_{1\leq i,j\leq n}\left(f_j(x_i)\right)=C\,W_R(x) , \end{equation} for some constant $C$, where {\allowdisplaybreaks \begin{align*} W_{B_n}(x)&=\prod_{i=1}^n\theta(x_i) \prod_{1\leq i<j\leq n}x_i^{-1}\theta(x_ix_j)\,\theta(x_i/x_j), \\ W_{B_n^\vee}(x)&=\prod_{i=1}^n x_i^{-1}\theta(x_i^2;p^2) \prod_{1\leq i<j\leq n}x_i^{-1}\theta(x_ix_j)\,\theta(x_i/x_j), \\ W_{C_n}(x)&=\prod_{i=1}^n x_i^{-1}\theta(x_i^2) \prod_{1\leq i<j\leq n}x_i^{-1}\theta(x_ix_j)\,\theta(x_i/x_j),\\ W_{C_n^\vee}(x)&=\prod_{i=1}^n\theta(x_i;p^{\frac 12}) \prod_{1\leq i<j\leq n}x_i^{-1}\theta(x_ix_j)\,\theta(x_i/x_j), \\ W_{BC_n}(x)&=\prod_{i=1}^n\theta(x_i)\,\theta(px_i^2;p^2) \prod_{1\leq i<j\leq n}x_i^{-1}\theta(x_ix_j)\,\theta(x_i/x_j), \\ W_{D_n}(x)&=\prod_{1\leq i<j\leq n}x_i^{-1}\theta(x_ix_j)\,\theta(x_i/x_j). \end{align*}}% \quad \quad \qed \end{Lemma} Rosengren and Schlosser show in \machSeite{RoScAC}\cite[Prop.~6.1]{RoScAC} that the famous Macdonald identities for affine root systems \machSeite{MacdAA}\cite{MacdAA} are equivalent to special cases of this lemma. We state the corresponding results below. \begin{Theorem}\label{mdp} The following determinant evaluations hold: $$ \det_{1\leq i,j\leq n}\left(x_i^{j-1} \theta((-1)^{n-1}p^{j-1}tx_i^n;p^n)\right)=\frac{(p;p)_\infty^{n}}{(p^n;p^n)_\infty^n} \,W_{A_{n-1}}(x), $$ \begin{multline*} \det_{1\leq i,j\leq n}\left(x_i^{j-n} \theta(p^{j-1}x_i^{2n-1};p^{2n-1}) -x_i^{n+1-j} \theta(p^{j-1}x_i^{1-2n};p^{2n-1}) \right)\\ =\frac{2(p;p)_\infty^{n}} {(p^{2n-1};p^{2n-1})_\infty^n}\,W_{B_{n}}(x), \end{multline*} \begin{multline*} \det_{1\leq i,j\leq n}\left(x_i^{j-n-1} \theta(p^{j-1}x_i^{2n};p^{2n}) -x_i^{n+1-j} \theta(p^{j-1}x_i^{-2n};p^{2n}) \right)\\ =\frac{2(p^2;p^2)_\infty(p;p)_\infty^{n-1}} {(p^{2n};p^{2n})_\infty^n}\,W_{B_{n}^\vee}(x), \end{multline*} \begin{multline*} \det_{1\leq i,j\leq n}\left(x_i^{j-n-1} \theta(-p^{j}x_i^{2n+2};p^{2n+2}) -x_i^{n+1-j} \theta(-p^{j}x_i^{-2n-2};p^{2n+2}) \right)\\ =\frac{(p;p)_\infty^{n}} {(p^{2n+2};p^{2n+2})_\infty^n}\, W_{C_{n}}(x), \end{multline*} \begin{multline*} \det_{1\leq i,j\leq n}\left(x_i^{j-n} \theta(-p^{j-\frac12}x_i^{2n};p^{2n}) -x_i^{n+1-j} \theta(-p^{j-\frac12}x_i^{-2n};p^{2n}) \right)\\ =\frac{(p^{\frac12};p^{\frac12})_\infty(p;p)_\infty^{n-1}} {(p^{2n};p^{2n})_\infty^n}\, W_{C_{n}^\vee}(x), \end{multline*} \begin{multline*} \det_{1\leq i,j\leq n}\left(x_i^{j-n} \theta(-p^{j}x_i^{2n+1};p^{2n+1}) -x_i^{n+1-j} \theta(-p^{j}x_i^{-2n-1};p^{2n+1}) \right)\\ =\frac{(p;p)_\infty^n}{(p^{2n+1};p^{2n+1})_\infty^n}\, W_{BC_{n}}(x), \end{multline*} \begin{multline*} \det_{1\leq i,j\leq n}\left(x_i^{j-n} \theta(-p^{j-1}x_i^{2n-2};p^{2n-2}) +x_i^{n-j} \theta(-p^{j-1}x_i^{2-2n};p^{2n-2}) \right)\\ =\frac{4(p;p)_\infty^n}{(p^{2n-2};p^{2n-2})_\infty^n} \,W_{D_{n}}(x), \qquad n\geq 2. \end{multline*} \quad \quad \qed \end{Theorem} Historically, aside from Frobenius' elliptic Cauchy identity \eqref{froaid}, the subject of elliptic determinant evaluations begins with Warnaar's remarkable paper \machSeite{WarnAG}\cite{WarnAG}. While the main subject of this paper is {\it elliptic hypergeometric series}, some elliptic determinant evaluations turn out to be crucial for the proofs of the results. Lemma~5.3 from \machSeite{WarnAG}\cite{WarnAG} extends one of the basic determinant lemmas listed in \machSeite{KratBN}\cite{KratBN}, namely \machSeite{KratBN}\cite[Lemma~5]{KratBN}, to the elliptic world, to which it reduces in the case $p=0$. We present this important elliptic determinant evaluation in the theorem below. \begin{Theorem} \label{bcdet} Let $x_1,x_2,\dots,x_n$, $a_1,a_2,\dots,a_n$ be indeterminates. For each $j=1,\dots,n$, let $P_j(x)$ be a $D_j$ theta function. Then there holds \begin{multline}\label{bcdetid} \det_{1\le i,j\le n}\left(P_{j}(x_i) \prod_{k=j+1}^n\theta(a_kx_i)\,\theta(a_k/x_i)\right)\\ =\prod_{i=1}^nP_{i}(a_i) \prod_{1\le i<j\le n}a_jx_j^{-1}\,\theta(x_jx_i)\,\theta(x_j/x_i). \end{multline} \quad \quad \qed \end{Theorem} Warnaar used this identity to obtain a summation formula for a {\it multidimensional elliptic hyper\-geometric series}. Further related applications may be found in \machSeite{RoseAA}% \machSeite{Ro}% \machSeite{RoScAB}% \machSeite{Sp}% \cite{RoseAA,Ro,RoScAB,Sp}. The relevant special case of the above theorem is the following (see \machSeite{WarnAG}\cite[Cor.~5.4]{WarnAG}). It is the elliptic generalisation of \machSeite{KratBN}\cite[Theorem~28]{KratBN}. In the statement, we use the notation \begin{equation} \label{eq:pqell} (a;q,p)_m=\theta(a;p)\,\theta(aq;p)\cdots \theta(aq^{m-1};p), \end{equation} which extends the notation for $q$-shifted factorials to the elliptic world. \begin{Theorem} \label{thm:Warn1} Let $X_1,X_2,\dots,X_n$, $A$, $B$ and $C$ be indeterminates. Then, for any non-negative integer $n$, there holds \begin{multline} \label{eq:Warn1} \det_{1\le i,j\le n}\(\frac {(AX_i;q,p)_{n-j}\,(AC/X_i;q,p)_{n-j}} {(BX_i;q,p)_{n-j}\,(BC/X_i;q,p)_{n-j}}\)\\=(Aq)^{\binom n2} \prod _{1\le i<j\le n} ^{}X_j\,\theta(X_i/X_j)\,\theta(C/X_iX_j) \prod _{i=1} ^{n}\frac {(B/A;q,p)_{i-1}\,(ABCq^{2n-2i};q,p)_{i-1}} {(BX_i;q,p)_{n-1}\,(BC/X_i;q,p)_{n-1}}. \end{multline} \quad \quad \qed \end{Theorem} Theorem~29 from \machSeite{KratBN}\cite{KratBN}, which is slightly more general than \machSeite{KratBN}\cite[Theorem~28]{KratBN}, can also be extended to an elliptic theorem by suitably specialising the variables in Theorem~\ref{bcdet}. \begin{Theorem} \label{thm:Warn1a} Let $X_1,X_2,\dots,X_n$, $Y_1,Y_2,\dots,Y_n$, $A$ and $B$ be indeterminates. Then, for any non-negative integer $n$, there holds \vbox{\noindent \begin{multline} \label{eq:Warn1a} \det_{1\le i,j\le n}\(\frac {(X_iY_j;q,p)_{j}\,(AY_j/X_i;q,p)_{j}} {(BX_i;q,p)_{j}\,(AB/X_i;q,p)_{j}}\)\\= q^{2\binom n3}(AB)^{\binom n2} \prod _{1\le i<j\le n} ^{}\theta(X_jX_i/A)\,\theta(X_j/X_i)\\ \times \prod _{i=1} ^{n}\frac {(ABY_iq^{i-2};q,p)_{i-1}\,(Y_i/Bq^{i-1};q,p)_{i-1}} {X_i^{i-1}\,(BX_i;q,p)_{n-1}\,(AB/X_i;q,p)_{n-1}}. \end{multline} \quad \quad \qed} \end{Theorem} Another, very elegant, special case of Theorem~\ref{bcdet} is the following elliptic Cauchy-type determinant evaluation. It was used by Rains \machSeite{RainAA}\cite[Sec.~3]{RainAA} in the course of deriving a {\it $BC_n\leftrightarrow BC_m$ integral transformation}. \begin{Lemma} \label{frobc} Let $x_1,x_2,\dots,x_n$ and $a_1,a_2,\dots,a_n$ be indeterminates. Then there holds \begin{equation}\label{frobcid} \det_{1\leq i,j\leq n}\left(\frac 1{\theta(a_jx_i)\,\theta(a_j/x_i)}\right) =\frac{\displaystyle\prod_{1\leq i<j\leq n} a_jx_j^{-1}\,\theta(x_jx_i)\,\theta(x_j/x_i)\,\theta(a_ia_j)\,\theta(a_i/a_j)} {\displaystyle\prod_{i,j=1}^n\theta(a_jx_i)\,\theta(a_j/x_i)}. \end{equation} \quad \quad \qed \end{Lemma} The remaining determinant evaluations in the current subsection, with the exception of the very last one, are all due to Rosengren and Schlosser \machSeite{RoScAC}\cite{RoScAC}. The first one is a further (however non-obvious) consequence of Theorem~\ref{bcdet} (see \machSeite{RoScAC}\cite[Cor.~4.3]{RoScAC}). Two related determinant evaluations, restricted to the polynomial case, were applied in \machSeite{SchlAB}\cite{SchlAB} and \machSeite{SchlAF}\cite{SchlAF} to obtain {\it multidimensional matrix inversions} that played a major role in the derivation of new {\it summation formulae for multidimensional basic hypergeometric series}. \begin{Theorem} \label{thm:cdet} Let $x_1,x_2,\dots,x_n$, $a_1,a_2,\dots,a_{n+1}$, and $b$ be indeterminates. For each $j=1,\dots,n+1$, let $P_j(x)$ be a $D_j$ theta function. Then there holds \begin{multline} P_{n+1}(b)\det_{1\le i,j\le n}\!\Bigg(P_{j}(x_i) \prod_{k=j+1}^{n+1}\big(\theta(a_kx_i)\,\theta(a_k/x_i)\big)\\ -\frac{P_{n+1}(x_i)}{P_{n+1}(b)}P_{j}(b) \prod_{k=j+1}^{n+1}\big(\theta(a_kb)\,\theta(a_k/b)\big)\Bigg)\\ =\prod_{i=1}^{n+1}P_{i}(a_i) \prod_{1\le i<j\le n+1}a_jx_j^{-1}\,\theta(x_jx_i)\,\theta(x_j/x_i), \end{multline} where $x_{n+1}=b$. \quad \quad \qed \end{Theorem} The next determinant evaluation is Theorem~4.4 from \machSeite{RoScAC}\cite{RoScAC}. It generalises another basic determinant lemma listed in \machSeite{KratBN}\cite{KratBN}, namely Lemma~6 from \machSeite{KratBN}\cite{KratBN}, to the elliptic case. It looks as if it is a limit case of Warnaar's in Theorem~\ref{bcdet}. However, limits are very problematic in the elliptic world, and therefore it does not seem that Theorem~\ref{bcdet} implies the theorem below. For a generalisation in a different direction than Theorem~\ref{bcdet} see \machSeite{TV}\cite[Appendix B]{TV} (cf.\ also \machSeite{RoScAC}\cite[Remark~4.6]{RoScAC}). \begin{Theorem}\label{adet} Let $x_1,x_2,\dots,x_n$, $a_1,a_2,\dots,a_n$, and $t$ be indeterminates. For each $j=1,\dots,n$, let $P_j(x)$ be an $A_{j-1}$ theta function of norm $ta_1\dotsm a_j$. Then there holds \begin{multline}\label{adetid} \det_{1\le i,j\le n}\left(P_j(x_i) \prod_{k=j+1}^n\theta(a_kx_i)\right)\\ =\frac{\theta(ta_1\dotsm a_nx_1\dotsm x_n)}{\theta(t)} \prod_{i=1}^nP_i(1/a_i) \prod_{1\le i<j\le n}a_jx_j\,\theta(x_i/x_j). \end{multline} \quad \quad \qed \end{Theorem} As is shown in \machSeite{RoScAC}\cite[Cor.~4.8]{RoScAC}, this identity implies the following determinant evaluation. \begin{Theorem} \label{adetcor} Let $x_1,x_2,\dots,x_n$, $a_1,a_2,\dots,a_{n+1}$ and $b$ be indeterminates. For each $j=1,\dots,n+1$, let $P_j(x)$ be an $A_{j-1}$ theta function of norm $ta_1\dotsm a_j$. Then there holds \begin{multline} P_{n+1}(b)\;\det_{1\le i,j\le n}\left(P_j(x_i) \prod_{k=j+1}^{n+1}\theta(a_kx_i)-\frac{P_{n+1}(x_i)}{P_{n+1}(b)} P_j(b)\prod_{k=j+1}^{n+1}\theta(a_kb)\right)\\ =\frac{\theta(tba_1\dotsm a_{n+1}x_1\dotsm x_n)}{\theta(t)} \prod_{i=1}^{n+1}P_i(1/a_i) \prod_{1\le i<j\le n+1}a_jx_j\,\theta(x_i/x_j), \end{multline} where $x_{n+1}=b$. \quad \quad \qed \end{Theorem} By combining Lemma~\ref{frobc} and Theorem~\ref{adet}, a determinant evaluation similar to the one in Theorem~\ref{thm:cdet}, but different, is obtained in \machSeite{RoScAC}\cite[Theorem~4.9]{RoScAC}. \begin{Theorem}\label{cdet} Let $x_1,x_2,\dots,x_n$, $a_1,a_2,\dots,a_n$, and $c_1,\dots, c_{n+2}$ be indeterminates. For each $j=1,\dots,n$, let $P_j(x)$ be an $A_{j-1}$ theta function of norm $(c_1\dotsm c_{n+2}a_{j+1}\dotsm a_n)^{-1}$. Then there holds \begin{multline}\label{dv} \det_{1\leq i,j\leq n}\left(x_i^{-n-1} P_j(x_i)\prod_{k=1}^{n+2}\theta(c_kx_i)\, \prod_{k=j+1}^n\theta(a_kx_i)\right.\\ \left.-x_i^{n+1}P_j(x_i^{-1})\prod_{k=1}^{n+2}\theta(c_kx_i^{-1})\, \prod_{k=j+1}^n\theta(a_kx_i^{-1})\right)\\ \kern-4cm=\frac{a_1\dotsm a_n} {x_1\dotsm x_n\,\theta(c_1\dotsm c_{n+2}a_1\dotsm a_n)} \prod_{i=1}^nP_i(1/a_i)\\\times \prod_{1\leq i<j\leq n+2}\theta(c_ic_j)\prod_{i=1}^n\theta(x_i^2) \prod_{1\le i<j\le n}a_jx_i^{-1}\,\theta(x_ix_j)\,\theta(x_i/x_j). \end{multline} \quad \quad \qed \end{Theorem} The last elliptic determinant evaluation which I present here is a surprising elliptic extension of a determinant evaluation due to Andrews and Stanton \machSeite{AnStAA}\cite[Theorem~8]{AnStAA} (see \machSeite{KratBN}\cite[Theorem~42]{KratBN}) due to Warnaar \machSeite{WarnAG}\cite[Theorem~4.17]{WarnAG}. It is surprising because in the former there appear $q$-shifted factorials {\it and\/} $q^2$-shifted factorials at the same time, but nevertheless there exists an elliptic analogue, and to obtain it one only has to add the $p$ everywhere in the shifted factorials to convert them to elliptic ones. \begin{Theorem} \label{thm:Warn2} Let $x$ and $y$ be indeterminates. Then, for any non-negative integer $n$, there holds \begin{multline} \label{eq:Warn2} \det_{0\le i,j\le n-1}\(\frac {(y/xq^{i};q^2,p)_{i-j}\, (q/yxq^i;q^2,p)_{i-j}\, (1/x^2q^{2+4i};q^2,p)_{i-j}} {(q;q,p)_{2i+1-j}\, (1/yxq^{2i};q,p)_{i-j}\, (y/xq^{1+2i};q,p)_{i-j}}\)\\ =\prod _{i=0} ^{n-1}\frac {(x^2q^{2i+1};q,p)_i\, (xq^{3+i}/y;q^2,p)_i\, (yxq^{2+i};q^2,p)_i} {(x^2q^{2i+2};q^2,p)_i\, (q;q^2,p)_{i+1}\, (yxq^{1+i};q,p)_i\, (xq^{2+i}/y;q,p)_i}. \end{multline} \quad \quad \qed \end{Theorem} In closing this final subsection, I remind the reader that, as was already said before, many Hankel determinant evaluations involving elliptic functions can be found in \machSeite{MilnAO}\cite{MilnAP} and \machSeite{MilnAP}\cite{MilnAO}. \section*{Acknowledgments} I would like to thank Anders Bj\"orner and Richard Stanley, and the Institut Mittag--Leffler, for giving me the opportunity to work in a relaxed and inspiring atmosphere during the ``Algebraic Combinatorics" programme in Spring 2005 at the Institut, without which this article would never have reached its present form. Moreover, I am extremely grateful to Dave Saunders and Zhendong Wan who performed the {\sl LinBox} computations of determinants of size $3840$, without which it would have been impossible for me to formulate Conjectures~\ref{prob:1}--\ref{prob:3} and \ref{prob:8}. Furthermore I wish to thank Josep Brunat, Adriano Garsia, Greg Kuperberg, Antonio Montes, Yuval Roichman, Hjalmar Rosengren, Michael Schlosser, Guoce Xin, and especially Alain Lascoux, for the many useful comments and discussions which helped to improve the contents of this paper considerably.
{ "timestamp": "2005-08-10T16:39:40", "yymm": "0503", "arxiv_id": "math/0503507", "language": "en", "url": "https://arxiv.org/abs/math/0503507" }
\section{INTRODUCTION} The theory of holonomic modules over the Weyl algebra and more general algebras of differential or $q$-difference operators is becoming increasingly important, both as a crucial part of the general theory of D-modules and in view of various applications (see, for example, \cite{BK,Cart,Gue,S}). Well-known pathological properties of differential operators over fields of positive characteristic make the available, for this case, analogs of the theory of D-modules much more complicated \cite{Bog,Lyub}. More importantly, the resulting structures are not connected with the existing analysis in positive characteristic based on a completely different algebraic foundation. Any non-discrete locally compact field of a positive characteristic $p$ is isomorphic to the field $K$ of formal Laurent series with coefficients from the Galois field $\mathbb F_q$, $q=p^\nu$, $\nu \in \mathbb Z_+$. The field $K$ is endowed with a non-Archimedean absolute value as follows. If $z\in K$, $$ z=\sum\limits_{i=m}^\infty \zeta_ix^i,\quad m\in \mathbb Z,\ \zeta_i\in \mathbb F_q ,\ \zeta_m\ne 0, $$ then $|z|=q^{-m}$. This valuation can be extended onto the field $\overline{K}_c$, the completion of an algebraic closure of $K$. Analysis over $K$ and $\overline{K}_c$, which was initiated in the great paper by Carlitz \cite{Carl} and developed subsequently by Wagner, Goss, Thakur, the author, and many others (see the bibliography in \cite{G2,Th3}) is very different from the classical calculus. An important feature is the availability of many non-trivial additive (actually, $\mathbb F_q$-linear) polynomials and power series of the form $u(t)=\sum\limits a_kt^{q^k}$. Taking into account the fact that the usual factorial $i!$, seen as an element of $K$, vanishes for $i\ge p$, Carlitz introduced the new factorial \begin{equation} D_i=[i][i-1]^q\ldots [1]^{q^{i-1}},\quad [i]=x^{q^i}-x\ (i\ge 1),\ D_0=1, \end{equation} the $\mathbb F_q$-linear logarithm and exponential (which obtained a wide generalization later, in the theory of Drinfeld modules), as well as an important polynomial system, the Carlitz polynomials. Subsequently many other $\mathbb F_q$-linear special functions, such as Thakur's hypergeometric function \cite{Th1,Th2,Th3} and further special polynomial systems, were introduced and investigated. The difference operator \begin{equation} \Delta u(t)=u(xt)-xu(t) \end{equation} introduced in \cite{Carl} became the main ingredient of the $\mathbb F_q$-linear calculus and analytic theory of differential equations over $K$ developed in \cite{K2,K3,K4}. The role of a derivative is played by the $\mathbb F_q$-linear operator $d=\sqrt[q]{}\circ \Delta$ ({\it the Carlitz derivative}). The latter appears also in the $\mathbb F_q$-linear umbral calculus \cite{K5} where an important role belongs to the following new analog of binomial coefficients \begin{equation} \binom{k}{m}_K=\frac{D_k}{D_mD_{k-m}^{q^m}},\quad 0\le m\le k. \end{equation} The meaning of a polynomial coefficient in a differential equation of the above type is not a usual multiplication by a polynomial, but the action of a polynomial in the Frobenius operator $\tau$, $\tau u=u^q$. With this notation, $d=\tau^{-1}\Delta$. The operator $d$ is defined on any $\mathbb F_q$-linear $\overline{K}_c$-valued continuous function; in particular, it decreases by one the ``$\mathbb F_q$-linear degree'' of any $\mathbb F_q$-linear polynomial (see the relation (8) below). The above developments show that in the positive characteristic case a natural counterpart of the Weyl algebra is, for the case of a single variable, the ring $\mathfrak A_1$ generated by $\tau,d$, and scalars from $\overline{K}_c$, with the relations \cite{K1} \begin{equation} d\tau -\tau d=[1]^{1/q},\quad \tau \lambda =\lambda^q\tau ,\quad , d\lambda =\lambda^{1/q}d\ (\lambda \in \overline{K}_c). \end{equation} Some algebraic properties of $\mathfrak A_1$ were studied in \cite{K3} -- it is left and right Noetherian, with no zero divisors. The aim of this paper is to initiate the dimension theory for modules over $\mathfrak A_1$ and more general ``several variable'' rings. The definition of the latter is not straightforward. If, for example, we consider the natural action of the Carlitz derivatives $d_s$ and $d_t$ on an $\mathbb F_q$-linear monomial $f(s,t)=s^{q^m}t^{q^n}$, we notice immediately that $d_s^mf$ is not a polynomial, nor even a holomorphic function in $t$, if $m>n$ (since the action of $d$ is not linear and involves taking the $q$-th root). Moreover, it follows from the relation $d(s^{q^m})=[m]^{1/q}s^{q^{m-1}}$ and the last commutation relation in (4) that $d_s$ and $d_t$ do not commute even on monomials $f$ with $m<n$. A reasonable generalization is inspired by Zeilberger's idea (see \cite{Cart}) to study holonomic properties of sequences of functions making a transform with respect to the discrete variables, which reduces the continuous-discrete case to the purely continuous one (simultaneously in all the variables). In our situation, if $\{ P_k(s)\}$ is a sequence of $\mathbb F_q$-linear polynomials with $\deg P_k\le q^k$, we set \begin{equation*} f(s,t)=\sum \limits_{k=0}^\infty P_k(s)t^{q^k},\tag{$*$} \end{equation*} and $d_s$ is well-defined. In the variable $t$, we consider not $d_t$ but the linear operator $\Delta_t$. The latter does not commute with $d_s$ either, but satisfies the commutation relations $$ d_s\Delta_t-\Delta_td_s=[1]^{1/q}d_s,\quad \Delta_t\tau -\tau \Delta_t=[1]\tau , $$ so that the resulting ring $\mathfrak A_2$ resembles a universal enveloping algebra of a solvable Lie algebra. Similarly we define $\mathfrak A_{n+1}$ for $n>1$. Introducing in $\mathfrak A_{n+1}$ an analog of the Bernstein filtration and considering filtered modules over $\mathfrak A_{n+1}$, we find that basic principles of the theory of algebraic D-modules \cite{Cout} carry over to this case without serious complications. However, the nonlinearity of $\tau$ and $d$ brings new phenomena. In particular, already the ring $\mathfrak A_1$ possesses non-trivial finite-dimensional representations. Therefore an analog of the Bernstein inequality does not hold here without some additional assumptions. In spite of this fact, the notion of a holonomic module (that is a module with the minimal possible GK dimension) seems to have a reasonable sense for the case of $\mathfrak A_{n+1}$-modules. The examples considered in this paper (both for $\mathfrak A_1$-modules and $\mathfrak A_{n+1}$-modules with $n\ge 1$) show that the cases of an anomalously small GK dimension may be seen as degenerate ones. In terms of applications to analysis, it appears that a remarkable phenomenon discovered by Zeilberger (see \cite{Cart}) -- that virtually all important special functions and sequences of classical analysis generate holonomic modules -- is maintained in the positive characteristic case, if a holonomic module is defined as a one with a minimal ``generic'' GK dimension, with degenerate cases excluded. In the author's opinion, such applications provide a sufficient justification for the definition of a quasi-holonomic module given in this paper (Sect. 3.2). Accordingly, the case we study in a greater detail is that of quasi-holonomic submodules of the $\mathfrak A_{n+1}$-module of $\mathbb F_q$-linear functions $u(s,t_1,\ldots ,t_n)$, polynomial in $s$ and holomorphic near the origin in $t_1,\ldots ,t_n$. Following \cite{Cart} we call a function $f$ quasi-holonomic if such is the module $\mathfrak A_{n+1}f$. We prove general conditions for a function $f$ to be quasi-holonomic and verify them for basic objects of this branch of analysis -- the Carlitz polynomials, Thakur's hypergeometric polynomials, and the $K$-binomial coefficients (3), making the above transformation (*) from discrete variables to continuous ones. Considering the $K$-binomial coefficients we use this occasion to prove also the fact that they belong to the ring of integers not only for the field $K$, but for any place of the global function field $\mathbb F_q (x)$. Together with the results of \cite{K5}, this property supports the case for considering the expressions (3) as ``proper'' analogs of the classical binomial coefficients. For other analogs of the latter see \cite{Th3}. \section{The Carlitz Ring} {\bf 2.1.} Denote by $\mathcal F_{n+1}$ the set of all germs of functions of the form \begin{equation} f(s,t_1,\ldots ,t_n)=\sum\limits_{k_1=0}^\infty \ldots \sum\limits_{k_n=0}^\infty \sum\limits_{m=0}^{\min (k_1,\ldots ,k_n)}a_{m,k_1,\ldots ,k_n}s^{q^m}t_1^{q^{k_1}}\ldots t_n^{q^{k_n}} \end{equation} where $a_{m,k_1,\ldots ,k_n}\in \overline{K}_c$ are such that all the series are convergent on some neighbourhoods of the origin. We do not exclude the case $n=0$ where $\mathcal F_1$ will mean the set of all $\mathbb F_q$-linear power series $\sum\limits_ma_ms^{q^m}$ convergent on a neighbourhood of the origin. $\widehat{\mathcal F}_{n+1}$ will denote the set of all polynomials from $\mathcal F_{n+1}$, that is the series (5) in which only a finite number of coefficients is different from zero. The ring $\mathfrak A_{n+1}$ is generated by the operators $\tau ,d_s,\Delta_{t_1},\ldots \Delta_{t_n}$ on $\mathcal F_{n+1}$ defined in the Introduction, and the operators of multiplication by scalars from $\overline{K}_c$. To simplify the notation, we will write $\Delta_j$ instead of $\Delta_{t_j}$ and identify a scalar $\lambda \in \overline{K}_c$ with the operator of multiplication by $\lambda$. The operators $\Delta_j$ are $\overline{K}_c$-linear, so that \begin{equation} \Delta_j\lambda =\lambda \Delta_j,\quad \lambda \in \overline{K}_c , \end{equation} while the operators $\tau ,d_s$ satisfy the commutation relations (4). In the action of each operator $d_s,\Delta_j$ (acting in a single variable), other variables are treated as scalars. The operator $\tau$ acts simultaneously on all the variables and coefficients, so that $$ \tau f=\sum a^q_{m,k_1,\ldots ,k_n}s^{q^{m+1}}t_1^{q^{k_1+1}}\ldots t_n^{q^{k_n+1}}. $$ It follows from (2) that \begin{equation} \Delta_jt_j^{q^k}=\begin{cases} [k]t_j^{q^k}, & \text{if $k\ge 1$};\\ 0, & \text{if $k=0$}; \end{cases} \end{equation} the second equality can be included in the first one, if we set $[0]=0$. Similarly \begin{equation} d_ss^{q^m}=[m]^{1/q}s^{q^{m-1}},\quad m\ge 0. \end{equation} Since $|[m]|=q^{-1}$ for any $m\ge 1$, the action of operators from $\mathfrak A_{n+1}$ does not spoil convergence of the series (5). The identity $[k+1]-[k]^q=[1]$, together with (7) and (8), implies the commutation relations \begin{equation} \Delta_j\tau -\tau \Delta_j=[1]\tau ,\quad d_s\Delta_j-\Delta_jd_s=[1]^{1/q}d_s,\quad j=1,\ldots ,n, \end{equation} verified by applying both sides of each equality to an arbitrary monomial. Using the commutation relations (4), (6), and (9), we can write any element $a\in \mathfrak A_{n+1}$ as a finite sum \begin{equation} a=\sum c_{l,\mu,i_1,\ldots ,i_n}\tau^ld_s^\mu \Delta_1^{i_1}\ldots \Delta_n^{i_n}. \end{equation} \medskip \begin{prop} The representation (10) of an element $a\in \mathfrak A_{n+1}$ is unique. \end{prop} \medskip {\it Proof}. Suppose that \begin{equation} \sum\limits_{l,\mu,i_1,\ldots ,i_n}c_{l,\mu,i_1,\ldots ,i_n}\tau^ld_s^\mu \Delta_1^{i_1}\ldots \Delta_n^{i_n}=0. \end{equation} Applying the left-hand side of (11) to the function $st_1^{q^{k_1}}\ldots t_n^{q^{k_n}}$ with $k_1,\ldots ,k_n>0$ we find that $$ \sum\limits_l\left( \sum\limits_{i_1,\ldots ,i_n}c_{l,0,i_1,\ldots ,i_n} [k_1]^{i_1q^l}\ldots [k_n]^{i_nq^l}\right) s^{q^l}t_1^{q^{k_1+l}}\ldots t_n^{q^{k_n+l}}=0 $$ whence $$ \sum\limits_{i_1,\ldots ,i_n}c_{l,0,i_1,\ldots ,i_n} [k_1]^{i_1q^l}\ldots [k_n]^{i_nq^l}=0 $$ for each $l$. Writing this in the form \begin{equation} \sum\limits_{i_n}\rho (i_n)y^{i_n}=0 \end{equation} where $$ \rho (i_n)=\sum\limits_{i_1,\ldots ,i_{n-1}}c_{l,0,i_1,\ldots ,i_n} [k_1]^{i_1q^l}\ldots [k_{n-1}]^{i_{n-1}q^l},\quad y=[k_n]^{q^l}, $$ and taking into account that (12) holds for arbitrary $k_n\ge 1$, that is for an infinite set of values of $y$, we find that $\rho (i_n)=0$. Repeating this reasoning we get the equality $c_{l,0,i_1,\ldots ,i_n}=0$ for all $l,0,i_1,\ldots ,i_n$. Suppose that $c_{l,\mu ,i_1,\ldots ,i_n}=0$ for $\mu \le \mu_0$ and arbitrary $l,i_1,\ldots ,i_n$. Then we apply the left-hand side of (11) to the function $s^{q^{\mu_0+1}}t_1^{q^{k_1}}\ldots t_n^{q^{k_n}}$ and proceed as before coming to the equality $c_{l,\mu_0+1,i_1,\ldots ,i_n}=0$ for all $l,i_1,\ldots ,i_n$. $\qquad \blacksquare$ \medskip It is easy to prove by induction with respect to $n$ (using the commutation relations (9) and the result from \cite{K3} regarding the case $n=0$) that $\mathfrak A_{n+1}$ has no zero-divisors. \medskip {\bf 2.2.} Let us introduce a filtration in $\mathfrak A_{n+1}$ denoting by $\Gamma_\nu$, $\nu \in \mathbb Z_+$, the $\overline{K}_c$-vector space of operators (10) with $\max \{ l+\mu +i_1+\cdots +i_n\} \le \nu$ where the maximum is taken over all the terms contained in the representation (10). It is clear that $\mathfrak A_{n+1}$ is a filtered ring (for the definitions see \cite{MR}). Setting $T_0=\overline{K}_c$, $T_\nu =\Gamma_\nu /\Gamma_{\nu -1}$, $\nu \ge 1$, we introduce the associated graded ring $$ \gr (\mathfrak A_{n+1})=\bigoplus\limits_{\nu=0}^\infty T_\nu . $$ It is generated by scalars $\lambda \in T_0$ and the images $\bar \tau,\bar d_s,\bar \Delta_1,\ldots ,\bar \Delta_n\in T_1$ of the elements $\tau ,d_s,\Delta_1,\ldots ,\Delta_n\in \Gamma_1$ respectively, which satisfy, by virtue of (4), (6), and (9), the relations \begin{gather*} \bar d_s\bar\tau-\bar\tau \bar d_s=0, \bar\tau \lambda =\lambda^q\bar\tau,\bar d_s\lambda =\lambda^{1/q}\bar d_s,\\ \bar d_s\bar \Delta_j-\bar \Delta_j\bar d_s=0, \bar \Delta_j\bar \tau -\bar\tau \bar \Delta_j=0, \bar \Delta_j\lambda =\lambda \bar \Delta_j\quad (j=1,\ldots ,n). \end{gather*} It is clear that $\mathfrak A_{n+1}$ is a (left and right) almost normalizing extension of the field $\overline{K}_c$ (see Chapter 1, \S 6 in \cite{MR}), so that the rings $\mathfrak A_{n+1}$ and $\gr (\mathfrak A_{n+1})$ are left and right Noetherian. Let us compute the dimension of the $\overline{K}_c$-vector space $\Gamma_\nu$. Note that $$ \dim \Gamma_\nu =\dim \bigoplus\limits_{j=1}^\nu T_j, $$ so that $\dim \Gamma_\nu$ coincides with the dimension of the appropriate space appearing in the natural filtration in $\gr (\mathfrak A_{n+1})$. \medskip \begin{lem} For any $\nu \in \mathbb N$ $$ \dim \Gamma_\nu = \binom{\nu +n+2}{n+2}. $$ \end{lem} \medskip {\it Proof}. The number $\dim \Gamma_\nu$ coincides with the number of non-negative integral solutions $(l,\mu ,i_1,\ldots ,i_n)$ of the inequality $l+\mu +i_1+\cdots +i_n\le \nu$, so that $$ \dim \Gamma_\nu =\sum\limits_{j=0}^\nu N(j,n+2) $$ where $N(j,k)$ is the number of different representations of $j$ as sums of $k$ non-negative integers. It is known (Proposition 6.1 in \cite{Lan}) that $N(j,k)=\dbinom{j+k-1}{k-1}$. Then (see Sect. 1.3 from \cite{Ri}) $$ \dim \Gamma_\nu =\sum\limits_{j=0}^\nu \binom{j+n+1}{n+1}= \sum\limits_{i=0}^\nu \binom{\nu+n+1-i}{n+1}=\binom{\nu +n+2}{n+2}, $$ as desired. $\qquad \blacksquare$ \section{Filtered Modules} {\bf 3.1.} Let $M$ be a left module over the Carlitz ring $\mathfrak A_{n+1}$. Suppose we have a filtration $\{ \mathfrak M_j\}$ of $M$, that is \begin{equation} \mathfrak M_0\subset \mathfrak M_1\subset \ldots \subset M,\quad M=\bigcup\limits_{j\ge 0}\mathfrak M_j, \end{equation} and $\Gamma_\nu \mathfrak M_j\subset \mathfrak M_{\nu +j}$ for any $\nu ,j\in \mathbb Z_+$. We assume that each $\mathfrak M_j$ is a finite-dimensional vector space over $\overline{K}_c$. Below we write $\mathfrak M_j=\{ 0\}$ and $\Gamma_\nu =\{ 0\}$ if $j<0$ and $\nu <0$. In a standard way \cite{Cout} we define the graded module $$ \gr (M)=\bigoplus\limits_{j\ge 0}\left( \mathfrak M_j/\mathfrak M_{j-1}\right) $$ over $\gr (\mathfrak A_{n+1})$, associated with the filtration (13). As usual, the filtration (13) is called {\it good}, if $\gr (M)$ is finitely generated. Main properties of filtered modules over the Weyl algebra (see \cite{Bj,Cout}) carry over to our situation without any substantial changes, both in their formulations and proofs. In fact, the only technical difference is that the operators $\tau$ and $d_s$ are semilinear, not linear. However, as it is explained in Appendix I to Chapter 2 of \cite{Bour}, basic notions of linear algebra remain valid for semilinear mappings -- a semilinear mapping of a vector space into itself can be interpreted as a linear mapping between two different vector spaces, and, for instance, dimensions of the kernel and cokernel are not changed in this interpretation. Note that everywhere in this paper we consider vector spaces over the algebraically closed field $\overline{K}_c$, on which $\tau$ induces an automorphism. Below, as before, $\dim$ means the dimension over $\overline{K}_c$. In particular, for a good filtration there exist a polynomial $\chi \in \mathbb Q[t]$ and a number $N\in \mathbb N$, such that $$ \dim \mathfrak M_s=\sum\limits_{i=0}^s\dim (\mathfrak M_i/\mathfrak M_{i-1})=\chi (s)\text{ for }s\ge N. $$ The number $d(M)=\deg \chi$, called the (Gelfand-Kirillov) {\it dimension} of $M$, and the leading coefficient of $\chi$ multiplied by $d(M)!$, called the {\it multiplicity} $m(M)$ of $M$, do not depend on the choice of a good filtration on $M$. A filtration $\{ \mathfrak M_i\}$ is good if and only if there exists such $k_0\in \mathbb N$ that $$ \mathfrak M_{i+k}=\Gamma_i\mathfrak M_k\text{ \ for all }k\ge k_0. $$ If $N$ and $M/N$ are a submodule and the corresponding quotient module, with the induced filtrations, then $d(M)=\max \{ d(N),d(M/N)\}$, and if $d(N)=d(M/N)$, then $m(M)=m(N)+m(M/N)$. For a direct sum $M=M_1\oplus \cdots \oplus M_k$ we have $d(M)=\max \{ d(M_1),\ldots ,d(M_k)\}$. In particular, if we consider $\mathfrak A_{n+1}$ as a left module over itself, then by Lemma 1 \begin{equation} d(\mathfrak A_{n+1})=n+2,\quad m(\mathfrak A_{n+1})=1. \end{equation} It follows from (14) and the above general facts that for any finitely generated left $\mathfrak A_{n+1}$-module \begin{equation} d(M)\le n+2. \end{equation} By (14), the bound in (15) in general cannot be improved. However, if $I$ is a non-zero left ideal in $\mathfrak A_{n+1}$, then \begin{equation} d(\mathfrak A_{n+1}/I)\le n+1. \end{equation} The proof of (16) is identical to the proof of Corollary 9.3.5 from \cite{Cout}. \medskip {\bf 3.2.} Let us consider the set $\widehat{\mathcal F}_{n+1}$ of polynomials (5) as a $\mathfrak A_{n+1}$-module. A filtration $$ \mathcal F^{(0)}_{n+1}\subset \mathcal F^{(1)}_{n+1}\subset \ldots \subset \widehat{\mathcal F}_{n+1} $$ can be introduced by setting $\mathcal F^{(j)}_{n+1}$ to be the collection of all the polynomials (5), in which the maximal indices $k_1,\ldots ,k_n$ corresponding to non-zero coefficients $a_{m,k_1,\ldots ,k_n}$ do not exceed $j$. This filtration is obviously good. \medskip \begin{prop} For the module $\widehat{\mathcal F}_{n+1}$, \begin{equation} d\left( \widehat{\mathcal F}_{n+1}\right) =n+1,\quad m\left( \widehat{\mathcal F}_{n+1}\right) =n! \end{equation} \end{prop} \medskip {\it Proof}. Let us compute $\dim \mathcal F^{(j)}_{n+1}$. For a fixed $\mu$, the quantity of $n$-tuples $(k_1,\ldots ,k_n)$ of non-negative integers, for which $\min (k_1,\ldots ,k_n)=\mu$, is added up from those $n$-tuples where $i$ numbers are equal to $\mu$ while $n-i$ numbers are strictly larger and can take $j-\mu$ values. Therefore the above quantity equals $\sum\limits_{i=1}^n\dbinom{n}{i}(j-\mu )^{n-i}$. Next, $\mu +1$ possible values of $m$ in (5) correspond to each $n$-tuple. Thus, $$ \dim \mathcal F^{(j)}_{n+1}=\sum\limits_{\mu =0}^j(\mu +1) \sum\limits_{i=1}^n\dbinom{n}{i}(j-\mu )^{n-i}=\sum\limits_{\mu =0}^j(\mu +1)\left\{ (j-\mu +1)^n-(j-\mu )^n\right\} . $$ Denote $r_\mu =(j-\mu +1)^n-(j-\mu )^n$, $R_i=r_0+r_1+\cdots +r_i=(j+1)^n-(j-i)^n$. Performing the Abel transformation we get \begin{multline*} \dim \mathcal F^{(j)}_{n+1}=(j+1)R_j-\sum\limits_{i=0}^{j-1}R_i =(j+1)^{n+1}-j(j+1)^n+\sum\limits_{i=0}^{j-1}(j-i)^n\\ =(j+1)^n+\sum\limits_{k=1}^jk^n=(j+1)^n+S_n(j+1) \end{multline*} where $S_n(N)=1^n+2^n+\cdots +(N-1)^n$. It is known (\cite{IR}, Chapter 15) that $$ S_n(N)=\frac{1}{n+1}\sum\limits_{k=0}^n\binom{n+1}{k}B_kN^{n+1-k} $$ where $B_k$ are the Bernoulli numbers. Therefore we find that $$ \dim \mathcal F^{(j)}_{n+1}=\frac{(j+1)^{n+1}}{n+1}+P_n(j) $$ where $P_n$ is a polynomial of the degree $n$. This implies (17). $\qquad \blacksquare$ \medskip It is natural to call an $\mathfrak A_{n+1}$-module $M$ {\it quasi-holonomic} if $d(M)=n+1$. Thus, $\widehat{\mathcal F}_{n+1}$ is an example of a quasi-holonomic module. \medskip {\bf 3.3.} Let us look at possible values of $d(M)$ for $\mathfrak A_1$-modules. The next result demonstrates a sharp difference from the case of modules over the Weyl algebras. \medskip \begin{teo} \begin{description} \item{{\rm (i)}} For any $k=1,2,\ldots$, there exists such a nontrivial $\mathfrak A_1$-module $M$ that $\dim M=k$ ($\dim$ means the dimension over $\overline{K}_c$), that is $d(M)=0$. \item{{\rm (ii)}} Let $M$ be a finitely generated $\mathfrak A_1$-module with a good filtration. Suppose that there exists a ``vacuum vector'' $v\in M$, such that $d_sv=0$ and $\tau^m(v)\ne 0$ for all $m=0,1,2,\ldots$. Then $d(M)\ge 1$. \end{description} \end{teo} \medskip {\it Proof}. (i) Let $M=(\overline{K}_c )^k$. Denote by $\mathbf e_1,\ldots ,\mathbf e_k$ the standard basis in $M$, that is $\mathbf e_j=(0,\ldots ,0,1,0,\ldots ,0)$, with 1 at the $j$-th place. Let $(\lambda_{ij})$ be a $k\times k$ matrix over $\overline{K}_c$, such that $\lambda_{ij}\in \mathbb F_q$ if $i\ne j$, while the diagonal elements satisfy the equation $\lambda^q-\lambda +[1]^{1/q}=0$. We define the action of $\tau$ and $d_s$ on $M$ as follows: $$ \tau (c\mathbf e_j)=c^q\mathbf e_j;\ d_s(\mathbf e_j)=\sum\limits_{i=1}^n\lambda_{ij}\mathbf e_i;\ d_s(c\mathbf e_j)=c^{1/q}\mathbf e_j,\quad c\in \overline{K}_c ,j=1,\ldots ,k, $$ with subsequent additive continuation onto $M$. If $x=\sum\limits_{j=1}^kc_j\mathbf e_j$, $c_j\in \overline{K}_c$, then we have $$ \tau d_s(x)=\sum\limits_{j=1}^kc_j\sum\limits_{i=1}^n\lambda_{ij}^q\mathbf e_i,\quad d_s\tau (x)=\sum\limits_{j=1}^kc_j\sum\limits_{i=1}^n\lambda_{ij}\mathbf e_i, $$ so that $$ d_s\tau (x)-\tau d_s(x)=[1]^{1/q}x, $$ and we have indeed an $\mathfrak A_1$-module. (ii) It follows from the relation $[d_s,\tau^m]=[m]^{1/q}\tau^{m-1}$ (see \cite{K3}) that $$ d_s\tau^mv=[m]^{1/q}\tau^{m-1}v,\quad m=1,2,\ldots , $$ that is $\tau^{m-1}v$ is an eigenvector of a linear operator $d_s\tau$ on $M$ (considered as a $\overline{K}_c$-vector space) corresponding to the eigenvalue $[m]^{1/q}$. Therefore the vectors $\tau^{m-1}v$ are linearly independent. It follows from the existence of the Hilbert polynomial $\chi$ implementing the dimension $d(M)$ that $d(M)\ge 1$. $\qquad \blacksquare$ \medskip \section{Holonomic Functions} {\bf 4.1.} Let $0\ne f\in \mathcal F_{n+1}$, $$ I_f=\left\{ \varphi \in \mathfrak A_{n+1}:\ \varphi (f)=0\right\} . $$ $I_f$ is a left ideal in $\mathfrak A_{n+1}$. The left $\mathfrak A_{n+1}$-module $M_f=\mathfrak A_{n+1}/I_f$ is isomorphic to the submodule $\mathfrak A_{n+1}f\subset \mathcal F_{n+1}$ -- an element $\varphi (f)\in \mathfrak A_{n+1}f$ corresponds to the class of $\varphi \in \mathfrak A_{n+1}$ in $M_f$. A natural good filtration in $M_f$ is induced from that in $\mathfrak A_{n+1}$ -- the subspace $\mathfrak M_j$ is generated by elements $\tau^ld_s^\mu \Delta_1^{i_1}\ldots \Delta_n^{i_n}f$ with $l+\mu +i_1+\cdots +i_n\le j$. As we know (see (16)), if $I_f\ne \{ 0\}$, then $d(M_f)\le n+1$. We call a function $f$ {\it quasi-holonomic} if the module $M_f$ is quasi-holonomic, that is $d(M_f)=n+1$. The condition $I_f\ne \{ 0\}$ means that $f$ is a solution of a ``differential equation'' $\varphi (f)=0$, $\varphi \in \mathfrak A_{n+1}$. For $n=0$, we have the following easy result. \medskip \begin{teo} If a non-zero function $f\in \mathcal F_1$ satisfies an equation $\varphi (f)=0$, $0\ne \varphi \in \mathfrak A_1$, then $f$ is quasi-holonomic. \end{teo} \medskip {\it Proof}. It is sufficient to show that $\dim M_f=\infty$. In fact, the sequence $\left\{ \tau^lf\right\}_{l=0}^\infty$ is linearly independent because otherwise we would have such a finite collection of elements $c_0,c_1,\ldots ,c_N\in \overline{K}_c$, some of which are different from zero, that \begin{equation} c_0f(s)+c_1f^q(s)+\cdots +c_Nf^{q^N}(s)=0 \end{equation} for all $s$ from a neighbourhood of the origin in $\overline{K}_c$. It follows from (18) that $f$ takes only a finite number of values. By the uniqueness theorem for non-Archimedean holomorphic functions, $f(s)\equiv \text{const}$ on some neighbourhood of the origin. Due to the $\mathbb F_q$-linearity, $f(s)\equiv 0$, and we have come to a contradiction. $\qquad \blacksquare$ \medskip In particular, any $\mathbb F_q$-linear polynomial of $s$ is quasi-holonomic, since it is annihilated by $d_s^m$, with a sufficiently large $m$. \medskip {\bf 4.2.} If $n>0$, the situation is more complicated. We call the module $M_f$ (and the corresponding function $f$) {\it degenerate} if $d(M_f)<n+1$ (by the Bernstein inequality, there is no degeneracy phenomena for modules over the complex Weyl algebra). We give an example of degeneracy for the case $n=1$. Let $f(s,t_1)=g(st_1)\in \mathcal F_2$ where the function $g$ belongs to $\mathcal F_1$ and satisfies an equation $\varphi (g)=0$, $\varphi \in \mathfrak A_1$. Then $f$ is degenerate. Indeed, by the general rule, $\mathfrak M_j$ is spanned by elements $\tau^ld_s^\mu \Delta_1^{i_i}f$ with $l+\mu +i_1\le j$. In the present situation, $$ \Delta_1f=g(xst_1)-xg(st_1)=\tau d_sg, $$ so that an element $\tau^ld_s^\mu \Delta_1^{i_i}f$ is a linear combination of elements $\left( \tau^{l+\lambda}d_s^{\mu +\nu}g\right) (s,t)$ with $\lambda \le i_1$, $\nu \le i_1$. Therefore $\mathfrak M_j$ is contained in the linear hull of elements $\tau^kd_s^mg$, $k+m\le 2j$. By Theorem 2, the $\overline{K}_c$-dimension of the latter does not exceed a linear function of $2j$, so that $d(M_f)\le 1$. On the other hand, since, as in the proof of Theorem 2, the system of functions $\left\{ \tau^lf\right\}_{l=0}^\infty$ is linearly independent, we find that $d(M_f)=1$. In order to exclude the degenerate case, we introduce the notion of a non-sparse function. A function $f\in \mathcal F_{n+1}$ of the form (5) is called {\it non-sparse} if there exists such a sequence $m_l\to \infty$ that, for any $l$, there exist sequences $k_1^{(i)},k_2^{(i)},\ldots ,k_n^{(i)}\ge m_l$ (depending on $l$), such that $k_\nu^{(i)}\to \infty$ as $i\to \infty$ ($\nu =1,\ldots ,n$), and $a_{m,k_1^{(i)},\ldots ,k_n^{(i)}}\ne 0$. \medskip \begin{lem} If a function $f$ is non-sparse, then the system of functions $(\tau d_s)^\lambda \Delta_1^{j_1}\ldots \Delta_n^{j_n}f$ ($\lambda ,j_1,\ldots ,j_n=0,1,2,\ldots$) is linearly independent over $\overline{K}_c$. \end{lem} \medskip {\it Proof}. Suppose that \begin{equation} \sum\limits_{\lambda =0}^\Lambda \sum\limits_{j_1=0}^{J_1}\ldots \sum\limits_{j_n=0}^{J_n}c_{\lambda ,j_1,\ldots ,j_n} (\tau d_s)^\lambda \Delta_1^{j_1}\ldots \Delta_n^{j_n}f=0 \end{equation} for some $c_{\lambda ,j_1,\ldots ,j_n}\in \overline{K}_c$, $\Lambda ,J_1,\ldots ,J_n\in \mathbb N$. Substituting (5) into (19) and collecting coefficients of the power series we find that \begin{equation} \sum\limits_{\lambda =0}^\Lambda \sum\limits_{j_1=0}^{J_1}\ldots \sum\limits_{j_n=0}^{J_n}c_{\lambda ,j_1,\ldots ,j_n} [m_l]^\lambda [k_1^{(i)}]^{j_1}\ldots[k_n^{(i)}]^{j_n}=0 \end{equation} for all $l,i$. We see from (20) that the polynomial $$ \sum\limits_{j_n=0}^{J_n}\left\{ \sum\limits_{\lambda =0}^\Lambda \sum\limits_{j_1=0}^{J_1}\ldots \sum\limits_{j_{n-1}=0}^{J_{n-1}}c_{\lambda ,j_1,\ldots ,j_n} [m_l]^\lambda [k_1^{(i)}]^{j_1}\ldots[k_{n-1}^{(i)}]^{j_{n-1}} \right\} z^{j_n} $$ has an infinite sequence of different roots, so that $$ \sum\limits_{\lambda =0}^\Lambda \sum\limits_{j_1=0}^{J_1}\ldots \sum\limits_{j_{n-1}=0}^{J_{n-1}}c_{\lambda ,j_1,\ldots ,j_n} [m_l]^\lambda [k_1^{(i)}]^{j_1}\ldots[k_{n-1}^{(i)}]^{j_{n-1}}=0 $$ for all $l,i$, and for each $j_n=0,1,\ldots ,J_n$. Repeating this reasoning we find that all the coefficients $c_{\lambda ,j_1,\ldots ,j_n}$ are equal to zero. $\qquad\blacksquare$ \medskip Now the above arguments regarding $d(M_f)$ yield the following result. \medskip \begin{teo} If a function $f$ is non-sparse, then $d(M_f)\ge n+1$. If, in addition, $f$ satisfies an equation $\varphi (f)=0$, $0\ne \varphi \in \mathfrak A_{n+1}$, then $f$ is quasi-holonomic. \end{teo} \medskip As in the classical situation, one can construct quasi-holonomic functions by addition. \medskip \begin{prop} If the functions $f,g\in \mathcal F_{n+1}$ are quasi-holonomic, and $f+g$ is non-sparse, then $f+g$ is quasi-holonomic. \end{prop} \medskip {\it Proof}. Consider the $\mathfrak A_{n+1}$-module $M_2=(\mathfrak A_{n+1}f)\oplus (\mathfrak A_{n+1}g)$. Since $f$ and $g$ are both quasi-holonomic, we have $d(M_2)=n+1$. Next, let $N_2$ be a submodule of $M_2$ consisting of such pairs $(\varphi (f),\varphi (g))$ that $\varphi (f)+\varphi (g)=0$. Then $d(M_2)=\max \{ d(N_2),d(M_2/N_2)\}$, so that $d(M_2/N_2)\le n+1$. On the other hand, we have an injective mapping $\mathfrak A_{n+1}(f+g)\to M_2/N_2$, which maps $\varphi (f+g)$ to the image of $(\varphi (f),\varphi (g))$ in $M_2/N_2$. Therefore $d(\mathfrak A_{n+1}(f+g))\le d(M_2/N_2)\le n+1$. It remains to use Theorem 3. $\qquad \blacksquare$ \medskip {\bf 4.3.} We use Theorem 3 to prove that the functions (5) obtained via the sequence-to-function transform ($*$) or its multi-index generalizations, from some well-known sequences of polynomials over $K$ are quasi-holonomic. a) {\it The Carlitz polynomials}. The sequence $$ f_k(s)=D_k^{-1}\prod \limits _{\genfrac{}{}{0pt}{1}{m\in \mathbb F_q [x]}{\deg m<k}}(s-m) \quad (k\ge 1),\quad f_0(s)=s, $$ of normalized Carlitz polynomials forms an orthonormal basis of the space of all $\mathbb F_q$-linear continuous functions on the ring of integers of the field $K$. Its transform ($*$), the function \begin{equation} C_s(t)=\sum\limits_{k=0}^\infty f_k(s)t^{q^k} \end{equation} called the {\it Carlitz module}, is one of the main objects of the function field arithmetic \cite{G2,Th3}. It is known \cite{Carl,G1} that $$ f_k(s)=\sum\limits_{i=0}^k\frac{(-1)^{k-i}}{D_iL_{k-i}^{q^i}}s^{q^i} $$ where $L_i=[i][i-1]\ldots [1]$ ($i\ge 1$), $L_0=1$. By (1), we have \begin{equation} |D_i|=q^{-\frac{q^i-1}{q-1}},\quad |L_i|=q^{-i}, \end{equation} so that $$ \left| D_iL_{k-i}^{q^i}\right| =q^{-\left( \frac{q^i-1}{q-1}+(k-i)q^i\right) },\quad 0\le i\le k. $$ For large values of $k$, an elementary investigation of the function $z\mapsto (k-z)q^z$, $z\le k$, shows that $$ \max\limits_{0\le i\le k}(k-i)q^i\le \alpha q^k,\quad \alpha >0, $$ so that $$ |f_k(s)|\le q^{\alpha q^k} $$ for all $s\in \overline{K}_c$ with $|s|\le q^{-1}$. Therefore the series (21) converges for small $|t|$, so that the Carlitz module function belongs to $\mathcal F_2$. Since $d_sf_i=f_{i-1}$ for $i\ge 1$, and $d_sf_0=0$ \cite{G1}, we see that $d_sC_s(t)=C_s(t)$. Clearly, the function $C_s(t)$ is non-sparse. Therefore the Carlitz module function is quasi-holonomic, jointly in both its variables. \medskip b) {\it Thakur's hypergeometric polynomials}. We consider the polynomial case of Thakur's hypergeometric function \cite{Th1,Th2,Th3}: \begin{equation} {}_lF_\lambda (-a_1,\ldots ,-a_l;-b_1,\ldots ,-b_\lambda ;z)=\sum\limits_m\frac{(-a_1)_m\ldots (-a_l)_m}{(-b_1)_m\ldots (-b_\lambda )_mD_m}z^{q^m} \end{equation} where $a_1,\ldots ,a_l,b_1,\ldots ,b_\lambda \in\mathbb Z_+$, \begin{equation} (-a)_m=\begin{cases} (-1)^{a-m}L_{a-m}^{-q^m}, & \text{if $m\le a$},\\ 0, & \text{if $m>a$},\end{cases},\quad a\in \mathbb Z_+. \end{equation} It is seen from (24) that the terms in (23), which make sense and do not vanish, are those with $m\le \min (a_1,\ldots ,a_l,b_1,\ldots ,b_\lambda )$. Let \begin{multline} f(s,t_1,\ldots ,t_l,u_1,\ldots ,u_\lambda )\\ = \sum\limits_{k_1=0}^\infty \ldots \sum\limits_{k_l=0}^\infty \sum\limits_{\nu_1=0}^\infty \ldots \sum\limits_{\nu_\lambda =0}^\infty {}_lF_\lambda (-k_1,\ldots ,-k_l;-\nu_1,\ldots ,-\nu_\lambda ;s) t_1^{q^{k_1}}\ldots t_l^{q^{k_l}}u_1^{q^{\nu_1}}\ldots u_\lambda^{q^{\nu_\lambda}}. \end{multline} We prove as above that all the series in (25) converge near the origin. Thus, $f\in \mathcal F_{l+\lambda +1}$. It is known (\cite{Th3}, Sect. 6.5) that \begin{equation} d_s{}_lF_\lambda (-k_1,\ldots ,-k_l;-\nu_1,\ldots ,-\nu_\lambda ;s)={}_lF_\lambda (-k_1+1,\ldots ,-k_l+1;-\nu_1+1,\ldots ,-\nu_\lambda +1;s) \end{equation} if all the parameters $k_1,\ldots ,k_l,\nu_1,\ldots ,\nu_\lambda$ are different from zero. If at least one of them is equal to zero, then the left-hand side of (26) equals zero. This property implies the identity $d_sf=f$, the same as that for the Carlitz module function. Since $f$ is non-sparse, it is quasi-holonomic. In the next section we will see that the $K$-binomial coefficients (3) correspond to a quasi-holonomic function satisfying a more complicated equation containing also the operator $\Delta_t$. \section{$K$-Binomial Coefficients} {\bf 5.1.} Let us consider the $K$-binomial coefficients (3). It follows from (22) that $$ \left| \binom{k}{m}_K\right| =1,\quad 0\le m\le k. $$ Since $\binom{k}{m}_K\in \mathbb F_q (x)$, it is natural to consider also other places of $\mathbb F_q (x)$, that is other non-equivalent absolute values on $\mathbb F_q (x)$. It is well known (\cite{Weil}, Sect. 3.1) that they are parametrized by monic irreducible polynomials $\pi \in \mathbb F_q [x]$. The absolute value $|t|_\pi$, $t\in \mathbb F_q (x)$, is defined as follows. We write $t=\pi^\nu \alpha /\alpha'$ where $m\in \mathbb Z$, $\alpha ,\alpha'\in \mathbb F_q [x]$, and $\pi$ does not divide $\alpha ,\alpha'$. Then $|t|_\pi =|\pi |_\pi^\nu$, $|\pi |_\pi =q^{-\delta}$ where $\delta =\deg \pi$; as usual, $|0|_\pi =0$. The absolute value $|\cdot |$ used elsewhere in this paper corresponds to $\pi (x)=x$. \medskip \begin{prop} For any monic irreducible polynomial $\pi \in \mathbb F_q[x]$, the $K$-binomial coefficients (3) satisfy the inequality $$ \left| \binom{k}{m}_K\right|_\pi \le 1,\quad 0\le m\le k. $$ \end{prop} \medskip {\it Proof}. First we compute $|D_m|_\pi$. It follows from Lemma 2.13 of \cite{LN} that $$ |[i]|_\pi =\begin{cases} q^{-\delta }, & \text{if $\delta$ divides $i$},\\ 1, & \text{otherwise}.\end{cases} $$ Writing $m=j\delta +i$, with $i,j\in \mathbb Z_+$, $0\le i<\delta$, we find that \begin{multline*} |D_m|_\pi =|[j\delta ]|_\pi^{q^i}|[(j-1)\delta ]|_\pi^{q^{\delta +i}}\ldots |[\delta ]|_\pi^{q^{(j-1)\delta +i}}=\left\{ q^{-\delta }\cdot \left( q^{-\delta}\right)^{q^\delta }\cdot \ldots \cdot \left( q^{-\delta}\right)^{q^{(j-1)\delta }}\right\}^{q^i} \\ =\left\{ \left( q^{-\delta}\right)^{1+q^\delta +\cdots +q^{(j-1)\delta}}\right\}^{q^i}=q^{-\delta q^i\frac{q^{j\delta } -1}{q^\delta -1}}. \end{multline*} Similarly we can write $k-m=\varkappa \delta +\lambda$, with $\varkappa ,\lambda \in \mathbb Z_+$, $0\le \lambda <\delta$, and get that $$ |D_{k-m}|=q^{-\delta q^\lambda \frac{q^{\varkappa \delta }-1}{q^\delta -1}}. $$ If $i+\lambda <\delta$, then we obtain a similar representation for $k$ simply by adding those for $m$ and $k-m$, so that \begin{multline*} \log_q\left| \binom{k}{m}_K\right|_\pi =-\frac{\delta}{q^\delta -1}\left\{ q^{i+\lambda }\left( q^{(j+\varkappa )\delta }-1\right) -q^i\left( q^{j\delta }-1\right) -q^\lambda \left( q^{\varkappa \delta }-1\right) q^{j\delta +i}\right\} \\ =-\frac{\delta}{q^\delta -1}q^i\left( 1+q^{\lambda +j\delta }-q^\lambda -q^{j\delta }\right) =-\frac{\delta}{q^\delta -1}q^i\left( q^\lambda -1\right) \left( q^{j\delta }-1\right) \le 0. \end{multline*} If $i+\lambda \ge \delta$, then $k=(j+\varkappa +1)\delta +\nu$ where $0\le \nu =i+\lambda -\delta <\delta$. In this case \begin{multline*} \log_q\left| \binom{k}{m}_K\right|_\pi =-\frac{\delta}{q^\delta -1}\left\{ q^\nu \left( q^{(j+\varkappa +1)\delta }-1\right) -q^i\left( q^{j\delta }-1\right) -q^\lambda \left( q^{\varkappa \delta }-1\right) q^{j\delta +i}\right\} \\ =-\frac{\delta}{q^\delta -1}\left( q^i+q^{\lambda +j\delta +i }-q^{i+j\delta }-q^\nu \right) <0, \end{multline*} since $\nu <i+\lambda$. $\qquad \blacksquare$ \medskip Below we will use only the valuation with $\pi (x)=x$, that is, as above, consider the field $K$. \medskip {\bf 5.2.} Let us derive, for the $K$-binomial coefficients (3), analogs of the classical Pascal and Vandermonde identities. \medskip \begin{prop} The identity \begin{equation} \binom{k}{m}_K=\binom{k-1}{m-1}_K^q+\binom{k-1}{m}_K^qD_m^{q-1} \end{equation} holds, if $0\le m\le k$ and it is assumed that $\dbinom{k}{-1}_K=\dbinom{k-1}{k}_K=0$. \end{prop} \medskip {\it Proof}. Let $e_m(t)=D_mf_m(t)$ be the ``non-normalized'' Carlitz polynomials. They satisfy the main $K$-binomial identity \cite{Carl,K5} \begin{equation} e_k(st)=\sum\limits_{m=0}^k\binom{k}{m}_Ke_m(s)\left\{ e_{k-m}(t)\right\}^{q^m}, \end{equation} which holds, for example, for any $s,t\in \mathbb F_q [x]$. It is known \cite{Carl,G1} that \begin{equation} e_k=e_{k-1}^q-D_{k-1}^{q-1}e_{k-1}. \end{equation} Let us rewrite the left-hand side of (28) in accordance with (29), and apply to each term the identity (28) with $k-1$ substituted for $k$. We have $$ e_{k-1}^q(st)=\sum\limits_{i=0}^{k-1}\binom{k-1}{i}_K^qe_i^q(s) e_{k-i-1}^{q^{i+1}}(t). $$ By (29), $e_i^q=e_{i+1}+D_i^{q-1}e_i$, $e_{k-i-1}^q=e_{k-i}+D_{k-i-1}^{q-1}e_{k-i-1}$, whence \begin{multline*} e_{k-1}^q(st) =\sum\limits_{j=1}^k\binom{k-1}{j-1}_K^qe_j(s)e_{k-j}^{q^j}(t) +\sum\limits_{i=0}^{k-1}\binom{k-1}{i}_K^q D_i^{q-1}e_i(s)e_{k-i}^{q^i}(t)\\ +\sum\limits_{i=0}^{k-1}\binom{k-1}{i}_K^q D_i^{q-1}D_{k-i-1}^{q^i(q-1)}e_i(s)e_{k-i-1}^{q^i}(t). \end{multline*} Note that \begin{equation} \binom{k-1}{i}_K^qD_i^{q-1}D_{k-i-1}^{q^i(q-1)}=D_{k-1}^{q-1}\binom{k-1}{i}_K. \end{equation} Indeed, the left-hand side of (30) equals $$ \frac{D_{k-1}^q}{D_i^qD_{k-i-1}^{q^{i+1}}}D_i^{q-1}D_{k-i-1}^{q^{i+1}-q^i}= \frac{D_{k-1}}{D_iD_{k-i-1}^{q^i}}D_{k-1}^{q-1} $$ and coincides with the right-hand side. Therefore the last sum in the expression for $e_{k-1}^q(st)$ equals $$ D_{k-1}^{q-1}\sum\limits_{i=0}^{k-1}\binom{k-1}{i}_Ke_i(s)e_{k-i-1}^{q^i}(t) =D_{k-1}^{q-1}e_{k-1}(st). $$ Using (29) again we find that $$ e_k(st)=\sum\limits_{i=0}^k\binom{k-1}{i-1}_K^qe_i(s)e_{k-i}^{q^i}(t) +\sum\limits_{i=0}^k\binom{k-1}{i}_K^qD_i^{q-1}e_i(s)e_{k-i}^{q^i}(t), $$ and the comparison with (28) yields $$ \sum\limits_{m=0}^k\left\{ \binom{k}{m}_K-\binom{k-1}{m-1}_K^q- \binom{k-1}{m}_K^qD_m^{q-1}\right\} e_m(s)e_{k-m}^{q^m}(t)=0 $$ for any $s,t$. Since the Carlitz polynomials are linearly independent, we obtain that $$ \left\{ \binom{k}{m}_K-\binom{k-1}{m-1}_K^q- \binom{k-1}{m}_K^qD_m^{q-1}\right\} e_{k-m}^{q^m}(t)=0 $$ for any $t$, and it remains to note that $e_{k-m}(t)\ne 0$ if $t\in \mathbb F_q [x]$, $\deg t\ge k$, by the definition of the Carlitz polynomials. $\qquad \blacksquare$ \medskip More generally, we have the following Vandermonde-type identity. Let $k,m$ be integers, $0\le m\le k$. \begin{prop} Define $c_{li}^{(m)}\in K$ by the recurrent relation \begin{equation} c_{l+1,i}^{(m)}=c_{l,i-1}^{(m)}+c_{li}^{(m)}D_{m-i}^{q-1} \end{equation} and the initial conditions $c_{li}^{(m)}=0$ for $i<0$ and $i>l$, $c_{00}^{(m)}=1$. Then, for any $l\le m$, \begin{equation} \binom{k}{m}_K=\sum\limits_{i=0}^lc_{li}^{(m)}\binom{k-l}{m-i}_K^{q^l}. \end{equation} \end{prop} \medskip {\it Proof}. The identity (32) is trivial for $l=0$. Suppose it has been proved for some $l$. Let us transform the right-hand side of (32) using the identity (27). Then we have \begin{multline*} \binom{k}{m}_K=\sum\limits_{i=0}^lc_{li}^{(m)}\binom{k-l-1}{m-i-1}_K^{q^{l+1}} +\sum\limits_{i=0}^lc_{li}^{(m)}\binom{k-l-1}{m-i}_K^{q^{l+1}}D_{m-i}^{q-1}\\ =\sum\limits_{j=1}^{l+1}c_{l,j-1}^{(m)}\binom{k-l-1}{m-j}_K^{q^{l+1}} +\sum\limits_{i=0}^lc_{li}^{(m)}\binom{k-l-1}{m-i}_K^{q^{l+1}}D_{m-i}^{q-1}. \end{multline*} Since we assume that $c_{l,-1}^{(m)}=c_{l,l+1}^{(m)}=0$, the summation in both the above sums can be performed from 0 to $l+1$. Using (31) we obtain the required identity (32) with $l+1$ substituted for $l$. $\qquad \blacksquare$ \medskip {\bf 5.3.} Now we consider a function $f\in \mathcal F_2$ associated with the $K$-binomial coefficients, that is \begin{equation} f(s,t)=\sum\limits_{k=0}^\infty \sum\limits_{m=0}^k\binom{k}{m}_Ks^{q^m}t^{q^k}. \end{equation} Obviously, $f$ is non-sparse. \medskip \begin{prop} The function (33) satisfies the equation \begin{equation} d_sf(s,t)=\Delta_tf(s,t)+[1]^{1/q}f(s,t), \end{equation} so that $f$ is quasi-holonomic. \end{prop} \medskip {\it Proof}. Let us compute $d_sf$. We have $$ d_sf(s,t)=\sum\limits_{k=1}^\infty \sum\limits_{m=1}^k\binom{k}{m}_K^{1/q}[m]^{1/q}s^{q^{m-1}}t^{q^{k-1}} =\sum\limits_{\nu =0}^\infty \sum\limits_{\mu =0}^\nu \binom{\nu +1}{\mu +1}_K^{1/q}[\mu +1]^{1/q}s^{q^\mu}t^{q^\nu}. $$ Using Proposition 5 we find that $d_sf=\Sigma_1+\Sigma_2$ where $$ \Sigma_1=\sum\limits_{\nu =0}^\infty \sum\limits_{\mu =0}^\nu \binom{\nu}{\mu}_K[\mu +1]^{1/q}s^{q^\mu}t^{q^\nu}, $$ $$ \Sigma_2=\sum\limits_{\nu =0}^\infty \sum\limits_{\mu =0}^\nu \binom{\nu}{\mu +1}_K[\mu +1]^{1/q}D_{\mu +1}^{1-q^{-1}}s^{q^\mu}t^{q^\nu}. $$ Note that $$ [\mu +1]^{1/q}=\left( x^{q^{\mu +1}}-x\right)^{1/q}=\left( x^{q^\mu }-x\right) +\left( x^q-x\right)^{1/q}=[\mu ]+[1]^{1/q}, $$ so that \begin{equation} \Sigma_1=\sum\limits_{\nu =0}^\infty \sum\limits_{\mu =0}^\nu \binom{\nu}{\mu}_K[\mu ] s^{q^\mu}t^{q^\nu}+[1]^{1/q}f(s,t). \end{equation} Next, we have $$ \binom{\nu}{\mu +1}_K[\mu +1]^{1/q}D_{\mu +1}^{1-q^{-1}}=\frac{D_\nu}{D_{\mu +1}D_{\nu -\mu -1}^{q^{\mu +1}}}D_{\mu +1}\left( \frac{[\mu +1]}{D_{\mu +1}}\right)^{1/q} =\frac{D_\nu}{D_\mu D_{\nu -\mu -1}^{q^{\mu +1}}}, $$ and also $$ D_{\nu -\mu -1}^q=\frac{1}{[\nu -\mu ]}[\nu -\mu ]D_{\nu -\mu -1}^q =\frac{D_{\nu -\mu}}{[\nu -\mu ]}, $$ whence $$ D_{\nu -\mu -1}^{q^{\mu +1}}=\frac{D_{\nu -\mu }^{q^\mu}}{[\nu -\mu ]^{q^\mu }}. $$ Therefore $$ \Sigma_2=\sum\limits_{\nu =0}^\infty \sum\limits_{\mu =0}^\nu \binom{\nu}{\mu}_K[\nu -\mu ]^{q^\mu}s^{q^\mu}t^{q^\nu}. $$ As above, $[\nu -\mu ]^{q^\mu }=\left( x^{q^{\nu -\mu}}-x\right)^{q^\mu}=[\nu ]-[\mu ]$, so that $$ \Sigma_2=\sum\limits_{\nu =0}^\infty \sum\limits_{\mu =0}^\nu \binom{\nu}{\mu}_K([\nu ]-[\mu ])s^{q^\mu}t^{q^\nu}. $$ Together with (35), this implies (34). $\qquad \blacksquare$ \newpage
{ "timestamp": "2006-05-01T11:20:17", "yymm": "0503", "arxiv_id": "math/0503398", "language": "en", "url": "https://arxiv.org/abs/math/0503398" }
\section{Introduction} In this paper we discuss the limiting theory for a novel, unifying class of non-parametric measures of the variation of financial prices. The theory covers commonly used estimators of variation such as realised volatility, but it also encompasses more recently suggested quantities like realised power variation and realised bipower variation. We considerably strengthen existing results on the latter two quantities, deepening our understanding and unifying their treatment. We will outline the proofs of these theorems, referring for the very technical, detailed formal proofs of the general results to a companion probability theory paper \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)}. Our emphasis is on exposition, explaining where the results come from and how they sit within the econometrics literature. Our theoretical development is motivated by the advent of complete records of quotes or transaction prices for many financial assets. Although market microstructure effects (e.g. discreteness of prices, bid/ask bounce, irregular trading etc.) mean that there is a mismatch between asset pricing theory based on semimartingales and the data at very fine time intervals it does suggest the desirability of establishing an asymptotic distribution theory for estimators as we use more and more highly frequent observations. Papers which directly model the impact of market microstructure noise on realised variance include \cite{BandiRussell(03)}, \cite{HansenLunde(03)}, \cite{ZhangMyklandAitSahalia(03)}, \cite% {BarndorffNielsenHansenLundeShephard(04)} and \cite{Zhang(04)}. Related work in the probability literature on the impact of noise on discretely observed diffusions can be found in \cite{GloterJacod(01a)} and \cite% {GloterJacod(01b)}, while \cite{DelattreJacod(97)} report results on the impact of rounding on sums of functions of discretely observed diffusions. In this paper we ignore these effects. Let the $d$-dimensional vector of the log-prices of a set of assets follow the process \begin{equation*} Y=\left( Y^{1},...,Y^{d}\right) ^{\prime }. \end{equation*} At time $t\geq 0$ we denote the log-prices as $Y_{t}$. Our aim is to calculate measures of the variation of the price process (e.g. realised volatility) over discrete time intervals (e.g. a day or a month). Without loss of generality we can study the mathematics of this by simply looking at what happens when we have $n$ high frequency observations on the time interval $t=0$ to $t=1$ and study what happens to our measures of variation as $n\rightarrow \infty $ (e.g., for introductions to this, \cite% {BarndorffNielsenShephard(02realised)}). In this case returns will be measured over intervals of length $n^{-1}$ as \begin{equation} \Delta _{i}^{n}Y=Y_{i/n}-Y_{(i-1)/n},\quad i=1,2,...,n, \label{return} \end{equation}% where $n$ is a positive integer. We will study the behaviour of the realised generalised bipower variation process \begin{equation} \frac{1}{n}\sum_{i=1}^{\left\lfloor nt\right\rfloor }g(\sqrt{n}~\Delta _{i}^{n}Y)h(\sqrt{n}~\Delta _{i+1}^{n}Y), \label{RGBP} \end{equation}% as $n$ becomes large and where $g$ and $h$ are two given, matrix functions of dimensions $d_{1}\times d_{2}$ and $d_{2}\times d_{3}$ respectively, whose elements have at most polynomial growth. Here $\left\lfloor x\right\rfloor $ denotes the largest integer less than or equal to $x$. Although (\ref{RGBP}) looks initially rather odd, in fact most of the non-parametric volatility measures used in financial econometrics fall within this class (a measure not included in this setup is the range statistic studied in, for example, \cite{Parkinson(80)}). Here we give an extensive list of examples and link them to the existing literature. More detailed discussion of the literature on the properties of these special cases will be given later. \begin{example} \label{Example: 1}\textbf{(a)} Suppose $g(y)=\left( y^{j}\right) ^{2}$ and $% h(y)=1$, then (\ref{RGBP}) becomes% \begin{equation*} \sum_{i=1}^{\left\lfloor nt\right\rfloor }\left( \Delta _{i}^{n}Y^{j}\right) ^{2}, \end{equation*}% which is called the realised quadratic variation process of $Y^{j}$ in econometrics, e.g. \cite{Jacod(94)}, \cite{JacodProtter(98)}, \cite% {BarndorffNielsenShephard(02realised)}, \cite% {BarndorffNielsenShephard(04multi)} and \cite{MyklandZhang(05)}. The increments of this quantity, typically calculated over a day or a week, are often called the realised variances in financial economics and have been highlighted by \cite{AndersenBollerslevDieboldLabys(01)} and \cite% {AndersenBollerslevDiebold(05)} in the context of volatility measurement and forecasting. \noindent \textbf{(b)} Suppose $g(y)=yy^{\prime }$ and $h(y)=I$, then (\ref% {RGBP}) becomes, after some simplification, \begin{equation*} \sum_{i=1}^{\left\lfloor nt\right\rfloor }\left( \Delta _{i}^{n}Y\right) \left( \Delta _{i}^{n}Y\right) ^{\prime }. \end{equation*}% \newline This is the realised covariation process. It has been studied by \cite% {JacodProtter(98)}, \cite{BarndorffNielsenShephard(04multi)} and \cite% {MyklandZhang(05)}. \cite{AndersenBollerslevDieboldLabys(03model)} study the increments of this process to produce forecast distributions for vectors of returns. \ \noindent \textbf{(c)} Suppose $g(y)=\left\vert y^{j}\right\vert ^{r}$ for $% r>0$ and $h(y)=1$, then (\ref{RGBP}) becomes \begin{equation*} n^{-1+r/2}\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\vert \Delta _{i}^{n}Y^{j}\right\vert ^{r}, \end{equation*}% which is called the realised $r$-th order power variation. When $r$ is an integer it has been studied from a probabilistic viewpoint by \cite% {Jacod(94)} while \cite{BarndorffNielsenShephard(03bernoulli)} look at the econometrics of the case where $r>0$. The increments of these types of high frequency volatility measures have been informally used in the financial econometrics literature for some time when $r=1$, but until recently without a strong understanding of their properties. Examples of their use include \cite{Schwert(90JB)}, \cite{AndersenBollerslev(98)} and \cite% {AndersenBollerslev(97jef)}, while they have also been informally discussed by \cite[pp. 349--350]{Shiryaev(99)}\ and \cite{MaheswaranSims(93)}. Following the work by \cite{BarndorffNielsenShephard(03bernoulli)}, \cite% {GhyselsSantaClaraValkoanov(04)} and \cite{ForsbergGhysels(04)} have successfully used realised power variation as an input into volatility forecasting competitions. \noindent \textbf{(d)} Suppose $g(y)=\left\vert y^{j}\right\vert ^{r}$ and $% h(y)=\left\vert y^{j}\right\vert ^{s}$ for $r,s>0$, then (\ref{RGBP}) becomes \begin{equation*} n^{-1+(r+s)/2}\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\vert \Delta _{i}^{n}Y^{j}\right\vert ^{r}\left\vert \Delta _{i+1}^{n}Y^{j}\right\vert ^{s}, \end{equation*}% which is called the realised $r,s$-th order bipower variation process. This measure of variation was introduced by \cite{BarndorffNielsenShephard(04jfe)}% , while a more formal discussion of its behaviour in the $r=s=1$ case was developed by \cite{BarndorffNielsenShephard(03test)}. These authors' interest in this quantity was motivated by its virtue of being resistant to finite activity jumps so long as $\max (r,s)<2$. Recently \cite% {BarndorffNielsenShephardWinkel(04)} and \cite{Woerner(04power)} have studied how these results on jumps extend to infinite activity processes, while \cite{CorradiDistaso(04)} have used these statistics to test the specification of parametric volatility models. \noindent \textbf{(e)} Suppose \begin{equation*} g(y)=\left( \begin{array}{cc} \left\vert y^{j}\right\vert & 0 \\ 0 & \left( y^{j}\right) ^{2}% \end{array}% \right) ,\quad h(y)=\left( \begin{array}{c} \left\vert y^{j}\right\vert \\ 1% \end{array}% \right) . \end{equation*}% Then (\ref{RGBP}) becomes,% \begin{equation*} \left( \begin{array}{c} \displaystyle\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\vert \Delta _{i}^{n}Y^{j}\right\vert \left\vert \Delta _{i+1}^{n}Y^{j}\right\vert \\ \displaystyle\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left( \Delta _{i}^{n}Y^{j}\right) ^{2}% \end{array}% \right) . \end{equation*}% \cite{BarndorffNielsenShephard(03test)} used the joint behaviour of the increments of these two statistics to test for jumps in price processes. \ \cite{HuangTauchen(03)} have empirically studied the finite sample properties of these types of jump tests. \cite% {AndersenBollerslevDiebold(03bipower)} \ and \cite{ForsbergGhysels(04)} use bipower variation as an input into volatility forecasting. \ \end{example} We will derive the probability limit of (\ref{RGBP}) under a general Brownian semimartingale, the workhorse process of modern continuous time asset pricing theory. Only the case of realised quadratic variation, where the limit is the usual quadratic variation QV (defined for general semimartingales), has been previously been studied under such wide conditions. Further, under some stronger but realistic conditions, we will derive a limiting distribution theory for (\ref{RGBP}), so extending a number of results previously given in the literature on special cases of this framework. The outline of this paper is as follows. Section 2 contains a detailed listing of the assumptions used in our analysis. Section 3 gives a statement of a weak law of large numbers for these statistics and the corresponding central limit theory is presented in Section 4. Extensions of the results to higher order variations is briefly indicated in Section 5. Section 6 illustrates the theory by discussing how it gives rise to tests for jumps in the price processes, using bipower and tripower variation. The corresponding literature which discusses various special cases of these results is also given in these sections. Section 8 concludes, while there is an Appendix which provides an outline of the proofs of the results discussed in this paper. For detailed, quite lengthy and highly technical formal proofs we refer to our companion probability theory paper \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)}. \section{Notation and models} We start with $Y$ on some filtered probability space $\left( \Omega ,% \mathcal{F},\left( \mathcal{F}_{t}\right) _{t\geq 0},P\right) $. In most of our analysis we will assume that $Y$ follows a $d$-dimensional Brownian semimartingale (written $Y\in \mathcal{BSM}$). It is given in the following statement. \noindent \textbf{Assumption (H): }We have \begin{equation} Y_{t}=Y_{0}+\int_{0}^{t}a_{u}\mathrm{d}u+\int_{0}^{t}\sigma _{u-}\mathrm{d}% W_{u}, \label{H} \end{equation}% where $W$ is a $d^{\prime }$-dimensional standard Brownian motion (BM), $a$ is a $d$-dimensional process whose elements are predictable and has locally bounded sample paths, and the spot covolatility $d,d^{\prime }$-dimensional matrix $\sigma $ has elements which have c\`{a}dl\`{a}g sample paths. Throughout we will write \begin{equation*} \Sigma _{t}=\sigma _{t}\sigma _{t}^{\prime }, \end{equation*}% the spot covariance matrix. Typically $\Sigma _{t}$ will be full rank, but we do not assume that here. We will write $\Sigma _{t}^{jk}$ to denote the $% j,k$-th element of $\Sigma _{t}$, while we write% \begin{equation*} \sigma _{j,t}^{2}=\Sigma _{t}^{jj}. \end{equation*} \begin{remark} Due to the fact that $t\mapsto \sigma _{t}^{jk}$ is c\`{a}dl\`{a}g all powers of $\sigma _{t}^{jk}$ are locally integrable with respect to the Lebesgue measure. \ In particular then $\int_{0}^{t}\Sigma _{u}^{jj}\mathrm{d% }u<\infty $ for all $t$ and $j$. \end{remark} \begin{remark} Both $a$ and $\sigma $ can have, for example, jumps, intraday seasonality and long-memory. \end{remark} \begin{remark} The stochastic volatility (e.g. \cite{GhyselsHarveyRenault(96)} and \cite% {Shephard(05)}) component of $Y$, \begin{equation*} \int_{0}^{t}\sigma _{u-}\mathrm{d}W_{u}, \end{equation*}% is always a vector of local martingales each with continuous sample paths, as $\int_{0}^{t}\Sigma _{u}^{jj}\mathrm{d}u<\infty $ for all $t$ and $j$. All continuous local martingales with absolutely continuous quadratic variation can be written in the form of a stochastic volatility process. This result, which is due to \cite{Doob(53)}, is discussed in, for example, \cite[p. 170--172]{KaratzasShreve(91)}. Using the Dambis-Dubins-Schwartz Theorem, we know that the difference between the entire continuous local martingale class and the SV class are the local martingales which have only continuous, not absolutely continuous\footnote{% An example of a continuous local martingale which has no SV representation is a time-change Brownian motion where the time-change takes the form of the so-called \textquotedblleft devil's staircase,\textquotedblright\ which is continuous and non-decreasing but not absolutely continuous (see, for example, \cite[Section 27]{Munroe(53)}). This relates to the work of, for example, \cite{CalvetFisher(02)} on multifractals.}, QV. The drift $% \int_{0}^{t}a_{u}\mathrm{d}u$ has elements which are absolutely continuous. This assumption looks ad hoc, however if we impose a lack of arbitrage opportunities and model the local martingale component as a SV process then this property must hold (\cite[p. 3]{KaratzasShreve(98)} and \cite[p. 583]% {AndersenBollerslevDieboldLabys(03model)}). Hence (\ref{H}) is a rather canonical model in the finance theory of continuous sample path processes. \end{remark} We are interested in the asymptotic behaviour, for $n\rightarrow \infty $, of the following volatility measuring process: \begin{equation} Y^{n}(g,h)_{t}=\frac{1}{n}\sum_{i=1}^{\left\lfloor nt\right\rfloor }g(\sqrt{n% }~\Delta _{i}^{n}Y)h(\sqrt{n}~\Delta _{i+1}^{n}Y), \label{XP} \end{equation}% where $g$ and $h$ are two given conformable matrix functions and recalling the definition of $\Delta _{i}^{n}Y$ given in (\ref{return}). \section{Law of large numbers} To build a weak law of large numbers for $Y^{n}(g,h)_{t}$ we need to make the pair $(g,h)$ satisfy the following assumption. \noindent \textbf{Assumption (K):} All the elements of $f$ on $\mathbf{R}% ^{d} $ are continuous with at most polynomial growth. This amounts to there being suitable constants $C>0$ and $p\geq 2$ such that \begin{equation} x\in \mathbf{R}^{d}\quad \Rightarrow \quad \left\Vert f(x)\right\Vert \leq C(1+\Vert x\Vert ^{p}). \label{G1} \end{equation} We also need the following notation. \begin{equation*} \rho _{\sigma }(g)=\mathrm{E}\left\{ g(X)\right\} ,\quad \text{where\quad }% X|\sigma \sim N(0,\sigma \sigma ^{\prime }), \end{equation*}% and \begin{equation*} \rho _{\sigma }(gh)=\mathrm{E}\left\{ g(X)h(X)\right\} . \end{equation*} \begin{example} \label{Example: second}\textbf{(a)} Let $g(y)=yy^{\prime }$ and $h(y)=I$, then $\rho _{\sigma }(g)=\Sigma $ and $\rho _{\sigma }(h)=I$. \noindent \textbf{(b)} Suppose $g(y)=\left\vert y^{j}\right\vert ^{r}$ then $% \rho _{\sigma }(g)=\mu _{r}\sigma _{j}^{r}$, where $\sigma _{j}^{2}$ is the $% j,j$-th element of $\Sigma $, $\mu _{r}=\mathrm{E}(\left\vert u\right\vert ^{r})$ and $u\sim N(0,1)$. \end{example} This setup is sufficient for the proof of Theorem 1.2 of \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)}, which is restated here. \begin{theorem} \label{TT1} Under (H) and assuming $g$ and $h$ satisfy (K) we have that \begin{equation} Y^{n}(g,h)_{t}~\rightarrow ~Y(g,h)_{t}:=\int_{0}^{t}\rho _{\sigma _{u}}(g)\rho _{\sigma _{u}}(h)\mathrm{d}u, \label{WLLN} \end{equation}% where the convergence is in probability, locally uniform in time. \newline \end{theorem} The result is quite clean as it is requires no additional assumptions on $Y$ and so is very close to dealing with the whole class of financially coherent continuous sample path processes. This Theorem covers a number of existing setups which are currently receiving a great deal of attention as measures of variation in financial econometrics. Here we briefly discuss some of the work which has studied the limiting behaviour of these objects. \begin{example} \textbf{(Example \ref{Example: 1}(a) continued)}. Then $g(y)=\left( y^{j}\right) ^{2}$ and $h(y)=1$, so (\ref{WLLN}) becomes% \begin{equation*} \sum_{i=1}^{\left\lfloor nt\right\rfloor }\left( \Delta _{i}^{n}Y^{j}\right) ^{2}\rightarrow ~\int_{0}^{t}\sigma _{j,u}^{2}\mathrm{d}u=[Y^{j}]_{t}, \end{equation*}% the quadratic variation (QV) of $Y^{j}$. This well known result in probability theory is behind much of the modern work on realised volatility, which is compactly reviewed in \cite{AndersenBollerslevDiebold(05)}. \noindent (\textbf{Example \ref{Example: 1}(b) continued}). As $% g(y)=yy^{\prime }$ and $h(y)=I$, then \begin{equation*} \sum_{i=1}^{\left\lfloor nt\right\rfloor }\left( \Delta _{i}^{n}Y\right) \left( \Delta _{i}^{n}Y\right) ^{\prime }\rightarrow ~\int_{0}^{t}\Sigma _{u}% \mathrm{d}u=[Y]_{t}, \end{equation*}% the well known multivariate version of QV. \noindent \textbf{(Example \ref{Example: 1}(c) continued).} Then $% g(y)=\left\vert y^{j}\right\vert ^{r}$ and $h(y)=1$ so \begin{equation*} n^{-1+r/2}\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\vert \Delta _{i}^{n}Y^{j}\right\vert ^{r}\rightarrow ~\mu _{r}\int_{0}^{t}\sigma _{j,u}^{r}\mathrm{d}u. \end{equation*}% This result is due to \cite{Jacod(94)} and \cite% {BarndorffNielsenShephard(03bernoulli)}. \noindent \textbf{(Example \ref{Example: 1}(d) continued).} Then $% g(y)=\left\vert y^{j}\right\vert ^{r}$ and $h(y)=\left\vert y^{j}\right\vert ^{s}$ for $r,s>0$, so \begin{equation*} n^{-1+(r+s)/2}\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\vert \Delta _{i}^{n}Y^{j}\right\vert ^{r}\left\vert \Delta _{i+1}^{n}Y^{j}\right\vert ^{s}\rightarrow ~\mu _{r}\mu _{s}\int_{0}^{t}\sigma _{j,u}^{r+s}\mathrm{d}u, \end{equation*}% a result due to \cite{BarndorffNielsenShephard(04jfe)}, who derived it under stronger conditions than those used here. \noindent \textbf{(Example \ref{Example: 1}(e) continued).} Then \begin{equation*} g(y)=\left( \begin{array}{cc} \left\vert y^{j}\right\vert & 0 \\ 0 & \left( y^{j}\right) ^{2}% \end{array}% \right) ,\quad h(y)=\left( \begin{array}{c} \left\vert y^{j}\right\vert \\ 1% \end{array}% \right) , \end{equation*}% so \begin{equation*} \left( \begin{array}{c} \displaystyle\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\vert \Delta _{i}^{n}Y^{j}\right\vert \left\vert \Delta _{i+1}^{n}Y^{j}\right\vert \\ \displaystyle\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left( \Delta _{i}^{n}Y^{j}\right) ^{2}% \end{array}% \right) \rightarrow \left( ~% \begin{array}{c} \mu _{1}^{2} \\ 1% \end{array}% \right) \int_{0}^{t}\sigma _{j,u}^{2}\mathrm{d}u. \end{equation*}% \cite{BarndorffNielsenShephard(03test)} used this type of result to test for jumps as this particular bipower variation is robust to jumps. \end{example} \section{Central limit theorem\label{sect:CLT}} \subsection{Motivation} It is important to be able to quantify the difference between the estimator $% Y^{n}(g,h)$ and $Y(g,h)$. In this subsection we do this by giving a central limit theorem for $\sqrt{n}(Y^{n}(g,h)-Y(g,h))$. We have to make some stronger assumptions both on the process $Y$ and on the pair $(g,h)$ in order to derive this result. \subsection{Assumptions on the process} We start with a variety of assumptions which strengthen (H) and (K) given in the previous subsection. \noindent \textbf{Assumption (H0):} We have (H) with \begin{equation} \sigma _{t}=\sigma _{0}+\int_{0}^{t}a_{u}^{\ast }\mathrm{d}% u+\int_{0}^{t}\sigma _{u-}^{\ast }\mathrm{d}W_{u}+\int_{0}^{t}v_{u-}^{\ast }% \mathrm{d}Z_{u}, \label{H'} \end{equation}% where $Z$ is a $d^{\prime \prime }$-dimensional L\'{e}vy process, independent of $W$. Further, the processes $a^{\ast }$, $\sigma ^{\ast }$, $% v^{\ast }$ are adapted c\`{a}dl\`{a}g arrays, with $a^{\ast }$ also being predictable and locally bounded. \noindent \textbf{Assumption (H1):} We have (H) with \begin{eqnarray} \sigma _{t} &=&\sigma _{0}+\int_{0}^{t}a_{u}^{\ast }\mathrm{d}% u+\int_{0}^{t}\sigma _{u-}^{\ast }\mathrm{d}W_{u}+\int_{0}^{t}v_{u-}^{\ast }% \mathrm{d}V_{u} \label{assumption (V)} \\ &&+\int_{0}^{t}\int_{E}\varphi \circ w(u-,x)\left( \mu -\nu \right) \left( \mathrm{d}u,\mathrm{d}x\right) +\int_{0}^{t}\int_{E}\left( w-\varphi \circ w\right) \left( u-,x\right) \mu \left( \mathrm{d}u,\mathrm{d}x\right) . \notag \end{eqnarray}% Here $a^{\ast }$, $\sigma ^{\ast }$, $v^{\ast }$ are adapted c\`{a}dl\`{a}g arrays, with $a^{\ast }$ also being predictable and locally bounded. $V$ is a $d^{\prime \prime }$-dimensional Brownian motion independent of $W$. $\mu $ is a Poisson measure on $\left( 0,\infty \right) \times E$ independent of $W$ and $V$, with intensity measure $\nu (\mathrm{d}t,\mathrm{d}x)=\mathrm{d}% t\otimes F(\mathrm{d}x)$ and $F$ is a $\sigma $-finite measure on the Polish space $\left( E,\mathcal{E}\right) $. $\varphi $ is a continuous truncation function on $R^{dd^{\prime }}$ (a function with compact support, which coincide with the identity map on the neighbourhood of $0$). Finally $% w(\omega ,u,x)$ is a map $\Omega \times \lbrack 0,\infty )\times E$ into the space of $d\times d^{\prime }$arrays which is $\mathcal{F}_{u}\otimes $ $% \mathcal{E}-$measurable in $(\omega ,x)$ for all $u$ and c\`{a}dl\`{a}g in $% u $, and such that for some sequences $\left( S_{k}\right) $ of stopping times increasing to $+\infty $ we have% \begin{equation*} \sup_{\omega \in \Omega ,u<S_{k}(\omega )}\left\Vert w(\omega ,u,x)\right\Vert \leq \psi _{k}(x)\quad \text{where\quad }\int_{E}\left( 1\wedge \psi _{k}(x)^{2}\right) F(\mathrm{d}x)<\infty . \end{equation*} \noindent \textbf{Assumption (H2): }$\Sigma =\sigma \sigma ^{\prime }$ is everywhere invertible. \begin{remark} Assumption (H1) looks quite complicated but has been setup so that the same conditions on the coefficients can be applied both to $\sigma $ and $\Sigma =\sigma \sigma ^{\prime }$. If there were no jumps then it would be sufficient to employ the first line of (\ref{assumption (V)}). The assumption (H1) is rather general from an econometric viewpoint as it allows for flexible leverage effects, multifactor volatility effects, jumps, non-stationarities, intraday effects, etc. \end{remark} \subsection{Assumptions on $g$ and $h$} In order to derive a central limit theorem we need to impose some regularity on $g$ and $h$. \noindent \textbf{Assumption (K1): }$f$ is even (that is $f(x)=f(-x)$ for $% x\in R^{d}$) and continuously differentiable, with derivatives having at most polynomial growth. In order to handle some of the most interesting cases of bipower variation, where we are mostly interested in taking low powers of absolute values of returns which may not be differentiable at zero, we sometimes need to relax (K1). The resulting condition is quite technical and is called (K2). It is discussed in the Appendix. \noindent \textbf{Assumption (K2):} $f$ is even and continuously differentiable on the complement $B^{c}$\ of a closed subset $B\subset \mathbb{R}^{d}$ and satisfies% \begin{equation*} ||y||\leq 1\Longrightarrow |f(x+y)-f(x)|\leq C(1+||x||^{p})||y||^{r} \end{equation*}% for some constants $C$, $p\geq 0$ and $r\in \left( 0,1\right] $. Moreover a) If $r=1$ then $B$ has Lebesgue measure $0$. b) If $r<1$ then $B$ satisfies \begin{equation} \left. \begin{array}{l} \text{for any positive definite }d\times d\text{ matrix }C\text{ and } \\ \text{any }N(0,C)\text{-random vector }U\text{ the distance }d(U,B) \\ \text{from }U\text{ to }B\text{ has a density }\psi _{C}\text{ on }R_{+},% \text{ such that } \\ sup_{x\in R_{+},|C|+|C^{-1}|\leq A}\psi _{C}(x)<\infty \text{ for all }% A<\infty ,% \end{array}% \right\} \label{K13} \end{equation} \qquad\ and we have \begin{equation} x\in B^{c},~\Vert y\Vert \leq 1\bigwedge {\frac{d(x,B)}{2}}~~\Rightarrow ~~\left\{ \begin{array}{l} \Vert \nabla f(x)\Vert \leq {\frac{C(1+\Vert x\Vert ^{p})}{d(x,B)^{1-r}}}, \\% [2.5mm] \Vert \nabla f(x+y)-\nabla f(x)\Vert \leq {\frac{C(1+\Vert x\Vert ^{p})\Vert y\Vert }{d(x,B)^{2-r}}}.% \end{array}% \right. \label{K11} \end{equation} \begin{remark} These conditions accommodate the case where $f$ equals $\left\vert x^{j}\right\vert ^{r}$: this function satisfies (K1) when $r>1$, and (K2) when $r\in (0,1]$ (with the same $r$ of course). When $B$ is a finite union of hyperplanes it satisfies (\ref{K13}). Also, observe that (K1) implies (K2) with $r=1$ and $B=\emptyset $. \end{remark} \subsection{Central limit theorem} Each of the following assumptions (J1) and (J2) are sufficient for the statement of Theorem 1.3 of \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)} to hold. \noindent \textbf{Assumption (J1):} We have (H1) and $g$ and $h$ satisfy (K1). \noindent \textbf{Assumption (J2):} We have (H1), (H2) and $g$ and $h$ satisfy (K2).\newline The result of the Theorem is restated in the following. \begin{theorem} \label{TT3}Assume at least one of (J1) and (J2) holds, then the process \begin{equation*} \sqrt{n}~(Y^{n}(g,h)_{t}-Y(g,h)_{t}) \end{equation*}% converges stably in law towards a limiting process $U(g,h)$ having the form% \begin{equation} U(g,h)_{t}^{jk}=\sum_{j^{\prime }=1}^{d_{1}}\sum_{k^{\prime }=1}^{d_{3}}\int_{0}^{t}\alpha (\sigma _{u},g,h)^{jk,j^{\prime }k^{\prime }}~% \mathrm{d}B_{u}^{j^{\prime },k^{\prime }}, \end{equation}% where% \begin{equation*} \sum_{l=1}^{d_{1}}\sum_{m=1}^{d_{3}}\alpha (\sigma ,g,h)^{jk,lm}\alpha (\sigma ,g,h)^{j^{\prime }k^{\prime },lm}=A(\sigma ,g,h)^{jk,j^{\prime }k^{\prime }}, \end{equation*}% and% \begin{eqnarray*} A(\sigma ,g,h)^{jk,j^{\prime }k^{\prime }} &=&\displaystyle% \sum_{l=1}^{d_{2}}\sum_{l^{\prime }=1}^{d_{2}}\left\{ \rho _{\sigma }\left( g^{jl}g^{j^{\prime }l^{\prime }}\right) \rho _{\sigma }\left( h^{lk}h^{l^{\prime }k^{\prime }}\right) +\rho _{\sigma }\left( g^{jl}\right) \rho _{\sigma }\left( h^{l^{\prime }k^{\prime }}\right) \rho _{\sigma }\left( g^{j^{\prime }l^{\prime }}h^{lk}\right) \right. \\ &&\displaystyle+\rho _{\sigma }\left( g^{j^{\prime }l^{\prime }}\right) \rho _{\sigma }\left( h^{lk}\right) \rho _{\sigma }\left( g^{jl}h^{l^{\prime }k^{\prime }}\right) \\ &&\displaystyle\left. -3\rho _{\sigma }\left( g^{jl}\right) \rho _{\sigma }\left( g^{j^{\prime }l^{\prime }}\right) \rho _{\sigma }\left( h^{lk}\right) \rho _{\sigma }\left( h^{l^{\prime }k^{\prime }}\right) \right\} . \end{eqnarray*}% Furthermore, $B$ is a standard Wiener process which is defined on an extension of $\left( \Omega ,\mathcal{F},\left( \mathcal{F}_{t}\right) _{t\geq 0},P\right) $ and is independent of the $\sigma $--field $\mathcal{F} $. \end{theorem} \begin{remark} Convergence stably in law is slightly stronger than convergence in law. It is discussed in, for example, \cite[pp. 512-518]{JacodShiryaev(03)}. \end{remark} \begin{remark} Suppose $d_{3}=1$, which is the situation looked at in Example \ref{Example: 1}(e). Then $Y^{n}(g,h)_{t}$ is a vector and so the limiting law of $\sqrt{n}% (Y^{n}(g,h)-Y(g,h))$ simplifies. It takes on the form of \begin{equation} U(g,h)_{t}^{j}=\sum_{j^{\prime }=1}^{d_{1}}\int_{0}^{t}\alpha (\sigma _{u},g,h)^{j,j^{\prime }}~\mathrm{d}B_{u}^{j^{\prime }}, \end{equation}% where% \begin{equation*} \sum_{l=1}^{d_{1}}\alpha (\sigma ,g,h)^{j,l}\alpha (\sigma ,g,h)^{j^{\prime },l}=A(\sigma ,g,h)^{j,j^{\prime }}. \end{equation*}% Here% \begin{eqnarray*} A(\sigma ,g,h)^{j,j^{\prime }} &=&\displaystyle\sum_{l=1}^{d_{2}}\sum_{l^{% \prime }=1}^{d_{2}}\left\{ \rho _{\sigma }(g^{jl}g^{j^{\prime }l^{\prime }})\rho _{\sigma }(h^{l}h^{l^{\prime }})+\rho _{\sigma }(g^{jl})\rho _{\sigma }(h^{l^{\prime }})\rho _{\sigma }(g^{j^{\prime }l^{\prime }}h^{l})\right. \\ &&\displaystyle+\left. \rho _{\sigma }(g^{j^{\prime }l^{\prime }})\rho _{\sigma }(h^{l})\rho _{\sigma }(g^{jl}h^{l^{\prime }})-3\rho _{\sigma }(g^{jl})\rho _{\sigma }(g^{j^{\prime }l^{\prime }})\rho _{\sigma }(h^{l})\rho _{\sigma }(h^{l^{\prime }})\right\} . \end{eqnarray*}% In particular, for a single point in time $t$, \begin{equation*} \sqrt{n}~(Y^{n}(g,h)_{t}-Y(g,h)_{t})\rightarrow MN\left( 0,\int_{0}^{t}A(\sigma _{u},g,h)\mathrm{d}u\right) , \end{equation*}% where $MN$ denotes a mixed Gaussian distribution. and $A(\sigma ,g,h)$ denotes a matrix whose $j,j^{\prime }$-th element is $A(\sigma ,g,h)^{j,j^{\prime }}$. \end{remark} \begin{remark} Suppose $g(y)=I$, then $A$ becomes \begin{equation*} A(\sigma ,g,h)^{jk,j^{\prime }k^{\prime }}=\rho _{\sigma }(h^{jk}h^{j^{\prime }k^{\prime }})-\rho _{\sigma }(h^{jk})\rho _{\sigma }(h^{j^{\prime }k^{\prime }}). \end{equation*} \end{remark} \subsection{Leading examples of this result} \begin{example} Suppose $d_{1}=d_{2}=d_{3}=1$, then \begin{equation} U(g,h)_{t}=\int_{0}^{t}\sqrt{A(\Sigma _{u},g,h)}~\mathrm{d}B_{u}, \end{equation}% where% \begin{equation*} A(\sigma ,g,h)=\rho _{\sigma }(gg)\rho _{\sigma }(hh)+2\rho _{\sigma }(g)\rho _{\sigma }(h)\rho _{\sigma }(gh)-3\left\{ \rho _{\sigma }(g)\rho _{\sigma }(h)\right\} ^{2}. \end{equation*}% We consider two concrete examples of this setup. \noindent \textbf{(i)} Power variation. Suppose $g(y)=1$ and $% h(y)=\left\vert y^{j}\right\vert ^{r}$ where $r>0$, then $\rho _{\sigma }(g)=1$, \begin{equation*} \rho _{\sigma }(h)=\rho _{\sigma }(gh)=\mu _{r}\sigma _{j}^{r},\quad \rho _{\sigma }(hh)=\mu _{2r}\sigma _{j}^{2r}. \end{equation*}% This implies that% \begin{eqnarray*} A(\sigma ,g,h) &=&\mu _{2r}\sigma _{j}^{2r}+2\mu _{r}^{2}\sigma _{j}^{2r}-3\mu _{r}^{2}\sigma _{j}^{2r} \\ &=&\left( \mu _{2r}-\mu _{r}^{2}\right) \sigma _{j}^{2r} \\ &=&v_{r}\sigma _{j}^{2r}, \end{eqnarray*}% where $v_{r}=\mathrm{Var}(\left\vert u\right\vert ^{r})$ and $u\sim N(0,1)$. When $r=2$, this yields a central limit theorem for the realised quadratic variation process, with \begin{equation*} U(g,h)_{t}=\int_{0}^{t}\sqrt{2\sigma _{j,u}^{4}}~\mathrm{d}B_{u}, \end{equation*}% a result which appears in \cite{Jacod(94)}, \cite{MyklandZhang(05)} and, implicitly, \cite{JacodProtter(98)}, while the case of a single value of $t$ appears in \cite{BarndorffNielsenShephard(02realised)}. For the more general case of $r>0$ \cite{BarndorffNielsenShephard(03bernoulli)} derived, under much stronger conditions, a central limit theorem for $U(g,h)_{1}$. Their result ruled out leverage effects, which are allowed under Theorem \ref{TT3}% . The finite sample behaviour of this type of limit theory is studied in, for example, \cite{BarndorffNielsenShephard(05tom)}, \cite% {GoncalvesMeddahi(04)} and \cite{NielsenFrederiksen(05)}. \noindent \textbf{(ii)} Bipower variation. Suppose $g(y)=\left\vert y^{j}\right\vert ^{r}$ and $h(y)=\left\vert y^{j}\right\vert ^{s}$ where $% r,s>0$, then% \begin{eqnarray*} \rho _{\sigma }(g) &=&\mu _{r}\sigma _{j}^{r},\quad \rho _{\sigma }(h)=\mu _{s}\sigma _{j}^{s},\quad \rho _{\sigma }(gg)=\mu _{2r}\sigma _{j}^{2r},\quad \\ \rho _{\sigma }(hh) &=&\mu _{2s}\sigma _{j}^{2s},\quad \rho _{\sigma }(gh)=\mu _{r+s}\sigma _{j}^{r+s}. \end{eqnarray*}% This implies that \begin{eqnarray*} A(\sigma ,g,h) &=&\mu _{2r}\sigma _{j}^{2r}\mu _{2s}\sigma _{j}^{2s}+2\mu _{r}\sigma _{j}^{r}\mu _{s}\sigma _{j}^{s}\mu _{r+s}\sigma _{j}^{r+s}-3\mu _{r}^{2}\sigma _{j}^{2r}\mu _{s}^{2}\sigma _{j}^{2s} \\ &=&\left( \mu _{2r}\mu _{2s}+2\mu _{r+s}\mu _{r}\mu _{s}-3\mu _{r}^{2}\mu _{s}^{2}\right) \sigma _{j}^{2r+2s}. \end{eqnarray*}% In the $r=s=1$ case \cite{BarndorffNielsenShephard(03test)} derived, under much stronger conditions, a central limit theorem for $U(g,h)_{1}$. Their result ruled out leverage effects, which are allowed under Theorem \ref{TT3}% . In that special case, writing \begin{equation*} \vartheta =\frac{\pi ^{2}}{4}+\pi -5, \end{equation*}% we have \begin{equation*} U(g,h)_{t}=\mu _{1}^{2}\int_{0}^{t}\sqrt{\left( 2+\vartheta \right) \sigma _{j,u}^{4}}~\mathrm{d}B_{u}. \end{equation*} \end{example} \begin{example} Suppose $g=I$, $h(y)=yy^{\prime }$. Then we have to calculate% \begin{equation*} A(\sigma ,g,h)^{jk,j^{\prime }k^{\prime }}=\rho _{\sigma }(h^{jk}h^{j^{\prime }k^{\prime }})-\rho _{\sigma }(h^{jk})\rho _{\sigma }(h^{j^{\prime }k^{\prime }}). \end{equation*}% However, \begin{equation*} \rho _{\sigma }(h^{jk})=\Sigma ^{jk},\quad \rho _{\sigma }(h^{jk}h^{j^{\prime }k^{\prime }})=\Sigma ^{jk}\Sigma ^{j^{\prime }k^{\prime }}+\Sigma ^{jj^{\prime }}\Sigma ^{kk^{\prime }}+\Sigma ^{jk^{\prime }}\Sigma ^{kj^{\prime }}, \end{equation*}% so \begin{eqnarray*} A(\sigma ,g,h)^{jk,j^{\prime }k^{\prime }} &=&\Sigma ^{jk}\Sigma ^{j^{\prime }k^{\prime }}+\Sigma ^{jj^{\prime }}\Sigma ^{kk^{\prime }}+\Sigma ^{jk^{\prime }}\Sigma ^{kj^{\prime }}-\Sigma ^{jk}\Sigma ^{j^{\prime }k^{\prime }} \\ &=&\Sigma ^{jj^{\prime }}\Sigma ^{kk^{\prime }}+\Sigma ^{jk^{\prime }}\Sigma ^{kj^{\prime }}. \end{eqnarray*}% This is the result found in \cite{BarndorffNielsenShephard(04multi)}, but proved under stronger conditions, and is implicit in the work of \cite% {JacodProtter(98)}. \end{example} \begin{example} \label{Example:vector case}Suppose $d_{1}=d_{2}=2$, $d_{3}=1$ and $g$ is diagonal. Then \begin{equation} U(g,h)_{t}^{j}=\sum_{j^{\prime }=1}^{2}\int_{0}^{t}\alpha (\sigma _{u},g,h)^{j,j^{\prime }}~\mathrm{d}B_{u}^{j^{\prime }}, \end{equation}% where% \begin{equation*} \sum_{l=1}^{2}\alpha (\sigma ,g,h)^{j,l}\alpha (\sigma ,g,h)^{j^{\prime },l}=A(\sigma ,g,h)^{j,j^{\prime }}. \end{equation*}% Here% \begin{eqnarray*} A(\sigma ,g,h)^{j,j^{\prime }} &=&\rho _{\sigma }(g^{jj}g^{j^{\prime }j^{\prime }})\rho _{\sigma }(h^{j}h^{j^{\prime }})+\rho _{\sigma }(g^{jj})\rho _{\sigma }(h^{j^{\prime }})\rho _{\sigma }(g^{j^{\prime }j^{\prime }}h^{j}) \\ &&+\rho _{\sigma }(g^{j^{\prime }j^{\prime }})\rho _{\sigma }(h^{j})\rho _{\sigma }(g^{jj}h^{j^{\prime }})-3\rho _{\sigma }(g^{jj})\rho _{\sigma }(g^{j^{\prime }j^{\prime }})\rho _{\sigma }(h^{j})\rho _{\sigma }(h^{j^{\prime }}). \end{eqnarray*} \end{example} \begin{example} \label{Example:joint BPV and RV}Joint behaviour of realised QV and realised bipower variation. This sets% \begin{equation*} g(y)=\left( \begin{array}{cc} \left\vert y^{j}\right\vert & 0 \\ 0 & 1% \end{array}% \right) ,\quad h(y)=\left( \begin{array}{c} \left\vert y^{j}\right\vert \\ \left( y^{j}\right) ^{2}% \end{array}% \right) . \end{equation*}% The implication is that \begin{equation*} \rho _{\sigma }(g^{11})=\rho _{\sigma }(g^{22}g^{11})=\rho _{\sigma }(g^{11}g^{22})=\mu _{1}\sigma _{j},\ \rho _{\sigma }(g^{22})=1,\ \rho _{\sigma }(g^{11}g^{11})=\sigma _{j}^{2},\ \rho _{\sigma }(g^{22}g^{22})=1, \end{equation*}% \begin{equation*} \rho _{\sigma }(h^{1})=\mu _{1}\sigma _{j},\ \rho _{\sigma }(h^{2})=\rho _{\sigma }(h^{1}h^{1})=\sigma _{j}^{2},\ \rho _{\sigma }(h^{1}h^{2})=\rho _{\sigma }(h^{2}h^{1})=\mu _{3}\sigma _{j}^{3},\ \rho _{\sigma }(h^{2}h^{2})=3\sigma _{j}^{4}, \end{equation*}% \begin{equation*} \rho _{\sigma }(g^{11}h^{1})=\sigma _{j}^{2},\ \rho _{\sigma }(g^{11}h^{2})=\mu _{3}\sigma _{j}^{3},\ \rho _{\sigma }(g^{22}h^{1})=\mu _{1}\sigma _{j},\ \rho _{\sigma }(g^{22}h^{2})=\sigma _{j}^{2}. \end{equation*}% Thus \begin{eqnarray*} A(\sigma ,g,h)^{1,1} &=&\sigma _{j}^{2}\sigma _{j}^{2}+2\mu _{1}\sigma _{j}\mu _{1}\sigma _{j}\sigma _{j}^{2}-3\mu _{1}\sigma _{j}\mu _{1}\sigma _{j}\mu _{1}\sigma _{j}\mu _{1}\sigma _{j} \\ &=&\sigma _{j}^{4}\left( 1+2\mu _{1}^{2}-3\mu _{1}^{4}\right) =\mu _{1}^{4}(2+\vartheta )\sigma _{j}^{4}, \end{eqnarray*}% while% \begin{equation*} A(\sigma ,g,h)^{2,2}=3\sigma _{j}^{4}+2\sigma _{j}^{4}-3\sigma _{j}^{4}=2\sigma _{j}^{4}, \end{equation*}% and% \begin{eqnarray*} A(\sigma ,g,h)^{1,2} &=&\mu _{1}\sigma _{j}\mu _{3}\sigma _{j}^{3}+\mu _{1}\sigma _{j}\sigma _{j}^{2}\mu _{1}\sigma _{j}+\mu _{1}\sigma _{j}\mu _{3}\sigma _{j}^{3}-3\mu _{1}\sigma _{j}\mu _{1}\sigma _{j}\sigma _{j}^{2} \\ &=&2\sigma _{j}^{4}\left( \mu _{1}\mu _{3}-\mu _{1}^{2}\right) =2\mu _{1}^{2}\sigma _{j}^{4}. \end{eqnarray*}% This generalises the result given in \cite{BarndorffNielsenShephard(03test)} to the leverage case. In particular we have that \begin{equation*} \left( \begin{array}{c} U(g,h)_{t}^{1} \\ U(g,h)_{t}^{2}% \end{array}% \right) =\left( \begin{array}{l} \displaystyle\mu _{1}^{2}\int_{0}^{t}\sqrt{2\sigma _{u}^{4}}\mathrm{d}% B_{u}^{1}+\mu _{1}^{2}\int_{0}^{t}\sqrt{\vartheta \sigma _{u}^{4}}\mathrm{d}% B_{u}^{2} \\ \displaystyle\int_{0}^{t}\sqrt{2\sigma _{u}^{4}}\mathrm{d}B_{u}^{1}.% \end{array}% \right) \end{equation*} \end{example} \section{Multipower variation\label{sect:multipower variation}} A natural extension of generalised bipower variation is to generalised multipower variation% \begin{equation*} Y^{n}(g)_{t}=\frac{1}{n}\sum_{i=1}^{\left\lfloor nt\right\rfloor }\left\{ \dprod\limits_{i^{\prime }=1}^{I\wedge \left( i+1\right) }g_{i^{\prime }}(% \sqrt{n}~\Delta _{i-i^{\prime }+1}^{n}Y)\right\} . \end{equation*}% This measure of variation, for the $g_{i^{\prime }}$ being absolute powers, was introduced by \cite{BarndorffNielsenShephard(03test)}. We will be interested in studying the properties of $Y^{n}(g)_{t}$ for given functions $\left\{ g_{i}\right\} $ with the following properties. \noindent \textbf{Assumption (K}$^{\ast }$\textbf{):} All the $\left\{ g_{i}\right\} $ are continuous with at most polynomial growth. The previous results suggests that if $Y$ is a Brownian semimartingale and Assumption (K$^{\ast }$) holds then \begin{equation*} Y^{n}(g)_{t}\rightarrow Y(g)_{t}:=\int_{0}^{t}\dprod\limits_{i=0}^{I}\rho _{\sigma _{u}}(g_{i})\mathrm{d}u. \end{equation*} \begin{example} \textbf{(a)} Suppose $I=4$ and $g_{i}(y)=\left\vert y^{j}\right\vert $, then $\rho _{\sigma }(g_{i})=\mu _{1}\sigma _{j}$ so \begin{equation*} Y(g)_{t}=\mu _{1}^{4}\int_{0}^{t}\sigma _{j,u}^{4}\mathrm{d}u, \end{equation*}% a scaled version of integrated quarticity. \newline \noindent \textbf{(b)} Suppose $I=3$ and $g_{i}(y)=\left\vert y^{j}\right\vert ^{4/3}$, then \begin{equation*} \rho _{\sigma }(g_{i})=\mu _{4/3}\sigma _{j}^{4/3} \end{equation*}% so \begin{equation*} Y(g)_{t}=\mu _{4/3}^{3}\int_{0}^{t}\sigma _{j,u}^{4}\mathrm{d}u. \end{equation*} \end{example} \begin{example} Of some importance is the generic case where $g_{i}(y)=\left\vert y^{j}\right\vert ^{2/I}$, which implies \begin{equation*} Y(g)_{t}=\mu _{2/I}^{I}\int_{0}^{t}\sigma _{j,u}^{2}\mathrm{d}u. \end{equation*}% Thus this class provides an interesting alternative to realised variance as an estimator of integrated variance. Of course it is important to know a central limit theory for these types of quantities. \ \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)} show that when (H1) and (H2) hold then \begin{equation*} \sqrt{n}\left[ Y^{n}(g)_{t}-Y(g)_{t}\right] \rightarrow \int_{0}^{t}\sqrt{% \omega _{I}^{2}\sigma _{j,u}^{4}}~\mathrm{d}B_{u}, \end{equation*}% where% \begin{equation*} \omega _{I}^{2}=\mathrm{Var}\left( \dprod\limits_{i=1}^{I}\left\vert u_{i}\right\vert ^{2/I}\right) +2\sum_{j=1}^{I-1}\mathrm{Cov}\left( \dprod\limits_{i=1}^{I}\left\vert u_{i}\right\vert ^{2/I},\dprod\limits_{i=1}^{I}\left\vert u_{i-j}\right\vert ^{2/I}\right) , \end{equation*}% with $u_{i}\sim NID(0,1)$. Clearly $\omega _{1}^{2}=2$, while recalling that $\mu _{1}=\sqrt{2/\pi }$, \begin{eqnarray*} \omega _{2}^{2} &=&\mathrm{Var}(\left\vert u_{1}\right\vert \left\vert u_{2}\right\vert )+2\mathrm{Cov}(\left\vert u_{1}\right\vert \left\vert u_{2}\right\vert ,\left\vert u_{2}\right\vert \left\vert u_{3}\right\vert ) \\ &=&1+2\mu _{1}^{2}-3\mu _{1}^{4}, \end{eqnarray*} \noindent and \begin{eqnarray*} \omega _{3}^{2} &=&\mathrm{Var}(\left( \left\vert u_{1}\right\vert \left\vert u_{2}\right\vert \left\vert u_{3}\right\vert \right) ^{2/3})+2% \mathrm{Cov}(\left( \left\vert u_{1}\right\vert \left\vert u_{2}\right\vert \left\vert u_{3}\right\vert \right) ^{2/3},\left( \left\vert u_{2}\right\vert \left\vert u_{3}\right\vert \left\vert u_{4}\right\vert \right) ^{2/3}) \\ &&+2\mathrm{Cov}(\left( \left\vert u_{1}\right\vert \left\vert u_{2}\right\vert \left\vert u_{3}\right\vert \right) ^{2/3},\left( \left\vert u_{3}\right\vert \left\vert u_{4}\right\vert \left\vert u_{5}\right\vert \right) ^{2/3}) \\ &=&\left( \mu _{4/3}^{3}-\mu _{2/3}^{6}\right) +2\left( \mu _{4/3}^{2}\mu _{2/3}^{2}-\mu _{2/3}^{6}\right) +2\left( \mu _{4/3}\mu _{2/3}^{4}-\mu _{2/3}^{6}\right) . \end{eqnarray*} \end{example} \begin{example} The law of large numbers and the central limit theorem also hold for linear combinations of processes like $Y(g)$ above. For example one may denote by $% \zeta^n_i$ the $d\times d$ matrix whose $(k,l)$ entry is $% \sum_{j=0}^{d-1}\Delta^n_{i+j}Y^k\Delta^n_{i+j}Y^l$. Then \begin{equation*} Z^n_t=\frac{n^{d-1}}{d!}\sum_{i=1}^{[nt]}\det(\zeta^n_i) \end{equation*} is a linear combinations of processes $Y^n(g)$ for functions $g_l$ being of the form $g_l(y)=y^jy^k$. It is proved in \cite{JacodLejayTalay(05)} that under (H) \begin{equation*} Z^n_t\rightarrow Z_t:=\int_0^t \det(\sigma_u\sigma^{\prime}_u)du \end{equation*} in probability, whereas under (H1) and (H2) the associated CLT is the following convergence in law: \begin{equation*} \sqrt{n}(Z^n_t-Z_t)\rightarrow \int_0^t\sqrt{\Gamma(\sigma_u)}~dB_u, \end{equation*} where $\Gamma(\sigma)$ denotes the covariance of the variable $% \det(\zeta)/d! $, and $\zeta$ is a $d\times d$ matrix whose $(k,l)$ entry is $\sum_{j=0}^{d-1}U_j^kU_j^l$ and the $U_j$'s are i.i.d. centered Gaussian vectors with covariance $\sigma\sigma^{\prime}$. This kind of result may be used for testing whether the rank of the diffusion coefficient is everywhere smaller than $d$ (in which case one could use a model with a $d^{\prime}<d$ for the dimension of the driving Wiener process $W$). \end{example} \section{Conclusion} This paper provides some rather general limit results for realised generalised bipower variation. In the case of power variation and bipower variation the results are proved under much weaker assumptions than those which have previously appeared in the literature. In particular the no-leverage assumption is removed, which is important in the application of these results to stock data. There are a number of open questions. It is rather unclear how econometricians might exploit the generality of the $g$ and $h$ functions to learn about interesting features of the variation of price processes. It would be interesting to know what properties $g$ and $h$ must possess in order for these statistics to be robust to finite activity and infinite activity jumps. A challenging extension is to construct a version of realised generalised bipower variation which is robust to market microstructure effects. Following the work on the realised volatility there are two leading strategies which may be able to help: the kernel based approach, studied in detailed by \cite% {BarndorffNielsenHansenLundeShephard(04)}, and the subsampling approach of \cite{ZhangMyklandAitSahalia(03)} and \cite{Zhang(04)}. In the realised volatility case these methods are basically equivalent, however it is perhaps the case that the subsampling method is easier to extend to the non-quadratic case. \section{Acknowledgments} Ole E. Barndorff-Nielsen's work is supported by CAF (\texttt{www.caf.dk}), which is funded by the Danish Social Science Research Council. Neil Shephard's research is supported by the UK's ESRC through the grant \textquotedblleft High frequency financial econometrics based upon power variation.\textquotedblright\ \section{Proof of Theorem \protect\ref{TT3}} \subsection{Strategy for the proof} Below we give a fairly detailed account of the basic techniques in the proof of Theorem \ref{TT3}, in the one-dimensional case and under some relatively minor simplifying assumptions. Throughout we set $h=1$ for the main difficulty in the proof is being able to deal with the generality in the $g$ function. Once that has been mastered the extension to the bipower measure is not a large obstacle. We refer the reader to \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)} for readers who wish to see the more general case. In this subsection we provide a brief outline of the content of the Section. The aim of this Section is to show that% \begin{equation} \sqrt{n}\left( \frac{1}{n}\,\sum_{i=1}^{[nt]}g\left( \sqrt{n}\,\triangle _{i}^{n}Y\right) -\int_{0}^{t}\rho _{\sigma _{u}}(g)\right) \rightarrow \int_{0}^{t}\sqrt{\rho _{\sigma _{u}}(g^{2})-\rho _{\sigma _{u}}(g)^{2}}\;% \mathrm{d}B_{u} \label{eqn 0} \end{equation}% where $B$\ is a Brownian motion independent of the process $Y$\ and the convergence is (stably) in law. This case is important for the extension to realised generalised bipower (and multipower) variation is relatively simple once this fundamental result is established. The proof of this result is done\ in a number of steps, some of them following fairly standard reasoning, others requiring special techniques. The first step is to rewrite the left hand side of (\ref{eqn 0}) as follows% \begin{eqnarray*} &&\sqrt{n}\left( \frac{1}{n}\,\sum_{i=1}^{[nt]}g(\sqrt{n}\,\triangle _{i}^{n}Y)-\int_{0}^{t}\rho _{\sigma _{u}}(g)\mathrm{d}u\right) \\ &=&\frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\left\{ g(\sqrt{n}\,\triangle _{i}^{n}Y)-\mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1% }{n}}\right] \right\} \, \\ &&+\sqrt{n}\left( \frac{1}{n}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\int_{0}^{t}\rho _{\sigma _{u}}(g)\mathrm{d}u\right) . \end{eqnarray*}% It is rather straightforward to show that the first term of the right hand side satisfies \begin{equation*} \frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\left\{ \,g(\sqrt{n}\,\triangle _{i}^{n}Y)-\mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1% }{n}}\right] \right\} \rightarrow \int_{0}^{t}\sqrt{\rho _{\sigma _{u}}(g^{2})-\rho _{\sigma _{u}}(g)^{2}}\mathrm{d}B_{u}. \end{equation*}% Hence what remains is to verify that% \begin{equation} \sqrt{n}\left( \frac{1}{n}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\int_{0}^{t}\rho _{\sigma _{u}}(g)\mathrm{d}u\right) \rightarrow 0. \label{2} \end{equation}% We have% \begin{eqnarray} &&\sqrt{n}\left( \frac{1}{n}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\int_{0}^{t}\rho _{\sigma _{u}}(g)\mathrm{d}u\right) \notag \\ &=&\frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\sqrt{n}% \sum_{i=1}^{[nt]}\int_{(i-1)/n}^{i/n}\rho _{\sigma _{u}}(g)\mathrm{d}u \notag \\ &&+\sqrt{n}\left( \sum_{i=1}^{[nt]}\int_{(i-1)/n}^{i/n}\rho _{\sigma _{u}}(g)% \mathrm{d}u-\int_{0}^{t}\rho _{\sigma _{u}}(g)\mathrm{d}u\right) \label{3} \end{eqnarray}% where% \begin{equation*} \sqrt{n}\left\{ \sum_{i=1}^{[nt]}\int_{(i-1)/n}^{i/n}\rho _{\sigma _{u}}(g)% \mathrm{d}u-\int_{0}^{t}\rho _{\sigma _{u}}(g)\mathrm{d}u\right\} \rightarrow 0. \end{equation*}% The first term on the right hand side of (\ref{3}) is now split into the difference of% \begin{equation} \frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\left\{ \mathrm{E}\left[ g(\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\rho _{\frac{i-1}{n}% }\right\} \label{4} \end{equation}% where \begin{equation*} \rho _{\frac{i-1}{n}}=\rho _{\sigma _{\frac{i-1}{n}}}(g)=\mathrm{E}\left[ g(\sigma _{\frac{i-1}{n}}\triangle _{i}^{n}W)\,|\,\mathcal{F}_{\frac{i-1}{n}}% \right] \end{equation*}% and% \begin{equation} \sqrt{n}\sum_{i=1}^{[nt]}\int_{(i-1)/n}^{i/n}\left\{ \rho _{\sigma _{u}}(g)% \mathrm{d}u-\rho _{\frac{i-1}{n}}\right\} \mathrm{d}u. \label{4b} \end{equation}% It is rather easy to show that (\ref{4}) tends to $0$ in probability uniformly in $t$. The challenge is thus to show the same result holds for (% \ref{4b}). To handle (\ref{4b}) one splits the individual terms in the sum into% \begin{equation} \sqrt{n}\ \Phi ^{\prime }\left( \sigma _{\frac{i-1}{n}}\right) \int_{(i-1)/n}^{i/n}\left( \sigma _{u}-\sigma _{\frac{i-1}{n}}\right) \mathrm{\,d}u \label{5} \end{equation}% plus% \begin{equation} \sqrt{n}\,\int_{(i-1)/n}^{i/n}\,\left\{ \Phi (\sigma _{u})-\Phi \left( \sigma _{\frac{i-1}{n}}\right) -\Phi ^{\prime }\left( \sigma _{\frac{i-1}{n}% }\right) \cdot \left( \sigma _{u}-\sigma _{\frac{i-1}{n}}\right) \right\} \,\,\mathrm{d}u, \label{6} \end{equation}% where $\Phi (x)$ is a shorthand for $\rho _{x}(g)$ and $\Phi ^{\prime }(x)$ denotes the derivative with respect to $x$.\ That (\ref{6})\ tends to $0$\ may be shown via splitting it into two terms, each of which tends to $0$\ as is verified by a sequence of inequalities, using in particular Doob's inequality. To prove that (\ref{5}) converges to $0$, again one splits, this time into three terms, using the differentiability of $g$\ in the relevant regions and the mean value theorem for differentiable functions. The two first of these terms can be handled by relatively simple means, the third poses the most difficult part of the whole proof and is treated via splitting it into seven parts. It is at this stage that the assumption that $% g$\ be even comes into play and is crucial. This section has six other subsections. In subsection \ref% {subsection:conventions} we introduce our basic notation, while in \ref% {subsection:model assumptions} we set out the model and review the assumptions we use. In subsection \ref{subsection:main result} we state the theorem we will prove. Subsections \ref{subsection: intermediate limiting results}, \ref{subsection:13b} and \ref{subsection:proof of 13a} give the proofs of the successive steps. \subsection{Notational conventions \label{subsection:conventions}} All processes mentioned in the following are defined on a given filtered probability space $(\Omega ,\mathcal{F},(\mathcal{F}_{t}),P)$. We shall in general use standard notation and conventions. For instance, given a process $(Z_{t})$ we write \begin{equation*} \triangle _{i}^{n}Z:=Z_{\frac{i}{n}}-Z_{\frac{i-1}{n}},\ \ \ i,n\geq 1. \end{equation*} We are mainly interested in convergence in law of sequences of c\`{a}dl\`{a}% g processes. In fact all results to be proved will imply convergence `stably in law' which is a slightly stronger notion. For this we shall use the notation \begin{equation*} (Z_{t}^{n})\rightarrow (Z_{t}), \end{equation*}% where $(Z_{t}^{n})$ and $(Z_{t})$ are given c\`{a}dl\`{a}g processes. Furthermore we shall write \begin{equation*} (Z_{t}^{n})\overset{P}{\rightarrow }0\ \ \ \text{meaning}\ \ \sup_{0\leq s\leq t}|Z_{s}^{n}|\rightarrow 0\ \ \mbox{\rm in\ probability\ for\ all}\ t\geq 0, \end{equation*}% \begin{equation*} (Z_{t}^{n})\overset{P}{\rightarrow }(Z_{t})\ \ \ \text{meaning}\ \ (Z_{t}^{n}-Z_{t})\overset{P}{\rightarrow }0. \end{equation*}% Often \begin{equation*} Z_{t}^{n}=\sum_{i=1}^{[nt]}a_{i}^{n}\ \ \ \text{for all}\ t\geq 0, \end{equation*}% where the $a_{i}^{n}$'s are $\mathcal{F}_{\frac{i-1}{n}}$-measurable. Recall here that given c\`{a}dl\`{a}g processes $(Z_{t}^{n}),\,(Y_{t}^{n})$ and $% (Z_{t})$ we have \begin{equation*} (Z_{t}^{n})\rightarrow (Z_{t})\ \ \text{if}\ \ (Z_{t}^{n}-Y_{t}^{n})\overset{% P}{\rightarrow }0\ \ \text{and}\ \ (Y_{t}^{n})\rightarrow (Z_{t}).\vspace{1mm% } \end{equation*} Moreover, for $h:\mathbf{R}\rightarrow \mathbf{R}$ Borel measurable of at most polynomial growth we note that $x\mapsto \rho _{x}(h)$ is locally bounded and continuous if $h$ is continuous at $0$.\vspace{1mm}\newline In what follows many arguments will consist of a series of estimates of terms indexed by $i,n$ and $t$. In these estimates we shall denote by $C$ a finite constant which may vary from place to place. Its value will depend on the constants and quantities appearing in the assumptions of the model but it is always independent of $i,n$ and $t$. \subsection{Model and basic assumptions \label{subsection:model assumptions}} Throughout the following $(W_{t})$ denotes a $((\mathcal{F}_{t}),P)$-Wiener process and $(\sigma _{t})$ a given c\`{a}dl\`{a}g $(\mathcal{F}_{t})$% -adapted process. Define \begin{equation*} Y_{t}:=\int_{0}^{t}\sigma _{s-}\,\mathrm{d}W_{s}\ \ \ \ \ t\geq 0, \end{equation*}% implying that is $(Y_{t})$ is a continuous local martingale. We have deleted the drift of the $\left( Y_{t}\right) $ process as taking care of it is a simple technical task, while its presence increase the clutter of the notation. Our aim is to study the asymptotic behaviour of the processes \begin{equation*} \{(X_{t}^{n}(g))\,|\,n\geq 1\,\} \end{equation*}% where% \begin{equation*} X_{t}^{n}(g)=\frac{1}{n}\,\sum_{i=1}^{[nt]}g(\sqrt{n}\,\triangle _{i}^{n}Y),\ \ \ t\geq 0,\,n\geq 1. \end{equation*}% Here $g:\mathbf{R}\rightarrow \mathbf{R}$ is a given continuous function of at most polynomial growth. We are especially interested in $g$'s of the form $x\mapsto |x|^{r}\ (r>0)$ but we shall keep the general notation since nothing is gained in simplicity by assuming that $g$ is of power form. We shall throughout the following assume that $g$ furthermore satisfies the following. \begin{assumption} \textbf{(K)}: $g$ is an even function and continuously differentiable in $% B^{c}$ where $B\subseteq \mathbf{R}$ is a closed Lebesgue null-set and $% \exists \ M,\,p\geq 1$ such that \begin{equation*} |g(x+y)-g(x)|\leq M(1+|x|^{p}+|y|^{p})\cdot |y|\ , \end{equation*}% for all $x,y\in \mathbf{R}$. \end{assumption} \begin{remark} The assumption (K) implies, in particular, that if $x\in B^{c}$ then \begin{equation*} |g^{\prime }(x)|\leq M(1+|x|^{p}).\vspace{1mm} \end{equation*}% Observe that only power functions corresponding to $r\geq 1$ do satisfy (K). The remaining case $0<r<1$ requires special arguments which will be omitted here\vspace{1mm} (for details see \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)}). \end{remark} In order to prove the CLT-theorem we need some additional structure on the volatility process $(\sigma _{t})$. A natural set of assumptions would be the following. \begin{assumption} \textbf{(H0)}: $(\sigma _{t})$ can be written as \begin{equation*} \sigma _{t}=\sigma _{0}+\int_{0}^{t}a_{s}^{\ast }\,\mathrm{d}% s+\int_{0}^{t}\sigma _{s}^{\ast }\,\mathrm{d}W_{s}+\int_{0}^{t}v_{s-}^{\ast }\,\mathrm{d}Z_{s} \end{equation*}% where $(Z_{t})$ is a $((\mathcal{F}_{t}),P)$-L\'{e}vy process independent of $(W_{t})$ and $(\sigma _{t}^{\ast })$ and $(v_{t}^{\ast })$ are adapted c% \`{a}dl\`{a}g processes and $(a_{t}^{\ast })$ a predictable locally bounded process. \end{assumption} However, in modelling volatility it is often more natural to define $(\sigma _{t}^{2})$ as being of the above form, i.e.% \begin{equation*} \sigma _{t}^{2}=\sigma _{0}^{2}+\int_{0}^{t}a_{s}^{\ast }\,\mathrm{d}% s+\int_{0}^{t}\sigma _{s}^{\ast }\,\mathrm{d}W_{s}+\int_{0}^{t}v_{s-}^{\ast }\,\mathrm{d}Z_{s}. \end{equation*}% Now this does not in general imply that $(\sigma _{t})$ has the same form; therefore we shall replace (H0) by the more general structure given by the following assumption. \begin{assumption} \textbf{(H1)}: $(\sigma _{t})$ can be written, for $t\geq 0$, as% \begin{equation*} \begin{array}{lll} \sigma _{t} & = & \displaystyle\sigma _{0}+\int_{0}^{t}a_{s}^{\ast }\,% \mathrm{d}s+\int_{0}^{t}\sigma _{s}^{\ast }\,\mathrm{d}W_{s}+% \int_{0}^{t}v_{s-}^{\ast }\,\mathrm{d}V_{s} \\ & & \displaystyle+\int_{0}^{t}\int_{E}q\circ \phi (s-,x)\,(\mu -\nu )(% \mathrm{d}s\,\mathrm{d}x) \\ & & \displaystyle+\int_{0}^{t}\int_{E}\ \left\{ \phi (s-,x)-q\circ \phi (s-,x)\right\} \,\mu (\mathrm{d}s\,\mathrm{d}x).% \end{array}% \end{equation*}% Here $(a_{t}^{\ast }),\,(\sigma _{t}^{\ast })$ and $(v_{t}^{\ast })$ are as in (H0) and $(V_{t})$ is another $((\mathcal{F}_{t}),P)$-Wiener process independent of $(W_{t})$ while $q$ is a continuous truncation function on $% \mathbf{R}$, i.e. a function with compact support coinciding with the identity on a neighbourhood of $0$. Further $\mu $ is a Poisson random measure on $(0,\infty )\times E$ independent of $(W_{t})$ and $(V_{t})$ with intensity measure $\nu (\mathrm{d}s\,\mathrm{d}x)=\mathrm{d}s\otimes F(% \mathrm{d}x)$, $F$ being a $\sigma $-finite mea\-sure on a measurable space $% (E,\mathcal{E})$ and \begin{equation*} (\omega ,s,x)\mapsto \phi (\omega ,s,x) \end{equation*}% is a map from $\Omega \times \,[\,0,\infty )\times E$ into $\mathbf{R}$ which is $\mathcal{F}_{s}\otimes \mathcal{E}$ measurable in $(\omega ,x)$ for all $s$ and c\`{a}dl\`{a}g in $s$, satisfying furthermore that for some sequence of stopping times $(S_{k})$ increasing to $+\infty $ we have for all $k\geq 1$ \begin{equation*} \int_{E}\left\{ 1\wedge \psi _{k}(x)^{2}\right\} \,F(\mathrm{d}x)<\infty , \end{equation*}% where% \begin{equation*} \psi _{k}(x)=\sup_{\omega \in \Omega ,\,s<S_{k}(\omega )}|\phi (\omega ,s,x)|. \end{equation*} \end{assumption} \begin{remark} (H1) is weaker than (H0), and if $(\sigma _{t}^{2})$ satisfies (H1) then so does $(\sigma _{t})$.\newline \end{remark} Finally we shall also assume a non-degeneracy in the model. \begin{assumption} \textbf{(H2)}: $(\sigma _{t})$ satisfies \begin{equation*} 0<\sigma _{t}^{2}(\omega )\ \text{for all}\ (t,\omega ).\vspace{1mm} \end{equation*} \end{assumption} According to general stochastic analysis theory it is known that to prove convergence in law of a sequence $(Z_{t}^{n})$ of c\`{a}dl\`{a}g processes it suffices to prove the convergence of each of the stopped processes $% (Z_{T_{k}\wedge t}^{n})$ for at least one sequence of stopping times $% (T_{k}) $ increasing to $+\infty $. Applying this together with standard localisation techniques (for details see \cite% {BarndorffNielsenGraversenJacodPodolskyShephard(04shiryaev)}), we may assume that the following more restrictive assumptions are satisfied. \begin{assumption} \textbf{(H1a)}: $(\sigma _{t})$ can be written as \begin{equation*} \sigma _{t}=\sigma _{0}+\int_{0}^{t}a_{s}^{\ast }\,ds+\int_{0}^{t}\sigma _{s-}^{\ast }\,\mathrm{d}W_{s}+\int_{0}^{t}v_{s-}^{\ast }\,\mathrm{d}% V_{s}+\int_{0}^{t}\int_{E}\phi (s-,x)(\mu -\nu )(\mathrm{d}s\,\mathrm{d}x)\ \ \ t\geq 0.\vspace{1mm} \end{equation*}% Here $(a_{t}^{\ast }),\,(\sigma _{t}^{\ast })$ and $(v_{t}^{\ast })$ are real valued uniformly bounded c\`{a}dl\`{a}g $(\mathcal{F}_{t})$-adapted proces\-ses; $(V_{t})$ is another $((\mathcal{F}_{t}),P)$-Wiener process independent of $(W_{t})$. Further $\mu $ is a Poisson random measure on $% (0,\infty )\times E$ independent of $(W_{t})$ and $(V_{t})$ with intensity measure $\nu (\mathrm{d}s\,\mathrm{d}x)=\mathrm{d}s\otimes F(\mathrm{d}x)$, $% F$ being a $\sigma $-finite mea\-sure on a measurable space $(E,\mathcal{E})$ and \begin{equation*} (\omega ,s,x)\mapsto \phi (\omega ,s,x) \end{equation*}% is a map from $\Omega \times \,[\,0,\infty )\times E$ into $\mathbf{R}$ which is $\mathcal{F}_{s}\otimes \mathcal{E}$ measurable in $(\omega ,x)$ for all $s$ and c\`{a}dl\`{a}g in $s$, satisfying furthermore \begin{equation*} \psi (x)=\sup_{\omega \in \Omega ,\,s\geq 0}|\phi (\omega ,s,x)|\leq M<\infty \ \ \text{and}\ \ \int \psi (x)^{2}\,F(\mathrm{d}x)<\infty .\vspace{% 1mm} \end{equation*} \end{assumption} Likewise, by a localisation argument, we may assume \begin{assumption} \textbf{(H2a)}: $(\sigma _{t})$ satisfies \begin{equation*} a<\sigma _{t}^{2}(\omega )<b\ \ \ \text{for all}\ (t,\omega )\ \text{for some% }\ a,b\in (0,\infty ).\vspace{1mm} \end{equation*} \end{assumption} Observe that under the more restricted assumptions $(Y_{t})$ is a continuous martingale having moments of all orders and $(\sigma _{t})$ is represented as a sum of three square integrable martingales plus a continuous process of bounded variation. Furthermore, the increments of the increasing processes corresponding to the three martingales and of the bounded variation process are dominated by a constant times $\triangle t$, implying in particular that \begin{equation} \mathrm{E}\left[ \,\left\vert \sigma _{v}-\sigma _{u}\right\vert ^{2}\right] \leq C\,(v-u),\ \ \ \ \text{for all}\ 0\leq u<v.\vspace{2mm} \label{8} \end{equation} \subsection{Main result \label{subsection:main result}\newline } As already mentioned, our aim is to show the following special version of the general CLT-result given as Theorem \ref{TT3}. \begin{theorem} \vspace{2mm}Under assumptions (K), (H1a) and (H2a), there exists a Wiener process $(B_{t})$ defined on some extension of $(\Omega ,\mathcal{F},(% \mathcal{F}_{t}),P)$ and independent of $\mathcal{F}$ such that% \begin{equation} \left( \sqrt{n}\left( \,\ \frac{1}{n}\,\sum_{i=1}^{[nt]}g(\sqrt{n}% \,\triangle _{i}^{n}Y)-\int_{0}^{t}\rho _{\sigma _{u}}(g)\,\mathrm{d}% u\,\right) \right) \rightarrow \int_{0}^{t}\sqrt{\rho _{\sigma _{u-}}(g^{2})-\rho _{\sigma _{u-}}(g)^{2}}\,\mathrm{d}B_{u}. \label{Main result} \end{equation}% \emph{\ } \end{theorem} Introducing the notation% \begin{equation*} U_{t}(g)=\int_{0}^{t}\sqrt{\rho _{\sigma _{u-}}(g^{2})-\rho _{\sigma _{u-}}(g)^{2}}\,\mathrm{d}B_{u}\ \ \ t\geq 0\vspace{1mm} \end{equation*}% we may reexpress (\ref{Main result}) as \begin{equation} \left( \sqrt{n}\,\left( X_{t}^{n}(g)-\int_{0}^{t}\sigma _{u}(g)\,\mathrm{d}% u\right) \,\right) \rightarrow (U_{t}(g)). \label{Main result reform} \end{equation}% To prove this, introduce the set of variables $\{\beta _{i}^{n}\,|\,i,\,n\geq 1\}$ given by \begin{equation*} \beta _{i}^{n}=\sqrt{n}\cdot \sigma _{\frac{i-1}{n}}\cdot \triangle _{i}^{n}W,\ \ \ i,\,n\geq 1. \end{equation*} The $\beta _{i}^{n}$'s should be seen as approximations to $\sqrt{n}% \,\triangle _{i}^{n}Y$. In fact, since \begin{equation*} \sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}=\sqrt{n}\,\int_{(i-1)/n}^{i/n}(% \sigma _{s}-\sigma _{\frac{i-1}{n}})\,\mathrm{d}W_{s} \end{equation*}% and $(\sigma _{t})$ is uniformly bounded, a straightforward application of (% \ref{8}) and the Burkholder-Davis-Gundy-inequalities (e.g. \cite[pp. 160-171]% {RevuzYor(99)}) gives for every $p>0$ the following simple estimates. \begin{equation} \mathrm{E}\left[ \,|\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|^{p}\,|\,% \mathcal{F}_{\frac{i-1}{n}}\right] \leq \frac{C_{p}}{n^{p\wedge 1}} \end{equation}% and \begin{equation} \mathrm{E}\left[ \,|\sqrt{n}\,\triangle _{i}^{n}Y|^{p}+|\beta _{i}^{n}|^{p}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \leq C_{p}\vspace{1mm} \label{12} \end{equation}% for all $i,n\geq 1$. Observe furthermore that \begin{equation*} \mathrm{E}\left[ g(\beta _{i}^{n})\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] =\rho _{\sigma _{\frac{i-1}{n}}}(g),\ \ \ \text{for all}\ i,\,n\geq 1.% \vspace{1mm} \end{equation*} Introduce for convenience, for each $t>0$ and $n\geq 1$, the shorthand notation \begin{equation*} U_{t}^{n}(g)=\frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\,\,\left\{ g(\sqrt{n}% \,\triangle _{i}^{n}Y)-\mathrm{E}\left[ g(\sqrt{n}\,\triangle _{i}^{n}Y)\,|\,% \mathcal{F}_{\frac{i-1}{n}}\right] \right\} \, \end{equation*}% and \begin{equation*} \tilde{U}_{t}^{n}(g)=\frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\,\,\left\{ g(\beta _{i}^{n})-\rho _{\sigma _{\frac{i-1}{n}}}(g)\right\} =\frac{1}{\sqrt{% n}}\,\sum_{i=1}^{[nt]}\,\,\left\{ g(\beta _{i}^{n})-\mathrm{E}\left[ g(\beta _{i}^{n})\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \right\} \,.\vspace{1mm} \end{equation*}% The asymptotic behaviour of $(\tilde{U}_{t}^{n}(g))$ is well known. More precisely under the the given assumptions\thinspace (\thinspace in fact much less is needed\thinspace ) we have \begin{equation*} (U_{t}^{n}(g))\rightarrow (U_{t}(g)).\vspace{1mm} \end{equation*}% This result is a rather straightforward consequence of \cite[Theorem IX.7.28]% {JacodShiryaev(03)}. Thus, if $(U_{t}^{n}(g)-\tilde{U}_{t}^{n}(g))\overset{P}% {\rightarrow }0$ we may deduce the following result. \begin{theorem} \label{theorem B}\ \ \emph{Let $(B_{t})$ and $(U_{t}(g))$ be as above. Then} \begin{equation*} (\tilde{U}_{t}^{n}(g))\rightarrow (U_{t}(g)).\vspace{1mm} \end{equation*} \end{theorem} \noindent \textbf{Proof.} As pointed out just above it is enough to prove that \begin{equation*} (U_{t}^{n}(g)-\tilde{U}_{t}^{n}(g))\overset{P}{\rightarrow }0. \end{equation*}% But for $t\geq 0$ and $n\geq 1$ \begin{equation*} U_{t}^{n}(g)-\tilde{U}_{t}^{n}(g)=\sum_{i=1}^{[nt]}\,\left( \xi _{i}^{n}-% \mathrm{E}\left[ \xi _{i}^{n}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \right) \end{equation*}% where \begin{equation*} \xi _{i}^{n}=\frac{1}{\sqrt{n}}\left\{ g(\sqrt{n}\triangle _{i}^{n}Y)-g(\beta _{i}^{n})\right\} ,\ \ \ i,n\geq 1. \end{equation*}% Thus we have to prove \begin{equation*} \left( \,\sum_{i=1}^{[nt]}\,\left\{ \xi _{i}^{n}-\mathrm{E}\left[ \xi _{i}^{n}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \right\} \right) \overset{P}{% \rightarrow }0. \end{equation*}% But, as the left hand side of this relation is a sum of martingale differences, this is implied by Doob's inequality (e.g. \cite[pp. 54-55]% {RevuzYor(99)}) if for all $t>0$ \begin{equation*} \sum_{i=1}^{[nt]}\,\mathrm{E}[(\xi _{i}^{n})^{2}]=\mathrm{E}% [\,\sum_{i=1}^{[nt]}\,\mathrm{E}[(\xi _{i}^{n})^{2}\,|\,\mathcal{F}_{\frac{% i-1}{n}}]\,]\rightarrow 0\ \ \ \text{as}\ n\rightarrow \infty .\vspace{1mm} \end{equation*}% Fix $t>0$. Using the Cauchy-Schwarz inequality and the Burkholder-Davis-Gundy inequalities we have for all $i,n\geq 1$. \begin{eqnarray*} \mathrm{E}\left[ (\xi _{i}^{n})^{2}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] &=&\frac{1}{n}\,\mathrm{E}\left[ \left\{ g(\sqrt{n}\triangle _{i}^{n}Y)-\beta _{i}^{n}+\beta _{i}^{n}-g(\beta _{i}^{n})\right\} ^{2}\,|\,% \mathcal{F}_{\frac{i-1}{n}}\right] \\ &\leq &\frac{C}{n}\,\mathrm{E}\left[ \,(1+|\sqrt{n}\triangle _{i}^{n}Y|^{p}+|\beta _{i}^{n}|^{p})^{2}\cdot (\sqrt{n}\triangle _{i}^{n}Y-\beta _{i}^{n})^{2}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \\ &\leq &\frac{C}{n}\,\sqrt{\mathrm{E}\left[ \,(1+|\sqrt{n}\triangle _{i}^{n}Y|^{2p}+|\beta _{i}^{n}|^{2p})\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] }\cdot \sqrt{\mathrm{E}\left[ (\sqrt{n}\triangle _{i}^{n}Y-\beta _{i}^{n})^{4}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] } \\ &\leq &C\,\sqrt{\mathrm{E}\left[ \,\left( \int_{(i-1)/n}^{i/n}\left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) \,\mathrm{d}W_{u}\right) ^{4}\,|\,% \mathcal{F}_{\frac{i-1}{n}}\right] } \\ &\leq &C\,\sqrt{\mathrm{E}\left[ \left( \int_{(i-1)/n}^{i/n}\left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) ^{2}\,\mathrm{d}u\right) ^{2}\,|\,% \mathcal{F}_{\frac{i-1}{n}}\right] }. \end{eqnarray*}% Thus \begin{eqnarray*} \sum_{i=1}^{[nt]}\,\mathrm{E}[(\xi _{i}^{n})^{2}] &\leq &Cn\,\frac{t}{n}% \,\sum_{i=1}^{[nt]}\mathrm{E}\,\left[ \sqrt{\mathrm{E}\,\left[ \left( \int_{(i-1)/n}^{i/n}\left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) ^{2}\,% \mathrm{d}u\right) ^{2}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] }\right] \, \\ &\leq &C\,tn\,\sqrt{\frac{1}{n}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ \left( \int_{(i-1)/n}^{i/n}\left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) ^{2}\,% \mathrm{d}u\right) ^{2}\right] } \\ &\leq &Ctn\,\sqrt{\frac{1}{n^{2}}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ \,\int_{(i-1)/n}^{i/n}\left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) ^{4}\,\mathrm{d}u\right] \,} \\ &\leq &Ct\,\sqrt{\,\sum_{i=1}^{[nt]}\int_{(i-1)/n}^{i/n}\mathrm{E}\left[ \left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) ^{2}\,\right] \mathrm{d}% u\,} \\ &\rightarrow &\,0\vspace{2mm}, \end{eqnarray*}% as $n\rightarrow \infty $ by Lebesgue's Theorem and the boundedness of $% (\sigma _{t})$. \noindent $\square $ To prove the convergence (\ref{Main result reform}) it suffices, using Theorem \ref{theorem B} above, to prove that \begin{equation*} \left( U_{t}^{n}(g)-\sqrt{n}\,\left\{ \,X_{t}^{n}(g)-\int_{0}^{t}\rho _{\sigma _{u}}(g)\,\mathrm{d}u\right\} \,\right) \overset{P}{\rightarrow }0.% \vspace{1mm} \end{equation*}% But as \begin{equation*} U_{t}^{n}(g)-\sqrt{n}\,X_{t}^{n}(g)=-\frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}% \mathrm{E}\left[ \,g(\sqrt{n}\,\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{% i-1}{n}}\right] \end{equation*}% and, as is easily seen, \begin{equation*} \left( \sqrt{n}\,\int_{0}^{t}\rho _{\sigma _{u}}\,\left( g\right) \mathrm{d}% u-\,\sum_{i=1}^{[nt]}\sqrt{n}\,\int_{(i-1)/n}^{i/n}\rho _{\sigma _{u}}(g)\,% \mathrm{d}u\right) \overset{P}{\rightarrow }0,\vspace{2mm} \end{equation*}% the job is to prove that \begin{equation*} \,\sum_{i=1}^{[nt]}\eta _{i}^{n}\,\overset{P}{\rightarrow }0\ \ \ \text{for all}\ t>0, \end{equation*}% where for $i,n\geq 1$ \begin{equation*} \eta _{i}^{n}=\,\frac{1}{\sqrt{n}}\,\mathrm{E}\,\left[ g(\sqrt{n}\,\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\sqrt{n}% \,\int_{(i-1)/n}^{i/n}\rho _{\sigma _{u}}(g)\,\mathrm{d}u.\vspace{1mm} \end{equation*}% Fix $t>0$ and write, for all $i,n\geq 1$, \begin{equation*} \eta _{i}^{n}=\eta (1)_{i}^{n}+\eta (2)_{i}^{n} \end{equation*}% where \begin{equation} \eta (1)_{i}^{n}=\frac{1}{\sqrt{n}}\,\,\left\{ \mathrm{E}\left[ \,g(\sqrt{n}% \,\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] -\rho _{\sigma _{\frac{i-1}{n}}}(g)\,\right\} \end{equation}% and \begin{equation} \eta (2)_{i}^{n}=\sqrt{n}\,\int_{(i-1)/n}^{i/n}\left\{ \rho _{\sigma _{u}}(g)-\rho _{\sigma _{\frac{i-1}{n}}}(g)\right\} \mathrm{d}u.\vspace{1mm} \end{equation} We will now separately prove \begin{equation} \,\eta (1)^{n}=\sum_{i=1}^{[nt]}\eta (1)_{i}^{n}\,\overset{P}{\rightarrow }0 \label{13a} \end{equation}% and% \begin{equation} \,\eta (2)^{n}=\,\sum_{i=1}^{[nt]}\eta (2)_{i}^{n}\,\overset{P}{\rightarrow }% 0\vspace{1mm}. \label{13b} \end{equation} \subsection{Some auxiliary estimates\label{subsection: intermediate limiting results}} In order to show (\ref{13a}) and (\ref{13b}) we need some refinements of the estimate (\ref{8}) above. To state these we split up $(\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})$ into several terms. By definition \begin{equation*} \sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}=\sqrt{n}\,\int_{(i-1)/n}^{i/n}% \,\left( \sigma _{u-}-\sigma _{\frac{i-1}{n}}\right) \,\mathrm{d}W_{u}% \vspace{1mm} \end{equation*}% for all $i,n\geq 1$. Writing \begin{equation*} E_{n}=\{x\in \mathrm{E}\,|\,|\Psi (x)|>1/\sqrt{n}\,\} \end{equation*}% the difference $\sigma _{u}-\sigma _{\frac{i-1}{n}}$ equals \begin{eqnarray*} &&\int_{(i-1)/n}^{u}a_{s}^{\ast }\,ds+\int_{(i-1)/n}^{u}\sigma _{s-}^{\ast }\,\mathrm{d}W_{s}+\int_{(i-1)/n}^{u}v_{s-}^{\ast }\,\mathrm{d}% V_{s}+\int_{(i-1)/n}^{u}\int_{E}\phi (s-,x)\,(\mu -\nu )(\mathrm{d}s\,% \mathrm{d}x)\vspace{1mm} \\ &=&\displaystyle\sum_{j=1}^{5}\xi (j)_{i}^{n}(u), \end{eqnarray*}% for $i,n\geq 1$ and $u\geq (i-1)/n$ where \begin{eqnarray*} \displaystyle\xi (1)_{i}^{n}(u) &=&\int_{(i-1)/n}^{u}a_{s}^{\ast }\,\mathrm{d% }s+\int_{(i-1)/n}^{u}\,\left( \sigma _{s-}^{\ast }-\sigma _{\frac{i-1}{n}% }^{\ast }\right) \,\mathrm{d}W_{s}+\int_{(i-1)/n}^{u}\,\left( v_{s-}^{\ast }-v_{\frac{i-1}{n}}^{\ast }\right) \,\mathrm{d}V_{s} \\ \displaystyle\xi (2)_{i}^{n}(u) &=&\sigma _{\frac{i-1}{n}}^{\ast }\,\left( W_{u}-W_{\frac{i-1}{n}}\right) +v_{\frac{i-1}{n}}^{\ast }\,\left( V_{u}-V_{% \frac{i-1}{n}}\right) \\ \displaystyle\xi (3)_{i}^{n}(u) &=&\int_{(i-1)/n}^{u}\int_{E_{n}^{c}}\phi (s-,x)\,(\mu -\nu )(\mathrm{d}s\,\mathrm{d}x) \\ \displaystyle\xi (4)_{i}^{n}(u) &=&\int_{(i-1)/n}^{u}\int_{E_{n}}\left\{ \phi (s-,x)-\phi \left( \frac{i-1}{n},x\right) \right\} \,(\mu -\nu )(% \mathrm{d}s\,\mathrm{d}x) \\ \displaystyle\xi (5)_{i}^{n}(u) &=&\int_{(i-1)/n}^{u}\int_{E_{n}}\phi \left( \frac{i-1}{n},x\right) \,(\mu -\nu )(\mathrm{d}s\,\mathrm{d}x)\vspace{1mm} \end{eqnarray*}% That is, for $i,n\geq 1$, \begin{equation} \sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}=\sum_{j=1}^{5}\xi (j)_{i}^{n} \label{17} \end{equation}% where \begin{equation*} \ \xi (j)_{i}^{n}=\sqrt{n}\,\int_{(i-1)/n}^{i/n}\,\xi (j)_{i}^{n}(u-)\,% \mathrm{d}W_{u}\ \ \ \ \mbox{\rm for}\ j=1,2,3,4,5.\vspace{1mm} \end{equation*}% The specific form of the variables implies, using Burkholder-Davis-Gundy inequalities, that for every $q\geq 2$ we have \begin{eqnarray*} \mathrm{E}[\,|\xi (j)_{i}^{n}|^{q}\,] &\leq &C_{q}\,n^{q/2}\,\mathrm{E}\,% \left[ \left( \int_{(i-1)/n}^{i/n}\,\xi (j)_{i}^{n}(u)^{2}\,\mathrm{d}% u\right) ^{q/2}\right] \\ &\leq &\displaystyle n\int_{(i-1)/n}^{i/n}\,\mathrm{E}[\,|\xi (j)_{i}^{n}(u)|^{q}\,]\mathrm{\,d}u \\ &\leq &\displaystyle\sup_{(i-1)/n\leq u\leq i/n}\,\mathrm{E}[\,|\xi (j)_{i}^{n}(u)|^{q}\,] \end{eqnarray*}% for all $i,n\geq 1$ and all $j$. These terms will now be estimated. This is done in the following series of lemmas where $i$ and $n$ are arbitrary and we use the notation \begin{equation*} d_{i}^{n}=\int_{(i-1)/n}^{i/n}\mathrm{E}\,\left[ \left( \sigma _{s-}^{\ast }-\sigma _{\frac{i-1}{n}}^{\ast }\right) ^{2}+\left( v_{s-}^{\ast }-v_{\frac{% i-1}{n}}^{\ast }\right) ^{2}+\int_{E}\left\{ \phi (s-,x)-\phi \left( \frac{% i-1}{n},x\right) \right\} ^{2}\,F(\mathrm{d}x)\right] \mathrm{d}s.\vspace{1mm% } \end{equation*} \begin{lemma} \label{lemma 1st} \begin{equation*} \mathrm{E}[\,(\xi (1)_{i}^{n})^{2}]\leq C_{1}\cdot (1/n^{2}+d_{i}^{n}). \end{equation*} \end{lemma} \begin{lemma} \label{lemma 2nd}% \begin{equation*} \mathrm{E}[\,(\xi (2)_{i}^{n})^{2}]\leq C_{2}/n. \end{equation*} \end{lemma} \begin{lemma} \begin{equation*} \mathrm{E}[\,(\xi (3)_{i}^{n})^{2}]\leq C_{3}\,\varphi (1/\sqrt{n})/n, \end{equation*}% where% \begin{equation*} \varphi (\epsilon )=\int_{\{\,|\Psi |\leq \epsilon \,\}}\Psi (x)^{2}\,F(% \mathrm{d}x). \end{equation*} \end{lemma} \begin{lemma} \begin{equation*} \mathrm{E}[\,(\xi (4)_{i}^{n})^{2}]\leq C_{4}\,d_{i}^{n}. \end{equation*} \end{lemma} \begin{lemma} \label{lemma 5th} \begin{equation*} \mathrm{E}[\,(\xi (5)_{i}^{n})^{2}]\leq C_{5}/n. \end{equation*} \end{lemma} The proofs of these five Lemmas rely on straightforward martingale inequalities. Observe that Lebesgue's Theorem ensures, since the processes involved are assumed c\`{a}dl\`{a}g and uniformly bounded, that as $n\rightarrow \infty $ \begin{equation*} \sum_{i=1}^{[nt]}d_{i}^{n}\,\rightarrow 0\ \ \ \ \text{for all}\ t>0.\vspace{% 1mm} \end{equation*} Taken together these statements imply the following result. \begin{corollary} \noindent \emph{For\ all\ $t>0$ as }$n\rightarrow \infty $\emph{\ } \begin{equation*} \sum_{i=1}^{[nt]}\,\left\{ \mathrm{E}[\,(\xi (1)_{i}^{n})^{2}]+\mathrm{E}% [\,(\xi (3)_{i}^{n})^{2}]+\mathrm{E}[\,(\xi (4)_{i}^{n})^{2}]\right\} \,)\,\rightarrow 0.\vspace{1mm} \end{equation*} \end{corollary} Below we shall invoke this Corollary as well as Lemmas \ref{lemma 2nd} and % \ref{lemma 5th}.\ \newline \subsection{Proof of $\,\protect\eta (2)^{n}\protect\overset{P}{\rightarrow }% 0$ \label{subsection:13b}} Recall we wish to show that \begin{equation} \,\eta (2)^{n}=\sum_{i=1}^{[nt]}\eta (2)_{i}^{n}\,\overset{P}{\rightarrow }0. \label{eqn 7} \end{equation}% >From now on let $t>0$ be fixed. We split the $\eta (2)_{i}^{n}$'s according to \begin{equation*} \eta (2)_{i}^{n}=\eta ^{\prime }(2)_{i}^{n}+\eta ^{\prime \prime }(2)_{i}^{n}\ \ \ \ i,n\geq 1 \end{equation*}% where, writing $\Phi (x)$ for $\rho _{x}(g)$, \begin{equation*} \eta ^{\prime }(2)_{i}^{n}=\sqrt{n}\ \Phi ^{\prime }\left( \sigma _{\frac{i-1% }{n}}\right) \int_{(i-1)/n}^{i/n}\left( \sigma _{u}-\sigma _{\frac{i-1}{n}% }\right) \,\mathrm{d}u \end{equation*}% and \begin{equation*} \eta ^{\prime \prime }(2)_{i}^{n}=\sqrt{n}\,\int_{(i-1)/n}^{i/n}\,\,\left\{ \Phi (\sigma _{u})-\Phi \left( \sigma _{\frac{i-1}{n}}\right) -\Phi ^{\prime }\left( \sigma _{\frac{i-1}{n}}\right) \cdot \left( \sigma _{u}-\sigma _{% \frac{i-1}{n}}\right) \right\} \,\,\mathrm{d}u. \end{equation*}% Observe that the assumptions on $g$ imply that $x\mapsto \Phi (x)$ is differentiable with a bounded derivative on any bounded interval not including $0$; in particular\thinspace (see (H2a)) \begin{equation} |\,\Phi (x)-\Phi (y)-\Phi ^{\prime }(y)\cdot (x-y)\,|\leq \Psi (|x-y|)\cdot |x-y|,\ \ \ x^{2},y^{2}\in (a,b), \label{eqn 8} \end{equation}% where $\Psi :\mathbf{R}_{+}\rightarrow \mathbf{R}_{+}$ is continuous, increasing and $\Psi (0)=0$. \vspace{1mm} With this notation we shall prove (\ref{eqn 7}) by showing% \begin{equation*} \,\sum_{i=1}^{[nt]}\eta ^{\prime }(2)_{i}^{n}\,\overset{P}{\rightarrow }0 \end{equation*}% and \begin{equation*} \ \,\sum_{i=1}^{[nt]}\eta ^{\prime \prime }(2)_{i}^{n}\,\overset{P}{% \rightarrow }0.\vspace{1mm} \end{equation*}% Inserting the description of $(\sigma _{t})$\thinspace (see (H1a)) we may write \begin{equation*} \eta ^{\prime }(2)_{i}^{n}=\eta ^{\prime }(2,1)_{i}^{n}+\eta ^{\prime }(2,2)_{i}^{n} \end{equation*}% where for all $i,n\geq 1$ \begin{equation*} \eta ^{\prime }(2,1)_{i}^{n}=\sqrt{n}\ \Phi ^{\prime }\left( \sigma _{\frac{% i-1}{n}}\right) \int_{(i-1)/n}^{i/n}\left( \int_{(i-1)/n}^{u}a_{s}^{\ast }\,% \mathrm{d}s\right) \,\,\mathrm{d}u \end{equation*}% and \begin{eqnarray*} \eta ^{\prime }(2,2)_{i}^{n} &=&\displaystyle\sqrt{n}\ \Phi ^{\prime }\left( \sigma _{\frac{i-1}{n}}\right) \int_{(i-1)/n}^{i/n}\,\left[ \int_{(i-1)/n}^{u}\,\sigma _{s-}^{\ast }\,\mathrm{d}W_{s}+\int_{(i-1)/n}^{u}% \,v_{s-}^{\ast }\,\mathrm{d}V_{s}\right. \, \\ &&+\displaystyle\left. \int_{E}\phi (s-,x)\,(\mu -\nu )(\mathrm{d}s\,\mathrm{% d}x)\right] \mathrm{d}u.\vspace{1mm} \end{eqnarray*}% By (H2a) and (\ref{eqn 8}) and the uniform boundedness of $(a_{t}^{\ast })$ we have \begin{equation*} |\eta ^{\prime }(2,1)_{i}^{n}|\leq C\,\sqrt{n}\,\int_{(i-1)/n}^{i/n}\left\{ u-(i-1)/n\right\} \,\mathrm{d}u\leq C/n^{3/2} \end{equation*}% for all $i,n\geq 1$ and thus \begin{equation*} \,\sum_{i=1}^{[nt]}\eta ^{\prime }(2,1)_{i}^{n}\,\overset{P}{\rightarrow }0.% \vspace{1mm} \end{equation*}% Since \begin{equation*} (W_{t}),\ (V_{t})\ \text{and}\ \left( \int_{0}^{t}\int_{E}\phi (s-,x)(\mu -\nu )(\mathrm{d}s\,\mathrm{d}x)\right) \vspace{1mm} \end{equation*}% are all martingales we have \begin{equation*} \mathrm{E}\left[ \eta ^{\prime }(2,2)_{i}^{n}\,|\,\mathcal{F}_{\frac{i-1}{n}}% \right] =0\ \ \ \text{for all}\quad i,n\geq 1.\vspace{1mm} \end{equation*} By Doob's inequality it is therefore feasible to estimate \begin{equation*} \sum_{i=1}^{[nt]}\,\mathrm{E}[\,(\eta ^{\prime }(2,2)_{i}^{n})^{2}].\vspace{% 1mm} \end{equation*}% Inserting again the description of $(\sigma _{t})$ we find, applying simple inequalities, in particular Jensen's, that \begin{eqnarray*} &&(\eta ^{\prime }(2,2)_{i}^{n})^{2} \\ &\leq &\displaystyle C\,n\,\left( \int_{(i-1)/n}^{i/n}\left\{ \int_{(i-1)/n}^{u}\,\sigma _{s-}^{\ast }\,\mathrm{d}W_{s}\right\} \,\mathrm{d% }u\right) ^{2}+C\,n\,\left( \,\int_{(i-1)/n}^{i/n}\left\{ \int_{(i-1)/n}^{u}\,v_{s-}^{\ast }\,\mathrm{d}V_{s}\right\} \,\mathrm{d}% u\right) ^{2} \\ &&\displaystyle+C\,n\,\left( \,\int_{(i-1)/n}^{i/n}\int_{(i-1)/n}^{u}\left\{ \int_{E}\phi (s-,x)\,(\mu -\nu )(\mathrm{d}s\,\mathrm{d}x)\right\} \mathrm{d}% u\right) ^{2} \\ &\leq &\displaystyle C\,\int_{(i-1)/n}^{i/n}\left( \,\int_{(i-1)/n}^{u}\,\sigma _{s-}^{\ast }\,\mathrm{d}W_{s}\,\right) ^{2}\,% \mathrm{d}u+C\,\int_{(i-1)/n}^{i/n}\left( \,\int_{(i-1)/n}^{u}\,v_{s-}^{\ast }\,\mathrm{d}V_{s}\,\right) ^{2}\,\mathrm{d}u \\ &&\displaystyle+C\,\int_{(i-1)/n}^{i/n}\left( \,\int_{(i-1)/n}^{u}\int_{E}\phi (s-,x)\,(\mu -\nu )(\mathrm{d}s\,\mathrm{d}% x)\,\right) ^{2}\,\mathrm{d}u.\vspace{1mm} \end{eqnarray*}% The properties of the Wiener integrals and the uniform boundedness of $% (\sigma _{t}^{\ast })$ and $(v_{t}^{\ast })$ ensure that \begin{equation*} \mathrm{E}\left[ \left( \,\int_{(i-1)/n}^{u}\,\sigma _{s-}^{\ast }\,\mathrm{d% }W_{s}\,\right) ^{2}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \leq C\cdot \left( u-\frac{i-1}{n}\right) \end{equation*}% and likewise \begin{equation*} \mathrm{E}\left[ \left( \,\int_{(i-1)/n}^{u}\,v_{s-}^{\ast }\,\mathrm{d}% V_{s}\,\right) ^{2}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \leq C\cdot \left( u-\frac{i-1}{n}\right) \vspace{1mm} \end{equation*}% for all $i,n\geq 1$. Likewise for the Poisson part we have \begin{eqnarray*} &&\displaystyle\mathrm{E}\left[ \left( \,\int_{(i-1)/n}^{u}\int_{E}\phi (s-,x)\,(\mu -\nu )(\mathrm{d}s\,\mathrm{d}x)\,\right) ^{2}\,|\,\mathcal{F}_{% \frac{i-1}{n}}\right] \\ &\leq &\displaystyle C\int_{(i-1)/n}^{u}\int_{E}\mathrm{E}[\phi ^{2}(s,x)\,|\,\mathcal{F}_{\frac{i-1}{n}}]\,F(\mathrm{d}x)\,\mathrm{d}s% \vspace{1mm} \end{eqnarray*}% yielding a similar bound. Putting all this together we have for all $i,n\geq 1$ \begin{eqnarray*} \mathrm{E}[\,(\eta ^{\prime }(2,2)_{i}^{n})^{2}\,|\,\mathcal{F}_{\frac{i-1}{n% }}] &\leq &C\,\int_{(i-1)/n}^{i/n}(u-(i-1)/n)\,\mathrm{d}u \\ &\leq &C/n^{2}. \end{eqnarray*}% Thus as $n\rightarrow \infty $ so \begin{equation*} \sum_{i=1}^{[nt]}\mathrm{E}[\,(\eta ^{\prime }(2,2)_{i}^{n})^{2}]\rightarrow 0.\vspace{1mm} \end{equation*}% and since \begin{equation*} \mathrm{E}\left[ \eta ^{\prime }(2,2)_{i}^{n}\,|\,\mathcal{F}_{\frac{i-1}{n}}% \right] =0\ \ \ \ \text{for all}\quad i,n\geq 1\vspace{1mm} \end{equation*}% we deduce from Doob's inequality that \begin{equation*} \,\sum_{i=1}^{[nt]}\eta ^{\prime }(2,2)_{i}^{n}\,\overset{P}{\rightarrow }0 \end{equation*}% proving\ altogether \begin{equation*} \,\sum_{i=1}^{[nt]}\eta ^{\prime }(2)_{i}^{n}\,\overset{P}{\rightarrow }0.% \vspace{1mm} \end{equation*}% Applying once more (H2a) and (\ref{eqn 8}) we have for every $\epsilon >0$ and every $i,n$ that \begin{eqnarray*} |\eta ^{\prime \prime }(2)_{i}^{n}| &\leq &\sqrt{n}\int_{(i-1)/n}^{i/n}\,% \Psi \left( \left\vert \sigma _{u}-\sigma _{\frac{i-1}{n}}\right\vert \right) \cdot \left\vert \sigma _{u}-\sigma _{\frac{i-1}{n}}\right\vert \,% \mathrm{d}u \\ &\leq &\sqrt{n}\,\Psi (\epsilon )\int_{(i-1)/n}^{i/n}\,\left\vert \sigma _{u}-\sigma _{\frac{i-1}{n}}\right\vert \,\mathrm{d}u+\sqrt{n}\,\Psi (2\sqrt{% b})/\epsilon \int_{(i-1)/n}^{i/n}\,\left\vert \sigma _{u}-\sigma _{\frac{i-1% }{n}}\right\vert ^{2}\,\mathrm{d}u.\vspace{1mm} \end{eqnarray*}% Thus from (\ref{8}) and its consequence \begin{equation*} \mathrm{E}\,\left[ \left\vert \sigma _{u}-\sigma _{\frac{i-1}{n}}\right\vert \,\right] \leq C/\sqrt{n} \end{equation*}% we get \begin{equation*} \sum_{i=1}^{[nt]}\mathrm{E}[\,|\eta ^{\prime \prime }(2)_{i}^{n}|\,]\leq Ct\,\Psi (\epsilon )+\frac{C\,\Psi (b)}{\sqrt{n}\,\epsilon } \end{equation*}% for all $n$ and all $\epsilon $. Letting here first $n\rightarrow \infty $ and then $\epsilon \rightarrow 0$ we may conclude that as $n\rightarrow \infty $ \begin{equation*} \sum_{i=1}^{[nt]}\mathrm{E}[\,|\eta ^{\prime \prime }(2)_{i}^{n}|\,]\rightarrow 0\ \end{equation*}% implying the convergence \begin{equation*} \,\sum_{i=1}^{[nt]}\eta (2)_{i}^{n}\,\overset{P}{\rightarrow }0.\vspace{1mm} \end{equation*}% Thus ending the proof of (\ref{13b}). \noindent $\square $ \subsection{Proof of $\protect\eta (1)^{n}\protect\overset{P}{\rightarrow }0$ \label{subsection:proof of 13a}} Recall we are to show that \textbf{\ } \begin{equation} \eta (1)^{n}=\,\sum_{i=1}^{[nt]}\eta (1)_{i}^{n}\,\overset{P}{\rightarrow }0. \end{equation}% Let still $t>0$ be fixed. Recall that \begin{eqnarray*} \eta (1)_{i}^{n} &=&\frac{1}{\sqrt{n}}\,\left\{ \mathrm{E}\left[ \,g(\sqrt{n}% \,\triangle _{i}^{n}Y)\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,-\rho _{\sigma _{\frac{i-1}{n}}}(g)\right\} \\ &=&\frac{1}{\sqrt{n}}\,\mathrm{E}\,\left[ g(\sqrt{n}\,\triangle _{i}^{n}Y)-g(\beta _{i}^{n})\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] .\vspace{% 1mm} \end{eqnarray*}% Introduce the notation\thinspace (recall the assumption (K)) \begin{equation*} A_{i}^{n}=\{\,|\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|>\,d(\beta _{i}^{n},B)/2\,\}.\vspace{1mm} \end{equation*}% Since $B$ is a Lebesgue null set and $\beta _{i}^{n}$ is absolutely continuous, $g^{\prime }(\beta _{i}^{n})$ is defined $a.s.$ and, by assumption, $g$ is differentiable on the interval joining $\triangle _{i}^{n}Y(\omega )$ and $\beta _{i}^{n}(\omega )$ for all $\omega \in A_{i}^{n\,c}$. Thus, using the Mean Value Theorem, we may for all $i,n\geq 1$ write \begin{eqnarray*} &&g(\sqrt{n}\,\triangle _{i}^{n}Y)-g(\beta _{i}^{n}) \\ &=&\left\{ g(\sqrt{n}\,\triangle _{i}^{n}Y)-g(\beta _{i}^{n})\right\} \cdot \mathbf{1}_{A_{i}^{n}} \\ &&+g^{\prime }(\beta _{i}^{n})\cdot (\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})\cdot \mathbf{1}_{A_{i}^{n\,c}} \\ &&+\left\{ g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})\right\} \cdot (\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})\cdot \mathbf{1}% _{A_{i}^{n\,c}} \\ &=&\sqrt{n}\,\left\{ \delta (1)_{i}^{n}+\delta (2)_{i}^{n}+\delta (3)_{i}^{n}\right\} ,\vspace{1mm} \end{eqnarray*}% where $\alpha _{i}^{n}$ are random points lying in between $\sqrt{n}% \,\triangle _{i}^{n}Y$ and $\beta _{i}^{n}$, i.e. \begin{equation*} \sqrt{n}\,\triangle _{i}^{n}Y\wedge \beta _{i}^{n}\leq \alpha _{i}^{n}\leq \sqrt{n}\,\triangle _{i}^{n}Y\vee \beta _{i}^{n}, \end{equation*}% and% \begin{equation*} \begin{array}{lll} \delta (1)_{i}^{n} & = & \left[ \,\left\{ g(\sqrt{n}\,\triangle _{i}^{n}Y)-g(\beta _{i}^{n})\right\} -g^{\prime }(\beta _{i}^{n})\cdot (% \sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})\,\right] \cdot \mathbf{1}% _{A_{i}^{n}}/\sqrt{n} \\ \delta (2)_{i}^{n} & = & \left\{ g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})\right\} \cdot (\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})\cdot \mathbf{1}_{A_{i}^{n\,c}}/\sqrt{n} \\ \delta (3)_{i}^{n} & = & g^{\prime }(\beta _{i}^{n})\cdot (\sqrt{n}% \,\triangle _{i}^{n}Y-\beta _{i}^{n})/\sqrt{n}.% \end{array}% \end{equation*}% Thus it suffices to prove \begin{equation*} \,\sum_{i=1}^{[nt]}\mathrm{E}\,\left[ \delta (k)_{i}^{n}\,|\,\mathcal{F}_{% \frac{i-1}{n}}\right] \,\,\overset{P}{\rightarrow }0,\ \ \ k=1,2,3. \end{equation*} Consider the case $k=1$. Using (K) and the fact that $\beta _{i}^{n}$ is absolutely continuous we have a.s.% \begin{eqnarray*} &&|g(\sqrt{n}\,\triangle _{i}^{n}Y)-g(\beta _{i}^{n})| \\ &\leq &M(1+|\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|^{p}+|\beta _{i}^{n}|^{p})\cdot |\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}| \\ &\leq &(2^{p}+1)M(1+|\sqrt{n}\,\triangle _{i}^{n}Y|^{p}+|\beta _{i}^{n}|^{p})\cdot |\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|, \end{eqnarray*}% and \begin{equation*} |\,g^{\prime }(\beta _{i}^{n})\cdot (\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})\,|\leq M(1+|\beta _{i}^{n}|^{p})\cdot |\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|.\vspace{1mm} \end{equation*}% By Cauchy-Schwarz's inequality $\mathrm{E}[\,|\delta (1)_{i}^{n}|\,]$ is therefore for all $i,n\geq 1$ less than \begin{equation*} C\cdot \mathrm{E}[\,1+|\sqrt{n}\,\triangle _{i}^{n}Y|^{3p}+|\beta _{i}^{n}|^{3p}]^{1/3}\cdot \mathrm{E}[\,(\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})^{2}/n\,]^{1/2}\cdot P(A_{i}^{n})^{1/6}\vspace{1mm} \end{equation*}% implying for fixed $t,$ by means of (\ref{2}), that \begin{eqnarray*} \mathrm{E}[\left[ \sum_{i=1}^{[nt]}|\,\delta (1)_{i}^{n}|\right] \, &\leq &C\cdot \sup_{i\geq 1}P(A_{i}^{n})^{1/6}\,\sum_{i=1}^{[nt]}\mathrm{E}% [\,(\triangle _{i}^{n}Y-\beta _{i}^{n})^{2}/n\,]^{1/2}\vspace{1mm} \\ &\leq &C\cdot \sup_{i\geq 1}P(A_{i}^{n})^{1/6}\,\sum_{i=1}^{[nt]}1/n \\ &\leq &Ct\cdot \sup_{i\geq 1}P(A_{i}^{n})^{1/6}.\vspace{1mm} \end{eqnarray*}% For all $i,n\geq 1$ we have for every $\epsilon >0$ \begin{eqnarray*} P(A_{i}^{n}) &\leq &P(A_{i}^{n}\cap \{d(\beta _{i}^{n},B)\leq \epsilon \})+P(A_{i}^{n}\cap \{d(\beta _{i}^{n},B)>\epsilon \})\vspace{1mm} \\ &\leq &P(d(\beta _{i}^{n},B)\leq \epsilon )+P(|\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|>\epsilon /2)\vspace{1mm} \\ &\leq &P(d(\beta _{i}^{n},B)\leq \epsilon )+\frac{4}{\epsilon ^{2}}\cdot \mathrm{E}[\,(\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n})^{2}]\vspace{1mm} \\ &\leq &P(d(\beta _{i}^{n},B)\leq \epsilon )+\frac{C}{n\,\epsilon ^{2}}. \end{eqnarray*}% But (H2a) implies that the densities of $\beta _{i}^{n}$ are pointwise dominated by a Lebesgue integrable function $h_{a,b}$ providing, for all $% i,n\geq 1$, the estimate \begin{eqnarray} P(A_{i}^{n}) &\leq &\int_{\{x\,|\,d(x,B)\leq \epsilon \}}h_{a,b}\,\mathrm{d}% \lambda _{1}+\frac{C}{n\,\epsilon ^{2}} \label{eqn 10} \\ &=&\alpha _{\epsilon }+\frac{C}{n\,\epsilon ^{2}}.\vspace{1mm} \notag \end{eqnarray}% Observe $\lim_{\epsilon \rightarrow 0}\alpha _{\epsilon }=0$. Taking now in (% \ref{eqn 10}) $\sup $ over $i$ and then letting first $n\rightarrow \infty $ and then $\epsilon \downarrow 0$ we get \begin{equation*} \lim_{n}\,\sup_{i\geq 1}\,P(A_{i}^{n})=0 \end{equation*}% proving that \begin{equation*} \mathrm{E}\left[ \,\sum_{i=1}^{[nt]}|\,\delta (1)_{i}^{n}|\right] \,\rightarrow 0 \end{equation*}% and\ thus \begin{equation*} \,\sum_{i=1}^{[nt]}\mathrm{E}\left[ \,\delta (1)_{i}^{n}\,|\,\mathcal{F}_{% \frac{i-1}{n}}\right] \,\overset{P}{\rightarrow }0.\vspace{1mm} \end{equation*} Consider next the case $k=2$. As assumed in (K), $g$ is continuously differentiable outside of $B$. Thus for each $A>1$ and $\epsilon >0$ there exists a function $G_{A,\,\epsilon }:(0,1)\rightarrow \mathbf{R}_{+}$ such that for given $0<\epsilon ^{\prime }<\epsilon /2$ \begin{equation*} \left\vert g^{\prime }(x+y)-g^{\prime }(x)\right\vert \leq G_{A,\,\epsilon }(\epsilon ^{\prime })\ \ \text{for all}\ |x|\leq A,\ |y|\leq \epsilon ^{\prime }<\epsilon <d(x,B).\vspace{1mm} \end{equation*}% Observe that $\lim_{\epsilon ^{\prime }\downarrow 0}G_{A,\,\epsilon }(\epsilon ^{\prime })=0$ for all $A$ and $\epsilon $.\vspace{1mm} Fix $A>1$ and $\epsilon \in (0,1)$. For all $i,n\geq 1$ we have \begin{eqnarray*} &&|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \mathbf{1}% _{A_{i}^{n\,c}} \\ &=&\displaystyle|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \mathbf{1}_{A_{i}^{n\,c}}\,(\mathbf{1}_{\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|>A\}}+\mathbf{1}_{\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}})\vspace{1mm} \\ &\leq &\displaystyle|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \frac{|\alpha _{i}^{n}|+|\beta _{i}^{n}|}{A}+|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \mathbf{1}% _{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}} \\ &\leq &\displaystyle\frac{C}{A}\cdot (1+|\alpha _{i}^{n}|^{p}+|\beta _{i}^{n}|^{p})^{2}+|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \mathbf{1}_{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}} \\ &\leq &\displaystyle\frac{C}{A}\cdot (1+|\sqrt{n}\,\triangle _{i}^{n}Y|^{2p}+|\beta _{i}^{n}|^{2p})+|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \mathbf{1}_{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}}.\vspace{1mm} \end{eqnarray*}% Now writing \begin{eqnarray*} 1 &=&\mathbf{1}_{\{d(\beta _{i}^{n},B)\leq \epsilon \}}+\mathbf{1}% _{\{d(\beta _{i}^{n},B)>\epsilon \}}\vspace{1mm} \\ &=&\mathbf{1}_{\{d(\beta _{i}^{n},B)\leq \epsilon \}} \\ &&+\mathbf{1}_{\{d(\beta _{i}^{n},B)>\epsilon \}\,\cap \,\{|\alpha _{i}^{n}-\beta _{i}^{n}|\leq \epsilon ^{\prime }\}} \\ &&+\mathbf{1}_{\{d(\beta _{i}^{n},B)>\epsilon \}\,\cap \,\{|\alpha _{i}^{n}-\beta _{i}^{n}|>\epsilon ^{\prime }\}} \end{eqnarray*}% for all $0<\epsilon ^{\prime }<\epsilon /2$ we have \begin{eqnarray*} \mathbf{1}_{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}} &\leq &\mathbf{1}_{\{d(\beta _{i}^{n},B)\leq \epsilon \}\,\cap \,A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}} \\ &&+\mathbf{1}_{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}\,\cap \,\{d(\beta _{i}^{n},B)>\epsilon \}\,\cap \,\{|\alpha _{i}^{n}-\beta _{i}^{n}|\leq \epsilon ^{\prime }\}} \\ &&+\mathbf{1}_{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}\,\cap \,\{d(\beta _{i}^{n},B)>\epsilon \}}\cdot \frac{|\alpha _{i}^{n}-\beta _{i}^{n}|}{\epsilon ^{\prime }}. \end{eqnarray*}% Combining this with the fact that \begin{eqnarray*} |g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})| &\leq &C(1+|\alpha _{i}^{n}|^{p}+|\beta _{i}^{n}|^{p}) \\ &\leq &CA^{p} \end{eqnarray*}% on $A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}$ we obtain that \begin{eqnarray*} &&|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot \mathbf{1}% _{A_{i}^{n\,c}\,\cap \,\{|\alpha _{i}^{n}|+|\beta _{i}^{n}|\leq A\}} \\ &\leq &CA^{p}\cdot \left( \,\mathbf{1}_{\{d(\beta _{i}^{n},B)\leq \epsilon \}}+\frac{|\alpha _{i}^{n}-\beta _{i}^{n}|}{\epsilon ^{\prime }}\right) \,+G_{A,\,\epsilon }(\epsilon ^{\prime })\vspace{1mm} \\ &\leq &CA^{p}\cdot (\,\mathbf{1}_{\{d(\beta _{i}^{n},B)\leq \epsilon \}}+% \frac{|\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|}{\epsilon ^{\prime }}% \,)+G_{A,\,\epsilon }(\epsilon ^{\prime }).\vspace{1mm} \end{eqnarray*} Putting this together means that \begin{eqnarray*} \sqrt{n}\,|\delta (2)_{i}^{n}| &=&|g^{\prime }(\alpha _{i}^{n})-g^{\prime }(\beta _{i}^{n})|\cdot |\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|\cdot \mathbf{1}_{A_{i}^{n\,c}} \\ &\leq &\left\{ \frac{C}{A}\cdot (1+|\sqrt{n}\,\triangle _{i}^{n}Y|^{2p}+|\beta _{i}^{n}|^{2p})+G_{A,\,\epsilon }(\epsilon ^{\prime })\right\} \cdot |\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|\vspace{1mm} \\ &&+\,CA^{p}\cdot \left( \mathbf{1}_{\{d(\beta _{i}^{n},B)\leq \epsilon \}}\cdot |\sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}|+\frac{|\sqrt{n}% \,\triangle _{i}^{n}Y-\beta _{i}^{n}|^{2}}{\epsilon ^{\prime }}\right) .% \vspace{1mm} \end{eqnarray*}% Exploiting here the inequalities (\ref{2}) and (\ref{3}) we obtain, for all $% A>1$ and $0<2\epsilon ^{\prime }<\epsilon <1$ and all $i,n\geq 1$, using H% \"{o}lder's inequality, the following estimate \begin{equation*} \mathrm{E}[\,|\delta (2)_{i}^{n}|\,]\leq C\left( \frac{1}{A\,n}+\frac{% G_{A,\,\epsilon }(\epsilon ^{\prime })}{n}+\frac{A^{p}\,\sqrt{\alpha _{\epsilon }}}{n}+\frac{A^{p}}{\epsilon ^{\prime }\,n^{3/2}}\right) \vspace{% 1mm} \end{equation*}% implying for all $n\geq 1$ and $t\geq 0$ that \begin{equation*} \sum_{i=1}^{[nt]}\mathrm{E}[\,|\delta (2)_{i}^{n}|\,]\leq Ct\left( \frac{1}{A% }+G_{A,\,\epsilon }(\epsilon ^{\prime })+A^{p}\,\sqrt{\alpha _{\epsilon }}+% \frac{A^{p}}{\epsilon ^{\prime }\,n^{1/2}}\right) .\vspace{1mm} \end{equation*}% Choosing in this estimate first $A$ sufficiently big, then $\epsilon $ small\thinspace (recall that $\lim_{\epsilon \rightarrow 0}\alpha _{\epsilon }=0$\thinspace ) and finally $\epsilon ^{\prime }$ small, exploiting that $% \lim_{\epsilon ^{\prime }\downarrow 0}G_{A,\,\epsilon }(\epsilon ^{\prime })=0$ for all $A$ and $\epsilon $, we may conclude that \begin{equation*} \lim_{n}\,\sum_{i=1}^{[nt]}\mathrm{E}\left[ \,|\delta (2)_{i}^{n}|\,\right] =0 \end{equation*}% and thus \begin{equation*} \sum_{i=1}^{[nt]}\mathrm{E}\left[ \,\delta (2)_{i}^{n}\,|\,\mathcal{F}_{% \frac{i-1}{n}}\right] \,\overset{P}{\rightarrow }0. \end{equation*} So what remains to be proved is the convergence \begin{equation*} \,\sum_{i=1}^{[nt]}\mathrm{E}\,\left[ \delta (3)_{i}^{n}\,|\,\mathcal{F}_{% \frac{i-1}{n}}\right] \,\,\overset{P}{\rightarrow }0. \end{equation*}% As introduced in (\ref{17}) \begin{equation*} \sqrt{n}\,\triangle _{i}^{n}Y-\beta _{i}^{n}=\sum_{j=1}^{5}\xi (j)_{i}^{n}=\psi (1)_{i}^{n}+\psi (2)_{i}^{n} \end{equation*}% for all $i,n\geq 1$ where \begin{equation*} \psi (1)_{i}^{n}=\xi (1)_{i}^{n}+\xi (3)_{i}^{n}+\xi (4)_{i}^{n}, \end{equation*}% \begin{equation*} \psi (2)_{i}^{n}=\xi (2)_{i}^{n}+\xi (5)_{i}^{n}, \end{equation*}% and as \begin{equation*} \delta (3)_{i}^{n}=g^{\prime }(\beta _{i}^{n})\cdot (\psi (1)_{i}^{n}+\psi (2)_{i}^{n})/\sqrt{n}\vspace{1mm} \end{equation*}% it suffices to prove \begin{equation*} \left( \,\sum_{i=1}^{[nt]}\mathrm{E}\left[ \,g^{\prime }(\beta _{i}^{n})\cdot \psi (k)_{i}^{n}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,\,/% \sqrt{n}\,\right) \overset{P}{\rightarrow }0,\ \ \ k=1,2.\vspace{1mm} \end{equation*} The case $k=1$ is handled by proving \begin{equation} \frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\mathrm{E}[\,|g^{\prime }(\beta _{i}^{n})\cdot \xi (j)_{i}^{n}|\,]\rightarrow 0,\ \ \ j=1,3,4.\vspace{1mm} \label{eqn 11} \end{equation}% Using Jensen's inequality it is easily seen that for $j=1,3,4$ \begin{equation*} \frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\mathrm{E}[\,|g^{\prime }(\beta _{i}^{n})\cdot \xi (j)_{i}^{n}|\,]\leq C\,t\cdot \sqrt{\frac{1}{n}% \,\sum_{i=1}^{[nt]}\mathrm{E}[\,g^{\prime }(\beta _{i}^{n})^{2}]}\,\cdot \,% \sqrt{\sum_{i=1}^{[nt]}\mathrm{E}[\,(\xi (j)_{i}^{n})^{2}]} \end{equation*}% and so using (\ref{12}) \begin{equation*} \frac{1}{\sqrt{n}}\,\sum_{i=1}^{[nt]}\mathrm{E}[\,|g^{\prime }(\beta _{i}^{n})\cdot \xi (j)_{i}^{n}|\,]\leq C\,t\cdot \,\sqrt{\sum_{i=1}^{[nt]}% \mathrm{E}[\,(\xi (j)_{i}^{n})^{2}]} \end{equation*}% since almost surely \begin{equation*} |g^{\prime }(\beta _{i}^{n})|\leq C\,(1+|\beta _{i}^{n}|^{p}) \end{equation*}% for all $i,n\geq 1$. From here, (\ref{eqn 11}) is an immediate consequence of Lemmas \ref{lemma 1st}-\ref{lemma 5th}.\vspace{1mm} The remaining case $k=2$ is different. The definition of $\psi (2)_{i}^{n}$ implies, using basic stochastic calculus, that $\psi (2)_{i}^{n}/\sqrt{n}$, for all $i,n\geq 1$, may be written as \begin{eqnarray*} &&\int_{(i-1)/n}^{i/n}\left\{ \sigma _{\frac{i-1}{n}}^{\prime }\,\left( W_{u}-W_{\frac{i-1}{n}}\right) +M(n,i)_{u}\right\} \,\mathrm{d}W_{u} \\ &=&\sigma _{\frac{i-1}{n}}^{\prime }\,\int_{(i-1)/n}^{i/n}\left( W_{u}-W_{% \frac{i-1}{n}}\right) \,\mathrm{d}W_{u} \\ &&+\triangle _{i}^{n}M(n,i)\cdot \triangle _{i}^{n}W \\ &&+\int_{(i-1)/n}^{i/n}\left( W_{u}-W_{\frac{i-1}{n}}\right) \,\mathrm{d}% M(n,i)_{u}, \end{eqnarray*}% where $(M(n,i)_{t})$ is the martingale defined by $M(n,i)_{t}\equiv 0$ for $% t\leq (i-1)/n$ and \begin{equation*} M(n,i)_{t}=v_{\frac{i-1}{n}}^{\ast }\,\left( V_{t}-V_{\frac{i-1}{n}}\right) +\int_{(i-1)/n}^{t}\int_{E_{n}}\phi \left( \frac{i-1}{n},x\right) (\mu -\nu )(\mathrm{d}s\,\mathrm{d}x)\vspace{1mm} \end{equation*}% otherwise. Thus for fixed $i,n\geq 1$ \begin{equation*} \mathrm{E}\left[ \,g^{\prime }(\beta _{i}^{n})\cdot \psi (2)_{i}^{n}\,|\,% \mathcal{F}_{\frac{i-1}{n}}\right] \,\,/\sqrt{n} \end{equation*}% is a linear combination of the following three terms \begin{equation*} \mathrm{E}\left[ g^{\prime }(\beta _{i}^{n})\cdot \sigma _{\frac{i-1}{n}% }^{\prime }\,\int_{(i-1)/n}^{i/n}\left( W_{u}-W_{\frac{i-1}{n}}\right) \,% \mathrm{d}W_{u}\,|\,\mathcal{F}_{\frac{i-1}{n}}\right] \,, \end{equation*}% \begin{equation*} \mathrm{E}\left[ g^{\prime }(\beta _{i}^{n})\cdot \triangle _{i}^{n}M(n,i)\cdot \triangle _{i}^{n}W\,|\,\mathcal{F}_{\frac{i-1}{n}}% \right] \, \end{equation*}% and% \begin{equation*} \mathrm{E}[\,g^{\prime }(\beta _{i}^{n})\cdot \int_{(i-1)/n}^{i/n}W_{u}\,% \mathrm{d}M(n,i)_{u}\,|\,\mathcal{F}_{\frac{i-1}{n}}\,]. \end{equation*}% But these three terms are all equal to $0$ as seen by the following arguments.\vspace{1mm} The conditional distribution of \begin{equation*} \left( W_{t}-W_{\frac{i-1}{n}}\right) _{t\geq \frac{i-1}{n}}|\mathcal{F}_{% \frac{i-1}{n}} \end{equation*}% is clearly not affected by a change of sign. Thus since $g$ being assumed even and $g^{\prime }$ therefore odd we have \begin{equation*} \mathrm{E}\left[ \,g^{\prime }(\beta _{i}^{n})\,\int_{(i-1)/n}^{i/n}\left( W_{u}-W_{\frac{i-1}{n}}\right) \,\mathrm{d}W_{u}\,|\,\mathcal{F}_{\frac{i-1}{% n}}\,\right] =0 \end{equation*}% implying the vanishing of the first term. \vspace{1mm} Secondly, by assumption, $\left( W_{t}-W_{\frac{i-1}{n}}\right) _{t\geq \frac{i-1}{n}}$ and $(M(n,i)_{t})_{t\geq \frac{i-1}{n}}$ are independent given $\mathcal{F}_{\frac{i-1}{n}}$. Therefore, denoting by $\mathcal{F}% _{i,n}^{\,0}$ the $\sigma $-field generated by \begin{equation*} \left( W_{t}-W_{\frac{i-1}{n}}\right) _{\frac{i-1}{n}\leq t\leq i/n}\ \ \ \text{and}\ \ \ \mathcal{F}_{\frac{i-1}{n}}, \end{equation*}% the martingale property of $(M(n,i)_{t})$ ensures that \begin{equation*} \mathrm{E}[\,g^{\prime }(\beta _{i}^{n})\cdot \triangle _{i}^{n}M(n,i)\cdot \triangle _{i}^{n}W\,|\,\mathcal{F}_{i,n}^{\,0}\,]=0\ \end{equation*}% and% \begin{equation*} \mathrm{E}[\left[ g^{\prime }(\beta _{i}^{n})\cdot \int_{(i-1)/n}^{i/n}W_{u}\,\mathrm{d}M(n,i)_{u}\,|\,\mathcal{F}_{i,n}^{\,0}% \right] \,=0. \end{equation*}% Using this the vanishing of \begin{equation*} \mathrm{E}\left[ \,g^{\prime }(\beta _{i}^{n})\cdot \triangle _{i}^{n}M(n,i)\cdot \triangle _{i}^{n}W\,|\,\mathcal{F}_{\frac{i-1}{n}}% \right] \end{equation*}% and% \begin{equation*} \mathrm{E}\left[ \,g^{\prime }(\beta _{i}^{n})\cdot \int_{(i-1)/n}^{i/n}W_{u}\,\mathrm{d}M(n,i)_{u}\,|\,\mathcal{F}_{\frac{i-1}{n% }}\right] \,\vspace{1mm} \end{equation*}% is easily obtained by successive conditioning.\vspace{1mm} The proof of (\ref{13a}) is hereby completed. \noindent $\square $ \bibliographystyle{chicago}
{ "timestamp": "2005-03-30T16:28:41", "yymm": "0503", "arxiv_id": "math/0503711", "language": "en", "url": "https://arxiv.org/abs/math/0503711" }
\section*{Introduction} In recent years, the study of singular Yang-Mills fields has been an extremely active area of research. Considering $\mbox{\rm SU}(2)$ instantons on four manifolds with codimension two singularities, it was found that these connections can admit non-trivial holonomy around arbitrarily small circles linking the embedded singular surface. An analytical theory for such instantons with holonomy singularity has been developed in \cite{KM,KM2}. Although we currently have an understanding of the moduli space for such singular instantons, the literature in this field has a conspicuous dearth of explicit examples. A singular solution on $T^{2}\times D^{2}$ is given in the appendix to \cite{KM}, although of greater interest are solutions on the standard model $S^{4}\setminus S^{2}$. The first example of a singular instanton was discovered by P.~Forg\'{a}cs, Z.~Horv\'{a}th, and L.~Palla. In their 1981 paper \cite{FHP1}, they describe a self-dual Yang-Mills field on $S^{4}\setminus S^{2}$ with the fractional Chern class $c_{2} = 3/2$. At first their result was not readily accepted, due to the resistance to the new idea of a fractional charge. They later published a second paper \cite{FHP2} defending their result, and since then over a decade of successful research into the field has eliminated any initial skepticism. Nevertheless, their construction itself remains poorly understood. The goal of this article is to elucidate and extend the work of Forg\'{a}cs {\em et al.\/}, writing their construction using simpler notation, explaining the motivation behind their formulae, and generalizing to obtain a family of singular instantons with varying holonomy parameter. To this work we shall contribute a mathematical perspective, exchanging indices and Pauli matrices for more invariant complex, quaternionic, and spinor notation, and offering geometric interpretations for the equations involved. Section 1 is devoted to the construction of instantons on $S^{4}$ employing the ansatz proposed by the physicists Corrigan, Fairlie, and Wilczek in 1976 and described in \cite{CF, JNR}. Starting with a positive real-valued function $\r$ on $\mathbb{R}^{4}$, known as the {\em super-potential}, we consider the Yang-Mills connection \begin{equation} \label{eq:ansatz} A = \sum_{\mu,\nu} i\bar{\sigma}_{\mu\nu}\,\partial^{\nu}\log\r\,dx^{\mu}. \end{equation} Here the anti-symmetric matrix $\bar{\sigma}_{\mu\nu}$ is defined as% \SSfootnote{We use the Greek indices $\mu,\nu$ when indexing over 4-space, while the Roman indices $i,j$ range over 1,2,3.} \begin{displaymath} \bar{\sigma}_{\mu\nu} = \left\{ \begin{array}{l} \bar{\sigma_{ij}} = \frac{1}{4i} \left[ \sigma_{i},\sigma_j \right] \\ \bar{\sigma_{i0}} = -\frac{1}{2}\sigma_{i} \end{array} \right. \end{displaymath} where the $\sigma_{i}$ are the standard Pauli matrices generating the Lie algebra $\mathfrak{su}(2)$. For such a connection, the self-duality equation $\ast F_{A} = F_{A}$ is equivalent to the condition $\Delta\rho = 0$. By reversing orientation, this construction can also be used to generate anti-self-dual connections from a harmonic super-potential. This harmonic function ansatz was used by 't Hooft to construct a class of instantons with $5n$ parameters, corresponding to the centers and scales of $n$ superimposed basic instantons. Since then, this ansatz has been shown to be the simplest case of a more general algebraic-geometric construction involving twistors discussed in \cite{AW}. More recently, both constructions have been eclipsed by the ADHM description of instantons given in \cite{ADHM}, which provides a complete construction for all ASD connections on $S^{4}$ up to gauge equivalence. In Section 1 we recast the harmonic function ansatz in terms of quaternionic notation. Not only does this greatly simplify the required calculations, but also it better exhibits the underlying structure. We then show how these connections arise naturally via conformal transformations. In Section 2 we introduce an $\mbox{\rm SO}(3)$-action on $S^{4}$. Taking advantage of the conformal equivalence $S^{4}\setminus S^{1} \cong \mathcal{H}^{2}\times S^{2}$, we show that the symmetric SD and ASD equations over $S^{4}$ are equivalent to the vortex and anti-vortex equations over hyperbolic space $\mathcal{H}^{2}$. This technique is known as dimensional reduction. The harmonic function ansatz for instantons then reduces to a similar ansatz for hyperbolic vortices, which we also derive using conformal transformations of hyperbolic space. After computing the vortex equivalents of the symmetric 't Hooft instantons, we use an equivariant version of the ADHM construction to provide a classification for all hyperbolic vortices. Examining gauge transformations, we obtain the surprising result that if two hyperbolic vortices constructed by the harmonic function ansatz are gauge equivalent, then they are both completely determined by the gauge transformation between them. We return to our primary task of constructing singular instantons in Section 3. Restricting our attention to $\mbox{\rm SO}(3)$-invariant connections on $S^{4}$, we can work instead with hyperbolic vortices. Using the unit disc model of $\mathcal{H}^{2}$, singular instantons correspond to vortices with a holonomy singularity at the origin. We then proceed to construct solutions on the cut disc using the harmonic function ansatz, patching them together with gauge transformations to form global solutions on the punctured disc. In \S\ref{fhp} we essentially rewrite the paper \cite{FHP1} in this context, and in the following section we construct our desired family of singular vortices. \section{The Harmonic Function ansatz} \label{instanton-ansatz} \subsection{Quaternionic Notation} \label{quaternionic-notation} For the duration of this section, we adopt the quaternionic notation as used in \cite{A}. Writing $x\in\H$ in the form $x = x^{0}+ix^{1}+jx^{2}+kx^{3}$, its conjugate is $\bar{x} = x^{0}-ix^{1}-jx^{2}-kx^{3}$, and the corresponding differentials are \begin{align*} dx = dx^{0} + i\,dx^{1} + j\,dx^{2} + k\,dx^{3} \qquad d\bar{x} = dx^{0} - i\,dx^{1} - j\,dx^{2} - k\,dx^{3}. \end{align*} By analogy with the complex case, we define the partial derivatives \begin{align*} \del{x} & = \frac{1}{2} \left( \del{x^{0}} - i\del{x^{1}} - j\del{x^{2}} - k\del{x^{3}} \right) \\ \del{\bar{x}} & = \frac{1}{2} \left( \del{x^{0}} + i\del{x^{1}} + j\del{x^{2}} + k\del{x^{3}} \right). \end{align*} In this notation the Laplacian takes the form \begin{displaymath} \Delta = -4\,\del{x}\del{\bar{x}} = -4\,\del{\bar{x}}\del{x}. \end{displaymath} As expected, the exterior derivative $d$ may be written as the sum of $\partial$ and $\bar{\partial}$ components, although there are now two distinct splittings \begin{displaymath} d = dx\,\del{x} + \del{\bar{x}}\,d\bar{x} = \del{x}\,dx + d\bar{x}\,\del{\bar{x}} \end{displaymath} due to the non-abelian nature of the operators involved. Expanding the 2-form $dx\wedge d\bar{x}$ in terms of coordinates as \begin{equation}\begin{split} dx\wedge d\bar{x} & = -2 \left[\, i\left(dx^{0}\wedge dx^{1} + dx^{2}\wedge dx^{3}\right)\,+\, j\left(dx^{0}\wedge dx^{2} + dx^{3}\wedge dx^{1}\right) \right. \nonumber \\ & \qquad \left.\mbox{\qquad}\,+\, k\left(dx^{0}\wedge dx^{3} + dx^{1}\wedge dx^{2}\right) \,\right], \label{eq:dxdxbar} \end{split}\end{equation} we see that $dx\wedge d\bar{x}$ is self-dual and likewise that $d\bar{x}\wedge dx$ is anti-self-dual. Rewriting the connection (\ref{eq:ansatz}) in terms of this new quaternionic notation, the harmonic function ansatz now takes the surprisingly familiar form \begin{theorem} \label{theorem-1} Given a positive real-valued super-potential $\r$ on $\mathbb{R}^{4}$, the Yang-Mills connection $A^{+}$ defined by \begin{equation} \label{eq:anti-self-dual} A^{+}= - \Im\left(\del{\bar{x}}\log \r\,d\bar{x} \right) = - \frac{1}{2} \left( \del{\bar{x}}\log \r\,d\bar{x} - dx\,\del{x}\log\r \right) \end{equation} is anti-self-dual and the connection $A^{-}$ defined by the conjugate expression \begin{equation} \label{eq:self-dual} A^{-} = - \Im \left(\del{x}\log \r\,dx \right) = - \frac{1}{2} \left( \del{x}\log \r\,dx - d\bar{x}\,\del{\bar{x}}\log\r \right) \end{equation} is self-dual if and only if the super-potential $\r$ is harmonic. \end{theorem} Before proceeding with the proof of this theorem the reader may want to verify that (\ref{eq:ansatz}) and (\ref{eq:self-dual}) both yield the same self-dual connection. Expanding (\ref{eq:self-dual}) using coordinates, we obtain the expression \begin{equation*}\begin{split} A^{-} & = \frac{1}{2} \left(\, (+i\partial_{1}+j\partial_{2}+k\partial_{3})\,dx^{0} + (-i\partial_{0}-k\partial_{2}+j\partial_{3})\,dx^{1} \right. \nonumber\\ & \qquad \left. \,\mbox{}+ (-j\partial_{0}+k\partial_{1}-i\partial_{3})\,dx^{2} + (-k\partial_{0}-j\partial_{1}+i\partial_{2})\,dx^{3} \,\right), \end{split}\end{equation*} writing $\partial_i$ as an abbreviation for $\partial_i\log\r$. \begin{proof}[Proof of Theorem 1] For the purposes of this proof, we restrict our attention to the potentially self-dual connection $A^{-}$ given in (\ref{eq:self-dual}), calling it $A$. Explicitly computing the two components of the curvature $F_{A} = dA+ A\wedge A$, we obtain \begin{align*} A\wedge A & = -\frac{1}{2} \left( \del{x}\log\r\,dx \wedge d\bar{x}\,\del{\bar{x}}\log\r + d\bar{x}\,\del{\bar{x}}\log\r \wedge \del{x}\log\r\,dx \right) \\ dA & = -\frac{1}{2} \left( \del{x}\,dx + d\bar{x}\,\del{\bar{x}} \right) \left( \del{x}\log \r\,dx - d\bar{x}\,\del{\bar{x}}\log\r \right) \\ & = -\frac{1}{2} \left( - \del{x}\,dx\wedge d\bar{x}\,\del{\bar{x}}\log\r + d\bar{x}\,\del{\bar{x}}\wedge\del{x}\log\r\,dx \right) . \end{align*} Recalling that the 2-forms $dx\wedge d\bar{x}$ and $d\bar{x}\wedge dx$ are self-dual and anti-self-dual respectively, the curvature of $A$ then splits as $F_{A} = F_{A}^{+} + F_{A}^{-}$ with \begin{align*} F_{A}^{+} & = \frac{1}{2} \left( \del{x}\,dx\wedge d\bar{x}\,\del{\bar{x}}\log\r\,-\, \del{x}\log\r\,dx \wedge d\bar{x}\,\del{\bar{x}}\log\r \right) \\ F_{A}^{-} & = -\frac{1}{2} \left( \del{\bar{x}}\del{x}\log\r\,+\, \del{\bar{x}}\log\r\,\del{x}\log\r \right) d\bar{x}\wedge dx. \end{align*} Noting the identity \begin{displaymath} \del{\bar{x}}\del{x}\log\r + \del{\bar{x}}\log\r\,\del{x}\log\r = \frac{1}{\r}\,\del{\bar{x}}\del{x}\r = -\frac{1}{4\r}\Delta\r, \end{displaymath} we see that the self-duality equation $F_{A}^{-} = 0$ is equivalent to the condition $\Delta\r = 0$ that the super-potential $\r$ be harmonic. \end{proof} We now calculate the curvature density $|F_{A}|^{2}$ of the self-dual connection (\ref{eq:self-dual}), from which we can construct the Yang-Mills functional $\|F_{A}\|^{2}$ and the Chern class $c_{2}(A)$. Using the decomposition $F_{A} = F_{A}^{+} + F_{A}^{-}$ given above, we first compute the anti-self-dual component $|F_{A}^{-}|^{2}$. From equation (\ref{eq:dxdxbar}) we observe that $(d\bar{x}\wedge dx)\wedge-(\overline{d\bar{x}\wedge dx}) = 24\,d\mu$, where $d\mu$ is the volume form, and we immediately obtain \begin{equation*} |F_{A}^{-}|^{2} = \frac{3}{8}\left(\frac{1}{\r}\Delta\r\right)^{2}, \end{equation*} which clearly vanishes if the super-potential $\r$ is harmonic. \newcommand{\log\r}{\log\r} \newcommand{\partial_{i}}{\partial_{i}} \renewcommand{\dj}{\partial_{j}} On the other hand, the self-dual component $|F_{A}^{-}|^{2}$ of the curvature density is significantly more difficult to compute. Again using the expansion (\ref{eq:dxdxbar}) for $dx\wedge d\bar{x}$, we have \begin{equation*}\begin{split} |F_{A}^{+}|^{2} & = 2 \left( \left| \del{x}\,i\,\del{\bar{x}}\log\r - \left(\del{x}\log\r\right) i\left(\del{\bar{x}}\log\r\right) \right|^{2}\right. \\ & \qquad \mbox{}+\left.\left| \del{x}\,j\,\del{\bar{x}}\log\r - \left(\del{x}\log\r\right) j\left(\del{\bar{x}}\log\r\right) \right|^{2}\right. \\ & \qquad \mbox{}+\left.\left| \del{x}\,k\,\del{\bar{x}}\log\r - \left(\del{x}\log\r\right) k\left(\del{\bar{x}}\log\r\right) \right|^{2}\right), \end{split}\end{equation*} which when fully expanded in terms of coordinates becomes \begin{equation*}\begin{split} |F_{A}^{+}|^{2} & = \frac{1}{8}\,\sum_{i,j}\left[ 4\,(\partial_{i}\dj\log\r)^{2} + 3\,(\partial_{i}\log\r)^{2}(\dj\log\r)^{2} \right. \\* & \mbox{\qquad} - \left. 8\,(\partial_{i}\dj\log\r)(\partial_{i}\log\r)(\dj\log\r) - (\partial_{i}^{2}\log\r)(\dj^{2}\log\r) \right. \\ & \rule{0in}{4ex}\mbox{\qquad} + \left. (\partial_{i}^{2}\log\r)(\dj\log\r)^{2} + (\partial_{i}\log\r)^{2}(\dj^{2}\log\r) \right]. \end{split}\end{equation*} If the super-potential $\r$ is harmonic, then we can take advantage of the identity $\sum_{i}\partial_{i}^{2}\log\r = -\sum_{i}(\partial_{i}\log\r)^{2}$ to simply this expression for $|F_{A}^{+}|^{2}$ to \begin{displaymath} |F_{A}^{+}|^{2} = \frac{1}{2}\,\sum_{i,j}\left[ (\partial_{i}\dj\log\r)^{2} - 2\,(\partial_{i}\dj\log\r)(\partial_{i}\log\r)(\dj\log\r) \right]. \end{displaymath} On the other hand, expanding the expression $\Delta\Delta\log\r$, we obtain \begin{equation*}\begin{split} \Delta\Delta\log\r & = \sum_{i,j}\dj^{2}\partial_{i}^{2}\log\r = -\sum_{i,j}\dj^{2}(\partial_{i}\log\r)^{2} \\ & = -2\,\sum_{i,j}\dj\left[(\partial_{i}\log\r)(\partial_{i}\dj\log\r)\right] \\ & = -2\,\sum_{i,j}\left[ (\partial_{i}\dj\log\r)^{2} + (\partial_{i}\log\r)(\partial_{i}\dj^{2}\log\r) \right] \\ & = -2\,\sum_{i,j}\left[ (\partial_{i}\dj\log\r)^{2} - (\partial_{i}\log\r)\partial_{i}(\dj\log\r)^{2} \right] \\ & = -2\,\sum_{i,j}\left[ (\partial_{i}\dj\log\r)^{2} - 2\,(\partial_{i}\log\r)(\dj\log\r)(\partial_{i}\dj\log\r) \right], \end{split}\end{equation*} again assuming that $\r$ is harmonic and using the identity $\sum_{i}\partial_{i}^{2}\log\r = -\sum_{i}(\partial_{i}\log\r)^{2}$ repeatedly. Hence if the super-potential $\r$ is harmonic, then the components of the curvature density $|F_{A}|^{2}$ for the self-dual connection (\ref{eq:self-dual}) are \begin{displaymath} |F_{A}^{+}|^{2} = -\frac{1}{4}\,\Delta\Delta\log\r, \qquad |F_{A}^{-}|^{2} = 0, \end{displaymath} and the Chern class $c_{2}(A)$ and $L^{2}$ norm $\|F_{A}\|^{2}$ are given by \begin{equation} \label{eq:c2} c_{2}(A) = \frac{1}{4\pi^{2}}\,\|F_{A}\|^{2} = -\frac{1}{16\pi^{2}} \int_{\mathbb{R}^{4}}\Delta\Delta\log\r\,d\mu. \end{equation} Here we have a factor of $4\pi^{2}$ instead of the customary $8\pi^{2}$ because the function $\xi\mapsto\mbox{Tr}(\xi^{2})$ on the Lie algebras $\mathfrak{su}(2)$ and $\mathfrak{so}(3)$ corresponds to the map $\xi\mapsto 2\xi^{2}$ in our quaternionic notation. Similarly, if we take the anti-self-dual connection (\ref{eq:anti-self-dual}) then the two components $|F_{A}^{+}|^{2}$ and $|F_{A}^{-}|^{2}$ are interchanged and the Chern class $c_{2}(A)$ switches sign. It is important to note that the scalar curvature density is gauge invariant. In other words, if two harmonic super-potentials $\r_{1}$ and $\r_{2}$ yield gauge equivalent connections via the ansatz of Theorem~\ref{theorem-1}, then they must satisfy the equation $\Delta\Delta\log\r_{1} = \Delta\Delta\log\r_{2}$. \subsection{The 't Hooft Construction} \label{tHooft} As an example of the harmonic function ansatz, we take for our super-potential the Green's functions of the Laplacian. Although these functions have $O(1/r^{2})$ poles, the corresponding singularities can be removed from the resulting connections by a gauge transformation. In the simplest case, consider the spherically symmetric harmonic function \begin{displaymath} \r = 1 + \frac{1}{|x|^{2}} = 1 + \frac{1}{x\bar{x}}, \end{displaymath} the sum of the Green's functions centered at the origin and infinity. Applying formula~(\ref{eq:anti-self-dual}), this super-potential generates the anti-self-dual connection \begin{equation} \label{eq:singular-gauge} A = \Im\left( \frac{\bar{x}^{-1}\,d\bar{x}}{1 + x\bar{x}} \right) = -\Im\left( \frac{dx\,x^{-1}}{1 + x\bar{x}} \right), \end{equation} which is simply the basic instanton with $c_{2} = 1$ expressed in the ``singular gauge''. Applying the gauge transformation $g = x^{-1}$, we can remove the $O(1/r)$ pole at the origin to obtain this instanton's customary form \begin{equation} \label{eq:standard-gauge} g(A) = \Im\left( - x^{-1}\,\frac{dx\,x^{-1}}{1+x\bar{x}}\,x - dx^{-1}\,x \right) = \Im\left( \frac{\bar{x}\,dx}{1+x\bar{x}} \right). \end{equation} Note that if we switch to coordinates around infinity by putting $x = y^{-1}$, then we simply interchange these two gauges (\ref{eq:singular-gauge}) and (\ref{eq:standard-gauge}). This connection therefore takes the same form about infinity as it does about the origin. More generally, we can modify the basic instanton~(\ref{eq:standard-gauge}) by applying a dilation and translation $x\mapsto\,\lambda^{-1}(x-a)$ with $\lambda>0$ real and $a\in\H$. The super-potential and associated connection then become \begin{equation}\label{eq:basic-instanton} \r = 1 + \frac{\lambda^{2}}{|x-a|^{2}},\qquad g_{a}(A) = \Im\left( \frac{ (\bar{x} - \bar{a})\,dx }{\lambda^{2} + |x-a|^{2}} \right). \end{equation} Here we have again used a gauge transformation $g_{a} = (x - a)^{-1}$ in order to remove the singularity at the point $x = a$. One of the interesting features of this ansatz is that it allows us to take the superposition of several such instantons simply by adding their super-potentials. For instance, the 't Hooft instantons with $c_{2} = k$ are constructed using the harmonic function \begin{equation}\label{eq:tHooft} \r = 1 + \frac{\lambda_{1}^{2}}{|x-a_{1}|^{2}} + \cdots + \frac{\lambda_{k}^{2}}{|x-a_{k}|^{2}}, \end{equation} combining $k$ basic instantons of the form (\ref{eq:basic-instanton}) with scales $\lambda_{1},\ldots,\lambda_{k}$ and distinct centers $a_{1},\ldots,a_{k}$. \subsection{Conformal Transformations} \label{conformal-instanton} In this section, we discuss a differential geometric interpretation of the harmonic function ansatz introduced in \S\ref{quaternionic-notation}. Treating the super-potential as a conformal transformation of flat Euclidean space, the connections (\ref{eq:anti-self-dual}) and (\ref{eq:self-dual}) arise naturally from the action of the Levi-Civita connection on the half-spin bundles. We can then express the curvatures of these two connections in terms of the decomposition of the Riemann curvature into its scalar, trace-free Ricci, and conformally invariant Weyl curvature components, thereby providing an alternative proof of Theorem \ref{theorem-1}. Starting with the flat Euclidean metric $g_{ij} = \delta_{ij}$ on $\mathbb{R}^{4}$, we consider the conformally equivalent metric $g' = \r^{2}g$, given a smooth, positive, real-valued super-potential $\r$. The condition that $\r$ be harmonic enters when calculating the scalar curvature of this new metric as in the following lemma. \begin{lemma} \label{lemma-scalar-curvature} The scalar curvature $R'$ of the conformally Euclidean metric $g'$ given by $g' = \r^{2}\delta_{ij}$ vanishes if and only if the super-potential $\r$ is harmonic. \end{lemma} \begin{proof} Using the expression for $R'$ computed in \cite[p. 125]{Au}, in dimension $n=4$ we have \begin{equation*}\begin{split} R'&=- \r^{-2} \left[ (n-1)\,\sum_{\nu}\partial_{\nu}^{2}\log\r^{2}\,+\, \frac{(n-1)(n-2)}{4}\, \sum_{\nu}\left(\partial_{\nu}\log\r^{2}\right)^{2} \right] \\ &=- 6 \r^{-2}\,\sum_{\nu} \left[ \partial_{\nu}^{2}\log\r + \left(\partial_{\nu}\log\r\right)^{2} \right] = 6 \r^{-3}\Delta\r.\rule{0in}{2.5ex} \end{split}\end{equation*} Hence $R' = 0$ if and only if $\Delta\r = 0$. \end{proof} Let $\{e_{0},\ldots,e_{3}\}$ be an orthonormal tangent frame for the original metric $g$. After applying the conformal transformation, the Levi-Civita connection for the metric $g'$ is given with respect to this frame by Christoffel's formula \begin{displaymath} \Gamma'^{j}_{ik} = \partial_{i}\log\r\:\delta^{j}_{k} + \partial_{k}\log\r\:\delta^{j}_{i} - \partial^{j}\log\r\:\delta_{ik}. \end{displaymath} In order to express this as an $\mathfrak{so}(4)$ connection, we must rescale the tangent frame so that it is again orthonormal with respect to the new metric $g'$. Switching to the frame $e'_{i} = \r^{-1}e_{i}$ introduces a factor of $-\partial_{i}\log\r\;\delta^{j}_{k}$ into the connection, cancelling the diagonal term and leaving us with an expression skew-symmetric in the indices $j$ and $k$. Taking the double cover $\mbox{\rm Spin}(4)$ of $\mbox{\rm SO}(4)$, we recall that the Lie algebra isomorphism $\mathfrak{so}(4) \cong \mathfrak{spin}(4)$ associates to a skew-symmetric matrix $a_{ij}$ the Clifford algebra element% \SSfootnote{The Clifford algebra $\mbox{Cl}(4)$ is the algebra generated by $\mathbb{R}^{4}$ subject to the relation $v\cdot w + w\cdot v = -2(v,w)$, and the Lie algebra $\mathfrak{spin}(4)$ is the subspace spanned by $\{e_{i}\cdot e_{j}\}_{i\neq j}$.} $-\frac{1}{4}\sum_{i,j} a_{ij}\,e_{i}\cdot e_{j}$ (see \cite{LM}). We may thus write the Levi-Civita connection in this $\mathfrak{spin}(4)$ notation as \begin{displaymath} A = \frac{1}{2} \sum_{i\neq j}\left( e'_{j}\,\partial^{j}\log\r \right) \cdot \left( e'_{i}\,dx^{i} \right). \end{displaymath} From the decomposition $\mbox{\rm Spin}(4) = \mbox{\rm Sp}(1) \times \mbox{\rm Sp}(1)$, we see that the complex 4-dimensional spin space splits as the direct sum $S = S^{+} \oplus S^{-}$ of two half-spin spaces, each of which is isomorphic to the quaternions $\H$. These spaces $S^{+}$ and $S^{-}$ are called the spaces of self-dual and anti-self-dual spinors respectively. The two half-spin representations $\gamma^{\pm}$ of the Lie algebra $\mathfrak{spin}(4)$ on $\H^{\pm}$ are then given by% \SSfootnote{Here the various signs are determined by the Clifford algebra relation $\gamma^{\pm}(v\cdot w + w\cdot v) = -2(v,w)$ and also by the convention that $\gamma^{+}(e'_{0}\cdot e'_{1} - e'_{2}\cdot e'_{3}) = 0$ and $\gamma^{-}(e'_{0}\cdot e'_{1} + e'_{2}\cdot e'_{3}) = 0$.} \begin{align*} \gamma^{+}:v\cdot w \mapsto -\gamma(v)\,\gamma^{\ast}(w) \qquad \gamma^{-}:v\cdot w \mapsto -\gamma^{\ast}(v)\,\gamma(w), \end{align*} where the Clifford action $\gamma(\cdot)$ is simply quaternion multiplication \begin{displaymath} \gamma(e'_{0}) = 1, \qquad \gamma(e'_{1}) = i, \qquad \gamma(e'_{2}) = j, \qquad \gamma(e'_{3}) = k, \end{displaymath} and $\gamma^{\ast}(\cdot)$ is its adjoint. Hence, the Levi-Civita connection for the conformally transformed metric $g' = \r^{2}g$ splits into the two $\sp(1)$ components \begin{equation*} A^{+}=-\Im\left(\del{\bar{x}}\log\r\,d\bar{x}\right) \qquad A^{-}=-\Im\left(\del{x}\log\r\,dx\right) \end{equation*} acting on the positive and negative half-spin spaces respectively. Note that these two connections agree with the connections $A^{+}$ and $A^{-}$ given in equations (\ref{eq:anti-self-dual}) and (\ref{eq:self-dual}). By definition, the Riemann curvature tensor $\mathcal{R}$ is an $\mathfrak{so}(4)$-valued 2-form. However, using the identification $\Lambda^{2} \cong \mathfrak{so}(4)$, we may view it as a self-adjoint linear map $\mathcal{R} : \Lambda^{2} \rightarrow\Lambda^{2}$ given in coordinates by \begin{displaymath} \mathcal{R}\left(dx^{i}\wedge dx^{j}\right) = \frac{1}{2} \sum_{k,l} R_{ijkl}\,dx^{k}\wedge dx^{l}. \end{displaymath} Relative to the familar decomposition $\Lambda^{2} = \Lambda^{2}_{+} \oplus \Lambda^2_{-}$ of the space of two-forms into its self-dual and anti-self-dual subspaces, the Riemann curvature can be written in the block matrix form \begin{displaymath} \mathcal{R} = \left( \begin{array}{c|c} \mathcal{W}^{+} - \frac{1}{12}R & R_{0} \\ \hline R_{0}^{\ast} & \mathcal{W}^{-} - \frac{1}{12}R \end{array} \right). \end{displaymath} Here $R$ denotes the scalar curvature multiplied by the identity matrix, while $R_{0} : \Lambda^{2}_{-}\rightarrow \Lambda^{2}_{+}$ is the trace-free Ricci curvature tensor, $R_{0}^{\ast} : \Lambda^{2}_{+}\rightarrow \Lambda^{2}_{-}$ is its adjoint, and $\mathcal{W} = \mathcal{W}^{+} + \mathcal{W}^{-}$ is the conformally invariant Weyl tensor. A standard reference for this material is \cite{AHS}. We now consider the Riemann curvature $\mathcal{R}'$ of the metric $g'$ discussed above. Since $g'$ is by definition conformally flat, we see that the Weyl tensor $\mathcal{W}'$ vanishes. We also recall from Lemma~\ref{lemma-scalar-curvature} that if our super-potential $\r$ is harmonic, then the scalar curvature $R'$ vanishes as well. All that remains is the trace-free Ricci tensor $R_{0}'$, and so the Riemann curvature is simply \begin{displaymath} \mathcal{R}' = \left( \begin{array}{c|c} 0 & R_{0}'\\ \hline R_{0}'^{\ast} & 0 \end{array} \right). \end{displaymath} Note that the splitting $\mathfrak{spin}(4) \cong \sp(1) \oplus \sp(1)$ which we used to construct the connections $A^{+}$ and $A^{-}$ is isomorphic to the decomposition $\Lambda^{2} = \Lambda^{2}_{+} \oplus \Lambda^{2}_{-}$. We can therefore read off the curvatures $F_{A^{+}}$ and $F_{A^{-}}$ of these connections directly from the block form of the Riemann curvature, giving us $$ F_{A^{+}}^{+} = 0 \qquad F_{A^{+}}^{-} = R_{0}' \qquad F_{A^{-}}^{+} = R_{0}'^{\ast} \qquad F_{A^{-}}^{-} = 0. $$ Hence the connections $A^{+}$ and $A^{-}$ are anti-self-dual and self-dual respectively as claimed in Theorem \ref{theorem-1}. \section{Hyperbolic Vortices} \subsection{Dimensional Reduction} \label{dimensional-reduction} In this section, we examine $\mbox{\rm SO}(3)$-invariant instantons, showing that the SD and ASD equations for Yang-Mills connections over $S^{4}$ with $\mbox{\rm SO}(3)$ symmetry are equivalent to the $\mbox{\rm U}(1)$ vortex equations over the hyperbolic plane $\mathcal{H}^{2}$. This is an example of dimensional reduction, whereby the Yang-Mills or (A)SD equations for a symmetric connection reduce to differential equations for a connection and Higgs fields (sections of the Lie algebra bundle) over a lower dimensional space. Viewing $S^{4}$ as the standard conformal compactification $\mathbb{R}^{4} \cup \{\infty\}$, we let $\mbox{\rm SO}(3)$ act via its fundamental representation on a three-dimensional subspace of $\mathbb{R}^{4}$. Expressing this using quaternionic notation, we see that an element $g\in \mbox{\rm Sp}(1)$ acts on $\H$ according to $g:x\mapsto g x g^{-1}$, fixing the real part of $x$ and acting by the adjoint representation on its imaginary part. Regarding $S^{4}$ as the quaternionic projective space $P(\H^{2})$ with homogeneous coordinates $(x:y)=(x\alpha:y\alpha)$, the embedding of $\H$ is simply the map $x\mapsto(1:x)$. The $\mbox{\rm Sp}(1)$-action given by \[g:(x:y)\mapsto(gx:gy) = (gxg^{-1}:gyg^{-1})\] then provides an extension of the above action on $\mathbb{R}^{4}$ to all of $S^{4}$. In order to discuss $\mbox{\rm SO}(3)$-invariant connections, we must lift this action on $S^{4}$ to an action on the Lie algebra bundle with fibres $\mathfrak{so}(3)$. There are two possible lifts: either $\mbox{\rm SO}(3)$ acts trivially on each fibre or it acts via the adjoint representation. For our purposes, we will consider this second, more interesting, action. Note that for any $g\in\mbox{\rm SO}(3)$, the adjoint action leaves fixed an $\mathbb{R} = \u(1)$ subalgebra. Again adopting quaternionic notation, any $x\in\H$ can be written in the form $x = t + r Q$, with $t,r$ real, $r \geq 0$, and Q pure imaginary with $Q^{2} = -1$. Note that $t,r$ coordinatize the upper half-plane, which we will later regard as hyperbolic space $\mathcal{H}^{2}$. If $A$ is an $\mbox{\rm Sp}(1)$-invariant connection, then its connection one-form satisfies $g A(t,r,Q) g^{-1} = A(t,r,gQg^{-1})$. The most general connection exhibiting this symmetry is of the form \begin{displaymath} A = \frac{1}{2}\,\bigl(Qa + \Phi_{1}\,dQ + \Phi_{2}\,Q\,dQ\bigr), \label{symmetric-connection} \end{displaymath} where $a = a_{t}\,dt + a_{r}\,dr$, and the $a_{t}, a_{r}, \Phi_{1}, \Phi_{2}$ are all real functions of $t,r$. The curvature $F_{A}$ of this connection $A$ is then \begin{align*} F_{A}&=\frac{1}{2} \left( Q\,da\,+\, \frac{1}{2} \left(\Phi_{1}^{2}+\Phi_{2}^{2}+2\Phi_{2} \right)dQ \wedge dQ\,+\right.\\* & \qquad\qquad\left.\rule{0in}{3ex} [ d\Phi_{1} - a ( \Phi_{2} + 1 ) ] \wedge dQ\,+\, ( d\Phi_{2} + a \Phi_{1} ) \wedge Q\,dQ \right). \end{align*} Putting $\Phi = \Phi_{1}+ i\,(\Phi_{2}+1)$ and writing $d_{a}\Phi = d\Phi + ia\Phi$, the curvature can be written much more simply as \begin{align*} F_{A}=\frac{1}{2} \left( Q\,da\,-\, \frac{1}{2}\,\left( 1 - |\Phi|^{2}\right)\,dQ \wedge dQ\,+ \Re(d_{a}\Phi) \wedge dQ \,+\, \Im(d_{a}\Phi) \wedge Q\,dQ \right). \end{align*} Note that multiplication by $Q$ here behaves like multiplication by $i$. From the above discussion, we see that an $\mbox{\rm SO}(3)$-invariant connection $A$ on $S^{4}$ gives rise in a natural way to a $\mbox{\rm U}(1)$ connection $ia$ and a complex scalar field $\Phi$ on the upper half-plane. The next step is to analyze the SD and ASD equations in terms of this dimensional reduction. To determine the action of the Hodge star operator, we consider the 2-form $dx\wedge d\bar{x}$ which we already know to be self-dual. In coordinates $t,r,Q$, we have \begin{align*} dx \wedge d\bar{x} &= (dt+Q\,dr+r\,dQ) \wedge (dt-Q\,dr-r\,dQ) \\* &= 2Q\,dt\wedge dr\,+\,r^{2}\,dQ\wedge dQ\,+\, 2r\,(dt\wedge dQ\,+\,dr\wedge Q\,dQ), \end{align*} and so the Hodge star operator acts according to \begin{align*} \ast\,Q\,dt\wedge dr & = \frac{r^{2}}{2}\,dQ\wedge dQ \\ \ast\,dt\wedge dQ & = dr\wedge Q\,dQ \\ \ast\,dr\wedge dQ & = -dt\wedge Q\,dQ. \end{align*} Furthermore, using the hyperbolic metric \begin{displaymath} h = \frac{1}{r^{2}}\,( dt^{2} + dr^{2} ) \end{displaymath} on the upper half-plane, the corresponding Hodge star operator $\ast_{h}$ satisfies $\ast_{h}dt \wedge dr = r^{2}$, $\ast_{h}dt = dr$, and $\ast_{h}dr = -dt$. Combining this with the usual Hodge star operator yields \begin{align*} \ast\,Q\,da = \frac{1}{2}\,(\ast_{h} da)\,dQ \wedge dQ \qquad \ast\,\Re\,(d_{a}\Phi) \wedge dQ = \ast_{h} \Re\,(d_{a}\Phi) \wedge Q\,dQ. \end{align*} Using complex notation with $z=t+ir$ and noting the identity $\ast_{h}dz = -i\,dz$, we observe that $2\,\bar{\partial}_{a}\Phi = d_{a}\Phi - i \ast_{h}d_{a}\Phi$. We therefore conclude that the $\mbox{\rm SO}(3)$-symmetric self-duality equation $F_{A} = \ast F_{A}$ on $S^{4}$ is equivalent to the following two equations on hyperbolic space $\mathcal{H}^{2}$: \begin{align} \bar{\partial}_{a} \Phi & = 0 \label{eq:vortex-a} \\ \ast_{h} i F_{a} & = 1 - |\Phi|^{2}, \label{eq:vortex-b} \end{align} where $F_{a} = i\,da$ is the curvature of the connection $ia$. These equations are known as the {\em vortex equations}. Similarly, the $\mbox{\rm SO}(3)$-symmetric anti-self-dual equation $F_{A} = -\ast F_{A}$ is equivalent to the {\em anti-vortex equations}: \begin{align} \partial_{a} \Phi & = 0 \label{eq:anti-vortex-a} \\ \ast_{h} i F_{a} & = |\Phi|^{2} - 1. \label{eq:anti-vortex-b} \end{align} The first equation in each pair is simply the condition that $\Phi$ be holomorphic (or anti-holomorphic) with respect to the holomorphic structure compatible with the connection $ia$. The second equation then expresses a form of duality between the connection and Higgs field. These vortex equations are discussed in great detail in \cite{JT}% \SSfootnote{After adjusting to the slightly different notation of \cite{JT}, using $a' = -a$ and $\Phi' = \bar{\Phi}$, the reader will find that equations (11.5a) and (11.5b) on p. 99 of \cite{JT} should be switched.}. Note that if we consider these vortex and anti-vortex equations over the plane $\mathbb{R}^{2}$ with the flat metric $h = \frac{1}{2}(dx^{2}+dy^{2})$, then we obtain the Euclidean vortex and anti-vortex equations in their customary form as given by (1.7) and (1.8) on p. 55 of \cite{JT}. We now compute the $L^{2}$ norm of the curvature $F_{A}$ using the standard metric $|\xi|^{2} = \xi\bar{\xi} = -\xi^{2}$ on the Lie algebra $\sp(1)$ of imaginary quaterions. The Yang-Mills action, or energy, of this $\mbox{\rm SO}(3)$-invariant connection is thus \begin{align*} \|F_{A}\|^{2} & = \int_{S^{4}} -\,F_{A}\wedge \ast F_{A} \\ & = \frac{1}{8}\int_{S^{4}} \Bigl( da\wedge \ast_{h} da \,+\, (1-|\Phi|^{2}) \wedge \ast_{h}(1-|\Phi|^{2}) \,+\, \\ & \qquad \qquad \rule{0in}{2ex} 2\,\Re d_{a}\Phi \wedge\ast_{h}\Re d_{a}\Phi \,+\,2\,\Im d_{a}\Phi \wedge\ast_{h}\Im d_{a}\Phi \Bigr) \wedge\,\left(\,dQ\wedge Q\,dQ\,\right). \end{align*} Note that the left factor of the integrand is independent of the variable $Q$. Since $Q$ parametrizes the unit 2-sphere with volume form $\frac{1}{2}(dQ\wedge Q\,dQ)$, we can integrate out a factor of $\int_{S^{2}}dQ\wedge Q\,dQ = 8\pi$, leaving an integral over hyperbolic space. The action then becomes \begin{displaymath} \|F_{A}\|^{2} = \pi \left( \|F_{a}\|^{2}_{h} \,+\, 2\,\|d_{a}\Phi\|^{2}_{h} \,+\, \|1-|\Phi|^{2}\|^{2}_{h} \right), \end{displaymath} which we recognize as the $\mbox{\rm U}(1)$ Yang-Mills-Higgs action on hyperbolic space, at least up to a constant. From this action, we see that a finite-energy $\mbox{\rm SO}(3)$-invariant connection on $S^{4}$ corresponds to a pair $(ia,\Phi)$ over $\mathcal{H}^2$ satisfying the boundary conditions \begin{displaymath} d_{a}\Phi(x) \rightarrow 0, \qquad |\Phi(x)| \rightarrow 1, \end{displaymath} as $|x| \rightarrow \infty$. Next we examine the relationship between the Chern classes of an $\mbox{\rm SO}(3)$-invariant connection $A$ on $S^{4}$ and those of the corresponding connection $ia$ over $\mathcal{H}^2$. Computing $c_{2}(A)$, we first note that the negative definite form $\xi\mapsto\mbox{Tr}(\xi)^2$ on the Lie algebra $\mathfrak{su}(2)$ corresponds to $\xi\mapsto 2\xi^{2}$ on $\sp(1)$. In this quaternionic notation we therefore have \begin{align*} c_{2}(A) & = -\frac{1}{4\pi^{2}} \int_{S^{4}}F_{A}\wedge F_{A} \\ & = -\frac{1}{16\pi^{2}} \int_{S^{4}} \left[\, ( 1-|\Phi|^{2} )\,da \,-\, 2\,\Re d_{a}\Phi\wedge\Im d_{a}\Phi \,\right] \wedge\, \left(\,dQ\wedge Q\,dQ\,\right) \\ & = \frac{i}{2\pi} \int_{\mathcal{H}^2} F_{a} = c_{1}(a). \end{align*} Here we again integrate out the $S^{2}$ factor $dQ\wedge Q\,dQ$, and on the last line we apply Stokes' theorem with the integrand \begin{align*} d\,(i\bar{\Phi}\,d\Phi) & = i\,d\bar{\Phi} \wedge d\Phi \,-\, \Phi\bar{\Phi}\,da \,-\, (\Phi\,d\bar{\Phi} + \bar{\Phi}d\Phi)\wedge a \\* & = -|\Phi|^2\,da - 2\,\Re d_{a}\Phi\wedge\Im d_{a}\Phi, \end{align*} assuming that $i\bar{\Phi}\,d\Phi$ vanishes at infinity. \subsection{Another Harmonic Function ansatz} \label{vortex-ansatz} We now return to the harmonic function ansatz that we discussed in Section~\ref{instanton-ansatz}. If we begin with a $\mbox{\rm SO}(3)$-invariant harmonic super-potential, then the resulting SD or ASD connection will also exhibit $\mbox{\rm SO}(3)$ symmetry, and from the previous section we know that such a connection is equivalent to a hyperbolic vortex or anti-vortex. In this section, we take advantage of this dimensional reduction to provide a similar harmonic function ansatz constructing solutions to the vortex equation over hyperbolic space. Our first step is to examine the relationship between the Laplacian on hyperbolic space $\mathcal{H}^{2}$ and the $\mbox{\rm SO}(3)$-symmetric Laplacian on $\mathbb{R}^{4}$. \begin{lemma} An $\mbox{\rm SO}(3)$-invariant function $\r$ on $\mathbb{R}^{4}$ is harmonic if and only if it can be written as $\r = r^{-1}\phi$, where $\phi$ is a harmonic function on $\mathcal{H}^2$. \end{lemma} \begin{proof} Using the quaternionic notation $x = t + rQ$ introduced in \S\ref{dimensional-reduction}, if $\r = \r(r,t)$ is an $\mbox{\rm SO}(3)$-invariant function on $\mathbb{R}^{4}$, then its Laplacian is \begin{equation} \label{eq:laplacian} \Delta \r = -\left( \delsq{t} + \delsq{r} + \frac{2}{r}\del{r} \right) \r. \end{equation} To cancel the unwanted linear term, we put $\r = r^{-1}\phi$, where $\phi$ is a function on $\mathcal{H}^{2}$. The Laplacian $\Delta$ then becomes \begin{displaymath} \Delta r^{-1}\phi = -\frac{1}{r} \left( \delsq{t} + \delsq{r} \right) \phi = r^{-3}\Delta_{h} \phi, \end{displaymath} where the Laplacian $\Delta_{h}$ on $\mathcal{H}^{2}$ is \begin{equation} \label{eq:hyperbolic-laplacian} \Delta_{h} = -r^{2}\left(\delsq{t} + \delsq{r}\right) \end{equation} These two Laplacians are therefore related by% \SSfootnote{In general, the conformal Laplacian is $L_{g}=\Delta+kR$, where $R$ is the scalar curvature and $k$ is a constant depending on the dimension. Taking the metric $g' = e^{2f}g$, it we have \begin{displaymath} L_{g'} = e^{-(d+2)f/2}L_{g}\,e^{(d-2)f/2}. \end{displaymath} Here the Euclidean metric on $R^{4}$ is $g=dt^{2}+dr^{2}+r^{2}dS^{2}$, where $dS^{2}$ is the metric on $S^{2}$. Taking the conformally equivalent metric $g' = r^{-2}(dt^{2}+dr^{2}) + dS^{2}$ on $\mathcal{H}^2 \times S^{2}$, Lemma \ref{lemma-scalar-curvature} tells us that $R' = 0$ since $r^{-1}$ is harmonic. It follows that $\Delta'=r^{3}\Delta r^{-1}$.} $\Delta$ = $r^{-3}\Delta_{h}r$ and thus $\r$ is harmonic on $\mathbb{R}^{4}$ if and only if $\phi$ is harmonic on $\mathcal{H}^{2}$. \end{proof} We recall from Theorem \ref{theorem-1} that the connection given by equation~(\ref{eq:self-dual}), \begin{displaymath} A = - \Im\left(\del{x}\log \r\,dx \right), \end{displaymath} is self-dual if and only if $\r$ is harmonic. In our current notation, the quaternionic differential and partial derivative in this expression are \begin{align*} dx = dt + Q\,dr + r\,dQ \qquad \del{x} = \frac{1}{2} \left( \del{t} - Q\,\del{r} - \cdots \right), \end{align*} where we have left out the portions of the partial derivative in the $Q$ directions as these vanish when applied to $\mbox{\rm SO}(3)$-invariant functions. Taking $\r = r^{-1}\phi$, we see that $\log \r = \log \phi - \log r$. Expanding equation~(\ref{eq:self-dual}) using these expressions, our $\mbox{\rm SO}(3)$-invariant self-dual connection becomes \begin{align*} A & = \frac{1}{2}\,\left(\,Q \left[\, \left( \del{r}\log\phi - \frac{1}{r}\right) dt \,-\, \del{t}\log\phi\,dr\, \right]\right. \\* & \qquad \left. \mbox{\qquad} -\,r\,\del{t}\log\phi\,dQ\,+\, \left( r\,\del{r}\log\phi - 1 \right) Q\,dQ\, \right). \end{align*} As we did in in \S\ref{dimensional-reduction}, we can extract from this connection the $\mbox{\rm U}(1)$ connection \begin{equation} \label{eq:vortex-da} d_a = d\,+\,i \left[\, \left( \del{r}\log\phi - \frac{1}{r}\right) dt - \del{t}\log\phi\,dr\, \right] \end{equation} with curvature \begin{align*} F_{a} = -i\left[ \left(\delsq{t} + \delsq{r}\right) \log\phi + \frac{1}{r^{2}} \right] dt\wedge dr = -i \left( 1 - \Delta_{h}\log\phi \right)\,r^{-2}dt\wedge dr, \end{align*} and the complex Higgs field \begin{equation} \label{eq:vortex-phi} \Phi = r \left( - \del{t}\log\phi + i\,\del{r}\log\phi \right) \end{equation} with norm \begin{displaymath} |\Phi|^{2} = r^{2}\left[ \left(\del{t}\log\phi\right)^{2} + \left(\del{r}\log\phi\right)^{2} \right] \\* = \left| \nabla\log\phi \right|_{h}^{2}. \end{displaymath} Writing the pair $(a,\phi)$ using complex notation with $z = t + ir$, we obtain the hyperbolic space analogue of Theorem \ref{theorem-1}. \begin{theorem} \label{theorem-vortex-ansatz} Given a positive real-valued super-potential $\phi$ on the hyperbolic upper half-plane $\mathcal{H}^{2}$, the connection and Higgs field pair $(a,\Phi)$ defined by \begin{align} \bar{\partial}_{a} = \bar{\partial}\,+\,\bar{\partial}\log\phi\,+\,\frac{d\bar{z}}{z-\bar{z}} \label{eq:vortex-dbar} \qquad \Phi = i\,(z-\bar{z})\,\del{z}\log\phi, \end{align} satisfies the vortex equations (\ref{eq:vortex-a}) and (\ref{eq:vortex-b}) and the pair $(a',\Phi')$ defined by \begin{align} \label{eq:anti-vortex-dbar} \partial_{a}' & = \partial\,+\,\partial\log\phi\,-\,\frac{dz}{z-\bar{z}} \\ \Phi' & = -i\,(z-\bar{z})\,\del{\bar{z}}\log\phi, \label{eq:anti-vortex-field} \end{align} satisfies the anti-vortex equations (\ref{eq:anti-vortex-a}) and (\ref{eq:anti-vortex-b}) if and only if the super-potential $\phi$ is harmonic. \end{theorem} \begin{proof} Recalling that $\ast_{h} 1 = r^{-2}dt\wedge dr$ with our hyperbolic metric, we see that the second of the vortex equations $iF_{a} = \ast_{h}\left( 1 - |\Phi|^{2} \right)$ reduces to \begin{displaymath} \frac{1}{\phi}\,\Delta_{h}\phi = \Delta_{h}\log\phi - \left| \nabla\log\phi \right|_{h}^{2} = 0. \end{displaymath} Using the complex form (\ref{eq:vortex-dbar}), it is easy to verify that the holomorphicity condition $\bar{\partial}_{a}\Phi = 0$ likewise reduces to $\Delta_{h} \phi = 0$. Hence the pair $(a, \Phi)$ satisfies the vortex equations if and only if the super-potential $\phi$ is harmonic. Similarly, the pair $(a',\phi')$ is derived by dimensional reduction from the ASD connection (\ref{eq:anti-self-dual}), and so it satisfies the anti-vortex equations if and only if $\phi$ is harmonic. \end{proof} Computing the Chern class $c_{1}$ for a pair $(a,\Phi)$ satisfying the second of the vortex equations (\ref{eq:vortex-b}), we have \begin{displaymath} c_{1}(a) = \frac{i}{2\pi}\int_{\mathcal{H}^{2}} F_{a} = \frac{1}{2\pi}\int_{\mathcal{H}^{2}} \ast_{h}\left(1-|\Phi|^{2}\right). \end{displaymath} For the vortex over the upper half-plane $\mathcal{H}^{2}$ constructed in (\ref{eq:vortex-da}) and (\ref{eq:vortex-phi}) using a harmonic super-potential $\phi$, this Chern class takes the form \begin{align*} c_{1}(a)=\frac{1}{2\pi}\int_{\mathbb{R}^{2}_{+}}\left( 1 - \Delta_{h}\log\phi \right)\,r^{-2} dt\wedge dr =\frac{1}{2\pi}\int_{\mathbb{R}^{2}_{+}}\left( \frac{1}{r^{2}} - 4\left|\del{z}\log\phi\right|^{2} \right)\,dt\wedge dr. \end{align*} Likewise, if we use the above ansatz to construct the anti-vortex corresponding to a harmonic super-potential, then the Chern class switches sign. As in \S\ref{quaternionic-notation}, we note that the curvature $F_{a}$ of a vortex is gauge invariant. Therefore, if two harmonic super-potentials $\phi_{1}$ and $\phi_{2}$ over hyperbolic space yield gauge equivalent vortices, then they must satisfy the equations \begin{displaymath} \left|\del{z}\log\phi_{1}\right| = \left|\del{z}\log\phi_{2}\right| \end{displaymath} and $\Delta_{h}\log\phi_{1} = \Delta_{h}\log\phi_{2}$. Note that it is significantly simpler to calculate $c_{1}$ directly from the vortex construction on $\mathcal{H}^{2}$ than it is by invoking dimensional reduction and computing the equivalent Chern class $c_{2}$ for the corresponding $\mbox{\rm SO}(3)$-invariant instanton over $S^{4}$. Indeed, by comparing the above expression for $c_{1}(a)$ with the expression (\ref{eq:c2}) for $c_{2}(A)$, we obtain a circuitous proof of the identity \begin{displaymath} \int_{\mathbb{R}^{2}_{+}} \left( 1 - \Delta_{h}\log\phi \right)\,r^{-2}dt \wedge dr = - \int_{\mathbb{R}^{2}_{+}} \left( \frac{1}{2}\,r^{2} \Delta\Delta\log \frac{\phi}{r} \right) dt \wedge dr \end{displaymath} for a harmonic function $\phi$ defined on the upper half-plane, where $\Delta_{h}$ is the Laplacian on $\mathcal{H}^{2}$ given by (\ref{eq:hyperbolic-laplacian}) and $\Delta$ is the Laplacian on $\mathbb{R}^{4}$ given by (\ref{eq:laplacian}). \subsection{Conformal Transformations Revisited} Instead of relying on dimensional reduction to derive the vortex ansatz of Theorem~\ref{theorem-vortex-ansatz}, we present here an interpretation of this construction that is entirely intrinsic to hyperbolic space. As we did in \S\ref{conformal-instanton}, we can treat the super-potential as a conformal transformation and then compute the Levi-Civita connection of the resulting metric. Since we are working on two-dimensional hyperbolic space, we can take advantage of complex notation to simplify our task. Let $\bar{\partial}$ be the standard holomorphic stucture on the complex upper half-plane. Choosing a holomorphic tangent frame (i.e., a single holomorphic section) $e$, consider the Hermitian metric $g$ specified by $(e,e)_{g} = \r^{2}$, where $\r$ is a smooth nonzero real-valued function. With respect to our holomorphic frame $e$, the unique connection compatible with both the holomorphic structure $\bar{\partial}$ and the metric $g$ is specified by the $(1,0)$-form \begin{displaymath} a = \r^{-2}(\partial \r^{2}) = 2\,\partial\log\r. \end{displaymath} To express this in the form of a unitary connection (in this case given by a purely imaginary complex 1-form), we must switch to a tangent frame that is orthonormal with respect to the metric $g$. In terms of the unitary frame $e' = \r^{-1}e$, the connection $a$ then becomes \begin{displaymath} a' = \partial\log\r - \bar{\partial}\log\r = 2i\,\Im\partial\log\r = -2i\,\Im\bar{\partial}\log\r \end{displaymath} and the new holomorphic structure is $\bar{\partial}'=\bar{\partial}-\bar{\partial}\log\r$, which we observe is compatible with the connection $a'$. In either frame, the curvature of this connection is given by \begin{displaymath} F_{a} = 2\,\bar{\partial}\partial\log\r = -i\,\Delta\log\r\,d\mu, \end{displaymath} where the volume form $d\mu$ and the Laplacian $\Delta$ are both taken here with respect to the Euclidean metric on $\mathbb{R}^{2}$. When working with $\mathcal{H}^2$, the hyperbolic metric $h$ on the upper half-plane corresponds to the function $\r = r^{-2}$. Taking a conformal transformation, we consider the metric $h'$ specified by a function of the form $\r = \phi^{2}/r^{2}$ with $\phi$ harmonic. The resulting unitary connection $a$ then splits into the $(0,1)$ and $(1,0)$ components \begin{equation*} \bar{\partial}_{a} = \bar{\partial}\,-\,\bar{\partial}\log\phi\,-\,\frac{d \bar{z}}{2ir} \qquad \partial_{a} = \partial \,+\,\partial\log\phi\,-\,\frac{dz}{2ir}, \end{equation*} noting that $r = (z - \bar{z})/2i$. The curvature of this connection is then \begin{displaymath} i F_{a} = \left( \Delta\log\phi - r^{-2} \right) d\mu = \left( \Delta_{h}\log\phi - 1 \right) d\mu_{h} \end{displaymath} where $d\mu_{h} = r^{-2}d\mu$ is the volume form and $\Delta_{h} = r^{2}\Delta$ is the Laplacian for the hyperbolic metric $h$. It is then easy to show that the complex Higgs field $\Phi$ defined by \begin{displaymath} \Phi = 2r\,\del{\bar{z}}\log\phi = 2\,\frac{r}{\phi}\,\del{\bar{z}}\phi \end{displaymath} satisfies the anti-vortex equations $\partial_{a}\Phi = 0$ and $\ast_{h} i F_{a} = |\Phi|^{2} - 1$ if the super-potential $\phi$ is harmonic. We observe that this pair $(a,\Phi)$ agrees with the anti-vortex (\ref{eq:anti-vortex-dbar}) and (\ref{eq:anti-vortex-field}) constructed by dimensional reduction of an anti-self-dual connection over the 4-sphere. Similarly, we can construct the vortex given by (\ref{eq:vortex-dbar}) by reversing orientation, thereby exchanging the holomorphic and anti-holomorphic structures $\bar{\partial}$ and $\partial$. \subsection{The Symmetric 't Hooft Construction} \label{vortex-tHooft} In \S\ref{tHooft}, as an illustration of the harmonic function ansatz, we constructed the 't Hooft instantons. These are the instantons formed by taking the superposition of multiple copies of the basic instanton with varying scales and distinct centers. For our super-potential, we used a sum of the Green's functions of the Laplacian, centered at the given points and weighted according to the corresponding scales. If we impose $\mbox{\rm SO}(3)$ symmetry on this class of instantons, we see that all of the centers must lie on a single real line. In fact, as we will demonstrate in the following section, {\em all\/} $\mbox{\rm SO}(3)$-invariant instantons can be constructed in this manner---as the superposition of basic instantons on a line. In this section, we examine the hyperbolic vortices associated to these symmetric 't Hooft instantons by dimensional reduction. We begin with the basic instanton with unit scale centered at the origin, which we recall is given by the $\mathbb{R}^{4}$ super-potential $\r = 1 + |x|^{-2}$. The corresponding super-potential for hyperbolic space $\mathcal{H}^{2}$ is then \begin{equation}\begin{split} \phi = r\r = r + \frac{r}{r^{2}+t^{2}} = \Im \left( z - \frac{1}{z} \right) = \frac{ (z-\bar{z})\,(1 + z\bar{z}) }{2i\,z\bar{z} }. \label{eq:simple-potential} \end{split}\end{equation} Taking the complex partial derivatives of its logarithm, we obtain \begin{align*} \del{z}\log\phi = + \frac{1}{z-\bar{z}}\,+\,\frac{\bar{z}}{1+z\bar{z}} \,-\,\frac{1}{z} \qquad \del{\bar{z}}\log\phi = - \frac{1}{z-\bar{z}}\,+\,\frac{z}{1+z\bar{z}} \,-\,\frac{1}{\bar{z}}. \end{align*} Inserting these expressions into the formula (\ref{eq:vortex-dbar}) gives us the connection and Higgs field \begin{equation*} \bar{\partial}_{a} = \bar{\partial}\,-\,\frac{d\bar{z}}{\bar{z}\,(1+z\bar{z})} \qquad \Phi = i\,\frac{\bar{z}\,(1+z^{2})}{z\,(1+z\bar{z})}, \end{equation*} satisfying the vortex equations (\ref{eq:vortex-a}) and (\ref{eq:vortex-b}). Note that as $z$ approaches the real axis, the Higgs field obeys the boundary condition $|\Phi|\rightarrow 1$. One of the primary results concerning solutions to the vortex equations is that they are uniquely specified up to gauge equivalence by the zeros of the Higgs field (see \cite[Chapter III]{JT}). In the example above, we see that $\Phi$ vanishes at the point $z = i$. If we alter the scale of our basic instanton and translate it along the real axis, the $\mathcal{H}^{2}$ super-potential becomes \begin{displaymath} \phi = \Im \left( z - \frac{\lambda}{z-a} \right) \end{displaymath} with $\lambda > 0$ and $a$ real. The corresponding vortex is then \begin{equation*} \bar{\partial}_{a} = \bar{\partial}\,-\,\frac{\lambda\,d\bar{z}} {(\bar{z}-\bar{a})\,(1+|z-a|^{2})} \qquad \Phi = i\,\frac{(\bar{z}-\bar{a})\,(1+(z-a)^{2})} {(z-a)\,(1+|z-a|^{2})}, \end{equation*} and we see that $\Phi$ vanishes at the point $z = a + i\sqrt{\lambda}$. Most generally, given a set of $k$ complex points $\{z_{i}\}$ in the upper half-plane, the super-potential \begin{displaymath} \phi = \Im \left( z - \sum_{i = 1}^{k} \frac{(\Im z_{i})^2}{z - \Re z_{i}} \right) \end{displaymath} generates the unique hyperbolic vortex with Higgs field vanishing at the points $\{z_{i}\}$. Hence the centers of the instantons correspond to the real parts of the complex zeros, while the scales correspond to their imaginary parts. \subsection{The Equivariant ADHM Construction} In this section we shall use an $\mbox{\rm SO}(3)$ equivariant version of the ADHM construction \cite{ADHM} in order to provide an alternative construction for the symmetric 't Hooft instantons discussed in the previous section. In addition, since the ADHM construction actually generates {\em all} possible anti-self-dual connections on bundles over $S^{4}$, we will then be able to show that every symmetric instanton must be gauge equivalent to one constructed using the 't Hooft ansatz. By dimensional reduction, this gives us a complete classification of hyperbolic vortices, proving that such vortices are uniquely determined up to a gauge transformation by the zeros of their Higgs fields. This is to be contrasted with Euclidean vortices, in which case the classification theorem may be proved using approximation techniques (see \cite[Chapter III]{JT}), but no explicit construction for the vortex solutions is known. \newcommand{W\!\otimes\!\H_{1}}{W\!\otimes\!\H_{1}} Here we use the quaternionic version of the construction as discussed in \cite{A}. When dealing with quaternionic vector spaces and linear maps, we use the convention that scalar multiplication acts on the {\em right}. Recall from \S\ref{dimensional-reduction} that under our $\mbox{\rm Sp}(1)$-action, we may view $S^{4}$ as the quaternionic projective space $P(\H_{1}^{2})$, where $\H_{1}$ is the fundamental representation with $\mbox{\rm Sp}(1)$ acting by {\em left} quaternion multiplication. To construct an $\mbox{\rm Sp}(1)$-invariant ASD connection on the bundle $E\rightarrow S^{4}$ with Chern class $c_{2}(E) = -k$, we introduce the $k+1$ dimensional {\em quaternionic\/} $\mbox{\rm Sp}(1)$ representation $V$ given by \begin{equation} \label{eq:V} V = \mbox{Ker}\,\mathcal{D}_{A}^{\ast} : \Gamma (S^{4}, E\otimes S^{-}\otimes S^{-}) \rightarrow \Gamma (S^{4}, E\otimes S^{+}\otimes S^{-}) \end{equation} and the $k$ dimensional {\em real\/} $\mbox{\rm Sp}(1)$ representation $W$ given by \begin{equation} \label{eq:W} W = \left( \mbox{Ker}\,\mathcal{D}_{A}^{\ast} : \Gamma (S^{4}, E\otimes S^{-}) \rightarrow \Gamma (S^{4}, E\otimes S^{+}) \right)_{\mathbb{R}}^{\ast}, \end{equation} where $A$ is an arbitrary connection on $E$ (the spaces $V,W$ are independent of the connection), $S^{\pm}$ are the two quaternionic half-spin bundles, and $\mathcal{D}_{A}^{\ast}$ is the adjoint of the Dirac operator with coefficients in $E\otimes S^{-}$ and $E$ respectively. Using these spaces $V,W$ the ADHM data consists of the the three maps: \begin{itemize} \samepage \item an arbitrary $\mbox{\rm Sp}(1)$ equivariant inclusion $W\!\otimes\!\H_{1} \hookrightarrow V$ \item an $\mbox{\rm Sp}(1)$ equivariant $\H$-linear map $B:W\!\otimes\!\H_{1}\rightarrowW\!\otimes\!\H_{1}$ satisfying $B^{\ast} = \bar{B}$ (i.e., $B$ is represented by a symmetric matrix) \item an $\mbox{\rm Sp}(1)$ equivariant $\H$-linear map $\Lambda : W\!\otimes\!\H_{1} \rightarrow V\,/\,W\!\otimes\!\H_{1}$. \end{itemize} If we fix the inclusion $W\!\otimes\!\H_{1} \hookrightarrow V$, then we say that two sets of ADHM data $(B,\Lambda)$ and $(B',\Lambda')$ are {\em equivalent\/} if \begin{displaymath} B' = U B U^{-1}, \qquad \Lambda' = v \Lambda U^{-1} \end{displaymath} for suitable $U\in O(W)$ and $v\in \mbox{\rm Sp}(V\,/\,W\!\otimes\!\H_{1})$. From the ADHM data, we construct an $\mbox{\rm Sp}(1)$ equivariant family of $\H$-linear maps $v(x) : W\!\otimes\!\H_{1} \rightarrow V$ parametrized by $x\in\H$, given by \begin{displaymath} v(x) = \left( \begin{array}{c} \Lambda \\ B - xI \end{array} \right) \end{displaymath} relative to the decomposition $V = (V\,/\,W\!\otimes\!\H_{1}) \oplus (W\!\otimes\!\H_{1})$. \begin{theorem}[ADHM] \label{ADHM} There is a one-to-one correspondence between equivalence classes of ADHM data $(B,\Lambda)$ satisfying the two conditions \begin{description} \item[non-degeneracy] $v(x)$ is injective for all $x\in\H$ \item[ADHM condition] $\Lambda^{\ast}\Lambda + B^{\ast}B : W\!\otimes\!\H_{1} \rightarrow W\!\otimes\!\H_{1}$ is real, \end{description} and gauge equivalence classes of $\mbox{\rm Sp}(1)$-invariant ASD connections on $E$. \end{theorem} To construct the connection associated to a set of ADHM data, we first observe that the non-degeneracy condition implies that $f : x \mapsto \mbox{Coker}\,v(x)\subset V$ is a smooth map from $S^{4}$ to the quaternionic projective space $P(V)$ (we map the point at $\infty$ to the line $V\,/\,W\!\otimes\!\H_{1}$). We then define the corresponding vector bundle $E$ and connection $A$ to be the pullback of the canonical quaternionic line bundle over $P(V)$ with its standard connection (induced by orthogonal projection from the trivial flat connection on V). The fact that $A$ is ASD follows from the ADHM condition. For a complete proof of the non-equivariant version of this theorem, see \cite[\S3.3]{DK} or \cite{A}. The proof of equivariant version then proceeds with minimal modification. We now give an even more precise description of $\mbox{\rm Sp}(1)$-invariant instantons, starting by examining the characters of the representations $V$ and $W$. \begin{lemma} The ADHM representations $V,W$ described in (\ref{eq:V}) and (\ref{eq:W}) corresponding to the bundle $E \rightarrow S^{4}$ with $c_{2}(E) = -k$ are given by \begin{displaymath} V = \H_{1}^{k+1}, \qquad W = \mathbb{R}_{0}^{k}, \end{displaymath} where $\H_{1}$ is the fundamental representation of $\mbox{\rm Sp}(1)$ acting by left multiplication and $\mathbb{R}_{0}$ is the trivial real representation. \end{lemma} \begin{proof} If $E$ is an $\mbox{\rm Sp}(1)$ equivariant vector bundle, then it is clearly also equivariant with respect to any one-parameter subgroup $S^{1}$ of $\mbox{\rm Sp}(1)$. We may therefore apply the results of \cite{Br} and \cite{BA}. The fixed point set for this $S^{1}$-action on $S^{4}$ is the sphere $S^{2}$, over which the bundle $E$ splits as $E|_{S^{2}}=L\oplus L^{\ast}$. Here $L$ is a complex line bundle with an $S^{1}$-action and $L^{\ast}$ is its dual. As in \cite{Br}, such $S^{1}$ equivariant bundles $E$ are characterized by a pair of constants $(m,k)$, where $m$ is the weight of the $S^{1}$-action on $L$ and $k = c_{1}(L^{\ast})$. In our case this $S^{1}$-action is derived from the fundamental representation of $\mbox{\rm Sp}(1)$, and so its weight is $m = 1/2$. In addition, noting that $2mk = -c_{2}(E)$, we see that the two definitions of $k$ agree. Using the equivariant index calculations of \cite{Br}, Braam and Austin compute the representations $V, W$ in equations (3.4) and (3.5) of \cite{BA}. For $m = 1/2$ we have $V=(\mathbb{C}_{1/2}\oplus\mathbb{C}_{-1/2})^{k+1}$ and $W=\mathbb{R}_{0}^{k}$ as representations of $S^{1}$. The corresponding $\mbox{\rm Sp}(1)$ representations are then $V = \H_{1}^{k+1}$ and $W = \mathbb{R}_{0}^{k}$. \end{proof} From this lemma, we see that the ADHM data $(B,\Lambda)$ is a pair of $\mbox{\rm Sp}(1)$ equivariant maps $B : \H_{1}^{k}\rightarrow\H_{1}^{k}$ and $\Lambda : \H_{1}\rightarrow\H_{1}$. We then have $gBg^{-1} = B$ and $g\Lambda g^{-1} = \Lambda$ for all $g\in\mbox{\rm Sp}(1)$, where $g$ acts by quaternion multiplication. It follows that both $B$ and $\Lambda$ must be {\em real\/} transformations. Recalling that $B$ is symmetric, we see that $B$ is diagonalizable with real eigenvalues. Choosing a suitable basis, we can therefore write the map $v(x) : \H_{1}^{k} \rightarrow \H_{1}^{k+1}$ as a matrix of the form \begin{displaymath} v(x) = \left( \begin{array}{ccc} \lambda_{1} & \cdots & \lambda_{k} \\ b_{1} - x & 0 & 0 \\ 0 & \ddots & 0 \\ 0 & 0 & b_{k} - x \end{array} \right) \end{displaymath} with real eigenvalues $\{b_{1},\ldots,b_{k}\}$ and real scales $\{\lambda_{1},\ldots,\lambda_{k}\}$ with $\lambda_{i} \geq 0$. The non-degeneracy condition of Theorem \ref{ADHM} implies that the $b_{i}$ are distinct and the $\lambda_{i}$ are nonzero, while the ADHM condition is automatically satisfied since $B$ and $\Lambda$ are real. Computing $\mbox{Coker}\,v(x) = \mbox{Ker}\,v(x)^{\ast}$, we obtain \begin{theorem} Given $k$ distinct real centers $\{b_{1},\ldots,b_{k}\}$ and $k$ positive real scales $\{\lambda_{1},\ldots,\lambda_{k}\}$, let $f : S^{4} \rightarrow P(\H_{1}^{k+1})$ be the map given by \begin{displaymath} f\,:\,x\,\mapsto\,\left( 1\,:\,\frac{\lambda_{1}}{x-b_{1}}\,:\, \cdots\,:\,\frac{\lambda_{k}}{x-b_{k}} \right). \end{displaymath} The $\mbox{\rm Sp}(1)$-invariant connection obtained by taking the pullback of the standard connection on the canonical bundle over $P(\H_{1}^{k+1})$ is then ASD and has Chern class $c_{2} = -k$. Furthermore, every $\mbox{\rm Sp}(1)$-invariant connection on a bundle $E\rightarrow S^{4}$ with $c_{2}(E) = -k$ is gauge equivalent to one of this form. \end{theorem} We note that the instantons constructed by the above theorem are the superposition of $k$ basic instantons with distinct centers along the real axis, which we recognize as the 't Hooft instantons from the previous section in a different guise. By dimensional reduction, we therefore obtain a constructive proof of the classification theorem for hyperbolic vortices. \begin{corollary} Given a set of $k$ distinct points $\{z_{1},\ldots,z_{k}\}$ in the complex upper half-plane, there exists a finite action solution $(a,\Phi)$ to the hyperbolic vortex equations, unique up to gauge transformation, such that $\{z_{1},\ldots,z_{k}\}$ is the zero set of the Higgs field $\Phi$. The Chern class of such a vortex is then $c_{1}(a) = k$, obtained by counting the zeros of the Higgs field. \end{corollary} \subsection{Symmetric Gauge Transformations} \label{gauge-transformation} Now that we understand the relationship between symmetric instantons and hyperbolic vortices, we examine how the notion of gauge equivalence behaves under this dimensional reduction. Starting with a symmetric $\mbox{\rm SU}(2)$ gauge transformation on $S^{4}$, we compute the resulting $\mbox{\rm U}(1)$ gauge transformation on hyperbolic space $\mathcal{H}^{2}$. Then, continuing to work in the simpler $\mathcal{H}^{2}$ picture, we examine the conditions under which two hyperbolic vortices given by the harmonic function ansatz of \S\ref{vortex-ansatz} are gauge equivalent. Using the quaternionic notation $x = t + rQ$, the most general $\mbox{\rm Sp}(1)$-invariant gauge transformation on $S^{4}$ has the form \begin{displaymath} g = e^{Q\,\chi(t,r)} = \cos\chi(t,r) + Q\sin\chi(t,r), \end{displaymath} where $\chi$ is a real-valued function on $\mathcal{H}^{2}$. Computing its differential, we have \begin{equation*} dg\,g^{-1} = Q\,d\chi\,+\,\sin\chi\:dQ\:e^{-Q\chi} = Q\,d\chi\,-\,\frac{1}{2}\,Q\,(e^{2Q\chi}-1)\,dQ. \end{equation*} Applying this gauge transformation to the general $\mbox{\rm SO}(3)$-invariant connection (\ref{symmetric-connection}), we obtain \begin{align*} g(A)&= \frac{1}{2} \, e^{Q\chi}\,\left[\, Qa\,+\,(\Phi_{1}+Q\Phi_{2})\,dQ \,\right]\,e^{-Q\chi} \,-\,dg\,g^{-1} \\ &=\frac{1}{2}\left[\, Q\,(a - 2\,d\chi) \,+\, \left( e^{2Q\chi}\,[\,\Phi_{1}+Q\,(\Phi_{2}+1)\,] - Q \right)dQ \,\right], \end{align*} noting that $g\,Q = Q\,g$ while $g\,dQ = dQ\,g^{-1}$. The associated connection $ia$ and Higgs field $\Phi = \Phi_{1} + i\,(\Phi_{2}+1)$ over $\mathcal{H}^{2}$ then transform according to \begin{displaymath} g(a) = a - 2\,d\chi, \qquad g(\Phi) = e^{2i\chi}\Phi. \end{displaymath} Hence, the corresponding gauge transformation over $\mathcal{H}^{2}$ is simply $g = e^{2i\chi}$. Suppose that we have two gauge equivalent hyperbolic vortices $(a_{+},\Phi_{+})$ and $(a_{-},\Phi_{-})$ constructed by the harmonic function ansatz of \S\ref{vortex-ansatz}, using equation (\ref{eq:vortex-dbar}) with the super-potentials $\phi_{+}$ and $\phi_{-}$ respectively. If these two vortices satisfy $a_{-} = g(a_{+})$ and $\Phi_{-} = g(\Phi_{+})$ with a gauge transformation of the form $g = e^{2i\chi}$ as discussed above, then we obtain the system of differential equations \begin{align} \label{eq:equivalent-connections} \del{\bar{z}} \log\phi_{-} & = \del{\bar{z}} \log\phi_{+}\,-\, \del{\bar{z}}\,2i\chi \\ \label{eq:equivalent-fields} \del{z} \log\phi_{-} & = e^{2i\chi}\,\del{z} \log\phi_{+}. \end{align} Note that equation (\ref{eq:equivalent-connections}) implies that $\chi$ is harmonic, \begin{displaymath} \del{z}\del{\bar{z}}\,\chi = 0, \end{displaymath} as it is the imaginary part of a holomorphic function. Taking the conjugate of (\ref{eq:equivalent-fields}) and inserting it into (\ref{eq:equivalent-connections}), we can eliminate either $\log\phi_{+}$ or $\log\phi_{-}$ from these equations to obtain \begin{align*} \del{\bar{z}}\log\phi_{+} = \frac{e^{2i\chi}}{e^{2i\chi}-1}\,\del{\bar{z}}\,2i\chi & \qquad \del{z}\log\phi_{+} = \frac{1}{e^{2i\chi}-1}\,\del{z}\,2i\chi \\ \del{\bar{z}}\log\phi_{-} = \frac{1}{e^{2i\chi}-1}\,\del{\bar{z}}\,2i\chi & \qquad \del{\bar{z}}\log\phi_{-} = \frac{e^{2i\chi}}{e^{2i\chi}-1}\,\del{z}\,2i\chi. \end{align*} These equations may also be written in either of the two simpler forms \begin{align*} \del{\bar{z}}\log\phi_{\pm} = \frac{e^{\pm i\chi}}{\sin\chi}\,\del{\bar{z}}\,\chi & \qquad \del{z}\log\phi_{\pm} = \frac{e^{\mp i\chi}}{\sin\chi}\,\del{z}\,\chi \end{align*} or \begin{align} \label{eq:dzbar} \del{\bar{z}}\log\phi_{\pm} & = \del{\bar{z}}\log\left( e^{\pm 2i\chi}-1 \right) \qquad \del{z}\log\phi_{\pm} = \del{z}\log\left( e^{\mp 2i\chi}-1 \right). \end{align} Substituting these formulae for the partial derivatives into (\ref{eq:vortex-dbar}), we can express the two gauge equivalent hyperbolic vortices completely in terms of the gauge transformation without reference to their super-potentials. Furthermore, computing the Laplacian of the super-potentials $\phi_{1},\phi_{2}$ in terms of the function $\chi$, we have \begin{equation*} \frac{1}{\phi_{\pm}}\,\del{\bar{z}}\del{z}\,\phi_{\pm} = \del{\bar{z}}\log\phi_{\pm}\,\del{z}\log\phi_{\pm} \,+\, \del{\bar{z}}\del{z}\log\phi_{\pm} = \frac{e^{\pm i\chi}}{\sin\chi}\,\del{z}\del{\bar{z}}\,\chi. \end{equation*} Hence, the requirements that $\phi_{1},\phi_{2}$ be harmonic reduce simply to the condition that $\chi$ be harmonic, which we have already established as a corollary to equation~(\ref{eq:equivalent-connections}). We have thus proved \begin{theorem} \label{vortex-construction} Let $\chi$ be a real-valued harmonic function on the hyperbolic upper half-plane $\mathcal{H}^2$. The two pairs $(a_{+},\Phi_{+})$ and $(a_{-},\Phi_{-})$ given by \begin{align*} \bar{\partial}_{a_{\pm}} & = \bar{\partial}\,+\, \bar{\partial}\log\left( e^{\pm 2i\chi}-1 \right) \,+\, \frac{d\bar{z}}{z-\bar{z}} \\ \Phi_{\pm} & = i\,(z-\bar{z})\, \del{z}\log\left( e^{\mp 2i\chi}-1 \right) \end{align*} then satisfy the hyperbolic vortex equations (\ref{eq:vortex-a}) and (\ref{eq:vortex-b}) and are related by the gauge transformation $g = e^{2i\chi}$. Conversely, any two gauge equivalent hyperbolic vortices constructed via the harmonic function ansatz of Theorem~\ref{theorem-vortex-ansatz} can be expressed in this form. \end{theorem} \subsection{The Unit Disc Model} \label{unit-disc} Until now, we have always used the upper half-plane model for hyperbolic space $\mathcal{H}^2$. In some circumstances, it will be more convenient to use the unit disc model. While the upper half-plane arises naturally by the dimensional reduction technique discussed in the previous sections, the calculations in the following section become much simpler and exhibit significantly more symmetry if we can work on the unit disk. Here we make the transition beween the two coordinate systems, showing how the formulae of the previous sections behave under the transformation. Letting $z$ be the complex coordinate for the upper half-plane and $w$ the coordinate for the unit disc, these two models are related by the conformal transformation \begin{equation} \label{eq:coordinate-transform} w = \frac{i-z}{i+z}, \qquad z = i\,\frac{1-w}{1+w}. \end{equation} This map takes the upper half-plane to the interior of the unit disc, mapping the real axis to the unit circle. In particular, the point $z = i$ maps to the origin, while the origin maps to $w = 1$ and the point at infinity maps to $w = -1$. The positive imaginary axis in $z$-coordinates maps to the interval $(-1,1)$ on the real axis in $w$-coordinates. The hyperbolic metric on the unit disk is given by \begin{displaymath} h = \frac{4}{\left(1-|w|^2\right)^2}\,dw\,d\bar{w}. \end{displaymath} The vortex equations remain fixed under this change of coordinates; equation (\ref{eq:vortex-a}) is preserved because the transformation is holomorphic, while equation (\ref{eq:vortex-b}) is a relationship between coordinate-invariant scalar quantities. To compute the new connection and Higgs fields generated by the harmonic function ansatz, we first note that \begin{displaymath} z - \bar{z} = 2i\,\frac{1-|w|^2}{|1+w|^2}, \end{displaymath} and that the partial derivatives transform according to \begin{displaymath} \del{w} = -\frac{2i}{(1+w)^2}\,\del{z}, \qquad \del{\bar{w}} = \frac{2i}{(1+\bar{w})^2}\,\del{\bar{z}}. \end{displaymath} Converting the formula (\ref{eq:vortex-dbar}) for the hyperbolic vortex associated to a harmonic super-potential $\phi$ to $w$-coordinates on the unit disc, Theorem~\ref{theorem-vortex-ansatz} then becomes \begin{theorem} \label{w-theorem-vortex-ansatz} Given a positive real-valued super-potential $\phi$ over the hyperbolic disc $\mathcal{H}^{2}$, the $\mbox{\rm U}(1)$ connection and Higgs field pair $(a, \Phi)$ defined by \begin{align} \label{eq:w-dbar} \bar{\partial}_{a} = \bar{\partial} \,+\, \bar{\partial}\log\phi \,+\, \frac{d\bar{w}}{1-|w|^2}\,\frac{1+w}{1+\bar{w}} \qquad \Phi =-i\,(1-|w|^2)\,\frac{1+w}{1+\bar{w}}\,\del{w}\log\phi, \end{align} satisfies the vortex equations (\ref{eq:vortex-a}) and (\ref{eq:vortex-b}) if and only if the super-potential $\phi$ is harmonic. \end{theorem} The Chern class $c_{1}(a)$ of this connection is given in these coordinates by \begin{equation}\begin{split} c_{1}(a) =\frac{1}{2\pi}\int_{\mathcal{H}^{2}} \ast_{h}\left(1-|\Phi|^{2}\right) =\frac{1}{2\pi}\int_{D^{2}}\left( \frac{4}{\left(1-|w|^{2}\right)^{2}} - 4\,\left|\del{w}\log\phi\right|^{2} \right) d\mu, \label{eq:c1-w} \end{split}\end{equation} where $d\mu$ is the volume element on the disc. We now return to the simple hyperbolic vortex constructed in \S\ref{vortex-tHooft}, which we obtained by a dimensional reduction of the basic $c_{2}=1$ instanton. In $w$-coordinates on the unit disc, the super-potential (\ref{eq:simple-potential}) becomes \begin{equation} \label{eq:w-simple-potential} \phi = 2\,\frac{1-|w|^{4}}{\left|1-w^{2}\right|^{2}}. \end{equation} Computing the partial derivatives of its logarithm, we obtain \begin{equation*} \del{w}\log\phi = \frac{2w}{1-|w|^{4}}\, \frac{1-\bar{w}^{2}}{1-w^{2}} \qquad \del{\bar{w}}\log\phi = \frac{2\bar{w}}{1-|w|^{4}}\, \frac{1-w^{2}}{1-\bar{w}^{2}}, \end{equation*} which when inserted into formula (\ref{eq:w-dbar}) above give the vortex \begin{equation*} \bar{\partial}_{a} = \bar{\partial} + \frac{d\bar{w}}{1+|w|^{2}}\, \frac{1+w}{1-\bar{w}}\, \qquad \Phi = -\frac{2iw}{1+|w|^{2}}\,\frac{1-\bar{w}}{1-w}. \end{equation*} Note that here the Higgs field $\Phi$ vanishes only at the origin, where it has a simple zero, and so the Chern class of this vortex should be $c_{1}(a) = 1$. Using (\ref{eq:c1-w}) to explicitly calculate this Chern class, we obtain the integral \begin{displaymath} c_{1}(a) = \frac{1}{2\pi} \int_{D^{2}} \left( \frac{4}{\left(1-|w|^{2}\right)^{2}} - \frac{16\,|w|^{2}}{\left(1-|w|^{4}\right)^{2}} \right) d\mu. \end{displaymath} Taking polar coordinate $r,\theta$ on the unit disc, this integral becomes \begin{equation*} c_{1}(a) = \int_{0}^{1} \left( \frac{4\,r}{\left(1-r^2\right)^2} - \frac{16\,r^3}{\left(1-r^4\right)^2} \right) dr = \left[ \frac{2}{1-r^2} - \frac{4}{1-r^4} \right]_{r=0}^{r=1}. \end{equation*} To evaluate this expression at $r=1$, we substitute $r=1-x$ and expand it about $x=0$, giving us \begin{align*} \frac{2}{1-(1-x)^{2}} - \frac{4}{1-(1-x)^4} &= \frac{2}{2x-x^2+O(x^3)} - \frac{4}{4x-6x^2+O(x^3)} \\* &= \left(\frac{1}{x} + \frac{1}{2} + O(x)\right) - \left(\frac{1}{x} + \frac{3}{2} + O(x)\right) \\* &= -1 + O(x) \rule{0in}{3.5ex}. \end{align*} We therefore see that the Chern class of this vortex is indeed $c_{1}(a) = 1$ as we predicted by counting the zeros of the Higgs field. The formulae from \S\ref{gauge-transformation} giving the two super-potentials in terms of the gauge transformation remain unchanged, except for replacing all the $z$'s with $w$'s. In particular, if the vortices determined by the super-potentials $\phi_{+}$ and $\phi_{-}$ are gauge equivalent by a transformation of the form $g = e^{2i\chi}$, then the differential equations (\ref{eq:equivalent-connections}) and (\ref{eq:equivalent-fields}) become \begin{align} \label{eq:w-equivalent-connections} \del{\bar{w}} \log\phi_{-} & = \del{\bar{w}} \log\phi_{+}\,-\, \del{\bar{w}}\,2i\chi \\ \label{eq:w-equivalent-fields} \del{w} \log\phi_{-} & = e^{2i\chi}\,\del{w} \log\phi_{+}. \end{align} The unit disc version of Theorem \ref{vortex-construction} is then \begin{theorem} \label{w-vortex-construction} Let $\chi$ be a real-valued harmonic function on the hyperbolic unit disc $\mathcal{H}^2$. The two pairs $(a_{+},\Phi_{+})$ and $(a_{-},\Phi_{-})$ given by \begin{align*} \bar{\partial}_{a_{\pm}} & = \bar{\partial}\,+\, \bar{\partial}\log\left( e^{\pm 2i\chi}-1 \right) \,+\, \frac{d\bar{w}}{1-|w|^2}\,\frac{1+w}{1+\bar{w}} \\ \Phi_{\pm} & = -i\,(1-|w|^2)\,\frac{1+w}{1+\bar{w}}\, \del{w}\log\left( e^{\mp 2i\chi}-1 \right) \end{align*} then satisfy the hyperbolic vortex equations (\ref{eq:vortex-a}) and (\ref{eq:vortex-b}) and are related by the gauge transformation $g = e^{2i\chi}$. Conversely, any two gauge equivalent hyperbolic vortices constructed via the harmonic function ansatz of Theorem~\ref{w-theorem-vortex-ansatz} can be expressed in this form. \end{theorem} \section{Holonomy Singularity} \subsection{The Forg\'{a}cs, Horv\'{a}th, Palla Instanton} \label{fhp} In this section, we construct the singular instanton described by P. Forg\'{a}cs, Z. Horv\'{a}th, and L. Palla in \cite{FHP1}. In order to obtain a connection on $S^4\setminus S^2$ with a holonomy singularity, Forg\'{a}cs {\em et al.\/} patch together two non-singular connections on overlapping simply connected regions using a gauge transformation. (This process is not unlike the clutching construction, which creates ``twisted'' vector bundles given their local trivializations and transition functions.) These two non-singular solutions are generated by the harmonic function ansatz of Section~\ref{instanton-ansatz}, and since the super-potentials they use are $\mbox{\rm SO}(3)$-invariant, the resulting connection can be analyzed in terms of the dimensional reduction to hyperbolic space $\mathcal{H}^2$ discussed in Section 2. Using the quaternionic notation $x = t + rQ$ with $t, r$ real, $r > 0$ and $Q$ pure imaginary satisfying $Q^{2} = -1$, we want to construct an $\mbox{\rm SO}(3)$-invariant self-dual connection singular along the 2-sphere $t=0, r=1$. By dimensional reduction, this translates into a vortex over the hyperbolic upper half-plane with non-trivial holonomy around the point $z = i$. If instead we work using the unit disc model of hyperbolic space, our task takes the more symmetric form of finding a hyperbolic vortex on the punctured disc with a holonomy singularity at the origin. We therefore set out to construct two gauge equivalent hyperbolic vortices on the punctured disc, using the harmonic function ansatz of \S\ref{vortex-ansatz}. For the two simply connected regions, we let $P_{1}$ be the disc with a cut along the positive real axis, and let $P_{2}$ be the disc with a cut along the negative real axis. The areas of overlap are then the upper and lower half-discs, excluding the real axis. Let $(a_{1},\Phi_{1})$ and $(a_{2},\Phi_{2})$ be the hyperbolic vortices corresponding to the super-potentials $\phi_{1}$ and $\phi_{2}$ on the regions $P_{1}$ and $P_{2}$ respectively. In light of Theorem~\ref{vortex-construction}, we begin by examining the gauge transformation between these two vortices, rather than focusing on the super-potentials. Here our gauge transformation $g$ is specified by a real-valued harmonic function $\chi$, and we take \begin{displaymath} g = \left\{ \begin{array}{ll} e^{+2i\chi} & \mbox{for $\Im w > 0$} \\ e^{-2i\chi} & \mbox{for $\Im w < 0$}. \end{array} \right. \end{displaymath} In other words, on the lower half-disc we use the inverse of the gauge transformation that we use on the upper half-disc. In the notation of \S\ref{gauge-transformation}, switching the sign of $2i\chi$ simply interchanges the resulting gauge equivalent super-potentials $\phi_{+}$ and $\phi_{-}$ determined by equation (\ref{eq:dzbar}). Note that although the resulting $g$ is undefined along the real axis, this does not pose a problem for our construction. Indeed, we use $g$ directly only on regions excluding the real axis, and we will find that the two hyperbolic vortices $(a_{1},\Phi_{1})$ and $(a_{2},\Phi_{2})$ constructed from $g$ are continuous across the negative and positive axes respectively. In \cite{FHP1}, Forg\'{a}cs {\em et al.\/} use the gauge transformation specified by \begin{equation} \label{eq:imlog} 2\chi = \left(\, \frac{\pi}{2}\,+\,2 \arctan \frac{T_{2}}{1-T_{1}} \,+\,2 \arctan \frac{T_{1}}{1-T_{2}} \,\right). \end{equation} We will define the $T_{i}$ below, but before doing so we first study the behavior of this gauge transformation for general values of $T_{i}$. In particular, we would like to coerce $T_{1}$ and $T_{2}$ into being the real and imaginary parts of a holomorphic (or anti-holomorphic) function $f(w)$. Then where $|f(w)| = 1$, the arguments of the arctans resemble the half-angle formula for $\tan(\theta)$. In such circumstances we have \begin{equation*} \arctan \frac{T_{2}}{1-T_{1}} = \Im\log\,(1-\bar{f}) \qquad \arctan \frac{T_{1}}{1-T_{2}} = \Im\log\,(1+if), \end{equation*} and we can write $\chi$ in terms of $f$ using \begin{displaymath} 2\chi = \Im\log\left( i\,\frac{(1+if)^{2}}{(1-f)^{2}} \right). \end{displaymath} We then see that $\chi$ is indeed harmonic as it is the imaginary part of a holomorphic (or anti-holomorphic) function. Exponentiating, we obtain \begin{equation} \label{eq:exp-2ichi} e^{2i\chi} = i\,\frac{(1-\bar{f})\,(1+if)}{(1-f)\,(1-i\bar{f})}, \end{equation} and the expressions $e^{2i\chi}-1$ and $e^{-2i\chi}-1$ take the form \begin{equation*} e^{+2i\chi} - 1 = -\frac{(1-i)\,(1-f\bar{f})}{(1-f)\,(1-i\bar{f})}, \qquad e^{-2i\chi} - 1 = -\frac{(1+i)\,(1-f\bar{f})}{(1-\bar{f})\,(1+if)}, \end{equation*} which we shall use in equation (\ref{eq:dzbar}). Note that if $|f| = 1$ then $\bar{f} = f^{-1}$, and we observe that $e^{\pm 2i\chi} = 1$. In equation (\ref{eq:imlog}) above, the $T_{i}$ are defined by \begin{align*} T_{1}&=\frac{1}{2\,S^{5}}\,\sqrt{(S+S_{-})^{2}-4}\,\left\{ \frac{1}{4} \left[4 - (S-S_{-})^2 \right]^{2}\,+\, S^2 S_{-}^2\,-\,3\,(z+\bar{z})^{2} \right\} \\ T_{2}&=\frac{1}{2\,S^{5}}\,\sqrt{4-(S-S_{-})^{2}}\,\left\{ \frac{1}{4} \left[4 - (S+S_{-})^2 \right]^{2}\,+\, S^2 S_{-}^2\,-\,3\,(z+\bar{z})^{2} \right\} \end{align*} with \begin{displaymath} S = |z + i|, \qquad S_{-} = |z - i|, \end{displaymath} using the complex coordinate $z$ on the upper half-plane. With formulae such as these, it is not surprising that the mathematical community was incredulous. Changing to the complex coordinate $w$ on the unit disc via the conformal transformation (\ref{eq:coordinate-transform}), we have \begin{displaymath} S = \frac{2}{|w+1|}, \qquad S_{-} = 2\,\frac{|w|}{|w+1|} = |w| S. \end{displaymath} Calculating the various components of the $T_{i}$, we obtain \begin{align*} \sqrt{(S+S_{-})^{2}-4} & = 2\,\frac{|\sqrt{w}-\sqrt{\bar{w}}|}{|w+1|} = \left|\sqrt{w}-\sqrt{\bar{w}}\right|S \\ \sqrt{4-(S-S_{-})^{2}} & = 2\,\frac{|\sqrt{w}+\sqrt{\bar{w}}|}{|w+1|} = \left|\sqrt{w}+\sqrt{\bar{w}}\right|S \\ \end{align*} and \begin{displaymath} z + \bar{z} = -2i\,\frac{w-\bar{w}}{|w+1|^{2}} = -\frac{i}{2}\,(w-\bar{w})\,S^{2}. \end{displaymath} Putting these pieces together, the $T_{i}$ are given much more simply by \begin{align*} T_{1} & = \frac{1}{2} \left|\sqrt{w}-\sqrt{\bar{w}}\right| \left[\, \frac{1}{4}\left(\sqrt{w}+\sqrt{\bar{w}}\right)^{4} \,+\, w\bar{w} \,+\, \frac{3}{4}\,(w-\bar{w})^{2} \,\right] \\* & = \frac{1}{2} \left|w^{1/2}-\bar{w}^{1/2}\right| \left( w^2 + w^{3/2}\bar{w}^{1/2} + w\bar{w} + w^{1/2}\bar{w}^{3/2} + \bar{w}^{2} \right) \\* & = \mp\frac{i}{2} \left(w^{5/2}-\bar{w}^{5/2}\right) = \left(\Im w^{5/2}\right) \left(\mbox{sign}\:\Im w^{1/2}\right), \end{align*} and \begin{align*} T_{2} & = \frac{1}{2} \left|\sqrt{w}+\sqrt{\bar{w}}\right| \left[\, \frac{1}{4}\left(\sqrt{w}-\sqrt{\bar{w}}\right)^{4} \,+\, w\bar{w} \,+\, \frac{3}{4}\,(w-\bar{w})^{2} \,\right] \\* & = \frac{1}{2} \left|w^{1/2}+\bar{w}^{1/2}\right| \left( w^2 - w^{3/2}\bar{w}^{1/2} + w\bar{w} - w^{1/2}\bar{w}^{3/2} + \bar{w}^{2} \right) \\* & = \pm\frac{1}{2} \left(w^{5/2}+\bar{w}^{5/2}\right) = \left(\Re w^{5/2}\right) \left(\mbox{sign}\:\Re w^{1/2}\right). \end{align*} Hence, we see that $T_{1}$ and $T_{2}$ are indeed the imaginary and real parts of the function $f(w)$ defined by \begin{displaymath} f(w) = \left\{ \begin{array}{rcrcl} w^{5/2} && 0 & \le\;\;\arg w\;\;\le & \pi \\ -\bar{w}^{5/2} && \pi & \le\;\;\arg w\;\;\le & 2\pi \end{array} \right. \end{displaymath} on the region $P_{1}$ or equivalently \begin{displaymath} f(w) = \left\{ \begin{array}{rcrcl} \bar{w}^{5/2} && -\pi & \le\;\;\arg w\;\;\le & 0 \\ w^{5/2} && 0 & \le\;\;\arg w\;\;\le & \pi \\ \end{array} \right. \end{displaymath} on the region $P_{2}$. This function $f(w)$ is holomorphic on the upper half-disc and anti-holomorphic on the lower half-disc, and we note that $f(w)$ is well defined and continuous over the whole unit disc. We now have all that we need to calculate the gauge equivalent connections and Higgs fields $a_{i},\Phi_{i}$ satisfying $a_{2} = g(a_{1})$ and $\Phi_{2} = g(\Phi_{1})$. We first consider the super-potential $\phi_{1}$ defined on the region $0 < \arg w < 2\pi$. Using the notation of \S\ref{gauge-transformation}, on the upper half-disc we have $\phi_{1} = \phi_{+}$ and $f = w^{5/2}$, giving us \begin{displaymath} \del{\bar{w}}\log\phi_{1} = \del{\bar{w}}\log\frac{1-|w|^{5}} {(1-w^{5/2})\,(1-i\,\bar{w}^{5/2})} =\frac{5}{2}\,\frac{i\,\bar{w}^{3/2}}{1-|w|^5}\,\frac{1+i\,w^{5/2}}{1-i\,\bar{w}^{5/2}}. \end{displaymath} On the lower half-disc, we obtain the same expression \begin{displaymath} \del{\bar{w}}\log\phi_{1} = \del{\bar{w}}\log\frac{1-|w|^{5}} {(1+w^{5/2})\,(1-i\,\bar{w}^{5/2})} = \frac{5}{2}\,\frac{i\,\bar{w}^{3/2}}{1-|w|^5}\,\frac{1+i\,w^{5/2}}{1-i\,\bar{w}^{5/2}}, \end{displaymath} although this time we take $\phi_{1}=\phi_{-}$ and $f=-\bar{w}^{5/2}$. Similarly, considering the super-potential $\phi_{2}$ defined for $-\pi < \arg w < \pi$, on the upper half-disc with $\phi_{2} = \phi_{-}$ and $f = w^{5/2}$ we have \begin{displaymath} \del{\bar{w}}\log\phi_{2} = \del{\bar{w}}\log\frac{1-|w|^{5/2}} {(1-\bar{w}^{5/2})\,(1+i\,w^{5/2})} = \frac{5}{2}\,\frac{\bar{w}^{3/2}}{1-|w|^5}\,\frac{1-w^{5/2}}{1-\bar{w}^{5/2}}, \end{displaymath} while on the lower half-disc, we again get the same expression \begin{displaymath} \del{\bar{w}}\log\phi_{2} = \del{\bar{w}}\log\frac{1-|w|^{5/2}} {(1-\bar{w}^{5/2})\,(1-i\,w^{5/2})} = \frac{5}{2}\,\frac{\bar{w}^{3/2}}{1-|w|^5}\,\frac{1-w^{5/2}}{1-\bar{w}^{5/2}}, \end{displaymath} taking $\phi_{2} = \phi_{+}$ and $f = \bar{w}^{5/2}$. Hence, as we predicted, the vortices determined by $\phi_{1}$ and $\phi_{2}$ extend continuously across the negative and positive real axes respectively, even though the gauge transformation $g$ between them does not. For the purposes of our vortex construction, we need not know the super-potentials explicitly; rather, all we require are the complex partial derivatives of their logarithms which we computed above. Nevertheless, here we present the super-potentials as given in \cite{FHP1}. There the two super-potentials take center stage, defined on their own instead of being constructed from the gauge transformation as we have done. Forg\'{a}cs {\em et al.\/} define $\phi_{1}$, $\phi_{2}$ by% \SSfootnote{Actually, \cite{FHP1} uses $+2\,S^{5}\,T_{1}$ and $+2\,S^{5}\,T_{2}$ in the denominators of $\phi_{1}$ and $\phi_{2}$ respectively. The resulting super-potentials are still harmonic, and they do indeed generate gauge equivalent hyperbolic vortices. However, this choice of sign is not consistent with the gauge transformation they use, given here by (\ref{eq:imlog}).} \begin{displaymath} \phi_{1}=\frac{S^{5}-S_{-}^{5}}{S^{5}+S_{-}^{5}-2\,S^{5}\,T_{1}}, \qquad \phi_{2}=\frac{S^{5}-S_{-}^{5}}{S^{5}+S_{-}^{5}-2\,S^{5}\,T_{2}}. \end{displaymath} Simplifying these expressions and writing them using the unit disc model of hyperbolic space, we obtain \begin{align} \label{eq:w-phi1} \phi_{1}&=\frac{1-|w|^{5}}{1\,-\,2\,\Im w^{5/2}\,+\,|w|^5} =\frac{1-|w|^{5}}{\left|1+i\,w^{5/2}\right|^{2}} \\ \label{eq:w-phi2} \phi_{2}&=\frac{1-|w|^{5}}{1\,-\,2\,\Re w^{5/2}\,+\,|w|^5} =\frac{1-|w|^{5}}{\left|1-w^{5/2}\right|^{2}}. \end{align} The reader may want to verify that the partial derivatives $\del{\bar{w}}\log\phi_{1}$ and $\del{\bar{w}}\log\phi_{2}$ agree with those calculated on the previous page. We now compute the Chern class $c_{1}$ of the vortex patched together from the two super-potentials $\phi_{1}$ and $\phi_{2}$. Using equation (\ref{eq:c1-w}), we have \begin{displaymath} c_{1}(a) = \frac{1}{2\pi}\int_{D^{2}}\left( \frac{4}{(1-|w|^{2})^{2}} - \frac{25\,|w|^{3}}{(1-|w|^{5})^{2}} \right) d\mu. \end{displaymath} Taking polar coordinates $r,\theta$ on the disc $D^{2}$, this integral becomes \begin{equation*} c_{1}(a)=\int_{0}^{1}\left( \frac{4\,r}{(1-r^{2})^{2}} - \frac{25\,r^{4}}{(1-r^{5})^{2}} \right) dr =\left[\frac{2}{1-r^{2}} - \frac{5}{1-r^{5}} \right]_{r=0}^{r=1}, \end{equation*} and evaluating this expression by the method used at the end of \S\ref{unit-disc} shows that $c_{1}(a) = 3/2$. We can also arrive at this same result by naively counting the zeros (with multiplicity) of the Higgs field, as both $\del{w}\log\phi_{1}$ and $\del{w}\log\phi_{2}$ vanish to order $3/2$ at the origin but are otherwise nonzero. \subsection{A Family of Singular Vortices} In this section we will generalize the construction of \cite{FHP1} to produce a family of hyperbolic vortices with varying Chern class $c_{1}$. This family includes both the standard $c_{1} = 1$ vortex of \S\ref{unit-disc} and the fractionally charged vortex of the previous section, as well as vortices with arbitrary real $c_{1}$. We will continue to work using the unit disc model of hyperbolic space. \newcommand{\epsilon}{\epsilon} \newcommand{\bar{\epsilon}}{\bar{\epsilon}} Observing the resemblance between the two super-potentials (\ref{eq:w-simple-potential}) and (\ref{eq:w-phi2}), we consider a more general super-potential of the form \begin{equation} \label{eq:phi-c} \phi = \frac{1-|w|^{2c}}{\left|1-w^{c}\right|^{2}}, \end{equation} where $c$ is a nonzero real constant. For non-integral $c$, this $\phi$ is not well defined over the whole unit disc; rather, we must restrict it to a simply-connected cut disc. If we loop around the origin once in the positive direction, crossing our cut in the unit disc, then this super-potential becomes \begin{equation} \label{eq:phi-c-epsilon} \phi' = \frac{1-|w|^{2c}}{\left|1-\epsilon w^{c}\right|^{2}}, \end{equation} introducing a factor of $\epsilon = e^{2\pi ic}$ in the denominator. In both of these cases, we note that $\phi$ and $\phi'$ vanish on on the unit circle, except at the roots of $w^{a} = 1$ (or $w^{a} = \bar{\epsilon}$) where they have simple poles. Taking the logarithmic derivatives of the super-potential $\phi$, we obtain \begin{equation*} \del{w}\log\phi = \frac{c\,w^{c-1}}{1-|w|^{2c}}\, \frac{1-\bar{w}^{c}}{1-w^{c}} \qquad \del{\bar{w}}\log\phi = \frac{c\,\bar{w}^{c-1}}{1-|w|^{2c}}\, \frac{1-w^{c}}{1-\bar{w}^{c}}, \end{equation*} which we can use in equation (\ref{eq:w-dbar}) to construct a hyperbolic vortex $(a,\Phi)$. After looping around the origin, the logarithmic derivatives become \begin{equation*} \del{w}\log\phi' = \frac{c\,\epsilon\,w^{c-1}}{1-|w|^{2c}}\, \frac{1-\bar{\epsilon}\,\bar{w}^{c}}{1-\epsilon\,w^{c}} \qquad \del{\bar{w}}\log\phi' = \frac{c\,\bar{\epsilon}\,\bar{w}^{c-1}}{1-|w|^{2c}}\, \frac{1-\epsilon\,w^{c}}{1-\bar{\epsilon}\,\bar{w}^{c}}, \end{equation*} and we let $(a',\Phi')$ be the corresponding hyperbolic vortex. In order to construct a single hyperbolic vortex over the whole of the unit disc, we would like to find a gauge transformation $g$ taking the vortex $(a,\Phi)$ to the vortex $(a',\Phi')$. Using equation (\ref{eq:w-equivalent-fields}) for gauge equivalent Higgs fields, we have \begin{displaymath} \del{w}\log\phi' = g\,\del{w}\log\phi, \end{displaymath} and so our gauge transformation must be \begin{equation} \label{eq:g} g = \epsilon\,\frac{1-\bar{\epsilon}\,\bar{w}^{c}}{1-\epsilon\,w^{c}}\, \frac{1-w^{c}}{1-\bar{w}^{c}}. \end{equation} We must also verify that this gauge transformation $g$ satisfies equation (\ref{eq:w-equivalent-connections}) for gauge equivalent connections \begin{displaymath} \del{\bar{w}}\log\phi' = \del{\bar{w}}\log\phi - \del{\bar{w}}\log g, \end{displaymath} which we leave to the reader. Hence when we loop around the origin, we obtain a vortex that is gauge equivalent to our original one, and so this vortex is well defined over the punctured disc. Suppose that instead of defining $\epsilon = e^{2\pi ic}$, we use an arbitrary constant $\epsilon$ with $|\epsilon| = 1$ in equations (\ref{eq:phi-c-epsilon}) and (\ref{eq:g}). In this case, the gauge transformation $g$ still maps between the two vortices corresponding to the super-potentials $\phi$ and $\phi'$. Indeed, if we take $c = 5/2$ and $\epsilon = -i$, then the resulting $\phi$, $\phi'$, and $g$ correspond to the super-potentials (\ref{eq:w-phi2}) and (\ref{eq:w-phi1}) and the gauge transformation (\ref{eq:exp-2ichi}) we used in \S\ref{fhp}. Writing out the connection $a$ and Higgs field $\Phi$ of this vortex explicitly using formula (\ref{eq:w-dbar}), we obtain \begin{align*} \bar{\partial}_{a}&= \bar{\partial} + \left( \frac{c\,\bar{w}^{c-1}}{1-|w|^{2c}}\, \frac{1-w^{c}}{1-\bar{w}^{c}} + \frac{1}{1-|w|^{2}}\,\frac{1+w}{1+\bar{w}} \right) d\bar{w} \\ \Phi &= -ic\,w^{c-1}\,\frac{1-|w|^{2}}{1-|w|^{2c}}\, \frac{1+w}{1+\bar{w}}\, \frac{1-\bar{w}^{c}}{1-w^{c}}. \end{align*} The Chern class $c_{1}(a)$ of this vortex is given by equation (\ref{eq:c1-w}), yielding the integral \begin{align*} c_{1}(a) & = \frac{1}{2\pi}\int_{D^{2}}\left( \frac{4}{\left(1-|w|^2\right)^{2}} - \frac{4c^{2}|w|^{2c-2}}{\left(1-|w|^{2c}\right)^{2}} \right) d\mu \\ & = \int_{0}^{1}\left( \frac{4r}{(1-r^{2})^{2}} - \frac{4c^{2}r^{2c-1}}{(1-r^{2c})^{2}} \right) dr = \left[ \frac{2}{1-r^{2}} - \frac{2c}{1-r^{2c}} \right]_{r=0}^{r=1}, \end{align*} where we have used polar coordinates $r,\theta$ on the unit disc. Evaluating this final expression by the method used at the end of \S\ref{unit-disc}, we have $c_{1}(a) = c - 1$. For $c = 2$, this construction yields the standard $c_{1} = 1$ hyperbolic vortex which we discussed in \S\ref{unit-disc}. If we take $c = 5/2$ then we obtain the $c_{1} = 3/2$ vortex given by the super-potential (\ref{eq:w-phi2}) on the cut disc $P_{2}$. Note that with our construction, it is no longer necessary to cover the disc with two overlapping regions as we did in \S\ref{fhp}. Rather, it is sufficient to take a single vortex on the cut disc and then study how it behaves across that cut. For integer values of $c$, the vortex is continuous across the cut, while for other values the vortex changes by a gauge transformation. The flat $c=1$ vortex in this family is \begin{equation*} \bar{\partial}_{a} = \bar{\partial} + \frac{2\,d\bar{w}}{1-\bar{w}^{2}} \qquad \Phi = -i\,\frac{1+w}{1+\bar{w}}\,\frac{1-\bar{w}}{1-w}, \end{equation*} which we readily see has $|\Phi| = 1$ and $F_{a} = 0$, and we therefore have $c_{1} = 0$ as expected. We observe that this vortex is equivalent to the standard flat vortex $(a=0, \Phi=1)$ using the gauge transformation (\ref{eq:g}) with $\epsilon = -1$. To compute the holonomy around loops circling the origin, we introduce polar coordinates $r, \theta$ on the unit disc. In these coordinates, the complex differentials $dw$ and $d\bar{w}$ are \begin{displaymath} dw = \frac{w}{|w|}\,dr + iw\,d\theta, \qquad d\bar{w} = \frac{\bar{w}}{|\bar{w}|} \, dr - i\bar{w}\,d\theta. \end{displaymath} In a small neighborhood of the origin, our connection is approximated by \begin{equation*} a \approx c \left( \bar{w}^{c-1}\,d\bar{w} - w^{c-1}\,dw \right) = -ic \left( w^{c} + \bar{w}^{c} \right) d\theta + \cdots \end{equation*} (only the $d\theta$ term is needed for calculating the holonomy). The contribution to the holonomy around the circle $|w| = r$ due to the connection $a$ is then given by the loop integral \begin{align*} \oint_{|w|=r}a &= \int_{0}^{2\pi}\!-ic r^{c} \left( e^{ic\theta} + e^{-ic\theta} \right) d\theta \\ &=r^{c} \left( -e^{2\pi ic} + e^{-2\pi ic} \right) =-2 r^{c} \sin 2\pi ic. \end{align*} For $c > 0$, this expression vanishes as $r \rightarrow 0$, and thus the limit holonomy around the origin comes entirely from the gauge transformation $g$ given by (\ref{eq:g}). Near the origin, we have $g\approx\epsilon=e^{2\pi ic}$. Hence, the limit of the holonomy around small loops centered at the origin is $e^{2\pi ic}$. By the dimensional reduction technique of Section 2, our family of singular hyperbolic vortices corresponds to a similar family of $\mbox{\rm SO}(3)$-invariant self-dual connections over $S^{4}\setminus S^{2}$. Using the results and quaternionic notation of \S\ref{gauge-transformation}, we see that the holonomy around small loops linking the singular surface $S^{2}$ is $e^{\pi Qc}$. Viewing these connections as $\mbox{\rm SU}(2)$ connections on a bundle $E$, we then obtain a splitting $E=L\oplus L^{\ast}$ on a neighborhood of $S^{2}$ with respect to which the holonomy takes the standard form \begin{displaymath} \left( \begin{array}{cc} e^{2\pi i\alpha} & 0 \\ 0 & e^{-2\pi i\alpha} \end{array} \right) \end{displaymath} for a constant $\alpha$ in the range $[0,1/2)$. In the $\mbox{\rm SO}(3)$-symmetric case, the complex line bundle $L$ on $S^{2}$ has Chern class $c_{1}(L)=-1$. For our family of singular solutions, the holonomy parameter is $\alpha=(c-\lfloor c\rfloor)/2$, where $\lfloor c\rfloor$ is the greatest integer less than or equal to $c$.
{ "timestamp": "2005-03-26T00:10:29", "yymm": "0503", "arxiv_id": "math/0503611", "language": "en", "url": "https://arxiv.org/abs/math/0503611" }
\section{Introduction} Entanglement, the nonclassical correlations between spatially separated particles, is typically a signature of interactions in the past or emergence from a common source. However, it can also arise as the interference of identical particles \cite{Yur92}. By postselecting experimental data based on the ``click'' of detectors \cite{Shi88,Ou88}, photons scattered at a beam splitter have violated a Bell inequality, even if they originated from independent sources \cite{Pit03,Fat04}. In reverse, triggered by an interferometric Bell-state measurement, entanglement has been swapped \cite{Zuk93} to initially uncorrelated photons of different Bell pairs \cite{Pan98a,Pan98b,Jen02}. The observation of these nonclassical interference effects is an important step on the road towards an optical approach of quantum information processing \cite{Kni01,Fra02}. Being furnished by interference, the ability of a beam splitter to entangle the polarizations of two independent photons depends on their indistinguishability \cite{Fey69}. One of the incident photons is horizontally polarized in state $|{\rm H};\psi\rangle$, the other vertically polarized in $|{\rm V};\phi\rangle$. The photons are partially distinguishable by their temporal degrees of freedom captured in the kets $|\psi\rangle$ and $|\phi\rangle$. Besides temporal which-path information inherited from incident photons, a scattered two-photon state possibly holds polarization which-path information. We make no assumptions about the scattering amplitudes connecting polarizations at the beam splitter, except that they constitute a unitary scattering matrix. Translated to a polarization-conserving beam splitter, this corresponds to incident photons in states $|\sigma;\psi\rangle$ and $|\sigma';\phi\rangle$ where $\sigma$, $\sigma'$ are arbitrary superpositions of ${\rm H}$, ${\rm V}$. Our analysis generalizes existing work on a polarization-conserving beam splitter where $\sigma={\rm H}$ and $\sigma'={\rm V}$ \cite{Bos02,Fat04}. The polarization-state $\rho$ of a scattered photon pair is established from the scattering amplitudes of the beam splitter, the shape and timing of photonic wavepackets ($|\psi\rangle$, $|\phi\rangle$) and the time-window of coincidence detection. If not erased by ultra-coincidence detection, an amount of temporal distinguishability of $(1-|\langle \psi|\phi \rangle|^{2})$ pertains corresponding to a mixed state $\rho$. We calculate both its concurrence and the Bell-CHSH parameter. The ability of the latter to witness entanglement can disappear in the presence of a Mandel dip. In terms of a polarization-conserving beam splitter, this corresponds to a deviation of $\sigma$, $\sigma'$ from $\sigma={\rm H}$ and $\sigma'={\rm V}$. \section{Formulation of the problem} In a second-quantized notation, the incident two-photon state $|{\rm H};\psi\rangle_{\rm L}|{\rm V};\phi\rangle_{\rm R}$ takes the form \begin{equation} |{\rm \Psi}_{\rm in}\rangle= {\rm \Psi}_{\rm H,L}^{\dagger}{\rm \Phi}_{\rm V,R}^{\dagger}|0\rangle, \label{Psiin} \end{equation} with field creation operators given by (see Fig. \ref{beamsplitter}) \begin{equation} {\rm \Psi}_{\rm H,L}^{\dagger}=\int d\omega\, a_{\rm H}^{\dagger}(\omega)\psi^{*}(\omega), \quad {\rm \Phi}_{\rm V,R}^{\dagger}=\int d\omega\, b_{\rm V}^{\dagger}(\omega)\phi^{*}(\omega). \end{equation} (The subscripts R,L indicate the two sides of the beam splitter.) The operators $a_{i}(\omega)$ with $i={\rm H},{\rm V}$ satisfy commutation rules \begin{equation} [a_{i}(\omega),a_{j}(\omega')]=0, \quad [a_{i}(\omega),a_{j}^{\dagger}(\omega')]=\delta_{ij}\delta(\omega-\omega'). \end{equation} The same commutation rules hold for the operators $b_{i}(\omega)$, with commutation among $a$ and $b$. The outgoing operators $c_{i}(\omega)$, $d_{i}(\omega)$ are related to the incoming ones $a_{i}(\omega)$, $b_{i}(\omega)$ by a $4 \times 4$ unitary scattering matrix $S$, decomposed in $2 \times 2$ reflection and transmission matrices $r$,$t$,$t'$,$r'$: \begin{equation} \left(\begin{array}{c} c(\omega) \\ d(\omega) \end{array} \right) = \left( \begin{array}{ll} r & t' \\ t & r' \end{array} \right) \left(\begin{array}{c} a(\omega) \\ b(\omega) \end{array} \right), \quad a(\omega) \equiv \left(\begin{array}{c} a_{\rm H}(\omega) \\ a_{\rm V}(\omega) \end{array} \right), \label{inout} \end{equation} and vectors $b(\omega)$, $c(\omega)$, $d(\omega)$ defined similarly. The scattering amplitudes are frequency-independent. \begin{widetext} The outgoing state $|{\rm \Psi}_{\rm out}\rangle$ can be conveniently written in a matrix notation \begin{equation} |{\rm \Psi}_{\rm out}\rangle= \int d\omega \int d\omega' \, \psi^{*}(\omega)\phi^{*}(\omega') \left(\begin{array}{c} c^{\dagger}(\omega) \\ d^{\dagger}(\omega) \end{array} \right)^{\rm T} \left( \begin{array}{ll} r\sigma_{\rm in}t'^{\rm T} & r\sigma_{\rm in}r'^{\rm T} \\ t\sigma_{\rm in}t'^{\rm T} & t\sigma_{\rm in}r'^{\rm T} \end{array} \right) \left(\begin{array}{c} c^{\dagger}(\omega') \\ d^{\dagger}(\omega') \end{array} \right)|0\rangle. \label{psimatout} \end{equation} \end{widetext} Here we used the unitarity of $S$ and $\sigma_{\rm in}=(\sigma_{x}+i\sigma_{y})/2$, with $\sigma_{x}$ and $\sigma_{y}$ Pauli matrices, corresponds to the polarizations of the incoming photons cf. Eq. (\ref{Psiin}). The matrix $\sigma_{\rm in}$ has rank 1 reflecting the fact that polarizations are not entangled prior to scattering. Since we make no assumptions about the scattering amplitudes (apart from the unitarity of $S$), the choice of $\sigma_{\rm in}$ is without loss of generality (see Appendix \ref{stostate}). \begin{figure} \includegraphics[width=8cm]{beamsplitter} \caption{ Schematic drawing of generation and detection of polarization-entanglement at a beam splitter. The independent sources SL and SR each create a photon in modes $\{a\}$ and $\{b\}$ cf. Eq. (\ref{Psiin}). The beam splitter with unitary $4 \times 4$ scattering matrix $S$ couples the polarization of incoming modes to the polarization of outgoing modes $\{c\}$ and $\{d\}$. Polarizations are mixed by $R_{\rm L},R_{\rm R}$ and absorbed by photodetectors DL,DR. A coincidence circuit C registers simultaneous detection of photons. \label{beamsplitter} } \end{figure} The joint probability per unit $\mbox{(time)}^{2}$ of absorbing a photon with polarization $i$ at detector DL and a photon with polarization $j$ at detector DR at times $t$ and $t'$ respectively is given by \cite{Gla63} \begin{equation} w_{ij}(t,t') \propto \langle {\rm \Psi}_{\rm out}|E_{i{\rm L}}^{(-)}(t)E_{j{\rm R}}^{(-)}(t')E_{j{\rm R}}^{(+)}(t')E_{i{\rm L}}^{(+)}(t) |{\rm \Psi}_{\rm out}\rangle, \end{equation} where $E_{i{\rm L}}^{(+)}(t)$ and $E_{i{\rm R}}^{(+)}(t)$ are the positive frequency field operators of polarization $i$ at detectors DL and DR. The probability $C_{ij}(t)$ of a coincidence event within time-windows $\tau$ around $t$ is given by \begin{equation} C_{ij}(t)=\int_{t-\frac{\tau}{2}}^{t+\frac{\tau}{2}} dt'\int_{t-\frac{\tau}{2}}^{t+\frac{\tau}{2}} dt''w_{ij}(t',t''). \label{Corstart} \end{equation} Experimentally, the time-window $\tau$ has typically a lower bound determined by the random rise time of an avalanche of charge carriers in response to a photon absorption event. The polarization-entanglement is detected by violation of the Bell-CHSH inequality \cite{Cla69}. This requires two local polarization mixers $R_{\rm L}$ and $R_{\rm R}$. The Bell-CHSH parameter ${\cal E}$ is \begin{equation} {\cal E}=|E(R_{\rm L},R_{\rm R})+E(R'_{\rm L},R_{\rm R})+E(R_{\rm L},R'_{\rm R})-E(R'_{\rm L},R'_{\rm R})|, \label{Bellpar} \end{equation} where $E(R_{\rm L},R_{\rm R})$ is related to the correlators $C_{ij}(R_{\rm L},R_{\rm R})$ by \begin{equation} E=\frac{C_{\rm HH}+C_{\rm VV}-C_{\rm HV}-C_{\rm VH}}{C_{\rm HH}+C_{\rm VV}+C_{\rm HV}+C_{\rm VH}}. \label{Edef} \end{equation} Substituting the correlators of Eq. (\ref{Corstart}) into Eq. (\ref{Edef}), we see that \begin{equation} E(R_{\rm L},R_{\rm R})={\rm Tr}\, \rho \, (R_{\rm L}^{\dagger}\sigma_{z}R_{\rm L}) \otimes (R_{\rm R}^{\dagger}\sigma_{z}R_{\rm R}), \label{Erho} \end{equation} where $\sigma_{z}$ is a Pauli matrix and $\rho$ a $4 \times 4$ polarization density matrix with elements \begin{widetext} \begin{equation} \rho_{ij,mn} = \frac{1}{\mathcal{N}} \left( (1 + |\alpha|^{2})(\gamma_{1})_{ij}(\gamma_{1})^{*}_{mn}+(1 - |\alpha|^{2}) (\gamma_{2})_{ij}(\gamma_{2})^{*}_{mn} \right). \end{equation} The parameter $\alpha$ is given by \begin{equation} \alpha = \frac{\int_{t-\frac{\tau}{2}}^{t+\frac{\tau}{2}} dt' \int d\omega \int d\omega' \phi(\omega)\psi^{*}(\omega')e^{i(\omega-\omega')t'}} {\sqrt{\left(\int_{t-\frac{\tau}{2}}^{t+\frac{\tau}{2}} dt' \int d\omega \int d\omega' \phi(\omega)\phi^{*}(\omega') e^{i(\omega-\omega')t'}\right) \left(\int_{t-\frac{\tau}{2}}^{t+\frac{\tau}{2}} dt' \int d\omega \int d\omega' \psi(\omega)\psi^{*}(\omega')e^{i(\omega-\omega')t'}\right)}} \end{equation} \end{widetext} and $\gamma_{1}$,$\gamma_{2}$ are $2 \times 2$ matrices related to the scattering amplitudes by \begin{equation} \gamma_{1} = r\sigma_{\rm in}r'^{\rm T}+t'\sigma_{\rm in}^{\rm T}t^{\rm T}, \quad \gamma_{2} = r\sigma_{\rm in}r'^{\rm T}-t'\sigma_{\rm in}^{\rm T}t^{\rm T}. \label{gammas} \end{equation} The normalization factor $\mathcal{N}$ takes the form \begin{equation} \mathcal{N}=(1 + |\alpha|^{2}){\rm Tr}\,\gamma_{1}^{\dagger}\gamma_{1}^{\vphantom{\dagger}}+ (1 - |\alpha|^{2}){\rm Tr}\,\gamma_{2}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}. \label{norm} \end{equation} The parameter $1-|\alpha|^{2} \in (0,1)$ represents the amount of temporal which-path information. Generally, the time-window $\tau$ is much larger than the coherence times or temporal difference of the wavepackets. We may then take the limit $\tau \rightarrow \infty$ and $\alpha$ reduces to the overlap of wavepackets \begin{equation} \alpha=\int d\omega \phi(\omega)\psi^{*}(\omega). \label{overlap} \end{equation} In the opposite limit of ultra-coincidence detection where $\tau \rightarrow 0$, temporal which-path information is completely erased corresponding to $|\alpha|^{2}=1$. \section{Entanglement of formation} \label{EOF} The entanglement of formation of the mixed state $\rho$ is quantified by the concurrence $\mathcal{C}$ \cite{Woo98} given by \begin{equation} \mathcal{C}={\rm max}\left(0,\sqrt{\lambda_{1}}-\sqrt{\lambda_{2}}-\sqrt{\lambda_{3}}-\sqrt{\lambda_{4}}\right). \end{equation} The $\lambda_{i}$'s are the eigenvalues of the matrix product $\rho\tilde{\rho}$, where $\tilde{\rho}=(\sigma_{y}\otimes\sigma_{y})\rho^{*}(\sigma_{y}\otimes\sigma_{y})$, in the order $\lambda_{1} \ge \lambda_{2} \ge \lambda_{3} \ge \lambda_{4}$. The concurrence ranges from 0 (no entanglement) to 1 (maximal entanglement). For simplicity of notation it is convenient to define $(\widehat{xy})_{ij,mn} \equiv x_{ij}y^{*}_{mn}$. The matrix $\tilde{\rho}$ can be written as \begin{equation} \tilde{\rho} = \frac{1}{\mathcal{N}} \left( (1 + |\alpha|^{2})\widehat{\tilde{\gamma}_{1}\tilde{\gamma}_{1}}+ (1 - |\alpha|^{2})\widehat{\tilde{\gamma}_{2}\tilde{\gamma}_{2}} \right), \end{equation} with $\tilde{\gamma} \equiv \sigma_{y}\gamma^{*}\sigma_{y}$. The product $\rho\tilde{\rho}$ takes the simple form \begin{equation} \rho\tilde{\rho}=\frac{{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}}{\mathcal{N}^{2}} \left((1 + |\alpha|^{2})^{2}\widehat{\gamma_{1}\tilde{\gamma}_{1}}- (1 - |\alpha|^{2})^{2}\widehat{\gamma_{2}\tilde{\gamma}_{2}} \right), \end{equation} where we have used the multiplication rule $\widehat{xy}\widehat{vw}=({\rm Tr}\,y^{\dagger}v) \widehat{xw}$ and \begin{equation} {\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}=-{\rm Tr}\,\gamma_{2}^{\dagger} \tilde{\gamma}_{2}^{\vphantom{\dagger}}, \quad {\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{2}^{\vphantom{\dagger}}={\rm Tr}\,\gamma_{2}^{\dagger} \tilde{\gamma}_{1}^{\vphantom{\dagger}}=0. \label{traces1} \end{equation} The results for the tilde inner products of Eq. (\ref{traces1}) hold since the photons are not polarization-entangled prior to scattering (${\rm Det}\,\sigma_{\rm in}=0$). \begin{widetext} The non-Hermitian matrix $\rho\tilde{\rho}$ has eigenvalue-eigenvector decomposition \begin{equation} \rho\tilde{\rho}=\frac{|{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}|^{2}}{\mathcal{N}^{2}} \left(\sum_{i=1,2} \widehat{\gamma_{i}s_{i}}\right) \left( (1 + |\alpha|^{2})^{2}\widehat{s_{1}s_{1}}+ (1 - |\alpha|^{2})^{2}\widehat{s_{2}s_{2}} \right) \left(\sum_{i=1,2} \widehat{\gamma_{i}s_{i}}\right)^{-1}, \label{simtrans} \end{equation} \end{widetext} where we have defined orthonormal states $s_{1}=(1/2)(\openone+\sigma_{z})$ and $s_{2}=(1/2)(\sigma_{x}+i\sigma_{y})$. The pseudo-inverse is easily seen to be \begin{equation} \left(\sum_{i=1,2} \widehat{\gamma_{i}s_{i}}\right)^{-1}= \frac{1}{({\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}})^{*}} \left(\widehat{s_{1}\tilde{\gamma}_{1}}-\widehat{s_{2}\tilde{\gamma}_{2}}\right). \end{equation} It follows that \begin{equation} \mathcal{C}=\frac{2|\alpha|^{2}|{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}|}{\mathcal{N}}. \label{Ctemp} \end{equation} The trace that appears in the numerator of Eq. (\ref{Ctemp}) is given by \begin{equation} |{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}| = 2\sqrt{{\rm Det}\,X^{\dagger}X\,{\rm Det}(\openone-X^{\dagger}X)}, \label{trace2} \end{equation} where we have defined a ``hybrid'' $2 \times 2$ matrix $X$ as \begin{equation} X=\left(\begin{array}{cc} r_{\rm HH} & t'_{\rm HV} \\ r_{\rm VH} & t'_{\rm VV} \end{array}\right). \end{equation} The normalization factor $\mathcal{N}$ given by Eq. (\ref{norm}) can be expressed in terms of $X$ using \begin{equation} {\rm Tr}\,\gamma_{1}^{\dagger}\gamma_{1}^{\vphantom{\dagger}} = {\rm Tr}\,X^{\dagger}X-2\,{\rm Per}\,X^{\dagger}X, \label{trace3} \end{equation} \begin{equation} {\rm Tr}\,\gamma_{2}^{\dagger}\gamma_{2}^{\vphantom{\dagger}} = {\rm Tr}\,X^{\dagger}X-2\,{\rm Det}\,X^{\dagger}X. \label{trace4} \end{equation} (``Per'' denotes the permanent of a matrix.) In the derivation of Eqs. (\ref{trace2},\ref{trace3},\ref{trace4}) we have made use of the unitarity of $S$. The concurrence becomes \begin{equation} \mathcal{C}=\frac{2|\alpha|^{2}\sqrt{{\rm Det}\,X^{\dagger}X\,{\rm Det}(\openone-X^{\dagger}X)}}{ {\rm Tr}\, X^{\dagger}X-(1 + |\alpha|^{2}){\rm Per}\, X^{\dagger}X-(1 - |\alpha|^{2}){\rm Det}\, X^{\dagger}X}. \label{Cfinal} \end{equation} Entanglement depends on the amount of temporal indistinguishability $|\alpha|^{2}$ and the Hermitian matrix \begin{equation} X^{\dagger}X=\left( \begin{array}{cc} |{\mathbf r}_{\rm H}|^{2} & {\mathbf r}_{\rm H}\cdot {\mathbf t}'_{\rm V} \\ ({\mathbf r}_{\rm H}\cdot {\mathbf t}'_{\rm V})^{*} & |{\mathbf t}'_{\rm V}|^{2} \end{array} \right), \end{equation} containing the states ${\mathbf r}_{\rm H}=(r_{\rm HH},r_{\rm VH})$ and ${\mathbf t}'_{\rm V}=(t'_{\rm HV},t'_{\rm VV})$ of a reflected and transmitted photon to the left of the beam splitter. The determinant of $X^{\dagger}X$ measures the size of the span of ${\mathbf r}_{\rm H}$ and ${\mathbf t}'_{\rm V}$ as \begin{equation} {\rm Det}\, X^{\dagger}X = |{\mathbf r}_{\rm H}|^{2} |{\mathbf t}'_{\rm V}|^{2}\left(1-\frac{|{\mathbf r}_{\rm H}\cdot{\mathbf t}'_{\rm V}|^{2}}{ |{\mathbf r}_{\rm H}|^{2} |{\mathbf t}'_{\rm V}|^{2}} \right). \end{equation} If ${\mathbf r}_{\rm H}$ and ${\mathbf t}'_{\rm V}$ are parallel (${\rm Det}X^{\dagger}X=0$), a scattered photon to the left of the beam splitter is in a definite state, giving rise to an unentangled two-photon state ($\mathcal{C}=0$). Similarly, \begin{equation} {\rm Det}(\openone-X^{\dagger}X)= |{\mathbf t}_{\rm H}|^{2} |{\mathbf r}'_{\rm V}|^{2}\left(1-\frac{|{\mathbf t}_{\rm H}\cdot {\mathbf r}'_{\rm V}|^{2}}{ |{\mathbf t}_{\rm H}|^{2} |{\mathbf r}'_{\rm V}|^{2}} \right) \end{equation} involves scattered states ${\mathbf t}_{\rm H}=(t_{\rm HH},t_{\rm VH})$ and ${\mathbf r}'_{\rm V}=(r'_{\rm HV},r'_{\rm VV})$ to the right of the beam splitter. The denominator of Eq. (\ref{Cfinal}) is the probability of finding a scattered state with one photon on either side of the beam splitter. It deviates from its classical value $(X^{\dagger}X)_{\rm HH}+(X^{\dagger}X)_{\rm VV}-2(X^{\dagger}X)_{\rm HH}(X^{\dagger}X)_{\rm VV}$ by an amount $-2|\alpha|^{2}|(X^{\dagger}X)_{\rm HV}|^{2}$ due to photon bunching. This reduction of coincidence count probability is the Mandel dip \cite{Hon87}. It measures the indistinguishability of a reflected and transmitted photon as the product of temporal indistinguishability $|\alpha|^{2}$ and polarization indistinguishability $|(X^{\dagger}X)_{\rm HV}|^{2}$. \section{Violation of the Bell-CHSH inequality} \label{BellCHSH} The maximal value ${\cal E}_{\rm max}$ of the Bell-CHSH parameter (\ref{Bellpar}) for an arbitrary mixed state was analyzed in Refs. \cite{Hor95,Ver02}. For a pure state with concurrence $\mathcal{C}$ one has simply ${\cal E}_{\rm max}=2\sqrt{1+\mathcal{C}^{2}}$ \cite{Gis91}. For a mixed state there is no one-to-one relation between $\mathcal{C}$ and ${\cal E}_{\rm max}$. Depending on the density matrix, ${\cal E}_{\rm max}$ can take on values between $2\mathcal{C}\sqrt{2}$ and $2\sqrt{1+\mathcal{C}^{2}}$. The dependence of ${\cal E}_{\rm max}$ on $\rho$ involves the two largest eigenvalues of the real symmetric $3 \times 3$ matrix $R^{\rm T}R$ constructed from $R_{kl}={\rm Tr}\rho\, \sigma_{k} \otimes \sigma_{l}$, where $\sigma_{1}=\sigma_{x}$,$\sigma_{2}=\sigma_{y}$ and $\sigma_{3}=\sigma_{z}$. In terms of $\gamma_{1}$ and $\gamma_{2}$, the elements $R_{kl}$ take the form \begin{equation} R_{kl}=\frac{(1+|\alpha|^{2})}{\mathcal{N}}{\rm Tr}\,\gamma_{1}^{\dagger}\sigma_{k}\gamma_{1}^{\vphantom{\dagger}} \sigma_{l}^{\rm T}+ \frac{(1-|\alpha|^{2})}{\mathcal{N}}{\rm Tr}\,\gamma_{2}^{\dagger}\sigma_{k}\gamma_{2}^{\vphantom{\dagger}} \sigma_{l}^{\rm T}. \label{Rdef} \end{equation} The matrix $\gamma_{2}$ has a polar decomposition $\gamma_{2}=U\sqrt{\xi}V$ where $U$ and $V$ are unitary matrices and $\xi$ is a diagonal matrix holding the eigenvalues of $\gamma_{2}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}$. The real positive $\xi_{i}$'s are determined by \begin{equation} \xi_{1}+\xi_{2}={\rm Tr}\,\gamma_{2}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}, \quad 2\sqrt{\xi_{1}\xi_{2}}=|{\rm Tr}\,\gamma_{2}^{\dagger}\tilde{\gamma}_{2}^{\vphantom{\dagger}}|. \label{xi} \end{equation} The matrix $\gamma_{1}$ can be conveniently expressed as (see Appendix \ref{semipol}) \begin{equation} \gamma_{1}=UQ\sqrt{\xi}V, \quad \mbox{where} \quad Q=\left( \begin{array}{cc} c_{1} & c_{2} \\ c_{3} & -c_{1} \end{array}\right). \label{gam1} \end{equation} The parameters $c_{1}$,$c_{2}$,$c_{3}$ are real numbers. The matrix $Q$ is traceless due to the orthogonality of $\gamma_{1}$ and $\tilde{\gamma}_{2}$. The number $c_{1} \in (-1,1)$ on the diagonal is related to the inner product of $\gamma_{1}$ and $\gamma_{2}$ and takes the form \begin{equation} c_{1}=\frac{{\rm Tr}\,\gamma_{1}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}}{\xi_{1}-\xi_{2}}, \quad \mbox{with} \quad {\rm Tr}\,\gamma_{1}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}={\rm Tr}\,\sigma_{z}X^{\dagger}X. \label{c1} \end{equation} The numbers $c_{2}$,$c_{3}$ are determined by the norm and tilde inner product of $\gamma_{1}$ and satisfy the relations \begin{equation} c_{1}^{2}+c_{2}c_{3}=1, \quad c_{1}^{2}(\xi_{1}+\xi_{2})+c_{2}^{2}\xi_{2}+c_{3}^{2}\xi_{1}= {\rm Tr}\, \gamma_{1}^{\dagger}\gamma_{1}^{\vphantom{\dagger}}. \label{c2&c3} \end{equation} We substitute $\gamma_{1}$ of Eq. (\ref{gam1}) and the polar decomposition of $\gamma_{2}$ in Eq. (\ref{Rdef}) and parameterize \begin{equation} U^{\dagger}\sigma_{k} U = \sum_{i=1}^{3} N_{ki}\sigma_{i}, \quad V\sigma_{k}^{\rm T} V^{\dagger} = \sum_{i=1}^{3} M_{ki} \sigma_{i}^{\rm T}, \end{equation} in terms of two $3 \times 3$ orthogonal matrices $N$ and $M$. The matrix $R$ takes the form \begin{equation} R=N R' M^{\rm T}, \end{equation} where $R'$ is given by Eq. (\ref{Rdef}) with substitutions $R \rightarrow R'$, $\gamma_{2} \rightarrow \sqrt{\xi}$ and $\gamma_{1} \rightarrow Q\sqrt{\xi}$. With the help of Eqs. (\ref{xi},\ref{c1},\ref{c2&c3}), the eigenvalues $u_{i}$ of $R^{\rm T}R$ can now be expressed as (see Appendix \ref{RtReig}) \begin{equation} u_{1}=\frac{1}{2\mathcal{N}^{2}}\left(\mathcal{T}+\sqrt{\mathcal{T}^{2}-4\mathcal{D}}\right), \label{u1} \end{equation} \begin{equation} u_{2}=\frac{1}{2\mathcal{N}^{2}}\left(\mathcal{T}-\sqrt{\mathcal{T}^{2}-4\mathcal{D}}\right), \label{u2} \end{equation} \begin{equation} u_{3}=4\frac{|\alpha|^{4}|{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}|^{2}}{\mathcal{N}^{2}}, \label{u3} \end{equation} where \begin{widetext} \begin{equation} \mathcal{T}=\mathcal{N}^{2}+4|{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}|^{2} -4(1-|\alpha|^{4})\left({\rm Tr}\,\gamma_{1}^{\dagger}\gamma_{1}^{\vphantom{\dagger}} {\rm Tr}\,\gamma_{2}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}- {\rm Tr}^{2}\,\gamma_{1}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}\right), \label{mathT} \end{equation} \begin{equation} \mathcal{D}=4|{\rm Tr}\,\gamma_{1}^{\dagger}\tilde{\gamma}_{1}^{\vphantom{\dagger}}|^{2}\left(\mathcal{N}^{2}-4(1-|\alpha|^{4}) {\rm Tr}\,\gamma_{1}^{\dagger}\gamma_{1}^{\vphantom{\dagger}} {\rm Tr}\,\gamma_{2}^{\dagger}\gamma_{2}^{\vphantom{\dagger}}\right). \label{mathD} \end{equation} \end{widetext} We can relate the $u_{i}$'s to $X^{\dagger}X$ and $|\alpha|^{2}$ using Eqs. (\ref{norm},\ref{trace2},\ref{trace3},\ref{trace4},\ref{c1}). The parameter ${\cal E}_{\rm max}$ depends on the two largest eigenvalues of $R^{\rm T}R$ as \begin{equation} {\cal E}_{\rm max}=2\sqrt{u_{1}+{\rm max}(u_{2},u_{3})}. \label{Emax} \end{equation} Generically, the expression for ${\cal E}_{\rm max}$ takes a complicated form where ordering of $u_{2}$ and $u_{3}$ depends on $X^{\dagger}X$ and $|\alpha|^{2}$. \section{Discussion} \label{Dis} The objective of the discussion is to reveal the role played by the Mandel dip $-2|\alpha|^{2}|(X^{\dagger}X)_{\rm HV}|^{2}$ in the connection between $\mathcal{C}$ and ${\cal E}_{\rm max}$. We first consider the case $|(X^{\dagger}X)_{\rm HV}|^{2}=0$. The concurrence of Eq. (\ref{Cfinal}) reduces to \begin{equation} \mathcal{C}=\frac{2|\alpha|^{2}\prod_{i={\rm H,V}}\sqrt{(X^{\dagger}X)_{ii}(1-(X^{\dagger}X)_{ii})}}{(X^{\dagger}X)_{\rm HH}+(X^{\dagger}X)_{\rm VV} -2(X^{\dagger}X)_{\rm HH}(X^{\dagger}X)_{\rm VV}}. \end{equation} The maximal value of the Bell-CHSH parameter takes the form \begin{equation} {\cal E}_{\rm max}=2\sqrt{1+\mathcal{C}^{2}} \label{Enomix} \end{equation} and $\mathcal{C} > 0$ implies ${\cal E}_{\rm max} > 2$. \begin{figure} \includegraphics[width=8cm]{regions} \caption{ Parameter space of a beam splitter with $(X^{\dagger}X)_{ii}=1/2$ spanned by $|\alpha|^{2} \in (0,1)$ and $|(X^{\dagger}X)_{\rm HV}|^{2} \in (0,1/4)$. All points correspond to a non-vanishing polarization-entanglement ($\mathcal{C} > 0$) except the line segments $|\alpha|^{2}=0$ and $|(X^{\dagger}X)_{\rm HV}|^{2}=1/4$ where entanglement vanishes ($\mathcal{C}=0$). Only in the shaded region, the Bell-CHSH parameter is able to detect entanglement (${\cal E}_{\rm max} > 2$). The lines correspond to the functions $f$, $g$ of Eqs. (\ref{f},\ref{g}) respectively. \label{regions} } \end{figure} In the presence of a Mandel dip ($|\alpha|^{2}|(X^{\dagger}X)_{\rm HV}|^{2} > 0$), the ability of ${\cal E}$ to witness entanglement can disappear. We consider the special case $(X^{\dagger}X)_{ii}=1/2$. The concurrence of Eq. (\ref{Cfinal}) reduces to \begin{equation} \mathcal{C}=\frac{|\alpha|^{2}\left(1-4|(X^{\dagger}X)_{\rm HV}|^{2}\right)}{1-4|\alpha|^{2}|(X^{\dagger}X)_{\rm HV}|^{2}}. \end{equation} To find ${\cal E}_{\rm max}$ we have to consider the ordering of $u_{2}$ and $u_{3}$ which depends on $|(X^{\dagger}X)_{\rm HV}|^{2}$ and $|\alpha|^{2}$. The function \begin{equation} f(|\alpha|^{2})=\frac{|\alpha|^{2}}{2(1+|\alpha|^{2})} \label{f} \end{equation} divides parameter space in the region $|(X^{\dagger}X)_{\rm HV}|^{2} \le f$ where ${\cal E}_{\rm max}=2\sqrt{u_{1}+u_{3}}$ and the region $|(X^{\dagger}X)_{\rm HV}|^{2} > f$ where ${\cal E}_{\rm max}=2\sqrt{u_{1}+u_{2}}$. The equation ${\cal E}_{\rm max}=2$ has a solution $g(|\alpha|^{2})$ for $|(X^{\dagger}X)_{\rm HV}|^{2}$ that lies in the region $|(X^{\dagger}X)_{\rm HV}|^{2} \le f$. The function $g$ takes the form \begin{equation} g(|\alpha|^{2})=\frac{1}{4}\left(1-|\alpha|^{2}+|\alpha|^{4}-(1-|\alpha|^{2})\sqrt{1+|\alpha|^{4}}\right) \label{g} \end{equation} and breaks parameter space in two fundamental regions: a region $|(X^{\dagger}X)_{\rm HV}|^{2}<g$ where ${\cal E}_{\rm max} > 2$ and a region $|(X^{\dagger}X)_{\rm HV}|^{2}>g$ where ${\cal E}_{\rm max} < 2$. We have drawn these regions in Fig. \ref{regions}. The maximal value of the Bell-CHSH parameter is given by \begin{equation} {\cal E}_{\rm max}=2\mathcal{C}|\alpha|^{-2}\sqrt{1+|\alpha|^{4}} \end{equation} in the region $|(X^{\dagger}X)_{\rm HV}|^{2} \le f$. \section{Conclusions} In summary, we have calculated the amount of polarization-entanglement (concurrence $\mathcal{C}$) and its witness (maximal value of the Bell-CHSH parameter ${\cal E}$) induced by two-photon interference at a lossless beam splitter. The ability of ${\cal E}$ to witness entanglement (${\cal E}_{\rm max} > 2$) depends on the Mandel dip $-2|\alpha|^{2}|(X^{\dagger}X)_{\rm HV}|^{2}$. In the absence of a Mandel dip, $\mathcal{C} > 0$ implies ${\cal E}_{\rm max} > 2$ cf. Eq. (\ref{Enomix}), whereas in its presence this is not necessarily true. In the latter case, as we have demonstrated in Sec. \ref{Dis} with $(X^{\dagger}X)_{ii}=1/2$, the witnessing ability of ${\cal E}$ depends on the individual contributions of temporal ($|\alpha|^{2}$) and polarization indistinguishability ($|(X^{\dagger}X)_{\rm HV}|^{2}$). Our results can be applied to interference of other kinds of particles, getting entangled in some $2 \otimes 2$ Hilbert space and being ``marked'' by an additional degree of freedom. However, determining the indistinguishability parameter $|\alpha|^{2}$ requires careful analysis of the detection scheme. In case of fermions, the matrices $\gamma_{1}$ and $\gamma_{2}$ of Eq. (\ref{gammas}) are to be interchanged. Systems without a time-reversal symmetry are captured by the analysis, as we did not make use of the symmetry of the scattering matrix. \acknowledgments I am grateful to C. W. J. Beenakker for discussions and advice. This work was supported by the Dutch Science Foundation NWO/FOM and by the U.S. Army Research Office (Grant No. DAAD 19-02-0086).
{ "timestamp": "2005-03-25T18:46:05", "yymm": "0503", "arxiv_id": "quant-ph/0503204", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503204" }
\section{Introduction} Numerical calculations of lattice QCD predict a transition from ordinary hadronic matter to a deconfined state of quarks and gluons when the temperature of the system is of the order of $T_{crit}\approx$ 0.17 GeV~\cite{latt}. The existence of such a phase transition manifests itself clearly in the QCD equation-of-state (EoS) on the lattice by a sharp jump of the (Stefan-Boltzmann) scaled energy density, $\varepsilon(T)/T^4$, at the critical temperature, reminiscent of a first-order phase change\footnote{The order of the phase transition itself is not exactly known: the pure SU(3) gauge theory is first-order whereas introduction of 2+1 flavours makes it of a fast cross-over type~\cite{latt}.}. The search for evidences of this deconfined plasma of quarks and gluons (QGP) is the main driving force behind the study of relativistic nuclear collisions at different experimental facilities in the last 20 years. Whereas several experimental results have been found consistent with the formation of the QGP both at CERN-SPS~\cite{sps_qgp} and BNL-RHIC~\cite{rhic_qgp} energies, it is fair to acknowledge that there is no incontrovertible proof yet of bulk deconfinement in the present nucleus-nucleus data. In this paper, we present a detailed study of the only experimental signature, thermal photons, that can likely provide direct information on the {\it thermodynamical} properties (and, thus, on the equation-of-state) of the underlying QCD matter produced in high-energy heavy-ion collisions. Electromagnetic radiation (real and virtual photons) emitted in the course of a heavy-ion reaction, has long~\cite{feinberg,shuryak_photons} been considered a privileged probe of the space-time evolution of the colliding system\footnote{Excellent reviews on photon production in relativistic nuclear collisions have been published recently~\cite{peitz_thoma_physrep,yellow_rep,gale_rep}.}, inasmuch as photons are not distorted by final-state interactions due to their weak interaction with the surrounding medium. Direct photons, defined as real photons not originating from the decay of final hadrons, are emitted at various stages of the reaction with several contributing processes. Two generic mechanisms are usually considered: (i) {\it prompt} (pre-equilibrium or pQCD) photon emission from perturbative parton-parton scatterings in the first tenths of fm/$c$ of the collision process, (ii) subsequent $\gamma$ emission from the {\it thermalized} partonic (QGP) and hadronic (hadron resonance gas, HRG) phases of the reaction.\\ \begin{sloppypar} Experimentally, direct $\gamma$ have been indeed measured in Pb+Pb collisions at CERN-SPS ($\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 17.3 GeV)~\cite{wa98_photons}. However, the relative contributions to the total spectrum of the pQCD, QGP and HRG components have not been determined conclusively. Different hydrodynamics calculations~\cite{srivastava_sps_rhic,alam_sps_rhic,peressou,steffen_sps_rhic_lhc,finnish_hydro} require ``non-conventional'' conditions: high initial temperatures ($T_{0}^{max}>$ $T_{crit}$), strong partonic and/or hadronic transverse velocity flows, or in-medium modifications of hadron masses, in order to reproduce the observed photon spectrum. However, no final conclusion can be drawn from these results due mainly to the uncertainties in the exact amount of radiation coming from primary parton-parton collisions. In a situation akin to that affecting the interpretation of high $p_T$ hadron data at SPS~\cite{dde_sps}, the absence of a concurrent baseline experimental measurement of prompt photon production in p+p collisions at the same $\sqrt{s}$ and $p_T$ range as the nucleus-nucleus data, makes it difficult to have any reliable empirical estimate of the actual thermal $\gamma$ excess in the Pb+Pb spectrum. In the theoretical side, the situation at SPS is not fully under control either: (i) next-to-leading-order (NLO) perturbative calculations are known to underpredict the experimental reference nucleon-nucleon $\gamma$ differential cross-sections below $\sqrt{s}\approx$ 30~GeV~\cite{aurenche} (a substantial amount of parton intrinsic transverse momentum $k_T$~\cite{wong}, approximating the effects of parton Fermi motion and soft gluon radiation, is required~\cite{apanasevich}), (ii) the implementation of the extra nuclear $k_T$ broadening observed in the nuclear data (``Cronin enhancement''~\cite{cronin} resulting from multiple soft and semi-hard interactions of the colliding partons on their way in/out the traversed nucleus) is model-dependent~\cite{dumitru,ina,levai} and introduces an additional uncertainty to the computation of the yields, and (iii) hydrodynamical calculations usually assume initial conditions (longitudinal boost invariance, short thermalization times, zero baryochemical potential) too idealistic for SPS energies. The situation at RHIC (and LHC) collider energies is undoubtedly far more advantageous. Firstly, the photon spectra for different centralities in Au+Au~\cite{ppg042} and in (baseline) p+p~\cite{ppg049} collisions at $\sqrt{s}$ = 200 GeV are already experimentally available. Secondly, the p+p baseline reference is well under control theoretically (NLO calculations do not require extra non-perturbative effects to reproduce the hard spectra at RHIC~\cite{ppg049,ppg024}). Thirdly, the amount of nuclear Cronin enhancement experimentally observed is very modest (high $p_T$ $\pi^0$ are barely enhanced in d+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV~\cite{dAu_phnx}), and one expects even less enhancement for $\gamma$ which, once produced, do not gain any extra $k_T$ in their way out through the nucleus. Last but not least, the produced system at midrapidity in heavy-ion reactions at RHIC top energies is much closer to the zero net baryon density and longitudinally boost-invariant conditions customarily presupposed in the determination of the parametrized photon rates and in the hydrodynamical implementations of the reaction evolution. In addition, the thermalization times usually assumed in the hydrodynamical models ($\tau_{\mbox{\tiny{\it{therm}}}}\lesssim$ 1 fm/$c$) are, for the first time at RHIC, above the lower limit imposed by the transit time of the two colliding nuclei ($\tau_0 = 2R/\gamma\approx$ 0.15 fm/$c$ for Au+Au at 200 GeV). As a matter of fact, it is for the first time at RHIC that hydrodynamics predictions agree {\it quantitatively} with most of the differential observables of bulk (``soft'') hadronic production below $p_T\approx$ 1.5 GeV/$c$ in Au+Au reactions~\cite{kolb_heinz_rep,teaney_hydro,hirano}.\\ \end{sloppypar} \begin{sloppypar} In this context, the purpose of this paper is three-fold. First of all, we present a relativistic Bjorken hydrodynamics model that reproduces well the identified hadron spectra measured at all centralities in Au+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV (and, thus, the centrality dependence of the total charged hadron multiplicity). Secondly, using such a model complemented with the most up-to-date parametrizations of the QGP and HRG photon emission rates, we determine the expected thermal photon yields in Au+Au reactions and compare them to the prompt photon yields computed in NLO perturbative QCD. The combined inclusive (thermal+pQCD) photon spectrum is successfully confronted to recent results from the PHENIX collaboration as well as to other available predictions. Thirdly, after discussing in which $p_T$ range the thermal photon signal can be potentially identified experimentally, we address the issue of how to have access to the thermodynamical properties (temperature, entropy density) of the radiating matter. We propose the correlation of two experimentally measurable quantities: the thermal photon slope and the multiplicity of charged hadrons produced in the reaction, as a direct method to determine the underlying degrees of freedom and the equation of state, $s(T)/T^3$, of the dense and hot QCD medium produced in Au+Au collisions at RHIC energies. \end{sloppypar} \section{Hydrodynamical model} \subsection{Implementation} \begin{sloppypar} Hydrodynamical approaches of particle production in heavy-ion collisions assume {\it local} conservation of energy and momentum in the hot and dense strongly interacting matter produced in the course of the reaction and describe its evolution using the equations of motion of perfect (non-viscous) relativistic hydrodynamics. These equations are nothing but the conservation of: \begin{description} \item (i) the energy-momentum tensor: $\partial_{\mu} T^{\mu\nu} = 0$ with $T^{\mu\nu} = (\varepsilon + p)u^{\mu}u^{\nu}-p\,g^{\mu\nu}$ [where $\varepsilon$, $p$, and $u^{\nu}=(\gamma,\gamma$v) are resp. the energy density, pressure, and collective flow 4-velocity fields, and $g^{\mu\nu}$=diag(1,-1,-1,-1) the metric tensor], and \item (ii) the conserved currents in strong interactions: $\partial_\mu J^{\mu}_{i} = 0$, with $J^{\mu}_{i}=n_{i}u^{\mu}$ [where $n_i$ is the number density of the net baryon, electric charge, net strangeness, etc. currents]. \end{description} \end{sloppypar} These equations complemented with three input ingredients: (i) the initial conditions ($\varepsilon_0$ at time $\tau_0$), (ii) the equation-of-state of the system, $p(\varepsilon,n_{i})$, relating the local thermodynamical quantities, and (ii) the freeze-out conditions, describing the transition from the hydrodynamics regime to the free streaming final particles, are able to reproduce most of the bulk hadronic observables measured in heavy-ion reactions at RHIC~\cite{kolb_heinz_rep,teaney_hydro,hirano}.\\ \begin{sloppypar} The particular hydrodynamics implementation used in this work is discussed in detail in~\cite{peressou}. We assume cylindrical symmetry in the transverse direction ($r$) and longitudinal ($z$) boost-invariant (Bjorken) expansion~\cite{bjorken} which reduces the equations of motion to a one-dimensional problem but results in a loss of the dependence of the observables on longitudinal degrees of freedom. Our results, thus, are only relevant for particle production within a finite range around midrapidity\footnote{The experimental $\pi^\pm$ and $K^\pm$ $dN/dy$ distributions at RHIC are Gaussians~\cite{brahms_hadrons}, as expected from perturbative QCD initial conditions~\cite{eskola_hydro}. Thus, although there is no Bjorken rapidity plateau, the widths of the distributions are quite broad and within $|y|\lesssim$ 2, deviations from boost invariance are not very large~\cite{eskola_hydro}.}. The equation-of-state used here describes a first order phase transition from a QGP to a HRG at $T_{crit}$ = 165 MeV with latent heat\footnote{Although the lattice results seem to indicate that the transition is of a fast cross-over type, the predicted change of $\Delta\varepsilon \approx$ 0.8 GeV/fm$^3$ in a narrow temperature interval of $\Delta T\approx$ 20 MeV~\cite{latt} can be interpreted as the latent heat of the transition.} $\Delta\varepsilon\approx$ 1.4 GeV/fm$^3$, very similar to that used in other works~\cite{kolb_heinz_rep}. The QGP is modeled as an ideal gas of massless quarks ($N_f$ = 2.5 flavours) and gluons with total degeneracy $g_{\mbox{\tiny{\it{QGP}}}} = (g_{\mbox{\tiny{\it{gluons}}}}+7/8\, g_{\mbox{\tiny{\it{quarks}}}})$ = 42.25. The corresponding EoS, $p=1/3\varepsilon-4/3B$ ($B$ being the bag constant), has sound velocity $c_s^2=\partial p/\partial \varepsilon = 1/3$. The hadronic phase is modeled as a non-interacting gas of $\sim$400 known hadrons and hadronic resonances with masses below 2.5 GeV/$c^2$. The inclusion of heavy hadrons leads to an equation of state significantly different than that of an ideal gas of massless pions: the velocity of sound in the HRG phase is $c_s^2\approx$ 0.15, resulting in a relatively soft hadronic EoS as suggested by lattice calculations~\cite{mohanty_cs}; and the effective number of degrees of freedom at $T_{c}$ is $g_{\mbox{\tiny{\it{HRG}}}}\approx$ 12 (as given by $g_{\ensuremath{\it eff}} = 45\,s/(2\pi^2\,T^3)$, see later). Both phases are connected via the standard Gibbs' condition of phase equilibrium, $p_{\mbox{\tiny{\it{QGP}}}}(T_{c}) = p_{\mbox{\tiny{\it{HRG}}}}(T_{c})$, during the mixed phase. The external bag pressure, calculated to fulfill this condition at $T_c$, is $B\approx$ 0.38 GeV/fm$^3$. The system of equations is solved with the MacCormack two-step (predictor-corrector) numerical scheme~\cite{maccormack} with time and radius steps: $\delta t$ = 0.02 fm/$c$ and $\delta r$ = 0.1 fm respectively.\\ \end{sloppypar} Statistical model analyses of particle production in nucleus-nucleus reactions~\cite{pbm_thermal} provide a very good description of the measured particle ratios at RHIC assuming that all hadrons are emitted from a thermalized system reaching chemical equilibrium at a temperature $T_{chem}$ with baryonic, strange and isospin chemical potentials $\mu_{i}$. In agreement with those observations, our specific hydrodynamical evolution reaches chemical freeze-out at $T_{chem}=150$ MeV with $\mu_{B}=25$ MeV (as given by the latest statistical fits to hadron ratios~\cite{andronic05}), and has $\mu_{S}=\mu_{I}=0$. For temperatures above $T_{chem}$ we conserve baryonic, strange and charge currents, but not particle numbers, while for temperatures below $T_{chem}$ we explicitly conserve particle numbers by introducing individual (temperature-dependent) chemical potentials for each hadron. The final differential hadron $dN/dp_T$ spectra are produced via a standard Cooper-Frye ansatz~\cite{cooper_frye} at the kinetic freeze-out temperature ($T_{\ensuremath{\it fo}}=120$ MeV) when the hydrodynamical equations lose their validity, i.e. when the microscopic length (the hadrons mean free path) is no longer small compared to the size of the system. Unstable resonances are then allowed to decay with their appropriate branching ratios~\cite{PDG}. Table I summarizes the most important parameters describing our hydrodynamic evolution. The only free parameters are the initial energy density $\varepsilon_0$ in the center of the reaction zone for head-on (impact parameter $b$ = 0 fm) Au+Au collisions at the starting time $\tau_0$, and the temperature at freeze-out time, $T_{\ensuremath{\it fo}}$. \subsection{Initialization} \begin{sloppypar} We distribute the initial energy density within the reaction volume according to the geometrical Glauber\footnote{The density of participant and colliding nucleons are obtained from the nuclear overlap function $T_{AA}(b)$ computed with a Glauber Monte Carlo code which parametrizes the Au nuclei with Woods-Saxon functions with radius $R$ = 6.38~fm and diffusivity $a$ = 0.54~fm~\cite{hahn}.} prescription proposed by Kolb {\it et al.}~\cite{kolb_heinz_finnishgroup}. Such an ansatz ascribes 75\% of the initial entropy production in a given centrality bin, $s_0(b)$, to soft processes (scaling with the transverse density of participant nucleons $N_{part}(b)$) and the remaining 25\% to hard processes (scaling with the density of point-like collisions, $N_{coll}(b)$, proportional to the nuclear overlap function $T_{AA}(b)$): \begin{equation} s(b) = C\cdot(0.25\cdot N_{part}(b) + 0.75 \cdot N_{coll}(b)), \end{equation} where $C$ is a normalization coefficient chosen so that we produce the correct particle multiplicity at $b$ = 0 fm. For each impact parameter, we construct an azimuthally symmetric hydrodynamical source from the (azimuthally deformed) initial Glauber entropy distribution, by defining a coordinate origin in the middle point between the centers of the two colliding nuclei and averaging the entropy density over all azimuthal directions. We then transform $\varepsilon_0(b)\propto s_0(b)^{4/3}$. This method provides a very good description of the measured centrality dependence of the final charged hadron rapidity densities $dN_{ch}/d\eta$ measured at RHIC as can be seen in Figure~\ref{fig:dNch}. Note that in our implementation of this prescription, we explicitly added the contribution of the particle multiplicity coming from hard processes (i.e. from hadrons having $p_T>$ 1 GeV/$c$) obtained from the scaled pQCD calculations (see later). Such a ``perturbative'' component accounts for a roughly constant $\sim$7\% factor of the total hadron multiplicity for all centralities. The good reproduction of the measured charged hadron integrated yields is an important result for our later use of $dN_{ch}/d\eta|_{\eta=0}$ as an empirical measure of the initial entropy density in different Au+Au centrality classes (see Section~\ref{sec:eos}).\\ \end{sloppypar} \begin{figure}[htbp] \begin{center} \psfig{figure=dNchdeta_vs_Npart.eps,width=9.cm} \end{center} \caption{Charged hadron multiplicity at midrapidity (normalized by the number of participant nucleon pairs) as a function of centrality (given by the number of participants, $N_{part}$) measured in Au+Au at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV by PHENIX~\protect\cite{ppg019} (circles), STAR~\protect\cite{star_Nch} (stars), PHOBOS~\protect\cite{phobos_Nch} (squares) and BRAHMS~\protect\cite{brahms_Nch} (crosses), compared to our hydrodynamics calculations (dashed line), our scaled pQCD ($p_T>$ 1 GeV/$c$) p+p yields~\protect\cite{vogel_hadrons} (dashed-dotted line), and to the sum hydro+pQCD (solid line).} \label{fig:dNch} \end{figure} \begin{table*}[htb] \caption{Summary of the thermodynamical parameters characterizing our hydrodynamical model evolution for central ($b = 0$ fm) Au+Au collisions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV. Input parameters are the (maximum) initial energy density $\varepsilon_0$ (with corresponding ideal-gas entropy densities $s_0$ and temperature $T_0$) at time $\tau_0$, the baryochemical potential $\mu_{B}$, and the chemical and kinetic freeze-out temperatures $T_{chem}$ and $T_{\ensuremath{\it fo}}$ (or energy density $\varepsilon_{\ensuremath{\it fo}}$). The energy densities at the end of the pure QGP ($\varepsilon_{\mbox{\tiny{\it{QGP}}}}^{\mbox{\tiny{\it{min}}}}$), and at the beginning of the pure hadron gas phase ($\varepsilon_{\mbox{\tiny{\it{HRG}}}}^{\mbox{\tiny{\it{max}}}}$) are also given for indication, as well as the average (over total volume) values of the initial energy density $\langle\varepsilon_0\rangle$, entropy density $\langle s_0\rangle$, and temperature $\langle T_0\rangle$.} \begin{center} \begin{tabular}{c|c|c|c|c|c|c|c|c|c} \hline\hline $\tau_0$ & $\varepsilon_0$ ($\langle\varepsilon_0\rangle)$ & $s_0$ ($\langle s_0\rangle$) & $T_0$ ($\langle T_0\rangle$) & $\varepsilon_{\mbox{\tiny{\it{QGP}}}}^{\mbox{\tiny{\it{min}}}}$ & $\varepsilon_{\mbox{\tiny{\it{HRG}}}}^{\mbox{\tiny{\it{max}}}}$ & $\mu_{B}$ & $T_{chem}$ & $T_{\ensuremath{\it fo}}$ & $\varepsilon_{\ensuremath{\it fo}}=\varepsilon_{\mbox{\tiny{\it{HRG}}}}^{\mbox{\tiny{\it{min}}}}$ \\ (fm/$c$) & (GeV/fm$^3$) & (fm$^{-3}$) & (MeV) & (GeV/fm$^3$) & (GeV/fm$^3$) & (MeV) & (MeV) & (MeV) & (GeV/fm$^3$) \\\hline 0.15 & 220 (72) & 498 (190) & 590 (378) & 1.7 & 0.35 & 25. & 150 & 120 & 0.10 \\ \hline\hline \end{tabular} \label{tab:hydro_parameters} \end{center} \end{table*} For the initial conditions (Table~\ref{tab:hydro_parameters}), we choose $\varepsilon_0$ = 220 GeV/fm$^3$ (maximum energy density at $b$ = 0 fm, corresponding to an {\it average} energy density over the total volume for head-on collisions of $\langle\varepsilon_0\rangle$ = 72 GeV/fm$^3$) at a time $\tau_0 = 2R/\gamma\approx$ 0.15 fm/$c$ equal to the transit time of the two Au nuclei at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV. The choice of this relatively short value of $\tau_0$, -- otherwise typically considered in other hydrodynamical studies of thermal photon production at RHIC~\cite{srivastava_sps_rhic,finnish_hydro,frankfurt_rhic_lhc} --, rather than the ``standard'' thermalization time of $\tau_{therm}$ = 0.6 fm/$c$~\cite{kolb_heinz_rep,teaney_hydro,hirano}, is driven by our will to consistently take into account within our space-time evolution the emission of photons from secondary ``cascading'' parton-parton collisions~\cite{bass,bass2} taking place in the {\it thermalizing} phase between prompt pQCD emission (at $\tau\sim 1/p_T\lesssim$0.15 fm/$c$) and full equilibration (see Sect.~\ref{sec:extra_gamma}). Though it may be questionable to identify such photons from second-chance parton-parton collisions as genuine {\it thermal} $\gamma$, it is clear that their spectrum reflects the momentum distribution of the partons during this thermalizing phase\footnote{Note also that it is precisely those secondary partonic interactions that are actually driving the system towards (local) thermal equilibrium.}. Additionally, recent theoretical works~\cite{berges04,arnold04} do seem to support the application of hydrodynamics equations in such ``pre-thermalization'' conditions. Our consequent space-time evolution leads to a value of the energy density of $\varepsilon\approx$ 30 GeV/fm$^3$ at $\tau_{therm}$ = 0.6 fm/$c$, in perfect agreement with other 2D+1 hydrodynamic calculations which do not invoke azimuthal symmetry~\cite{kolb_heinz_rep,teaney_hydro} as well as more numerically involved 3D+1 approaches~\cite{hirano}. Thus, our calculations reproduce the final hadron spectra as well, at least, as those other works do. As a matter of fact, by using $\tau_0$ = 0.15 fm/$c$ (rather than 0.6 fm/$c$), the system has a few more tenths of fm/$c$ to develop some extra transverse collective flow and there is no need to consider in our initial conditions a supplemental input radial flow velocity parameter, $v_{r_0}$, as done in other works~\cite{peressou,kolb_rapp_flow} in order to reproduce the hadron spectra. \subsection{Comparison to hadron data} Figure~\ref{fig:hadron_spectra_AuAu200GeV} shows the pion, kaon, and proton\footnote{For a suitable comparison to the (feed-down corrected) PHENIX~\cite{ppg026}, PHOBOS~\cite{phobos_lowpt_had} and BRAHMS~\cite{brahms_hadrons} yields, the STAR proton spectra~\cite{star_hadrons} have been appropriately corrected for a $\sim$40\% ($p_T$-independent) contribution from weak decays~\cite{star_hadrons2}.} transverse spectra measured by PHENIX~\cite{ppg026}, STAR~\cite{star_hadrons,star_hadrons2}, PHOBOS~\cite{phobos_lowpt_had} and BRAHMS~\cite{brahms_hadrons} in central (0--10\% corresponding to $\mean{b}$ = 2.3 fm) and peripheral (60--70\% corresponding to $\mean{b}$ = 11.9 fm) Au+Au collisions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV, compared to our hydrodynamical predictions (dashed lines) and to properly scaled p+p NLO pQCD expectations~\cite{vogel_hadrons} (dotted lines). At low transverse momentum, the agreement data--hydro is excellent starting from the very low $p_T$ PHOBOS data ($p_T<$ 100 MeV) up to at least $p_T\approx$ 1.5 GeV/$c$. Above this value, contributions from perturbative processes (parton fragmentation products) start to dominate over bulk hydrodynamic production. Indeed, particles with transverse momenta $p_T\gtrsim$ 2 GeV/$c$ are mostly produced in primary parton-parton collisions at times of order $\tau\sim 1/p_T\lesssim$ 0.15 fm/$c$ (i.e. during the interpenetration of the colliding nuclei and {\it before} any sensible time estimate for equilibration), and as such, they are {\it not} in thermal equilibrium with the bulk particle production. Therefore, one does not expect hydrodynamics to reproduce the spectral shapes beyond $p_T\approx$ 2 GeV/$c$. The dotted lines of Fig.~\ref{fig:hadron_spectra_AuAu200GeV} show NLO predictions for $\pi$, $K$ and $p$ production in p+p collisions at $\sqrt{s}$ = 200 GeV~\cite{vogel_hadrons} scaled by the number of point-like collisions ($N_{coll}\propto T_{AA}$) times an empirical quenching factor, $R_{AA}$ = 0.2 (0.7) for 0-10\% central (60-70\% peripheral) Au+Au, to account for the observed constant suppression factor of hadron yields at high $p_T$~\cite{phenix_hiptpi0_200,star_hipt_200} (such a suppression is not actually observed in the $p,\bar{p}$ spectra at intermediate $p_T\approx$ 3 -- 5 GeV/$c$, see discussion below).\\ \begin{figure*}[htbp] \psfig{figure=hadron_spec_AuAu200GeV_cent.eps,width=8cm,height=9cm \psfig{figure=hadron_spec_AuAu200GeV_periph.eps,width=8cm,height=9cm \caption{Transverse momentum spectra for $\pi^{\pm,0}$,$K^{\pm,0}$, and protons measured in the range $p_T$ = 0 -- 5.5 GeV/$c$ by PHENIX~\protect\cite{ppg026}, STAR ($K^0_s$ are preliminary)~\protect\cite{star_hadrons,star_hadrons2}, PHOBOS~\protect\cite{phobos_lowpt_had} and BRAHMS~\protect\cite{brahms_hadrons} in central (0-10\% centrality, left) and peripheral (60-70\%, right) Au+Au collisions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV, compared to our hydrodynamics calculations (dashed lines), to the scaled pQCD p+p rates~\protect\cite{vogel_hadrons} (dotted lines), and to the sum hydro+pQCD (solid lines).} \label{fig:hadron_spectra_AuAu200GeV} \end{figure*} Fig.~\ref{fig:ratios_hadron_spectra_theory} shows more clearly (in linear rather than log scale as the previous figure) the relative agreement between the experimental hadron transverse spectra and the hydrodynamical plus (quenched) pQCD yields presented in this work. The data-over-theory ratio plotted in the figure is obtained by taking the quotient of the pion, kaon and proton data measured in central Au+Au reactions (shown in the left plot of Fig.~\ref{fig:hadron_spectra_AuAu200GeV}) over the corresponding sum of hydrodynamical plus perturbative results (solid lines in Fig.~\ref{fig:hadron_spectra_AuAu200GeV}). In the low $p_T$ range dominated by hydrodynamical production, there exist some local $p_T$-dependent deviations between the measurements and the calculations. However, the same is true within the independent data sets themselves and, thus, those differences are indicative of the amount of systematic uncertainties associated with the different measurements. High $p_T$ hadro-production, dominated by perturbative processes, agrees also well within the $\sim$20\% errors associated with the standard scale uncertainties for pQCD calculations at this center-of-mass energy. It is, thus, clear from Figs.~\ref{fig:hadron_spectra_AuAu200GeV} and~\ref{fig:ratios_hadron_spectra_theory} that identified particle production at $y$ = 0 in nucleus-nucleus collisions at RHIC can be fully described in their whole $p_T$ range and for all centralities by a combination of hydrodynamical (thermal+collective boosted) emission plus (quenched) prompt perturbative production. An exception to this rule are, however, the (anti)protons~\cite{phnx_ppbar}. Although due to their higher masses, they get an extra push from the hydrodynamic flow up to $p_T\sim 3 $ GeV/$c$, for even higher transverse momenta the combination of hydro plus (quenched) pQCD still clearly undershoots the experimental proton spectra. This observation has lent support to the existence of an additional mechanism for baryon production at intermediate $p_T$ values ($p_T\approx$ 3 -- 5 GeV/$c$) based on quark recombination~\cite{recomb}. This mechanism will not, however, be further considered in this paper since it has no practical implication for photon production and/or for the overall hydrodynamical evolution of the reaction. The overall good theoretical reproduction of the differential $\pi,K,p$ experimental spectra for all centralities is obviously consistent with the previous observation that our calculated total integrated hadron multiplicities agree very well with the experimental data measured at mid-rapidity by the four different RHIC experiments (Fig.~\ref{fig:dNch}). \begin{figure*}[htbp] \psfig{figure=ratio_hadron_data_hydro+pQCD_AuAu_cent.eps,height=7.cm} \caption{Ratio of $\pi^{\pm,0}$,$K^{\pm,0}$, and protons yields measured in the range $p_T$ = 0 -- 5.5 GeV/$c$ by PHENIX~\protect\cite{ppg026}, STAR (note that $K^0_s$ are preliminary)~\protect\cite{star_hadrons,star_hadrons2}, PHOBOS~\protect\cite{phobos_lowpt_had} and BRAHMS~\protect\cite{brahms_hadrons} in 0-10\% most central Au+Au collisions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV, over the sum hydro ($\mean{b}$ = 2.3 fm) plus (quenched) pQCD. Theoretical calculations above $p_T\approx$ 2 GeV/$c$ have an overall $\pm$20\% uncertainty (not shown) dominated by pQCD scale uncertainties.} \label{fig:ratios_hadron_spectra_theory} \end{figure*} \section{Direct photon production} As in the case of hadron production, the total direct photon spectrum in a given Au+Au collision at impact parameter $b$ is obtained by adding the primary production from perturbative parton-parton scatterings to the thermal emission rates integrated over the whole space-time volume of the produced fireball. Three sources of direct photons are considered corresponding to each one of the phases of the reaction: prompt production, partonic gas emission, and hadronic gas radiation. \subsection{Prompt photons} \begin{sloppypar} For the prompt $\gamma$ production we use the NLO pQCD predictions of W.~Vogelsang~\cite{vogel_gamma} scaled by the corresponding Glauber nuclear overlap function at $b$, $T_{AA}(b)$, as expected for hard processes in A+A collisions unaffected by final-state effects (as empirically confirmed for photon production in Au+Au~\cite{ppg042}). This pQCD photon spectrum is obtained with CTEQ6M~\cite{cteq6} parton distribution function (PDF), GRV~\cite{grv_photons} parametrization of the $q,g\rightarrow\gamma$ fragmentation function (FF), and renormalization-factorization scales set equal to the transverse momentum of the photon ($\mu = p_T$). Such NLO calculations provide an excellent reproduction of the inclusive direct $\gamma$~\cite{ppg049} and large-$p_T$ $\pi^0$~\cite{ppg024} spectra measured by PHENIX in p+p collisions at $\sqrt{s}$ = 200 GeV without any additional parameter (in particular, at variance with results at lower energies~\cite{wong}, no primordial $k_T$ is needed to describe the data). We do {\it not} consider any modification of the prompt photon yields in Au+Au collisions due to partially counteracting initial-state (IS) effects such as: (i) nuclear modifications (``shadowing'') of the Au PDF ($<20$\%, in the relevant ($x,Q^2$) kinematical range considered here~\cite{frankfurt_rhic_lhc,ina,jamal}), and (ii) extra nuclear $k_T$ broadening (Cronin enhancement) as described e.g. in~\cite{dumitru}. Both IS effects are small and/or approximately cancel each other at mid-rapidity at RHIC as evidenced experimentally by the barely modified nuclear modification factor, $R_{dAu}\lesssim$1.1, for $\gamma$ and $\pi^0$ measured in d+Au collisions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV~\cite{qm05}. Likewise, we do {\it not} take into account any possible final-state (FS) {\it photon} suppression due to energy loss of the jet-fragmentation (aka. ``anomalous'') component of the prompt photon cross-section~\cite{dumitru,jamal,arleo04,dde_hq04}, which, if effectively present (see~\cite{zakharov04} and discussion in Sect.~\ref{sec:extra_gamma}), can be in principle experimentally determined by detailed measurements of the isolated and non-isolated direct photon baseline spectra in p+p collisions at $\sqrt{s}$ = 200 GeV~\cite{dde_hq04}. \end{sloppypar} \subsection{Thermal photon rates} \begin{sloppypar} For the QGP phase we use the most recent full leading order (in $\alpha_{em}$ and $\alpha_s$ couplings) emission rates from Arnold {\it et al.}~\cite{arnold}. These calculations include hard thermal loop diagrams to all orders and Landau-Migdal-Pomeranchuk (LPM) medium interference effects. The parametrization given in~\cite{arnold} assumes zero net baryon density (i.e. null quark chemical potential, $\mu_q$ = 0), and {\it chemical} together with thermal equilibrium. Corrections of the QGP photon rates due to net quark densities are $\mathcal{O}[\mu_q^2/(\pi T)^2]$~\cite{traxler} i.e. marginal at RHIC energies where the baryochemical potential is close to zero at midrapidity ($\mu_B=3\,\mu_q\sim$ 25 MeV) and neglected here. Similarly, although the early partonic phase is certainly not chemically equilibrated (the first instants of the reaction are strongly gluon-dominated) the two main effects from chemical non-equilibrium composition of the QGP: reduction of quark number and increase of the temperature, nearly cancel in the photon spectrum~\cite{yellow_rep,gelis} and have not been considered either. For the HRG phase, we use the latest improved parametrization from Turbide {\it et al.}~\cite{turbide} which includes hadronic emission processes not accounted for before in the literature. In all calculations, we use a temperature-dependent parametrization of the strong coupling\footnote{According to this parametrization, $\alpha_S(T)$ = 0.3 -- 0.6 in the range of temperatures of interest here ($T\approx$ 600 -- 150 MeV).}, $\alpha_s(T) = 2.095/\{\frac{11}{2\pi}\ln{(Q/\Lambda_{\overline{MS}})} + \frac{51}{22\pi}\ln{[2\ln(Q/\Lambda_{\overline{MS}})]}\}$ with $Q = 2\pi T$, obtained from recent lattice results~\cite{karsch_alphaS}. \end{sloppypar} \subsection{Extra photon contributions} \label{sec:extra_gamma} Apart from the aforementioned (prompt and thermal) photon production mechanisms, S.~Bass {\it et al.}~\cite{bass,bass2} have recently evaluated within the Parton Cascade Model (PCM), the contribution to the total Au+Au photon spectrum from secondary (cascading) parton-parton collisions taking place before the attainment of thermalization (i.e. between the transit time of the two nuclei, $\tau \approx$ 0.15 fm/$c$, and the standard $\tau_{therm}$ = 0.6 fm/$c$ considered at RHIC). Since such cascading light emission is due to second-chance partonic collisions which are, simultaneously, driving the system towards equilibrium, we consider not only ``valid'' (see the discussion of refs.~\cite{berges04,arnold04}) but more self-consistent within our framework to account for this contribution with our hydrodynamical evolution alone. We achieve this by starting hydrodynamics (whose photon rates also include the expected LPM reduction of the secondary rates~\cite{bass2}) at $\tau_{0}$ = 0.15 fm/$c$. By doing that, at the same time that we account for this second-chance emission, our initial plasma temperature and associated thermal photon production can be considered to be at their {\it maximum values} for RHIC energies.\\ Likewise, we do not consider the conjectured extra $\gamma$ emission due to the passage of quark jets (Compton-scattering and annihilating) through the dense medium~\cite{fries_rhic_lhc,fries_gammajet2,zakharov04} since such contribution is likely partially compensated by: (i) the concurrent non-Abelian energy loss of the parent quarks going through the system~\cite{turbide05}, plus (ii) a possible {\it photon} suppression due to energy loss of the ``anomalous'' component of the prompt photon cross-section~\cite{dumitru,jamal,dde_hq04,arleo04}. As a matter of fact, some approximate cancellation of all those effects must exist since the experimental Au+Au photon spectra above $p_T\approx$ 4 GeV/$c$ turn out to be well reproduced by primary (pQCD) hard processes alone for all centralities, as can be seen in the comparison of pQCD NLO predictions with PHENIX data~\cite{ppg042} (Fig.~\ref{fig:photon_spec_AuAu_cent_periph}). The apparent agreement between the experimental spectra above $p_T\approx$ 4 GeV/$c$ and the NLO calculations does not seem to leave much room for extra radiation contributions. A definite conclusion on the existence or not of FS effects on photon production will require in any case precision $\gamma$ data in Au+Au, d+Au and p+p collisions. The more critical issue of the role of the jet bremsstrahlung component needs to be estimated, for example, via measurements of isolated and non-isolated direct photon baseline p+p spectra as discussed in~\cite{dde_hq04}. Additional IS effects not considered so far due, for example, to isospin corrections\footnote{Direct photon cross-sections depend on the light quark electric charges and are thus disfavoured in a nucleus target less rich in up quarks than the standard proton reference~\cite{arleo05}.} will require a careful analysis and comparison of Au+Au to reference d+Au photon cross-sections too.\\ \begin{figure*}[htbp] \psfig{figure=photon_spec_AuAu200GeV_cent.eps,height=9.cm} \psfig{figure=photon_spec_AuAu200GeV_periph.eps,height=9.cm} \caption{Photon spectra for central (0--10\%, left) and peripheral (60--70\%, right) Au+Au reactions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV as computed with our hydrodynamical model [with the contributions for the QGP and hadron resonance gas (HRG) given separately] compared to the expected NLO pQCD p+p yields for the prompt $\gamma$~\protect\cite{vogel_gamma} (scaled by the corresponding nuclear overlap function), and to the experimental photon yields measured by the PHENIX collaboration~\protect\cite{ppg042}.} \label{fig:photon_spec_AuAu_cent_periph} \end{figure*} \subsection{Total direct photon spectra} Figure~\ref{fig:photon_spec_AuAu_cent_periph} shows our computed total direct photon spectra for central (left) and peripheral (right) Au+Au collisions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV, with the pQCD, QGP, and HRG components differentiated\footnote{We split the mixed phase contribution onto QGP and HRG components calculating the relative proportion of QGP (HRG) matter in it.}. In central reactions, thermal photon production (mainly of QGP origin) outshines the prompt pQCD emission below $p_T\approx$ 3 GeV/$c$. Within $p_T\approx$ 1 -- 4 GeV/$c$, thermal photons account for roughly 90\% -- 50\% of the total photon yield in central Au+Au, as can be better seen in the ratio total-$\gamma$/pQCD-$\gamma$ shown in Fig.~\ref{fig:RAA_photon}. Photon production in peripheral collisions is, however, clearly dominated by the primary parton-parton radiation. In both cases, hadronic gas emission prevails only for lower $p_T$ values. In Figure~\ref{fig:photon_spec_AuAu_cent_periph} we also compare our computed spectra to the inclusive Au+Au photon spectra published recently by the PHENIX collaboration~\cite{ppg042}. The total theoretical (pQCD+hydro) differential cross-sections are in good agreement with the experimental yields, though for central reactions our calculations tend to ``saturate'' the upper limits of the data in the range below $p_T\approx$ 4 GeV/$c$ where thermal photons dominate. New preliminary PHENIX Au+Au direct-$\gamma^*$ results~\cite{qm05,qm05_akiba} are also systematically above (tough still consistent with) these published spectra in the range $p_T\approx$ 1 -- 4 GeV/$c$ and, if confirmed, will bring our results to an even better agreement with the data.\\ To better distinguish the relative amount of thermal radiation in the theoretical and experimental {\it total} direct photon spectra in central Au+Au collisions, we present in Figure~\ref{fig:RAA_photon} the nuclear modification factor $R_{AA}^\gamma$ defined as the ratio of the total over prompt (i.e. $T_{AA}$-scaled p+p pQCD predictions) photon yields: \begin{equation} R_{AA}^{\gamma}(p_T)\;=\;\frac{dN_{AuAu}^{total\;\gamma}/dp_{T}}{T_{AA}\cdot d\sigma_{pp}^{\gamma\;pQCD}/dp_{T}}. \label{eq:RAA} \end{equation} \begin{figure}[htbp] \psfig{figure=photon_RAA_AuAu200GeV_cent.eps,height=6.cm} \caption{Direct photon ``nuclear modification factor'', $R_{AA}^\gamma$ (Eq.~\ref{eq:RAA}), obtained as the ratio of the total over the prompt $\gamma$ spectra for 0--10\% most central Au+Au reactions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV. The solid line is the ratio resulting from our hydro+pQCD model. The points show the PHENIX data~\protect\cite{ppg042} over the same NLO yields and the dashed-dotted curves indicate the theoretical uncertainty of the NLO calculations (see text).} \label{fig:RAA_photon} \end{figure} A value $R_{AA}^\gamma \approx$ 1 would indicate that all the photon yield can be accounted for by the prompt production alone. Of course, since our total direct-$\gamma$ result for central Au+Au includes thermal emission from the QGP and HRG phases, we theoretically obtain $R_{AA}^\gamma \approx$ 10 -- 1 in the $p_T\approx$ 1 -- 4 GeV/$c$ region where the thermal component is significant (Fig.~\ref{fig:RAA_photon}). In this very same $p_T$ range, although the available PHENIX results have still large uncertainties\footnote{Technically, the PHENIX data points below $p_T$ = 4 GeV/$c$ have ``lower errors that extend to zero'', i.e. a non-zero direct-$\gamma$ signal is indeed observed in the data but the associated errors are larger than the signal itself~\cite{ppg042}.}, the central value of most of the data points is clearly consistent with the existence of a significant excess over the NLO pQCD expectations. A note of caution is worth here, however, regarding the $R_{AA}^\gamma\gg$ 1 value observed for both the theoretical and experimental spectra below $p_T\approx$ 4 GeV/$c$ since it is not yet clear to what extent the NLO predictions, entering in the denominator of Eq.~(\ref{eq:RAA}), are realistic in this thermal-photon ``region of interest''. Indeed, in this comparatively low $p_T$ range the theoretical prompt yields are dominated by the jet bremsstrahlung contribution~\cite{dde_hq04} which is intrinsically non-perturbative (i.e. not computable) and determined solely from the parametrized parton-to-photon GRV~\cite{grv_photons} FF which is relatively poorly known in this kinematic range. The standard scale uncertainties in the NLO pQCD calculations are $\pm$20\% above $p_T\approx$ 4 GeV/$c$ but we have assigned a much more pessimistic $_{+50}^{-200}$\% uncertainty to these calculations in the range $p_T\approx$ 1 -- 4 GeV/$c$ (dashed-dotted lines in Fig.~\ref{fig:RAA_photon}). Precise measurements of the direct-$\gamma$ baseline spectrum in p+p collisions at $\sqrt{s}$ = 200 GeV above $p_T$ = 1 GeV/$c$ are mandatory before any definite conclusion can be drawn on the existence or not of a thermal excess from the Au+Au experimental data.\\ \begin{figure}[htbp] \begin{center} \psfig{figure=photon_spec_AuAu200GeV_cent_thermal_comparison.eps,height=8.5cm} \end{center} \caption{Thermal photon predictions for central Au+Au reactions at $\sqrt{s_{_{NN}}}$ = 200 GeV as computed with different hydrodynamical~\protect\cite{srivastava_sps_rhic,alam_sps_rhic,finnish_hydro} or ``dynamical fireball''~\protect\cite{turbide} models, compared to (i) our hydro calculations (dashed curve), (ii) the expected perturbative $\gamma$ yields ($T_{AA}$-scaled NLO p+p calculations~\protect\cite{vogel_gamma}), and (iii) the experimental total direct photon spectrum measured by PHENIX~\protect\cite{ppg042}.} \label{fig:compare} \end{figure} As a final cross-check of our computed hydrodynamical photon yields, we have compared them to previously published predictions for thermal photon production in Au+Au collisions at top RHIC energy: D.~K.~Srivastava {\it et al.}~\cite{srivastava_sps_rhic} (with initial conditions $\tau_0\approx$ 0.2 fm/$c$ and $T_0\approx$ 450 -- 660 MeV), Jan-e Alam {\it et al.}\footnote{Alam {\it et al.} have recently~\cite{alam05} recomputed their hydrodynamical yields using higher initial temperatures ($T_0$ = 400 MeV at $\tau_0$ = 0.2 fm/$c$) and getting a better agreement with the data.}~\cite{alam_sps_rhic} ($\tau_0$ = 0.5 fm/$c$ and $T_0$ = 300 MeV), F.~D.~Steffen and M.~H.~Thoma~\cite{steffen_sps_rhic_lhc} ($\tau_0$ = 0.5 fm/$c$ and $T_0$ = 300 MeV), S.~S.~Rasanen {\it et al.}~\cite{finnish_hydro} ($\tau_0$ = 0.17 fm/$c$ and $T_0$ = 580 MeV), N.~Hammon {\it et al.}~\cite{frankfurt_rhic_lhc} ($\tau_0$ = 0.12 fm/$c$ and $T_0$ = 533 MeV), and Turbide {\it et al.}\footnote{Note that {\it stricto senso} Turbide's spectra are not obtained with a pure hydrodynamical computation but using a simpler ``dynamical fireball'' model which assumes constant acceleration in longitudinal and transverse directions.}~\cite{turbide} ($\tau_0$ = 0.33 fm/$c$ and $T_0$ = 370 MeV). For similar initial conditions, the computed total thermal yields in those works are compatible within a factor of $\sim$2 with those presented here. Some of those predictions are shown in Figure~\ref{fig:compare} confronted to our calculations. Our yields are, in general, above all other predictions since, as aforementioned, both our initial thermalization time and energy densities (temperatures) have the most ``extreme'' values possible consistent with the RHIC charged hadron multiplicities. They agree specially well with the hydrodynamical calculations of the Jyv\"askyl\"a group~\cite{finnish_hydro} which have been computed with the same up-to-date QGP rates used here. Given the current (large) uncertainties of the available published data, all thermal photon predictions are consistent with the experimental results. However, as aforementioned, newer (preliminary) PHENIX direct-$\gamma^\star$ measurements have been reported very recently~\cite{qm05,qm05_akiba} and indicate a clear excess of direct photons over NLO pQCD for Au +Au at $\sqrt{s_{_{NN}}}$ = 200 GeV in this $p_T$ range in excellent agreement with our thermal photon calculations. \section{Thermal photons and the QCD equation-of-state} \label{sec:eos} In order to experimentally isolate the thermal photon spectrum one needs to subtract from the total direct $\gamma$ spectrum the non-equilibrated ``background'' of prompt photons. The prompt $\gamma$ contribution emitted in a given Au+Au centrality can be measured separately in reference p+p (or d+Au) collisions at the same $\sqrt{s}$, scaled by the corresponding nuclear overlap function $T_{AA}(b)$, and subtracted from the total Au+Au $\gamma$ spectrum~\cite{dde_hq04}. The simpler expectation is that the remaining photon spectrum for a given impact parameter $b$ \begin{equation} \frac{dN_{AuAu}^{thermal\;\gamma}(b)}{dp_T} \; = \; \frac{dN_{AuAu}^{total\;\gamma}(b)}{dp_T} - T_{AA}(b)\cdot\frac{d\sigma_{pp}^{\gamma}}{dp_T}\;, \label{eq:thermal_spec} \end{equation} will be just that due to thermal emission from the partonic and hadronic phases of the reaction. Such a subtraction procedure can be effectively applied to all the $\gamma$ spectra measured in different centralities as long as both the total Au+Au and baseline p+p photon spectra are experimentally measured with reasonable ($\lesssim$15\%) point-to-point (systematical and statistical) uncertainties~\cite{dde_hq04}. The subtracted spectra~(\ref{eq:thermal_spec}) can be therefore subject to scrutiny in terms of the thermodynamical properties of the radiating medium. \subsection{Determination of the initial temperature} \begin{sloppypar} Due to their weak electromagnetic interaction with the surrounding medium, photons produced in the reaction escape freely the interaction region immediately after their production. Thus, even when emitted from an equilibrated source, they are not reabsorbed by the medium and do not have a black-body spectrum at the source temperature. Nonetheless, since all the theoretical thermal $\gamma$ rates~\cite{arnold,turbide} have a general functional dependence of the form\footnote{The $T^2$ factor is just an overall normalization factor in this case (since its temporal variation is small compared to the short emission times) and does not significantly alter the exponential shape of the spectra.} $E_\gamma\,dR_{\gamma}/d^3p\propto T^2\cdot \exp{(-E_{\gamma}/T)}$, one would expect the final spectrum to be locally exponential with an inverse slope parameter strongly correlated with the (local) temperature $T$ of the radiating medium. Obviously, such a general assumption is complicated by several facts. On the one hand, the final thermal photon spectrum is a sum of exponentials with different temperatures resulting from emissions at different time-scales and/or from different regions of the fireball which has strong temperature gradients (the core being much hotter than the ``periphery''). On the other hand, collective flow effects (stronger for increasingly central collisions) superimpose on top of the purely thermal emission leading to an effectively larger inverse slope parameter ($T_{\ensuremath{\it eff}}\approx\sqrt{(1+\beta)/(1-\beta)}\,T$)~\cite{peressou}. One of the main results of this paper is to show that, based upon a realistic hydrodynamical model, such effects do not destroy completely the correlation between the apparent photon temperature and the maximal temperature actually reached at the beginning of the collision process. We will show that such a correlation indeed exists and that the local inverse slope parameter obtained by fitting to an exponential, at high enough $p_T$, the thermal photon spectrum obtained via the expression (\ref{eq:thermal_spec}), indeed provides a good proxy of the initial temperature of the system without much distortion due to collective flow (and other) effects.\\ \end{sloppypar} To determine to what extent the thermal slopes are indicative of the original temperature of the system, we have fitted the thermal spectra obtained from our hydrodynamical calculations in different Au+Au centralities to an exponential distribution in different $p_T$ ranges. Since, -- according to our Glauber prescription for the impact-parameter dependence of the hydrodynamical initial conditions --, different centralities result in different initial energy densities, we can in this way explore the dependence of the apparent thermal photon temperature on the maximal initial temperatures $T_{0}$ (at the core) of the system. The upper plot of Figure~\ref{fig:thermal_slopes} shows the obtained local slope parameter, $T_{\ensuremath{\it eff}}$, as a function of the initial $T_{0}$ for our default QGP+HRG hydrodynamical evolution (Table~\ref{tab:hydro_parameters}). We find that although all the aforementioned effects smear the correlation between the apparent and original temperatures, they do not destroy it completely. The photon slopes are indeed approximately proportional to the initial temperature of the medium, $T_{0}$. There is also an obvious anti-correlation between the $p_T$ of the radiated photons and their time of emission. At high enough $p_T$ the hardest photons issuing from the hottest zone of the system swamp completely any other softer contributions emitted either at later stages and/or from outside the core region of the fireball. Thus, the higher the $p_T$ range, the closer is $T_{\ensuremath{\it eff}}$ to the original $T_{0}$ at the center of the system. According to our calculations, empirical thermal slopes measured above $p_T\approx$ 4 GeV/$c$ in central Au+Au collisions are above $\sim$400 MeV i.e. only $\sim$30\% lower than the ``true'' maximal (local) temperature of the quark-gluon phase. On the other hand, local $\gamma$ slopes in the range below $p_T\approx$ 1 GeV/$c$ have almost constant value $T_{\ensuremath{\it eff}}\sim$ 200 MeV (numerically close to $T_{crit}$) for all centralities and are almost insensitive to the initial temperature of the hydrodynamical system but mainly specified by the exponential prefactors in the hadronic emission rates, plus collective boost effects.\\ \begin{figure}[htbp] \centerline{\psfig{figure=Teff_vs_T0_qgp+hrg.eps,height=6.cm,width=9.5cm}} \centerline{\psfig{figure=Teff_vs_T0_hrg.eps,height=5.cm,width=9.5cm}} \caption{Local photon slope parameters $T_{\ensuremath{\it eff}}$ (obtained from exponential fits of the thermal photon spectrum in different $p_T$ ranges) plotted versus the initial (maximum) temperature $T_{0}$ of the fireball produced at different centralities in Au+Au collisions at $\sqrt{s}=200$ GeV. Upper plot - hydrodynamical calculations with QGP+HRG EoS (Table~\ref{tab:hydro_parameters}), bottom - HRG EoS (with initial conditions: $\varepsilon_0 = 30$ GeV/fm$^3$ at $\tau_0=0.6$~fm/$c$).} \label{fig:thermal_slopes} \end{figure} To assess the dependence of the thermal photon spectra on the underlying EoS, we have rerun our hydro evolution with just the EoS of a hadron resonance gas. We choose now as initial conditions: $\varepsilon_0 = 30$ GeV/fm$^3$ at $\tau_0=0.6$~fm/$c$, which can still reasonably describe the experimental hadron spectra. Obviously, any description in terms of hadronic degrees of freedom at such high energy densities is unrealistic but we are interested in assessing the effect on the thermal photon slopes of a non ideal-gas EoS as e.g. that of a HRG-like system with a large number of heavy resonances (or more generally, of any EoS with exponentially rising number of mass states). The photon slopes for the pure HRG gas EoS (Fig.~\ref{fig:thermal_slopes}, bottom) are lower ($T_{\ensuremath{\it eff}}^{\ensuremath{\it max}}\approx$ 220 MeV) than in the default QGP+HRG evolution, not only because the input HRG $\varepsilon_0$ is smaller (the evolution starts at a later $\tau_0$) but, specially because for the same initial $\varepsilon_0$ the effective number of degrees of freedom in a system with a HRG EoS is higher than that in a QGP\footnote{Note that $g(T)\propto \varepsilon/T^4$ increases exponentially with $T$ for a HRG-like EoS, and at high enough temperatures will clearly overshoot the QGP constant number of degrees of freedom.} and therefore the initial temperatures are lower. A second difference is that, for all $p_T$ ranges, we find almost the same exact correlation between the local $\gamma$ slope and $T_{0}$ indicating a single underlying (hadronic) radiation mechanism dominating the transverse spectra at all $p_T$.\\ Two overall conclusions can be obtained from the study of the hydrodynamical photon slopes. First, the observation in the data, via Eq.~(\ref{eq:thermal_spec}), of a thermal photon excess above $p_T\approx$ 2.5 GeV/$c$ with exponential slope $T_{\ensuremath{\it eff}}\gtrsim$ 250 MeV is an unequivocal proof of the formation of a system with maximum temperatures above $T_{crit}$ since no realistic collective flow mechanism can generate such a strong boost of the photon slopes, while simultaneously reproducing the hadron spectra. Secondly, pronounced $p_T$ dependences of the local thermal slopes seem to be characteristic of space-time evolutions of the reaction that include an ideal-gas QGP radiating phase. \subsection{Determination of the QCD Equation of State (EoS)} As we demonstrated in the previous section, $T_{\ensuremath{\it eff}}$ is approximately proportional to the maximum temperature reached in a nucleus-nucleus reaction. One can go one step further beyond the mere analysis of the thermal photon slopes and try to get a direct handle on the equation of state of the radiating medium by looking at the correlation of $T_{\ensuremath{\it eff}}$ with experimental observables related to the initial energy or entropy densities of the system. For example, assuming an isentropic expansion (which is implicit in our perfect fluid hydrodynamical equations with zero viscosity) one can estimate the {\it initial} entropy density $s$ at the time of photon emission from the total {\it final} particle multiplicity $dN/dy$ measured in the reaction. Varying the centrality of the collision, one can then explore the form of the dependence $s\,=\,s(T)$ at the first instants of the reaction, extract the underlying equation of state of the radiating system and trace any signal of a possible phase transition. Indeed, the two most clear evidences of QGP formation from QCD calculations on the lattice are: (i) the sharp rise of $\varepsilon(T)/T^4$, or equivalently of $s(T)/T^3$, at temperatures around $T_{crit}$, and (ii) the flattening of the same curve above $T_{crit}$. The sharp jump is of course due to the sudden release of a large number of (partonic) degrees of freedom at $T_{crit}$. The subsequent plateau is due to the full formation of a QGP with a {\it fixed} (constant) number of degrees of freedom.\\ We propose here to use $T_{\ensuremath{\it eff}}$ as a proxy for the initial temperature of the system, and directly study the evolution, versus $T_{\ensuremath{\it eff}}$, of the effective number of degrees of freedom defined as\footnote{Units are in GeV and fm. $\zeta(4)$ = $\pi^4/90$, where $\zeta(n)$ is the Riemann zeta function.} \begin{equation} g(s,\,T)\,=\,\frac{\pi^2}{4\,\zeta(4)}\frac{s}{T^3}\,(\hbar c)^3=\,\frac{45}{2 \pi^2}\,\frac{s}{T^3}(\hbar c)^3, \label{eq:ndf} \end{equation} which coincides with the degeneracy of a weakly interacting gas of massless particles. [In a similar avenue, B.~Muller and K.~Rajagopal~\cite{muller_rajagopal05} have recently proposed a method to estimate the number of thermodynamic degrees of freedom via $g_{\ensuremath{\it eff}}\propto s^4/\epsilon^3$, where $s$ is also determined from the final hadron multiplicities]. The dashed line in Fig.~\ref{fig:EoS} (top) shows the evolution of the {\it true} number of degrees of freedom $g_{\ensuremath{\it hydro}}(s_0,T_0)$ computed via Eq.~(\ref{eq:ndf}), as a function of the (maximal) temperatures and entropies directly obtained from the initial conditions of our hydrodynamical model in different Au+Au centralities\footnote{In the most peripheral reactions, the bag entropy has been subtracted to make more apparent the drop near $T_c$.}. The first thing worth to note is that $g(s,T)$ remains constant at the expected degeneracy $g_{\ensuremath{\it hydro}}$ = 42.25 of an ideal gas of $N_f$ = 2.5 quarks and gluons for basically {\it all} the maximum temperatures accessible in the different centralities of Au+Au at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV. This indicates that at top RHIC energies and for most of the impact parameters, $T_0$ is (well) above $T_{crit}$ and the hottest parts of the initial fireball are in the QGP phase. The expected drop in $g_{\ensuremath{\it hydro}}$ related to the transition to the hadronic phase is only seen, if at all, for the very most peripheral reactions (with $T_0\approx T_c$). Thus, direct evidence of the QGP-HRG phase change itself via the study of the centrality dependence of any experimentally accessible observable would only be potentially feasible at RHIC in Au+Au reactions at {\it lower} center-of-mass energies~\cite{dde_dima}.\\ \begin{figure}[htbp] \begin{center} \psfig{figure=ndf_dNchdeta_vs_Teff_qgp+hrg.eps,height=6.cm,width=9.cm} \includegraphic [height=4.5cm,width=9.cm,clip=true,viewport=0 0 567 310]{ndf_dNchdeta_vs_Teff_hrg.eps} \end{center} \caption{Effective initial number of degrees of freedom obtained from our hydrodynamical calculations with a QGP+HRG EoS (upper plot), and with a pure HRG EoS (bottom), plotted as a function of the temperature ($T_{0}$) or thermal photon slope ($T_{\ensuremath{\it eff}}$) in different Au+Au centrality classes at $\sqrt{s_{_{NN}}}$ = 200 GeV. The number of degrees of freedom are computed respectively: (i) From our initial thermodynamical conditions $(s_0,\,T_0)$ via Eq.~(\protect\ref{eq:ndf}) (dashed line), (ii) from the obtained charged hadron multiplicity $dN_{ch}/d\eta$ and the {\it true} initial temperature $T_0$ via Eq.~(\protect\ref{eq:geff}) (dotted-dashed line); and (iii) from $dN_{ch}/d\eta$ and the thermal photon slopes $T_{\ensuremath{\it eff}}$ measured in different $p_T$ ranges via Eq.~(\protect\ref{eq:geff}) (solid lines). For illustrative purposes, the open squares indicate the approximate position of the different Au+Au centrality classes (in 10\% percentiles) for the values of $g_{\ensuremath{\it eff}}$ obtained using the thermal photon slopes measured above $p_T$ = 4 GeV/$c$.} \label{fig:EoS} \end{figure} As aforementioned, we can empirically trace the QCD EoS shown in Fig.~\ref{fig:EoS} (and eventually determine the temperature-evolution of the thermodynamic degrees of freedom of the produced medium) using the estimate of the initial temperature given by the thermal photon slopes, $T_{\ensuremath{\it eff}}$, and a second observable closely related to the initial entropy of the system such as the final-state hadron multiplicity, $dN/dy$. Although one could have also considered to obtain $g_{\ensuremath{\it eff}}$ via $\varepsilon/T^4 \propto (dE_{T}/dy)/T_{\ensuremath{\it eff}}^4$, using the transverse energy per unit rapidity $dE_{T}/dy$ measured in different Au+Au centralities~\cite{ppg019}, we prefer to use the expression (\ref{eq:ndf}) which contains the entropy-, rather than the energy-, density for two reasons: \begin{description} \item (i) the experimentally accessible values of $dN/dy$ remain constant in an isentropic expansion (i.e. $dN/dy \propto s_0$) whereas, due to longitudinal work, the measured final $dE_{T}/dy$ provides only a {\it lower limit} on the initial $\varepsilon$ ($dE_{T}/dy \lesssim \varepsilon_0$); and \item (ii) $g_{\ensuremath{\it eff}}\propto s/T_{\ensuremath{\it eff}}^3$ is less sensitive to experimental uncertainties associated to the measurement of $T_{\ensuremath{\it eff}}$ than $g_{\ensuremath{\it eff}}\propto \varepsilon/T_{\ensuremath{\it eff}}^4$ is. \end{description} \begin{sloppypar} Again, in the absence of dissipative effects, the space-time evolution of the produced system in a nucleus-nucleus reaction is isentropic and the entropy density (per unit rapidity) at the thermalization time $\tau_{0}$ can be directly connected (via $s \,\approx\, 4 \,\rho$~\cite{wong_book}) to the final charged hadron pseudo-rapidity density\footnote{This formula uses $N_{tot}/N_{ch}$ = 3/2, and the Jacobian $|d\eta/dy|=E/p\approx$ 1.2.}: \begin{equation} s\,\approx\,4\cdot\frac{dN}{dV}\,\approx\, \frac{7.2}{\mean{A_\perp}\cdot\tau_0}\cdot\frac{dN_{ch}}{d\eta} \label{eq:entropy} \end{equation} where we have written the volume of the system, $dV=\mean{A_\perp}\tau_0\,d\eta$, as the product of the (purely geometrical) average transverse overlap area for each centrality times the starting proper time of our hydro evolution ($\tau_0=0.15$ fm/$c$), and where $dN_{ch}/d\eta$ is the {\it charged} hadron multiplicity customarily measured experimentally at mid-rapidity\footnote{Note again that both the photon slopes and the charged hadron multiplicities are proxies of the thermodynamical conditions of the system {\it at the same time} $\tau_{0}$.}. By combining, Eqs.~(\ref{eq:ndf}) and (\ref{eq:entropy}), we obtain an estimate for the number of degrees of freedom of the system produced in a given A+A collision at impact parameter $b$: \begin{equation} g_{\ensuremath{\it eff}}\left(\frac{dN_{ch}(b)}{d\eta},\,T_{\ensuremath{\it eff}}(b)\right)\,\approx \,\frac{150}{\pi^2}\cdot\frac{(\hbar c)^3}{\mean{A_\perp(b)}\cdot\tau_0\cdot T^{3}_{\ensuremath{\it eff}}(b)}\cdot \frac{dN_{ch}(b)}{d\eta}\;, \label{eq:geff} \end{equation} which can be entirely determined with two experimental observables: $dN_{ch}/d\eta$ and $T_{\ensuremath{\it eff}}$.\\ \end{sloppypar} Let us first assess to what extent the ansatz~(\ref{eq:geff}) is affected by the assumption that Eq.~(\ref{eq:entropy}) indeed provides a good experimental measure of the initial entropy density $s$. The dotted-dashed curve in Fig.~\ref{fig:EoS} has been obtained via Eq.~(\ref{eq:geff}) using the $(dN_{ch}/d\eta)/\mean{A_\perp}$ values obtained from our hydrodynamical model, and the {\it true} (input) initial temperature of the system $T_0$, and thus it is only sensitive to the way we estimate the entropy density. The resulting curve is a factor of $\sim$3 below the expected ``true'' $g_{\ensuremath{\it hydro}}$ curve, i.e. $g_{\ensuremath{\it eff}}\left(dN_{ch}/d\eta,T_{0}\right)\approx 3\cdot g_{\ensuremath{\it hydro}}(s_0,T_0)$, indicating that Eq.~(\protect\ref{eq:geff}) underestimates by the same amount the maximal entropy of the original medium. This is so because our estimate $(dN_{ch}/d\eta)/\mean{A_\perp}$ specifies the entropy density averaged over the {\it whole} Glauber transverse area $\mean{A_\perp}$, whereas the maximal entropy area in the {\it core} of the system (from where the hardest thermal photons are emitted) is $\sim$3 times {\it smaller}. Although one could think of a method to correct for this difference, this would introduce an extra model-dependence that we want to avoid at this point. We prefer to maintain the simple (geometrical overlap) expression of the transverse area $\mean{A_\perp(b)}$ in Eq.~(\protect\ref{eq:geff}), and exploit the fact that, although such an equation does not provide the true {\it absolute} number of degrees of freedom, it does provide a very reliable indication of the dependence of $g_{\ensuremath{\it eff}}$ on the temperature of the system and, therefore, of the exact {\it form} of the underlying EoS.\\ Finally, let us consider the last case where we use Eq.~(\ref{eq:geff}) with the values of $dN_{ch}/d\eta$ {\it and} $T_{\ensuremath{\it eff}}$ that can be actually experimentally measured. The different solid curves in the upper plot of Fig.~\ref{fig:EoS} show the effective degeneracy $g_{\ensuremath{\it eff}}$, computed using Eq.~(\ref{eq:geff}) and the local photon slopes $T_{\ensuremath{\it eff}}$ measured in different $p_T$ ranges for our default QGP+HRG evolution. As one could expect from Fig.~\ref{fig:thermal_slopes}, the best reproduction of the shape of the underlying EoS is obtained with the effective temperatures measured in higher $p_T$ bins. For those $T_{\ensuremath{\it eff}}$, the computed $g_{\ensuremath{\it eff}}$'s show a relatively constant value in a wide range of centralities as expected for a weakly interacting QGP. Deviations from this ideal-gas plateau appear for more central collisions, due to an increasing difference between the (high) initial temperatures, $T_0$, and the apparent temperature given by the photon slopes (Fig.~\ref{fig:thermal_slopes}). Such deviations do not spoil, however, the usefulness of our estimate since, a non-QGP EoS would result in a considerably different dependence of $g_{\ensuremath{\it eff}}$ on the reaction centrality. Indeed, the different curves in the bottom plot of Fig.~\ref{fig:EoS} obtained with a pure hadron resonance gas EoS clearly indicate\footnote{Accidentally, $g_{\ensuremath{\it eff}}\gtrsim g_{\ensuremath{\it hydro}}$ in the case of a HRG EoS, because the underestimation of the apparent temperature (raised to the cube) compensates for the aforementioned area averaging of the entropy.} that a HRG EoS, or in general any EoS with exponentially increasing number of mass states, would bring about a much more dramatic rise of $g_{\ensuremath{\it eff}}$ with $T_{\ensuremath{\it eff}}$.\\ In summary, the estimate~(\ref{eq:geff}) indeed provides a direct experimental handle on the {\it form} of the EoS of the strongly interacting medium produced in the first instants of high-energy nuclear collisions. More quantitative conclusions on the possibility to extract the exact shape of the underlying EoS and/or the absolute number of degrees of freedom of the produced medium require more detailed theoretical studies (e.g. with varying lattice-inspired EoS's~\cite{dde_dima} and/or using more numerically involved 3D+1 hydrodynamical approaches). In any case, we are confident that by experimentally measuring the thermal photon slopes in different Au+Au centralities and correlating them with the associated charged hadron multiplicities as in Eq.~(\ref{eq:geff}), one can approximately observe the expected ``plateau'' in the number of degrees of freedom indicative of QGP formation above a critical value of $T$. \section{Conclusions} We have studied thermal photon production in Au+Au reactions at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV using a Bjorken hydrodynamic model with longitudinal boost invariance. We choose the initial conditions of the hydrodynamical evolution so as to efficiently reproduce the observed particle multiplicity in central Au+Au collisions at RHIC and use a simple Glauber prescription to obtain the corresponding initial conditions for all other centralities. With such a model we can perfectly reproduce the identified soft pion, kaon and proton $p_T$-differential spectra measured at RHIC. Complementing our model with the most up-to-date parametrizations of the QGP and HRG thermal photon emission rates plus a NLO pQCD calculation of the prompt $\gamma$ contribution, we obtain direct photon spectra which are in very good agreement with the Au+Au direct photon (upper limit) yields measured by the PHENIX experiment. In central collisions, a thermal photon signal should be identifiable as a factor of $\sim$8 -- 1 excess over the pQCD $\gamma$ component within $p_T\approx$ 1 -- 4 GeV/$c$, whereas pure prompt emission clearly dominates the photon spectra at all $p_T$ in peripheral reactions. The local inverse slope parameter of the thermal photon spectrum is found to be directly correlated to the maximum temperature attained in the course of the collision. The experimental measurement of local thermal photon slopes above $p_T\approx$ 2.5 GeV/$c$, with values $T_{\ensuremath{\it eff}}\gtrsim$ 250 MeV and with pronounced $p_T$ dependences can only be reproduced by space-time evolutions of the reaction that include a QGP phase.\\ Finally, we have proposed and tested within our framework, an empirical method to determine the effective thermodynamical number of degrees of freedom of the produced medium, $g(s,T)\propto s(T)/T^3$, by correlating the thermal photon slopes with the final-state charged hadron multiplicity measured in different centrality classes. We found that one can clearly distinguish between the equation of state of a weakly interacting quark-gluon plasma and that of a system with rapidly rising number of mass states with $T$. Stronger quantitative conclusions on the exact shape of the underlying EoS and/or the absolute number of degrees of freedom of the produced medium require more detailed theoretical studies as well as high precision photon data in Au+Au and baseline p+p, d+Au collisions. In any case, the requirement for hydrodynamical models of concurrently describing the experimental bulk hadron and thermal photon spectra for different Au+Au centralities at $\sqrt{s_{\mbox{\tiny{\it{NN}}}}}$ = 200 GeV, imposes very strict constraints on the form of the equation of state of the underlying expanding QCD matter produced in these reactions. \section{Acknowledgments} We would like to thank Werner Vogelsang for providing us with his NLO pQCD calculations for photon production in p+p collisions at $\sqrt{s}$ = 200 GeV; Sami Rasanen for valuable comments on hydrodynamical photon production; and Helen Caines and Olga Barannikova for useful discussions on (preliminary) STAR hadron data. D.P. acknowledges support from MPN of Russian Federation under grant NS-1885.2003.2.
{ "timestamp": "2006-02-09T13:01:30", "yymm": "0503", "arxiv_id": "nucl-th/0503054", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503054" }
\subsection*{Notation} We use the standard notation for the coproduct-insertion maps: we say that an ordered set is a pair of a finite set $S$ and a bijection $\{1,\ldots,|S|\} \to S$. For $I_1,\dots,I_m$ disjoint ordered subsets of $\{1,\dots,n\}$, $(U,\Delta)$ a Hopf algebra and $a \in U^{\otimes m}$, we define $$a^{I_1,\dots,I_n}= \sigma_{I_1,\ldots,I_m} \circ (\Delta^{|I_1|}\otimes \cdots \otimes \Delta^{|I_n|})(a), $$ with $\Delta^{(1)}=\on{id}$, $\Delta^{(2)}=\Delta$, $\Delta^{(n+1)}=({\on{id}}^{\otimes n-1} \otimes \Delta)\circ \Delta^{(n)}$, and $\sigma_{I_1,\ldots,I_m} : U^{\otimes \sum_i |I_i|} \to U^{\otimes n}$ is the morphism corresponding to the map $\{1,\ldots,\sum_i |I_i|\} \to \{1,\ldots,n\}$ taking $(1,\ldots,|I_1|)$ to $I_1$, $(|I_1| + 1,\ldots,|I_1| + |I_2|)$ to $I_2$, etc. When $U$ is cocommutative, this definition depends only on the sets underlying $I_1,\ldots,I_m$. \subsection*{Acknowledgements} We would like to thank V. Dolgushev, P. Etingof and L.-C. Li for discussions. \section{Solutions of the functional twist equations} If ${\mathfrak{g}}$ is a Lie algebra, we denote by ${\mathcal O}_{{\mathfrak{g}}^*} = \wh S({\mathfrak{g}})$ the formal series ring of functions on the formal neighborhood of $0$ in ${\mathfrak{g}}^*$. We define by ${\mathfrak{m}}_{{\mathfrak{g}}^*} \subset {\mathcal O}_{{\mathfrak{g}}^*}$ the maximal ideal of this ring. If $k$ is an integer $\geq 1$, we denote by ${\mathcal O}_{({\mathfrak{g}}^*)^k} = \wh S({\mathfrak{g}})^{\wh\otimes k}$ the \footnote{$\wh\otimes$ is the completed tensor product, defined by $V_0[[x_1,\ldots,x_n]] \wh\otimes W_0[[y_1,\ldots,y_n]] := V_0\otimes W_0[[x_1,\ldots,y_n]]$, where $V_0,W_0$ are vector spaces.} ring of formal functions functions on $({\mathfrak{g}}^*)^k$, by ${\mathfrak{m}}_{({\mathfrak{g}}^*)^k}$ its maximal ideal and by ${\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^i$ the $i$th power of this ideal. If $f,g\in {\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^2$, then the series $f \star g = f + g + {1\over 2} \{f,g\} + \cdots + B_n(f,g) + \cdots$ is convergent, where $\sum_{i\geq 1} B_i(x,y)$ is the Baker-Campbell-Hausdorff series specialized to the Poisson bracket of ${\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^2$. The product $\star$ defines a group structure on ${\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^2$. If $f\in {\mathcal O}_{{\mathfrak{g}}^*}^{\wh\otimes n}$ and $P_1,\dots, P_m$ are disjoint subsets of $\{1,\dots,m\}$, one defines $f^{P_1,\dots,P_n}$ as in the Introduction using the cocommutative coproduct of ${\mathcal O}_{{\mathfrak{g}}^*}$ (dual to the addition of ${\mathfrak{g}}^*$). \medskip Let ${\mathfrak{g}}$ be a Lie algebra and $Z\in \wedge^3({\mathfrak{g}})^{\mathfrak{g}}$. \begin{proposition} \label{lift1} There exists $\varphi\in ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3})^{\mathfrak{g}} (\subset {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^2)$, the image of which under the map ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} \to ({\mathfrak{m}}_{{\mathfrak{g}}^*} / {\mathfrak{m}}_{{\mathfrak{g}}^*}^2)^{\otimes 3} = {\mathfrak{g}}^{\otimes 3} \stackrel{\on{Alt}}{\to} \wedge^3({\mathfrak{g}})$ equals $Z$ (here $\on{Alt}$ is the total antisymmetrization map) and satisfying the functional pentagon equation $$ \varphi^{1,2,34} \star \varphi^{12,3,4} = \varphi^{2,3,4} \star \varphi^{1,23,4} \star \varphi^{1,2,3}. $$ Such a $\varphi$ (we call it a lift of $Z$) is unique up to the action of an element of $({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2})^{\mathfrak{g}}$ by $\sigma \cdot \varphi = \sigma^{2,3} \star \sigma^{1,23} \star \varphi \star (-\sigma)^{12,3} \star (-\sigma)^{1,2}$. \end{proposition} {\em Proof.} In \cite{Dr:QH}, Proposition 3.10, Drinfeld constructed a solution $\Phi\in U({\mathfrak{g}})^{\otimes 3}[[\hbar]]$ of the pentagon equation \begin{equation} \label{pent} \Phi^{1,2,34} \Phi^{12,3,4} = \Phi^{2,3,4} \Phi^{1,23,4} \Phi^{1,2,3} \end{equation} such that $\varepsilon^{(2)}(\Phi)=1$ and $\Phi = 1^{\otimes 3} + O(\hbar)$ (here $\varepsilon^{(2)} = \id\otimes \varepsilon\otimes\id$; applying $\varepsilon$ to the first and third factors of (\ref{pent}), we also get $\varepsilon^{(1)}(\Phi) = \varepsilon^{(3)}(\Phi)=1$). In \cite{EH2}, we stated that $\Phi$ can be transformed into an admissible solution $\Phi'$ of the same equations, using an invariant twist. In Appendix \ref{app:A}, we explain why the proof given in \cite{EH2} is wrong and we give a correct proof. The classical limit of $\hbar\log(\Phi')$ then satisfies the functional pentagon equation. This gives the existence of $\varphi$. One can also construct $\varphi$ directly using cohomological methods, as it will be done for $\rho$ later. Let us prove uniqueness: let $\varphi$ and $\varphi'$ be two lifts of $Z$. The classes of $\varphi$ and $\varphi'$ are the same in ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}/({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3} ^3)$, as this space is $0$. Let $N$ be an integer $\geq 3$; assume that we have found $\sigma_N \in ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2})^{\mathfrak{g}}$ such that $\sigma_N \cdot \varphi$ and $\varphi'$ are equal modulo ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^N$. Write $\varphi'=\sigma_N \cdot \varphi+\psi$, with $\psi \in ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^N)^{\mathfrak{g}}$. We will use the following lemma (see \cite{EGH}, p. 2477): \begin{lemma} \label{lemmetech} For any $k\geq 1$ and $n \geq 2$, $f,h \in {\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^2$ and $g \in {\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^n$, one has $$ f \star (h + g) = f \star h + g, \quad (f+g)\star h = f \star h +g \hbox{\ modulo\ }{\mathfrak{m}}_{({\mathfrak{g}}^*)^k}^{n+1}. $$ \end{lemma} Let $\overline\psi$ be the class of $\psi$ in $({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap{\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^N)^{\mathfrak{g}} / ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^{N+1})^{\mathfrak{g}} = (S^{>0}({\mathfrak{g}})^{\otimes 3})^{\mathfrak{g}}_N$. Then $\overline\psi^{1,2,34} +\overline\psi^{12,3,4} = \overline\psi^{2,3,4} + \overline\psi^{1,23,4} + \overline\psi^{1,2,3}$, which means that $\overline\psi$ is a cocycle in the subcomplex $((S^{>0}({\mathfrak{g}})^{\otimes \cdot})^{\mathfrak{g}},d)$ of the co-Hochschild\footnote{We denote by $S({\mathfrak{g}})$ the symmetric algebra of ${\mathfrak{g}}$, by $S^{>0}({\mathfrak{g}})$ is positive degree part; the index $N$ means the part of total degree $N$.} complex $(S({\mathfrak{g}})^{\otimes \cdot},d)$. Using \cite{Dr:QH}, Proposition 3.11, one can prove that the $k$th cohomology group of this complex is $\wedge^k({\mathfrak{g}})^{\mathfrak{g}}$ and that the antisymmetrization map coincides with the canonical map from the space of cocycles to the cohomology. For $N=3$, the hypothesis implies that $\on{Alt}(\overline\psi) = 0$, so $\overline\psi$ is a coboundary of an element $\overline\tau_3 \in (S^{>0}({\mathfrak{g}})^{\otimes 2})^{\mathfrak{g}}_3$. For $N >3$, $\overline\psi$ is the a coboundary of an element $\overline\tau_N\in (S^{>0}({\mathfrak{g}})^{\otimes 2})^{\mathfrak{g}}_N$, since the degree $N$ part of the relevant cohomology group vanishes. We then set $\sigma_{N+1} = \sigma_N + \tau_N$, where $\tau_N \in ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^2}^N)^{\mathfrak{g}}$ is a lift of $\overline\tau_N$. Then $\sigma_{N+1} \cdot \varphi$ and $\varphi'$ are equal modulo ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^{N+1}$. The sequence $(\sigma_N)_{N\geq 3}$ has a limit $\sigma$. Then $\sigma \cdot \varphi = \varphi'$. \hfill \qed \medskip We now construct a lift of $r$: \begin{theorem} \label{lift2} There exists $\rho\in {\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2}$, the image of which in ${\mathfrak{g}}^{\otimes 2}$ under the square of the projection ${\mathfrak{m}}_{{\mathfrak{g}}^*} \to {\mathfrak{m}}_{{\mathfrak{g}}^*} / {\mathfrak{m}}_{{\mathfrak{g}}^*}^2 = {\mathfrak{g}}$ equals $r$, and such that \begin{equation} \label{twistequation} \rho^{1,2}\star \rho^{12,3} = \rho^{2,3} \star \rho^{1,23} \star \varphi. \end{equation} Such a $\rho$ (we call it a lift of $r$) is unique up to the action of ${\mathfrak{m}}_{{\mathfrak{g}}^*}$ by $\lambda \cdot \rho = \lambda^{1} \star \lambda^{2} \star \rho \star (-\lambda)^{12}$. We call equation (\ref{twistequation}) the functional cocycle equation. \end{theorem} {\em Proof.} Let us construct $\rho$ by induction: we will construct a convergent sequence $\rho_N \in {\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2}$ ($N\geq 2$) satisfying (\ref{twistequation}) in ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} / ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^N)$. When $N = 3$, we take for $\rho_2$ any lift of $r$ to ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2}$; then equation (\ref{twistequation}) is automatically satisfied. Let $N$ be an integer $\geq 3$; assume that we have constructed $\rho_N$ in ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2}$ satisfying equation (\ref{twistequation}) in ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}/ ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap{\mathfrak{m}}_{({\mathfrak{g}}^*)^3} ^N)$. Set $\alpha_N := \rho_N^{1,2}\star \rho_N^{12,3} -\rho_N^{2,3} \star \rho_N^{1,23} \star \varphi$. Then $\alpha_N$ belongs to ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap{\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^N$, and the following equalities hold in ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 4}/({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 4} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^4}^{N+1})$: \begin{align*} \alpha_N^{12,3,4} = &~ \rho_N^{1,2} \star \alpha_N^{12,3,4}= \rho_N^{1,2} \star \rho_N^{12,3}\star \rho_N^{123,4} - \rho_N^{1,2} \star \rho_N^{3,4} \star \rho_N^{12,34} \star \varphi^{12,3,4} \\ & ~(\hbox{using Lemma \ref{lemmetech}}) \\ =&~(\alpha_N^{1,2,3}+ \rho_N^{2,3} \star \rho_N^{1,23}\star \varphi^{1,2,3}) \star \rho_N^{123,4}-\rho_N^{3,4}\star \rho_N^{1,2}\star \rho_N^{12,34} \star \varphi^{12,3,4} \\ =&~\alpha_N^{1,2,3} + \rho_N^{2,3} \star \rho_N^{1,23} \star \rho_N^{123,4}\star \varphi^{1,2,3} -\rho_N^{3,4}\star (\rho_N^{2,34}\star \rho_N^{1,234} \star \varphi^{1,2,34} +\alpha_N^{1,2,34}) \star \varphi^{12,3,4} \cr &~(\hbox{using Lemma \ref{lemmetech}, the invariance of }\varphi\hbox{ and the definition of }\alpha_N^{1,2,34})\cr =&~\alpha_N^{1,2,3} + \rho_N^{2,3} \star (\alpha_N^{1,23,4}+ \rho_N^{23,4} \star \rho_N^{1,234} \star \varphi^{1,23,4})\star \varphi^{1,2,3}\cr &- \alpha_N^{1,2,34}-\rho_N^{3,4}\star \rho_N^{2,34}\star \rho_N^{1,234} \star \varphi^{1,2,34} \star \varphi^{12,3,4} \cr &~(\hbox{using the definition of }\alpha_N^{1,23,4} \hbox{ and Lemma \ref{lemmetech}})\cr =&~\alpha_N^{1,2,3} + \alpha_N^{1,23,4}+ (\rho_N^{3,4} \star \rho_N^{2,34}\star \varphi^{2,3,4}+ \alpha_N^{2,3,4} )\star \rho_N^{1,234} \star \varphi^{1,23,4}\star \varphi^{1,2,3}\cr &- \alpha_N^{1,2,34}-\rho_N^{3,4}\star \rho_N^{2,34}\star \rho_N^{1,234} \star \varphi^{1,2,34} \star \varphi^{12,3,4} \cr &~ (\hbox{using the definition of }\alpha_N^{2,3,4} \hbox{ and Lemma \ref{lemmetech}}) \cr &~ = \alpha_N^{1,2,3} + \alpha_N^{1,23,4} - \alpha_N^{1,2,34} + \alpha_N^{2,3,4} \end{align*} (using Lemma \ref{lemmetech}, the invariance of $\varphi$ and the fact that $\varphi$ satisfies the functional pentagon equation). Let us denote by $\overline\alpha_N$ the image of $\alpha_N$ in $({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap{\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^N) / ({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}\cap{\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^{N+1}) = (S^{>0}({\mathfrak{g}})^{\otimes 3})_N$, then we get $$ \overline\alpha_N^{12,3,4} + \overline\alpha_N^{1,2,34} =\overline\alpha_N^{1,2,3} + \overline\alpha_N^{1,23,4} +\overline\alpha_N^{2,3,4}. $$ This means that $\alpha$ is a cocycle for the subcomplex $(S^{>0}({\mathfrak{g}})^{\otimes\cdot},d)$ of the co-Hochschild complex. Using \cite{Dr:QH}, Proposition 3.11, one proves that the $k$th cohomology group of this subcomplex is $\wedge^k({\mathfrak{g}})$, and that antisymmetrization coincides with the canonical projection from the space of cocycles to the cohomology group. For $N=3$, the equation $\on{CYB}(r)=Z$ implies $\on{Alt}(\overline\alpha_3)=0$, hence $\overline\alpha_3$ is the coboundary of an element $\overline\beta_3\in (S^{>0}({\mathfrak{g}})^{\otimes 2})_3$. For $N>3$, $\overline\alpha_N$ is the coboundary of an element $\overline\beta_N \in (S^{>0}({\mathfrak{g}})^{\otimes 2})_N$, since the degree $N$ part of the cohomology vanishes. We then set $\rho_{N+1} := \rho_N + \beta_N$, where $\beta_N\in{\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^2}^N$ is a representative of $\overline\beta_N$. Then $\rho_{N+1}$ satisfies (\ref{twistequation}) in ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} /({\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3} \cap {\mathfrak{m}}_{({\mathfrak{g}}^*)^3}^{N+1})$. The sequence $(\rho_N)_{N\geq 2}$ has a limit $\rho$, which then satisfies (\ref{twistequation}). The second part of the theorem can be proved either by analyzing the choices for $\overline\beta_N$ in the above proof, or following the proof of the previous proposition. \hfill \qed \medskip \begin{remark} \label{rem:sigma} If $\varphi$ is replaced by $\varphi' = \sigma \star \varphi$, then a solution of (\ref{twistequation}) is $\rho' = \rho \star (-\sigma)$. \end{remark} \section{Isomorphism of formal Poisson manifolds ${\mathfrak{g}}^* \simeq G^*$} Let us assume that ${\mathfrak{g}}$ is a finite dimensional coboundary Lie bialgebra. the following result was proved in \cite{EEM} when ${\mathfrak{g}}$ is quasitriangular; the result of \cite{EEM} is itself a generalization of the formal version of the Ginzburg-Weinstein isomorphism (\cite{GW,A,Bo}). \begin{corollary} \label{coroprinc} There exists an isomorphism of formal Poisson manifolds ${\mathfrak{g}}^* \simeq G^*$. \end{corollary} {\em Proof.} Let $P : \wedge^2({\mathcal O}_{{\mathfrak{g}}^*}) \to {\mathcal O}_{{\mathfrak{g}}^*}$ be the Poisson bracket on ${\mathcal O}_{{\mathfrak{g}}^*}$ corresponding to the Lie-Poisson\footnote{or Kostant-Kirillov-Souriau, or linear} Poisson structure on ${\mathfrak{g}}^*$. Then $(\O_{{\mathfrak{g}}^*},m_0,P,\Delta_0)$ is a Poisson formal series Hopf (PFSH) algebra; it corresponds to the formal Poisson-Lie group $({\mathfrak{g}}^*,+)$ equipped with its Lie-Poisson structure. Set ${}^{\rho}\Delta(f) = \rho\star\Delta_0(f)\star (-\rho)$ for any $f\in{\mathcal O}_{{\mathfrak{g}}^*}$. It follows from the fact that $\rho$ satisfies the functional cocycle equation that $({\mathcal O}_{{\mathfrak{g}}^*},m_0,P,{}^{\rho}\Delta_0)$ is a PFSH algebra. Let us denote by ${\bf PFSHA}$ and ${\bf LBA}$ the categories of PSFH algebras and Lie bialgebras. We have a category equivalence $c : {\bf PFSHA} \to {\bf LBA}$, taking $({\mathcal O},m,P,\Delta)$ to the Lie bialgebra $(\c,\mu,\delta)$, where $\c := {\mathfrak{m}}/{\mathfrak{m}}^2$ (${\mathfrak{m}}\subset{\mathcal O}$ is the maximal ideal), the Lie cobracket of $\c$ is induced by $\Delta - \Delta^{2,1} : {\mathfrak{m}}\to \wedge^2({\mathfrak{m}})$, and the Lie bracket of $\c$ is induced by the Poisson bracket $P : \wedge^2({\mathfrak{m}}) \to {\mathfrak{m}}$. The inverse of the functor $c$ takes $(\c,\mu,\delta)$ to ${\mathcal O} = \wh S(\c)$ equipped with its usual product; $\Delta$ depends only on $\delta$ and $P$ depends on $(\mu,\delta)$. Then $c$ restricts to a category equivalence $c_{\on{fd}} : {\bf PFSHA}_{\on{fd}} \to {\bf LBA}_{\on{fd}}$ of subcategories of finite-dimensional objects (in the case of ${\bf PFSH}$, we say that ${\mathcal O}$ is finite-dimensional iff ${\mathfrak{m}}/{\mathfrak{m}}^2$ is). Let $\on{dual} : {\bf LBA}_{\on{fd}} \to {\bf LBA}_{\on{fd}}$ be the duality functor. It is a category antiequivalence; we have $\on{dual}({\mathfrak{g}},\mu,\delta) = ({\mathfrak{g}}^*,\delta^t,\mu^t)$. Then $\on{dual} \circ c_{\on{fd}} : {\bf PFSHA}_{\on{fd}} \to {\bf LBA}_{\on{fd}}$ is a category antiequivalence. Its inverse it the usual functor ${\mathfrak{g}}\mapsto U({\mathfrak{g}})^*$. If $G$ is the formal Poisson-Lie group with Lie bialgebra ${\mathfrak{g}}$, one sets ${\mathcal O}_G = U({\mathfrak{g}})^*$. Let us apply the functor $c$ to $({\mathcal O}_{{\mathfrak{g}}^*},m_0,P,{}^\rho\Delta_0)$. We obtain $\c = {\mathfrak{m}}/{\mathfrak{m}}^2 = {\mathfrak{g}}$; the Lie bracket is unchanged w.r.t. the case $\rho=0$, so it is the Lie bracket of ${\mathfrak{g}}$; the Lie cobracket is given by $\delta(x) = [r,x\otimes 1 + 1\otimes x]$ since the reduction of $\rho$ modulo $({\mathfrak{m}}_{{\mathfrak{g}}^*})^2\wh\otimes {\mathfrak{m}}_{{\mathfrak{g}}^*} + {\mathfrak{m}}_{{\mathfrak{g}}^*} \wh\otimes ({\mathfrak{m}}_{{\mathfrak{g}}^*})^2$ is equal to $r$. Then applying $\on{dual} \circ c_{\on{fd}}$ to $({\mathcal O}_{{\mathfrak{g}}^*},m_0,P,{}^\rho\Delta_0)$, we obtain the Lie bialgebra ${\mathfrak{g}}^*$. So this PFSH algebra is isomorphic to the PFSH algebra of the formal Poisson-Lie group $G^*$. In particular, the Poisson algebras ${\mathcal O}_{{\mathfrak{g}}^*}$ and ${\mathcal O}_{G^*}$ are isomorphic. It is easy to check that the map ${\mathfrak{g}} = {\mathfrak{m}}_{{\mathfrak{g}}^*} / {\mathfrak{m}}_{{\mathfrak{g}}^*}^2 \to {\mathfrak{m}}_{G^*}/{\mathfrak{m}}_{G^*}^2 = {\mathfrak{g}}$ induced by this isomorphism is the identity (here ${\mathfrak{m}}_{G^*} \subset {\mathcal O}_{G^*}$ is the maximal ideal). \hfill \qed \medskip \begin{remark} When ${\mathfrak{g}}$ is infinite dimensional, one can define ${\mathcal O}_{G^*}$ as the image of ${\mathfrak{g}}$ under ${\bf LBA} \to {\bf PFSHA}$ and then show that the Poisson algebras ${\mathcal O}_{G^*}$ and ${\mathcal O}_{{\mathfrak{g}}^*} = (\wh S({\mathfrak{g}}),$ linear Poisson structure) are isomorphic. \end{remark} \section{The morphism $S({\mathfrak{g}}^*)^{\mathfrak{g}} \hookrightarrow U({\mathfrak{g}}^*)$} In this section, ${\mathfrak{g}}$ is a finite dimensional coboundary Lie bialgebras. The following fact is well-known (\cite{STS2}): \begin{lemma} \label{lemma:O} ${\mathcal O}_G^{\mathfrak{g}} \subset {\mathcal O}_G$ is a Poisson commutative subalgebra. \end{lemma} Here the action of ${\mathfrak{g}}$ on ${\mathcal O}_G$ corresponds to adjoint action of $G$. We recall the proof: if $f,g\in {\mathcal O}_G$, then $\{f,g\} = m(({\bf L} - {\bf R})(r)(f\otimes g))$, where ${\bf L}, {\bf R}$ are the infinitesimal left and right actions and $m$ is the product map. If $\varphi \in {\mathcal O}_G^{\mathfrak{g}}$, then ${\bf L}(a)(\varphi) = {\bf R}(a)(\varphi)$ for any $a\in{\mathfrak{g}}$, therefore if $f,g\in{\mathcal O}_G^{\mathfrak{g}}$, then $({\bf L} - {\bf R})(r)(f\otimes g) = 0$, hence $\{f,g\}=0$. The inclusion ${\mathcal O}_G^{\mathfrak{g}} \subset {\mathcal O}_G$ is a morphism of Poisson algebras with a decreasing filtration. By passing to the associated graded, we obtain: \begin{lemma} \label{lemma:poisson} $S({\mathfrak{g}}^*)^{\mathfrak{g}} \subset S({\mathfrak{g}}^*)$ is a Poisson commutative subalgebra. \end{lemma} {\em Another proof.} If $\alpha,\beta\in{\mathfrak{g}}^*$, then $[\alpha,\beta] = \on{ad}^*(R(\beta))(\alpha) - \on{ad}^*(R(\alpha))(\beta)$, where $R : {\mathfrak{g}}^* \mapsto {\mathfrak{g}}$ is given by $R(\xi) = (\on{id}\otimes\xi)(r)$. Let $f,g \in S({\mathfrak{g}}^*)^{\mathfrak{g}}$ be of degrees $k$ and $\ell$. Write $f = \sum_\alpha a_1^\alpha \cdots a_k^\alpha$, $g = \sum_\beta b_1^\beta \cdots b_\ell^\beta$. Then $$ \{f,g\} = \sum_\beta \sum_{j=1}^\ell b_1^\beta \cdots \check b_j^\beta\cdots b_\ell^\beta \on{ad}^*(R(b_j^\beta))(f) - \sum_\alpha \sum_{i=1}^k a_1^\alpha \cdots \check a_i^\alpha\cdots a_k^\alpha \on{ad}^*(R(a_i^\alpha))(g). $$ When $f$ and $g$ are both invariant, this bracket vanishes. \hfill \qed \medskip We now prove that $S({\mathfrak{g}}^*)^{\mathfrak{g}} \subset S({\mathfrak{g}}^*)$ is also the associated graded of an inclusion of noncommutative algebras with an increasing filtration: \begin{theorem} \label{theoprinc} There exists a morphism of filtered algebras: $$\theta : S({\mathfrak{g}}^*)^{\mathfrak{g}} \to U({\mathfrak{g}}^*),$$ the associated graded morphism of which is the canonical inclusion $S({\mathfrak{g}}^*)^{\mathfrak{g}} \subset S({\mathfrak{g}}^*)$. \end{theorem} {\em Proof.} Let us denote by ${\bf FSHA}$ the category of formal series Hopf (FSH) algebras and by ${\bf FilAlg}$ the category of filtered algebras. There is a contravariant functor (restricted duality) ${\bf FSHA} \to {\bf FilAlg}$, defined by ${\mathcal O}\mapsto {\mathcal O}^\circ$, where ${\mathcal O}^\circ = \{\ell\in {\mathcal O}^* | \exists n\geq 0, \ell({\mathfrak{m}}^n) = 0\} \subset {\mathcal O}^*$; here ${\mathfrak{m}}\subset{\mathcal O}$ is the maximal ideal of ${\mathcal O}$. The algebra structure of ${\mathcal O}^\circ$ is defined by $(\ell_1 \cdot \ell_2)(f) = (\ell_1 \otimes \ell_2)(\Delta(f))$; its filtration is defined by $({\mathcal O}^\circ)_{\leq n} = \{\ell\in{\mathcal O}^* | \ell({\mathfrak{m}}^{n+1}) = 0\}$. Note that we have a category equivalence ${\bf FSHA} \to {\bf LCA}$, where ${\bf LCA}$ is the category of Lie coalgebras, taking ${\mathcal O}$ to ${\mathfrak{m}}/{\mathfrak{m}}^2$, equipped with the cobracket induced by $\Delta - \Delta^{2,1}$. Then the composed functor ${\bf LCA} \to {\bf FSHA} \to {\bf FilAlg}$ is $\c\mapsto U(\c^*)$ (recall that $\c^*$ is a Lie algebra). $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0) = (\wh S({\mathfrak{g}}),\Delta_0)$ is a graded FSH algebra. Its restricted dual is the graded algebra $S({\mathfrak{g}}^*)$. Recall that ${\mathcal O}_{{\mathfrak{g}}^*}$ is also a Poisson algebra. We define the set of Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}$ as the subspace of all $\ell\in{\mathcal O}_{{\mathfrak{g}}^*}^\circ$, such that $\ell(\{u,v\}) = 0$ for any $u,v\in{\mathcal O}_{{\mathfrak{g}}^*}$. Then $\{$Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}\} \subset {\mathcal O}_{{\mathfrak{g}}^*}^\circ$ identifies with $S({\mathfrak{g}}^*)^{\mathfrak{g}} \subset S({\mathfrak{g}}^*)$; this is a graded subalgebra of ${\mathcal O}_{{\mathfrak{g}}^*}^\circ$. This defines a graded algebra structure on $\{$Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}\}$. Consider the FSH algebra $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)$. It is isomorphic (as a filtered vector space) to $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)$, and this isomorphism induces an algebra isomorphism between their associated graded FSH algebras. It follows that we have an isomorphism of filtered vector spaces between the filtered algebra $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ and $S({\mathfrak{g}}^*)$, and the associated graded of this morphism is an algebra isomorphism $\on{gr} (({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ) \to S({\mathfrak{g}}^*)$. Recall that the vector spaces underlying $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)^\circ$ and $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ are the same (i.e., ${\mathcal O}_{{\mathfrak{g}}^*}^\circ$). We claim that the canonical inclusion $\{$Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}\} \subset ({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ is a morphism of filtered algebras. Indeed, let us denote by $\cdot_\rho$ (resp., $\cdot$) the product of $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ (resp., $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)^\circ$). Let $\ell_1,\ell_2$ be Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}$. Then for any $x\in{\mathcal O}_{{\mathfrak{g}}^*}$, we have $(\ell_1 \cdot_\rho \ell_2)(x) = (\ell_1 \otimes \ell_2) (\rho \star \Delta_0(f) \star (-\rho))$. Now Leibniz's rule implies that $(\ell_1\otimes \ell_2)(\{u,v\})=0$ for any $u,v\in{\mathcal O}_{{\mathfrak{g}}^*}^{\wh\otimes 2}$, therefore $(\ell_1 \cdot_\rho \ell_2)(x) = (\ell_1\otimes \ell_2)(\Delta_0(x)) = (\ell_1\cdot \ell_2)(x)$. So $\{$Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}\} \subset ({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ is an algebra morphism. Since the filtrations on the vector spaces underlying $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)^\circ$ and $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ are the same, and since the filtration on $\{$Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}\}$ is induced by that of $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)^\circ$, this morphism is filtered, and its associated graded is the canonical inclusion $S({\mathfrak{g}}^*)^{\mathfrak{g}} \subset S({\mathfrak{g}}^*)$. Now the FSH algebra isomorphism ${\mathcal O}_{G^*} \simeq ({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)$ (Corollary \ref{coroprinc}) induces a filtered algebra isomorphism $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ \to {\mathcal O}_{G^*}^\circ = U({\mathfrak{g}}^*)$. The fact that the associated graded of this morphism is the canonical isomorphism $S({\mathfrak{g}}^*) \to \on{gr}(U({\mathfrak{g}}^*))$ follows from the fact that the completed graded of the FSH algebras ${\mathcal O}_{G^*}$ and $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)$ are both $({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)$. We now compose the filtered algebra morphism $\{$Poisson traces on ${\mathcal O}_{{\mathfrak{g}}^*}\} \subset ({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ with the filtered algebra isomorphism $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ \to {\mathcal O}_{G^*}^\circ$ and obtain a filtered algebra morphism $S({\mathfrak{g}}^*)^{\mathfrak{g}} \to U({\mathfrak{g}}^*)$, whose associated graded is the canonical inclusion $S({\mathfrak{g}}^*)^{\mathfrak{g}} \subset S({\mathfrak{g}}^*)$. The situation may be summarized as follows: $$ \begin{matrix} & & ({\mathcal O}_{{\mathfrak{g}}^*},\Delta_0)^\circ = S({\mathfrak{g}}^*) & & \\ & \scriptstyle{(a)} \nearrow & & & & \\ S({\mathfrak{g}}^*)^{\mathfrak{g}} = \{\hbox{Poisson\ traces\ on\ }{\mathcal O}_{{\mathfrak{g}}^*}\} & & \scriptstyle{(c)}\uparrow & & & \\ & \scriptstyle{(b)}\searrow & & & \\ & & ({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ & \stackrel{(d)}{\to} & {\mathcal O}_{G^*}^\circ = U({\mathfrak{g}}^*) \end{matrix} $$ Here $S({\mathfrak{g}}^*)^{\mathfrak{g}}$ and $S({\mathfrak{g}}^*)$ are graded algebras, $({\mathcal O}_{{\mathfrak{g}}^*},{}^\rho\Delta_0)^\circ$ and ${\mathcal O}_{G^*}^\circ$ are filtered algebras; $(a)$ is a morphism of graded algebras, $(c)$ is an isomorphism of filtered vector spaces, $(b)$ and $(d)$ are morphisms of filtered algebras ($(d)$ is an isomorphism). The associated graded of $(c)$ is an isomorphism of graded algebras. \hfill \qed \medskip \begin{remark} The restricted dual of the isomorphism ${\mathcal O}_{{\mathfrak{g}}^*} \to {\mathcal O}_{G^*}$ appearing in the above proof is an isomorphism of filtered vector spaces $\sigma : S({\mathfrak{g}}^*) \to U({\mathfrak{g}}^*)$, whose associated graded is the canonical isomorphism $S({\mathfrak{g}}^*) \to \on{gr}(U({\mathfrak{g}}^*))$. These properties are also satisfied by the symmetrization map $\on{Sym}$, however $\sigma$ depends on $\rho$, so in general $\on{Sym}$ and $\sigma$ are different. \end{remark} \begin{remark} One can check that the morphism $\theta$ is independent on the choice of $(\rho,\varphi)$ (these choices are described in Remark \ref{rem:sigma} and in Theorem \ref{lift2}). \end{remark} \section{Duality of QUE and QFSH algebras} \label{sect:duality} In this section, we recall some facts from \cite{Dr:QG} (proofs can be found in \cite{Gav}). Let us denote by ${\bf QUE}$ the category of quantized universal enveloping (QUE) algebras and by ${\bf QFSH}$ the category of quantized formal series Hopf (QFSH) algebras. We denote by ${\bf QUE}_{\on{fd}}$ and ${\bf QFSH}_{\on{fd}}$ the subcategories corresponding to finite dimensional Lie bialgebras. We have contravariant functors ${\bf QUE}_{\on{fd}} \to {\bf QFSH}_{\on{fd}}$, $U\mapsto U^*$ and ${\bf QFSH}_{\on{fd}} \to {\bf QUE}_{\on{fd}}$, ${\mathcal O}\mapsto {\mathcal O}^\circ$. These functors are inverse to each other. $U^*$ is the full topological dual of $U$, i.e., the space of all continuous (for the $\hbar$-adic topology) ${\mathbb{K}}[[\hbar]]$-linear maps $U \to {\mathbb{K}}[[\hbar]]$. ${\mathcal O}^\circ$ the space of continuous ${\mathbb{K}}[[\hbar]]$-linear forms ${\mathcal O}\to {\mathbb{K}}[[\hbar]]$, where ${\mathcal O}$ is equipped with the ${\mathfrak{m}}$-adic topology (here ${\mathfrak{m}}\subset {\mathcal O}$ is the maximal ideal). We also have covariant functors ${\bf QUE} \to {\bf QFSH}$, $U\mapsto U'$ and ${\bf QFSH} \to {\bf QUE}$, ${\mathcal O}\mapsto {\mathcal O}^\vee$. There functors are also inverse to each other. $U'$ is a subalgebra of $U$, while ${\mathcal O}^\vee$ is the $\hbar$-adic completion of $\sum_{k\geq 0} \hbar^{-k} {\mathfrak{m}}^k \subset {\mathcal O}[1/\hbar]$. We also have canonical isomorphisms $(U')^\circ \simeq (U^*)^\vee$ and $({\mathcal O}^\vee)^* \simeq ({\mathcal O}^\circ)'$. If $\a$ is a finite dimensional Lie bialgebra and $U = U_\hbar(\a)$ is a QUE algebra quantizing $\a$, then $U^* = {\mathcal O}_{A,\hbar}$ is a QFSH algebra quantizing the Poisson-Lie group $A$ (with Lie bialgebra $\a$), and $U' = {\mathcal O}_{A^*,\hbar}$ is a QFSH algebra quantizing the Poisson-Lie group $A^*$ (with Lie bialgebra $\a^*$). If now ${\mathcal O} = {\mathcal O}_{A,\hbar}$ is a QFSH algebra quantizing $A$, then ${\mathcal O}^\circ = U_\hbar(\a)$ is a QUE algebra quantizing $\a$ and ${\mathcal O}^\vee = U_\hbar(\a^*)$ is a QFSH algebra quantizing $\a^*$. We now compute these functors explicitly in the case of cocommutative QUE and commutative QFSH algebras. If $U = U(\a)[[\hbar]]$ with cocommutative coproduct (where $\a$ is a Lie algebra), then $U'$ is a completion of $U(\hbar \a[[\hbar]])$; this is a flat deformation of $\wh S(\a)$ equipped with its linear Lie-Poisson structure. If $G$ is a formal group with function ring ${\mathcal O}_G$, then ${\mathcal O} := {\mathcal O}_G[[\hbar]]$ is a QFSH algebra, and ${\mathcal O}^\vee$ is a commutative QUE algebra; it is a quantization of $(S({\mathfrak{g}}^*)$, commutative product, cocommutative coproduct, co-Poisson structure induced by the Lie bracket of ${\mathfrak{g}})$. \section{Relation between twist quantization and its functional version} \label{sect:rel} Let us define a twist quantization of the coboundary Lie bialgebra $({\mathfrak{g}},r,Z)$ as a pair $(J,\Phi)$, $J\in U({\mathfrak{g}})^{\otimes 2}[[\hbar]]$, $\Phi \in U({\mathfrak{g}})^{\otimes 3}[[\hbar]]$, such that $\Phi$ is invariant, and $(J,\Phi)$ satisfies the twisted cocycle relation \begin{equation} \label{eq1} J^{1,2}J^{12,3}=J^{2,3}J^{1,23}\Phi, \end{equation} $(\varepsilon\otimes\id)(J)=(\id\otimes\varepsilon)(J)=1$, $J = 1^{\otimes 2} + O(\hbar)$, $\Phi = 1^{\otimes 3} + O(\hbar)$, $\on{Alt}((J-1^{\otimes 2})/\hbar) = r + O(\hbar)$, $\on{Alt}((\Phi-1^{\otimes 3})/\hbar^2) = Z + O(\hbar)$. These conditions imply that $\Phi$ satisfies the pentagon relation, as well as $\varepsilon^{(i)}(\Phi) = 1^{\otimes 2}$, $i=1,2,3$. (We know that such a twist quantization always exists when ${\mathfrak{g}}$ is triangular or quasi-triangular.) Our purpose is to relate twist quantization with its functional version. The first step is to show that $(J,\Phi)$ can be transformed into an admissible pair, in a sense which we now precise. \begin{definition} \label{admissible} {\it An element $x$ in a QUE algebra $U$ is admissible if $x\in 1 + \hbar U$, and if $\hbar \log x$ is in $U' \subset U$.} \end{definition} We will use the isomorphism $U({\mathfrak{g}})^{\otimes k}[[\hbar]] \simeq U({\mathfrak{g}}^{\oplus k})[[\hbar]]$ to view $U({\mathfrak{g}})^{\otimes k}[[\hbar]]$ as a QUE algebra. \begin{proposition} \label{propadmi} Any twist quantization $(J,\Phi)$ of a coboundary Lie bialgebra $({\mathfrak{g}},r,Z)$ is gauge equivalent to an admissible twist quantization $(J',\Phi')$ (i.e., such that $J'$ and $\Phi'$ are admissible). \end{proposition} {\em Proof.} Let us set $U = U({\mathfrak{g}})[[\hbar]]$. According to Proposition \ref{prop:assoc}, one can find an invariant $F \in U^{\wh\otimes 2}$, such that $F \in 1^{\otimes 2} + \hbar U_0^{\wh\otimes 2}$ and $\Phi' := {}^F\Phi = F^{2,3} F^{1,23} \Phi (F^{1,2}F^{12,3})^{-1}$ is admissible. In particular, $\Phi' \in 1^{\otimes 3} + \hbar^2 U_0^{\wh\otimes 3}$. Then if we set $J_0 := JF$, we have $J_0^{1,2} J_0^{12,3} = J_0^{2,3} J_0^{1,23} \Phi'$, and $J_0 \in 1^{\otimes 2} + \hbar U_0^{\wh\otimes 2}$. For any $u\in 1^{\otimes 3} + \hbar U_0$, ${}^uJ_0 := u^1 u^2 J_0 (u^{12})^{-1}$ is such that $({}^uJ_0,\Phi')$ is a twist quantization of $({\mathfrak{g}},r,Z)$. It remains to find $u$ such that $J' := {}^u J_0$ is admissible. We will construct $u$ as a product $\cdots u_2 u_1$, where $u_n\in 1 + \hbar^n U_0$, in such a way that if $J_n := {}^{u_n\cdots u_1}J_0$, then $\hbar\log(J_n) \in U_0^{\prime\wh\otimes 2} + \hbar^{n+2} U_0^{\wh\otimes 2}$. We have already $\hbar\log(J_0) \in \hbar^2 U_0^{\wh\otimes 2}$. Expand $J_0 = 1^{\otimes 2} + \hbar j_1 + \cdots$, then $\on{Alt}(j_1) = r$. Moreover, the coefficient of $\hbar$ in $J_0^{1,2}J_0^{12,3} = J_0^{2,3}J_0^{1,23}\Phi$ yields $d(j_1) = 0$, where $d : U({\mathfrak{g}})_0^{\otimes 2} \to U({\mathfrak{g}})_0^{\otimes 3}$ is the co-Hochschild differential. It follows that for some $a_1\in U({\mathfrak{g}})_0$, we have $j_1 = r+d(a_1)$. Then if we set $u_1 := \exp(\hbar a_1)$ and $J_1 = {}^{u_1}J_0$, we get $J_1 \in 1^{\otimes 2} + \hbar r + \hbar^2 U_0^{\wh\otimes 2}$. Then $\hbar\log(J_1) \in \hbar^2 r + \hbar^3 U_0^{\wh\otimes 2} \subset U_0^{\prime\wh\otimes 2} + \hbar^3 U_0^{\wh\otimes 3}$. Assume that for $n\geq 2$, we have constructed $u_1,\ldots,u_{n-1}$ such that $\alpha_{n-1} := \hbar\log(J_{n-1}) \in U_0^{\prime\wh\otimes 2} + \hbar^{n+1} U_0^{\wh\otimes 2}$. Let us denote by $\bar\alpha$ the image of the class of $\alpha_{n-1}$ in $U({\mathfrak{g}})_0^{\otimes 2} / (U({\mathfrak{g}})_0^{\otimes 2})_{\leq n+1}$ under the isomorphism of this space with $(U_0^{\prime\wh\otimes 2} + \hbar^{n+1} U_0^{\wh\otimes 2}) / (U_0^{\prime\wh\otimes 2} + \hbar^{n+2} U_0^{\wh\otimes 2})$ (see Lemma \ref{lemma:quot}). Let $\alpha\in U({\mathfrak{g}})_0^{\otimes 2}$ be a representative of $\bar\alpha$, then $\alpha_{n-1} = \alpha' + \hbar^{n+1}\alpha$, where $\alpha'\in U_0^{\prime\wh\otimes 2} + \hbar^{n+2} U_0^{\wh\otimes 2}$. Let us set $\varphi' := \hbar\log(\Phi')$, then the twist equation gives $$ (-\alpha'-\hbar^{n+1}\alpha)^{1,23} \star_\hbar (-\alpha'-\hbar^{n+1}\alpha)^{2,3} \star_\hbar (\alpha'+\hbar^{n+1}\alpha)^{1,2} \star_\hbar (\alpha'+\hbar^{n+1}\alpha)^{12,3} = \varphi', $$ where $\star_\hbar$ is defined as in Appendix \ref{app:A}. According to Lemma \ref{lemma:approx}, the image of this equality in $(U^{\wh\otimes 3} + \hbar^{n+1} U^{\prime\wh\otimes 3}) / (U^{\wh\otimes 3} + \hbar^{n+2} U^{\prime\wh\otimes 3}) \simeq U({\mathfrak{g}})^{\otimes 3} / (U({\mathfrak{g}})^{\otimes 3})_{\leq n+1}$ is $d(\bar\alpha)$, where $d$ is the co-Hochschild differential on $U({\mathfrak{g}})_0^{\otimes \cdot} / (U({\mathfrak{g}})_0^{\otimes\cdot})_{\leq n+1}$. Since $n\geq 2$, the relevant cohomology group vanishes, so $\bar\alpha = d(\bar\beta)$, where $\bar\beta\in U({\mathfrak{g}})_0 /(U({\mathfrak{g}})_0)_{\leq n+1}$. Let $\beta\in U({\mathfrak{g}})_0$ be a representative of $\bar\beta$ and set $u_n := \exp(\hbar^n\beta)$, $J_n := {}^{u_n}J_{n-1}$, $\alpha_n := \hbar\log(J_n)$. Then $$ \alpha_n = (\hbar^{n+1}\beta)^1 \star_\hbar (\hbar^{n+1}\beta)^2 \star_\hbar \alpha_{n-1} \star_\hbar (-\hbar^{n+1}\beta)^{12}. $$ According to Lemma \ref{lemma:approx}, the image of $\alpha_n$ in $$ (U_0^{\wh\otimes 2} + \hbar^{n+1} U_0^{\prime\wh\otimes 2}) / (U_0^{\wh\otimes 2} + \hbar^{n+2} U_0^{\prime\wh\otimes 2}) \simeq U({\mathfrak{g}})_0^{\otimes 2}/(U({\mathfrak{g}})^{\otimes 2}_0)_{\leq n+1} $$ is $\bar\alpha - d(\bar\beta)=0$. So $\alpha_n$ belongs to $U_0^{\wh\otimes 2} + \hbar^{n+2} U_0^{\prime\wh\otimes 2}$, as required. This proves the induction step. \hfill \qed \medskip If now $(J',\Phi')$ is an admissible twist quantization, then $\rho := \hbar \log(J')_{|\hbar=0}$ and $\varphi := \hbar \log(\Phi')_{|\hbar=0}$ are formal functions on ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 2}$ and ${\mathfrak{m}}_{{\mathfrak{g}}^*}^{\wh\otimes 3}$, solutions of the functional twist equation. \section{Quantization of ${\mathcal O}_G^{\mathfrak{g}} \subset {\mathcal O}_G$} \label{sect:O} Using a (non necessarily admissible) twist quantization, we construct a formal noncommutative deformation of the inclusion of algebras of Lemma \ref{lemma:O}: \begin{proposition} We have an injective algebra morphism ${\mathcal O}_G^{\mathfrak{g}}[[\hbar]] \hookrightarrow {\mathcal O}_{G,\hbar}$ deforming ${\mathcal O}_G^{\mathfrak{g}}\subset {\mathcal O}_G$, where ${\mathcal O}_{G,\hbar}$ is a quantization of the PFSH algebra ${\mathcal O}_G$ and ${\mathcal O}_G^{\mathfrak{g}}[[\hbar]]$ is the trivial deformation of the commutative algebra ${\mathcal O}_G^{\mathfrak{g}}$ (it is also commutative). \end{proposition} {\em Proof.} Let us first construct the QFSH algebra ${\mathcal O}_{G,\hbar}$. For $x\in U({\mathfrak{g}})[[\hbar]]$, set ${}^J\Delta_0(x) = J\Delta_0(x) J^{-1}$, where $\Delta_0$ is the usual cocommutative coproduct. Then $U_\hbar({\mathfrak{g}}) = (U({\mathfrak{g}})[[\hbar]],m_0,{}^J\Delta_0)$ is a quantization of the Lie bialgebra ${\mathfrak{g}}$ (here $m_0$ is the product on $U({\mathfrak{g}})$). The dual ${\mathcal O}_{G,\hbar} := U_\hbar({\mathfrak{g}})^*$ of this QUE algebra is a QFSH algebra quantizing the PFSH algebra ${\mathcal O}_G$. The product in this QFSH algebra is defined by $(f \star g)(x) = (f \otimes g)(J\Delta_0(x)J^{-1})$ for $f,g\in U_\hbar({\mathfrak{g}})^*$ and $x\in U_\hbar({\mathfrak{g}})$. On the other hand, the FSH algebra ${\mathcal O}_G$ is equal to $U({\mathfrak{g}})^*$, and its product is defined by $(fg)(x) = (f\otimes g)(\Delta_0(x))$ for $f,g\in U({\mathfrak{g}})^*$ and $x\in U({\mathfrak{g}})$. We say that $f\in U({\mathfrak{g}})^*$ is a trace iff $f(xy) = f(yx)$ for any $x,y\in U({\mathfrak{g}})$. Then the inclusion $\{$traces on $U({\mathfrak{g}})\} \subset U({\mathfrak{g}})^*$ identifies with ${\mathcal O}_G^{\mathfrak{g}} \subset {\mathcal O}_G$. In the same way, we define $\{$traces on $U({\mathfrak{g}})[[\hbar]]\}$; this is a subalgebra of $U({\mathfrak{g}})[[\hbar]]^* \simeq {\mathcal O}_G[[\hbar]]$, which identifies with ${\mathcal O}_G^{\mathfrak{g}}[[\hbar]]$. The canonical map $\{$traces on $U({\mathfrak{g}})[[\hbar]]\} \to U_\hbar({\mathfrak{g}})^*$ is an algebra morphism. Indeed, if $f_1,f_2$ are traces on $U({\mathfrak{g}})[[\hbar]]$, then $f_1\otimes f_2$ is a trace on $U({\mathfrak{g}})^{\otimes 2}[[\hbar]]$, so $(f_1 \star f_2)(x) = (f_1\otimes f_2)(J\Delta_0(x) J^{-1}) = (f_1\otimes f_2)(\Delta_0(x)) = (f_1 f_2)(x)$ for any $x\in U({\mathfrak{g}})[[\hbar]]$, so $f_1\star f_2 = f_1 f_2$. So we have obtained an algebra morphism ${\mathcal O}_G^{\mathfrak{g}}[[\hbar]] \to U_\hbar({\mathfrak{g}})^* = {\mathcal O}_{G,\hbar}$. It is clearly a deformation of the canonical inclusion ${\mathcal O}_G^{\mathfrak{g}}\subset {\mathcal O}_G$. \hfill \qed \medskip \section{Quantization of $S({\mathfrak{g}}^*)^{\mathfrak{g}} \hookrightarrow U({\mathfrak{g}}^*)$} \label{sect:SG} Assume now that $(J,\Phi)$ is an admissible twist quantization. We will construct a formal deformation of the inclusion of algebras of Theorem \ref{theoprinc}. \begin{theorem} There is an injective algebra morphism: $$ \theta_\hbar : S({\mathfrak{g}}^*)^{\mathfrak{g}}[[\hbar]] \hookrightarrow U_\hbar({\mathfrak{g}}^*), $$ where $U_\hbar({\mathfrak{g}}^*)$ is a quantization of ${\mathfrak{g}}^*$. Its reduction modulo $\hbar$ coincides with the morphism $S({\mathfrak{g}}^*)^{\mathfrak{g}} \hookrightarrow U({\mathfrak{g}})$ from Theorem \ref{theoprinc}. \end{theorem} {\em Proof.} Recall that $U({\mathfrak{g}})[[\hbar]]'$ is a cocommutative QFSH algebra; we denote by $m_0$, $\Delta_0$ its product and coproduct. Since $(\varepsilon\otimes\id)(J) = (\id\otimes\varepsilon)(J)=1$, we have $\hbar\log(J) \in {\mathfrak{m}}_0^{\wh\otimes 2}$, where ${\mathfrak{m}}_0\subset U({\mathfrak{g}})[[\hbar]]'$ is the kernel of the counit. According to \cite{EH1}, Proposition 3.1, this implies that the inner automorphism $z\mapsto JzJ^{-1}$ of $U({\mathfrak{g}})^{\otimes 2}[[\hbar]]$ restricts to an automorphism of $U({\mathfrak{g}})^{\otimes 2}[[\hbar]]'$. We then equip $U({\mathfrak{g}})[[\hbar]]'$ with the coproduct ${}^J\Delta : x\mapsto J \Delta_0(x) J^{-1}$. Then $(U({\mathfrak{g}})[[\hbar]]',m_0,{}^J\Delta_0)$ is a QFSH algebra. Its classical limit is the PFSH algebra $({\mathcal O}_{{\mathfrak{g}}^*},m_0,P,{}^\rho\Delta_0)$. We have seen that this PSFH algebra is isomorphic to ${\mathcal O}_{G^*}$, hence $(U({\mathfrak{g}})[[\hbar]]',m_0,{}^J\Delta_0)$ is a quantization of ${\mathcal O}_{G^*}$. It now follows from Section \ref{sect:duality} that $(U({\mathfrak{g}})[[\hbar]]',m_0,{}^J\Delta_0)^\circ$ is a quantization of $U({\mathfrak{g}}^*)$, which we denote by $U_\hbar({\mathfrak{g}}^*)$. Let us say that $\varphi\in (U({\mathfrak{g}})[[\hbar]]')^\circ$ is a trace if $\varphi(xy) = \varphi(yx)$ for any $x,y\in U({\mathfrak{g}})[[\hbar]]'$. Then $\{$traces on $U({\mathfrak{g}})[[\hbar]]'\} \subset (U({\mathfrak{g}})[[\hbar]]')^\circ$ is a subalgebra. Indeed, if $\ell_1,\ell_2$ are traces then $\ell_1\otimes \ell_2$ is also a trace, so for $x,y\in U({\mathfrak{g}})[[\hbar]]'$, we have $(\ell_1\ell_2)(xy) = (\ell_1\otimes \ell_2)(\Delta(x)\Delta(y)) = (\ell_1\ell_2)(yx)$. This inclusion identifies with the inclusion $({\mathcal O}_G[[\hbar]]^\vee)^{\mathfrak{g}} \subset {\mathcal O}_G[[\hbar]]^\vee$. Indeed, the Drinfeld functors have the property that $(U')^\circ = (U^*)^\vee$ for any QUE algebra $U$. Now we show that the map $\{$traces on $U({\mathfrak{g}})[[\hbar]]'\} \subset (U({\mathfrak{g}})[[\hbar]]',{}^J\Delta_0)^\circ$ is also an algebra morphism. Indeed, let $\cdot_J$ be the product of the latter algebra. If $\ell_1,\ell_2$ are traces and $x,y\in U({\mathfrak{g}})[[\hbar]]$, then $(\ell_1 \cdot_J \ell_2)(x) = (\ell_1\otimes \ell_2)(J\Delta_0(x)J^{-1}) = (\ell_1\otimes \ell_2)(\Delta_0(x)) = (\ell_1\ell_2)(x)$, so $\ell_1 \cdot_J \ell_2 = \ell_1 \ell_2$. So we have constructed an algebra morphism $({\mathcal O}_G[[\hbar]]^\vee)^{\mathfrak{g}} \to U_\hbar({\mathfrak{g}})$. It is clearly a deformation of the morphism constructed in Theorem \ref{theoprinc}. Recall that ${\mathcal O}_G[[\hbar]]^\vee$ is the $\hbar$-adic completion of $\sum_{k\geq 0} \hbar^{-k} {\mathfrak{m}}_G^k \subset {\mathcal O}_G((\hbar))$.\footnote{${\mathcal O}_G[[\hbar]]^\vee$ may be also be viewed as the formal Rees algebra associated to the decreasing filtration ${\mathcal O}_G \supset {\mathfrak{m}}_G \supset {\mathfrak{m}}_G^2 \cdots$.} Then ${\mathcal O}_G[[\hbar]]^\vee$ is a topologically free ${\mathbb{K}}[[\hbar]]$-commutative algebra; its specialization at $\hbar=0$ is ${\mathcal O}_G[[\hbar]]^\vee / \hbar {\mathcal O}_G[[\hbar]]^\vee \simeq S({\mathfrak{g}}^*)$. The action of ${\mathfrak{g}}$ on ${\mathcal O}_G$ induces an action of ${\mathfrak{g}}$ on ${\mathcal O}_G[[\hbar]]^\vee$. Then $({\mathcal O}_G[[\hbar]]^\vee)^{\mathfrak{g}}$ is the $\hbar$-adic completion of $\sum_{k\geq 0} \hbar^{-k} ({\mathfrak{m}}_G^k)^{\mathfrak{g}}$. We have an inclusion of topologically free ${\mathbb{K}}[[\hbar]]$-algebras $({\mathcal O}_G[[\hbar]]^\vee)^{\mathfrak{g}} \subset {\mathcal O}_G[[\hbar]]^\vee$. Now the dual of the symmetrization map induces an algebra isomorphism $\wh S({\mathfrak{g}}^*) = {\mathcal O}_{\mathfrak{g}}\simeq {\mathcal O}_G$ (dual to the exponential map ${\mathfrak{g}}\to G$). This isomorphism induces a ${\mathfrak{g}}$-equivariant isomorphism of ${\mathcal O}_G[[\hbar]]^\vee$ with the $\hbar$-adic completion of $\sum_{k\geq 0} \hbar^{-k} {\mathfrak{m}}_{\mathfrak{g}}^k \subset {\mathcal O}_{\mathfrak{g}}((\hbar))$. So we have an algebra isomorphism ${\mathcal O}_G[[\hbar]]^\vee \simeq S({\mathfrak{g}}^*)[[\hbar]]$. It restricts to an isomorphism $({\mathcal O}_G[[\hbar]]^\vee)^{\mathfrak{g}} \simeq S({\mathfrak{g}}^*)^{\mathfrak{g}}[[\hbar]]$. Composing its inverse with the morphism $({\mathcal O}_G[[\hbar]]^\vee)^{\mathfrak{g}} \to U_\hbar({\mathfrak{g}}^*)$, we get the announced morphism $S({\mathfrak{g}}^*)^{\mathfrak{g}}[[\hbar]] \to U_\hbar({\mathfrak{g}}^*)$. \hfill \qed \medskip \section{The quasitriangular case} \label{sect:Sem} A quasitriangular Lie bialgebra (QTLBA) is a pair $({\mathfrak{g}},r')$, where ${\mathfrak{g}}$ is a Lie algebra and $r'\in {\mathfrak{g}}^{\otimes 2}$ is such that $\on{CYB}(r')=0$ and $t:= r'+r^{\prime 2,1} \in S^2({\mathfrak{g}})^{\mathfrak{g}}$. Any QTLBA gives rise to a coboundary Lie bialgebra $({\mathfrak{g}},r,Z)$, where $r=(r'-r^{\prime 2,1})/2$ and $Z = [t^{1,2},t^{2,3}]/4$. We call a QTLBA {\it nondegenerate} if ${\mathfrak{g}}$ is finite dimensional and $t$ is nondegenerate. Let $D : {\mathfrak{g}}^* \to {\mathfrak{g}}^*$ be the composition of the Lie cobracket $\delta : {\mathfrak{g}}^* \to \wedge^2({\mathfrak{g}}^*)$ with the Lie bracket of ${\mathfrak{g}}^*$. It is a derivation and a coderivation, and it induces a derivation of $U({\mathfrak{g}}^*)$, which we also denote by $D$ (or sometimes $D_{{\mathfrak{g}}^*}$). \begin{proposition} For any scalar $s$, $C_s := \on{Ker}(\delta - s (D\otimes \id)\circ \Delta_0)$ is a commutative subalgebra of $U({\mathfrak{g}}^*)$. \end{proposition} {\em Proof.} The condition $\ell\in C_s$ means that for any $u,v\in {\mathcal O}_{G^*}$, we have $\ell(\{u,v\} - s D^*(u)v) = 0$ (here $D^*$ is the derivation of ${\mathcal O}_{G^*}$ dual to the coderivation $D$). Let $\ell_1,\ell_2$ belong to $C_s$. Then for any $u,v\in {\mathcal O}_{G^*}$, \begin{align*} (\ell_1\ell_2)(\{u,v\} - s D^*(u)v) & = (\ell_1\otimes \ell_2)(\{\Delta(u),\Delta(v)\} - s \Delta(D^*(u)) \Delta(v)) \\ & = (\ell_1\otimes \ell_2)(\{u^{(1)},v^{(1)}\} \otimes u^{(2)}v^{(2)} + u^{(1)}v^{(1)} \otimes \{u^{(2)},v^{(2)}\} \\ & - s D^*(u^{(1)})v^{(1)} \otimes u^{(2)}v^{(2)} - u^{(1)}v^{(1)} \otimes sD^*(u^{(2)})v^{(2)}) = 0, \end{align*} hence $\ell_1\ell_2\in C_s$. Here $\Delta$ is the coproduct of ${\mathcal O}_{G^*}$. Moreover, we constructed in \cite{EGH} an element $\varrho\in {\mathfrak{m}}_{G^*}^{\wh\otimes 2}$, such that $\Delta'(u) = \varrho \star \Delta(u) \star (-\varrho)$ for any $u\in{\mathcal O}_{G^*}$; if $(U_\hbar({\mathfrak{g}}),{\mathcal R})$ is any quantization of $({\mathfrak{g}},r')$, then $\hbar\log({\mathcal R}) \in {\mathfrak{m}}_\hbar^{\wh\otimes 2}$, where ${\mathfrak{m}}_\hbar \subset U_\hbar({\mathfrak{g}})'$ is the augmentation ideal, and the reduction of $\hbar\log({\mathcal R})$ mod $\hbar$ equals $\varrho$. Then it follows from $(S^2\otimes S^2) ({\mathcal R}) = {\mathcal R}$ that $(S_{\mathcal O}^2\otimes S_{\mathcal O}^2)(\hbar\log{\mathcal R}) = \hbar\log{\mathcal R}$, where $S$ is the antipode of $U_\hbar({\mathfrak{g}})$ and $S_{\mathcal O} = S_{|U_\hbar({\mathfrak{g}})'}$ is the antipode of $U_\hbar({\mathfrak{g}})'\subset U_\hbar({\mathfrak{g}})$; since the specialization for $\hbar=0$ of $\hbar^{-1}(S_{\mathcal O}^2 -\id)$ is $D^*$, we get $(D^* \otimes \id + \id \otimes D^*)(\varrho) = 0$. Then if $\ell_1,\ell_2\in C_s$, then $(\ell_2\ell_1)(u) = (\ell_1\otimes \ell_2)(\Delta'(u)) = (\ell_1\otimes \ell_2)(\varrho \star \Delta(u) \star (-\varrho)) = (\ell_1\otimes \ell_2)(\Delta(u)) + \sum_{n\geq 1} (1/n!) (\ell_1\otimes \ell_2)(\{\varrho,\{\varrho,\ldots,\{\varrho,\Delta(u)\}\})$. Now if $f\in {\mathcal O}_{G^*}^{\wh\otimes 2}$, then $(\ell_1\otimes \ell_2) (\{\varrho,f\}) = s(\ell_1\otimes \ell_2)((D^*\otimes \id + \id \otimes D^*)(\varrho)f) = 0$. It follows that $\ell_2\ell_1 = \ell_1 \ell_2$. \hfill \qed \medskip \begin{remark} If $A$ is a quasitriangular Hopf algebra with antipode $S$, set $C_{s,A} := \{\ell\in A^* | \forall a,b\in A, \ell(ab) = \ell(bS^{-2s}(a))\}$ for any $s\in{\mathbb{Z}}$. Then it follows from \cite{Dr:coco} that $C_{s,A}$ is a commutative algebra, and that we have isomorphisms $C_s \simeq C_{s+2}$ for any $s\in{\mathbb{Z}}$. The isomorphism takes $\ell\in C_s$ to $\overline\ell\in C_{s+2}$ defined by $\overline\ell(x) = \ell(xu^{-1}S(u))$, where $u = m\circ (\id\otimes S)(R)$ ($m,R$ are the product and $R$-matrix of $A$). The definition of $C_{s,A}$ can be generalized to $s\in{\mathbb{K}}$ when $A = (U_\hbar({\mathfrak{g}}),{\mathcal R})$ is a quasitriangular QUE Hopf algebra. Define $U_\hbar({\mathfrak{g}}^*)$ as $(U_\hbar({\mathfrak{g}})')^\circ = (U_\hbar({\mathfrak{g}})^*)^\vee \supset U_\hbar({\mathfrak{g}})^*$. Then $C_{s,\hbar} := \{\ell\in (U_\hbar({\mathfrak{g}})')^\circ | \forall a,b\in U_\hbar({\mathfrak{g}})', \ell(ab) = \ell(b (S^2)^{-s}(a)) \}$ is a commutative subalgebra of $U_\hbar({\mathfrak{g}}^*)$, and its reduction modulo $\hbar$ is contained in $C_s$. In this case, $u^{-1}S(u)$ does not necessarily belong to $U_\hbar({\mathfrak{g}})'$, therefore $C_{s,\hbar}$ and $C_{s+2,\hbar}$ are not necessarily isomorphic. \end{remark} \begin{remark} If $({\mathfrak{g}},r,Z)$ is a coboundary Lie bialgebra, then $r$ is $D$-invariant iff $(\mu\otimes \id)(Z)$ is symmetric (where $\mu$ is the Lie bracket of ${\mathfrak{g}}$). Otherwise, if we set $\varrho := \rho^{2,1} \star (-\rho)$, then $(D^*\otimes \id + \id \otimes D^*)(\varrho) \neq 0$, so unless $s=0$, one cannot prove that $C_s$ is commutative. \end{remark} For each nondegenerate QTLBA $({\mathfrak{g}},r')$, Semenov-Tian-Shansky defined an algebra morphism $\Theta : Z(U({\mathfrak{g}})) \to U({\mathfrak{g}}^*)$, where $Z(A)$ denotes the center of an algebra $A$ (\cite{STS1}). Let us recall the construction of $\Theta$. There are unique Lie algebra morphisms $L,R : {\mathfrak{g}}^* \to {\mathfrak{g}}$, defined by $L(\ell) = (\ell\otimes \id)(r')$, $R(\ell) = -(\id\otimes \ell)(r')$ for any $\ell\in{\mathfrak{g}}^*$. We denote by $\alpha : U({\mathfrak{g}}^*) \to U({\mathfrak{g}})$ the composed map $U({\mathfrak{g}}^*) \stackrel{\Delta_0}{\to} U({\mathfrak{g}}^*)^{\otimes 2} \stackrel{L \otimes (S_0\circ R)}{\to} U({\mathfrak{g}})^{\otimes 2} \stackrel{m_0}{\to} U({\mathfrak{g}})$. Here $m_0,\Delta_0$ are the standard product and coproduct maps, we still denote by $L,R$ the algebra morphisms induced by $L,R$, and $S_0$ denotes the antipode of $U({\mathfrak{g}})$. The associated graded of the map $\alpha$ is the isomorphism $S({\mathfrak{g}}^*) \to S({\mathfrak{g}})$ induced by $t$, hence $\alpha$ is an isomorphism. Then $\Theta : Z(U({\mathfrak{g}})) \to U({\mathfrak{g}}^*)$ is defined as the restriction of $\alpha^{-1}$ to $Z(U({\mathfrak{g}}))$; one can prove that it is an algebra morphism. We will show, together with Proposition \ref{8:10}: \begin{proposition} \label{prop:sem} \label{8:4} $\on{Im}(\Theta) = C_1 \subset U({\mathfrak{g}}^*)$. The associated graded of $C_1$ (for the degree filtration of $U({\mathfrak{g}}^*)$) is $S({\mathfrak{g}}^*)^{\mathfrak{g}}$. \end{proposition} \begin{remark} Let $\theta$ be as in Theorem \ref{theoprinc}. The image of $\theta : S({\mathfrak{g}}^*)^{\mathfrak{g}} \to U({\mathfrak{g}}^*)$ is $\{$Poisson traces on ${\mathcal O}_{G^*}\}$, i.e., this is $\on{Ker}(\delta)$, where $\delta : U({\mathfrak{g}}^*) \to \wedge^2 U({\mathfrak{g}}^*)$ is the co-Poisson map of $U({\mathfrak{g}}^*)$. So the images of $\Theta$ and $\theta$ do not coincide. \hfill \qed \medskip \end{remark} Let us now construct a deformation $\Theta_\hbar$ of $\Theta$. The following lemma is proved in \cite{Dr:coco}. \begin{lemma} \label{8:9} Let $(A,\Delta,R)$ be a quasitriangular Hopf algebra with antipode $S$. Define a linear map $\alpha_A : A^* \to A$ by $\alpha_A(\ell) = (\ell\otimes \id)(R^{21}R)$. Then $\alpha_A$ induces an algebra morphism $C_{1,A} \to Z(A)$. \end{lemma} \begin{lemma} \label{lemma:p7} Assume moreover that $A$ is finite dimensional and $R^{2,1}R$ is nondegenerate. Then the map $C_1 \to Z(A)$ is a linear isomorphism. Its inverse induces an algebra morphism $\Theta_A : Z(A) \to A^*$. \end{lemma} {\em Proof.} We have to check that if $\ell\in A^*$ is such that $\alpha_A(\ell) \in Z(A)$, then $\ell$ is a trace. The condition $\alpha_A(\ell) \in Z(A)$ means that for any $a\in A$, we have $(\ell \otimes \id)([R^{2,1}R,1\otimes a])=0$. It follows that for any $a\in A$, we have $S^{-1}(a^{(4)}) (\ell\otimes \id)([R^{2,1}R, a^{(2)}S^{-1}(a^{(1)}) \otimes a^{(3)}]) =0$. Since $R^{2,1}R$ commutes with the image of $\Delta_A$, $(\ell\otimes \id)((a^{(2)} \otimes S^{-1}(a^{(4)}) a^{(3)}) [R^{2,1}R, S^{-1}(a^{(1)}) \otimes 1]) =0$. Therefore $(\ell\otimes \id)((a^{(2)} \otimes 1) R^{2,1}R (S^{-1}(a^{(1)}) \otimes \id)) = \varepsilon(a) (\ell\otimes \id)(R^{2,1} R)$. Since $R^{2,1}R$ is nondegenerate, this means that for any $b\in A$, we have $\ell(a^{(2)} b S^{-1}(a^{(1)})) = \varepsilon(a)\ell(b)$. Replacing $a\otimes b$ by $a^{(1)} \otimes S(a^{(2)}) b$, we get $\ell(bS^{-1}(a)) = \ell(S(a)b)$, so that $\ell\in C_1$. \hfill \qed \medskip The QUE algebra version of these lemmas is 1), 2) of the following proposition. Let $({\mathfrak{g}},r')$ be a QTLBA and let $(U_\hbar({\mathfrak{g}}),\Delta,{\mathcal R})$ be a quantization of $({\mathfrak{g}},r')$. \begin{proposition} \label{8:10} 1) The linear map $U_\hbar({\mathfrak{g}})^* \to U_\hbar({\mathfrak{g}})$, $\ell\mapsto (\ell\otimes \id)({\mathcal R}\cR^{2,1})$ extends to a map $\alpha_\hbar : U_\hbar({\mathfrak{g}}^*) \to U_\hbar({\mathfrak{g}})$. 2) If $({\mathfrak{g}},r')$ is nondegenerate, then $\alpha_\hbar$ is a linear isomorphism, and it restricts to an algebra isomorphism $C_{1,\hbar} \to Z(U_\hbar({\mathfrak{g}}))$. 3) Proposition \ref{8:4} is true. \end{proposition} {\em Proof.} Let us prove 1). Define $L_\hbar, R'_\hbar : U_\hbar({\mathfrak{g}})^* \to U_\hbar({\mathfrak{g}})$ by $L_\hbar(\xi) = (\xi\otimes \id)({\mathcal R})$, $R'_\hbar(\xi) = (\id\otimes \xi)({\mathcal R})$. According to \cite{EH1}, $\hbar\log({\mathcal R}) \subset (U_\hbar({\mathfrak{g}})'_0)^{\wh\otimes 2} \subset U_\hbar({\mathfrak{g}})'_0 \wh\otimes \hbar U_\hbar({\mathfrak{g}})_0$, so that $\log({\mathcal R}) \in U_\hbar({\mathfrak{g}})'_0 \wh\otimes U_\hbar({\mathfrak{g}})_0$. According to \cite{EGH}, appendix, the image of $\log({\mathcal R})$ in $({\mathfrak{m}}_{G^*}/{\mathfrak{m}}_{G^*}^2) \wh\otimes U({\mathfrak{g}})_0$ (by reduction mod $\hbar$ followed by projection) is $r'$. It follows that ${\mathcal R}\in U_\hbar({\mathfrak{g}})' \wh\otimes U_\hbar({\mathfrak{g}})$, therefore $L_\hbar$ extends to a map $U_\hbar({\mathfrak{g}}^*) \to U_\hbar({\mathfrak{g}})$; this map is necessarily a QUE algebra morphism. The quasitriangularity identities imply that the image of ${\mathcal R}$ in ${\mathcal O}_{G^*} \wh\otimes U({\mathfrak{g}})$ has the form $\on{exp}(\rho)$, where $\rho\in {\mathfrak{m}}_{G^*}\otimes {\mathfrak{g}}$ is a lift of $r$. It follows that the reduction mod $\hbar$ of $L_\hbar$ is the morphism induced by ${\mathfrak{g}}^*\to{\mathfrak{g}}$, $\ell \mapsto (\ell\otimes\id)(r)$. In the same way, $R'_\hbar$ extends to a (anti)morphism $U_\hbar({\mathfrak{g}}^*) \to U_\hbar({\mathfrak{g}})$. Define $\alpha_\hbar : U_\hbar({\mathfrak{g}}^*) \to U_\hbar({\mathfrak{g}})$, by $x\mapsto m \circ (L_\hbar\otimes R'_\hbar) \circ \Delta$. Then $\alpha_\hbar$ extends $\ell\mapsto (\ell\otimes \id)({\mathcal R}\cR^{2,1})$. Let us prove 2). The reduction mod $\hbar$ of $\alpha_\hbar$ is $\alpha$, which is a linear isomorphism; hence $\alpha_\hbar$ is a linear isomorphism. The second part is proved as Lemma \ref{8:9}. Let us prove Proposition \ref{8:4}. Assume that $U_\hbar({\mathfrak{g}})$ is as in \cite{EK}, hence $U_\hbar({\mathfrak{g}}) \simeq U({\mathfrak{g}})[[\hbar]]$ as algebras. Then $Z(U_\hbar({\mathfrak{g}})) \simeq Z(U({\mathfrak{g}}))[[\hbar]]$. 2) implies that $\alpha$ induces an isomorphism $(\on{mod\ }\hbar)(C_{1,\hbar}) \to Z(U({\mathfrak{g}}))$; here $(\text{mod\ }\hbar)$ is the reduction modulo $\hbar$. On the other hand, $(\text{mod\ }\hbar)(C_{1,\hbar})\subset C_1$, therefore $\Theta(Z(U({\mathfrak{g}}))) \subset C_1$. The map $\delta - (D\otimes \id)\circ \Delta_0 : U({\mathfrak{g}}^*) \to U({\mathfrak{g}}^*)^{\otimes 2}$ is filtered, and its associated graded is the dual $\delta : S({\mathfrak{g}}^*) \to \wedge^2(S({\mathfrak{g}}^*))$ of the Poisson bracket of $S({\mathfrak{g}})$. We have a surjective morphism $S({\mathfrak{g}})_{\mathfrak{g}} = S({\mathfrak{g}}) / \{{\mathfrak{g}},S({\mathfrak{g}})\} \twoheadrightarrow S({\mathfrak{g}})/\{S({\mathfrak{g}}),S({\mathfrak{g}})\}$ to the cokernel of this Poisson bracket, hence $\on{Ker}(\delta) \hookrightarrow (S({\mathfrak{g}})_{\mathfrak{g}})^* = S({\mathfrak{g}}^*)^{\mathfrak{g}}$. We have $\on{gr}(C_1) \subset \on{Ker}(\delta)$, hence $\on{gr}(C_1) \subset S({\mathfrak{g}}^*)^{\mathfrak{g}}$. Now since $\Theta$ is filtered and its associated graded takes $\on{gr}(Z(U({\mathfrak{g}}))) \simeq S({\mathfrak{g}})^{\mathfrak{g}}$ to $S({\mathfrak{g}}^*)^{\mathfrak{g}}$, we get $\on{gr}(C_1) = S({\mathfrak{g}}^*)^{\mathfrak{g}}$ and $\Theta(Z(U({\mathfrak{g}}))) = C_1$. \hfill \qed \medskip We denote by $\Theta_\hbar : Z(U_\hbar({\mathfrak{g}})) \to U_\hbar({\mathfrak{g}}^*)$ the algebra morphism inverse to $\alpha_\hbar$. $\Theta_\hbar$ is the QUE algebra version of $\Theta_A$ defined above. The image of $\Theta_\hbar$ is $C_{1,\hbar}$. When the quantization is an in \cite{EK}, $U_\hbar({\mathfrak{g}}) \simeq U({\mathfrak{g}})[[\hbar]]$, so this image is not the same as that of $\theta_\hbar$, which is $\{$traces on $U_\hbar({\mathfrak{g}})'\} = C_{0,\hbar}$. Therefore in this case, the images of $\theta_\hbar$ and $\Theta_\hbar$ do not coincide. \section{On the canonical derivation of ${\mathcal O}_{G^*}$} \label{sect:D} Let $(\a,\mu_\a,\delta_\a)$ be a finite dimensional Lie bialgebra. Then ${\mathcal O}_A$ is a Poisson-Lie group, dual to $U(\a)$. Set $D_\a := \mu_\a \circ \delta_\a$, then $D_\a$ is a derivation of $U(\a)$, such that if $U_\hbar(\a)$ is any quantization of $U(\a)$ with antipode $S$, then $D_\a = \hbar^{-1}(S^2-\id)_{|\hbar=0}$ (see \cite{Dr:coco}). It follows that the dual derivation $D_\a^*$ of ${\mathcal O}_A$ has the same property. When $\a = ({\mathfrak{g}},r')$ is a quasitriangular Lie bialgebra, $D_\a$ is inner, given by $D_\a(x) = -[\mu(r'),x]$ for any $x\in U({\mathfrak{g}})$; here $\mu$ is the Lie bracket of ${\mathfrak{g}}$ (see \cite{Dr:coco}). \begin{proposition} \label{prop:inner} If ${\mathfrak{g}}$ is a nondegenerate quasitriangular Lie bialgebra, then the derivation $D_{{\mathfrak{g}}^*}^*$ of ${\mathcal O}_{G^*}$ is inner, i.e., there exists a function $h\in {\mathcal O}_{G^*}$ such that $D_{{\mathfrak{g}}^*}^*(f) = \{h,f\}$ for any $f\in {\mathcal O}_{G^*}$. \end{proposition} {\em Proof.} We assume that ${\mathfrak{g}}$ is the double $\a_+\oplus \a_-$ of a Lie bialgebra $\a_+$ (here $\a_- = \a_+^*$); the general case is similar. Then ${\mathfrak{g}}^*$ is (as a Lie algebra) the direct sum $\a_+\oplus \a_-$. Let $A_\pm$ be the formal groups corresponding to $\a_\pm$. The morphism $\alpha : U({\mathfrak{g}}^*) \to U({\mathfrak{g}})$ is now $U(\a_+) \otimes U(\a_-) \to U({\mathfrak{g}})$, $x_+ \otimes x_- \mapsto x_- S(x_+)$. The dual morphism $\alpha^* : {\mathcal O}_{G} \to {\mathcal O}_{G^*}$ takes $F\in {\mathcal O}_G$ to $f\in {\mathcal O}_{G^*}$ given by $f(g_+,g_-) := F(g_- g_+^{-1})$. \begin{lemma} Let $D^*_{\mathfrak{g}}$, $D^*_{{\mathfrak{g}}^*}$ be the canonical derivations of ${\mathcal O}_G$ and ${\mathcal O}_{G^*}$. Then $\alpha^* \circ D^*_{{\mathfrak{g}}} = D^*_{{\mathfrak{g}}^*} \circ \alpha^*$. Moreover, $D^*_{{\mathfrak{g}}} = {\bold L}_{\mu(r')} - {\bold R}_{\mu(r')}$, where $\mu$ is the Lie bracket of ${\mathfrak{g}}$ and ${\bold L}_a f(g) = (d/d\varepsilon)_{|\varepsilon=0} F(e^{\varepsilon a}g)$, ${\bold R}_a f(g) = (d/d\varepsilon)_{|\varepsilon=0} F(ge^{\varepsilon a})$. \end{lemma} {\em Proof of Lemma.} $D_{{\mathfrak{g}}^*}$ is a coderivation, so $\Delta_0 : U({\mathfrak{g}}^*) \to U({\mathfrak{g}}^*)^{\otimes 2}$ intertwines $D_{{\mathfrak{g}}^*}$ and $D_{{\mathfrak{g}}^*} \otimes \id + \id\otimes D_{{\mathfrak{g}}^*}$; $L$ and $R$ are Lie bialgebra morphisms, so they intertwine $D_{{\mathfrak{g}}^*}$ and $D_{{\mathfrak{g}}}$; $S$ commutes with $D_{{\mathfrak{g}}}$; and $D_{\mathfrak{g}}$ is a derivation, so $m_0$ intertwines $D_{{\mathfrak{g}}} \otimes \id + \id \otimes D_{\mathfrak{g}}$ with $D_{\mathfrak{g}}$. Hence $\alpha \circ D_{{\mathfrak{g}}^*} = D_{\mathfrak{g}} \circ \alpha$. The first part follows. According to \cite{Dr:coco}, $D_{\mathfrak{g}}(x) = -[\mu(r'),x]$, which implies the second part. \hfill \qed \medskip In \cite{STS2}, the image of the Poisson bracket on $G^*$ under the formal isomorphism $\alpha : G^*\to G$ dual to $\alpha^*$ was computed. Let $f,h\in {\mathcal O}_{G^*}$ and $F = (\alpha^*)^{-1}(f)$, $H = (\alpha^*)^{-1}(h)$, then \begin{align} \label{PB:sem} (\alpha^*)^{-1}(\{f,h\})(g) & = \langle (d_{{\bold R}} - d_{\bold L}) F(g) \otimes d_{\bold L} H(g), r' \rangle + \langle (d_{{\bold R}} - d_{\bold L}) F(g) \otimes d_{\bold R} H(g), (r')^{2,1} \rangle \\ & \nonumber = \langle (d_{\bold L} - d_{\bold R}) F(g), L(d_{\bold R} H(g)) - R(d_{\bold L} H(g)) \rangle \end{align} where $g\in G$, $d_{\bold L} F(g), d_{\bold R} F(g) \in {\mathfrak{g}}^*$ are the left and right differentials defined by $\langle d_{\bold L} F(g) , a\rangle = ({\bold L}_a F)(g)$, $\langle d_{\bold R} F(g) , a\rangle = ({\bold R}_a F)(g)$ for any $a\in {\mathfrak{g}}$. \begin{lemma} \label{9:3} There exists a function $H(g)\in {\mathcal O}_G$ such that $L(d_{{\bold R}} H(g)) - R(d_{\bold L} H(g)) = \mu(r')$. \end{lemma} {\em Proof of lemma.} We prove this when ${\mathfrak{g}}$ is the double $\a_+ \oplus \a_-$ of a Lie bialgebra $\a_+$. Then set $a = (a_+,a_-)$ where $a_\pm\in \a_\pm$. We have ${\mathfrak{g}}^* = \a_+ \oplus \a_-$, and we should solve: $d_{{\bold R}} H_a(g) = \mu(r')_- + u_+(g)$, $d_{{\bold L}} H_a(g) = \mu(r')_+ + u_-(g)$, where $u_\pm(g)$ are functions $G \to \a_\pm$. Now $d_{\bold L} H(g) = \operatorname{Ad}(g) (d_{\bold R} H(g))$, hence $\mu(r')_+ + u_-(g) = \operatorname{Ad}(g)(\mu(r')_- + u_+(g))$. Let us decompose $g = g_- g_+^{-1}$, where $g_\pm \in A_\pm = \exp(\a_\pm)$, we get $\operatorname{Ad}(g_+^{-1})(u_+(g)) - \operatorname{Ad}(g_-^{-1})(u_-(g)) = \operatorname{Ad}(g_-^{-1})( \mu(r')_+) - \operatorname{Ad}(g_+^{-1})(\mu(r')_-)$. Therefore $$ u_+(g) = \operatorname{Ad}(g_+) \big( \operatorname{Ad}(g_-^{-1})(\mu(r')_+) - \operatorname{Ad}(g_+^{-1})(\mu(r')_-)\big)_+ $$ and the condition is $$ d_{\bold R} H(g) = \operatorname{Ad}(g_+) \big( \operatorname{Ad}(g_+^{-1})(\mu(r')_-)\big)_- + \operatorname{Ad}(g_+) \big( \operatorname{Ad}(g_-^{-1})(\mu(r')_+) \big)_+ , $$ i.e., \begin{equation} \label{flatness} {\bold R}_\alpha H_a(g) = \langle \mu(r')_-, \operatorname{Ad}(g_+)\big((\operatorname{Ad}(g_+^{-1})(\alpha))_+\big) \rangle + \langle \mu(r')_+, \operatorname{Ad}(g_-) \big( (\operatorname{Ad}(g_+^{-1})(\alpha))_- \big) \rangle \end{equation} for any $\alpha\in {\mathfrak{g}}$. Let us denote by $A_\alpha(g)$ the r.h.s. of (\ref{flatness}). Let us compute ${\bold R}_\alpha A_\beta - {\bold R}_\beta A_\alpha$, for $\alpha,\beta\in{\mathfrak{g}}$. Recall that $g = g_- g_+^{-1}$, then we have ${\bold R}_\alpha(g) = g \alpha$, so ${\bold R}_\alpha(g_\pm^{-1}) = \pm (\operatorname{Ad}(g_\pm^{-1})(\alpha))_\pm g_\pm^{-1}$. After computations, we find: $$ {\bold R}_\alpha A_\beta - {\bold R}_\beta A_\alpha = A_{[\beta,\alpha]} + B_{\alpha,\beta}, $$ where \begin{align*} & B_{\alpha,\beta}(g) = - \langle [(\operatorname{Ad}^*(g_+^{-1})(\beta))_+, (\operatorname{Ad}^*(g_+^{-1})(\alpha))_+], (\operatorname{Ad}(g_+^{-1})(\mu(r')_-))_-\rangle \\ & + \langle [(\operatorname{Ad}^*(g_+^{-1})(\beta))_-,(\operatorname{Ad}^*(g_+^{-1})(\alpha))_-], (\operatorname{Ad}(g_-^{-1})(\mu(r')_+))_+\rangle . \end{align*} Now for $u,v\in\a_+$, we have \begin{align*} & \langle [u,v], (\operatorname{Ad}(g_+^{-1})(\mu(r')_-))_- \rangle = \langle [u,v], \operatorname{Ad}(g_+^{-1})(\mu(r')_-) \rangle \\ & = \langle [\operatorname{Ad}^*(g_+)(u), \operatorname{Ad}^*(g_+)(v)], \mu(r')_- \rangle = \langle [\operatorname{Ad}^*(g_+)(u), \operatorname{Ad}^*(g_+)(v)], \mu(r') \rangle \\ & = \langle \operatorname{Ad}^*(g_+)(u) \otimes \operatorname{Ad}^*(g_+)(v), \delta(\mu(r')) \rangle =0, \end{align*} since $\delta(\mu(r')) = 0$ (see \cite{Dr:coco}). In the same way, the second term of $B_{\alpha,\beta}(g)$ vanishes. Hence the system (\ref{flatness}) has a solution (it is unique if we impose that $H$ vanishes at the origin). \hfill \qed \medskip {\em End of proof of Proposition \ref{prop:inner}.} Now if $h = -\alpha^*(H)$ with $H$ as in Lemma \ref{9:3} and for any $f\in{\mathcal O}_{G^*}$, we have \begin{align*} & D_{{\mathfrak{g}}^*}(f) = \alpha^*(D_{\mathfrak{g}}^*(F)) = \alpha^*(({\bold L}_{\mu(r')} - {\bold R}_{\mu(r')})(F)) \\ & = \alpha^*(\langle (d_{\bold L} - d_{\bold R})(F)(g), R(d_{\bold L} H(g)) - L(d_{\bold R} H(g)) \rangle) = \{h,f\}. \end{align*} \hfill \qed \medskip
{ "timestamp": "2005-03-25T17:52:25", "yymm": "0503", "arxiv_id": "math/0503608", "language": "en", "url": "https://arxiv.org/abs/math/0503608" }
\section{Introduction} In the search for a suitable system for quantum information processing, certain requirements have to be met \cite{04}, such as scalability of the physical system, the capability of initializing and reading out the qubits, and the possibility of having a set of universal logic gates. Neutral atoms are one of the most promising candidates for storing and processing quantum information. A qubit can be encoded in the internal or motional state of an atom, and several qubits can be entangled using atom-light interactions or atom-atom interactions. Schemes for quantum gates for neutral atoms have been theoretically proposed, that rely on dipole-dipole interactions \cite{02,qg1,qg2,qg3} or controlled collisions \cite{03,jaksch,10,12}. Such schemes can be implemented in optical lattices with a controlled filling factor, as shown in ref. \cite{bloch} where multi-particle entanglement via controlled collisions was demonstrated. Presently a major challenge is to combine controlled collisions with the loading and the addressing of individually trapped atoms. Recently techniques to confine single atoms in micron-sized \cite{07,RS,01} or larger \cite{meschede} dipole traps have been experimentally demonstrated. A set of qubits can be obtained by creating an array of such dipole traps, each one storing a single atom \cite{register}. Gate operations require the addressability of individual trapping sites and reconfigurability of the array. Arrays of dipole traps, each containing many atoms, were obtained using either arrays of micro-lenses \cite{09} or holograms \cite{salomon}. Actually, holographic techniques allow one to realize arrays of very small dipole traps \cite{grier}, which can trap single atoms. Holographic optical tweezers use a computer designed diffractive optical element to split a single collimated beam into several beams, which are then focused by a high numerical aperture lens into an array of tweezers. Recently holographic optical tweezers for individual Rubidium atoms have been implemented by using computer-driven liquid crystal Spatial Light Modulators (SLM) \cite{Bergamini2004}. The advantage of these systems is that the holograms corresponding to various arrays of traps can be designed, calculated and optimized on a computer. As a consequence, the trap array can be (slowly) controlled and reconfigured by writing these holograms on the SLM in real-time. Here we want to combine such an holographic array with a fast moving tweezer, in order to implement quantum gates based on a state selective collision between two atoms, by using a Feschbach resonance. Optimizing the control of the atoms motion is then of crucial importance, and is the subject of the present paper. \section{Quantum register with holographic dipole traps} The present approach for neutral atoms quantum gates is related to several schemes which have been proposed for trapped ions \cite{wineland,zoller}, and it uses a quantum register made of individual atoms stored in an array of holographic dipole traps. The atoms encoding the qubit will be stored in this register, which can be slowly reconfigured to move the atoms around, but does not allow fast precise motion, which is required to implement a controlled collision between two atoms. As a consequence, the register has to be combined with one (or several) fast tweezers, which can rapidly move an atom from one place to an other. There are then several options~: either there is an atom in the moving tweezer, which can be entangled and disentangled with the atoms in the register (``moving head" scheme, similar to the one proposed in ref. \cite{zoller}). One can also consider a configuration with two tweezers, which catch two atoms in the register and bring them to interaction. The fast tweezer (or tweezers) consist of a laser beam passing through an acousto-optical modulator (AOM), which allow to control simultaneously the deflection and the intensity of the beam with high accuracy. In the present paper, we will consider only two such tweezers, each containing one atom, and we will show that a quantum gate can be implemented with high fidelity by using optimal control techniques. The parameters of the calculations will be inspired by the experiment described in ref. \cite{01,07,RS,Bergamini2004}, but the scheme may work as well in a large range of parameter values. Typically, the size of the beam waist for the tweezer will be less than a micron, resulting in oscillation frequencies of 130 kHz in the radial directions, and about 30 kHz in the axial direction. In addition, we will assume that a standing wave is added along the propagation axis. This has two important consequences~: first, the axial oscillation frequency is increased up to a value which is typically close or above the radial oscillation frequency; second, it will confine the two atoms within the same ``pancake", therefore maximizing the non-linear phase shift acquired during a controlled cold collision. In the following, we will also assume that the two atoms have been prepared in the ground state of the tweezer. Though this was not implemented yet, it can in principle be done, by using either side band cooling, or evaporative cooling down to the single atom level \cite{RS}. \section{Atom transport in a time-dependent double-well potential} The transport mechanism is discussed in~\cite{calarco2004} for atoms in a time dependent, optical super lattice which has the form of a periodic array of double well potentials. Here, however, we consider a one-dimensional system with a single double well potential of the form \begin{equation} V(x,t) = -A(t) \; e^{-x^2/2w^2} - B(t) \; e^{-(x+d(t))^2/2w^2}. \label{potential} \end{equation} \begin{figure}[h!] \begin{center} \includegraphics[]{potential.eps} \end{center} \caption{ The double well potential at (a) the initial time $t=0$, (b) at an intermediate time $t=T/2$ and (c) at the end of the transport process $t=T$. The position is given in units of $a\equiv\sqrt{\hbar/m\omega}$ and the energy in units of $\varepsilon\equiv{\hbar^2/2ma^2}$ as described in the text. The horizontal, dashed lines indicate the eigenenergies of the system. The solid lines are the (real) eigenfunctions of $H(t)$ corresponding to the moving and the register atom, i.e. $\psi_2(x,t)$ and $\psi_0(x,t)$, respectively. The shown potentials correspond to experimental parameters as described in \cite{01,07,RS}. \label{fig:sequence}} \end{figure} In the geometry described above, it is sufficient to consider the one-dimensional case, where the position coordinate $x$ corresponds to the distance between the two tweezers. We will thus assume that the motional state along the two other axis does not change during the transport process to be described (this point is further discussed later in this section). The location and the depth of the minima of the potential~(\ref{potential}) is determined by the time dependent control parameters $A(t),B(t)$ and $d(t)$. The time evolution of the motional degrees of freedom of a single particle in the trap is governed by the time dependent Schr\"odinger equation \begin{equation} \mathrm{i} \hbar \frac{d}{dt}\psi(x,t) = H(t)\psi(x,t) \label{eq:SG} \end{equation} with \begin{equation} H(t) = -\frac{\hbar^2}{2m}\frac{d^2}{dx^2} + V(x,t). \end{equation} In the following, distances are measured in units of a harmonic oscillator length $a\equiv\sqrt{\hbar/m\omega}$ and energies in units of $\varepsilon\equiv{\hbar^2/2ma^2}$. In case of ${}^{87}$Rb and $\omega=2\pi\times 100$kHz this defines a length scale of $a = 34\,\mathrm{nm}$ and an energy scale of $\varepsilon = 2\pi\hbar\times50\,\mathrm{kHz}$. As in~\cite{calarco2004}, we assume that there is initially one atom in the ground state of each well and that the barrier is sufficiently high to prevent tunnelling between the wells. This allows to raise rapidly the left potential well, such that at time $t=0$ the situation depicted in figure~\ref{fig:sequence}(a) can be created: The lowest motional state of the left atom corresponds to the second excited state of the double well system while the right atom is in the ground state. The moving atom can then be adiabatically transported to the right well by lowering the left well and the barrier simultaneously (see figure~\ref{fig:sequence}(b)), and raising the barrier again while the left well is further lowered leading to the final configuration at $t=T$ shown in figure~\ref{fig:sequence}(c). During the whole process the moving atom stays always in the second excited instantaneous eigenstate of the system which corresponds eventually to the first excited state of the right well while the register atom remains in the ground state. \begin{figure}[t!] \begin{center} \includegraphics[]{pulses.eps} \end{center} \caption{% (a) The lowest instantaneous eigenenergies of the double-well system during the transport process (the ground state energy is set to zero). The transported atom corresponds to the solid line and the static atom to the zero line. (b)(c) The pulse functions $A(t)$, $B(t)$ and $d(t)$ control the depth and the distance of the two wells. In (a)-(c), the parameters for $t=0$, $t=T/2$ and $t=T$ are the same as in figure~\ref{fig:sequence}. \label{fig:pulses}} \end{figure} The adiabatic transport is possible since by choosing appropriate pulse functions $A(t),B(t)$ and $d(t)$, level crossings of the eigenenergies of the system during the process can be avoided. An example for such pulse functions, as well as the corresponding instantaneous eigenenergies of $H(t)$, are shown in figure~\ref{fig:pulses}. In figure~\ref{fig:pulses}(a) the time dependent energy of the moving atom is given by the bold line (the ground state energy is set to zero). Back transport of the moving atom to its original position is obtained by time inversion of the pulses. In order to study the dynamics of the transport process we introduce the occupation probabilities \begin{equation} P_n^A(t) = \left\vert \int dx\, {\psi^A}^*(x,t) \psi_\mathrm{n}(x,t) \right\vert^2, \end{equation} where $\psi_\mathrm{n}(x,t)$ with $n=0,1,2,...$ is the $n$th instantaneous eigenfunction of the double well potential. The superscript $A\in\{M,R\}$ indicates the wavefunction of the atom to be transported (moving atom) and the atom which is supposed to stay located at its well (register atom), i.e. $\psi^M(x,t)$ [$\psi^R(x,t)$] is the solution of the time dependent single particle Schr\"odinger equation (\ref{eq:SG}) with initial condition $\psi^M(x,0)=\psi_\mathrm{2}(x,0)$ [$\psi^R(x,0)=\psi_\mathrm{0}(x,0)$] as shown in figure~\ref{fig:sequence}(a). The fidelities of the processes are then given by $F^M \equiv P^M_2(T)$ and $F^R \equiv P^R_0(T)$. \begin{figure}[t!] \begin{center} \includegraphics[]{fidelities.eps} \end{center} \caption{ (a) Fidelity $F^M(t)$ corresponding to the moving atom during the transport for $T=250\hbar/\varepsilon$ (i), $T=350\hbar/\varepsilon$ (ii), $T=400\hbar/\varepsilon$ (iii), $T=500\hbar/\varepsilon$ (iv), $T=1000\hbar/\varepsilon$ (v), $T=5000\hbar/\varepsilon$ (vi). (b) Fidelity $F^R(t)$ corresponding to the register atom during the transport for $T=500\hbar/\varepsilon$. (c) Probabilities $P_0^M(t)$ (solid line), $P_1^M(t)$ (dash-dotted line), $P_3^M(t)$ (dashed line) and $P_4^M(t)$ (dotted line) of finding the moving atom in the respective eigenstates during its transport. \label{fig:fidelity}} \end{figure} For the example shown in Fig.~\ref{fig:pulses} we get $F^M = 99.7\%$ for propagating the moving atom wavefunction and $F^R = 99.9\%$ for propagating the register atom wavefunction from $t=0$ to $t=T=500\hbar/\varepsilon$. In the case of Rubidium this would correspond to a time $T=1.6\,\mathrm{ms}$. Figure~\ref{fig:fidelity} shows the probability $P^M_2(t)$ that the moving atom remains in the second instantaneous eigenstate during the transport for various operation times $T$. For $T=500\hbar/\varepsilon$ the fidelity is always greater than $98.7\%$ and the corresponding occupation probability $P^R_0(t)$ of the register atom (shown in figure \ref{fig:fidelity}(b)) is always larger than $99.9\%$. The corresponding probabilities of finding the moving atom in the ground, the first excited, the third excited and the fourth excited instantaneous eigenstates are displayed in figure~\ref{fig:fidelity}(c). The occupation probabilities of higher excited states are smaller than $4\times10^{-5}$ and are not shown. An excitation energy of $100$kHz, which would be required for the excitation of radial motional states, would correspond at least roughly to the eighth excited state along the axis of motion. The occupation probability of this (and higher) state(s) is found to be smaller than $5\times10^{-9}$ which justifies the one-dimensional model used in this paper. In the following we will discuss the influence of experimental imperfections, especially variations in the laser intensities, which are proportional to the pulse functions $A(t)$ and $B(t)$, and variations in the distance of the lasers creating the double well potential, which affect the pulse function $d(t)$. Motivated by the experimental conditions we assume variations of the pulse functions of the form \begin{eqnarray} \tilde d(t) &=& d(t) + \delta d \sin(\Omega t) \\ \tilde A(t) &=& A(t) + \delta A \sin(\Omega t) \\ \tilde B(t) &=& B(t) + \delta B \sin(\Omega t) \end{eqnarray} with $\Omega = 2\pi/1$ms. Assuming a drift $\delta d = 1nm$ while $\delta A=\delta B=0$ leads for $T=500\hbar/\varepsilon$ to a slight reduction of the fidelity to $F^M = 99.4\%$ and a higher fidelity $F^R$. A variation of $\delta A = 0.1\varepsilon$ while $\delta B = 0$ results in $F^M = 99.5\%$ and $F^R = 99.9 \%$. If $A(t)$ and $B(t)$ undergo the {\em same} perturbation, i.e. $\delta A = \delta B$, the shape of the potential~(\ref{potential}) does not change significantly if the variation is not too large (except for an approximately constant shift of the potential). Therefore, the level spacing as shown in figure~\ref{fig:pulses}(a) remains roughly the same and it is expected that the transport can be done as fast as without fluctuations. This behavior is confirmed by numerical simulations: Assuming $\delta A = \delta B = 10\varepsilon$, which corresponds to variations of the laser intensity of approximately $1\%$, we get the fidelities $F^M = 99.7\%$ and $F^R = 99.9\%$. This analysis shows that the current transport scheme is relatively insensitive to noise which affects both parameters, $A(t)$ and $B(t)$, in the same way, while it is more sensitive to different perturbations in these parameters. In this case, level crossings in the energy diagram~\ref{fig:pulses} can appear, leading to significant leakage into higher excited states, which would require a more sophisticated engineering to be controlled and will be a subject of future investigations. \section{Quantum gates by optimal control of molecular interactions} Performing gate operations requires a strong molecular interaction between atoms. They can be coupled to molecular states either by means of Feshbach resonances \cite{Feshbach} or through Raman photo-association laser pulses \cite{photoassociation}. For the sake of concreteness, we focus here on Feshbach resonances -- however, all of our arguments can be adapted, e.g., to Raman photo-association. We consider $^{87}$Rb atoms. Feshbach resonances occur when a bound molecular state $|n\rangle$ crosses the dissociation threshold for a state having the same quantum numbers \cite{Feshbach} while changing an external magnetic field $B$. Close to resonance, the scattering length varies as \begin{equation} A(B)=A_{bg}\left( 1-\frac{\Delta_n}{B-B_0}\right) , \end{equation}% where $A_{bg}$ is a non resonant background scattering length, $B_0$ is the resonant magnetic field, and $\Delta_n$ is the width of the resonance. The resonance energy varies almost linearly with the field \begin{equation} \varepsilon_n(B)=s_n(B-B_0), \label{resenergy} \end{equation}% with a slope $s_n$. We are interested in the dynamics of such a system in a confined geometry. Following \cite{Mies00}, we shall model it by the effective Hamiltonian \begin{equation} H_{\rm res}=\varepsilon_n(B)|n\rangle\langle n|+\sum_v(v\hbar\nu|v\rangle\langle v|+V_v|v\rangle\langle n|+{\rm h.c.}), \end{equation} where the $|v\rangle$'s are the trapped relative-motion atomic eigenstates of an isotropic harmonic oscillator trap having frequency $\nu$. The couplings to the resonance are \begin{equation} \label{couplings} V_v=2\hbar\nu \sqrt{\sqrt{4v+3}\;a_{bg}\delta_n/\pi} \end{equation} with $a_{bg}\equiv A_{bg}\sqrt{m\nu/\hbar}$, $\delta_n\equiv\Delta_n s_n/(\hbar\nu)$. In a different geometry, for instance in an elongated trap characterized by a ratio $\gamma$ between the ground level spacings in the transverse and in the longitudinal potential, the couplings can be calculated by projection on the corresponding eigenstates \cite{calarco2004}. Accurate values for the resonance parameters $\Delta_n$ and $B_0$, as well as for $A_{bg}$, are now available from both theoretical calculations and recent measurements \cite{Verhaar02}. The possibility of controlling the resonance energy via an external magnetic field, as described by Eq.~(\ref{resenergy}), provides a straightforward way to steer the interaction between the atoms. Indeed, the coupling to a specific resonant state $|n\rangle$ is only effective for a particular entrance channel, {\em i.e.} a specific combinations of atomic hyperfine states (that is, of logical qubit states in our case), while in general all other channels will be unaffected by the resonance. Thus the resonance-induced energy shift will cause a two-particle phase to appear only for that particular two-qubit computational basis state. \begin{figure}[t!] \begin{center} \includegraphics[]{gate.eps} \end{center} \caption{Two-qubit gate operation via optimal magnetic field control: optimized field time dependence (top left); overlap between initial and evolved state (bottom left); accumulated two-particle phase $\varphi$ (top right); decrease of the infidelity with increasing iterations (bottom right). \label{fig:gate}} \end{figure} We will identify our qubit logical states with the clock-transition states \begin{equation} \left\vert 0 \right\rangle \equiv \left\vert F=1,m_{F}=0\right\rangle,\quad \left\vert 1 \right\rangle \equiv \left\vert F=2,m_{F}=0\right\rangle. \end{equation}% The main advantage of this choice is that the qubit states are not sensitive to the magnetic field, and hence not subject to decoherence due to its fluctuations. We will use the resonance for the channel $|00\rangle$ occurring around $B_0=685$ G, having a width $\Delta_n=16$ mG. For obtaining a two-qubit gate, the magnetic field is ramped across $B_0$, and eventually tuned out of the Feshbach resonance again, getting the following truth table for the operation: \begin{eqnarray} \ket{00}&\rightarrow &e^{\mathrm{i} \varphi}\ket{00}, \nonumber\\ \ket{01}&\rightarrow &\ket{01}, \nonumber\\ \ket{10}&\rightarrow &\ket{10}, \nonumber\\ \ket{11}&\rightarrow &\ket{11}, \end{eqnarray} where we included the phase $\varphi$ accumulated by state $|00\rangle$ during the ramping process due to the interaction energy shift, whose value can be adjusted by controlling the magnetic field. If $\varphi=\pi$, a C-phase gate between the two atoms is obtained. Note that laser addressing of single qubits is never required throughout the procedure. The magnetic ramping process can be even performed non-adiabatically, provided that all population is finally returned to the trapped atomic ground state. This can be accomplished via a quantum optimal control technique in analogy with the above discussion for the transport process. The control parameter in this case is the external magnetic field $B$. Care has to be taken in optimizing not only the absolute value of the overlap of the final state onto the goal state, but also its phase $\varphi$. Fig.~\ref{fig:gate} shows the optimization results for a trap with transverse frequencies of 100 kHz and a longitudinal frequency of about 25 kHz, corresponding to the right well of Fig.~\ref{fig:sequence}. The final infidelity is about $2\times10^{-5}$ in this case. \section{Outlook} We have described a scheme using moving tweezers and state-dependent controlled collisions, which is able to implement a quantum gate between two individual atoms with a high fidelity. The sensitivity of the scheme to intensity or position fluctuations has been examined, and the controlled motion is found to be very tolerant to ``common-mode" noise between the two tweezers. It is even relatively tolerant to differential noise, because the overall process is close to adiabatic, and designed in such a way as to avoid unwanted level crossings. The magnetic field which is used here to obtain the Feshbach resonance would be ultimately very advantageously replaced by an optical field \cite{schlyap,grimm}, which can be switched on and off with high speed and precision, and which will not perturb the neighboring atoms stored in the holographic array. Though the overall scheme is clearly not easy to implement, optimal control techniques as used here certainly help to make it closer to realistic. \ack This research was supported by a Marie Curie Intra-European Fellowship within the 6th European Community Framework Programme, the RTN ``CONQUEST'', and by the IST/FET/QIPC projects ``QGATES'' and ``ACQP''. The Institut d'Optique group acknowledges partial support from ARDA/NSA. \vspace{0.5truecm}
{ "timestamp": "2005-03-22T12:40:11", "yymm": "0503", "arxiv_id": "quant-ph/0503180", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503180" }
\section{Introduction} One of the most remarkable lattices in Euclidean space is the Leech lattice, the unique even unimodular lattice $\Gamma _1\subset ({\mathbb{R}} ^{24}, (,)) $ of dimension 24 that does not contain vectors of square length 2. Here a lattice $\Lambda \subset ({\mathbb{R}}^n,(,))$ is called {\em unimodular}, if $\Lambda $ equals its {\em dual lattice} $$\Lambda ^{\#} := \{ x\in {\mathbb{R}} ^n \mid (x,\lambda ) \in {\mathbb{Z}} \mbox{ for}\mbox{ all } \lambda \in \Lambda \} $$ and {\em even}, if the quadratic form $x\mapsto (x,x) $ takes only even values on $\Lambda $. \cite{SieMod} studies spaces of Siegel modular forms generated by the Siegel theta-series of the 24 isometry classes of lattices in the genus of $\Gamma _1$. The present paper extends this investigation to further genera of lattices, closely related to $\Gamma _1$. A unified construction is given in \cite{RS98s}: Consider the Matthieu group $M_{23} \leq \mbox{\rm Aut} (\Gamma _1)$, where the {\em automorphism group} of a lattice $\Lambda \subset ({\mathbb{R}} ^n,(,)) $ is $\mbox{\rm Aut}(\Lambda ):= \{ g\in O(n) \mid \Lambda g = \Lambda \} .$ Let $g\in M_{23}$ be an element of square-free order $l:=|\langle g \rangle |$. Then $$l\in \{ 1,2,3,5,6,7,11,14,15,23 \} =: {\cal N} = \{ n\in {\mathbb{N}} \mid \sigma_1(n):=\sum _{d \mid n} d \mbox{ divides } 24 \} $$ and for each $l\in {\cal N}$, there is an up to conjugacy unique cyclic subgroup $\langle g \rangle \leq M_{23}$ of order $l$. Let $\Gamma _l := \{ \lambda \in \Gamma _1 \mid \lambda g = \lambda \}$ denote the fixed lattice of $g$. Then $\Gamma _l$ is an extremal strongly modular lattice of level $l$ and of dimension $2k_l$, where $$k_l:=12 \sigma_0(l)/\sigma _1(l)$$ and $\sigma_0(l)$ denotes the number of divisors of $l$. In particular $\Gamma _1$ is the Leech lattice, $\Gamma _2$ the 16-dimensional Barnes-Wall lattice and $\Gamma _3$ the Coxeter-Todd lattice of dimension 12. Let $\Lambda $ be an even lattice. The minimal $l\in {\mathbb{N}} $ for which $\sqrt{l} \Lambda ^{\# }$ is even, is called the {\em level} of $\Lambda $. Then $l\Lambda ^{\#} \subset \Lambda $. For an exact divisor $d$ of $l$ let $$\Lambda ^{\#,d} := \Lambda ^{\#} \cap \frac{1}{d} \Lambda $$ denote the {\em $d$-partial dual} of $\Lambda $. A lattice $\Lambda $ is called {\em strongly $l$-modular}, if $\Lambda $ is isometric to $\sqrt{d} \Lambda ^{\# ,d}$ for all exact divisors $d$ of the level $l$ of $\Lambda $. If $l$ is a prime, this coincides with the notion of {\em modular} lattices, which just means that the lattice is similar to its dual lattice. The Siegel theta-series $$\Theta ^{(m)} _{\Lambda } (Z) := \sum _{(\lambda _1,\ldots , \lambda _m) \in \Lambda ^m }\exp ( i \pi \mbox{\rm trace} ((\lambda _i,\lambda _j)_{i,j} Z)) $$ (which is a holomorphic function on the Siegel halfspace ${\cal H}^{(m)} = $ $ \{ Z\in \mbox{\rm Sym} _m({\mathbb{C}}) \mid $ $ \Im (Z) $ positive definite $ \} $) of a strongly $l$-modular lattice is a modular form for the $l$-th congruence subgroup $\Gamma _0^{(m)}(l)$ of $ \mbox{\rm Sp}_{2m}({\mathbb{Z}} )$ (to a certain character) invariant under all Atkin-Lehner-involutions (cf. \cite{Andrianov}). In particular for $m=1$ and $l\in {\cal N}$ the relevant ring of modular forms is a polynomial ring in 2 generators as shown in \cite{Quebmodular}, \cite{Quebstmodular}. Explicit generators of this ring allow to bound the minimum of an $n$-dimensional strongly $l$-modular lattice $\Lambda $ with $l\in {\cal N}$, $$ \min (\Lambda ) := \min _{0\neq \lambda \in \Lambda }(\lambda, \lambda ) \leq 2 + 2 \lfloor \frac{n}{2k_l} \rfloor .$$ Lattices $\Lambda $ achieving this bound are called {\em extremal}. For all $l \in {\cal N}$ there is a unique extremal strongly $l$-modular lattices of dimension $2k_l$ and this is the lattice $\Gamma _l$ described above. All the genera are presented in the nice survey article \cite{SchaSchuPi}. In this paper we investigate the spaces of Siegel modular forms generated by the Siegel theta-series of the lattices in the genus ${\cal G} (\Gamma _l)$ for $l\in {\cal N}$ using similar methods as for the case $l=1$ which is treated in \cite{SieMod}. The vector space ${\cal V} := {\cal V}({\cal G})$ of all complex formal linear combinations of the isometry classes of lattices in any genus ${\cal G}$ forms a finite dimensional commutative ${\mathbb{C}} $-algebra with positive definite Hermitian scalar product. Taking theta-series defines linear operators $\Theta ^{(m)}$ from ${\cal V}$ into a certain space of modular forms and hence a filtration of ${\cal V}$ by the kernels of these operators. This filtration behaves nicely under the multiplication and is invariant under all Hecke-operators. With the Kneser neighbouring process we construct a family of commuting self-adjoint linear operators on ${\cal V}$. Their common eigenvectors provide explicit examples of Siegel cusp forms. The genera ${\cal G} (\Gamma _l) $ ($l\in {\cal N}$) share the following properties: \begin{kor} Let $l\in {\cal N}$ and let $p$ be the smallest prime not dividing $l$. The mapping $\Theta ^{(k_l)}$ is injective on ${\cal V}({\cal G}(\Gamma _l))$. For $l\neq 7$, the construction described in \cite{BFW} (see Paragraph \ref{BFW}) gives a non-zero cusp form $\mbox{\rm BFW} (\Gamma _l,p) = \Theta ^{(k_l)}(\mbox{\rm Per} (\Gamma _l,p))$. The eigenvalue of the Kneser operator $K_2$ at the eigenvector $\mbox{\rm Per} (\Gamma _l,p) $ is the negative of the number of pairs of minimal vectors in $\Gamma _l$ which is also the minimal eigenvalue of $K_2$. \end{kor} \begin{remark} In Section \ref{results} we also list the eigenvalues of some of the operators $T(q)$ defined in Subsection \ref{Hecke}. These eigenvalues suggest that for even values of $k_l$, the cusp form $\mbox{\rm BFW} (\Gamma _l,p)$ is a generalized Duke-Imamoglu-Ikeda lift (see \cite{Ikeda}) of the elliptic cusp form of minimal weight $k_l$. \end{remark} {\bf Acknowledgement.} We thank R. Schulze-Pillot for helpful comments, suggestions and references. \section{Methods} The general method has already been explained in \cite{SieMod} (see also \cite{SchuPi0},\cite{SchuPi1}, \cite{SchuPi2} and \cite{Birch} for similar strategies). \subsection{The algebra ${\cal V} = {\cal V}({\cal G})$} Let ${\cal G}$ be a genus of lattices in the Euclidean space $({\mathbb{R}} ^{2k},(,))$. Then ${\cal G}$ is the disjoint union of finitely many isometry classes $${\cal G} = [\Lambda _1] \cup \ldots \cup [\Lambda _h ] .$$ Let ${\cal V} := {\cal V}({\cal G})\cong {\mathbb{C}} ^h$ be the complex vector space with basis ($[\Lambda _1], \ldots , [\Lambda _h])$. Let ${\cal V}_{{\mathbb{Q}} } = \langle [\Lambda _1], \ldots , [\Lambda _h] \rangle _{{\mathbb{Q}} } \cong {\mathbb{Q}} ^h$ be the rational span of the basis. The space ${\cal V}$ can be identified with the algebra ${\cal A} $ of complex functions on the double cosets $G({\mathbb{Q}}) \backslash G({\mathbb{A}}) / \mbox{\rm Stab} _{G({\mathbb{A}} )} (\Lambda _{{\mathbb{A}} } ) = \cup _{i=1}^h G({\mathbb{Q}}) x_i \mbox{\rm Stab} _{G({\mathbb{A}} )} (\Lambda _{{\mathbb{A}} }) $ where $G$ is the integral form of the real orthogonal group $G({\mathbb{R}} )= O_{2k}$ defined by $\Lambda _1$, ${\mathbb{A}} $ denotes the ring of rational ad\`eles and $\Lambda _{{\mathbb{A}}} $ the ad\'elic completion of $\Lambda _1$. If $\chi _{i}$ denotes the characteristic function mapping $G({\mathbb{Q}}) x_j \mbox{\rm Stab} _{G({\mathbb{A}} )} (\Lambda _{{\mathbb{A}} }) $ to $\delta _{ij}$ and $\Lambda _i = x_i \Lambda _1$ ($i=1,\ldots , h$) then the isomorphism maps $[\Lambda _i ]$ to $|\mbox{\rm Aut} (\Lambda _i)| \chi _i $. The usual Petersson scalar product then translates into the Hermitian scalar product on ${\cal V}$ defined by $$< [\Lambda _i], [\Lambda _j] > := \delta _{ij} |\mbox{\rm Aut} (\Lambda _i)| $$ and the multiplication of ${\cal A}$ defines a commutative and associative multiplication $\circ $ on ${\cal V}$ with $$[ \Lambda _i ] \circ [ \Lambda_j ] := \# (\mbox{\rm Aut}(\Lambda _i )) \delta_{i,j} [ \Lambda _i ]$$ (see for instance \cite[Section 1.1]{Boe}). Note that the Hermitian form $\langle , \rangle $ is associative, i.e. $$\langle v_1 \circ v_2 , v_3 \rangle = \langle v_1 , v_2 \circ v_3 \rangle \mbox{ for}\mbox{ all } v_1,v_2,v_3\in {\cal V} . $$ \subsection{The two basic filtrations of ${\cal V}$} For simplicity we now assume that ${\cal G}$ consists of even lattices. Let $l$ be the level of the lattices in ${\cal G}$. Taking the degree-$n$ Siegel theta-series $\Theta _{\Lambda _i }^{(n)} $ ($n=0,1,2,\ldots $) of the lattices $\Lambda _i$ ($i=1,\ldots , h $) then defines a linear map $$ \Theta ^{(n)} : {\cal V}\rightarrow M_{n,k}(l) \mbox{ by } \Theta ^{(n)}(\sum _{i=1 }^h c_i [ \Lambda _i] ):= \sum _{i=1}^h c_{i } \Theta _{\Lambda _i }^{(n)} $$ with values in a space of modular forms of degree $n$ and weight $k$ for the group $\Gamma _{0}^{(n)}(l)$ (see \cite{Andrianov}). For $n=0,\ldots, 2k $ let ${\cal V}_{n}:= \ker (\Theta ^{(n)})$ be the kernel of this linear map. Then we get the filtration $${\cal V}=:{\cal V}_{-1} \supseteq {\cal V}_0 \supseteq {\cal V}_1 \supseteq \ldots \supseteq {\cal V}_{2k} = \{ 0 \}$$ where ${\cal V}_0 = \{ v =\sum _{i=1}^h c_{i } [\Lambda _i ] \mid \sum _{i=1}^h c_{i } = 0 \}$ is of codimension 1 in ${\cal V}$. Clearly $\Theta ^{(n)}({\cal V}_{n-1})$ is the kernel of the Siegel $\Phi $-operator mapping $\Theta ^{(n)}({\cal V})$ onto $\Theta ^{(n-1)} ({\cal V})$. For square-free level one even has \begin{theorem} (see \cite[Theorem 8.1]{BoeSchuPi}) {\label{cusp}} If $l$ is square-free, then $\Theta ^{(n)} ({\cal V}_{n-1}) $ is the space of cusp forms in $\Theta ^{(n)} ({\cal V})$. \end{theorem} Let ${\cal W}_n := {\cal V}_n^{\perp }$ be the orthogonal complement of ${\cal V}_n$. We then have the ascending filtration $$0={\cal W}_{-1} \subseteq {\cal W}_0 \subseteq {\cal W}_1 \subseteq \ldots \subseteq {\cal W}_{2k} = {\cal V}.$$ By \cite[Proposition 2.3, Corollary 2.4]{SieMod} one has the following lemma: \begin{lemma}\label{mult} $${\cal W}_n \circ {\cal W}_m \subset {\cal W}_{n+m} \mbox{ for}\mbox{ all } m,n \in \{ -1,\ldots , 2k \} $$ and $${\cal W}_n \circ {\cal V}_m \subset {\cal V}_{m-n} \mbox{ for}\mbox{ all } m>n \in \{ -1,\ldots , 2k \} .$$ \end{lemma} Since theta-series have rational coefficients, both filtrations are rational, i.e. ${\cal V}_n = {\mathbb{C}} \otimes ({\cal V}_n \cap {\cal V}_{{\mathbb{Q}} })$ and ${\cal W}_n = {\mathbb{C}} \otimes ({\cal W}_n \cap {\cal V}_{{\mathbb{Q}} })$, hence the same statements hold when ${\cal V}$ is replaced by ${\cal V}_{{\mathbb{Q}} }$. \subsection{The Borcherds-Freitag-Weissauer cusp form}{\label{BFW}} The article \cite{BFW} gives a quite general construction of a cusp form of degree $k$. Let $\Lambda $ be a $2k$-dimensional even lattice and choose some prime $p$ such that the quadratic space $(\Lambda / p\Lambda , Q_p) $ (where $Q_p(x) := \frac{1}{2} (x,x) + p{\mathbb{Z}}$) is isometric to the sum of $k$ hyperbolic planes. Fix a totally isotropic subspace $F$ of $\Lambda / p\Lambda $ of dimension $k$. For $\lambda := (\lambda _1,\ldots, \lambda _k) \in \Lambda ^k$ we put $E(\lambda ):=\langle \lambda _1,\ldots, \lambda _k \rangle + p\Lambda $ and $S(\lambda ) := \frac{1}{p} ((\lambda _i, \lambda _j )_{i,j}) \in \mbox{\rm Sym} _k ({\mathbb{R}} )$. Define $ \epsilon (E(\lambda) ) = \epsilon (\lambda ):= (-1) ^{\dim (F\cap E(\lambda ))} $ if $E(\lambda )$ is a $k$-dimensional totally isotropic subspace of $\Lambda /p \Lambda $ and $\epsilon (E(\lambda) ) = \epsilon (\lambda ) := 0$ otherwise. \begin{defi} $\mbox{\rm BFW}(\Lambda ,p) (Z) := \sum _{\lambda \in \Lambda ^k} \epsilon(\lambda ) \exp (i \pi \mbox{\rm trace} (S(\lambda )Z )) $. \end{defi} By \cite{BFW} the form $\mbox{\rm BFW}(\Lambda, p)$ is a linear combination of Siegel theta-series of lattices in the genus of $\Lambda $: For any $k$-dimensional totally isotropic subspace $E$ of $\Lambda / p\Lambda $ let $\Gamma (E) := \langle E , p\Lambda \rangle $ be the full preimage of $E$. Dividing the scalar product by $p$, one obtains a lattice $\ ^{1/p} \Gamma (E) := (\Gamma (E) , \frac{1}{p} (,) ) \in {\cal G}$. Then we define $$\mbox{\rm Per} (\Lambda ,p):= \sum _{E} \epsilon(E) [ \ ^{1/p} \Gamma (E) ] \in {\cal V}$$ where the sum runs over all $k$-dimensional totally isotropic subspaces of $\Lambda / p\Lambda $. As $\epsilon $ is only defined up to a sign, also $\mbox{\rm Per} (\Lambda ,p) $ is only well defined up to a factor $\pm 1 $. It is shown in \cite[Theorem 2]{BFW} that $$\Theta ^{(k)} (\mbox{\rm Per} (\Lambda ,p)) = \mbox{\rm BFW} (\Lambda , p ).$$ In analogy to the notation in \cite{KV} we call $\mbox{\rm Per} (\Lambda , p)$ the {\em perestroika} of $\Lambda $. Clearly $\mbox{\rm BFW} (\Lambda ,p )$ is in the kernel of the $\Phi $-operator and hence a cusp form, if the level of $\Lambda $ is square-free by Theorem \ref{cusp}. \subsection{Hecke-actions}\label{Hecke} Strongly related to the Borcherds-Freitag-Weissauer construction are the Hecke operators $T(p)$ which define self-adjoint linear operators on ${\cal V}$ and whose action on theta series coincides with the one of $T(p)$ in \cite[Theorem IV.5.10]{Freitag} and \cite[Proposition 1.9]{Yoshida} up to a scalar factor (depending on the degree of the theta series). Assume that the genus ${\cal G}$ consists of even $2k$-dimensional lattices of level $l$. For primes $p$ not dividing $l$ we define $T(p) : {\cal V} \to {\cal V}$ by $$T(p) ([\Lambda ]):= \sum _{E} [ \ ^{1/p} \Gamma (E) ] $$ where the sum runs over all $k$-dimensional totally isotropic subspaces of $(\Lambda /p\Lambda , Q_p)$. Note that $T(p)$ is 0 if $(\Lambda /p\Lambda , Q_p)$ is not isomorphic to the sum of $k$ hyperbolic planes. The following operators commute with the $T(p)$ and are usually easier to calculate using the Kneser neighbouring-method (see \cite{Kneser}): For a prime $p$ define the linear operator $K_p$ by $$K_p ([ \Lambda ] ) := \sum _{\Gamma } [ \Gamma ], \mbox{ for}\mbox{ all } \Lambda \in {\cal G}$$ where the sum runs over all lattices $\Gamma $ in ${\cal G}$ such that the intersection $ \Lambda \cap \Gamma$ has index $p$ in $\Lambda $ and in $\Gamma $. If $p$ does not divide the level $l$ \cite[Proposition 1.10]{Yoshida} shows that the operators $K_p$ are essentially the Hecke operators $T^{(m-1)}(p^2)$ (up to a summand, which is a multiple of the identity and a scalar factor). Also if $p$ divides $l$, the operators $K_p$ are self-adjoint: For $\Lambda $ and $\Gamma $ in ${\cal G}$, the number $n(\Gamma,[\Lambda ])$ of neighbours of $\Gamma $ that are isometric to $\Lambda $ equals the number of rational matrices $X\in \mbox{\rm GL} _{2k}({\mathbb{Z}} ) \mbox{\rm diag} ( p^{-1},1^{2k-1},p) \mbox{\rm GL}_{2k }({\mathbb{Z}} )$ solving $$I(\Gamma , \Lambda ): \ \ X F_{\Gamma } X^{tr} = F_{\Lambda }$$ (where $F_{\Gamma }$ and $F_{\Lambda }$ denote fixed Gram matrices of $\Gamma $ respectively $\Lambda $) divided by the order of the automorphism group of $\Lambda $ (since one only counts lattices, $X$ and $gX$ have to be identified for all $g\in \mbox{\rm GL} _{2k}({\mathbb{Z}})$ with $g F_{\Lambda } g^{tr} = F_{\Lambda }$). Mapping $X$ to $X^{-1}$ gives a bijection between the set of solutions of $I(\Gamma , \Lambda )$ and $I(\Lambda , \Gamma )$. Therefore $$n(\Gamma , [\Lambda ]) |\mbox{\rm Aut} (\Lambda )| = n(\Lambda , [\Gamma ]) |\mbox{\rm Aut} (\Gamma )| .$$ Hence the linear operators $K_p$ and $T(p)$ generate a commutative subalgebra $${\cal H}:= \langle T(q), K_p \mid q,p \mbox{ primes }, q \teiltnicht l \rangle \leq \mbox{\rm End}^s ({\cal V}) $$ of the space of self-adjoint endomorphisms of ${\cal V}$ and ${\cal V}$ has an orthogonal basis $(d_1,\ldots , d_h)$, consisting of common eigenvectors of ${\cal H}$. For each $1\leq i \leq h$ we define $v(i)\in \{ -1,\ldots, 2k-1 \}$ by $d_i \in {\cal V}_{v(i)},\ \ d_i \not\in {\cal V}_{v(i)+1} .$ \\ Analogously let $w(i) \in \{ 0,\ldots ,2k \}$ be defined by $d_i \in {\cal W}_{w(i)},\ \ d_i \not\in {\cal W}_{w(i)-1} .$ \begin{lemma}(\cite[Lemma 2.5]{SieMod})\label{mult1} Let $1\leq i \leq h$ and assume that $d_i$ generates a full eigenspace of ${\cal H}$. Then $w(i) = v(i)+1$. \end{lemma} If the genus ${\cal G}$ is strongly modular of level $l$, by which we mean that $\sqrt{d} \Lambda ^{\# , d} \in {\cal G}$ for all $\Lambda \in {\cal G}$ and all exact divisors $d$ of $l$, then the Atkin-Lehner involutions $$ W_d: [\Lambda ] \mapsto [\sqrt{d} \Lambda ^{\# ,d } ] $$ for exact divisors $d$ of $l$ define further self-adjoint linear operators on ${\cal V}$. In this case let $$\hat{{\cal H}}:= \langle {\cal H}, W_{d} \mid d \mbox{ exact divisor of } l \rangle . $$ If all lattices in ${\cal G}$ are strongly modular, then $W_d = 1 $ for all $d$ and $\hat{\cal H} = {\cal H}$ is commutative. Again, the Hecke action is rational on ${\cal V}_{{\mathbb{Q}} }$ hence the ${\mathbb{Q}}$-algebras ${\cal H}_{{\mathbb{Q}} }$ and $\hat{\cal H}_{{\mathbb{Q}}}$ spanned by the $K_p$ respectively the $K_p$ and $W_d$ act on ${\cal V}_{{\mathbb{Q}} }$. \begin{remark} For $v\in \{ -1,0,\ldots , 2k-1 \} $ let $${\cal D}_v := \langle d_i \mid v(i) = v \rangle .$$ If all eigenspaces of ${\cal H}$ are 1-dimensional, the decomposition ${\cal V} = \oplus _{v=-1}^{2k-1} {\cal D}_v$ is preserved under any semi-simple algebra ${\cal A}$ with ${\cal H} \leq {\cal A} \leq \mbox{\rm End} (V) $ that respects the filtration. \end{remark} \section{Results}\label{results} The explicit calculations are performed in MAGMA (\cite{MAGMA}). Fix $l\in {\cal N}$, let ${\cal G}:= {\cal G}(\Gamma _l)$, ${\cal V} = {\cal V}({\cal G})$ and denote by $\Lambda _1 := \Gamma _l, \Lambda _2,\ldots, \Lambda _h$ representatives of the isometry classes of lattices in ${\cal G}$. We find that in all cases ${\cal H} = \langle K_2,K_3 \rangle \cong {\mathbb{C}} ^h$ is a maximal commutative subalgebra of $\mbox{\rm End} ({\cal V}) $. Therefore the common eigenspaces are of dimension one and it is straightforward to calculate an explicit orthogonal basis $(d_1,\ldots , d_h)$ of ${\cal V}$ consisting of eigenvectors of ${\cal H}$. In particular $v(i) = w(i) -1 $ for all $i=1,\ldots, h$ by Lemma \ref{mult1}. Here we choose $d_1 := \sum _{i=1} ^h | Aut(\Lambda _i)|^{-1} [\Lambda _i ] \in {\cal V}_0 \setminus {\cal V}_1$ to be the unit element of ${\cal V}$ and (for $l\neq 7$) $d_h = \mbox{\rm Per} (\Gamma _l , p) \in {\cal V}_{k_l-1}$, where $p$ is the smallest prime not dividing $l$. We then determine some Fourier-coefficients of the series $\Theta ^{(n)} (d_i)$ $(n=0,1,\ldots, k_l)$ to get upper bounds on $v(i)$. In all cases the degree-$k_l$ Siegel theta-series of the lattices are linearly independent hence ${\cal V}_{k_l} = \{ 0 \}$. Moreover ${\cal V}_{{k_l}-1} = \langle d_h \rangle $ if $l\neq 7$. We also know that $w(1)=0$ and we may choose $d_2$ such that $w(2) =1 $. By Lemma \ref{mult} and \ref{mult1} the product $d_j \circ d_i $ lies in $ {\cal W} _{w(i)+w(j)}$. If the coefficient of $d_h$ in the product is non-zero, this yields lower bounds on the sum $w(i)+w(j)$ which often yield sharp lower bound for $w(i)$ and $w(j)$. The method is illustrated in \cite[Section 3.2]{SieMod} and an example is given in Paragraph \ref{BW}. \subsection{The genus of the Barnes-Wall lattice in dimension 16.}\label{BW} The lattices in this genus are given in \cite{BW}. The class number is $h=24$ and we find $$\langle K_2 , K_3 \rangle = {\cal H}_{{\mathbb{Q}} } \cong {\mathbb{Q}} ^{13} \oplus F_1 \oplus F_2 \oplus F_3 $$ where the totally real number fields $F_i \cong {\mathbb{Q}}[x] / (f_i(x)) $ are given by $$\begin{array}{ll} f_1 = & x^3 - 11496x^2 + 41722560x - 47249837568 \\ f_2 = & x^3 - 1704x^2 + 400320x + 173836800 \\ f_3 = & x^5 - 11544x^4 + 42868800x^3 - 53956108800x^2 + 1813238784000x \\ & + 20094119608320000 \end{array} $$ and $\langle K_2 , K_3, W_2 \rangle = \hat{{\cal H}}_{{\mathbb{Q}} } \cong {\mathbb{Q}}^{13} \oplus \mbox{\rm Mat} _3 ( {\mathbb{Q}}) \oplus \mbox{\rm Mat} _3({\mathbb{Q}} ) \oplus \mbox{\rm Mat} _5 ({\mathbb{Q}} ).$ Let $\alpha _i$, $\beta _i$ and $\gamma _j$ ($ i=1,\ldots 3, j=1,\ldots, 5$) denote the complex roots of the polynomials $f_1$, $f_2$ respectively $f_3$. Let $\epsilon _i$ ($i=1,\ldots , 3$) denote the primitive idempotents of ${\cal H}_{{\mathbb{Q}} }$ with ${\cal H}_{{\mathbb{Q}}} \epsilon _i \cong F_i$. Since the image of ${\cal V}_{{\mathbb{Q}} }$ under $\Theta ^{(n)}$ has rational Fourier-coefficients, the functions $v$ and $w$ are constant on the eigenspaces $E_i = {\cal V} \epsilon _i $ ($i=1,2,3$). We therefore give their values in one line in the following tabular: \begin{theorem} The functions $v$ and and the eigenvalues of $ev_2$ and $ev_3$ of $K_2$ respectively $K_3$ on $(d_1,\ldots , d_{24})$ are as follows: \\ \begin{center} \begin{tabular}{|l|c|r|r|c|l|c|r|r|} \hline $i$ & $v(i)$ & $ev_2$ & $ev_3$ & & $i$ & $v(i)$ & $ev_2$ & $ev_3$ \\ \hline $1$ & $-1$ & $34560$ & $7176640$ & & $15$ & $3,4$ & $1320$ & $8640$ \\ $2$ & $0$ & $ 16200$ & $2389440$ & & $E_2$ & $4$ & $\beta _j$ & $31680$ \\ $3$ & $1$ & $8760$ & $792000$ & & $19$ & $3,4,5$ & $1080$ & $-45120$ \\ $4$ & $1$ & $7128$ & $804288$ & & $20$ & $3,4,5$ & $312$ & $4032$ \\ $E_1$ & $2$ & $\alpha _j$ & $266688$ & & $21$ & $5$ & $-216$ & $8640$ \\ $8$ & $3$ & $2664$ & $90048$ & & $22$ & $5$ & $-216$ & $20928$ \\ $9$ & $3$ & $1320$ & $77760$ & & $23$ & $6$ & $-936$ & $13248$ \\ $E_3$ & $3$ & $\gamma _j$ &$ 100800$ & & $24$ & $7$ & $-2160$ & $39360$ \\ \hline \end{tabular} \end{center} For the dimensions of ${\cal D}_v$ one finds $$ \begin{array}{|l|ccccccccc|} \hline v & -1 & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 \\ \hline \dim ({\cal D}_v) & 1 & 1 & 2 & 3 & 7\mbox{-}10 & 3\mbox{-}5 & 2\mbox{-}4 & 1 & 1 \\ \hline \end{array} $$ \end{theorem} \underline{Proof.} By explicit calculations of the Fourier-coefficients the values given in the table are upper bounds for the $v(i)$. By Lemma \ref{mult1} they also provide upper bounds on the $w(i) = v(i) + 1 $. We see that $$d_i \circ d_j = A_{ij} d_{24} + \sum _{m=1}^{23} b_{ij}^m d_m $$ with a nonzero coefficient $A_{ij}$ for the following pairs $(i,j)$: $$(23,2),\ (22,3),\ (21,4),\ (E_1,E_2),\ (E_3,E_3),\ (8,8), \ (9,9) $$ (where $(E_1,E_2)$ means that there is some vector in $E_1$ and some in $E_2$ such that this coefficient is non-zero, similarly $(E_3,E_3)$). Since $d_m \in {\cal W}_7$ for all $m\leq 23$ and $d_j\circ d_i \in {\cal W}_{w(i)+w(j)}$ the inequality $w(i)+w(j) \leq 7$ together with $A_{ij} \neq 0$ implies that $d_{24} \in {\cal W}_7$ which is a contradiction. Hence $w(i) + w(j) \geq 8$ for all pairs $(i,j)$ above. This yields equality for all values $v(i)$ and $v(j)$ for these pairs. Similarly we get $3\leq v(i) $ for $i=15,19,20$ since $A_{i,i} \neq 0$ for these $i$. \hfill{q.e.d.} \begin{Conjecture} $v(19) = 5$ and $v(20) = 5 $. \end{Conjecture} Since $d_{15}\circ d_2 = \sum _{m=1}^{18} c_m d_m + A_1 d_{19} + A_2 d_{20} $ with $A_1 \neq 0 \neq A_2$, we get $w(15) + 1 \geq \max (w(19),w(20))$. \begin{remark} If the conjecture is true, then $v(15) = 4$ and $\dim ({\cal D}_3 )= 7$, $\dim ({\cal D}_4 )= 4$, and $\dim ({\cal D}_5 )= 4$. \end{remark} Using the formula in \cite[Korollar 3]{Krieg} (resp. \cite[Proposition 1.9]{Yoshida}) we may calculate the eigenvalues of $T^{(m-1)}(3^2)$ from the one of $K_3$ and compare them with the ones given in \cite[formula (7)]{BK}. The result suggests that $\Theta ^{(2)}(d_4)$, $\Theta ^{(4)}(v)$ (for some $v\in E_3$), $\Theta ^{(6)}(d_{19})$ and $\Theta ^{(8)}(d_{24})$ are generalized Duke-Imamoglu-Ikeda-lifts (cf. \cite{Ikeda}) of the elliptic cusp forms $\delta _8 \theta _{D_4}^i$ (i=3,2,1,0) where $\delta _8 = \frac{1}{96} (\theta _{D_4}^4-\theta _{\Gamma _3})$ is the cusp form of $\Gamma _0(2)$ of weight 8 and $\theta _{D_4}$ the theta series of the 4-dimensional 2-modular root lattice $D_4$. This would imply that $v(19)=5$ and, with Lemma \ref{mult}, $v(15)=4$. \subsection{The genus of the Coxeter-Todd lattice in dimension 12.} For $l=3$ one has $h=10$, all lattices in this genus are modular, and ${\cal H} _{{\mathbb{Q}} } = \langle K_2 \rangle \cong {\mathbb{Q}} ^{10} = \hat{\cal H} _{{\mathbb{Q}} }$ \begin{theorem} There is some $a\in \{ 0,1 \}$ such that the function $v$ and the eigenvalues $ev_2$ of $K_2$ and $e_2$ of $T(2)$ are as follows: \begin{center} \begin{tabular}{|c|c|r|r|c|c|c|r|r|} \hline $i$ & $v(i)$ & $ev_2$ & $e_2$ & & $i$ & $v(i)$ & $ev_2 $ & $e_2$ \\ \hline $1$ & $-1$ & $2079$ & $151470 $ & & $6$ & $3-a$ & $234$ & $7560 $ \\ $2$ & $0$ & $1026$ & $-27540 $ & & $7$ & $3$ & $126$ & $2376 $ \\ $3$ & $1$ & $594$ & $17820 $ & & $8$ & $3$ & $-36$ & $432 $ \\ $4$ & $1$ & $432$ & $3240 $ & & $9$ & $4$ & $-144$ & $-864 $ \\ $5$ & $2$ & $288$ & $-5400 $ & & $10$ & $5$ & $-378$ & $1944 $ \\ \hline \end{tabular} \end{center} For the dimensions of ${\cal D}_v$ one finds $$ \begin{array}{|l|ccccccc|} \hline v & -1 & 0 & 1 & 2 & 3 & 4 & 5 \\ \hline \dim ({\cal D}_v) & 1 & 1 & 2 & 1+a & 3-a & 1 & 1 \\ \hline \end{array} $$ \end{theorem} We conjecture that $a=0$ but cannot prove this using Lemma \ref{mult}. The eigenvalues of $T(2)$ suggest that $\Theta ^{(2)}(d_3)$, $\Theta ^{(4)}(d_6)$ and $\Theta ^{(6)}(d_{10})$ are generalized Duke-Imamoglu-Ikeda-lifts (cf. \cite{Ikeda}) of the elliptic cusp forms $\delta _6 \theta _{A_2}^2$, $\delta _6 \theta _{A_2}$, respectively $\delta _6 $, where $\delta _6 = \frac{1}{36} (\theta _{A_2}^6-\theta _{\Gamma _3})$ is the cusp form of $\Gamma _0(3)$ of weight 6 and $\theta _{A_2}$ the theta series of the hexagonal lattice $A_2$. This would imply $v(3) = 1$, $v(6)=3$ and $v(10)=5$ and hence $a=0$. \subsection{The genus of the 5-modular lattices in dimension 8.} The class number of this genus is $h=5$, all lattices in this genus are modular, and ${\cal H} _{{\mathbb{Q}} } = \langle K_2 \rangle \cong {\mathbb{Q}} ^5 = \hat{\cal H} _{{\mathbb{Q}} }$ \begin{theorem} For $l=5$ one has $\dim ({\cal D}_{v} ) = 1 $ for $v=-1,0,1,2,3$. The function $v$ and the eigenvalues $ev_2$ of $K_2$ and $e_p$ of $T(p)$ ($p=2,3$) are given in the following table: \begin{center} \begin{tabular}{|c|c|r|r|r|c|c|c|r|r|r|} \hline $i$ & $v(i)$ & $ev_2$ & $e_2$ & $ e_3$ & & $i$ & $v(i)$ & $ev_2 $ & $e_2$ & $ e_3$ \\ \hline $1$ & $-1$ & $135$ & $270$ & $2240 $ & & $4$ & $2$ & $-8$ & $-16 $ & $56 $ \\ $2$ & $0$ & $70$ & $-120 $ & $ 160 $ & & $5$ & $3$ & $-60$ & $10$ & $ 420 $ \\ $3$ & $1$ & $42$ & $84 $ & $ 256 $ & & & & & & \\ \hline \end{tabular} \end{center} \end{theorem} \subsection{The genus of the strongly 6-modular lattices in dimension 8.} The class number of ${\cal G}(\Gamma _6)$ is $h=8$, the Hecke-algebras are $\hat{{\cal H}} _{{\mathbb{Q}} } = \langle K_2 , W_2 \rangle \cong {\mathbb{Q}}^5 \oplus \mbox{\rm Mat} _3({\mathbb{Q}} ) $ and ${\cal H}_{{\mathbb{Q}} }= \langle K_2 \rangle \cong {\mathbb{Q}}^5 + {\mathbb{Q}}[x]/(f(x))$ where $$f(x) = x^3 - 66x^2 - 216x + 31104.$$ Let $\delta _i \in {\mathbb{R}} $ ($i=1,2,3$) denote the roots of $f$. \begin{theorem} Then the function $v$ and the eigenvalues $ev_2$ of $K_2$ and $e_5$ of $T(5)$ are given in the following table: \begin{center} \begin{tabular}{|c|c|r|r|c|c|c|r|r|} \hline $i$ & $v(i)$ & $ev_2$ & $ e_5 $ & & $i$ & $v(i)$ & $ev_2$ & $ e_5 $ \\ \hline $1$ & $-1$ & $144$ & $39312 $ & & $E$ & $1$ & $\delta _j $ & $ 1872 $ \\ $2$ & $0$ & $54$ & $1872 $ & & $7$ & $2$ & $-6$ & $432 $ \\ $3$ & $1$ & $ 18 $ & $ 1008 $ & & $8$ & $3$ & $-36$ & $ 4752 $ \\ \hline \end{tabular} \end{center} Hence $\dim ({\cal D}_v) = 1$ for $v=-1,0,2,3$ and $\dim({\cal D}_1) = 4$. \end{theorem} \subsection{The genus of the 7-modular lattices in dimension 6.} The class number is $h=3$, all lattices are modular, and $\hat{\cal H} _{{\mathbb{Q}} } = {\cal H} _{{\mathbb{Q}} } = \langle K_2 \rangle \cong {\mathbb{Q}}^3$. In contrast to the other genera, the perestroika $\mbox{\rm Per} (\Gamma _7,2) $ and hence also $\mbox{\rm BFW} (\Gamma _7,2)$ vanishes due to the fact that the image of $\mbox{\rm Aut} (\Gamma _7)$ in $GO_6^+(2)$ is not contained in the derived subgroup $O_6^+(2)$. In fact, $\Theta ^{(2)} $ is already injective. Since the discriminant of the space is not a square modulo $3$ and $5$, the Hecke operators $T(3)$ and $T(5)$ vanish. \begin{theorem} We have $v(i) = i-2$ for $i=1,2,3$ and hence $\dim ({\cal D}_v) = 1 $ for $v=-1,0,1 $. The eigenvalues of $K_2$ are $35$, $19$, and $5$, the ones of $T(2)$ are $30$, $-18$, and $10$, and $T(11)$ has eigenvalues $2928$, $ -144$, and $ 248$. \end{theorem} \subsection{The genus of the strongly $l$-modular lattices in dimension 4 for $l=11,14,15$.} For $l=11,14,15$ the genus ${\cal G}(\Gamma _l)$ consists of 3 isometry classes and ${\cal H} _{{\mathbb{Q}} } = \langle K_2 \rangle \cong {\mathbb{Q}} ^3 = \hat{\cal H} _{{\mathbb{Q}} }$ since all lattices in the genus are strongly modular. \begin{theorem} For $l=11,14,15$ one has $\dim ({\cal D}_{v}) = 1 $ for $v=-1,0,1 $. The eigenvalues $ev_2$ of $K_2$ and $e_p$ of $T(p)$ for primes $p\leq 7$ not dividing $l$ are given in the following table: \begin{center} \begin{tabular}{|c|c||r|r|r|r|r||r|r|r||r|r|r|} \hline & & \multicolumn{5}{|c||}{$l=11$} & \multicolumn{3}{|c||}{$l=14$} & \multicolumn{3}{|c|}{$l=15$} \\ \hline $i$ & $v(i)$ & $ev_2$ & $e_2$ & $e_3$ & $e_5$ & $e_7$ & $ev_2$ & $e_3$ & $e_5$ & $ev_2$ & $e_2$ & $e_7$ \\ \hline $1$ & $-1$ & $9$ & $6$ & $8$ & $12$ & $16$ & $8$ & $8$ & $12$ & $9$ & $6$ & $16$ \\ $2$ & $0$ & $4$ & $-4$ & $-2 $ & $2 $ & $-4 $ & $2$ & $-4$ & $0$ & $1$ & $ -2 $ & $ 0 $ \\ $3$ & $1$ & $ -6 $ & $1$ & $3 $ & $ 7 $ & $ 6 $ & $ -4 $ & $2$ & $6$ & $ -3 $ & $2$ & $8 $ \\ \hline \end{tabular} \end{center} \end{theorem} \subsection{The genus of the $23$-modular lattices in dimension 2.} In the smallest possible dimension $2$ the genus ${\cal G}(\Gamma _{23})$ consists of only 2 isometry classes and ${\cal H} _{{\mathbb{Q}} } = \langle K_2 \rangle \cong {\mathbb{Q}} ^2 = \hat{\cal H} _{{\mathbb{Q}} }$ for the same argument that all lattices in the genus are modular. \begin{theorem} For $l=23$ one has $\dim ({\cal D}_{v}) = 1 $ for $v=-1,0$. One has $v(1) = -1$, $v(2)= 0$, $d_1 K_2 = 2 d_1 $ and $d_2 K_2 = - d_2 $. For the $T(p)$ for primes $p<23$ we find $T(2)=T(3)=T(13) = K_2$ and $T(5)=T(7)=T(11)=T(17)=T(19)=0$. \end{theorem}
{ "timestamp": "2005-03-22T08:59:06", "yymm": "0503", "arxiv_id": "math/0503447", "language": "en", "url": "https://arxiv.org/abs/math/0503447" }
\section{Introduction} Einstein-Podolsky-Rosen (EPR) \cite{einstein} suggested that quantum mechanics was incomplete and that hidden variables may be needed for completion. This was the subject of an extensive debate with Bohr \cite{bohr} who denied the existence of such hidden variables using his well known reasoning that is at the basis of the Copenhagen interpretation of quantum mechanics. A mathematical non-existence proof for these hidden variables was presented by von Neumann within the framework of quantum mechanics. Bell \cite{bellbook}, however, showed that von Neumann's proof assumed the simultaneous measurability of certain quantities that could not possibly be simultaneously measured. Subsequently Bell himself presented a non-existence proof \cite{bell} in form of inequalities that are derived by using more complex assumptions that did not necessarily include simultaneous measurability. The inequalities of Bell \cite{bell} are derived by the use of probability theory, in essence by the use of very elementary facts about random variables within the framework pioneered by Kolmogorov. The derivation of the inequalities does not involve physics or quantum mechanics, yet the inequalities have assumed an important role for the foundations of quantum mechanics. This role is a consequence of the assumptions for the random variables (and other possible variables) that are used to derive the Bell inequalities. It is commonly believed that these assumptions are needed and moreover are equivalent to the basic postulates of an ``objective local" parameter space \cite{leggett}. These postulates or conditions encompass Einstein-locality and the existence of elements of reality in the sense of Mach that co-determine the outcome of measurements. The fact that some results of quantum mechanics violate the Bell inequalities has therefore led to the belief that no objective local parameter space can exist that explains all the results of quantum mechanics. Subsequent to these discussions experiments were proposed \cite{ba}, \cite{bohm} and later realized \cite{eprex} that confirmed the theoretical result of quantum mechanics. This seemed to leave only difficult options for theoretical physics such as (i) to deny that the microscopic entities of physics have objective reality (non-existence of objective local parameters) or (ii) to assert that an influence can be propagated faster than the speed of light \cite{moore}. Additional options involving the validity of counterfactual reasoning have also been suggested and will be discussed below. We show in this paper that because of the historical sequence of events, viz. the development of Bell's theory before the performance of the actual experiments, some very important facts of probability theory have not been considered and/or misinterpreted . These facts are connected with the concept of a probability space that is important to link probability theory and mathematical statistics to the evaluation of actual experiments. We reconsider here these concepts and show that violations of the Bell inequalities have a purely mathematical reason, in particular that the Bell inequalities represent a special case of theorems given earlier by Bass \cite{bass}, Vorob'ev \cite{vorob1}, \cite{vorob} and Schell \cite{schell}. These theorems permit us to deduce that, for all the possible Bell inequalities to be valid, it is a necessary and sufficient condition that the random variables involved in their proof are defined on one common probability space. Beyond this we show that the requirement of the use of one common probability space does not follow from the requirements of objective local spaces and vice versa. In fact, we show that there exist objective local random variables that can not be defined on a common probability space and therefore do not need to obey the Bell inequalities. As mentioned, Bell has dismissed von Neumann's proof that assumed from the start the simultaneous measurability of the observables that correspond to our random variables and that can not be measured simultaneously. Our contribution here is that we also dismiss non-existence proofs of certain systems of random variables that need to be defined on one common probability space when it is clear from the outset that these systems of random variables can not be defined on any common probability space. We believe that our results give additional options of explanation for Aspect-type experiments without violating relativity or denying objective reality. \section{Mathematical model of a singlet spin state EPR experiment} As outlined above, probability theory provides a probability space and random variables and the link to the statistical treatment of the data from an actual experiment. Bell \cite{bell} considered an experimental situation advocated by Bohm and Aharonov \cite{ba}, and by Bohm and Hiley \cite {bohm}. This proposal was transformed into an actual experiment by Aspect et al. \cite{eprex} including the suggestion of Bell and others that a rapid change of the settings needed to be implemented to accomplish a delayed choice situation \cite{eprex}. We develop now an idealized experiment and a probability model for this actual experiment by Aspect et al within the framework of Kolmogorov. We note that our procedure also applies to other related experiments. We first recall the concepts of a (discrete) probability space and of a random variable defined on it. As Feller states \cite{feller}``If we want to speak about experiments or observations in a theoretical way and without ambiguity, we first must agree on the simple events representing the thinkable outcomes; {\it they define the idealized experiment}...... By definition {\it every indecomposable result of the (idealized) experiment is represented by one, and only one, sample point}. The aggregate of all sample points will be called the {\it sample space}." In our case of the idealized EPR experiment, the simple event can be chosen, for example, as the event of sending out one (and only one) correlated pair. With this event we associate an element $\omega$. In order to avoid mathematical technicalities that are not needed for the purpose of our paper we assume that the sample space $\Omega$ is at most countable. Each simple event $\omega \in \Omega$ is assigned the probability $P(\omega)$ that $\omega$ occurs. $P$ is a set function, defined for all subsets of $\Omega$, that satisfies the usual axioms, such as countable additivity and it assigns to $\Omega$ the value $P(\Omega) = 1$. The pair $(\Omega, P)$ is called a probability space. A random variable is a real-valued function on $\Omega$, but if needed it also can assume values in high-dimensional space. We now turn to the specifics of an idealized EPR-experiment. A correlated spin pair in the singlet state is sent out from a source in opposite directions toward measurement stations. These stations are characterized by certain randomly and rapidly switched settings which we denote by three dimensional unit vectors ${\bf a}, {\bf b}, {\bf c}, ...$. The measurements in the stations are mathematically represented by random variables $A = \pm 1, B = \pm 1, C = \pm 1, ...$ that may in turn be functions of other random variables e.g. a source parameter $\Lambda$ that characterizes all the properties of the particles sent out from the source. $A$ indicates that the measurements that correspond to the outcomes of random variable $A$ have been performed using the setting $\bf a$ and similarly for $B$ and $C$. We perform three categories of experiments, each with a different pair of setting vectors. The first category is characterized by the vectors $\bf a$ in station $S_1$ and $\bf b$ in station $S_2$. According to our notational convention we denote the pair of measurements $(A, B)$ and the joint probability density of $A$ and $B$ by $f_1$. Thus $f_1$ is given by \begin{equation} f_1 (+1, +1) = P(A=+1, B=+1) \text{ , }f_1(-1, +1) = P(A=-1, B=+1) \nonumber \end{equation} \begin{equation} f_1 (+1, -1) = P(A=+1, B=+-1) \text{ , }f_1(-1, -1) = P(A=-1, B=-1) \label{bvc1} \end{equation} The second category of experiments will be characterized by the vectors $\bf a$ in $S_1$ and $\bf c$ in $S_2$ with the resulting pair of measurements $(A, C)$ having density $f_2$, and the third category by the vectors $\bf b$ in $S_1$ and $\bf c$ in $S_2$ resulting in a pair of measurements $(B, C)$ with density $f_3$. The measurement outcomes on both sides need to be completely random and $\pm 1$ with equal probability, i.e. all three distributions have identical marginals. This is dictated by the rules of quantum mechanics and verified by experiment. From this it follows that the $f_i, i =1, 2, 3$ have the center of gravity for their point masses at the origin $(0, 0)$ and \begin{equation} f_i(+1,-1) + f_i(+1,+1) = \frac {1} {2} = f_i(+1,+1) + f_i(-1,+1)\text{ for } i = 1,2,3 \label{bv1} \end{equation} The idealized mathematical model with exactly the properties described above and used within the framework of Kolmogorov is the basis for all our further considerations and we call it the Ma-EPR model. \section{Ma-EPR and the theorems of Bass and Vorob'ev} We start with an example that illustrates the theorems of Bass \cite{bass} and Vorob'ev \cite{vorob} for the special case of the Ma-EPR model. The essential point of these theorems is that, in general, it is not possible to find three random variables $A, B$ and $C$, defined on a common probability space such that the three pairs $(A,B)$, $(A,C)$, and $(B,C)$ of random variables have their joint densities equal to $f_1, f_2$ and $f_3$, respectively. Hence the notation that is commonly used and that we also introduced in section 2 above is misleading in the sense that it suggests there exist three random variables $A, B$ and $C$ that can reproduce the joint densities $f_1$, $f_2$ and $f_3$, when in fact they can not. Here is a modification of an example of Vorob'ev \cite{vorob}. \begin{table}[ht] \centering \begin{tabular}{|l||r|r|r|r|}\hline & $(+1,+1)$ & $(+1,-1)$ & $(-1,+1)$ & $(-1,-1)$ \\ \hline $f_1(.,.)$ & $3/8$ & $1/8$ & $1/8$ & $3/8$\\ \hline $f_2(.,.)$ & $3/8$ & $1/8$ & $1/8$ & $3/8$\\ \hline $f_3(.,.)$ & $1/8$ & $3/8$ & $3/8$ & $1/8$\\ \hline \end{tabular} \caption{Vorob'ev-type example \cite{vorob}.}\label{TA:ma} \end{table} Clearly Eq(\ref{bv1}) holds. Suppose now that three such random variables $A, B$ and $C$ exist and are defined on one common probability space. Then the first two rows would imply that $P(A = B) = \frac {3} {4} = P(A = C)$, and so $P(B \neq C) \leq \frac {1} {2}$ , contradicting the fact that according to the third row $P(B = C) = \frac {1} {4}$. Another easy way to see that three such random variables cannot be defined on a common probability space follows from the fact that, for instance, it is not possible to assign a probability to the event $(A = 1, B = 1, C = 1)$. According to the first entry of the third row this probability could not exceed $\frac {1} {8}$. Subtracting this value from the first entry of the first row we obtain that P(A = 1, B = 1, C = -1) would have to be at least $\frac {1} {4}$. But this is in conflict with the second entry of the second row. The reason for this phenomenon is that, picturesquely speaking, the three pair distributions form a closed loop. The joint densities of $(A, B)$ and of $(A, C)$ already contain some information about the joint density of $(B, C)$. Hence we do not have complete freedom to choose the latter one. This was shown for three general pair distributions by Jean Bass \cite{bass} and independently by Schell \cite{schell} who also investigated the connection with certain problems in economics. Vorob'ev \cite{vorob1}, \cite{vorob} established necessary and sufficient conditions that any complex of distributions must possess so that these distributions can be realized as marginal distributions of a set of random variables defined on a common probability space. \begin{table}[ht] \centering \begin{tabular}{|l||r|r|r|r|}\hline & $(+1,+1)$ & $(+1,-1)$ & $(-1,+1)$ & $(-1,-1)$ \\ \hline $f_1(.,.)$ & ${\frac {1} {4}}(1 + \sigma_1)$ & ${\frac {1} {4}}(1 - \sigma_1)$ & ${\frac {1} {4}}(1 - \sigma_1)$ & ${\frac {1} {4}}(1 + \sigma_1)$\\ \hline $f_2(.,.)$ & ${\frac {1} {4}}(1 + \sigma_2)$ & ${\frac {1} {4}}(1 - \sigma_2)$ & ${\frac {1} {4}}(1 - \sigma_2)$ & ${\frac {1} {4}}(1 + \sigma_2)$\\ \hline $f_3(.,.)$ & ${\frac {1} {4}}(1 + \sigma_3)$ & ${\frac {1} {4}}(1 - \sigma_3)$ & ${\frac {1} {4}}(1 - \sigma_3)$ & ${\frac {1} {4}}(1 + \sigma_3)$\\ \hline \end{tabular} \caption{Pair densities in terms of covariances}\label{TA:ob} \end{table} It is easy to show that under the assumption of Eq(\ref{bv1}) the joint pair densities can be expressed in terms of the covariances $\sigma_i, i=1,2,3$ defined by these pair densities. The pair densities are then given by Table \ref{TA:ob} (see also the Lemma below). Note that the covariances $\sigma_i$ do not exceed $1$ in absolute value. Suppose now that there exist three random variables $A, B, C$ defined on one common probability space that reproduce the densities $f_1, f_2, f_3$ in Table \ref{TA:ob}. Then $\sigma_1 = E(AB)$, $\sigma_2 = E(AC)$, and $\sigma_3 = E(BC)$ where $E$ denotes the expectation value. Expressing the entries of Table \ref{TA:ob} in terms of the eight unknown probabilities $P(A = \pm 1, B = \pm 1, C = \pm1)$ will result in a system of twelve linear equations in these eight unknowns that can be solved in an elementary way. In particular, solving this system shows that these eight probabilities can be expressed in terms of the three covariances $\sigma_i, i = 1,2,3$. It turns out that five of these twelve linear equations are redundant. Thus this system has infinitely many solutions. Taking into account that the solutions of this system represent probabilities $P \geq 0$ we obtain in a straightforward way that the following four inequalities are necessary and sufficient conditions for the solvability of the consistency problem for the three pair distributions given in Table \ref{TA:ob}: \begin{equation} 1 + \sigma_1 + \sigma_2 + \sigma_3 \geq 0 \label{bvcc1} \end{equation} \begin{equation} 1 + \sigma_1 - \sigma_2 - \sigma_3 \geq 0 \label{bvcc2} \end{equation} \begin{equation} 1 - \sigma_1 + \sigma_2 - \sigma_3 \geq 0 \label{bvcc3} \end{equation} \begin{equation} 1 - \sigma_1 - \sigma_2 + \sigma_3 \geq 0 \label{bvcc4} \end{equation} Of course, the necessity part of this conclusion can be shown directly and trivially by modifying the standard proofs of the Bell inequality along the lines shown in \cite{bell}. \section{Bell's inequalities as a special case of Bass-Vorob'ev} Replacing the covariances $\sigma$ by the corresponding expectation values, one obtains from Eqs.(\ref{bvcc2}-\ref{bvcc3}): \begin{equation} E(AB) - E(AC) \leq 1 - E(BC) \label{bv6} \end{equation} and \begin{equation} -E(AB) + E(AC) \leq 1 - E(BC) \label{bv7} \end{equation} Eqs.(\ref{bv6}) and (\ref{bv7}) give \begin{equation} |E(AB) - E(AC)| \leq 1 - E(BC) \label{bv8} \end{equation} This is, of course, one of the celebrated Bell inequalities. Five more can be obtained in analogous fashion from Eqs.(\ref{bvcc1}-\ref{bvcc4}) giving a total of 6 (4 choose 2). These can also be obtained by cyclic permutation in Eq.(\ref{bv8}) and replacing both minus signs by plus signs. Bass \cite{bass} proved that for three general pair distributions the consistency problem can be solved if and only if the triple $(\sigma_1, \sigma_2, \sigma_3)$ considered as a point in $R^3$ belongs to a certain domain. In the special case we have been considering this domain reduces to the tetrahedron defined by the inequalities of Eqs.(\ref{bvcc1}-\ref{bvcc4}). We shall call it the covariance tetrahedron. It is diplayed in Fig. \ref{fig:Tetrahedra}. \begin{figure}[htbp] \centering \includegraphics[width=0.30\textwidth]{Tetrahedra.eps} \caption{Covariance Tetrahedron. The solid point represents a choice of values that violates the Bell inequalities.} \label{fig:Tetrahedra} \end{figure} We formulate now our findings for the Ma-EPR experiment as a theorem. We first collect a few facts of section 3 above in form of a lemma. Lemma: Let $f$ be a density supported on the four vertices $(\pm 1, \pm 1)$ of a square. Suppose that \begin{equation} f(+1, +1) + f(+1, -1) = f(+1, +1) + f(-1, +1) = \frac {1} {2} \label{bvcc5} \end{equation} Then \begin{equation} \sum x f(x, y) = \sum y f(x, y) = 0 \label{bvcc6} \end{equation} where the sums are extended over the four points $(x, y) = (\pm 1, \pm 1)$. Conversely, if $f$ satisfies Eq.(\ref{bvcc6}) then $f$ also satisfies Eq.(\ref{bvcc5}). Set \begin{equation} \sigma := \sum x y f(x, y) \label{bvcc7} \end{equation} with the same proviso for the sum. Then $f$ can be expressed in terms of $\sigma$ by the equations \begin{equation} f(+1, +1) = f(-1, -1) = {\frac {1} {4}}(1 + \sigma) \label{bvcc8} \end{equation} \begin{equation} f(-1, +1) = f(+1, -1) = {\frac {1} {4}}(1 - \sigma) \label{bvcc9} \end{equation} Theorem1: Let $f_1, f_2, f_3$ be three probability densities satisfying the hypotheses of the Lemma with corresponding covariances $\sigma_1, \sigma_2, \sigma_3$. Then the following statements are equivalent \begin{enumerate} \item[(I)] The point $(\sigma_1, \sigma_2, \sigma_3) \in R^3$ satisfies the system of inequalities Eqs.(\ref{bvcc1}-\ref{bvcc4}) and therefore belongs to the covariance tetrahedron. \item[(II)] The point $(\sigma_1, \sigma_2, \sigma_3)$ satisfies the following six Bell-type inequalities \begin{equation} |\sigma_1 - \sigma_2| \leq 1 - \sigma_3 \text{ , }|\sigma_1 + \sigma_2| \leq 1 + \sigma_3 \nonumber \end{equation} \begin{equation} |\sigma_1 - \sigma_3| \leq 1 - \sigma_2 \text{ , }|\sigma_1 + \sigma_3| \leq 1 + \sigma_2 \nonumber \end{equation} \begin{equation} |\sigma_2 - \sigma_3| \leq 1 - \sigma_1 \text{ , }|\sigma_2 + \sigma_3| \leq 1 + \sigma_1 \label{bvcc10} \end{equation} \item[(III)] There exist three random variables $A, B, C$ defined on a single common probability space with the following properties. The joint probability densities of $(A, B), (A, C)$ and $(B, C)$ are $f_1, f_2$ and $f_3$ respectively. In particular, the expectation values equal \begin{equation} E(A) = E(B) = E(C) = 0 \label{bvcc11} \end{equation} the covariances equal \begin{equation} E(AB) = \sigma_1 \text{ , }E(AC) = \sigma_2 \text{ , }E(BC) = \sigma_3 \label{bvcc12} \end{equation} Using Eqs.(\ref{bvcc10}) and (\ref{bvcc12}) one obtains the six Bell inequalities for the expectation values $E(AB), E(AC), E(BC)$. \end{enumerate} Proofs: The proof of the Lemma is straightforward. The proof that conditions (I) and (II) of Theorem1 are equivalent can be done by inspection. The proof that (III) implies (II) or the six Bell inequalities obtained from Eq.(\ref{bvcc10}) and Eq.(\ref{bvcc12}) can be carried out by a simple modification of the standard proof of the Bell inequalities \cite{bell}. Finally, the proof that (I) implies (III) was outlined at the end of section 3. We have shown therefore the following. The inequalities of Bell are a special case of the theorems of Bass and Vorob'ev for the Ma-EPR experiment. If the 6 Bell inequalities are valid then it is possible to find three random variables $A, B$ and $C$, defined on one common probability space that reproduce the three joint pair densities $f_i$ of Table \ref{TA:ob} and their covariances $\sigma_1 = E(AB), \sigma_2 = E(AC)$ and $\sigma_3 = E(BC)$. These covariances satisfy the Bell inequalities. Therefore, if quantum mechanics predicts that, for a given idealized experiment involving random variables $A, B$ and $C$ and $E(AB)$, $E(AC)$, and $E(BC)$, one of the six Bell inequalities in Eq.(\ref{bvcc10}) will be violated or equivalently if the point with coordinates $(E(AB), E(AC), E(BC)) \in R^3$ does not belong to the covariance tetrahedron of Fig. \ref{fig:Tetrahedra}, then the random variables $A, B$ and $C$, that are supposed to form the basis for the model of this idealized experiment, can not be defined on one common probability space. We note that the work of Fine \cite{fine} has already shown the importance of a joint density and therefore of a common probability space in the derivations of Bell-type inequalities. The importance of a common probability space was also stressed more recently in \cite{entrop1} and other publications. In summary, we have shown that the definability of $A, B$ and $C$ on one common probability space (OCPS) is a necessary and sufficient condition for the validity of Bell's inequalities and that this condition is of a purely mathematical nature and has nothing to do with the questions of non-locality or counterfactual reasoning that usually surround discussions of the Bell inequalities. The condition is, however, related to some of the physics of EPR experiments in a variety of ways that will be discussed in section 6. We add here that other inequalities of similar type such as the Clauser-Horne-Holt-Shimony (CHHS) \cite{chhs} inequalities can be treated similarly, although with greater algebraic exertion (16 linear equations in 16 unknowns). Their validity is again a necessary and sufficient reason that all involved random variables are defined on one common probability space. In fact, a theorem analogous to Theorem1 above holds, with the covariance tetrahedron replaced by a four-dimensional polytope. This polytope equals the intersection of the four-dimensional parallelepiped, defined by the four CHHS inequalities, and the four-dimensional cube with vertices $(\pm 1, \pm 1,\pm 1,\pm1)$. The details will be published elsewhere. \section{Bell-type proofs and Bass-Vorob'ev} In view of the OCPS condition and the theorems of Bass and Vorob'ev, the proofs for the Bell inequalities as given by Bell and others become obvious and at the same time lacking physical justification. Consider Bell's original proof \cite{bell}. Here Bell assumes that all random variables $A, B, C$ are in turn functions of a single random variable $\Lambda$. Then it is clear that $A, B, C$ are defined on one common probability space and therefore the inequalities can not be violated by the pair expectation values as explained above. It is clear that no $\Lambda$ can exist that leads to a violation of the inequalities for purely mathematical reasons as already found by Bass much earlier. Bell's physical justification is wanting because he attempts to show that the inequalities follow from the fact that $\Lambda$ does not depend on the settings ${\bf a}, {\bf b},...$. In fact, it does not matter on what $\Lambda$ depends as long as the resulting $A, B$ and $C$ are random variables defined on one probability space. We will discuss this in more detail below. Other well known proofs \cite{leggett} invoke ``counterfactual" reasoning of the following kind: If, for example, $A$ is measured given a certain information that we denote by $\lambda$ (a value that $\Lambda$ assumes in a given experiment) and that is carried by the correlated spin pair, then one could have measured with another setting, say $\bf b$ and the same $\lambda$. As we have explained in more detail previously \cite{hpnp}, it is permissible to ask the question of what would have been obtained if the measurement had been performed with a different setting. It is also permissible to hypothesize the existence of an element of reality related to that different setting if that different setting had been chosen. However, to assume then, as is always done in Bell type proofs, that all these possible different measurement results are actually contained in the data set of actual outcomes of the idealized experiment is arbitrary and against all the rules of modelling and simulation especially for the particular case of the Aspect-type experiment and all other known EPR experiments \cite{hpnp}. Naturally, we do not have to pay for all items on a restaurant's menu just because we could have chosen them. We call this latter assumption the extended counterfactual assumption (ECA). ECA is equivalent to the assumption that $A, B$ and $C$ only depend on one random variable $\Lambda$ and is therefore an assumption, not a proof. As a consequence, ECA implies that $A, B$ and $C$ are defined on one common probability space. In view of the Bass-Vorob'ev theorem it leads to a contradiction from the outset irrespective and independent of any physical considerations. \section{EPR-physics and probability spaces} A number of physical conditions have been given in the past that have been thought to be necessary and sufficient for the Bell inequalities to be valid. Most prominently among these conditions ranks the definition of an objective local parameter space \cite{peres}, \cite{leggett}. This definition involves several conditions that are automatically fulfilled in our Kolmogorovian model as has been outlined before \cite{hpnp}; it further implies the existence of elements of reality that contain information related to the spin (represented by the random variable $\Lambda$) and, most importantly Einstein locality. Armed with the knowledge that the validity of the Bell inequalities as described above is equivalent to the assumption that $A, B, C$ can be defined on one common probability space, we must now ask the question how this fact can be related to the condition of an objective local parameter space i.e. essentially to Einstein locality and the existence of elements of reality. We first deal with the question of the relation between the elements of reality that are ``carried" by the correlated spin pair and the elements $\omega$ of a probability space. Part of the work around the Bell theorem concentrates on the question whether elements of reality that determine (or at least co-determine) the outcome of the spin measurement can exist. Is not $\omega$ such an element of reality and do we not assume then its existence to start with? The answer is that the $\omega$'s represent only a necessary tool to count and average all measurements correctly. Whether or not the outcome of a single measurement is the causal consequence of an element of reality is, at this point, not discussed. The symbol $\omega$ represents just the choice of the goddess Tyche (Fortuna) for the given experiment. Of course, if an element of reality exists, $\omega$ can just represent this element. The question of whether such elements of reality can exist in nature and do explain the EPR experiments was, of course, a subject of the Einstein-Bohr debate and is also subject of our discussion here. To explore this question using the Bell inequalities we need to explore whether there exist physical reasons that demand the definition of $A, B, C$ on one probability space. \subsection{Physical reasons for definition on one probability space for source parameters only} A very important and broadly applicable physical reason for the definition of $A, B, C$ on one common probability space arises for the case in which all random variables are characterized only by the information emanating from a common source. If in addition this information is stochastically independent of the settings (delayed choice arguments), then in line with our notational convention $A, B, C$ are completely determined by one random variable $\Lambda$ corresponding to the elements of reality $\lambda$. These elements of reality can be viewed as the value the random variable $\Lambda$ assumes for the experiment that we denoted by $\omega$ i.e. we have the relation $\Lambda(\omega) = \lambda$. The settings may, of course, also be treated as random variables and may be defined on a separate probability space. However, because $\Lambda$ and the settings are stochastically independent, all random variables can be defined on one common probability space namely the product space. We have discussed details of these facts in \cite{hpnp}. Under these conditions the Bell inequalities will hold and the mathematical model obeying these conditions is in contradiction to quantum mechanics. We will show in the next section how this contradiction can be resolved by still using a classical space-time framework and just adding time and setting dependent equipment random variables in addition to the source random variable $\Lambda$. We would like to emphasize, however, that even though the system consisting of source parameters only correctly can be ruled out, this fact does not necessarily have anything to do with Einstein locality. For example, we can introduce a source parameter represented by a random variable $\Lambda_1$ that operates only if the settings ${\bf a}, {\bf b}$ and $\bf c$ are employed and $\Lambda_1$ ``knows" of these settings by action at a distance. Similarly we admit a completely different source parameter $\Lambda_2$ that operates and operates only if the three different settings ${\bf d}, {\bf e}$ and $\bf f$ are going to be chosen. Again $\Lambda_2$ ``knows" of these settings ${\bf d}, {\bf e}$, $\bf f$ by action at a distance. As long as $\Lambda_1$ is a random variable defined on some probability space and $\Lambda_2$ is a random variable defined on some possibly different probability space, the Bell inequalities formed as before for the settings ${\bf a}, {\bf b}$, $\bf c$ respectively for ${\bf d}, {\bf e}$,$\bf f$ are valid in spite of the assumption of action at a distance. Thus a contradiction exists between the results of quantum mechanics and the physical assumptions that have just been described and that appear, on the surface, to be very general. This contradiction has therefore been explained by some authors invoking violations of Einstein locality \cite{bellbook}. Others have given more reasonable, albeit noncommittal, explanations by postulating that (i) the elements of reality simply do not exist and/or (ii) there exists a ``contextuality" as discussed in \cite{peres}. Different contexts of measurements provide then different probability spaces. There were also other choices to explain the difficult situation such as (iii) counterfactual reasoning was held responsible for the difficulties \cite{peres}. As we have shown, no counterfactual reasoning is necessary to derive the inequalities and the extended counterfactual reasoning (ECA) described above is flawed from the viewpoint of mathematical modelling. We will show in the next section that explanations (i) and (ii) can, in principle, be reformulated in such a way as to have a natural explanation in the space-time of relativity. We note that (i) and (ii) contain in essence Bohr's interpretation: the spin is determined in the moment of measurement and, with respect to measurements in any of the two wings of the experiment, there is essentially the question of ``an influence on the very conditions which define the possible types of prediction..." \cite{bohr}. \subsection{A space-time interpretation of Ma-EPR that agrees with Bohr in essence} As the basis for our reasoning in this section, we will assume or postulate certain properties for the parameters and random variables of the probability theory that are in harmony with special relativity. We define with each basic experiment that corresponds to an element $\omega$ of the probability space a pair of light-cones corresponding to locations and time at which the experiments are performed as shown in Fig \ref{fig:LightCones}. \begin{figure}[htbp] \centering \includegraphics[width=0.30\textwidth]{LightCones.eps} \caption{Light cone figure} \label{fig:LightCones} \end{figure} The elements of reality and the corresponding random variables of the mathematical model are permitted to be functions of the space-time coordinates of the respective light-cones. As parameter random variables we admit not only source parameters but also equipment parameters for each measurement station. What we introduce below is a dependence of the equipment parameters of a given station on the setting vector in the light-cone of that station and an additional dependence on the time of measurement of a clock in the inertial frame of the equipment. All we need to achieve is to derive a model for Ma-EPR within the space-time of relativity that is not refuted by Bell-type inequalities and agrees with (i) and (ii) in spirit (if not the letter). For this it is only necessary to find an Einstein local model with random variables $A, B, C$ that can not be defined on one common probability space. To show that this is possible we revert to the standard notation using the settings as subscripts and denoting the functions in the two experimental wings by $A_{\bf a}, A_{\bf b}$ on one side and $B_{\bf b}, B_{\bf c}$ on the other. We continue to permit all functions to be functions of a source parameter $\Lambda$ which may have a time dependence e.g. $\Lambda$ may depend on the time of emission of the correlated pair. However, we also add equipment random variables. Of course, equipment parameters have been discussed before in many research articles. But none of them considered the role of time dependencies of these equipment parameters except our work (see \cite{hpnp}). We permit that the probability densities of these additional random variables depend on the time of measurement $t_m$ as shown by a local clock and also to depend on the local setting. We indicate this latter fact by denoting the additional random variable e.g. for setting $\bf a$ by $\Lambda_{\bf a}(t_m)$ we then have $A_{\bf a} = A_{\bf a}(\Lambda, \Lambda_{\bf a}(t_m))$ and similar for the other settings and the $B$'s on the other side e.g. $B_{\bf b} = B_{\bf b}(\Lambda, \Lambda_{\bf b}(t_m))$. Notice that all light-cones for different measurement times may contain different $\Lambda_{\bf a}(t_m)$ even though the setting is the same. No matter how a probabilistic model is conceived, different light-cones can certainly support different probability distributions for the elements of reality. Assume now that as in the Aspect-type experiment the settings on each side are sequentially changed. Because according to relativity this change of settings to take place requires a time interval of length bounded away from 0 by a positive constant $c_0$, all light-cone pairs of a sequence of measurements are different and each such experiment may be on a different probability space with a different density of the involved random variables. Furthermore, let $\eta$ be an element of a probability space that determines the random times of measurement i.e. $t_m(\eta)$ is the actual measurement time of a given experiment. We now show that physical reasons, derived from the framework of relativity only, necessitate the involvement of different probability spaces if one postulates the existence of time and setting dependent Einstein local equipment parameters. Theorem2: Assume that there exist a source parameter $\Lambda$ and equipment parameters $\Lambda_{\bf a}, \Lambda_{\bf b}$ and $\Lambda_{\bf c}$ such that $\Lambda_{\bf a}, \Lambda_{\bf b}$ and $\Lambda_{\bf c}$ not only depend on the setting vectors ${\bf a}, \bf b$ and $\bf c$, respectively, but also on the time $t_m$ of a given measurement. Here we consider $t_m$ to be a random variable $t_m = t_m(\eta)$. Thus \begin{equation} \Lambda_{\bf a} = \Lambda_{\bf a}(t_m(\eta)) \text{ , }\Lambda_{\bf b} = \Lambda_{\bf b}(t_m(\eta)) \text{ , }\Lambda_{\bf c} = \Lambda_{\bf c}(t_m(\eta)) \label{ccc1} \end{equation} The source parameter $\Lambda = \Lambda(\omega)$ is permitted to depend on emission time. We assume further that the random variables corresponding to the measurements of spin $A_{\bf a}, A_{\bf b}, B_{\bf b}$ and $B_{\bf c}$ all equal to $\pm 1$ are functions of the source parameter and of the equipment parameters $\Lambda_{\bf a}, \Lambda_{\bf b}$ and $\Lambda_{\bf c}$. Then, under the assumption that the velocity of light in vacuo is an upper limit for the velocities by which the settings can be changed, there is no probability space on which all of \begin{equation} A_{\bf a} = A_{\bf a}(\Lambda(\omega), \Lambda_{\bf a}(t_m(\eta)) \nonumber \end{equation} \begin{equation} A_{\bf b} = A_{\bf b}(\Lambda(\omega), \Lambda_{\bf b}(t_m(\eta)) \nonumber \end{equation} \begin{equation} B_{\bf b} = B_{\bf b}(\Lambda(\omega), \Lambda_{\bf b}(t_m(\eta)) \nonumber \end{equation} \begin{equation} B_{\bf c} = B_{\bf c}(\Lambda(\omega), \Lambda_{\bf c}(t_m(\eta)) \label{ccc2} \end{equation} can be consistently defined. Proof: Let $I$ be any time interval of length $|I| \leq \frac {1} {2} c_0$. Let $M$ be a measurable set in the range of $\Lambda$ and let $F, G$ and $H$ be sets in the ranges of $\Lambda_{\bf a}$, $\Lambda_{\bf b}$ and $\Lambda_{\bf c}$ respectively. Then \begin{equation} [(\omega, \eta): t_m(\eta) \in I, \Lambda(\omega) \in M, \Lambda_{\bf a}(t_m(\eta)) \in F, \Lambda_{\bf b}(t_m(\eta)) \in G, \Lambda_{\bf c}(t_m(\eta)) \in H] \label{bvcc13} \end{equation} is the impossible event and therefore has probability 0. Recall that $\omega$ signifies the sending out of a particular particle pair from the source. This result simply reflects the impossibility in the space-time of relativity to accomplish two different settings on both sides within the same short time interval and all for the same $\omega$. Hence for each time interval $I$ each of the sixteen probabilities \begin{equation} P[(\omega, \eta): t_m(\eta) \in I, A_{\bf a}(\cdot) = \pm 1, A_{\bf b}(\cdot) = \pm 1, B_{\bf b}(\cdot) = \pm 1, B_{\bf c}(\cdot) = \pm 1] = 0 \label{bvcc15} \end{equation} must vanish. Here $(\cdot)$ denotes the dependence on source and equipment parameters that in turn depend on $\omega$ and $\eta$ respectively just as in Eq.(\ref{ccc2}). Now let $J$ be a finite but arbitrarily long time interval. Then $J$ can be split up into a large but finite number $N$ of intervals $I_i, i = 1,2,...,N$ with length $|I_i| \leq \frac {1} {2} c_0$. Then the probability in Eq.(\ref{bvcc15}) with $I$ replaced by $J$ also must vanish because of finite additivity and thus $A_{\bf a}, A_{\bf b}, B_{\bf b}$ and $B_{\bf c}$ cannot be defined on a common probability space as claimed. In other words, not only must we have different probability spaces involved in the Aspect-type experiment for mathematical reasons, we must have different probability spaces for physical reasons, the requirements of relativity. We emphasize that none of the assumptions in the above proof imply any synchronization of the measurement times with certain settings. Both settings and measurement times can be chosen randomly, only the measurement times in $S_1$ and $S_2$ are correlated for any given photon pair. Note that the essence of Bohr's discussion is not violated by the above. We just need to view both spin and measurement equipment in the sense of information theory: the measurement outcome is really not the single consequence of the source information $\lambda$ that characterizes particle properties but also that of the measurement equipment and the corresponding $\lambda_{\bf a}(t_m)$ etc.. These equipment parameters correspond to the use of decoding machines in information theory \cite{shannon}. Both the source information content together with that of the decoding machines or equipment parameters (that involve different probability spaces) determine the measurement outcomes i.e. the values $\pm 1$ that the functions $A_{\bf a}$ etc. assume. In a larger sense this fulfills the spirit of Bohr. The spin does not really exist before the measurement, but only in the very moment of measurement is the outcome determined (decoded) and can not be separated from the equipment and act of measurement. The contextuality is implicitly contained in the dependence of the probability densities of the various variables on measurement time. For example, it is now incorrect to say that it makes no difference if one measures with setting $\bf b$ or setting $\bf c$ on the other side. It does make a difference because one necessarily makes these different measurements during different time intervals. The measurements in both wings are also performed at the same clock-time or at least at correlated clock-times which opens the possibility of correlations between the two wings even though the settings are randomly chosen. What is the meaning then of the Aspect et al. \cite{eprex} experiment in view of the above discussions? If one assumes that this experiment is free from any problems related to non-ideal experimental conditions and if one assumes that a space-time explanation must be possible then the Aspect et al. experiment has proven the existence of setting and time dependent equipment parameters. \section{Conclusion} We have shown that the inequalities of Bell can be derived as special cases of a more general theorem found by Bass ten years earlier. We have further shown that the Bell inequalities are valid if and only if the three random variables involved can actually be defined on a common probability space. As a consequence the Bell theorem is correct at least for the following systems of hidden variables, in the sense that these systems can be ruled out: \begin{enumerate} \item{} Source parameter $\Lambda$ only, \item{} Source parameter $\Lambda$ and equipment parameters $\Lambda_{\bf a}$, $\Lambda_{\bf b}$ and $\Lambda_{\bf c}$ that depend only on the respective settings. \end{enumerate} On the other hand, equipment parameters that depend on the measurement times as well as on the respective instrument settings can not be ruled out. A space-time explanation of the Aspect et al. experiment is therefore not ruled out by Bell's inequalities. Any such space-time explanation can not rely on source parameters only but must involve a certain type of time and setting dependent equipment parameter random variables. Thus, the validity of the Bell inequalities for objective local parameter spaces has not been proven by any of the proofs reported in the literature \cite{bellbook}, \cite{leggett}, \cite{peres}. \section{Acknowledgement} The authors would like to thank M. Aschwanden for creating the figures of the manuscript and helpful suggestions. Support of the Office of Naval Research (N00014-98-1-0604) is gratefully acknowledged.
{ "timestamp": "2005-03-03T22:01:05", "yymm": "0503", "arxiv_id": "quant-ph/0503044", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503044" }
\section{Introduction} To obtain information about a system, a measurement has to be made. Based on the results of this measurement we assign to the system our state of knowledge. For a classical system this state takes the form of a probability distribution $P(x',t)$, while for a quantum system we have a state matrix $\rho(t)$. \footnote{Here we are not concerned with where the division between classical systems and quantum systems occurs. Instead we recognize that both descriptions are valid and the system dynamics determine which is appropriate.} In this paper we are concerned with efficient simulation techniques for {\em partly} observed systems; that is, systems for which the observer cannot obtain enough information to assign the system a pure state, $P(x',t)=\delta[x'-x(t)]$ or $\rho(t)=\ket{\psi(t)}\bra{\psi(t)}$. The chief motivation for wishing to know the conditional state of a system is for the purpose of feedback control \cite{Jac93,Bel87,WisMil94,DohJac99}. That is because for cost functions that are additive in time, the optimal basis for controlling the system is the observer's state of knowledge about the system. Even if such a control strategy is too difficult to implement in practice, it plays the important role of bounding the performance of any strategy, which helps in seeking the best practical strategy. It is well known that the quantum state of an open quantum system, given continuous-in-time measurements of the bath, follows a stochastic trajectory through time \cite{BelSta92}. In the quantum optics community this is referred to as a quantum trajectory \cite{Car93a,GarParZol92,MolCasDal93,WisMil93a,GoeGra93,GoeGra94,Wis96,GamWis01,WisDio01,GarZol00}. The form of this trajectory can be either jump-like in nature or diffusive depending on how we choose to measure the system; that is, the arrangement of the measuring apparatus. In this paper we review quantum trajectory theory for partially observed systems by presenting a simple model: A three level atom that emits into two separate environments, only one of which is accessible to our detectors. Such partially observed systems cannot be described by a stochastic Schr\"odinger~~equation (SSE) \cite{Car93a, GarParZol92,MolCasDal93}, but rather requires a more general form of a quantum trajectory that has been called a stochastic master equation (SME) \cite{WisMil93a}. This is an instance of the fact that the most general form of quantum measurement theory requires the full Kraus representation of operations \cite{Kra83,BraKha92}, rather than just measurement operators \cite{BraKha92}. It is also well known that if we have a classical system and we make measurements on it with a measurement apparatus that has associated with it a Gaussian noise, then the evolution of this classical state in the continuous-in-time limit obeys a Kushner-Stratonovich equation (KSE) \cite{Mcg74}. To review these dynamics for partially observed classical systems we present the KSE for a system that experiences an `internal' unobservable white noise process. That is, the evolution in the absence of the measurements is given by a Fokker-plank equation \cite{Gar85}. This is the classical analogue to the quantum master equation. The new work in this paper is a simple numerical technique that allows us to reduce the numerical resources required to calculate the continuous-in-time trajectories. This method relies on the implementation of linear or `ostensible' \cite{Wis96} measurement theory, classical \cite{Mcg74} and quantum \cite{GoeGra93,GoeGra94,Wis96,GamWis01}. For the classical case our method reduces the problem from solving the KSE for the probability distribution to simulating the ensemble average of two coupled stochastic differential equations (SDE). For the quantum case our method reduces the problem from solving a conditional SME to simulating the ensemble average of a SSE plus a c-number SDE. Thus in both the classical and the quantum case, our method reduces the size of the problem by a factor of $N,$ the number of basis states required to represent the system. Recently Brun and Goan \cite{Brun} have used a similar idea to investigate a partially observed quantum system. However, since they did not use measurement theory with ostensible probabilities, their claim that they can generate a typical trajectory conditioned on some partial record ${\bf R}$ is not valid. This is demonstrated in detail in \ref{AppendixQuant}. (In their method, the record ${\bf R}$ can only be generated randomly, and can be found only by doing the ensemble average over the fictitious noise, but that is not the issue of concern here.) Finally, we combine these theories to consider the following case: a quantum system is monitored continuously in time by a classical system but we only have access to the results of non-ideal measurements performed on the classical system. Note that such joint systems have recently been studied by Warszawski {\em et al} \cite{WarWisMab01,WarWis03a,WarGamWis04} and Oxtoby {\em et al} \cite{Neil}. Warszawski {\em et al} considered continuous-in-time monitoring of a quantum optical system with realistic photodections while Oxtoby {\em et al} considered continuous-in-time monitoring of a quantum solid-state system with a quantum point contact. We show that our ostensible numerical technique can be applied to these types of systems, greatly simplifying the simulations. The format of this paper is as follows. In Secs.~\ref{measQ} and \ref{measC} we review quantum and classical measurement theory respectively. This is essential as it allows us to define both the notations and the physical insight that will be used throughout this paper. In Secs.~\ref{quantum}, \ref{class}, and \ref{both} we investigate the above mentioned quantum, classical and joint systems respectively, and present our ostensible numerical technique for each specific case. Finally in Sec.~\ref{con} we conclude with a discussion. \section{Quantum Measurement theory (QMT)}\label{measQ} \subsection{General theory} In quantum mechanics the most general way we can represent the state of the system is via a state matrix $\rho(t)$. This is a positive semi-definite operator that acts in the system Hilbert space ${\cal H}_{\rm s}$. In this paper we take the view that this represents our state of knowledge of the system. Taking this view allows us to simply interpret the ``collapse of the wavefunction'', upon measurement, as an update in the observer's knowledge of the system \cite{CavFucSch02,Fuc02}. If we now assume that we have a measurement apparatus that allows us to measure observable $R$ of the system, then the conditional state $\rho_{r}(t')$ of the system given result $r$ is determined by \cite{Kra83} \begin{equation}\label{QuantumUpdate} \rho_{r}(t')=\frac{\hat{\cal O}_r(t',t)\rho(t)} {P(r,t')}, \end{equation} where $P(r,t')$ is the probability of getting result $r$ at time $t'=t+T$, where $T$ is the measurement duration time. Here $\hat{\cal O}_r(t',t)$ is known as the operation of the measurement and is a completely positive superoperator and for efficient measurements can be defined by \begin{equation}\label{OperationDef} \hat{\cal O}_r(t',t)\rho(t)=\hat{\cal J}[\hat{M}_{r}(T)]\rho(t)=\hat{M}_{r}(T)\rho(t)\hat{M}_{r}^\dagger(T), \end{equation} where $\hat{M}_r(T)$ is called a measurement operator. The probability of getting result $r$ is given by \begin{eqnarray}\label{QuantumProb} P(r,t')={\rm Tr}[\hat{\cal O}_r(t',t)\rho(t)]={\rm Tr}[\hat{F}_r(T)\rho(t)], \end{eqnarray} where the set $\{\hat{F}_r(T)=\hat{M}_r^\dagger(T)\hat{M}_r(T)\}$ is the positive operator measure (POM) for observable $R$. By completeness, the sum of all the POM elements satisfies \begin{equation}\label{QuantumMeas} \sum_r \hat{F}_r(T)=\hat{1}. \end{equation} So far we have only considered efficient, or purity-preserving measurements. That is if $\rho(t)$ was initially $\ket{\psi(t)}\bra{\psi(t)}$ then the state after the measurement would also be of this form. In a more general theory we must dispense with the measurement operator $\hat{M}_r(T)$ and define the Kraus operator $\hat{K}_{r,f}(T)$ \cite{Kra83}. This has the effect of changing the definition of the operation of the measurement $\hat{\cal O}_r(t',t)$ [\erf{OperationDef}] to \begin{equation}\label{OperationDefComplete} \hat{\cal O}_r(t',t)=\sum_f\hat{\cal J}[\hat{K}_{r,f}(T)], \end{equation} and the POM elements for this measurement are now given by \begin{equation} \label{KrausEffec} \hat{F}_{r}(T)=\sum_f\hat{K}_{r,f}^\dagger(T)\hat{K}_{r,f}(T). \end{equation} Note $\hat{F}_{r}(T)$ still satisfies the completeness condition [\erf{QuantumMeas}]. We can think of $f$ as labelling results of fictitious measurement. If one is only interested in the average evolution of the system, this can be found via \begin{equation} \label{AverageO} \rho(t')=\sum_r P(r)\rho_r(t')=\hat{\cal O}(t',t)\rho(t), \end{equation} where $\hat{{\cal O}}(t',t)=\sum_r\hat{\cal O}_r(t',t)$ is the non-selective operation. \subsection{Quantum trajectory theory} Quantum trajectory theory is simply quantum measurement theory applied to a continuous in-time monitored system \cite{Car93a,GarParZol92,WisMil93a,MolCasDal93,GoeGra93,GoeGra94,Wis96,GamWis01,WisDio01,GarZol00}. In continuous monitoring, repeated measurements of duration $T=d t$ are performed on the system. This results in the state being conditioned on a record ${\bf R}}%_{(0,t]}$, which is a string containing the results $r_k$ of each measurement from time 0 to $t$ but not including time 0. Here the subscript $k$ refers to a measurement completed at time $t_{k}=k d t$. From the record ${\bf R}}%_{(0,t]}$, the conditioned state at time $t$ can be written as \begin{equation} \rho_{\bf R}(t)=\frac{\tilde\rho_{\bf R}(t)}{{P}({\bf R}}%_{(0,t]})}, \end{equation} where $\tilde\rho_{\bf R}(t)$ is an unnormalized state defined by \begin{equation} \tilde\rho_{\bf R}(t)=\hat{\cal O}_{r_{k}}(t_k,t_{k-1})\ldots\hat{\cal O}_{r_2}(t_2,t_1)\hat{\cal O}_{r_1}(t_1,0)\rho(0). \end{equation} The probability of observing the record ${\bf R}}%_{(0,t]}$ is \begin{equation} P({\bf R}}%_{(0,t]})={\rm Tr}[\tilde\rho_{{\bf R}}%_{(0,t]}}(t)]. \end{equation} If we now assume that the coupling between the apparatus (bath) and the system is Markovian then the average state \begin{eqnarray} \rho(t)=\hat{\cal O}(t_k,t_{k-1})\ldots\hat{\cal O}(t_2,t_1)\hat{\cal O}(t_1,0)\rho(0) \end{eqnarray} is equivalent to the reduced state \begin{equation} \rho_{\rm red}(t)={\rm Tr}_{\rm bath}[\ket{\Psi(t)}\bra{\Psi(t)}], \end{equation} which itself obeys the Master equation \cite{Lin76} \begin{equation}\label{QuantumMaster} \dot{\rho}(t)=\hat{\cal L}\rho(t)=-i[\hat{H},\rho(t)]+\sum_j\gamma_j\hat{\cal D}[\hat{L}_j]\rho(t). \end{equation} Here $\hat{\cal D}[\hat{A}]$ is the superoperator defined by \begin{equation}\label{DampSuper} \hat{\cal D}[\hat{A}]\rho=\hat{A}\rho\hat{A}^\dagger-\hat{A}^\dagger\hat{A}\rho/2-\rho\hat{A}^\dagger\hat{A}/2, \end{equation} and represents dissipation of information about the system into the baths. \subsection{Fictitious quantum trajectories: the ostensible numerical technique} \label{OstensibleQ} If the system is only partly observed ($f$ in \erf{OperationDefComplete} represents the unobservable processes) this state will be mixed. This is not a problem for simple systems but for a large system a numerical simulation for $\rho_{{\bf R}}(t)$ would be impractical. This brings us to the goal of this section which is to demonstrate that $\rho_{{\bf R}}(t)$ can be numerically simulated by using SSEs, requiring less space to store on a computer. To do this we assume that a fictitious measurement with record ${\bf F}}%_{(0,t]}$ is actually made on the unobservable process. Then we can expand $\rho_{{\bf R}}(t)$ to \begin{equation}\label{eq29} \rho_{{\bf R}}(t)=\sum_{{\bf F}}%_{(0,t]}}\rho_{{\bf R, F}}(t){ P}({\bf F}}%_{(0,t]}|{\bf R}}%_{(0,t]}), \end{equation} where \begin{equation}\label{eq30} \rho_{{\bf R, F}}(t)=\ket{{\psi}_{{\bf R, F}}(t)}\bra{{\psi}_{{\bf R, F}}(t)}. \end{equation} Here $\ket{{\psi}_{{\bf R, F}}(t)}$ is a normalised state conditioned on both ${\bf F}}%_{(0,t]}$ and ${\bf R}}%_{(0,t]}$. In quantum trajectory theory this is defined as \begin{equation}\label{eq31} \ket{{\psi}_{{\bf R, F}}(t)}=\frac{\ket{\tilde{\psi}_{{\bf R, F}}(t)}}{\rt{{P}({\bf F}}%_{(0,t]},{\bf R}}%_{(0,t]})}}, \end{equation} where \begin{equation}\label{eq32} \ket{\tilde{\psi}_{{\bf R, F}}(t)}=\hat{M}_{r_{k},f_{k}}(dt)....\hat{M}_{r_{1},f_{1}}(dt)\ket{\psi(0)}. \end{equation} Here $r_{k}$ and $f_{k}$ are the results of the measurement operator \begin{equation}\label{eq33} \hat{M}_{r_k,f_k}(dt)=\bra{r_k}\bra{f_k}\hat{U}(t_k,t_{k-1})\ket{0}\ket{0}, \end{equation} where $\ket{0}\ket{0}$ is the initial bath state. This indicates that given that we have a real record ${\bf R}}%_{(0,t]}$ we can calculate $\rho_{{\bf R}}(t)$ from averaging over an ensemble of pure states $\ket{{\psi}_{{\bf R, F}}(t)}$. But as shown in \ref{AppendixQuant} the fact that future real results are not necessarily independent from the current fictitious results means that we cannot generate single trajectories without knowing the full solution. However by using quantum measurement theory with ostensible distributions we can get around this problem. Under ostensible quantum trajectory theory \cite{GoeGra94,Wis96,GamWis01} we can define a state, $\ket{\bar{\psi}_{{\bf R, F}}(t)}$ as, \begin{equation}\label{eq38} \ket{\bar{\psi}_{{\bf R, F}}(t)}=\frac{\ket{\tilde{\psi}_{{\bf R, F}}(t)}}{\rt{\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})}}, \end{equation} where $\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})$ is an ostensible probability distribution. This is simply a guessed distribution that only has the requirement that it be a probability distribution and be non-zero when $P({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})$ is non-zero. Note this state is no longer normalized to one and this is why we signify it with the bar. The true probability can be related to the ostensible probability by \begin{equation}\label{eq40} {P}({\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]})=\bra{\bar{\psi}_{{\bf R, F}}(t)}{{\bar{\psi}_{{\bf R, F }}(t)}}\rangle\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]}), \end{equation} which is a generalized Girsanov transformation \cite{BelSta92,GoeGra94,Wis96,GamWis01,GatGis91}. Going back to \erf{eq29} and using the above equations we can write $\rho_{{\bf R}}(t)$ as \begin{equation}\label{eq42} \rho_{{\bf R}}(t)=\frac{\sum_{{\bf F}}%_{(0,t]}}\ket{\bar{\psi}_{{\bf R, F}}(t)}\bra{\bar{\psi}_{{\bf R, F}}(t)}\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})}{{ P}({\bf R}}%_{(0,t]})}, \end{equation} where \begin{equation}\label{eq43} {P}({\bf R}}%_{(0,t]})=\sum_{{\bf F}}%_{(0,t]}}\langle\bar{\psi}_{{\bf R}, {\bf F}}(t) \ket{\bar{\psi}_{{\bf R, F}}(t)} \Lambda({\bf F}}%_{(0,t]},{\bf R}}%_{(0,t]}). \end{equation} Note that the sum containing $\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})$ in the above equations simply represents the ensemble average over all possible fictitious records. Thus we can rewrite \erf{eq42} as \begin{equation}\label{MainEquation} \rho_{{\bf R}}(t)=\frac{{\rm E}_{\bf F} \Big{[}\ket{\bar{\psi}_{{\bf R, F}}(t)}\bra{\bar{\psi}_{{\bf R, F}}(t)}\Big{]}} {{\rm E}_{\bf F}\Big{[}\langle\bar{\psi}_{{\bf R}, {\bf F}}(t) \ket{\bar{\psi}_{{\bf R, F}}(t)}\Big{]}}. \end{equation} \section{Classical Measurement theory (CMT)} \label{measC} \subsection{General theory} In this paper when considering what we call a classical system, we are referring to a system that can be described by the probability distribution $P(x,t)$ (i.e a vector of probabilities) rather than a state matrix. That is, with respect to a fixed basis $x$ the coherences (off diagonal elements) are always zero. If we now measure observable $R$ of the system, then after a measurement which yielded result $r$, the state of the system is given by \cite{Bayes} \begin{equation}\label{BayesTheorem} P_r(x,t)=\frac{P(r,t|x,t)P(x,t)}{P(r,t)}, \end{equation} where \begin{equation}\label{Pr} P(r,t)=\int dx P(r,t|x,t)P(x,t). \end{equation} This is known as Bayes' theorem. Here $ P_r(x,t)\equiv P(x,t|r,t)$ is called a conditional state and represents our new state of knowledge given that we observed result $r$. Here we have only considered minimally disturbing classical measurements. That is, there is no back action acting on the system in the measurement process. To generalize Bayes' theorem to deal with measurements which incur back action we mathematically split the measurement into a two stage process. The first is the Bayesian update, followed by a second stage described by $ B_r(x',t'|x,t)$, the probability for the measurement to cause the system to make a transition from $x$ at time $t$ to $x'$ at time $t'=t+T$, given the result $r$. Thus for all $x'$, $x$ and $r$ \begin{eqnarray}\label{Bpropeties1} B_r(x',t'|x,t)&\geq& 0,\\ \int dx' B_r(x',t'|x,t)&=& 1\label{Bpropeties}. \end{eqnarray} Now by defining the operation \begin{eqnarray} {\cal O}_r(x',t'|x,t)&=& B_r(x',t'|x,t)P(r,t|x,t) \end{eqnarray} the conditional system state after the measurement becomes \begin{equation}\label{GeneralBayessTheorem} P_r(x',t')=\frac{ \int dx {\cal O}_r(x',t'|x,t)P(x,t)}{P(r,t')}, \end{equation} where \begin{eqnarray}\label{Pr2} P(r,t')&=&\int dx'\int dx {\cal O}_r(x',t'|x,t) P(x,t). \end{eqnarray} Using \erf{Bpropeties} this can be rewritten as \begin{eqnarray}\label{Pr3} P(r,t')&=&\int dx F_r(x,t) P(x,t), \end{eqnarray}where $F_r(x,t)=P(r,t|x,t)$, which by definition satisfies \begin{equation} \sum_r F_r(x,t)=1, \end{equation} is the classical analogue of the POM element. The average evolution of the system is given by \begin{eqnarray}\label{Ave} P(x',t')&=&\sum_{r}P_r(x',t')P(r,t')\nonumber\\&=&\int dx {\cal O}(x',t'|x,t)P(x,t), \end{eqnarray} where ${\cal O}(x',t'|x,t)=\sum_r{\cal O}_r(x',t'|x,t)$ is the non-selective operation. Note that for any $B_r(x',t'|x,t)$ that satisfies \erfs{Bpropeties1}{Bpropeties} we can rewrite it as \begin{equation}\label{Bansatz} B_r(x',t'|x,t)=\sum_f \delta [x'-x_{r,f}(t')]P(f,t'|x,t;r,t), \end{equation} where $x_{r,f}(t')$ is the new system configuration $x'$ at time $t'$ given the measurement result $r$ and extra noise $f$ (the stochastic part of the back action). The parameter $f$ is analogous to the fictitious measurement results in the quantum case. Thus the operation for the measurement can be written as \begin{eqnarray}\label{orf} {\cal O}_r(x',t'|x,t)&=& \sum_f \delta [x'-x_{r,f}(t')]P(f,t';r,t|x,t),\nonumber\\ &=&\sum_f {\cal J}_{r,f}(x',t'|x,t). \end{eqnarray} This is the classical equivalent of \erf{OperationDefComplete}. In the above we have purposely structured QMT and CMT so that the theories appear to be similar and as a general rule we will push this point of view throughout the rest of this paper. However, it is important to point out the key differences between these theories. In the quantum case we can always write the measurement operator (or Kraus operator) as $\hat{M}_r=\hat{U}_r\rt{\hat{F}_r}$ where $U_r$ is a unitary operator. That is we can always interpret a measurement as a two stage process, where $\rt{\hat{F}_r}$ is responsible for the wavefunction collapse and the gain in information by the observer and $\hat{U}_r$ is some extra evolution that entails no information gain (as the entropy of the system is not changed by this evolution). It simply adds surplus back action to the system. In the classical case we can also write the measurement as a two stage process. However, the first process by definition has no back action; it is simply the update in the observer's knowledge of the system. Furthermore the second stage is not necessary unitary evolution (and as such can change the entropy of the system). Thus back action in the quantum and classical case are physically different processes and one can not separate all the back action in the quantum case from the observer's information gain. Mathematically speaking, the difference arises from the fact that a quantum state is represented by a positive {\em matrix}, the state matrix, while a classical state is represented by a positive {\em vector}, the vector of probabilities. \subsection{Classical trajectory theory} To achieve continuous-in-time measurements theory for a classical system we simply let the measurement time tend to $dt$ and extend the number of consecutive measurements to $t/dt$. Then the state of the classical system given the measurement record ${\bf R}}%_{(0,t]}$ is \begin{equation} P_{{\bf R}}%_{(0,t]}}(x,t)=\frac{\tilde{P}_{{\bf R}}%_{(0,t]}}(x,t)}{{P}({\bf R}}%_{(0,t]})}, \end{equation} where $\tilde{P}_{{\bf R}}%_{(0,t]}}(x,t)$ is an unnormalized state defined by \begin{eqnarray} &&\hspace{-.8cm} \tilde{P}_{{\bf R}}%_{(0,t]}}(x,t)=\int dx_{k-1}...\int dx_1\int dx_0 \nn \\ &&\times{\cal O}_{r_{k}}(x,t |x_{k-1},t_{k-1}) \ldots {\cal O}_{r_{2}}(x_2,t_2 |x_1,t_1) \nn \\ &&\times{\cal O}_{r_{1}}(x_1,t_1|x_0,0)P(x_0,0). \end{eqnarray} The probability of observing the record ${\bf R}}%_{(0,t]}$ is \begin{equation} P({\bf R}}%_{(0,t]})=\int dx \tilde{P}_{{\bf R}}%_{(0,t]}}(x,t). \end{equation} If we now assume that the noise added by the measurement apparatus is white, and the form of the back action is independent of the results ${\bf R}$, then the unconditional state \begin{eqnarray} &&\hspace{-.8cm} {P}(x,t)= \int dx_{k-1}...\int dx_1\int dx_0 \nn \\ &&\times{\cal O}(x,t |x_{k-1},t_{k-1}) \ldots {\cal O}(x_2,t_2 |x_1,t_1) \nn \\ &&\times{\cal O}(x_1,t_1|x_0,0)P(x_0,0) \end{eqnarray} is the solution of the Fokker Plank Equation \cite{Gar85} \begin{equation}\label{Eq.FPE2} \partial_t P(x,t)=-\partial_{x}[A(x,t) P(x,t)]+\smallfrac{1}{2}\partial^2_{x}[ D^2(x,t)P(x,t)], \end{equation} where $A(x,t)$ determines the amount of drift and $D(x,t)$ determines the amount of diffusion. \subsection{Fictitious classical trajectories: The ostensible numerical technique}\label{sec.fitcla} The basic principle behind this technique is that we assume that the unobservable process, ${\bf F}}%_{(0,t]}$, that generates the back action part of the measurement is fictitiously simulated. To be more specific we can define \begin{eqnarray}\label{ptilde} &&\hspace{-.8cm} \tilde{P}_{{\bf R, F}}%_{(0,t]}}(x,t)=\int dx_{k-1}...\int dx_1\int dx_0 \nn \\ &&\times{\cal J}_{r_{k},f_{k}}(x,t |x_{k-1},t_{k-1}) \ldots {\cal J}_{r_{2},f_{2}}(x_2,t_2 |x_1,t_1) \nn \\ &&\times{\cal J}_{r_{1},f_{1}}(x_1,t_1|x_0,0)P(x_0,0). \end{eqnarray} where ${\cal J}_{r,f}(x',t'|x,t)$ is defined implicitly in \erf{orf}. From this the conditional state, $P_{{\bf R}}%_{(0,t]}}(x,t)$, is given by \begin{equation}\label{conclass} {P}_{{\bf R}}%_{(0,t]}}(x,t)= \sum_{{\bf F}}%_{(0,t]}}{P}_{{\bf R, F}}%_{(0,t]}}(x,t){{P}({{\bf F}}%_{(0,t]}|{\bf R}}%_{(0,t]}})}, \end{equation} where \begin{equation} {P}_{{\bf R, F}}%_{(0,t]}}(x,t)=\frac{\tilde{P}_{{\bf R, F}}%_{(0,t]}}(x,t)}{{P}({{\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]}})}. \end{equation} But as in the quantum case this cannot be directly calculated and as a result we must use an ostensible theory. We define the ostensible state by \begin{equation}\label{pbar} \bar{P}_{{\bf R, F}}%_{(0,t]}}(x,t)=\frac{\tilde{P}_{{\bf R, F}}%_{(0,t]}}(x,t)}{{\Lambda}({{\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]}})}, \end{equation} and the true probability can be related to the ostensible by \begin{equation} {{P}({{\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]}})}=\int dx \bar{P}_{{\bf R, F}}%_{(0,t]}}(x,t){{\Lambda}({{\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]}})}, \end{equation} the classical Girsanov transformation. Using the above we can rewrite \erf{conclass} as \begin{equation}\label{Pxgrvialin} {P}_{{\bf R}}%_{(0,t]}}(x,t)= \frac{\sum_{{\bf F}}%_{(0,t]}}\bar{P}_{{\bf R, F}}%_{(0,t]}}(x,t){{\Lambda}({{\bf F}}%_{(0,t]},{\bf R}}%_{(0,t]}})}}{P({\bf R}}%_{(0,t]})}, \end{equation} where \begin{equation} {P({\bf R}}%_{(0,t]})}=\sum_{{\bf F}}%_{(0,t]}}\int dx \bar{P}_{{\bf R, F}}%_{(0,t]}}(x,t){{\Lambda}({{\bf F}}%_{(0,t]},{\bf R}}%_{(0,t]}})}. \end{equation} As in the quantum case we can rewrite \erf{Pxgrvialin} as \begin{equation}\label{EPxgrvialin} {P}_{{\bf R}}%_{(0,t]}}(x,t)= \frac{{\rm E}_{\bf F}\Big{[}\bar{P}_{{\bf R, F}}%_{(0,t]}}(x,t)\Big{]}}{{\rm E}_{\bf F}\Big{[}\int dx\bar{P}_{{\bf R, F}}%_{(0,t]}}(x,t)\Big{]} }, \end{equation} where $\bar{P}_{{\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]}}(x,t)$ is an unnormalized pure classical state. That is, it is of the form $\bar{P}(x,t)=p_{{\bf R, F}}%_{(0,t]}}\delta[x-x_{{\bf R, F}}%_{(0,t]}}(t)]$, where $p_{{\bf R, F}}%_{(0,t]}}$ is the norm of the ostensible state. To show this we consider a system initially in the state $\bar{P}(x,0)=p\delta(x-x_0)$ then by using \erfs{pbar}{ptilde} with ${\cal J}_{r_{1},f_{1}}(x',t_1|x,0)$ defined implicitly in \erf{orf} we can rewrite $\bar{P}_{r_1,f_1}(x',t_1)$ as \begin{equation} \bar{P}_{r_1,f_1}(x',t_1)=p_{r_1,f_1}(t_1)\delta [x'-x_{f_1,r_1}(t_1)], \end{equation} which is still of the $\delta$-function form. Here $p_{r_1,f_1}(t_1)$ is given by \begin{equation}\label{pdifff} p_{r_1,f_1}(t_1) = P(f_1,t_1;r_1,0|x_0,0)p(0)/\Lambda({r_1,f_1}), \end{equation} and $x_{f_1,r_1}(t_1)$ is determined by the underlying dynamics. That is, we can simulate the distribution by solving the two coupled SDEs, $\dot{x}_{\bf R, F}(t)$ and $\dot{{p}}_{\bf R, F}(t)$. \section{A Quantum system with an unobserved process}\label{quantum} To illustrate a quantum system where a complete measurement can not be performed, due to some physical constraint, the system in Fig.~\ref{fig1} was considered. This system is a three level atom with lowering operators $\hat{L}_1=\ket{1}\bra{3}$ and $\hat{L}_2=\ket{2}\bra{3}$, and decay rates $\gamma_{1}$ and $\gamma_{2}$ respectively. \begin{figure}\begin{center} \includegraphics[width=0.25\textwidth]{PartlyObservedEvolutionFig01} \caption{\label{fig1} A simple system (a three level atom) which has two outputs due to the to lowering operators $\hat{L}_1$ and $\hat{L}_2$.}\end{center} \end{figure} \subsection{Master equation} With no external driving [$\hat{H}=0$ in \erf{QuantumMaster}], the solution of the master equation can be determined analytically. To illustrate a non-trivial solution we calculated this solution for the initial condition $\ket{\psi(0)}=0.4123\ket{1}+0.1\ket{2}+(0.9+0.1i)\ket{3}$ and coupling constants $\gamma_{1}=0.5$ and $\gamma_{2}=1$. This is shown in Fig.~\ref{fig2}. In this figure it is observed that as time goes on, the state becomes mixed. This is seen as the purity $p(t)={\rm Tr}[\rho^2(t)]$ of the state decays (although not monotonically) as time increases. This figure also shows that the state becomes a mixture of the two ground states, with the ground state associated with the larger coupling constant being weighted more heavily, even though it started with less weight. \begin{figure}[t]\begin{center} \includegraphics[width=0.45\textwidth]{PartlyObservedEvolutionFig02} \vspace{0.2cm} \caption{\label{fig2} The solution to the master equation. The first subplot shows $\rho_{33}(t)$ (solid line), $\rho_{22}(t)$ (dashed line) and $\rho_{11}(t)$ (dotted line). The second and third subplot show the real and imaginary parts respectively of $\rho_{12}(t)$ (solid line), $\rho_{31}(t)$ (dashed line) and $\rho_{32}(t)$ (dotted line). The fourth subplot illustrates the purity of this state. This is all for the initial condition $\ket{\psi(0)}=0.4123\ket{1}+0.1\ket{2}+(0.9+0.1i)\ket{3}$ and $\gamma_{1}=0.5$ and $\gamma_{2}=1$.}\end{center} \end{figure} \subsection{Conditional evolution: The quantum trajectory} \label{Incomplete} In this section we consider the trajectory $\rho_{{\bf R}}%_{(0,t]}}(t)$ which occurs when output $\hat{L}_1$ is monitored using homodyne-$x$ detection and output $\hat{L}_2$ is un-monitored. A schematic of this measurement process is shown in Fig.~\ref{fig3}. Because this arrangement is an inefficient measurement we have to use the operation defined in \erf{OperationDefComplete}. To determine the Kraus operators we need to present the underlying dynamics in more detail. For the interaction of this system with a Markovian bath (and under the rotating wave approximation and in the interaction frame) the total Hamiltonian is \begin{eqnarray}\label{eq7} H(t)&=&i\hbar\rt{\gamma_{1}}\int \delta(t-t')[\hat{L}_1\hat{b}_r^\dagger(t')-\hat{L}_1^\dagger\hat{b}_r(t')]dt' \nn \\ &&+i\hbar\rt{\gamma_{2}}\int \delta(t-t')[\hat{L}_2\hat{b}_f^\dagger(t')-\hat{L}_2^\dagger\hat{b}_f(t')]dt'.\nonumber\\ \end{eqnarray} Here $\hat{b}_r(t)$ and $\hat{b}_f(t)$ are the temporal-mode annihilation operators for the detected ($\hat{b}_r$) and non-detected ($\hat{b}_f$) fields (baths). Since these fields are Markovian there will be a commutator relationship for the field of the following form \begin{equation}\label{eq8} [{\hat{b}_{i}(t),\hat{b}_{j}^\dagger(s)}]=\delta(t-s)\delta_{i,j}, \end{equation} where $i$, $j$ denotes either of the two baths. This indicates that the field operators are gaussian white noise operators. Thus they obey It\^o~ calculus and the infinitesimal evolution operator is \cite{GarParZol92,GarZol00} \begin{eqnarray}\label{U} \hat{U}(t+dt,t)&=&\exp\Big{\{}\rt{\gamma_{1}}[\hat{L}_1d\hat{B}_r^\dagger(t)-\hat{L}_1^\dagger d\hat{B}_r(t)]\nn \\ && +\rt{\gamma_{2}} [\hat{L}_2d\hat{B}_f^\dagger(t)-\hat{L}_2^\dagger d\hat{B}_f(t)]\Big{\}}, \end{eqnarray} where $d\hat{B}_i$ satisfies the commutator relation \begin{equation} [d\hat{B}(t)_{i},d\hat{B}_{j}^\dagger(t)]=dt\delta_{i,j}. \end{equation} Thus $\hat{U}(t+dt,t)$ is an operator acting in the Hilbert space ${\cal H}_{s}\otimes {\cal H}_{r}\otimes {\cal H}_{f}$, where ${\cal H}_{s}$, ${\cal H}_{r}$ and ${\cal H}_{f}$ are the Hilbert spaces for the system, detected field and non detected field respectively. \begin{figure}\begin{center} \includegraphics[width=0.4\textwidth]{PartlyObservedEvolutionFig03} \vspace{0.2cm} \caption{\label{fig3} A schematic representing homodyne measurement of one of the outputs of the three level atom. In an ordinary homodyne measurement the signal is coupled to a classical local oscillator (LO) via a low reflective beam splitter and then detected using a photoreceiver.}\end{center} \end{figure} Now, given that a projective measurement is made on bath field $\hat{b}_r(t)$ and bath field $\hat{b}_f(t)$ is completely unobserved the state of the system after this measurement (time $dt$ later) is given by \erfs{QuantumUpdate}{OperationDefComplete} with $T=dt$, and the Kraus operator is \begin{equation}\label{Kraus2} \hat{K}_{r,f}(dt)=\bra{f}_f\bra{r}_r\hat{U}(t+dt,t)\ket{0}_{r}\ket{0}_f. \end{equation} Here $\{\ket{r}_r\}$ is the set of orthogonal states the bath is projected into, while $\{\ket{f}_f\}$ is any arbitrary orthogonal basis set. For a homodyne-$x$ measurement of bath $\hat{b}_r(t)$ the set $\{\ket{r}_r\}$ corresponds to the eigenset of the operator $d\hat{B}_{r}(t)+d\hat{B}_{r}^\dagger(t)$ \cite{GoeGra94} and the results $r$ are the corresponding eigenvalues. Note we have assumed that initially the baths, for all the temporal-modes, are in the vacuum state. After some simple rearrangement and using $(rdt)^2=dt$, the POM elements for this measurement are of the form \begin{equation} \hat{F}_{r}(dt)=|\bra{r}{0}\rangle_{r}|^2[1+\rt{\gamma_{1}}r(t+dt)dt \hat{x}_1], \end{equation} where $\hat{x}_1=\hat{L}_1+\hat{L}_1^\dagger$. Thus \begin{eqnarray}\label{eq20} { P}(r,t+dt)&=&|\bra{r}{0}\rangle|^2[1+rdt\rt{\gamma_{1}}\langle \hat{x}_1\rangle_t], \end{eqnarray} where $\langle \hat{x}_1\rangle_t={\rm Tr}[\hat{x}_1\rho(t)]$. Using the fact that $\ket{r}$ is a temporal-quadrature state, \begin{equation}\label{eq22} |\bra{r}{0}\rangle_{r}|^2=\rt{\frac{dt}{2\pi}}\exp\Big{(}-\frac{r^{2}}{2/dt}\Big{)}, \end{equation} we can rearrange this to \begin{equation}\label{eq23} { P}(r,t+dt)=\rt{\frac{dt}{2\pi}}\exp\Big{[}-\frac{[r-\rt{\gamma_{1}}\langle \hat{x}_1\rangle_t]^{2}}{2/dt}\Big{]}. \end{equation} This implies that the random variable associated with this distribution, $r(t+dt)dt$, is a gaussian random variable (GRV) of mean $\rt{\gamma_{1}}\langle \hat{x}_1\rangle_t dt$ and variance $dt$. That is, \begin{equation}\label{eq24} r(t+dt) dt=dW(t)+dt\rt{\gamma_{1}}\langle \hat{x}_1\rangle_t, \end{equation} where $dW(t)$ is a Wiener increment \cite{Gar85}. Using the above and \erfs{QuantumUpdate}{OperationDefComplete} the stochastic master equation for this system is \begin{eqnarray}\label{eq16} d\rho_{\bf R}(t+dt)&&=dt\Big{(}\gamma_{2}{\cal D}[\hat{L}_2] +\gamma_{1}{\cal D}[\hat{L}_1] \nn \\ &&+dW(t)\rt{\gamma_{1}}{\cal H}[\hat{L}_1]/dt\Big{)}\rho_{\bf R}(t),\nonumber\\ \end{eqnarray} where ${\cal H}[\hat{A}]$ is the superoperator \begin{equation}\label{eq17} {\cal H}[\hat{A}]\rho=\hat{A}\rho+\rho\hat{A}^\dagger -{\rm Tr}[\hat{A}\rho+\rho\hat{A}^\dagger]\rho. \end{equation} To illustrate an example quantum trajectory, \erf{eq16} was solved for a randomly chosen record ${\bf R}}%_{(0,t]}$ and the same parameters used in Fig. $\ref{fig2}$. This is shown in Fig. \ref{fig4}. It is observed that this state evolution is stochastic in time and becomes mixed (but not as mixed as the average evolution). It is interesting to note that by performing this measurement the coherence $\rho_{12, {\bf R}}(t)$, which was a constant of motion for the average state, becomes comparable to the other coherence and does not decay with time. \begin{figure}[t]\begin{center} \includegraphics[width=0.45\textwidth]{PartlyObservedEvolutionFig04} \vspace{0.2cm} \caption{\label{fig4} The solution to $\rho_{{\bf R}}$ written in matrix elements. The first subplot shows $\rho_{33,{\bf R}}(t)$ (solid line), $\rho_{22,{\bf R}}(t)$ (dashed line), $\rho_{11,{\bf R}}(t)$ (dotted line). The second and third subplot show the real and imaginary parts respectively of $\rho_{12,{\bf R}}(t)$ (solid line), $\rho_{31,{\bf R}}(t)$ (dashed line) and $\rho_{32,{\bf R}}(t)$ (dotted line). The fourth subplot illustrates the purity. We have used the same parameters as in Fig. \ref{fig2}}\end{center} \end{figure} \subsection{The ostensible numerical technique} In Sec.~\ref{OstensibleQ} we observed that the conditional evolution of a partly monitored system could be simulated by assuming that fictitious measurements are made on the unobservable process. For this system we assume that a fictitious homodyne-$x$ measurement is made on output $\hat{L}_2$. Note we could have chosen any unraveling for ${\bf F}}%_{(0,t]}$. To determine the SSE for the ostensible state $\ket{\bar{\psi}_{{\bf R, F}}(t)}$ [the state which we substitute into \erf{MainEquation} to determine the actual conditional evolution] we have to derive the measurement operator for the combined real and fictitious measurements, as well as make a convenient choice for $\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})$. Using \erf{eq33} and the fact that we are performing homodyne-$x$ measurements the measurement operator is \begin{eqnarray}\label{eq34} \hat{M}_{r,f}(dt)&=&\bra{f}0\rangle\bra{r}0\rangle\Big{(} 1+\rt{\gamma_{1}}rdt\hat{L}_1+\rt{\gamma_{2}}fdt\hat{L}_2\nn \\ &&\hspace{-1cm}-\gamma_{1}dt \hat{L}_1^\dagger\hat{L}_1/2-\gamma_{2}dt \hat{L}_2^\dagger\hat{L}_2/2\Big{)}, \end{eqnarray} where the bath states $\ket{f}$ and $\ket{r}$ are temporal quadrature states acting in Hilbert spaces ${\cal H}_f$ and ${\cal H}_r$ respectively. To derive this we have expanded \erf{U} to first order in $dt$ and used the fact that $(fdt)^2=(rdt)^2=dt$. Since the real distribution is Gaussian (with a variance $1/dt$) a convenient choice for $\Lambda({\bf F}}%_{(0,t]}, {\bf R}}%_{(0,t]})$ is $\Lambda({\bf F}}%_{(0,t]})\Lambda({\bf R}}%_{(0,t]})$ where $\Lambda({\bf F}}%_{(0,t]})=\Lambda(f_k)\ldots\Lambda(f_1)$ and $\Lambda({\bf R}}%_{(0,t]})=\Lambda(r_k)\ldots\Lambda(r_1)$ with \begin{eqnarray}\label{o1} \Lambda(r) &=& \rt{\frac{dt}{2\pi}}\exp\Big{[}-\frac{(r-\lambda)^{2}}{2/dt}\Big{]} \\ \Lambda(f) &=& \rt{\frac{dt}{2\pi}}\exp\Big{[}-\frac{(f-\mu)^2}{2/dt}\Big{]}\label{ostenFit}. \end{eqnarray} Here $\lambda$ and $\mu$ are arbitrary parameters. With these ostensible distributions, \erf{eq34}, and \erf{eq38}, the ostensible SSE is \begin{eqnarray}\label{eq50} d\ket{\bar{\psi}_{\bf R, F}(t)}&=&dt \Big{(}[r-\lambda](\rt{\gamma_{1}}\hat{L}_1-\lambda/2)+ [f-\mu]\nn \\ &&\times(\rt{\gamma_{1}}\hat{L}_2-\mu/2) -\smallfrac{1}{2} [\gamma_{1}\hat{L}_1^\dagger\hat{L}_1+\gamma_{2}\hat{L}_2^\dagger\hat{L}_2 \nn \\ &&-\rt{\gamma_{1}}\lambda\hat{L}_1-\rt{\gamma_{2}}\mu\hat{L}_2+\lambda^2/4+\mu^2/4] \Big{)} \nn \\ &&\times \ket{\bar{\psi}_{\bf R, F}(t)}. \end{eqnarray} Now since we are interested in calculating $\rho_{{\bf R}}(t)$ based on an assumed known real record ${\bf R}}%_{(0,t]}$, we can rewrite \erf{eq50} as \begin{eqnarray}\label{eq52} dc_{1}&=& c_{3}[\rt{\gamma_{1}}(r-\lambda)dt+dt\lambda/2]-c_1[\rt{\gamma_2} d{\cal W}\mu\nn \\ &&+\rt{\gamma_1}(r-\lambda)dt\lambda+dt\lambda^2/4+dt\mu^2/4]/2,\\ dc_{2}&=&c_{3}[\rt{\gamma_{2}}d{\cal W}+dt\mu/2]-c_2[\rt{\gamma_2} d{\cal W}\mu\nn \\ &&+\rt{\gamma_1}(r-\lambda)dt\lambda+dt\lambda^2/4+dt\mu^2/4]/2\\ dc_{3}&=&c_3[-\gamma dt+\rt{\gamma_2} d{\cal W}\mu+\rt{\gamma_1}(r-\lambda)dt\lambda\nn \\ &&+dt\lambda^2/4+dt\mu^2/4]/2, \end{eqnarray} where $\gamma=\gamma_1+\gamma_2$. Here we have used the identity \begin{equation}\label{eq47} \ket{\bar{\psi}(t)}=c_{1}\ket{1}+c_{2}\ket{2}+c_{3}\ket{3}, \end{equation} and replaced $fdt$ with $d{\cal W}(t)+\mu dt$, where $d{\cal W}(t)$ is a Wiener increment. To illustrate the convergence of our method the ensemble average of the above ostensible SSE for $\lambda=\mu=0$ was calculated for $n=10$ and $n=1000$. To quantify how closely the ensemble method reproduces $\rho_{{\bf R}}%_{(0,t]}}(t)$ we used the fidelity measure, which for two different quantum states is defined as \begin{equation}\label{felquant} F^{\rm (Q)}(t)={\rm Tr}[\rt{\rt{\rho_1(t)}\rho_2(t)\rt{\rho_1(t)}}]. \end{equation} Note this measure ranges from 0 to 1 with 0 indicating two orthogonal states and 1 indicating the same state. The result of this measure for the actual $\rho_{{\bf R}}%_{(0,t]}}(t)$ and the ensemble version are shown in part A of figure \ref{felquantf}. Here we see that for larger ensemble size the fidelity is closer to one, indicating that as we increase the ensemble size our ostensible method approaches the actual $\rho_{{\bf R}}%_{(0,t]}}(t)$. To illustrate the effect of choosing different ostensible distributions we considered the case when $\lambda=0$ and \begin{equation}\label{eq422} \mu=\rt{\gamma_2}\frac{\bra{\bar{\psi}_{\bf R, F}(t)}\hat{L}_2+\hat{L}_2^\dagger \ket{\bar{\psi}_{\bf R, F}(t)}}{\bra{\bar{\psi}_{\bf R, F}(t)}\bar{\psi}_{\bf R, F}(t)\rangle}. \end{equation} That is, the ostensible probability for the $k^{th}$ fictitious results is the true probability we would expect based on the past real and fictitious results up to, but not including the time $kdt$. The motivation for this choice is that with $\mu=0$, the improbable trajectories, ones that tend towards being inconsistent with the full real record, will have norms that are very small and as such have little contribution to the ensemble average. By contrast, using \erf{eq422}, the improbable trajectories are less likely to be generated, so avoiding useless simulations. With this ostensible distribution the fidelity measure was calculated for $n=10$ and $n=1000$. These results are shown in part $B$ of figure \ref{felquantf}. Here we see that for the smaller ensemble size the fidelity is closer to one than that observed using the first ostensible case. This indicates that the rate of convergence for this case is greater than the $\lambda=\mu=0$ case. \begin{figure}[t]\begin{center} \includegraphics[width=0.45\textwidth]{PartlyObservedEvolutionFig05} \vspace{0.2cm} \caption{\label{felquantf} This figure shows the fidelity between the actual $\rho_{{\bf R}}%_{(0,t]}}(t)$ and our ensemble method for ensembles sizes 10 (dotted) and 1000 (solid). Part $A$ corresponds to a linear ostensible distribution while part $B$ refers to the non-linear ostensible distribution. The same parameters were used as in Fig. \ref{fig2}.}\end{center} \end{figure} \section{A Classical system with an internal unobserved process}\label{class} In this section we consider continuous-in-time measurements with Gaussian precision of a classical system driven by an unobservable noise process. This for example could correspond to a measurement of the voltage across a resistor that is driven by a noisy classical current. \subsection{The average evolution} We restrict ourselves to unconditional state evolution described by the Fokker Plank equation \erf{Eq.FPE2}. This equation has as its solution a distribution that diffuses and drifts though time. Using Eq. (\ref{Ave}) and only considering one interval in time we can write \begin{equation} P(x',t+dt)=\int dx {\cal O}(x',t+dt|x,t) P(x,t), \end{equation} which when compared to \erf{Eq.FPE2} implies that RHS of the above equation equals \begin{eqnarray}\label{Eq.FPE3} &&\hspace{-.8cm} \int dx [1 -dt\partial_{x'}A(x,t) +dt\partial^2_{{x'}}D^2(x,t)/2]\delta(x'-x) \nn \\ &&\times P(x,t). \end{eqnarray} By introducing an arbitrary Gaussian distribution $P(f,t+dt)$ with mean $m(t)$ and variance $1/dt$, that is \begin{equation}\label{RealFclass} P(f,t+dt)= \rt{\frac{dt}{2\pi}}\exp\Big{[}-\frac{[f-m(t)]^{2}}{2/dt}\Big{]}, \end{equation} \erf{Eq.FPE3} can be rewritten as \begin{eqnarray}\label{Eq.FPE4} &&\hspace{-.8cm} \int df P(f,t+dt) \int dx [1 -dt\partial_{x'}A(x,t) -dt [f-m(t)]\nn \\ &&\times \partial_{x'}D_f(x,t)+dt\partial^2_{{x'}}D^2(x,t)/2]\delta(x'-x) P(x,t).\nonumber\\ \end{eqnarray} By using It\^o~ calculus and a Taylor expansion this can be rewritten as \begin{eqnarray}\label{Eq.FPE6} &&\hspace{-.8cm} \int dx {\rm E}_f \Big{\{}\delta[x'-x-dt A(x,t)- dt[f(t+dt)-m(t)] \nn \\ &&\times D(x,t)]\Big{\}}P(x,t). \end{eqnarray} where $f(t+dt)dt=m(t)dt +d {\cal W}(t)$. Thus \begin{eqnarray}\label{OO} {\cal O}(x',t+dt|x,t)={\rm E}_f \Big{\{}\delta[x'-x_f(t+dt)]\Big{\}}, \end{eqnarray} where $x_f(t+dt)$ is determined by the following SDE \begin{equation} \label{Eq.Linear1} d x_{\bf F}(t)=dt A[x_{\bf F}(t),t]+ dt[f(t+dt)-m(t)]D[x_{\bf F}(t),t]. \end{equation} Note here we have written the SDE for the complete record ${\bf F}$. \subsection{Conditional evolution: The Kushner-Stratonovich equation}\label{KSEsec} To derive the KSE we start by deriving ${\cal O}_r(x',t'|x,t)$ and $P(r,t+dt)$. For the case when the classical measurement has a back action that is independent of the result $r(t+dt)$, the operation for the measurement is given by \begin{equation} {\cal O}_r(x',t'|x,t)={\cal O}(x',t'|x,t)P(r,t|x,t), \end{equation} where $ {\cal O}(x',t'|x,t)=B(x',t'|x,t)$. Thus to derive ${\cal O}_r(x',t'|x,t)$ we need only $P(r,t|x,t)$. For a measurement that has a precision limited by Gaussian white noise it follows that \begin{equation}\label{Eq.PIgivenx} P(r,t|x,t)=F_r(x,t)=\frac{\sqrt{dt}}{\sqrt{2\pi \beta}}\exp[-(r-x)^2 dt/2\beta], \end{equation} where $\beta$ is a constant characterizing the classical measurement strength. To find $P(r,t)$ we substitute \erf{Eq.PIgivenx} into \erf{Pr3}. This gives \begin{equation}\label{Eq.PI} P(r,t+dt)=\int dx\frac{\sqrt{dt}}{\sqrt{2\pi \beta }}\exp[-(r-x)^2 dt/2\beta]P(x,t). \end{equation} After some simple stochastic algebra and using $r^2=\beta/dt$ this can be simplified to \cite{WarWis03a} \begin{equation}\label{Eq.PI5} P(r,t+dt) =\frac{\sqrt{dt}}{\sqrt{2\pi \beta}} \exp[-(r-\langle x\rangle_t)^2dt/2\beta ], \end{equation} where for the classical system $\langle x\rangle_t=\int x P(x,t)dx$. From \erf{Eq.PI5} the stochastic representation of $r(t+dt)$ is a Gaussian random variable with mean $\langle x\rangle_t$ and variance $\beta dt$. That is, \begin{equation}\label{Eq.I} r(t+dt)=\langle x\rangle_t+\sqrt{\beta}dW(t)/dt. \end{equation} With all the above information and \erf{GeneralBayessTheorem} the conditional state at time $t'=t+dt$ is \begin{eqnarray} && \hspace{-0.8cm} P_r(x',t+dt)=\int dx {\rm E}_f \Big{\{}\delta[x'-x_f(t+dt)]\Big{\}} \Big{\{}1\nn \\ &&+[x-\langle x\rangle_t] [r-\langle x\rangle_t]dt/\beta\Big{\}}P(x,t). \end{eqnarray} Here we have expanded the exponentials in \erf{Eq.PI5} and \erf{Eq.PIgivenx} to second order in $dt$ and used $r^2=\beta/dt$. Taylor expanding the delta function and averaging over the $f(t+dt)$ [using \erf{RealFclass}] for each step in time gives the KSE \begin{eqnarray} \label{KS} P_{\bf R}(x,t+dt)&=&P_{\bf R}(x,t)+dt[x-\langle x\rangle_t][r(t+dt)-\langle x\rangle_t]\nn \\ &&\times P_{\bf R}(x,t)/\beta -dt\partial_x [{A}({x},t)P_{\bf R}(x,t)]\nn \\ &&+\smallfrac{1}{2} dt\partial^2_x [D^2({x},t)P_{\bf R}(x,t)] \end{eqnarray} and $\langle x\rangle_t$ becomes $\int x P_{\bf R}(x,t)dx$. In general to solve this equation we need to solve for all $x$. For some ${A}({x},t)$ and $D(x,t)$ this can be a rather lengthy numerical problem. In the following section we will present our ostensible technique which allows us to reformulate the problem to solving two coupled SDEs, at the cost of performing an ensemble average. \subsection{The ostensible numerical technique}\label{sec.linearclass} As shown in Sec. \ref{sec.fitcla}, if we consider the unobservable process ${\bf F}$ as actually occurring then we can simulate the KSE by using \erf{EPxgrvialin}, and $\bar{P}_{{\bf R}}%_{(0,t]},{\bf F}}%_{(0,t]}}(x,t)$ is determined by solving two coupled SDEs. For the case when the classical measurement has Gaussian precision and the back action only depends on the white noise process $f(t)$, we can rewrite $P(f,t';r,t|x,t)$ in \erf{pdifff} as $P(f,t')P(r,t|x,t)$ where $P(f,t')$ is given by \erf{RealFclass} and $P(r,t|x,t)$ is given by \erf{Eq.PIgivenx}. Thus $\dot{x}_{\bf R,F}(t)$ becomes $\dot{x}_{\bf F}(t)$ and is given by \erf{Eq.Linear1}. To find the differential equation for $\dot{p}_{\bf R, F}(t)$ we need to assume a form for the ostensible distribution $\Lambda(f,r)$. We use $\Lambda(f,r)=\Lambda(f)\Lambda(r)$, where \begin{eqnarray}\label{ostenRit} \Lambda(r) &=& \rt{\frac{dt}{2\pi}}\exp\Big{[}-\frac{(r-\lambda)^{2}}{2\beta/dt}\Big{]} \end{eqnarray} and $\Lambda(f)$ is given by \erf{ostenFit}. Extending \erf{pdifff} to continuous measurements gives \begin{eqnarray}\label{Eq.Linear2} d p_{\bf R, F}(t)&=& dt[m(t)-\mu][f(t+dt)-\mu]p_{\bf R, F}(t)\nn \\ &&+dt[x(t)-\lambda][r(t+dt)-\lambda]p_{\bf R, F}(t)/\beta.\nonumber\\ \end{eqnarray} Thus to determine $\bar{P}_{\bf R, F}(x,t)$ we only need to simulate \erfs{Eq.Linear1}{Eq.Linear2} with ${\bf R}$ assumed known and $f(t+dt)$ given by \erf{ostenFit}. $P_{\bf R}(x,t)$ is then determined by \erf{Pxgrvialin}. Since the theory requires $\bar{P}_{\bf R, F}(x,t)$ to be a delta function, one might conclude that this method is only valid for initial conditions of the form ${P}(x,0)\rho(0)=\delta(x-x_0)$. Infact, we are not limited to this case. To consider other initial conditions we simply choose the initial value $x_0$ in \erf{Eq.Linear1} from the distribution ${P}(x,0)$. \subsection{A simple example} To illustrate the classical theory we consider a Gaussian measurement of a classical system that is driven by an an unobservable white noise process with $m(t)=0$ and drift and diffusion functions given by \begin{eqnarray} \label{cond1} A({x},t) &=& -kx+l, \\ D({x},t) &=& b.\label{cond2} \end{eqnarray} If this is the case then $P_{{\bf R}}%_{(0,t]}}(x,t)$ has a Gaussian solution with a mean $\langle x_{\bf R}\rangle_t$ and variance $\nu_{\bf R}(t)$ given by \begin{eqnarray} d\langle x_{\bf R}\rangle_{t}&=&dt\{\nu_{\bf R}(t)[r(t+dt)-\langle x_{\bf R}\rangle_t ]/\beta- k \langle x_{\bf R}\rangle_t\nn \\ &&+l\},\label{Eq.bKS22}\\ \label{Eq.vKS22} d\nu_{\bf R}(t)&=&dt[-\nu_{\bf R}^2(t)/\beta-2k\nu_{\bf R}(t)+b^2], \end{eqnarray} and $r(t+dt)=\langle x_{\bf R}\rangle_t+dW(t)$. That is, as time increases the measurement has the effect of reducing the variance but the diffusive coefficient $b$ causes this variance to increase. The mean, however contains both the deterministic evolution and a random term due the measurement. To illustrate this solution we have simulated \erfs{Eq.bKS22}{Eq.vKS22} for the case when $ A(x,t) = 1-x$, $D=1$ and $\beta=1$. The results of this simulation are shown in Fig. \ref{Fig.PxGivenIEquations} as a solid line. Here we see that the mean follows some stochastic path conditioned on the record ${\bf R}}%_{(0,t]}$, while the variance follows a smooth function. \begin{figure}\begin{center} \includegraphics[width=0.45\textwidth]{PartlyObservedEvolutionFig06} \caption{\label{Fig.PxGivenIEquations} The mean and variance of ${P}_{{\bf R}}%_{(0,t]}}(x,t)$ when $\beta= 1$, $A=1-x$ and $D=1$ for both ${P}_{{\bf R}}%_{(0,t]}}(x,t)$ calculated exactly (solid) and via the linear method for an ensemble size of 10 000 (dotted). }\end{center} \end{figure} To illustrate our ostensible method we use the above record and solve numerically \erfs{Eq.Linear1}{Eq.Linear2} with $\lambda=\mu=0$. The mean and variance is then found via \begin{eqnarray} \langle x_{\bf R}\rangle_{t}&=&\frac{{\rm E}_{\bf F}\Big{[}{x}_{\bf F}(t)p_{\bf R, F}(t)\Big{]}}{ {\rm E}_{\bf F}\Big{[}p_{\bf R, F}(t)\Big{]}} \\ \nu_{\bf R}(t)&=&\frac{{\rm E}_{\bf F}\Big{[}x_{\bf F}^2(t)p_{\bf R, F}(t)\Big{]}}{ {\rm E}_{\bf F}\Big{[}p_{\bf R, F}(t)\Big{]}}-\langle x_{\bf R}\rangle^2_{t}\nonumber\\ \end{eqnarray} where ${\rm E}_{\bf F}$ denotes an ensemble average over all possible fictitious records. The numerical values for the mean and variance are shown in Fig. \ref{Fig.PxGivenIEquations} (dotted) for an ensemble size of 10 000. To get an indication of the numerical error in the solution from our method, the difference from the exact solution is shown in Fig. \ref{Fig.EPxGivenIF1000}. The dotted line corresponds to an ensemble of 100 and the solid to one of 10 000. Here we see that the ostensible method solution agrees well with the exact solution and as we increase the ensemble size the difference between these solutions decreases. To get a better indication of how well our method reproduces the actual ${P}_{{\bf R}}%_{(0,t]}}(x,t)$, we also calculated the classical fidelity, which is defined by \begin{equation} \label{felclass} F^{(C)}(t)=\int dx \rt{P_1(x,t)}\rt{P_2(x,t)}. \end{equation} This was calculated under the assumption that the state calculated via the ostensible method was also Gaussian. This is illustrated in Fig. \ref{Fig.EPxGivenIF1000}, where we see that for the larger ensemble the fidelity is very close to one, implying that the distributions are almost identical. \begin{figure}\begin{center} \includegraphics[width=0.45\textwidth]{PartlyObservedEvolutionFig07} \caption{\label{Fig.EPxGivenIF1000} The first and second plot show the difference between the mean and variance of ${P}_{{\bf R}}%_{(0,t]}}(x,t)$ calculated by the linear method and the know result for ensemble sizes 100 (dotted) and 10 000 (solid). The third plot shows the Fidelity between ${P}_{{\bf R}}%_{(0,t]}}(x,t)$ calculated by the linear method and the know result for ensemble sizes 100 (dotted) and 10 000 (solid). The parameters are the same as in Fig. \ref{Fig.PxGivenIEquations}.}\end{center} \end{figure} \section{An unobservable quantum system driving a Classical system}\label{both} In this section we consider the following situation: a quantum system is monitored continuously in time by a classical system. This in turn is measured with Gaussian precision, and these are the only results to which we have access. This for example occurs when the signal from the quantum system enters a detector with a bandwidth $B$, resulting in the state of the detector being related to ${\bf F}}%_{(0,t]}$ by \cite{WarWis03a} \begin{equation}\label{pollysyst} x(t)=\int_{-\infty}^{t} ds B\exp[-B(t-s)]{f}(s). \end{equation} Thus in a measurement that reveals $x(t)$ with perfect precession [eg $F_r(x)=P(r,t|x,t)=\delta(r-x)$] we could determine ${\bf F}}%_{(0,t]}$ (the quantum signal) by inverting the convolution in \erf{pollysyst}. But if this measurement has Gaussian precision [\erf{Eq.PIgivenx}] then we must treat the state of the detector as a classical probability distribution and use a mixture of CMT and QMT to describe the conditional state of the supersystem (classical and quantum system). To denote the supersystem we use the notation $\rho(x,t)$, where $x$ refers to the classical configuration space and $\rho$ denotes an object acting on a Hilbert space. This has the interpretation whereby $P(x,t)={\rm Tr}[\rho(x,t)]$ is the (marginal) classical state and $\rho(t)=\int \rho(x,t)dx$ is the (reduced) quantum state. For uncorrelated quantum and classical states, $\rho(x,t)=P(x,t)\rho(t)$. \subsection{Conditional evolution} \label{ConSec} We denote the state of the supersystem conditioned on the classical result $r$ at time $t+dt$ as $\rho_{r}(x,t+dt)$. Assuming that the quantum system is not affected by the classical system, this can be expanded as \begin{equation}\label{s} \rho_{r}(x,t+dt)=\sum_{f}P_r(f,t+dt)P_{r,f}(x,t+dt)\rho_f(t+dt), \end{equation} where $\rho_f(t+dt)$ is the state that an observer who had access to all the quantum information would ascribe to the quantum system. That is, $f(t+dt)$ can be regarded as really existing (with the collapse of the wavefunction occurring at this level); it is just that the real observer does not have access to this information. The state of knowledge of this real observer is different from, but consistent with, that of the hypothetical observer who has access to {\bf F}. In terms of the operation of the measurement, the conditional state can be written as \begin{eqnarray}\label{ss} \rho_{r}(x',t+dt)&=&\frac{\tilde{\rho}_{r}(x',t+dt)}{P(r,t+dt)}, \end{eqnarray} where \begin{eqnarray}\label{sss} \tilde{\rho}_{r}(x',t+dt)&=&\int dx \sum_{f} {\cal J}_{r,f}(x',t+dt|x,t) \nn \\ &&\times\hat{\cal O}_f(t+dt,t) \rho(x,t)/P(f,t+dt) \hspace{.8cm} \end{eqnarray} and \begin{eqnarray}\label{ssss} P(r,t+dt)&=&\int dx' {\rm Tr}\Big{[}\tilde{\rho}_{r}(x',t+dt) \Big{]}. \end{eqnarray} The quantum part of the operation of measurement in defined by \erf{OperationDefComplete} and the classical part is defined in \erf{orf} with the replacement of $P(f,t';r,t|x,t)\rightarrow P(f,t')P(r,t|x,t)$ because in this system the quantum signal does not depend on the classical state. To illustrate the above we consider the case when we are monitoring with Gaussian precision the classical system defined by \erf{pollysyst} which is in turn monitoring the $x$ quadrature flux coming from a classically driven two level atom (TLA). This is the same as the system considered in Ref \cite{WarWis03a} and as such we will simply list the important equations. The quantum part of operation is given by $\hat{\cal O}_f(t+dt,t)=\hat{\cal J}[\hat{M}_f(dt)]$ where \begin{equation}\label{m} \hat{M}_f(dt)=\bra{f}0\rangle[1-dt(i\hat{H}-\rt{\gamma}f\hat{\sigma}+\gamma \hat{\sigma}^\dagger\hat{\sigma}/2)]. \end{equation} The fictitious quantum signal statistic obeys \begin{equation} P(f,t+dt)=\int dx {\rm Tr}[\hat{\cal O}_f(t+dt,t)\rho(x,t)], \end{equation} which for a homodyne-$x$ measurement can be shown to be of the form displayed in \erf{RealFclass} with $m(t)=\rt{\gamma}{\rm Tr} [(\hat \sigma+\hat\sigma^\dagger)\rho(t)]$. Here $\hat{\sigma}$ is the lowering operator for the TLA and $\gamma$ is the decay rate. Note here we have assumed all the quantum signal is fed into the classical system, if we wanted to simulate some inefficiency we would simply use the Kraus represention, and for the case where this inefficiency is a constant, $\eta$, we simply replace $\sigma$ in the above equations by $\rt{\eta}\sigma$. As shown in Sec. \ref{class} for a classical measurement with Gaussian precision and a back action that does not depend on the results of the measurement, the classical part of the operation is \begin{equation}\label{p} {\cal J}_{r,f}(x',t+dt|x,t)=\delta[x'-x_f(t+dt)]P(f,t+dt)P(r,t|x,t), \end{equation} where $P(r,t|x,t)$ is defined in \erf{Eq.PIgivenx} and $x_f(t+dt)$ is given by \erf{Eq.Linear1}. For the system we are considering, to find $A(x,t)$ and $D(x,t)$ we simply differentiate \erf{pollysyst} and equate this with \erf{Eq.Linear1}. Doing this gives \begin{eqnarray}\label{Ad} A(x,t)&=&-B x +B m(t),\\ D(x,t)&=&B. \label{Dd} \end{eqnarray} Combining the quantum and classical parts of the operation and using the same techniques as in Sec. \ref{KSEsec} allows us to rewrite \erf{ss} for continuous-in-time measurements as \begin{eqnarray}\label{supertrajextory} d\rho_{\bf R}(x,t)&=&dt \Big{(} B\partial_{x}x +\smallfrac{1}{2}B^2\partial^2_{x} +\hat{\cal L}\Big{)}\rho_{\bf R}(x,t) \nn \\ &&+dt \Big{(}\frac{[x-\langle x_{\bf R}\rangle_t][r(t+dt)-\langle x_{\bf R}\rangle_t]}{\beta}\Big{)}\rho_{\bf R}(x,t) \nn \\ &&-dt\rt{\gamma}\partial_x B [\hat\sigma\rho_{\bf R}(x,t)+\rho_{\bf R}(x,t)\hat\sigma^\dagger], \end{eqnarray} where $\langle x_{\bf R}\rangle_t=\int x{\rm Tr}[ \rho_{\bf R}(x,t)]dx $ and \begin{equation}\label{rboth} r(t+dt)dt= \langle x_{\bf R}\rangle_tdt+\rt{\beta}dW(t). \end{equation} This equation (\ref{supertrajextory}) has been labeled the Superoperator-Kushner-Stratonovich equation \cite{WarWis03a} and represents the evolution of the combined supersystem. The first line contains the free evolution for both the quantum and the classical systems. For this quantum system \begin{equation} \hat{\cal L}[\hat\sigma]\rho=\frac{-i \Omega}{2}[\hat{\sigma}_x,\rho] +\gamma\hat{\cal D}[\hat\sigma]\rho, \end{equation} where $\Omega$ is the Rabi frequency and $\hat{\cal D}$ is the damping superoperator and is defined in \erf{DampSuper}. The second line of Eq.~(\ref{supertrajextory}) describes the gaining of knowledge about the state of classical system via Gaussian measurements. Lastly the third line describes the coupling of the quantum and classical system. For a TLA we can write the state of the supersystem as \begin{equation}\label{superrho} \rho(x,t)=\smallfrac{1}{2}[P(x,t)\hat{1}+X(x,t)\hat{\sigma}_x+Y(x,t)\hat{\sigma}_y+Z(x,t)\hat{\sigma}_z]. \end{equation} Note that $P(x,t)$ is the marginal state of knowledge for the classical system (found via tracing out the quantum degrees of freedom). Substituting this into \erf{supertrajextory} gives the following four coupled partial differential equations \begin{eqnarray}\label{classical} \dot{P}_{\bf R}&=&[x-\langle x_{\bf R}\rangle_t][r-\langle x_{\bf R}\rangle_t]P_{\bf R}/\beta+B\partial_x[x P_{\bf R}\nn \\ &&-\rt{\gamma}X_{\bf R}] +\smallfrac{1}{2}B^2\partial_x^2P_{\bf R}\\ \dot{X}_{\bf R}&=&[x-\langle x_{\bf R}\rangle_t][r-\langle x_{\bf R}\rangle_t]X_{\bf R}/\beta+\smallfrac{1}{2}B^2\partial_x^2X_{\bf R}\nn \\ &&+B\partial_x[x X_{\bf R}-\rt{\gamma}P_{\bf R}-\rt{\gamma}Z_{\bf R}] -\smallfrac{1}{2}\gamma X_{\bf R}\\ \dot{Y}_{\bf R}&=&[x-\langle x_{\bf R}\rangle_t][r-\langle x_{\bf R}\rangle_t]Y_{\bf R}/\beta+B\partial_x[x Y_{\bf R}]\nn \\ &&+\smallfrac{1}{2} B^2\partial_x^2Y_{\bf R}-\Omega Z_{\bf R}-\smallfrac{1}{2}\gamma Y_{\bf R}\\ \dot{Z}_{\bf R}&=&[x-\langle x_{\bf R}\rangle_t][r-\langle x_{\bf R}\rangle_t]Z_{\bf R}/\beta+B\partial_x[x Z_{\bf R}\nn \\ &&+\rt{\gamma}X_{\bf R} ]+\smallfrac{1}{2}B^2\partial_x^2Z_{\bf R}+\Omega Y_{\bf R}-\gamma(P_{\bf R}+Z_{\bf R}).\nonumber \\ \label{classical4} \end{eqnarray} To determine the state of knowledge for the quantum system we simply integrate out the classical degrees of freedom. To illustrate a trajectory for this supersystem the following parameters were used; $\beta=0.5$, $B=2$, $\gamma=1$ and $\Omega=5$. The results are shown in Fig. \ref{figsuperI} (solid line) for a randomly chosen record ${\bf R}}%_{(0,t]}$. This figure displays the mean and the variance of the classical trajectory found via tracing over the quantum degrees of freedom as well as the quantum state in Bloch representation after we have integrated out the classical degrees of freedom. \begin{figure}\begin{center} \includegraphics[width=0.45\textwidth]{PartlyObservedEvolutionFig08} \caption{\label{figsuperI} ${\rho}_{{\bf R}}%_{(0,t]}}(x,t)$ calculated via numerical integration (solid) and via the ostensible method for an ensemble size of 10 000 (dotted). The parameters are $\beta=0.5$, $B=2$, $\gamma=1$ and $\Omega=5$ and initial conditions ${\rho}(x,0)=P(x)\ket{g}\bra{g}$ where $P(x)$ is a Gaussian with mean zero and variance 0.1. }\end{center} \end{figure} \subsection{Fictitious trajectories: The ostensible numerical technique} In the above section we observed that to be able to calculate the supersystem trajectory we needed to solve four coupled partial differential functions (each involving derivatives with respect to a classical configuration coordinate $x$). This is a rather lengthy calculation which for higher dimensional ($d$) quantum systems will require $d^2$ partial differential equations. Here we present our linear method that allows us to reduce the problem to $d+2$ couple differential equations. The expense, again, is that an ensemble average must be performed. To do this we simply note that we can define the following quantum and classical states \begin{equation}\label{linearquantum} \bar\rho_f(t+dt)=\frac{\hat{\cal O}_f(t+dt,t)\rho(t)}{\bar{\Lambda}(f)} \end{equation} and \begin{equation}\label{linearclassical} \bar P_{r,f}(x',t+dt)=\frac{\int dx \bar{\cal O}_{r,f}(x',t+dt|x,t) P(x,t)}{\Lambda(f)\Lambda(r)}, \end{equation} where \begin{equation} \bar{\cal O}_{r,f}(x',t+dt|x,t)=\delta[x'-x_{f}(t+dt)]P(r,t|x,t)\bar{\Lambda}(f). \end{equation} Note the bar above $\bar{\Lambda}(f)$ means that the ostensible distribution used to scale the quantum state does not have to be the same as that used to scale the classical state. Here for simplicity we consider only the case when they are the same (as no numerical advantage is gain by different choices). Using the above equations we can rewrite \erfs{ss}{ssss} as \begin{eqnarray}\label{li} \rho_{r}(x',t+dt)&=&\frac{\sum_{f}\Lambda(f)\Lambda(r)\bar{P}_{r,f}(x',t+dt)\bar\rho_f(t+dt)}{P(r,t+dt)},\nonumber\\\\ P(r,t+dt)&=&\int dx \sum_f{\rm Tr} [\Lambda(f)\Lambda(r)\bar{P}_{r,f}(x,t+dt)\nn \\ &&\times\bar\rho_f(t+dt)]. \end{eqnarray} Thus to simulate $\rho_{\bf R}(x,t)$ we need only to calculate $\bar{P}_{\bf R, F}(x,t)$ and $\bar\rho_{\bf F}(t)$ for a specific record ${\bf R}}%_{(0,t]}$. For the above TLA-classical detector system with $\Lambda(r)$ and $\Lambda(f)$ defined by \erfs{ostenRit}{ostenFit} respectively, $\bar P_{\bf R, F}(x',t)$ has a solution of the form $p_{\bf R}(t)\delta[x'-x_{\bf R, F}(t)]$ where $x_{\bf F}(t)$ is given by \begin{equation} \label{Eq.Linear1a} d{x}_{\bf F}(t)=dt [-B x_{\bf F}(t) +B f(t+dt)] \end{equation} and $p_{\bf R, F}(t)$ is given by \begin{equation}\label{Eq.Linear3} dp_{\bf R, F}(t)=dt[x_{\bf F}(t)-\lambda][r(t+dt)-\lambda]p_{\bf R, F}(t)/\beta,\hspace{1cm}. \end{equation} Thus we can rewrite \erf{li} as \begin{equation}\label{liMain} \rho_{\bf R}(x,t)=\frac{{\rm E}_{\bf F}\Big{[}\delta[x-x_{\bf F}(t)]p_{\bf R,F}(t)\bar\rho_{\bf F}(t)\Big{]}} {{\rm E}_{\bf F}\Big{[}p_{\bf R,F}(t)\check{p}_{\bf F}(t) \Big{]}}, \end{equation} where $\check{p}_{\bf F}(t) ={\rm Tr}[\bar\rho_{\bf F}(t)]$. To determine the evolution of the ostensible quantum state we simply substitute the measurement operator defined in \erf{m} with $\hat{H}=\Omega\hat{\sigma}_x/2$ and the ostensible distribution ${\Lambda}(f)$ into \erf{linearquantum}. Doing this gives \begin{eqnarray}\label{linearquantumtraj} d\bar\rho_{\bf F}(t)&=&dt\frac{-i\Omega}{2}[\hat{\sigma}_x,\rho_{\bf F}(t)]+dt\gamma\hat{\cal D}[\hat{\sigma}]\rho_{\bf F}(t)+\nn \\ && dt [f(t+dt)-{\mu}][\rt{\gamma}\hat{\sigma}\rho_{\bf F}(t) \nn \\ &&+\rt{\gamma}\rho_{\bf F}(t)\hat{\sigma}^\dagger-{\mu}\rho_{\bf F}(t)]. \end{eqnarray} However since we have assumed that all the quantum signal is fed into the detector the evolution of the ostensible quantum state can be written as an ostensible SSE. That is, \begin{eqnarray} d\ket{\bar{\psi}_{\bf F}(t)}&=&dt \Big{(}-\frac{i\Omega}{2}\hat{\sigma}_x+[f(t+dt)-\mu](\rt{\gamma}\hat{\sigma}-\mu/2)\nn \\ && -\smallfrac{1}{2} [\gamma\hat{\sigma}^\dagger\hat{\sigma} -\rt{\gamma}\mu\hat{\sigma}+\mu^2/4] \Big{)} \ket{\bar{\psi}_{\bf F}(t)}.\hspace{.8cm} \end{eqnarray} Thus to determine $\rho_{\bf R}(x,t)$ all we need to do is solve the above SSE and \erfs{Eq.Linear1a}{Eq.Linear3} for ${\bf R}$ assumed known and $f(t+dt)dt=d{\cal W} +dt \mu$ where $d{\cal W}$ is a Wiener increment. Once solved the quantum state conditioned on ${\bf R}}%_{(0,t]}$ is given by \begin{eqnarray} {\chi}_{\bf R}(t)&=&\frac{{\rm E}_{\bf F}\Big{[}p_{\bf R, F}(t)\check{\chi}_{\bf F}(t)\Big{]}}{ {\rm E}_{\bf F}\Big{[}p_{\bf R, F}(t)\check{p}_{\bf F}(t)\Big{]}}, \end{eqnarray} where ${\chi}_i =\{\check{x}_i,\check{y}_i,\check{z}_i\}$ are the Bloch vectors of the quantum state. The moments of the classical state are given by \begin{eqnarray} \langle x^m_{\bf R} \rangle_{t}&=&\frac{{\rm E}_{\bf F}\Big{[}{x}^m_{\bf F}(t)p_{\bf R, F}(t)\check{p}_{\bf F}(t)\Big{]}}{ {\rm E}_{\bf F}\Big{[}p_{\bf R, F}(t)\check{p}_{\bf F}(t)\Big{]}} . \hspace{.5cm} \end{eqnarray} To illustrate this method we considered two choices for the ostensible distributions. The first is $\lambda=\mu=0$; that is, all the ostensible distributions are Gaussian distributions of mean zero and variance $dt$. The second case corresponds to the situation when $\lambda=0$ and $\mu=\rt{\gamma}{\rm Tr} [(\hat \sigma+\hat\sigma^\dagger)\rho_{\bf F}(t)]$; that is, the fictitious distribution is treated as the real unobservable distribution. Both cases were simulated to show the robustness of our numerical technique and to demonstrate that while any ostensible distributions can be chosen a more realistic choice will result in a faster convergence. To demonstrate this we numerically solved \erf{supertrajextory} and used this as our reference solution. Then we compared the mean and variance of the classical marginal states and the fidelity for quantum reduced states (using \erf{felquant} once the classical space has been removed) for both ostensible cases and with ensemble sizes of 100 and 10 000. These results are shown Fig. \ref{figsuper2} where it is observed that for the larger ensemble size the difference in the classical marginal state is small and the quantum fidelity is close to one, indicating that our ostensible method has reproduced the known result and is converging. Furthermore it is observed that for the second case for the same ensemble size this difference is smaller thereby indicating that the second method convergence is faster. \begin{figure}\begin{center} \includegraphics*[width=0.45\textwidth]{PartlyObservedEvolutionFig09} \caption{\label{figsuper2} This figure shows the quantum and classical fidelity between the actual solution and our ostensible solution for ${\rho}_{{\bf R}}%_{(0,t]}}(x,t)$. Part A corresponds to the $\lambda=\mu=0$ case while part B represents the $\lambda=0$ and $\mu=\rt{\gamma}{\rm Tr} [(\hat \sigma+\hat\sigma^\dagger)\rho_{\bf F}(t)]$ case. In both case an ensemble size of $n=100$ (dotted) and $n=10 000$ (solid) was used. The system parameters are the same as in Fig. \ref{figsuperI}.}\end{center} \end{figure} \section{Discussion and Conclusion}\label{con} The central topic of this paper was to investigate the conditional dynamics of partially observed systems (classical and quantum). Due to the fact that the information obtained is incomplete we have to assign a mixed state to the system. For a quantum system this means the state of knowledge given result $r$ is given by the state matrix $\rho_{r}(t)$ and for a classical system a probability distribution $P_{r}(x,t)$ has to be used. If we consider a joint system (for example a classical detector is used to monitor a quantum system) the conditional state is given by $\rho_{r}(x,t)$. Even when we consider continuous-in-time monitoring we can still have incomplete information because of unobserved processes. For this case the conditional state trajectories obey either a stochastic master equation (for a quantum system), a Kushner Stratonovich equation (for a classical system) or a superoperator Kushner-Stratonovich equation (for the joint system). That is, to simulate the conditional state we have to solve a rather numerically expensive equation. In this paper we showed that by introducing a fictitious record ${\bf F}}%_{(0,t]}$ for the unobserved processes and ostensible measurement theory we can reduce this problem to solving pure states (stochastic Schr\"odinger~~ equations for the quantum system or stochastic differential equations for the classical system) conditioned on both ${\bf R}}%_{(0,t]}$ and ${\bf F}}%_{(0,t]}$. Then by averaging over all possible ${\bf F}}%_{(0,t]}$ we get the require conditional state. That is the numerical memory requirements are decreased by a factor of $N$, the number of basis states for the system. However, this is at the cost of an ensemble average. In summary, our ostensible method will be useful for investigating realistic situations where the dimensions of the systems are large. It is also much easier to implement numerically than the standard technique, so we expect it to find immediate applications. \acknowledgments We would like to acknowledge the interest shown and help provided by K. Jacobs and N. Oxtoby. This work was supported by the Australian Research Council (ARC) and the State of Queensland.
{ "timestamp": "2007-04-06T23:01:15", "yymm": "0503", "arxiv_id": "quant-ph/0503241", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503241" }
\section{Introduction} \hspace{3ex} It is interesting to consider the elastic rod of a large mass $M$, the left end of which is joined with mass $m << M$ and body of mass $m$ is fixed to the right end of the rod. Then, it is interesting to study the consequences of the application of the the force of the delta-function form to the left side of the rod. The delta-function is chosen for simplicity. This function can be replaced by the different functions. We show that the internal motion of the elastic rod medium is controlled by the wave equation. We derive the mathematical form of the mechanical motion of the considered string or rod. Our problem represents the missing problem in the Newton ``Principia mathematica'' [1] and in any textbook on mechanics. The relation of our theory to the quark-string model of mesons is evident. \section{Classical theory of interaction of particle with an impulsive force} We will first show that use of the impulsive force of the delta-function form is physically meaningful in a classical mechanics of a point particle. We idealize the impulsive force by the Dirac $\delta$-function. Newton's second law in the one-dimensional form for the interaction of a massive particle with mass $m$ with force $F$ $$ma = F\eqno(1)$$ with $F$ being an impulsive force $P\delta(t)$ is as follows: $$m\frac {d^{2}x}{dt^{2}} = P\delta(\alpha t),\eqno(2)$$ where $P$ and $\alpha$ are some constants, with MKSA dimensionality [$P$] = ${\rm kg.m.s}^{-2}$, [$\alpha$] = ${\rm s}^{-1}$. We put $|\alpha| = 1$. Using the Laplace transform [2] in the last equation, with $$\int_{0}^{\infty}e^{-st}x(t)dt \stackrel{d}{=} X(s), \eqno(3)$$ $$\int_{0}^{\infty}e^{-st}{\ddot x}(t)dt = s^{2}X(s) - sx(0) - {\dot x}(0), \eqno(4)$$ $$\int_{0}^{\infty}e^{-st}\delta(\alpha t)dt = \frac{1}{\alpha}, \eqno(5)$$ we obtain: $$ms^{2}X(s) - msx(0) - m\dot x(0) = P/\alpha.\eqno(6)$$ For a particle starting from the rest with $\dot x(0) = 0, x(0) = 0$, we get $$X(s) = \frac{P}{ms^{2}\alpha}.\eqno(7)$$ Using the inverse Laplace transform, we obtain $$x(t) = \frac{P}{m\alpha}t\eqno(8)$$ and $$\dot x(t) = \frac{P}{m\alpha}.\eqno(9)$$ In case of the harmonic oscillator with the damping force and under influence of the general force $F(t)$, the Newton law is as follows: $$m\frac {d^{2}x(t)}{dt^{2}} + b{\dot x}(t) + kx(t) = F(t).\eqno(10)$$ After application of the Laplace transform (3) and with regard to the same initial conditions as in the preceding situation, $\dot x(0) = 0, x(0) = 0$, we get the following algebraic equation: $$ms^{2}X(s) +bsX(s) + kX(s) = F(s),\eqno(11)$$ or, $$X(s) = \frac{F(s)}{m\omega_{1}} \frac {\omega_{1}}{(s + b/2m)^{2} + \omega_{1}^{2}} \eqno(12)$$ with $\omega_{1}^{2} = k/m - b^{2}/4m^{2}$. Using inverse Laplace transform denoted by symbol ${\cal L}^{-1}$ applied to multiplication of functions $f_{1}(s)f_{2}(s)$, $${\cal L}^{-1}(f_{1}(s)f_{2}(s)) = \int_{0}^{t} d\tau F_{1}(t-\tau)F_{2}(\tau), \eqno(13)$$ we obtain with $f_{1}(s) = F(s)/m\omega_{1}, \quad f_{2}(s) = \omega_{1}/ ((s + b/2m)^{2} + \omega_{1}^{2}), \quad F_{1}(t) = F(t)/m\omega_{1},\\ F_{2}(t) = \exp{(-bt/2m)}\sin\omega_{1}t$. $$x(t) = \frac {1}{m\omega_{1}}\int_{0}^{t}F(t-\tau)e^{-\frac {b}{2m}\tau} \sin(\omega_{1}\tau)d\tau.\eqno(14)$$ For impulsive force $F(t) = P\delta(\alpha t)$, we have from the last formula $$x(t) = \frac {(P/\alpha)}{m\omega_{1}}e^{-\frac {b}{2m}t}\sin\omega_{1}t. \eqno(15)$$ \section{The pulse propagating in a rod} \hspace{3ex} In this section we will solve the motion of a string or rod with the massive ends (the body with mass $m$ is fixed to the every end of the string) on the assumption that the tension in the string is linear and the applied force is of the Dirac delta-function. First, we will derive the Euler wave equation from the Hook law of tension and then we will give the rigorous mathematical formulation of the problem. Linearity of the wave equation enables to solve this problem by the Laplace transform method. We follow [3] and the author preprint [4] where this method was used to solve the Gassendi model of gravity. Although Gassendi [5] is known in physics as the founder of the modern atomic theory of matter, his string model of gravity was not accepted. The Newton reaction to this model was empirical. He said: ``Hypotheses non fingo''. It seems that Gassendi ideas was applied later by Faraday in his theory of electromagnetism. We know also that Gassendi was independent thinker and he was persecuted. Every independent thinker is persecuted in any society. The present problem can be also defined as a central collision of two bodies (balls). While in the basic mechanics the central collision is considered as a contact collision of the two balls, here, the collision is mediated by the string, or, rod. To our knowledge, the present problem is not involved in the textbooks of mathematical physics or in the mathematical journals. This problem was not possible to define and solve in the Newton period, because the method of solution is based on the Euler partial wave equation, the Laplace transform, The Riemann-Mellin transform, the Bromwich integral and Bromwich contour and other ingredients of the operator calculus which was elaborated after the Newton period. So, this is why the problem is not involved in the Newton ``Principia Mathematica'' [1]. Now, let us consider the rod (or string) of the length $L$, the left end of which is joined with mass $m$ and the right end is joined with mass $m$. The force of the delta-function form is applied to the left end and the initial state of the rod is the sate of equilibrium. The deflection of the rod element $dx$ at point $x$ and time $t$ let be $u(x,t)$ where $x\in(0,L)$. The differential equation of motion of string elements can be derived by the following way [3]. We suppose that the force acting on the element $dx$ of the string is given by the law: $$T(x,t) = ES\left(\frac {\partial \*u}{\partial x}\right), \eqno(16)$$ where $E$ is the modulus of elasticity, $S$ is the cross section of the string. We easily derive that $$T(x+dx)-T(x) = ESu_{xx}dx. \eqno(17)$$ The mass $dm$ of the element $dx$ is $\varrho ESdx$, where $\varrho = const$ is the mass density of the string matter and the dynamical equilibrium gives $$\varrho\*Sdx u_{tt} = ESu_{xx}dx. \eqno(18)$$ So, we get $$\frac {1}{c^2}u_{tt} - u_{xx} = 0; \quad c = \left(\frac {E}{\varrho}\right)^{1/2}. \eqno(19)$$ Now, we get the problem of the mathematical physics in the form: $$u_{tt} = c^{2}u_{xx}\eqno(20)$$ with the initial conditions $$u(x,0) = 0; \quad u_{t}(x,0) = 0\eqno(21)$$ and with the boundary conditions $$mu_{tt}(0,t) = au_{x}(0,t) + P\delta(\alpha t);\quad mu_{tt}(L,t) = au_{x}(L,t),\eqno(22)$$ where we have put $$a = - ES;\quad P = {\rm some\; constant}.\eqno(23)$$ The delta-function can be approximatively realized by the strike of the hammer to the left end of the rod. The equation (20) with the initial and boundary conditions (21) and (22) represents one of the standard problems of the mathematical physics and can be easily solved using the Laplace transform [2]: $$\hat L u(x,t) \stackrel{d}{=}\int_{0}^{\infty}e^{-pt} u(x,t)dt \stackrel{d}{=} \varphi(x,p). \eqno(24)$$ Using (24) and (20) we get: $$\hat Lu _{tt}(x,t) = p^{2}\varphi(x,p) - pu(x,0) - u_{t}(x,0) = p^{2}\varphi(x,p),\eqno(25)$$ $$\hat L u_{xx}(x,t) = \varphi_{xx}(x,p);\quad \hat L \delta(\alpha, t) = 1/\alpha . \eqno(26)$$ After elementary mathematical operations we get the differential equation for $\varphi$ in the form $$\varphi_{xx}(x,p) - k^{2}\varphi(x,p) = 0; \quad k = p/c. \eqno(27)$$ with the boundary condition in eq. (22). We are looking for the the solution of eq. (27) in the form $$\varphi(x,p) =c_{1}\cosh k x + c_{2}\sinh k x .\eqno(28)$$ We get from the boundary conditions in eq. (22) $$c_{1} = \frac {1}{p}\;\frac {ac(P/\alpha) \cosh (pL/c) - (P/\alpha)mpc^2\sinh (pL/c)} {\sinh(pL/c)(a^2 - m^2 p^2 c^2)},\eqno(29)$$ $$c_{2} = -\frac {(P/\alpha)c}{ap} + \frac {(P/\alpha)mac^2\cosh(pL/c) - (P/\alpha)pm^2c^3\sinh (pL/c)}{a\sinh (pL/c) (a^2 - m^2 c^2 p^2)}.\eqno(30)$$ The corresponding $\varphi(x,p)$ is of the form: $$\varphi(x,p)= \frac {1}{p}\;\frac {ac(P/\alpha)\cosh (pL/c) - (P/\alpha)mpc^2\sinh(pL/c)} {\sinh (pL/c)(a^2 - m^2 p^2 c^2)}\cosh (px/c)\quad + $$ $$\left[-\frac {(P/\alpha)c}{ap} + \frac {a(P/\alpha)mc^2\cosh(pL/c) - bpm^2c^3\sinh (pL/c)}{a\sinh (pL/c)(a^2 - m^2 c^2 p^2)}\right]\sinh(px/c). \eqno(31)$$ The corresponding function $u(x,t)$ follows from the theory of the Laplace transform as the mathematical formula (res is residuum)[2]: $$u(x,t) = \frac {1}{2\pi i}\oint e^{pt}\varphi(x,p) dp = \sum_{p=p_{n}}{\rm res}\;e^{pt}\varphi(x,p) = $$ $$\sum _{p=p_{n}}{\rm res}\;e^{pt} \frac {1}{p}\;\frac {ac(P/\alpha) \cosh (pL/c)} {\sinh (pL/c)(a^2 - m^2 p^2 c^2)}\cosh (px/c) \quad -$$ $$\sum _{p=p_{n}}{\rm res}\;e^{pt} \frac {(P/\alpha)mc^2} {(a^2 - m^2 p^2 c^2)}\cosh (px/c)\quad - $$ $$\sum _{p=p_{n}}{\rm res}\;e^{pt} \left[\frac {(P/\alpha)c}{ap}\right]\sinh (px/c)\quad + $$ $$\sum _{p=p_{n}}{\rm res}\;e^{pt} \left[\frac {m(P/\alpha)c^2\cosh (pL/c)}{\sinh (pL/c)(a^2 - m^2 p^2 c^2)}\right]\sinh( px/c)\quad - $$ $$\sum _{p=p_{n}}{\rm res}\;e^{pt} \left[\frac{(P/\alpha)pm^2c^3}{a}\;\frac {1}{(a^2 - m^2 c^2p^2)}\right] \sinh (px/c)\quad =$$ $$u_{1} - u_{2} - u_{3} + u_{4} - u_{5}, \eqno(32)$$ where $$u_{j} = \sum {\rm res}\; e^{pt}\frac{A_{j}}{B_{j}}; \quad j = 1, 2, 3, 4, 5 \eqno(33)$$ and $$A_{1} = {ac(P/\alpha) \cosh (pL/c)}\cosh (px/c);\quad B_{1} = p\sinh (pL/c)(a^2 - m^2 p^2 c^2)\eqno(34)$$ $$A_{2} = (P/\alpha)mc^2\cosh (px/c); \quad B_{2} = (a^2 - m^2 p^2 c^2)\eqno(35)$$ $$\quad A_{3} = (P/\alpha)c\sinh(px/c); \quad B_{3} = ap \eqno(36)$$ $$A_{4} = (P/\alpha)mc^2\cosh (pL/c)\sinh (px/c); \quad B_{4} = \sinh (pL/c)(a^2 - m^2 p^2 c^2)\eqno(37)$$ $$A_{5} = (P/\alpha)pm^2c^3\sinh (px/c); \quad B_{5} = a(a^2 - m^2 p^2 c^2).\eqno(38)$$ We know from the theory of the complex functions that if the pole of some function $f(z)/g(z)$ is simple and it is at point $a$, then the residuum is as follows [2]: $${\rm residuum} = \frac{f(a)}{g'(a)}.\eqno(39)$$ If the pole at point $a$ of the function $f(z)$ is multiply of the order $m$, then the residuum is defined as follows: $${\rm residuum} = \frac{1}{(m-1)!} \lim_{z\to a}\frac{d^{m-1}}{dz^{m-1}} \left[(z-a)^m f(z) \right].\eqno(40)$$ Let us first determined the function $$u_{1} = \sum {\rm res}\; e^{pt}\frac{A_{1}}{B_{1}}.\eqno(41)$$ Poles of $B_{1}$ are at points $p=0$, this is pole of the order 2, $p = + a/mc, p = -a/mc$ and $p_{n} = + i\pi nc/L, p_{n} = -i\pi nc/L, n = 1, 2, 3, ...$. So, the function $u_{1}$ is as follows: $$u_{1} = \frac{(P/\alpha)c^2}{La}t - \frac{(P/\alpha)c}{a}\cosh\left(\frac{aL}{mc^2}\right) \cosh\left(\frac{ax}{mc^2}\right)\sinh\left(\frac{at}{mc}\right) \quad + $$ $$\sum_{n=1}^{n= \infty} \frac{2a(P/\alpha)c}{\pi n} \frac{L^2}{a^2 L^2 + m^2\pi^2 n^2 c^4}\cos\left(\frac{\pi nx}{L}\right) \sin\left(\frac{\pi n c t}{L}\right). \eqno(42)$$. For the function $u_{2}$ we get: $$u_{2} = \left(-\frac{(P/\alpha)c}{a}\right)\sinh\left(\frac{at}{mc}\right)\cosh \left(\frac{ax}{mc^2}\right).\eqno(43)$$ $$u_{3} = 0 .\eqno(44)$$ For $u_{4}$ and $u_{5}$ we get: $$u_{4} = \left(-\frac{(P/\alpha)c}{a}\right) \coth\left(\frac{aL}{mc^2}\right)\sinh\left(\frac{ax}{mc^2}\right) \sinh\left(\frac{at}{mc}\right)\eqno(45)$$ $$u_{5} = \left(-\frac{(P/\alpha)c}{a}\right) \sinh\left(\frac{ax}{mc^2}\right) \sinh\left(\frac{at}{mc}\right)\eqno(46)$$ The dimensionality of $u$ is [$u$] = m and $u(x,0) = 0$. The momentum of a left particle $p = mu(0,t)$, or right particle $p = mu(L,t)$ is not conserved. Only the total momentum of a system is conserved. \section{Discussion} Our problem is the modification of some problems involved in the textbooks on mathematical physics. However, our approach is pedagogically original in the sense that we use the initial force of a delta-function form to show the internal motion of the string, or, rod. The delta-function form of electromagnetic pulse was used by author in [6] and [7] to discuss the quantum motion of an electron in the laser pulse. We have considered here the real strings and rods in the real space and we do not use extra-dimensions and unrealistic strings. The M-dimensional geometrical object cannot be realized in N-dimensional space for M $>$ N [8]. The mathematical theory of unrealistic strings is well known as the string theory in particle physics. Our problem with the real strings and rods can be generalized for the two-dimensional and three-dimensional situation. It can be also generalized to the situation with the dissipation of waves in the strings and rods. In this case it is necessary to write the wave equation with the dissipative term and then to solve this problem ``ab initio''. While we have solved the problem for the situation where the pulse was generated by the force of the delta-form, we give some general ideas following from the wave equation. It is well known that the solution of this equation is in general in the form [9]: $$f\left(t - \frac{x}{c}\right); \quad g\left(t + \frac{x}{c}\right) \eqno(47)$$ where functions $f, g$ are general. It means it involves also the function of he delta-form. For the wave propagating from the left side to the right side, we take function $f$. The corresponding tension in the rod is $$T = ESu_{x}(x, t) = ESf'\left(t - \frac{x}{c}\right) \left(\frac{-1}{c}\right).\eqno(48)$$ We easily see that $T(x = 0, t= 0) = T(x = L,t = L/c )$, and it means that when the pulse force is created at the left end of the rod then it propagates in the rod and after time $L/c$ it is localized in the right end of the rod. However, we have seen that the pulse force generated in the system with the massive ends of strings or rods develops in time according to laws of the mathematical physics of strings and rods and cannot be intuitively predicted. Only rigorous solution of the dynamics of the system can give the answer on the real motion of tension in the string. There is no information in the Newton ``Principia mathematica '' [1] and in any textbook on mechanics on the central collision of two particles where the force is mediated by string or rod. Similarly, there is not the solution of our problem in the famous monograph by Pars [10]. So, This is the missing problem in the textbooks on mechanics. The propagation of a pulse in one direction was confirmed experimentally by author using the heavy elastic rod (the segment of a rail). The delta-form force (tension) was generated approximately by the strike of hammer. The experiment was performed as the table experiment and it can be repeated by any theorist. The proposed model with the string with massive ends can be also related in the modified form to the problem of the radial motion of quarks bound by a strings, and used to calculate the excited states of such system. The resent analysis of such problem was performed by Lambiase and Nesterenko [11] and Nesterenko and Pirozhenko [12], and others. So, can we hope that our approach and their approach will be unified to generate the new revolution of the string theory of matter and space-time? Why not? \vspace{5mm}
{ "timestamp": "2005-03-01T23:12:12", "yymm": "0503", "arxiv_id": "math-ph/0503003", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503003" }
\section{Introduction} It is a well-known consequence of the simultaneous uniformisation theorem of Bers \cite{Bers} that given two abstractly isomorphic Fuchsian groups $G_1\subset PSL_2(\mbox{$\mathbb R$})$ and $G_2\subset PSL_2(\mbox{$\mathbb R$})$, acting on the upper and lower complex half-planes respectively, each having limit set $\hat{\mathbb R}={\mathbb R} \cup \infty$, and such that the action of $G_1$ on $\hat{\mathbb R}$ is topologically conjugate to that of $G_2$, the actions of $G_1$ and $G_2$ can be {\it mated} to obtain a quasifuchsian Kleinian group $G \subset PSL_2(\mbox{$\mathbb C$})$. This {\it mating} is a group which is abstractly isomorphic to both $G_1$ and $G_2$, it has limit set $\Lambda(G)$ a simple closed (fractal) curve, and the actions of $G$ on the two components of $\Omega={\hat{\mathbb C}}-\Lambda$ are conformally conjugate to those of $G_1$ on ${\mathcal U}$ and $G_2$ on ${\mathcal L}$. \medskip It is also well-known that given two polynomial maps $P$ and $Q$ of the same degree $n$, in appropriate circumstances one can find a rational map $R$ which realises a {\it mating} between the actions of $P$ and $Q$ on their filled Julia sets, in a precise sense as defined for example in \cite{HT}. A necessary condition for a mating between two quadratic polynomials $P:z \to z^2+c$ and $Q:z \to z^2+c'$ to exist is that $c$ and $c'$ should not belong to conjugate limbs of the connectivity locus (the Mandelbrot Set) in parameter space: this was first shown also to be a sufficient condition in the case that $P$ and $Q$ are {\it postcritically finite} \cite{ree,T}, and subsequently for much more general classes of $P$ and $Q$ \cite{HT}. \medskip In \cite{BP} the first examples of holomorphic correspondences realising {\it matings} between Fuchsian groups and polynomials were presented. {\it Holomorphic correspondences} on the Riemann sphere are multi-valued maps $f:z\to w$ defined by polynomial equations $p(z,w)=0$. Examples of holomorphic correspondences are those defined by a union of the graphs of some finite set of M\"obius transformations, or by the graph of a rational map (or its inverse). We say that such a correspondence has {\it bidegree} $(m:n)$ if a generic point $z$ has $n$ images $w$ and a generic point $w$ has $m$ inverse images $z$. \medskip {\bf Definition} {\it Let $q_c:z \to z^2+c$ be a quadratic map with connected filled Julia set $K(q_c)$. A holomorphic correspondence $f:z \to w$ of bidegree $(2:2)$ is called a {\it mating} between $q_c$ and the modular group $PSL_2(\mbox{$\mathbb Z$})$ if: \medskip (a) there exists a completely invariant open simply-connected region $\Omega \subset \hat{\mbox{$\mathbb C$}}$ and a conformal bijection $h$ from $\Omega$ to the upper half-plane conjugating the two branches of $f\vert_\Omega$ to the pair of generators $z \to z+1,\ z \to z/(z+1)$ of $PSL_2(\mbox{$\mathbb Z$})$; \medskip (b) the complement of $\Omega$ is the union of two closed sets $\Lambda_-$ and $\Lambda_+$, which intersect in a single point and are equipped with homeomorphisms $h_\pm: \Lambda_\pm \to K(q_c)$, conformal on interiors, respectively conjugating $f$ restricted to $\Lambda_-$ as domain and codomain to $q_c$ on $K(q_c)$, and conjugating $f$ restricted to $\Lambda_+$ as domain and codomain to $q_c^{-1}$ on $K(q_c)$.} \medskip In \cite{BP} the one parameter family of correspondences $$\left(\frac{az+1}{z+1}\right)^2+\left(\frac{az+1}{z+1}\right) \left(\frac{aw-1}{w-1}\right)+\left(\frac{aw-1}{w-1}\right)^2=3 \leqno(1) $$ was shown to contain examples of matings between quadratic maps and the modular group. The following conjecture is implicit in the discussion in Sections 1 and 6 of that paper. \medskip {\bf Conjecture 1} {\it The family $(1)$ of $(2:2)$ correspondences contains matings between $PSL_2(\mbox{$\mathbb Z$})$ and {\it every} quadratic polynomial having a connected Julia set, that is to say every $z \to z^2+c$ with $c \in {\mathcal M}$, the {\it Mandelbrot set}.} \medskip Supporting evidence was provided by proofs for particular examples and numerical experiments suggesting the resemblance between the space of matings and the Mandelbrot set. However difficulties in adapting the theory of {\it polynomial-like maps} \cite{DH} to the setting of {\it pinched polynomial-like maps} prevented a proof. \medskip A different question turned out to be easier to answer. The modular group may be considered as a representation of the free product $C_2*C_3$ of cyclic groups, of orders two and three, in $PSL_2(\mbox{$\mathbb C$})$. Up to conjugacy there is a one parameter family of such representations and in the parameter space there is a set ${\mathcal D}$, homeomorphic to a once-punctured closed disc, for which the representation is discrete and faithful. The modular group corresponds to a particular {\it boundary} point of $\mathcal D$. Let $r$ be any representation of $C_2*C_3$ corresponding to a parameter value in the {\it interior} ${\mathcal D}^\circ$ of $\mathcal D$. The ordinary set $\Omega(r)$ of the Kleinian group defined by such a representation $r$ is connected and the limit set $\Lambda(r)$ is a Cantor set. In \cite{BH} the notion of a mating between such a representation $r$ of $C_2*C_3$ and a quadratic polynomial $q_c: z \to z^2+c$ was introduced: $\Lambda_-$ and $\Lambda_+$ are now disjoint, and their complement $\Omega$ is canonically associated to $\Omega(r)$ (see Section 2.2). By the application of polynomial-like mapping theory the following analogue of Conjecture 1 was proved in \cite{BH}. \medskip \begin{thm}\label{mating} For every quadratic map $q_c:z \to z^2 + c$ with $c \in {\mathcal M}$ and every faithful discrete representation $r$ of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ having connected ordinary set, there exists a polynomial relation $p(z,w)=0$ defining a $(2:2)$ correspondence which is a mating between $q_c$ and $r$. \end{thm} An outline of the proof of Theorem \ref{mating} is presented in Section 2.2, as a prelude to applying pinching techniques to the matings it shows to exist. \medskip We describe an involution $J$ on $\hat{\mathbb C}$ as {\it compatible} with a mating $f$ if $(J\circ f) \cup I_{\hat{\mathbb C}}$ is an equivalence relation, where $I_{\hat{\mathbb C}}$ denotes the identity map on $\hat{\mathbb C}$ and $(J\circ f) \cup I_{\hat{\mathbb C}}$ denotes the $3:3$ correspondence defined by the algebraic curve $$p(z,J(w))(z-w)=0$$ (Here $p(z,w)=0$ is the curve defining $f$.) \begin{prop}\label{compatible} Every mating with a compatible involution is conjugate to a correspondence in the following two parameter family (also considered in \cite{BP}): $$\left(\frac{az+1}{z+1}\right)^2+\left(\frac{az+1}{z+1}\right) \left(\frac{aw-1}{w-1}\right)+\left(\frac{aw-1}{w-1}\right)^2=3k \leqno(2) $$ \end{prop} \medskip As we shall see, the matings constructed in \cite{BH} have compatible involutions, so they have representatives in the family (2), a fact observed in \cite{BH} but for which Proposition \ref{compatible} (proved in Section 2) provides a more conceptual setting. \medskip The basic idea of pinching can be seen in the process by which the modular group can be obtained from any chosen standard representation $r_*$ of $C_2*C_3$ lying in the interior of $\mathcal D$, that is to say a faithful discrete representation with connected ordinary set $\Omega(r_*)$ (and therefore limit set a Cantor set). We first recall that each Kleinian representation of $C_2*C_3$ comes equipped with a canonical involution $\chi$ which conjugates the generators $\sigma\in C_2$ and $\rho\in C_3$ to their inverses (see Section 2.1); we let $G$ denote the group $<\chi,\sigma,\rho>$. For each rational number $p/q$ there is an arc $\delta_{p/q}$ on the orbit space $\Sigma=\Omega(r_*)/G$ which lifts to simple closed geodesic $\tilde{\delta}_{p/q}$ of winding number $p/q$ on a certain torus $\tilde{\Sigma}$ double-covering $\Sigma$ (see Lemma \ref{arcs-exist} in Section 3.1 for details). The arc $\delta_{p/q}$ lifts to an arc $\alpha_{p/q}$ on $\Omega(r_*)$ together with its translates under $G$. This arc $\alpha_{p/q}$ is {\it precisely $<<g>>$-invariant} for any loxodromic element $g \in G$ which stabilises it. (Here $<<g>>$ denotes the maximal elementary subgroup of $G$ containing $g$, and saying that an arc $\alpha$ is {\it precisely $<<g>>$-invariant} means that $<<g>>\alpha=\alpha$ and $h(\alpha)\cap \alpha = \emptyset$ for all $h\in G$ not in $<<g>>$). In this situation Maskit's Theorem \cite{M} states that the representation of $G$ in $PSL_2(\mbox{$\mathbb C$})$ can be deformed to one in which $\alpha_{p/q}$ and its translates under $G$ are pinched to points and $g$ becomes parabolic. We deduce that we may pinch $\delta_0$, and hence its lift $\alpha_0$, to a point, thereby deforming the representation $r_*$ of $C_2*C_3$ to the representation $PSL_2(\mbox{$\mathbb Z$})$, which lies on the boundary of the deformation space $\mathcal D$. Similarly for $p/q \ne 0$ we may pinch $\delta_{p/q}$ to a point and so deform the representation $r_*$ to a faithful discrete representation which we denote $r_{p/2q}$. This has ordinary set a disjoint union of a countable infinity of open round discs, and limit set a circle-packing. The representation $r_{p/2q}$ depends only on the value of $p/2q$ mod $2$: pinching $\delta_{(2nq+p)/q}$ in place of $\delta_{p/q}$ amounts to approaching the same limit representation $r_{p/2q}$ but by a non-isotopic path in ${\mathcal D}$. We remark that by a deep result of McMullen \cite{mc2} the representations $r_{p/2q}$ are dense in the boundary of ${\mathcal D}$. \medskip Recently, Ha\"{\i}ssinsky \cite{H2}, Cui \cite{cui} and Ha\"{\i}ssinsky and Tan \cite{HT} proved analogous results to Maskit's in the context of rational maps, showing that, under appropriate hypotheses, given a rational map $R$ and an $R$-invariant union of arcs joining attracting to repelling cycles, one can continuously deform the map in such a way that the arcs, and their pre-images, are pinched to points and the cycles become parabolic. \medskip In this paper we adapt the techniques of \cite{H2} and \cite{HT} to apply them to the holomorphic correspondences constructed in \cite{BH}. In Section 3, for any correspondence $p_0(z,w)=0$ which is a mating between $r_*$ and $q_c$, and for any rational number $p/q$, we identify an arc $\gamma_{p/q}$ such that the grand orbit of $\gamma_{p/q}$ under the correspondence is a union of infinitely many disjoint copies of $\gamma_{p/q}$ (or of copies of a quotient of $\gamma_{p/q}$ by an involution), and such that pinching each connected component of this union to a point corresponds to deforming the representation $r_*$ to $r_{p/2q}$. We describe the pinching process formally as follows. \medskip {\bf Definition} {\it A convergent pinching deformation for $\gamma_{p/q}$ is a family of quasiconformal maps $(\varphi_t)_{0\le t <1}$ of the Riemann sphere such that the conjugate correspondences $p_t$ defined by $$p_t(z,w)= p_0(\varphi_t^{-1}(z),\varphi_t^{-1}(w))$$ are holomorphic and satisfy the following\,: \begin{itemize} \item $(p_t,\varphi_t)$ are uniformly convergent to a pair $(p_1,\varphi_1)$ as $t$ tends to $1$ , \item the non-trivial fibres of $\varphi_1$ are exactly the closure of the connected components of the orbit of $\gamma_{p/q}$. \end{itemize}} \medskip There are two technical conditions that we require the quadratic map $q_c$ to satisfy in order to apply the techniques of \cite{HT} to $\gamma_0$: \medskip (i) if the critical point $0$ of $q_c$ is recurrent, the $\beta$-fixed point of $q_c$ is not in the $\omega$-limit set of $0$; \medskip (ii) $q_c$ is weakly hyperbolic, that is, there are constants $r>0$ and $\delta<\infty$ such that, for all $z\in J_q\smallsetminus \{\text{preparabolic points}\}$, there is a subsequence of iterates $(q^{n_k})_k$ such that $$\mbox{deg}(W_k(z)\stackrel{q^{n_k}}{\longrightarrow} D(q^{n_k}(z),r) )\le \delta\,$$ where $W_k(z)$ is the connected component of $q^{-n_k}(D(q^{n_k}(z),r) )$ containing $z$. \medskip In Section 4 we prove: \medskip \begin{thm}\label{simple} Let $p_0(z,w)$ be a mating between the representation $r_*$ and $q_c$, where $q_c$ satisfies conditions (i) and (ii) above. Then there exists a pinching deformation of $p_0$ such that $(p_t)_{0 \le t <1}$ converges uniformly to a mating $p_1$ between $PSL_2(\mbox{$\mathbb Z$})$ and $q_c$.\end{thm} \begin{figure} \begin{center} \includegraphics{pinchfig1.eps} \caption{A mating of a representation of $C_2*C_3$ with a Douady rabbit (and zoom). The arc $\gamma_0$ and its images are shown. Pinching these gives a mating of $PSL_2(\mbox{$\mathbb Z$})$ with the rabbit, by Theorem 2.} \end{center} \end{figure} \medskip {\bf Corollary} {\it Conjecture 1 is true for all quadratic maps $q_c$ which satisfy conditions (i) and (ii).} \medskip The class of {\it weakly hyperbolic} quadratic maps is quite large: for example it contains all quadratic maps which satisfy the {\it Collet-Eckmann} condition \cite{pr}, and those which contain parabolic points. \medskip We next investigate pinching $\gamma_{p/q}$, for $p/q \ne 0$. In Section 3.2, we define the notion of a mating between the circle-packing representation $r_{p/2q}$ of $C_2*C_3$ and $q_c$. This generalises our earlier definition of a {\it mating between $PSL_2(\mbox{$\mathbb Z$})$ and $q_c$}, replacing $K(q_c)$ by a certain identification space $K(q_c)/\sim_{p/q}$ and replacing the condition that $\Lambda_+\cap\Lambda_-$ be a point by the condition that it consist of $q$ points (the $p/q$ {\it Sturmian orbit} on the boundary of $K(q_c)$). We show that a mating between $r_{p/2q}$ and $q_c$ depends only on $p/q$ mod $1$. To avoid technical difficulties we restrict attention to the special case that the quadratic map is $q_0:z \to z^2$. Using the techniques of \cite{H2}, we prove the following: \begin{thm}\label{rat} Let $p_0(z,w)$ be a mating between the representation $r_*$ and $q_0$, and let $p/q$ be any rational. Then there exists a pinching deformation of $p_0$ such that $(p_t)_{0 \le t <1}$ converges uniformly to a mating $p_1$ between the circle-packing representation $r_{p/2q}$ of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ and $q_0$. \end{thm} The following is the natural generalisation of Conjecture 1. \medskip {\bf Conjecture 2} {\it For every $0\le p/q<1$, the family $(2)$ of $(2:2)$ correspondences contains matings between the circle-packing representation $r_{p/2q}$ and every quadratic polynomial $q_c$ which has $c \in {\mathcal M}\setminus{\mathcal M}_{1-p/q}$, where ${\mathcal M}_{1-p/q}$ denotes the $(1-p/q)$-limb of the Mandelbrot set ${\mathcal M}$.} \medskip The condition that $c$ does not lie in ${\mathcal M}_{1-p/q}$ is necessary for elementary topological reasons. One might hope to generalise the techniques of the present paper to prove Conjecture 2 in the case that $q_c$ satisfies conditions (i) and (ii) of the hypotheses of Theorem \ref{simple}, but the technical details could be formidable. \medskip {\bf Warning} As will already be apparent, certain of the constructions and results in this article depend on $p/q \in {\mathbb Q}$, certain depend only on $p/q$ mod $1$ (the class of $p/q$ in ${\mathbb Q}/{\mathbb Z}$), and certain on $p/q$ mod $2$. We shall try to make the dependence clear in each case, but briefly the situation may be summed up as follows. A circle-packing representation $r_{p/2q}$ of $C_2*C_3$ depends on $p/q$ mod $2$ but the route to it (in the moduli space ${\mathcal D}$) given by pinching $\delta_{p/q}$ depends on $p/q \in {\mathbb Q}$. A mating between $r_{p/2q}$ of $C_2*C_3$ and $q_c$ depends only on $p/q$ mod $1$, but again the route to it (in mating space) given by pinching $\gamma_{p/q}$ depends on $p/q \in {\mathbb Q}$. \section{Matings between quadratic maps and representations of $C_2*C_3$} We define what we mean by {\it matings} between quadratic maps and representations of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ which lie in ${\mathcal D}^o$, we recall the main ideas of the proof \cite{BH} of Theorem \ref{mating}, we prove Proposition \ref{compatible}, and we present a group-theoretic description of the `ordinary set' $\Omega(f)$ of a mating. \subsection{Faithful discrete representations with connected ordinary sets} Up to conjugacy each representation $r$ of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ is determined by a single complex parameter, the cross-ratio between the fixed points on $\hat{\mathbb C}$ of the action of the generator $\sigma$ of $C_2$ and those of the generator $\rho$ of $C_3$. Such a representation comes equipped with a (unique) involution $\chi$ which exchanges the two fixed points of $\sigma$ and also those of $\rho$, so that $\chi\sigma=\sigma\chi$ and $\chi\rho=\rho^{-1}\chi$. On the Poincar\'e $3$-ball $\chi$ is simply rotation through $\pi$ around the common perpendicular to the axes of $\sigma$ and $\rho$. Write $G$ for the group $<\sigma,\rho,\chi>$, and note that it has ordinary set $\Omega(G)$ the same as that of $<\sigma,\rho>$. \medskip The faithful discrete actions $r:C_2*C_3 \subset PSL_2(\mbox{$\mathbb C$})$ with connected ordinary set $\Omega(G)$ form a single quasiconformal conjugacy class, the class of representations for which one can find simply-connected fundamental domains for $\sigma$ and $\rho$ with interiors together covering the whole Riemann sphere (the conditions of the simplest form of the Klein Combination Theorem are satisfied) \cite{mas}. Such fundamental domains may be constructed as illustrated in figure 2. \begin{figure} \begin{center} \input{pinchfig2.pstex_t} \caption{A fundamental domain $D_G$ for the group $G=<\sigma,\rho,\chi>$.} \label{fund} \end{center} \end{figure} \medskip Here $P$ and $P'$ are the fixed points of $\rho$, $Q$ and $Q'$ are the fixed points of $\sigma$, $R$ is a fixed point of (the involution) $\chi\rho$ and $S$ and $S'$ are the fixed points of $\chi\sigma$. The lines $l,m$ and $n$, joining $R$ to $S$, $Q$ to $S$ and $R$ to $P$, are chosen such that they are smooth and remain non-intersecting in the quotient orbifold $\Omega(G)/G$. The region bounded by $n,\rho n, \chi n$ and $\chi\rho n$ is a fundamental domain for $\rho$, and the region exterior to the loop made up of $m,\sigma m,\chi m$ and $\chi\sigma m$ is a fundamental domain for $\sigma$. The intersection of these two regions is a fundamental domain for the (faithful) action of $C_2*C_3$ on $\Omega(G)$, and the half $D_G$ of this intersection bounded by $n,l,m,\sigma m, \chi l$ and $\rho n$ is a fundamental domain for the action of $G$. The union of all translates of $D_G$ under elements of $C_2*C_3$ is a topological disc $D$ which is a fundamental domain for the action of $\chi$ on $\Omega(G)$. The complement $\Lambda(G)$ of $\Omega(G)= D \cup \chi(D)$ in $\hat{\mathbb C}$ is a Cantor set. \medskip The orbifold $\Omega(G)/G$ is a sphere $\Sigma$, which has a complex structure with four cone points, which we may also label $P,Q,R,S$, where $P$ has angle $2\pi/3$ and $Q,R,S$ each have angle $\pi$. For a given representation of $C_2*C_3$, a set of lines $l,m,n$ as in figure 2 descend to an isotopy class of non-intersecting paths joining the corresponding cone points in $\Sigma$. By considering the choices we may make of the various labels and lines in figure 2 we can obtain a description of $\widetilde{\mathcal D}^0$, the universal cover of the moduli space ${\mathcal D}^0$. \begin{lemma}\label{Kleinian-markings} There is a homeomorphism $\Phi$ between ${\mathcal D}^o$ and the space $\mathcal S$ of spheres $\Sigma$ having a complex structure with four marked cone points $P,Q,R,S$ where $P$ has angle $2\pi/3$ and $Q,R,S$ each have angle $\pi$. This homeomorphism $\Phi$ lifts to a homeomorphism $\tilde{\Phi}$ between $\widetilde{\mathcal D}^o$ and the space $\tilde{\mathcal S}$ of spheres $\Sigma \in {\mathcal S}$ marked with an isotopy class of non-intersecting paths $PR$, $RS$ and $SQ$. \end{lemma} {\bf Proof.} For $r\in {\mathcal D}^o$ define $\Phi(r)$ to be the orbifold $\Omega(G)/G$, where $G=<\sigma,\rho,\chi>$ is the subgroup of $PSL_2(\mbox{$\mathbb C$})$ corresponding to the representation $r$. Clearly $\Phi$ is continuous as ${\mathcal D}^o$ is endowed with the topology induced by its parametrisation by the cross-ratio $(Q,Q';P,P')$. To define an inverse to $\Phi$, observe that given any $\Sigma \in {\mathcal S}$, we may obtain a representation $r$ by regarding $\Sigma$ as a quasiconformal deformation of the orbifold corresponding to $r_*$, lifting the corresponding ellipse field to $\hat{\mbox{$\mathbb C$}}$, and applying the Measurable Riemann Mapping Theorem. \medskip To lift $\Phi$ to a homeomorphism $\tilde{\Phi}$ we have to consider markings. Note that given a representation of $C_2*C_3$ which lies in ${\mathcal D}^o$, there is only one choice for which of the pair $P,P'$ (in figure 2) to label $P$, namely the fixed point of $\rho$ around which the rotation is {\it anticlockwise}. There is also just one choice (up to isotopy) for the arc $n$. The labels $Q$ and $Q'$ are interchangeable (provided that we also interchange the labels $S$ and $S'$), but once a choice has been made for $Q$ the arc $m$ is determined, and even if the labels $Q$ and $Q'$ are exchanged the arc $QS$ in the orbifold $\Sigma$ is unchanged up to isotopy. This just leaves us a choice of the arc $l$ in figure 2. We can alter $l$ to wind an extra $n$ times around the central `hole' for any integer $n$, or $n+1/2$ times if we switch the labels $Q$ and $Q'$. Changing the winding number of $l$ corresponds to choosing a different isotopy class of paths between the points labelled $R$ and $S$ in the orbifold $\Sigma$. $\square$ \medskip Let $t_\alpha$ denote the automorphism of $\widetilde{\mathcal D}^o$ corresponding to turning the internal boundary of figure 2 through an angle $2\pi\alpha$. Note that $t_{1/4}$ moves the pair of points labelled $Q,Q'$ to the pair labelled $S,S'$ and vice versa. Let $\iota:{\mathcal D}^o \to {\mathcal D}^o$ denote the involution obtained by replacing the generating pair $\{\sigma,\rho\}$ of $C_2*C_3$ by $\{\sigma',\rho \}$, where $\sigma'=\chi\sigma$. This corresponds to composing the representation with an outer automorphism of $C_2*C_3$. The following result is now self-evident. \begin{lemma}\label{quarter-twist} The automorphism $t_{1/4}:\widetilde{\mathcal D}^o \to \widetilde{\mathcal D}^o$ covers $\iota:{\mathcal D}^o \to {\mathcal D}^o$, and $t_{1/2}$ generates the group of covering transformations of $\widetilde{\mathcal D}^o \to {\mathcal D}^o$. $\square$ \end{lemma} \subsection{Matings between $q_c$ and $r \in {\mathcal D}^o$} As in the previous subsection, $G$ denotes the group $<\sigma,\rho,\chi>$. \medskip {\bf Definition} {\it A $(2:2)$ holomorphic correspondence $f:z \to w$ is called a {\it mating} between a faithful discrete representation $r$ of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ having connected ordinary set $\Omega(G)$ and a polynomial $q_c:z \to z^2+c$ having connected filled Julia set $K(q_c)$, if the Riemann sphere $\hat{\mathbb C}$ is the disjoint union of a connected open set $\Omega(f)$ and a closed set $\Lambda(f)$ made up of two components, $\Lambda_+(f)$ and $\Lambda_-(f)$ such that each of $\Omega(f)$ and $\Lambda(f)$ is completely invariant under $f$ and: \medskip (a) the action of $f$ on $\Omega(f)$ is discontinuous and there is a conformal bijection between the grand orbit space $\Omega(f)/f$ and $\Omega(G)/G$; \medskip (b) there is a quasiconformal homeomorphism defined from a neighbourhood of $\Lambda_-(f)$ onto a neighbourhood of $K(q_c)$ in $\mbox{$\mathbb C$}$, which realises a hybrid equivalence, conjugating $f$ to $q_c$. Similarly there is a hybrid equivalence between $(f^{-1},\Lambda_+(f))$ and $(q_c,K(q_c))$, this time conjugating $f^{-1}$ to $q_c$.} \medskip (See \cite{DH} for the definition of the term `hybrid equivalence'.) \bigskip The construction of a holomorphic correspondence which realises a mating between given $q_c$ and $r$ proceeds as follows (see \cite{BH} for more details). \medskip We first associate an annulus $A$ to $q_c:z \to z^2+c$. There is a holomorphic conjugacy (the B\"ottcher coordinate) from $z \to z^2$ to $q_c$ on a neighbourhood of $\infty$, fixing the point $\infty$ and tangent to the identity map there \cite{DH1}. An {\it equipotential} for $q_c$ is the image of a circle $\{Re^{2\pi it}: 0 \le t <1\}$ under this conjugacy. It is a smooth Jordan curve parameterized by {\it external angle} $t$. The region bounded by such an equipotential is a simply-connected domain $V$, mapped $2:1$ by $q_c$ onto a larger domain $U \supset \overline{V}$ which also has boundary an equipotential parametrised by external angle. Let $A$ denote the annulus $U-V$, and denote its inner and outer boundaries by $\partial_1 A$ and $\partial_2 A$ respectively. The map $q_c$ sends $\partial_1 A$ two-to-one onto $\partial_2 A$. There is an involution $i:z \to -z$ on $V$ sending each $z \in V$ to the other point which has the same image in $U$ under $q_c$, and there are many choices possible of an orientation-reversing smooth involution $j$ on $\partial_2 A$, a canonical choice being given by $t \to 1-t$ on external angles. \begin{figure} \begin{center} \input{pinchfig3.pstex_t} \caption{The set $D_G \cup \rho(D_G) \cup \rho^{-1}(D_G)$ and its quotient the annulus $B$.} \label{ann} \end{center} \end{figure} \medskip The next ingredient is an annulus $B$ associated to $r$. Recall the fundamental domain $D_G$ constructed above for the group $G=<\sigma,\rho,\chi>$ and illustrated in figure 2. Let $B$ denote the annulus consisting of the three copies $D_G \cup \rho D_G \cup \rho^{-1}D_G$ of $D_G$, with the boundary identifications (induced by $\chi$) indicated in figure 3. The rotations $\rho$ and $\rho^{-1}$ mapping $D_G \cup \rho D_G \cup \rho^{-1} D_G$ to itself descend to a $2:2$ correspondence $g$ on $B$, mapping each $z \in B$ to the pair $\{\rho z, \rho^{-1}z\}$ (or rather to their equivalence classes under the action of $\chi$). The set $D_G$ descends to a `fundamental domain' for the action of $g$ on $B$. The boundary of $B$ is divided into three segments (two inner and one outer, figure 3), each of which is mapped to the other two by $g$. Thus when its domain is restricted to the inner boundary $\partial_1 B$, and its range is restricted to the outer boundary $\partial_2 B$, the correspondence $g$ defines a two-to-one map. When restricted to a correspondence from the inner boundary to itself, $g$ defines a (fixed point free) bijection. Moreover the involution $\sigma$ descends to an involution (which we also denote $\sigma$) on the outer boundary $\partial_2 B$ of $B$, having fixed points $Q$ and $S$. \medskip \begin{lemma}\label{AB} There exists a quasiconformal homeomorphism $h$ from $A$ to $B$ which restricts to a smooth homeomorphism from $\partial A$ to $\partial B$ conjugating the boundary maps $(q_c:\partial_1 A \to \partial_2 A,\ j:\partial_2 A \to \partial_2 A)$ to the boundary maps $(\sigma \circ g:\partial_1 B \to \partial_2 B,\ \sigma: \partial_2 B \to \partial_2 B)$. \end{lemma} \medskip This lemma is proved \cite{BH} by applying standard techniques of Ahlfors and Bers. Now to construct a mating between $q_c$ and $r$ first glue together $U$ and a second copy $U'$ of $U$, via the boundary involution $j$, to obtain a sphere $U \cup U'$, equipped with an involution (also denoted $j$) exchanging $U$ with $U'$ and restricting to the original $j$ on the common boundary. Inside $U'$ is a simply-connected subdomain $V'$ corresponding to $V \subset U$. Let $q_c'=j\circ q_c \circ j:V' \to U'$ denote the quadratic map corresponding to $q_c:V \to U$ and $A'$ denote the annulus $U'-V'$. To define a $2:2$ topological correspondence $f$ on $U \cup U'$ we fit together: $\bullet$ $q_c:V \to U$ (a $2:1$ correspondence); $\bullet$ $(q_c')^{-1}=j\circ q_c^{-1} \circ j:U' \to V'$ (a $1:2$ correspondence); $\bullet$ $j\circ i:V \to V'$ (a $1:1$ correspondence), and $\bullet$ $j\circ g:A \to A'$ (a $2:2$ correspondence), where $g:A \to A$ is the $2:2$ correspondence constructed earlier. Now define an ellipse field on $A$ by using Lemma \ref{AB} to transport the standard complex structure from the annulus $B$. Using $j$ extend this ellipse field to $A'$ and pulling back via $q_c^{-1}$ and $q_c'^{-1}$ extend it to an ellipse field on the whole of $\hat{\mathbb C}-(K(q_c) \cup K(q_c'))$, which transforms equivariantly under the action of the $2:2$ correspondence $f$. Extend this ellipse field to the whole of $\hat{\mathbb C}$ by using the standard complex structure on $K(q_c) \cup K(q_c')$. By applying the Measurable Riemann Mapping Theorem we obtain a complex structure respected by $f$, completing our outline proof of Theorem \ref{mating}. \medskip For any mating $f$ constructed by the method of the proof above, the $3:3$ correspondence $(j\circ f)\cup I_{\hat{\mathbb C}}$ sends each $z\in V$ to the triple of points $\{z,i(z),jq_c(z)\}$, each $z\in A$ to the triple $\{z,g(z)\}$ (recall that $g$ is $2:2$ so $g(z)$ contains two points), and each $z \in U'$ to the triple $\{z,q_c^{-1}j(z)\}$. It is easily checked that each of these triples is the grand orbit under $(j\circ f)\cup I_{\hat{\mathbb C}}$ of any one of its elements, in other words the $3:3$ correspondence is an equivalence relation. The involution $j$ is therefore {\it compatible} with the mating $f$ in the sense defined in Section 1. To show that $f$ is conjugate to a correspondence in the family $(2)$ it now only remains to prove Proposition \ref{compatible}. But a holomorphic correspondence which is an equivalence relation is necessarily the covering correspondence of a rational map, and so there is a rational map $Q$ of degree three such that $(J\circ f)\cup I_{\hat{\mathbb C}}=Cov^Q$ where $$Cov^Q: z \to w \quad \Leftrightarrow \quad Q(w)-Q(z)=0.$$ We deduce that $f=J\circ Cov^Q_0$, where $$Cov^Q_0: z \to w \quad \Leftrightarrow \quad \frac{Q(w)-Q(z)}{w-z}=0.$$ Counting singular points of $f$ now tells us that $Q$ has one double and two single critical points, and that therefore up to pre- and post-compositions by M\"obius transformations $Q$ is the polynomial $Q(z)=z^3-3z$. It follows that up to conjugacy we may write $f$ in the form $$z \to w \quad \Leftrightarrow \quad (Jw)^2+(Jw)z+z^2=3.$$ It is easy to see that if we apply a further conjugacy to transform $J$ to the involution $J(z)=-z$, the equation defining the correspondence $f$ becomes a member of the family $(2)$. This completes the proof of Proposition \ref{compatible}. \subsection{A group-theoretic description of $\Omega$ for a mating} We shall be pinching unions of arcs in $\Omega(f)$ which are lifts of simple closed curves in the grand orbit space $\Omega(f)/f$, where $f$ is one of the matings provided by Theorem 1. With a view to describing these arcs we examine the structure of $\Omega(f)$ and its relationship with $\Omega(G)$. Our first step will be to find a {\it Fuchsian} uniformisation for $\Omega(G)/G$. \medskip Let $\Gamma$ denote the abstract group $<\sigma,\rho,\tau: \sigma^2=\rho^3=\tau^2=(\sigma\rho\tau)^2=1>$. \medskip Let ${\mathcal F}$ denote the moduli space of conjugacy classes of faithful discrete co-compact representations of $\Gamma$ in $PSL_2(\mbox{$\mathbb R$})$ (recall that a Fuchsian group is said to be {\it co-compact} if the quotient of the Poincar\'e disc by its action is compact). An example of a representation of $\Gamma$ which lies in ${\mathcal F}$ is illustrated in figure 4. Let ${\mathcal F}_0$ denote the path-component of ${\mathcal F}$ containing the representation illustrated. Thus a faithful discrete representation of $\Gamma$ lies in ${\mathcal F}_0$ if and only if there is a fundamental domain $D_\Gamma$ for $\Gamma$ isotopic to that illustrated in figure 4, with boundary passing through the fixed points of the corresponding elements of $\Gamma$, in the same order but with the intervening boundary segments no longer necessarily geodesic. \begin{figure} \begin{center} \input{pinchfig4.pstex_t} \caption{A Fuchsian representation of $\Gamma$. Here $P,Q,R$ and $S$ are the fixed points of $\rho,\sigma,\tau$ and $\sigma\rho\tau$. The heavy lines indicate the boundary of $D_{\Gamma_1}$.} \label{Gamma1} \end{center} \end{figure} \medskip Let $\sigma'=\rho\tau\sigma$. Then $\sigma'$, $\rho$ and $\tau$ together generate $\Gamma=<\sigma,\rho,\tau>$, and satisfy the same relations. Changing to the new generating set amounts to applying an (outer) automorphism, which we denote $\beta$, to $\Gamma$. Let $\psi$ be the automorphism of ${\mathcal F}_0$ induced by composing the representation $\gamma \to PSL_2(\mbox{$\mathbb R$})$ with $\beta$ and replacing the boundary of $D_\Gamma$ in figure 3 with that given by moving $S$ up to $Q$ (the fixed point of $\sigma'\rho\tau$), and $Q$ up to $\sigma(S)$ (the fixed point of $\sigma'$), but keeping $P$ and $R$ unchanged. \medskip Recall our description in Section 2.1 of the universal cover $\widetilde{{\mathcal D}^o}$ of the space ${\mathcal D}^o$ of conjugacy classes of faithful discrete representations of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ having connected ordinary set. \begin{prop}\label{Kleinian-Fuchsian} There is a homeomorphism $\Psi:\widetilde{{\mathcal D}^o} \to {\mathcal F}_0$, which carries $t_{1/4}$ to $\psi$ and hence induces homeomorphisms: (i) ${\mathcal D}^o \to {\mathcal F}_0/<\psi^2>$; (ii) ${\mathcal D}^o/\iota \to {\mathcal F}_0/<\psi>$. \end{prop} {\bf Proof.} By Lemma \ref{Kleinian-markings}, Section 2.1, a point of $\widetilde{\mathcal D}^o$ corresponds to an element of $\tilde{\mathcal S}$, that is to say a sphere equipped with a complex structure having cone points $P$ of angle $2\pi/3$, and $Q,R$ and $S$ all of angle $\pi$, together with an isotopy class of paths $PR$, $RS$ and $SQ$. Obviously it suffices to define a homeomorphism between $\tilde{\mathcal S}$ and ${\mathcal F}_0$. \medskip To do this we uniformise each marked orbifold $\Sigma\in \tilde{\mathcal S}$ as a quotient of the Poincar\'e disc $\Delta$ by isometries. The marked arcs on $\Sigma$ lift to a union of arcs, tiling $\Delta$ by translates of a polygon isotopic to that labelled $D_\Gamma$ in figure 4. The group of covering transformations of the projection from $\Delta$ to $\Sigma$ is isomorphic to $\Gamma$ by Poincar\'e's polygon theorem \cite{B}. Conversely, given a faithful discrete representation of $\Gamma$ lying in ${\mathcal F}_0$, its quotient orbifold $\Sigma$ is an element of $\tilde{\mathcal S}$. Thus we have a bijection $\tilde{\mathcal S} \to {\mathcal F}_0$ which, by construction, is continuous and has a continuous inverse. Since $t_{1/4}$ and $\psi$ have identical effects on $\Sigma$, our composite homeomorphism $\Psi:\widetilde{{\mathcal D}^o} \to {\mathcal F}_0$ carries $t_{1/4}$ to $\psi$, and the assertions (i) and (ii) are immediate corollaries. $\square$ \medskip {\bf Remark.} The question of finding explicit formulae for bijections between moduli spaces of representations of Kleinian groups and Fuchsian groups, such as the bijection provided by Proposition \ref{Kleinian-Fuchsian}, is in general highly non-trivial, a classical example being to relate each Schottky group to a Fuchsian group representing the same surface. \bigskip Now let $\Gamma_1 \subset \Gamma$ be the subgroup generated by $\rho\tau$ (which has infinite order), the involution $\rho^{-1}\tau\rho$, and all involutions of the form $W\rho^{-1}\tau\rho W^{-1}$, where $W$ runs through those words in $\sigma$ and $\rho$ which have rightmost letter $\sigma$. Then $\Gamma_1$ has as fundamental domain the region $D_{\Gamma_1}$ bounded by heavy lines in figure 4. Note that $D_{\Gamma_1}/\Gamma_1$ is a topological cylinder, the top edge of the region $D_{\Gamma_1}$ in figure 4 being identified with the bottom edge, each of the arcs on the left hand edge being folded in onto an interval, and each of the arcs on the right hand edge also being folded in onto an interval. \medskip Suppose $f$ is a $2:2$ holomorphic correspondence which is a mating, constructed as in Theorem 1, between a faithful discrete representation of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ having connected ordinary set and a quadratic map $z \to z^2+c$ having connected Julia set. Let $\Gamma\subset PSL_2(\mbox{$\mathbb R$})$ be the Fuchsian representation associated to it by Lemma \ref{Kleinian-Fuchsian}, and let $\Gamma_1$ be the subgroup of $\Gamma$ defined above. \begin{prop}\label{groupify} There is a bi-analytic homeomorphism $$D_{\Gamma_1}/{\Gamma_1}\cong \Delta/{\Gamma_1} \to \Omega(f)$$ carrying the action of the pair $\{\sigma \rho, \sigma\rho^{-1}\}$ on $D_{\Gamma_1}/{\Gamma_1}$ to that of the correspondence $f$ on $\Omega(f)$. \end{prop} {\bf Proof.} From the construction of the mating $f$ in our outline proof of Theorem \ref{mating} (in Section 2.2), it is apparent that $(\Delta,{\Gamma_1})$ uniformises $\Omega(f)$: the set $D_{\Gamma_1} \cup \rho D_{\Gamma_1} \cup \rho^{-1} D_{\Gamma_1}$ in figure 4, when quotiented by the boundary identifications induced by $\Gamma_1$, becomes the annulus $B$ of figure 3, and the maps $\sigma\rho$ and $\sigma\rho^{-1}$ become the two `branches' of the correspondence $f$ on $\Omega(f)$. $\square$ \begin{cor}\label{iota} A mating between $q_c$ and $r\in {\mathcal D}^o$ constructed by the method of Theorem \ref{mating} is canonically isomorphic to a mating between $q_c$ and $\iota(r)$. \end{cor} {\bf Proof.} The outer automorphism $\beta$ defined by replacing the generator $\sigma$ of $\Gamma$ by $\sigma'=\rho\tau\sigma$ stabilises $\Gamma_1$, and the correspondence induced by $\{\sigma\rho,\sigma\rho^{-1}\}$ on $\Delta/\Gamma_1$ is the same as that induced by $\{\sigma'\rho,\sigma'\rho^{-1}\}$, since $\sigma'\rho=\rho\tau\sigma\rho$ and $\sigma'\rho^{-1}=\rho\tau\sigma\rho^{-1}$. $\square$ \medskip {\bf Remarks} \medskip 1. The idea of regarding $\Omega(f)$ as a quotient of $\Delta$ by an infinitely generated Fuchsian group is originally due to Chris Penrose. \medskip 2. We can recover the action of the Kleinian group $G=<\sigma,\rho,\chi>$ on $\Omega(G)$ from the action of the corresponding Fuchsian group $\Gamma=<\sigma,\rho,\tau>$ on $\Delta$, as follows. Take the polygon $D_{\Gamma_2}=D_{\Gamma_1}\cup \rho\tau(D_{\Gamma_1})$ formed by two copies of $D_{\Gamma_1}$, one above the other, identify the top and bottom edges of this polygon to form a cylinder, then fold and glue the left-hand edge together and fold and glue the right hand edge together, to form a sphere. The quotient $D_{\Gamma_2}/\sim$, which can also be described as an orbit space $\Delta/\Gamma_2$ for an appropriate infinitely generated subgroup $\Gamma_2 \subset \Gamma$, is conformally equivalent to $\Omega(G)$. Indeed $\Gamma_2\cong \pi_1(\Omega(G))$, and the projection $\Delta \to \Delta/\Gamma_2$ is the universal cover for $\Omega(G)$. Under the bijection from $D_{\Gamma_2}/\sim$ to $\Omega(G)$ the ends of $D_{\Gamma_2}$ (the cusps) become the points of the limit set $\Lambda(G)$ of the action of the Kleinian group $G$ on $\hat{\mbox{$\mathbb C$}}$. \section{The pinching deformation} \subsection{The arcs to be pinched} To describe the arcs that we shall pinch later, we first fix a standard faithful discrete representation $r_*$ of $C_2*C_3$ having connected ordinary set, and a path $l$ from a fixed point $R$ of $\chi\rho$ to a fixed point $S$ of $\chi\sigma$ (so $R$ and $S$ are as illustrated in figure 2). For convenience we may choose $r_*$ and $l$ so that the corresponding group $\Gamma$ has the reflection symmetry in the horizontal axis apparent in figure 4. Now consider the double cover $\tilde{\Sigma}$ of the orbifold $\Sigma$ ramified at all four cone points. This is a torus, with a single cone point $P$ of angle $4\pi/3$, represented by the central hexagon $D_\Gamma \cup \sigma D_\Gamma$ illustrated in figure 4, with the top edge identified with the bottom edge, and the left-hand edge identified with the right-hand edge. While $\tilde{\Sigma}$ is not itself a quotient of the unit disc $\Delta$ by a subgroup of $PSL_2(\mbox{$\mathbb C$})$ (since the cone point is not of angle $2\pi/n$), nevertheless we may equip ${\tilde{\Sigma}}$ with the metric induced by the restriction of the hyperbolic metric on $\Delta$ to the hexagon $D_\Gamma \cup \sigma D_\Gamma$. The involution $\sigma$ (on $\Delta$) induces an involution $\tilde\sigma$ on $\tilde{\Sigma}$ such that ${\tilde{\Sigma}}/ \tilde{\sigma}={\Sigma}$. \begin{lemma}\label{arcs-exist} For each rational number $p/q$ there is a geodesic arc $\delta_{p/q}$ in ${\Sigma}$ which has end points two of the three cone points of angle $\pi$, which misses the other cone point of angle $\pi$ and the cone point of angle $2\pi/3$, and which has lift $\tilde{\delta}_{p/q}$ to $\tilde{\Sigma}$ a simple closed geodesic of winding number $p/q$. \end{lemma} {\bf Proof.} For each such $p/q$ (in lowest terms), there is a simple closed curve of winding number $p/q$ on the torus ${\tilde{\Sigma}}$, passing through (i) the cone points $Q$ and $S$ if $q$ is even, (ii) the cone points $Q$ and $R$ if $p$ and $q$ are both odd, and (iii) the cone points $R$ and $S$ if $p$ is even. Examples are illustrated in figure 5 for typical cases of each type. \begin{figure} \begin{center} \input{pinchfig5.pstex_t} \caption{The arcs $\tilde{\delta}_{p/q}$ for $p/q=1/2$, $1/3$, $2/3$ and $4/3$ respectively.} \label{winding-nunmbers} \end{center} \end{figure} \medskip Note that when we add an even integer to $p/q$ the new $\delta_{p/q}$ is an arc between the same two cone points on ${\Sigma}$. But when we add an odd integer the roles of $Q$ and $S$ are interchanged. \medskip In every case the simple closed curve on $\tilde{\Sigma}$ can be chosen to be invariant under $\tilde\sigma$. Since it passes through the lifts of two cone points, it descends to an arc on ${\Sigma}$ joining these two points. We define $\delta_{p/q}$ to be a representative of shortest length in the isotopy class of this arc, relative to its end points and the other two cone points on ${\Sigma}$. Note that there must exist such a minimal length example, as arcs which pass through one or both of the other cone points have lengths which are local maxima (since all the cone points have cone angle less that $2 \pi$). $\square$ \medskip Let $A_{p/q}$ denote the lift of $\delta_{p/q}$ to the cylinder $(D_\Gamma \cup \sigma D_\Gamma)/\Gamma_1$ constructed by identifying the top and bottom of the hexagon. Thus $A_{p/q}$ consists of $q$ arcs each running from one boundary circle of this cylinder to the other. Consider the union $\Gamma A_{p/q}$ of all lifts of $\delta_{p/q}$. Recall that $D_{\Gamma_1}/{\Gamma_1}$ is a cylinder, with ends corresponding to $\partial \Lambda_-$ and $\partial \Lambda_+$ (by Proposition \ref{groupify}), that the correspondence $f$ acts on $\partial \Lambda_-$ as a quotient of the doubling map, and that $f^{-1}$ acts on $\partial \Lambda_+$ as a quotient of the doubling map. For simplicity of description assume that $\partial \Lambda_-$ is a topological circle and the action of $f$ on it is that of the doubling map (this is the case when the quadratic map in the mating corresponds to a value of $c$ in the interior of the main cardioid of the Mandelbrot set): obvious adaptations are possible for the cases where $\partial \Lambda_-$ is a proper quotient of the circle. \begin{figure} \begin{center} \input{pinchfig6.pstex_t} \caption{The three arcs linking $\Lambda_-$ to $\Lambda_+$ in the case $p/q=1/3$ (the other images of these arcs under $\Gamma$ are not shown).} \label{arcs-on-omega} \end{center} \end{figure} \medskip If we label the ends of $\partial D_{\Gamma_1}$ by binary sequences as indicated in figure 6 then the folding identifications induced by $\Gamma_1$ impose the usual quotient from the space of binary sequences to the unit circle, carrying the shift to the doubling map. Thus, under our assumption that $\partial \Lambda_-$ is the circle, points of $\partial \Lambda_-$ are labelled (figure 6) in such a way that $f^{-1}:\partial \Lambda_- \to \partial \Lambda_-$ (a $1:2$ correspondence) is defined by ``right shift and insert $0$ or $1$'' according as the branch of $f^{-1}$ is $\rho\sigma$ or $\rho^{-1}\sigma$ respectively, and points of $\partial \Lambda_+$ are labelled in such a way that $f:\partial \Lambda_+ \to \partial \Lambda_+$ (also a $1:2$ correspondence) is defined by ``right shift and insert $0$ or $1$'' according as the branch of $f$ is $\sigma\rho$ or $\sigma\rho^{-1}$ respectively. We adopt the usual notational convention that a bar over a symbol (or group of symbols) indicates the infinite repetition of that symbol (or group of symbols). \medskip {\bf Definition} {\it An infinite sequence of $0$'s and $1$'s is known as {\it Sturmian} if the binary number it represents on the circle has orbit under the doubling map a sequence of points arranged in the same order around the circle as for a rigid rotation.} \medskip One may assign a rotation number to each Sturmian sequence $s$, namely the limit as $n$ tends to infinity of the proportion of the first $n$ digits of $s$ which are $1$'s, or equivalently the rotation number of the rigid rotation having orbit points in the same order as those of $s$. Note that such a rotation number is only defined mod $1$. For each rational $p/q$ (mod $1$) there is a unique periodic Sturmian orbit of rotation number $p/q$ (this was observed by Morse and Hedlund, who introduced the notion of Sturmian sequences). We remark that the points of each periodic Sturmian orbit ${\mathcal O}$ must be contained in an interval of length less than $1/2$ on the circle $\mbox{$\mathbb R$} /\mbox{$\mathbb Z$}$, as the doubling map must preserve the cyclic order of ${\mathcal O}$ (see \cite{BS} for more about this and other properties of Sturmian sequences). \medskip {\bf Examples} \medskip The infinite sequences $\overline{01}$, $\overline{001}$ and $\overline{00101}$ are Sturmian, of rotation numbers $1/2,1/3$ and $2/5$ respectively. \begin{prop}\label{landing-points} $(\Gamma A_{p/q}\cap D_{\Gamma_1})/\Gamma_1$ contains exactly $q$ arcs which join $\Lambda_-$ to $\Lambda_+$. These land on $\partial\Lambda_-$ at points of the unique Sturmian orbit of rotation number $p/q$ (mod $1$) of the $2:1$ map $f:\partial\Lambda_- \to \partial\Lambda_-$ and at the other end they land on $\partial \Lambda_+$ at points of the unique Sturmian orbit of $f^{-1}$ of rotation number $p/q$ (mod $1$). \end{prop} {\bf Proof.} The fact that there are exactly $q$ arcs joining $\Lambda_-$ to $\Lambda_+$ follows at once from the fact that exactly $q$ arcs in $(\Gamma A_{p/q}\cap D_{\Gamma_1})/\Gamma_1$ cross the equator circle of the central cylinder $(D_\Gamma \cup \sigma D_\Gamma)/\Gamma_1$ (the vertical line in the central hexagon in $D_{\Gamma_1}$). The action of the correspondence $f^{-1}=\{\rho^{-1}\sigma, \rho\sigma\}$ on these arcs is to map the $j$th arc to the $(j+p)$th arc for each $j$, where the arcs are counted modulo $q$, from the bottom of the central hexagon upwards. Thus the action of $f^{-1}$ on the landing point of the $j$th arc on $\Lambda_+$ is to send it to the landing point of the $(j+p)$th arc, for each $j$. Similarly $f$ sends the $j$th landing point on $\Lambda_-$ to the $(j+p)$th. $\square$ \medskip {\bf Definition of the arc $\gamma_{p/q}$.} {\it For each $p/q$ we pick as $\gamma_{p/q}$ one of the $q$ components of $(\Gamma A_{p/q}\cap D_{\Gamma_1})/\Gamma_1$ which cross the equator circle of the central cylinder and therefore join $\Lambda_-$ to $\Lambda_+$. For definiteness, when $q$ is odd we take $\gamma_{p/q}$ to be the component which passes through $R$ (the fixed point of $\tau$) and when $q$ is even we take it to be the component which passes through $S$ (the fixed point of $\sigma\rho\tau$). We remark that in the case $p/q=0$ there is just one component crossing the vertical symmetry line of the central hexagon, and it passes through both of these points.} \medskip In figure 6 we illustrate $\gamma_{1/3}$, which joins $\overline{010}\in \Lambda_-$ to $\overline{100}\in \Lambda_+$, and its two images which also join $\Lambda_-$ to $\Lambda_+$. These join $\overline{100}\in \Lambda_-$ to $\overline{010}\in \Lambda_+$, and $\overline{001}\in \Lambda_-$ to $\overline{001}\in \Lambda_+$ respectively. Arcs $\gamma_{(3n+1)/3}$ for values of $n$ other than $0$, and their images, join the same pairs of points in $\Lambda_-$ and $\Lambda_+$, but wind a different number of times around the cylinder $D_{\Gamma_1}/\Gamma_1$. \medskip For general rational $p/q$ we have the following: \medskip {\bf Algorithm} {\it Each point in $\Lambda_-$ represented by a Sturmian $p/q$ word $\overline{u_1\ldots u_q}$ is joined (by $\gamma_{p/q}$ or one of its images) to the point in $\Lambda_+$ represented by the Sturmian $p/q$ word $\overline{u_{q-1}u_{q-2}\ldots u_1u_q}$.} \medskip {\bf Proof.} Both $\sigma \rho$ and $\sigma\rho^{-1}$ map the fixed point $P$ of $\rho$ to $\sigma P$. It follows that $f$ maps the pair of geodesics landing on $\Lambda_-$ either side of $\bar 1$ to the pair of geodesics landing on $\Lambda_+$ either side of $\bar 1$ (figure 6). The pair of landing points either side of $\bar 1$ are represented by the maximum and minimum Sturmian $p/q$ words, $M_{p/q}$ and $m_{p/q}$ respectively, so the arcs landing at these points of $\Lambda_-$ have their opposite ends at the points of $\Lambda_+$ represented by $s(m_{p/q})$ and $s(M_{p/q})$ respectively, where $s$ denotes left shift (i.e. `forget the first digit'). Since it is easily proved from the {\it staircase algorithm} for Sturmian words \cite{BS} that the minimum word $m_{p/q}=\overline{v_q\ldots v_1}$ is the reverse of the maximum word $M_{p/q}=\overline{v_1 \ldots v_q}$, the result follows. Indeed we may regard the $q$ arcs joining $\Lambda_-$ to $\Lambda_+$ as indexed by a marked digit in a bi-infinite Sturmian word, and the action of $f$ and $f^{-1}$ on these arcs as moving the marker left and right. $\square$ \bigskip {\bf Remarks.} \medskip 1. Which two of the three cone points on ${\Sigma}$ of cone angle $\pi$ are the end points of the arc $\delta_{p/q}$ is determined by the reflection symmetries of the bi-infinite periodic Sturmian word of rotation number $p/q$ mod $1$. Each such word has reflection symmetries of exactly two of four possible types: reflection at a $0$, or at a $1$, or between two adjacent $0$'s or $1$'s. Which two types occur depends on whether (after reduction of $p/q$ mod $1$) $p$ is even, $q$ is even, or $p$ and $q$ are both odd. For example the bi-infinite word generated by $\overline{00101}$, a case where $p$ is even, has reflection points between the first two $0$'s and at the third $0$. The stabiliser of any lift of $\delta_{p/q}$ to $\Delta$ is an infinite dihedral group, generated by a pair of involutions fixing adjacent lifts of cone points on the arc, and indeed isomorphic to the group of symmetries of the bi-infinite periodic Sturmian word. \medskip 2. The same construction of geodesic arcs crossing the central hexagon can be followed through for {\it irrational} slope $\nu$ in place of $p/q$. One then obtains a lamination on $D_{\Gamma_1}/{\Gamma_1}$, with singular leaves passing through the fixed point of $\rho$ and its translates. In this case the leaves crossing the hexagon join a Cantor set in $\partial \Lambda_-$, the unique closed invariant Sturmian set of rotation number $\nu$ mod $1$, to the analogous Cantor set in $\partial \Lambda_+$. The algorithm above also applies in this case to tell us which points are joined to which; we omit details here. \bigskip It remains to describe the grand orbit of $\gamma_{p/q}$ under the correspondence $f$. \medskip We start with the special case $p/q=0$. The arc $\gamma_0$ is the lower boundary component of the region $D_{\Gamma_1}$ in figure 4. Under $f$ this component maps to itself and to the boundary component of $D_{\Gamma_1}$ which passes through the point $\sigma(T)$. The grand orbit of $\gamma_0$ under $f$ is the union of all the boundary components of $D_{\Gamma_1}$, and quotienting by $f$, or equivalently by $\Gamma_1$, folds all these components (except the original one) into ``spikes''. \medskip We now turn to general $p/q$. From the explicit construction of matings in Section 2.2 it follows that the branch of $f$ mapping $\Lambda_-$ to $\Lambda_+$ is defined as follows: given a word $W$ in $0$'s and $1$'s representing a point in $\partial\Lambda_-$ the $f$-image in $\partial\Lambda_+$ of that point is represented by the word $\phi(W)$ obtained by changing the parity of the first digit of $W$. It is now a straightforward computation that when $q$ is even the set of $q$ arcs joining $\Lambda_-$ to $\Lambda_+$ is mapped two to one by this branch to a set of $q/2$ ``concentric'' arcs connecting pairwise the $q$ points of $\Lambda_+$ obtained by applying the operation $\phi$ to the Sturmian $p/q$ orbit (i.e. the points of the circle {\it opposite} to points of the Sturmian orbit). When $q$ is odd, the set of $q$ arcs joining $\Lambda_-$ to $\Lambda_+$ is mapped by this branch of $f$ to a set of $(q-1)/2$ concentric arcs together with an innermost spike (figure 7) which lands on $\Lambda_+$ at a single point, the point opposite to the middle point of the Sturmian $p/q$ orbit. This spike arises from the fact that for $q$ odd the geodesic $\gamma_{p/q}$ passes through the fixed point of the involution $\tau$. Hence its image under the branch of $f$ we are considering passes through the fixed point of an involution in the group $\Gamma_1$. This fixed point is on the boundary of $D_{\Gamma_1}$ (indeed in figure 4 it is the point $\sigma(T)$), and becomes the end point of a spike in the quotient $D_{\Gamma_1}/\Gamma_1\cong \Omega(f)$. \begin{figure} \begin{center} \input{pinchfig7.pstex_t} \caption{The Sturmian orbits of rotation number $2/5$ on $\Lambda_-$ and $\Lambda_+$, the five arcs joining them, and the first images of these under the correspondence and its inverse (subsequent images are not shown).} \label{images_of_arcs} \end{center} \end{figure} \medskip Applying $f$ again arbitrarily may times to our ``concentric'' set of $q/2$ arcs (or $(q-1)/2$ arcs plus a spike, if $q$ is odd), we obtain smaller and smaller copies around $\partial\Lambda_+$, and applying $\sigma$ to these copies we obtain similar copies around $\partial\Lambda_-$, together making up the grand orbit under $f$ of our original set of $q$ arcs. \subsection{Matings between $q_0$ and circle-packing representations of $C_2*C_3$} We can now define precisely what we mean by the {\it mating} between $q_0$ and $r_{p/2q}$ referred to in the statement of Theorem \ref{rat}. After the arcs which make up the grand orbit of $\gamma_{p/q}$ have been pinched, the intersection $\Lambda_+\cap \Lambda_-$ is no longer empty, but consists of the $p/q$ Sturmian orbit of the correspondence on $\partial\Lambda_+$, identified with the same orbit (in the opposite direction) on $\partial\Lambda_-$. The set $\Omega$ for the pinched correspondence has $q$ components whose boundaries meet this orbit. These form what we call the {\it principal cycle} of components of $\Omega$. Together with $\Lambda_-\cap\Lambda_+$ itself, they separate the Riemann sphere into two parts, one containing $\Lambda_-\setminus(\Lambda_+\cap\Lambda_-)$ and the other containing $\Lambda_+\setminus (\Lambda_+\cap\Lambda_-)$. The stabilizer (under the iterated pinched correspondence) of each of the components of the principal cycle is a group, since these components do not contain ``fold'' points. Moreover it is not hard to see that this group is isomorphic to $C_2*C_3$. \medskip {\bf Definition.} {\it A holomorphic correspondence is said to be a mating between $r_{p/2q}$ and $q_0$ if it is topologically conjugate to a correspondence obtained by pinching to a point each component of the grand orbit of $\gamma_{p/q}$ for a mating between $r_*$ and $q_0$, and if moreover the action of the stabiliser of each component of the principal cycle of the correspondence is conformally conjugate to the action of $PSL_2(\mbox{$\mathbb Z$})$ on the upper half-plane.} \medskip In a mating between $q_0$ and $r_{p/2q}$, the sets $\Lambda_+$ and $\Lambda_-$ are no longer copies of $K(q_0)$ (the unit disc) but are now each homeomorphic to a quotient $K(q_0)_{p/q}$ of $K(q_0)$ by an equivalence relation $\sim_{p/q}$ on $\partial K(q_0)$ (the unit circle) which may be described as follows. Let $\omega'_{p/q}$ denote the points of the circle opposite to points of the Sturmian $p/q$ orbit $\omega_{p/q}$, so $\omega_{p/q}$ and $\omega'_{p/q}$ are contained in disjoint intervals. To define the relation $\sim_{p/q}$ we identify the `outermost' pair of points of $\omega'_{p/q}$, and similarly we identify the next pair of points from the outside, and so on, folding the points of $\omega'_{p/q}$ together in pairs. Similarly we identify in pairs the corresponding inverse images of points of $\omega'_{p/q}$ under the doubling map, and repeat so that the relation $\sim_{p/q}$ becomes invariant under this inverse. \bigskip {\bf Remarks.} \medskip 1. The justification for describing the construction in the definition as ``a mating between $q_0$ and $r_{p/2q}$'' is two-fold. Firstly, both the construction and $r_{p/2q}$ are obtained by pinching the same simple closed curve $\delta_{p/q}$ on the same orbifold $\Sigma$, and secondly the definition agrees with our earlier definition for a mating between $q_0$ and the modular group. However when $p/q\notin {\mathbb Z}$ the most direct relationship we know of between $\Omega(r_{p/2q})$ and $\Omega(f)$ for the correspondence pinched along $\gamma_{p/q}$ is that given by pinching $\delta_{p/q}$ in the Fuchsian picture of $\Omega(r_*)$, described in Remark 2 following Corollary \ref{iota} (in Section 2.4). \medskip 2. Corollary \ref{iota} implies that a mating between $q_0$ and $r_{p/2q}$ is isomorphic to a mating between $q_0$ and $r_{(p+q)/2q}$. For example a mating between $q_0$ and $r_{1/2}$ is isomorphic to one between $q_0$ and the modular group. This example is easily understood directly, since $r_{1/2}$ is the faithful discrete representation of $C_2*C_3$ for which the limit set is a single round circle, like $PSL_2(\mbox{$\mathbb Z$})$, but for which the generator $\sigma$ of $C_2$ acts by interchanging the two components of the complement. We remark that $r_{p/2q}$ and $r_{(p+q)/2q}$ always have the same limit set, since the second representation is obtained from the first by composing with an (outer) automorphism of $C_2*C_3$. \subsection{Invariant collar neighbourhoods of arcs} For the proofs of Theorem \ref{simple} and Theorem \ref{rat} we shall need well-behaved neighbourhoods of our arcs on which to support the pinching deformations. We define an {\it invariant collar neighbourhood} of an arc $A$ joining $\Lambda_-$ to $\Lambda_+$ to be a closed set ${\mathcal N}(A)$ containing $A$, bounded by a pair of arcs joining the end points of $A$, such that under the action of $f$ the set ${\mathcal N}(A)$ has stabiliser isomorphic to the infinite dihedral group, and ${\mathcal N}(A)$ is {\it precisely invariant} under the action of this stabiliser. (Strictly speaking, ${\mathcal N}(A)$ is not a topological neighbourhood of $A$, since the end points of $A$ are on the boundary of ${\mathcal N}(A)$.) \begin{lemma}\label{collars_exist} The arc $\gamma_{p/q}$ has an invariant collar neighbourhood. \end{lemma} {\bf Proof.} A collar neighbourhood of each of the $q$ arcs which join $\Lambda_-$ to $\Lambda_+$ is obtained by lifting any collar neighbourhood of the $p/q$ geodesic $\delta_{p/q}$ on the orbifold $\Sigma$. It is immediate from the action of $\sigma\rho$ and $\sigma\rho^{-1}$ on the lift of such a neighbourhood that its stabiliser under the action of $f$ is an infinite dihedral group, generated by the appropriate branch of $f^q$ and by $\sigma$ (which is a branch of $f^{-1}ff^{-1}$) composed with a branch of whichever $f^r$ maps the $\sigma$ image of the arc back to the arc. This lifted collar neighbourhood is precisely invariant under the action of the stabiliser. $\square$ \medskip The small copies of the $q$ arcs have collar neighbourhoods that are the images of the original collar neighbourhoods under appropriate branches of forward or backward iterates of $f$. These images are each either a bijective copy, or (in the case of a ``spike'') a quotient by an involution, of one of the original collar neighbourhoods. In the case of the arc $\gamma_0$, joining the fixed points of the doubling map on $\partial \Lambda_-$ and $\partial \Lambda_+$, {\it all} the images are such quotients. \subsection{A pinching deformation} Let us consider a correspondence $p$ which represents the mating of a quadratic polynomial $q$ with a faithful and discrete representation of $C_2 * C_3$ with connected ordinary set, and let $f:\Lambda_-\to\Lambda_-$ be the $2:1$-branch of $p$. We fix the curve of rotation number $p/q$ and consider its lifts ${\cal R}$ (for red) to $\overline{\mbox{$\mathbb C$}}$. Thus $\gamma=\gamma_{p/q}$ is one of the connected components of ${\cal R}$ which joins $\Lambda_-$ to $\Lambda_+$. Let us denote its collar neighbourhood defined above by ${\cal N}(\gamma)$. Then $\mbox{Stab}\,_p({\cal N}(A))$ is isomorphic to the infinite dihedral group. Let $B_-$ and $B_+$ be both components of ${\cal N}(A)\setminus \gamma$. \medskip We will first define an appropriate quasiconformal deformation on a model strip and then implement it on the dynamical plane \cite{HT}. \medskip Our model space will be a closed horizontal strip on the upper half-plane. Choose a collection of numbers $0< L_y < L_r$ (the indices $y,r$ are colours yellow and red respectively), and then an increasing $C^1$-function $\tau:[0,1[\to [L_r,+\infty[$. Let $M\subset \mbox{$\mathbb R$}^2$ be the closed subset bounded by $$([0,1]\times\{0\})\cup (\{0\}\times [0,L_r])\cup (\{1\}\times [0,+\infty[ ) \cup ( \{(t,\tau(t)),t\in [0,1[\})\ .$$ Choose $v_t(y)$ so that $v_t(y)=y$ for $0\le y\le L_y$ and that $(t,y)\mapsto (t,v_t(y))$ is a $C^1$-diffeomorphism from $[0,1]\times [0,L_r]\smallsetminus \{(1,L_r) \}\to M$. \begin{figure} \begin{center} \input{deform.pstex_t} \caption{The diffeomorphism $(t,y)\mapsto (t, v_t(y))$.} \label{deform} \end{center} \end{figure} \medskip We also make the following technical assumption: for any $L'<L_r$, there is $t(L')\in ]0,1[$ with $t(L')\to 1$ as ${L'\to L_r}$, such that for any $(s,y) \in\, ]t(L'),1] \times [0,L']$, we have $v_s(y)=v_{t(L')}(y)$. Now on the straight strip $\{0\le x \le L_r\}$, and for every $t\in [0,1]$ , set $$\widetilde{P}_t(x+iy)=x+ i\cdot v_t(y)\ .$$ This map satisfies the following properties\,: \begin{enumerate} \item It commutes with the translation by $1$ (and by any other real number). \item It is the identity on the sub-strip $\{0\le y\le L_y\}$. \item The coefficient of the Beltrami form $$\left.\frac{\partial \widetilde{P}_t /\partial \bar{z}}{\partial \widetilde{P}_t /\partial z}\right|_{x+iy}= \frac{1-\frac{\partial}{\partial y}v_t(y)}{1+\frac{\partial}{\partial y}v_t(y)}$$ is continuous on $(t,x+iy)\in [0,1]\times \{0\le y\le L_r\}$, whose norm is locally uniformly bounded from $1$ if $(t,y)\ne (1,L_r)$ and tends to $1$ as $(t,y)\to (1,L_r)$. \end{enumerate} Define conformal maps $\psi_\pm :B_\pm\to \mbox{$\mathbb R$}\times (0,L_r)$ which map $\gamma$ to $\mbox{$\mathbb R$}\times\{L_r\}$. For $t\in [0,1[$, set $\sigma'_t=(\widetilde{P}_t\circ \psi_\pm)^*(\sigma_0)$ to be the pull-back of the standard complex structure on $B_\pm$. Since the action is properly discontinuous on $\mbox{$\Omega$} (f)$, we may spread $\sigma_t'$ to the whole orbit of ${\cal N}(\gamma)$ under the correspondence $p$. We let $\sigma_t$ be the extension of this almost complex structure to the whole Riemann sphere by setting $\sigma_t=\sigma_0$ on the complement. It is a $p$-invariant complex structure. We let ${\cal Y}$ (for yellow) be the set of points $z$ such that $\sigma_t(z)$ is not the standard conformal structure for some $t$. \medskip The family of $p$-invariant complex structures $(\sigma_t)_{t\in [0,1)}$ defines a pinching deformation supported on ${\cal R}$. We let $h_t$ be the quasiconformal map given by the Measurable Riemann Mapping Theorem applied to $\sigma_t$ normalised so that $h_t$ fixes both critical points of $f|_{\Lambda_-}$ and $f^{-1}|_{\Lambda_+}$ and the point at infinity as well. The correspondence $p_t$ defined by $p_t(z,w)=p(h_t^{-1}(z),h_t^{-1}(w))$ is holomorphic by construction, and the family of pairs $(p_t,h_t)_{t\in [0,1)}$ defines a marked pinching deformation. \section{Convergence of the pinching deformation} The proofs of both Theorem \ref{simple} and Theorem \ref{rat} follow essentially the same lines. We must prove that the pinching deformation defined in the previous section converges uniformly in each case, and we must prove that in each case the limit correspondence has as stabiliser of each of the components of the principal cycle of $\Omega$ a group conformally equivalent to $PSL_2(\mbox{$\mathbb Z$})$. The strategy for proving uniform convergence is inspired by \cite{H2,HT} where analogous statements are proved for rational maps and where detailed proofs can be found. \medskip We proceed to prove both theorems simultaneously as far as possible. We refer to \cite{H2} and \cite{HT} when we can, instead of repeating the detailed arguments presented in these papers. The parts of the proofs which differ for the two theorems are postponed to $4.1$ and $4.2$. In particular we delay the proof of the key Lemma \ref{nbhd} (stated below). The first step in the proof of the theorems is to prove that the path of quasiconformal homeomorphisms $(h_t)$ is equicontinuous. We will apply the following criterion the proof of which is elementary (cf. Lemma 2.5 in \cite{HT}). \begin{lemma}{\em\bf (Equicontinuity criterion at a point)} Let ${\cal H}=\{h:\mbox{$\mathbb D$}\to \mbox{$\mathbb C$}\}$ be a family of continuous injective maps such that $\cup_{h \in{\cal H}}h(\mbox{$\mathbb D$})$ avoids at least 2 points in $\mbox{$\mathbb C$}$. Let $(U_{n})_{n\ge 0}$ be a nested sequence of disc-like neighbourhoods of the origin in the unit disc $\mbox{$\mathbb D$}$ such that $A_{n}=\mbox{$\mathbb D$}\smallsetminus \overline{U_{n}}$ is an annulus. If there exists a sequence $\eta_n \nearrow +\infty$ such that $$\forall\,h\in{\cal H},\ \forall\,n\ge 0,\ \mbox{\em mod\,}h(A_{n})\,\ge\, \eta_n,$$ then ${\cal H}$ is equicontinuous at the origin. \end{lemma} This means that we need to get infinitely many annuli with controlled moduli. The assumption on the fixed point $\beta$ will give us information on the support of the deformation\,: this will enable us to prove the following lemma in the respective cases. \begin{lemma}\label{nbhd}{\em\bf (One good annulus around each Julia point)} Fix $r>0$. \begin{itemize}\item[(i)] For any $x\in \partial\Lambda_-\cup\partial\Lambda_+\smallsetminus{\cal R}$, there are two open neighbourhoods $N'(x)$ and $N(x)$ of $x$ in $ D(x,\frac{r}{4})$ and $m>0$ such that ${\rm mod}\, h_t(N(x)\smallsetminus \overline{N'(x)})\ge m$ for all $t$. \item[(ii)] For any $x=\beta_\gamma \in {\cal R}\cap ( \partial\Lambda_-\cup\partial\Lambda_+)$, with $\gamma$ an ${\cal R}$-component, there is a sequence $(t_n)$ in $[0,1)$ tending to $1$, a nested sequence of annuli $(A_n)_n$ surrounding $\gamma$, and a constant $m > 0$ such that ${\rm mod}\, h_t(A_n)\ge m/n$ for $t\ge t_n$. \end{itemize}\end{lemma} Then the weak hyperbolicity condition is used to spread these annuli at every point and at every scale and therefore to imply the equicontinuity of $(h_t)$ (cf. the proof of the Proposition 2.3 in \cite{HT} or \S 3 in \cite{H2}). The estimates of the conformal moduli also enable us to analyse the structure of the fibres of any limit map and to conclude that its fibres are exactly the closures of the connected components of ${\cal R}$. \medskip Any limit $h_1$ satisfies the conclusion of the theorem and we may also extract a convergent sequence $(p_{t_n})$ of the correspondences to a correspondence $p_1$ (cf. Appendix A in \cite{HT}). \medskip Since the fibre structure is well understood, it follows that if there are other limits $(\widehat{h},\widehat{p})$, then $\widehat{h}\circ h_1^{-1}$ defines a conjugacy which is conformal off $h_1(\partial\Lambda_-\cup\partial\Lambda_+)$ (cf. Lemma A.2 in \cite{HT}). \medskip Now it can be shown as in \cite{HT} that all the limit correspondences satisfy the ``weak hyperbolicity'' condition on the image of $\partial\Lambda_-\cup\partial\Lambda_+$. Since $\partial\Lambda_-\cup\partial\Lambda_+$ has no interior, a standard argument of Sullivan implies that the Lebesgue measure of $h_1(\partial\Lambda_-\cup\partial\Lambda_+)$ is zero (cf. Theorem 4.1 \cite{H1}). Furthermore, the weak hyperbolicity condition on $p_1$ implies that the following rigidity statement holds. \begin{prop}\label{conjugacy-conformal-off-limits} Let $p_0$ and $p_1$ be two correspondences which are matings of weakly hyperbolic polynomials with discrete representations of $C_2*C_3$. If $p_0$ and $p_1$ are conjugate by a topological homeomorphism which is conformal off the limit sets, then the conjugacy is a M\"obius transformation.\end{prop} The proof of this proposition follows the same lines as Proposition 6.3 and Theorem 0.2 in \cite{H1}. $\square$ \medskip Thus $\widehat{h}\circ h_1^{-1}$ is a M\"obius transformation, whence the uniqueness of the limits $(p_t,h_t)$ as $t$ tends to $1$. \medskip To complete the proofs of Theorem \ref{simple} and Theorem \ref{rat} it now remains only to prove Lemma \ref{nbhd} in both cases, and to prove that in each case the limit of the family of pinching deformations corresponds to the mating we are looking for. \subsection{The simple case (winding number zero)} We shall make use of the statements proved in \cite{HT} for simple pinchings of rational maps, so we have to show how to get to that setting. \medskip Using McMullen's gluing lemma (Proposition 5.5 in \cite{mc1}), we may construct a rational map $R$ of degree $2$ which induces a partition of the sphere $\overline{\mbox{$\mathbb C$}}= K\sqcup {\cal F}$ where $K$ is the filled-in Julia set of a quadratic-like map induced by a restriction of $R$ hybrid-equivalent to $q$, and ${\cal F}$ is the basin of attraction of a fixed point at infinity of multiplier $1/2$. For the domains of the quadratic-like map, we first choose a linearising disc $D$ for the point at infinity which contains the critical value, and set $V=\overline{\mbox{$\mathbb C$}}\setminus \overline{D}$. If $V'=R^{-1}(V)$, then $R:V'\to V$ is quadratic-like. Furthermore, we may find a forward-invariant Jordan arc $\kappa$ in ${\cal F}$ joining the point at infinity with the corresponding $\beta$-fixed point which only cuts $\partial V$ once, and then transversally. Let $\widehat{{\cal R}}$ be the grand orbit of $\kappa$ for $R$. It follows that $(\widehat{{\cal R}}\setminus\kappa)\cap \partial V=\emptyset$. \begin{prop}\label{comp} There is a quasiconformal $\Phi:\overline{\mbox{$\mathbb C$}}\to\overline{\mbox{$\mathbb C$}}$ such that \begin{itemize} \item $\Phi(\Lambda_-)=K$ and $\Phi({\cal R})=\widehat{{\cal R}}$, \item $\Phi\circ f= R\circ\Phi$ in a neighbourhood of $\Lambda_-$, \item $\overline{\partial}\Phi =0$ a.e. on $\Lambda_-$. \end{itemize}\end{prop} {\bf Proof.} We already know that there is a quasiconformal map $\phi:\overline{\mbox{$\mathbb C$}}\to\overline{\mbox{$\mathbb C$}}$ which fulfills the conclusions of the Proposition except for the condition on the curves. We let $U'\subset \subset U$ be simply connected domains such that the extension $f:U'\to U$ of the branch of the correspondence $f:\Lambda_-\to\Lambda_-$ is a quadratic-like map hybrid-equivalent to $q$. It follows from the construction of $f$ that we may assume that $U$ is a fundamental domain for the involution $J$. Furthermore, we may also assume that $\phi(U)=V$. \medskip We let $\phi_0:U\to V$ be a quasiconformal homeomorphism isotopic to $\phi$ rel. $\Lambda_-$ through an isotopy which maps $\partial U$ to $\partial V$ throughout, and such that $$\phi_0(\gamma_0\cap (\overline{U}\setminus U'))= \kappa\cap (\overline{V}\setminus V')\mbox{ and } R\circ \phi_0|_{\partial U'}=\phi_0\circ f|_{\partial U'}.$$ This is possible since both sets $U\setminus \Lambda_-$ and $V\setminus K$ are annuli and since the action of the maps $f$ and $R$ are 2:1 coverings. Define $(\phi_n)$ inductively so that $\phi_{n+1}\circ f=R\circ\phi_n$ so that $\phi_n|_{\Lambda_-}=\phi|_{\Lambda_-}$ and $\phi_n|_{\overline{U}\setminus U'}=\phi_0|_{\overline{U}\setminus U'}$. This sequence is a normal family quasiconformal mappings which admits at least one limit $\Phi:U\to V$. This map satisfies the conclusion of the proposition. $\square$ \medskip We now provide a proof of Lemma \ref{nbhd} under the assumptions of Theorem \ref{simple}. \medskip {\bf Proof of Lemma \ref{nbhd}.} We first assume that $q$ is not conjugate to $z\mapsto z^2 +1/4$. Then by Lemma 2.7 in \cite{HT} we have the result we seek but for $\widehat{{\cal R}}$ and the rational map $R$ in place of ${\cal R}$ and the correspondence. By Proposition \ref{comp} this is all we need, except for the case of the only ${\cal R}$-component, $\gamma_0$, which is not contained in the neighbourhood $U$ of $\Lambda_-$. But $\gamma_0$ is a double cover of any other component $\gamma$ of ${\cal R}$ by a branch of the correspondence, and $\gamma_0$ has a neighbourhood which is a double cover of a disc neighbourhood of $\gamma$, by the same branch. \medskip We now deal with $q(z)=z^2+1/4$. Let us denote by $p$ the mating of $ q$ with $C_2*C_3$ and let us define $q_0(z)=z^2$, $p_0$ and $R_0$ the corresponding mating and rational map. We let $(p_t,\widehat{h}_t)$ be the simple pinching of $p_0$ considered above, and $\Phi_0:\Lambda_-(p_0)\to \overline{\mbox{$\mathbb D$}}$ be given by Proposition \ref{comp}. It follows from Corollary 3.10 in \cite{HT} that there is a $\mu$-homeomorphism , in the sense of David, $\phi:\mbox{$\mathbb C$}\to\mbox{$\mathbb C$}$, conjugating $p_0|_{\mbox{$\Omega$} (p_0)}$ conformally to $p|_{\mbox{$\Omega$} (p)}$. Furthermore, a constant $K_0\ge 1$ exists such that the set of points $z\in \mbox{$\mathbb C$}$ for which the dilatation ratio $K_\phi(z)$ is at least $K_0$ is contained in the disjoint union of the orbit of an invariant sector $S\subset int(\Lambda(p_0))$ with vertex $\beta$ (see Lemma 2.1 \cite{H} for details). \medskip We claim that the image under $\phi$ of the controlled annuli for $p_0$ have also controlled moduli. For points outside the red set, this is because the set where $K_\phi$ is large is contained in the union of sectors so that the Key lemma in \cite{HT}, which implies the bounds on the moduli, also holds for these domains. \medskip For points in the red set, we must be more precise and use intermediate results which are established for the proof of Lemma 2.7 in \cite{HT}. We refer to \S 2.5 in \cite{HT} for the details. We let $Y$ be the connected component of ${\cal Y}(p_0)$ which contains $\gamma_0$. In the proof of the equicontinuity at those points, it is shown that there is a sequence $\psi_n: A_n \to (-C-(n+1),C+(n+1))^2\setminus[-C-n,C+n]^2$ of homeomorphisms, where $C$ is a fixed positive real number, such that, for $t\ge t_n$, $\psi_n\circ \widehat{h}_t^{-1}$ is uniformly quasiconformal off ${\cal Y}\setminus Y$. Moreover, $\psi_n$ maps $\Phi_0(S)\cap A_n$ onto a rectangle $Q_n=[-C-(n+1),-C-n]\times [C_1,C_2]$ for fixed constants $C_1$ and $C_2$. \medskip The bound on the moduli for the cauliflower map $z\mapsto z^2+1/4$ comes from a length-area argument provided by metrics $(\rho_n^t)$ defined as follows. Let $t\ge t_n$; on $\widehat{h}_t({\cal Y}(p_0)\setminus Y)$, we let $\rho_n^t=0$ and on its complement we define $$\rho_n^t=\frac{1}{|\partial_z \widehat{h}_t\circ\psi_n^{-1}|-|\partial_ {\bar z}\widehat{h}_t\circ\psi_n^{-1}|} \circ (\psi_n\circ \widehat{h}_t^{-1})\,.$$ This kind of metric is used to prove the quasi-invariance of moduli of annuli for quasiconformal maps. This metric yields the bound $\mod \widehat{h}_t(A_n)\ge m/n$ where $m>0$ is independent of $n$. \medskip Similarly, we let $\widehat{\rho}^t_n=0$ for points in $h_t\circ\phi({\cal Y}(p_0)\setminus Y)$ and on the complement, we let $$\widehat{\rho}^t_n=\frac{\rho_n^t}{|\partial_z \phi_t|-|\partial_{\bar z} \phi_t|}\circ\phi_t^{-1}\,,$$ where $\phi_t= h_t\circ \phi\circ \chi_t^{-1}$. It follows from the construction of $\phi$ that $K_{\phi}\asymp n$ on $Q_n $ (see Lemma 2.1 in \cite{H}), so that the area of $h_t(\phi(Q_n))$ is at most a multiple of $n$, as the area of $h_t(\phi(A_n\setminus Q_n))$, for the metric $\widehat{\rho}^t_n$. Thus, we get $\mod h_t(\phi_0(A_n))\ge c/n$. Whence we obtain the estimates of the moduli for these points also. $\square$ \medskip The following proposition now completes the proof of Theorem \ref{simple}. \begin{prop}\label{simple-mating} Under the assumptions of Theorem \ref{simple}, the limit $p_1$ of $(p_t)$ is a mating of $q$ with $PSL_2(\mbox{$\mathbb Z$})$. \end{prop} {\bf Proof.} The limiting correspondence $p_1$ inherits a compatible involution $J$ from $p_0$, so by Proposition {\ref{compatible} (Section 1) this correspondence is conjugate to some member of the family (2), or equivalently to $J \circ Cov_0^Q$ for $Q(z)=z^3-3z$ and $J$ some (M\"obius) involution. The proof of the Proposition now follows the same steps as the proof of Theorem 7.1 in \cite{bf}, which states an analogous result for the degree $4$ Chebyshev polynomial in place of $Q$. We summarise the steps but refer the reader to \cite{bf} for technical details. The topological dynamics of $p_1$ ensure that there exist a transversal $D_Q$ for $Q$ and a fundamental domain $D_J$ for $J$ such that the complement of the union of the interiors of $D_Q$ and $D_J$ consists precisely of the fixed point $\Lambda_+ \cap \Lambda_-$. This fixed point is parabolic for $f$ and it follows from local anaysis that in a neighbourhood the boundaries of $D_Q$ and $D_J$ may be chosen to be smooth curves, tangent to one another at the fixed point. The set $D_Q\cap D_J$ is a fundamental domain for the action of $f|_\Omega$, and since $f|_\Omega$ and $f^{-1}|_\Omega$ have no critical points (only double points) we know that $f|_\Omega$ is conformally conjugate to $\{\sigma\rho,\sigma\rho^{-1}\}$ for some Fuchsian representation of $C_2*C_3$ acting on the open upper half of the complex plane. To show that this action is indeed that of $PSL(2,\mbox{$\mathbb Z$})$ it suffices to show that in the upper half-plane the images of $\partial D_Q$ and $\partial D_J$ converge to the same point on the real axis. This can be shown to follow from the fact that $\partial D_Q$ and $\partial D_J$ are smooth curves which meet tangentially (see \cite{bf}). $\square$ \subsection{Pinching arcs of non-zero rational winding number} Let $p$ be a correspondence which is a mating between $z\mapsto z^2$ and a faithful discrete representation of $C_2*C_3$ in $PSL_2(\mbox{$\mathbb C$})$ with connected ordinary set. In this section we prove Lemma \ref{nbhd} for curves in $\mbox{$\Omega$} (p)$ with non-zero rational rotation number. The fact that the Julia set is a quasicircle will be crucial in the proof, which closely follows the argument in \S 3 of \cite{H2}. \medskip The first step is to straighten the limit set and the support of the pinching. Figure \ref{chi} illustrates an example. \begin{lemma} There is a quasiconformal map $\chi:\overline{\mbox{$\mathbb C$}}\to\overline{\mbox{$\mathbb C$}}$ such that $\chi(\partial \Lambda_-)=\mbox{$\mathbb S$}^1$, which satisfies the following properties\,: \begin{itemize} \item $\chi$ is conformal on the interior of $\Lambda_-$\,; \item $\chi$ conjugates $f$ to $z\mapsto z^2$ in a neighbourhood of the interior of $\Lambda_-$\,; \item components of ${\cal Y}$ which are attached at two points $x$ and $y$ to $\Lambda_-$ are mapped into rectangles in (log)-polar coordinates with base $[\chi(x),\chi(y)]$; \item components $Y$ of ${\cal Y}$ which are attached at a single point $x$ to $\Lambda_-$ are mapped into sectors based at $\chi(x)$; \end{itemize}\end{lemma} {\bf Proof.} The restriction of $\chi$ to $\Lambda_-$ is given by the B\"ottcher coordinates of $f$. The extension of $\chi$ to the outside makes use of a pull-back argument (see pp.\,14-15 in \cite{H2}). $\square$ \medskip \begin{figure} \begin{center} \input{chi.pstex_t} \caption{Image under $\chi$ of the collars of the first two generations of the orbit of $\gamma_{p/q}$, in the case $p/q=2/5$ (cf. fig. 7).} \label{chi} \end{center} \end{figure} \medskip The next step of the proof is to control the moduli of many annuli. We place ourselves in the coordinates given by $\chi$. As in \cite{H2}, we may define annuli bounded by rectangles in the log-polar coordinates which avoid the image of ${\cal Y}$ under $\chi$. \medskip As in the case of simple pinchings, there is no problem with the curves which link both components of $\Lambda$, because they cover other components which do not. This enables us to prove Lemma \ref{nbhd} (cf. Proposition 3.3 and 3.4 in \cite{H2}). \medskip Finally, the following proposition completes the proof of Theorem \ref{rat}. \begin{prop}\label{rational-mating} Under the assumptions of Theorem \ref{rat}, the limit $p_1$ of $(p_t)$ is a mating of $z\mapsto z^2$ with the circle-packing representation $r_{p/2q}$ of $C_2*C_3$. \end{prop} {\bf Proof.} As in the proof of Proposition \ref{simple-mating} the limiting correspondence $p_1$ is necessarily conjugate to some member of the family (2), or equivalently to $J \circ Cov_0^Q$ for $Q(z)=z^3-3z$ and $J$ some (M\"obius) involution. Once again we can now follow the same steps as in the proof of Theorem 7.1 in \cite{bf}. Transversals $D_Q$ and $D_J$ can be chosen this time such that the complement of the union of their interiors consists precisely of the period $q$ parabolic orbit $\Lambda_+ \cap \Lambda_-$, and such that in a neighbourhood of any point of this orbit the boundaries of these transversals are smooth curves, tangent to one another at the orbit point. From the fact that $\Omega$ is now a countable union of topological discs and our knowledge of the topological dynamics of $f$ (using convergence of the pinching deformation) we know that $f|_\Omega$ and $f^{-1}|_\Omega$ have no critical points (only double points) and that for any component of $\Omega$ which meets the period $q$ orbit $\Lambda_+\cap\Lambda_-$ the iterated branches of $f$ which stabilise the component are conformally conjugate to the elements of the group generated by $\{\sigma\rho,\sigma\rho^{-1}\}$ for some Fuchsian representation of $C_2*C_3$ acting on the open upper half of the complex plane. As in the proof of Proposition \ref{simple-mating} the properties of the boundaries of $D_Q$ and $D_J$ again ensure that this representation is indeed conformally conjugate to $PSL_2(\mbox{$\mathbb Z$})$. $\square$
{ "timestamp": "2005-03-30T15:10:07", "yymm": "0503", "arxiv_id": "math/0503706", "language": "en", "url": "https://arxiv.org/abs/math/0503706" }
\section*{References} \noindent\REF{[1]}S. Sachdev, {\it Quantum Phase Transitions} (Cambridge Univ. Press, Cambridge, 1999). \REF{[2]} J. A. Hertz, Phys. Rev. B {\bf 14}, 1165 (1976). \REF{[3]} F. lachello and A. Arima, {\it The Interacting Boson Model} (Cambridge University Press, Cambridge, 1987). \REF{[4]} A. Borh and B. R. Mottelson, {\it Nuclear Structure} Vol. I (Benjamin, New York, 1969); Vol. II (Benjamin, New York, 1975). \REF{[5]} F. lachello, AIP Conf. Proc. {\bf 726}, 111 (2004). \REF{[6]} A. E. L. Dieperink, O. Scholten, and F. lachello, Phys. Rev. Lett. {\bf 44}, 1747 (1980). \REF{[7]} D. H. Feng, R. Gilmore, and S. R. Deans, Phys. Rev. {\bf C23}, 1254 (1981). \REF{[8]} O. S. Van Roosmalen, {\it Algebraic Description of Nuclear and Molecular Rotation-Vibration Spectra}, Ph.D. Thesis, University of Groningen, The Netherlands, 1982. \REF{[9]} R. Gilmore and D. H. Feng, Nucl. Phys. {\bf A301}, 189 (1978); R. Gilmore, J. Math. Phys. {\bf 20}, 89 (1979). \REF{[10]} R. F. Casten, in {\it Interacting Bose-Fermi System}, ed. F. Iachello (Plenum, 1981). \REF{[11]} F. Iachello, Phys. Rev. Lett. {\bf 85}, 3580 (2000). \REF{[12]} R. M. Clark, M. Cromaz, M. A. Deleplanque, M. Descovich, R. M. Diamond, P. Fallon, I. Y. Lee, A. O. Macchiavelli, H. Mahmud, E. Rodriguez-Vieitez, F. S. Stephens, and D. Ward, Phys. Rev. {\bf C69}, 064322 (2004). \REF{[13]} J. M. Arias, J. Dukelsky, and J. E. Garcia-Ramos, Phys. Rev. Lett. {\bf 91}, 162502 (2003). \REF{[14]} D. J. Rowe, Phys. Rev. Lett. {\bf 93}, 122502 (2004). \REF{[15]} F. Iachello, N. V. Zamfir, Phys. Rev. Lett. {\bf 92}, 212501 (2004). \REF{[16]} P. Cejnar, S. Heinze, and J. Dobe{\v{s}}, Phys. Rev. C {\bf 71}, 011304 (2005). \REF{[17]} J. M. Arias, C. E. Alonso, A. Vitturi, J. E. Garcia-Ramos, J. Dukelsky, A. Frank, Phys. Rev. {\bf C68}, 041302 (2003). \end{document} \begin{figure} \begin{center} \epsfig{file=fig6.eps,width=3.8cm}~~~ \epsfig{file=fig7.eps,width=3.8cm} \epsfig{file=fig8.eps,width=3.8cm} \epsfig{file=fig9.eps,width=3.8cm} \end{center} {\scriptsize Fig. 2. Overlaps of the ground state wavefunction, where the full line shows the overlap $|\langle 0_g;x|0_g;x=0\rangle|$, and the dotted line shows the overlap $|\langle 0_g;x|x_g;x=1\rangle|$.} \end{figure}
{ "timestamp": "2005-03-29T15:04:57", "yymm": "0503", "arxiv_id": "nucl-th/0503071", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503071" }
\section*{Introduction} Myosin V is a motor protein involved in different forms of intracellular transport \cite{Reck-Peterson.Mercer2000,Vale2003b}. Because it was the first discovered processive motor from the myosin superfamily and due to its unique features, including a very long step size, it has drawn a lot of attention in recent years and now belongs to the best studied motor proteins. The experiments have characterized it mechanically \cite{Mehta.Cheney1999,Rock.Spudich2000,Rief.Spudich2000,Veigel.Molloy2002,Purcell.Sweeney2002}, biochemically \cite{De_La_Cruz.Sweeney1999,De_La_Cruz.Ostap2000a,De_La_Cruz.Ostap2000b,Yengo.Sweeney2002,Purcell.Sweeney2002}, optically \cite{Ali.Ishiwata2002,Forkey.Goldman2003,Yildiz.Selvin2003} and structurally \cite{Walker.Knight2000,Burgess.Trinick2002,Wang.Sellers2003,Coureux.Houdusse2003}. These studies have shown that myosin V walks along actin filaments in a hand-over-hand fashion \cite{Yildiz.Selvin2003} with an average step size of about 35 nm, roughly corresponding to the periodicity of actin filaments \cite{Mehta.Cheney1999,Rief.Spudich2000,Veigel.Molloy2002,Ali.Ishiwata2002}, a stall force of around 2 pN \cite{Rief.Spudich2000} and a run length of a few microns \cite{Rief.Spudich2000,Sakamoto.Sellers2003,Baker.Warshaw2004}. Under physiological conditions, ADP release was shown to be the time limiting step in the duty cycle \cite{De_La_Cruz.Sweeney1999,Rief.Spudich2000}. Two stages of the power stroke have been resolved: one about 20nm, possibly connected with the release of phosphate, and another one of 5nm, probably occurring upon release of ADP \cite{Veigel.Molloy2002}. Despite all this progress, the definite answer to the questions how the mechanical and the chemical cycle are coupled and how the heads communicate with each other to coordinate their activity has not yet been found. Theoretical models for processive molecular motors can follow different goals. What most models have in common is that they identify a few long-living states in the mechanochemical cycle and assume stochastic (Markovian) transitions between them. The differences between models start in the way these states are chosen. An approach that has been applied to myosin V \cite{Kolomeisky.Fisher2003}, kinesin \cite{peskin95,schief2001,Thomas.Tawada2002}, as well as to other biological mechanisms of force generation, including actin polymerization \cite{peskin93} and RNA polymerase \cite{wang-oster98}, models the motors as stochastic steppers. These models describe the whole motor as an object that can go through a certain number of conformations (typically a few) with different positions along the track. After the completion of one cycle (which is, in models for myosin V and kinesin, tightly coupled to the hydrolysis of one ATP molecule), the motor advances by one step. All steps are reversible and at loads above the stall the motor is supposed to walk backwards and thereby regenerate ATP. The approach has been particularly useful for interpreting the measured force-velocity relations and relating them to the kinetic parameters and positions of substeps \cite{schief2001,Fisher.Kolomeisky2001,Kolomeisky.Fisher2003}. A limitation of such models is that they assume coordinated activity of both heads rather than explaining it. They also assume that the motor strictly follows the regular cycle and there is no place for events like steps of variable length and dissociation from the track, although the latter can be incorporated into the models by proposing a different dissociation rate for each state in the cycle. In this Article we present a physical model for the processive motility of myosin V. The basic building block of our model is an individual head, which we model in a similar way as the models for conventional myosins do \cite{hill74}, albeit with different rate constants. The head is connected to the lever arm, which we model as an elastic rod, whose geometry we infer from electron microscopy studies \cite{Walker.Knight2000,Burgess.Trinick2002}. The two lever arms are connected through a flexible joint and this is the exclusive way of communication between them. We will derive the properties of the dimer from those of the individual head. \section*{The Model} To describe each myosin V head we use a model based on the 4-state cycle as postulated by \citet{lymn71} and used in many quantitative muscle models \cite{hill74} (Fig.~\ref{fig:2}A). We restrict ourselves to the long-living states in the cycle: detached with ADP.Pi, bound with ADP.Pi, bound with ADP, detached with ADP and bound without a nucleotide. The bound state with ATP and the free state with ATP have both been found to be very short-lived \cite{De_La_Cruz.Sweeney1999} and we therefore omit them in our description, i.e., we assume that binding of ATP to a bound head leads to immediate detachment and ATP hydrolysis. The detached state without a nucleotide is very unlikely to be occupied because of the low transition rates leading to it and we omit it from our scheme as well. One question that has not yet been definitely answered, is whether Pi release occurs before or during the power stroke, i.e., whether a head which is mechanically restrained form conducting its power stroke can release Pi or not. The 4-state model assumes a tight linkage between the Pi release and the power stroke. While the 4-state model has been successfully applied to myosin II (e.g., \citet{duke99,vilfan2003b}), recent experimental evidence suggests that the lead head can release Pi before the power stroke \cite{Rosenfeld.Sweeney2004}. We therefore also discuss an alternative 5-state model. In the 5-state model we introduce an additional state ADP$'$ in which the phosphate is already released, but the lever-arm is still in the pre-powerstroke state. The next transition, ADP release, however, is still linked to the completion of the full power-stroke. This is necessary in order to explain head coordination and also in agreement with experiments that show a strain-dependence in the ADP release rate in single-headed molecules \cite{Veigel.Molloy2002}. The extended duty cycle of a head is shown in Fig.~\ref{fig:2}B. \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure1} } \mycaption{The myosin V dimer is modeled as two heads, each connected to a lever arm which leaves the head at a certain angle $\phi$, depending on the state of the head. The two lever arms, modeled as elastic beams, are connected with a flexible joint, which is also connected to the external load.} \label{fig:1} \end{figure} \begin{figure} \figurecontents{ A)\includegraphics{Figure2a}\\ B)\includegraphics{Figure2b} } \mycaption{A) The mechanochemical cycle of each individual head. The head attaches to actin in the state with ADP and Pi bound on it, undergoes a large conformational change upon Pi release, another smaller conformational change upon ADP release, then binds ATP and enters the very weakly bound state, which dissociates quickly. B) The mechanochemical cycle in the 5-state model. In this scenario, the phosphate release and the power stroke are two separate transitions.} \label{fig:2} \end{figure} \begin{table} \mycaption{Geometric parameters of a myosin V head (see also Fig.~\ref{fig:1} for their definition).} \label{tab:I} \begin{center} \begin{tabular}{lll} \hline Lever arm length & $L$ & $26\,{\rm nm}$\\ Lever arm start & $R$ & $8\,{\rm nm}$\\ Lever arm start & $\delta_{\rm ADP.Pi}$ & $0\,{\rm nm}$\\ Lever arm start & $\delta_{\rm ADP,apo}$ & $3.5\,{\rm nm}$\\ Angle ADP.Pi & $\phi_{\rm ADP.Pi}$ & $115^{\circ}$\\ Angle ADP & $\phi_{\rm ADP}$ & $50^{\circ}$\\ Angle apo & $\phi_{\rm apo}$ & $40^{\circ}$\\ \hline \end{tabular} \end{center} \end{table} A head always binds to an actin subunit in the same relative position. In each state, the proximal end of the lever arm leaves the head in a fixed direction in space, determined by the polar angle $\phi$ towards the filament plus end and the azimuthal angle $\theta=\theta_0 i$ of the actin subunit $i$ to which the head is bound. The geometry of the molecule and the angles were inferred from images obtained with electron microscopy \cite{Walker.Knight2000,Burgess.Trinick2002}. They are summarized in Table \ref{tab:I}. In our calculations we assume a 13/6 periodicity of the actin helix (6 rotations per 13 subunits), which means $\theta_0=2\pi \times 6/13$. We assume that the lever arm has the properties of a linear, uniform and isotropic elastic rod, described with the bending modulus $EI$. Then the local curvature $\kappa$ is determined from $M=EI {\mathbf\kappa}$, where $M$ is the local bending moment (torque). The lever arms from both heads are joined together (and to the tail) with a flexible joint which allows free rotation in all directions. For a certain configuration of chemical states, binding sites of both heads and a given external force, the three-dimensional shape and the bending energy of both lever arms can be calculated numerically as described in the Appendix. Some of the calculated shapes are shown in Fig.~\ref{fig:3}. \begin{figure*}[htbp] \figurecontents{ \includegraphics{Figure3} } \mycaption{Calculated shapes and bending energies of dimers, bound $i$ subunits apart ($i=-2,2,\ldots,15$) and in different states: first in post-, second in the pre-powerstroke state (upper row), both in the post-powerstroke state (middle row) and both in the pre-powerstroke state (bottom row). Each configuration is shown in side and front view. If both heads are in the same state (bottom two rows) there is a significant cost in elastic energy needed to buckle one of the lever arms. Binding of the lead head before the trail head undergoes the power stroke is therefore unlikely. } \label{fig:3} \end{figure*} We calculate the free energy of a dimer state as \begin{equation} G=G_1 + G_2 + U_1 + U_2 + F x\;, \label{eq:1} \end{equation} where $G_1$ and $G_2$ are the intrinsic free energies of both heads (which depend on the chemical state of the head and the concentrations of nucleotides), $U_1$ and $U_2$ are the energies stored in the elastic deformation of each lever arm, and $F x$ is the work done against the external load ($x$ denotes the coordinate of the flexible joint along the filament axis with positive values towards the plus end, while positive values of $F$ denote a force pulling towards the minus end, against the direction of motion of an unloaded motor). \subsection*{Transition rates} There are two exact statements we can make about the kinetic rates of the duty cycle that follow from the principle of detailed balance. The first statement relates the forward and the backward rate of any reaction to the free energy difference between the initial and the final state. For any transition the principle of detailed balance states that \begin{equation} \label{eq:2} \frac{k_{+i}}{k_{-i}} = \frac{k^0_{+i}}{k^0_{-i}} e^{-\frac{\Delta U + F \Delta x}{k_B T}} \end{equation} where $\Delta U$ denotes the change in elastic energy of the dimer and $F\Delta x$ the work performed against the external load. The second exact statement can be derived by multiplying together the detailed balance conditions for a monomer in the absence of any external force along a closed pathway in Fig.~\ref{fig:2}. After one cycle the free energy of the bound monomeric head returns to its initial value, while the total free energy change in the system equals the amount gained from the hydrolysis of one ATP molecule. The resulting relation reads \begin{multline} \label{eq:3} \frac{k^0_{\rm +A} k^0_{\rm -Pi} k^0_{\rm -ADP} k_{\rm +ATP} [{\rm ATP}]}{k_{\rm -A} k^0_{\rm +Pi} [{\rm Pi}] k^0_{\rm +ADP} [{\rm ADP}] k^0_{\rm -ATP}} \\ =e^{\frac{\Delta G_{ATP}}{k_B T}}=e^{\frac{\Delta G^0}{k_B T}} \frac{[{\rm ATP}]}{[{\rm ADP}][{\rm Pi}]} \end{multline} and provides an important constraint on the kinetic rates of the model. In the 5-state model, we obtain an equivalent equation, \begin{multline} \label{eq:4} \frac{k^0_{\rm +A} k^0_{\rm -Pi} k^0_{\rm +PS} k^0_{\rm -ADP} k_{\rm +ATP} [{\rm ATP}]}{k_{\rm -A} k^0_{\rm +Pi} [{\rm Pi}] k^0_{\rm -PS} k^0_{\rm +ADP} [{\rm ADP}] k^0_{\rm -ATP}} \\ =e^{\frac{\Delta G^0}{k_B T}} \frac{[{\rm ATP}]}{[{\rm ADP}][{\rm Pi}]}\;. \end{multline} A similar statement also holds for the rates along the inner loop in the reaction scheme, which involves attachment, power stroke and detachment, all in the ADP state. Because we assume that the detachment rate in the pre-powerstroke and the post-powerstroke state are both the same ($k'_{\rm -A}$), the relation reads \begin{equation} \label{eq:5} \frac{ k^{0\prime\prime}_{\rm +A} k^0_{\rm +PS} } { k^{0\prime}_{\rm +A} k^0_{\rm -PS} }=1 \;. \end{equation} When it comes to the actual force dependence of transition rates we have to rely on approximations. An approach that is most widely used when modeling motor proteins, but also other conformational changes, like the gating of ion channels, involves the Arrhenius theory of reaction rates \cite{hill74}. It proposes that the protein has to reach an activation point ($x_a$) somewhere between the initial ($x_i$) and the final state ($x_f$) by thermal diffusion, but completes the reaction rapidly after that. Therefore, the force dependence of the forward rate can be modeled as \begin{equation} \label{eq:6} k_{+i}=k_{+i}^0 e^{-\frac {U(x_a)-U(x_i)}{k_B T}} \qquad k_{-i}=k_{-i}^0 e^{-\frac {U(x_a)-U(x_f)}{k_B T}} \end{equation} where $U(x)$ means the total potential (bending of both lever-arms and work done against the external load) which a head has to overcome to bring the lever arm angle into a given state. We use the variable $\epsilon$ to denote the relative position of the activation point between the initial and the final state, so that $x_a=(1-\epsilon) x_i +\epsilon x_f$. Unless otherwise noted, we will assume $\epsilon=0.5$. Not precisely identical, but useful for practical purposes is also the approximation $U(x_a)=(1-\epsilon)U(x_i)+\epsilon U(x_f)$. Therefore we get the following expression for the force-dependence of the transition rate: \begin{equation} \label{eq:7} k_{+i}=k_{+i}^0 e^{\frac{\epsilon \Delta U}{k_B T}} \end{equation} For reactions that involve the binding and unbinding of a head, Eq.~\ref{eq:2} is valid, but one expects the activation point to be much closer to the bound state. The strain-dependence of the detachment rate for heads in the ADP and ATP.Pi state has not yet been measured and we therefore neglect it, assuming that the detachment rate is force-independent, $k_{\rm -A}\equiv k_{\rm -A}^0$. The attachment rate then relates to the potential difference as \begin{equation} k_{\rm +A}=k_{\rm +A}^0 e^{-\frac{\Delta U}{k_B T}}\;. \end{equation} \subsection*{Choice of kinetic parameters} \begin{table*}[htbp] \mycaption{Kinetic parameters of the model} \label{tab:II} \figurecontents{ \begin{tabular}{lp{5.5cm}llp{6cm}} \multicolumn{2}{c}{Parameter} & \multicolumn{2}{c}{Value}& Source\\ &&4-state&5-state&\\ \hline\\ $k^0_{\rm +A}$ & actin binding with ADP.Pi & $5000\,{\rm s^{-1}}$& $5000\,{\rm s^{-1}}$&est.~from run length \\ % $k_{\rm -A}$ & actin release with ADP.Pi & $1\,{\rm s^{-1}}$& $50\,{\rm s^{-1}}$&est.~from run length \\ % $k^{0\prime}_{\rm +A}$ & actin binding with ADP & $5000\,{\rm s^{-1}}$& $5000\,{\rm s^{-1}}$& $\approx k^0_{\rm +A}$ \cite{De_La_Cruz.Sweeney1999} \\ % $k^{\prime}_{\rm -A}$ & actin release with ADP & $0.1\,{\rm s^{-1}}$ & $0.1\,{\rm s^{-1}}$ & $0.032\,{\rm s^{-1}}$ \cite{De_La_Cruz.Sweeney1999}, $1.1 \,{\rm s^{-1}}$ \cite{Baker.Warshaw2004}\\ % $k^0_{\rm -Pi}$ & Pi release & $200\,{\rm s}^{-1}$ & $200\,{\rm s}^{-1}$ & $>250\,{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999}, $110\,{\rm s}^{-1}$ \cite{Yengo.Sweeney2004}, $228\,{\rm s}^{-1}$ \cite{Rosenfeld.Sweeney2004}\\ $\epsilon_{\rm -Pi}$ & activation point & $0.3$ &--&F-v relation at high loads\\ % $k^0_{\rm +Pi}$ & Pi binding & $10^{-4}\,{\mu \rm M}^{-1}{\rm s}^{-1}$ & $10^{-2}\,{\mu \rm M}^{-1}{\rm s}^{-1}$ & guess \\ % $k^0_{\rm +PS}$ & power stroke & -- & $10^{4}\,{\rm s}^{-1}$ & guess \\ % $k^0_{\rm -PS}$ & reverse stroke & -- & $0.05\,{\rm s}^{-1}$ & $k^0_{\rm +PS}/k^0_{\rm -PS}$ from the stall force \\ % $k^0_{\rm -ADP}$ & ADP release & $20\,{\rm s}^{-1}$ & $20\,{\rm s}^{-1}$ & $k_{\rm -ADP}=13\,{\rm s}^{-1}$ for dimers \cite{Rief.Spudich2000}\\ % $k^0_{\rm +ADP}$ & ADP binding & $12\,{\mu \rm M}^{-1}{\rm s}^{-1}$ & $12\,{\mu \rm M}^{-1}{\rm s}^{-1}$ & $12.6\,{\mu \rm M}^{-1}{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999}, $14\,{\mu \rm M}^{-1}{\rm s}^{-1}$ \cite{Wang.Sellers2000}\\ % $k_{\rm +ATP}$ & ATP binding, actin release & $0.7 \,{\mu \rm M}^{-1}{\rm s}^{-1}$ & $0.7 \,{\mu \rm M}^{-1}{\rm s}^{-1}$ & $0.9 \,{\mu \rm M}^{-1}{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999,Rief.Spudich2000}, $0.6-1.5 \,{\mu \rm M}^{-1}{\rm s}^{-1}$ \cite{Veigel.Molloy2002}\\ % $k^0_{\rm -ATP}$ & actin binding with ATP release & $0.07\,{\rm s}^{-1}$ & $1.2\,{\rm s}^{-1}$ & Eq.~\ref{eq:3}, Eq.~\ref{eq:4} \\ \end{tabular} } \end{table*} Some of the transition rates in the cycle are well known from the literature. $k_{\rm -ADP}$ is the limiting rate both for running myosin V molecules and for single-headed constructs at low ATP concentrations. The measured values are $13\,{\rm s}^{-1}$ \cite{Rief.Spudich2000} for dimers and $12\,{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999}, $13$--$22\,{\rm s^{-1}}$ \cite{Trybus.Freyzon1999}, and $4.5$--$7\,{\rm s}^{-1}$ \cite{Molloy.Veigel2003} for monomers. Because the actual rate in a dimer is slowed down as compared to the monomer, we use the value $k^0_{\rm -ADP}=20\,{\rm s}^{-1}$. The reverse rate, $k_{\rm +ADP}$ can be determined from the inhibitory effect of ADP on the velocity and has been estimated as $12.6\,{\rm \mu M}^{-1}{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999}, $4.5\,{\rm \mu M}^{-1}{\rm s}^{-1}$ \cite{Rief.Spudich2000}, $14\,{\rm \mu M}^{-1}{\rm s}^{-1}$ \cite{Wang.Sellers2000}. Equally well known is the rate for ATP binding, $k_{\rm +ATP}$, which has been measured as $0.9\,{\rm \mu M}^{-1}{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999,Rief.Spudich2000}, $0.6$--$1.5\,{\rm \mu M}^{-1}{\rm s}^{-1}$ \cite{Veigel.Molloy2002}. For the Pi release rate the estimates range from $k_{\rm -Pi}>250\,{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999} to $110\,{\rm s}^{-1}$ \cite{Yengo.Sweeney2004}. We therefore use the value $k_{\rm -Pi}=200\,{\rm s}^{-1}$. There is some more discrepancy between the current values for the release rate from actin in the ADP state. While direct measurements gave $k'_{-\rm A}=0.032\,{\rm s}^{-1}$ \cite{De_La_Cruz.Sweeney1999} and $0.08\,{\rm s}^{-1}$ \cite{Yengo.Sweeney2004}, a recent estimate from the run length led to a higher value of $1.1\,{\rm s}^{-1}$ \cite{Baker.Warshaw2004}. We use an intermediate value of $k'_{-\rm A}=0.1\,{\rm s}^{-1}$. For the attachment rate in the ADP state, we set $k^{0 \prime}_{+\rm A}\approx k^0_{+\rm A}$, based on kinetic measurements \cite{De_La_Cruz.Sweeney1999}. This leaves us with a total of 4 unknown kinetic rates, of which 3 need to be estimated from the measured stepping behavior and run length data, while one can be determined from Eq.~\ref{eq:3}. \section*{Results} \subsection*{Choice of the value for the bending modulus} There are two ways to estimate the bending stiffness of the myosin V lever arm - one from its structure and analogy with similar molecules and the other one from the observed behavior of the dimeric molecule. The lever arm consists of 6 IQ motifs, forming an $\alpha$-helix, surrounded by 6 calmodulin or other light chains \cite{Wang.Sellers2003,Terrak.Dominguez2003}. One possible estimate for the stiffness of the lever arm can be obtained by approximating it with a coiled-coil domain, as has been done by \citet{Howard.Spudich1996}. Generally, the stiffness of a semiflexible molecule is related to its persistence length $\ell_p$ as $EI=\ell_p k_B T$. Howard and Spudich estimated the persistence length of a coiled-coil domain as $100\,{\rm nm}$, which yields $EI\approx 400\,{\rm pN\,nm}^2$. Other researchers report values of $\ell_p=130\,{\rm nm}$ for myosin \cite{Hvidt.Ferry1982} and $\ell_p=150\,{\rm nm}$ for tropomyosin \cite{Swenson.Stellwagen1989,Phillips.Chacko1996}. On the other hand, we can estimate the stiffness from the force a lever arm has to bear under conditions close to stall. We do this by calculating the distribution of binding probabilities to different sites at $F=1.8\,{\rm pN}$, which is close to stall force. We assume that the binding rate to each site is proportional to its Boltzmann weight, $\exp(- G / k_B T)$, which is equivalent to assuming that the activation point of the binding process is close to the final state and that the reverse reaction (detachment in the state with ADP.Pi) has no force-dependence in its rate. The expectation value of the binding position of the lead head relative to the trail head is shown in Fig.~\ref{fig:4}. It shows that a stiffness of $EI\gtrsim 1000\,{\rm pN\, nm^2 }$ is necessary to allow stepping at loads of this magnitude. \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure4} } \mycaption{The average step size under a load of $F=1.8\,{\rm pN}$ as a function of the lever arm elasticity $EI$. The step size was calculated from attachment probabilities of the lead head (ADP.Pi state) relative to the bound trail head (ADP state).} \label{fig:4} \end{figure} For these reasons, we use the value $EI=1500\,{\rm pN\,nm^2}$. This corresponds to an elastic constant (measured at the joint) of \begin{equation} \label{eq:9} k=3 EI/L^3=0.25\,{\rm pN/nm}\;. \end{equation} The elastic constant for longitudinal forces (with respect to the lever arm) is much higher. If we approximate the lever arm with a homogeneous cylinder of radius $r=1\,{\rm nm}$, we can estimate it as $k_L=4 EI /(r^2 L)=230\,{\rm pN/nm}$. We therefore neglect the longitudinal extensibility of the lever arm in all calculations. A similar value ($EI=1300\,{\rm pN\,nm^2}$) has also been obtained by analyzing data from optical trap experiments on single-headed myosin V molecules with different lever arm lengths \cite{Moore.Warshaw2004}. Even though it is somewhat larger (about 3 times) than the values estimated for myosin II \citep{Howard.Spudich1996}, there is no solid evidence that the structures with different light chains have the same bending stiffness. On the other hand, there could have been some evolutionary pressure to increase the lever arm stiffness, as it is directly related to the stall force of myosin V. While we are not able to give a definite answer to the question whether the lever arm behaves like a uniform elastic rod or whether there is a pliant region close to the head, we favor the first hypothesis because the estimated lever arm elasticity already is more than sufficient to explain the mechanical properties of the dimeric molecule. \subsection*{Step size distribution} Figure \ref{fig:3} shows the energies stored in the elastic distortions of the lever arms of both heads in the pre-powerstroke or the post-powerstroke state. For example, if the first head is in the ADP.Pi state and the second head binds before the first one undergoes a power stroke, this is connected with an energy cost of $6.6\,k_B T$. The attachment rate of the lead head before the power stroke in the trail head is therefore more than 100 times slower than after the power stroke. Because the lead head normally attaches to actin while the trail head is in the ADP state, we can determine the probability that the lead head binds to an actin site $i$ subunits in front of the trail head from the Boltzmann factors formed from the bending energy in the final configuration, $P_i \propto \exp (- (U_1+U_2)/k_B T)$. Here $U_1+U_2$ denotes the sum of elastic energies stored in both lever arms if the trail head is in the ADP state and the lead head in the ADP.Pi state, bound $i$ sites in front of the trail head. The resulting distributions for different lever arm lengths are shown in Fig.~\ref{fig:5}. For the lever arm consisting of 6 IQ motifs, the result is a mixture of 11 and 13 subunit steps, whereby 13 subunits dominate. Azimuthal distortion plays a major role in the bending energy, therefore binding is only likely to sites 2, 11, 13 and 15, on which the azimuthal angles of both heads differ by not more than $27^\circ$. \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure5} } \mycaption{Step size distribution for 4 different lever arm lengths L: 10nm (2IQ), 18nm (4IQ), 26nm (6IQ) and 34nm (8IQ) and no external load. The histograms show the probability that a lead head (ADP.Pi state) will bind $i$ sites in front of the trail head in the post-powerstroke ADP state. The probabilities were determined from the Boltzmann factors, resulting from the elastic distortion energy of the configuration. Azimuthal distortion plays a crucial role role in determining the step size, which is the reason why the binding is always concentrated on sites 2, 11, 13 and 15. Taking into account the fluctuations in the actin would lead to a broader distribution, in better agreement with experiments \citep{Walker.Knight2000}.} \label{fig:5} \end{figure} \subsection*{The gated step in the cycle} \label{sec:} A question that has been a subject of intense discussion is which step in the cycle is deciding for the coordination of the two heads. A currently often favored hypothesis proposes that the lead head undergoes its power stroke immediately after binding, thereby storing energy into elastic deformation of its lever arm and releasing it after the unbinding of the trail head. An alternative hypothesis proposes that the release of the rear head is necessary for the power stroke in the front head. As we will show below, our model favors this picture. In the 4-state scenario, this implies that the lead head is waiting in the ADP.Pi. In the 5-state scenario it is in the ADP$'$ state (the pre-powerstroke ADP state). The trail head spends most of its cycle in the ADP state in both scenarios at saturating ADP concentrations. Because this model challenges the currently prevailing view, we should first critically review the arguments supporting it. One argument includes the direct observation of telemark-shaped molecules, with the leading head leaning forward and then the lever arm tilted strongly backwards \cite{Walker.Knight2000}. A more detailed image analysis, however, showed that the converter of the leading head is in the pre-powerstroke state \cite{Burgess.Trinick2002}. Another piece of evidence comes from experiments by \citet{Forkey.Goldman2003} which show a fraction of tags on the lever arm (30-50\%) that do not tilt while moving, but again the data provide no conclusive proof because the method does not allow detection of tilts symmetric with respect to the vertical axis. To conclude, one cannot say that the present experimental evidence excludes any of the two hypotheses about the moment of phosphate release and of the power stroke. From the theoretical side, we will argue that in a model with linear elasticity the mechanism with immediate power stroke in the lead head cannot work under loads for which the motor is known to be operational. It is known that the monomeric constructs of myosin V undergo a normal duty cycle \cite{De_La_Cruz.Sweeney1999,Yengo.Sweeney2002}, which means that no step in the cycle requires mechanical work from the outside for its completion (which would be, for example, the case if the head needed to be pulled away from actin to complete the cycle). This excludes the possibility that the free energy gain connected with binding and the power stroke exceeds $\Delta G_{\rm ATP}=100\,{\rm pN nm}$, the total available energy for one cycle. Because this and other transitions in the cycle need to be forward-running, we use the still conservative estimate that the free energy gain from binding and the power stroke cannot exceed $80\,{\rm pN nm}$. On the other hand, we can estimate the free energy that would be necessary for a head to bind to a site 13 units ahead and then undergo a conformational change. The amount of energy needed to bring the dimer into the hypothetical state with both heads in the post-powerstroke state and a strong distortion, especially of the leading lever arm, is plotted in Fig.~\ref{fig:6}. The calculation shows that the binding of the front head with the subsequent power stroke before the rear head detaches (for a load of $F=1.8\,{\rm pN}$) is only possible for values of $EI\lesssim 450\,{\rm pN\,nm^2}$, which is inconsistent with the lower estimate based on the observed step size (Fig.~\ref{fig:4}). Of course, we cannot rule out that there is some additional state in the middle of the power stroke which is occupied immediately while the lead head waits for the trail head to detach. But within the scope of the geometrical model with a single power stroke connected with the Pi release, we consider the scenario where the lead head instantaneously undergoes the power stroke without waiting for the detachment of the trail head unrealistic. \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure6} } \mycaption{The amount of energy needed for the binding of the lead head and the subsequent power-stroke, plotted against the lever arm elasticity. The load pulling on the tail is $F=1.8\,{\rm pN}$. The lower curve shows the energy needed to pull the external load and distort the lever arms in order to bind the new lead head 13 sites in front of the trailing head. Note that most of this work will be performed by Brownian motion, but the potential well in the bound state still has to be strong enough to stabilize the bound state. The middle curve shows the energy needed mainly for the distortion of the lever arms when the lead head undergoes a power-stroke before the trailing head detaches. Since the sum of both cannot be higher than $80\,{\rm pN\,nm}$, we estimate that this hypothetical scenario would only be possible if the lever arm stiffness was $EI \lesssim 450\,{\rm pN\,nm^2}$. This is inconsistent with other requirements of the model, so we rule this scenario out.} \label{fig:6} \end{figure} \subsection*{Hidden power strokes in the dimer configuration} \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure7} } \mycaption{For a single head, the $x$-component of the power-stroke upon ADP release equals 3.3nm (for zero load). In the dimer with both heads bound, only 0.07nm of that power stroke reach the load. As a consequence, the load-dependence of transition rates between states with both heads bound is negligible.} \label{fig:7} \end{figure} An immediate consequence of the elastic lever arm model is that the tail position is mainly determined by the geometry of the triangle and less by the conformations of individual heads. For a monomeric head or a dimer bound by a single head, the power-stroke upon ADP release has an $x$-component (in the direction of the actin filament) of about $3.3\,{\rm nm}$ (Fig.~\ref{fig:7}). If the lead head is attached, however, the power stroke as measured on the tail is reduced by about a factor of 50. The tail movement is also closely related to the force-dependence of transition rates, which means that transitions between states with both heads bound do not show any significant load dependence. In the kinetic scheme we use here this implies that the rates of ADP release and ATP binding (the two rate limiting steps at low or forward loads) are both constant, in agreement with the flat F-v curve measured by \citet{Mehta.Cheney1999}. \begin{figure*}[htbp] \figurecontents{ \begin{tabular}{ll} \hspace*{-0.5cm}\includegraphics{Figure8a}& \includegraphics{Figure8b} \\ \hspace*{-0.5cm}A)&B) \end{tabular} } \mycaption{ Most probable kinetic pathways for a dimer in the 4-state (A) and in the 5-state model (B). The thick arrows denote the regular pathway and the thin arrows side branches that can result in dissociation from actin. Note that the simulation was not restricted to the pathways shown here, but included all possible combinations of transitions between monomer states. } \label{fig:8} \end{figure*} \begin{figure}[htbp] \figurecontents{ A) \includegraphics{Figure9a}\\ B) \includegraphics{Figure9b}\\ } \mycaption{A) Force-velocity curves in the 4-state model, obtained from a stochastic simulation. The solid curve shows the values for $1000\mu \rm M$ ATP and the dashed curve for $1 \mu \rm M$ ATP. Both curves are compared with the prediction of the simplified analytical expression (Eq.~\ref{eq:13}), dotted lines. The minor deviation is mainly due to cycles taking other pathways, neglected force-dependence of the ADP release rate and variation in the step size. Note that the velocities above $\sim 2.5\,{\rm pN}$ are not well defined because the dissociation time becomes comparable with the step time. B) Inhibition by ADP and Pi. The force-velocity relation with 1mM ATP is shown by the continuous line. The dashed line shows the same relation with additional $10\mu{\rm M}$ ADP and the dotted line with 1mM phosphate. The velocity reduction through ADP occurs at low or negative loads, while the inhibition by Pi only becomes significant close to stall conditions. } \label{fig:9} \end{figure} \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure10} } \mycaption{Force-velocity relation of the 5-state model with 1mM ATP (solid), 1mM ATP+ 10$\mu$M ADP (dashed) and 1$\mu$M ATP (dotted). Note the sharper drop at high loads as compared to the 4-state coupled model (Fig.~\ref{fig:9}).} \label{fig:10} \end{figure} \subsection*{Force-velocity and run length curves} The bending energies, calculated for each possible dimer configuration, and the transition rates were fed into a kinetic simulation to determine the average velocity of a dimeric motor and its dissociation rate from actin. The most probable kinetic pathway of the dimer is indicated by thick arrows in Fig.~\ref{fig:8}, while the thin arrows indicate some of the possible side branches that can lead to dissociation. Figure \ref{fig:9} shows the resulting force-velocity curves and Fig.~\ref{fig:11} the dissociation rates. An analytical solution of the 4-state model would, in theory, require solving the occupation probabilities for a system with about $6+8\times 3 \times 3=78$ states (6 states with one head bound, plus configurations with both heads bound, where each head can occupy 3 different states and the relative positions of both heads can have 8 different values). Such a system could easily be solved numerically, but would be too complex for obtaining an insightful analytical expression. However, we will show that a simplified pathway can already lead to expressions that agree reasonably well with simulation data and are therefore useful for fitting model parameters to experimental data. In the following, we give approximate expressions for the most significant steps in the mechanochemical cycle. The average time it takes for a head in the state 0 to bind an ATP molecule can be estimated as \begin{equation} \left< t_{\rm +ATP} \right> = \frac{1}{k_{\rm +ATP} [{\rm ATP}] } \left(1+ \frac{k_{\rm +ADP} [{\rm ADP}]}{k_{-\rm ADP}} \right) \end{equation} where the second term takes into account a reduction of the forward rate due to ADP rebinding. The second rate limiting process (especially at hight loads) is the release of phosphate. The average dwell time in the state with one head free and the other one in the ADP.Pi state is \begin{equation} \label{eq:11} \left< t_{\rm -Pi} \right> = \frac {1}{k_{\rm -Pi}} \end{equation} The third rate limiting step is the ADP release, with the time constant \begin{equation} \left< t_{\rm -ADP} \right> = \frac 1 {k_{\rm -ADP}}\;. \end{equation} With these three average dwell times, the motor velocity can be calculated as \begin{equation} \label{eq:13} v=\frac{\left< d \right>}{\left< t_{\rm -Pi} \right>+ \left< t_{\rm -ADP} \right> + \left< t_{\rm +ATP} \right> }\;, \end{equation} where $\left< d \right>$ denotes the average step size, which is about $35\,{\rm nm}$. The individual rates that appear in this expression can be estimated as follows: $k_{\rm -Pi}\approx k_{\rm -Pi}^0 \exp(-F \epsilon_{\rm -Pi}d_{\rm PS}/k_B T)$ with $d_{\rm PS}=L (\cos \phi_{\rm ADP}-\cos \phi_{\rm ADP.Pi})+\delta$ and $k_{\rm -ADP}\approx k_{\rm -ADP}^0 \exp(-\Delta U_{\rm -ADP}/2k_B T) \approx 0.65 k_{\rm -ADP}^0 $. The results for two different ATP concentrations are shown in Fig.~\ref{fig:9}A and compared with a simulation result. The analytical expression reproduces the simulation result well, with a small deviation being mainly the result of alternative pathways, neglected force-dependence of the ADP release rate and variation in the step size. The experimentally measured force-velocity curves \citep{Mehta.Cheney1999,Uemura.Ishiwata2004} are also well reproduced, although the experiments show a more abrupt drop in velocity at high loads, with no measurable effect up to about 1pN. In the 5-state model the power-stroke can be fast and reversible, in which case the pre- and the post-powerstroke state can reach an equilibrium and the limiting rate is proportional to the probability of the post-powerstroke state $1/(1+ \exp(F d_{\rm PS} /k_B T))$ - a significantly sharper load dependence than the 4-state model (Fig.~\ref{fig:10}). \subsection*{Inhibition by ADP and phosphate} It is a well established observation that ADP can slow down myosin V by binding to heads in the state with no nucleotide and thereby preventing them from accepting an ATP molecule. The rate of ADP rebinding is already taken into account in the kinetic constants and the model naturally reproduces the observed behavior, as shown in Figs.~\ref{fig:9} and \ref{fig:12} for the 4-state model and in Fig.~\ref{fig:10} for the 5-state model. Not yet experimentally investigated has been the inhibition by phosphate. Its intensity depends on the reverse power-stroke rate, which is one of the open parameters of our model. In the 4-state model, Pi re-binding is necessary for the reverse power stroke and therefore some inhibition effect can be expected at high loads. The simulation shows clearly that the phosphate concentration has no effect on zero-load velocity, but it does slow down the motor close to stall (Fig.~\ref{fig:9}B). A similar effect of Pi on isometric force has also been observed in muscle \cite{Cooke.Pate1985}. In the 5-state model Pi rebinding is not mechanically sensitive and its effect is roughly force-independent. However, with the parameters chosen here, it is negligible. \subsection*{Three dissociation pathways} As we can see from the kinetic scheme (Fig.~\ref{fig:8}), there are three significant pathways in the cycle that can lead to the dissociation of the myosin V dimer from an actin filament. The first pathway leaves the cycle if a dimer bound with one head in the ADP.Pi state detaches before the second head can attach. The second pathway runs through a state in which the bound head releases ADP and binds a new ATP molecule before the free head can bind. With the third pathway we denote all processes that involve the detachment of a head in the ADP state. This is the pathway favored by recent results of \citet{Baker.Warshaw2004}. Figure \ref{fig:11} shows the dissociation rate, separated by contributions of the three pathways. They have the following characteristics: \emph{Pathway 1:} With this pathway we denote the dissociation of a head in the ADP.Pi state. Because this state is long-lived at high loads in the 4-state, but short-lived in the 5-state model, the resulting force-dependence of the dissociation rate differs significantly in both scenarios. In the 4-state model, the contribution to the dissociation probability per step shows a strong load-dependence, but no significant dependence on the ATP concentration. It can be estimated as \begin{equation} \label{eq:14} P_{\rm diss}\approx \frac {k_{\rm -A}}{k_{\rm -Pi}}\approx \frac{k_{\rm -A}}{k^0_{\rm -Pi}} e^{\frac{F \epsilon_{\rm -Pi} d_{\rm PS}}{k_B T}} \end{equation} with $d_{\rm PS}=L (\cos \phi_{\rm ADP}-\cos \phi_{\rm ADP.Pi})+\delta$. The dissociation rate is higher for positive loads. From the estimated run length at $1\,{\rm pN}$ load and saturating ATP concentration of about 15 steps \cite{Clemen.Rief2003}, we can estimate the unbinding rate as $k_{\rm -A}\approx 1\,{\rm s}^{-1}$. In order to account for reported run lengths of over 50 steps at low loads, we tentatively assign $k_{\rm +A}^0\approx 5000\,{\rm s}^{-1}$. In the 5-state model, the situation is reversed. There the dissociation process on path 1 takes place if the trail head releases ADP before the lead head releases Pi, which can happen in two different ways: on one the rate is approximately force-independent, on the other it grows with negative (forward) loads. In order to obtain a significant contribution to the detachment rate on this pathway, we choose a higher detachment rate $k_{-A}$ than in the 4-state model ($50\,{\rm s}^{-1}$ instead of $1\,{\rm s}^{-1}$). \emph{Pathway 2:} Because the process of unbinding requires an ATP molecule, the per-step dissociation rate grows with the ATP concentration. In addition, it is proportional to the ratio of the $ADP$ dissociation rate and the actin binding rate, $k_{\rm -ADP}/k_{\rm +A}$, which is higher for negative (forward) loads. This holds in both the 4- and the 5-state scenario. \emph{Pathway 3:} The dissociation probability on pathway 3 is proportional to the detachment rate in the ADP state, $k'_{\rm -A}$. Of all three pathways, this one shows the weakest load-dependence, although it is higher for forward loads. We expect that systematic data on mean run length as a function of load and nucleotide concentrations will be helpful to determine the remaining model parameters. \begin{figure} \figurecontents{ \includegraphics{Figure11a}~\\ \includegraphics{Figure11b}~ } \mycaption{Dissociation rate of myosin V dimers from actin under a high (top) and a low (bottom) ATP concentration (4-state model). The continuous line shows the total dissociation rate, the dashed line the dissociation via pathway 1, the dot-dashed line via pathway 2 and the dotted line via pathway 3.} \label{fig:11} \end{figure} \begin{figure} \figurecontents{ \includegraphics{Figure12} } \mycaption{Velocity (continuous, left scale) and mean run length (dashed, right scale) as a function of ADP concentration in the 4-state model for zero load and 1mM ATP.} \label{fig:12} \end{figure} \begin{figure}[htbp] \figurecontents{\includegraphics{Figure13}} \mycaption{Force-dependence of the dissociation rate in the 5-state model. The load dependence for positive loads is much weaker than in the 4-state model (Fig.~\ref{fig:11})} \label{fig:13} \end{figure} \subsection*{ Reverse stepping in the 5-state model } As a consequence of both the reversibility of the power stroke and the slower dissociation rate at high loads, the motor can step backwards under loads exceeding the stall force (Fig.~\ref{fig:14}). Note that these steps are not the simple reversal of forward steps (which would involve ATP synthesis), but rather indicate a different pathway in the kinetic scheme, in which both heads stay in the ADP state and alternately release actin at the leading position and rebind at the trailing. The time scale of reverse stepping is determined by the dissociation rate of a head in the ADP state, $k'_{\rm -A}$, which we chose as $0.1\,{\rm s}^{-1}$. With a higher value of $k'_{\rm -A}$, especially for the pre-powerstroke state (so far we assumed that the rate is equal in both ADP states), faster stepping would also be possible, although there is an upper limit on $k'_{\rm -A}$, imposed by the dissociation rate on pathway 3. \begin{figure}[htbp] \figurecontents{ \includegraphics{Figure14} } \mycaption{Reverse stepping in the 5-state model under a high load (4.5pN), $10\,{\mu \rm M}$ ATP and $1\,{\mu \rm M}$ ADP. There is also some creeping motion between the steps, which results from the attachment and detachment of the two heads on neighboring sites, and only takes place if myosin V is allowed to follow a helical path on actin. If binding is constrained to one side of the actin filament (like on a coverslip), then only regular reverse steps with the periodicity of the helix are observed (not shown).} \label{fig:14} \end{figure} \section*{Discussion} We used the geometrical data of the myosin V molecule as obtained from EM images to calculate the conformations and elastic energies in all dimer configurations. These data were first used in a model with a four-state cycle and subsequently in a five-state model. The first result, which follows directly from the bending potentials and is independent of the underlying cycle is that the elastic lever arm model explains two key components of the coordination between heads: why the lead head does not bind to actin before the power stroke in the trail head and why it does not undergo its power stroke before the trail head detaches. It also allows us to calculate the distribution of step sizes. The results for different lever-arm lengths (Fig.~\ref{fig:5}) give realistic values, in agreement with step size and helicity measurements \cite{Purcell.Sweeney2002,Ali.Ishiwata2002}, even though they have a slight tendency towards underestimation and also show a narrower distribution than direct electron microscopy observations \cite{Walker.Knight2000}. A possible explanation for the broader distribution than predicted by the model lies in the fact that in reality the actin structure does not follow the perfect helix, as assumed in our model, but has angular deviations of up to $10^{\circ}$ per subunit \cite{Egelman.DeRosier1982}. Taking these fluctuations into account would clearly broaden the distribution of our step sizes, but alone it cannot explain the tendency towards longer steps. The most straightforward explanation for the longer steps is that the power stroke has an additional right-handed azimuthal component. Then the configuration with the lowest energy is reached if the lead head is twisted to the right relatively to the trail head, which is the case if it is bound further away along the helix. The observation that the actin repeat is often somewhat longer than 13 subunits (some results suggest a structure closer to a 28/13 helix \citep{Egelman.DeRosier1982}) could also partially explain the deviation. An issue that has been much discussed is the contribution of Brownian motion and the power stroke to the total step size. With the geometric data used in this study, the power stroke, i.e., the distance of the lever arm tip movement between the states ADP.Pi and ADP, is about 31nm, or 5nm less than the average step size. Note that the second, smaller power stroke connected with ADP release does not contribute to the step size because it is normally followed by the detachment of the same head. Its function could be suppressing premature dissociation before the lead head binds and thus improving the processivity. The remaining 5nm can be overcome by Brownian motion before the lead head binds. However, at low loads, the binding of the lead head does not move the load, but rather stores the energy into bent lever arms. This energy gets released when the rear head detaches, which leads to an elastic power stroke immediately preceding the power stroke upon Pi release. At higher loads the situation is different, because the 5nm load movement occurs when the lead head binds. In neither case we expect the 5nm power stroke to be resolvable under normal conditions because it always immediately precedes or follows the large power stroke. However, it is possible that the substeps become observable in the presence of chemicals that slow down the power stroke \citep{Uemura.Ishiwata2004}. In order to fully reproduce the substeps as reported by \citet{Uemura.Ishiwata2004}, some modifications would be necessary to the model. First, part of the power-stroke would have to occur immediately upon Pi release, resulting in a lever arm move of about $12\,\rm nm$ (first substep). This step would need a very strong force-dependence in its transition rate (activation point near the final state). The subsequent longer power stroke (ADP'$\to$ADP) would then need a slower rate ($\sim 200\,\rm s^{-1}$) with less force dependence (activation point close to the initial state). However, the finding that the substep position is independent of force remains difficult to explain, because the substep involves transition between a stiff configuration, bound on both heads, and a more compliant state, bound on a single head. The main value of both models (4- and 5-state) is that they provide a quantitative explanation of the coordinated head-over-head motility of the dimeric molecule while using only the properties of a single head as input. Both models also explain the observed force-velocity curves at high and low ATP concentration and the effect of additional ADP, but these features already reveal some testable differences between the two scenarios. One of them is the shape of the force-velocity curve. In the 4-state scenario the reverse power-stroke needs the rebinding of a phosphate molecule. This makes the cutoff behavior at high loads dependent on the Pi concentration: the velocity drop is more gradual at low, but might become sharper at high Pi concentrations (Fig.~\ref{fig:9}B). In the 5-state scenario the velocity decline is more abrupt regardless of the Pi concentration. This is the first suggestion how experiments with improved precision and a wider range of chemical conditions could help distinguishing between the two scenarios. The main difference between the two scenarios is the predicted shape of the run length. Because the dissociation can take place on three different pathways, its rate depends on a number of parameters, of which a few cannot yet be determined by other methods. In the 4-state model the dissociation rate at high loads is dominated by detachment of a head in the ADP.Pi state and it therefore depends on the ratio $k^0_{\rm -A}/ k^0_{\rm -Pi}$ (Eq.~\ref{eq:14}). A strong increase with the load is characteristic for the 4-state model, because the load slows down the phosphate release and prolongs the dwell time in the state that is most vulnerable to dissociation. Dissociation at negative (forward) loads is dominated by pathways 2 (ATP mediated actin release in one head before the other head has bound) and 3 (dissociation of a head with ADP). In the 5-state model all three pathways can contribute towards the dissociation rate, but there is no significant increase for positive loads - in fact, the dissociation rate can even decrease. The run length shortens with an increasing ADP concentration in both scenarios. The decrease in run length is weaker than the decrease in the velocity (Fig.~\ref{fig:12}), which is consistent with recent observations \cite{Baker.Warshaw2004}. However, we cannot reproduce the reported complete saturation of run length at hight ADP concentrations. \citet{Baker.Warshaw2004} explain this saturation with a big difference (50-fold) between the attachment rates of the lead head depending whether the trail head is in the ADP or apo state, which we currently cannot reproduce with the relatively small power stroke ($10^{\circ}$) upon ADP release in our model. An interesting difference between the 4- and the 5-state model is also that the 5-state model allows backward steps at high loads (above the stall force), while the 4-state model predicts rapid dissociation. In general, there are three possibilities how backward steps can occur: (i) The motor hydrolyzes ATP, but runs backwards. (ii) The motor slips backwards without hydrolyzing ATP --- this is the case in our model. (iii) The motor synthesizes ATP from ADP and phosphate while being pulled backwards, as assumed by tightly coupled stochastic stepper models \citep[e.g.,][]{Kolomeisky.Fisher2003}. It is possible to test these three possibilities experimentally: If (i) is the case, the backward sliding velocity should show a Michealis-Menten type dependence on ATP concentration. This mechanism would, however, require an even looser mechanochemical coupling, so that not only the release of Pi, but also the release of ADP and binding of ATP would be possible without completing the power stroke. In case (iii) it should depend on ADP as well as on Pi concentration, but not on ATP. In case (ii), which is favored by our study, the backward stepping occurs when both heads have ADP bound on them and they successively release actin at the lead position and rebind it at the new trail position. Even though this stepping requires no net reaction between the nucleotides, a certain (low) ADP concentration is still required to prevent the heads from staying locked in the rigor (no nucleotide) state. The application of the elastic lever-arm approach developed here should not be limited to simple geometries and longitudinal loads. A natural extension of the present work will be the influence of perpendicular forces on the activity of the motor. One will also be able to study the stepping behavior in more complex geometries, for example when passing a branching site induced by the Arp2/3 complex \cite{Machesky.Gould1999}. After completion of this manuscript, it has been brought to my attention that \citet{Lan.Sun2005} have also published a model for myosin V, based on the elasticity of the lever arm. In contrast to our model, they do not describe it as an isotropic rod, but use a weaker in-plane stiffness, combined with a strong (phenomenological) azimuthal term that prevents binding of both heads to adjacent sites on actin. Another difference is that their study explicitly excludes dissociation events, whereas we use the dissociation rate to determine some of the model parameters. \section*{Acknowledgment} I would like to thank Erwin Frey and Jaime Santos for help with calculating the lever-arm shape, Peter Knight for help with the geometry of the molecule, and Matthias Rief and Mojca Vilfan for helpful discussions. This work was supported by the Slovenian Office of Science (Grants No.~Z1-4509-0106-02 and P0-0524-0106). \section*{Appendix} \subsection*{Numerical solution for the lever arm shape} The aim of this calculation is to determine the shape of the dimeric molecule for a given set of binding sites (trailing head bound on the site with the index $i_1$, leading head with $i_2$), nucleotide states, which determine the lever arm starting angles $\phi_1$ and $\phi_2$, and a given external load $F$. We start this task by deriving a function that numerically determines the endpoint of a lever arm as a function of the force acting on it: $\mathbf{x}_{j}(\mathbf{F}_{j}, \phi_{j})$ ($j=1,2$). The shape of the whole molecule can then be determined numerically from the conditions that the endpoints of the two lever arms coincide, $\mathbf{x}_1=\mathbf{x}_2$, and from the force equilibrium in that point \begin{equation} \label{eq:15} \mathbf{F}_1+\mathbf{F}_2 = -F \hat{e}_x\;. \end{equation} In many cases the function $\mathbf{x}_{j}$ will have more than one solution. Then we solve the system with all possible combinations and then choose the solution with the lowest energy $U=U_1+U_2+Fx$, where $U_1$ and $U_2$ denote the energy stored in the distortion of each lever arm and $Fx$ the work performed against the applied load. For a head bound at site $i$, the position of the proximal end of its lever arm in Cartesian coordinates reads \begin{equation} \label{eq:16} \mathbf{x}^0=\left( \begin{array}{c} i a+\delta\\ - R \sin(\theta) \\ R \cos(\theta) \end{array} \right) \end{equation} and its initial tangent \begin{equation} \label{eq:17} \hat{t}^0=\left( \begin{array}{c} \cos(\phi) \\ - \sin(\phi) \sin(\theta) \\ \sin(\phi) \cos(\theta) \end{array} \right) \end{equation} where $\phi$ is the lever arm tilt (a function of the nucleotide state), $\delta$ is the relative position of the lever arm proximal end ($0$ or $3.5\,{\rm nm}$) and $\theta$ is the azimuthal angle of the actin subunit to which the head is bound, $\theta=\theta_0 i$ with $\theta_0\approx \frac{6}{13} \times 360^{\circ} \approx 166^{\circ}$. The helix rise per subunit is $a= 2.75\,{\rm nm}$. If the force $\mathbf{F}$ acts on a lever arm that leaves the head in the direction $\hat{t}^0$, the whole lever arm will be bent in a plane spanned by the vectors $\hat{t}^0$ and $\mathbf{F}$. We can introduce a new two-dimensional orthogonal coordinate system in this plane, so that \begin{align} \tilde{\hat{t}}^0&=\left( \begin{array}{c}0\\1\end{array} \right) & \tilde{\mathbf{F}}&=\left( \begin{array}{c}\tilde F _x \\ \tilde F _y \end{array}\right) \\ \tilde F_y&=\mathbf{F} \hat{t}_0 & \tilde F_x&=\left| \mathbf{F}-\hat{t}_0 ( \mathbf{F} \hat{t}_0 ) \right| \end{align} In this coordinate system the shape can be determined by solving the equations \begin{align} \label{eq:20} M(s)&=\tilde{\mathbf{F}} \wedge ( \tilde{\mathbf{x}}(L)-\tilde{\mathbf{x}}(s) )= EI \frac{d\phi(s)}{ds} \\ \frac{d\tilde{\mathbf{x}}}{ds} &= \hat{\tilde{t}} \qquad \hat{\tilde{t}}=\left( \begin{array}{c} \sin(\phi) \\ \cos(\phi) \end{array} \right) \end{align} with the boundary condition $\phi(0)=0$. The symbol ``$\wedge$'' denotes the outer product, which is the out-of-plane component of the vector product. If we differentiate Eq.~\ref{eq:20} by $\phi$ we get \begin{equation} \label{eq:22} EI \frac{d^2 \phi}{ds^2} = - \tilde{F}_x \cos(\phi) + \tilde{F}_y \sin(\phi) \end{equation} Through partial integration and taking into account the boundary condition $M(L)=0$, we finally obtain \begin{equation} \label{eq:23} \begin{split} &\frac{EI}{2}\left( \frac{d\phi}{ds} \right)^2 \\&= \tilde{F}_x (\sin\phi_L-\sin\phi)+\tilde{F}_y (\cos \phi_L - \cos \phi )\\ &\equiv F \sin\left( \frac{\phi_L-\phi}{2}\right) \sin \left( \phi_F -\frac{\phi_L+\phi}{2} \right) \end{split} \end{equation} Here we introduced the force angle $\phi_F$, so that $\tilde{F}_x=F \sin(\phi_F)$ and $\tilde{F}_y=F \cos(\phi_F)$. \begin{figure}[tbp] \figurecontents{ \includegraphics{Figure15} } \mycaption{ The shapes of an elastic beam anchored at one end and pulled by a given force $\mathbf F$ on its other end. The dashed line shows the unloaded beam. According to the sign of the initial curvature and the final angle $\phi_L$ the solutions can be divided into 4 classes. The beam corresponds to the myosin V lever arm, which is anchored in the head at one end and connected to a flexible joint at the other end. Note that the bending shown is exaggerated in comparison with realistic dimer configurations. } \label{fig:15} \end{figure} Because of the ambiguity of a quadratic equation, Eq.~\ref{eq:23} generally has two solutions for a given set of values for $\phi(s)$, $F$, $\phi_L$ and $\phi_F$. As we have defined the coordinate system in a way that $\tilde{F}_x \ge 0$, we have $0 \le \phi_F \le \pi$. We also restrict ourselves to solutions with $\left| \phi(s) \right|<2\pi$, i.e., we do not consider any spiraling solutions, because they always have a higher bending energy than the straighter solution with the same endpoint. There are four classes of functions $\phi(s)$ that satisfy the condition that the RHS of Eq.~\ref{eq:23} be positive: \begin{center} \begin{tabular}{cccc} Solution& $\phi_L$ & $\phi(s \to 0)$ & conditions \\ \hline I & + & + & $0 \le \phi_L \le \phi_F$ \\ II & - & - & $\phi_F-2\pi \le \phi_L \le 2\phi_F-2\pi$ \\ III & + & - & $0\le \phi_L \le \phi_F$ \\ IV & - & + & $\phi_F-2\pi \le \phi_L \le 2\phi_F-2\pi$ \\ \end{tabular} \end{center} The solutions III and IV have a turning point at $\phi_0=-2 (\pi-\phi_F) -\phi_L$, where $d\phi/ds$ changes sign. Eq.~\ref{eq:23} can finally be transformed to \begin{align} \label{eq:24} L&=\frac{1}{2} \sqrt{\frac{EI}{F}} I(\phi_L) \qquad (\text{cases I and II})\\ L&=\frac{1}{2} \sqrt{\frac{EI}{F}} (2 I(\phi_0)+I(\phi_L)) \qquad (\text{cases III and IV})\\ I(\phi_x)&=\left| \int_0^{\phi_x} \left( \sin\left( \frac{\phi_L-\phi}{2}\right) \sin \left( \phi_F -\frac{\phi_L+\phi}{2} \right) \right) ^{-1/2} d\phi \right| \nonumber \end{align} Note that for classes II and III the RHS of Eq.~\ref{eq:24} is not monotonous in $\phi_L$ and there can be two solutions for a given $L$. Taking this into account, we obtain a total of up to 6 solutions. A situation in which all cases are represented is shown in Fig.~\ref{fig:15}. The configuration of the dimer is determined by solving Eq.~\ref{eq:15} for all possible combinations of modes and taking the one with the lowest potential. The numerical integration and solution were performed using NAG libraries (Numerical Algorithms Group) and the 3-d graphical representation of the calculated shapes was made with POV-Ray (www.povray.org).
{ "timestamp": "2005-03-14T13:44:10", "yymm": "0503", "arxiv_id": "physics/0503109", "language": "en", "url": "https://arxiv.org/abs/physics/0503109" }
\section{Introduction} The relative entropy of states of quantum systems is a measure of how well one quantum state can be operationally distinguished from another. Defined as \begin{eqnarray}\nonumber S(\rho||\sigma)= \mathop{\rm Tr}\nolimits[\rho(\log \rho - \log \sigma) ] \end{eqnarray} for states $\rho$ and $\sigma$, it quantifies the extent to which one hypothesis $\rho$ differs from an alternative hypothesis $\sigma$ in the sense of quantum hypothesis testing \cite{ohya,Wehrl,disti,disti2,disti3}. Dating back to work by Umegaki \cite{umegaki}, the relative entropy is a quantum generalisation of the Kullback-Leibler relative entropy for probability distributions in mathematical statistics \cite{Kullback}. The quantum relative entropy plays an important role in quantum statistical mechanics \cite{Wehrl} and in quantum information theory, where it appears as a central notion in the study of capacities of quantum channels \cite{Schumacher,Holevo,Schumacher2,prop} and in entanglement theory \cite{prop,Plenio,Plenio2}. In {\it finite-dimensional Hilbert spaces}, the relative entropy functional is manifestly continuous \cite{Wehrl}, see also footnotes \footnote{ For states on infinite-dimensional Hilbert spaces the relative entropy functional is not trace norm continuous any more, but -- as the von-Neumann entropy -- lower semi-continuous. That is, for sequences of states $\{\sigma_n\}_n$ and $\{\rho_n\}_n$ converging in trace norm to states $\sigma$ and $\rho$, i.e., $\lim_{n\rightarrow\infty} \| \sigma_n-\sigma\|_1 =0$ and $\lim_{n\rightarrow\infty} \| \rho_n-\rho\|_1 =0$, we merely have $ S(\rho||\sigma)\leq \liminf_{n\rightarrow\infty} S(\rho_n||\sigma_n)$. However, for systems for which the Gibbs state exists, these discontinuities can be tamed \cite{Wehrl} when considering compact subsets of state space with finite mean energy. In a similar manner, entropic measures of entanglement can become trace norm continuous on subsets with bounded energy \cite{Jensito}.}, \footnote{ For considerations of the continuity of the relative entropy in classical contexts, see Ref.\ \cite{naudts}.}. In particular, if $\{\sigma_n\}_{n}$ is a sequence of states of fixed finite dimension satisfying \begin{equation}\nonumber \lim_{n\rightarrow \infty} || \sigma_n - \sigma ||_1 = \lim_{n\rightarrow \infty} \mathop{\rm Tr}\nolimits | \sigma_n - \sigma | =0 \end{equation} for a given state $\sigma$, then \begin{eqnarray}\nonumber \lim_{n\rightarrow\infty} S(\sigma_n|| \sigma)=0. \end{eqnarray} In practical contexts, however, more precise estimates can be necessary, in particular in an asymptotic setting. Consider a state $\rho$ on a Hilbert space ${\cal H}$, and a sequence $\{\sigma_n\}_n$, where $\sigma_n$ is a state on ${\cal H}^{\otimes n}$, the $n$-fold tensor product of ${\cal H}$. The sequence is said to asymptotically approximate $\rho$ if $\sigma_n$ tends to $\rho^{\otimes n}$ for $n\rightarrow \infty$. More precisely, one typically requires that \begin{equation}\nonumber \lim_{n\rightarrow\infty} \| \sigma_n - \rho^{\otimes n}\|_1 =0. \end{equation} Now, as an alternative to the trace norm distance, one can consider the use of the Bures distance. The Bures distance $D$ is defined as $$ D(\rho_1,\rho_2) = 2\left(1-F(\rho_1,\rho_2)\right)^{1/2}, $$ in terms of the Uhlmann fidelity $$ F(\rho_1,\rho_2) = \mathop{\rm Tr}\nolimits( \rho_1^{1/2} \rho_2 \rho_1^{1/2})^{1/2}. $$ Because of the inequalities \cite{Hayden} \begin{equation}\label{hayden} 1-F(\rho_1,\rho_2) \le \mathop{\rm Tr}\nolimits|\rho_1-\rho_2| \le \left(1-F^2(\rho_1,\rho_2)\right)^{1/2}, \end{equation} the trace norm distance tends to zero if and only if the Bures distance tends to zero, which shows that, for the purpose of state discrimination, both distances are essentially equivalent and one can use whichever is most convenient. A natural question that now immediately arises is whether the same statement is true for the relative entropy. To find an answer to that one would need inequalities like (\ref{hayden}) connecting the quantum relative entropy, used as a distance measure, to the trace norm distance, or similar distance measures. In this paper, we do just that: we find upper bounds on the relative entropy functional in terms of various norm differences of the two states. As such, the presented bounds are very much in the same spirit as Fannes' inequality, sharpening the notion of continuity for the von Neumann entropy \cite{fannes}. It has already to be noted here that one of the main stumbling blocks in this undertaking is the well-known fact that the relative entropy is not a very good distance measure, as it gives infinite distance between non-identical pure states. However, we will present a satisfactory solution, based on using the minimal eigenvalue of the state that is the second argument of the relative entropy. Apart from the topic of upper bounds, we also study lower bounds on the relative entropy, giving a complete picture of the relation between norm based distances and relative entropy. We start in Section II with presenting a short motivation of how this paper came about. Section III contains the relevant notations, definitions and basic results that will be used in the rest of the paper. In Section IV we discuss some properties of unitarily invariant (UI) norms that will allow us to consider all UI norms in one go. The first upper bounds on the relative entropy $S(\rho||\sigma)$ are presented in Section V, one bound being quadratic in the trace norm distance of $\rho$ and $\sigma$ and the other logarithmic in the minimal eigenvalue of $\sigma$. Both bounds separately capture an essential behaviour of the relative entropy, and it is argued that finding a single bound that captures both behaviours at once is not a trivial undertaking. Nevertheless, we will succeed in doing this in Section VII by constructing upper bounds that are as sharp as possible for given trace norm distance \textit{and} minimal eigenvalue of $\sigma$. In Section VI we use similar techniques to derive lower bounds that are as sharp as possible. Finally, in Section IX, we come back to the issue of state discrimination mentioned at the beginning. \section{Background} In Ref.\ \cite{brat2} (Example 6.2.31, p.\ 279) we find the following upper bound on the relative entropy, valid for all $\rho$ and for non-singular $\sigma$: \begin{equation}\label{bound_brat} S(\rho||\sigma) \le \frac{||\rho-\sigma||_\infty}{\lambda_{\min}(\sigma)}. \end{equation} This bound is linear in the operator norm distance between $\rho$ and $\sigma$ and has a $1/x$ dependence on $\lambda_{\min}(\sigma)$. For several purposes, such a bound is not necessarily sharp enough. The impetus for the present paper was given by the observation that sharper upper bounds on the relative entropy should be possible than (\ref{bound_brat}). Specifically, there should exist bounds that are \begin{enumerate} \item {\em quadratic} in $\rho-\sigma$, and/or \item depend on $\lambda_{\min}(\sigma)$ in a {\em logarithmic} way. \end{enumerate} A simple argument shows that a logarithmic dependence on $\lambda_{\min}(\sigma)$ can be achieved instead of an $1/x$ dependence. Note that $0\ge\log\sigma\ge\mathbbm{1}\cdot\log\lambda_{\min}(\sigma)$. Thus, \begin{eqnarray} S(\rho||\sigma) &=& \mathop{\rm Tr}\nolimits[\rho(\log\rho-\log\sigma)] \nonumber \\ &\le& -S(\rho)-\log\lambda_{\min}(\sigma) \nonumber \\ &\le& -\log\lambda_{\min}(\sigma). \label{bound1} \end{eqnarray} Concerning the quadratic dependence on $\rho-\sigma$, we can put $\rho=\sigma+\varepsilon\Delta$, with $\mathop{\rm Tr}\nolimits[\Delta]=0$, and calculate the derivative $$\lim_{\varepsilon\rightarrow0} S(\sigma+\varepsilon\Delta||\sigma) / \varepsilon $$ and find that this turns out to be zero for any non-singular $\sigma$. Indeed, the gradient of the relative entropy $S(\rho||\sigma)$ with respect to $\rho$ is $\mathbbm{1}+\log\rho-\log\sigma$ (see Lemma \ref{lemma1}). Hence, for $\rho=\sigma$ and $\mathop{\rm Tr}\nolimits[\Delta]=0$, \begin{eqnarray}\nonumber \lim_{\varepsilon\rightarrow0} S(\sigma+\varepsilon\Delta||\sigma)/\varepsilon =\mathop{\rm Tr}\nolimits[\Delta(\mathbbm{1}+\log\sigma-\log\sigma)]=0. \end{eqnarray} This seems to imply that for small $\varepsilon$, $S(\sigma+\varepsilon\Delta||\sigma)$ must at least be quadratic in $\varepsilon$, and, therefore, upper bounds might exist that indeed are quadratic in $\varepsilon$. Furthermore, Ref.\ \cite{ohya} contains the following quadratic lower bound (Th.\ 1.15) \begin{equation}\nonumber \label{bound_ohya} S(\rho||\sigma) \ge \frac{1}{2}||\rho-\sigma||_1^2. \end{equation} The rest of the paper will be devoted to finding firm evidence for these intuitions, by exploring the relation between relative entropy and norm based distances, culminating in a number of bounds that are the sharpest possible. \section{Notation} In this paper, we will use the following notations. We will use the standard vector and matrix bases: $e^i$ is the vector with the $i$-th element equal to 1, and all other elements being equal to $0$. $e^{i,j}$ is the matrix with $i,j$ element equal to 1 and all other elements 0. For any diagonal matrix $A$, we write $A_i$ as a shorthand for $A_{i,i}$, and $\mathop{\rm Diag}\nolimits(a_1,a_2,\ldots)$ is the diagonal matrix with $a_i$ as diagonal elements. We reserve two symbols for the following special matrices: \begin{equation}\nonumber E:=\mathop{\rm Diag}\nolimits(1,0,\ldots,0) = e^{1,1}, \end{equation} and \begin{equation}\nonumber F:=\mathop{\rm Diag}\nolimits(1,-1,0,\ldots,0) = e^{1,1}-e^{2,2}. \end{equation} The positive semi-definite order is denoted using the $\ge$ sign: $A\ge B$ iff $A-B\ge 0$ (positive semi-definite). The (quantum) relative entropy is denoted as $S(\rho||\sigma) = \mathop{\rm Tr}\nolimits[\rho(\log\rho-\log\sigma)]$. All logarithms in this paper are natural logarithms. When $\rho$ and $\sigma$ are both diagonal (i.e., when we encounter the commutative, classical case) we use the shorthand \begin{eqnarray*} \lefteqn{S((r_1,r_2,\ldots)||(s_1,s_2,\ldots))} \\ &:=& S(\mathop{\rm Diag}\nolimits(r_1,r_2,\ldots)||\mathop{\rm Diag}\nolimits(s_1,s_2,\ldots)). \end{eqnarray*} \begin{lemma}\label{lemma1} The gradient of the relative entropy $S(\rho||\sigma)$ with respect to its first argument $\rho$, being non-singular, is given by $\mathbbm{1}+\log\rho-\log\sigma$. \end{lemma} \textit{Proof.} The calculation of this derivative is straightforward. Since the classical entropy function $x\longmapsto h(x) := -x \log x $ is continuously differentiable on $(0,1)$, and therefore, \begin{eqnarray*} \frac{\partial}{\partial\varepsilon}\Big|_{\varepsilon=0} S(\rho+\varepsilon \Delta ) = \mathop{\rm Tr}\nolimits[\Delta h'(\rho) ], \end{eqnarray*} we can write \begin{eqnarray} \label{log_deriv} \lim_{\varepsilon\rightarrow0} S(\rho+\varepsilon\Delta||\sigma)/\varepsilon &=& \mathop{\rm Tr}\nolimits[\Delta(\mathbbm{1}+\log\rho-\log\sigma)].\nonumber \end{eqnarray} \hfill$\square$\par\vskip24pt Finally, we recall a number of series expansions related to the logarithm, which are valid for $-1<y<1$, \begin{eqnarray} \log(1-y) &=& -\sum_{k=1}^\infty \frac{y^k}{k},\nonumber\\ \log(1+y)+\log(1-y) &=& -\sum_{k=1}^\infty\frac{y^{2k}}{k} ,\nonumber\\ \log(1+y)-\log(1-y) &=& 2\sum_{k=0}^\infty\frac{y^{2k+1}}{2k+1}. \nonumber \end{eqnarray} These expansions will be made extensive use of. \section{Unitarily Invariant Norms} In this section we collect the main definitions and known results about unitarily invariant norms along with a number of refinements that will prove to be very useful for the rest of the paper. A \textit{unitarily invariant norm} (UI norm), denoted with $|||.|||$, is a norm on square matrices that satisfies the property \begin{equation}\nonumber |||UAV|||=|||A||| \end{equation} for all $A$ and for unitary $U$, $V$ (\cite{bhatia}, Section IV.2). Perhaps the most important property of UI norms is that they only depend on the singular values of the matrix $A$. If $A$ is positive semi-definite, then $|||A|||$ depends only on the eigenvalues of $A$. A very important class of UI norms are the \textit{Ky Fan norms} $||.||_{(k)}$, which are defined as follows: for any given square $n\times n$ matrix $A$, with singular values $s_j^\downarrow(A)$ (sorted in non-increasing order) and $1\le k\le n$, the $k$-Ky Fan norm is the sum of the $k$ largest singular values of $A$: $$ ||A||_{(k)} = \sum_{j=1}^k s_j^\downarrow(A). $$ Two special Ky Fan norms are the \textit{operator norm} and the \textit{trace norm}, \begin{equation}\nonumber ||A||_\infty = ||A||_{(1)},\,\, ||A||_{\mbox{Tr}} = ||A||_1 = ||A||_{(n)}. \end{equation} The importance of the Ky Fan norms derives from their leading role in Ky Fan's \textit{Dominance Theorem} (Ref.\ \cite{bhatia}, Theorem IV.2.2): \begin{theorem}[Ky Fan Dominance] Let $A$ and $B$ be any two $n\times n$ matrices. If $B$ majorises $A$ in all the Ky Fan norms, $$ ||A||_{(k)}\le ||B||_{(k)}, $$ for all $k=1,2,...$ , then it does so in all other UI norms as well, $$ |||A|||\le|||B|||. $$ \end{theorem} From Ky Fan's Dominance Theorem follows the following well-known norm dominance statement. \begin{lemma} For any matrix $A$, and any unitarily invariant norm $|||.|||$, $$ ||A||_\infty \le \frac{|||A|||}{|||E|||} \le ||A||_1. $$ \end{lemma} \textit{Proof.} We need to show that, for every $A$, $$ |||(||A||_\infty)E||| \le |||A||| \le |||(||A||_1)E|||, $$ holds for every unitarily invariant norm. By Ky Fan's dominance theorem, we only need to show this for the Ky Fan norms. All the Ky Fan norms of $E$ are 1, and $$ ||A||_\infty =||A||_{(1)} \le ||A||_{(k)} \le ||A||_{(d)}=||A||_1 $$ follows from the definition of the Ky Fan norms. \hfill$\square$\par\vskip24pt The main mathematical object featuring in this paper is not the state, but rather the difference $\Delta$ of two states, $\Delta: =\rho-\sigma$, and for that object a stronger dominance result obtains. We first show that the largest norm difference between two states occurs for orthogonal pure states. Indeed, by convexity of norms, $|||\rho-\sigma|||$ is maximal in pure $\rho$ and $\sigma$. A simple calculation then reveals that, for any unitarily invariant norm, $$ |||\,\ket{\psi}\bra{\psi} - \ket{\phi}\bra{\phi}\,||| = \left( 1-|\langle\psi|\phi\rangle|^2 \right)^{1/2} |||F|||. $$ This achieves its maximal value $|||F|||$ for $\psi$ orthogonal to $\phi$, showing that it makes sense to normalise a norm distance $|||\rho-\sigma|||$ by division by $|||F|||$. We will call this a \textit{rescaled} norm. We now have the following dominance result for rescaled norms of differences of states: \begin{lemma}\label{lemma:dom} For any Hermitian $A$, with $\mathop{\rm Tr}\nolimits[A]=0$, $$ \frac{||A||_\infty}{||F||_\infty} \le \frac{|||A|||}{|||F|||} \le \frac{||A||_1}{||F||_1}. $$ \end{lemma} Note that equality can be obtained for any value of $|||A|||$, by setting $A=cF$. \textit{Proof.} We need to show, for all traceless Hermitian $A$, that \begin{equation}\label{eq:dom} |||(||A||_\infty)F||| \le |||A||| \le |||(||A||_1/2) F ||| \end{equation} holds for every unitarily invariant norm. Again by Ky Fan's dominance theorem, we only need to do this for the Ky Fan norms $||.||_{(k)}$. Since $$ ||F||_{(k)} = \left\{\begin{array}{cc}1,&k=1,\\2,&k>1,\end{array}\right. $$ and \begin{eqnarray*} ||X||_\infty = ||X||_{(1) }\le ||X||_{(k)} \le ||X||_{(d)}=||X||_1, \end{eqnarray*} the inequalities (\ref{eq:dom}) follow trivially for $k>1$. The case $k=1$ is covered by Lemma \ref{lemma:k1} below. \hfill$\square$\par\vskip24pt \begin{lemma}\label{lemma:k1} For any Hermitian $A$, with $\mathop{\rm Tr}\nolimits[A]=0$, $$ ||A||_1 \ge 2||A||_\infty. $$ \end{lemma} \textit{Proof.} Let the Jordan decomposition of $A$ be \begin{equation}\nonumber A=A_+ - A_-, \end{equation} with $A_+$, $A_-\ge0$. Since $\mathop{\rm Tr}\nolimits[A]=0$, clearly $\mathop{\rm Tr}\nolimits[A_+]=\mathop{\rm Tr}\nolimits[A_-]$ holds. Thus, $||A||_1=\mathop{\rm Tr}\nolimits|A|=\mathop{\rm Tr}\nolimits[A_+]+\mathop{\rm Tr}\nolimits[A_-]=2\mathop{\rm Tr}\nolimits[A_+]$. Also, \begin{equation}\nonumber ||A||_\infty = \max(||A_+||_\infty,||A_-||_\infty). \end{equation} Hence, $||A||_\infty \le \max(||A_+||_1,||A_-||_1) = \mathop{\rm Tr}\nolimits[A_+] = ||A||_1/2$. \hfill$\square$\par\vskip24pt In this paper, we will also be dealing with $\Delta=\rho-\sigma$ under the constraint $\sigma\ge\beta\mathbbm{1}$. Obviously we have $$ \beta\le 1/d. $$ We now show that under this constraint, any rescaled norm of $\Delta$ is upper bounded by $1-\beta$. \begin{lemma}\label{lem:maxt} For any state $\rho$, and states $\sigma$ such that $\sigma\ge\beta\mathbbm{1}$, \begin{equation}\nonumber T:=|||\rho-\sigma|||/|||F||| \le 1-\beta. \end{equation} \end{lemma} \textit{Proof.} We proceed by maximising $T$ under the constraint $\sigma\ge\beta\mathbbm{1}$. Convexity of norms yields that $T$ is maximal when $\rho$ and $\sigma$ are extremal \cite{Convex}, hence in $\rho$ being a pure state $\ket{\phi}\bra{\phi}$ and $\sigma$ being of the form \begin{equation}\nonumber \sigma = \beta\mathbbm{1} + (1-\beta d)\ket{\psi}\bra{\psi}. \end{equation} Fixing $\phi=e^1$, we need to maximise $$ |||e^{1,1}-\beta\mathbbm{1}-(1-\beta d)\ket{\psi}\bra{\psi}\, ||| $$ over all $\psi$. Put $\psi=(\cos\alpha,\sin\alpha,0,\ldots,0)^T$, then the eigenvalues of the matrix are $$ \lambda_\pm=\left((d-2)\beta\pm \left( {(\beta d)^2+4(1-\beta d)\sin^2\alpha} \right)^{1/2} \right)/2 $$ and $-\beta$ (with multiplicity $d-2$). One finds that, for $d>2$, $\lambda_+,|\lambda_-|\ge\beta$, for any value of $\alpha$, and both $\lambda_+$ and $|\lambda_-|$ are maximal for $\alpha=\pi/2$, as would be expected. The maximal Ky Fan norms of this matrix are therefore \begin{eqnarray*} ||.||_{(1)} &=& \lambda_+ = 1-\beta ,\\ ||.||_{(k)} &=& \lambda_+ + |\lambda_-|+(k-2)\beta = (2-\beta d) + (k-2)\beta, \end{eqnarray*} for $k>1$. Hence, for every Ky Fan norm, the maximum norm value is obtained for orthogonal $\phi$ and $\psi$. By the Ky Fan dominance theorem, this must then hold for any UI norm. In case of the trace norm, as well as of the operator norm, the rescaled value of the maximum is $1-\beta$. By Lemma \ref{lemma:dom}, this must then be the maximal value for any rescaled norm. \hfill$\square$\par\vskip24pt \textit{Remark.} For the Schatten $q$-norm, $|||F||| = 2^{1/q}$. The largest value of $|||F|||$ is 2, obtained for the trace norm, and the smallest value is 1, for the operator norm. \section{Some simple upper bounds} In this Section we present our first attempts at finding upper bounds that capture the essential features of relative entropy. In Subsection A we present a bound that is indeed quadratic in the trace norm distance, the existence of which was already hinted at in Section II. Likewise, in Subsection B, we find a bound that is logarithmic in the minimal eigenvalue of $\sigma$, again in accordance with previous intuition. Combining the two bounds into one that has both of these features turns out to be not so easy. In fact, in Subsection C a number of arguments are given that initially hinted at the impossibility of realising sich a combined bound. Nevertheless, we will succeed in finding a combined bound later on in the paper, by using techniques from optimisation theory \cite{Convex}. \subsection{A quadratic upper bound} \begin{lemma} For any positive definite matrix $A$ and Hermitian $\Delta$ such that $A+\Delta $ is positive definite, $$ \log(A+\Delta)-\log(A) \le \int_0^\infty dx(A+x)^{-1}\Delta(A+x)^{-1}. $$ \end{lemma} {\em Proof.} Since the logarithm is strictly matrix concave \cite{HJII}, for all $t\in[0,1]$: $$ \log((1-t)A+tB) \ge (1-t)\log(A)+t\log(B). $$ Setting $B=A+\Delta$ and rearranging terms then gives $$ \frac{\log(A+t\Delta)-\log(A)}{t} \ge \log(A+\Delta)-\log(A), $$ for all $t\in[0,1]$. A fortiori, this holds in the limit for $t$ going to zero, and then the left-hand side is just the Fr\'echet derivative of $\log$ at $A$ in the direction $\Delta$. \hfill$\square$\par\vskip24pt This Lemma allows us to give a simple upper bound on $S(\sigma+\Delta||\sigma)$. Note that if $A\ge B$ then $\mathop{\rm Tr}\nolimits [CA]\ge\mathop{\rm Tr}\nolimits [CB]$ for any $C\ge 0$. Therefore, we arrive at \begin{eqnarray*} S(\rho||\sigma) &=& \mathop{\rm Tr}\nolimits\left[(\sigma+\Delta)(\log(\sigma+\Delta)-\log(\sigma))\right] \\ &\le& \int_0^\infty dx\mathop{\rm Tr}\nolimits[(\sigma+\Delta)(\sigma+x)^{-1}\Delta(\sigma+x)^{-1}] \\ &=& \int_0^\infty dx\mathop{\rm Tr}\nolimits[(\sigma+x)^{-1}\sigma(\sigma+x)^{-1}\Delta] \\ & + &\int_0^\infty dx\mathop{\rm Tr}\nolimits[\Delta(\sigma+x)^{-1}\Delta(\sigma+x)^{-1}]. \end{eqnarray*} The first integral evaluates to $\mathop{\rm Tr}\nolimits[\Delta]$, because $$ \int_0^\infty dx \frac{s}{(s+x)^2} = 1 $$ for any $s>0$, and therefore gives the value $0$. The second integral can be evaluated most easily in a basis in which $\sigma$ is diagonal. Denoting by $s_i$ the eigenvalues of $\sigma$, we get \begin{eqnarray} &&\int_0^\infty dx\mathop{\rm Tr}\nolimits[\Delta(\sigma+x)^{-1} \Delta(\sigma+x)^{-1}] \nonumber \\ &=& \sum_{i,j} \Delta_{i,j}\Delta_{j,i} \int_0^\infty dx (s_i+x)^{-1} (s_j+x)^{-1} \nonumber\\ &=& \sum_{i\neq j} \Delta_{i,j}\Delta_{j,i} \frac{\log s_i -\log s_j }{s_i-s_j} + \sum_i (\Delta_{i,i})^2 \frac{1}{s_i}.\label{dec} \end{eqnarray} The coefficients of $\Delta_{i,j}\Delta_{j,i}$ are easily seen to be always positive, and furthermore, bounded from above by $1/\lambda_{\min}(\sigma)$. Hence we get the upper bound $$ \int_0^\infty dx\mathop{\rm Tr}\nolimits[\Delta(\sigma+x)^{-1}\Delta(\sigma+x)^{-1}] \le \frac{\mathop{\rm Tr}\nolimits[\Delta^2]}{\lambda_{\min}(\sigma)}, $$ yielding an upper bound on the relative entropy which is, indeed, quadratic in $\Delta$: \begin{theorem} For states $\rho$ and $\sigma$ with $\Delta=\rho-\sigma$, $T=||\Delta||_2$ and $\beta=\lambda_{\min}(\sigma)$, \begin{equation}\nonumber \label{bound_quad} S(\rho||\sigma) \le \frac{T^2}{\beta}. \end{equation} \end{theorem} \subsection{An upper bound that is logarithmic in the minimum eigenvalue of $\sigma$} We have already found a sharper bound than (\ref{bound_brat}) concerning its dependence on $\lambda_{\min}(\sigma)$. However, the bound (\ref{bound1}) is not sharp at all concerning its dependence on $\rho-\sigma$. A slight modification can greatly improve this. First note \begin{eqnarray*} |\mathop{\rm Tr}\nolimits[\Delta\log\sigma]| &\le& ||\Delta||_1 \cdot ||\log\sigma||_{\infty} \\ &=& \mathop{\rm Tr}\nolimits|\Delta|\cdot|\log\lambda_{\min}(\sigma)|. \end{eqnarray*} This inequality can be sharpened, since $\mathop{\rm Tr}\nolimits[\Delta]=0$ and $\sigma$ is a state. Let $\Delta=\Delta_+ - \Delta_-$ be the Jordan decomposition of $\Delta$, then \begin{equation}\nonumber |\mathop{\rm Tr}\nolimits[\Delta\log\sigma]| \le ||\Delta_+||_1\cdot|\log\lambda_{\min}(\sigma)|, \end{equation} and hence \begin{eqnarray*} |\mathop{\rm Tr}\nolimits[\Delta\log\sigma]| &\le& \mathop{\rm Tr}\nolimits|\Delta|/2 \cdot|\log\lambda_{\min}(\sigma)|. \end{eqnarray*} Furthermore, we have Fannes' continuity of the von Neumann entropy \cite{fannes}, $$ |S(\sigma+\Delta)-S(\sigma)| \le T \log d +\min\left(-T\log T,\frac{1}{e}\right), $$ where $d$ is the dimension of the underlying Hilbert space and $T:=\mathop{\rm Tr}\nolimits |\Delta| $. Combining all this with $$ S(\sigma+\Delta||\sigma) = -(S(\sigma+\Delta)-S(\sigma)) - \mathop{\rm Tr}\nolimits[\Delta\log\sigma] $$ gives rise to the subsequent upper bound, logarithmic in the smallest eigenvalue of $\sigma$. \begin{theorem} For all states $\rho$ and $\sigma$ on a $d$-dimensional Hilbert space, with $T=||\rho-\sigma||_1$ and $\beta=\lambda_{\min}(\sigma)$, \begin{equation} S(\rho||\sigma) \le T \log d + \min\biggl (-T\log T,\frac{1}{e}\biggr) - \frac{T\log\beta}{2}. \label{bound_log} \end{equation} \end{theorem} \subsection{A combination of two bounds?} The following question comes to mind almost automatically: can we combine the two bounds (\ref{bound_quad}) and (\ref{bound_log}) into a single bound that is both quadratic in $\Delta$ and logarithmic in $\lambda_{\min}(\sigma)$? This would certainly be a very desirable feature for a good upper bound. For instance, could it be true that $$ S(\rho||\sigma) \le C\cdot \mathop{\rm Tr}\nolimits[(\rho-\sigma)^2]\cdot |\log\lambda_{\min}(\sigma)|, $$ for some constant $C>0$? Unfortunately, the answer to this first attempt is negative. In fact, the proposed inequality is violated no matter how large the value of $C$. \begin{proposition} For any $r>0$ there exist states $\sigma$ and $\rho$ such that \begin{equation} \label{bound_bad} S(\rho||\sigma) > r\cdot \mathop{\rm Tr}\nolimits[(\rho-\sigma)^2]\cdot | \log\lambda_{\min}(\sigma)|. \end{equation} \end{proposition} \textit{Proof.} It suffices to consider the case that $\sigma,\rho$ are states acting on the Hilbert space ${\mathbb{C}}^2$, and that $\sigma$ and $\rho$ commute. Hence, the statement must only be shown for two probability distributions \begin{eqnarray*} P=(p,1-p),\,\,\,Q=(q,1-q). \end{eqnarray*} Without loss of generality we can require $q$ to be in $[0,1/2]$. Then, one has to show that for any $r>0$ there exist $p,q$ such that the $C^\infty$-function $f$, defined as \begin{eqnarray*} f(p,q,r) &=& r \left((p-q)^2 + (2-p-q)^2\right)\, | \log(q)| \\ &-& \left( p \log(p/q) +(1-p) \log[(1-p)/(1-q)] \right), \end{eqnarray*} assumes a negative value. Now, for any $r>1$, fix a $q\in(0,1/2)$ such that $- 4 r (q \log q)<1$. Clearly, $$ f(q,q,r) =0,\,\,\,\,\,\, \frac{\partial}{\partial p}\bigr|_{p=q} f(p,q,r)=0. $$ Then \begin{eqnarray*} \frac{\partial^2}{\partial p^2} \bigr|_{p=q} f(p,q,r) &=& -\frac{1}{1-q} - \frac{1}{q} - 4 r \log(q) \\ &<& - \frac{1}{q} - 4 r \log(q)< 0 . \end{eqnarray*} This means that there exists an $\varepsilon>0$ such that $f(p,q,r)<0$ for $p\in[q,q+\varepsilon]$, which in turn proves the validity of (\ref{bound_bad}). \hfill$\square$\par\vskip24pt The underlying reason for this failure is that the two bounds (\ref{bound_log}) and (\ref{bound_quad}) are incompatible, in the sense that there are two different regimes where either one or the other dominates. To see when the logarithmic dependence dominates, let us again take the basis where $\sigma$ is diagonal, with $s_i$ being the main diagonal elements. When keeping $\Delta=\rho-\sigma$ fixed while $s_1=\lambda_{\text{min}}(\sigma)$ tends to zero, then \begin{eqnarray}\nonumber \lim_{s_1\rightarrow 0} S(\sigma+\Delta||\sigma)/ |\log s_1| = \Delta_{1,1}<\infty. \end{eqnarray} Hence, in the regime where $\lambda_{\min}(\sigma)$ tends to zero and $\rho-\sigma$ is fixed, the bound (\ref{bound_log}) is the appropriate one. The other regime is the one where $\sigma$ is fixed and $\rho-\sigma$ tends to zero. This can be intuitively seen by considering the case where the states $\rho$ and $\sigma$ commute (the classical case). Let $p_i$ and $q_i$ be the diagonal elements of $\rho$ and $\sigma$, respectively, in a diagonalising basis, and $r_i=p_i-q_i$. Then \begin{eqnarray}\nonumber S(\rho||\sigma) = \sum_i (q_i+r_i)\log(1+r_i/q_i). \end{eqnarray} We can develop $S(\rho||\sigma)$ as a Taylor series in the $r_i$, giving $$ S(\rho||\sigma) = \sum_i \frac{r_i^2}{2q_i} + O(r_i^3). $$ Hence, in the regime where $\rho-\sigma$ tends to zero and $\sigma$ is otherwise fixed, the relative entropy exhibits the behaviour of bound (\ref{bound_quad}). In terms of the matrix derivatives, this notion can be made more precise as follows. Denote the bound (\ref{bound_log}) as $$ g(\rho||\sigma) = \frac{\mathop{\rm Tr}\nolimits[(\rho-\sigma)^2]}{\lambda_{\min}(\sigma)} $$ for states $\rho,\sigma$, then clearly $$ \lim_{\varepsilon \rightarrow 0} g(\sigma +\varepsilon \Delta ||\sigma)/\varepsilon=0. $$ On using the integral representation of the second Fr\'echet derivative of the matrix logarithm \cite{ohya}, \begin{eqnarray*} \lefteqn{\frac{\partial^2}{\partial \varepsilon^2}\Big|_{\varepsilon=0} \log(\sigma + \varepsilon \Delta)} \\ & =& -2 \int_0^\infty dx (\sigma +x)^{-1} \Delta (\sigma+x)^{-1} \Delta (\sigma+x)^{-1}, \end{eqnarray*} one obtains \begin{eqnarray*} \lefteqn{\frac{\partial^2}{\partial \varepsilon^2}\Big|_{\varepsilon=0} S(\sigma+\varepsilon \Delta||\sigma)} \\ &=& -2 \mathop{\rm Tr}\nolimits\biggl[\sigma \int_0^\infty dx (\sigma +x)^{-1} \Delta (\sigma+x)^{-1} \Delta (\sigma+x)^{-1}\biggr]\\ & & +2 \mathop{\rm Tr}\nolimits\left[\Delta\int_0^\infty dx (\sigma+x)^{-1} \Delta (\sigma+x)^{-1}\right]. \end{eqnarray*} The right hand side is bounded from above by \begin{eqnarray*} \lefteqn{\frac{\partial^2}{\partial \varepsilon^2}\Big|_{\varepsilon=0} S(\sigma+\varepsilon \Delta||\sigma)} \nonumber\\ &\leq& \int_0^\infty dx\mathop{\rm Tr}\nolimits[\Delta(\sigma+x)^{-1}\Delta(\sigma+x)^{-1}], \end{eqnarray*} see Ref.\ \cite{lieb,ohya}. This bound can be written as in Eq.\ (\ref{dec}). Therefore, one can conclude that \begin{eqnarray*} \frac{\partial^2}{\partial \varepsilon^2}\Big|_{\varepsilon=0} S(\sigma+\varepsilon \Delta ||\sigma) = \frac{\partial^2}{\partial \varepsilon^2}\Big|_{\varepsilon=0} g(\sigma+\varepsilon \Delta ||\sigma) \end{eqnarray*} holds for all $\Delta$ satisfying $\mathop{\rm Tr}\nolimits[\Delta]=0$ if and only if $\sigma=\mathbbm{1}/d$, where $d$ is the dimension of the underlying Hilbert space. These considerations seem to spell doom for any attempt at ``unifying'' the two kinds of upper bounds. However, below we will see how a certain change of perspective will allow us to get out of the dilemma. \section{A sharp lower bound in terms of norm distance} We define $S_{\min}(T)$ with respect to a norm to be the smallest relative entropy between two states that have a distance of exactly $T$ in that norm, that is \begin{equation} \label{eq:def_smin} S_{\min}(T) = \min_{\rho,\sigma} \left\{S(\rho||\sigma): |||\rho-\sigma||| = T \right\}. \end{equation} When one agrees to assign $S(\rho||\sigma)=+\infty$ for non-positive $\rho$, the definition of $S_{\min}$ can be rephrased as \begin{equation} \label{eq:def_smin2} S_{\min}(T) = \min_{\Delta,\sigma} \left\{S(\sigma+\Delta||\sigma): |||\Delta||| = T, \mathop{\rm Tr}\nolimits[\Delta]=0 \right\}. \end{equation} Intuitively one would guess that $S_{\min}$ is monotonously increasing with $T$. The following lemma shows that this is true, but some care is required in proving it. \begin{lemma}\label{lem:mono} For $T_1\le T_2$, $S_{\min}(T_1)\le S_{\min}(T_2)$. \end{lemma} \textit{Proof.} Keep $\sigma$ fixed and define $$ f_\sigma(T) = \min_{\Delta} \left\{S(\sigma+\Delta||\sigma): |||\Delta||| = T, \mathop{\rm Tr}\nolimits[\Delta]=0\right\}, $$ so that $S_{\min}(T) = \min_\sigma f_\sigma(T)$. Considering $S(\sigma+\Delta||\sigma)$ as a function of $\Delta$, it is convex and minimal in the origin $\Delta=0$. Furthermore, for the norm balls \begin{equation}\nonumber {\cal B}(T):=\{\Delta: |||\Delta|||\le T, \mathop{\rm Tr}\nolimits[\Delta]=0\} \end{equation} we have \begin{equation}\nonumber \{0\}={\cal B}(0)\subseteq{\cal B}(T_1)\subseteq{\cal B}(T_2). \end{equation} This is sufficient to prove that $0=f_\sigma(0)\le f_\sigma(T_1)\le f_\sigma(T_2)$. Now, since this holds for any $\sigma$, it also holds when minimising over $\sigma$, and that is just the statement of the Lemma. \hfill$\square$\par\vskip24pt As a direct consequence, a third equivalent definition of $S_{\min}(T)$ is \begin{equation} \label{eq:def_smin3} S_{\min}(T) = \min_{\Delta,\sigma} \left \{S(\sigma+\Delta||\sigma): |||\Delta||| \ge T, \mathop{\rm Tr}\nolimits[\Delta]=0 \right\}. \end{equation} We now show that one can restrict oneself to the commutative case. \begin{lemma} The minimum in Eq.\ (\ref{eq:def_smin2}) is obtained for $\sigma$ and $\Delta$ commuting. \end{lemma} \textit{Proof.} Fix $\Delta$ and consider a basis in which $\Delta$ is diagonal. Let $\rho\longmapsto\mathop{\rm Diag}\nolimits(\rho)$ be the completely positive trace-preserving map which, in that basis, sets all off-diagonal elements of $\rho$ equal to zero. Thus $\mathop{\rm Diag}\nolimits(\Delta)=\Delta$. By monotonicity of the relative entropy, $$ S(\sigma+\Delta||\sigma) \ge S(\mathop{\rm Diag}\nolimits(\sigma)+\Delta||\mathop{\rm Diag}\nolimits(\sigma)). $$ Minimising over all states $\sigma$ then gives \begin{eqnarray*} \min_\sigma S(\sigma+\Delta||\sigma) &\ge& \min_\sigma S(\mathop{\rm Diag}\nolimits(\sigma)+\Delta||\mathop{\rm Diag}\nolimits(\sigma)) \\ &=& \min_\sigma\left \{S(\sigma+\Delta||\sigma):[\sigma,\Delta]=0\right \}. \end{eqnarray*} On the other hand, the states $\sigma$ that commute with $\Delta$ are included in the domain of minimisation of the left-hand side, hence equality holds. \hfill$\square$\par\vskip24pt For later reference we define the auxiliary function \begin{equation}\label{eq:sofx} s(x) := \min_{0<r<1-x} S((r+x,1-r-x)||(r,1-r)), \end{equation} for $0\le x< 1$. An equivalent expression for this function is given by \begin{equation}\nonumber s(x) := \min_{x<r<1} S((r-x,1-r+x)||(r,1-r)). \end{equation} The first three non-zero terms in its series expansion around $x=0$ are given by: \begin{equation}\nonumber s(x) = 2 x^2 + \frac{4}{9}x^4 + \frac{32}{135}x^6 + O(x^{8}) \label{bound_ohya_better} \end{equation} (obtained using a computer algebra package). Further calculations reveal that some of the higher-order coefficients are negative, the first one being the coefficient of $x^{62}$. One can easily prove \cite{csiszar} that the lowest order expansion $2x^2$ is actually a lower bound. It is, therefore, the sharpest quadratic lower bound. For values of $x$ up to $1/2$, the error incurred by considering only the lowest order term in (\ref{bound_ohya_better}) is at most 6.5\%. For larger values of $x$, the error increases rapidly. In fact, when $x$ tends to its maximal value of 1, $s(x)$ tends to infinity, as can easily be seen from the minimisation expression ($r$ tends to 0); accordingly, the series expansion diverges. For values of $x>4/5$, $s(x)$ is well approximated by its upper bound \begin{eqnarray*} s(x) &\le & \lim_{r\rightarrow 1-x} S((r+x,1-r-x)||(r,1-r))\nonumber\\ & = & -\log(1-x). \end{eqnarray*} This is illustrated in Figure \ref{fig1}. \begin{figure} \includegraphics[width=3.4in]{s2.eps} \caption{\label{fig1}The function $s$ defined in Eq.\ (\ref{eq:sofx}) (middle curve), the lower bound $2x^2$ (lower curve), and the upper bound $-\log(1-x)$ (upper curve).} \end{figure} Let us now come back to Eq.\ (\ref{eq:def_smin2}), with $\sigma$ and $\Delta$ diagonal, and $|||.|||$ any unitarily invariant norm. Let $\sigma$ and $\Delta$ have diagonal elements $\sigma_k$ and $\Delta_k$, respectively. Fixing $\Delta$, we minimise first over $\sigma$. This is a convex problem and any local minimum is automatically a global minimum \cite{Convex}. The corresponding Lagrangian is \begin{equation}\nonumber {\cal L} = \sum_k \sigma_k (1+\Delta_k/\sigma_k)\log(1+\Delta_k/\sigma_k) - \nu \left(\sum_k\sigma_k-1\right). \end{equation} The derivative of the Lagrangian with respect to $\sigma_k$ is \begin{equation}\label{eq:deriv2} \frac{\partial{\cal L}}{\partial \sigma_k} = \log(1+\Delta_k/\sigma_k)-\Delta_k/\sigma_k-\nu. \end{equation} This must vanish in a critical point, giving the expression \begin{equation}\nonumber \log(1+\Delta_k/\sigma_k) = \Delta_k/\sigma_k+\nu. \label{eq:de2} \end{equation} Now note that the equation $\log(1+x)-x=b$, for $b<0$ has only two real solutions, one positive and one negative, and none for $b>0$. Therefore, for any $k$ $\Delta_k/\sigma_k$ can assume only one of these two possible values. Let $K$ be an integer between 1 and $d-1$. Without loss of generality we can set \begin{equation}\label{eq:ds} \Delta_k/\sigma_k = \left\{\begin{array}{ll}c_p,& 1\le k\le K, \\ -c_m,& K<k\le d,\end{array} \right. \end{equation} where $c_p$ and $c_m$ are positive numbers, to be determined along with $K$. The requirement $\sum_k \Delta_k=0$ imposes $$ c_p \sum_{k=1}^K \sigma_k - c_m \sum_{k=K+1}^d \sigma_k = 0, $$ which upon defining \begin{equation}\label{eq:r} r: = \sum_{k=1}^K \sigma_k, \end{equation} turns into \begin{equation}\label{eq:cc} c_p r = c_m (1-r) =: c. \end{equation} Substituting Eqs.\ (\ref{eq:ds}) and (\ref{eq:r}), the function to be minimised becomes $$ r (1+c_p)\log(1+c_p) +(1-r)(1-c_m)\log(1-c_m), $$ which, given Eq.\ (\ref{eq:cc}), can be rewritten as $$ S((r+c,1-r-c)||(r,1-r)). $$ The one remaining constraint $|||\Delta|||=T$ likewise becomes $$ |||(c_p\sigma_1,\ldots,c_p\sigma_K,-c_m\sigma_{K+1},\ldots,-c_m\sigma_d)|||=T. $$ Defining \begin{eqnarray*} \tau' &:=& (\sigma_1,\ldots,\sigma_K)/r , \\ \tau'' &:=& (\sigma_{K+1},\ldots,\sigma_d)/(1-r), \end{eqnarray*} this turns into $$ T=|||(c_p r\tau' ; -c_m(1-r)\tau'')|||=c |||(\tau';\tau'')|||, $$ where we have exploited the homogeneity of a norm. Note that by their definition, $\tau'$ and $\tau''$ are vectors consisting of positive numbers adding up to 1. The minimisation itself thus turns into $$ S_{\min}(T) = \min_{r,\tau',\tau''} S((r+c,1-r-c)||(r,1-r)), $$ where $c :=T/|||(\tau';\tau'')|||$. Quite obviously, the minimum over $c$ is obtained for the smallest possible $c$, hence \begin{eqnarray*} S_{\min}(T) & =& \min_{r} S((r+T/\gamma,1-r-T/\gamma)||(r,1-r))\nonumber\\ & = & s(T/\gamma), \end{eqnarray*} with $$ \gamma=\max_{\tau',\tau''} |||(\tau';\tau'')|||. $$ By convexity of a norm, this maximum is obtained in an extreme point, so $$ \gamma=|||F|||. $$ Incidentally, by Lemma \ref{lem:maxt}, this value is also the maximum $$ \max_{\rho,\sigma}|||\rho-\sigma|||, $$ over all possible states $\rho$ and $\sigma$, i.e., $\gamma$ is the largest possible value of $T$ for the given norm. We have thus proven \begin{theorem} For any unitarily invariant norm $|||.|||$, we have the sharp lower bound \begin{equation}\nonumber S(\rho||\sigma) \ge s(|||\rho-\sigma|||/|||F|||). \end{equation} \end{theorem} A few remarks are in order at this point: \begin{enumerate} \item Within the setting of finite-dimensional systems, this theorem generalises a result of Hiai, Ohya and Tsukada \cite{hot,ohya} for the trace norm to all unitarily invariant norms. This paper also uses the technique of getting lower bounds by projecting on an abelian subalgebra and then exploiting the case of a two-dimensional support as the worst case scenario. \item If we take the Hiai-Ohya-Tsukada result for granted and combine it with Lemma \ref{lemma:dom}, we immediately get \begin{eqnarray*} S(\rho||\sigma) &\ge & s(||\rho-\sigma||_1/||F||_1)\nonumber\\ & \ge & s(|||\rho-\sigma|||/|||F|||). \end{eqnarray*} \item The divergence of $s$ at $x=1$ is easily understood. The largest norm difference between two states occurs for orthogonal pure states, in which case their relative entropy is infinite. \end{enumerate} \section{Sharp upper bounds in terms of norm distance} Let now $S_{\max}(T,\beta)$ be the largest relative entropy between $\rho$ and $\sigma$ that have a normalised distance of exactly $T$ and $\lambda_{\min}(\sigma)=\beta$, so let \begin{equation} \label{eq:def_smax} S_{\max}(T,\beta) := \max_{\rho,\sigma} \left \{S(\rho||\sigma): \frac{|||\rho-\sigma|||}{|||F|||} = T, \lambda_{\min}(\sigma)=\beta\right \}. \end{equation} The need for the extra parameter $\beta$ arises because for $\beta=0$, $S_{\max}$ is infinite, as can be seen by taking different pure states for $\rho$ and $\sigma$. We can rephrase this definition as \begin{eqnarray} \label{eq:def_smax2} S_{\max}(T,\beta) &=& \max_{\Delta,\sigma} \Biggl \{S(\sigma+\Delta||\sigma): \frac{|||\Delta|||}{|||F|||} = T, \mathop{\rm Tr}\nolimits[\Delta]=0, \nonumber \\ && \sigma+\Delta\ge0,\lambda_{\min}(\sigma)=\beta\Biggr \}. \end{eqnarray} Because $\Delta$ commutes with the identity matrix, there is a unique common least upper bound on $\beta\mathbbm{1}$ and $-\Delta$, which we will denote by $\max(\beta\mathbbm{1},-\Delta)$. In the eigenbasis of $\Delta$, this is a diagonal matrix with diagonal elements $\max(\beta,-\Delta_{i})$. The constraints $\sigma\ge\beta$ and $\sigma+\Delta\ge 0$ can therefore be combined into the single constraint \begin{equation}\label{eq:sigcon} \sigma \ge \max(\beta\mathbbm{1},-\Delta). \end{equation} The extremal $\sigma$ obeying this constraint are \begin{equation}\nonumber \sigma = \max(\beta\mathbbm{1},-\Delta) + \eta \ket{\psi}\bra{\psi}, \end{equation} where $\psi$ is any state vector, and \begin{equation}\nonumber \eta := 1-\mathop{\rm Tr}\nolimits[\max(\beta\mathbbm{1},-\Delta)]. \end{equation} Therefore, the constrained maximisation over $\sigma$ can be replaced by an unconstrained maximisation over all pure states of the function \begin{eqnarray} S(\Delta+\max(\beta\mathbbm{1},-\Delta) + \eta \ket{\psi}\bra{\psi}\,\, || \nonumber \\ \qquad \max(\beta\mathbbm{1},-\Delta) + \eta \ket{\psi}\bra{\psi}). \label{eq:maximand} \end{eqnarray} Of course, all of this puts constraints on $\Delta$ as well. Indeed, in order that states $\sigma$ obeying (\ref{eq:sigcon}) exist, $\max(\beta\mathbbm{1},-\Delta)$ must obey the condition \begin{equation}\nonumber \mathop{\rm Tr}\nolimits[\max(\beta\mathbbm{1},-\Delta)]\le 1. \end{equation} We now have to distinguish between two cases: the case $d=2$, and the case $d>2$. \subsection{The case $d=2$} For the $d=2$ case, the maximisation over $\Delta$ is trivial. In its eigenbasis, $\Delta$ is a multiple of $\mathop{\rm Diag}\nolimits(1,-1)=F$. Hence, fixing the eigenbasis of $\Delta$ (which we can do because of unitary invariance of the relative entropy), and fixing \begin{equation}\nonumber |||\Delta|||/|||F|||=T, \end{equation} actually leaves just one possibility for $\Delta$, namely $\Delta=TF$. The term $\max(\beta\mathbbm{1},-\Delta)$ leads to two cases: $T\le\beta$ and $T>\beta$. The condition $T\le\beta$ implies, by Lemma \ref{lemma:dom}, that $||\Delta||_\infty\le\beta$ and, hence, \begin{eqnarray*} \max(\beta\mathbbm{1},-\Delta) &=& \mathop{\rm Diag}\nolimits(\beta,\beta) ,\\ \eta &=& 1-2\beta. \end{eqnarray*} The remaining maximisation of (\ref{eq:maximand}) is therefore given by \begin{eqnarray} \max_\psi S(\mathop{\rm Diag}\nolimits(\beta+T,\beta-T)+(1-2\beta)\ket{\psi}\bra{\psi} \, || \nonumber\\ \qquad \mathop{\rm Diag}\nolimits(\beta,\beta)+(1-2\beta)\ket{\psi}\bra{\psi}).\label{eq:maxi1a} \end{eqnarray} Positivity of $\eta$ requires $\beta\le 1/2$. By unitary invariance of the relative entropy, and invariance of diagonal states under diagonal unitaries (phase factors), we can restrict ourselves to vectors $\psi$ of the form $\psi=(\cos\alpha,\sin\alpha)^T$. \begin{lemma}\label{lem:convex} For a state vector $\psi=(\cos\alpha,\sin\alpha)^T$, the function to be maximised in (\ref{eq:maxi1a}) is convex in $\cos(2\alpha)$. \end{lemma} \textit{Proof.} Let $D_1$ be the determinant of the first argument. It is linear in $t:=\cos(2\alpha)$: $$ D_1 = \beta^2-T^2+(1-2\beta)(\beta-T t). $$ After some basic algebra involving eigensystem decompositions of the states, the function to be maximised in (\ref{eq:maxi1a}) is found to be given by \begin{eqnarray*} f(x) &:=& ((1-x)\log(1-x)+(1+x)\log(1+x))/2 \\ &+& (-1+2\beta-2T t)(\log(1-\beta)-\log\beta)/2 \\ &-& (\log(4-4\beta)+\log\beta)/2, \end{eqnarray*} where $x=(1-4D_1)^{1/2}$. We will now show that this function is convex in $t$. Since the second and third terms are linear in $t$, we only need to show convexity for the first term. The series expansion of the first term is $$ ((1-x)\log(1-x)+(1+x)\log(1+x))/2 = \sum_{k=1}^\infty \frac{x^{2k}}{2k(2k-1)}. $$ Every term in the expansion is a positive power of $x^2$ with positive coefficient and is therefore convex in $x^2$, which itself is linear in $t$. The sum is therefore also convex in $t$. \hfill$\square$\par\vskip24pt By the above Lemma, the maximum of the maximisation over $\psi$ is obtained for extremal values of $t$, that is: either $\psi=(1,0)^T$ or $\psi=(0,1)^T$. Evaluation of the maximum is now straightforward and it can be checked that the choice $\psi=(1,0)^T$ always yields the largest value of the relative entropy. We will now more specifically look at the case where $T>\beta$. In this case, we get \begin{eqnarray*} \max(\beta\mathbbm{1},-\Delta) &=& \mathop{\rm Diag}\nolimits(\beta,T), \\ \eta &=& 1-\beta-T, \end{eqnarray*} and the remaining maximisation of (\ref{eq:maximand}) is given by \begin{eqnarray} \max_\psi S(\mathop{\rm Diag}\nolimits(\beta+T,0)+(1-\beta-T)\ket{\psi}\bra{\psi} \, || \nonumber\\ \qquad \mathop{\rm Diag}\nolimits(\beta,T)+(1-\beta-T)\ket{\psi}\bra{\psi}).\label{eq:maxi1} \end{eqnarray} Positivity of $\eta$ requires $\beta\le 1/2$ and $T\le 1-\beta$. Again, we can restrict ourselves to states $\psi=(\cos\alpha,\sin\alpha)^T$. We also have the equivalent of Lemma \ref{lem:convex}, which needs more work in this case: \begin{lemma} For a state vector $\psi=(\cos\alpha,\sin\alpha)^T$, the function to be maximised in (\ref{eq:maxi1}) is convex in $\cos(2\alpha)$. \end{lemma} \textit{Proof.} Let $D_1$ and $D_2$ be the determinant of the first and second argument, respectively. Both are linear in $t:=\cos(2\alpha)$: \begin{eqnarray*} D_1 &=& (1-\beta-T)(\beta+T)(1-t)/2 \\ D_2 &=& ((\beta+T-\beta^2-T^2)+(1-\beta-T)(T-\beta)t)/2. \end{eqnarray*} In the $(D_1,D_2)$-plane, this describes a line segment with gradient $$ K:= -\frac{T-\beta}{T+\beta}, $$ which lies in the interval $[-1,0]$. Again, after some basic algebra, the function to be maximised in (\ref{eq:maxi1}) is identified to be $f((1-4D_1)^{1/2}, (1-4D_2)^{1/2})$, where \begin{eqnarray*} f(x,y) &:=& ((1-x)\log(1-x)+(1+x)\log(1+x))/2 \\ &+& ((x^2+y^2-2y-4T^2)\log(1-y) \\ &-& (x^2+y^2+2y-4T^2)\log(1+y))/4 y. \end{eqnarray*} We will now show that $f((1-4D_1)^{1/2}, (1-4D_2)^{1/2})$ is convex in $t$. First, note that \begin{eqnarray*} f(x,y)=f_0(x,y)+T^2 f_1(y). \end{eqnarray*} The term $f_1(y)$ is itself convex in $t$: its series expansion is $$ f_1(y) = (\log(1+y)-\log(1-y))/y = 2\sum_{k=0}^\infty \frac{y^{2k}}{2k+1}, $$ which by the positivity of all its coefficients is convex in $y^2$, and $y^2$ is linear in $t$. The other term, $f_0(x,y)$ is given by a sum of three terms \begin{eqnarray*} f_0(x,y) &=& \frac{1}{2}((1-x)\log(1-x)+(1+x)\log(1+x)) \\ &+& \frac{1}{4}((y-2)\log(1-y)-(y+2)\log(1+y)) \\ &-& \frac{x^2}{4}(\log(1+y)-\log(1-y))/y. \end{eqnarray*} Replacing each of the three terms by its series expansion yields \begin{eqnarray*} f_0(x,y) &=& \sum_{k=1}^\infty \frac{x^{2k}}{2k(2k-1)} \\ &+& \sum_{k=1}^\infty (k-1)\frac{y^{2k}}{2k(2k-1)} \\ &-& \frac{x^2}{2}\sum_{k=0}^\infty \frac{y^{2k}}{2k+1}. \end{eqnarray*} To show that this function is convex in $t$, we will evaluate it along the curve \begin{eqnarray*} x^2 &=& u+p \\ y^2 &=& v+Kp, \end{eqnarray*} with gradient $K$ between 0 and $-1$, and $u$ and $v$ lying in the interval $[0,1]$, and check positivity of its second derivative with respect to $p$ at $p=0$: \begin{eqnarray*} &&{\left. \frac{\partial^2}{\partial p^2}\right |_{p=0} f_0(x,y)} =\sum_{k=2}^\infty \frac{k-1}{2k-1}u^{k-2} \\ &&+(k-1)\left(K\frac{(k-1)K-2}{2k-1} - K^2 \frac{k}{2k+1}u \right)v^{k-2}. \end{eqnarray*} The coefficient of $u^{k-2}$ is clearly positive, hence the derivative is positive if the coefficient of $v^{k-2}$ is positive for all allowed values of $u$ and $K$. The worst case occurs for $u=1$, yielding a coefficient $$ K\frac{(k-1)K-2}{2k-1} - K^2 \frac{k}{2k+1} = \frac{-K (2 + 4 k + K)}{(2k-1)(2k+1)}. $$ For values of $K$ between 0 and $-1$, this is indeed positive. \hfill$\square$\par\vskip24pt By the above Lemma, the maximum of the maximisation over $\psi$ is obtained for extremal values of $t$, that is: either $\psi=(1,0)^T$ or $\psi=(0,1)^T$. Evaluation of the maximum is again straightforward, and calculations show that sometimes $\psi=(1,0)^T$ yields the larger value, and sometimes $\psi=(0,1)^T$. In this way we have obtained the upper bounds: \begin{theorem}\label{th:ub2b} Let $\Delta=\rho-\sigma$, $T=|||\Delta|||/|||F|||$ and $\beta=\lambda_{\min}(\sigma)$. For $d=2$, and $T\le\beta$, \begin{eqnarray} S(\rho||\sigma) &\le& (T+1-\beta)\log\frac{T+1-\beta}{1-\beta} \nonumber \\ &+& (\beta-T)\log(1-T/\beta) \label{bound_quad2}. \end{eqnarray} For $d=2$, and $T>\beta$, \begin{eqnarray} S(\rho||\sigma) &\le& \max(-\log(1-T) , \nonumber\\ && (\beta+T)\log(1+T/\beta)+\nonumber\\ && (1-\beta-T)\log(1-T/(1-\beta))). \end{eqnarray} \end{theorem} It is interesting to study the behaviour of the bound in the case of large $\beta$. More specifically, an approximation for bound (\ref{bound_quad2}), valid for $T \ll \beta$, is \begin{eqnarray} S(\rho||\sigma) &\le& \sum_{k=2}^\infty \frac{T^k}{k(k-1)} \left(\frac{1}{\beta^k}-\frac{(-1)^k}{(1-\beta)^k}\right) \nonumber \\ &\approx& \frac{T^2}{2\beta(1-\beta)}, \label{bound_quad2_approx} \end{eqnarray} Figure \ref{fig2} illustrates the combined upper bounds of Theorem \ref{th:ub2b} ($d=2$) for various values of $\beta$. \begin{figure*} \begin{tabular}{cc} \includegraphics[width=3.4in]{ud2b0_1.eps} & \includegraphics[width=3.4in]{ud2b0_2.eps} \\ (a) & (b) \\ \includegraphics[width=3.4in]{ud2b0_3.eps} & \includegraphics[width=3.4in]{ud2b0_5.eps} \\ (c) & (d) \end{tabular} \caption{\label{fig2}Upper bounds of Theorem \ref{th:ub2b} on $S=S(\rho||\sigma)$ vs.\ the rescaled norm distance $T=|||\rho-\sigma|||/|||F|||$, for $d=2$, and for values of smallest eigenvalue of $\sigma$ (a) $\beta=0.1$, (b) $0.2$, (c) $0.3$, and (d) $0.5$. The two regimes $T\le\beta$ and $\beta\le T\le 1-\beta$ can be clearly identified. For ease of comparison, each curve is shown superimposed on the curves for $\beta=0.1$, $0.2$, $0.3$, $0.4$ and $0.5$ (in grey).} \end{figure*} \subsection{The case $d>2$} In case $d$ is larger than 2, it is not clear how to proceed in the most general setting, for general UI norms, as the maximisation over $\Delta$ must explicitly be performed. In the following, we will restrict ourselves to using the trace norm, which is in some sense the most important one anyway. That is, the requirements on $\Delta$ are \begin{eqnarray} ||\Delta||_1 &=& 2T, \label{eq:delta1} \\ \mathop{\rm Tr}\nolimits[\Delta] &=& 0 ,\label{eq:delta2} \\ \mathop{\rm Tr}\nolimits[\max(\beta\mathbbm{1},-\Delta)] &\le& 1. \label{eq:delta3} \end{eqnarray} The following very simple Lemma will prove to be a powerful tool. \begin{lemma}\label{lem:sabc} For all $A$, $B$, and $C$, positive semi-definite operators, $$ S(A+C||B+C) \le S(A||B). $$ \end{lemma} \textit{Proof.} First note that for any $a>0$, \begin{eqnarray*} S(aA||aB) &=& \mathop{\rm Tr}\nolimits[aA(\log(aA)-\log(aB))] \\ &=& a S(A||B). \end{eqnarray*} This, together with joint convexity of the relative entropy in its arguments (which need not be normalised to trace 1), leads to \begin{eqnarray*} S(A+C||B+C) &=& 2 S(\frac{A+C}{2} || \frac{B+C}{2}) \\ &\le& S(A||B) + S(C||C) \\ &=& S(A||B). \end{eqnarray*} \hfill$\square$\par\vskip24pt The Lemma immediately yields an upper bound on (\ref{eq:maximand}): letting $$ \sigma:=\max(\beta\mathbbm{1},-\Delta) + \eta \ket{\psi}\bra{\psi}, $$ such that we obtain \begin{eqnarray} S(\Delta+\sigma \,||\,\sigma) &\le& S(\Delta+\max(\beta\mathbbm{1},-\Delta)\,||\, \max(\beta\mathbbm{1},-\Delta)) \nonumber \\ &=& S((\Delta+\beta\mathbbm{1})_+ \,||\, \beta\mathbbm{1} + (\Delta+\beta\mathbbm{1})_-).\label{eq:s11} \end{eqnarray} To continue, we consider two cases. {\it Case 1 :} When $T\le\beta$, the requirement (\ref{eq:delta3}) is automatically satisfied, and $\max(\beta\mathbbm{1},-\Delta)=\beta\mathbbm{1}$. Let $\Delta_+$ and $\Delta_-$ be the positive and negative part of $\Delta$, respectively. That is, $\Delta = \Delta_+ - \Delta_-$, with $\Delta_+$ and $\Delta_-$ non-negative and orthogonal. Because we are using the trace norm we can rewrite the conditions on $\Delta$ as \begin{eqnarray*} ||\Delta||_1 &=& \mathop{\rm Tr}\nolimits[\Delta_+] + \mathop{\rm Tr}\nolimits[\Delta_-] = 2T ,\\ \mathop{\rm Tr}\nolimits[\Delta] &=& \mathop{\rm Tr}\nolimits[\Delta_+ ]- \mathop{\rm Tr}\nolimits[\Delta_- ]= 0, \end{eqnarray*} hence $$ \mathop{\rm Tr}\nolimits[\Delta_+ ]= \mathop{\rm Tr}\nolimits[\Delta_-]= T. $$ By Lemma \ref{lem:sabc}, (\ref{eq:maximand}) is upper bounded by $S(\Delta+\beta\mathbbm{1}||\beta\mathbbm{1})$. By convexity, its maximum over $\Delta_+$, $\Delta_-\ge0$, with $\mathop{\rm Tr}\nolimits[\Delta_+ ]= \mathop{\rm Tr}\nolimits[\Delta_-] = T$, is obtained in $\Delta_+$ and $\Delta_-$ of rank 1, giving as upper bound $$ S(\Delta + \sigma||\sigma) \leq (\beta+T)\log\frac{\beta+T}{\beta} +(\beta-T)\log\frac{\beta-T}{\beta}. $$ The upper bound can be achieved in dimensions $d\ge3$ for all values of $T\le\beta$ by setting $\Delta=TF$ and $\psi=e^3$. {\it Case 2:} In the other case, when $T>\beta$, we have to deal with condition (\ref{eq:delta3}). To do that we split $\Delta$ into three non-negative parts, \begin{equation}\nonumber \Delta = \Delta_+ -\Delta_0-\Delta_-, \end{equation} with $\Delta_+$, $\Delta_0$ and $\Delta_-$, operating on orthogonal subspaces $V_+$, $V_0$ and $V_-$, respectively, with \begin{eqnarray*} \phantom{-}\Delta_+ &\ge& 0, \\ 0 &\ge& -\Delta_0\, \ge -\beta\mathbbm{1}_0, \\ -\beta\mathbbm{1}_- &\ge& -\Delta_-. \end{eqnarray*} We denote the projectors on these subspaces by $\mathbbm{1}_+$, $\mathbbm{1}_0$, and $\mathbbm{1}_-$. Then $$ (\Delta+\beta)_+ = \Delta_+ - \Delta_0 + \beta\mathbbm{1}_{+0}, $$ where $\mathbbm{1}_{+0}:=\mathbbm{1}_++\mathbbm{1}_0$. The conditions on $\Delta$, $\mathop{\rm Tr}\nolimits[\Delta]=0$ and $\mathop{\rm Tr}\nolimits[ |\Delta| ]=2T$ translate to $$ \mathop{\rm Tr}\nolimits[\Delta_+ ]= \mathop{\rm Tr}\nolimits[\Delta_0]+\mathop{\rm Tr}\nolimits[\Delta_- ]= T. $$ Due to the orthogonality of positive and negative part, (\ref{eq:s11}) can be simplified to $ S((\Delta+\beta\mathbbm{1})_+ \,||\, \beta\mathbbm{1}_{+0})$. After subtracting $\beta\mathbbm{1}_0-\Delta_0$ from both arguments, we get $$ S(\Delta_++\beta\mathbbm{1}_+ || \beta\mathbbm{1}_++\Delta_0), $$ which is an upper bound on (\ref{eq:s11}), by Lemma \ref{lem:sabc}. Ignoring condition (\ref{eq:delta3}) on $\Delta$, we get $$ S_{\max} \le \max_{\Delta_+\ge 0 \atop \mathop{\rm Tr}\nolimits[\Delta_+ ]= T} S(\Delta_++\beta\mathbbm{1}_+ || \beta\mathbbm{1}_+). $$ By convexity, the maximum is obtained for $\Delta_+$ rank 1, giving the upper bound $$ S_{\max}\le(T+\beta)\log((T+\beta)/\beta). $$ To see that this bound is sharp for (almost) any value of $T$, consider the two states \begin{eqnarray*} \rho &=& \mathop{\rm Diag}\nolimits(T+\beta,0,0^{\times J},\beta^{\times K},\beta+\eta) , \\ \sigma &=& \mathop{\rm Diag}\nolimits(\beta,T-J\beta,\beta^{\times J},\beta^{\times K},\beta+\eta) ,\\ \eta &:=& 1-T-(d-1-J)\beta. \end{eqnarray*} Here, $J$ is an integer between 0 and $d-3$ and $k=d-3-J$. Conditions on $J$ are $J\beta\le T$ (so that $\sigma\ge0$) and $T\le 1-(d-1-J)\beta$ (so that $\eta\ge0$). This choice of states can thus be obtained for $\beta\le T\le 1-2\beta$. It can be seen that $||\rho-\sigma||_1=2T$ and \begin{equation}\nonumber S(\rho||\sigma)=(T+\beta)\log((T+\beta)/\beta). \end{equation} The result of the foregoing can be subsumed into the following theorem. \begin{theorem} Let $\Delta=\rho-\sigma$, $T=||\Delta||_1/2$ and $\beta=\lambda_{\min}(\sigma)$. If $T \le \beta$ then \begin{equation}\nonumber S(\rho||\sigma) \le (\beta+T)\log\frac{\beta+T}{\beta} +(\beta-T)\log\frac{\beta-T}{\beta}, \label{bound_quad2_bis} \end{equation} and this upper bound is sharp when $d>2$. If $\beta\le T \le 1-\beta$ then \begin{equation}\nonumber S(\rho||\sigma) \le (\beta+T)\log\frac{\beta+T}{\beta}. \label{bound_quad3} \end{equation} When $d>2$, this bound is sharp for (at least) $\beta\le T\le 1-2\beta$. \end{theorem} Figure \ref{fig3} illustrates these bounds and shows their superiority to the previously obtained bound (\ref{bound_log}). \begin{figure} \includegraphics[width=3.4in]{ud3b2.eps} \caption{\label{fig3}Comparison between upper bounds (\ref{bound_log}) and (\ref{bound_quad2_bis})-(\ref{bound_quad3}) on $S=S(\rho||\sigma)$ vs.\ the trace norm distance $T=||\rho-\sigma||_1/2$, for various values of $\beta$, the smallest eigenvalue of $\sigma$. The upper set of dashed curves depict bound (\ref{bound_log}) (with $d=3$) for $\beta=0.1$ (lower curve), $0.2$, $0.3$, $0.4$ and $0.5$ (upper curve). The lower set of full line curves depict bounds (\ref{bound_quad2_bis})-(\ref{bound_quad3}) for $\beta=0.1$ (upper curve), $0.2$, $0.3$, $0.4$ and $0.5$ (lower curve). The two regimes $T\le\beta$ and $\beta\le T\le 1-\beta$ can be clearly seen. } \end{figure} Again, it is interesting to look at the bound for large $\beta$. An approximation for bound (\ref{bound_quad2_bis}), valid for $T \ll \beta$, is given by \begin{eqnarray} S(\rho||\sigma) \le \sum_{k=1}^\infty \frac{T^{2k}}{k(2k-1) \beta^{2k-1}} \approx \frac{T^2}{\beta}. \label{bound_quad2_bis_approx} \end{eqnarray} \section{Application to state approximation} In the following paragraph we will give an application of our bounds to state approximation. Consider a state $\rho$ on a Hilbert space ${\cal H}$, and a sequence $\{\sigma_n\}_n$ where $\sigma_n$ is a state on ${\cal H}^{\otimes n}$. As before, the sequence is said to asymptotically approximate $\rho$ if for $n$ tending to infinity, $ \| \sigma_n-\rho^{\otimes n}\|_1= \mathop{\rm Tr}\nolimits|\sigma_n-\rho^{\otimes n}|$ tends to zero. Let us define $T_n$ as $$ T_n := \mathop{\rm Tr}\nolimits|\rho^{\otimes n}-\sigma_n|/2. $$ Because of the lower bound (\ref{bound_ohya}), we get $$ S_n:=S(\rho^{\otimes n}||\sigma_n) \ge 2 T_n^2, $$ and this bound is sharp. Hence, $T_n$ goes to zero if $S_n$ does. On the other hand, $T_n$ going to zero does not necessarily imply $S_n$ going to zero. Indeed, $S_n$ can be infinite for any finite value of $n$ when $\rho^{\otimes n}$ is not restricted to the range of $\sigma_n$. In particular, the relative entropy distance between two pure states is infinite unless the states are identical. At first sight, this seems to render the relative entropy useless as a distance measure. Nevertheless, sense can be made of it by imposing an additional requirement that the range of $\sigma_n$ must contain the range of $\rho^{\otimes n}$. Let us then restrict $\sigma_n$ to the range of $\rho^{\otimes n}$, as the relative entropy only depends on that part of $\sigma_n$. Letting $d$ be the rank of $\rho$, the dimension of the range of $\rho^{\otimes n}$ is $d^n$. Let $\beta_n$ be the smallest non-zero eigenvalue of $\sigma_n$ on that range; $\beta_n$ is at most $1/d^n$. The behaviour of the relative entropy then very much depends on the relation between $\beta_n$ and $T_n$. Since $\beta_n$ decreases at least exponentially, we only need to consider the case $T_n\ge\beta_n$, and use the bound (\ref{bound_quad3}) $$ S_n \le (\beta_n+T_n)\log \biggl (1+\frac{T_n}{\beta_n}\biggr). $$ In the worst-case behaviour of $T_n$ ($T_n/\beta_n$ tending to infinity) the bound can be approximated by \begin{eqnarray*} S_n & \le & T_n\log\frac{T_n}{\beta_n} = T_n(\log T_n-\log\beta_n)\nonumber\\ & \approx & T_n|\log\beta_n|. \end{eqnarray*} To guarantee convergence of $S_n$ we therefore need $T_n$ to converge to 0 at least as fast as $1/|\log\beta_n|$, which in the best case goes as $1/n$. Note that bound (\ref{bound_log}) yields the same requirement, but as this bound is not a sharp one it could have been too strong a requirement. This gives us the subsequent theorem. \begin{theorem} Consider a state $\rho$ on a finite-dimensional Hilbert space ${\cal H}$ and a sequence $\{\sigma_n\}_n$ of states $\sigma_n$ on ${\cal H}^{\otimes n}$. The sequence $\{\sigma_n\}_n$ asymptotically approximates $\rho$ in the trace norm, if \begin{equation}\nonumber \lim_{n\rightarrow\infty} S(\rho^{\otimes n}||\sigma_n)=0. \end{equation} Conversely, if the range of $\sigma_n$ includes the range of $\rho^{\otimes n}$ and $||\rho^{\otimes n}-\sigma_n||_1$ converges to zero faster than $1/|\log\beta_n|$, where $\beta_n$ is the minimal eigenvalue of $\sigma_n$ restricted to the range of $\rho^{\otimes n}$, then $\lim_{n\rightarrow\infty} S(\rho^{\otimes n}||\sigma_n)=0$. \end{theorem} \section{Summary} In this paper, we have discussed several lower and upper bounds on the relative entropy functional, thereby sharpening the notion of continuity of the relative entropy for states which are close to each other in the trace norm sense. The main results are the sharp lower bound from Theorem 4, and the sharp upper bounds of Theorems 5 ($d=2$) and 6 ($d>2$). Theorems 4 and 5 give the relation between relative entropy and norm distances based on any unitarily invariant norm, while Theorem 6 holds only for the trace norm distance. These results have been obtained employing methods from optimisation theory. \begin{acknowledgments} This work was supported by the Alexander-von-Humboldt Foundation, the European Commission (EQUIP, IST-1999-11053), the DFG (Schwerpunktprogramm QIV), the EPSRC QIP-IRC, and the European Research Councils (EURYI). \end{acknowledgments}
{ "timestamp": "2005-10-14T18:15:20", "yymm": "0503", "arxiv_id": "quant-ph/0503218", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503218" }
\subsection*{Isospectral billiards.} All known isospectral billiards can be obtained by unfolding triangle-shaped tiles \cite{BusConDoySem94, OkaShu01}. The way the tiles are unfolded can be specified by three permutation matrices $M^{(\mu)}$, $1\leq \mu\leq 3$, associated to the three sides of the triangle: $M^{(\mu)}_{ij}=1$ if tiles $i$ and $j$ are glued by their side $\mu$ (and $M^{(\mu)}_{ii}=1$ if the side $\mu$ of tile $i$ is the boundary of the billiard), and 0 otherwise \cite{OkaShu01, Tha04, Gir04}. Following \cite{OkaShu01}, one can sum up the action of the $M^{(\mu)}$ in a graph with coloured edges: each copy of the base tile is associated to a vertex, and vertices $i$ and $j$, $i\neq j$, are linked by an edge of colour $\mu$ if and only if $M^{(\mu)}_{ij}=1$ (see \ref{graphs}). \begin{figure}[ht] \begin{center} \includegraphics[width=0.66\linewidth]{fig2.eps} \end{center} \caption{The graphs corresponding to a pair of isospectral billiards: if we label the sides of the triangle by $\mu=1,2,3$, the unfolding rule by symmetry with respect to side $\mu$ can be represented by edges made of $\mu$ braids in the graph. From a given pair of graphs, one can construct infinitely many pairs of isospectral billiards by applying the unfolding rules to any triangle. Note that a different labeling of the tiles would just induce a permutation of the labelings of points and blocks in the Fano plane.} \label{graphs} \end{figure} In the same way, in the second member of the pair, the tiles are unfolded according to permutation matrices $N^{(\mu)}$, $1\leq \mu\leq 3$. Two billiards are said to be transplantable if there exists an invertible matrix $T$ (the {\it transplantation matrix}) such that $\forall\mu\ \ T M^{(\mu)}= N^{(\mu)} T$. One can show that transplantability implies isospectrality (if the matrix $T$ is not merely a permutation matrix, in which case the two domains would just have the same shape). The underlying idea is that if $\psi^{(1)}$ is an eigenfunction of the first billiard and $\psi^{(1)}_i$ its restriction to triangle $i$, then one can build an eigenfunction $\psi^{(2)}$ of the second billiard by taking $\psi^{(2)}_i=\sum_j T_{ij}\psi^{(1)}_j$. Obviously $\psi^{(2)}$ verifies Schr\"odinger equation; it can be checked from the commutation relations that the function is smooth at all edges of the triangles, and that boundary conditions at the boundary of the billiard are fulfilled \cite{OkaShu01}.\\ Suppose we want to construct a pair of isospectral billiards, starting from any polygonal base shape. Our idea is to start from the transplantation matrix, and choose it in such a way that the existence of commutation relations $T M^{(\mu)}= N^{(\mu)} T$ for some permutation matrices $M^{(\mu)}, N^{(\mu)}$ will be known {\it a priori}. As we will see, this is the case if $T$ is taken to be the incidence matrix of a FPS; the matrices $M^{(\mu)}$ and $N^{(\mu)}$ are then permutations on the points and the hyperplanes of the FPS. \subsection*{Finite projective spaces.} For $n\geq 2$ and $q=p^h$ a power of a prime number, consider the $(n+1)$-dimensional vector space $\mathbb{F}_{q}^{n+1}$, where $\mathbb{F}_{q}$ is the finite field of order $q$. The {\it finite projective space} $PG(n,q)$ of {\it dimension} $n$ and {\it order} $q$ is the set of subspaces of $\mathbb{F}_{q}^{n+1}$: the points of $PG(n,q)$ are the 1-dimensional subspaces of $\mathbb{F}_{q}^{n+1}$, the lines of $PG(n,q)$ are the 2-dimensional subspaces of $\mathbb{F}_{q}^{n+1}$, and more generally $(d+1)$-dimensional subspaces of $\mathbb{F}_{q}^{n+1}$ are called {\it $d$-spaces} of $PG(n,q)$; the $(n-1)$-spaces of $PG(n,q)$ are called hyperplanes or {\it blocks}. A $d$-space of $PG(n,q)$ contains $(q^{d+1}-1)/(q-1)$ points. In particular, $PG(n,q)$ has $(q^{n+1}-1)/(q-1)$ points. It also has $(q^{n+1}-1)/(q-1)$ blocks \cite{Hir79}. As an example, \ref{fano} shows the finite projective plane (FPP) of order $q=2$, or Fano plane, $PG(2,2)$. \begin{figure}[ht] \begin{center} \includegraphics[width=0.58\linewidth]{fig3.eps} \caption{The Fano plane $PG(2,2)$ and its corresponding incidence matrix $T$. The Fano plane has $(q^3-1)/(q-1)=7$ points and 7 lines; each line contains $q+1=3$ points and each point belongs to 3 lines. Any pair of points belongs to one and only one line.} \label{fano} \end{center} \end{figure} A $(N,k, \lambda)-${\it symmetric balanced incomplete block design} (SBIBD) is a set of $N$ points, belonging to $N$ subsets (or {\it blocks}); each block contains $k$ points, in such a way that any two points belong to exactly $\lambda$ blocks, and each point is contained in $k$ different blocks \cite{DinSti92}. One can show that $PG(n,q)$ is a $(N,k, \lambda)$-SBIBD with $N=(q^{n+1}-1)/(q-1)$, $k=(q^{n}-1)/(q-1)$ and $\lambda=(q^{n-1}-1)/(q-1))$. For example, the Fano plane is a $(7,3,1)-$SBIBD. The points and the blocks can be labeled from $0$ to $N-1$. For any $(N,k, \lambda)-$SBIBD one can define an $N\times N$ {\it incidence matrix} $T$ describing to which block each point belongs. The entries $T_{ij}$ of the matrix are $T_{ij}=1$ if point $j$ belongs to line $i$, $0$ otherwise. The matrix $T$ verifies the relation $T T^{t}=\lambda J+(N-k)\lambda/(k-1)I$, where $J$ is the $N\times N$ matrix with all entries equal to 1 and $I$ the $N\times N$ identity matrix \cite{DinSti92}. In particular, the incidence matrix of $PG(n,q)$ verifies \begin{equation} \label{tt} T T^{t}=\lambda J+(k-\lambda)I \end{equation} with $k$ and $\lambda$ as given above. For example, the incidence matrix of the Fano plane given in \ref{fano} corresponds to a labeling of the lines such that line 0 contains points 0,1,3, and line 1 contains points 1,2,4, etc. A {\it collineation} of a FPS is a bijection that preserves incidence, that is a permutation of the points that takes $d$-spaces to $d$-spaces (in particular, it takes blocks to blocks). Any permutation $\sigma$ on the points can be written as a $N\times N$ {\it permutation matrix} $M$ defined by $M_{i\sigma(i)}=1$ and the other entries equal to zero. If $M$ is a permutation matrix associated to a collineation, then there exists a permutation matrix $N$ such that \begin{equation} \label{commutation} TM=NT. \end{equation} In words, (\ref{commutation}) means that permuting the columns of $T$ (i.e. the blocks of the space) under $M$ is equivalent to permuting the rows of $T$ (i.e. the points of the space) under $N$. The commutation relation (\ref{commutation}) is a related to an important feature of projective geometry, the so-called ''principle of duality'' \cite{Hir79}. This principle states that for any theorem which is true in a FPS, the dual theorem obtained by exchanging the words ''point'' and ''block'' is also true. As we will see now, this symmetry between points and blocks in FPSs is the central reason which accounts for known pairs of isospectral billiards.\\ Let us consider a FPS ${\mathcal P}$ with incidence matrix $T$. To each block in ${\mathcal P}$ we associate a tile in the first billiard, and to each point in ${\mathcal P}$ we associate a tile in the second billiard. If we choose permutations $M^{(\mu)}$ in the collineation group of ${\mathcal P}$, then the commutation relation (\ref{commutation}) will ensure that there exist permutations $N^{(\mu)}$ verifying $TM^{(\mu)}=N^{(\mu)}T$. These commutation relations imply transplantability, and thus isospectrality, of the billiards constructed from the graphs corresponding to $M^{(\mu)}$ and $N^{(\mu)}$. If the base tile has $r$ sides, we need to choose $r$ elements $M^{(\mu)}$, $1\leq\mu\leq r$, in the collineation group of ${\mathcal P}$. This choice is constrained by several factors. Since $M^{(\mu)}$ represents the reflexion of a tile with respect to one of its sides, it has to be of order 2 (i.e. an involution). In order that the billiards be connected, no point should be left out by the matrices $M^{(\mu)}$ (in other words, the graph associated to the matrices $M^{(\mu)}$ should be connected). Finally, if we want the base tile to be of any shape, there should be no loop in the graph. We now need to characterize collineations of order 2. \subsection*{Collineations of finite projective spaces.} Let $q=p^h$ be a power of a prime number. Each point $P$ of $PG(n,q)$ is a 1-dimensional subspace of $\mathbb{F}_{q}^{n+1}$, spanned by some vector $v$. We write $P=P(v)$. An {\it automorphism} is a bijection of the points $P(v_i)$ of $PG(n,q)$ obtained by the action of an automorphism of $\mathbb{F}_{q}$ on the coordinates of the $v_i$. If $q=p^h$, the automorphisms of $\mathbb{F}_{q}$ are $t\mapsto t^{p^i}$, $0\leq i \leq h-1$. A {\it projectivity} is a bijection of the points $P(v_i)$ of $PG(n,q)$ obtained by the action of a linear map $L$ on the $v_i$. There are $q-1$ matrices $t L$, with $t\in\mathbb{F}_{q}\setminus\{0\}$ and $L\in GL_{n+1}(\mathbb{F}_{q})$ (the group of $(n+1)\times (n+1)$ invertible matrices with coefficients in $\mathbb{F}_{q}$), yielding the same projectivity $P(v_i)\mapsto P(L v_i)$. The {\it Fundamental theorem of projective geometry } states that any collineation of $PG(n,q)$ can be written as the composition of a projectivity by an automorphism \cite{Hir79}. The converse is obviously true since projectivities and automorphisms are collineations. Therefore the set of all collineations is obtained by taking the composition of all the non-singular $(n+1)\times(n+1)$ matrices with coefficients in $\mathbb{F}_{q}$ by all the $h$ automorphisms of $\mathbb{F}_{q}$. The collineation group of $PG(n, q)$ has $[h\prod_{k=0}^{n}(q^{n+1}-q^k)]/(q-1)$ elements, among which we only want to consider elements of order 2. In the case of FPPs ($n=2$), there are various known properties characterizing collineations of order 2. A {\it central collineation}, or {\it perspectivity}, is a collineation fixing each line through a point $C$ (called the centre). By ''fixing'' we mean that the line is invariant but the points can be permuted within the line. One can show that the fixed points of a non-identical perspectivity are the centre itself and all points on a line (called the axis), while the fixed lines are the axis and all lines through the centre. If the centre lies off the axis a perspectivity is called a {\it homology} (and has $q+2$ fixed points), whereas if the centre lies on the axis it is called an {\it elation} (and has $q+1$ fixed points). Perspectivities in dimension $n=2$ have following properties \cite{Bon04}: {\scshape Proposition 1.} A perspectivity of order two of a FPP of order $q$ is an elation or a homology according to whether $q$ is even or odd. {\scshape Proposition 2.} A collineation of order two of a FPP of order $q$ is a perspectivity if $q$ is not a square; it is a collineation fixing all points and lines in a subplane if $q$ is a square (a subplane is a subset of points having all the properties of a FPP). When the order of the FPP is a square, there is the following result \cite{Hir79}: {\scshape Proposition 3.} $PG(2, q^2)$ can be partitioned into $q^2-q+1$ subplanes $PG(2, q)$. \subsection*{Generating isospectral pairs.} Let us assume we are looking for a pair of isospectral billiards with $N=(q^3-1)/(q-1)$ copies of a base tile having the shape of a $r$-gon, $r\geq 3$. We need to find $r$ collineations of order 2 such that the associated graph is connected and without loop. Such a graph connects $N$ vertices and thus requires $N-1$ edges. From propositions 1-3, we can deduce the number $s$ of fixed points of a collineation of order 2 for any FPP. Since a collineation is a permutation, it has a cycle decomposition as a product of transpositions. For permutations of order 2 with $s$ fixed points, there are $(N-s)/2$ independent transpositions in this decomposition. Each transposition is represented by an edge in the graph. As a consequence, $q$, $r$ and $s$ have to fulfill the following condition: $r(q^2+q+1-s)/2=q^2+q$. Let us examine the various cases. {\it If $q$ is even and not a square}, propositions 1 and 2 imply that any collineation of order 2 is an elation and therefore has $q+1$ fixed points. Therefore, $q$ and $r$ are constrained by the relation $r q^2/2=q^2+q$. The only integer solution with $r\geq 3$ and $q\geq 2$ is $(r=3, q=2)$. These isospectral billiards correspond to $PG(2,2)$ and will be made of $N=7$ copies of a base triangle. {\it If $q$ is odd and not a square}, propositions 1 and 2 imply that any collineation of order 2 is a homology and therefore has $q+2$ fixed points. Therefore, $q$ and $r$ are constrained by the relation $r (q^2-1)/2=q^2+q$. The only integer solution with $r\geq 3$ and $q\geq 2$ is $(r=3, q=3)$. These isospectral billiards correspond to $PG(2,3)$ and will be made of $N=13$ copies of a base triangle. {\it If $q=p^2$ is a square}, propositions 2 and 3 imply that any collineation of order 2 fixes all points in a subplane $PG(2,p)$ and therefore has $p^2+p+1$ fixed points. Therefore, $p$ and $r$ are constrained by the relation $r (p^4-p)/2=p^4+p^2$. There is no integer solution with $r\geq 3$ and $q\geq 2$. However, one can look for isospectral billiards with loops: this will require the base tile to have a shape such that the loop does not make the copies of the tiles come on top of each other when unfolded. If we tolerate one loop in the graph describing the isospectral billiards, then there are $N$ edges in the graph instead of $N-1$ and the equation for $p$ and $r$ becomes $r (p^4-p)/2=p^4+p^2+1$, which has the only integer solution $(r=3, p=2)$. These isospectral billiards correspond to $PG(2,4)$ and will be made of $N=21$ copies of a base triangle. We can now generate all possible pairs of isospectral billiards whose transplantation matrix is the incidence matrix of $PG(2,q)$, with $r$ and $q$ restricted by the previous analysis. All pairs must have a triangular base shape ($r=3$). $PG(2, 2)$ provides 3 pairs (made of 7 tiles), $PG(2, 3)$ provides 9 pairs (made of 13 tiles), $PG(2, 4)$ provides 1 pair (made of 21 tiles). It turns out that the pairs obtained here are exactly those obtained by \cite{BusConDoySem94} and \cite{OkaShu01} by other methods. Let us now consider spaces $PG(n,q)$ of higher dimensions. The smallest one is $PG(3,2)$, which contains 15 points. Since the base field for $PG(3,2)$ is $\mathbb{F}_{2}$, the Fundamental Theorem of projective geometry states that the collineation group of $PG(3,2)$ is essentially the group $GL_4(\mathbb{F}_{2})$ of $4\times 4$ non-singular matrices with coefficients in $\mathbb{F}_{2}$. Generating all possible graphs from the 316 elements of order 2 in $GL_4(\mathbb{F}_{2})$, we obtain four pairs of isospectral billiards with 15 triangular tiles, which completes the list of all pairs found in \cite{BusConDoySem94} and \cite{OkaShu01}.\\ Our method explicitly gives the transplantation matrix $T$ for all these pairs: each one is the incidence matrix of a FPS. The transplantation matrix explicitly provides the mapping between eigenfunctions of both billiards. The inverse mapping is given by $T^{-1}=(1/q^{n-1})(T^{t}-(\lambda/k)J)$. For all pairs, isospectrality can therefore be explained by the symmetry between points and blocks in FPSs. We do not know if it is possible to find isospectral billiards for which isospectrality would not rely on this symmetry. Our construction furthermore allows to generalize the results we obtained in \cite{Gir04}, where a relation between the Green functions of the billiards in an isospectral pair was derived. A similar relation can be found for all other pairs obtained by point-block duality. Let $M^{(\mu)}$ and $N^{(\mu)}$ be the matrices describing the gluings of the tiles. To any path $p$ going from a point to another on the first billiard, one can associate the sequence $(\mu_1,...,\mu_n)$, $1\leq\mu_i\leq 3$, of edges of the triangle hit by the path. The matrix $M=\prod M^{(\mu_i)}$ is such that $M_{ij}=1$ if path $p$ can be drawn from $i$ to $j$. If $N=T^{-1}MT$, relations (\ref{tt}) and (\ref{commutation}) imply $\sum_{k,l}T_{ik}T_{jl}M_{kl}=\lambda+(k-\lambda)N_{ij}$. Since Green functions can be written as a sum over all paths, the relation between the Green functions $G^{(2)}(a,i;b,j)$ and $G^{(1)}(a,i';b,j')$ is $\sum_{i',j'}T_{i,i'}T_{j,j'}G^{(1)}(a,i';b,j')= (k-\lambda)G^{(2)}(a,i;b,j)+\lambda G^{t}(a;b)$, where $G^{t}$ is the Green function of the triangle, and a point $(a,i)$ is specified by a tile number $i$ and its position $a$ inside the tile (see \cite{Gir04} for further detail). More precisely, the term $\sum_{k,l}T_{ik}T_{jl}M_{kl}$ can be interpreted as the number of pairs of tiles $(i,j)$ in the first billiard such that a path identical to $p$ can go from $i$ to $j$. A given diffractive orbit going from $i$ to $j$ in the second billiard corresponds to a matrix $N$ such that $N_{ij}=1$: it is therefore constructed from a superposition of $k=(q^n-1)/(q-1)$ identical diffractive orbits in the first billiard. (Note that these results correspond to Neumann boundary conditions. It is easy to obtain similar relations for Dirichlet boundary conditions by conjugating all matrices with a diagonal matrix $D$ with entries $D_{ii}=\pm 1$ according to whether tile $i$ is like the initial tile or like its mirror inverse.)
{ "timestamp": "2005-03-31T15:25:36", "yymm": "0503", "arxiv_id": "nlin/0503069", "language": "en", "url": "https://arxiv.org/abs/nlin/0503069" }
\section{Introduction} Nematic liquid crystals (nematics) represent the simplest anisotropic fluid. The description of the dynamic behavior of the nematics is based on well established equations. The description is valid for low molecular weight materials as well as nematic polymers. The coupling between the preferred molecular orientation (director $\Vec{\Hat{n}}$) and the velocity field leads to interesting flow phenomena. The orientational dynamics of nematics in flow strongly depends on the sign of the ratio of the Leslie viscosity coefficients $\lambda = \alpha_3 / \alpha_2$. In typical low molecular weight nematics $\lambda$ is positive ({\em flow-aligning materials}). The case of the initial director orientation perpendicular to the flow plane has been clarified in classical experiments by Pieranski and Guyon \cite{Pieranski:SSC:1973, Pieranski:PRA:1974} and theoretical works of Dubois-Violette and Manneville (for an overview see \cite{PF:Ch4}). An additional external magnetic field could be applied along the initial director orientation. In Couette flow and low magnetic field there is a homogeneous instability \cite{Pieranski:SSC:1973}. For high magnetic field the homogeneous instability is replaced by a spatially periodic one leading to rolls \cite{Pieranski:PRA:1974}. In Poiseuille flow, as in Couette flow, the homogeneous instability is replaced by a spatially periodic one with increasing magnetic field \cite{Manneville:JPh:1979}. All these instabilities are stationary. Some nematics (in particular near a nematic-smectic transition) have negative $\lambda$ ({\em non-flow-aligning materials}). For these materials in steady flow and in the geometry where the initial director orientation is perpendicular to the flow plane only spatially periodic instabilities are expected \cite{Pieranski:CPh}. These materials demonstrate also tumbling motion \cite{Cladis:PRL:1975} in the geometry where the initial director orientation is perpendicular to the confined plates that make the orientational behavior quite complicated. Most previous theoretical investigations of the orientational dynamics of nematics in shear flow were carried out under the assumption of strong anchoring of the nematic molecules at the confining plates. However, it is known that there is substantial influence of the boundary conditions on the dynamical properties of nematics in hydrodynamic flow \cite{Kedney:LC:V24:P613:Y1998, Nasibullayev:MCLC:V351:P395:Y2000, Tarasov:LC:V28:N6:P833:Y2001,Nasibullayev:CR:2001}. Indeed, the anchoring strength strongly influences the orientational behavior and dynamic response of nematics under external electric and magnetic fields. This changes, for example, the switching times in bistable nematic cells \cite{Kedney:LC:V24:P613:Y1998}, which play an important role in applications \cite{Chigrinov:1999}. Recently the influence of the boundary anchoring on the homogeneous instabilities in steady flow was investigated theoretically \cite{Tarasov:LC:V28:N6:P833:Y2001}. In this paper we study the combined action of steady flow (Couette and Poiseuille) and external fields (electric and magnetic) on the orientational instabilities of the nematics with initial orientation perpendicular to the flow plane. We focus on {\em flow-aligning} nematics. The external electric field is applied across the nematic layer and the external magnetic field is applied perpendicular to the flow plane. We analyse the influence of weak azimuthal and polar anchoring and of external fields on both homogeneous and spatially periodic instabilities. In section II the formulation of the problem based on the standard set of Ericksen-Leslie hydrodynamic equations \cite{Leslie:MCLC:1976} is presented. Boundary conditions and the critical Fre\'edericksz field in case of weak anchoring are discussed. In section III equations for the homogeneous instabilities are presented. Rigorous semi-analytical expressions for the critical shear rate $a_c^2$ for Couette flow (section III A), the numerical scheme for finding $a_c^2$ for Poiseuille flow (section III B) and approximate analytical expressions for both types of flows (section III C) are presented. In section IV the analysis of the spatially periodic instabilities is given and in section V we discuss the results. In particular we will be interested in the boundaries in parameter space (anchoring strengths, external fields) for the occurrence of the different types of instabilities. \section{Basic equations} \begin{figure} \epsfig{file=cell.eps,width=8cm} \caption{Geometry of NLC cell ($a$). Couette ($b$) and Poiseuille ($c$) flows.} \label{fig:cell} \end{figure} Consider a nematic layer of thickness $d$ sandwiched between two infinite parallel plates that provide weak anchoring (Fig. \ref{fig:cell} $a$). The origin of the Cartesian coordinates is placed in the middle of the layer with the $z$ axis perpendicular to the confining plates ($z = \pm d / 2$ for the upper/lower plate). The flow is applied along $x$. Steady Couette flow is induced by moving the upper plate with a constant speed (Fig. \ref{fig:cell} $b$). Steady Poiseuille flow is induced by applying a constant pressure difference along $x$ (Fig. \ref{fig:cell} $c$). An external electric field $E_0$ is applied along $z$ and a magnetic field $H_0$ along $y$. The nematodynamic equations have the following form \cite{deGennes} \begin{eqnarray} \label{eqn:EL:velocity} &&\rho (\partial_t + \Vec{v} \cdot \Vec{\nabla}) v_i = - p_{,i} + [T^v_{ji} + T^e_{ji}]_{,j},\\ \label{eqn:EL:director} &&\gamma_1 \Vec{N} = - (1 - \Vec{n} \Vec{n} \cdot) (\gamma_2 A \cdot \Vec{n} + \Vec{h}), \end{eqnarray} where $\rho$ is the density of the NLC and $p_{,i} = \Delta P / \Delta x$ the pressure gradient; $\gamma_1 = \alpha_3 - \alpha_2$ and $\gamma_2 = \alpha_3 + \alpha_2$ are rotational viscosities; $\Vec{N} = \Vec{n}_{,t} + \Vec{v} \cdot \Vec{\nabla} \Vec{n} - (\nabla \times \Vec{v}) \times \Vec{n}/ 2$ and $A_{ij} = (v_{i,j} + v_{j,i}) / 2$, $h_i = \delta F / \delta n_i$. The notation $f_{,i}\equiv \partial_i f$ is used throughout. The viscous tensor $T^v_{ij}$ and elastic tensor $T^e_{ij}$ are \begin{eqnarray} &T^v_{ij} = &\alpha_1 n_i n_j A_{km} n_k n_m + \alpha_2 n_i N_j + \alpha_3 n_j N_i \nonumber\\ && + \alpha_4 A_{ij} + \alpha_5 n_i n_k A_{ki} + \alpha_6 A_{ik} n_k n_j,\\ &T^e_{ij} = & - \frac{\partial F}{\partial n_{k,i}} n_{k,j}, \end{eqnarray} where $\alpha_i$ are the Leslie viscosity coefficients. The bulk free energy density $F$ is \begin{eqnarray} &F = & \frac12 \Bigl\{ K_{11} (\nabla \cdot \Vec{n})^2 + K_{22} [\Vec{n} \cdot (\nabla \times \Vec{n})]^2 \nonumber\\ && + K_{33} [\Vec{n} \times (\nabla \times \Vec{n})]^2 - \varepsilon_0 \varepsilon_a (\Vec{n} \cdot \Vec{E}_0)^2 \\ && - \mu_0 \chi_a (\Vec{n} \cdot \Vec{H}_0)^2 \Bigr\}.\nonumber \end{eqnarray} Here $K_{ii}$ are the elastic constants, $\varepsilon_a$ the anisotropy of the dielectric permittivity and $\chi_a$ is the anisotropy of the magnetic susceptibility. In addition one has the normalization equation \begin{equation} \label{eqn:normalization} \Vec{n} = 1 \end{equation} and incompressibility condition \begin{equation} \label{eqn:incompressibility} \nabla \cdot \Vec{v} = 0. \end{equation} The basic state solution of equations \eqref{eqn:EL:velocity} and \eqref{eqn:EL:director} has the following form \begin{equation} \label{eqn:base} \Vec{n}_0=(0,\:1,\:0),\:\Vec{v}_0=(v_{0x}(z),\:0,\:0), \end{equation} where $v_{0x}=V_0(1/2+z/d)$ for Couette and $v_{0x} = (\Delta P/\Delta x)[d^2 / \alpha_4][1 / 4 - z^2 / d^2]$ for Poiseuille flow. In order to investigate the stability of the basic state \eqref{eqn:base} with respect to small perturbations we write: \begin{equation} \label{def:hom} \Vec{n}=\Vec{n}_0+\Vec{n}_1(z) e^{\sigma t} e^{i q y},\: \Vec{v}=\Vec{v}_0+\Vec{v}_1(z) e^{\sigma t} e^{i q y}; \end{equation} We do not expect spatial variation along $x$ for steady flow. The case $q = 0$ corresponds to a homogeneous instability. Here we analyse stationary bifurcations, thus the threshold condition is $\sigma = 0$. Introducing the dimensionless quantities in terms of layer thickness $d$ (typical length) and director relaxation time $\tau_d = (-\alpha_2) d^2 / K_{22}$ (typical time) the linearised equations \eqref{eqn:EL:velocity} and \eqref{eqn:EL:director} can be rewritten in the form \begin{subequations} \label{eqn:rolls} \begin{align} \label{eqn:rolls:1}&(\eta_{13} - 1) q^2 S n_{1z} + i q (\eta_{13} q^2 - \partial_z^2) v_{1x} = 0,\\ \label{eqn:rolls:2}&\partial_z [\eta_{52} q^2 + (1 - \eta_{32}) \partial_z^2] (S n_{1x}) \nonumber\\ &\quad + (\eta_{12} q^4 - \eta_{42} q^2 \partial_z^2 + \partial_z^4) v_{1y} = 0,\\ \label{eqn:rolls:3}&(\partial_z^2 - k_{32} q^2 - h) n_{1x} + S n_{1z} + i q v_{1x} = 0,\\ \label{eqn:rolls:4}&\partial_z(k_{12} \partial_z^2 - k_{32} q^2 - h + k_{12} e) n_{1z} \nonumber\\ &\quad + \lambda \partial_z (S n_{1x}) - (q^2 + \lambda \partial_z^2) v_{1y}= 0,\\ \label{eqn:rolls:5}&v_{1z,z} = - i q v_{1y}. \end{align} \end{subequations} where $\eta_{ij} = \eta_i / \eta_j$, $\eta_1 = (\alpha_4 + \alpha_5 - \alpha_2) / 2$, $\eta_2 = (\alpha_3 + \alpha_4 + \alpha_6) / 2$, $\eta_3 = \alpha_4 / 2$, $\eta_4 = \alpha_1 + \eta_1 + \eta_2$, $\eta_5 = - (\alpha_2 + \alpha_5) / 2$, $k_{ij} = K_{ii} / K_{jj}$, $\lambda = \alpha_3 / \alpha_2$, $h = \pi^2 H_0^2 / H_F^2$, $e = \sgn(\varepsilon_a) \pi^2 E_0^2/ E_F^2$ and $H_F = (\pi / d) \sqrt{K_{22} / (\mu_0 \chi_a)}$, $E_F = (\pi / d) \sqrt{K_{11} / (\varepsilon_0 |\varepsilon_a|)}$ are the critical Fr\'eedericksz transition fields for strong anchoring. For the shear rate $S$ one has, for Couette flow, \begin{equation} S = a^2,\: a^2 = \dfrac{V_0 \tau_d}{d} \end{equation} and for Poiseuille flow \begin{equation} S = -a^2 z,\: a^2 = -\dfrac{\Delta P}{\Delta x} \dfrac{\tau_d d}{\eta_3}. \end{equation} The anchoring properties are characterised by a surface energy per unit area, $F_s$, which has a minimum when the director at the surface is oriented along the {\em easy} axis (parallel to the $y$ axis in our case). A phenomenological expression for the surface energy $F_s$ can be written in terms of an expansion with respect to $(\Vec{n} - \Vec{n}_0)$. For small director deviations from the easy axis one obtains \begin{equation} \label{F_s} F_s = \frac12 W_a n_{1x}^2 + \frac12 W_p n_{1z}^2,\quad W_a > 0,\: W_p > 0, \end{equation} where $W_a$ and $W_p$ are the ``azimuthal'' and ``polar'' anchoring strengths, respectively. $W_a$ characterizes the surface energy increase due to distortions within the surface plate and $W_p$ relates to distortions out of the substrate plane. The boundary conditions for the director perturbations can be obtained from the torques balance equation \begin{equation} \pm \frac{\partial F}{\partial (\partial n_{1i}/\partial z)} + \frac{\partial F_s}{\partial n_{1i}} = 0, \end{equation} with ``$\pm$'' for $z = \pm d/2$. The boundary conditions \eqref{F_s} can be rewritten in dimensionless form as: \begin{equation} \label{bc:director:dim} \pm \beta_a n_{1x,z} + n_{1x} = 0,\: \pm \beta_p n_{1z,z} + n_{1z} = 0, \end{equation} with ``$\pm$'' for $z = \pm 1/2$. Here we introduced dimensionless anchoring strengths as ratios of the characteristic anchoring length ($K_{ii} / W_i$) over the layer thickness $d$: \begin{equation} \label{eqn:anchoring} \beta_a = K_{22} / (W_a d),\: \beta_p = K_{11} / (W_p d). \end{equation} In the limit of strong anchoring, $(\beta_a,\:\beta_p) \to 0$, one has $n_{1x} = n_{1z} = 0$ at $z = \pm 1/2$. For torque-free boundary conditions, $(\beta_a,\:\beta_p)\to \infty$, one has $n_{1x,z} = n_{1z,z} = 0$ at the boundaries. From \eqref{eqn:anchoring} one can see that by changing the thickness $d$, the dimensionless anchoring strengths $\beta_a$ and $\beta_p$ can be varied with the ratio $\beta_a/\beta_p$ remaining constant. The boundary conditions for the velocity field (no-slip) are \begin{align} \label{bc:vx} &v_{1x}(z = \pm 1 / 2) = 0;\\ \label{bc:vy} &v_{1y}(z = \pm 1 / 2) = 0;\\ \label{bc:vz} &v_{1z}(z = \pm 1 / 2) = v_{1z,z}(z = \pm 1 / 2) = 0. \end{align} The existence of a nontrivial solution of the linear ordinary differential equations \eqref{eqn:rolls} with the boundary conditions \eqref{bc:director:dim}, (\ref{bc:vx} -- \ref{bc:vz}) gives values for the shear rate $S_0(q)$ (neutral curve). The critical value $S_c(q_c)$, above which the basic state \eqref{eqn:base} becomes unstable, are given by the minimum of $S_0$ with respect to $q$. \begin{table} \caption{\label{table:sym:all}Symmetry properties of the solutions of equations \eqref{eqn:rolls} under $\{z \to -z\}$.} \begin{ruledtabular} \begin{tabular}{ccccc} \multicolumn{1}{c}{}&\multicolumn{2}{c}{Couette flow}&\multicolumn{2}{c}{Poiseuille flow}\\ \multicolumn{1}{c}{Perturbation}& ``odd'' & ``even'' & ``odd'' & ``even''\\ \hline $n_{1x}$ & odd & even & odd & even\\ $n_{1z}$ & odd & even & even & odd\\ $v_{1x}$ & odd & even & odd & even\\ $v_{1y}$ & even & odd & odd & even\\ $v_{1z}$ & odd & even & even & odd\\ \end{tabular} \end{ruledtabular} \end{table} The symmetry properties of the solutions of equations \eqref{eqn:rolls} under the reflection $z \to - z$ is shown in the Table \ref{table:sym:all}. We will always classify the solutions by the $z$ symmetry of the $x$ component of the director perturbation $n_{1x}$ (first row in Table I). In case of positive $\varepsilon_a$, for some critical value of the electric field the basic state loses its stability already in the absence of flow ({\em Fre\'edericksz transition}). Clearly the Fre\'edericksz transition field depends on the polar anchoring strength. There is competition of the elastic torque $K_{11} n_{1z,zz}$ and the field-induced torque $\varepsilon_a \varepsilon_0 E_0^2 n_{1z}$. The solution of Eq. \eqref{eqn:rolls:4} with $n_{1x} = 0$, $v_{1y} = 0$ for $h = 0$ has the form \begin{equation} \label{def:Eweak} n_{1z} = C \cos(\pi \delta z / d), \end{equation} where $\delta = E_F^{weak} / E_F$ and $E_F^{weak}$ is the actual Fr\'eedericksz field. After substituting $n_{1z}$ into the boundary conditions \eqref{bc:director:dim} we obtain the expression for $\delta$: \begin{equation} \label{eqn:Eweak} \tan\dfrac{\pi \delta}2 = \dfrac1{\pi \beta_p \delta}. \end{equation} One easily sees that $\delta \to 1$ for $\beta_p \to 0$ and $\delta \to \sqrt{2 / \beta_p} / \pi$ for $\beta_p \to \infty$. For $\beta_p =1$ one gets $E_F^{weak} = 0.42 E_F$. \section{Homogeneous instability} In order to obtain simpler equations we use the renormalized variables as in Ref. \cite{Tarasov:LC:V28:N6:P833:Y2001}: \begin{align} \label{def:renorm} &\Tilde{S} = \beta^{-1} S,\: N_{1x} = \beta^{-1} n_{1x},\: N_{1z} = n_{1z},\: V_{1x} = \beta^{-1} v_{1x},\nonumber\\ & V_{1y} = (\beta^2 \eta_{23})^{-1} v_{1y},\: V_{1z} = (\beta^2 \eta_{23})^{-1} v_{1z} \end{align} with \begin{equation} \beta^2 = \alpha_{32} k_{21} \eta_{32},\: \alpha_{i j} = \dfrac{\alpha_i}{\alpha_j}. \end{equation} In the case of homogeneous perturbations ($q = 0$) Eqs. \eqref{eqn:rolls} reduce to $V_{1z} = 0$ and \begin{subequations} \label{eqn:set:hom} \begin{align} \label{eqn:set:hom:1} &V_{1y,zz} - (1 - \eta_{23}) (\Tilde{S} N_{1x})_{,z} = 0,\\ \label{eqn:set:hom:2}&\Tilde{S} N_{1z} - N_{1x,zz} + h N_{1x} = 0,\\ \label{eqn:set:hom:3}&\eta_{23} \Tilde{S} N_{1x} + N_{1z,zz} - V_{1y,z} - (k_{21} h - e) N_{1z} = 0. \end{align} \end{subequations} \subsection{Couette flow} For Couette flow we can obtain the solution of \eqref{eqn:set:hom} semi-analytically. For the ``odd'' solution one gets \begin{align} &N_{1x} = C_1 \sinh(\xi_1 z) + C_2 \sin(\xi_2 z),\\ &N_{1z} = C_3 \sinh(\xi_1 z) + C_4 \sin(\xi_2 z),\\ &V_{1y} = C_5 \cosh(\xi_1 z) + C_6 \cos(\xi_2 z) + C_7. \end{align} Taking into account the boundary conditions (\ref{bc:director:dim}, \ref{bc:vy}) the solvability condition for the $C_i$ (``boundary determinant'' equal to zero) gives an expression for the critical shear rate $a_c$: \begin{multline} \label{eqn:couette:odd} (h + \xi_2^2) [\xi_1 \beta_a \cosh(\xi_1 / 2) + \sinh(\xi_1 / 2)] \\ \times [\xi_2 \beta_p \cos(\xi_2 / 2) + \sin(\xi_2 / 2)] \\ - (h - \xi_1^2)[\xi_2 \beta_a \cos(\xi_2 / 2) + \sin(\xi_2 / 2)] \\ \times [\xi_1 \beta_p \cosh(\xi_1 / 2) + \sinh(\xi_1 / 2)] = 0. \end{multline} where \begin{align} &\xi_1^2 = \dfrac{[ (1 + k_{12}) h - k_{12} e] + \xi}{2 k_{12}},\\ &\xi_2^2 = \dfrac{- [(1 + k_{12}) h - k_{12} e] + \xi}{2 k_{12}},\\ &\xi = \sqrt{ [(1 - k_{12}) h - k_{12} e]^2 + 4 k_{12}^2 a^4}. \end{align} For the ``even'' solution one obtains: \begin{align} &N_{1x} = C_1 \cosh(\xi_1 z) + C_2 \cos(\xi_2 z) + C_3,\\ &N_{1z} = C_4 \cosh(\xi_1 z) + C_5 \cos(\xi_2 z) + C_6,\\ &V_{1y} = C_7 \sinh(\xi_1 z) + C_8 z. \end{align} The boundary conditions (\ref{bc:vx}-\ref{bc:vz}) now lead to the following condition (``boundary determinant''): \begin{widetext} \begin{equation} \label{eqn:couette:even} \begin{vmatrix} 1 & h & \dfrac{\eta_{23}}{2}\left(\dfrac{h(h-k_{12}e)}{a^4k_{12}\eta_{23}}-1\right)\\ - \xi_2 \beta_a \sin(\xi_2 / 2) + \cos(\xi_2 / 2) & (h + \xi_2^2)[ - \xi_2 \beta_p \sin(\xi_2 / 2) + \cos(\xi_2 / 2)] & \dfrac{1 - \eta_{23}}{\xi_2}\sin(\xi_2 / 2)\\ \xi_1 \beta_a \sinh(\xi_1 / 2) + \cosh(\xi_1 / 2) & (h - \xi_1^2)[\xi_1 \beta_p \sinh(\xi_1 / 2) + \cosh(\xi_1 / 2)] & \dfrac{1 - \eta_{23}}{\xi_1}\sinh(\xi_1 / 2) \end{vmatrix}=0. \end{equation} \end{widetext} \subsection{Poiseuille flow} In the case of Poiseuille flow the system \eqref{eqn:set:hom} with $\Tilde{S} = - z a^2 / \beta$ admits an analytical solution only in the absence of external fields (in terms of Airy functions) \cite{Tarasov:LC:V28:N6:P833:Y2001}. In the presence of fields we solve the problem numerically. In the framework of the Galerkin method we expand $N_{1x}$, $N_{1z}$ and $V_{1y}$ in a series \begin{align} \label{poise:full:galerkin} &N_{1x} = \sum\limits_{n=1}^{\infty} C_{1,n} f_n(z),\nonumber\\ &N_{1z} = \sum\limits_{n=1}^{\infty} C_{2,n} g_n(z),\\ &V_{1y} = \sum\limits_{n=1}^{\infty} C_{3,n} u_n(z),\nonumber \end{align} where the trial functions $f_n$, $g_n$ and $u_n$ satisfy the boundary conditions \eqref{bc:director:dim}, \eqref{bc:vy}. For the ``odd'' solution we write \begin{equation} f_n(z) = \zeta_n^o(z;\beta_a),\; g_n(z) = \zeta_n^e(z;\beta_p),\: u_n(z) = \nu_n^o(z) \end{equation} and for the ``even'' solution \begin{equation} f_n(z) = \zeta_n^e(z;\beta_a),\; g_n(z) = \zeta_n^o(z;\beta_p),\: u_n(z) = \nu_n^e(z). \end{equation} The functions $\zeta_n^o(z;\beta)$, $\zeta_n^e(z;\beta)$, $\nu_n^o(z)$, $\nu_n^e(z)$ are given in Appendix A. In our calculations we have to truncate the expansions \eqref{poise:full:galerkin} to a finite number of modes. After substituting \eqref{poise:full:galerkin} into the system \eqref{eqn:set:hom} and projecting the equations on the trial functions $f_n(z)$, $g_n(z)$ and $u_n(z)$ one gets a system of linear homogeneous algebraic equations for $\Vec{X} = \{C_{i,n}\}$ in the form $(A - a^2 B) \Vec{X} = 0$. We have solved this eigenvalue problem for $a^2$. The lowest (real) eigenvalue corresponds to the critical shear rate $a_c^2$. According to the two types of $z$-symmetry of the solutions (and of the set of trial functions) one obtains the threshold values of $a_c^2$ for the ``odd'' and ``even'' instability modes. The number of Galerkin modes was chosen such that the accuracy of the calculated eigenvalues was better than 1\% (we took ten modes in case of ``odd'' solution and five modes for ``even'' solution). \subsection{Approximate analytical expression for the critical shear rate} In order to obtain an {\em easy-to-use} analytical expression for the critical shear rate as a function of the surface anchoring strengths and the external fields we use the lowest-mode approximation in the framework of the Galerkin method. By integrating \eqref{eqn:set:hom:1} over $z$ one can eliminate $V_{1y,z}$ from \eqref{eqn:set:hom:3} which gives \begin{equation} \label{eqn:set:director} \Tilde{S} N_{1x} + N_{1z,zz} + (k_{21} h - e) N_{1z} = K, \end{equation} where $K$ is an integration constant. Taking into account the boundary conditions for $V_{1y}$ one has \begin{equation} \label{eqn:K} K - (1 - \eta_{32}) \int\limits_{-1/2}^{1/2} S N_{1x}(z) \:dz = 0. \end{equation} We choose for the director components $N_{1x}$, $N_{1z}$ the one-mode approximation \begin{equation} \label{eqn:leading} N_{1x} = C_1 f(z),\: N_{1z} = C_2 g(z), \end{equation} Substituting \eqref{eqn:leading} into \eqref{eqn:set:hom:2} and \eqref{eqn:set:director} and projecting the first equation on $f(z)$ and the second one on $g(z)$ we get algebraic equations for $C_i$. The solvability condition [together with \eqref{eqn:K}] gives the expression for the critical shear rate $a_c^2$ \begin{equation} \label{eqn:one:ac} a_c^2 = \sqrt{\dfrac{c_1 c_2} {c_3}}, \end{equation} where $c_1 =\langle ff''\rangle - h \langle f^2\rangle $, $c_2 = \langle gg''\rangle - (h/k_{12} - e) \langle g^2\rangle $, $c_3 = \langle sfg\rangle [\langle sfg\rangle - (1 - \eta_{23}) \langle sf\rangle \langle g\rangle ]$, where $\langle \dots\rangle $ denotes a spatial average \begin{equation} \label{def:int} \langle \dots\rangle =\int\limits_{-1/2}^{1/2}(\dots)\:dz. \end{equation} The values for the integrals $\langle \dots\rangle $ are given in Appendix B. In Table \ref{table:trial_func:appr} and Appendix A the trial functions used are given. Equation \eqref{eqn:one:ac} can be used for both Couette and Poiseuille flow by choosing the function $s(z)$ [where $s(z) = 1$ for Couette flow and $s(z) = - z$ for Poiseuille flow] and the trial functions $f(z)$ and $g(z)$ with appropriate symmetry. \begin{table} \caption{\label{table:trial_func:appr}Trial functions for the homogeneous solutions.} \begin{ruledtabular} \begin{tabular}{ccccc} \multicolumn{1}{c}{}&\multicolumn{2}{c}{Couette flow}&\multicolumn{2}{c}{Poiseuille flow}\\ \multicolumn{1}{c}{Function}& ``odd'' & ``even'' & ``odd'' & ``even''\\ \hline $f(z)$ & $\zeta_1^o(z;\:\beta_a)$ & $\zeta_1^e(z;\:\beta_a)$ & $\zeta_1^o(z;\:\beta_a)$ & $\zeta_1^e(z;\:\beta_a)$\\ $g(z)$ & $\zeta_1^o(z;\:\beta_p)$ & $\zeta_1^e(z;\:\beta_p)$ & $\zeta_1^e(z;\:\beta_p)$ & $\zeta_1^o(z;\:\beta_p)$\\ \end{tabular} \end{ruledtabular} \end{table} For the material MBBA in the case of Couette flow the one-mode approximation \eqref{eqn:one:ac} for the ``odd'' solution gives an error that varies from 2.5\% to 16\% when $H_0 / H_F$ varies from 0 to 4. The ``even'' solution has an error of $0.6\% \div 8\%$ for $0 \leqslant H_0 / H_F \leqslant 3$ and of $0.6\% \div 12\%$ for $0 \leqslant E_0 / E_F \leqslant 0.6$. For Poiseuille flow for ``odd'' solution the error is $29\%$ in the absence of fields. For the ``even'' solution the error is $12\% \div 15\%$ for magnetic fields $0 \leqslant H_0 / H_F \leqslant 0.5$. For both Couette and Poiseuille flow the accuracy of the formula \eqref{eqn:one:ac} decreases with increasing field strengths. \section{Spatially periodic instabilities} We used for Eqs. \eqref{eqn:rolls} again the renormalized variables \eqref{def:renorm}. The system \eqref{eqn:rolls} has no analytical solution. Thus we solved the problem numerically in the framework of the Galerkin method: \begin{eqnarray} \label{Ansatz:rolls} N_{1x} = e^{iqy} \sum\limits_{n=1}^{\infty} C_{1,n} f_n(z),\: N_{1z} = e^{iqy} \sum\limits_{n=1}^{\infty} C_{2,n} g_n(z),\\ V_{1x} = e^{iqy} \sum\limits_{n=1}^{\infty} C_{3,n} u_n(z),\: V_{1z} = e^{iqy} \sum\limits_{n=1}^{\infty} C_{4,n} w_n(z). \end{eqnarray} After substituting \eqref{Ansatz:rolls} into the system \eqref{eqn:rolls} and projecting on to the trial functions $\{f_n(z),\:g_n(z),\:u_n(z),\:w_n(z)\}$ we get a system of linear homogeneous algebraic equations for $\Vec{X} = \{C_{i,n}\}$. This system has the form $[A(q) - a^2(q) B(q)] \Vec{X} = 0$. We have solved the eigenvalue problem numerically to find the marginal stability curve $a(q)$. For the numerical calculations we have chosen the trial functions shown in Table \ref{table:trial_func:rolls} and Appendix A. \begin{table} \caption{\label{table:trial_func:rolls}Trial functions for the spatially periodic solutions.} \begin{ruledtabular} \begin{tabular}{ccccc} \multicolumn{1}{c}{}&\multicolumn{2}{c}{Couette flow}&\multicolumn{2}{c}{Poiseuille flow}\\ \multicolumn{1}{c}{Function}& ``odd'' & ``even'' & ``odd'' & ``even''\\ \hline $f(z)$ & $\zeta_n^o(z;\:\beta_a)$ & $\zeta_n^e(z;\:\beta_a)$ & $\zeta_n^o(z;\:\beta_a)$ & $\zeta_n^e(z;\:\beta_a)$\\ $g(z)$ & $\zeta_n^o(z;\:\beta_p)$ & $\zeta_n^e(z;\:\beta_p)$ & $\zeta_n^e(z;\:\beta_p)$ & $\zeta_n^o(z;\:\beta_p)$\\ $u(z)$ & $\nu_n^o(z)$ & $\nu_n^e(z)$ & $\nu_n^o(z)$ & $\nu_n^e(z)$\\ $w(z)$ & $\varsigma_n^o(z)$ & $\varsigma_n^e(z)$ & $\varsigma_n^e(z)$ & $\varsigma_n^o(z)$\\ \end{tabular} \end{ruledtabular} \end{table} In order to get an approximate expression for the threshold we use the leading-mode approximation in the framework of the Galerkin method. We used the same scheme described above for the single mode and get the following formula for the critical shear rate: \begin{equation} \label{rolls:one:common} a_c^2 = \sqrt{ \eta_{23} f_x f_z / (\tilde{\alpha}_2\tilde{\alpha}_3) }, \end{equation} with \begin{align} &f_x = \langle ff''\rangle - (q^2 k_{32} + h) \langle f^2\rangle ,\\ &f_z = \langle gg''\rangle - (q^2 k_{31} + k_{12} h - e) \langle g^2\rangle ,\\ &\tilde{\alpha}_2 = [ \langle fsg\rangle - q^2 (1 - \eta_{31}) \langle fu\rangle \langle gsu\rangle / \gamma ],\\ &\tilde{\alpha}_3 = \langle fsg\rangle + [ \alpha_{23} q^2 \langle gw\rangle + \alpha_3 \langle gw''\rangle ] \\ & \quad \times [ (1 - \eta_{32}) \langle w[sf]''\rangle - \eta_{52} q^2 \langle wsf\rangle ] / r,\\ &\gamma = q^2 \langle uu\rangle - \eta_{31} \langle uu''\rangle ,\\ &r = \langle ww^{(4)}\rangle - \eta_{42} q^2 \langle ww''\rangle + \eta_{12} q^4 \langle ww\rangle . \end{align} The values of the integrals $\langle \dots\rangle $ appearing in the expression \eqref{rolls:one:common} are given in Appendix C. In the case of strong anchoring an approximate analytical expression for $a_c^2 = a_c^2(q_c)$ was obtained by Manneville \cite{Manneville:JdPh:1976:285} using test functions that satisfy free-slip boundary conditions. The formula \eqref{rolls:one:common} is more accurate because we chose for $v_{1z}$ Chandrasekhar functions that satisfy the boundary conditions \eqref{bc:vz}. For calculations we used material parameters for MBBA. The accuracy of \eqref{rolls:one:common} is better than 1\% for Couette flow and better than 3\% for Poiseuille flow. Note, that Eq. \eqref{eqn:one:ac} for the homogeneous instability is more accurate than \eqref{rolls:one:common} for $q = 0$ because \eqref{rolls:one:common} was obtained by solving four equations \eqref{eqn:rolls} by approximating all variables, whereas \eqref{eqn:one:ac} was obtained by solving the reduced equations \eqref{eqn:set:hom} by approximating only two variables. \section{Discussion} For the calculations we used parameters for MBBA at 25 $^\circ$C \cite{mp:MBBA}. Calculations were made for the range of anchoring strengths $\beta_a = 0 \div 1$ and $\beta_p = 0 \div 1$. \subsection{Couette flow} \begin{figure} \epsfig{file=couette_e.eps,height=21cm} \caption{Contour plot of the critical shear rate $a_c^2$ for Couette flow vs. $\beta_a$ and $\beta_p$. $a$: $E_0 = 0$; $b$: $E_0 = E_F^{weak}$, $\varepsilon_a < 0$; $c$: $E_0 = E_F^{weak}$, $\varepsilon_a > 0$. $E_F$ is defined after Eq. \eqref{def:Eweak} and calculated in Eq. \eqref{eqn:Eweak}.} \label{fig:couette0} \end{figure} We found that without and with an additional electric field the critical shear rate $a_c^2$ for the ``even'' type homogeneous instability (EH) is systematically lower than the threshold for other types of instability (Fig. \ref{fig:couette0}a--c). Note, that in the presence of the field the symmetry with respect to the exchange $\beta_a \leftrightarrow \beta_p$ is broken. In Fig. \ref{fig:couette0} contour plots for the critical value $a_c^2$ vs. anchoring strengths $\beta_a$ and $\beta_p$ for different values of the electric field are shown. The differences between $a_c^2$ obtained from the exact, semi-analytical solution \eqref{eqn:couette:even} and from the one-mode approximation \eqref{eqn:one:ac} are indistinguishable in the figure. \begin{figure} \epsfig{file=couette_h.eps,height=21cm} \caption{Critical shear rates and phase diagram for the instabilities under Couette flow with additional magnetic field. $a$: $H_0 / H_F = 3$; $b$: $H_0 / H_F = 3.5$; $c$: $H_0 / H_F = 4$. Boundaries for occurrence of instabilities are given by thick solid lines (full numerical) and thick dashed lines (one-mode approximation).} \label{fig:couette1} \end{figure} In Fig. \ref{fig:couette1} contour plots of $a_c^2$ (thin dashed lines) and the boundaries where the type of instability changes [the solid lines are obtained numerically, the thick dashed lines from \eqref{eqn:one:ac}] for different values of magnetic field are shown. For not too strong magnetic field in the region of weak anchoring the ``odd'' type homogeneous instability (OH) takes place (Fig. \ref{fig:couette1}a). Increasing the magnetic field the OH region expands toward stronger anchoring strengths. Above $H_0 \approx 3.2$ a region with lowest threshold corresponding to the ``even'' roll mode (ER) appears. This region has borders with both types of the homogeneous instability (Fig. \ref{fig:couette1}b). With increasing magnetic field the ER region increases (Fig. \ref{fig:couette1}c) and above $H_0 / H_F = 4$ the ER instability has invaded the whole investigated parameter range. For strong anchoring and $H_0 / H_F = 3.5$ the critical wave vector is $q_c = 5.5$. It increases with increasing magnetic field and decreases with decreasing anchoring strengths. With increasing magnetic field the threshold for the EH instability becomes less sensitive to the surface anchoring. Leslie has pointed out (using an approximate analytical approach) that for strong anchoring a transition from a homogeneous state without transverse flow (EH) to one with such flow (OH) as the magnetic field is increased is not possible in MBBA because of the appearance of the ER type instability \cite{Leslie:MCLC:1976}. This is consistent with our results. We find that the EH--OH transition in MBBA is possible only in the region of weak anchoring (Figs. \ref{fig:couette1}a--c). \begin{figure} \epsfig{file=aq.eps,width=8cm} \caption{$a_c$ vs. $q$. Couette flow, $\beta_a = 0.1$, $\beta_p = 0.1$. $a$: $H_0 / H_F = 3$; $b$: $H_0 / H_F = 3.4$; $c$: $H_0 / H_F = 4$.} \label{fig:couette2} \end{figure} In Fig. \ref{fig:couette2} marginal stability curves for different values of the magnetic field and fixed anchoring strengths is shown (solid line for ER and dashed lines for OR). There are always two minima for the even mode; one of them at $q = 0$ that corresponds to the homogeneous instability EH. For small magnetic field the absolute minimum is at $q = 0$ (line a). The OR curve is systematically higher than ER. In a small range of $q$ (dotted lines) a stationary ER solution does not exist but we have OR instead. With increasing magnetic field the critical amplitude for the EH minimum ($q = 0$) increases more rapidly then the one for the ER minimum ($q \neq 0$) so that for $H_0 / H_F > 3.4$ the ER solution is realized (lines b and c). The range of $q$ where ER is replaced by OR expands with increasing magnetic field. For the ER instability in the absence of fields and strong anchoring we find $a_c^2 = 12.15$ from the semi-analytical expression \eqref{eqn:couette:even} as well as from the one-mode approximation \eqref{eqn:one:ac} and also \eqref{rolls:one:common} with $q = 0$. The only available experimental value for $a_c^2$ is $6.3 \pm 0.3$ \cite{Pieranski:SSC:1973}. We suspect that the discrepancy is due to deviations from the strong anchoring limit and the difference in the material parameters of the substance used in the experiment. Assuming $\beta_a \ll 1$ one would need $\beta_p \approx 1$ to explain the experimental value. \subsection{Poiseuille flow} \begin{figure} \epsfig{file=poise_e.eps,height=21cm} \caption{Critical shear rates and phase diagram for the instabilities in Poiseuille flow. $a$: $E_0 = 0$; $b$: $E_0 = E_0^{weak}$, $\varepsilon_a < 0$; $c$: $E_0 = E_0^{weak}$, $\varepsilon_a > 0$. Thin dashed lines: full numerical threshold; dotted lines: one-mode approximation for threshold. Boundaries for occurrence of instabilities are given by thick solid lines (full numerical) and thick dashed lines (one-mode approximation).} \label{fig:poise0} \end{figure} In Fig. \ref{fig:poise0} the contour plot for $a_c^2$ [thin dashed lines from the full numerical calculation, dotted lines from the one-mode approximations \eqref{eqn:one:ac} and \eqref{rolls:one:common}] and the boundary for the various types of instabilities [thick solid line: numerical; thick dashed line: \eqref{eqn:one:ac} and \eqref{rolls:one:common}] are shown. In Poiseuille flow the phase diagram is already very rich in the absence of external fields. In the region of large $\beta_a$ one has the EH instability. For intermediate anchoring strengths rolls of type OR occur [Fig. \ref{fig:poise0}$a$]. Note, that even in the absence of the field there is no symmetry under exchange $\beta_a \leftrightarrow \beta_p$, contrary to Couette flow. The one-mode approximations \eqref{eqn:one:ac} and \eqref{rolls:one:common} not give the transition to EH for strong anchoring. Here we should note that in that region the difference between the EH and the OR instability thresholds is only about 5\%. By varying material parameters [increase $\alpha_2$ by 10\% or decrease $\alpha_3$ by 20\% or $\alpha_5$ by 25\% or $K_{33}$ by 35\%] it is possible to change the type of instability in that region. Application of an electric field leads for $\varepsilon_a < 0$ ($\varepsilon_a > 0$) to expansion (contraction) of the EH region [Figs. \ref{fig:poise0}$b$ and \ref{fig:poise0}$c$]. At $E_0 / E_F = 1$ and $\varepsilon_a < 0$ rolls vanish completely and the EH instability occurs in the whole area investigated. For $\varepsilon_a > 0$ the instability of OH type appears in the region of large $\beta_p$. In this case, increasing the electric field from $E_F^{weak}$ to $E_F$ cause an expansion of the OH region. Note that for $\beta_p > 1$, which is in the OH region, the Fre\'edericksz transition occurs first . \begin{figure} \centering \epsfig{file=poise_h.eps,width=6.3cm} \caption{Phase diagram for the instabilities under Poiseuille flow with an additional magnetic field ($H_0 / H_F = 0.4$).} \label{fig:poise1} \end{figure} An additional magnetic field suppresses the homogeneous instability (Fig. \ref{fig:poise1}). Above $H_0 / H_F \approx 0.5$ the OR instability (Fig. \ref{fig:poise1}) occurs for all anchoring strengths investigated. The wave vector $q_c$ in the absence of fields is $1.4$. Application of an electric field decreases $q_c$ whereas the magnetic field increases $q_c$. The wave vector decreases with decreasing anchoring strengths. In the absence of fields and strong anchoring we find for the EH instability $a_c = 102$ [Eq. \eqref{eqn:one:ac} gives 110 and Eq. \eqref{rolls:one:common} with $q = 0$ gives 130]. The experimental value is 92 \cite{Guyon:JdPh:1975}. Thus, theoretical calculations and experimental results are in good agreement. Note, that in the experiments \cite{Guyon:JdPh:1975} actually not steady but oscillatory flow with very low frequency was used ($f = 5 \cdot 10^{-3}$ Hz). In summary, the orientational instabilities for both steady Couette (semi-analytical for homogeneous instability and numerical for rolls) and Poiseuille flow (numerical) were analysed rigorously taking into account weak anchoring and the influence of external fields. Easy-to-use expressions for the threshold of all possible types of instabilities were obtained and compared with the rigorous calculations. In particular the region in parameter space where the different types of instabilities occurred were determined. \acknowledgments Financial support from DFG (project Kr690/22-1 and EGK ``Non-equilibrium phenomena and phase transition in complex systems'').
{ "timestamp": "2005-03-10T23:19:31", "yymm": "0503", "arxiv_id": "physics/0503091", "language": "en", "url": "https://arxiv.org/abs/physics/0503091" }
\section{Introduction} \label{S:Intro} The best-known application of quantum cryptography is quantum key distribution (QKD) \cite{BB84}. The goal of QKD is to allow two parties, Alice and Bob, to share a common string of secret in the presence of an eavesdropper, Eve. Such a key can subsequently be used for, for example, perfectly secure communications via the so-called one-time-pad. Unlike conventional cryptography, the security of QKD is guaranteed by the fundamental law of physics---the Heisenberg uncertainty principle. The best-known protocol for QKD is the Bennett-Brassard protocol (BB84) \cite{BB84}. In BB84, Alice sends Bob a sequence of single photons in one of the four polarizations (vertical, horizontal, 45-degree and 135-degree) and Bob randomly performs a measurement in one of the two conjugate bases. In principle, the security of QKD has been proven in a number of papers including \cite{proofs,Ben-Or,ShorPreskill}. For practical implementations, an attenuated laser pulse (a so-called weak coherent state) is often used as the source. The security of QKD with a rather generic class of imperfect devices has been proven in GLLP \cite{GLLP}, following the earlier work \cite{ilm} Recently, Hwang \cite{HwangDecoy} has proposed a decoy state idea for improving the performance (i.e., the key generation rate and distance) of QKD systems. We \cite{Decoy} have demonstrated rigorously how the decoy state idea can be combined with GLLP to obtain a key generation rate (per pulse emitted by Alice) which is lower bounded by: \begin{equation}\label{refinedkeyrate} S \geq Q_{signal} \{- H_2(E_{signal}) + \Omega_1 [ 1- H_2(e_1)] \}, \end{equation} where $Q_{signal}$ and $E_{signal}$ are respectively the gain and quantum bit error rate (QBER) of the signal state, $\Omega_1$ and $e_1$ are respectively the fraction and QBER of detection events by Bob that have originated from single-photon signals emitted by Alice. Here, the gain means the ratio of Bob's detection events to Alice's total number of emitted signals. [Decoy state QKD has subsequently been investigated by Wang \cite{WangDecoy} and by Harrington \cite{Harrington}.] The key goal of this paper is to increase the above key generation rate in Eq.~\ref{refinedkeyrate} by a term $Q_{signal} \Omega_0$ where $\Omega_0$ is the fraction of detection events of Bob that have originated from vacua emitted by Alice. More concretely, we have the following main Theorem. {\bf Theorem~1} The key generation of an efficient BB84 scheme is given by: \begin{equation}\label{newkeyrate} S \geq Q_{signal} \{- H_2(E_{signal}) + \Omega_0 + \Omega_1 [ 1- H_2(e_1)] \}, \end{equation} where $\Omega_0$ is the fraction of detection events by Bob that has originated from the vacuum signals emitted by Alice. In other words, we find that each detection event by Bob that has originated from a vacuum (i.e., nothing) emitted by Alice automatically contributes to a bit of secure key over and above the prior art result (Eq.~\ref{refinedkeyrate}) presented in \cite{GLLP} and also \cite{Decoy}. Before we embark on a detailed discussion, let us check for the consistency of our result. Naively, one might think that our suggestion that the vacuum will contribute to a secure key is an insane idea because if nothing is emitted by Alice, what is the origin of security? We remark that the vacuum {\it alone} does {\it not} contribute to a secure key. More concretely, suppose all the signals sent by Alice are vacua and there are no background events. Then, $\Omega_0 =1$, $\Omega_1 =0$, and $E_{signal} = 1/2$. Therefore, from Eq.~\ref{newkeyrate}, we get the lower bound $0$ for the key generation rate. The reason is that the term $\Omega_0$ is exactly cancelled by the error correction term $- H_2(E_{signal})$. What Eq.~\ref{newkeyrate} does say is that no privacy amplification is needed for the vacua state. This is intuitively clear because Eve cannot have any a priori information on Alice's bit, if nothing is emitted from Alice's laboratory. Now, let us prove our main result (Eq.~\ref{newkeyrate}). We shall use the method of communication complexity. As noted by by Ben-Or \cite{Ben-Or} and by Renner and Koenig \cite{RennerKoenig}, the number of rounds of universal hashing needed for privacy amplification in QKD is at most given by any upper bound to the size of Eve's quantum memory which contains information on the key. In other words, we have, informally: {\bf Theorem~2} \cite{Ben-Or,RennerKoenig}: The key generation rate in QKD \begin{equation} S \geq N-{\cal S}_{Eve} \end{equation} where $N$ is the size of the sifted key shared between Alice and Bob and ${\cal S}_{Eve}$ is the size of Eve's quantum memory. [A more formal definition involving the relevant $\epsilon$ and $\delta$ can be found as Eq.~(11) in \cite{RennerKoenig}.] {\it Remark}: Note that Theorem~2 only gives a lower bound to the key generation rate because it does not consider the possibility of advantage distillation in QKD \cite{twoway}. In summary, all we need to compute (a lower bound to) the key generation rate is to work out the size of Eve's quantum memory. {\bf Proof of Theorem~1}: Now, note that Eve has two pieces of information on the key. The first piece, which is strictly quantum, comes from Eve's eavesdropping attack during the quantum transmission from Alice to Bob. The second piece is classical and comes from the classical error correction part. We argue that the first piece, from eavesdropping the quantum transmission, consists of two parts: single-photon part and multi-photon part. It should be emphasized that the vacua signals do {\it not} contribute at all. This is because, since Alice is emitting nothing, Eve cannot possibly learn anything about Alice's key. Eve can influence and, in fact, decide on Bob's key by sending her own photons into Bob's detector. However, Bob's key does not really tell Eve anything about Alice's key. Let us consider the multi-photon part first. We take the most conservative assumption that Eve has all the information on all multi-photon signals. Her quantum memory size on the multi-photon part is then given by $Q_{signal} \Omega_m$ . Here, $\Omega_m$ is the fraction of detection events of Bob that have originated from multi-photon signals. Note that $\Omega_0 + \Omega_1 + \Omega_m = 1$. The single-photon part is given by simply $ Q_{signal} \Omega_1 H_2 (e_1^{phase})$, where $e_1^{phase}$ is the phase error rate of the single-photon signals. From Shor-Preskill's proof \cite{ShorPreskill}, $e_1^{phase} = e_1$, which is the bit-flip error rate for the single-photon signals. So, the quantum memory for single-photon part is actually given by $ Q_{signal} \Omega_1 H_2 (e_1)$. Adding the two parts, the first piece of Eve's information has a memory size $Q_{signal} [ \Omega_1 H_2 (e_1) + \Omega_m ]$. The second piece of Eve's information, which comes from classical error correction, is asymptotically given by $Q_{signal} H_2(E_{signal})$. In summary, adding the two pieces together, the total quantum memory size of Eve is given by ${\cal S}_{Eve}= Q_{signal} [ H_2(E_{signal}) + \Omega_1 H_2 (e_1) + \Omega_m ]$. Now, the length of the sifted key (per pulse emitted by Alice) shared by Alice and Bob is $N= Q_{signal} [ \Omega_0 + \Omega_1 + \Omega_m ]$. Therefore, the number of secure key bits (per pulse emitted by Alice) is given by \begin{eqnarray} S &\geq& N-{\cal S}_{Eve} \nonumber \\ &=& Q_{signal} \{ - H_2(E_{signal}) + \Omega_0 + \Omega_1 [ 1- H_2(e_1)] \}. \end{eqnarray} which is precisely Eq.~(\ref{newkeyrate}). This concludes the proof of our Theorem~1. In summary, we have increased the key generation from Eq.~(\ref{refinedkeyrate}) in the prior art result \cite{GLLP} to Eq.~(\ref{newkeyrate}) by showing that, rather counter-intuitively, the detection events due to vacua contribute directly to the secure key. What is interesting about this result is that it is based on a communication complexity approach and is not entirely clear whether it can be derived from an entanglement distillation approach. In future, it will perhaps be interesting to rephrase this result in the general framework of $\Gamma$ states \cite{Gammastates}, which generalizes the entanglement distillation approach. \section*{Acknowledgements} \noindent We thank helpful discussions with colleagues including J. Batuwantudawe, Jean-Christian Boileau, Debbie Leung, John Preskill and Kiyoshi Tamaki. This part is financially supported in part by funding agencies including CFI, CIPI, CRC program, NSERC, OIT, and PREA. Parts of this paper were written during visits to the Institute of Quantum Information (IQI) at Caltech and to the Isaac Newton Institute, Cambridge, UK, whose kind hospitality is acknowledged.
{ "timestamp": "2005-03-01T03:16:28", "yymm": "0503", "arxiv_id": "quant-ph/0503004", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503004" }
\section*{1. Introduction} The Faraday Experiment is probably the first non equilibrium pattern forming system that has been investigated scientifically, namely by Michael Faraday in 1831 \cite{Faraday}. Nevertheless it was only recently that it was possible to determine the complete Fourier spectrum of the deformed surface state \cite{Kityk04}. While an experimental analysis of the full mode spectrum in other pattern forming model systems like Rayleigh-Benard or Taylor-Couette has been standard technique for a long time, it is the refraction of light at the free surface of a liquid that renders the analysis of surface waves so difficult. Quantitative information about the patterned state up to higher orders is important not only to verify the validity of theoretical calculations \cite{kumar94,zhang96} but also to gain insight on the resonance mechanisms that \textit{form} the patterns. The Faraday experiment is especially known for its richness of different patterns that are observed \cite{miles90,milner90,Christiansen92,edwards94,binks97,kudrolli97}. By using complex liquids \cite{wagner99}, very low fill heights \cite{wagner00} or by introducing additional driving frequencies, highly complex ordered states like superlattices \cite{kudrolli98,arbell98} have been observed recently. But we will demonstrate that even simple patterns like lines, squares and hexagons observed in a single driving frequency experiment can still unveil unknown surprising characteristics. A discussion of the different attempts to reveal quantitative information on the the surface elevation profile $\zeta(r,t)$ of Faraday waves is given in \cite{Wernet01}. The main difficulties in determining the surface elevation profile of capillary waves at the the free surface are the difference in refractive indexes of the liquid and the air, and the fact that the interface diffuses almost no light but rather reflects or transmits incoming light completely. To our knowledge there is only one optical method that overcomes this problem with the use Polystyrene colloids to provide light scatterers within the fluid \cite{Wright96}, but the method was only used in a turbulent regime. Another powerful method for the investigation of capillary waves on ferrofluids based on x-ray absorption was presented in \cite{Richter01}, but the related costs and efforts might be justified for fully opaque liquids only. To bypass the problems associated with light refraction and reflection on a liquid-air interface we chose to study the interface between two index matched liquids. The upper fluid is transparent, the lower one is dyed. In the presence of surface deformations the instantaneous thickness can be deduced from the intensity of the light transmitted trough the colored layer. From a hydrodynamic point of view the replacement of the air by a second liquid is nothing but a change of viscosity and density, though the low kinematic viscosity of air simplifies the theoretical calculations. However, the first exact theoretical analysis of the linear stability problem by Kumar and Tuckerman \cite{kumar94} was carried out for the more general case of a system of two layers of liquid. \section*{2. Experimental setup} The experimental setup is shown in Fig.~\ref{fig1}. The container consists of an aluminium ring (diameter $D=$18 cm) seperating two parallel glass windows by a gap of 10 mm. It is filled by two unmixible liquids: a silicone oil (SOIL, Dow Corning, viscosity $\eta = 20$ mPas, density $\rho=949$ kg/m$^3$) and an aqueous solution of sugar and NiSO$_4$(WSS, $\eta = 7.2 $ mPas, $\rho=1185$ kg/m$^3$). The liquid liquid interfacial tension has been determined with the pending droplet method to $35 \pm 2$ dynes/cm. The ratio of the filling heights SOIL/WSS was 8.4/1.6. The choice of heights was made in order to obtain a variety of different patterns, including a transition from squares to hexagons \cite{binks97b}. The sugar concentration has been adapted to match the refractive index to that of the covering silicone oil ($n \simeq 1.405$) to a precision of $5 \times 10^{-4}$. The Ni$^{2+}$-ions produce a broad absorbtion band in the spectral region 600-800 nm and provide high contrast patterns projected onto the diffusive screen. The container is illuminated from below with parallel light, and a band pass filter in front of the camera was used to detect only wavelengths $\lambda = (655 \pm 5) nm$. By varying the intensity of the lamp the flat interface has been set to a level of about $50\%$ of the maximum optical transmission. At a NiSO$_4$ concentration of $17 \%$ by weight the contrast between the light intensity passing through crests and valleys of the wave pattern was optimum. The associated coefficient of optical absorption was measured as $\alpha = 5.2 \pm 0.1 cm^{-1}$. In order to avoid uncontrolled changes of the viscosity, density and interfacial tension of both liquids all the measurements were performed at a constant temperature (23$\pm$0.1$^o$C). The Faraday waves were excited by an electromagnetic shaker vibrating vertically with an acceleration in the form $a(t) = a_0$cos$(\Omega t)$. The driving signal came from a computer via a D/A-converter and the acceleration has been measured by piezoelectric sensor. A self developed closed-loop algorithm was used to suppress higher harmonics $n(\Omega t)$ in the driving signal to guarantee a purely harmonic driving. Faraday patterns were recorded in the following way: a high speed (250 Hz) 8-bit CCD camera was mounted some distance above the diffusive screen. \begin{figure} \includegraphics[ width=0.5\linewidth]{fig1.eps} \caption{Experimental setup: L - halogen lamp, Ln - lens, C - container filled by two liquids: SOIL and WSS with the same refractive indices, DS - diffusive screen, IF - interference filter, CCD - high speed CCD camera. } \label{fig1} \end{figure} Pictures were taken synchronous to the external driving. For a certain instant $t_o$ the surface elevation of the Faraday patterns $h (x,y,t_o)$ is given by: \begin{equation} h(x,y,t_o)= \frac {1}{\alpha}ln \frac{I_r(x,y)}{I_p(x,y,t_o)} \label{eqn1} \end{equation} where $I_r(x,y)$ and $I_p(x,y,t_o)$ are 2D intensity distributions captured by the camera for the reference picture (flat interface, $a_o=0$) and for the Faraday pattern ($a_o \ne 0$), respectively. Finally, the surface elevation function $h (x,y,t)$ are Fourier transformed and the time evolution of the Fourier amplitudes and phases of spatial modes is extracted. The use of a high speed camera compared to the earlier measurements by some of the co-authors \cite{Kityk04} allows for a better temporal resolution and the method is not sensitive to distortions (defects) on time scales of several periods. The logarithm of the intensity profile renders the dynamic range nonlinear, and with an 8-bit dynamical range the resolution is approximately $1\%$ ($2\%$) at small (high) surface elevations, relative to the maximal surface heights. The validity of the method has been also checked with flat layers of colored liquids of different thicknesses and the same accuracy was found. However, one should note the the Fourier transformation integrates over many pixels and a significant better resolution is to expect. \section*{3. The linear regime} The experiments have been performed by quasistatically ramping the driving amplitude for the frequencies $f=\Omega/2\pi=12,16,20,29$ and $57 Hz$ from slightly below the critical acceleration $a_c$ ($\varepsilon = (a-a_c)/a_c= -0.02)$ up to just below the acceleration where the interface disintegrates and droplets form. In the same form a ramp was driven down to check for hysteretic effects, of which none were found. For each amplitude step a series of pictures where taken and Fourier transformed. Typically the pattern occurs in the center region of the container first but evolves in a range of $\Delta\varepsilon = 0.02$. From the Fourier transformation of the pictures at $-0.02<\varepsilon < 0.1$ (Fig. \ref{squ}) the critical acceleration $a_c$ and the critical wavenumber $k_c$ has been determined (Fig. \ref{lin1}). \begin{figure} \includegraphics[ width=0.95\linewidth]{fig2.eps} \caption{Critical acceleration $a_c$ and critical wave number $k_c$ for different driving frequencies $\Omega$. The symbols mark experimental data, the lines the theoretical linear stability analysis. The size of the symbols coincide with the size of the error bars.}\label{lin1} \end{figure} The experimental data can be compared with the results from the theoretical linear stability analysis that has been performed using the algorithm proposed by Kumar and Tuckerman\cite{kumar94}. The agreement between theory and experiment is very good, similar to former studies at the liquid-air interface \cite{bechhoefer95,wagner97}. But with our new technique we are now able to verify more details of the predictions of the linear theory, e.g. the temporal spectra at onset of the instability. It is a particular feature of the Faraday-Experiment, that at onset only one wave number $k_c$ becomes unstable, but the temporal spectrum already contains multiples of the fundamental oscillation frequency $\omega$ at onset. We are in the regime of subharmonic response and the fundamental oscillation frequency at onset is always $\omega=\Omega/2$ but the spectrum also contains $(n+1/2) \Omega$ frequency components. More precisely we can write the surface deformation $h({\bf r,t})$ as \begin{equation} h({\bf r}, t)= \frac{1}{4} \, \sum_{i=1}^{N} {(A_i e^{{\rm i} {\bf k}_i \cdot {\bf r}} } + c.c.) \, \sum_{n=-\infty}^{+\infty} \zeta_n e^{{\rm i} (n+1/2) \Omega t} \label{eqn2} \end{equation} Here ${\bf r}=(x,y)$ are the horizontal coordinates. The set of \textit{complex} Fourier coefficients $\zeta_n$ are the components of the eigenvector related to the linear stability problem and determine the subharmonic time dependence. The spatial modes are characterized by the wave vectors ${\bf k}_i$ , each carrying an individual \textit{complex} amplitude $A_i$. These quantities are determined by the nonlinearities of the problem. In principle the ${\bf k}_i$ can have any length and orientation but at onset the relation $|{\bf k}_i|= k_c$ holds. The number N of participating modes determines the degree of rotational symmetry of the pattern: $N=1$ corresponds to lines, $N=2$ to squares, $N=3$ to hexagons or triangles, etc. It can be shown \cite{douady89} that the $\zeta_n$ and $\zeta_{-n}$ are coupled in a way that $\zeta_n=\zeta_{-n}$ so that heterodyning of right and left travelling waves always result in \textit{standing} waves. Equation \ref{eqn1} then reads \begin{multline} h({\bf r}, t)= \sum_{i=1}^{N} {( |A_i| cos{ {\bf k}_i \cdot {\bf r}+\phi_i} }) \\ \times \sum_{n=0}^{+\infty} {(|\zeta_n| cos{(n+1/2) \Omega t+\psi_n})} \label{eqn3} \end{multline} The complex eigenvectors $\zeta_n$ can be calculated modulo a constant factor and the ratio of the amplitudes $|\zeta_n|/|\zeta_{n+l}|$ as well as the temporal phases $\psi_n$ can be compared with experimental data. They are obtained in the following way: For each step in the driving amplitude a series of snapshots of the surface state (Fig. \ref{squ}) is taken. The primary pattern consists of squares and their formation is governed by the nonlinearities of the problem and one of our goals is to identify how far from onset the predictions from the linear theory hold. An analysis of the Fourier transformation of the pictures yield amplitudes $A(t)({\bf k}(ij))$ that are shown in Fig. \ref{12hzsq1}. For the wave vectors ${\bf k}(ij)$ the nomenclature from crystallography is used, e.g. ${\bf k}(10)$ and ${\bf k}(01)$ are the vectors that generate the simple unit cell of the square pattern (Fig. \ref{squ}). The temporal evolution of the amplitude of one of the critical modes $A(t){\bf k}(10)$ with $|{\bf k}(10)|=\textit{k}_c$ (Fig. \ref{12hzsq1} is then again Fourier transformed and a typical spectrum is shown in Fig. \ref{12hzspec}a. These data are taken for all driving strengths $\varepsilon$ (Fig. \ref{12hzbif1}a) and we always find the same values for $A(t){\bf k}(10)$ and $A(t){\bf k}(01)$ within the experimental resolution. In agreement with former investigations \cite{Wernet01} in a system with a larger aspect ratio (container size to wave length) our study reveals also that the fundamental spatial mode $|{\bf k}(10)|=\textit{k}_c$ for all $\varepsilon$. We can now extract the ratio of $A(\Omega3/2,{\bf k}(10))/A(\Omega/2,{\bf k}(10))$ that is shown in the inset of Fig. \ref{lin2}. \begin{figure} \includegraphics[ width=0.95\linewidth]{fig3.eps} \caption{The ratio of the amplitudes $A(\Omega3/2)/A(\Omega/2)$ of the ${\bf k}(10)$ mode at $\varepsilon=0$ for different driving frequencies. The values are extrapolated from measurements at $\varepsilon>0$ shown in the inset: the amplitude ratios at $\Omega/2\pi=12Hz$ and $29 Hz$ as a function of the driving strength $\varepsilon$.} \label{lin2} \end{figure} \begin{figure} \includegraphics[ width=0.95\linewidth]{fig4.eps} \caption{The temporal phases $\psi$ ($\Omega/2)$ and $\psi(3\Omega/2)$ of the ${\bf k}(10)$ mode for two driving frequencies versus driving strength $\varepsilon$. The symbols mark experimental, the lines theoretical data. Squares and broken line: $\Omega/2\pi=57 Hz$. Circles and dotted line: $\Omega/2\pi=12 Hz$; in the range $0.2<\varepsilon<0.28$ a transition from squares to hexagons takes place and in this disordered state an extraction of phases is not possible.}\label{lin3} \end{figure} The contribution of higher harmonics is in the order of $5$ to $10\%$ and increases slightly with the driving strength. The agreement between the experimental data and the linear theory is again very good up to driving strength $\varepsilon >0.5$, especially for lower driving frequencies. It is very surprising that the agreement even holds up to secondary patterned surface states, where a transition from a square to a hexagonal state has been taken place and, as we show later, strong nonlinear contributions participate in the dynamics of the system. The experimental data show also that at driving strength as low as $\varepsilon = 0.02$ the surface state consists of no measurable higher spatial Fourier modes (see Fig. \ref{12hzbif1}) but of higher temporal harmonics in perfect agreement with the linear theory. This allows an extrapolation of $A(\Omega3/2,{\bf k}(10))/A(\Omega/2,{\bf k}(10))$ to the neutral situation $\varepsilon = 0$ for all driving frequencies $\Omega$ (Fig. \ref{lin2}). The frequency ratio decreases first with increasing frequency and has a minimum at $\Omega/(2 \pi)\approx 40 Hz$. This characteristic shape reflects the amount of damping present in the system. At low driving frequencies the ration between fill hight and wave number is small. In this regime damping from the bottom, that increases with decreasing frequency, is most significant. For larger driving frequencies the damping from the bulk of the liquid (a function increasing with the frequency) is the strongest contribution. This behavior is also reflected in the critical accelerations (compare with Fig. \ref{lin1}). The ratio $A(\Omega5/2,{\bf k}(10))/A(\Omega/2,{\bf k}(10))$ has been evaluated too, but the experimental resolution is not sufficient here for a conclusive comparison between theory and experiment. In the same way the temporal phases $\psi_n$ can be extracted from the Fourier spectrum and once more a good agreement between the theoretical predictions and experimental data is obtained, at least for the fundamental $\Omega/2$ component. For the $3 \Omega/2$ component the scatter of the experimental data is very large and we find significant differences between experiment and theory. But besides the large scatter we observe a pronounced nonmonotonic behavior of the $\psi_{3/2\Omega}$ component at $57 Hz$ and this part of the spectrum seems to be governed by nonlinear interactions. \section*{4. The nonlinear surface state at $\Omega =12 Hz$} \subsection {The square state} \begin{figure} \includegraphics[ width=0.95\linewidth]{fig5.eps} \vspace{0.3cm} \caption{Snapshots of the surface state and the power spectra at $\Omega/2\pi=12Hz$ and $\varepsilon=$0.17 ($a_0=$30.0 m/s$^2$) for two different temporal phases a) at maximum and b) minimum surface elevation as indicated in Fig. \ref{12hzsq1}}. \label{squ} \end{figure} \begin{figure} \includegraphics[ width=0.95\linewidth]{fig6.eps} \vspace{-0.3cm} \caption{The absolute amplitudes $A$ of different spatial modes and the driving signal $a(t)$ in the square state at $\Omega/2\pi=12Hz$ and $\varepsilon=$0.17 ($a_0=$30.0 m/s$^2$). Please note that unlike in Ref. \cite{Kityk04} not the square root of the power spectra but the amplitude A of a deformation $h({\bf r},t)=A cos({\bf k}(ij) \cdot {\bf r})$ is shown.} \label{12hzsq1} \end{figure} The primary pattern near onset ($0<\varepsilon<0.28$) consists of squares, shown in Fig. \ref{squ}. Their formation is determined by the minimum of the Lyaponov functional of the according amplitude equation of the critical modes \cite{cross93} and a quantitative theoretical prediction of the expected pattern can be given by inspection of the cubic coupling coefficient \cite{zhang96}. To our knowledge, for a two liquid system there has not yet been an attempt to calculate this coefficient, but squares are a common pattern in free surface experiments with low viscous liquids. The amplitude equations follow from a solvability condition of a weakly nonlinear analysis of the underlying constitutive equation, and its principal form is determined by the symmetries of the system. For the subharmonic response one can write \begin{equation} \label{ampleqn1} \tau \partial_t A({\bf k}_i)= \epsilon A({\bf k}_i) - \sum_{j=1}^{N} \Gamma(\theta_{ij})|{A({\bf k}_j)}|^2 A({\bf k}_i), \end{equation} with $\tau$ the linear relaxation time and $\Gamma(\theta_{ij})$ the cubic coupling coefficient that depends on the angle $\theta_{ij}$ between the modes ${\bf k}_j$ and ${\bf k}_j$ with $|{\bf k}_{i,j}|=k_c$. The amplitudes $A({\bf k}_i)$ are modulated with a subharmonic $(n+1/2)\Omega$ time spectrum given by the $\zeta_n$ from the linear eigenvectors. Equation \ref{ampleqn1} predicts a pitchfork bifurcation and in order to study this scenario one has to extract the different temporal Fourier modes of the measured $A({\bf k}_i,t)$ (Fig. \ref{12hzsq1}) first. The result is shown in Figs. \ref{12hzspec} and \ref{12hzbif1}a. As long as the pattern consists of squares there are no harmonic time dependencies in the basic spatial modes to observe, but a continued growth of $\Omega/2$ and $3\Omega/2$ contributions. The $5\Omega/2$ contribution is very weak and only slightly larger than the noise. The square of the sum of the amplitudes $A_s=A(\Omega/2) + A(3\Omega/2)$ yields a straight line if plotted versus the driving strength $\varepsilon$ (Fig. \ref{12hzbif1}c) as one would expect for the case of a pitchfork bifurcation. From the slope we can extract the cubic coupling coefficient $A_s=\varepsilon/\Gamma(90^\circ)$, and we find $\Gamma(90^\circ)=0.179 mm^{-2}$. \begin{figure} \begin{flushleft} \includegraphics[ width=1.05\linewidth]{fig7.eps} \end{flushleft} \vspace{-0.3cm} \caption{The temporal spectra of the amplitudes $A({\bf k}(ij,\omega))$ of different spatial modes at $\Omega/2\pi=12Hz$ and a driving strength: a) $\varepsilon$=0.18 in the square state and b) $\varepsilon=$0.38 in the hexagonal state. } \label{12hzspec} \end{figure} \begin{figure} \includegraphics[ width=0.52\linewidth]{fig8a.eps} \hspace{-0.6cm} \includegraphics[ width=0.52\linewidth]{fig8b.eps} \includegraphics[ width=0.78\linewidth]{fig8c.eps} \caption{The amplitudes $A(n/2\Omega)$ of the a) ${\bf k}(10)$ and b) ${\bf k}(11)$ mode at $\Omega/2\pi=12Hz$ as a function of the driving strength $\varepsilon$. c): the square of the sum of the subharmonic components of the $A(10)$ mode versus $\varepsilon$. As expected for a forward bifurcation the data can be linearly fitted, at least up to driving strength of $\varepsilon \approx 0.1$ } \label{12hzbif1} \end{figure} Now we can inspect the next higher harmonic spatial modes $A({\bf k}(11))$ and $A({\bf k}(20))$. Their temporal evolution is shown in Fig. \ref{12hzsq1}. Both modes are a result of an interaction of two fundamental modes, ${\bf k}(11)={\bf k}(10)+{\bf k}(01)$ and ${\bf k}(20)={\bf k}(10)+{\bf k}(10)$. Quadratic coupling does not appear in the amplitude equations, but they are a natural consequence of nonlinear spatial wave interaction and it is no surprise that we find that they obey harmonic oscillations, shown in Fig. \ref{12hzspec}a. The striking result of our analysis is rather the constant offset that we find in the $A({\bf k}(11))$ and $A({\bf k}(20))$ spectrum (Fig. \ref{12hzsq1}). This means that in addition to the oscillatory part, the interfacial profile is also composed of contributions of constant deformations of the form $h({\bf r}, t)= |A_i| cos{ {\bf k}_i \cdot {\bf r}}$. This might surprisingly first, but please note that this does not violate the mass conservation. Actually, it is a simple consequence of the quadratic coupling of a \textit{real} standing surface wave oscillation, $\Re (e^{i {\bf k}_i \cdot {\bf r}} e^{i {\bf k}_i \cdot {\bf r}} e^{i\Omega/2}e^{-i\Omega/2})=cos{2 {\bf k}_i \cdot {\bf r}} (1+cos{\Omega})$ (compare also Fig. \ref{lines-ampl}c). \begin{figure} \includegraphics[ width=0.75\linewidth]{fig9.eps} \caption{The amplitudes $A({\bf k}(20))$ and $A({\bf k}(11))$ versus the square of the amplitude of the fundamental mode $A({\bf k}(10))$ or the product of $A({\bf k}(10))$ and $A({\bf k}(01))$ respectively ($\Omega/2\pi=12Hz$). } \label{A1versA2} \end{figure} This quadratic coupling scheme can be verified by plotting $A({\bf k}(20))$ and $A({\bf k}(11))$ versus the square of the amplitude of the fundamental mode $A({\bf k}(10))$ or the product $A({\bf k}(10))\times A({\bf k}(01))$ respectively. The data can be perfectly reproduced by a linear fit. From the slope one gets the strength of this nonlinear coupling and we do find the same values for all frequencies $\Omega/2\pi = 16,20,29$ Hz where squares are to be observed. Finally our Fourier analysis yields that the imaginary part of the coupling scheme obeys the same resonance conditions, and the spatial phase of the higher harmonic modes is given by $\phi({\bf k}(20))=2\phi({\bf k}(10))$ and $\phi({\bf k}(10))+\phi({\bf k}(01))$. \subsection {The hexagonal state} \begin{figure} \includegraphics[ width=0.99\linewidth]{fig10.eps} \vspace{0.3cm} \caption{Snapshots of the surface state and the Power spectra at $\Omega/2\pi=12Hz$ and $\varepsilon=$0.37 ($a_0=$39.3 m/s$^2$) for three different temporal phases. a) down hexagons, b) minimal surface elevation, c) up hexagons. See Refs. \cite{wagner00,Kityk04} for further explanations on the switch from up to down hexagons in the Faraday-Experiment. } \label{hex} \end{figure} In the range ($0.20<\varepsilon<0.28$) the pattern becomes disordered and transforms at higher driving strength to a hexagonal state (see Fig. \ref{hex}) that consists of three fundamental spatial Fourier modes ${\bf k}_{1,2,3}$. But please note that for the construction of the crystallographic simple unit cell two vectors ${\bf k}(10)={\bf k}_1$ and ${\bf k}(\bar{1}1)={\bf k}_2$ are sufficient (${\bf k}(\bar{1}1)+{\bf k}(10)={\bf k}(0\bar{1})={\bf k}_3$, as indicated in Fig. \ref{kvectorshex}). \begin{figure} \includegraphics[ width=0.5\linewidth]{fig11.eps} \caption{Vector diagram of the interacting modes for the hexagonal surface state. } \label{kvectorshex} \end{figure}\begin{figure} \includegraphics[ width=0.99\linewidth]{fig12.eps} \vspace{0.3cm} \caption{Snapshots of the surface state and the power spectra at a,b) $\Omega/2\pi=$20Hz, $\varepsilon=$0.6 ($a_0=$54.4 m/s$^2$) c,d) $\Omega/2\pi=$20Hz, $\varepsilon=$0.08 ($a_0=$36.7 m/s$^2$) e,f) $\Omega/2\pi=$57Hz, $\varepsilon=$0.11 ($a_0=$116.3 m/s$^2$). } \label{sqline} \end{figure} \begin{figure} \includegraphics[ width=0.75\linewidth]{fig13a.eps} \includegraphics[ width=0.75\linewidth]{fig13b.eps} \vspace{0.3cm} \caption{The amplitudes $A(\Omega/2)$ of the a) ${\bf k}(10)$ and ${\bf k}(01)$ mode at $\Omega/2\pi=29Hz$ and b) ${\bf k}(1)$ mode at $\Omega/2\pi=57Hz$ as a function of the driving strength $\varepsilon$. The shaded region in a) indicates the crossed rolls state.} \label{sqlinebif} \end{figure} An analysis of the temporal behavior of the amplitudes $A({\bf k}(ij))$ of the spatial modes reveals, besides the the striking offset with the according constant spatial sinusoidal surface deformation, both harmonic and subharmonic time dependencies (see Fig. \ref{12hzspec}). While harmonic ($n\Omega$) temporal contributions in the higher spatial harmonics ${\bf k}(20,11)$ appear in a similar manner to that of the square pattern, the resonance between ${\bf k}(10)+{\bf k}(0\bar{1})={\bf k}(1\bar{1})$ results in harmonic ($n\Omega$) contributions in the critical mode $|{\bf k}(1\bar{1})|=\textit{k}_c$. Consequently the $n \Omega$ contributions couple with $(n+1/2) \Omega$ contributions back into the spectrum of the ${\bf k}(20,11)$ modes and result in subharmonic contributions. Harmonic contributions do not appear in the temporal spectra of the linear unstable modes $\textit{k}_c$ and quadratic interactions of ($n \Omega/2$) components do not appear in the amplitude equations. Nevertheless the hexagonal state allows for a \textit{spatial} resonance between linear unstable modes. In other words, this means that - within the framework of the weakly nonlinear approximation - we have here the interesting case where the system has a broken temporal symmetry that is driven by spatial resonances. It is not to be observed at any point in the quadratic state, where spatial resonances between linear unstable modes are forbidden too. This particular violation of the weakly nonlinear resonance conditions can best be seen in Fig. \ref{12hzbif1} where clearly the amplitudes of the $\Omega$ components of the $|{\bf k(10)}|=\textit{k}_c$ mode grow from zero at the transition point from squares to hexagons, and similarly the $\Omega/2$ components of the $|{\bf k(11)}|=2\textit{k}_c$ modes. \section*{5. Pattern dynamics at $\Omega/2\pi> 12 Hz$} \begin{figure} \includegraphics[ width=0.95\linewidth]{fig14a.eps} \includegraphics[ width=0.95\linewidth]{fig14b.eps} \vspace{0.3cm} \caption{a,b): The absolute amplitudes $A$ of different spatial modes in the line state at $\Omega/2\pi=$57Hz and $\varepsilon=$0.11 ($a_0=$116.3 m/s$^2$). In b) the temporal constant offset $A(2k,0)$ is indicated by the dotted line. c) The amplitude $A(2k,0)$ versus $A^2(k,\Omega/2)$} \label{lines-ampl} \end{figure} The pattern dynamics at driving frequencies $\Omega/2\pi> 12 Hz$ are characterized by a transition to lines. At $\Omega/2\pi= 16 Hz$ the pattern still consists only of squares, while at $\Omega/2\pi= 20$ and $29 Hz$ the primary pattern consists of (slightly distorted) lines (Fig. \ref{sqline}c,d). At higher driving strengths $\varepsilon$, a second Fourier mode perpendicular to the first one starts to grow (Fig. \ref{sqlinebif}a) and the pattern evolves to a square state (Fig. \ref{sqline}a,b). For $\Omega/2\pi= 57 Hz$ a pure line state is stable for all driving strengths (Fig. \ref{sqline}e,f and \ref{sqlinebif}b). The pronounced constant offset in the $A(2{\bf k},t)$ (Fig. \ref{lines-ampl}b) mode is now larger than the temporal oscillation period and $A(2{\bf k},t)$ never crosses the zero line. Similar like for the $A({\bf k} 20,\Omega)$ or $A({\bf k} 11,\Omega)$ modes in the squares state this quadratic coupling scheme holds also for the zero frequency modes as shown in (Fig. \ref{lines-ampl}c). \section*{6. Conclusion} We have demonstrated a new technique to measure quantitatively the spatio-temporal Fourier spectrum of Faraday waves on a two liquid interface. With this technique it is now possible to test theoretical predictions, especially those from numerical simulations. To our knowledge there are still no full Navier Stokes numerical simulation of the 3D problem and quantitative tests for future work are most important. In this sense we would like to encourage such attempts. But with our technique we are also able to verify known predictions from the linear stability analysis and we find good agreement up to high driving strength of $\varepsilon \approx 0.5$. In the nonlinear state the most pronounced result is the identification of strong temporal constant sinusodial surface deformations in the spectrum. And with our possibility to access any Fourier component separately we can identify several resonance mechanisms, including an interesting case of a temporal resonance violation by use of spatial resonances. \begin{acknowledgments} This work was supported by the German Science Foundation project Mu 912. \end{acknowledgments}
{ "timestamp": "2005-03-10T14:37:32", "yymm": "0503", "arxiv_id": "nlin/0503023", "language": "en", "url": "https://arxiv.org/abs/nlin/0503023" }
\subsection*{Introduction} This paper presents a quantum protocol based on public key cryptogrpahy for secure transmission of data over a public channel. The security of the protocol derives from the fact that Alice and Bob each use secret keys in the multiple exchange of the qubit. Unlike the BB84 protocol [1] and its many variants (e.g. [2]-[4]), where the qubits are transmitted in only one direction and classical information exchanged thereafter, the communication in the proposed protocol remains quantum in each stage. In the BB84 protocol, each transmitted qubit is in one of four different states; in the proposed protocol, the transmitted qubit can be in any arbitrary state. \subsection*{The Protocol} Consider the arrangement of Figure 1 to transfer state $X$ from Alice to Bob. The state $X$ is one of two orthogonal states, such as $0\rangle$ and $|1\rangle$, or $\frac{1}{\sqrt2} (| 0 \rangle + | 1\rangle )$ and $\frac{1}{\sqrt2} (| 0 \rangle - | 1\rangle )$, or $\alpha |0\rangle + \beta |1\rangle$ and $\beta |0\rangle - \alpha |1\rangle$. The orthogonal states of $X$ represent $0$ and $1$ by prior mutual agreement of the parties, and this is the data or the cryptographic key being transmitted over the public channel. Alice and Bob apply secret transformations $U_A$ and $U_B$ which are commutative, i.e., $U_A U_B = U_B U_A$. An example of this would be $U_A = R(\theta)$ and $U_B = R (\phi)$, each of which is the rotation operator: \vspace{0.2in} $R(\theta) = \left[ \begin{array}{cc} cos \theta & - sin \theta \\ sin \theta & cos \theta \\ \end{array} \right]$ \vspace{0.2in} \begin{figure} \hspace*{0.2in}\centering{ \psfig{file=cryp.eps,width=10cm}} \caption{Three-stage protocol for quantum cryptography where $U_A U_B = U_B U_A$} \end{figure} \vspace{0.2in} \noindent The sequence of operations in the protocol is as follows: \begin{enumerate} \item Alice applies the transformation $U_A$ on $X$ and sends the qubit to Bob. \item Bob applies $U_B$ on the received qubit $U_A (X)$ and sends it back to Alice. \item Alice applies $U_A^\dagger$ on the received qubit, converting it to $U_B (X)$, and forwards it to Bob. \item Bob applies $U_B^\dagger$ on the qubit, converting it to $X$. \end{enumerate} At the end of the sequence, the state $X$, which was chosen by Alice and transmitted over a public channel, has reached Bob. Eve, the eavesdropper, cannot obtain any information by intercepting the transmitted qubits, although she could disrupt the exchange by replacing the transmitted qubits by her own. This can be detected by \begin{itemize} \item appending parity bits, and/or \item appending previously chosen bit sequences, which could be the destination and sending addresses or their hashed values, or some other mutually agreed sequence. \end{itemize} Since the $U$ transformations can be changed as frequently as one pleases, Eve cannot obtain any statistical clues to their nature by intercepting the qubits. \subsection*{Key distribution protocol} A related key distribution protocol is given in Figure 2. Unlike the previous case, $X$ is a fixed public state (say $|0\rangle$ or $\frac{1}{\sqrt2} (|0\rangle + |1\rangle)$). The objective is to generate a key that is a function of the transformations involved, which is not chosen in advance by either party. The protocol consists of two stages: \begin{figure} \hspace*{0.2in}\centering{ \psfig{file=cryp2.eps,width=8cm}} \caption{Key distribution protocol, where $U_A U_B = U_B U_A$.} \end{figure} \begin{enumerate} \item Alice and Bob use secret transformations, $U_A$ and $U_B$, on the known state $X$, and exchange these qubits. \item They again apply the same transformations on the received qubits, thereby each getting $U_A U_B (X)$, since $U_A U_B = U_B U_A$. It is assumed that neither Alice or Bob will measure the received qubits, and will use them as the input to a quantum register. \end{enumerate} In a variant of this scheme, two copies of the unknown state $X$ may be supplied to Alice and Bob by a key registration authority. \subsection*{Conclusion} The three-stage protocol provides perfect security in the exchange of data over a public channel under the assumptions that a separate classical protocol ensures the identity of the two parties, and errors (deliberate or random) are detected by means of parity check and confirming that a known bit sequence that was appended to the bits has arrived correctly. Since the proposed protocol does not use classical communication, it is immune to the man-in-the-middle attack on the classical communication channel which BB84 type quantum cryptography protocols suffers from [5]. On the other hand, implementation of this protocol may be harder because the qubits get exchanged multiple times. \section*{References} \begin{description} \item [1] M.A. Nielsen and I.L. Chuang, {\it Quantum Computation and Quantum Information}. Cambridge University Press, 2000. \item [2] A.K. Ekert, ``Quantum cryptography based on Bell's theorem.'' Phys. Rev. Lett., {\bf 67,} 661-663 (1991). \item [3] S. Kak, ``Quantum key distribution using three basis states.'' Pramana, {\bf 54,} 709-713 (2000); also quant-ph/9902038. \item [4] A. Poppe {\it et al}, ``Practical quantum key distribution with polarization entangled photons.'' quant-ph/0404115. \item [5] K. Svozil, ``The interlock protocol cannot save quantum cryptography from man-in-the-middle attacks.'' quant-ph/0501062. \end{description} \end{document} \end{enumerate} \end{document}
{ "timestamp": "2005-03-02T23:27:29", "yymm": "0503", "arxiv_id": "quant-ph/0503027", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503027" }
\section*{References}
{ "timestamp": "2005-03-09T13:33:26", "yymm": "0503", "arxiv_id": "quant-ph/0503087", "language": "es", "url": "https://arxiv.org/abs/quant-ph/0503087" }
\section{Introduction} In modern condensed matter research most interesting subjects are only subtle effects which can be investigated only by thorough and systematic studies of large numbers of samples. Even though first investigations have to be done by hand, a lot of time can be saved with automated measurement setups. Such automated systems are already widely used in large scale experiments, but most small laboratory experiments, even though computer controlled, do not allow for automated measurements. Automation is present, to some extent, in order to facilitate measurements in that there frequently are possibilities to have the system execute a certain measurement automatically, but covering easily large parameter spaces is often not possible. Commercially available complete measurement systems, on the other hand, seldom come with sophisticated control software without possibilities for programming long measurement sequences. Of course such software systems are the result of expensive software development which is beyond the possibilities of an ordinary research laboratory. Even though there are commercial programs available for some of the instruments that constitute an experimental setup, the important part is the interplay between them. Consequently most control software is written by the scientists themselves, who face the lack of time, money and manpower to develop extensive automation software. In this paper we present an easy way of creating control software which offers possibilities of programming complex sequences and automatically executes them\cite{sourcecode}. This is shown to be achieved with moderate development effort using a common laboratory programming language. We will first present the different architecture approach needed to achieve this goal, after which the addition of automation is a small step. \section{software for laboratory equipment} Programs created for control of experiments need to perform several tasks. Firstly, they have to be able to send control commands to the instruments and receive the measured data. Secondly, these data are to be processed and displayed and eventually user input needs to be translated to control commands. Various development platforms offer vast libraries of procedures to interface instruments, create user interfaces and perform complicated data processing. These help to reduce the workload associated with creating such software. LabVIEW\texttrademark \cite{labview}, a development environment from National Instruments\texttrademark\ for creating programs (called virtual instruments or shortly VIs) in its own graphical programming language ``G'', is probably best known and most widely used for such applications. ``G'' offers all the flow control structures like loops and conditional branches found in any other programming language. Moreover, any VI can easily be used in any other VI as a subVI. LabVIEW VIs consist of a user interface (UI) and a block diagram (BD) containing the actual code. Programming is done by modelling data flow, where graphical representations of functions and procedures are interconnected by lines, usually called wires. The designation VI stems from the similarity of such a program to an actual instrument, the UI obviously corresponding to the instrument's front panel and the BD to its internal wiring. A usual way of creating LabVIEW software for measurement control is by writing a main VI containing the UI and the logic for acting appropriately on the user input as well as processing, displaying and saving the data. Communication with the instruments is performed by driver subVIs which are regularly executed by the main VI. When such a driver VI is called to perform a query on an instrument it sends the necessary command to the instrument, waits some time for the instrument to prepare the answer and finally reads this response from the instrument. Usually this process takes tens to hundreds of milliseconds. Assuming the whole measurement setup consists of several instruments, the main VI may be organised in two different ways. Either all driver VIs are called sequentially, causing the time needed to collect all data to grow with the number of instruments. Another apprach would be to call the driver VIs in parallel, which is possible thanks to the inherently multithreading architecture of LabVIEW. In this case, however, all drivers would attempt to access the instruments at the same time. This would result in a ``traffic jam'' in case the instruments are connected to a single interface bus. Some drivers would be forced to wait until the others have finished their writing to the bus. Moreover, as some instruments take measurements less often than others, many operations on the bus would be unnecessary because no new data would be obtained. In this paper we present the use of independent driver VIs, which we call handlers, running in parallel and communicating with a main VI by means offered by LabVIEW. This allows for a more efficient use of the interface bus employed to connect the instruments and results in a higher data acquisition rate. Moreover, by employing a ``state machine'' (SM) architecture such programs become easier to extend in functionality, to maintain and most importantly allow for the control by a separate program and consequently automation. \section{experimental setup} \begin{figure} \center \includegraphics[width=80mm]{fig1} \caption{\label{torquesetup}Torque measurement setup overview, which was automated using the presented software. A cryostat is placed between the poles of an iron yoke magnet, which is freely rotatable. The torque sensor is inserted into the cryostat and connected to readout electronics. All instruments needed to control the experiment's state are connected to a personal computer.} \end{figure} The programs presented here were developed to control and automatise a torque magnetometry apparatus which was built in our group\cite{Willemin1998a,Willemin1998b}. Such a device is used to measure a sample's magnetic moment $\vect{m}$ by the torque \begin{equation} \mbox{\eulerbold{\char28}} = \mu_0 \vect{m} \times \vect{H} \end{equation} it experiences due to a magnetic field $\vect{H}$. It is well suited for investigation of anisotropic magnetic phenomena as found in most high temperature superconductors. Torque magnetometry is complementary to most other magnetometry techniques in that it is only sensitive to the part $m_\perp$ of \vect{m} perpendicular to the applied field. A torque measurement is fast --- one measurement taking a fraction of a second only --- and due to the proportionality $\tau \propto H$ reaches high sensitivities for $m_\perp$ in high fields. Our home made torque magnetometer system, shown schematically in Fig.~\ref{torquesetup}, consists of a flow cryostat between the poles of an iron yoke magnet which is sitting on a rotatable support. The torque sensor with a sample mounted on it is inserted into the cryostat and connected to a Lock-In Amplifier (LIA) for read out. Details of the measurement principle are beyond the scope of this article and are described elsewhere \cite{Willemin1998a,Willemin1998b}. All devices needed to control and measure the system's state are connected to a Windows PC via an IEEE-488 General Purpose Interface Bus (GPIB), RS-232 serial connections and indirectly via additional analog and digital input and output ports present in the LIA instrument. The main parts are the EG\&G Model 7265 LIA, a Lakeshore DRC 93A temperature controller, and the Bruker BH-15 magnetic field controller. Additional devices such as a pressure transducer with read out electronics for monitoring the exchange gas pressure in the cryostat or current sources and volt meters for specialized applications may also be connected via the GPIB. The GPIB is an interface bus which is widely used in scientific instruments. It features 8-bit parallel data transfer, handshaking and real-time response capabilities. \section{software system architecture} \begin{figure} \center \includegraphics{fig2} \caption{\label{overview}Architecture of the torque control software system. All VIs (torque.vi, dataserver.vi and the handler*.vis) execute in parallel. Commands are sent along the solid right pointing arrows and data propagates back along the dashed left pointing arrows.} \end{figure} The architecture of the newly developed control software is shown in Fig.~\ref{overview}. Each instrument connected to the system is represented by a VI counterpart called \emph{handler.vi}. All handlers are managed by the \emph{dataserver.vi} VI which communicates with the \emph{torque.vi} VI, which is the main application. All these VIs run independently in parallel. This way each \emph{handler.vi} can be optimised to take best advantage of the instrument it is built for. This includes the waiting times needed for communication, an optimized data rate based on varying needs as well as the use of each instruments ability to signal special events via the GPIB. Since all \emph{handler.vis} run in parallel, their individual write--wait--read cycles needed to talk to the instruments are interlaced, thus reducing the bus' idle time. Moreover each instrument is talked to only when necessary thus reducing the bus occupation while retaining data quality. This can be optimised particularly well by exploiting the service request (SRQ) functionality of the GPIB. Each instrument can signal a number of events to the GPIB controller by asserting the special SRQ line. Such events might be error conditions but can also be indicators of data availability. As an example the Lakeshore temperature controller is programmed to assert the SRQ line whenever a new temperature reading is ready. As this occurs only every two seconds, the instrument is read only when really necessary instead of reading the same data several times per second. Even instruments not offering such functionality can be optimised by reducing the rate at which the \emph{handler.vi} is instructed to read the instrument. This enables the more crucial measurements to be read more often resulting in data taken at a higher rate and resulting in better quality. Because the \emph{handler.vis} are not called as subVIs by the main VI a special means of communication needs to be established. Here we present the use of \emph{queues} for sending commands to the \emph{handler.vis} and \emph{DataSockets} for receiving the measured data. A \emph{queue} is a first--in--first--out style memory construct which is offered by LabVIEW. It may contain a fixed or unlimited number of string entries, in our case commands. By use of special subVIs any VI can append commands to a \emph{queue's} end or retrieve the oldest commands. Any read entry is automatically removed. \emph{Queues} are identified by a name, making access to them fairly easy. In most applications a given \emph{queue} is read by only one VI whereas several VIs may write to it. \emph{DataSockets} are memory constructs as well, identified by a unique name, but only contain the most recent datum. Their data type can be freely chosen among the data types in LabVIEW. The \emph{DataSockets} used in our case are arrays of floating point numbers containing a \emph{handler.vi's} main data. The \emph{dataserver.vi} mentioned above serves as an intermediate VI which collects all the \emph{handler.vi's} data and puts all together in a separate \emph{DataSocket} which is then read by the \emph{torque.vi} main VI. Thus the main VI needs no knowledge about which data to obtain from which instrument. \begin{figure} \center \includegraphics{fig3} \caption{\label{FIGparser}Schematic illustration of the VI's basic structure. An all enclosing main loop executes infinitely. The logic inside consists of a command stack whose first element is divided into instruction and argument. The instruction is used as the selector value into a case structure containing the code for the individual instructions. This results in the command parsing functionality needed for the operation. Internal data needed for the VI's execution is passed through each iteration and can be read and modified by each command case.} \end{figure} In order for the VIs to be able to act accordingly on the possible commands they must be given some command parsing functionality. In fact such a command parser is every VI's core part: Even the regular operations performed by the VIs are put into commands which are executed repeatedly. Essentially, all VIs are designed as command driven state machines (SM). The use of the SM paradigm in LabVIEW programs was already proposed at several occasions and given LabVIEW's capabilities this is not surprising. Nevertheless, to our knowledge only few applications make use of this architecture. The basic idea is that by being executed, a program goes through various named states. The order in which these states are visited may be fixed and defined in advance or the state to follow might be determined based on the current state's result. The implementation in LabVIEW is fairly simple and schematically shown in Fig.~\ref{FIGparser}. An infinitely running loop contains a case structure consisting of all the states. These states are identified by character strings and are therefore easily human readable. In contrast to other methods, where the identification is by numbers or special enumeration data types, this makes the structure easy to extend and maintain. Additionally to these structures the VIs contain a command stack and some internal data needed for execution. Upon startup, when the command stack is empty, a default case (state) is executed. Usually this is the ``GetCommands'' case. This case contains the code needed to empty this VI's \emph{queue} and a set of default commands which are put onto the command stack. When the main loop is iterated for the second time, the oldest command is removed from the stack, split into an instruction and optional arguments, whereupon this instruction is fed into the case structure selector, defining the case to be executed. This case may add more commands to the stack or simply perform a specific task. When the case is finished, the main loop iterates again, the next command is removed from the stack and so on. Whenever the stack becomes empty, the default case ``GetCommands'' is executed again and refills it. Because the \emph{handler.vis} are independent programs not having to rely on being called regularly by a master VI they can be used to carry out more complex tasks than just talking to the instruments. As an example \emph{handlerLakeshore.vi}, the \emph{handler.vi} for the Lakeshore temperature controller contains logic to control the temperature by software through control of the coolant flow in the cryostat. The flow controller is connected to a separate digital--to--analog converter (DAC), thus enabling the \emph{handlerLakeshore.vi} to control it by sending commands to the DAC's \emph{handler.vi} (\emph{handlerDAC.vi}). Keeping track of the last few seconds of measured data, calculating their time trends and publishing it to the \emph{DataSocket} is coded into a command and performed by the \emph{handler.vis} as well. \section{Automation} As mentioned earlier, all VIs are organized as state machines, even the main VI \emph{torque.vi}. As shown in Fig.~\ref{automation} every user action (button press, value change) on its user interface (UI) is transformed into a command by the UI-handler which is then sent to and processed in the SM. The SM then sends appropriate commands to the \emph{dataserver.vi} and the \emph{handler.vis} (wide arrow (1) in Fig.~\ref{automation}). These two parts (UI-handler and SM) are independently running components of \emph{torque.vi}. The communication between them is again ensured via \emph{queues}. This enables other VIs, such as the \emph{sequencer.vi} shown in Fig.~\ref{automation} to be used to control the SM in \emph{torque.vi} programmatically by sending these commands directly to the SM (wide arrow (2) in Fig.~\ref{automation}). \begin{figure} \center \includegraphics{fig4} \caption{\label{automation}All VIs consist of a User interface (UI) and a block diagram (BD). In contrast to all other VIs the \emph{torque.vi's} BD consists of the UI-handler part and the state machine (SM) itself, both running in parallel. In normal, interactive operation of the torque system, user actions on the \emph{torque.vi's} UI are translated by the UI-handler into commands which are sent to the SM via a queue and then propagate on to the dataserver and handler VIs (wide arrow (1)). If an automated measurement is run, the \emph{sequencer.vi's} SM retrieves commands from the text sequence on its UI, sends them via a queue to the \emph{torque.vi's} SM from where they propagate on to the dataserver and handler VIs (wide arrow (2)). The \emph{torque.vi's} SM sends confirmation messages back to the \emph{sequencer.vi}. Solid black arrows indicate direct access between the BD and the UI, whereas dotted arrows represent data transmission via \emph{queues} and \emph{DataSockets}.} \end{figure} When automatic measurements are required, a sequence text file is written containing the commands needed to accomplish these measurements which is then read by the \emph{sequencer.vi}. Additionally to the commands of \emph{torque.vi's} SM the \emph{sequencer.vi} understands a set of flow control instructions such as ``if'', ``while'' and ``for'' which are useful for creating short sequences for repetitive tasks, as well as the use of variables and their arithmetic manipulation and comparison. The \emph{sequencer.vi} parses through the sequence file by looking for known keywords -- the commands. Any strings which are not recognized as a keyword are treated as arguments to the preceding keyword. The string \texttt{settemp 20 waittemp} present in a sequence file would instruct the torque software to change the temperature to 20\,K and wait for the cryostat to stabilize at this temperature. In this example \texttt{settemp} and \texttt{waittemp} are keywords and \texttt{20} is the argument to the keyword \texttt{settemp}. Such sequencing possibilities are already well known in control software of commercially available measurement equipment (eg.\ SQUID magnetometers or the \emph{Quantum Design} Physical Property Measurement System\cite{Quantum}). Now such efficient and flexible data taking is also possible with our home made torque magnetometer. \section{Example of Application} In order to demonstrate the possibilities of such an automatable measurement system we present some results of a systematic study\cite{Kohout2005} of the so called lock-in transition in the high temperature superconductor La$_{2-x}$Sr$_{x}$CuO$_{4}$. Details about this effect can be obtained from various other sources and are not discussed here\cite{Blatter1994,Steinmeyer1994}. Most easily this effect is visible in angle dependent torque measurements and manifests itself as a deviation from an otherwise smooth behaviour. An example of such a measurement is shown in Fig.~\ref{torque1}, where the measured data points close to 90$^\circ$ deviate from a theoretical curve\cite{kogan} which fits well to the remaining angle range. The same model can also be used to describe data taken as a function of magnetic field magnitude $H$ at a fixed angle. It is commonly accepted that in first approximation the magnetic moment $m=\tau/H$ of a superconductor is proportional to $\ln(H)$. \begin{figure} \center \includegraphics{fig5} \caption{\label{torque1}Angle dependent torque measurement (circles) of an underdoped crystal of La$_{2-x}$Sr$_{x}$CuO$_{4}$\ with $x=0.07$ ($\ensuremath{T_\mathrm{c}}=17\U{K}$), performed at $T=8\U{K}$ in a magnetic field $\mu_0H=1\U{T}$. The solid line is a fit of a model derived by Kogan\cite{kogan}. The deviation close to $\theta \approx 90\ensuremath{^\circ}$ stems from the lock-in transition.} \end{figure} Within our study we measured six La$_{2-x}$Sr$_{x}$CuO$_{4}$\ single microcrystals with varying Sr content $0.07 \le x \le 0.23$ and critical temperatures \ensuremath{T_\mathrm{c}}\ varying from 17\U{K} to 35\U{K}. They were mounted on a highly sensitive torque sensor and cooled below \ensuremath{T_\mathrm{c}}. Field dependent measurements ($\mu_0H=0\ldots1.5\U{T}$ at $5\U{mT}$ steps with increasing and decreasing field) were taken at 60 field orientations ($\theta=-90\ensuremath{^\circ}\ldots90\ensuremath{^\circ}$ with varying steps) and at about ten temperatures below the critical temperature \ensuremath{T_\mathrm{c}}. We emphasize that such extensive measurements would hardly be possible without our software's automation possibilities. As each field scan takes about six minutes, without automation user interaction would be necessary at this interval during \emph{one week} to collect all these data for one crystal. After writing the sequence and starting its execution, the measurement system, on the other hand, finishes such a measurement set within about \emph{three days} with no need of intervention. The experiment is finished faster, because less time is lost between consecutive field scans and because the measurement is running day and night. \begin{figure} \center \includegraphics{fig6} \caption{\label{torque2}Field dependent measurement $\tau(H)$ of the same La$_{2-x}$Sr$_{x}$CuO$_{4}$\ crystal as was used for the measurement in Fig.\ \ref{torque1}. The angle of the magnetic field was fixed at $\theta=75\ensuremath{^\circ}$ and $\theta=80\ensuremath{^\circ}$. The measurements are plotted as $\tau/H$ vs.\ $\ln(H)$. The lines are guides to the eye to show the two linear regions I (low field) and II (high field).} \end{figure} We present here only one dataset of a single crystal taken at one particular temperature. Such a dataset consists of 60 field scans taken at various orientations. The two field scans shown in Fig.~\ref{torque2} illustrate the deviations of field dependent data due to the lock-in transition. Clearly visible are two regions (I and II) where $\tau/H$ is proportional to $\ln(H)$. A comparison of these measurements to angle dependent measurements at similar conditions indicate that region I corresponds to the part where lock-in takes place, whereas data in region II are well described by the theoretical curve in Fig.~\ref{torque1}. By analysing the whole data set it is now easy to investigate the evolution of these two regions as a function of angle $\theta$. The result is shown in Fig.~\ref{torque3}, where the extents of the two regions, obtained from field dependent measurements, are plotted vs.\ the angle $\theta$. The horizontal line A indicates the cut of the measurement in Fig.~\ref{torque1} and the vertical lines B and C the measurements shown in Fig.~\ref{torque2}. The observed region, separating regions I and II manifests the lock-in transition and can be understood in terms of a model proposed by Feinberg and Villard\cite{Feinberg}. \begin{figure} \center \includegraphics{fig7} \caption{\label{torque3}Summary of field dependent measurements performed on a La$_{2-x}$Sr$_{x}$CuO$_{4}$\ single crystal at $T=8\U{K}$. Only the extents of the linear regions such as shown in Fig.\ \ref{torque2} as a function of field orientation $\theta$ are shown. The enhancement of the low-field region I close to the ab-plane ($\theta \approx 90\ensuremath{^\circ}$) is clearly visible. The horizontal line A indicates the position of the measurement shown in Fig.~\ref{torque1}. The vertical lines B and C indicate the position of the measurements shown in Fig.~\ref{torque2}.} \end{figure} \section{Acknowledgements} This work was supported in part by the Swiss National Science Foundation. \newpage
{ "timestamp": "2005-03-01T11:16:07", "yymm": "0503", "arxiv_id": "physics/0503005", "language": "en", "url": "https://arxiv.org/abs/physics/0503005" }
\section{Introduction.} In quantum probability there exist several natural notions of independence, see \cite{muraki03} and the references therein. These allow to define new convolutions for probability measures, cf.\ \cite{voiculescu+dykema+nica92,voiculescu97,speicher+woroudi93,muraki00}. Bercovici \cite{bercovici04} defined multiplicative monotone convolutions for probability measures on the unit circle and on the half line. He showed that with an appropriate function of the Cauchy transform these multiplicative convolutions can be calculated by composition of those functions, similar to Muraki's result \cite[Theorem 3.1]{muraki00} for the additive monotone convolution. In this paper we give a new proof of Bercovici's result based on the combinatorics of moments, see Theorem \ref{thm-operators}. Using Berkson and Porta's \cite{berkson+porta78} characterization of composition semigroups, one can deduce a characterization of continuous convolution semigroups for the monotone convolution, see \cite[Theorem 4.6]{bercovici04} or Theorem \ref{thm-levy-khintchine-circle} for the case of probability measures on the unit circle. This paper is organized as follows. In Section \ref{sec-mon} we recall the definition of monotone independence and the monotone product of algebraic and quantum probability spaces. In Section \ref{sec-mon-cond} we show that the monotone product is actually a special case of the conditionally free product introduced in \cite{bozejko+speicher91b,bozejko+leinert+speicher96}. Sections \ref{sec-op}, \ref{sec-conv}, and \ref{sec-levy} contain the main results on the multiplicative monotone convolution. We formulate a slightly modified version of a theorem by Bercovici that shows that these convolutions can be calculated by taking the composition of appropriate functions of the Cauchy transform of the measures, see Theorem \ref{thm-operators} and Corollaries \ref{cor-unitary} and \ref{cor-pos}. We also state a L\'evy-Khintchine type characterization of all continuous convolution semigroups for the monotone convolution of probability measures on the unit circle, see Theorem \ref{thm-levy-khintchine-circle}. In Section \ref{sec-galton}, we show that the problem of embedding a probability measure on the unit circle into a continuous monotone convolution semigroup is very similar to the problem of embedding a discrete-time Markovian branching process (or Galton-Watson process) into a continuous-time Markovian branching process. In Section \ref{sec-embed} we adapt a characterization of embeddable branching processes due to Gorya\u{\i}nov \cite{goryainov93} to our situation. Finally, in the Appendix we discuss the multiplicative monotone convolution of probability measures on the half line and show that there exist two natural, but inequivalent definitions. One of them is equivalent to the definition due to Bercovici and can be treated by similar methods as the multiplicative monotone convolution of measures on the unit circle., cf.\ \cite{bercovici04}. \section{Monotone Independence.}\label{sec-mon} In this section we present the definition of monotone independence and its main properties. By an {\em algebraic probability space} we mean a pair $(\mathcal{A},\varphi)$ consisting of a unital algebra $\mathcal{A}$ and a unital functional $\varphi:\mathcal{A}\to\mathbb{C}$. Assume that we have two algebraic probability spaces $(\mathcal{A}_1,\varphi_1)$ and $(\mathcal{A}_2,\varphi_2)$, such that the first algebra has a decomposition $\mathcal{A}_1=\mathbb{C}\mathbf{1}\oplus\mathcal{A}_1^0$ (direct sum as vector spaces), where $\mathcal{A}_1^0$ is a subalgebra of $\mathcal{A}_1$. Then we define the algebraic monotone product $(\mathcal{A},\varphi)$ of $(\mathcal{A}_1,\varphi_1)$ and $(\mathcal{A}_2,\varphi_2)$ as follows, see also \cite{muraki01,muraki03}. The algebra $\mathcal{A}=\mathcal{A}_1\coprod \mathcal{A}_2$ is the free product of $\mathcal{A}_1$ and $\mathcal{A}_2$ with identification of the units of $\mathcal{A}_1$ and $\mathcal{A}_2$. The unital functional $\varphi=\varphi_1\triangleright\varphi_2:\mathcal{A}\to\mathbb{C}$ is determined by the condition \begin{equation}\label{def-alg-mon} \varphi(b_1a_1b_2\cdots a_{n-1}b_n)=\varphi_1(a_1\cdots a_{n-1})\varphi_2(b_1)\cdots \varphi_2(b_n) \end{equation} for $n\in\mathbb{N}$ and all $a_1,\ldots,a_{n-1}\in\mathcal{A}_1^0$, $b_1,\ldots,b_n\in\mathcal{A}_2$. Let now $\mathcal{A}_1,\mathcal{A}_2\subseteq\mathcal{B}$ be two such algebras, which are contained in an algebraic probability space $(\mathcal{B},\Phi)$ and denote by $j_1:\mathcal{A}_1\to\mathcal{B}$, $j_2:\mathcal{A}_2\to\mathcal{B}$ the inclusion maps. Then the universal property of the free product of algebras implies that there exists a unique homomorphism $j:\mathcal{A}_1\coprod\mathcal{A}_2\to B$ such that the following diagram commutes \[ \xymatrix{ & \mathcal{B} & \\ \mathcal{A}_1\ar[ur]^{j_1}\ar[r]_{i_1} & \mathcal{A}_1\coprod\mathcal{A}_2 \ar[u]|-j & \mathcal{A}_2\ar[ul]_{j_2}\ar[l]^{i_2} } \] where are $i_1:\mathcal{A}_1\to\mathcal{A}_1\coprod\mathcal{A}_2$ and $i_2:\mathcal{A}_2\to\mathcal{A}_1\coprod\mathcal{A}_2$ are the canonical inclusion maps. The subalgebras $\mathcal{A}_1,\mathcal{A}_2$ are called {\em monotonically independent} w.r.t.\ $\Phi$, if \[ \Phi\circ j = (\Phi\circ j_1)\triangleright (\Phi\circ j_2) \] cf.\ \cite{franz02} We will call a triple $(\mathcal{A},\mathcal{H},\Omega)$ consisting of a Hilbert space $\mathcal{H}$, a unit vector $\Omega\in\mathcal{H}$, and a subalgebra $\mathcal{A}\subseteq\mathcal{B}(\mathcal{H})$ a {\em quantum probability space}. If we have an algebraic probability space $(\mathcal{A},\varphi)$, whose algebra has an involution such that $\Phi$ is even a state, and if for all $a\in\mathcal{A}$ there exists a constant $C_a\ge 0$ such that the inequality \[ \Phi(x^*a^*ax)\le C_a\Phi(x^*x) \] holds for all $x\in\mathcal{A}$, then the GNS representation $(H_\varphi,\pi_\varphi,\Omega_\varphi)$ of $(\mathcal{A},\Phi)$ yields a quantum probability space $(\pi_\varphi(\mathcal{A}),H_\varphi,\Omega_\varphi)$. If two subalgebras $\mathcal{A}_1=\mathbb{C}\mathbf{1}\otimes\mathcal{A}_1^0,\mathcal{A}_2\subseteq\mathcal{A}$ are monotonically independent in $(\mathcal{A},\varphi)$, then $\pi_\varphi(\mathcal{A}_1^0)$ and $\pi_\varphi(\mathcal{A}_2)$ are monotonically independent in $(\pi_\varphi(\mathcal{A}),H_\varphi,\Omega_\varphi)$ in the sense of the following definition. \begin{definition}\label{def-mon-indep} Let $\mathcal{H}$ be a Hilbert space, $\Omega\in\mathcal{H}$ a unit vector, and define a state $\Phi:\mathcal{B}(\mathcal{H})\to\mathbb{C}$ on the algebra of bounded operators on $\mathcal{H}$ by \[ \Phi(X)=\langle \Omega, X\Omega\rangle, \qquad\mbox{ for } X\in\mathcal{B}(\mathcal{H}). \] Two subalgebras $\mathcal{A}_1,\mathcal{A}_2\subseteq\mathcal{B}(\mathcal{H})$ are called {\em monotonically independent} w.r.t.\ $\Omega$, if the following two conditions are satisfied. \begin{description} \item[(a)] For all $X,Z\in\mathcal{A}_1$, $Y\in\mathcal{A}_2$, we have \[ XYZ = \Phi(Y)XZ. \] \item[(b)] For all $Y\in\mathcal{A}_1$, $X,Z\in\mathcal{A}_{2}$, \[ \Phi(XYZ)=\Phi(X)\Phi(Y)\Phi(Z). \] \end{description} Two operators $X,Y\in\mathcal{B}(\mathcal{H})$ are called monotonically independent w.r.t.\ $\Omega$, if the subalgebras $\mathcal{A}_1={\rm alg}(X)={\rm span}\{X^k|k=1,2,\ldots\}$ and $\mathcal{A}_2={\rm alg}(Y)={\rm span}\{Y^k|k=1,2,\ldots\}$ are monotonically independent. \end{definition} \begin{proposition}\label{prop-mon-prod} Let $(\mathcal{A}_i,\mathcal{H}_i,\Omega_i)$, $i=1,2$, be two quantum probability spaces, and denote the states associated to $\Omega_1$ and $\Omega_2$ by $\Phi_1$ and $\Phi_2$, respectively. Then there exists a quantum probability space $(\mathcal{A},\mathcal{H},\Omega)$ and two injective state-preser\-ving homomorphisms $J_i:\mathcal{A}_i\to\mathcal{A}$, $i=1,2$, such that the images $J_1(\mathcal{A}_1)$ and $J_2(\mathcal{A}_2)$ are monotonically independent w.r.t.\ $\Omega$. \end{proposition} \Proof We set $\mathcal{H}=\mathcal{H}_1\otimes\mathcal{H}_2$ and $\Omega=\Omega_1\otimes\Omega_2$. Denote by $P_2$ the orthogonal projection on $\mathbb{C}\Omega_2\subseteq\mathcal{H}_2$. We define the embeddings $J_i:\mathcal{A}_i\to\mathcal{B}(\mathcal{H})$ by \begin{eqnarray*} J_1(X) &=& X\otimes P_2, \qquad \mbox{ for } X\in\mathcal{A}_1, \\ J_2(X) &=& \mathbf{1}\otimes X, \qquad \mbox{ for } X\in\mathcal{A}_2. \end{eqnarray*} For $\mathcal{A}$ we take the subalgebra generated by $J_1(\mathcal{A}_1)$ and $J_2(\mathcal{A}_2)$. It is clear that $J_1$ and $J_2$ are injective, state-preserving homomorphisms. A simple calculation shows that $J_1(\mathcal{A}_1)$ and $J_2(\mathcal{A}_2)$ are monotonically independent w.r.t.\ $\Omega$. E.g., for products of the form $J_1(X_1)J_2(Y)J_1(X_2)$, $X_1,X_2\in\mathcal{A}_1$, $Y\in\mathcal{A}_2$, we get \begin{eqnarray*} J_1(X_1)J_2(Y)J_1(X_2) &=& (X_1\otimes P_2)(\mathbf{1}\otimes Y)(X_1\otimes P_2) = (X_1X_2)\otimes P_2YP_2 \\ &=& \Phi\big(J_2(Y)\big) J_1(X_1)J_1(X_2). \end{eqnarray*} On the other hand, for $J_2(Y_1)J_1(X)J_2(Y_2)$, $X\in\mathcal{A}_1$, $Y_1,Y_2\in\mathcal{A}_2$, we get \begin{eqnarray*} \Phi\big(J_2(Y_1)J_1(X)J_2(Y_2)\big) &=& \langle\Omega_1\otimes\Omega_2,(\mathbf{1}\otimes Y_1)(X\otimes P_2)(\mathbf{1}\otimes Y_2)\Omega_1\otimes\Omega_2\rangle \\ &=& \langle\Omega_1\otimes\Omega_2,X\otimes (Y_1PY_2)\Omega_1\otimes\Omega_2\rangle \\ &=& \Phi_1(X)\Phi_2(Y_1)\Phi_2(Y_2) =\Phi\big(J_2(Y_1)\big)\Phi\big(J_1(X)\big)\Phi\big(J_2(Y_2)\big). \end{eqnarray*} \endproof We will call the quantum probability space $(\mathcal{A},\mathcal{H},\Omega)$ constructed in the previous proposition the {\em monotone product} of $(\mathcal{A}_1,\mathcal{H}_1,\Omega_1)$ and $(\mathcal{A}_2,\mathcal{H}_2,\Omega_2)$. When there is no danger of confusion, we shall identify the algebras $\mathcal{A}_1$ and $\mathcal{A}_2$ with their images $J_1(\mathcal{A}_1)$ and $J_2(\mathcal{A}_2)$, respectively. The monotone product is associative and can be extended to more than two factors, see \cite{franz01}. But it is not commutative. The embedding $J_1:\mathcal{A}_1\to\mathcal{A}$ is not unital and the product is not trace-preserving. If $\Phi_1|_{\mathcal{A}_1}$ is not identically equal to zero, then the calculation \[ \Phi_1(X)\Phi_2(Y_1Y_2)=\Phi(XY_1Y_2)=\Phi(Y_2XY_1)=\Phi_1(X)\Phi_2(Y_1)\Phi_2(Y_2) \] for all $X\in\mathcal{A}_1$, $Y_1,Y_2\in\mathcal{A}_2$ shows that $\Phi$ can only be a trace on $\mathcal{A}$, if $\Phi_2|_{\mathcal{A}_2}$ is a homomorphism. \section{Relation of monotone independence and conditional free independence.}\label{sec-mon-cond} We recall now the definition of the conditional free product of algebraic probability spaces and show that the monotone product is contained as a special case. Let $(\mathcal{A}_1,\varphi_1,\psi_1)$ and $(\mathcal{A}_2,\varphi_2,\psi_2)$ be two unital algebras, equipped with two unital functionals. Recall that the conditionally free product\cite{bozejko+speicher91b,bozejko+leinert+speicher96} of $(\mathcal{A}_1,\varphi_1,\psi_1)$ and $(\mathcal{A}_2,\varphi_2,\psi_2)$ is defined as the triple $(\mathcal{A},\varphi,\psi)$, where $\mathcal{A}=\mathcal{A}_1\coprod \mathcal{A}_2$ is the free product of $\mathcal{A}_1$ and $\mathcal{A}_2$ with identification the units of $\mathcal{A}_1$ and $\mathcal{A}_2$. The unital functionals $\varphi$ and $\psi$ on $\mathcal{A}=\mathcal{A}_1\coprod \mathcal{A}_2$ can be defined by the conditions \begin{equation}\label{def-cond-free} \varphi(a_1a_2\cdots a_n)=\varphi_{\epsilon(1)}(a_1)\cdots\varphi_{\epsilon(n)}(a_n) \quad \mbox{ and }\quad \psi(a_1a_2\cdots a_n)=0 \end{equation} for all $n\in \mathbb{N}$ and all $a_i\in\mathcal{A}_{\epsilon(i)}$ with $\epsilon(i)\in\{1,2\}$, $\epsilon(1)\not=\epsilon(2)\not=\cdots\not=\epsilon(n)$ and $\psi_{\epsilon(1)}(a_1)=\cdots=\psi_{\epsilon(n)}(a_n)=0$. The functional $\psi$ is simply the free product $\psi_1*\psi_2$ of $\psi_1$ and $\psi_2$, cf.\ \cite{voiculescu+dykema+nica92,voiculescu97}. We will denote $\varphi$ by \[ \varphi=\varphi_1\,{}_{\psi_1}\kern-.5em*\kern-.3em{}_{\psi_2}\,\varphi_2. \] The product defined in this way for triples $(\mathcal{A},\varphi,\psi)$ can be shown to be commutative and associative, cf. \cite{bozejko+speicher91b,bozejko+leinert+speicher96}. Taking pairs of the form $(\mathcal{A}_1,\varphi_1,\varphi_1)$ and $(\mathcal{A}_2,\varphi_2,\varphi_2)$, one obtains the free product also for the first functional, i.e. \[ \varphi_1\,{}_{\varphi_1}\kern-.5em*\kern-.3em{}_{\varphi_2}\,\varphi_2 = \varphi_1*\varphi_2. \] Suppose now that the algebras $\mathcal{A}_1$ and $\mathcal{A}_2$ have decompositions $\mathcal{A}_i=\mathbb{C}\mathbf{1}\oplus\mathcal{A}_i^0$, $i=1,2$, as a direct sum of vector spaces, such that the $\mathcal{A}_i^0$ are even subalgebras. If one defines functionals $\delta_i:\mathcal{A}_i\to\mathbb{C}$ by \begin{equation}\label{delta} \delta_i(\lambda\mathbf{1}+a_0)=\lambda \end{equation} for $\lambda\in\mathbb{C}$, $a_0\in\mathcal{A}_i^0$, $i=1,2$, then one obtains the boolean product \[ \varphi_1\,{}_{\delta_1}\kern-.5em*\kern-.3em{}_{\delta_2}\,\varphi_2=\varphi_1\diamond\varphi_2, \] cf.\ \cite{speicher+woroudi93,bozejko+leinert+speicher96}. Since the conditionally free product of triples of the form $(\mathcal{A},\varphi,\delta)$ can be shown to be again of the same form, the commutativity and associativity of the boolean product follow immediately from this construction. One can also obtain the monotone product from the conditionally free product. \begin{proposition} Let $(\mathcal{A}_1,\varphi_1)$ and $(\mathcal{A}_2,\varphi_2)$ be two algebraic quantum probability spaces and assume $\mathcal{A}_1$ has a decomposition $\mathcal{A}_1=\mathbb{C}\mathbf{1}\oplus\mathcal{A}_1^0$, where $\mathcal{A}_1^0$ is a subalgebra of $\mathcal{A}_1$. Define a unital functional $\delta_1:\mathcal{A}_1\to\mathbb{C}$ as in Equation (\ref{delta}). Then we have \[ \varphi_1\triangleright\varphi_2=\varphi_1\,{}_{\delta_1}\kern-.5em*\kern-.3em{}_{\varphi_2}\,\varphi_2 \] \end{proposition} \Proof Let $n\in\mathbb{N}$, $\epsilon(1),\ldots,\epsilon(n)\in\{1,2\}$ such that $\epsilon(1)\not=\epsilon(2)\not=\cdots\not=\epsilon(n)$, and $a_1\in\mathcal{A}_{\epsilon(1)},\cdots,a_n\in\mathcal{A}_{\epsilon(n)}$ such that $\delta_1(a_k)=0$ if $\epsilon(k)=1$ and $\varphi_2(a_k)=0$ if $\epsilon(k)=2$. This implies $a_k\in\mathcal{A}_1^0$ for $\epsilon(k)=1$ and therefore by Equation (\ref{def-alg-mon}) \[ \varphi_1\triangleright\varphi_2(a_1a_2\cdots a_n)=\prod_{k:\epsilon(k)=2} \varphi_2(a_k)=0 \] (If the product $a_1a_2\cdots a_n$ does not begin or end with an element of $\mathcal{A}_2$, add $\mathbf{1}\in\mathcal{A}_2$ in order to apply Equation (\ref{def-alg-mon})). Therefore $\varphi_1\triangleright\varphi_2$ satisfies condition (\ref{def-cond-free}) that defines the conditionally free product $\varphi_1\,{}_{\delta_1}\kern-.5em*\kern-.3em{}_{\varphi_2}\,\varphi_2$. \endproof With this observation, Muraki's formula \cite[Theorem 3.1]{muraki00} for the additive monotone convolution can be deduced from the analytic theory of the additive conditionally free convolution developed in \cite{bozejko+leinert+speicher96}. \section{Products of monotonically independent operators.}\label{sec-op} For a bounded operator $X$ in a quantum probability space $(\mathcal{B}(\mathcal{H}),\mathcal{H},\Omega)$ we define \[ \psi_X(z)=\left\langle \Omega, \frac{zX}{1-zX} \Omega\right\rangle \] and \[ K_X(z)= \frac{\psi_X(z)}{1+\psi_X(z)} \] for $|z|<1/||X||$. The following theorem is similar to \cite[Theorem 2.2]{bercovici04}. Below we provide a new proof. \begin{theorem}\label{thm-operators} Let $(B(\mathcal{H}),\mathcal{H},\Omega)$ be a quantum probability space and $\mathcal{A}_1,\mathcal{A}_2 \subseteq B(\mathcal{H})$ two monotonically independent subalgebras. Let $V_1,V_2\in \mathbb{C}\mathbf{1}+\mathcal{A}_1$, such that $V_2V_1-\mathbf{1}\in \mathcal{A}_1$ and $W\in\mathcal{A}_2$. Then we have \[ K_{V_1WV_2}(z)=K_{V_1V_2}\big(K_W(z)\big) \] for all $|z|<\min(1/||V_1WV_2||,1/||W||)$. \end{theorem} \Proof Let $M=\max\big(||V_1WV_2||,||W||(||V_1V_2||+2)\big)$ and $|z|<1/M$. Then we have \begin{eqnarray*} \frac{zV_1WV_2}{1-zV_1WV_2} &=& \sum_{n=1}^\infty (zV_1WV_2)^n = \sum_{n=1}^\infty z^n V_1 \underbrace{W (X+\mathbf{1}) W \cdots W(X+\mathbf{1})} W V_2 \\ &&\hspace{55mm}n-1 \mbox{ times} \\ &=& \sum_{n=1}^\infty z^n \sum_{k=1}^n \sum_{{\nu_1,\ldots,\nu_k\ge 1}\atop{\nu_1+\cdots+\nu_k=n}} V_1 W^{\nu_1} X W^{\nu_2} X \cdots X W^{\nu_k} V_2, \end{eqnarray*} where $X=V_2V_1-\mathbf{1}$. Using properties (a) and (b) in Definition \ref{def-mon-indep}, we get \begin{eqnarray*} \psi_{V_1WV_2}(z) &=& \left\langle\Omega, \frac{zV_1WV_2}{1-zV_1WV_2}\Omega\right\rangle \\ &=& \sum_{n=1}^\infty z^n \sum_{k=1}^n \sum_{{\nu_1,\ldots,\nu_k\ge 1}\atop{\nu_1+\cdots+\nu_k=n}}\left\langle\Omega,V_1X^{k-1}V_2\Omega\right\rangle\left\langle\Omega,W^{\nu_1}\Omega\right\rangle\cdots\left\langle\Omega,W^{\nu_k}\Omega\right\rangle \\ &=& \sum_{k=1}^\infty \left\langle\Omega,V_1(V_2V_1-\mathbf{1})^{k-1}V_2\Omega\right\rangle \big(\psi_W(z)\big)^k \\ &=& \sum_{k=1}^\infty \left\langle\Omega,V_1V_2(V_1V_2-\mathbf{1})^{k-1}\Omega\right\rangle \big(\psi_W(z)\big)^k \\ &=& \sum_{k=1}^\infty\left\langle\Omega,\psi_W(z)V_1V_2\frac{1}{\mathbf{1}-\psi_W(z)(V_1V_2-\mathbf{1})}\Omega\right\rangle \\ &=&\sum_{k=1}^\infty\left\langle\Omega,\frac{\frac{\psi_W(z)}{1+\psi_W(z)}V_1V_2}{\mathbf{1}-\frac{\psi_W(z)}{1+\psi_W(z)}V_1V_2}\Omega\right\rangle = \psi_{V_1V_2}\big(K_W(z)\big). \end{eqnarray*} By uniqueness of analytic continuation, we get \[ K_{V_1WV_2}(z)=K_{V_1V_2}\big(K_W(z)\big) \] for all $|z|<\min(1/||V_1WV_2||,1/||W||)$. \endproof \begin{corollary}\label{cor-unitary} Let $U,V$ be two unitary operators such that $U-\mathbf{1}$ and $V$ are monotonically independent with respect to $\Omega$. Then we have \[ K_{UV}(z)=K_{VU}(z)=K_U\big(K_V(z)\big) \] for all $|z|\in\mathbb{D}=\{z\in\mathbb{C}:|z|<1\}$. \end{corollary} \begin{corollary}\label{cor-pos} Let $X,Y$ be two positive operators such that $X-\mathbf{1}$ and $Y$ are monotonically independent with respect to $\Omega$. Then we have \[ K_{\sqrt{X}Y\sqrt{X}}(z)=K_X\big(K_Y(z)\big) \] for all $|z|<\min(1/||\sqrt{X}Y\sqrt{X}||,1/||Y||)$. \end{corollary} \section{Multiplicative monotone convolution for probability measures on the unit circle.}\label{sec-conv} For a probability measure $\mu$ on $S^1$ we define \[ \psi_\mu(z)=\int_{S^1} \frac{zx}{1-zx}{\rm d}\mu(x) \quad \mbox{ and }\quad K_\mu(z)=\frac{\psi_\mu(z)}{1+\psi_\mu(z)} \] for $z\in\mathbb{D}=\{z\in\mathbb{C}:|z|<1\}$. We will call $K_\mu$ the {\em K-transform} of $\mu$, it characterizes the measure $\mu$ completely. Furthermore, for a holomorphic function $K:\mathbb{D}\to\mathbb{D}$ there exists a probability measure $\mu$ on the unit circle $S^1$ such that $K=K_\mu$ if and only if $K(0)=0$. This follows from the Herglotz representation theorem, the proof is similar to \cite[Proposition 3.3]{franz04}. It is clear that the composition of two K-transforms is again a K-transform of some probability measure on the unit circle. In view of Corollary \ref{cor-unitary} this suggests the following definition. \begin{definition} Let $\mu,\nu$ be two probability measures on $S^1$, with transforms $K_\mu$ and $K_\nu$. Then the unique probability measure $\mu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\nu$ on $S^1$ with \[ K_{\mu\kern0.17em\lower0.1ex\hbox{\rule{0.025em}{0.43em}}\kern-.105em\gtrdot\nu}=K_\mu\circ K_\nu \] is called the {\em monotone convolution} of $\mu$ and $\nu$. \end{definition} \begin{remark} \begin{enumerate} \item The monotone convolution is weakly continuous. \item The monotone convolution is associative, i.e. \[ (\lambda\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\mu)\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\nu=\lambda\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot(\mu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\nu) \] for all $\lambda,\mu,\nu$, but not commutative, i.e., in general $\mu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\nu\not=\nu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\mu$. \item The Dirac measure $\delta_1$ at $1$ is a two-sided unit, $\delta_1\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\mu=\mu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\delta_1=\mu$ for all $\mu$. Right convolution by a Dirac measure $\delta_x$ acts as translation, i.e.\ $\mu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\delta_x=T_x\mu$, where $T_x:S^1\to S^1$ is defined by $T_x(y)=xy$ for $x\in S^1$. But $\delta_x\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\mu\not=T_x\mu$ in general. \item The monotone convolution is affine in the first argument. Together with weak continuity this implies the following formula \[ \mu\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\nu = \int_{S^1} {\rm d}\mu(x) \delta_x\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\nu. \] \end{enumerate} \end{remark} \section{L\'evy-Khintchine formula for monotone convolution semigroups.}\label{sec-levy} We call a weakly continuous one-parameter family $(\mu_t)_{t\ge0}$ of probability measures on the unit circle a continuous monotone convolution semigroup, if \[ \mu_0=\delta_1 \qquad \mbox{ and } \qquad \mu_s\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\mu_t=\mu_{s+t} \] for all $s,t\ge 0$. By definition a one-parameter family $(\mu_t)_{t\ge}$ is a continuous monotone convolution semigroup if and only if the $K$-transforms $K_t=K_{\mu_t}$, $t\ge 0$ form a continuous semigroup w.r.t.\ to composition. The continuity of the $K$-transforms is uniform in $z$ on compact sets. Our main tool for characterizing continuous monotone convolution semigroups will be Berkson and Porta's \cite{berkson+porta78} characterisation of composition semigroups of holomorphic maps. \begin{theorem}\label{thm-levy-khintchine-circle} \cite[Theorem 4.6]{bercovici04} Let $(\mu_t)_{t\ge 0}$ be a weakly continuous family of probability measures on the unit circle, with K-transforms $(K_t)_{t\ge 0}$. Then the following are equivalent. \begin{description} \item[(a)] $(\mu_t)_{t\ge 0}$ is a continuous monotone convolution semigroup. \item[(b)] $(K_t)_{t\ge 0}$ is a continuous semigroups w.r.t.\ to composition. \item[(c)] There exists a holomorphic function $u:\mathbb{D}\to\mathbb{C}$ with $\Re\, u(z)\ge 0$ for $z\in\mathbb{D}$ such that $(K_t)_{t\ge 0}$ is the (unique) solution of \[ \frac{{\rm d}K_t(z)}{{\rm d}t} = -K_t(z)u\big(K_t(z)\big) \] for $z\in\mathbb{D}$ and $t\ge 0$, with initial condition $K_0(z)=z$. \end{description} \end{theorem} \Proof The equivalence between (a) and (b) follows from the definition and the continuity properties of the monotone convolution. The equivalence between (b) and (c) is an immediate consequence of \cite[Theorem (3.3)]{berkson+porta78}, it suffices to identify the fixed point at zero as the Denjoy-Wolff point of the $K_t$. \endproof \begin{remark} \begin{enumerate} \item The function $u$ in (c) can be computed from the derivative of $(K_t)_{t\ge 0}$ in $t=0$ by \[ u(z) = -\frac{1}{z}\left.\frac{{\rm d}}{{\rm d}t}\right|_{t=0} K_t(z), \] we will call it the {\em generator} of $(K_t)_{t\ge 0}$. \item Such a function $u$ has a unique Herglotz representation \[ u(z)=ib + \int_{S^1} \frac{w+z}{w-z}{\rm d}\rho(w), \] where $b$ is a real number and $\rho$ a finite measure on $S^1$. \end{enumerate} \end{remark} \section{Relation to Galton-Watson processes.}\label{sec-galton} A probability measure $\mu$ on the unit circle is called infinitely divisible w.r.t.\ to the monotone convolution, if for all $n\in\mathbb{N}$ there exists a probability measure $\mu_n$ on the unit circle such that \begin{eqnarray*} \mu&=&\underbrace{\mu_n\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\cdots\kern0.3em\rule{0.04em}{0.52em}\kern-.35em\gtrdot\mu_n}. \\ && \qquad n\mbox{ times} \end{eqnarray*} Bercovici has shown in \cite[Theorem 4.7]{bercovici04} that all infinitely divisible probability measures can be embedded into a continuous monotone convolution semigroup, i.e.\ if $\mu$ is infinitely divisible w.r.t.\ to the monotone convolution, then there exists a continuous monotone convolution semigroup $(\mu_t)_{t\ge 0}$ such that $\mu=\mu_t$ for some $t\ge0$. And from the previous section it is clear this implies that the K-transform $K_\mu$ can be embedded into a continuous composition semigroup of K-transforms. A similar problem has been studied in the theory of Galton-Watson processes. Let $X_{n,k}$, $n,k=1,2,\ldots$ be independent, identically distributed random variables with values in $\mathbb{N}$ with generating function \[ \varphi(z)=\mathbb{E}(z^{X_{n,k}})=\sum_{m=0}^\infty p_m z^m,\quad \text{ for }z\in\mathbb{D} \] where $p_m=\mathbb{P}(X_{n,k}=m)$. Then the associated Galton-Watson process $(Y_n)_{n\ge 0}$ is defined by $Y_0=1$, and \[ Y_{n+1}=\sum_{k=1}^{Y_n} X_{n,k}, \qquad \mbox{ for } n\ge 1. \] This process describes the evolution of a population where after each step each individual produces a random number of offspring according to the probabilities $(p_m)_{m\ge 0}$. Its generating functions form a discrete composition semigroup, \[ \mathbb{E}(z^{Y_n})=\varphi^n(z),\qquad \text{ for }z\in\mathbb{D},\quad n\in\mathbb{N}. \] If $\mathbb{P}(X_{n,k}=0)=0$ (i.e.\ no individual dies without offspring), then $\varphi(0)=0$ and $\varphi$ is the $K$-transform of a probability measure $\mu$ on $S^1$. If $(Y_n)_{n\ge 0}$ can be embedded into a continuous-time Markovian branching process (or equivalently, if $(\varphi^n)_{n\ge 0}$ can be embedded into a continuous composition semigroup $(\varphi_t)_{t\ge0}$ of generating functions), then $\mu$ is infinitely divisible for the monotone convolution and can be embedded into a continuous monotone convolution semigroup. The problem of embedding Galton-Watson processes has been studied by Gorya\u{\i}nov\cite{goryainov93,goryainov00}. \begin{example} Continuous-time Markovian branching processes with extinction probability $0$ can be obtained by choosing infinitesimal offspring probabilities $\lambda_j\ge 0$ for $j\ge 2$ such that $\alpha=\sum_{j=2}^\infty \lambda_j<\infty$, setting \[ v(z)=\sum_{j=2}^\infty \lambda_jz^j-\alpha z, \qquad \mbox{ for } |z|\le 1, \] and solving the differential equation \[ \frac{{\rm d}}{{\rm d}t}\varphi_t(z) = v\big(\varphi_t(z)\big) \] with initial condition $\varphi_0(z)=z$, cf.\ \cite[Theorem 4]{goryainov93}. A simple example is the Yule process, where $v(z)=\alpha(z^k-z)$ and \[ \varphi_t(z) =\frac{ze^{-\alpha t}}{\sqrt[k-1]{1-\left(1-e^{-\alpha(k-1)t}\right)z^{k-1}\,}}, \qquad t\ge 0, \] for some $k\in\mathbb{N}$, $k\ge 2$. This process describes a population were the individuals are replaced by $k$ new individuals after an exponentially distributed random time. \end{example} \section{On the embedding of probability measures into continuous monotone convolution semigroups.}\label{sec-embed} \cite[Theorem 6]{goryainov93} and \cite[Theorem 7]{goryainov93} characterize probability generating functions that can be embedded into composition semigroups of probability generating functions. In this section we give a similar characterization for K-transforms of probability measures on the unit circle that can be embedded into continuous monotone convolution semigroups. Let $(K_t)_{t\ge 0}$ be a continuous composition semigroups of K-transforms. By \cite{berkson+porta78}, $K_t$ is differentiable w.r.t.\ $t$ and satisfies the differential equation \begin{equation}\label{eq-4} \frac{{\rm d}}{{\rm d}t} K_t(z) = v\big(K_t(z)\big) = v(z)K'_t(z) \end{equation} for $t\ge 0$, $z\in\mathbb{D}$, with $v$ given by \[ v(z)=\left.\frac{{\rm d}}{{\rm d}t}\right|_{t=0} K_t(z). \] This equation follows from the semigroup property $K_{s+t}=K_s\circ K_t=K_t\circ K_s$ by differentiation w.r.t.\ $s$ at $s=0$. By Theorem \ref{thm-levy-khintchine-circle}, the function $v$ is of the form $v(z)=-zu(z)$, with a holomorphic function $u:\mathbb{D}\to\mathbb{C}$ such that $\Re\,u(z)\ge 0$ for $z\in\mathbb{D}$. We will need the following lemma. \begin{lemma}\label{lemma-unique} Let $u:\mathbb{D}\to\mathbb{C}$, $u\not\equiv0$, be a holomorphic function such that $\Re\,u(z)\ge 0$ for $z\in\mathbb{D}$ and set $\beta=u(0)$, $v(z)=-zu(z)$ for $z\in\mathbb{D}$. Then, for all $t\ge 0$, the equation \begin{equation}\label{eq-5} v\big(f(z)\big) = v(z)f'(z), \qquad z\in\mathbb{D}, \end{equation} has a unique solution $f$ with $f'(0)=e^{-t\beta}$. \end{lemma} \Proof The proof of this lemma is borrowed from \cite[Lemma 2]{goryainov93}. Let $(K_t)_{t\ge 0}$ be a composition semigroup of K-transforms with generator $u$. Then all $K_t$, $t\ge 0$ satisfy Equation (\ref{eq-5}). Furthermore, the differential equation that the $K_t$ satisfy, implies \[ \frac{{\rm d}}{{\rm d}t}K'_t(0)=-u(0)K'_z(0) \] and therefore $K'_t(0)=e^{-t\beta}$, since $K_0(z)=z$ and $K'_0(0)=1$. This proves existence. Let now $f$ be an arbitrary solution of Equation (\ref{eq-5}) with $f'(0)=e^{-t\beta}$. Since $v$ has no zeros inside $\mathbb{D}$ other than $z=0$, we get $f(0)=0$ by substituting $z=0$ into Equation (\ref{eq-5}). Differentiation Equation (\ref{eq-5}) $k$ times, we can calculate the higher derivatives of $f$ at zero from $f'(0)=e^{-t\beta}$ and the derivatives of $v$ at zero. This proves uniqueness. \endproof \begin{remark} Let $(K_t)_{t\ge0}$ be the K-transforms of a continuous monotone convolution semigroup $(\mu_t)_{t\ge 0}$ with generator $u$. Then $K'_t(0)=e^{-tu(0)}$ is the first moment of $\mu_t$, i.e.\ \[ e^{-tu(0)}=\int_{S^1} x {\rm d}\mu_t, \qquad \mbox{ for }t\ge 0. \] \end{remark} We come to the main result of this section. \begin{theorem} Let $\mu$ be a probability measure on the unit circle $S^1$ that is not concentrated in one point. Then $\mu$ can be embedded into a continuous monotone convolution semigroup if and only if $K'_\mu(z)\not=0$ for all $z\in\mathbb{D}$ and there exists a locally uniform limit \[ \lim_{n\to\infty} -\frac{K^n_\mu(z)}{(K^n_\mu)'(z)}=v(z), \] in $\mathbb{D}$ that is of the form $v(z)=\alpha z u(z)$ with a non-zero constant $\alpha\in\mathbb{C}$ and a holomorphic function $u:\mathbb{D}\to\mathbb{C}$ such that $\Re\,u(z)\ge 0$ for $z\in\mathbb{D}$ and $K'_\mu(0)=e^{-t_0u(0)}$ for some some $t_0\ge 0$. \end{theorem} \Proof The proof of this theorem is similar to that of \cite[Theorem 6]{goryainov93}. Suppose that $\mu$ can be embedded into a continuous monotone convolution semigroup. Then $K_{\mu}$ can be embedded into a composition semigroup of K-transforms $(K_t)_{t\ge 0}$. Therefore all $K_t$ are injective and $K'_t(z)\not=0$ for all $z\in\mathbb{D}$, $t\ge0$, cf.\ \cite{berkson+porta78}. Denote by $u$ the generator of $(K_t)_{t\ge 0}$ and define $v$ by $v(z)=-zu(z)$ for $z\in\mathbb{D}$. By the Denjoy-Wolff theorem we get $\lim_{t\to\infty}K_t(z)=0$ and $\lim_{t\to\infty}K'_t(z)=0$ locally uniformly for all $z\in\mathbb{D}$. Therefore \[ \lim_{t\to \infty} \frac{v\big(K_t(u)\big)}{K_t(z)}=v'(0)=-u(0). \] With the right-hand-side of Equation (\ref{eq-4}) this implies \[ \lim_{n\to\infty} -\frac{K^n_\mu(z)}{(K^n_\mu)'(z)}=\lim_{t\to\infty} -\frac{K_t(z)}{K'_t(z)} =\lim_{t\to\infty}-\frac{K_t(z)v(z)}{v\big(K_t(z)\big)}= -\frac{v(z)}{v'(0)}=-z\frac{u(z)}{u(0)}. \] The limit is of the form required in the theorem with the constant $\alpha=-1/u(0)$. To show the converse, let now $K_\mu$ be a K-transform satisfying the conditions of the theorem with $v(z)=\alpha z u(z)$, $\alpha$ and $u$ as described in the theorem. Let $(K_t)_{t\ge 0}$ be the composition semigroup of K-transforms with generator $u$. Then the $K_t$ satisfy \[ v\big(K_t(z)\big) = v(z)K'_t(z), \qquad \mbox{ for } t\ge 0, \quad z\in\mathbb{D}. \] The conditions of the theorem imply that $K_\mu$ is also a solution of the same equation, \[ v(z)=\lim_{t\to\infty} -\frac{K_\mu^{n+1}(z)}{(K_\mu^{n+1})'(z)} =\lim_{t\to\infty} -\frac{K_\mu^n\big(K_\mu(z)\big)}{K'_\mu(z)(K_\mu^n)'(z)} = \frac{v\big(K_\mu(z)\big)}{K'_\mu(z)}. \] The uniqueness in Lemma \ref{lemma-unique} now implies $K_\mu=K_{t_0}$. \endproof \begin{remark} Let $\mu=\delta_x$ be concentrated in one point $x=e^{i\varphi}\in S^1$. Then we get $\psi_{\delta_x}(z)=\frac{xz}{1-xz}$ and $K_\mu(z)=e^{i\varphi}z$ and $\mu$ can be embedded into the continuous convolution semigroups $(\mu_t^{(k)})_{t\ge 0}$ given by $\mu^{(k)}_t=\delta_{e^{it(\varphi+2\pi k)}}$, $k\in\mathbb{Z}$. \end{remark} \section{Appendix: Multiplicative monotone convolution for probability measures on $\mathbb{R}_+$.} Just as there are many different ways to define multiplicatively a positive operator from two given positive operators, there are different definitions of multiplicative monotone convolutions of two probability measures $\mu$ and $\nu$ on $\mathbb{R}_+$. Two possible choices are to take positive self-adjoint operators $X$ and $Y$, whose distributions are given by $\mu$ and $\nu$, resp., such that $X-\mathbf{1}$ and $Y-\mathbf{1}$ are monotonically independent, and to define the convolution of $\mu$ and $\nu$ as the distributions of $\sqrt{X}Y\sqrt{X}$ or $\sqrt{Y}X\sqrt{Y}$. By Corollary \ref{cor-pos} the K-transform of $\sqrt{X}Y\sqrt{X}$ is equal to the composition of the K-transforms of $X$ and $Y$. Therefore this gives a definition which is equivalent to the one chosen by Bercovici, cf.\ \cite{bercovici04}. We will show below that choosing the distribution of $\sqrt{Y}X\sqrt{Y}$ as the convolution of the distributions of $X$ and $Y$ leads to an inequivalent definition. The operators $\sqrt{X}Y\sqrt{X}$ and $\sqrt{Y}X\sqrt{Y}$ have the same spectrum, except for $0$. More precisely, $\sigma(\sqrt{X}Y\sqrt{X})\backslash\{0\}=\sigma(\sqrt{Y}X\sqrt{Y})\backslash\{0\}$, since $\sqrt{X}Y\sqrt{X}=AB$ and $\sqrt{Y}X\sqrt{Y}=BA$ with $A=\sqrt{X}\sqrt{Y}$ and $B=\sqrt{Y}\sqrt{X}$. But the following example shows that, unlike in the free case where one works with tracial states, here the distributions of $\sqrt{X}Y\sqrt{X}$ and $\sqrt{Y}X\sqrt{Y}$ are in general different and therefore we have two different multiplicative monotone convolutions for probability measures on $\mathbb{R}_+$. \begin{example} Consider the positive definite $2\times 2$-matrix \[ M(a)=\left(\begin{array}{cc} 1 & a \\ a & 1\end{array}\right)=\frac{1}{\sqrt{2}}\left(\begin{array}{cc} 1 & 1 \\ 1 & -1 \end{array}\right)\left(\begin{array}{cc} 1+a & 0 \\ 0 & 1-a \end{array}\right)\frac{1}{\sqrt{2}}\left(\begin{array}{cc} 1 & 1 \\ 1 & -1 \end{array}\right), \] with $a\in(0,1)$. Then we have \[ \left\langle\left(\begin{array}{c} 0 \\ 1 \end{array}\right), A^k \left(\begin{array}{c} 0 \\ 1 \end{array}\right)\right\rangle = \frac{1}{2}\left((1-a)^k+(1+a)^k\right) \] for $k\in\mathbb{N}$, i.e.\ the distribution of $A$ in the vector state given by $\omega=\left(\begin{array}{c} 0 \\ 1 \end{array}\right)$ is equal to $\frac{1}{2}(\delta_{1-a}+\delta_{1+a})$. A simple calculation yields \begin{equation}\label{sqrt} \sqrt{M(a)}=\frac{1}{2}\left(\begin{array}{ccc} \sqrt{1+a}+\sqrt{1-a} & & \sqrt{1+a}-\sqrt{1-a} \\ \sqrt{1+a}-\sqrt{1-a} & & \sqrt{1+a}+\sqrt{1-a} \end{array}\right). \end{equation} Let $a,b\in(0,1)$ and consider the pair of positive definite matrices \begin{eqnarray*} X&=&\mathbf{1}\otimes\mathbf{1}+\big(M(a)-\mathbf{1}\big)\otimes P_\omega = \left(\begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & a \\ 0 & 0 & a & 1 \end{array}\right), \\ Y&=&\mathbf{1}\otimes M(b) = \left(\begin{array}{cccc} 1 & 0 & b & 0 \\ 0 & 1 & 0 & b \\ b & 0 & 1 & 0 \\ 0 & b & 0 & 1 \end{array}\right), \end{eqnarray*} in $\mathcal{M}_2(\mathbb{C})\otimes \mathcal{M}_2(\mathbb{C})\cong\mathcal{M}_4(\mathbb{C})$ where $P_\omega$ denotes the orthogonal projection onto $\omega=\left(\begin{array}{c} 0 \\ 1 \end{array}\right)$. With respect to the vector state given by $\omega\otimes \omega$, $X-\mathbf{1}\otimes\mathbf{1}$ and $Y-\mathbf{1}\otimes\mathbf{1}$ are monotonically independent, with distributions given by $\frac{1}{2}(\delta_{1-a}+\delta_{1+a})$ and $\frac{1}{2}(\delta_{1-b}+\delta_{1+b})$, respectively. As in Equation (\ref{sqrt}), we compute \begin{eqnarray*} \sqrt{X} &=& \left(\begin{array}{ccccccc} 1 & & 0 & & 0 & & 0 \\ 0 & & 1 & & 0 & & 0 \\ 0 & & 0 & & \frac{\sqrt{1+a}+\sqrt{1-a}}{2} & & \frac{\sqrt{1+a}-\sqrt{1-a}}{2} \\ 0 & & 0 & & \frac{\sqrt{1+a}-\sqrt{1-a}}{2} & & \frac{\sqrt{1+a}+\sqrt{1-a}}{2}\end{array}\right), \\ \sqrt{Y} &=& \frac{1}{2}\left(\begin{array}{ccccccc} \sqrt{1+b}+\sqrt{1-b} & & 0 & & \sqrt{1+b}-\sqrt{1-b} & & 0 \\ 0 & & \sqrt{1+b}+\sqrt{1-b} & & 0 & & \sqrt{1+b}-\sqrt{1-b} \\ \sqrt{1+b}-\sqrt{1-b} & & 0 & & \sqrt{1+b}+\sqrt{1-b} & & 0 \\ 0 & & \sqrt{1+b}-\sqrt{1-b} & & 0 & & \sqrt{1+b}+\sqrt{1-b}\end{array}\right). \end{eqnarray*} The eigenvalues of both $\sqrt{X}Y\sqrt{X}$ and $\sqrt{X}Y\sqrt{X}$ are \begin{eqnarray*} \lambda_1 &=& 1+\frac{a}{2}+\frac{1}{2}\sqrt{a^2+4(1+a)b^2}, \\ \lambda_2 &=& 1+\frac{a}{2}-\frac{1}{2}\sqrt{a^2+4(1+a)b^2}, \\ \lambda_3 &=& 1-\frac{a}{2}+\frac{1}{2}\sqrt{a^2+4(1-a)b^2}, \\ \lambda_4 &=& 1-\frac{a}{2}-\frac{1}{2}\sqrt{a^2+4(1-a)b^2}, \end{eqnarray*} and therefore their distributions have the same support. But their distributions in the vector state $\omega=\left(\begin{array}{c} 0 \\ 0 \\ 0 \\ 1 \end{array}\right)$ are different. For example their second moments differ, \begin{eqnarray*} \langle\omega,\left(\sqrt{X}Y\sqrt{X}\right)^2\omega\rangle &=& 1+b^2+a^2, \\ \langle\omega,\left(\sqrt{Y}X\sqrt{Y}\right)^2\omega\rangle &=& 1+b^2+\frac{a^2}{2}\left(1+\sqrt{1-b^2}\right), \end{eqnarray*} (recall that we assumed $a\not=0$, $b\not=0$). \end{example} \section*{Acknowledgements.} I presented the results of this paper, in particular Theorems \ref{thm-operators} and \ref{thm-levy-khintchine-circle} at the conference ``Quantum Probability and Infinite Dimensional Analysis'' in B\c{e}dlewo, Poland, in June 2004. I wish to thank Marek Bo{\.z}ejko and Janusz Wysoczanski who indicated Reference \cite{bercovici04} to me. I am also indebted to W.\ Hazod for bringing Gorya\u{\i}nov's work \cite{goryainov93,goryainov00} to my attention.
{ "timestamp": "2005-03-25T16:28:22", "yymm": "0503", "arxiv_id": "math/0503602", "language": "en", "url": "https://arxiv.org/abs/math/0503602" }
\section{INTRODUCTION} In recent years entanglement has been recognized as a physical resource central to quantum information processing. As a result, a remarkable research effort has been devoted to classifying and quantifying it. The first achievement in this direction was the identification of the entropy of entanglement \cite{bennett1}, $E_{E}$, as the unique measure of entanglement for pure bipartite states in the asymptotic limit. It was shown that that $m$ copies of a pure state $\ket{\psi}$ can be reversibly converted into $n$ copies of $\ket{\phi}$ by local operations and classical communication (LOCC) if, and only if, $mE_{E}(\ket{\psi}) = nE_{E}(\ket{\phi})$. This reversibility is lost however when one considers the more general picture of mixed states. In this case two different entanglement measures, associated with the formation and distillation processes, respectively, have to be taken into account. On one hand the entanglement cost, $E_{C}(\rho)$ \cite{bennett1}, is the minimal number of singlets necessary to create the state $\rho$ by LOCC in the asymptotic regime. On the other, the distillable entanglement, $E_{D}(\rho)$ \cite{bennett1}, is the maximum number of singlets that can be extracted by LOCC from $\rho$. Another important measure connected to asymptotic properties is the relative entropy of entanglement, $E_{R}$ \cite{vedral1}. It is related to how distinguishable an entangled state is from a separable one and gives bounds to $E_{C}$ and $E_{F}$. The finite copy case is more complex and the entropic quantities considered above are not applicable anymore. For bipartite pure states, where the reversibility is already lost, the minimum set of entanglement measures characterizing deterministic and probabilistic transformations were derived \cite{nielsen1,vidal1,jonathan1}. The mixed case, however, is known only for very restricted situations and remains mainly unsolved. Another approach for the quantification of entanglement is to measure the usefulness of a state to perform a given quantum information task. For example, the maximal fidelity of teleportation achieved by single copy LOCC \cite{horodecki1}, the maximal secret-key rate attainable by local measurements in a cryptographic protocol \cite{curty1} and the capacity of dense coding \cite{bruss1}, despite not being equal to any of the measures discussed so far, are clearly the best quantifiers when one of these protocols is analyzed. Entanglement in multi-partite systems exhibits a much richer structure than the bipartite case and its study is even more challenging. Already in the pure three qubits case, there are two different manners for a state to be entangled, in the sense that there are states that cannot be converted, even with a certain probability, in each other \cite{dur1}. From the measures considered above only $E_{R}$ is unambiguously defined to multi-partite systems, although it is not the only one. It is thus clear that entanglement is a highly complex phenomenon, which cannot be quantified by only one measure. Then, a natural way to measure it is to use any quantie which satisfies some particular properties, being the monotonicity under LOCC the most important \cite{vidal2,vedral1}. In this axiomatic approach any measure which does not increase, on average, under LOCC, called an entanglement monotone \cite{vidal2}, is a good measure of entanglement and, conversely, any meaningful quantifier has to be an entanglement monotone, or at least has some sort of weaker monotonicity under LOCC. A closely related problem to the quantification problem is the characterization of entanglement. The very fundamental question whether a given mixed state is entangled or not is extremely difficult, being actually NP-HARD \cite{gurvits}. A possible approach is then to consider sufficient criteria for entanglement, such as the Peres-Horodecki \cite{peres} and the alligment \cite{chen} tests. Nonetheless, the strongest manner to characterize entanglement is using entanglement witnesses (EW) \cite{horodecki2,terhal1}. They are Hermitian operators whose expectation value is positive in every separable state. Therefore, a negative expectation value in a measurement of a witness operator in an arbitrary state is a direct indication of entanglement in this state. Furthermore, it was shown by the Horodeckis that a state is entangled if, and only if, it is detected by an EW \cite{horodecki2}. A great deal of research has been devoted to the study of EWs, varying from the their classification and optimization \cite{lewenstein1,terhal2,terhal1} to their use in the characterization of entanglement in important, even macroscopic, physical systems \cite{toth1,brukner1,wu1}. Also optimal set-ups for local measurements of witnesses \cite{toth2,guhne1} and experimental realizations of witnessing entanglement were realized \cite{bou}. In spite of the determination of EWs for all states being also computationally intractable \cite{brandao1}, different methods from convex optimization theory can be applied to the problem, leading to efficient approximative procedures to determine and even optimize EWs for arbitrary states \cite{brandao1,doherty1,eisert1}. The main objective of this paper is to show that EWs can be very helpful also in the quantification of entanglement. The first measure related to EWs was due to Bertlmann {\it et al} \cite{bert} and was shown to be equal to the Hilbert-Schmidt distance from the set of separable states. Brandao and Vianna \cite{brandao2} took another significant step in this direction, showing that a measure derived from optimal EWs of the most studied group of witnesses so far, the group of EWs with unit trace, was in fact equal to the random robustness, which led to the establishment of properties still unknown for the later, such as its monotonicity under separable trace-preserving superoperators. Besides the obvious benefit of increasing the number of entanglement measures known, EWs based quantifiers are particular interesting due to the possibility of performing experimental measurements of them, which could be important to the extension of entanglement to other areas of physics, such as thermodynamics and statistical mechanics. Moreover, despite being necessary in general a complete tomography of a state to the determination of its degree of entanglement based on an EW measure, any EW provides a lower bound to it, even when no information at all about the state is available. The paper is structured as follows. In Sec. II we briefly review the basic properties of multi-partite optimal entanglement witnesses. In Sec. III we define a class of entanglement measures based on EWs, which includes several important already known quantities such as the negativity and the concurrence, and introduce a new infinite family of entanglement monotones having the generalized robustness and the best separable approximation measure as its limits. In Sec IV we present further properties of the considered measures and relate them to the localizable entanglement. In Sec. V it is shown how the methods developed in the last years to the characterization of entanglement based on convex optimization can be used to calculate approximately a large number of measures based on EWs. In Sec. VI possible extensions of our approach to Gaussian states are discussed. In Sec. VII we consider how the measures and their calculation are modified in states with symmetries. In Sec. VIII the questions of the amount of entanglement and of nonlocality in the presence of a super-selection rule are answered from the perspective of the studied measures. In Sec. IX it is shown that the three most successful approaches to the quantification of entanglement in systems of indistinguishable particles can be easily accessed from the EWs based quanties. In Sec. X and XI the questions of bounds on the teleportation distance and on the distillable entanglement of a given quantum state is review using our measures. It is shown that they provide sharper bounds than the negativity for the majority of states. In Sec. XII we derive lower bounds to the entanglement of formation with any EW. Possible applications of the measures are exemplified in Sec. XIII, where the derivation of two thermodynamic \textit{equations of state} which take into account entanglement is presented. Finally, in Sec. XIV we summarize our results and discuss future perspectives. \section{MULTIPARTITE SYSTEMS AND OPTIMAL ENTANGLEMENT WITNESSES} We consider a system shared by $N$ parties ${\cal f} A_{i} {\cal g}_{i=1}^{N}$. Following \cite{dur2}, we call a $k-$partite split a partition of the system into $k \leq N$ sets ${\cal f} S_{i} {\cal g}_{i=1}^{k}$, where each may be composed of several original parties. Given a density operator $\rho_{1...k} \in {\cal B}(H_{1} \otimes ... \otimes H_{k})$ (the Hilbert space of bounded operators acting on $H_{1}\otimes ... \otimes H_{k}$) associated with some $k-$partite split, we say that $\rho_{1...k}$ is a $m$-separable state if it is possible to find a convex decomposition for it such that in each pure state term at most $m$ parties are entangled among each other, but not with any member of the other group of $n - m$ parties. For example, every 1-separable state, also called fully-separable, can be written as: \begin{equation} \rho_{1...k} = \sum_{i}p_{i} \ket{\psi_{i}}_{1}\bra{\psi_{i}}\otimes ... \otimes \ket{\psi_{i}}_{k}\bra{\psi_{i}}. \end{equation} Another example is the 2-separable states of a 3-partite split given by: \begin{equation} \rho_{1:2:3} = \sum_{i}p_{i}\rho_{i}, \end{equation} where each $\rho_{i}$ is separable with respect to at least one of the three possible partitions (A:BC, AB:C and AC:B). For each kind of separable state there is a different kind of entanglement associated to it. We will say that a state is $(m+1)$-partite entangled if it is not $m$-separable. It is clear that if a state is $m$-separable it cannot be $n$-entangled for all $n > m$. It is possible to detect ($m$+1)-partite entanglement using entanglement witnesses. In order to do that, consider the index set $P = {\cal f}1, 2, ..., k{\cal g}$. Let $P^{m}$ be a subset of $P$ which has at most $m$ elements. Then $W$ is a $(m+1)$-partite entanglement witness if \begin{center} \begin{equation} \begin{array}{c} _{P^{m}_{v}}\bra{\psi}\otimes ... \otimes \hspace{0.07 cm}_{P^{m}_{1}}\bra{\psi}W\ket{\psi}_{P^{m}_{1}} \otimes ... \otimes \ket{\psi}_{P^{m}_{v}} \geq 0 \\ \\ \forall \hspace{0.2 cm} P^{m}_{1}, ..., P^{m}_{v} \hspace{0.2 cm} $such that$ \\ \\ \bigcup_{k=1}^{v}P^{m}_{k} = P \hspace{0.2 cm} $and$ \hspace{0.2 cm} P^{m}_{k} \bigcap P^{m}_{l} = {\cal f}{\cal g}. \end{array} \end{equation} \end{center} Equation $(3)$ assures that the operator $W$ is positive for all $m$-separable states. Thus, as the subspace of $m$-separable density operators is convex and closed, a state $\rho$ is $(m+1)$-entangled if and only if there is a Hermitian operator satisfying equation (3) such that $Tr(W\rho) < 0$ \cite{horodecki3}. Usually one is interest in a selected group of witnesses operators called {\it optimal}. Two different definitions of optimal entanglement witness (OEW) exist. The first, introduced by Lewenstein {\it et al} \cite{lewenstein1}, is based on how much entangled states a given entanglement witness (EW) $W$ is able to detect: $W$ is optimal iff there is no other EW which detects all the states detected by $W$ and some other states not detected by $W$. The second definition, due to Terhal \cite{terhal2}, establish the concept of OEW relative to a chosen entangled state $\rho$. The $\rho-$optimal entanglement witness $W_{\rho}$ is given by \begin{equation} Tr(W_{\rho}\rho) = \min_{W \in {\cal M}} Tr(W\rho), \end{equation} where ${\cal M}$ is the intersection of the set of entanglement witnesses, denoted by ${\cal W}$, with some other set ${\cal C}$ such that ${\cal M}$ is compact \cite{foot1}. Note that every $\rho$-OEW is also an OEW accordingly to the first definition, whereas the converse may not be true. A general expression for entanglement witnesses was presented in \cite{lewenstein2}. Every EW acting on $k$-partite Hilbert space can be written as: \begin{equation} W = P + \sum_{i=1}^{k}Q_{i}^{T_{i}} - \epsilon I, \end{equation} where $P$ and the $Q_{i}$'s are positive semi-definite, $\epsilon \geq 0$, $I$ is the identity operator and $T_{i}$ is the partial transposition with respect to partie $i$. Note that even ($m$+1)-partite EWs can be written in the form of equation (5) \cite{foot2}. An important class of EW is the decomposable entanglement witnesses (d-EW), which can always be written as: \begin{equation} W = P + \sum_{i=1}^{k}Q_{i}^{T_{i}}. \end{equation} This class will be particularly important in our discussion, since the set of entangled states detected by d-EW is invariant under LOCC \cite{doherty2}. \section{Definitions and basic properties} In this section we show how $\rho$-optimal EWs can be used to quantify all the different kinds of multipartite entanglement. First, an unifying approach, which includes several important entanglement measures (EM), will be presented. Then we will consider a new infinite family of entanglement monotones \cite{vidal2}. A general expression for the quantification of entanglement via EWs is defined as: \begin{equation} E(\rho) = \max {\cal f}0, -\min_{W \in {\cal M}} Tr(W\rho) {\cal g}, \end{equation} where ${\cal M} = {\cal W} \cap {\cal C}$, and the set ${\cal C}$ is what distinguish the quantities. We call {\it witnessed entanglement} any measure expressed by equation (7). Some well known EM can be expressed as (7). The first, introduced by Bertlmann {\it et al} \cite{bert}, is: \begin{equation} B(\rho) = \max_{||W - I||_{2} \leq 1}[\min_{\sigma \in {\cal S}}Tr(W\sigma) - Tr(W\rho)], \end{equation} where $W \in {\cal W}$. $B(\rho)$ was shown to be monotonic decreasing under mixing enhancing maps \cite{hayden1} and to be equal to the ${\cal H}_{s}$-distance of $\rho$ to the set ${\cal S}$ of fully-separable states: \begin{equation} B(\rho) = D(\rho) = \min_{\sigma \in {\cal S}}||\rho - \sigma||_{2}. \end{equation} The second is the negativity, i.e., the sum of the negative eigenvalues of $\rho^{T_{A}}$ (the partial transpose of $\rho$ with respect to subsystem A) \cite{vidal3,eisert2,zyc}. It is easily seen that ${\cal N}$ can be written as: \begin{equation} {\cal N}(\rho) = \max {\cal f} 0, - \min_{0 \leq W \leq I}Tr(W^{T_{A}}\rho) {\cal g}. \end{equation} Another quantie is the maximal fidelity of distillation under PPT-protocols, introduced by Rains \cite{rains1}, \begin{equation} F_{d}(\rho) = \frac{I}{d} + \max {\cal f} 0, - \min_{W \in {\cal M}}Tr(W^{T_{A}}\rho) {\cal g}, \end{equation} where ${\cal M} = {\cal f}W \hspace{0.1 cm}| \hspace{0.1 cm} (1 - d)I/d \leq W \leq I/d, \hspace{0.3 cm} 0 \leq W^{T_{A}} \leq 2I/d {\cal g}$ \cite{foot3}. The last is the celebrated Wootter's concurrence of two qubits, which can be written, accordingly to Verstraete \cite{verstraete1}, as \begin{equation} C(\rho) = \max {\cal f} 0, -\min_{A \in SL(2,C)}Tr((\ket{A}\bra{A})^{T_{B}}\rho) {\cal g}, \end{equation} where $\ket{A}$ denotes the unnormalized state $(A \otimes I)\ket{I}$ with $\ket{I} = \sum_{i} \ket{ii}$, det$(A) = 1$. Assuming that the set ${\cal C}$ is also convex, which is the case of all the quantities considered in this paper, except the concurrence, it is possible to apply the concept of Lagrange duality from the theory of convex optimization to the problems represented by equation (7) \cite{boyd}. Remarkably, the {\it dual} measures obtained are those related to mixing properties, such as the robustness of entanglement \cite{vidal4}, introduced by Vidal and Tarach, and the best separable approximation measure \cite{karnas}, introduced by Karnas and Lewenstein. Moreover, since in all the cases considered here there always exist a strictly feasible point, i.e, a $W \in \textbf{relint} {\cal M}$ (denoted in the convex optimization literature by Slater condition), the optimal solution of the primal and dual problems are the same, i.e., the primal and dual measures are equal \cite{boyd, ree}. We now show that the dual representation of the generalized robustness of entanglement $R_{G}(\rho)$ \cite{steiner}, i.e., the minimal $s$ such that \begin{equation} \frac{\rho + s\pi}{1 + s} \end{equation} is separable, where $\pi$ is any, not necessarily separable, density matrix, is given by (7) with ${\cal M} = {\cal f} W \in {\cal W} \hspace{0.1 cm} | \hspace{0.1 cm} W \leq I {\cal g}$. Following \cite{boyd}, the Lagrangian of the problem is given by \begin{equation} \begin{split} L(W, g(\sigma), Z) = Tr(W\rho) + Tr(WZ) \\ - Tr(Z) - \int\limits_{\sigma \in {\cal S}} g(\sigma)Tr(W\sigma)d\sigma, \end{split} \end{equation} where $Z$, $g(\sigma)$ are the Lagrange multipliers associated with the constraints $W \leq I$ and $Tr(W\sigma) \geq 0 \hspace{0.1 cm} \forall \sigma \in {\cal S}$, respectively. Note that since the definition of EW is a composition of infinite constraints, its Lagrange multiplier is a generalized function \cite{ree}. The dual problem is then \begin{eqnarray} \quad \mbox{minimize} \quad & Tr(Z) \\ \quad \mbox{subject to} \quad & Z \geq 0 \nonumber \\ & g(\sigma) \geq 0, \hspace{0.2 cm} \forall \sigma \in {\cal S} \nonumber \\ & \rho + Z = {\displaystyle \int\limits_{\sigma \in {\cal S}}} g(\sigma)\sigma d\sigma. \nonumber \end{eqnarray} Since $g(\sigma) \geq 0$, the integral in the constraints above is a separable state. Conversely, any separable state $\sigma_{o}$ is obtained with the choice of $g(\sigma) = \delta(\sigma - \sigma_{o})$. It is then easily seen that the result of (15) is the generalized robustness. The dual representation of the robustness of entanglement, $R(\rho)$, has, instead of $W \leq I$, the constraint $Tr(W\sigma) \leq 1, \hspace{0.1 cm} \forall \hspace{0.1 cm} \sigma \in S$. The best separable approximation measure $BSA(\rho)$ \cite{karnas} is the minimum $\lambda$ such that there exist a separable state $\sigma$ and a density operator $\delta \rho$ satisfying \begin{equation} \rho = (1 - \lambda)\sigma + \lambda \delta \rho. \end{equation} It can been seen that the dual representation of $BSA(\rho)$ is given by (7) with ${\cal M} = {\cal f} W \in {\cal W} \hspace{0.1 cm} | \hspace{0.1 cm} W \geq -I {\cal g}$. In \cite{brandao2} it was shown that the random robustness $R_{r}(\rho)$ \cite{vidal4}, i.e., to the minimal $s$ such that \begin{equation} \frac{\rho + s(I/D)}{1 + s} \end{equation} is separable, is equal to equation (7), with ${\cal M} = {\cal f} W \in {\cal W} \hspace{0.1 cm} | \hspace{0.1 cm} Tr(W) = D {\cal g}$. This result can also be easily derived using the concept of Lagrange duality. In the next subsection we will introduce our new family of entanglement monotones \subsection{A New Family of Entanglement Monotones} If we let ${\cal C}$ be the set of Hermitian matrices $W$ such that $-nI \leq W \leq mI$, where $n, m \geq 0$, then the quantity derived from (7) will be denoted by $E_{n:m}$. \begin{proposition} $E_{n:m}$ is an entanglement monotone for every $n, m \geq 0$, i.e. \begin{equation} \sum_{i}p_{i}E_{n:m}(\rho_{i}') \leq E_{n:m}(\rho) \end{equation} where $\rho_{i}'$ is the final state conditional on the occurrence of the classical variable $i$, which occurs with probability $p_{i}$ at the end of a LOCC protocol. \end{proposition} \begin{proof} It suffices to consider final states of the form \begin{equation} \rho_{i}' = A_{i}\rho A_{i}^{\cal y} / p_{i} \end{equation} with $p_{i} = Tr[A_{i}\rho A_{i}^{\cal y}]$, where the Kraus operators $A_{1}, ..., A_{M}$ are given by $A_{i} = A_{i}^{1} \otimes ... \otimes A_{i}^{k}$ and satisfy $\sum_{i=1}^{M}A_{i}^{\cal y}A_{i} \leq I$: \begin{equation} \begin{array}{c} \sum_{i}p_{i}E_{W}(\rho_{i}') = \sum_{i}p_{i} \max{\cal f}0, -Tr(W_{\rho_{i}'}\rho_{i}') {\cal g} \\ = \sum_{k} -Tr(A_{k}^{\cal y}W_{\rho_{k}'}A_{k}\rho) \leq -Tr(W_{\rho}\rho) = E_{W}(\rho) \end{array} \end{equation} where $k$ sums only the terms such that $\max{\cal f}0, -Tr(W_{\rho_{i}'}\rho_{i}'){\cal g}$ is different from zero. In the last inequality we used that $W = \sum_{k}A_{k}^{\cal y}W_{\rho_{k}'}A_{k} \leq m\sum_{k}A_{k}^{\cal y}A_{k} \leq mI$, $W = \sum_{k}A_{k}^{\cal y}W_{\rho_{k}'}A_{k} \geq -n\sum_{k}A_{k}^{\cal y}A_{k} \geq -nI$, and that $W_{\rho}$ is optimal. \end{proof} Note that the proof of proposition (2), with minors modifications, applies also to ${\cal N}$ and $F_{d}$. The dual representation of $E_{m:n}(\rho)$ is \begin{eqnarray} \quad \mbox{minimize} \quad & ms + nt \\ \quad \mbox{subject to} \quad & \rho + s\pi_{1} = (1 + s - t)\sigma + t\pi_{2} \nonumber \end{eqnarray} where $\pi_{i}$ are density matrices, $\sigma$ is a separable state and $s, t \geq 0$. From (24) we find that \begin{equation*} \lim_{m \rightarrow \infty}E_{n:m}(\rho) = nBSA(\rho), \hspace{0.4 cm} \lim_{n \rightarrow \infty}E_{n:m}(\rho) = mR_{G}(\rho) \end{equation*} Actually, the equalities above are already valid when one of the numbers is sufficiently larger than the other. The elements of this new family of EMs can be interpreted as intermediate measures between the generalized robustness and the best separable approximation. Note that for every distinct rational number $n/m$ within a certain finite interval, the $E_{m:n}$ are genuine different EMs, meaning that there is no positive number $c$ such that $E_{m:n} = cE_{m':n'}$ if $n/m \neq n'/m'$. If we consider that ${\cal C}$ is the intersection of set of Hermitian matrices $W$ such that $-nI \leq W \leq mI$ with the set of decomposable entanglement witnesses, a new family of entanglement monotones, denoted by $E^{PPT}_{n:m}$ is defined. To see that they are indeed EMs, all we have to note is that for every $A_{i} = A^{1}_{i} \otimes ... \otimes A^{k}_{i}$, $A^{\cal y}_{i}WA_{i}$ is a decomposable EW whenever $W$ is. Therefore, proposition (2) also applies to them. It is possible to derive several other families of EMs considering intersections of the sets ${\cal C}$ of different entanglement measures which can be written as equation (7), such as those given by equations (10-12). \section{Multipartite Entanglement Hierarchy} We now discuss more about the different kinds of multipartite entanglement introduced in the second section. Usually the set of separable states is regarded to be composed of all sates which can be created by LOCC protocols. In this sense, given a specific split and considering that each part of the split can perform global quantum operations on its subsystems, only 1-separable states can be properly identified as separable. However, one might also be interest in the situation where some of the parties are allowed to perform join operations. In this case, the different types of entanglement play an important role. Consider, for example, the situation where $k$ parties want to create a common quantum state and each one is connected to the others via a quantum channel. If they all agree in using their channels, every state can be prepared and the situation becomes trivial. However, suppose that they agree that only $m \leq k$ parties will use their quantum channels, where the probabilities of which parties will be involved are given by $p_{i}$. At the end of the protocol they will share an ensemble of states, ${\cal f}\rho_{i}, p_{i}{\cal g}$, which clearly does not have $m + 1$-partite entanglemnt. Now, since erasing classical information cannot create entanglement, we are lead to consider the different kinds of entanglement discussed before. This property is reflected in the condition that every goog entanglement measure should be convex, which we show for every quantity defined according to equation (7) \begin{proposition} $E$ is a convex function for any choice of ${\cal C}$, i.e., \begin{equation} E\left(\sum_{i}p_{i}\rho_{i} \right) \leq \sum_{i}p_{i}E(\rho_{i}), \end{equation} whenever the $\rho_{i}$ are Hermitian, and $p_{i} \geq 0$ with $\sum_{i}p_{i} = 1$. \end{proposition} Proposition (3) follows from the convexity and the concavity of the \textit{max} and \textit{min} functions, respectively. Consider a given $k$-partite split of a multi-partite system $\rho$. It is possible to attribute $(k - 1)$ numbers, $E^{m}$, $1 \leq m \leq k - 1$, where each one quantifies one type of multi-partite entanglement of the system. It is easy to see from equation (3) that all constraints imposed to an EW which detects $m$-partite entanglement ($m$-EW), are also imposed to every $n$-EW, with $n \geq m$. Hence, the following order between the $E^{m}$ holds: \begin{equation} E^{m}(\rho) \geq E^{n}(\rho), \hspace{0.4 cm} \forall \hspace{0.2 cm} n \geq m. \end{equation} $E^{1}(\rho)$, formed by the OEW with respect to the fully separable states is an upper-bound to all other $E(\rho)$, including those with respect to other splits formed by grouping several original parties into one. This means, for example, that in a 3-split, $E^{1}(\rho)$ is greater or equal to the bipartite entanglement of any of the three 2-splits, namely A-BC, AB-C and AC-B. Actually, it is possible to establish a complete hierarchy in the proposed measures \cite{brandao3}. An interesting measure of entanglement for multi-partite systems is the {\it localizable entanglement}, introduced by Verstraete {\it et al} \cite{verstraete2}. Given a quantum system of $n$ parties $\rho$, the {\it localizable entanglement} $E_{ij}(\rho)$ is the maximal amount of entanglement that can be created, on average, between the parties $i$ and $j$ by performing a single-copy LOCC protocol in the system \cite{foot4}. More specifically, if at the end of a LOCC protocol we have an ensemble of states $\mu = {\cal f}p_{k}, \rho_{k}^{ij}{\cal g}$, where $p_{k}$ is the probability that the reduced state of the parties $i$ and $j$ is $\rho_{k}^{ij}$, the LE is then given by \begin{equation} E^{ij} = \max_{\mu}\sum_{k}p_{k}E(\rho_{k}^{ij}), \end{equation} where $E(\rho)$ represents, in this paper, one measure based on OEWs. The LE has the operational meaning which applies to situations in which out of some multipartite entangled state one would like to concentrate as much entanglement as possible in two particular parties \cite{verstraete2}, which could be used later, for instance, in some quantum information task. \begin{proposition} Consider a multi-partite state $\rho$. Then \begin{equation} E^{ij}_{n:m} \leq E^{1}_{n:m}(\rho), \hspace{0.4 cm} \forall \hspace{0.2 cm} i, j, n, m. \end{equation} \end{proposition} \begin{proof} As in the proof of proposition 2, it suffices to consider final states of the form \begin{equation} \rho_{l}^{ij} = Tr_{/ij}(A_{l}\rho A_{l}^{\cal y} / p_{l}) \end{equation} with $p_{l} = Tr[A_{l}\rho A_{l}^{\cal y}]$, where $Tr_{/ij}$ stands for the partial trace of all parties, except i and j. The Kraus operators $A_{1}, ..., A_{M}$ are given by $A_{l} = A_{l}^{1} \otimes ... \otimes A_{l}^{k}$ and satisfy $\sum_{i=1}^{M}A_{i}^{\cal y}A_{i} \leq I$. \begin{equation} \begin{array}{c} E^{ij}_{n:m} = \sum_{l}p_{l}E_{n:m}(\rho_{l}^{ij}) = \sum_{l} \max{\cal f}0, -Tr(I\otimes W_{\rho_{l}^{ij}}\rho_{l}) {\cal g} \\ = \sum_{k} -Tr(A_{k}^{\cal y}I\otimes W_{\rho_{k}^{ij}}A_{k}\rho) \leq -Tr(W_{\rho}\rho) = E_{n:m}(\rho) \end{array} \end{equation} where $k$ sums only the terms such that $\max{\cal f}0, -Tr(W_{\rho_{l}'}\rho_{l}'){\cal g}$ is different from zero. In the last inequality we used that the EW $W = \sum_{k}A_{k}^{\cal y}I\otimes W_{\rho_{k}}^{ij}A_{k} \leq m\sum_{k}A_{k}^{\cal y}A_{k} \leq mI$, $W = \sum_{k}A_{k}^{\cal y}I\otimes W_{\rho_{k}}^{ij}A_{k} \geq -n\sum_{k}A_{k}^{\cal y}A_{k} \geq -nI$ and that $W_{\rho}$ is optimal. Note that proposition (5) also applies to $E_{n:m}^{PPT}$. \end{proof} The following relation between the negativity, ${\cal N}(\rho)$, and $E_{\infty:1}^{PPT}(\rho) = R_{G}^{PPT}(\rho)$ holds: \begin{proposition} \begin{equation} {\cal N}(\rho) \leq E_{\infty:1}^{PPT}(\rho) \leq d{\cal N}(\rho) \end{equation} \end{proposition} \begin{proof} For every positive operator $M$, we have \begin{equation} \lambda_{max}(M^{T_{A}}) \leq \lambda_{max}(M) \leq d\lambda_{max}(M^{T_{A}}) \end{equation} where the first (second) inequality is saturated for separable (singlet) states. Hence, as $0 \leq W \leq 1$ implies $W^{T_{A}} \leq 1$, we find \begin{eqnarray} {\cal N}(\rho) = -\min_{0 \leq W \leq I}Tr(W^{T_{A}}\rho) \nonumber \\ \leq -\min_{\substack{W^{T_{A}} \leq I \\ W \geq 0}}Tr(W^{T_{A}}\rho) = E_{\infty,1}^{PPT}(\rho) \end{eqnarray} where we have used that the optimal decomposable EW for a bipartite system has always the form $W^{T_{A}}$, $W \geq 0$. From equation (32) we also find that $W \geq 0$, $W^{T_{A}} \leq I$ implies $0 \leq W \leq dI$. Thus, \begin{eqnarray} E_{\infty:1}^{PPT}(\rho) = -\min_{\substack{W^{T_{A}} \leq I \\ W \geq 0}}Tr(W^{T_{A}}\rho) \nonumber \\ \leq -\min_{0 \leq W \leq dI}Tr(W^{T_{A}}\rho) = d{\cal N}(\rho) \end{eqnarray} \end{proof} The second inequality is strict for example on the state \begin{equation} \rho = \frac{I - d(P^{+})^{T_{A}}}{d^{2} - d} \end{equation} where $P^{+}$ is the maximal $d$ x $d$ entangled state. \section{Numerical calculation} The lack of an operational procedure to calculate entanglement measures in general is ultimately related to the complexity of distinguish entangled from separable mixed states, which was shown to be NP-HARD \cite{gurvits}. Since an operational measure, which has positive value in every entangled state, would also be a necessary and sufficient test for separability, we should not expect to find one. Nonetheless, some approximative numerical methods based on convex optimization have been proposed to the separability problem \cite{brandao1,doherty1,eisert1}. What we will show in this section is that these methods can also be used to calculate, approximately, the {\it witnessed entanglement}. The first one, proposed in \cite{brandao1} by Brandao and Vianna, linked the optimization of EWs with a class of convex optimization problems known as robust semidefinite programs (RSDP). Although RSDPs belong to NP-HARD, some well known probabilistic relaxations, which transforms the problem in a semidefinite program (SDP), were applied, leading to a method of optimizing {\it pseudo}-EWs (operators which are positive in almost all separable states) to every multipartite state and with respect to all types of entanglement. The second approach, due to Doherty {\it et al} \cite{doherty1,doherty2} , was actually the first method to the separability problem based on SDP. Using the existence of symmetric extensions for separable states and the concept of duality in convex optimization, a hierarchy of SDPs, where the $(k + 1)^{th}$ test is at least as powerful as $(k)^{th}$ (but demands more computational effort), was constructed to detect entanglement. In each step k, an OEW with respect to a restricted set of EWs, which converges to the whole set of EWs in the limit of $k \rightarrow \infty $, is obtained. This method can be used, however, only for the entanglement with respect to the fully-separable states, $E^{1}$. Note that the further constraints that we demand to the EWs can be incorporated in the SDP, since they are linear matrix (in)equalities. The last method, introduced by Eisert {\it et al} \cite{eisert1}, is based on recently developed relaxations of non-convex polynomial problems of degree three in a hierarchy of SDPs, which converges to the solution of the original problem as the dimension of the SDP reaches the infinity \cite{lassere}. One of the applications of this method is the minimization of the expectation values of EWs with respect to pure product states. Therefore, it can be used together with the second method discussed to lower the value of $Tr(W\rho)$, where $W$ is a non-optimal EW determined by some step of the hierarchy. We would like to stress the complementary character of these methods. Whereas the first method usually provides upper bounds to $E(\rho)$, since it only determines {\it pseudo}-EWS, the second and third provides lower bounds to $E(\rho)$, as the EWs resulting from them are non-optimal. Although only the first one can calculate $E^{m}(\rho)$, for $m > 1$, the number of constrains imposed grows exponentially with $m$. Thus, in most of cases, we will restrict ourselves to the determination of $E^{1}$, which is an upper bound to all other types of multi-partite entanglement (see section IV). Note that measures restricted to decomposable EWs can always be exactly calculated, in the worst case, by a semidefinite program. \subsection{Example I} \begin{figure} \begin{center} \includegraphics[scale=0.45]{WGHZ1.eps} \caption{(Color online) $E^{1}_{n:1}(\rho_{q})$ for $0 \leq n \leq 4$ and $0 \leq q \leq 1$.} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.45]{WGHZ2.eps} \caption{(Color online) $E^{2}_{n:1}(\rho_{q})$ for $0 \leq n \leq 4$ and $0 \leq q \leq 1$.} \end{center} \end{figure} As a first example we calculated $E_{n:1}^{1}$ and $E_{n:1}^{2}$ , $0 \leq n \leq 1$, for the following family of states \begin{equation} \rho_{q} = q\ket{W}\bra{W} + (1- q)\ket{GHZ}\bra{GHZ}, \hspace{0.2 cm} 0 \leq q \leq 1 \end{equation} where $\ket{W} = (\ket{001} + \ket{010} + \ket{100})/\sqrt{3}$ and $\ket{GHZ} = (\ket{000} + \ket{111})/\sqrt{2}$. The results are plotted in figures (1) and (2). When $n << 1$, $E_{n:1}^{1} = nBSA$ and we see that for all q there is not product vectors and even biseparable vectors in the range of $\rho_{q}$. In the other limit, where $E_{n:1}^{1} = R_{G}$, we find that the generalized robustness of entanglement with respect to biseparable states is the same for all $\rho_{q}$ with $q \leq 0.7$. \subsection{Example II} The classes of entangled states equivalent by SLOCC for 2 X 2 X n systems were determined in \cite{miyake} and can be represented by the states (1-5) of figure (3). The arrows indicate which transformations are probabilistic possibles. \begin{figure} \begin{center} \includegraphics[scale=0.4]{verstraete1.eps} \caption{Representative states of the five distinct classes of 3-entangled states.} \end{center} \end{figure} $E_{n:1}^{1}$ , $0 \leq n \leq 1$, was calculated for each of these and plotted in figure (4). \begin{figure} \begin{center} \includegraphics[scale=0.4]{2x2x4.eps} \caption{$E^{1}_{n:1}(\rho_{q})$, $0 \leq n \leq 4.5$, for the states (1-5) of figure (3).} \end{center} \end{figure} Note that for all n considered, the incomparable states (2-3) and (4-5) with respect to state transformations have approximately the same $E^{1}_{n:1}$. \subsection{Example III} As a final example, we present a numerical comparison between ${\cal N}$ and $E_{\infty,1} = R_{G}^{PPT}$. The bipartite PPT-generalized robustness can be determined as easily as the negativity. Actually, it can be written as \begin{equation} R_{G}^{PPT}(\rho) = \frac{1}{\lambda_{max}(P^{T_{A}})}{\cal N}(\rho) \end{equation} where $\lambda_{max}(P^{T_{A}})$ is the maximum eigenvalue of the partial transposed projector onto the negative eigenspace of $\rho^{T_{A}}$. We have generated $10^{5}$ random states using the algorithm presented in \cite{zyc} and plotted in figures (5) and (6). \begin{figure} \begin{center} \includegraphics[scale=0.45]{NEw1.eps} \caption{(Color online) $R_{G}^{PPT} x \hspace{0.1 cm} {\cal N}$ for $10^{5}$ 4 x 4 random sates.} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.45]{NEw2.eps} \caption{(Color online) $R_{G}^{PPT} x \hspace{0.1 cm} {\cal N}$ for $10^{5}$ 6 x 6 random sates.} \end{center} \end{figure} Although $d{\cal N} \geq R_{G}^{PPT} \geq {\cal N}$ (see section IV), we see from figures 5 and 6 that $R_{G}^{PPT} \leq 2{\cal N}$ for the majority of states. \section{Gaussian states} We considered $n$ distinguishable infinite dimensional subsystems, each with local Hilbert space ${\cal H} = {\cal L}^{2}({\cal R}^{n})$. A Gaussian state is characterized by a density operator whose characteristic function $\chi_{\rho}(x) = Tr[\rho W(x)]$ is a Gaussian function \cite{giedke1}. We can write, for every Gaussian state $\rho$, \begin{equation} \rho = \pi^{-n}\int_{{\cal R}^{2n}}dx e^{-\frac{1}{4}x^{T}\gamma x + id^{T}x}W(x), \end{equation} where $W(x) = \exp[-ix^{T}R]$ are the displacement operators and $R = (X_{1}, P_{1}, X_{2}, ..., P_{n})$, with $[X_{k}, P_{l}] = i\delta_{kl}$. The matrix $\gamma \geq iJ_{n}$ is a 2$n$ x 2$n$ real matrix called correlation matrix (CM) and $d$ is an 2$n$ real vector called displacement \cite{giedke1}. The symplectic matrix is given by \begin{equation} J_{n} = \bigoplus^{n}_{k = 1}J_{1}, \hspace{0.3 cm} J_{1} = \left( \begin{array}{cc} 0 & -1 \\ 1 & 0 \end{array} \right) \end{equation} Note that the displacement of a state can always be adjusted to $d = 0$ by a sequence of unitaries applied to individual modes. This implies that d is irrelevant for the study of entanglement. Thus, we set $d = 0$ for now on without loss of generality. The optimization of EWs for states of infinite dimension is completely infeasible. Nonetheless, we can still obtain meaningful quanties if we restrict it to a simpler, but sufficiently large, set of operators. An obvious choice would be the restriction to Gaussian entanglement witnesses (GEW), i.e., Gaussian operators which are positive in separable Gaussian states. Unfortunately, none Gaussian entangled state is detected by a GEW. Assuming that G is a GEW with covariance matrix $\Gamma$, we find \begin{equation} Tr(\rho G) = \int_{{\cal R}^{2n}}dx e^{-\frac{1}{4}x^{T}(\Gamma + \gamma)x + c} \geq 0. \end{equation} Another possible class of operators then is given by \begin{equation} {\cal W}_{\cal G} = {\cal f} Q \in {\cal B}({\cal H} \otimes {\cal H}) \hspace{0.1 cm} | \hspace{0.1 cm} Q = 2^{n}I - G {\cal g}, \end{equation} where $G$ is a Gaussian operator and $I$ the identity operator \cite{foot5}. In the next proposition we show that $E_{\infty:m}^{G}(\rho)$ given by \begin{equation} E_{\infty:1}^{G}(\rho) = \max {\cal f} 0, -\min_{Q \in {\cal W}_{\cal G}, \hspace{0.1 cm} Q \leq I} Tr(Q \rho) {\cal g} \end{equation} can be very efficiently numerically calculated by a simple semidefinite program. \begin{proposition} \begin{equation} E_{\infty:m}^{G}(\rho) = \max {\cal f}0, \hspace{0.1 cm} \det(\Gamma + \gamma)^{-1/2} - 2^{n} {\cal g} \end{equation} where the matrix $\Gamma \in M_{2n}({\cal R})$ is obtained by the following SDP determinant maximization problem \begin{equation*} \max_{\Gamma, \tau, \Delta, s_{ij}} \hspace{0.1 cm} \det(\Gamma + \gamma)^{-1/2} \end{equation*} \begin{center} subject to \hspace{0.3 cm} $-1 \leq \tau \leq 1, \hspace{0.6 cm} \Gamma \geq 0$ \end{center} \begin{eqnarray} \left( \begin{array}{cc} \tilde{\Gamma}_{2} + i\tau J & \tilde{\Gamma}_{12}^{T} \\ \tilde{\Gamma}_{12} & \tilde{\Gamma}_{1} - iJ \end{array} \right) \geq 0 \end{eqnarray} \begin{equation*} \left( \begin{array}{cc} J_{n} + \Gamma & \Delta \\ \Delta & D(\Delta) \end{array} \right) \geq 0, \hspace{0.4 cm} \left( \begin{array}{cc} s_{k-1,2l-1} & s_{k,l} \\ s_{k,l} & s_{k-1,2l} \end{array} \right) \geq 0 \end{equation*} \begin{center} $s_{k-1,2l-1} \geq 0, \hspace{0.11cm} s_{k-1,2l} \geq 0, \hspace{0.11cm} s_{kl} \geq 2$, \hspace{0.11 cm} {\small $k = 1..l, i = 1...2^{l-k}$} \end{center} where $\Delta$ is a n x n lower triangular matrix comprised of additional variables, D($\Delta$) is a diagonal matrix with same diagonal entries as those of $\Delta$, l is the smallest number such that $2^{l} \geq n$, and $s_{0,i} = \Delta_{ii}$ if $1 \leq i \leq n$ and $s_{0,i} = 2$ if $n \leq i \leq 2^{l}$. \end{proposition} \begin{proof} Consider the following structure for the bipartite Gaussian operator $G$ \begin{equation} G = \int_{{\cal R}^{2n}}dx e^{-\frac{1}{4}x^{T}\Gamma x}W(x), \end{equation} where $\Gamma^{T} = \Gamma \geq 0 \in M_{2n}({\cal R})$, with modes $1$ to $m$ and $m + 1$ to $n$ belonging to Alice and Bob, respectively. The optimization objective $Tr(Q \rho)$, where $Q = 2^{n}I - G$, can be written as \begin{equation} 2^{n} - \pi^{-n}\int_{{\cal R}^{2n}}dx e^{-\frac{1}{4}x^{T}(\Gamma + \gamma)x} = 2^{n} - \det(\Gamma + \gamma)^{-1/2} \end{equation} From the Jamiolkowski isomorphism, $Q$ is an EW iff the map ${\cal Q}$ defined as $Q = I \otimes {\cal Q}(P^{+}) = 2^{n}I - I \otimes {\cal G}(P^{+})$ \cite{foot6} is positive, which is equivalent to $\rho' = I \otimes {\cal G}(\rho) \leq 2^{n}I$, for every density operator $\rho$. The covariance matrix of $\rho'$, $\gamma'$, can be written as \cite{simon} \begin{equation} \gamma' = S^{T}\sigma S , \end{equation} where $S \in Sp(2n,{\cal R})$ and $\sigma$ is the covariance matrix \begin{equation*} \sigma = \mbox{diag}(\mu_{1},\mu_{1},...,\mu_{n},\mu_{n}) \end{equation*} corresponding to a tensor product of states diagonal in the number basis given by \begin{equation} M' = \bigotimes_{i}\frac{2}{\mu_{i} + 1}\sum_{k=0}^{\infty}\left(\frac{\mu_{i} - 1}{\mu_{i} + 1}\right)\ket{k}_{i}{}_{i}\bra{k}, \end{equation} $\ket{k}_{i}$ being the {\it k}-th number state of the Fock space ${\cal H}_{i}$ \cite{adesso}. The symplectic transformation (47) is reflected in the Hilbert space level by an unitary transformation: $G = U(S)^{\cal y}G'U(S)$. Since we are considering bounded operators, the $\mu_{i}$ must be non-negative. We also see that the positiveness of ${\cal Q}$ is equivalent to \begin{equation} \lambda_{max}(\rho') = \prod_{j = 1}^{n} \left( \frac{1}{1 + \mu_{j}} \right) \leq 1, \hspace{0.2 cm} \forall \rho' \end{equation} Thus, since we are only considering Gaussian operators, equation (49) is satisfied iff $\gamma' \geq 0$ for every $\gamma \geq iJ$, where $\gamma'$ and $\gamma$ are the covariance matrices of $\rho'$ and $\rho$, respectively. Following Giedke and Cirac \cite{giedke1}, one finds that ${\cal Q}$ is positive iff \begin{equation} \min_{z \in{\cal C}^{2n}}\max{\cal f}z^{\cal y}(M + iJ)z, z^{\cal y}(M - iJ)z{\cal g} \geq 0 \end{equation} where $M = \tilde{\Gamma}_{2} - \tilde{\Gamma}_{12}^{T}(\tilde{\Gamma}_{1})^{-1}\tilde{\Gamma}_{12}$ and $\tilde{\Gamma} = (I \oplus \Lambda)\Gamma(I \oplus \Lambda)$, with $\Lambda = $ diag$(1,-1,1,-1,...,-1)$. The matrices $\Gamma_{i}$ are such that \begin{equation*} \Gamma = \left( \begin{array}{cc} \Gamma_{1} & \Gamma_{12} \\ \Gamma_{12}^{T} & \Gamma_{2} \end{array} \right) \end{equation*} We now express condition (50) as a linear matrix inequality. Equation (50) is equivalent to \begin{eqnarray} z^{\cal y}(M + iJ)z \geq 0 \quad \mbox{$\forall \hspace{0.1 cm} z \in {\cal C}^{2n}$ \hspace{0.1 cm}s.t.} \quad z^{\cal y}(M - iJ)z \leq 0 \nonumber \\ z^{\cal y}(M - iJ)z \geq 0 \quad \mbox{$ \forall \hspace{0.1 cm} z \in {\cal C}^{2n}$ \hspace{0.1 cm} s.t.} \quad z^{\cal y}(M + iJ)z \leq 0 \nonumber \end{eqnarray} An important theorem of matrix analysis, known as {\cal S}-procedure, can be stated as follows: a quadratic function in the variable $x$, $G(x)$, is positive for all $x$ such that $H(x) \geq 0$, where $H(x)$ is another quadratic function, iff there exists a positive real number $\tau$ such that $G - \tau H \geq 0$, for all $x$ \cite{boyd2}. Applying it to the two conditions above we find that equation (50) holds iff there exists a positive number $\tau$ such that \begin{equation} M + \tau iJ \geq 0, \hspace{0.3 cm} -1 \leq \tau \leq 1 \end{equation} We now use another fact of matrix analysis which says that the constraints on the Schur complement $R > 0, \hspace{0.3 cm} Q - SR^{-1}S^{T} \geq 0$ and $\ker(R) \subseteq \ker(Q)$ are equivalent to \begin{equation*} \left( \begin{array}{cc} Q & S \\ S^{T} & R \end{array} \right) \geq 0 \end{equation*} Hence, applying it to equation (51), we find that a Gaussian operator ${\cal G}$ is an EW iff there exists a real number $-1 \leq \tau \leq 1$ such that equation (44) holds. From the Williason decomposition, we see that $Q \leq I$ is equivalent to \begin{equation} \prod_{i=1}^{n}\left(\frac{2}{\eta_{i} + 1}\right) = 2^{n}\det(I + (S^{T})^{-1}\Gamma S^{-1})^{-1} \leq 1, \end{equation} where $\eta_{i}$ ate the symplectic eigenvalues of $Q$. Since $S$ is symplectic, one has $S^{T}J_{n}S = J_{n}$, so that $\det(I + (S^{T})^{-1}\Gamma S^{-1})^{-1} = \det[{J_{n}^{T}(S^{T})^{-1}(J_{n} + \Gamma)S^{-1}}]^{-1} = \det(J_{n} + \Gamma)^{-1}$ . The proposition then follows from reference \cite{ben}, which presents a LMI representation for the inequality $\det(A)^{1/m} \geq t$, where $A$ is a positive $m$ x $m$ real matrix. \end{proof} \section{States with symmetry} Entanglement measures have usually their calculation greatly simplified when the state in question has certain symmetries. Following \cite{vollbrecht1}, let $G$ be a closed group of product unitary operators of the form $U = U_{1} \otimes U_{2}$. Defining the projection \begin{equation} \textbf{P}(A) = \int dU \hspace{0.2 cm} UAU^{*}, \end{equation} for any operator $A$ on $H_{1}\otimes H_{2}$ \cite{foot7}, where $dU$ is the Haar measure of $G$, we say that an operator $M$ is invariant under $G$ if $\textbf{P}(M) = M$, which is equivalent to $[U, M] = 0$ for all $U \in G$. Consider now the determination of any measure expressed by equation (7). If the state in question has the property $\textbf{P}(\rho) = \rho$, then one can restrict the optimization in (7) to operators with the this same symmetry. More specifically, \begin{eqnarray} E(\rho) = -\min_{W \in {\cal M}} Tr(W\rho) = -\min_{W \in {\cal M}} Tr(W\textbf{P}(\rho)) \nonumber \\ = -\min_{W \in {\cal M}} Tr(\textbf{P}(W)\rho) = -\min_{W \in \textbf{P}({\cal M})} Tr(W\rho), \end{eqnarray} where $\textbf{P}({\cal M}) = {\cal f}W \in {\cal M} \hspace{0.1 cm} | \hspace{0.1 cm} W = \textbf{P}(W){\cal g}$ \cite{foot10}. As an example, consider the isotropic states $\rho_{p}$ on ${\cal C}^{d}\otimes{\cal C}^{d}$ \begin{equation} \rho_{p} = pP^{+} + (1-p)\frac{I}{d^{2}}, \end{equation} where $P^{+} = \ket{\Phi^{+}}\bra{\Phi^{+}}$ is the maximally entangled state. It can be shown that $P^{+}$ and the identity are the only operators which commute with all unitaries of the form $U\otimes U^{*}$. Hence, the OEWs for $\rho_{p}$ can be written as \begin{equation} W(\rho_{p}) = a(p)P^{+} + b(p)I. \end{equation} Since $\bra{\psi}P^{+}\ket{\psi} \leq 1/d$, for every separable state $\ket{\psi}$, we find \begin{eqnarray} E_{n:1}(\rho_{p}) = \begin{cases} (n + 1)p + \frac{(1 - p)(n + 1)}{d^{2}} - 1 & n \leq d-1, \\ dp + \frac{1 - p}{d} - 1 & n \geq d. \end{cases} \end{eqnarray} As the OEWs for this family of states are decomposable, equation (57) is also valid to $E^{PPT}_{n:1}$. \section{Superselection Rules} The effect of superselection rules (SSR) in theory of entanglement has been studied recently under a number of different perspectives \cite{verstraete3,bartlett1,terhal3,schuch}. Two striking features emerge from the existence of a SSR. The entanglement of a given state under SSR is usually reduced \cite{bartlett1} and the notion of nonlocality has to be redefined, as there exists separable states that cannot be created by LOCC \cite{verstraete3}. In this section we show how the {\it witnessed entanglemet} fit in each of these scenarios. Following Bartlett and Wiseman \cite{bartlett1}, we define a SSR as a restriction on the allowed local operations on a system, associated with a group of physical transformations $G$. An operation ${\cal O}$ is $G$-covariant if \begin{equation} {\cal O}[T(g)\rho T^{\cal y}(g)] = T(g){\cal O}[\rho]T^{\cal y}(g), \end{equation} for all group elements $g \in G$ and all density operators $\rho$. Then the SSR associated to $G$ is the restriction on the allowed operations on the system to those $G$-invariant. As these restrictions make a state $\rho$ indistinguishable from the states $T(g)\rho T^{\cal y}(\rho)$ for all $g \in G$, it is convenient to describe $\rho$ by the $G$-invariant state \begin{equation} {\cal G}(\rho) = \int\limits_{G}dgT(g)\rho T^{\cal y}(g), \end{equation} where $dg$ is the group-invariant Haar measure. For multipartite systems, where the SSRs are local, we have ${\cal G}[\rho] = {\cal G}\otimes ... \otimes{\cal G}[\rho]$. As it was shown in \cite{bartlett1}, the maximal amount of entanglement which can be produced by LOCC in a register shared by all the parties, initially in a product state and not subjected to SSRs, from a state $\rho$, constrained by a $G$-SSR, is given by the entanglement they can produce from ${\cal G}(\rho)$ by unconstrained LOCC. If $E$ is an entanglement monotone, any LOCC applied to ${\cal G}(\rho)$, can, on average, at most maintain $E({\cal G}[\rho])$. Since it is always possible to reach this bound applying local swap operators, we have that the maximal amount of entanglement produced by SSR is exactly $E({\cal G}[\rho])$. Hence, from section (VII), it follows that, under a $G$-SSR \begin{equation} E(\rho) = \max {\cal f}0, -\min_{W \in {\cal G}[{\cal M}]} Tr(W\rho) {\cal g}, \end{equation} where ${\cal G}[{\cal M}] = {\cal f}W \in {\cal M} \hspace{0.1 cm} | \hspace{0.1 cm} W = {\cal G}[W]{\cal g}$ \cite{foot8}. We now consider the effect of SSRs in the notion of locality. The states that can be prepared locally in the presence of a $G$-SSR are those which can be written as (1), with each $\ket{\psi_{i}}_{k}$ being $G$-invariant. It is possible to detect nonlocal states with witness operators, defining a $G$-nonlocality witness (GW) as a Hermitian operator which satisfies equation (3), with $\ket{\psi}_{P_{k}^{m}}{}_{P_{k}^{m}}\bra{\psi} = {\cal G}[\ket{\psi}_{P_{k}^{m}}{}_{P_{k}^{m}}\bra{\psi}]$, for all $i$ and $k$. This nonlocal character of some states in the presence of a SSR can be quantified \cite{schuch}. We can then, as it was done with entanglement, use GWs to perform this quantification. A {\it witnessed nonlocality measure}, $N_{G}$, will be any quantie given by equation (7), with the set of EWs substituted by the set of nonlocality witnesses. It is easy to see that all properties discussed for $E$ are valid for $N_{G}$. As an example, considered the following state \begin{eqnarray} \rho = \frac{1}{4}(\ket{0}_{A}\bra{0}\otimes \ket{0}_{B}\bra{0} + \ket{1}_{A}\bra{1}\otimes \ket{1}_{B}\bra{1}) \nonumber \\ + \frac{1}{2}\ket{\Psi_{+}}_{AB}\bra{\Psi_{+}}, \end{eqnarray} where $\ket{\Psi_{+}}_{AB} = (\ket{0}_{A}\ket{1}_{B} + \ket{1}_{A}\ket{1}_{B})/\sqrt(2)$. Verstraete and Cirac have shown \cite{verstraete3} that, although this state has a separable decomposition, it is not local when a particle number SSR exists, since all possible separable decompositions have local states involving superpositions of a different number of particles. Any Hermitian matrix with positive diagonal entries is a G-nonlocality witness in this case. This should be contrasted with the case of a general EW, where an infinite number of inequalities are necessary for its characterization. Calculating, for example, the nonlocal measure equivalent to $E_{\infty:1} = R_{G}$, \begin{equation*} N_{G}(\rho) = \max{\cal f}0, -\min_{G \in {\cal G}}Tr(G\rho){\cal g}, \end{equation*} where ${\cal G} = {\cal f}G \hspace{0.1 cm} | \hspace{0.1 cm} G_{ii} \geq 0, G \leq I {\cal g}$, we find $N_{G})(\rho) = 1/2$, with $G = -\ket{01}\bra{10} - \ket{10}\bra{01}$. \section{Indistinguishable particles} The study of entanglement in systems of indistinguishable particles has been the subject of recent controversy \cite{zanardi1,gittings,pask,schliemann,wiseman1}. At least three different approaches to the problems have been proposed, namely, the {\it entanglement of modes} \cite{zanardi1}, the {\it quantum correlations} \cite{schliemann} and the {\it entanglement of particles} \cite{wiseman1}. Each of these has its own advantages and drawbacks, and no consensus has been reached on which one is the most suitable. In this section we show how the proposed measures based on EW can be used to quantify entanglement in each of the three methods. We start with the {\it entanglement of modes}, proposed by Zanardi \cite{zanardi1}, which suggests that the entanglement of indistinguishable particles should be calculated by any regular entanglement measure, using the density matrix in the mode-occupation, or Fock, representation. In this case it is clear that the determination of the witnessed entanglement follows straightforwardly. The {\it quantum correlations}, introduced by Schliemann {\it et al} \cite{schliemann}, is motivated by the believe that no quantum correlations due to symmetrization (for bosons) or anti-symmetrization (for fermions) should be considered as true entanglement. Then, the characterization of entanglement, for pure states, is determined by the Slater rank of the state, as opposed to the Schimdt rank usually considered in distinguishable particles. Furthermore, Schliemann {\it et al} have shown that the concept of entanglement witness is also applicable to multipartite systems of indistinguishable particles \cite{foot11} and, thus, the witnessed entanglement are well defined in this case too. The last approach, due to Wiseman and Vaccaro \cite{wiseman1}, is probably the best motivated one. The {\it entanglement of particles} is defined as the maximal amount of entanglement, computed by a standard measure, which Alice and Bob can produce between a quantum register, shared by them, composed of distinguishable particles by local operations. The amount of entanglement will clearly depends on the physical constraints imposed, which are in most of cases expressed as a SSR. Therefore the approach presented in the previous section to SSR can also be applied in this case. \section{Teleportation distance} In this section we derive lower bounds to the teleportation distance, using $E_{n:m}$. We consider a quantum state $\rho$ shared by $k$ parties and ask what is the best possible teleportation distance attained by a LOCC protocol when the parties form two groups and teleport a quantum state from one group to the other. Consider a teleportation protocol where a bipartite state $\rho_{AB}$ is used as a quantum channel between Alice and Bob. Following the approach of Vidal and Werner to the negativity \cite{vidal3}, we will first consider the {\it single distance} of a bipartite state defined as: \begin{equation} \Delta(P_{+}, \rho) \equiv \inf_{P}||P_{+} - P(\rho)||_{1}, \end{equation} where $P_{+}$ is the maximally entangled state and the infimum is taken over LOCC protocols $P$. Using the convexity and the invariance under unitary transformations in the two terms of the absolute distance, and the invariance of $P_{+}$ under unitary transformations of the form $U\otimes U^{*}$, we may assume that the optimal state which minimizes equation (62), $P_{opt}(\rho)$, has undergone a twirling operation \cite{foot9} and, therefore, is a {\it noise singlet}, \begin{equation} \rho_{p} = pP_{+} + (1-p)\frac{I\otimes I}{d^{2}}. \end{equation} The absolute distance of $\rho_{p}$ is given by $||P_{+} - \rho_{p}||_{1} = 2(1 - p)(d^{2} - 1)/d^{2}$. From equation (57), \begin{equation} ||P_{+} - \rho_{p}||_{1} = 2(1 - \frac{1 + E_{n:1}^{PPT}(\rho_{p})}{d}). \end{equation} From the monotonicity of $E_{n:1}^{PPT}$ under LOCC, we find that, for $n \geq d$, \begin{proposition} \begin{equation} \Delta(P_{+}, \rho) \geq 2(1 - \frac{1 + E_{n:1}^{PPT}(\rho)}{d}). \end{equation} \end{proposition} Since $E_{n:1}^{PPT}(\rho_p) = 2{\cal N}(\rho_{p})$, for $n \geq d$, we see that $E_{n:1}^{PPT}$ provides, when $E_{n:1}^{PPT} \leq 2{\cal N}$, a sharper bound than the one derived from the negativity. In the limit case $n \rightarrow \infty$ already, where $E_{n:1}^{PPT}$ is equal to the PPT-generalized robustness, we see from section (V) that the new bound is indeed sharper for the majority of states. A measure of the degree of performance of a quantum channel is the teleportation distance \begin{equation} d(\Lambda) = \int d\phi ||\phi - \Lambda(\phi)||_{1}. \end{equation} As it was shown by the Horedecki family \cite{horodecki1}, the minimal teleportation distance that can be achieved when using the bipartite state $\rho$ to construct an arbitrary teleportation channel is given by \begin{equation} d_{min}(\rho) = \frac{d}{d + 1}\Delta(P_{+}, \rho). \end{equation} Therefore, \begin{equation} d_{min}(\rho) \leq \frac{2d}{d + 1}(1 - \frac{1 + E_{n:1}^{PPT}(\rho)}{d}), \hspace{0.1 cm} n \geq d. \end{equation} Until now we have just adapted Vidal and Werner's reasoning for the negativity to the $E_{n:1}^{PPT}$. Nevertheless, as opposed to ${\cal N}$, $E_{n:1}^{PPT}$ are also defined to multi-partite systems. \begin{proposition} Consider a quantum state $\rho$ shared by $k$ parties. Let $\rho^{1..m:(m+1)..k}$ denote a bipartite split of the system, where the parties 1 to $m$ and $m$ + 1 to $k$ form two groups. Then, \begin{equation} d_{min}(\rho^{1..m:(m+1)..k}) \leq \frac{2D}{D + 1}(1 - \frac{1 + (E_{n:1}^{PPT})^{1}(\rho)}{D}), \end{equation} $\forall \hspace{0.1 cm} 1 < m < k$, where $D$ stands for the minimum of the dimensions of the two groups. \end{proposition} Proposition (8) follows from the upper bound to all types of entanglement provided by $E^{1}$. Equation (57) is saturated, for example, on the $k$-partite GHZ state $\ket{\Psi_{GHZ}} = 1/\sqrt{2}(\ket{00...0} + \ket{11...1})$. \section{Upper Bounds for the Distillable Entanglement} We now move on to show another application of the family $E_{1:m}$, namely bounds to the distillable entanglement of bipartite mixed states. We first derive the following additivity property \begin{proposition} \begin{equation} E_{n:1}(\rho \otimes \rho) \leq E_{n:1}(\rho)^{2} + 2E_{n:1}(\rho) , \hspace{0.2 cm} \forall \hspace{0.1 cm} n \geq 1. \end{equation} \end{proposition} \begin{proof} Consider the dual representation (24) of $E_{n:1}$. Let s, t, $\sigma$, $\pi_{1}$ and $\pi_{2}$ be variables which minimize (24). Then we find $E_{n:1}(\rho) = s + nt$, with \begin{equation} \rho = (1 + s - t)\sigma + t\pi_{2} - s\pi_{2}. \end{equation} Thus, \begin{eqnarray} \rho \otimes \rho = (1 + s - t)^{2}\sigma \otimes \sigma + t(1 + s - t)\sigma \otimes \pi_{2} \nonumber \\ - s(1 + s - t) \sigma \otimes \pi_{1} + t(1 + s - t)\pi_{2} \otimes \sigma + t^{2}\pi_{2} \otimes \pi_{2} \nonumber \\ - st\pi_{2}\otimes \pi_{1} - s(1 + s - t)\pi_{1}\otimes \sigma - st\pi_{1}\otimes \pi_{2} + s^{2}\pi_{1}\otimes \pi_{1} \nonumber \\ = [1 + (2s + s^{2} + t^{2}) - (2t + 2st)]\sigma \otimes \sigma \nonumber \\ + [t(\rho \otimes \pi_{2} + \pi_{2} \otimes \rho) + st(\pi_{1}\otimes \pi_{2} + \pi_{2}\otimes \pi_{1})] \nonumber \\ - [ s(\rho \otimes \pi_{1} + \pi_{1} \otimes \rho) + s^{2} \pi_{1}\otimes \pi_{1} + t^{2}\pi_{2} \otimes \pi_{2}], \nonumber \end{eqnarray} where in the last two lines we used that \begin{equation*} \sigma = \frac{1}{1 + s - t}\left(\rho + s\pi_{1} - t\pi_{2}\right). \end{equation*} It is therefore easily seen that if $E_{n:1}(\rho \otimes \rho) = s' + nt'$, then $s' + nt'\leq s^{2} + t^{2} + 2s + n(2t + 2st)$. Hence, as $n \geq 1$, \begin{eqnarray} E_{n:1}(\rho)^{2} + 2E_{n:1}(\rho) - E_{n:1}(\rho \otimes \rho) \nonumber \\ = s^{2} + n^{2}t^{2} + 2nst + 2s + 2nt - s' - nt' \nonumber \\ \geq s^{2} + n^{2}t^{2} + 2nst + 2s + 2nt - s^{2} - t^{2} - 2s - 2nt - 2nst \nonumber \\ = t^{2}(n^{2} - 1) \geq 0. \nonumber \end{eqnarray} \end{proof} We can define a family of quantities close related to $E_{n:1}$ by \begin{equation} LE_{n:1}(\rho) = \log_{2}(1 + E_{n:1}(\rho)). \end{equation} The $LE_{n:1}(\rho)$ are non-increasing under trace preserving separable operations. From proposition (9) we find that they are also weakly-subadditive. Indeed, for $n \geq 1$, \begin{equation} LE_{n:1}(\rho \otimes \rho) \leq \log_{2}((1 + E_{n:1}(\rho))^{2}) = 2LE_{n:1}(\rho). \end{equation} Note that the same results are also valid to $E^{PPT}_{n:1}$. We now can state the main result of this section \begin{proposition} \begin{equation} E_{D}(\rho) \leq LE_{n:1}(\rho) , \hspace{0.1 cm} \forall \hspace{0.1 cm} n \geq 1. \end{equation} where $E_{D}(\rho)$ is the distillable entanglement of the bipartite state $\rho$. \end{proposition} \begin{proof} The proof of proposition (10) is basically an application of a theorem due to the Horodeckis \cite{horodecki5} which can be stated as follows: any function B satisfying the conditions 1)-3) below is an upper bound for the entanglement of distillation. \begin{enumerate} \item Weak monotonicity: $B(\rho) \geq B(\Lambda(\rho))$ where $\Lambda$ is any trace-preserving superoperator realizable by means of LOCC operations. \item Partial subadditivity: $B(\rho^{n}) \leq nB(\rho)$. \item Continuity for isotropic states $\rho_{p}$ given by equation (55). Suppose that we have a sequence of isotropic states $\rho_{p}$\ such that $Tr(\rho_{p}P^{+}) \rightarrow 1$, if $d \rightarrow \infty$. Then we require \begin{equation} \lim_{d \rightarrow \infty}\frac{1}{\log_{2}d}B(\rho_{p}) \rightarrow 1 \end{equation} \end{enumerate} We have already shown that $LE_{n:1}(\rho)$, for $n \geq 1$, satisfies conditions (1) and (2). From equation (54) \begin{equation} LE_{n:1}(\rho_{p}) = \log_{2}(dp + \frac{1 - p}{d}), \hspace{0.1 cm} \forall \hspace{0.1 cm} n \geq 1 \end{equation} By evaluating this expression now for large $d$, we easily obtain that condition (3) is satisfied. \end{proof} It is also possible to state a proposition like (8) to the bounds on the distillable entanglement. $E^{1}$ will in this case provide an upper bound to $E_{D}$ of all bipartite partitions. \section{Lower bounds for the entanglement of formation} One of the most celebrated entanglement measures is the entanglement of formation \cite{bennett2} \begin{equation} E_{F}(\rho) = \min_{p_{1}, \psi_{i}}\sum_{i}p_{i}E_{E}(\ket{\psi_{i}}), \end{equation} where $E_{E}$ is the entropy of entanglement. Although this measure has a very meaningful physical interpretation and good properties, its calculation has been done only for a very class of states \cite{terhal5}. We show in this section that any entanglement witness can be used to provide lower bounds to the entanglement of formation. Let $\rho = \ket{\Psi}\bra{\Psi}$ be a pure bipartite state with the following Schimdt decomposition: \begin{equation} \ket{\Psi} = \sum_{j=1}^{d} c_{j} \ket{jj} , \hspace{0.4 cm} c_{1} \geq c_{2} \geq ... \geq c_{d}. \end{equation} An analytic expression for the random robustness $R_{r}$ and the generalized robustness $R_{G}$ of a pure state given by equation (78) is \cite{vidal4} \begin{equation} R_{G}(\rho) = \left(\sum_{j=1}^{d}c_{j} \right)^{2} - 1, \end{equation} \begin{equation} R_{r}(\rho) = c_{1}c_{2}. \end{equation} We start with two bounds for the entropy of entanglement, i.e., the von Neumann entropy of the reduced density matrix of a pure state $\ket{\psi}$. In the case of two qubits, Wootters has shown that \cite{wootters1} \begin{equation*} H\left( \frac{1 + \sqrt{1 - 4c_{1}^{2}c_{2}^{2}}}{2}\right) = E_{E}(\ket{\psi}), \end{equation*} where $H(x) = -x\log(x) - (1 - x)\log(1 - x)$. That is a particular case of the more general inequality \begin{equation} H\left(\frac{1 + \sqrt{1 - 4c_{1}^{2}c_{2}^{2}}}{2}\right) \leq -\sum_{i}^{d}c_{i}^{2}\log(c_{i}^{2}), \hspace{0.2 cm} \sum_{i}^{d}c_{i}^{2} = 1. \end{equation} Another similar inequality is \begin{equation} \frac{\log(d) - 1}{d}\left[\left(\sum_{i}^{d}c_{i}\right)- 1\right] \leq -\sum_{i}^{d}c_{i}^{2}\log(c_{i}^{2}). \end{equation} Equations (72-73) can be proved maximizing the L.H.S. minus the R.H.S. and noting that the maximum is null in both cases. Choosing ${\cal f}p_{i}, \ket{\psi_{i}}{\cal g}$ to be an optimal ensemble in equation (71), we have \begin{eqnarray} E_{F}(\rho) = \sum_{i}p_{i}E_{E}(\ket{\psi_{i}}) \geq \sum_{i}p_{i}H\left(\frac{1 + \sqrt{1 - 4(c_{1}^{2})_{i}(c_{2}^{2})_{i}}}{2}\right) \nonumber \\ \geq H\left(\frac{1 + \sqrt{1 - 4R_{r}^{2}(\rho)}}{2}\right), \nonumber \end{eqnarray} where we have used the convexity of $R_{r}$ and $f(x) = H\left(\frac{1 + \sqrt{1 - 4x^{2}}}{2}\right)$. Similarly, we find \begin{equation} E_{F}(\rho) \geq \frac{\log(d) - 1}{d}R_{G}(\rho). \end{equation} The bound derived from $R_{r}$ is suitable for slightly entangled states, where the Schimdt coefficients of the optimal $\ket{\psi_{i}}$ decay fast enough, making the truncation in the second Schimdt coefficient a good approximation. As an first example, we consider the Horodecki 3x3 states \cite{horodecki6}. These states exhibit bound entanglement, since they have positive partial transposition. They are given by \begin{equation} \rho(a) = \left[ \begin{array}{ccccccccc} a & 0 & 0 & 0 & a & 0 & 0 & 0 & a \\ 0 & 0 & a & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & a & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & a & 0 & 0 & 0 & 0 \\ a & 0 & 0 & 0 & a & 0 & 0 & 0 & a \\ 0 & 0 & 0 & 0 & 0 & a & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & \frac{1 + a}{2} & 0 & \frac{\sqrt{1 - a^{2}}}{2}\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & a & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & \frac{\sqrt{1 - a^{2}}}{2} & 0 & \frac{1 + a}{2} \end{array} \right] \end{equation} This family of states is interesting to test the first bound since their entanglement of formation was numerically calculated by Audenaert {\it et al} \cite{audenaert} and was found to be very low. Figure (7) shows the bound provided by $R_{r}$ for the states $\rho' = e\rho(a) + (1 - e)(I/D)$. \begin{figure} \begin{center} \includegraphics[scale=0.45]{horodecki.eps} \caption{(Coloronline) Lower bound for the Horodecki states using $R_{r}$.} \end{center} \end{figure} For our second example we consider one of the unique states for which a analytic formula for $E_{F}$ is know. It was shown by Terhal and Vollbrecht \cite{terhal5} that for the isotropic states \cite{foot13} \begin{equation*} \rho_{F} = \frac{1 - F}{d^{2} - 1}\left( I - P^{+} \right) + FP^{+} \end{equation*} \begin{equation} E_{F}(\rho_{F}) = \frac{d\log(d - 1)}{d - 2}(F - 1) + \log(d), \hspace{0.2 cm} F \in \left[\frac{4(d - 1)}{d^{2}}, 1\right] \end{equation} From section (VII) we find \begin{equation} E_{F}(\rho_{F}) \geq (\log(d) - 1)\left(F - \frac{1}{d} \right) \end{equation} We see that, in this case, for sufficiently large d, the difference of the bound and the actual value of $E_{F}$ is always less than F. Note that every entanglement witness $W$, after being normalized such that either $Tr(W) = 1$ or $W \leq I$ holds, can be used to deliver lower bounds to the entanglement of formation. \section{Entanglement, Thermodynamics and Lattice Systems} The study of entanglement properties of many-body systems, manly condensed matter, has received much attention recently \cite{nielsen2,arnesen,wang1,wang5,bose1,connor1,osborne1,osterloh,toth1,vidal6,ghose,brukner1}. Several important models have been analyzed and connections with thermodynamic variables, such as internal energy and magnetization, have been raised \cite{wang5,bose1,connor1,osborne1,toth1}. The negativity and concurrence have been the most used measures, partly due their easy computation, but also because they made possible the derivation of some interesting simple {\it thermodynamics like} equations. This can be understood from the view of the {\it witnessed entanglement}. Every quantie derived from (7) not only defines a measure of the degree of entanglement, but also gives a Hermitian operator, which vary for each state, whose expectation value quantifies the entanglement of the state in question. It is exactly the possibility of measure experimentally the amount of entanglement, which is a feature shared by all common thermodynamics variables, that makes quantities expressed by (7) useful to the study of entanglement thermodynamical properties. In this section we present, as an example, the study of entanglement in the XXX Heisenberg model with and without a magnetic field using $E_{\infty:1} = R_{G}$. The corresponding Hamiltonian is given by \begin{equation} H_{XXX} = J\sum_{i=1}^{N}\vec{\sigma}_{i}.\vec{\sigma}_{i+1} + B\sum_{i=1}^{N}\sigma_{i}^{z}. \end{equation} We first consider $B = 0$, in which case both Hamiltonians have $SU(2)$ symmetry. According to section (VII), we can restrict the EWs in (7) to the ones that also have $SU(2)$ symmetry. Then, from a standard result from representation theory \cite{eggeling1,weyl}, we find that all EWs with this symmetry can be written as \begin{equation} W = \sum_{i}\mu_{i}V_{i}, \end{equation} where $V_{i}$ are unitary permutation operators. From analytic and numerical studies for the $XXX$ Heisenberg model of odd $N$ in the fundamental state and in the thermodynamical limit \cite{sakai}, we find that all other correlators are very small compared to the first neighbor correlators. We, thus, use the following ansatz for the optimal entanglement witness for the thermal states, at very low temperatures, of Hamiltonians (79) and (80) \begin{equation} W = \left(NI + \sum_{i=1}^{N}\left( \sigma_{i}^{x}\sigma_{i+1}^{x} + \sigma_{i}^{y}\sigma_{i+1}^{y} + \sigma_{i}^{z}\sigma_{i+1}^{z}\right)\right)/2N, \end{equation} where the factor 2$N$ in the denominator comes from $W \leq I$. Note that this is the EW introduced by Toth {\it et al} \cite{toth5}. Assuming that $|B| << |J|$, and using the continuity of OEWs, we find that for the XXX Heisenberg model, at temperatures sufficiently close to zero, \begin{equation} R_{G} \approx \frac{U - BM}{2NJ} - \frac{1}{2}, \end{equation} where the magnetization and the internal energy are given, respectively, by $M = \sum_{i}\left<\sigma_{i}^{z}\right>$ and $U = \left<H\right>$. We now proceed analysing the relation between entanglement and the magnetic susceptibility ($\chi$) in thermal states of $H_{2}$. According to Brukner, Vedral and Zeilinger \cite{brukner1}, under temperatures close to zero and at zero external magnetic field, $\chi = (g^{2}\mu_{B}^{2}/kT)[NI + (1/3)\sum_{i}\vec{\sigma_{i}}.\vec{\sigma_{i+1}}]$. Thus, \begin{equation} \chi \approx \frac{2Ng^{2}\mu_{B}^{2}}{3kT} + \frac{2NR_{G}}{3kT}. \end{equation} Remarkably, we see that the susceptibility is given by a term which resembles the classical Curie law more a term which takes into account the entanglement presented in the state. The equation above can be seen as a quantitative version of the experimental result of Ghose {\it et al} \cite{ghose}, who shown that at very low temperatures the magnetic susceptibility of certain materials is affected by the existence of entanglement. \section{Conclusion} Summarizing, we have presented a new perspective to the quantification of entanglement based on witness operators. Several important EMs were shown to fit into this scenario and a new infinite family of EMs was introduced. The usefulness of the {\it witnessed entanglement} was illustrated by the study of diverse features of entanglemnt, including super-selection rules constraints and efficiency of quantum information protocols. Finally, we have shown some interesting preliminary results in the study of thermodynamical properties of entanglement in macroscopic systems. We believe the results presented in this paper are only preliminary. The quantification on entanglement with EWs might be a very fruitful approach to development of the theory of entanglement, specially in the new applications of entanglement, such as in identifying quantum phase transitions and improving the approximation of mean field theories \cite{vedral5}. \section{acknowledgment} The author would like to thank J. Eisert and D. Santos for helpful comments and especially Michal Horodecki for pointing out a mistake in a earlier version of this draft. Financial support from CNPQ is also acknowledged.
{ "timestamp": "2006-07-29T10:18:03", "yymm": "0503", "arxiv_id": "quant-ph/0503152", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503152" }
\section{Introduction} Recent Au+Au collision experiments at the Relativistic Heavy Ion Collider (RHIC) saw a dramatic suppression of hadrons with high transverse momenta (``high-$p_T$ suppression'') \cite{Whitepapers}, and the quenching of jets in the direction opposite to a high-$p_T$ trigger particle \cite{STARjetqu,PHENIXjetqu}, when compared with p+p and d+Au collisions. This is taken as evidence for the creation of a very dense, color opaque medium of deconfined quarks and gluons \cite{QGP3jetqu}. Independent evidence for the creation of dense, thermalized quark-gluon matter, yielding comparable estimates for its initial energy density ($\langle e\rangle\agt 10$\,GeV/fm$^3$ at time $\tau_{\rm therm}\alt0.6$\,fm/$c$ \cite{QGP3v2}), comes from the observation of strong elliptic flow in non-central Au+Au collisions \cite{Whitepapers}, consistent with ideal fluid dynamical behaviour of the bulk of the matter produced in these collisions. These two observations raise the question what happens, in the small fraction of collision events where a hard scattering produces a pair of high-$p_T$ partons, to the energy lost by the parton travelling through the medium. The STAR Collaboration has shown that, while in central Au+Au collisions there are no {\em hard} particles left in the direction opposite to a high-$p_T$ trigger particle, one sees enhanced production (compared to p+p) of {\em soft} (low-$p_T$) particles, broadly distributed over the hemisphere diametrically opposite to the trigger particle \cite{STARjet_therm}. This shows that the energy of the fast parton originally emitted in the direction opposite to the trigger particle is not lost, but severely degraded by interactions with the medium. As the impact parameter of the collisions decreases, the average momentum of the particles emitted opposite to the trigger particle approaches the mean value associated with {\em all} soft hadrons, i.e. the $\langle p_T\rangle$ of the thermalized medium \cite{STARjet_therm}. This suggests that the energy lost by the fast parton has been largely thermalized. Nevertheless, this energy is deposited locally along the fast parton's trajectory, leading to local energy density inhomogeneities which, if thermalized, should in turn evolve hydrodynamically. This would modify the usual hydrodynamic expansion of the collision fireball as observed in the overwhelming number of soft collision events where no high-$p_T$ partons are created. Since the fast parton moves at supersonic speed, it was suggested in Ref.~\cite{shuryak} that a Mach shock (``sonic boom'') should develop, resulting in {\em conical flow} and preferred particle emission at a specific angle away from the direction of the fast parton which lost its energy. This Mach angle is sensitive to the medium's speed of sound $c_s$ and thus offers the possibility to measure one of its key properties. A recent analysis by the PHENIX Collaboration \cite{Adler:2005ee} of azimuthal di-hadron correlations in $200\,A$\,GeV Au+Au collisions revealed structures in the angular distribution which might be suggestive of conical flow. The idea of Mach shock waves travelling through compressed nuclear matter was first advocated 30 years ago \cite{shock,SGM74}, but RHIC collisions for the first time exhibit \cite{QGP3v2} the kind of ideal fluid behaviour which might make an extraction of the speed of sound conceivable. An alternate scenario, in which the color wake field generated by the fast colored parton travelling through a quark-gluon plasma accelerates soft colored plasma particles in the direction perpendicular to the wake front \cite{Stocker,RM05}, leads to an emission pattern which is sensitive to the propagation of {\em plasma} rather than {\em sound} waves \cite{RM05}. In a strongly coupled plasma with overdamped plasma oscillations, which seems to be the preferred interpretation of RHIC data \cite{SQCD,sQGP}, the wake field scenario should reduce to the hydrodynamic Mach cone picture. We here study the dynamical consequences of the latter, going beyond the discussion of linearized hydrodynamic equations in a static background offered in \cite{shuryak}. We assume that just before hydrodynamics become applicable, a pair of high-$p_T$ partons is produced near the surface of the fireball. One of them moves outward and escapes, forming the trigger jet, while the other enters into the fireball along, say, the $-x$ direction. The fireball is expanding and cooling. The ingoing parton travels at the speed of light and loses energy in the fireball which thermalizes and acts as a source of energy and momentum for the fireball medium. We model this medium as an ideal fluid with vanishing net baryon density. Its dynamics is controlled by the energy-momentum conservation equations \begin{equation} \label{1} \partial_\mu T^{\mu\nu} =J^\nu, \end{equation} where the energy-momentum tensor has the ideal fluid form $T^{\mu\nu}\eq(\varepsilon{+}p)u^\mu u^\nu -p g^{\mu\nu}$, with energy density $\varepsilon$ and pressure $p$ being related by the equation of state (EOS) $p{\eq}p(\varepsilon)$, $u^\mu\eq\gamma(1,v_x,v_y,0)$ is the fluid 4-velocity, and the source current $J^\nu$ is given by \begin{eqnarray} \label{2} &&J^\nu(x)=J(x)\,\bigl(1,-1,0,0\bigr),\\ \label{3} &&J(x) = \frac{dE}{dx}(x)\, \left|\frac{dx_{\rm jet}}{dt}\right| \delta^3(\bm{r}-\bm{r}_{\rm jet}(t)). \end{eqnarray} Massless partons have light-like 4-momentum, so the current $J^\nu$ describing the 4-momentum lost and deposited in the medium by the fast parton is taken to be light-like, too. $\bm{r}_{\rm jet}(t)$ is the trajectory of the jet moving with speed $|dx_{\rm jet}/dt|\eq{c}$. $\frac{dE}{dx}(x)$ is the energy loss rate of the parton as it moves through the liquid. It depends on the fluid's local rest frame particle density. Taking guidance from the phenomenological analysis of parton energy loss observed in Au+Au collisions at RHIC \cite{Eloss} we take \begin{equation} \label{4} \frac{dE}{dx} = \frac{s(x)}{s_0} \left.\frac{dE}{dx}\right|_0 \end{equation} where $s(x)$ is the local entropy density without the jet. The measured suppression of high-$p_T$ particle production in Au+Au collisions at RHIC was shown to be consistent with a parton energy loss of $\left.\frac{dE}{dx}\right|_0\eq14$\,GeV/fm at a reference entropy density of $s_0\eq140$\,fm$^{-3}$ \cite{Eloss}. For comparison, we also perform simulations with ten times larger energy loss, $\left.\frac{dE}{dx}\right|_0\eq140$\,GeV/fm. For the hydrodynamic evolution we use AZHYDRO \cite{AZHYDRO,QGP3v2}, the only publicly available relativistic hydrodynamic code for anisotropic transverse expansion. This algorithm is formulated in $(\tau,x,y,\eta)$ coordinates, where $\tau{=}\sqrt{t^2{-}z^2}$ is the longitudinal proper time, $\eta{=}\frac{1}{2}\ln\left[\frac{t{+}z}{t{-}z}\right]$ is space-time rapidity, and $\bm{r}_\perp{\,=\,}(x,y)$ defines the plane transverse to the beam direction $z$. AZHYDRO employs longitudinal boost invariance along $z$ but this is violated by the source term (\ref{3}). We therefore modify the latter by replacing the $\delta$-function in (\ref{3}) by \begin{eqnarray} \label{5} \delta^3(\bm{r}-\bm{r}_{\rm jet}(t)) &\longrightarrow& \frac{1}{\tau}\,\delta(x-x_{\rm jet}(\tau))\,\delta(y-y_{\rm jet}(\tau)) \nonumber\\ &\longrightarrow&\frac{1}{\tau} \, \frac{e^{-(\bm{r}_\perp-\bm{r}_{\perp,{\rm jet}}(\tau))^2/(2\sigma^2)}} {2\pi\sigma^2} \end{eqnarray} with $\sigma{\,=\,}0.35$\,fm. Intuitively, this replaces the ``needle'' (jet) pushing through the medium at one point by a ``knife'' cutting the medium along its entire length along the beam direction. The resulting ``wedge flow'' is expected to leave a stronger signal in the azimuthal particle distribution $dN/d\phi$ than ``conical flow'' induced by a single parton, since in the latter case one performs an implicit $\phi$-average when summing over all directions of the cone normal vector. While a complete study of this would require a full (3+1)-dimensional hydrodynamic calculation, the present boost-invariant simulation should give a robust upper limit for the expected angular signatures. We show that the angular structures predicted from wedge flow are too weak to explain the experimentally observed $\phi$-modulation \cite{Adler:2005ee}. The modified hydrodynamic equations in $(\tau,x,y,\eta)$ coordinates read \cite{AZHYDRO} \begin{eqnarray} \label{6} \partial_\tau \tilde{T}^{\tau \tau} + \partial_x(\tilde{v}_x \tilde{T}^{\tau \tau}) + \partial_y(\tilde{v}_y \tilde{T}^{\tau \tau}) &=& - p + \tilde{J}, \\ \label{7} \partial_\tau \tilde{T}^{\tau x} + \partial_x(v_x \tilde{T}^{\tau x}) + \partial_y(v_y \tilde{T}^{\tau x}) &=& - \partial_x \tilde{p} - \tilde{J}, \quad \\ \label{8} \partial_\tau \tilde{T}^{\tau y} + \partial_x(v_x \tilde{T}^{\tau y}) + \partial_y(v_y \tilde{T}^{\tau y}) &=& -\partial_y \tilde{p}, \quad \end{eqnarray} where $\tilde{T}^{\mu\nu}\eq\tau T^{\mu\nu}$, $\tilde{v}_i{\eq}T^{\tau i}/T^{\tau\tau}$, $\tilde p\eq\tau p$, and $\tilde{J}\eq\tau J$. To simulate central Au+Au collisions at RHIC, we use the standard initialization described in \cite{QGP3v2} and provided in the downloaded AZHYDRO input file \cite{AZHYDRO}, corresponding to a peak initial energy density of $\varepsilon_0\eq30$\,GeV/fm$^3$ at $\tau_0\eq0.6$\,fm/$c$. We use the equation of state EOS-Q described in \cite{QGP3v2,AZHYDRO} incorporating a first order phase transition and hadronic chemical freeze-out at a critical temperature $T_c{\,=\,}164$\,MeV. The hadronic sector of EOS-Q is soft with a squared speed of sound $c_s^2 \approx 0.15$. \begin{figure}[t] \includegraphics[bb=14 50 581 820,width=0.99\linewidth,clip]{Fig1.eps} \caption{Contours of constant local energy density in the $x$-$y$ plane at three different times, $\tau{\,=\,}4.6,$ 8.6, and 12.6 fm/$c$. In each case the position of the fast parton, along with the integrated energy loss $\Delta E=\int J(x) dxdy d\tau$ up to this point, is indicated at the top of the figure. Diagrams (a)-(c) in the left column were calculated with a reference energy loss $dE/dx|_0{\,=\,}14$\,GeV/fm, those in the right column (panels (d)-(f)) with a 10 times larger value. } \label{F1} \end{figure} In our study the quenching jet starts from $x_{\rm jet}\eq6.4$\,fm at $\tau_0\eq0.6$\,fm, moving left towards the center with constant speed $v_{\rm jet}\eq{c}$. For an upper limit on conical flow effects, the fast parton is assumed to have sufficient initial energy to emerge on the other side of the fireball. To simulate cases where the fast parton has insufficient energy to fully traverse the medium we have also done simulations where the parton loses all its energy within an (arbitrarily chosen) distance of 6.4 fm. We further compared with a run where the parton moved at (constant) subsonic speed ($v_\mathrm{jet}{\,=\,}0.2\,c{\,<\,}c_s$). The resulting evolution of the energy density of the QGP fluid is shown in Figs.~\ref{F1} and \ref{F2}. The left column of Fig.~\ref{F1} shows results for a phenomenologically acceptable value $\left.\frac{dE}{dx}\right|_0\eq14$\,GeV/fm \cite{Eloss} for the reference parton energy loss whereas in all other columns we use a ten times larger energy loss. The width of the Gaussian source (see Eq.~(\ref{5})) is $\sigma$=0.7 fm. In the left column of Fig.~\ref{F1} the effects of the energy deposition from the fast parton are hardly visible. Only for a much larger energy loss (right column) we recognize a clear conical flow pattern. The accumulating wave fronts from the expanding energy density waves build up a ``sonic boom'' shock front which creates a Mach cone. The right columns in Figs.~\ref{F1} and \ref{F2} show that the cone normal vector forms an angle $\theta_M$ with the direction of the quenching jet that is qualitatively consistent with expectations from the theoretical relation $\cos\theta_M{\eq}c_s/v_\mathrm{jet}$. However, this angle is not sharply defined since the cone surface curves due to inhomogeneity and radial expansion of the underlying medium. This differs from the static homogeneous case \cite{shuryak}. \begin{figure}[t] \includegraphics[bb=14 50 581 820,width=0.99\linewidth,clip]{Fig2.eps} \caption{ As in Fig.\ref{F1} but for a parton moving at subsonic speed $v_\mathrm{jet}{\,=\,}0.2\,c$ (left column) and for a fast parton ($v_\mathrm{jet}{\,=\,}c$) which loses all its energy within the first 6.4\,fm (right column). } \label{F2} \end{figure} When the parton travels at a subsonic speed $v_\mathrm{jet}{\,=\,}0.2\,c{\,<\,}c_s$ (left column in Fig.~\ref{F2}), it doesn't get very far before the fireball freezes out due to longitudinal expansion. In this case one only observes an accumulation of energy around the parton but no evidence of Mach cone formation. When the parton travels with $v_\mathrm{jet}{\,=\,}c$ but looses all its energy after 6.4\,fm before fully traversing the fireball (right column in Fig.~\ref{F2}), the fireball evolution beyond $\tau$=7\,fm is not affected by the fast parton directly but only indirectly through the propagation of earlier deposited energy. Still, Figs.~\ref{F1}f and \ref{F2}f show that the late time evolution of the fireball is quite similar in both cases, demonstrating that energy deposition by the fast parton during the late fireball stages is small, due to dilution of the matter, and can almost be neglected. Tests with different values for the width $\sigma$ of the Gaussian profile in Eq.~(\ref{5}) for the deposited energy show that the cone angle gets better defined for smaller source size $\sigma$. Note that the quenching jet destroys the azimuthal symmetry of the initial energy density distribution but leaves the azimuthally symmetric energy contours to the left of the jet unaffected. We close with a discussion of observable consequences of conical flow. One expects \cite{shuryak} azimuthally anisotropic particle emission, peaking at angles $\phi\eq\pi{\,\pm\,}\theta_M$ relative to the trigger jet where $\theta_M$ is the Mach angle. Using the standard Cooper-Frye prescription, we have computed the angular distribution of directly emitted pions at a freeze-out temperature $T_\mathrm{fo}{\,=\,}100$\,MeV \cite{QGP3v2}. \begin{figure}[t] \includegraphics[bb=14 13 550 800,width=0.99\linewidth,clip=]% {Fig3.eps} \vspace{-4.5cm} \caption{Azimuthal distribution $dN/dy d\phi$ of negative pions per unit rapidity. In the upper panel we integrate over all $p_T$ while the lower panel shows only pions with $p_T{\,>\,}1$\,GeV/$c$. Different symbols refer to different parameters as indicated. For better visibility the $\phi$-independent rate in the absence of the quenching jet has been subtracted. Filled symbols show the realistic case $dE/dx|_0{\,=\,}14$\,GeV/fm, enhanced by a factor 10 for visibility. } \label{F3} \end{figure} Figure~\ref{F3} shows the azimuthal distribution of $\pi^-$ from a variety of different simulations. [The $\phi$-independent constant $(dN_{\pi^-}/dy d\phi)_\mathrm{no\ jet}=27$ from central collisions without jets has been subtracted.] For $dE/dx|_0=14$\,GeV/fm the azimuthal modulation is very small. In none of the cases studied we find peaks at the predicted Mach angle with an associated dip in the direction of the quenched jet at $\phi\eq\pi$. One rather sees a {\em peak} at $\phi\eq\pi$, broadened by shoulders on both sides which turn into small peaks at (relative to the quenching jet) backward angles if thermal smearing is reduced by considering only high-$p_T$ pions. The peak at $\phi\eq\pi$ is absent when the parton loses all its energy halfway through the fireball or is too slow to get to the other side before freeze-out, suggesting that it reflects the directed momentum imparted on the medium by the fast parton. It is slightly more accentuated for smaller $\sigma$ and higher $p_T$. We also found that the width of the shoulders is almost independent of the speed of sound of the medium and can not be used to diagnose the stiffness of its equation of state. The shoulders exist even for subsonic parton propagation ($v_\mathrm{jet}\eq0.2\,c$, upright triangles), showing that other mechanisms (such as backsplash from the hard parton hitting the fluid and a general bias of the energy deposition towards the right side of the fireball due to the higher density of the medium at early times) have a strong influence on the angular distribution of the emitted particles which interferes with the position of the Mach peaks. The shoulder resulting from this combination of effects is much broader than the angular structures seen in the data \cite{Adler:2005ee}. The absence of a clear dip at $\phi\eq\pi$ in our simulations is all the more troubling since it should have been stronger for the ``wedge flow'' studied here than for real ``conical'' flow. Our calculation does not average over the initial production points of the trigger particle, i.e. it ignores that in most cases its quenching partner does not travel right through the middle of the fireball cylinder, but traverses it semi-tangentially. This should further decrease the prominence of the shoulders in $dN/d\phi$. We conclude that conical flow may be able to explain the broadening of the away-side peak in the hadron angular correlation function around $\phi\eq\pi$ pointed out by the STAR Collaboration \cite{STARjet_therm}, but is unlikely to be responsible for the relatively sharp structures near $\phi\eq\pi{\pm}1$ seen by PHENIX \cite{Adler:2005ee}. This conclusion extends to other conical flow phenomena, such as those generated by color wake fields \cite{Stocker,RM05}. This work was supported by the U.S. Department of Energy under contract DE-FG02-01ER41190.
{ "timestamp": "2006-04-10T14:47:39", "yymm": "0503", "arxiv_id": "nucl-th/0503028", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503028" }
\section{Introduction} Let $(X,o)$ be a normal complex surface singularity germ. Let $\Gamma$ and $\Sigma$ denote the resolution graph and the link of $(X,o)$, respectively. We assume that $X$ is homeomorphic to the cone over $\Sigma$. It is known that the resolution graph and the link of the singularity determine each other (\cite{neumann.plumbing}). Assume that the link $\Sigma$ is a $\mathbb Q$-homology sphere, or equivalently, that the exceptional set of a good resolution is a tree of rational curves. Then $H_1(\Sigma, \mathbb Z)$ is finite. A morphism $(Y,o) \to (X,o)$ of germs of normal surface singularities is called a {\itshape universal abelian covering} if it induces an unramified Galois covering $Y \setminus \{o\} \to X \setminus \{o\}$ with covering transformation group $H_1(\Sigma,\mathbb Z)$. By our assumption, the universal abelian covering $(Y,o) \to (X,o)$ must exist; in fact, the link of $Y$ is the universal abelian cover of $\Sigma$ in the topological sense. We are interested in the analytic properties of $Y$ and a way to construct $Y$ explicitly. In the case that $(X,o)$ is quasihomogeneous, Neumann \cite{neumann.abel} proved that the universal abelian cover $(Y,o)$ is a Brieskorn-Pham complete intersection, by writing down the explicit equations from the data of $\Sigma$ (it is known that $\Gamma$ is star-shaped in this case). Neumann and Wahl generalized Brieskorn-Pham complete intersections and the way to construct them, and obtained numerous interesting results; see \cite{nw-uac}, \cite{nw-HSL}, \cite{nw-qcusp}, \cite{nw-CIuac}. They introduced the {\itshape splice diagram equations} (or {\itshape forms}) associated with a weighted tree called a {\itshape splice diagram} satisfying the ``semigroup condition''. From an arbitrary resolution graph corresponding to a $\mathbb Q$-homology sphere link, we can construct a splice diagram. Let $\widetilde Y$ denote the singularity defined by the splice diagram equations obtained from $\Gamma$. They proved that $\widetilde Y$ is an isolated complete intersection surface singularity, and that (under ``congruence condition'') if the equations are chosen so that the {\itshape discriminant group} $G$ ($\cong H_1(\Sigma, \mathbb Z)$) for $\Gamma$ naturally acts on $\widetilde Y$, then the quotient $\widetilde Y/G$ is a normal surface singularity (it is called a {\itshape splice-quotient singularity}) with resolution graph $\Gamma$, and the quotient morphism is the universal abelian covering. Neumann and Wahl conjectured that rational singularities and minimally elliptic singularities with $\mathbb Q$-homology sphere links are splice-quotient singularities. However, it is not known whether the splice diagrams obtained from the resolution graphs of those singularities satisfy the semigroup condition. In this paper we prove that the universal abelian cover of a rational or minimally elliptic singularity is a complete intersection singularity defined by certain special functions. Let $\pi\: M \to X$ be a good resolution with the exceptional set $A$. Under a topological condition (\condref{c:A}), we can associate a collection of certain special polynomials and a system of weights with each node of $A$. These polynomials are quasihomogeneous with respect to the weights. We call the union of those collections over all nodes a Neumann-Wahl system, which is an analogue of the system of splice diagram equations. Though the definition of a Neumann-Wahl system is very similar to that of splice diagram equations, they are not the same (see \remref{r:diff}). We suspect that a Neumann-Wahl system is a special type of system of splice diagram forms. Our first result is the following (see \thmref{t:V}). \begin{thm}\label{t:Vintro} Let $(V,o)$ be a singularity defined by a Neumann-Wahl system. Then $(V,o)$ is an isolated complete intersection surface singularity. A singularity defined by functions obtained by adding ``higher terms'' to the Neumann-Wahl system is an equisingular deformation of $(V,o)$. \end{thm} We call a singularity defined by a Neumann-Wahl system a Neumann-Wahl complete intersection. If in addition a certain analytic condition (\condref{c:B}) and a topological condition (\condref{c:C}), which is stronger than \condref{c:A}, are satisfied, then the universal abelian cover $Y$ is an equisingular deformation of a Neumann-Wahl complete intersection (\thmref{t:main-general}). The equations of $Y$ are constructed on the resolution space $M$. Our equations for the deformation are automatically equivalent with respect to a natural action of the discriminant group $G$, and the action is free on nonsingular locus. An important point is that \condref{c:C} and \ref{c:B} are satisfied in case $(X,o)$ is a rational or minimally elliptic singularity (it can be easily verified!). Thus we have the following (see \thmref{t:main} and \proref{p:esquotient}). \begin{thm}\label{t:mainintro} If $(X,o)$ is a rational or minimally elliptic singularity, then its universal abelian cover $(Y,o)$ is an equisingular deformation of a Neumann-Wahl complete intersection singularity $(Y_0,o)$. Moreover the $(X,o)$ is an equisingular deformation of $(Y_0/G,o)$. \end{thm} This paper is organized as follows. In \sref{s:pre}, we briefly review fundamental results on the universal abelian covers of normal surface singularities in \cite{o.uac-rat}. We recall there that $\mathcal O_{Y,o}$ is isomorphic to an $\mathcal O_{X,o}$-algebra $\mathcal A:=\bigoplus _{b \in \mathcal B}H^0(-L^{(b)})$, where $\mathcal B$ is a group isomorphic to $G$ and $L^{(b)}$ are suitable divisors on $M$. In \sref{s:NWS}, we first define monomial cycles. The semigroup of monomial cycles is naturally isomorphic to that of monomials; the variables are associated with the ends of $A$. We show that \condref{c:C} implies \condref{c:A}, and is satisfied for rational or minimally elliptic singularities. Then we introduce the Neumann-Wahl systems and the weights. In \sref{s:VdefbyF}, we prove a slight generalization of \thmref{t:Vintro}; there we consider a system of polynomials obtained by substituting some power of the variables into the variables of a Neumann-Wahl system. We will apply some ideas in \cite[\S 2]{nw-CIuac} for the proof. In the last section, we prove \thmref{t:mainintro}. We construct the equations by looking at certain relations of sections of $H^0(-L^{(b)})$'s. We can take a good basis of the algebra $\mathcal A$ by \condref{c:B}, and can explicitly obtain the relations by \condref{c:C}. \section{Preliminaries}\label{s:pre} In this section, we recall some fundamental results on the universal abelian covers of surface singularities; see \cite{o.uac-rat} for details. Let $(X,o)$ be a normal complex surface singularity germ. We assume that the link $\Sigma$ of $(X,o)$ is a $\mathbb Q$-homology sphere and that $X$ is homeomorphic to a cone over $\Sigma$. Then $X$ has a unique universal abelian cover. Let $\pi\: M \to X$ be a resolution of the singularity, and let $A=\bigcup _iA_i$ be the decomposition of the exceptional set $A=\pi^{-1}(o)$ into irreducible components. Assume that $\pi$ is a good resolution, i.e., $A$ is a divisor having only simple normal crossings. Then the condition that $\Sigma$ is a $\mathbb Q$-homology sphere is equivalent to that $H^1(\mathcal O_A)=0$, i.e., $A$ is a tree of nonsingular rational curves. We call a divisor supported in $A$ a cycle. Let $A_{\mathbb Z}$ denote the group of cycles. An element of $A_{\mathbb Q}:=A_{\mathbb Z}\otimes \mathbb Q$ is called a $\mathbb Q$-cycle. Let $\du i \in A_{\mathbb Q}$ denote the dual cycle of $A_i$, i.e., the $\mathbb Q$-cycle satisfying $\du i\cdot A_j=-\delta _{ij}$, where $\delta _{ij}$ denotes the Kronecker delta. We denote by $\du{\mathbb Z}$ the subgroup of $A_{\mathbb Q}$ generated by $\du i$'s. Recall that the first homology group $H_1(\Sigma, \mathbb Z)$, the Galois group of the universal abelian covering of $X$, is isomorphic to the group $\du {\mathbb Z}/A_{\mathbb Z}$ called the discriminant group. The order of the group is $|\det(A_i\cdot A_j)|$. Let $D$ be a $\mathbb Q$-divisor on $M$. We denote by $\nu(D)$ the $\mathbb Q$-cycle satisfying $(\nu(D)-D)\cdot A_i=0$ for all $A_i$. We say that $D$ is $\pi$-anti-nef if $-D$ is $\pi$-nef, i.e., $D\cdot A_i \le 0$ for all $A_i$. Let $ \mathcal F (D)$ denote the set of $\pi$-anti-nef $\mathbb Q$-divisors $F$ satisfying $F-D\in A_{\mathbb Z}$. Note that $\mathcal F(D)$ has the minimum with respect to ``$\ge$''. We take effective $\mathbb Q$-cycles $E_1, \ldots , E_s$ such that if $\mathcal E_i$ denotes the cyclic subgroup of $\du {\mathbb Z}/A_{\mathbb Z}$ generated by $E_i$, then $\du {\mathbb Z}/A_{\mathbb Z}=\mathcal E_1 \oplus \cdots \oplus \mathcal E_s$. Let $r_i$ be the order of $\mathcal E_i$. Then for $1 \le i \le s$ there exists a divisor $L_i$ and a function $f_i$ on a suitable neighborhood of $A$ such that $r_iL_i- r_iE_i=\di (f_i)$. For any $b=(b_1,\dots,b_s)\in \mathbb Z^s$, we define a divisor $L^{(b)}$ by $$ L^{(b)}=\sum_{j=1}^s b_jL_j-\left[\sum_{j=1}^sb_jE_j\right]. $$ Let $\bar b_i$ denote the smallest nonnegative integer such that $r_i \mid b_i -\bar b_i$. Let $\bar b= (\bar b_1, \dots , \bar b_s)$ and $\mathcal B=\{\bar b| b \in \mathbb Z^s\}$; the set $\mathcal B$ is identified with the discriminant group $\du {\mathbb Z}/A_{\mathbb Z}$. We define a sheaf $\bar \mathcal A$ of $\mathcal O_M$-modules by $$ \bar \mathcal A=\bigoplus _{b \in \mathcal B}\mathcal O_M(-L^{(b)}). $$ The $\mathcal O_M$-algebra structure of $\bar \mathcal A $ is given by the composite $$ \mathcal O_M(-L^{(b)}) \otimes \mathcal O_M(-L^{(b')}) \to \mathcal O_M(-L^{(b)}-L^{(b')}) \subset \mathcal O_M(-L^{(b+b')}) $$ and the isomorphism $$ \mathcal O_M(-L^{(b+b')}) \to \mathcal O_M(-L^{(\overline{b+b'})}) $$ given by multiplying by $\prod _{b_i+b_i'\neq\overline{b_i+b'_i}}f_i^{-1}$. Then the natural projection $$ Y:=\specan_X \pi_*\bar\mathcal A \to X $$ is the universal abelian covering (see \cite[Theorem 3.4]{o.uac-rat}). The local ring $\mathcal O_{Y,o}$ of the singularity $(Y,o)$ is isomorphic to $$ \mathcal A:=(\pi_*\bar\mathcal A )_o=\bigoplus _{b \in \mathcal B}H^0(-L^{(b)}), $$ where $H^0(-L^{(b)})=\dlim _U H^0(U,\mathcal O_M(-L^{(b)}))$, $U$ varies over all open neighborhoods of $A$. We write $\mathcal A_b=H^0(-L^{(b)})$. Let $h \in \mathcal A_b$. If $\di (h)-L^{(b)}-D$ has no component of $A$ for some cycle $D\in A_{\mathbb Z}$, then we write $(h)_A=\nu(L^{(b)})+D$. \begin{lem}\label{l:product} Let $\sigma_i \in \mathcal A_{b^i}$, $i=1,2$, and let $\sigma_1\cdot \sigma_2\in \mathcal A_{\overline{b^1+b^2}}$ be the product of $\sigma_1$ and $\sigma_2$ in the algebra $\mathcal A$. Suppose that a divisor $L \in \mathcal F(L^{( \overline{b^1+b^2})})$ satisfies $\nu(L)=(\sigma_1)_A+(\sigma_2)_A$. Then $\sigma_1\cdot \sigma_2$ is a section of $H^0(-L) $ and $(\sigma_1\cdot \sigma_2)_A=\nu(L)$. \end{lem} \begin{proof} It follows from the definition of the algebra structure of $\mathcal A$. \end{proof} Let $D$ be a reduced and connected cycle. A component $A_i$ of $D$ is called an {\itshape end} of $D$ if $(D-A_i)\cdot A_i\le 1$. We denote by $\mathcal E(D)$ the set of ends of $D$. Assume that $\mathcal E(A)=\{A_1, \dots ,A_m\}$. For any $1 \le i \le m$, there uniquely exists a divisor $L^i$ such that $\nu (L^i)=\du i$ and $L^i \in \mathcal F(L^{(b)})$ for some $b \in \mathcal B$. If $(X,o)$ is rational, then there exists $y_i \in H^0(-L^i)$ such that $(y_i)_A=\nu(L^i)$ (cf. \lemref{l:satisfyB}). The next theorem follows from \cite[Theorem 7.5]{o.uac-rat} and its proof. \begin{thm}\label{t:repOY} Assume that $(X,o)$ is rational. Let $y_i \in H^0(-L^i)$, $1 \le i \le m$, be as above. Let $S=\mathbb C\{x_1, \dots ,x_m\}$ be the convergent power series ring. We define a homomorphism $\psi \: S \to \mathcal A=\mathcal O_{Y,o}$ of $\mathbb C$-algebras by $\psi(x_i)=y_i$. Then $\psi$ is surjective. \end{thm} In the last section, we give a proof of \thmref{t:repOY}, which is different from that in \cite{o.uac-rat}. In fact it is shown that the assertion also holds true for minimally elliptic singularities. \section{Neumann-Wahl systems}\label{s:NWS} In this section we will introduce a Neumann-Wahl system associated with the exceptional set $A$. It is a set of certain polynomials, and an analogue of the system of splice diagram equations in Neumann and Wahl's work (\cite{nw-uac}, \cite{nw-HSL}, \cite{nw-CIuac}). We use the notation of the preceding section, and keep the assumption that $H^1(\mathcal O_A)=0$. First assume that the set $\mathcal E(A)$ consists of the components $A_1, \dots ,A_m$. Let $\mathbb C[x_1, \ldots ,x_m]$ be the polynomial ring. \thmref{t:repOY} suggests us the following \begin{defn} Let $D=\sum a_i\du i \in A_{\mathbb Q}$, $a_i \ge 0$. If $a_i=0$ for all $A_i \notin \mathcal E(A)$, then we call $D$ a {\itshape $\mathbb Q$-monomial cycle}; if in addition $a_i \in \mathbb Z$ for all $i$, we call $D$ a {\itshape monomial cycle}. For any monomial cycle $D=\sum _{i=1}^ma_i\du i$, we associate a monomial $$ x(D):=\prod _{i=1}^mx_i^{a_i} \in \mathbb C[x_1, \ldots ,x_m]. $$ The $x$ induces an isomorphism between the semigroup of monomial cycles and that of monomials of $x_1, \ldots, x_m$. Formally, we may also consider the $\mathbb Q$-monomial $x(D)$ for a $\mathbb Q$-monomial cycle $D$. \end{defn} \begin{defn} For any $F=\sum a_kA_k \in A_{\mathbb Q}$, we write $m_{A_k}(F)=a_k$. For any component $A_j$, we define the {\itshape $A_{j}$-weight} of $x_i$ to be $m_{A_{j}}(\du i)$. Then the {\itshape $A_{j}$-degree} of a monomial $x(D)$ is defined to be $m_{A_{j}}(D)$. \end{defn} A connected component of $A-A_i$ is called a {\itshape branch} of $A_i$. A component $A_i$ is called a {\itshape node} if $(A-A_i)\cdot A_i\ge 3$. We consider the following conditions concerning the weighted dual graph of $A$. \begin{cond}\label{c:A} For any branch $C$ of any node $A_i$, there exists a monomial cycle $D$ such that $D-\du i$ is an effective integral cycle supported on $C$; in this case, we say that $D$ (or monomial $x(D)$) belongs to the branch $C$. \end{cond} Note that in general there may exist more than one monomials belonging to a branch. \begin{cond}\label{c:C} $A$ is star-shaped, or for any branch $C$ of any component $A_i\notin \mathcal E(A)$, the fundamental cycle $Z_C$ supported on $C$ satisfies $Z_C\cdot A_i=1$. \end{cond} \begin{lem}\label{l:existmonomials} \condref{c:C} implies \condref{c:A}, and is satisfied in the following cases: \begin{enumerate} \item $(X,o)$ is a rational singularity; \item $(X,o)$ is a minimally elliptic singularity, and the minimally elliptic cycle is supported on $A$ (this condition is satisfied on the minimal good resolution). \end{enumerate} \end{lem} \begin{proof} Assume that the condition (1) or (2) is satisfied. From basic results on the computation sequences for the fundamental cycle (\cite{la.rat}, \cite{la.me}), we obtain \condref{c:C}. Suppose that $C_{1}$ is a branch of a node $A_{i_1}$. Then $D_1:=\du {i_1}+Z_{C_1}$ is $\pi$-anti-nef and $D_1 \cdot A_j=0$ for every $A_j \le A-C_1$. If $D_1 \cdot A_{i_2}<0$ for $A_{i_2} \notin \mathcal E(A)$, then take a branch $C_2$ of $A_{i_2}$, not containing $A_{i_1}$, and put $D_2:=D_1+Z_{C_2}$. In this manner we obtain a finite sequence $\{D_1, \ldots ,D_n\}$ of $\pi$-anti-nef cycles, which ends with a monomial cycle belonging to $C_1$. Thus \condref{c:A} is satisfied. These arguments also show that \condref{c:A} holds in case $A$ is star-shaped. \end{proof} \begin{defn}\label{d:CSAM} Assume that \condref{c:A} is satisfied, and that $A_1, \ldots ,A_s$ are all of the nodes of $A$. Let $C_1, \dots ,C_p$ be the branches of $A_{1}$. (1) A monomial $x(D)$ is called an {\itshape admissible monomial} at the node $A_{1}$ if it belongs to one of the branches of $A_1$. A set of monomials $\{x(D_1), \dots ,x(D_p)\}$ is called a {\itshape complete system} of admissible monomials at $A_{1}$ if $D_i$ belongs to $C_i$ for $i=1, \ldots ,p$. (2) Let $\{m_1, \ldots ,m_p\}$ be any complete system of admissible monomials at $A_{1}$. Let $F=(c_{ij})$, $c_{ij} \in \mathbb C$, be a $((p-2) \times p)$-matrix such that every maximal minor of it has rank $p-2$. We define polynomials $f_1, \ldots ,f_{p-2}$ by $$ \begin{pmatrix} f_1\\ \vdots \\ f_{p-2} \end{pmatrix} = F \begin{pmatrix} m_1\\ \vdots \\ m_p \end{pmatrix}. $$ We call each $f_i$ an {\itshape admissible form} at $A_1$ and the set $\{f_1 , \dots ,f_{p-2}\}$ a {\itshape Neumann-Wahl system} at $A_{1}$. (3) Let $\mathcal F_i$ denote a Neumann-Wahl system at a node $A_{i}$. Then we call the set $ \bigcup _{i=1}^s\mathcal F_i$ a {\itshape Neumann-Wahl system} associated with $A$; it is an empty set in case $A$ has no nodes. \end{defn} \begin{rem}\label{r:normal} The matrix $F$ above can be reduced to the following matrix by row operations: $$ \begin{pmatrix} 1 & 0 & \dots & 0 & a_1 & b_1 \\ 0 & 1 & \dots & 0 & a_2 & b_2 \\ \vdots & \vdots & \ddots & \vdots & \vdots & \vdots \\ 0 & 0 & \dots & 1 & a_{p-2} & b_{p-2} \end{pmatrix}, $$ where $a_ib_j-a_jb_i\neq 0$ for $i\neq j$, and all $a_i$ and $b_i$ are nonzero. \end{rem} The admissible forms at a node $A_i$ are quasihomogeneous polynomials with respect to the $A_i$-weight. The following lemma is needed in the next section. \begin{lem}\label{l:higher} Let $D$ be a $\pi$-anti-nef $\mathbb Q$-cycle such that $m_{A_i}(D)>m_{A_i}(\du i)$ for some $i$. Then for any component $A_j$, we have $m_{A_j}(D)>m_{A_j}(\du i)$. \end{lem} \begin{proof} First we will show that $D \ge \du i$. There exist effective $\mathbb Q$-cycles $D_1$ and $D_2$ such that $D-\du i=D_1-D_2$ and that $D_1$ and $D_2$ have no common components. If $D_2>0$, then $$ m_{A_i}(D_2)=-\du i\cdot D_2=-D\cdot D_2+D_1\cdot D_2-D_2\cdot D_2>0. $$ It contradicts the assumption of the lemma. Hence $D-\du i\ge 0$. Let $a=m_{A_i}(D-\du i)$ and $C$ a branch of $A_i$. If a component $A_k$ of $C$ intersects $A_i$, then $(\du i+aA_i)\cdot A_k=a>0$. Since $D$ is $\pi$-anti-nef, there exists a $\mathbb Q$-cycle $C'$ supported on $C$ such that $C'\cdot A_l \le 0$ for all $A_l \le C$ and that $D \ge \du i+aA_i+C'$. \end{proof} \begin{rem}\label{r:diff} The definition of a Neumann-Wahl system is very similar to that of a system of splicing diagram equations in the Neumann and Wahl's work. However those are not the same. Let us consider a weighted graph $\Gamma$ represented as in \figref{f:Gamma}; the vertex \rule{2mm}{2mm} has weight $-4$ and other vertices $\bullet$ have weight $-2$. \begin{figure}[htbp] \begin{center} \setlength{\unitlength}{0.5cm} \begin{picture}(11,2)(0,0) \multiput(0,0)(0,2){2}{\circle*{0.25}} \multiput(5,1)(2,0){2}{\circle*{0.25}} \put(2.8,0.8){\rule{2mm}{2mm}} \multiput(9,0.5)(0,1){2}{\circle*{0.25}} \multiput(11,0)(0,2){2}{\circle*{0.25}} \put(3,1){\line(1,0){4}} \put(0,0){\line(3,1){3}} \put(0,2){\line(3,-1){3}} \put(7,1){\line(4,1){4}} \put(7,1){\line(4,-1){4}} \end{picture} \caption{\label{f:Gamma}} \end{center} \end{figure} The graph $\Gamma$ is realized as the resolution graph of an elliptic singularity with $\mathbb Q$-homology sphere link. The splice diagram $\Delta$ associated with $\Gamma$ is represented as in \figref{f:Delta}. \begin{figure}[htbp] \begin{center} \setlength{\unitlength}{0.5cm} \begin{picture}(10,2)(0,0) \multiput(0,0)(0,2){2}{\circle*{0.25}} \multiput(3,1)(4,0){2}{\circle*{0.25}} \multiput(10,0)(0,2){2}{\circle*{0.25}} \put(3,1){\line(1,0){4}} \put(0,0){\line(3,1){3}} \put(0,2){\line(3,-1){3}} \put(7,1){\line(3,1){3}} \put(7,1){\line(3,-1){3}} \put(2.2,1.5){$2$} \put(2.2,0){$2$} \put(7.5,1.5){$3$} \put(7.5,0){$3$} \put(3.5,1.2){$3$} \put(6,1.2){$20$} \end{picture} \caption{\label{f:Delta}} \end{center} \end{figure} We see that $\Delta$ satisfies semigroup condition. However, $\Gamma$ does not satisfy \condref{c:A}. Therefore we cannot define Neumann-Wahl systems in this case, though the splice diagram equations are defined. That might indicate a Neumann-Wahl system is a special type of system of splice diagram equations. \end{rem} \section{Varieties defined by Neumann-Wahl systems}\label{s:VdefbyF} In this section we prove that any Neumann-Wahl system defines a complete intersection surface with an isolated singularity at the origin; so we will call such a singularity a {\itshape Neumann-Wahl complete intersection singularity}. In fact, we will prove the assertion for a slight generalization of a Neumann-Wahl system. We note that some of methods in this section are discussed in \cite{nw-CIuac}. We use the notation of the preceding section. Assume that \condref{c:A} is satisfied and that $A$ has at least one node. As in the preceding section, we associate ends of $A$ with the variables $x_1, \ldots ,x_m$, where $m=\# \mathcal E(A)$. Suppose that $A_1, \ldots ,A_s$ are all of the nodes of $A$. Let $d_i$ denote the number of branches of a node $A_i$. Let $\mathcal M_i=\{m_{i1}, \ldots ,m_{id_i}\}$ denote a complete system of admissible monomials and $\mathcal F_i=\{f_{i1}, \ldots ,f_{id_i-2}\}$ a Neumann-Wahl system at a node $A_i$, where each $f_{ij}$ is a linear form of monomials of $\mathcal M_i$. By counting the numbers of the ends, the nodes and the edges around the nodes of the dual graph of $A$, we see that $$ \sum _{i=1}^s(d_i-2)=m-2. $$ Let $C_1, \ldots ,C_{d_1}$ denote the branches of $A_1$. Without loss of generality, we may assume the following. \begin{enumerate} \item For $1\le j \le d_1-1$, $C_j$ is a chain of curves; in this case, $A_1$ is an end of the minimal reduced connected cycle containing all nodes of $A$. \item For $1\le j \le d_1-1$, the variable $x_j$ corresponds to the end of $C_j$. \item For $2\le i \le s$, the admissible monomial $m_{id_i}$ belongs to the branch of $A_i$ containing $A_1$. \item For $1\le i \le s$, the admissible forms of $\mathcal F_i$ are given by the matrix as in \remref{r:normal}; we write \begin{align*} f_{ij} &= m_{ij}+a_{ij}m_{id_i-1}+b_{ij}m_{id_i}, & &1\le i \le s, \; 1\le j \le d_i-2, \\ m_{1j} & = x_j^{\alpha_j}, & &1\le j \le d_1-1. \end{align*} \end{enumerate} Now we slightly modify the admissible forms. This modification is needed for the induction step of the proof of the main theorem. Let $\mathbb N$ denote the set of positive integers. A vector $v \in \mathbb N^m$ is said to be {\itshape primitive} if $v$ cannot be written as $v=c v'$ with $v' \in \mathbb N^m$ and $c \in \mathbb N$, $c>1$. Fix an arbitrary vector $\delta=(\delta_1, \ldots ,\delta_m) \in \mathbb N^m$. For each node $A_i$, let $e_i$ denote the positive integer such that $$ \mathbf w_i:=(\adeg i(x_1)\cdot e_i/\delta_1, \ldots , \adeg i(x_m)\cdot e_i/\delta_m) \in \mathbb N^m $$ is primitive. Let $S=\mathbb C\{x_1, \dots ,x_m\}$ be the convergent power series ring. \begin{defn}\label{d:wdeg} Let $\mathbf w =(w_1, \ldots ,w_m)\in \mathbb N^m$. We define the {\itshape $\mathbf w$-degree} of a $\mathbb Q$-monomial $m=\prod _{k=1}^mx_k^{a_k}$ to be $\wdeg{}(m)=\sum a_kw_k$. Let $f=\sum_{k\ge 1} f_k \in S$, where $f_1\neq 0$ and each $f_k$ is a quasihomogeneous polynomial with respect to $\mathbf w$ such that $\wdeg {}(f_k)<\wdeg{}(f_{k+1})$. We call $f_1$ the {\itshape leading form} of $f$, and denote it by $\LF_{\mathbf w}(f)$. Then $f-\LF_{\mathbf w}(f)$ is called the {\itshape higher term} of $f$. We define {\itshape $\mathbf w$-order} of $f$ to be $\mathbf w\text{-}\!\ord (f)=\mathbf w\text{-}\!\deg (\LF_{\mathbf w}(f))$. We set $\mathbf w\text{-}\!\ord (0)=\infty$. \end{defn} We write $\mathbf x_k=x_k^{\delta_k}$ and $\mathbf f=f(\mathbf x_1, \ldots ,\mathbf x_m)$ for $f=f(x_1, \ldots ,x_m) \in S$. We call $\mathbf f$ the {\itshape $\delta$-lifting} of $f$. Any monomial can be thought as the $\delta$-lifting of a $\mathbb Q$-monomial. A polynomial $f$ is quasihomogeneous with respect to $A_i$-weight if and only if so is $\mathbf f$ with respect to the weight $\mathbf w_i$; in fact, for a $\mathbb Q$-monomial $m$, we have \begin{equation}\label{eq:deg} \wdeg i(\mathbf m)=\adeg i(m)\cdot e_i. \end{equation} For each $\mathbf f_{ij}$, we take a convergent power series $f_{ij}^+ \in S$ satisfying $$ \wdeg i(\mathbf f_{ij})<\word i(f_{ij}^+). $$ For each $t \in \mathbb C$, we set $$ \mathcal F_t=\{\mathbf f _{ij}+tf_{ij}^+ | 1 \le i \le s, \; 1\le j \le d_i-2\}. $$ We note that if $m_{ij}=x(D_{ij})$, then for $i \ge 2$ and $1 \le j \le d_i-1$, $$ m_{A_1}(D_{id_i})>m_{A_1}(D_{ij})=m_{A_1}(\du i). $$ If the $\delta$-lifting of a $\mathbb Q$-monomial $x(D)$ is contained in $f_{ij}^+$, then $m_{A_i}(D)>m_{A_i}(\du i)$. By \lemref{l:higher}, $m_{A_1}(D)>m_{A_1}(\du i)$. Thus we obtain the following: \begin{align*} \LF_{\mathbf w_1}(\mathbf f_{1j}+tf_{1j}^+)&=\mathbf f_{1j}, \\ \LF_{\mathbf w_1}(\mathbf f_{ij}+tf_{ij}^+)& =\mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}, \quad \text{for $i \ge 2$}. \end{align*} Therefore the set $$ \LF_{\mathbf w_1}\mathcal F:=\{\LF_{\mathbf w_1}(f)|f \in \mathcal F_t\} $$ is independent of $t \in \mathbb C$; it can be shown that $\LF_{\mathbf w_i}\mathcal F$ is also independent of $t$ for $2 \le i \le s$. \begin{defn}[{Wahl \cite{wahl.es}, cf. \cite[V]{la.simul}}] Let $\omega \: \widetilde X \to T$ be a deformation of a normal surface singularity $\widetilde X_o=\omega^{-1}(o)$, $o \in T$. Suppose that each fiber $\widetilde X_t$ has only one singular point and that there exists a simultaneous resolution $\bar \omega\: \widetilde M \to \widetilde X$ with the exceptional set $\widetilde A$. If the restriction $(\omega \circ \bar \omega)|_{\widetilde A}$ is a locally trivial deformation of the exceptional divisor of $\widetilde M_o$, then we call $\omega \circ \bar \omega$ (resp. $\omega$) an {\itshape equisingular deformation}. $\bar \omega$ is called a {\itshape weak simultaneous resolution} of $\omega$. \end{defn} For a subset $B$ of any commutative ring, let $I(B)$ denote the ideal generated by the elements of $B$. Let $(V_t, o) \subset (\mathbb C^m,o)$ denote the singularity defined by the ideal $I(\mathcal F_t) \subset S$. The main result of this section is the following. \begin{thm}[cf. {\cite[Theorem 2.6]{nw-CIuac}}]\label{t:V} The singularity $(V_t ,o)$ is an isolated complete intersection surface singularity for each $t \in \mathbb C$. Furthermore, the family $\{V_t | t \in \mathbb C\}$ is an equisingular deformation. \end{thm} First we show that every $V_t$ is a complete intersection. \begin{lem}[cf. {\cite[Theorem 3.1]{nw-CIuac}}]\label{l:curveC} For any variable $x_k$, let $A_{i_k}$ denote the node nearest to the end corresponding to $x_k$. Let $C \subset \mathbb C^{m}$ be the affine variety defined by the ideal $I(\LF_{\mathbf w_{i_k}}\mathcal F\cup \{x_k\})$. \begin{enumerate} \item If $x_j=0$ ($j\neq k$) at $p \in C$, then $p$ is the origin. \item $C$ is a complete intersection curve and smooth except for the origin. \end{enumerate} \end{lem} \begin{proof} Without loss of generality, we may assume that $k=1$. Let $$ c_{j}=\begin{cases} b_{11}/a_{11} & \text{if $j=d_1-1$}, \\ b_{1j}-a_{1j}b_{11}/a_{11} & \text{if $2 \le j\le d_1-2$}. \end{cases} $$ Then $c_j\neq 0$ and $C$ is a subvariety of the affine space $\mathbb C^{m-1}$ with coordinates $x_2, \dots, x_m$ defined by the equations \begin{equation}\label{eq:C} \begin{split} \mathbf x_j^{\alpha_j}+c_j\mathbf m_{1d_1}=0, \quad & 2 \le j \le d_1-1, \\ \mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}=0, \quad & 2\le i \le s, 1\le j \le d_i-2. \end{split} \end{equation} If a monomial appearing in \eqref{eq:C} vanishes at $p \in C$, then so does every monomial in the equations at the same node. On the other hand, for each $2 \le i \le m$, some power of $x_i$ appears in \eqref{eq:C} because each end of $A$ is a unique end of a branch of the nearest node. Thus if a variable $x_i$ ($i \ge 2$) vanishes on $C$, then so do all monomials appearing in \eqref{eq:C}, since $A$ is a connected tree of curves. Hence we obtain (1). Since $\#\LF_{\mathbf w_1}\mathcal F=m-2$, it follows from (1) that $\{x_1, x_i\} \cup \LF_{\mathbf w_1}\mathcal F$, where $i\neq 1$, is a regular sequence. Hence $C$ is one-dimensional and complete intersection. The argument above also shows that an ideal $I(\{x_1, x_i-1\} \cup \LF_{\mathbf w_1}\mathcal F)$ defines a nonsingular zero-dimensional variety. Since $C $ is defined by quasihomogeneous polynomials, $C$ is smooth except for the origin. \end{proof} \begin{cor}[cf. {\cite[Corollary 3.4]{nw-CIuac}}]\label{c:CI} $V_t$ is a complete intersection surface singularity, and the support of $V_t \cap \{x_j=x_k=0\}$, $j\neq k$, is the origin. \end{cor} \begin{proof} By \lemref{l:curveC}, $\LF_{\mathbf w_{i_k}}\mathcal F\cup \{x_j,x_k\}$ is a regular sequence. Hence so are $\mathcal F_t$ and $\mathcal F_t\cup\{x_j,x_k\}$. \end{proof} \begin{defn} We define the {\itshape weighted dual graph} of a normal surface singularity to be that of the exceptional set of the minimal good resolution of the singularity. \end{defn} Let $(W,o)$ be a germ of a normal surface singularity and $W' \to W$ the minimal resolution. The {\itshape canonical cycle} on $W'$ is a $\mathbb Q$-cycle which is numerically equivalent to the canonical divisor $K_{W'}$. The self-intersection number of the canonical cycle is an invariant of the singularity and determined by the weighted dual graph; we denote it by $K^2(W)$. Let $\omega \: \widetilde X \to T \subset \mathbb C$ be a deformation of surface singularities. The invariance of $K^2(\widetilde X_t)$ implies the existence of the simultaneous canonical model (or simultaneous RDP resolution) of $\omega$; first the Gorenstein case was proved by Laufer (\cite[Theorem 4.3]{la.simul}), and the general case by Ishii (\cite[Corollary 1.10]{is.simul}). By \cite{bri.simul}, the singularities of the simultaneous canonical model are simultaneously resolved after a suitable finite base change. Thus we obtain the following theorem by the arguments of \cite[VI]{la.simul}. \begin{thm}[Laufer, Ishii]\label{t:la} Let $\omega \: \widetilde X \to T\subset \mathbb C$ be a deformation of a normal surface singularity. If the weighted dual graphs of $\widetilde X_t$, $t\in T$, are the same, then $\omega$ is an equisingular deformation, and it admits a simultaneous resolution such that each fiber is the minimal good resolution. \end{thm} For a divisor $D$, we denote by $D_{red}$ the reduced divisor with $\supp (D_{red})=\supp (D)$. For the induction step of the proof of the main theorem, we need the following. \begin{lem}\label{l:es-div} Let $\omega \: \widetilde X \to \mathbb C$ be an equisingular deformation of a germ of a normal surface singularity $\widetilde X_0=\omega^{-1}(0)$. Let $\widetilde D$ be a reduced divisor on $\widetilde X$, which contains the singular locus of $\widetilde X$. Suppose that $\omega |_{\widetilde D}$ is a locally trivial deformation of a reduced divisor $\widetilde D_0:=\t D|_{\widetilde X_0}$ on $\widetilde X_0$. Then there exists a simultaneous resolution $\bar \omega\: \widetilde M \to \widetilde X$ with the exceptional set $\widetilde A$ such that $(\omega \circ \bar \omega)|_{(\bar \omega^*\widetilde D)_{red}}$ is a locally trivial deformation (hence so is $(\omega \circ \bar \omega)|_{\widetilde A}$). \end{lem} \begin{proof} There exists a weak simultaneous resolution $\bar \omega\:\widetilde M \to \widetilde X$ with the exceptional set $\widetilde A$ such that each $\widetilde A_t:=\widetilde A |_{\widetilde M_t}$ has only simple normal crossing and $(\omega \circ \bar \omega)|_{\widetilde A}$ is a locally trivial deformation. Let $\widetilde F$ denote the strict transform of $\widetilde D$ on $\widetilde M$. Let $\widetilde A^1$ (resp. $\widetilde F^1$) be a divisor on $\widetilde M$, which is the total space of the deformation of an irreducible component of $\widetilde A_0$ (resp. $\widetilde F_0:=\widetilde F |_{ \widetilde M_0}$). Then the intersection number $\widetilde A^1_t\cdot \widetilde F_t^1$ is constant. We may assume $\# (\widetilde A^1_t\cap \widetilde F_t^1) \le 1$. If $\widetilde A^1_t\cdot \widetilde F_t^1\ge 2$, then take the blowing up of $\widetilde M$ along the curve $\widetilde A^1\cap \widetilde F^1$. By taking blowing ups successively in a similar way, we obtain a simultaneous resolution $\bar \omega'\: \widetilde M' \to \widetilde X$ such that each divisor $((\bar \omega'|_{\widetilde M'_t})^*\widetilde D_t)_{red}$ has only normal crossings and that the weighted dual graph of the divisor is independent of $t \in \mathbb C$. \end{proof} \begin{lem}\label{l:G} Let $\omega\: \widetilde X \to \mathbb C$ be as in \lemref{l:es-div}. Suppose that $\widetilde X$ is embedded in an open subset of $\mathbb C^n\times \mathbb C$ such that the singular locus of $\widetilde X$ is $\{o\}\times \mathbb C$, and that $\omega$ is the composite of this embedding and the projection $\mathbb C^n\times \mathbb C \to \mathbb C$. Let $G$ be a finite subgroup of the unitary group $\mathrm U(n)\subset \mathrm {GL}(\mathbb C^n)$. Then $G$ acts on $\mathbb C^n\times \mathbb C$ by $g\cdot(z,t)=(g\cdot z, t)$, $g \in G$. Assume the action induces an action on $\widetilde X$ which is free on $\widetilde X \setminus \{o\}\times \mathbb C$. Then the morphism $\mathcal Omega\: \widetilde X /G \to \mathbb C$ obtained from $\omega$ is an equisingular deformation of $(\widetilde X_t/G,o)$, $t \in \mathbb C$. \end{lem} \begin{proof} Let $\mathbf S_c \subset \mathbb C^n$, $c>0$, denote the $(2n-1)$-sphere of radius $c$. Let $t_0 \in \mathbb C$ be an arbitrary point. Then there exist an open neighborhood $U$ of $t_0$ and a positive number $\epsilon \in \mathbb R$ such that $\Sigma_t:=\mathbf S_{\epsilon}\cap \widetilde X_t \subset \mathbb C^n$ is the link of $\widetilde X_t$ for every $t \in U$ and the family $\{\Sigma_t | t \in U\}$ is topologically trivial. By the assumption, $G$ acts on $\{\Sigma_t | t \in U\}$ freely. Thus we obtain a family $\{\Sigma_t/{G} | t \in U\}$ which is topologically trivial. Recall that the weighted dual graph of a surface singularity is determined by its link (Neumann \cite{neumann.plumbing}). By \thmref{t:la}, we obtain the assertion. \end{proof} We mention the weighted blowing up which is needed in the proof of the theorem. Let $\mathbf w=(w_1, \ldots ,w_m) \in \mathbb N^m$ be a primitive vector, and let $\beta\: Z \to \mathbb C^m$ be the weighted blowing up with respect to the weight $\mathbf w$. It is a projective morphism inducing an isomorphism $Z \setminus \beta^{-1}(o) \to \mathbb C^m \setminus \{o\}$, and $\beta^{-1}(o)=\P(\mathbf w)$, the weighted projective space of type $\mathbf w$. The variety $Z$ is covered by affine varieties $Z_1, \dots ,Z_m$; each $Z_i$ is a quotient of $W_i=\mathbb C^m$ by a cyclic group $\mathcal C_i$ of order $w_i$ determined by the weight $\mathbf w$. Let $\{x_1, \ldots ,x_m\}$ and $\{z_1, \ldots ,z_m\}$ be the coordinates of $\mathbb C^m$ and $W_1$, respectively. The action of the group $\mathcal C_1$ on $W_1$ is given by the diagonal matrix $$ \diag [e(-1/w_1), e(w_2/w_1), \, \ldots \, ,e(w_m/w_1)], $$ where $e(q)=\exp(2\pi \sqrt{-1}q)$. Since $\mathbf w$ is primitive, $\mathcal C_1$ is trivial or the fixed locus is a proper subvariety of the hyperplane $\{x_1=0\}$ which is the exceptional locus. The morphism $W_1 \to \mathbb C^m$, which is the composite of the quotient morphism $W_1 \to Z_1$ and $\beta\:Z_1 \to \mathbb C^m$, is given by $$ x_1=z_1^{w_1}, \quad x_i=z_1^{w_i}z_i \quad (i=2, \ldots ,m). $$ \begin{proof}[Proof of \thmref{t:V}] We prove the theorem by induction on the number of nodes $s$ of $A$. We have to show the isolated singularity of each $V_t$ and the equisingularity of the family $\{V_t | t \in \mathbb C\}$. First assume that $s=1$. Then $V_0$ is the so-called Brieskorn-Pham complete intersection singularity; it is known that $V_0$ has an isolated singularity (it is also easily checked by using the Jacobian criterion). We fix $t \in \mathbb C$. Since $\LF_{\mathbf w_1}\mathcal F$ is a regular sequence, it follows from the theory of filtered rings that there exists an equisingular deformation of $V_0$ with general fiber $V_t$ (cf. \cite[\S 6]{tki-w}, \cite{wahl.defqh}). Hence $V_t$ is an isolated complete intersection singularity, and the weighted dual graphs of $V_0$ and $V_t$ are the same. By \thmref{t:la}, the family $\{V_t | t \in \mathbb C\}$ is an equisingular deformation. Next assume that $s\ge 2$. Let $\beta\: Z\times \mathbb C \to \mathbb C^m\times \mathbb C$ be the trivial family of the weighted blowing up $Z \to \mathbb C^m$ with respect to the weight $\mathbf w_1=(w_1, \ldots ,w_m)$. The family $\{V_t|t \in \mathbb C\}$ is naturally embedded in $\mathbb C^m\times\mathbb C$. Let $W_i=\mathbb C^m$ be as above; however we write $x_i$ instead of $z_i$. Then the cyclic group $\mathcal C_i$ acts on $W_i\times \mathbb C$ as in \lemref{l:G}. Let $V_t^i\subset W_i\times {\{t\}}$ be the strict transform of $V_t$. Recall that $V_t\cap \{x_1=x_2=0\}=\{o\}$ by \corref{c:CI} and that the action of the cyclic group $\mathcal C_i$ on $V_t^i$ is free outside the exceptional locus. Thus to prove that $V_t$ has an isolated singularity at the origin, it suffices to show that any component of singular loci of $V_t^1$ and $V_t^2$, intersecting the exceptional set, is an isolated point. Note that the exceptional divisor is singular at singular points of $V_t^i$. By \thmref{t:la}, \lemref{l:es-div} and \ref{l:G}, it is enough to prove the following three claims (the claims on $V_t^2$ are proved in the same way). \begin{clm}\label{clm:1} The family of exceptional divisors $E_t \subset V_t^1$ is trivial. \end{clm} \begin{clm}\label{clm:2} Let $p \in E_0$ be a singular point and let $p_t \in E_t$ denote the point $p$ under the identification $E_0=E_t$. Then each $(V_t^1, p_t)$ is an isolated singularity and the weighted dual graph of $(V_t^1,x_t)$ is independent of $t$. \end{clm} \begin{clm}\label{clm:3} If $q_1, \ldots ,q_k \in V_0^1$ be the fixed points of $\mathcal C_1$-action, then the fixed locus of the family $\{V_t^1|t \in \mathbb C\}$ is $\bigcup _j\{q_j\}\times \mathbb C$. If $\mathcal C_{1,j} \subset \mathcal C_1$ denotes the isotropy group of $\{q_j\}\times \mathbb C$, then \lemref{l:G} applies to the family $\{(V_t^1,q_j)|t \in \mathbb C\}$ with $\mathcal C_{1,j}$-action for every $q_j$. \end{clm} The exceptional divisor $E_t$ is defined by $x_1=0$ in $V_t^1$. The ideal of $V_t^1 \subset \mathbb C^m= W_1\times {\{t\}}$ is generated by the following functions: \begin{align*} F_{11} & =1+a_{11} \mathbf x_{d_1-1}^{\alpha_{d_1-1}}+b_{11} \mathbf m_{1d_1}+t\bar f_{11}^+, & & \\ F_{1j} & =\mathbf x_j^{\alpha_j}+a_{1j} \mathbf x_{d_1-1}^{\alpha_{d_1-1}}+b_{1j} \mathbf m_{1d_1} +t\bar f_{1j}^+, & & 2 \le j \le d_1-2, \\ F_{ij} & =\mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}+b_{ij}\bar \mathbf m_{id_i}+t\bar f_{ij}^+, & & i \ge 2, \; 1 \le j \le d_i-2, \end{align*} where $$ \bar f_{ij}^+=f_{ij}^+(x_1^{w_1}, x_1^{w_2}x_2, \ldots ,x_1^{w_m}x_m)/x_1^{\word 1(\mathbf f_{ij})}, $$ and $\bar \mathbf m_{id_i}$ is obtained in the same way from $\mathbf m _{id_i}$. Recall the condition on the order of $f_{ij}^+$, and that $$ \adeg 1(m_{ij})<\adeg 1(m_{id_i}), \quad i \ge 2, \; 1 \le j \le d_i-2. $$ Then we see that the ideal of $E_t \subset \mathbb C^m$ is generated by $x_1$ and the following polynomials: \begin{align} &1+a_{11} \mathbf x_{d_1-1}^{\alpha_{d_1-1}}+b_{11} \mathbf m_{1d_1},\label{eq:tF1} & & \\ &\mathbf x_j^{\alpha_j}+a_{1j} \mathbf x_{d_1-1}^{\alpha_{d_1-1}}+b_{1j} \mathbf m_{1d_1}, & & 2 \le j \le d_1-2, \label{eq:tFj}\\ &\mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}, & & i\ge 2, \; 1 \le j \le d_i-2. \label{eq:gij} \end{align} These polynomials do not contain $t$; thus \clmref{clm:1} is verified. Since the fixed points of the $\mathcal C_j$-action lie on $E_t$, the first assertion of \clmref{clm:3} follows. By looking at the action explicitly, we see that the $\mathcal C_{1,j}$-action on $W_1\times \mathbb C$ is unitary around $q_j$. Thus \clmref{clm:3} follows from \clmref{clm:2}. It is easy to see that there are $m-d_1$ polynomials in \eqref{eq:gij} and any of them contains no variables $x_1, \ldots ,x_{d_1-1}$. As in the proof of \lemref{l:curveC}, we can show that the functions of \eqref{eq:gij} define a complete intersection curve $C' \subset \mathbb C^{m-d_1+1}$ which is smooth except for the origin. Furthermore \eqref{eq:tF1} and \eqref{eq:tFj} define a tower of cyclic coverings over $C'$. Hence the singularities of $E_t$ are lying above the point $(0,\ldots ,0)$ of $C'$. Therefore at each singular point of $E_t$, the only $d_1-2$ variables $x_2,\ldots ,x_{d_1-1}$ are nonzero, and others are zero. We fix $t \in \mathbb C$. Let $p$ be a singular point of $E_t$. At $p$, the Jacobian matrix $$ \partial (F_{11}, \ldots ,F_{1d_1-2})/\partial(x_2, \ldots ,x_{d_1-1}) $$ is regular, and thus $x_2, \ldots ,x_{d_1-1}$ can be expressed by a convergent power series with nonzero constant terms in $x_1, x_{d_1}, \ldots ,x_m$. By substituting them into $F_{ij}$, $i\ge 2$, we obtain new defining functions for the germ $(V_t^1, p)$ in $(\mathbb C^{m-d_1+2}, o)$ as follows: \begin{equation}\label{eq:neweq} F'_{ij}:=\mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}+b_{ij}'\bar \mathbf m_{id_i}'+r_{ij}, \quad i\ge 2, \quad 1 \le j \le d_i-2, \end{equation} where $\bar \mathbf m_{id_i}'$ is a monomial obtained by substituting $1$ into $x_2, \ldots, x_{d_1-1}$ of $\bar \mathbf m_{id_i}$. We see that \begin{equation}\label{eq:neq} a_{ij_1}b'_{ij_2}\not=a_{ij_2}b'_{ij_1} \; (j_1\neq j_2) \quad \text{and} \quad b'_{ij} \neq 0. \end{equation} By \eqref{eq:deg}, for every $\mathbb Q$-monomial $m=x(D)$, we have $$ \wdeg 1(\mathbf m)=e_1\cdot \adeg 1(m)=e_1\cdot m_{A_1}(D). $$ Suppose that $m_{id_i}=x(D_i)$ for $i\ge 2$. Since $\adeg 1(m_{id_i-1})=m_{A_1}(\du i)$, the exponent of $x_1$ in $\bar \mathbf m_{id_i}'$ is \begin{equation}\label{eq:exponent} \wdeg 1(\mathbf m_{id_i})- \wdeg 1(\mathbf m_{id_i-1})= e_1m_{A_1}(D_i-\du i)>0. \end{equation} Since $D_i-\du i \in A_{\mathbb Z}$, it follows that $\bar \mathbf m_{id_i}'$ is a monomial of $x_1^{e_1}, \mathbf x_{d_1}, \ldots ,\mathbf x_m$. We denote by $h_{i}$ the monomial obtained by replacing $x_1^{e_1},\mathbf x_{d_1}, \ldots ,\mathbf x_m$ of $\bar \mathbf m_{id_i}'$ with $x_1, x_{d_1}, \cdots ,x_m$, respectively. Let $\delta'=(e_1,\delta_{d_1},\ldots ,\delta_m) \in \mathbb N^{m-d_1+2}$ and let $$ f'_{ij}=m_{ij}+a_{ij}m_{id_i-1}+b_{ij}'h_{i}, \quad i\ge 2, \quad 1 \le j \le d_i-2. $$ Then the polynomial $\mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}+b_{ij}'\bar \mathbf m_{id_i}'$ is the $\delta'$-lifting of $f'_{ij}$. Let $A_{e}$ be the component of the branch $C_{d_1}$ of the node $A_1$, which intersects $A_1$; see \figref{f:diag}. \begin{figure}[htbp] \begin{center} \setlength{\unitlength}{0.5cm} \begin{picture}(14,3)(-7,-2) \put(0,-1.2){\framebox(6,2){}} \put(-2,0){\line(1,0){4}} \put(-2,0){\circle*{0.25}} \put(-2.2,-1){$A_1$} \put(0.8,-1){$A_e$} \put(2.3,0){ . . .} \put(1,0){\circle*{0.25}} \put(6.5,-0.5){$C_{d_1}$} \put(-4,-1){\line(2,1){2}} \put(-6,-2){\framebox(2,1){}} \put(-8,-2){$C_{d_1-1}$} \put(-4,1){\line(2,-1){2}} \put(-6,1){\framebox(2,1){}} \put(-7.5,1){$C_1$} \put(-5,-0.5){$\vdots$} \end{picture} \caption{\label{f:diag}} \end{center} \end{figure} We may assume that the $A_e$ is an end of $C_{d_1}$; take the blowing up at $A_1 \cap A_e$ if necessary. Let $A'=C_{d_1}$. Then $A'\subset M$ can be blown down to a normal surface singularity. We consider admissible forms concerning $A'$. We associate the end $A_e$ with the variable $x_1$, while keeping the correspondence between the other ends and the variables $x_{d_1}, \ldots, x_m$. \begin{clm}\label{clm:smallgraph} We have the following. \begin{enumerate} \item $A'$ has $s-1$ nodes $A_2, \cdots, A_s$. \item $A'$ satisfies \condref{c:A}, and for each $2\le i \le s$, the set of polynomials $$ \{f'_{i1}, \ldots ,f'_{i d_i-2}\} $$ is a Neumann-Wahl system concerning $A'$ at a node $A_i$. \item In this new situation, we define a weight $\mathbf w_i' \in \mathbb N^{m-d_1+2}$ ($2\le i \le s$) with respect to $\delta'$ as in the sentence before \defref{d:wdeg}. Then $$ \wpdeg i(\mathbf m_{ij}+a_{ij}\mathbf m_{id_i-1}+b_{ij}'\bar \mathbf m_{id_i}') <\wpord i(r_{ij}). $$ \end{enumerate} \end{clm} The inductive hypothesis and \clmref{clm:smallgraph} imply that $(V_t^1,p)$ is an isolated singularity and is an equisingular deformation of the singularity defined by the $\delta'$-lifting of the Neumann-Wahl system $\{f'_{ij}\}$; thus \clmref{clm:2} follows. Now we prove \clmref{clm:smallgraph}. First, the assertion (1) is obvious. It is clear that the number of branches of a node $A_i$ ($i\ge 2$) in $A'$ is the same as that of $A_i$ in $A$. For $D=\sum a_jA_j \in A_{\mathbb Q}$, we write $\res(D)= \sum _{A_j\le A'}a_jA_j$. Let $\t A_k$ denote the dual cycle of $A_k$ on $A'$. Let $D_{ij}$ ($i \ge 2$) be the monomial cycle such that $x(D_{ij})=m_{ij}$, and let $$ E_{ij}=\res(D_{ij}-\du i). $$ Since $(D_{ij}-\du i)\cdot A_e=0$, we have $-E_{ij}\cdot A_e=m_{A_1}(D_{ij}-\du i)$; this is the exponent of $x_1$ in $h_{i}$ (see \eqref{eq:exponent}). By computing the intersection numbers $(E_{ij}+\t A_i)\cdot A_l$ for every $A_l \le A'$, we see that $E_{ij}+\t A_i$ is a monomial cycle belonging to a branch of $A_i$ and that $$ x(E_{id_i}+\t A_i)=h_{i} \quad \text{and} \quad x(E_{ij}+\t A_i)=m_{ij} \quad \text{for $1\le j <d_i$}. $$ By \eqref{eq:neq}, we have (2) of \clmref{clm:smallgraph}. We can generalize the argument above as follows. Let $m$ be any monomial in $\mathbf f_{ij}+tf_{ij}^+$. Suppose that $m$ is the $\delta$-lifting of a $\mathbb Q$-monomial $x(D)$ associated with a $\mathbb Q$-monomial cycle $D \in A_{\mathbb Q}$. Let $m'$ be the $\delta'$-lifting of the monomial $x(\res(D-\du i)+\t A_{i})$. Then, by the operation which changes $\mathbf f_{ij}+tf_{ij}^+$ into the function $F'_{ij}$, the monomial $m$ is changed into a function of the form $u m'$, where $u \in S$ is a unit. We have that $m_{A_i}(\res(D-\du i)+\t A_{i})=m_{A_i}(D)$ for a node $A_i$ $(i \ge 2)$. Now it is easy to see that (3) holds. Thus we have proved \clmref{clm:smallgraph}. \end{proof} Let $\hat S=\mathbb C[[x_1, \ldots ,x_m]]$ denote the formal power series ring. By a similar argument as in the proof of \thmref{t:V}, we obtain the following. \begin{thm}\label{t:formal} Let $\mathbf f_{ij}$ be as above. Take a formal power series $g_{ij} \in \hat S$ such that $$ \wdeg i(\mathbf f_{ij})<\word i(g_{ij})\quad \text{ for} \quad 1 \le i\le s\;, 1 \le j \le d_i-2. $$ Let $J \subset \hat S$ be the ideal generated by all $\mathbf f_{ij}+g_{ij}$'s. Then $\hat S/J$ is a two-dimensional complete intersection ring, and has only a singularity at the maximal ideal. \end{thm} \section{The main results}\label{s:mainresults} In this section we will prove the following. \begin{thm}\label{t:main} If $(X,o)$ is a rational or minimally elliptic singularity, then its universal abelian cover $(Y,o)$ is an equisingular deformation of a Neumann-Wahl complete intersection singularity. The deformation is defined by the functions of the form $f+tf^+$, where $\{f\}$ is a Neumann-Wahl system associated with the exceptional set $A$, $f^+$ is a function with order greater than the degree of $f$ and $t$ is the parameter. \end{thm} We start without the assumption of the theorem. We use the notation of \sref{s:pre}. Write $\mathcal A_b=H^0(-L^{(b)})$ and $\mathcal O_{Y,o}=\bigoplus _{b \in \mathcal B}\mathcal A_b$. Recall that for any component $A_i$ of $A$, there exist a divisor $L^i$ and an element $b^i \in \mathcal B$ such that $\nu (L^i)=\du i$ and $L^i - L^{(b^i)} \in A_{\mathbb Z}$. We consider the following condition, which depends on the resolution $\pi\: M \to X$. \begin{cond}\label{c:B} For each end $A_i \in \mathcal E(A)$, there exists a section $y_i \in H^0(-L^i)$ such that $(y_i)_A=\nu(L^i)$. \end{cond} \begin{lem}\label{l:satisfyB} If $(X,o)$ is rational, or if $(X,o)$ is minimally elliptic and the minimally elliptic cycle $E$ is supported on $A$, then \condref{c:B} is satisfied. \end{lem} \begin{proof} Let $D$ be any $\pi$-nef divisor on $M$. If $(X,o)$ is rational, then $D$ is $\pi$-free; see \cite[4.17]{chap}. Assume that $(X,o)$ is minimally elliptic and the minimally elliptic cycle $E$ is supported on $A$. Since $E$ is 2-connected, $H^0(D)$ has no fixed component on $A$ if $D\cdot A\ge 1$ by \cite[4.23, Remark]{chap}. Thus the assertion follows. \end{proof} Let $\frak m_Y$ (resp. $\frak m_X$) denote the maximal ideal of $\mathcal O_{Y,o}$ (resp. $\mathcal O_{X,o}$). We may identify $\frak m_Y$ as $\frak m_X \bigoplus (\bigoplus _{b \neq 0}\mathcal A_b)$ (cf. \cite[\S 6]{o.uac-rat}). \begin{lem}\label{l:powerofmax} Let $\{Z_k \in A_{\mathbb Z}|k \in \mathbb N\}$ be a sequence of cycles such that $Z_{k+1}>Z_k>0$ for every $k \in \mathbb N$. Then there exists a function $\alpha\:\mathbb N \to \mathbb N$ such that for each $b \in \mathcal B$, $$ H^0(-L^{(b)}-Z_k)\subset \frak m_Y^{\alpha(k)} \quad \text{and} \quad \lim_{k \to \infty} \alpha (k)= \infty. $$ \end{lem} \begin{proof} We only give an outline. We can take a positive integer $a$ so that for any $\pi$-nef divisor $D$ on $M$ and a cycle $Z:=a \sum_{A_i \le A} \du i$, the natural map $$ H^0(D-Z)\otimes H^0(-Z) \to H^0(D-2Z) $$ is surjective (cf. \cite[III]{la.simul}). Let $\beta$ be a nonnegative integer such that $-L^{(b)}-\beta Z$ is $\pi$-nef for every $b \in \mathcal B$. Then we obtain $H^0(-L^{(b)}-(\beta+k)Z)\subset \frak m_Y^k$. We may assume $Z_1>(\beta+1)Z$. Now define $\alpha(l)= \max\{k \in \mathbb N |(\beta+k)Z \le Z_l\}$. \end{proof} \begin{ass} From now on, we assume that \condref{c:C} and \ref{c:B} are satisfied. \end{ass} If $A$ is a chain of curves, then $(X,o)$ is a cyclic quotient singularity and $\mathcal O_{Y,o}=\mathbb C\{y_1,y_2\}$, where $y_i$'s are as in \condref{c:B} (if $A$ is irreducible, then $y_1,y_2 \in H^0(-L^1)$). Let $m=\#\mathcal E(A)$. Assume that $m\ge 3$. Then we can define admissible monomials at each node by associating each end with a variable as in \sref{s:NWS}. We define the homomorphism $$ \psi\: S=\mathbb C\{x_1, \ldots ,x_m\} \to \mathcal A=\mathcal O_{Y,o} $$ of $\mathbb C$-algebras by $\psi (x_i)=y_i$. We denote by $\hat{}$ the maximal-ideal-adic completions of local rings. Let $\hat \psi\:\hat S=\mathbb C[[x_1, \ldots ,x_m]] \to \hat {\mathcal O}_{Y,o}$ be the induced homomorphism. By the definition of the set $\mathcal B$, we may regard $\mathcal B$ as the discriminant group $G:=\du {\mathbb Z}/A_{\mathbb Z}$ in the natural way. Let $S_b \subset S$ (resp. $\hat S_b \subset \hat S$), $b \in \mathcal B$, denote the set of power series represented as the sum of monomials $x(D)$ satisfying $D\pmod{A_{\mathbb Z}}=b$. Then we have $S=\bigoplus _{b \in \mathcal B}S_b$ and the $\psi$ becomes a homomorphism of $\mathcal B$-graded (or $G$-graded) algebras. The same holds for $\hat S$ and $\hat{\psi}$. Let $A_1, \ldots ,A_s$ be all of the nodes of $A$, and $d_i$ the number of branches of a node $A_i$. Let $\mathcal M_i=\{m_{i1}, \ldots ,m_{id_i}\}$ denote a complete system of admissible monomials at a node $A_i$. Let $\mathbb C\mathcal M_i \subset S$ denote the $\mathbb C$-linear subspace spanned by the monomials of $\mathcal M_i$. \begin{lem}\label{l:constructCSAF} For any node $A_i$, let $\mu_i$ denote the composite of homomorphisms $$ \mathbb C\mathcal M_{i} \overset{\psi}\longrightarrow H^0( \mathcal O_M(-L^{i})) \to H^0(\mathcal O_{A_{i}}(-L^{i})). $$ Then $\mu_i$ is surjective. We have $h^0(\mathcal O_{A_{i}}(-L^{i}))=2$ and $\dim \Ker \mu_i=d_i-2$. Let $\mathcal F_i=\{f_{i1}, \ldots ,f_{id_i-2}\}$ be a basis of $\Ker \mu_i$. Then $\mathcal F_i$ is a Neumann-Wahl system at the node $A_i$. \end{lem} \begin{proof} Since $L^i\cdot A_i=-1$, we have $h^0(\mathcal O_{A_{i}}(-L^{i}))=2$. Suppose $m_{ij}=x(D_{ij})$ for a monomial cycle $D_{ij}$ belonging to a branch $C_{ij}$ of $A_i$. Since $(D_{ij}-\du {i})\cdot A_{i}=1$, it follows from \lemref{l:product} that $\mu_i(x(D_{ij}))$ has a zero of order one at $A_{i} \cap C_{ij}$. Thus $\mu_i(x(D_{id_i-1}))$ and $\mu_i(x(D_{id_i}))$ generate $H^0(\mathcal O_{A_{i}}(-L^{i}))$. Therefore we obtain a complete system of admissible forms expressed by a $((p-2) \times p)$-matrix as in \remref{r:normal}, which is a basis of $\Ker \mu_i$. \end{proof} Let us recall that polynomials of $\mathcal F_i$ are quasihomogeneous with respect to the $A_i$-weight. \begin{lem}\label{l:higherterms} Let $A_i$ be a node and $h\in H^0(-L^i)$. Then there exists $\bar h \in \hat S_{b^i}$ such that $\hat \psi(\bar h)=h$ in $\hat {\mathcal O}_{Y,o}$. Suppose that $h=\psi (\bar h_0)$ for an admissible form $\bar h_0 \in \Ker \mu _i$. Then we can take the $\bar h$ so that $\adeg i (\bar h_0)<\aord i(\bar h)$. If in addition $\psi $ is surjective, such $\bar h$ can be taken from $S_{b^i}$. \end{lem} \begin{proof} We use the notation of \lemref{l:constructCSAF}. Suppose that $h \neq 0$. Let $F_0$ be the divisor such that $\nu (F_0)=(h)_A$ and $F_0-L^i \in A_{\mathbb Z}$. Let $c_0=-F_0\cdot A_i$. Then $c_0\ge 0$. By \condref{c:C} and the proof of \lemref{l:existmonomials}, there exists a cycle $F'_0\ge F_0$ such that $m_{A_i}(F'_0-F_0)=0$, $(F'_0-F_0)\cdot A_i=0$ and $F'_0\cdot A_j=0$ if $j\neq i$ and $A_j$ is not an end. Let $D'_{ij}=D_{ij}-\du i$. Then an arbitrary cycle of the form $E_{\mathbf a}:=F'_0+\sum a_{j}D'_{ij}$ is a monomial cycle, where $\mathbf a=(a_1, \ldots ,a_m) \in \mathbb Z^m$ with $\sum a_j=c_0$ and $a_j \ge 0$. It is easy to see that the $\mu_i(x(E_{\mathbf a}))$'s span $H^0(\mathcal O_{A_i}(-F_0))$. Thus we have a quasihomogeneous polynomial $\bar h_1$, which is a linear form of $x(E_{\mathbf a})$'s, with respect to $A_i$-weight such that $h-\psi(\bar h_1) \in H^0(-F_0-A_i)$. Let $F_1$ be the divisor such that $\nu (F_1)=(h-\psi(\bar h_1))_A$ and $F_1-L^i \in A_{\mathbb Z}$. Then it follows from the argument above that there exists a quasihomogeneous polynomial $\bar h_2$ such that $$ h-\psi(\bar h_1)-\psi(\bar h_2) \in H^0(-F_1-A_i). $$ Thus we obtain a sequence $\{\bar h_k|k \in \mathbb N\}$ of quasihomogeneous polynomials and a sequence $\{F_k | k \in \mathbb N\}$ of divisors satisfying the following: for all $k \in \mathbb N$, \begin{enumerate} \item $h-\psi(\bar h_1+\cdots +\bar h_k) \in H^0(-F_k)$, \item $F_{k+1}>F_k$, \item $\adeg i(\bar h_k)<\adeg i(\bar h_{k+1})$. \end{enumerate} By \lemref{l:powerofmax} there exists a function $\alpha\: \mathbb N \to \mathbb N$ such that $H^0(-F_k) \subset \frak m_Y^{\alpha(k)}$ and that $\lim_{k\to \infty}\alpha(k) = \infty$. Now put $\bar h=\sum \bar h_i \in \hat S_{b^i}$. Then $\hat \psi(\bar h)=h$. Suppose that $h=\psi(\bar h_0)$ with an admissible form $\bar h_0 \in \Ker \mu_i$. Since $\psi(\bar{h_0}) \in H^0(-L^i-A_i)$, we have $$ \adeg i(\bar h_0)<m_{A_i}(F_0)=\adeg i(\bar h_{1}). $$ If $\psi$ is surjective, then the maps $(x_1, \cdots ,x_m)^k \to \frak m_Y^k$ and $S_{b} \to \mathcal A_b$ are surjective for every $k \in \mathbb N$ and $b \in \mathcal B$. Therefore, for sufficiently large $k$, there exists $\bar h' \in S_{b^i}$ such that $\aord i(\bar h')>\adeg i(\bar h_0)$ and $h=\psi(\bar h_1+\cdots +\bar h_k+\bar h')$. \end{proof} \begin{prop}\label{p:surj} The homomorphism $\psi\: S \to \mathcal O_{Y,o}$ is surjective. \end{prop} \begin{proof} We fix a node $A_i$. By \cite[Proposition 5.1]{nw-CIuac}, the group $G=\du {\mathbb Z}/A_{\mathbb Z}$ is generated by $\{\du j|A_j \in \mathcal E(A)\}$. Hence for each $b \in \mathcal B$, there exists a monomial $m_b \in S$ such that $\mathcal A_b\cdot m_b \subset H^0(-L^i)$. By \lemref{l:higherterms}, we have $\mathcal A_b \subset \hat\psi(\hat S_b)\cdot m_b^{-1}$. Therefore $$ \hat \psi (\hat S) \subset \hat {\mathcal O}_{Y,o} \subset \sum_{b \in \mathcal B} \hat \psi (\hat S) \cdot m_b^{-1}. $$ Then it follows that $\hat S /\Ker \hat \psi$ is a two-dimensional domain. By \lemref{l:higherterms}, for any $f_{kl} \in \bigcup \mathcal F_j$, there exists $\tilde f_{kl} \in \hat S$ such that $\LF_{A_k}(\tilde f_{kl})=f_{kl}$ and $\tilde f_{kl} \in \Ker \hat \psi$. Let $\tilde I \subset \hat S$ denote the ideal generated by all $\tilde f_{kl}$'s. Then it follows from \thmref{t:formal} that $\hat S/\tilde I$ is a two-dimensional normal domain. Since $\tilde I \subset \Ker \hat{\psi}$, we obtain that $\hat S/\tilde I\cong \hat \psi (\hat S)$. Since $\hat S$ is Noetherian, $ \hat {\mathcal O}_{Y,o} $ is finitely generated $\hat \psi(\hat S)$-module. By the normality of $\hat \psi (\hat S)$, we obtain $\hat \psi (\hat S)=\hat {\mathcal O}_{Y,o}$. This implies that $\psi$ is surjective. \end{proof} It follows from \lemref{l:higherterms} and \proref{p:surj} that for each $f_{ij} \in \mathcal F_i$, there exists $f_{ij}^+ \in S_{b^i}$ such that $f_{ij}+f_{ij}^+ \in \Ker \psi$ and $\adeg i( f_{ij})<\aord i(f_{ij}^+)$. For each $t \in \mathbb C$, we denote by $I_t \subset S$ the ideal generated by the functions $$ f_{ij}+tf_{ij}^+, \quad 1\le i \le s, 1 \le j \le d_{i}-2. $$ As in the proof of \proref{p:surj}, by \thmref{t:V}, we see that $\Ker \psi=I_1$. \begin{cor} $\mathcal O_{Y,o} \cong S/I_1$. \end{cor} Again by \thmref{t:V}, we obtain the following. \begin{thm}\label{t:main-general} If \condref{c:C} and \ref{c:B} are satisfied, then the universal abelian cover $(Y,o)$ of $(X,o)$ is an equisingular deformation of a Neumann-Wahl complete intersection singularity. The deformation is defined by the functions $$ f_{ij}+T f_{ij}^+ \in S[T], \quad 1\le i \le s, 1 \le j \le d_{i}-2, $$ where $S[T]$ is the polynomial ring over $S$. \end{thm} Now \thmref{t:main} follows from \thmref{t:main-general}, \lemref{l:existmonomials} and \ref{l:satisfyB}. \begin{cor}\label{c:QGor} If \condref{c:C} and \ref{c:B} are satisfied, then $(X,o)$ is $\mathbb Q$-Gorenstein. If in addition the link of $(X,o)$ is a homology sphere, then $(X,o)$ is an equisingular deformation of a Neumann-Wahl complete intersection singularity. \end{cor} \begin{rem} There is ambiguity in the choice of the complete systems of admissible monomials $\mathcal M_i$. The proof of \lemref{l:higherterms} shows that if $m$ and $m'$ belong to the same branch of a node $A_i$, then there exists $h \in S_{b^i}$ such that $x(m)-x(m')-h \in \Ker \psi$ and $\aord i(h)>\adeg i(x(m))$. Therefore, whether $(Y,o)$ is a Neumann-Wahl complete intersection depends not only on the choice of the sections $\{y_i\}$, but also on the choice of the complete systems of admissible monomials. \end{rem} In the rest of this section, we describe the action of the Galois group of the universal abelian covering. Recall that the Galois group $H_1(\Sigma, \mathbb Z)$ of the universal abelian covering $(Y,o) \to (X,o)$ is isomorphic to the discriminant group $G=\du {\mathbb Z}/A_{\mathbb Z}$. For any $\mathbb Q$-cycle $D=\sum a_i\du i=\sum b_iA_i$, we have $D\cdot A_j=-a_j$ and $D\cdot \du j=-b_j$. Thus we obtain the following \begin{lem}\label{l:du} Let $D \in A_{\mathbb Q}$. Then \begin{enumerate} \item $D \in \du {\mathbb Z}$ if and only if $\{D\cdot A_i|A_i \le A\} \subset \mathbb Z$, \item $D \in A_{\mathbb Z}$ if and only if $\{D\cdot \du i|A_i \le A\} \subset \mathbb Z$. \end{enumerate} \end{lem} For any $D \in \du {\mathbb Z}$, we denote by $(D)$ the class of $D$ modulo $A_{\mathbb Z}$. Then the action of $G$ on $S=\mathbb C\{x_1,\ldots ,x_m\}$ is expressed as follows (cf. \cite[\S 5]{nw-CIuac}). For $(D) \in G$ and a monomial cycle $F$, \begin{equation}\label{eq:action} (D)\cdot x(F):=\exp (2\pi\sqrt{-1}D\cdot F)x (F). \end{equation} By \lemref{l:du} (1), this is well-defined. Let $(Y_t,o)$ denote the singularity defined by the ideal $I_t \subset S$. We have seen that $\{Y_t|t \in \mathbb C\}$ is an equisingular family. Since $I_t$ is generated by homogeneous elements of $G$-graded algebra $S$, the group $G$ naturally acts on $S/ I_t =\mathcal O_{Y_t,o}$. It follows from \lemref{l:du} (2) that $\mathcal O_{X,o}=(\mathcal O _{Y,o})^{G}$. By \corref{c:CI} and \cite[Proposition 5.2]{nw-CIuac}, the action is free on $Y_t\setminus \{o\}$ (see also \cite[Theorem 7.2 (2)]{nw-CIuac}). Therefore $X_t:=Y_t/G$ is a normal singularity. The linear action of $G$ on $\mathbb C^m$ determined by \eqref{eq:action} is unitary. By \lemref{l:G} and the uniqueness of the universal abelian covering, we obtain the following \begin{thm}\label{p:esquotient} The family $\{X_t|t \in \mathbb C\}$ is an equisingular deformation, and each $Y_t \to X_t$ is the universal abelian covering. \end{thm}
{ "timestamp": "2006-03-16T10:51:12", "yymm": "0503", "arxiv_id": "math/0503733", "language": "en", "url": "https://arxiv.org/abs/math/0503733" }
\section{Introduction} The theoretical framework of self-consistent mean-field models enables a description of the nuclear many-body problem in terms of universal energy density functionals. By employing global effective interactions, adjusted to empirical properties of symmetric and asymmetric nuclear matter, and to bulk properties of spherical nuclei, the current generation of self-consistent mean-field models has achieved a high level of accuracy in the description of ground states and properties of excited states in arbitrarily heavy nuclei, exotic nuclei far from $\beta$-stability, and in nuclear systems at the nucleon drip-lines \cite{BHR.03}. The relativistic mean-field (RMF) models, in particular, are based on concepts of non-renormalizable effective relativistic field theories and density functional theory. They have been very successfully applied in studies of nuclear structure phenomena at and far from the valley of $\beta$-stability. For a quantitative analysis of open-shell nuclei it is necessary to consider also pairing correlations. Pairing has often been taken into account in a very phenomenological way in the BCS model with the monopole pairing force, adjusted to the experimental odd-even mass differences. This approach, however, presents only a poor approximation for nuclei far from stability. The physics of weakly-bound nuclei necessitates a unified and self-consistent treatment of mean-field and pairing correlations. This has led to the formulation and development of the relativistic Hartree-Bogoliubov (RHB) model, which represents a relativistic extension of the conventional Hartree-Fock-Bogoliubov framework, and provides a basis for a consistent microscopic description of ground-state properties of medium-heavy and heavy nuclei, low-energy excited states, small-amplitude vibrations, and reliable extrapolations toward the drip lines \cite{VALR.05}. In most applications of the RHB model \cite{VALR.05} the pairing part of the well known and very successful Gogny force~\cite{BGG.84} has be employed in the particle-particle ($pp$) channel: \begin{equation} V^{pp}(1,2)~=~\sum_{i=1,2}e^{-((\mathbf{r}_{1}-\mathbf{r}_{2})/{\mu_{i}})^{2}% }\,(W _{i}~+~B_{i}P^{\sigma}-H_{i}P^{\tau}-M_{i}P^{\sigma}P^{\tau})\;, \end{equation} with the set D1S \cite{BGG.91} for the parameters $\mu_{i}$, $W_{i}$, $B_{i}$, $H_{i}$, and $M_{i}$ $(i=1,2)$. This force has been very carefully adjusted to the pairing properties of finite nuclei all over the periodic table. In particular, the basic advantage of the Gogny force is the finite range, which automatically guarantees a proper cut-off in momentum space. However, the resulting pairing field is non-local and the solution of the corresponding Dirac-Hartree-Bogoliubov integro-differential equations can be time-consuming, especially in the case of deformed nuclei. Another possibility is the use of a zero-range, possibly density-dependent, $\delta$-force in the $pp$-channel of the RHB model \cite{Meng.98}. This choice, however, introduces an additional cut-off parameter in energy and neither this parameter, nor the strength of the interaction, can be determined in a unique way. The effective range of the interaction is determined by the energy cut-off, and the strength parameter must be chosen accordingly in order to reproduce empirical pairing gaps. In a series of recent papers \cite{Bul1,Bul2,Bul3} A. Bulgac and Y. Yu have introduced a simple scheme for the renormalization of the Hartree-Fock-Bogoliubov equations in the case of zero-range pairing interaction. The scheme is equivalent to a simple energy cut-off with a position dependent coupling constant. In this work we use the prescription of Refs.~\cite{Bul1,Bul2} to implement a regularization scheme for the relativistic Hartree-Bogoliubov equations with zero-range pairing. We analyze the resulting $^1S_0$ pairing gap in isospin-symmetric nuclear matter and apply the RHB model to the calculation of ground-state pairing properties of finite spherical nuclei. In Sec.~\ref{SecRHB} we present an outline of the RHB model and introduce the renormalization scheme for the case of zero-range pairing. The model is applied in Sec.~\ref{SecNM} to pairing in isospin-symmetric nuclear matter. Ground-state pairing properties of Sn nuclei are analyzed in Sec.~\ref{SecSn}. Sec.~\ref{SecSum} contains the summary and conclusions. \section{\label{SecRHB} Relativistic Hartree-Bogoliubov model with zero-range pairing} A detailed review of the relativistic Hartree-Bogoliubov model can be found, for instance, in Ref.~\cite{VALR.05}. In this section we include those features which are essential for the discussion of the renormalization of the RHB equations. The model can be derived within the framework of covariant density functional theory. When pairing correlations are included, the energy functional depends not only on the density matrix $\hat{\rho}$ and the meson fields $\phi_{m}$, but in addition also on the anomalous density $\hat{\kappa}$ \begin{equation} E_{RHB}[\hat{\rho},\hat{\kappa},\phi_{m}]=E_{RMF}[\hat{\rho},\phi _{m}]+E_{pair}[\hat{\kappa}]\;,\label{ERHB}% \end{equation} where $E_{RMF}[\hat{\rho},\phi]$ is the RMF energy density functional and the pairing energy $E_{pair}[\hat{\kappa}]$ is given by \begin{equation} E_{pair}[\hat{\kappa}]=\frac{1}{4}\mathrm{Tr}\left[ \hat{\kappa}^{\ast}% V^{pp}\hat{\kappa}\right] . \end{equation} $V^{pp}$ denotes a general two-body pairing interaction. The equation of motion for the generalized density matrix \begin{equation} \mathcal{R}=\left( \begin{array} [c]{cc}% \rho & \kappa\\ -\kappa^{\ast} & 1-\rho^{\ast} \end{array} \right) \;, \end{equation} reads \begin{equation} i\partial_{t}\mathcal{R}=\left[ \mathcal{H}(\mathcal{R}),\mathcal{R}\right] \;.\label{TDRHB} \end{equation} The generalized Hamiltonian $\mathcal{H}$ is a functional derivative of the energy with respect to the generalized density \begin{equation} \mathcal{H}_{RHB}~=~\frac{\delta E_{RHB}}{\delta\mathcal{R}}~=~\left( \begin{array} [c]{cc}% \hat{h}_{D}-m-\mu & \hat{\Delta}\\ -\hat{\Delta}^{\ast} & -\hat{h}^{\ast}_{D}+m+\mu \end{array} \right) \;.\label{HFB-hamiltonian}% \end{equation} The self-consistent mean field $\hat{h}_{D}$ is the Dirac Hamiltonian, and the pairing field reads \begin{equation} \Delta_{ab}(\mathbf{r},\mathbf{r}^{\prime})={\frac{1}{2}}\sum\limits_{c,d}% V_{abcd}^{pp}(\mathbf{r},\mathbf{r}^{\prime})\kappa_{cd}(\mathbf{r}% ,\mathbf{r}^{\prime}),\label{equ.2.5}% \end{equation} where $a,b,c,d$ denote quantum numbers that specify the Dirac indices of the spinors, and $V_{abcd}^{pp}(\mathbf{r},\mathbf{r}^{\prime})$ are the matrix elements of a general two-body pairing interaction. Pairing effects in nuclei are restricted to an energy window of a few MeV around the Fermi level, and their scale is well separated from the scale of binding energies, which are in the range of several hundred to thousand MeV. There is no experimental evidence for any relativistic effect in the nuclear pairing field $\hat{\Delta}$. Therefore, pairing can be treated as a non-relativistic phenomenon, and a hybrid RHB model with a non-relativistic pairing interaction can be employed. For a general two-body interaction, the matrix elements of the relativistic pairing field read \begin{equation} \hat{\Delta}_{a_1 p_1, a_2 p_2} = {\frac{1}{2}}\sum\limits_{a_3 p_3, a_4 p_4} \langle a_1 p_1, a_2 p_2 |V^{pp}|a_3 p_3, a_4 p_4\rangle_a~ \kappa_{a_3 p_3, a_4 p_4}\; , \end{equation} where the indices ($p_1,p_2,p_3,p_4 = +,-$) refer to the large and small components of the quasiparticle Dirac spinors. In most applications of the RHB model, only the large components of the spinors $U_{k}({\bf r})$ and $V_{k}({\bf r})$ have been included in the non-relativistic pairing tensor $\hat{\kappa}$ in Eq. (\ref{kappa0}). The resulting pairing field reads \begin{equation} \hat{\Delta}_{a_1 +, a_2 +} = {\frac{1}{2}}\sum\limits_{a_3 +, a_4 +} \langle a_1 +, a_2 + |V^{pp}|a_3 +, a_4 +\rangle_a~ \kappa_{a_3 +, a_4 +}\; . \end{equation} The other components: $\hat{\Delta}_{+-}$, $\hat{\Delta}_{-+}$, and $\hat{\Delta}_{--}$ are neglected, in accordance with the results that are obtained with a relativistic zero-range force \cite{SR.02}. The ground state of an open-shell nucleus is described by the solution of the relativistic Hartree-Bogoliubov equations \begin{equation} \left( \begin{array} [c]{cc}% \hat{h}_{D}-m-\mu & \hat{\Delta}\\ -\hat{\Delta}^{\ast} & -\hat{h}^{\ast}_{D}+m+\mu \end{array} \right) \left( \begin{array} [c]{c}% U_{k}(\mathbf{r})\\ V_{k}(\mathbf{r}) \end{array} \right) =E_{k}\left( \begin{array} [c]{c}% U_{k}(\mathbf{r})\\ V_{k}(\mathbf{r}) \end{array} \right) \;, \label{eqhb}% \bigskip \end{equation} which correspond to the stationary limit of Eq. (\ref{TDRHB}). The chemical potential $\mu$ is determined by the particle number subsidiary condition in order that the expectation value of the particle number operator in the ground state equals the number of nucleons. The column vectors denote the quasiparticle wave functions, and $E_{k}$ are the quasiparticle energies. The RHB wave functions determine the hermitian single-particle density matrix \begin{equation} \hat{\rho}_{ll^{\prime}}=(V^{\ast}V^{T})_{ll^{\prime}},% \label{rho0}% \end{equation} and the antisymmetric anomalous density \begin{equation} \hat{\kappa}_{ll^{\prime}}=(V^{\ast}U^{T})_{ll^{\prime}}. \label{kappa0}% \end{equation} The calculated nuclear ground-state properties sensitively depend on the choice of the effective Lagrangian and pairing interaction. Over the years many parameter sets of the mean-field Lagrangian have been derived that provide a satisfactory description of nuclear properties along the $\beta $-stability line. The most successful RMF effective interactions are purely phenomenological, with parameters adjusted to reproduce the nuclear matter equation of state and a set of global properties of spherical closed-shell nuclei. This framework has recently been extended to include effective Lagrangians with explicit density-dependent meson-nucleon couplings. In a number of studies it has been shown that this class of global effective interactions provides an improved description of asymmetric nuclear matter, neutron matter and finite nuclei far from stability. In the present analysis of ground-state properties of Sn isotopes the density-dependent effective interaction DD-ME1~\cite{Nik1.02} will be employed in the particle-hole ($ph$) channel of the RHB model. In the following we extend the regularization scheme of Bulgac and Yu \cite{Bul1,Bul2,Bul3} to the solution of the relativistic Hartree-Bogoliubov equations for a zero-range pairing interaction \begin{equation} V^{pp}(\mathbf{r},\mathbf{r}^{\prime}) = g \delta(\mathbf{r} - \mathbf{r}^{\prime})\; . \label{pair_int} \end{equation} In Refs.~\cite{Bul1,Bul2} it has been shown that in this case the renormalized pairing field can be expressed as \begin{equation} \Delta (\mathbf{r}) = -g_{\mathit{eff}}(\mathbf{r}) \kappa_c (\mathbf{r}) \;, \label{eq:Delta} \end{equation} where $\kappa_c(\mathbf{r})$ denotes the cut-off anomalous density \begin{equation} \kappa_c(\mathbf{r}) =\sum\limits_{E_k>0}^{E_c} V_{k}^{\dagger}(\mathbf{r})U_{k}(\mathbf{r}) \;. \label{pair_c} \end{equation} The cut-off energy $E_c$ defines the two corresponding momenta $k_c$ and $l_c$ \begin{eqnarray} \sqrt{k_c^2(\mathbf{r}) + {m^*}^2(\mathbf{r})}+ V(\mathbf{r}) - m &=&E_c +\mu \label{cut1} ,\\ \sqrt{l_c^2(\mathbf{r}) + {m^*}^2(\mathbf{r})}+ V(\mathbf{r}) - m &=&-E_c+\mu \label{cut2}\;. \end{eqnarray} $m^*(\mathbf{r}) = m +S(\mathbf{r})$ is the Dirac mass, and $S(\mathbf{r})$ and $V(\mathbf{r})$ are, respectively, the scalar and vector single-nucleon potentials contained in the Dirac Hamiltonian $\hat{h}_{D}$. The chemical potential $\mu$ determines the local Fermi momentum \begin{equation} \sqrt{k_f^2(\mathbf{r}) + {m^*}^2(\mathbf{r})} + V(\mathbf{r}) - m =\mu\;. \label{kfermi} \end{equation} The effective, position-dependent coupling in Eq.~(\ref{eq:Delta}) reads \begin{equation} \frac{1}{ g_{\mathit{eff}}(\mathbf{r})} = \frac{1}{g} + F_1(\mathbf{r}) + F_2(\mathbf{r}) \;, \label{g_eff} \end{equation} with \begin{eqnarray} F_1(\mathbf{r}) &=& -{{k_c(\mathbf{r})\sqrt{k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{4\pi^2}} \left [ 1 - {{k_f(\mathbf{r})}\over{k_c(\mathbf{r})}}~{\rm Ar~cth} {{k_c(\mathbf{r})}\over{k_f(\mathbf{r})}} \right ] \nonumber \\ &&-{{k_c(\mathbf{r})\sqrt{k_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{8\pi^2}} + {{\sqrt{k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{4\pi^2}}~k_f(\mathbf{r}) ~{\rm Ar~cth} {{k_c(\mathbf{r})\sqrt{k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over {k_f(\mathbf{r})\sqrt{k_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}} \nonumber \\ &&-{{2k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}\over {8\pi^2}} ~{\rm ln} {{k_c(\mathbf{r})+ \sqrt{k_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over {{m^*}(\mathbf{r})}} \label{F1Re} \end{eqnarray} \begin{eqnarray} F_2(\mathbf{r}) &=& -{{l_c(\mathbf{r})\sqrt{k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{4\pi^2}} \left [ 1 - {{k_f(\mathbf{r})}\over{l_c(\mathbf{r})}}~{\rm Ar~th} {{l_c(\mathbf{r})}\over{k_f(\mathbf{r})}} \right ] \nonumber \\ &&-{{l_c(\mathbf{r})\sqrt{l_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{8\pi^2}} + {{\sqrt{k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{4\pi^2}}~k_f(\mathbf{r}) ~{\rm Ar~th} {{l_c(\mathbf{r})\sqrt{k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over {k_f(\mathbf{r})\sqrt{l_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}} \nonumber \\ &&-{{2k_f^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}\over {8\pi^2}} ~{\rm ln} {{l_c(\mathbf{r})+ \sqrt{l_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over {{m^*}(\mathbf{r})}} \label{F2Re} \end{eqnarray} $F_1 + F_2$ is the relativistic generalization of the corresponding correction to the coupling constant $g$, as defined in Eq. (16) of Ref.~\cite{Bul1}. \section{\label{SecNM}Pairing properties of symmetric nuclear matter} A zero-range pairing interaction leads to a particularly simple expression for the gap equation in symmetric nuclear matter \begin{equation} \frac{1}{ g_{\mathit{eff}}} = - \frac{1}{4\pi^2}\int_{l_c}^{k_c} dk~ {k^2\over \sqrt{\left [ \sqrt{k^2+{m^*}^2} - \sqrt{k_f^2+{m^*}^2}~ \right ]^2 + \Delta^2}} \; , \label{gap} \end{equation} The momenta $k_c$ and $l_c$ are determined by the cut-off energy $E_c$ Eqs. (\ref{cut1},~\ref{cut2}), and the effective coupling $g_{\mathit{eff}}$ is defined in Eq.~(\ref{g_eff}). In the left panel of Fig.~\ref{FigA} we display the density dependence of the resulting pairing gap in nuclear matter (dashed curve). The single-particle spectrum has been calculated with the relativistic effective interaction DD-ME1~\cite{Nik1.02}, and the coupling constant of the zero-range pairing interaction Eq.~(\ref{pair_int}) $g = -330$ MeV fm$^{3}$ is typical for the values used by Bulgac and Yu in their analyses. The pairing gap is shown in comparison to the gap calculated with the effective Gogny interaction D1S \cite{BGG.91} (dots). The corresponding single-particle spectrum has been computed in the Hartree-Fock approximation for the Gogny interaction. The density dependence of the two gaps is completely different. The pairing gap of the renormalized zero-range interaction increases uniformly with density, whereas the gap of the Gogny interaction display the characteristic maximum of $\approx 2.5$ MeV at low density $\rho = 0.03 - 0.04$ fm$^{-3}$ (corresponding to a Fermi momentum of approximately 0.8 fm$^{-1}$) and decreases at higher densities. The bell-shaped form of the pairing gap as a function of the density was, in fact, obtained already more than forty years ago \cite{ES.60}. This density dependence is not characteristic only of the phenomenological finite-range interactions, but is also obtained when the gap is calculated with bare nucleon-nucleon potentials adjusted to the empirical nucleon-nucleon phase shifts and deuteron properties (for a recent review see Ref.~\cite{DH-J.03}). The decrease of the gap at Fermi momenta $k_f > 0.8$ fm$^{-1}$ simply reflects the repulsive character of the nucleon-nucleon interaction at short distances \cite{SRR.02}. Of course there is no repulsive component in the zero-range force with constant coupling Eq.~(\ref{pair_int}), and the corresponding pairing gap displays the unphysical uniform increase with density. We notice, however, that in the range of densities shown in Fig.~\ref{FigA}, i.e. up to nuclear matter saturation density, the values of the pairing gap of the renormalized zero-range interaction are comparable with those of the Gogny pairing gap. As will be shown in the next section, this means that the renormalization scheme for the zero-range interaction with constant coupling can be safely applied to the calculation of pairing correlations in finite nuclei, provided an appropriate choice is made for the strength parameter $g$. On the other hand, there is no particular reason why the strength parameter $g$ of the zero-range pairing interaction should be a constant. In fact, in many applications to finite nuclei an explicit density dependence is introduced, and in this way pairing correlations partially include finite-range effects. For instance, in one of the first applications \cite{BE.91} Bertsch and Esbensen used a density-dependent contact interaction, together with a simple energy cut-off, in a description of pairing correlations in weakly bound neutron-rich nuclei. They also compared the corresponding pairing gap in symmetric nuclear matter with the result of a Hartree-Fock calculation using the Gogny interaction. In the present anaysis we have adjusted a density-dependent strength parameter $g(\rho)$ of the zero-range pairing interaction Eq.~(\ref{pair_int}), in such a way that the pairing gap of the renormalized zero-range interaction Eq.~(\ref{gap}), reproduces the density dependence of the Gogny pairing gap. The resulting density dependence can be approximated by the following analytic expression \begin{equation} g(\rho) = \frac{1}{a_0 + a_1 \rho^{1/3} + a_2 \rho^{2/3}} \; , \label{g_rho} \end{equation} with $a_0 = -0.064$ fm$^{-2}$, $a_1 = 0.447$ fm$^{-1}$, and $a_2 =-3.693$. The resulting pairing gap, displayed in the left panel of Fig.~\ref{FigA} (solid line), is in very good agreement with the one calculated using the Gogny interaction. A very similar procedure was employed in Ref.\cite{YuB.03}, where the density dependence of the ``bare coupling constant" $g(\rho)$ was adjusted to a specific formula for the pairing gap in low-density homogeneous neutron matter. In the right panel of Fig.~\ref{FigA} we display the effective couplings $g_{\mathit{eff}}$ calculated using the constant $g = -330$ MeV fm$^{3}$, and the density dependent coupling of Eq. (\ref{g_rho}). The density dependence of the two effective couplings is completely different. In order to prevent an unphysical growth of the pairing gap with density, the density dependence of the pairing strength Eq. (\ref{g_rho}) ensures that the effective coupling becomes weaker with increasing nucleon density. A very strong effective coupling in the low-density region produces a peak in the corresponding pairing gap shown in the left panel. On the other hand, $g_{\mathit{eff}}$ calculated using the constant coupling increases in absolute value with density, i.e. the resulting pairing gap increases uniformly with density. However, rather similar values for the two effective couplings $g_{\mathit{eff}}$ are calculated in the region of densities characteristic for the bulk of finite nuclei. One should not, therefore, expect very different results for the pairing properties of finite nuclei calculated with the zero-range interaction with constant coupling, or with the density-dependent coupling of Eq. (\ref{g_rho}). In the next section we will show that this is really not true in weakly-bound nuclei far from stability. The renormalization prescription must, of course, lead to a pairing field which is independent of the cut-off energy $E_c$, if the latter is chosen large enough. This is illustrated in Fig.~\ref{FigB}, where we plot the pairing gap, calculated using the density-dependent coupling of Eq.~(\ref{g_rho}), for a number of characteristic values $E_c$ in the interval between 5 MeV and 60 MeV. The pairing gap shows a weak dependence on the cut-off energy only for the two lowest values of $E_c$. When the cut-off is increased beyond 10 MeV, the corresponding pairing gaps cannot be distinguished. Thus already for $E_c \geq 10$ MeV the pairing gap of the renormalized zero-range interaction in symmetric nuclear matter converges. This is in agreement with the results obtained in the analysis of the pairing gap in homogeneous neutron matter~\cite{Bul1}. \section{\label{SecSn}Ground-state pairing properties of spherical nuclei} In this section the renormalization scheme is tested in the calculation of ground-state pairing properties of Sn isotopes. The DD-ME1 mean-field Lagrangian is employed for the $ph$ channel, and the zero-range interaction Eq.~(\ref{pair_int}) is used in the $pp$ channel. The renormalization procedure described in the previous section is carried out for the zero-range interaction with constant pairing strength $g = -330$ MeV fm$^{3}$, and and for the density-dependent coupling of Eq.~(\ref{g_rho}). In the latter case the density dependence of the pairing strength has been adjusted to reproduce the Gogny D1S pairing gap in symmetric nuclear matter. In the following we denote by RCC the case of the renormalized constant coupling, and by RDDC the results obtained with the renormalized density-dependent coupling. While in the symmetric nuclear matter the Fermi momentum is always real (see Eq. (\ref{kfermi})), in the surface region of finite nuclei it becomes imaginary. In Ref.~\cite{Bul1} it has been shown that also in this case the renormalized anomalous density is real. The effective coupling $g_{\mathit{eff}}$ is still given by Eq.~(\ref{g_eff}), but \begin{eqnarray} F_1(\mathbf{r}) &=& -{{k_c(\mathbf{r})\sqrt{-|k_f(\mathbf{r})|^2+ {m^*}^2(\mathbf{r})}}\over{4\pi^2}} \left [ 1 - {{k_f(\mathbf{r})}\over{k_c(\mathbf{r})}}~{\rm Ar~ctg} {{k_c(\mathbf{r})}\over{k_f(\mathbf{r})}} \right ] \nonumber \\ &&-{{k_c(\mathbf{r})\sqrt{k_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over{8\pi^2}} + {{\sqrt{-|k_f(\mathbf{r})|^2+ {m^*}^2(\mathbf{r})}}\over{4\pi^2}}~k_f(\mathbf{r}) ~{\rm Ar~ctg} {{k_c(\mathbf{r})\sqrt{-|k_f(\mathbf{r})|^2+ {m^*}^2(\mathbf{r})}}\over {k_f(\mathbf{r})\sqrt{k_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}} \nonumber \\ &&-{{(-2|k_f(\mathbf{r})|^2+ {m^*}^2(\mathbf{r}))}\over {8\pi^2}} ~{\rm ln} {{k_c(\mathbf{r})+ \sqrt{k_c^2(\mathbf{r})+ {m^*}^2(\mathbf{r})}}\over {{m^*}(\mathbf{r})}} \;, \label{F1Im} \end{eqnarray} and \begin{equation} F_2(\mathbf{r}) = 0\;. \label{F2Im} \end{equation} However, if either $k_c$ or $l_c$ becomes imaginary, the corresponding terms in the effective coupling should be omitted. The rate of convergence of the renormalization scheme is illustrated in Fig.~\ref{FigC} where, for the nucleus $^{114}$Sn, we display the average pairing gaps and the pairing energies as functions of the cut-off energy $E_c$. The average gaps shown in the left panel, are defined as \begin{equation} < \Delta_N > = {{\sum_{nlj} \Delta_{nlj} v_{nlj}^2}\over {\sum_{nlj} v_{nlj}^2}} \; , \label{ang} \end{equation} where $v_{nlj}^2$ are the occupation probabilities of the neutron states in the canonical basis. Both the pairing gaps and the pairing energies converge already for $E_c \geq 10$ MeV. We also notice that, even though the renormalized constant coupling and the renormalized density-dependent coupling lead to very different pairing gaps in symmetric nuclear matter, in $^{114}$Sn they produce similar average pairing gaps and virtually identical pairing energies. The corresponding pairing fields as functions of the radial coordinate, and $g_{\mathit{eff}}$ Eq.~(\ref{g_eff}) as functions of the density, are plotted in Fig.~\ref{FigD} for a series of values of the energy cut-off. In both cases the calculation of the pairing field and $g_{\mathit{eff}}$ shows convergence for $E_c > 10$ MeV. While the renormalized constant coupling and the renormalized density-dependent coupling produce very similar average pairing gaps and pairing energies, the dependence of the corresponding pairing fields on the radial coordinate is rather different. The RCC pairing field (upper left panel) is concentrated in the bulk of the nucleus, whereas the RDDC pairing field (lower left panel) exhibits a pronounced peak on the surface. This behavior reflects the difference between the effective couplings $g_{\mathit{eff}}$, already shown in the right panel of Fig.~\ref{FigA} for the case of symmetric nuclear matter. In the panels on the right of Fig.~\ref{FigD} we plot the effective couplings $g_{\mathit{eff}}(r(\rho))$ as functions of the nucleon density in $^{114}$Sn. The $g_{\mathit{eff}}$ which corresponds to the density-dependent coupling of Eq.~(\ref{g_rho}) decreases steeply in the region of very low density, i.e., on the surface of the nucleus. Consequently, also the pairing field displays a peak in the surface region. In both the RCC and RDDC cases the pronounced discontinuity of the effective coupling $g_{\mathit{eff}}$ at very low density corresponds to the transition from real to imaginary Fermi momentum $k_f$. This is illustrated in Fig.~\ref{FigE}, where we plot the effective single-nucleon potential (left panel) and the correction to the coupling originating from the renormalization of the anomalous density (right panel). The effective single-nucleon potential is determined by the sum of the vector and scalar potentials $V_{cen}(\mathbf{r})=S(\mathbf{r})+V(\mathbf{r})$. For real values of the Fermi momentum (the effective potential is below the chemical potential $\mu$) the correction to the coupling $F_1(\mathbf{r})+F_2(\mathbf{r})$ is calculated from Eqs. (\ref{F1Re}) and (\ref{F2Re}), and for imaginary values of the Fermi momentum (the effective potential is above the chemical potential $\mu$) from Eqs. (\ref{F1Im}) and (\ref{F2Im}). In the region where the Fermi momentum changes from real to imaginary the correction $F_1(\mathbf{r})+F_2(\mathbf{r})$ displays a very sharp peak, which is reflected in the discontinuities of the effective couplings. The importance of possible surface effects is illustrated in Fig.~\ref{FigF}, where we plot the calculated average pairing gaps and pairing energies for the chain of even-even Sn isotopes with $110 \le A \le 160$. Although both the RCC and RDCC schemes lead to comparable values of the average pairing gaps for the entire isotopic chain, the pairing energies differ significantly for isotopes beyond the doubly closed-shell $^{132}$Sn. For example, the pairing energy of $^{150}$Sn calculated with renormalized density-dependent coupling (RDDC) is almost 25 MeV larger than the one calculated with the renormalized constant coupling (RCC). The large increase in the pairing energy for the RDDC case is caused by the dominant role of the surface region for the very neutron-rich Sn isotopes, and because the effective coupling is especially strong at very low densities. In the panels on the left of Figs.~\ref{FigG} and~\ref{FigH} we plot the self-consistent solutions for the cut-off anomalous densities Eq.~(\ref{pair_c}) for the isotopes $^{114}$Sn, $^{124}$Sn and $^{150}$Sn, calculated using the RDDC and RCC effective couplings, respectively. The corresponding effective couplings $g_{\mathit{eff}}$ are shown in the panels on the right of Figs.~\ref{FigG} and~\ref{FigH}. The anomalous densities for $^{114}$Sn and $^{124}$Sn are concentrated in the nuclear volume ($r \le 6$ fm), where the effective couplings $g_{\mathit{eff}}$ have comparable values. Therefore, the corresponding pairing energies are similar for the RDDC and RCC cases. In $^{150}$Sn, on the other hand, the anomalous densities extend to the region $r\ge 8$ fm, where the RDDC effective coupling becomes much stronger than the one calculated with the constant coupling (RCC). Hence, the pairing energy for $^{150}$Sn, calculated using the renormalized density-dependent coupling is much larger than the one obtained with the renormalized constant coupling. \section{\label{SecSum}Conclusions} A simple renormalization scheme for the Hartree-Fock-Bogoliubov equations with zero-range pairing has recently been introduced \cite{Bul1,Bul2,Bul3}. In the present work we have implemented this renormalization scheme for the relativistic Hartree-Bogoliubov equations with a zero-range pairing interaction. The procedure is equivalent to a simple energy cut-off with a position dependent coupling constant. We have verified that the resulting average pairing gaps and pairing energies do not depend on the cut-off energy $E_c$, if the latter is chosen large enough. Convergence is achieved already for values $E_c \ge 10$ MeV, both in nuclear matter and for finite nuclei. If the strength parameter of the zero-range pairing is a constant, the resulting pairing gap in symmetric nuclear matter displays an unphysical increase with density. We have therefore adjusted a density-dependent strength parameter of the zero-range pairing in such a way that the renormalization procedure reproduces in symmetric nuclear matter the pairing gap of the phenomenological Gogny interaction. In this sense the present study goes beyond the simple extension of the renormalization scheme of Ref.~\cite{Bul1} to the relativistic framework. However, the resulting effective coupling is too strong in the region of low density, and this leads to large pairing energies in open-shell nuclei with very diffuse surfaces, e.g. in neutron-rich Sn isotopes. One must therefore be careful when applying the renormalized HFB or RHB models with zero-range pairing to nuclei far from stability. Adjusting the strength parameter to the pairing gap in symmetric nuclear matter obviously does not provide enough information about the density dependence of the zero-range pairing to be used in very neutron-rich nuclei. \leftline{\bf ACKNOWLEDGMENTS} This work has been supported in part by the Bundesministerium f\"ur Bildung und Forschung - project 06 MT 193, by the Alexander von Humboldt Stiftung, and by the Croatian Ministry of Science - project 0119250. \bigskip
{ "timestamp": "2005-03-30T10:04:40", "yymm": "0503", "arxiv_id": "nucl-th/0503078", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503078" }
\section{Introduction} Search for spatiotemporal solitons in diverse optical media, alias ``light bullets" (LBs) \cite{Yaron}, is a challenge to fundamental and applied research in nonlinear optics, see original works \cite{KR,Koshiba,chi2,Miriam,tandem,Frank,Isaac,saturable,Wagner} and a very recent review \cite{review}. Stationary solutions for LBs can easily be found in the cubic ($\chi ^{(3)}$) multi-dimensional nonlinear Schr\"{o}dinger (NLS) equation \cite{Yaron}, but their stability is a problem, as they are unstable against spatiotemporal collapse \cite{Berge}. The problem may be avoided by resorting to milder nonlinearities, such as saturable \cite{saturable}, cubic-quintic \cite{CQ}, or quadratic ($\chi ^{(2)}$) \cite{KR,Koshiba,chi2,Miriam,tandem,Frank,Isaac}. Despite considerable progress in theoretical studies, three-dimensional (3D) LBs in a bulk medium have not yet been observed in an experiment. The only successful experimental finding reported thus far was a stable quasi-2D spatiotemporal soliton in $\chi ^{(2)}$ crystals \cite{Frank} (the tilted-wavefront technique \cite{Paolo}, used in that work, precluded achieving self-confinement in one transverse direction). On the other hand, it was predicted \cite{Isaac} that a spatial cylindrical soliton may be stabilized in a bulk medium composed of layers with alternating signs of the Kerr coefficient. Similar stabilization was then predicted for what may be regarded as 2D\ solitons in Bose-Einstein condensates (BECs), with the coefficient in front of the cubic nonlinear term subjected to periodic modulation in time via the \textit{Feshbach resonance} in external ac magnetic field \cite{FR,Castilla}. However, no stable 3D soliton could be found in either realization (optical or BEC) of this setting. Serious difficulties encountered in the experimental search for LBs in 3D media is an incentive to look for alternative settings admitting stable\emph{\/} 3D optical solitons. With the Kerr nonlinearity, a possibility is to use a layered structure that periodically reverses the sign of the local group-velocity dispersion (GVD), without affecting the $\chi ^{(3)}$ coefficient. This resembles a well-known scheme in fiber optics, known as \textit{dispersion management}\ (DM), see, e.g., Refs. \cite{DM} and review \cite{Progress}. A 2D generalization of the DM scheme was recently proposed, assuming a layered planar waveguide of this type, uniform in the transverse direction \cite{we,Spain}. As a result, large stability regions for the 2D spatiotemporal solitons were identified, including double-peaked breathers; however, a 3D version of the same model could not give rise to any stable soliton \cite{we}. It was also shown in Ref. \cite{we} that no stable 3D soliton could be found in a more sophisticated model, which combines the DM and periodic modulation of the Kerr coefficient in the longitudinal direction. Another approach to the stabilization of multidimensional solitons was developed in the context of the self-attracting BEC. It is based on the corresponding Gross-Pitaevskii equation which includes a periodic potential created as an optical lattice (OL, i.e., an interference pattern produced by illuminating the condensate by counter-propagating coherent laser beams). It has been demonstrated that 2D \cite{BBB,Yang,Estoril} and 3D \cite{BBB} solitons can be easily stabilized by the OL of the same dimension. Moreover, stable solitons can also be readily supported by \emph{low-dimensional\/} OLs, i.e., 1D and 2D ones in the 2D \cite{Estoril,BBB2} and 3D \cite{Estoril,BBB2,Barcelona} cases, respectively; additionally, a 3D soliton can be stabilized by a cylindrical (\textit{Bessel}) lattice \cite{Dumitru}, similar to one introduced, in the context of 2D models, in Ref. \cite{Bessel}. On the other hand, 3D solitons cannot be stabilized by a 1D periodic potential \cite{BBB2}; however, the 1D lattice potential in combination with the above-mentioned time-periodic modulation of the nonlinearity, provided by the Feshbach resonance in the ac magnetic field, supports single- and multi-peaked stable 3D solitons in vast areas of the respective parameter space \cite{we-new}. The above results suggest a possibility of existence of stable 3D ``bullets" in a $\chi ^{(3)}$ medium with the DM in the longitudinal direction ($z$), additionally equipped with an effective lattice potential (i.e., periodic modulation of the refractive index) in one transverse direction ($y$), while in the remaining transverse direction ($x$) the medium remains uniform. If this is possible, stable LBs will be definitely possible too in a medium with the periodic modulation of the refractive index in both transverse directions; however, the setting with one uniform direction is more interesting in terms of steering solitons and studying collisions between them \cite{Estoril,BBB2}. The objective of the present work is to predict such 3D spatiotemporal solitons and investigate their stability. Our first consideration of this possibility is based on the variational approximation (VA); systematic simulations of the 3D model are quite complicated, and will be presented in a follow-up work. It is relevant to mention that the existence and stability of 3D solitons in the Gross-Pitaevskii equation with the quasi-2D periodic potential, which were originally predicted by the VA\cite{Estoril,BBB2}, was definitely confirmed by direct simulations \cite{Estoril,BBB2,Barcelona}, which suggests that in the present model the 3D solitons may easily be stable too. The model is based on the normalized NLS equation describing the evolution of the local amplitude $u$ of the electromagnetic wave, which is a straightforward extension of the 2D model put forward in Ref. \cite{we}: \begin{equation} i\frac{\partial u}{\partial z}+\left[ \frac{1}{2}\left( \frac{\partial ^{2}}{\partial x^{2}}+\frac{\partial ^{2}}{\partial y^{2}}+D(z)\frac{\partial ^{2}}{\partial \tau ^{2}}\right) +\varepsilon \cos (2y)+|u|^{2}\right] u=0. \label{general} \end{equation}Here, $\varepsilon $ is the strength of the transverse modulation (the modulation period is normalized to be $\pi $), while $\tau $ and $D(z)$ are the same reduced temporal variable and local GVD coefficient as in the fiber-optic DM models \cite{DM,Progress}. Equation (\ref{general}) implies the propagation of a linearly polarized wave, with the single component $u$; a more general situation will be described by a two-component (vectorial) version of Eq. (\ref{general}), with the two polarization coupled, as usual, by the cubic cross-phase-modulation terms. We do not expect that the vectorial model will produce results qualitatively different form those presented below. As usual, the NLS equation assumes the applicability of the paraxial approximation, i.e., the spatial size of solitons (see below) must be much larger than the underlying wavelength of light, which is definitely a physically relevant assumption \cite{review}, and the temporal part of the equation implies that the higher-order GVD is negligible (previous considerations have demonstrated that the higher-order dispersion does not drastically alter DM solitons \cite{TOD}). As is commonly adopted, we assume a symmetric \textit{DM map}, with equal lengths $L$ of the normal- and anomalous-GVD segments (usually, the results are not sensitive to the map's asymmetry), \begin{equation} D(z)=\left\{ \begin{array}{l} \overline{D}+D_{\mathrm{m}}>0,\,0<z<L, \\ \overline{D}-D_{\mathrm{m}}<0,\,L<z<2L,\end{array}\right. \label{D(z)} \end{equation}the average dispersion being much smaller than the modulation amplitude, $\left\vert \overline{D}\right\vert \ll D_{\mathrm{m}}$. Using the scaling invariances of Eq. (\ref{general}), we fix $L\equiv 1$ and $D_{\mathrm{m}}\equiv 1$. Recently, a somewhat similar 2D model was introduced in Ref. \cite{Salerno}. The most important difference is that it has the variable coefficient $D(z)$ multiplying \emph{both} the GVD and diffraction terms, $u_{\tau \tau }$ and $u_{xx}$. Actually, that model was motivated by a continuum limit of some discrete systems; in the present context, it would be quite difficult to implement the periodic reversal of the sign of the transverse diffraction. \section{The variational approximation} Aiming to apply the VA for the search of LB solutions (a review of the variational method can be found in Ref. \cite{Progress}), we adopt the Gaussian \textit{ansatz}, \begin{eqnarray} u &=&A(z)\exp \left\{ \mathrm{i}\phi (z)-\frac{1}{2}\left[ \frac{x^{2}}{W^{2}(z)}+\frac{y^{2}}{V^{2}(z)}+\frac{\tau ^{2}}{T^{2}(z)}\right] \right. + \nonumber \\ &&+\left. \frac{\mathrm{i}}{2}\left[ b(z)\,x^{2}+c(z)\,y^{2}+\beta (z)\,\tau ^{2}\right] \right\} , \label{ansatz} \end{eqnarray}where $A$ and $\phi $ are the amplitude and phase of the soliton, $T$ and $W,V$ are its temporal and two transverse spatial widths, and $\beta $ and $b,c$ are the temporal and two spatial chirps. The Lagrangian from which Eq. (\ref{general}) can be derived is \begin{eqnarray} L &=&\frac{1}{2}\int_{-\infty }^{+\infty }\,\mathrm{d}x\int_{-\infty }^{+\infty }\,\mathrm{d}y\int_{-\infty }^{+\infty }\,\mathrm{d}\tau \left[ \mathrm{i}\left( u_{z}u^{\ast }-u_{z}^{\ast }u\right) -\left\vert u_{x}\right\vert ^{2}-\left\vert u_{y}\right\vert ^{2}-D\left\vert u_{\tau }\right\vert ^{2}\right. \\ &&\left. +2\varepsilon \cos (2y)|u|^{2}+|u|^{4}\right] . \end{eqnarray}The substitution of the ansatz (\ref{ansatz}) in this expression and integrations lead to an \textit{effective Lagrangian}, with the prime standing for $d/dz$: \begin{eqnarray} (4/\pi ^{3/2})L_{\mathrm{eff}} &=&A^{2}WVT\left[ 4\phi ^{\prime }-b^{\prime }W^{2}-c^{\prime }V^{2}-\beta ^{\prime }T^{2}-W^{-2}-V^{-2}-DT^{-2}\right. \nonumber \\ &&\left. -b^{2}W^{2}-c^{2}V^{2}+\varepsilon \exp \left( -V^{2}\right) -D(z)\beta ^{2}T^{2}+A^{2}/\sqrt{2}\right] , \label{effL} \end{eqnarray} The first variational equation, $\delta L_{\mathrm{eff}}/\delta \phi =0$, applied to Eq. (\ref{effL}) yields the energy conservation, $dE/dz=0$, with \begin{equation} E\equiv A^{2}WVT. \label{E} \end{equation}The conservation of $E$ is used to eliminate $A^{2}$ from the set of subsequent equations, $\delta L_{\mathrm{eff}}/\delta \left( W,V,T,b,c,\beta \right) =0$. They can be arranged so as, first, to eliminate the chirps, \begin{equation} b=W^{\prime }/W,\,c=V^{\prime }/V,\beta =D^{-1}T^{\prime }/T. \label{betab} \end{equation}the remaining equations for the spatial and temporal widths being \begin{eqnarray} W^{\prime \prime } &=&\frac{1}{W^{3}}-\frac{E}{2\sqrt{2}W^{2}VT}, \label{variat1} \\ V^{\prime \prime } &=&\frac{1}{V^{3}}-4\varepsilon V\exp \left( -V^{2}\right) -\frac{E}{2\sqrt{2}WV^{2}T}, \label{variat2} \\ \left( \frac{T^{\prime }}{D}\right) ^{\prime } &=&\frac{D}{T^{3}}-\frac{E}{2\sqrt{2}WVT^{2}}. \label{variat3} \end{eqnarray}{The Hamiltonian of these equations, which is a dynamical invariant in the case of constant $D$, is \[ {\mathcal{H}}=\left( W^{\prime }\right) ^{2}+\left( V^{\prime }\right) ^{2}+\frac{\left( T^{\prime }\right) ^{2}}{D}+\frac{1}{W^{2}}+\frac{1}{V^{2}}+\frac{D}{T^{2}}-4\varepsilon \exp (-V^{2})-\frac{E}{\sqrt{2}WVT} \]} In the case of the piece-wise constant modulation, such as in (\ref{D(z)}), the variables $W$, $W^{\prime }$, $V$, $V^{\prime }$, $T$ and $\beta $ must be continuous at junctions between the segments with $D_{\pm }\equiv \overline{D}\pm D_{\mathrm{m}}$. As it follows from Eq. (\ref{betab}), the continuity of the temporal chirp $\beta (z)$ implies a jump of the derivative $T^{\prime }$ when passing from $D_{-}$ to $D_{+}$, or vice versa: \begin{equation} \left( T^{\prime }\right) _{D=D_{+}}=\left( D_{+}/D_{-}\right) \left( T^{\prime }\right) _{D=D_{-}}. \label{jump} \end{equation} In the case of a continuous DM map, rather than the one (\ref{D(z)}), Eq. (\ref{variat3}) has a formal singularity at the points where $D(z)$ vanishes, changing its sign. However, it is known that there is no real singularity in this case, as $T^{\prime }$ vanishes at the same points, which cancels the singularity out \cite{Progress}. In the absence of the DM and transverse modulation, i.e., $D\equiv +1$ and $\varepsilon =0$, three equations (\ref{variat1}) - (\ref{variat3})\ reduce to one, which is tantamount to the variational equation derived in Ref. \cite{Sweden} from the spatiotemporally isotropic ansatz [cf. Eq. (\ref{ansatz})], $u=A\exp \left[ \mathrm{i}\phi -(1/2)\left( W^{-2}+\mathrm{i}b\right) \left( x^{2}+y^{2}+\tau ^{2}\right) \right] $. In particular, this single equation correctly predicts the asymptotic law of the \textit{strong collapse} in the 3D case, which is stable against anisotropic perturbations \cite{Russia}, $V=W=T\approx \left( 5E/3\sqrt{2}\right) ^{1/5}\left( z_{0}-z\right) ^{2/5}$, $z=z_{0}$ being the collapse point. The location of this point is determined by initial conditions, but, in any case, it belongs to an interval $D>0$, where the GVD is anomalous. Another possible collapse scenario is an effectively two-dimensional (weak) one, with two widths shrinking to zero as $z_{0}-z\rightarrow 0$, while the third one remains finite. For instance, the corresponding asymptotic law may be\begin{equation} V=T=A\left( z_{0}-z\right) ^{1/2},~W=\frac{\sqrt{2}E}{4+A^{4}}-\frac{\left( 4+A^{4}\right) ^{2}}{4\sqrt{2}EA^{2}}\left( z_{0}-z\right) \ln \left( z_{0}-z\right) , \label{2Dcollapse} \end{equation}where $A$ is a positive constant or else $V=2/T\sim (z_{0}-z)^{1/2}$ and $W\rightarrow W_{0}$ (in this case too, the collapse point $z_0$ must belong to a segment with $D>0$). In direct simulations of Eqs. (\ref{variat1}) - (\ref{variat3}), we actually observed only the latter scenario. However, we did not specially try to find initial conditions that could initiate a solution corresponding to the strong 3D collapse, as our objective is not the study of the collapse, but rather search for solitons stable against collapse. In fact, known results for the solitons in the 3D Gross-Pitaevskii equation with the OL potential suggest that, while the VA may be incorrect in the description of the collapse, as a singular solution, it provides for quite accurate predictions for the stability of solitons as \emph{regular solutions} \cite{Estoril,BBB2}. If the DM is absent, and the constant GVD is normal, i.e., $D\equiv -1$, only the 2D collapse in the transverse plane would be possible, so that (cf. Eq. (\ref{2Dcollapse}))\[ V=W=A\left( z_{0}-z\right) ^{1/2},~T=\frac{\sqrt{2}E}{4+A^{4}}+\frac{\left( 4+A^{4}\right) ^{2}}{4\sqrt{2}EA^{2}}\left( z_{0}-z\right) \ln \left( z_{0}-z\right) . \]However, we did not observed this collapse scenario in our simulations. The same comment as one given above pertains to this case as well. A possibility of the stabilization of the 3D soliton by a sufficiently strong lattice can be understood noticing that, for large $\varepsilon $, one may keep only the first two terms on the right-hand side of Eq. (\ref{variat2}). This approximation yields a nearly constant value $V_{0}$ of $V$, which is a smaller root of the corresponding equation, \begin{equation} 4\varepsilon V_{0}^{4}\exp \left( -V_{0}^{2}\right) =1 \label{static} \end{equation}(a larger root corresponds to an unstable equilibrium). The two roots exist provided that \begin{equation} \varepsilon >\varepsilon _{\mathrm{\min }}=\mathrm{e}^{2}/16\approx \allowbreak 0.46, \label{min} \end{equation}the relevant one being limited by $V_{0}<2$. Then, the substitution of $V=V_{0}$ in the remaining equations (\ref{variat1}) and (\ref{variat3}) leads to essentially the same VA-generated dynamical system as derived for the 2D DM model in Ref. \cite{we}, which was shown to give rise to stable spatiotemporal solitons. On the other hand, it was demonstrated in Ref. \cite{we} too that, in the case of $\varepsilon =0$, the 3D VA equations, as well as the full underlying 3D model, have no stable soliton solutions. The stabilization of the LB in the present model for large $\varepsilon $ can also be understood in a different way, without resorting to VA: in a very strong lattice, the soliton is trapped entirely in a single ``valley" of the periodic potential, and the problem thus reduces to a nearly 2D one, where spatiotemporal solitons may be stable, cf. a similar stabilization mechanism for the solitons in the Gross-Pitaevskii equations developed in \cite{Castilla}. From this point of view, a really interesting issue is to find an \emph{actual} minimum $\varepsilon _{\min }$ of the lattice's strength which is necessary for the stabilization of the 3D solitons, as at $\varepsilon $ close enough to $\varepsilon _{\mathrm{\min }} $ the stabilized solitons are truly 3D objects, rather than their nearly-2D counterparts. \section{Results} We explored the parameter space of the variational system (\ref{variat1}) - (\ref{variat3}), $\left( E,\varepsilon ,\overline{D}\right) $, by means of direct simulations of the equations (with regard to the jump condition (\ref{jump})). As a result, it was possible to identify regions where the model admits \emph{stable} solitons featuring regular oscillations in $z$ with the DM-map period. An example of such a regime is shown in Fig. \ref{fig1} (oscillations in the evolution of $W$ are not visible in the figure because, as an estimate demonstrates, their amplitude is $\simeq 0.001$). \begin{figure}[tbp] \includegraphics[width=13.5cm]{fig1.eps} \caption{An example of the stable evolution of solutions to the variational equations (\protect\ref{variat1}) - (\protect\ref{variat3}). The soliton's widths in the direction $x,y$ and $\protect\tau $, i.e., $W,V$ and $T$, are shown as functions of $z$, for $E=0.5$, $\protect\varepsilon =1$, and $\overline{D}=0$.} \label{fig1} \end{figure} Systematic results obtained from the simulations are summarized in stability diagrams displayed in Figs. \ref{fig2} and \ref{fig3}. A remarkable fact, apparent in Fig. \ref{fig2}, is that the minimum value of the lattice's strength, $\varepsilon _{\min }=0.46$, at which the solitons may be stable, coincides with the analytical prediction (\ref{min}), up to the available numerical accuracy. \begin{figure}[tbp] \includegraphics[width=13.5cm]{fig2.eps} \caption{The stability area for the 3D spatiotemporal solitons in the $\left( E,\protect\varepsilon \right) $ plane, with $\overline{D}=0$, is shown by light-gray shading. In gray and dark-gray regions, the 3D soliton is predicted, respectively, to spread out and collapse. The vertical line corresponds to the analytically predicted threshold (\protect\ref{min}).} \label{fig2} \end{figure} \begin{figure}[tbp] \includegraphics[width=13.5cm]{fig3.eps} \caption{The stability area in the $\left( E,\overline{D}\right) $ plane, with $\protect\varepsilon =1$, is shown by light--gray shading. In the gray region, the 3D soliton is predicted to spread out.} \label{fig3} \end{figure} The existence of a maximum value $E_{\max }$ of the energy admitting the stable LBs is, essentially, a quasi-2D feature, which can be understood assuming that the potential lattice is strong. Indeed, as explained above, in such a case the value of $V$ is approximately fixed as the smaller root of Eq. (\ref{static}). Within a segment where the GVD coefficient keeps the constant value, $D_{+}=\overline{D}+D_{\mathrm{m}}>0$, which corresponds to anomalous dispersion (see Eq. (\ref{D(z)}), the remaining equations (\ref{variat1}) and (\ref{variat3}) are tantamount to those for a uniform 2D Kerr-self-focusing medium, hence the energy is limited by the value $E_{\max }$ corresponding to the \textit{Townes soliton}; the soliton will collapse if $E>E_{\max }$ \cite{Berge}. The fact that the region of stable solitons is also limited by a minimum energy, $E_{\min }$, except for the case of $\overline{D}=0$, when $E_{\min }=0$ (see Fig. \ref{fig3})), is actually a quasi-1D feature, which is characteristic to the DM solitons in optical fibers. In that case, the term $\sim E$ in the evolution equation for $T(z)$, cf. Eq. (\ref{variat3}), is necessary to balance the average GVD coefficient $\overline{D}$, so that $E_{\min }$ and $\overline{D}$ vanish simultaneously \cite{DM}. It is noteworthy too that, as well as in the case of the 1D DM solitons in fibers, the stability area in Fig. \ref{fig3} includes a part with \emph{normal} average GVD, $\overline{D}<0$, which seems counterintuitive, but can be explained \cite{DM}. This part extends in Fig. \ref{fig3} up to $\left( -\overline{D}\right) _{\max }\approx 0.005$. \section{Conclusions} In this work, we have proposed a possibility to stabilize spatiotemporal solitons (``light bullets") in three-dimensional self-focusing Kerr media by means of the dispersion management (DM), which means that the local group-velocity dispersion coefficient alternates between positive and negative values along the propagation direction, $z$. Recently, it was shown that the DM alone can stabilize solitons in 2D (planar) waveguides, but in the bulk (3D) DM medium the ``bullets" are unstable. In this work, we have demonstrated that the complete stabilization can \ be provided if the longitudinal DM is combined with periodic modulation of the refractive index in one transverse direction ($y$), out of the two. The analysis was based on the variational approximation (systematic results of direct simulations will be reported in a follow-up paper). A stability area for the light bullets was identified in the model's parameter space. Its salient features are a necessary minimum strength of the transverse modulation of the refractive index, and minimum and maximum values $E_{\min ,\max }$ of the soliton's energy. The former feature can be accurately predicted (see Eq. (\ref{min})) in an analytical form from the evolution equation for the width of the soliton in the $y$-direction. The existence of $E_{\min }$, which vanishes when we assume zero average dispersion, can be explained in the same way as for the temporal solitons in DM optical fibers. Also, similar to the case of DM solitons in fibers, we find that the stability area extends to a region of \emph{normal} average dispersion \cite{DM}. On the other hand, the existence of $E_{\max }$ can be understood similarly to as it was recently done in the 2D counterpart of the present model (the strong transverse lattice can squeeze the system to a nearly 2D shape). The results presented in this work suggest a new approach to the challenging problem of the creation of 3D spatiotemporal optical solitons. The model also opens a way to address advanced issues, such as collisions between the LBs, and the existence and stability of solitons with different symmetries (for instance, LBs which are odd in the longitudinal and/or transverse directions). These issues will be considered elsewhere. \section{Acknowlegdements} M.M., M.T. and E.I. acknowledge support from KBN Grant No. 2P03 B4325. B.A.M. acknowledges the hospitality of the Physics Department and Soltan Institute for Nuclear Studies at the Warsaw University, and partial support from the Israel Science Foundation grant No. 8006/03. This author also appreciates the help of A. Desyatnikov in making Ref. \cite{Estoril} available on the internet.
{ "timestamp": "2005-03-08T12:03:11", "yymm": "0503", "arxiv_id": "physics/0503058", "language": "en", "url": "https://arxiv.org/abs/physics/0503058" }
\section{Notations} Let ${p}=e^{-\frac{\pi K'}{K}}$, $q=-e^{-\frac{\pi \lambda}{2K}}$ and $\zeta=e^{-\frac{\pi \lambda u}{2K}}$. We introduce $x$, $\tau$ and $r$ by $x=-q$, $\tau=\frac{2iK}{K'}$ and $r=\frac{K'}{\lambda}$. Then $p=e^{-\frac{2\pi i}{ \tau}}=x^{2r}$. Through this paper, we assume ${\rm Im}\tau >0$. Let ${\tilde{p}}=e^{2\pi i \tau}=e^{-\pi \frac{I'}{I}}$, where $I=\frac{K'}{2},\ I'=2K$. We use the theta functions \begin{eqnarray*} &&\vartheta_1(u|\tau)=2\tilde{p}^{1/8}(\tilde{p};\tilde{p})_\infty\sin\pi u \prod_{n=1}^\infty(1-2\tilde{p}^n\cos2\pi u+\tilde{p}^{2n}),\\ &&\vartheta_0(u|\tau)=-ie^{\pi i(u+\tau/4)}\vartheta_1\left(u+\frac{\tau}{2}\Big|\tau\right),\\ &&\vartheta_2(u|\tau)=\vartheta_1\left(u+\frac{1}{2}\Big|\tau\right),\\ &&\vartheta_3(u|\tau)=e^{\pi i(u+\tau/4)}\vartheta_1\left(u+\frac{\tau+1}{2}\Big|\tau\right) \end{eqnarray*} and Jacobi's elliptic functions \begin{eqnarray*} &&{\rm sn}\lambda u=\frac{\vtf{3}{0}{\tau}\vtf{1}{\frac{\lambda u}{2I}}{\tau}}{\vtf{2}{0}{\tau}\vtf{0}{\frac{\lambda u}{2I}}{\tau}} =\frac{\vtf{3}{0}{\tau}\vtf{1}{\frac{u}{r}}{\tau}}{\vtf{2}{0}{\tau}\vtf{0}{\frac{u}{r}}{\tau}},\\ &&{\rm cn}\lambda u=\frac{\vtf{0}{0}{\tau}\vtf{2}{\frac{\lambda u}{2I}}{\tau}}{\vtf{2}{0}{\tau}\vtf{0}{\frac{\lambda u}{2I}}{\tau}} =\frac{\vtf{0}{0}{\tau}\vtf{2}{\frac{u}{r}}{\tau}}{\vtf{2}{0}{\tau}\vtf{0}{\frac{u}{r}}{\tau}},\\ &&{\rm dn}\lambda u=\frac{\vtf{0}{0}{\tau}\vtf{3}{\frac{\lambda u}{2I}}{\tau}}{\vtf{3}{0}{\tau}\vtf{0}{\frac{\lambda u}{2I}}{\tau}} =\frac{\vtf{0}{0}{\tau}\vtf{3}{\frac{u}{r}}{\tau}}{\vtf{3}{0}{\tau}\vtf{0}{\frac{u}{r}}{\tau}}. \end{eqnarray*} We also use the symbol $[u]$ defined by \begin{eqnarray*} &&[u]=x^{\frac{u^2}{r}-u}\Theta_{x^{2r}}(x^{2u})=C\vt{1}{u}{\tau}, \quad C=x^{-\frac{r}{4}}e^{-\frac{\pi i}{4}}\tau^{\frac{1}{2}} \end{eqnarray*} and abbreviation \begin{eqnarray*} &&\vth{1,2}{u}=\vth{1}{u}\vth{2}{u}=\vtf{0}{0}{\tau}\vt{1}{u}{\tau}, \end{eqnarray*} etc.. \section{Fusion of Baxter's $R$-matrix} Baxter's elliptic $R$-matrix is given by\cite{Baxter} \begin{eqnarray} &&{R}(u)= {R}_0(u)\left(\matrix{a(u)&&&d(u)\cr &b(u)&c(u)&\cr &c(u)&b(u)&\cr d(u)&&&a(u)\cr}\right), \end{eqnarray} where \begin{eqnarray} R_0(u)&=&z^{-\frac{r-1}{2r}}\frac{(px^2z;x^4,p)_\infty(x^2z;x^4,p)_\infty (p/z;x^4,p)_\infty(x^4/z;x^4,p)_\infty}{(px^2/z;x^4,p)_\infty (x^2/z;x^4,p)_\infty(pz;x^4,p)_\infty(x^4z;x^4,p)_\infty},\lb{evR}\\ a(u)&=&\frac{\vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}}\vtf{2}{\frac{ u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}\vtf{2}{\frac{1+u}{2r}}{\frac{\tau}{2}}},\qquad b(u)=\frac{\vtf{2}{\frac{1 }{2r}}{\frac{\tau}{2}}\vtf{1}{\frac{ u}{2r}}{\frac{\tau}{2}}} {\vtf{2}{0}{\frac{\tau}{2}}\vtf{1}{\frac{1+u}{2r}}{\frac{\tau}{2}}},\\ c(u)&=&\frac{\vtf{1}{\frac{ 1}{2r}}{\frac{\tau}{2}}\vtf{2}{\frac{ u}{2r}}{\frac{\tau}{2}}} {\vtf{2}{0}{\frac{\tau}{2}}\vtf{1}{\frac{1+u}{2r}}{\frac{\tau}{2}}},\qquad d(u)=-\frac{\vtf{1}{\frac{ 1}{2r}}{\frac{\tau}{2}}\vtf{1}{\frac{ u}{2r}}{\frac{\tau}{2}}} {\vtf{2}{0}{\frac{\tau}{2}}\vtf{2}{\frac{1+u}{2r}}{\frac{\tau}{2}}} \end{eqnarray} with $z=\zeta^2=x^{2u}$. Let $V=\C v_{\varepsilon_1}\oplus \C v_{\varepsilon_2},\ \varepsilon_1,\varepsilon_2=+,-$. We regard $R(u)\in {\rm End}(V\otimes V)$. The $R$-matrix \eqref{evR} satisfies \begin{eqnarray} &&R(u)PR(u)P=\hbox{id},\\ &&R(-u-1)=(\sigma^{y }\otimes 1)^{-1}\ (PR(u)P)^{t_1} \ \sigma^y\otimes 1,\\ &&R(0)=P, \qquad \lim_{u\to -1}R(u)=P-\hbox{id}. \end{eqnarray} Here ${^{t_1}}$ denotes the transposition with respect to the first vector space in the tensor product and $P(\varepsilon_1\otimes \varepsilon_2)=\varepsilon_2\otimes \varepsilon_1$. Fusion of $R(u)$ was considered systematically in \cite{DJKMO}. Let $V^{(2)}$ be the space of the symmetric tensors in $V\otimes V$ spanned by $v^{(2)}_{2}\equiv v_+\otimes v_+,\ v^{(2)}_{0}\equiv \frac{1}{2}(v_+\otimes v_-+v_-\otimes v_+),\ v^{(2)}_{-2}\equiv v_-\otimes v_-$.The projection operator of the space $V\otimes V$ on $V^{(2)}$ is given by $\Pi=\frac{1}{2}(P+\hbox{id})$. Let $V_1, V_2, V_{{\bar{1}}}, V_{{\bar{2}}}$ be the copies of V. Define \begin{eqnarray} &&R^{(2,1)}_{12, \bar{j}}(u)=\Pi_{12}R_{1{\bar{j}}}(u+1)R_{2{\bar{j}}}(u) \ \in {\rm End}(V^{(2)} \otimes V_{{\bar{j}}}). \lb{hfusion} \end{eqnarray} It follows that \begin{eqnarray} &&R^{(2,1)}_{12,{\bar{j}}}(u)\Pi_{12}=R^{(2,1)}_{12,{\bar{j}}}(u). \end{eqnarray} The 2$\times$2 fusion of the $R$-matrix is then given by \begin{eqnarray} &&R^{(2,2)}(u)=\Pi_{{\bar{1}}{\bar{2}}}R^{(2,1)}_{12,{\bar{2}}}(u)R^{(2,1)}_{12,{\bar{1}}}(u-1) \ \in {\rm End}(V^{(2)} \otimes V^{(2)}).\lb{vhfusion} \end{eqnarray} This satisfies the Yang-Baxter equation on $V^{(2)} \otimes V^{(2)} \otimes V^{(2)}$. We calculate the matrix elements of $R^{(2,1)}(u)$ and $R^{(2,2)}(u)$ defined by \begin{eqnarray} &&R^{(2,1)}(u) v^{(2)}_{\mu}\otimes v^{}_{\varepsilon}=\sum_{\mu'=2,0,-2 \atop \varepsilon'=+,-} R^{(2,1)}(u)^{\mu \varepsilon}_{\mu' \varepsilon'}\ v^{(2)}_{\mu'}\otimes v^{}_{\varepsilon'},\\ &&R^{(2,2)}(u) v^{(2)}_{\mu_1}\otimes v^{(2)}_{\mu_2} =\sum_{\mu'_1,\mu'_2=2,0,-2} R^{(2,2)}(u)^{\mu_1 \mu_2}_{\mu'_1 \mu'_2}\ v^{(2)}_{\mu'_1}\otimes v^{(2)}_{\mu'_2}. \end{eqnarray} From \eqref{hfusion}, we have \begin{eqnarray} &&R^{(2,1)}(u)^{\mu\ \bar{\varepsilon}}_{\mu'\ \bar{\varepsilon}'}=\sum_{\varepsilon_2',\bar{\varepsilon}''=\pm 1}R(u+1)^{\mu-\varepsilon_2\ \bar{\varepsilon}''}_{\mu'-\varepsilon_2'\ \bar{\varepsilon}'}R(u)^{\varepsilon_2\ \bar{\varepsilon}}_{\varepsilon_2'\ \bar{\varepsilon}''}, \end{eqnarray} where we set $\mu=\varepsilon_1+\varepsilon_2,\ \mu'=\varepsilon_1'+\varepsilon_2'$. Evaluating the summation explicitly, we obtain \begin{prop} \begin{eqnarray*} &&R^{(2,1)}(u)=R^{(2,1)}_0(u)\left(\matrix{R^{+2+}_{+2+}&0&0&R^{\ 0-}_{+2+}&R^{-2+}_{+2+}&0\cr 0&R^{+2-}_{+2-}&R^{\ 0+}_{+2-}&0&0&R^{-2-}_{+2-}\cr 0&R^{+2-}_{\ 0+}&R^{0+}_{0+}&0&0&R^{-2-}_{\ 0+}\cr R^{+2+}_{\ 0-}&0&0&R^{\ 0-}_{\ 0-}&R^{-2+}_{\ 0-}&0\cr R^{+2+}_{-2+}&0&0&R^{\ 0-}_{-2+}&R^{-2+}_{-2+}&0\cr 0&R^{+2-}_{-2-}&R^{\ 0+}_{-2-}&0&0&R^{-2-}_{-2-}\cr}\right), \end{eqnarray*} where \begin{eqnarray*} &&R^{(2,1)}_0(u)=R_0(u+1)R_0(u)=-\frac{[u+1]}{[u]},\\ &&R(u)^{+2+}_{+2+}=R(u)^{-2-}_{-2-}=\frac{\vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}}^2\vtf{2}{\frac{u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}^2\vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{\ 0-}_{+2+}=R(u)^{\ 0+}_{-2-} =-\frac{\vtf{1}{\frac{1}{2r}}{\frac{\tau}{2}}\vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}} \vtf{1}{\frac{u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}^2\vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{-2+}_{+2+}=R(u)^{+2-}_{-2-}=-\frac{\vtf{1}{\frac{1}{2r}}{\frac{\tau}{2}}^2\vtf{2}{\frac{u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}^2\vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{+2-}_{+2-}=R(u)^{-2+}_{-2+}=\frac{\vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}}^2\vtf{1}{\frac{u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}^2\vtf{1}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{\ 0+}_{+2-}=R(u)^{\ 0-}_{-2+}=\frac{\vtf{1}{\frac{1}{2r}}{\frac{\tau}{2}}\vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}} \vtf{2}{\frac{u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}^2\vtf{1}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{-2-}_{+2-}=R(u)^{+2+}_{-2+}=- \frac{\vtf{1}{\frac{1}{2r}}{\frac{\tau}{2}}^2\vtf{1}{\frac{u}{2r}}{\frac{\tau}{2}}}{\vtf{2}{0}{\frac{\tau}{2}}^2\vtf{1}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{+2-}_{0\ +}=R(u)^{-2+}_{0\ -}= \frac{\vtf{1}{\frac{1}{r}}{\frac{\tau}{2}}\vtf{2}{\frac{u+1}{2r}}{\frac{\tau}{2}}^2} {\vtf{2}{0}{\frac{\tau}{2}} \vtf{1}{\frac{u+2}{2r}}{\frac{\tau}{2}}\vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{+2+}_{0\ -}=R(u)^{-2-}_{0\ +}=- \frac{\vtf{1}{\frac{1}{r}}{\frac{\tau}{2}}\vtf{1}{\frac{u+1}{2r}}{\frac{\tau}{2}}^2} {\vtf{2}{0}{\frac{\tau}{2}} \vtf{1}{\frac{u+2}{2r}}{\frac{\tau}{2}}\vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}}},\\ &&R(u)^{0\ +}_{0\ +}=R(u)^{0\ -}_{0\ -}= \frac{\vtf{2}{\frac{1}{r}}{\frac{\tau}{2}}\vth{1,2}{u+1}} {\vtf{2}{0}{\frac{\tau}{2}}\vth{1,2}{u+2}}. \end{eqnarray*} \end{prop} Similarly, from \eqref{vhfusion}, we obtain \begin{eqnarray*} &&R^{(2,2)}(u)^{\mu_1\ \mu_2}_{\mu_1'\ \mu_2'}=\sum_{\mu''=0, \pm2 \atop \bar{\varepsilon}_1'=\pm1}R^{(2,1)}(u)^{ \mu''\ {\mu}_2-\bar{\varepsilon}_1}_{\mu_1'\ {\mu}_2'-\bar{\varepsilon}_1'}R^{(2,1)}(u-1)^{ \mu_1\ \bar{\varepsilon}_1}_{\mu''\ \bar{\varepsilon}_1'}, \end{eqnarray*} where $\mu_1=\varepsilon_1+\varepsilon_2,\ \mu_1'=\varepsilon_1'+\varepsilon_2'$ and $\mu_2=\bar{\varepsilon}_{{1}}+\bar{\varepsilon}_{{2}},\ \mu_2'=\bar{\varepsilon}_{{1}}' +\bar{\varepsilon}_{{2}}'$. \begin{prop} \begin{eqnarray*} &&R^{(2,2)}(u)=R^{(2,2)}_0(u)\left(\matrix{ G&0&A&0&B&0&A&0&H\cr 0&F&0&C&0&C^*&0&D&0\cr A^*&0&G^*&0&B^*&0&H^*&0&A^*\cr 0&C&0&F&0&D&0&C^*&0\cr I&0&I^*&0&E&0&I^*&0&I\cr 0&C^*&0&D&0&F&0&C&0\cr A^*&0&H^*&0&B^*&0&G^*&0&A^*\cr 0&D&0&C^*&0&C&0&F&0\cr H&0&A&0&B&0&A&0&G\cr }\right), \end{eqnarray*} where \begin{eqnarray*} R^{(2,2)}_0(u)&=&R^{(2,1)}_0(u-1)R^{(2,1)}_0(u)=\frac{[u+1]}{[u-1]},\\ A&=&R(u)^{+2-2}_{+2+2}=R(u)^{-2+2}_{-2-2}=R(u)^{-2+2}_{+2+2}=R(u)^{+2-2}_{-2-2}{\nonumber}\\ &=&- \frac{ \vtf{1}{\frac{1}{r}}{\frac{\tau}{2}} \vtf{2}{\frac{u}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{1}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u}{2r}}{\frac{\tau}{2}} }{ \vtf{2}{0}{\frac{\tau}{2}}^3 \vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u+1}{2r}}{\frac{\tau}{2}} },\\ B&=&R(u)^{\ 0\ 0}_{+2+2}=R(u)^{\ 0\ 0}_{-2-2}=- \frac{ \vtf{2}{\frac{1}{r}}{\frac{\tau}{2}} \vtf{1}{\frac{u}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{1}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u}{2r}}{\frac{\tau}{2}} }{ \vtf{2}{0}{\frac{\tau}{2}}^3 \vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u+1}{2r}}{\frac{\tau}{2}} },\\ C&=&R(u)^{+2\ 0}_{\ 0+2}=R(u)^{-2\ 0}_{\ 0-2}=R(u)^{\ 0+2}_{+2\ 0}=R(u)^{\ 0-2}_{-2\ 0}= \frac{ \vtf{2}{\frac{u}{2r}}{\frac{\tau}{2}}^2 \vtf{1,2}{\frac{1}{r}}{\frac{\tau}{2}} } { \vtf{2}{0}{\frac{\tau}{2}}^2 \vth{1,2}{u+2} },\\ D&=&R(u)^{+2\ 0}_{-2\ 0 }=R(u)^{-2\ 0}_{+2\ 0 }=R(u)^{\ 0+2}_{\ 0-2}=R(u)^{\ 0 -2}_{\ 0+2} =-\frac{ \vtf{1}{\frac{1}{r}}{\frac{\tau}{2}}^2 \vth{1,2}{u} } { \vtf{2}{0}{\frac{\tau}{2}}^2 \vth{1,2}{u+2} },\\ E&=&R(u)^{\ 0\ 0 }_{\ 0\ 0 }{\nonumber}\\ &=&\frac{ \vtf{1}{\frac{1}{r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{1}{2r}}{\frac{\tau}{2}} \left( \vtf{2}{\frac{u+1}{2r}}{\frac{\tau}{2}}^3 \vtf{2}{\frac{u-1}{2r}}{\frac{\tau}{2}}+ \vtf{1}{\frac{u+1}{2r}}{\frac{\tau}{2}}^3 \vtf{1}{\frac{u-1}{2r}}{\frac{\tau}{2}} \right) }{ \vtf{2}{0}{\frac{\tau}{2}}^3 \vtf{1,2}{\frac{u+2}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u+1}{2r}}{\frac{\tau}{2}}}{\nonumber}\\ &&+\frac{ \vtf{2}{\frac{1}{r}}{\frac{\tau}{2}}^2\vth{1,2}{u} }{ \vtf{2}{0}{\frac{\tau}{2}}^2\vth{1,2}{u+2} } ,\\ F&=&R(u)^{\ 0+2}_{\ 0 +2}=R(u)^{\ 0-2}_{\ 0 -2}=R(u)^{+2\ 0}_{+2\ 0}=R(u)^{-2\ 0}_{-2\ 0} =\frac{ \vtf{2}{\frac{1}{r}}{\frac{\tau}{2}}^2 \vth{1,2}{u} } { \vtf{2}{0}{\frac{\tau}{2}}^2 \vth{1,2}{u+2} }, \end{eqnarray*} \begin{eqnarray*} G&=&R(u)^{+2+2}_{+2+2}=R(u)^{-2-2}_{-2-2}{\nonumber}\\ &=& \frac{ \vtf{2}{\frac{u}{2r}}{\frac{\tau}{2}} \left(\vtf{1}{\frac{1}{2r}}{\frac{\tau}{2}}^4\vtf{1}{\frac{u-1}{2r}}{\frac{\tau}{2}} \vtf{2}{\frac{u+1}{2r}}{\frac{\tau}{2}}+ \vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}}^4\vtf{2}{\frac{u-1}{2r}}{\frac{\tau}{2}} \vtf{1}{\frac{u+1}{2r}}{\frac{\tau}{2}} \right) } { \vtf{2}{0}{\frac{\tau}{2}}^4 \vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u+1}{2r}}{\frac{\tau}{2}} } , \end{eqnarray*} \begin{eqnarray*} H&=&R(u)^{+2+2}_{-2-2}=R(u)^{-2-2}_{+2+2}= \frac{ \vtf{1}{\frac{1}{r}}{\frac{\tau}{2}} \vtf{1}{\frac{u}{2r}}{\frac{\tau}{2}}^3 \vth{1,2}{1} }{ \vtf{2}{0}{\frac{\tau}{2}}^3 \vtf{2}{\frac{u+2}{2r}}{\frac{\tau}{2}} \vth{1,2}{u+1} }, \end{eqnarray*} \begin{eqnarray*} I&=&R(u)^{+2+2}_{\ 0\ 0 }=R(u)^{-2-2}_{\ 0\ 0}{\nonumber}\\ &=&-\frac{ \vtf{1}{\frac{1}{r}}{\frac{\tau}{2}} \left( \vtf{2}{\frac{1}{2r}}{\frac{\tau}{2}}^2 \vtf{1}{\frac{u+1}{2r}}{\frac{\tau}{2}}^3 \vtf{2}{\frac{u-1}{2r}}{\frac{\tau}{2}}+ \vtf{1}{\frac{1}{2r}}{\frac{\tau}{2}}^2 \vtf{2}{\frac{u+1}{2r}}{\frac{\tau}{2}}^3 \vtf{1}{\frac{u-1}{2r}}{\frac{\tau}{2}} \right) }{ \vtf{2}{0}{\frac{\tau}{2}}^3 \vtf{1,2}{\frac{u+2}{2r}}{\frac{\tau}{2}} \vtf{1,2}{\frac{u+1}{2r}}{\frac{\tau}{2}}}{\nonumber}\\ &&-\frac{ \vtf{1}{\frac{u}{2r}}{\frac{\tau}{2}}^2\vtf{1,2}{\frac{1}{r}}{\frac{\tau}{2}} }{ \vtf{2}{0}{\frac{\tau}{2}}^2\vth{1,2}{u+2} }. \end{eqnarray*} The $*$-ed matrix element is obtained from a corresponding non-$*$-ed element by replacing the theta functions only depending on $u$ in the following rule. \begin{eqnarray*} &&\vartheta_1 \to -\vartheta_2,\quad \vartheta_2 \to \vartheta_1. \end{eqnarray*} \end{prop} From this expression, one can easily see the following symmetries. \begin{eqnarray*} &&R^{(2,2)}(u)^{\varepsilon_1 \varepsilon_2}_{\varepsilon_1' \varepsilon_2'}=R^{(2,2)}(u)^{\varepsilon_2 \varepsilon_1}_{\varepsilon_2' \varepsilon_1'}, \qquad\qquad ( P{\rm-invariance})\\ &&R^{(2,2)}(u)^{\varepsilon_1 \varepsilon_2}_{\varepsilon_1' \varepsilon_2'}=R^{(2,2)}(u)^{-\varepsilon_1 -\varepsilon_2}_{-\varepsilon_1' -\varepsilon_2'} \qquad\qquad ({\rm \Z_2-symmetry}). \end{eqnarray*} In 1980, Fateev proposed the 21-vertex model as the spin one extension of Baxter's eight vertex model\cite{Fateev}. Solving the Yang-Baxter equation, he obtained the following $R$-matrix. \begin{eqnarray*} &&R_{F}(u)=\tilde{F}(u)\left(\matrix{ s_1&0&0&0&\mu&0&0&0&\nu\cr 0&t&0&r&0&0&0&0&0\cr 0&0&T&0&0&0&R&0&0\cr 0&r&0&t&0&0&0&0&0\cr \mu&0&0&0&s_2&0&0&0&\rho\cr 0&0&0&0&0&a&0&q&0\cr 0&0&R&0&0&0&T&0&0\cr 0&0&0&0&0&q&0&a&0\cr \nu&0&0&0&\rho&0&0&0&s_3\cr } \right) \end{eqnarray*} where \begin{eqnarray*} &&s_1=\cn2\lambda+\frac{{\rm sn}\lambda\ \sn2\lambda}{{\rm sn}\lambda u\ {\rm sn}\lambda (u+1)},\\ &&s_2=\cn2\lambda+\dn2\lambda-1+\frac{{\rm sn}\lambda\ \sn2\lambda}{{\rm sn}\lambda u\ {\rm sn}\lambda( u+1)},\\ &&s_3=\dn2\lambda+\frac{{\rm sn}\lambda\ \sn2\lambda}{{\rm sn}\lambda u\ {\rm sn}\lambda( u+1)},\\ &&T=1,\qquad\qquad t=\cn2\lambda,\qquad\qquad a=\dn2\lambda,\\ &&r=\frac{{\rm cn}\lambda u\ \sn2\lambda}{{\rm sn}\lambda u},\qquad\qquad \mu=-\frac{{\rm cn}\lambda( u+1)\ \sn2\lambda}{{\rm sn}\lambda( u+1)},\\ &&R=\frac{ \sn2\lambda}{{\rm sn}\lambda u},\qquad\qquad \nu=-\frac{ \sn2\lambda}{{\rm sn}\lambda( u+1)},\\ &&q=\frac{{\rm dn}\lambda u\ \sn2\lambda}{{\rm sn}\lambda u},\qquad\qquad \rho=-\frac{{\rm dn}\lambda( u+1)\ \sn2\lambda}{{\rm sn}\lambda( u+1)} \end{eqnarray*} and $\tilde{F}(u)$ satisfies \begin{eqnarray*} &&\tilde{F}(u)=\tilde{F}(-u-1),\\ &&\tilde{F}(u)\tilde{F}(-u)=\frac{{\rm sn}^2\lambda u}{{\rm sn}^2\lambda u-{\rm sn}^22\lambda}. \end{eqnarray*} $R_{F}(u)$ has the following symmetries. \begin{eqnarray*} &&R_{F}(u)^{ij}_{kl}=R_{F}(u)^{ji}_{lk} \qquad\qquad ( P{\rm-invariance})\\ &&R_{F}(u)^{ij}_{kl}=R_{F}(u)_{ij}^{kl},\qquad\qquad (T{\rm-invariance})\\ &&R_{F}(u)^{ij}_{kl}=R_{F}(-u-1)_{il}^{kj}\qquad\qquad ({\rm Crossing\ symmetry}). \end{eqnarray*} We find \begin{prop} \begin{eqnarray*} &&\tilde{F}(u)=R^{(2,2)}_0(u)\frac{\vt{0}{2}{\tau}\vt{1}{u}{\tau}}{\vtf{0}{0}{\tau}\vt{1}{u+2}{\tau}}. \end{eqnarray*} \end{prop} Then the following theorem is essentially due to Jimbo\cite{Jimbo}. \begin{thm} $R^{(2,2)}(u)$ is gauge equivalent to $R_{F}(u)$. Namely, \begin{eqnarray*} &&R_{F}(u)=U\otimes U \ R^{(2,2)}(u)\ (U\otimes U)^{-1}, \end{eqnarray*} where \begin{eqnarray*} &&U=\left(\matrix{1&0&0\cr 0&x&0\cr 0&0&y\cr}\right)\left(\matrix{\frac{1}{\sqrt{2}}&0&\frac{1}{\sqrt{2}}\cr 0&1&0\cr \frac{1}{\sqrt{2}}&0&-\frac{1}{\sqrt{2}}\cr}\right), \end{eqnarray*} and \begin{eqnarray*} &&x^2={-\frac{1}{2}\frac{\vtf{0}{0}{\tau}\vt{3}{1}{\tau}}{\vt{0}{1}{\tau}\vtf{3}{0}{\tau}}},\quad y^2={-\frac{\vtf{2}{0}{\tau}\vt{3}{1}{\tau}}{\vt{2}{1}{\tau}\vtf{3}{0}{\tau}}}. \end{eqnarray*} \end{thm} Combining the crossing symmetry of $R_{F}(u)$ and the $P$-invariance of $R^{(2,2)}(u)$, we find the following crossing symmetry formula for $R^{(2,2)}(u)$. \begin{cor} \begin{eqnarray} &&R^{(2,2)}(-u-1)=Q^{-1}\otimes 1\ (P^{(2)}R^{(2,2)}(u)P^{(2)})^{t_1}\ Q\otimes 1,\lb{crossing} \end{eqnarray} where \begin{eqnarray*} &&Q=U^t U=\frac{1}{2}\left(\matrix{1+y^2&0&1-y^2\cr 0&x^2&0\cr 1-y^2&0&1+y^2\cr}\right) \end{eqnarray*} and $P^{(2)}$ is the permutation operator $P^{(2)}( v^{(2)}_{\varepsilon_1}\otimes v^{(2)}_{\varepsilon_2})= v^{(2)}_{\varepsilon_2}\otimes v^{(2)}_{\varepsilon_1}$. \end{cor} \noindent {\it Remark :}\ \noindent The crossing symmetry of the elliptic $R$-matrix is related to the dual module of the finite dimensional module of the elliptic algebra ${\cal A}_{q,p}(\widehat{\goth{sl}}_2)$, or the module of the Sklyanin algebra. See \cite{JM} for the case $U_q(\widehat{\goth{sl}}_2)$. Let $V_\zeta$ be the 3-dimensional module of ${\cal A}_{q,p}(\widehat{\goth{sl}}_2)$, and $V_\zeta^*$ its dual module. The above $Q$-matrix gives an isomorphism between $V_\zeta$ and $V_\zeta^*$. \section{The Vertex-Face Correspondence } The vertex-face correspondence is a relationship between Baxter's $R$-matrix and the SOS face weight $W\BW{a_1}{a_2}{a_4}{a_3}{u}$ given by \begin{eqnarray} &&{W}\left(\left. \begin{array}{cc} n&n\pm 1\\ n\pm1&n\pm2 \end{array}\right|u\right)={R}_0(u),{\nonumber}\\ &&{W}\left(\left. \begin{array}{cc} n&n\pm 1\\ n\pm1&n \end{array}\right|u\right)={R}_0(u)\frac{[n\mp u][1]}{[ n][ 1+u]},\lb{face}\\ &&{W}\left(\left. \begin{array}{cc} n&n\pm 1\\ n\mp 1&n \end{array}\right|u\right)= {R}_0(u) \frac{[ n\pm 1][u] }{[ n][1+u]}.{\nonumber} \end{eqnarray} Let us consider the following vectors in $V$ \begin{eqnarray} &&\psi(u)^a_b=\psi_+(u)^a_b\ v_+ + \psi_-(u)^a_b\ v_-,\qquad \\ &&{\psi}_+(u)^a_b=\vtf{0}{\frac{(a-b)u+a}{2r}}{\frac{\tau}{2}},\qquad \psi_-(u)=\vtf{3}{\frac{(a-b)u+a}{2r}}{\frac{\tau}{2}} \lb{intertwinvec}{\nonumber} \end{eqnarray} with $|a-b|=1$. Baxter showed the following identity\cite{Baxter}. \begin{eqnarray} &&\sum_{\varepsilon_1',\varepsilon_2'} R(u-v)_{\varepsilon_1 \varepsilon_2}^{\varepsilon_1' \varepsilon_2'}\ \psi_{\varepsilon_1'}(u)_{b}^{a} \psi_{\varepsilon_2'}(v)_{c}^{b} =\sum_{b' \in {\mathbb{Z}}} \psi_{\varepsilon_2}(v)_{b'}^{a} \psi_{\varepsilon_1}(u)_{c}^{b'} W\left(\left. \begin{array}{cc} a&b\\ b'&c \end{array}\right|u-v\right).\lb{vertexface} \end{eqnarray} \subsection{Fusion} Following Date et al.\cite{DJKMO}, we consider the fusion of the Vertex-Face correspondence relation \eqref{vertexface}. The fusion of the SOS weights is briefly summarized as follows. The SOS weight \eqref{face} satisfies \begin{eqnarray} &&W\BW{a}{b}{d}{c}{0}=\delta_{b,d},\\ &&W\BW{a}{b}{d}{c}{-1}=0\qquad {\rm if}\ |a-c|=2,\\ &&W\BW{a}{a\pm 1}{a\pm 1}{a}{-1}=-W\BW{a}{a\pm 1}{a\mp 1}{a}{-1}. \end{eqnarray} Then if one defines \begin{eqnarray} &&W_{21}\BW{a}{b}{d}{c}{u}=\sum_{d'}W\BW{a}{a'}{d}{d'}{u+1}W\BW{a'}{b}{d'}{c}{u}, \lb{fhfusion} \end{eqnarray} one can verify the following statements. (i) The RHS of \eqref{fhfusion} is independent of the choice of $a'$ provided $|a-a'|=|a'-b|=1$. (ii) For all $a, b, c, d$, \begin{eqnarray*} &&W_{21}\BW{a}{b}{d}{c}{-1}=0. \end{eqnarray*} The $2\times 2$ fusion of the SOS weight is then given by the formula \begin{eqnarray} &&W_{22}\BW{a}{b}{d}{c}{u}=\sum_{a'}W_{21}\BW{a}{b}{a'}{b'}{u-1}W_{21}\BW{a'}{b'}{d}{c}{u} \lb{fvhfusion}. \end{eqnarray} Here the RHS is independent of the choice of $b'$ provided $|b-b'|=|b'-c|=1$. Now the dynamical variables $a, b, c, d$ satisfies the extended admissible condition; $a_j-a_k\in \{2, 0, -2\}$ for any two adjacent local heights $a_j, a_k$. Furthermore the resultant SOS weight $W_{22}$ satisfies the face type YBE and defines the $2\times 2$ fusion SOS model. Explicit expressions of $W_{22}\BW{a}{b}{d}{c}{u}$ are given, for example, in \cite{KKW}. It satisfies the unitarity and crossing symmetry relations \begin{eqnarray} &&\sum_{s}W_{22}\BW{a}{s}{d}{c}{-u} W_{22}\BW{a}{b}{s}{c}{u}=\delta_{b,d},\label{SOSunitarity}\\ &&W_{22}\BW{d}{c}{a}{b}{u}=\frac{(b,c)_2 \ g_a g_c}{(a,d)_2\ g_b g_d} \, W_{22}\BW{a}{d}{b}{c}{-1-u}. \label{SOScrossing} \end{eqnarray} Here $g_a=\varepsilon_a\sqrt{[a]}$\, $\varepsilon_a=\pm1,\ \varepsilon_a\varepsilon_{a+1}=(-)^a$ and \begin{eqnarray*} &&(a,b)_M=(b,a)_M=\left[\matrix{M\cr \frac{a-b+M}{2}\cr}\right]^{-1} \frac{\left[\frac{a+b-M}{2}, \frac{a+b+M}{2} \right]}{\sqrt{[a][b]}},\\ &&\left[\matrix{A\cr B\cr}\right]=\frac{[A][A-1]\cdots [A-B+1]}{[B][B-1]\cdots [1]},\\ &&[A,B]=[A][A+1]\cdots [B] \quad (A<B),\qquad [A,A-1]=1. \end{eqnarray*} The fusion of the intertwining vectors is given by \begin{eqnarray} &&\psi^{(2)}(u)^a_b=\Pi\ \psi(u+1)^a_c\otimes \psi(u)^c_b \ \in V\otimes V.\lb{fusionpsi} \end{eqnarray} The RHS is independent of the choice of $c$ provided $|a-c|=|c-b|=1$. Then using \eqref{vertexface}, \eqref{hfusion}, \eqref{vhfusion}, \eqref{fhfusion}, \eqref{fvhfusion} and \eqref{fusionpsi}, pne can show \begin{eqnarray} &&\sum_{\mu_1',\mu_2'} R^{(2,2)}(u-v)_{\mu_1 \mu_2}^{\mu_1' \mu_2'}\ \psi^{(2)}_{\mu_1'}(u)_{b}^{a} \psi^{(2)}_{\mu_2'}(v)_{c}^{b} =\sum_{b' \in {\mathbb{Z}}} \psi^{(2)}_{\mu_2}(v)_{b'}^{a} \psi^{(2)}_{\mu_1}(u)_{c}^{b'} W_{22}\left(\left. \begin{array}{cc} a&b\\ b'&c \end{array}\right|u-v\right).{\nonumber}\\ &&\lb{fusionvertexface} \end{eqnarray} Explicitly, the vector $\psi^{(2)}(u)^a_b$ is calculated as follows\cite{KKW}. \begin{prop} \begin{eqnarray*} \left(\begin{array}{c} \psi^{(2)}_2(u)_{n+2}^n\\ \psi^{(2)}_0(u)_{n+2}^n\\ \psi^{(2)}_{-2}(u)_{n+2}^n \end{array}\right)= \left(\begin{array}{c} \vartheta_0\left(\left. \frac{u-n+1 }{2r} \right|\frac{\tau}{2} \right) \vartheta_0 \left( \left.\frac{u-n-1}{2r}\right|\frac{\tau}{2} \right)\\ {2} \vartheta_0\left(\left. \frac{u-n}{r}\right| \tau\right) \vartheta_0\left( \left.\frac{1}{r}\right|\tau \right) \\ \vartheta_3\left(\left. \frac{u-n+1 }{2r} \right|\frac{\tau}{2} \right) \vartheta_3 \left( \left.\frac{u-n-1}{2r}\right|\frac{\tau}{2} \right) \end{array}\right), \end{eqnarray*} \begin{eqnarray*} \left(\begin{array}{c} \psi^{(2)}_2(u)_{n}^n\\ \psi^{(2)}_0(u)_{n}^n\\ \psi^{(2)}_{-2}(u)_{n}^n \end{array}\right)= \left(\begin{array}{c} \vartheta_0\left(\left. \frac{u-n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_0\left( \left.\frac{u+n+1}{2r}\right|\frac{\tau}{2} \right)\\ {2} \vartheta_0\left(\left. \frac{n}{r}\right| \tau \right) \vartheta_0 \left( \left.\frac{u+1}{r} \right| \tau \right)\\ \vartheta_3\left(\left. \frac{u-n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_3\left( \left.\frac{u+n+1}{2r}\right|\frac{\tau}{2} \right) \end{array} \right), \end{eqnarray*} \begin{eqnarray*} \left(\begin{array}{c} \psi^{(2)}_2(u)_{n-2}^n\\ \psi^{(2)}_0(u)_{n-2}^n\\ \psi^{(2)}_{-2}(u)_{n-2}^n \end{array}\right)= \left(\begin{array}{c} \vartheta_0\left(\left. \frac{u+n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_0\left( \left.\frac{u+n-1}{2r}\right|\frac{\tau}{2} \right) \\ {2}\vartheta_0\left(\left. \frac{u+n}{r}\right| \tau \right) \vartheta_0\left( \left.\frac{1}{r}\right| \tau \right) \\ \vartheta_3\left(\left. \frac{u+n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_3\left( \left.\frac{u+n-1}{2r}\right|\frac{\tau}{2} \right)\end{array}\right). \end{eqnarray*} \end{prop} \subsection{The dual intertwining vectors and their fusion} Let us consider the dual vector $\psi^*(u)^a_b$ defined by \begin{eqnarray} &&\psi^*(u)^a_b\ v_{\varepsilon}= \psi^*_\varepsilon(u)^a_b,\qquad \psi^*_\varepsilon(u)^a_b=-\varepsilon\frac{a-b}{2[b][u]}C^2\ \psi_{-\varepsilon}(u-1)^a_b \lb{dualintvec} \end{eqnarray} with $|a-b|=1$. By a direct calculation, we verify the inversion relations \begin{eqnarray} &&\sum_{\varepsilon=\pm}\psi_\varepsilon^*(u)^a_b\psi_\varepsilon(u)^b_c=\delta_{a,c},\lb{inversiona}\\ &&\sum_{a=b\pm1}\psi_{\varepsilon'}^*(u)^a_b\psi_\varepsilon(u)^b_a=\delta_{\varepsilon',\varepsilon}.\lb{inversionb} \end{eqnarray} Hence we call $\psi^*(u)^a_b$ the dual intertwining vector. From the crossing symmetry properties of $R$ and $W$ the following vertex-face correspondence is held. \begin{eqnarray} &&\sum_{\varepsilon_1',\varepsilon_2'} R(u-v)^{\varepsilon_1 \varepsilon_2}_{\varepsilon_1' \varepsilon_2'}\ \psi^*_{\varepsilon_1'}(u)_{b}^{a} \psi^*_{\varepsilon_2'}(v)_{c}^{b} =\sum_{s \in {\mathbb{Z}}} \psi^*_{\varepsilon_2}(v)_{b'}^{a} \psi^*_{\varepsilon_1}(u)_{c}^{b'} W\left(\left. \begin{array}{cc} c&b'\\ b&a \end{array}\right|u-v\right).\lb{vertexfacedual} \end{eqnarray} The fusion of the dual intertwining vectors is given by\cite{KKW} \begin{eqnarray} \psi^{*(2)}(u)_a^b= \sum_{c=a\pm 1} \psi^*(u+1)_a^c \otimes \psi^*(u)_c^b. \lb{fusiondualint} \end{eqnarray} Then $\psi^{*(2)}(u)_a^b$ satisfies \begin{eqnarray} &&\Pi\ \psi^{*(2)}(u)_a^b=\psi^{*(2)}(u)_a^b\ \Pi. \lb{prdualint} \end{eqnarray} In the components, \eqref{fusiondualint} yields \begin{eqnarray} \psi^{*(2)}_{\mu}(u)_a^b= \sum_{c=a\pm 1} \psi^*_{\varepsilon_1}(u+1)_a^c \psi^*_{\varepsilon_2}(u)_c^b. \lb{compfusiondualint} \end{eqnarray} The relation \eqref{prdualint} indicates that the RHS of \eqref{compfusiondualint}is independent of the choice of $\varepsilon_1, \varepsilon_2$ proivided $\mu=\varepsilon_1+\varepsilon_2$. Then using \eqref{inversiona} and \eqref{inversionb}, it is easy to verify the following inversion relations. \begin{prop} \begin{eqnarray} &&\sum_{\varepsilon=0,\pm2}\psi^{*(2)}_\varepsilon(u)^a_b\psi^{(2)}_\varepsilon(u)^b_c=\delta_{a,c}, \\ &&\sum_{a=b,b\pm 2}\psi^{*(2)}_{\varepsilon'}(u)^a_b\psi^{(2)}_\varepsilon(u)^b_a=\delta_{\varepsilon', \varepsilon}. \end{eqnarray} \end{prop} Furthermore, in the similar way to the derivation of \eqref{fusionvertexface}, we obtain the fused form of \eqref{vertexfacedual} \begin{eqnarray} &&\sum_{\mu_1',\mu_2'} R^{(2,2)}(u-v)_{\mu_1 \mu_2}^{\mu_1' \mu_2'}\ \psi^{*(2)}_{\mu_1'}(u)_{b}^{a} \psi^{*(2)}_{\mu_2'}(v)_{c}^{b} =\sum_{b' \in {\mathbb{Z}}} \psi^{*(2)}_{\mu_2}(v)_{b'}^{a} \psi^{*(2)}_{\mu_1}(u)_{c}^{b'} W_{22}\left(\left. \begin{array}{cc} c&b'\\ b&a \end{array}\right|u-v\right).{\nonumber}\\ &&\lb{fusionvertexfacedual} \end{eqnarray} The expression of $\psi^{*(2)}_\mu(u)_a^b\ (\mu=2, 0, -2)$ is evaluated as follows. \begin{prop} \begin{eqnarray} \left(\begin{array}{c} \psi^{*(2)}_2(u)_{n+2}^n\\ \psi^{*(2)}_0(u)_{n+2}^n\\ \psi^{*(2)}_{-2}(u)_{n+2}^n \end{array}\right)=\frac{C^4}{4[n+1][n+2][u][u+1]} \left(\begin{array}{c} \vartheta_3\left(\left. \frac{u-n-1 }{2r} \right|\frac{\tau}{2} \right)^2\\ -\vartheta_3\left(\left. \frac{u-n-1}{2r}\right| \frac{\tau}{2}\right) \vartheta_0\left( \left.\frac{u-n-1}{2r}\right|\frac{\tau}{2} \right) \\ \vartheta_0\left(\left. \frac{u-n-1 }{2r} \right|\frac{\tau}{2} \right)^2 \end{array}\right),{\nonumber} \end{eqnarray} \begin{eqnarray} &&\left(\begin{array}{c} \psi^{*(2)}_2(u)_{n}^n\\ \psi^{*(2)}_0(u)_{n}^n\\ \psi^{*(2)}_{-2}(u)_{n}^n \end{array}\right)=-\frac{C^5}{4[n][n+1][n-1][u][u+1]}{\nonumber}\\ &&\qquad\times \left(\begin{array}{c} \vartheta_3\left(\left. \frac{u+n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_3\left( \left.\frac{u-n-1}{2r}\right|\frac{\tau}{2} \right)\vt{1}{n-1}{\tau}+\vartheta_3\left(\left. \frac{u-n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_3\left( \left.\frac{u+n-1}{2r}\right|\frac{\tau}{2} \right)\vt{1}{n+1}{\tau}\\ -\vartheta_1\left(\left. \frac{n}{r}\right| \frac{\tau}{2} \right) \vartheta_2 \left( \left.\frac{1}{r} \right| \frac{\tau}{2} \right) \vartheta_0 \left( \left.\frac{u}{r} \right|{\tau} \right)\\ \vartheta_0\left(\left. \frac{u+n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_0\left( \left.\frac{u-n-1}{2r}\right|\frac{\tau}{2} \right)\vt{1}{n-1}{\tau}+\vartheta_0\left(\left. \frac{u-n+1}{2r}\right|\frac{\tau}{2} \right) \vartheta_0\left( \left.\frac{u+n-1}{2r}\right|\frac{\tau}{2} \right)\vt{1}{n+1}{\tau} \end{array} \right),{\nonumber} \end{eqnarray} \begin{eqnarray} \left(\begin{array}{c} \psi^{*(2)}_2(u)_{n-2}^n\\ \psi^{*(2)}_0(u)_{n-2}^n\\ \psi^{*(2)}_{-2}(u)_{n-2}^n \end{array}\right)=\frac{C^4}{4[n-1][n-2][u][u+1]} \left(\begin{array}{c} \vartheta_3\left(\left. \frac{u+n-1}{2r}\right|\frac{\tau}{2} \right)^2 \\ -\vartheta_3\left(\left. \frac{u+n-1}{2r}\right| \frac{\tau}{2} \right) \vartheta_0\left( \left.\frac{u+n-1}{2r}\right| \frac{\tau}{2} \right) \\ \vartheta_0\left(\left. \frac{u+n-1}{2r}\right|\frac{\tau}{2} \right)^2 \end{array}\right).{\nonumber} \end{eqnarray} \end{prop} Applying the crossing symmetry relations \eqref{crossing} and \eqref{SOScrossing} twice to \eqref{fusionvertexface}, we obtain the relation which should be compared with \eqref{fusionvertexfacedual}. Then fixing the suitable normalization function, we obtain\begin{thm} \begin{eqnarray*} &&\psi^{*(2)}_{\varepsilon}(u)^a_b=-\frac{C^4}{4[u][u+1]}\frac{\vtf{3}{0}{\tau}}{\vt{3}{1}{\tau}}\frac{g_a}{g_b (a,b)_2} \sum_{\varepsilon'=0,\pm2}Q^{\varepsilon'}_{\varepsilon}\psi^{(2)}_{\varepsilon'}(u-1)^a_b. \end{eqnarray*} \end{thm} \vspace{3mm} ~\\ {\Large\bf Acknowledgements}~~ \noindent The author would like to thank Michio Jimbo for sending his note and for stimulating discussions. He is also grateful to Takeo Kojima and Robert Weston for their collaboration in the work \cite{KKW}.
{ "timestamp": "2005-03-31T06:47:18", "yymm": "0503", "arxiv_id": "math/0503726", "language": "en", "url": "https://arxiv.org/abs/math/0503726" }
\section{Introduction} The one dimensional Discrete Nonlinear Schr\"odinger Equation (DNLS), \begin{eqnarray} \label{DNLS} i\dot{\psi}_n+\epsilon(\psi_{n-1}-2\psi_n+\psi_{n+1})+\gamma |\psi_n|^2\psi_n=0, \end{eqnarray} represents an infinite ($n\in\mathbb{Z}$), or a finite ($|n|\leq K$), one-dimensional array of coupled anharmonic oscillators, coupled to their nearest neighbors with a coupling strength $\epsilon$. Here $\psi_n(t)$ stands for the complex mode amplitude of the oscillator at lattice site $n$, and $\gamma$ denotes an anharmonic parameter. Setting $\epsilon=1/(\Delta x)^2$, reminds that the model includes a finite spacing between molecules, and the formal continuum limit, the NLS partial differential equation, is obtained by taking $\Delta x\rightarrow 0$. The DNLS equation is one of the most inportant inherently discrete models, having a crucial role in modelling of a great variety of phenomena, ranging from solid state and condensed matter physics to biology, \cite{Aubry,Eil,FlachWillis,HennigTsironis,Yuri}. Depending on the size of the lattice, we have to deal with an infinite or finite system of ordinary differential equations, respectively. The gauge invariance of the nonlinearity, allows for the support of special solutions of (\ref{DNLS}) of the form $\psi_n(t)=\phi_n\exp (i\omega t)$, $\omega>0$. These solutions are called {\em breather solutions}, due to their periodic time behavior. Inserting the ansatz of a breather solution into (\ref{DNLS}), it follows that $\phi_n$ satisfies the nonlinear system of algebraic equations \begin{eqnarray} \label{breather} -\epsilon(\phi_{n-1}-2\phi_n+\phi_{n+1})+\omega\phi_n=\gamma |\psi_n|^2\psi_n. \end{eqnarray} The problem of existence and stability properties of breather solutions of coupled oscillators, has been developed as a fascinating sublect of research, from the derivation of the stationary DNLS equation \cite{Holstein}, the derivation of stationary solutions for the (coupled) DNLS, by numerical continuation from the so-called anticontinuum limit (the case $\epsilon\rightarrow 0$) \cite{Eil2}, to the ingenious construction of localized time-periodic or quasiperiodic solutions of general discrete systems, starting from periodic solutions of the corresponding anticontinuum limit equations \cite{Aubry, RSMackayAubry}. We refer to \cite{Eil,Kevrekidis} for a review of the existing results and the history of the problem as for a long list of references. Motivated by \cite[Section 3]{Eil} and \cite{bang}, in this work we consider higher dimensional generalizations of DNLS, involving an arbitrary power law nonlinearity, and site dependence of the anharmonic parameter $\gamma$. For this particular case of nonlinearity, we also refer to \cite{Mol1,Mol2,Mol3}. For instance, we seek for breather solutions of the DNLS equation in infinite higher dimensional lattices ($n=(n_1,n_2,\ldots,n_N)\in\mathbb{Z}^N$), \begin{eqnarray} \label{DNLSh} i\dot{\psi}_n+(\mathbf{A}\psi)_n+ \gamma_n|\psi_n|^{2\sigma}\psi_n=0, \end{eqnarray} where \begin{eqnarray} \label{DiscLap} (\mathbf{A}\psi)_{n\in\mathbb{Z}^N}&=&\psi_{(n_{1}-1,n_2,\ldots ,n_N)}+\psi_{(n_1,n_{2}-1,\ldots ,n_N)}+\cdots+ \psi_{(n_1,n_{2},\cdots ,n_N-1)}\nonumber\\ &&-2N\psi_{(n_{1},n_2,\ldots ,n_N)} +\psi_{(n_{1}+1,n_2,\ldots ,n_N)}\nonumber\\ &&+\psi_{(n_1,n_{2}+1,\ldots ,n_N)}+\cdots+ \psi_{(n_1,n_{2},\cdots ,n_N+1)}, \end{eqnarray} In this case, equation (\ref{DNLSh}), could be viewed as the discretization of the NLS partial differential equation \begin{eqnarray} \label{NLSh} i\psi_t+\Delta\psi+\gamma(x)|\psi|^{2\sigma}\psi=0,\;\;x\in\mathbb{R}^N. \end{eqnarray} As in the one dimensional case, it can be easily seen that any breather solution $\psi_n(t)=\phi_n\exp (i\omega t)$, of the DNLS equation (\ref{DNLSh}), satisfies the infinite nonlinear system of agebraic equations \begin{eqnarray} \label{swe} -(\mathbf{A}\phi)_{n}+\omega\phi_n=\gamma_n|\phi_n|^{2\sigma}\phi_n,\;\;n\in\mathbb{Z}^N. \end{eqnarray} Based on a variational approach, which makes use of the famous Mountain Pass Theorem (MPT), we give a simple proof on the existence of (nontrivial) breather solutions for (\ref{DNLSh}), by showing that the energy functional associated to (\ref{swe}), has a critical point of ``mountain pass type''. Our main assumption is that $\gamma_n$ decays in an appropriate rate, in the sense that $\gamma_n$ is in an appropriate sequence space. This restriction enables us to use a compact inclusion between ordinary sequence spaces and {\em weighted} sequence spaces, in order to justify one of the crucial steps needed for for the application of MPT, namely the Palais-Smale condition. This is an important difference with the case of constant anharmonic parameter as the analysis of our recent work \cite{AN} shows: The latter is associated with lack of compactness, and restricted our study for a finite dimensional problem (in 1-D lattice, assuming Dirichlet boundary conditions). The application of MPT to (\ref{swe}) gives rise to some restrictions, which possibly have some physical interpretation, if viewed as local estimates for some ``energy quantities'' associated with the breather solution. On the other hand, it is shown that nontrivial solutions of (\ref{swe}) do not exist, in a sufficiently small ball of the energy space. The proof is based on a fixed point argument used also in \cite{AN}. This result could have the implementation, that we should not expect the existence of breather solutions, if the energy of the excitations of the lattice is sufficiently small. If the estimates derived by the application of the MPT, do not appear just as a technical step for the proof, they could be combined with that of the non-existence result, to derive a ``dispersion relation'' of nonlinearity exponent $\sigma$ vs the frequency $\omega$ of the breather solution, providing indication on the behavior of the associated energy quantities. For a detailed discussion on the breather problem in higher dimensional lattices and the dependence of the frequency $\omega$ on the conserved quantities of DNLS, we refer to \cite[Section 6]{Eil}. \section{Preliminaries} In this introductory section, we describe the functional setting needed for the treatment of the infinite nonlinear system (\ref{swe}). We also introduce some weighted sequence spaces, and we prove a compact inclusion between the ordinary sequence spaces and their weighted counterparts. For some positive integer $N$, we consider the complex sequence spaces \begin{equation} \label{ususeqs} \ell^p=\left\{ \begin{array}{ll} &\phi=\{\phi_n\}_{n\in\mathbb{Z}^{N}},\;n=(n_1,n_2,\ldots,n_N)\in\mathbb{Z}^N,\;\;\phi_n\in\mathbb{C},\\ &||\phi||_{\ell^p}=\left(\sum_{n\in\mathbb{Z}^N}|\phi_n|^p\right)^{\frac{1}{p}}<\infty \end{array} \right\}. \end{equation} Between $\ell^p$ spaces the following elementary embedding relation \cite{ree79} holds, \begin{eqnarray} \label{lp1} \ell^q\subset\ell^p,\;\;\;\; ||\phi||_{\ell^p}\leq ||\phi||_{\ell^q}\,\;\; 1\leq q\leq p\leq\infty. \end{eqnarray} Note that the contrary holds for the $L^p(\Omega)$-spaces if $\Omega\subset\mathbb{R}^N$ has finite measure. For $p=2$, we get the usual Hilbert space of square-summable sequences, which becomes a real Hilbert space if endowed with the real scalar product \begin{eqnarray} \label{lp2} (\phi,\psi)_{\ell^2}=\mathrm{Re}\sum_{{n\in\mathbb{Z}^N}}\phi_n\overline{\psi_n},\;\;\phi,\,\psi\in\ell^2. \end{eqnarray} For a sequence {\em of positive real numbers}\ \ $\delta=\{\delta_n\}_{n\in\mathbb{Z}^{N}}$, we define the weighted sequence spaces $\ell^2_{\delta}$ \begin{equation} \label{ususeqsw} \ell^p_{\delta}=\left\{ \begin{array}{ll} &\phi=\{\phi_n\}_{n\in\mathbb{Z}^{N}},\;n=(n_1,n_2,\ldots,n_N)\in\mathbb{Z}^N,\;\;\phi_n\in\mathbb{C},\\ &||\phi||_{\ell^p_{\delta}}=\left(\sum_{n\in\mathbb{Z}^N}\delta_n|\phi_n|^p\right)^{\frac{1}{p}}<\infty \end{array} \right\}. \end{equation} For the case $p=2$, it is not hard to see that $\ell^2_{\delta}$ is a Hilbert space, with scalar product \begin{eqnarray} \label{weightscal} (\phi,\psi)_{\ell^2_{\delta}}=\mathrm{Re}\sum_{n\in\mathbb{Z}^N}\delta_n\phi_n\overline{\psi_n},\;\;\phi,\,\psi\in\ell^2_\delta. \end{eqnarray} For a certain class of weight $\delta$, we have the following lemma which shall play a crucial role in our analysis. \begin{lemma} \label{compactness} We assume that the positive sequence of real numbers $\delta\in\ell^{\rho}$, $\rho=\frac{q-1}{q-2}$ for some $q>2$. Then $\ell^2\hookrightarrow\ell^2_{\delta}$ with compact inclusion. \end{lemma} {\bf Proof:} We use the ideas of \cite[Lemma 2.3, pg. 79]{KJB} and (\ref{lp1}). We consider a bounded sequence $\phi_k\in\ell^2$ and we denote by $(\phi_k)_n$ the $n$-th coordinate of this sequence. It suffices to show that the sequence $\phi_k$ is a Cauchy sequence in $\ell^2_{\delta}$. For some $q>2$ we consider its H\"older conjugate through the relation $p^{-1}+q^{-1}=1$. Then for all positive integers $k,l$, we have \begin{eqnarray} \label{lem1} ||\phi_k-\phi_l||^2_{\ell^2_{\delta}}&=&\sum_{n\in\mathbb{Z}^{N}}\delta_n|(\phi_k)_n-(\phi_l)_n|^2\\ &\leq& \left(\sum_{n\in\mathbb{Z}^{N}}\delta_n|(\phi_k)_n-(\phi_l)_n|^p\right)^{\frac{1}{p}} \left(\sum_{n\in\mathbb{Z}^{N}}\delta_n|(\phi_k)_n-(\phi_l)_n|^q\right)^{\frac{1}{q}}.\nonumber \end{eqnarray} Since $\phi_k$ is a bounded sequence in $\ell^2$, it follows from (\ref{lp1}) that $\phi_k$ is bounded in $\ell^q$. Then from (\ref{lem1}) we have that there exists a positive constant $c$, such that \begin{eqnarray} \label{lem2} ||\phi_k-\phi_l||^2_{\ell^2_{\delta}}\leq c \left(\sum_{n\in\mathbb{Z}^{N}}\delta_n|(\phi_k)_n-(\phi_l)_n|^p\right)^{\frac{1}{p}}. \end{eqnarray} Since $\delta\in\ell^{\rho}$, it holds that for any $\epsilon_1>0$, there exists $K_0(\epsilon_1)$ such that for all $K>K_0(\epsilon_1)$ $$\sum_{|n|> K}|\delta_n|^{\rho}<\epsilon_1.$$ Thus, using the boundedness of $\phi_k$ in $\ell^q$ once again, we have \begin{eqnarray} \label{lem5} \sum_{|n|> K}\delta_n|(\phi_k)_n-(\phi_l)_n|^p&\leq& \left(\sum_{|n|> K}|\delta_n|^{\rho}\right)^{\frac{1}{\rho}} \left(\sum_{|n|> K}|(\phi_k)_n-(\phi_l)_n|^q\right)^{\frac{p}{q}}\nonumber\\ &<& c\epsilon_1^{\frac{1}{\rho}}. \end{eqnarray} On the other hand, since the sequence $\phi_k$ is a Cauchy sequence in the finite dimensional space $\mathbb{C}^{(2K+1)^N}$, we get that for $k$ and $l$ sufficiently large and for any $\epsilon_2>0$, holds that \begin{eqnarray} \label{lem4} \sum_{|n|\leq K}\delta_n|(\phi_k)_n-(\phi_l)_n|^p < \epsilon_2. \end{eqnarray} Inequality (\ref{lem2}) can be rewritten as \begin{eqnarray} \label{lem3} ||\phi_k-\phi_l||^{2p}_{\ell^2_{\delta}}\leq c\left\{\sum_{|n|\leq K}\delta_n|(\phi_k)_n-(\phi_l)_n|^p +\sum_{|n|> K}\delta_n|(\phi_k)_n-(\phi_l)_n|^p\right\}. \end{eqnarray} Now from (\ref{lem5})-(\ref{lem3}), and appropriate choices of $\epsilon_1$ and $ \epsilon_2$, we may derive that for sufficiently large $k$ and $l$, $$||\phi_k-\phi_l||_{\ell^2_{\delta}}<\epsilon,\;\;\mbox{for any}\;\;\epsilon>0.$$ That is $\phi_k$ is a Cauchy sequence in $\ell^2_{\delta}$.\ \ $\diamond$ \vspace{0.2cm} Let $\mathbf{A}:D(\mathbf{A})\subseteq X\rightarrow X$ a $\mathbb{C}$-linear, self-adjoint\ $\leq 0$ operator with dense domain $D(\mathbf{A})$ on the Hilbert space $X$, equipped with the scalar product $(\cdot ,\cdot)_{X}$. The space $X_{\mathbf{A}}$ is the completion of $D(\mathbf{A})$ in the norm $||u||_{\mathbf{A}}^2=||u||^2_X-(\mathbf{A}u,u)_X$, for $u\in X_{\mathbf{A}}$, and we denote by $X_{\mathbf{A}}^*$ its dual and by $\mathbf{A}^*$ the extension of $\mathbf{A}$ to the dual of $D(\mathbf{A})$, denoted by $D(\mathbf{A})^*$ (Friedrichs extension theory \cite{Davies1}, \cite[Vol. II/A]{zei85}). Considering the operator $\mathbf{A}$ defined by (\ref{DiscLap}), we observe that for any $\phi\in\ell^2$ \begin{eqnarray} \label{preA} ||\mathbf{A}\phi||_{\ell^2}^2\leq 4N||\phi||_{\ell^2}^2, \end{eqnarray} that is, $\mathbf{A}:\ell^2\rightarrow\ell^2$ is a continuous operator. Now we consider the discrete operator $\mathbf{L}^+:\ell^2\rightarrow\ell^2$ defined by \begin{eqnarray} \label{discder1} (\mathbf{L}^+\psi)_{n\in\mathbb{Z}^N}&=&\left\{\psi_{(n_{1}+1,n_2,\ldots ,n_N)}-\psi_{(n_{1},n_2,\ldots ,n_N)}\right\}\nonumber\\ &+&\left\{\psi_{(n_{1},n_2+1,\ldots ,n_N)}-\psi_{(n_{1},n_2,\ldots ,n_N)}\right\}\nonumber\\ &\vdots&\nonumber\\ &+&\left\{\psi_{(n_{1},n_2,\ldots ,n_N+1)}-\psi_{(n_{1},n_2,\ldots ,n_N)}\right\}, \end{eqnarray} and $\mathbf{L}^{-}:\ell^2\rightarrow\ell^2$ defined by \begin{eqnarray} \label{discder2} (\mathbf{L}^-\psi)_{n\in\mathbb{Z}^N}&=&\left\{\psi_{(n_{1}-1,n_2,\ldots ,n_N)}-\psi_{(n_{1},n_2,\ldots ,n_N)}\right\}\nonumber\\ &+&\left\{\psi_{(n_{1},n_2-1,\ldots ,n_N)}-\psi_{(n_{1},n_2,\ldots ,n_N)}\right\}\nonumber\\ &\vdots&\nonumber\\ &+&\left\{\psi_{(n_{1},n_2,\ldots ,n_N-1)}-\psi_{(n_{1},n_2,\ldots ,n_N)}\right\}. \end{eqnarray} Setting \begin{eqnarray} \label{discder3} (\mathbf{L}^+_{\nu}\psi)_{n\in\mathbb{Z}^N}=\psi_{(n_{1},n_2,\ldots , n_{\nu-1},n_{\nu}+1,n_{\nu+1},\ldots ,n_N)}-\psi_{(n_{1},n_2,\ldots ,n_N)},\\ \label{discder4} (\mathbf{L}^-_{\nu}\psi)_{n\in\mathbb{Z}^N}=\psi_{(n_{1},n_2,\ldots , n_{\nu-1},n_{\nu}-1,n_{\nu+1},\ldots ,n_N)}-\psi_{(n_{1},n_2,\ldots ,n_N)}, \end{eqnarray} we observe that the operator $\mathbf{A}$ satisfies the relations \begin{eqnarray} \label{diffop2} (-\mathbf{A}\psi_1,\psi_2)_{\ell^2}&=&\sum_{\nu=1}^N(\mathbf{L}_\nu^+\psi_1,\mathbf{L}_\nu^+\psi_2)_{\ell^2},\;\;\mbox{for all}\;\;\psi_1,\psi_2\in\ell^2,\\ \label{diffop3} (\mathbf{L}_\nu^+\psi_1,\psi_2)_{\ell^2}&=&(\psi_1,\mathbf{L}_\nu^-\psi_2)_{\ell^2},\;\;\mbox{for all}\;\;\psi_1,\psi_2\in\ell^2. \end{eqnarray} From (\ref{diffop2}), it is clear that $\mathbf{A}:\ell^2\rightarrow\ell^2$ defines a self adjoint operator on $\ell^2$, and $\mathbf{A}\leq 0$. The graph-norm \begin{eqnarray*} ||\phi||_{D(\mathbf{A})}^2=||\mathbf{A}\phi||_{\ell^2}^2+||\phi||_{\ell^2}^2, \end{eqnarray*} is an equivalent with that of $\ell^2$, since \begin{eqnarray*} ||\phi||_{\ell^2}^2\leq ||\phi||^2_{D(\mathbf{A})}\leq (4N+1)||\phi||_{\ell^2}^2. \end{eqnarray*} In our case, it appears that $X_{\mathbf{A}}=\ell^2$ equipped with the norm $$||\phi||_{\mathbf{A}}^2=||\phi||_X^2-(\mathbf{A}\phi,\phi)_X=\sum_{\nu=1}^{N}||\mathbf{L}^+_{\nu}\phi||_{\ell_2}^2+ ||\phi||^2_{\ell^2},$$ for $\phi\in\ell^2$, and is an equivalent norm with the usual one of $\ell^2$. Moreover, $D(\mathbf{A})=X=\ell^2=D(\mathbf{A})^*$. Obviously $\mathbf{A}^*=\mathbf{A}$ and the operator $i\mathbf{A}:\ell^2\rightarrow \ell^2$ defined by $(i\mathbf{A})\phi=i\mathbf{A}\phi$ for $\phi\in \ell^2$, is $\mathbb{C}$-linear and skew-adjoint and $i\mathbf{A}$ generates a group $(\mathcal{T}(t))_{t\in\mathbb{R}}$, of isometries on $\ell^2$ (see \cite{cazS}). The analysis of the operator $\mathbf{A}$ is useful if one would like to consider the DNLS equation (\ref{DNLSh}) as an abstract evolution equation \cite{AN}, and holds for other discrete operators which are not necessary discretizations of the Laplacian (for example as those of \cite{SZ2}). \subsection{Existence of non trivial breather solutions in the case of decaying anharmonic parameter} We shall seek for nontrivial breather solutions as critical points of the functional \begin{eqnarray} \label{Enegfun} \mathbf{E}(\phi )=\frac{1}{2}\sum_{\nu=1}^{N}||\mathbf{L}^+_{\nu}\phi||_{\ell_2}^2+\frac{\omega^2}{2}\sum_{n\in\mathbb{Z}^N}|\phi_n|^2-\frac{1}{2\sigma +2}\sum_{n\in\mathbb{Z}^N}\gamma_n|\phi_n|^{2\sigma +2}. \end{eqnarray} To establish differentiability of the functional $\mathbf{E}:\ell^2\rightarrow\mathbb{R}$, we shall use the following discrete version of the dominated convergence Theorem, provided by \cite{Bates2}. \begin{theorem} \label{dc} Let $\{\psi_{i,k}\}$ be a double sequence of summable functions, $$\sum_{i\in\mathbb{Z}^N}|\psi_{i,k}|<\infty,$$ and $\lim_{k\rightarrow\infty}\psi_{i,k}=\psi_{i}$, for all $i\in\mathbb{Z}^N$. If there exists a summable sequence $\{g_{i}\}$ such that $|\psi_{i,k}|\leq g_{i}$ for all $i,k$'s, we have that $\lim_{k\rightarrow\infty}\sum_{i\in\mathbb{Z}^N}\psi_{i,k}=\sum_{i\in\mathbb{Z}^N}\psi_{i}$. \end{theorem} We then have the following Lemma. \begin{lemma} \label{derivative} Let $(\phi_n)_{n\in\mathbb{Z}^N}=\phi\in\ell^{2\sigma+2}$ for some $0<\sigma <\infty$. Moreover we assume that $\gamma_n\in\ell^{\rho}$, $\rho=\frac{q-1}{q-2}$ for some $q>2$. Then the functional $$\mathbf{F}(\phi)=\sum_{n\in\mathbb{Z}^N}\gamma_n|\phi_n|^{2\sigma +2},$$ is a $\mathrm{C}^{1}(\ell^{2\sigma +2},\mathbb{R})$ functional and \begin{eqnarray} \label{gatdev} <\mathbf{F}'(\phi),\psi>=(2\sigma +2)\mathrm{Re}\sum_{n\in\mathbb{Z}^N}\gamma_n|\phi_n|^{2\sigma}\phi_n\overline{\psi_n},\;\;\psi=(\psi_n)_{n\in\mathbb{Z}^N}\in\ell^{2\sigma +2}. \end{eqnarray} \end{lemma} {\bf Proof:}\ \ We assume that $\phi,\,\psi\in\ell^{2\sigma +2}$. Then for any $n\in\mathbb{Z}^N$, $0<s<1$, we get \begin{eqnarray} \label{mv} &&\frac{\mathbf{F}(\phi +s\psi)-\mathbf{F}(\psi)}{s}=\frac{1}{s}\mathrm{Re}\sum_{n\in\mathbb{Z}^N}\gamma_n\int_{0}^{1}\frac{d}{d\theta}|\phi_n + \theta s\psi_n|^{2\sigma +2}d\theta\\ &&\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;=(2\sigma +2)\mathrm{Re}\sum_{n\in\mathbb{Z}^N}\gamma_n\int_{0}^{1}|\phi_n+s\theta\psi_n|^{2\sigma} (\phi_n+s\theta\psi_n)\overline{\psi_n} d\theta.\nonumber \end{eqnarray} Since $\gamma_n$ is in $\ell^{\rho}$ it follows from (\ref{lp1}) that \begin{eqnarray} \label{boundV} \mathrm{sup}_{n\in\mathbb{Z}^N}|\gamma_n|=M<\infty. \end{eqnarray} On the other hand we have the inequality \begin{eqnarray} \label{mv1} &&\sum_{n\in\mathbb{Z}^N}|\phi_n+\theta s\psi_n|^{2\sigma +1}|\psi_n| \leq \sum_{n\in\mathbb{Z}^N}\left(|\phi_n|+|\psi_n|\right)^{2\sigma +1}|\psi_n|\;\;\;\;\;\;\;\nonumber\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;&&\leq\left(\sum_{n\in\mathbb{Z}^N}(|\phi_n|+|\psi_n|)^{2\sigma +2}\right)^{\frac{2\sigma +1}{2\sigma +2}} \left(\sum_{n\in\mathbb{Z}^N}|\psi_n|^{2\sigma +2}\right)^{\frac{1}{2\sigma +2}}. \end{eqnarray} Now by using (\ref{boundV}) and inserting (\ref{mv1}) into (\ref{mv}), we see that Lemma \ref{dc} is applicable: Letting $s\rightarrow 0$, we get the existence of the Gateaux derivative (\ref{gatdev}) of the functional $\mathbf{F}:\ell^{2\sigma +2}\rightarrow\mathbb{R}$. We show next that the functional $\mathbf{F}':\ell^{2\sigma +2}\rightarrow\ell^{\frac{2\sigma +2}{2\sigma +1}}$ is continuous. For $\phi\in\ell^{2\sigma +2}$, we set $(F_1(\phi))_{n\in\mathbb{Z}^N}=|\phi_n|^{2\sigma}\phi_n$. Let us note that for any $F\in \mathrm{C}(\mathbb{C},\mathbb{C})$ which takes the form $F(z)=g(|z|^2)z$, with $g$ real and sufficiently smooth, holds \begin{eqnarray} \label{GL} F(\phi_1)-F(\phi_2)=\int_{0}^{1}\left\{(\phi_1-\phi_2)(g(r)+rg'(r))+(\overline{\phi}_1-\overline{\phi}_2)\Phi^2 g'(r)\right\}d\theta, \end{eqnarray} for any $\phi_1,\;\phi_2\in \mathbb{C}$,where $\Phi=\theta \phi_1+(1-\theta)\phi_2$, $\theta\in (0,1)$ and $r=|\Phi|^2$ (see \cite[pg. 202]{GiVel96}). Applying (\ref{GL}) for the case of $F_1$ $(g(r)= r^{\sigma})$ one obtains that \begin{eqnarray*} F_1(\phi_1)-F_1(\phi_2)=\int_0^1[(\sigma +1)(\phi_1-\phi_2)|\Phi|^{2\sigma} +\sigma(\overline{\phi}_1-\overline{\phi}_2)\Phi^2|\Phi|^{2\sigma -2}]d\theta, \end{eqnarray*} which implies the inequality \begin{eqnarray} \label{GL2} |F_1(\phi_1)-F_1(\phi_2)|\leq (2\sigma +1)(|\phi_1|+|\phi_2|)^{2\sigma}|\phi_1-\phi_2|. \end{eqnarray} We consider a sequence $\phi_m\in\ell^{2\sigma +2}$ such that $\phi_m\rightarrow \phi$ in $\ell^{2\sigma +2}$. Using (\ref{boundV}), we get the inequality \begin{eqnarray} \label{hoin} \left|\left<\mathbf{F}'(\phi_m)-\mathbf{F}'(\phi),\,\psi\right>\right|&\leq& c(M)||F_1(\phi_m)-F_1(\phi)||_{\ell^{q}}||\psi||_{\ell^p},\\ &&q=\frac{2\sigma +2}{2\sigma +1},\;\;p=2\sigma +2.\nonumber \end{eqnarray} We denote by $(\phi_m)_n$ the $n$-th coordinate of the sequence $\phi_m\in\ell^2$. By setting $\Phi_n=(|(\phi_m)_n|+|\phi_n|)^{2\sigma}$, we get from (\ref{GL2}), that for some constant $c>0$ \begin{eqnarray*} &&\sum_{n\in\mathbb{Z}^N}|F_1((\phi_m)_n)-F_1(\phi_n)|^{q}\leq c\sum_{n\in\mathbb{Z}^N}(\Phi_n)^q|(\phi_m)_n-\phi_n|^{q}\nonumber\\ &&\leq c\left(\sum_{n\in\mathbb{Z}^N}|(\phi_m)_n-\phi_n|^{2\sigma +2}\right)^{\frac{1}{2\sigma +1}} \left(\sum_{n\in\mathbb{Z}^N}(\Phi_n)^{\frac{\sigma +1}{\sigma}}\right)^{\frac{2\sigma}{2\sigma +1}}\rightarrow 0, \end{eqnarray*} as $m\rightarrow\infty$.\ \ \ $\diamond$ By using (\ref{diffop2}), we may easily get that the rest of the terms of the functional $\mathbf{E}$ given by (\ref{Enegfun}), define $\mathrm{C}^1(\ell^2,\mathbb{R})$ functionals. Since Lemma \ref{derivative} holds for any $\phi\in\ell^2$ (by (\ref{lp1})), we finally obtain that the functional $\mathbf{E}$ is $\mathrm{C}^1(\ell^2,\mathbb{R})$. Moreover, by using the analysis of Section 1 for the self-adjoint operator $\mathbf{A}:\ell^2\rightarrow\ell^2$, it appears that any solution of (\ref{swe}), satisfies the formula \begin{eqnarray*} (-\mathbf{A}\phi,\psi)_{\ell^2}+\omega(\phi,\psi)_{\ell^2}=(\gamma_nF_1(\phi),\psi)_{\ell^2},\;\;\mbox{for all}\;\;\psi\in\ell^2, \end{eqnarray*} and vice versa. Equivalently, due to the differentiability of the functional $\mathbf{E}$, any solution of (\ref{swe}) is a critical point of $\mathbf{E}$. For convenience, we recall \cite[Definition 4.1, pg. 130]{CJ} (PS-condition) and \cite[Theorem 6.1, pg. 140]{CJ} or \cite[Theorem 6.1, pg. 109]{struwe} (Mountain Pass Theorem of Ambrosseti-Rabinowitz \cite{Amb}). \begin{definition} \label{condc} Let $X$ be a Banach space and $\mathbf{E}:X\rightarrow\mathbb{R}$ be $\mathrm{C}^1$. We say that $\mathbf{E}$ satisfies condition $(PS)$ if, for any sequence $\{\phi_n\}\in X$ such that $|\mathbf{E}(\phi_n)|$ is bounded and $\mathbf{E}'(\phi_n)\rightarrow 0$ as $n\rightarrow\infty$, there exists a convergent subsequence. If this condition is only satisfied in the region where $\mathbf{E}\geq\alpha >0$ (resp $\mathbf{E}\leq -\alpha <0$) for all $\alpha >0$, we say $\mathbf{E}$ satisfies condition $(PS^+)$ (resp. $(PS^-)$). \end{definition} \begin{theorem} \label{mpass} Let $\mathbf{E}:X\rightarrow\mathbb{R}$ be $C^1$ and satisfy (a) $\mathbf{E}(0)=0$, (b) $\exists\rho >0$, $\alpha >0:\;||\phi||_X=\rho$ implies $\mathbf{E}(\phi)\geq\alpha$, (c) $\exists \phi_1\in X :\;||\phi_1||_X\geq\rho$ and $\mathbf{E}(\phi_1)<\alpha$. Define $$\Gamma=\left\{\gamma\in \mathrm{C}^0([0,1],X):\;\gamma (0)=0,\;\;\gamma (1)=\phi_1\right\}.$$ Let $F_{\gamma}=\{\gamma(t)\in X:\;0\leq t\leq 1\}$ and $\mathcal{L}=\{F_\gamma :\;\gamma\in \Gamma\}$. If $\mathbf{E}$ satisfies condition $(PS)$, then $$\beta:=\inf_{F_{\gamma}\in \mathcal{L}}\sup\{\mathbf{E}(v):v\in F_{\gamma}\}\geq\alpha$$ is a critical point of the functional $\mathbf{E}$. \end{theorem} For fixed $\omega>0$, we shall consider a norm in $\ell^2$ defined by \begin{eqnarray} \label{moup1} ||\phi||_{\ell^2_{\omega}}^2=\sum_{\nu=1}^{\nu=N}||\mathbf{L}^+_{\nu}\phi||_{\ell_2}^2+\omega ||\phi||^2_{\ell^2},\;\;\phi\in\ell^2. \end{eqnarray} The norm (\ref{moup1}) is an equivalent norm with the usual one of $\ell^2$, since \begin{eqnarray} \label{moup2} \omega ||\phi||^2_{\ell^2}\leq ||\phi||_{\ell^2_\omega}^2\leq (2N+\omega)||\phi||^2_{\ell^2}. \end{eqnarray} We first check the behavior of the functional $\mathbf{E}$. Using (\ref{moup2}), we observe that \begin{eqnarray} \label{moup3} |\mathbf{F}(\phi)|&\leq& M\sum_{n\in\mathbb{Z}^N}|\phi_n|^{2\sigma+2} \leq M||\phi||_{\ell^2}^{2\sigma+2}\nonumber\\ &\leq&\frac{M}{\omega^{\sigma+1}}||\phi||^{2\sigma+2}_{\ell^2_{\omega}}. \end{eqnarray} Now setting $M_0=M/\omega^{\sigma+1}$ we observe that \begin{eqnarray} \label{moup4} \mathbf{E}(\phi)&=&\frac{1}{2}||\phi||^2_{\ell^2_{\omega}}-\frac{1}{2\sigma+2}\mathbf{F}(\phi)\nonumber\\ &&\geq \frac{1}{2}||\phi||^2_{\ell^2_{\omega}}-\frac{M_0}{2\sigma+2}||\phi||^{2\sigma+2}_{\ell^2_\omega}. \end{eqnarray} Now we select some $\phi\in\ell^2$ such that $||\phi||_{\ell^2_{\omega}}=R>0$. Then, if \begin{eqnarray} \label{disp1} 0<R < \left(\frac{\sigma +1}{M_0}\right)^{\frac{1}{2\sigma}}=\left(\frac{(\sigma +1)\omega^{\sigma+1}}{M}\right)^{\frac{1}{2\sigma}}:=E_{\ell^2_{\omega}}^*(\sigma,\omega,M), \end{eqnarray} it follows from (\ref{moup4}) that \begin{eqnarray} \label{moup5} \mathbf{E}(\phi) \geq \alpha>0,\;\;\alpha=R^2\left(\frac{1}{2}-\frac{M_0}{2\sigma +2}R^{2\sigma}\right). \nonumber \end{eqnarray} We assume that $\gamma_n>0$ for all $n\in\mathbf{S}_+\subseteq\mathbb{Z}^N$. We shall consider next, some $\psi\in\ell^2$ such that $||\psi||_{\ell^2_{\omega}}=1$ and \begin{eqnarray*} \{\psi_n\}_{n\in\mathbb{Z}}=\{\psi_n\}_{n\in\mathbf{S}_+}+\{\psi_n\}_{n\in(\mathbb{Z}^N\setminus\mathbf{S}_+)},\;\;\mbox{where}\;\; \left\{ \begin{array}{rlr} &\{\psi_n\}_{n\in\mathbf{S}_+}&>0,\nonumber \\ &\{\psi_n\}_{n\in(\mathbb{Z}^N\setminus\mathbf{S}_+)}&=0. \end{array} \right. \end{eqnarray*} For some $t>0$ we considet the element $\chi=t\psi\in\ell^2$. We have that \begin{eqnarray} \label{moup6} \mathbf{E}(\chi)=\frac{t^2}{2}-\frac{1}{2\sigma +2}t^{2\sigma +2}\sum_{n\in\mathbf{S}_+}\gamma_n|\psi_n|^{2\sigma+2}. \end{eqnarray} Now letting $t\rightarrow +\infty$ we get that $\mathbf{E}(t\psi)\rightarrow -\infty$. For fixed $\phi\neq 0$ and choosing $t$ sufficiently large, we may set $\phi_1=t\phi$ to satisfy the second condition of Theorem \ref{mpass}. To conclude with the existence of a non-trivial breather solution, it remains to show that the functional $\mathbf{E}$ satisfies Lemma \ref{condc}. To this end, we consider a sequence $\phi_m$ of $\ell^2$ be such that $|\mathbf{E}(\phi_m)|<M'$ for some $M'>0$ and $\mathbf{E}'(\phi_m)\rightarrow 0$ as $m\rightarrow\infty$. By using (\ref{Enegfun}) and Lemma \ref{derivative}, we observe that for $m$ sufficiently large \begin{eqnarray} \label{boundP.S} M'\geq \mathbf{E}(\phi_m)-\frac{1}{2\sigma +2}\left<\mathbf{E}'(\phi_m),\phi_m\right>= \left(\frac{1}{2}-\frac{1}{2\sigma +2}\right)||\phi_m||^2_{\ell^2_{\omega}}. \end{eqnarray} Therefore the sequence $\phi_m$ is bounded. Thus, we may extract a subsequence, still denoted by $\phi_m$, such that \begin{eqnarray} \label{weakcon} \phi_m\rightharpoonup \phi\;\;\mbox{in}\;\;\ell^2,\;\;\mbox{as}\;\;m\rightarrow\infty. \end{eqnarray} For this subsequence it follows once again from (\ref{Enegfun}) and Lemma \ref{derivative} that \begin{eqnarray} \label{moup8} ||\phi_m-\phi||_{\ell^2_{\omega}}^2&=&\left<\mathbf{E}'(\phi_m)-\mathbf{E}'(\phi),\phi_m-\phi\right>\nonumber\\ &&+\sum_{n\in\mathbb{Z}^N}\gamma_n[|(\phi_m)_n|^{2\sigma}(\phi_m)_n-|\phi_n|^{2\sigma}\phi_n]((\phi_m)_n-\phi_n)).\;\;\; \end{eqnarray} Another assumption on the sequence $\gamma_n$ is that the sequence $|\gamma_n|=(\delta_n)_{n\in\mathbb{Z}^N}$ satisfies the assumptions of Lemma \ref{compactness}. We consider the associated Hilbert space $\ell^2_{\delta}$. Now by using the inequality (\ref{GL2}), we get for the second term of right hand side of (\ref{moup8}), the estimate \begin{eqnarray} \label{moup9} &&\sum_{n\in\mathbb{Z}^N}\gamma_n[|(\phi_m)_n|^{2\sigma}(\phi_m)_n-|\phi_n|^{2\sigma}\phi_n]((\phi_m)_n-\phi_n))\;\;\;\;\;\nonumber\\ &&\;\;\;\;\leq c\sum_{n\in\mathbb{Z}^N}\Phi_n|\gamma_n|\,\,|(\phi_m)_n-\phi_n|^2\nonumber\\ &&\;\;\;\;\leq c\sup_{n\in\mathbb{Z}^N}\Phi_n\sum_{n\in\mathbb{Z}^N}|\gamma_n|\,|(\phi_m)_n-\phi_n|^2 =c_2||\phi_m-\phi||_{\ell^2_\delta}^2, \end{eqnarray} where $c_2=c\sup_{n\in\mathbb{Z}^N}\Phi_n$. Obviously, $\phi_m$ is bounded in $\ell^2_{\delta}$ and by Lemma \ref{compactness} it follows that \begin{eqnarray} \label{weakcon2} \phi_m\rightarrow \phi\;\;in\;\;\ell^2_{\delta},\;\;\mbox{as}\;\;m\rightarrow\infty. \end{eqnarray} Combining (\ref{moup8}), (\ref{moup9}) and (\ref{weakcon2}), we obtain that $$||\phi_m-\phi||_{\ell^2_{\omega}}\rightarrow 0,\;\;\mbox{as}\;\;m\rightarrow\infty.$$ Hence $\phi_m$ has a (strongly) convergent subsequence. The assumptions of Theorem \ref{mpass} are satisfied, and we may summarize in the following \begin{theorem} \label{COMP} Assume that the site-dependent anharmonic parameter $\gamma_n>0$ in some $\mathbf{S}_+\subseteq\mathbb{Z}^N$. Moreover, we assume that $|\gamma_n|=\delta_n\in\ell^{\rho}$, $\rho=\frac{q-1}{q-2}$ for some positive integer $q>2$. Then for any $\omega>0$, there exists a nontrivial breather solution $\psi_n(t)=\phi_n\exp (i\omega t)$ of the DNLS equation (\ref{DNLSh}). \end{theorem} \vspace{0.2cm} We remark here that the assumptions on the sequence of anharmonic parameters $\gamma_n$, are crucial for the derivation of the strong convergence of the subsequence $\phi_m$. If $\gamma_n$ is constant for all $n\in\mathbb{Z}^N$, then due to the lack of the Schur property for the space $\ell^{2}$ (in contrast with the space $\ell^1$ which posses this property-weak convergence coincides with strong convergence), we may not conclude the strong convergence of the subsequence, from its weak convergence. Of course the strong convergence, is valid in the case of a finite lattice: In this case, the problem is formulated in finite dimensional spaces where weak is equivalent to strong convergence \cite{AN}. Inequality (\ref{disp1}) could have some physical interpretation with respect to the nontrivial breather solutions of frequency $\omega>0$, if one considers (\ref{disp1}) as a possible local upper bound for the ``energy'' quantity defined by (\ref{moup1}). It contains information on the type of nonlinearity and the sequence of anharmonic parameters, through its dependence on the nonlinearity exponent $\sigma$ and $M$. Such type of relations seem to be reasonable, as the next result concerning nonexistence of nontrivial breather solutions shows. The restriction on the energy of the excitations for nonexistence, combined with the upper bound (\ref{disp1}) above, could provide us with some indicative information, on the behavior of energy quantities, associated with the nontrivial breather solution. For the shake of completeness, we recall \cite[Theorem 18.E, pg. 68]{zei85} (Theorem of Lax and Milgram). This theorem will be used to establish existence of solutions for an auxiliary infinite linear system of algebraic equations related to (\ref{swe}). \begin{theorem} \label{LMth} Let $X$ be a Hilbert space and $\mathbf{A}:X\rightarrow X$ be a linear continuous operator. Suppose that there exists $c^*>0$ such that \begin{eqnarray} \label{strongmonot} \mathrm{Re}(\mathbf{A}u,u)_X\geq c^*||u||^2_X,\;\;\mbox{for all}\;\;u\in X. \end{eqnarray} Then for given $f\in X$, the operator equation $\mathbf{A}u=f,\;\;u\in X$, has a unique solution \end{theorem} The non existence result can be stated as follows \begin{theorem} \label{notri} There exist no nontrivial breather solution of energy less than \begin{eqnarray} \label{disp2} E_{\mathrm{min}}(\omega,\sigma,M):=\frac{1}{2}\left(\frac{\omega}{M(2\sigma +1)}\right)^{1/2\sigma}. \end{eqnarray} \end{theorem} {\bf Proof:}\ \ For some $\omega> 0$, we consider the operator $\mathbf{A}_{\omega}:\ell^2\rightarrow\ell^2$, defined by \begin{eqnarray} \label{strongop1} (\mathbf{A}_{\omega}\phi)_{n\in\mathbb{Z}^N}&=&(\mathbf{A}\phi)_{n\in\mathbb{Z}^N}+\omega\phi_n. \end{eqnarray} It is linear and continuous and satisfies assumption (\ref{strongmonot}) of Theorem \ref{LMth}: Using (\ref{diffop2}), we get that \begin{eqnarray} \label{check} (\mathbf{A}_{\omega}\phi,\phi)_{\ell^2}=\sum_{\nu=1}^N|\mathbf{L}^+_{\nu}\phi||^2_{\ell^2}+\omega ||\phi||^2\geq \omega ||\phi||^2_{\ell^2}\;\;\mbox{for all}\;\;\phi\in\ell^2. \end{eqnarray} For given $z\in\ell^2$, we consider the linear operator equation \begin{eqnarray} \label{linear} (\mathbf{A}_{\omega}\phi)_{n\in\mathbb{Z}^N}=\gamma_n|z_n|^{2\sigma}z_n. \end{eqnarray} For the map \begin{eqnarray} \label{nolimap} (\mathbf{T}(z))_{n\in\mathbb{Z}^N}=\gamma_n|z_n|^{2\sigma}z_n, \end{eqnarray} we observe that \begin{eqnarray*} ||\mathbf{T}(z)||^2_{\ell^2}\leq M^2\sum_{n\in\mathbb{Z}^N}|z_n|^{4\sigma +2} \leq M^2||z||_{\ell^2}^{4\sigma +2}. \end{eqnarray*} Hence, the assumptions of Theorem \ref{LMth} are satisfied, and (\ref{linear}) has a unique solution $\phi\in\ell^2$. For some $R>0$, we consider the closed ball of $\ell^2$, $B_R:=\{z\in\ell^2\;:||z||_{\ell^2}\leq R\}$, and we define the map $\mathcal{P}:\ell^2\rightarrow\ell^2$, by $\mathcal{P}(z):=\phi$ where $\phi$ is the unique solution of the operator equation (\ref{linear}). Clearly the map $\mathcal{P}$ is well defined. Let $\zeta$, $\xi\in B_R$ such that $\phi=\mathcal{P}(\zeta)$, $\psi=\mathcal{P}(\xi)$. The difference $\chi:=\phi-\psi$ satisfies the equation \begin{eqnarray} \label{claim2} (\mathbf{A}_{\omega}\chi)_{n\in\mathbb{Z}^N}=(\mathbf{T}(z))_{n\in\mathbb{Z}^N}-(\mathbf{T}(\xi))_{n\in\mathbb{Z}^N}. \end{eqnarray} The map $\mathbf{T}:\ell^2\rightarrow\ell^2$ is locally Lipschitz, since we may use (\ref{GL2}) once again, to get \begin{eqnarray} \label{claim3} ||\mathbf{T}(\zeta)-\mathbf{T}(\xi)||_{\ell^2}^2&\leq& (2\sigma+1)^2M^2\sum_{n\in\mathbb{Z}^N}(|\zeta_n|+|\xi_n|)^{2\sigma})^2|\zeta_n-\xi_n|^2\nonumber\\ &\leq&(2\sigma+1)^2M^2[\sup_{n\in\mathbb{Z}^N}(|\zeta_n|+|\xi_n|)^{2\sigma})]^2\sum_{n\in\mathbb{Z}^N}|\zeta_n-\xi_n|^2\nonumber\\ &\leq& M_1^2R^{4\sigma}||\zeta-\xi||^2_{\ell^2}, \end{eqnarray} whith $M_1=2^{2\sigma}M(2\sigma+1)$. Taking now the scalar product of (\ref{claim2}) with $\chi$ in $\ell^2$ and using (\ref{claim3}), we have \begin{eqnarray} \label{cmap1a} \sum_{\nu=1}^N||\mathbf{L}_{\nu}^+\chi||^2_{\ell^2}+\omega ||\chi||^2_{\ell^2}&\leq& ||\mathbf{T}(\zeta)-\mathbf{T}(\xi)||_{\ell^2}||\chi||_{\ell^2}\nonumber\\ &\leq&M_1R^{2\sigma}||\zeta-\xi||_{\ell^2}||\chi||_{\ell^2}\nonumber\\ &\leq&\frac{\omega}{2}||\chi||_{\ell^2}^2+\frac{1}{2\omega}M^2_1R^{4\sigma}||z-\xi||_{\ell^2}^2. \end{eqnarray} From (\ref{cmap1a}), we obtain the inequality \begin{eqnarray} \label{claim4} ||\chi||_{\ell^2}^2=||\mathcal{P}(z)-\mathcal{P}(\xi)||_{\ell^2}^2 \leq \frac{1}{\omega^2}M^2_1R^{4\sigma}||z-\xi||^2_{\ell^2}. \end{eqnarray} Since $\mathcal{P}(0)=0$, from inequality (\ref{claim4}), we derive that for $R< E_{\mathrm{min}}$, the map $\mathcal{P}:B_R\rightarrow B_R$ and is a contraction. Therefore $\mathcal{P}$, satisfies the assumptions of Banach Fixed Point Theorem and has a unique fixed point, the trivial one. Hence, for $R<E_{\mathrm{min}}$ the only breather solution is the trivial. \ \ $\diamond$ \vspace{0.2cm} \newline If the energy of the excitation is less that $E_{\mathrm{min}}$ the lattice may not support a standing wave of frequency $\omega$. This time, relation (\ref{disp2}) could be seen as some kind of dispersion relation of frequency vs energy for the nonexsistence of breather solutions of the DNLS equation (\ref{DNLSh}). The dependence $E_{\ell^2_{\omega}}^*$ and $E_{min}$ on $\omega,\sigma, M$ as it appears from inequalities (\ref{disp1}), (\ref{disp2}), could be a point of departure for investigations on the relation of the energy quantity defined by (\ref{moup1}) and the $\ell^2$-norm of the nontrivial breather solution (the power), as well as on their behavior. For example, the inequality $E_{\mathrm{min}}<E_{\ell^2_{\omega}}^*$, is satisfied if \begin{eqnarray} \label{disp3} \left(\frac{1}{2^{2\sigma}(\sigma+1)(2\sigma+1)}\right)^{\frac{1}{\sigma}}<\omega. \end{eqnarray} In the case $\sigma=1$ (cubic nonlinearity) we get a lower bound $\omega>24^{-1}\sim 0.04166$, for the frequency of the nontrivial breather solution $\psi_n(t)=\phi_n\exp (i\omega t)$, satisfying \begin{eqnarray} \label{disp4} ||\phi||_{\ell^2_{\omega}}>E_{\mathrm{min}}. \end{eqnarray} Let us also remark that a similar nonexistence result as Theorem \ref{notri}, can be proved in the case (a) of an infinite lattice with $\gamma=\mathrm{const}, \cite{AN}$ and (b) the case of finite lattice (assuming Dirichlet boundary conditions). Numerical simulations, for testing restriction (\ref{disp2}) or (\ref{disp3})-(\ref{disp4}), could be of interest. Further developments could consider DNLS equations with site dependence on the coupling strength, or operators which are not necessarily discretizations of the Laplacian (for examples of such operators see \cite{SZ2}). \vspace{0.2cm} {\bf Acknowledgements}. I would like to thank Professors J. C. Eilbeck, and J. Cuevas, for their valuable discussions (especially for resolving the significance of relation (\ref{disp3})) and their interest, improving considerably the presentation of the final version of the manuscript, and my colleagues A. N. Yannacopoulos and H. Nistazakis for their suggestions. I would like also to thank the referee for his useful comments. This work was partially supported by the research project proposal ``Pythagoras I-Dynamics of Discrete and Continuous Systems and Applications''- National Technical University of Athens and University of the Aegean.
{ "timestamp": "2005-06-27T18:01:20", "yymm": "0503", "arxiv_id": "nlin/0503031", "language": "en", "url": "https://arxiv.org/abs/nlin/0503031" }
\section{Introduction} \label{intro} In this paper, we study the spectrum of one-dimensional perturbed periodic Schr{\"o}dinger operators. Precisely, we consider the Schr{\"o}dinger operator defined on $L^{2}(\mathbb{R})$ by: \begin{equation} \label{eqpa} H_{\varphi,\varepsilon}=-\frac{d^{2}}{dx^{2}}+[V(x)+W(\varepsilon x+\varphi)], \end{equation} where $\varepsilon>0$ is a small positive parameter, $\varphi$ is a real parameter, and $V$ is a real valued 1-periodic function. We also assume that $V$ is $L^{2}_{\textrm{loc}}$ and that $W$ is a fast-decaying function.\\ The operator $H_{\varphi,\varepsilon}$ can be regarded as an adiabatic perturbation of the periodic operator $H_{0}$: \begin{equation} \label{opper} H_{0}=-\triangle+V. \end{equation} The spectrum of the periodic operator $H_{0}$ is absolutely continuous and consists of intervals of the real axis called the spectral bands, separated by the gaps.\\ If the perturbation $W$ is relatively compact with respect to $H_{0}$, there are in the gaps of $H_{0}$ some eigenvalues \cite{Zhe, RB}. We intend to locate these eigenvalues, called impurity levels.\\ The equation \begin{equation} \label{eqp} H_{\varphi,\varepsilon}\psi=E\psi \end{equation} depends on two parameters $\varepsilon$ et $\varphi$. We study the operator $H_{\varphi,\varepsilon}$ in the adiabatic limit, i.e as $\varepsilon\rightarrow 0$. The periodicity of $V$ implies that the eigenvalues of $H_{\varphi,\varepsilon}$ are $\varepsilon$-periodic in $\varphi$. We shall shift $\varphi$ in the complex plane and we shall assume that $W$ is analytic in a strip of the complex plane.\\ If $V=0$, there are many results. The case when $W$ is a well has been studied; in the interval $]\inf\limits_{\mathbb{R}} W,0[$, there is a quantified sequence of eigenvalues \cite{Fe}. We shall give an analogous description of the eigenvalues of $H_{\varphi,\varepsilon}$ in an interval $J$ out of the spectrum of $H_{0}$. Precisely, when $W$ and $J$ satisfy some additional conditions described in sections \ref{assW1}, \ref{assW2} et \ref{assJ}, we show that the eigenvalues of $H_{\varphi,\varepsilon}$ oscillate around some quantized energies. The quantization is given by a Bohr-Sommerfeld quantization rule; the amplitude of oscillation is exponentially small and is determined by a tunneling coefficient.\\ \subsection{Physical motivation} The operator $H_{\varphi,\varepsilon}$ is an important model of solid state physics. The function $\psi$ is the wave function of an electron in a crystal with impurities. $V$ represents the potential of the perfect crystal; as such it is periodic. The potential $W$ is the perturbation created by impurities. In the semiconductors, this perturbation is slow-varying \cite{Zi}. It is natural to consider the semi-classical limit. \subsection{Perturbation of periodic operators} In $\mathbb{R}^{d}$, the spectral theory of the perturbations of a periodic operator \begin{equation} \label{hp} H_{P}=H_{0}+P \end{equation} has motivated numerous studies with different view points.\\ The characterization of the existence of eigenvalues is not easy: particularly, in any dimension, \cite{KuVa2} deals with the existence of embedded eigenvalues in the bands. On the real axis, the situation is simpler. When the perturbation is integrable, the eigenvalues are necessarily in the adherence of the gaps (\cite{RoBe, HiSh1}).\\ To count the eigenvalues in the gaps, many results have been obtained thanks to trace formulas. In the large coupling constant limit, i.e when $P=\lambda U$, with $\lambda\rightarrow+\infty$, \cite{ADH, Bi1, So2} have studied $\lim\limits_{\lambda\mapsto+\infty}\textrm{tr}(P_{[E,E']}^{(\lambda)})$, where $P_{[E,E']}^{(\lambda)}=1_{[E,E']}H_{\lambda}$ (spectral projector of $H_{\lambda}$ on an interval $[E,E']$ of a gap of $H_{0}$). In the semi-classical case, \cite{Di1} has given, under assumptions close to mine, an asymptotic expansion of $\textrm{tr} [f(H_{\varphi,\varepsilon})]$, for $f\in C_{0}^{\infty}(\mathbb{R})$ and $\textrm{Supp }f$ in a gap of $H_{0}$. These formulas are valid in any dimension but are less accurate. For example, in the expansion obtained in \cite{Di1}, the accuracy depends on the successive derivatives of the function $f$; the formula does not give an exponentially precise localization of the eigenvalues.\\ In the one-dimensional case, the scattering theory, well-known in the case $V=0$, has been developed in \cite{Fi1, New} for the periodic case. Precisely, we construct some particular solutions of equation \eqref{hp}, which tend to zero as $x$ tends to infinity. We call these functions recessive functions. The eigenvalues of equation \eqref{eqp} are given by a relation of linear dependence between these solutions. \subsection{Main steps of the study} \label{ppalet} We give here the main ideas of the paper. An important difficulty is the dependence of the equation on the parameters $\varepsilon$ and $\varphi$; particularly, one has to decouple the ``fast'' variable $x$ and the ``slow'' variable $\varepsilon x$. The new idea developed in \cite{FK1, FK2} is the following : we construct some particular solutions of \eqref{eqp}, satisfying an additional relation called the consistency condition: \begin{equation} \label{coh} f(x+1,\varphi,E,\varepsilon)=f(x,\varphi+\varepsilon,E,\varepsilon). \end{equation} This condition relates their behavior in $x$ and their behavior in $\varphi$.\\ To find a recessive solution of \eqref{eqp}, it suffices to construct a solution of \eqref{eqp} which satisfies \eqref{coh} and which tends to $0$ as $|\mbox{Re }\varphi|$ tends to $+\infty$. First, we build on the horizontal half-strip $\{\varphi\in \mathbb{C}\ ;\ \varphi\in]-\infty,-A]+i[-Y,Y]\}$ a solution $h_{-}^{g}$ of equation \eqref{eqp} which is consistent and which tends to $0$ as $\mbox{Re }\varphi$ tends to $-\infty$. Similarly, we construct $h_{+}^{d}$ for $\{\varphi\in \mathbb{C}\ ;\ \varphi\in[A,+\infty[+i[-Y,Y]\}$ (Theorem \ref{jostthm}). These functions are recessive for the variable $x$. The characterization of the eigenvalues is given by the relation of linear dependence between $h_{-}^{g}$ and $h_{+}^{d}$: $$w(h_{-}^{g},h_{+}^{d})=0.$$ In the above-mentioned equation, $w$ represents the Wronskian whose definition is recalled in \eqref{wronsk}.\\ It remains to compute $w(h_{-}^{g},h_{+}^{d})$. To do that, we use the complex WKB method developed by A. Fedotov and F. Klopp. This method consists in describing some complex domains, called canonical domains, on which we construct some functions satisfying \eqref{coh} and having a particular asymptotic behavior: \begin{equation} \label{stdas} f_{\pm}(x,\varphi,E,\varepsilon)=e^{\pm\frac{i}{\varepsilon}\int^{\varphi}\kappa}(\psi_{\pm}(x,\varphi,E)+o(1)),\quad \varepsilon\rightarrow 0. \end{equation} In equation \eqref{stdas}, the function $\kappa$ is a analytic multi-valued function, defined in \eqref{momcompa}; the functions $\psi_{\pm}$ are some particular solutions of equation $$H_{0}\psi=(E-W(\varphi))\psi,$$ analytic in $\varphi$ on these canonical domains and called Bloch solutions. We will prove the existence of such functions in section \ref{cansolbloch}.\\ A. Fedotov and F. Klopp prove the existence of functions with standard asymptotic only on compact domains of the complex plane. We shall extend some results on infinite strips of the complex plane. The consistency condition implies that the function $h_{-}^{g}$ satisfies the standard asymptotic \eqref{stdas} to the left of $-A$ and that $h_{+}^{d}$ satisfies an analogous property to the right of $A$. Thus, the computation of $w(h_{-}^{g},h_{+}^{d})$ is similar to the calculations of A. Fedotov and F. Klopp. We must find a sufficiently large domain of the complex plane, in which we know the Wronskian of $h_{-}^{g}$ and $h_{+}^{d}$.\\ The methods used in their works underline some topological obstacles, which change the standard asymptotic \eqref{stdas}; these obstacles depend on $W$ and $E$. We give precise assumptions in sections \ref{assW} and \ref{assJ}. \section{The main results} \label{resppaux} In this section, we describe the general context and the main results of the paper.\\ First, we present the assumptions on the potentials $V$ and $W$, and on the interval $J$. There are mainly three kinds of assumptions. Firstly, the study requires some assumptions on the decay of $W$ to develop the scattering theory. Then, in view of the hypotheses of the complex WKB method of \cite{FK1}, we assume that $W$ is analytic in some domain of the complex plane. Finally, we shall depict the geometric framework and particularly the subset $(E-W)^{-1}(\mathbb{R})$.\\ We obtain an equation for the eigenvalues in terms of geometric objects depending on $H_{0}$, $W$ and $E$: the phases and action integrals, defined in sections \ref{splecross}. \subsection{The potential $V$} \label{assV} We assume that $V$ has the following properties:\\ \\ {\bf ($\mathbf{H_{V,p}}$) $\mathbf{V}$ is $\mathbf{L^{2}_{\textrm{loc}}}$, $\mathbf{1}$-periodic.}\\ \\ We consider \eqref{eqp} as a perturbation of the periodic equation: \begin{equation} -\frac{d^{2}}{dx^{2}}\psi(x)+V(x)\psi(x)=(E-W(\varphi))\psi(x).\label{esp} \end{equation} We shall use some well known facts about periodic Schr{\"o}dinger operators. They are described in detail in section \ref{opepera}.\\ We just recall elementary results on $H_{0}$. The operator $H_{0}$ defined in \eqref{opper} is a self-adjoint operator on $H^{2}(\mathbb{R})$. The spectrum of $H_{0}$ consists of intervals of the real axis: \begin{equation} \label{band} \sigma(H_{0})=\bigcup\limits_{n\in\mathbb{N}}[E_{2n+1},E_{2n+2}], \end{equation} such that: $$ E_{1}<E_{2}\leq E_{3}< E_{4}...E_{2n}\leq E_{2n+1}< E_{2n+2}...,\quad E_{n}\rightarrow + \infty,n\rightarrow +\infty.$$ These intervals $[E_{2n+1},E_{2n+2}]$ are called the {\it spectral bands}. We set $E_{0}=-\infty$. The intervals $(E_{2n},E_{2n+1})$ are called the {\it spectral gaps}. If $E_{2n}\neq E_{2n+1}$, we say that the gap is open.\\ Furthermore, we assume that $V$ satisfies: \\ {\bf ($\mathbf{H_{V,g}}$) Every gap of $\mathbf{H_{0}}$ is not empty.}\\ \\ This assumption is ``generic'', we refer to \cite{ReSi4} section XIII.16. An important object of the theory of one-dimensional periodic operators is the Bloch quasi-momentum $k$ (see section \ref{qm}). This function is a multi-valued analytic function; its branch points are the ends of the spectrum, they are of square root type. We shall give a few details about this function in section \ref{opepera}. Finally, we suppose:\\ \\ {\bf ($\mathbf{H_{V}}$) $\mathbf{V}$ satisfy ($\mathbf{H_{V,p}}$) and ($\mathbf{H_{V,g}}$).}\\ \subsection{The perturbation $W$} \label{assW} \subsubsection{Smoothness assumptions} \label{assW1} We assume that $W$ is such that:\\ \\ {\bf ($\mathbf{H_{W,r}}$) There exists $\mathbf{Y>0}$ such that $\mathbf{W}$ is analytic in the strip $\mathbf{S_{Y}=\{|\mbox{Im }(\xi)|\leq Y\}}$ and there exists $\mathbf{s>1}$ et $C>0$ such that for $\mathbf{z\in S_{Y}}$, we have:} \begin{equation} \mathbf{|W(z)|\leq\frac{C}{1+|z|^{s}}}.\end{equation} These assumptions are essential to develop the complex WKB method. The analyticity of the perturbation is crucial in the theory of \cite{FK1}. The decay of $W$ replaces the compactness resulting from periodicity in \cite{FK1}.\\ We begin with presenting the complex momentum. This main object of the complex WKB method shows the importance of $W^{-1}(\mathbb{R})$. \subsubsection{The complex momentum and its branch points} \label{momcompb} We put: $$\mathbb{C}_{+}=\{\varphi\in\mathbb{C}\ ;\ \mbox{Im }\varphi\geq 0\}\textrm{ and } \mathbb{C}_{-}=\{\varphi\in\mathbb{C}\ ;\ \mbox{Im }\varphi\leq 0\}.$$ For equation \eqref{eqp}, we consider the analytic function $\kappa$ defined by \begin{equation} \label{momcompa} \kappa(\varphi)=k(E-W(\varphi)). \end{equation} We recall that the function $k$ is presented in section \ref{assV}. The function $\kappa$ is called the complex momentum. It plays a crucial role in adiabatically perturbed problems, see \cite{Bu, FK1}.\\ $\mathbb{N}$ is the set of non-negative integers. We define: \begin{equation} \label{nupsilon} \Upsilon(E)=\{\varphi\in S_{Y}\ ;\ \exists\ n\in\mathbb{N}^{*}\ /\ E-W(\varphi)=E_{n}\} \end{equation} The set of branch points of $\kappa$ is clearly a subset of $\Upsilon(E)$. The following result gives a characterization of the branch points of $\kappa$ among the points of $\Upsilon(E)$: \begin{lem} Let $\varphi$ be a point of $\Upsilon(E)$. If $\inf\{q\ ;\ W^{(q)}(\varphi)\neq 0\}\in 2\mathbb{N}+1$, then $\varphi$ is a branch point of $\kappa$. \end{lem} This result follows from the fact that the ends of the spectrum are of square root type. \subsubsection{Geometric assumptions} \label{assW2} \label{descw} The spectrum $\sigma(H_{0})$ consists of real intervals. Fix $E\in\mathbb{R}$. If $E-W(\varphi)$ is in the spectrum $\sigma(H_{0})$, then $W(\varphi)$ is real. The spectral study of (\ref{eqp}) is then tightly connected with the geometry of $W^{-1}(\mathbb{R})$.\\ We state now the geometric assumptions for $W$. These assumptions are mainly a description of $W^{-1}(\mathbb{R})$ in a strip containing the real axis. We call strictly vertical a line whose slope does not vanish; for precise definitions, we refer to section \ref{vertdef}. \\ {\bf ($\mathbf{H_{W,g}}$)\begin{enumerate} \item $\mathbf{W_{|\mathbb{R}}}$ is real and has a finite number of extrema, which are non-degenerate. \item There exists $\mathbf{Y>0}$ and a finite sequence of strictly vertical lines containing an extremum of $\mathbf{W}$, such that: \begin{equation} \mathbf{W^{-1}(\mathbb{R})\cap S_{Y}=\bigcup\limits_{i\in \{1\ldots p\}}(\Sigma_{i})\cup\mathbb{R}}. \end{equation}\end{enumerate}} \subsection{Some remarks} \begin{itemize} \item Since $W$ is real analytic, we know that $W(\overline{\varphi})=\overline{W(\varphi)}$; this implies that $W^{-1}(\mathbb{R})$ is symmetric with respect to the real axis. \item We define $\Sigma_{i}^{+}=\Sigma_{i}\cap\mathbb{C}_{+}$ and $\Sigma_{i}^{-}=\Sigma_{i}\cap\mathbb{C}_{-}$. \item Figure \ref{exW1} shows an example of the pre-image of the real axis by such a potential. \end{itemize} As we have explained in section \ref{ppalet}, we cover the strip $S_{Y}$ with local canonical domains. On these domains, we construct consistent functions with standard behavior (ie satisfying \eqref{coh} and \eqref{stdas}).\\ To compute the connection between the bases associated with different domains, we get round the branch points (for analog studies, we refer the reader to \cite{FR, FK2}). We will now state some more accurate assumptions about the configuration of the branch points; in particular, these assumptions specify $(E-W)^{-1}(\sigma(H_{0}))$ when $E$ is real. The spectral results of A. Fedotov and F. Klopp on perturbed periodic equation have shown the importance of the relative positions of $J$ and $\sigma(H_{0})$.\\ \input{fig1} \subsection{Assumptions on the interval $J$} \label{assJ} Now, we describe the interval $J$ on which we study equation \eqref{eqp}. \subsubsection{Hypotheses} We assume that the interval $J$ is a compact interval satisfying:\\ \\ {\bf $\mathbf{(H_{J})}$ \begin{enumerate} \item For any $\mathbf{E\in J}$, there exists only one band $\mathbf{B}$ of $\mathbf{\sigma(H_{0})}$ such that the pre-image $\mathbf{C:=(E-W)^{-1}(B)}$ is not empty. \item For any $\mathbf{E\in J}$, $\mathbf{C:=(E-W)^{-1}(B)}$ is connected and compact and $\mathbf{(E-W)^{-1}(\stackrel{\circ}{B})}$ contains exactly one real extremum of $\mathbf{W}$. \end{enumerate}} \subsubsection{Consequences} \begin{itemize} \item $(H_{J})$ implies that $J$ is included in a gap.\item The band $B$ in $(H_{J})$ (1) depends a priori on $E$. But, since $J$ is connected, the band $B$ is fixed for any $E\in J$.\item Similarly, the extremum of $W$ in assumption $(H_{J})$ (2) depends on $E$, but by connectedness, it is the same for any $E\in J$.\end{itemize} \subsubsection{Notations} Put $B=[E_{2n-1},E_{2n}]$, for $n\in\mathbb{N}^{*}$. Moreover, we can always change $W$ or $\varphi$ so that the extremum of $W$ in $(2)$ is $0$.\\ Then $(H_{J})$ has the following consequences: \begin{enumerate}\item For any $E\in J$, $(E-W)^{-1}(\sigma(H_{0}))\cap S_{Y}=(E-W)^{-1}(B)\cap S_{Y}$\item Let $E_{r}\in\{E_{2n-1},E_{2n}\}$ be the end of $B$ satisfying $E_{r}\in(E-W)(\mathbb{R})$ for any $E\in J$. We define $E_{i}$ such that $\{E_{i},E_{r}\}=\{E_{2n-1},E_{2n}\}$.\item There are exactly four branch points $(\varphi_{r}^{-},\varphi_{r}^{+})\in\mathbb{R}^{2}$ and $(\varphi_{i}, \overline{\varphi_{i}})$ in $S_{Y}$ related to $E_{r}$ and $E_{i}$. They satisfy: $$E-W(\varphi_{r}^{+})=E_{r},E-W(\varphi_{r}^{-})=E_{r},\ \varphi_{r}^{-}<0<\varphi_{r}^{+},$$ $$E-W(\varphi_{i})=E-W(\overline{\varphi_{i}})=E_{i},\ \mbox{Im }\varphi_{i}>0.$$\item There exists a strictly vertical line $\sigma$ containing $0$ and connecting $\overline{\varphi_{i}}$ to $\varphi_{i}$, such that $(E-W)^{-1}(B)\cap S_{Y}=[\varphi_{r}^{-},\varphi_{r}^{+}]\cup\sigma$. We define $\sigma_{+}=\sigma\cap\mathbb{C}_{+}$ and $\sigma_{-}=\sigma\cap\mathbb{C}_{-}$. We let $\Sigma=(E-W)^{-1}(\mathbb{R})\backslash\mathbb{R}$, $\sigma\subset\Sigma$. \end{enumerate} These objects are described in figure \ref{pbf}. \input{fig2} \subsubsection{Remarks and examples} We first give a few comments on assumption $(H_{J})$. \begin{itemize} \item We call $C$ the cross. \item This assumption means intuitively that, in $S_{Y}$, we see the band $B$ only near the extremum $0$. \end{itemize} To illustrate these technical assumptions, we give a few examples of potentials $W$ and intervals $J$. We have depicted some examples in figure \ref{exW2}.\begin{itemize}\item The simplest case is when $W$ has only a non-degenerate minimum $W_{-}$ (see figure \ref{exW2} A).\\ in concrete terms, we can think of the example: $$ W(x)=\frac{-\alpha}{1+x^{2}},\quad \alpha>0,$$ Then, if we fix $B=[E_{2n-1},E_{2n}]$ and $Y<1$, we can choose $J=[a,b]$ such that: $$\max\{E_{2n-2},E_{2n-1}-\alpha,E_{2n}-\frac{\alpha}{1-Y^{2}}\}< a<b<\min\{E_{2n-1},E_{2n}-\alpha,E_{2n+1}-\frac{\alpha}{1-Y^{2}}\}$$ \item We can assume that $W$ has a maximum $W_{+}$ and a minimum $W_{-}$, if $J$ is chosen to see the band only near the maximum (see figure \ref{exW2} B).\\ $$W(x)=\frac{2}{1+x^{2}}-\frac{1}{1+(x-5)^{2}} $$ $$J\subset]E_{2n-1}+W_{+},E_{2n-2}+W_{+}[\cup]E_{2n},E_{2n+1}+W_{-}[,\quad |J|\leq|E_{2n-2}-E_{2n-1}|$$ Consider this example a little further. The choice of $Y$ is more complicated in this case. The study of equation $W(u)=w$ for $w>W_{+}$ shows that there exists only one solution in the strip $\{\mbox{Im } u\in]0,1[\}$ that we call $Z(w)$ ; we choose $Y\in]\sup\limits_{E\in J}Z(E-E_{2l-1}),\inf\limits_{E\in J}Z(E-E_{2l-2})[$. \item In fact, we could adapt our method to weaker assumptions. For example, we can assume that we do not see the branch points $\varphi_{i}$ and $\overline{\varphi_{i}}$ (incomplete cross), which means that the vertical line $\sigma$ does not contain any branch points of $\kappa$. We refer to section \ref{unccross} for some details. \item For the sake of simplicity, we have assumed that all the extrema of $W$ are non degenerate. Actually, it suffices to assume that only the extremum of $W$ in $0$ is non degenerate. \item Similarly, we could weaken assumption $(H_{V,g})$. We only have to assume that the gaps adjoining the band $B$ of $(H_{J})$ are not empty. \end{itemize} \input{fig3} \subsection{Phases and action} \label{splecross} In this section, we define the tunneling coefficient $t$ and the phases $\Phi$ et $\Phi_{d}$; these analytic objects play an essential role in the location of the eigenvalues. These coefficients are represented as integrals of the complex momentum $\kappa$ in the $\varphi$ plane.\\ In the strip $S_{Y}$, we consider $\kappa$ a branch of the complex momentum, continuous on $C$. \subsubsection{Definition and properties} \label{chem} We introduce the action $S$ and the phases $\Phi$ and $\Phi_{d}$ related to the branch $\kappa$. \begin{defn} We define the phase: \begin{equation} \label{phi} \Phi(E)=\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}(\kappa(u)-\kappa(\varphi_{r}^{-}))du, \end{equation} the action: \begin{equation} \label{action} S(E)=i\int_{\sigma}(\kappa(u)-\kappa(\varphi_{i}))du, \end{equation} the second phase: \begin{equation} \label{phid} \Phi_{d}(E)=\int_{\varphi_{r}^{-}}^{0}(\kappa(u)-\kappa(\varphi_{r}^{-}))du+\int_{\varphi_{r}^{+}}^{0}(\kappa(u)-\kappa(\varphi_{r}^{+}))du+\int_{\sigma_{+}}(\kappa(u)-\kappa(\varphi_{i}))du-\int_{\sigma_{-}}(\kappa(u)-\kappa(\overline{\varphi_{i}}))du. \end{equation} \end{defn} In section \ref{anares}, we prove the following result on the behavior of the coefficients $\Phi$, $S$ and $\Phi_{d}$. \begin{lem} \label{phaseactint} There exists a branch $\tilde{\kappa}_{i}$ such that the phases and action integrals have the following properties: \begin{enumerate} \item $\Phi$, $S$, $\Phi_{d}$ are analytic in $E$ in a complex neighborhood of the interval $J$. \item $\Phi$, $S$, $\Phi_{d}$ take real values on $J$. $\Phi$ and $S$ are positive on $J$. \item $\forall E\in J,\quad \Phi'(E)(E_{i}-E_{r})>0,\quad S(E)\leq 2\pi\ \mbox{Im }(\varphi_{i}(E)).$ \end{enumerate} \end{lem} We define {\it the tunneling coefficient} : \begin{equation} t(E,\varepsilon)=\exp(-S(E)/\varepsilon). \end{equation} $t$ is exponentially small. \subsubsection{Remark} The phase and action are simply a generalization of the coefficients of the form $\int\sqrt{E-W(\varphi)}d\varphi$, well-known in the case $V=0$ (we refer to \cite{Fe, FR, Ra}).\\ We point out that the coefficient $\Phi$ depend only on the value of $W$ on the real axis, whereas $S$ and $\Phi_{d}$ depend on the values of $W$ in the complex plane. The phase $\Phi$ is independent of the analyticity of $W$ unlike $S$ and $\Phi_{d}$.\\ Now, we state the equation for eigenvalues for \eqref{eqp}. \subsubsection{The main result} \begin{thm} \label{eigenloc} Equation for eigenvalues.\\ Let $V$, $W$ and $J$ satisfy assumptions $(H_{V})$, $(H_{W,r})$, $(H_{W,g})$ and $(H_{J})$. Fix $Y_{0}\in]0,Y[$.\\ There exists a complex neighborhood $\mathcal{V}$ of $J$, a real number $\varepsilon_{0}>0$ and two functions $\widetilde{\Phi}$ and $\widetilde{\Phi_{d}}$ with complex values, defined on $\mathcal{V}\times]0,\varepsilon_{0}[$ such that: \itemize{\item The functions $\widetilde{\Phi}(\cdot,\varepsilon)$ and $\widetilde{\Phi_{d}}(\cdot,\varepsilon)$ are analytic on $\mathcal{V}$. Moreover, $\widetilde{\Phi}$ and $\widetilde{\Phi_{d}}$ satisfy: $$\widetilde{\Phi}(E,\varepsilon)=\Phi(E)+h_{0}(E,\varepsilon)\quad\textrm{ and }\quad\widetilde{\Phi_{d}}(E,\varepsilon)=\Phi_{d}(E)+h_{1}(E,\varepsilon),$$ where $\rho$ is a real coefficient, $h_{0}(E,\varepsilon)=o(\varepsilon)$ and $h_{1}(E,\varepsilon)=o(\varepsilon)$ uniformly in $E\in\mathcal{V}$. \item If we define the energy levels $\{E^{(l)}(\varepsilon)\}$ in $J$ by: \begin{equation} \frac{\widetilde{\Phi}(E^{(l)}(\varepsilon),\varepsilon)}{\varepsilon}=l\pi+\frac{\pi}{2},\quad\quad\forall l\in\{L_{-}(\varepsilon),\ldots,L_{+}(\varepsilon)\}, \end{equation} then, for any $\varepsilon\in]0,\varepsilon_{0}[$,\itemize{\item the spectrum of $H_{\varphi, \varepsilon}$ in $J$ consists in a finite number of eigenvalues, that is to say \begin{equation} \sigma(H_{\varphi, \varepsilon})\cap J =\bigcup\limits_{l\in\{L_{-}(\varepsilon),\ldots,L_{+}(\varepsilon)\}}\{E_{l}(\varphi,\varepsilon)\}, \end{equation}\item these eigenvalues satisfy {\footnotesize\begin{equation}\label{compcross} E_{l}(\varphi,\varepsilon)=E^{(l)}(\varepsilon)+\varepsilon(-1)^{l+1}\frac{t(E^{(l)}(\varepsilon),\varepsilon)}{\Phi'(E^{(l)}(\varepsilon))}\left[\cos\left(\frac{\widetilde{\Phi_{d}}(E^{(l)}(\varepsilon),\varepsilon)+2\pi\varphi+\rho\varepsilon}{\varepsilon}\right)+t(E^{(l)}(\varepsilon),\varepsilon)r(E^{(l)}(\varepsilon),\varphi,\varepsilon)\right], \end{equation}} where there exists $c>0$ such that $$\sup\limits_{E\in\mathcal{V},\varphi\in\mathbb{R}}r(E,\varphi,\varepsilon)<\frac{1}{c}e^{-\frac{c}{\varepsilon}}.$$}} \end{thm} We prove this result in section \ref{anares}. \subsubsection{Remark} \label{unccross} If we only assume that $\sigma$ does not contain any branch points, asymptotic \eqref{compcross} is replaced by the estimate: $$|E_{l}(\varphi,\varepsilon)-E^{(l)}(\varepsilon)|< C e^{-\frac{2\pi Y}{\varepsilon}}$$ where $2 Y$ is the width of the strip $S_{Y}$. \subsubsection{Application : asymptotic expansion of the trace} By using the previous result, we can compute the first terms in the asymptotic expansion of the trace formula , and partially recover a result of \cite{Di1}. \begin{cor} \label{cordim} Let $f\in C_{0}^{\infty}(\mathbb{R})$ be a real function such that $\textrm{Supp }f\in J$. Then the function $f(H_{\varphi,\varepsilon})$ is $\varepsilon$-periodic in $\varphi$ and its Fourier expansion satisfies: \begin{equation} \textrm{tr }[f(H_{\varphi,\varepsilon})]=\frac{1}{\varepsilon}\int_{0}^{\varepsilon}[f(H_{u,\varepsilon})]du+O(e^{-S/\varepsilon}) \end{equation} \begin{equation} \int_{0}^{\varepsilon}[f(H_{u,\varepsilon})]du=\frac{1}{2\pi}\int_{\mathbb{R}_{u}}\int_{[-\pi,\pi]}f(W(u)+E(\kappa))d\kappa du+o(\varepsilon) \end{equation} where $S=\inf\limits_{e\in\textrm{Supp }f}S(e)>0$ \end{cor} We give more details and the proof of this corollary in section \ref{trform2}. \section{Main steps of the study} \label{schem} Here, we explain the main ideas of the paper. \subsection{One-dimensional perturbed periodic operators} \subsubsection{} \label{rapp} We consider equation \eqref{eqp} as a perturbation of the periodic equation \begin{equation} \label{espa} H_{0}\psi=E\psi \end{equation} where the operator $H_{0}$ is defined in \eqref{opper}. To do that, we shall describe the spectral theory of periodic operators in section \ref{opepera}. \begin{itemize} \item \label{rappa} For the moment, we simply introduce the Bloch solutions of equation \eqref{esp}. We call a {\it Bloch solution} of \eqref{esp} a function $\Psi$ satisfying \eqref{esp} and: \begin{equation} \label{blochsola} \forall x\in\mathbb{R},\quad\Psi(x+1,E)=\lambda(E)\Psi(x,E), \end{equation} with $\lambda\neq 0$ independent of $x$. The coefficient $\lambda(E)$ is called {\it Floquet multiplier}. We represent $\lambda(E)$ in the form $\lambda(E)=e^{i k(E)}$; $k$ is the quasi-momentum presented in section \ref{assV} and described in section \ref{qm}. If $E\notin\sigma(H_{0})$, there exist two linearly independent Bloch solutions of \eqref{espa}(see section \ref{bloch}). We call them $\widetilde{\Psi}_{+}$ et $\widetilde{\Psi}_{-}$; the associated Floquet multipliers are inverse of each other and the functions $\widetilde{\Psi}_{\pm}$ are represented in the form: $$ \widetilde{\Psi}_{\pm}(x,E)=e^{\pm ik(E)x}p_{\pm}(x,E)\quad\textrm{avec}\quad p_{\pm}(x+1,E)=p_{\pm}(x,E).$$ For $\mbox{Im } k(E)>0$, the function $\widetilde{\Psi}_{+}(x,E)$ tends to $0$ as $x$ tends to $+\infty$ and the function $\widetilde{\Psi}_{-}(x,E)$ tends to $0$ as $x$ tends to $-\infty$. Actually, equation \eqref{blochsola} defines the functions $\widetilde{\Psi}_{+}$ and $\widetilde{\Psi}_{-}$ except for a multiplicative coefficient. Precisely, equation \eqref{blochsola} defines two one-dimensional vector spaces that we call {\it Bloch sub-spaces}.\\ To study the eigenvalues of perturbations of periodic operators, \cite{Fi1} and \cite{New} introduce, for $\mbox{Im } k(E)>0$, two functions $(x,\varphi,E,\varepsilon)\mapsto F_{+}(x,\varphi,E,\varepsilon)$ and $(x,\varphi,E,\varepsilon)\mapsto F_{-}(x,\varphi,E,\varepsilon)$ solutions of \eqref{eqp} satisfying: \begin{equation} \label{jostcond} \lim\limits_{x\rightarrow +\infty}[F_{+}(x,\varphi,E,\varepsilon)-\widetilde{\Psi}_{+}(x,E)]=0,\quad \lim\limits_{x\rightarrow -\infty}[F_{-}(x,\varphi,E,\varepsilon)-\widetilde{\Psi}_{-}(x,E)]=0 \end{equation} Condition \eqref{jostcond} guarantees the uniqueness of $F_{+}$ (resp. of $F_{-}$) since the function $\widetilde{\Psi}_{+}$ (resp. $\widetilde{\Psi}_{-}$) tends to $0$ as $x$ tends to $+\infty$ (resp. $-\infty$). These functions are called Jost functions; they are generally constructed as solutions of a Lippman-Schwinger integral equation. This construction is an adaptation of the usual theory of scattering (chapter XI of \cite{ReSi3}) for a perturbation of laplacian; it consists in looking for particular solutions of \eqref{eqp} from the solutions of the periodic equation.\\ We call {\it Jost sub-spaces} the sub-spaces $\mathcal{J}_{+}$ and $\mathcal{J}_{-}$ generated by $F_{+}$ and $F_{-}$.\\ $\mathcal{J}_{+}$ (resp $\mathcal{J}_{-}$) is the set of solutions of \eqref{eqp} being a member of $L^{2}([0,\infty))$ (resp. $L^{2}((-\infty,0])$). \item Let $f$ and $g$ be two derivable functions, the {\it Wronskian} of $f$ and $g$ called $w(f,g)$ is defined by: \begin{equation} \label{wronsk} w(f,g)=f'g-fg' \end{equation} We recall that if $f$ and $g$ are the solutions of a second-order differential equation, their Wronskian is independent of $x$. The spectral interest of the Jost sub-spaces is the following: \begin{prop}\label{carvp} We assume that $\mbox{Im } k(E)>0$. Let $h^{-}_{g}\in\mathcal{J}_{-}$ and $h^{+}_{d}\in\mathcal{J}_{+}$ be two nontrivial Jost solutions of \eqref{eqp}. $E$ is an eigenvalue of $H_{\varphi,\varepsilon}$ if and only if: \begin{equation} w(h^{+}_{d},h^{-}_{g})=0 \end{equation} \end{prop} To compute the eigenvalues, it suffices to construct the Jost sub-spaces. \end{itemize} \subsection{Construction of consistent Jost solutions} We denote by $(H_{J}^{0})$ the following assumption: \\ \\ {\bf $\mathbf{(H_{J}^{0})}$ There exists $n\in\mathbb{N}$ such that $\mathbf{J}$ is a compact interval of $\mathbf{]E_{2n},E_{2n+1}[}$.}\\ \\ Clearly, $(H_{J}^{0})$ is weaker than $(H_{J})$.\\ We introduce a new notation.\\ For a function $f:\ \mathcal{U}\subset\mathbb{C}^{n}\rightarrow\mathbb{C}^{p}$, we define the function $f^{*}:\ \overline{\mathcal{U}}\rightarrow\mathbb{C}^{p}$: \begin{equation} \label{eqconj1} f^{*}(Z)=\overline{f(\overline{Z})}. \end{equation} As we have explained in section \ref{intro}, an useful idea to study \eqref{eqp} is the construction of consistent solutions, i.e. satisfying \eqref{coh}. First, we choose in $\mathcal{J}_{-}$ and $\mathcal{J}_{+}$ some consistent bases. We shall prove the following result: \begin{thm}\label{jostthm} We assume that $(H_{V})$, $(H_{W,r})$ and $(H_{J}^{0})$ are satisfied. Fix $X>1$. Then, there exist a complex neighborhood $\mathcal{V}=\overline{\mathcal{V}}$ of $J$, a real $\varepsilon_{0}>0$, two points $m_{g}$ and $m_{d}$ in $\mathbb{C}$, two real numbers $A_{g}$ and $A_{d}$ and two functions $(x,\varphi,E,\varepsilon)\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$, $(x,\varphi,E,\varepsilon)\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$ such that: \begin{itemize} \item The functions $(x,\varphi,E,\varepsilon)\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$ and $(x,\varphi,E,\varepsilon)\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$ are defined and consistent on $\mathbb{R}\times S_{Y}\times \mathcal{V}\times]0,\varepsilon_{0}[$. \item For any $x\in[-X,X]$ and $\varepsilon\in]0,\varepsilon_{0}[$, $(\varphi,E)\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$ and $(\varphi,E)\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$ are analytic on $S_{Y}\times \mathcal{V}$. \item The function $x\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$ (resp. $x\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$) is a basis of $\mathcal{J}_{-}$ (resp. $\mathcal{J}_{+}$). \item The functions $h_{-}^{g}$ and $h_{+}^{d}$ have the following asymptotic behavior: \begin{equation} \label{asj1} h_{-}^{g}(x,\varphi,E,\varepsilon)=e^{\frac{-i}{\varepsilon}\int_{m_{g}}^{\varphi}\kappa(u)du}\psi_{-}(x,\varphi,E)(1+R_{g}(x,\varphi,E,\varepsilon)), \end{equation} and \begin{equation} \label{asjda1} h_{+}^{d}(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int_{m_{d}}^{\varphi}\kappa(u)du}\psi_{+}(x,\varphi,E)(1+R_{d}(x,\varphi,E,\varepsilon)), \end{equation} where \begin{itemize} \item $R_{g}$ and $R_{d}$ satisfy: $$\sup\limits_{x\in]-X,X[,\ \mbox{Re }\varphi<A_{g},\\ E\in\mathcal{V}}\max\{|R_{g}(x,\varphi,E,\varepsilon)|,|\partial_{x}R_{g}(x,\varphi,E,\varepsilon)|\}\leq r(\varepsilon).$$ $$\sup\limits_{x\in]-X,X[,\ \mbox{Re }\varphi>A_{d},\\ E\in\mathcal{V}}\max\{|R_{d}(x,\varphi,E,\varepsilon)|,|\partial_{x}R_{d }(x,\varphi,E,\varepsilon)|\}\leq r(\varepsilon),$$ with $$\lim\limits_{\varepsilon\rightarrow 0}r(\varepsilon)=0.$$ \item The functions $\psi_{+}$ and $\psi_{-}$ are the Bloch canonical solutions of the periodic equation \eqref{espa} defined in section \ref{cansolbloch}. \end{itemize} \item There exist two real numbers $\sigma_{g}\in\{-1,1\}$, $\sigma_{d}\in\{-1,1\}$, an integer $p$ and two functions $E\mapsto\alpha_{g}(E)$ and $E\mapsto\alpha_{d}(E)$ such that: \begin{enumerate} \item For any $\varepsilon\in]0,\varepsilon_{0}[$, $x\in\mathbb{R}$, $E\in\mathcal{V}$,et $\varphi\in S_{Y}$ ,we have: \begin{equation} \label{stargj} \alpha_{g}^{*}(E)(h_{-}^{g})^{*}(x,\varphi,E,\varepsilon)=i\sigma_{g}e^{-\frac{i}{\varepsilon}2p\pi x} \alpha_{g}(E)h_{-}^{g}(x,\varphi,E,\varepsilon) \end{equation} \begin{equation} \label{stardj} \alpha_{d}^{*}(E)(h_{+}^{d})^{*}(x,\varphi,E,\varepsilon)=i\sigma_{d}e^{\frac{i}{\varepsilon}2p\pi x} \alpha_{d}(E)h_{+}^{d}(x,\varphi,E,\varepsilon) \end{equation} \item The functions $\alpha_{g}$ and $\alpha_{d}$ are analytic and given by \eqref{renormconstg} and \eqref{renormconstd}. They do not vanish on $\mathcal{V}$. \end{enumerate} \end{itemize} \end{thm} We immediately deduce from Theorem \ref{jostthm} and Proposition \ref{carvp} that the eigenvalues of $H_{\varphi,\varepsilon}$ are characterized by: \begin{equation} w(h_{-}^{g}(\cdot,\varphi,E,\varepsilon),h_{+}^{d}(\cdot,\varphi,E,\varepsilon))=0 \end{equation} Theorem \ref{jostthm} is the consequence of two main ideas: \begin{itemize} \item First, we adapt the construction of Jost functions developed by \cite{Fi1,Ne}. Indeed, this construction proves that asymptotic \eqref{as1} is only satisfied on domains which depend on $\varepsilon$ (see section \ref{scattheory}). \item We must understand how this asymptotic evolves on a domain which does not depend on $\varepsilon$. To do that, we extend the continuation results of Fedotov and Klopp in a non compact frame (see section \ref{infwkb}). \end{itemize} \subsection{Conclusion} To finish the computations, it suffices to apply the methods of \cite{FK1}. We have to go through the cross (see figure \ref{pbf}). We will show that there exists in the neighborhood of the cross a consistent basis $f_{\pm}^{i}$ with standard asymptotic. We shall express the functions $h_{-}^{g}$ and $h_{+}^{d}$ on this basis (section \ref{calcmattransf}). \section{Periodic Schr{\"o}dinger operators on the real line} \label{opepera} We now discuss the periodic operator \eqref{opper} where $V$ is a 1-periodic, real-valued, $\mathbf{L^{2}_{\textrm{loc}}}$-function. We collect known results needed in the present paper (see \cite{Ma, McK, Ti}). \subsection{Bloch solutions} \label{bloch} Let $\widetilde{\Psi}$ be a solution of the equation \begin{equation} \label{espc} H_{0}\widetilde{\Psi}=\mathcal{E}\widetilde{\Psi} \end{equation} satisfying the relation \begin{equation} \widetilde{\Psi}(x+1)=\lambda\widetilde{\Psi}(x),\quad\forall \in\mathbb{R} \end{equation} for some complex number $\lambda\neq 0$ independent of $x$. Such a solution is called a {\it Bloch solution}, and the number $\lambda$ is called the {\it Floquet multiplier}. Let us discuss the analytic properties of Bloch solutions.\\ In \eqref{band}, we have denoted by $[E_{1}, E_{2}],\ldots ,[E_{2n+1}, E_{2n+2}],\ldots$ the spectral bands of the periodic Schr{\"o}dinger equation. Consider $\Gamma_{\pm}$ two copies of the complex plane $\mathcal{E}\in \mathbb{C}$ cut along the spectral bands. Paste them together to get a Riemann surface with square root branch points. We denote this Riemann surface by $\Gamma$.\\ One can construct a Bloch solution $\widetilde{\Psi}(x,\mathcal{E})$ of equation \eqref{espc} meromorphic on $\Gamma$. The poles of this solution are located in the spectral gaps. Precisely, each spectral gap contains precisely one simple pole. This pole is situated either on $\Gamma_{+}$ or on $\Gamma_{-}$. The position of the pole is independent of $x$. For the details, we refer to \cite{Fi1}.\\ Except at the edges of the spectrum (i.e. the branch points of $\Gamma$), the restrictions $\widetilde{\Psi}_{\pm}$ of $\widetilde{\Psi}$ on $\Gamma_{\pm}$ are linearly independent solutions of \ref{espc}. Along the gaps, these functions are real and satisfy : \begin{equation} \label{gaprel} \overline{\widetilde{\Psi}_{\pm}(x,\mathcal{E}-i0)}=\widetilde{\Psi}_{\pm}(x,\mathcal{E}+i0),\quad\forall \mathcal{E}\in]E_{2n},E_{2n+1}[,\ n\in\mathbb{N}. \end{equation} Along the bands, we have : \begin{equation} \label{bandrel} \overline{\widetilde{\Psi}_{\pm}(x,\mathcal{E}-i0)}=\widetilde{\Psi}_{\mp}(x,\mathcal{E}+i0),\quad\forall \mathcal{E}\in]E_{2n+1},E_{2n+2}[,\ n\in\mathbb{N}. \end{equation} \subsection{Bloch quasi-momentum} \subsubsection{} Consider the Bloch solution $\widetilde{\Psi}(x,\mathcal{E})$ introduced in the previous subsection. The corresponding Floquet multiplier $\lambda(\mathcal{E})$ is analytic on $\Gamma$. Represent it in the form: \begin{equation} \lambda(\mathcal{E})=\exp(ik(\mathcal{E})). \end{equation} The function $k(\mathcal{E})$ is called Bloch quasi-momentum. It has the same branch points as $\widetilde{\Psi}(x,\mathcal{E})$, but the corresponding Riemann surface is more complicated. \\ To describe the main properties of $k$, consider the complex plane cut along the real line from $E_{1}$ to $+\infty$. Denote the cut plane by $\mathbb{C}_{0}$. One can fix there a single valued branch of the quasi-momentum by the condition \begin{equation} ik_{0}(\mathcal{E})<0,\quad \mathcal{E}<E_{1}. \end{equation} All the other branches of the quasi-momentum have the form $\pm k_{0}(\mathcal{E})+2\pi m, m\in\mathbb{Z}$. The $\pm$ and the number $m$ are indexing these branches. The image of $\mathbb{C}_{0}$ by $k_{0}$ is located in the upper half of the complex plane. \begin{equation} \mbox{Im } k_{0}(\mathcal{E})>0,\quad \mathcal{E}\in\mathbb{C}_{0}. \end{equation} In figure \ref{qm}, we drew several curves in $\mathbb{C}_{0}$ and their images under transformation $E\mapsto k_{0}(E)$. The quasi-momentum $k_{0}(E)$ is real along the spectral zones, and, along the spectral gaps, its real part is constant; in particular, we have \begin{equation}\label{qm1} k_{0}(E_{1})=0\quad k_{0}(E_{2l}\pm i0)=k_{0}(E_{2l+1}\pm i0)=\pm\pi l,\ \ \ l\in\mathbb{N}. \end{equation} All the branch points of $k$ are of square root type. Let $E_{m}$ be one of the branch points of $k$. Then, each function: \begin{equation} \label{qm2} f_{m}^{\pm}(\mathcal{E})=(k_{0}(\mathcal{E}\pm i0)-k_{0}(E_{m}\pm i0))/\sqrt{\mathcal{E}-E_{m}},\ \ \ E\in\mathbb{R} \end{equation} can be analytically continued in a small vicinity of the branch point $E_{m}$.\\ Finally, we note that \begin{equation} k_{0}(\mathcal{E})=\sqrt{\mathcal{E}}+O(1/\sqrt{\mathcal{E}}),\ \ \ |\mathcal{E}|\rightarrow\infty \end{equation} where $E\in\mathbb{C}_{0}$ and $0<\arg E<2\pi$. The values of the quasi-momentum $k_{0}$ on the two sides of the cut $[E_{1},+\infty)$ are related to each other by the formula: \begin{equation} \label{ko} \forall \mathcal{E}\in]E_{1},+\infty[,\quad k_{0}(\mathcal{E}+i0)=-\overline{k_{0}(\mathcal{E}-i0)},\ \ \ E_{1}\leq \mathcal{E}. \end{equation} Consider the spectral gap $(E_{2l},E_{2l+1}),l \in\mathbb{N}$. Let $\mathbb{C}_{l}$ be the complex plane cut from $-\infty$ to $E_{2l}$ and from $E_{2l+1}$ to $+\infty$. Denote by $k_{l}$ the branch of the quasi-momentum analytic on $\mathbb{C}_{l}$ and coinciding with $k_{0}$ for $\mbox{Im } E >0$. Then, one has: \begin{equation} \label{kl} \forall\mathcal{E}\in]-\infty,E_{2l}[\cup]E_{2l+1},+\infty[,\quad\quad k_{l}(\mathcal{E}+i0)+\overline{k_{l}(\mathcal{E}-i0)}=2\pi l. \end{equation} \input{fig4} \subsection{Periodic components of the Bloch solution} Let $D$ a simply connected domain that does not contain any branch point of $k$. On $D$, we fix an analytic branch of $k$. Consider two copies of $D$, denoted by $D_{\pm}$, corresponding to two sheets of $\mathcal{G}$. Now we redefine $\widetilde{\Psi}_{\pm}$ to be the restrictions of $\widetilde{\Psi}$ to $D_{\pm}$. They can be represented in the form: \begin{equation} \widetilde{\Psi}_{\pm}(x,\mathcal{E})=e^{\pm i k(\mathcal{E})x}p_{\pm}(x,\mathcal{E}),\ \ \ \mathcal{E}\in D \end{equation} where $p_{l}^{\pm}(x,\mathcal{E})$ are 1-periodic in $x$, \begin{equation} p_{\pm}(x+1,\mathcal{E})=p_{\pm}(x,\mathcal{E}),\ \ \ \forall x\in\mathbb{R} \end{equation} \subsection{Analytic solutions of \eqref{espc}} To describe the asymptotic formulas of the complex WKB method for equation \eqref{eqp}, one needs specially normalized Bloch solutions of the equation \eqref{espc}.\\ Let $\mathcal{D}$ be a simply connected domain in the complex plane containing no branch point of the quasi-momentum $k$. We fix on $\mathcal{D}$ a continuous determination of $k$. We fix $\mathcal{E}_{0}\in\mathcal{D}$. We recall the following result (\cite{FK1, FK4}). \begin{lem}\label{anasol} We define the functions $g_{\pm}$ : \begin{equation} \label{fu} g_{\pm}\ :\ \mathcal{D}\rightarrow\mathbb{C} \ ;\ \mathcal{E}\mapsto -\frac{\int_{0}^{1}p_{\mp}(x,\mathcal{E})\partial_{\mathcal{E}}p_{\pm}(x,\mathcal{E})dx}{\int_{0}^{1}p_{+}(x,\mathcal{E})p_{-}(x,\mathcal{E})dx}, \end{equation} and the functions $\psi_{\pm}^{0}$ : \begin{equation} \label{blochsol} \psi_{\pm}^{0}\ :\ \mathbb{R}\times\mathcal{D}\rightarrow\mathbb{C}\ ;\ (x,\mathcal{E})\mapsto\sqrt{k'(\mathcal{E})}e^{\int_{\mathcal{E}_{0}}^{\mathcal{E}}g_{\pm}(e)de}\widetilde{\Psi}_{\pm}(x,\mathcal{E}). \end{equation} The functions $\mathcal{E}\mapsto\psi_{\pm}^{0}(x,\mathcal{E})$ are analytic on $\mathcal{D}$, for any $x\in\mathbb{R}$. The functions $\psi_{\pm}^{0}$ are called {\it analytic Bloch solutions normalized at the point $\mathcal{E}_{0}$} of \eqref{espc}. \end{lem} Sometimes, we shall denote $\psi_{\pm}^{0}(x,\mathcal{E},\mathcal{E}_{0})$ to specify the normalization. We refer to section 1.4.4 of \cite{FK4} for the details of the proof. The proof follows from the study of the poles of $\widetilde{\Psi}_{\pm}$ and the zeros of $k'$. The poles of $g_{\pm}$ are simple and exactly situated at the singularities of $\sqrt{k'}\widetilde{\Psi}_{\pm}$. The computation of the residues of $g_{\pm}$ at these points completes the proof. \subsection{Useful formulas} We end this section with some useful formulas. We recall that the functions $g_{\pm}$ are given in \eqref{fu}. Fix $n\in\mathbb{N}$. Equations \eqref{gaprel} and \eqref{bandrel} lead to the following relations: \begin{equation}\label{symgap} g_{\pm}^{*}(\mathcal{E})=g_{\pm}(x,\mathcal{E}),\quad\forall \mathcal{E}\in]E_{2n},E_{2n+1}[. \end{equation} \begin{equation}\label{symband} g_{\pm}^{*}(\mathcal{E})=g_{\mp}(x,\mathcal{E}),\quad\forall \mathcal{E}\in]E_{2n+1},E_{2n+2}[. \end{equation} \section{Main tools of the complex WKB method} \label{wkbconst} In this section, we recall the main tools of the complex WKB method on compact domains. The idea of the method is to construct some consistent functions of \eqref{eqp} with asymptotic behavior \eqref{stdas}. This construction is not possible on any domain of the complex plane but on some domains called canonical.\\ We apply the results of \cite{FK1, FK2, FK3} to the assumptions $(H_{W,g})$ and $(H_{J})$. We build a neighborhood of the cross, in which we construct a consistent basis with standard behavior \eqref{stdas}. In this section, we fix $Y$ such that the assumptions $(H_{W,g})$ and $(H_{J})$ are satisfied in the strip $S_{Y}$. \subsection{Canonical domains} The canonical domain is the main geometric notion of the complex WKB method. \subsubsection{The complex momentum} \label{vertdef} The canonical domains can be described in terms of the complex momentum $\kappa(\varphi)$. Remind that this function is defined by formula \eqref{momcompa}. We have described $\kappa$ in section \ref{momcompb}. The properties of $\kappa$ depend on the spectral parameter $E$ and of the analytic properties of $W$.\\ We first formulate some definitions (\cite{FK1}). \subsubsection{Vertical, strictly vertical curves} \begin{defn} We say that a curve $\gamma$ is $\textit{vertical}$ if it intersects the lines $\mbox{Im } z=\textrm{Const}$ at non-zero angles $\theta$.\\ We say that a curve $\gamma$ is $\textit{strictly vertical}$ if there is a positive number $\delta$ such that, at any point of $\gamma$, the intersection angle $\theta$ satisfies the inequality: \begin{equation} \delta<\theta<\pi-\delta. \end{equation} \end{defn} \subsubsection{Canonical, strictly canonical curves} \label{lc} Let $\gamma$ be a vertical curve which does not contain any branch point. On $\gamma$, fix a continuous branch of the momentum of $\kappa$. \begin{defn} We call $\gamma$ $\textit{canonical}$ if, along $\gamma$, \begin{itemize} \item $\mbox{Im }\varphi\mapsto\mbox{Im }\int^{\varphi}\kappa(u)du$ is strictly increasing. \item $\mbox{Im }\varphi\mapsto\mbox{Im }\int^{\varphi}(\kappa(u)-\pi)du$ is strictly decreasing. \end{itemize} \end{defn} Assume that $\gamma$ is strictly vertical. If there is a positive number $\delta$ such that, along $\gamma$: \begin{equation} \label{integra} \mbox{Im }\int_{\varphi}^{\varphi'}\kappa(u)du\geq\delta\mbox{Im }(\varphi'-\varphi)\quad\forall(\varphi,\varphi')\in\gamma^{2}, \end{equation} and \begin{equation}\label{integrb} \mbox{Im } \int_{\varphi}^{\varphi'}(\pi-\kappa(u))du\geq\delta\mbox{Im }(\varphi'-\varphi)\quad\forall(\varphi,\varphi')\in\gamma^{2}, \end{equation} we call $\gamma$ $\delta-\textit{strictly canonical}$.\\ We identify the complex numbers with vectors in $\mathbb{R}^{2}$. To construct canonical lines, we have to study the vector fields $\kappa$ and $\kappa-\pi$, or rather their integral curves. For $\varphi\in D$, $S(\varphi)$ denotes the sector of apex $\varphi$ such that, for any vector $z\in S(\varphi)$, we have: \begin{equation} \mbox{Im }(i\overline{\kappa(\varphi)}(z-\varphi))>0\textrm{ et } \mbox{Im }(i(\overline{\kappa(\varphi)}-\pi)(z-\varphi))<0. \end{equation} Let $\gamma\in D$ a curve which does not contain any branch point. For all $\varphi\in\gamma$, we denote $t(\varphi)$ the vector tangent to $\gamma$ in $\varphi$ and oriented upward. The curve $\gamma\in D$ is canonical for the determination $\kappa$ if and only if for any $\varphi\in\gamma$, the vector $t(\varphi)$ belongs to $S(\varphi)$ (see figure \ref{can}). The cone $S(\varphi)$ depends on the determination of $\kappa$. For example, if $\kappa$ satisfies $\mbox{Re }\kappa\in]0,\pi[$, this cone is not empty. \subsubsection{} In what follows, $\xi_{1}$ and $\xi_{2}$ are two points in $\mathbb{C}$ such that $$ \mbox{Im } \xi_{1}<\mbox{Im }\xi_{2}.$$ We shall denote by $\gamma$ a smooth curve going from $\xi_{1}$ to $\xi_{2}$; this curve will always be oriented from $\xi_{1}$ to $\xi_{2}$. \input{fig5} \subsubsection{Definition of the canonical domain} Let $K$ be a simply connected domain in $\left\{\mbox{Im }\varphi\in[\mbox{Im }\xi_{1},\mbox{Im }\xi_{2}]\right\}$ containing no branch points of the complex momentum. On $K$, fix a continuous branch $\kappa$. \begin{defn} We call $K$ a $\textit{canonical domain for }\kappa,\ \xi_{1}\textit{ and }\xi_{2}$ if it is the union of curves that are connecting $\xi_{1}$ and $\xi_{2}$ and that are canonical with respect to $\kappa$.\\ If there is $\delta>0$ such that $K$ is a union of $\delta$-strictly canonical curves, we call $K$ $\delta-\textit{strictly canonical}$. \end{defn} \subsubsection{} Assume that $K$ is a canonical domain. Denote by $\partial K$ its boundary. Fix a positive number $\delta$. We call the domain $$\mathcal{C}=\{z\in K\ ;\ \textrm{dist}(z,\partial K)>\delta\}$$ an admissible sub-domain of $K$.\\ Note that the branch points of the complex momentum are outside of $\mathcal{C}$, at a distance greater than $\delta$. \subsection{Canonical Bloch solutions} \label{cansolbloch}To describe the asymptotic formulas of the complex WKB method for equation \eqref{eqp}, we shall use the analytic Bloch solutions of \eqref{espc}, defined in Lemma \ref{anasol} for the parameter $\mathcal{E}=E-W(\varphi)$. Precisely, we consider the unperturbed periodic equation: \begin{equation} \label{espb} H_{0}\psi=(E-W(\varphi))\psi. \end{equation} \subsection{} Let $D$ be a simply connected domain in $S_{Y}$, containing no branch points of $\kappa$. The mapping $\varphi\mapsto E-W(\varphi)$ maps $D$ onto a domain $\mathcal{D}\subset\mathbb{C}$. The domain $\mathcal{D}$ does not contain any branch point of $k$.\\ Fix $\varphi_{0}\in D$, such that $k'(E-W(\varphi_{0}))\neq 0$. In Lemma \ref{anasol}, we have built the analytic Bloch solutions $\{\psi_{\pm}^{0}\}$ of equation (\ref{espc}), normalized in $E-W(\varphi_{0})$. For $\varphi\in D$, we define: \begin{equation} \psi_{\pm}(x,\varphi,E)=\psi_{\pm}^{0}(x,E-W(\varphi)),\quad\forall u\in\mathbb{R},\quad\forall \varphi\in D. \end{equation} In \cite{FK1}, it is proved that the functions $\varphi\mapsto\psi_{\pm}(x,\varphi,E)$ can be analytically continued to $D$. $\psi_{\pm}$ are called the {\it canonical Bloch solutions} of equation \eqref{espb}. Sometimes, we shall precise $\psi_{\pm}(x,\varphi,E,\varphi_{0})$ to specify the normalization.\\ We define \begin{equation} \label{omega} \omega_{\pm}(\varphi,E)=-W'(\varphi)g_{\pm}(E-W(\varphi)). \end{equation} We also define: \begin{equation} \label{racq} q(\varphi)=\sqrt{k'(E-W(\varphi)} \end{equation} \subsection{The consistency relation} \subsubsection{Consistent functions and consistent bases} We recall that we say that $f$ is a {\it consistent} function if it satisfies \eqref{coh}.\\ We say also that a basis $\{f_{\pm}\}$ of solutions of \eqref{eqp} is {\it a consistent basis} if: \begin{itemize} \item The functions $f_{+}$ and $f_{-}$ are consistent. \item Their Wronskian is independent of $\varphi$. \end{itemize} \subsubsection{Analyticity and consistency} First, we define the width of a set. \begin{defn} Fix $Y_{0}>0$ and $M\subset S_{Y_{0}}$ a set of points. We define $l(M,Y_{0})$: \begin{equation} \label{larga} l(M,Y_{0})=\inf\limits_{y\in[-Y_{0},Y_{0}]}\sup\left\{|\mbox{Re }\varphi-\mbox{Re }\varphi'| ; (\varphi,\varphi')\in M^{2}\textrm{ such that } \mbox{Im }\varphi=\mbox{Im }\varphi'=y \right\} \end{equation} $l(M,Y_{0})$ is called the {\it width} of $M$ in $S_{Y_{0}}$ \end{defn} One has: \begin{lem} \label{anacoh} Fix $E$. We consider $X>0$, $\tilde{Y}\in]0,Y[$, $\varepsilon_{0}>0$ and $K$ a complex domain such that $l(K,\tilde{Y})>\varepsilon_{0}$. We assume that for any $\varepsilon\in]0,\varepsilon_{0}[$, $f(\cdot,\varphi,E,\varepsilon)$ is a consistent solution of \eqref{eqp} for $\varphi\in K$ and that for any $x\in[-X,X]$, the function $\varphi\mapsto f(x,\varphi,E,\varepsilon)$ is analytic on $K$. Then, for any $\varepsilon\in]0,\varepsilon_{0}[$ and any $x\in[-X,X]$, the function $\varphi\mapsto f(x,\varphi,E,\varepsilon)$ is analytic on $S_{\tilde{Y}}$. \end{lem} This result is proved in \cite{FK3, FK4}. \subsection{The theorem of the complex WKB method on a compact domain} \label{wkbth} In this section, we recall the main result of the complex WKB method. \subsubsection{Standard asymptotic behavior} \label{cptmtasstd} We briefly introduce the notion of standard asymptotic behavior (see \cite{FK4}). Speaking about a solution having standard asymptotic behavior, we mean first of all that this solution has the asymptotics \eqref{stdas} and other properties that we present now.\\ Fix $E_{0}\in\mathbb{C}$. Let $D\subset\mathbb{C}$ a simply connected domain containing no branch points. Let $\kappa$ be a branch of the complex momentum continuous in $D$ and $\psi_{\pm}$ the canonical Bloch solutions normalized in $\varphi_{0}\in D$.\\ We say that a consistent solution $f$ has standard behavior $f\sim e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}\psi_{+}(x,\varphi,E)$, respectively $f\sim e^{-\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}\psi_{-}(x,\varphi,E)$ in $D$ if \begin{itemize}\item there exists a complex neighborhood $\mathcal{V}_{0}$ of $E_{0}$ and $X>0$ such that $f$ is a consistent solution of equation \eqref{eqp} for any $(x,\varphi,E)\in[-X,X]\times D\times \mathcal{V}_{0}$; \item for any $x\in[-X,X]$, the function $((\varphi,E)\mapsto f(x,\varphi,E,\varepsilon))$ is analytic on $D\times \mathcal{V}_{0}$; \item for any $A$, a sub-admissible domain of $D$, there is a neighborhood $\mathcal{V}_{A}$ of $E_{0}$ such that \begin{equation} \label{cptmtasstda} \forall(x,\varphi,E)\in[-X,X]\times D\times \mathcal{V}_{A},\quad f(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi,E)+o(1)),\quad\varepsilon\rightarrow 0 \end{equation} respectively \begin{equation} \label{cptmtasstdb} \forall(x,\varphi,E)\in[-X,X]\times D\times \mathcal{V}_{A},\quad f(x,\varphi,E,\varepsilon)=e^{-\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{-}(x,\varphi,E)+o(1)),\quad\varepsilon\rightarrow 0 \end{equation} \item the asymptotics \eqref{cptmtasstda} and \eqref{cptmtasstdb} are uniform on $[-X,X]\times D\times \mathcal{V}_{A}$; \item the asymptotics \eqref{cptmtasstda} and \eqref{cptmtasstdb} can be differentiated once in $x$. \end{itemize} \subsubsection{} Let us formulate the Theorem WKB on a compact domain. \begin{thm}\cite{FK1,FK4}\\ \label{finwkbthm} We assume that $V$ satisfies $(H_{V})$ and that $W$ satisfies $(H_{W,r})$. Fix $X>1$ and $E_{0}\in\mathbb{C}$. Let $K\subset S_{Y}$ be a bounded canonical domain with respect to $\kappa$. There exists $\varepsilon_{0}>0$ and a consistent basis $\{f_{+}(x,\varphi,E,\varepsilon),f_{-}(x,\varphi,E,\varepsilon)\}$ of solutions of \eqref{eqp}, having the standard behavior \eqref{cptmtasstda} et \eqref{cptmtasstdb} in $K$.\\ For any fixed $x\in\mathbb{R}$, the functions $\varphi\mapsto f_{\pm}(x,\varphi,E,\varepsilon)$ are analytic in $K$. \end{thm} \subsection{The main geometric tools of the complex WKB method} \label{constgeo} In this section, we introduce the main geometric tools of the complex WKB method. To do that, we recall some ideas of \cite{Fe, FK1, FK3, Wa}. \subsubsection{Stokes lines} The definition of the Stokes lines is fairly standard, \cite{Fe,FK1}. The integral $\varphi\mapsto\int^{\varphi}\kappa(u)du$ has the same branch points as the complex momentum. Let $\varphi_{0}$ be one of them. Consider the curves beginning at $\varphi_{0}$, and described by the equation \begin{equation} \mbox{Im }\int_{\varphi_{0}}^{\varphi}(\kappa(\xi)-\kappa(\varphi_{0}))d\xi=0 \end{equation} These curves are the {\it Stokes lines} beginning at $\varphi_{0}$. According to equation \eqref{ko} and equation \eqref{kl}, the Stokes line definition is independent of the choice of the branch of $\kappa$.\\ Assume that $W'(\varphi_{0})\neq 0$. Equation \eqref{qm2} implies that there are exactly three Stokes lines beginning at $\varphi_{0}$. The angle between any two of them at this point is equal to $\frac{2\pi}{3}$. \subsection{Lines of Stokes type} \label{ligtypsto} We recall that $D\subset S_{Y}$ is a simply connected domain containing no branch points. Let $\gamma\subset D$ be a smooth curve. We say that $\gamma$ is a line of Stokes type with respect to $\kappa$ if, along $\gamma$, we have $$\textrm{ either }\mbox{Im }\left(\int^{\varphi}\kappa(u) du\right)=\textrm{Const}\quad\textrm{ or }\quad\mbox{Im }\left(\int^{\varphi}(\kappa(u)-\pi) du\right)=\textrm{Const}$$ \subsection{Pre-canonical lines} Let $\gamma\subset D$ be a vertical curve. We call $\gamma$ $\textit{pre-canonical}$ if it consists of union of bounded segments of canonical curves and/or lines of Stokes type. \subsection{Some branches of the complex momentum} In this section, we describe different branches of $\kappa$ near the branch points described in \ref{assJ}. The geometrical configuration is similar to the one studied in \cite{FK2}. \subsubsection{Different cases} \label{poss} We assume that $(H_{W,r})$, $(H_{W,g})$, and $(H_{J})$ are satisfied. To study the geometrical tools of the WKB complex method, one needs to specify the properties of $\mbox{Im }\kappa$ and $\mbox{Re }\kappa$. We know that $\kappa(\varphi_{r}^{\pm})\equiv 0[\pi]$ (see section \ref{opepera}). We consider two cases: either $\kappa(\varphi_{r}^{\pm})\equiv 0[2\pi]$ or $\kappa(\varphi_{r}^{\pm})\equiv \pi[2\pi]$.\\ We define $S_{-}$ the open domain delimited by the real line at the bottom and by $\Sigma_{+}$ to the right: \begin{equation} \label{smoins} S_{-}=\{\varphi-r\ ;\ \varphi\in\Sigma_{+}^{*},r\in\mathbb{R}_{+}^{*}\}\cap S_{Y} \end{equation} Similarly, we define $S_{+}$ the open domain delimited by the real line at the bottom and by $\Sigma_{+}$ to the left: \begin{equation} \label{splus} S_{+}=\{\varphi+r\ ;\ \varphi\in\Sigma_{+}^{*},r\in\mathbb{R}_{+}^{*}\}\cap S_{Y} \end{equation} The domains $S_{+}$ and $S_{-}$ are shown in figure \ref{smoinsa}.\\ We prove the following result. \begin{lem} \label{detpos}\label{dtepos}There exists a branch $\kappa_{i}$ of the complex momentum such that \begin{enumerate} \item $\mbox{Im }\kappa_{i}(\varphi)>0$ for $\varphi\in S_{-}$, $\kappa_{i}(\varphi_{r}^{-}+i0)=0$ and $\kappa_{i}(\varphi_{i}-0)=\pi$,\\ ou \item $\mbox{Im }\kappa_{i}(\varphi)<0$ for $\varphi\in S_{-}$, $\kappa_{i}(\varphi_{r}^{-}+i0)=\pi$ and $\kappa_{i}(\varphi_{i}-0)=0$. \end{enumerate} \end{lem} \begin{dem} \begin{itemize}\item First, we specify the sign of $\mbox{Im }\kappa_{i}$. The set $(E-W)(\mathbb{R}-[\varphi_{r}^{-},\varphi_{r}^{+}]))$ belongs to a gap $G$. We define $\Lambda_{-}=(E-W)(S_{-})$. We prove that $\Lambda_{-}$ is a connected domain which intersects with $\mathbb{R}$ only in the gap $G$. According to assumption $(H_{W,g})$ (\ref{assW2}), there exists a sequence of vertical curves $\widetilde{\Sigma}_{k}$ such that: $$\Lambda_{-}\cap\mathbb{R}=(E-W)((-\infty,\varphi_{r}^{-}]\cup[\varphi_{r}^{+},+\infty))\cup(E-W)(\widetilde{\Sigma}_{k}^{+}).$$ $(E-W)(\widetilde{\Sigma}_{k}^{+})$ is a connected domain of $\mathbb{R}$; it contains at least a point of $G$ and does not intersect with $\partial\sigma(H_{0})$. Consequently, $(E-W)(\widetilde{\Sigma}_{k}^{+})$ belongs to $G$ and: $$\Lambda_{-}\cap\mathbb{R}=G.$$ We fix on $\Lambda_{-}$ a continuous branch of the quasi momentum $k$. The sign of $\mbox{Im } k$ does not change since $(E-W)(S_{-})$ does not intersect with $\sigma(H_{0})$. If we define $\kappa_{i}(\varphi)=k(E-W(\varphi))$, $\mbox{Im }\kappa_{i}$ we can assume that $\mbox{Im }\kappa_{i}>0$ on $S_{-}$. \item Now, we consider $\mbox{Re }\kappa_{i}$. According to section \ref{qm}, we can choose the branch $\kappa_{i}$ such that $\kappa_{i}(\varphi_{r}^{-}+i0)\in\{0,\pi\}$. First, we study the case $\kappa_{i}(\varphi_{r}^{-})=0$; this assumption implies two possibilities. \begin{enumerate}\item The point $0$ is a minimum for $W$ and the band $B$ in $(H_{J})$ is in the form $[E_{4l+1},E_{4l+2}]$. The points $E_{r}$ and $E_{i}$ satisfy $E_{r}=E_{4l+1}$ and $E_{i}=E_{4l+2}$. There exists a neighborhood $V$ of $[\varphi_{r}^{-},0]\cup\sigma$ such that $(E-W)(S_{-}\cap V)\subset\mathbb{C}_{+}\backslash\mathbb{R}$. Actually, in the neighborhood of $0$, we have $\mbox{Im }(E-W(\varphi))\geq 0$; according to $(H_{W,g})$, there exists a neighborhood $V$ of $[\varphi_{r}^{-},0]\cup\sigma$ such that $(E-W)(S_{-}\cap V)$ does not intersect $\mathbb{R}$. By continuity of the mapping $\varphi\mapsto\mbox{Im }(E-W(\varphi))$, the sign of $\mbox{Im }(E-W(\varphi))$ remains positive on $S_{-}\cap V$. There exists a branch $k$ of the quasi-momentum such that $$\mbox{Im } k(\mathcal{E})>0 \textrm{ for }\mbox{Im }\mathcal{E}>0 \textrm{ and } k(E_{n}+i0)=0,\ k(E_{p}+i0)=\pi.$$ We define $\kappa_{i}(\varphi)=k(E-W(\varphi))$. \item The point $0$ is a maximum for $W$ and the band $B$ is in the form $[E_{4l+3},E_{4l+4}]$; the points $E_{r}$ and $E_{i}$ satisfy $E_{r}=E_{4l+4}$ and $E_{i}=E_{4l+3}$. Let $k$ be the branch of the quasi-momentum such that $\mbox{Im } k(\mathcal{E})>0$ for $\mbox{Im }\mathcal{E}<0$, and $k(E_{n})=0$; then $k(E_{p})=\pi$. We define $\kappa_{i}(\varphi)=k(E-W(\varphi))$. \end{enumerate} The case $\kappa_{i}(\varphi_{r}^{-})=\pi$ is similar. \end{itemize} This completes the proof of the lemma. \end{dem}\\ For the sake of clarity, for all the proofs, we shall consider the case: \begin{equation} \label{premcassc} \kappa_{i}(\varphi_{r}^{-}+i0)=0\textrm{ et }\kappa_{i}(\varphi_{i}-0)=\pi \end{equation} The arguments in the second case are similar and we will not give the details. \subsubsection{Other branches of the complex momentum} \label{compmom} \label{detsc} The properties of the complex momentum near the branch points $\varphi_{i},\ \overline{\varphi_{i}},\ \varphi_{r}^{\pm}$ are determined by the behavior of $k$ near $E_{r}$ and $E_{i}$.\\ Now, we describe other branches of $\kappa$, which are obtained from the branch $\kappa_{i}$ (Lemma \ref{dtepos}) by analytic continuation. We consider the case \eqref{premcassc}. \begin{itemize} \item We denote by $\kappa_{g}$ the continuation of $\kappa_{i}$ to the domain $\{\mbox{Re }(\varphi)<\mbox{Re }(\varphi_{r}^{-})\}$. $\kappa_{g}$ satisfies $$\begin{array}{c}\mbox{Im }(\kappa_{g}(\varphi))>0\textrm{ for }\{\mbox{Re }(\varphi)<\mbox{Re }(\varphi_{r}^{-})\}\\\mbox{Re }(\kappa_{g})(\varphi)\rightarrow 0\textrm{ as }\mbox{Re }(\varphi)\rightarrow -\infty. \end{array}$$ $\kappa_{g}$ is the continuation of $\kappa_{i}$ through $(-\infty,\varphi_{r}^{-}]$. \item We consider the strip $S_{Y}$ cut along $(\Sigma\backslash\sigma)\cup(\overline{\Sigma}\backslash\overline{\sigma})\cup(-\infty,\varphi_{r}^{-})\cup(\varphi_{r}^{+},+\infty)$. We always denote by $\kappa_{i}$ the continuation of $\kappa_{i}$ through $C$. \item On $\{\mbox{Re }(\varphi)>\mbox{Re }(\varphi_{r}^{+})\}$, we fix a continuous branch $\kappa_{d}$ with the conditions: $$\begin{array}{c}\mbox{Im }(\kappa_{d}(\varphi))>0\textrm{ for }\{\mbox{Re }(\varphi)>\mbox{Re }(\varphi_{r}^{+})\}\\\mbox{Re }(\kappa_{d})(\varphi)\rightarrow 0\textrm{ as }\mbox{Re }(\varphi)\rightarrow +\infty. \end{array}$$ $\kappa_{d}$ is the continuation of $\kappa_{i}$ through $\overline{S_{+}}$. \end{itemize} Here, we describe the behavior of the different branches of $\kappa$. \begin{equation} \label{kgki} \forall\varphi\in S_{-},\quad\kappa_{g}(\varphi)=\kappa_{i}(\varphi)\quad;\quad\forall\varphi\in\overline{S_{-}},\quad\kappa_{g}(\varphi)=-\kappa_{i}(\varphi). \end{equation} \begin{equation} \label{kdki} \forall\varphi\in S_{+},\quad\kappa_{d}(\varphi)=-\kappa_{i}(\varphi)\quad;\quad\forall\varphi\in\overline{S_{+}},\quad\kappa_{d}(\varphi)=\kappa_{i}(\varphi). \end{equation} \input{fig6} \subsection{Stokes lines} \label{stline} This section is devoted to the description of the Stokes lines under assumptions $(H_{W,g})$ and $(H_{J})$. We describe the Stokes lines beginning at $\varphi_{r}^{-}$, $\varphi_{r}^{+}$, $\varphi_{i}$ and $\overline{\varphi_{i}}$. Since $W$ is real on the real line, the set of the Stokes lines is symmetric with respect to the real line.\\ First, $\kappa_{i}$ is real on the interval $[\varphi_{r}^{-},\varphi_{r}^{+}]\subset\mathbb{R}$; therefore, $[\varphi_{r}^{-},\varphi_{r}^{+}]$ is a part of a Stokes line beginning at $\varphi_{r}^{-}$. The two other Stokes lines beginning at $\varphi_{r}^{-}$ are symmetric with respect to the real line. We denote by $b$ the Stokes line going upward and by $\bar{b}$ its symmetric. Similarly, we denote by $a$ and $\bar{a}$ the two other Stokes lines beginning at $\varphi_{r}^{+}$; $a$ goes upwards.\\ Consider the Stokes lines beginning at $\varphi_{i}$. The angles between the Stokes lines at this point are equal to $2\pi/3$. So, one of the Stokes lines is situated between $\Sigma$ and $e^{\frac{2i\pi}{3}}\Sigma$. It is locally going to the right of $\Sigma$; we denote by $d$ this line. Similarly, we denote by $e$ the Stokes line between $\Sigma$ and $e^{-\frac{2i\pi}{3}}\Sigma$. Finally, we denote by $c$ the third Stokes line beginning at $\varphi_{i}$; $c$ is going upwards.\\ By symmetry, we denote by $\bar{c}$, $\bar{d}$ and $\bar{e}$ the Stokes lines beginning at $\overline{\varphi_{i}}$.\\ We describe the behavior of $a$, $b$, $c$, $d$ and $e$ in the strip $S_{Y}$. We have represented these lines in figure \ref{ls}. In this figure, we have precised the values of $\kappa$ in the branch points. \begin{lem} \label{stlinea} We assume that $V$, $W$ and $J$ satisfy $(H_{V})$, $(H_{W})$ and $(H_{J})$. Then, the Stokes lines described in figure \ref{ls} have the following properties: \begin{enumerate} \item $a$ stays vertical; it intersects $\{\mbox{Im }(\varphi)=Y\}$. \item $b$ stays vertical; it intersects $\{\mbox{Im }(\varphi)=Y\}$. \item $d$ intersects $a$ above $\varphi_{r}^{+}$; the segment between $\varphi_{i}$ and this intersection with $a$ is vertical. \item $e$ intersects $b$ above $\varphi_{r}^{-}$; the segment between $\varphi_{i}$ and this intersection with $b$ is vertical. \item $c$ stays vertical; it intersects $\{\mbox{Im }(\varphi)=Y\}$ and does not intersect $\sigma$. \item $a$ and $c$ do not intersect one another in the strip $S_{Y}$. \item $b$ and $c$ do not intersect one another in the strip $S_{Y}$. \end{enumerate} \end{lem} \begin{dem} First, we note that a Stokes line can become horizontal only at a point where $\mbox{Im } \kappa=0$, i.e. at a point of the pre-image of a spectral band. Besides, a Stokes line beginning at $\varphi_{1}^{\pm}$ (respectively at $\varphi_{2}$ or $\overline{\varphi_{2}}$) is locally orthogonal to $i\ \overline{\kappa(\varphi)}$ (respectively $i\ \overline{(\pi-\kappa(\varphi))}$).\\ We first prove 1). According $(H_{_{J}})$, the pre-image of the spectrum is $[\varphi_{r}^{-},\varphi_{r}^{+}]\cup\sigma$. So, $a$ becomes horizontal only if it intersects $\sigma$. Let us prove by contradiction that it is impossible. Let us assume that $a$ intersects $\sigma$ in $\varphi_{a}$, then: $$\mbox{Im }\int_{\varphi_{r}^{+}}^{\varphi_{a}}\kappa(u)du=0=\mbox{Im }\int_{0,\textrm{ along }\sigma}^{\varphi_{a}}\kappa(u)du$$ $$=\int_{0}^{\varphi_{a}}(\mbox{Re }\kappa(u))d(\mbox{Im }(u))\leq-k_{1}(E-W_{-})\mbox{Im }\varphi_{a}<0$$ which is impossible. Therefore, $a$ stays vertical. Moreover, as $\varphi\rightarrow\infty,\ \varphi\in S_{Y},\ \mbox{Im }(i\bar{\kappa})\rightarrow 0$. Thus, $a$ admits a vertical asymptote and intersects $\{\mbox{Im }(\varphi)=Y\}$.\\ Similarly, we prove 2).\\ To prove 3), we consider the Stokes line $d$. If $a$ and $d$ do not intersect one another, then $d$ intersect either $\sigma$ or $[0,\varphi_{r}^{+}]$. In this case, we denote by $\varphi_{d}$ the intersection between $d$ and $\sigma$ and we have: $$\mbox{Im }\int_{\varphi_{d}}^{\varphi_{i}}(\kappa(u)-\pi)du=0=\int_{\varphi_{d}}^{\varphi_{i}}\mbox{Re }(\kappa(u)-\pi)d(\mbox{Im }(u))<0$$ Consequently, $d$ and $a$ do not intersect one another. Before its intersection with $a$, $d$ does not intersect the pre-image of a spectral band and it stays vertical. We prove similarly the properties of $e$.\\ We prove now 5). $c$ is going upwards. $c$ does not intersect the pre-image of a spectral band in $\{\mbox{Im }\varphi\in]\mbox{Im }\varphi_{i},Y[\}$ and $c$ stays vertical.\\ We prove 6) by contradiction. Let us assume that there is $\varphi_{a}\in a\cap c$. Then, we compute: $$\mbox{Im }\int_{0}^{\varphi_{a}}\kappa(u)du=0=\mbox{Im }\int_{\sigma}\kappa(u)du+\mbox{Im }\int_{\varphi_{i}}^{\varphi_{a}}\kappa(u)du$$ First, $\mbox{Im }\int_{\varphi_{i}}^{\varphi_{a}}\kappa(u)du=\pi\mbox{Im }(\varphi_{a}-\varphi_{i})>0$ and $\mbox{Im }\int_{\sigma}\kappa(u)du=\int_{\sigma}\mbox{Re }\kappa(u)d(\mbox{Im } u)>0$.\\ which is impossible. So, $a$ and $c$ do not intersect one another in $S_{Y}$. \end{dem}\\ In the following, we choose $\widetilde{Y}\in]\sup\limits_{E\in J}\mbox{Im }\varphi_{i}(E),Y[$. \input{fig7} \subsection{Construction of a consistent basis with standard behavior in the neighborhood of the cross} In this section, we begin with constructing a canonical line near the cross. To do that, we follow the methods developed in \cite{FK4}. \subsubsection{General constructions} We first recall some general geometric tools presented in \cite{FK4}, section 4.1. \begin{itemize} \item We first introduce the idea of enclosing canonical domain. \begin{defn} Let $\gamma\subset D$ be a line canonical with respect to $\kappa$. Denote its ends by $\xi_{1}$ and $\xi_{2}$. Let a domain $K\subset D$ be a canonical domain corresponding to the triple $\kappa,\ \xi_{1}$ and $\xi_{2}$. If $\gamma\subset K$, then $K$ is called a canonical domain {\it enclosing $\gamma$}. \end{defn} We have the following property: \begin{lem}\cite{FK2}\\ \label{enccan} One can always construct a canonical domain enclosing any given compact canonical curve located in an arbitrarily small neighborhood of that curve. \end{lem} Such canonical domains, whose existence is established using this lemma are called {\it local}. \item To construct a canonical domain, we need a canonical line to start with. To construct such a line, we first build pre-canonical lines made of some ``elementary'' curves. Let $\gamma\subset D$ be a vertical curve. We call $\gamma$ {\it pre-canonical} if it is a finite union of bounded segments of canonical lines and/or lines of Stokes type. The interest of pre-canonical curves is the following: \begin{lem}\cite{FK2}\\ \label{precan} Let $\gamma$ be a pre-canonical curve. Denote the ends of $\gamma$ by $\xi_{a}$ and $xi_{b}$. Fix $V\subset D$, a neighborhood of $\gamma$ and $V_{a}\subset D$ a neighborhood of $\xi_{a}$. Then, there exists a canonical line $\tilde{\gamma}$ connecting the point $\xi_{b}$ to some point in $V_{a}$. \end{lem} \end{itemize} \subsubsection{Constructing a canonical line near the cross} Here, we mimic the construction of \cite{FK4}, section 4.2. We assume that assumptions $(H_{V})$, $(H_{W,r})$, $(H_{W,g})$ and $(H'_{J})$ are satisfied. We now explain the construction of a canonical line going from $\{\mbox{Im }\xi=-Y\}$ to $\{\mbox{Im }\xi=Y\}$. First, we consider the curve $\beta$ which is the union of the Stokes line $\bar{b}$, the segment $[\varphi_{r}^{-},0]$ of the real line, the closed curve $\sigma_{+}$ and the Stokes line $c$.\\ We now construct $\alpha$ a pre-canonical line close to the line $\beta$. We prove: \begin{prop} \label{cancurva} Fix $\delta>0$. In the $\delta$-neighborhood of $\beta$, there exists $\alpha$ a pre-canonical line with respect to the branch $\kappa$ connecting $\xi_{1}$ to $\xi_{2}$ and having the following properties: \begin{itemize} \item at its upper end, $\mbox{Im }\xi_{2}=Y$, \item at its lower end, $\mbox{Im }\xi_{1}=-Y$, \item it goes around the branch points of the complex momentum as the curve shown in figure \ref{cancurve}; \item it contains a canonical line which stays in $S_{-}$, goes downward from a point in $S_{-}$ to the curve $\sigma$ and then continues along this curve until it intersects the real line. \end{itemize} \end{prop} \begin{dem} The proof of this Proposition is completely similar to the proof of Proposition 4.2 in \cite{FK4}. It consists in breaking down $\alpha$ in ``elementary'' segments. We do not give the details. \end{dem}\\ An immediate consequence of Proposition \ref{cancurva} is the following result: \begin{prop} \label{cancurvb} In arbitrarily small neighborhood of the pre-canonical line $\alpha$, there exists a canonical line $\gamma$ which has all the properties of the line $\alpha$ listed in Proposition \ref{cancurva}. \end{prop} \input{fig8} \subsubsection{Some continuation tools} In this section, we recall some continuation tools; these tools are developed in \cite{FK4}. \begin{enumerate} \item Now, we present the continuation lemma on compact domains. We recall that $q$ is defined in \eqref{racq}. \begin{lem} \label{lemcontfin} \cite{FK1} Let $\varphi_{-}, \varphi_{+}, \varphi_{0}$ be fixed points such that \begin{itemize} \item $\mbox{Im }\varphi_{-}=\mbox{Im }\varphi_{+}$; \item there is no branch point of $\varphi\mapsto\kappa(\varphi)$ on the interval $[\varphi_{-}, \varphi_{+}]$; \item $\varphi_{0}\in(\varphi_{-} \varphi_{+}), q(\varphi_{0})\neq 0.$ \end{itemize} Fix a continuous branch of $\kappa$ on $[\varphi_{-}, \varphi_{+}]$. Let $f(x,\varphi,E,\varepsilon)$, $ f _{\pm}(x,\varphi,E,\varepsilon)$ be solutions of \eqref{eqp} for $\varphi\in[\varphi_{-}, \varphi_{+}]$ and $x\in[-X,X]$ satisfying \eqref{coh} and such that: \begin{enumerate} \item $f(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int_{\varphi_{0}}^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi,E)+o(1))$ pour $\varphi\in[\varphi_{-}, \varphi_{0}]$ for $\varphi\in[\varphi_{-}, \varphi_{0}]$ when $\varepsilon\rightarrow 0$ and the asymptotic is differentiable in $x$; \item $f_{\pm}(x,\varphi,E,\varepsilon)=e^{\pm\frac{i}{\varepsilon}\int_{\varphi_{0}}^{\varphi}\kappa(u)du}(\psi_{\pm}(x,\varphi,E)+o(1))$ for $\varphi\in[\varphi_{-}, \varphi_{+}]$ when $\varepsilon\rightarrow 0$, and the asymptotic is differentiable in $x$. \end{enumerate} Here, $\psi_{\pm}$ are canonical Bloch solutions associated to the complex momentum $\kappa$.\\ Then,\\ \begin{itemize} \item if $\mbox{Im }(\kappa(\varphi))>0$ for all $\varphi\in [\varphi_{-}, \varphi_{+}]$, there exists $C>0$ such that, for $\varepsilon>0$ small enough, \begin{equation} \left|\frac{df}{dx}(x,\varphi,E,\varepsilon)\right|+|f(x,\varphi,E,\varepsilon)|\leq C e^{\frac{1}{\varepsilon}\int_{\varphi}^{\varphi_{0}}|\mbox{Im }\kappa(u)|du},\quad\varphi\in [\varphi_{0}, \varphi_{+}]; \end{equation} \item if $\mbox{Im }(\kappa(\varphi))<0$ for all $\varphi\in[\varphi_{-}, \varphi_{+}]$, then \begin{equation} f(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int_{\varphi_{0}}^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi,E)+o(1)),\quad\varphi\in[\varphi_{0}, \varphi_{+}], \end{equation} and the asymptotic is differentiable in $x$. \end{itemize} \end{lem} Intuitively, this lemma means that a function $f$ has the standard behavior along a horizontal line as long as the leading term of its asymptotics is growing along that line. For analogous results with real WKB method, we refer to \cite{Vo}. \item The estimate we obtained in Lemma \ref{lemcontfin} can be far from optimal. The Adjacent Canonical Domain Principle gives a more precise result: \begin{prop}\cite{FK3} \label{adjdom} Assume that a solution $f$ has standard behavior in either the left hand side or the right hand side of a constant neighborhood of a vertical curve $\gamma$. Assume that $\gamma$ is canonical with respect to some branch of the complex momentum. Then $f$ has standard behavior in any bounded canonical domain enclosing $\gamma$. \end{prop} \item The last tool we shall need in the sequel is the Stokes Lemma.\\ Notations and assumptions:\\ Assume that $\xi_{0}$ is a branch point of the complex momentum such that $W'(\xi_{0})\neq 0$. There are three Stokes lines beginning at $\xi_{0}$. The angles between them at $\xi_{0}$ are equal to $2\pi/3$. We denote these lines by $\sigma_{1}$, $\sigma_{2}$ and $\sigma_{3}$, so that $\sigma_{1}$ is vertical at $\xi_{0}$. Let $V$ be a neighborhood of $\xi_{0}$; assume that $V$ is so small that $\sigma_{1}$, $\sigma_{2}$ and $\sigma_{3}$ divide it into three sectors. We denote them by $S_{1}$, $S_{2}$ and $S_{3}$ so that $S_{1}$ be situated between $\sigma_{1}$ and $\sigma_{2}$, and the sector $S_{2}$ be between $\sigma_{2}$ and $\sigma_{3}$ (see figure \ref{lsbp}).\\ We recall now the result: \begin{lem}\cite{FK4}.\label{stoklemma} Let $V$ be sufficiently small. Let $f$ be a solution that has standard behavior $f=e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi)+o(1))$ inside the sector $S_{1}\cup\sigma_{2}\cup S_{2}$ of $V$. Moreover, assume that, in $S_{1}$ near $\sigma_{1}$, one has $\mbox{Im }\kappa>0$ if $S_{1}$ is to the left of $\sigma_{1}$ and $\mbox{Im }\kappa<0$ otherwise. Then, $f$ has standard behavior $f=e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi)+o(1))$ inside $V\backslash\sigma_{1}$, the asymptotics being obtained by analytic continuation from $S_{1}\cup\sigma_{2}\cup S_{2}$ to $V\backslash\sigma_{1}$. \end{lem} \end{enumerate} \input{fig9} \subsubsection{Construction of a basis with standard asymptotic behavior near the cross} We prove the existence of a consistent basis with standard asymptotic behavior near the canonical line $\alpha$. Let $\alpha$ be the curve described in Proposition \ref{cancurvb}. According to Lemma \ref{enccan}, we can construct a local canonical domain $K_{i}$ enclosing $\alpha$. \begin{prop} \label{bcki} Assume that $(H_{V})$, $(H_{W,r})$, $(H_{W,g})$ and $(H_{J})$ are satisfied. Fix $E_{0}\in J$, $X>1$ and $\tilde{Y}\in]0,\tilde{Y}[$. Then, there exist a complex neighborhood $\mathcal{U}_{0}$ of $E_{0}$, a real number $\varepsilon_{0}>0$ and a function $f_{i}$ satisfying the following properties: \begin{itemize} \item The function $(x,\varphi,E,\varepsilon)\mapsto f_{i}(x,\varphi,E,\varepsilon)$ is defined on $\mathbb{R}\times S_{\tilde{Y}}\times\mathcal{U}_{0}\times]0,\varepsilon_{0}[$. \item For any $x\in\mathbb{R}$, for any $\varepsilon\in]0,\varepsilon_{0}[$, the function $((\varphi,E)\mapsto f_{i}(x,\varphi,E,\varepsilon))$ is analytic on $S_{\tilde{Y}}\times\mathcal{U}_{0}$. \item For $(x,\varphi,E)\in [-X,X]\times K_{i}\times\mathcal{U}_{0}$, the function $f_{i}$ has the asymptotic behavior: \begin{equation} \label{asca} f_{i}(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int_{0}^{\varphi}\kappa(u) du}\left(\psi_{\pm}(x,\varphi,E)+o(1)\right),\quad\varepsilon\rightarrow 0. \end{equation} \item The asymptotics \eqref{asca} are uniform in $(x,\varphi,E)\in [-X,X]\times K_{i}\times\mathcal{U}_{0}$. \item The asymptotics \eqref{asca} can be differentiated once in $x$. \item There exists a real number $\sigma_{i}\in\{-1,1\}$ such that the function $f_{i}$ satisfies the relation: $$w(f_{i},f_{i}^{*})=w(f_{i}(\cdot,\varphi,E,\varepsilon),f_{i}(\cdot,\overline{\varphi},\overline{E},\varepsilon))=\sigma_{i}(k'_{i}w_{i})(E-W(0))$$ \end{itemize} \end{prop} The end of this section is devoted to the proof of Proposition \ref{bcki}. This Proposition mainly follows from Theorem \ref{finwkbthm}. \subsubsection{Existence of $f_{i}$} The domain $K_{i}$ is a local canonical domain. According to Theorem \ref{finwkbthm}, we can build a function $f_{i}$ such that, on $K_{i}$, $f_{i}$ has the following asymptotic behavior: $$f_{i}\sim e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa_{i}}\psi_{+}.$$ Let us normalize $f_{i}$ in $0$. \subsubsection{Computation of the Wronskian $w(f_{i},f_{i}^{*})$} \label{precsigma} To finish the proof of Proposition \ref{bcki}, it remains to compute $w(f_{i},f_{i}^{*})$.\\ Let $R_{+}$ be a small enough rectangle to the left of $\alpha_{+}$, so that $R_{+}\subset K_{i}\cap S_{\tilde{Y}}$. We define $R=R_{+}\cup R_{-}$; we study the behavior of $f_{i}$ and $(f_{i})^{*}$ in $R$. \begin{itemize} \item First, by construction, in $R_{+}$, the function $f_{i}$ satisfies: $$ f_{i}\sim e^{\frac{i}{\varepsilon}\int_{0}^{\varphi}\kappa_{i}}\psi_{+}^{i}.$$ \item To the right of $\alpha_{-}$, the function $f_{i}$ satisfies $$ f_{i}\sim e^{\frac{i}{\varepsilon}\int_{0}^{\varphi}\kappa_{i}}\psi_{+}^{i},$$ with $\mbox{Im }\kappa_{i}<0$. According to Lemma \ref{lemcontfin}, we know that, in $\bar{S_{-}}$, the function $f_{i}$ admits the asymptotic behavior: $$ f_{i}\sim e^{\frac{i}{\varepsilon}\int_{0}^{\varphi}\kappa_{i}}\psi_{+}^{i}.$$ \item Thus, the function $f_{i}$ has the standard asymptotic behavior in $R$. \item Now, we study the behavior of $f_{i}^{*}$. To do that, we start with describing the main objects related to $\kappa_{i}$ in $R$. Let $k_{i}$ be the branch of the quasi-moment of \eqref{espc}, analytically continued through $[E_{r},E_{i}]$ and satisfying: \begin{equation*} k_{i}(E_{r})=0\quad\textrm{ and }\quad k_{i}(E_{i})=\pi \end{equation*} $k_{i}$ is real on $[E_{r},E_{i}]$. Therefore, $k_{i}$ satisfies: $$k_{i}(\overline{\mathcal{E}})=\overline{k_{i}(\mathcal{E})}.$$ The branch $\kappa_{i}$ satisfies $\kappa_{i}(\varphi)=k_{i}(E-W(\varphi))$. The associated canonical Bloch solutions $\Psi_{\pm}^{i}$ are such that: $$\overline{\Psi_{+}^{i}(x,\overline{\varphi})}=\Psi_{-}^{i}(x,\varphi).$$ Therefore, we have in $R$: \begin{equation} \label{kappai} \kappa_{i}^{*}(\varphi)=\kappa_{i}(\varphi)\quad (\Psi_{+}^{i})^{*}(\varphi)=\Psi_{-}^{i}(\varphi)\quad(\omega_{+}^{i})^{*}(\varphi)=\omega_{-}^{i}(\varphi)\quad \forall\varphi\in R. \end{equation} Besides, since $k'_{i}$ is real on the band, there exists a real number $\sigma_{i}\in\{-1,1\}$ such that: \begin{equation} q_{i}^{*}(\varphi)=\sigma_{i}q_{i}(\varphi). \end{equation} We shall precise this coefficient in section \ref{deco}.\\ We compute: $$w(f_{+}^{i}(\cdot,\varphi,E,\varepsilon),(f_{+}^{i})^{*}(\cdot,\varphi,E,\varepsilon))=q_{i}(0)q_{i}^{*}(0)w(\Psi_{+}^{i}(\cdot,0),\Psi_{-}^{i}(\cdot,0))g(\varphi,E,\varepsilon).$$ Since $\overline{w(\Psi_{+}^{i}(\cdot,0),\Psi_{-}^{i}(\cdot,0))}=-w(\Psi_{+}^{i}(\cdot,0),\Psi_{-}^{i}(\cdot,0))$, the term $g(\varphi,E,\varepsilon)$ satisfies: $$g^{*}(\varphi,E,\varepsilon)=\overline{g(\bar{\varphi},\bar{E},\varepsilon)}=g(\varphi,E,\varepsilon),$$ $$g(\varphi,E,\varepsilon)=[1+o(1)].$$ Since the Wronskian is analytic and $\varepsilon$-periodic, this asymptotic is valid in $S_{\widetilde{Y}}$. \\ Since $g^{*}=g$ and $g=[1+o(1)]$, there exists an analytic function $(\varphi,E)\mapsto h(\varphi,E,\varepsilon)$ on $S_{\widetilde{Y}}\times\mathcal{U}$ such that:\\ - $g(\varphi,E,\varepsilon)=h(\varphi,E,\varepsilon)h^{*}(\varphi,E,\varepsilon)$,\\ - $h(\varphi,E,\varepsilon)=[1+o(1)]$.\\ We slightly deform $f_{i}$, i.e., we replace $f_{i}$ by $\frac{f_{i}}{h(\varphi,E,\varepsilon)}$; the basis $\{f_{i},f_{i}^{*}\}$ is consistent. \end{itemize} This ends the proof of Proposition \ref{bcki}.\\ \section{Consistent Jost solutions of \eqref{eqp}} \label{scattheory} This section is devoted to the proof of the following result. \begin{prop} \label{propconstinf} We assume that $(H_{V})$, $(H_{W,r})$ and $(H_{J}^{0})$ are satisfied. Fix $X>1$ and $\lambda>1$. Then, there exist a complex neighborhood $\mathcal{V}=\overline{\mathcal{V}}$ of $J$, a real $\varepsilon_{0}>0$, a constant $C>0$, two complex numbers $m_{g},\ m_{d}$ and two functions $(x,\varphi,E,\varepsilon)\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$, $(x,\varphi,E,\varepsilon)\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$ such that, if we define $$B_{\varepsilon}^{g}=\left\{\varphi\in S_{Y}\ ;\ \mbox{Re }\varphi<-C\varepsilon^{-\frac{\lambda}{s-1}}\right\}\textrm{ et }B_{\varepsilon}^{d}=\left\{\varphi\in S_{Y}\ ; \ \mbox{Re }\varphi>C\varepsilon^{-\frac{\lambda}{s-1}}\right\},$$ then \begin{itemize} \item The functions $(x,\varphi,E,\varepsilon)\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$ and $(x,\varphi,E,\varepsilon)\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$ are clearly defined and consistent on $\mathbb{R}\times S_{Y}\times \mathcal{V}\times]0,\varepsilon_{0}[$. \item For any $x\in[-X,X]$ and $\varepsilon\in]0,\varepsilon_{0}[$, $(\varphi,E)\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$ and $(\varphi,E)\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$ are analytic on $S_{Y}\times \mathcal{V}$. \item The function $x\mapsto h_{-}^{g}(x,\varphi,E,\varepsilon)$ (resp. $x\mapsto h_{+}^{d}(x,\varphi,E,\varepsilon)$) is a basis of $\mathcal{J}_{-}$ (resp. $\mathcal{J}_{+}$). \item The functions $h_{-}^{g}$ and $h_{+}^{d}$ have the following asymptotic behavior: \begin{equation} \label{as1} h_{-}^{g}(x,\varphi,E,\varepsilon)=e^{\frac{-i}{\varepsilon}\int_{m_{g}}^{\varphi}\kappa(u)du}\psi_{-}(x,\varphi,E)(1+R_{g}(x,\varphi,E,\varepsilon)), \end{equation} and \begin{equation} \label{asda1} h_{+}^{d}(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int_{m_{d}}^{\varphi}\kappa(u)du}\psi_{+}(x,\varphi,E)(1+R_{d}(x,\varphi,E,\varepsilon)), \end{equation} where \begin{itemize} \item $R_{g}$ and $R_{d}$ satisfy: $$\exists M>0,\quad\forall\varepsilon\in]0,\varepsilon_{0}[,\quad,\forall x\in[-X,X],\quad\forall E\in\mathcal{V},\quad\forall\varphi\in B_{\varepsilon}^{g},\quad|R_{g}(x,\varphi,E,\varepsilon)|\leq\frac{M}{\varepsilon|\mbox{Re }\varphi|^{s-1}},$$ $$\exists M>0,\quad\forall\varepsilon\in]0,\varepsilon_{0}[,\quad,\forall x\in[-X,X],\quad\forall E\in\mathcal{V},\quad\forall\varphi\in B_{\varepsilon}^{d},\quad|R_{d}(x,\varphi,E,\varepsilon)|\leq\frac{M}{\varepsilon|\mbox{Re }\varphi|^{s-1}}.$$ \item The functions $\psi_{+}$ and $\psi_{-}$ are the Bloch canonical solutions of the periodic equation \eqref{espa} defined in section \ref{cansolbloch}. \end{itemize} \item The asymptotics (\ref{as1}) and (\ref{asda1}) may be differentiated once in $x$. \item There exist two real numbers $\sigma_{g}\in\{-1,1\}$, $\sigma_{d}\in\{-1,1\}$, an integer $p$ and two functions $E\mapsto\alpha_{g}(E)$ and $E\mapsto\alpha_{d}(E)$ such that: \begin{enumerate} \item For any $\varepsilon\in]0,\varepsilon_{0}[$, $x\in\mathbb{R}$, $E\in\mathcal{V}$,et $\varphi\in S_{Y}$ ,we have: \begin{equation} \label{starg} \overline{\alpha_{g}(E)h_{-}^{g}(x,\overline{\varphi},\overline{E},\varepsilon)}=i\sigma_{g}e^{-\frac{i}{\varepsilon}2p\pi x} \alpha_{g}(E)h_{-}^{g}(x,\varphi,E,\varepsilon) \end{equation} \begin{equation} \label{stard} \overline{\alpha_{d}(E)h_{+}^{d}(x,\overline{\varphi},\overline{E},\varepsilon)}=i\sigma_{d}e^{\frac{i}{\varepsilon}2p\pi x} \alpha_{d}(E)h_{+}^{d}(x,\varphi,E,\varepsilon) \end{equation} \item The functions $\alpha_{g}$ and $\alpha_{d}$ are analytic and given by \eqref{renormconstg} and \eqref{renormconstd}. They do not vanish on $\mathcal{V}$. \end{enumerate} \end{itemize} \end{prop} We shall construct some consistent Jost solutions of \eqref{eqp}. To do that, we regard equation \eqref{eqp} as a perturbation of equation \eqref{esp} with $\mathcal{E}=E$. We adapt the construction of Jost functions developed in \cite{Fi1, New}. Precisely, we look for solutions of \eqref{esp} in the form : $$F_{-}^{g}=e^{- ik(E)\varphi/\varepsilon}\psi_{-}^{0}(x,E)(1+o(1)),\quad x\rightarrow -\infty,$$ $$ F_{+}^{d}=e^{ik(E)\varphi/\varepsilon}\psi_{+}^{0}(x,E)(1+o(1)),\quad x\rightarrow +\infty.$$ Since the functions $(x,\varphi,E,\varepsilon)\mapsto e^{\pm ik(E)\varphi/\varepsilon}\psi_{\pm}(x,E)$ are consistent, they allow us to construct a consistent resolvent for the periodic equation. Using this property and the fact that equation \eqref{eqp} is invariant by the consistency transformation $(x,\varphi)\mapsto(x-1,\varphi+\varepsilon)$, we obtain the consistency of the Jost functions. \subsection{Construction of the Jost functions} \label{scathyp} We start with constructing $F_{-}^{g}$. The construction of $F_{+}^{d}$ is similar. Since the parameter $E$ lies in the neighborhood of a gap, $\mbox{Im } k(E)$ is non zero; the function $F_{-}^{g}$ is therefore exponentially decreasing and goes to zero as $x$ goes to $-\infty$. Such a solution is called recessive. \subsubsection{} On a small enough complex neighborhood of $J$, $\mathcal{V}=\overline{\mathcal{V}}$, one can fix a determination $k$ of the quasi-momentum such that: $$\mbox{Im } k(E)\geq\beta>0,\quad\forall E\in\mathcal{V}.$$ Fix $m_{g}$ in $S_{Y}$ such that: \begin{itemize} \item The point $m_{g}$ is not a branch point of $\kappa$. \item It satisfies $\mbox{Im } m_{g}>0$, $k'_{E}(m_{g})\neq 0$. \item The domain $\{\varphi\in S_{Y}\ ;\ \mbox{Re }(\varphi-m_{g})<0\textrm{ and }\mbox{Im }(\varphi-m_{g})>0\}$ does not contain any branch point of $\kappa$. \end{itemize} We define $E_{g}=E-W(m_{g})$. We denote by $\psi_{\pm}^{0}$ the analytic Bloch solutions of equation \eqref{esp} normalized at the point $E_{g}$ ( $k'(E_{g}))\neq 0$). These solutions are constructed in Lemma \ref{anasol}. $$ \psi_{\pm}^{0}(x,E)=e^{\pm ik(E)x}p_{\pm}^{0}(x,E)\quad\textrm{with}\quad p_{\pm}^{0}(x+1,E)=p_{\pm}^{0}(x,E).$$ We define : $$\widetilde{\psi_{\pm}}(x,\varphi,E,\varepsilon)=e^{\pm ik(E)(x+\frac{\varphi}{\varepsilon})}p_{\pm}^{0}(x,E)=e^{\pm ik(E)\frac{\varphi}{\varepsilon}}\psi_{\pm}^{0}(x,E).$$ We consider the resolvent $R$ of $H_{0}$ : $$(Rg)(x)=-\int_{-\infty}^{x}\frac{\psi_{+}^{0}(x,E)\psi_{-}^{0}(x',E)-\psi_{+}^{0}(x',E)\psi_{-}^{0}(x,E)}{(k'w_{0})(E_{g})}g(x')dx'$$ \subsection{} Since $\widetilde{\psi_{-}}$ goes to zero as $x$ goes to $-\infty$, we look for a recessive consistent solution $\widetilde{f}$ of \eqref{eqp} in the form : \begin{equation} \label{eqres} \tilde{f}(x,\varphi,E,\varepsilon)=\widetilde{\psi_{-}}(x,\varphi,E,\varepsilon)+R[W(\varepsilon x+\varphi)\tilde{f}(x,\varphi,E,\varepsilon)]. \end{equation} We define $\tilde{f}(x,\varphi,E,\varepsilon)=e^{-ik(E)(x+\frac{\varphi}{\varepsilon})}f(x,\varphi,E,\varepsilon)$; equation \eqref{eqres} is transformed into: \begin{equation} \label{eqres_a} f(x,\varphi,E,\varepsilon)=p_{-}^{0}(x,E)+\int_{-\infty}^{x}A(x,x',E)W(\varepsilon x'+\varphi)f(x',\varphi,E,\varepsilon)dx' \end{equation} where the function $A$ satisfies: \begin{equation} \label{noy} A(x,x',E)=\frac{e^{2ik(E)(x-x')}p_{+}^{0}(x,E)p_{-}^{0}(x',E)-p_{+}^{0}(x',E)p_{-}^{0}(x,E)}{(k'w_{0})(E_{g})}. \end{equation} Since $\mbox{Im } k(E)\geq \beta>0$ for $E\in\mathcal{V}$, there exists a constant $C>0$ such that: \begin{equation} \forall x>x',\quad\forall E\in\mathcal{V},\quad |A(x,x',E)|\leq C \end{equation} \subsubsection{} Fix $X_{0}\in\mathbb{R}$ and $a>0$. If $I$ is a real interval, we define: $$R_{I}=\{\varphi\in S_{Y}\ ;\ \mbox{Re }\varphi\in I\}.$$ Let $B((-\infty,X_{0}]\times R_{[-a,a]})$ the set of bounded functions $\{f\ :\ (x,\varphi)\mapsto f(x,\varphi)\}$ on $(-\infty,X_{0}]\times R_{[-a,a]}$. The set $B((-\infty,X_{0}]\times R_{[-a,a]})$ equipped with the norm $$ \|f\|_{\infty}=\sup\limits_{x\in(-\infty,X_{0}],\mbox{Re }\varphi\in[-a,a]}|f(x,\varphi)|$$ is a Banach space.\\ We define the integral operator $T_{E}$ by: $$\begin{array}{ccccc}T_{E}:&B((-\infty,X_{0}]\times R_{[-a,a]})&\rightarrow &B((-\infty,X_{0}]\times R_{[-a,a]})&\\ &f&\mapsto &F \end{array}$$ \begin{equation} \label{opea} \textrm{ where } F(x,\varphi)=\int_{-\infty}^{x}A(x,x',E)W(\varepsilon x'+\varphi)f(x',\varphi)dx'. \end{equation} The operator $T_{E}$ is a bounded operator on $B((-\infty,X_{0}]\times R_{[-a,a]})$ and satisfies the estimate: $$\forall x\in(-\infty,X_{0}],\ \forall\varphi\in R_{[-a,a]},\quad |T_{E}(f)(x,\varphi)|\leq C\|f\|_{\infty}\int_{-\infty}^{x}|W(\varepsilon x'+\mbox{Re } (\varphi)+i\mbox{Im }(\varphi))|dx'.$$ $$\|T_{E}(f)\|_{\infty}\leq\frac{M}{\varepsilon}\sup\limits_{x\in(-\infty,X_{0}],\mbox{Re }\varphi\in[-a,a]}\frac{1}{|\varepsilon x+\mbox{Re }(\varphi)|^{s-1}}.$$ \subsubsection{} Fix $\lambda>1$. There exists a constant $C>0$ such that: $$ |X_{0}|> C\varepsilon^{-\frac{\lambda s}{s-1}}\Rightarrow\||T_{E}\||<\varepsilon^{s(\lambda-1)}.$$ We rewrite \eqref{eqres_a} in the form: \begin{equation} \label{eqres_b} (1-T_{E})f=p_{-}^{0}(x,E) \end{equation} The operator is then invertible on $B((-\infty,X_{0}]\times R_{[-a,a]})$. We define: \begin{equation} \label{eqresc} F_{-}^{g}(x,\varphi,E,\varepsilon)=(1-T_{E})^{-1}p_{-}^{0}(x,\varphi,E,\varepsilon) \end{equation} We now give some properties of $F_{-}^{g}$. \subsection{Properties of $F_{-}^{g}$} \subsubsection{Asymptotic behavior in $x$} Substituting $$(1-T_{E})^{-1}=1-(1-T_{E})^{-1}T_{E}$$ in equation (\ref{eqresc}), we obtain: \begin{equation} \label{eqresd} F_{-}^{g}(x,\varphi,E,\varepsilon)=e^{-ik(E)\varphi/\varepsilon}\psi_{-}(x,E)(1+R_{g}(x,\varphi,E,\varepsilon)), \end{equation} with $$ |R_{g}(x,\varphi,E,\varepsilon)|\leq\frac{M}{\varepsilon|\varepsilon x|^{s-1}},$$ for $x\in(-\infty,X_{0}]$ and $\varphi\in R_{[-a,a]}$.\\ The function $F_{-}^{g}$ is therefore in the Jost subspace $\mathcal{J}_{-}$ of equation \eqref{eqp}. \subsubsection{Study of the consistency} We assume that $a>1$ and $\varepsilon<1$. We now prove that the function $F_{-}^{g}$ is consistent.\\ We denote by $G$ the function: $$G\ :\ (x,\varphi,E,\varepsilon)\mapsto G(x,\varphi,E,\varepsilon)=F_{-}^{g}(x+1,\varphi-\varepsilon,E,\varepsilon)$$ $G$ is defined for $x\in(-\infty,X_{0}-1]$ and $\varphi\in R_{[-a+1,a-1]}$. Moreover, the function $G$ belongs to $B((-\infty,X_{0}-1]\times R_{[-a+1,a-1]})$.\\ We define the operator: $$\begin{array}{ccccc}\widetilde{T_{E}}:&B((-\infty,X_{0}-1]\times R_{[-a+1,a-1]})&\rightarrow &B((-\infty,X_{0}-1]\times R_{[-a+1,a-1]})&\\ &f&\mapsto &F \end{array}$$ \begin{equation} \label{opeb} \textrm{ where } F(x,\varphi)=\int_{-\infty}^{x}A(x,x',E)W(\varepsilon x'+\varphi)f(x',\varphi)dx'. \end{equation} Since $B((-\infty,X_{0}]\times R_{[-a,a]})\subset B((-\infty,X_{0}-1]\times R_{[-a+1,a-1]})$ and according to equations \eqref{opea} and \eqref{opeb}, the operator $\widetilde{T_{E}}$ is an extension of the operator $T_{E}$. Let us denote by $\widetilde{F_{-}^{g}}$ the restriction of $F_{-}^{g}$ to $(-\infty,X_{0}-1]\times R_{[-a+1,a-1]}$.\\ We compute in $ B((-\infty,X_{0}-1]\times R_{[-a+1,a-1]})$: $$(\widetilde{T_{E}}(G))(x,\varphi,E,\varepsilon)=(T_{E}(F_{-}^{g}))(x+1,\varphi-\varepsilon,E,\varepsilon)$$ This leads to: $$((1-\widetilde{T_{E}})(G))(x+1,\varphi-\varepsilon,E,\varepsilon)=((1-T_{E})(F_{-}^{g}))(x+1,\varphi-\varepsilon,E,\varepsilon)=p^{0}_{-}(x+1,E)=p^{0}_{-}(x,E),$$ The functions $\widetilde{F_{-}^{g}}$ and $G$ satisfy the relation: $$((1-\widetilde{T_{E}})(G))=((1-\widetilde{T_{E}})(\widetilde{F_{-}^{g}})).$$ For a sufficiently small $\varepsilon_{0}$, the operator $\widetilde{T_{E}}$ satisfies, for any $\varepsilon\in]0,\varepsilon_{0}[$: $$\||\widetilde{T_{E}}\||<\frac{1}{2}.$$ The operator $(1-\widetilde{T_{E}})$ is invertible in $B((-\infty,X_{0}-1]\times R_{[-a+1,a-1]})$ and: $$\widetilde{F_{-}^{g}}=G.$$ For $\varphi\in R_{[-a+1,a-1]}$, the functions $F_{-}^{g}$ and $G$ coincide on $(-\infty,X_{0}-1]$; according to the Cauchy-Lipschitz Theorem, they coincide for $x\in\mathbb{R}$. Fix $x\in\mathbb{R}$; $F_{-}^{g}$ and $G$ coincide for $\varphi\in R_{[-a+1,a-1]}$. By analyticity, they are equal for $\varphi\in S_{Y}$. \subsubsection{Asymptotic behavior in $\varphi$} We use now the consistency of $F_{-}^{g}$ to compute its asymptotics as $\mbox{Re }\varphi$ goes to $-\infty$. Fix $X>0$. We study $F_{-}^{g}$ for $x\in[-X,X]$. The function $F_{-}^{g}$ is consistent, and: $$F_{-}^{g}(x,\varphi,E,\varepsilon)=F_{-}^{g}(x+\frac{[\mbox{Re }(\varphi)]}{\varepsilon},\varphi-[\mbox{Re }(\varphi)],E,\varepsilon)$$ $$=e^{-ik(E)(x+\varphi/\varepsilon)}p_{-}^{0}(x,E)\left(1+O(\frac{1}{\varepsilon|\varepsilon x+[\mbox{Re }\varphi]|^{s-1}})\right).$$ As a result, there exists a constant $C$ such that: $$\mbox{Re }\varphi< -C\varepsilon^{-\frac{\lambda}{s-1}}\Rightarrow F_{-}^{g}(x,\varphi,E,\varepsilon)=e^{-ik(E)(x+\varphi/\varepsilon)}p_{-}(x,E)(1+\widetilde{R_{g}}(x,\varphi,E,\varepsilon)),$$ where $$|\widetilde{R_{g}}(x,\varphi,E,\varepsilon)|\leq\frac{M}{\varepsilon|\mbox{Re }\varphi|^{s-1}},$$ for $x\in[-X,X]$ and $\mbox{Re }\varphi< -C\varepsilon^{-\frac{\lambda}{s-1}}$.\\ We define $B_{\varepsilon}^{g}=\{\varphi\in S_{Y}\ ;\ \mbox{Re }\varphi< -C\varepsilon^{-\frac{\lambda}{s-1}}\}$. \subsection{Renormalization of $F_{-}^{g}$} We now renormalize $F_{-}^{g}$. We define: \begin{equation} \label{renorm} f_{-}^{g}(x,\varphi,E,\varepsilon)=e^{-\frac{i}{\varepsilon}\int_{m_{g}}^{-\infty}[\kappa(u)-k(E)]du}F_{-}^{g}(x,\varphi,E,\varepsilon), \end{equation} where the integral $\int_{m_{g}}^{-\infty}[\kappa(u)-k(E)]du$ is taken in the upper half plane. The function $E\mapsto\int_{m_{g}}^{-\infty}[\kappa-k(E)]$ is analytic on $\mathcal{V}$. For $\varphi\in B_{\varepsilon}^{g}$, we have: $$f_{-}^{g}(x,\varphi,E,\varepsilon)=e^{-\frac{i}{\varepsilon}\int_{m_{g}}^{\varphi}[\kappa(u)-k(E)]du}e^{-\frac{i}{\varepsilon}\int_{\varphi}^{-\infty}[\kappa-k(E)]}e^{-\frac{ik(E)\varphi}{\varepsilon}}\psi_{-}^{0}(x,E)(1+o(1))$$ Since the function $\psi_{-}$ is analytic and since $W(\varphi)=O(\varepsilon^{\frac{\lambda s}{s-1}})$ for $\varphi\in B_{\varepsilon}^{g}$, we get: \begin{equation} \forall\varphi\in B_{\varepsilon}^{g},\quad \psi_{-}(x,\varphi,E)=\psi_{-}^{0}(x,E-W(\varphi))=\psi_{-}^{0}(x,E)(1+o(1)) \end{equation} We finally obtain that, for $x\in[-X,X]$ and $\varphi\in B_{\varepsilon}^{g}$: \begin{equation*} f_{-}^{g}(x,\varphi,E,\varepsilon)=e^{-\frac{i}{\varepsilon}\int_{m_{g}}^{\varphi}\kappa(u)du}\psi_{-}(x,\varphi,E)(1+o(1)) \end{equation*} \subsubsection{Symmetries} Let $\gamma$ be a complex path and $f$ be an analytic function on $\gamma$. We have: \begin{equation} \label{symint} \int_{\gamma}f(z)dz=\overline{\int_{\overline{\gamma}}f^{*}(z)dz}. \end{equation} Since $J$ satisfies $(H_{J}^{0})$, according to equation \eqref{kl}, there exists an integer $p$ such that: \begin{equation} \label{entierp} k(E)+k^{*}(E)=2 p \pi. \end{equation} We recall that the functions $\omega_{\pm}$ associated to $\kappa$ are defined by equation \eqref{omega}. We consider a path $\widetilde{\gamma}_{g}$ such that: \begin{itemize} \item The path $\widetilde{\gamma}_{g}$ connects $\overline{m_{g}}$ to $m_{g}$ and is symmetric with respect to the real axis. \item The path $\widetilde{\gamma}_{g}$ does not contain any branch point of $\kappa$ and any pole of $\omega_{\pm}$. \end{itemize} We fix a continuous determination $q_{g}$ of $\sqrt{k'_{E}}$ on $\gamma_{g}$. According to relation \eqref{kl}, we have $(k^{*})'=-k'$, which implies that there exists $\sigma_{g}\in\{-1,1\}$ such that: \begin{equation} \label{symraccarrg} q_{g}^{*}=i\sigma_{g}q_{g} \end{equation} The functions $\psi_{\pm}(x,\varphi,E,m_{g})$ satisfy the relation: \begin{equation} \label{symblochnorm} \psi_{\pm}^{*}(x,\varphi,E,m_{g})=i\sigma_{g}e^{\pm \frac{2i p\pi x}{\varepsilon}}e^{\int_{\widetilde{\gamma_{g}}}\omega_{\pm}^{g}}\psi_{\pm}^{*}(x,\varphi,E,m_{g}). \end{equation} Besides, equations \eqref{symint} and \eqref{symgap} lead to the following relations: \begin{equation} \overline{\int_{\widetilde{\gamma_{g}}}\omega_{+}}=-\int_{\widetilde{\gamma_{g}}}\omega_{+}\ ;\ \overline{\int_{\widetilde{\gamma_{g}}}\omega_{-}}=-\int_{\widetilde{\gamma_{g}}}\omega_{-}. \end{equation} According to $(H_{W,r})$, $W^{*}=W$. By using \eqref{noy}, we compute: $$ A(x,x',E)=\overline{A(x,x',\bar{E})}.$$ The operator $T_{E}$ satisfies: $$T_{E}(f^{*})=[T_{E}(f)]^{*}.$$ Consequently, according to \eqref{eqres_b} and \eqref{symblochnorm}, we obtain that, for $E$ in $\mathcal{V}$, $x$ in $\mathbb{R}$ and $\varphi$ in $B_{\varepsilon}^{g}$, \begin{equation} (F_{-}^{g})^{*}(x,\varphi,E,\varepsilon)=\overline{F_{-}^{g}(x,\bar{\varphi},\bar{E},\varepsilon)}=i\sigma_{g}e^{-\frac{i}{\varepsilon}2p\pi x}e^{\int_{\widetilde{\gamma_{g}}}\omega_{-}^{g}}F_{-}^{g}(x,\varphi,E,\varepsilon). \end{equation} This leads to: $$(h_{-}^{g})^{*}=i\sigma_{g}e^{-\frac{i}{\varepsilon}2p\pi x}\frac{\alpha_{g}(E)}{\alpha_{g}^{*}(E)}h_{-}^{g},$$ where \begin{equation} \label{renormconstg} \alpha_{g}(E)=e^{-\frac{i}{2\varepsilon}\left(\int_{\widetilde{\gamma}_{g}}(\kappa(u)-p\pi)du+p\pi(m_{g}+\overline{m_{g}})\right)}e^{\frac{1}{2}\int_{\widetilde{\gamma}_{g}}\omega_{-}^{g}} \end{equation} Similarly, we fix $m_{d}$ in $S_{Y}$ such that: \begin{itemize} \item The point $m_{d}$ is not a branch point of $\kappa$. \item It satisfies $\mbox{Im } m_{d}>0$, $k'_{E}(m_{d})\neq 0$. \item The domain $\{\varphi\in S_{Y}\ ;\ \mbox{Re }(\varphi-m_{d})>0\textrm{ and }\mbox{Im }(\varphi-m_{d})>0\}$ does not contain any branch point of $\kappa$. \end{itemize} We define $E_{d}=E-W(m_{d})$ and we define the function $h_{+}^{d}$ by: \begin{equation} \label{renormd} h_{+}^{d}(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int_{m_{d}}^{+\infty}[\kappa(u)-k(E)]du+\frac{ik(E)m_{d}}{\varepsilon}}F_{+}^{d}(x,\varphi,E,\varepsilon) \end{equation} where the integral $\int_{m_{d}}^{+\infty}[\kappa(u)-k(E)]du$ is taken in the upper half plane.\\ We consider the path $\widetilde{\gamma}_{d}$ such that: \begin{itemize} \item The path $\widetilde{\gamma}_{d}$ connects $\overline{m_{d}}$ to $m_{d}$ and is symmetric with respect to the real axis. \item The path $\widetilde{\gamma}_{d}$ does not contain any branch point of $\kappa$ and any pole of $\omega_{\pm}$. \end{itemize} We fix a continuous branch $q_{d}$ of $\sqrt{k'_{E}}$ on $\gamma_{d}$. There exists a real number $\sigma_{d}$ such that: \begin{equation} \label{symraccarrd} q_{d}^{*}=i\sigma_{d}q_{d} \end{equation} The function $h_{+}^{d}$ satisfies: $$(h_{+}^{d})^{*}=i\sigma_{d}e^{\frac{i}{\varepsilon}2p\pi x}\frac{\alpha_{d}(E)}{\alpha_{d}^{*}(E)}h_{+}^{d},$$ where \begin{equation} \label{renormconstd} \alpha_{d}(E)=e^{\frac{i}{2\varepsilon}\left(\int_{\widetilde{\gamma}_{d}}(\kappa(u)-p\pi)du+p\pi(m_{d}+\overline{m_{d}})\right)}e^{\frac{1}{2}\int_{\widetilde{\gamma}_{d}}\omega_{+}^{d}} \end{equation} We define {\it the transmission coefficient}: \begin{equation} \label{transmcoeff} d(\varphi,E,\varepsilon)=w(\alpha_{g}h_{-}^{g}(\cdot,\varphi,E,\varepsilon),\alpha_{d}h_{+}^{d}(\cdot,\varphi,E,\varepsilon)) \end{equation} We immediately deduce from Proposition \ref{carvp} and Proposition \ref{propconstinf} that the eigenvalues of $H_{\varphi,\varepsilon}$ are characterized by: \begin{equation} d(\varphi,E,\varepsilon)=0 \end{equation} \subsection{Some remarks} \subsubsection{} The assumption $(H_{W,r})$ is not optimal. Actually, it suffices to assume that $W$ is analytic real in $S_{Y}$ and that there exists a function $f\in L^{1}(\mathbb{R})$ such that : $$\forall x\in\mathbb{R}\quad\sup\limits_{y\in[-Y,Y]}|W(x+iy)|\leq f(x).$$ \subsubsection{} In equations \eqref{starg} and \eqref{stard}, we could have included the numbers $i\sigma_{g}$ and $i\sigma_{d}$ into the functions $\alpha_{g}$ and $\alpha_{d}$, but we prefer showing the relations between $q_{g}$ and $q_{g}^{*}$, $q_{d}$ and $q_{d}^{*}$. \subsubsection{} Note that this construction differs from the constructions of canonical domains in \cite{FK1}. Indeed, the domains on which we construct these functions depend on $\varepsilon$. We shall extend these asymptotics on a fixed strip in the neighborhood of the real line (section \ref{infwkb}). \section{WKB Theorem on non compact domains} \label{infwkb} In this section, we prove a continuation result on non compact domains of $S_{Y}$. This result is a generalization on non compact domains of the method developed in \cite{FK1} and particularly of Lemma \ref{lemcontfin}.\\ We prove that the continuation of asymptotics stay valid on some half-strips $\{\varphi\in S_{Y}\ ;\ |\mbox{Re }\varphi|>A\}$. To do that, we cover these domains by a countable union of small local overlapping canonical domains, called $\delta$-chain (see section \ref{deltcha}).\\ This principle follows the recent developments and improvements of the WKB method (see \cite{FK3}). The idea is to get over the local notion of canonical domain in favor of maximal domains. These domains, constructed as union of local canonical domains are some domains on which a function keeps the standard behavior (see \cite{FK3}). \subsection{Continuation Theorem on non compact domains} \subsubsection{The main result} We shall prove the following result: \begin{thm} Continuation Theorem on non compact domains.\\ \label{infcontle} Fix $\tilde{Y}\in]0,Y[$. Assume that $V$ satisfies $(H_{V})$, that $W$ satisfies $(H_{W,r})$ and that $J$ satisfies $(H_{J}^{0})$. Then, there exist a real $\varepsilon_{0}>0$, a complex neighborhood $\mathcal{V}$ of $J$ and two real numbers $A_{g}$ and $A_{d}$ such that, if $f$ has the following properties: \begin{itemize} \item The function $f(\cdot,\varphi,E,\varepsilon)$ is a consistent solution of \eqref{eqp}. \item The function $(\varphi,E)\mapsto f(x,\varphi,E,\varepsilon)$ is analytic on $S_{\tilde{Y}}\times \mathcal{V}$ for any $x\in[-X,X]$ and any $\varepsilon\in]0,\varepsilon_{0}[$. \end{itemize} Then, \begin{enumerate} \item There exists $\kappa$ a continuous branch on $\{\mbox{Re }\varphi<A_{g}\}$ such that $\mbox{Im }\kappa>0$. Moreover, for any $C<B<A_{g}$, if the function $f$ satisfies the asymptotic behavior \begin{equation} \label{asprolac} f(x,\varphi,E,\varepsilon)=e^{-\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{-}(x,\varphi,E)+r_{C}(x,\varphi,E,\varepsilon)) \end{equation} with $\lim\limits_{\varepsilon\rightarrow 0}\sup\limits_{[-X,X]\times R_{(-\infty,C]}\times\mathcal{V}}\max\{|r_{C}(x,\varphi,E,\varepsilon)|,|\partial_{x}r_{C}(x,\varphi,E,\varepsilon)|\}=0$,\\ then, this behavior stays valid until $B$. Precisely: \begin{equation} \label{asprola} f(x,\varphi,E,\varepsilon)=e^{-\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{-}(x,\varphi,E)+r_{B}(x,\varphi,E,\varepsilon)) \end{equation} with $\lim\limits_{\varepsilon\rightarrow 0}\sup\limits_{[-X,X]\times R_{(-\infty,B]}\times\mathcal{V}}\max\{|r_{B}(x,\varphi,E,\varepsilon)|,|\partial_{x}r_{B}(x,\varphi,E,\varepsilon)|\}=0$.\\ \item There exists $\kappa$ a continuous branch on $\{\mbox{Re }\varphi>A_{d}\}$ such that $\mbox{Im }\kappa>0$. Moreover, for any $C>B>A_{d}$, if $f$ satisfies the asymptotic behavior \begin{equation} \label{asprolbc} f(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi,E)+r_{C}(x,\varphi,E,\varepsilon)) \end{equation} with $\lim\limits_{\varepsilon\rightarrow 0}\sup\limits_{[-X,X]\times R_{[C,+\infty)}\times\mathcal{V}}\max\{|r_{C}(x,\varphi,E,\varepsilon)|,|\partial_{x}r_{C}(x,\varphi,E,\varepsilon)|\}=0$,\\ then this behavior stays valid until $B$. Precisely: \begin{equation} \label{asprolb} f(x,\varphi,E,\varepsilon)=e^{\frac{i}{\varepsilon}\int^{\varphi}\kappa(u)du}(\psi_{+}(x,\varphi,E)+r_{B}(x,\varphi,E,\varepsilon)) \end{equation} with $\lim\limits_{\varepsilon\rightarrow 0}\sup\limits_{[-X,X]\times R_{[B,+\infty)}\times\mathcal{V}}\max\{|r_{B}(x,\varphi,E,\varepsilon)|,|\partial_{x}r_{B}(x,\varphi,E,\varepsilon)|\}=0$. \end{enumerate} \end{thm} Theorem \ref{infcontle} and Proposition \ref{propconstinf} clearly imply Theorem \ref{jostthm}. \subsubsection{Some remarks} \label{hypW} We shall prove Theorem \ref{infcontle} as $W$ satisfies the weaker assumptions:\\ {\bf (H1) $\mathbf{W}$ is an analytic real function in $\mathbf{S_{Y}}$}.\\ {\bf (H2) $\mathbf{\exists\ C>0,\quad \exists\ s>1\textrm{ such that }\forall\ z\in S_{Y},\quad |W'(z)|\leq\frac{C}{1+|z|^{s}}}$}\\ {\bf (H3) $\mathbf{\exists\ f \in L^{1}(\mathbb{R})\textrm{ such that }\forall x\in\mathbb{R}\quad \sup\limits_{y\in[-Y,Y]}|W(x+iu)|\leq f(x)}$}\\ The following lemma relates $(H_{W,r})$ and $(H1)$, $(H2)$ and $(H3)$: \begin{lem} \label{hypfaib} Let $W$ satisfy $(H_{W,r})$ on $S_{Y}$. Fix $\tilde{Y}\in]0,Y[$. Then $W$ satisfies $(H1)$, $(H2)$ and $(H3)$ on $S_{\tilde{Y}}$. \end{lem} \begin{dem} Assume that $W$ satisfy $(H_{W,r})$ on $S_{Y}$. We prove that $W$ satisfies $(H2)$ on $S_{\tilde{Y}}$ by using the following lemma: \begin{lem} \label{der} Let $f$ be an analytic function on $S_{Y}$ such that $|f(z)|\leq\frac{C}{1+|z|^{s}}$, $C>0$.\\ Fix $\eta>0$. Then, $$\forall p\in\mathbb{N}^{*}\quad \exists C_{p}>0/\quad \forall z\in S_{Y-\eta}\quad|f^{(p)}(z)|\leq\frac{C_{p}}{1+|z|^{s}}.$$ \end{lem} \begin{dem}\\ This result is a consequence of the Cauchy formula. We do not give the details. \end{dem}\\ - Clearly, $W$ satisfies $(H1)$ on $S_{Y}$.\\ - $W$ satisfies $(H3)$ with $f(x)=\frac{C}{1+|x|^{s}}$.\\ This completes the proof of Lemma \ref{hypfaib}. \end{dem} \subsubsection{} Let us briefly outline the ideas of the proof. We shall concentrate on $B_{g}=\{\varphi\in S_{Y};\ \mbox{Re }(\varphi)<A_{g}\}$. There are three steps.\\ First we cover $B_{g}$ with an union of overlapping local compact canonical domains $K_{m}$.\\ In each canonical domain $K_{m}$, we can construct a consistent local basis thanks to Theorem \ref{finwkbthm}. To compute the connection between the consistent bases of $K_{m}$ and $K_{m+n}$, it suffices to do the product of the $n$ transfer matrices between the canonical bases of two successive domains. The accuracy of the rest cannot be better than the sum of the accuracies obtained on each domain. Theorem \ref{finwkbthm} gives an estimate in $o(1)$; this accuracy is insufficient when $n$ goes to infinity.\\ A refinement of the calculation of asymptotics in Theorem \ref{finwkbthm} is therefore necessary. We prove it by using the integrability of $W$. \subsubsection{Branch points} The following result specifies the location of the branch points of $\kappa$. We recall that $\Upsilon(E)$ is defined in \eqref{nupsilon}. \begin{lem} Let $\mathcal{V}$ be a complex neighborhood of the interval $J$. Assume that $W$ satisfies $$\lim\limits_{x\rightarrow +\infty}\sup\limits_{y\in[-Y,Y]}|W(x+iy)|=0,$$ then: \begin{equation*} \label{bplem} \exists A>0\textrm{ such that }\forall\ E\in\mathcal{V},\quad\varphi\in\Upsilon(E)\cap S_{Y}\Rightarrow|\mbox{Re }(\varphi)|<A. \end{equation*} \end{lem} \begin{dem} Since $\overline{\mathcal{V}}\cap\partial\sigma(H_{0})=\emptyset$, there exists $\alpha>0$ such that: $$\forall\ E\in\mathcal{V},\quad\forall\ p\in\mathbb{N}^{*},\quad|E-E_{p}|\geq\alpha.$$ If $\varphi_{p}(E)$ satisfies $E-W(\varphi_{p}(E))=E_{p}$, we get: $$\forall\ E\in\mathcal{V},\quad\forall\ p\in\mathbb{N}^{*},\quad|W(\varphi_{p}(E))|\geq\alpha.$$ Finally, $\{u\in S_{Y}\ ;\ |W(u)|\geq\alpha\}$ is a subset of a compact of $S_{Y}$. This completes the proof of Lemma \ref{bplem}. \end{dem} \subsubsection{Uniform asymptotics on a $\delta$-chain} \label{deltcha} First, we introduce a new definition. We remind that the width of a complex subset is defined in \eqref{larga}. \begin{defn}$\delta$-$\textit{chain of strictly canonical domains}$\\ Fix $\widetilde{Y}\in]0,Y[$. Fix $E$. Let $D$ be a simply connected domain of $S_{\widetilde{Y}}$ containing no branch points of the complex momentum. We fix on $D$ a continuous branch $\kappa$ of the complex momentum. Let $\{\tau_{n}\}_{n\in\mathbb{N}}$ be a sequence of real numbers and $K$ be a compact of $S_{\widetilde{Y}}$.\\ $\{K+\tau_{n}\}_{n\in\mathbb{N}}$ is called a $\delta$-chain for $E$, $\kappa$ and $D$ if it satisfies the following properties: \begin{enumerate} \item $\bigcup\limits_{n=0}^{\infty}(K+\tau_{n})=D.$ \item $\exists\tau>0\textrm{ such that }\forall n\in\mathbb{N}\quad l((K+\tau_{n})\cap(K+\tau_{n+1}),\widetilde{Y})>\tau.$ \item The domain $K$ is an union of curves $\gamma$ such that, for any $n$, $\gamma+\tau_{n}$ is a $\delta$-strictly canonical curve for $\kappa$. \end{enumerate} \end{defn} $K$ is called the fundamental domain of the $\delta$-chain. Now, we have the intermediate result: \begin{prop} \label{infwkbprop} Assume that $V$ satisfies $(H_{V})$ and that $W$ satisfies $(H_{1})$, $(H_{2})$ and $(H_{3})$. Fix $\tilde{Y}\in]0,Y[$. Let $\mathcal{V}$ a complex neighborhood of $J$ and $D\subset S_{\tilde{Y}}$ a domain with the following properties: \begin{itemize} \item $\inf\limits_{p\in\mathbb{N}^{*},E\in\mathcal{V}} \textrm{dist}\{D,\varphi_{p}(E)\}\geq C,$ \item there exists $\{\tau_{n}\}_{n\in\mathbb{N}}$ such that, for any $E\in\mathcal{V}$, $\{K+\tau_{n}\}_{n\in\mathbb{N}}$ is a $\delta$-chain for $E$ and $D$. \end{itemize} Fix $\varphi_{0}\in D$.\\ Then, there exists $\varepsilon_{0}>0$ such that, for any $n\in\mathbb{N}$, there exist two functions $(x,\varphi,E,\varepsilon)\mapsto\psi_{\pm}^{n}(x,\varphi,E,\varepsilon)$ with the following properties: \begin{itemize} \item The functions $(x,\varphi,E,\varepsilon)\mapsto\psi_{\pm}^{n}(x,\varphi,E,\varepsilon)$ are defined on $\mathbb{R}\times(K+\tau_{n})\times\mathcal{V}\times]0,\varepsilon_{0}[$ and form a consistent basis. \item for any fixed $x\in\mathbb{R},\ \varepsilon\in]0,\varepsilon_{0}[$, the functions $(\varphi,E)\mapsto\psi_{\pm}^{n}(x,\varphi,E,\varepsilon)$ are analytic on $(K+\tau_{n})\times\mathcal{V}$. \item for $x\in[-X,X]$, $\varphi\in (K+\tau_{n})$ and $E\in\mathcal{V}$, the functions $\psi_{\pm}^{n}$ have the asymptotic behavior: \begin{equation} \label{asc} \psi_{\pm}^{n}(x,\varphi,E,\varepsilon)=e^{\pm\frac{i}{\varepsilon}\int_{\varphi_{0}}^{\varphi}\kappa du}\left(\psi_{\pm}(x,\varphi,E)+\frac{1}{1+|\tau_{n}|^{s}}o(1)\right). \end{equation} \item The asymptotics (\ref{asc}) are uniform in $x,\ \tau$, $\varphi\in K+\tau_{n}$ et $E\in\mathcal{V}$. \item The asymptotics can be differentiated once in $x$. \end{itemize} \end{prop} The proof of Proposition \ref{infwkbprop} mimics this of Theorem 1.1 in \cite{FK1}. We omit the details and we refer to \cite{FK1}, section 4 for an analogous statement. \subsection{Construction of a $\delta$-chain of strictly canonical domains} In this section, we shall construct a $\delta$-chain under assumptions $(H_{1})$, $(H_{2})$ and $(H_{3})$. \begin{prop} \label{constdeltachain} Fix $\tilde{Y}\in]0,Y[$. Assume that $V$ satisfies $(H_{V})$, that $W$ satisfies $(H_{1})$, $(H_{2})$ and $(H_{3})$ and that $J$ satisfies $(H_{J}^{0})$. Then, there exist a complex neighborhood $\mathcal{V}$ of $J$, two real numbers $(A_{g},A_{d})\in\mathbb{R}^{2}$, a domain $K\subset S_{\tilde{Y}}$ and two real sequences $\{\tau_{n}^{1}\}_{n\in\mathbb{N}}$, $\{\tau_{n}^{2}\}_{n\in\mathbb{N}}$ such that: \begin{itemize}\item for any $E\in\mathcal{V}$, there exists a continuous branch $\kappa$ on $\{\varphi\in S_{\tilde{Y}}\ ;\ \mbox{Re }\varphi\in(-\infty,A_{g}]\}$ (resp. on $\{\varphi\in S_{\tilde{Y}}\ ;\ \mbox{Re }\varphi\in[A_{d},\infty)\}$), \item for any $E\in\mathcal{V}$, $\{K+\tau_{n}^{1}\}_{n\in\mathbb{N}}$ (resp. $\{K+\tau_{n}^{2}\}_{n\in\mathbb{N}}$) is a $\delta$-chain for $\kappa$, $E$ and $\{\mbox{Re }\varphi\in(-\infty,A_{g}]\}$ (resp. $\{\mbox{Re }\varphi\in[A_{d},+\infty)$)). \end{itemize} \end{prop} The rest of the section \ref{deltcha} is devoted to the proof of Proposition \ref{constdeltachain}. This proof is based on elementary geometrical arguments. We prove the construction for $\mbox{Re }\varphi\in(-\infty,A_{g}]$. \subsubsection{Construction of $\delta$-strictly canonical straight-lines} We have defined the canonical lines in section \ref{lc} and described them in terms of the vector $t(\varphi)$.\\ We set $\alpha=\frac{1}{2}\inf\limits_{E\in J}\mbox{Im } k(E)$ and $m=2\sup\limits_{E\in J}|\mbox{Re } k(E)|$.\\ Since the mapping $(E,\varphi)\mapsto E-W(\varphi)$ is continuous and since $W(\varphi)$ goes to zero when $\mbox{Re }\varphi$ goes to infinity, there exist a complex neighborhood $\mathcal{V}$ of $J$ and a real number $A_{g}$ such that: $$\forall E\in\mathcal{V},\quad\forall\varphi\in(-\infty,A_{g}],\quad \mbox{Re } k(E-W(\varphi))\in[-m,m],\quad \mbox{Im } k(E-W(\varphi))>\alpha$$ We set $B_{g}=(-\infty,A_{g}]+i[-\tilde{Y},\tilde{Y}]$. The canonical curves for $\mbox{Re }\varphi$ in the neighborhood of $-\infty$ are described by: \begin{lem} \label{geom} There exists $\theta_{0}\in]0,\pi/2[$ such that, if $\gamma$ is a smooth curve in $B_{g}$ satisfying: \begin{equation} \label{geoma} \forall\varphi\in\gamma,\quad \arg[t(\varphi)]\in]\theta_{0},\pi/2-\theta_{0}[. \end{equation} then, $\gamma$ is a canonical line for $\kappa$. \end{lem} \begin{dem} For $\arg(u)=\theta$ and $\cot\theta\in]-\frac{m-\delta}{\alpha},\frac{\pi+m-\delta}{\alpha}[$, we have: $$\mbox{Im }(\overline{(\kappa-\delta)}u)>0\quad\textrm{et}\quad\mbox{Im }(\overline{(\pi-\kappa+\delta)}u)>0$$ Consequently, $\cot\theta_{0}=\frac{m-\delta}{\alpha}$ implies that (\ref{geoma}) is satisfied. \end{dem} \subsubsection{The fundamental domain $K$} Let $\xi_{1}=-i\tilde{Y}$ and $\xi_{2}=i\tilde{Y}$. We denote by $K$ the lozenge bounded by the straight lines containing $\xi_{1}$ and $\xi_{2}$ whose guiding vectors have the affixes $e^{i\theta_{0}}$ and $e^{i(\pi-\theta_{0})}$.\\ We set $[-u_{0},u_{0}]=K\cap\{y=0\}$. $K$ is shown in figure \ref{element}. Fix $x$ such that $K+x\subset B_{g}$; we shall show that $K+x$ is a $\delta$-strictly canonical domain. According to Lemma \ref{geom}, it suffices to write $K$ as an union of smooth curves satisfying (\ref{geoma}).\\ For any $u\in K$, we consider a vertical segment $[\overline{\xi},\xi]$ containing $u$ and included in $K$ (see figure \ref{element}). The broken line $[\xi_{1},\overline{\xi}]\cup[\overline{\xi},\xi]\cup[\xi,\xi_{2}]$ satisfies (\ref{geoma}). The relation (\ref{geoma}) is stable under small $C^{1}$-perturbation; we slightly deform the line $[\xi_{1},\overline{\xi}]\cup[\overline{\xi},\xi]\cup[\xi,\xi_{2}]$ to get a smooth curve which satisfies (\ref{geoma}).\\ Consequently, $K$ satisfies the following properties:\\ - $K\cap S_{\tilde{Y}}$ contains a rectangle of width $4\eta>0$.\\ - $l((K-n\eta)\cap(K-(n+1)\eta),\tilde{Y})>\eta$.\\ - $K$ is the union of curves $\gamma$ such that $\gamma-n\eta$ is $\delta$-strictly canonical for any sufficiently large $n$. \subsubsection{Conclusion} To finish the proof, it suffices to adapt the proof of Lemma \ref{lemcontfin} in section 5.9 of \cite{FK1}, by using Proposition \ref{constdeltachain} and Proposition \ref{infwkbprop}. The convergence of the series of general term $\frac{1}{1+|\tau_{n}|^{s}}$ replaces the compactness. We do not give the details. \input{fig10} \section{Transmission coefficient. Equation for eigenvalues} \label{calcmattransf}\label{anares} In Theorem \ref{jostthm}, we have constructed two functions $h_{-}^{g}$ and $h_{+}^{d}$. We have defined the transmission coefficient $d(E,\varphi,\varepsilon)$. We choose $m_{g}=-0+i0$ and $m_{d}=0+i0$.\\ In Proposition \ref{bcki}, we have introduced a consistent basis $(f_{i},f_{i}^{*})$ near the cross. To compute $d(E,\varphi,\varepsilon)$, we shall project the functions $h_{-}^{g}$ and $h_{+}^{d}$ onto the basis $(f_{i},f_{i}^{*})$.\\ \subsection{Preliminaries} \subsubsection{Introduction. Notations} Fix $\widetilde{Y}<Y$ and $E_{0}\in J$. We have described in section \ref{wkbconst} the complex momentum $\kappa$ and the related geometric objects. We recall that we consider the case \eqref{premcassc}. We use the notations introduced in section \ref{wkbconst}. The branch points are called $\varphi_{r}^{\pm}$ and $\varphi_{i},\ \overline{\varphi_{i}}$. We have described the Stokes lines in section \ref{stline}.\\ We have described in sections \ref{compmom} and \ref{poss} the different branches $\kappa_{i}$, $\kappa_{g}$ and $\kappa_{d}$. The branch $\kappa_{g}$, resp. $\kappa_{d}$, is defined and continuous on the domain $\{\varphi\in S_{Y};\ \mbox{Re }\varphi<\varphi_{r}^{-}\}$, resp. $\{\mbox{Re }\varphi>\varphi_{r}^{+}\}$. The branch $\kappa_{i}$ is defined and continuous on a neighborhood of the cross. The domain $(E-W)\left(\{\varphi\in S_{Y};\ \mbox{Re }\varphi<\varphi_{r}^{-}\}\right)$ is a simply connected domain which intersects with real axis in only one gap. Thus, we can fix a determination $k_{g}$ of the quasi-momentum such that: $$k_{g}(E-W(\varphi))=\kappa_{g}(\varphi).$$ Similarly, we fix the branches $k_{i}$ and $k_{d}$ of the quasi-momentum such that : $$ k_{i}(E-W(\varphi))=\kappa_{i}(\varphi),\quad k_{d}(E-W(\varphi))=\kappa_{d}(\varphi).$$ Finally, we set: $$ q_{i}(\varphi)=\sqrt{k'_{i}(E-W(\varphi))},\quad q_{g}(\varphi)=\sqrt{k'_{g}(E-W(\varphi))}\quad q_{d}(\varphi)=\sqrt{k'_{d}(E-W(\varphi))}.$$ Let $\varphi_{g}\in\mathbb{R}$ such that $\varphi_{g}<\varphi_{r}^{-}$ and such that the interval $[\varphi_{g},\varphi_{r}^{-}]$ does not contain any pole of $\omega_{\pm}$. We define the path $\gamma_{g}$ in the complex plane by: $$\gamma_{g}=[-0+i0,\varphi_{g}+i0]\cup[\varphi_{g}-i0,-0-i0].$$ Similarly, fix $\varphi_{d}\in\mathbb{R}$ such that $\varphi_{d}>\varphi_{r}^{+}$ and such that the interval $[\varphi_{r}^{+},\varphi_{d}]$ does not contain any pole of $\omega_{\pm}$. We define the path $\gamma_{d}$ in the complex plane by: $$\gamma_{d}=[0+i0,\varphi_{d}+i0]\cup[\varphi_{d}-i0,0-i0].$$ In the following section, we explain the choice of the determinations $q_{i}$, $q_{g}$ and $q_{d}$. \subsubsection{The determination $q$} \label{deco} We recall that there exists a real number $\sigma_{i}\in\{-1,1\}$ such that: \begin{equation} \label{sigmareli} \frac{q_{i}^{*}}{q_{i}}=\sigma_{i}. \end{equation} We refer to section \ref{precsigma}.\\ The Wronskian satisfies $w(f_{i},(f_{i})^{*})=\sigma_{i}(w_{0}k_{i}')(E-W(0)).$\\ The number $\sigma_{i}$ depends on the sign of $k'$ along the band $B$: \begin{itemize} \item If the band $B$ can be written $[E_{4p+1},E_{4p+2}]$, then $k'>0$ on $B$ and $\sigma_{i}=1$. \item If the band $B$ can be written $[E_{4p+3},E_{4p+4}]$, then $k'<0$ on $B$ and $\sigma_{i}=-1$. \end{itemize} We fix the branch $q_{g}$ such that $q_{g}=q_{i}$ in $S_{-}$ and such that $q_{g}$ is analytically continued in $\{\varphi\in S_{Y};\ \mbox{Re }\varphi<\varphi_{r}^{-}\}$. According to relation \eqref{sigmareli}, the branch $q_{g}$ satisfies: \begin{equation} \label{sigmarelg} q_{g}^{*}=i\sigma_{i}q_{g} \end{equation} Similarly, we fix $q_{d}$ such that $q_{d}=q_{i}$ in $\overline{S_{+}}$ and such that $q_{d}$ is analytically continued in $\{\varphi\in S_{Y};\ \mbox{Re }\varphi>\varphi_{r}^{+}\}$. The branch $q_{d}$ satisfies: \begin{equation} \label{sigmareld} q_{d}^{*}=i\sigma_{i}q_{d} \end{equation} According to equations \eqref{sigmarelg} and \eqref{sigmareld}, we have also: $$\sigma_{g}=\sigma_{i}\quad;\quad\sigma_{d}=\sigma_{i}.$$ We denote by $\widetilde{\Psi_{\pm}}^{g}(x,\mathcal{E})$, $\widetilde{\Psi_{\pm}}^{i}(x,\mathcal{E})$ and $\widetilde{\Psi_{\pm}}^{d}(x,\mathcal{E})$ the Bloch solutions described in section \ref{bloch}. We set: {\small $$\Psi_{\pm}^{i}(x,\varphi,E)=\widetilde{\Psi_{\pm}}^{i}(x,E-W(\varphi))\ ;\ \Psi_{\pm}^{g}(x,\varphi,E)=\widetilde{\Psi_{\pm}}^{g}(x,E-W(\varphi))\ ;\ \Psi_{\pm}^{d}(x,\varphi,E)=\widetilde{\Psi_{\pm}}^{d}(x,E-W(\varphi)).$$} We define the functions $\omega_{\pm}^{g}$, $\omega_{\pm}^{i}$ and $\omega_{\pm}^{d}$ associated by \eqref{omega} to the branches $k_{g}$, $k_{i}$ and $k_{d}$. \subsubsection{Ideas of the method} The computation is similar to this done in \cite{FK2, FK1, FK4}. It is based on some elementary principles that we outline now. \begin{enumerate} \item Periodicity.\\ The consistency condition \eqref{coh} implies that the Wronskians are $\varepsilon$-periodic in $\varphi$. To get a total control of the Wronskians in a horizontal strip, we only need to control them in some vertical sub-strip of width $\varepsilon$. \item Analyticity.\\ Since the functions $(\varphi,E)\mapsto f_{-}^{g}(x,\varphi,E,\varepsilon)$, $(\varphi,E)\mapsto f_{+}^{d}(x,\varphi,E,\varepsilon)$, $(\varphi,E)\mapsto f_{\pm}^{i}(x,\varphi,E,\varepsilon)$ are analytic on $S_{\tilde{Y}}\times \mathcal{U}$, their Wronskians are analytic in $(\varphi,E)\in S_{\tilde{Y}}\times \mathcal{U}$. This allows us to expand them into exponentially converging series. \\ Let $w(\varphi,E,\varepsilon)$ be an analytic function in $(\varphi,E)$ which is $\varepsilon$-periodic in $\varphi$. We set: $$ w(\varphi,E,\varepsilon)=\sum\limits_{k\in\mathbb{Z}}w_{k}(E,\varepsilon)e^{\frac{2i\pi\varphi}{\varepsilon}}$$ The Cauchy formula gives an estimate of the Fourier coefficients: \begin{equation} \label{fourgen} w_{k}(E,\varepsilon)=\frac{1}{\varepsilon}\int_{\varphi_{0}}^{\varphi_{0}+\varepsilon}w(\varphi,E,\varepsilon)e^{-\frac{2i k\pi\varphi}{\varepsilon}}d\varphi,\quad\forall k\in\mathbb{N},\quad\forall\varphi_{0}\in S_{\tilde{Y}}. \end{equation} By moving $\mbox{Im }\varphi_{0}$ in $[-\tilde{Y},\tilde{Y}]$, we get a control of positive and negative coefficients. \end{enumerate} \subsection{Asymptotic expansion of $d(\varphi,E,\varepsilon)$} In this section, we shall establish the following result. \begin{prop} \label{colina} For any $E_{0}$ in $J$, there exist a complex neighborhood $\mathcal{U}_{0}$ of $E_{0}$ and two functions $(\varphi,E,\varepsilon)\mapsto b_{g}^{-}(\varphi,E,\varepsilon)$ and $(\varphi,E,\varepsilon)\mapsto b_{d}^{+}(\varphi,E,\varepsilon)$ such that: \begin{itemize} \item The coefficient $d$ defined in \eqref{transmcoeff} can be written: \begin{equation} d(\varphi,E,\varepsilon)=i\sigma_{i}w(f_{i},(f_{i})^{*})[b_{g}^{-}(b_{d}^{+})^{*}-(b_{g}^{-})^{*}b_{d}^{+}]. \end{equation} \item The functions $(\varphi,E)\mapsto b_{g}^{-}(\varphi,E,\varepsilon)$ and $(\varphi,E)\mapsto b_{d}^{+}(\varphi,E,\varepsilon)$ are analytic on $S_{Y}\times \mathcal{U}_{0}$. \item The functions $\varphi\mapsto b_{g}^{-}(\varphi,E,\varepsilon)$ and $\varphi\mapsto b_{d}^{+}(\varphi,E,\varepsilon)$ are $\varepsilon$-periodic and admit the following Fourier asymptotic expansion, when $\varepsilon\rightarrow 0$: \begin{equation} \label{bscoeffaaa} b_{g}^{-}(\varphi,E,\varepsilon)=\sum\limits_{k\in\mathbb{Z}}(b_{g}^{-})_{k}(E,\varepsilon)e^{\frac{2ik\pi\varphi}{\varepsilon}}, \end{equation} with \begin{equation} \label{bscoeffaa} (b_{g}^{-})_{0}(E,\varepsilon)=\sigma_{i}e^{-\frac{i}{\varepsilon}\int_{0}^{\varphi_{r}^{-}}\kappa_{i}}e^{\frac{1}{2}\int_{0}^{\varphi_{r}^{-}}(\omega_{+}^{i}-\omega_{-}^{i})}[1+o(1)], \end{equation} and \begin{equation} \label{bscoeffbb}\forall k \neq 0,\quad |(b_{g}^{-})_{k}(E,\varepsilon)|<C e^{-\alpha/\varepsilon}e^{\frac{-2|k|\pi Y_{0}}{\varepsilon}}, \end{equation} \begin{equation} \label{bscoeffbbb} b_{d}^{+}(\varphi,E,\varepsilon)=\sum\limits_{k\in\mathbb{Z}}(b_{d}^{+})_{k}(E,\varepsilon)e^{\frac{2ik\pi\varphi}{\varepsilon}}, \end{equation} with \begin{equation} \label{bscoeffcc}(b_{d}^{+})_{0}(E,\varepsilon)=i\sigma_{i}e^{\frac{i}{\varepsilon}\int_{0}^{\varphi_{r}^{+}}\kappa_{i}}e^{\frac{1}{2}\int_{0}^{\varphi_{r}^{+}}(\omega_{+}^{i}-\omega_{-}^{i})}[1+o(1)],\end{equation} \begin{equation} \label{bscoeffdd}(b_{d}^{+})_{1}(E,\varepsilon)=-i\sigma_{i}e^{\frac{i}{\varepsilon}\int_{\varphi_{r}^{+}}^{0}\kappa_{i}}e^{\frac{2i}{\varepsilon}\int_{0}^{\overline{\varphi_{i}}}(\kappa_{i}-\pi)}e^{\frac{1}{2}\int_{\varphi_{r}^{+}}^{0}\omega_{+}^{i}-\omega_{-}^{i}}e^{\int_{0}^{\overline{\varphi_{i}}}(\omega_{+}^{i}-\omega_{-}^{i})}[1+o(1)],\end{equation} et \begin{equation} \label{bscoeffee}\forall k>1,\quad |(b_{d}^{+})_{k}(\varphi,E,\varepsilon)|<C |(b_{d}^{+})_{1}(E,\varepsilon)|e^{-\alpha/\varepsilon}e^{\frac{-2|k-1|\pi Y_{0}}{\varepsilon}},\end{equation} \begin{equation} \label{bscoeffff}\forall k<0,\quad |(b_{d}^{+})_{k}(\varphi,E,\varepsilon)|<C e^{-\alpha/\varepsilon}e^{\frac{-2|k|\pi Y_{0}}{\varepsilon}},\end{equation} \end{itemize} \end{prop} The rest of the section is devoted to the proof of Proposition \ref{colina}.\\ Fix $E_{0}\in J$. According to the choice of $\kappa_{g}$ and $\kappa_{d}$ (sections \ref{compmom} and \ref{poss}), there exist two analytic functions $\alpha_{g}(E)$ and $\alpha_{d}(E)$ such that: $$(\alpha_{g}h_{-}^{g})^{*}=i\sigma_{g}\alpha_{g}h_{-}^{g},$$ $$(\alpha_{d}h_{+}^{d})^{*}=i\sigma_{d}\alpha_{d}h_{+}^{d}.$$ Now, we use the function $f_{i}$ constructed in Proposition \ref{bcki}. There exists a neighborhood $\mathcal{U}_{0}$ of $E_{0}$ such that we can write: $$\alpha_{g}h_{-}^{g}=-i\sigma_{g}(b_{g}^{-})^{*}f_{i}+b_{g}^{-}(f_{i})^{*},$$ $$\alpha_{d}h_{+}^{d}=-i\sigma_{d}(b_{d}^{+})^{*}f_{i}+b_{d}^{+}(f_{i})^{*}.$$ The coefficients $\alpha_{g}$ and $\alpha_{d}$ are defined in equations \eqref{renormconstg} and \eqref{renormconstd}. We compute: $$\int_{\gamma_{g}}\omega_{-}^{g}=\int_{0}^{\varphi_{r}^{-}}(\omega_{+}^{i}-\omega_{-}^{i}),$$ $$\int_{\gamma_{d}}\omega_{+}^{d}=\int_{0}^{\varphi_{r}^{+}}(\omega_{+}^{i}-\omega_{-}^{i}).$$ This leads to: $$\alpha_{g}(E)=e^{-\frac{i}{\varepsilon}\int_{0}^{\varphi_{r}^{-}}\kappa_{i}}e^{\frac{1}{2}\int_{0}^{\varphi_{r}^{-}}(\omega_{+}^{i}-\omega_{-}^{i})},$$ $$\alpha_{d}(E)=e^{\frac{i}{\varepsilon}\int_{0}^{\varphi_{r}^{+}}\kappa_{i}}e^{\frac{1}{2}\int_{0}^{\varphi_{r}^{+}}(\omega_{+}^{i}-\omega_{-}^{i})}.$$ The coefficients $b_{g}^{-}$ and $b_{d}^{+}$ satisfy: $$b_{g}^{-}=\alpha_{g}a_{g}^{-},\quad b_{d}^{+}=\alpha_{d}a_{d}^{+},$$ where the coefficients $a_{g}^{-}$ and $a_{d}^{+}$ are given by: \begin{equation} \label{asag} a_{g}^{-}=\frac{w(f_{i},h_{-}^{g})}{w(f_{i},f_{i}^{*})}, \end{equation} and: \begin{equation} \label{asad} a_{d}^{+}=\frac{w(f_{i},h_{+}^{d})}{w(f_{i},f_{i}^{*})}. \end{equation} We compute, for $E\in\mathcal{U}_{0}$: $$d(\varphi,E,\varepsilon)=w(\alpha_{g}h_{-}^{g},\alpha_{d}h_{+}^{d})$$ $$=\left[b_{g}^{-}(b_{d}^{+})^{*}-b_{d}^{+}(b_{g}^{-})^{*}\right]i\sigma_{i}w(f_{i},(f_{i})^{*}).$$ \subsubsection{Continuation diagram of $f_{i}$} First, we describe the asymptotic behavior of the function $f_{i}$ in some domains of the complex plane. \begin{lem} \label{contdiag} We suppose that the assumptions of Proposition \ref{bcki} are satisfied. Fix $\tilde{Y}<Y$. Fix $\varphi_{g}<\varphi_{r}^{-}$ and $\varphi_{d}>\varphi_{r}^{+}$. There exists $y_{0}\in]0,\mbox{Im }\varphi_{i}[$ such that the function $f_{i}$ has the following asymptotic behavior: \begin{itemize} \item For $\varphi\in\{\varphi\in S_{\tilde{Y}};\ \mbox{Re }\varphi\in[\varphi_{g},\varphi_{r}^{-}]\}$, $f_{i}$ has the standard asymptotic behavior: $$f_{i}=q_{g}e^{\frac{i}{\varepsilon}\int_{0}^{\varphi_{r}^{-}}\kappa_{i}}e^{\frac{i}{\varepsilon}\int_{\varphi_{r}^{-}}^{\varphi}\kappa_{g}}e^{\int_{0}^{\varphi_{r}^{-}}\omega_{+}^{i}}e^{\int_{\varphi_{r}^{-}}^{\varphi}\omega_{+}^{g}}\left(\Psi_{+}^{g}+o(1)\right).$$ \item For $\varphi\in\{\varphi\in S_{\tilde{Y}};\ \mbox{Re }\varphi\in[\varphi_{r}^{+},\varphi_{d}];\ \mbox{Im }\varphi>-y_{0}\}$, $f_{i}$ has the standard asymptotic behavior: $$f_{i}=iq_{d}e^{\frac{i}{\varepsilon}\int_{0}^{\varphi_{r}^{+}}\kappa_{i}}e^{-\frac{i}{\varepsilon}\int_{\varphi_{r}^{+}}^{\varphi}\kappa_{d}}e^{\int_{0}^{\varphi_{r}^{+}}\omega_{+}^{i}}e^{\int_{\varphi_{r}^{+}}^{\varphi}\omega_{-}^{d}}\left(\Psi_{-}^{d}+o(1)\right).$$ \item For $\varphi\in\{\varphi\in S_{\tilde{Y}};\ \mbox{Re }\varphi\in[\varphi_{r}^{+},\varphi_{d}];\ \mbox{Im }\varphi<-y_{0}\}$, $f_{i}$ has the standard asymptotic behavior: $$f_{i}=-iq_{d}e^{\frac{i}{\varepsilon}\int_{0}^{\varphi_{i}^{-}}\kappa_{i}}e^{\frac{i}{\varepsilon}\int_{\varphi_{i}^{-}}^{\varphi}(2\pi-\kappa_{d})}e^{\int_{0}^{\varphi_{i}^{-}}\omega_{+}^{i}}e^{\int_{\varphi_{i}^{-}}^{\varphi}\omega_{-}^{d}}\left(\Psi_{-}^{d}+o(1)\right).$$ \end{itemize} \end{lem} \begin{dem} This lemma is similar to the continuation diagram presented in section 6 of \cite{FK4}. Thus, we give only the main ideas of the study and refer to this paper for the details. The continuation diagram is represented in figure \ref{contdiagfig}. In this figure, the straight arrows indicate the use of continuation lemma (Lemma \ref{lemcontfin}), the circular arrows the use of the Stokes lemma (Lemma \ref{stoklemma}) and the hatched zones the use of the Adjacent Canonical Domain Principle (Lemma \ref{adjdom}). To complete the proof, it remains to explain the connections between the different objects of the WKB method. \begin{itemize} \item According to the definitions given in section \ref{compmom}, the branches $\kappa_{i}$ and $\kappa_{g}$ are equal in $S_{-}$ and, for all $\varphi\in S_{-}$, we have: \begin{equation} \label{bda} \kappa_{i}(\varphi+0)=\kappa_{g}(\varphi-0),\quad\Psi_{\pm}^{i}(\varphi+0)=\Psi_{\pm}^{g}(\varphi-0),\quad \omega_{\pm}^{i}(\varphi+0)=\omega_{\pm}^{g}(\varphi-0). \end{equation} Besides, it remains to link $q_{i}$ and $q_{g}$. Section \ref{deco} implies: $$\forall\varphi\in S_{-},\quad q_{i}(\varphi)=q_{g}(\varphi).$$ \item Similarly, we have, for all $\varphi\in \overline{S_{+}}$: \begin{equation} \label{bdd} \kappa_{i}(\varphi-0)=-\kappa_{d}(\varphi+0), \quad\Psi_{\pm}^{i}(x,\varphi-0)=\Psi_{\mp}^{d}(x,\varphi+0), \end{equation} \begin{equation*} \omega_{\pm}^{i}(\varphi-0)=\omega_{\mp}^{d}(\varphi+0),\quad q_{d}(\varphi+0)=-i\ q_{i}(\varphi-0) \end{equation*} \item We study finally the link between $\kappa_{i}$ and $\kappa_{d}$ along the Stokes line $\bar{c}$ beginning at $\overline{\varphi_{i}}$. We consider the quasi-momenta $k_{d}$ and $k_{i}$ associated to $\kappa_{d}$ and $\kappa_{i}$. Equation \eqref{kl} for $k_{d}$ and $k_{i}$, on either side of $[E_{2},E_{3}]$, implies that $\kappa_{d}$ and $\kappa_{i}$ satisfy the following relations, for $\varphi\in c$, \begin{equation} \label{bdf} \kappa_{d}(\varphi+0)=2\pi-\kappa_{i}(\varphi-0),\quad\Psi_{\pm}^{d}(x,\varphi+0)=\Psi_{\mp}^{i}(x,\varphi-0) \end{equation} \begin{equation*} \omega_{\pm}^{d}(\varphi+0)=\omega_{\mp}^{i}(\varphi-0)\quad q_{d}(\varphi+0)=iq_{i}(\varphi-0) \end{equation*} \end{itemize} \end{dem} \input{fig11} \subsubsection{Computation of $b_{g}^{-}$ and $b_{d}^{+}$} Now, we compute the coefficients $b_{g}^{-}$ and $b_{d}^{+}$ given by \eqref{asag} and \eqref{asad}.\\ According to Theorem \ref{infcontle}, we know that the asymptotic behavior of the function $h_{-}^{g}$ remains valid in the domain $\{\varphi\in S_{Y};\ \mbox{Re }(\varphi)\in[\varphi_{g},\varphi_{r}^{-}]\}$. Lemma \ref{contdiag} gives the asymptotic behavior of $f_{i}$ in this domain and we get : \begin{equation} \forall \varphi\in S_{Y},\quad a_{g}^{-}(\varphi,E,\varepsilon)=\sigma_{i}[1+o(1)]. \end{equation} Fix $Y_{0}\in]0,Y[$. In the strip $S_{Y_{0}}$, we write: \begin{equation} a_{g}^{-}(\varphi,E,\varepsilon)=\sum\limits_{n\in \mathbb{Z}}\alpha_{n}e^{\frac{2i\pi n\varphi}{\varepsilon}} \end{equation} The coefficients $\alpha_{n}$ satisfy: \begin{equation} \label{four} \alpha_{n}=\frac{1}{\varepsilon}\int_{\varphi_{0}}^{\varphi_{0}+\varepsilon}a_{g}^{-}(\varphi,E,\varepsilon)e^{-2i\pi n\frac{\varphi}{\varepsilon}}d\varphi,\quad \forall n\in\mathbb{N},\quad\forall\varphi_{0}\in\{-Y_{0}\leq\mbox{Im }\varphi\leq Y_{0}\}. \end{equation} Fix $n>0$. We estimate $|\alpha_{n}|$. We use formula (\ref{four}) for $\mbox{Im }\varphi_{0}=-(Y-\delta) $, and we get: $$ |\alpha_{n}|\leq C e^{-2\pi n(Y-\delta)/\varepsilon}e^{\frac{C\delta}{\varepsilon}}.$$ We treat similarly the case $n<0$ with $\mbox{Im }\varphi_{0}=(Y-\delta)$ and we obtain: $$ |\alpha_{n}|\leq C e^{2\pi n(Y-\delta)/\varepsilon}e^{\frac{C\delta}{\varepsilon}}.$$ Besides, we have: $$\alpha_{0}=\sigma_{i}[1+o(1)].$$ We fix $\delta<\frac{2\pi(Y-Y_{0})}{C+2\pi}$. For a constant $C$ such that $\alpha<2\pi(Y-Y_{0})-\delta(C+2\pi)$, we obtain the estimates \eqref{bscoeffaa} and \eqref{bscoeffbb}.\\ The arguments for the coefficients $a_{d}^{+}$ and $b_{d}^{+}$ are similar. \subsection{Proof of Lemma \ref{phaseactint}} \label{phaseactintdem} Now, we want to express the coefficient $d$ in a more understandable form. We begin with proving Lemma \ref{phaseactint}. We recall that we denote by $\varphi_{r}^{\pm}$ and $\varphi_{i},\ \overline{\varphi_{i}}$ the branch points of the complex momentum, and by $E_{r}$ and $E_{i}$ the related ends of $\sigma(H_{0})$. We shall prove the lemma in the case \eqref{premcassc}. Let $\kappa_{i}$ be the branch described in section \ref{poss}. $\kappa_{i}$ satisfies \eqref{kappai}.\\ We shall prove Lemma \ref{phaseactint} for the branch $\widetilde{\kappa_{i}}=\kappa_{i}$. \begin{itemize} \item First, we express $\Phi$, $S$ and $\Phi_{d}$ as integrals of the complex momentum along complex paths. Let $\gamma$ be an oriented curve, we call $\gamma^{\dagger}$ the curve oriented in the opposite direction. Fix $\varphi_{d}\in\mathbb{R}$ and $\varphi_{g}\in\mathbb{R}$ such that: $$\varphi_{d}>\varphi_{r}^{+}\ ;\ \varphi_{g}<\varphi_{r}^{-}.$$ We define the complex paths $\gamma_{\Phi}$, $\gamma_{S}$ and $\gamma_{g,d}$: $$\gamma_{\Phi}=[\varphi_{r}^{-}+i0,\varphi_{r}^{+}+i0]\cup[\varphi_{r}^{+}-i0,\varphi_{r}^{-}-i0],$$ $$\gamma_{S}=(\sigma+0)\cup(\sigma^{\dagger}-0),$$ $$\gamma_{g,d}=[\varphi_{g}+i0,0+i0]\cup(\sigma_{+}-0)\cup(\sigma_{+}^{\dagger}+0)\cup[0+i0,\varphi_{d}+i0].$$ These paths are represented in figure \ref{uscoeff}. We have the following result: \begin{lem} \label{contour} The coefficients $\Phi$, $\Phi_{d}$ and $S$ can be written: $$\Phi=\frac{1}{2}\oint_{\gamma_{\Phi}}\kappa(u)du,$$ $$S=\frac{1}{2i}\oint_{\gamma_{S}}\kappa(u)du,$$ $$\Phi_{d}=\frac{1}{2}\left(\int_{\gamma_{g,d}}(\kappa(u)-\pi)du+\int_{\overline{\gamma_{g,d}}}(\widetilde{\kappa}(u)-\pi)du\right)+\pi(\varphi_{g}-\varphi_{d}).$$ where $\kappa=\kappa_{i}$ in $S_{-}$ and $\kappa$ is analytically continued along each path; $\widetilde{\kappa}=\kappa_{i}$ in $\overline{S_{-}}$ and $\widetilde{\kappa}$ is analytically continued along $\overline{\gamma_{g,d}}$. \end{lem} \begin{dem} \begin{itemize} \item First, let us justify the fact that integrals along $\gamma_{\Phi}$ and $\gamma_{S}$ can be considered along closed curves. It suffices to show that $\kappa$ can be analytically continued along $\gamma_{\Phi}$ and $\gamma_{S}$.\\ We consider the curve $\gamma_{\Phi}$. We have taken the cut of $\gamma_{\Phi}$ in $\varphi_{r}^{-}$. We show that $\kappa$ has the same values on each side of the cut. $\kappa=\kappa_{i}$ on $[\varphi_{r}^{-}+i0,\varphi_{r}^{+}+i0]$, since $\kappa$ is continuous to the right of $\varphi_{r}^{+}$, we obtain that $\kappa=-\kappa_{i}$ on $[\varphi_{r}^{-}-i0,\varphi_{r}^{+}-i0]$. In addition, $\kappa(\varphi_{r}^{-}+i0)=0=\kappa(\varphi_{r}^{-}-i0)$, which proves that the integral can be taken on the closed curve $\gamma_{\Phi}$ .\\ The arguments for $\gamma_{S}$ are similar. \item We compute: $$\frac{1}{2}\oint_{\gamma_{\Phi}}\kappa(u)du=\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}\kappa_{i}(u)du=\Phi(E).$$ Similarly, for the coefficient $S(E)$, $$\frac{1}{2i}\oint_{\gamma_{S}}\kappa(u)du=\frac{1}{i}[\int_{\sigma_{+}}(\pi-\kappa_{i}(u))du+\int_{\sigma_{-}}(\pi-\kappa_{i}(u))du].$$ \item It remains to study $\Phi_{d}$. We introduce the branch $\kappa_{i}$ and we cut $\gamma_{g,d}$ in elementary segments: $$\int_{\gamma_{g,d}}(\kappa(u)-\pi)du+\int_{\overline{\gamma_{g,d}}}(\widetilde{\kappa}(u)-\pi)du$$ $$=2\int_{\varphi_{g}}^{0}(\kappa_{i}(u)-\pi)du+2\int_{\sigma_{+}}(\kappa_{i}(u)-\pi)du+\int_{\varphi_{d}}^{0}(\kappa_{i}(u)-\pi)du-2\int_{\sigma_{-}}(\kappa_{i}(u)-\pi)du$$ $$=2\int_{\varphi_{r}^{-}}^{0}(\kappa_{i}(u)-\pi)du+2\int_{\sigma_{+}}(\kappa_{i}(u)-\pi)du+\int_{\varphi_{r}^{+}}^{0}(\kappa_{i}(u)-\pi)du-2\int_{\sigma_{-}}(\kappa_{i}(u)-\pi)du-2\pi(\varphi_{g}-\varphi_{r}^{-})+2\pi(\varphi_{d}-\varphi_{r}^{+})$$ $$=2\Phi_{d}(E)+2\pi(\varphi_{d}-\varphi_{g})$$ \end{itemize} This ends the proof of Lemma \ref{contour}. \end{dem} \item We use Lemma \ref{contour} to prove the analyticity of $\Phi$, $S$ and $\Phi_{d}$.\\ First, we consider $\Phi$. We can deform $\gamma_{\Phi}$ to a closed curve going around $[\varphi_{r}^{-},\varphi_{r}^{+}]$ and staying at a nonzero distance from this interval. Besides, $\kappa$ is analytic in $E$ on the integration contour when $E$ is close enough to $J$. The analysis of the coefficient $S$ is done in the same way. To prove that $\Phi_{d}$ is analytic, we deform the curves $\gamma_{g,d}$ and $\overline{\gamma_{g,d}}$ to stay at a nonzero distance of the cross. \item Fix $E\in J$. On the interval $[\varphi_{r}^{-},\varphi_{r}^{+}]$, the branch $\kappa_{i}$ satisfies $\kappa_{i}\in[0,\pi]$. Thus, the function $\Phi(E)$ is real positive on $J$.\\ Now, we give a simplified expression of $S$: \begin{equation} \label{exps} S(E)=-i\left[\int_{\sigma_{+}}(\pi-\kappa_{i}(u))du+\int_{\sigma_{-}}(\pi-\kappa_{i}(u))du\right]=2\mbox{Im }\int_{\sigma_{+}}(\pi-\kappa_{i}(u))du \end{equation} On $\sigma$, the branch $\kappa_{i}$ satisfies $\kappa_{i}\in[0,\pi]$. According to \eqref{action} and \eqref{exps}, we obtain that $0< S(E)\leq 2\pi\mbox{Im }\varphi_{i}(E)$.\\ Finally, we have $\int_{\sigma_{-}}(\kappa_{i}(u)-\pi)du=-\overline{\int_{\sigma_{+}}(\kappa_{i}(u)-\pi)du}$. Consequently, the coefficient $\Phi_{d}(E)$ is real. \item Now, we compute $S'$ and $\Phi'$ on $J$. Let $k$ be the branch of the Bloch momentum continuous through $[E_{r},E_{i}]$, then $\kappa(\varphi)=k(E-W(\varphi))$ and $$\Phi'(E)=\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}k'(E-W(u))du+k(E-W(\varphi_{r}^{+}))-k(E-W(\varphi_{r}^{-}))=\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}k'(E-W(u))du.$$ We recall that $k$ has some branch points of square root type at the ends of spectral bands (see section \ref{qm2}); consequently, the integral $\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}k'(E-W(u))du$ is convergent. In the interval $[E_{r},E_{i}]$, $k'(\mathcal{E})>0$ and $(E_{i}-E_{r})\Phi'$ takes positive values on $J$.\\ The analysis of $S'$ is similar. \item We complete this section with the following formulas: \begin{equation} \label{phida} \Phi_{d}(E)+iS(E)=\int_{\varphi_{r}^{-}}^{0}\kappa_{i}(u)du-2\int_{\sigma_{-}}(\kappa_{i}-\pi)(u)du+\int_{\varphi_{r}^{+}}^{0}\kappa_{i}(u)du \end{equation} \begin{equation} \label{phidb} -\Phi_{d}(E)+iS(E)=-\int_{\varphi_{r}^{-}}^{0}\kappa_{i}(u)du-2\int_{\sigma_{+}}(\kappa_{i}-\pi)(u)du-\int_{\varphi_{r}^{+}}^{0}\kappa_{i}(u)du \end{equation} When $\kappa(\varphi_{r}^{-})=\pi$, the proof is analogous for the branch $\widetilde{\kappa_{i}}=2\pi-\kappa_{i}$.\\ \end{itemize} \input{fig12} \subsubsection{Further computations} \label{comp} We recall that the functions $\omega_{+}^{i}$ and $\omega_{-}^{i}$ are defined in \eqref{omega}. We consider the integrals of $\omega_{+}^{i}$ and $\omega_{-}^{i}$ along some paths of the complex plane. We have the following relations: \begin{lem} \label{coeffut} The integrals of $\omega_{+}^{i}$ and $\omega_{-}^{i}$ satisfy: \begin{equation} \label{omegaa} \forall E\in J,\quad\int_{[\varphi_{r}^{-},\varphi_{r}^{+}]}\omega_{+}^{i}(u,E)du=0,\quad\int_{[\varphi_{r}^{-},\varphi_{r}^{+}]}\omega_{-}^{i}(u,E)du=0 \end{equation} \begin{equation} \label{omegab} \forall E\in J,\quad\int_{\sigma}\omega_{+}^{i}(u,E)du=0,\quad\int_{\sigma}\omega_{-}^{i}(u,E)du=0\end{equation} There exists a real number $\rho$ such that: \begin{equation} \label{omegac} \forall E\in J,\quad\int_{[\varphi_{r}^{+},0]\cup\sigma_{+}}(\omega_{+}^{i}(u,E)-\omega_{-}^{i}(u,E))du-\int_{\sigma_{-}\cup[0,\varphi_{r}^{-}]}(\omega_{+}^{i}(u,E)-\omega_{-}^{i}(u,E))du=i\rho \end{equation} \end{lem} \begin{dem} We consider the case \eqref{premcassc}. \begin{itemize} \item We first prove \eqref{omegaa}. According to \eqref{omega}, we compute: $$\int_{[\varphi_{r}^{-},\varphi_{r}^{+}]}\omega_{+}^{i}(u,E)du=-\int_{[\varphi_{r}^{-},\varphi_{r}^{+}]}g_{+}^{i}(E-W(u))W'(u)du=\int_{E-W([\varphi_{r}^{-},\varphi_{r}^{+}])}g_{+}^{i}(e)de=0$$ Indeed, for $E\in J$, the subset $E-W([\varphi_{r}^{-},\varphi_{r}^{+}])$ is a complex path of energies connecting $E_{r}$ to $E_{r}$ and containing $(E-W(0))\in]E_{1},E_{2}[)$. We have shown this path in figure \ref{imagchem}A. Particularly, $E-W([\varphi_{r}^{-},\varphi_{r}^{+}])$ is a closed path and does not surround any pole of the meromorphic function $g_{+}^{i}$. Consequently, the integral is zero. We prove similarly that $$\int_{[\varphi_{r}^{-},\varphi_{r}^{+}]}\omega_{-}^{i}(u,E)du=0.$$ \item We consider now \eqref{omegab}. We write: $$\int_{\sigma}\omega_{+}^{i}(u,E)du=-\int_{E-W(\sigma)}g_{+}^{i}(e)de$$ The image of the path $\sigma$ is shown in figure \ref{imagchem}B. We deal with $\omega_{-}^{i}$ similarly. \item Finally, we compute: $$\int_{[\varphi_{r}^{+},0]\cup\sigma_{+}}(\omega_{+}^{i}(u,E)-\omega_{-}^{i}(u,E))du-\int_{\sigma_{-}\cup[0,\varphi_{r}^{-}]}(\omega_{+}^{i}(u,E)-\omega_{-}^{i}(u,E))du$$ $$=\int_{E-W([\varphi_{r}^{+},0]\cup\sigma_{+})}(g_{+}^{i}(e)-g_{-}^{i}(e))de-\int_{E-W(\sigma_{-}\cup[0,\varphi_{r}^{-}])}(g_{+}^{i}(e)-g_{-}^{i}(e))de $$ The images $E-W([\varphi_{r}^{+},0]\cup\sigma_{+})$ and $E-W(\sigma_{-}\cup[0,\varphi_{r}^{-}])$ are two paths of energies connecting $E_{r}$ to $E_{i}$ (see figure \ref{imagchem}C). By analyticity of $(g_{+}^{i}-g_{-}^{i})$ in the domain $\mbox{Re }(e)\in]E_{r},E_{i}[$ , we obtain that: \begin{equation} \label{coeffrho} \int_{[\varphi_{r}^{+},0]\cup\sigma_{+}}(\omega_{+}^{i}(u,E)-\omega_{-}^{i}(u,E))du-\int_{\sigma_{-}\cup[0,\varphi_{r}^{-}]}(\omega_{+}^{i}(u,E)-\omega_{-}^{i}(u,E))du=2\int_{E_{r}}^{E_{i}}(g_{+}^{i}-g_{-}^{i})(e)de \end{equation} It remains to show that this coefficient is purely imaginary. To do that, we point out that $(g_{-}^{i})^{*}=g_{+}^{i}$, according to \eqref{symband}. Equation \eqref{coeffrho} becomes: $$\int_{[\varphi_{r}^{+},0]\cup\sigma_{+}}(\omega_{+}^{i}-\omega_{-}^{i})(u,E)du-\int_{\sigma_{-}\cup[0,\varphi_{r}^{-}]}(\omega_{+}^{i}-\omega_{-}^{i})(u,E)du$$ $$=2\left(\int_{E_{r}}^{E_{i}}g_{+}^{i}(e)de-\int_{E_{r}}^{E_{i}}(g_{+}^{i})^{*}(e)de\right)=2\left(\int_{E_{r}}^{E_{i}}g_{+}^{i}(e)de-\overline{\int_{E_{r}}^{E_{i}}g_{+}^{i}(e)de}\right)$$ \end{itemize} This ends the proof of Lemma \ref{coeffut}. \end{dem} \input{fig13} \subsection{Equation for the eigenvalues} \label{demeigeneq} The following result gives a characterization of the eigenvalues of $H_{\varphi,\varepsilon}$. \begin{prop} \label{eigeneq2} We assume that $(H_{V})$, $(H_{W,r})$, $(H_{W,g })$ and $(H_{J})$ are satisfied.\\ There exist $\varepsilon_{0}>0$, a neighborhood $\mathcal{V}=\overline{\mathcal{V}}$ of $J$, two functions $(E,\varepsilon)\mapsto\widetilde{\Phi}(E,\varepsilon)$ and $(E,\varepsilon)\mapsto\widetilde{\Phi_{d}}(E,\varepsilon)$ defined on $\mathcal{V}\times]0,\varepsilon_{0}[$ and two functions $(\varphi,E,\varepsilon)\mapsto F(\varphi,E,\varepsilon)$ and $(\varphi,E,\varepsilon)\mapsto R_{2}(\varphi,E,\varepsilon)$ defined on $\mathbb{R}\times\mathcal{V}\times]0,\varepsilon_{0}[$ such that: \begin{enumerate}\item $E$ is an eigenvalue of $H_{\varphi,\varepsilon}$ if and only if: $$ F(\varphi,E,\varepsilon)=0$$ \item The function $F$ satisfies: $$\forall\varphi\in \mathbb{R},\ \forall E\in\mathcal{V},\ \forall\varepsilon\in]0,\varepsilon_{0}[,\quad F^{*}(\varphi,E,\varepsilon)=\overline{F(\overline{\varphi},\overline{E},\varepsilon)}=F(\varphi,E,\varepsilon).$$ \item The function $\varphi\mapsto F(\varphi,E,\varepsilon)$ is $\varepsilon$-periodic and its Fourier expansion is written: \begin{equation} \label{decsfoud} F(\varphi,E,\varepsilon)=\cos\left(\frac{\widetilde{\Phi}(E)}{\varepsilon}\right)+e^{-S(E)/\varepsilon}\cos\left(\frac{\widetilde{\Phi_{d}}(E)}{\varepsilon}+\frac{2\pi\varphi}{\varepsilon}+\rho\right)+e^{-S(E)/\varepsilon}R_{2}(\varphi,E,\varepsilon) \end{equation} \item The functions $\widetilde{\Phi}$, $\widetilde{\Phi_{d}}$ satisfy the following properties for any $\varepsilon\in]0,\varepsilon_{0}[$: \begin{itemize}\item $E\mapsto\widetilde{\Phi}(E,\varepsilon)$ and $E\mapsto\widetilde{\Phi_{d}}(E,\varepsilon)$ are analytic on $\mathcal{V}$. \item $\widetilde{\Phi}(E,\varepsilon)=\Phi(E)+o(\varepsilon)$ and $\widetilde{\Phi_{d}}(E,\varepsilon)=\Phi_{d}(E)+o(\varepsilon)$ uniformly for $E\in\mathcal{V}$. \end{itemize} \item For any $\varepsilon\in]0,\varepsilon_{0}[$, the function $(\varphi,E)\mapsto R_{2}(\varphi,E,\varepsilon)$ is analytic on $\mathbb{R}\times\mathcal{V}$. Besides, there exists a constant $\alpha>0$ such that,for all $\varepsilon\in]0,\varepsilon_{0}[$, and all $E$ in $\mathcal{V}$, the function $R_{2}$ satisfies the following properties: $$ \int_{0}^{\varepsilon}R_{2}(u,E,\varepsilon)du=0,\quad\int_{0}^{\varepsilon}R_{2}(u,E,\varepsilon)e^{\frac{ 2i\pi u}{\varepsilon}}du=0,\quad\int_{0}^{\varepsilon}R_{2}(u,E,\varepsilon)e^{\frac{-2i\pi u}{\varepsilon}}du=0,$$ $$\sup\limits_{\varphi\in\mathbb{R},E\in\mathcal{V}}|R_{2}(\varphi,E,\varepsilon)|\leq e^{-\frac{\alpha}{\varepsilon}}$$ \end{enumerate} The functions $\Phi$, $\Phi_{d}$, $S$ are defined in Lemma \ref{phaseactint}. $\rho$ is a real number defined in \eqref{omegac}. \end{prop} Now, we prove Proposition \ref{eigeneq2}. \begin{itemize}\item Now, it suffices to compute the Fourier expansion of: $$b_{g}^{-}(b_{d}^{+})^{*}(\varphi,E,\varepsilon)=\sum\limits_{n\in\mathbb{Z}}\gamma_{n}(E,\varepsilon)e^{\frac{2i n\pi\varphi }{\varepsilon}}.$$ By using the asymptotic expansion of the coefficients $b_{g}^{-}$ and $b_{d}^{+}$ given in Lemma \ref{colina}, we prove that: $$\gamma_{0}=-ie^{-\frac{i}{\varepsilon}\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}\kappa_{i}} [1+o(1)].$$ $$\gamma_{1}=+ie^{-\frac{i}{\varepsilon}(\int_{\varphi_{r}^{+}}^{0}\kappa_{i}+\int_{\varphi_{r}^{-}}^{0}\kappa_{i})}e^{\frac{2i}{\varepsilon}\int_{0}^{\overline{\varphi_{i}}}(\kappa_{i}-\pi)}e^{\int_{0}^{\varphi_{r}^{+}}(\omega_{+}^{i}-\omega_{-}^{i})}e^{\int_{0}^{\varphi_{i}}(\omega_{-}^{i}-\omega_{+}^{i})}[1+o(1)].$$ $$\left|\sum\limits_{n\in\mathbb{Z}\backslash\{0,1\}}\gamma_{n}e^{\frac{2i n\pi\varphi }{\varepsilon}}\right|=O(e^{-\alpha/\varepsilon})\quad\textrm{ pour }\varphi\in S_{Y_{0}}.$$ Actually, $$\gamma_{0}=\alpha_{0}\beta_{0}+\sum\limits_{n\neq 0}\alpha_{n}\beta_{-n}=-i e^{-\frac{i}{\varepsilon}\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}\kappa_{i}}e^{\frac{1}{2}\left[\int_{0}^{\varphi_{r}^{-}}(\omega_{+}^{i}-\omega_{-}^{i})+\int_{0}^{\varphi_{r}^{-}}(\omega_{-}^{i}-\omega_{+}^{i})\right]}[1+o(1)].$$ According to \eqref{omegaa}, we simplify: $$\left[\int_{0}^{\varphi_{r}^{-}}(\omega_{+}^{i}-\omega_{-}^{i})+\int_{0}^{\varphi_{r}^{-}}(\omega_{-}^{i}-\omega_{+}^{i})\right]=0.$$ According to \eqref{phi}, $\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}\kappa_{i}=\Phi(E)$. Consequently, $$\gamma_{0}=-i e^{\frac{i\Phi(E)}{\varepsilon}}[1+o(1)].$$ We compute: $$\gamma_{1}=\alpha_{0}\beta_{1}+\sum\limits_{n\neq 1}\alpha_{n}\beta_{1-n}$$ We start with computing $\alpha_{0}\beta_{1}$. To do that, we deduce from equation \eqref{phida} that: $$\int_{\varphi_{r}^{+}}^{0}\kappa_{i}(\varphi)d\varphi+\int_{\varphi_{r}^{-}}^{0}\kappa_{i}(\varphi)d\varphi-\int_{\sigma_{-}}(\kappa_{i}(\varphi)-\pi)d\varphi=\Phi_{d}(E)+iS(E).$$ $$\alpha_{0}\beta_{1}=ie^{-\frac{i}{\varepsilon}(\int_{\varphi_{r}^{+}}^{0}\kappa_{i}+\int_{\varphi_{r}^{-}}^{0}\kappa_{i})}e^{\frac{2i}{\varepsilon}\int_{0}^{\overline{\varphi_{i}}}(\kappa_{i}-\pi)}e^{\int_{0}^{\varphi_{r}^{+}}\omega_{+}^{i}+\int_{\varphi_{r}^{-}}^{0}\omega_{-}^{i}-\int_{\sigma_{+}}(\omega_{+}^{i}-\omega_{-}^{i})}[1+o(1)]+O(e^{\frac{-\alpha}{\varepsilon}})e^{-S(E)/\varepsilon}.$$ Equation \eqref{phida} leads to: $$\int_{\varphi_{r}^{+}}^{0}\kappa_{i}(\varphi)d\varphi+\int_{\varphi_{r}^{-}}^{0}\kappa_{i}(\varphi)d\varphi-\int_{\sigma_{-}}(\kappa_{i}(\varphi)-\pi)d\varphi=\Phi_{d}(E)+iS(E).$$ Besides, according to Lemma \ref{coeffut}, we have: $$\int_{0}^{\varphi_{r}^{+}}\omega_{+}^{i}+\int_{\varphi_{r}^{-}}^{0}\omega_{-}^{i}-\int_{\sigma_{+}}(\omega_{+}^{i}-\omega_{-}^{i})=i\rho.$$ and: $$\alpha_{0}\beta_{1}=ie^{-S/\varepsilon}e^{-i\Phi_{d}/\varepsilon}e^{i\rho}[1+o(1)].$$ Since $S(E)\leq 2\pi\mbox{Im }\varphi_{i}(E)$, we estimate the remainder in the expansion: $$|\sum\limits_{n\neq 0}\alpha_{n}\beta_{-n}|=o(e^{-S/\varepsilon}).$$ Finally, for $p\neq 0,1$, we estimate: $$\gamma_{p}=\sum\limits_{n\in\mathbb{Z}}\alpha_{n}\beta_{p-n}.$$ For $p>1$, we have: $$|\gamma_{p}|=e^{-S/\varepsilon}e^{-\alpha/\varepsilon}O(e^{-\frac{2\pi Y_{0}(p-1)}{\varepsilon}}).$$ Similarly, we estimate for $p<0$, $$|\gamma_{p}|=e^{-S/\varepsilon}e^{-\alpha/\varepsilon}O(e^{-\frac{2\pi Y_{0}(|p|-1)}{\varepsilon}}).$$ \item Now, we consider $\varphi\in\mathbb{R}$. We compute the Fourier asymptotic expansion of the coefficient $d(E,\varphi,\varepsilon)$ in a neighborhood $\mathcal{U}_{0}$ of $E_{0}$: $$d(\varphi,E,\varepsilon)=iw(f_{i},\sigma_{i}(f_{i})^{*})\left(\lambda_{0}(E,\varepsilon)+\sum\limits_{n\in\mathbb{N}^{*}}(\lambda_{n}(E,\varepsilon)e^{\frac{2i n\pi\varphi}{\varepsilon}}+(\lambda_{n})^{*}(E,\varepsilon)e^{\frac{-2i n\pi\varphi}{\varepsilon}})\right)$$ $$=i(w_{0}k'_{i})(E-W(0))\sum\limits_{n\in\mathbb{N}}u_{n}(\varphi,E,\varepsilon).$$ where $u_{n}(\varphi,E,\varepsilon)=\lambda_{n}(E,\varepsilon)e^{\frac{2i n\pi\varphi}{\varepsilon}}+(\lambda_{n})^{*}(E,\varepsilon)e^{\frac{-2i n\pi\varphi}{\varepsilon}}$, pour $n\in\mathbb{N}^{*}$, et $u_{0}(\varphi,E,\varepsilon)=\lambda_{0}(E,\varepsilon)$.\\ We have: $$u_{0}(\varphi,E,\varepsilon)=\gamma_{0}(E,\varepsilon)-\gamma_{0}^{*}(E,\varepsilon)=-ie^{i\frac{\Phi}{\varepsilon}}g(E,\varepsilon)-ie^{-i\frac{\Phi}{\varepsilon}}g^{*}(E,\varepsilon).$$ where $g(E,\varepsilon)=1+o(1)$.\\ We define $g(E,\varepsilon)=r_{g}(E,\varepsilon)e^{i\theta_{g}(E,\varepsilon)}$ where the functions $E\mapsto r_{g}(E,\varepsilon)$ and $E\mapsto \theta_{g}(E,\varepsilon)$ are analytic and satisfy $$ r_{g}^{*}=r_{g},\quad r_{g}=1+o(1)\quad \theta_{g}^{*}=\theta_{g},\quad\theta_{g}=o(1).$$ We simplify: $$u_{0}(\varphi,E,\varepsilon)=-i r_{g}(E,\varepsilon)\cos\left(\frac{\Phi(E)}{\varepsilon}+\theta_{g}(E,\varepsilon)\right).$$ Similarly, we compute: $$u_{1}(\varphi,E,\varepsilon)=i r_{h}(E,\varepsilon)e^{-S(E)/\varepsilon}\cos\left(\frac{\Phi_{d}+2\pi\varphi}{\varepsilon}+\rho+\theta_{h}(E,\varepsilon)\right).$$ where the functions $E\mapsto r_{h}(E,\varepsilon)$ and $E\mapsto \theta_{h}(E,\varepsilon)$ are analytic and satisfy $$ r_{h}^{*}=r_{h},\quad r_{h}=1+o(1)\quad \theta_{h}^{*}=\theta_{h},\quad\theta_{h}=o(1).$$ In addition, we have the following estimate of the remainder: $$\left|\sum\limits_{p\geq 2}u_{p}(\varphi,E,\varepsilon)\right|\leq Ce^{\frac{-S(E)}{\varepsilon}}e^{\frac{-\alpha}{\varepsilon}}\quad\textrm{ pour }\varphi\in \mathbb{R}.$$ \item We have proved that, for $E$ in a neighborhood of $E_{0}$, the Fourier expansion of $d(E,\varphi,\varepsilon)$ can be written: \begin{equation} \label{decsfou} \frac{d(\varphi,E,\varepsilon)}{i(w_{0}k'_{i})(E-W(0))}=-i[1+o(1)]\cos\left(\frac{\Phi(E)}{\varepsilon}+o(1)\right) \end{equation} $$+i[1+o(1)]e^{\frac{-S(E)}{\varepsilon}}\cos\left(\frac{\Phi_{d}+2\pi\varphi}{\varepsilon}+\rho+o(1)\right)+e^{\frac{-S(E)}{\varepsilon}}O(e^{\frac{-\alpha}{\varepsilon}}).$$ The compactness of $J$ implies that there exists a finite number of intervals $\{J_{k}\}_{k\in\{1\cdots p \}}$ such that: \begin{enumerate} \item $J\subset\bigcup\limits_{k\in\{1\cdots p \}}J_{k}$ \item For any $k\in\{1,\cdots ,p-1 \}$, the intervals $J_{k}$ and $J_{k+1}$ overlap.\item For any $k\in\{1,\cdots, p \}$, there exists a complex neighborhood $\mathcal{U}_{k}$ of $J_{k}$ such that the expansion (\ref{decsfou}) is satisfied on $\mathcal{U}_{k}$. \end{enumerate} We shall prove that we can define some functions $\widetilde{\Phi}$ and $\widetilde{\Phi}_{d}$ on the whole neighborhood $\mathcal{V}=\bigcup\limits_{k\in\{1\cdots p \}}\mathcal{U}_{k}$. To do that, we shall ``stick'' the expansions obtained on each interval.\\ The coefficient $u_{0}$ is written: $$\forall E\in J_{k},\quad u_{0}(E,\varepsilon)=r_{0,k}(E,\varepsilon)\cos\left(\frac{\Phi(E)}{\varepsilon}+\theta_{0,k}(E,\varepsilon)\right)$$ $$\forall E\in J_{k+1},\quad u_{0}(E,\varepsilon)=r_{0,k+1}(E,\varepsilon)\cos\left(\frac{\Phi(E)}{\varepsilon}+\theta_{0,k+1}(E,\varepsilon)\right)$$ where $r_{0,k}(E,\varepsilon)=1+o(1)$ and $\theta_{0,k}(E,\varepsilon)=o(1)$ (resp. $r_{0,k+1}(E,\varepsilon)=1+o(1)$ and $\theta_{0,k+1}(E,\varepsilon)=o(1)$) for $E\in J_{k}$ (resp. $E\in J_{k+1}$).We get that: $$r_{0,k}(E,\varepsilon)=r_{0,k+1}(E,\varepsilon)=r_{0}(E,\varepsilon)\textrm{ et } \theta_{0,k}(E,\varepsilon)=\theta_{0,k+1}(E,\varepsilon)=\theta_{0}(E,\varepsilon)\textrm{ for }E\in J_{k}\cap J_{k+1}$$ The function $\widetilde{\Phi}$ defined by its restrictions to each $\mathcal{U}_{k}$ is analytic on $\mathcal{V}$.\\ The case of $\widetilde{\Phi_{d}}$ is treated similarly. \end{itemize} Defining $$F(\varphi,E,\varepsilon)=\frac{d(\varphi,E,\varepsilon)}{i(w_{0}k'_{i})(E-W(0))r_{0}(E,\varepsilon)},$$ we finish the proof of Proposition \ref{eigeneq2}. \subsection{Localization of the eigenvalues} In this section, we deduce Theorem \ref{eigenloc} from Proposition \ref{eigeneq2}. \\ We solve equation $F(\varphi,E,\varepsilon)=0$, where $F$ is described in (\ref{decsfoud}). \subsubsection{Energy levels $E^{(l)}(\varepsilon)$} \label{koe} For $E\in \mathcal{V}$, we start with solving: \begin{equation} \label{phasemodif} \cos\frac{\widetilde{\Phi}(E,\varepsilon)}{\varepsilon}=0 \end{equation} $E\mapsto\widetilde{\Phi}(E,\varepsilon)$ is a real analytic function. For a sufficiently small $\varepsilon_{0}$, by Lemma \ref{phaseactint}, there exists a constant $m>0$ such that: \begin{equation} \label{phasemodifdiff} \forall E\in\mathcal{V},\quad\forall\varepsilon\in]0,\varepsilon_{0}[,\quad |\widetilde{\Phi}'(E,\varepsilon)|\geq m \end{equation} Consequently, equation \eqref{phasemodif} has a finite number of zeros in $J$. We denote them by $E^{(l)}(\varepsilon)$, for $l\in\{L_{-}(\varepsilon),\dots,L_{+}(\varepsilon)\}$. They are given by: \begin{equation} \label{zeroa} \frac{\widetilde{\Phi}(E^{(l)}(\varepsilon),\varepsilon)}{\varepsilon}=l\pi+\frac{\pi}{2},\quad\quad\forall l\in\{L_{-}(\varepsilon),\ldots,L_{+}(\varepsilon)\}. \end{equation} and satisfy: \begin{equation} \label{ecart} E^{(l+1)}(\varepsilon)-E^{(l)}(\varepsilon)=\frac{1}{\widetilde{\Phi}'(E^{(l)}(\varepsilon))}\pi\varepsilon+o(\varepsilon). \end{equation} The distances between two consecutive zeros are of order $\varepsilon$. Precisely, by combining \eqref{phasemodifdiff} with \eqref{ecart}, we obtain that there exists a constant $c>0$ such that: \begin{equation} \label{ecarta} \frac{1}{c}\varepsilon<|E^{(l+1)}(\varepsilon)-E^{(l)}(\varepsilon)|<c\varepsilon,\quad \forall l\in \{L_{-}(\varepsilon),\ldots,L_{+}(\varepsilon)-1\} \end{equation} First, we prove that the zeros of $F$ are in an exponentially small neighborhood of the points $E^{(l)}(\varepsilon)$. \subsubsection{First order approximation} We give a first order approximation of the zeros of $F$.\\ We set $$a_{0}(E,\varepsilon)=\cos\frac{\widetilde{\Phi}(E,\varepsilon)}{\varepsilon}.$$ We can assume that the neighborhood $\mathcal{V}$ is sufficiently small and such that, for any $E\in\mathcal{V}$, $$\mbox{Re }(S(E))>\beta>0.$$ Then, there exists a positive constant $A$ such that $$|F(\varphi,E,\varepsilon)-a_{0}(E,\varepsilon)|<Ae^{-\beta/\varepsilon}.$$ In addition, we have the following inequality: \begin{equation} \label{cosinus} \exists C>0/\quad \left|\cos\frac{\widetilde{\Phi}(E,\varepsilon)}{\varepsilon}\right|\geq\frac{C}{\varepsilon}d(E,\bigcup\limits_{l\in\{L_{-},\cdots,L_{+}\}}E^{(l)}(\varepsilon)). \end{equation} Actually, there exists a constant $c>0$ such that: $$|\cos\theta|\geq c d(\theta,\pi\mathbb{Z}+\pi/2).$$ By using \eqref{phasemodifdiff}, we obtain the relation: $$|\widetilde{\Phi}(E,\varepsilon)-\widetilde{\Phi}(E^{(l)}(\varepsilon),\varepsilon)|\geq m |E-E^{(l)}(\varepsilon)|$$ and finally: $$\left|\cos\frac{\widetilde{\Phi}(E,\varepsilon)}{\varepsilon}\right|\geq\frac{C}{\varepsilon}d\left(E,\bigcup\limits_{l\in\{L_{-},\cdots,L_{+}\}}E^{(l)}(\varepsilon)\right).$$ For $z_{0}\in\mathbb{C}$ and $r>0$, we define $$D(z_{0},r)=\{z\in\mathbb{C}\ ;\ |z-z_{0}|<r\}.$$ Inequality (\ref{cosinus}) implies that there are no zeros of $F$ outside exponentially small neighborhoods of the points $E^{(l)}(\varepsilon)$. Precisely, there exists a positive constant $D$ such that, if $r\geq D\varepsilon e^{-\beta/\varepsilon}$, then for any $E\in \partial D(E^{(l)}(\varepsilon),r)$, we have: $$|F(\varphi,E,\varepsilon)-a_{0}(E,\varepsilon)|<|a_{0}(E,\varepsilon)|.$$ Rouch{\'e}'s Theorem implies that, for any $l$, $F$ has exactly one zero $E_{l}(\varphi,\varepsilon)$, in each neighborhood $D(E^{(l)}(\varepsilon),D\varepsilon e^{-\beta/\varepsilon})$ of $E^{(l)}(\varepsilon)$. The relation $F=F^{*}$ allows us to recover that the eigenvalues are real. Indeed, if $F(E)=0$, $\overline{E}$ is also a zero of $F$. By uniqueness, we obtain that $E=\overline{E}$.\\ We set: $$E_{l}(\varphi,\varepsilon)=E^{(l)}(\varepsilon)+\varepsilon\lambda_{l}(\varphi,\varepsilon).$$ We know that $\lambda_{l}(\varphi,\varepsilon)$ is exponentially small. Now, we compute its asymptotic behavior. \subsubsection{Second order approximation} We define: $$a_{1}(\varphi,E,\varepsilon)=F(\varphi,E,\varepsilon)-a_{0}(E,\varepsilon).$$ We write $$e^{-S(E_{l}(\varphi,\varepsilon))/\varepsilon}=e^{-S(E^{(l)}(\varepsilon))/\varepsilon}(1+O(\lambda_{l}(\varphi,\varepsilon))).$$ Similarly, with the help of the modified phase $\widetilde{\Phi_{d}}$, we obtain the expansion: $$\cos\left(\frac{\widetilde{\Phi_{d}}(E_{l}(\varphi,\varepsilon))+2\pi\varphi+\rho\varepsilon}{\varepsilon}\right)=\cos\left(\frac{\widetilde{\Phi_{d}}(E^{(l)}(\varepsilon))+2\pi\varphi+\rho\varepsilon}{\varepsilon}\right)+O(e^{-\beta/\varepsilon})$$ The expansion of $a_{1}$ can be written: $$a_{1}(\varphi,E_{l}(\varphi,\varepsilon),\varepsilon)=a_{1}(\varphi,E^{(l)}(\varepsilon),\varepsilon)(1+r(\varphi,E^{(l)}(\varepsilon),\varepsilon)).$$ Moreover, we use the first order Taylor's expansion of the function $E\mapsto a_{0}(E,\varepsilon)$: $$a_{0}(E_{l}(\varphi,\varepsilon),\varepsilon)=(-1)^{l+1}\widetilde{\Phi}'(E^{(l)}(\varepsilon),\varepsilon)\lambda_{l}(\varphi,\varepsilon)(1+r(\varphi,E^{(l)},\varepsilon))=(-1)^{l+1}\Phi'(E^{(l)})\lambda_{l}(\varphi,\varepsilon)(1+o(1)).$$ By combining these computations, we finally obtain: $$\lambda_{l}(\varphi,\varepsilon)=\frac{(-1)^{l+1}}{\Phi'(E^{(l)}(\varepsilon))}e^{-S(E^{(l)}(\varepsilon))/\varepsilon}\left(\cos\left(\frac{\widetilde{\Phi_{d}}(E^{(l)}(\varepsilon),\varepsilon)+2\pi\varphi}{\varepsilon}+\rho\right)+o(1)\right).$$ \subsection{Application to the trace formula} \label{trform2} In \cite{Di1}, the author proves the existence of an asymptotic expansion of ${\mathrm tr} [f(H_{\varphi,\varepsilon})]$, for $f\in C_{0}^{\infty}$, when $\textrm{Supp }f$ is disjoint from the bands of $H_{0}$; in addition, he computes explicitly the first and second terms of this expansion.\\ Corollary \ref{cordim} allows us to recover these terms. \subsubsection{} Let $J$ be an interval satisfying $(H_{J})$. Particularly, $J$ is such that $J\cap(\sigma_{ac}\cup\sigma_{sc})=\emptyset$. For $f\in C_{0}^{\infty}$, with $\textrm{Supp }f\subset J$, we compute: $${\mathrm tr}[f(H_{\varphi,\varepsilon})]=\sum\limits_{l\in\{L_{-}(\varepsilon),...,L_{+}(\varepsilon)\}}f(E_{l}(\varphi,\varepsilon)).$$ Let $\beta>0$ be such that $S(E)>\beta$ for any $E\in J$; according to Theorem \ref{eigenloc}, we know that there exists a constant $C>0$ such that: $$\forall u\in[0,\varepsilon],\quad\left|{\mathrm tr}[f(H_{\varphi,\varepsilon})]-{\mathrm tr}[f(H_{u,\varepsilon})]\right|<C\sum\limits_{l\in\{L_{-}(\varepsilon),...,L_{+}(\varepsilon)\}}\varepsilon e^{-\beta/\varepsilon}.$$ By integrating with respect to $u$, we obtain that: $$\textrm{tr }[f(H_{\varphi,\varepsilon})]=\frac{1}{\varepsilon}\int_{0}^{\varepsilon}\textrm{tr }[f(H_{u,\varepsilon})]du+O(e^{-\beta/\varepsilon}).$$ According to Theorem \ref{eigenloc}, we know that there exists a constant $C$ such that $$\forall u\in[0,\varepsilon],\quad\left|{\mathrm tr} [f(H_{u,\varepsilon})]-\sum\limits_{l\in\{L_{-}(\varepsilon),...,L_{+}(\varepsilon)\}}f(E^{(l)}(\varepsilon))\right|<C e^{-\beta/\varepsilon}$$ By integration, we obtain: \begin{equation} \label{formtrace} \frac{1}{\varepsilon}\int_{0}^{\varepsilon}\textrm{tr }[f(H_{u,\varepsilon})]du=\sum\limits_{l\in\{L_{-}(\varepsilon),...,L_{+}(\varepsilon)\}}f(E^{(l)}(\varepsilon))+O(e^{-\beta/\varepsilon}) \end{equation} Now, we estimate: $$\sum\limits_{l\in\{L_{-}(\varepsilon),...,L_{+}(\varepsilon)\}}f(E^{(l)}(\varepsilon))=\sum\limits_{l\in\{L_{-}(\varepsilon),...,L_{+}(\varepsilon)\}}f\circ\widetilde{\Phi}^{-1}(\varepsilon(l\pi+\pi/2))$$ \subsubsection{} Now, we compute this last term. \begin{lem} \label{trace} Let $f$ be a function in $C_{0}^{\infty}$ such that $\textrm{Supp }f\subset J$. The trace of $H_{\varphi,\varepsilon}$ has the following asymptotic behavior: $$\int_{0}^{\varepsilon}{\mathrm tr} [f(H_{u,\varepsilon})]du=\frac{1}{\pi}\int_{J} f(\mathcal{E})\widetilde{\Phi}'(\mathcal{E},\varepsilon)d\mathcal{E}+O(\varepsilon^{\infty})$$ \end{lem} \begin{dem} The proof of this Lemma is based on elementary results of real analysis. \begin{itemize} \item We apply the Poisson formula to the function $f\circ\widetilde{\Phi}^{-1}\in C_{0}^{\infty}$: $$\varepsilon\sum\limits_{l\in\mathbb{Z}}f\circ\widetilde{\Phi}^{-1}(\varepsilon(l\pi+\pi/2))=2\sum\limits_{n\in\mathbb{Z}}(-1)^{n}\widehat{(f\circ\widetilde{\Phi}^{-1})}\left(\frac{2n}{\varepsilon}\right).$$ Besides, the Fourier transform of $f\circ\widetilde{\Phi}^{-1}$ satisfies the estimates: $$\forall\nu>1,\quad\exists\ C_{\nu}>0,\textrm{ such that }\left|\widehat{(f\circ\widetilde{\Phi}^{-1})}\left(\frac{2n}{\varepsilon}\right)\right|\leq C_{\nu}\frac{\varepsilon^{\nu}}{n^{\nu}}.$$ Actually, since $f\circ\widetilde{\Phi}^{-1}$ is $C^{\nu}$, $|\xi^{\nu}\widehat{f\circ\widetilde{\Phi}^{-1}}(\xi)|$ is bounded.\\ This leads to: $$\varepsilon\sum\limits_{p\in\mathbb{Z}}f\circ\widetilde{\Phi}^{-1}(\varepsilon(p\pi+\pi/2))=2\widehat{(f\circ\widetilde{\Phi}^{-1})}(0)+O(\varepsilon^{\infty}).$$ \item It remains to prove that: $$2\widehat{(f\circ\widetilde{\Phi}^{-1})}(0)=\frac{1}{\pi}\int f\circ\widetilde{\Phi}^{-1}(u)du.$$ With the substitution $u=\widetilde{\Phi}(\mathcal{E})$, we obtain that: $$2\widehat{(f\circ\widetilde{\Phi}^{-1})}(0)=\frac{1}{\pi}\int_{J} f(\mathcal{E})\widetilde{\Phi}'(\mathcal{E},\varepsilon)d\mathcal{E}$$ This completes the proof of Lemma \ref{trace}. \end{itemize} \end{dem} \subsubsection{Conclusion} To get an asymptotic expansion of the trace at any order, it suffices to know an asymptotic expansion of the modified phase at any order. Our computations are not accurate enough, but we know that $\widetilde{\Phi}'(\mathcal{E})=\Phi'(E)+o(\varepsilon)$, hence: $$\frac{1}{\pi}\int_{J} f(\mathcal{E})\widetilde{\Phi}'(\mathcal{E})d\mathcal{E}=\frac{1}{\pi}\int_{J} f(\mathcal{E})\Phi'(\mathcal{E})d\mathcal{E}+o(\varepsilon).$$ To transform the right member of previous equality, we do the substitution $(\kappa,u)\mapsto(E(\kappa)+W(u),u)$, which implies: $$\frac{1}{\pi}\int_{J} f(\mathcal{E})\Phi'(\mathcal{E})d\mathcal{E}=\frac{1}{2\pi}\int_{[-\pi,\pi]}\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}f(E(\kappa)+W(u))d\kappa du$$ We finally obtain: $$\int_{0}^{\varepsilon}\textrm{tr }[f(H_{u,\varepsilon})]du=\frac{1}{2\pi}\int_{[-\pi,\pi]}\int_{\varphi_{r}^{-}}^{\varphi_{r}^{+}}f(E(\kappa)+W(u))d\kappa du+o(\varepsilon)$$ This ends the proof of Corollary \ref{cordim}. \subsection{Asymptotic behavior of the eigenvalues} Now, we give a second application of Theorem \ref{eigenloc} for the computation of the asymptotic behavior of the eigenvalues of $H_{\varphi,\varepsilon}$. Such a computation is outlined in \cite{CDS}, in the case $V=0$. We obtain an explicit result at first order.\\ Under the assumptions of Theorem \ref{eigenloc}, $E_{r}$ is the only end of $\sigma(H_{0})$ belonging to $(E-W)(\mathbb{R})$. We define: \begin{equation} \label{rnp} d_{p}(E_{n})=\lim\limits_{\stackrel{E\rightarrow E_{n},}{ E\in[E_{n},E_{p}]}}\frac{k(E)-k(E_{n})}{\sqrt{E-E_{n}}} \end{equation} \begin{cor} Let $H_{\varphi, \varepsilon}$ verify the assumptions of Theorem \ref{eigenloc}. The eigenvalues $E^{(l)}(\varphi,\varepsilon)$ of $H_{\varphi,\varepsilon}$ have the following asymptotic behavior: $$E^{(l)}(\varphi,\varepsilon)=\widetilde{\Phi}^{-1}(\varepsilon(l\pi+\pi/2))+0(\varepsilon^{\infty}).$$ Particularly, $E^{(l)}(\varphi,\varepsilon)$ has the following Taylor expansion at first order in $\varepsilon$ : $$E^{(l)}(\varphi,\varepsilon)=E_{r}+W(0)+\sqrt{\frac{W"(0)}{2}}\frac{1}{d_{i}(E_{r})}(2l+1)\varepsilon+o(\varepsilon),$$ where $d_{i}(E_{r})$ is defined by \eqref{rnp}. \end{cor} \begin{dem} The first equality is obvious. It suffices to give an expansion of $\widetilde{\Phi}^{-1}(\varepsilon(l\pi+\pi/2))$. To do that, we compute an expansion at first order of: $$\Phi(E_{r}+W(0)+\alpha)=\int_{\varphi_{r}^{-}(E_{r}+W_{-}+\alpha)}^{\varphi_{r}^{+}(E_{r}+W_{-}+\alpha)}k(E_{r}+W(0)+\alpha-W(u))du.$$ The mapping $W$ is a bijection from $[0,\varphi_{r}^{+}]$ to $[W(0),E_{r}]$. By the substitution $\alpha v=W(0)+\alpha-W(u)$, we get that: $$\int_{0}^{\varphi_{r}^{+}(E_{r}+W(0)+\alpha)}=\alpha\int_{0}^{1}\frac{k(E_{r}+\alpha v)}{W'\circ W^{-1}(W(0)+\alpha(1-v))}dv.$$ But, $\lim\limits_{\alpha\rightarrow 0}\frac{k(E_{r}+\alpha v)}{W'\circ W^{-1}(W_{-}+\alpha(1-v))}=\frac{d_{i}(E_{r})}{\sqrt{2 W"(0)}}\frac{\sqrt{v}}{\sqrt{1-v}}$.\\ Similarly, on $[\varphi_{r}^{-},0]$, we have: $$\Phi(E_{r}+W(0)+\alpha)=d_{i}(E_{r})\frac{\pi}{2}\sqrt{\frac{2}{W"(0)}}\alpha[1+o(1)].$$ Consequently, by inverting the expansion of $\widetilde{\Phi}$ in the neighborhood of $E_{r}+W(0)$, we prove the result. \end{dem}\\ We point out that, as in \ref{trform2}, a more accurate asymptotic expansion of $\widetilde{\Phi}$ would give a better result on the eigenvalues. \bibliographystyle{plain}
{ "timestamp": "2005-03-12T19:58:38", "yymm": "0503", "arxiv_id": "math-ph/0503031", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503031" }
\section{Introduction} As is well known, \cite{UhlOrn}, \cite{Chandra}, \cite{Bal}, the ensemble average of the stochastic evolutions in {velocity space} of a Brownian test particle\footnote{For the beginnings of the theory of Brownian motion, see the collection of Einstein's papers with commentary \cite{Einstein}.} of unit mass, immersed in a drifting uniform heat bath of fixed temperature $T$ and constant drift velocity $\uV$, is governed by the Fokker--Planck equation with prescribed constant coefficients of diffusion and (linear) friction, \begin{equation} \partial_t f (\vV;t) = \partial_{\vV}\cdot\Big({T}\partial_{\vV}f(\vV;t) + \big(\vV - \uV\big) f(\vV;t)\Big). \label{FPbrown} \end{equation} Here, $f(\, .\, ;t):\mathbb{R}^3\to\mathbb{R}_+$ is the ensemble's probability density function on velocity space at time $t\in \mathbb{R}_+$, and an overall constant has been absorbed in the time scale. Of course, we could also shift $\vV$ to obtain $\uV=\mathbf{0}$, then rescale $\vV$, $t$, and $f$ to obtain $T = 1$; however, for pedagogical purposes we refrain from doing so. The solution $f (\vV;t)$ of \refeq{FPbrown} is given by $f (\vV;t) = \int_{\mathbb{R}^3}G_t(\wV,\vV|\uV;T)f_0(\wV)\mathrm{d}^3\wV$, where $f_0(\vV)\equiv f (\vV;0)$ and \begin{equation} G_t(\wV,\vV|\uV;T) = \left({2\pi T(1-e^{-2t})}\right)^{-\frac{3}{2}} \exp\left( -\frac{1}{2 T}\frac{|\vV-\uV -\wV e^{-t}|^2}{1-e^{-2t}} \right) \label{OUkernel} \end{equation} is the Green function for \refeq{FPbrown}, see \cite{UhlOrn}, \cite{Chandra}. In its standard form, i.e. with $ T = 1$ and $\uV=\mathbf{0}$, \refeq{OUkernel} is known as the (Mehler) kernel of the adjoint Ornstein-Uhlenbeck semigroup (a.k.a. Fokker--Planck semigroup). Over the years, the Ornstein-Uhlenbeck semigroup and its adjoint have come to play an important role in several branches of probability theory \cite{Hsu} related, in some form, to Brownian motions. The fact that the explicitly known kernel \refeq{OUkernel} of the Fokker--Planck semigroup readily lends itself to analytical estimates has led to useful applications also outside the realm of probability theory. In particular, in recent years the Fokker--Planck semigroup has found applications in kinetic theory, the subfield of transport theory which is concerned with the approach to equilibrium and the response to driving external forces of individual continuum systems not in local thermal equilibrium; see, for instance, the review \cite{Vil}. However, the linear Fokker--Planck equation itself, \refeq{FPbrown}, usually is not thought of as a {kinetic} equation for the particle density function on velocity space of an \textit{individual, isolated} space-homogeneous system of particles in some compact domain, which perform a microscopic autonomous dynamics that may be deterministic or stochastic but should satisfy the usual conservation laws of mass (particle number), energy and, depending on the shape of the domain in physical space and its boundary conditions, also momentum and angular momentum. Evidently the very meaning of $f$ and the parameters $\uV$ and $T$ in \refeq{FPbrown} voids this interpretation. Yet, with a re-interpretation of $f$, $\uV$ and $T$ it \emph{is} possible to assign to \refeq{FPbrown} a kinetic meaning. Incidentally, the first result showing that at least a partial re-interpretation of \refeq{FPbrown} in this direction is possible can be found in a paper by Villani \cite{Vil98} who, in his study of the space-homogeneous Landau equation for the weak deflection (i.e. Landau) limit of a gas of particles with Maxwellian molecular interactions, discovered that for isotropic velocity distribution functions $f$ (and only for these) the Landau equation is identical to \refeq{FPbrown}, with parameters $\uV=\vect{0}$ and $T$ matched to guarantee energy conservation. For general non-isotropic data the Landau equation for Maxwell molecules is identical to a more complicated equation than \refeq{FPbrown}. To pave the ground for a complete re-interpretation of \refeq{FPbrown}, which requires re-assigning the meaning of $f$, $\uV$ and $T$, we first note that by the linearity of \refeq{FPbrown} we can scale $f$ to any positive normalization we want. We now introduce the following functionals of~$f$, \noindent the ``mass of $f$'' \begin{equation} m(f) = \int_{\mathbb{R}^3} f(\vV;t)\mathrm{d}^3\vV\,, \label{mOFf} \end{equation} the ``momentum of $f$'' \begin{equation} \pV(f) = \int_{\mathbb{R}^3} \vV f(\vV;t)\mathrm{d}^3\vV\,, \label{pOFf} \end{equation} and the ``energy of $f$'' \begin{equation} e(f) = \int_{\mathbb{R}^3} \frac{1}{2}|\vV|^2 f(\vV;t)\mathrm{d}^3\vV\,. \label{eOFf} \end{equation} The ``angular momentum of $f$'' for a space-homogeneous $f(\vV;t)$ is simply $\jV(f) = \xV_{\mathrm{CM}}\times\pV(f)$, with $\xV_{\mathrm{CM}}$ the center of mass of the system, but this does not add any further insight and hence will not be considered explicitly. The functionals \refeq{mOFf}, \refeq{pOFf}, and \refeq{eOFf} inherit some time dependence from the solution $f(\,.\,;t)$ of \refeq{FPbrown}, but to find this dependence explicitly it is not necessary to solve for $f$ first. Indeed, it is an elementary exercise in integration by parts to extract from \refeq{FPbrown} the following linear evolution equations with constant coefficients for $m$, $\pV$, and $e$, \begin{equation} \dot{m} = 0\,, \label{mDOT} \end{equation} \begin{equation} \dot{\pV} = m\uV - \pV\,, \label{pDOT} \end{equation} \begin{equation} \dot{e} = 3 T - 2e + \uV\cdot\pV\,, \label{eDOT} \end{equation} \newpage \noindent which, beside the conservation of mass, i.e. $m(f) = m(f_0)$, describe the exponentially fast convergence to a stationary state $\pV(f) \leadsto m(f_0)\uV$ and $e(f)\leadsto \frac{3}{2} T+ \frac{1}{2}m(f_0)|\uV|^2$. While all this is of course quite trivial and well known, the relevant fact to realize here is that whenever the energy and the momentum of the initial $f_0$ equal these asymptotically stationary values, viz. if $\pV(f_0) = m(f_0)\uV$ and $e(f_0) = \frac{3}{2} T+ \frac{1}{2}m(f_0)|\uV|^2$, then beside mass $m$, also energy $e$ and momentum $\pV$ will be conserved. Conservation of mass, energy, and momentum for such a large subset of initial data $f_0$ does not yet mean that we may already think of the linear equation \refeq{FPbrown} as a kinetic equation, which should conserve mass, energy, and (depending on the shape of the domain in physical space and its boundary conditions) also momentum for \emph{all} initial data, no matter what their mass, energy and momentum are; moreover, a genuine kinetic equation for particles with (pair or higher order) interactions must express the time derivative of $f$ in terms of an at least\footnote{The Boltzmann, the Landau, and the Vlasov kinetic equations have bilinear ``interaction operators,'' the Balescu--Lenard--Guernsey equation has a higher order nonlinearity which reduces to the bilinear format in the long wavelength regime.} bilinear operator in $f$. However, with the help of \refeq{mOFf}, \refeq{pOFf} and \refeq{eOFf} we now replace $T$ and $\uV$ in \refeq{FPbrown} to obtain just such a kinetic equation. Indeed, consider the {\emph{a priori}} nonlinear Fokker--Planck equation \begin{equation} \partial_t f (\vV;t) = \partial_{\vV}\cdot\Big(\frac{1}{3}\big(2e(f)m(f)- |\pV(f)|^2\big) \partial_{\vV}f(\vV;t) + \big(m(f)\vV - \pV(f)\big) f(\vV;t)\Big), \label{FPkin} \end{equation} where $f(\, .\, ;t):\mathbb{R}^3\to\mathbb{R}_+$ now is a particle density function on velocity space at time $t\in \mathbb{R}_+$. The right-hand side of \refeq{FPkin} is a sum of a bilinear and a trilinear operator acting on $f$ which now guarantees conservation of mass, momentum, and energy for \emph{all} initial data $f_0\geq 0$, as verified by repeating the easy exercise in elementary integrations by parts using \refeq{FPkin} to find $\dot{m} = 0$ as well as $\dot{\pV} = m\pV - \pV m = \mathbf{0}$ and $\dot{e} = 2em - |\pV|^2 m - 2em + |\pV|^2m = 0$. Of course, \emph{after this fact} of mass, momentum, and energy conservations the {\emph{a priori}} nonlinear equation \refeq{FPkin} in effect becomes just a completely and explicitly solvable linear\footnote{In this sense \refeq{FPkin} is ``almost nonlinear,'' or ``essentially linear,'' depending on one's viewpoint.} Fokker--Planck equation \refeq{FPbrown}, only now with parameters $\uV$ and $T$ which are not prescribed but determined through the initial data $f_0$, viz. $\uV = \pV(f_0)/m(f_0)\equiv \uV_0$ and $\frac{3}{2} T = e(f_0) - |\pV(f_0)|^2/2m(f_0)\equiv \varepsilon_0$; we also set $m(f_0)\equiv m_0$ and $e(f_0) = e_0$. Accordingly, \refeq{FPkin} inherits from \refeq{FPbrown} the feature that, as $t\to\infty$, its solutions $f$ converge pointwise exponentially fast to the Maxwellian equilibrium state \begin{equation} f_{\mathrm{M}}(\vV) = m_0 \left(\frac{3}{4\pi\varepsilon_0}\right)^{\frac{3}{2}} \exp\left( -\frac{3|\vV -\uV_0|^2}{4\varepsilon_0} \right), \end{equation} with monotonically increasing relative entropy \begin{equation} S(f|f_{\mathrm{M}}) = - \int_{\mathbb{R}^3} f(\vV;t)\ln \frac{f(\vV;t)}{f_{\mathrm{M}}(\vV)}\mathrm{d}^3\vV \end{equation} which in fact approaches its maximum value $0$ exponentially fast. Since (\ref{FPkin}) displays all the familiar features of a kinetic equation (formal nonlinearity; conservation laws of mass, momentum, energy; an $H$-Theorem; approach to equilibrium; Maxwellian equilibrium states), at this point we may legitimately contemplate \refeq{FPkin} as a kinetic equation of some spatially homogeneous, isolated system of $N$ interacting particles in a compact spatial domain compatible with momentum conservation (e.g. a rectangle with periodic boundary conditions). In the remainder of this paper we show explicitly how \refeq{FPkin} arises from the Kolmogorov equation\footnote{In the physics literature, the Kolmogorov equation for an $N$-particle Markov process is traditionally called ``master equation".} for the adjoint evolution of an underlying $N$-particle Markov process in the limit $N\to\infty$. We use the strategy originally introduced by Kac \cite{Kac} in 1956 in the context of his work on a caricature of the Boltzmann equation; for important recent work on Kac's original program, see \cite{CarLoss}. As Kac realized, the crucial property that needs to be established in order to validate the $N\to\infty$ limit is what he called ``propagation of chaos," which loosely speaking means that if the particle velocities are uncorrelated at $t=0$, they remain uncorrelated at later times; this can be rigorously true only on the continuum scale in the limit $N\to\infty$. Interestingly enough, by adding some suitable lower order terms to the putatively simplest $N$-particle Markov process that leads to the (kinetic) Fokker--Planck equation in the limit $N\to\infty$, the corresponding Kolmogorov equation for an ensemble of such isolated $N$-particle systems can be simplified to be just the diffusion equation on the $3N-4$-dimensional manifold (a sphere) of constant energy and momentum. Since therefore both the finite-$N$ and the infinite-$N$ equations are exactly solvable, the kinetic limit $N\to\infty$ can be carried out explicitly and studied in great detail. For this reason we actually defer the discussion of the underlying $N$-particle process to Appendix Ab while in the main part of our paper we analyse the diffusion equation on $\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}$ and derive from it the kinetic Fokker--Planck equation on $\mathbb{R}^3$. Technically, we apply the Laplace--Beltrami operator to a probability density on $\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}$ and then integrate out $N-n$ velocities over their constrained domain of accessibility. Taking next the limit $N\to\infty$ yields a Fokker--Planck operator acting on the $n$-th marginal density on $\mathbb{R}^{3n}$. Thus we obtain a linear Fokker--Planck hierarchy of equations indexed by $n$. Using the Hewitt--Savage decomposition theorem, the hierarchy is seen to be generated by the single, {\emph{a priori}} nonlinear kinetic Fokker--Planck equation \refeq{FPkin} which in view of the conservation laws is equivalent to the essentially linear Fokker--Planck equation \refeq{FPbrown} with constant parameters which are determined by the initial data. Experts in probability theory may have noticed a similarity between the first part of our program and what has been called the ``Poincar\'e limit'' \cite{Bak}; in fact, our approach is ``dual'' to Bakry's approach. More specifically, Bakry \cite{Bak} has shown that the action of the Laplace--Beltrami operator for $\mathbb{S}^N_{\sqrt{N}}\hookrightarrow \mathbb{R}^{N+1}$ on a probability density function over a ``radial'' coordinate axis of $\mathbb{S}^N_{\sqrt{N}}$ becomes identical, in the limit $N\to\infty$, to the action of the Ornstein--Uhlenbeck operator on the same density viewed as a function over $\mathbb{R}$. Obviously, whenever the ``radial" function is obtained by taking the marginal of a probability density over $\mathbb{S}^N_{\sqrt{N}}$, i.e. by integrating out the $N-1$ Cartesian coordinates of the embedding space which are perpendicular to a fixed ``radial'' direction, the Ornstein--Uhlenbeck operator acts on the limiting marginal density as $N\to\infty$. This relationship between the operators is reflected at the spectral level by the convergence of the whole structure of orthogonal eigenfunctions of the Laplacian on $\mathbb{S}^N_{\sqrt{N}}$ (hyper-spherical harmonics) to the orthogonal eigenfunctions of the Ornstein--Uhlenbeck operator on $\mathbb{R}$ (Hermite polynomials multiplied by the square root of their Gaussian weight function); one of the earliest works is \cite{Mehler}, while more recent works on the Poincar\'e limit, containing interesting connections with the theory of Markov semigroups, are \cite{Bak} and \cite{BakMaz}. Our procedure is ``dual'' to Bakry's approach in the sense that we integrate out subsets of the Cartesian variables of the embedding space \emph{after} having applied the Laplace--Beltrami operator to a probability density on the high-dimensional sphere, thereby obtaining the \emph{adjoint} Ornstein--Uhlenbeck operator acting on the respective marginals; in addition, while Bakry considers only mass and energy conservation, we consider conservation of mass, energy, and momentum. Incidentally, our work is not inspired by Bakry's works on the Poincar\'e limit, nor by Villani's discovery about the isotropic evolution of the space-homogeneous Landau equation, about both of which we learned only after our own findings. Rather, our study of the diffusion equations on the $3N-C$-dimensional spheres of constant energy ($C=1$), respectively energy and momentum ($C=4$), which began in \cite{KieLan04}, was originally conceived of as a \emph{technically simpler primer} for our investigation (also in \cite{KieLan04}) of the Balescu--Prigogine master equation for Landau's kinetic equation. And while the present paper is also a technical continuation of \cite{KieLan04}, in the sense that here we supply various calculations that we had announced in \cite{KieLan04}, the main purpose of the present paper is to amplify the conceptual spin-off of our technical investigations, the \emph{new physical interpretation} of one of the simplest and best known linear transport equations as an (almost nonlinear) kinetic equation. As should be clear from our discussion in this introduction, this kinetic theory interpretation of the prototype Fokker--Planck equation may have been suspected by others long ago, yet we have not been able to find the whole story in the literature. In what follows, for the sake of simplicity we set $m_0 =1$, and accordingly\footnote{Setting $m_0=1$ means we should now speak of the energy per particle $e_0$, the thermal energy per particle $\varepsilon_0$, and the momentum per particle $\pV_0 (=\uV_0)$.} obtain $\pV(f_0) \equiv \uV_0$ and $e(f_0) - |\pV(f_0)|^2/2 = e_0 - |\uV_0|^2/2 \equiv \varepsilon_0$. With these simplifications \refeq{FPkin} now becomes \begin{equation} \partial_t f (\vV;t) = \partial_{\vV}\cdot\Big(\frac{2}{3}\varepsilon_0 \partial_{\vV}f(\vV;t) + \big(\vV - \uV_0\big) f(\vV;t)\Big). \label{FPkinSIMPLE} \end{equation} While \refeq{FPkinSIMPLE} is essentially a linear PDE, it should just be kept in mind that $\varepsilon_0$ and $\uV_0$ are functionals of $f$ which are determined by the initial data $f_0$ and not chosen independently.\footnote{The identification of \refeq{FPkin} with \refeq{FPkinSIMPLE} is valid only for isolated systems that can freely translate. If a driving external force field $\mathbf{F}$ is applied, then $e(f)$ and $\pV(f)$ are no longer constant and \refeq{FPkin} -- with the addition of the forcing term $-\mathbf{F}\cdot\partial_\vV f$ to its r.h.s. -- is the relevant equation.} We next shall derive \refeq{FPkinSIMPLE} from the diffusion equation equation on $\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}$ in the spirit of Kac's program. \section{The Finite-$N$ Ensembles} Consider an infinite ensemble of i.i.d. random vectors $\{\vVN_\alpha\}_{\alpha =1}^\infty$ where each $\vVN =(\vV_1,...,\vV_N)\in \mathbb{R}^{3N}$ represents a possible micro-state of an individual system of $N$ particles with velocities $\vV_i=(v_{i1},v_{i2},v_{i3})\in\mathbb{R}^3$ and particle positions assumed to be uniformly distributed over a periodic box; hence, particle positions will not be considered explicitly. Each $\vVN$ takes values in the $3N-4$-dimensional manifold of constant energy $e_0$ and momentum $\uV_0$, \begin{equation} \mathbb{M}^{3N-4}_{\uV_0,e_0} = \Big\{\vVN\; :\;\sum_{k=1}^N \vV_k=N\mathbf{u}_0, \; \sum_{k=1}^N\frac{1}{2} \abs{\vV_k}^2=Ne_0, \; e_0 > \frac{1}{2}|\uV_0|^2 \Big\}. \end{equation} The manifold $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ is identical to a $3N-4$-dimensional sphere of radius $\sqrt{2N\varepsilon_0}$ (where $\varepsilon_0$ appears above \refeq{FPkinSIMPLE}), centered at $\uVN = (\uV_0,...,\uV_0)$ and embedded in the $3(N-1)$-dimensional affine linear subspace of $\mathbb{R}^{3N}$ given by $\uVN + \mathbb{L}^{3N-3}$, where $\mathbb{L}^{3N-3} \equiv \mathbb{R}^{3N}\cap\big\{\vVN\in\mathbb{R}^{3N}: \sum_{k=1}^N \vV_k= \mathbf{0}\big\}$ is the space of velocities in any center-of-mass frame. The ensemble at time $\tau$ is characterized by a probability density $F^{(N)}(\vVN;\tau)$ on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$, the evolution of which is determined by the diffusion equation \begin{equation} \partial_\tau F^{(N)}(\vVN;\tau) = \Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}\,F^{(N)}(\vVN;\tau), \label{heat} \end{equation} where $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ is the Laplace--Beltrami operator on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$. Since all particles are of the same kind, we consider only solutions to (\ref{heat}) which are invariant under the symmetric group $S_N$ applied to the $N$ components in $\mathbb{R}^3$ of $\vVN$. Clearly, permutation symmetry is preserved by the evolution.\footnote{In what follows, for the sake of notational simplicity we will not enforce this symmetry explicitly, but the reader should be aware that (for instance) all the eigenfunctions that appear below in the solution for $F^{(N)}$ can be easily symmetrized.} We will show that the diffusion equation (\ref{heat}), here viewed as a master equation, leads precisely to the essentially linear Fokker--Planck equation \refeq{FPkinSIMPLE} in the sense of Kac's program: (a) the Fokker--Planck equation \refeq{FPkinSIMPLE} arises as the $N\to\infty$ limit of the equation for the first marginal of $F^{(N)}(\vVN;\tau)$ derived from (\ref{heat}), and (b) propagation of chaos holds. In this section we prepare the ground by discussing the finite-$N$ equation \refeq{heat}. The limit $N\to\infty$ is carried out in the next section, while propagation of chaos is discussed in the final section. For the sake of completeness, we begin by listing some general facts about the diffusion equation. We note that the Laplacian $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ is a positive semi-definite, essentially self-adjoint operator on the dense domain $\mathfrak{C}^\infty (\mathbb{M}^{3N-4}_{\uV_0,e_0}) \subset\mathfrak{L}^2 (\mathbb{M}^{3N-4}_{\uV_0,e_0})$, thus it has a unique self-adjoint extension with domain $\mathfrak{H}^2(\mathbb{M}^{3N-4}_{\uV_0,e_0})$. Its self-adjoint extension is the generator of a non-expansive semigroup on $\mathfrak{L}^2 (\mathbb{M}^{3N-4}_{\uV_0,e_0})$ which is strictly contracting on the $\mathfrak{L}^2$ orthogonal complement of the constant functions. Thus, we may ask that the initial condition $\lim_{t\downarrow 0}F^{(N)}(\,.\,;\tau) = F_0^{(N)}(\,.\,)\in\mathfrak{L}^2 (\mathbb{M}^{3N-4}_{\uV_0,e_0})$ (which implies $F_0^{(N)}\in \mathfrak{L}^1 (\mathbb{M}^{3N-4}_{\uV_0,e_0})$). Yet, as is well-known, the diffusion semigroup is so strongly regularizing that we may even take $F_0^{(N)}(\,.\,)\in\mathfrak{M}_{+,1} (\mathbb{M}^{3N-4}_{\uV_0,e_0})$, a probability measure, and obtain $F^{(N)}(\,.\,;\tau) \in \mathfrak{C}^\infty(\mathbb{M}^{3N-4}_{\uV_0,e_0})$ for all $\tau>0$. In fact, the solutions of \refeq{heat} can be computed quite explicitly in terms of an eigenfunction expansion. Since via translation by $\uVN$ (choosing a center-of-mass frame) and scaling by $\sqrt{2N\varepsilon_0}$ (choosing a convenient unit of energy) the manifold $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ can be identified with the unit sphere centered at the origin of the linear subspace $\mathbb{L}^{3N-3}\subset\mathbb{R}^{3N}$, the complete spectrum of $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ and an orthogonal basis of eigenfunctions can be obtained from the well-known eigenvalues and eigenfunctions for the Laplacian on the unit sphere $\mathbb{S}^{3N-4}\hookrightarrow\mathbb{R}^{3N-3}$. Of course, in our case the embedding is $\mathbb{S}^{3N-4}\hookrightarrow\mathbb{L}^{3N-3}$ with $\mathbb{L}^{3N-3}$ isomorphic by a rotation to standard $\mathbb{R}^{3N-3}$. Thus we start from $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ and we first carry out a rotation in $\mathbb{R}^{3N}$ that transforms $\vVN$ to $\wVN=\mathcal{U}\vVN$ in such a way that $\mathbb{L}^{3N-3}$ is mapped to the $3N-3$-dimensional linear subspace $\big\{\wVN\ :\ \wV_N=\mathbf{0}\big\}$. Obviously, $\mathcal{U}^T$ is the linear transformation that diagonalizes the projection operator onto $\mathbb{L}^{3N-3}$. A complete orthonormal set of eigenvectors for such a projection is readily calculated and leads to \begin{eqnarray} \wV_1 &=& \sqrt{\frac{N-1}{N}}\left[\vV_1-\frac{1}{N-1}\sum_{i=2}^N\vV_i\right] \nonumber\\ &\vdots&\nonumber\\ \wV_n&=&\sqrt{\frac{N-n}{N-n+1}}\left[\vV_n-\frac{1}{N-n}\sum_{i=n+1}^N\vV_i \right]\nonumber\\ &\vdots&\nonumber\\ \wV_{N-1}\!\!\!\!\!\!&=&\frac{1}{\sqrt{2}}[\vV_{N-1}-\vV_N]\nonumber\\ \wV_N&=&\frac{1}{\sqrt{N}}\sum_{i=1}^N\vV_i \label{rotation} \end{eqnarray} It is easily checked that the matrix associated with this transformation is indeed orthogonal, and that $\wV_N$ vanishes whenever $\vVN\in\mathbb{L}^{3N-3}$. More generally, the affine subspace $\uVN+\mathbb{L}^{3N-3}$ is mapped to the linear manifold $\big\{\wVN\ :\ \wV_N=\sqrt{N}\uV_0\big\}$ and $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ is mapped to \begin{equation} \bigg\{\wVN\; :\; \wV_N=\sqrt{N}\uV_0,\quad \sum_{i=1}^{N-1}\abs{\wV_i}^2=2Ne_0-N\abs{\uV_0}^2=2N\varepsilon_0 \bigg\} \end{equation} which implies that the truncated vector $(\wV_1,\dots,\wV_{N-1})$ belongs to the sphere $\mathbb{S}_{\sqrt{2N\varepsilon_0}}^{3N-4}\hookrightarrow\mathbb{R}^{3N-3}$ (in $\wV_k$-coordinates). Thus, the transform $\mathcal{U}$ allows one to analyse the $N$-particle system with energy and momentum conservation (``periodic box" setup) in terms of an $(N-1)$-particle system with only energy conservation (a ``container with reflecting walls" setup).\footnote{The gas in such a container was discussed in our earlier work \cite{KieLan04}, but without detailed calculations. Our calculations with the $\wV$ variables here now supply the relevant details.} For future reference, we also observe that for $n$ fixed and $N\to\infty$ the effect of $\mathcal{U}$ reduces to a translation of each of the $n$ velocities by $\uV_0$, in the following sense. Consider a consistent hierarchy of vectors of increasing size $N$, in which lower-$N$ vectors can be obtained from the higher-N ones by truncation (i.e. projection). Suppose that the vectors belong to $\uVN+\mathbb{L}^{3N-3}$ for all $N$, apply the transformation in (\ref{rotation}) and look at the $n$-th component. Since $\sum_{i=n+1}^N\vV_i=N\uV_0-\sum_{i=1}^n\vV_i$, where $\sum_{i=1}^n\vV_i$ is independent of $N$, we find \begin{equation} \lim_{N\to\infty}\wV_n=\vV_n-\uV_0. \label{wntovn} \end{equation} We now recall that the Laplacian is invariant under Euclidean transformations. Thus, under our orthogonal transformation $\mathcal{U}$, the Laplacian $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ becomes the Laplacian on $\mathbb{S}_{\sqrt{2N\varepsilon_0}}^{3N-4}$ in $\mathbb{R}^{3N-3}$, the space of truncated vectors $(\wV_1,\dots,\wV_{N-1})$ (which will also be denoted by $\wVN$, at the price of abusing the notation). Since $\Delta_{\mathbb{S}_{\sqrt{2N\varepsilon_0}}^{3N-4}}= \frac{1}{2N\varepsilon_0}\Delta_{\mathbb{S}^{3N-4}}$, and the Laplacian on the unit sphere $\mathbb{S}^{3N-4}$ has spectrum $j(j + 3N -5)$, $j=0,1,\dots$, the spectrum of $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ is \begin{equation} \lambda^{(j)}_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}= \frac{j(j + 3N -5)}{2N\varepsilon_0},\qquad j=0,1,\dots\ . \medskip \label{eigenvaluesSN} \end{equation} The eigenspace on $\mathbb{S}^{3N-4}$ for the $j$-th eigenvalue has dimension \begin{equation} \mathcal{N}(j,3N-3)=\frac{(3N-5+2j)(3N-6+j)!}{j!(3N-5)!} \end{equation} and is spanned by an orthogonal basis of hyper-spherical harmonics\footnote{The hyper-spherical harmonics on $\mathbb{S}^n$ are restrictions to $\mathbb{S}^n\subset \mathbb{R}^{n+1}$ of homogeneous harmonic polynomials in $\mathbb{R}^{n+1}$. For $j>0$ the restriction has to be non-constant, since $\widetilde{Y}_{0,1}\equiv\ const.$.} on $\mathbb{S}^{3N-4}\subset \mathbb{R}^{3N-3}$ of order $j$, here denoted $\widetilde{Y}_{j,\ell}(\omV;3N-3)$, with $\ell\in\mathbb{D}_j=\{1,\dots,\mathcal{N}(j,3N-3)\}$ and with $\omV\in\mathbb{S}^{3N-4}$. The indexing of our $\widetilde{Y}_{j,\ell}(\omV;3N-3)$ follows the convention of \cite{Mul} for his $Y_{j,\ell}$ and differs from what might have been anticipated from the familiar convention for spherical harmonics on $\mathbb{S}^2$. Our reason for using tildes atop the function symbols is to remind the reader that we will use a normalization of the $\widetilde{Y}_{j,\ell}(\omV;3N-3)$ which conveniently suits our purposes and does not seem to agree with any of the existing conventions, such as in \cite{Mul} or for the spherical harmonics on $\mathbb{S}^2$. Our convention is motivated by the analysis of the large $N$ behavior of the eigenfunctions, carried out in Appendix B. Hence, the eigenspace of $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ associated with the $j$-th eigenvalue in (\ref{eigenvaluesSN}) is spanned by the eigenfunctions $\widetilde{Y}_{j,\ell}\left({\wVN}/{\sqrt{2N\varepsilon_0}};3N-3\right)$, $\ell\in\mathbb{D}_j$, where $\wVN$ is given by (\ref{rotation}) for $n=1,\dots,N-1$. To shorten the notation we introduce \begin{equation} G_{j,\ell}^{(N)}(\vVN) \equiv \abs{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{-1} \widetilde{Y}_{j,\ell}\left({\wVN}/{\sqrt{2N\varepsilon_0}}\,;3N-3\right); \label{eigenfunctions} \end{equation} here, the factor $\abs{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{-1}$ is introduced for later convenience. In terms of the eigenfunctions $G_{j,\ell}^{(N)}(\vVN)$, the solution to equation \refeq{heat} is simply given by the generalized Fourier series \begin{equation} F^{(N)}(\vVN;\tau) = \abs{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{-1} + \sum_{j\in\mathbb{N}} \sum_{\ell\in\mathbb{D}_j} F_{j,\ell}^{(N)} G_{j,\ell}^{(N)}(\vVN)\, e^{- \textstyle{\frac{j(j +3N -5)}{2N\varepsilon_0}}\tau} \label{FNevolution} \end{equation} with Fourier coefficients $F_{j,\ell}^{(N)}$ given by \begin{equation} F_{j,\ell}^{(N)} = \frac{\langle F^{(N)}_0|G_{j,\ell}^{(N)}\rangle} {\langle G^{(N)}_{j,\ell}|G_{j,\ell}^{(N)}\rangle} \label{FourierCOEFF} \end{equation} where $\langle\,.\,|\,.\,\rangle$ denotes the inner product in $\mathfrak{L}^2(\mathbb{M}^{3N-4}_{\uV_0,e_0})$. Notice, though, that the numerator $\langle F^{(N)}_0|G_{j,\ell}^{(N)}\rangle$ can be extended to mean the canonical pairing of the $G_{j,\ell}^{(N)}$s with an element of their dual space, which allows us to take $F^{(N)}_0$ to be a measure. In particular, we may take $F^{(N)}_0$ to be the Dirac measure concentrated at any particular point of $\mathbb{M}^{3N-4}_{\uV_0,e_0}$. The formula \refeq{FNevolution} then describes the fundamental solution of the diffusion equation \refeq{heat}. In any event, whatever $F^{(N)}_0$, \refeq{FNevolution} makes it evident that when $\tau\to\infty$ the ensemble probability density function on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ decays exponentially fast to the uniform probability density $\abs{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{-1}= \abs{\mathbb{S}_{\sqrt{2N\varepsilon_0}}^{3N-4}}^{-1}= F_{0,1}^{(N)} G_{0,1}^{(N)}(\vVN)$, which is the constant eigenfunction corresponding to the smallest non-degenerate eigenvalue $0$ of the Laplacian. \section{Evolution of the Marginals} To study the limit $N\to\infty$ for the time-evolution of the ensemble measure, we need to consider the hierarchy of $n$-velocity marginal distributions \begin{equation} F^{(n|N)}(\vV_1,\dots,\vV_n;\tau) \equiv \int_{\Omega^{3(N-n)-4}_{\mathbf{u}_0,e_0}} F^{(N)}(\vVN;\tau)\, d\vV_{n+1}\dots d\vV_N \end{equation} where $\Omega^{3(N-n)-4}_{\mathbf{u}_0,e_0}$ is given by all the $(\vV_{n+1},\dots ,\vV_N)$ such that \begin{equation} \sum_{i=n+1}^N\!\! \vV_k=N\uV_0-\sum_{i=1}^n\vV_k, \quad \sum_{i=n+1}^N \abs{\vV_k}^2=2Ne_0- \sum_{i=1}^n \abs{\vV_k}^2 \end{equation} and $F^{(n|N)}$ has domain $\{(\vV_1,\dots,\vV_n): \sum_{k=1}^n |\vV_k-\uV_0|^2\leq {4(N-n)\varepsilon_0}\}\subset\mathbb{R}^{3n}$. The evolution equation for the $n$-th marginal $F^{(n|N)}(\vV_1,\dots,\vV_n;\tau)$ is obtained by integrating \refeq{heat} over $(\vV_{n+1},\dots,\vV_N)\in \mathbb{R}^{3N-3n}$, using the representation of the Laplace--Beltrami operator given in (\ref{heat1}) of Appendix Aa. Then, a straightforward calculation (previously presented in \cite{KieLan04}) shows that $F^{(n|N)}$ satisfies \begin{eqnarray} \partial_\tau F^{(n|N)}\!\!&=&\!\! \sum_{i=1}^{n} \frac{\partial}{\partial\vV_i}\cdot\frac{\partial F^{(n|N)}}{\partial\vV_i}- \frac{1}{N}\sum_{k=1}^3\sum_{i,j=1}^{n} \frac{\partial^2 F^{(n|N)}}{\partial v_{ik}\partial v_{jk}} \nonumber\\ && -\frac{1}{2N\varepsilon_0} \sum_{i,j=1}^{n}\frac{\partial}{\partial\vV_i}\cdot \left((\vV_i-\uV_0)\,(\vV_j-\uV_0)\cdot \frac{\partial F^{(n|N)}}{\partial\vV_j}\right) \nonumber\\ && +\frac{3(N-n)}{2\varepsilon_0N}\sum_{i=1}^{n} \frac{\partial}{\partial\vV_i}\cdot\Big((\vV_i-\uV_0) F^{(n|N)}\Big). \label{nDIFFhierarchyEQ} \end{eqnarray} Clearly, to obtain the solutions of these equations it is advisable to integrate the series solution for $F^{(N)}(\vVN;\tau)$, (\ref{FNevolution}). For this purpose, it will be convenient to calculate the marginals in terms of the rotated variables $\wVN$. Changing the integration variables\footnote{Note that (\ref{rotation}) defines a one-to-one linear map with determinant $\sqrt{\frac{N}{N-n}}$ between $(\vV_{n+1},\dots\vV_N)$ and $(\wV_{n+1},\dots\wV_{N-1},\vect{z}_N)$, where $\vect{z}_N\equiv\wV_N-\frac{1}{\sqrt{N}}\sum_{i=1}^n\vV_i$.} gives \begin{equation} F^{(n|N)}(\vV_1,\dots,\vV_n;\tau) = {\textstyle{\sqrt{\frac{N}{N-n}}}} \int F^{(N)}(\vVN;\tau)\,d\wV_{n+1}\dots d\wV_{N-1} \label{marginW} \end{equation} where the integral is over $\mathbb{S}^{3(N-n)-4}_{\sqrt{2N\varepsilon_0-\sum_{i=1}^n\abs{\wV}_i^2}}$, and we abused the notation $F^{(N)}(\vVN;\tau)$ by applying it to what is now regarded as a function of $(\vV_1,...,\vV_n,\wV_{n+1},...,\wV_{N-1})$. To obtain the series solution for $F^{(n|N)}(\vVN;\tau)$ we need to express (\ref{FNevolution}) in the variables $(\vV_1,...,\vV_n,\wV_{n+1},...,\wV_{N-1})$ and then integrate term by term in the spirit of \refeq{marginW}. To accomplish this we need to choose explicitly a basis of spherical harmonics $\widetilde{Y}_{j,\ell}$ on $\mathbb{S}^{3N-4}$. It is convenient to do this in an iterative fashion, by assuming that a basis is known for the spherical harmonics with one independent variable less, here $\widetilde{Y}_{k,m}(\omV_{3N-5};3N-4)$ with $\omV_{3N-5}\in\mathbb{S}^{3N-5}$. Then, the desired basis is obtained \cite{Mul} by taking all the elements in the given lower-dimensional basis and multiplying them by associated Legendre functions of the ``extra" variable. In our case the $(3N-3$)-th variable will be $w_{11}/\sqrt{\scriptstyle{{2N\vareps_0}}}$, the first component of $\wVN/\sqrt{\scriptstyle{{2N\vareps_0}}}$, and $\omV_{3N-5}$ will be a unit vector in the space of the remaining $3N-4$ components, denoted by $(\wVN)_{3N-4}/\sqrt{\scriptstyle{{2N\vareps_0}}}$; thus, \begin{equation} \widetilde{Y}_{j,\ell}\left(\frac{\wVN}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-3\right) = \widetilde{Y}_{k,m}\left(\frac{(\wVN)_{3N-4}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-4\right) \, \widetilde{P}_j^k\left(\frac{w_{11}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-3\right) \label{basis1} \end{equation} where $k=0,1,\dots, j$, $m=1,\dots,\mathcal{N}(k,3N-4)$ and each choice of the pair $k,m$ is associated with a value of the degeneracy index $\ell$ for the basis $\widetilde{Y}_{j,\ell}$; moreover, $\widetilde{P}_j^k$ is an associated Legendre function \cite{Mul}, suitably normalized (see Appendix B). By repeating this process $3n$ times, we write out the eigenfunctions in the form \begin{eqnarray} &&\hskip-.9truecm \widetilde{Y}_{j,\ell}\left(\frac{\wVN}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-3\right) = \widetilde{Y}_{k_{3n},m}\left(\frac{(\wVN)_{3N-3n-3}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-3n-3\right) \times\phantom{spaaace} \\[0.4cm] &\times&\!\!\!\!\!\! \widetilde{P}_{k_{3n-1}}^{k_{3n}}\!\! \left(\frac{w_{n3}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-3n-2\right) \widetilde{P}_{k_1}^{k_2}\!\left(\frac{w_{12}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-4\right) \cdots \widetilde{P}_j^{k_1}\!\left(\frac{w_{11}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-3\right) \nonumber \label{basis2} \end{eqnarray} where $0\leq k_{3n}\leq \dots\leq k_1\leq j$ and $m=1,\dots,\mathcal{N}(k_{3n},3N-3n-3)$. Now let $g_{j,\ell}^{(n|N)}(\vV_1,\dots,\vV_n)$ denote the $n$-th ``marginal'' of $G_{j,\ell}^{(N)}(\vVN)$ (as for $F^{(N)}$ in (\ref{marginW})), and set $N^*\equiv N - n -1$. We find \begin{eqnarray} g_{j,\ell}^{(n|N)}(\vV_1,\dots,\vV_n) = \abs{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{-1} \int \widetilde{Y}_{k_{3n},m}\left(\frac{(\wVN)_{3N^*}} {\sqrt{\scriptstyle{{2N\vareps_0}}}};3N^*\right) d\wV_{n+1}\dots d\wV_{N-1} \nonumber\\[0.4cm] \times {\textstyle{\sqrt{\frac{N}{N-n}}}} \widetilde{P}_{k_{3n-1}}^{k_{3n}}\!\! \left(\frac{w_{n3}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N\!-\!3n\!-\!2\right)\!\!\! \cdots \widetilde{P}_j^{k_1}\!\left(\frac{w_{11}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N\!-\!3\right) \end{eqnarray} where the integral is over the same domain as in \refeq{marginW}. The integral of $\widetilde{Y}_{k_{3n},m}$ is non-zero if and only if $k_{3n}=0$ and $m=1$, and the integrals over $\widetilde{Y}_{0,1}$ are determined only up to the overall factor $\widetilde{Y}_{0,1}$, which we may choose to be unity without loss of generality. Accordingly, $g_{j,\ell}^{(n|N)}\equiv 0$ unless $\ell\in\widetilde\mathbb{D}_j\subset\mathbb{D}_j$, where $\widetilde\mathbb{D}_j$ contains the indices of the basis functions that ``descend'' from the uniform harmonic in $\mathbb{R}^{3N-3n-3}$. For such $\ell$'s the integrated eigenfunctions then become \begin{eqnarray} g_{j,\ell}^{(n|N)}(\vV_1,\dots,\vV_n) = \widetilde{P}_j^{k_1}\!\left(\frac{w_{11}}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N\!-\!3\right) \cdots \widetilde{P}_{k_{3n-1}}^{0}\!\!\left(\frac{w_{n3}}{\sqrt{\scriptstyle{{2N\vareps_0}}}}, 3N\!-\!3n\!-\!2\right) \nonumber\\[0.4cm] \times \sqrt{\frac{N}{N-n}}\, \frac{\abs{\mathbb{S}^{3(N-n)-4}}} {\abs{\mathbb{S}^{3N-4}}} \frac{1}{{\sqrt{\scriptstyle{{2N\vareps_0}}}}^{3n}}\, \Big(1-\frac{1}{\sqrt{\scriptstyle{{2N\vareps_0}}}}\sum_{i=1}^n|\wV_i|^2\Big)^{\frac{3(N-n)-4}{2}}. \label{eigmarg2} \end{eqnarray} The series for the $n$-th marginal $F^{(n|N)}(\,.\,;\tau)$ (the integrated \refeq{FNevolution}) is a series in the functions \refeq{eigmarg2}, viz. \begin{equation} F^{(n|N)}(\vV_1,\dots,\vV_n;\tau) = \sum_{j\in \mathbb{N}\cup\{0\}} \sum_{\ell\in\widetilde\mathbb{D}_j} F_{j,\ell}^{(N)} g_{j,\ell}^{(n|N)}(\vV_1,\dots,\vV_n)\, e^{- \textstyle{\frac{j(j +3N -5)}{2N\varepsilon_0}}\tau}. \label{FnNevolution} \end{equation} \section{The Limit $N\to\infty$} We are now ready to take the infinitely many particles limit. First of all, we observe that the evolution equation for the marginal velocity densities $f^{(n)}(\vV_1,\dots,\vV_n;\tau)\equiv \lim_{N\to\infty}F^{(n|N)}(\vV_1,\dots,\vV_n;\tau)$ which obtains in the formal limit $N\to\infty$ from \refeq{nDIFFhierarchyEQ} is the essentially linear Fokker--Planck equation in $\mathbb{R}^{3n}$, \begin{equation} \partial_\tau f^{(n)} = \sum_{i=1}^{n} \frac{\partial}{\partial\vV_i}\cdot\Big(\frac{\partial f^{(n)}} {\partial\vV_i}+\frac{3}{2\varepsilon_0}(\vV_i-\uV_0)\,f^{(n)}\Big). \label{nDIFFhierarchyEQlim} \end{equation} We now show that the series expansion for the time-evolved finite-$N$ marginals $F^{(n|N)}(\,.\,;\tau)$ converge under natural conditions to solutions of these equations. Beginning with the spectrum of $\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$, we note that the limit $N\to\infty$ yields \begin{equation} \lim_{N\to\infty} \Bigl\{ \lambda_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{(j)} \Bigr\}_{j=0}^\infty = \Big\{\textstyle{ \frac{3j}{2\varepsilon_0}} \Big\}_{j=0}^\infty. \label{LIMspectrumNEw} \end{equation} Thus, the limit spectrum is discrete. In particular, there is a spectral gap separating the origin from the rest of the spectrum. As a result, the time evolution of the limit $N\to\infty$ continues to approach a stationary state exponentially fast when $\tau\to\infty$. Coming to the eigenfunctions, the expression on the second line in (\ref{eigmarg2}) contains the $n$-velocity marginal distribution of the uniform density $\abs{\mathbb{M}^{3N-4}_{\uV_0,e_0}}^{-1}$ (the $j=0$ case). As is well-known at least since the time of Boltzmann, this distribution converges pointwise when $N\to\infty$ to the $n$-velocity drifting Maxwellian on $\mathbb{R}^{3n}$, \begin{equation} f_{\mathrm{M}}^{\otimes{n}}(\vV_1,...,\vV_n) = \left({\frac{3}{4\pi\varepsilon_0}}\right)^{\frac{3n}{2}} \prod_{i=1}^n \exp\left( -{\textstyle{\frac{3}{4\varepsilon_0}}}|\vV_i-\uV_0|^2 \right) \label{nMaxwellian} \end{equation} (recall (\ref{wntovn})). In terms of eigenfunctions this means that the ``projection'' onto $\mathbb{R}^{3n}$ of the $j=0$ eigenfunction of the Laplace--Beltrami operator on $\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}$ converges pointwise (in fact, even uniformly) to the $j=0$ eigenfunction of the linear Fokker--Planck operator in $\mathbb{R}^{3n}$, appearing in the r.h.s. of (\ref{nDIFFhierarchyEQlim}). The connection between the eigenfunctions generalizes to the cases $j\neq 0$; cf. \cite{BakMaz} for the special case $\uV_0=\vect{0}$. The asymptotic behavior for $N\to\infty$ of the associated Legendre functions in (\ref{eigmarg2}), which is discussed in Appendix B, together with (\ref{wntovn}), yields that $g_{j,\ell}^{(n)}(\vV_1,\dots,\vV_n) \equiv \lim_{N\to\infty}g_{j,\ell}^{(n|N)}(\vV_1,\dots,\vV_n)$ exists pointwise for all $(\vV_1,\dots,\vV_n)\in\mathbb{R}^{3n}$, with \setlength{\arraycolsep}{0.5mm} \begin{eqnarray} \!\!\!\!\!\! g_{j,\ell}^{(n)}(\vV_1,\dots,\vV_n) &=& \frac{(-1)^j}{2^{j/2}} H_{j-k_1}\!\left({\textstyle{\sqrt{\frac{3}{4\varepsilon_0}}}} (v_{11}\!-u_1)\!\right) \cdots H_{k_{3n-1}}\!\left({\textstyle{\sqrt{\frac{3}{4\varepsilon_0}}}} (v_{n3}\!-u_3) \!\right) \label{eigmarg3} \nonumber\\ &&\times \left({\frac{3}{4\pi\varepsilon_0}}\right)^{\frac{3n}{2}}\, \prod_{i=1}^n \exp\left( -{\textstyle{\frac{3}{4\varepsilon_0}}}|\vV_i-\uV_0|^2 \right) \nonumber\\ &\equiv&\frac{(-1)^j}{2^{j/2}}\! \left({\textstyle{\frac{3}{4\pi\varepsilon_0}}}\right)^{\!\!\frac{3n}{2}}\! \prod_{i=1}^n e^{-\frac{3}{4\varepsilon_0}|\vV_i-\uV_0|^2} \prod_{l=1}^3 H_{m_{i\cdot l}}\!\left({\textstyle{\sqrt{\frac{3}{4\varepsilon_0}}}} (v_{il}\!-u_l)\!\right) \label{nEIGlim} \end{eqnarray} for all $\ell\in\widetilde\mathbb{D}_j$, where $H_m(x)$ is the Hermite polynomial of degree $m$ on $\mathbb{R}$, and we defined $m_1=j-k_1,m_2=k_1-k_2,\dots, m_{3n}=k_{3n-1}$. In terms of the $m_i$'s, the index set $\widetilde\mathbb{D}_j$ counts all the choices of integers $0\leq m_1,\dots,m_{3n}\leq j$ such that $\sum_{i=1}^{3n}m_i=j$. For $n=1$ one readily recognizes the well-known eigenfunctions \cite{Risk} for the linear Fokker--Planck operator in $\mathbb{R}^3$, viz. r.h.s.(\ref{FPkinSIMPLE}) with constant $\varepsilon_0$ and $\uV_0$, easily calculated by separation of variables. In fact, what we have recovered are precisely the eigenfunctions for the linear Fokker--Planck operator in $\mathbb{R}^{3n}$, see (\ref{nDIFFhierarchyEQlim}). Now assume that one can choose sequences of initial conditions $F^{(N)}_0$ such that, for each fixed $j$ and $\ell$, the Fourier coefficients $F_{j,\ell}^{(N)}$ converge to a limit $F_{j,\ell}$ \emph{such that} each initial $n$-velocity marginal density, $n\in\mathbb{N}$, converges in $(\mathfrak{L}^2\cap\mathfrak{L}^1)(\mathbb{R}^{3n})$ to \begin{equation} f^{(n)}(\vV_1,\dots,\vV_n;0) = f_{\mathrm{M}}^{\otimes{n}}(\vV_1,...,\vV_n) + \sum_{j\in\mathbb{N}} \sum_{\ell\in\widetilde\mathbb{D}_j} F_{j,\ell} g_{j,\ell}^{(n)}(\vV_1,\dots,\vV_n); \label{fnNULL} \end{equation} it then follows that the subsequent evolution of the $n$-velocity marginal densities is given by \begin{equation} f^{(n)}(\vV_1,\dots,\vV_n;\tau) = f_{\mathrm{M}}^{\otimes{n}}(\vV_1,...,\vV_n) + \sum_{j\in\mathbb{N}} \sum_{\ell\in\widetilde\mathbb{D}_j} F_{j,\ell} g_{j,\ell}^{(n)}(\vV_1,\dots,\vV_n) e^{- \textstyle{\frac{3j}{2\varepsilon_0}}\tau}. \label{fnSOL} \end{equation} Formula \refeq{fnSOL} describes an exponentially fast approach to equilibrium in the ensemble of infinite systems. The $f^{(n)}(\,.\,;\tau)\in(\mathfrak{L}^2\cap\mathfrak{L}^1)(\mathbb{R}^{3n})$, and in addition they automatically satisfy \begin{eqnarray} \int_{\mathbb{R}^{3n}} f^{(n)}(\vV_1,\dots,\vV_n;\tau) \, d\vV_1\dots d\vV_n \!&=&\! 1 \label{fnMASS} \\ \int_{\mathbb{R}^{3n}}(\vV_1+\dots+\vV_n) f^{(n)}(\vV_1,\dots,\vV_n;\tau) \, d\vV_1\dots d\vV_n \!&=&\! n\uV_0 \label{fnMOMENTUM} \\ \int_{\mathbb{R}^{3n}} \frac12 (|\vV_1|^2+\dots +|\vV_n|^2) f^{(n)}(\vV_1,\dots,\vV_n;\tau)\, d\vV_1\dots d\vV_n \!&=&\! n e_0 \label{fnENERGY} \end{eqnarray} for all $\tau\geq 0$ (recall that $e_0 = \varepsilon_0 + |\uV|_0^2/2$). In fact, \refeq{fnSOL} solves \refeq{nDIFFhierarchyEQlim}, which now implies that $f^{(n)}(\,.\,;\tau)$ can also be expressed through integration of the initial data against the $n$-fold tensor product of \refeq{OUkernel}. The upshot is that $f^{(n)}(\,.\,;\tau)\in \mathfrak{S}(\mathbb{R}^{3n})\ \forall\tau>0$ (Schwartz space). To vindicate these conclusions, for us it remains to show that the infinitely many constraints on each $F_{j,\ell}$ implied by \refeq{fnNULL}, viz. \begin{equation} F_{j,\ell} = \frac{\langle f^{(n)}_0|g_{j,\ell}^{(n)}\rangle} {\langle g^{(n)}_{j,\ell}|g_{j,\ell}^{(n)}\rangle}\qquad \forall n\in\mathbb{N}, \label{FjlCONSTRAINTS} \end{equation} where $\langle\,.\,|\,.\,\rangle$ now means inner product in $\mathfrak{L}^2(\mathbb{R}^{3n})$, do not impose impossible consistency requirements. To show this, recall that the $f^{(n)}_0$ by definition satisfy \begin{equation} \int_{\mathbb{R}^3} f_{0}^{(n+1)}(\vV_1,\dots,\vV_{n+1}) d\vV_{n+1} = f_{0}^{(n)}(\vV_1,\dots,\vV_{n}), \label{fMARG} \end{equation} which in view of \refeq{fnNULL} implies that the hierarchy of the $g_{j,\ell}^{(n)}$ must satisfy \begin{equation} \int_{\mathbb{R}^3} g_{j,\ell}^{(n+1)}(\vV_1,\dots,\vV_{n+1}) d\vV_{n+1} = g_{j,\ell}^{(n)}(\vV_1,\dots,\vV_{n})\prod_{i=1}^3 \delta_{k_{3(n+1)-i},0}, \label{gMARG} \end{equation} which is readily verified by explicit integration of \refeq{nEIGlim}. Thus, the constraints \refeq{FjlCONSTRAINTS} are automatically consistent, and this vindicates our initial assumption. \section{Propagation of Chaos} Setting $n=1$ in \refeq{nDIFFhierarchyEQlim}, and changing the time scale by setting $\tau = \frac{2}{3}\varepsilon_0t$, we recover (\ref{FPkinSIMPLE}), with $f^{(1)}$ in place of $f$. However, (\ref{FPkinSIMPLE}) (or \refeq{FPkin} for that matter) cannot be said to have been shown to be a kinetic equation yet. Note that propagation of chaos has not entered the derivation of \refeq{nDIFFhierarchyEQlim}. In fact, (\ref{nDIFFhierarchyEQlim}) for $n=1,2,\dots$ constitutes a ``Fokker--Planck hierarchy'' analogous to the the well-known Boltzmann, Landau and Vlasov hierarchies which arise in the validation of kinetic theory \cite{SpoBOOK,CIPbook} using ensembles. In our case the hierarchy has the very simplifying feature that the $n$-th equation in the hierarchy is decoupled from the equation for the $n+1$-th marginal. Since all the hierarchies used in the validation of kinetic theory are by construction \emph{linear}\footnote{More precisely, they are only essentially linear, for the parameters $\varepsilon_0$ and $\uV_0$, which also enter any of the other hierarchies whenever they describe ensembles of systems conserving mass, momentum, and energy, are all tied up with the initial conditions.} in the ``vector'' of the $f^{(n)}$, whenever one has a decoupling hierarchy one obtains closed linear equations for the $f^{(n)}$. In particular, our equation \refeq{nDIFFhierarchyEQlim} with $n=1$ is already a closed linear equation for $f^{(1)}$. However, at this point, any $f^{(n)}$ is still in general an ensemble superposition of states; in particular, $f^{(1)}$ still describes a statistical ensemble of pure states $f$ with same mass, momentum, and energy. By ignoring this fact one can mislead oneself into thinking that \refeq{nDIFFhierarchyEQlim} with $n=1$ and $f^{(1)}$ in place of $f$ is already the kinetic equation we sought. The final step in extracting \refeq{FPkinSIMPLE} as kinetic equation for the pure states involves the Hewitt--Savage \cite{HewSav} decomposition theorem. This theorem says that in the continuum limit any $f^{(n)}$ is a unique convex linear superposition of extremal (i.e. pure) $n$ particle states, and that these pure states are products of $n$ identical one-particle functions $f$ evaluated at $n$ generally different velocities. Each of the $f$ in the support of the superposition measure represents the velocity density function of an actual individual member of the infinite statistical ensemble of infinitely-many-particles systems. In formulas, at $\tau =0$ the initial data for $f^{(n)}$ read \begin{equation} f^{(n)}(\vV_1,\dots,\vV_n;0) = \langle f_0^{\otimes{n}}(\vV_1,...,\vV_n)\rangle, \label{HWinitially} \end{equation} where $\langle\,.\,\rangle$ is the Hewitt--Savage \cite{HewSav} ensemble decomposition measure on the space of initial velocity density functions $f_0$ of {individual physical systems} with same mass $m(f_0)(=1)$, momentum $\pV(f_0) = \uV_0$ and energy $e(f_0) = e_0 = \varepsilon_0 + |\uV_0|^2/2$. To extend this representation to $\tau>0$, let $U_\tau^{(n)}$ denote the one-parameter evolution semigroup for \refeq{nDIFFhierarchyEQlim}, i.e. $f^{(n)}(\vV_1,\dots,\vV_n;\tau)=U_\tau^{(n)}f^{(n)}_0(\vV_1,\dots,\vV_n)$. Noting now that the Hewitt--Savage measure is of course invariant under the evolution, and that by the linearity of \refeq{nDIFFhierarchyEQlim} it commutes with the linear operator $U_\tau^{(n)}$ for all $\tau\geq 0$, it follows that at later times $\tau >0$ the $n$ point density of the ensemble is given by \begin{equation} f^{(n)}(\vV_1,\dots,\vV_n;\tau) = \langle U_\tau^{(n)}f^{\otimes{n}}_0(\vV_1,...,\vV_n)\rangle. \end{equation} This so far simply states that, if the ensemble is initially a statistical mixture of pure states (product states), then at later times it is a statistical mixture of time-evolved initially pure states. Next we note that by inspection of \refeq{nDIFFhierarchyEQlim} it follows that \begin{equation} U_\tau^{(n)}f^{\otimes{n}}_0(\vV_1,...,\vV_n) = (U_\tau^{(1)}f_0)^{\otimes{n}}(\vV_1,...,\vV_n), \end{equation} viz. pure states evolve into pure states. Every factor $f(\vV_k;\tau)= U_\tau^{(1)}f_0(\vV_k)$ solves \refeq{FPkinSIMPLE} with $\tau = \frac{2\varepsilon_0}{3}t$, obeying the desired conservation laws. At last one can legitimately say that (\ref{FPkinSIMPLE}) has been derived as a full-fledged kinetic equation valid for almost every (w.r.t. $\langle\,.\,\rangle$) individual member of the limiting ensemble. \section{Summary and Outlook} In summary, the diffusion equation on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ can be interpreted as the simplest ``master equation'' for an underlying $N$-body Markov process with single-particle and pair terms. The $N\to\infty$ limit for the marginal densities of solutions to the diffusion equation is well-defined and can be carried out explicitly. After invoking the Hewitt--Savage decomposition, the limit $N\to\infty$ is seen to produce solutions of the ``kinetic Fokker--Planck equation'' describing individual isolated systems conserving mass, momentum, and energy. The Fokker--Planck equation \refeq{FPkin} is exactly solvable and displays correctly the qualitative behavior of a typical kinetic equation. In this sense, (\ref{FPkin}) really can be regarded as the simplest example of a kinetic equation of the ``diffusive" type, in the same family as, for instance, the much more complex Landau and Balescu-Lenard-Guernsey equations. Our work raises many new questions. 1) In particular, in Appendix Ab we have only written down the generator for the adjoint process of the underlying $N$-particle Markov process; hence, what is the explicit characterization of this process? 2) A derivation of a kinetic equation \`a la Kac is an intermediate step towards a full validation from some deterministic (Hamiltonian) microscopic model, which is in general a very difficult program, see the rigorous derivations of kinetic equations in \cite{SpoBOOK,CIPbook}. The substitute Markov process is usually chosen to preserve some of the essential features of the deterministic dynamics which (formally) leads to the same kinetic equation. Here we have only identified a stochastic model which leads to \refeq{FPkin}. Villani's work \cite{Vil98} suggests that a deterministic model may exist which in the kinetic regime leads to \refeq{FPkin}. Can one indentify this model? 3) In this paper, we conveniently assumed that the Fourier coefficients ensure convergence of the marginal density functions in $\mathfrak{L}^2\cap\mathfrak{L}^1$ and subsequently upgraded the regularity to Schwartz functions. What are the explicit conditions on the Fourier coefficients of the initial functions on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ which ensure convergence in $\mathfrak{L}^2\cap\mathfrak{L}^1$, in Schwartz space, in some topology for measures? 4) Since the PDEs in our finite-$N$ Fokker--Planck hierarchy are already self-contained for each $n$ (viz., they do not involve the usual coupling to $f^{(n+1)}$), the finite-$N$ corrections to the limiting evolutions can be studied in great detail; hence, for instance, how do the explicit corrections to propagation of chaos look? 5) We already mentioned in a footnote that the kinetic Fokker--Planck equation can easily be generalized to situations where the system is exposed to some external driving force by adding a forcing term. Can one derive this equation from some suitable ensemble of driven systems? Under which conditions do there exist stationary non-equilibrium states, and what are their stability properties? 6) Finally, our derivation is only valid for the space-homogeneous Fokker--Planck equation without driving force term; hence, can one extend our derivation to obtain the space-inhomogeneous generalization of the kinetic Fokker--Planck equation, first without and then with driving force term? These are many interesting questions which should be answered in future works. \medskip \noindent \textbf{Acknowledgment} We thank the referees for drawing our attention to \cite{BakMaz} and for their constructive criticisms. Thanks go also to Michael Loss for pointing out Mehler's paper \cite{Mehler}. Kiessling was supported by NSF Grant DMS-0103808. Lancellotti was supported by NSF Grant DMS-0318532. \newpage \section*{Appendix} \subsection*{A. Two useful representations of the Laplacian on spheres} \subsubsection*{Aa. Extrinsic representation in divergence form} For the purpose of obtaining equations for the marginals by integrating (\ref{heat}), it is advantageous to express the Laplacian on the right-hand side in terms of the projection operator $P_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ from $\mathbb{R}^{3N}$ to the fibers of the tangent bundle of the embedded manifold $\mathbb{M}^{3N-4}_{\uV_0,e_0}$. It is easy to verify \cite{KieLan04} that \begin{equation} \Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}} F^{(N)} = \nabla\cdot[P_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}\nabla F^{(N)}] \label{LapBel2} \end{equation} In order to have an explicit expression for $P_{\mathbb{M}^{3N-4}_{\uV_0,e_0}}$ we introduce an orthogonal basis for the orthogonal complement of the tangent space to $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ at $\vVN\in\mathbb{M}^{3N-4}_{\uV_0,e_0}\subset \mathbb{R}^{3N}$. Clearly, such orthogonal complement is spanned by the four vectors $\vVN$ and $\eVN_\sigma=(\eV_{\sigma},\dots,\eV_{\sigma})$, $\sigma=1,2,3$, where the $\eV_{\sigma}$ are the standard unit vectors in $\mathbb{R}^3$. The vectors $\eVN_\sigma$ are orthogonal to each other but not to $\vVN$; projecting away the non-orthogonal component of $\vVN$ yields \begin{equation} \biggl( \mathbf{I}_{3N} -\frac{1}{N}\sum_{\sigma=1}^3 \eVN_\sigma\otimes \eVN_\sigma \biggr) \cdot\vVN = \vVN - \uVN. \end{equation} The vectors $\{\vVN-\uVN, \eVN_1, \eVN_2, \eVN_3\}$ form the desired orthogonal basis; their magnitudes are $\abs{\eVN_\sigma}=\sqrt{N}$ and $\abs{\vVN-\uVN}=\sqrt{2N\varepsilon_0}$. Finally, (\ref{LapBel2}) becomes \begin{equation} \!\!\!\!\Delta_{\mathbb{M}^{3N-4}_{\uV_0,e_0}} F^{(N)} = {\partial_\vVN}\cdot \!\left[\!\left(\!\mathbf{I}_{3N}- \frac{1}{N} \sum_{\sigma=1}^3\eVN_\sigma\otimes\eVN_\sigma- \frac{1}{2N\varepsilon_0} (\vVN\!-\!\uVN)\otimes(\vVN\!-\!\uVN)\!\right)\!{\partial_\vVN F^{(N)}} \!\right]\! \label{heat1} \end{equation} \subsubsection*{Ab. Representation for the $N$-Body Markov Process} In the main part of this paper we started from the diffusion equation on the manifold $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ of $N$-body systems with same energy (per particle) $e_0$ and momentum (per particle) $\uV_0$, then took the limit $N\to\infty$, obtaining the kinetic Fokker--Planck equation \refeq{FPkinSIMPLE}, which rewrites into \refeq{FPkin} in view of the conservation laws. The Laplace--Beltrami operator on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ is the generator of the adjoint semigroup of the underlying stochastic Markov process that rules the microscopic dynamics of an individual $N$-body system. Here we show that this generator can be written as a sum of single particle and two-particle operators, thus characterizing the Markov process as a mixture of individual stochastic motions and stochastic binary interactions. Moreover, we show that the binary particle operators are the only ones that do not vanish in the $N\to\infty$ limit. This means that the kinetic Fokker--Planck equation can also be derived in terms of an $N$-body stochastic process with purely binary interactions, which is more satisfactory from a physical point of view. Recall that in section 2 we explained that $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ can be identified with the sphere $\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}$ centered at the origin of $\mathbb{L}^{3N-3}$ (which itself is an affine linear subspace of the space of all velocities, $\mathbb{R}^{3N}$). Recall that $\Delta_{\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}} = \frac{1}{2N\varepsilon_0}\Delta_{\mathbb{S}^{3N-4}}$. Note the well-known representation \begin{equation} \Delta_{\mathbb{S}^{3N-4}} = \!\!\!\!\!\!\!\!{\sum_{\qquad 1\leq k < l\leq 3(N-1)}}\!\! \Big( w_{k} \partial_{w_{l}} -w_{l} \partial_{w_{k}} \Big)^2, \label{LAPopDECOMPOSED} \end{equation} where ${w_{k}}$ is the $k$-th Cartesian component of $\wVN\in\mathbb{S}^{3N-4}\subset\mathbb{R}^{3(N-1)}$ (note that in section 2 we used $\wVN\in\mathbb{S}^{3N-4}_{\sqrt{2N\varepsilon_0}}$, but note furthermore that the r.h.s. of \refeq{LAPopDECOMPOSED} is invariant under $\wVN\to\lambda\wVN$). Grouping the components of $\wVN$ into blocks of vectors $\wV_k\in\mathbb{R}^3$, $k=1,...,N-1$, the r.h.s. of \refeq{LAPopDECOMPOSED} can be recast as \begin{eqnarray} \Delta_{\mathbb{S}^{3N-4}} = &&\!\!\!\! \Big. \sum_{k=1}^{N-1} \sum_{\stackrel{l=1}{l\ne k}}^{N-1} \Big(3\wVk\cdot\partial_{\wV_k}+\abs{\wVk}^2\partial_{\wV_l}\cdot\partial_{\wV_l} - \big( \wVk\cdot\partial_{\wV_k} \big) \big(\wVl\cdot\partial_{\wV_l} \big) \Big) \nonumber\\ &&- \sum_{k=1}^{N-1} \big( \wVk\times\partial_{\wV_k} \big)^2 \Big. , \label{MASTERopDECOMPOSED} \end{eqnarray} containing one-body terms as well as binary terms. Note however that the first term in the binary sum is effectively a sum of two-body terms in disguise, which scale with factor $N-2$ and thus survive in the limit $N\to\infty$, while the true one-body sum (second line) drops out in that limit. This implies that the kinetic Fokker--Planck equation \refeq{FPkinSIMPLE} can be derived from a master equation on $\mathbb{M}^{3N-4}_{\uV_0,e_0}$ which contains \emph{only} the binary terms (first line) in \refeq{MASTERopDECOMPOSED}. This in turn implies that \refeq{FPkinSIMPLE} is the kinetic equation for an underlying system of $N$ particles with stochastic pair interactions. \subsection*{B. High-Dimension Asymptotics of Associated Legendre Functions} In \refeq{eigmarg2} the associated Legendre functions of degree $s=0,1,2,...$ and order $r=0,...,s$ in $q$ dimensions occur. They are defined on the interval $[-1,1]$ and given by \begin{equation} \widetilde{P}_s^r(t;q) = \sqrt{q}^{s+r} \frac{s!}{2^r}\,\Gamma\left(\frac{q-1}{2}\right) \sum_{l=0}^{\intgpart} \left(-\frac{1}{4}\right)^l \frac{(1-t^2)^{l+\frac{r}{2}}\, t^{s-r-2l}} {l!\,(s-r-2l)!\,\Gamma\left(l+r+\frac{q-1}{2}\right)} \label{LEGENDREf} \end{equation} which differ from the $P_s^r(t;q)$ in \cite{Mul} in their normalization. In our investigation, $q = 3N - p$ and $t = \frac{w}{\sqrt{\scriptstyle{{2N\vareps_0}}}}$, and we are interested in the limit $N\to\infty$. The familiar asymptotics of Euler's Gamma function gives us \begin{equation} \frac{\Gamma\left(x\right)} {\Gamma\left(a+x\right)} = x^{-a} + O\left(x^{-(a+1)}\right). \end{equation} for $x\gg 1$. Applying this asymptotics with $2x = q-1=3N-p-1$ and $a=l+r$ to \refeq{LEGENDREf}, we find that given $p\in\mathbb{N}$ and $w\in \mathbb{R}$ (which implies $N> \max\{p/3,w^2/(2\varepsilon_0)\}$), when $N\gg 1$ we have \begin{eqnarray} \widetilde{P}_s^r\left(\frac{w}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N\!-\!p\right) =&&\!\! \sqrt{2}^{r-s} \sum_{l=0}^{\intgpart} (-1)^l \frac{s!}{l!\,(s-r-2l)!} \left(\sqrt{{\textstyle{\frac{3}{\varepsilon_0}}}}\,w\right)^{s-r-2l} \nonumber \\ &&\!\! +\; O\!\left(\frac{1}{\sqrt{N}} \right). \end{eqnarray} By comparing with the formula for the Hermite polynomial of degree $k$ on $\mathbb{R}$, \begin{equation} H_{k}(x) = \sum_{l=0}^{\big\lfloor\!\frac{{\scriptstyle k}}{2}\!\big\rfloor} (-1)^{l+k}\frac{ s!}{l!\,(k-2l)!}(2x)^{k-2l}, \end{equation} we see that, given $p\in\mathbb{N}$ and $w\in \mathbb{R}$, we have \begin{eqnarray} \widetilde{P}_s^r\left(\frac{w}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-p\right) = \left(-\sqrt{2}\right)^{r-s}\, H_{s-r}\left(\sqrt{{\textstyle{\frac{3}{4\varepsilon_0}}}}w\right) +\; O\!\left(\frac{1}{\sqrt{N}} \right) \end{eqnarray} when $N\gg 1$. Hence, for all fixed $p$ we now find that pointwise for any $w\in\mathbb{R}$, \begin{equation} \lim_{N\to\infty} \widetilde{P}_s^r\left(\frac{w}{\sqrt{\scriptstyle{{2N\vareps_0}}}};3N-p\right) = \left(-\sqrt{2}\right)^{r-s}\, H_{s-r}\left(\sqrt{{\textstyle{\frac{3}{4\varepsilon_0}}}}w\right) \end{equation} where again it is understood that $N> \max\{p/3,w^2/(2\varepsilon_0)\}$ in the expression under the limit in the left-hand side. Equation (\ref{eigmarg3}) in the main text follows. \newpage
{ "timestamp": "2005-11-03T23:18:24", "yymm": "0503", "arxiv_id": "math-ph/0503073", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503073" }
\section{Introduction} In the last years, several theoretical approaches have predicted strong medium effects on the pion pion interaction in the scalar isoscalar ($\sigma$) channel. In Ref. \cite{Hatsuda:1999kd}, Hatsuda et al. studied the $\sigma$ propagator in the linear $\sigma$ model and found an enhanced and narrow spectral function near the $2\pi$ threshold caused by the partial restoration of the chiral symmetry, where $m_\sigma$ would approach $m_\pi$. The same conclusions were reached using the nonlinear chiral Lagrangians in Ref. \cite{Jido:2000bw}. Similar results, with large enhancements in the $\pi\pi$ amplitude around the $2\pi$ threshold, have been found in a quite different approach by studying the $s-$wave, $I=0$ $\pi\pi$ correlations in nuclear matter \cite{Schuck:1988jn,Rapp:1996ir,Aouissat:1995sx,Chiang:1998di}. In these cases the modifications of the $\sigma$ channel are induced by the strong $p-$wave coupling of the pions to the particle-hole ($ph$) and $\Delta$-hole ($\Delta h$) nuclear excitations. It was pointed out in \cite{Aouissat:2000ss,Davesne:2000qj} that this attractive $\sigma$ selfenergy induced by the $\pi$ renormalization in the nuclear medium could be complementary to additional $s$-wave renormalizations of the kind discussed in \cite{Hatsuda:1999kd,Jido:2000bw} calling for even larger effects. On the experimental side, there are also several results showing strong medium effects in the $\sigma$ channel at low invariant masses in the $A(\pi,2\pi)$ \cite{Bonutti:1996ij,Bonutti:1998zw,Camerini:1993ac,bonutti,Starostin:2000cb} and $A(\gamma,2\pi)$ \cite{Messchendorp:2002au} reactions. At the moment, the cleanest signal probably corresponds to the $A(\gamma,2\pi^0)$ reaction, which shows large density effects that had been predicted in both shape and size in Ref. \cite{Roca:2002vd}, using a model for the $\pi\pi$ final state interaction along the lines of the present work. Note, however, that a part of the spectrum modification could be due to quasielastic collisions of the pion \cite{Muhlich:2004zj}. Our aim in this paper is to study the $\pi\pi$ scattering in the scalar isoscalar ($\sigma$) channel at finite densities in the context of the model developed in \cite{Dobado:1990qm,Dobado:1993ha,Oller:1998ng,Oller:1999hw,Oller:1999zr,Oller:1997ti}. These works, which provide an economical and successful description of a wide range of hadronic phenomenology, use as input the lowest orders of the Lagrangian of Chiral Perturbation Theory ($\chi PT$) \cite{Gasser:1985ux} and calculate meson meson scattering in a coupled channels unitary way. Some nuclear medium effects, namely the $p-$wave coupling of the pions to the particle hole ($ph$) and Delta hole ($\Delta h$) excitations, were implemented in this framework in Refs. \cite{Chiang:1998di,Oset:2000ev}. As in other approaches, large medium effects were found as reflected in the imaginary part of the $\pi \pi$ scattering amplitude which showed a clear shift of strength towards low energies as the density increases. Although this model was able to predict the size of the medium effects on the $(\gamma, 2\pi)$ reaction \cite{Roca:2002vd}, it was pointed out that some probably large contributions related to nucleon tadpole diagrams \cite{Jido:2000bw} and some vertex corrections \cite{Meissner:2001gz} were missing. In this work, we will include those pieces and analize its influence in the $\pi\pi$ scattering amplitude at finite nuclear densities. In the next section we present, for the sake of completeness, a brief description of the model used for the $\pi \pi$ interaction both in vacuum and in a dense medium, which is already published elsewhere \cite{Chiang:1998di,Oset:2000ev}. In Section 3 we consider further contributions to the $\pi\pi$ interaction in the nuclear medium, associated to higher order terms in the chiral Lagrangian than those included in Refs. \cite{Chiang:1998di,Oset:2000ev}, and some baryonic vertex corrections advocated in Ref. \cite{Meissner:2001gz}. \section{$\pi \pi$ interaction} In this section we summarize the method of Ref. \cite{Oller:1997ti} for $\pi \pi$ interaction in vacuum and Refs. \cite{Chiang:1998di,Oset:2000ev} for the nuclear medium effects. Additional information on this and related approaches for different spin isospin channels can be found in Refs. \cite{Oller:1997ti,Oller:1998ng,Oller:1999hw,Nieves:2000bx}. \subsection{Vacuum} The basic idea is to solve a Bethe Salpeter (BS) equation, which guarantees unitarity, matching the low energy results to $\chi PT$ predictions. We consider two coupled channels, $\pi \pi$ and $K \bar{K}$ and neglect the $\eta \eta $ channel which is not relevant at the low energies we are interested in. The BS equation is given by \begin{equation} \label{eq:BS} T=V+VGT. \end{equation} Eq. (\ref{eq:BS}) is a matrix integral equation which involves the two mesons one loop divergent integral (see Fig.~\ref{fig:BSF}), where $V$ and $T$ appear off shell. However, for this channel both functions can be factorized on shell out of the integral. The remaining off shell part can be absorbed by a renormalization of the coupling constants as it was shown in Refs. \cite{Oller:1997ti,Nieves:1999hp}. Thus, the BS equation becomes purely algebraic and the $VGT$ term, originally inside the loop integral, becomes then the product of $V$, $G$ and $T$, with $V$ and $T$ the on shell amplitudes independent of the integration variables, and $G$ given by the expression \begin{equation} G_{ii}(P) = i \int \frac{d^4 q}{(2 \pi)^4} \frac{1}{q^2 - m_{1i}^2 + i \epsilon} \; \; \frac{1}{(P - q)^2 - m_{2i}^2 + i \epsilon} \end{equation} where $P$ is the momentum of the meson meson system. This integral is regularized with a cut-off ($\Lambda$) adjusted to optimize the fit to the $\pi\pi$ phase shifts ($\Lambda=1.03$ GeV). \begin{figure} \begin{center} \epsfig {figure=fig_1.eps,width=12.cm} \caption{Diagrammatic representation of the Bethe Salpeter equation.} \label{fig:BSF} \end{center} \end{figure} The potential $V$ appearing in the BS equation is taken from the lowest order chiral Lagrangian \begin{equation} {\cal L}_2 = \frac{1}{12 f^2} \langle (\partial_\mu \Phi \Phi - \Phi \partial_\mu \Phi)^2 + M \Phi^4 \, \rangle \end{equation} \noindent where the symbol $\langle \rangle$ indicates the trace in flavour space, $f$ is the pion decay constant and $\Phi$, $M$ are the pseudoscalar meson and mass $SU(3)$ matrices. This model reproduces well phase shifts and inelasticities up to about 1.2 GeV. The $\sigma$ and $f_0 (980)$ resonances appear as poles of the scattering amplitude in $L=0$, $I=0$. The coupling of channels is essential to produce the $f_0 (980)$ resonance, while the $\sigma$ pole is little affected by the coupling of the pions to $K \bar{K}$ \cite{Oller:1997ti}. \subsection{\label{sec:nucmed}The nuclear medium} As we are mainly interested in the low energy region, which is not very sensitive to the kaon channels, we will only consider the nuclear medium effects on the pions. The main changes of the pion propagation in the nuclear medium come from the $p-$wave selfenergy, produced basically by the coupling of pions to particle-hole ($ph$) and Delta-hole ($\Delta h$) excitations. For a pion of momentum $q$ it is given by \begin{equation} \label{eq:self} \Pi(q)= {{\left({D+F}\over{2f}\right)^2 \vec q\,^2 U(q)} \over {1-\left({D+F}\over{2f}\right)^2 g' U(q)}} \end{equation} with $g'=0.7$ the Landau-Migdal parameter, $U(q)$ the Lindhard function and $(D+F)=1.257$. The expressions for the Lindhard functions are taken from Ref. \cite{Oset:1990ey}. Thus, the in-medium BS equation will include the diagrams of Fig. \ref{fig:BSF2} where the solid line bubbles represent the $ph$ and $\Delta h$ excitations. \begin{figure}[htb] \begin{center} \epsfig{height=2.2cm,width=12.1cm,angle=0, figure=fig_2.eps} \caption{Terms of the meson meson scattering amplitude accounting for $ph$ and $\Delta h$ excitation.} \label{fig:BSF2} \end{center} \end{figure} In fact, as it was shown in \cite{Chanfray:1999nn}, the contact terms with the $ph$ ($\Delta h$) excitations of diagrams (b-d) cancel the off-shell contribution from the meson meson vertices in the term of Fig. \ref{fig:BSF2}(a). Hence, we just need to calculate the diagrams of the free type (Fig. \ref{fig:BSF}) and those of Fig. \ref{fig:BSF2}(a) with the amplitudes factorized on shell. Therefore, at first order in the baryon density, we are left with simple meson propagator corrections which can be readily incorporated by changing the meson vacuum propagators by the in medium ones. The $\pi\pi$ scattering amplitude obtained using this model exhibits a strong shift towards low energies. In Fig. \ref{fig:MED1}, we show the imaginary part of this amplitude for several densities. Quite similar results have been found using different models \cite{Aouissat:1995sx} and it has been suggested that this accumulation of strength, close to the pion threshold, could reflect a shift of the $\sigma$ pole which would approach the mass of the pion. \begin{figure}[htb] \begin{center} \epsfig{width=12.1cm, figure=fig_3.eps} \caption{Imaginary part of the $\pi\pi$ scattering amplitude at several densities.} \label{fig:MED1} \end{center} \end{figure} Other pion selfenergy contributions related to $2ph$ excitations, and thus proportional to $\rho^2$, can be incorporated in the pion propagator. As we are most interested in the region of low energies we can take as estimation the corresponding piece of the optical potentials obtained from pionic atoms data, following the procedure of Ref. \cite{Chiang:1998di} and substituting in Eq. (\ref{eq:self}) \begin{equation} \left({D+F}\over{2f}\right)^2 U(q) \;\longrightarrow\; \left({D+F}\over{2f}\right)^2 U(q) -4\pi C_0^* \rho ^2 \label{eq:self2} \end{equation} with $\rho$ the nuclear density and $C_0^*=(0.105+i 0.096) m_\pi^{-6}$. Its effects are small except at large densities as can be appreciated by comparing Fig. \ref{fig:MED1}, with Fig. 7 of Ref. \cite{Chiang:1998di} where this piece is included. \section{Further contributions} \subsection{\label{sec:tadpoles}Higher order tadpole and related terms} The chiral Lagrangian generates tadpole terms that could contribute to the pion selfenergy and also in the form of vertex corrections as in Fig. \ref{fig:NUCTAD}. At the lowest order these terms vanish in isospin symmetric nuclear matter \cite{Oset:2000ev}. However, at next order there are terms which provide some contribution. The complete structure of the higher order Lagrangian adapted to the $\pi N$ system can be seen in \cite{Bernard:1995dp}. The medium corrections associated to these new Lagrangian terms in the $\pi$ nucleus interaction were studied in \cite{Thorsson:1995rj} and interpreted in terms of changes of the time and space components of $f$ and changes of the pion mass in the medium. Further developments in this direction are done in \cite{Meissner:2001gz}. The repercussion of these terms in $\pi\pi$ scattering in the nuclear medium has been considered in \cite{Jido:2000bw} and we follow here the same steps. We start from the second order $\pi N$ Lagrangian relevant for the isoscalar sector \begin{eqnarray} \label{LpiN2} {\cal L}_{\pi N}^{(2)} & = & c_3 \bar{N} (u_{\mu} u^{\mu}) N + (c_2 - {g_A^2 \over 8 m_N}) \bar{N}({\rm v}_{\mu} u^{\mu})^2 N \nonumber \\ & & + c_1 \bar{N}N {\rm Tr} (U^{\dagger} \chi + \chi^{\dagger} U) + \cdot \cdot \cdot , \end{eqnarray} where $u_{\mu} = i u^{\dagger} \partial_{\mu} U u^{\dagger}$, with $U = u^2 = {\rm exp}(i \tau^a \phi^a /f)$ in the $SU(2)$ formalism used there, ${\rm v}_{\mu}$ is the four velocity of the nucleon, $g_A$ the axial charge of the nucleon and $\chi = {\rm diag}(m_{\pi}^2,m_{\pi}^2)$. The pion nucleon amplitude obtained from the Lagrangian in Eq. (\ref{LpiN2}) is \begin{eqnarray} \label{TpiNfromLpiN2} t_{\pi N} &=& \frac{4 c_1}{f^2} m_{\pi}^2 - \frac{2 c_2}{f^2} (q^0)^2 - \frac{2 c_3}{f^2} q^2 \nonumber \\ &=& ( \frac{4 c_1}{f^2} m_{\pi}^2 - \frac{2 c_2}{f^2} \omega(q)^2 - \frac{2 c_3}{f^2} m_{\pi}^2 ) \nonumber \\ & & - \frac{2 c_2 + 2 c_3}{f^2} (q^2-m_{\pi}^2) = t_{\pi N}^{on} + t_{\pi N}^{off} \,\,\, , \end{eqnarray} where in the last part of the equation we have separated what we call the on-shell part and the off-shell part of the amplitude (term with $(q^2-m_{\pi}^2)$). This $s-$wave $\pi N$ interaction produces a modification of the pion propagator which we shall consider later in the solution of the Bethe Salpeter equation in the medium. In \cite{Jido:2000bw} and \cite{Thorsson:1995rj} the medium effects are recast at the mean field level in terms of a medium Lagrangian given by \begin{eqnarray} \label{mean-f} \langle {\cal L} \rangle & = & ( {f^2 \over 4} + {c_3 \over 2} \rho)\ {\rm Tr} [\partial_{\mu} U \partial^{\mu} U^{\dagger}] \nonumber \\ & & + \ \ ( {c_2 \over 2} - {g_A^2 \over 16 m_N})\ \rho \ {\rm Tr} [\partial_0 U \partial_0 U^{\dagger}] \nonumber \\ & & + \ \ ({ f^2 \over 4} + {c_1 \over 2} \rho) \ {\rm Tr} (U^{\dagger} \chi + \chi^{\dagger} U) \,\,\, . \end{eqnarray} The different corrections to the $\pi\pi$ scattering amplitude coming from the $\partial_{\mu} U \partial^{\mu} U^{\dagger}$, $\partial_0 U \partial^0 U^{\dagger}$ terms and the mass term in Eq. (\ref{mean-f}) ($c_3$, $c_2$ and $c_1$ terms) are given by \begin{eqnarray} \label{correc_derivative} \delta t_{\pi\pi}^{(t)} &=&- \frac{1}{f^2} \lbrace \frac{2 c_3}{f^2} \rho (s-\frac{4}{3}m_{\pi}^2) + \frac{2 c_2}{f^2} \rho (s-\frac{1}{3} \sum_i \omega_i(q)^2) + \frac{c_1}{f^2}\rho\frac{5}{6} m_{\pi}^2 \rbrace \nonumber \\ & &+ \frac{1}{f^2} \lbrace \frac{2 c_3}{f^2} \rho \frac{1}{3} \sum_i (q_i^2-m_{\pi}^2) + \frac{2 c_2}{f^2} \rho \frac{1}{3} \sum_i (q_i^2-m_{\pi}^2) \rbrace \,\,\, , \end{eqnarray} where we have also separated the on-shell part from the off-shell part. These are the corrections coming from the many body tadpole diagram of Fig. \ref{fig:NUCTAD}, which are included in the $\rho$ dependent terms of Eq. (\ref{mean-f}). Note that in the chiral unitary approach that we follow, the external legs are placed on shell ($q_i^2=m_{\pi}^2$). This is the case even when the diagrams appear in loops, as in Fig. \ref{fig:NUCTADLOOP}, since the underlying physics is the use of a dispersion relation using the $N/D$ method \cite{Oller:1998zr,Oller:2000fj} which determines the diagram contribution in terms of its imaginary part. In the case of Fig. \ref{fig:NUCTADLOOP} the cut corresponds to two free pions on shell, like in the vacuum. Hence, we shall use only the on shell part of the correction of Eq. (\ref{correc_derivative}). \begin{figure}[htb] \begin{center} \epsfig{width=6cm,figure=fig_4.eps} \caption{Nucleon tadpole term correction to the $\pi\pi$ interaction.} \label{fig:NUCTAD} \end{center} \end{figure} \begin{figure}[htb] \begin{center} \epsfig{width=8cm,figure=fig_5.eps} \caption{$\pi\pi$ rescattering diagram with tadpole vertex correction showing the $\pi\pi$ cut.} \label{fig:NUCTADLOOP} \end{center} \end{figure} As mentioned before, at the same time, when solving the Bethe-Salpeter equation, we have also to take into account the $s-$wave selfenergy insertion from the Lagrangian of Eq. (\ref{LpiN2}) in the pion propagators as depicted in Fig. \ref{fig:NUCTADLOOPSERIES}. This is easily accounted for, at lowest order in $\rho$, adding to each pion propagator, $D_{\pi}$, the correction $D_{\pi} t_{\pi N} \rho D_{\pi}$. A technically simple way to account for that is to add to the scalar isoscalar $\pi\pi$ vertex from ${\cal L}_2$, $t_{\pi\pi}$, the correction \begin{equation} \label{swaveinsertion} \delta t_{\pi\pi}^{(s)} = (t_{\pi N}^{on}+t_{\pi N}^{off}) \rho \frac{1}{q^2-m_{\pi}^2} t_{\pi\pi} \end{equation} for the two pion propagator lines to the left of the $\pi\pi$ vertex. Now the separation of the on-shell and off-shell parts of $t_{\pi N}$ is most useful since the pion propagator in Eq. (\ref{swaveinsertion}) is cancelled out by the $(q^2-m_{\pi}^2)$ factor of the off-shell part of $t_{\pi N}$. Thus we have \begin{equation} \label{swaveinsertion2} \delta t_{\pi\pi}^{(s)} = t_{\pi N}^{on} \, \rho \, \frac{1}{q^2-m_{\pi}^2} t_{\pi\pi} - \frac{2 c_2 + 2 c_3}{f^2} \, \rho \, t_{\pi\pi} \,\,\, . \end{equation} This means that with the Lagrangian used, on top of the corrections in the loops from the $s-$wave (on-shell) pion selfenergy, we have an additional correction (second term of Eq. (\ref{swaveinsertion2})) of the same topology as the tadpole term considered before. Considering the pion selfenergy insertion in either of the two pion propagators, we obtain for this \begin{equation} \label{swaveinsertion3} \delta t_{\pi\pi}^{(st)} = -\frac{4c_2 + 4c_3}{f^2} \, \rho \, t_{\pi\pi} \,\,\, . \end{equation} Thus, we are left with the usual contribution in the pion loops of the ordinary on-shell $s-$wave pion selfenergy, plus the tadpole correction of Eq. (\ref{correc_derivative}), plus the tadpole equivalent of Eq. (\ref{swaveinsertion3}). \begin{figure}[htb] \begin{center} \epsfig{width=10cm,figure=fig_6.eps} \caption{Nucleon tadpole correction in the pion propagator.} \label{fig:NUCTADLOOPSERIES} \end{center} \end{figure} There are still further contributions belonging to the same family. Indeed, the $t_{\pi\pi}$ amplitude in the scalar isoscalar channel, \begin{equation} \label{tpipiatrhozero} t_{\pi\pi} = -\frac{1}{f^2} (s-\frac{m_{\pi}^2}{2} - \frac{1}{3}\sum_i(q_i^2-m_{\pi}^2)) \,\,\,, \end{equation} is also split in on- and off-shell parts. In \cite{Chiang:1998di,Chanfray:nn} it was shown that the off-shell pieces could be removed from the loop calculations for both the free pion case and the pion with a $p-$wave selfenergy. However, the diagrams in Fig. \ref{fig:NUCTADLOOPSERIES}(a) have one free pion and a pion with a $s-$wave medium selfenergy insertion, hence the imaginary part of the two-pion loop is not the same as in the mentioned cases. It is again easy to take into account this correction and we have, from the $s-$wave selfenergy insertions in the pion propagators \begin{equation} \label{swaveinsertion4} \delta t_{\pi\pi}^{(so)} = t_{\pi N}\rho \frac{1}{q^2-m_{\pi}^2} \frac{1}{3 f^2} (q^2-m_{\pi}^2) \equiv \frac{1}{3 f^2} t_{\pi N}\rho \end{equation} for each pion line. Next we separate the on-shell and off-shell parts of $t_{\pi N}$. For the on-shell part we get \begin{equation} \label{swaveinsertion5} \frac{1}{3 f^2} (\frac{4 c_1}{f^2}m_{\pi}^2-\frac{2 c_2}{f^2}\omega(q)^2- \frac{2 c_3}{f^2}m_{\pi}^2) \rho \,\,\, , \end{equation} which compared at threshold to the free $t_{\pi\pi}$ amplitude, $t_{\pi\pi}= -\frac{1}{f^2} \frac{7}{2} m_{\pi}^2$, gives \begin{equation} \label{swaveinsertion6} \frac{\delta t_{\pi\pi}}{t_{\pi\pi}} \simeq \frac{1}{21 f^2} (8 c_1 - 4 c_2 - 4 c_3) \rho \,\,\, , \end{equation} which with respect to Eq. (\ref{swaveinsertion3}) gets a reduction of a factor 21, plus an extra reduction from the near on-shell cancellation of the isoscalar $t_{\pi N}$. Hence, this correction is negligible and we take advantage of this large reduction factor $21$ to also neglect the part involving simultaneously the off-shell parts of $t_{\pi\pi}$ and $t_{\pi N}$. In order to proceed we have to decide upon the $c_i$ coefficients to be used. It is well known that the Lagrangian of Eq. (\ref{LpiN2}) leads to a part of $p-$wave pion selfenergy \cite{Kirchbach:1996xy}, but we are explicitly taking a $p-$wave selfenergy insertion accounting for $ph$ and $\Delta h$ excitations. There is a work which uses the same Lagrangian of Eq. (\ref{LpiN2}), and in addition takes into account explicitly the $\Delta$ degrees of freedom \cite{Fettes:2000bb}. Thus, we stick to the values of the $c_i$ coefficients obtained there from two fits, with and without using the $\sigma$ term as a constraint, shown in Table \ref{ci}. For comparison, the values of the coefficients $c_i$ without including the $\Delta$ are of the order of $c_1=-1.53$ GeV$^{-1}$, $c_2=3.22$ GeV$^{-1}$ and $c_3=-6.20$ GeV$^{-1}$ \cite{Fettes:1998ud}. \begin{table}[ht] \begin{center} \begin{tabular}{|l|l|l|} coef.& set I (GeV$^{-1}$) & set II (GeV$^{-1}$) \\ \hline $c_1$ & $-0.35$ & $-0.32$ \\ \hline $c_2$ & $-1.49$ & $-1.59$ \\ \hline $c_3$ & $0.93$ & $1.15$ \\ \hline \end{tabular} \caption{\footnotesize{$c_i$ coefficients from Ref. \cite{Fettes:1998ud}.}} \label{ci} \end{center} \end{table} As stressed in \cite{Fettes:1998ud} the values of the coefficient $c_i$ with the explicit contribution of the $\Delta$ are of natural order, while those obtained without its consideration are too large and a source of problems in chiral perturbative calculations \cite{Epelbaum:2003gr}. But in our case, as pointed above, the choice is mandatory. We can estimate the size of the correction of Eq. (\ref{correc_derivative}) at pion threshold, and taking advantage of the reduction factor $1/3$ in the term $\frac{1}{3} \omega_i^2(q)$ in front of $s \simeq 4 m_{\pi}^2$, we approximate $\omega_i(q) \simeq m_{\pi}$. So we get \begin{equation} \label{delta} \frac{\delta t_{\pi\pi}}{t_{\pi\pi}} = \frac{32}{21 f^2}(c_2 + c_3) \rho + \frac{10}{21 f^2}c_1 \rho \,\,\, , \end{equation} which for the two sets of parameters of Table \ref{ci} gives \begin{eqnarray} \label{deltanum} \frac{\delta t_{\pi\pi}}{t_{\pi\pi}} &=& -0.154 \, \rho / \rho_0 \,\,\, \textrm{(set I)} \nonumber \\ &=& -0.124 \, \rho / \rho_0 \,\,\, \textrm{(set II)} \,\,\,. \end{eqnarray} Let us note that the correction is negative, reducing effectively the strength of the $\pi\pi \to \pi\pi$ vertex in the medium. Note that should we have used the values of $c_i$ without explicit $\Delta$ we would obtain a value for the ratio of Eq. (\ref{delta}) of $-0.80 \rho / \rho_0$, certainly too large, but also negative. Next we consider the contribution from Eq. (\ref{swaveinsertion3}). This correction has opposite sign to the former one. When adding the two corrections we find, again taking the threshold for comparison, \begin{equation} \label{deltatotal} \frac{\delta t_{\pi\pi}}{t_{\pi\pi}} = -\frac{52}{21 f^2}(c_2 + c_3) \rho + \frac{10}{21 f^2} c_1 \rho \,\,\, , \end{equation} which for the two sets of values of Table \ref{ci} gives \begin{eqnarray} \label{deltatotalnum} \frac{\delta t_{\pi\pi}}{t_{\pi\pi}} &=& 0.18 \, \rho / \rho_0 \,\,\, \textrm{(set I)} \nonumber \\ &=& 0.14 \, \rho / \rho_0 \,\,\, \textrm{(set II)} \,\,\,. \end{eqnarray} We can see that the sign of the correction is now reversed and, altogether, we find now an effective increase of the $\pi\pi$ vertex in the medium by a moderate amount. Apart from the vertex corrections, we need to include the effect of the on-shell $s-$wave pion selfenergy in the pion propagators in the loops, produced by the nucleon tadpole diagram. Since we have a broad range of pion energies in the loop, we have used the $t \rho$ approximation for the $s-$wave pion selfenergy and the amplitude $t$ has been taken from the experimental fit to data \cite{Arndt:1995bj}. This is a more realistic approach than to take the expression from the model used here which gives a too large $s-$wave scattering amplitude at high energies, and in any case produces a minor effect. The considered corrections are included in the $\pi\pi$ amplitude by modifying the kernel of the BS equation with the on-shell part of Eq. (\ref{correc_derivative}) and Eq. (\ref{swaveinsertion3}), namely \begin{equation} \label{BStad} T = \frac{V+\delta t_{\pi\pi}^{(t)\,on}} {1-(V+\delta t_{\pi\pi}^{(t)\,on}+\delta t_{\pi\pi}^{(st)})\,G} \,\,\, , \end{equation} and modifying the pion propagators in the calculation of the two-pion loop function, $G$, as explained in Section \ref{sec:nucmed}. \subsection{\label{sec:newmech}Vertex corrections from baryonic loops} In the previous sections we have considered relevant medium effects, according to the pion nucleus phenomenology, which describe correctly the pion in the medium in a wide range of energies. These mechanisms lead to $s-$ and $p-$ wave pion selfenergies in the propagators of the BS equation and some associated vertex corrections. In Ref. \cite{Meissner:2001gz}, other vertex corrections to the $\pi\pi$ amplitude which could provide some effect at low energies, where the leading $p-$wave pion selfenergy is not so strong, were studied. \begin{figure} \begin{center} \includegraphics[width=10cm,height=5cm]{fig_7.eps} \caption{\label{new_mech}(a) $ph$ bubble exchange in the $t$ channel; (b) Box diagram.} \end{center} \end{figure} The mechanisms considered in \cite{Meissner:2001gz} relevant for $\pi\pi$ $s-$wave scattering, modifying the kernel of the Bethe Salpeter equation in the $\pi \pi$ interaction, are shown in Fig. \ref{new_mech}. The $\pi \pi$ isoscalar contribution in the $s$ channel for the $p h$ excitation in the $t$ channel in Fig. \ref{new_mech}(a) is given, with the unitary normalization, by \begin{equation} -i t = - \left( \frac{1}{4 f^2} \right) ^2 (p^0_1+k^{0}_1)(p^0_2+k^0_2) U(q) = \tilde{t} \, U(q) \label{textra} \end{equation} where $U(q)$ is the ordinary Lindhard function for $ph$ excitation, including a factor 2 of isospin (see Appendix of \cite{Oset:1990ey}). The second equation in Eq. (\ref{textra}) defines $\tilde{t}$. We have neglected the isoscalar $\pi N$ amplitude in Eq. (\ref{textra}) since it is very small compared with the isovector one \cite{Schroder:uq}. The magnitude of $t$ was shown in \cite{Meissner:2001gz} to be comparable to the $s-$wave $V$ of the lowest order chiral Lagrangian at densities of the order of the nuclear density. Yet, there are some observations to be made: First, at pion threshold the diagram of Fig. \ref{new_mech}(a) is proportional to $U(q^0=0,\vec{q}=\vec{0})$. This quantity is evaluated in \cite{Meissner:2001gz} using the ordinary limit of the Lindhard function at $q^0=0$ and $|\vec{q}| \to 0$, which is finite and larger in size than for any finite value of $|\vec{q}|$. This limit is however quite different from the value of the response function at $\vec{q}=\vec{0}$ in finite nuclei which is strictly zero, as already noted in \cite{Meissner:2001gz,Oset:xm,Oset:sm}. We take into account the fact that the isovector $\pi N$ amplitude reflects the exchange of a $\rho$ in the $t$ channel \cite{Ericson:gk} and multiply $(4f^2)^{-2}$ by a factor reflecting the two $\rho$ propagators, $F(q)=(M_{\rho}^2/(M_{\rho}^2+\vec{q}\,^2))^2$. \begin{figure} \begin{center} \includegraphics[width=10cm,height=5cm]{fig_8.eps} \caption{\label{oneloopBS}Loop contributions to the Bethe Salpeter equation at first order in the nuclear density, including the $t-$channel $ph$ excitation.} \end{center} \end{figure} In order to estimate the importance of this contribution as compared to the $p-$wave pion selfenergy insertions, we have evaluated the diagrams (a-e) in Fig. \ref{oneloopBS}. Details of the calculation are given in Appendix II. The results are shown in Fig. \ref{res:oneloopBS} for the imaginary part of the resulting amplitude. We find that the contribution of diagrams (d,e) is smaller than the changes produced by the insertion of the $p-$wave pion selfenergy in the pion propagators. Similar results are found for the real part of the amplitude. \begin{figure} \begin{center} \includegraphics[width=0.7\textwidth]{fig_9.eps} \caption{\label{res:oneloopBS}Imaginary part of the $\pi\pi$ amplitude from the terms in Fig. \ref{oneloopBS}, as indicated in the legend. The calculation for the dashed and dotted lines is done for $\rho=\rho_0/2$.} \end{center} \end{figure} The consideration of this mechanism in the BS equation proceeds by adding the tree level term and modifying the kernel in the loop terms with an effective potential $\delta V$ defined as \begin{equation} \frac{\delta V(s,\rho)}{V} = \frac{{\cal F}(s,\rho)}{V G V} \,\,\, , \label{Veff} \end{equation} where ${\cal F}(s,\rho)$ is the amplitude corresponding to diagram (b) and $VGV$ gives the amplitude of diagram (a) in Fig. \ref{oneloopBS}. In this sense, by substituting $V$ by $V+\delta V$ in the $\pi\pi$ vertex, the loop function of diagram in Fig. \ref{oneloopBS}(a) would account correctly for all the diagrams (a-e) at the first order in the nuclear density. One of the reasons for the small size of this contribution is that the Lindhard function behaves roughly as $q^{-2}$ for large values of $q$ and we should expect a large cancellation of this piece in the loops. This would be in contrast with the $ph$ excitations leading to the $p-$wave $\pi$ selfenergy in Fig. \ref{oneloopBS}(b,c), since there one has the combination $\vec{q}\,^2 U(q)$ and a priori this type of $ph$ excitation should be more important, as it is indeed the case. Thus, the $t-$channel $ph$ exchange mechanism leads to a sizeable correction to the tree level $\pi\pi$ scattering amplitude and a small vertex correction in the calculation of the loops appearing in the unitarization procedure\footnote{This mechanism would play an even smaller role in the position of the $\sigma$ pole \cite{VicenteVacas:2002se}, which is determined by the vanishing of the denominator of the BS solution, where the tree level term does not appear.}. Next we consider the box diagram of Fig. \ref{new_mech}(b). This term was found to be smaller in strength than the $ph$ exchange in \cite{Meissner:2001gz}, particularly at small energies, where the $p-$wave character of the vertices made the contribution negligible. The consideration of this mechanism at the pion loop level, necessary to include it in the BS equation, makes its contribution small since, apart from the reduction of the box diagram for large values of $q$, there is a further cancellation of terms as we show in Appendix III. Similar analytic treatments are done in \cite{Herrmann:1993za,Cabrera:2000dx}. For all these reasons this contribution should be even smaller than the one previously evaluated and one can safely neglect it for practical purposes. \begin{figure} \begin{center} \includegraphics[width=0.7\textwidth]{fig_10.eps} \caption{\label{fig:tadeffect}Real and imaginary parts of $T$, as obtained in Eq. (\ref{BStad}), for the two sets of parameters in Table \ref{ci} (set I, dash-dotted line, set II, dashed line) and $\rho=\rho_0/2$. The solid line corresponds to the result of the model in Sec. \ref{sec:nucmed} and the dotted line is the result in vacuum.} \end{center} \end{figure} \section{Results} We solve the BS equation including the corrections discussed in Sec. \ref{sec:tadpoles}, as they appear in Eq. (\ref{BStad}). The results are shown in Fig. \ref{fig:tadeffect}. The new terms considered modify little the results from Ref. \cite{Chiang:1998di}. In comparison, the imaginary part of the $\pi\pi$ amplitude exhibits a small increase of strength at low invariant energies whereas the real part decreases over all the calculated range of energies. Altogether, the basic effect of the nuclear medium, as found in Ref. \cite{Chiang:1998di}, is a strong depletion of the interaction at energies around $500$~MeV, where the vacuum $\sigma$ pole is found, and some accumulation of strength close to the $2\pi$ threshold, as it can be seen in Fig. \ref{fig:tadeffectT2}, where the squared modulus of the amplitude is depicted. \begin{figure} \begin{center} \includegraphics[width=0.7\textwidth]{fig_11.eps} \caption{\label{fig:tadeffectT2}Squared modulus of $T$. Lines as in Fig. \ref{fig:tadeffect}.} \end{center} \end{figure} The contribution of the terms discussed in Sec. \ref{sec:newmech} is shown in Fig. \ref{fig:newmech} for the imaginary part of the $\pi\pi$ amplitude. We find a strong reduction of the amplitude at energies close to the $2\pi$ threshold, basically produced by the repulsive tree level term in Fig. \ref{new_mech}(a). At these energies the amplitude stays closer to the vacuum case. A similar reduction of the nuclear medium effects as compared to the results of Ref. \cite{Chiang:1998di} is found in the real part of the amplitude. \begin{figure} \begin{center} \includegraphics[width=0.7\textwidth]{fig_12.eps} \caption{\label{fig:newmech}Imaginary part of $T$ including the mechanism described in Sec. \ref{sec:newmech} at $\rho=\rho_0/2$ (dashed line). The solid line corresponds to the result of the model in Sec. \ref{sec:nucmed} and the dotted line is the result in vacuum.} \end{center} \end{figure} Finally, we have included together the contributions of the tadpole terms, Section \ref{sec:tadpoles}, and the $t-$channel $ph$ exchange, Section \ref{sec:newmech}, in the BS equation, and the results are depicted in Fig. \ref{fig:all} for the real and imaginary parts of the $\pi\pi$ amplitude. We observe, compared to the model of Ref. \cite{Chiang:1998di} in which the basic medium effect is due to the $p-$wave pion selfenergy, a considerable reduction of strength close to the two pion threshold. The global effect in both calculations is still a sizable depletion of the interaction at higher energies and a certain accumulation of strength below the $\sigma$ pole position in vacuum which, as suggested in \cite{VicenteVacas:2002se}, could be reflecting a change in the $\sigma$ pole position to lower energies as a function of the nuclear density. \begin{figure} \begin{center} \includegraphics[width=0.7\textwidth]{fig_13.eps} \caption{\label{fig:all}Real and imaginary parts of $T$ including the mechanisms described in Secs. \ref{sec:tadpoles} and \ref{sec:newmech} at $\rho=\rho_0/2$, using set I (dashed line). The solid line corresponds to the result of the model in Sec. \ref{sec:nucmed} and the dotted line is the result in vacuum.} \end{center} \end{figure} \section{Conclusions} In summary, we have considered in this work the contribution of some new terms to the $\pi\pi$ interaction in the scalar isoscalar channel at finite densities, starting from a previous work \cite{Chiang:1998di,Oset:2000ev} in which only medium effects associated to the $p-$wave pion selfenergy had been accounted for. Tadpole insertions, sometimes advocated as a possible source of a large attraction, have been shown to affect little the $\pi\pi$ amplitude once the Bethe Salpeter equation is solved. This is partly due to certain cancellations which take place between vertices and internal pion propagator insertions. We have also taken into account new terms in the driving kernel of the Bethe Salpeter equation, which have been found important in a study based on a chiral power counting in the many body problem. We could see that these new terms, although large at tree level, when appearing inside loops were not as important as one could guess from their comparison with the lowest order chiral $\pi \pi$ amplitude in the case that all pions are on shell. As a consequence, their consideration barely changed the results for the $\pi\pi$ interaction in the medium. Altogether, the final results are quite similar to those obtained previously in \cite{Chiang:1998di,Oset:2000ev}, namely a strong reduction of the interaction at energies around 400~MeV and beyond, and some increase of strength around the 2$\pi$ threshold. This confirms the leading role of the strong $p-$wave pion selfenergy in the medium modification of the $\pi\pi$ interaction in the scalar isoscalar channel. These results are also satisfactory because a prediction on the $(\gamma,2\pi)$ reaction in nuclei \cite{Roca:2002vd} based on the previous calculation \cite{Chiang:1998di,Oset:2000ev} of the two-pion final state interaction has been later confirmed by experimental data \cite{Messchendorp:2002au}. The much larger medium effects obtained at threshold energies in other approaches are incompatible with the observed effect in the $(\gamma,2\pi)$ reaction. \section*{Acknowledgements} This work is partly supported by DGICYT contract no. BFM2003-00856. D.~Cabrera acknowledges financial support from MEC. \section*{Appendix I} We quote in this section the Lindhard function, with an energy gap $\Delta$, separated into the direct and crossed contributions, $U=U_d+U_c$. From \cite{Oset:sm} we have \begin{equation} U_d(q^0,\vec{q},\Delta;\rho) = 4 \int \frac{d^3 p}{(2\pi)^3} \frac{n(\vec{p}) \lbrack 1 - n(\vec{p}+\vec{q}) \rbrack} {q^0 + \varepsilon (\vec{p}) - \varepsilon (\vec{p}+\vec{q}) - \Delta + i \epsilon} \label{U_dgap} \end{equation} and $U_c(q^0,\vec{q},\Delta;\rho) \equiv U_d(-q^0,\vec{q},\Delta;\rho)$. In the following we shall use the definitions \begin{eqnarray} x = \frac{q}{k_{F}} \,\,\, , \,\,\, \nu = \frac{2Mq^0}{k_F^2} \nonumber \\ \delta = \frac{2M\Delta}{k_F^2} \,\,\, , \,\,\, \rho=\frac{2}{3\pi^2}k_F^3 \label{defs} \end{eqnarray} with $M$ the mass of the nucleon, $k_F$ the Fermi momentum and $q \equiv |\vec{q}|$. Once the integration in Eq. (\ref{U_dgap}) is done, the real part of $U_d$ reads, for $x \leq 2$, \begin{eqnarray} \textrm{Re} \, U_d(q^0,\vec{q},\Delta;\rho) = - \frac{2Mk_F}{\pi^2} \frac{1}{2x} \bigg \lbrace \frac{x}{2} - \frac{\nu - \delta}{4} + \frac{\nu - \delta}{2} \ln \bigg | \frac{\nu-\delta+x^2-2x}{\nu-\delta}\bigg | \nonumber \\ + \frac{1}{2}\left[ 1-\frac{1}{4} \left( \frac{\nu-\delta}{x}-x \right) ^2 \right] \ln \bigg | \frac{\nu-\delta-x^2-2x}{\nu-\delta+x^2-2x} \bigg | \bigg \rbrace \label{Reless2} \end{eqnarray} and, for $x > 2$, \begin{eqnarray} \textrm{Re} \, U_d(q^0,\vec{q},\Delta;\rho) = - \frac{2Mk_F}{\pi^2}\frac{1}{2x} \bigg \lbrace \frac{-\nu+\delta+x^2}{2x} \nonumber \\ + \frac{1}{2}\left[ 1 - \frac{1}{4} \left( \frac{\nu-\delta}{x}-x\right)^2 \right] \ln \bigg | \frac{\nu-\delta-x^2-2x}{\nu-\delta-x^2+2x} \bigg | \bigg \rbrace \,\,\, . \label{Regt2} \end{eqnarray} The imaginary part of $U_d$ is given by Im $U_d(q^0,\vec{q},\Delta;\rho)=$ Im $\tilde{U}(q^0-\Delta,\vec{q};\rho) \Theta(q^0-\Delta)$, with \begin{eqnarray} \textrm{Im}\, \tilde{U}(q^0,\vec{q};\rho)= -\frac{3}{4} \pi \rho \frac{M}{q k_F} \lbrack (1-z^2)\Theta(1-|z|) - (1-z'\,^2)\Theta(1-|z'|) \rbrack \frac{q^0}{|q^0|} \,\,\, , \label{Utilda} \end{eqnarray} where $\Theta$ is the Heaviside step function and the $z$, $z'$ variables are defined as \begin{equation} z=\frac{M}{q k_F}\left[ q^0-\frac{q^2}{2M}\right] \,\,\, , \,\,\, z'=\frac{M}{q k_F}\left[ -q^0-\frac{q^2}{2M}\right] \,\,\, . \end{equation} \section*{Appendix II} \begin{figure} \begin{center} \includegraphics[width=6cm]{fig_14.eps} \caption{\label{loop2}Loop contribution of the $ph$ exchange in the $t$ channel.} \end{center} \end{figure} The amplitude corresponding to the diagram in Fig. \ref{loop2} is given by \begin{equation} -i T = \int \frac{d^4q}{(2\pi)^4} (-i \tilde{t}) \frac{i}{(q+p)^2 - m_{\pi}^2 + i\epsilon} \frac{i}{(q-p')^2 - m_{\pi}^2 + i\epsilon} (-i V(s)) \, i U(q) \,\,\, . \label{new_T} \end{equation} In order to perform the integral it is most useful to separate $U(q)$ into the direct and crossed parts, $U(q)=U_d(q)+U_c(q)$, given their different analytical structure. \begin{figure} \begin{center} \includegraphics[width=13cm]{fig_15.eps} \caption{\label{analyt}Analytical structure of the integrand in Eq. (\ref{new_T}). The poles of the pion propagators are represented by '$x$' and '$o$' symbols, and the dotted lines correspond to the analytical cuts of the Lindhard function. The arrows indicate the circuit used for the integration of each term.} \end{center} \end{figure} In Fig. \ref{analyt} we depict the pole and cut structure for the different terms and the path followed for the integration in the complex plane. The poles are located at \begin{eqnarray} q^0 = -p^0 + \omega(\vec{p}+\vec{q}) - i\epsilon \,\,\, , \,\,\, q^0 = -p^0 - \omega(\vec{p}+\vec{q}) + i\epsilon \nonumber \\ q^0 = p^0 + \omega(\vec{p}+\vec{q}) - i\epsilon \,\,\, , \,\,\, q^0 = p^0 - \omega(\vec{p}+\vec{q}) + i\epsilon \,\,\, . \label{poles} \end{eqnarray} The integration over the $q^0$ variable is done by closing the contour in the complex plane in the upper half plane for the $U_d$ part and in the lower half plane for the $U_c$ part. The result of the integration is \begin{eqnarray} T = - \left( \frac{1}{4f^2}\right)^2 (2p^0)^2 V(s) \int \frac{d^3q}{(2\pi)^3} \frac{1}{4\omega^2} \bigg \lbrace \frac{U_c(p^0+\omega,\vec{q})}{p^0+\omega} \nonumber \\ -\frac{U_d(p^0-\omega,\vec{q})}{p^0-\omega+i\epsilon} + \frac{U_d(p^0-\omega,\vec{q})-U_c(p^0+\omega,\vec{q})}{p^0} \bigg \rbrace \bigg ( \frac{M_{\rho}^2}{M_{\rho}^2+\vec{q}\,^2} \bigg ) ^2 \,\,\, , \end{eqnarray} where $\omega \equiv \omega(\vec{p}+\vec{q})$ and we have explicitly written the $\rho$ meson exchange form factor arising from each $\pi\pi NN$ vertex. Let us note that we have factorized the $\pi N \to \pi N$ vertex on shell. This is done in analogy to what is done in \cite{Oset:1997it} where one shows that the off shell part can be cast into a renormalization of the lowest order diagram (no meson loop in this case). An alternative justification using dispersion relations, which require only the on shell information, is given in \cite{Oller:2000fj}. \section*{Appendix III} \begin{figure} \begin{center} \includegraphics[width=5cm]{fig_16.eps} \caption{\label{box}Box diagram with two of the pions as a part of a loop.} \end{center} \end{figure} We evaluate the loop function of Fig. \ref{box} containing the box diagram of Fig. \ref{new_mech}(b) plus all the different time orderings, which we can see in Fig. \ref{box_orderings}. In all the diagrams the internal nucleon lines are particle lines. This means we are taking only the terms of order $\rho$, which are obtained when the external lines are folded to give a single hole line in Fig. \ref{box_orderings}. \begin{figure} \begin{center} \includegraphics[width=7cm]{fig_17.eps} \caption{\label{box_orderings}Set of different time orderings of diagram in Fig. \ref{box}. The initial and final nucleon lines correspond to a hole propagator.} \end{center} \end{figure} {\bf Diagrams (a), (b)}. For diagrams (a), (b) in Fig. \ref{box_orderings} for $\vec{P}=\vec{p}_1+\vec{p}_2=\vec{0}$ the intermediate nucleon line after the two pion vertices has the same momentum as the hole line (belonging to the Fermi sea) and hence they both vanish. Next we observe some strong cancellations in the other diagrams. The set of the two meson propagators, which is common to all of them, can be written as \begin{equation} \frac{1}{2 P^0 \omega} \bigg \lbrace \frac{1}{q^0-\omega+i\epsilon} \; \frac{1}{q^0+P^0+\omega-i\epsilon} - \frac{1}{q^0+\omega-i\epsilon} \; \frac{1}{q^0+P^0-\omega+i\epsilon} \bigg \rbrace \,\,\, , \label{props} \end{equation} with $\omega=\omega(\vec{q})$. {\bf Diagram (c)}. The diagram (c) contains three nucleon propagators. By making a heavy baryon approximation and neglecting the kinetic energy of the nucleons we find for the product \begin{equation} \frac{1}{-p_1^0} \; \frac{1}{q^0+p_2^0+i\epsilon} \; \frac{1}{q^0+i\epsilon} \,\,\, . \label{approx_c} \end{equation} Hence, multiplying this by the pion propagators and closing the contour on the upper half plane to perform the $q^0$ integration, we find that the integral is \begin{equation} A \frac{1}{p_1^0} \bigg \lbrace -\frac{1}{p_1^0+\omega}\; \frac{1}{P^0+\omega} + \frac{1}{\omega} \; \frac{1}{p_2^0-\omega+i\epsilon} \bigg \rbrace \,\,\ . \label{integ_c} \end{equation} There, the first term, which comes from the negative energy components of the mesons, is small and has no imaginary part. The second term can lead to an imaginary part and a more sizeable real part from the principal value. {\bf Diagram (e)}. The set of nucleon propagators in the heavy baryon approximation is now \begin{equation} \frac{1}{p_2^0} \; \frac{1}{p_2^0+q^0+i\epsilon} \; \frac{1}{-p_1^0} \label{approx_e} \end{equation} and hence by closing the contour in the upper half of the complex $q^0$ plane we find for the $q^0$ integration \begin{equation} A \frac{1}{p_1^0}\bigg \lbrace \frac{1}{p_2^0} \; \frac{1}{p_1^0+\omega} - \frac{1}{p_2^0} \; \frac{1}{p_2^0-\omega+i\epsilon} \bigg \rbrace \,\,\, , \label{integ_e} \end{equation} with the same $A$ as in Eq. (\ref{integ_c}). The first term is again small, coming from the negative energy components of the pions, and has opposite sign to the first term from diagram (c). The second term above is the same but with opposite sign to the second term of diagram (c) at $\omega=p_2^0$, which is the singular point. Hence there are strong cancellations in the principal part of the integral and the imaginary part from this source vanishes. {\bf Diagram (d)}. Repeating the same arguments as above we find now \begin{equation} -A \frac{1}{\omega}\bigg \lbrace \frac{1}{p_1^0+\omega} \; \frac{1}{P^0+\omega} + \frac{1}{P^0-\omega+i\epsilon} \; \frac{1}{p_2^0-\omega+i\epsilon} \bigg \rbrace \,\,\, . \label{integ_d} \end{equation} {\bf Diagram (f)}. For this diagram we find \begin{equation} A \frac{1}{p_1^0}\bigg \lbrace \frac{1}{\omega} \; \frac{1}{p_2^0+\omega} - \frac{1}{P^0-\omega+i\epsilon} \; \frac{1}{p_1^0-\omega+i\epsilon} \bigg \rbrace \,\,\, . \label{integ_f} \end{equation} Once again the first two terms from (d), (f), coming from the negative energy part of the pion propagators, give a small contribution and partly cancel, and the second terms which provide an imaginary part and a larger real part from the principal value, also show cancellations. Indeed for $p_1^0=p_2^0=\omega$ the imaginary parts corresponding to the poles $p_1^0=p_2^0=\omega$ cancel and the real parts from the principal value would also largely cancel. At the $P^0=\omega$ pole the cancellation would only be partial. We thus see that when considering all the time orderings for the coupling of the two pions and the loop with the two pion propagators there are large cancellations of terms. In addition we have the $(p_i / M)^2$ factor of the $p-$wave couplings for the initial pions, which make this contribution small at small momenta of the pions. We have looked at strong cancellations of terms in the heavy baryon approximation, which holds for small values of momenta. At large momenta we must note that we have two extra nucleon propagators which bring two extra powers of $q$ in the denominator, with respect to the ordinary Lindhard function. This makes up for the two extra $p-$wave vertices, and hence we have a similar behaviour altogether as the one of Fig. \ref{loop2} which lead to small contributions when evaluated into the loop. All these elements discussed above would render this piece far smaller than the ones of Fig. \ref{oneloopBS}(d,e) and, given the smallness of the effects found there, this can also be neglected.
{ "timestamp": "2005-03-04T17:55:35", "yymm": "0503", "arxiv_id": "nucl-th/0503014", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503014" }
\section{Introduction} \newtheorem{Definition}{Definition} \newtheorem{Lemma}{Lemma} \newtheorem{Theorem}{Theorem} \newtheorem{Proposition}{Proposition} \newtheorem{Corollary}{Corollary} Let $\Omega$ be a bounded homogeneous domain in $\mbox{\Bbb C}^{n}.$ The class of all holomorphic functions with domain $\Omega$ will be denoted by $H(\Omega).$ Let $\phi $ be a holomorphic self-map of $\Omega,$ the composition operator $C_{\phi}$ induced by $\phi$ is defined by $$(C_{\phi}f)(z)=f(\phi(z)),$$ for $z$ in $\Omega$ and $f\in H(\Omega)$. Let $K(z,z)$ be the Bergman kernel function of $\Omega$, the Bergman metric $H_{z}(u,u)$ in $\Omega$ is defined by $$H_{z}(u,u)=\displaystyle\frac{1}{2} \sum\limits^{n}_{j,k=1} \displaystyle\frac{\partial^{2}\log K(z,z)}{\partial z_{j} \partial {\overline {z}}_{k}}u_{j}{\overline u}_{k},$$ where $z\in\Omega$ and $u=(u_{1},\ldots,u_{n})\in \mbox{\Bbb C}^{n}.$ Following Timoney [1], we say that $f\in H(\Omega)$ is in the Bloch space ${\cal B}(\Omega),$ if $$\|f\|_{{\cal B}(\Omega)}=\sup\limits_{z\in \Omega}Q_{f}(z)<\infty,$$ where \begin{equation}Q_{f}(z)=\sup\left\{\displaystyle\frac {|\bigtriangledown f(z)u|}{H^{\frac{1}{2}}_{z}(u,u)}: u\in \mbox{\Bbb C}^{n}-\{0\}\right\},\label{1}\end{equation} and $\bigtriangledown f(z) =\left(\frac{\partial f(z)}{\partial z_{1}}, \ldots, \frac{\partial f(z)}{\partial z_{n}} \right), \bigtriangledown f(z)u =\sum\limits^{n}_{l=1}\frac{\partial f(z)}{\partial z_{l}}u_{l}.$ The little Bloch space ${\cal B}_0(\Omega)$ is the closure in the Banach space ${\cal B}(\Omega)$ of the polynomial functions. Let $\partial\Omega$ denote the boundary of $\Omega$. Following Timoney [2], for $\Omega=B_n$ the unit ball of $\mbox{\Bbb C}^n$, ${\cal B}_0(B_n)=\left\{f\in{\cal B}(B_n): Q_f(z)\to 0, \mbox{as}\hspace*{2mm}z\to\partial B_n \right\};$ for $\Omega=\cal D$ the bounded symmetric domain other than the ball $B_n$, $\left\{ f\in{\cal B}({\cal D}): Q_f(z)\to 0, \mbox{as}\hspace*{2mm} z\to\partial{\cal D}\right\}$ is the set of constant functions on $\cal D.$ So if $\cal D$ is a bounded symmetric domain other than the ball, we denote the ${\cal B}_{0*}({\cal D})= \left\{f\in{\cal B}({\cal D}): Q_f(z)\to 0, \mbox{as}\hspace*{2mm} z\to\partial^*{\cal D}\right\}$ and call it little star Bloch space, here $\partial^*{\cal D}$ means the distinguished boundary of $\cal D$. The unit ball is the only bounded symmetric domain $\cal D$ with the property that $\partial^*{\cal D}=\partial{\cal D}.$ Let $U^n$ be the unit polydisc of $\mbox{\Bbb C}^n$. Timoney [1] shows that $f\in{\cal B}(U^n)$ if and only if $$\|f\|_1=|f(0)|+\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\frac{\partial f} {\partial z_k}(z)\right|(1- |z_k|^2)<+\infty,$$ where $f\in H(U^n).$ This definition was the starting point for introducing the $p$-Bloch spaces. Let $p>0,$ a function $f\in H(U^n)$ is said to belong to the $p$-Bloch space ${\cal B}^p(U^n)$ if $$\|f\|_p=|f(0)|+\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\frac{\partial f} {\partial z_k}(z)\right|\left(1- |z_k|^2\right)^p<+\infty.$$ It is easy to show that ${\cal B}^p(U^n)$ is a Banach space with the norm $\|\cdot\|_p.$ Just like Timoney [2], if $$\lim_{z\to\partial U^n}\sum\limits^n_{k=1}\left|\frac{\partial f}{\partial z_k}(z)\right|(1- |z_k|^2)^p=0,$$ it is easy to show that $f$ must be a constant. Indeed, for fixed $z_1\in U,$ $\displaystyle\frac{\partial f}{\partial z_1}(z)(1-|z_1|^2)^p$ is a holomorphic function in $z'=(z_2,\cdots,z_n)\in U^{n-1}$. If $z\to\partial U^n$, then $z'\to\partial U^{n-1},$ which implies that $$\lim\limits_{z'\to\partial U^{n-1}}\left|\frac{\partial f} {\partial z_1}(z)\right|\left(1- |z_1|^2\right)^p=0.$$ Hence, $\frac{\partial f} {\partial z_1}(z)\left(1- |z_1|^2\right)^p\equiv 0$ for every $z'\in \partial U^{n-1},$ and for each $z_1\in U,$ and consequently $\frac{\partial f} {\partial z_1}(z)=0$ for every $z\in U^n.$ Similarly, we can obtain that $\frac{\partial f} {\partial z_j}(z)=0$ for every $z_j\in U^n$ and each $j\in\{2,\cdots,n\},$ therefore $f\equiv const .$ So, there is no sense to introduce the corresponding little $p$-Bolch space in this way. We will say that the little $p$-Bolch space ${\cal B}_0^p(U^n)$ is the closure of the polynomials in the $p$-Bolch space. If $f\in H(U^n)$ and $$\sup\limits_{z\in \partial^*U^n}\sum\limits^n_{k=1} \left|\displaystyle\frac{\partial f} {\partial z_k}(z)\right|\left(1- |z_k|^2\right)^p=0,$$ we say $f$ belongs to little star $p$-Bolch space ${\cal B}_{0*}^p(U^n).$ Using the same methods as that of Theorem 4.14 in reference [2], we can show that ${\cal B}^p_{0}(U^n)$ is a proper subspace of ${\cal B}^p_{0*}(U^n)$ and ${\cal B}^p_{0*}(U^n)$ is a non-separable closed subspace of ${\cal B}^p(U^n).$ Let $\phi $ be a holomorphic self-map of $U^n,$ the composition operator $C_{\phi}$ induced by $\phi$ is defined by $(C_{\phi}f)(z)=f(\phi(z))$ for $z$ in $U^n$ and $f\in H(U^n)$. For the unit disc $U\subset\mbox{\Bbb C},$ Madigan and Matheson [3] proved that $C_{\phi}$ is always bounded on ${\cal B}(U)$ and bounded on ${\cal B}_0(U)$ if and only if $\phi\in{\cal B}_0(U).$ They also gave the sufficient and necessary conditions that $C_{\phi}$ is compact on ${\cal B}(U)$ or ${\cal B}_0(U).$ More recently, [4,5,7] gave some sufficient and necessary conditions for $C_{\phi}$ to be compact on the Bloch spaces in polydisc. We recall that the essential norm of a continuous linear operator $T$ is the distance from $T$ to the compact operators, that is, \begin{equation}\|T\|_e=\inf\{\|T-K\|: K \mbox{ is compact}\}.\label{2}\end{equation} Notice that $\|T\|_e=0$ if and only if $T$ is compact, so that estimates on $\|T\|_e$ lead to conditions for $T$ to be compact. In this paper, we give some estimates of the essential norms of bounded composition operators $C_{\phi}$ between ${\cal B}^p(U^n)$ (${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$ ) and ${\cal B}^q(U^n)$ (${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$). As their consequences, some necessary and sufficient conditions for the bounded composition operators $C_{\phi}$ to be compact from ${\cal B}^p(U^n)$ (${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$ ) into ${\cal B}^q(U^n)$ (${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) are obtained. The fundamental ideals of the proof are those used by J. H. Shpairo [8] to obtain the essential norm of a composition operator on Hilbert spaces of analytic functions (Hardy and weighted Bergman spaces) in terms of natural counting functions associated with $\phi$. This paper generalizes the result on the Bloch space in [10] to the Bloch-type space in polydisk. Throughout the remainder of this paper $C$ will denote a positive constant, the exact value of which will vary from one appearance to the next. Our main results are the following: \begin{Theorem} Let $\phi=(\phi_1, \phi_2, \cdots, \phi_n)$ be a holomorphic self-map of $U^n$ and $\|C_{\phi}\|_e$ the essential norm of a bounded composition operator $C_{\phi}:$ ${\cal B}^p(U^n)$ ( ${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow$ $ {\cal B}^q(U^n)$ ( ${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) , then \begin{eqnarray}&&\displaystyle\frac{1}{n}\lim\limits_{\delta\to 0} \sup\limits_{dist(\phi(z),\partial U^n)<\delta}\sum\limits^n_{k,l=1} \left|\displaystyle\frac{\partial \phi_{l}}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}\nonumber\\ &&\leq\|C_{\phi}\|_e \leq 2\lim\limits_{\delta\to 0} \sup\limits_{dist(\phi(z),\partial U^n)<\delta}\sum\limits^n_{k,l=1} \left|\displaystyle\frac{\partial \phi_{l}}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}.\label{3}\end{eqnarray} \end{Theorem} By Theorem 1 and the fact that $C_{\phi}:$ ${\cal B}^p(U^n)$ (or ${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow$ $ {\cal B}^q(U^n)$ (or ${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) is compact if and only if $\|C_{\phi}\|_e=0$, we obtain Theorem 2 at once. \begin{Theorem}\hspace{2mm}Let $\phi=(\phi_1, \ldots, \phi_n)$ be a holomorphic self-map of $U^{n}.$ Then the bounded composition operator $C_{\phi}:$ ${\cal B}^p(U^n)$ (${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow$ $ {\cal B}^q(U^n)$ (${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) is compact if and only if for any $\varepsilon>0,$ there exists a $\delta$ with $0<\delta<1,$ such that \begin{equation}\sup\limits_{dist(\phi(z),\partial U^n)<\delta} \sum\limits^n_{k,l=1} \left|\displaystyle\frac{\partial \phi_{l}}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}<\varepsilon.\label{4}\end{equation} \end{Theorem} When $n=1,$ on ${\cal B}(U)$ we obtain Theorem 2 in [3]. Since $\partial U=\partial^* U,$ ${\cal B}_0(U)={\cal B}_{0*}(U),$ we can also obtain Theorem 1 in [3]. By Theorem 2 and Lemmas 3, 4 and 5 in next part, we can get the following three Corollaries. \begin{Corollary} Let $\phi=(\phi_1, \ldots, \phi_n)$ be a holomorphic self-map of $U^{n}.$ Then\\ $C_{\phi}:{\cal B}^p(U^{n})$(${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow{\cal B}^q(U^{n})$ is compact if and only if $$\sum\limits^n_{k,l=1} \left|\displaystyle\frac{\partial \phi_{l}}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}\leq C$$ for all $z\in U^n$ and (\ref{4}) holds.\end{Corollary} {\bf Proof}\hspace*{4mm} By Lemma 3 in next part, we know $C_{\phi}:{\cal B}^p(U^{n})$(${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow{\cal B}^q(U^{n})$ is bounded. It follows from Theorem 2 that $C_{\phi}:{\cal B}^p(U^{n})$(${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow{\cal B}^q(U^{n})$ is compact. Conversely, if $C_{\phi}:{\cal B}^p(U^{n})$(${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow{\cal B}^q(U^{n})$ is compact, it is clear that $C_{\phi}:{\cal B}^p(U^{n})$(${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow{\cal B}^q(U^{n})$ is bounded, by Theorem 2, (\ref{4}) holds. \begin{Corollary} Let $\phi=(\phi_1, \ldots, \phi_n)$ be a holomorphic self-map of $U^{n}.$ Then \\ $C_{\phi}:$${\cal B}^p_{0*}(U^{n})$(${\cal B}^p_{0}(U^n)$)$\rightarrow{\cal B}^q_{0*}(U^{n})$ is compact if and only if $\phi_l\in {\cal B}^q_{0*}(U^n)$ for every $l=1,2,\cdots, n$ and (\ref{4}) holds.\end{Corollary} {\bf Proof}\hspace*{4mm} Note that Lemma 4 in next part, similar to the proof of Corollary 1, the Corollary follows. \begin{Corollary} Let $\phi=(\phi_1, \ldots, \phi_n)$ be a holomorphic self-map of $U^{n}.$ Then \\ $C_{\phi}:$${\cal B}^p_{0}(U^{n})\rightarrow{\cal B}^q_{0}(U^{n})$ is compact if and only if $\phi_l\in {\cal B}^q_{0}(U^n)$ for every $l=1,2,\cdots, n$ and (\ref{4}) holds.\end{Corollary} {\bf Proof}\hspace*{4mm} Note that Lemma 5 in next part, similar to the proof of Corollary 1, the Corollary follows. \section{Some Lemmas} In order to prove Theorem 1, we need some Lemmas. \begin{Lemma} Let $f\in{\cal B}^p(U^n),$ then (1) If $0\leq p<1,$ then $\|f(z)|\leq |f(0)|+\displaystyle\frac{n}{1-p}\|f\|_p;$ (2) If $p=1,$ then $|f(z)|\leq \left(1+\displaystyle\frac{1}{n\ln 2}\right)\left(\sum\limits^n_{k=1}\ln \displaystyle\frac{2}{1-|z_k|^2}\right)\|f\|_p.$ (3) If $p>1,$ then $|f(z)|\leq \left(\displaystyle\frac{1}{n}+\displaystyle\frac{2^{p-1}}{p-1}\right) \sum\limits^n_{k=1}\displaystyle\frac{1}{(1-|z_k|^2)^{p-1}}\|f\|_p.$ \end{Lemma} {\bf Proof}\hspace{2mm} This Lemma can be proved by some integral estimates (if necessary, the proof can be omitted). By the definition of $\|.\|_{p}$, $$|f(0)|\leq \|f\|_{p},\hspace*{4mm}\left|\displaystyle\frac{\partial f(z) }{\partial z_l}\right|\leq \displaystyle\frac{\|f\|_{p}}{(1-|z_l|^2)^p} \hspace*{4mm}(l\in\{1,2,\cdots,n\})$$ and \begin{eqnarray*}&&f(z)-f(0)= \int^1_0 \displaystyle\frac{d f(tz)}{d t}dt=\sum\limits^n_{l=1}\int^1_0 z_l\displaystyle\frac{\partial f}{\partial\zeta_l}(tz)dt,\end{eqnarray*} So\begin{eqnarray}&&|f(z)|\leq |f(0)|+\sum\limits^n_{l=1}|z_l|\int^1_0\displaystyle\frac{\|f\|_p}{\left(1-t^2|z_l|^2\right)^p}dt \nonumber\\ &&\leq \|f\|_p+\|f\|_p\sum\limits^n_{l=1}\int^{|z_l|}_0\displaystyle\frac{1}{\left(1-t^2\right)^p}dt.\label{5} \end{eqnarray} If $p=1,$ \begin{equation}\int^{|z_l|}_0\displaystyle\frac{1}{\left(1-t^2\right)^p}dt=\displaystyle\frac{1}{2} \ln\displaystyle\frac{1+|z_l|}{1-|z_l|}\leq \displaystyle\frac{1}{2} \ln\displaystyle\frac{4}{1-|z_l|^2}.\label{6}\end{equation} It is clear that $\ln\displaystyle\frac{4}{1-|z_l|^2}>\ln 4=2\ln 2,$ so\begin{equation}1\leq \displaystyle\frac{1}{2\ln 2} \ln\displaystyle\frac{4}{1-|z_l|^2}\leq \displaystyle\frac{1}{2n\ln 2} \sum\limits^n_{l=1}\ln\displaystyle\frac{4}{1-|z_l|^2}.\label{7}\end{equation} Combining (\ref{5}),(\ref{6}) and (\ref{7}), we get $$|f(z)|\leq \left(\displaystyle\frac{1}{2}+\displaystyle\frac{1}{2n\ln 2}\right)\left(\sum\limits^n_{l=1}\ln\displaystyle\frac{4}{1-|z_l|^2}\right)\|f\|_p.$$ If $p\neq 1,$ \begin{eqnarray}&&\int^{|z_l|}_0\displaystyle\frac{1}{\left(1-t^2\right)^p}dt =\int^{|z_l|}_0\displaystyle\frac{1}{(1-t)^p}\cdot \displaystyle\frac{1}{(1+t)^p}dt\nonumber\\ &&\leq \int^{|z_l|}_0\displaystyle\frac{1}{(1-t)^p}dt=\displaystyle\frac{1-(1-|z_l|)^{-p+1}}{1-p}.\label{8} \end{eqnarray} If $0<p<1,$ (\ref{8}) gives that $\int^{|z_l|}_0\displaystyle\frac{1}{\left(1-t^2\right)^p}dt\leq\displaystyle\frac{1}{1-p},$ it follows from (\ref{5}) that $|f(z)|\leq\left(1+\displaystyle\frac{n}{1-p}\right)\|f\|_p.$ If $p>1,$ (\ref{8}) gives that $$\int^{|z_l|}_0\displaystyle\frac{1}{\left(1-t^2\right)^p}dt\leq \displaystyle\frac{1-(1-|z_l|^{p-1})}{(p-1)(1-|z_l|)^{p-1}} \leq\displaystyle\frac{2^{p-1}}{(p-1)(1-|z_l^2|)^{p-1}},$$ it follows from (\ref{5}) that \begin{eqnarray*}|f(z)|&\leq&\|f\|_p+\displaystyle\frac{2^{p-1}}{p-1}\left(\sum\limits^n_{l=1} \displaystyle\frac{1}{(1-|z_l|^2)^{p-1}}\right)\|f\|_p\\ &\leq& \left(\displaystyle\frac{1}{n}+\displaystyle\frac{2^{p-1}}{p-1}\right)\left(\sum\limits^n_{l=1} \displaystyle\frac{1}{(1-|z_l|^2)^{p-1}}\right) \|f\|_{p}.\end{eqnarray*}Now the Lemma is proved. \begin{Lemma} Set $$f_w(z)=\int_0^{z_l}\frac {dt}{(1-\bar wt)^p},$$ where $w\in U.$ Then $f\in {\cal B}^p_0(U^n)\subset{\cal B}^p_{0*}(U^n)\subset {\cal B}^p(U^n).$ \end{Lemma} {\bf Proof}\hspace{2mm}Since $$\displaystyle\frac{\partial f_w}{\partial z_l} =\left(1-\overline wt\right)^{-p}, \hspace*{4mm} \displaystyle\frac{\partial f_w}{\partial z_i}=0, \hspace*{4mm}(i\neq l),$$ it follows that $$|f(0)|+\sum\limits^n_{k=1}\left|\displaystyle\frac{\partial f_w}{\partial z_k}(z)\right|(1-|z_k|^2)^p =\displaystyle\frac{(1-|z_l|^2)^p}{|1-\overline wt|^p}\leq(1+|z_l|^2)^p\leq 2^p.$$ Hence $f_w\in {\cal B}^p(U^n).$ Now we prove that $f_w\in{\cal B}^p_0(U^n).$ Using the asymptotic formula $$(1-\bar w t)^{-p}=\sum\limits^{+\infty}_{k=0}\frac{p(p+1)\cdots (p+k-1)}{k!}(\bar w)^kt^k,$$ we obtain $$f_w(z)=\sum\limits^{+\infty}_{k=0}\frac{p(p+1)\cdots (p+k-1)}{k!}(\bar w)^k\int^{z_l}_0 t^kdt.$$ Denote $P_n(z)=\sum\limits^{n}_{k=0}\frac{p(p+1)\cdots (p+k-1)}{k!}(\bar w)^k\int^{z_l}_0t^kdt,$ it is easy to see that $$f_w(z)-P_n(z)=\sum\limits^{+\infty}_{k=n+1}\frac{p(p+1)\cdots (p+k-1)}{k!}(\bar w)^k\int^{z_l}_0t^k dt,$$ $$\left|\frac{\partial (f_w-P_n)}{\partial z_l}\right|\leq \sum\limits^{+\infty}_{k=n+1}\frac{p(p+1)\cdots (p+k-1)}{k!}|w|^k \to 0, \mbox{as}\ \ n\to\infty,$$ \begin{eqnarray*}\|f_w-P_n\|_p&=&|f_w(0)-P_n(0)|+\sup\limits_{z\in U^n}\left|\frac{\partial (f_w-P_n)}{\partial z_l}\right|(1-|z_l|^2)^p \\ &\leq& \sup\limits_{z\in U^n}\left|\frac{\partial (f_w-P_n)}{\partial z_l}\right|\to 0, \end{eqnarray*} it shows that $f_w\in {\cal B}^p_0(U^n).$ So $f\in {\cal B}^p_0(U^n)\subset{\cal B}^p_{0*}(U^n)\subset {\cal B}^p(U^n).$ \begin{Lemma} Let $\phi=(\phi_1, \ldots, \phi_n)$ be a holomorphic self-map of $U^n$, $p,q>0.$ Then $C_{\phi}: {\cal B}^p(U^n) ({\cal B}^p_0(U^n)$ or ${\cal B}^p_{0*}(U^n))\longrightarrow {\cal B}^q(U^n) $ is bounded if and only if there exists a constant $C$ such that \begin{equation}\sum\limits^n_{k,l=1}\left|\displaystyle\frac{\partial \phi_{l}} {\partial z_k}(z)\right|\displaystyle\frac{\left(1-|z_k|^2\right)^q} {\left(1-|\phi_l(z)|^2\right)^p}\leq C ,\label{9}\end{equation}for all $z\in U^n.$\end{Lemma} {\bf Proof}\hspace*{4mm}First assume that condition (\ref{9}) holds. Let $f\in{\cal B}^p(U^n)({\cal B}^p_0(U^n)$ or ${\cal B}^p_{0*}(U^n)),$ by Lemma 1, we know the evaluation at $\phi(0)$ is a bounded linear functional on ${\cal B}^p(U^n),$ so $|f(\phi(0))|\leq C\|f\|_p.$ On the other hand we have \begin{eqnarray}&&\sum\limits^n_{k=1}\left|\frac{\partial \left(C_{\phi}f(z)\right)} {\partial z_k}\right| (1-|z_k|^2)^q =\sum\limits^n_{k=1}\left|\sum\limits^n_{l=1}\frac{\partial f}{\partial \phi_l}(\phi(z)) \frac{\partial \phi_l}{\partial z_k}(z)\right|(1-|z_k|^2)^q\nonumber\\ &&\leq \sum\limits^n_{k, l=1}\left|\frac{\partial f}{\partial \phi_l}(\phi(z))\frac{\partial \phi_l}{\partial z_k}(z)\right|(1-|z_k|^2)^q\nonumber\\ &&\leq \sum\limits^n_{l=1}\left|\frac{\partial f}{\partial \phi_l}(\phi(z))\right|\left(1-|\phi_l(z)|^2\right)^p \sum\limits^n_{k,l=1}\left|\frac{\partial \phi_{l}} {\partial z_k}(z)\right|\frac{\left(1-|z_k|^2\right)^q} {\left(1-|\phi_l(z)|^2\right)^p}\label{10}\\ &&\leq\|f\|_p\sum\limits^n_{k,l=1}\left|\frac{\partial \phi_l} {\partial z_k}(z)\right| \frac{\left(1-|z_k|^2\right)^q}{\left(1-|\phi_l(z)|^2\right)^p}.\label{11} \end{eqnarray} From (\ref{11}) it follows that $$\sum\limits^n_{k=1}\left|\frac{\partial \left(C_{\phi}f(z)\right)} {\partial z_k}\right| (1-|z_k|^2)^q\leq C\|f\|_p.$$ So $C_{\phi}: {\cal B}^p(U^n)\to {\cal B}^q(U^n) $ is bounded. For the converse, assume that $C_{\phi}: {\cal B}^p(U^n) ({\cal B}^p_0(U^n)$ or ${\cal B}^p_{0*}(U^n))\longrightarrow {\cal B}^q(U^n) $ is bounded, with \begin{equation}\|C_{\phi}f\|_q\leq C\|f\|_p\label{12}\end{equation} for all $f\in{\cal B}^p(U^n)({\cal B}^p_0(U^n)$ or ${\cal B}^p_{0*}(U^n)).$ For fixed $l (1\leq l\leq n),$ we will make use of a family of test functions $\{f_{w}: w\in\mbox{\Bbb C}, |w|<1\}$ in ${\cal B}(U^n)$ defined as follows: If $p>0$, let $$f_w(z)=\int^{z_l}_0\left(1-\overline wz_l\right)^{-p}dt.$$ It follows from Lemma 2 that $$f_w\in {\cal B}^p_0(U^n)\subset{\cal B}^p_{0*}(U^n)\subset {\cal B}^p(U^n).$$ For $z\in U^n,$ it follows from (\ref{12}) that \begin{equation}\sum\limits^n_{k=1}\left|\sum\limits^n_{l=1}\displaystyle\frac{\partial f_w(\phi(z))}{\partial \phi_l} \displaystyle\frac{\partial \phi_l}{\partial z_k}(z)\right|(1-|z_k|^2)^q\leq C.\label{13}\end{equation} Let $w=\phi_l(z),$ then $$\sum\limits^n_{k=1}\left|\displaystyle\frac{\partial \phi_{l}} {\partial z_k}(z)\right| \displaystyle\frac{\left(1-|z_k|^2\right)^q}{\left(1-|\phi_l(z)|^2\right)^p} \leq C.$$ Now the proof of Lemma 3 is completed. \begin{Lemma} Let $\phi=(\phi_1, \phi_2, \cdots, \phi_n)$ be a holomorphic self-map of $U^n.$ Then \\ $C_{\phi}:{\cal B}^p_{0*}(U^{n}) ({\cal B}^p_0(U^n)) \rightarrow{\cal B}^q_{0*}(U^{n})$ is bounded if and only if $\phi_l\in{\cal B}^q_{0*}(U^n)$ for every $l=1,2,\cdots, n$ and (\ref{9}) holds.\end{Lemma} {\bf Proof}\hspace*{4mm}If $C_{\phi}:{\cal B}^p_{0*}(U^{n}) ({\cal B}^p_0(U^n)) \rightarrow {\cal B}^q_{0*}(U^{n})$ is bounded , it is clear that, for every $l=1,2,\cdots, n$, $f_l(z)=z_l\in{\cal B}^p_0(U^n)\subset{\cal B}^q_{0*}(U^n),$ so $C_{\phi}f_l=\phi_l\in{\cal B}^q_{0*}(U^n).$ In the proof of Lemma 3, note that the test functions $f_w\in {\cal B}^p_0(U^n)\subset{\cal B}^p_{0*}(U^n),$ we know $(\ref{9})$ holds. In order to prove the Converse, we first prove that if $\phi_l\in{\cal B}^q_{0*}(U^n)$ for every $l=1,2,\cdots,n.,$ then $f\circ\phi\in{\cal B}^q_{0*}(U^n)$ for any $f\in{\cal B}^p_{0*}(U^n).$ Without loss of generality, we prove this result when $n=2.$ For any sequence $\{z^j=(z^j_1, z^j_2)\}\subset U^n$ with $z^j\to\partial^* U^n$ as $j\to\infty,$ then $$|z^j_1|\to 1, |z^j_2|\to 1.$$ Since $|\phi_1(z^j)|<1$ and $|\phi_2(z^j)|<1,$ there exists a subsequence $\{z^{j_s}\}$ in $\{z^j\}$ such that $$|\phi_1(z^{j_s})|\to \rho_1, |\phi_2(z^{j_s})|\to\rho_2,$$ as $s\to\infty .$ It is clear that $0\leq\rho_1, \rho_2\leq 1.$ \begin{eqnarray}&&\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\nonumber\\ &&\leq\left| \displaystyle\frac{\partial f}{\partial w_1}(\phi(z^{j_s}))\right| \left|\displaystyle\frac {\partial\phi_1}{\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q+\left| \displaystyle\frac{\partial f}{\partial w_2}(\phi(z^{j_s}))\right| \left|\displaystyle\frac {\partial\phi_2}{\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\nonumber\\ &&=\left| \displaystyle\frac{\partial f}{\partial w_1}(\phi(z^{j_s})) \right|(1-|\phi_1(z^{j_s})|^2)^p \left|\displaystyle\frac {\partial\phi_1}{\partial z_k}(z^{j_s})\right|\displaystyle\frac{(1-|z^{j_s}_k|^2)^q} {(1-|\phi_1(z^{j_s})|^2)^p}\nonumber\\ &&+\left| \displaystyle\frac{\partial f}{\partial w_2}(\phi(z^{j_s}))\right| (1-|\phi_2(z^{j_s})|^2)^p \left|\displaystyle\frac {\partial\phi_2}{\partial z_k}(z^{j_s})\right|\displaystyle\frac{(1-|z^{j_s}_k|^2)^q} {(1-|\phi_2(z^{j_s})|^2)^p},\label{14}\end{eqnarray} $k=1,2.$ Now we prove the left of $(\ref{14})\to 0$ as $s\to\infty$ according to four cases. Case 1. If $\rho_1<1$ and $\rho_2<1.$ It is clear that there exist $r_1$ and $r_2$ such that $\rho_1<r_1<1$ and $\rho_2<r_2<1,$ so as $j$ is large enough, $|\phi_1(z^{j_s})|\leq r_1$ and $|\phi_2(z^{j_s})|\leq r_2.$ By $\phi_1, \phi_2\in{\cal B}^q_{0*}(U^n)$ and (\ref{14}), we get \begin{eqnarray*}\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q &\leq& \|f\|_p \displaystyle\frac{1}{(1-r_1^2)^p}\left|\displaystyle\frac {\partial\phi_1}{\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\\ &&+\|f\|_p\displaystyle\frac{1}{(1-r_2^2)^p}\left|\displaystyle\frac {\partial\phi_2}{\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\\ & &\to 0\end{eqnarray*} as $s\to\infty.$ Case 2. If $\rho_1=1$ and $\rho_2=1.$ Then $\phi(z^{j_s})\to\partial^*U^n,$ by (\ref{9}) and $f\in{\cal B}^p_{0*}(U^n)$, (\ref{14}) gives that \begin{eqnarray*}&&C\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\\ &&C\leq \left| \displaystyle\frac{\partial f}{\partial w_1}(\phi(z^{j_s})) \right|(1-|\phi_1(z^{j_s})|^2)^p +\left| \displaystyle\frac{\partial f}{\partial w_2}(\phi(z^{j_s}))\right| (1-|\phi_2(z^{j_s})|^2)^p\to 0\end{eqnarray*} as $s\to\infty.$ Case 3. If $\rho_1<1$ and $\rho_2=1.$ Similar to Case 1, we can prove that \begin{eqnarray}&&\left| \displaystyle\frac{\partial f}{\partial w_1}(\phi(z^{j_s})) \right|(1-|\phi_1(z^{j_s})|^2)^p\left|\displaystyle\frac {\partial\phi_1}{\partial z_k}(z^{j_s})\right|\displaystyle\frac{(1-|z^{j_s}_k|^2)^q} {(1-|\phi_1(z^{j_s})|^2)^p}\nonumber\\ &&\leq\|f\|_p\displaystyle\frac{1}{(1-r_1^2)^p} \left|\displaystyle\frac {\partial\phi_1}{\partial z_k}(z^{j_s})\right|\displaystyle\frac{(1-|z^{j_s}_k|^2)^q} {(1-|\phi_1(z^{j_s})|^2)^p} \to 0 \label{15}\end{eqnarray} as $s\to\infty.$ On the other hand, for fixed $s,$ let $w^{j_s}_2=\phi_2(z^{j_s}),$ then $|w^{j_s}_2|<1.$ Denote $$F(w_1)=\displaystyle\frac{\partial f}{\partial w_2}(w_1, w^{j_s}_2).$$ It is clear that $F(w_1)$ is holomorphic on $|w_1|<1,$ choose $R_{j_s}\to 1$ with $r_1\leq R_{j_s}<1.$ $|\phi_1(z^{j_s})|\leq r_1,$ so $$|F(\phi_1(z^{j_s}))|\leq\max\limits_{|w_1|\leq r_1}|F(w_1)| \leq\max\limits_{|w_1|\leq R_{j_s}}|F(w_1)|= \max\limits_{|w_1|=R_{j_s}}|F(w_1)|=|F(w^{j_s}_1)|,$$ where $|w^{j_s}_1|=R_{j_s}\to 1.$ This means that $\left|\displaystyle\frac{\partial f}{\partial w_2}(\phi_1(z^{j_s}), \phi_2(z^{j_s})) \right|\leq \left|\displaystyle\frac{\partial f}{\partial w_2}(w^{j_s}_1, w^{j_s}_2) \right|. $ Since $|w^{j_s}_1|\to 1, |w^{j_s}_2|\to\rho_2=1$ and $f\in{\cal B}^p_{0*}(U^n),$ $$\left|\displaystyle\frac{\partial f}{\partial w_2}(w^{j_s}_1, w^{j_s}_2)\right| (1-|w^{j_s}_2|^2)^p\to 0$$ as $s\to\infty,$ so by (\ref{9}), \begin{eqnarray}&&\left| \displaystyle\frac{\partial f}{\partial w_2}(\phi(z^{j_s}))\right| (1-|\phi_2(z^{j_s})|^2) ^p\left|\displaystyle\frac {\partial\phi_2}{\partial z_k}(z^{j_s})\right|\displaystyle\frac{(1-|z^{j_s}_k|^2)^q} {(1-|\phi_2(z^{j_s})|^2)^p}\nonumber\\ &&\leq C\left| \displaystyle\frac{\partial f}{\partial w_2}(w^{j_s}_1, w^{j_s}_2)\right| (1-|w^{j_s}_2|^2)^p\to 0\label{16}\end{eqnarray} as $s\to\infty.$ By (\ref{15}) and (\ref{16}), (\ref{14}) gives $$\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\to 0,$$ as $s\to\infty.$ Case 4. If $\rho_1=1$ and $\rho_2<1.$ Similar to Case 3, we can prove $$\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\to 0,$$ as $s\to\infty.$ Combining Case 1, Case 2, Case 3 and Case 4, we know there exists a subsequence $\{z^{j_s}\}$ in $\{z^j\}$ such that $$\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\to 0,$$ as $s\to\infty$ for $k=1,2.$ We claim that $$\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^j)\right|(1-|z^j_k|^2)^q\to 0,$$ as $j\to\infty.$ In fact, if it fails, then there exists a subsequence $\{z^{j_s}\}$ such that \begin{equation}\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\to \varepsilon>0\label{17}\end{equation} for $k=1$ or $2$. But from the above discussion, we can find a subsequence in $\{z^{j_s}\}$ we still write $\{z^{j_s}\}$ with $$\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j_s})\right|(1-|z^{j_s}_k|^2)^q\to 0,$$ it contradicts with (\ref{17}). So for any sequence $\{z^j\}\subset U^n$ with $z^j\to\partial^* U^n$ as $j\to\infty,$ we have $$\left|\displaystyle\frac{\partial(f\circ\phi)} {\partial z_k}(z^{j})\right|(1-|z^{j}_k|^2)^q\to 0$$ for $k=1,2.$ By (\ref{9}) and Lemma 3, it is clear that $f\circ\phi\in{\cal B}^q(U^n),$ so $f\circ\phi\in{\cal B}^q_{0*}(U^n).$ For any $f\in {\cal B}^p_0(U^n)).$ Since ${\cal B}^p_0(U^n))\subset {\cal B}^p_{0*}(U^n)),$ then $f\circ\phi\in{\cal B}^q_{0*}(U^n).$ By closed graph theorem we known that $$C_{\phi}:{\cal B}^p_{0*}(U^{n}) ({\cal B}^p_0(U^n)) \rightarrow{\cal B}^q_{0*}(U^{n})$$ is bounded. This ends the proof of Lemma 4. {\bf Remark 1}\hspace*{4mm}For the case $C_{\phi}:{\cal B}^p(U^n)\to{\cal B}^q_{0*}(U^n)$, the necessity is also true, but we can't guaranty that the sufficiency is true because we can't sure that $C_{\phi}f\in{\cal B}^q_{0*}(U^n)$ for all $f\in{\cal B}^p(U^n$. \begin{Lemma} Let $\phi=(\phi_1, \phi_2, \cdots, \phi_n)$ be a holomorphic self-map of $U^n.$ Then $$C_{\phi}:{\cal B}^p_0(U^n) \rightarrow{\cal B}^q_0(U^{n})$$ is bounded if and only if if and only if $\phi^{\gamma}\in {\cal B}_0^q(U^n)$ for every multi-index $\gamma$, and (\ref{9}) holds.\end{Lemma} {\bf Proof} \hspace{2mm}Sufficiency. From (\ref{9}) and by Theorem 1 we know that $C_{\phi}:{\cal B}^p(U^n)\to{\cal B}^q(U^n)$ is bounded, in particular $$\|C_\phi f\|_q\leq \|C_\phi\|_{{\cal B}^p(U^n)\to {\cal B}^q(U^n)}\|f\|_p,\quad \mbox{for all}\; f\in{\cal B}_0^p(U^n).$$ The boundedness of $C_{\phi}: {\cal B}_0^p(U^n)\to {\cal B}_0^q(U^n)$ directly follows, if we prove $C_\phi f\in{\cal B}_0^q(U^n)$ whenever $f\in {\cal B}_0^p(U^n).$ So, let $f\in {\cal B}_0^p(U^n).$ By the definition of ${\cal B}_0^p(U^n)$ it follows that for every $\varepsilon>0$ there is a polynomial $p_\varepsilon$ such that $\|f-p_\varepsilon\|_p<\varepsilon.$ Hence \begin{equation}\|C_\phi f-C_\phi p_\varepsilon\|_q\leq \|C_\phi\|_{{\cal B}^p(U^n)\to {\cal B}^q(U^n)}\|f-p_\varepsilon\|_p<\varepsilon \|C_\phi\|_{{\cal B}^p(U^n)\to {\cal B}^q(U^n)}.\label{a}\end{equation} Since $\phi^{\gamma}\in{\cal B}_0^q(U^n)$ for every multi-index $\gamma,$ we obtain $C_\phi p_\varepsilon\in{\cal B}_0^q(U^n).$ From this and (\ref{a}) the result follows. If $C_{\phi}:{\cal B}_0^p(U^n)\to{\cal B}_0^q(U^n)$ is bounded, then (\ref{9}) can be proved as in Lemma 3, since the test functions appearing there belong to ${\cal B}_0^p(U^n).$ Since the polynomials $z^\gamma\in {\cal B}_0^p(U^n)$ for every multi-index $\gamma,$ we get $C_\phi z^\gamma\in {\cal B}_0^q(U^n),$ as desired. {\bf Remark 2}\hspace*{4mm}For the case $C_{\phi}:{\cal B}^p(U^n)\;\; ({\cal B}^p_{0*}(U^n))\to{\cal B}^q_{0}(U^n)$, similar to Remark 1, the necessity is also true, but we can't guaranty that the sufficiency is true. \begin{Lemma} If $\{f_k\}$ is a bounded sequence in ${\cal B}^p(U^n)$, then there exists a subsequence $\{f_{k_l}\}$ of $\{f_k\}$ which converges uniformly on compact subsets of $U^n$ to a holomorphic function $f\in{\cal B}^p(U^n)$. \end{Lemma} {\bf Proof}\hspace{2mm} Let $\{f_k\}$ be a bounded sequence in ${\cal B}^p(U^n)$ with $\|f_k\|_p\leq C.$ By Lemma 1, $\{f_j\}$ is uniformly bounded on compact subsets of $U^n$ and hence normal by Montel's theorem. Hence we may extract subsequence $\{f_{j_k}\}$ which converges uniformly on compact subsects of $U^n$ to a holomorphic function $f$. It follows that $\displaystyle\frac{\partial f_{j_k}}{\partial z_l}\to\displaystyle\frac{\partial f}{\partial z_l}$ for each $l\in\{1,2,\cdots,n\}$, so $$\sum\limits^n_{l=1}\left|\displaystyle\frac{\partial f}{\partial z_l}\right|(1-|z_l|^2)^p= \lim\limits_{k\to\infty}\sum\limits^n_{l=1}\left|\displaystyle\frac{\partial f_{j_k}}{\partial z_l}\right|(1-|z_l|^2)^p=\leq \sup\limits_{k}\|f_{j_k}\|_p\leq C,$$ which implies $f\in{\cal B}^p(U^n)$. The Lemma is proved. \begin{Lemma} Let $\Omega$ be a domain in $\mbox{\Bbb C}^n,$ $f\in H(\Omega).$ If a compact set $K$ and its neighborhood $G$ satisfy $K\subset G\subset\subset \Omega$ and $\rho=dist(K, \partial G)>0,$ then $$\sup\limits_{z\in K}\left|\displaystyle\frac{\partial f}{\partial z_j}(z) \right|\leq\displaystyle\frac{\sqrt{n}}{\rho}\sup\limits_{z\in G}|f(z)|.$$ \end{Lemma} {\bf Proof}\hspace{2mm}Since $\rho=dist(K, \partial G)>0,$ for any $a\in K,$ the polydisc $$P_a=\left\{(z_1, \cdots, z_n)\in\mbox{\Bbb C}^n: |z_j-a_j| <\displaystyle\frac{\rho}{\sqrt{n}}, j=1,\cdots,n\right\}$$ is contained in $G.$ By Cauchy's inequality, $$\left|\displaystyle\frac{\partial f}{\partial z_j}(a) \right|\leq\displaystyle\frac{\sqrt{n}}{\rho} \sup\limits_{z\in\partial^* P_a}|f(z)|\leq \displaystyle\frac{\sqrt{n}}{\rho}\sup\limits_{z\in G}|f(z)|.$$ Taking the supremum for $a$ over $K$ gives the desired inequality. \section{The Proof of Theorem 1} Now we turn to the proof of Theorem 1. The lower estimate. It is clear that $\{m^{p-1}z^m_1\}\subset{\cal B}^p_0(U^n) \subset{\cal B}_{0*}(U^n) \subset{\cal B}(U^n)$ for $m=1,2,\cdots,$ and this sequence converges to zero uniformly on compact subsets of the unit polydisc $U^n.$ \begin{equation}\|m^{p-1}z^m_1\|_p =\sup\limits_{z\in U^n} (1-|z_1|^2)^p|m^pz^{m-1}_1|.\label{18}\end{equation} Let $p(x)=m^p(1-x^2)^px^{m-1},$ then $$p'(x)=-m^px^{m-2}(1-x^2)^{p-1}\left[(2p+m-1)x^2-(m-1)\right],$$ so $p'(x)\leq 0$ for $x\in \left[\sqrt{\displaystyle\frac{m-1}{2p+m-1}},1\right],$ and $p'(x)\geq 0$ for $x\in \left[0,\sqrt{\displaystyle\frac{m-1}{2p+m-1}}\right].$ That is, $p(x)$ is a decreasing function for $x\in \left[\sqrt{\displaystyle\frac{m-1}{2p+m-1}},1\right]$ and $p(x)$ is a increasing function for $x\in \left[0,\sqrt{\displaystyle\frac{m-1}{2p+m-1}}\right].$ Hence $$\max\limits_{x\in [0,1]}p(x)=p\left(\sqrt{\displaystyle\frac{m-1}{2p+m-1}}\right).$$ It follows from (\ref{18}) that $$\|m^{p-1}z^m_1\|_p =p\left(\sqrt{\displaystyle\frac{m-1}{2p+m-1}}\right)=\left(\displaystyle\frac{2p}{2p+m-1}\right)^pm^p \left(\displaystyle\frac{m-1}{2p+m-1}\right)^{\frac{m-1}{2}} \to\left(\displaystyle\frac{2p}{e}\right)^p,$$ as $m\to\infty.$ Therefore, the sequence $\{m^{p-1}z^m_1\}_{m\geq 2}$ is bounded away from zero. Now we consider the normalized sequence $\{f_m=\displaystyle\frac{m^{p-1}z^m_1}{\|m^{p-1}z^m_1\|_p}\}$ which also tends to zero uniformly on compact subsets of $U^n.$ For each $m\geq 2,$ we define $$A_m=\{z=(z_1, \ldots, z_n)\in U^n: r_m\leq |z_1|\leq r_{m+1}\},$$ where $r_m=\sqrt{\displaystyle\frac{m-1}{2p+m-1}}.$ So \begin{eqnarray*}&&\min\limits_{A_m} \sum\limits^n_{l=1} \left\{\left|\displaystyle\frac{\partial f_m} {\partial z_l}(z)\right|(1-|z_l|^2)^p\right\} =\min\limits_{A_m}\left|\displaystyle\frac{\partial f_m} {\partial z_1}(1-|z_1|^2)^p\right|\\ &&=\displaystyle\frac{ (1-r^2_{m+1})^p|m^pr^{m-1}_{m+1}|}{\|m^{p-1}z^m_1\|_p} =\left(\displaystyle\frac{2p+m-1}{2p+m}\right) \left(\displaystyle\frac{m(2p+m-1)}{(m-1)(2p+m)}\right)^{\frac{m-1}{2}}=c_m. \end{eqnarray*} It is easy to show that $c_m$ tends to 1 as $m\to\infty$. For the moment fix any compact operator $K:{\cal B}^p(U^n){\cal B}^p(U^n) ({\cal B}^p_0(U^n)$ or ${\cal B}^p_{0*}(U^n))\longrightarrow {\cal B}^q(U^n)$ $({\cal B}^q_0(U^n)$ or ${\cal B}^q_{0*}(U^n)).$ The uniform convergence on compact subsets of the sequence $\{f_m\}$ to zero and the compactness of $K$ imply that $\|Kf_m\|_q\to 0.$ It is easy to show that if a bounded sequence that is contained in ${\cal B}^p_{0*}(U^n)$ converges uniformly on compact subsets of $U^n,$ then it also converges weakly to zero in ${\cal B}^p_{0*}(U^n)$ as well as in ${\cal B}^p(U^n).$ Since $\|f_m\|_p=1$, we have \begin{eqnarray*}&&\|C_{\phi}-K\|\geq \limsup\limits_{m}\|(C_{\phi}-K)f_m\|_q\nonumber\\ &&\geq\limsup\limits_{m}\left(\|C_{\phi}f_m\|_q -\|Kf_m\|_q\right) =\limsup\limits_{m}\|C_{\phi}f_m\|_q\nonumber\\ &&\geq \limsup\limits_{m}\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left\{\left|\displaystyle\frac{\partial(f_m\circ\phi)}{\partial z_k}\right| (1-|z_k|^2)^q \right\}\nonumber\\ &&=\limsup\limits_{m}\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\displaystyle\frac{\partial f_m}{\partial w_1}(\phi(z))\right| \left|\displaystyle\frac {\partial\phi_1}{\partial z_k}(z)\right|(1-|z_k|^2)^q\nonumber\\ &&=\limsup\limits_{m}\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\displaystyle\frac{\partial\phi_1}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_1(z)|^2)^p}\left|\displaystyle\frac {\partial f_m}{\partial w_1}(\phi(z))\right|(1-|\phi_1(z)|^2)^p\nonumber\\ &&\geq\limsup\limits_{m}\sup\limits_{\phi(z)\in A_m} \sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_1}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_1(z)|^2)^p}\left|\displaystyle\frac {\partial f_m}{\partial w_1}(\phi(z))\right|(1-|\phi_1(z)|^2)^p\nonumber\\ &&\geq\limsup\limits_{m}\sup\limits_{\phi(z)\in A_m}\sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_1}{\partial z_k}(z)\right| \displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_1(z)|^2)^p}\nonumber\\ & &\times\liminf\limits_{m}\min\limits_{\phi(z)\in A_m} \left|\displaystyle\frac{\partial f_m} {\partial w_1}(\phi(z))\right|(1-|\phi_1(z)|^2)^p\nonumber\\ &&\geq\limsup\limits_{m} \sup\limits_{\phi(z)\in A_m}\sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_1}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_1(z)|^2)^p}\liminf\limits_m c_m\nonumber\\ &&\geq\limsup\limits_{m} \sup\limits_{\phi(z)\in A_m}\sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_1}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_1(z)|^2)^p}. \end{eqnarray*} So \begin{eqnarray}\|C_{\phi}\|_e &=&\inf\{\|C_{\phi}-K\|: K \mbox{ is compact}\} \nonumber\\ &\geq&\limsup\limits_{m} \sup\limits_{\phi(z)\in A_m}\sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_1}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_1(z)|^2)^p}.\label{19} \end{eqnarray} For each $l=1,2,\cdots, n, $ define \begin{equation}a_l=\lim\limits_{\delta\to 0} \sup\limits_{dist(\phi(z), \partial U^n)<\delta} \sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}.\label{20}\end{equation} For any $\varepsilon>0,$ (\ref{20}) shows that there exists a $\delta_0$ with $0<\delta_0<1,$ such that \begin{equation}\sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}>a_l-\varepsilon,\label{21}\end{equation} whenever $dist(\phi(z),\partial U^n)<\delta_0$ and $l=1,2,\cdots,n.$ Since $r_m\to 1$ as $m\to\infty,$ so as $m$ is large enough, $r_m>1-\delta_0.$ If $\phi(z)\in A_m,$ $r_m\leq |\phi_1(z)|\leq r_{m+1},$ so $1-r_{m+1}<1-|\phi_1(z)|<1-r_m<\delta_0,$ $dist(\phi_1(z),\partial U)<\delta_0.$ There exists $w_1$ with $|w_1|=1$ such that $dist(\phi_1(z),w_1)=dist(\phi_1(z),\partial U)<\delta_0.$ Let $w=(w_1, \phi_2(z),\ldots, \phi_n(z)),$ $w\in\partial U^n$, then $$dist(\phi(z),\partial U)\leq dist(\phi(z), w)=dist(\phi_1(z),w_1)<\delta_0.$$ By (\ref{21}), (\ref{19}) implies that $$\|C_{\phi}\|_e\geq a_1-\varepsilon.$$ Similarly, if we choose $g_m(z)=\displaystyle\frac{m^{p-1}z^{m}_l}{\|m^{p-1}z^m_l\|}$, we have $$\|C_{\phi}\|_e\geq a_l-\varepsilon,$$ for every $l=2\cdots,n.$ So \begin{eqnarray*}\|C_{\phi}\|_e&\geq&\displaystyle\frac{1}{n} \sum\limits^n_{l=1} (a_l-\varepsilon)\\ &=&\displaystyle\frac{1}{n}\sum\limits^n_{l=1} (\lim\limits_{\delta\to 0} \sup\limits_{dist(\phi(z), \partial U^n)<\delta} \sum\limits^n_{k=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}-\varepsilon)\\ &\geq&\displaystyle\frac{1}{n} \lim\limits_{\delta\to 0} \sup\limits_{dist(\phi(z), \partial U^n)<\delta} \sum\limits^n_{k,l=1} \left|\displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}- \varepsilon.\end{eqnarray*} Let $\varepsilon\to 0,$ the low estimate follows. The upper estimate. To obtain the upper estimate we first prove the following proposition. \begin{Proposition} Let $\phi=(\phi_1, \ldots, \phi_n)$ a holomorphic self-map of $U^{n}.$ The operators $K_m$ ($m\geq 2$) as follows: $$K_mf(z)=f(\displaystyle\frac{m-1}{m}z),$$ for $f\in H(U^n).$ Then the operators $K_m$ have the following properties: (i)\hspace*{2mm} For any $f\in H(U^n),$ $K_mf\in {\cal B}^p_0(U^n)\subset {\cal B}^p_{0*}(U^n)\subset {\cal B}^p(U^n).$ (ii)\hspace*{2mm} If $C_{\phi}:$ ${\cal B}^p(U^n)$ ( ${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow$ $ {\cal B}^q(U^n)$ ( ${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) is bounded, then $C_{\phi}K_mf\in{\cal B}^q(U^n)$ (${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) for all $f\in H(U^n).$ (iii) \hspace*{2mm}For fixed $m$, the operator $K_m$ is compact on ${\cal B}^p(U^n)$ (${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$). (iv)\hspace*{2mm} If $C_{\phi}:$ ${\cal B}^p(U^n)$ ( ${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$)$\rightarrow$ $ {\cal B}^q(U^n)$ ( ${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n)$) is bounded, then $C_{\phi}K_mf\in{\cal B}^q(U^n)$ (${\cal B}^q_{0}(U^n)$ or ${\cal B}^q_{0*}(U^n))$ is compact. (v) \hspace*{2mm} $\|I-K_m\|\leq 2.$ (vi) \hspace*{2mm}$(I-K_m)f$ converges uniformly to zero on compact subset of $U^n$.\end{Proposition} {\bf Proof}\hspace*{2mm} (i)\hspace*{2mm} Let $f\in H(U^n),$ $r_m=\displaystyle\frac{m-1}{m}, (0<r_m<1)$ and $f_m(z)=K_mf(z)=f(r_mz).$ First note that \begin{eqnarray}\|f_m\|_p&=&|f(0)|+\sup\limits_{z\in U^n}\sum\limits^n_{k=1} r_m\left|\frac{\partial f} {\partial z_k}(r_mz)\right|\left(1- |z_k|^2\right)^p\nonumber\\ &\leq&|f(0)|+\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\frac{\partial f} {\partial z_k}(r_mz)\right|\left(1- |r_mz_k|^2\right)^p\leq\|f\|_p.\label{b}\end{eqnarray} On the other hand, $f_m\in H(\frac{1}{r_m}U^n).$ $0<\displaystyle\frac{2}{1+r_m}<\displaystyle\frac{1}{r_m},$ $\displaystyle\frac{2}{1+r_m}\overline{U^n}\subset \displaystyle\frac{1}{r_m}U^n.$ which implies that for fixed $m,$ and $\varepsilon=\displaystyle\frac{1}{j}, j=1,2,\cdots,$ there is a polynomial $P^{(j)}_m$ such that $$\sup_{z\in \frac{2}{1+r_m}\overline{U^n}}|f_m(z)-P^{(j)}_m(z)|<(1-r_m)^2\displaystyle\frac{1}{j}.$$ Let $K=\overline{U^n},$ $G=\displaystyle\frac{2}{1+r_m}U^n,$ $\Omega=\displaystyle\frac{1}{r_m}U^n,$ then $K\subset G\subset\subset\Omega$, and $\rho=dist(K,\partial G)=\displaystyle\frac{1-r_m}{1+r_m}>0$, so $\forall w\in U^n$, $k\in\{1,\cdots,n\}$, it follows from Lemma 7 that \begin{eqnarray*} &&\Big|\frac{\partial (f_m-P_m^{(j)})}{\partial w_k}(w)\Big|\leq \sup_{w\in K} \Big|\frac{\partial (f_m-P_m^{(j)})}{\partial w_k}(w)\Big|\\[6pt] &\leq& \frac{\sqrt{n}(1+r_m)}{1-r_m}\sup_{w\in G}|f_m(w)-P_m^{(j)}(w)|\\[6pt] &\leq& \frac{\sqrt{n}(1+r_m)}{1-r_m}(1-r_m^2)\frac{1}{j}\leq 4\sqrt{n}\frac{1}{j}. \end{eqnarray*} Therefore $$ \sum_{k=1}^n\Big|\frac{\partial (f_m-P_m^{(j)})}{\partial w_k}(w)\Big|(1-|w_k|^p)^p \leq 4n\sqrt{n}\frac{1}{j}\to 0 $$ as $j\to \infty.$ that is, $$ ||f_m-P_m^{(j)}||_{{\cal B}^p}=|f_m(0)-P_m^{(j)}(0)|+\sup_{w\in U^n}\sum_{k=1}^n\Big|\frac{\partial (f_m-P_m^{(j)})}{\partial w_k}(w)\Big|(1-|w_k|^p)^p\to 0. $$ $P_m^{(j)}(w)\in {\cal B}^p_0(U^n)$ implies that $f_m\in{\cal B}^p_0(U^n)$. (ii)\hspace*{2mm} By (i), as desired. (iii) \hspace{2mm} For any sequence $\{f_{j}\}\subset{\cal B}^p(U^n)$ (${\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n)$) with $\|f_j\|_p\leq M,$ by (i), $\{K_mf_{j}\}\in {\cal B}^p_{0}(U^n).$ By Lemma 6, there is a subsequence $\{f_{j_s}\}$ of $\{f_j\}$ which converges uniformly on compact subsets of $U^n$ to a holomorphic function $f\in{\cal B}^p(U^n)$ and $\|f\|_p\leq M.$ $\left\{\displaystyle\frac{\partial f_{j_s}}{\partial z_i} \right\}, i=1,2,\cdots,n,$ also converges uniformly on compact subsets of $U^n$ to the holomorphic function $\displaystyle\frac{\partial f} {\partial z_i}.$ So as $s$ is large enough, for any $w\in E=\{\frac{m-1}{m}z: z\in \overline{U^n}\}\subset U^n$ \begin{equation}\left|\displaystyle\frac {\partial (f_{j_s}-f)}{\partial w_l}(w)\right|<\varepsilon,\label{23}\end{equation} for every $l=1,2,\cdots,n.$ So \begin{eqnarray}&&\left\|K_mf_{j_s}-K_mf\right\|_p =\left\|f_{j_s}(\displaystyle\frac{m-1}{m}z)- f(\displaystyle\frac{m-1}{m}z)\right\|_p\nonumber\\ &&=\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left\{\left|\displaystyle\frac{\partial \left[(f_{j_s}-f) (\displaystyle\frac{m-1}{m}z)\right]} {\partial z_k}\right| (1-|z_k|^2)^p \right\}+|f_{j_s}(0)-f(0)|\nonumber\\ &&\leq\sup\limits_{z\in U^n}\sum\limits^n_{k=1}\sum\limits^n_{l=1}\left| \displaystyle\frac{\partial (f_{j_s}-f)}{\partial w_l} (\displaystyle\frac{m-1}{m}z)\right|\displaystyle\frac{m-1}{m}+|f_{j_s}(0)-f(0)| \nonumber\\ &&\leq n\sup\limits_{w\in E_1}\displaystyle\frac{m-1}{m}\sum\limits^n_{l=1} \left|\displaystyle\frac {\partial (f_{j_s}-f)}{\partial w_l}(w)\right|+|f_{j_s}(0)-f(0)|\to 0,\label{24} \end{eqnarray} as $s\to\infty.$ This shows that $\{K_mf_{j_s}\}$ converges to $g=K_mf\in{\cal B}^p_{0}(U^n)\subset {\cal B}^p_{0*}(U^n)\subset {\cal B}^p(U^n).$ So $K_m$ is compact on ${\cal B}^p(U^n)$(${\cal B}^p_0(U^n)$ or ${\cal B}^p_{0*}(U^n)).$ (iv)\hspace*{2mm} By (i) and (iii), the result is obvious. (v)\hspace*{2mm}In fact, for any $f\in{\cal B}^p(U^n)({\cal B}^p_{0}(U^n)$ or ${\cal B}^p_{0*}(U^n))$, note that $(I-K_m)f(0)=0$, so \begin{eqnarray*}&&\|(I-K_m)f\|_p =\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\displaystyle\frac{\partial (I-K_m)f}{\partial z_k}(z) \right|(1-|z_k|^2)\\ &&=n\sup\limits_{z\in U^n}\max\limits_{1\leq k\leq n} \left|\displaystyle\frac{\partial f}{\partial z_k}(z) -(1-\frac{1}{m}) \displaystyle\frac{\partial f}{\partial z_k}((1-\frac{1}{m})z) \right|(1-|z_k|^2)^p\\ &&\leq\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left|\displaystyle\frac{\partial f}{\partial z_k}(z)\right|(1-|z_k|^2)^p\\ &&+(1-\frac{1}{m})\sup\limits_{z\in U^n}\sum\limits^n_{k=1}\left|\displaystyle\frac{\partial f}{\partial z_k}((1-\frac{1}{m})z) \right|(1-|(1-\frac{1}{m})z_k|^2)^p\\ &&\leq \|f\|_p+\|f\|_p=2\|f\|_p, \end{eqnarray*} so $\|I-K_m\|\leq 2.$ (vi)\hspace*{2mm} For any compact subset $E\subset U^n$, $\exists r,$ $0<r<1$ such that $E\subset rU^n\subset \subset U^n$. For $\forall z\in E$, \begin{eqnarray*} |(I-K_m)f(z)|&=&|f(z)-f_m(z)|=|f(z)-f(r_mz)|\\ &=&\left|\int_{r_m}^1\frac{d}{dt}(f(tz))dt\right|=\left|\int_{r_m}^1\sum_{k=1}^n \frac{\partial f}{\partial w_k}(tz)\cdot z_kdt\right|\\ &\leq& \sum_{k=1}^n\int_{r_m}^1\left|\frac{\partial f}{\partial w_k}(tz)\right|dt. \end{eqnarray*} $t\in[r_m,1]$, $\forall z\in U^n,\hspace*{4mm} |tz_k|=t|z_k|<|z_k|<r,$\hspace*{4mm} so $\displaystyle\frac{\partial f}{\partial w_k}(w)$ is bounded in $r\overline{U^n}$, i.e., $\forall z\in E,$\hspace*{4mm} $\left|\displaystyle\frac{\partial f} {\partial w_k}(tz)\right|\leq M$. So $$ |(I-K_m)f(z)|\leq nM(1-r_m)\to 0$$ as $m\to \infty$, the results follows. Now return to the upper estimate. For the convenience, we denote $\|f\|=\|f\|_p.$ \begin{eqnarray}&&\|C_{\phi}\|_e \leq\|C_{\phi}-C_{\phi}K_m\| =\|C_{\phi}(I-K_m)\| =\sup\limits_{\|f\|=1}\|C_{\phi}(I-K_m)f\|_q\nonumber\\ &&=\sup\limits_{\|f\|=1}\left(\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \left\{\left|\displaystyle\frac{\partial (I-K_m)(f\circ\phi)} {\partial z_k}\right|(1-|z_k|^2)^q\right\}+\left|(I-K_m)f(\phi(0))\right|\right) \nonumber\\ &&\leq\sup\limits_{\|f\|=1}\sup\limits_{z\in U^n}\sum\limits^n_{k=1} \sum\limits^n_{l=1}\left|\displaystyle\frac{\partial (I-K_m)f} {\partial w_l}(\phi(z))\right| \left|\displaystyle\frac {\partial\phi_l}{\partial z_k}(z)\right|(1-|z_k|^2)^q\nonumber\\ && +\sup\limits_{\|f\|=1} \left|f(\phi(0))-f(\frac{m-1}{m}\phi(0))\right|\nonumber\\ &&\leq \sup\limits_{\|f\|=1}\sup\limits_{z\in U^n} \sum\limits^n_{k, l=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right| \displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p} \left|\displaystyle\frac{\partial (I-K_m)f}{\partial w_l}(\phi(z)) \right|(1-|\phi_l(z)|^2)^p\nonumber\\ && +\sup\limits_{\|f\|=1} \left|f(\phi(0))-f(\frac{m-1}{m}\phi(0))\right|.\label{26}\end{eqnarray} Denote $G_1=\{z\in U^n: dist(\phi(z), \partial U^n)<\delta\},$ $G_2=\{z\in U^n: dist(\phi(z), \partial U^n)\geq\delta\},$ $G=\{w\in U^n: dist(w, \partial U^n)\geq\delta\}$, where $G$ is a compact subset of $\mbox{\Bbb C}^n.$ Then by Lemma 3, Lemma 4 and Lemma 5, condition (9) holds, so \begin{eqnarray}\|C_{\phi}\|_e &\leq& \sup\limits_{\|f\|=1}\sup\limits_{z\in G_1} \sum\limits^n_{k, l=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p} \left|\displaystyle\frac {\partial (I-K_m)f}{\partial w_l}(\phi(z))\right|(1-|\phi_l(z)|^2)^q \nonumber\\ &&+C\sup\limits_{\|f\|=1}\sup\limits_{z\in G_2} \sum\limits^n_{l=1}(1-|\phi_l(z)|^2)^p \left|\displaystyle\frac{\partial (I-K_m)f}{\partial w_l}(\phi(z))\right|\nonumber\\ && +\sup\limits_{\|f\|=1} \left|f(\phi(0))-f(\frac{m-1}{m}\phi(0))\right| \nonumber\\ &\leq& \|I-K_m\|\sup\limits_{z\in G_1} \sum\limits^n_{k, l=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}\nonumber\\ &&+C\sup\limits_{\|f\|=1}\sup\limits_{z\in G_2}\sum\limits^n_{l=1}(1-|\phi_l(z)|^2)^p \left|\displaystyle\frac{\partial (I-K_m)f}{\partial w_l}(\phi(z))\right|\nonumber\\ && +\sup\limits_{\|f\|=1} \left|f(\phi(0))-f(\frac{m-1}{m}\phi(0))\right|\nonumber\\ &\leq& 2\sup\limits_{z\in G_1} \sum\limits^n_{k, l=1}\left| \displaystyle\frac{\partial\phi_l}{\partial z_k}(z)\right|\displaystyle\frac {(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}\nonumber\\ &&+C\sup\limits_{\|f\|=1}\sup\limits_{z\in G_2} \sum\limits^n_{l=1}(1-|\phi_l(z)|^2)^p \left|\displaystyle\frac{\partial (I-K_m)f}{\partial w_l}(\phi(z))\right|\nonumber\\ && +\sup\limits_{\|f\|=1} \left|f(\phi(0))-f(\frac{m-1}{m}\phi(0))\right|.\label{27} \end{eqnarray} Denote the second term and third term of the right hand side of (\ref{27}) by $I_1$ and $I_2$. Then Theorem 1 is proved if we can prove $$\lim\limits_{m\to\infty}I_1=0\hspace*{4mm} \mbox{and}\hspace*{4mm} \lim\limits_{m\to\infty}I_2=0.$$ To do this, let $z\in G_2$ and $w=\phi(z),$ then $w\in G$ \begin{eqnarray}I_1&\leq&C\sup\limits_{\|f\|=1} \sup\limits_{w\in G}\sum\limits^n_{l=1}(1-|w_l|^2)^p \left|\displaystyle\frac{\partial f}{\partial w_l}(w)- (1-\frac{1}{m})\displaystyle\frac{\partial f}{\partial w_l} ((1-\frac{1}{m})w)\right|\nonumber\\ &\leq& C\sup\limits_{\|f\|=1} \sup\limits_{w\in G}\sum\limits^n_{l=1}(1-|w_l|^2)^p \left|\displaystyle\frac{\partial f}{\partial w_l}(w)- \displaystyle\frac{\partial f}{\partial w_l} ((1-\frac{1}{m})w)\right|\nonumber\\ & &+\displaystyle\frac{C}{m}\sup\limits_{\|f\|=1} \sup\limits_{w\in G}\sum\limits^n_{l=1}(1-|w_l|^2)^p \left|\displaystyle\frac{\partial f}{\partial w_l} ((1-\frac{1}{m})w)\right|\nonumber\\ &\leq& C\sup\limits_{\|f\|=1} \sup\limits_{w\in G}\sum\limits^n_{l=1}(1-|w_l|^2)^p \left|\displaystyle\frac{\partial f}{\partial w_l}(w)- \displaystyle\frac{\partial f}{\partial w_l} ((1-\frac{1}{m})w)\right|+\displaystyle\frac{C}{m}.\label{28} \end{eqnarray} Let $w=(w_1,w_2,\cdots,w_{n-1},w_n),$ for $m$ large enough, we have \begin{eqnarray}&&\left|\displaystyle\frac{\partial f}{\partial w_l}(w)- \displaystyle\frac{\partial f}{\partial w_l}((1-\frac{1}{m})w) \right|\nonumber\\ &&\leq\sum\limits^n_{j=1}\left|\displaystyle\frac{\partial f} {\partial w_l}\left((1-\frac{1}{m})w_1,\cdots, (1-\frac{1}{m})w_{j-1}, w_j,\cdots, w_n\right)\right.\nonumber\\ & &-\left.\displaystyle\frac{\partial f}{\partial w_l} \left((1-\frac{1}{m})w_1,\cdots,(1-\frac{1}{m}) w_j,w_{j+1},\cdots,w_n\right)\right| \nonumber\\ &&=\sum\limits^n_{j=1}\left|\int^{w_j}_{(1-\frac{1}{m})w_j} \displaystyle\frac{\partial^2 f}{\partial w_l\partial w_j} \left((1-\frac{1}{m})w_1,\cdots, (1-\frac{1}{m})w_{j-1},\zeta, w_{j+1},\cdots, w_n\right)d\zeta\right|\nonumber\\ &&\leq\frac{1}{m}\sum\limits^n_{j=1} \sup\limits_{w\in G}\left|\displaystyle\frac {\partial^2 f}{\partial w_l\partial w_j}(w)\right|.\label{29}\end{eqnarray} Denote $G_3=\left\{w\in U^n:dist(w,\partial U^n)> \displaystyle\frac{\delta}{2}\right\},$ then $G\subset G_3\subset\subset U^n.$ Since $dist(G, \partial G_3)=\displaystyle\frac{\delta}{2},$ then by Lemma 7, (\ref{29}) gives \begin{equation}\left|\displaystyle\frac{\partial f}{\partial w_l}(w)- \displaystyle\frac{\partial f}{\partial w_l} ((1-\frac{1}{m})w)\right| \leq\displaystyle\frac{2n\sqrt{n}} {m\delta}\max\limits_{z\in G_3} \left|\displaystyle\frac{\partial f}{\partial w_l}(w)\right|.\label{30}\end{equation} On the other hand, on the unit ball of ${\cal B}^p(U^n)$, we have $$\sup\limits_{z\in G_3}(1-|w_l|^2)^p\left|\displaystyle\frac{\partial f} {\partial w_l}(w)\right|=\sup\limits_{dist(w,\partial U^n)>\frac{\delta}{2}} (1-|w_l|^2)^p\left|\displaystyle\frac{\partial f} {\partial w_l}(w)\right|\leq \|f\|_p=1,$$ namely \begin{equation}\sup\limits_{z\in G_3}\left|\displaystyle\frac{\partial f} {\partial w_l}(w)\right| \leq\displaystyle\frac{1}{1-\left(\frac{\delta}{2}\right)^2}= \displaystyle\frac{4}{4-\delta^2}.\label{31}\end{equation} Combining (\ref{28}), (\ref{30}) and (\ref{31}), imply $$I_1\leq\displaystyle\frac{2n\sqrt{n}C}{m\delta} \displaystyle\frac{4}{4-\delta^2}+\displaystyle\frac{C}{m}$$ and $\lim\limits_{m\to\infty}I_1=0.$ Now we can prove $\lim\limits_{m\to\infty}I_2=0$. In fact, \begin{eqnarray*}&&f(\phi(0))-f(\frac{m-1}{m}\phi(0))= \int^{1}_{\frac{m-1}{m}}\displaystyle\frac{d f(t\phi(0))}{d t}dt=\sum\limits^n_{l=1}\int^{1}_{\frac{m-1}{m}} \phi_l(0)\displaystyle\frac{\partial f}{\partial\zeta_l}(t\phi(0))dt.\end{eqnarray*} By Lemma 1, it follows that for any compact subset $K\subset U^n$, $|f(z)|\leq C_K \|f\|_p=C_K.$ Let $K=\{z\in U^n: |z_i|\leq|\phi_i(0)|\},$ So\begin{eqnarray*}&&|f(\phi(0))-f(\frac{m-1}{m}\phi(0))|\leq \sum\limits^n_{l=1}|\phi_l(0)|\int^{1}_{\frac{m-1}{m}}C_K dt \leq nC_K(1-\frac{m-1}{m})=\frac{nC_K}{m}, \end{eqnarray*} so $I_2\leq \frac{nC_K}{m}\to 0.$ Thus let first $m\to\infty,$ then $\delta\to 0$ in (\ref{27}), we get the upper estimate of $\|C_{\phi}\|_e$: $$\|C_{\phi}\|_e\leq 2\lim\limits_{\delta\to 0} \sup\limits_{dist(\phi(z),\partial U^n)<\delta}\sum\limits^n_{k,l=1} \left|\displaystyle\frac{\partial \phi_{l}}{\partial z_k}(z)\right| \displaystyle\frac{(1-|z_k|^2)^q}{(1-|\phi_l(z)|^2)^p}.$$ Now the proof of Theorem 1 is finished.
{ "timestamp": "2005-12-27T16:13:13", "yymm": "0503", "arxiv_id": "math/0503723", "language": "en", "url": "https://arxiv.org/abs/math/0503723" }
\section{Introduction} Quantum billiards, that is, closed compact domains in the two-dimensional Euclidean plane, are the simplest model of a quantum system corresponding to physical instances such as quantum dots or microstructures. The statistical properties of the quantum energy levels of such systems have been investigated, and it turns out that the statistical quantum behaviour can be related to the classical properties of the system. It is believed that systems whose classical motion is chaotic have energy levels behaving like eigenvalues of random matrix ensembles \cite{BohGiaSch84}, whereas the energy levels of systems whose classical motion is integrable are Poisson distributed, i.e. they behave like independent uniformly distributed random variables \cite{BerTab77a}. Both numerical evidence and some analytical results support these conjectures \cite{AndAlt95, AgaAndAlt95, Mar98}.\\ Among systems which are classically neither chaotic nor integrable, some systems have been found to display an eigenvalue statistics which is intermediate between the Poisson and the Random matrix distribution. The characteristics of such intermediate statistics are \cite{BogGerSch99} level repulsion, exponential decrease of the nearest-neighbour spacing distribution at infinity and linear asymptotic behaviour of the number variance (which is related to a non-vanishing form factor at small arguments). The form factor at the origin is equal to $1$ for classically integrable systems, to $0$ for chaotic systems, and it is found numerically to take values between 0 and 1 for intermediate statistics, the case $\overline{K_2(0)}=1/2$ corresponding to semi-Poisson statistics \cite{BogGerSch99}. Numerous quantum systems have been found to display numerically intermediate statistics: for example, pseudo-integrable systems such as rational polygonal billiards (polygons in which all angles are commensurate with $\pi$) \cite{CasPro99}, or quantum maps \cite{GirMarOke04}.\\ An analytical approach to the study of level statistics is the semiclassical trace formula, which gives an expansion of the density of energy levels as a sum over periodic orbits \cite{BalBlo72, Gut89}, or families of periodic orbits in the case of integrable systems \cite{BerTab76}. For diffractive systems, the trace formula can be modified to include diffractive orbits contributions \cite{Sie99, BogPavSch00}. It can be argued however (see \cite{BogGirSch01} for a discussion) that only the periodic orbits contribute to the semiclassical form factor at small arguments, $\overline{K_2(0)}$. The calculation of this quantity therefore only requires to find the periodic orbits and the areas occupied by the pencils of periodic orbits in a given system. Unfortunately, this is not a simple task. For instance is is not known whether any acute triangle has a periodic orbit. In the case of rational polygonal billiards, it has been shown \cite{Mas90} that the number ${\mathcal N}(L)$ of periodic orbits of length less than $L$ is quadratically bounded, namely there exist $c_1$ and $c_2$ such that $c_1L^2\leq {\mathcal N}(L)\leq c_2L^2$, but even for general rational polygonal billiards exact asymptotics is not known. There exist however certain specific rational polygonal billiards for which more precise statements are known. For instance for Veech billiards \cite{Vee89, Vor96}, a special class of rational polygonal billiards (whose stabilizer is a discrete cofinite subgroup of $SL(2, \mathbb{R})$), precise asymptotics for ${\mathcal N}(L)$ is known, and in \cite{BogGirSch01} it was possible to calculate analytically the form factor at the origin for triangular Veech billiards.\\ This paper presents the calculation of the semiclassical form factor at the origin for a billiard which does not have this special Veech property, the barrier billiard. The barrier billiard is one of the simplest pseudo-integrable billiards. It was introduced by Hannay and McCraw \cite{HanMcc90} and consists of a rectangle $[0,a]\times[0,b]$ containing a barrier described by the segment $\{\epsilon_0 a\}\times[0,\alpha b]$ with $0\leq\epsilon_0, \alpha<1$ (see Figure \ref{billard} left). It is a rational polygonal billiard with six angles equal to $\pi/2$ and one angle equal to $2\pi$. It is therefore a pseudo-integrable billiard \cite{BerRic81}, and the movement in phase space takes place on a surface of genus 2. When the height of the barrier is such that $\alpha\in\mathbb{Q}$ then the barrier billiard is a Veech billiard. But when $\alpha$ is irrational the billiard loses this property. Nevertheless, from results obtained in \cite{EskMasSch01}, it is still possible to work out the distribution of the periodic orbits in this latter case, and thus calculate analytically the semiclassical form factor at the origin, provided the position of the barrier is a rational number with respect to the size of the side: $\epsilon_0=p/q$ with $p,q\in\mathbb{N}$ coprime. We will first devise a method to obtain a complete characterization of the periodic orbit pencils in the non-Veech barrier billiard (Section \ref{section3}). We then rigourously derive asymptotics for each family of periodic orbit pencils (Section \ref{onpeutremplacer}), then use this result to calculate the semiclassical form factor at small arguments (Section \ref{calculff}). Previously obtained analytical results show that the semi-classical form factor at the origin takes non-universal values between 0 and 1. For Veech triangular billiards with angles $(\pi/2, \pi/n, \pi/2-\pi/n)$, the value $K_2(0)=\frac{1}{3}(n+\epsilon(n))/(n-2)$ with $\epsilon(n)=0,2$ or $6$ was found \cite{BogGirSch01}. For a rectangular billiard perturbed by an Aharonov-Bohm flux line, we obtained $K_2(0)=1-\kappa\overline{\alpha}+4\overline{\alpha}^2$ where $\overline{\alpha}\in[0,1/2[$ is the strength of the magnetic flux and $\kappa$ a rational depending on the position of the flux line in the billiard (for irrational positions, $\kappa=3$) \cite{BogGirSch01}. For a circular billiard perturbed by an Aharonov-Bohm flux line, a similar result $K_2(0)=1-\kappa\overline{\alpha}(1-\overline{\alpha})$, with $\kappa\in[0,2]$ an explicit function of the position of the flux, was derived \cite{TheseGir02}. In the case of the barrier billiard, we obtain $K_2(0)=1/2+1/q$. This value depends on the position of the barrier inside the rectangle, which reflects the fact that the structure and the properties of periodic orbits strongly depend on it. This analytical expression for $K_2(0)$ extends previous results to the case of non-Veech polygonal billiards. \section{Periodic orbits in the barrier billiard} \label{section3} The aim of this section is to characterize periodic orbits in a barrier billiard. We first begin by the simple case of a rectangular billiard. \subsection{Periodic orbits in the rectangular billiard} \label{casrectangle} Let us consider a rectangle of area ${\mathcal A}=a\times b$ with Dirichlet boundary conditions. It is easy to work out the density of the lengths of periodic orbits. Any orbit in the rectangle can be unfolded into a straight line in a torus (a rectangle with periodic boundary conditions) of size $2a\times 2b$; a periodic orbit is therefore defined by two integers $M$ and $N$ and has length \begin{equation} \label{lprectangle} l_p=\sqrt{(2 M a)^2+(2 N b)^2}. \end{equation} If we restrict ourselves to $(M,N)$ in the upper right quadrant, each family of periodic orbits occupies an area $4{\mathcal A}$ ($2{\mathcal A}$ for the orbit itself, $2{\mathcal A}$ for its time-reverse). The number ${\mathcal N}(l)$ of pencils of length less than $l$ is just the number of lattice points $(2Ma, 2Nb)$ within a (quarter of a) disk of radius $l$. It has the asymptotic expression ${\mathcal N}(l)\sim\pi l^2/16{\mathcal A}$. The corresponding density of periodic orbits is the derivative of ${\mathcal N}(l)$: \begin{equation} \label{rhol} \rho(l)\sim\frac{\pi l}{8{\mathcal A}}. \end{equation} The density of primitive periodic orbits is given by (see e.g. \cite{BogGirSch01}) \begin{equation} \label{densiterectangle} \rho_{pp}(l)\sim\frac{3 l}{4\pi{\mathcal A}}. \end{equation} We want to obtain a similar result for the barrier billiard. In the rest of this section we investigate the periodic orbits of the barrier billiard, and Section \ref{onpeutremplacer} leads to Equation \eqref{rhoppf} which gives the density of primitive periodic orbits for the barrier billiard. \subsection{The translation surface} \label{transsurf} Instead of studying directly the barrier billiard itself, we will consider the equivalent problem of studying the translation surface associated to this billiard \cite{GutJud00}. \begin{figure}[ht] \begin{center} \epsfig{file=fig1.eps,width=11cm} \end{center} \caption{The barrier billiard and its translation surface} \label{billard} \end{figure} A construction due to Zemlyakov and Katok \cite{ZemKat76} shows that the translation surface associated to a generic rational polygonal billiard is obtained by unfolding the polygon with respect to each of its sides, which means gluing to the initial polygon its images by reflexion with respect to each of its sides and repeating the operation. If the angles of the polygon are $\alpha_i=\pi m_i/n_i$ and $N$ is the least common multiple of the $n_i$, then $2N$ copies of the initial billiard are needed. Here all the angles are multiples of $\pi/2$, therefore only 4 copies are needed, and the translation surface $S$ obtained by this construction is represented in Figure \ref{billard} (right). In this surface, all opposite sides are identified. Any trajectory in the barrier billiard can be unfolded to a straight line on the translation surface. The surface $S$ is of genus 2: there are two singular angles of measure $4\pi$ that we will represent respectively by $z_1$ (a dot in Figure \ref{billard}) and $z_2$ (a cross in Figure \ref{billard}). The two singularities are traditionally called saddles \cite{EskMasZor03} and a geodesic joining them is called a saddle-connexion. \subsection{Periodic orbits in the barrier billiard} \label{pobb} In this subsection, our aim is to describe qualitatively the periodic orbits in the barrier billiard in a given direction. On translation surfaces the periodic orbits occur in pencils, or cylinders, of periodic orbits of same length. These cylinders are bounded by saddle-connexions and are characterized by their length and their height. Let us consider a 'rational direction' on the translation surface $S$: \begin{equation} \label{vecteurs} {\bf v}=(2 M a/q, 2 N b), \end{equation} with $M$ and $N$ two coprime positive integers. The length of the vector ${\bf v}$ is \begin{equation} \label{lgbarr} l_p=\sqrt{(2a M/q)^2+( 2 b N)^2}. \end{equation} Let us label by the integers $k=0$, 1,..., $q-1$ the positions on the translation surface such that the barrier on the ''left'' of the translation surface in Figure \ref{billard} be at position $p-1$ and the barrier on the ''right'' in Figure \ref{billard} be at position $q-p$ (see Figure \ref{unfolded}). Since the opposite sides on the translation surface are identified, then when a trajectory hits the barrier at position $p-1$ it reappears at position $q-p$, and vice-versa. The translation by vector ${\bf v}$ induces a permutation $\sigma_{{\bf v}}$ of the positions $\{0, 1,..., q-1\}$. Let us define \begin{equation} w_1=\min\left\{k\in{\mathbb{N}}; \sigma^{k}_{{\bf v}}(p-1)\in\{p-1, q-p\}\right\} \end{equation} and in the same way \begin{equation} w_2=\min\left\{k\in{\mathbb{N}}; \sigma^{k}_{{\bf v}}(q-p)\in\{p-1, q-p\}\right\}. \end{equation} A translation by the vector $w_1 {\bf v}$ takes $z_1$ to itself and defines a saddle-connexion of length $w_1 l_p$. The second saddle-connexion joining $z_1$ to itself starts at position $q-p$ and its length is $w_2 l_p$.\\ Figure \ref{unfolded} shows, as an example, the two saddle-connexions going from $z_1$ to itself in the direction $(9,2)$ for $p/q=1/3$. The translation by the vector ${\bf v}$ induces the permutation $(012)\mapsto (021)$. One of the saddle-connexions goes from the position 0 to itself and has a length $l_p$; the other goes from position 2 to itself via position 1 and has a length $2l_p$. In any direction, there are always two saddle-connexions going from $z_1$ to itself, and, in the same way, two from $z_2$ to itself. These four saddle-connexions form the boundary of three cylinders of periodic orbits (see Figure \ref{3cylindres}). \begin{figure}[ht] \begin{center} \epsfig{file=fig2.eps,width=13cm} \end{center} \caption{Starting from points of abscissa 0, 1 or 2 ( for $q=3$) in the direction $(M=9,N=2)$, one arrives at 0, 2 or 1: there are two saddle-connexions $0\to 0$ and $2\to 1\to 2$.} \label{unfolded} \end{figure} The lengths of these cylinders are necessarily of the form $w_1 l_p$, $w_2 l_p$ and $(w_1+w_2)l_p$, with $w_i\in\mathbb{N}$, and their heights $(2b/M)h_i$ are such that $h_1+h_3\in\mathbb{Z}$, $h_2+h_3\in\mathbb{Z}$ and $h_3-\sigma \delta_2\in\mathbb{Z}$ for some $\sigma=\pm 1$ and $\delta_2=\{M\alpha\}$, the fractional part of $M\alpha$. For instance in Figure \ref{3cylindres}, there is one cylinder immediately above the saddle-connexion $0\to 0$, one immediately above the saddle-connexion $2\to 1\to 2$, and the third cylinder is below both. \begin{figure}[ht] \begin{center} \epsfig{file=fig3.eps,width=11cm} \end{center} \caption{Three cylinders of periodic orbits bounded by four saddle-connexions in the case $M=4, N=1$.} \label{3cylindres} \end{figure} The results of this section can be summed up as follows. We set \begin{eqnarray} s_1=h_1+h_3\nonumber\\ s_2=h_2+h_3, \end{eqnarray} so that $s_1$ and $s_2$ are integers. Then in each direction ${\bf v}$ defined by $(M,N)$ with $M$ and $N$ coprime, there are three cylinders of periodic orbits of lengths $w_i l_p$ and heights $(2b/M)h_i$ with $i=1,2,3$. The cylinders can be described by the following five characteristic numbers: \begin{itemize} \item[-] the integers $w_1$ and $w_2$ (giving the lengths of the two short cylinders and the length $(w_1+w_2)l_p$ of the long cylinder) \item[-] the real number $h_3$ (giving the height $(2b/M)h_3$ of the long cylinder) \item[-] the integers $s_1$ and $s_2$ (giving the heights $(2b/M)(s_1-h_3)$ and\\ $(2b/M)(s_2-h_3)$ of the short cylinders). \end{itemize} Note that by definition of the $s_i$ we need to have $0<h_3<\min(s_1,s_2)$. Also note that the condition that the sum of the areas of the cylinders be $4{\mathcal A}$ can be expressed as $s_1 w_1+s_2 w_2=q$. \section{Asymptotics for the periodic orbit lengths} \label{onpeutremplacer} Let us define ${\mathcal F}$ as the set of all 4-uples $(w_1,w_2,s_1,s_2)\in {(\mathbb{N}^{*})}^4$ such that $(s_1, s_2)$ are coprime and $s_1 w_1+s_2 w_2=q$. We say that a direction ${\bf v}$ belongs to the family $f\in{\mathcal F}$ if the three cylinders in the direction ${\bf v}$ have the characteristic numbers $w_1,w_2,s_1,s_2$. The goal of this section is to calculate, for a fixed family $f\in{\mathcal F}$ and a fixed interval $I\subset[0,\min(s_1,s_2)[$, the asymptotics for the number ${\mathcal N}^{(q)}_{f,I}(l)$ of directions ${\bf v}$ belonging to the family $f$, such that $||{\bf v}||<l$ and such that the height $h_3$ of the third cylinder in the direction ${\bf v}$ belongs to the interval $I$. \subsection{Counting periodic orbits} The asymptotics for the number ${\mathcal N}^{(q)}(l)$ of cylinders of length less than $l$ have been calculated in \cite{EskMasSch01}. These asymptotics are obtained by applying a Siegel-Veech formula to the space ${\mathcal M}_q(1,1)$ of $q$-fold coverings of the torus with two branch points and area 1. If $V(S)$ is the set of vectors associated with cylinders of periodic orbits on a 'stable' $q$-fold torus cover $S$, then it is shown that there is a constant $\kappa(S)$ depending only on the connected component ${\mathcal M}(S)$ of ${\mathcal M}_q(1,1)$ containing $S$, such that \begin{equation} \label{quadraticasymptotics1} |V(S)\cap B(T)|\sim\pi\kappa T^2, \end{equation} where $|.|$ denotes the cardinal of a set and $B(T)$ is the ball of radius $T$ centered at the origin. The constant $\kappa$ is given by the following Siegel-Veech formula: for any continuous compactly supported $\varphi:{\mathbb{R}}^2\to\mathbb{R}$, \begin{equation} \label{quadraticasymptotics2} \frac{1}{\tilde{\mu}({\mathcal M}(S))}\int_{{\mathcal M}(S)}\hat{\varphi}\ d\tilde{\mu} =\kappa\int_{{\mathbb{R}}^2}\varphi, \end{equation} where $\hat{\varphi}$ is the Siegel-Veech transform of $\varphi$ defined by \begin{equation} \label{quadraticasymptotics3} \hat{\varphi}(S)=\sum_{v\in V(S)}\varphi(v) \end{equation} and $\tilde{\mu}$ is the measure on ${\mathcal M}_q(1,1)$ (Theorem 2.4 of \cite{EskMasSch01}). This Theorem applies to the translation surface constructed from the barrier billiard, provided $S$ be a stable $q$-fold torus cover, which is true only if the height $\alpha$ of the barrier is irrational. It is shown that in this case ${\mathcal M}(S)$ is the set ${\mathcal P}_q(1,1)\subset{\mathcal M}_q(1,1)$ of primitive torus covers. The following asymptotics are then obtained (here we have a factor 1/16 differing from the factor in \cite{EskMasSch01} because of our conventions for the counting of the time-reverse partner of a periodic orbit): \begin{equation} \label{eskin1} {\mathcal N}^{(q)}(l)\sim c\frac{\pi l^2}{16{\mathcal A}}. \end{equation} The constant $c$ is given by \begin{equation} \label{eskin2} c=\frac{q}{N_q}\sum_{r|q}\mu(r) \hspace{-.5cm}\sum_{\genfrac{}{}{0pt}{}{(s_1,s_2)=1}{s_1 u_1+s_2 u_2=q/r}}\hspace{-.5cm} u_1 u_2 (u_1+u_2) \min(s_1,s_2)\left(\frac{1}{u_1^2}+\frac{1}{u_2^2}+\frac{1}{(u_1+u_2)^2}\right) \end{equation} (the gcd of $s, s'$ will be noted either $\gcd(s, s')$ or simply $(s,s')$), and \begin{equation} \label{nqp} N_q=\sum_{r|q}\mu(r)r^2\hspace{-.5cm}\sum_{\genfrac{}{}{0pt}{}{(s_1,s_2)=1}{s_1 u_1+s_2 u_2=q/r}} \hspace{-.5cm}u_1 u_2 (u_1+u_2) \min(s_1,s_2) \end{equation} (Proposition 4.14 of \cite{EskMasSch01}). The constant $N_q$ is the number of primitive covers of degree $q$ of a surface of genus 2 with 2 branch points. \subsection{Siegel-Veech formula} The proof leading to Equation \eqref{eskin1} can be adapted to any subset of $V(S)$ provided it varies linearly under $SL(2, \mathbb{R})$ action, i.e. provided the subset verifies $\forall g\in SL(2, \mathbb{R})$, $V(g S)=g V(S)$ (see Section 2 of \cite{EskMas01} for more detail). To obtain the asymptotics for a fixed pair $F=(f,I)$ with $f=(w_1,w_2,s_1,s_2)\in{\mathcal F}$ and $I$ an interval, $I\subset[0,\min(s_1,s_2)[$, let us define $V_{F}(S)$ the set of vectors ${\bf v}\in{\mathbb{R}}^2$ defined by \eqref{vecteurs}, such that the triple of cylinders in the direction ${\bf v}$ belongs to the family $f$, with $h_3\in I$. Then along the same lines of the proof of Theorem 2.4 in \cite{EskMasSch01}, one can show that when the height $\alpha$ of the barrier is irrational, the translation surface $S$ of the barrier billiard is a stable $q$-fold torus cover and \begin{equation} \label{vb} |V_F(S)\cap B(T)|\sim\pi\kappa_F T^2, \end{equation} where the constant $\kappa_F$ is given by the Siegel-Veech formula \begin{equation} \label{svf2} \frac{1}{\tilde{\mu}({\mathcal P}_q(1,1))}\int_{{\mathcal P}_q(1,1)}\hat{\varphi_F}\ d\tilde{\mu} =\kappa_F\int_{{\mathbb{R}}^2}\varphi_F, \end{equation} with $\hat{\varphi_F}$ the Siegel-Veech transform \begin{equation} \label{svfF} \hat{\varphi_F}(S)=\sum_{v\in V_F(S)}\varphi(v), \end{equation} for some continuous compactly supported $\varphi:{\mathbb{R}}^2\to\mathbb{R}$. \subsection{Asymptotics for a family of periodic orbits} Following the steps leading from the Siegel-Veech formula \eqref{quadraticasymptotics1}-\eqref{quadraticasymptotics3} to the asymptotics \eqref{eskin1}-\eqref{eskin2} in \cite{EskMasSch01}, we can now derive asymptotics for the number of cylinders in each family $(f, I)$. Recall that ${\mathcal N}^{(q)}_{f,I}(l)$ is the number of directions ${\bf v}$ belonging to a family characterized by the numbers $f=(w_1, w_2, s_1, s_2)$, with a height $h_3\in I$, and such that $l_p$ given by Equation \eqref{lgbarr} is less than $l$. (Note that ${\mathcal N}^{(q)}_{f,I}(l)$ is a number of directions and not a number of cylinders.) Let $\rho_{f, I}(l)$ be the corresponding density. According to Equation \eqref{vb}, ${\mathcal N}^{(q)}_{f,I}(l)$ is proportional to $l^2$; we define the constant $c_{f,I}$ by \begin{equation} \label{eskin1f} {\mathcal N}^{(q)}_{f,I}(l)\sim c_{f,I}\frac{\pi l^2}{16{\mathcal A}}. \end{equation} The proof leading to the asymptotics for ${\mathcal N}^{(q)}_{f,I}(l)$ is essentially the same as the proof in \cite{EskMasSch01}, section 4.4, provided we replace the counting functions of the cylinders in \cite{EskMasSch01} by counting functions of directions in which the cylinders belong to the family $(f, I)$ we are interested in. We take $\varphi$ to be the characteristic function of a disc of radius $\epsilon$ in ${\mathbb{R}}^2$. Therefore its Siegel-Veech transform $\hat{\varphi}$, as defined by \eqref{svfF}, counts the number of directions on $S$ in which the cylinders belong to family $(f,I)$ and such that $l_p<\epsilon$. For $\epsilon$ small enough, the Siegel-Veech formula \eqref{svf2} is equivalent to \begin{equation} \label{svf3} \pi\epsilon^2 c_{f, I}=\zeta(2)\frac{1}{\tilde{\mu}({\mathcal P}_q(1,1))} \int_{{\mathcal P}_q(1,1)}\tilde{\varphi_F}\ d\tilde{\mu}, \end{equation} where $\zeta$ is the Riemann Zeta function and \begin{equation} \tilde{\varphi}_{f, I}(S)=\left\{ \begin{array}{cl} 1&\textrm{ \ \ \ if the cylinders in the horizontal direction belong}\cr &\textrm{to the family $f$, if $h_3\in I$ and if}\ ||{\bf v}||<\epsilon\cr 0&\textrm{ \ \ \ otherwise.} \end{array} \right. \end{equation} We define $\chi_{f, I}:{\mathbb{R}}^2\mapsto\mathbb{R}$ by \begin{equation} \chi_{f, I}(v)=\left\{ \begin{array}{cl} 1&\textrm{ \ \ \ if the cylinders in the horizontal direction belong}\cr &\textrm{to the family $f$, if $h_3\in I$ and if}\ ||{\bf v}||<\epsilon\sqrt{q}\cr 0&\textrm{ \ \ \ otherwise.} \end{array} \right. \end{equation} Following \cite{EskMasSch01}, we parametrize ${\mathcal P}_q(1,1)$ and perform the integration in \eqref{svf3}. Part of it can be related to the integral over $\chi_{f, I}$, which is $\int_{{\mathbb{R}}^2}\chi_{f, I}(v)\ dv=\pi\epsilon^2 q$. The integration yields \begin{equation} \int_{{\mathcal P}_q(1,1)}\tilde{\varphi_F}\ d\tilde{\mu}= \frac{\pi\epsilon^2 q}{q\zeta(2)}\sum_{r|(w_1, w_2)} \frac{\mu(r)}{r}w_1 w_2 (w_1+w_2)|I|. \end{equation} From \cite{EskMasSch01} we get $\tilde{\mu}({\mathcal P}_q(1,1))=N_q/q$, with $N_q$ given by \eqref{nqp}. Equation \eqref{svf3} finally gives \begin{equation} \label{eskin2f} c_{f,I}=\frac{q}{N_q}\sum_{r|(w_1, w_2)}\frac{\mu(r)}{r}w_1 w_2 (w_1+w_2)|I|, \end{equation} where $|I|$ is the length of the interval $I$. Equation \eqref{eskin2f} shows that $c_{f,I}$ depends on $I$ only through its length. Is is therefore convenient to introduce the density of directions $p=(M,N)$ corresponding to a family $f$ and such that $l_p<l$ and $h_3=h$: \begin{equation} \label{eskin1fh} {\mathcal N}^{(q)}_{f,h}(l)\sim c_{f,h}\frac{\pi l^2}{16{\mathcal A}}, \end{equation} with $c_{f,h}$ given by \begin{equation} \label{eskin2fh} c_{f,h}=\frac{q}{N_q}\sum_{r|q}\frac{\mu(r)}{r}w_1 w_2 (w_1+w_2)\theta_f(h) \end{equation} for any family $f=(w_1, w_2, s_1, s_2)$ of ${\mathcal F}$ and $h\in \mathbb{R}$. The function $\theta_f$ is the characteristic function of the interval $[0,\min(s_1,s_2)[$. The density of primitive periodic orbit lengths for the family $f\in{\mathcal F}$ and $h\in\mathbb{R}$ is \begin{equation} \label{rhoppf} \rho_{pp,f, h}(l)\sim c_{f,h} \frac{3 l}{4\pi{\mathcal A}}. \end{equation} It is easy to verify that the expression \eqref{eskin1fh} of ${\mathcal N}^{(q)}_{f,h}(l)$ is consistent with the total number ${\mathcal N}^{(q)}(l)$ of pencils of periodic orbits with length less than $l$. This comes from the fact that any pencil of periodic orbits contributing to ${\mathcal N}^{(q)}(l)$ belongs to a certain family $f$ and has a length $w_i l_{p}\leq l$, which implies that $l_{p}\leq l/w_i$. Therefore \begin{eqnarray} {\mathcal N}^{(q)}(l)&=&\sum_{f\in{\mathcal F}}\int dh\sum_{i=1}^{3} {\mathcal N}^{(q)}_{f,h}(l/w_i)\nonumber\\ &\sim&\sum_{f\in{\mathcal F}}\int dh\ c_{f,h}\frac{\pi l^2}{16{\mathcal A}} \left(\frac{1}{w_1^2}+\frac{1}{w_2^2}+\frac{1}{(w_1+w_2)^2}\right). \end{eqnarray} Using Equation \eqref{eskin2fh}, we obtain, after integration over $h$, \begin{eqnarray} {\mathcal N}^{(q)}(l)&\sim&\frac{\pi l^2}{16{\mathcal A}} \frac{q}{N_q}\sum_{\genfrac{}{}{0pt}{}{(s_1,s_2)=1}{s_1 w_1+s_2 w_2=q}} \sum_{r|(w_1,w_2)}\frac{\mu(r)}{r} w_1 w_2 (w_1+w_2)\nonumber\\ &&\times\min(s_1,s_2) \left(\frac{1}{w_1^2}+\frac{1}{w_2^2}+\frac{1}{(w_1+w_2)^2}\right). \end{eqnarray} Making the substitution $w_i=r u_i$ and inverting the two sums, we get exactly the expression given by Equations (\ref{eskin1}) and (\ref{eskin2}). \section{Calculation of the form factor at $\tau=0$} \label{calculff} \subsection{Definitions} \label{section2} The spectrum $\{E_n, n\in\mathbb{N}\}$ of a quantum billiard can be described by the density \begin{equation} d(E)\equiv\sum_n \delta(E-E_n). \end{equation} The two-point correlation form factor is defined as the Fourier transform of the two-point correlation function of the density of states: \begin{equation} \label{formfactor} K_2(\tau)=\int_{-\infty}^{\infty}\frac{d\epsilon}{\bar{d}}\langle d(E+\epsilon/2)d(E-\epsilon/2)\rangle_{\textrm{c}} e^{2 i \pi \bar{d} \tau\epsilon}. \end{equation} Here the product of the densities is averaged over an energy window of width $\Delta E\gg 1/\bar{d}$ centered around $E=k^2$ and such that $\Delta E \ll E$. If ${\mathcal A}$ is the area of the billiard, $\bar{d}={\mathcal A}/4\pi$ is the non-oscillating part of the density of states. The subscript c means that one only considers the connected part of the correlation function. It can be argued that in the case of pseudo-integrable systems, the leading term of the semiclassical expansion of $K_2(\tau)$ at small argument ($\tau\to 0$) is given in the diagonal approximation by the contribution of periodic orbits only: $K_2(\tau)=K^{\textrm{diag}}(\tau)+O(\tau)$, with \begin{equation} K^{\textrm{diag}}(\tau)=\frac{1}{8\pi^2\bar{d}} \sum_{p} \frac{|S_{p}|^2}{l_p} \delta (l_{p}-4\pi k \bar{d} \tau) \label{k0} \end{equation} (see \cite{BogGirSch01} for the derivation of this expression, based on heuristic arguments). The sum is performed over all pencils of periodic orbits $p$ of length $l_p$. In general, there can be several pencils having exactly the same length: in Equation \eqref{k0}, $S_p$ is the sum of the areas occupied by all pencils having, when (possibly) multiply repeated, a length $l_p$. The aim of the present section is to calculate the semiclassical form factor at small arguments \eqref{k0}, using the result (\ref{rhoppf}) for the distribution of pencils of periodic orbits in the barrier billiard. Let us take a $C^{\infty}$, compactly supported test function and integrate the distribution $K^{\textrm{diag}}(\tau)$ over $\tau$. If the density of periodic pencils depends linearly on $l$ (as is the case for the barrier billiard or the rectangular billiard), the integration over families of periodic orbits yields $K^{\textrm{diag}}(\tau)=\lambda\Theta(\tau)$, where $\lambda$ is a constant and $\Theta$ the Heaviside step function. In such a case, we define $\overline{K_2(0)}=\lambda$. As an introduction, we first deal with the simpler case of a rectangular billiard. \subsection{Rectangular billiard} In the case of the rectangular billiard, discussed in section \ref{casrectangle}, the periodic orbits have lengths $n l_{pp}$, where $l_{pp}$ is given by \eqref{lprectangle} with $(M,N)$ coprime, and $n\in\mathbb{N}$ is the repetition number. The area of each pencil of primitive periodic orbits $pp$ is $A_{pp}=4{\mathcal A}$. When the sides $a$ and $b$ of the rectangle are incommensurable, there is only one pencil of length $l_{pp}$ and therefore in Equation (\ref{k0}) $S_p=4{\mathcal A}$. Equation (\ref{k0}) becomes \begin{equation} K^{\textrm{diag}}(\tau)=\frac{1}{8\pi^2\bar{d}} \sum_{pp}\sum_{n}\frac{|4{\mathcal A}|^2}{n^2 l_{pp}} \delta (l_{pp}-4\pi k \bar{d} \tau/n) \end{equation} hence (using the fact that $\bar{d}={\mathcal A}/4\pi$ and turning the sum over $pp=(M,N)$ with $M$ and $N$ coprime into an integral over $l$ with density $\rho_{pp}(l)$) \begin{equation} \label{k2rec} K^{\textrm{diag}}(\tau)=\frac{8{\mathcal A}}{\pi}\sum_n \int_{0}^{\infty}dl\ \frac{1}{n^2 l}\rho_{pp}(l) \delta (l-4\pi k \bar{d} \tau/n) \end{equation} The density $\rho_{pp}(l)$ of periodic orbits is given by Equation (\ref{densiterectangle}) and yields $K^{\textrm{diag}}(\tau)=1$, as expected for integrable systems. \subsection{Barrier billiard} In the case of the barrier billiard, the periodic orbits have a length of the form $n w l_p$ with $l_p$ given by \eqref{lgbarr}: here the primitive length is $w l_p$ and $n$ is the repetition number. Two pencils of periodic orbits $p$ and $p'$ have the same length provided there exist repetition numbers $n$ and $n'$ such that $n w l_{p}=n' w' l_{p'}$. When $a$ and $b$ are incommensurable, this implies $p=p'$, i.e. two pencils can have same length only if they are in the same direction. For a given direction $(M,N)$ with $M$ and $N$ coprime (which will now be labeled by $p$), there are three cylinders of area ${\mathcal A}_i$ and length $w_i l_{p}$, $1\leq i \leq 3$, and therefore $w,w'$ belong to the set $\{w_1, w_2, w_1+w_2\}$. Equation (\ref{k0}) becomes \begin{equation} \label{dpo2barr} K^{\textrm{diag}}(\tau)=\frac{1}{8\pi^2\bar{d}} \sum_{p}\sum_{n}\frac{|S_{p,n}|^2}{n l_p} \delta (n l_{p}-4\pi k \bar{d} \tau) \end{equation} where $l_p$ is given by \eqref{lgbarr} and $S_{p,n}$ is the sum over the ${\mathcal A}_i$ corresponding to a $w_i$ which divides $n$: \begin{equation} \label{sppn} S_{p,n}=\sum_{i=1}^{3}{\mathcal A}_i \delta_{w_i|n}, \end{equation} with $\delta_{r|t}=1$ if $r$ divides $t$, 0 otherwise. Each area ${\mathcal A}_i$ is equal to $(2b/M)h_i\times (w_i l_{p})\cos\varphi_{p}$ ($\varphi_{p}$ is the angle between the orbit and the horizontal). This can be rewritten as \begin{equation} \label{area} {\mathcal A}_i=\frac{4{\mathcal A}}{q}h_i w_i \end{equation} (note that since $\sum_ih_iw_i=s_1w_1+s_2 w_2=q$, one has $\sum_i{\mathcal A}_i=4{\mathcal A}$, i.e. the total area of the translation surface, as expected). Therefore $S_{p,n}$ only depends of the five numbers $f=(w_1, w_2,s_1,s_2)$ and $h_3$, and can be rewritten: \begin{equation} S_{f,h_3,n}=\frac{4{\mathcal A}}{q}\left[(s_1-h_3)w_1\delta_{w_1|n}+(s_2-h_3)w_2\delta_{w_2|n} +h_3(w_1+w_2)\delta_{(w_1+w_2)|n}\right]. \end{equation} The sum (\ref{dpo2barr}) over all periodic orbits can be partitioned into sums running over primitive pencils of periodic orbits $p(f,h)$ belonging to a family $f$ with a height of the long cylinder in $[h, h+dh[$; \eqref{dpo2barr} becomes \begin{equation} K^{\textrm{diag}}(\tau)=\sum_{f\in{\mathcal F}}\int dh\sum_{p(f,h)}\sum_{n} \frac{|S_{p,n}|^2}{8\pi^2 n^2 l_{p} \bar{d}} \delta(l_{p}-\frac{4\pi k \bar{d} \tau}{n}). \end{equation} Each of the sums corresponding to a family $f$ can be replaced, as in (\ref{k2rec}), by an integral with density $\rho_{pp,f,h}(l)$, and $S_{p,n}$ by $S_{f,h,n}$: \begin{equation} K^{\textrm{diag}}(\tau)=\sum_{f\in{\mathcal F}}\int dh\sum_{n} \frac{|S_{f,h,n}|^2}{8\pi^2 n^2\bar{d}} \int_{0}^{\infty}dl\ \frac{\rho_{pp,f,h}(l)}{l} \delta(l-\frac{4\pi k \bar{d} \tau}{n}). \end{equation} Replacing the density $\rho_{pp,f,h}$ by its expression (\ref{rhoppf}), the integration over $l$ becomes straightforward and yields $K^{\textrm{diag}}(\tau)=\overline{K_2(0)}\Theta(\tau)$, where \begin{eqnarray} \overline{K_2(0)}=\frac{1}{q^2}\frac{6}{\pi^2}\sum_{n}\frac{1}{n^2}\sum_{f\in{\mathcal F}}\int dh \left[(s_1-h)w_1\delta_{w_1|n}\right.\\ \nonumber +\left.(s_2-h)w_2\delta_{w_2|n} +h(w_1+w_2)\delta_{(w_1+w_2)|n}\right]^2 c_{f,h} \end{eqnarray} (we have used the fact that $\bar{d}={\mathcal A}/4\pi$). Expanding the square, we can perform the summation over $n$, using the identity \begin{equation} \sum_{n=1}^{\infty}\frac{\delta_{w_1|n}\delta_{w_2|n}}{n^2}=\frac{\pi^2}{6} \frac{\gcd(w_1,w_2)^2}{w_1^2 w_2^2}. \end{equation} The form factor can therefore be written, after simplifications using the fact that $\gcd(w_1,w_1+w_2)=\gcd(w_2,w_1+w_2)=\gcd(w_1,w_2)$, as \begin{eqnarray} \overline{K_2(0)}&=&\frac{1}{q^2}\sum_{f\in{\mathcal F}}\int dh\ c_{f,h} \left[3h^2-2\left(s_1+s_2+q\frac{\gcd(w_1,w_2)^2}{w_1 w_2 (w_1+w_2)}\right)h\right.\nonumber\\ &+&\left.s_1^2+s_2^2+2 s_1 s_2\frac{\gcd(w_1,w_2)^2}{w_1 w_2}\right]. \end{eqnarray} Replacing the weight $c_{f,h}$ by its expression (\ref{eskin2fh}), we can easily perform the integration over $h$, which consists of terms of the form \begin{equation} \int_{0}^{\min(s_1, s_2)}dh\ h^{\nu}=\frac{\min(s_1,s_2)^{\nu+1}}{\nu+1} \end{equation} for $\nu=0,1,2$. The form factor becomes \begin{eqnarray} \overline{K_2(0)}&=&\frac{1}{q N_q}\sum_{\genfrac{}{}{0pt}{}{(s_1,s_2)=1}{s_1 w_1+s_2 w_2=q}} \sum_{r|(w_1,w_2)}\frac{\mu(r)}{r} w_1 w_2 (w_1+w_2) \left[\min(s_1,s_2)^3\right.\nonumber\\ &-&\left(s_1+s_2+q\frac{\gcd(w_1,w_2)^2}{w_1 w_2 (w_1+w_2)}\right)\min(s_1,s_2)^2\nonumber\\ &+&\left.\left(s_1^2+s_2^2+2 s_1 s_2\frac{\gcd(w_1,w_2)^2}{w_1 w_2}\right)\min(s_1,s_2)\right]. \end{eqnarray} This sum can be evaluated with some cumbersome arithmetic manipulations; the calculation is given in the Appendix, and the final result is unexpectedly simple: \begin{equation} \label{resultatfinal} \overline{K_2(0)}=\frac{1}{2}+\frac{1}{q}. \end{equation} There are several comments to make concerning this value. First, it is close to the result corresponding to semi-Poisson statistics $\overline{K_2(0)}=1/2$ \cite{BogGirSch01}. This result is not valid for $q=2$, since in that case there is an additional symmetry in the billiard, with respect to the barrier, and the spectrum has to be desymmetrized. The calculation in this case has been done in \cite{Wie02} for a height of the barrier equal to $b/2$ (half the height of the rectangle), and yields $\overline{K_2(0)}=1/2$. The calculation for $q=2$ and a barrier with any height has been done in \cite{TheseGir02} using a different method, and also yields $\overline{K_2(0)}=1/2$. The result (\ref{resultatfinal}) is similar to previously obtained results \cite{BogGirSch01} for rational polygonal billiards having the Veech property. For instance for triangular billiards with angles $(\pi/2, \pi/n, \pi/2-\pi/n)$ the form factor at the origin was found to be between $1/3$ and $3/5$ \cite{BogGirSch01}. Here the form factor lies between $1/2$ and $5/6$, which again is close to the semi-Poisson result. \section*{Acknowledgments} The author thanks Professor Alex Eskin for helpful discussions. The funding of the Leverhulme trust and the Department of Physics of the University of Bristol, where most of this work has been done, are gratefully acknowledged for their support. The funding of post-doctoral CNRS fellowship and the theoretical physics laboratory of the University of Toulouse have made the completion of this work possible. \section*{Appendix} In this appendix, we want to evalute the quantity \begin{equation} \label{depart} \overline{K_2(0)}=\frac{1}{q N_q}\sum_{\genfrac{}{}{0pt}{}{(s_1,s_2)=1}{s_1 w_1+s_2 w_2=q}} \sum_{r|(w_1,w_2)}\frac{\mu(r)}{r}f(s_1, s_2, w_1, w_2,q), \end{equation} where \begin{eqnarray} \label{homogeneite} f(s_1, s_2, w_1, w_2,q)&=& w_1 w_2 (w_1+w_2)\left[\min(s_1,s_2)^3\right.\nonumber\\ &-&\left.\left(s_1+s_2+q\frac{\gcd(w_1,w_2)^2}{w_1 w_2 (w_1+w_2)}\right)\min(s_1,s_2)^2\right.\nonumber\\ &+&\left.\left(s_1^2+s_2^2+2 s_1 s_2\frac{\gcd(w_1,w_2)^2}{w_1 w_2}\right)\min(s_1,s_2)\right]. \end{eqnarray} The function $f$ is homogeneous, in the sense that it verifies \begin{equation} f(s_1, s_2,\lambda w_1,\lambda w_2, \lambda q)=f(\lambda s_1,\lambda s_2,w_1, w_2,\lambda q) =\lambda^3 f( s_1, s_2,w_1,w_2,q). \end{equation} In (\ref{depart}), the first sum goes over all integers $w_i\geq 1$ and $s_i\geq 1$, $i=1,2$, verifying $s_1 w_1+s_2 w_2=q$ and $\gcd(s_1,s_2)=1$. The number $N_q$ is given by (\ref{nqp}). The first step is to exchange the sum over $(s_i, w_i)$ and the sum over $r$ in (\ref{depart}), and substitute $w_i=r u_i$: using the homogeneity of $f$, we get \begin{equation} \label{stari} \overline{K_2(0)}=\frac{1}{q N_q}\sum_{r|q}\mu(r)r^2 \sum_{\genfrac{}{}{0pt}{}{(s_1,s_2)=1}{s_1 u_1+s_2 u_2=q/r}}f(s_1, s_2, u_1, u_2, \frac{q}{r}), \end{equation} To get rid of the co-primality condition on $(s_1, s_2)$ we use the exclusion-inclusion principle, which for any function $\varphi$ gives \begin{equation} \label{excluinclu} \sum_{(s,s')=1}\varphi(s,s')=\sum_{s, s'=1}^{\infty}\sum_{t=1}^{\infty}\mu(t)\varphi(t s, t s'). \end{equation} This allows to rewrite the form factor as \begin{eqnarray} \overline{K_2(0)}&=&\frac{1}{q N_q}\sum_{r|q}\mu(r)r^2\sum_{t=1}^{\infty}\mu(t) \hspace{-.5cm}\sum_{t s_1 u_1+t s_2 u_2=q/r}f(t s_1, t s_2,u_1, u_2, \frac{q}{r})\hspace{-.5cm}\\ &=&\frac{1}{q N_q}\sum_{r|q}\mu(r)r^2\sum_{t|q}\mu(t)t^3 \hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=q/(rt)}f(s_1, s_2, u_1, u_2, \frac{q}{r t})\hspace{-.5cm}\nonumber \end{eqnarray} Here the sum over $t$ from 1 to $\infty$ has been replaced by a sum over $t|q$ since for all the other values of $t$ there is no value of $(s_1, s_2,u_1, u_2)$ fulfilling the condition $t s_1 u_1+t s_2 u_2=q/r$. Again, the homogeneity of $f$ (Equation (\ref{homogeneite})) has been used. Setting $d=r t$ we get \begin{eqnarray} \label{starf} \overline{K_2(0)}=\frac{1}{q N_q}\sum_{d|q}\left(\sum_{t|d}\mu(t)\mu(\frac{d}{t}) d^2 t\right) \sum_{s_1 u_1+s_2 u_2=q/d}f(s_1, s_2, u_1, u_2, \frac{q}{d}).\nonumber \end{eqnarray} We need to evaluate \begin{equation} \label{k0total} \overline{K_2(0)}=\frac{1}{q N_q}\sum_{d|q}\left(\sum_{t|d}\mu(t)\mu(\frac{d}{t}) d^2 t\right)(G_{q/d}+H_{q/d}), \end{equation} where \begin{eqnarray} G_n&\equiv&\hspace{-.5cm} \sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2 (u_1+u_2)\left[\min(s_1,s_2)^3\right.\\ &-&\left.\left(s_1+s_2\right)\min(s_1,s_2)^2+\left(s_1^2+s_2^2+2 s_1 s_2\right)\min(s_1,s_2) \right]\nonumber \end{eqnarray} and \begin{eqnarray} H_n&\equiv&\hspace{-.5cm} \sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2 (u_1+u_2) \left[-q\frac{\gcd(u_1,u_2)^2}{u_1 u_2 (u_1+u_2)}\min(s_1,s_2)^2\right.\nonumber\\ &&\hspace{3cm}+\left.2 s_1 s_2\frac{\gcd(u_1,u_2)^2}{u_1 u_2}\min(s_1,s_2)\right]. \end{eqnarray} The quantities $G_n$ and $H_n$ will be evaluated separately. This evaluation will require the use of a theorem proved in \cite{HuaOuSpeWil02}:\\ {\bf Theorem.} Let $f:\mathbb{Z}$$^{4}\rightarrow \mathbb{C}$ such that \begin{equation} f(a,b,x,y)-f(x,y,a,b)=f(-a, -b, x,y)-f(x,y,-a,-b) \end{equation} for all integers $a,b,x$ and $y$. Then for $n\in\mathbb{N}$, $n\geq 1$, \begin{eqnarray} \label{theoreme} \sum_{\genfrac{}{}{0pt}{}{a,b,x,y\geq 1}{a x+b y=n}} \left[f(a,b,x,-y)-f(a,-b,x,y)+f(a,a-b,x+y,y)\right.\nonumber\\ -\left. f(a,a+b,y-x,y)+f(b-a, b, x, x+y)-f(a+b, b, x, x-y)\right]\nonumber\\ =\sum_{d|n}\sum_{x=1}^{d-1}\left[ f(0, \frac{n}{d}, x, d)+f(\frac{n}{d},0,d, x)+f(\frac{n}{d},\frac{n}{d},d-x,-x)\right.\nonumber\\ -\left. f(x, x-d, \frac{n}{d},\frac{n}{d})-f(x,d,0,\frac{n}{d})-f(d,x,\frac{n}{d},0)\right]. \end{eqnarray} \subsection*{a. Evaluation of $G_n$} We can immediately point out that the identity \begin{equation} \min(s_1,s_2)^2-(s_1+s_2)\min(s_1,s_2)=-s_1 s_2, \end{equation} valid for any integers $s_1$ and $s_2$, allows to simplify $G_n$. We now need to evaluate the sum \begin{equation} G_n=\hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm} u_1 u_2 (u_1+u_2)\min(s_1,s_2)\left(s_1^2+s_2^2-s_1 s_2\right). \end{equation} for any integer $n$. Writing $\min(a,b)=\frac{1}{2}(a+b-|a-b|)$, we have \begin{eqnarray} \label{gn2sommes} G_n&=&\frac{1}{2}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2\left[ s_1^3 u_2+s_2^3 u_1-\frac{1}{3}(s_1^3 u_1+s_2^3 u_2)\right.\\ &-&\left.\vphantom{\frac{1}{2}}(u_1+u_2)(s_1^2+s_2^2-s_1 s_2)|u_1-u_2|\right] +\frac{2}{3}\hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2(s_1^3 u_1+s_2^3 u_2).\nonumber \end{eqnarray} The first sum in (\ref{gn2sommes}) can be evaluated by applying Theorem (\ref{theoreme}) to the function \begin{equation} f(a,b,x,y)=\frac{1}{3}\left(x y-\frac{|x y|}{2}\right)\left|(a-b)(x-y)\right|(a^2+b^2-a b) \end{equation} and is equal to \begin{equation} \frac{n^2 (n-1)}{18}\sum_{d|n}d. \end{equation} The second sum in (\ref{gn2sommes}) can be evaluated by applying Theorem (\ref{theoreme}) to the function $f(a,b,x,y)=b^2y^4-b^2 x y^3$ (see \cite{HuaOuSpeWil02}). It gives \begin{equation} \frac{4}{3}\sum_{a x+b y=n}\hspace{-.3cm}a^3 x^2 y =\frac{n^2}{18}\sum_{d|n}\left(3d^3+(1-4n)d\right). \end{equation} Finally we get \begin{equation} \label{gn} G_n=\frac{n^2}{6}\sum_{d|n}\left(d^3-n d\right). \end{equation} If we now evaluate the quantity \begin{eqnarray} \sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm} u_1 u_2 (u_1+u_2)\min(s_1,s_2)=\frac{1}{2}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2 \left[s_1 u_2+s_2 u_1\vphantom{\frac{1}{2}}\right.\nonumber\\ \left.-\frac{1}{3}(s_1 u_1+s_2 u_2) -(u_1+u_2)|u_1-u_2|\right] +\frac{2n}{3}\hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2, \end{eqnarray} the first sum is given by Theorem (\ref{theoreme}) applied to the function \begin{equation} f(a,b,x,y)=\frac{1}{3}\left(x y-\frac{|x y|}{2}\right)\left|(a-b)(x-y)\right| \end{equation} and the second one is given by Theorem (\ref{theoreme}) applied to the function\\ $f(a,b,x,y)=n x y/3$; altogether, this gives \begin{equation} \sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm} u_1 u_2 (u_1+u_2)\min(s_1,s_2)=\frac{n}{3}\sum_{d|n}\left(d^3-n d\right). \end{equation} Together with Equation (\ref{gn}) we get \begin{equation} G_{n}=\frac{n}{2}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm}u_1 u_2 (u_1+u_2)\min(s_1,s_2). \end{equation} \subsection*{b. Evaluation of $H_n$} We want to evaluate \begin{eqnarray} \label{departH} H_n&=&\hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=n}\hspace{-.5cm} u_1 u_2 (u_1+u_2)\min(s_1,s_2)\hspace{4cm}\nonumber\\ &&\hspace{2cm}\left(-n\frac{\min(s_1,s_2)}{ u_1 u_2 (u_1+u_2)}+\frac{2 s_1 s_2}{u_1 u_2}\right) \gcd(u_1, u_2)^2 \end{eqnarray} for any integer $n$. Summing over all the possible values $r$ of the $\gcd$ of $u_1$ and $u_2$, and substituting $u_i=r v_i$, we have \begin{equation} H_n=\sum_{r|n}\hspace{-.2cm} \sum_{\genfrac{}{}{0pt}{}{(v_1,v_2)=1}{s_1 v_1+s_2 v_2=n/r}}\hspace{-.2cm} r^3 v_1 v_2 (v_1+v_2)\min(s_1,s_2) \left(-\frac{n}{r}\frac{\min(s_1,s_2)}{v_1 v_2 (v_1+v_2)}+\frac{2 s_1 s_2}{v_1 v_2}\right). \end{equation} Then, as before, the co-primality condition can be expressed by a sum over $t$ (see Equation (\ref{excluinclu})). Restricting the sum over $t$ as before, we get \begin{eqnarray} H_n&=&\sum_{r|n}\sum_{t|n}\mu(t)r^3 t \hspace{-.5cm}\sum_{s_1 v_1+s_2 v_2=n/(r t)}\hspace{-.5cm} v_1 v_2 (v_1+v_2)\min(s_1,s_2)\hspace{2cm}\nonumber\\ &&\hspace{5cm}\left(-\frac{n}{r t}\frac{\min(s_1,s_2)}{v_1 v_2 (v_1+v_2)}+\frac{2 s_1 s_2}{v_1 v_2}\right). \end{eqnarray} Setting $d=rt$ we get \begin{eqnarray} \label{arriveeH} H_n&=&\sum_{d|n}\left(\sum_{t|d}\mu(t)\frac{d^3}{t^2}\right)\sum_{s_1 v_1+s_2 v_2=n/d}\hspace{-.5cm} v_1 v_2 (v_1+v_2)\min(s_1,s_2)\hspace{2cm}\nonumber\\ &&\hspace{5cm}\left(-\frac{n}{d}\frac{\min(s_1,s_2)}{v_1 v_2 (v_1+v_2)}+\frac{2 s_1 s_2}{v_1 v_2}\right). \end{eqnarray} Let us now evaluate, for any integer $m$, the quantity \begin{eqnarray} K_m&=&\sum_{s_1 v_1+s_2 v_2=m}v_1 v_2 (v_1+v_2)\min(s_1,s_2) \left(-m\frac{\min(s_1,s_2)}{v_1 v_2 (v_1+v_2)}+\frac{2 s_1 s_2}{v_1 v_2}\right)\nonumber\\ &=&-m\hspace{-.5cm}\sum_{s_1 v_1+s_2 v_2=m}\hspace{-.5cm}\min(s_1,s_2)^2 +2\hspace{-.5cm}\sum_{s_1 v_1+s_2 v_2=m}\hspace{-.5cm}s_1 s_2 (v_1+v_2)\min(s_1,s_2). \end{eqnarray} Let \begin{eqnarray} L_m&=&\hspace{-.5cm}\sum_{s_1 v_1+s_2 v_2=m}\hspace{-.5cm}(2 s_1 s_2-v_1 v_2)(v_1+v_2)\min(s_1,s_2)\\ &=&\hspace{-.5cm}\sum_{s_1 v_1+s_2 v_2=m}\hspace{-.5cm} (2 v_1 v_2(s_1+s_2)\min(v_1,v_2)-v_1 v_2(v_1+v_2)\min(s_1,s_2))\nonumber \end{eqnarray} after exchanging $(s_1, s_2)$ and $(v_1, v_2)$ in the first half of the right member. Writing $\min(a,b)=\frac{1}{2}(a+b-|a-b|)$ and applying Theorem (\ref{theoreme}) to the function \begin{equation} f(a,b,x,y)=-\frac{1}{2}\left|a b (a-b)(x-y)\right| \end{equation} one gets \begin{equation} L_m=m\sum_{d|m}\sum_{x=1}^{d-1}x(d-x). \end{equation} Applying Theorem (\ref{theoreme}) to the function \begin{equation} f(a,b,x,y)=\frac{a b-\left|a b\right|}{2} \end{equation} one gets \begin{equation} \sum_{s_1 v_1+s_2 v_2=m}\hspace{-.5cm}\min(s_1,s_2)^2=\sum_{d|m}\sum_{x=1}^{d-1}x(d-x). \end{equation} This proves that \begin{equation} K_m=\hspace{-.5cm}\sum_{s_1 v_1+s_2 v_2=m}\hspace{-.5cm}v_1 v_2(v_1+v_2)\min(s_1,s_2) \end{equation} and therefore \begin{equation} H_n=\sum_{d|n}\left(\sum_{t|d}\mu(t)\frac{d^3}{t^2}\right) \sum_{s_1 v_1+s_2 v_2=n/d}\hspace{-.5cm}v_1 v_2 (v_1+v_2)\min(s_1,s_2). \end{equation} \subsection*{c. Calculation of $\overline{K_2(0)}$} The evaluation of (\ref{k0total}) will require to introduce the functions \begin{equation} f(n)=\frac{\mu(n)}{n}\ \ \ \ \ \ \ \textrm{and}\ \ \ \ g(n)=\sum_{d|n}\frac{\mu(d)}{d^2} \end{equation} For $f_1$ and $f_2$ two arithmetic functions, the Dirichlet convolution is defined by \begin{equation} f_1*f_2(n)=\sum_{d|n}f_1(d)f_2(\frac{n}{d}). \end{equation} Replacing the expressions found for $G_{q/d}$ and $H_{q/d}$ in Equation (\ref{k0total}) we get \begin{eqnarray} \label{deuxtermes} \overline{K_2(0)}&=&\frac{1}{q N_q}\sum_{d|q}(f*\mu)(d)d^3 \left\{\sum_{s_1 u_1+s_2 u_2=q/d}\hspace{-.3cm}\left(\frac{q}{2d}\right)u_1 u_2 (u_1+u_2)\min(s_1,s_2) \right.\nonumber\\ &+&\left.\sum_{d'|q/d}d'^3g(d') \hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=q/dd'}\hspace{-.5cm}u_1 u_2 (u_1+u_2)\min(s_1,s_2)\right\}. \end{eqnarray} Rewriting the constant $N_q$ given by (\ref{nqp}), using the inclusion-exclusion principle and following the steps from Equations (\ref{stari}) to (\ref{starf}), we get \begin{equation} N_q=\sum_{d|q}(f*\mu)(d)d^2 \hspace{-.5cm}\sum_{s_1 u_1+s_2 u_2=q/d}\hspace{-.5cm}u_1 u_2 (u_1+u_2)\min(s_1,s_2). \end{equation} We see that the first term in (\ref{deuxtermes}) is equal to 1/2. If we set $\delta=d d'$, the second term gives \begin{equation} \label{2emeterm} \frac{1}{q N_q}\sum_{\delta|q}(\frac{\delta}{d})^3\left(\sum_{d|\delta}(f*\mu)(d)g(\frac{\delta}{d})\right) \sum_{s_1 u_1+s_2 u_2=q/\delta}\hspace{-.5cm}u_1 u_2 (u_1+u_2)\min(s_1,s_2). \end{equation} But \begin{equation} \sum_{d|\delta}(f*\mu)(d)g(\frac{\delta}{d})=[(f*\mu)*g](\delta)=[f*(\mu*g)](\delta) \end{equation} by associativity of Dirichlet convolution, and \begin{equation} (\mu*g)(\delta)=\frac{\mu(\delta)}{\delta^2} \end{equation} by Moebius inversion formula. Finally the term (\ref{2emeterm}) simplifies to $1/q$, which completes the proof.
{ "timestamp": "2005-03-31T14:55:13", "yymm": "0503", "arxiv_id": "nlin/0503067", "language": "en", "url": "https://arxiv.org/abs/nlin/0503067" }
\section{Introduction} Quantum entanglement and quantum nonlocality are two striking aspects of quantum mechanics. They are introduced by Einstein, Podolsky, and Rosen (EPR) in their famous paper \cite{epr}. The relationship between them has been paid much attention. They play an essential role in the modern understanding of quantum phenomena, and quantum information transmission and processing. Bell proposed a remarkable inequality imposed by a local hidden variable theory \cite{bel}, which enables a quantitative test on quantum nonlocality. The quantum nonlocality test can be performed on an entangled system composed of two coherent systems. This entangled system can be used as a quantum entangled channel for quantum information transfer. Numerous theoretical studies and experimental demonstrations have been carried out to understand nonlocal properties of quantum states. Various versions of Bell's inequality \cite{chs,ch} are proposed. Gisin and Peres found pairs of observable whose correlations violate Bell's inequality for a discrete $N$-dimensional entangled state \cite{gis} . Banaszek and W\'{o}dkiewicz studied Bell's inequality for continuous-variable states in terms of Wigner representation in phase space based upon parity measurement and displacement operation. Recently, Chen {\it et al.} studied Bell's inequality of continuous-variable states \cite{chen1} using their newly defined pseudospin Bell operators, and they showed that the EPR state can maximally violate Bell's inequality in their framework. Recently, $N$-photon entangled states and their superposition states are paid much attention. They are widely used to realize quantum lithography \cite{kuan,kok,bjo}, super-resolving phase measurements \cite{mit}, and quantum teleportation \cite{coc}. In this letter, we study quantum entanglement and quantum nonlocality of the following $N$-photon entangled states and their superpositions \begin{eqnarray} \label{e1} |\psi_{N m}\rangle && =\mathcal{N}_{m}[\cos\gamma|N-m\rangle_{1}|m\rangle_{2} \nonumber \\ &&+e^{i\theta_{m}}\sin\gamma|m\rangle_{1}|N-m\rangle_{2}], \end{eqnarray} where $m$ takes its values from zero to $N$, the normalization factor is given by $\mathcal{N}_{m}^{-2}=1+\cos\theta_{m}\sin 2\gamma \delta_{N,2m}$ with $\gamma$ and $\theta_{m}$ being an entanglement angle and a relative phase, respectively. We will calculate the von Neumann entropy and study the Bell's inequality for quantum states $|\psi_{N m}\rangle$ and their superposition states. We will show that quantum superpositions for the two modes may increase the amount of entanglement, and $N$-photon entangled states can maximally violate Bell's inequality. This letter is organized as follows. In Sec. II, we study quantum entanglement of $N$-photon entangled states and their superposition states through analyzing their von Neumann entropy. Quantum nonlocality of $N$-photon entangled states and their superposition states is investigated through discussing the violation of the Bell's inequality in the pseudospin Bell-operator formalism developed by Chen {\it et al.} \cite{chen1} in Sec. III. The last section is devoted to summary and conclusion. \section{Quantum entanglement for $N$-photon entangled states} In this section, we study properties of quantum entanglement of the $N$-photon entangled states given by Eq. (\ref{e1}). In order to this, we consider the nontrivial case of $N\neq 2m$, in which the normalization constant $\mathcal{N}_{m}=1$. The degree of the entanglement can be described by the von Neumann entropy defined by \begin{eqnarray} \label{e2} E(\rho_{1})=-Tr_{1}(\rho_{1}log\rho_{1}), \end{eqnarray} where $\rho_{1}$ is the reduced density operator of the first mode. For the $N$-photon entangled states defined by Eq. (\ref{e1}) we find the von Neumann entropy to be \begin{eqnarray} \label{e3} E_{1}&=&-\cos^{2}\gamma\log\cos^2\gamma - \sin^{2}\gamma\log\sin^2\gamma, \end{eqnarray} which indicates that quantum entanglement of the $N$-photon entangled states given by Eq. (\ref{e1}) is independent of the superposition phases $\theta_m$ and the total photon number of the two modes $N$, and changes periodically with respect to the entanglement angle $\gamma$. In particular, the amount of entanglement reaches the maximal value of $E_{1}=1$ when the entanglement angle takes values by $\gamma=k\pi + \pi/4$ with $k$ being an integer. In Fig. 1 we plot the change of the von Neumann entropy with respect to the entanglement angle. \begin{figure}[htp] \center \includegraphics[width=3.3in,height=2.1in]{fig1.eps} \caption{ The von Neumann entropy of the $N$-photon entangled state is plotted against the entanglement angle $\gamma$.} \end{figure} We then study quantum entanglement of superposition states based on the $N$-photon entangled states given by Eq. (\ref{e1}). Firstly, we consider a two-state superposition state defined by \begin{eqnarray} \label{e4} |\Psi_{2}\rangle=\frac{1}{\sqrt{2}}(|\psi_{30}\rangle + |\psi_{31}\rangle), \end{eqnarray} which leads to the von Neumann entropy \begin{eqnarray} \label{e5} E_{2}=1-\cos^{2}\gamma\log\cos^2\gamma - \sin^{2}\gamma\log\sin^2\gamma, \end{eqnarray} which implies that the amount of entanglement of the two-component superposition state only depends on the entanglement angle $\gamma$ of its basis. Especially, comparing Eq. (\ref{e5}) with (\ref{e5}) we find that the difference of the amount of entanglement between the two-component superposition state (\ref{e4}) and the basis state (\ref{e1}) is a positive constant, i.e., $E_2-E_1=1$. This indicates that the entanglement amount of the superposition state is always larger than that of the basis state for an arbitrary entanglement angle $\gamma$. In other words, starting with basis states defined by (\ref{e1}) one can construct quantum superposition states with larger amount of entanglement than that of the basis states. Hence, we may conclude that quantum superpositions for the two modes may increase the amount of entanglement. In order to further demonstrate the above idea of quantum superpositions increasing quantum entanglement, in what follows we take into account a multi-component quantum superposition state consisting of $N$ bases with a fixed photon number $N$, but with different distributions $m$, \begin{eqnarray} \label{e6} |\psi_{N}\rangle&=&A\sum_{m=0}^{N}\alpha_{m}|\psi_{Nm}\rangle, \end{eqnarray} where $A$ is a normalization factor. It is straightforward to express the $N$-component superposition state as the following number-sum Bell state, \begin{eqnarray} \label{e7} |\Psi_{N}\rangle&=&\sum_{m=0}^{N}d_{m}|N-m\rangle_{1}|m\rangle_{2}, \end{eqnarray} which is the eigenstate of the number-sum Bell operators $\hat{N}=\hat{N}_1+\hat{N}_2$ with $\hat{N}_i$ being the number operators of the respective modes. The coefficients in Eq. (\ref{e7}) are given by \begin{eqnarray} \label{e8} d_{m}=A(\alpha_{m}\mathcal{N}_{m}\cos\gamma+ \alpha_{N-m}\mathcal{N}_{N-m}e^{i\theta_{N-m}}\sin\gamma), \end{eqnarray} where the normalization of the superposition state (\ref{e8}) implies that $d_{m}$ satisfy the condition $\sum_{m=0}^{N}|d_{m}|^{2}=1$. In order to calculate the von Neumann entropy of the superposition state (\ref{e7}), we need the reduced density operator \begin{eqnarray} \label{e9} \rho_{1}&=&\sum_{m=0}^{N}|d_{m}|^{2}|N-m\rangle_{1}\langle N-m|, \end{eqnarray} which leads to the following von Neumann entropy \begin{eqnarray} \label{e10} E_{N}&=&-\sum_{m=0}^{N}|d_{m}|^{2}\log|d_{m}|^{2}. \end{eqnarray} \begin{figure}[htp] \center \includegraphics[width=3.5in,height=2.1in]{fig2.eps} \caption{ The von Neumann entropy of multi-component superposition states based on $|\psi_{Nm}\rangle$ is plotted against the entanglement angle $\gamma$ for $N=1, 2, 3$, and $4$, respectively.} \end{figure} For the sake of simplicity and without the loss of generality, we consider the situation of equal-weight superposition in which we choose the superposition coefficients $\alpha_{m}=1/\sqrt{N+1}$ and the relative phases $\theta_{m}=2\pi m/N$. In this case, we have \begin{eqnarray} \label{e11} |d_{m}|^{2}=A^2|\mathcal{N}_{m}|^{2}\{1+\cos[2\pi(N-m)/N]\sin2\gamma\}, \end{eqnarray} where the normalization constant is given by \begin{eqnarray} \label{e12} A^{-2}=\sum_{m=0}^{N}|\mathcal{N}_{m}|^{2}\{1+\cos[2\pi(N-m)/N]\sin2\gamma\}. \end{eqnarray} Then the von Neumann entropy of the superposition state can be directly obtained through substituting (\ref{e11}) and (\ref{e12}) into (\ref{e10}). From Eqs. (\ref{e10})-(\ref{e12}) we can find the maximal value of the von Neumann entropy of the superposition state to be $E_{N,max}=\log_2(N+1)$ when $\gamma=k\pi/2$ with $k$ being an arbitrary integer. This implies that the maximal amount of entanglement of the $N$-component superposition state only depends on the total photon number $N$, and increases with increasing the total photon number $N$. In order to clearly see the influence of the number of components $N$ and the entanglement angle $\gamma $, in Fig. 2 we plot the von Neumann entropy of the superposition states when $N=1, 2, 3$, and $4$, respectively. From Fig. 2 we can see that the entanglement amount increases with increasing the number of components, i.e., the photon number $N$, and changes periodically with respect to the entanglement angle $\gamma $. \section{Quantum nonlocality for $N$-photon entangled states} In this section, we study quantum nonlocality of $N$-photon entangled states and their superposition states through discussing the violation of the Bell's inequality in the pseudospin Bell-operator formalism developed by Chen and coworkers \cite{chen1}. Let us begin with a brief review of the pseudospin-operator formalism \cite{chen1}. For a single-mode boson field, the pseudospin operators can be defined in terms of project operators in a Fock space in the following form \begin{eqnarray} \label{e13} S_{z}&=&\sum_{n=0}^{\infty}[| 2n+1 \rangle\langle 2n+1 |-|2n\rangle\langle 2n |],\nonumber\\ S_{-}&=&\sum_{n=0}^{\infty}[| 2n \rangle\langle 2n+1 |]=(S_{j+})^{\dag}, \end{eqnarray} where $ |n\rangle$ are the usual Fock states of the boson mode. The operator $S_{z}=-(-1)^{\hat{N}}$ with $\hat{N}$ being the number operator and $(-1)^{N} $ being the parity operator, $S_{+}$ and $S_{-}$ being the ``parity-flip" operators. They satisfy the commutation relations of the $su(2)$ Lie algebra \begin{eqnarray} \label{e14} [S_{z},S_{ \pm}]=\pm 2S_{\pm},\hspace{0.3cm} [S_{+},S_{-}]&=S_{z}. \end{eqnarray} For an arbitrary vector living on the surface of a unit sphere $\vec{a}$ = ($\sin\theta_{a}\cos\varphi_{a}, \sin\theta_{a}\sin\varphi_{a}, \cos\theta_{a}$), we have the following dot product \begin{eqnarray} \label{e15} \vec{a}\cdot\vec{S}=S_{z}\cos\theta_{a}+\sin\theta_{a}(e^{i\varphi_{a}}S_{-}+e^{-i\varphi_{a}}S_{+}). \end{eqnarray} Then for a two-mode boson field, the Bell operator due to Clauser, Horne, Shimony, and Holt (CHSH) \cite{cla} can be defined by \begin{eqnarray} \label{e16} B&=& (\vec{a}\cdot\vec{S_{1}}) \otimes (\vec{b}\cdot\vec{S_{2}})+(\vec{a}\cdot\vec{S_{1}})\otimes (\vec{b'}\cdot\vec{S_{2}}) \nonumber \\ &&+(\vec{a'}\cdot\vec{S_{1}})\otimes (\vec{b}\cdot\vec{S_{2}})-(\vec{a'}\cdot\vec{S_{1}})\otimes (\vec{b'}\cdot\vec{S_{2}}), \end{eqnarray} where $ \vec{a'}, \vec{b}$, and $\vec{b'}$ are three unit vectors similarly defined as $\vec{a}$, $\vec{S_{1}}$ and $\vec{S_{2}}$ are defined as in Eq. (\ref{e13}). As well known, local hidden variable theories impose the Bell-CHSH inequality $|\langle B \rangle|\leq 2$ where $\langle B \rangle$ is the mean value of the Bell operator with respect to a given quantum state. However, in the quantum theory it is found that $|\langle B \rangle|\leq 2\sqrt{2}$, which implies that the Bell-CHSH inequality is violated. In particular, when $|\langle B \rangle|= 2\sqrt{2}$ for a given quantum state, we say that the Bell-CHSH inequality is maximally violated by the quantum state. Quantum nonlocality of a quantum state can be described by the violation of the Bell-CHSH inequality. The expectation value of the Bell operator with respect to a quantum state $|\psi \rangle$ can be expressed in terms of the correlation functions as \begin{eqnarray} \label{e17}\langle B \rangle&=&E(\theta_{a},\theta_{b})+E(\theta_{a},\theta_{b'})+E(\theta_{a'},\theta_{b})\nonumber \\&&-E(\theta_{a'},\theta_{b'}), \end{eqnarray} where the correlation functions are defined by \begin{equation} \label{e18} E(\theta_{a},\theta_{b})=\langle \psi|S_{\theta_{a}}^{(1)}\otimes S_{\theta_{b}}^{(2)}|\psi\rangle, \end{equation} with \begin{equation} \label{e19} S_{\theta_{a}}^{(j)}=S_{jz}\cos\theta_{a}+S_{jx}\sin\theta_{a}. \end{equation} We now investigate quantum nonlocality of the $N$-photon entangled state given by Eq. (\ref{e1}). For this multi-photon entangled state we find the correlation function to be \begin{eqnarray} \label{e20} E_{Nm}(\theta_{a},\theta_{b})&=&\mathcal{N}_{m}^{2}\{[(-1)^{N}+K(\theta_{m},\gamma)\delta_{N,2m}]\nonumber\\ &&\times\cos\theta_{a}\cos\theta_{b} +\delta_{N,2m\pm 1}\nonumber\\ &&\times K(\theta_{m},\gamma)\sin\theta_{a}\sin\theta_{b}\}, \end{eqnarray} where we have introduced the following effective state parameter \begin{eqnarray} \label{e21}K(\theta_{m},\gamma)=\cos\theta_{m}\sin 2\gamma, \end{eqnarray} which describes the effect of the basis state defined by (\ref{e1}) on the correlation functions. Making use of the correlation function (\ref{e20}), from Eq. (\ref{e17}) we can get the expectation value of the Bell operator for arbitrary values of all azimuthal angles $\theta_{a}$, $\theta_{b}$, $\theta_{a'}$ and $\theta_{b'}$. As a concrete example, we consider the situation of $\theta_{a}=0$, $\theta_{a'}=\pi/2$ and $\theta_{b}=-\theta_{b'}$. In this case, from Eqs. (\ref{e17}), (\ref{e20}) and (\ref{e21}) we can obtain the expectation value of the Bell operator given by \begin{eqnarray} \label{e22}\langle B \rangle&=&2\mathcal{N}_{m}^{2}\{[(-1)^{N}+K(\theta_{m},\gamma)\delta_{N,2m}]\cos\theta_{b}\nonumber\\ &&+\delta_{N,2m\pm 1}K(\theta_{m},\gamma)\sin\theta_{b}\}. \end{eqnarray} In order to observe quantum nonlocality of the $N$-photon entangled state (\ref{e1}), we consider two different cases of $N=2m$ and $N=2m \pm 1$, respectively. When $N=2m$, the quantum state given by Eq. (\ref{e1}) is disentangled, and reduces to \begin{eqnarray} \label{e23}|\psi_{N m}\rangle=\mathcal{N}_{m}(\cos\gamma+e^{i\theta_{m}}\sin\gamma)|m\rangle_{1}|m\rangle_{2}. \end{eqnarray} Making use of Eqs. (\ref{e17}), (\ref{e18}) and (\ref{e20}), we obtain the expectation value of the Bell operator with respect to the disentangled state (\ref{e23}) $\langle B \rangle=2\cos\theta_{b}$, which means that $|\langle B \rangle|\leq 2$. Hence, the unentangled state (\ref{e23}) cannot produce a violation of Bell's inequality. On the other hand, when $N=2m\pm 1$, the expectation value of the Bell operator with respect to the state (\ref{e1}) is given by \begin{eqnarray} \label{e24}\langle B \rangle=2[K(\theta_{m},\gamma)\sin\theta_{b}-\cos\theta_{b}], \end{eqnarray} which indicates that for a fixed entanglement angle $\gamma$ and a fixed relative phase $\theta_m$, when $\theta_{b}=-\tan^{-1}K$, the expectation value of the Bell operator $\langle B \rangle$ reaches its maximum given by \begin{eqnarray} \label{e25} \langle B \rangle_{max}=2\sqrt{1+K^{2}(\theta_{m},\gamma)}, \end{eqnarray} which implies that $\langle B \rangle_{max}\geq 2$ due to $|K(\theta_{m},\gamma)|\leq1.$ Thus, the $N$-photon entangled state always violates the Bell's inequality if $K(\theta_{m},\gamma)\neq 0$. From Eqs. (\ref{e21}) and(\ref{e24}) we can see that the degree of violation of the Bell's inequalities depends upon both the entangling angle $\gamma$ and the relative phase $\theta_{m}$, and it changes periodically with respect to both $\gamma$ and $\theta_{m}$. In particular, we note that the relative phase $\theta_{m}$ seriously affects the mean value of the Bell operator although it dos not affect quantum entanglement of the quantum state given by (\ref{e1}). It is easy to find that when $|\psi_{N m}\rangle$ reaches maximal entanglement i.e., $\gamma=\pi/4$, and $\theta_{m}=0$, we have $K(\theta_{m},\gamma)=1$. Under these conditions, we can reach the maximal violation of Bell's inequality with $ \langle B \rangle_{max}=2\sqrt{2}$. In Fig. 3 we plot the degree of the violation of the Bell's inequality for $N$-photon entangled state against $\gamma$ and $\theta_{m}$. From Fig. 3 we can see that the mean value of the Bell operator with respect to the multi-photon entangled state (1) changes periodically with both of the entanglement angle and the relative phase, and the Bell's inequality is always violated for arbitrary values of the entanglement angle and the relative phase except $\gamma=k\pi/2$ and $\theta=(2k+1)\pi/2$ with $k$ is an integer. \begin{figure}[htp] \center \includegraphics[width=3.3in,height=3.1in]{fig3.eps} \caption{The degree of the maximal violation of the Bell's inequality for $N$-photon entangled state $\langle B \rangle_{max}$ is plotted against the entanglement angle $\gamma$ and the relative phase $\theta_{m}$. } \end{figure} Finally, we consider quantum nonlocality of quantum superposition states which is formed using the multi-photon entangled state (\ref{e1}). In order to calculate the expectation value of the Bell operator with respect to the superposition states based on entangled states given by (\ref{e1}), we need the following correlation function \begin{eqnarray} \label{e26} E_{m m'}=\langle\psi_{N m}|S_{\theta_{a}}^{(1)}\otimes S_{\theta_{b}}^{(2)}|\psi_{N m'}\rangle, \end{eqnarray} which is given by the following expression \begin{eqnarray} \label{e27}E_{m m'}&=&\mathcal{N}_{m}\mathcal{N}_{m'}^{*}\cos\theta_{a}\cos\theta_{b}\nonumber \\ &&\times \left \{(-1)^{N}\delta_{m,m'}\delta_{N,m+m'}\cos\gamma \sin\gamma \right.\nonumber\\ & &\times [\cos^{2}\gamma+e^{i (\theta_{m'}-\theta_{m})} \sin^{2}\gamma] \nonumber\\ && + (-1)^{m+m'}(e^{i \theta_{m'}} +e^{-i\theta_{m}})\}\nonumber\\ && + \mathcal{N}_{m}\mathcal{N}^*_{m'}\sin\theta_{a}\sin\theta_{b} \nonumber \\ && \times \left\{[\cos^{2}\gamma + e^{i(\theta_{m'}-\theta_{m})}\sin^{2}\gamma]\delta_{m,m'\pm 1} \right. \nonumber\\ && \left. + \cos \gamma \sin \gamma(e^{i\theta_{m'}}+e^{-i\theta_{m}})\delta_{N,m+m'\pm 1}\right \}. \end{eqnarray} As a simple example of analyzing quantum nonlocality of superposition states, we consider the following two-component superposition state, \begin{eqnarray} \label{e28}|\psi_{3}\rangle=C(\alpha_{0}|\psi_{30}\rangle+\alpha_{1}|\psi_{31}\rangle), \end{eqnarray} where $C$ is the normalization constant. Making use of Eq. (\ref{e27}), we can obtain the expectation value of the Bell operator for this superposition state \begin{eqnarray} \label{e29}\langle B \rangle&=&-2\cos\theta_{b} + 4\sin\theta_{b}(|\alpha_{0}|^{2}+|\alpha_{1}|^{2})^{-1}\nonumber \\&&\times Re[\alpha_{0}^{*}\alpha_{1}(\cos^{2}\gamma +e^{i (\theta_{1}-\theta_{0})} \sin^{2}\gamma)], \end{eqnarray} which indicates that $\langle B \rangle$ depends upon not only the superposition coefficients $\alpha_{0},\alpha_{1}$, the azimuthal angles $\theta_{a}, \theta_{b}$, but also upon the state parameters of the basis states given by Eq. (\ref{e1}), $\theta_{0}, \theta_{1}$ and $\gamma$. In order to observe the maximal violation of the Bell inequality, we can choose $\alpha_{0}=\alpha_{1}=1$, the Bell function $\langle B \rangle$ given by Eq. (\ref{e29}) becomes \begin{eqnarray} \label{e30}\langle B \rangle&=&2\sin\theta_{b}[\cos^{2}\gamma +\cos(\theta_{1}-\theta_{0})\sin^{2}\gamma]\nonumber\\ &&-2\cos\theta_{b}. \end{eqnarray} Making use of Eq. (\ref{e30}), we can show that the maximal violation of the Bell inequality can be reached by controlling the azimuthal angle $\theta_{b}$, the state parameters of the basis states given by Eq. (1), $\theta_{0}, \theta_{1}$ and $\gamma$. In fact, when $\theta_{1}-\theta_{0}=2k\pi$ with $k$ being an arbitrary integer, we can arrive at the following Bell function \begin{eqnarray} \label{e31}\langle B \rangle&=2(\sin\theta_{b}-\cos\theta_{b}) \end{eqnarray} which means that the values of the Bell function of $\langle B \rangle$ are independent of the entangling parameter of the basis state defined in (1), and we can reach the maximal violation of the Bell inequality $\langle B \rangle_{max}=2\sqrt{2}$ when $\theta_{b}=-\pi/4$. On the other hand, when the phase difference takes $\theta_{1}-\theta_{0}=(2k+1)\pi$ with $k$ being an arbitrary integer, the Bell function given by Eq. (30) becomes \begin{eqnarray} \label{e32}\langle B \rangle&=2[\sin\theta_{b}\cos(2\gamma)-\cos\theta_{b}], \end{eqnarray} from which we can see that when the entanglement angle satisfies $\gamma=k\pi$ with $k$ being an arbitrary integer, Eq. (\ref{e32}) becomes Eq. (\ref{e31}), hence the maximal violation of the Bell inequality is reached with $\theta_{b}=-\pi/4$. However, when the entanglement angle satisfies $\gamma=(2k+1)\pi/2$ with $k$ being an arbitrary integer, the Bell function (\ref{e32}) reduces to \begin{eqnarray} \label{e33}\langle B \rangle&=-2(\sin\theta_{b}+\cos\theta_{b}), \end{eqnarray} which implies that the violation of the Bell inequality reaches its maximal value $\langle B \rangle_{max}=2\sqrt{2}$ when $\theta_{b}=\pi/4$. Therefore, we can conclude that superposition states based on entangled states (\ref{e1}) can provide us with more ways of reaching the maximal violation of the Bell's inequality. \section{Concluding remarks} In summary, we have studied quantum entanglement and quantum nonlocality of $N$-photon entangled states for the two modes defined in Eq. (1) and their superpositions through investigating the von Neumann entropy and the violation of the Bell's inequality in the pseudospin Bell-operator formalism. For the multi-photon entangled states defined in Eq. (1) we have indicated that the von Neumann entropy is independent of the relative phase $\theta_{m}$. Hence, quantum entanglement of these states only depends on the entanglement angle. However, quantum nonlocality of these quantum states exhibits different dependence upon the state parameters. We have shown that both of the entanglement angle and the relative phase seriously affect the violation of the Bell's inequality. And we have indicated that under certain conditions the maximal violation of the Bell's inequality can be reached. It is worthwhile to mention that multi-component superposition states made from $N$-photon entangled states for the two modes defined in Eq. (1) exhibit some interesting characteristics on their quantum entanglement and quantum nonlocality. Firstly, we have found that these multi-component superposition states have larger amount of entanglement than that of the basis states defined by (1). Hence, quantum superpositions for the two modes can increase the amount of entanglement. This indicates the possibility of obtaining entangled states with a larger amount of entanglement starting from entangled states with a smaller amount of entanglement. Secondly, we have found that for these quantum superposition states there are more ways to make corresponding Bell's inequality reach the maximal violation, and revealed that quantum nonlocality can be controlled and manipulated by adjusting the state parameters of $|\psi_{N m}\rangle$, superposition coefficients, and the azimuthal angles of the Bell operator. We hope that these results obtained in present paper would find their applications in quantum information processing \cite{nie,che,lu,zhou2} and the test of quantum nonlocality \cite{jeo,pit}. \acknowledgments This work was supported in part the National Fundamental Research Program (2001CB309310), the National Natural Science Foundation of China under Grant Nos. 90203018, 10325523,10347128 and 10075018, the foundation of the Education Ministry of China, and the Educational Committee of Human Province under Grant Nos. 200248 and 02A026.
{ "timestamp": "2005-03-10T02:32:56", "yymm": "0503", "arxiv_id": "quant-ph/0503099", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503099" }
\section{Introduction} Consider the problem of finding a zero of a function $\varphi : \mathbb{R} \to \mathbb{R}$. If there are several zeros, it is required to find at least one of them. It is supposed that the function can be measured at any point, with some random error. The standard algorithm of stochastic approximation consists in calculating successive approximations of the required value, $x_0$,\, $x_1$,\, $x_2, \ldots$, according to the rule \begin{equation}\label{eqal1} x_{t} = x_{t-1} - \gamma_{t-1} y_t, \quad t=1,\ 2,\ldots, \end{equation} where \begin{equation}\label{eqal2} y_t = \varphi(x_{t-1}) + \xi_t \hspace{26mm} \end{equation} is the value of $\varphi$ measured at $x_{t-1}$,\, $\xi_t$ is the measurement error; \,$\gamma_0$,\, $\gamma_1$,\, $\gamma_2, \ldots$ is the sequence of step sizes of the algorithm. Usually it is assumed that the step sizes are positive real numbers satisfying the relations $\sum \gamma_t = \infty$,\, $\sum \gamma_t^2 < \infty$. Then, under some additional assumptions on $\varphi$ and $\xi_t$, the algorithm a.s. converges to a zero point of $\varphi$ (see, e.g., \cite{b007,b001}). In practice, however, the convergence rate of this algorithm may prove to be unsatisfactory, therefore, when solving practical tasks, various modifications of the algorithm are used. There are widely utilized heuristical algorithms using random, rather than deterministic, step size, which is corrected in the course of the algorithm, according to the current data \cite{a005,a011,a100,a006}. In particular, there is used the idea that prescribes to decrease the step size if the sequence of increments $x_{t} - x_{t-1}$ changes the sign often enough, indicating that the current value $x_t$ is close to the set of zeros of $\varphi$, and hence, the measurement error $\xi_t$ of the function is big enough with respect to the function itself $\varphi(x_{t-1})$. Alternatively, one should increase the step size, or leave it unchanged. So, Kesten in the theoretical work \cite{a008} considered an algorithm using (\ref{eqal1}), (\ref{eqal2}), and the rule of modification of $\gamma_t$: \begin{equation}\label{eqK} \gamma_t = \gamma(s_t), \ \ \ \ \ \ s_t = \left\{ \begin{array}{lll} s_{t-1} & \textrm{ if } & y_{t-1} y_t > 0\\ s_{t-1} + 1 & \textrm{ if } & y_{t-1} y_t \le 0, \end{array} \right.\quad t=2,3,\ldots. \end{equation} where $s_0 = 0$,\, $s_1 = 1$;\, $\gamma(0)$,\, $\gamma(1)$,\, $\gamma(2), \ldots$ is a sequence of positive numbers satisfying the relations $\sum \gamma(m) = \infty$,\, $\sum \gamma^2(m) < \infty$. Thus, the step size cannot increase in the course of algorithm; it can only decrease or remain unchanged. It is supposed that there is a unique zero of $\varphi$. Kesten proved that $x_t$ a.s. converges to this zero point. A multidimensional version of this algorithm is considered in \cite{a003}. There are also heuristical procedures (in particular, in artificial neural networks), where at each moment $t$ the step size is multiplied by a positive constant less than 1, if the measurement data indicate that $x_t$ is close enough to the zero set of $\varphi$, and by a constant more than 1, elsewhere \cite{a067,a001,a100,a101}. This kind of rules ensure sufficiently high convergence rate, however the step size converges like a geometric progression, therefore $\sum \gamma_t < \infty$, which means that the limit of $\{ x_t \}$ need not be a zero point of $\varphi$, but instead, the sequence may "get stuck" on its way to the set of zeros of $\varphi$. Nevertheless, such a procedure may be justified if it gives a value close enough to one of the zeros of $\varphi$. In the present paper, a stochastic approximation algorithm utilizing this rule of step size modification is considered. Namely, the rule (\ref{eqal1}), (\ref{eqal2}), jointly with the following rule \begin{eqnarray} \gamma_t = \left\{ \begin{array}{lll} \min\{u\, \gamma_{t-1},\, \bar{\mathrm{g}}\} & \textrm{ if } & y_{t-1} y_t > 0,\\ d\, \gamma_{t-1} & \textrm{ if } & y_{t-1} y_t \le 0, \end{array} \right.\quad t=2,3,\ldots. \label{eqal3} \end{eqnarray} is used. Here $0 < d < 1 < u$,\, $0 < \gamma_0$,\, $\gamma_1 \le \bar{\mathrm{g}}$,\, $\bar{\mathrm{g}}$ is a positive constant. Let us point out the main differences between (\ref{eqal3}) and Kesten's rule (\ref{eqK}). First, according to (\ref{eqal3}), $\gamma_t$ can both decrease and increase. Second, in Kesten's algorithm one always has $\sum \gamma_t = \infty$. On the other hand, it looks likely that in the case of convergence of the algorithm (\ref{eqal1}), (\ref{eqal2}), (\ref{eqal3}), $\gamma_t$ converges like a geometric progression (this conjecture will be justified in the section 3), therefore the limit of algorithm may not be a zero point of $\varphi$. Suppose that $\{ \xi_t \}$ is a sequence of i.i.d.r.v. with zero mean, besides $\mathrm{P} (\xi_t > 0) = \mathrm{P} (\xi_t < 0)$. Under some additional assumptions on $\varphi$, $\xi_t$, and $\bar{\mathrm{g}}$, stated below, the process defined by (\ref{eqal1}), (\ref{eqal2}), (\ref{eqal3}) a.s. diverges if $ud > 1$, and converges if $ud < 1$, moreover the limit of $\{ x_t \}$ belongs to $\mathcal{U} (\frac{\ln u}{-\ln d})$. Here $\mathcal{U}(\lambda)$,\, $0 < \lambda < 1$, is a monotone decreasing family of sets of real numbers, besides every set $\mathcal{U}(\lambda)$ contains the set $\mathrm{Z}$ of zeros of $\varphi$, and $\partial (\mathcal{U}(\lambda), \mathrm{Z}) \to 0$ as $\lambda \to 1^-$. (Here by definition $\partial (A,B) = \sup_{x\in A} \inf_{y\in B} |x-y|$ for any two sets of real numbers $A$ and $B$.) This statement is a consequence of the main theorem, which will be stated in section 2 and proved in section 3. Thus, by adjusting the parameters $u$ and $d$ (for example, fixing $u$ and letting $d \to 1/u - 0$), one can reach necessary precision of the algorithm; higher precision is obtained at the expense of lower convergence rate. \section{Definition of the algorithm and statement of the main result} Consider the algorithm given by (\ref{eqal1}), (\ref{eqal2}), (\ref{eqal3}). The rule (\ref{eqal3}) means that at each instant $t$, step size is multiplied by $u$ or by $d$, if the result of multiplication is less than $\bar{\mathrm{g}}$; otherwise, step size is set to be $\bar{\mathrm{g}}$. Thus, the maximal possible value of step size equals $\bar{\mathrm{g}}$. The rule (\ref{eqal3}) can be written in the form \begin{equation}\label{eqal4n} \begin{array}{l@{ = }l} \ln \tilde \gamma_t & \ln \gamma_{t-1} + \ln u \cdot \;\mathbb{I}(y_{t-1} y_t>0) + \ln d \cdot \;\mathbb{I}(y_{t-1} y_t \le 0), \\ \ln \gamma_t & \min\{ \ln \tilde \gamma_t, \ln \bar{\mathrm{g}}\}. \end{array} \end{equation} Let us take the following assumptions: \begin{description} \item [A1] Denote ${\cal F}_t$, $t = 0,1,2,\ldots$ the $\sigma$-algebra generated by $x_i$, $\gamma_i$, and $\xi_i$, $0\le i \le t$; then $\xi_{t+1}$ does not depend on ${\cal F}_t$. \item [A2] The values $\xi_t$ are identically distributed, with zero mean and finite variance: $\mathrm{E} \xi_t = 0$,\, $\mathrm{Var} \xi_t =: S < +\infty$. \item [A3] (a) There exists $L >0$ such that for any interval $I \subset [-L,\, L]$, $\mathrm{P}(\xi_1 \in I)>0$;\\ \hspace*{-2mm}(b) $\mathrm{P}(\xi_1 = 0) = 0$. \item [A4] $\varphi \in \mathbb{C}^1(\mathbb{R})$ and $\sup_x |\varphi'(x)| =: M < \infty$. \item [A5] $\bar{\mathrm{g}} < 2/M$. \item [A6] There exists $R>0$ such that \begin{itemize} \item[(a)] $x \varphi(x) > 0$ as $|x| \ge R$, and \item[(b)] $\displaystyle \inf_{|x|\ge R} \varphi^2(x) > \frac{\bar{\mathrm{g}} M S}{2-\bar{\mathrm{g}} M}$. \end{itemize} \end{description} \begin{remark From \A{A4} and \A{A6}\,(a) it follows that the set $\mathrm{Z}$ is non-empty and is contained in $(-R,\, R)$. \end{remark} \begin{remark Note that assumptions \A{A4}--\A{A6} guarantee convergence of the deterministic counterpart of algorithm (\ref{eqal1}), (\ref{eqal2}), (\ref{eqal3}) (that is, of the algorithm with $\xi_t \equiv 0$). Moreover, under these conditions, any deterministic algorithm $x_t = x_{t-1} - \gamma_{t-1} \varphi(x_{t-1})$ converges, whatever the sequence $\{\gamma_t\}$ satisfying $\gamma_t \le \bar{\mathrm{g}}$. \end{remark} Introduce the functions: \begin{equation} \label{ast1} k_+(z) := \lim_{\epsilon\to0^+} \sup\{ \mathrm{P}( (\varphi_1+\xi_1)(\varphi_2+\xi_2) > 0),\ |\varphi_1 - z| < \epsilon,\ |\varphi_2 - z| < \epsilon\}, \end{equation} \begin{equation} \label{ast2} k_-(z) := \lim_{\epsilon\to0^+} \inf\{ \mathrm{P}( (\varphi_1+\xi_1)(\varphi_2+\xi_2) > 0),\ |\varphi_1 - z| < \epsilon,\ |\varphi_2 - z| < \epsilon\}; \end{equation} one has $k_+(z) \ge 1/2$,\, $0 \le k_{\pm}(z) \le 1$,\, $\lim_{z\to\infty} k_{\pm}(z)=1$. Further, define the sets of real numbers \begin{equation}\label{ast3} V_{\pm}^{(a)} := \{ x : k_\pm(\varphi(x)) < a \}, \quad V_{\pm}^{[a]} := \{ x : k_\pm(\varphi(x)) \le a \}; \end{equation} obviously, $V_+^{(a)} \subset V_-^{(a)}$,\, $V_\pm^{(a)} \subset V_\pm^{[a]}$ for any $a$. Note that $V_+^{(a)}$ is open. Indeed, let $x \in V_+^{(a)}$, then there exists $\epsilon > 0$ such that $$ \sup\{ \mathrm{P}( (\varphi_1+\xi_1)(\varphi_2+\xi_2) > 0),\ |\varphi_1 - \varphi(x)| < \epsilon,\ |\varphi_2 - \varphi(x)| < \epsilon\} =: c < a. $$ Then for $x'$ close enough to $x$ one has $|\varphi(x')- \varphi(x)| < \varepsilon/2$, hence $$ \sup\{ \mathrm{P}( (\varphi_1+\xi_1)(\varphi_2+\xi_2) > 0),\ |\varphi_1 - \varphi(x')| < \epsilon/2,\ |\varphi_2 - \varphi(x')| < \epsilon/2 \} \le c < a. $$ This implies that $k_+(\varphi(x')) < a$, hence $x' \in V_+^{(a)}$. Denote also \begin{equation}\label{eq5} \mathrm{k} := \frac{\ln(1/d)}{\ln(u/d)}. \end{equation} Denote by $\mathrm{Z}$ the set of zeros of $\varphi$, i.e., $\mathrm{Z}:=\{ x : \varphi(x)=0 \}$. Suppose that $x\in V_+^{(\mathrm{k})}$,\, $x_{t-2} \in (x-\epsilon,\, x+\epsilon) \subset V_+^{(\mathrm{k})}$, and $\gamma_{t-2} < \epsilon$, where $\epsilon$ is a small positive number. Then, with a probability close to 1,\, $x_{t-1}$ also belongs to a small (possibly larger) neighborhood of $x$ contained in $V_+^{(\mathrm{k})}$, and taking into account (\ref{ast1}) and (\ref{ast3}), one gets \[ \begin{array}{l} \mathrm{P}( y_{t-1} y_t > 0 \,{\Big |}\, |x_{t-2} - x| < \epsilon, \gamma_{t-2} < \epsilon ) =\\ =\mathrm{P}( (\varphi(x_{t-2}) + \xi_{t-1}) (\varphi(x_{t-1}) + \xi_{t}) > 0 \,{\Big |}\, |x_{t-2}-x|<\epsilon, \gamma_{t-2} < \epsilon) < \mathrm{k}. \end{array} \] Then, using (\ref{eqal4n}) and (\ref{eq5}), one obtains \[ \begin{array}{l} \mathrm{E}[ \ln \gamma_t-\ln \gamma_{t-1} \,{\Big |}\, |x_{t-2}-x| < \epsilon, \gamma_{t-2} < \epsilon ] \le \\ \ln u \cdot \mathrm{P}(y_{t-1} y_t > 0 \,{\Big |}\, |x_{t-2}-x| < \epsilon, \gamma_{t-2} < \epsilon ) + \ln d \cdot \mathrm{P}(y_{t-1} y_t \le 0 \,{\Big |}\, |x_{t-2} - x| < \epsilon, \gamma_{t-2} < \epsilon ) \\ < \ln u \cdot \mathrm{k} + \ln d \cdot (1-\mathrm{k}) = 0. \end{array} \] Thus, in a sense, the set $V_+^{(\mathrm{k})}$ can be regarded to be a \textit{domain of decrease of step size}: if several consecutive values of $x_t$ belong to $V_+^{(\mathrm{k})}$ and are close enough to each other, and if the first term of the sequence of corresponding step sizes $\gamma_t$ is small enough, then the sequence of their mean values $E \gamma_t$ decreases. Now, suppose that $x \in \mathbb{R} \setminus V_-^{[\mathrm{k}]}$,\, $x_{t-2} \in (x-\epsilon,\, x+\epsilon) \subset \mathbb{R} \setminus V_-^{[\mathrm{k}]}$, and that $\gamma_{t-2} < \epsilon$. Analogously, for $\epsilon$ small enough, one has \[ \mathrm{P}( y_{t-1} y_t > 0 \,{\Big |}\, |x_{t-2} - x| < \epsilon, \gamma_{t-2} < \epsilon ) > \mathrm{k}, \] and then, using again (\ref{eqal4n}) and (\ref{eq5}) and taking into account that for $\epsilon < \bar{\mathrm{g}}/u^2$,\, $\tilde\gamma_t = \gamma_t$, one obtains \[ \begin{array}{l} \mathrm{E}[ \ln \gamma_t - \ln \gamma_{t-1} \,{\Big |}\, |x_{t-2} - x| < \epsilon, \gamma_{t-2} < \epsilon] = \\ \ln u \cdot \mathrm{P}( y_{t-1} y_t > 0 \,{\Big |}\, |x_{t-2} - x | < \epsilon, \gamma_{t-2} < \epsilon]) + \ln d \cdot \mathrm{P}( y_{t-1} y_t \le 0 \,{\Big |}\, |x_{t-2} - x| < \epsilon, \gamma_{t-2} < \epsilon ]) \\ > \ln u \cdot \mathrm{k} + \ln d \cdot (1-\mathrm{k}) = 0. \end{array} \] Thus, the set $\mathbb{R} \setminus V_-^{[\mathrm{k}]}$ can be regarded as a \textit{domain of increase of step size}: if several consecutive values of $x_t$ belong to $\mathbb{R}\setminus V_-^{[\mathrm{k}]}$ and are close enough to each other, and if the first of the corresponding values of $\gamma_t$ is small enough, then the sequence of their mean values $E \gamma_t$ increases. Note that if $\mathrm{k} > k_+(0)$ then, by virtue of (\ref{ast3}), $\mathrm{Z} \subset V_+^{(\mathrm{k})}$, that is, all the zeros of $\varphi$ belong to the region of decrease of step size. On the other hand, if $\mathrm{k} < \inf_z k_-(z)$ then $V_-^{[\mathrm{k}]} = \emptyset$, which means that the region of increase of step size coincides with $\mathbb{R}$. It seems likely that in the first case the algorithm can converge, and in the second one, cannot. This conjecture is confirmed by the following theorem, which is the main result of the paper. \vspace{2mm} \textbf{Theorem} \textit{ Let the assumptions \A{A1}--\A{A6} be satisfied; consider the process $\{ x_t,\ \gamma_t \}$ defined by (\ref{eqal1}), (\ref{eqal2}), (\ref{eqal3}). Recall that $\mathrm{k} = \frac{\ln(1/d)}{\ln(u/d)}$. Then\\ (a) If $\mathrm{k} > k_+(0)$ then $\{x_t\}$ a.s. converges to a point from $V_-^{[\mathrm{k}]}$.\\ (b) If $\mathrm{k} < \inf_z k_-(z)$ then $\{ x_t \}$ a.s. diverges. } \vspace{2mm} Suppose that $\mathrm{P} (\xi_1 = x) = 0$ for any real $x$ and that $\mathrm{P} (\xi_1 > 0) = \mathrm{P} (\xi_1 < 0)$. Then the function $k(\cdot) := k_+(\cdot)$ coincides with $k_-(\cdot)$, is continuous, and is given by $$ k(z) = \mathrm{P} ((z + \xi_1)(z + \xi_2) > 0); $$ $z = 0$ is the unique minimum of $k(\cdot)$, and $k(0) = \inf_z k(z) = 1/2$. After a simple algebra, one can rewrite the hypotheses of theorem in the form (a) $ud < 1$, (b) $ud > 1$. Denote $\mathcal{U}(\lambda) := V^{[\frac{1}{1+\lambda}]} = \{ x :\, k(\varphi(x)) \le \frac{1}{1 + \lambda} \}$; \, $\mathcal{U}(\lambda)$,\, $1 < \lambda < 1$ is a monotone decreasing family of sets containing $\mathrm{Z}$ and tending to $\mathrm{Z}$ as $\lambda \to 1^-$. Thus, one comes to \vspace{2mm} \textbf{Corollary} \textit{ Let, in addition to assumptions \A{A1}--\A{A6}, $\mathrm{P} (\xi_1 = x) = 0$ for any $x \in \mathbb{R}$, and $\mathrm{P} (\xi_1 > 0) = \mathrm{P} (\xi_1 < 0) = 1/2$. Consider the process defined by (\ref{eqal1}), (\ref{eqal2}), (\ref{eqal3}). Then there exists a monotone decreasing family of sets $\mathcal{U}(\lambda)$,\, $0 < \lambda < 1$ such that $\mathcal{U}(\lambda) \supset \mathrm{Z}$,\, $\partial (\mathcal{U}(\lambda), \mathrm{Z}) \to 0$ as $\lambda \to 1^-$, and\\ (a) if $ud < 1$ then $\{x_t\}$ a.s. converges to a point from $\mathcal{U} (\frac{\ln u}{-\ln d})$;\\ (b) if $ud > 1$ then $\{x_t\}$ a.s. diverges. } \begin{remark Theorem does not give any information about behavior of the algorithm for the values $u$,\, $d$ such that $$ \inf\nolimits_z k_-(z) \le \frac{\ln(1/d)}{\ln(u/d)} \le k_+(0). $$ In particular, under the hypotheses of corollary, the case $ud = 1$ remains unexplored. These issues will be addressed elsewhere. \end{remark} \section{Proof of theorem} First we prove 10 auxiliary lemmas, and then, basing on them, we prove theorem. Here all statements about random variables are supposed to be true almost surely. In the sequel, we shall mainly designate random values by Greek letters, and real numbers and functions from $\mathbb{R}$ to $\mathbb{R}$, by Latin ones; the letters $t$,\, $i$,\, $j$,\, $s$ will denote integer non-negative numbers. The function $\varphi$ and the random values $x_t$,\, $y_t$ are exceptions; also, traditional notation $\epsilon$,\, $\delta$ for small positive numbers will be used. \begin{lemma} If $\sum_t \gamma_t < \infty$ then the sequence $\{ x_t \}$ converges. \end{lemma} \textit{Proof.} Note that without loss of generality one can assume that $x_0$ is bounded. Indeed, replacing $x_0$ by $\tilde x_0 = x_0 \cdot \;\mathbb{I}(|x_0|<X)$ changes the process only with probability $\mathrm{P}(|x_0|>X)$. By taking $X$ large enough, one can make this probability arbitrarily small. Let $C>0$; define the stopping time $\tau_C = \inf \{ t : \sum_{i=0}^t \gamma_i > C\}$ and introduce the new process $x_t^C$, $\gamma_t^C$ by \[ \begin{array}{l} x_t^C = x_t, \quad \gamma_t^C=\gamma_t \textrm{ as } t < \tau_c, \textrm{ and } \\ x_t^C = x_{\tau_C}, \quad \gamma_t^C=0 \textrm{ as } t \ge \tau_c. \end{array} \] First, let us prove that the sequence $\{x_t^C\}$ is bounded. Designate $M_R := \sup_{|x|\ge R} \frac{\varphi(x)}{x}$; from \A{A4} it follows that $M_R<\infty$. One has \begin{equation}\label{eq9} |x_t^C| \le |x_{t-1}^C - \gamma_{t-1}^C \varphi(x^C_{t-1})| + \gamma_{t-1}^C |\xi_t|. \end{equation} Using that $\gamma_{t-1}^C \le C$ and $|\varphi(x_{t-1})^C| \le |\varphi(0)| + M |x_{t-1}^C|$, one obtains \begin{equation}\label{eq10} |x_t^C| \le |x_{t-1}^C|(1 + CM) + \gamma_{t-1}^C(|\varphi(0)|+|\xi_t|). \end{equation} If $\gamma_{t-1}^C \le 2/M_R$, an even more precise estimate for $x_t^C$ can be obtained. We shall distinguish between two cases: (i) $|x_{t-1}|\le R$ and (ii) $|x_{t-1}^C| > R$. In case (i), designating $\bar{b} := \sup_{|x|\le R} |\varphi(x)|$, one has \begin{equation}\label{eq11} |x_{t-1}^C - \gamma_{t-1}^C \varphi(x_{t-1}^C)| \le |x_{t-1}^C| + \gamma_{t-1}^C \bar{b}. \end{equation} In the case (ii) one has \[ 0 \le \gamma_{t-1}^C \frac{\varphi(x_{t-1}^C)}{x_{t-1}^C} \le \frac{2}{M_R} M_R = 2, \] hence \begin{equation}\label{eq12} |x_{t-1}^C - \gamma_{t-1}^C\varphi(x_{t-1}^C)| \le |x_{t-1}^C|. \end{equation} Thus, in both cases (i) and (ii), from (\ref{eq9}), (\ref{eq11}), and (\ref{eq12}) one gets \begin{equation}\label{eq13} |x_t^C| \le |x_{t-1}^C| + \gamma_{t-1}^C ( \bar{b} + |\xi_t| ). \end{equation} The overall number of values of $t$ such that $\gamma_{t-1}^C \le 2/M_R$ is less than $CM_R/2$; therefore, using (\ref{eq10}) and (\ref{eq13}), one concludes that \begin{equation}\label{eq14} |x_t^C| \le \left( |x_0| + \sum_{i=1}^t \gamma_{i-1}^C(\bar{b}+|\varphi(0)|+|\xi_i|) \right) \cdot (1+CM)^{CM_R/2}. \end{equation} Denote $c_0 := \bar{b} + |\varphi(0)|+\mathrm{E}|\xi_1|$ and $\zeta_t := |\xi_t| - \mathrm{E}|\xi_t|$; using that $\sum_1^\infty \gamma_{i-1}^C \le C$ one gets \begin{equation}\label{eq15} |x_t^C| \le \left( |x_0| + C\, c_0 + \sum_{i=1}^t \gamma_{i-1}^C \zeta_i \right)\cdot(1+CM)^{CM_R/2}. \end{equation} Using that $\sum_1^\infty \mathrm{E}(\gamma_{t-1}^C \zeta_t)^2 = \mathrm{E} \zeta_1^2 \cdot \sum_1^\infty \mathrm{E}(\gamma_{t-1}^C)^2 < \infty$, one obtains that the martingale $\sum_1^t \gamma_{i-1}^C \zeta_i$ is bounded; the value $x_0$ is also bounded, so, by (\ref{eq15}), one concludes that the sequence $\{x_t^C\}$ is bounded. Now, let us show that $\{x_t^C\}$ converges. From the definition of $x_t^C$ and $\gamma_t^C$ it follows that \[ x_t^C = x_0 - \sum_1^t \gamma_{i-1}^C \varphi(x_{i-1}^C) - \sum_1^t \gamma_{i-1}^C \xi_i. \] Using that the sequence $\{ \varphi(x_{i-1}^C)\}$ is bounded and that $\sum_1^\infty \gamma_{i-1}^C \le C$, one gets that the series $\sum_1^\infty \gamma_{i-1}^C \varphi(x_{i-1}^C)$ converges. Further, one has \[ \sum_1^\infty \mathrm{E}(\gamma_{t-1}^C \xi_t)^2 = S\cdot \sum_1^\infty \mathrm{E}(\gamma_{t-1}^C)^2 < \infty, \] hence the martingale $\sum_1^t \gamma_{i-1}^C \xi_i$ converges. This implies that $\{x_t^C\}$ also converges. Define the events $A_C = \{ \sum_t \gamma_t \le C\}$ and $A_\infty = \{ \sum_t \gamma_t < \infty \}$. One has $A_\infty = \cup_C A_C$. If $\sum_t \gamma_t \le C$ then $x_t^C=x_t$ for any $t$; this means that $\;\mathbb{I}(A_C)\cdot(x_t^C - x_t)=0$ for any $t$ and $C$. The sequence $\{ \;\mathbb{I}(A_C) x_t^C\}$ converges, therefore the sequence $\{\;\mathbb{I}(A_C)x_t\}$ also converges, and passing to the limit $C\to\infty$ one obtains that $\{ \;\mathbb{I}(A_\infty) x_t\}$ converges. This means exactly that if $\sum_t \gamma_t < \infty$ then $\{ x_t \}$ converges. \hfill$\Box$ \begin{lemma} If $\lim_{t\to\infty} x_t = x$ then $x \in V_-^{[\mathrm{k}]}$. \end{lemma} \textit{Proof.} Note that, using \A{A3}\,(a), it is easy to show that there exists $\delta_0 >0$ such that $\mathrm{P}(\xi_1 \not \in [x-L/2,\, x+L/2]) > \delta_0$, whatever $x \in \mathbb{R}$. Next, for any $x\not \in V_-^{([\mathrm{k}])}$ there exist $w(x)>0$ and $0 < \epsilon(x) < L/4$ such that the following holds: for any two random variables $\phi_1$ and $\phi_2$ satisfying the relations $|\phi_l - \varphi(x)| \le \epsilon(x)$,\, $l=1,2$ one has \[ \mathrm{P}( (\phi_1+\xi_1)(\phi_2+\xi_2) >0 ) > \frac{\ln(1/d)+w(x)}{\ln u + \ln(1/d)}. \] Choose a countable set of intervals $U_i = (\varphi(x_i)-\epsilon(x_i),\ \varphi(x_i)+\epsilon(x_i))$ covering the set $\varphi(\mathbb{R}\setminus V_-^{[\mathrm{k}]})$, and denote $w_i := w(x_i)$. Fix $i$ and $s\in\{0,\, 1,\, 2,\ldots\}$, and define the auxiliary process $x_t^{(is)}$, $\gamma_t^{(is)}$ by formulas: \vspace{1mm} if $t<s$ then $x_t^{(is)} = x_t$, \ and \ if $t \ge s$ then \begin{eqnarray}\label{eq16} x_t^{(is)} = \left\{ \begin{array}{ll} x_{t-1}^{(is)} - \gamma_{t-1}^{(is)}\, y_t^{(is)} & \textrm{ if } \ \varphi(x_{t-1}^{(is)} - \gamma_{t-1}^{(is)}\, y_t^{(is)}) \in U_i,\\ x_i & \textrm{ elsewhere}; \end{array} \right. \end{eqnarray} \begin{equation}\label{eq17} y_t^{(is)} = \varphi(x_{t-1}^{(is)}) + \xi_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \hspace{35mm} \end{equation} \begin{eqnarray}\label{eq18} \gamma_t^{(is)} = \left\{ \begin{array}{l@{\textrm{ if }}l} \min\{u\gamma_{t-1}^{(is)},\, \bar{\mathrm{g}}\} \ \ \ & \ y_{t-1}^{(is)}\, y_t^{(is)} > 0,\\ d \gamma_{t-1}^{(is)} \ \ \ & \ y_{t-1}^{(is)}\, y_t^{(is)} \le 0. \end{array} \right. \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \end{eqnarray} So, as $t \ge s$,\, $\varphi(x_t^{(is)})$ is forced to be contained in $U_i$. For $t\ge s+2$, using that $y_{t-1}^{(is)} = \varphi(x_{t-2}^{(is)})+\xi_{t-1}$,\, $y_t^{(is)} = \varphi(x_{t-1}^{(is)}) + \xi_t$,\, $\varphi(x_{t-2}^{(is)}) \in U_i$, one obtains that \[ \mathrm{P}( y_{t-1}^{(is)}\, y_t^{(is)} > 0) > \frac{\ln(1/d) + w_i}{\ln u + \ln(1/d)} \] and \[ \mathrm{P}( y_{t-1}^{(is)}\, y_t^{(is)} \le 0 ) < \frac{\ln u - w_i}{\ln u + \ln (1/d)}, \] hence $$ \displaystyle \mathrm{E}[\ln u\cdot \;\mathbb{I}(y_{t-1}^{(is)}\, y_t^{(is)}>0)\, +\, \ln d\cdot \;\mathbb{I}(y_{t-1}^{(is)}\, y_t^{(is)}\le 0) ] > $$ $$ \displaystyle > \ln u \cdot \frac{\ln(1/d) + w_i}{\ln u + \ln(1/d)}\ +\ \ln d \cdot \frac{\ln u - w_i}{\ln u + \ln(1/d)} = w_i. $$ Consider variables $\phi_1=f_1(\xi_1, \xi_2)$ and $\phi_2 = f_2(\xi_1,\xi_2)$ providing a solution of the (deterministic) minimization problem: \[ (\phi_1 + \xi_1)(\phi_2+\xi_2) \to \min, \] subject to \[ \begin{array}{l} |\phi_1 - \varphi(x_i)| \le \epsilon(x_i) \\ |\phi_2 - \varphi(x_i)| \le \epsilon(x_i),\\ \end{array} \] and denote $Y_{t-1}^1 = f_1(\xi_{t-1},\xi_t) + \xi_{t-1}$, $Y_{t}^2 = f_2(\xi_{t-1},\xi_t) + \xi_{t}$, $\eta_t = \ln u \cdot \;\mathbb{I}(Y_{t-1}^1 Y_{t-1}^2 > 0) + \ln d \cdot \;\mathbb{I}( Y_{t-1}^1 Y_{t-1}^2 \le 0)$. One has (i) $\eta_t \le \ln u \cdot \;\mathbb{I}( y_{t-1}^{(is)}\, y_t^{(is)} > 0) + \ln d \cdot \;\mathbb{I}( y_{t-1}^{(is)}\, y_t^{(is)} \le 0)$; (ii) $\eta_t$ are identically distributed, and $\mathrm{E} \eta_t \ge w_i$; (iii) the set of random variables $\{\eta_t,\ t \textrm{ even},\ t \ge s+2 \}$ as well as the set $\{\eta_t,\ t \textrm{ odd},\ t \ge s+2 \}$, are mutually independent. From (ii)--(iii) it follows that almost surely $\sum_t \eta_t = +\infty$, and from (i) it follows that \[ \sum_t [\ln u \cdot \;\mathbb{I}(y_{t-1}^{(is)}\, y_t^{(is)} > 0 ) + \ln d \cdot \;\mathbb{I}( y_{t-1}^{(is)}\, y_t^{(is)} \le 0)]= +\infty, \] so, by virtue of (\ref{eq18}), $\gamma^{(is)}$ does not go to zero. Thus, there exists a random value $\chi > 0$ such that for infinitely many values of $t$,\, $\gamma_t^{(is)} \ge \chi$. Define a sequence of stopping times $\tau_0$, $\tau_1$, $\tau_2, \ldots$ inductively, letting $\tau_0=0$ and $\tau_j = \inf\{ t > \tau_{j-1} : \gamma_t^{(is)} \ge \chi\}$ for $j\ge 1$. The events $B_j = \{ |\xi_{\tau_j+1} + \varphi(x_i)|>L/2\}$ happen with probability more that $\delta_0$ (recall the remark done in the beginning of proof), and every event $B_j$, $j\ge 2$ does not depend on the set of events $\{ B_1, \ldots, B_{j-1} \}$. Therefore, for infinitely many values of $j$, $B_j$, takes place, i.e., $|\xi_{\tau_j+1} + \varphi(x_i)| > L/2$, and hence, taking into account that $|y_{\tau_j+1}| \ge |\xi_{\tau_j+1} + \varphi(x_i)| - |\varphi(x_{\tau_j}) - \varphi(x_i)|$ and $|\varphi(x_{\tau_j})-\varphi(x_i)| < \epsilon(x_i) < L/4$, for these values of $j$ one has $|y_{\tau_j+1}| \ge L/4$. Thus, one concludes that \begin{equation}\label{eqast} \textrm{for infinitely many values of }j, \ \ |\gamma_{\tau_j} y_{\tau_j + 1}| \ge \chi\, L/4. \end{equation} Suppose that $x_t$ converges to a point from $\mathbb{R}\setminus V_-^{[\mathrm{k}]}$, then for some $i$ and $s$ one has $x_t \in U_i$ as $t \ge s$, hence the process $x_t^{(is)}$, $\gamma_t^{(is)}$ coincides with $x_t$,\, $\gamma_t$, and therefore $\gamma_t\, y_{t+1} \to 0$ as $t \to \infty$. The last relation contradicts (\ref{eqast}), thus Lemma 2 is proved. \hfill$\Box$ \begin{lemma} Let $\sum_t \gamma_t = \infty$. Then for any open set ${\cal O}$ containing $\mathrm{Z}$ there exists a positive constant $g=g({\cal O})$ such that either (i) for some $t$,\, $x_t\in{\cal O}$, or (ii) for some $t$, $|x_t|<R$ and $\gamma_t > g$. \end{lemma} \textit{Proof.} Designate by $f$ the primitive of $\varphi$ such that $\inf_x f(x) = 0$. Define the stopping time \[ \tau = \tau({\cal O},g) := \inf \{ t : \textrm{ either (i) } x_t \in {\cal O}, \textrm{ or (ii) } |x_t|<R \textrm{ and } \gamma_t \ge g\}. \] The value of $g \in (0,\bar{\mathrm{g}})$ will be specified below. Consider the sequence $\mathrm{E}_t = \mathrm{E}[ f(x_t) \;\mathbb{I}(t<\tau)]$. Introducing shorthand notation $f(x_t) =: f_t$,\, $\;\mathbb{I}(t<\tau)=: I_t$,\, $f'(x_t)=:f_t'=\varphi_t$, and using that $I_t \le I_{t-1}$, one gets \begin{equation}\label{eq20} E_t - E_{t-1} = \mathrm{E}[f_t \;\mathbb{I}_t - f_{t-1} \;\mathbb{I}_{t-1}]\, \le\, \mathrm{E}[(f_t - f_{t-1}) \;\mathbb{I}_{t-1}]. \end{equation} Next, we utilize the Taylor decomposition $$ f_t = f(x_{t-1} - \gamma_{t-1} y_t) = f_{t-1} - f'_{t-1}\, \gamma_{t-1} y_t + \frac 12\, f''(x')\, \gamma_{t-1}^2 y_t^2, $$ $x'$ being some point between $x_{t-1}$ and $x_t$. Substituting $y_t = \varphi_{t-1}+\xi_t$ and recalling that $f'_{t-1}=\varphi_{t-1}$ and $f''(x')=\varphi'(x') \le M $, one obtains \begin{equation}\label{eq21} f_t - f_{t-1} \le -\gamma_{t-1}\, \varphi_{t-1} (\varphi_{t-1} + \xi_t) + {M\over 2}\, \gamma_{t-1}^2\, (\varphi_{t-1} + \xi_t)^2. \end{equation} Using (\ref{eq20}) and (\ref{eq21}) and taking into account that each of the values $\gamma_{t-1}$,\, $\varphi_{t-1}$,\, $\;\mathbb{I}_{t-1}$ is mutually independent with $\xi_t$ (see \A{A1}), one gets \begin{equation}\label{eq23} \begin{array}{l} E_t - E_{t-1} \le \mathrm{E}[ (-\gamma_{t-1}\, \varphi_{t-1}^2 - \gamma_{t-1}\, \varphi_{t-1} \xi_t + {M\over 2} \gamma_{t-1}^2\, \varphi_{t-1}^2 + M\gamma_{t-1}^2\, \varphi_{t-1} \xi_t + {M\over 2} \gamma_{t-1}^2\, \xi_t^2) \;\mathbb{I}_{t-1}] =\\ = \mathrm{E}[ (-\varphi_{t-1}^2 + \frac M2 \gamma_{t-1}\, \varphi_{t-1}^2 + {M\over 2} \gamma_{t-1} S) \gamma_{t-1} \;\mathbb{I}_{t-1}] = \\ = \mathrm{E}[ (-\varphi_{t-1}^2(1-M\gamma_{t-1}/2) + M\gamma_{t-1} S/2) \gamma_{t-1} \;\mathbb{I}_{t-1}]. \end{array} \end{equation} If $\;\mathbb{I}_{t-1}=1$ then~ either~ (i)~ $x_{t-1} \in [-R,R]\setminus{\cal O}$ and $\gamma_{t-1} < g$,~~ or~ (ii)~ $|x_{t-1}| \ge R$. In the case (i) one has \begin{equation}\label{eq23.1} -\varphi_{t-1}^2(1-M\gamma_{t-1}/2) + M\gamma_{t-1}S/2 \le -c_0(1-Mg/2) + M g S/2 =: -c'_g, \end{equation} where $c_0 := \inf\{ |\varphi(x)| : x\in[-R,R]\setminus{\cal O} \}$; obviously, $c_0>0$. Let us fix a $g \in (0,\bar{\mathrm{g}})$ such that $c'_g>0$. In the case (ii), designating $b_0 := \inf_{|x|\ge R} \varphi^2(x)$, one has \begin{equation}\label{eq23.2} -\varphi_{t-1}^2(1-M\gamma_{t-1}/2) + M\gamma_{t-1} S/2 \le -b_0(1-M\bar{\mathrm{g}}/2)+M\bar{\mathrm{g}} S/2 =: -c''. \end{equation} Using \A{A6}, one gets that $c''>0$. Denote $c=\min\{c'_g,c''\}$. The relations (\ref{eq23.1}) and (\ref{eq23.2}) imply that if $\;\mathbb{I}_{t-1}=1$ then $-\varphi_{t-1}^2(1-M\gamma_{t-1}/2) + M\gamma_{t-1} S/2 \le -c < 0$, hence, by virtue of (\ref{eq23}), \begin{equation}\label{eq24} E_t - E_{t-1} \le -c \cdot \mathrm{E}[\gamma_{t-1} \;\mathbb{I}_{t-1}]. \end{equation} Summing up both sides of (\ref{eq24}) over $t=1,\ldots,s$ and denoting $\;\mathbb{I}_\infty = \;\mathbb{I}(\tau=\infty)=\min_t \;\mathbb{I}_t$, one obtains \[ \mathrm{E}_s - \mathrm{E}_0 \le -c \cdot \mathrm{E} \left[\sum_{i=0}^{s-1} \gamma_i \cdot \;\mathbb{I}_\infty \right]. \] One has $\mathrm{E}_s \ge 0$, and $x_0$ is bounded, hence $E_0 < \infty$. Thus, for arbitrary $s$ \[ \mathrm{E} \left[\sum_{i=0}^{s-1} \gamma_i \cdot \;\mathbb{I}_\infty \right] \le \frac{\mathrm{E}_0}{c} < \infty. \] This implies that a.s. either $\sum_0^\infty \gamma_i < \infty$, or $\tau = \infty$. Lemma 3 is proved. \hfill$\Box$ \vspace{2mm} Denote $c_1 := 1- M \bar{\mathrm{g}}/2$. Recall that $f$ is the primitive of $\varphi$ such that $\inf_x f(x) = 0$; the assumption \A{A6} implies that $\lim_{x\to\pm \infty} f(x) = +\infty$. Denote $H:= \sup_{|x|\le R} f(x)$. Denote also~ $c_3:= \bar{\mathrm{g}} \cdot \sup\{ |\varphi(x)| : f(x) \le H\} + 1$,~ $z^{l} := \inf \{ x : f(x) \le H \} - c_3$,\ $z^{r} := \sup\{ x : f(x) \le H\} + c_3$,~ $c_2 := \inf\{ |\varphi(x)| : x \in [z^{l},\, z^{r}] \setminus{\cal O}\}$,~ and~ $\mathrm{K} := \sup\{ |\varphi(x)| : x\in[z^{l},\, z^{r}]\}$.~ Obviously, $c_1 > 0$ and $\mathrm{K}\ge c_2 > 0$. Fix an open set ${\cal O}$ containing $\mathrm{Z}$. Let $g > 0$,\, $0 < w < 1$. We shall say that a (finite or infinite) deterministic sequence $\{z_0, z_1, z_2, \ldots \}$ is $(g,\, w)$-admissible if $|z_0|\le R$ and there exist deterministic sequences $\{q_t\},$ $\{h_t\}$ such that 1) $|h_t| \le w$; 2) if $\{ z_0,z_1,\ldots,z_t\} \subset[z^{l},\,z^{r}]\setminus{\cal O}$~ then~ $g d^2 \le q_s \le \bar{\mathrm{g}}$,~ $s=0,1,\ldots,t$; 3) $z_t = z_{t-1} - q_{t-1}\, \varphi(z_{t-1})-h_t$,~ $t=1,2,\ldots$. \begin{proposition} There exists constants $t_0$ and $w$ such that any $(g,\, w)$-admissible sequence $\{z_t,\ t=0,\, 1,\ldots,t_0\}$ has non-empty intersection with ${\cal O}$. \end{proposition} \textit{Proof.} Let $w:= \min\{ 1,\, g d^2 c_2^2 c_1/(2\mathrm{K})\}$. Designate $\tilde{t}=\inf\{ t:z_t\in{\cal O}\}$;\, $\tilde{t}$ takes values from $\{0,\, 1, \ldots, t_0,\, +\infty\}$. We shall use shorthand notation $f_t := f(z_t)$, $f'_t= \varphi_t := \varphi(z_t)$. One has \begin{equation}\label{eq25} f_t = f(z_{t-1}-q_{t-1} \varphi_{t-1} - h_t) = f(z_{t-1} - q_{t-1} \varphi_{t-1}) - f'(\tilde z).h_t, \end{equation} where $\tilde z$ is a point between $z_{t-1} - q_{t-1} \varphi_{t-1}$ and $z_{t-1} - q_{t-1} \varphi_{t-1} -h_t$. Next, one has \begin{equation}\label{eq26} f(z_{t-1}-q_{t-1} \varphi_{t-1}) = f_{t-1} - f'_{t-1} q_{t-1} \varphi_{t-1} + \frac 12 f''(\hat z)\, q_{t-1}^2 \varphi_{t-1}^2, \end{equation} where $\hat z$ is a point between $z_{t-1}$ and $z_{t-1}-q_{t-1} \varphi_{t-1}$. We are going to prove by induction that \begin{equation}\label{eq29} \textrm{if } 0\le s \le \tilde{t} \ \textrm{ then } \ f_s \le H-s\cdot g d^2 c_2^2 c_1 / 2. \end{equation} For $s=0$,~ (\ref{eq29}) follows from the condition $|z_0|\le R$ and the definition of $H$. Now, let $1 \le t \le \tilde{t}$;~ suppose that formula (\ref{eq29}) is true for $0 \le s \le t-1$ and prove it for $s = t$. For $0 \le s \le t-1$, one has $f(z_s) \le H$,\, $z_s \not \in {\cal O}$, therefore $z_s \in [z^{l},\, z^{r}] \setminus {\cal O}$; hence, by virtue of 2), $g d^2 \le q_s \le \bar{\mathrm{g}}$ for $0 \le s \le t-1$. One has $f(z_{t-1}) \le H$,\, $|q_{t-1} \varphi_{t-1}| \le \bar{\mathrm{g}} \cdot \sup\{ |\varphi(x)| : f(x) \le H\}$, and $|h_t| \le w \le 1$, hence $|q_{t-1} \varphi_{t-1}| \le c_3$,\, $|q_{t-1} \varphi_{t-1} + h_t| \le c_3$, and so, $z_{t-1} - q_{t-1} \varphi_{t-1} \in [z^{l},\, z^{r}]$,\, $z_{t-1}-q_{t-1}\varphi_{t-1} -h_t \in [z^{l},\, z^{r}]$, thus $\tilde z$ also belongs to $[z^{l},\, z^{r}]$. This implies that $|\varphi(\tilde z)| = |f'(\tilde z)| \le \mathrm{K}$. Then, combining (\ref{eq25}) and (\ref{eq26}) and using that $|h_t| \le w$ and $|f''(\hat z)| = |\varphi'(\hat z)|\le M$, one obtains \begin{equation}\label{eq27} f_t \le f_{t-1} - q_{t-1} \varphi^2_{t-1}(1-{1\over 2} q_{t-1} M) + w\mathrm{K}. \end{equation} One has $z_{t-1} \in [z^{l},\, z^{r}] \setminus {\cal O}$, hence $|\varphi(z_{t-1})| = |\varphi_{t-1}| \ge c_2$. Using also that $q_{t-1} \ge g d^2$,\, $1-{1 \over 2} q_{t-1} M \ge c_1$, and $w\mathrm{K} \le g d^2 c_2^2 c_1/2$, one gets from (\ref{eq27}) that \[ f_t \le f_{t-1} - g d^2 c_2^2 c_1 / 2, \] and using the induction hypothesis, one concludes that \[ f_t \le H - t \cdot g d^2 c_2^2 c_1 /2. \] Formula (\ref{eq29}) is proved. Let $t_0 := \lfloor 2H/(g d^2 c_2^2 c_1) \rfloor + 1$; here $\lfloor z \rfloor$ stands for the integral part of $z$. Then, taking into account that $f_s \ge 0$, from (\ref{eq29}) one concludes that $\tilde{t} < t_0$, thus Proposition 1 is proved. \hfill$\Box$. \begin{proposition} If $\gamma_{t-1} < 1/(3M)$,\, $|\xi_t|<c_2$,\, $|\xi_{t+1}|< c_2$,\, $x_{t-1}$ and $x_t$ belong to $[z^{l},\, z^{r}] \setminus {\cal O}$,~ then $\gamma_{t+1} \ge \gamma_t$. \end{proposition} \textit{Proof.} Using notation $\varphi_t := \varphi(x_t)$, one gets $$ \varphi_t = \varphi(x_{t-1} - \gamma_{t-1}(\varphi_{t-1}+\xi_t)) = \varphi_{t-1} - \varphi'(\tilde x) \cdot \gamma_{t-1}(\varphi_{t-1}+\xi_t), $$ where $\tilde x$ is a point between $x_{t-1}$ and $x_t$. Therefore, \[ \varphi_{t-1} \varphi_t = \varphi^2_{t-1} \cdot[ 1- \varphi'(\tilde x) \gamma_{t-1} \cdot(1+\xi_t / \varphi_{t-1})]. \] Using that $|\varphi'(\tilde x)| \le M$,\, $\gamma_{t-1} < 1/(3M)$,\, $|\xi_t| < c_2$,\, $|\varphi_{t-1}| \ge c_2$, one obtains $1-\varphi'(\tilde x)\, \gamma_{t-1} \cdot (1 + \xi_t / \varphi_{t-1}) \ge 1/3$, hence $\varphi_{t-1} \varphi_t >0$. Further, using that $|\xi_t| < c_2$,\, $|\xi_{t+1}|<c_2$,\, $|\varphi_{t-1}| \ge c_2$,\, $|\varphi_t| \ge c_2$, one gets \[ y_{t}\, y_{t+1} = \varphi_{t-1} \varphi_t \cdot(1+\xi_t/\varphi_{t-1})(1+\xi_{t+1}/\varphi_t) > 0. \] This implies that $\gamma_{t+1} = \min\{ u \gamma_t, \bar{\mathrm{g}}\} \ge \gamma_t$. \hfill$\Box$ \begin{lemma For any open set ${\cal O}$, containing $\mathrm{Z}$, and any $g > 0$~ there exists $\delta = \delta({\cal O}, g) > 0$ such that \[ \text{if } \ |x_0| \le R, \ \gamma_0 \ge g \ \text{ then } \ \ \mathrm{P}(\textrm{for some } t, \ x_t \in {\cal O}) \ge \delta. \] \end{lemma} \textit{Proof.} Without loss of generality suppose that $g < 1/(3M)$. Define the event \[ A := \{ |\xi_i| < \min\{ c_2,\, w/\bar{\mathrm{g}} \}, \ i=1,2,\ldots,t_0\}, \] where $w$ and $t_0$ are the same as in the proof of Proposition 1:~ $w = \min\{ 1,\, g d^2 c_2^2 c_1/(2\mathrm{K})\}$,\ $t_0 = \lfloor 2H/(g d^2 c_2^2 c_1) \rfloor + 1$. Denote \[ \delta := P(A) = ( \mathrm{P}( |\xi_1| < \min\{ c_2,\, w/\bar{\mathrm{g}}\}))^{t_0}; \] by virtue of \A{A3}\,(a), $\delta > 0$. Let us show that for any elementary event $\omega \in A$, the sequence $\{ z_t = x_t(\omega),\ t=0, 1, \ldots, t_0\}$ is $(g,\, w)$-admissible. One has $|z_0|=|x_0(\omega)| < R$. Further, one has $z_t = z_{t-1} - q_{t-1} \varphi(z_{t-1})- h_t$, with $q_{t-1} = \gamma_{t-1}(\omega)$,\, $h_t = \gamma_{t-1}(\omega)\, \xi_t(\omega)$, and using that $\gamma_{t-1}(\omega) \le \bar{\mathrm{g}}$ and $|\xi_t(\omega)| < \omega / \bar{\mathrm{g}}$, one gets $|h_t| \le w$. Thus, conditions 1) and 3) are verified. Now, let $\{z_0, z_1, \ldots,z_t\} \subset [z^{l},\, z^{r}] \setminus {\cal O}$,~ $t \le t_0$. Let $s_0 \in \{ 0,1,2,\ldots,t\}$ be the minimal value such that $q_{s_0} = \min\{ q_0, q_1, \ldots, q_t \}$. If $s_0=0$ then $\min\{ q_0, q_1,\ldots, q_t \} = q_0 = \gamma_0(\omega)\ge g \ge g d^2$. If $s_0=1$ then $\min\{ q_0, q_1, \ldots, q_t\} = q_1 = \gamma_1(\omega) \ge g d \ge g d^2$. If $s_0 \ge 2$ then $\gamma_{s_0-2}(\omega) \ge 1/(3M)$; otherwise, using that $|\xi_{s_0-1}| < c_2$,\, $|\xi_{s_0}| < c_2$,\, $x_{s_0-2}(\omega)$ and $x_{s_0-1}(\omega)$ belong to $[z^{l},\, z^{r}] \setminus {\cal O}$, and applying Proposition 2, one would conclude that $\gamma_{s_0}(\omega) \ge \gamma_{s_0-1} (\omega)$, which contradicts the definition of $s_0$. Thus, $\gamma_{s_0}(\omega) \ge 1/(3M) \cdot d^2 \ge g d^2 $, and therefore, $\min\{ q_0, q_1, \ldots, q_t \} = \gamma_{s_0}(\omega) \ge g d^2$. So, the condition 2) is also verified. Now, applying Proposition 1 to the $(g,\, w)$-admissible sequence $\{z_t\}$, one concludes that there exists a non-negative $\tau \le t_0$ such that $z_{\tau} = x_\tau(\omega) \in {\cal O}$. This implies that \[ \mathrm{P}(\textrm{for some } t, \ x_t \in {\cal O}) \ge \mathrm{P}(A) = \delta. \] \hfill$\Box$ \begin{lemma If $\sum_t \gamma_t = \infty$ then for any open set ${\cal O}$ containing $\mathrm{Z}$ there exists $t$ such that $x_t \in {\cal O}$. \end{lemma} \textit{Proof.} Let us fix an open set ${\cal O} \supset \mathrm{Z}$, and denote $\delta = \delta({\cal O}, g({\cal O}))$. Combining Lemma 3 and Lemma 4, one concludes that for any ${\cal O} \supset \mathrm{Z}$ there exists $\delta > 0$ such that whatever the initial conditions $x_0$, $\gamma_0$, $\gamma_1$, \[ \mathrm{P}(\textrm{for some } t, \ x_t \in {\cal O} \,{\Big |}\, \sum_t \gamma_t = \infty) > \delta. \] Then one can choose a measurable integer-valued function $n(\cdot,\cdot,\cdot)$ defined on $\mathbb{R} \times (0,\bar{\mathrm{g}}] \times (0,\bar{\mathrm{g}}]$ such that for $\nu=n(x_0,\gamma_0,\gamma_1)$ one will have \[ \mathrm{P}(\textrm{for some } t\le \nu, \ x_t \in {\cal O} \,{\Big |}\, \sum_t \gamma_t=\infty) > \delta/2 \] Designate \[ \bar p = \sup \mathrm{P}(\textrm{for all } t, \ x_t \not \in {\cal O} \,{\Big |}\, \sum_t \gamma_t = \infty), \] the supremum being taken over all the initial conditions $x_0$, $\gamma_0$, $\gamma_1$. Fix $x_0$, $\gamma_0$, $\gamma_1$, then \begin{equation}\label{eq30} \begin{array}{l} \mathrm{P}(\textrm{for all } t, \ x_t \not \in {\cal O} \,{\Big |}\, \sum_t \gamma_t = \infty ) =\\ = \mathrm{P}(\textrm{for all } t > \nu, \ x_t \not \in {\cal O} \,{\Big |}\, \textrm{for all } t \le \nu, \ x_t \not \in {\cal O} \textrm{ and } \sum_t \gamma_t =\infty) \cdot \\ \cdot \mathrm{P}(\textrm{for all } t \le \nu, \ x_t \not \in {\cal O}\, | \sum_t \gamma_t = \infty) \le \bar p\, (1-\delta/2). \end{array} \end{equation} Taking supremum of the left hand side of (\ref{eq30}) over all $(x_0, \gamma_0, \gamma_1) \in \mathbb{R} \times (0,\bar{\mathrm{g}}] \times (0,\bar{\mathrm{g}}]$, one obtains $\bar p \le \bar p\, (1- \delta/2)$, hence $\bar p = 0$. Lemma~5 is proved. \hfill$\Box$. \vspace{2mm} Denote ${\cal O}_* = \{ x: |\varphi(x)| < L/2 \}$. \begin{lemma For any open bounded sets $\mathcal{O}$,\, $\mathcal{O}_1$ such that $\bar{\mathcal{O}} \subset \mathcal{O}_1 \subset {\cal O}_*$ and for any $w > 0$ there exists $\delta = \delta({\cal O}, {\cal O}_1, w)> 0$ such that $$ \text{if } \ x_0 \in \mathcal{O} \text{ then } \ \mathrm{P}(\textrm{for some } n,\ x_{n} \in {\cal O}_1 \text{ and } \gamma_{n} < w) \ge \delta. $$ \end{lemma} \textit{Proof.} Denote $n = \lfloor \frac{\ln\bar{\mathrm{g}} - \ln w}{\ln (1/d)} \rfloor + 2$. Denote also $$ \varepsilon = \min \left\{ \frac{L}{2},\ \frac{ \partial (\mathcal{O},\, \mathbb{R} \setminus \mathcal{O}_1)}{n \bar{\mathrm{g}}} \right\}, $$ where $\partial (A,B) := \sup_{x\in A} \inf_{y\in B} |x-y|$ for arbitrary sets of real numbers $A$,\, $B$. Using assumption \A{A3}\,(a), one obtains that there exists $\delta_1 > 0$ such that for any $x \in {\cal O}_1$ and for any integer $t$, $$ \mathrm{P} \left( (-1)^{t-1} \varphi(x) < (-1)^t \xi_1 < (-1)^{t-1} \varphi(x) + \varepsilon \right) \ge \delta_1. $$ This implies that if $x_0 \in {\cal O}$ then $$ \mathrm{P} (0 < (-1)^t y_t < \varepsilon,\ \text{dist}(x_{t-1},\, {\cal O}) < (t - 1) \bar{\mathrm{g}} \varepsilon,\ t = 1,\, 2,\ldots, n+1) \ge \delta_1^{n+1}. $$ Denoting $\delta = \delta_1^{n+1}$, one concludes that the following statements (i) and (ii) hold with probability at least $\delta$: (i) dist$(x_n,\, {\cal O}) < n \bar{\mathrm{g}} \varepsilon \le$ dist$({\cal O},\, \mathbb{R} \setminus {\cal O}_1)$, hence $x_n \in {\cal O}_1$; (ii) as $t = 2,\, 3,\ldots, n+1$, one has $y_{t-1} y_t < 0$, hence $\gamma_t = d \gamma_{t-1}$, therefore $\gamma_{n} = d^{n-1} \gamma_1 \le d^{n-1} \bar{\mathrm{g}} < w$.\\ Lemma~6 is proved. \hfill$\Box$ \begin{lemma If $\sum_t \gamma_t = \infty$,\ ${\cal O}$ is an open set containing $\mathrm{Z}$, and $w > 0$ then for some $t$,~ $x_{t-1} \in {\cal O}$ and $\gamma_t < w$. \end{lemma} \textit{Proof.} Without loss of generality, suppose that ${\cal O}$ is bounded and ${\cal O} \subset {\cal O}_*$. Choose an open set ${\cal O}_1$ such that $\mathrm{Z} \subset {\cal O}_1$,\, $\bar{\cal O}_1 \subset {\cal O}$;~ applying Lemmas 5 and 6, one gets that for $\delta = \delta({\cal O}_1, {\cal O}, w)$ and for arbitrary initial conditions, $$ \mathrm{P} (\textrm{for some } t,\ x_{t} \in {\cal O} \textrm{ and } \gamma_t < w) > \delta. $$ Repeating the argument of Lemma 5, one concludes that there exists $t$ such that $x_{t} \in {\cal O}$ and $\gamma_t < w$. \hfill$\Box$ \vspace{2mm} From now on we suppose that $\mathrm{k} > k_+(0)$. Choose $k'$ such that $k_+(0) < k' < \mathrm{k}$; using \A{A3}\,(b), one obtains that for some $\varepsilon_0 > 0$,\, $\mathrm{P} ( \xi_1 \xi_2 > 0, \text{ or } |\xi_1| < \varepsilon_0, \text{ or } |\xi_2| < \varepsilon_0) \le k'$. Denote ${\cal O}_0 = \{ x:\, |\varphi(x)| < \varepsilon_0 \}$ and $\tau = \inf \{ t: \ x_t \not\in {\cal O}_0 \}$. Without loss of generality, suppose that ${\cal O}_0$ is bounded. \begin{lemma Suppose that $\mathrm{k} > k_+(0)$,~ then there exist a constant $b > 0$ and a monotone decreasing function $p(\cdot)$ such that $\lim_{a\to+\infty} p(a) = 0$ and $$ \text{if } \ \gamma_0 < w \text{ then } \ \mathrm{P} (\ln\gamma_t < \ln v - bt \text{ for all } t < \tau ) > 1 - p(v/w). $$ \end{lemma} \textit{Proof.} Define the sequences $\{ \rho_t \}$ and $\{ \sigma_t \}$ by \begin{eqnarray*} \rho_t &=& \ln u \cdot \;\mathbb{I}(\xi_{t-1} \xi_t > 0, \text{ or } |\xi_{t-1}| < \varepsilon_0, \text{ or } |\xi_t| < \varepsilon_0) +\\ &+& \ln d \cdot \;\mathbb{I}(\xi_{t-1} \xi_t \le 0 \ \, \& \, \ |\xi_{t-1}| \ge \varepsilon_0 \ \, \& \ \, |\xi_t| \ge \varepsilon_0), \end{eqnarray*} $$ \sigma_t = \ln w + \sum_{i=1}^t \rho_i. $$ Using (\ref{eqal4n}) and definition of $\tau$, one obtains that for all $t < \tau$,\, $\gamma_t \le \sigma_t$. The variables $\rho_t$ are identically distributed, take the values $\ln u$ and $\ln d$, and \begin{eqnarray*} E\rho_t &=& \ln u \cdot \mathrm{P}(\xi_{t-1} \xi_t > 0, \text{ or } |\xi_{t-1}| < \varepsilon_0, \text{ or } |\xi_t| < \varepsilon_0) +\\ &+& \ln d \cdot \mathrm{P}(\xi_{t-1} \xi_t \le 0 \ \, \& \, \ |\xi_{t-1}| \ge \varepsilon_0 \ \, \& \ \, |\xi_t| \ge \varepsilon_0) \le\\ &\le& \ln u \cdot k' + \ln d \cdot (1 - k') < \ln u \cdot \mathrm{k} + \ln d \cdot (1-\mathrm{k}) = 0. \end{eqnarray*} Moreover, the variables in the set $\{ \rho_t, \ t \text{ even} \}$, as well as the variables in the set $\{ \rho_t, \ t \text{ odd} \}$, are independent. Denote $b = -E\rho_t/2$. One has $$ \mathrm{P} (\ln\gamma_t < \ln v - b t \ \text{ for all } t < \tau ) \ge \mathrm{P} (\sigma_t < \ln v - b t \ \text{ for all } t) = $$ $$ = \mathrm{P} (\sum_{i=1}^t (\rho_i + 2b) < \ln v - \ln w + b t \ \text{ for all } t ) \ge 1 - p(v/w), $$ where $p(a) = p_1(a) + p_2(a)$, $$ p_1(a) = \mathrm{P} \left( {\sum_{1\le i\le t}}' (\rho_i + 2 b) \ge \frac {\ln a}2 + \frac b2\, t \ \text{ for all } t \right), $$ $$ p_2(a) = \mathrm{P} \left( {\sum_{1\le i\le t}}'' (\rho_i + 2 b) \ge \frac {\ln a}2 + \frac b2\, t \ \text{ for all } t \right); $$ the sum $\sum'$ ($\sum''$) is taken over the even (odd) values of $i$. Both $\sum'$ and $\sum''$ are sums of i.i.d.r.v. with zero mean, hence both $p_1(a)$ and $p_2(a)$ tend to zero as $a \to +\infty$. Lemma 8 is proved. \hfill$\Box$ Define the stopping times $\tau_v = \inf \{ t: \ x_t \not\in {\cal O}_0 \text{ or } \ln\gamma_t \ge \ln v - bt \}$. Recall that $f$ is the primitive of $\varphi$ such that $\inf_x f(x) = 0$. Fix an open set ${\cal O}'$ such that $\mathrm{Z} \subset {\cal O}' \subset {\cal O}_0$ and $\sup_{x\in{\cal O}'} f(x) < \inf_{x\not\in{\cal O}_0} f(x)$, and denote $\delta = \inf_{x\not\in{\cal O}_0} f(x) - \sup_{x\in{\cal O}'} f(x)$. \begin{lemma Let $\mathrm{k} > k_+(0)$,\ $x_0 \in {\cal O}'$, and $\gamma_0 < w$, then $$ \mathrm{P} (\tau_v < \infty ) \le K\, v^2 + p(v/w); $$ here $K$ is a positive constant, and $p(\cdot)$ satisfies the statement of lemma 8. \end{lemma} \textit{Proof.} We shall use shorthand notation of Lemma 3: $f_t := f(x_t)$ and $\varphi_t := \varphi(x_t)$. According to (\ref{eq21}), one has $$ f_t - f_{t-1} \le -\gamma_{t-1} \varphi_{t-1} (\varphi_{t-1} + \xi_t) + {M\over 2}\, \gamma_{t-1}^2(\varphi_{t-1} + \xi_t)^2 \le $$ $$ \le -\gamma_{t-1} \varphi_{t-1} \xi_t + M \gamma_{t-1}^2 (\varphi_{t-1}^2 + \xi_t^2). $$ This implies that $f_t - f_1 \le Q_t' + Q_t''$, with $$ Q_t' = \big| \sum_{i=2}^{t} \gamma_{i-1} \varphi_{i-1} \xi_i \big|, \ \ \ \ \ Q_t'' = M \sum_{i=2}^{t} \, \gamma_{i-1}^2 (\varphi^{2}_{i-1} + \xi_i^2). $$ Using Lemma 8, one gets $$ \mathrm{P} (\tau_v < \infty)\, \le\, p(v/w) + P' + P'', $$ where $$ P' = \mathrm{P} (Q'_{\tau_v} \ge {\delta}/{2}) \ \ \text{ and } \ \ P'' = \mathrm{P} (Q''_{\tau_v} \ge {\delta}/{2}). $$ According to the Chebyshev inequality, $$ P'\, \le\, \frac{4}{\delta^2}\, EQ'^2_{\tau_v} = \frac{4}{\delta^2} \sum_{i,j=1}^{\infty} E_{ij}, $$ where $$ E_{ij}\, =\, E \left[ \gamma_{i-1} \varphi_{i-1} \xi_i\, \;\mathbb{I}(i-1 < \tau_v) \cdot \gamma_{j-1} \varphi_{j-1} \xi_j\, \;\mathbb{I}(j-1 < \tau_v) \right]. $$ Using that the values $\gamma_i$,\, $\varphi_i$,\, $\xi_i$, and $\;\mathbb{I} (i < \tau_v)$ are $\mathcal{F}_{i}$-measurable, and using assumptions \A{A1} and \A{A2}, one obtains that for $i \ne j$,\, $E_{ij} = 0$, and for $i = j$, $$ E_{ii} = E \left[ \gamma_{i-1}^2 \varphi_{i-1}^2 \;\mathbb{I}(i-1 < \tau_v) \cdot \xi_i^2 \right] \le v^2 e^{-2bi} \sup_{x\in{\cal O}_0} \varphi^2(x) \cdot S. $$ Therefore, $$ P'\, \le\, \frac{4}{\delta^2} \sum_{i=2}^\infty E_{ii} \le \frac{4v^2 S}{\delta^2}\ \frac{e^{-4b}}{1 - e^{-2b}}\ \sup_{x\in{\cal O}_0} \varphi^2(x). $$ Similarly, $$ P''\, \le\, \frac{2}{\delta}\, EQ_{\tau_v}'' = \frac{2M}{\delta} \sum_{i=2}^\infty E \left[ \gamma^2_{i-1} (\varphi^{2}_{i-1} + \xi_i^2) \;\mathbb{I}(i-1 < \tau_v) \right] \le $$ $$ \le \frac{2M v^2}{\delta} \sum_{i=2}^\infty e^{-2bi} \left( \sup_{x\in{\cal O}_0} \varphi^2(x) + S \right) = \frac{2M v^2}{\delta}\ \frac{e^{-4b}}{1 - e^{-2b}} \left( \sup_{x\in{\cal O}_0} \varphi^2(x) + S \right). $$ Taking $$ K\, =\, \left[ \frac{4S}{\delta^2} \sup_{x\in{\cal O}_0} \varphi^2(x)\, +\, \frac{2M}{\delta} \left( \sup_{x\in{\cal O}_0} \varphi^2(x) + S \right) \right] \frac{e^{-4b}}{1 - e^{-2b}}, $$ one gets that $P' + P'' \le K\, v^2$. Lemma 9 is proved. \hfill$\Box$ \begin{lemma If $\mathrm{k} > k_+(0)$ then $\sum_t \gamma_t < \infty$. \end{lemma} \textit{Proof.} From the definition of $\tau_v$ one easily sees that if $\tau_v = \infty$ for some $v > 0$, then $\sum_t \gamma_t < \infty$. This implies that for any $v > 0$ \begin{equation}\label{1point} \mathrm{P} \left(\sum \gamma_t\, =\, \infty \right) \le \mathrm{P}(\tau_v = \infty). \end{equation} Further, by virtue of Lemma 9, if $x_0 \in {\cal O}'$ and $\gamma_0 < w$ then \begin{equation}\label{2points} \mathrm{P} (\tau_{\sqrt{w}}\, <\, \infty)\, \le\, Kw + p(1/\sqrt{w}). \end{equation} Combining (\ref{1point}) and (\ref{2points}), one gets that for any $w > 0$ \begin{equation}\label{3points} \mathrm{P} \left(\sum \gamma_t = \infty\ |\ x_0 \in {\cal O}' \text{ and } \gamma_0 < w\right)\, \le\, Kw\, +\, p(1/\sqrt{w}). \end{equation} Define the event $\mathcal{A}_w = \{ \text{ for some } t,\ x_t \in {\cal O}' \text{ and } \gamma_t < w \}$, then by virtue of (\ref{3points}), \begin{equation}\label{101} \mathrm{P} \left(\sum \gamma_t = \infty\ \big|\ \mathcal{A}_w \right) \le Kw + p(1/\sqrt{w}). \end{equation} Denote by $\bar\mathcal{A}_w$ the complementary event, $\bar\mathcal{A}_w = \{ \text{ for any } t,\ x_t \not\in {\cal O}' \text{ or } \gamma_t \ge w \}$. By virtue of Lemma 7, \begin{equation}\label{102} \mathrm{P} \left(\sum \gamma_t = \infty \ \, \& \, \ \bar\mathcal{A}_w \right) = 0. \end{equation} Using (\ref{101}) and (\ref{102}), one gets \begin{eqnarray*} \mathrm{P} \left( \sum \gamma_t = \infty \right) = \mathrm{P} \left( \sum \gamma_t = \infty \, \ \& \ \, \mathcal{A}_w \right) + \mathrm{P} \left( \sum \gamma_t = \infty \ \, \& \ \, \bar\mathcal{A}_w \right) \le \end{eqnarray*} $$ \le (Kw + p(1/\sqrt{w})) \cdot \mathrm{P} (\mathcal{A}_w). $$ Taking into account that $w$ can be chosen arbitrarily small and that $Kw + p(1/\sqrt{w}) \to 0$ as $w \to 0^+$, one concludes that $\mathrm{P} \left( \sum_t \gamma_t = \infty \right) = 0$. \hfill$\Box$ \vspace{2mm} Now, we are in a position to prove the theorem. Suppose that $\mathrm{k} < \inf_z k_-(z)$, then $V_-^{[\mathrm{k}]} = \emptyset$, and by Lemma 2, $\{ x_t \}$ diverges. So, the statement (b) of Theorem is proved. On the other hand, according to Lemma 10, if $\mathrm{k} > k_+(0)$ then $\sum_t \gamma_t < \infty$, and by Lemmas 1 and 2, the sequence $\{ x_t \}$ converges to a point from $V_-^{[\mathrm{k}]}$. Thus, the statement (a) of theorem is also established. \section*{Acknowledgements} This work was partially supported by the R\&D Unit CEOC (Center for Research in Optimization and Control). The second author (PC) also gratefully acknowledges the financial support by the Portuguese program PRODEP `Medida 5 - Ac\c c\~ao 5.3 - Forma\c c\~ao Avan\c cada de Docentes do Ensino Superior - Concurso nr. 2/5.3/PRODEP/2001'.
{ "timestamp": "2005-03-21T19:44:42", "yymm": "0503", "arxiv_id": "math/0503434", "language": "en", "url": "https://arxiv.org/abs/math/0503434" }
\section{Introduction} \subsection{} This work arose from an attempt to understand the results of the paper ~\cite{gl} of A.~Givental and Y.-P.~Lee where the authors perform some computations related to ``quantum $K$-theory" of flag varieties (as well as some results from ~\cite{neok} related to 5d $SU(n)$-gauge theory compactified on a circle) in the framework of representation theory. Similar approach to quantum cohomology of flag varieties (and to partition functions of 4d gauge theory) is discussed in ~\cite{b} and ~\cite{be}. In \cite{gl} the authors consider the moduli spaces $\fQ_{\ul{d}}$ introduced by G.~Laumon in ~\cite{la1}, ~\cite{la2}. These are certain closures of the moduli spaces of based maps of degree $\ul{d}$ from $\BP^1$ to the flag variety $\CB$ of $\mathfrak{sl}_n$. A Cartan torus $T$ of $SL_n$ acts on $\fQ_{\ul{d}}$. The multiplicative group $\BC^*$ of dilations of $\BP^1$ (loop rotations) also acts on $\fQ_{\ul{d}}$. The formal character of the (infinite dimensional) $T\times\BC^*$-module $R\Gamma(\fQ_{\ul{d}},\CO_{\ul{d}})$ turns out to be a rational function on $T\times\BC^*$. One may form a certain generating function $\fJ$ of these rational functions for all degrees $\ul{d}$. Computing the function $\fJ$ presumably should give rise to a computation of the $SL_n$-equivariant quantum $K$-theory ring of $\CB$ (which to the best of the authors' knowledge has not yet been defined in the literature). A.~Givental and Y.-P.~Lee prove that $\fJ$ satisfies a certain $v$-difference version of the quantum Toda lattice equations (here $v$ stands for the tautological character of $\BC^*$). Moreover, they suggest another way to construct solutions of the $v$-difference Toda system: as the Shapovalov scalar product of the Whittaker vectors in the universal Verma module for the quantum group $U_v(\mathfrak{sl}_n)$. The latter construction was worked out independently in ~\cite{e}, ~\cite{s2}. \subsection{} The principal goal of the present paper is to identify these two constructions of solutions of the $v$-difference Toda system. Namely, we prove that the natural correspondences between the moduli spaces $\fQ_{\ul{d}}$ (for the degrees differing by a simple root) give rise to the action of the standard generators of $U_v(\mathfrak{sl}_n)$ on the localized equivariant $K$-theory $\oplus_{\ul{d}}\ul{K}^{T\times\BC^*} (\fQ_{\ul{d}})$. Here the localization is taken with respect to the $K^{T\times\BC^*}(\cdot)=\BC[T\times\BC^*]$, that is, we tensor everything with the fraction field of $\BC[T\times\BC^*]$. This is needed since the above correspondences are not proper, but the subspaces of their $T\times\BC^*$-fixed points are proper (in fact, they are finite), so their action is well defined only in the localized equivariant $K$-theory. This way we get a $U_v(\mathfrak{sl}_n)$-module, and we identify it with the universal Verma module $M$. We also compute in geometric terms the Shapovalov scalar product on $M$, and the Whittaker vectors. It turns out that the generating function for the Shapovalov scalar product of the Whittaker vectors is a simple modification of the Givental-Lee generating function $\fJ$. Thus we reprove the Main Theorem of Givental-Lee. \subsection{} There is a similar generating function $J$ for equivariant integrals of the unit cohomology classes of $\fQ_{\ul{d}}$ which controls the $T$-equivariant quantum cohomology of $\CB$. It satisfies the quantum Toda lattice differential system, as proved originally by A.~Givental and B.~Kim. For the simple Lie algebras $\fg$ other than $\mathfrak{sl}_n$ there is no analogue of the Laumon moduli spaces $\fQ_{\ul{d}}$ but there is Drinfeld's moduli space of Quasimaps $\CZ_{\ul{d}}(\fg)$. It also exists for the case of affine Lie algebras, under the name of Uhlenbeck compactification. In the $\mathfrak{sl}_n$ case $\fQ_{\ul{d}}$ is a small resolution of $\CZ_{\ul{d}}(\mathfrak{sl}_n)$. In the affine $\widehat{\mathfrak{sl}}_n$ case $\CZ_{\ul{d}}(\widehat{\mathfrak{sl}}_n)$ possesses a semismall resolution of singularities: the moduli space $\CP_{\ul{d}}$ of torsion free parabolic sheaves on $\BP^1\times\BP^1$ endowed with some additional structures. Thus in the affine case we can define an analog of the function $J$ which we denote by $J_{\aff}$ (this is discussed in \cite{b}). The generating function $J$ (for any simple $G$) is known to satisfy the quantum (differential) Toda equations (cf. \cite{gk} and \cite{kim}). In the work ~\cite{b}, the generating function $J$ for equivariant integrals of the unit cohomology classes of $\CZ_{\ul{d}}(\fg)$ was proved to satisfy the quantum Toda lattice by constructing the action of the Langlands dual Lie algebra $\check\fg$ in the equivariant Intersection Cohomology of the Drinfeld compactifications. Also in the affine case the function $J_{\aff}$ was shown to satisfy some non-stationary analog of ``the most basic" (quadratic) Toda equation. Thus ~\cite{b} offered a representation theoretic explanation of the Givental-Kim results as well as generalized them to the affine case. And the present work is a multiplicative analogue of ~\cite{b} in the simplest case of $\mathfrak{sl}_n$. \subsection{} It would be extremely interesting to extend our work to other simple and affine Lie algebras. It would require something like an equivariant ``IC $K$-theory'' of $\CZ_{\ul{d}}(\fg)$ which is not defined at the moment. In case of $\mathfrak{sl}_n$ the IC cohomology of $\CZ_{\ul{d}}(\mathfrak{sl}_n)$ coincides with the cohomology of the small resolution $\fQ_{\ul{d}}$, while in the affine case of $\widehat{\mathfrak{sl}}_n$ the IC cohomology of $\CZ_{\ul{d}}(\widehat{\mathfrak{sl}}_n)$ is a direct summand in the cohomology of the semismall resolution $\CP_{\ul{d}}$. Accordingly, one might look for the correct ``IC $K$-theory'' of $\CZ_{\ul{d}}(\widehat{\mathfrak{sl}}_n)$ as an appropriate direct summand of the usual $K$-theory of $\CP_{\ul{d}}$. This is sketched in the Section ~\ref{p}. Namely, similarly to the case of Laumon spaces, the quantum affine group $U_v(\widehat{\mathfrak{sl}}_n)$ acts by the natural correspondences on the direct sum of localized equivariant $K$-groups $\oplus_{\ul{d}}\ul{K}^{T\times\BC^*\times\BC^*} (\CP_{\ul{d}})$. However, this module looks more like the universal Verma module for $U_v(\widehat{\mathfrak{gl}}_n)$, and we have to specify a certain submodule isomorphic to the universal Verma module for $U_v(\widehat{\mathfrak{sl}}_n)$. Then we construct geometrically the Shapovalov scalar product, and the Whittaker vectors. It turns out that the Shapovalov scalar product of the Whittaker vectors can be expressed via the formal characters of the global sections $R\Gamma(\CP_{\ul{d}},\CO_{\ul{d}})$ as in the case of $\mathfrak{sl}_n$. However, we were unable to derive any $v$-difference equation for the affine version of the generating function $\fJ$. \subsection{Acknowledgments} M.F. is obliged to V.~Schechtman, A.~Stoyanovsky, B.~Feigin, E.~Vasserot, and R.~Bezrukavnikov who, ever since the appearance of ~\cite{fk}, urged him to consider its equivariant $K$-theory analogue. While trying to guess the correct formulae in the low ranks, we profited strongly from the computational help of V.~Dotsenko, V.~Golyshev, A.~Kuznetsov. We are also grateful to P.~Etingof and A.~Joseph for very useful explanations; to M.~Kashiwara for bringing the reference ~\cite{nz} to our attention, and to the referee for the valuable comments. We would like to thank the Weizmann Institute and RIMS, Kyoto, as well as the University of Chicago, for the hospitality and support. M.F. was partially supported by the CRDF award RM1-2545-MO-03. A.B. was partially supported by the NSF grant DMS-0300271. \section{Laumon spaces and quantum groups} \subsection{} We recall the setup of ~\cite{fk}. Let $\bC$ be a smooth projective curve of genus zero. We fix a coordinate $z$ on $\bC$, and consider the action of $\BC^*$ on $\bC$ such that $v(z)=v^{-2}z$. We have $\bC^{\BC^*}=\{0,\infty\}$. We consider an $n$-dimensional vector space $W$ with a basis $w_1,\ldots,w_n$. This defines a Cartan torus $T\subset G=SL_n\subset Aut(W)$. We also consider its $2^{n-1}$-fold cover, the bigger torus $\widetilde{T}$, acting on $W$ as follows: for $\widetilde{T}\ni\ul{t}=(t_1,\ldots,t_n)$ we have $\ul{t}(w_i)=t_i^2w_i$. We denote by $\CB$ the flag variety of $G$. \subsection{} Given an $(n-1)$-tuple of nonnegative integers $\ul{d}=(d_1,\ldots,d_{n-1})$, we consider the Laumon's quasiflags' space $\CQ_{\ul{d}}$, see ~\cite{la2}, ~4.2. It is the moduli space of flags of locally free subsheaves $$0\subset\CW_1\subset\ldots\subset\CW_{n-1}\subset\CW=W\otimes\CO_\bC$$ such that $\on{rank}(\CW_k)=k$, and $\deg(\CW_k)=-d_k$. It is known to be a smooth projective variety of dimension $2d_1+\ldots+2d_{n-1}+\dim\CB$, see ~\cite{la1}, ~2.10. \subsection{} We consider the following locally closed subvariety $\fQ_{\ul{d}}\subset\CQ_{\ul{d}}$ (quasiflags based at $\infty\in\bC$) formed by the flags $$0\subset\CW_1\subset\ldots\subset\CW_{n-1}\subset\CW=W\otimes\CO_\bC$$ such that $\CW_i\subset\CW$ is a vector subbundle in a neighbourhood of $\infty\in\bC$, and the fiber of $\CW_i$ at $\infty$ equals the span $\langle w_1,\ldots,w_i\rangle\subset W$. It is known to be a smooth quasiprojective variety of dimension $2d_1+\ldots+2d_{n-1}$. \subsection{} \label{fixed points} The group $G\times\BC^*$ acts naturally on $\CQ_{\ul{d}}$, and the group $\widetilde{T}\times\BC^*$ acts naturally on $\fQ_{\ul{d}}$. The set of fixed points of $\widetilde{T}\times\BC^*$ on $\fQ_{\ul{d}}$ is finite; we recall its description from ~\cite{fk}, ~2.11. Let $\widetilde{\ul{d}}$ be a collection of nonnegative integers $(d_{ij}),\ i\geq j$, such that $d_i=\sum_{j=1}^id_{ij}$, and for $i\geq k\geq j$ we have $d_{kj}\geq d_{ij}$. Abusing notation we denote by $\widetilde{\ul{d}}$ the corresponding $\widetilde{T}\times\BC^*$-fixed point in $\fQ_{\ul{d}}$: $\CW_1=\CO_\bC(-d_{11}\cdot0)w_1,$ $\CW_2=\CO_\bC(-d_{21}\cdot0)w_1\oplus\CO_\bC(-d_{22}\cdot0)w_2,$ $\ldots\ \ldots\ \ldots\ ,$ $\CW_{n-1}=\CO_\bC(-d_{n-1,1}\cdot0)w_1\oplus\CO_\bC(-d_{n-1,2}\cdot0)w_2 \oplus\ldots\oplus\CO_\bC(-d_{n-1,n-1}\cdot0)w_{n-1}.$ \subsection{} For $i\in\{1,\ldots,n-1\}$, and $\ul{d}=(d_1,\ldots,d_{n-1})$, we set $\ul{d}+i:=(d_1,\ldots,d_i+1,\ldots,d_{n-1})$. We have a correspondence $\CE_{\ul{d},i}\subset\CQ_{\ul{d}}\times \CQ_{\ul{d}+i}$ formed by the pairs $(\CW_\bullet,\CW'_\bullet)$ such that for $j\ne i$ we have $\CW_j=\CW'_j$, and $\CW'_i\subset\CW_i$, see ~\cite{fk}, ~3.1. In other words, $\CE_{\ul{d},i}$ is the moduli space of flags of locally free sheaves $$0\subset\CW_1\subset\ldots\CW_{i-1}\subset\CW'_i\subset\CW_i\subset \CW_{i+1}\ldots\subset\CW_{n-1}\subset\CW$$ such that $\on{rank}(\CW_k)=k$, and $\deg(\CW_k)=-d_k$, while $\on{rank}(\CW'_i)=i$, and $\deg(\CW'_i)=-d_i-1$. According to ~\cite{la1}, ~2.10, $\CE_{\ul{d},i}$ is a smooth projective algebraic variety of dimension $2d_1+\ldots+2d_{n-1}+\dim\CB+1$. We denote by $\bp$ (resp. $\bq$) the natural projection $\CE_{\ul{d},i}\to\CQ_{\ul{d}}$ (resp. $\CE_{\ul{d},i}\to\CQ_{\ul{d}+i}$). We also have a map $\br:\ \CE_{\ul{d},i}\to\bC,$ $$(0\subset\CW_1\subset\ldots\CW_{i-1}\subset\CW'_i\subset\CW_i\subset \CW_{i+1}\ldots\subset\CW_{n-1}\subset\CW)\mapsto\on{supp}(\CW_i/\CW'_i).$$ The correspondence $\CE_{\ul{d},i}$ comes equipped with a natural line bundle $\CL_i$ whose fiber at a point $$(0\subset\CW_1\subset\ldots\CW_{i-1}\subset\CW'_i\subset\CW_i\subset \CW_{i+1}\ldots\subset\CW_{n-1}\subset\CW)$$ equals $\Gamma(\bC,\CW_i/\CW'_i)$. Finally, we have a transposed correspondence $^\sT\CE_{\ul{d},i}\subset \CQ_{\ul{d}+i}\times\CQ_{\ul{d}}$. \subsection{} Restricting to $\fQ_{\ul{d}}\subset\CQ_{\ul{d}}$ we obtain the correspondence $\fE_{\ul{d},i}\subset\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}$ together with line bundle $\fL_i$ and the natural maps $\bp:\ \fE_{\ul{d},i}\to\fQ_{\ul{d}},\ \bq:\ \fE_{\ul{d},i}\to\fQ_{\ul{d}+i},\ \br:\ \fE_{\ul{d},i}\to\bC-\infty$. We also have a transposed correspondence $^\sT\fE_{\ul{d},i}\subset \fQ_{\ul{d}+i}\times\fQ_{\ul{d}}$. It is a smooth quasiprojective variety of dimension $2d_1+\ldots+2d_{n-1}+1$. \subsection{} We denote by ${}'M$ the direct sum of equivariant (complexified) $K$-groups: ${}'M=\oplus_{\ul{d}}K^{\widetilde{T}\times\BC^*}(\fQ_{\ul{d}})$. It is a module over $K^{\widetilde{T}\times\BC^*}(pt)=\BC[\widetilde{T}\times\BC^*]= \BC[t_1,\ldots,t_n,v\ :\ t_1\cdots t_n=1]$. We define $M=\ {}'M\otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))$. We have an evident grading $M=\oplus_{\ul{d}}M_{\ul{d}},\ M_{\ul{d}}=K^{\widetilde{T}\times\BC^*}(\fQ_{\ul{d}}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))$. \subsection{} \label{operators} The grading and the correspondences $^\sT\fE_{\ul{d},i},\fE_{\ul{d},i}$ give rise to the following operators on $M$ (note that though $\bp$ is not proper, $\bp_*$ is well defined on the localized equivariant $K$-theory due to the finiteness of the fixed point sets): $K_i=t_{i+1}t_i^{-1}v^{2d_i-d_{i-1}-d_{i+1}+1}:\ M_{\ul{d}}\to M_{\ul{d}}$; $L_i=t_1^{-1}\cdots t_i^{-1}v^{d_i+\frac{1}{2}i(n-i)}:\ M_{\ul{d}}\to M_{\ul{d}}$; $f_i=\bp_*\bq^*:\ M_{\ul{d}}\to M_{\ul{d}-i}$; $F_i=t_{i+1}^it_i^{-i}v^{2id_i-id_{i-1}-id_{i+1}-i}\bp_*\bq^*:\ M_{\ul{d}}\to M_{\ul{d}-i}$; $e_i=-t_{i+1}^{-1}t_i^{-1}v^{d_{i+1}-d_{i-1}}\bq_*(\fL_i\otimes\bp^*):\ M_{\ul{d}}\to M_{\ul{d}+i}$, $E_i=-t_{i+1}^{-i-1}t_i^{i-1}v^{(i-1)d_{i-1}+(i+1)d_{i+1}-2id_i-i} \bq_*(\fL_i\otimes\bp^*):\ M_{\ul{d}}\to M_{\ul{d}+i}$. \subsection{} We recall the notations and results of ~\cite{s} in the special case of quantum group of $SL_n$ type. $U$ is the $\BC[v,v^{-1}]$-algebra with generators $E_i,L_i^{\pm1},K_i^{\pm1},F_i,\ 1\leq i\leq n-1$, subject to the following relations: \begin{equation} \label{och} L_iL_j=L_jL_i,\ K_1=L_1^2L_2^{-1},\ K_i=L_{i-1}^{-1}L_i^2L_{i+1}^{-1},\ K_{n-1}=L_{n-2}^{-1}L_{n-1}^2 \end{equation} \begin{equation} \label{ochev} L_iE_jL_i^{-1}=v^{\delta_{i,j}}E_j,\ L_iF_jL_i^{-1}=v^{-\delta_{i,j}}F_j \end{equation} \begin{equation} \label{ochevidno} E_iF_j-F_jE_i=\delta_{i,j}\frac{K_i-K_i^{-1}}{v-v^{-1}} \end{equation} \begin{equation} \label{Serre1} |i-j|>1\ \Longrightarrow\ E_iE_j-E_jE_i=0=F_iF_j-F_jF_i \end{equation} \begin{equation} \label{Serre2} |i-j|=1\ \Longrightarrow\ E_i^2E_j-(v+v^{-1})E_iE_jE_i+E_jE_i^2=0= F_i^2F_j-(v+v^{-1})F_iF_jF_i+F_jF_i^2 \end{equation} Sevostyanov considers elements $e_i,f_i\in U$ depending on a choice of $(n-1)\times(n-1)$-matrices $n_{ij},\ c_{ij}$. We make the following choice: \begin{equation} \label{vybor} n_{i,i}=-2i;\ n_{i,i+1}=n_{i,i-1}=i, \end{equation} otherwise $n_{ij}=0$. \begin{equation} \label{Rossii} i<n-1\ \Longrightarrow\ c_{i,i+1}=-1,\ c_{i+1,i}=1, \end{equation} otherwise $c_{ij}=0$. In other words, $c_{ij}=n_{ij}-n_{ji}$. Then we have \begin{equation} \label{Seva} f_i:=L_{i-1}^iL_i^{-2i}L_{i+1}^iF_i=K_i^{-i}F_i,\ e_i:=E_iL_{i-1}^{-i}L_i^{2i}L_{i+1}^{-i}=E_iK_i^i. \end{equation} Clearly, the algebra $U$ is generated by $e_i,L_i^{\pm1},K_i^{\pm1},f_i,\ 1\leq i\leq n-1$, and the relations ~(\ref{ochev})--~(\ref{Serre2}) above are equivalent to the relations ~(\ref{ochev'})--~(\ref{Serre2'}) below. \begin{equation} \label{ochev'} L_ie_jL_i^{-1}=v^{\delta_{i,j}}e_j,\ L_if_jL_i^{-1}=v^{-\delta_{i,j}}f_j \end{equation} \begin{equation} \label{ochevidno'} e_if_j-v^{c_{ij}}f_je_i=\delta_{i,j}\frac{K_i-K_i^{-1}}{v-v^{-1}} \end{equation} \begin{equation} \label{Serre1'} |i-j|>1\ \Longrightarrow\ e_ie_j-e_je_i=0=f_if_j-f_jf_i \end{equation} \begin{equation} \label{Serre2'} |i-j|=1\ \Longrightarrow\ e_i^2e_j-v^{c_{ij}}(v+v^{-1})e_ie_je_i+v^{2c_{ij}}e_je_i^2=0= f_i^2f_j-v^{c_{ij}}(v+v^{-1})f_if_jf_i+v^{2c_{ij}}f_jf_i^2 \end{equation} \subsection{Remark} The elements $f_i$ of the subalgebra $U_{\leq0}$ generated by $F_1,\ldots,F_{n-1},K_1,\ldots,K_{n-1}$ were introduced by C.~M.~Ringel in ~\cite{r}. They are the natural generators of the Hall algebra of the $A_{n-1}$-quiver with the set of vertices $1,\ldots,n-1$, and orientation $i\longrightarrow i+1$. More generally, Ringel's construction works for an arbitrary orientation of an $ADE$ quiver, and produces Sevostyanov's generators $f_i$ (in the simply laced case). It can be seen easily that the set of Sevostyanov's matrices $c_{ij}$ (parametrizing the choices of his ``Coxeter realizations'') is in a natural bijection with the set of orientations of the corresponding quiver. \subsection{} We are finally able to formulate our main theorem. Recall the operators $E_i,e_i,L_i^{\pm1},K_i^{\pm1},F_i,f_i$ on $M$ defined in ~\ref{operators}. \begin{thm} \label{main} The operators $E_i,L_i^{\pm1},K_i^{\pm1},F_i,\ 1\leq i\leq n-1$, on $M$ satisfy the relations ~(\ref{och})--~(\ref{Serre2}). Equivalently, the operators $e_i,L_i^{\pm1},K_i^{\pm1},f_i,\ 1\leq i\leq n-1$, on $M$ satisfy the relations ~(\ref{och}), ~(\ref{ochev'})--~(\ref{Serre2'}). \end{thm} The relations ~(\ref{och}) and ~(\ref{ochev}) are evident. The relation ~(\ref{ochevidno}) for $i\ne j$ follows from a transversality property formulated in the next subsection. \subsection{} We consider the subvarieties $\bp_{12}^{-1}(\fE_{\ul{d},i})$ and $\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ in $\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}\times\fQ_{\ul{d}+i-j}$. Similarly, we consider the subvarieties $\bp_{12}^{-1}(\ ^\sT\fE_{\ul{d}-j,j})$ and $\bp_{23}^{-1}(\fE_{\ul{d}-j,i})$ in $\fQ_{\ul{d}}\times\fQ_{\ul{d}-j}\times\fQ_{\ul{d}+i-j}$. \begin{lem} \label{trans} For $i\ne j$ the intersection (a) $\bp_{12}^{-1}(\fE_{\ul{d},i})\cap\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ in $\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}\times\fQ_{\ul{d}+i-j}$ (resp. (b) $\bp_{12}^{-1}(\ ^\sT\fE_{\ul{d}-j,j})\cap\bp_{23}^{-1}(\fE_{\ul{d}-j,i})$ in $\fQ_{\ul{d}}\times\fQ_{\ul{d}-j}\times\fQ_{\ul{d}+i-j}$) is transversal. (c) $\bp_{12}^{-1}(\fE_{\ul{d},i})\cap\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j}) \simeq \bp_{12}^{-1}(\ ^\sT\fE_{\ul{d}-j,j})\cap\bp_{23}^{-1}(\fE_{\ul{d}-j,i})$. \end{lem} \begin{proof} We prove (a). By definition, $\bp_{12}^{-1}(\fE_{\ul{d},i})$ is the moduli space of pairs of flags $$(0\subset\CW'_1=\CW_1\subset\CW'_2=\CW_2\subset \ldots\subset\CW'_i\subset\CW_i \subset\ldots\subset\CW'_{n-1}=\CW_{n-1}\subset\CW,$$ $$0\subset\CW'''_1\subset\CW'''_2\subset\ldots\subset\CW'''_{n-1}\subset\CW)$$ of prescribed ranks and degrees, while $\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ is the moduli space of pairs of flags $$(0\subset\CW_1\subset\ldots\subset\CW_{n-1}\subset\CW,$$ $$0\subset\CW'_1=\CW'''_1\subset\CW'_2=\CW'''_2\subset \ldots\subset\CW'_j\subset \CW'''_j\subset\ldots\subset\CW'_{n-1}=\CW'''_{n-1}\subset\CW)$$ of prescribed ranks and degrees. Their intersection is the moduli space of flags (say, $i<j$) $$0\subset\CW'_1=\CW_1=\CW'''_1\subset\ldots\subset\CW'_i=\CW'''_i\subset\CW_i \subset\ldots\subset\CW'_j=$$ $$=\CW_j\subset\CW'''_j\subset\ldots\subset \CW'_{n-1}=\CW_{n-1}=\CW'''_{n-1}\subset\CW$$ of prescribed ranks and degrees which is smooth according to ~\cite{la1}, ~2.10. This implies that at any closed point of the scheme-theoretic intersection $\bp_{12}^{-1}(\fE_{\ul{d},i})\cap\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ the Zariski tangent space to $\bp_{12}^{-1}(\fE_{\ul{d},i})\cap\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ is the intersection of tangent spaces to $\bp_{12}^{-1}(\fE_{\ul{d},i})$ and $\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$. Comparing the dimensions we conclude that the sum of tangent spaces to $\bp_{12}^{-1}(\fE_{\ul{d},i})$ and $\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ must coincide with the tangent space to $\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}\times\fQ_{\ul{d}+i-j}$. Hence the intersection is transversal. This completes the proof of (a). In (b) we prove similarly that $\bp_{12}^{-1}(\ ^\sT\fE_{\ul{d}-j,j})\cap\bp_{23}^{-1}(\fE_{\ul{d}-j,i})$ is the moduli space of flags (say, $i<j$) $$0\subset\CW_1=\CW'''_1=\CW''_1\subset\ldots\subset\CW'''_i\subset\CW_i =\CW''_i \subset\ldots\subset\CW_j\subset$$ $$\subset\CW'''_j=\CW''_j\subset \ldots\subset \CW_{n-1}=\CW'''_{n-1}=\CW''_{n-1}\subset\CW$$ of prescribed ranks and degrees which is smooth according to ~\cite{la1}, ~2.10. Hence the intersection is transversal by the same argument as in the proof of (a). This completes the proof of (b). Part (c) was proved in ~\cite{fk}, ~3.6. We just recall that the mutually inverse isomorphisms send a triple $(\CW_\bullet,\CW'_\bullet,\CW'''_\bullet)$ to $(\CW_\bullet,\CW''_\bullet,\CW'''_\bullet)$ where $\CW''_\bullet:=\CW_\bullet+\CW'''_\bullet$, and a triple $(\CW_\bullet,\CW''_\bullet,\CW'''_\bullet)$ to $(\CW_\bullet,\CW'_\bullet,\CW'''_\bullet)$ where $\CW'_\bullet:=\CW_\bullet\cap\CW'''_\bullet$. \end{proof} \subsection{} We return to the proof of relation ~(\ref{ochevidno}) for $i\ne j$. The composition $E_iF_j$ is given by the action of correspondence $$f(\ul{t})g(v)\bp_{13*}(\bp_{12}^*\fL_i \stackrel{L}{\otimes}_{\CO_{\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}\times \fQ_{\ul{d}+i-j}}}\bp_{23}^*\CO_{^\sT\fE_{\ul{d}+i-j,j}})$$ where $f$ (resp. $g$) is a certain monomial in $\ul{t}$ (resp. $v$). Because of the transversality in ~\ref{trans}(a), $\bp_{12}^*\fL_i \stackrel{L}{\otimes}_{\CO_{\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}\times \fQ_{\ul{d}+i-j}}}\bp_{23}^*\CO_{^\sT\fE_{\ul{d}+i-j,j}}$ is a line bundle $\fL_{i,j}$ on $\bp_{12}^{-1}(\fE_{\ul{d},i})\cap\bp_{23}^{-1}(\ ^\sT\fE_{\ul{d}+i-j,j})$ whose fiber at a point $(\CW_\bullet,\CW'_\bullet,\CW'''_\bullet)$ is equal to $\Gamma(\bC,\CW_i/\CW'''_i)$. Similarly, due to the transversality in ~\ref{trans}(b), the composition $F_jE_i$ is given by the action of correspondence $$f'(\ul{t})g'(v)\bp_{13*}(\fL'_{i,j})$$ where $f'$ (resp. $g'$) is a certain monomial in $\ul{t}$ (resp. $v$), and $\fL'_{i,j}$ is a line bundle on $\bp_{12}^{-1}(\ ^\sT\fE_{\ul{d}-j,j})\cap\bp_{23}^{-1}(\fE_{\ul{d}-j,i})$ whose fiber at a point $(\CW_\bullet,\CW''_\bullet,\CW'''_\bullet)$ is equal to $\Gamma(\bC,\CW_i/\CW'''_i)$. Now the isomorphism in ~\ref{trans}(c) clearly takes $\fL_{i,j}$ to $\fL'_{i,j}$, and a routine check shows that $f(\ul{t})g(v)=f'(\ul{t})g'(v)$. This completes the proof of the relations ~(\ref{ochevidno}) for $i\ne j$. \subsection{} \label{matrix} To prove the relation ~(\ref{ochevidno}) for $i=j$ we use the localization to the fixed points. According to the Thomason localization theorem (see e.g. ~\cite{cg}), restriction to the $\widetilde{T}\times\BC^*$-fixed point set induces an isomorphism $$K^{\widetilde{T}\times\BC^*}(\fQ_{\ul{d}}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))\to K^{\widetilde{T}\times\BC^*}(\fQ_{\ul{d}}^{\widetilde{T}\times\BC^*}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))$$ (resp. $$K^{\widetilde{T}\times\BC^*}(\fE_{\ul{d},i}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))\to K^{\widetilde{T}\times\BC^*}(\fE_{\ul{d},i}^{\widetilde{T}\times\BC^*}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt)))$$ The classes of the structure sheaves $[\widetilde{\ul{d}}]$ of the $\widetilde{T}\times\BC^*$-fixed points $\widetilde{\ul{d}}$ (see ~\ref{fixed points}) form a basis in $\oplus_{\ul{d}}K^{\widetilde{T}\times\BC^*} (\fQ_{\ul{d}}^{\widetilde{T}\times\BC^*}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)}\on{Frac} (K^{\widetilde{T}\times\BC^*}(pt))$. In order to compute the matrix coefficients of $E_i,F_i$ in this basis, we have to know the character of the $\widetilde{T}\times\BC^*$-action in the tangent spaces $\CT_{\widetilde{\ul{d}}}\fQ_{\ul{d}}$ and also in the tangent spaces to the fixed points in the correspondences. This is the subject of the following Proposition. \subsection{} Note that a point $(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')$ lies in the correspondence $\fE_{\ul{d},i}$ if and only if $d_{k,j}=d'_{k,j}$ with a single exception $d'_{i,j}=d_{i,j}+1$ for certain $j\leq i$. \begin{prop} \label{zanudstvo} a) The character $\chi_{\widetilde{\ul{d}}}$ of $\widetilde{T}\times\BC^*$ in the tangent space $\CT_{\widetilde{\ul{d}}}\fQ_{\ul{d}}$ equals $$\sum_{1\leq k,j\leq n-1}t_k^2t_j^{-2}\left(\sum_{l=0}^{d_{k-1,j}}v^{2l}- \sum_{l=d_{k-1,j}-d_{k,k}+1}^{d_{k-1,j}}v^{2l}+ \sum_{n-1\geq i\geq k,j}\sum_{l=d_{i,j}-d_{i,k}+1}^{d_{i,j}-d_{i+1,k}}v^{2l} \right)-\sum_{1\leq j<k\leq n-1}t_k^2t_j^{-2}$$ where we set $d_{n,k}=0$. b) The character $\chi_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}$ of $\widetilde{T}\times\BC^*$ in the tangent space $\CT_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}\fE_{\ul{d},i}$ equals $$\chi_{\widetilde{\ul{d}}}+ \sum_{k\leq i}t_j^2t_k^{-2}v^{2d'_{i,k}-2d_{i,j}}- \sum_{k\leq i-1}t_j^2t_k^{-2}v^{2d_{i-1,k}-2d_{i,j}}$$ if $d'_{i,j}=d_{i,j}+1$ for certain $j\leq i$. c) The character $\lambda_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}$ of $\widetilde{T}\times\BC^*$ in the fiber of $\fL_i$ at the point $(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')$ equals $t_j^2v^{-2d_{i,j}}$ if $d'_{i,j}=d_{i,j}+1$. \end{prop} \begin{proof} Let $\CQ$ be the moduli space of flags of locally free subsheaves $$0\subset\CW_1\subset\CW_2\subset\ldots\subset\CW_r\subset\CW$$ of fixed ranks. Then the tangent space $\CT_{\CW_\bullet}\CQ$ equals the kernel of $$\sum_{1\leq l<r}p_{l-1}^*\otimes\on{Id}-\on{Id}\otimes q_l:\ \oplus_l\on{Hom}(\CW_l,\CW/\CW_l)\twoheadrightarrow \oplus_l\on{Hom}(\CW_l,\CW/\CW_{l+1})$$ where $p_l:\ \CW_l\hookrightarrow\CW_{l+1};\ q_l:\ \CW/\CW_l\twoheadrightarrow\CW/\CW_{l+1}$ (see e.g. ~\cite{gl}, ~3.2). The parts a), b) follow easily. The part c) is obvious. \end{proof} \subsection{} Let us denote by $S\chi_{\widetilde{\ul{d}}}=\Lambda^{-1}\chi_{\widetilde{\ul{d}}}$ (resp. $S\chi_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}= \Lambda^{-1}\chi_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}$) the character of $\widetilde{T}\times\BC^*$ in the symmetric algebra $\on{Sym}^\bullet\CT_{\widetilde{\ul{d}}}\fQ_{\ul{d}}$ (resp. $\on{Sym}^\bullet\CT_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')} \fE_{\ul{d},i}$). It is the inverse of the character of the corresponding exterior algebra, thus it lies in the fraction field $\on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))$. According to the Bott-Lefschetz fixed point formula, the matrix coefficient $\bp_*\bq^*_{[\widetilde{\ul{d}}{}',\widetilde{\ul{d}}]}$ of $\bp_*\bq^*:\ M_{\ul{d}'}\to M_{\ul{d}}$ with respect to the basis elements $[\widetilde{\ul{d}}]\in K^{\widetilde{T}\times\BC^*} (\fQ_{\ul{d}}),\ [\widetilde{\ul{d}}{}']\in K^{\widetilde{T}\times\BC^*} (\fQ_{\ul{d}'})$ (see ~\ref{matrix}) equals $S\chi_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}/ S\chi_{\widetilde{\ul{d}}{}'}$. Similarly, the matrix coefficient $\bq_*(\fL_i\otimes\bp^*)_{[\widetilde{\ul{d}},\widetilde{\ul{d}}{}']}$ of $\bq_*(\fL_i\otimes\bp^*):\ M_{\ul{d}}\to M_{\ul{d}'}$ equals $\lambda_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')} S\chi_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}/ S\chi_{\widetilde{\ul{d}}}$. Hence, the matrix coefficient $E_{i[\widetilde{\ul{d}},\widetilde{\ul{d}}{}']}$ of $E_i:\ M_{\ul{d}}\to M_{\ul{d}'}$ equals $$-t_{i+1}^{-i-1}t_i^{i-1}v^{(i-1)d_{i-1}+(i+1)d_{i+1}-2id_i-i} \lambda_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')} S\chi_{(\widetilde{\ul{d}},\widetilde{\ul{d}}{}')}/ S\chi_{\widetilde{\ul{d}}}.$$ And the matrix coefficient $F_{i[\widetilde{\ul{d}},\widetilde{\ul{d}}{}']}$ of $F_i:\ M_{\ul{d}}\to M_{\ul{d}'}$ equals $t_{i+1}^it_i^{-i}v^{2id_i-id_{i-1}-id_{i+1}-i} S\chi_{(\widetilde{\ul{d}}{}',\widetilde{\ul{d}})}/ S\chi_{\widetilde{\ul{d}}}$. Thus, Proposition ~\ref{zanudstvo} admits the following Corollary. \begin{cor} \label{coefficients} $$E_{i[\widetilde{\ul{d}},\widetilde{\ul{d}}{}']}= -t_{i+1}^{-i-1}t_i^{i-1}v^{(i-1)d_{i-1}+(i+1)d_{i+1}-2id_i-i} t_j^2v^{-2d_{i,j}}\times$$ $$(1-v^2)^{-1}\prod_{j\ne k\leq i}(1-t_j^2t_k^{-2}v^{2d_{i,k}-2d_{i,j}})^{-1} \prod_{k\leq i-1}(1-t_j^2t_k^{-2}v^{2d_{i-1,k}-2d_{i,j}})$$ if $d'_{i,j}=d_{i,j}+1$ for certain $j\leq i$; $$F_{i[\widetilde{\ul{d}},\widetilde{\ul{d}}{}']}= t_{i+1}^it_i^{-i}v^{2id_i-id_{i-1}-id_{i+1}-i}\times$$ $$(1-v^2)^{-1}\prod_{j\ne k\leq i}(1-t_k^2t_j^{-2}v^{2d_{i,j}-2d_{i,k}})^{-1} \prod_{k\leq i+1}(1-t_k^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,k}})$$ if $d'_{i,j}=d_{i,j}-1$ for certain $j\leq i$; All the other matrix coefficients of $E_i,F_i$ vanish. \end{cor} Now the relation ~(\ref{ochevidno}) boils down to the following identity. \begin{prop} \label{mrak} $$\frac{t_it_{i+1}^{-1}v^{d_{i-1}-2d_i+d_{i+1}-1}- t_i^{-1}t_{i+1}v^{-d_{i-1}+2d_i-d_{i+1}+1}}{v-v^{-1}} (1-v^2)^2v^{d_{i-1}-d_{i+1}}t_it_{i+1}=$$ $$\sum_{j\leq i}t_j^2v^{-2d_{i,j}+2} (1-t_i^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,i}}) (1-t_{i+1}^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,i+1}})\times$$ $$\times\prod_{k\leq i}^{k\ne j}(1-t_k^2t_j^{-2}v^{2d_{i,j}-2d_{i,k}})^{-1} (1-t_j^2t_k^{-2}v^{2d_{i,k}-2d_{i,j}+2})^{-1}\times$$ $$\times\prod_{k\leq i-1}(1-t_k^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,k}}) (1-t_j^2t_k^{-2}v^{2d_{i-1,k}-2d_{i,j}+2})-$$ $$-\sum_{j\leq i}t_j^2v^{-2d_{i,j}} (1-t_i^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,i}+2}) (1-t_{i+1}^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,i+1}+2})\times$$ $$\times\prod_{k\leq i}^{k\ne j}(1-t_k^2t_j^{-2}v^{2d_{i,j}-2d_{i,k}+2})^{-1} (1-t_j^2t_k^{-2}v^{2d_{i,k}-2d_{i,j}})^{-1}\times$$ $$\times\prod_{k\leq i-1}(1-t_k^2t_j^{-2}v^{2d_{i,j}-2d_{i+1,k}+2}) (1-t_j^2t_k^{-2}v^{2d_{i-1,k}-2d_{i,j}}).$$ \end{prop} \begin{proof} We introduce the new variables $q:=v^2;\ s_j:=t_j^2v^{-2d_{ij}},\ 1\leq j\leq i;\ r_k:=t_k^2v^{-2d_{i+1,k}},\ 1\leq k\leq i+1;\ p_k:=t_k^2v^{-2d_{i-1,k}},\ 1\leq k\leq i-1$. Then the LHS of ~\ref{mrak} equals $$(1-q)\left(q\prod_{k=1}^{i+1}r_k\prod_{j=1}^is_j^{-1}- \prod_{j=1}^is_j\prod_{k=1}^{i-1}p_k^{-1}\right)$$ while the RHS of ~\ref{mrak} equals $$\prod_{j=1}^is_j\prod_{k=1}^{i-1}p_k^{-1} \left(q\sum_{j\leq i}s_j^{-2}\prod_{k=1}^{i+1}(s_j-r_k) \prod_{k=1}^{i-1}(p_k-qs_j) \prod_{k\leq i}^{k\ne j}(s_j-s_k)^{-1}(s_k-qs_j)^{-1}\right.-$$ $$\left.\sum_{j\leq i}s_j^{-2}\prod_{k=1}^{i+1}(s_j-qr_k) \prod_{k=1}^{i-1}(p_k-s_j) \prod_{k\leq i}^{k\ne j}(s_j-qs_k)^{-1}(s_k-s_j)^{-1}\right)$$ Dividing both the LHS and the RHS by $\prod_{j=1}^is_j\prod_{k=1}^{i-1}p_k^{-1}$ we arrive at $$(1-q)(q\prod_{j=1}^is_j^{-2}\prod_{k=1}^{i-1}p_k\prod_{k=1}^{i+1}r_k-1)=$$ $$q\sum_{j\leq i}s_j^{-2}\prod_{k=1}^{i+1}(s_j-r_k) \prod_{k=1}^{i-1}(p_k-qs_j) \prod_{k\leq i}^{k\ne j}(s_j-s_k)^{-1}(s_k-qs_j)^{-1}-$$ $$\sum_{j\leq i}s_j^{-2}\prod_{k=1}^{i+1}(s_j-qr_k) \prod_{k=1}^{i-1}(p_k-s_j) \prod_{k\leq i}^{k\ne j}(s_j-qs_k)^{-1}(s_k-s_j)^{-1}.$$ If we subtract the LHS from the RHS we obtain a rational expression in $s_j$ of degree 0, that is, the degree of numerator is not bigger than the degree of denominator. We see easily that as $s_j$ tends to $\infty$, the difference of the RHS and the LHS tends to 0. The possible poles of the difference can occur at $s_j=0,\ s_j=s_k,\ s_j=qs_k,\ s_j=q^{-1}s_k$. We see easily that the principal parts of the difference at these points vanish. We conclude that the difference is identically 0. This completes the proof of the Proposition. \end{proof} \subsection{} \label{import} To finish the proof of relation ~(\ref{ochevidno}) we note that the commutator correspondence $E_iF_i-F_iE_i$ is concentrated on the diagonal of $\fQ_{\ul{d}}\times\fQ_{\ul{d}}$. This is proved exactly as in Lemma ~\ref{trans}. In other words, $E_iF_i-F_iE_i$ is given by tensor product $?\mapsto?\stackrel{L}{\otimes}X_i$ for certain $X_i\in M_{\ul{d}}$. This means that in the basis $[\widetilde{\ul{d}}]$ the operator $E_iF_i-F_iE_i$ is diagonal. Now the Proposition ~\ref{mrak} computes the matrix coefficient $(E_iF_i-F_iE_i)_{[\widetilde{\ul{d}},\widetilde{\ul{d}}]}$ and proves that it equals $\frac{K_i-K_i^{-1}}{v-v^{-1}}|_{M_{\ul{d}}}$. This completes the proof of the relation ~(\ref{ochevidno}). \subsection{} \label{alternate} Alternatively, the relation ~(\ref{ochevidno}) follows from the next Conjecture. We consider a 2-dimensional vector space with a basis $\fw_1,\fw_2$. Let $\fT$ be a torus acting on $\fw_1$ (resp. $\fw_2$) via a character $\tau_1^2$ (resp. $\tau_2^2$). Let $\fZ_{\fd_1,\fd_2}$ be the moduli stack of flags of coherent sheaves $\fW_1\subset\fW_2$ on $\bC$ locally free at $\infty\in\bC$, equipped with a trivialization $\fW_1|_{\infty}=\langle\fw_1\rangle,\ \fW_2|_{\infty}=\langle\fw_1,\fw_2\rangle$, and such that $\deg\fW_1=-\fd_1,\ \deg\fW_2/\fW_1=-\fd_2$. We have a natural correspondence $\fE_{\fd_1}\subset\fZ_{\fd_1,\fd_2}\times\fZ_{\fd_1+1,\fd_2-1}$ formed by the pairs $(\fW_1,\fW_2;\fW'_1,\fW'_2)$ such that $\fW'_1\subset\fW_1\subset\fW_2=\fW'_2$. The projection $\fE_{\fd_1}\to\fZ_{\fd_1,\fd_2}$ (resp. $\fE_{\fd_1}\to\fZ_{\fd'_1,\fd_2}$) is denoted by $\bp$ (resp. $\bq$). Finally, $\fE_{\fd_1}$ is equipped with the line bundle $\fL_{\fd_1}$ whose fiber at the point $(\fW_1,\fW_2;\fW'_1,\fW'_2)$ equals $\Gamma(\bC,\fW_1/\fW'_1)$. The stack $\fZ_{\fd_1,\fd_2}$ is smooth, and acted upon by $\fT\times\BC^*$. So it makes sense to consider the operators $$f:=\bp_*\bq^*:$$ $$K^{\fT\times\BC^*}(\fZ_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))\to K^{\fT\times\BC^*}(\fZ_{\fd_1-1,\fd_2+1}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt)),$$ $$e:=-\tau_1^{-1}\tau_2^{-1}v^{\fd_1+\fd_2} \bq_*(\fL_{\fd_1}\otimes\bp^*):$$ $$K^{\fT\times\BC^*}(\fZ_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))\to K^{\fT\times\BC^*}(\fZ_{\fd_1+1,\fd_2-1}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt)),$$ $$K=\tau_1^{-1}\tau_2v^{\fd_1-\fd_2+1}:$$ $$K^{\fT\times\BC^*}(\fZ_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))\to K^{\fT\times\BC^*}(\fZ_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))$$ \begin{conj} \label{inteligent} $ef-fe=\frac{K-K^{-1}}{v-v^{-1}}$. \end{conj} \subsection{} \label{derivation} To derive the relation ~(\ref{ochevidno}), or equivalently, ~(\ref{ochevidno'}) for $j=i$ from Conjecture ~\ref{inteligent} we consider the map $$\fz_{\ul{d}}:\ \fQ_{\ul{d}}\to\fZ_{d_i-d_{i-1},d_{i+1}-d_i},\ \CW_\bullet\mapsto(\CW_i/\CW_{i-1},\CW_{i+1}/\CW_{i-1}).$$ Then we have $$(\fQ_{\ul{d}}\times\fZ_{d_i-d_{i-1}+1,d_{i+1}-d_i-1}) \times_{\fZ_{d_i-d_{i-1},d_{i+1}-d_i}\times\fZ_{d_i-d_{i-1}+1,d_{i+1}-d_i-1}} \fE_{d_i-d_{i-1}}=\fE_{\ul{d},i}\subset\fQ_{\ul{d}}\times\fQ_{\ul{d}+i}.$$ We also have the natural maps $$\fe_{\ul{d},i}:\ \fE_{\ul{d},i}\to\fE_{d_i-d_{i-1}},$$ $$^\sT\fe_{\ul{d},i}:\ {}^\sT\fE_{\ul{d},i}\to\ {}^\sT\fE_{d_i-d_{i-1}},$$ $$\fh_{\ul{d},i}:\ \fE_{\ul{d}-i,i}\circ\ {}^\sT\fE_{\ul{d}-i,i}\to \fE_{d_i-d_{i-1}-1}\circ\ {}^\sT\fE_{d_i-d_{i-1}-1},$$ $$'\fh_{\ul{d},i}:\ {}^\sT\fE_{\ul{d},i}\circ\fE_{\ul{d},i}\to\ {}^\sT\fE_{d_i-d_{i-1}}\circ\fE_{d_i-d_{i-1}}.$$ We may consider $e_i$ (resp. $f_i,e,f$) as an element of $K^{\widetilde{T}\times\BC^*}(\fE_{\ul{d},i})$ (resp. $K^{\widetilde{T}\times\BC^*}(\ {}^\sT\fE_{\ul{d},i}),\\ K^{\widetilde{T}\times\BC^*}(\fE_{d_i-d_{i-1}}),\ K^{\widetilde{T}\times\BC^*}(\ {}^\sT\fE_{d_i-d_{i-1}})$). We evidently have $$\fe_{\ul{d},i}^*e=e_i,\ {}^\sT\fe_{\ul{d},i}^*f=f_i.$$ Moreover, according to ~\cite{n}, ~8.2 (Restriction of the convolution to submanifolds), we have \begin{equation} \label{nak} \fh_{\ul{d},i}^*(e*f)=e_i*f_i,\ '\fh_{\ul{d},i}^*(f*e)=f_i*e_i. \end{equation} We already know from the argument in ~\ref{import} that the correspondence $e_i*f_i-f_i*e_i$ acts as tensor multiplication with a certain class $X_i\in M_{\ul{d}}$. Similarly, the correspondence $e*f-f*e$ acts in $K^{\widetilde{T}\times\BC^*}(\fZ_{d_i-d_{i-1},d_{i+1}-d_i}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))$ as tensor multiplication with a certain class $\fX\in K^{\widetilde{T}\times\BC^*}(\fZ_{d_i-d_{i-1},d_{i+1}-d_i}) \otimes_{K^{\widetilde{T}\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*}(pt))$ By ~(\ref{nak}) we must have $X_i=\fz_{\ul{d}}^*\fX$. Thus the relation ~(\ref{ochevidno'}) for $j=i$ follows from Conjecture ~\ref{inteligent}. \subsection{} \label{univerma} To complete the proof of Theorem ~\ref{main} it remains to check the relations ~(\ref{Serre1}), ~(\ref{Serre2}). To this end we consider the algebra $\widetilde{U}$ given by the generators $E_i,L_i^{\pm1},K_i^{\pm1},F_i,\ 1\leq i\leq n-1$, and the relations ~(\ref{och})--~(\ref{ochevidno}). Thus, $U$ is the quotient of $\widetilde{U}$ by the Serre relations. We extend the scalars to $\on{Frac}(\BC[L_1^{\pm1},\ldots,L_{n-1}^{\pm1}])$: we set $$U'=U\otimes_\BC \on{Frac}(\BC[L_1^{\pm1},\ldots,L_{n-1}^{\pm1}]),\ \widetilde{U}{}'=\widetilde{U}\otimes_\BC \on{Frac}(\BC[L_1^{\pm1},\ldots,L_{n-1}^{\pm1}])$$ Note that $\widetilde{U}{}'$ acts in $M$, so $U'$ acts in the quotient $\overline{M}$ of $M$ by the two-sided ideal $\CI$ in $\widetilde{U}{}'$ generated by the Serre relations. So it suffices to check that $\overline{M}=M$, or equivalently, $\CI M=0$. Now $M$ has the size of the universal Verma module over $U'$ which is an irreducible $U'$- (and $\widetilde{U}{}'$-) module. In effect, a bijection between the set $\{[\widetilde{\ul{d}}]\}$, and the set of Kostant partitions for $\mathfrak{sl}_n$ is defined e.g. in ~\cite{fk}, 2.11. Hence we only have to check that $\CI M\ne M$. But any element $x\in\CI$ of principal grading degree 0 annihilates the lowest weight vector $[(0,\ldots,0)]$ of $M$ since we may shift the generators $e_i$ in the expression of $x$ to the right. This completes the proof of the Serre relations in $M$ along with the proof of Theorem ~\ref{main}. \subsection{Remark} \label{joseph} (A.~Joseph) We have constructed a basis $\{[\widetilde{\ul{d}}]\}$ in the universal Verma module $M$ over $U$. Though we can not identify it with any known type of basis, the parametrization of this basis coincides with the polyhedral realization of the crystal base of $U^+_v(\mathfrak{sl}_n)$ corresponding to the reduced expression in the Weyl group of $SL_n$: $$w_0=s_{n-1}s_{n-2}\ldots s_1s_{n-1}s_{n-2}\ldots s_2\ldots s_{n-1}s_{n-2} s_{n-1}$$ (see ~\cite{nz}). \subsection{} \label{Shapovalov} Recall that the universal Verma module $M$ over $U$ is equipped with the symmetric Shapovalov form $(,)$ with values in $\on{Frac}(\BC[\widetilde{T}\times\BC^*])$. It is characterized by the properties (a) $([\widetilde{\ul{d}}{}_0],[\widetilde{\ul{d}}{}_0])=1$ where $[\widetilde{\ul{d}}{}_0]=[(0,\ldots,0)]$ is the lowest weight vector; (b) $(E_ix,y)=(x,F_iy)\ \ \forall\ x,y\in M$. \medskip We will write down a geometric expression for the Shapovalov form. Evidently, the different weight spaces of $M$ are orthogonal with respect to the Shapovalov form. We consider the line bundle $\D_i$ on $\fQ_{\ul{d}}$ whose fiber at the point $(\CW_\bullet)$ equals $\det R\Gamma(\bC,\CW_i)$. We also define the line bundle $\D:=\bigotimes_{i=1}^{n-1}\D_i$. \begin{prop} \label{Shapoval} For $\CG_1,\CG_2\in M_{\ul{d}}$ we have $$(\CG_1,\CG_2)=(-1)^{\sum_{i=1}^{n-1}d_i} v^{\sum_{i=1}^{n-1}2id_i^2-\sum_{i=2}^{n-1}(2i-1)d_id_{i-1}} \prod_{i=1}^nt_i^{(2i-1)(d_{i-1}-d_i)}[R\Gamma(\fQ_{\ul{d}}, \CG_1\otimes\CG_2\otimes\D)]$$ \end{prop} \begin{proof} Since $\det R\Gamma$ is multiplicative in short exact sequences, we have an equality of line bundles on the correspondence $\fE_{\ul{d},i}:\ \bp^*\D=\bq^*\D\otimes\fL_i$. Now the projection formula shows that the operators $\bp_*\bq^*$ and $\bq_*(\fL_i\otimes\bp^*)$ are adjoint with respect to the pairing $\CG_1,\CG_2\mapsto R\Gamma(\fQ_{\ul{d}},\CG_1\otimes\CG_2\otimes\D)$. Finally, it is easy to see that the $v,t$-factor takes care of the scaling coefficients of our $E_i,F_i$. \end{proof} \subsection{} \label{conjugate} While the operators $E_i,F_i$ are conjugate to each other with respect to the Shapovalov form, the operators $e_i,f_i$ are not. In fact, obviously, $e_i^*=K_i^{2i}f_i$. It is known that a completion of the universal Verma module $M$ contains a unique vector $\fk=\sum_{\ul{d}}\fk_{\ul{d}}$ (resp. $\fw=\sum_{\ul{d}}\fw_{\ul{d}}$) such that $\fk_{(0,\ldots,0)}=\fw_{(0,\ldots,0)}=[(0,\ldots,0)]$, and $f_i\fk=(1-v^2)^{-1}\fk$ (resp. $e_i^*\fw=(1-v^2)^{-1}\fw$) for any $i$ (the {\em Whittaker vectors}). The following proposition gives a geometric construction of the Whittaker vectors $\fk,\fw\in M$. \begin{prop} \label{hlop} a) $\fk_{\ul{d}}=[\CO_{\ul{d}}]$ (the class of the structure sheaf of $\fQ_{\ul{d}}$); b) $\fw_{\ul{d}}= v^{\sum_{i=1}^{n-1}(1-2i)d_i^2-\sum_{i=2}^{n-1}(2-2i)d_id_{i-1}- \sum_{i=1}^{n-1}d_i} \prod_{i=1}^nt_i^{(2-2i)(d_{i-1}-d_i)}[\D^{-1}_{\ul{d}}]$. \end{prop} \begin{proof} a) We have $\bq^*\CO_{\ul{d}+i}=\CO_{\fE_{\ul{d},i}}$. Furthermore, since $\bp\times\br:\ \fE_{\ul{d},i}\to\fQ_{\ul{d}}\times(\bC-\infty)$ is proper and birational, and both the source and the target are smooth, we have $(\bp\times\br)_*[\CO_{\fE_{\ul{d},i}}]= [\CO_{\ul{d}}]\boxtimes[\CO_{\bC-\infty}]$. In effect, $(\bp\times\br)_*\CO_{\fE_{\ul{d},i}}= \CO_{\ul{d}}\boxtimes\CO_{\bC-\infty}$, and the higher direct images $R^{>0}(\bp\times\br)_*\CO_{\fE_{\ul{d},i}}$ vanish. Finally, $pr_*[\CO_{\ul{d}}\boxtimes\CO_{\bC-\infty}]= (1-v^2)^{-1}[\CO_{\ul{d}}]$ where $pr:\ \fQ_{\ul{d}}\times(\bC-\infty)\to \fQ_{\ul{d}}$ is the projection to the first factor. b) Recall that $e_i^*=K_i^{2i}f_i$. Thus we have to check that $f_i[\D^{-1}_{\ul{d}+i}]=t_i^2v^{2d_{i-1}-2d_i}(1-v^2)^{-1}[\D^{-1}_{\ul{d}}]$. Furthermore, recall that on $\fE_{\ul{d},i}$ we have a canonical isomorphism $\bq^*\D^{-1}_{\ul{d}+i}=\fL_i\otimes\bp^*\D^{-1}_{\ul{d}}$. By the projection formula we are reduced to \begin{equation} \label{zhelaem} \bp_*[\fL_i]=t_i^2v^{2d_{i-1}-2d_i}(1-v^2)^{-1}[\CO_{\ul{d}}] \end{equation} This can be calculated in the basis $[\widetilde{\ul{d}}]$ where we already know the matrix coefficients of our operators (see Corollary ~\ref{coefficients}). More precisely, by the Bott-Lefschetz fixed point formula, we have to check $$\sum_{j\leq i}t_j^2v^{-2d_{ij}}(1-v^2)^{-1} \prod_{j\ne k\leq i}(1-t_j^2t_k^{-2}v^{2d_{i,k}-2d_{i,j}})^{-1} \prod_{k\leq i-1}(1-t_j^2t_k^{-2}v^{2d_{i-1,k}-2d_{i,j}})=$$ $$=t_i^2v^{2d_{i-1}-2d_i}(1-v^2)^{-1}$$ Recall the change of variables we used in the proof of Proposition ~\ref{mrak}: $s_j:=t_j^2v^{-2d_{ij}},\ 1\leq j\leq i;\ p_k:=t_k^2v^{-2d_{i-1,k}},\ 1\leq k\leq i-1$. Then we have to prove $$\sum_{j\leq i}s_j\prod_{k\leq i}^{k\ne j}(1-s_js_k^{-1})^{-1} \prod_{k\leq i-1}(1-s_jp_k^{-1})=s_1\cdots s_ip_1^{-1}\cdots p_{i-1}^{-1}$$ This follows immediately from the well known identity $$\sum_{j\leq i}\prod_{k\leq i-1}(p_k-s_j) \prod_{k\leq i}^{k\ne j}(s_k-s_j)^{-1}=1.$$ This completes the proof of the Proposition. \end{proof} \begin{cor} \label{vot tebe} The Shapovalov scalar product of the Whittaker vectors equals $(\fk_{\ul{d}},\fw_{\ul{d}})=(-1)^{\sum_{i=1}^{n-1}d_i} v^{\sum_{i=1}^{n-1}d_i^2-\sum_{i=2}^{n-1}d_id_{i-1}-\sum_{i=1}^{n-1}d_i} \prod_{i=1}^nt_i^{d_{i-1}-d_i}[R\Gamma(\fQ_{\ul{d}},\CO_{\ul{d}})]$. \end{cor} \subsection{} \label{generating} According to the works ~\cite{e}, ~\cite{s2}, the appropriate generating function of the Shapovalov scalar product of the Whittaker vectors satisfies a $v$-deformed ($v$-difference) version of the quantum Toda lattice equations. Let us recall the required notations and results. We introduce the formal variables $\sz_1,\ldots,\sz_n$, and we set $\sQ_i=\exp(\sz_i-\sz_{i+1}),\ i=1,\ldots,n-1$. We set $\hbar=\log(v)$, so that $v=\exp(\hbar)$. We introduce the shift operators $\sT_i,\ i=1,\ldots,n$, acting on the space of functions of $\sz_1,\ldots,\sz_n$ invariant with respect to the simultaneous translations $f(\sz_1,\ldots,\sz_n)=f(\sz_1+\sz,\ldots,\sz_n+\sz)$. Namely, we set $\sT_if(\sz_1,\ldots,\sz_n)= f(\sz_1,\ldots,\sz_i+\hbar,\ldots,\sz_n)$. We define the following $v$-difference operators: \begin{equation} \label{Etingof} {\mathfrak S}:=\sum_{j=1}^n\sT_j^2+v^{-2}\sum_{i=1}^{n-1}\sQ_i\sT_i\sT_{i+1} \end{equation} \begin{equation} \label{Givental} {\mathfrak G}:=\sT_1^2+\sT_2^2(1-\sQ_1)+\ldots+\sT_n^2(1-\sQ_{n-1}) \end{equation} We also consider the following generating functions: \begin{equation} \label{etingof} {\mathfrak I}:=\prod_{i=1}^{n-1}\sQ_i^{\frac{-\log(t_1\cdots t_i)}{\hbar}} \sum_{\ul{d}}(\fk_{\ul{d}},\fw_{\ul{d}})\sQ_1^{d_1}\cdots\sQ_{n-1}^{d_{n-1}} \end{equation} \begin{equation} \label{givental} {\mathfrak J}:=\prod_{i=1}^{n-1}\sQ_i^{\frac{-\log(t_1\cdots t_i)}{\hbar}} \sum_{\ul{d}}[R\Gamma(\fQ_{\ul{d}}, \CO_{\ul{d}})]\sQ_1^{d_1}\cdots\sQ_{n-1}^{d_{n-1}} \end{equation} Then according to the last formula of ~\cite{s2} (or equivalently, the formula ~(5.7) of ~\cite{e}), we have \begin{equation} \label{etisev} {\mathfrak S}{\mathfrak I}=\left(\sum_{i=1}^nt_i^2\right){\mathfrak I} \end{equation} In effect, the seeming discrepancy between the formula ~(\ref{Etingof}) above, and the formula ~(5.7) of ~\cite{e} is explained by the fact that (a) our $v$ corresponds to $q$ of ~\cite{e}; (b) our Whittaker vectors have eigenvalue $(1-v^2)^{-1}$, whereas the Whittaker vectors of ~\cite{e} have eigenvalue 1, which takes care of the factor $(q-q^{-1})^2$ in the second summand of the formula ~(5.7) of ~\cite{e}. Now the argument of ~\cite{e}, section ~6 (see the formula ~(6.5)) together with Corollary ~\ref{vot tebe}, establishes \begin{equation} \label{talgiven} {\mathfrak G}{\mathfrak J}=\left(\sum_{i=1}^nt_i^2\right){\mathfrak J} \end{equation} thus reproving the Main Theorem ~2 of ~\cite{gl}. \section{Parabolic sheaves and affine quantum groups} In this section we want to generalize the previous results to the affine setting. \label{p} \subsection{Parabolic sheaves} We recall the setup of ~\cite{fgk}. Let $\bX$ be another smooth projective curve of genus zero. We fix a coordinate $x$ on $\bX$, and consider the action of $\BC^*$ on $\bX$ such that $u(x)=u^{-2}x$. We have $\bX^{\BC^*}=\{0_\bX,\infty_\bX\}$. Let $\bS$ denote the product surface $\bC\times\bX$. Let $\bD_\infty$ denote the divisor $\bC\times\infty_\bX\cup\infty_\bC\times\bX$. Let $\bD_0$ denote the divisor $\bC\times0_\bX$. Given an $n$-tuple of nonnegative integers $\ul{d}=(d_0,\ldots,d_{n-1})$, we say that a {\em parabolic sheaf} $\CF_\bullet$ of degree $\ul{d}$ is an infinite flag of torsion free coherent sheaves of rank $n$ on $\bS:\ \ldots\subset\CF_{-1}\subset\CF_0\subset\CF_1\subset\ldots$ such that: (a) $\CF_{k+n}=\CF_k(\bD_0)$ for any $k$; (b) $ch_1(\CF_k)=k[\bD_0]$ for any $k$: the first Chern classes are proportional to the fundamental class of $\bD_0$; (c) $ch_2(\CF_k)=d_i$ for $i\equiv k\pmod{n}$; (d) $\CF_0$ is locally free at $\bD_\infty$ and trivialized at $\bD_\infty:\ \CF_0|_{\bD_\infty}=W\otimes\CO_{\bD_\infty}$; (e) For $-n\leq k\leq0$ the sheaf $\CF_k$ is locally free at $\bD_\infty$, and the quotient sheaves $\CF_k/\CF_{-n},\ \CF_0/\CF_k$ (both supported at $\bD_0=\bC\times0_\bX\subset\bS$) are both locally free at the point $\infty_\bC\times0_\bX$; moreover, the local sections of $\CF_k|_{\infty_\bC\times \bX}$ are those sections of $\CF_0|_{\infty_\bC\times \bX}=W\otimes\CO_\bX$ which take value in $\langle w_1,\ldots,w_{n-k}\rangle\subset W$ at $\infty_\bX\in \bX$. \medskip According to ~\cite{fgk}, ~3.5, the fine moduli space $\CP_{\ul{d}}$ of degree $\ul{d}$ parabolic sheaves exists and is a smooth connected quasiprojective variety of dimension $2d_0+\ldots+2d_{n-1}$. The group $\widetilde{T}\times\BC^*\times\BC^*$ acts naturally on $\CP_{\ul{d}}$, and its fixed point set is finite. \subsection{Correspondences} If the collections $\ul{d}$ and $\ul{d}'$ differ at the only place $i\in I:=\BZ/n\BZ$, and $d'_i=d_i+1$, then we consider a correspondence $\sE_{\ul{d},i}\subset\CP_{\ul{d}}\times\CP_{\ul{d}'}$ formed by the pairs $(\CF_\bullet,\CF'_\bullet)$ such that for $j\not\equiv i\pmod{n}$ we have $\CF_j=\CF'_j$, and for $j\equiv i\pmod{n}$ we have $\CF'_j\subset\CF_j$. It is a smooth quasiprojective algebraic variety of dimension $2\sum_{i\in I}d_i+1$. In effect, the argument of ~\cite{fgk}, ~Lemma~3.3, reduces this statement to the corresponding fact about Laumon correspondences (see ~\cite{la1}, ~2.10). We denote by $\bp$ (resp. $\bq$) the natural projection $\sE_{\ul{d},i}\to\CP_{\ul{d}}$ (resp. $\sE_{\ul{d},i}\to\CP_{\ul{d}'}$). For $j\equiv i\pmod{n}$ the correspondence $\sE_{\ul{d},i}$ is equipped with a natural line bundle $\sL_j$ whose fiber at $(\CF_\bullet,\CF'_\bullet)$ equals $\Gamma(\bC,\CF_{j-n}/\CF'_{j-n})$. Finally, we have a transposed correspondence $^\sT\sE_{\ul{d},i}\subset\CP_{\ul{d}'}\times\CP_{\ul{d}}$. \subsection{} We denote by ${}'\CM$ the direct sum of equivariant (complexified) $K$-groups: ${}'\CM=\oplus_{\ul{d}}K^{\widetilde{T}\times\BC^*\times\BC^*}(\CP_{\ul{d}})$. It is a module over $K^{\widetilde{T}\times\BC^*\times\BC^*}(pt) =\BC[\widetilde{T}\times\BC^*\times\BC^*]= \BC[t_1,\ldots,t_n,v,u\ :\ t_1\cdots t_n=1]$. We define $\CM=\ {}'\CM\otimes_{K^{\widetilde{T}\times\BC^*\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*\times\BC^*}(pt))$. We have an evident grading $\CM=\oplus_{\ul{d}}\CM_{\ul{d}},\ \CM_{\ul{d}}=K^{\widetilde{T}\times\BC^*\times\BC^*}(\CP_{\ul{d}}) \otimes_{K^{\widetilde{T}\times\BC^*\times\BC^*}(pt)} \on{Frac}(K^{\widetilde{T}\times\BC^*\times\BC^*}(pt))$. \subsection{} \label{operators'} The grading and the correspondences $^\sT\sE_{\ul{d},i},\sE_{\ul{d},i}$ give rise to the following operators on $\CM$ (note that though $\bp$ is not proper, $\bp_*$ is well defined on the localized equivariant $K$-theory due to the finiteness of the fixed point sets): $K_i=t_{i+1}t_i^{-1}u^{\delta_{0,i}}v^{2d_i-d_{i-1}-d_{i+1}+1}:\ \CM_{\ul{d}}\to\CM_{\ul{d}}$, $C=uv^n$, For $i=0,\ldots,n-1$ we define $L_i=t_1^{-1}\cdots t_i^{-1}v^{d_i+\frac{1}{2}i(n-i)}:\ M_{\ul{d}}\to M_{\ul{d}}$ (that is, $L_0=v^{d_0}$), $f_i=\bp_*\bq^*:\ \CM_{\ul{d}}\to\CM_{\ul{d}-i}$; For $n>2$ and $i=0,\ldots,n-1$ we define $F_i=t_{i+1}^{-1}v^{d_{i+1}-d_i+\frac{n+1}{2}-i}\bp_*\bq^*:\ \CM_{\ul{d}}\to\CM_{\ul{d}-i}$, For $n=2$ we define $F_i=f_i$, $e_i=-t_i^{-1}t_{i+1}^{-1}u^{\delta_{0,i}}v^{d_{i+1}-d_{i-1}} \bq_*(\sL_i\otimes\bp^*):\ \CM_{\ul{d}}\to\CM_{\ul{d}+i}$, For $n>2$ and $i=0,\ldots,n-1$ we define $E_i=-t_i^{-1}u^{\delta_{0,i}}v^{d_i-d_{i-1}+\frac{1-n}{2}+i} \bq_*(\sL_i\otimes\bp^*):\ \CM_{\ul{d}}\to\CM_{\ul{d}+i}$, For $n=2$ we define $E_i=e_i$. \subsection{Sevostyanov's form of affine quantum $SL_n$} Let $I$ denote the set $\BZ/n\BZ$ of residue classes modulo $n$. $\CU$ is the $\BC[v,v^{-1}]$-algebra with generators $E_i,L_i^{\pm1},K_i^{\pm1},C^{\pm1},F_i,\ i\in\BZ/n\BZ$, subject to the following relations: \begin{equation} \label{ev} L_iL_j=L_jL_i,\ K_i=L_i^2L_{i+1}L_{i-1}C^{\delta_{i,0}} \end{equation} \begin{equation} \label{evid} L_jE_iL_j^{-1} =v^{\delta_{i,j}}E_i,\ L_jF_iL_j^{-1} =v^{-\delta_{i,j}}F_i,\ C\hphantom{m}\on{is}\hphantom{m}\on{central} \end{equation} \begin{equation} \label{evident} E_iF_j-F_jE_i=\delta_{i,j}\frac{K_i-K_i^{-1}}{v-v^{-1}} \end{equation} \begin{equation} \label{Ser1} |i-j|>1\ \Longrightarrow\ E_iE_j-E_jE_i=0=F_iF_j-F_jF_i \end{equation} \begin{equation} \label{Ser2} n>2\ \&\ |i-j|=1\ \Longrightarrow\ E_i^2E_j-(v+v^{-1})E_iE_jE_i+E_jE_i^2=0= F_i^2F_j-(v+v^{-1})F_iF_jF_i+F_jF_i^2 \end{equation} \begin{equation} \label{Ser22} n=2\ \&\ |i-j|=1\ \Longrightarrow\ E_i^3E_j-(v^2+1+v^{-2})E_i^2E_jE_i+ (v^2+1+v^{-2})E_iE_jE_i^2-E_jE_i^3=0 \end{equation} \begin{equation} \label{Ser22F} n=2\ \&\ |i-j|=1\ \Longrightarrow\ F_i^3F_j-(v^2+1+v^{-2})F_i^2F_jF_i+ (v^2+1+v^{-2})F_iF_jF_i^2-F_jF_i^3=0 \end{equation} For $n>2$ we also consider elements $e_i,f_i\in\CU$ depending on the following choice of $n\times n$-matrices $n_{ij},\ c_{ij}$ (cf. ~\cite{s}, ~Remark ~3): \begin{equation} \label{choice} n_{i,i}=1,\ n_{i,i+1}=-1,\ n_{i+1,i}=0, \end{equation} otherwise $n_{ij}=0$. \begin{equation} \label{Russia} c_{i,i+1}=-1,\ c_{i+1,i}=1, \end{equation} otherwise $c_{ij}=0$. Then we set \begin{equation} \label{Sevo} f_i:=L_iL_{i+1}^{-1}F_i,\ e_i:=E_iL_i^{-1}L_{i+1}. \end{equation} Clearly, the algebra $\CU$ is generated by $e_i,L_i^{\pm1},K_i^{\pm1}, C^{\pm1},f_i,\ i\in\BZ/n\BZ$, and the relations ~(\ref{evid})--~(\ref{Ser2}) above are equivalent to the relations ~(\ref{evid'})--~(\ref{Ser2'}) below. \begin{equation} \label{evid'} L_je_iL_j^{-1} =v^{\delta_{i,j}}e_i,\ L_jf_iL_j^{-1} =v^{-\delta_{i,j}}f_i,\ C\hphantom{m}\on{is}\hphantom{m}\on{central} \end{equation} \begin{equation} \label{evident'} e_if_j-v^{c_{ij}}f_je_i=\delta_{i,j}\frac{K_i-K_i^{-1}}{v-v^{-1}} \end{equation} \begin{equation} \label{Ser1'} |i-j|>1\ \Longrightarrow\ e_ie_j-e_je_i=0=f_if_j-f_jf_i \end{equation} \begin{equation} \label{Ser2'} |i-j|=1\ \Longrightarrow\ e_i^2e_j-v^{c_{ij}}(v+v^{-1})e_ie_je_i+v^{2c_{ij}}e_je_i^2=0= f_i^2f_j-v^{c_{ij}}(v+v^{-1})f_if_jf_i+v^{2c_{ij}}f_jf_i^2 \end{equation} \subsection{} The following is an affine analogue of Theorem ~\ref{main}. Recall the operators $E_i,e_i,K_i^{\pm1},L_i^{\pm1},C^{\pm1},F_i,f_i,\ i\in I$, on $\CM$ defined in ~\ref{operators'}. \begin{conj} \label{main'} The operators $E_i,K_i^{\pm1},L_i^{\pm1},C^{\pm1},F_i,\ i\in I$, on $\CM$ satisfy the relations ~(\ref{ev})--~(\ref{Ser22F}). Equivalently, if $n>2$, the operators $e_i,K_i^{\pm1},L_i^{\pm1},C,f_i,\ i\in I$, satisfy the relations ~(\ref{ev}), ~(\ref{evid'})--~(\ref{Ser2'}). \end{conj} \subsection{} \label{long} We can prove Conjecture ~\ref{main'} for $n>2$ . Let us sketch this proof. It is parallel to the proof of Theorem ~\ref{main}. In effect, the relation ~(\ref{evident'}) for $i\ne j$ follows from the transversality statement absolutely similar to Lemma ~\ref{trans}. More precisely, the argument of ~\cite{fgk} ~(Lemma ~3.3), reduces the required smoothness to that proved in Lemma ~\ref{trans}. The relation ~(\ref{evident'}) for $j=i$ follows from Conjecture ~\ref{inteligent} by the argument of ~\ref{derivation}. Since we can not prove Conjecture ~\ref{inteligent} at the moment, we will derive the relation ~(\ref{evident'}) for $j=i$ from its weaker but accessible form. To this end we consider the following closed substack $\fZ'_{\fd_1,\fd_2}\subset\fZ_{\fd_1,\fd_2}$. Recall that a coherent sheaf $\fW_1$ (resp. $\fW_2$) contains the maximal torsion subsheaf $\fW_1^{tors}$ (resp. $\fW_2^{tors}$) with the locally free quotient sheaf $\fW_1^{free}$ (resp. $\fW_2^{free}$). Moreover, we have $\fW_1\simeq\fW_1^{tors}\oplus\fW_1^{free}$ (resp. $\fW_2\simeq\fW_2^{tors}\oplus\fW_2^{free}$). The closed substack $\fZ'_{\fd_1,\fd_2}\subset\fZ_{\fd_1,\fd_2}$ classifies the flags of coherent sheaves (with trivialization at $\infty\in\bC$) $\fW_1\subset\fW_2$ such that $\deg\fW_1^{free}\leq0\geq\deg\fW_2^{free}$. We define $K^{\fT\times\BC^*}(\fZ'_{\fd_1,\fd_2})$ as the $K$-group of $\fT\times\BC^*$-equivariant coherent sheaves on the smooth stack $\fZ_{\fd_1,\fd_2}$ supported on the closed substack $\fZ'_{\fd_1,\fd_2}$. Note that for any $\ul{d}=(d_1,\ldots,d_{n-1})$ the map $\fz_{\ul{d}}:\ \fQ_{\ul{d}}\to\fZ_{d_i-d_{i-1},d_{i+1}-d_i}$ factors through the same named map into the closed substack $\fZ'_{d_i-d_{i-1},d_{i+1}-d_i}$. Similarly, for any $\ul{d}=(d_0,d_1,\ldots,d_{n-1})$ the map $$\fz\fz_{\ul{d}}:\ \CP_{\ul{d}}\to\fZ_{d_i-d_{i-1},d_{i+1}-d_i},\ \CF_\bullet\mapsto(\CF_i/\CF_{i-1},\CF_{i+1}/\CF_{i-1})$$ factors through the same named map into the closed substack $\fZ'_{d_i-d_{i-1},d_{i+1}-d_i}$. Let $(\fW_1\subset\fW_2)$ be a $\fT\times\BC^*$-fixed point of $\fZ'_{\fd_1,\fd_2}$. Let $\iota_{(\fW_1\subset\fW_2)}$ denote its locally closed embedding into $\fZ'_{\fd_1,\fd_2}$. Let $Aut_{(\fW_1\subset\fW_2)}$ stand for its automorphisms' group. One can easily check the following \begin{lem} \label{easily} There exists $n,\ i,\ 1\leq i\leq n-1,\ \ul{d}=(d_1,\ldots,d_n)$, and a fixed point $\widetilde{\ul{d}}\in\fQ_{\ul{d}}^{\widetilde{T}\times\BC^*}$ such that (a) $\fz_{\ul{d}}(\widetilde{\ul{d}})=(\fW_1\subset\fW_2)$; (b) $\fz_{\ul{d}}(\widetilde{T}\times\BC^*)$ is a maximal torus of $Aut_{(\fW_1\subset\fW_2)}$. \end{lem} \subsection{} \label{birka} One way to prove Conjecture ~\ref{inteligent} would be to reverse the argument of ~\ref{derivation} and derive it from the relations ~(\ref{ochevidno'}) for all $n,i$. In effect, we must compute (notations of ~\ref{derivation}) $\fX\in K^{\fT\times\BC^*}(\fZ_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)} \on{Frac}(K^{\fT\times\BC^*}(pt))$ while we know $\fz_{\ul{d}}^*\fX$ for all $n,i,\ul{d}$ such that $d_i-d_{i-1}=\fd_1,\ d_{i+1}-d_i=\fd_2$ (also, the homomorphism of tori $\widetilde{T}_n\to\fT$ acts on the characters as $\tau_1=t_i,\ \tau_2=t_{i+1}$). Let us denote by $\fY\in K^{\fT\times\BC^*}(\fZ'_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)} \on{Frac}(K^{\fT\times\BC^*}(pt))$ the restriction of $\fX$ to $\fZ'_{\fd_1,\fd_2}$. The Lemma ~\ref{easily} implies that the kernel $Ker_1$ of the direct product of inverse images $$\prod_{n,i,\ul{d}}\fz_{\ul{d}}^*:\ K^{\fT\times\BC^*}(\fZ'_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))\to \prod_{n,i,\ul{d}}M_{\ul{d}}$$ coincides with the kernel $Ker_2$ of the direct product of restrictions $$\prod_{(\fW_1\subset\fW_2)\in(\fZ'_{\fd_1,\fd_2})^{\fT\times\BC^*}} \iota^*_{(\fW_1\subset\fW_2)}:\ K^{\fT\times\BC^*}(\fZ'_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))\to$$ $$\to\prod_{(\fW_1\subset\fW_2)\in(\fZ'_{\fd_1,\fd_2})^{\fT\times\BC^*}} K^{\fT\times\BC^*\times Aut_{(\fW_1\subset\fW_2)}}(pt) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))$$ It follows that for any $n,\ i,\ 0\leq i\leq n-1,\ \ul{d}=(d_0,\ldots,d_n)$, such that $\fd_1=d_i-d_{i-1},\ \fd_2=d_{i+1}-d_i$, the kernel $Ker_1=Ker_2$ is contained in the kernel $Ker_3$ of the inverse image $$\fz\fz_{\ul{d}}^*:\ K^{\fT\times\BC^*}(\fZ'_{\fd_1,\fd_2}) \otimes_{K^{\fT\times\BC^*}(pt)}\on{Frac}(K^{\fT\times\BC^*}(pt))\to \CM_{\ul{d}}$$ By the argument of ~\ref{derivation} we know that $\fY=\frac{K-K^{-1}}{v-v^{-1}}$ modulo $Ker_1$, and hence the same holds modulo $Ker_3$. The argument of {\em loc. cit.} then shows that the relation ~(\ref{evident'}) for $j=i$ holds in $\CM$. \subsection{} \label{longer} It remains to check the Serre relations. The relations for negative generators follow from the relations for positive generators because they are adjoint with respect to the nondegenerate Shapovalov form, see ~\ref{Shapo} below. So it suffices to consider the relations ~(\ref{Ser1}), ~(\ref{Ser2}) between $E_i,E_j,\ i\ne j$. It is here that we need the assumption $n>2$ for technical reasons. Namely, for $n>2$ we can find $k\in I$ such that $i\ne k\ne j$. We consider an $n$-dimensional vector space with a basis $\fw_1,\ldots,\fw_n$, and a torus $\fT$ acting on $\fw_l$ by the character $\tau^2_l$. Let $\fZ_n$ be the moduli stack of flags of coherent sheaves $\fW_1\subset\ldots\subset\fW_n$ on $\bC$ locally free at $\infty\in\bC$, equipped with compatible trivializations $\fW_l|_\infty=\langle\fw_1,\ldots,\fw_l\rangle$. Note that $\fZ_n$ has connected components numbered by the degrees of $\fW_l$, which for $n=2$ coincide with the stacks $\fZ_{\fd_1,\fd_2}$. Absolutely similarly to ~\ref{alternate} we introduce the correspondences between various connected components, which give rise to the operators $E^\fZ_1,\ldots,E^\fZ_{n-1}$ on the localized equivariant $K$-theory of $\fZ_n$. As in ~\ref{long} above, we have a closed substack $\fZ'_n\subset\fZ_n$ classifying the flags such that $\deg\fW_l^{free}\leq0,\ 1\leq l\leq n$. We have a map $$\fz\fz_k:\ \CP_{\ul{d}}\to\fZ_n,\ (\CF_\bullet)\mapsto (\CF_{k+1}/\CF_k\subset\ldots\subset\CF_{k+n}/\CF_k)$$ factoring through the same named map $\CP_{\ul{d}}\to\fZ'_n$. For any $N\geq n$, and $m$ such that $0\leq m\leq N-n$, and $\ul{d}=(d_1,\ldots,d_N)$, we also have a map $$\fz_{m,\ul{d}}:\ \fQ_{\ul{d}}\to\fZ_n,\ (\CW_\bullet)\mapsto(\CW_{m+1}/\CW_m\subset\ldots\subset\CW_{m+n}/\CW_m)$$ factoring through the same named map $\fQ_{\ul{d}}\to\fZ'_n$. Now the argument of ~\ref{derivation} shows that the Serre relation between $E_i,E_j$ would follow from the Serre relation between $E^\fZ_{i'},E^\fZ_{j'}$ for certain $i',j'$. Though we cannot establish the latter relations, the argument of ~\ref{birka} shows that they hold modulo the subspace $Ker_1$ (because we already know the Serre relations for $\mathfrak{sl}_N$ with arbitrary $N$), and also shows that this suffices to derive the former relations. This completes the proof of the Serre relations for $n>2$. Thus, Conjecture ~\ref{main'} is proved for $n>2$. \subsection{} Similarly to ~\ref{Shapoval}, we will write down a geometric expression for a Shapovalov form on $\CM$, that is a symmetric $\operatorname{Frac}(\BC[\widetilde{T}\times\BC^*\times\BC^*])$-valued bilinear form on $\CM$ such that $(E_im_1,m_2)=(m_1,F_im_2)$ for any $i\in I$, and $m_1,m_2\in\CM$. The different weight spaces of $\CM$ will be orthogonal with respect to this geometric Shapovalov form. For $i=0,\ldots,n-1$, we consider the line bundle $\D_i$ on $\CP_{\ul{d}}$ whose fiber at the point $(\CF_\bullet)$ equals $\det R\Gamma(\bS,\CF_{i-n})$. We also define the line bundle $\D_{\ul{d}}:=\bigotimes_{i=0}^{n-1}\D_i$. For $\CG_1,\CG_2\in\CM_{\ul{d}}$, we set \begin{equation} \label{Chapo} (\CG_1,\CG_2):=(-1)^{\sum_{i=0}^{n-1}d_i} v^{-\sum_{i=0}^{n-1}d_i^2+\sum_{i=0}^{n-1}d_id_{i+1} +\sum_{i=0}^{n-1}(n-2i)d_i}u^{-d_0} \prod_{i=0}^{n-1}t_i^{d_i-d_{i-1}}R\Gamma(\CP_{\ul{d}}, \CG_1\otimes\CG_2\otimes\D_{\ul{d}}) \end{equation} Clearly, the form $(,)$ is nondegenerate, since the classes of the structure sheaves of the $\widetilde{T}\times\BC^*\times\BC^*$-fixed points form an orthogonal basis of $\CM$. The following proposition is proved exactly as ~\ref{Shapoval}. \begin{prop} \label{Shapo} For $i\in I,\ \CG_1\in\CM_{\ul{d}},\ \CG_2\in\CM_{\ul{d}+i}$ we have $(E_i\CG_1,\CG_2)=(\CG_1,F_i\CG_2)$. \end{prop} \subsection{} We define a formal sum in a completion of $\CM$ as follows: $\fn=\sum_{\ul{d}}\fn_{\ul{d}}:=\sum_{\ul{d}}[\CO_{\ul{d}}] =\sum_{\ul{d}}[\CO_{\CP_{\ul{d}}}]$. We also consider the following formal sum: $\fu=\sum_{\ul{d}}\fu_{\ul{d}}$ where \begin{equation} \fu_{\ul{d}}=v^{2\sum_{i=0}^{n-1}d_i^2-2\sum_{i=0}^{n-1}d_id_{i+1} -\sum_{i=0}^{n-1}(n-2i+1)d_i}u^{2d_0} \prod_{i=1}^nt_i^{2d_{i-1}-2d_i}[\D^{-1}_{\ul{d}}] \end{equation} \begin{prop} \label{vykusi} a) $\fn$ is a common eigenvector of the operators $f_i$ with the eigenvalue $(1-v^2)^{-1}$; b) $\fu$ is a common eigenvector of the operators $e_i^*$ with the eigenvalue $(1-v^2)^{-1}$. \end{prop} \begin{proof} a) is proved exactly as Proposition ~\ref{hlop} ~(a). To check b) we argue as in the proof of Proposition ~\ref{hlop} ~(b), and reduce it to \begin{equation} \label{hotim} \bp_*[\sL_i]=t_i^2u^{-2\delta_{0,i}}v^{2d_{i-1}-2d_i} (1-v^2)^{-1}[\CO_{\ul{d}}] \end{equation} To verify this we recall the setup of ~\ref{alternate}, and claim that in the notations of {\em loc. cit.} we have \begin{equation} \label{imeem} \bp_*[\fL_{\fd_1}]=\tau_1^2v^{-2\fd_1}(1-v^2)^{-1} [\CO_{\fZ_{\fd_1,\fd_2}}] \end{equation} In effect, ~(\ref{imeem}) is deduced from ~(\ref{zhelaem}) by the argument of ~\ref{birka}. Finally, ~(\ref{hotim}) is deduced from ~(\ref{imeem}) by the argument of ~\ref{derivation}. The Proposition is proved. \end{proof} \begin{cor} \label{vot te} The Shapovalov scalar product of the Whittaker vectors equals $(\fn_{\ul{d}},\fu_{\ul{d}})=(-1)^{\sum_{i=0}^{n-1}d_i} v^{\sum_{i=0}^{n-1}d_i^2-\sum_{i=0}^{n-1}d_id_{i-1}-\sum_{i=0}^{n-1}d_i} u^{d_0}\prod_{i=1}^nt_i^{d_{i-1}-d_i}R\Gamma(\CP_{\ul{d}},\CO_{\ul{d}})$. \end{cor} \subsection{} \label{new} We define $\CM'\subset\CM$ as a minimal $\CU$-submodule containing the lowest weight vector $[0,\ldots,0]$. The relations ~(\ref{evident'}) show that $\CM'$ is generated from $[0,\ldots,0]$ by the action of operators $e_i,\ i\in I$. Clearly, $\CM'$ is isomorphic to a universal Verma module over $\CU$. \begin{conj} \label{nakos} The class of the structure sheaf $[\CO_{\ul{d}}]$ lies in $\CM'_{\ul{d}}$. \end{conj} In what follows we shall assume the validity of Conjecture ~\ref{main'} (as was explained above this is actually not an assumption for $n>2$). \begin{prop} \label{nakosi} The class of $[\D^{-1}_{\ul{d}}]$ lies in $\CM'_{\ul{d}}$. \end{prop} \begin{proof} We have $\CM=\CM'\oplus\CM''$ where $\CM''$ is the orthogonal complement of $\CM'$ in $\CM$ with respect to the Shapovalov form. We have to prove that $[\D^{-1}_{\ul{d}}]$ is orthogonal to $\CM''$. Let $A\in\CM''_{\ul{d}}$. Suppose $A=e_iB$ for some $i\in I$ and $B\in\CM''_{\ul{d}-i}$. Then $(A,[\D^{-1}_{\ul{d}}])=(B,e_i^*[\D^{-1}_{\ul{d}}])$. Thus up to (an invertible) monomial in $t,u,v$ we have $(A,[\D^{-1}_{\ul{d}}])=(B,[D^{-1}_{\ul{d}-i}])$. Hence, arguing by induction in $\ul{d}$ we may assume that $A\in\CM''_{\ul{d}}$ is orthogonal to the image of any $e_i$. Then $e_i^*A=0$ or, equivalently, $f_iA=0$ for any $i\in I$. Up to (an invertible) monomial in $t,u,v$ we have $(A,[\D^{-1}_{\ul{d}}])=R\Gamma(\CP_{\ul{d}},A)$. Thus we are reduced to the following claim for $\ul{d}\ne(0,\ldots,0)$: \begin{equation} \label{ponjatno} f_iA=0\ \forall\ i\in I\ \Longrightarrow\ R\Gamma(\CP_{\ul{d}},A)=0. \end{equation} We will derive ~(\ref{ponjatno}) from the corresponding claim in the equivariant (complexified) Borel-Moore homology $H^{\widetilde{T}\times\BC^*\times\BC^*}_{BM}(\CP_{\ul{d}})$. Let $\on{Td}_{\CP_{\ul{d}}}$ denote the equivariant Todd class in the completion of the equivariant cohomology. Let also $\on{ch}_*$ denote the homological Chern character map from the equivariant $K$-theory to the completion of the equivariant Borel-Moore homology (see e.g. ~\cite{cg}). We define $$a:=\on{Td}_{\CP_{\ul{d}}}\cup\on{ch}_*A\in \widehat{H}{}^{\widetilde{T}\times\BC^*\times\BC^*}_{BM}(\CP_{\ul{d}})$$ By the bivariant Riemann-Roch Theorem (see e.g. ~\cite{cg}, ~5.11.11) we have $\on{ch}_*(f_iA)=\on{ch}_*(\bp_*\bq^*A)=\bp_*\bq^*a$ where in the RHS $\bp_*$ and $\bq^*$ refer to the operations in the (localized and completed) equivariant Borel-Moore homology. We also have $R\Gamma(\CP_{\ul{d}},A)=\int_{\CP_{\ul{d}}}a$. Since $\on{ch}_*$ is injective, and the operation $?\mapsto\on{Td}_{\CP_{\ul{d}}}\cup?$ is invertible, the claim ~(\ref{ponjatno}) follows from the corresponding claim in the equivariant Borel-Moore homology $\widehat{H}{}^{\widetilde{T}\times\BC^*\times\BC^*}_{BM}(\CP_{\ul{d}})$: \begin{equation} \label{clear} {\mathfrak f}_ia=0\ \forall\ i\in I\ \Longrightarrow\ \int_{\CP_{\ul{d}}}a=0. \end{equation} Here ${\mathfrak f}_i=\bp_*\bq^*$ is a part of the action of the affine Lie algebra $\widehat{\mathfrak{sl}}_n$ on ${\mathfrak M}:=\oplus_{\ul{d}} \widehat{\ul{H}}{}^{\widetilde{T}\times\BC^*\times\BC^*}_{BM}(\CP_{\ul{d}})$ (localized and completed equivariant Borel-Moore homology). The positive generators act as ${\mathfrak e}_i=-\bq_*\bp^*$. This can be checked along the lines of ~\ref{long}--\ref{longer} but simpler. Reversing the argument in the beginning of the proof, we see that ~(\ref{clear}) is equivalent to the statement that the fundamental cycle $[\CP_{\ul{d}}]\in \widehat{\ul{H}}{}^{\widetilde{T}\times\BC^*\times\BC^*}_{BM}(\CP_{\ul{d}})$ is contained in the subspace ${\mathfrak M}'$ of ${\mathfrak M}$ generated by the action of ${\mathfrak e}_i,\ i\in I$, from $[\CP_{(0,\ldots,0)}]$. Recall the semismall resolution morphism $\pi_{\ul{d}}:\ \CP_{\ul{d}}\to\frP_{\ul{d}}$ to the Uhlenbeck flag space, see ~\cite{fgk}. By the Decomposition Theorem of Beilinson-Bernstein-Deligne-Gabber, the direct sum of (localized and completed) equivariant Intersection Homology $'{\mathfrak M}:=\oplus_{\ul{d}} \widehat{\ul{IH}}{}^{\widetilde{T}\times\BC^*\times\BC^*}(\frP_{\ul{d}})$ is a direct summand of $\mathfrak M$. Now ~\cite{b} defines the action of $\widehat{\mathfrak{sl}}_n$ on $'{\mathfrak M}$, and one can check that the action of ~\cite{b} is the restriction of the above $\widehat{\mathfrak{sl}}_n$-action on $\mathfrak M$. It follows that $'{\mathfrak M}={\mathfrak M}'$. Finally, it is proved in ~\cite{b} that $[\CP_{\ul{d}}]\in\ '{\mathfrak M}$. This completes the proof of the Proposition. \end{proof} \subsection{} We conclude that $\fu$ is the unique Whittaker vector in the completion of the Verma module $\CM'$ with the lowest weight component $\fu_{(0,\ldots,0)}=[(0,\ldots,0)]$ (the common eigenvector of $e_i^*,\ i\in I$, with the eigenvalue $(1-v^2)^{-1}$). Let $\fn'\in\widehat{\CM}{}'$ be the unique common eigenvector of $f_i,\ i\in I$, with the eigenvalue $(1-v^2)^{-1}$ and with the lowest weight component $\fn'_{(0,\ldots,0)}=[(0,\ldots,0)]$. Then $\fn'$ is the orthogonal projection of $\fn$ onto $\widehat{\CM}{}'$ along $\widehat{\CM}{}''$. Hence the Corollary ~\ref{vot te} yields the following \begin{cor} \label{last} One has $$(\fn'_{\ul{d}},\fu_{\ul{d}})=(-1)^{\sum_{i=0}^{n-1}d_i} v^{\sum_{i=0}^{n-1}d_i^2-\sum_{i=0}^{n-1}d_id_{i-1}-\sum_{i=0}^{n-1}d_i} u^{d_0}\prod_{i=1}^nt_i^{d_{i-1}-d_i}[R\Gamma(\CP_{\ul{d}},\CO_{\ul{d}})].$$ \end{cor} \subsection{Some further remarks} The next natural step would be to study the generating function of all $[R\Gamma(\CP_{\ul{d}},\CO_{\ul{d}})]$'s in a way similar to subsection ~\ref{generating}; let us denote this function by $\fJ_{\aff}$. The cohomology (as opposed to $K$-theory) analogue of this is performed in \cite{b} and \cite{be}. In particular, in \cite{b} it is shown that such a function is an eigen-function of a certain linear differential operator of 2nd order (the ``non-stationary analogue" of the quadratic affine Toda hamiltonian). This fact is used in \cite{be} in order to show that certain asymptotic of this function is given by the {\it Seiberg-Witten prepotential} of the corresponding classical affine Toda system. This agrees well with the results of \cite{neok} about a similar asymptotic of the partition function of N=2 supersymmetric gauge theory in 4 dimensions. Unfortunately, in the present ($K$-theoretic) case we can't derive any good equation for the the function $\fJ_{\aff}$. Thus we do not know how to generalize the results of \cite{be} to this case. One can probably show that the results of \cite{neok} on 5d gauge theory imply that a similar asymptotic (when the classical affine Toda lattice is replaced by the classical affine relativistic Toda) is valid for the function $\fJ_{\aff}$, but we do not know how to derive it from Corollary ~\ref{last}.
{ "timestamp": "2005-06-13T11:04:42", "yymm": "0503", "arxiv_id": "math/0503456", "language": "en", "url": "https://arxiv.org/abs/math/0503456" }
\section{Introduction} Recently, Szmytkowski \cite{Szmytkowski} proposed a new formulation of the eigenchannel method for quantum scattering from Hermitian short-range potentials, different from that presented by Danos and Greiner \cite{Danos}. Some ideas leading to this method were drawn from works on electromagnetism theory by Garbacz \cite{Garbacz1} and Harrington and Mautz \cite{Harrington1}. This method was further extended to the case of zero-range potentials for Schr\"odinger particles by Szmytkowski and Gruchowski \cite{Szmytkowski2} and then for Dirac particles by Szmytkowski \cite{Szmytkowski3} (see also \cite{Szmytkowski5}). On the other hand, it is the well-known fact that separable potentials, since they provide analytical solutions to the Lippmann-Schwinger equations \cite{LippSchw}, have found applications in many branches of physics, both in the non-relativistic and relativistic cases \cite{Zast}. (It should be noted that much larger effort has been devoted to the separable potentials in the non-relativistic regime.) Especially, their utility was confirmed in nuclear physics by successful use for describing nucleon-nucleon interactions \cite{NN}. Moreover, methods allowing one to approximate an arbitrary non-local potential by a separable one are known \cite{metody}. In view of what has been said above, it seems interesting to pose the question: {\it how does the new method apply to quantum scattering from non-local separable potentials?} Partially, the answer was given by the author by applying the method to quantum scattering of Schr\"odinger particles from separable potentials \cite{moja}. In the present contribution, we extend considerations from \cite{moja} to the case of Dirac particles. This paper is organized as follows. In Section 2 some facts and notions from the theory of potential scattering of Dirac particles (see \cite{Thaller}) are provided. In Section 3 we concentrate on the special class of non-local potentials, namely, separable potentials. In this context, expressions for the bispinor as well as matrix scattering amplitudes are provided. Section 4 contains main ideas and results. Here, we define {\it eigenchannel vectors}, directly related to eigenchannels, as solutions to a certain weighted eigenproblem. Moreover, we introduce eigenphase-shifts, relating them to eigenvalues of this spectral problem. Within this approach, we also calculate expressions for the scattering amplitude and the average total cross section. In Section 5, scattering from a rank one delta-like separable potential is discussed as an illustrative example. The paper ends with two appendices. \section{Quantum scattering of Dirac particles from non-local potentials} \label{SecII} Let us assume that a free Dirac particle of energy $E$ (with $|E|>mc^{2}$) described by the following monochromatic plane wave \begin{equation}\label{I.1} \phi_{i}(\wektor{r})\equiv\braket{\wektor{r}}{\wektor{k}_{i}\chi_{i}}= U_{i}(\wektor{k}_{i})e^{i\wektor{k}_{i}\cdot\wektor{r}}, \end{equation} where \begin{equation}\label{I.2} U_{i}(\wektor{k}_{i})=\frac{1}{\sqrt{1+\varepsilon^{2}}} \left( \begin{array}{c} \chi_{i} \\*[0.2ex] \displaystyle\varepsilon\wektor{\sigma}\cdot\wersor{k}_{i}\,\chi_{i} \end{array} \right), \end{equation} \begin{equation} \varepsilon=\sqrt{\frac{E-mc^{2}}{E+mc^{2}}} \end{equation} is being scattered from a non-local potential given by a kernel $\mathsf{V}(\wektor{r},\wektor{r}')$, which in general may be a $4\times 4$ matrix. In the above equation, $\chi_{i}$ stands for a normalized pure spin-$\frac{1}{2}$ state belonging to $\mathbb{C}^{2}$. Orientation of the spin in $\mathbb{R}^{3}$ will be denoted by $\wektor{\nu}_{i}$ and is related to $\chi_{i}$ by $\wektor{\nu}_{i}=\chi^{\dagger}_{i}\wektor{\sigma}\chi_{i}$, where $\wektor{\sigma}$ is a vector consisting of the standard Pauli matrices, i.e., \begin{equation}\label{I.3} \wektor{\sigma}= \left[ \left( \begin{array}{cc} 0 & 1\\ 1 & 0 \end{array} \right), \left( \begin{array}{cc} 0 & -i\\ i & 0 \end{array} \right), \left( \begin{array}{cc} 1 & 0\\ 0 & -1 \end{array} \right) \right]. \end{equation} Moreover, $\wektor{p}_{i}=\hbar\wektor{k}_{i}$ is a momentum of the incident particle and $k$ denotes the Dirac wave number and is given by \begin{equation}\label{DiracWaveNumber} k=\mathrm{sgn}(E)\sqrt{\frac{E^{2}-\left(mc^{2}\right)^{2}}{c^{2}\hbar^{2}}}. \end{equation} Thereafter, we shall consider only Hermitian potentials, i.e., those with kernels obeying $\mathsf{V}(\wektor{r},\wektor{r}')=\mathsf{V}^{\dagger}(\wektor{r}',\wektor{r})$. For this scattering process we may write the Lippmann-Schwinger equation \cite{LippSchw} of the form \begin{eqnarray}\label{I.4} &&\hspace{-1.4cm}\psi(\wektor{r})=\phi_{i}(\wektor{r})\nonumber\\ &&\hspace{-1cm}-\calkaob{r}'\calkaob{r}''\:G(E,\wektor{r},\wektor{r}')\mathsf{V}(\wektor{r}',\wektor{r}'') \psi(\wektor{r}''). \end{eqnarray} Function $G(E,\wektor{r},\wektor{r}')$ appearing above is the relativistic free--particle outgoing Green function given by \begin{equation}\label{I.6} G(E,\wektor{r},\wektor{r}')= \frac{1}{4\pi c^{2}\hbar^{2}}\left( -ic\hbar\wektor{\alpha}\cdot\wektor{\nabla}+\beta mc^{2}+E\jed_{4} \right) \frac{e^{ik|\wektor{r}-\wektor{r}'|}}{|\wektor{r}-\wektor{r}'|}, \end{equation} and formally is a kernel of the relativistic outgoing Green operator defined as \begin{equation}\label{I.7} \hat{G}(E)=\lim_{\epsilon\downarrow 0}[\hat{\mathcal{H}}_{0}-E-i\epsilon]^{-1}, \end{equation} with $\hat{\mathcal{H}}_{0}=-ic\hbar\wektor{\alpha}\cdot\wektor{\nabla}+ \beta mc^{2}$ being a Dirac free--particle Hamiltonian. Here \begin{equation}\label{I.5} \wektor{\alpha}= \left( \begin{array}{cc} 0 & \wektor{\sigma}\\ \wektor{\sigma} & 0 \end{array} \right), \qquad \beta= \left( \begin{array}{cc} \jed_{2} & 0\\ 0 & -\jed_{2} \end{array} \right),\qquad \jed_{2}=\left( \begin{array}{cc} 1 & 0\\ 0 & 1 \end{array} \right) \end{equation} and $\jed_{4}=\jed_{2}\ot\jed_{2}$. It is worth noticing that within the relativistic regime the Green function (\ref{I.6}) is a $4\times 4$ matrix. For purposes of further analysis, it is useful to introduce the following projector: \begin{equation}\label{I.8} \mathcal{P}(\wektor{k})=\frac{c\hbar \wektor{\alpha}\cdot\wektor{k}+\beta mc^{2}+E\jed_{4}}{2E}, \end{equation} which, as one can immediately infer, may be decomposed in the following way \begin{eqnarray}\label{I.9} \mathcal{P}(\wektor{k})&=&\Theta_{+}(\wektor{k}) \Theta_{+}^{\dagger}(\wektor{k})+\Theta_{-}(\wektor{k})\Theta_{-}^{\dagger}(\wektor{k})\nonumber\\ &=&\frac{1}{1+\varepsilon^{2}} \left( \begin{array}{cc} \mathbbm{1}_{2} & \varepsilon \wektor{\sigma}\cdot\wersor{k}\\ \varepsilon \wektor{\sigma}\cdot\wersor{k} & \varepsilon^{2}\mathbbm{1}_{2}, \end{array} \right) \end{eqnarray} with $\Theta_{\pm}(\wektor{k})$ being defined as \begin{equation}\label{I.10} \Theta_{\pm}(\wektor{k})= \frac{1}{\sqrt{1+\varepsilon^{2}}}\left( \begin{array}{c} \theta_{\pm}\\ \varepsilon\wektor{\sigma}\cdot\wersor{k}\,\theta_{\pm} \end{array} \right). \end{equation} Spinors $\theta_{\pm}$ constitute an arbitrary orthonormal basis in $\mathbb{C}^{2}$, i.e., $\theta_{s}^{\dagger}\theta_{t}=\delta_{st}$ $(s,t=-,+)$ and $\sum_{s=-}^{+}\theta_{s}\theta_{s}^{\dagger}=\jed_{2}$. What is important for further considerations, the matrix (\ref{I.8}) possesses the obvious property that $\mathcal{P}(\wektor{k})\Theta_{\pm}(\wektor{k})=\Theta_{\pm}(\wektor{k})$ and therefore \begin{equation}\label{I.11} \mathcal{P}(\wektor{k}_{i})U_{i}(\wektor{k}_{i})=U_{i}(\wektor{k}_{i}). \end{equation} We shall be exploiting this property in later analysis. Considering scattering processes we usually tend to find expressions for a scattering amplitude and various cross sections. To this aim we need to find an asymptotic behavior of the relativistic outgoing Green function. From Eq. (\ref{I.6}), using the projector (\ref{I.8}), we have \begin{equation}\label{I.12} G(E,\wektor{r},\wektor{r}')\stackrel{r\to\infty}{\sim} \frac{E}{2\pi c^{2}\hbar^{2}} \mathcal{P}(\wektor{k}_{f})\frac{e^{ikr}}{r} e^{-i\wektor{k}_{f}\cdot\wektor{r}'}, \end{equation} where $\wektor{k}_{f}=k\wektor{r}/r$ is a wave vector of the scattered particle. Notice that due to the fact that we deal with elastic processes $|\wektor{k}_{i}|=|\wektor{k}_{f}|=k$. After application of Eq. (\ref{I.12}) to Eq. (\ref{I.4}), we obtain \begin{equation}\label{I.13} \psi(\wektor{r})\stackrel{r\to\infty} {\sim}\underset{r\to\infty}{\mathrm{asymp}}\, \phi_{i}(\wektor{r}) +\Amplituda\frac{e^{ikr}}{r}, \end{equation} where $\Amplituda$ is the bispinor scattering amplitude and is defined through the relation \begin{eqnarray}\label{I.14} &&\hspace{-1.5cm}\Amplituda=-\frac{E}{2\pi c^{2}\hbar^{2}}\mathcal{P}(\wektor{k}_{f})\nonumber\\ &&\hspace{-1cm}\times\calkaob{r}'\calkaob{r}''\,e^{-i\wektor{k}_{f}\cdot\wektor{r}'} \mathsf{V}(\wektor{r}',\wektor{r}'')\psi(\wektor{r}'') \end{eqnarray} and, in general, is of the form \begin{equation}\label{I.15} \Amplituda=\frac{1}{\sqrt{1+\varepsilon^{2}}}\left( \begin{array}{c} \chi_{f}\\ \varepsilon\wektor{\sigma}\cdot\wersor{k}_{f}\chi_{f} \end{array} \right). \end{equation} Here $\chi_{f}$ is a spinor transformed from the initial spinor $\chi_{i}$ by the scattering process. Vector $\wektor{\nu}_{f}=(\chi_{f}^{\dagger}\wektor{\sigma}\chi_{f})/(\chi_{f}^{\dagger}\chi_{f})$ responds for an orientation of the spin of the scattered particle. Therefore let us assume that there exist a matrix, such that $\chi_{f}=\Amplitudaaa\chi_{i}$. Then it is easy to verify that the bispinor scattering amplitude may be written in the form \begin{equation}\label{I.16} \Amplituda=\Amplitudaa U_{i}(\wektor{k}_{i}), \end{equation} where the matrix $\Amplitudaa$ is related to $\Amplitudaaa$ by \begin{equation}\label{I.17} \Amplitudaa=\frac{1}{1+\varepsilon^{2}} \left( \begin{array}{cc} \Amplitudaaa &\quad \varepsilon\Amplitudaaa\wektor{\sigma}\cdot\wersor{k}_{i}\\ \varepsilon\wektor{\sigma}\cdot\wersor{k}_{f}\Amplitudaaa &\quad \varepsilon^{2}\wektor{\sigma}\cdot\wersor{k}_{f}\Amplitudaaa\wektor{\sigma}\cdot\wersor{k}_{i} \end{array} \right). \end{equation} Henceforth matrices $\Amplitudaa$ and $\Amplitudaaa$ will be called the matrix scattering amplitudes. The differential cross section for scattering from the direction $\wektor{k}_{i}$ and the spin arrangement $\wektor{\nu_{i}}$ onto $\wektor{k_{f}}$ and $\wektor{\nu_{f}}$ is defined as \begin{equation}\label{I.18} \frac{\mathrm{d}\sigma}{\mathrm{d}\Omega_{f}} =\chi_{f}^{\dag}\chi_{f}=\chi_{i}^{\dagger} \Amplitudaaa^{\dagger}\Amplitudaaa\chi_{i}, \end{equation} Subsequently, after integration the above over all the directions of $\wektor{k}_{f}$, we arrive at the total cross section \begin{equation}\label{I.19} \sigma(\wektor{k}_{i},\wektor{\nu}_{i})= \calkapow{k}_{f}\:\chi_{f}^{\dag}\chi_{f}. \end{equation} Finally, averaging over all directions of incidence $\hat{\wektor{k}}_{i}$ and the initial spin orientation $\hat{\wektor{\nu}}_{i}$, one finds the average total cross section \begin{equation}\label{I.20} \sigma_{t}(E)= \frac{1}{(4\pi)^{2}}\calkapow{k}_{i}\calkapow{\nu}_{i}\calkapow{k}_{f} \:\chi_{f}^{\dag}\chi_{f}. \end{equation} Obviously all the mentioned cross sections may be expressed in terms of all the scattering amplitudes $\Amplituda$, $\Amplitudaa$, and $\Amplitudaaa$. \section{Special class of non--local separable potentials} \setcounter{equation}{0} \noindent\indent In this section we employ the above considerations to the special class of non--local separable potentials. As previously mentioned, such a class of potentials allows to find solutions to the Lippmann--Schwinger equations in an analytical way. Consider the following class of potential kernels: \begin{equation}\label{II.1} \mathsf{V}(\wektor{r},\wektor{r}')=\sum_{\mu}\omega_{\mu} \mathsf{u}_{\mu}(\wektor{r})\mathsf{u}_{\mu}^{\dag}(\wektor{r}') \end{equation} where it is assumed that in general $\mu$ may denote the arbitrary finite set of indices, i.e., $\mu=\{\mu_{1},\ldots,\mu_{k}\}$ and all the coefficients $\omega_{\mu}$ different from zero. Functions $\mathsf{u}_{\mu}(\wektor{r})$ are assumed to be four--element columns. Substitution of Eq. (\ref{II.1}) to Eq. (\ref{I.4}) leads us to the Lippmann--Schwinger equation for the separable potentials: \begin{eqnarray}\label{II.2} &&\hspace{-0.2cm}\psi(\wektor{r})=\phi_{i}(\wektor{r})-\sum_{\mu}\omega_{\mu} \calkaob{r}'\,G(E,\wektor{r},\wektor{r}')\mathsf{u}_{\mu}(\wektor{r}') \nonumber\\ &&\hspace{1cm}\times\calkaob{r}''\,\mathsf{u}_{\mu}^{\dag}(\wektor{r}'')\psi(\wektor{r}''), \end{eqnarray} which may be equivalently rewritten as a set of linear algebraic equations. Indeed, using the Dirac notation one finds \begin{equation}\label{II.3} \sum_{\mu} \left[ \delta_{\nu\mu}+ \avg{\mathsf{u}_{\nu}}{\hat{G}(E)}{\mathsf{u}_{\mu}} \omega_{\mu} \right] \braket{\mathsf{u}_{\mu}}{\psi}=\braket{\mathsf{u}_{\nu}}{\phi_{i}}. \end{equation} For further convenience we introduce the following notations \begin{eqnarray}\label{II.4} &&\big<\mathsf{\mathsf{u}}\big|\varphi\big>= \left( \begin{array}{c} \big<\mathsf{u}_{1}\big|\varphi\big>\\*[0.5ex] \big<\mathsf{u}_{2}\big|\varphi\big>\\ \vdots \end{array} \right),\nonumber\\ &&\big<\varphi\big|\mathsf{\mathsf{u}}\big>=\big<\mathsf{u}\big|\varphi\big>^{\dagger}= \left( \big<\varphi\big|\mathsf{u}_{1}\big>\; \big<\varphi\big|\mathsf{u}_{2}\big>\; \dots \right). \end{eqnarray} Consequently, we may rewrite Eq. (\ref{II.3}) as a matrix equation $(\mathbbm{1}+\mathsf{G}\mathsf{\Omega}) \big<\mathsf{u}\big|\psi\big>=\big<\mathsf{u}\big|\phi_{i}\big>$ or equivalently as \begin{equation}\label{II.6} \big<\mathsf{u}\big|\psi\big>=(\mathbbm{1}+\mathsf{G}\mathsf{\Omega})^{-1} \big<\mathsf{u}\big|\phi_{i}\big>, \end{equation} with $\mathsf{G}$ being a matrix composed of the elements $\avg{\mathsf{u}_{\nu}}{\hat{G}(E)}{\mathsf{u}_{\mu}}$ and $\sf{\Omega}=\mathrm{diag}[\omega_{\mu}]$. Similarly, substituting Eq. (\ref{II.1}) to Eq. (\ref{I.14}) and again using Eq. (\ref{I.8}), we arrive at the bispinor scattering amplitude for the separable potentials in the form \begin{eqnarray}\label{II.7} &&\hspace{-1.5cm}\Amplituda=\frac{-E}{2\pi c^{2}\hbar^{2}}\mathcal{P}(\wektor{k}_{f})\sum_{\mu}\omega_{\mu}\calkaob{r}\:e^{-i\wektor{k}_{f}\cdot\wektor{r}} \mathsf{u}_{\mu}(\wektor{r})\nonumber\\ &&\times \calkaob{r}'\:\mathsf{u}_{\mu}^{\dag}(\wektor{r}')\psi(\wektor{r}'), \end{eqnarray} which, by virtue of Eqs. (\ref{I.9}) and (\ref{II.6}), reduces to \begin{equation}\label{II.8} \Amplituda=\frac{-E}{2\pi c^{2}\hbar^{2}} \sum_{s=-}^{+}\Theta_{s}(\wektor{k}_{f})\braket{\wektor{k}_{f}\theta_{s}}{\mathsf{u}} \mathsf{\Omega}\left(\mathbbm{1}+\mathsf{G}\mathsf{\Omega}\right)^{-1} \big<\mathsf{u}\big|\phi_{i}\big> \end{equation} and, utilizing the fact that for all invertible matrices $\mathsf{X}$ and $\mathsf{Y}$ the relation $(\mathsf{X}\mathsf{Y})^{-1}=\mathsf{Y}^{-1}\mathsf{X}^{-1}$ is satisfied, finally to \begin{equation}\label{II.9} \Amplituda=\frac{-E}{2\pi c^{2}\hbar^{2}} \sum_{s=-}^{+}\Theta_{s}(\wektor{k}_{f})\braket{\wektor{k}_{f}\theta_{s}}{\mathsf{u}} \left(\mathsf{\Omega}^{-1}+\mathsf{G}\right)^{-1} \big<\mathsf{u}\big|\phi_{i}\big>. \end{equation} Subsequently, using the fact that (\ref{I.11}), we obtain the bispinor scattering amplitude in the following form \begin{eqnarray}\label{II.10} &&\hspace{-1.02cm}\Amplituda=\frac{-E}{2\pi c^{2}\hbar^{2}}\sum_{s,t=-}^{+}\Theta_{s}(\wektor{k}_{f})\braket{\wektor{k}_{f}\theta_{s}}{\mathsf{u}}\nonumber\\ &&\times \left(\mathsf{\Omega}^{-1}+\mathsf{G}\right)^{-1} \braket{\mathsf{u}}{\wektor{k}_{i}\theta_{t}}\Theta_{t}^{\dagger}(\wektor{k}_{i}) U_{i}(\wektor{k}_{i}), \end{eqnarray} which, after comparison with Eq. (\ref{I.16}), gives the formulae for the $4\times4$ matrix scattering amplitude: \begin{eqnarray}\label{II.11} &&\hspace{-2.02cm}\Amplitudaa=\frac{-E}{2\pi c^{2}\hbar^{2}} \sum_{s,t=-}^{+}\Theta_{s}(\wektor{k}_{f})\braket{\wektor{k}_{f}\theta_{s}}{\mathsf{u}}\nonumber\\ &&\hspace{-1cm}\times \left(\mathsf{\Omega}^{-1}+\mathsf{G}\right)^{-1} \braket{\mathsf{u}}{\wektor{k}_{i}\theta_{t}}\Theta_{t}^{\dagger}(\wektor{k}_{i}), \end{eqnarray} and finally, after straightforward movements, for the $2\times 2$ matrix scattering amplitude as \begin{equation}\label{II.12} \Amplitudaaa=\frac{-E}{2\pi c^{2}\hbar^{2}} \sum_{s,t=-}^{+}\theta_{s} \braket{\wektor{k}_{f}\theta_{s}}{\mathsf{u}} \left(\mathsf{\Omega}^{-1}+\mathsf{G}\right)^{-1} \braket{\mathsf{u}}{\wektor{k}_{i}\theta_{t}}\theta_{t}^{\dagger}. \end{equation} \section{The eigenchannel method}\label{III} \setcounter{equation}{0} \noindent\indent Now we are in position to apply the eigenchannel method proposed recently by Szmytkowski \cite{Szmytkowski} to scattering of the Dirac particles from potentials of the form (\ref{II.1}). As we shall see below, such a class of potentials allows us to formulate this method in a simplified algebraic form. We start from the decomposition of the matrix $\mathsf{\Omega}^{-1}+\mathsf{G}_{\mathrm{D}}$ into its Hermitian and non-Hermitian parts, i.e., \begin{equation}\label{III.1} \mathsf{\Omega}^{-1}+\mathsf{G}=\mathsf{A}+i\mathsf{B}, \end{equation} where matrices $\sf{A}$ and $\sf{B}$ are defined through relations \begin{equation}\label{III.2} \mathsf{A}=\mathsf{\Omega}^{-1}+\frac{1}{2} \left(\mathsf{G}+\mathsf{G}^{\dagger}\right),\qquad \mathsf{B}=\frac{1}{2i}\left(\mathsf{G}- \mathsf{G}^{\dagger}\right). \end{equation} It is evident from these definitions that both matrices $\sf{A}$ and $\sf{B}$ are Hermitian. Moreover, utilizing the fact that \begin{equation}\label{III.4} \wektor{\nabla}\frac{\mathrm{e}^{ik|\wektor{r}-\wektor{r}'|}}{|\wektor{r}-\wektor{r}'|}= \frac{\wektor{r}-\wektor{r}'}{|\wektor{r}-\wektor{r}'|}\left(\frac{ik\mathrm{e}^{ik|\wektor{r}-\wektor{r}'|}} {|\wektor{r}-\wektor{r}'|}-\frac{\mathrm{e}^{ik|\wektor{r}-\wektor{r}'|}}{|\wektor{r}-\wektor{r}'|^{2}}\right) \end{equation} where $\wektor{\varrho}=\wektor{r}-\wektor{r}'$, the straightforward calculations lead us to their matrix elements of the form \begin{eqnarray}\label{III.6} &&\mathsf{A}_{\nu\mu}= \omega^{-1}_{\nu}\delta_{\nu\mu}-\frac{k}{4\pi c^{2}\hbar^{2}} \calkaob{r}\calkaob{r}'\mathsf{u}_{\nu}^{\dagger}(\wektor{r})\nonumber\\ &&\times\left[ ic\hbar k\wektor{\alpha}\cdot\frac{\wektor{\varrho}}{|\wektor{\varrho}|} y_{1}(k|\wektor{\varrho}|)+(\beta mc^{2}+E)y_{0}(k|\wektor{\varrho}|) \right]\mathsf{u}_{\mu}(\wektor{r}')\nonumber\\ \end{eqnarray} and \begin{eqnarray}\label{III.7} &&\mathsf{B}_{\nu\mu}=\frac{k}{4\pi c^{2}\hbar^{2}} \calkaob{r}\calkaob{r}'\mathsf{u}_{\nu}^{\dagger}(\wektor{r})\nonumber\\ &&\times\left[ic\hbar k\wektor{\alpha}\cdot\frac{\wektor{\varrho}}{|\wektor{\varrho}|} j_{1}(k|\wektor{\varrho}|)+(\beta mc^{2}+E)j_{0}(k|\wektor{\varrho}|) \right]\mathsf{u}_{\mu}(\wektor{r}'),\nonumber\\ \end{eqnarray} where $j_{0}(z)$, $j_{1}(z)$, $y_{0}(z)$ and $y_{1}(z)$ are, respectively, the Bessel and Neumann spherical functions \cite{AbrStegun}. Recall that in general $j_{0}(z)=(\sin z)/z$, $y_{0}(z)=(\cos z)/z$ and \begin{equation} j_{1}(z)=-\frac{\cos z}{z}+\frac{\sin z}{z^{2}}, \quad y_{1}(z)=-\frac{\sin z}{z}-\frac{\cos z}{z^{2}}. \end{equation} The main idea of the present paper, adopted from \cite{Szmytkowski}, is to construct the following weighted spectral problem: \begin{equation}\label{III.8} \mathsf{A}X_{\gamma}(E)=\lambda_{\gamma}(E)\mathsf{B} X_{\gamma}(E), \end{equation} where $X_{\gamma}(E)$ and $\lambda_{\gamma}(E)$ is, respectively, an eigenvector and an eigenvalue. Thereafter the eigenvectors $\{X_{\gamma}(E)\}$ will be called {\it eigenchannel vectors}. They are directly related to eigenchannels defined in \cite{Szmytkowski} as state vectors. In fact, they constitute a projection of eigenchannels onto subspace spanned by $\mac{u}_{\mu}(\wektor{r})$. Using the fact that matrices $\mathsf{A}$ and $\mathsf{B}$ are Hermitian and, as it is proven in Appendix \ref{AppA}, the matrix $\mathsf{B}$ is positive semi-definite, one finds that the eigenvalues $\{\lambda_{\gamma}(E)\}$ are real, i.e., $\lambda_{\gamma}^{*}(E)=\lambda_{\gamma}(E)$. Moreover, the eigenchannels associated with different eigenvalues obey the orthogonality relation \begin{equation}\label{III.9} X_{\gamma'}^{\dagger}(E)\mathsf{B}X_{\gamma}(E)=0 \qquad (\lambda_{\gamma'}(E)\neq\lambda_{\gamma}(E)). \end{equation} In case of degeneration of some eigenvalues one may always choose the corresponding eigenvectors to be orthogonal according to the above relation. Then, imposing the normalization $X_{\gamma}^{\dagger}(E)\mathsf{B}X_{\gamma}(E)=1$, one obtains the following orthonormality relation \begin{equation}\label{III.10} X_{\gamma'}^{\dagger}(E)\mathsf{B}X_{\gamma}(E)=\delta_{\gamma'\gamma}. \end{equation} From Eqs. (\ref{III.8}) and (\ref{III.10}) one infers that the eigenvalues $\{\lambda_{\gamma}(E)\}$ may be related to the matrix $\mathsf{A}$ as follows \begin{equation}\label{III.11} \lambda_{\gamma}(E)=X_{\gamma}^{\dagger}(E)\mathsf{A}X_{\gamma}(E). \end{equation} Similar reasoning may be carried out employing the matrices $\mathsf{A}$ and $\mathsf{\Omega}^{-1}+\mathsf{G}$. Indeed, after algebraic manipulations we arrive at \begin{eqnarray}\label{III.12} &&X_{\gamma'}^{\dagger}(E)\mathsf{A}X_{\gamma}(E)= \lambda_{\gamma}(E)\delta_{\gamma'\gamma},\nonumber\\ &&\hspace{-1cm}X_{\gamma'}^{\dagger}(E)(\mathsf{\Omega}^{-1}+\mathsf{G})X_{\gamma}(E)= [i+\lambda_{\gamma}(E)]\delta_{\gamma'\gamma}, \end{eqnarray} and $\lambda_{\gamma}(E)=X_{\gamma}^{\dagger}(E)(\mathsf{\Omega}^{-1}+\mathsf{G})X_{\gamma}(E)-i$. Since the eigenchannels $\{X_{\gamma}(E)\}$ are the solutions of the Hermitian eigenvalue problem, they may satisfy the following closure relations \begin{eqnarray}\label{III.14} &\displaystyle \sum_{\gamma}X_{\gamma}(E)X_{\gamma}^{\dagger}(E)\mathsf{B}=\jed,&\nonumber\\ &\displaystyle \sum_{\gamma}\lambda_{\gamma}^{-1}(E) X_{\gamma}(E)X_{\gamma}^{\dagger}(E)\mathsf{A}=\jed,& \end{eqnarray} and \begin{equation}\label{III.15} \sum_{\gamma} \frac{1}{i+\lambda_{\gamma}(E)}X_{\gamma}(E)X_{\gamma}^{\dagger}(E) (\mathsf{\Omega}^{-1}+\mathsf{G})=\jed, \end{equation} where $\jed$ is an identity matrix, which dimension depends on the dimension of the matrix $\mathsf{G}$. For purposes of further analyzes the above closure relations are assumed to hold. Below, we employ the above reasoning to the derivation of the scattering amplitudes. From Eq. (\ref{III.15}) one deduces that \begin{equation}\label{III.16} (\mathsf{\Omega}^{-1}+\mathsf{G})^{-1}=\sum_{\gamma} \frac{1}{i+\lambda_{\gamma}(E)} X_{\gamma}(E)X_{\gamma}^{\dagger}(E). \end{equation} After substitution of Eq. (\ref{III.16}) to Eq. (\ref{II.11}) and rearranging terms, we have \begin{eqnarray}\label{III.17} &&\hspace{-0.5cm}\Amplitudaa=\frac{-E}{2\pi c^{2}\hbar^{2}} \sum_{\gamma}\frac{1}{i+\lambda_{\gamma}(E)}\sum_{s=-}^{+}\Theta_{s}(\wektor{k}_{f}) \braket{\wektor{k}_{f}\theta_{s}}{\mathsf{u}}X_{\gamma}(E)\nonumber\\ &&\hspace{0.5cm}\times \sum_{t=-}^{+}X_{\gamma}^{\dagger}(E) \braket{\mathsf{u}}{\wektor{k}_{i}\theta_{t}}\Theta_{t}^{\dagger}(\wektor{k}_{i}). \end{eqnarray} Let us define the following angular functions \begin{equation}\label{III.18} \mathcal{Y}_{\gamma}(\wektor{k})=\sqrt{\frac{Ek}{8\pi^{2}c^{2}\hbar^{2}}} \sum_{s=-}^{+}\Theta_{s}(\wektor{k})\braket{\wektor{k}\theta_{s}}{\mathsf{u}} X_{\gamma}(E), \end{equation} hereafter termed the {\it eigenchannel bispinor harmonics}. The functions $\{\mathcal{Y}_{\gamma}(\wektor{k})\}$ are orthonormal on the unit sphere (for proof, see Appendix \ref{AppB}), i.e., \begin{equation}\label{III.19} \calkapow{k}\,\mathcal{Y}_{\gamma'}^{\dagger}(\wektor{k})\mathcal{Y}_{\gamma}(\wektor{k}) =\delta_{\gamma'\gamma}. \end{equation} Application of Eq. (\ref{III.18}) to Eq. (\ref{III.17}) yields \begin{equation}\label{III.20} \Amplitudaa=\frac{4\pi}{k}\sum_{\gamma}e^{i\delta_{\gamma}(E)}\sin\delta_{\gamma}(E) \mathcal{Y}_{\gamma}(\wektor{k}_{f})\mathcal{Y}_{\gamma}^{\dagger}(\wektor{k}_{i}), \end{equation} where $\{\delta_{\gamma}(E)\}$ are called {\it eigenphase-shifts} and are related to $\{\lambda_{\gamma}(E)\}$ according to \begin{equation}\label{III.21} \lambda_{\gamma}(E)=-\cot\delta_{\gamma}(E). \end{equation} Similar considerations may be carried out for the $2\times 2$ matrix scattering amplitude $\Amplitudaaa$. Indeed, in virtue of Eq. (\ref{I.17}) we may rewrite Eq. (\ref{II.12}) in the form \begin{equation}\label{III.22} \Amplitudaaa=\frac{4\pi}{k}\sum_{\gamma}e^{i\delta_{\gamma}(E)} \sin\delta_{\gamma}(E)\Upsilon_{\gamma}(\wektor{k}_{f})\Upsilon_{\gamma}^{\dagger}(\wektor{k}_{i}), \end{equation} where the angular functions $\{\Upsilon_{\gamma}(\wektor{k})\}$, hereafter called {\it eigenchannel spinor harmonics}, are defined as follows \begin{equation}\label{III.23} \Upsilon_{\gamma}(\wektor{k})=\sqrt{\frac{Ek}{8\pi^{2}c^{2}\hbar^{2}}} \sum_{s=-}^{+}\theta_{s}\braket{\wektor{k}\theta_{s}}{\mathsf{u}} X_{\gamma}(E). \end{equation} Moreover, they are orthogonal on the unit sphere (for proof, see Appendix \ref{AppB}) \begin{equation}\label{III.24} \calkapow{k}\,\Upsilon_{\gamma'}^{\dagger}(\wektor{k})\Upsilon_{\gamma}(\wektor{k})=\delta_{\gamma'\gamma}, \end{equation} and, as one can verify, are related to the eigenchannel bispinor harmonics $\{\mathcal{Y}_{\gamma}(\wektor{k})\}$ via the relation \begin{equation}\label{III.25} \mathcal{Y}_{\gamma}(\wektor{k})=\frac{1}{\sqrt{1+\varepsilon^{2}}} \left( \begin{array}{c} \Upsilon_{\gamma}(\wektor{k})\\ \varepsilon\wektor{\sigma}\cdot\wersor{k}\Upsilon_{\gamma}(\wektor{k}) \end{array} \right). \end{equation} Now we are in position to compute scattering cross-sections. Substitution of Eq. (\ref{III.22}) to Eq. (\ref{I.18}) and integration over all directions of scattering $\hat{\wektor{k}}_{f}$, by virtue of relation (\ref{III.24}), yields \begin{equation}\label{III.26} \sigma(\wektor{k}_{i},\wektor{\nu}_{i})=\frac{16\pi^{2}}{k^{2}} \sum_{\gamma}\sin^{2}\delta_{\gamma}(E)\left|\chi_{i}^{\dagger}\Upsilon_{\gamma}(\wektor{k}_{i})\right|^{2}. \end{equation} To compute the total cross-section averaged over all arrangements of spin of the incident particle, we have to notice that the projector onto the pure state $\chi_{i}$ may be written as $\chi_{i}\chi_{i}^{\dagger}=(1/2)[\jed_{2}+\wektor{\nu}_{i}\cdot\wektor{\sigma}]$ with $|\wektor{\nu}_{i}|=1$. Therefore, substituting of the above to Eq. (\ref{III.26}) and averaging over all directions of $\boldsymbol{\nu}_{i}$, we arrive at \begin{equation}\label{III.28} \sigma(\wektor{k}_{i})= \frac{8\pi^{2}}{k^{2}}\sum_{\gamma}\sin^{2}\delta_{\gamma}(E)\Upsilon_{\gamma}^{\dagger}(\wektor{k}_{i}) \Upsilon_{\gamma}(\wektor{k}_{i}). \end{equation} Finally, averaging the above scattering cross-section over all directions of incidence $\hat{\wektor{k}}_{i}$, again by virtue of Eq. (\ref{III.24}), we get the total cross-section in the form \begin{equation}\label{III.29} \sigma_{t}(E)=\frac{2\pi}{k^{2}}\sum_{\gamma}\sin^{2}\delta_{\gamma}(E). \end{equation} It should be emphasized that all the above considerations respecting scattering cross-sections may be repeated using the eigenchannel bispinor harmonics $\{\mathcal{Y}_{\gamma}(\wektor{k})\}$ instead of the eigenchannel spinor harmonics $\{\Upsilon_{\gamma}(\wektor{k})\}$. The significant difference is that then the integrals over $\hat{\wektor{k}}_{f}$ and $\hat{\wektor{k}}_{i}$ need to be calculated using relation (\ref{III.19}) instead of (\ref{III.24}). \section{Example} \setcounter{equation}{0} We conclude our considerations providing here an illustrative example concerning the scattering from a spherical shell of radius $R$, centered at the origin of the coordinate system. Due to the assumption of non-locality of potentials under consideration, we shall simulate this process by using a potential of the form \begin{equation}\label{Ex1} \mac{V}(\wektor{r},\wektor{r}')=\omega v(\wektor{r})v(\wektor{r}')\jed_{4},\qquad v(\wektor{r})=\frac{1}{\sqrt{4\pi}}\frac{\delta(r-R)}{R^{2}}, \end{equation} where $\omega\neq 0$. Notice that the potential defined above is the special case of that proposed recently by de Prunel\'e \cite{Prunele} (see also \cite{Prunele1}). Scattering of the Dirac particles from delta-like potentials was also studied e.g. in Refs. \cite{Dombey,Loewe}. However, in these papers the authors considered only local potentials and not non-local ones. At the very beginning, we need to bring the potential (\ref{Ex1}) to the previously postulated form (\ref{II.1}). To this aim, let $\mac{e}_{1}$ and $\mac{e}_{2}$ constitute a standard basis in $\mathbb{C}^{2}$, i.e., $\mac{e}_{1}=(1\;0)^{T}$ and $\mac{e}_{2}=(0\;1)^{T}$. Moreover, let $\mac{e}_{ij}=\mac{e}_{i}\otimes\mac{e}_{j}$ and then by virtue of the fact that $\jed_{4}=\sum_{i,j=1}^{2}\mac{e}_{ij}\mac{e}_{ij}^{\dagger}$, we may rewrite (\ref{Ex1}) as \begin{equation}\label{Ex2} \mac{V}(\wektor{r},\wektor{r}')=\omega\sum_{i,j=1}^{2}\mac{u}_{ij}(\wektor{r})\mac{u}_{ij}^{\dagger}(\wektor{r}),\qquad \mac{u}_{ij}(\wektor{r})=v(\wektor{r})\mac{e}_{ij}. \end{equation} Now, we are in position to compute the matrix $\mac{G}$. Using Eqs. (\ref{I.6}) and (\ref{Ex1}), after straightforward integrations we have \begin{eqnarray}\label{Ex3} \mac{G}=ikj_{0}(kR)h_{0}^{(+)}(kR) \left( \begin{array}{cc} \eta_{+}\jed_{2} & 0\\ 0 & \eta_{-}\jed_{2} \end{array} \right), \end{eqnarray} where $\eta_{\pm}=(E\pm mc^{2})/c^{2}\hbar^{2}$ and $h_{0}^{(+)}(z)=j_{0}(z)+iy_{0}(z)$ is the spherical Hankel function of the first kind. Hence, by the definitions given in Eq. (\ref{III.2}), we find that the explicit forms of matrices $\mac{A}$ and $\mac{B}$ are \begin{equation}\label{Ex4} \mac{A}=\left( \begin{array}{cc} [\omega^{-1}-kj_{0}(kR)y_{0}(kR)\eta_{+}]\jed_{2} &\hspace{-1.7cm} 0\\ 0 & \hspace{-1.7cm} [\omega^{-1}-kj_{0}(kR)y_{0}(kR)\eta_{-}]\jed_{2} \end{array} \right) \end{equation} and \begin{equation}\label{Ex5} \mac{B}=kj_{0}^{2}(kR) \left( \begin{array}{cc} \eta_{+}\jed_{2} & 0\\ 0 & \eta_{-}\jed_{2} \end{array} \right). \end{equation} According to the method formulated in Sec. \ref{III}, we may construct the following spectral problem \begin{equation}\label{Ex6} \mac{A}X_{\gamma}(E)=\lambda_{\gamma}(E)\mac{B}X_{\gamma}(E)\qquad (\gamma=1,2,3,4), \end{equation} which, as one can easily verify, has two different eigenvalues \begin{equation}\label{Ex7} \lambda_{\pm}(E)=\frac{\omega^{-1}-kj_{0}(kR)y_{0}(kR)\eta_{\pm}}{kj_{0}^{2}(kR)\eta_{\pm}} \end{equation} and respective eigenvectors \begin{eqnarray}\label{Ex8} &\displaystyle X_{+}^{(1(2))}(E)=\frac{1}{\sqrt{k\eta_{+}}j_{0}(kR)}\,\mac{e}_{1}\otimes\mac{e}_{1(2)},&\nonumber\\ &\displaystyle X_{-}^{(1(2))}(E)=\frac{1}{\sqrt{k\eta_{-}}j_{0}(kR)} \,\mac{e}_{2}\otimes\mac{e}_{1(2)}.& \end{eqnarray} Then, using Eq. (\ref{III.18}) and by virtue of the fact that \begin{eqnarray}\label{Ex9} &&\hspace{-1.5cm}\braket{\wektor{k}\chi}{\mac{\bf{u}}}=\sqrt{\frac{4\pi}{1+\varepsilon^{2}}}j_{0}(kR) \nonumber\\ &&\hspace{-1cm}\times \left( \chi^{\dagger}\mac{e}_{1}\;\; \chi^{\dagger}\mac{e}_{2}\;\;\varepsilon\chi^{\dagger}\wektor{\sigma}\cdot\wersor{k}\, \mac{e}_{1}\;\; \varepsilon\chi^{\dagger}\wektor{\sigma}\cdot\wersor{k}\, \mac{e}_{2}\right), \end{eqnarray} we arrive at the four eigenchannel bispinor harmonics $\{\mathcal{Y}_{\gamma}(\wektor{k})\}$ in the form \begin{equation}\label{Ex10} \mathcal{Y}_{+}^{(1(2))}(\wektor{k})= \frac{1}{\sqrt{4\pi(1+\varepsilon^{2})}}\, \left( \begin{array}{c} \mac{e}_{1(2)}\\ \varepsilon\wektor{\sigma}\cdot\wersor{k}\,\mac{e}_{1(2)} \end{array} \right) \end{equation} and \begin{equation} \mathcal{Y}_{-}^{(1(2))}(\wektor{k})=\frac{1}{\sqrt{4\pi(1+\varepsilon^{2})}}\, \left( \begin{array}{c} \varepsilon\wektor{\sigma}\cdot\wersor{k}\,\mac{e}_{1(2)}\\ \mac{e}_{1(2)} \end{array} \right). \end{equation} Then, by virtue of Eq. (\ref{III.23}), one obtains the eigenchannel spinor harmonics $\{\Upsilon_{\gamma}(\wektor{k})\}$ in the form \begin{equation}\label{Ex11} \Upsilon_{+}^{(1(2))}(\wektor{k})= \frac{1}{\sqrt{4\pi}}\,\mac{e}_{1(2)},\quad \Upsilon_{-}^{(1(2))}(\wektor{k})=\frac{1}{\sqrt{4\pi}}\,\wektor{\sigma}\cdot\wersor{k}\,\mac{e}_{1(2)}. \end{equation} The latter may be equivalently obtained combining Eqs. (\ref{III.25}) and (\ref{Ex11}). Moreover, as one may easily verify, functions given by Eqs. (\ref{Ex10}) and (\ref{Ex11}) are orthonormal, respectively, in the sense (\ref{III.19}) and (\ref{III.24}). Before we find an expression for total cross section, we compute the scattering amplitude. Since, as shown in Sec. \ref{SecII}, the bispinor and both matrix scattering amplitudes are mutually related, we restrict our considerations to the $2\times 2$ scattering amplitude. Thus, combining Eqs. (\ref{III.22}), (\ref{Ex5}), and (\ref{Ex11}) we obtain \begin{eqnarray}\label{Ex12} &&\hspace{-0.9cm}\Amplitudaaa=-j_{0}^{2}(kR)\left[ \frac{\jed_{2}}{ik j_{0}(kR)h_{0}^{(+)}(kR)+(\omega\eta_{+})^{-1}}\right.\nonumber\\ &&\left.+\frac{(\wektor{\sigma}\cdot\wersor{k}_{f}) (\wektor{\sigma}\cdot\wersor{k}_{i})} {ikj_{0}(kR)h_{0}^{(+)}(kR)+(\omega\eta_{-})^{-1}} \right]. \end{eqnarray} Finally, substitution of Eqs. (\ref{Ex8}) and (\ref{Ex11}) to Eq. (\ref{III.26}) with the aid of Eq. (\ref{III.21}) yields \begin{eqnarray}\label{Ex13} &&\hspace{-1cm}\sigma(\wektor{k}_{i},\wektor{\nu}_{i})=\frac{4\pi}{k^{2}} j_{0}^{4}(kR)\nonumber\\ &&\hspace{-0.5cm}\times\left\{ \frac{1}{[(k\omega \eta_{+})^{-1}-j_{0}(kR)y_{0}(kR)]^{2}+j_{0}^{4}(kR)}\right.\nonumber\\ &&\hspace{-0.5cm}+\left.\frac{1}{[(k\omega\eta_{-})^{-1}-j_{0}(kR)y_{0}(kR)]^{2} +j_{0}^{4}(kR)}\right\}. \end{eqnarray} Here it is evident that $\sigma(\wektor{k}_{i},\wektor{\nu}_{i})=\sigma(\wektor{k}_{i})=\sigma_{t}(E)$. In order to illustrate the obtained results, the eigenphaseshifts for two different values of $\omega$, derived from Eqs. (\ref{III.21}) and (\ref{Ex7}), are plotted in Fig. 1 and 2. Figures 3 and 4 present partial $\sigma_{\pm}(E)$ as well as total $\sigma_{t}(E)$ cross sections. It seems interesting to investigate the behavior of both eigenvalues $\lambda_{\pm}(E)$ in the non-relativistic limit, i.e., when $c\to\infty$. From (\ref{DiracWaveNumber}) one concludes that \begin{equation} \eta_{+}\stackrel{c\to\infty}{\LRA}\frac{2m}{\hbar^{2}},\qquad \eta_{-}\stackrel{c\to\infty}{\LRA}0 \end{equation} and therefore \begin{equation} \lambda_{+}(E)\stackrel{c\to\infty}{\LRA}\frac{(\hbar^{2}/2m\omega)-kj_{0}(kR)y_{0}(kR)}{kj_{0}^{2}(kR)} \end{equation} and \begin{equation} \lambda_{-}(E)\stackrel{c\to\infty}{\LRA}\mathrm{sgn}(\omega)\infty. \end{equation} This means that $\delta_{-}(E)\to n\pi$ $(n\in \mathbb{Z})$ in the limit of $c\to\infty$. Therefore the cross section $\sigma_{-}(E)$ vanishes in the non-relativistic limit and in this sense it has a purely relativistic character leading to the fact that the resonance appearing in Fig. 4 at about $1.25 mc^{2}$ is purely relativistic effect. One sees that in the non-relativistic limit the cross section (\ref{Ex13}) reduces to \begin{eqnarray} &&\hspace{-1cm}\sigma_{t}(E)\stackrel{c\to\infty}{\LRA}\frac{4\pi}{k^{2}} j_{0}^{4}(kR)\nonumber\\ &&\hspace{-0.8cm}\times\left\{ \frac{1}{[(\hbar^{2}/2m k\omega)-j_{0}(kR)y_{0}(kR)]^{2}+j_{0}^{4}(kR)}\right\}. \end{eqnarray} The above cross section may also be obtained using non-relativistic formulation of the present method given in Ref. \cite{moja}. \begin{figure}[ht] \includegraphics[width=6cm]{fig1.eps} \caption{Behavior of eigenphaseshifts $\delta_{+}(E)$ (solid curve) and $\delta_{-}(E)$ (dashed curve) as functions of energy $E$ (in units of $mc^{2}$) for $\omega=-\hbar^{3}/m^{2}c$ and $R=\hbar/mc$. The eigenphaseshift $\delta_{+}(E)$ has been constrained to the range $[-\pi/2,\pi/2]$. } \end{figure} \begin{figure}[ht] \includegraphics[width=6cm]{fig2.eps} \caption{Behavior of eigenphaseshifts $\delta_{+}(E)$ (solid curve) and $\delta_{-}(E)$ (dashed curve) as functions of energy $E$ (in units of $mc^{2}$) for $\omega=-5\hbar^{3}/m^{2}c$ and $R=\hbar/mc$. Both eigenphaseshifts have been constrained to the range $[-\pi/2,\pi/2]$.} \end{figure} \begin{figure}[ht] \includegraphics[width=6cm]{fig3.eps} \caption{Partial $\sigma_{+}(E)$ (dashed curve), $\sigma_{-}(E)$ (dotted curve), and total $\sigma_{t}(E)$ (solid curve) cross sections (all in units of $R^{2}$) as functions of energy $E$ (in units of $mc^{2}$) for $\omega=-\hbar^{3}/m^{2}c$ and $R=\hbar/mc$.} \end{figure} \begin{figure}[ht] \includegraphics[width=6cm]{fig4.eps} \caption{Partial $\sigma_{+}(E)$ (dashed curve), $\sigma_{-}(E)$ (dotted curve), and total $\sigma_{t}(E)$ (solid curve) cross sections (all in units of $R^{2}$) as functions of energy $E$ (in units of $mc^{2}$) for $\omega=-5\hbar^{3}/m^{2}c$ and $R=\hbar/mc$. } \end{figure} \section{Conclusions} In this work, an application of the recently proposed eigenchannel method \cite{Szmytkowski} to the scattering of Dirac particles from non-local separable potentials has been presented. Application of such a particular case of the non-local potentials reduces naturally the general weighted eigenvalue problem stated in Ref. \cite{Szmytkowski} to its matrix counterpart given by Eq. (\ref{III.8}) leading to the definition of eigenchannel vectors. Using the notion of the eigenchannel vectors the definitions of eigenchannel spinor as well as bispinor harmonics have been given. The latter provide us with the formulas for scattering amplitudes similar to that well-known for central potentials generalizing them at the same time to the case of non-local separable potentials. The general considerations have been extended with an illustrative example in which the Dirac particles are scattered from non-local, delta-like potential. In this particular case, the general eigenvalue problem (\ref{III.8}) become just a $4\times 4$ matrix equation and therefore is easily solvable (notice that in the case of non-relativistic scattering it would be just a one-dimensional problem). The eigenvalues of this problem are two-fold degenerated and therefore give two different eigenphase-shifts from which one has a purely relativistic character in the sense that it tends to $n\pi$ $(n\in \mathbb{Z})$ whenever $c\to\infty$ giving no contribution to total cross sections in non-relativistic limit. One sees also that even such a simple example of non-local potentials may lead to some resonances (see Fig. 4). The next step in our considerations will be to investigate the applicability of the new formulation of the eigenchannel method in the case of inelastic scattering from separable potentials. Moreover it seems also interesting to investigate the applicability of the method to the other, more complicated examples of separable potentials. \section*{Acknowledgments} I am grateful to R.~Szmytkowski for very useful discussions, suggestions and commenting on the manuscript. Discussions with M.~Czachor are also acknowledged.
{ "timestamp": "2007-11-05T18:23:44", "yymm": "0503", "arxiv_id": "physics/0503196", "language": "en", "url": "https://arxiv.org/abs/physics/0503196" }
\section{\label{sec:level1}Introduction} Recent advances in quantum information science have shown that, on one hand, photons are ideal carriers of quantum information, and on the other hand, atoms represent reliable and long-lived storage and processing units. In recent years quantum light storage is one of the extensively studied tasks of quantum optics. Basic idea of light storage is electromagnetically induced transparency (EIT) \cite{EIT}. Electromagnetically induced transparency is a coherent interaction process in which a coupling laser field is used to made the optical dense media transparent for the probe field. Since its discovery, a number of new effects and techniques for light-matter interaction have appeared [2-6]. Most notably, from the point of view of the work presented here, particular attention has been devoted to ultraslow light propagation and light storage techniques [3-5]. The key concept of EIT is the dark state and population trapping \cite% {arrimon}. The dark state is a specific coherent superposition state which does not contain excited short-living atomic level due to\ destructive interference between two interaction paths. The dark-state is eigenstate of the light -- atom interaction Hamiltonian, so the atom prepared in a dark-state can not be excited and cannot leave the dark-state if the interaction is adiabatic (Fig.1). The population trapping via applying strong coupling leads to the adiabatic formation of the dark-state. Since the interaction is realized by the light pulses the infuence of nonadiabatic corrections may become important \cite{adiabatica}, \cite{Nasha} . This infuence has been studied e.g. in \cite{adiabatica}. In particular, the first nonadiabatic correction connects the dark state and the bright state, so the depletion of the bright state, because of optical pumping, can affect the dark state. Despite of large amount of experimental and theoretical papers concerning light storage (see \cite{revlukin}\ and citations there), and applications in quantum information science, the influence of decoherence level width on information carried by stored light is studied insufficiently. In this work we present theoretical study which discusses and explains influence of all relaxations on probe propagation both analytically and numerically (it is essential in especially, solid state systems \cite{solid}% ). The goal is to study comprehensively how the depletion of bright state will affect the pulse propagation in an EIT\ media and light storage in particular.\ By solving the coupled system of Maxwell and density matrix equations to the first order of the nonstationary pertrubation theory with respect to nonadiabaticity and decoherence we obtain analytical solution which completely describes the probe pulse propagation and is consistent with the recent light storage experiments. The paper is organized as follows. In section II the basic equations are written down and the probe pulse propagation equation is derived and analyzed. In section III and Appendixes the analytical solution for counterintuitive pulse switching order and for matched pulses are obtained and the asymptotic solutions discussed. Section IV deals with the physical consequences of the obtained solution, namely the necessary conditions of the pulse storage and retrieving, also the numerical results are demonstrated. In section V we consider the transverse relaxation of the coherence induced in the medium. Section VI concludes the paper. \section{\label{basic}Basic Equations} Figure 1 shows a schematic diagram of the atomic system in the EIT basis: media of three level atoms interacting with two laser pulses $% E_{p}=A_{p}\cos \left( \omega _{p}t-k_{p}z+\varphi _{p}\right) $ (probe)\ and $E_{c}=A_{c}\cos \left( \omega _{c}t-k_{c}z+\varphi _{c}\right) $\emph{\ }(coupling). The probe field resonantly connects the state $|1\rangle $ to the state $|3\rangle $ and the coupling field connects $|2\rangle $ to $% |3\rangle $. The Hamiltonian of the system in the rotating wave approximation is: \begin{equation*} H=\hbar \Delta \sigma _{33}-\hbar \Omega _{p}\sigma _{31}-\hbar \Omega _{p}\sigma _{32}+H.c.\text{,} \end{equation*}% where $\Omega _{p,c}=\dfrac{A_{p,c}\mu _{3i}}{\hbar }$ ($i=1,2$) are the respective Rabi frequencies, $\sigma _{ij}=|i\rangle \langle j|$ are the atomic transition operators, $\Delta =\omega _{p}-\omega _{31}=$ $\omega _{c}-\omega _{32}$ is the detuning of the pulse frequencies from the upper level and $\mu _{3i}$ ($i=1,2$) are the dipole moments of corresponding transitions. We assume\ that: (i) the probe field is weak as compared to the coupling pulse field $\Omega _{p}<<\Omega _{c}$; (ii) the interaction is adiabatic ($% \Omega _{c}T>>1$, where $T$ is the interaction duration). Then, the atomic density matrix equation may be written as \begin{eqnarray} \overset{.}{\rho }_{31} &=&-\Gamma \rho _{31}+i\Omega _{p}+i\Omega _{c}\rho _{21}, \notag \\ \overset{.}{\rho }_{21} &=&i\Omega _{c}^{\ast }\rho _{31}, \label{bloch} \\ \rho _{11} &=&1, \notag \\ \rho _{22} &=&\rho _{33}=\rho _{32}=0, \notag \end{eqnarray}% where $\Gamma $ is the width of the upper level which is the sum of the spontaneous decay and transverse relaxations rates. It is supposed that interaction is fast enough to neglect the decoherence between metastable levels (sec. III, IV), or to take it into account to the first order (sec. V). The propagation of the pulses is governed by the Maxwell equation for slowly varying amplitudes,% \begin{eqnarray} \left( \frac{\partial }{\partial x^{\prime }}+\frac{1}{c}\frac{\partial }{% \partial t^{\prime }}\right) \Omega _{p} &=&iq_{p}\rho _{31}, \label{prop} \\ \text{ \ }\left( \frac{\partial }{\partial x^{\prime }}+\frac{1}{c}\frac{% \partial }{\partial t^{\prime }}\right) \Omega _{p} &=&iq_{c}\rho _{32}, \notag \end{eqnarray}% where $q_{i}=\dfrac{2\pi \mu _{3i}\omega _{i}N}{\hbar c}$, $N$ is the atomic number density. System of equations (\ref{bloch}) can be reduced to one equation for $\rho _{31}$% \begin{equation} \overset{..}{\rho }_{31}-\overset{.}{\rho }_{31}\frac{\overset{.}{\Omega }% _{c}}{\Omega _{c}}+\rho _{31}\Omega _{c}^{2}+\Gamma \left( \overset{.}{\rho }% _{31}-\rho _{31}\frac{\overset{.}{\Omega }_{c}}{\Omega _{c}}\right) =i\Omega _{c}\overset{.}{\theta } \label{dens} \end{equation}% where $\theta =\dfrac{\Omega _{p}}{\Omega _{c}}$ is the common used notation for the so called mixing angle. Influence of the first two terms in (\ref% {dens}) can be neglected if we confine to\ only first terms with respect to the nonadiabaticity (i.e. $\left( \Omega _{c}T\right) ^{-2}<<1$ is neglected). Influence of the fourth term in (\ref{dens}) is essential parameter only under the assumption% \begin{equation} \Gamma T>>1. \label{largeg} \end{equation}% The meaning of the condition (\ref{largeg})\ is obvious: under the condition of complete adiabaticity relaxation does not affect the pulse propagation (dark-state), but taking into account first nonadiabatic correction, has essential influence. The relaxation can be neglected when $\Gamma T\lesssim 1 $. Finally, by substituting the Maxwell equation (\ref{prop}) into (\ref{dens}) one gets pulse propagation equation in wave variables $x=x^{\prime }$, $% t=t-x^{\prime }/c$,% \begin{equation} \frac{1}{\Gamma _{1}}\frac{\partial ^{2}\theta }{\partial x\partial t}+\frac{% q_{p}}{\Omega _{c}^{2}}\frac{\partial \theta }{\partial t}+\frac{\partial \theta }{\partial x}=0, \label{probe_prop} \end{equation} where notation $\Gamma _{1}=\dfrac{\Omega _{c}^{2}}{\Gamma }$ is used. In this connection the coherence dynamics is governed by the following equation:% \begin{equation} \overset{.}{\rho }_{21}=-\Gamma _{1}\rho _{21}-\Gamma _{1}\theta \label{coh_dyn} \end{equation} Thus $\Gamma _{1}$ is the coherence decay rate, or width of EIT\ resonance, due to applied coupling. Under the condition% \begin{equation} \Gamma _{1}T>>1 \label{bright} \end{equation}% equation (\ref{coh_dyn}) has the well known quasi-stationary solution $\rho _{21}=-\theta $ \cite{polariton}, and the equation (\ref{probe_prop}) passes to the dark state polariton propagation equation. The condition (\ref{bright}% ) means, that width of EIT\ resonance exceeds the spectral width of the probe. \section{\label{solution of propag}Solution of propagation equation} The obtained probe pulse propagation equation (\ref{probe_prop}) is solved by the method presented in \cite{Mostowski}. Since (\ref{probe_prop}) is linear in $\theta $ and $\Omega _{c}\left( t\right) $ is independent of $x$, it can be solved by using the Laplace transform with respect to $x$. The solution of (\ref{probe_prop}) for $\theta $'s Laplace image can be found easily:% \begin{equation} \overset{\symbol{126}}{\theta }\left( s,t\right) =\int\limits_{-\infty }^{t}% \frac{\overset{.}{\theta }_{0}+\Gamma _{1}\theta _{0}}{s+q_{p}/\Gamma }% B\left( s,t,t_{1}\right) dt_{1}+c\left( s\right) B\left( s,t,-\infty \right) \label{image_sol} \end{equation}% where $B\left( s,t,t_{1}\right) =\exp \left( -\dfrac{s}{s+q_{p}/\Gamma }% \int\limits_{t_{1}}^{t}\Gamma _{1}dt^{\prime }\right) $, $c(s)$ is an integration constant that is determined by the initial condition, $c\left( s\right) =\overset{\symbol{126}}{\theta }\left( s,-\infty \right) $. If pulses are switched in counterintuitive sequence (coupling turns on earlier than the probe does) then $c\left( s\right) =0$, since $\theta \left( z,-\infty \right) =\theta _{0}\left( -\infty \right) =0$ (see appendix A). Space time evolution of the probe pulse is obtained by implementing the reverse Laplace transform in (\ref{image_sol}).% \begin{eqnarray} \theta \left( z,t\right) &=&\int\limits_{-\infty }^{t}dt_{1}\left( \theta _{0}\left( t\right) \Gamma _{1}+\overset{.}{\theta }_{0}\left( t\right) \right) \times \label{solution} \\ &&\times \exp \left( -z-\alpha \left( t_{1},t\right) \right) I_{0}\left( 2% \sqrt{z\alpha \left( t_{1},t\right) }\right) , \notag \end{eqnarray} where $z=\dfrac{q_{p}x}{\Gamma }$ is propagation distance normalized to linear absorption factor and, for convenience, the notation $\alpha \left( t_{1},t\right) =\int\limits_{t_{1}}^{t}\Gamma _{1}\left( t^{\prime }\right) dt^{\prime }$ is used. \ By using the condition $z\alpha \left( t_{1},t\right) >>1$ one can substitute the modified Bessel function by its asymptote, so the solution (\ref{solution})\ reduces to the following: \begin{eqnarray} \theta \left( z,t\right) &=&\frac{1}{2\sqrt{\pi }}\int\limits_{-\infty }^{t}dt_{1}\left( \theta _{0}\left( t\right) \Gamma _{1}\left( t_{1}\right) +% \overset{.}{\theta }_{0}\left( t\right) \right) \times \label{gauss} \\ &&\times \exp \left( -\left( \sqrt{z}-\sqrt{\alpha \left( t_{1},t\right) }% \right) ^{2}\right) \left( z\alpha \left( t_{1},t\right) \right) ^{-1/4}. \notag \end{eqnarray} Depending on optical propagation distance $z$ two simple asymptotes for (\ref% {gauss}) can be obtained (see appendix B). The first is the case where% \begin{equation} \frac{\Gamma _{1m}T}{\sqrt{z}}>>4\sqrt{\ln 2} \label{polariton_cond} \end{equation}% $\Gamma _{1m}$ is the maximal value of $\Gamma _{1}\left( t\right) $ (see also \cite{polariton}). Under condition (\ref{polariton_cond}), solution (% \ref{solution}) reduces to the dark-state polariton propagation solution with correction in (\ref{bright}):% \begin{equation} \theta \left( z,t\right) =\theta _{0}\left( \xi \right) +\frac{1}{\Gamma _{1}\left( \xi \right) }\overset{.}{\theta }_{0}\left( \xi \right) \label{polariton} \end{equation}% where $\xi $ is the non-linear time determined by $\int\limits_{\xi }^{t}\Omega _{c}^{2}\left( t_{1}\right) dt_{1}=q_{p}x$ \cite{Nasha}. Note, that turning off the coupling $\Omega _{c}\left( t\right) $ does not reduce $% \Gamma _{1}\left( \xi \right) $\ to zero, since $\xi $ retards from $t$. In this case, as it will be shown below, the information stored in the medium can be well retrieved (see sec. IV). In the case of condition reversed to (\ref{polariton_cond}),% \begin{equation} \frac{\Gamma _{1m}T}{\sqrt{z}}<<4\sqrt{\ln 2}. \label{bluring_cond} \end{equation}% the\emph{\ }solution (\ref{solution}) reduces to \begin{equation} \theta \left( z,t\right) =R\exp \left( -\left( \sqrt{z}-\sqrt{\alpha \left( t_{0},t\right) }\right) ^{2}\right) \left( z\alpha \left( t_{0},t\right) \right) ^{-1/4} \label{bluring} \end{equation}% where $t_{0}$ is the maximal value of $\theta _{0}\left( t\right) $, $% R=\int\limits_{-\infty }^{\infty }\Gamma _{1}\left( t^{\prime }\right) \theta _{0}\left( t^{\prime }\right) dt^{\prime }$ and does not depend on time. We emphasize that for propagation distances meeting the condition (\ref% {bluring_cond}), the obtained pulse loses all the information about its initial temporal shape, since the right hand side in (\ref{bluring}) does not contain time dependent $\theta _{0}$. \section{\label{calc}Discussion} In this section the probe pulse propagation dynamics obtained from the analytical solution (\ref{solution}) is presented. First of all we consider the case of constant coupling field. Shape of the initial pulse is chosen to be double-humped in order to visualize the propagation dynamics. For the small propagation distances when the condition (\ref{polariton_cond})\ is met influence of $\Gamma $ is negligible (Fig.2a). When the condition (\ref% {polariton_cond}) is violated, the influence of upper level width becomes essential as one can see from Figs. 2b,c. Thus influence of $\Gamma $ breaks the adiabaton propagation regime. As it was mentioned above, propagation over very long distances (\ref% {bluring_cond}) leads to the lost of the information on the initial pulse temporal shape. This can be seen in Fig 3, where propagation over the same distance of two pulses with different temporal shapes but with the same initial area is depicted. By propagating over very long distance (\ref% {bluring_cond}) they lose any information about their initial temporal shapes. In Fig. 4 we show that the increase of $\Gamma _{1}$ suppresses the smearing of the probe. This is caused by the decrease of the bright state population and hence leads to the decrease the influence of $\Gamma $ on pulse propagation. Note, that in the dark-state propagation regime the pulse temporal shape does not depend on coupling field amplitude or on unstable level width. Summarizing presented results one can see that to minimize the pulse smearing during its propagation one has to either increase $\Gamma _{1}$ or decrease the propagation distance $z$. The situation changes dramatically for the light storage and retrieving process ($\Omega _{c}\neq const$). It is known, that pulse can be completely stored and retrieved from the medium if the medium length and $\Gamma _{1}$ meet the condition (see for example \cite{Nasha}):% \begin{equation} z\gtrsim \Gamma _{1m}T. \label{fitting} \end{equation} For efficient storage and retrieving the condition (\ref{polariton_cond}) also has to be met. Combining this two nonequalities one gets that to completely store and well retrieve the light pulse, $\Gamma _{1}$ has to meet the following condition:$\ $% \begin{equation} \Gamma _{1m}T>>16\ln 2>>1. \label{good_storage} \end{equation} Therefore, influence of the second term in (\ref{polariton}) is insufficient when the condition (\ref{good_storage}) is met. Storage and retrieving of the light pulse for different propagation distances under the condition (\ref{good_storage}) is depicted in Fig 5. For small propagation distances when the condition (\ref{fitting}) is violated only the falling edge of the pulse is stored and can be retrieved (Fig. 5a), because when this edge enters the medium, the leading edge emerges already. For larger propagation distances when the condition (\ref{fitting}) is satisfied the whole pulse can be stored and then well retrieved by turning on the coupling field. Let us now consider the case when the condition (\ref{good_storage}) is not met (Fig. 6). For small propagation distances when the condition (\ref% {polariton_cond}) is satisfied but (\ref{fitting})\ is not, only the falling edge of the pulse can be stored and retrieved. Propagation over longer distances brings to satisfying of (\ref{fitting})\ and violation of (\ref% {polariton_cond}). Thus the whole pulse can be stored but the retrieved pulse temporal shape is smeared. We present finally comparison of the experimental results with our analytical solution. In Fig 7a the experimental data of storage and retrieving of light pulse from \cite{experiment} are presented. Curve in Fig 7b is plotted from our analytical solution (\ref{solution}): all parameters correspond to the conditions of the experiment. One can see good consistency between experimental data and our analytical solution (Note that the storage in case of experiment () is incomplete as was discussed above). \section{\label{sec_V}Consideration of $\protect\rho _{21}$ transverse decay} In this sections we take into acount the quantity $\gamma T$ in first order. This leads, instead of (\ref{bloch}), to the equations. \begin{eqnarray*} \overset{.}{\rho }_{31} &=&-\Gamma \rho _{31}+i\Omega _{p}+i\Omega _{c}\rho _{21}, \\ \overset{.}{\rho }_{21} &=&-\gamma \rho _{21}+i\Omega _{c}^{\ast }\rho _{31}, \\ \rho _{11} &=&1, \\ \rho _{22} &=&\rho _{33}=\rho _{32}=0, \end{eqnarray*} Thus, probe pulse propagation equation is written as follows:% \begin{equation} \frac{1}{\Gamma _{1}}\frac{\partial ^{2}\theta }{\partial x\partial t}+\frac{% \partial \theta }{\partial x}+\frac{q_{p}}{\Gamma _{1}\left( \Gamma +\gamma \right) }\frac{\partial \theta }{\partial t}+\frac{q_{p}\gamma }{\Gamma _{1}\left( \Gamma +\gamma \right) }\theta =0, \label{V_2} \end{equation} where $\Gamma _{1}\left( t\right) $ now is% \begin{equation} \Gamma _{1}\left( t\right) =\frac{\Omega _{c}^{2}+\gamma \left( \Gamma +% \dfrac{\dot{\Omega}_{c}}{\Omega _{c}}\right) }{\Gamma +\gamma }. \label{V_3} \end{equation} As results from (\ref{V_3}), to completely stop the light in the medium $% \left( \Gamma _{1}=0\right) $ one should turn off the coupling field $\Omega _{c}$ at the rate $\Gamma $ (i.e., $\Gamma +\dfrac{\dot{\Omega}_{c}}{\Omega _{c}}=0$). By performing the stated above Laplace transform procedure one obtains analytical solution of the equation (\ref{V_2}) in the form% \begin{eqnarray} \theta \left( z,t\right) &=&\int\limits_{-\infty }^{t}dt_{1}\left( \theta _{0}+\Gamma _{1}\dot{\theta}_{0}\right) \times \label{V_4} \\ &&\times \exp \left( -z-\int\limits_{t_{1}}^{t}\Gamma _{1}dt^{\prime }\right) I_{0}\left( 2\sqrt{z\alpha \left( t_{1},t\right) }\right) \notag \end{eqnarray}% where $\alpha \left( t_{1},t\right) =\int\limits_{t_{1}}^{t}\Gamma _{1}\left( t^{\prime }\right) -\gamma dt^{\prime }$ and $z=\dfrac{q_{p}x}{% \Gamma +\gamma }$. More detailed analysis of the expression (\ref{V_2}) will be performed in a subsequent publication. \section{Conclusion} We have considered the propagation, storage and retrieving of the light pulse in EIT media by taking into account all dephasing rates. From coupled system of Maxwell and density matrix equations we derive the probe pulse propagation equation, which in particular case passes into the dark-state polariton propagation equation. We find an analytical solution and analyzed its physical consequences. We derived a simple asymptotes of the solution, and showed strong dependence of light pulse temporal shape on optical propagation distance in the presence of relaxations. We demonstrated that an efficient storage of light is possible by choosing appropriate coupling intensities and optical propagation distances. Finally, we compared our solution with experimental data and showed that our solution is well consistent with the recent experiments. \begin{acknowledgments} We are grateful to Prof. M.Fleischhauer, Prof. V.Chaltykyan and Prof. Yu.Malakyan for helpful discussions. The work was supported by the ISTC Grant No. \#A-1095. \end{acknowledgments} \bigskip
{ "timestamp": "2005-04-01T23:34:38", "yymm": "0503", "arxiv_id": "quant-ph/0503209", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503209" }
\section{introduction} \vskip -0.15in Since the early realization of sub-micron atom lithography \cite{timp}, the subject of focusing neutral atoms by use of light fields continues to attract a great deal of attention. The basic principle of atom lithography relies on the possibility of concentrating the atomic flux in space utilizing a spatially modulated atom-light interaction. In the conventional atom-lithographic schemes, a standing wave (SW) of light is used as a mask on atoms to concentrate the atomic flux periodically and create desired patterns at the nanometer scale \cite{review}. The technique has been applied to many atomic-species in one \cite{{prentis},{mcc1},{sodium},{mcc2},{chr},{alu},{ces},{ytt},{iro}} as well as two-dimensional \cite{twod} pattern formations. There are two ways to focus a parallel beam of atoms by light masks in close correspondence with conventional optics. In the thin-lens approach, atoms are focused outside the region of light field which happens for low intensity light beams. On the other hand, the atoms can be focused within the light beam when its intensity is high. This is known as thick-lens regime and is very similar to the graded-index lens of traditional optics. The laser focusing of atoms depends on parameters such as thickness of light beam, velocity spread of atoms, detuning of laser frequency from the atomic transition frequencies, etc,. Experimentally, atomic nanostructures have been reported with sodium \cite{{timp},{sodium}}, chromium \cite{{mcc2},{chr}}, aluminium \cite{alu}, cesium \cite{ces}, ytterbium \cite{ytt}, and iron \cite{iro} atoms. Most of the theoretical studies on atom lithography employ a particle optics approach to laser focusing of atoms \cite{{prentis},{mcc1},{ashkin}}. The classical trajectories of atoms in the potentials induced by light fields suffice to study the focal properties of light lens. In the case of direct laser-guided atom deposition, the diffraction resolution limit will be ultimately determined by the de Broglie wavelength of atoms, and may reach several picometers for typical atomic beams \cite{lee}. In practice, however, this limit has never been relevant because of the surface diffusion process, the quality of the atomic beam, and severe aberrations due to anharmonicity of the sinusoidal dipole potential. As a result, all current atom lithography schemes suffer from a considerable background in the deposited structures. A possible way to overcome the aberration problem was suggested in \cite{sch}, by using nanofabricated mechanical masks that block atoms passing far from the minima of the dipole potential. However, this complicates considerably the setup and reduces the deposition rate. Therefore, there is a considerable need in a pure atom optics solution for the enhanced focusing of an atomic beam having a significant angular spread. In paraxial approximation, the steady-state propagation of an atomic beam through a standing light wave is closely connected to the problem of the time-dependent lateral motion of atoms subject to a spatially periodic potential of an optical lattice. From this point of view, enhanced focusing of the atomic beam can be considered as a squeezing process on atoms in the optical lattice. In recent work \cite{mleib}, novel squeezing technique has been introduced for atoms in a pulsed optical lattice. The approach considered a time modulation of the SW with a series of short laser pulses. Based on specially designed aperiodic sequence of pulses, it has been shown that atoms can be squeezed to the minima of the light-induced potential with reduced background level. Oskay {\it et al.} \cite{raizen} have verified this proposal experimentally using Cs atoms in an optical lattice. In Refs. \cite{{mleib},{raizen}}, the atoms were loaded into the optical lattice and the dynamics of atoms along the direction of SW was studied as a time-dependent problem. The aim of the present work is to extend the focusing scenario of Ref. \cite{mleib} to the beam configuration employed for atomic nanofabrication. We generalize the results on atomic squeezing in the pulsed SW to a system involving the atomic-beam traversing several layers of light masks. In particular, we will investigate prospects for reducing spherical and chromatic aberrations in atom focusing with double-layer light masks. High-resolution deposition of chromium atoms will be considered as an example. The plan of the paper is as follows. In Sec. II, the basic framework of the problem is defined and the linear focusing of atoms by a double-layer light mask is studied using the particle optics approach in paraxial approximation. In Sec. III, we examine the optimal squeezing scheme of \cite{mleib} in application to the atomic-beam traversing two layers of light masks. The effects of beam collimation and chromatic aberrations are considered in Sec. IV. Here, we optimize the double lens performance and give parameters for the minimum spot-size in the atom deposition. Finally, in Sec. V, we summarize our main results. \section{squeezing of atoms by multi-layer light masks - classical treatment} \vskip -0.15in The focusing property of a single SW light has been studied in great details by McClelland {\it et al.} \cite{{mcc1},{mcc2}}. The light acts like an array of cylindrical lenses for the incident atomic beam, focusing the atoms into a grating on the substrate. However, because of the non-parabolic nature of the light-induced potential, the focusing of atoms is subject to spherical aberrations giving a finite width to the deposited features \cite{mcc1}. A doublet of light masks made from two standing light waves may, in principle, reduce the focusing imperfections due to a clear physical mechanism. In this configuration, the first SW prefocuses the atoms towards the minima of the sinusoidal potential. When the pre-focused atoms cross the second SW, they see closely the parabolic part of the potential which should result in a reduction of the over-all spherical aberrations. To test this scheme, we consider the propagation of an atomic-beam through a combination of two SWs formed by counter-propagating laser beams. The two SWs are identical except for their intensities and are assumed to be formed along the x-direction. Atoms are described as two-level systems with transition frequency $\omega_o$. We take the direction of propagation of atoms through the SW fields along the z-direction. If the atoms move sufficiently slow (adiabatic conditions) through the light fields, the internal variables of atoms maintain a steady state during propagation \cite{cohen}. In this approximation, the atoms can be described as point-like particles moving under the influence of an average dipole-force. The potential energy of interaction is given by \cite{{ashkin},{conserve}} \begin{equation} U(x,z) = \frac{\hbar \Delta}{2}~\hbox{ln}[1 + p(x,z)]~, \label{poten} \end{equation} where \begin{equation} p(x,z) = \frac{\gamma^2}{\gamma^2 + 4 \Delta^2}~\frac{I(x,z)}{I_s}~. \label{pxz} \end{equation} In Eq. (\ref{pxz}), $\Delta$ is the detuning of the laser frequency from the atomic resonance, $I(x,z)$ is the light intensity, $\gamma$ is \vskip -0.2 in \begin{figure}[t] \epsfxsize=220pt \centerline{ \epsfbox{nfig1.eps} } \end{figure} \vskip -0.1in \noindent FIG. 1. Schematic representation of the laser focusing of atoms by a double layer of Gaussian standing waves. The intensity profile shows the Gaussian envelopes along the z-axis and the sinusoidal variations along the x-axis. \vskip 0.2in \noindent the spontaneous decay rate of excited level, and $I_s$ is the saturation intensity associated with the atomic transition. For the arrangement of two SW light masks (denoted by 1 and 2) with separation $S$ between them, the net intensity profile of light is given by \begin{eqnarray} I(x,z) &=& \left[I_1 \exp(-2 z^2/\sigma_z^2) + I_2 \exp(-2 {(z - S)}^{2}/\sigma_z^2) \right] \nonumber \\ &&~~~~~ \times \sin^2(k x)~. \end{eqnarray} Here, $\sigma_z$ is the $1/e^2$ radius and $\lambda = 2 \pi/k$ is the wavelength of laser beams forming the SWs. We consider Gaussian intensity profiles and ignore any y-dependence of laser intensities as the force on atoms along the y-direction is negligible compared to that along the direction of SW (x-axis). $I_1$ and $I_2$ denote the maximum intensity of the standing light waves 1 and 2, respectively. We neglect the overlap and interference between two SWs. The intensity profile of light and the focusing of atoms by light fields are shown schematically in Fig. 1. The classical trajectories of atoms in the potential ($\ref{poten}$) induced by the double-layer light masks obey the Newton's equations of motions~: \begin{equation} \frac{d^2 x}{d t^2} + \frac{1}{m} \frac{\partial U(x,z)}{\partial x} = 0~,~~~~ \frac{d^2 z}{d t^2} + \frac{1}{m} \frac{\partial U(x,z)}{\partial z} = 0~. \end{equation} Using the conservation of energy, we can combine the above two equations and solve for $x$ as a function of $z$. This results in two first-order coupled differential equations for $x(z)$, $\alpha \equiv dx(z)/dz$~~: \begin{eqnarray} \frac{dx(z)}{dz} &=& \alpha~~, \label{newton} \\ \frac{d\alpha(z)}{dz} &=& \frac{1 + \alpha^2}{2 (E - U)} \left(\alpha \frac{dU}{dz} - (1 + \alpha^2) \frac{dU}{dx} \right)~~. \nonumber \end{eqnarray} Here, $E$ represents the total energy of the incoming atoms (the kinetic energy in the field-free region) and $\alpha$ gives the slope of the trajectory $x(z)$. \vskip -0.2 in \begin{figure}[t] \centerline{ \epsfxsize=220 pt \epsfbox{nfig2ab.eps}} \centerline{ \epsfxsize=220 pt \epsfbox{nfig2c.eps}} \end{figure} \vskip -0.2in \noindent FIG. 2. Numerically calculated trajectories of atoms for laser focusing by a single- (a) and double-layer (b) light masks. The parameters used are $I_1/I_s = 1000$, $I_2/I_s = 0$ (a) and $I_1/I_s = 1000$, $I_2 = I_1$, $S = 500$ (b). All other parameters are the same as in Table I. The solid (dashed) curve in graph (c) shows the probability density of atoms at the focal point $z = z_f \approx 650~(700)$ of the double (single) light lens. The region to the right of origin $(x = 0)$ in graph (c) is zoomed and shown in the inset. \vskip 0.2in We first study the focal properties of the light fields, and solve numerically Eq. ($\ref{newton}$) for an atomic beam that is initially parallel to the z-axis. The linear focal points and principal-plane locations can be obtained by tracing paraxial trajectories as discussed in \cite{mcc1}. Some typical results are shown in Fig. 2, where we present the numerical calculation of a series of atomic trajectories entering the nodal region of both single $(I_1 \neq 0,I_2 \equiv 0)$ and double $(I_1,I_2 \neq 0)$ light masks. Table I lists the parameters used in dimensionless units, in which length is expressed in units of $\lambda$, and frequency is in units $\omega_{r} \equiv \hbar k^2 /2 m $ corresponding to the recoil energy. We have considered the intensities of light SWs to be equal in the case of double light masks. For the other variables, the values close to the experimental parameters of the chromium atom-deposition \cite{dirk} are taken as an example, though the general conclusions to be drawn should apply to other atoms. It is seen from Fig. 2, that a sharp focal spot appears in the flux of focused atoms \cite{pflux}. Despite the small size of the focal spot, the overall localization of atoms in the focal plane is not very marked. Atomic background in the focal plane indeed gets reduced with double light masks as shown in the inset of Fig. 2(c), however this effect is not very pronounced. To take full advantage from double-mask arrangement, we have to replace the concept of linear focusing (useful for paraxial trajectories only) by the notion of optimized nonlinear spatial squeezing. \section{optimal squeezing theory - \\application to atom nanolithography} \vskip -0.1in We have seen that the double light lens leads to some improvement in feature contrast in the focal plane in comparison to the single light lens. However, even for a single SW, the best squeezing of atoms (maximal spatial compression) is achieved not at the focal plane, but after the linear focusing phenomenon takes place. To characterize the spatial localization of atoms we use a convenient figure of merit, the localization factor \cite{mleib}: \begin{eqnarray} L(z) &=& 1 - <\cos\left(2 k x(z,x_o)\right)> \nonumber \\ &\equiv& \frac{2}{\lambda} \int_{-\lambda/4}^{\lambda/4} dx_o \left[1 - \cos\left(2 k x(z,x_o)\right) \right]~, \label{local} \end{eqnarray} where $x(z,x_o)$ is the solution of the differential equations ($\ref{newton}$) satisfying the initial condition $x\rightarrow x_{0}$ at $z \rightarrow - \infty$. The average in Eq. ($\ref{local}$) is taken over the random initial positions of atoms and the localization factor is measured as a function of distance $z$ from the center $(z=0)$ of the first SW. The localization factor equals zero for an ideally localized atomic ensemble, and is proportional to the mean-square variation of the x coordinate (modulo standing wave period) in the case of well-localized distribution $(L << 1)$. \vskip 0.1in Ref. \cite{mleib} considered the squeezing process in the time-domain by analyzing the action of pulsed SWs on atoms. In the Raman-Nath approximation, this corresponds to the thin-lens regime (in space domain) for interaction of a propagating atomic-beam with multiple layers of light masks. According to the optimal squeezing strategy \cite{mleib}, the time sequence of pulses applied to the atomic system is determined by minimizing the localization factor. To apply this procedure to the atom squeezing by multi-layer light masks, we should minimize the localization factor ($\ref{local}$) in the parameter space: the separations between the light SWs, their intensities, and the relative distance of substrate surface with respect to the layers of light masks. This optimization can be done numerically \vskip 0.1in \begin{table}[htb] TABLE I. Parameters in scaled units. Frequency is measured in recoil units, and length in units of the optical wavelength. Energy is given in the units of recoil energy $\hbar \omega_r$. \vskip 0.03in \begin{tabular}[t]{lc} ~~~~~~~~~~Parameter & Numerical value \\ \hline Spontaneous emission rate $\gamma$ & 238 \\ Detuning $\Delta$ & 9500 \\ 1/$e^2$ radius of SW $\sigma_z$ & 120 \\ Energy of the incoming atoms $E$ & 3$~\times 10^9$ \\ \end{tabular} \end{table} \begin{figure}[t] \epsfxsize=220pt \centerline{ \epsfbox{nfig3.eps} } \end{figure} \vskip -0.2in \noindent FIG. 3. Localization factor of the atomic distribution for squeezing by a single- (dashed curve) and double-layer (solid curve) light masks. The parameters used are $I_1/I_s = 1000$, $I_2/I_s = 0$ (dashed curve) and $I_1/I_s = 1000$, $I_2 = I_1$, $S = S_m \approx 1000$ (solid curve). All other parameters are the same as in Table I. The minimal value of $L(z)$ is 0.31 (0.42) and it occurs at $z = z_m \approx 1450~(1300)$ for the solid (dashed) curve. In the case of the double light mask, the point $(z_m,S_m)$ corresponds to the numerically found global minimum of the localization factor. \vskip 0.1in \noindent using the established simplex-search method. Our numerical analysis shows that the localization factor exhibits multiple local minima even for the simplest case of double light masks. In Fig. 3, we plot the localization factor as a function of distance $z$ both for single and double light masks around its global minimum $(z_m,S_m)$. The intensities of SWs have been chosen to be equal and satisfy the thin-lens condition of atom-light interaction \cite{thin}. The graph shows that the localization factor gets a sizable reduction with double light masks indicating for an enhanced focusing of atoms. The minimum values of $L(z)$ in Fig. 3 are in conformity with the values obtained for optimal squeezing of atoms with single and double pulses in the time-dependent problem \cite{mleib}. We emphasize that the best squeezing (localization) of atoms does not occur at the focal point. Figure 4 displays the spatial distribution of atoms at the point of best localization. Instead of a single focal peak, a two-peaked spatial distribution of atoms near the potential minima is observed in Fig. 4. The origin of these peaks can be related to the formation of rainbows in the wave optics and quantum mechanics, and it is discussed in detail in \cite{{mleib},{rain}}. Moreover, on comparing the inset of Figs. 2(c) and (4), it is seen that the optimized separation between layers of the double light mask results in a considerable reduction of atomic deposition in the background. This also leads to an overall increased concentration of atomic flux near the potential minima. We note, that according to \cite{{mleib},{raizen}}, further squeezing of atoms can be achieved by increasing the number of identical SWs in the multi-layer light masks. For the best localization, again the optimized values for the separations between light masks should be used. \vskip 0.8in \begin{figure}[h] \epsfxsize=220pt \centerline{ \epsfbox{nfig4.eps} } \end{figure} \vskip -0.5in \noindent FIG. 4. Probability density of atoms at the point of the best squeezing by a single- (dashed curve) and double-layer (solid curve) light masks. The parameters used are $I_1/I_s = 1000$, $I_2/I_s = 0$, $z = z_m \approx 1300$ (dashed curve) and $I_1/I_s = 1000$, $I_2 = I_1$, $S = S_m \approx 1000$, $z = z_m \approx 1450$ (solid curve). All other parameters are the same as in Table I. The region to the right of origin $(x = 0)$ is zoomed and shown in the inset. \vskip -0.1in In the above analysis, we have considered the case of equal intensities for the light lenses and the problem has been studied in the thin-lens \cite{firstthin} regime of atom focusing by light masks. However, in many current atom-lithographic schemes, focusing of atoms is generally achieved using an intense SW light. This corresponds to the thick-lens regime of atom-light interaction. In this limit, the focal point is within or close to the region of laser fields and hence a detailed information on atomic motion within the light is required for a full description \cite{mcc1}. For the chromium atoms deposition, the focusing of atoms to the center of an intense SW has been extensively studied both theoretically \cite{mcc1} and experimentally \cite{mcc2}. We show here that a combination of a thin and thick lenses can result in the enhanced localization of atoms \begin{figure}[h] \epsfxsize=215pt \centerline{ \epsfbox{nfig5.eps} } \end{figure} \vskip -0.5in \noindent FIG. 5. Minimal localization factor (maximal squeezing) of the atomic distribution as a function of the relative intensity $I_r \equiv I_2/I_1$ of standing light waves in a double light mask. The parameters used are same as in Table I with $I_1/I_s = 1500$. \noindent with minimal background structures. For illustration, we consider the focusing of atoms by a doublet of light masks made of a thin and a thick lens. We fix the intensity of the first SW light mask to satisfy the thin-lens limit and study the best localization of atoms that can be achieved by varying the intensity of the second SW. A plot of the minimal value of the localization factor versus the relative intensity of the second light mask is shown in Fig. 5. The graph shows that the localization factor becomes almost insensitive to the variation in relative intensity after the intensity ratio reaches the value of 5, and it approaches a small value of $L=0.15$. This result is to be compared with the value of $L=0.31$ for the optimal squeezing by two thin lenses, and $L = 0.42$ achievable by a single thin lens. Fig. 6 shows the corresponding trajectories of atoms and a plot of atomic distribution at \begin{figure}[h] \centerline{ \epsfxsize=185 pt \epsfbox{nfig6a.eps}} \centerline{ \epsfxsize=215 pt \epsfbox{nfig6b.eps}} \end{figure} \vskip -0.25in \noindent FIG. 6. (a) Numerical trajectory calculation for laser focusing by a double light mask. The parameters used for the calculation are $I_1/I_s = 1500$, $I_2 = 25 I_1$, and $S = S_m \approx 1500$. All other parameters are the same as in Table I. (b) Probability density of the atomic distribution at the point $(z_m,S_m)$ of maximal squeezing by the double light mask. The parameters used are the same as those of (a) with $z = z_m \approx 1550$. The point $(z_m,S_m)$ is the numerically found global minimum of the function $L(z)$ with respect to the variables $(z,S)$. The dashed curve in graph (b) shows the atomic distribution at the point $z = z_m$ of the best squeezing by a single thick light lens with parameters $I_1/I_s = 37500$, $I_2/I_s = 0$, $z_m \approx 90$. The region to the right of origin $(x = 0)$ is zoomed and shown in the inset. \noindent the point of the best localization. Note that the optimized double light mask reduces the atomic background by a factor of three in the midpoint $(x = 0.25)$ between two deposition peaks (see the inset of Fig. 6). Moreover, the background in the optimized double mask scheme is five times smaller compared to the usual atom deposition in the focal plane (graph not shown) of a single thick lens. \section{parameters for optimal squeezing of a thermal atomic beam} \vskip -0.11in The effects that have been discussed so far assume an initially collimated $(\alpha = 0)$ beam of atoms with fixed velocity (or energy). However, in atom optics experiments involving thermal atomic beams, the atoms possess a wide range of velocities along the longitudinal (z-axis) and transverse (x-axis) directions. In order to characterize the atom spatial squeezing under such conditions, we need to average the localization factor Eq. (\ref{local}) over the random initial velocities and angles of the beam. The averaging can be done by using the normalized probability density \cite{mcc1} \begin{equation} P(v,\alpha) = \frac{1}{2 \sqrt{2\pi}}~\frac{1}{\alpha_o v_o^5}~v^4 \exp\left[- \frac{v^2}{2 v_o^2} \left(1 + \frac{\alpha^2}{\alpha_o^2}\right)\right]~, \label{aver} \end{equation} where $v_o$ is the root mean square speed of atoms with average energy $\bar{E} \equiv m v_o^2/2$. In the above equation, the term proportional to $v^3 \exp(-v^2/2v_o^2)dv$ represents the thermal flux probability of having a longitudinal velocity $v$ along the z-direction. The probability of having a transverse velocity $v_x = \alpha v$ along the x-direction is proportional to the Gaussian distribution $\exp(-v_x^2/2v_o^2 \alpha_o^2) dv_x$, where $\alpha_o$ is the FWHM of the angular distribution. Using the probability density ($\ref{aver}$), the averaged localization factor is thus given by \begin{eqnarray} L(z) &=& \frac{2}{\lambda} \int_{\alpha = -\infty}^{\alpha = \infty} \int_{v=0}^{v=\infty} \int_{x_o = -\lambda/4}^{x_o = \lambda/4} P(v,\alpha) \nonumber \\ &&~~~~~\times \left[1 - \cos\left(2 k x(z,x_o)\right) \right] dx_o dv d\alpha \label{mainlocal}~. \end{eqnarray} Here, $x(z,x_o)$ represents the solution of differential equations ($\ref{newton}$) for varying initial conditions $(x_o,v,\alpha)$ at $z \rightarrow -\infty$ of atoms. Note that the solution of Eq. ($\ref{newton}$) depends on the initial conditions $(v,\alpha)$ through the energy term $E \equiv m v^2 (1 + \alpha^2)/2$ as well. Since the focal length of light masks depends on velocity of the incoming atoms, the velocity spread in the atomic beam leads to the broadening of the deposited feature size. In the particle optics context of atom focusing, this is referred to as chromatic aberration. In addition, the initial angular divergence $(\alpha \neq 0)$ of the atomic beam degrades greatly the focusing of atoms. We are interested in the extent to which the velocity and angular spreads degrade the optimal squeezing of atoms. The best feature contrast in the presence of aberrations is again defined by minimizing the localization factor, Eq. (\ref{mainlocal}). We have carried out the triple integration in Eq. (\ref{mainlocal}) \begin{figure}[h] \centerline{ \epsfxsize=210 pt \epsfbox{nfig7ab.eps}} \centerline{ \epsfxsize=215 pt \epsfbox{nfig7c.eps}} \end{figure} \vskip -0.2in \noindent FIG. 7. Localization factor of the atomic distribution for squeezing by a single- (a) and double-layer (b) light masks. The parameters used are $\bar{E} = 3 \times 10^9$, $\alpha_o = 10^{-4}$, $I_1/I_s = 1500$, and (a) $I_2/I_s = 0$, (b) $I_2 = I_1$, $S = S_m \approx 800$. All other parameters used are the same as in Table I. The minimal value of $L(z)$ is 0.67~[0.8] and it occurs at $z = z_m \approx 1350~[975]$ in the graph (b)~[(a)]. The dashed and solid curves in graph (c) give the atomic distribution at the point $(z_m,S_m)$ of best squeezing by the single- and double-layer light masks with the parameters of (a) and (b). The region to the right of origin $(x = 0)$ in graph (c) is zoomed and shown in the inset. \vskip 0.2in \noindent numerically and optimized the localization factor $L(z)$ in the parameter space $(z,S)$ for the case of the double-layer light masks. Figures 7 and 8 display atomic distribution at the point of the best squeezing by thin-thin and thin-thick lenses configurations. On comparing the results with those ones for a single thin or thick lens, it is seen that the thin-thick lens combination provides the smallest feature size for the atom deposition. In the case of thin-thin lenses, the effects of chromatic aberrations are greater because of the strong dependence of focal length on the atomic velocity. We note that, though the initial velocity and angular spread of thermal beam worsen the optimal squeezing of atoms, the effects may become less important with increasing the number of layers in the multi-layer light masks. Further, chromatic aberrations can be greatly reduced by employing low-temperature supersonic beams of highly collimated atoms. \section{summary} \vskip -0.2in In this paper, we presented the particle-optics analysis for atom lithography using multiple layers of SW light \begin{figure}[h] \centerline{ \epsfxsize=210 pt \epsfbox{nfig8ab.eps}} \centerline{ \epsfxsize=215 pt \epsfbox{nfig8c.eps}} \end{figure} \vskip -0.24in \noindent FIG. 8. Localization factor of the atomic distribution for squeezing by a single- (a) and double-layer (b) light masks. The parameters used are $\bar{E} = 3 \times 10^9$, $\alpha_o = 10^{-4}$, and (a) $I_1/I_s = 37500$, $I_2/I_s = 0$, (b) $I_1/I_s = 1500$, $I_2 = 25 I_1$, $S = S_m \approx 1200$. All other parameters used are the same as in Table I. The minimal value of $L(z)$ is 0.5~[0.66] and it occurs at $z = z_m \approx 1350~[160]$ in the graph (b) [(a)]. The dashed and solid curves in graph (c) give the atomic distribution at the point $(z_m,S_m)$ of best squeezing by the single- and double-layer light masks with the parameters of (a) and (b). The region to the right of origin $(x = 0)$ in graph (c) is zoomed and shown in the inset. \vskip 0.2in \noindent masks. In particular, we studied the spatial squeezing of atoms by a double layer of standing light waves with particular reference to minimizing the feature size of atom deposition. At first, linear focusing of atoms using paraxial approximation was considered. This showed an improvement in feature contrast at the focal plane, but the effect was rather modest. We then applied the approach of optimal squeezing that was suggested recently for the enhanced localization of atoms in a pulsed SW \cite{mleib}. We showed that this approach works effectively for atomic nanofabrication and can considerably reduce the background in the atom deposition. Based on the optimal squeezing approach, a new figure of merit, the localization factor, was introduced to characterize the atomic localization. Both, thin-thin and thin-thick lens regimes of atom focusing were considered for monoenergetic as well as thermal beams of atoms. The parameters for the smallest feature size were found by minimizing the localization factor. We have shown that using a proper choice of lens parameters, it is possible to narrow considerably the atomic spatial distribution using the double-layer light mask instead of the single-layer one. Finally, we note that our model calculations neglect the effects of atomic recoil due to spontaneous emission and the dipole force fluctuations. These effects are generally beyond the scope of the classical particle optics analysis and can be treated by means of a fully quantum approach. A detailed quantum mechanical study of the optimal atomic squeezing in application to nanofabrication will be published elsewhere. \begin{center} {\bf ACKNOWLEDGMENTS} \end{center} This work was supported by German - Israeli Foundation for Scientific Research and Development.
{ "timestamp": "2005-03-22T13:20:06", "yymm": "0503", "arxiv_id": "quant-ph/0503181", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503181" }
\section{Introduction and main result} Dirichlet's theorem on diophantine approximation tells us that we can approximate any real number by rational numbers quite well, namely: \begin{thm}\label{theorem1} For any real $\theta$ and any positive integer $N$, there exist integers $a$ and $q$, with $1 \leq q \leq N$, such that \[ \left| \theta - \frac{a}{q} \right| < \frac {1}{qN}. \] \end{thm} Moreover, the bound $1/(qN)$ is best possible, apart from the constant factor. To see this, it suffices to consider the golden ratio $\theta = (\sqrt{5} - 1)/2$ (see \cite[\S 11.8]{HW}). During his work in \cite{C}, the first author accidentally stumbled across the following analogous question: \begin{question}\label{q1} For any real $\theta$ and any positive integer $N$, give an upper bound for \[ \min_{ \substack{ a_1, a_2, q_1, q_2 \in \mathbb{Z}\\ 1 \leq q_1, q_2 \leq N}} \left| \frac{a_1}{q_1} + \frac{a_2}{q_2} - \theta \right|. \] \end{question} With the golden ratio in mind, we know that the upper bound can be no better than $O\big( 1/(q_1 q_2 N^2) \big)$. So, what is the best possible upper bound? More generally, \begin{question}\label{q2} Let $k$ be a positive integer. For any real $\theta$ and any positive integer $N$, give an upper bound for \[ \min_{ \substack{ a_1, \dots, a_k, q_1, \dots, q_k \in \mathbb{Z}\\ 1 \leq q_1, \dots, q_k \leq N}} \left| \frac{a_1}{q_1} + \dots + \frac{a_k}{q_k} - \theta \right|. \] \end{question} To these, we have the following result: \begin{thm}\label{theorem2} Let $k$ be a positive integer. For any real $\theta$ and any positive integer $N$, there exist integers $a_1, \dots, a_k$, $q_1, \dots, q_k$, with $1 \le q_1, \dots q_k \le N$, such that \[ \left| \frac {a_1}{q_1} + \dots + \frac {a_k}{q_k} - \theta \right| \ll N^{-k}. \] \end{thm} The bound $N^{-k}$ is best possible in the sense that, for some $\theta$, the minimum in Question \ref{q2} can be as large as $N^{-k}$. For example, if one considers $\theta = 1/(2N^k)$, \[ \left| \frac{a_1}{q_1} + \cdots + \frac{a_k}{q_k} - \theta \right| \ge \frac 1{2N^k} \] for any choice of $a_1, \dots, a_k, q_1, \dots, q_k$, with $1 \le q_1, \dots, q_k \le N$. However, one expects such pathological examples to be relatively rare, and so one may wonder if it is possible to obtain a sharper upper bound involving the $q_i$'s. For example, is it possible to replace $N^{-k}$ by $(q_1 \cdots q_k)^{-1}N^{-k}$ in Theorem \ref{theorem2}? We shall briefly address this issue in the last section. \section{Proof of Theorem \ref{theorem2}} \begin{lem} \label{lemma1} Suppose that $k \ge 1$ is an integer. There is a number $x_0(k) \ge 1$ such that \[ \sideset{}{^*} \sum_{1 \le q_1, \dots, q_k \le x} q_1 \cdots q_k \gg x^{2k}, \] whenever $x \ge x_0(k)$. Here, $\sum^*$ denotes a summation over the $k$-tuples $q_1, \dots, q_k$ such that $(q_i, q_j) = 1$ whenever $1 \le i < j \le k$. \end{lem} \begin{proof} It suffices to show that \begin{equation}\label{1} \sum_{ \substack{ n \le x\\ (n, m) = 1}} \phi(n)^{\alpha}n^{1 - \alpha} \gg x^2\phi(m)m^{-1}, \end{equation} whenever $0 \le \alpha \le k - 1$, $1 \le m \le x^{k - 1}$, and $x \ge x_0(k)$. The conclusion of the lemma will then follow by successive applications of \eqref{1} with $\alpha = 0, 1, \dots, k - 1$ to the summations over $q_k, q_{k - 1}, \dots, q_1$. We now proceed to establish \eqref{1}. We start by showing that \begin{equation}\label{2} \sum_{ \substack{ n \le x\\ (n, m) = 1}} \left( \frac n{\phi(n)} \right)^{\alpha} \ll x \phi(m)m^{-1}. \end{equation} Define the multiplicative functions \[ f(n) = \begin{cases} \big( n/\phi(n) \big)^{\alpha} & \text{if } (n, m) = 1, \\ 0 & \text{if } (n, m) > 1, \end{cases} \qquad g(n) = \sum_{d \mid n} f(d)\mu(n/d). \] Then $g(n) \ge 0$, and \begin{align*} \sum_{n \le x} f(n) &= \sum_{n \le x} \sum_{d \mid n} g(d) = \sum_{d \le x} g(d) \left\lfloor \frac xd \right\rfloor \le x \sum_{d \le x} g(d)d^{-1}\\ &\le x \prod_{p \le x} \sum_{\nu = 0}^{\infty} g(p^{\nu})p^{-\nu} = x \prod_{p \le x} \left( 1 - p^{-1} \right) \sum_{\nu = 0}^{\infty} f(p^{\nu})p^{-\nu} \\ &\le x \prod_{ \substack{ p \mid m\\ p \le x}} \left( 1 - p^{-1} \right) \prod_p \left( 1 + \frac {p^{\alpha} - (p - 1)^{\alpha}}{p(p - 1)^{\alpha}} \right) \\ &\le x \prod_{p \mid m} \left( 1 - p^{-1} \right) \prod_p \left( 1 + \frac {p^{\alpha} - (p - 1)^{\alpha}}{p(p - 1)^{\alpha}} \right) + O(1). \end{align*} The last inequality follows on noting that $m$ has at most $k - 2$ prime divisors $p > x$, and hence, \[ \prod_{ \substack{ p \mid m\\ p > x}} \left( 1 - p^{-1} \right) = 1 + O \big( x^{-1} \big). \] This proves \eqref{2}. On the other hand, when $\alpha = 0$, we have \[ \sum_{ \substack{ n \le x\\ (n, m) = 1}} n = \sum_{d \mid m} \mu(d) d \sum_{k \le x/d} k = \frac {\phi(m)}{2m} x^2 + O(x\tau(m)), \] whence \begin{equation}\label{3} \sum_{ \substack{ n \le x\\ (n, m) = 1}} n^{1/2} \ge x^{-1/2} \sum_{ \substack{ n \le x\\ (n, m) = 1}} n \gg x^{3/2}\phi(m)m^{-1}. \end{equation} Finally, \eqref{1} follows from \eqref{2}, \eqref{3}, and Cauchy's inequality: \[ \sum_{ \substack{ n \le x\\ (n, m) = 1}} \phi(n)^{\alpha}n^{1 - \alpha} \ge \bigg\{ \sum_{ \substack{ n \le x\\ (n, m) = 1}} n^{1/2} \bigg\}^2 \bigg\{ \sum_{ \substack{ n \le x\\ (n, m) = 1}} \left( \frac n{\phi(n)} \right)^{\alpha} \bigg\}^{-1} \gg x^2\phi(m)m^{-1}. \] \end{proof} \begin{proof}[Proof of Theorem \ref{theorem2}] For $0 < \Delta < 1/2$, define \[ t(x) = \max \big( 1 - |x|/\Delta, 0 \big), \qquad g(x) = \sum_{n = -\infty}^{\infty} t(x - n). \] The function $g$ has a Fourier expansion \[ g(x) = \sum_{h = -\infty}^{\infty} \hat g_h e(h x), \qquad \hat g_h = \Delta \left( \frac{\sin \pi \Delta h}{\pi \Delta h} \right)^2. \] We consider the sum \begin{equation}\label{4} \mathcal S = \sideset{}{^*} \sum_{1 \le q_1, \dots, q_k \le N} \sum_{a_1 = 1}^{q_1} \dots \sum_{a_k = 1}^{q_k} g \left( \frac{a_1}{q_1} + \dots + \frac{a_k}{q_k} - \theta \right), \end{equation} where $\sum^*$ has the same meaning as in the Lemma. Putting in the Fourier expansion for $g$, we get \begin{align}\label{5} \mathcal S =& \sideset{}{^*} \sum_{1 \le q_1, \dots, q_k \le N} \sum_{a_1 = 1}^{q_1} \dots \sum_{a_k = 1}^{q_k} \sum_{h = -\infty}^{\infty} \hat g_h e \left( h \left ( \frac{a_1}{q_1} + \dots + \frac {a_k}{q_k} - \theta \right) \right) \\ =& \sideset{}{^*} \sum_{1 \le q_1, \dots, q_k \le N} \sum_{h = -\infty}^{\infty} \hat g_h e(-h \theta) \sum_{a_1 = 1}^{q_1} e \big( ha_1/ q_1 \big) \dots \sum_{a_k = 1}^{q_k} e \big( ha_k/ q_k \big) \notag\\ =& \sideset{}{^*} \sum_{1 \le q_1, \dots, q_k \le N} q_1 \cdots q_k \sum_{ \substack{ h = -\infty\\ q_1 \cdots q_k \mid h}}^{\infty} \hat g_h e(-h \theta) \notag, \end{align} as \[ \sum_{a = 1}^{q} e(ha/q) = \begin{cases} q & \text{if } q \mid h, \\ 0 & \text{otherwise}. \end{cases} \] If $m$ is a positive integer and $\Delta \le m^{-1}$, we have \begin{equation}\label{6} \begin{split} \sum_{h \ne 0} \big| \hat g_{mh} \big| &\le 2 \bigg\{ \sum_{h = 1}^H \Delta + \frac 1{\Delta m^2} \sum_{h = H + 1}^{\infty} h^{-2} \bigg\} \\ &\le 2 \bigg( H\Delta + \frac 1{H\Delta m^2} \bigg) \le 6m^{-1}, \end{split} \end{equation} where $H = \big\lceil (\Delta m)^{-1} \big\rceil$; whereas if $\Delta > m^{-1}$, we have \begin{equation}\label{7} \sum_{h \ne 0} \big| \hat g_{mh} \big| \le \frac {2\zeta(2)}{\Delta m^2} \le 4m^{-1}. \end{equation} Putting \eqref{6} and \eqref{7} (with $m = q_1 \cdots q_k$) into \eqref{5}, we obtain \[ \mathcal S = \Delta \sideset{}{^*} \sum_{1 \le q_1, \dots, q_k \le N} q_1 \cdots q_k + O \big( N^k \big), \] the $O$-implied constant being absolute (in fact, it is $6$). Therefore, upon choosing $\Delta = cN^{-k}$ with a sufficiently large $c > 0$, it follows from the Lemma that $\mathcal S > 0$. Hence, by \eqref{4}, \[ g\left( \frac {a_1}{q_1} + \dots + \frac {a_k}{q_k} - \theta \right) > 0 \] for some integers $a_1, \dots, a_k, q_1, \dots, q_k$ with $1 \le q_1, \dots, q_k \le N$. Then, by the definition of $g$, \[ \left| \frac {a_1}{q_1} + \dots + \frac {a_k}{q_k} - n - \theta \right| \le cN^{-k} \] for some integer $n$. This establishes the theorem. \end{proof} \section{Closing remarks} We conclude this note with a short discussion of possible improvement on the bound $N^{-k}$ in Theorem \ref{theorem2}. For example, is it possible to replace $N^{-k}$ by $(q_1 \cdots q_k)^{-1} N^{-k}$? While such a result may appear to be the right generalization of Dirichlet's theorem, it is not true in general. Indeed, suppose that for any real $\theta$, there exist integers $a_1, \dots, a_k, q_1, \dots, q_k$, with $1 \le q_1, \dots, q_k \le N$, such that \begin{equation}\label{8} \left| \frac{a_1}{q_1} + \cdots + \frac{a_k}{q_k} - \theta \right| \le \frac C{q_1 \cdots q_kN^k}. \end{equation} Then \begin{equation}\label{9} [0, 1] \subseteq \bigcup_{q \in \mathcal D_k(N)} \bigcup_{0 \le a \le q} \left\{ \theta \in \mathbb R \; : \; |\theta - a/q| \le C/(qN^k) \right\}, \end{equation} where $\mathcal D_k(N)$ denotes the set of least common denominators of the sums appearing on the left side of \eqref{8}. By a result of Erd\"os \cite{E}, $\mathcal D_k(N)$ has cardinality \[ |\mathcal D_k(N)| \ll N^k(\log N)^{-c} \] for some constant $c = c(k) > 0$, so it follows from \eqref{9} that \[ 1 \le \sum_{q \in \mathcal D_k(N)} \sum_{0 \le a \le q} \frac {2C}{qN^k} \le 4CN^{-k}|\mathcal D_k(N)| \ll (\log N)^{-c}, \] which is impossible when $N \to \infty$. On the other hand, one may hypothesize that the set of fractions with denominators in $\mathcal D_k(N)$ is distributed similarly to the set of all fractions $a/q$ with denominators $q \le N^k$. Under such a hypothesis, one might hope for an estimate with $|\mathcal D_k(N)|^{-1}$ in place of the term $(\log 3N)^cN^{-k}$ on the right side of \eqref{10} below, and such an estimate, if true, would be essentially best possible. However, upon observing that \[ |\mathcal D_k(N)| \ge \sum_{q_1 \le N} \cdots \sum_{q_k \le N} d(q_1 \cdots q_k)^{-1} \ge \bigg\{ \sum_{q \le N} d(q)^{-1} \bigg\}^{k} \gg N^k(\log 3N)^{-k}, \] we will take a more cautious approach and pose the following \begin{question} Let $k$ be a positive integer. Determine the least value of $c_k$ such that for any real $\theta$ and any positive integer $N$, there exist integers $a_1, \dots, a_k$, $q_1, \dots, q_k$, with $1 \le q_1, \dots, q_k \le N$, such that \begin{equation}\label{10} \left| \frac {a_1}{q_1} + \dots + \frac {a_k}{q_k} - \theta \right| \ll \frac {(\log 3N)^{c_k}}{q_1 \cdots q_k N^k}. \end{equation} \end{question} We leave the answer to this question to the future. \bigskip \begin{acknowledgement} The first author would like to thank the American Institute of Mathematics for support. The second author would like to thank Jeff Vaaler for several enlightening conversations on this and related topics. \end{acknowledgement}
{ "timestamp": "2005-04-12T18:25:08", "yymm": "0503", "arxiv_id": "math/0503440", "language": "en", "url": "https://arxiv.org/abs/math/0503440" }
\section{Introduction} Let $G$ be a connected reductive group over $\ensuremath{{\mathbb R}}$, and $T$ a maximal torus in $G$. Assume that $G$ has a discrete series of representations. Let $A$ be the split part of $T$, and $M$ the centralizer of $A$ in $G$. It is a Levi subgroup of $G$ containing $T$. Let $E$ be a finite-dimensional representation of $G(\ensuremath{{\mathbb C}})$, and consider the packet $\Pi_E$ of discrete series representations $\pi$ of $G(\ensuremath{{\mathbb R}})$ which have the same infinitesimal and central characters as $E$. Write $\Theta_{\pi}$ for the character of $\pi$, and put \[ \Theta^E=(-1)^{q(G)} \sum_{\pi \in \Pi_E} \Theta_{\pi}. \] \noindent Here $q(G)$ is half the dimension of the symmetric space associated with $G$. Note that $\Theta^E(\gamma)$ will not extend to all elements $\gamma \in T(\ensuremath{{\mathbb R}})$, in particular to $\gamma=1$. Define the number $D_M^G(\gamma)$ by \[ D_M^G(\gamma) = \det(1-\Adj(\gamma); \Lie(G)/ \Lie(M)). \] Then a result of Arthur and Shelstad [1] states that the function \[ \gamma \mapsto |D_M^G(\gamma)|^{\frac{1}{2}} \Theta^E(\gamma), \] \noindent defined on the set of regular elements $T_{\reg}(\ensuremath{{\mathbb R}})$ extends continuously to $T(\ensuremath{{\mathbb R}})$. We denote this extension by $\Phi_M(\gamma,\Theta^E)$. These quantities give the contribution from the real place to the $L^2$-Lefschetz numbers of Hecke operators in [1] and [2]. An expression for $\Phi_M(\gamma,\Theta^E)$ as essentially a sum over elements in the Weyl group $W$ of $T$ in $G$ appears in the proof of Lemma 4.1 in [2]. Although this expression suffices to prove the lemma, it can be considerably refined when $\gamma$ is in the maximal elliptic subtorus $T_e(\ensuremath{{\mathbb R}})$ of $T(\ensuremath{{\mathbb R}})$. \bigskip The following theorem is proved in section 4. \bigskip {\bf Theorem 1.} {\it If $\gamma \in T_e(\ensuremath{{\mathbb R}})$, then \it} \[ \Phi_M(\gamma,\Theta^E)=(-1)^{q(L)}\cdot |W_L| \cdot \sum_{\omega \in W^{LM}} \varepsilon(\omega) \cdot \tr(\gamma;V^M_{\omega(\lambda_B+\rho_B)-\rho_B}). \] Here we write $L$ for the centralizer of $T_c$ in $G$, where $T_c$ is the maximal compact subtorus of $T$. Also write $W_L$ and $W_M$ for the Weyl groups of $T$ in $L$ and $M$. The latter are subgroups of $W$ which commute and have trivial intersection. Here $W^{LM}$ is a certain set of representatives for the cosets $(W_L \times W_M) \backslash W$. It is defined explicitly in section 5. We write $\varepsilon$ for the sign character of $W$. Finally by $V^M_{\omega(\lambda_B+\rho_B)-\rho_B}$ we denote the irreducible finite-dimensional representation of $M$, with highest weight $\omega(\lambda_B+\rho_B)-\rho_B$, where $\lambda_B$ is the $B$-dominant highest weight of $E$. \bigskip In particular, we obtain the extremely simple expression, \[ \Phi_A(1,\Theta^E)=(-1)^{q(G)} \cdot |W|, \] \noindent in the case of a split torus $T=A$. \bigskip We now describe the organization of this paper. \bigskip In section 2, we spell out the relationship between the root systems of $G$, $L$, and $M$. There are two distinct systems of chambers in $X_*(A) \otimes_{\ensuremath{{\mathbb Z}}} \ensuremath{{\mathbb R}}$ obtained from these root systems which are important to understand. \bigskip In section 3, we take the aforementioned lemma a step further to express $\Phi_M(\gamma,\Theta^E)$ explicitly as a linear combination of characters. (Actually we do the computation for any stable virtual character $\Theta$, as it is no more difficult.) The sum over $W$ simplifies to a sum over Kostant representatives $W^M$. \bigskip In section 4, in which we deal specifically with $\Phi_M(\gamma,\Theta^E)$, we distill out the action of $W_L$. A sum over $W^{LM}$ remains. At a key step we use a result of section 5, the computation of an alternating sum of stable discrete series constants. \bigskip In section 5, we prove the result mentioned above, in the context of abstract root systems. It is independent from the rest of the paper. \bigskip I am indebted to my advisor Robert Kottwitz for suggesting the problem and many useful comments. I also thank Christian Kaiser for some helpful conversations. This project was carried out during a stay at the Max-Planck-Institut f\"{u}r Mathematik in Bonn, and I am grateful to the Institut for its support and hospitality. \section{$L$-chambers and $\ensuremath{{\mathcal P}}$-chambers} Let $G$ be a connected reductive group over $\ensuremath{{\mathbb R}}$, and $T$ a maximal torus of $G$. Assume that $G$ has a discrete series, or equivalently, that $G$ has an elliptic maximal torus. \bigskip Write $T_c$, respectively $A$, for the maximal compact, resp. split, subtori of $T$ with centralizers $L$, resp. $M$, in $G$. Write $R$ for the root system of $T$ in $G$, and $R_L$, resp. $R_M$, for the set of roots of $T$ in $L$, resp. $M$. Then $R_L$ is the subset of $R$ consisting of real roots, and $R_M$ is the subset of imaginary roots. Write $W_L$ and $W_M$ for the respective Weyl groups. They are commuting subgroups of $W$ with trivial intersection. Note that $W_L$ fixes each root in $R_M$. \bigskip $A$ is contained as a split maximal torus in $L_{\der}$, the derived group of $L$, and we may identify $R_L$ with the set of roots of $A$ in $L_{\der}$. \bigskip Write $\ensuremath{{\mathfrak a}}_M$ for $X_*(A) \otimes_{\ensuremath{{\mathbb Z}}} \ensuremath{{\mathbb R}}$. For any $\alpha \in R \backslash R_M$ the root hyperplane $H_{\alpha}$ of $X^*(T)_{\ensuremath{{\mathbb R}}}:=X_*(T) \otimes \ensuremath{{\mathbb R}}$ gives a hyperplane in $\ensuremath{{\mathfrak a}}_M$. Let us consider two kinds of chambers in $\ensuremath{{\mathfrak a}}_M$ obtained from these. Define $\ensuremath{{\mathcal P}}$-chambers to be those obtained by deleting from $\ensuremath{{\mathfrak a}}_M$ all the hyperplanes $H_{\alpha}$, with $\alpha \in R \backslash R_M$. Define $L$-chambers to be those obtained by deleting all the $H_{\alpha}$ with $\alpha \in R_L$. The latter are the Weyl chambers for $A$ in $L_{\der}$; therefore $W_L$ acts simply transitively on them. \bigskip Observe that $R_L \subset (R \backslash R_M)$. Any additional hyperplanes coming from roots in $R \backslash (R_L \cup R_M)$ divide the $L$-chambers into $\ensuremath{{\mathcal P}}$-chambers. Thus every $\ensuremath{{\mathcal P}}$-chamber is contained in a unique $L$-chamber. \bigskip Write $\ensuremath{{\mathcal P}}(M)$ for the set of parabolic subgroups of $G$ admitting $M$ as a Levi component. There is a one-to-one correspondence between $\ensuremath{{\mathcal P}}(M)$ and the set of $\ensuremath{{\mathcal P}}$-chambers in $\ensuremath{{\mathfrak a}}_M$, obtained as follows: for $P=MN \in \ensuremath{{\mathcal P}}(M)$, the corresponding $\ensuremath{{\mathcal P}}$-chamber is \[ \ensuremath{{\mathfrak a}}_P^+= \{x \in \ensuremath{{\mathfrak a}}_M : \langle \alpha,x \rangle > 0 \text{, for all } \alpha \in R_N \}, \] \noindent where $R_N$ denotes the set of roots of $T$ in $\Lie(N)$. \bigskip Recall that the set of $L$-chambers is in bijection with the set of Borel subgroups of $L$ containing $T$, or equivalently the set of positive root systems $R_L^+$ in the root system $R_L$. \bigskip Now let $C_P$ be a $\ensuremath{{\mathcal P}}$-chamber, and let $P=MN$ be the corresponding element of $\ensuremath{{\mathcal P}}(M)$. It is easy to see that $R_N \cap R_L$ is a positive system in $R_L$, and this corresponds to an $L$-chamber $C_L$. Thus we have defined a map $C_P \mapsto C_L$ from the set of $\ensuremath{{\mathcal P}}$-chambers to the set of $L$-chambers. It is the obvious one which associates to $C_P$ the unique $L$-chamber containing $C_P$. \section{A Linear Combination of Characters} A stable virtual character is a finite $\ensuremath{{\mathbb Z}}$-linear combination $\Theta$ of characters $\Theta_{\pi}$ so that \[ \Theta(\gamma)=\Theta(\gamma^{\prime}) \] \noindent whenever $\gamma$ and $\gamma^{\prime}$ are regular, stably conjugate elements of $G(\ensuremath{{\mathbb R}})$. In Lemma 4.1 of [2], it is proved that for a stable virtual character $\Theta$ on $G(\ensuremath{{\mathbb R}})$, the function \[ \gamma \mapsto |D^G_M(\gamma)|^{\frac{1}{2}}\Theta(\gamma) \] \noindent on $T_{\reg}(\ensuremath{{\mathbb R}})$ extends continuously to $T(\ensuremath{{\mathbb R}})$. A key ingredient of the proof is the fact that the expression at the bottom of page 497 is a linear combination of irreducible finite-dimensional representations of $M$. In this section we will compute explicitly the coefficients and the representations involved, in the case where the element $a$ appearing in the proof is equal to $1$. \bigskip We translate the set-up of the proof in [2] as follows. We take $\Gamma$ to be the identity component of $T(\ensuremath{{\mathbb R}})$. The root system $R_{\Gamma}$ is then simply $R_L$. Fix a positive root system $R_L^+$ in $R_L$, and let $C$ be the corresponding $L$-chamber in $\ensuremath{{\mathfrak a}}_M$. We then choose a parabolic subgroup $P=MN$ so that $R_L \cap R_N \subseteq R_L^+$. Note that $R_L \cap R_N$ is also a system of positive roots, so this condition is equivalent to having $R_L \cap R_N=R_L^+$. Thus we simply require that the $\ensuremath{{\mathcal P}}$-chamber corresponding to $P$ be contained in $C$. \bigskip Although at the end of our computations we will allow $\gamma$ to be nonregular, we choose now $\gamma$ to be a regular element of $\Gamma= T_c(\ensuremath{{\mathbb R}}) \cdot \exp(\bar{C})$. \bigskip The expression is \begin{equation} \sum_B m(B) \frac{\Delta_P(\gamma) \cdot \lambda_B(\gamma)}{\Delta_B(\gamma)}. \end{equation} \noindent The sum ranges over Borels containing $T$, which correspond to elements of $W$. \bigskip \noindent Here $\lambda_B$ is the $B$-dominant highest weight of $E$, \[ \Delta_B=\prod_{\alpha >0} (1-\alpha^{-1}) \text{, and } \Delta_P=\prod_{\alpha \in R_N}(1-\alpha^{-1}). \] Fix now a Borel $B$ of $G$ with $T \subseteq B \subseteq P$, for the rest of this paper. \bigskip Recall the set of Kostant representatives $W^M$ for the Weyl group $W_M$ of $M$, relative to B. It is the set $\{ w \in W| w^{-1}R_M^+ \subset R^+ \}$. \bigskip If $w \in W$, write $w * B$ for $wBw^{-1}$. \bigskip We will use the observation that for $\omega \in W^M, (\omega * B)_M=B_M$. Indeed, if $\alpha \in R^+ \cap R_M$, then $\omega^{-1} \alpha \in R^+$, which implies that $\alpha \in \omega R^+ \cap R_M$. \bigskip Our sum (1) breaks up as \begin{equation} \label{break} \sum_{\omega \in W^M} m(\omega * B) \cdot \Delta_P(\gamma) \cdot \sum_{w_M \in W_M} \frac{w_M (\omega\lambda_B)(\gamma)}{\Delta_{w_M \omega * B}(\gamma)}. \end{equation} We would prefer the denominator inside the sum to be $\Delta_{ w_M * B_M}(\gamma)$. (Recall that $B_M=B \cap M$.) Note that $\Delta_P \cdot \Delta_{B_M} = \Delta_B$, since $R^+$ is the disjoint union of $R_M^+$ and $R_N$. \bigskip So we consider the quantity \begin{equation} \label{quantity} \frac{ \Delta_P \cdot \Delta_{w_M * B_M}}{\Delta_{w_M \omega * B}}= \frac{ \Delta_B \cdot \Delta_{w_M * B_M}}{\Delta_{B_M} \cdot \Delta_{w_M \omega * B}} . \end{equation} \noindent Observe that if $\BB$ is a Borel, $\Delta_{\BB}=\delta_{\BB} \cdot \rho_{\BB}^{-1}$, where $\delta_{\BB}=\prod_{\alpha > 0} (\alpha^{\frac{1}{2}}-\alpha^{-\frac{1}{2}})$ and $\rho_{\BB}$ is the usual half sum of positive roots. Since $\delta_{w * \BB}=\varepsilon(w)\delta_{\BB}$, we compute that \[ \frac{\Delta_{w * \BB}}{\Delta_{\BB}}=\varepsilon(w) \cdot (\rho_{\BB}-w\rho_{\BB}). \] Thus \eqref{quantity} becomes \[ \varepsilon(\omega) (w_M(\omega \rho_B-\rho_{B_M})-\rho_B+\rho_{B_M}). \] Next observe that for $w_M \in W_M$, \[ w_M(\rho_B-\rho_{B_M})=\rho_B-\rho_{B_M}. \] Indeed, the roots of $R^+$ not in $R_M^+$ are in $R_N$, and are thus normalized by $W_M$. So the above expression simplifies to \[ \varepsilon(\omega) \cdot w_M(\omega \rho_B-\rho_B). \] We can therefore rewrite \eqref{break} as \begin{equation} \label{rewrite} \sum_{\omega \in W^M} m(\omega * B) \cdot \varepsilon(\omega) \cdot \sum_{w_M \in W_M} \frac{w_M (\omega(\lambda_B+\rho_B)-\rho_B)(\gamma)}{\Delta_{w_M * B_M}(\gamma)}. \end{equation} Since $\omega$ is a Kostant representative, the weight $\omega(\lambda_B+\rho_B)-\rho_B$ is positive for $B_M$, and we may use the Weyl character formula to rewrite this as \begin{equation} \label{Weyl} \sum_{\omega \in W^M} m(\omega * B) \cdot \varepsilon(\omega) \cdot \tr(\gamma; V^M_{\omega(\lambda_B+\rho_B)-\rho_B}). \end{equation} \noindent Here $V^M_{\omega(\lambda_B+\rho_B)-\rho_B}$ denotes the irreducible finite-dimensional representation of $M$ with highest weight $\omega(\lambda_B+\rho_B)-\rho_B$. \section{A Formula for $\Phi_M(\gamma,\Theta^E)$} To identify \eqref{Weyl} with $\Phi_M(\gamma,\Theta^E)$, we replace $m(\omega * B)$ with $n(\gamma,\omega * B)$ as on page 500 of [2], and multiply it by the factor $\delta_P^{\frac{1}{2}}(\gamma)$: \begin{equation} \label{factor} \delta_P^{\frac{1}{2}}(\gamma) \cdot \sum_{\omega \in W^M} n(\gamma,\omega * B) \cdot \varepsilon(\omega) \cdot \tr(\gamma; V^M_{\omega(\lambda_B+\rho_B)-\rho_B}). \end{equation} \noindent Here $\delta_P$ is the modulus character of $P$. (We are still only considering regular $\gamma$.) \bigskip Write $A_G$ for the split component of the center of $G$. Let $\lambda_0 \in X^*(A_G)$ denote the character by which $A_G$ acts on $E$. It extends to $X^*(T)_{\ensuremath{{\mathbb R}}}$ in the usual way, and is $W$-invariant. \bigskip Let $T_e$ denote the subtorus of $T$ generated by $T_c$ and $A_G$. It is the maximal subtorus of $T$ which is elliptic in $G$. \bigskip Write $p_M$ for the projection from $X^*(T)_{\ensuremath{{\mathbb R}}}$ to $X^*(A)_{\ensuremath{{\mathbb R}}}$, and note that it is $W_L$-invariant. The group $W_L$ fixes each root of $M$, thus it acts on $W^M$. For every orbit of this action, there is a unique member $\omega$ so that $p_M(\omega(\lambda_B + \rho_B-\lambda_0))$ is dominant with respect to $C$. We denote the set of these elements by $W^{LM}$, one element for each orbit of $W_L$ on $W_M$. \bigskip If $\lambda \in X^*(T)$ and $w_L \in W_L$, then plainly $w_L \lambda-\lambda \in \ensuremath{{\mathfrak a}}_M^*$. Write $(\chi_{w_L,\omega,B},\ensuremath{{\mathbb C}}_{w_L,\omega,B})$ for the one-dimensional representation of $M$, acting through $A$, with weight $w_L\omega(\lambda_B+\rho_B)-\omega(\lambda_B+\rho_B)$. Note that $T_c$ and $A_G$ act trivially on $\ensuremath{{\mathbb C}}_{w_L,\omega,B}$, thus so does $T_e$. \bigskip Thus we have \[ V^M_{ w_L\omega(\lambda_B+\rho_B)-\rho_B} \cong V^M_{ \omega(\lambda_B+\rho_B)-\rho_B} \otimes \ensuremath{{\mathbb C}}_{w_L,\omega,B}. \] \bigskip Our formula \eqref{factor} is now (replacing $\omega \in W^M$ with $w_L \omega$, where $\omega$ is now in $W^{LM}$): \begin{equation} \label{replace} \delta_P^{\frac{1}{2}}(\gamma) \cdot \sum_{\omega \in W^{LM}} \varepsilon(\omega) \cdot \tr(\gamma; V^M_{\omega(\lambda_B+\rho_B)-\rho_B}) \cdot \sum_{w_L \in W_L} \varepsilon(w_L) \cdot \chi_{w_L,\omega,B}(\gamma) \cdot n(\gamma, w_L \omega * B), \end{equation} Of course we now wish to simplify the inner sum. Recall from page 500 of [2] that \[ n(\gamma,w_L \omega * B)= \bar{c}(x,p_M(w_L \omega \lambda_B+ w_L\omega \rho_B - \lambda_0)), \] \noindent where $x$ is in the interior of $C$. Here $\bar{c}(x,\lambda)$ is the integer-valued ``stable discrete series constant'' on \[ (X_*(A/A_G)_{\ensuremath{{\mathbb R}}})_{\reg} \times (X^*(A/A_G)_{\ensuremath{{\mathbb R}}})_{\reg}, \] \noindent as defined, for instance, on page 493 of [2]. Recall that $\lambda_0 \in X^*(T)_{\ensuremath{{\mathbb R}}}$ is obtained from the character $\lambda_0 \in X^*(A_G)$ by which $A_G$ acts on $E$, and is thus $W$-invariant. \bigskip As $p_M$ commutes with $w_L$, the inner sum of \eqref{replace} is now \begin{equation} \label{inner} \sum_{w_L \in W_L} \varepsilon(w_L) \cdot \bar{c}(x,w_L \Lambda) \cdot \chi_{w_L,\omega,B}(\gamma), \end{equation} \noindent where $\Lambda=p_M(\omega \lambda_B+ \omega \rho_B - \lambda_0)$. \bigskip We would like to consider the limit of \eqref{inner} as $x$ approaches $0$. Recall we can write $\gamma=\gamma_c \cdot \exp(x)$, with $\gamma_c \in T_c(\ensuremath{{\mathbb R}})$ and $x$ in $\bar{C}$. Also recall that $\gamma$ is still regular (not for long!). Consider the above formula with $\gamma_c$ fixed and $x$ going to $0$ along regular elements of $\bar{C}$. Fix some element $x_0$ in the interior of $C$. The value \[ \bar{c}(x,w_L \Lambda)=\bar{c}(x_0,w_L \Lambda) \] is unchanged, but $\chi_{w_L,\omega,B}(\gamma)$ approaches $\chi_{w_L,\omega,B}(\gamma_c)=1$. Thus \eqref{inner} converges to \[ \sum_{w_L \in W_L} \varepsilon(w_L) \cdot \bar{c}(x_0,w_L \Lambda) \] \noindent for some $x_0 \in C$. \bigskip But this is simply $(-1)^{q(L)}|W_L|$, by Proposition 1(ii) in Section 5 below. Here we use that $\omega \in W^{LM}$. Note that $-1$ is in the Weyl group of the root system by the argument on page 499 of [2]. \bigskip It is easy to modify this argument to get the same limit as $x$ approaches an element of $X_*(A_G)_{\ensuremath{{\mathbb R}}}$. \bigskip Finally note that $\delta_P$ is a positive character and therefore trivial on the compact group $T_c(\ensuremath{{\mathbb R}})$. It is thus trivial on $T_e(\ensuremath{{\mathbb R}})$. \bigskip Now consider irregular $\gamma$. We take the limit in \eqref{replace} and obtain our theorem: \begin{thm} If $\gamma \in T_e(\ensuremath{{\mathbb R}})$, then \begin{equation} \label{main} \Phi_M(\gamma,\Theta^E) = (-1)^{q(L)}\cdot |W_L| \cdot \sum_{w \in W^{LM}} \varepsilon(w) \cdot \tr(\gamma; V^M_{w(\lambda_B+\rho_B)-\rho_B}). \end{equation} \end{thm} \bigskip \noindent For the reader's convenience, we review the definition of $W^{LM}$. \bigskip The definition depends on the choice of a parabolic $P=MN$ and a Borel subgroup $B$ with $T \subseteq B \subseteq P$. The choice of $B$ gives a set of positive roots $R^+$ for $R$ and a set of positive roots $R_M^+$ for $R_M$. It also gives $B$-dominant elements $\lambda_B$ and $\rho_B$ of $X^*(T)_{\ensuremath{{\mathbb R}}}$. The choice of $P$ determines an $L$-chamber $C$ as in Section 2. Recall the character $\lambda_0$ determined by $A_G$ on $E$ and the projection $p_M$ from $X^*(T)_{\ensuremath{{\mathbb R}}}$ to $X^*(A)_{\ensuremath{{\mathbb R}}}$. Then \[ W^{LM} = \{ w \in W | w^{-1}R_M^+ \subseteq R^+ \text{ and } p_M(w(\lambda_B + \rho_B - \lambda_0)) \text{ is dominant w.r.t. } C \}. \] \bigskip We now evaluate \eqref{main} for $\Phi_M$ on the extreme cases for $T$. If $T=A$ is split, then $M=A$, $L=G$, $W^{LM}$ is trivial, but so is $T_c$. We conclude that for $z \in A_G(\ensuremath{{\mathbb R}})$, \[ \Phi_A(z,\Theta^E)=(-1)^{q(G)} \cdot |W| \cdot \lambda_0(z). \] If $T$ is elliptic, then $M=G$, $L=T$, $W^{LM}$ is again trivial, and so for $\gamma \in T$, \[ \Phi_G(\gamma,\Theta^E)= \tr(\gamma; E). \] \noindent Note that this agrees with the results of Theorems 5.1 and 5.2 of [2], since \[ \tr(\gamma^{-1}; E^*)= \tr(\gamma; E). \] \section{The Sum of the Stable Discrete Series Constants} Let $(X,X^*,R,\check{R})$ be a root system. Write $W$ for the Weyl group of the root system, and $\varepsilon$ for its sign character. Assume that $R$ generates the real vector space $X$ and that $-1 \in W$. Write $q(R)$ for $(|R^+|+\dim(X)) / 2$, as in [2]. Let $x_0$ be a regular element of $X$, and $\lambda$ a regular element of $X^*$. Write $C_0$ for the chamber of $X$ containing $x_0$, and $C_0^{\vee}$ for its dual chamber in $X^*$. Recall the stable discrete series constants $\bar{c}_R(x_0,\lambda)$ from section 3 of [2]. \begin{prop} We have the following formulas for sums of discrete series constants: \begin{itemize} \item[(i)] For all such $ \lambda, \sum_{w \in W} \bar{c}_R(wx_0,\lambda)=|W|.$ \item[(ii)] For $\lambda=\lambda_0 \in C_0^{\vee}$, we have $\sum_{w \in W} \varepsilon(w) \cdot \bar{c}_R(wx_0,\lambda_0)=(-1)^{q(R)}|W|.$ \end{itemize} \noindent The same formulas hold if we sum over the $W$-orbit of $\lambda$ rather than that of $x_0$. \end{prop} We make a few comments before beginning the proof. The proof begins by using the ``inductive'' property (4) of the discrete series constants from page 493 of [2], to change the sum over chambers into a sum over certain facets of $X$. In fact we consider those facets which separate the chambers of $X$, i.e., those which span the root hyperplanes $Y$ of $X$. \bigskip In the course of the proof, we (mis)use the term ``facet'' only in reference to these particular facets, of codimension $1$. So a facet in this sense will be the common face of two adjacent chambers. \bigskip The hyperplanes $Y$ have their own chambers, and we examine the relationship between the facets and these smaller chambers. Not every facet is equal to such a chamber, as in the case of $B_3$ when $Y$ is the root hyperplane of a long root. The facets in $Y$ give a $B_2$ system, but the chambers of $R_Y$ give an $A_1 \times A_1$ system. \bigskip Finally induction on the rank of the root system gives the calculation. \begin{proof} \noindent The second formula follows from the first by applying Theorem 3.2(2) on page 494 of [2]. \bigskip We induce on $r=\dim X$. The proposition is clear when $r=0$. \bigskip We associate these discrete series constants with the various chambers and facets of $X$, and introduce some appropriate notation. \bigskip Write $c(\ensuremath{{\mathcal C}})$ for $\bar{c}_R(x,\lambda)$, when $x$ is in the interior of a chamber $\ensuremath{{\mathcal C}}$. \bigskip Suppose $F$ is a facet in $X$, $y$ is in the interior of $F$, and $\bar{F}:=\Span(F)=Y$. Then write $c(F)=\bar{c}_{R_Y}(y,\lambda_Y)$, notation as on page 493 of [2]. \bigskip Thus if $F$ is the common face of distinct chambers $\ensuremath{{\mathcal C}}$ and $\ensuremath{{\mathcal C}}^{\prime}$, then \[ 2c(F)=c(\ensuremath{{\mathcal C}})+c(\ensuremath{{\mathcal C}}^{\prime}). \] Each chamber has $r$ faces, and it follows that \begin{equation} \label{summing} r \cdot \sum_{\ensuremath{{\mathcal C}}} c(\ensuremath{{\mathcal C}})= 2 \sum_{F} c(F), \end{equation} \noindent where we are summing over all chambers and then all facets. \bigskip We show the right hand side of \eqref{summing} is equal to $r \cdot |W|$ to prove the proposition. \bigskip Now every facet is on some root hyperplane $X_{\alpha}=X_{-\alpha}$, so we have \[ 2 \sum_F c(F)= \sum_{\alpha \in R} \sum_{\bar{F}=X_{\alpha}} c(F). \] We now work with the inner sum. There is a root system on $X_{\alpha}$ whose set of coroots is $\check{R} \cap X_{\alpha}$, which defines chambers $\ensuremath{{\mathcal C}}_{\alpha}$ in $X_{\alpha}$ and constants $c_{\alpha}(\ensuremath{{\mathcal C}}_{\alpha})$. Write $W_{\alpha}$ for the Weyl group of $X_{\alpha}$. We have \[ \sum_{\bar{F}=X_{\alpha}} c(F)=\sum_{\ensuremath{{\mathcal C}}_{\alpha}} \sum_{F \subset \ensuremath{{\mathcal C}}_{\alpha}} c_{\alpha}(\ensuremath{{\mathcal C}}_{\alpha})= \sum_{\ensuremath{{\mathcal C}}_{\alpha}} \sum_{W_{\alpha} \backslash \{F \subset X_{\alpha} \}} c_{\alpha}(\ensuremath{{\mathcal C}}_{\alpha}) = \sum_{W_{\alpha} \backslash \{ F \subset X_{\alpha} \} } \sum_{\ensuremath{{\mathcal C}}_{\alpha}}c_{\alpha}(\ensuremath{{\mathcal C}}_{\alpha}). \] \noindent For the first equality, note that every facet $F$ with $\bar{F}=X_{\alpha}$ is contained in a some chamber $\ensuremath{{\mathcal C}}_{\alpha}$. \bigskip The second equality follows because $W_{\alpha}$ acts transitively on the chambers $C_{\alpha}$. \bigskip Write $n(\alpha)$ for the order of $W_{\alpha} \backslash \{ F \subset X_{\alpha} \}$. It is equal to the number of facets in a given chamber $\ensuremath{{\mathcal C}}_{\alpha}$. Then by induction the above is merely \[ n(\alpha) \cdot |W_{\alpha}|, \] \noindent which is exactly the number of facets in $X_{\alpha}$. It follows that \eqref{summing} is simply equal to twice the total number of facets in $X$. \bigskip Since $W$ has $r$ orbits on the set of facets in $X$, and the stabilizer in $W$ of any facet has order $2$, we conclude that the total number of facets is half of $r \cdot |W|$, as desired. \end{proof}
{ "timestamp": "2005-03-24T10:50:13", "yymm": "0503", "arxiv_id": "math/0503524", "language": "en", "url": "https://arxiv.org/abs/math/0503524" }
\section{Introduction} Experiments in cooling and trapping of neutral gases have paved the way toward a new parameter regime of ionized gases, namely the regime of ultracold neutral plasmas (UNPs). Experimentally, UNPs are produced by photoionizing a cloud of laser-cooled atoms collected in a magneto-optical trap \cite{Kil99}, with temperatures down to 10 $\mu$K. By tuning the frequency of the ionizing laser, initial electron kinetic energies of $E_{\rm{e}}/k_{\rm B} = 1 \mbox{K} - 1000 \mbox{K}$ have been achieved. The time evolution of several quantities characterizing the state of the plasma, such as the plasma density \cite{Kil99,Kul00,Sim04}, the rate of expansion of the plasma cloud into the surrounding vacuum \cite{Kul00}, the energy-resolved population of bound Rydberg states formed through recombination \cite{Kil01}, or electronic \cite{Rob04,Van04} as well as ionic \cite{Sim04} temperature have been measured using various plasma diagnostic methods. Despite the low typical densities of $\approx 10^9$ cm$^{-3}$, the very low initial temperatures suggest that these plasmas have been produced well within the strongly coupled regime, with Coulomb coupling parameters up to $\Gamma_{\rm{e}} = 10$ for the electrons and even $\Gamma_{\rm{i}} = 30000$ for the ions. Thus, UNPs seem to offer a unique opportunity for a laboratory study of neutral plasmas where, depending on the initial electronic kinetic energy, either one component (namely the ions) or both components (ions and electrons) may be strongly coupled. Moreover, the plasma is created in a completely uncorrelated state, i.e.\ far away from thermodynamical equilibrium. The relaxation of a strongly correlated system towards equilibrium is an interesting topic in non-equilibrium thermodynamics and has been studied for decades. The history of this problem must be traced back to the important contributions of Klimontovich \cite{Kli72,Kli73,Kli82}, who pointed out that kinetic energy conserving collision integrals such as the Boltzmann, Landau and Lenard-Balescu collision integrals are not appropriate for such a situation, and derived non-Markovian kinetic equations taking correctly into account total energy conservation of the system. In the following years this problem has attracted much attention and the relaxation of nonequilibrium strongly coupled plasmas has been studied by a variety of different methods \cite{Wal78,Bel96,Bon98,Zwi99}. The very low densities of UNPs make it now possible to directly observe the dynamical development of spatial correlations, which may serve as the first experimental check of the present understanding of the strongly coupled plasma dynamics. Moreover, it turns out that the timescale of the plasma expansion, the correlation time as well as the relaxation time of the ions are almost equal. Therefore Bogoliubov's functional hypothesis, usually used in kinetic theory, breaks down under the present conditions, which may lead to a very interesting relaxation behavior but also causes some difficulties in the theoretical description of these systems, since the plasma dynamics can not be divided into different relaxation stages. \section{Theoretical approach} A full molecular dynamics simulation of ultracold neutral plasmas over experimentally relevant timescales is infeasible with present-day computer resources due to the large number of particles ($N \approx 10^5$) and the long observation times ($t \approx 10^{-4}$ s) involved. In order to model the evolution of UNPs, we have developed a hybrid molecular dynamics (HMD) approach which treats electrons and ions on different levels of sophistication, namely in a hydrodynamical approximation on the one hand (for the electrons) and on a full molecular dynamics level on the other hand (for the ions) \cite{PPR04}. For the electrons, it has been shown that several heating effects, such as continuum threshold lowering \cite{Hah02}, build-up of correlations \cite{Kuz02}, and, predominantly, three-body recombination \cite{Rob02} rapidly increase the electronic temperature. As a consequence, the electrons are always weakly coupled, $\Gamma_{\rm{e}} < 0.2$, over the whole course of the system evolution. Moreover, due to the small electron-to-ion mass ratio, the relaxation timescale of the electrons is much smaller than that of the ions as well as the timescale of the plasma expansion. Hence, an adiabatic approximation may safely be applied, assuming instant equilibration of the electrons. This allows for the use of much larger timesteps than in a full MD simulation since the electronic motion does not need to be resolved. It is this adiabatic approximation for the electrons which makes a molecular dynamics treatment of the ionic motion in UNPs computationally feasible. The electronic density is determined self-consistently from the Poisson equation. The fact that the potential well created by the ions which is trapping the electrons has a finite depth is taken into account by using a King-type distribution \cite{Kin66} known from simulations of globular clusters rather than a Maxwell-Boltzmann distribution for the electron velocities, with the electronic temperature $T_{\rm{e}}$ obtained from energy conservation. The finite well depth also leads to evaporation of a fraction of the free electrons in the very early stage of the plasma evolution, which is accounted for by determining the fraction of trapped electrons from the results of \cite{Kil99}. The dynamics of the heavy particles is described in the framework of a chemical picture, where inelastic processes, namely three-body recombination and electron impact ionization, excitation and deexcitation, are taken into account on the basis of Boltzmann-type collision integrals \cite{Kli81,Kli82}, with the transition rates taken from \cite{Man69}. Numerically, the resulting collision integrals are evaluated using a Monte Carlo sampling as described in \cite{Rob03,PPR04,PPR04c}. The ions and recombined atoms are then propagated individually in a molecular dynamics simulation, taking into account the electronic mean-field potential and the full interaction potential of the remaining ions\footnote{In order to bring out clearly the role of ionic correlations, it is also possible to neglect them in the HMD approach by propagating the ions in the mean-field potential created by all charges rather than the full ionic interaction.}. In order to allow for larger particle numbers, the most time-consuming part of the HMD simulation, namely the calculation of the interionic forces, is done using a treecode procedure originally designed for astrophysical problems \cite{Bar90}, which scales like $N_{\rm{i}} \ln N_{\rm{i}}$ rather than $N_{\rm{i}}^2$ with the number $N_{\rm{i}}$ of ions. As shown in several publications \cite{PPR04,PPR04a,PPR04b,PPR04c}, the HMD approach outlined above provides a powerful method for the description of UNPs, taking full account of ionic correlation effects. However, due to the large numerical effort involved, it is limited to particle numbers of $N_{\rm{i}} \approx 10^5$. While this permits a direct simulation of many, particularly of the early, experiments, an increasing number of experiments is performed with larger particle numbers up to $10^7$. Thus, an alternative method which is able to treat such larger systems is desirable. Such a method is indeed available \cite{PPR04}, based on a hydrodynamical description of both electrons and ions similar to that introduced in \cite{Rob02,Rob03}. Starting from the first equation of the BBGKY hierarchy, one obtains the evolution equations for the one-particle distribution functions $f$ of the electrons and ions. Neglecting again electron-electron as well as electron-ion correlations, and employing the same adiabatic approximation for the electrons already used in the HMD approach, a quasineutral approximation \cite{Dor98} permits expressing the mean-field electrostatic potential in terms of the ionic density, leading to a closed equation for the ion distribution function which contains the electron temperature as a parameter. A Gaussian ansatz for the ion distribution function, \begin{equation} \label{e1} f_{\rm{i}} \propto \exp{\left(-\frac{r^2}{2\sigma^2}\right)}\exp{\left(- \frac{m_{\rm{i}}\left({\bf{v}}-\gamma{\mathbf{r}}\right)^2}{2k_{\rm{B}} T_{\rm{i}}}\right)} \; , \end{equation} which corresponds to the initial state of the plasma cloud, is then inserted into the evolution equations for the second moments $\langle r^2 \rangle$, $\langle \bf{r} \bf{v} \rangle$ and $\langle v^2 \rangle$ of the ion distribution function. In this way, evolution equations for the width $\sigma$ of the cloud, the parameter $\gamma$ of the hydrodynamical expansion velocity $\bf{u} = \gamma{\mathbf{r}}$ and the ionic temperature $T_{\rm{i}}$ are obtained. Ionic correlations are taken into account in an approximate way using a local density approximation together with a gradient expansion, reducing the description of their influence on the plasma dynamics to the evolution of a single macroscopic quantity, namely the correlation energy $U_{\rm{ii}}$ of a homogeneous plasma. The relaxation behavior of $U_{\rm{ii}}$ is modeled using a correlation-time approximation \cite{Bon96} with a correlation time equal to the inverse of the ionic plasma frequency, $\tau_{\rm corr} = \omega_{\rm{p,i}}^{-1}$, together with an analytical expression for the equilibrium value of $U_{\rm{ii}}$ \cite{Cha98}. Finally, inelastic processes such as three-body recombination and electron impact ionization, excitation and deexcitation are incorporated on the basis of rate equations, and the influence of the recombined Rydberg atoms on the expansion dynamics is taken into account assuming equal hydrodynamical velocities for atoms and ions. The final set of evolution equations then reads \begin{subequations} \label{e2} \begin{eqnarray} \label{e2a} \dot{\sigma}&=&\gamma\sigma\;,\\ \label{e2b} \dot{\gamma}&=&\frac{N_{\rm{i}}\left(k_{\rm{B}}T_{\rm{e}}+k_{\rm{B}}T_{\rm{i}}+ \frac{1}{3}U_{\rm{ii}}\right)}{\left(N_{\rm{i}}+N_{\rm{a}}\right)m_{\rm{i}} \sigma^2}-\gamma^2\;,\\ \label{e2c} k_{\rm{B}}\dot{T}_{\rm{i}}&=&-2\gamma k_{\rm{B}}T_{\rm{i}}-\frac{2}{3}\gamma U_{\rm{ii}}-\frac{2}{3}\dot{U}_{\rm{ii}}\;,\\ \label{e2d} \dot{U}_{\rm{ii}}&=&-\omega_{\rm{p,i}}\left(U_{\rm{ii}}-U_{\rm{ii}}^{\rm{(eq)}} \right)\\ \label{e2e} \dot{{\cal{N}}}_{\rm{a}}(n)&=&\sum_{p\neq n}\left[R_{\rm{bb}}{(p,n)} {\cal{N}}_{\rm{a}}(p)-R_{\rm{bb}}{(n,p)}{\cal{N}}_{\rm{a}}(n)\right] \nonumber\\&&+R_{\rm{tbr}}(n)N_{\rm{i}}-R_{\rm{ion}}{(n)}{\cal{N}}_{\rm{a}}(n) \end{eqnarray} and the electronic temperature is determined by energy conservation, \begin{equation} \label{e2f} N_{\rm{i}}k_{\rm{B}}T_{\rm{e}}+\left[N_{\rm{i}}+N_{\rm{a}}\right] \left[k_{\rm{B}}T_{\rm{i}}+m_{\rm{i}}\gamma^2\sigma^2\right]+\frac{2}{3} N_{\rm{i}}U_{\rm{ii}}-\frac{2}{3}\sum_n{\cal{N}}_{\rm{a}}(n) \frac{{\cal{R}}}{n^2}={\rm{const.}}\;, \end{equation} \end{subequations} where ${\cal{N}}_{\rm{a}}(n)$ defines the population of Rydberg states, $N_{\rm{a}}=\sum_n{\cal{N}}_{\rm{a}}(n)$ is the total number of atoms and ${\cal{R}}=13.6$eV is the Rydberg constant. The preceeding hydrodynamical method is much more approximate than the HMD approach, but, on the other hand, it is much simpler and quicker. For particle numbers of $N_i \approx 10^5$, it requires about two orders of magnitude less CPU time. Since its computational effort is independent of the number of particles, it allows for a simulation of larger plasma clouds corresponding to a number of current experiments. Moreover, and maybe equally important, it provides physical insight into the plasma dynamics since it is based on a few simple evolution equations for the macroscopic observables characterizing the state of the plasma. As we have investigated in detail in \cite{PPR04}, there is generally surprisingly good agreement between the hydrodynamical simulation and the more sophisticated HMD calculation as long as macroscopic, i.e.\ spatially averaged, quantities such as electronic temperature, expansion velocity, ionic correlation energy etc.\ are considered. \begin{figure}[tb] \centerline{\psfig{figure=f1a.eps,width=6.3cm} \hfill \psfig{figure=f1b.eps,width=6.3cm}} \caption{\label{f1} Electronic temperature $T_{\rm{e}}(t)$ for an expanding plasma of $40000$ Sr ions with an initial average density of $\rho_{\rm{i}}=10^9$cm$^{-3}$ and an initial electron kinetic energy of $20\:$K, obtained from the HMD simulation (a) and from eqs.\ (\ref{e2}) (b), with (solid) and without (dotted) the inclusion of ionic correlations. The inset shows the ratio of the electron temperatures obtained from both methods.} \end{figure} As an example, we show in figure \ref{f1} the time evolution of the electronic temperature for a plasma of 40000 Sr ions with an initial average density of $10^9$cm$^{-3}$ and an initial electron kinetic energy of $20\:$K, obtained from the HMD simulation (a) and from eqs.\ (\ref{e2}) (b). During the whole system evolution, the agreement between the two simulation methods is better than about 8\%, and it becomes even better at later times. Thus, we conclude that, for the present type of experimental setups, the hydrodynamical method outlined above, and in particular the approximate treatment of ionic correlations, is well suited for the description of the behavior of UNPs. \section{Results and discussion} \subsection{Comparison with experiments} In fact, fig.\ \ref{f1} only shows good agreement between the two theoretical simulation methods, without comparison with experiment. Such a comparison is now also possible, since measurements of the electron temperature dynamics have recently been reported in \cite{Rob04}. Fig.\ \ref{f2} shows the time evolution of the electronic temperature for a Xenon plasma with $N_{\rm{i}}(0)=1.2\cdot10^6$, $\rho_{\rm{i}}(0)=1.35\cdot10^9$cm$^{-3}$ and two different initial temperatures of $T_{\rm{e}}(0)=66.67$K and $T_{\rm{e}}(0)=6.67$K. In addition to the full hydrodynamical simulation according to equations (\ref{e2}), fig.\ \ref{f2} also shows corresponding calculations where the effect of inelastic electron-ion collisions, eq.\ (\ref{e2e}), is neglected (dashed lines). (The plasmas in these experiments are too large to be simulated using the HMD approach.) \begin{figure}[tb] \centerline{\psfig{figure=f2.eps,width=9cm}} \caption{\label{f2} Electronic temperature $T_{\rm{e}}(t)$ for a plasma of $1.2\cdot10^6$ Xenon ions with an initial average density of $1.35\cdot10^9$cm$^{-3}$ for two different initial electron temperatures, $T_{\rm{e}} =6.67$K (filled dots) and $T_{\rm{e}} = 66.67$K (open dots). The lines show the hydrodynamical simulation (solid lines: including inelastic collisions, dashed lines: without inelastic collisions), the dots the experiment \cite{Rob04}, scaled down by 26\% (see text).} \end{figure} \begin{figure}[tb] \centerline{\psfig{figure=f3a.eps,height=3.8cm} \hfill \psfig{figure=f3b.eps,height=3.8cm}} \caption{\label{f3} Time evolution of the average electron density of a Xenon plasma of 500000 ions with an initial average density of $10^9$cm$^{-3}$ and an initial electron temperature of $T_{\rm{e}}=210$K (a) and $T_{\rm{e}}=2.6$K (b). The lines show the results of the model equations (\ref{e2}) (solid lines: including inelastic collisions, dashed lines: without inelastic collisions) and the dots the experimental data from \cite{Kul00}.} \end{figure} For the high initial temperature, there is close agreement between the two corresponding simulations, showing that inelastic processes are almost negligible in this case. Indeed, it is known that the high-temperature plasma expansion is well described by the collisionless plasma dynamics, and the hydrodynamical model is expected to accurately reproduce the plasma dynamics in this regime. Since an overall systematic error of about $70\%$ for the temperature measurement has been reported in \cite{Rob04}, we have exploited this fact to calibrate the measured temperatures by scaling down both experimental data sets by $26\%$ in order to match the high-temperature results to our calculations. As can be seen in the figure, there is excellent agreement between simulation and experiment also for the lower temperature. (We stress that there is no further scaling of the low-temperature data in order to achieve quantitative agreement, the same calibration factor as in the high-temperature case is used.) In this case, inelastic collisions play a decisive role for the evolution of the system. More specifically, as has been found already in \cite{Rob02}, three-body recombination heats the plasma and significantly changes its behavior, leading to a weakly coupled electron gas, as discussed above in connection with the omission of electronic correlation effects in the numerical treatment. Moreover, there has been some discussion in the literature whether the collision rates of \cite{Man69} would still be applicable at these ultralow temperatures, or whether three-body recombination would be significantly altered. The close agreement between the present simulation and the experimental data in fig.\ \ref{f2} suggests that the rates of \cite{Man69}, while ultimately diverging $\propto T_{\rm{e}}^{-9/2}$ for $T_{\rm{e}} \to 0$, still adequately describe three-body recombination processes in the temperature range under consideration. As a second example, figure \ref{f3} shows the time evolution of the electronic density for a Xenon plasma of $500000$ ions with an initial average density of $10^9$cm$^{-3}$ and two different initial electron temperatures of $T_{\rm{e}}(0)=210$K and $T_{\rm{e}}(0)=2.6$K \cite{Kul00}. Again, it can be seen that the model equations nicely reproduce the density evolution in both temperature regimes, in agreement with \cite{Rob02} where it was shown that the low-temperature enhancement of the expansion velocity \cite{Kul00} is caused by recombination heating and is not due to strong-coupling effects of the electrons. \subsection{Role of ionic correlations} Having thus established the validity of our numerical methods for the description of UNPs, we can now turn to a more detailed investigation of the role of ionic correlations in these systems. It is found that, for situations corresponding to the type of experiments \cite{Kil99,Rob04,Van04}, they hardly influence the macroscopic expansion behavior of the plasma. This becomes evident, e.g., in fig.\ \ref{f1}, where the ``full'' simulations as described above (solid lines) are compared to a mean-field treatment of the system completely neglecting correlation effects (dotted lines). The correlation-induced heating of the ions \cite{Mur01,Ger03a,Ger03b} leads to a slightly faster expansion of the plasma, which in turn results in a slightly faster adiabatic cooling of the electrons \cite{PPR04}. However, the overall effect is almost negligible. \begin{figure}[bt] \centerline{\psfig{figure=f4.eps,width=9cm}} \caption{\label{f4} Spatial density $\rho_{\rm{i}}$ (solid) of the ions, at $t=3\:\mu$s, compared to the Gaussian profile assumed for the kinetic model (dashed). Additionally, $\rho_{\rm{i}}$ obtained from the particle simulation using the mean-field interaction only is shown as the dotted line. Initial-state parameters are the same as in fig.\ \ref{f1}.} \end{figure} A closer look, on the other hand, reveals that certain aspects of the expansion dynamics are indeed significantly affected by the strong ion-ion interaction, as can be seen in figure \ref{f4}. There, the spatial density of the ions is shown after $t=3\:\mu$s for the same plasma as in fig.\ \ref{f1}. A mean-field treatment of the particle interactions \cite{Rob03} predicts that a shock front should form at the plasma edge, seen as the sharp spike in fig.\ \ref{f4} (dotted line). Apparently, with ionic correlations included (solid line) the peak structure is much less pronounced than in mean-field approximation. This is due to dissipation caused by ion-ion collisions which are fully taken into account in the HMD simulation. As shown in \cite{Sac85}, by adding an ion viscosity term to the hydrodynamic equations of motion, dissipation tends to stabilize the ion density and prevents the occurrence of wavebreaking which was found to be responsible for the diverging ion density at the plasma edge in the case of a dissipationless plasma expansion. Furthermore, the initial correlation heating of the ions largely increases the thermal ion velocities, leading to a broadening of the peak structure compared to the zero-temperature case. Another obvious aspect where ionic correlations play a dominant role is the behavior of the ionic temperature. Considering the huge ionic coupling constants suggested by the low initial ion temperatures, this temperature turns out to be an important quantity since it directly determines the value of $\Gamma_{\rm{i}}$. According to a mean-field treatment, the ions would remain the (near) zero temperature fluid they are initially. However, as has been pointed out before, the ions are created in a completely uncorrelated non-equilibrium state, and they quickly heat up through the build-up of correlations as the system relaxes toward thermodynamical equilibrium. As shown in \cite{PPR04}, even at early times the ionic velocity distribution is locally well described by a Maxwell distribution corresponding to a (spatially) local temperature, justifying the definition of a --- due to the spherical symmetry of the plasma --- radius-dependent ion temperature $T_i(r,t)$. Moreover, if the spatially averaged temperature is identified with the ion temperature determined by the model equations (\ref{e2}) one can find again good agreement between both approaches concerning the timescale of the initial heating as well as the magnitude of the ion temperature, even at later times \cite{PPR04}. However, as becomes apparent from fig.\ \ref{f5}, the HMD simulations show temporal oscillations of the ionic temperature, which can, of course, not be described by the linear ansatz of the correlation-time approximation. \begin{figure}[bt] \centerline{\psfig{figure=f5.eps,width=9cm}} \caption{\label{f5} Time evolution of the density-scaled average ionic temperature for a plasma consisting of $400000$ ions with an initial electronic Coulomb coupling parameter of $\Gamma_{\rm{e}}(0)=0.07$.} \end{figure} Such temporal oscillations of the temperature during the initial relaxation stage are known from molecular dynamics simulations of homogeneous one-component \cite{Zwi99} and two-component \cite{Mor03} plasmas, which are clearly caused by the strongly coupled collective ion dynamics, since they increase in strength with increasing $\Gamma_{\rm{i}}$ and disappear for $\Gamma_{\rm{i}}(0)<0.5$ \cite{Zwi99}. \begin{figure}[tb] \centerline{\psfig{figure=f6a.eps,width=6.1cm} \hfill \psfig{figure=f6b.eps,width=6.1cm}} \caption{\label{f6} Time evolution of the ion temperature determined from a central sphere with a radius of half of the plasma width $\sigma$ (a) and time dependence of the amplitude of the corresponding oscillations (b). The initial-state parameters are the same as in fig.\ \ref{f5}.} \end{figure} \begin{figure}[t] \centerline{\psfig{figure=f7.eps,width=8cm}} \caption{\label{f7} Temporally and spatially resolved time evolution of the ion temperature. The initial-state parameters are the same as in fig.\ \ref{f5}.} \end{figure} Despite the fact that the maximum initial coupling constant used in \cite{Zwi99} is $\Gamma_{\rm{i}}=5$, while a value of $\Gamma_{\rm{i}}(0)\approx40000$ is considered in the case of fig.\ \ref{f5}, the oscillations observed in \cite{Zwi99} are much more pronounced and persist much longer than in the present case. It becomes apparent that the rapid damping of the temperature oscillations can be traced to the inhomogeneity of the Gaussian density profile by looking at the central part of the plasma only, where the ionic density is approximately constant (figure \ref{f6}). The temperature oscillations with an oscillation period of half of the inverse plasma frequency $\nu_0=\omega_{{\rm{p,i}}}(0)/2\pi$ defined in the central plasma region are much more pronounced in this case, showing an exponential decay with a characteristic damping rate of $\nu_0$ (fig.\ \ref{f6}(b)). The temporally and spatially resolved temperature evolution shown in fig.\ \ref{f7} shows that the radially decreasing ion density leads to local temperature oscillations with radially increasing frequencies, thereby causing also spatial oscillations of the local ion temperature. Therefore, the seemingly enhanced damping rate, which has also been observed in recent experiments, is purely an effect of the averaging of these local oscillations over the total plasma volume. \subsection{Coulomb crystallization through laser cooling} The above considerations show that, while not dramatically affecting the overall expansion behavior of the plasma cloud, strong-coupling effects play an important role in different aspects of the evolution of UNPs. Thus, UNPs provide a prime example of laboratory realizations of strongly nonideal plasmas. Moreover, the HMD approach developed in \cite{PPR04} is well suited for an accurate description of these systems over experimentally relevant timescales, allowing for direct comparison between experiment and theory. Many interesting aspects of the relaxation behavior of these non-equilibrium plasmas may thus be studied in great detail. However, while effects of strong ionic coupling become apparent in UNPs, the naively expected regime with $\Gamma > 100$ can not be reached with the current experimental setups. For the electrons, it is predominantly three-body recombination which heats them by several orders of magnitude, so that $\Gamma_{\rm{e}} < 0.2$ during the whole system evolution. The ionic component, on the other hand, is heated by the correlation-induced heating until $\Gamma_{\rm{i}} \approx 1$, i.e.\ just at the border of the strongly coupled regime \cite{Mur01,Sim04}. Thus, it is the very build-up of ionic correlations one wishes to study that eventually shuts off the process and limits the amount of coupling achievable in these systems. \begin{figure}[bt] \centerline{\psfig{figure=f8.eps,height=5cm}} \caption{\label{f8} Radial density and a central slice of a plasma with $N_{\rm{i}}(0)=80000$, $\Gamma_{\rm{e}}(0)=0.08$, cooled with a damping rate of $\beta=0.2\omega_{\rm{p,i}}(0)$ at a time of $\omega_{\rm{p,i}}(0)t=216$. (For better contrast, different cuts have been overlayed.)} \end{figure} As soon as the reason for this ionic heating became clear, several proposals have been made in order to avoid or at least reduce the effect, among them ({\em i}) using fermionic atoms cooled below the Fermi temperature in the initial state, so that the Fermi hole around each atom prevents the occurrence of small interatomic distances \cite{Mur01}; ({\em ii}) an intermediate step of exciting atoms into high Rydberg states, so that the interatomic spacing is at least twice the radius of the corresponding Rydberg state \cite{Ger03a}; and ({\em iii}) the continuous laser-cooling of the plasma ions after their initial creation, so that the correlation heating is counterbalanced by the external cooling \cite{Kil03,PPR04a}. We have simulated the latter scenario using the HMD method, extended to allow for the description of laser cooling, as well as elastic electron-ion collisions which are negligible for the free plasma expansion but not necessarily in the laser-cooled case \cite{PPR04a,PPR04c}. Laser cooling is modeled by adding a Langevin force \begin{equation} {\bf{F}}_{\rm cool} =-m_i\beta{\bf{v}}+\sqrt{2\beta k_{\rm{B}}T_c m_i} {\bm{\xi}} \end{equation} to the ion equation of motion, where ${\bf{v}}$ is the ion velocity, ${\bm{\xi}}$ is a stochastic variable with $\left<{\bm{\xi}}\right> ={\bf{0}}$, $\left<{\bm{\xi}}(t){\bm{\xi}}(t+\tau)\right>=3\delta(\tau)$, and the cooling rate $\beta$ and the corresponding Doppler temperature $T_c$ are determined by the properties of the cooling laser \cite{Met99}. Elastic electron-ion collisions are taken into account on the basis of the corresponding Boltzmann collision integral, which is again evaluated by a Monte-Carlo procedure \cite{PPR04c}. \begin{figure}[tb] \centerline{\psfig{figure=f9a.eps,width=3.8cm} \hfill \psfig{figure=f9b.eps,width=3.7cm} \hfill \psfig{figure=f9c.eps,width=3.7cm}} \caption{\label{f9} Arrangement of the ions on the first (a), third (b) and fifth (c) shell of the plasma of fig.\ \ref{f8}.} \end{figure} It is found that laser cooling leads to qualitative changes of the plasma dynamics. In particular, it significantly decelerates the expansion of the plasma, whose width is found to increase only as $\sigma\propto t^{1/4}$, in contrast to freely expanding plasmas which behave as $\sigma\propto t$. It is this drastic slow-down of the expansion which favors the development of strong ion correlations, compared to a free plasma where the expansion considerably disturbs the relaxation of the system. The simulations show further that strongly coupled expanding plasmas can indeed be created under realistic conditions, with ionic coupling constants far above the crystallization limit for homogeneous plasmas of $\Gamma_{\rm{i}}\approx174$ \cite{Dub99}. Here we find, depending on the initial conditions, i.e.\ ion number and initial electronic Coulomb coupling parameter, strong liquid-like short-range correlations or even the onset of a radial crystallization of the ions. This is demonstrated in fig.\ \ref{f8}, showing the radial density and a central slice of a plasma with $N_{\rm{i}}(0)=80000$, $\Gamma_{\rm{e}}(0)=0.08$, cooled with a damping rate of $\beta=0.2\omega_{\rm{p,i}}(0)$, at a scaled time of $\omega_{\rm{p,i}}(0)t=216$. The formation of concentric shells in the center of the cloud is clearly visible. As illustrated in fig.\ \ref{f9}, beside the radial ordering there is also significant intra-shell ordering, namely a formation of hexagonal structures on the shells, which are, however, considerably disturbed by the curvature of the shells. \section{Conclusions} In summary, we have used an HMD approach to study the behavior of ultracold neutral plasmas on long time scales. We have shown that effects of strong interionic coupling are indeed visible in such systems, e.g.\ most prominently in the relaxation behavior of the ion temperature, which is connected with transient temporal as well as spatial oscillations. Nevertheless, the strongly coupled regime of $\Gamma > 100$ is not reached with the current experimental setups. We have demonstrated, however, that additional continuous laser cooling of the ions during the plasma evolution qualitatively changes the expansion behavior of the system and should allow for the Coulomb crystallization of the plasma \cite{PPR04a,PPR04c}. It will be an interesting subject for further investigation to study in detail the dynamics of this crystallization process, which differs from the shell structure formation observed in trapped nonneutral plasmas \cite{Dub99} as explained in \cite{PPR04a}. In particular, the influence of the plasma expansion, which presumably causes the transition from liquid-like short-range correlation to the radial ordering, deserves more detailed studies. Other future directions include the study of effects induced by additional magnetic fields, or of ways to confine the plasma in a trap. We gratefully acknowledge many helpful discussions with J.M.\ Rost, as well as conversations with T.C.\ Killian and F.\ Robicheaux.
{ "timestamp": "2005-03-01T20:13:39", "yymm": "0503", "arxiv_id": "physics/0503018", "language": "en", "url": "https://arxiv.org/abs/physics/0503018" }
\section{Introduction}\label{sec:int} A simple generalization of a closed space curve is the notion of a ribbon. An ideal narrow ribbon in three-dimensional space is specified by the position of one edge at each point along its length, together with the unit normal vector to the ribbon at each point of this edge. If the ribbon is a closed loop (with two faces, not one as a M{\"o}bius band), then the two edges are non-intersecting closed curves in space which may wind around each other if the looped ribbon is twisted. Such ideal twisted ribbon loops are important in applications, for instance modelling circular duplex DNA molecules (Fuller 1971, 1978; Pohl 1980; Bauer \textit{et al.} 1980; Hoffman \textit{et al.} 2003), magnetic field lines (Moffatt \& Ricca 1992), phase singularities (Winfree \& Strogatz 1983; Dennis 2004), rotating body frames (Hannay 1998; Starostin 2002), and various aspects of geometric phase theory (Chiao \& Wu 1986; Kimball \& Frisch 2004). A fundamental result in the geometry of twisted, closed ribbon loops is C\u{a}lug\u{a}\-rea\-nu's theorem (C\u{a}lug\u{a}reanu 1959, 1961; Moffatt \& Ricca 1992) (also referred to as White's formula (White 1969; Pohl 1980; Kauffman 2001; Eggar 2000), and the C\u{a}lug\u{a}reanu-White-Fuller theorem (Adams 1994; Hoffman \textit{et al.} 2003), which is expressed as \begin{equation} Lk = Tw + Wr. \label{eq:cwf} \end{equation} $Lk$ is the topological linking number of the two edge curves; it is the classical Gauss linking number of topology (described, for example, by Epple (1998)). The theorem states that this topological invariant is the sum of two other terms whose proper definitions will be given later and which individually depend on geometry rather than topology: the twist $Tw$ is a measure of how much the ribbon is twisted about its own axis, and the writhe $Wr$ is a measure of non-planarity (and non-sphericity) of the axis curve. The formula actually has a very simple interpretation in terms of `views' of the ribbon from different projection directions (this is hinted at in the discussion of Kauffman (2001)). Such interpretations of $Lk$ and $Wr$ go back to Fuller (1971, 1978) and Pohl (1968{\em a}, {\em b}), but do not seem to have been extended to $Tw.$ This is our first result. Our second uses this picture to construct, for any curve, a particular ribbon on it which has zero $Lk.$ This is the writhe framing ribbon, anticipated algebraically by J. H. Maddocks (private communication, unpublished notes; also see Hoffman \textit{et al.} 2003). The topological argument can be paraphrased very simply. Consider the two edges of a ribbon loop, for example that in figure \ref{fig:loop}. Viewing a particular projection as in the picture, there are a number of places where one edge crosses the other. A positive direction around the ribbon is assigned arbitrarily, so that the two edges can be given arrows. At each crossing between the two edge curves, a sign ($\pm$) can be defined according to the the sense of rotation of the two arrows at the crossing ($+1$ for right handed, $-1$ for left handed, as shown in figure \ref{fig:loop}). Summing the signs ($\pm 1$) of the crossings gives twice the linking number $Lk$ of the ribbon edges (the sign of $Lk$ is positive for a ribbon with planar axis and a right-handed twist, and is negative for a planar ribbon with a left-handed twist). Since the ribbon is two-sided, the total number of crossings must be even, ensuring that $Lk$ is an integer. \begin{figure} \begin{center} \includegraphics*[width=8cm]{loops.eps} \end{center} \caption{A projection of a twisted ribbon exhibiting the types of crossing described in the text. The two edges of the ribbon are represented in different colours, and their linking number $Lk$ is +1. At the left and right sides of the figure, when the ribbon is edge-on, there are `local crossings' (a `right-handed' or positive crossing on the right, a `left-handed' negative crossing on the left). Two more positive crossings occur in the middle (where the ribbon crosses over itself); these correspond to the two crossings of the ribbon edges with themselves (i.e. a nonlocal crossing).} \label{fig:loop} \end{figure} The crossings between the two edge curves naturally fall into two types: `local,' which will be associated with $Tw,$ and `nonlocal,' which will be associated with $Wr.$ Local crossings are where the ribbon is edge-on to the viewing direction: one edge of the ribbon is crossing its own other edge. If the ribbon is arbitrarily narrow, then the two edges are indefinitely close in space as they cross in the projection. Nonlocal type crossings between the two edges are caused by the ribbon crossing over itself. More precisely, they occur when a single edge curve of the ribbon crosses over itself. This does not itself count towards the linking number, but such a crossing implies that there are two crossings with the other edge close by, as in the centre of figure \ref{fig:loop}. These two counts, the signed local crossings and signed nonlocal crossings (which come in pairs as described) add up to twice $Lk$ by definition. These quantities are integers, but depend on the choice of projection direction (although their sum does not). It is well known (Fuller 1971, 1978; Adams 1994; Kauffman 2001) that the number of self-crossings of the ribbon axis curve, signed appropriately, and averaged democratically over the sphere of all projection directions, equals the writhe. Therefore, the signed sum of nonlocal crossings between opposite edges is twice the writhe. We claim that the local crossing counterpart (that is, the average over all projection directions of the signed local crossing number) is twice $Tw.$ The C\u{a}lug\u{a}reanu theorem then follows automatically. This is described in the following section. Section \ref{sec:formalism} is a review of standard material formalising the notions of linking and writhe; this is used in section \ref{sec:writhe} to construct the natural writhe framing of any closed curve. The interpretation of twist and writhe in terms of local and nonlocal crossings gives insight into a common application of the theorem, namely `supercoiling' of elastic ribbons (such as DNA, or telephone cords; see, for example, Adams (1994), Pohl (1980), Bauer \textit{et al.} (1980) and Hoffman \textit{et al.} (2003)). Repeated local crossings (i.e. a high twist) are energetically unfavourable, whereas nonlocal crossings (represented by writhe), where the ribbon repeatedly passes over itself (`supercoiling'), are preferred elastically. \section{Local crossings, twist, and a proof of the C\u{a}lug\u{a}reanu theorem}\label{sec:proof} In order to describe the argument more formally, it is necessary to introduce some notation. We will represent the edges of the narrow ribbon by two closed curves, $\mathcal{A}$ and $\mathcal{B},$ whose points are $\mathbf{a}(s)$ and $\mathbf{b}(s)$ (and where no confusion will ensue, just $s$). $s$ is an arbitrary parameterization (giving the sense of direction along the curves), and $\rd \bullet/\rd s$ is denoted $\dot{\bullet}.$ $\mathcal{A}$ will be referred to as the {\em axis curve} and will play the primary role of the two curves. Its unit tangent $\mathbf{t}(s)$ is proportional to $\dot{\mathbf{a}}(s).$ The other curve $\mathcal{B}$ will be regarded as derived from $\mathcal{A}$ via a {\em framing} of $\mathcal{A},$ by associating at each point $\mathbf{a}(s)$ a unit vector $\mathbf{u}(s)$ perpendicular to the tangent ($\mathbf{t}\cdot\mathbf{u} = 0$). Then we define \begin{equation} \mathbf{b}(s) = \mathbf{a}(s) + \varepsilon \mathbf{u}(s) \label{eq:bdef} \end{equation} for $\varepsilon$ arbitrarily small, ensuring that the ribbon nowhere intersects itself. $Tw$ may now be defined formally, as the integral around the curve $\mathcal{A}$ of the rate of rotation of $\mathbf{u}$ about $\mathbf{t}:$ \begin{equation} Tw = \frac{1}{2\pi} \int_{\mathcal{A}} \rd s \, (\mathbf{t} \times \mathbf{u}) \cdot \dot{\mathbf{u}}. \label{eq:twdef} \end{equation} (We adopt the convention of writing the integrand at the end of an integration throughout.) $Tw$ is local in the sense that it is an integral of quantities defined only by $s$ on the curve, and clearly depends on the choice of framing (ribbon). A simple example of a framing is the {\em Frenet framing}, where $\mathbf{u}(s)$ is the direction of the normal vector $\dot{\mathbf{t}}(s)$ to the curve (where defined). This framing plays no special role in the following. As described in the introduction, the crossings in any projection direction can be determined to be either nonlocal (associated with writhe) or local. Choosing an observation direction $\mathbf{o},$ there is a local crossing at $\mathbf{a}(s)$ (the ribbon appears edge-on), when $\mathbf{o}$ is linearly dependent on $\mathbf{t}(s)$ and $\mathbf{u}(s),$ i.e. there exists some $\theta$ between 0 and $\pi$ such that \begin{equation} \mathbf{o} = \mathbf{t} \cos \theta + \mathbf{u} \sin \theta \qquad \hbox{at a local crossing.} \label{eq:local} \end{equation} We therefore define the vector \begin{equation} \mathbf{v}(s, \theta) = \mathbf{t}(s) \cos \theta + \mathbf{u}(s) \sin \theta \label{eq:vdef} \end{equation} dependent on parameters $s,$ labelling a point of $\mathcal{A},$ and angle $\theta$ with $0 \le \theta \le \pi.$ We claim that twice $Tw$ is the average, over all projection directions $\mathbf{o},$ of the local crossing number. For each $\mathbf{o},$ the local crossing number is defined as the number of coincidences of $\mathbf{v}$ with $\mathbf{o}$ or $-\mathbf{o}.$ $Tw$ itself (rather than its double) can therefore be determined by just counting coincidences of $\mathbf{v}$ with $\mathbf{o}$ (not $-\mathbf{o}$). The sign of the crossing is determined by which way the tangent plane of the ribbon at $s$ sweeps across $\mathbf{o}$ as $s$ passes through the edge-on position (which side of the ribbon is visible before the crossing and which after). Thus the vector ($\partial_{\theta} \mathbf{v} \times \partial_s \mathbf{v}$) is either parallel or antiparallel to $\mathbf{v},$ and this decides the crossing sign: i.e. the sign of $(\partial_{\theta} \mathbf{v} \times \partial_s \mathbf{v})\cdot \mathbf{v}$ is minus the sign of the crossing. This sign is opposite to the usual crossing number (described, for instance, in the next section), because $\theta$ increases in the opposite direction to $s$ along the ribbon. The average over all projection directions $\mathbf{o}$ can be replaced by an integral over $s$ and $\theta,$ since the only projection directions which count are those for which there exist $s, \theta$ such that $\mathbf{o} = \mathbf{v}(s,\theta).$ This transformation of variables gives rise to a jacobian factor of $|(\partial_{\theta} \mathbf{v} \times \partial_s \mathbf{v})\cdot\mathbf{v}|$ (the modulus of the quantity whose sign gives the crossing sign). Our claim (justified in the following), that $Tw$ is the spherical average of crossing numbers, is therefore \begin{equation} Tw = \frac{1}{4\pi} \int_{\mathcal{A}} \rd s \int_0^{\pi} \rd \theta \, (\partial_{\theta} \mathbf{v} \times \partial_s \mathbf{v})\cdot\mathbf{v}. \label{eq:twdef1} \end{equation} The two expressions for twist, equations (\ref{eq:twdef}) and (\ref{eq:twdef1}), are equal; this can be seen by integrating $\theta$ in equation (\ref{eq:twdef1}), \begin{multline} \frac{1}{4\pi} \int_0^{\pi} \rd \theta \, (\partial_{\theta} \mathbf{v} \times \partial_s \mathbf{v})\cdot\mathbf{v} \\ \begin{aligned} &= \frac{1}{4\pi} \int_0^{\pi} \rd \theta \,((\mathbf{t} \cos \theta + \mathbf{u} \sin \theta) \times (-\mathbf{t} \sin \theta + \mathbf{u} \cos \theta) ) \cdot (\dot{\mathbf{t}} \cos \theta + \dot{\mathbf{u}} \sin \theta) \\ &= \frac{1}{4\pi} \int_0^{\pi} \rd \theta \, (\cos^2 \theta + \sin^2 \theta) (\mathbf{t} \times \mathbf{u})\cdot (\dot{\mathbf{t}} \cos \theta + \dot{\mathbf{u}} \sin \theta) \\ &= \frac{1}{2\pi} (\mathbf{t} \times \mathbf{u})\cdot\dot{\mathbf{u}}, \end{aligned} \label{eq:twder} \end{multline} which is the integrand of equation (\ref{eq:twdef}). Thus the two expressions for twist, the conventional one (equation (\ref{eq:twdef})) and the local crossing count averaged over viewing directions (equation (\ref{eq:twdef1})), are the same. The preceding analysis of $Tw$ is reminiscent of the more abstract analysis of Pohl (1980), albeit with a different interpretation. Using the notion of direction-averaged crossing numbers, the C\u{a}lug\u{a}reanu theorem follows immediately, as we now explain. For a sufficiently narrow ribbon, it is straightforward to decompose the crossings between different edges into local and nonlocal for each projection of the ribbon. It is well known (Fuller (1971, 1978), Adams (1994)) that writhe $Wr$ equals the sum of signed nonlocal crossings, averaged over direction, i.e. twice the average of self-crossings of the axis curve with itself (a proof is provided in the next section). The average of local crossings (between different edges, counting with respect to both $\mathbf{o}$ and $-\mathbf{o}$), has been shown to equal twice the twist $Tw$ defined in equation (\ref{eq:twdef}). The sum of local plus nonlocal crossings is independent of the choice of projection direction, and is twice the linking number $Lk$ of the two curves. So, averaging the crossings over the direction sphere, $Lk = Tw + Wr.$ Any ambiguity as to whether a crossing is local or nonlocal arises only when the projection direction $\bf{o}$ coincides with the (positive or negative) tangent $\pm \mathbf{t}.$ However, the set of such projection directions is only one-dimensional (parameterised by $s$), and so does not contribute (has zero measure) to the total two-dimensional average over the direction sphere. \section{Formalism for writhe and linking number}\label{sec:formalism} We include the present section, which reviews known material (e.g. Fuller 1971, 1978; Adams 1994; Kauffman 2001; Hannay 1998), to provide some formal geometrical tools for the next section, as well as providing further insight into the proof from the previous section. The linking number $Lk$ between the curves $\mathcal{A}$ and $\mathcal{B}$ is related to the system of (normalised) {\em cross chords} \begin{equation} \mathbf{c}_{\mathcal{A}\mathcal{B}}(s,s') = \frac{\mathbf{a}(s) - \mathbf{b}(s')}{| \mathbf{a}(s) - \mathbf{b}(s') |}. \label{eq:cabdef} \end{equation} $Lk$ is represented mathematically using Gauss's formula: \begin{eqnarray} Lk & =& \frac{1}{4\pi} \int_{\mathcal{A}} \rd s \int_{\mathcal{B}} \rd s' \frac{( \dot{\mathbf{a}}(s) \times \dot{\mathbf{b}}(s') )\cdot(\mathbf{a}(s) - \mathbf{b}(s'))}{| \mathbf{a}(s) - \mathbf{b}(s') |^3} \nonumber \\ & =& \frac{1}{4\pi} \int_{\mathcal{A}} \rd s \int_{\mathcal{B}} \rd s' \, (\partial_s \mathbf{c}_{\mathcal{A}\mathcal{B}} \times \partial_{s'} \mathbf{c}_{\mathcal{A}\mathcal{B}})\cdot \mathbf{c}_{\mathcal{A}\mathcal{B}}. \label{eq:lk} \end{eqnarray} This bears some similarity to the formula for twist in equation (\ref{eq:twdef}). $Lk$ is invariant with respect to reparameterization of $s$ and $s',$ and, of course, any topological deformation avoiding intersections. The domain of integration in equation (\ref{eq:lk}), $\mathcal{A} \times \mathcal{B},$ is the cross chord manifold (secant manifold) of pairs of points on the two curves, topologically equivalent to the torus. The mapping $\mathbf{c}_{\mathcal{A}\mathcal{B}}(s,s')$ takes this torus smoothly to the sphere of directions, with the torus `wrapping around' the sphere an integer number of times (the integer arises since the cross chord manifold has no boundary, and the mapping is smooth); the wrapping is a two-dimensional generalization of the familiar `winding number' of a circle around a circle. This wrapping integer, the integral in equation (\ref{eq:lk}), is the linking number of the two curves (it is the degree of the mapping; see, for example, Madsen \& Tornhave (1997), Epple (1998)). It is easy to see that this interpretation of $Lk$ agrees with that defined earlier in terms of crossings. Choosing the observation direction $\mathbf{o},$ the crossings in the projection are precisely those chords $\mathbf{c}_{\mathcal{AB}}(s,s')$ coinciding with $\mathbf{o}$ where the sign of the scalar triple product $(\partial_s \mathbf{c}_{\mathcal{A}\mathcal{B}} \times \partial_{s'} \mathbf{c}_{\mathcal{A}\mathcal{B}})\cdot \mathbf{c}_{\mathcal{A}\mathcal{B}}$ gives the sign of the crossing. As with $Tw,$ a crossing is only counted in the integral when $\mathbf{c}_{\mathcal{AB}}(s,s')$ is parallel to $\mathbf{o}$ (not antiparallel). Since the cross chord manifold is closed (has no boundary), the total sum of signed crossings does not depend on the choice of $\mathbf{o}.$ \begin{figure} \begin{center} \includegraphics*[width=11cm]{writhes.eps} \end{center} \caption{Illustrating the writhe mesh construction. a) The curve on the torus $(-0.3 \cos(2s), \cos(s)(1+0.3 \sin(2s)), \sin(s)(1+0.3 \sin(2s)).$ The points on the curve are represented by different colours on the colour wheel. b) The writhe mesh for the curve in a). The black lines are the tangent curves $\pm \mathbf{t}(s),$ and the coloured lines are the chord fans $\mathcal{C}_s,$ whose colours correspond to the points $s$ on the curve in a). Note that the total writhe in this example is less than $4\pi$ (i.e. the writhe mesh does not completely cover the direction sphere).} \label{fig:writhe} \end{figure} Writhe has a similar interpretation to link. We define the chords between points of the same curve $\mathcal{A},$ \begin{equation} \mathbf{c}_{\mathcal{A}}(s,s') = \frac{\mathbf{a}(s) - \mathbf{a}(s')}{| \mathbf{a}(s) - \mathbf{a}(s') |}, \label{eq:cadef} \end{equation} noting that, as $s'\to s$ from above, $\mathbf{c}_{\mathcal{A}} \to \mathbf{t}(s),$ and as $s'\to s$ from below, $\mathbf{c}_{\mathcal{A}} \to -\mathbf{t}(s).$ The writhe is the total area on the direction sphere traversed by the vector as $s$ and $s'$ are varied; this two-dimensional surface embedded on the sphere will be referred to as the {\em writhe mesh}: \begin{eqnarray} Wr &=& \frac{1}{4\pi} \int_{\mathcal{A}} \rd s \int_{\mathcal{A}} \rd s' \frac{( \dot{\mathbf{a}}(s) \times \dot{\mathbf{a}}(s') )\cdot(\mathbf{a}(s) - \mathbf{a}(s'))}{| \mathbf{a}(s) - \mathbf{a}(s') |^3} \nonumber \\ &=& \frac{1}{4\pi} \int_{\mathcal{A}} \rd s \int_{\mathcal{A}} \rd s' \, (\partial_s \mathbf{c}_{\mathcal{A}} \times \partial_{s'} \mathbf{c}_{\mathcal{A}})\cdot \mathbf{c}_{\mathcal{A}}. \label{eq:wr} \end{eqnarray} The interpretation of crossings applies to the integral for writhe as well as link and twist; however, for writhe, the visual crossings are uniquely associated with a chord (the chord from $s$ to $s',$ and its reverse, each have separate associated viewing directions). Thus, for writhe, each crossing is counted exactly once, and the average over projections is not divided by 2 (unlike link and twist). (However, as shown in figure \ref{fig:loop}, at every nonlocal crossing of curve $\mathcal{A}$ with itself, there are two crossings (of the same sign) with the other curve.) Unlike the cross chord manifold $\mathbf{c}_\mathcal{AB},$ the writhe mesh $\mathbf{c}_{\mathcal{A}}$ has a boundary: it is topologically equivalent to an annulus (i.e. a disk with a hole). The two boundary circles map to the tangent indicatrix curves (i.e. the loops $\pm \mathbf{t}(s)$ on the direction sphere, as $s$ varies). For each $s,$ the following locus on the sphere, referred to as the {\em chord fan} \begin{equation} \mathcal{C}_s = \mathbf{c}_{\mathcal{A}}(s,s') \quad \hbox{(varying $s'$ from $s$ round to $s$ again)}, \label{eq:csetdef} \end{equation} follows the directions of all the chords to the point $s,$ starting in the positive tangent direction $+\mathbf{t}(s),$ and ending at its antipodal point $-\mathbf{t}(s).$ Topologically, it is a `radial' line joining the two edges of the writhe mesh annulus, although on the direction sphere it may have self-intersections. In addition to the partial covering bounded by the tangent indicatrix curves $\pm\mathbf{t}(s),$ the total writhe mesh may cover the direction sphere an integer number of times. The signed number of crossings, as a function of viewing direction $\mathbf{o},$ changes (by $\pm2$) when $\mathbf{o}$ crosses $\pm \mathbf{t}(s).$ The writhe integral (equation (\ref{eq:wr})) is its average value over all viewing directions. The writhe mesh construction is illustrated in figure \ref{fig:writhe}. The approach of section \ref{sec:proof} may also naturally be interpreted topologically on the direction sphere. The vector $\mathbf{v}(s,\theta),$ dependent on two parameters, also defines a mesh on the direction sphere, the {\em twist mesh}. For fixed $s,$ the locus of points on the twist mesh is a semicircle, \begin{equation} \mathcal{S}_s = \mathbf{v}(s,\theta) \qquad \hbox{($0 \le \theta \le \pi$)} \label{eq:ssetdef} \end{equation} whose endpoints are at $\pm\mathbf{t}(s)$ and whose midpoint is the framing vector $\mathbf{u}(s).$ As $s$ varies around the curve, the semicircle $\mathcal{S}_s$ sweeps out the solid angle $Tw.$ Like the writhe mesh, the twist mesh is topologically an annulus, with $\mathcal{S}_s$ the `radial' line labelled by $s;$ its boundary is again the set of tangent directions $\pm \mathbf{t}(s).$ Therefore, a topological visualization of $Tw + Wr$ is the union of the meshes for twist and writhe; each mesh has the same boundary, so the closure of the union is the join of two annuli along their boundary, topologically a torus. This torus therefore wraps around the sphere the same number of times (has the same degree) as the cross chord manifold $\mathbf{c}_{\mathcal{AB}}.$ It is therefore possible to interpret the loop $\mathcal{L}_s$ on the manifold of cross chords $\mathcal{A}\times\mathcal{B},$ labelled by $s$ on $\mathcal{A}$ (varying $s'$ on $\mathcal{B}$), in terms of the twist semicircle $\mathcal{S}_s$ and the chord fan $\mathcal{C}_s.$ The loop $\mathcal{L}_s,$ mapped to the direction sphere, is the set of cross chords directions between fixed $\mathbf{a}(s)$ and all $\mathbf{b}(s')$ on $\mathcal{B}.$ When the ribbon is vanishingly thin, for $s'$ outside the neighbourhood of $s,$ the chords $\mathbf{c}_{\mathcal{AB}}(s,s')$ can be approximated by the chords of $\mathbf{c}_{\mathcal{A}}(s,s').$ When $s'$ is in the neighbourhood of $s,$ $\mathbf{c}_{\mathcal{AB}}(s,s')$ is approximated by $\mathbf{v}(s,\theta)$ for some $\theta$ (exact at $\theta = \pi/2$ when $s' = s$). As $\varepsilon \to 0,$ these approximations improve, and $\mathcal{L}_s$ approaches the union of $\mathcal{C}_s$ and $\mathcal{S}_s.$ Link is therefore the area on the direction sphere swept out by the loop union of $\mathcal{C}_s$ and $\mathcal{S}_s$ for $s$ varying around the curve; since this family of loops generates the closed torus, the direction sphere is enveloped an integer number of times. Thus, C\u{a}lug\u{a}reanu's theorem may be interpreted as a natural decomposition of the integrand in Gauss's formula (\ref{eq:lk}) in terms of the writhe mesh and twist mesh. This is very close, in a different language, to White's proof (White 1969, also see Pohl 1968{\em a}, and particularly Pohl 1980), where objects analogous to the writhe mesh and the twist mesh appears as boundaries to a suitably regularised (blown-up) 3-manifold of chords from $\mathcal{A}$ to points on surface of the ribbon with boundary $\mathcal{A}, \mathcal{B}.$ \section{The writhe framing}\label{sec:writhe} In this section, we use the description of $Tw$ and $Wr$ to define, for any non-self-intersecting closed curve in space, a natural framing (i.e. ribbon) whose linking number is zero. Such a framing is useful since it can be used as a reference to determine the linking number for any other framing: if $\mathbf{u}_0(s)$ represents this zero framing, and $\mathbf{u}(s)$ any other framing with linking number $Lk,$ then \begin{equation} Lk = \frac{1}{2\pi} \int_{\mathcal{A}} \rd s\, \arccos \mathbf{u}\cdot\mathbf{u}_0. \label{eq:zerolinkref} \end{equation} Clearly the twist of a zero framing is equal to minus the writhe by C\u{a}lug\u{a}reanu's theorem, and therefore we will refer to our natural framing as the {\em writhe framing}. Expressing $Lk$ in terms of an integral involving the difference between two framings is reminiscent of C\u{a}lug\u{a}reanu's original proof (C\u{a}lug\u{a}reanu 1959, 1961; also see Moffatt \& Ricca 1992), in which the twist is defined as the total number of turns the framing vector makes with respect to the Frenet framing around the curve (provided the normal to the curve is everywhere defined). The total twist is therefore the sum of this with the integrated torsion around the curve (i.e. $Tw$ of the Frenet framing). As stated before, the Frenet framing plays no role in our construction. It is easy to construct an unnatural zero framing by taking any framing, cutting anywhere, and rejoining with a compensating number of twists locally. In contrast, the framing we construct here is natural in the canonical sense that, at any point $s,$ the definition of $\mathbf{u}_0(s)$ and its corresponding semicircle $\mathcal{S}_s$ (defined in equation (\ref{eq:ssetdef})) depends only on the view from $s$ of the rest of the closed curve, that is on the chord fan $\mathcal{C}_s$ (defined in equation (\ref{eq:csetdef})). The rate of rotation of the resulting $\mathbf{u}_0$ exactly compensates the corresponding writhe integrand, that is, the integrands (with respect to $s$) of equation (\ref{eq:wr}) and the twist of the writhe framing (equation (\ref{eq:twdef})) are equal and opposite. \begin{figure} \begin{center} \includegraphics*[width=12.15cm]{framings.eps} \end{center} \caption{Illustrating the writhe framing construction, using the same example curve as figure \ref{fig:writhe}. a) The chord fan for $s = 0,$ is represented in red, the semicicle which bisects its area represented in pink. The tangent curves are also represented. b) The same as part a), and in addition the chord fan with $s = -0.4$ is shown in dark blue, and its bisecting semicircle in light blue. No area is swept out by these closed curves as $s$ evolves.} \label{fig:framing} \end{figure} This may be interpreted topologically as follows. $\mathcal{C}_s$ and $\mathcal{S}_s$ correspond topologically to `radial' lines of their respective annular meshes; their union is a closed loop on the direction sphere. The sum of the twist and writhe integrands, by the discussion in section \ref{sec:proof}, is the area swept out by this changing closed loop as $s$ develops; it was proved that the total area swept out by this loop is $4\pi Lk.$ The writhe framing construction of $\mathbf{u}_0$ below arranges that the loop has a constant area (say zero), and therefore the rate of area swept by this loop is zero as $s$ evolves, giving zero total area swept (which is $4\pi Lk$). The direction of $\mathbf{u}_0(s)$ for the writhe framing is that for which the semicircle $\mathcal{S}_s$ bisects the chord fan $\mathcal{C}_s$ in the following sense. Since $\mathcal{C}_s$ is a curve on the direction sphere with endpoints $\pm\mathbf{t}(s),$ it may be closed by an arbitrary semicircle with the same endpoints (i.e. any rotation of $\mathcal{S}_s$ about $\pm\mathbf{t}(s)$). The total closed curve encloses some area on the sphere (mod 4$\pi$); the (unique) semicircle which gives zero area (mod $4\pi$) defines $\mathbf{u}_0.$ Since the spherical area enclosed by this closed curve is contant, the curve does not sweep out any area as it evolves (since as much leaves as enters). An example of the writhe framing is represented in figure \ref{fig:framing}. The rate of area swept out by this curve is the integrand with respect to $s$ in the $Lk$ expression (equation (\ref{eq:lk})), (i.e. the sum of the $Tw, Wr$ integrands). Since it has been shown that this is zero, the writhe framing $\mathbf{u}_0$ indeed has zero linking number. Of course, although the choice of zero area is most natural, the writhe faming vector $\mathbf{u}_0$ could be defined such that the area between the chord fan $\mathcal{C}_s$ and the twist semicircle $\mathcal{S}_s$ is any fixed value - the important feature in the construction is that the area does not change with $s.$ \begin{acknowledgements} We are grateful to John Maddocks for originally pointing out to us the problem of the writhe framing, and useful correspondence. MRD acknowledges support from the Leverhulme Trust and the Royal Society of London. \end{acknowledgements}
{ "timestamp": "2005-06-10T11:48:58", "yymm": "0503", "arxiv_id": "math-ph/0503012", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503012" }
\section{Introduction} The Weizs\"acker-Bethe mass formula has played an important role for many years. It offers a guideline for developing modern mass formulas. Indeed, modern mass formulas, such as the droplet mass formula (FRDM) \cite{Moller} (excluding the Hartree-Fock approaches \cite{Aboussir,Goriely}) largely retain the original form. They include the pairing and symmetry energy terms, in addition to the volume, surface and Coulomb energy terms. While the pairing energy has been investigated by combining with the microscopic pairing correlations, the symmetry energy has not necessarily been investigated from the point of view of the shell model. A recent detailed work \cite{Duflo} succeeded in obtaining a precise mass formula. That work is based on the rigorous microscopic guidelines given in Ref. 5), which considers the monopole field and pairing structure providing the dominant terms of the mass formula. The study presented in Ref. 5) considers general properties of the shell model Hamiltonian. However, while the pairing interaction as the origin of the pairing energy determines the structure of wave functions, the symmetry energy is not treated symmetrically with the pairing energy. In the present paradigm for the nuclear mass formula, the symmetry energy is regarded as a basic concept. In the shell model, however, there is no approach other than describing the symmetry energy in terms of nuclear correlations. The asymmetrical treatment of the symmetry energy and pairing energy leaves a missing link in relating mass formulas to nuclear structure. This paper proposes an alternative approach, in which the ``symmetry energy" is not treated as a fundamental concept and explains the symmetry energy in terms of certain correlations, as the pairing energy is determined in terms of the pair correlations. We wish to understand the symmetry energy derived from the mean field theory from the point of view of the shell model. Our treatment begins from the $jj$ coupling shell model based on a $Z=N$ doubly-closed shell core, and we do not discuss Strutinski's prescription \cite{Strut}. The purpose of this paper is not to give a new mass formula better than modern sophisticated mass formulas, but to present a useful understanding of the mass formula. We therefore start from an old fashioned mass formula in order to clearly show the basic idea. In the $jj$ coupling shell model with an effective interaction, the energy depending on the total isospin $T$ comes from the interactions between valence nucleons in $j$ orbits. The corresponding correlations are not yet reduced to a mean field but determine wave functions or structure of nuclei, in the shell model description. We consider such correlations in even-even $N=Z$ ($A_0+m\alpha$) nuclei that give no contribution of the symmetry energy. Here, $A_0$ represents a doubly-closed-shell core, $\alpha$ is a two-neutron-two-proton ($2n-2p$) quartet with $T=0$, and $m$ is an integer. We show that the interaction energies of the $A_0+m\alpha$ nuclei characterize the binding energies of $N \approx Z$ nuclei, excluding the bulk energy depending on mass $A$. The strength of the correlations can be evaluated in terms of the difference between the mass of an $A_0+m\alpha +2n$ ($A_0+m\alpha +2p$) nucleus and the average mass of its neighboring nuclei with $A=A_0+m\alpha$ and $A=A_0+(m+1)\alpha$. With this indicator, Gambhir, Ring and Schuck \cite{Gamb} studied a superfluid state of many $\alpha$ {\it particles}. The term $\alpha$, however, represents only a $T=0$ $2n-2p$ quartet, not the spatial $\alpha$ cluster. We call the correlations ``$T=0$ $2n-2p$ correlations" in the sense of many-body correlations. (We use ``$\alpha$-like" as a concise term for the superfluid state.) This paper shows that the symmetry energy is derived from the nonparticipation of redundant nucleons in the $T=0$ $2n-2p$ correlations, in parallel with the pairing energy derived from the nonparticipation of an odd nucleon in the $T=1$ pair correlations. The energy of valence nucleons is separated from the binding energy, and the leading role of $T=0$ $2n-2p$ correlations is discussed in $\S$2. Section 3 discusses the fundamental $T=0$ $2n-2p$ correlated structure in $N \approx Z$ nuclei, (which is called ``$\alpha$-like superfluidity"). Section 4 explains the mass differences between even-even nuclei in terms of multi-pair structure on the base of $\alpha$-like superfluidity. In $\S$5, we discuss how the pairing energy should be evaluated. Section 6 gives concluding remarks. \section{Correlations of valence nucleons buried in the binding energy} \subsection{Extraction of the energy of valence nucleons} In the old fashioned mass formula, the bulk of the binding energy is written in terms of the volume, surface and Coulomb energies as \begin{equation} B_{VSC}(A) = - a_V A + a_S A^{2/3} + a_C Z^2 / A^{1/3}. \label{eq:1} \end{equation} We can consider that these main terms basically represent a nuclear potential in the shell model picture, while the other terms of the mass formula are related to the shell model interactions. It must be stressed that the symmetry energy depending on the total isospin $T$ is attributed to the shell model interactions in this picture. Let us estimate the interaction energy by subtracting $B_{VSC}(A)$ from the experimental binding energy $B(A)$ \cite{Audi} for $A_0+m\alpha$ nuclei with $T=0$. (Note that the sign of the binding energy $B(A)$ is negative in this paper.) The values $B(A)-B_{VSC}(A)$ calculated with a few mass formulas with simple forms \cite{Duflo,Yagi,Ring,Samanta} are listed in Table \ref{table1}. [In the third line, we used the six-parameter mass formula in Ref. 4). The Coulomb energy term in Refs. 4) and 11) is expressed in terms of different functions of the proton number $Z$.] In Table \ref{table1}, we tabulate the values $B(A)-B_{VSC}(A)$ for seven $A=A_0+m\alpha$ nuclei with $T=0$, where the symmetry energy makes no contribution. These values display variation depending on $A$. \begin{table}[b] \caption{The values of $B(A)-B_{VSC}(A)$ for $A_0+m\alpha$ nuclei with $T=0$, calculated using a few mass formulas.} \begin{center} \begin{tabular}{c|rrrrrrr} \hline ref. & $^{16}$O & $^{20}$Ne & $^{28}$Si & $^{40}$Ca & $^{44}$Ti & $^{56}$Ni & $^{64}$Ge \\ \hline \cite{Yagi} & $-5.81$ & $-2.13$ & $-4.82$ & $-2.88$ & $-1.35$ & $-7.99$ & $-4.63$ \\ \cite{Ring} & $-12.80$ & $-10.21$ & $-14.92$ & $-15.79$ & $-15.17$ & $-24.50$ & $-22.94$ \\ \cite{Duflo} & $-4.85$ & $-0.87$ & $-2.97$ & $-0.15$ & $ 1.67$ & $-4.10$ & $-0.17$ \\ \cite{Samanta} & $-7.21$ & $-3.52$ & $-6.08$ & $-3.85$ & $-2.21$ & $-8.49$ & $-4.91$ \\ \hline \end{tabular} \end{center} \label{table1} \end{table} Table \ref{table1} shows that the mass formula \cite{Ring} fitted for heavy nuclei is not good for $N \approx Z$ nuclei but that the other three display parallel and interesting behavior (dips at $^{28}$Si and $^{56}$Ni) as $A$ increases. According to ordinary mass formulas, there remains the pairing term $\delta_P$, which contributes to the $T=0$ even-even nuclei under consideration. However, the deviations of the experimental binding energies from $B_{VSC}$ shown in Table \ref{table1} are much larger than the pairing effect. The characteristic behavior of $B(A_0+m\alpha)-B_{VSC}(A_0+m\alpha)$ cannot be explained as a simple variation of $\delta_P$ depending on $A$, like $\delta_P \propto A^p$. The existing mass formulas do not describe the behavior. The characteristic behavior of $B(A_0+m\alpha)-B_{VSC}(A_0+m\alpha)$ must reflect correlations stronger than the pairing correlations from the point of view of the shell model. This is worth investigating further. Let us start from the old simple mass formula \cite{Yagi} with the parameters $a_V=15.56$, $a_S=17.23$ and $a_C=0.6986$ in MeV. It is noticed in Table \ref{table1} that the values in the first line \cite{Yagi} resemble those in the fourth line obtained with the mass formula \cite{Samanta}, and the mass formula \cite{Duflo} has an elaborate form, so that the deviations $B(A_0+m\alpha)-B_{VSC}(A_0+m\alpha)$ may be small. The values in Table \ref{table1} indicate the insufficiency of $B_{VSC}$ for the doubly-closed-shell nuclei $^{16}$O and $^{40}$Ca, which are the bases for the shell model calculations. We suppose that the deviations in $^{16}$O and $^{40}$Ca require adjustments of the depth of the shell model potential. The adjustment parameter $\delta_P$ for even-even nuclei has the form $A^{-3/4}$ in the old convention. If we adopt the $A^{-3/4}$ adjustment for the potential depth, we can fix its parameter so as to make the deviations $B(A_0+m\alpha)-B_{VSC}(A_0+m\alpha)$ nearly zero for $^{16}$O and $^{40}$Ca. We assume that the main part of the mass formula corresponding to the shell model potential is approximated by \begin{eqnarray} & {} & B_0(A) = B_{VSC}(A) + \delta U_{pot}(A), \\ \label{eq:2} & {} & \delta U_{pot}(A) = - 46.4 / A^{3/4}. \label{eq:3} \end{eqnarray} The deviation $B(A)-B_0(A)$ could be regarded as the energy of valence nucleons outside a doubly-closed-shell core, \begin{equation} E(A) = B(A) - B_0(A). \label{eq:4} \end{equation} The energies $E(A_0+m\alpha)$ for the even-even $N=Z$ nuclei with $T=0$ are plotted in Fig. \ref{fig1}, which displays the characteristic behavior of the binding energies $B(A_0+m\alpha)$ mentioned above. In the shell model calculation, the experimental energy of correlated valence nucleons outside a doubly-closed-shell core $A_0=(N_0,Z_0)$ is evaluated using \begin{equation} E_{shl}(N,Z) = B(N,Z) - B(N_0,Z_0) - \lambda (A-A_0) - \Delta E_C(N,Z), \label{eq:5} \end{equation} where $\Delta E_C(N,Z)$ is a Coulomb energy correction for the valence nucleons. For instance, the correction $\Delta E_C(N,Z) = p(Z-Z_0)+q(Z-Z_0)(Z-Z_0-1)+r(Z-Z_0)(N-N_0)$ was used by Caurier et al. \cite{Caurier} in the shell model calculations for $f_{7/2}$ shell nuclei. We calculated the energy $E_{shl}(N,Z)$ using the parameter values $p=7.279$, $q=0.15$, $r=-0.065$ and $\lambda =-12.45$ (in MeV) for the $pf$ shell nuclei and $p=3.54$, $q=0.20$, $r=0.0$ and $\lambda =-11.2$ (in MeV) for the $sd$ shell nuclei. The calculated values are indicated by the dotted curves in Fig. \ref{fig1}. We can see that $E(A_0+m\alpha) \approx E_{shl}(A_0+m\alpha)$ in the first half of the shells, which supports our assumption that $E(A)$ in Eq. (\ref{eq:4}) represents the energy of valence nucleons. The disagreement between $E(A_0+m\alpha)$ and $E_{shl}(A_0+m\alpha)$ is large in the latter half of the shells. (Note that the disagreement is smaller in the heavier $pf$ shell nuclei.) The hole picture would be better for the latter half of the shells. \begin{figure}[t] \begin{center} \includegraphics[width=7cm,height=7cm]{fig1.eps} \caption{The energies of $E(A_0+m\alpha)=B(A_0+m\alpha)-B_0(A_0+m\alpha)$ and the curve connecting them, which defines the $T=0$ base level.} \label{fig1} \end{center} \end{figure} \subsection{The leading role of the $T=0$ $2n-2p$ correlations} From the above, we find that the energy $E(A)$ defined by Eq. (\ref{eq:4}) approximately represents the total energy of valence nucleons outside the doubly-closed-shell core $^{16}$O or $^{40}$Ca. To obtain a guide for the discussion given in the following sections, let us consider the main features of $E(A)$ near $^{40}$Ca, where $E(A)$ corresponds well with the shell model energy $E_{shl}(A)$. Figure \ref{fig2} depicts the ground state energies $E(A)$ of $^{40}$Ca, an $A=41$ system with $T=1/2$, an $A=42$ system with $T=1$, $^{44}$Ti and $^{48}$Cr. (This figure is similar to the diagram for the pairing vibrations.) The energy $E(A)$ in Eq. (\ref{eq:4}) does not give exactly the same energy to $^{41}$Ca and $^{41}$Sc. A correction term representing the Coulomb energy is necessary in the final stage. In Fig. \ref{fig2}, we show the average energy of $E(^{41}{\rm Ca})$ and $E(^{41}{\rm Sc})$ for the $A=41$ system with $T=1/2$. Similarly, we show the average energy of $E(^{42}{\rm Ca})$ and $E(^{42}{\rm Ti})$ [which is approximately equal to $E(^{42}{\rm Sc})$] for the $A=42$ system with $T=1$. The energy $E(A=41)$ represents an effective single-particle energy $e_{sp}$ in the nuclear potential represented by $B_0(A)$, whose depth is adjusted to be zero for $^{40}$Ca. If there are no interactions between valence nucleons, the energy of the $A=40+n_v$ system ($n_v$ being the number of valence nucleons) is $n_v e_{sp}$. However, the real energy $E(A=42)$ lies substantially below the line $n_v e_{sp}$ in Fig. \ref{fig2}. The difference is the pair correlation energy. Half of the absolute pair correlation energy is called the (three-point) odd-even mass difference $\Delta$. The value $\Delta$ is often used as the indicator of the pair correlations. We show the odd-even mass difference $\Delta$ at $A=41$ in Fig. \ref{fig2}. The definition of $\Delta$ given in Eq. (\ref{eq:30}) explains the geometrical relations shown in Fig. \ref{fig2}. \begin{figure}[t] \begin{center} \includegraphics[width=7.5cm,height=7.5cm]{fig2.eps} \caption{Schematic depiction of the pair correlations and $T=0$ $2n-2p$ correlations and their indicators for the nuclei $^{40}$Ca, $A=41$ with $T=1/2$, $A=42$ with $T=1$, $^{44}$Ti and $^{48}$Cr.} \label{fig2} \end{center} \end{figure} The energy of $^{44}$Ti measured from the line $n_v e_{sp}$ is the interaction energy of the four nucleons with $T=0$ outside the $^{40}$Ca core, which is denoted by the dotted line in Fig.~\ref{fig2}. Figure \ref{fig2} shows that the $T=0$ four nucleon correlations, which we call $T=0$ $2n-2p$ correlations, experience an energy gain much larger than that of the pair correlations $2 \Delta$. The strong $T=0$ $2n-2p$ correlations cause the four nucleons to form an $\alpha$-like quartet outside the $^{40}$Ca core, as described by the core plus $\alpha$-cluster model. The study of $\alpha$-like correlations has a long history. \cite{Maru,Danos,Arim,Kamim,Tomo,Cauvin,Duss1,Gamb,Jensen,Apos,Duss2,Curut,Suga,Hase1,Hase1B,Merm,Hase2,Sakuda} (Also, see other references cited in Ref. 28), especially for the core plus $\alpha$-cluster model.) Sometimes the indicator of the $T=0$ $2n-2p$ correlations was evaluated after subtracting the symmetry energy from the binding energy (for instance, see Ref. 20)). As shown below, however, different involvements of valence nucleons in the $T=0$ $2n-2p$ correlations result in the ordering of the ground-state energies of $N \approx Z$ nuclei according to that of the total isospin $T$ (which results in the symmetry energy in the mean field theory). If we limit the comparative study of different $T$ nuclei by excluding the symmetry energy from the binding energy at the beginning, we miss a substantial energy gain due to the underlying $T=0$ $2n-2p$ correlations, and we therefore cannot understand the formation of the $\alpha$-cluster in $^{20}$Ne and $^{44}$Ti in contrast to the absence of the $\alpha$-cluster outside the core in $^{20}$O and $^{44}$Ca. Let $e_s$ denote the energy of the $T=1$ correlated pair and $l$ the number of correlated pairs. If there is no interaction between the $2n$ and $2p$ pairs, the energy of $^{44}$Ti is expected to be on the line $l e_s$. The real energy of $^{44}$Ti lies far below the line $l e_s$ in Fig. \ref{fig2}. The difference is the interaction energy between the $2n$ and $2p$ pairs, which is denoted by the dot-dashed line in Fig. \ref{fig2}. Half of the absolute interaction energy between the $2n$ and $2p$ pairs is a good indicator of the $T=0$ $2n-2p$ correlations. This indicator is called the ``ODD-EVEN mass difference for the $\alpha$-like correlations" by Gambhir, Ring and Schuck \cite{Gamb}, where ODD and EVEN are used for the number of pairs. We define it in Eq. (\ref{eq:13}) and Eq. (\ref{eq:14}), and we write it as $\delta M(A_0+m\alpha +2)$ and $\delta W(A_0+m\alpha +2:T=1)$ in the respective cases. The value $\delta M$ is indicated by the solid line at $A=42$ in Fig. \ref{fig2}. The definitions given in Eq. (\ref{eq:13}) and Eq. (\ref{eq:14}) explain the geometrical relations. The total interaction energy of the $T=0$ $2n-2p$ quartet is $-2(2 \Delta + \delta M)$. Figure \ref{fig2} clearly shows that the ODD-EVEN mass difference $\delta M$ for the $T=0$ $2n-2p$ correlations is much larger than the odd-even mass difference $\Delta$ for the pair correlations. The ODD-EVEN mass difference $\delta M$ for the $A=42$ system with $T=1$, which is measured from the $T=0$ line, reflects the symmetry energy. In other words, the symmetry energy in the framework of the mass formula can be explained in terms of the $T=0$ $2n-2p$ correlations from the point of view of the shell model. It should be noticed that the symmetry energy and pairing energy in the mass formula are treated on the same footing here. Because the $T=0$ $2n-2p$ correlations are very strong, we believe that the $T=0$ $2n-2p$ quartet is approximately a good excitation mode. Let its energy be $e_\alpha$. Then the energy of the $A=40+m \alpha$ system is expected to be nearly $m e_\alpha$. This expectation holds roughly for $^{48}$Cr, as shown in Fig. \ref{fig2}, where $E(^{48}{\rm Cr})$ lies below but near the line $m e_\alpha$. It is shown in Ref. 29) that the $^{48}$Cr nucleus is described quite well by the $^{40}$Ca core plus two $\alpha$-cluster model. Figure \ref{fig1} shows that the energies $E(40+m \alpha)$ are below the line $m e_\alpha$. This result indicates the important point that the interaction between the $T=0$ $2n-2p$ quartets is attractive, and the $T=0$ $2n-2p$ correlations are collective in systems of many quartets. Figure \ref{fig1} shows that $^{56}$Ni in the middle of the $pf$ shell is different from the typical doubly-closed-shell nuclei $^{16}$O and $^{40}$Ca, but it resembles $^{28}$Si in the middle of the $sd$ shell. Because $B_0(A_0+m\alpha)$ as a function of $m$ is monotonic in the regions $A=16-36$ and $A=40-72$ of the $A_0+m\alpha$ nuclei, we cannot attribute the difference between the $^{56}$Ni nucleus and the $^{16}$O and $^{40}$Ca nuclei to special behavior of $B_0(A)$. The difference is due to correlations of valence nucleons or a shell effect. The rigid core of $^{16}$O and $^{40}$Ca is supported by the successful description of $^{20}$Ne and $^{44}$Ti with the core plus $\alpha$-cluster model. Figure \ref{fig1} suggests structure of $^{56}$Ni ($^{28}$Si) that differs from a rigid core. We suppose that $^{16}$O and $^{40}$Ca have rather rigid cores and that the other $N \approx Z$ nuclei are described as systems of correlated valence nucleons outside the respective cores, as is done in ordinary shell model calculations. \subsection{Examination by means of the shell model calculation} Let us examine the above picture by carrying out shell model calculations with a realistic effective interaction in the $pf$ shell nuclei outside the $^{40}$Ca core. The shell model Hamiltonian describing valence nucleons outside the core is composed of the single-particle energy part and the effective interaction: \begin{equation} H = H_{sp} + H_{int}. \label{eq:H1} \end{equation} We adopt the Honma interaction, \cite{Honma} which accurately describes the $pf$ shell nuclei near $^{56}$Ni, and we consider systems in the $jj$ coupling scheme. To compare with Fig. \ref{fig1}, we use the same parameter value $\lambda = -12.45$ MeV, as in Eq. (\ref{eq:5}), though a somewhat different Coulomb energy correction is used in Ref. 30). The adopted single-particle energies are $e(f_{7/2})=3.862$, $e(p_{3/2})=6.7707$, $e(p_{1/2})=8.313$ and $e(f_{5/2})=11.0671$ in MeV. It is useful to decompose the effective interaction $H_{int}$ into the monopole part and the residual part \cite{Dufour}. The $T=0$ monopole field defined by the following equation is especially important, because it determines the main part of the interaction energy (expectation value $\langle H_{int} \rangle$): \begin{eqnarray} & {} & H_{mp}^{T=0} = - k^0 \sum_{a \leq b} \sum_{JM} A^\dagger_{JM00}(ab) A_{JM00}(ab), \\ \label{eq:H2} & {} & k^0 = \frac{\sum_{ab} \overline{V}(ab:T=0)} {\sum_{ab}1} \label{eq:H3} \end{eqnarray} with \begin{equation} \overline{V}(ab:T=0) = \frac{ \sum_J (2J+1) V(abab:J,T=0) }{\sum_J (2J+1)}, \label{eq:H4} \end{equation} where $A^\dagger_{JMTK}(ab)$ is the creation operator of a nucleon pair with spin $JM$ and isospin $TK$ in the single-particle orbits ($a,b$) and $V(abab:JT)$ is a diagonal two-body interaction matrix element. Let us write the effective interaction as \begin{equation} H_{int} = H_{mp}^{T=0} + H_{res} \quad ( H_{res} = H_{int} - H_{mp}^{T=0}). \label{eq:H5} \end{equation} The monopole field $H_{mp}^{T=0}$ is expressed exactly as \begin{equation} H_{mp}^{T=0} = - \frac{k^0}{2} \Big\{ \frac{\hat n_v}{2} \big(\frac{\hat n_v}{2} +1\big) - {\hat T}({\hat T}+1) \Big\}, \label{eq:H6} \end{equation} where ${\hat n_v}$ stands for the number of valence nucleons, and ${\hat T}$ stands for the total isospin. It is well known that realistic effective interactions have large and comparable values of the centroids $\overline{V}(ab:T=0)$. The expression (\ref{eq:H6}) with a large average value $k^0$ (for instance, $k^0=1.44$ MeV for $^{56}$Ni) shows that the symmetry energy comes mainly from $H_{mp}^{T=0}$ with the $T(T+1)$ term \cite{Kaneko}. The monopole field in the form (\ref{eq:H6}) can be regarded as an additional term to the Hartree-Fock mean field, in a sense. However, the residual interaction $H_{res}$, which determines the microscopic structure, contributes significantly to the symmetry energy \cite{Kaneko}. The symmetry energy cannot be reduced to a simple mean field but is affected by dynamical interactions in the shell model. \begin{figure}[b] \begin{center} \includegraphics[width=6.8cm,height=7.5cm]{fig3a.eps} \includegraphics[width=6.8cm,height=7.5cm]{fig3b.eps} \caption{Expectation values of $H_{sp}$, $H_{sp}+H_{mp}^{T=0}$ and $H$ for (a) even-even $N=Z$ nuclei with $A=40+m\alpha$ and (b) odd-$A$ nuclei with $A=40+m\alpha+1n$. The residual interaction energy $\langle H_{res}\rangle$ is denoted by the solid-line arrows, and the monopole contribution $\langle H_{mp}^{T=0} \rangle$ is denoted by the dashed-line arrows.} \label{fig3} \end{center} \end{figure} We carried out numerical calculations using Mizusaki's code, \cite{Mizusaki,Mizusaki2} which makes large-scale shell model calculations possible by means of extrapolation. The calculated results for the $A=40+m\alpha$ nuclei from $^{44}$Ti to $^{64}$Ge are illustrated in Fig. \ref{fig3}(a), where $\langle H_{sp} \rangle$, $\langle H_{sp}+H_{mp}^{T=0} \rangle$ and $\langle H \rangle$ denote their expectation values for the ground state. The behavior of the ground-state energy $\langle H \rangle$ corresponds well with that of the energy $E(A_0+m\alpha)$ seen in Figs.~\ref{fig1} and \ref{fig2}. Figure \ref{fig3}(a) shows that the energy $E(A_0+m\alpha)$ represents the ground-state energies of the even-even $N=Z$ nuclei and, moreover, that $E(A)$ hides significant correlations in the background. This figure supports the schematic explanation for the large energy gains of the $A_0+m\alpha$ nuclei in Fig. \ref{fig2}. Even if we regard the monopole field as a part of the mean field, the residual interaction energy $\langle H_{res} \rangle$ is still large in Fig. \ref{fig3}(a). The residual interaction energy $\langle H_{res} \rangle$ is essential for bringing the values of $E(A_0+m\alpha)$ ($ \approx \langle H \rangle$) close to the zero line. In $^{56}$Ni, for instance, $\langle H_{res} \rangle$ is approximately 20 MeV, which overwhelms the single-particle energy gap between $f_{7/2}$ and $p_{3/2}$. The closed-shell configuration $(f_{7/2})^{16}$ does not exceed 68\% in the wave function of the ground state, according to Ref. 30). The new Tamm-Dancoff solution for the $J=T=0$ four-particle excitation mode indicates that the ground state of $^{56}$Ni cannot be described within a perturbation expansion starting with the closed-shell configuration \cite{Hase3}. This situation is called ``$\alpha$-like superfluidity" in the next section. The $^{56}$Ni nucleus can be regarded as a correlated state of valence nucleons outside the $^{40}$Ca core. Figure \ref{fig3}(a) shows the upward turn of $\langle H \rangle$ from $^{56}$Ni to $^{60}$Zn resulting from an energy loss of four additional nucleons occupying the upper orbits beyond the semi-magic number $Z=N=28$. The variation of $E(A_0+m\alpha)$ is therefore related to a shell effect as well as correlations. However, it should be noted that the position of $\langle H(^{60}\mbox{Zn}) \rangle$ in Fig. \ref{fig3}(a) depends on significant collapse of the $^{56}$Ni core. Figure \ref{fig3}(b) displays the shell model results for odd-mass nuclei with $A_0+m\alpha+1n$. This figure is very similar to Fig. \ref{fig3}(a). The ground-state energy $\langle H \rangle$ exhibits a dip at $^{57}$Ni in Fig. \ref{fig3}(b), like the dip at $^{56}$Ni in Fig. \ref{fig3}(a). In the shell model results for $^{57}$Ni, the occupation probabilities of the respective orbits indicate strong correlations of valence nucleons and collapse of the $^{56}$Ni core, which is contrary to a simple picture of the $^{56}$Ni core plus one neutron. The single-particle energy gap between $f_{7/2}$ and $p_{3/2}$ is negligible as compared with the interaction energy, though the energy loss of four additional nucleons occupying the upper orbits ($p_{3/2}$, $p_{1/2}$, $f_{5/2}$) causes an upward turn of $\langle H \rangle$ from $^{57}$Ni to $^{61}$Zn. The present shell model also explains the behavior of the experimental energy $E(A_0+m\alpha+1n)$ shown in Fig. \ref{fig4}(b). Thus, Figs. \ref{fig3}(a) and \ref{fig3}(b) support our picture, which regards nuclei around $^{56}$Ni as correlated states of valence nucleons outside the $^{40}$Ca core. It is interesting that a $A_0+m\alpha+1n$ nucleus appears to be composed of an $A_0+m\alpha$ system and a last neutron with an effective single-particle energy. We can assume roughly the same correlations forming a common structure in the two nuclei with $A=A_0+m\alpha$ and $A=A_0+m\alpha+1n$. The correlations buried in the energies $E(A)$ of the $A_0+m\alpha$ nuclei have been considered little in the framework of the mass formula. The inconspicuous values of $E(A_0+m\alpha)$, however, hide the $T=0$ $2n-2p$ correlations which are stronger than the $T=1$ pair correlations, in the background. The energy $E(A_0+m\alpha)$ should be considered explicitly in the mass formulas. Figure \ref{fig2} suggests a description of the energies $E(A)$ of $N \ne Z$ nuclei with the indicators $\delta M$ and $\Delta$ on the $T=0$ line connecting the $A_0+m\alpha$ nuclei. We now note the importance of the $T=0$ line in Fig. \ref{fig1}. Ignoring the variation of the $T=0$ line affects the masses of all $N \approx Z$ nuclei. The values of $E(A_0+m\alpha)$ are not zero, and the substantial deviations should not be ignored. \section{Fundamental $T=0$ $2n-2p$ correlated structure} \begin{figure}[t] \begin{center} \includegraphics[width=6.8cm,height=7.4cm]{fig4a.eps} \includegraphics[width=6.8cm,height=7.4cm]{fig4b.eps} \caption{Experimental energies $E(A)$ defined in Eq. (\ref{eq:4}) for (a) even-even nuclei with $N \ge Z$ and (b) nuclei with $A_0+m\alpha$, $A_0+m\alpha+1n$ and $A_0+m\alpha+1n1p$ around $^{56}$Ni.} \label{fig4} \end{center} \end{figure} We have seen significant correlations buried in $E(A)$ of $A_0+m\alpha$ and $A_0+m\alpha+1n$ nuclei. What is the nature of the correlations in other $N \neq Z$ nuclei? To answer this, let us consider Fig.~\ref{fig4}(a), in which experimental values of $E(A)$ are plotted for even-even nuclei with $N \ge Z$. Even-even nuclei with $N > Z$ can be classified according to the number of neutron pairs added to the $A_0+m\alpha$ systems, such as $A=A_0+m\alpha +2n$, $A=A_0+m\alpha +4n$, $\cdots$, and each series of them with increasing $m$ has the same $T$. Figure \ref{fig4}(a) indicates a parallelism with regard to $E(A)$ between even-even nuclei with $T>0$ and $T=0$. More precisely, every $T$ line connecting a series of nuclei is parallel to the $T=0$ line of the $A_0+m\alpha$ nuclei. The same parallelism is seen in the experimental values of $E(A)$ for even-even $N \le Z$ nuclei ($A=A_0+m\alpha$, $A=A_0+m\alpha +2p$, $A=A_0+m\alpha +4p$, $\cdots$), though they are omitted in Fig. \ref{fig4}(a) for simplicity. The experimental energy $E(A)$ for the $T=1$, $0^+$ states of odd-odd $N=Z$ nuclei also varies in a manner parallel to $E(A_0+m\alpha)$, as shown in Fig. \ref{fig4}(b). Realistic shell model calculations faithfully reproduce the parallelism of the experimental $E(A)$ in Fig. \ref{fig4}. This parallelism is expressed in terms of the symmetry energy in ordinary mass formulas. It should, however, be noted that the characteristic behavior of $E(A_0+m\alpha)$, which hides important correlations and resulting structure, appears in the $T>0$ lines of other nuclei. The parallel variations of $E(A)$ suggest the existence of a common structure formed in nuclei with $A=A_0+m\alpha$, $A=A_0+m\alpha +2n(2p)$, $A=A_0+m\alpha +4n(4p)$, etc., and also in $A_0+m\alpha +1n1p$ nuclei with $T=1$. It is notable in Fig. \ref{fig4}(b) that parallel variation of $E(A)$ appears also in the odd-$A$ nuclei with $A=A_0+m\alpha +1n$, which is seen in the shell model results of Fig. \ref{fig3}. The energy difference $E(A_0+m\alpha+1n)-E(A_0+m\alpha)$ is related to the pairing energy in the ordinary mass formulas. This parallelism again suggests a common structure in the two nuclei with $A=A_0+m\alpha$ and $A=A_0+m\alpha +1n$. This common structure is probably the $T=0$ $2n-2p$ correlated structure of the $A_0+m\alpha$ nuclei. \subsection{Multi-quartet structure of even-even $N=Z$ nuclei} The study of nuclear structure has clarified the importance of the $T=0$ $2n2p$ correlations in $N \approx Z$ nuclei. Recall that the core plus $\alpha$ cluster (two $\alpha$ cluster) model accurately describes $^{20}$Ne, $^{24}$Mg, $^{44}$Ti and even $^{48}$Cr. In a simplified picture ignoring the spatial correlations of $\alpha$, the ground states of $A_0+m\alpha$ nuclei can be approximated in the following way \cite{Hase1}: \begin{equation} |\Phi_0(A_0+m\alpha) \rangle \approx \frac{1}{\sqrt{N(A_0+m\alpha)}} (\alpha_{J=T=0}^\dagger)^m |A_0 \rangle, \label{eq:6} \end{equation} where $\alpha_{J=T=0}^\dagger$ consists of a linear combination of four valence nucleons $(c^\dagger)^4_{J=T=0}$ determined by the Tamm-Dancoff equation $[H, \alpha_{J=T=0}^\dagger ] \approx e_\alpha \alpha_{J=T=0}^\dagger$, $N(A_0+m\alpha)$ is a normalization constant, and $|A_0 \rangle$ denotes the doubly-closed-shell core. When the Hamiltonian has only two-body interactions, the energy $E(A_0+m\alpha)$ can be calculated as \begin{eqnarray} & {} & E(A_0+m\alpha) = m \frac{\langle A_0| (\alpha_0)^m (\alpha_{J=T=0}^\dagger)^{m-1} [H, \alpha_{J=T=0}^\dagger ]|A_0 \rangle}{N(A_0+m\alpha)} \nonumber \\ & {} & + \frac{1}{2} m(m-1) \frac{\langle A_0| (\alpha_0)^m (\alpha_{J=T=0}^\dagger)^{m-2} [[H, \alpha_{J=T=0}^\dagger ], \alpha_{J=T=0}^\dagger ]|A_0 \rangle}{N(A_0+m\alpha)}. \label{eq:7} \end{eqnarray} The microscopic calculations of Eq. (\ref{eq:7}) in Ref. 25) approximately reproduce the experimental energies $E_{shl}(A_0+m\alpha)$ [and hence $E(A_0+m\alpha)$] for $^{44}$Ti, $^{48}$Cr, $^{52}$Fe and $^{56}$Ni. Therefore, the $A_0+m\alpha$ nuclei have a multi-quartet structure approximated by Eq. (\ref{eq:6}), at least up to $^{56}$Ni. The characteristic behavior of $E(A_0+m\alpha)$ in Fig. \ref{fig1} indicates the leading role of the multi-quartet structure. The shell model results in $\S$2.3 allow us to imagine the multi-quartet structure for the $A_0+m\alpha$ nuclei beyond $^{56}$Ni. In order to get a simple formula, let us transfer the fermion equation (\ref{eq:7}) into an interacting $\alpha$-boson model, \begin{equation} E_{IBM}(A_0+m\alpha) = m e_\alpha - \frac{1}{2} m(m-1) G_{\alpha \alpha}, \label{eq:8} \end{equation} where $e_\alpha$ is the energy of the $\alpha$-boson and $G_{\alpha \alpha}$ denotes the interaction between the $\alpha$-bosons. Because the $\alpha$-boson is regarded as being mapped from the four correlated fermions, the interaction strength $G_{\alpha \alpha}$ should reflect the Pauli principle. The stable doubly-closed-shell nuclei $^{16}$O and $^{40}$Ca suggest that the four correlated nucleons are mainly in one major shell \cite{Tomo}. As the number of $\alpha_{J=T=0}^\dagger$ increases in the major shell, the Pauli principle applied to $\alpha_{J=T=0}^\dagger$ must restrict the degrees of freedom for $\alpha_{J=T=0}^\dagger$. Let us take the Pauli principle effect into account by expressing the interaction strength $G_{\alpha \alpha}$ in the form of a decreasing function of $m$ (the number of $\alpha_{J=T=0}^\dagger$). The simplest way to do this is to approximate the decline function by a linear function of $m$, such as \begin{equation} G_{\alpha \alpha} = g_{\alpha \alpha} \{ 1 - C_\alpha (m-1) \}. \label{eq:9} \end{equation} The factor $C_\alpha$ represents something like the scale of the subspace $\{(\alpha_{J=T=0}^\dagger)^m|A_0 \rangle \}$, depending on the shell structure. This interacting $\alpha$-boson model can reproduce the experimental values of $E(A_0+m\alpha)$ up to $^{56}$Ni, as shown in Fig. \ref{fig1}, where the values denoted by the open squares are obtained with the parameter values $e_\alpha=2.75$, $g_{\alpha \alpha}=5.3$, $C_\alpha=0.21$ in MeV for $^{20}$Ne to $^{28}$Si and $e_\alpha=1.35$, $g_{\alpha \alpha}=3.75$, $C_\alpha=0.18$ in MeV for $^{44}$Ti to $^{56}$Ni. The interacting $\alpha$-boson model describes the peaks at $A=A_0+\alpha$ ($^{20}$Ne and $^{44}$Ti) and the decline toward $^{28}$Si and $^{56}$Ni. The most important point here is that the interaction between the composite quartets $\alpha_{J=T=0}^\dagger$ is attractive and quite strong. Other composite fermion units, like Cooper pairs and vibrational phonons with $J=2$ and $J=3$ in nuclear physics, have repulsive interactions between them and do not actually have a boson-like property, because of the Pauli principle. Only the $\alpha$-like quartet with $J=T=0$ has the possibility to resemble a boson, because of a special mechanism in couplings of spin and isospin. The large energy gain due to the strong $T=0$ $2n-2p$ correlations and the attractive interaction between the $\alpha$-like quartets cause the effect of the (collective) $T=0$ $2n-2p$ correlations on the nuclear mass to be rather inconspicuous. The interacting $\alpha$-boson model (\ref{eq:8}) with (\ref{eq:9}) roughly reproduces the energies $E(^{32}$S$)$ and $E(^{60}$Zn$)$ given in Fig. \ref{fig1}. The shell model calculation in $\S$2.3, however, shows that the upward turn to $^{60}$Zn in the graph of $E(A_0+m\alpha)$ is due to a shell effect. The formula (\ref{eq:8}) cannot be applied to the regions $32<A<40$ and $60<A<80$, because the expression (\ref{eq:9}) is not valid there. A hole picture is probably suitable for these latter halves of the $sd$ and $pf$ shells. Then we obtain the same type of states as in Eq. (\ref{eq:6}) by replacing $\alpha_{J=T=0}^\dagger$ with a linear combination of four holes $(c_h^\dagger)^4_{J=T=0}$. The corresponding boson picture, the interacting $\alpha$-hole boson model, may give a formula similar to (\ref{eq:8}). It is, however, difficult to obtain a simple formula that reproduces the variation of $E(A_0+m\alpha)$ including a shell effect in the entire region of $N=Z$ nuclei. We therefore abandon this problem and instead adopt the experimental values given in Fig. \ref{fig1} for the energy $E(A_0+m\alpha)$ in this paper. The approximation (\ref{eq:6}) is very simplified in comparison with the realistic shell model. Adding other collective modes of the $T=0$ $2n-2p$ correlations with $J>0$ is necessary to better reproduce the variation of $E(A_0+m\alpha)$ for the $f_{7/2}$ shell nuclei \cite{Hase1B}. We should consider the $T=0$ $2n-2p$ correlations as correlations of collective $T=0$ $2n-2p$ quartets with various $J$ in an improved approximation. In fact, although we usually imagine the multi $J$$=$$0$ pair structure for the $T=1$ pair correlated state, a realistic shell model wave function includes components of various $J>0$ pairs in nuclear physics. We use the term ``$T=0$ $2n-2p$ correlations" in such a broader sense. In the following sections, we express the $T=0$ $2n-2p$ correlated states as \begin{equation} |\Phi_0(A_0+m\alpha) \rangle \propto (\alpha_{T=0}^\dagger)^m |A_0 \rangle, \label{eq:10} \end{equation} where $\alpha_{T=0}^\dagger$ represents a quartet of $T=0$ $2n-2p$ correlated nucleons or holes. We formally use the expression (\ref{eq:10}) also in the hole regions, $32<A<40$ and $60<A<80$. \subsection{Superfluid state induced by the $T=0$ $2n-2p$ correlations} In Fig. \ref{fig1}, the line connecting the energies $E(A_0+m\alpha)$ of the $T=0$ nuclei plays an important role in the mass formula, because the energies of the other nuclei with $T>0$ are measured from this line. For $E(A_0+m\alpha)$, we provisionally use the experimental values evaluated from $B(A_0+m\alpha)-B_0(A_0+m\alpha)$, as mentioned above. Let us extrapolate the line for $E(A_0+m\alpha)$ to nuclei with $A \neq A_0+m\alpha$ as follows: \begin{eqnarray} E_{T=0}(A_0+m\alpha+2) & = & (E(A_0+m\alpha) + E(A_0+m\alpha+\alpha)) /2, \nonumber \\ E_{T=0}(A_0+m\alpha+1) & = & (E(A_0+m\alpha) + E_{T=0}(A_0+m\alpha+2)) /2, \nonumber \\ E_{T=0}(A_0+m\alpha+3) & = & (E_{T=0}(A_0+m\alpha+2) + E(A_0+m\alpha+\alpha)) /2. \label{eq:11} \end{eqnarray} These equations define the $T=0$ plane as the base level of energy in the mass table. At this stage, the binding energy is written \begin{equation} B(A) = B_0(A) + E_{T=0}(A) + W(A). \label{eq:12} \end{equation} The pairing energy, symmetry energy, Wigner energy and a correction for odd-odd nuclei are included in the residual energy $W(A)$ in Eq. (\ref{eq:12}). The $T=0$ $2n-2p$ correlations are related to the ``$\alpha$-like superfluidity" proposed in Ref. 7), where, by analogy to pairing superfluidity, $\alpha$-like superfluidity is indicated by the following mass difference corresponding to the odd-even mass difference $\Delta$: \begin{eqnarray} \delta M(A_0+m\alpha+2) & = & (B(A_0+m\alpha+2n)+B(A_0+m\alpha+2p))/2 \nonumber \\ & - & (B(A_0+m\alpha)+B(A_0+m\alpha+\alpha))/2. \label{eq:13} \end{eqnarray} This quantity is called the ODD-EVEN mass difference in $\S$2.2. The average energy of the $A_0+m\alpha+2n$ and $A_0+m\alpha+2p$ nuclei is used so as to remove the Coulomb energy effect. Although the $\alpha$-like superfluidity in heavy nuclei is discussed in Ref. 7), we are concerned with $N \approx Z$ nuclei, for which the isospin is a good quantum number. In place of Eq. (\ref{eq:13}), we define the following quantity as an indicator of the $\alpha$-like superfluidity: \begin{eqnarray} \delta W(A_0+m\alpha+2:T=1,K) & = & W(A_0+m\alpha+2:T=1,K) \nonumber \\ & - & (W(A_0+m\alpha)+W(A_0+m\alpha+\alpha))/2, \label{eq:14} \end{eqnarray} where $K=1$, 0 and $-1$ correspond to the $T=1$ states of the $A_0+m\alpha+2n$, $A_0+m\alpha+1n1p$ and $A_0+m\alpha+2p$ nuclei, respectively. The residual energy $W(A)$ is measured from the $T=0$ plane [$E_{T=0}(A)$], and hence we have $W(A_0+m\alpha)=W(A_0+m\alpha+\alpha)=0$. It is thus seen that $W(A_0+m\alpha+2:T=1,K)$ is identically the ODD-EVEN mass difference for $\alpha$-like superfluidity, \begin{equation} \delta W(A_0+m\alpha+2:T=1,K) = W(A_0+m\alpha+2:T=1,K). \label{eq:15} \end{equation} The average of $W(A_0+m\alpha+2:T=1,K=1)$ and $W(A_0+m\alpha+2:T=1,K=-1)$ corresponds with $\delta M(A_0+m\alpha+2)$ in Eq. (\ref{eq:13}), which represents the $n-p$ (mainly $T=0$) interaction energy between $2n$ and $2p$ in an $A_0+m\alpha+\alpha$ nucleus \cite{Kaneko}. The values of $W(A_0+m\alpha+2:T=1,K)$ for the $A_0+m\alpha+2n$, $A_0+m\alpha+1n1p$ and $A_0+m\alpha+2p$ nuclei are plotted by the dotted curves in Fig. \ref{fig5}. Figure \ref{fig5} shows that the ODD-EVEN mass difference for the $T=0$ $2n-2p$ correlations is larger than the odd-even mass difference $\Delta$ for the pairing superfluidity. The shell model calculation in $\S$2.3 shows that there is a significant contribution of the monopole field $H_{mp}^{T=0}$ to the ODD-EVEN mass difference [in Eq. (\ref{eq:14})]. The $H_{mp}^{T=0}$ contribution is $3k^0/2$. For $A\approx 58$, for instance, the value is about 2.14 MeV, while $W(A=58:T=1,K=1) \approx 3.4$ MeV. The $H_{res}$ contribution to the ODD-EVEN mass difference is about 1.25 MeV, which is comparable to the odd-even mass difference $\Delta \approx 1.34$ MeV near $A=58$. It should be noted here that the $T=1$ pair correlations of neutron and proton pairs joining in the formation of the $T=0$ $2n-2p$ quartet are not included in the ODD-EVEN mass difference. We can say that the strong $T=0$ $2n-2p$ correlations cause a superfluid state, like the pairing superfluid state, as claimed by Gambhir {\it et al.}\cite{Gamb} It is notable that the ODD-EVEN mass difference for $N \approx Z$ nuclei is larger than that for $N>Z$ nuclei, with regard to which the term ``$\alpha$-superfluidity" was first used \cite{Gamb}. We call the strongly correlated state an ``$\alpha$-like superfluid state". As discussed in the previous subsection, the $\alpha$-like superfluid state has the multi-quartet structure (\ref{eq:6}) in the $A_0+m\alpha$ nuclei, at least up to $^{56}$Ni. Figure \ref{fig5} displays the systematic differences among the $A_0+m\alpha+2n$, $A_0+m\alpha+1n1p$ and $A_0+m\alpha+2p$ nuclei. This indicates that the effect of the Coulomb interaction remains after subtracting the Coulomb energy term $a_C Z^2/A^{1/3}$. It may be necessary for a practical mass formula to add some correction terms in order to remove the differences between the states with different $K$. In fact, modern mass formulas do have such correction terms. However, we leave this problem and employ different parameters for neutrons and protons in this paper, where we aim to explain our basic idea. \begin{figure}[t] \begin{center} \includegraphics[width=7cm,height=7cm]{fig5.eps} \caption{The energy $W(A_0+m\alpha+2:T=1,K)$ for $A_0+m\alpha+2$ nuclei with $T=1$. This quantity represents the ODD-EVEN mass difference for $\alpha$-like superfluidity.} \label{fig5} \end{center} \end{figure} \subsection{Bogoliubov transformation for the $\alpha$-like superfluid state} We consider $A_0+m\alpha+2$ nuclei with $T=1$, where structure is roughly expressed as \begin{eqnarray} & {} & |A=A_0+m\alpha+2:T=1,K \rangle \propto S_K^\dagger (\alpha_{T=0}^\dagger)^m |A_0 \rangle , \label{eq:16} \\ & {} & S_K^\dagger \propto (c^\dagger c^\dagger)_{J=0,T=1,K} . \nonumber \end{eqnarray} The large values of $W(A_0+m\alpha+2:1K)$ allow us to regard the state $|\Phi_0(A_0+m\alpha) \rangle $ as an $\alpha$-like superfluid state $(\alpha_{T=0}^\dagger)^m |A_0 \rangle$. After a kind of the Bogoliubov transformation, the $\alpha$-like superfluid state is the vacuum state $|0(\alpha) \rangle$ for a quasi-pair $\mbox{\boldmath $S$}_K^\dagger$, which is transformed from $S_K^\dagger$. In this picture, the state (\ref{eq:16}) is regarded as a quasi-pair state, like the quasi-particle state in the BCS theory, \begin{equation} |A=A_0+m\alpha+2:1,K \rangle = \mbox{\boldmath $S$}_K^\dagger |0(\alpha) \rangle . \label{eq:17} \end{equation} Measuring the energy $W(A)$ from the $T=0$ plane [$E_{T=0}(A)$] defined in Eq. (\ref{eq:11}) corresponds to the above transformation for the wave functions. The energy $W(A_0+m\alpha+2:1K)$ is the energy of the quasi-pair $\mbox{\boldmath $S$}_K^\dagger$. This discussion is parallel to the quasi-particle picture concerning the pairing energy, as seen below. The above transformation is, in fact, difficult to carry out for the four composite fermions $\alpha_{T=0}^\dagger$. Instead, we illustrate our plan using the interacting boson model (IBM), as done by Gambhir {\it et al.} \cite{Gamb}. The IBM for $N \approx Z$ nuclei is called the IBM3. The IBM3 Hamiltonian is expressed in terms of the $s$ boson ($J=0$) and $d$ boson ($J=2$) with $T=1$ \cite{Thompson,Hase4}, \begin{equation} s_K^\dagger = s_{J=0,T=1,K}^\dagger , \quad d_{MK}^\dagger = d_{2M1K}^\dagger. \label{eq:18} \end{equation} The $sd$ boson image of $\alpha_{J=T=0}^\dagger$ is given by \begin{equation} \alpha_{J=T=0}^\dagger \Rightarrow x (s^\dagger s^\dagger)_{J=T=0} + \sqrt{1-x^2} (d^\dagger d^\dagger)_{J=T=0}. \label{eq:19} \end{equation} For the $\alpha$-like superfluid state, the quasi $s$ and $d$ bosons ($\mbox{\boldmath $s$}_K$, $\mbox{\boldmath $d$}_{MK}$) are introduced through the Bogoliubov transformation \begin{eqnarray} s_K^\dagger &=& U_s \mbox{\boldmath $s$}_K^\dagger + V_s (-)^{1-K} \mbox{\boldmath $s$}_{-K} , \nonumber \\ d_{MK}^\dagger &=& U_d \mbox{\boldmath $d$}_{MK}^\dagger + V_d (-)^{2-M}(-)^{1-K}\mbox{\boldmath $d$}_{-M-K}. \label{eq:20} \end{eqnarray} Here, we have $U_i^2 - V_i^2 =1$ ($i=s$, $d$). We have a boson-type gap equation and can calculate the quasi-boson energies $e_s$ and $e_d$ using an appropriate IBM3 Hamiltonian. In this quasi-boson picture, the quasi-pair state (\ref{eq:17}) of the $A_0+m\alpha+2$ nuclei is written \begin{equation} |A=A_0+m\alpha+2:1,K \rangle \Rightarrow \mbox{\boldmath $s$}_K^\dagger |0(\alpha)), \label{eq:21} \end{equation} where the $\alpha$-like superfluid vacuum state is replaced by that for the quasi-bosons ($\mbox{\boldmath $s$}_K$, $\mbox{\boldmath $d$}_{MK}$). There is a well-defined IBM3 Hamiltonian for the $f_{7/2}$ shell nuclei \cite{Thompson}. Using the IBM3 Hamiltonian, we evaluated the quasi-boson energy $e_s$, which should be equal to $W(A_0+m\alpha+2:T=1)$ given in Eq. (\ref{eq:15}). The calculated values of $e_s$ for $^{46}$Ti, $^{50}$Cr and $^{54}$Fe are plotted in Fig. \ref{fig5}. It is seen that the quasi-boson energies $e_s$ accurately reproduce the experimental values of $W(A_0+m\alpha+2n)$ which is least sensitive to the Coulomb interaction effect. This success supports our interpretation of the $\alpha$-like superfluidity of $A_0+m\alpha+2$ nuclei. \section{Multi-pair structure on the base of $\alpha$-like superfluidity} Because the picture of $\alpha$-like superfluidity is good, the $J=0$ ground states of even-even nuclei can be approximated by \begin{eqnarray} |A=A_0+m\alpha+2l:T=l \rangle & \propto & (S^\dagger)^l |\Phi_0(A_0+m\alpha) \rangle , \nonumber \\ & \Rightarrow & (\mbox{\boldmath $s$}^\dagger)^l |0(\alpha)). \label{eq:22} \end{eqnarray} Similar wave functions are considered in the microscopic derivation of a mass formula \cite{Zuker}. Now we have reached the second stage, which can be compared with the first stage considering the multi-quartet state $(\alpha_{T=0}^\dagger)^m|A_0 \rangle$. We have another interacting boson picture for the Cooper pair, \begin{equation} W(A_0+m\alpha+2l:T=l) = l e_s + \frac{1}{2} l(l-1) g_{ss}. \label{eq:23} \end{equation} The interaction between the Cooper pairs (like-nucleon pairs) is repulsive because of the Pauli principle. The repulsive interaction between the quasi-$s$-bosons gives a quadratic increase of the mass, depending on the boson number $l$. The quasi-$s$-boson $\mbox{\boldmath $s$}_K^\dagger$ increases the isospin of the state by 1, and the number of $\mbox{\boldmath $s$}_K^\dagger$ can be replaced with the isospin $T$ in Eq. (\ref{eq:23}). Let us write Eq. (\ref{eq:23}) in the ordinary form \begin{eqnarray} W(A_0+m\alpha+2l:T=l) & = & a_{sym} T^2 + b_{Wig} T , \label{eq:24} \\ a_{sym} & = & \frac{1}{2}g_{ss}, \quad b_{Wig}=e_s-\frac{1}{2}g_{ss}. \label{eq:25} \end{eqnarray} The first term here is called the symmetry energy, and the second term is called the Wigner energy in the mass formulas. Our interacting boson picture for the multi-pair states explains the structural origins of the symmetry energy and Wigner energy. \begin{figure}[b] \begin{center} \includegraphics[width=8cm,height=8cm]{fig6.eps} \caption{Energies $W(m\alpha + 2ln:T=l)$ of multi-quasi-neutron-pair states $(S_n^\dagger)^{l=T}|\Phi_0(m\alpha) \rangle$. Experimental values (flat dots) are compared with the theoretical values obtained with the parameters (\ref{eq:26}) (which are at the intersections of the solid and dotted curves).} \label{fig6} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=8cm,height=5.71cm]{fig7.eps} \caption{Energies $W(m\alpha + 2lp:T=l)$ of multi-quasi-proton-pair states $(S_p^\dagger)^{l=T}|\Phi_0(m\alpha) \rangle$, shown in the same manner as Fig. \ref{fig5}. The theoretical values (at the intersections of the two curves) are obtained with the parameters (\ref{eq:27}).} \label{fig7} \end{center} \end{figure} Figures \ref{fig6} and \ref{fig7} display the experimental energies of the multi-quasi-pair states (\ref{eq:22}), $W(A_0+m\alpha+2ln)$ and $W(A_0+m\alpha+2lp)$. We can fix the parameter values $e_s$ and $g_{ss}$ in the approximation (\ref{eq:23}) [$a_{sym}$ and $b_{Wig}$ in Eq. (\ref{eq:24})] from the experimental values of $W(A_0+m\alpha+2l:T=l)$. They can be expressed in the same form as that determined in the microscopic mass formula \cite{Duflo}. The parameters, which are fixed separately for neutrons and protons, are \begin{eqnarray} a_{sym}^{(n)} & = & 116 (1-1.52/A^{1/3})/A, \nonumber \\ b_{Wig}^{(n)} & = & 218 (1-1.52/A^{1/3})/A, \label{eq:26} \\ a_{sym}^{(p)} & = & 82 (1-1.0/A^{1/3})/A, \nonumber \\ b_{Wig}^{(p)} & = & 92 (1-1.0/A^{1/3})/A. \label{eq:27} \end{eqnarray} The quasi-$s$-boson energies $e_s^{(n)}$ and $e_s^{(p)}$ are plotted by the solid and dash-dot curves in Fig. \ref{fig5} [where a curve fitted to $W(A_0+m\alpha+1n1p)$ is also shown]. We see in Figs. \ref{fig6} and \ref{fig7} that the approximation (\ref{eq:24}) with the parameters (\ref{eq:26}) and (\ref{eq:27}) is very good, and hence our interacting boson picture for Cooper pairs is also good. From the parameters (\ref{eq:26}) and (\ref{eq:27}), the coefficient of the Wigner energy is larger than that of the symmetry energy ($b_{Wig}^{(n)} \approx 1.88 a_{sym}^{(n)})$ for neutrons and $b_{Wig}^{(p)} \approx a_{sym}^{(p)}$ for protons. The symmetry energy coefficient $a_{sym}^{(n)}$ is nearly equal to that determined in Ref. 4). The values of these parameters depend on the manner of evaluating the Coulomb energy. We point out that the quasi-$s$-boson energy is, for instance, $e_s^{(n)} \approx 6.4$ MeV for $A=20$ and $e_s^{(n)} \approx 4.3$ MeV for $A=44$, as obtained from Fig.~\ref{fig5}. [The interaction energy between the quasi-$s$-bosons is $g_{ss}^{(n)} \approx 5.1$ MeV for $A=20$ and $g_{ss}^{(n)} \approx 3.0$ MeV for $A=44$, from Eq. (\ref{eq:26}).] The quasi-$s$-boson energy $e_s$ is larger than the $\alpha$-boson energy $e_\alpha$ ($e_\alpha =1.35$ MeV for the $sd$ shell nuclei and $e_\alpha =2.75$ MeV for the $pf$ shell nuclei). The fact that $e_\alpha$ is much smaller than $e_s^{(n)}+e_s^{(p)}$ indicates the very large energy gain of the $\alpha$-like quartet. Moreover, while the $\alpha$-like quartet interaction is attractive, the quasi-pair interaction is repulsive. The characteristic patterns in Figs. \ref{fig6} and \ref{fig7} are due to the repulsive interaction between the quasi-pairs (a quasi-pair transfers isospin 1). These are contrast with the inconspicuous effect of the multi-quartet structure on the energy $E(m \alpha)$, shown in Fig.~\ref{fig1}. (The $\alpha$-like quartet transfers no quantum number other than the nucleon number.) If there was no such great energy gain caused by the collective $T=0$ $2n-2p$ correlations, the nuclear mass table would be different. \section{Structure having an unpaired neutron and/or an unpaired proton} Before ending the second stage treating the multi-pair states, let us write our mass formula as \begin{equation} B(A) = B_0(A) + E_{T=0}(A) + w_T(A) + w_v(A). \label{eq:28} \end{equation} We extend the $T$-dependent energy $W(A_0+m\alpha+2l:T=l)$ in Eq. (\ref{eq:24}) to odd-$A$ nuclei and odd-odd nuclei, as we extended $E(A_0+m\alpha)$ to $E_{T=0}(A)$, expressing it as \begin{equation} w_T(A) = a_{sym}(A) T^2 + b_{Wig}(A) T, \label{eq:29} \end{equation} where we permit $T$ to be a half integer for odd-$A$ nuclei. The last term, $w_v(A)$ in Eq. (\ref{eq:28}), represents the energy of unpaired nucleon(s). The subscript $v$ is the seniority quantum number. It should be noted that in the mass formula (\ref{eq:28}), the energy $w_v(A)$ of an odd-$A$ (or odd-odd) nucleus is measured from the base {\it curve} given by (\ref{eq:29}). \subsection{Shifted quasi-particle energy for odd-mass nuclei} The strength of the pairing correlations in an odd-$A$ nucleus is usually evaluated with the odd-even mass difference. We define it using $W(A)$ of Eq. (\ref{eq:12}) in the same form as Eq. (\ref{eq:14}), \begin{equation} \Delta(A=A_0+m\alpha+2l+1) = W(A) - ( W(A-1) + W(A+1) ) /2. \label{eq:30} \end{equation} This relation is illustrated in Fig. \ref{fig2}. Substituting the relation (\ref{eq:29}) for $W(A-1)$ and $W(A+1)$, we obtain the approximate relation \begin{equation} \Delta(A=A_0+m\alpha+2l+1) \approx W(A) - ( w_{T=l+1/2}(A) + a_{sym}(A)/4 ). \label{eq:31} \end{equation} The energy $w_{v=1}(A)$ for an odd-$A$ nucleus with $A=A_0+m\alpha+2l+1$ in Eq. (\ref{eq:28}) is given by \begin{eqnarray} w_{v=1}(A) & \equiv & W(A) - w_{T=l+1/2}(A) \nonumber \\ & \approx & \Delta(A) + \frac{1}{4} a_{sym}(A). \label{eq:32} \end{eqnarray} The energy shift $a_{sym}(A)/4$ is inevitable when we measure the energy $w_{v=1}(A)$ from the base curve (\ref{eq:29}). \begin{figure}[b] \begin{center} \includegraphics[width=6.8cm,height=7.2cm]{fig8a.eps} \includegraphics[width=6.8cm,height=7.2cm]{fig8b.eps} \caption{Energies $w_{v=1}(A)=d_n(A)$ for odd-$N$ nuclei and $w_{v=1}(A)=d_p(A)$ for odd-$Z$ nuclei.} \label{fig8} \end{center} \end{figure} In the last stage, we consider $A_0+ m\alpha +2l+1$ nuclei with $T=l+1/2$, which have the structure \begin{equation} |A=A_0+m\alpha+2l+1 \rangle \propto c^\dagger (S^\dagger)^l |\Phi_0(A_0+m\alpha) \rangle . \label{eq:33} \end{equation} This structure is expressed approximately as a direct product of the three modules $|\Phi_0(A_0+m\alpha) \rangle$, $(S^\dagger)^l$, and the last odd nucleon $c^\dagger$. We regard the multi-pair structure $(S^\dagger)^l$ as the pairing superfluid structure as usual. After the Bogoliubov transformation, the odd-$A$ nucleus is regarded as the one quasi-particle state, \begin{equation} |A=A_0+m\alpha+2l+1 \rangle = a^\dagger |0(lS)\otimes 0(A_0+m\alpha) \rangle . \label{eq:34} \end{equation} In this picture, the energy $w_{T=l}(A)$ given in Eq. (\ref{eq:28}) represents the energy of the pairing superfluid state $|0(lS) \rangle$, which is the vacuum for the quasi-particle $a^\dagger$, and the quantity $\Delta (A)$ in Eq. (\ref{eq:30}) can be regarded as the quasi-particle energy. Let us rewrite the ``shifted quasi-particle energy" $w_{v=1}(A)$ as \begin{equation} d_n(A) = \Delta_n(A) + a_{sym}^{(n)}/4, \quad d_p(A) = \Delta_p(A) + a_{sym}^{(p)}/4. \label{eq:35} \end{equation} Experimental values of $w_{v=1}(A)$ calculated with the experimental values of $E_{T=0}(A)$ in Eq. (\ref{eq:28}) are plotted in Fig. \ref{fig8}. With the approximate relation (\ref{eq:32}), we can parameterize the quantity $\Delta(A)$ in the same form ($\propto A^{-1/3}$) as that of the microscopic mass formula \cite{Duflo}, {\it i.e.}, \begin{equation} \Delta_n(A) = 5.18/A^{1/3}, \quad \Delta_p(A) = 4.6/A^{1/3}. \label{eq:36} \end{equation} The neutron value, $\Delta_n(A)$, is equal to that given in Ref. 4), and the proton value, $\Delta_p(A)$, is smaller than $\Delta_n(A)$. It should be noted that $\Delta(A)$ is a measure of the $T=1$ pair correlations and $W(A:T=1)=e_s$ is approximately a measure of the $T=0$ $n-p$ correlations between the $T=1$ pairs \cite{Kaneko,Janecke}. This leads to the different $A$ dependences of $\Delta(A)$ and $e_s$. According to Ref. 39), because the symmetry energy contribution is cancelled by the curvature contribution from a smooth density of states in the Strutinsky method, the three-point odd-even mass difference $\Delta_n(A)$ in Eq. (\ref{eq:30}) is a good indicator of the pairing gap, which is approximately equal to the quasi-particle energy. It is notable that, in contrast to the $A$-dependence $5.18/A^{1/3}$ of $\Delta_n(A)$, the $A$-dependence of $w^{(n)}_{v=1}(A)=d_n(A)$ can be expressed as $12/\sqrt{A}$, as shown in Fig. \ref{fig8}(a). The curve $12/\sqrt{A}$ is known to represent the $A$-dependence of the pairing energy of the semi-empirical mass formula \cite{Zeldes}, which is estimated with the four-point odd-even mass difference. The shifted quasi-particle energy $d_n(A)$, therefore, corresponds to the pairing energy of the semi-empirical mass formula or the four-point odd-even mass difference. The classical mass formulas, having the pairing energy term $\delta_{pair}$ and the symmetry energy term $a_T T^2$, lead to the relation $\Delta_n(A) \approx \delta_{pair}- a_T/4$. Combining this relation and Eq. (\ref{eq:35}), we confirm the equivalence $d_n=\delta_{pair}$. Equation (\ref{eq:35}) indicates that the so-called pairing energy $d_n=\delta_{pair}$ contains a symmetry energy contribution. We can now distinguish the two curves $5.18/A^{1/3}$ and $12/\sqrt{A}$: The former represents the three-point odd-even mass difference $\Delta_n(A)$ (which is the quasi-particle energy or the pairing gap), and the latter represents the four-point odd-even mass difference equal to $d_n(A)$, including the symmetry energy contribution $a_{sym}^{(n)}/4$. \subsection{Seniority $v=2$ states of odd-odd nuclei} The remaining task is to determine whether the mass formula (\ref{eq:28}) is effective for odd-odd nuclei. The ground state of an odd-odd nucleus is the seniority $v=2$ state, except in the case of some $N=Z$ nuclei. (The exceptional state with $v=0$ and $T=1$ is the $1n1p$ pair state $S_{K=0}^\dagger |\Phi_0(A_0+m\alpha) \rangle$ considered in Fig. \ref{fig5}.) The seniority $v=2$ state is composed of a quasi-neutron and a quasi-proton, \begin{equation} |A=A_0+m\alpha+2l+n+p \rangle = a_n^\dagger a_p^\dagger |0(lS)\otimes 0(A_0+m\alpha) \rangle . \label{eq:37} \end{equation} The energy $w_{v=2}(A)$ for this state is defined by $w_{v=2}(A)=W(A)-w_{T=l}(A)$. Let us evaluate its experimental value in a manner similar to Eq. (\ref{eq:30}): \begin{eqnarray} & {} & w_{v=2}(A=A_0+m\alpha+2l+n+p) \nonumber \\ & {} & \ \ \ = W(N,Z) - ( W(N-1,Z-1) + W(N+1,Z+1) ) /2. \ \ \label{eq:38} \end{eqnarray} The calculated values are plotted in Fig. \ref{fig9}. It is seen that there is a difference between the odd-odd $N=Z$ nuclei and the other odd-odd nuclei. The data indicate the relations \begin{eqnarray} w_{v=2}(N=Z) \approx d_n + d_p , \label{eq:39} \\ w_{v=2}(N \neq Z) \approx \Delta_n + \Delta_p . \label{eq:40} \end{eqnarray} \begin{figure}[t] \begin{center} \includegraphics[width=7cm,height=7cm]{fig9.eps} \caption{Energies $w_{v=2}(A)$ for odd-odd nuclei.} \label{fig9} \end{center} \end{figure} The parameters $\Delta_n$ and $\Delta_p$ in Eq. (\ref{eq:36}) [$d_n$ and $d_p$ in Eq.~(\ref{eq:35})] fitted for the odd-$A$ nuclei can reproduce the experimental energies $w_{v=2}$ of odd-odd nuclei, though it is not clear why $w_{v=2}(N=Z)$ is different from $w_{v=2}(N \neq Z)$. This point is possibly related to the condition that there is no Cooper pair in odd-odd $N=Z$ nuclei, while the other odd-odd nuclei have one or more Cooper pairs. Sometimes, correction terms are added to mass formulas for odd-odd nuclei. The correction for the odd-odd $N=Z$ nuclei is included in Eq. (\ref{eq:39}) in contrast to Eq. (\ref{eq:40}) for odd-odd $N \neq Z$ nuclei. We ignore another correction, which represents an additional $n-p$ interaction, because the deviations from the fitted curves in Fig. \ref{fig9} are of a similar or smaller magnitude than the deviations in Figs. \ref{fig5}--\ref{fig9}. \section{Concluding remarks} We have shown the essential role of the $T=0$ $2n-2p$ correlations in the nuclear mass by considering concrete nuclear structure based on the $jj$ coupling shell model. We find that explicitly taking account of the effects of the $T=0$ $2n-2p$ correlations, which have been overlooked in the past, is important for understanding the nuclear mass formula. We have rearranged the mass formula by treating the $T=0$ $2n-2p$ correlations and the $T=1$ pair correlations as the most important correlations in nuclei. Let us write it again: \begin{equation} B(A) = B_{VSC}(A) + \delta U_{pot}(A) + E_{T=0}(A) + w_T(A) + w_v(A). \nonumber \end{equation} We have discussed the fact that the last three terms $E_{T=0}(A)$, $w_T(A)$ and $w_v(A)$ represent the three modules of the structured wave functions sketched in Eqs. (\ref{eq:10}), (\ref{eq:22}) and (\ref{eq:33}). The systematic formulation of the $T=0$ $2n-2p$ and $T=1$ pair correlations on the same footing makes it clear that the energy $E_{T=0}(A)$ of the multi-quartet structure should be added to the energy $w_T(A)$ of the multi-pair structure. The $T=0$ energy plane $E_{T=0}(A)$ supplies the base level for the measurement of the $T$-dependent energy $w_T(A)$. The interacting boson model for the $T=1$ Cooper pair on the $\alpha$-like superfluid base provides a structural explanation for the origins of the symmetry energy and Wigner energy. The two standard curves $5.18/A^{1/3}$ and $12/\sqrt{A}$ for the pairing energy are distinguished and identified as representing the quasi-particle energy or the pairing gap (three-point odd-even mass difference) and the shifted quasi-particle energy (four-point odd-even mass difference), respectively. The $E_{T=0}(A)$ term as the base level affects the binding energies of all nuclei. Adding $E_{T=0}(A)$ to existing mass formulas could improve the precision. We can estimate the precision using the parameters in Eqs. (\ref{eq:26}), (\ref{eq:27}) and (\ref{eq:36}) and the experimental values of $E_{T=0}(A)$. The average of the root-mean-square (rms) errors estimated is 1.42 MeV for even-even nuclei, 1.37 MeV for odd-$A$ nuclei, and 1.11 MeV for odd-odd nuclei. These values are, of course, larger than the rms errors for modern mass formulas. The average of the rms errors for the FRDM, for instance, is 1.08 MeV for even-even nuclei, 1.13 MeV for odd-$A$ nuclei, and 1.12 MeV for odd-odd nuclei in the region $17 \le Z,N \le 36$. However, it should be noted that these FRDM values are larger than the average of the rms errors for all nuclei, 0.67 MeV. This suggests a flaw in the FRDM mass formula for $N \approx Z$ nuclei. The advantage of our treatment is clear if we consider nuclei near the $N=Z$ line. For $T<4$ nuclei, the average of the rms errors becomes 0.63 MeV for even-even nuclei and 1.07 MeV for odd-$A$ nuclei. The good parallelism from the $T=0$ line to the $T=3$ line in Fig. \ref{fig4} reveals this mechanism. By contrast, the FRDM mass formula does not show such a reduction when the number of $T=|N-Z|/2$ is limited. There seems to be room to take into account the energy $E_{T=0}(A)$ of the fundamental $T=0$ $2n-2p$ correlated structure in the modern mass formulas.
{ "timestamp": "2005-05-21T02:49:56", "yymm": "0503", "arxiv_id": "nucl-th/0503006", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503006" }
\section*{Introduction} After J.J. Thomson [1] discovered the small corpuscle which soon became known as the electron an enormous amount of theoretical work has been done to explain the existence of the electron. Some of the most distinguished physicists have participated in this effort. Lorentz [2], Poincar\'{e} [3], Ehrenfest [4], Einstein [5], Pauli [6], and others showed that it is fairly certain that the electron cannot be explained as a purely electromagnetic particle. In particular it was not clear how the electrical charge could be held together in its small volume because the internal parts of the charge repel each other. Poincar\'{e} [7] did not leave it at showing that such an electron could not be stable, but suggested a solution for the problem by introducing what has become known as the Poincar\'{e} stresses whose origin however remained unexplained. These studies were concerned with the static properties of the electron, its mass m(e$^\pm$) and its electric charge e. In order to explain the electron with its existing mass and charge it appears to be necessary to add to Maxwell's equations a non-electromagnetic mass and a non-electromagnetic force which could hold the electric charge together. We shall see what this mass and force is. The discovery of the spin of the electron by Uhlenbeck and Goudsmit [8] increased the difficulties of the problem in so far as it now had also to be explained how the angular momentum $\hbar$/2 and the magnetic moment $ \mu_e$ come about. The spin of a point-like electron seemed to be explained by Dirac's [9] equation, however it turned out later [10] that Dirac type equations can be constructed for any value of the spin. Afterwards Schr\"{o}dinger [11] tried to explain the spin and the magnetic moment of the electron with the so-called Zitterbewegung. Later on many other models of the electron were proposed. On p.74 of his book ``The Enigmatic Electron" Mac Gregor [12] lists more than thirty such models. At the end none of these models has been completely successful because the problem developed a seemingly insurmountable difficulty when it was shown through electron-electron scattering experiments that the radius of the electron must be smaller than $10^{-16}$\,cm, in other words that the electron appears to be a point particle, at least by three orders of magnitude smaller than the classical electron radius r$_e$ = e$^2$/mc$^2$ = 2.8179$\cdot10^{-13}$\,cm. This, of course, makes it very difficult to explain how a particle can have a finite angular momentum when the radius goes to zero, and how an electric charge can be confined in an infinitesimally small volume. If the elementary electrical charge were contained in a volume with a radius of O($10^{-16}$)\,cm the Coulomb self-energy would be orders of magnitude larger than the rest mass of the electron, which is not realistic. The choice is between a massless point charge and a finite size particle with a non-interacting mass to which an elementary electrical charge is attached. We propose in the following that the non-electromagnetic mass which seems to be necessary in order to explain the mass of the electron consists of neutrinos. This is actually a necessary consequence of our standing wave model [13] of the masses of the mesons and baryons. And we propose that the non-electromagnetic force required to hold the electric charge and the neutrinos in the electron together is the weak nuclear force which, as we have suggested in [13], holds together the masses of the mesons and baryons and also the mass of the muons. Since the range of the weak nuclear force is on the order of $10^{-16}$\,cm the neutrinos can only be arranged in a lattice with the weak force extending from each lattice point only to the nearest neighbors. The size of the neutrino lattice in the electron does not at all contradict the results of the scattering experiments, just as the explanation of the mass of the muons with the standing wave model does not contradict the apparent point particle characteristics of the muon, because neutrinos are in a very good approximation non-interacting and therefore are not noticed in scattering experiments with electrons. \section{ The mass and charge of the electron} The rest mass of the electron is m(e$^\pm$) = 0.510\,998\,92 $\pm$ 4$\cdot10^{-8}$\,MeV/c$^2$ and the electrostatic charge of the electron is e = 4.803\,204\,41$\cdot10^{-10}$\,esu, as stated in the Review of Particle Physics [14]. Both are known with great accuracy. The objective of a theory of the electron must be the explanation of both values. We will first explain the rest mass of the electron making use of what we have learned from the standing wave model, in particular of what we have learned about the explanation of the mass of the $\mu^\pm$\,mesons in [13]. The muons are leptons, just as the electrons, that means that they interact with other particles exclusively through the electric force. The muons have a mass which is 206.768 times larger than the mass of the electron, but they have the same elementary electric charge as the electron or positron and the same spin. Scattering experiments tell that the $\mu^\pm$\,mesons are point particles with a size $<$\,$10^{-16}$\,cm, just as the electron. In other words, the muons have the same characteristics as the electrons and positrons but for a mass which is about 200 times larger. Consequently the muon is often referred to as a ``heavy" electron. If a non-electromagnetic mass is required to explain the mass of the electron then a non-electromagnetic mass 200 times as large as in the electron is required to explain the mass of the muons. These non-electromagnetic masses must be \emph{non-interacting}, otherwise scattering experiments could not find the size of either the electron or the muon at 10$^{-16}$\,cm. We have already explained the mass of the muons with the standing wave model [13]. According to this model the muons consist of an elementary electric charge and a lattice of neutrinos which, as we know, do not interact with charge or mass. Neutrinos are the only non-interacting matter we know of. In the muon lattice are, according to [13], (N\,-\,1)/4 = N$^\prime$/4 muon neutrinos $\nu_\mu$ (respectively anti-muon neutrinos $\bar{\nu}_\mu$), N$^\prime$/4 electron neutrinos $\nu_e$ and the same number of anti-electron neutrinos $\bar{\nu}_e$, one elementary electric charge and the energy of the lattice oscillations. The letter N stands for the number of all neutrinos and antineutrinos in the cubic lattice of the $\pi^\pm$ mesons [13,\,Eq.(15)] \begin{equation} \mathrm{N} = 2.854\cdot10^9\,. \end{equation} It is, according to [13], a necessary consequence of the decay of the $\mu^-$ muon $\mu^- \rightarrow$ e$^- + \bar{\nu}_e + \nu_\mu$ that there must be N$^\prime$/4 electron neutrinos $\nu_e$ in the emitted electron, where N$^\prime$ = N - 1 $\cong$ N [13]. For the mass of the electron neutrinos and anti-electron neutrinos we found in Eq.(34) of [13] that \begin{equation} \mathrm{m}(\nu_e) = \mathrm{m}(\bar{\nu}_e) = 0.365 \,\mathrm{milli\,eV/c^2}\,. \end{equation} \noindent The sum of the energies in the rest masses of the N$^\prime$/4 neutrinos or antineutrinos in the lattice of the electron or positron is then \begin{equation} \sum{\,\mathrm{m(\nu_e)c^2}} = \mathrm{N}^\prime/4\cdot\mathrm{m}(\nu_e)\mathrm{c}^2 = 0.260\,43\,\mathrm{MeV} = 0.5096\,\mathrm{m(e^\pm)}\mathrm{c}^2\,. \end{equation} To put this in other words, one half of the rest mass of the electron comes from the rest masses of electron neutrinos. The other half of the rest mass of the electron must originate from the energy in the electric charge carried by the electron. From pair production $\gamma$ + M $\rightarrow$ e$^-$ + e$^+$ + M, (M being any nucleus), and from conservation of neutrino numbers follows necessarily that there must also be a neutrino lattice composed of N$^\prime$/4 anti-electron neutrinos, which make up the lattice of the positrons, which lattice has, because of Eq.(2), the same rest mass as the neutrino lattice of the electron, as it must be for the antiparticle of the electron. Fourier analysis dictates that a continuum of high frequencies must be in the electrons or positrons created by pair production in a timespan of $10^{-23}$ seconds. We will now determine the energy E$_\nu$(e$^\pm$) contained in the oscillations in the interior of the electron. Since we want to explain the \emph{rest mass} of the electron we can only consider the frequencies of non-progressive waves, either standing waves or circular waves. The sum of the energies of the lattice oscillations is, in the case of the $\pi^\pm$\,mesons, given by \begin{equation} \mathrm{E}_\nu(\pi^\pm) = \frac{\mathrm{h}\nu_0\mathrm{N}}{2\pi(\mathrm{e^{h\nu/kT}}\,\mathrm{-}\,1)} \,\int\limits_{-\pi}^{\pi}\,\phi\,d\phi\,. \end{equation} \noindent This is Eq.(14) combined with Eq.(16) in [13] where they were used to determine the oscillation energy in the $\pi^0$ and $\pi^\pm$ mesons. This equation was introduced by Born and v.\,Karman [15] in order to explain the internal energy of cubic crystals. In Eq.(4) h is Planck's constant, $\nu_0$ = c/2$\pi\emph{a}$ is the reference frequency with the lattice constant \emph{a} = $10^{-16}$ cm, N is the number of all oscillations, $\phi = 2\pi\emph{a}/\lambda$ and T is the temperature in the lattice, for which we found in [13] the value T = 2.38\,$\cdot$\,$10^{14}$\,K. If we apply Eq.(4) to the oscillations in the electron which has N$^\prime$/4 electron neutrinos $\nu_e$ we arrive at E$_\nu$(e$^\pm)$ = 1/4$\cdot$E$_\nu(\pi^\pm$), which is mistaken because E$_\nu(\pi^\pm$) $\approx$ m($\pi^\pm$)c$^2$/2 and m($\pi^\pm$) $\approx$ 273\,m(e$^\pm$). Eq.(4) must be modified in order to be suitable for the oscillations in the electron. It turns out that we must use \begin{equation} \mathrm{E}_\nu(\mathrm{e}^\pm) = \frac{\mathrm{h}\nu_0\mathrm{N}\cdot\alpha_f}{2\pi(\mathrm{e^{h\nu/kT}}\, \mathrm{-}\,1)}\,\int\limits_{-\pi}^{\pi}\,\phi\,d\phi\,, \end{equation} \noindent where $\alpha_f$ is the fine structure constant. The appearance of $\alpha_f$ in Eq.(5) indicates that the nature of the oscillations in the electron is different from the oscillations in the $\pi^0$ or $\pi^\pm$ lattices. With $\alpha_f$ = e$^2/\hbar$c and $\nu_0$ = c/2$\pi$\emph{a} we have \begin{equation}h\nu_0\alpha_f = e^2/\emph{a}\, \end{equation} that means that the oscillations in the electron are \emph{electric oscillations}. There must be N$^\prime$/2 oscillations of the elements of the electric charge in e$^\pm$, because we deal with non-progressive waves, the superposition of two waves. As we will see later the spin requires that the oscillations are circular. That means that 2$\times$N$^\prime$/4 $\cong$ N/2 oscillations are in Eq.(5). From Eqs.(4,5) then follows that \begin{equation} \mathrm{E_\nu(e}^\pm) = \alpha_f/2\cdot\mathrm{E}_\nu(\pi^\pm)\,. \end{equation} \noindent E$_\nu(\pi^\pm)$ is the oscillation energy in the $\pi^\pm$\,mesons which can be calculated with Eq.(4). According to Eq.(27) of [13] it is \begin{equation} \mathrm{E}_\nu(\pi^\pm) = 67.82 \,\mathrm{MeV} = 0.486\,\mathrm{m}(\pi^\pm)\mathrm{c}^2 \approx \mathrm{m}(\pi^\pm)\mathrm{c}^2/2\,. \end{equation} With E$_\nu(\pi^\pm$) $\approx$ m($\pi^\pm$)c$^2$/2 = 139.57/2\,MeV and $\alpha_f$ = 1/137.036 follows from Eq.(7) that \begin{equation} \mathrm{E_\nu(e}^\pm) = \frac{\alpha_f}{2}\cdot \frac{\mathrm{m}(\pi^\pm)\mathrm{c}^2}{2} = 0.254\,62\,\mathrm{MeV} = 0.99657\,\mathrm{m(e^\pm)}\mathrm{c}^2/2\,. \end{equation} \noindent We have determined the value of the oscillation energy in e$^\pm$ from the product of the very accurately known fine structure constant and the very accurately known rest mass of the $\pi^\pm$\,mesons. \emph{One half of the energy in the rest mass of the electron comes from the electric oscillations in the electron}. The other half of the energy in the rest mass of the electron is in the rest masses of the neutrinos in the electron. We can confirm Eq.(9) using Eq.(5) or Eq.(13) with N/2 = 1.427$\cdot10^9$, e = 4.803$\cdot10^{-10}$\,esu, $\emph{a}$ = 1$\cdot10^{-16}$\,cm, f(T) = 1/1.305$\cdot10^{13}$, and with the integral being $\pi^2$ we obtain E$_\nu$(e$^\pm$) = 0.968\,m($\mathrm{e}^\pm)$c$^2$/2. This calculation involves more parameters than Eq.(9) and is consequently less accurate than Eq.(9). In a good approximation the oscillation energy of e$^\pm$ in Eq.(9) is equal to the sum of the energies in the rest masses of the electron neutrinos in the e$^\pm$ lattice in Eq.(3). Since \begin{equation} \mathrm{m(e}^\pm)\mathrm{c}^2 = \mathrm{E}_\nu(\mathrm{e}^\pm) + \sum{\,\mathrm{m}(\nu_e)\mathrm{c}^2} = \mathrm{E_\nu(e^\pm)} + \mathrm{N^\prime/4\cdot m(\nu_e)c^2}\,, \end{equation} \noindent it follows from Eqs.(3) and (9) that \begin{equation} \mathrm{m(e^\pm)c^2(theor)} = 0.5151\,\mathrm{MeV} = 1.0079\,\mathrm{m(e^\pm)c^2(exp)}\,. \end{equation} The measured rest mass of the electron or positron agrees within the accuracy of the parameters N and m($\nu_e)$ with the theoretically predicted rest masses. From Eq.(7) follows with E$_\nu(\pi^\pm)$ $\cong$ m($\pi^\pm$)c$^2$/2 that \vspace{0.5cm} \centerline{2E$_\nu$(e$^\pm)$ $\cong$ m(e$^\pm$)c$^2$ = $\alpha_f$E$_\nu(\pi^\pm)$ = $\alpha_f$m$(\pi^\pm)$c$^2$/2\,,} \vspace{0.5cm} \noindent or that \begin{equation} \mathrm{m(e^\pm)}\cdot2/\alpha_f = 274.072\,\mathrm{m(e^\pm)} \cong \mathrm{m(\pi^\pm)}\,, \end{equation} whereas the actual ratio of the mass of the $\pi^\pm$\,mesons to the mass of the electron is m($\pi^\pm$)/m(e$^\pm$) = 273.132 or 0.9965$\cdot$2/$\alpha_f$. We have here recovered the ratio m($\pi^\pm$)/m(e$^\pm$) which we found with the standing wave model of the $\pi^\pm$\,mesons, Eq.(65) of [13]. This seems to be a necessary condition for the validity of our model of the electron. We have thus shown that the \emph{rest mass of the electron can be explained} by the sum of the rest masses of the electron neutrinos in a cubic lattice with N$^\prime$/4 electron neutrinos $\nu_e$ and the mass in the sum of the energy of N/2 electric oscillations in the lattice, Eq.(9). The one oscillation added to the 2$\times$N$^\prime$/4 oscillations is the oscillation at the center of the lattice, Fig.(1). From this model follows, since it deals with a cubic neutrino lattice, that \emph{the electron is not a point particle}, which is unlikely to begin with, because at a true point the self-energy would be infinite. However, since neutrinos are non-interacting their presence will not be detected in electron-electron scattering experiments. \begin{figure}[h] \vspace{0.5cm} \hspace{2.2cm} \includegraphics{elat.eps} \vspace{-0.2cm} \begin{quote} Fig.1. Horizontal or vertical section through the central part of\\ \indent\hspace{1.1cm} the electron lattice. \end{quote} \end{figure} The \emph{rest mass of the muon} has been explained similarly with an oscillating lattice of muon and electron neutrinos [13]. We found that m($\mu^\pm)$/m(e$^\pm$) is\\ $\cong$ 3/2$\alpha_f$ = 205.55, nearly equal to the actual mass ratio 206.768, in agreement with what Nambu [16] found empirically. The heavy weight of the muon is primarily a consequence of the heavy weight of the N$^\prime$/4 muon neutrinos in the muon lattice. The mass of the muon neutrino is related to the mass of the electron neutrino through m($\nu_e)$ = $\alpha_f$m($\nu_\mu$), Eq.(39) of [13]. In order to confirm the \emph{validity} of our preceding explanation of the mass of the electron we must show that the sum of the charges of the electric oscillations in the interior of the electron is equal to the elementary electric charge of the electron. We recall that Fourier analysis requires that, after pair production, there must be a continuum of frequencies in the electron and positron. With h$\nu_0\alpha_f$ = e$^2$/\emph{a} from Eq.(6) follows from Eq.(5) that the oscillation energy in e$^\pm$ is the sum of 2$\times$(N$^\prime$/4 + 1) $\cong$ N/2 electric oscillations \begin{equation} \mathrm{E}_\nu(\mathrm{e}^\pm) = \frac{\mathrm{N}}{2} \cdot \frac{\mathrm{e^2}}{\emph{a}}\cdot \frac{f(T)}{2\pi}\, \int\limits_{-\pi}^{\pi}\,\phi\,\mathrm{d}\phi\,, \end{equation} with f(T) = 1/(e$^{h\nu/kT} \mathrm{-}$ 1) = 1/1.305$\cdot10^{13}$ from p.17 in [13]. Inserting the values for N, f(T) and \emph{a} we find that E$_\nu$(e$^\pm$) = 0.968\,m(e$^\pm$)c$^2$/2. The discrepancy between m(e$^\pm$)c$^2$/2 and E$_\nu$(e$^\pm$) so calculated must originate from the uncertainty of the parameters N, f(T) and \emph{a} in Eq.(13). We note that it follows from the factor e$^2$/\emph{a} in Eq.(13) that the oscillation energy is the same for electrons and positrons, as it must be. We replace the integral divided by 2$\pi$ in Eq.(13), which has the value $\pi$/2, by the sum $\Sigma\,\phi_k\Delta\,\phi$, where k is an integer number with the maximal value k$_m$ = (N/4)$^{1/3}$. $\phi_k$ is equal to k$\pi$/k$_m$ and we have \begin{displaymath} \Sigma\,\phi_k\,\Delta\phi = \sum_{k=1}^{k_m}\,\frac{k\pi}{k_m}\cdot\frac{1}{k_m} = \frac{ k_m(k_m + 1)\pi}{2\,k_m^2} \cong \frac{\pi}{2}\,, \end{displaymath} \noindent as it must be. The energy in the individual electric oscillation with index k is then \begin{equation} \Delta\mathrm{E}_\nu(k) = \phi_k\,\Delta\phi = k\pi/k_m^2\,. \end{equation} Suppose that the energy of the electric oscillations is correctly described by the self-energy of an electrical charge \begin{equation}\mathrm{U} = 1/2\,\cdot\,\mathrm{e}^2/\mathrm{r}\,. \end{equation} The self-energy of the elementary electrical charge is normally used to determine the mass of the electron from its charge, here we use Eq.(15) the other way around, we determine the charge from the energy in the oscillations. The charge of the electron is contained in the electric oscillations. That means that \emph{the electric charge is not concentrated in a point} but is distributed over N/4 = O($10^9)$ charge elements Q$_k$. \emph{The charge elements are distributed in a cubic lattice} and the resulting electric field is cubic, not spherical. For distances large as compared to the sidelength of the cube, (which is O($10^{-13}$)\,cm), say at the first Bohr radius which is on the order of $10^{-8}$\,cm, the deviation of the cubic field from the spherical field will be reduced by about $10^{-10}$. The charge in all electric oscillations is \begin{equation} \mathrm{Q} = \sum_{k}\,\mathrm{Q}_\mathrm{k}\,. \end{equation} Setting the radius r in the formula for the self-energy equal to 2\,\emph{a} we find, with Eqs.(13,14,15), that the charge in the individual electric oscillations is \begin{equation} \mathrm{Q_k} = \pm\,\sqrt{2\pi\,N\,e^2f(T)/k_m^2}\,\cdot\,\sqrt{k}\,. \end{equation} \noindent and with k$_m$ = 1/2\,(N/4)$^{1/3}$ = 447 and \begin{displaymath} \sum_{k=1}^{k_m}\,\sqrt{k} = 6310.8\, \end{displaymath} \noindent follows, after we have doubled the sum over $\sqrt{k}$, because for each index k there is a second oscillation on the negative axis of $\phi$, that \begin{equation} \mathrm{Q} = \Sigma\,\mathrm{Q_k} = \pm\,5.027\cdot10^{-10}\,\,\mathrm{esu}\,,\end{equation} \noindent whereas the elementary electrical charge is e = $\pm$\,4.803\,$\cdot\,10^{-10}$\,esu. That means that our theoretical charge of the electron is 1.047 times the elementary electrical charge. Within the uncertainty of the parameters the theoretical charge of the electron agrees with the experimental charge e. We have confirmed that it follows from our explanation of the mass of the electron that the electron has, within a 5\% error, the correct electrical charge. Each element of the charge distribution is surrounded in the horizontal plane by four electron neutrinos as in Fig.(1), and in vertical direction by an electron neutrino above and also below the element. The electron neutrinos hold the charge elements in place. We must assume that the charge elements are bound to the neutrinos by the weak nuclear force. The weak nuclear force plays here a role similar to its role in holding, for example, the $\pi^\pm$ or $\mu^\pm$ lattice together. It is not possibe, in the absence of a definitive explanation of the neutrinos, to give a theoretical explanation for the electro-weak interaction between the electric oscillations and the neutrinos. However, the presence of the range \emph{a} of the weak nuclear force in e$^2$/\emph{a} is a sign that the weak force is involved in the electric oscillations. The attraction of the charge elements by the neutrinos overcomes the Coulomb repulsion of the charge elements. The weak nuclear force is the missing non-electromagnetic force or the Poincar\'{e} stress which holds the elementary electric charge together. The same considerations apply for the positive electric charge of the positron, only that then the electric oscillations are all of the positive sign and that they are bound to anti-electron neutrinos. Finally we learn that Eq.(13) precludes the possibility that the charge of the electron sits only on its surface. The number N in Eq.(13) would then be on the order of $10^6$, whereas N must be on the order of $10^9$ so that E$_\nu$(e$^\pm$) can be m($\mathrm{e}^\pm)$c$^2$/2 as is necessary. In other words, the charge of the electron must be distributed throughout the interior of the electron, as we assumed. Summing up: The rest mass of the electron and positron originates from the sum of the rest masses of N$^\prime$/4 electron neutrinos or anti-electron neutrinos in cubic lattices plus the mass in the energy of N$^\prime$/2 electric oscillations in the neutrino lattices. That means that neither the electron nor the positron are point particles. The electric oscillations are attached to the neutrinos by the weak nuclear force. The sum of the charge elements of the electric oscillations accounts for the elementary charge of the electron, respectively positron. \section{The spin and magnetic moment \\ of the electron} The model of the electron we have proposed in the preceding chapter has, in order to be valid, to pass a crucial test; the model has to explain satisfactorily the spin and the magnetic moment of the electron. When Uhlenbeck and Goudsmit [8] (U\&G) discovered the existence of the spin of the electron they also proposed that the electron has a magnetic moment with a value equal to Bohr's magnetic moment $\mu_B$ = e$\hbar$/2m$(\mathrm{e}^\pm)$c. Bohr's magnetic moment results from the motion of an electron on a circular orbit around a proton. The magnetic moment of the electron postulated by U\&G has been confirmed experimentally, but has been corrected by about 0.11\% for the so-called anomalous magnetic moment. If one tries to explain the magnetic moment of the electron with an electric charge moving on a circular orbit around the particle center, analogous to the magnetic moment of hydrogen, one ends up with velocities larger than the velocity of light, which cannot be, as already noted by U\&G. It remains to be explained how the magnetic moment of the electron comes about. We will have to explain the spin of the electron first. The spin, or the intrinsic angular momentum of a particle is, of course, the sum of the angular momentum vectors of all components of the particle. In the electron these are the neutrinos and the electric oscillations. Each neutrino has spin 1/2 and in order for the electron to have s = 1/2 all, or all but one, of the spin vectors of the neutrinos in their lattice must cancel. If the neutrinos are in a simple cubic lattice as in Fig.(1) and the center particle of the lattice is not a neutrino, as in Fig.(1), the spin vectors of all neutrinos in the lattice cancel, $\Sigma\,j(n_i)$ = 0, provided that the spin vectors of the electron neutrinos of the lattice point in opposite direction at their mirror points in the lattice. Otherwise the spin vectors of the neutrinos would add up and make a very large angular momentum. We follow here the procedure we used in [17] to explain the spin of the muons. The spin vectors of all electron neutrinos in the electron cancel just as the spin vectors of all muon and electron neutrinos in the muons cancel because there is a neutrino vacancy at the center of their lattices, (Fig.(1) of [17]). We will now see whether the electric oscillations in the electron contribute to its angular momentum. As we said in context with Eq.(7) there must be two times as many electric oscillations in the electron lattice than there are neutrinos. The oscillation pairs can either be the two oscillations in a standing wave or they can be two circular oscillations. Both the standing waves and the circular oscillations are non-progressive and can be part of the \emph{rest mass} of a particle. We will now assume that the electric oscillations are circular. Circular oscillations have an angular momentum $\vec{j} = m\,\vec{r}\times\vec{v}$. And, as in the case of the spin vectors of the neutrinos, all or all but one of the O$(10^9)$ angular momentum vectors of the electric oscillations must cancel in order for the electron to have spin 1/2. As in [13] we will describe the superposition of the two circular oscillations by \begin{equation} x(t) = exp[i\omega t] + exp[-\,i(\omega t + \pi)]\,,\end{equation} \begin{equation} y(t) = exp[i(\omega t + \pi/2)] + exp[-\,i(\omega t + 3\pi/2)]\,\,, \end{equation} \noindent that means by the superposition of a circular oscillation with the frequency $\omega$ and a second circular oscillation with the frequency $\mathrm{-}\,\omega$. The latter oscillation is shifted in phase by $\pi$. Negative frequencies are permitted solutions of the equations of motion in a cubic lattice, Eqs.(7,13) of [13]. As is well-known oscillating electric charges should emit radiation. However, this rule does already not hold in the hydrogen atom, so we will assume that the rule does not hold in the electron either. In circular oscillations the kinetic energy is always equal to the potential energy and the sum of both is the total energy. From \begin{equation} \mathrm{E}_{pot} + \mathrm{E}_{kin} = 2\,\mathrm{E}_{kin} =\mathrm{E}_{tot} \end{equation} follows with E$_{kin}$ = $\Theta\,\omega^2$/2 and E$_{tot}$ = $\hbar\omega$ that 2\,E$_{kin}$ = $\Theta\,\omega^2$ = $\hbar\omega$. ${\Theta}$ is the moment of inertia. When we superpose the two circular oscillations with $\omega$ and $\mathrm{-}\,\omega$ of Eqs.(19,20) we have \begin{equation} 2\times2\,\mathrm{E}_{kin} = 2\,\Theta\,\omega^2 = \hbar\omega\,,\end{equation} from which follows that the angular momentum is \begin{equation} j = \Theta\,\omega = \hbar/2\,. \end{equation} That means that each of the O$(10^9$) pairs of superposed circular oscillations has an angular momentum $\hbar$/2. The circulation of the oscillation pairs in Eqs.(19,20) is opposite for all $\omega$ of opposite sign. It follows from the equation for the displacements u$_n$ of the lattice points \begin{equation} u_n = Ae^{i(\omega\,t\, +\, n\phi)}\,, \end{equation} \noindent (Eq.(5) in [13]) that the velocities of the lattice points are given by \begin{equation} v_n = \dot{u}_n = i\,\omega_n\,u_n\,. \end{equation} The sign of $\omega_n$ changes with the sign of $\phi$ because the frequencies are given by Eq.(13) of [13], that means by \begin{equation} \omega_n = \pm\,\omega_0\,[\,\phi_n + \phi_0\,]\,. \end{equation} Consequently the circulation of the electric oscillations is opposite to the circulation at the mirror points in the lattice and the angular momentum vectors cancel, but for the angular momentum vector of the electric oscillation at the\,\emph{ center of the lattice}. The center circular oscillation has, as all other electric oscillations, the angular momentum $\hbar$/2 as Eq.(23) says. The angular momentum of the entire electron lattice is therefore \begin{equation} j(\mathrm{e}^\pm) = \sum\,j(n_i) + \sum\,j(el_i) = j(el_0) =\hbar/2\,, \end{equation} as it must be for spin s = 1/2. The explanation of the spin of the electron given here follows the explanation of the spin of the baryons in [13], as well as the explanation of the absence of spin in the mesons. A valid explanation of the spin must be applicable to all particles, in particular to the electron, the prototype of a particle with spin. We will now turn to the magnetic moment of the electron which is known with extraordinary accuracy, $\mu(\mathrm{e}^\pm)$ = 1.001\,159\,652\,187\,$\mu_B$, according to the Review of Particle Physics [14], with $\mu_B$ being the Bohr magneton. The decimals after 1.00\,$\mu_B$ are caused by the anomalous magnetic moment which we will not consider. As is well-known the magnetic dipole moment of a particle with spin is, in Gaussian units, given by \begin{equation} \vec{\mu} = g\,\frac{e\hbar}{2mc}\,\vec{s}\,, \end{equation} where g is the dimensionless Land\'{e} factor, m the rest mass of the particle and $\vec{s}$ the spin vector. The g-factor has been introduced in order to bring the magnetic moment of the electron into agreement with the experimental facts. As U\&G postulated and as has been confirmed experimentally the g-factor of the electron is 2. With the spin s = 1/2 of the electron the magnetic dipole moment of the electron is then \begin{equation} \mu(\mathrm{e}^\pm) = \mathrm{e}\hbar/2\mathrm{m}(\mathrm{e}^\pm)c\,, \end{equation} or one Bohr magneton in agreement with the experiments, neglecting the anomalous moment. For a structureless point particle Dirac [9] has explained why g = 2 for the electron. However we consider here an electron with a finite size and which is at rest, which means that the velocity of the center of mass is zero. When it is at rest the electron has still its magnetic moment. Dirac's theory does therefore not apply here. The only part of Eq.(28) that can be changed in order to explain the g-factor of an electron with structure is the ratio e/m which deals with the spatial distribution of charge and mass. In the classical electron models the mass originates from the charge. However that is not necessarily always so. If part of the mass of the electron is non-electrodynamic and the non-electrodynamic part of the mass does not contribute to the magnetic moment of the electron, which to all that we know is true for neutrinos, then the ratio e/m in Eq.(28) is not e/m($\mathrm{e}^\pm$) in the case of the electron. The elementary charge e certainly remains unchanged, but e/m depends on what fraction of the mass is of electrodynamic origin and what fraction of m is non-electrodynamic, just as the mass of a current loop does not contribute to the magnetic moment of the loop. From the very accurately known values of $\alpha_f$, m($\pi^\pm$)c$^2$ and m(e$^\pm$)c$^2$ and from Eq.(9) for the energy in the electric oscillations in the electron E$_\nu$(e$^\pm$) = $\alpha_f$/2\,$\cdot$\,m($\pi^\pm$)c$^2$/2 = 0.996570\,m(e$^\pm$)c$^2$/2 follows that very nearly one half of the mass of the electron is of electric origin, whereas the other half of m($\mathrm{e}^\pm)$ is made of neutrinos which do not contribute to the magnetic moment. That means that in the electron the mass in e/m is practically m($\mathrm{e}^\pm$)/2. The magnetic moment of the electron is then \begin{equation} \vec{\mu}_e = g \frac{e\hbar}{2m(e^\pm)/2\cdot c}\vec{s}\,, \end{equation} and with s = 1/2 we have $\mu(\mathrm{e}^\pm)$ = g\,e$\hbar$/2m($\mathrm{e}^\pm$)c. Because of Eq.(29) the g-factor must be equal to one and is unnecessary. In other words, if the electron is composed of the neutrino lattice and the electric oscillations as we have suggested, then the electron has the correct magnetic moment $\mu_e$ = e$\hbar/2\mathrm{m(e^\pm)c}$, if exactly 1/2 of the electron mass consists of neutrinos. The preceding explanation of the magnetic moment of the electron has to pass a critical test, namely it has to be shown that the same considerations lead to a correct explanation of the magnetic moment of the muon $\mu_\mu$ = e$\hbar$/2m($\mu^\pm$)c, which is about 1/200th of the magnetic moment of the electron but is known with nearly the same accuracy as $\mu_e$. Both magnetic moments are related through the equation \begin{equation} \frac{\mu_\mu}{\mu_e} = \frac{\mathrm{m(e}^\pm)}{\mathrm{m}(\mu^\pm)} = \frac{1}{206.768}\,, \end{equation} as follows from Eq.(28) applied to the electron and muon. This equation agrees with the experimental results to the sixth decimal. The muon has, as the electron, an anomalous magnetic moment of about 0.11\,\% $\mu_\mu$, which is too small to be considered here. In the standing wave model [13] the muons consist of a lattice of N$^\prime$/4 muon neutrinos $\nu_\mu$, respectively anti-muon neutrinos $\bar{\nu}_\mu$, of N$^\prime$/4 electron neutrinos and the same number of anti-electron neutrinos plus an elementary electric charge. For the explanation of the magnetic moment of the muon we follow the same reasoning we have used for the explanation of the magnetic moment of the electron. We say that m($\mu^\pm)$ consists of two parts, one part which causes the magnetic moment and another part which does not contribute to the magnetic moment. The part of m($\mu^\pm$) which causes the magnetic moment must contain circular electric oscillations without which there would be no magnetic moment. It becomes immediately clear from the small mass of the electron neutrinos and from Eq.(5) for the energy of the electric oscillations in the electron that $\Sigma$\,m($\nu_e$) and E$_\nu$(e$^\pm$) are too small, as compared to the energy in the rest masses of all neutrinos in the muons, to make up m($\mu^\pm)$/2. However, the oscillations in the $\mu^\pm$\,mesons do not follow Eq.(5) for the oscillation energy in the electron, but rather Eq.(4) for the oscillation energy in the muons. Both differ by the factor $\alpha_f$ in Eq.(5). But even when the oscillation energy in the muons as given by Eq.(4) is considered, the energy of the electric oscillations in the muons would be only E$_\nu(\mu^\pm)$/4 = 16.955\,MeV, if the electric oscillations are attached to N/4 electron neutrinos, as is the case in the electron. It appears to be necessary to consider the case that the electric oscillations in the $\mu^\pm$\,mesons are attached to \emph{all} neutrinos of the electron neutrino type in the $\mu^\pm$ lattice, regardless whether they are electron neutrinos or anti-electron neutrinos. That would mean that the electric charge is distributed uniformly in the $\mu^\pm$ lattice. There are, as has been shown in the paragraph below Eq.(31) of [13], 3/4$\cdot$N neutrinos of the electron neutrino type in the muons, of which N/4 neutrinos originate from the charge e$^\pm$ carried by $\mu^\pm$. If the electric oscillations are attached to 3/4$\cdot$N electron neutrinos, regardless of their type, then the energy in all electric oscilllations or the energy in the electric charge is, with Eq.(8) and E$_\nu(\mu^\pm$) = E$_\nu(\pi^\pm$) = 67.82\,MeV from Eq.(31) in [13], as well as with m($\mu^\pm$)c$^2$ = 105.6583\,MeV, given by \vspace{1cm} \begin{eqnarray} \lefteqn{3/4\cdot\mathrm{E}_\nu(\mu^\pm) = 3/4\cdot 67.82\,\mathrm{MeV} = 50.865\,\mathrm{MeV}}\nonumber\\& = & 0.4814\,\mathrm{m}(\mu^\pm)\mathrm{c}^2 \cong 1/2\cdot \mathrm{m}(\mu^\pm)\mathrm{c}^2\,. \end{eqnarray} In other words, the energy in the electric oscillations or the electric charge makes up, in a good approximation, 1/2 of the mass of the muons. The other half of the rest mass of the muons consists of the sum of the rest masses of the neutrinos in the muon lattice plus the oscillation energy of the muon neutrinos, neither of which contributes to the magnetic moment. It is \begin{equation} 1/4\cdot\mathrm{E}_\nu(\mu^\pm) + \mathrm{N}/4\cdot\mathrm{m}(\nu_\mu)\mathrm{c}^2 + 3/4\cdot\mathrm{Nm}(\nu_e)\mathrm{c}^2 \\ = 53.347\,\mathrm{MeV} = 0.50490\,\mathrm{m}(\mu^\pm)\mathrm{c}^2\,. \end{equation} The theoretical total energy in the rest mass of the muons is then E(m($\mu^\pm$)) = 0.9863\,m($\mu^\pm$)c$^2$(exp). In simple terms, if E$_\nu(\pi^\pm)$ = E$_\nu(\mu^\pm$) = 1/2$\cdot$m($\pi^\pm$), not 0.486\,m($\pi^\pm$) as in Eq.(27) of [13], then it follows from 3/4$\cdot$E$_\nu(\mu^\pm$) = 3/8$\cdot$m($\pi^\pm$) and from the neutral part of the muon mass in Eq.(33), which is likewise $\approx$3/8$\cdot$m($\pi^\pm$), that the rest mass of the muons is m($\mu^\pm$) $\cong$ 3/8$\cdot$m($\pi^\pm$) + 3/8$\cdot$m($\pi^\pm$) = 3/4$\cdot$ m($\pi^\pm$), as it must be in a first approximation, whereas the actual m($\mu^\pm$) is 1.00937$\cdot$3/4$\cdot$m($\pi^\pm$). That means that in a good approximation the charged part of the rest mass of the muons is 1/2 of the mass of the muons. If the charged part of the muon mass as expressed by Eq.(32) makes up 1/2 of the mass of the muons and if the other part of the muon mass does not contribute to the magnetic moment, then the magnetic moment of the muon is given by \begin{equation} \vec{\mu}_\mu = \frac{\mathrm{e}\hbar} {2\mathrm{m}(\mu^\pm)/2\cdot\mathrm{c}}\cdot\vec{s}\,. \end{equation} With s = 1/2 we have $\mu_\mu$ = e$\hbar$/2m$(\mu^\pm)$c as it must be, without the artificial g-factor. \section*{Conclusions} One hundred years of sophisticated theoretical work have made it abundantly clear that the electron is not a purely electromagnetic particle. There must be something else in the electron but electric charge. It is equally clear from the most advanced scattering experiments that the ``something else" in the electron must be non-interacting, otherwise it could not be that we find that the radius of the electron must be smaller than $10^{-16}$\,cm. The only non-interacting matter we know of with certainty are the neutrinos. So it seems to be natural to ask whether neutrinos are not part of the electron. Actually we have not introduced the neutrinos in an axiomatic manner but rather as a consequence of our standing wave model of the stable mesons, baryons and $\mu$-mesons. It follows necessarily from this model that after the decay of the $\mu^-$\,meson there must be electron neutrinos in the emitted electron, and that they make up one half of the mass of the electron. The other half of the energy in the electron originates from the energy of electric oscillations. The theoretical rest mass of the electron agrees, within 1\% accuracy, with the experimental value of m(e$^\pm$). We have learned that the charge of the electron is not concentrated in a single point, but rather is distributed over O(10$^9$) elements which are held together with the neutrinos by the weak nuclear force. The sum of the charges in the electric oscillations is, within the accuracy of the parameters, equal to the elementary electrical charge of the electron. From the explanation of the mass and charge of the electron follows, as it must be, the correct spin and magnetic moment of the electron, the other two fundamental features of the electron. With a cubic lattice of anti-electron neutrinos we also arrive with the same considerations as above at the correct mass, charge, spin and magnetic moment of the positron. \section*{Acknowledgements} Contributions of Professor J. Zierep and of Dr. T. Koschmieder are gratefully acknowledged. \section*{References} \noindent [1] Thomson, J.J. Phil.Mag. {\bfseries44},293 (1897). \smallskip \noindent [2] Lorentz, H.A. \emph{Enzykl.Math.Wiss.} Vol.{\bfseries5},188 (1903). \noindent [3] Poincar\'{e}, H. Compt.Rend. {\bfseries 140},1504 (1905). Translated in:\\ \indent Logunov, A.A. \emph{On The Articles by Henri Poincare\\ \indent ``On The Dynamics of the Electron"}, Dubna JINR (2001). \smallskip \noindent [4] Ehrenfest, P. Ann.Phys. {\bfseries24},204 (1907). \smallskip \noindent [5] Einstein, A. Sitzungsber.Preuss.Akad.Wiss. {\bfseries20},349 (1919). \smallskip \noindent [6] Pauli, W. \emph{Relativit\"{a}tstheorie}, B.G. Teubner (1921). Translated in:\\ \indent Theory of Relativity, Pergamon Press (1958). \smallskip \noindent [7] Poincar\'{e}, H. Rend.Circ.Mat.Palermo {\bfseries21},129 (1906). \smallskip \noindent [8] Uhlenbeck, G.E. and Goudsmit, S. Naturwiss. {\bfseries13},953 (1925). \smallskip \noindent [9] Dirac, P.A.M. Proc.Roy.Soc.London A{\bfseries117},610 (1928). \smallskip \noindent [10] Gottfried, K. and Weisskopf, V.F. \emph{Concepts of Particle Physics},\\ \indent \,\,Vol.1,\,p.38. Oxford University Press (1984). \smallskip \noindent [11] Schr\"{o}dinger, E. Sitzungsber.Preuss.Akad.Wiss. {\bfseries24},418 (1930). \smallskip \noindent [12] Mac Gregor, M.H. \emph{The Enigmatic Electron}, Kluwer (1992). \smallskip \noindent [13] Koschmieder, E.L. http://arXiv.org/phys/0602037 (2006). \smallskip \noindent [14] Eidelman, S. et al. Phys.Lett.B {\bfseries 592},1 (2004). \smallskip \noindent [15] Born, M. and v.\,Karman, Th. Phys.Z. {\bfseries13},297 (1912). \smallskip \noindent [16] Nambu, Y. Prog.Th.Phys. {\bfseries7},595 (1952). \smallskip \noindent [17] Koschmieder, E.L. http://arXiv.org/physics/0308069 (2003),\\ \indent\,\, \emph{Muons: New Research}, Nova (2005). \end{document}
{ "timestamp": "2006-09-26T18:26:57", "yymm": "0503", "arxiv_id": "physics/0503206", "language": "en", "url": "https://arxiv.org/abs/physics/0503206" }
\section*{Introduction} Symplectic geometry (like many fields of geometry before it) has received a tremendous infusion of ideas from work on a single physics-inspired example. The celebrated paper of Atiyah and Bott \cite{AB82} created a boom in symplectic geometry when they interpreted the moduli space of a Riemann surface as the symplectic reduction of the space of connections over that surface considered as a Hamiltonian space. They were able to read off information about its cohomology by considering the square of the moment map as a Morse function (that this Morse function is exactly the Yang-Mills action reveals part of the physics inspiration behind their work). Kirwan \cite{Kirwan84} built on these ideas to prove that for any compact Hamiltonian space the square of the moment map is an equivariantly perfect Morse function (or Kirwan-Morse function, a weaker but sufficient notion), and used this to give an algorithm for computing the Betti numbers of the symplectic quotient. Not long afterwards inspiration came a second time from the same example: Witten \cite{Witten92} used quantum field theory ideas and some inventive symplectic geometry to find surprising formulas for the intersection pairings of the cohomology of the moduli space of a Riemann surface. Even for finite-dimensional Hamiltonian spaces Witten's techniques were not entirely rigorous, since he assumed regularities of the critical points of the squared moment map that do not typically hold. Jeffrey and Kirwan were able to reproduce his key results both for moduli space \cite{JK98a} and for general compact Hamiltonian spaces \cite{JK95} by replacing his main technique (which he called nonabelian localization) with an older technique of Duistermaat-Heckman \cite{DH83} called abelian localization. Specifically they were able to relate intersection pairings in the rational cohomology of a symplectic quotient to certain integrals of equivariant forms on the full Hamiltonian space. In the intervening decade symplectic geometry has been a booming field with much of the work centering on the cohomology of the symplectic quotient and its relationship to the topology of the original Hamiltonian space (a sampling includes \cite{Kalkman95,GK96,Vergne96,BV97,LMTW98,MS99,Paradan00,Kiem04,TW03,JKKW03,BTW04}). A remarkable feature of the work that has been driven by this one example is that none of the work actually applies to the original example (this is not entirely true: as mentioned above Jeffrey and Kirwan \cite{JK98a} manage to prove Witten's formulas for moduli space, but the geometry of the space of connections which inspired these results is entirely circumvented). The Hamiltonian space of interest is the space of connections, a space which is not just noncompact but in fact infinite dimensional, and thus a far cry from the finite-dimensional compact Hamiltonian spaces on which we usually focus. The failure of the results to apply to noncompact Hamiltonian spaces is particularly striking since the two most basic examples of Hamiltonian manifolds, $T^*G$ for $G$ a Lie group and any symplectic representation of a Lie group, are both noncompact. Naturally one would like to extend Jeffrey and Kirwan's approach to the noncompact setting. This may very well be possible, but one of the key benefits of reducing to the abelian case is the convexity results of Atiyah \cite{Atiyah82}, which apply only to compact spaces. More precisely, Witten's results apply to Hamiltonian spaces for which $0$ is a regular value of the moment map. In this situation a neighborhood of $\mu^{-1}(0)$ is always a Hamiltonian space with no fixed points for any subtorus, so any attempt to reduce questions of the topology of the reduced space to questions about the fixed points seems doomed. In particular, a number of authors have extended Duistermaat-Heckman localization to noncompact settings \cite{PW94,Paradan00,Libine05}. All these versions of Duistermaat-Heckman describe the induced measure on the Lie algebra at points away from $0,$ while in the case of a neighborhood of $\mu{-1}(0)$ when $0$ is regular the integrals in question give measures on the Lie algebra with support entirely at $0.$ Witten's original approach to nonabelian localization, however, makes no apparent reliance on compactness. In fact, if one were willing to join Witten in ignoring the analytic details and gave a sketchy introduction to equivariant cohomology in the Cartan model, Witten's argument could fit into a first year graduate course in differential geometry. The intrinsic simplicity of his argument suggests that even in the compact case it may be illuminating and productive to work out the analytic details, assuming they are tractable. They are indeed tractable, and this is the approach we take in this paper. What makes them tractable, and in fact not very difficult, is the ability to avoid the central problem: That the critical points of the square of the moment map, to which the integrals in question are supposed to localize, are in general singular spaces to which the differential geometry of forms and integration do not readily apply. It seems likely that Witten's nonabelian localization has much more to tell us, but that to make further progress will require understanding these singular spaces. for example, Paradan \cite{Paradan00} argues that the contribution to the Basic Integral from $\mu^{-1}(0)$ is still a polynomial even when $0$ is not a regular value. This suggests that it still localizes to an integral of some sort of cohomology class over the (now singular) reduced space. Since the higher critical sets can be built easily from $\mu^{-1}(0)$ of a related Hamiltonian space, we could then hope that the same is true for all critical sets. The paper is organized as follows. Section 1 gives the local characterization of an arbitrary Hamiltonian space due to Guillemin and Sternberg \cite{GS84a,GS84b}. It uses the local characterization to describe the critical points of the square of the moment map, and extend Kirwan's proof that the square of the moment map is an equivariantly perfect Morse-Kirwan function. Section 2 reviews the Cartan model for equivariant cohomology, expressing the Cartan and Kirwan maps explicitly in this language. Section 3 defines the Basic Integral and computes key estimates for it. Section 4 assumes that $0$ is a regular value of the moment map and proves the main theorem, that in this case the Basic Integral is the integral over the reduced space of the image under the Kirwan map of a certain form (which has polynomial dependence on $\epsilon$) plus additional contributions which are exponentially damped in $\epsilon.$ Thus cohomological integrals on the reduced space can be calculated by computing the Basic Integral over the full Hamiltonian space. \section{The Local Structure Of Hamiltonian Spaces}\label{sc_local} Let $M$ be a finite-dimensional smooth Hamiltonian space (not necessarily compact): That is a smooth manifold with symplectic form $\omega,$ acted on symplectomorphically by the compact Lie group $G$ with Lie algebra $\lieg,$ and with moment map $\mu \colon M \to \lieg^*.$ If $p \in M$ and $\xi \in \lieg$ we will write $V\!\xi$ for the vector field associated to the infinitesimal action of $\lieg$ on $M$ and $V_p\xi$ for the value of this vector field at the point $p.$ Then the moment map condition is \begin{equation} \label{eq:moment} \omega \intprod V\!\phi = d\mu \dotprod \phi \end{equation} for all $\phi \in \lieg,$ where we use $\intprod$ to represent the interior product between a vector field and a form (or a tangent vector and a form at a point) with the convention that $v \intprod \omega = (-1)^{\deg(\omega)}\omega \intprod v.$ Choose an invariant inner product $\bracket{\, \cdot \, , \, \cdot \,}_\lieg$ on $\lieg.$ This inner product determines an identification $\star \colon \lieg \to \lieg^*$ whose inverse we will also call $\star,$ so that $\mu^\star\colon M \to \lieg.$ Finally, choose an almost complex structure for $M$ compatible with the group action, that is to say an invariant metric $\bracket{\, \cdot \, , \, \cdot \,}_M$ and an operator $J$ on the tangent space such that $J^2=-1$ and $\bracket{x,y}=Jx\intprod \omega\intprod y.$ \subsection{Local Characterization} Guillemin and Sternberg (\cite{GS84a,GS84b}) give a local characterization of a Hamiltonian space which will be crucial for what follows. They show that for any point $p \in M,$ the Hamiltonian space $M$ is determined in an equivariant neighborhood of $p$ by the value of $\mu$ at $p,$ the Lie subgroup $H$ fixing $p$ and its Lie subalgebra $\lieh,$ and the symplectic action of $H$ on the tangent space $T_pM.$ More specifically, let $p \in M,$ with isotropy group $H \subset G,$ whose Lie algebra is $\lieh,$ and define $\beta=\mu^\star_p$ and $K \subset G$ the stabilizer of $\beta,$ with $\liek$ its Lie algebra (so that $\lieh \subset \liek$). Let $Y$ be the subspace of $T_pM$ of vectors which are omega-orthogonal and orthogonal to $V\liek,$ the space of directions $V\!\phi$ for $\phi \in \liek.$ This is an $H$ representation and the symplectic form restricts to a symplectic form $\omega_X$ on $X.$ On the space \[G \times ( \liek \oplus X)\] define the action of $G$ by the left action on the first component, and define the action of $H$ diagonally from the right action on the first component (applied to the inverse), the natural action on the second component, and the adjoint action on the third component. Define a closed invariant two-form $\omega$ at the point $(g,\nu + x)$ by \begin{equation}\label{eq:standard_omega} \omega=\bracket{d\nu - \frac{1}{2} [\beta+\nu,g^{-1}dg],g^{-1}dg}_G + \frac{1}{2} dx\intprod \omega_X \intprod dx. \end{equation} and moment maps for the two actions \begin{equation}\label{eq:G_moment} \mu_G \phi = \bracket{\Ad_g(\beta + \nu),\phi}_G \end{equation} \begin{equation}\label{eq:H_moment} \mu_H \phi = \bracket{\nu,\phi}_G + \frac{1}{2}x \intprod \omega \intprod \phi x. \end{equation} The symplectic reduction by $\mu_H$ (i.e., the quotient of $\mu^{-1}(0)$ by the action of $H$ is $G$-Hamiltonian space isomorphic to \[G \times_H (\liek/\lieh \oplus X)\] where we will interpret $\liek/\lieh$ as the subspace of $\liek$ perpendicular to $\lieh.$ One easily checks that the point $(1,0) \in G \times ( \liek \oplus X)$ is in $\mu_H^{-1}(0),$ has isotropy group $H,$ a tangent space isomorphic to $T_pM$ as an $H$-space, and moment value $\mu_G=\beta^\star.$ Therefore by \cite{GS84b}[Thm. 41.2] there is an isomorphism of Hamiltonian spaces from $G \times_H (\liek/\lieh \oplus X)$ to a neighborhood of $p$ in $M.$ In the future we will refer to the choice of such an isomorphism as ``choosing a standard neighborhood of $p.$'' \subsection{The Square of the Moment Map} We are interested in the critical sets of the nonnegative function $\bracket{\mu,\mu}_\lieg=|\mu|^2$ on $M.$ A point $ p\in M$ is a critical point for $|\mu|^2$ means that at $x,$ $d\bracket{\mu^\star,\mu^\star}_\lieg= 2\bracket{d\mu^\star,\mu^\star}_\lieg =d\mu \mu^\star= \omega \intprod V\! \mu^\star =0.$ Since $\omega$ is nondegenerate, to say that the one-form $V_p\mu^\star \intprod \omega$ is zero at $p$ is to say that $V\!\mu^\star$ is zero at $p,$ and thus the critical points of $|\mu|^2$ are exactly the zeros of the vector field $V\!\mu^\star.$ Equivalently, critical points are the zeros of the one-form \begin{equation}\label{eq_lambda} \lambda\intprod v \defequals \bracket{V\!\mu^\star, v}_M \end{equation} for $v$ a tangent vector on $M.$ If $p \in M$ and $G \times_H (\liek/\lieh \oplus X)$ is a standard neighborhood around $p$ then $p$ is a critical point for the square of the moment map if and only if the $H$-orbit of $(1,0)$ is critical in the standard neighborhood. This is equivalent to saying $V\!\beta=0,$ or $\beta \in \lieh.$ \begin{proposition} \label{pr_local_critical} Let $Z$ be the set of $x \in X$ such that $\beta x=0$ and $Q_x=0,$ where $\bracket{Q_x,\phi}= \frac{1}{2} x \intprod \omega_X \intprod \phi x$ for all $\phi \in \lieh.$ The connected component of the critical set of $|\mu|^2$ in $G \times_H(\liek/\lieh \oplus X)$ containing the $H$-orbit of $(1,0)$ is the $H$-orbit of all points $(g, z)$ where $g \in G $ and $z \in Z$. This space is an algebraic variety. \end{proposition} \begin{pf} Critical points of $|\mu|^2$ are points where $V\mu^\star=0.$ At a point $(g, \nu + x)$ in $G \times (\liek/\lieh \oplus X)$ we have \[\mu \cdot \phi=\bracket{Ad_{a}(\beta + \nu),\phi} + \frac{1}{2} \phi x\intprod \omega_X \intprod x.\] For a point $(g, \nu+x)$ in $\mu_H^{-1}(0)$ to descend to a point for which $V_G\phi=0$ means $V_G\phi$ is in $V_H \lieh.$ This requires that $\Ad_{g^{-1}}\phi \in \lieh,$ $[\Ad_{g^{-1}}\phi,\nu]=0,$ and $\Ad_{g^{-1}}\phi\dotprod x=0.$ The first condition when $\phi=\mu^\star$ implies that $\nu=0.$ The second implies nothing additional, and the third implies that $(\beta+Q_x)x=0.$ Thus the critical points are in general those for which $\nu=0$ and $(\beta+Q_x)x.$ The latter condition implies that $\bracket{\beta+Q_x,Q_x}=0,$ which in turn implies that $|\mu|^2 = |\beta+Q_x|^2 = |\beta|^2-|Q_x|^2.$ If there is a path of critical points connecting this to $(1,0),$ the value of $|\mu|^2$ would be constant, which implies that $Q_x=0$ and hence $\beta x=0.$ Of course all points of this form are critical and are obviously path connected to $(1,0),$ so the connected component of the critical set includes these points. We have only to show all other solutions are separated from this set. All other solutions have $\beta x \neq 0.$ Since $\beta$ commutes with $\lieh \subset \liek,$ write $X$ as a sum of orthogonal and $\omega$-orthogonal symplectic submodules $X_\alpha,$ on each of which $|\beta x| =\beta_\alpha |x|$ for some positive $\beta_\alpha.$ If $x$ satisfies $(\beta + Q_x) x=0,$ but not $\beta x=0$ then breaking $x$ into its components there is a nonzero $x_\alpha$ such that $-\beta x_\alpha = Q_x x_\alpha \neq 0.$ Since $|\beta x_\alpha|= \beta_\alpha |x_\alpha|$ at every solution $|Q_x| \geq \min_\alpha \beta_\alpha$ holds. Thus $|Q_x| > \frac{2}{3} \min_\alpha \beta_\alpha$ and $|Q_x| < \frac{1}{3} \min_\alpha \beta_\alpha$ separate $Z$ from all other solutions. \end{pf} \begin{corollary} \label{cr:describe_critical} The set of critical points of $|\mu|^2$ on the Hamiltonian space $M$ is a discrete union of closed connected components, each of which is locally an algebraic variety and on each one of which the value of $\mu$ lies in a single coadjoint orbit. \end{corollary} \begin{corollary} If $\mu$ is proper (that is the inverse image of compact sets is compact) then $|\mu|^2$ is a minimally degenerate equivariantly perfect Morse function in the sense of Kirwan \cite{Kirwan84}. \end{corollary} \begin{pf} In order for $|\mu|^2$ to be minimally degenerate we need to show that the critical set is a discrete union of compact sets on each of which $|\mu|^2$ is constant, and that for each of the sets there is a locally closed submanifold $\Sigma$ containing the critical set as a minimum and at each point in the critical set the tangent space to $\Sigma$ is a maximal subspace of the full tangent space on which the Hessian of $|\mu|^2$ is positive semidefinite. The description of the critical sets is exactly the content of the previous corollary, together with the properness of $\mu.$ The existence of such a $\Sigma$ follows by the argument given by Kirwan unmodified, as does the equivariant perfection of this function. \end{pf} \section{Equivariant de Rham Cohomology} An excellent reference on equivariant cohomology is \cite{GS99}, which gives a more complete and sophisticated treatment of everything in sections 2.1 and 2.2. \subsection{Equivariant Forms} Let $\mathcal{P}(\lieg)$ be the (graded) algebra of all complex-valued polynomial functions of $\lieg,$ $\mathcal{S}(\lieg)$ be the algebra of complex-valued Schwartz functions on $\lieg$ (that is, any combination of derivatives of the function times any power of $|\phi|$ approaches $0$ as $\phi\to \infty,$ with the supremums of these products as seminorms), $\mathcal{D}(\lieg)$ be the space of complex-valued tempered distributions, which is to say continuous linear functionals on $\mathcal{S}(\lieg),$ and $\mathcal{F}(\lieg)$ be the space of all continuous linear functionals on $\mathcal{P}(\lieg).$ Here and in the sequel we represent functions on $\lieg$ as formulas in a dummy variable $\phi \in \lieg.$ Each of the function spaces ($\mathcal{P}(\lieg),$ $\mathcal{S}(\lieg)$) is an algebra and $\mathcal{P}(\lieg)$ acts by multiplication on $\mathcal{S}(\lieg),$ inducing various actions of the function spaces on the dual spaces ($\mathcal{D}(\lieg),$ $\mathcal{F}(\lieg)$) all represented by multiplication. Also there are natural embeddings $\mathcal{P} \subset \mathcal{D},$ $\mathcal{S} \subset \mathcal{D},$ and $\mathcal{S} \subset \mathcal{F},$ sending $f(\phi)$ to $f(\phi) \dint \phi,$ where $\dint\phi$ represents Haar measure on $\lieg.$ By analogy with this embedding we will represent the pairing between a function space and its dual by $\int_\lieg \,\cdot.$ If $M$ is a smooth manifold and $\mathcal{X}$ represents one of $\mathcal{P}, \mathcal{S}, \mathcal{D}, \mathcal{F}$ we can define $\\Omega(M) \hat{\tensor} \mathcal{X}(\lieg^*)$ to be smooth sections of the bundle over $M$ which at each point $p \in M$ is the tensor product of $\Lambda(T_pM) \tensor \mathcal{X}(\lieg).$ Here smooth means that when any element of the given space dual to $\mathcal{X}$ is paired with the second factor, the result is a smooth ordinary form. When $\mathcal{X}$ is $\mathcal{P},$ this is an algebra graded by the form degree plus twice the polynomial degree. If $G$ acts smoothly on $M$ then $G$ acts naturally and consistently on these bundles (with the diagonal action of $G$ acting naturally on forms and by the dual of the adjoint action on functions on $\lieg$), so we may speak of the the $G$-invariant elements of each space. These are respectively the \emph{$\mathcal{X}$-equivariant forms on $M,$} though when $\mathcal{X}$ is $\mathcal{P}$ we drop the $\mathcal{P}$ and simply say \emph{equivariant forms on $M.$} The exterior derivative $d$ is defined on all four bundles, so consider the \emph{equivariant derivative} \begin{equation}\label{eq_equivariant_derivative} D \alpha= d\alpha + i V\!\phi\intprod \alpha \end{equation} where $V\!\phi$ represents the linear map from $\lieg$ to vector fields on $M.$ Note that $D$ is an equivariant map which increases degree by one and satisfies $D^2=0$ on invariant elements. Thus there are four equivariant cohomologies $H^{*,\mathcal{X}}_G(M),$ where in the case $\mathcal{X}=\mathcal{P}$ (the only case where the cohomology has an integer grading, the others have only a $\ZZ/2$ grading) we drop the $\mathcal{P}$ and write $H^{*}_G(M),$ the equivariant cohomology of $M.$ Equivariant differential forms give a model for the cohomology with complex coefficients of the homotopy quotient $M_G,$ which is the geometric significance of everything we do in this paper, but which is mentioned for the last time here. We say an $\mathcal{X}$-equivariant form has compact support if the closure of the set of points in $M$ where $\alpha$ is a nonzero function on $\lieg$ is compact. Equivariant $D$ preserves both these concepts and we call the cohomology generated by compactly-supported $\mathcal{X}$-equivariant forms $H^{*,\mathcal{X}}_{G,\compact}(M)$. The various products among $\mathcal{P},$ $\mathcal{S},$ $\mathcal{D},$ and $\mathcal{F}$ extend to products on the various equivariant forms by wedging the form component. For example if $\alpha$ is an $\mathcal{S}$-equivariant form and $\beta$ is a $\mathcal{D}$-equivariant form then $\alpha\beta$ is an $\mathcal{F}$-equivariant form. Note that equivariant $D$ satisfies the Leibniz rule on all such products. Finally, an $\mathcal{F}$-equivariant form can be paired with $1$ (i.e. integrated) to get an ordinary form, which can be integrated over an invariant submanifold $N$ (assuming $M$ is oriented, and that either the original form was compactly supported or $N$ is compact) by taking on the component of appropriate degree. Because $V\!\phi \intprod$ lowers form degree, \[\int_N \int_\lieg D\alpha \dint \phi= \int_N \int_\lieg d\alpha \dint \phi= \int_{\partial N} \int_\lieg \alpha \dint \phi\] which is zero when $N$ has no boundary. This fact can be viewed as an equivariant version of Stokes theorem and when applied to $N=M$ descends for example to a well-defined pairing on cohomology \[H^{*,\mathcal{S}}_{G}(M)\times H^{*,\mathcal{D}}_{G,\compact}(M) \to \CC\] and likewise with the compact subscript on the other factor. \subsection{The Cartan and Kirwan Maps} If the group action is locally free, the homotopy quotient retracts to the ordinary quotient and thus the equivariant cohomology is isomorphic to the ordinary cohomology of the quotient. This isomorphism can be made completely explicit on the level of equivariant forms. Let $P \to N$ be an orbifold principal $G$-bundle, which is to say locally $P$ can be identified with $G \times_H V$ where $H$ is a finite subgroup of $G$ and $V$ is an $H$-module, so that the $G$ orbit of each point in $G \times_H V$ is a fiber of the map $P \to N.$ Let $A$ be a connection for this bundle, i.e. an equivariant $\lieg$-valued one-form on $P$ such that $A\intprod V\!\phi=\phi$ for all $\phi \in \lieg.$ Let $P_A$ be the operator on $TP$ which sends a tangent vector $v$ to its projection onto the $A=0$ subspace, $P_Av=v - V\!A\intprod v.$ If $\alpha$ is a form on $P,$ define $P_A^*\alpha$ so that $v\intprod P_A^*\alpha = P_A^*(P_A(v) \intprod \alpha),$ i.e. $P_A^*\alpha$ is $\alpha$ projected onto the subspace of forms zero on all vertical vectors. This map extends naturally to equivariant forms. Define the \emph{Cartan map} $\Cartan \colon \Omega(P) \hat{\tensor} \mathcal{P}(\lieg) \to \Omega(P)$ by \begin{equation}\label{weyl_map} \Cartan(\alpha(\phi))=P_A^*(\alpha(iF_A)) \end{equation} where $F_A$ refers to the $\lieg$-valued curvature two-form of the connection and its placement in parentheses denotes substituting its value for $\phi$ in the second tensor factor of $\alpha,$ thus producing a form to be wedged with the first tensor factor. \begin{proposition} The Cartan map descends to a grade-preserving isomorphism from the complex of equivariant forms to that of basic (i.e. horizontal and invariant) forms on $P$ inverting the natural imbedding. Composing with the natural isomorphism of the complex of basic forms on $P$ with ordinary forms on $N,$ we get a map which descends to an isomorphism \[\Cartan \colon H^*_G(P)\to H^*(N).\] \end{proposition} \begin{pf} If $\alpha$ is an equivariant form on $P,$ it is clear that $\Cartan(\alpha)$ is invariant, by the equivariance of $P_A$ and $F_A.$ It is also clear that $\Cartan(\alpha)$ is horizontal, since $F_A$ is horizontal and the range of $P_A^*$ is horizontal. Finally, it is clear that the Cartan map is an algebra homomorphism. So for the homomorphism of complexes we need only show that the Cartan map intertwines the equivariant and ordinary derivatives, which can be checked locally. To do this consider a chart $V$ on which a finite subgroup $H$ of $G$ acts, and an equivariant isomorphism of $G \times_H V$ with a neighborhood in $P.$ $A$ defines an $H$-invariant one-form $\tau$ on $V$ with values in $\lieg,$ by $\tau_{v}\intprod \xi= A_{(1,v)}\intprod (0,\xi).$ A form on $G \times_H V$ is an $H$-invariant form on $G \times V.$ Since $G$ only acts on the first factor, \[\left[\left(\Omega(G\times V)\right)^H \hat{\tensor} {\mathcal P}(\lieg^*)\right]^G \iso \left(\left[\Omega(G) \hat{\tensor} {\mathcal P}(\lieg^*) \right]^G\times \Omega(V)\right)^H\] as complexes. Since $D$ and $d$ satisfy the Leibniz rule, we can check the intertwining on generators of the complex. These are forms on $V,$ one-forms on $G$ $\bracket{\xi,g^{-1} dg }_G$ for $\xi \in \lieg,$ and functions $\bracket{\xi, g^{-1}\phi g}_G$ for $\xi \in \lieg.$ That $D$ and $d$ are intertwined on the first set of generators is obvious. For the second class \begin{eqnarray*} \Cartan(D(\bracket{\xi,g^{-1}dg })) &=& \Cartan(\bracket{\xi, g^{-1} dg g^{-1} dg } + i \bracket{\xi, g^{-1} \phi g})\\ &=& \bracket{\xi, \frac{1}{2}[\tau, \tau]} -\bracket{\xi, d\tau + \frac{1}{2}[\tau,\tau]}\\ &=& -\bracket{\xi,d\tau}\\ &=& -d(\bracket{\xi,\tau})\\ &=& d(\Cartan(\bracket{\xi,g^{-1}dg })). \end{eqnarray*} For the third class \begin{eqnarray*} \Cartan(D(\bracket{\xi,g^{-1} \phi g }))&=& \Cartan(\bracket{\xi, [g^{-1}dg ,g^{-1} \phi g] } \\ &=& -i\bracket{\xi, [\tau, d\tau +\frac{1}{2} [\tau, \tau]]}\\ &=& i \bracket{\xi, [d\tau,\tau]}\\ &=& i d(\bracket{\xi,d\tau + \frac{1}{2}[\tau,\tau]})\\ &=& d(\Cartan(\bracket{\xi,g^{-1} \phi g })). \end{eqnarray*} Finally, to see that its inverse is the natural embedding of basic forms into equivariant forms, since it is the identity on basic forms, we need only check that every closed equivariant form is cohomologous to a basic form. This requires defining certain operators on the complex of equivariant forms. We write $\frac{\partial}{\partial \phi}$ for the formal derivative with respect to $\phi,$ which we view as a function on $\lieg$ with values in equivariant forms. Thus the operator \[\Phi=A\cdot \frac{\partial }{\partial \phi}\] denotes (viewing the connection $A$ as a form tensored with a Lie algebra element) applying this operator on the equivariant form to the second tensor factor and wedging the first tensor factor with the result. By a similar logic $VA \intprod $ applies $V$ to the second tensor factor to get a tangent vector, takes the interior product with the form on which the operator acts to get a new form, and wedges the first factor with the result. Now a straightforward calculation shows \[ D \Phi + \Phi D = dA \cdot \frac{\partial}{\partial \phi} + i( \phi \cdot \frac{\partial}{\partial \phi} + V\!A \intprod \,)\] where the two new operators in the above expression are defined similarly. The first operator in the parentheses ($\phi \cdot \frac{\partial}{\partial \phi}$) multiplies any homogenous polynomial by its degree, and thus gives a grading of the space of equivariant forms into eigenvalues. Similarly the second operator in the parentheses ($V\!A \intprod\,$) grades the space into eigenspaces with nonnegative integral eigenvalues, representing the ``number of form degrees in vertical directions.'' Since the two commute, they give a grading by their sum, call it the total degree, such that the total degree zero piece consists of basic forms on $P.$ Notice that the term not in parentheses ($dA \cdot \frac{\partial}{\partial \phi}$) strictly lowers total degree. Thus if $\alpha$ is a closed equivariant form whose maximum total degree piece has degree $p>0,$ then $\alpha + \frac{i}{p}D(\Phi\alpha)$ has strictly lower degree, and thus by induction $\alpha$ is cohomologous to a total degree zero form. \end{pf} Now suppose that $M$ is a Hamiltonian space with a proper moment map, and that $0$ is a regular value of $M,$ i.e. that $d\mu$ is onto for all points with $\mu=0.$ \begin{proposition}\label{pr:orbifold} If $d\mu$ is onto for each point of $\mu^{-1}(0),$ then $G$ acts on $\mu^{-1}(0)$ with finite stabilizers. In this case $\mu^{-1}(0)$ is a smooth manifold and an orbifold principal bundle over the quotient $M_{\text{red}}=\mu^{-1}(0)/G,$ which has an orbifold symplectic structure $\omega_0.$ \end{proposition} \begin{pf} If $d\mu$ is onto at some point, then by the moment map condition $V\!\phi$ is nonzero for all $\phi \in \lieg,$ so that the isotropy group must be finite. If the isotropy group is finite at some $z$ with $\mu_z=0,$ then a standard neighborhood looks like \[G \times_H (\lieg \oplus X)\] where $H$ is a finite subgroup and $X$ is a symplectic vector space on which $H$ acts. The subspace on which $\mu=0$ is $G \times_H X.$ The image of this space in the quotient by $G$ is isomorphic to $X/H,$ the quotient of a vector space by a finite-dimensional group action. If another standard neighborhood $G \times_K X'$ contains $z,$ we argue the diffeomorphism of standard neighborhoods lifts to a local diffeomorphism of $X$ and $X'.$ This guarantees that a covering collection of standard neighborhoods form an orbifold atlas for $M_{\text{red}}=\mu^{-1}(0)/G.$ To see this, we can assume by equivariance that the standard neighborhood $G \times_K X'$ is chosen so that $z$ is the image of a point $(1,x).$ Then $H$ is the subgroup of $K$ which fixes $x,$ so that a neighborhood of $x$ is a representation $V$ of $H,$ and does not intersect with its image under any other elements of $K.$ Then $G \times_K X'$ is diffeomorphic locally to $G \times_H V,$ and this induces a diffeomorphism between $V$ and $X.$ A symplectic structure on an orbifold is a choice of $H$-invariant symplectic form on $V$ for each chart $(V,H),$ which is preserved by the overlap maps. Clearly $\omega_X$ is an invariant form on each vector space $X,$ and it is immediate that it is preserved by the overlap map. \end{pf} The imbedding of $\mu^{-1}(0)$ into $M$ gives a map of equivariant cohomology which when composed with the Cartan map gives the \emph{Kirwan map} $\Kirwan \colon H^*_G(M) \to H^*(M_{\text{red}}).$ The fact that the square of the moment map is equivariantly perfect means that this map is surjective. \section{Equivariant Integration and Localization} For this section let $M$ be a Hamiltonian space with a proper moment map, and $\epsilon$ be a positive real parameter. \subsection{Localization} The moment map condition guarantees that the equivariant form \[\omega + i \mu \dotprod \phi\] is closed, and thus represents an element of $H^*_G(M).$ Here the exponentiation is interpreted as its power series. Suppose now that $\alpha$ is an equivariant form on $M$, so that \[\alpha \exp(\omega + i \mu \dotprod \phi - \frac{\epsilon}{2} |\phi|^2)\] is an $\mathcal{S}$-equivariant form which is closed and/or compactly-supported if $\alpha$ is. On the other hand consider an invariant ordinary one-form $\lambda$ on $M$ (which is therefore also an equivariant one-form). \begin{lemma}\label{lm:Dlambda} For each nonnegative real $t$ \[\int_0^t \exp(sD\lambda) \dint s \lambda\] gives a $\mathcal{D}$-equivariant form satisfying $D\left(\int_0^t \exp(sD\lambda) \dint s \lambda\right)= \exp(tD\lambda)-1.$ Thus $\exp(tD\lambda)$ is a closed $\mathcal{D}$-equivariant form which is $\mathcal{D}$-cohomologous to $1.$ Further, on a submanifold of $M$ on which $\lambda\intprod V\!\phi$ is never the zero functional on $\phi,$ the limit of this integral as $t$ approaches infinity exists in the $\mathcal{D}$-topology and satisfies $D\left(\int_0^\infty \exp(sD\lambda) \dint s \lambda\right)= -1.$ \end{lemma} \begin{pf} We interpret the exponential and the integral in terms of power series, and at a point in $M$ write $\lambda \intprod V\!\phi$ as $\bracket{\xi,\phi}$ for some $\xi\in \lieg,$ so that the integral is a sum of terms of the form \[\bracket{\text{FORM}} \int_0^t s^k \exp(i s \bracket{\xi, \phi}) \dint s \] which as a functional on some test function $f(\phi) \in \mathcal{S}(\lieg)$ is \[\bracket{\text{FORM}}\int_0^t s^k \int_\lieg f(\phi) \exp(is \bracket{\xi, \phi}) \dint \phi \dint s = \int_0^t s^k \widehat{f}(s\xi) \dint s \] where $\widehat{f}$ is the Fourier transform of $f$ (ignoring arbitrary constants) and thus is well-defined. So $\int_0^t \exp(sD\lambda) \dint s \lambda$ is a $\mathcal{D}$-equivariant form whose equivariant derivative is \[\int_0^t \exp(sD\lambda) D\lambda \dint s = 1-\exp(tD\lambda).\] If $\lambda \intprod V\!\phi$ is never zero then $\xi\in \lieg$ as defined in the previous paragraph is never zero, so we get \[\int_0^\infty s^k \widehat{f}(s\xi) \dint s\] which converges since $\widehat{f}$ is Schwartz. \end{pf} \subsection{The Basic Integral} Since $\mu$ is proper by Corollary \ref{cr:describe_critical} identify the critical values of $|\mu|^2$ as \[0 \leq r_1 < r_2 < \cdots\] (the sequence may be finite or infinite) and as long as $r\in \RR^+$ is regular, i.e. satisfies $r \neq r_i$ $\forall i \in \NN$ then \[M_r \defequals \{p \in M \,|\, |\mu_p|^2 \leq r\}\] is a compact manifold with compact boundary. Recall that the symplectic form gives a natural orientation to $M$ and hence $M_r$ and thus integration over $M_r$ when $r$ is a regular value of $|\mu|^2$ is well-defined. Let $\lambda$ be the invariant one-form on $M$ which for any tangent vector $v$ gives \begin{equation}\label{eq:lambda_def} \lambda \intprod v = \bracket{V\!\mu^\star, v}. \end{equation} The $\lambda\intprod V\!\phi$ is zero exactly when $V\!\mu^\star$ is zero, which in turn happens exactly at the critical points of $|\mu|^2.$ For a equivariant form $\alpha,$ for any nonnegative real number $t$ and for any regular value $r$ of $|\mu|^2$ define the \emph{Basic Integral} \begin{equation}\label{eq:regularized_int} \BI(\alpha, r, t) \defequals \frac{1}{K}\int_\lieg \int_{M_r} \alpha \exp[\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2 + t D\lambda] \dint \phi \end{equation} where $\lambda$ defined in Equation (\ref{eq:lambda_def}) and \begin{equation}\label{eq:int_factor} K= \vol(G) (2\pi)^{\dim(G)}. \end{equation} The following estimates are crucial to the calculations that follow. \begin{lemma}\label{lm:bound_higher} Suppose $\alpha$ is an equivariant form and $r$ and $s$ are regular values of $|\mu|^2$ with $s<r.$ Then \[|\BI(\alpha,r,0)-\BI(\alpha,s,0)| < \text{POLYNOMIAL}(\epsilon^{\pm 1/2})\exp(-\frac{s}{2\epsilon})\] where the coefficients of the polynomial depend on $r.$ In other words the contribution to the Basic Integral at $t=0$ of points with large values of $|\mu|$ is exponentially damped. \end{lemma} \begin{pf} \begin{eqnarray*} & &\left|\frac{1}{K}\int_\lieg \int_{M_r-M_s} \alpha(\phi) \exp(\omega + i \mu \dotprod \phi - \frac{\epsilon}{2} |\phi|^2)\dint \phi\right| \\ &=& \frac{1}{K} \left| \int_{M_r-M_s} \exp(\omega) \int_\lieg \alpha(\phi) \exp(i \mu \dotprod \phi - \frac{\epsilon}{2} |\phi|^2) \dint \phi \right|\\ &=& \frac{1}{K} \left| \int_{M_r-M_s} \exp(\omega) \exp(-\frac{1}{2\epsilon} |\mu|^2) \int_\lieg \alpha(\phi+i \mu^\star/\epsilon) \exp(-\frac{\epsilon}{2} |\phi|^2) \dint \phi \right|\\ &=& \frac{1}{K} \left| \int_{M_r-M_s} \exp(\omega) \exp(-\frac{1}{2\epsilon} |\mu|^2) \text{POLYNOMIAL}(\epsilon^{\pm1/2}) \right|\\ &\leq& \exp(-\frac{s}{2\epsilon})\left| \text{POLYNOMIAL}(\epsilon^{\pm1/2})\right|. \end{eqnarray*} Here the coefficients of the polynomial can be bounded by certain integrals over $M_r.$ \end{pf} \begin{lemma}\label{lm:large_t} Suppose that $\alpha$ is a \emph{closed} equivariant form and that $r$ is a regular value of $|\mu|^2.$ Then \[ \lim_{t\to \infty} \BI(\alpha, r, t)\] exists and differs from $\BI(\alpha, r, 0)$ by \[ \text{POLYNOMIAL}(\epsilon^{\pm 1/2})\exp(-\frac{C}{2\epsilon})\] where the coefficients of the polynomial and $C$ depend on $r.$ \end{lemma} \begin{pf} Suppose $t_1<t_2 \in \RR.$ \begin{eqnarray*} &&|\BI(\alpha,r, t_2)-\BI(\alpha, r, t_1)|\\ &=&\frac{1}{K}\left| \int _\lieg \int_{M_r} \alpha(\phi) \exp(\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2) (\exp(t_2D\lambda)-\exp( t_1 D\lambda) ) \dint \phi \right| \\ &=& \frac{1}{K}\left| \int_\lieg \int_{M_r} \alpha(\phi) \exp(\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2) D\left(\int_{t_1}^{t_2} \exp(sD\lambda) \lambda \dint s\right) \dint \phi\right| \\ &=&\frac{1}{K}\left| \int_\lieg \int_{\partial {M_r}} \alpha(\phi) \exp(\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2) \int_{t_1}^{t_2} \exp(sD\lambda) \lambda \dint s \dint \phi \right| \\ &=& \frac{1}{K}\Big| \int_{\partial {M_r}} \int_{t_1}^{t_2} \exp(\omega +sd\lambda)\lambda \int_\lieg \alpha(\phi) \exp(i \bracket{\mu^\star + sV^\star V\!\mu^\star, \phi} -\frac{\epsilon}{2} |\phi|^2) \dint \phi \dint s\Big|. \end{eqnarray*} Completing the square yields \begin{eqnarray*} &=& \frac{1}{K}\Big| \int_{\partial {M_r}} \int_{t_1}^{t_2} \exp(\omega +sd\lambda)\lambda \int_\lieg \alpha(\phi + \frac{1}{2\epsilon}(\mu^\star + sV^\star V\!\mu^\star) \exp(-\frac{\epsilon}{2} |\phi|^2)\dint \phi \\ &&\qquad \cdot \exp( -\frac{1}{2\epsilon} |\mu|^2 -\frac{s}{2\epsilon} |V\!\mu^\star|^2 -\frac{s^2}{2\epsilon} |V^\star V\!\mu^\star|^2) \dint s \Big| \\ &=& \frac{1}{K}\Big| \int_{\partial {M_r}} \int_{t_1}^{t_2} \exp(\omega +sd\lambda)\lambda \text{POLYNOMIAL}(\epsilon^{\pm1/2},s) \\ &&\qquad \cdot \exp( -\frac{1}{2\epsilon} |\mu|^2 -\frac{s}{2\epsilon} |V\!\mu^\star|^2 -\frac{s^2}{2\epsilon} |V^\star V\!\mu^\star|^2) \dint s \Big| \\ &\leq & \frac{1}{K}\Big| \int_{\partial {M_r}} \exp( -\frac{1}{2\epsilon} |\mu|^2) \text{FORM} \int_{t_1}^{t_2} \text{POLYNOMIAL}(\epsilon^{\pm1/2},s) \\ &&\qquad \cdot \exp( -\frac{s^2}{2\epsilon} |V^\star V\!\mu^\star|^2) \dint s \Big|. \end{eqnarray*} For a fixed $\epsilon,$ since the $s$ integral is a polynomial times a Gaussian, this quantity is bounded by $\exp(-t_1^2 \min(|V^\star V\!\mu^\star|^2)/(2\epsilon)),$ where $ |V^\star V\!\mu^\star|^2$ is bounded below since $r$ is a regular value. The limit of the difference can be written as a telescoping sum of such differences, which decrease hypergeometrically and hence the sum converges. On the other hand choosing $t_1=0$ we see that there is a polynomial times $\epsilon^{\pm 1/2}$ which times $\exp(-C/(2\epsilon))$ bounds the difference regardless of $\epsilon$ or $t_2.$ \end{pf} \subsection{The Basic Integral as a Sum of Contributions} The large $t$ limit of the Basic Integral is a sum of contributions from the critical points of $|\mu|^2,$ as is illustrated in the following. \begin{lemma} \label{lm:r_independence} If $r$ and $s$ are regular values of $|\mu|^2$ with no critical values between them and $\alpha$ is a closed equivariant form then \[\lim_{t\to \infty} \BI(\alpha, r, t)=\lim_{t\to \infty} \BI(\alpha, s, t).\] \end{lemma} \begin{pf} This follows directly from Lemma \ref{lm:Dlambda}. \end{pf} \begin{corollary} \label{cr:contribution} For each $i$ choose $r_i'$ and $r_i''$ such that $r_{i-1}<r_i' < r_i < r_i'' < r_{i+1}.$ Define $r_1'=0$ and if $r_i$ is the maximum critical value choose any $r_i''>r_i.$ Then given a closed equivariant form $\alpha$ the quantity \[C_i(\alpha)=\lim_{t\to \infty} \BI(\alpha, r_i'',t) - \BI(\alpha, r_i'',t)\] exists and is independent of the choice of $r_i'$ and $r_i''.$ Further, for any regular value $r$ of $|\mu|^2$ \[\lim_{t\to \infty} \BI(\alpha, r, t)= \sum_{r_i<r} C_i.\] In other words the large $t$ limit of the Basic Integral up to $r$ is the sum of the contributions from each critical set below $r.$ The contribution $C_i(\alpha)$ when $r_i>0$ is bounded by \[\text{POLYNOMIAL}(\epsilon^{\pm 1/2} \exp(-\frac{r_i-\delta}{2\epsilon})\] where $\delta$ can be made as small as we like. \end{corollary} \begin{lemma} \label{lm:lambda_invariance} Let $\alpha$ be any closed equivariant form, let $r_i$ be a critical value of $|\mu|^2,$ let $N$ be a compact manifold with boundary containing a neighborhood of the critical set corresponding to $r_i$ and no other critical points of $|\mu|^2,$ and let $\lambda'$ be the result of an isotopy of $\lambda$ such that the points of $M$ at which $\lambda \colon V\!\phi$ is the zero functional on $\lieg$ remain fixed through the isotopy. Then \[ C_i(\alpha) = \lim_{t\to \infty} \int_\lieg \int_N \alpha \exp(\omega + i \mu \phi + t D\lambda' -\frac{\epsilon}{2}|\phi|^2) \dint \phi. \] \end{lemma} \begin{pf} By Lemma \ref{lm:Dlambda} the limit above with $\lambda$ replacing $\lambda'$ is equal to $C_i(\alpha).$ Define $\lambda''$ to agree with $\lambda'$ in a neighborhood of the critical set but to agree with $\lambda $ near the boundary of $N.$ Then \begin{eqnarray*} &&\frac{1}{K}\int_\lieg \int_{N} \alpha \exp(\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2) \\ && \qquad \cdot \left(\exp(tD\lambda) - \exp(tD\lambda'')\right)\dint \phi\\ &=& \frac{1}{K}\int_\lieg \int_{N} D\Big( \alpha \exp(\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2) \\ &&\qquad \cdot \left( \int_0^t \exp(sD\lambda) \dint s \lambda - \int_0^t \exp(sD\lambda'') \dint s \lambda'' \right)\Big)\dint \phi\\ &=& \frac{1}{K}\int_\lieg \int_{\partial N} \alpha \exp(\omega + i \mu \dotprod \phi -\frac{\epsilon}{2} |\phi|^2)\\ &&\qquad \cdot \left( \int_0^t \exp(sD\lambda) \dint s \lambda - \int_0^t \exp(sD\lambda'') \dint s \lambda'' \right)\dint \phi\\ &=& 0 \end{eqnarray*} so that $\lambda$ and $\lambda''$ give the same contribution. On the other hand by replacing $N$ by a smaller neighborhood (again by Lemma \ref{lm:Dlambda}), we can assure that $\lambda'$ and $\lambda''$ agree on $N$ and thus give the same contribution. \end{pf} \begin{proposition} \label{pr:bound_higher} Suppose that $\alpha$ is a closed equivariant form and $r $ is a regular value of $|\mu|^2.$ Then the large $t$ limit of the Basic Integral (\ref{eq:regularized_int}) is equal to its contribution $C_0(\alpha)$ of the critical set with $\mu=0$ (as defined in Corollary \ref{cr:contribution}) plus a contribution bounded by $\exp(-c/\epsilon)$ for some $c.$ \end{proposition} \begin{pf} This follows immediately from Lemma \ref{lm:bound_higher}. \end{pf} \section{When Zero is a Regular Value of the Moment Map} The proof of the following result in the case of trivial isotropy group appears in \cite{GS84b}, the full statement appears in \cite{Jeffrey99}. While the statement and proof are widely known to experts, to the author's knowledge no proof appears in the literature, so for the sake of completeness it is included here. \begin{proposition}\label{pr:normal_form} If $0$ is a regular value of $\mu$ (i.e. if $d\mu$ is onto for each point of $\mu^{-1}(0)$) recall by Proposition \ref{pr:orbifold} the map $\pi\colon \mu^{-1}(0) \to M_{\text{red}}=\mu^{-1}(0)/G$ is a principal orbifold bundle and $M_{\text{red}}$ has an orbifold symplectic structure $\omega_0.$ Given a connection $A$ on this bundle, there is an isomorphism of Hamiltonian spaces between a neighborhood of $\mu^{-1}(0)$ in $M$ and the Hamiltonian space $\mu^{-1}(0) \times \lieg,$ with symplectic form and moment map at $(p,\nu) \in \mu^{-1}(0) \times \lieg$ given by \begin{equation} \widetilde{\omega}=\pi^* \omega_0 + d\bracket{\nu,A} \end{equation} \begin{equation} \widetilde{\mu}=\nu^\star. \end{equation} \end{proposition} \begin{pf} One readily checks that $\widetilde{\omega}$ defines a closed form which is nondegenerate at $\mu^{-1}(0),$ and therefore in a neighborhood. Also $\widetilde{\omega}$ is manifestly $G$-invariant (with the diagonal action of $G$ on $\mu^{-1}(0) \times \lieg$) and satisfies the moment map condition with $\widetilde{\mu}.$ By Guilleman and Sternberg's local characterization \cite{GS84b}[Thm.41.2], it suffices to give an equivariant symplectic isomorphism between the zeros of the moment map in each case, and then extend it to an equivariant identification of the normal bundles which preserves $d\mu.$ The equivariant symplectic isomorphism is of course the natural imbedding of ${\widetilde{\mu}}^{-1}(0)= \mu^{-1}(0) \times \{0\}$ into $M.$ Its equivariance is by naturality and it preserves $\omega$ by inspection. Because $d\mu$ is onto at every point it gives a trivialization of the normal bundle, identifying it with $\mu^{-1}(0) \times \lieg.$ This identification clearly is equivariant and takes $d\widetilde{\mu}$ to $d\mu.$ \end{pf} \begin{theorem} \label{th:zero_contribution} Suppose $\alpha$ is a closed equiviariant form, and $0$ is a regular value for the moment map. Then the contribution $C_0(\alpha)$ to the Basic Integral (\ref{eq:regularized_int}) from $\mu^{-1}(0)$ is \[\int_\lieg \dint \phi \int_{M_{\text{red}}} \Kirwan(\alpha) exp(\omega_0 + \frac{\epsilon}{2} c_2)\] where $\Kirwan$ is the Kirwan map, $M_{\text{red}}$ is the orbifold quotient $\mu^{-1}(0)/G$ and $c_2$ is the second Chern class of the bundle $\mu^{-1}(0) \to M_{\text{red}}.$ In particular it has polynomial dependence on $\epsilon.$ \end{theorem} \begin{pf} The contribution to the basic integral of $\mu^{-1}(0)$ is \[\lim_{t\to \infty}\frac{1}{K} \int_N \int_\lieg \alpha(\phi)\exp(\omega + i \mu \cdot \phi - \frac{\epsilon}{2}|\phi|^2 + t D\lambda) \dint \phi\] where $N$ is a neighborhood of $\mu^{-1}(0)$ containing no other critical points in its closure. By Proposition \ref{pr:normal_form} we can take $N$ isomorphic to a neighborhood of $\mu^{-1}(0)$ in $\mu^{-1}(0) \times \lieg.$ The integral is unchanged if we replace $\alpha$ by something cohomologous, so using an equivariant homotopy we can replace $\alpha$ with a form that agrees with $\iota^*(\alpha)\times 1$ in a neighborhood of $\mu^{-1}(0)$ in $\mu^{-1}(0) \times \lieg$ ($\iota$ being the inclusion of $\mu^{-1}(0)$). By making $N$ sufficiently small this form agrees with $\iota^*(\alpha)\times 1$ (which we will abbreviate $\iota^*(\alpha)$) everywhere. Thus \[=\lim_{t\to \infty}\frac{1}{K} \int_{N \subset \mu^{-1}(0) \times \lieg} \int_\lieg \iota^*(\alpha)(\phi) \exp(\pi^* \omega_0 + d\bracket{\nu,A} + i \bracket{\nu,\phi} - \frac{\epsilon}{2}|\phi|^2 + t D\lambda) \dint \phi.\] By Lemma \ref{lm:lambda_invariance}, we may isotope $\lambda$ provided the zeros of $\lambda\intprod V\!\phi$ do not change. Since $\bracket{V\!\mu^\star, \,\cdot \,}_M$ and $\bracket{\nu, A}_G$ are both positive on the vector $V\!\mu^\star,$ interpolating between them linearly does not change the zeros. Thus replacing $\lambda$ with $\bracket{\nu,A}$ does not change the limit, giving \begin{eqnarray*} &=& \lim_{t\to \infty}\frac{1}{K} \int_{N \subset \mu^{-1}(0) \times \lieg} \int_\lieg \iota^*(\alpha)(\phi) \\ && \qquad \cdot \exp(\pi^* \omega_0 + d\bracket{\nu,A} + td\bracket{\nu,A} + i \bracket{\nu,\phi} + it \bracket{\nu,\phi} - \frac{\epsilon}{2}|\phi|^2) \dint \phi. \end{eqnarray*} Notice that $\nu$ occurs throughout with the factor $(1+t)$ (because $\iota^*(\alpha)$ does not depend $\nu$) and thus we can rescale to eliminate $t$ except for the dependence of the region of integration. In the large $t$ limit this becomes the integral over all $\lieg$ \begin{eqnarray*} &=& \frac{1}{K} \int_{\mu^{-1}(0) \times \lieg} \int_\lieg \iota^*(\alpha(\phi)) \\ &&\qquad \cdot \exp(\pi^* \omega_0 + d\bracket{\nu,A} + i \bracket{\nu,\phi} - \frac{\epsilon}{2}|\phi|^2) \dint \phi. \end{eqnarray*} Completing the square \begin{eqnarray*} &=&\frac{1}{K} \int_{\mu^{-1}(0) \times \lieg} \left(\int_\lieg \iota^*(\alpha)(\phi+\frac{i}{\epsilon} \nu) \exp(-\frac{\epsilon}{2}|\phi|^2) \dint \phi\right)\\ &&\qquad \cdot \exp(\pi^* \omega_0 + d\bracket{\nu,A} - \frac{1}{2\epsilon}|\nu|^2)\\ &=& \frac{1}{K} \int_{\mu^{-1}(0) \times \lieg} \left(\int_\lieg \iota^*(\alpha)(\phi+ \frac{i}{\epsilon} \nu) \exp(-\frac{\epsilon}{2}|\phi|^2) \dint \phi\right)\\ &&\qquad \cdot \exp(\pi^* \omega_0 + \bracket{d\nu, A} + \bracket{\nu,dA} - \frac{1}{2\epsilon}|\nu|^2). \end{eqnarray*} Note the only occurrence of $d\nu$ is in $\exp(\bracket{d\nu,A}).$ Consider a basis of tangent vector at some point in $\mu^{-1}(0) \times \lieg$ which consists of an orthonormal basis of $\lieg,$ the image of this orthonormal basis under $V,$ and a basis of $A$-horizontal vectors in $\mu^{-1}(0).$ The top dimensional piece of this multiform is a sum of terms with $\bracket{d\nu,A}$ raised to various powers, but the only terms which are nonzero when applied to this basis are those where $\bracket{d\nu,A}$ is raised to $\dim(G),$ and on those terms the value is unchanged if $P_A^*$ is applied to all other forms in the product. Thus \begin{eqnarray*} &=&\frac{1}{K} \int_{\mu^{-1}(0) \times \lieg} \left(\int_\lieg P_A^*\circ \iota^*(\alpha)(\phi+ \frac{i}{\epsilon} \nu) \exp(-\frac{\epsilon}{2}|\phi|^2) \dint \phi\right)\\ &&\qquad \cdot \exp(\pi^* \omega_0 + \bracket{\nu,P_A^*(dA)} - \frac{1}{2\epsilon}|\nu|^2)\bracket{d\nu,A}^{\dim(G)}\\ &=& \frac{1}{K} \int_{\mu^{-1}(0) \times \lieg} \left(\int_\lieg P_A^*\circ \iota^*(\alpha)(\phi+ \frac{i}{\epsilon} \nu) \exp(-\frac{\epsilon}{2}|\phi|^2) \dint \phi\right)\\ &&\qquad \cdot \exp(\pi^* \omega_0 + \bracket{\nu,F_A} - \frac{1}{2\epsilon}|\nu|^2)\bracket{d\nu,A}^{\dim(G)}\\ &=& \frac{1}{K} \int_{\mu^{-1}(0) \times \lieg} \left(\int_\lieg P_A^*\circ \iota^*(\alpha)(\phi+ \frac{i}{\epsilon} \nu +i F_A) \exp(-\frac{\epsilon}{2}|\phi|^2- \frac{1}{2\epsilon}|\nu|^2) \dint \phi\right)\\ &&\qquad \cdot \exp(\pi^* \omega_0 + \frac{\epsilon}{2} |F_A|^2 ) \bracket{d\nu,A}^{\dim(G)}\\ &=&\frac{\vol(G)}{K} \int_{\mu^{-1}(0)/G} \left(\int_\lieg \int_\lieg P_A^*\circ \iota^*(\alpha)(\phi+ \frac{i}{\epsilon} \nu +i F_A) \exp(-\frac{\epsilon}{2}|\phi|^2 - \frac{1}{2\epsilon}|\nu|^2) \dint \phi \dint \nu\right)\\ &&\qquad \cdot \exp( \omega_0 + \frac{\epsilon}{2} |F_A|^2 ) \end{eqnarray*} where we have completed the square on $\nu$ and integrated the result over the vertical fibers, noting that the integral is constant in these directions and that the measure $ \bracket{d\nu,A}^{\dim(G)}$ is equal to Haar measure on the vertical fiber times Lebesgue measure $\dint \nu$ on $\nu .$ Now changing the $\nu$ and $\phi$ variables to a single complex variable $z=\sqrt{\epsilon} \phi + i \nu/\sqrt{\epsilon}$ and noting that the integral of any complex polynomial against a complex Gaussian measure gives its constant term yields \begin{eqnarray*} &=&\frac{\vol(G)}{K} \int_{\mu^{-1}(0)/G} \left(\int_{\lieg + i \lieg} P_A^*\circ \iota^*(\alpha)(\frac{1}{\sqrt{\epsilon}}z +i F_A) \exp(-|z|^2) \dint z\right)\exp( \omega_0 + \frac{\epsilon}{2} |F_A|^2 ) \\ &=&\frac{\vol(G)(2 \pi)^{\dim(G)})}{K} \int_{\mu^{-1}(0)/G} P_A^*\circ \iota^*(\alpha)(i F_A) \exp( \omega_0 + \frac{\epsilon}{2} |F_A|^2 ) \\ &=&\frac{\vol(G)(2 \pi)^{\dim(G)})}{K} \int_{\mu^{-1}(0)/G} \Kirwan(\alpha) \exp( \omega_0 + \frac{\epsilon}{2} |F_A|^2 ) \end{eqnarray*} \end{pf} \begin{corollary}\label{cr:reduction_of_integral} If $\alpha$ is a closed equivariant form, $r$ is a regular value of $|\mu|^2,$ $\mu$ is proper and $0$ is a regular value of $\mu$ then the Basic Integral $\BI(\alpha, r, 0)$ can be written uniquely as a sum of a polynomial in $\epsilon$ plus a term bounded by $\exp(-c/\epsilon)$ for some $c>0,$ the polynomial piece representing the contribution from $\mu^{-1}(0)$ as in Theorem \ref{th:zero_contribution}. \end{corollary} \begin{pf} We know that $\BI(\alpha, r, 0)$ differs from the large $t$ limit $\lim_{t\to \infty} \BI(\alpha, r, t)$ by a quantity bounded by $\exp(-c/\epsilon)$ for some $c>0$ by Lemma \ref{lm:large_t}. On the other hand the large $t$ limit is a sum of contributions from $r=0$ and higher critical values by Corollary \ref{cr:contribution}. The former is a polynomial in $\epsilon$ by Theorem \ref{th:zero_contribution}, the latter is bounded by $\exp(-c/\epsilon)$ for some $c>0$ by Proposition \ref{pr:bound_higher}. Since a function can only be written in one way as a polynomial plus a term bounded by $\exp(-c/\epsilon),$ the result follows. \end{pf} In general we have no reason to believe that the integral over all of $M,$ that is the large $r$ limit of $\BI(\alpha, r, t)$ exists for a fixed $t$ or the large $t$ limit. however, if it exists and converges sufficiently rapidly, the same results as above apply. For example \begin{proposition} Suppose that $M$ is a Hamiltonian space with proper moment map and $0$ is a regular value of $\mu.$ Suppose also the symplectic volume of $M_r$ as a function of $r$ is such that $|\partial \Vol(M_r)/\partial r | < \exp(c \sqrt{r})$ for some $c>0.$ Suppose also that for some almost complex structure the supremum over all of $M_r$ of the norm of $\alpha$ (the norm as an ordinary form at each point times the norm as a symmetric tensor in $\lieg^*$) is also bounded by $\exp(c \sqrt{r}).$ Then \[\lim_{r\to \infty} \int_{M_r} \int_\lieg \alpha \exp(\omega + i \mu \phi -\frac{\epsilon}{2}|\phi|^2)\dint \phi\] exists and is of the form a polynomial in $\epsilon$ plus a term exponentially damped in $\epsilon,$ the polynomial \end{proposition} \begin{pf} Fix a regular value $r_0$ of $|\mu|^2.$ \begin{eqnarray*} &&\left| \lim_{r\to \infty} \BI(\alpha, r, 0)- \BI(\alpha, r_0, 0)\right|\\ &=& \frac{1}{K} \int_{M-M_{r_0}} \int_{\lieg} \alpha(\phi) \exp(\omega + i \mu \phi -\frac{\epsilon}{2}|\phi|^2)\dint \phi\\ &=& \frac{1}{K} \left|\int_{M-M_{r_0}} \exp(\omega - \frac{1}{2\epsilon}|\mu|^2)\int_{\lieg} \alpha(\phi +_ i \mu^\star/\epsilon) \exp( -\frac{\epsilon}{2}|\phi|^2)\dint \phi\right|\\ &\leq& \frac{1}{K} \left|\int_{r_0}^\infty \exp(-\frac{r}{2\epsilon} + c \sqrt{r}) \text{POLY}(\sqrt{r}, \epsilon^{\pm 1/2}) \partial \Vol(M_r)/\partial r \dint r\right|\\ &\leq& \frac{1}{K} \left|\int_{r_0}^\infty \exp(-\frac{r}{2\epsilon} + 2c \sqrt{r}) \text{POLY}(\sqrt{r}, \epsilon^{\pm 1/2}) \dint r\right|\\ &\leq& \exp(-\frac{k}{\epsilon}) \text{POLY}( \epsilon^{\pm 1/2}) \end{eqnarray*} for some positive constant $k.$ Applying this to Corollary \ref{cr:reduction_of_integral} gives the result. \end{pf} \bibliographystyle{alpha} \def$'${$'$}
{ "timestamp": "2005-03-18T18:16:22", "yymm": "0503", "arxiv_id": "math/0503385", "language": "en", "url": "https://arxiv.org/abs/math/0503385" }
\section{Introduction} \label{intro} Pattern dynamics in non-equilibrium systems have been studied over decades, in fluid, chemical reaction-diffusion systems, liquid crystal, and so forth. The extensive studies in the field have elucidated a rich variety of pattern dynamics, together with the advances in theoretical analysis \cite{mikhailov 3} $\sim$ \cite{pearson}. One field in the pattern dynamics that is not so well explored in comparison with the above examples, is a discharge system. When a strong voltage is applied between the electrodes, electric discharge appears through the gas filled in the chamber. The discharge often forms a complex spatiotemporal pattern, as has been studied theoretically and experimentally (\cite{2 layers model 1}$\sim$\cite{Germany group 3}). In such discharge system with a variable resister, there exists some global constraint among each local discharge processes. Viewed as pattern dynamics, this means existence of global coupling among local dynamical processes. On the other hand, global coupling among nonlinear units often shows a non-trivial collective motion, as has been extensively studied as collective motion in globally coupled dynamical systems. Now in the discharge system, there exists interplay between collective motion in globally coupled systems and pattern dynamics by local nonlinear dynamics. As a study of nonlinear system, it is interesting to search for some novel dynamic state, due to this interplay. Indeed, there is a beautiful experiment by Nasuno, suggesting such novel, non-trivial dynamics\cite{Nasuno's experiment}\cite{Nasuno's experiment2}. He set up an experiment consisting of two parallel plates between which electric charge flows. By controlling the voltage and current, he found formation of spots, organization of 'molecular-like' structures of spots, complex motions of them, including the dynamics what he called teleportation. In the present paper we introduce a phenomenological model on 2-dimensional (2D) gas discharge system, as a coupled electric circuit subjecting a global constraint, to describe the salient features observed in Nasuno's experiment, and to make further predictions. As a theoretical model, we adopt a two-layer model that describes well a glow discharge system \cite{2 layers model 1}$\sim$ \cite{gas pressure effect 2}. This two-layer model consists of a non-linear resistive part and a linear resistive part. By extending this two-layer model, we construct a coupled dynamical system consisting of elementary circuit systems, each of which represents the process at the interior between square electrodes. By taking spatial continuum limit, this model is reduced to a two-dimensional reaction diffusion (RD) system with global constraint. Through extensive simulations of the model, we will show several phases of the pattern dynamics, and reproduce some salient features in Nasuno's experiment. On the one hand, our model is an abstract and simplified model, and it may not completely correspond to the discharge system by Nasuno. On the other hand, the model shows several novel interesting patten dynamics as a reaction-diffusion system with global coupling, which itself deserves investigation. The outline of the paper is as follows. In Sec.\ref{Experiment}, the experiment by Nasuno is briefly described. In Sec.\ref{Model}, we present some assumptions in order to reproduce the experiment and then construct a coupled circuit system as a model on the gas discharge experiment. In Sec.\ref{Phenomena}, results of extensive simulations are presented, with classification of distinctive pattern dynamics as in the phase diagram. Formation of spots, and a cluster of spots are numerically found, as well as the dynamics of creation and annihilation of spots, while the dynamical systems mechanism of the process is discussed. In Sec.\ref{S and D}, summary, discussion and future perspectives are presented. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.53]{fig1.jpg} \caption{Schematic figure of 2 layers model for glow discharge. We consider that the glow discharge structure consists of two parts, the non-linear resistive region and the linear one.} \label{glow discharge} \end{center} \end{figure} \section{Experiment} \label{Experiment} Here we briefly describe a remarkable experiment by Nasuno \cite{Nasuno's experiment},\cite{Nasuno's experiment2}. In the experiment, two square electrodes are immersed in low pressure air $P$ and connected to a dc power source $\epsilon$ and a variable resistor $R_{ex}$. The power source $\epsilon$ can supply either a constant voltage of up to 2 kV or a constant current of up to 150 mA via a series resistor $Rex$ 35.2 k$\Omega$. He adopted the parameter region of the gas discharge experiment so that the product of pressure $P$ and distance $d$ between electrodes satisfies $P \times d \sim$0.04 (Torr$\cdot$cm), by using $Air$. The parameter region is located in the vicinity of the minimum of paschen curve, according to Fig.7.3 of chapter7 in \cite{discharge physics text}. It is known that, under the experimental condition, current-voltage characteristics has monotonous curve \cite{discharge physics text}$\sim$\cite{negative slope problems 2}, and that no pattern dynamics is usually observed. He fixed the total current $I$ instead of voltage, which is different from the conventional discharge experiments. However, when the current $I$ is fixed as a constant, $V$ is determined uniquely by a monotonous current-voltage characteristic so that $V$ is also constant. Indeed, according to the experiment, current-voltage characteristic $I-V$ is monotonous, so that $I$ and $V$ remain fixed\cite{Nasuno's experiment2}. However, local current and charge density between electrodes can change in space and time, to produce nontrivial dynamics. \paragraph{} By controlling $I$ and $P$, complex pattern dynamics in glow discharge region was observed. By increasing current $I$, the pattern dynamics of discharge changes as follows. \begin{itemize} \item Just after onset of glow discharge, one isolated spot with a typical size (that is, a humped bell-like light intensity distribution) appears. The spot wanders irregularly on the square plane. \item With the increase of $I$, the number of isolated and wandering spots increases. \item As $I$ is increased further, some spots form a molecule-like localized structure, which is called \em cluster\rm. This cluster also wanders through the plane. \item With the further increase of $I$, the number of spots in the cluster increases leading to various clusters with a variety of configurations of spots. Switching among different clusters occurs intermittently. \item At much higher $I$, the excited domain (the area with high light intensity) forms a string or a closed loop. \item A remarkable phenomenon which he termed 'Teleportation' was reported, where annihilation of a spot and immediate recreation of a different spot at a distant place was observed. For example, when 3 spots existed, one of them suddenly was disappeared, and then a new spot immediately appeared at a distant position, located in the neighbor of a remaining spot, so that the 3 spot-state is recovered \cite{Nasuno's experiment}. \end{itemize} The phase diagram for the above behaviors is displayed in terms of $I$ and $P$ in the paper \cite{Nasuno's experiment}. \section{Model} \label{Model} \subsection{Assumption} Here we introduce an electric circuit model corresponding to the above experiment. The model is essentially based on a circuit model introduced by Purwins et al.\cite{2 layers model 3}$\sim$\cite{gas pressure effect 2}, derived from an analogy between an electric circuit and a reduced equation from plasma physics. The model is at a macroscopic phenomenological level, where discharge process between 2D electrodes is described as \em charge transfer process by a coupled circuit system \rm. For each element, we adopt the so-called ``2 layer model'' which consists of non-linear resistive part and linear resistive one, as schematically shown in Fig.\ref{glow discharge}. The model has succeeded in explaining several discharge experiments\cite{2 layers model 1}$\sim$\cite{gas pressure effect 2}. Here we revise their model so that it meets with the experimental condition by Nasuno. To set up the coupled circuit model, the following assumptions are made following Purwins et al. \begin{itemize} \item The experimental condition is set at the region of glow discharge. \item We describe the phenomena in terms of current density $i(x,y)$ and space charge density $q(x,y)$ because we focus only on the \em charge transfer process \rm between two electrodes. \item \em Gas pressure effect \rm between electrodes known experimentally is taken into account. According to \cite{gas pressure effect 1}\cite{gas pressure effect 2}, this effect is effectively described as diffusion of current density that flows between electrodes. Based on the experimental results, we assume a monotonous relation between gas pressure and the diffusion of current density. \end{itemize} In considering Nasuno's experiment, we make further assumptions: \begin{itemize} \item \em Global \rm current-voltage characteristic shows a \em monotonous \rm cubic curve in a logarithmic scale \cite{footnote 2}, as will be given later(Fig.\ref{nullcline}(b)). \item The total current that flows into electrodes is kept as \em constant \rm. Therefore, the voltage drop between electrodes also remains constant. \item In electrodes (represented by 2-dimensional planes), local voltage-current density characteristic is assumed to be cubic in a logarithmic scale with a \em negative slope. \rm \end{itemize} \subsection{Specific Model} The first set of assumptions underlie the coupled circuit model by Purwins et al.\cite{2 layers model 3}$\sim$\cite{gas pressure effect 2}, while we arrange it to meet with the latter set of assumptions to adapt to Nasuno's experiment. \paragraph{} The discharge model which we consider here consists of two square electrodes (area $S$) and the interior, which are under ``External circuit'' that consists of the dc power source $\epsilon$ and the variable resistor $R_{ex}$, as shown in Fig.\ref{circuit model}. The total current $I$ flowing into the electrodes is controlled to be constant\cite{Nasuno's experiment}. Thus $V$ is uniquely determined to be a constant. Further, each elementary circuit $k$ is connected with the neighboring ones by linear resistor $R$ and is also contacted with ``External circuit''. As shown in Fig.\ref{circuit model}, this elementary circuit consists of a linear resistor, capacitance, linear coil and a nonlinear resistor given by a specific $i$-$v$ characteristic. The linear resistor $r$, the region U and the intermediate region correspond to linear resistive layer, the non-linear resistive layer, and the interface in 2-layer model, respectively. Each part is characterized by a linear resistor $r$, capacitance C, the constant self-inductance $l$ of the coil, while the voltage drop $v_{k}$ at the non-linear layer follows the $i$-$v$ characteristic given by a log-scaled cubic function of $i_{k}$. The total current $I$ is distributed to each element with a current $J_k$, while its current at B and U is given by $j_{k}$ and $i_{k} + \dot{q_{k}}$ respectively, as shown in Fig.\ref{circuit model}. Now, the main discharge system is described by a set of $N$ ordinary differential equations as $N$ coupled circuits, as shown in Fig \ref{circuit model}. In the model it is assumed that there are some charge flow into the interface (i.e. at "B" in the figure) from the non-linear resistive layer. This flow of charge density is represented \em by \rm $\dot{s_{k}}$. The voltage drop at the interface is represented by $\frac{s_{k}}{C}$ Since the total space charge $Q$ should be constant in time between electrodes, there is a global constraint on $\sum_k s_k$ as will be discussed again in Sec.\ref{Flowing charge density}. Considering the law of the conservation of charge (the first Kirchoff's rule) at the point A, we get \begin{eqnarray} \label{model-1} J_{k}-j_{k}+[-\frac{1}{RC}\{(q_{k}+s_{k})-(q_{k-1}+s_{k-1}) \\ \nonumber +(q_{k}+s_{k})-(q_{k+1}+s_{k+1})\}]=0, \end{eqnarray} while the law of the conservation of charge at the point B gives \begin{equation} \label{model-2} j_{k}-(i_{k}+\frac{dq_{k}}{dt})=\frac{ds_{k}}{dt}. \end{equation} At the region U, the voltage balance (the second Kirchoff's rule) leads to the following equation \begin{equation} \label{model-3} l\frac{di_{k}}{dt}+v_{k}-\frac{q_{k}}{C}=0. \end{equation} In addition to (\ref{model-1})$\sim$(\ref{model-3}), we need to consider the constraint on the constant total current $I$ and constant voltage drop $V$. \begin{gather} \label{model-I} I=\sum_{k=1}^{N}J_{k}=const. \\ \label{model-V} V=rJ_{k}+\frac{q_{k}}{C}+\frac{s_{k}}{C}=const. \end{gather} Here, to solve the above equations, we assume a constraint for the description of $s_{k}$ as a function of $\{i_{m},q_{m}\},(m=1,2,...,N)$. Accordingly we must define $s_{k}$ satisfying (\ref{model-1})$\sim$(\ref{model-V}). After some calculations, with $v_{k}=v(i_{k})$ and $q^{\prime}_{k}=q_{k}+s_{k}$, $s_{k}$ are determined so as to satisfy \begin{gather} \frac{dq^{\prime}_{k}}{dt}=\frac{V}{r}-\frac{q^{\prime}_{k}}{rC}-i_{k}+\frac{1}{RC}(q^{\prime}_{k-1}+q^{\prime}_{k+1}-2q^{\prime}_{k}) \tag*{} \\ \frac{di_{k}}{dt}=\frac{q^{\prime}_{k}}{lC}-\frac{s_{k}}{lC}-\frac{v(i_{k})}{l} \tag*{} \\ I=\sum_{k=1}^{N}i_{k}=const. \tag*{} \\ Q^{\prime}=\sum_{k=1}^{N}q^{\prime}_{k}=Cv(\frac{I}{N})N=const.\tag*{} \\ V=r\frac{I}{N}+\frac{1}{C}\frac{Q^{\prime}}{N}=const.\tag*{} \end{gather} We define $s_{k}$ as \[s_{k}\equiv\frac{i_{k}}{I}(Q^{\prime}-C\sum_{k=1}^{N}v(i_{k})).\] (See appendix for details.) By taking spatial continuum limit, we obtain the following 2-component RD equation, with $D_{q^{\prime}}= \frac{1}{RC}, Q^{\prime}=\int q^{\prime} dS$, as \begin{subequations} \begin{eqnarray} \label{model-4} \frac{dq^{\prime}}{dt}=\frac{V}{r}-\frac{q^{\prime}}{rC}-i+D_{q^{\prime}}\triangle q^{\prime} \\ \label{model-5} \frac{di}{dt}=\frac{q^{\prime}}{lC}-\frac{s}{lC}-\frac{v(i)}{l}+D_{i}\triangle i \\ \label{model-6} s\equiv\frac{i}{I}\{Q^{\prime}-C\int v(i) dS\} \\ \label{model-7} I=\int i dS =const. \hspace{.2in} \\ \label{model-8} Q^{\prime}=\int q^{\prime} dS =Cv(\frac{I}{S})S=const. \\ \label{model-9} V=r\frac{I}{S}+v(\frac{I}{S})=const. \end{eqnarray} \end{subequations} By using dimensionless variables \paragraph{} $\tilde{i}=\frac{i}{i_{c}}$, $\tilde{v}=\frac{v}{v_{c}}$, $\tilde{q}=\frac{q^{\prime}}{Cv_{c}}$, $\tilde{t}=\frac{t}{\frac{l}{r}}$ \hspace{.1in}and redefining $\tilde{i}\rightarrow i$, $\tilde{v}\rightarrow v$, $\tilde{q}\rightarrow q$, $\tilde{t}\rightarrow t$,\hspace{.2in} the RD equation (\ref{model-4})$\sim$(\ref{model-9}) is written as \begin{subequations} \begin{eqnarray} \label{mujigen-1} \tau\frac{dq}{dt}=V-q-ai+\triangle q \\ \label{mujigen-2} \frac{di}{dt}=\frac{1}{a}\{q-s-v(i)\}+D\triangle i \\ \label{mujigen-3} s\equiv\frac{i}{I}\{Q-\int v(i) dS\} \\ \label{mujigen-4} I=\int i dS =const. \hspace{.2in} \\ \label{mujigen-5} Q=\int q dS =v(\frac{I}{S})S=const. \\ \label{mujigen-6} V=a\frac{I}{S}+v(\frac{I}{S})=const., \end{eqnarray} \end{subequations} \paragraph{} where $\tau=\frac{rC}{\frac{l}{r}},a=r\frac{i_{c}}{v_{c}},D^{\prime}=\frac{D_{i}}{D_{q}},D=\frac{D^{\prime}}{\tau},\\ \hspace{.32in}\xi^{2}=rCD_{q}$ (characteristic length), $S=\frac{1}{\xi^{2}}$. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.55]{fig2.jpg} \caption{Schematic figure of our 2-dimensional discharge model. Our model system consists of $N$ coupled circuits. The element $k$ is connected with neighbor ones by linear resistor $R$ and with ``External circuit''. Although displayed in a one-dimensional representation, what we actually simulated in the paper is a two-dimensional case. } \label{circuit model} \end{center} \end{figure} \subsection{Flowing charge density} \label{Flowing charge density} Here we make some remarks on the flowing charge $s_{k}$ at the point B in Fig \ref{circuit model}. The point B corresponds to the interface of the 2-layer model, that is, the position between the ``Negative glow" and Faraday dark space, which exists for the steady glow discharge (Fig \ref{glow discharge}). In this region, electrons flowing from Cathode layer (non-linear resistive layer) combine with positive ions flowing from ``Positive column", or anode (i.e., the linear resistive layer). There, some complex processes occur through diffusion, excitation, ionization, and recombination of molecules. Hence there is a longitudinal flow of charge, and the flowing charge $s_{k}$ at B is distributed. For a homogeneous steady glow discharge, there is no charge in this region. Furthermore, $s_{k}$ at B is zero, as the initial state before the discharge. Hence, if a spatially homogeneous state were stable, the flow term $s_{k}$ would remain to be 0. However, as will be shown, for almost all the parameter regions, such homogeneous state is unstable, where charge density $q$ is distributed inhomogeneously in the non-linear resistive layer. In this case, flowing charge density $s$ is also distributed in the interface. Hence we need to consider this term. After taking a spatial continuum limit, the total flowing charge is given by the spatial integration, \begin{equation} h\equiv \int s dS (= Q-\int v(i) dS ), \end{equation} where $h$ means the total charge flowing from the non-linear resistive layer into the interface, which is a macroscopic dynamical variable characterizing the global charge transfer process between the non-linear layer and the interface. \paragraph{} In Nasuno's experiment \cite{Nasuno's experiment},\cite{Nasuno's experiment2}, due to the short distance between electrodes, the Positive column does not appear. Hence the anode and the anode surface correspond to the linear part and the interface, respectively. Since the anode has very small resistance in the experiment, $s$ is assumed to be distributed over the anode surface. \subsection{Parameter setting} Hereafter we study the behaviors of the present model given by the equations (\ref{mujigen-1})$\sim$(\ref{mujigen-6}), by controlling the parameters $I$ and $D$, which are the total current into the system, and a parameter for the gas pressure effect, respectively. Hence the control parameters $I$ and $D$ correspond to those in the experiment, i.e., the total current and the pressure $P$, respectively. As initial conditions of $i(x,y)$ and $q(x,y)$, we mostly choose ``uniform state $i_{u}, q_{u}$" with small fluctuations $\eta(x,y)$ as a uniform random number over [-0.01,0.01]. Accordingly total current $I$ (= $i_{u}S$), total charge $Q (=\int q dS)$ and voltage $V$ are determined initially (See appendix). The phenomena to be discussed is always observed from these initial conditions, \em i.e. \rm the pattern dynamics is attracted to a global attractor. Unless otherwise mentioned, the other parameters are fixed as \paragraph{} $\tau=20, \xi^{2}=0.004 $ (characteristic length),$ a=400\frac{i_{c}}{v_{c}}=0.0108$,$ S=\frac{1}{\xi^{2}}=250$. \paragraph{} Numerical simulations are carried out by the mesh size $128\times128$. Unless otherwise mentioned, we choose a periodic boundary condition. These parameters are chosen so that a single element satisfies the behavior for a simple discharge (without spatial pattern). Before describing the pattern dynamics of the model, we first show how a single element behaves. Fig \ref{nullcline} is the nullcline of the single element dynamics (Fig.\ref{nullcline}(a)), and global $I-V$ characteristic (Fig.\ref{nullcline}(b)). We mainly focus on the case that the steady state ($i_{u},q_{u}$) is among the critical point (1,1) in Fig.\ref{nullcline}(a) \cite{icvc}. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.7]{fig3.jpg} \caption{Nullcline of the single element dynamics (a), and global $I-V$ characteristic (b). We mainly focus on the case that the steady state ($i_{u},q_{u}$) is among the critical point (1,1), where it is stable if $i_{u} < 1$, whereas destabilized if $i_{u}> 1$ and then the element shows limit cycle oscillation. $I-V$ characteristic is monotonous so that a steady state is realized. Q is also a function of $i_{u}$. $v(i)$ is cubic and log-scaled function of $i$. Note the relation $I=i_{u}S$.} \label{nullcline} \end{center} \end{figure} \section{Phenomena} \label{Phenomena} \subsection{Global phase diagram} According to our interpretation of the present model (\ref{mujigen-1})$\sim$(\ref{mujigen-6}), the discharge occurs at local site if $i(x,y) \geq i_{u}$. Hence we show the pattern dynamics by displaying only the pixels that satisfy $i(x,y)>i_{u}$, an active region with discharge, by gray. Here we study the pattern dynamics by changing $i_{u} (=\frac{I}{S})$ and $D$, by focusing on the parameter region with $i_{u} \sim 1$, where the discharge phenomena occur. As a 2-dimensional phase diagram with regard to $i_{u}$ and $D$, the pattern dynamics we have observed is classified into five phases, as displayed in Fig.\ref{global phase diagram}. We first give a brief description of each phase, and will discuss the characteristic of each phase in detail later. For small $i_{u}(\sim 1.0)$, there are three phases, depending on the value of $D$, that are "distributed spot phase (DS)", "localized spots phase (LS)", and moving spot phase (MS), respectively, as shown in Fig \ref{global phase diagram}. In these phases, spots are formed. Discharge occurs locally, only within these spots. Indeed, when $i(x,y)$ is plotted in space and time, it shows a stepwise increase at the border of each spot, and shows a high plateau within each spot. The pattern dynamics here will be discussed from the configuration of spot patterns and their motion, with which the three phases are classified. For all these three phases, the number of spots increases with $i_u$. For a larger value of $i_{u}$, there appears a periodic pattern in space, which we call Periodic Wave (PW) phase. For a further large value of $i_u$, there is a uniform glow (UG), without spatial inhomogeneity. The behavior of each phase is summarized as follows; \begin{itemize} \item DS phase --- Spots are arranged with some distance on the average, which decreases with the increase of $i_u$. For small $i_u$, a few spots are isolated with some distance, while hexagonal pattern of spots appears as $i_{u}$ is increased. After transient time, these spots are fixed, and do not move. \item LS phase --- Spots form several clusters similar to molecular structures, as reported in the experiment by Nasuno. The clusters show a dynamic, complex behavior with the formation, collapse, and regeneration. \item MS phase --- A spot can move by itself in the space, which shows a soliton-like behavior without collapse. \item PW phase --- There appears a periodic wave pattern in space and time. \item UG phase --- There is a uniform glow, without spatial inhomogeneity. \end{itemize} \begin{figure} \begin{center} \includegraphics[scale=0.45]{fig4.jpg} \caption{Phase diagram as function of $i_{u} (=\frac{I}{S})$ and $D$. Vertical axis is log-scaled. Global property of the model is roughly classified into five phases under $i_{u}> 0.9 $.} \label{global phase diagram} \end{center} \end{figure} \paragraph{Collective variable -- $\int{s}dS$ ($\equiv h$) --} As a global measure characterizing the pattern dynamics, the integration of $s$, $h \equiv \int{s}dS$ is often useful. In the subsequent subsections, the pattern dynamics will be characterized by the motion of $h$, which is a collective variable. Now we discuss the pattern dynamics, by referring to the change of this collective variable $h$. (Note that $h \ll Q$ in the present parameter setting.) \subsection{DS phase} \paragraph{Isolated steady spots} Starting with initial condition $i(x,y)=i_{u}+\eta(x,y)$, competition for current among elements occurs, resulting in instability of the homogeneous state. This leads to the formation of spots with localized high current. These spots are of the same size. Initially, each spot moves very slowly, and is arranged so that each is located with an equal distance, until a regular, stationary spot pattern is formed. \paragraph{Hexagonal pattern} As $i_{u}$ is increased, the number of spots is increased, so that the distance between spots is decreased. As $i_u$ is increased from $1.5$ to $2.0$, a regular hexagonal pattern of spots is formed, as shown in Fig.\ref{alpha-phase}. The formation of this regular lattice of spots is understood through weak repulsion between neighboring spots. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.35]{fig5.jpg} \caption{Steady patterns observed in the DS phase. Spots are of the same size under fixed $D$. Initially, each spot moves very slowly, and is arranged so that it is located with the equal distance. As $i_{u}$ is increased, the number of spots is increased. As the number is increased with the increase of $i_u$, the distance between spots is decreased. Left--Isolated steady spots. $i_{u}=$0.87,0.93,1.0 in the order from the top to bottom with, $D$=0.01. Right--Hexagonal spots pattern. $i_{u}=$1.2,1.5,2.0 in the order from the top to bottom with $D$=0.01. The mesh size here is 256$\times$256 } \label{alpha-phase} \end{center} \end{figure} \paragraph{Behaviors at the boundary between DS and other phases} As $i_{u}$ is increased toward the boundary to the PW phase, the spots start to breathe, and their sizes (and shape) change, while the distance between spots still remains almost equal as in Fig.\ref{intermittency}(a-1). The collective variable $h$ changes intermittently, corresponding to irregular pattern dynamics (Fig.\ref{intermittency}(a-2)). Around the boundary between the DS and LS phases, the characteristic wavelength disappears, and the configuration of spots is irregular, while spots start to show a bursting behavior, as well as splitting, as in (Fig \ref{intermittency}(b-1)). The change of $h$ as well as the pattern dynamics is irregular(Fig \ref{intermittency}(b-2)). \paragraph{} In the experiment by Nasuno, the behavior corresponding to this phase is not observed. We expect that the range of the change of pressure in the experiment may not be sufficient to cover low values needed and to detect this phase. \begin{figure} \begin{center} \includegraphics[scale=0.47]{fig6.jpg} \caption{Behaviors at the boundary between the DS and PW phases(a) and at the boundary between the DS and LS phases (b). The left figures (1) give snapshots of pattern dynamics, while the time series of the integration of flowing charge $h(=\int{s}dS)$ corresponding to them are plotted in (2)(right). In (a), the spots start to breathe, so that their sizes change, while the distance still remains almost equal. In (b), the distance is no longer equal, and an irregular pattern is formed.\hspace{.2in}(a) $i_{u}=2.4,D=0.01$,\hspace{.1in} (b) $i_{u}=2.0,D=0.012$, mesh size 256$\times$256 } \label{intermittency} \end{center} \end{figure} \subsection{LS phase} \label{betaphase} For small $i_{u}$($\sim$ 0.9), a single spot appears. As $i_{u}$ is increased, the number of spots increases, through successive split of the original spot. Spots are not completely stable in time, and some of them collapse after breathing. As replication and extinction of spots are repeated, a cluster of spots is formed. For small $i_u$ , the number of spots changes between 1 and 2, while the range of the number of spots that the system can take increases with $i_u$. Some of the generation processes are displayed in Fig \ref{split}. \paragraph{Phase diagram within the LS phase} By measuring the dependence of the maximal number of spots on $i_u$ and $D$, the phase diagram within the LS phase is depicted as shown in Fig.\ref{beta phase diagram}, where the maximal number of spots is plotted in the $D-i_{u}$ parameter space. Hereafter denote \em the case ``$M_{k}$'' \rm when $k$ spots exist at maximum. On the other hand, if $n$ spots exist at a moment, we call \em the state ``$sp_{n}$''\rm . Indeed, the configuration of the phases in the diagram of Fig.\ref{beta phase diagram} obtained from our model agree rather well with that obtained experimentally by Nasuno\cite{Nasuno's experiment}. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.55]{fig7.jpg} \caption{Phase diagram of the LS phase. The behaviors in the LS phase are classified in terms of the maximal number of spots, and plotted as a phase diagram in the $D-i_{u}$ parameter space. Note that, in this phase diagram, ``$M_{n}$'' shows the case in which the maximal number of spots is $n$.} \label{beta phase diagram} \end{center} \end{figure} \paragraph{Localized cluster of spots; molecule-like structure} In the following, we discuss dynamics of a localized cluster of spots. After spots are generated by splitting, they separate up to some distance. Thus they form a chain-like structure, a localized structure like a molecule. As the distance between spots is increased and a spot is separated from a cluster very slowly, it starts breathing. Following the amplification of this oscillation, the spot collapses. After this collapse, split of some other spot(s) follows, and the number of spots returns to the maximal under the given value of $i_{u}$. In this way, a cluster of spots is sustained. When the maximal number of spots $\geq$ 6, there appears a variety of configurations for the cluster, as shown in Fig \ref{localized structures}. Depending on which spot splits to which direction, there appears a different configuration. Over a long time span, one can observe a variety of spot configurations. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.55]{fig8.jpg} \caption{Snapshots of typical clusters,$sp_{2}\sim sp_{8}$. They are the snapshots of the clusters with the maximal number of spots under the given $i_{u}$ and $D$. $D$=$0.014.\hspace{.03in}sp_{2}i_{u}$=$0.93$,$sp_{3}i_{u}$=$0.934$,$sp_{4}$(a)$i_{u}$=$0.98,\hspace{.02in} \newline sp_{4}$(b)$i_{u}$=$1.01,\hspace{.02in}sp_{5}$(a)$i_{u}$=$1.1$ $sp_{5}$(b)$i_{u}$=$1.06, sp_{6}$(a) \newline $i_{u}$=$1.1,\hspace{.02in}sp_{6}$(b)$i_{u}$=$1.1$,$sp_{6}$(c)$i_{u}$=$1.1,\hspace{.02in}sp_{6}$(d)$i_{u}$=$1.1,\hspace{.02in} \newline sp_{8}$(a)$i_{u}$=$1.15$,$sp_{6}$(b)$i_{u}$=$1.15$.} \label{localized structures} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.7]{fig9.jpg} \caption{Generation processes of the clusters $sp_{3} \sim sp_{8}$ through successive split of the spots. $sp_{3}$ : $sp_{1} \rightarrow sp_{2} \Rightarrow sp_{2}\rightarrow sp_{3} \Rightarrow sp_{3}$ cluster ($i_{u}=0.939,D=0.014$). $sp_{4}$ : $sp_{1} \rightarrow sp_{2} \Rightarrow sp_{2} \rightarrow sp_{4} \Rightarrow sp_{4}$ cluster ($i_{u}=0.98,D=0.014$). $sp_{5}$ : $sp_{2} \rightarrow sp_{4} \Rightarrow sp_{4} \rightarrow sp_{5} \Rightarrow sp_{5}$ cluster ($i_{u}=1.06,D=0.013$). $sp_{6}$ : $sp_{2} \rightarrow sp_{4} \Rightarrow sp_{4} \rightarrow sp_{6} \Rightarrow sp_{6}$ cluster ($i_{u}=1.08,D=0.013$). $sp_{7}$ : $sp_{2} \rightarrow sp_{4} \Rightarrow sp_{4} \rightarrow sp_{7} \Rightarrow sp_{7}$ cluster ($i_{u}=1.1,D=0.013$). $sp_{8}$ : $sp_{2} \rightarrow sp_{4} \Rightarrow sp_{4} \rightarrow sp_{8} \Rightarrow sp_{8}$ cluster ($i_{u}=1.1,D=0.013$).} \label{split} \end{center} \end{figure} In a long time scale, a cluster wanders slowly on a 2-dimensional space, since the configuration after reconstruction of a cluster is slightly different from the original, even if they are similar. After the split of a spot, there remains some asymmetry between the two spots, which causes a wandering motion of a cluster. We show three examples of such wandering cluster, by displaying successive snapshot patterns in Fig. \ref{sp2-snap-series}, where the maximum number of spots is 2 ($M_{2}$) for Fig \ref{sp2-snap-series}(a), 3 ($M_{3}$) for Fig \ref{sp2-snap-series}(b) and 4 ($M_{4}$) for Fig, \ref{sp2-snap-series} (c), respectively. The motion of cluster is not unidirectional, but its direction changes irregularly. This irregular motion arises since the two spots after split are not completely identical. \paragraph{Dynamics of the collective variable $h$ corresponding to the change in the spot number} When the number of spots changes, the transfer of charge is altered drastically. Hence, it is relevant to study the number change in relationship with the collective variable $h$, i.e., the integration of the flowing charge. In Fig \ref{Time series of h for sp2 to sp4}(a),(b), and (c), we show the time series $h$ for $M_{2}$, $M_{3}$ and $M_{4}$ cases. We plot the time series of $h$ (in (a-1),(b-1),(c-1)), as well as the orbit in the phase space, by embedding the time series of $h$ into three dimensional phase space ($h(t),h(t+1),h(t+2)$). For all the cases, $h$ increases with the spot number $n$. When the spot number stays at some value, $h$ remains almost constant. Note that $h$ value is mainly determined by the spot number of the moment $sp_n$, and is almost independent of the maximal number $M_k$ for a given condition. For example, $h$ for $sp_{2}$ under $M_{2}$ almost equals to that of $sp_{2}$ under $M_{4}$. If the system stays at a state with given $sp_{n}$ ($n \leq N$), it is clearly seen as each plateau in the time series of $h$, or a region with residence of an orbit around a fixed point in the phase space. With the breathing of spots, $h$ also starts to oscillate, and in the phase space picture, the orbit spirals out of a fixed point, leading to a switch to a state with a different number of spots. Consider the case M3 or M4. There the decrease in the number of spots occurs as $(sp_{4}\rightarrow ) sp_{3}\rightarrow sp_{2} \rightarrow sp_{1}$, successively. During this decrease, the system stays at each state with an intermediate spot number, for some time interval, while the return process $sp_{1}\rightarrow sp_{2} \rightarrow sp_{3}$ (or $sp_{4}$), is relatively rapid, without staying long at each intermediate state. (The state $sp_1$ is unstable, and the orbit exits immediately). These decrease and increase in the spot number are repeated. We also note that the switching process observed here is true for a state with a higher number of spots, as have been numerically confirmed up to $sp_{8}$. In Fig.\ref{M6-M8}, we display an example for $M_{6}\sim M_{8}$ case. In the three-dimensional representation of the phase space from the time series of $h$, each state of a given spot number seems to be regarded as a saddle which has one-dimensional stable manifold, and two-dimensional unstable manifold, with an unstable focus. The attraction and repulsion of each orbit around a saddle appears whenever the orbit passes through a given $sp_{n}$ state. Each state of a given spot number is approached from a certain direction, and the orbit spirals out from it. This process is analogous to that observed in Shilnikov chaos\cite{Shilnikov chaos}. In the Shilnikov chaos, however, the spiral motion is stable, giving a stable 2-dimensional manifold for a focus, while a third direction gives an unstable manifold. In contrast, in the present case, the former is unstable, and the third direction gives a stable manifold. It is expected that the instability here leads to irregular wandering of spots. However, when $h$ increases, the orbits path through several saddle points with three real eigenvalues in this representation of the phase space, as long as $sp_n$ is less than its maximal spot number (as can be seen in Fig. \ref{Time series of h for sp2 to sp4}). Note that the orbits approaching the saddle and leaving it take different paths, depending on either if $h$ is increased or decreased, as shown in Fig.\ref{saddle}. In other words, the three-dimensional phase-space representation is insufficient, and the high-dimensionality in the original dynamics leads to the dependence on the history of the change in $h$. In this sense, the itinerant motion over different spot numbers will be better described as chaotic itinerancy\cite{GCM,chaotic scenario}, in the sense that several low dimensional ordered states are visited through high-dimensional chaotic motion. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.6]{fig10.jpg} \caption{A sequence of snapshots for $sp_{2},sp_{3},sp_{4}$ state. We see that the positions of $sp_{2},sp_{3},sp_{4}$ shift in time due to the asymmetry for splitting. In a long time scale, the clusters appear to wander. (a) $sp_{2}$:$i_{u}$=$0.93$($M_{2}$case), (b) $sp_{3}$:$i_{u}$=$0.939$($M_{3}$case), (c) $sp_{4}$:$i_{u}$=$0.98$($M_{4}$case), $D$=$0.014$} \label{sp2-snap-series} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.42]{fig11.jpg} \caption{Dynamics of $h$. (1) Its time series and (2) orbits embedded into a 3-dimensional phase. (a) The case $M_{2}$ ($i_{u}$=0.93), (b) The case $M_{3}$ ($i_{u}$=0.939), (c) The case $M_{4}$ ($i_{u}$=0.98). Each plateau value of $h$ corresponds to each state $sp_{1}\sim sp_{4}$ with respect to different $i_{u}$ under fixed $D$. $D=0.014$ } \label{Time series of h for sp2 to sp4} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.4]{fig12.jpg} \caption{Motion around a saddle in high dimensional phase space. The nature of the saddle point $sp_{m}$ is different depending on if $h$ increases or decreases. The approach to and deviation from a saddle changes according to the in(de)crease of $h$. $i_{u}=0.939,D=0.014$ ($M_{3}$ case, m=2) } \label{saddle} \end{center} \end{figure} \paragraph{Transition rules between several clusters} Here we study transition-rules between clusters in some detail. As the increase in the spot number occurs through the splitting of a spot, there exists a certain transition rule from a state with a lower number of spots to that with a higher number. Here we display the paths of generating $sp_{n}(n=2,...,8)$ through the splitting process (see Fig.\ref{keito}). As shown, each cluster with a larger number of spots is systematically generated from a \em specific \rm state with a smaller number. Note that there are several states with different configuration of spots for a given number of them, when it is larger than or equal to 6 \cite{footnote 3}. This transition rule of clusters is enriched with the increase of $i_{u}$. In Fig.\ref{process}, Schematic figure for the change of the transition rule with $i_{u}$ is displayed, corresponding to the pattern in Fig \ref{keito}, where the solid upward and broken downward arrows show multiplication of spots by splitting and their collapse by breathing, respectively, while the thickness of the arrows shows the frequency of such processes observed. The transition rules are summarized as follows. \begin{itemize} \item When $i_{u}$ is small, only the transition $sp_{1}\rightleftharpoons sp_{2}$ occurs. \item As $i_{u}$ is increased, the states $sp_{3}$ and $sp_{4}^{(1)}$ (which is one type of the 4-spot state) appear, with the transitions $sp_{1}\sim sp_{3}$, $sp_{1}\sim sp_{4}^{(1)}$. \item With the further increase of $i_{u}$, another 4-spot cluster with asymmetric configuration appears, denoted by $sp_{4}^{(2)}$. Now there are transitions $sp_{1}\sim sp_{3}$, $sp_{1}\sim sp^{(1)}_{4}$, $sp_{1}\sim sp^{(2)}_{4}$. \item As $i_{u}$ is increased further, the states $sp_{5}\sim sp_{8}$ appear, with a variety of transitions. These transitions are divided into two groups as in Fig.\ref{keito}, that is, those among $sp_{1}\sim sp^{(1)}_{4}$ and among $sp_{1}\sim sp^{(2)}_{4}$. \end{itemize} Note that with the increase in the spot number, both the configurations and transitions are diversified, while $h$ shows complex dynamics corresponding to the diversification (Fig.\ref{M6-M8}). \begin{itemize} \item When the maximum number of spots is larger than or equal to 4, there exist clusters with the same spot number but different configurations. Those states give almost the same value of $h$, and they are degenerated in the representation of $h$. \item When the maximum number of spots is larger than or equal to 6, the collapse process is diversified, and indeed is more diverse than that displayed in Fig \ref{process}. \item The higher the symmetry in the configuration of the clusters is, the higher is the frequency of the appearance of such configuration. Indeed, this frequent appearance of configuration with higher symmetry is also observed in the experiment (\cite{Nasuno's experiment}\cite{Nasuno's experiment2}). \item The transitions between the degenerate states, i.e., different configuration states with the same number of spots, cannot occur directly. Only after the increase or decrease in the spot number, they can mutually change, as in Fig. \ref{process}. \end{itemize} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.42]{fig13.jpg} \caption{Time series of $h$ for $i_{u}$=$1.1,D$=$0.013.$ $h$ shows complex dynamics, corresponding to the diversification of the cluster configurations with the increase of the spot number.} \label{M6-M8} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.4]{fig14.jpg} \caption{Paths to create state with a higher number of spots, $sp_{n}$ (n=2,...,8) by splitting. Note that only the states generated by splitting are displayed, and some other complicated ones, which are generated as a result of extinction of spots by breathing, are not displayed here.} \label{keito} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.38]{fig15.jpg} \caption{Schematic figure for transitions among quasi-stationary states. The solid upward and broken downward arrows show multiplication processes of spots by splitting and extinction process of them by breathing, respectively. The thickness of the arrows indicates the frequency of the processes. The typical cluster configurations in the present figure are shown in Fig \ref{keito}.} \label{process} \end{center} \end{figure} \paragraph{Correspondence with the experiment} \label{Correspondence with the experiment} The pattern dynamics we observed here agree with that observed experimentally \cite{Nasuno's experiment}\cite{Nasuno's experiment2} as is summarized as follows. \begin{itemize} \item The global phase diagram on the spot number with regards to $i_u$ and $D$ (pressure) agree rather well. \item The cluster of spots forms a molecule-like structure, whose structures are identical. These structures are similar to that observed on the experiment. (In the experiment, the molecular state is so far reported up to the spot number 6 mainly). In our case, the number of spots shows intermittent decrease, while by comparing the state that is dominant in time, the agreement is clear. \item When the maximal spot number is larger than or equal to 6, a variety of forms of clusters appear, both in our model and in experiment. \item An interpretation of teleportation in the term of Nasuno: the reported phenomena in which one of the two spots disappears, and right after it, a new spot appears around the remaining spot. Since it is observed as if one of the spots moved to a distant place within a very short time scale, he called this phenomenon as 'teleportation'. According to our result, we can give a possible interpretation to this phenomenon. In the $M_2$ regime, one of the spots disappears intermittently as we have mentioned. Following this disappearance of one spot, the remaining spot divides into two. Note that the time interval between the disappearance of one spot and the division of the remaining spot is very short. Then, within the resolution of experimental measurement, this process is observed as if one spot teleportated to the location close to another spot. \end{itemize} \subsection{Moving spot phase} At this phase a single spot moves in the space. For $0.93\leq i_{u}\leq 1.0$ in Fig \ref{global phase diagram}, a single \em moving spot \rm appears, while multiple moving spots are observed with the increase of $i_{u}$. \paragraph{Soliton-like spot} Here a single spot (abbreviated by \em ``$ms_{1}$'' \rm) moves without split or collapse. The locus of a spot is displayed in Fig \ref{Dependence under periodic condition}, with a periodic boundary condition, where a solid line shows the center of mass of $ms_{1}$. The center of mass is defined by the mean position among sites with $i(x,y)\geq i_{u}$. Of course, if the spot is completely symmetric in shape, it cannot move. The motion of a spot is due to asymmetry in the shape of $ms_{1}$, and indeed the spot is deviated from a circle slightly. The nature of motion changes with the increase of $i_u$. For small $i_u$ value, that is just above the onset of the appearance of $ms_{1}$, the motion is linear with a constant speed. (See Fig \ref{Dependence under periodic condition},$i_{u}=0.925$). With the increase of $i_{u}$, the motion starts to have a curvature, and the locus is bended. This curvature is increased with $i_u$, and the locus shows a circle, as in Fig \ref{Dependence under periodic condition}, $i_{u}=0.934,0.94$. With the increase of curvature, the speed also increases. For this motion of spot, the choice of boundary condition may be more crucial. For example, to compare with an experiment, the periodic boundary condition may not be relevant. Hence, we have also made some simulations with Neumann boundary condition, as shown in Fig \ref{The orbits of center of mass on Neumann boundary condition}. With reflection at the boundary, the motion becomes more complex. \begin{figure} \begin{center} \includegraphics[scale=0.45]{fig16.jpg} \caption{Dependence of the orbits of a single moving spot on the value of $i_{u}$, under periodic boundary condition where a solid line shows the center of mass of the spot. We show the direction of movement of ms1 with an arrow. $D=$0.027, mesh size 256$\times$256.} \label{Dependence under periodic condition} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.5]{fig17.jpg} \caption{The loci of a single moving spot under the Neumann boundary condition. Arrows show the time course of the spot center. Because the spot reflects at the boundary, the motion is complex as shown for $i_{u}=0.94$. On the other hand, it often shows periodic orbits as given for the loci for $i_{u}=0.93,0.935,0.945.$ For $i_{u}=0.95,0.97$, it shows almost circular motion, but the center of the circular orbits shifts gradually. These behaviors depend on the ratio of the radius of curvature of the spot locus to the electrode size $S^{\frac{1}{2}}$. $D=$0.027 } \label{The orbits of center of mass on Neumann boundary condition} \end{center} \end{figure} \paragraph{Multiple moving spots} As $i_{u}$ is increased further, a single moving spot splits, leading to moving multiple spots(Fig \ref{ms2-snap}). Here, the moving spot is first distorted to extend its tail as spiral. When the tail is small, the spot rotates with it, while as it is larger, the spot is divided into two, which moves to the opposite direction (\em generation of $ms_{2}$\rm). After the motion of these two spots, collapse of one spot occurs as in the LS phase, and the system comes back to a single spot state. These processes are repeated to lead to a complex motion of spots. \begin{figure} \begin{center} \includegraphics[scale=0.4]{fig18.jpg} \caption{Successive snapshots of multiple moving spots. We show the direction of movement of the moving spots with arrows. (1)A single moving spot just before splitting. Here the size starts to expand. (2)It splits into two spots, but one of them is too small and vanishes. The other distorted spot survives and soon begins splitting again. (3)Just before splitting of the distorted spot. (4)Just after splitting of the distorted spot, leading to the state $ms_{2}$. Each spot is separated from each other in the opposite direction. (5)The state $ms_{2}$ is maintained, but one of them starts breathing and will vanish before long. $i_{u}=1.05,D=0.03$ } \label{ms2-snap} \end{center} \end{figure} As shown in Fig.\ref{ms2-snap}, the spot shape starts to be distorted. This distortion is much clearer, as $i_u$ is further increased, where the system approaches the boundary with the PW phase. There the spot size is larger, and forms a distorted "banana-like" shape (Fig.\ref{banana-snap}). This "banana-like" spot moves straight without collapse. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.5]{fig19.jpg} \caption{Banana-like spot. A spot is spread in the direction perpendicular to the direction of its motion, and finally the banana-like shape is formed. The banana-like spot moves straight without collapse. $i_{u}=1.2,D=0.03$ } \label{banana-snap} \end{center} \end{figure} \paragraph{Periodic wave phase} In general, as $i_{u}$ is increased, each site tends to synchronize with each other, since the effective (attractive) coupling among oscillator elements is stronger. Accordingly, in this regime of a large current, spots no longer exist, and they are replaced by a pattern periodic both in space and time (Fig \ref{delta-phase}(a)). \paragraph{Traveling wave} One typical pattern observed here is plane wave(Fig \ref{delta-phase}(b)). As shown, an extended string of discharged region propagates with a constant speed here. A string of concentrated electric discharge is formed over the plate, and it travels with a constant speed. \begin{figure} \begin{center} \includegraphics[scale=0.42]{fig20.jpg} \caption{ (a) Periodic pattern. The parameter values are $i_{u}=5.0$, $D=0.015$, while the similar patterns appear over a wide range of parameters. (b) Plane wave. $i_{u}=5.0$, $D=0.03$. This plane wave moves in one direction as shown by the arrow.} \label{delta-phase} \end{center} \end{figure} \paragraph{Stripe pattern} Here we briefly describe pattern observed at a different set of parameter values of $\tau$ and $\xi$ than that adopted so far. For some parameter values of $\tau$ and $\xi$, we have found a steady pattern, which is different from the DS phase. As an example, we discuss briefly the pattern observed for $\tau=35$ and $\xi$=0.008. As $i_u$ is increased, the connected spots form a string and form a stripe pattern as shown in Fig \ref{Stripe pattern}. The length of string is increased with $i_{u}$, and finally a stripe pattern covers the whole space as in Fig.\ref{Stripe pattern}. Through the time evolution, a steady pattern is formed, without temporal change. Such pattern formation is also common to that observed in Benard convection with a large aspect ratio. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.4]{fig21.jpg} \caption{Steady stripe patten. Spots are extended and connected to form a \em string \rm ($i_{u}=2.2,4.2$). As $i_{u}\nearrow$, many strings are formed, and as a result they are connected with the neighboring strings as shown (\em stripe\rm). Plotted for $i_{u}=6.2,8.2$.} \label{Stripe pattern} \end{center} \end{figure} \section{Summary and Discussion} \label{S and D} \subsection{Summary} To sum up we have introduced a coupled circuit model $(\ref{mujigen-1})\sim (\ref{mujigen-6})$, corresponding to the discharge experiment by Nasuno. Mathematically, the model belongs to a class of reaction-diffusion system with global coupling due to the global constraint on the conservation of charge and current. When the total current is large, the system shows a homogeneous glow, while this homogeneous state is destabilized with the decrease in the current, and a variety of pattern dynamics of discharge is observed, as are classified into four phases, depending on the value of total current and the diffusion constant that corresponds to the pressure value in the discharge experiment. For the regime with a lower current, the regions with concentrated discharge form spots. They are classified as \begin{itemize} \item {\bf Distributed Spot phase}: Spots are arranged with some distance, to form a regular array of spots. The phase appears for small $D$. \item {\bf Localized Spot phase}: A few number of local spots exist, which form a cluster. With the increase in the current the spot number increases. In the cluster spots are arranged like a molecule, while there are several configurations of spots when their number is larger than or equal to 4. Also, we have observed intermittent collapse of spots to decrease their number, and immediate recovery by the splitting of spots. These processes are repeated as a cycle, and with this cycle, the cluster of spots wanders throughout the space. \item {\bf Moving Spot phase}: For a larger value of $D$, a shape of a spot is asymmetric, and each spot starts to move by itself. With the increase of the total current the motion changes from linear to circular, while complex motion is observed when the boundary condition is not periodic. \end{itemize} Besides these spot phases, we have observed another phase at a high current region: \begin{itemize} \item {\bf Periodic Wave phase}, where a string of discharged region propagates in space. \end{itemize} These pattern dynamics are characterized by introducing the collective variable $h=\int s dS$, the integration of the flowing charge $s_{k}$, which expresses the global charge transfer between the non-linear resistive region and the interface in the direction of the gap. Indeed, $h$ turns out to characterize the number of spots, and the birth-and-death dynamics of spots are represented by the temporal change of $h$. \subsection{Comparison with the experiment by Nasuno} The behaviors observed in the LS phase reproduce the phenomena observed in experiments. They include: \begin{itemize} \item Increase of spot number with the current. \item Molecule-like structures of spots and their variety in shapes. \item Disappearance and recreation of spots, as is termed 'Teleportation' by Nasuno. \item Wandering motion of spots. \end{itemize} \paragraph{\em loop \rm} In the experiment, at a high current region, a loop structure that moves in space is observed. Although a connected string observed in the PW phase is similar to the loops observed in the experiment, one difference here is that in our case the strings are connected through the whole space. Since we have chosen the periodic boundary condition, the edges of the traveling string in Fig \ref{delta-phase} are connected, to form a loop, but such traveling string observed is extended to the whole space in our model. So far, we are not yet confident if the strings in the PW phase correspond to the loops in the experiment. As the loops are not small in contrast to localized spot structures, the influence of the boundary may be important. A suitable choice of boundary condition should be necessary. Instead of periodic or Neumann boundary condition, the Dirichlet boundary may be more relevant to make more accurate correspondence with the experiment. This problem is left for the future. \paragraph{Related studies} \paragraph{} There are some recent studies concerning Nasuno's experiment on gas discharge. Static pattern of particle structure is discussed by modifying Gray-Scott equation in \cite{kobayashi}, while its relationship with discharge system is not so straightfoward. In \cite{obstructed discharge}, pattern dynamics caused by local current heating of the gas are studied by using the left-branch of Paschen curve, termed as obstructed discharge. In our study, we consider surface charge on the electrode in the obstructed discharge in Subsection \ref{Flowing charge density}, and study instability of the homogeneous steady state at the left-branch of Paschen curve. \subsection{Discussion} \paragraph{Novel phenomena in reaction-diffusion system} Since our model belongs to a class of reaction-diffusion systems, many of the pattern dynamics observed here are common with those studied therein; formation of spots, array of spots, and strings. Still, the molecule structure of spot cluster, cyclic process of collapse and split of spots, wandering motion of spots are rather novel and characteristic to the present model. For these phenomena, inclusion of global coupling into local reaction-diffusion equations is essential. Systems both with local and global couplings have also been studied in surface catalytic reactions \cite{mikhailov 1}, coupled maps\cite{Glocal}, and so forth. Search for a novel class of pattern dynamics by global coupling will be interesting in the future. \paragraph{Itinerancy over clusters with different configurations of spots} Through split and collapse of spots, their number changes, and also transitions between quasi-stationary states with different clusters take place. This transition among quasi-stationary states is found to be governed by a specific rule with regard to the increase or decrease in the spot number. With the aid of the phase-space representation by the macroscopic variable $h$, this process of itinerancy over different quasi-stationary states with different spot configurations is understood as follows: Each quasi-stationary state corresponds to a state with a different configuration or a different number of spots, $sp_n$. This state is also represented by a ``saddle'' point in the high dimensional dynamical systems of the collective coordinates $h$. There is a stable manifold to this saddle connecting from a state with a different number of spots, while the orbit spirals out of the saddle, corresponding to the breathing motion of a spot. Although this low-dimensional dynamical description of quasi-stationary state seems to be rather effective, the transition, indeed, occurs in a high-dimensional dynamical system. Such switching among effectively low-dimensional states within high-dimensional dynamical system is studied as chaotic itinerancy\cite{GCM,chaotic scenario,milnor}. The present model gives a novel example of chaotic itinerancy, whose mechanism has to be elucidated in future. When the number of spots is larger than or equal to 6, a variety of transition processes appear due to drastic increase in possible configurations of spots. As the spot number increases beyond 6, the combinatorial explosion of the cofiguration of spots sets in, so that stable manifolds connecting many saddle points come to be entangled. This leads to a variety of transitions over quasi-stable spot states, resulting in chaotic itinerancy. In \cite{magic7_1,magic7_2} it is shown that combinatorial explosion which appears beyond the degrees of freedom $5\sim 9$ leads to complex dynamics behavior with chaotic itinerancy. Complex dynamics for the spot number beyond 6 may be discussed from this viewpoint. \paragraph{Acknowledge} We would like to thank Masaki Sano, Masashii Tachikawa, Akinori Awazu, Koichi Fujimoto, Shin'ichi Sasa, and Teruhisa Komatsu, for discussions and suggestions. The present paper is dedicated to the memory of the late Dr. Satoru Nasuno.
{ "timestamp": "2009-04-30T18:29:19", "yymm": "0503", "arxiv_id": "nlin/0503034", "language": "en", "url": "https://arxiv.org/abs/nlin/0503034" }
\section{Introduction} \label{sect1} \input{cakeintro} \setcounter{equation}{0} \section{Cake Baking as a Diffusion Process} \label{sect2} \subsection{Cake Baking from a Quantitative Point of View} \label{sect2_1} \input{cake2_1} \subsection{The G\'enoise} \label{sect2_2} \input{cake2_2} \subsection{Theory from a Naive Perspective} \label{sect2_3} \input{cake2_3} \setcounter{equation}{0} \section{Experiment} \label{sect3} \subsection{Procedure} \label{sect3_1} \input{cake3_1} \subsection{Revised Model} \label{sect3_3} \input{cake3_2} \input{cake3_3} \setcounter{equation}{0} \section{Estimating the Baking Time of a Cake} \label{sect4} \input{cake4} \setcounter{equation}{0} \section{An Irrelevant but Intriguing Digression} \label{sect5} \input{cake5} \setcounter{equation}{0} \section{Conclusions} \label{sect6} \input{cakeconcl} \newpage
{ "timestamp": "2005-11-14T17:06:49", "yymm": "0503", "arxiv_id": "physics/0503210", "language": "en", "url": "https://arxiv.org/abs/physics/0503210" }
\section{Introduction} Elastic hadron electromagnetic form factors (FFs) are fundamental quantities for the understanding of nucleon structure. They contain information on the nucleon ground state, and constitute a further severe test for the models of nucleon structure, which already reproduce the static properties of the nucleon, such as masses and magnetic moments. Moreover, the dependence of FFs on the momentum transfer squared, $q^2=-Q^2$, should reflect the transition from the non perturbative regime, where effective degrees of freedom describe the nucleon structure, to the asymptotic region, where QCD applies. The magnetic proton FF, which is the dominant term in the elastic $ep$ cross section, has been measured at $Q^2$ values up to 31 GeV$^2$ in the space-like (SL) region \cite{Ar86}, and from $p \bar p$ or $e^+e^-$ annihilation up to 18 GeV$^2$ in the time-like (TL) region \cite{An03}. Large progress has been recently done in the determination of the electric and magnetic proton form factors, based on the idea, firstly suggested in Ref. \cite{Re68}, to measure the polarization of the recoil proton in $\vec e p$ elastic scattering, when the electron is longitudinally polarized. Experiments, based on this method, have been performed at JLab up to $Q^2=5.6$ GeV$^2$ \cite{Jo00,Ga02}. A similar method, applied to the reaction $d(e,e'n)$p in quasi-elastic kinematics, has allowed the measurement of the neutron electric FF up to $Q^2$=1 GeV$^2$ using a polarized deuteron target \cite{Day} and up to $ Q^2$=1.47 GeV$^2$, measuring the polarization of the outgoing neutron \cite{Madey}. The polarization method has been also successfully applied at low $Q^2$, for a precise determination of the neutron FFs, at Mainz, and shows that $G_{En}$ is definitely different from zero (\cite{Gl04} and refs therein). These results have been obtained thanks to the availability of high intensity, highly polarized electron beams and polarized targets, and to the optimization of hadron polarimeters in the GeV range. An extension of the measurement of the polarization transfer in $\vec e +p\to e+\vec p$ up to 9 GeV$^2$ is in preparation \cite{04108}. More data are expected in future, in SL region, after the upgrade of Jlab, and in TL region, at Frascati and at the future FAIR facility at Darmstadt \cite{GSI}. In the TL region \cite{An03}, due to the poor statistics, the determination of FFs requires to integrate the differential cross section over a wide angular range. One typically assumes that the $G_E$ contribution plays a minor role in the cross section at large $q^2$ and the experimental results are usually given in terms of $|G_M|$, under the hypothesis that $G_E=0$ or $|G_E|=|G_M|$. The first hypothesis is an arbitrary one. The second hypothesis is strictly valid at threshold only, i.e., for $\tau=q^2/(4m^2)=1$, but there is no theoretical argument which justifies its validity at any other momentum transfer, where $q^2\neq 4m^2$ ($m$ is the nucleon mass). The measurement of the differential cross section for the process $p+\overline{p}\to \ell^+ +\ell^-$ at a fixed value of the total energy $s$, and for two different angles $\theta$, allowing the separation of the two FFs, $|G_M|^2$ and $|G_E|^2$, is equivalent to the well known Rosenbluth separation for the elastic $ep$-scattering. However, in TL region, this procedure is simpler, as it requires to change only one kinematical variable, $\cos\theta$, whereas, in SL region it is necessary to change simultaneously two kinematical variables: the energy of the initial electron and the electron scattering angle, fixing the momentum transfer squared, $Q^2$. Due to the limited statistics, the Rosenbluth separation of the $|G_E|^2$ and $|G_M|^2$ contributions has not yet been realized in TL region. Early attempts showed that the large error bars prevent to discriminate between the two hypothesis on $|G_E|$ and $|G_M|$ quoted above \cite{Bi83,Ba94}. The $|G_M|$ values depend, in principle, on the kinematics where the measurement was performed and the angular range of integration. However, it turns out that these two assumptions for $G_E$ lead to comparable values for $|G_M|$. In the SL region the situation is different. The cross section for the elastic scattering of electrons on protons is sufficiently large to allow the measurements of the angular distribution and/or of polarization observables. Data on $G_M$ are available up to the highest measured value, $Q^2\simeq$ 31 GeV$^2$ \cite{Ar86} and this FF is often approximated according to a dipole behavior: \begin{equation} G_M(Q^2)/\mu_p=G_d,~\mbox{with}~ G_d=\left [1+{Q^2}/{ m_d^2 }\right ]^{-2},~m_d^2=0.71~\mbox{GeV}^2, \label{eq:dipole} \end{equation} where $\mu_p$ is the magnetic moment of the proton. It should be noted that the independent determination of both FFs, $G_M$ and $G_E$, from the unpolarized $e^- +p$-cross section, has been done up to $Q^2=$ 8.7 GeV$^2$ \cite{And94}, and the further extraction of $G_M$ assumes $G_E=G_M/\mu_p$. The behavior of $G_{Ep}$, deduced from polarization experiments, in which, more precisely, the ratio $G_{Ep}/G_{Mp}$ is directly related to the longitudinal and transversal component of the scattered proton polarization, differs from $G_M/\mu_p$, with a deviation up to 70\% at $Q^2$=5.6 GeV$^2$ \cite{Ga02}. This is the maximum momentum at which new, precise data are available. The recent experimental data have inspired many new theoretical developments, and shown the necessity of a global representation of FFs in the full region of momentum transfer squared. FFs are analytical functions of $q^2$, being real functions in the SL region (due to the hermiticity of the electromagnetic Hamiltonian) and complex functions in the TL region. The discussion of the constraints and consequences of a description in the full kinematical domain was firstly done in Ref. \cite{Bi93} and more recently in Refs. \cite{ETG01,Br03,Ia03,Wa04}. The extension of the nucleon models developed for the SL region to the TL region is straightforward for VMD inspired models, which may give a good description of all FFs in the whole kinematical region, after a fitting procedure involving a certain number of parameters \cite{Du03,Bij04}. The purpose of this paper is to update and compare some of the available models on the world data set in both TL and SL regions, and to predict time-like polarization observables, in framework of these models. The paper is organized as follows. In section II the expressions for the relevant polarization observables, in the process $\bar p+p\to \ell^+ +\ell^- $, $\ell=e$ or $\mu$, are given as a function of the electromagnetic FFs, in Section III we update some of the fits of nucleon FFs on the available data, and discuss their extension to the TL region. In section IV we give the predictions of the considered models in TL region. \section{Observables in TL region} We develop a simple and transparent formalism for the study of polarization phenomena for $p+\overline{p}\to \ell^+ +\ell^-$, in framework of one-photon mechanism. The calculations of the cross section and of the polarization observables for the process $\bar p+p\to \ell^+ +\ell^- $, $\ell=e$ or $\mu$, in the annihilation channel are more conveniently performed in the center of mass system (CMS), Fig. \ref{fig:cms}. The momenta of the particles are indicated in the figure and $|\vec k_1|=|\vec k_2| =|\vec k|$ and $|\vec p_1|=|\vec p_2| =|\vec p|$. Let us choose the $z$ axis along the direction of the incoming antiproton, the $y$ axis normal to the scattering plane, and the $x$ axis to form a left-handed coordinate system. The components of the unity vectors are therefore $\hat{\vec p}=(0,0,1)$ and $\hat{\vec k}=(\sin\theta,0,\cos\theta)$ with $\hat{\vec p}\cdot\hat{\vec k}=\cos\theta$, where $\theta$ is the electron production angle in CMS. The relevant kinematical variable is the antiproton energy, $E$, which is related to the four momentum transfer, $q^2=s=(k_1+k_2)^2=4E^2$, as, in CMS, $\vec k_1+\vec k_2=0$. In the laboratory (Lab) system, one finds $q^2=2m^2+2mE_L$, where $E_L$ is the Lab antiproton energy. The observables are calculated in the approximation of zero electron mass. The starting point of the analysis of the reaction $p+\overline{p}\to e^+ +e-$ is the standard expression of the matrix element in framework of one-photon exchange mechanism: \begin{equation} {\cal M}=\displaystyle\frac {e^2}{q^2}\overline{u}(-k_2)\gamma_{\mu}u(k_1) \overline{u}(p_2)\left [F_{1N}(q^2)\gamma_{\mu}- \displaystyle\frac{\sigma_{\mu\nu}q_{\nu}}{2m}F_{2N}(q^2)\right] u(-p_1), \label{eq:mat} \end{equation} where $p_1$, $p_2$, $k_1$ and $k_2$ are the four-momenta of initial antiproton and proton and the final electron and positron respectively, $q^2>4m^2$, $q=k_1+k_2=p_1+p_2$. $F_{1N}$ and $F_{2N}$ are the Dirac and Pauli nucleon electromagnetic FFs, which are complex functions of the variable $q^2$ - in the TL region of momentum transfer. \begin{center} \begin{figure}[ht] \mbox{\epsfxsize=8cm\leavevmode\epsffile{cms.eps}} \caption{The kinematics of the process $p+\overline{p}\to e^- + e^+$ in the reaction CMS.} \label{fig:cms} \end{figure} \end{center} In framework of one-photon exchange, the matrix element is written as the product of the leptonic and hadronic currents: \begin{equation} {\cal M}=\displaystyle\frac{e^2}{q^2} \ell_{\mu}{\cal J}_{\mu}= \displaystyle\frac{e^2}{q^2} (\ell_0{\cal J}_0- \vec\ell\cdot\vec{\cal J}) =-\displaystyle\frac{e^2}{q^2} \vec\ell\cdot\vec{\cal J}, \label{eq:eq1} \end{equation} where $\ell_0{\cal J}_0=0$, due to the conservation of the leptonic and hadronic currents\footnote{The conservation of the current implies that $\ell\cdot q=0$, i.e., $\ell_0 q_0-\vec\ell\cdot\vec q =0$, but $\vec q=\vec k_1+\vec k_2=0 $ in CMS. Therefore, $\ell_0 q_0=0$ for any energy $q_0$, i.e., $\ell _0=0.$}. The expression for the leptonic current is: \begin{equation} \vec\ell=\sqrt{q^2}\phi^{\dagger}_2(\vec\sigma-\hat{\vec k}\vec\sigma\cdot\hat{\vec k})\phi_1, \label{eq:eq2} \end{equation} where $\phi_1(\phi_2)$ is the two-component spinor of the electron (positron), $\hat{\vec k}$ is the unit vector along the final electron three-momentum, and for the hadronic current: \begin{equation} \vec{\cal J}=\sqrt{q^2}\chi^{\dagger}_2\left [ G_M(q^2)(\vec\sigma- \hat{\vec p}\vec\sigma\cdot\hat{\vec p})+\displaystyle\frac{1}{\sqrt\tau}G_E(q^2)\hat{\vec p}\vec\sigma\cdot\hat{\vec p} \right ]\chi_1, \label{eq:eq3} \end{equation} where $\chi_1$ and $\chi_2$ are the two-component spinors of the antiproton and the proton, $\hat{\vec p} $ is the unit vector along the three momentum of the antiproton in CMS. \begin{center} \begin{figure} \mbox{\epsfxsize=9.cm\leavevmode\epsffile{borns.ps}} \caption{One-photon mechanism for $p+\overline{p}\to e^- + e^+$ (with notation of four particle four-momenta).} \label{fig:borns} \end{figure} \end{center} From this expression one can see the physical meaning of the particular relation between the nucleon electromagnetic FFs at threshold: $$ G_{E}(q^2)=G_{M}(q^2),~q^2= 4 m^2. $$ The structure $\hat{\vec p}\vec\sigma\cdot\hat{\vec p}$ describes the $\overline{p} +p$ annihilation from $D$-wave, i.e., with angular momentum $\ell$=2. At threshold, where $\tau\to 1$, the finite radius of the strong interaction allows only the S-state, and $G_{M}(q^2)-\displaystyle\frac{1}{\sqrt\tau}G_{E}(q^2)=0$. From Eqs. (\ref{eq:eq1}), (\ref{eq:eq2}), and (\ref{eq:eq3}) one can find the formulas for the unpolarized cross section, the angular asymmetry and all the polarization observables. \subsection{The cross section} To calculate the cross section when all particles are unpolarized, one has to sum over the polarization of the final particles and to average over the polarization of initial particles: $$ \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0=\displaystyle\frac{| \overline{\cal M}|^2}{64\pi^2 q^2} \displaystyle\frac{k}{p},~ k=\displaystyle\frac{\sqrt{(q^2)}}{2},~p=\sqrt{\displaystyle\frac{(q^2)}{4}-m^2}, $$ \begin{equation} | \overline{\cal M}|^2=\displaystyle\frac{1}{4}\displaystyle\frac{e^4}{q^4} \ell_{ab} {\cal J}_{ab},~\ell_{ab}=\ell_a\ell_b^*,~ {\cal J}_{ab}={\cal J}_a{\cal J}_b^*. \label{eq:eq12} \end{equation} Using the expressions (\ref{eq:eq2}) and (\ref{eq:eq3}), the formula for the cross section in CMS is: \begin{equation} \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0={\cal N}\left [(1+\cos^2\theta)|G_M|^2+\displaystyle\frac{1}{\tau}\sin^2\theta|G_E|^2\right ], \label{eq:eq7} \end{equation} where ${\cal N}=\displaystyle\frac{\alpha^2}{4\sqrt{q^2(q^2-4m^2)}}$, $\alpha=e^2/(4\pi)\simeq 1/137 $, is a kinematical factor. This formula was firstly obtained in Ref. \cite{Zi62}. The angular dependence of the cross section, Eq. (\ref{eq:eq7}), results directly from the assumption of one-photon exchange, where the photon has spin 1 and the electromagnetic hadron interaction satisfies the $P-$invariance. Therefore, the measurement of the differential cross section at three angles (or more) would also allow to test the presence of $2\gamma$ exchange \cite{Re03}. The electric and the magnetic FFs are weighted by different angular terms, in the cross section, Eq. (\ref{eq:eq7}). One can define an angular asymmetry, ${\cal R}$, with respect to the differential cross section measured at $\theta=\pi/2$, $\sigma_0$ \cite{ETG01}: \begin{equation} \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0= \sigma_0\left [ 1+{\cal R} \cos^2\theta \right ], \label{eq:asym} \end{equation} where ${\cal R}$ can be expressed as a function of FFs: \begin{equation} {\cal R}=\displaystyle\frac{\tau|G_M|^2-|G_E|^2}{\tau|G_M|^2+|G_E|^2}. \end{equation} This observable should be very sensitive to the different underlying assumptions on FFs, therefore, a precise measurement of this quantity, which does not require polarized particles, would be very interesting. The $q^2$ dependence of the total cross section can be presented as follows: \begin{equation} \sigma(q^2)={\cal N}\displaystyle\frac{8}{3}\pi \left [2|G_M|^2+ \displaystyle\frac{1}{\tau}|G_E|^2\right ]. \label{eq:eq13} \end{equation} Polarization phenomena will be especially important in $p+\overline{p}\to \ell^+ +\ell^-$. The dependence of the cross section on the polarizations $\vec P_1$ and $\vec P_2$ of the colliding antiproton and proton can be written as follows: \begin{eqnarray} \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0(\vec P_1,\vec P_2) &=& \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0 [1+A_y(P_{1y}+ P_{2y})+A_{xx} P_{1x}P_{2x}+A_{yy} P_{1y}P_{2y}+A_{zz} P_{1z}P_{2z}\nonumber \\ && +A_{xz} (P_{1x}P_{2z}+P_{1z}P_{2x})], \label{eq:eq13a} \end{eqnarray} where the coefficients $A_i$ and $A_{ij}$ $(i,j=x,y,z)$, analyzing powers and correlation coefficients, depend on the nucleon FFs. Their explicit form is given in the following sections. The dependence (\ref{eq:eq13a}) results from the P-invariance of hadron electrodynamics. \subsection{Single spin polarization observables} In case of polarized antiproton beam with polarization $\vec P_1$, the contribution to the cross section can be calculated as: \begin{equation} \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0 \vec A_1=-\ell_{ab}\displaystyle\frac{1}{4} Tr {\cal J}_a\vec\sigma {\cal J}_b^*. \label{eq:eq14} \end{equation} Here the terms related to $|G_E|^2$ and $|G_M|^2$ vanish. For the interference terms, the only non zero analyzing power is related to $P_y$: \begin{equation} \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0 A_{1,y}=\displaystyle\frac{\cal N}{\sqrt{\tau}}\sin2\theta Im(G_MG_E^*). \label{eq:eq15} \end{equation} When the target is polarized, one writes: $$ \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0\vec A_2=\ell_{ab}\displaystyle\frac{1}{4} Tr {\cal J}_a{\cal J}_b^* \vec\sigma. $$ Again the terms related to $|G_E|^2$ and $|G_M|^2$ vanish. Moreover, one can find $\vec A_2=\vec A_1=\vec A$. Eq. (\ref{eq:eq15}) has been proved also in Ref. \cite{Zi62}. One can see that this analyzing power, being T-odd, does not vanish in $p+\overline{p}\to \ell^+ +\ell^-$, even in one-photon approximation, due to the fact FFs are complex in time-like region. This is a principal difference with elastic $ep$ scattering. Let us note also that the assumption $G_E=G_M$ implies $A_y=0$, independently from any model taken for the calculation of FFs. \subsection{Double spin polarization observables} The contribution to the cross section, when both colliding particles are polarized is calculated through the following expression: $$ \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0 A_{ab}=-\displaystyle\frac{1}{4} \ell_{mn} Tr {\cal J}_m\sigma_a{\cal J}_n^{\dagger}\sigma_b, $$ where $a$ and $b=x,y,z$ refer to the $a(b)$ component of the projectile (target) polarization. Among the nine possible terms, $A_{xy}=A_{yx}=A_{zy}=A_{yz}=0$, and the nonzero components are: \begin{eqnarray} \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0A_{xx}&=& \sin^2\theta\left (|G_M|^2 +\displaystyle\frac{1}{\tau}|G_E|^2\right ){\cal N},\nonumber \\ \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0A_{yy}&=& -\sin^2\theta\left (|G_M|^2 -\displaystyle\frac{1}{\tau}|G_E|^2\right ){\cal N},\nonumber\\ \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0A_{zz}&=& \left [(1+\cos^2\theta)|G_M|^2- \displaystyle\frac{1}{\tau}\sin^2\theta |G_E|^2\right ]{\cal N},\nonumber\\ \left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0A_{xz}&=&\left (\displaystyle\frac{d\sigma}{d\Omega}\right )_0A_{zx}= \displaystyle\frac{1}{\sqrt{\tau}}\sin 2\theta Re G_E G_M^* {\cal N}. \label{eq:pol} \end{eqnarray} One can see that the double spin observables depend on the moduli squared of FFs, besides $A_{xz}$. Therefore, in order to determine the relative phase of FFs, in TL region, the interesting observables are $A_y$, and $A_{xz}$ which contain, respectively, the imaginary and the real part of the product $G_EG_M^*$. \section{Results and discussion} \subsection{The data} The nucleon FFs world data were collected and listed in Table I for proton FFs and Table II for neutron FFs in SL region. In Fig. \ref{fig:fig1} the nucleon FFs world data in SL region are shown: the ratio $\mu_p G_{Ep}/G_{Mp} $ (Fig. \ref{fig:fig1}a) , the magnetic proton FF normalized to the dipole FF and divided by $\mu_p$ (Fig. \ref{fig:fig1}b), the electric neutron FF (Fig. \ref{fig:fig1}c), and the magnetic neutron FF normalized to the dipole FF and divided by $\mu_n$ (Fig. \ref{fig:fig1}d). For the electric proton FF, the discrepancy among the data measured with the Rosenbluth methods (stars) and the polarization method (solid squares) appears clearly in Fig. \ref{fig:fig1}a. This problem has widely been discussed in the literature, (for a recent discussion see, for instance, Ref. \cite{Pu05}) and rises fundamental issues. If the trend indicated by polarization measurements is confirmed at higher $Q^2$ \cite{04108}, not only the electric and magnetic charge distribution in the nucleus are different and deviate, classically, from an exponential charge distribution, but also the electric FF has a zero and becomes eventually negative. This scenario will change our view on the nucleon structure and will favor VMD inspired models like \cite{Du03,Bij04}, which can reproduce such behavior. We included data issued from both kind of measurements in the fit, although if a consensus seems to appear that FFs extracted from polarization measurements are more reliable, as less affected by all kinds of radiative corrections. Our purpose here is not to get the best $\chi^2$, but to get a global description of the overall data. The precision and the number of points is very different for the different FFs, therefore one can obtain a good $\chi^2$ for a model that reproduces well, for example, the electric and magnetic FFs in the SL region and fails in giving the trend of $G_{En}$ in TL region. We included in the fit the data on proton magnetic FFs which were published after 1973, and we did not include the data on the neutron electric FFs from Refs. \cite{Ha73,St66,Br95,Hu65,Ak64} as data of much better precision were, later, available, in the same $Q^2$ range. The data in the TL region are drawn in In Fig. \ref{fig:fig2}a, b for the proton and in Fig. \ref{fig:fig2}c, d for the neutron, respectively and summarized in Table III. As no separation has been done for electric and magnetic FFs, the data are extracted under the hypothesis that $|G_{EN}|=|G_{MN}|$. Concerning the neutron, the first and still unique measurement was done at Frascati, by the collaboration FENICE \cite{An98}. \subsection{The models} Among the existing models of nucleon FFs, we consider some parametrizations, which have an analytical expression that can be continued in TL region: predictions of pQCD, in a form generally used as simple fit to experimental data, a model based on vector meson dominance (VMD) \cite{Ia73}, and a third model based on an extension of VMD, with additional terms in order to satisfy the asymptotic predictions of QCD \cite{Lomon}, in the form called GKex(02L). We also considered the Hohler parametrization \cite{Ho76} and the Bosted empirical fit \cite{Bo95}. In order to help the reader, we report in the Appendix the explicit forms of the parametrizations previously published, with the parameters corresponding to the present fit, compared to the published ones. The pQCD prediction, based on counting rules, follows the dipole behavior (\ref{eq:dipole}) in SL region, and can be extended in TL region as \cite{Le80}: \begin{equation} |G_M|=\frac{A(N)}{q^4\ln^2(q^2/\Lambda^2)}, \label{eq:eqtp} \end{equation} where $\Lambda=0.3$ GeV is the QCD scale parameter and $A$ is a free parameter. This simple parametrization is taken to be the same for proton and neutron. The best fit ( Fig. \ref{fig:fig2}, dashed line) is obtained with a parameter $A(p)$= 56.3 GeV$^4$ for proton and $A(n)$= 77.15 GeV$^4$ for neutron, which reflects the fact that in TL region, neutron FFs are larger than for proton. One should note that errors are also larger in TL region. A possible explanation of the fact that FFs are systematically larger in TL region than in SL region (which is true also in the proton case) is the presence of a resonance in the $N\overline{N}$ system, just below the $N\overline{N}$ threshold \cite{Ga96}. More pQCD inspired parametrizations exist for the form factor ratio $F_2/F_1$, which include logarithmic corrections, and have been recently discussed in Ref. \cite{Br03}. However, some of these analytical forms have problems related to the asymptotic behavior. This will be discussed in a future paper. The analytical continuation to TL region of the other models is based on the following relations: \begin{equation} Q^2=-q^2=q^2e^{-i\pi}~\Longrightarrow~\left\{\begin{array}{c} \ln(Q^2)=ln(q^2)-i\pi\\ \sqrt{Q^2}=e^{\frac{-i\pi}{2}}\sqrt{q^2}\\ \end{array} \right. \end{equation} Most of the models predict a different behavior for the electric and the magnetic FFs in TL region, whereas, as already mentioned, no individual determination of electric and magnetic FFs has been done yet. We chose to fit the data assuming that they correspond to the magnetic FFs for proton and neutron, Fig. \ref{fig:fig2}a and \ref{fig:fig2}c, respectively. Therefore, the curves for the electric FFs, in Figs. \ref{fig:fig2}b and \ref{fig:fig2}d have to be considered predictions from the models. Including or not the data on neutron FFs, in TL region, influence very little the fitting procedure. The parametrization from Ref. \cite{Ia73} is shown as a dotted line, in Figs. \ref{fig:fig1} and \ref{fig:fig2}. This model is based on a view of the nucleon as composed by an inner core with a small radius (described by a dipole term) surrounded by a meson cloud. While it reproduces very well the proton data in SL region (and particularly the polarization measurements), it fails in reproducing the large $Q^2$ behaviour of the magnetic neutron FF in SL region. The present fit constrained on the TL data and on the recent SL data does not improve the situation. In framework of this model a good global fit in SL region has been obtained with a modification including a phase in the common dipole term. However, the TL region is less well reproduced \cite{Bij04}. Therefore, the curves drawn in all the figures correspond to the original parameters, which give, in our opinion, a better representation of the whole set of data. The result from an update fit based on the parametrization GKex(02L) \cite{Lomon} is shown in Figs. \ref{fig:fig1} and \ref{fig:fig2} (solid line). It is possible to find a good overall parametrization, with parameters not far from those found in the original paper for the SL region only. The agreement is very good, for both proton and neutron FFs. The Hohler parametrization \cite{Ho76}, contains also pole terms with adjustable parameters. The $\rho$-exchange contribution, however, is fully determined, with constants fixed on $\pi N$ data. The model contains 17 parameters, already, so we did not try to readjust or refit the $\rho$-contribution. As noted in the original paper, such model is not suited to the extrapolation to TL region, because poles appear in the physical region. Constraining the parameters, in order to avoid these instabilities, worsens the description in the SL region. Therefore, we give only a fit on all FFs, in SL region, Fig. \ref{fig:fig1} (dash-dotted line), corresponding to $\chi^2/ndf\simeq 1.7$. The formulas as well as the original and updated parameters are also given in Appendix. Parametrization \cite{Du03} can be considered a successful generalization, in TL region, based on unitarity and analyticity. It requires the modelization of ten resonances, five isoscalar and five isovector. The Bosted parametrization \cite{Bo95} is an empirical fit to nucleon FFs, in the SL region, based on simple formulas which are useful for fast estimations. It does not seem possible to find a unique function, which describes satisfactorily both the magnetic nucleon FFs and the electric proton FF, so the parameters are specific to each FFs. In the extension to TL region, as for the Hohler parametrization, one can not avoid poles and instabilities, and attempts to obtain a description in SL and TL regions remained unsuccessful. Therefore, we give the fit for the SL region, only, as dashed line in Fig. \ref{fig:fig1}, and report in the Appendix the useful formulas and the updated parameters. As one can see from the table, they do not differ more than 20\% from the published ones and the fit corresponds to $\chi^2/ndf\simeq 2$. \section{Predictions in TL region} We give the predictions for the cross section asymmetry and the polarization observables, for those models, described above, which give a good overall description of the available FFs data in SL and TL regions. The calculation is based on Eqs. (\ref{eq:asym}), (\ref{eq:eq15}) and (\ref{eq:pol}), for a fixed value of the angle $\theta=\pi/4$. As shown in Fig. \ref{fig:fig3}, all these observables are, generally, quite large. The model \cite{Ia73} predicts the largest (absolute) value at $q^2\simeq$ 15 GeV $^2$ for all observables, except $A_{xz}$, which has two pronounced extrema. All observables manifest a different behavior, according to the different models. The sign, also, can be opposite for VMD inspired models and pQCD. The model \cite{Lomon} is somehow intermediate between the two representations, as it contains the asymptotic predictions of QCD (at the expenses of a large number of parameters). The fact that single spin observables in annihilation reactions are discriminative towards models, especially at threshold, was already pointed out in Ref. \cite{Dub96}, for the process $e^++e^-\to p+\overline{p}$ on the basis of two versions of a unitary VDM model. The present results, (Fig. \ref{fig:fig3}), for the inverse reaction $p+\overline{p}\to e^++e^-$ confirm this trend and show that experimental data will be extremely useful, particularly in the kinematical region around $q^2\simeq$ 15 GeV $^2$. \section{Summary} The measurement of polarization observables and the possibility to access individual nucleon FFs in TL and SL regions at larger $Q^2$ and/or with higher precision is foreseen in next future. A general analysis of the experimental data on nucleon electromagnetic FFs, extracted from elastic scattering and annihilation reactions, has been performed in the available kinematical region. Expressions of the experimental observables in the reaction $p+\overline{p}\to e^++e^-$ have been derived in terms of the electromagnetic FFs, as a function of the momentum transfer squared. Some of the models on nucleon FFs have been reviewed, extended in TL region and used to give predictions on experimental observables which should be useful to plan future experiments. Many questions are still open. Recent data in the SL region show that the ratio $G_{Ep}/G_{Mp}$ deviates from the expected dipole behavior. In the TL region, the values of $|G_M|$, obtained under the assumption that $G_E=G_M$, are larger than the corresponding SL values. This has been considered as a proof of the non applicability of the Phr\`agmen-Lindel\"of theorem, (up to $s$=18 GeV$^2$, at least) or as an evidence that the asymptotic regime is not reached \cite{Bi93}. The presence of a large relative phase of magnetic and electric proton FFs in the TL region, if experimentally proved at relatively large momentum transfer, can be considered a strong indication that these FFs have a different behavior. In particular, it will allow a test of the Phr\`agmen-Lindel\"of theorem \cite{Bi93}. Large progress in view of a global interpretation of the nucleon FFs is expected from future experiments with antiproton beams: it will be possible, at the future FAIR facility at GSI, to separate the electric and magnetic FFs in a wide region of $s$ and to extend the measurement of FFs up to the largest available energy, corresponding to $s\simeq 30$ GeV$^2$. The angular distribution of the produced leptons will allow the separation of the electric and magnetic FFs. The measurement of the asymmetry ${\cal R}$ (from the angular dependence of the differential cross section for $p+\overline{p}\leftrightarrow \ell^+ +\ell^-$) is sensitive to the relative value of $|G_M|$ and $|G_E|$. In particular, the $\theta$-dependence of the single spin and double spin polarization observables is very sensitive to existing models of the nucleon FFs, which reproduce equally well the data in SL region. Similar information can be obtained from the final polarization in $\ell^++\ell^- \to \vec p+\overline{p}$ \cite{Dub96}, but in this case one has to deal with the problem of hadron polarimetry, in conditions of very small cross sections. Only the study of the processes $p+\overline{p}\to \pi^0+ \ell^+ +\ell^-$ and $p+\overline{p}\to \pi^++\pi^-+\ell^+ +\ell^-$, \cite{Re65,Dub95} will allow to measure proton FFs in the unphysical region (for $s\le 4m^2$, where the vector meson contribution plays an important role) and to determine the relative phase of pion and nucleon FFs. \section{Appendix} The Sachs FFs are expressed in terms of the Pauli and Dirac FFs as: $$ G^N_{E}=F_1^N(Q^2)+\tau F_2^N(Q^2),~G^N_{M}=F_1^N(Q^2)+F_2^N(Q^2).$$ One can introduce the isoscalar and isovector FFs $F_i^{s}$ and $F_i^{v}$, $i=1,2$ as: $2F^p_i=F_i^{s}+F_i^{v}$, $2F^n_i=F_i^{s}-F_i^{v}$. Then, the isoscalar and isovector currents can be parametrized in terms of meson propagators, effective FFs, and/or terms which insure specific properties, according to the different models. \subsection{Model from Iachello, Jackson and Land\'e \protect\cite{Ia73} and Iachello and Wan \protect\cite{Wa04} } FFs are parametrized following the work \cite{Ia73} , with a modification that consists in adding a phase in the dipole term, $g(Q^2)$, for the extension in TL region. \begin{eqnarray*} F_1^s(Q^2)&=& \displaystyle\frac{g(Q^2)}{2} \left[(1-\beta_\omega-\beta_\phi)+\beta_\omega\displaystyle\frac{\mu_\omega^2}{\mu_\omega^2+Q^2}+\beta_\phi \displaystyle\frac{\mu_\phi^2}{\mu_\phi^2+Q^2}\right],\\ F_1^v(Q^2)&=&\displaystyle\frac{g(Q^2)}{2} \left[(1-\beta_\rho)+\beta_\rho \displaystyle\frac{\mu_\rho^2+8\Gamma_\rho\mu_\pi/\pi} {(\mu_\rho^2+Q^2)+(4\mu_\pi^2+Q^2)\Gamma_\rho\alpha(Q^2)/\mu_\pi}\right],\\ F_2^s(Q^2)&=& \displaystyle\frac{g(Q^2)}{2} \left[(\mu_p+\mu_n-1-\alpha_\phi) \displaystyle\frac{\mu_\omega^2} {\mu_\omega^2+Q^2}+\alpha_\phi\displaystyle\frac{\mu_\phi^2}{\mu_\phi^2+Q^2}\right],\\ F_2^v(Q^2)&=&\displaystyle\frac{g(Q^2)}{2} \left[(\mu_p-\mu_n-1) \displaystyle\frac{\mu_\rho^2+8\Gamma_\rho\mu_\pi/\pi}{(\mu_\rho^2+Q^2)+(4\mu_\pi^2+Q^2) \Gamma_\rho\alpha(Q^2)/\mu_\pi}\right], \end{eqnarray*} with $g(Q^2)=\displaystyle\frac{1}{(1+\gamma e^{i\theta}Q^2)^2}$ and $\alpha(Q^2)=\displaystyle\frac{2}{\pi} \sqrt{\displaystyle\frac{Q^2+4\mu_\pi^2}{Q^2}} ln\left[\displaystyle\frac{\sqrt{(Q^2+4\mu_\pi^2)}+\sqrt{Q^2}}{2\mu_\pi}\right]$, with the standard values of the masses $m=0.939$~GeV, $\mu_\rho=0.77$~GeV, $\mu_\omega=0.78$~GeV, $\mu_\phi=1.02$~GeV, $\mu_\pi=0.139$~GeV and the $\rho$ width $\Gamma_\rho=0.112$~GeV. The values of the six parameters are given in Table \ref{table:IJL}. \subsection{Model from Lomon \protect\cite{Lomon}} \begin{eqnarray*} F_1^{v}(Q^2)&=& \displaystyle\frac{N}{2} \left [ \displaystyle\frac{1.0317+0.0875(1+Q^2/0.3176)^{-2}}{(1+Q^2/0.5496)}+ \frac{g_{\rho'}}{f_{\rho'}}\displaystyle\frac{m_{\rho'}^2}{m_{\rho'}^2+Q^2} \right ] F_1^\rho(Q^2)+\\ &&\left(1-1.1192\displaystyle\frac{N}{2}-\displaystyle\frac{g_{\rho'}}{f_{\rho'}}\right)F_1^D(Q^2),\\ F_2^{v}(Q^2)&=&\displaystyle\frac{N}{2} \left [\displaystyle\frac{5.7824+0.3907(1+Q^2/0.1422)^{-1}}{(1+Q^2/0.5362)}+ \kappa_{\rho'}\displaystyle\frac{g_{\rho'}}{f_{\rho'}}\displaystyle\frac{m_{\rho'}^2}{m_{\rho'}^2+Q^2} \right ] F_2^\rho(Q^2)+\\ && \left(\kappa_\nu-6.1731\displaystyle\frac{N}{2}-\kappa_{\rho'}\displaystyle\frac{g_{\rho'}}{f_{\rho'}}\right)F_2^D(Q^2),\\ F_1^{s}(Q^2)&=& \left (\displaystyle\frac{g_\omega}{f_\omega}\displaystyle\frac{m_{\omega}^2}{m_{\omega}^2+Q^2}+ \displaystyle\frac{g_{\omega '}}{f_{\omega '}}\displaystyle\frac{m_{\omega '}^2}{m_{\omega '}^2+Q^2}\right ) F_1^\omega(Q^2)+\\ &&\displaystyle\frac{g_\phi}{f_\phi}\displaystyle\frac{m_{\phi}^2}{m_{\phi}^2+Q^2}F_1^\phi(Q^2)+ \left(1-\displaystyle\frac{g_\omega}{f_\omega} -\displaystyle\frac{g_{\omega '}}{f_{\omega '}} \right)F_1^D(Q^2),\\ F_2^{s}(Q^2)&=& \left (\kappa_\omega\displaystyle\frac{g_\omega}{f_\omega}\displaystyle\frac{m_{\omega}^2}{m_{\omega}^2+Q^2} +\kappa_{\omega '}\displaystyle\frac{g_{\omega '}}{f_{\omega '}} \displaystyle\frac{m_{\omega '}^2}{m_{\omega '}^2+Q^2}\right ) F_2^\omega(Q^2)+\kappa_\phi\displaystyle\frac{g_\phi}{f_\phi}\displaystyle\frac{m_{\phi}^2}{m_{\phi}^2+Q^2}F_2^\phi(Q^2)+\\ && \left(\kappa_s- \kappa_\omega\displaystyle\frac{g_\omega}{f_\omega}- \kappa_{\omega '}\displaystyle\frac{g_{\omega '}}{f_{\omega '}}- \kappa_\phi\displaystyle\frac{g_\phi}{f_\phi}\right)F_2^D(Q^2),\\ \end{eqnarray*} with \begin{eqnarray*} F_1^{\alpha,D}(Q^2)&=&\displaystyle\frac{\Lambda_{1,D}^2}{\Lambda_{1,D}^2+\widetilde Q^2}\displaystyle\frac{\Lambda_{2}^2}{\Lambda_{2}^2+\widetilde Q^2}, ~\alpha=\rho,~ \omega~ and~ \Lambda_{1,D}\equiv \Lambda_1~ \mbox{for} F_i^\alpha, ~\Lambda_{1,D}\equiv\Lambda_D ~for~F_i^D \\ F_2^{\alpha,D}(Q^2)&=&\displaystyle\frac{\Lambda_{1,D}^2}{\Lambda_{1,D}^2+\widetilde Q^2}\left(\displaystyle\frac{\Lambda_{2}^2}{\Lambda_{2}^2+\widetilde Q^2}\right)^2,~ F_1^\phi(Q^2)=F_1^\alpha\left(\displaystyle\frac{Q^2}{\Lambda_1^2+Q^2}\right)^{1.5},~\\ F_2^\phi(Q^2)&=&F_2^\alpha\left(\displaystyle\frac{\Lambda_1^2}{\mu_\phi^2}\displaystyle\frac{Q^2+\mu_\phi^2}{\Lambda_1^2+Q^2}\right)^{1.5} ,~ \widetilde Q^2=Q^2\displaystyle\frac{ln[(\Lambda_D^2+Q^2)/\Lambda_{QCD}^2]}{ln(\Lambda_D^2/\Lambda_{QCD}^2)}. \end{eqnarray*} The set of parameters is reported in Table \ref{table:lomon}. \subsection{Model from Hohler \protect\cite{Ho76}} This model is also based on a VMD parametrization: \begin {eqnarray*} F_1^\rho (Q^2)& = & 0.5 \left [0.955+\displaystyle\frac{0.09}{\left(1+{Q^2}/{0.355} \right)^2} \right] \displaystyle\frac{1}{1+{Q^2}/{0.536}}, \\ F_2^\rho (Q^2)& = & 0.5 \left [5.335+\displaystyle\frac{0.962}{\left(1+{Q^2}/{0.268} \right)^2}\right] \displaystyle\frac{1}{1+{Q^2}/{1.603}}, \\ F_{i}^{(s)}(Q^2) & = & \sum_j\displaystyle\frac{a_j^{(i,s)}}{b_j^{(s)}+Q^2}, \\ F_{i}^{(v)}(Q^2) & = & F_{i}^\rho(Q^2)+ \sum_j\displaystyle\frac{a_j^{(i,v)}}{b_j^{(v)}+Q^2}. \end{eqnarray*} The parameters are given in Table \ref{table:Hohler}. \subsection{Model from Bosted \protect\cite{Bo95}} The analytical expressions are inverse of polynomes as functions of $Q$, whereas $G_{En}$ is described by a different function, as suggested by Galster \cite{Ga71}: \begin{equation} F^j=\displaystyle\frac{1}{1+\sum_i a^j_i Q^{i}}, \end{equation} \begin{equation} G_E^n=\displaystyle\frac{\alpha\mu_n\tau G_D(Q^2)}{1+\beta\tau }, \end{equation} with $a^j_i$, $\alpha$ and $\beta$ free parameters. In the present notation $j=1,2,3$ corresponds to $G_{Ep}$ and $G_{Mn}$ and $G_{Mp}$, respectively. The inverse polynomes are of fourth order for $G_{Ep}$ and $G_{Mn}$ and of fifth order for $G_{Mp}$. The parameters are given in Table \ref{table:Bosted}. \begin{figure}[pht] \begin{center} \includegraphics[width=16cm]{sl.eps} \caption{\label{fig:fig1} Nucleon Form Factors in Space-Like region: (a) proton electric FF, scaled by $\mu_p G_{Mp}$ (b) proton magnetic FF scaled by $\mu_p G_D$ , (c) neutron electric FF, (d) neutron magnetic FF, scaled by $\mu_n G_D$. The predictions of the models are drawn: from Ref. \cite{Ia73} (dotted line), from Ref. \cite{Lomon} (solid line), model from Ref. Ref. \cite{Ho76} (dash-dotted line), from \cite{Bo95} (dashed line). } \end{center} \end{figure} \begin{figure}[pht] \begin{center} \includegraphics[width=16cm]{tl.eps} \caption{\label{fig:fig2} Form Factors in Time-Like region and predictions of the models: pQCD-inspired (dashed line), from Ref. \cite{Ia73} (dotted line), from Ref. \cite{Lomon} (solid line).} \end{center} \end{figure} \begin{figure}[pht] \begin{center} \includegraphics[width=16cm]{obs.eps} \caption{\label{fig:fig3} Angular asymmetry and polarization observables, according to Eqs. (\protect\ref{eq:eq15}) and (\protect\ref{eq:pol}), for a fixed value of $\theta=45^0$. Notations as in Fig. \protect\ref{fig:fig2}.} \end{center} \end{figure}
{ "timestamp": "2005-03-01T15:25:09", "yymm": "0503", "arxiv_id": "nucl-th/0503001", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503001" }
\section{Introduction} \label{S:intro} Let $\mathcal{M}$ be a von Neumann algebra, $\mathcal{P}(\mathcal{M})$ its projections, and $\sim$ the relation of Murray-von Neumann equivalence on $\mathcal{P}(\mathcal{M})$. The description of the quotient $\mbox{$(\p(\M)/\sim)$}$ is known as the \textit{dimension theory} for $\mathcal{M}$. In this paper we prove basic results about three aspects of dimension theory: topology, parameterization, and order. The second section of the paper contains background which is relevant for all three topics. Section 3 deals with topology; Sections 4 and 5 with parameterization; Sections 6 and 7 with order structure. Except for one or two references, these three groupings are independent from each other. In the remainder of this introduction we explain the problems which motivate our investigations. \smallskip \textsc{Topology.} The first goal requires little explanation. \begin{problem} Study the topology that $\mbox{$(\p(\M)/\sim)$}$ inherits from the strong (equivalently, the weak) topology on $\mathcal{M}$. \end{problem} \noindent Some of the results are used in the author's recent work on unitary orbits (\cite{S}). \smallskip \textsc{Parameterization.} It is easy to check that $\mbox{$(\p(\M)/\sim)$}$ also inherits a well-defined partial order from $\mathcal{P}(\mathcal{M})$. Classical work of Murray and von Neumann (\cite{MvN1}) and Dixmier (\cite{Di1949,Di1952}) shows that $\mbox{$(\p(\M)/\sim)$}$ can be naturally parameterized by a subset of the extended positive cone of the center, at least when $\mathcal{M}$ is $\sigma$-finite. This parameterization map, called a \textit{dimension function}, can be extended to all of $\mathcal{M}_+$, and the extension is called an \textit{extended center-valued trace}. The existence of a dimension function on a non-$\sigma$-finite von Neumann algebra is also classical, though less-known. It was originally studied in connection with spatial isomorphisms by Griffin (\cite{G1953,G1955}) and Pallu de la Barri\`{e}re (\cite{P}), and eventually given a representation-free foundation by Tomiyama (\cite{To}). There is a noticeable gap between the last two objects. \begin{problem} Is there a version of the extended center-valued trace which extends the dimension function on a non-$\sigma$-finite von Neumann algebra? \end{problem} \noindent One might expect (and dread) technical constructions involving cardinals and limits. We show how to avoid most of this by simply marrying Tomiyama's dimension function to the equivalence relation of Kadison and Pedersen (\cite{KP}). In fact, the main point to settle does not involve cardinals. \smallskip \textsc{Order.} The range of Tomiyama's map consists of certain cardinal-valued order-continuous functions on the spectrum of the center. Tomi-yama assumed pointwise order and arithmetic on the range, then gave some examples to show that his map lacks basic continuity properties. In fact the pointwise operations (on infinite sets of functions) do not behave well, and it seems to us that these are essentially the wrong operations to be considering. Our viewpoint here is more algebraic. This repairs certain degeneracies and allows us to resolve affirmatively the basic \begin{problem} \label{P:complat} Is $\mbox{$(\p(\M)/\sim)$}$ always a complete lattice? \end{problem} We recall that a \textit{lattice} (resp. \textit{complete lattice}) is a partially-ordered set in which one may take meets and joins of finitely (resp. arbitrarily) many elements. $\mathcal{P}(\mathcal{M})$ is a complete lattice, but it does not induce lattice operations on $\mbox{$(\p(\M)/\sim)$}$: for example, $[p] \wedge [q]$ is not well-defined as $[p \wedge q]$. Nonetheless the comparison theorem for projections readily implies that $\mbox{$(\p(\M)/\sim)$}$ is a lattice. And in a finite von Neumann algebra, the dimension function identifies $\mbox{$(\p(\M)/\sim)$}$ with a complete sublattice of $\mathcal{Z}(\mathcal{M})_1^+$. Problem \ref{P:complat} asks about the existence of meets and joins of arbitrarily large sets of equivalence classes coming from arbitrarily large von Neumann algebras. Its answer has a somewhat surprising reformulation in terms of representations. \section{Background} \label{S:back} Let $\mathcal{M}$ be a von Neumann algebra of arbitrary type and cardinality. We write $\mathcal{Z}(\mathcal{M})$ for its center, and we occasionally symbolize the strong and weak topologies by $s$ and $w$. The central support of an operator is $c(\cdot)$. We use the standard terminology and results from \cite[Section V.1]{T} for projections, including $p^\perp$ for $(1-p)$. Besides $p \sim q$, we write $p \preccurlyeq q$ for subequivalence, and $p \prec q$ for $p \preccurlyeq q$ but not $p \sim q$. Notice that for pairwise orthogonal sets $\{p_\alpha\}, \{q_\alpha\} \subset \mathcal{P}(\mathcal{M})$, \begin{equation} \label{E:addeq} p_\alpha \sim q_\alpha, \: \forall \alpha \Rightarrow \left(\sum p_\alpha\right) \sim \left(\sum q_\alpha\right), \end{equation} \begin{equation} \label{E:addeq2} p_\alpha \preccurlyeq q_\alpha, \: \forall \alpha \Rightarrow \left(\sum p_\alpha\right) \preccurlyeq \left(\sum q_\alpha\right). \end{equation} Among the many adjectives which may be applied to a single projection, we specify one which may cause confusion. A nonzero projection $p$ is \textit{properly infinite} if $zp$ is infinite or zero for any central projection $z$. (An alternative definition: $p$ is properly infinite if it can be decomposed into a countably infinite sum of projections, each of which is equivalent to $p$.) Any adjective can be applied to an algebra when the adjective describes the identity projection of the algebra. According to \eqref{E:addeq}, we can sum unambiguously any set in $\mbox{$(\p(\M)/\sim)$}$ for which there are mutually orthogonal representatives, simply by taking the equivalence class of the sum of representatives. This determines a partial order on $(\mathcal{P}(\mathcal{M})/\sim)$: $[p] \leq [q]$ if there exists a projection $r$ with $[p] + [r] = [q]$. One may also induce the same order directly, since the quotient operation respects the order in $\mathcal{P}(\mathcal{M})$. By this we mean $$[p_1] \leq [p_2] \iff \exists q_1, q_2 \text{ with } q_1 \sim p_1, \: q_2 \sim p_2, \: q_1 \leq q_2.$$ So $[p_1] \leq [p_2]$ means nothing other than $p_1 \preccurlyeq p_2$. Actually the comparison theorem for projections (\cite[Theorem V.1.8]{T}) implies that $(\mathcal{P}(\mathcal{M})/\sim)$ is a lattice. For $p,q \in \mathcal{P}(\mathcal{M})$, let $z$ be a central projection with $zp \preccurlyeq zq$, $z^\perp p \succcurlyeq z^\perp q$. Then \begin{equation} \label{E:lattice} [p] \wedge [q] = [zp + z^\perp q], \qquad [p] \vee [q] = [z^\perp p + zq]. \end{equation} Next we recall basic properties of the extended center-valued trace. This material is due to Dixmier (\cite{Di1949,Di1952}), but for the reader's convenience (presumably), we give citations from Takesaki's book \cite{T}. \begin{definition} (\cite[Definition V.2.33]{T}) Let $\mathcal{M}$ be an arbitrary von Neumann algebra, and let $\Omega(\mathcal{Z}(\M))$ be the spectrum of the abelian $C^*$-algebra $\mathcal{Z}(\mathcal{M})$. By $\widehat{\mathcal{Z}(\mathcal{M})}_+$ we mean the partially-ordered monoid of $[0,+\infty]$-valued continuous functions on $\Omega(\mathcal{Z}(\M))$. $\mathcal{Z}(\mathcal{M})_+$ is contained in $\widehat{\mathcal{Z}(\mathcal{M})}_+$ and acts on it by multiplication. An \textbf{extended center-valued trace} on $\mathcal{M}$ is an additive map $T: \mathcal{M}_+ \to \widehat{\mathcal{Z}(\mathcal{M})}_+$ which commutes with the action of $\mathcal{Z}(\mathcal{M})_+$ and satisfies $T(x^*x) = T(xx^*)$ for $x \in \mathcal{M}_+$. $T$ is \textit{faithful} if $T(x^*x)=0 \Rightarrow x=0, \: \forall x \in \mathcal{M}_+$. $T$ is \textit{normal} if \begin{equation} \label{E:normal} T(\sup x_\alpha) = \sup T(x_\alpha) \end{equation} for any bounded increasing net $\{x_\alpha\} \subset \mathcal{M}_+$. $T$ is \textit{semifinite} if $\{x \in \mathcal{M} \mid T(x^*x) \in \mathcal{Z}(\mathcal{M})_+\}$ is $\sigma$-weakly dense in $\mathcal{M}$. \end{definition} Here we wish to draw attention to a point which will be amplified in Sections \ref{S:cont} and \ref{S:comp}. What is the meaning of the expression $\sup T(x_\alpha)$ in \eqref{E:normal}? The pointwise supremum of an increasing family of $[0,+\infty]$-valued continuous functions on $\Omega(\mathcal{Z}(\M))$ may not be continuous, and some kind of algebraic supremum is required instead. Dixmier showed that such a supremum exists, using the fact that $\Omega(\mathcal{Z}(\M))$ is stonean (\cite{Di1951}). He also mentions specifically that other methods, including a purely formal one, could reach the same goal (\cite[p.25]{Di1952}). We suppose that our technique in Section \ref{S:comp} is similar to the formal approach that he had in mind. Semifinite von Neumann algebras - those with no summand of type III - are characterized by the existence of a faithful normal semifinite extended center-valued trace (\cite[Theorem V.2.34]{T}). Such a map $T$ is unique up to multiplication by an element of $\widehat{\mathcal{Z}(\mathcal{M})}_+$ which takes finite values on an open dense subset of $\Omega(\mathcal{Z}(\M))$, so all are equally useful in calculations. A projection $p$ is finite if and only if $T(p)$ takes finite values on an open dense subset of $\Omega(\mathcal{Z}(\M))$ (\cite[Proposition V.2.35]{T}). From all this $p \preccurlyeq q \Rightarrow T(p) \leq T(q)$, and the converse holds if $p$ is finite. If $\mathcal{M}$ is finite, there is a \textit{unique} faithful extended center-valued trace $T$ with $T(1_\mathcal{M}) = 1_\mathcal{M}$ (\cite[Theorem V.2.6]{T}). Such a map is automatically normal, and the linear extension which is defined on all of $\mathcal{M}$ is called simply a \textit{center-valued trace}. \begin{convention} \label{C} Whenever we talk of an ``extended center-valued trace" $T$ on $\mathcal{M}_+$ in the sequel, it is assumed that \begin{itemize} \item $T$ is normal and faithful; \item on the finite summand of $\mathcal{M}$, $T$ agrees with the center-valued trace; \item on the semifinite summand of $\mathcal{M}$, $T$ is semifinite; \item on the infinite type I summand of $\mathcal{M}$, $T$ maps an abelian projection to its central support. \end{itemize} Therefore $T(p) = (+\infty) c(p)$ for a projection supported on the type III summand. \end{convention} A word about operator topologies on $\mathcal{M}$: the strong, $\sigma$-strong, weak, and $\sigma$-weak topologies can all be defined spatially. The $\sigma$-strong and $\sigma$-weak topologies are independent of the choice of (faithful normal) representation, and this is not true for the strong and the weak. But on \textit{bounded} sets, we have the agreements strong=$\sigma$-strong and weak=$\sigma$-weak; we therefore permit ourselves the small linguistic abuse of referring to the strong (or weak) topology on a bounded subset of $\mathcal{M}$. For $\mathcal{M}$ finite, the normality of the center-valued trace is equivalent to $\sigma$-weak-$\sigma$-weak continuity. It will be more useful for us that this map is also $\sigma$-strong-$\sigma$-strong continuous, and therefore strong-strong continuous on bounded sets. (See \cite[Theorem 13]{G1953}, \cite[I.4.Th\'{e}or\`{e}me 2 and p.250]{Di1969}, or \cite{R} in connection with this. In fact the strong-strong or weak-weak continuity on all of $\mathcal{M}$ does depend on the representation (\cite[Theorem 8]{G1953}).) \bigskip Here are some examples of $\mbox{$(\p(\M)/\sim)$}$. \begin{enumerate} \item When $\mathcal{M}$ is a type $\text{I}_n$ factor, $(\mathcal{P}(\mathcal{M})/\sim)$ is isomorphic to the initial segment of cardinals $\leq n$, via the map that sends a projection to its rank. \item When $\mathcal{M}$ is a type $\text{II}_1$ factor, $(\mathcal{P}(\mathcal{M})/\sim) \simeq [0,1]$. \item When $\mathcal{M}$ is a $\sigma$-finite type $\text{II}_\infty$ factor, $(\mathcal{P}(\mathcal{M})/\sim) \simeq [0, +\infty]$. \item When $\mathcal{M}$ is a $\sigma$-finite type $\text{III}$ factor, $(\mathcal{P}(\mathcal{M})/\sim) \simeq \{0, +\infty\}$. \end{enumerate} The isomorphisms in (2) and (3) are implemented by a (bounded or unbounded) trace. When $\mathcal{M}$ is a non-factor with separable predual, $(\mathcal{P}(\mathcal{M})/\sim)$ is naturally viewed as a direct integral of the lattices above. When $\mathcal{M}$ is finite, $(\mathcal{P}(\mathcal{M})/\sim)$ is isomorphic to a sublattice of $\mathcal{Z}(\mathcal{M})_1^+$ via the center-valued trace (see \cite[Theorem 8.4.4]{KR2}). Continuous (type II) and degenerate (type III) dimension theory were part of the original appeal for Murray and von Neumann: what happens at large cardinality? Since $\mbox{$(\p(\M)/\sim)$}$ is totally ordered if and only if $\mathcal{M}$ is a factor, this is the scenario closest to set theory. Do type II and III factors contain ``quantum cardinal arithmetic" which diverges from the usual cardinal arithmetic of a type I factor? \smallskip The questions above are answered neatly by the parameterization of $\mbox{$(\p(\M)/\sim)$}$ as developed by Griffin (\cite{G1953,G1955}), Pallu de la Barri\`{e}re (\cite{P}), and especially as formulated by Tomiyama (\cite{To}). The main point is a structure theorem allowing us to break a properly infinite von Neumann algebra into direct summands, each of which has a well-defined size. This is in direct analogy to the structure theorem for type I von Neumann algebras, but we use $\sigma$-finiteness instead of abelianness as the ``unit of measurement". \begin{definition} \label{D:homog} $\text{(\cite[Definition 1]{To})}$ Let $\kappa$ be a cardinal. We say that a nonzero projection $p$ in a von Neumann algebra $\mathcal{M}$ is \textbf{$\kappa$-homogeneous} if $p$ is the sum of $\kappa$ mutually equivalent projections, each of which is the sum of centrally orthogonal $\sigma$-finite projections. We also define $$\kappa_\mathcal{M} = \sup \{\kappa \mid \text{$\mathcal{M}$ contains a $\kappa$-homogeneous projection}\}.$$ \end{definition} \begin{remark} The terminology here is conflicting. We follow Tomiyama, but elsewhere ``$\kappa$-homogeneous projection" means a central projection which is the sum of $\kappa$ equivalent abelian projections (e.g. \cite[p.299]{T}). \end{remark} A projection can be $\kappa$-homogeneous for at most one $\kappa \geq \aleph_0$; also for $\kappa \geq \aleph_0$, two $\kappa$-homogeneous projections with identical central support are necessarily equivalent (\cite{G1955,To}). $\kappa_\mathcal{M}$ is not larger than the dimension of a Hilbert space on which $\mathcal{M}$ is faithfully represented. The fundamental result for us is a m\'{e}lange of two theorems of Griffin, one covering the semifinite case (slightly adapted to our setting, and also proved by Pallu de la Barri\`{e}re) and one covering the purely infinite. It was rewritten in the non-spatial setting by Tomiyama. \begin{theorem} \label{T:griffin} $($\cite[Theorem 3]{G1953}, \cite[Theorem 1]{G1955}, also \cite[I.5]{P} and \cite[Theorem 1]{To}$)$ Let $\mathcal{M}$ be a properly infinite von Neumann algebra. Then uniquely $$1_\mathcal{M} = \sum_{\aleph_0 \leq \kappa \leq \kappa_\mathcal{M}} z_\kappa,$$ where each $z_\kappa$ is either zero or a $\kappa$-homogeneous central projection. \end{theorem} Let $T$ be an extended center-valued trace on a von Neumann algebra $\mathcal{M}$ (following Convention \ref{C}). Given any $p \in \mathcal{P}(\mathcal{M})$, let $z^f$ be the largest central projection such that $z^f p$ is finite. By applying Theorem \ref{T:griffin} to $(1-z^f) p\mathcal{M} p$, there are unique central projections $(z_\kappa)_{\aleph_0 \leq \kappa \leq \kappa_\mathcal{M}}$ such that $\sum z_\kappa p = (1-z^f)p$ and any nonzero $z_\kappa p $ is $\kappa$-homogeneous. Make the formal assignment \begin{equation} \label{E:gdf} p = \left( z^f p + \sum_{\aleph_0 \leq \kappa \leq \kappa_\mathcal{M}} z_\kappa p \right) \mapsto \left( T(z^f p) + \sum_{\aleph_0 \leq \kappa \leq \kappa_\mathcal{M}} \kappa z_\kappa \right). \end{equation} From our earlier comments this assignment is a complete invariant for the equivalence class of $p$. Under the isomorphism $\mathcal{Z}(\mathcal{M}) \simeq C(\Omega(\mathcal{Z}(\M)))$, projections correspond to clopen subsets of $\Omega(\mathcal{Z}(\M))$, so elements on the right-hand side of \eqref{E:gdf} can be interpreted as partially-defined functions on $\Omega(\mathcal{Z}(\M))$. The range is in $([0,+\infty) \cup \{\kappa \mid \aleph_0 \leq \kappa \leq \kappa_\mathcal{M}\})$, and the functions are (order) continuous on their domains, which are easily shown to be open and dense. Tomiyama showed (\cite[Lemma 5]{To}) that such functions extend uniquely to continuous functions on all of $\Omega(\mathcal{Z}(\M))$. \begin{definition} \label{T:defgdf} (\cite{To}) The assignment described above, from $\mathcal{P}(\mathcal{M})$ to the continuous $([0,+\infty) \cup \{\kappa \mid \aleph_0 \leq \kappa \leq \kappa_\mathcal{M}\})$-valued functions on $\Omega(\mathcal{Z}(\M))$, is a \textbf{(generalized) dimension function} of $\mathcal{M}$. \end{definition} \begin{theorem} \label{T:gdf} $($\cite{To}$)$ Let $D$ be a dimension function of $\mathcal{M}$. Then $D$ is additive on pairs of orthogonal projections, provided that one incorporates the positive reals into cardinal arithmetic in the obvious way. We have $$p \preccurlyeq q \iff D(p) \leq D(q), \qquad \forall p,q \in \mathcal{P}(\mathcal{M}),$$ where we use the pointwise ordering of functions on the right-hand side. \end{theorem} It follows that $D$ factors as $$\mathcal{P}(\mathcal{M}) \twoheadrightarrow \mbox{$(\p(\M)/\sim)$} \overset{\sim}{\to} D(\mathcal{P}(\mathcal{M})).$$ Here the second map is an embedding in a function space, preserving order, sums (when they exist), and the multiplicative $\mathcal{P}(\mathcal{Z}(\mathcal{M}))$-action. \begin{corollary} \label{T:factor} ${}$ \begin{enumerate} \item In a factor of type $\text{II}_\infty$, the totally ordered set $(\mathcal{P}(\mathcal{M})/\sim)$ is isomorphic to $$[0, +\infty) \cup \{\kappa \mid \aleph_0 \leq \kappa \leq \kappa_\mathcal{M}\}.$$ \item In a factor of type III, the totally ordered set $(\mathcal{P}(\mathcal{M})/\sim)$ is isomorphic to $$\{0\} \cup \{\kappa \mid \aleph_0 \leq \kappa \leq \kappa_\mathcal{M}\}.$$ \end{enumerate} \end{corollary} So any interest in ``quantum cardinal arithmetic" wanes here: infinite quantum cardinals are (isomorphically) just cardinals. For the reader interested in axiomatic treatments of $\mbox{$(\p(\M)/\sim)$}$ and more general algebraic structures obtained as quotients of lattices, see \cite{L,M,F}. \section{The topology of $(\mathcal{P}(\mathcal{M})/\sim)$} \label{S:top} If we want $\mbox{$(\p(\M)/\sim)$}$ to inherit a topology from $\mathcal{P}(\mathcal{M})$, there really are not so many interesting choices. The quotient of the norm topology is the discrete topology, since $\|p - q\| < 1$ implies that $p$ and $q$ are unitarily equivalent (\cite[5.2.6-10]{W-O}). And all of the ``operator" topologies (notably, the strong and the weak) are equivalent when restricted to $\mathcal{P}(\mathcal{M})$ (\cite[Ex. 5.7.4]{KR1}). We point out, however, that $(\mathcal{P}(\mathcal{M}), \text{strong})$ is complete, while $(\mathcal{P}(\mathcal{M}), \text{weak})$ may not be; completeness is not a topological property. We will denote the resulting quotient strong/weak operator topology on $(\mathcal{P}(\mathcal{M})/\sim)$ by ``$QOT$". In the rest of this section, all closures and convergences in $\mbox{$(\p(\M)/\sim)$}$ are to be understood in this topology. We need a few lemmas. \begin{lemma} \label{T:liminf} Let $\{x_\alpha\}$ be a net in a semifinite von Neumann algebra $\mathcal{M}$ equipped with an extended center-valued trace $T$. If $x_\alpha^* x_\alpha = y_1$ is fixed, while $x_\alpha x_\alpha^* \overset{w}{\to} y_2$, then $T(y_1) \geq T(y_2)$ in $\widehat{\mathcal{Z}(\mathcal{M})}_+$. \end{lemma} \begin{proof} Fix any $\varphi \in \mathcal{Z}(\mathcal{M})_*^+$. Then $\varphi \circ T$ is a semifinite normal weight, so weakly lower-semicontinuous (\cite{H}). We have \begin{align*} \varphi \circ T(y_2) &= \varphi \circ T(w-\lim x_\alpha x_\alpha^*) \leq \liminf \varphi \circ T(x_\alpha x_\alpha^*) \\ &= \liminf \varphi \circ T(x_\alpha^* x_\alpha) = \varphi \circ T(y_1). \end{align*} Since $\varphi$ is arbitrary, the conclusion follows. \end{proof} \begin{lemma} \label{T:proj} Let $p,q,r \in \mathcal{P}(\mathcal{M})$, with $\mathcal{M}$ and $p$ properly infinite. \begin{enumerate} \item If $p \sim q_j$ for a countable set $\{q_j\}$, then $p \sim \vee q_j$. \item If $zq \prec zr$ for all nonzero central projections $z$, then $q^\perp \sim 1_\mathcal{M}$. \end{enumerate} \end{lemma} \begin{proof} $\quad$ (1) It is clear that $p \preccurlyeq \vee q_j$. Write $p = \sum p_j$, where each $p_j \sim p$. Let $v_j$ be a partial isometry between $p_j$ and $q_j$. The operator $\sum v_j/2^j$ has right support $\vee q_j$ and left support $\leq p$, so also $p \succcurlyeq \vee q_j$. (2) We compare $q$ and $q^\perp$. If there were a nonzero central projection $z$ with $zq \succcurlyeq zq^\perp$, then $zq$ would be properly infinite (else a nonzero central projection would be the sum of two finite projections). Write $zq = zq_1 + zq_2$, where $zq_1 \sim zq_2 \sim zq$. By \eqref{E:addeq2}, $$zq \preccurlyeq z = (zq + zq^\perp) \preccurlyeq (zq_1 + zq_2) = zq,$$ so that $zq \sim z$. Now $zr \succ zq \sim z$, which is impossible. Thus $q \preccurlyeq q^\perp$. By the same argument, $q^\perp$ is properly infinite and $q^\perp \sim 1_\mathcal{M}$. \end{proof} \begin{theorem} \label{T:sinfin} If $\mathcal{M}$ is a finite von Neumann algebra, the center-valued trace induces a homeomorphism from $(\mbox{$(\p(\M)/\sim)$}, QOT)$ to a subspace of $(\mathcal{Z}(\mathcal{M})_1^+, \text{\textnormal{strong}})$. Consequently $$\overline{\{[p]\}} = \{[p]\}, \qquad p \in \mathcal{P}(\mathcal{M}).$$ \end{theorem} \begin{proof} Let $T$ be the center-valued trace. If $[p_\alpha] \to [p]$, then there exist $q_\alpha \sim p_\alpha$ with $q_\alpha \overset{s}{\to} p$. By the strong-strong continuity of $T$ on bounded sets, we have $T(p_\alpha) = T(q_\alpha) \overset{s}{\to} T(p)$. On the other hand, suppose $p_\alpha, p$ are projections such that $T(p_\alpha) \overset{s}{\to} T(p)$. Let $q_\alpha \leq p$ be projections with $T(q_\alpha) = T(p_\alpha) \wedge T(p)$, where the meet is taken in $\mathcal{Z}(\mathcal{M})_1^+$. Let $r_\alpha \perp q_\alpha$ be projections with $T(r_\alpha) = T(p_\alpha - p) \vee 0$. It follows that $r_\alpha$ is centrally orthogonal to $(p-q_\alpha)$, and by comparing center-valued traces $(q_\alpha + r_\alpha) \sim p_\alpha$. When $\mathcal{M}$ is $\sigma$-finite, the strong topology on bounded sets is generated by the norm $x \mapsto \tau(x^*x)^{1/2}$, for $\tau$ any faithful tracial state (\cite[Proposition III.V.3]{T}). A general finite algebra is a direct sum of $\sigma$-finite ones (\cite[Corollary V.2.9]{T}), so it suffices to show convergence for the seminorms coming from a family of traces, each of which is faithful on a $\sigma$-finite summand. We now take such a trace $\tau$ and compute \begin{align*} \tau([(q_\alpha + r_\alpha) - p]^2) &= \tau([r_\alpha - (p - q_\alpha)]^2) = \tau (r_\alpha + (p-q_\alpha)) \\ &= \tau (T(r_\alpha) + T(p-q_\alpha)) = \tau(|T(r_\alpha) - T(p-q_\alpha)|) \\ &= \tau(|T((q_\alpha + r_\alpha) - p)|) = \tau(|T(p_\alpha - p)|) \\ &\leq \tau(|T(p_\alpha - p)|^2)^{1/2} \to 0. \qedhere \end{align*} \end{proof} Regarding Theorem \ref{T:sinfin}, we remind the reader that typically we do \textit{not} have an equivalence between the strong and weak topologies on $\mathcal{Z}(\mathcal{M})_1^+$. \begin{theorem} \label{T:sininf} Let $p$ be a projection in a properly infinite von Neumann algebra $\mathcal{M}$. If $p$ is finite, \begin{equation} \label{E:finclos} \overline{\{[p]\}} = \{[q] \mid [q] \leq [p]\}. \end{equation} If $p$ is properly infinite and $c(p)=1_\mathcal{M}$, \begin{equation} \label{E:infclos} \overline{\{[p]\}} = (\mathcal{P}(\mathcal{M})/\sim). \end{equation} Equations \eqref{E:finclos} and \eqref{E:infclos} may be synthesized into \begin{equation} \label{E:seg} \overline{\{[p]\}} = \{[q] \mid T(q) \leq T(p)\}, \qquad \forall p \in \mathcal{P}(\mathcal{M}), \end{equation} for any extended center-valued trace $T$. \end{theorem} \begin{proof} First consider a finite projection $p$. We may assume that $c(p) = 1$ and so $\mathcal{M}$ is semifinite; let $T$ be an extended center-valued trace. If $p_\alpha \sim p$ and $p_\alpha \overset{w}{\to} q$, then by Lemma \ref{T:liminf}, $T(q) \leq T(p)$. We have assumed $p$ finite, so $q$ is as well and $p \succcurlyeq q$. For the other containment, choose any $q$ with $[q] \leq [p]$. Write $p = p_0 + p_1$, with $p_0 \sim q$. Since $q$ is finite, $q^\perp \sim 1$ is properly infinite, and we may write $q^\perp = \sum_{k=1}^\infty q_k$, with $q_k \sim q^\perp \sim 1$. Let $p_1 \sim r_k \leq q_k$. Then $p = (p_0 + p_1) \sim (q + r_k) \overset{s}{\to} q$. This proves \eqref{E:finclos}. Now consider arbitrary $q$ and properly infinite $p$ with $c(p)=1$. Find the largest central projection $z$ with $zp \preccurlyeq zq$. Consider the nonempty net $\{zp_\alpha \mid zp \sim zp_\alpha \leq zq\}$, with order inherited from $\mathcal{P}(\mathcal{M})$. It is upward directed by Lemma \ref{T:proj}(1), applied to two projections. Its supremum is $zq$. By Lemma \ref{T:proj}(2) $z^\perp q^\perp \sim z^\perp$, which is properly infinite, so we may write $z^\perp q^\perp$ as the countable sum $\sum q_j$, with each $q_j \sim z^\perp q^\perp \sim z^\perp$. Write $z^\perp p = z^\perp p_0 + z^\perp p_1$, where $z^\perp p_0 \sim z^\perp q$. Also let $z^\perp p_1 \sim r_j \leq q_j$. Then $z^\perp p = (z^\perp p_0 + z^\perp p_1) \sim (z^\perp q + r_k) \overset{w}{\to} z^\perp q$. Combining the results for $zp$ and $z^\perp p$ and considering the product net, we see that $q$ is a strong limit of projections equivalent to $p$. This proves \eqref{E:infclos}. Equation \eqref{E:seg} follows from \eqref{E:finclos} and \eqref{E:infclos} by breaking off the largest central summand where $p$ is properly infinite with full central support. \end{proof} \begin{corollary} \label{T:Eclos} Let $\mathcal{M}$ be a factor and $E \subseteq \mbox{$(\p(\M)/\sim)$}$. We consider an extended center-valued trace $T$ on $\mathcal{M}$ to be a $[0,+\infty]$-valued function. If $\mathcal{M}$ is finite, $$\overline{E} = \{[q] \mid T(q) \in \overline{\{T(p) \mid [p] \in E\}}\}.$$ If $\mathcal{M}$ is properly infinite, $$\overline{E} = \{[q] \mid T(q) \leq \sup_{[p] \in E} T(p)\}.$$ \end{corollary} Corollary \ref{T:Eclos} follows readily from the preceding arguments, and its easy proof is left to the interested reader. \begin{corollary} $QOT$ is a $T_1$ topology exactly when $\mathcal{M}$ is finite. \end{corollary} \begin{proof} This is a direct consequence of Theorems \ref{T:sinfin} and \ref{T:sininf}. A topology is $T_1$ if for any two distinct points $x,y$, there is a closed set which contains $x$ and not $y$. If $\mathcal{M}$ is not finite, let $x$ be the equivalence class of a properly infinite projection, and let $y$ be $[0]$. Since $y$ belongs to the closure of $x$, no such separating closed set exists. (In general, a topology is $T_1$ iff singletons are closed.) \end{proof} It turns out to be more useful for our applications elsewhere (\cite{S}) to know when $QOT$ is $T_0$. A topology is $T_0$ if for any two distinct points, there exists a closed set which contains exactly one of them. \begin{proposition} \label{T:T0} For a von Neumann algebra $\mathcal{M}$, the following conditions are equivalent. \begin{enumerate} \item $QOT$ is a $T_0$ topology on $(\mathcal{P}(\mathcal{M})/\sim)$. \item For any $p,q \in \mathcal{P}(\mathcal{M})$, $[p] \in \overline{\{[q]\}} \Rightarrow p \preccurlyeq q$. \item $\kappa_\mathcal{M} \leq \aleph_0$. \item $\mathcal{M}$ is a (possibly uncountable) direct sum of $\sigma$-finite von Neumann algebras. \item $\mathcal{M}$ does not contain $\mathcal{B}(\mathfrak{H}_1)$, where $\mathfrak{H}_1$ is a Hilbert space of dimension $\aleph_1$. \end{enumerate} \end{proposition} \begin{proof} The equivalence of conditions (3)-(5) follows from the definitions and Theorem \ref{T:griffin}. We therefore focus on the equivalence of (1)-(3). (1) $\to$ (3): If (3) fails, let $q$ be an $\aleph_1$-homogeneous projection, and let $p$ be an $\aleph_0$-homogeneous projection with $c(p)=c(q)$. Then $[p] \in \overline{\{[q]\}}$ and $[q] \in \overline{\{[p]\}}$, but $[p] \neq [q]$. Clearly there is no closed set separating the two. (3) $\to$ (2): When $\kappa_\mathcal{M} \leq \aleph_0$, $T |_{\mathcal{P}(\mathcal{M})}$ can be identified with $D$. By Theorems \ref{T:sinfin} and \ref{T:sininf} we have $$[p] \in \overline{\{[q]\}} \Rightarrow T(p) \leq T(q) \Rightarrow D(p) \leq D(q) \Rightarrow p \preccurlyeq q.$$ (2) $\to$ (1): Suppose (2) holds. Given $[p],[q] \in \mbox{$(\p(\M)/\sim)$}$, they can be separated by a closed set if $[p] \notin \overline{\{[q]\}}$ or $[q] \notin \overline{\{[p]\}}$. If neither of these is true, then $$[p] \in \overline{\{[q]\}}, \: [q] \in \overline{\{[p]\}} \: \Rightarrow p \preccurlyeq q, \: q \preccurlyeq p \: \Rightarrow [p]=[q]. \qedhere$$ \end{proof} \section{From dimension function to trace in full generality} \label{S:trace} Let $T$ be an extended center-valued trace on a von Neumann algebra $\mathcal{M}$, with $D$ the induced dimension function. We will create a map which extends $D$ to the entire positive cone and so is a trace which distinguishes among infinite cardinalities. (In case $\kappa_\mathcal{M} \leq \aleph_0$, this process simply recovers $T$.) The main tool is \begin{definition} (\cite{KP}) For two elements $h,k \in \mathcal{M}_+$, we write $h \approx k$ if and only if there exists a family $\{x_\alpha\} \subset \mathcal{M}$ such that $h = \sum x_\alpha^* x_\alpha$ and $k = \sum x_\alpha x_\alpha^*$. We write $h \lessapprox k$ to mean that there exists $k' \leq k$ with $h \approx k'$. For $h \in \mathcal{M}_+$, we say that $h$ is \textit{finite} if $h \approx k \leq h \Rightarrow k=h$. \end{definition} The following facts are shown in \cite{KP}. \begin{itemize} \item The relation $\approx$ is an equivalence relation. It is homogeneous ($h \approx k \Rightarrow \lambda h \approx \lambda k, \: \lambda \in \mathbb{R}_+$) and completely additive in the sense that $$h_\alpha \approx k_\alpha, \: \forall \alpha \quad \Rightarrow \quad \sum h_\alpha \approx \sum k_\alpha$$ (when the two sums exist in $\mathcal{M}$). \item The relation $\lessapprox$ gives a partial order on equivalence classes. In particular, \begin{equation} \label{E:pord} h \lessapprox k, \: k \lessapprox h \: \Rightarrow \: h \approx k, \qquad h,k \in \mathcal{M}_+. \end{equation} \item For projections, $p \approx q \iff p \sim q$. \item For $h,k \in \mathcal{M}_+$, $h \lessapprox k \Rightarrow T(h) \leq T(k)$, and the converse holds if $h$ is finite. \end{itemize} We will also say that nonzero $h \in \mathcal{M}_+$ is \textit{properly infinite} if $zh$ is finite and nonzero for no central projection $z$. For projections, the usage here of ``finite" and ``properly infinite" coincides with the usual meaning; in fact proper infiniteness of (nonzero) $h$ in either case is characterized by $T(h)$ being $\{0,+\infty\}$-valued. \begin{lemma} \label{T:mult} ${}$ \begin{enumerate} \item Let $\lambda \in ((0,1) \cup (1,\infty))$, and let $p$ be a projection. Then $$p \text{ is properly infinite } \iff p \approx \lambda p.$$ \item Let $h,k \in \mathcal{M}_+$ have equal central support, with $k$ properly infinite and $h$ a countable sum of finite elements. Then $h \lessapprox k$. \item Let $h,k \in \mathcal{M}_+$ be properly infinite with equal central support, and suppose that each is a countably infinite sum of finite elements. Then $h \approx k$. \end{enumerate} \end{lemma} \begin{proof} (1) If $p \approx \lambda p$, then $T(p)$ must be $\{0,+\infty\}$-valued. For the opposite implication, we first check rational multiples. Let $m,n \in \mathbb{N}$. By proper infiniteness, we may write $$p = \sum_{i=1}^m p_i = \sum_{j=1}^n p'_n, \qquad p_i \sim p \sim p'_j, \: \forall i,j.$$ Then \begin{align*} p &= \sum_{i=1}^m p_i \approx \sum_{i=1}^m p = mp = \left( \frac{m}{n} \right) np = \left( \frac{m}{n} \right) \left(\sum_{j=1}^n p \right) \\ &\approx \left( \frac{m}{n} \right) \left( \sum_{j=1}^n p'_n \right) = \left( \frac{m}{n} \right) p. \end{align*} Find two positive rationals $\lambda_1, \lambda_2$ with $\lambda_1 \leq \lambda \leq \lambda_2$: $$p \approx \lambda_1 p \leq \lambda p \leq \lambda_2 p \approx p \; \Rightarrow \; p \approx \lambda p,$$ using \eqref{E:pord}. (2) Write $h = \sum_{j=1}^\infty h_j$, where each $h_j$ is finite. Since $T(h_1) \leq T(k)$, there is an operator $k_1$ with $h_1 \approx k_1 \leq k$. We continue in this way: since $T(h_n) \leq T(k - \sum_{j=1}^{n-1} k_j)$, find $k_n$ with $h_n \approx k_n \leq (h - \sum_{j=1}^{n-1} k_j)$. Now each $(\sum_{j=1}^n h_j) \approx (\sum_{j=1}^n k_j)$, and these terms are finite and increasing to $h$ and some $k'$, respectively. It follows from \cite[Lemma 3.3]{KP} that $h \approx k' \leq k$. (3) Both $h \lessapprox k$ and $h \gtrapprox k$ follow from the previous part; apply \eqref{E:pord}. \end{proof} \begin{proposition} \label{T:appproj} Let $h \in \mathcal{M}_+$ be properly infinite. Then there exists $p \in \mathcal{P}(\mathcal{M})$ such that $h \approx p$. \end{proposition} \begin{proof} It does no harm to assume that $h$ has full central support, and therefore $\mathcal{M}$ is properly infinite. Write the identity as $1_\mathcal{M} = \sum_{n=-\infty}^\infty p_n, \: 1_\mathcal{M} \sim p_n$, and let $r_0 \leq p_0$ be an $\aleph_0$-homogeneous projection with full central support. Now make the decomposition $$h = \sum_{n=1}^\infty (2^{-n} \|h\|) q_n,$$ where $q_n$ is the spectral projection for $h$ corresponding to $$\bigcup_{j=1}^{2^{n-1}} \left( (2j-1)2^{-n}\|h\|, (2j) 2^{-n} \|h\| \right].$$ For each $n \geq 1$, let $z^f_n$ be the largest central projection such that $z^f_n q_n$ is finite. Using Lemma \ref{T:mult}(1) and then conjugating by a partial isometry from $(1 - z^f_n)$ to $(1-z^f_n)p_n$, find a projection $r_n$ with $$(1-z_n^f)(2^{-n} \|h\|) q_n \approx (1-z^f_n) q_n \sim r_n \leq p_n.$$ Conjugating by a partial isometry from $z^f_n$ to $z^f_n p_{-n}$, let $r_{-n}$ be any operator (necessarily finite, but not necessarily a projection) with $$z^f_n (2^{-n} \|h\|) q_n \approx r_{-n} \in p_{-n} \mathcal{M} p_{-n}.$$ By construction we have $h \approx \sum_{n=1}^\infty (r_n + r_{-n})$. Set $z_0 = \wedge z^f_n$. We will complete the proof by showing that $z_0 h$ and $z_0^\perp h$ are both (Kadison-Pedersen) equivalent to projections. First, $$z_0 h \approx z_0 \left( \sum_{n=1}^\infty (r_n + r_{-n}) \right) = z_0 \left( \sum_{n=1}^\infty r_{-n} \right).$$ The left-hand side has central support $z_0$, and is either zero or properly infinite because $h$ is properly infinite. The right-hand side is a countable sum of finite elements. By Lemma \ref{T:mult}(3), $$z_0 h \approx z_0 r_0.$$ Second, $$z_0^\perp \left( \sum_{n=1}^\infty r_n \right) \sim z_0^\perp \left( r_0 + \sum_{n=1}^\infty r_n \right),$$ since the central supports are equal and the left-hand side is a properly infinite projection. (For example, this follows by evaluating the dimension function on both sides and noting that adding $\aleph_0$ does not change an infinite cardinal.) On the other hand, Lemma \ref{T:mult}(2) implies $$z_0^\perp \left( \sum_{n=1}^\infty r_{-n} \right) \lessapprox z_0^\perp r_0.$$ We put these together: \begin{align*} z_0^\perp h &\approx z_0^\perp \left( \sum_{n=1}^\infty (r_n + r_{-n}) \right) \lessapprox z_0^\perp \left( r_0 + \sum_{n=1}^\infty r_n \right) \\ &\sim z_0^\perp \left(\sum_{n=1}^\infty r_n \right) \lessapprox z_0^\perp \left( \sum_{n=1}^\infty (r_n + r_{-n}) \right) \approx z_0^\perp h. \end{align*} Then all terms above are (Kadison-Pedersen) equivalent, and the middle two are projections. \end{proof} \begin{corollary} \label{T:absorb} Under the same hypotheses as in Lemma \ref{T:mult}(2), $k \approx \lambda k$ for any $\lambda \in (0, \infty)$, and $(h + k) \approx k$. \end{corollary} \begin{proof} By Proposition \ref{T:appproj} and Lemma \ref{T:mult}(1), there is a properly infinite projection $p$ with $k \approx p \approx \lambda p \approx \lambda k$. By Lemma \ref{T:mult}(2), $$(h + k) \lessapprox 2k \approx k \lessapprox (h+k) \: \Rightarrow \: (h+k) \approx k.$$ \end{proof} We are now ready to define our map. \begin{definition} With $T$ (and $D$) given, we construct a \textbf{fully extended center-valued trace} $\widehat{T}$ on $\mathcal{M}$ as follows. For any $h \in \mathcal{M}_+$, let $z^f$ be the largest central projection so that $z^f h$ is finite. Let $p$ be a projection with $p \approx (1 -z^f) h$. Such a $p$ exists by Proposition \ref{T:appproj}, and all choices belong to the same Murray-von Neumann equivalence class. We define \begin{equation} \widehat{T}(h) = T(z^f h) + D((1- z^f) p), \end{equation} which we view as a continuous $([0,+\infty) \cup \{\kappa \mid \aleph_0 \leq \kappa \leq \kappa_\mathcal{M}\})$-valued function on $\Omega(\mathcal{Z}(\M))$. \end{definition} \begin{theorem} \label{T:Tfull} The map $\widehat{T}$ extends $D$, is additive, commutes with the multiplicative action of $\mathcal{Z}(\mathcal{M})_+$, and satisfies \begin{equation} \label{E:pres} h \lessapprox k \iff \widehat{T}(h) \leq \widehat{T}(k), \qquad h,k \in \mathcal{M}_+. \end{equation} (We are allowing cardinal arithmetic to incorporate the positive reals in the obvious way.) \end{theorem} \begin{proof} Clearly $\widehat{T}$ extends $D$. By the properties of $D$ and $T$ we have $h \approx k \iff \widehat{T}(h) = \widehat{T}(k)$. In saying that $\widehat{T}$ is additive, we mean that \begin{equation} \label{E:add} \widehat{T}(h + k) = \widehat{T}(h) + \widehat{T}(k), \qquad h,k \in \mathcal{M}_+. \end{equation} For $h,k$ finite, \eqref{E:add} follows from additivity of $T$. For $h,k$ properly infinite, the projection representing $h + k$ may be constructed as the sum of orthogonal representing projections for $h$ and $k$; \eqref{E:add} then follows from the additivity of $D$. Finally, let $h$ and $k$ have the same central support, with $h$ finite and $k$ properly infinite. In this case $\widehat{T}(h)$ is bounded above by $\aleph_0$, while $\widehat{T}(k) \geq \aleph_0$ where it is nonzero. So $\widehat{T}(h) + \widehat{T}(k) = \widehat{T}(k)$. Since $(h + k) \approx k$ by Corollary \ref{T:absorb}, $\widehat{T}(h + k) = \widehat{T}(k)$ as well. In saying that $\widehat{T}$ commutes with the action of $\mathcal{Z}(\mathcal{M})_+$, we mean \begin{equation} \label{E:modmap} y \widehat{T}(h) = \widehat{T}(yh), \qquad y \in \mathcal{Z}(\mathcal{M})_+, \: h \in \mathcal{M}_+. \end{equation} Clearly \eqref{E:modmap} holds for finite elements, since the analogous formula is true for $T$. It therefore suffices to prove \eqref{E:modmap} under the assumption that $h$ and $y$ have full central support, with $h$ properly infinite. In this case $y \widehat{T}(h) = \widehat{T}(h)$, so we are left to show that $yh \approx h$. If $y \geq \lambda c(y)$ for some $\lambda > 0$, then by Corollary \ref{T:absorb} $$h \approx \lambda h \leq yh \leq \|y\| h \approx h \: \Rightarrow \: h \approx yh.$$ The general conclusion follows by writing $y$ as a central sum of operators which are invertible on their supports. As for \eqref{E:pres}, the forward implication is a consequence of additivity. For the reverse implication, we look at central summands: where $h$ is finite, this is a property of $T$; where $h$ and $k$ are both infinite, this is a property of $D$. \end{proof} From Theorem \ref{T:Tfull}, we see that $\widehat{T}$ factors as $$\mathcal{M}_+ \twoheadrightarrow (\mathcal{M}_+/\approx) \overset{\sim}{\to} \widehat{T}(\mathcal{M}_+).$$ Here the second map is an embedding in a function space, preserving order, sums, and the multiplicative $\mathcal{Z}(\mathcal{M})_+$-action. More generally, we may say that an arbitrary completely additive map on $\mathcal{M}_+$ which respects the $\mathbb{R}_+$-action is \textit{tracial} if and only if it factors through the quotient $\mathcal{M}_+ \twoheadrightarrow (\mathcal{M}_+/\approx)$. Numerical (completely additive) traces result when the range is $[0,+\infty]$; they are ``one-dimensional representations" of $(\mathcal{M}_+/\approx)$. \begin{remark} \label{R:trace} Kadison and Pedersen observed that all extended center-valued traces on semifinite algebras can be generated in the following manner (\cite[Theorem 3.8]{KP}). Fix a finite projection $p$ with full central support, and assume that $p$ is the identity on the finite summand and abelian on the infinite type I summand (to match Convention \ref{C}). Then for finite $h \in \mathcal{M}_+$, $T(h)$ is the unique element of the extended center with $h \approx T(h)p$. Already this requires a small extension of $\approx$ to unbounded sums. With a further extension involving cardinals, $\widehat{T}$ can also be defined in this way. For general $\mathcal{M}$, let $p$ be the identity on the finite summand, abelian on the infinite type I summand, finite on the type II summand, and $\aleph_0$-homogeneous on the type III summand; of course $p$ should have full central support. For $h \in \mathcal{M}_+$, one can define $\widehat{T}(h)$ as the unique formal sum (as in \eqref{E:gdf}) such that $h \approx \widehat{T}(h)p$ and $\widehat{T}(h)$ takes no finite nonzero values on the type III summand. Probably this is more interesting to mention than to carry out, so we omit the details. \end{remark} \section{Continuity} \label{S:cont} In the remaininder of the paper we assume that $T$, $D$, and $\widehat{T}$ are given on $\mathcal{M}$. The order-preserving embeddings of $\mbox{$(\p(\M)/\sim)$}$ and $(\mathcal{M}_+/\approx)$ in a function space (albeit cardinal-valued) make pointwise operations available. From Theorems \ref{T:gdf} and \ref{T:Tfull} we know that for finite sets, addition in the quotient structures agrees with addition of functions. One may likewise add up infinite sets of functions, but there is no guarantee that the sum will be continuous. Tomiyama gave an example (\cite[Example 2]{To}) to show that for a pairwise orthogonal set $\{p_\alpha\}$, one cannot expect an identity between $\sum D(p_\alpha)$ and $D(\sum p_\alpha)$, so that $D$ is not completely additive. This is really an artifact of the function representation. $\mbox{$(\p(\M)/\sim)$}$ carries a natural (partially-defined) sum operation, given by $$\sum [p_\alpha] \triangleq \left[ \sum q_\alpha \right]$$ whenever there exists a set of pairwise orthogonal projections $\{q_\alpha\}$ with $q_\alpha \sim p_\alpha$. A similar definition is possible for sums in $\widehat{T}(\mathcal{M}_+)$, where we simply require that the representatives sum to an element of $\mathcal{M}_+$. Note that there is no ambiguity in these definitions, by \eqref{E:addeq} and the definition of $\approx$, and as an immediate consequence, the maps $\mathcal{P}(\mathcal{M}) \twoheadrightarrow \mbox{$(\p(\M)/\sim)$}$ and $\mathcal{M}_+ \twoheadrightarrow (\mathcal{M}_+/\approx)$ are completely additive. It is of course possible to transport these sum operations to $D(\mathcal{P}(\mathcal{M}))$ and $\widehat{T}(\mathcal{M}_+)$. \smallskip Pointwise lattice operations on pairs in $D(\mathcal{P}(\mathcal{M}))$ match \eqref{E:lattice} and so agree with the operations in $\mbox{$(\p(\M)/\sim)$}$, but meets and joins of infinite sets of continuous functions need not be continuous. For bounded real-valued functions on a stonean space, a regularization corrects this problem (\cite[Section III.1]{T}), but the situation for cardinal-valued functions is less clear. Normality for $D$ and $\widehat{T}$ means an appropriate analogue of \eqref{E:normal}. So how do we interpret an expression like $\sup D(p_\alpha)$, where $\{p_\alpha\}$ is an increasing net in $\mathcal{P}(\mathcal{M})$? As we just mentioned, the pointwise supremum need not lie in $D(\mathcal{P}(\mathcal{M}))$. In the next section we show that the supremum always does make sense in $\mbox{$(\p(\M)/\sim)$}$, but unfortunately normality is to much to ask. Tomiyama gave a simple example (\cite[Example 1]{To}) of an uncountable increasing family of projections $\{p_\alpha\}$ for which the pointwise supremum of $D(p_\alpha)$ does lie in $D(\mathcal{P}(\mathcal{M}))$, and yet $\sup D(p_\alpha) \neq D(\sup p_\alpha)$. This represents a phenomenon which really occurs in the quotient map $\mathcal{P}(\mathcal{M}) \twoheadrightarrow \mbox{$(\p(\M)/\sim)$}$, as we already saw in the proof of \eqref{E:infclos}. For $p$ a properly infinite projection, the elements of $[p]$, under the operator ordering, form an increasing net which converges strongly to $c(p)$. One obtains a counterexample to normality whenever $c(p) \notin [p]$, and such counterexamples exist when $\kappa_\mathcal{M} > \aleph_0$. On the other hand, if $\kappa_\mathcal{M} \leq \aleph_0$, the quotient maps are given by the extended center-valued trace, which we know to be normal. We conclude \begin{proposition} \label{T:normal} Another equivalent condition in Proposition \ref{T:T0} is \begin{enumerate} \item[(6)] The quotient maps $\mathcal{P}(\mathcal{M}) \twoheadrightarrow \mbox{$(\p(\M)/\sim)$}$ and $\mathcal{M}_+ \twoheadrightarrow (\mathcal{M}_+/\approx~)$ are normal. \end{enumerate} \end{proposition} In contrast, a pointwise criterion for normality of $D$ and $\widehat{T}$ holds if and only if $\kappa_\mathcal{M} \leq \aleph_0$ and the center of $\mathcal{M}$ is finite-dimensional. We do not bother to prove this explicitly, but we mention an example. Let $\mathcal{M} = \ell^\infty$, and take $p_n$ to be the sum of the first $n$ elements of the standard basis. Since $\sup D(p_n)$ does not agree with $D(\sup p_n)$ at any point of $(\beta \mathbb{N} \setminus \mathbb{N}) \subset \beta \mathbb{N} \simeq \Omega(\mathcal{Z}(\M))$, pointwise normality fails. And here $D$ is the identity! \smallskip Our conclusion from all this is that the pointwise lattice and addition operations on functions in the range of $D$ and $\widehat{T}$ should be shelved in favor of the induced quotient structures on $\mbox{$(\p(\M)/\sim)$}$ and $(\mathcal{M}_+/\approx)$. With this interpretation the assertion ``$D$ and $\widehat{T}$ are normal" is also equivalent to the conditions in Proposition \ref{T:T0}. \section{$\mbox{$(\p(\M)/\sim)$}$ is a complete lattice} \label{S:comp} Having just been warned about the degeneracies of the pointwise ordering, we omit the last step of Tomiyama's construction for $D$ and stick with a more algebraic language. We follow the right-hand side of \eqref{E:gdf}, further dividing $T(z^f p)$ into the pieces where it lies between consecutive finite cardinals. This allows us to write the typical element of $\widehat{T}(\mathcal{M}_+)$ as \begin{equation} \label{E:simpler} \sum_{\kappa \leq \kappa_\mathcal{M}} g_\kappa z_\kappa. \end{equation} The meaning of this expression is as follows. If $\kappa$ is an infinite cardinal, then $g_\kappa = \kappa$. If $\kappa$ is a nonnegative integer, $g_\kappa$ is an element of $\mathcal{Z}(\mathcal{M})_+$ satisfying $(\kappa-1)z_\kappa \leq g_\kappa \leq \kappa z_\kappa$ and $c(g_\kappa - (\kappa-1)z_\kappa) = z_\kappa$. The central projections $z_\kappa$ sum to 1, and the decomposition is unique. The partial order, pairwise sum operation, and pairwise lattice operations are all easily implemented for expressions of the form \eqref{E:simpler}. Conversely, such an expression belongs to $\widehat{T}(\mathcal{M}_+)$ if it is $\{0, +\infty\}$-valued on the type III summand and less than some finite multiple of $\widehat{T}(1_\mathcal{M})$. To belong to $D(\mathcal{P}(\mathcal{M})) \subseteq \widehat{T}(\mathcal{M}_+)$, an expression must be $\leq D(1_\mathcal{M})$ and appropriately valued on both the type I and type III summands. \begin{theorem} \label{T:complat} $(\mathcal{M}_+/\approx)$ and $\mbox{$(\p(\M)/\sim)$}$ are complete lattices. \end{theorem} \begin{proof} We show how to perform lattice operations on formal sums of the form \eqref{E:simpler}. Our constructions will preserve all conditions mentioned in the paragraph before Theorem \ref{T:complat}, so they are well-defined operations in $(\mathcal{M}_+/\approx)$ and $\mbox{$(\p(\M)/\sim)$}$. Let us find the supremum of an arbitrary set $\{f^\alpha\}$, where $$f^\alpha = \sum g_\kappa^\alpha z_\kappa^\alpha.$$ For each cardinal $\kappa \leq \kappa_\mathcal{M}$, set $$y_{\leq \kappa} = \bigwedge_\alpha \left(\sum_{\lambda \leq \kappa} z^\alpha_\lambda \right);$$ $y_{\leq \kappa}$ is ``where all $f^\alpha$ are $\leq \kappa$". Note that $y_{\leq \kappa}$ is increasing in $\kappa$ and $y_{\leq \kappa_\mathcal{M}} = 1$. Next define, for each cardinal $\kappa \leq \kappa_\mathcal{M}$, $$z_\kappa = y_{\leq \kappa} - \bigvee_{\lambda < \kappa} y_{\leq \lambda}.$$ The $z_\kappa$ are pairwise disjoint: if $\kappa_1 < \kappa_2$, then $$z_{\kappa_1} \leq y_{\leq \kappa_1} \perp z_{\kappa_2}.$$ Notice also that $\sum z_\kappa = 1$. For if there were $z \in \mathcal{P}(\mathcal{Z}(\mathcal{M}))$ with $z \perp (\sum z_\kappa)$, then let $\lambda$ be the least cardinal with $z y_{\leq \lambda} \neq 0$; by definition $z z_\lambda \neq 0$ as well, which contradicts the assumption. We claim that \begin{equation} \label{E:sup} \sup_\alpha f^\alpha = \sum g_\kappa z_\kappa \triangleq f, \end{equation} where $g_\kappa = \kappa$ when $\kappa$ is infinite, and otherwise $g_\kappa = \sup_\alpha (g_\kappa^\alpha z_\kappa)$, which exists as the supremum of a bounded set in $\mathcal{Z}(\mathcal{M})_+$. Next we show that $f \geq f^\alpha$ for any $\alpha$. Fixing a cardinal $\lambda \leq \kappa_\mathcal{M}$, \begin{equation} \label{E:cut} z_\lambda f^\alpha = z_\lambda \left( \sum g_\kappa^\alpha z_\kappa^\alpha \right) = (z_\lambda y_{\leq \lambda}) \left( \sum g_\kappa^\alpha z_\kappa^\alpha \right) \leq z_\lambda \left( \sum_{\kappa \leq \lambda} g_\kappa^\alpha z_\kappa^\alpha \right). \end{equation} When $\lambda$ is infinite, we continue \eqref{E:cut} as $$\leq \lambda z_\lambda = z_\lambda f.$$ When $\lambda$ is finite, we continue \eqref{E:cut} as $$\leq z_\lambda (\lambda - 1) \left( \sum_{\kappa < \lambda} z_\kappa^\alpha \right) + z_\lambda g_\lambda^\alpha z_\lambda^\alpha \leq z_\lambda g_\lambda = z_\lambda f.$$ Since $z_\lambda f^\alpha \leq z_\lambda f$ for all $\lambda$, $f \geq f^\alpha$. Finally we check that if $h = \sum h_\kappa x_\kappa$ satisfies $h \geq f^\alpha$, $\forall \alpha$, then necessarily $h \geq f$. Fixing a cardinal $\lambda \leq \kappa_\mathcal{M}$, \begin{align*} h_\lambda x_\lambda = x_\lambda h \geq x_\lambda f^\alpha, \: \forall \alpha \quad &\Rightarrow \quad x_\lambda \leq \sum_{\kappa \leq \lambda} z^\alpha_\kappa, \: \forall \alpha \\ &\Rightarrow \quad x_\lambda \leq \bigwedge_\alpha \left(\sum_{\kappa \leq \lambda} z^\alpha_\kappa \right) = y_{\leq \lambda}. \end{align*} This last inequality implies \begin{equation} \label{E:cut2} x_\lambda f = x_\lambda y_{\leq \lambda} f \leq x_\lambda \left( \sum_{\kappa \leq \lambda} g_\kappa z_\kappa \right). \end{equation} When $\lambda$ is infinite, we continue \eqref{E:cut2} as $$\leq \lambda x_\lambda = x_\lambda h.$$ When $\lambda$ is finite, we continue \eqref{E:cut2} as $$\leq x_\lambda (\lambda - 1) \left( \sum_{\kappa < \lambda} z_\kappa \right) + x_\lambda g_\lambda z_\lambda$$ and the inequality $h \geq f^\alpha$, $\forall \alpha$, allows us to compute further $$ = x_\lambda (\lambda - 1) \left( \sum_{\kappa < \lambda} z_\kappa \right) + x_\lambda \left( \sup_\alpha g_\lambda^\alpha z_\lambda \right) \leq x_\lambda h.$$ Since $x_\lambda f \leq x_\lambda h$ for all $\lambda$, $f \leq h$. This completes the proof that $f = \sup f^\alpha$. As for the infimum of the $f^\alpha$, we first point out that we cannot write anything like $$\bigwedge f^\alpha = 1 - \left(\bigvee (1-f^\alpha)\right),$$ which is a useful duality in $\mathcal{P}(\mathcal{M})$. There is no complementation in the lattices $\mbox{$(\p(\M)/\sim)$}$ and $(\mathcal{M}_+/\approx)$, at least when $\mathcal{M}$ is not finite. Instead we define $$y_{\leq \kappa} = \bigvee_\alpha \left(\sum_{\lambda \leq \kappa} z_\lambda^\alpha \right)$$ and complete the rest of the proof similarly to the proof for the supremum. (The substitute for \eqref{E:cut} should begin with ``$z_\lambda^\alpha f = \dots$"; for \eqref{E:cut2} should begin with ``$z_\lambda h = \dots$".) \end{proof} \section{Application to representation theory} \label{S:rep} In this section we reinterpret Theorem \ref{T:complat} in terms of the (normal Hilbert space) representations of $\mathcal{M}$. Unless noted otherwise, we use ``isomorphism" in the sense of normed $\mathcal{M}$-modules, i.e. $$\{\pi_1, \mathfrak{H}_1\} \simeq \{\pi_2, \mathfrak{H}_2\} \iff$$ $$\exists \text{ unitary }U: \mathfrak{H}_1 \to \mathfrak{H}_2: \qquad U\pi_1(x)U^* = \pi_2(x), \qquad \forall x \in \mathcal{M}.$$ It follows from the basic theory (see \cite[Sections 2.1-2]{JS} or \cite[Section 2]{S2003}) that any representation is (isomorphically) contained in a direct sum of copies of the standard form $\{\text{id}, L^2(\mathcal{M})\}$. We view $\oplus_I L^2(\mathcal{M})$ as a row vector and think of the $\mathcal{M}$-action as multiplication on the left. The commutant is right multiplication by $\mathcal{B}(\ell^2_I) \overline{\otimes} \mathcal{M}$, and the closed submodules are of the form $(\oplus_I L^2(\mathcal{M})) q$, where $q \in \mathcal{P}(\mathcal{B}(\ell^2_I) \overline{\otimes} \mathcal{M})$. Two submodules are isomorphic if and only if the corresponding projections are equivalent. This means that the isomorphism class of a representation corresponds to an equivalence class of projections in some amplification of $\mathcal{M}$. Adding representations corresponds to adding equivalence classes. As we have mentioned, the partial order can be defined in terms of the sum, so provided we make some kind of size restriction, we get an isomorphism of ordered monoids. For example, if $\mathcal{M}$ is $\sigma$-finite, we obtain an identification between $(\mathcal{P}(\mathcal{B}(\ell^2) \overline{\otimes} \mathcal{M})/\sim)$ and isomorphism classes of countably generated Hilbert $\mathcal{M}$-modules. This all works for $L^p$ modules (\cite{JuS}), too, and is closely related to the $\mathcal{K}_0$ functor (\cite{Han,W-O}). (Most of the ideas of the preceding two paragraphs were discussed by Breuer (\cite{B1968,B1969}), without making reference to standard forms. He focused on the monoid generated by equivalence classes of finite projections, because the associated Grothendieck group, called the \textit{index group} of $\mathcal{M}$, is the natural carrier for the Fredholm theory of $\mathcal{M}$. Olsen (\cite{O}) later combined Breuer's work with Tomiyama's dimension function to give a very general version of index theory in von Neumann algebras.) \begin{corollary} Let $\{\pi_\alpha, \mathfrak{H}_\alpha\}$ be a set of representations of a fixed von Neumann algebra $\mathcal{M}$. Then there is a maximal representation of $\mathcal{M}$ which is (isomorphically) contained in all of these, and there is a minimal representation which (isomorphically) contains all of these. Both are unique up to $\mathcal{M}$-module isomorphism. \end{corollary} \begin{proof} Choose a large enough set $I$ so that for all $\alpha$, $\{\pi_\alpha, \mathfrak{H}_\alpha\}$ is a subrepresentation of $\{\text{id}, \oplus_I L^2(\mathcal{M})\}$. The corollary follows from the preceding discussion and the fact that $(\mathcal{P}(\mathcal{B}(\ell^2_I) \overline{\otimes} \mathcal{M})/\sim)$ is a complete lattice. \end{proof} In the early years of the subject, von Neumann algebras were generally given on Hilbert spaces, and the notion of $\mathcal{M}$-module isomorphism was therefore not in use. Instead, one classified represented algebras up to the slightly weaker notion of \textit{spatial isomorphism}, which allows for an arbitrary isomorphism between the algebras. (An $\mathcal{M}$-module isomorphism between representations $\{\pi_1, \mathfrak{H}_1\}$ and $\{\pi_2, \mathfrak{H}_2\}$ is a spatial isomorphism between von Neumann algebras $\{\pi_1(\mathcal{M}), \mathfrak{H}_1\}$ and $\{\pi_2(\mathcal{M}), \mathfrak{H}_2\}$ which induces the algebra isomorphism $\pi_2 \circ \pi_1^{-1}$.) The question ``When is an algebraic isomorphism of represented von Neumann algebras spatial?", which is a noncommutative version of the fundamental problem of unitary equivalence for normal operators, is treated in detail in \cite{K1957}. Also see \cite{Dig} for a projection-based approach to the existence of spatial isomorphisms. Having said that, equivalence classes of representations/represented algebras were first studied by Murray and von Neumann (\cite[Chapter III]{MvN4}), using the coupling constant for finite factors. The generalizations to coupling functions and arbitrary algebras were the motivations for the Griffin and Pallu de la Barri\`{e}re results featured in Section \ref{S:trace}. The space-free approach was notably developed by the Japanese school of the 1950's. Modulo spatial isomorphism, the set of equivalence classes of representations of a fixed von Neumann algebra may not even be partially ordered. We mention the relevant example. Let $\mathcal{M}$ be a type $\text{II}_\infty$ factor with dimension function $D$ and fundamental group $\Gamma \notin \{\{1\},(0,\infty)\}$. (The existence of such an $\mathcal{M}$ remained in doubt until a breakthrough of Connes in 1980 (\cite{C}). The fundamental group of a $\text{II}_\infty$ factor can be defined as $$\{\lambda \in (0,\infty) \mid \exists \alpha \in \text{Aut}(\mathcal{M}), \; D \circ \alpha = \lambda D\},$$ with the group operation being multiplication.) Kadison (\cite{K1955}) showed that for nonzero finite projections $p,q$, $L^2(\mathcal{M})p$ is spatially isomorphic to $L^2(\mathcal{M})q$ if and only if $\frac{D(p)}{D(q)} \in \Gamma$. Since $\Gamma \neq (0,\infty)$, we may find nonzero finite projections $p,p'$ with $\frac{D(p)}{D(p')} \notin \Gamma$. And $\Gamma \neq \{1\}$, so we may find spatial isomorphisms $L^2(\mathcal{M})q_1 \simeq L^2(\mathcal{M})p' \simeq L^2(\mathcal{M})q_2$ with $q_1 \lneqq p \lneqq q_2$. Therefore the spatial equivalence class of $L^2(\mathcal{M})p$ both dominates and is dominated by that of $L^2(\mathcal{M})p'$, yet the two are not equal. At least for factors, this kind of pairing - $\text{II}_\infty$ algebra, $\text{II}_1$ commutant - is the only case where the two notions of equivalence differ. Not coincidentally, the only choice required for $T$, $D$, and $\widehat{T}$ which cannot be standardized is the normalization on the finite elements in a $\text{II}_\infty$ summand. (On a $\text{II}_\infty$ summand, one possible definition for ``normalization" is the inverse image of the identity, which is nothing but the equivalence class of the projection $p$ discussed in Remark \ref{R:trace}.)
{ "timestamp": "2005-03-31T20:16:20", "yymm": "0503", "arxiv_id": "math/0503747", "language": "en", "url": "https://arxiv.org/abs/math/0503747" }
\section{Introduction and statement of main results} We consider volume-preserving or symplectic diffeomorphisms on a compact connected Riemannian manifold $M$. Let $\mbox{{\rm Diff}$_\mu^r(M)$}$ be the the set of all $C^r$ diffeomorphisms preserving a smooth volume $\mu$ on $M$. If $r$ is not an integer, $r= k + \alpha$ for some positive integer $k$ and $0 < \alpha < 1$, it is understood that the functions in $\mbox{{\rm Diff}$_\mu^r(M)$}$ are $C^k$ functions with $\alpha$-H\"older $k$-th derivatives. An invariant set $\Lambda \subset M$ is said to be {\em hyperbolic}\/ if there is a continuous splitting of $T_xM = E^s_x \oplus E^u_x$ for every $x \in \Lambda$ and constants $C>0$, $\lambda >1$ such that \begin{eqnarray} df_x(E^s_x) &=& E^s_{f(x)} \; \mbox{ and } \; df_x(E^u_x) = E^u_{f(x)} \nonumber \\ |df^n_x v^s_x| &\leq& C \lambda^{-n} |v^s|,\; \mbox{ for all } v^s \in E^s_x, \; n \in \mathbb{N} \nonumber \\ |df^{-n}_x v^u_x| &\leq& C \lambda^{-n} |v^u|,\; \mbox{ for all } v^u \in E^u_x, \; n \in \mathbb{N} \nonumber \end{eqnarray} If the whole manifold $M$ is hyperbolic for some $f \in \mbox{{\rm Diff}$_\mu^r(M)$}$, then $f$ is said to be {\em Anosov}. Not all manifolds can support Anosov diffeomorphisms. Typical examples of hyperbolic invariant sets are Cantor sets as in Smale's horseshoe map. The following simple proposition explains why this is the case. \begin{prop} Let $f \in \mbox{{\rm Diff}$_\mu^r(M)$}$, $r \geq 1$, be a volume-preserving diffeomorphism on a compact manifold $M$. Let $\Lambda \subset M$ be a closed hyperbolic invariant set. If the interior of $\Lambda$ is non-empty, then $f$ is Anosov on $M$ and $\Lambda =M$. \label{prop} \end{prop} This proposition and its simple proof, given in the next section, will motivate our main result of this paper. The proof uses the fact that the recurrent points are dense on the manifold. This is a consequence of the volume-preserving property. Without the volume-preserving or the dense recurrent points condition, the proposition is not true, we refer to Fisher \cite{Fisher04} for a counter-example. Fisher also give a proof of the above proposition. On the other hand, it is an open problem whether there are any Anosov diffeomporphisms with wandering domains. A natural question one asks is whether there is any hyperbolic invariant set with a positive measure for a volume-preserving non-Anosov diffeomorphism. The answer is yes for $C^1$ diffeomorphisms, as Bowen's example of fat horseshoe shows \cite{Bowen75}, see also Robinson \& Young \cite{RY80}. However, if the map is assumed to be $C^{1 + \alpha}$ for some $\alpha >0$, then the answer is no. This is the main result of this paper. \begin{thm} Let $f \in \mbox{{\rm Diff}$_\mu^r(M)$}$, $r>1$, be a volume preserving diffeomorphism on a compact manifold $M$. Let $\Lambda \subset M$ be a closed hyperbolic invariant set. If $\mu(\Lambda) >0$, then $f$ is Anosov on $M$ and $\Lambda = M$. \label{thm} \end{thm} It is not surprising that the map is required to be $C^{1+ \alpha}$. As various examples show, measure-theoretical properties are often not respected by $C^1$ maps. Additional smoothness, even though very little, guarantees certain regularities in measure. Our proof uses a special type of measure density points different from the Lebesgue density points. The density basis for our density points are dynamically defined. It is similar to the juliennes defined by Pugh \& Shub \cite{PS00} \cite{PS03}. But our case is much simpler. Our method also provides a direct proof of the ergodicity of $C^{1 + \alpha}$ volume-preserving Anosov diffeomorphisms, without using the Hopf arguments or the Birkhoff ergodic theorem. However, we do use the absolute continuity of stable and unstable foliations. This is given in the last section of the paper. Another result of this paper Lemma \ref{lemf} is of interest in its own right. We showed that for a $C^{1+ \alpha}$ hyperbolic or partially hyperbolic volume-preserving diffeomorphism, if a set is invariant, then almost every point of the stable manifold (or unstable manifold) of almost every point is in the invariant set. This is also true for non-unifomly hyperbolic invariant set in Pesin theory. This result can find its applications in various other problems. \vspace{1ex} We are grateful to M. Viana for pointing out that Theorem \ref{thm} was also proved, with a different method, by Bochi and Viana \cite{BV03}. We are also grateful to F. Ledrappier for showing us another possible proof of Theorem \ref{thm}. \section{Proof of the proposition} In this section, we give a simple proof of the Proposition \ref{prop}. Let $U$ be the interior of the hyperbolic invariant set $\Lambda$. By the assumption of the proposition, $U \neq \emptyset$. Clearly, $U$ is invariant. We want to prove that the closure of $U$, $\bar{U}$ is the whole manifold. We know that $\bar{U}$ is closed, it suffices to show that $\bar{U}$ is also open. For any $x \in \bar{U}$, there exists a sequence of points $x_n \in U$, $n \in \mathbb{N}$, such that $x_n \rightarrow x$ as $n \rightarrow \infty$. As $f$ is volume preserving, by Poincar\'e recurrence theorem, almost every point is both forward and backward recurrent. Moreover, the set of periodic points is dense in $U$, since $U$ is hyperbolic. We may choose $\{x_n\}_{n \in \mathbb{N}}$ to be periodic points. Since $U$ is invariant and each $x_n$ is an interior point in $U$, then $W^s(x_n)$ and $W^u(x_n)$ are in $U$ for all $n \in \mathbb{N}$. For each fixed $\delta >0$ small, let $W^s_\delta(x)$ and $W^u_\delta(x)$ be, respectively, the local stable manifold and unstable manifold of $x$. As $x_n \rightarrow x$, as $n \rightarrow \infty$, we have that $W^s_\delta(x_n) \rightarrow W^s_\delta(x)$ and $W^u_\delta(x_n) \rightarrow W^u_\delta(x)$ as $n \rightarrow \infty$. This implies that each point on $W^s_\delta(x)$ or on $W^u_\delta(x)$ is also in the closure of $U$. Let $y \in W^s_\delta(x)$ and $z \in W^u_\delta(x)$, the same argument shows that $W^u_\delta(y)$ and $W^s_\delta(z)$ are both in the closure of $U$. Consequently, $$W^u_\delta(y) \cap W^s_\delta(z) \in \bar{U},$$ i.e., $\bar{U}$ has the product structure. This implies that $x$ is in the interior of $\bar{U}$. Consequently, the set $\bar{U}$ is open. Since $\bar{U}$ is also closed and $M$ is connected, we have $\bar{U} = M$. i.e., $f$ is hyperbolic on $M$. This proves the proposition. \section{Proof of the Theorem} The proof of Theorem \ref{thm} uses a similar idea to the proof of Proposition \ref{prop}, but the details are much more complicated. Here the interior points are replaced by density points. One may regard the density points as measure theoretical interior points for a set with positive measure. We need some preliminary results from standard smooth ergodic theory. It is well-known that the stable and unstable foliations for a $C^1$ Anosov diffeomorphism may not be absolutely continuous. However, for $C^{1 +\alpha}$ diffeomorphisms, these foliations are absolutely continuous (Anosov \cite{Anosov67}). Moreover, the stable and unstable foliations over a hyperbolic (even non-uniformly, cf Pesin \cite{Pesin77}) invariant set are also absolutely continuous for $C^{1 + \alpha}$ diffeomorphisms. In fact, the absolute continuity of the foliations is proved by showing that the holonomy maps of these foliations are absolutely continuous. We also need some results on density basis and density points of a measurable set. Let $A \subset \mathbb{R}^n$ be a measurable set with the standard Lebesgue measure $m$. A point $x \in \mathbb{R}^n$ is said to be a Lebesgue density point if $$\lim_{\epsilon \rightarrow 0} \frac{m(B(x, \epsilon) \cap A)}{m(B(x, \epsilon))} =1,$$ where $B(x, \epsilon)$ is the $\epsilon$-ball centered at $x$. Lebesgue density theorem states that almost every point of $A$ is a density point for $A$. To prove our theorem, we need a different definition of density point. The Lebesgue density point is defined by a basis of $\epsilon$-balls. We replace it by a dynamically defined basis. Let $\Lambda \subset M$ be a hyperbolic invariant set, we first define a basis on the unstable manifold for each point $\Lambda$. For a fixed small real number $\delta >0$, let $W^u_\delta(x)$ be the local unstable manifold of a point $x \in \Lambda$. Let $\mu_u$ and $\mu_s$ respectively be the induced measures of the smooth volume form $\mu$ on the unstable leaves and stable leaves. Let $n_u$ and $n_s$ respectively be the dimensions be the unstable and stable leaves. For any positive integer $k$, let $B^u_k(x)$ be a subset of $W^u(x)$ defined by $$B^u_k(x) = f^{-k}(W^u_\delta(f^k(x))). $$ Clearly, the cubes $B^u_k(x)$, $k \in \mathbb{N}$ shrinks to the point $x$ as $k \rightarrow \infty$. We call the collection of the sets $\{B^u_k(x) \; | \; k \in \mathbb{N}, \; x \in \Lambda \}$ the unstable density basis. Similarly, we can define the stable density basis $\{B^s_k(x) \; | \; k \in \mathbb{N}, \; x \in \Lambda \}$, by defining $$B^s_k(x) = f^{k}(W^u_\delta(f^{-k}(x))). $$ The density basis we defined has infinite eccentricity. A point $x \in \Lambda$ is said to be a dynamical density point, or simply density point, on the unstable foliation if $$ \lim_{k \rightarrow \infty} \frac{\mu_u(B^u_k(x) \cap \Lambda)}{\mu_u(B^u_k(x))} =1.$$ Similarly, we can define the dynamical density points on the stable foliation. \begin{prop} The set of points on $\Lambda$ that are both density points on the stable foliations and unstable foliations has the full measure in $\Lambda$. \label{propd} \end{prop} Our definitions of density points can be regarded as simplified versions of the juliennes density points defined by Pugh \& Shub \cite{PS00} \cite{PS03}. The Pugh-Shub density points are defined by the julienne density basis and they are much more complicated than what we have here. The above proposition follows from the proof for the Pugh-Shub's juliennes density point. The proof itself is similar to the proof of the Lebesgue Density Theorem. The key properties for the density basis are scaling and engulfing defined as follows. (a) Scaling: for any fixed $k \geq 0$, $m(B^u_n(x)) /m( B^u_{n+k}(x))$ is unformly bounded as $n \rightarrow \infty$. (b) Engulfing: there is a unifom $L$ such that $$B^u_{n+L}(x) \cap B^u_{n+L}(y) \neq \emptyset \; \Rightarrow \; B^u_{n+L}(x) \cup B^u_{n+L}(y) \subset B^u_n(x). $$ These two properties can be easily verified. We remark that our density points are defined on the stable and unstable foliations, we freely used the fact that the stable and unstable foliations are absolutely continuous. Let $A$ be a subset of $\Lambda$ such that for any $x \in A$, $x$ is a density point of $\Lambda$ on both stable foliation and unstable foliation; and $x$ is a recurrent point, both forward and backward. By Poincar\'e recurrence theorem, Proposition \ref{propd} and the absolute continuity of the foliations, the set $A$ has the full measure in $\Lambda$. The following is the main lemma in proving our theorem. \begin{lem} Assume that $f$ is a $C^{1+\alpha}$ volume-preserving diffeomorphism, for some positive number $\alpha >0$. Fix $x \in A$ and a positive number $\delta >0$. Then for any $\epsilon >0$, there exists a positive integer $k_0$, depending on $x$ and $\epsilon$, such that for $k \geq k_0$, $$\mu_s(W^s_\delta(f^{-k}(x)) \cap A) \geq (1-\epsilon) \mu_s(W^s_\delta(f^{-k}(x)))$$ and $$\mu_u(W^u_\delta(f^k(x)) \cap A) \geq (1-\epsilon) \mu_u(W^u_\delta(f^k(x)))$$ i.e., for sufficiently large $k$, the set $A$ has a very high density in $W^s_\delta(f^{-k}(x))$ and $W^u_\delta(f^k(x))$. \end{lem} \noindent {\it Proof of the lemma}: We first prove the lemma for the unstable foliation. Let $n_u$ be the dimension of the leaves of the foliation, local unstable manifold $W^u_\delta(x)$ can be identified with a cube in $E^u_x = \mathbb{R}^{n_u}$ by the exponential map from $E^u_x$ to $W^u(x)$. Since the leaves of the unstable foliation is smooth, the conditional measure $\mu_u$ are smoothly equivalent to the standard Lebesgue measure $m$ on $\mathbb{R}^{n_u}$. i.e., for any point $x \in \Lambda$, there is a smooth function $g_u(y)$ defined for $y \in W^u_\delta(x)$ on the local unstable manifold, uniformly bounded away from zero and infinity, such that $$\mu_u(E) = \int_{E} g_u dm,$$ where $E$ is a measurable set in $W^u_\delta(x)$ and $m$ is the standard Lebesgue measure in $\mathbb{R}^{n_u}$. For any positive integer $k$, we want to estimate the measure of the set $W^u_\delta(f^k(x)) \cap A$. Let $B^k_0 = f^{-k}(W^u_\delta(f^k(x)))$, obviously $B^k_0 \subset W^u_\delta(x)$. In fact, $B^k_0$ is the set $B^u_k(x)$ in our definition of density basis. We iterate $B^k_0$ under $f$ and obtain a sequence of sets $B^k_i = f^i(B^k_0)$, for $i=1, 2, \ldots, k$. The last set in the sequence is $B^k_k = W^u_\delta(f^k(x))$. Let $\eta_k = 1- \mu_u(B^k_0 \cap A) / \mu_u(B^k_0)$, then $0 \le \eta_k \leq 1$. As $x$ is a density point on the unstable foliation, $$\lim_{k \rightarrow \infty} \frac{\mu_u(B^k_0 \cap A)}{\mu_u(B^k_0)} =1,$$ The number $\eta_k$ is small for large $k$, and $\lim_{k \rightarrow \infty} \eta_k =0$. Since $f$ is $C^{1 + \alpha}$, there exists a constant $C_1 > 0$ such that $||df_y - df_z|| \leq C_1 |y-z|^\alpha$. Here we abuse the notation a little by writing $|y-z|$ as the distance between $y$ and $z$. Let $\rho^k_i$ be the maximum distance from $f^i(x)$ to the boundary of $B^k_i$, i.e., $$\rho^k_i = \max_{y \in B^k_i} \{ d(f^i(x), y) \}. $$ For any $y \in B^k_0$, $||df_y - df_x|| \leq C_1(\rho^k_0)^\alpha$. In general, for any $y \in B^k_i$, $||df_y - df_{f^{i}(x)}|| \leq C_1 (\rho^k_i)^\alpha$. Let $J_u(y)=|\det(df_y|_{E^u_y})|$ be the Jacobian of the map $f$ at $y$ restricted on the unstable manifold of $y$. Then $|J_u(x)-J_u(y)| \leq C_2 |x-y|^\alpha$, for some positive constant $C_2 >0$. Let $D_0= B^k_0 \backslash A$ and $D_i = B^k_i \backslash A$, for $i =1, 2, \ldots, k$. These are the complements of $A$ in $B^k_i$. By the definition of $\eta_k$, $\mu_u(D_0) = \eta_k \mu_u(B^k_0)$. We need to estimate the measure of $D_1$. For any set $E \subset B^k_0$, $$\mu_u(f(E)) = \int_{f(E)} g_u dm =\int_{E} J_u (g_u\cdot f) dm = \int_{E} J_u (g_u\cdot f) g_u^{-1} d\mu_u$$ Since the functions $g_u$ and $g_u^{-1}$ are smooth on any unstable manifold, the integrand in the above integral is $C^{1+\alpha}$, there is a constant $C_3 >0$ such that $$|J_u(y) (g_u\cdot f)(y) g_u^{-1}(y) - J_u(x) (g_u\cdot f)(x) g_u^{-1}(x)| \leq C_3 |x - y|^\alpha , $$ for all $x\in \Lambda$, $y \in W^u_\delta(x)$. Therefore, $$ |\mu_u(f(E)) - (J_u(x) (g_u\cdot f)(x) g_u^{-1}(x)) \mu_u(E)| \leq C_3 (\rho^k_0)^\alpha \mu_u(E).$$ Consequently, $$\mu_u(D_1) \leq (J_u(x) (g_u\cdot f)(x) g_u^{-1}(x)) \mu_u(D_0) + C_3 (\rho^k_0)^\alpha \mu_u(D_0)$$ and $$\mu_u(B^k_1\backslash D_1) \geq (J_u(x) (g_u\cdot f)(x) g_u^{-1}(x)) \mu_u(B^k_0 \backslash D_0) - C_3 (\rho^k_0)^\alpha \mu_u(B^k_0\backslash D_0)$$ and therefore $$\mu_u(D_1) \leq \eta_k \frac{(1 + C_3 (\rho^k_0)^\alpha)}{(1 - C_3 (\rho^k_0)^\alpha)} \mu_u(B^k_1).$$ By induction on $i$, we have $$\mu_u(D_k) \leq \eta_k \frac{(1 + C_3 (\rho^k_0)^\alpha)}{(1 - C_3 (\rho^k_0)^\alpha)}\frac{(1 + C_3 (\rho^k_1)^\alpha)}{(1 - C_3 (\rho^k_1)^\alpha)} \cdots \frac{(1 + C_3 (\rho^k_{k-1})^\alpha)}{(1 - C_3 (\rho^k_{k-1})^\alpha)}\mu_u(B^k_k). $$ The map $df: T_{\Lambda}M \rightarrow T_{\Lambda}M$ uniformly expands vectors on the unstable splitting. That uniform expansion extends to local unstable manifolds $W^u_\delta(x)$, $x \in \Lambda$ if $\delta$ is chosen small enough. There exist positive real numbers $C_4 > C >0$ and $\lambda> \lambda_1 >1$ (here $C$ and $\lambda$ are the same as those in the definition of the hyperbolic invariant set) such that $\rho^k_{k-1} \leq C_4 \lambda_1^{-1} \delta$ and $\rho^k_{i} \leq C_4 \lambda_1^{-(k-i)} \delta$, for $i=0, 1, \ldots, k-1$. This implies that \begin{eqnarray} \mu_u(D_k) & \leq & \eta_k \frac{(1 + C_3 (C_4 \lambda_1^{-k}\delta)^\alpha)}{(1 - C_3 (C_4 \lambda_1^{-k}\delta)^\alpha)}\frac{(1 + C_3 (C_4 \lambda_1^{-k+1}\delta)^\alpha)}{(1 - C_3 (C_4 \lambda_1^{-k+1}\delta)^\alpha)} \nonumber \\ && \cdots \frac{(1 + C_3 (C_4 \lambda_1^{-1}\delta)^\alpha)}{(1 - C_3 (C_4 \lambda_1^{-1}\delta)^\alpha)}\mu_u(B^k_k) \nonumber \\ &<& \eta_k \mu_u(B^k_k) \prod_{i=1}^\infty (1+ C_3(C_4\lambda_1^{-i}\delta)^\alpha )/ \prod_{i=1}^\infty (1- C_3(C_4\lambda_1^{-i}\delta)^\alpha ) \nonumber\end{eqnarray} Since $\lambda > 1$, then $\lambda^\alpha >1$, the infinite products converge. We have $$\mu_u(D_k) < C_5 \eta_k \mu_u(B^k_k).$$ for some constant $C_5 >0$. Choose a positive integer $k_0$ such that for $k \geq k_0$, $\eta_k < \epsilon /C_5$, then we have $$\mu_u(W^u_\delta(f^k(x)) \cap A) \geq (1-\epsilon) \mu_u(W^u_\delta(f^k(x))),$$ for all $k \geq k_0$. This proves the statement of the lemma on the unstable foliation. The part on the stable foliation can be proved in the same way by considering $f^{-1}$. This proves the lemma. \vs{1ex} We return to the proof of the theorem. Let $E$ be the closure of $A$ in $M$. We claim that if $y \in E$, then $W^s_\delta(y) \subset E$ and $W^u_\delta(y) \subset E$. Suppose that this is not true. i.e., there is a point $z \in W^u(y)$ and a small $\epsilon_1$-ball around $z$, $B(z, \epsilon_1) \subset M$ such that $B(z, \epsilon_1) \cap A = \emptyset$. Consequently, there are constants $\epsilon_2>0$, depending on $\epsilon_1$, and $C_6 >0$, independent of $\epsilon_1$, such that if $x \in \Lambda$, $|x-y| \leq \epsilon_2$, then $$\mu_u(W^u_\delta(f^k(x)) \cap A) < (1- C_6 \epsilon^{n_u}_1) \mu_u(W^u_\delta(f^k(x))),$$ for all $k \in \mathbb{N}$. On the other hand, since $y \in E$, there is a sequence of points $x_i \in A$, $i \in \mathbb{N}$ such that $x_i \rightarrow y$ and $i \rightarrow \infty$. For the above $\epsilon_1$, there is positive integer $i_0$ such that if $i \geq i_0$, $d(x_i, y) \leq \epsilon_1/3$. For any fixed $\epsilon >0$, by the lemma above and the recurrence of $x_i$, there is a positive integer $k >0$ such that $d(x_i, f^k(x_i)) \leq \epsilon_1 /3$ and $$\mu_u(W^u_\delta(f^k(x_i)) \cap A) \geq (1-\epsilon) \mu_u(W^u_\delta(f^k(x_i))).$$ Since $d(y, f^k(x_i)) \leq \frac{2\epsilon_2}{3} < \epsilon_2$, choosing $\epsilon = C_6 \epsilon^{n_u}_1$ leads to a contradiction. This contradiction show that if $y \in E$, then $W^u_\delta(y) \subset E$. Similarly by considering $f^{-1}$, we have $W^s_\delta(y) \subset E$. To conclude our proof, for any $y \in E$, $W^u_\delta(y) \subset E$ and $$V = \bigcup_{z \in W^u_\delta(y)} W^s_\delta(z) \subset E.$$ Since $V$ is hyperbolic, $y$ is in the interior of $V$. This implies that $E$ is an open set. But $V$ is also closed and non-empty. The connectness of $M$ implies that $E=M$. This implies that $f$ is Anosov. Finally, the reason that $\Lambda=M$ in the first place is that $f$ is ergodic, any invariant set with positive measure must have full measure and its closure must be the whole manifold. This proves the theorem. \section{Ergodicity of volume preserving Anosov diffeomorphisms} In this section, we give a direct proof of the ergodicity of $C^{1 + \alpha}$ volume-preserving Anosov diffeomorphisms, without using the usual Hopf arguments or the Birkhoff ergodic theorem. We first prove the following lemma. \begin{lem} Let $f \in \mbox{{\rm Diff}$_\mu^r(M)$}$, $r>1$, be a volume preserving diffeomorphism on a compact manifold $M$. Let $\Lambda \subset M$ be a hyperbolic invariant set (not neccessarily closed). If $\mu(\Lambda) >0$, then for a.e. $x \in \Lambda$, $W^s(x) \subset \Lambda$ modulus a $\mu_s$-measure zero set in $W^s(x)$ and $W^u(x) \subset \Lambda$ modulus a $\mu_u$-measure zero set in $W^u(x)$, where $\mu_s$ and $\mu_u$ are respectively the induced measures of $\mu$ on the stable and unstable foliations. \label{lemf} \end{lem} \noindent {\it Proof}: For any fixed positive integer $k$ and positive number $\eta>0$, let $\Lambda_{(\eta, k)} \subset \Lambda$ be the set such that $$\frac{\mu_u(B^u_i(x) \cap \Lambda)}{\mu_u(B^u_i(x))} > (1 -\eta), \; \mbox{ for all } i \geq k$$ Since almost every point of $\Lambda$ is a density point of $\Lambda$, for any $\eta>0$, $$\lim_{k\rightarrow \infty} \mu(\Lambda_{(\eta, k)}) = \mu(\Lambda).$$ By Poincar\'e recurrence theorem, for a.e. $x \in \Lambda_{(\eta, k)}$, there exists a sequence of integers $n_i \rightarrow \infty$ such that $f^{-n_i}(x) \in \Lambda_{(\eta, k)}$. By the distortion estimates from the last section, $$\mu_u(W^u_\delta(x) \cap \Lambda) \geq (1-C_5\eta) \mu_u(W^u_\delta(x)),$$ where $C_5$ and $\delta$ are fixed constants from the last section. Since the above estimate is independent of $k$, it must hold for almost all $x \in \cup_{k=1}^\infty \Lambda_{(\eta, k)}$. Since $\mu(\cup_{k=1}^\infty \Lambda_{(\eta, k)}) = \mu(\Lambda)$ for any fixed $\eta >0$, this implies that for a.e. $x \in \Lambda$, $$\mu_u(W^u_\delta(x) \cap \Lambda) \geq (1-C_5\eta) \mu_u(W^u_\delta(x)).$$ This is true for all $\eta >0$, therefore, for a.e. $x \in \Lambda$, $$\mu_u(W^u_\delta(x) \cap \Lambda) = \mu_u(W^u_\delta(x)).$$ This proves the lemma for the unstable foliations. The results on the stable foliation can be proved in the same way by considering $f^{-1}$. This proves the lemma. \vs{1ex} Now we can prove the following ergodicity theorem. \begin{thm} Let $f \in \mbox{{\rm Diff}$_\mu^r(M)$}$, $r>1$, be an Anosov volume preserving diffeomorphism on a compact manifold $M$. Then $f$ is ergodic. \end{thm} \noindent {\it Proof}: Let $\Lambda \subset M$ be an invariant set and $\mu(\Lambda) >0$. Let $\Lambda_s \subset \Lambda$ be the set such that for each $x \in \Lambda_s$, a.e. ($\mu_s$) $y$ in $W^s_\delta(x)$ is in $\Lambda$. The above lemma shows that $\mu(\Lambda_s) = \mu (\Lambda)$. Also by the above lemma, for a.e. $z \in \Lambda$, a.e. ($\mu_u$) $x$ in $W^u_\delta(z)$ is in $\Lambda_s$. By the absolute continuity of the stable and unstable foliations, the set $$\bigcup_{y \in W^u_\delta(z)} W^s_\delta(y)$$ has the full measure in the $\delta$ neighborhood of $z$ and therefore $\Lambda$ has the full measure in a $\delta$ neighborhood of $z$ for a.e. $z \in \Lambda$. This implies that $\Lambda$ has the full measure in $M$. Since $\Lambda$ is an arbitrary positive measure set, this implies ergodicity. This proves the theorem. Finally we remark that Lemma \ref{lemf} is also true for partially hyperbolic invariant and non-uniformly hyperbolic invariant sets (sets with non-zero Liapunov exponents). The proof is exactly the same.
{ "timestamp": "2005-08-26T22:13:29", "yymm": "0503", "arxiv_id": "math/0503437", "language": "en", "url": "https://arxiv.org/abs/math/0503437" }
\chapter{Analysis Methods for pions, kaons, proton and anti-proton} \chapter{Analysis Methods} \label{chp:analysis} \section{Trigger} The detector used for these studies was the Solenoidal Tracker at RHIC (STAR). The main tracking device is the Time Projection Chamber (TPC) which provides momentum information and particle identification for charged particles up to $p_{T}\sim1.1$ GeV/c by measuring their ionization energy loss ({\it dE/dx})~\cite{tpc}. Detailed descriptions of the TPC and d+Au run conditions have been presented in Ref.~\cite{stardau,tpc}. A prototype time-of-flight detector (TOFr) based on multi-gap resistive plate chambers (MRPC)~\cite{startof} was installed in STAR for the d+Au and p+p runs. It extends particle identification up to $p_{T}\sim3$ GeV/c for $p$ and $\bar{p}$. TOFr covers $\pi/30$ in azimuth and $-1\!<\!\eta\!<\!0$ in pseudorapidity at a radius of $\sim220$ cm. It contains 28 MRPC modules which were partially instrumented during the 2003 run. Since the acceptance of TOFr is small, a special trigger selected events with a valid pVPD coincidence and at least one TOFr hit. A total of 1.89 million and 1.08 million events were used for the analysis from TOFr triggered d+Au and non-singly diffractive (NSD) p+p collisions, representing an integrated luminosity of about 40 $\mathrm{{\mu}b}^{-1}$ and 30 $\mathrm{nb}^{-1}$, respectively. Minimum-bias d+Au and p+p collisions that did not require pVPD and TOFr hits were also used to study the trigger bias and enhancement, and the TOFr efficiency and acceptance. The d+Au minimum-bias trigger required an equivalent energy deposition of about 15 GeV in the Zero Degree Calorimeter in the Au beam direction~\cite{stardau}. The trigger efficiency was determined to be $95\pm3\%$. Minimum-bias p+p events were triggered by the coincidence of two beam-beam counters (BBC) covering $3.3< |\eta|<5.0$~\cite{starhighpt}. The NSD cross section was measured to be $30.0\pm3.5$ mb by a van der Meer scan and PYTHIA~\cite{pythia} simulation of the BBC acceptance~\cite{starhighpt}. \subsection{Centrality tagging} Centrality tagging of d+Au collisions was based on the charged particle multiplicity in $-3.8<\eta<-2.8$, measured by the Forward Time Projection Chamber in the Au beam direction~\cite{stardau,ftpc}. The TOFr triggered d+Au events were divided into three centralities: most central $20\%$, $20-40\%$ and $40-\sim100\%$ of the hadronic cross section. The average number of binary collisions $\langle N_{bin}\rangle$ for each centrality class and for the combined minimum-bias event sample is derived from Glauber model calculations and listed in Table~\ref{centrality}. Table~\ref{centrality} also lists the uncorrected FTPC east reference multiplicity ranges for centrality definitions. \begin{table}[h] \centering \begin{tabular}{|c|c|c|c|} \hline Centrality Bin & Uncorr. FTPCRefMult Range & Uncorr. $N_{charge}$ & $N_{bin}$ \\ \hline M.B. & & 10.2 & $7.5\pm0.4$ \\ \hline 0\%-20\% & FTPCRefMult $\geq$ 17 & 17.58 & $15.0\pm1.1$ \\ \hline 20\%-40\% & 10 $\leq$ FTPCRefMult $<$ 17 & 12.55 & $10.2\pm1.0$ \\ \hline 40\%-100\% & 0 $\leq$ FTPCRefMult $<$ 10 & 6.17 & $4.0\pm0.3$ \\ \hline \end{tabular} \caption{Centrality definitions for different uncorrected FTPC east reference multiplicity ranges. Uncorrected $N_{charge}$ stands for the average value of uncorrected reference multiplicity in certain centrality bin. The fourth column represents the number of binary collisons $\langle N_{bin}\rangle$ calculated from Glauber model.} \label{centrality} \end{table} \subsection{Trigger bias study} Since we set up a special trigger which selected events with a valid pVPD coincidence and at least one TOFr hit, the study of $p_{T}$ dependence of trigger bias is necessary. Figure~\ref{PtRatioRealTOFAcceptance} shows there is negligible trigger bias on $p_{T}$ dependence at $p_{T}>$ 0.3 GeV/c from simulation. In this figure, pVPD means that pVPD is required to fire in minimum-bias collisions. TOF means that TOFr is required to fire in minimum-bias collisions, and pVPD $\&$ TOF means that pVPD and TOFr are required to fire in minimum-bias collisions. From this figure, if we required pVPD and TOFr to fire, we can see the ratio is flat with $p_{T}$ when $p_{T}$ is larger than 0.3 GeV/c by comparison through the $p_{T}$ distribution in minimum-bias collisions. That means the trigger bias for $p_T$ distribution is negligible at $p_{T}>$ 0.3 GeV/c. \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{PtRatioRealTOFAcceptanceMod.eps} \caption{The $p_{T}$ dependence plot of the trigger bias.} \label{PtRatioRealTOFAcceptance} \end{figure} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{NchbiasNew.eps} \caption{The enhancement factor and $\langle N_{ch}\rangle$ bias in minimum-bias and centrality selected d+Au collisions. } \label{NchbiasNew} \end{figure} Minimum-bias d+Au and p+p collisions are used to study the trigger bias and enhancement. Figure~\ref{NchbiasNew} shows the trigger bias and enhancement in d+Au minimum-bias collisions and three centrality bins. In this figure, TOFr means that TOFr is required to fire in minimum-bias events. pVPD means that TOFr and pVPD are required to fire in minimum-bias events. Minbias means the minimum-bias triggered events. For enhancement study, TOFr/pVPD is the ratio of the number of events in which TOFr is required to fire over the number of events in which TOFr and pVPD are required to fire, and Minbias/pVPD is the ratio of the number of minimum-bias triggered events over the number of events in which TOFr and pVPD are required to fire. The enhancement factor for TOFr is (Minbias/pVPD)/(TOFr/pVPD). For example, in minimum-bias collision, Minbias/pVPD is equal to 28.7, while TOFr/pVPD is 2.87, so in minimum-bias collisions, the enhancement of TOFr trigger is 10. For $\langle N_{ch}\rangle$ bias study, TOFr/pVPD is the ratio of $\langle N_{ch}\rangle$ in the events where TOFr is required to fire over the $\langle N_{ch}\rangle$ in the events where TOFr and pVPD are required to fire. Since in our triggered events, TOFr and pVPD are required to fire, TOFr/pVPD is our $\langle N_{ch}\rangle$ bias factor. The curves in this figure show the charged particle multiplicity at mid-rapidity in TOFr events and in TOFr and pVPD events individually. Table~\ref{triggerbiastable} lists the enhancement factor and trigger bias in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions. \begin{table}[h] \begin{tabular}{|c|c|c|c|} \hline Centrality Bin & TOFr triggered events & enhancement factor & $\langle N_{ch}\rangle$ bias \\ \hline 0\%-100\% & 1.80 M & 10.0 & 1.02 \\ \hline 0\%-20\% & 0.523 M & 5.75 & 1.04 \\ \hline 20\%-40\% & 0.500 M & 8.03 & 1.03 \\ \hline 40\%-100\% & 0.479 M & 15.8 & 0.965 \\ \hline p+p & 0.995 M& 37.4 & 1.19 \\ \hline \end{tabular} \caption{Trigger bias study. The $\langle N_{ch}\rangle$ bias and enhancement factor in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{triggerbiastable} \end{table} \section{Track selection and calibration} The TPC and TOFr are two independent systems. In the analysis, hits from particles traversing the TPC were reconstructed as tracks with well defined geometry, momentum, and {\it dE/dx} ~\cite{tpc}. The particle trajectory was then extended outward to the TOFr detector plane. The pad with the largest signal within one pad distance to the projected point was associated with the track for further time-of-flight and velocity ($\beta$) calculations. \subsection{Calibration} \subsubsection{pVPD calibration} For TOFr, we use pVPD as our start-timing detector. In d+Au and p+p collisions, at least one east pVPD and one west pVPD were required to fire. In d+Au collisions, to calibrate east pVPD, we required 3 east pVPD to fire; to calibrate west pVPD, we required 3 west pVPD to fire. In p+p collisions, to calibrate east pVPD, we required 2 east pVPD to fire; to calibrate west pVPD, we required 2 west pVPD to fire. Let's take the east pVPD calibration in d+Au collisions as an example. The label for 3 pVPD are pVPD1, pVPD2, pVPD3, the adc and tdc value for pVPD1 are $a1$, $t1$, and the slewing correction function is $f1$; the adc and tdc value for pVPD2 are $a2$, $t2$, and the slewing correction function is $f2$; the adc and tdc value for pVPD3 are $a3$, $t3$, and the slewing correction function is $f3$. We use $t1-((t2-f2)+(t3-f3))/2$ vs $a1$ to get the slewing correction for pVPD1; use $t2-((t3-f3)+(t1-f1))/2$ vs $a2$ to get the slewing correction for pVPD2; use $t3-((t1-f1)+(t2-f2))/2$ vs $a3$ to get the slewing correction for pVPD3. At the beginning, $f1=f2=f3=0$, we got 3 curves of $t1-((t2-f2)+(t3-f3))/2$ vs $a1$, $t2-((t3-f3)+(t1-f1))/2$ vs $a2$ and $t3-((t1-f1)+(t2-f2))/2$ vs $a3$. The 3 curves corresponded to the 3 slewing functions $f(a1), f(a2), f(a3)$; For the second step, $f1=f(a1), f2=f(a2), f3=f(a3)$, also plot $t1-((t2-f2)+(t3-f3))/2$ vs $a1$, $t2-((t3-f3)+(t1-f1))/2$ vs $a2$ and $t3-((t1-f1)+(t2-f2))/2$ vs $a3$. And we got the new three slewing curves $f'(a1), f'(a2), f'(a3)$. For the third step, $f1=f'(a1), f2=f'(a2), f3=f'(a3)$, also plot $t1-((t2-f2)+(t3-f3))/2$ vs $a1$, $t2-((t3-f3)+(t1-f1))/2$ vs $a2$ and $t3-((t1-f1)+(t2-f2))/2$ vs $a3$. And we got another new three slewing curves $f''(a1), f''(a2), f''(a3)$. And so on and so forth till the resolution of $t1-f1-((t2-f2)+(t3-f3))/2, t2-f2-((t3-f3)+(t1-f1))/2$ and $t3-f3-((t1-f1)+(t2-f2))/2$ converged. The looping method is to subtract the correlation of different pVPD tubes in the same direction. The function for the slewing correction we use is $y=par[0]+par[1]/\sqrt{x}+par[2]/x+par[3]\times{x}$. In Figure~\ref{pvpdslewingplotforthesis}, the left plot shows the pVPD2 slewing plot and the right plot shows that the timing is independent on the ADC value after the slewing correction. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{pvpdslewingplot.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{pvpdslewingplot1.eps} \end{minipage} \caption{pVPD slewing correction.} \label{pvpdslewingplotforthesis} \end{figure} After the slewing correction, we got the corrected timing of east pVPD and west pVPD. For each side, the timing difference should be shifted to zero. That's to say the mean value in the distribution of $t1-f1-(t2-f2)$ and $t1-f1-(t3-f3)$ were shifted to zero. Also we need to correct for the effect caused by the different numbers of fired pVPD in different events. What we did was shifting the mean value of the distribution of ($\sum{te})/Ne-(\sum{tw})/Nw-2.\times{Vz/c}$ to zero, where the $\sum{te}$, $\sum{tw}$ means the sum of the corrected timing of east fired pVPD and west fired pVPD respectively, $Ne, Nw$ means the number of east fired pVPD and west fired pVPD, $Vz$ is the $z$ value of primary vertex of the event, and $c$ is the light velocity. \subsubsection{TOFr calibration} After the slewing correction for pVPD, we use this variable as our start timing: \begin{equation} T_{start}=\frac{{\sum_{i=1}^{Ne}{te}}+{\sum_{i=1}^{Nw}{tw}}-(Ne-Nw)\times{Vz}/c}{Ne+Nw} \end{equation} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{dedxplottmp.eps} \caption{dE/dx vs $p$ plot from d+Au collisions. The line represents that $dE/dx=0.028\times{10^{-4}}$ GeV/cm in this momentum range $0.3<p<0.6$ GeV/c.} \label{dAudedxplot} \end{figure} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{slewingplot.eps} \caption{The slewing correction.} \label{slewingplot} \end{figure} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{ZFit_forthesis.eps} \caption{The z position correction.} \label{ZFit_forthesis} \end{figure} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{dAutiming.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{pptiming.eps} \end{minipage} \caption{The overall timing resolution after the calibration.} \label{timeresolution} \end{figure} The difference between TOFr timing $T_{tofr}$ and start timing $T_{start}$ is our time of flight $tof=T_{tofr}-T_{start}$. To calibrate the $tof$, the pure pion sample was chosen by selecting the particle energy loss $dE/dx$ in TPC at $dE/dx<0.028\times{10^{-4}}$ GeV/cm in the momentum range $0.3<p<0.6$ GeV/c. Figure~\ref{dAudedxplot} shows dE/dx vs $p$ plot from d+Au collisions. Firstly the so called $T_{0}$ correction was done due to the different cable lengths for different read-out channels, which was done by shifting the mean value of the distribution of $tof-T_{\pi}$ to zero channel by channel, where $T_{\pi}$ is the calculation timing assuming the particle was pion particle. Secondly, the slewing correction due to correlation between timing and signal amplitude of the electronics was done by getting the curve of $tof'-T_{\pi}$ vs $adc$ for each channel, where the $tof'$ was the time of flight after the $T_{0}$ correction and $adc$ was the ADC value of TOFr. The slewing curve is like the plot shown in Figure~\ref{slewingplot}. The function of the slewing correction is $y=par[0]+par[1]/\sqrt{x}+par[2]/x+par[3]/\sqrt{x}/x+par[4]/x/x$. The z position correction was also done since the different hit positions on the read-out strip will generate different transmission timing. This was done by getting the function of $tof''-T_{\pi}$ versus $Z_{local}$, where the $tof''$ is the time of flight after the $T_{0}$ and slewing correction, and $Z_{local}$ is the the hit local z position of the TOFr. The function for the z position correction is $y=\sum_{i=0}^{7}{(par[i]\times{x^{i}})}$. The z position correction for all the channels is shown in Figure~\ref{ZFit_forthesis}. After the z position was done, the calibration for TOFr was finished. The overall resolution of TOFr was 120 ps and 160 ps in d+Au and p+p collisions respectively, where the effective timing resolution of the pVPDs was 85 ps and 140 ps, respectively. Figure~\ref{timeresolution} shows the overall resolution of TOFr in d+Au and p+p collisions. \section{Raw yield} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{tofr_beta_p_prplot0910.eps} \caption{$1/\beta$ vs. momentum for $\pi^{\pm}$, $K^{\pm}$, and $p(\bar{p})$ from 200 GeV d+Au collisions. Separations between pions and kaons, kaons and protons are achieved up to $p_{T}\simeq1.6$ and $3.0$ GeV/c, respectively. The insert shows $m^{2}=p^{2}(1/\beta^{2}-1)$ for $1.2<p_{T}<1.4$ GeV/c. Clear separation of $\pi$, $K$ and $p$ is seen.} \label{beta} \end{figure} From the timing information $t$ from TOFr after the calibration and the pathlength $L$ from TPC, the velocity $\beta$ of the particle can be easily got by $\beta=L/t/c$. Figure~\ref{beta} shows $1/\beta$ from TOFr measurement as a function of momentum ($p$) calculated from TPC tracking in TOFr triggered d+Au collisions. The raw yields of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ are obtained from Gaussian fits to the distributions in $m^{2}=p^{2}(1/\beta^{2}-1)$ in each $p_{T}$ bin. \subsection{$\pi$ raw yield extraction} For $\pi^{\pm}$, the rapidity range is $-0.5<y_{\pi}<0.$. After $|N_{\sigma\pi}|<2$ was required, the mass squared $m^{2}=p^{2}(1/\beta^{2}-1)$ distributions in different $p_{T}$ bin in d+Au minimum-bias collisions are shown is Figure~\ref{pionplusrawyieldplot} and Figure~\ref{pionminusrawyieldplot}. At $p_{T}<0.8$ GeV/c, the single Gaussian function was used to fit the distribution of $m^{2}$ to get the raw yield. At the same time, the counting result by counting the track number at the range $-0.1<m^{2}<0.1$ $(GeV/c^{2})^2$ was also used to compare with the raw yield from the fitting method. The difference between them was found in one sigma range. The raw yield we quote is from the fitting method. At $p_{T}>0.8$ GeV/c, the double Gaussian function was used to extract the raw yield. The raw signals in each $P_{T}$ bin are shown in Table~\ref{pionplustable} and Table~\ref{pionminustable}. Also shown in the tables are those in centrality selected d+Au collisions and minimum-bias p+p collisions. \subsection{$K$ raw yield extraction} For $K^{\pm}$, the rapidity range is $-0.5<y_{K}<0$. After $|N_{\sigma K}|<2$ was required, the mass squared $m^{2}=p^{2}(1/\beta^{2}-1)$ distributions in different $p_{T}$ bin in d+Au minimum-bias collisions are shown is Figure~\ref{kaonplusrawyieldplot} and Figure~\ref{kaonminusrawyieldplot}. At $p_{T}<0.8$ GeV/c, the single Gaussian function was used to fit the distribution of $m^{2}$ to get the raw yield. At the same time, the counting result by counting the track number at the range $0.16<m^{2}<0.36$ $(GeV/c^{2})^2$ was also used to compare with the raw yield from the fitting method. The difference between them was found in one sigma range. The raw yield we quote is from the fitting method. At $p_{T}>0.8$ GeV/c, the double Gaussian function was used to extract the raw yield. The raw signals in each $P_{T}$ bin are shown in Table~\ref{kaonplustable} and Table~\ref{kaonminustable}. Also shown in the tables are those in centrality selected d+Au collisions and minimum-bias p+p collisions. \subsection{$p$ and $\bar{p}$ raw yield extraction} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{pbardca1.0.eps} \caption{the ratio of $\bar{p}$ at $dca<1.0$ cm over $\bar{p}$ at $dca<3.0$ cm.} \label{pbardcaratio} \end{figure} For $\bar{p}$, the rapidity range is $-0.5<y_{\bar{p}}<0$. After $|N_{\sigma p}|<2$ was required, the mass squared $m^{2}=p^{2}(1/\beta^{2}-1)$ distributions in different $p_{T}$ bin in d+Au minimum-bias collisions are shown is Figure~\ref{pbarrawyieldplot}. At $p_{T}<1.6$ GeV/c, the single Gaussian function was used to fit the distribution of $m^{2}$ to get the raw yield. At the same time, the counting result by counting the track number at the range $0.64<m^{2}<1.44$ $(GeV/c^{2})^2$ was also used to compare with the raw yield from the fitting method. The difference between them was found in one sigma range. The raw yield we quote is from the fitting method. At $p_{T}>1.6$ GeV/c, the double Gaussian function was used to extract the raw yield. The raw signals in each $P_{T}$ bin are shown in Table~\ref{pbartable}. For the $p$, the raw yield extraction method is the same as $\bar{p}$ except that at $p_{T}<1.6$ GeV/c, we use the method $Np=Np_{dca<1.cm}\times{(N\bar{p}_{dca<3.cm}/N\bar{p}_{dca<1.cm}) }$ to reject the background, where $Np$ and $N\bar{p}$ are the number of the $p$ and $\bar{p}$ tracks individually, and $N\bar{p}_{dca<1.cm}/N\bar{p}_{dca<3.cm}$ is the ratio of $\bar{p}$ tracks at $dca<1.0$ cm over those at $dca<3.0$ cm. In Figure~\ref{protonrawyieldplot}, the first 10 $p_{T}$ bins are for $dca<1.0$ cm, the last 4 $p_{T}$ bins are for $dca<3.0$ cm. Figure~\ref{pbardcaratio} shows the ratio of $\bar{p}$ tracks at $dca<1.0$ cm over those at $dca<3.0$ cm. After this correction of $Np=Np_{dca<1.cm}\times{(N\bar{p}_{dca<3.cm}/N\bar{p}_{dca<1.cm}) }$, the $p$ raw signals in each $P_{T}$ bin are shown in Table~\ref{protontable}. \section{Efficiency and acceptance correction} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{pluseff.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{minuseff.eps} \end{minipage} \caption{TPC reconstruction efficiency of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ as a function of $p_{T}$. The left plot for charged plus particle and the right for charged minus particle. } \label{tpceff} \end{figure} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{plusMatchingEff.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{minusMatchingEff.eps} \end{minipage} \caption{Matching efficiency from TOFr to TPC of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ as a function of $p_{T}$, including detector response. The left plot for charged plus particle and the right for charged minus particle.} \label{matcheff} \end{figure} Acceptance and efficiency were studied by Monte Carlo simulations and by matching TPC track and TOFr hits in real data. TPC tracking efficiency was studied by Monte Carlo simulations. The simulated $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ are generated using a flat $p_T$ and a flat $y$ distribution and pass through GSTAR~\cite{long:01} (the framework software package to run the STAR detector simulation using GEANT~\cite{geant:01,geant:02}) and TRS (the TPC Response Simulator~\cite{long:01}). The simulated $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ are then combined with a real raw event and we call this combined event a simulated event. This simulated event is then passed through the standard STAR reconstruction chain and we call this event after reconstruction a reconstructed event. The reconstructed information of those particles in the reconstructed event is then associated with the Monte-Carlo information in the simulated event. And then we get the total number of simulated $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ from simulated events in a certain transverse momentum bin. Also we can get the total number of associated tracks in the reconstructed events in this transverse momentum bin~\cite{Haibin:03}. In the end, take the ratio of the number of associated $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ over the number of simulated $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ and this ratio is the TPC reconstruction efficiency for a certain transverse momentum bin in the mid-rapidity range. Figure~\ref{tpceff} shows the TPC reconstruction efficiency of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ as a function of $p_{T}$. \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{tofr_detecting_efficiency.eps} \caption{The TOFr response efficiency as a function of $p_{T}$.} \label{detectorresponse} \end{figure} The Matching Efficiency from TPC to TOFr were studied in real data, and the formula are \begin{equation} Eff_{Match}=\frac{TofrMatchedTracks/dAuTOFrEvents} {(MinBiasTracks/MinBiasEvents)_{pVPD}\times{factor1}\times{factor2}} \end{equation} where the $TofrMatchedTracks/dAuTOFrEvents$ is the number of TOFr matched tracks per dAuTOFr trigger event, $(MinBiasTracks/MinBiasEvents)_{pVPD}$ is the number of minimum-bias tracks per minimum-bias event by requiring the pVPD to fire, $factor1$ is the enhancement factor of dAuTOFr trigger, and $factor2$ is the other factors such as the TOFr trip factor. The $Eff_{Match}$ includes the detector response efficiency. Figure~\ref{matcheff} shows the matching efficiency of different particle species including the detector response versus $p_{T}$. The detector response efficiency, including the material absorption and scattering effect between TPC and TOFr, as a function of $p_{T}$ is shown in Figure~\ref{detectorresponse}, which is around 90\% at $p_{T}>$ 0.3 GeV/c. After the material absorption and scattering effect correction, the detector response efficiency is around 95\%. \section{Background correction} \begin{figure}[h] \centering \includegraphics[height=18pc,width=18pc]{pionbackgroundMod.eps} \caption{$\pi$ background contribution as a function of $p_{T}$. The circled symbols represent the total $\pi$ background contribution including feed-down and $\mu$ misidentification. The squared and triangled symbols represent the week-decay and $\mu$ misidentification contributions individually.} \label{pionbackground} \end{figure} Weak-decay feeddown (e.g. $K_{s}^{0}\rightarrow\pi^{+}\pi^{-}$) to pions is $\sim12\%$ at low $p_{T}$ and $\sim5\%$ at high $p_{T}$, and was corrected for using PYTHIA~\cite{pythia} and HIJING~\cite{hijing} simulations, as shown in Figure~\ref{pionbackground}. For $\pi$ spectra, the $\mu$ misidentification was also corrected for, which is also shown in Figure~\ref{pionbackground}. \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{protonScatter_dca1.0.eps} \caption{The $p$ scattering effect contribution when we cut $dca<1.0$ cm.} \label{protonScatter} \end{figure} Inclusive $p$ and $\bar{p}$ production is presented without hyperon feeddown correction. $p$ and $\bar{p}$ from hyperon decays have the same detection efficiency as primary $p$ and $\bar{p}$~\cite{antiproton} and contribute about 20\% to the inclusive $p$ and $\bar{p}$ yield, as estimated from the simulation. However, for $p$, there is still some scattering contribution which comes from the beam pipe interaction after the cut of $dca<1.0$ cm. Figure~\ref{protonScatter} shows the contribution of scattering effect for proton when we cut $dca<1.0$ cm. The correction is done at $p_{T}<$ 1.1 GeV/c and negligible at higher $p_{T}$. \section{Energy loss correction} The energy loss effect due to the interaction with the detector material was also corrected for. This was studied by simulation. Figure~\ref{eloss} shows the momentum and transverse momentum correction for energy loss effect. At $p_{T}>$0.35 GeV/c, for $\pi$, the energy loss effect is negligible while for kaon and proton, the energy loss correction is non-negligible at lower $p_{T}$ and negligible at higher $p_{T}$. The correction was done by shifting the position of $p_{T}$ in the $p_{T}$ spectra. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{ptotthetadiff.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=18pc]{ptdiff.eps} \end{minipage} \caption{(left) p energy loss correction of different particle species as a function of p. $p_{rec}$ is the reconstructed momentum before the energy loss correction, $p_{MC}$ is the momentum after energy loss correction from simulation, $\theta$ is the angle between the reconstructed momentum and beam line. (right) $p_{T}$ energy loss correction of different particle species as a function of $p_{T}$. $p_{T}(rec)$ is the reconstructed transverse momentum before the energy loss correction, $p_{T}(MC)$ is the transverse momentum after energy loss correction from simulation. } \label{eloss} \end{figure} \section{Normalization} The efficiency including vertex efficiency and trigger efficiency is 91\% in d+Au minimum-bias collisions and 85\% in p+p and 40-100\% d+Au collisions. In 0\%-20\% and 20\%-40\% d+Au collisions, the efficiency is 100\%. Since the statistic of p+p minimum-bias events in run 3 is not good enough for us to get very precise enhancement factor and $N_{ch}$ bias factor. We compare the $\pi$ spectra in the first 5 $p_{T}$ bin with those from the paper~\cite{olga} and get the additional normalization factor for p+p collisions. \begin{figure}[h] \hspace{-3pc} \includegraphics[height=40pc,width=40pc]{pionplusrawyield_forthesis.eps} \caption{$\pi^{+}$ raw yields versus mass squared distribution. The histograms are our data. The curves are Gaussian fits.} \label{pionplusrawyieldplot} \end{figure} \begin{figure}[h] \hspace{-3pc} \includegraphics[height=40pc,width=40pc]{pionminusrawyield_forthesis.eps} \caption{$\pi^{-}$ raw yields versus mass squared distribution. The histograms are our data. The curves are Gaussian fits.} \label{pionminusrawyieldplot} \end{figure} \begin{figure}[h] \hspace{-3pc} \includegraphics[height=40pc,width=40pc]{kaonplusrawyield_forthesis.eps} \caption{$K^{+}$ raw yields versus mass squared distribution. The histograms are our data. The curves are Gaussian fits.} \label{kaonplusrawyieldplot} \end{figure} \begin{figure}[h] \hspace{-3pc} \includegraphics[height=40pc,width=40pc]{kaonminusrawyield_forthesis.eps} \caption{$K^{-}$ raw yields versus mass squared distribution. The histograms are our data. The curves are Gaussian fits.} \label{kaonminusrawyieldplot} \end{figure} \begin{figure}[h] \hspace{-3pc} \includegraphics[height=40pc,width=40pc]{protonrawyield_forthesis.eps} \caption{$p$ raw yields versus mass squared distribution. The histograms are our data. The curves are Gaussian fits.} \label{protonrawyieldplot} \end{figure} \begin{figure}[h] \hspace{-3pc} \includegraphics[height=40pc,width=40pc]{pbarrawyield_forthesis.eps} \caption{$\bar{p}$ raw yields versus mass squared distribution. The histograms are our data. The curves are Gaussian fits.} \label{pbarrawyieldplot} \end{figure} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|} \hline $p_{T}$ (GeV/c) & d+Au Trigger & 0\%-20\% & 20\%-40\% & 40\%-100\% & p+p \\ \hline 0.3-0.4 & $2.929e+04\pm171.4$ & $9219\pm96.21$ & $8735\pm93.6$ & $8604\pm92.84$ & $1.806e+04\pm134.4$ \\ \hline 0.4-0.5 & $2.185e+04\pm147.8$ & $6894\pm83.03$ & $6657\pm81.59$ & $6325\pm79.53$ & $1.274e+04\pm114.3$ \\ \hline 0.5-0.6 & $1.592e+04\pm126.2$ & $5162\pm71.85$ & $4901\pm70.3$ & $4534\pm67.34$ & $9180\pm95.81$ \\ \hline 0.6-0.7 & $1.166e+04\pm108$ & $3832\pm62.19$ & $3556\pm59.64$ & $3311\pm57.54$ & $6531\pm80.82$ \\ \hline 0.7-0.8 & $8556\pm92.5$ & $2909\pm53.93$ & $2628\pm51.26$ & $2368\pm48.67$ & $4447\pm66.74$ \\ \hline 0.8-0.9 & $6198\pm78.86$ & $2099\pm45.85$ & $1936\pm44.33$ & $1693\pm41.17$ & $2973\pm54.57$ \\ \hline 0.9-1 & $4520\pm67.25$ & $1487\pm38.57$ & $1361\pm36.9$ & $1276\pm35.74$ & $2132\pm46.31$ \\ \hline 1-1.1 & $3312\pm57.61$ & $1147\pm33.9$ & $1033\pm32.17$ & $845.9\pm29.15$ & $1386\pm37.71$ \\ \hline 1.1-1.2 & $2406\pm49.35$ & $788.6\pm28.19$ & $752.5\pm27.58$ & $652.7\pm25.7$ & $959.2\pm31.93$ \\ \hline 1.2-1.4 & $3227\pm58.17$ & $1132\pm34.28$ & $934.4\pm30.98$ & $831.5\pm29.82$ & $1183\pm40.11$ \\ \hline 1.4-1.6 & $1756\pm45.2$ & $573.7\pm26.2$ & $543.7\pm24$ & $412.8\pm21.83$ & $625.5\pm30.16$ \\ \hline 1.6-1.8 & $1046\pm39$ & $337.9\pm20.42$ & $309.8\pm18.62$ & $234\pm16.58$ & $364.5\pm32.64$ \\ \hline \end{tabular} \caption{$\pi^{+}$ raw signal table in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{pionplustable} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|} \hline $p_{T}$ (GeV/c) & d+Au Trigger & 0\%-20\% & 20\%-40\% & 40\%-100\% & p+p \\ \hline 0.3-0.4 & $2.861e+04\pm169.4$ & $8922\pm94.66$ & $8507\pm92.38$ & $8519\pm92.4$ & $1.715e+04\pm131$ \\ \hline 0.4-0.5 & $2.139e+04\pm146.3$ & $6805\pm82.49$ & $6458\pm80.35$ & $6306\pm79.41$ & $1.28e+04\pm113.1$ \\ \hline 0.5-0.6 & $1.611e+04\pm126.9$ & $5327\pm72.98$ & $4873\pm69.8$ & $4605\pm67.86$ & $9189\pm95.86$ \\ \hline 0.6-0.7 & $1.166e+04\pm108$ & $3831\pm61.9$ & $3550\pm59.5$ & $3355\pm57.92$ & $6362\pm79.69$ \\ \hline 0.7-0.8 & $8447\pm91.91$ & $2837\pm53.64$ & $2540\pm50.4$ & $2387\pm48.86$ & $4154\pm64.5$ \\ \hline 0.8-0.9 & $5950\pm77.17$ & $2076\pm45.61$ & $1780\pm42.2$ & $1646\pm40.68$ & $2899\pm53.87$ \\ \hline 0.9-1 & $4284\pm65.46$ & $1446\pm38.03$ & $1317\pm36.31$ & $1171\pm34.29$ & $1924\pm44.01$ \\ \hline 1-1.1 & $3296\pm57.47$ & $1123\pm33.55$ & $1014\pm31.9$ & $897.5\pm30.02$ & $1372\pm37.77$ \\ \hline 1.1-1.2 & $2464\pm49.88$ & $812.4\pm28.58$ & $762.4\pm27.69$ & $650.8\pm25.72$ & $1005\pm33.16$ \\ \hline 1.2-1.4 & $3136\pm57.28$ & $1027\pm32.54$ & $972.9\pm31.71$ & $828.8\pm29.65$ & $1243\pm39.78$ \\ \hline 1.4-1.6 & $1716\pm45.79$ & $612.5\pm25.87$ & $539.3\pm25.67$ & $422.7\pm24.22$ & $603.8\pm30.49$ \\ \hline 1.6-1.8 & $1033\pm39.74$ & $375.3\pm21.21$ & $306.1\pm19.59$ & $239.9\pm48.55$ & $337.1\pm28.82$ \\ \hline \end{tabular} \caption{$\pi^{-}$ raw signal table in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{pionminustable} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|} \hline $p_{T}$ (GeV/c) & d+Au Trigger & 0\%-20\% & 20\%-40\% & 40\%-100\% & p+p \\ \hline 0.4-0.5 & $1410\pm37.54$ & $417.2\pm20.44$ & $420.9\pm20.52$ & $354.9\pm18.84$ & $753.2\pm27.44$ \\ \hline 0.5-0.6 & $1588\pm39.85$ & $486.3\pm22.06$ & $461\pm21.48$ & $435\pm20.87$ & $729.3\pm27$ \\ \hline 0.6-0.7 & $1499\pm38.71$ & $465.9\pm21.59$ & $445.5\pm21.17$ & $395\pm19.87$ & $710.1\pm26.65$ \\ \hline 0.7-0.8 & $1346\pm36.69$ & $423.9\pm20.59$ & $419.8\pm20.62$ & $335.7\pm18.32$ & $579\pm24.06$ \\ \hline 0.8-0.9 & $1105\pm33.59$ & $369.7\pm19.3$ & $317.9\pm18.43$ & $282.5\pm17.03$ & $496.2\pm22.44$ \\ \hline 0.9-1 & $969.1\pm31.19$ & $283.9\pm16.86$ & $305.1\pm17.52$ & $258.7\pm16.23$ & $381.2\pm19.87$ \\ \hline 1-1.1 & $799.3\pm28.41$ & $278.6\pm16.79$ & $224\pm15.04$ & $192.3\pm14.04$ & $301.5\pm18.39$ \\ \hline 1.1-1.2 & $656.7\pm26.24$ & $199.1\pm14.33$ & $186.2\pm14$ & $155.7\pm12.94$ & $267.8\pm18.71$ \\ \hline 1.2-1.4 & $1013\pm34.43$ & $335\pm19.51$ & $283.1\pm17.71$ & $234.1\pm17.68$ & $421.1\pm29.14$ \\ \hline 1.4-1.6 & $605.9\pm30.14$ & $191.1\pm17.77$ & $174.1\pm14.77$ & $148.8\pm15.48$ & $238.8\pm13.97$ \\ \hline 1.6-1.8 & $382.9\pm28.9$ & $---$ & $---$ & $---$ & $---$ \\ \hline \end{tabular} \caption{$K^{+}$ raw signal table in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{kaonplustable} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|} \hline $p_{T}$ (GeV/c) & d+Au Trigger & 0\%-20\% & 20\%-40\% & 40\%-100\% & p+p \\ \hline 0.4-0.5 & $1341\pm36.62$ & $411.6\pm20.29$ & $367.6\pm19.19$ & $378\pm19.44$ & $682.6\pm26.13$ \\ \hline 0.5-0.6 & $1498\pm38.7$ & $460\pm21.45$ & $411.5\pm20.32$ & $420.5\pm20.51$ & $740.7\pm27.21$ \\ \hline 0.6-0.7 & $1410\pm37.55$ & $436.2\pm20.89$ & $398.4\pm19.97$ & $361.9\pm19.02$ & $616.8\pm24.83$ \\ \hline 0.7-0.8 & $1207\pm34.74$ & $366\pm19.14$ & $349.7\pm18.7$ & $350\pm18.75$ & $557.4\pm23.61$ \\ \hline 0.8-0.9 & $1057\pm32.66$ & $317.4\pm18.12$ & $332.4\pm18.37$ & $268.2\pm16.66$ & $432.9\pm20.93$ \\ \hline 0.9-1 & $863.7\pm29.42$ & $256.9\pm16.09$ & $267.2\pm16.43$ & $223.8\pm15.03$ & $368.9\pm19.59$ \\ \hline 1-1.1 & $635.2\pm25.35$ & $198.2\pm14.35$ & $183.1\pm13.61$ & $187.9\pm13.88$ & $320.4\pm19.42$ \\ \hline 1.1-1.2 & $543\pm23.92$ & $166.4\pm13.14$ & $143\pm12.22$ & $154.1\pm12.89$ & $248.5\pm18.84$ \\ \hline 1.2-1.4 & $895\pm32.45$ & $302.1\pm18.3$ & $258\pm17.06$ & $206.7\pm16.17$ & $377.1\pm26.67$ \\ \hline 1.4-1.6 & $645.5\pm31.61$ & $202.4\pm16.78$ & $141\pm15.58$ & $166.6\pm17.87$ & $237.6\pm20.52$ \\ \hline 1.6-1.8 & $351.3\pm29.79$ & $---$ & $---$ & $---$ & $---$ \\ \hline \end{tabular} \caption{$K^{-}$ raw signal table in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{kaonminustable} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|} \hline $p_{T}$ (GeV/c) & d+Au Trigger & 0\%-20\% & 20\%-40\% & 40\%-100\% & p+p \\ \hline 0.4-0.5 & $1377\pm107.5$ & $412.6\pm40.73$ & $403.5\pm40.11$ & $422.7\pm41.42$ & $657.5\pm64.94$ \\ \hline 0.5-0.6 & $1527\pm89.71$ & $492.2\pm36.58$ & $428.6\pm33.09$ & $424.8\pm32.89$ & $752.6\pm59.91$ \\ \hline 0.6-0.7 & $1456\pm80.28$ & $437.9\pm31.47$ & $420\pm30.56$ & $401.3\pm29.6$ & $704.9\pm54.1$ \\ \hline 0.7-0.8 & $1336\pm73.68$ & $410.6\pm29.43$ & $387.9\pm28.28$ & $385\pm28.14$ & $670\pm55.36$ \\ \hline 0.8-0.9 & $1278\pm72.45$ & $387.6\pm28.57$ & $371.4\pm27.72$ & $312\pm24.57$ & $498.8\pm45.17$ \\ \hline 0.9-1 & $1124\pm68.87$ & $349.1\pm27.39$ & $322.1\pm25.87$ & $288.1\pm23.94$ & $448.3\pm44.89$ \\ \hline 1-1.1 & $954.5\pm62.39$ & $288.2\pm24.38$ & $285.3\pm24.23$ & $265.9\pm23.11$ & $365\pm40.02$ \\ \hline 1.1-1.2 & $832.1\pm58.34$ & $257.2\pm23.06$ & $245.4\pm22.25$ & $213.3\pm20.25$ & $273.8\pm34.08$ \\ \hline 1.2-1.4 & $1268\pm72.57$ & $441.9\pm31.19$ & $362.6\pm27.1$ & $320.2\pm24.77$ & $393.8\pm41.64$ \\ \hline 1.4-1.6 & $806.8\pm57.58$ & $306.9\pm26.38$ & $221.5\pm20.77$ & $190.8\pm18.71$ & $210.3\pm31.59$ \\ \hline 1.6-1.8 & $540.8\pm23.27$ & $170.7\pm13.06$ & $146.2\pm12.14$ & $116.3\pm10.81$ & $126\pm11.31$ \\ \hline 1.8-2 & $314.2\pm17.8$ & $119.2\pm10.97$ & $81.7\pm9.764$ & $68.35\pm9.093$ & $93.98\pm10.15$ \\ \hline 2-2.5 & $388.1\pm21.21$ & $148.4\pm12.48$ & $135.7\pm12.01$ & $89.74\pm10.02$ & $109\pm12.33$ \\ \hline 2.5-3 & $109.1\pm12.92$ & $36.33\pm6.809$ & $34.3\pm8.488$ & $30.64\pm7.487$ & $24.22\pm5.422$ \\ \hline 3-4 & $82.18\pm12.30$ & $---$ & $---$ & $---$ & $---$ \\ \hline \end{tabular} \caption{$p$ raw signal table in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{protontable} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|} \hline $p_{T}$ (GeV/c) & d+Au Trigger & 0\%-20\% & 20\%-40\% & 40\%-100\% & p+p \\ \hline 0.4-0.5 & $692.6\pm26.33$ & $215.1\pm14.67$ & $183\pm13.53$ & $202.8\pm14.26$ & $421.7\pm20.56$ \\ \hline 0.5-0.6 & $1009\pm31.76$ & $310.8\pm17.63$ & $304\pm17.43$ & $268.5\pm16.39$ & $526.6\pm22.95$ \\ \hline 0.6-0.7 & $1098\pm33.17$ & $317.9\pm17.84$ & $305.6\pm17.51$ & $327\pm18.11$ & $561.5\pm23.71$ \\ \hline 0.7-0.8 & $1062\pm32.59$ & $340.2\pm18.44$ & $307\pm17.53$ & $285.5\pm16.9$ & $435.1\pm20.86$ \\ \hline 0.8-0.9 & $992.2\pm31.5$ & $315.1\pm17.81$ & $284.4\pm16.96$ & $244.4\pm15.63$ & $376.4\pm19.4$ \\ \hline 0.9-1 & $827.5\pm28.76$ & $288.9\pm17$ & $225\pm15.01$ & $202.6\pm14.24$ & $310.4\pm17.62$ \\ \hline 1-1.1 & $724\pm26.91$ & $240.2\pm15.5$ & $181.5\pm13.48$ & $192.2\pm13.87$ & $246.4\pm15.7$ \\ \hline 1.1-1.2 & $608.5\pm24.67$ & $161.3\pm12.7$ & $184.3\pm13.61$ & $149.8\pm12.25$ & $192\pm13.87$ \\ \hline 1.2-1.4 & $914.9\pm30.24$ & $301.5\pm17.36$ & $269.7\pm16.42$ & $214.1\pm14.63$ & $269.6\pm16.42$ \\ \hline 1.4-1.6 & $575.8\pm24$ & $204.9\pm14.32$ & $160.9\pm12.71$ & $120.5\pm10.98$ & $138.6\pm12.01$ \\ \hline 1.6-1.8 & $407.2\pm20.18$ & $127.3\pm11.29$ & $108.1\pm10.43$ & $89.9\pm9.497$ & $100.6\pm10.63$ \\ \hline 1.8-2 & $257.3\pm16.26$ & $73.85\pm8.992$ & $92.22\pm9.69$ & $46.81\pm7.802$ & $71.23\pm8.92$ \\ \hline 2-2.5 & $305.6\pm18.45$ & $114\pm11.01$ & $83.43\pm9.464$ & $77.84\pm9.16$ & $64.02\pm10.44$ \\ \hline 2.5-3 & $111\pm12.79$ & $28.91\pm6.26$ & $29.29\pm8.869$ & $20.55\pm7.198$ & $25.71\pm6.856$ \\ \hline 3-4 & $67.05\pm11.87$ & $---$ & $---$ & $---$ & $---$ \\ \hline \end{tabular} \caption{$\bar{p}$ raw signal table in minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions.} \label{pbartable} \end{scriptsize} \end{table} \chapter{{\hspace{3.5cm}STAR Collaboration}} \begin{figure}[htb] \hspace{-6pc} \includegraphics[width=45pc]{sci-july03-01.eps} \end{figure} \begin{figure}[htb] \hspace{-6pc} \includegraphics[width=45pc]{sci-july03-02.eps} \end{figure} \chapter{Conclusion and Outlook} \label{chp:conclusion} \section{Conclusion} In summary, we have reported the identified particle spectra of pions, kaons, protons and anti-protons at mid-rapidity from 200 GeV minimum-bias, centrality selected d+Au collisions and NSD p+p collisions. The time-of-flight detector, based on novel multi-gap resistive plate chamber technology, was used for particle identification. This is the first time that MRPC detector was installed to take data as a time-of-flight detector in the collider experiment. The calibration method was set up in the STAR experiment for the first time and has been applied to the data taken later successfully. The intrinsic timing resolution of the MRPC was 85 ps after the calibration. In 2003 run, the pion/kaon can be separated up to transverse momentum 1.6 GeV/c while proton can be identified up to 3.0 GeV/c. The spectra of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ in d+Au and p+p collisions provide an important reference for those in Au+Au collisions. The initial state in d+Au collisions is similar to that in Au+Au collisions, and, it's believed that the quark-gluon plasma doesn't exist in d+Au collisions. These results from d+Au collisions are very important for us to judge whether the quark-gluon plasma exists in Au+Au collisions or not and to understand the property of the dense matter created in Au+Au collisions. We observe that the spectra of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ are considerably harder in d+Au than those in p+p collisions. In $\sqrt{s_{_{NN}}} = 200$ GeV d+Au collisions, the $R_{dAu}$ of protons rise faster than $R_{dAu}$ of pions and kaons. The $R_{dAu}$ of proton is larger than 1 at intermediate $p_T$ while the proton production follows binary scaling at the same $p_T$ range in 200 GeV Au+Au collisions. These results further prove that the suppression observed in Au+Au collisions at intermediate and high $p_T$ is due to final state interactions in a dense and dissipative medium produced during the collision and not due to the initial state wave function of the Au nucleus. Additionally, the particle-species dependence of the Cronin effect is found to be significantly smaller than that from lower energy p+A collisions. In $\sqrt{s_{_{NN}}} = 200$ GeV d+Au collisions, the ratio of the nuclear modification factor $R_{dAu}$ between $(p+\bar{p})$ and charged hadrons ($h$) in the $p_{T}$ range $1.2< p_{T}<3.0$ GeV/c was measured to be $1.19\pm0.05$(stat)$\pm0.03$(syst) in minimum-bias collisions. Both the $R_{dAu}$ values and $(p+\bar{p})/h$ ratios show little centrality dependence, in contrast to previous measurements in Au+Au collisions at $\sqrt{s_{NN}}$ = 130 and 200 GeV. The ratios of protons over charged hadrons in d+Au and p+p collisions are found to be about a factor of 2 lower than that from Au+Au collisions, indicating that the relative baryon enhancement observed in heavy ion collisions at RHIC is due to the final state effects in Au+Au collisions. The identified particle spectra in d+Au and p+p collisions not only provide the reference for those in Au+Au collisions, but also provide a chance to see the mechanism of the Cronin effect itself clearly. Usually the Cronin effect has been explained to be the initial state effect only since 1970s~\cite{accardi}. However, we compare our pion and proton spectra in minimum-bias and centrality-selected d+Au collisions with the recombination model~\cite{hwayang}. The recombination model can reproduce both the pion spectra and proton spectra. This recombination model is built on the hadronization process, which is a final-state effect, while the initial multiple parton scattering model~\cite{accardi} can't reproduce the difference of the Cronin effect between pions and protons. From these comparisons, we conclude that the Cronin effect in $\sqrt{s_{_{NN}}} = 200$ GeV d+Au collisions is not the initial state effect only, and that final state effect plays an important role. The integral yield $dN/dy$ and $\langle p_T \rangle$ in p+p and d+Au collisions were estimated from the power law fit and thermal model fit. The integral yield $R_{dAu}$ of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ are observed to be smaller than 1 while those of $p$ and $\bar{p}$ are close to 1. The $\pi^{-}/\pi^{+}$, $K^{-}/K^{+}$ and $\bar{p}/p$ ratios as a function of $p_{T}$ are observed to be flat with $p_T$ within the errors in d+Au and p+p minimum-bias collisions and show little centrality dependence in d+Au collisions. The integral yield ratios of $K^{-}/\pi^{-}$ and $\bar{p}/\pi^{-}$ as a function of $dN/d\eta$ were also presented in p+p and d+Au collisions. \section{Outlook} For the outlook, I will discuss whether the Cronin effect is mass dependent or baryon/meson dependent at 200 GeV. What other physics topic have we done from MRPC-TOFr in d+Au and p+p collisions in 2003 run? If we have the full time-of-flight (Full-TOF) coverage, what can we do? Also I will discuss a little bit about the low energy 63 GeV Au+Au run. \subsection{Cronin effect at 200 GeV: Mass dependent or baryon/meson dependent?} We know that recombination model can reproduce the spectra of pions and protons in d+Au collisions. Also the $R_{CP}$ of identified particles in Au+Au collisions suggest that the degree of suppression depends on particle species(baryon/meson) at intermediate $p_T$. Does the Cronin effect in 200 GeV d+Au collisions depend on the particle species (baryon/meson) or depend on the particle mass? From our data, it shows the Cronin effect for proton is bigger than those for pion and kaon. And the Cronin effect of pion shows little difference from that of kaon at $p_T<1.5$ GeV/c. In order to see the Cronin effect is baryon/meson dependent or mass dependent, we can compare the Cronin effect of proton with those of $K^*$ and $\phi$ since the mass of $K^*$ and $\phi$ are close to that of proton while $K^*$ and $\phi$ are mesons and proton is a baryon. The preliminary results show that the Cronin effect of $K^*$ and $\phi$~\cite{kstarphi} are similar to that of pion and different from that of proton. However, the final results from $K^*$ and $\phi$ are needed to confirm this issue. \subsection{Electron PID from MRPC-TOFr} The production and spectra of hadrons with heavy flavor are sensitive to initial conditions and the later stage dynamical evolution in high energy nuclear collisions, and may be less affected by the non-perturbative complication in theoretical calculations~\cite{charm1}. Charm production has been proposed as a sensitive measurement of parton distribution function in nucleon and the nuclear shadowing effect by systematically studying p+p, and p+A collisions~\cite{lin96}. The relatively reduced energy loss of heavy quark traversing a quark-gluon plasma will help us distinguish the medium in which the jet loses its energy~\cite{dokshitzer01}. A possible enhancement of charmonia ($J/\Psi$) production can be present at RHIC energies~\cite{jpsi} due to the coalescence of the copiously produced charm quarks~\cite{opencharm}. \begin{figure}[h] \centering \includegraphics[height=20pc,width=24pc]{electronpid.eps} \caption{$dE/dx$ in TPC versus $p$ without(the upper panel) or with (the lower pannel) the TOFr velocity cut $|1/\beta- 1|\le0.03$. The insert shows $dE/dx$ distribution for 1 $\le p \le$ 1.5 GeV/c. } \label{electron} \end{figure} The recent STAR results on the absolute open charm cross section measurements from direct charmed hadron $D^0$ reconstruction~\cite{Haibin:03} in d+Au collisions and electrons from charm semileptonic decay in both p+p and d+Au collisions at 200 GeV were presented~\cite{opencharm}. Based on the capability of hadron identification~\cite{startof1} from the MRPC-TOFr tray in 2003, electrons could be identified at low momentum ($p_{T}\le3$ GeV/c) by the combination of velocity ($\beta$) from TOFr~\cite{startof} and the particle ionization energy loss ($dE/dx$) from TPC~\cite{tpc}. Figure~\ref{electron} shows that the electrons are clearly identified as a separate band in the $dE/dx$ versus momentum ($p$) with a selection on $\beta$ at $|1/\beta-1|\le0.03$ in d+Au collisions. At higher $p_{T}$ (2--4 GeV/c), negative electrons were also identified directly by TPC since hadrons have lower $dE/dx$ due to the relativisitic rise of electron $dE/dx$. Based on the clear electron identification, the open-charm-decayed electron spectra was derived~\cite{opencharm}. Combined with $D^0$ measurement from TPC, the total charm cross section was obtained~\cite{opencharm} in d+Au collisions. \subsection{Full-TOF Physics} Based on the hadron PID and electron PID of MRPC-TOFr in 2003, we can imagine how many physics we can do if we have full time-of-flight coverage based on MRPC technology. The proposal~\cite{tofproposal} for large area time-of-flight system for STAR has been proposed. Since the pion/kaon can be separated up to transverse momentum 1.6 GeV/c and proton can be identified up to 3.0 GeV/c from time-of-flight system. The resonance spectra measured from hadronic decay will be extended to much higher $p_T$. The direct open charm spectra from its hadronic decay channel will reach higher $p_T$ with much more precise measurement. Since the electron can be clearly identified up to transverse momentum 3$\sim$4 GeV/c by the combination of velocity ($\beta$) from TOFr ~\cite{startof} and the particle ionization energy loss ($dE/dx$) from TPC ~\cite{tpc}, the electron spectra from charm-semi-leptonic decay will be measured precisely. As we know that the measurement of the di-leptonic decays of vector mesons are very difficult since the branch ratios are too small and it's really hard to subtract the background. But with TOF upgrading together with the SVT and micro-vertex detector upgrading, the di-leptonic decays of vector mesons will be measured much more easily, which will bring the direct information of QGP since the electron is a lepton and the cross section of interaction between electrons and hadrons is little. Thus we can see directly the property of quark-gluon plasma such as the temperature and the chiral symmetry restoration. This will be the most interesting and meaningful thing for the QGP search~\cite{xzb}. Besides, there are many other physics topics~\cite{tofproposal} such as identified particle correlation and fluctuation, particle composition of jet fragmentation, and anti-nuclei etc. \subsection{63 GeV Au+Au collisions at RHIC} The bulk properties such as elliptic flow $v_2$ and particle production show smooth trend from AGS, SPS to RHIC energy. One energy point $\sqrt{s_{_{NN}}} = 63$ GeV, which is between SPS and full RHIC energy, was selected since high quality charged-particle and $\pi^0$ inclusive spectra have been measured in p+p collisions at 63 GeV at Intersecting Storage Rings (ISR) and will serve as the reference spectra for computing the nuclear modification factor for Au+Au collisions measured at the same energy. The $v_2$ and particle production at 63 GeV will be studied at RHIC. Besides, the nuclear modification factor as a function of $p_T$ will also be studied in this collision system in which the hard scattering component has been significantly reduced. The results from 63 GeV Au+Au collisions will be helpful for us to understand the property of dense medium created in 200 GeV Au+Au collisions. \chapter{Discussion} \label{chp:discussion} \section{Cronin effect} The identified particle spectra in d+Au and p+p collisions not only provide the reference for those in Au+Au collisions at 200 GeV, but also provide a chance to see the mechanism of the Cronin effect itself clearly. Cronin effect was observed 30 years ago~\cite{cronin}. It is the enhancement of particle production at high $p_{T}$. The enhancement was explained by initial multiple parton scattering. Also the recent experimental results of Cronin effect on inclusive charged hadron are consistent with the predictions based on initial multiple parton scattering~\cite{accardi}. It suggests the suppression at intermediate $p_{T}$ in Au+Au collisions is due to final state effects. However, the initial multiple parton scattering with the independent fragmentation function will result in the same Cronin effect for $p(\bar{p})$ and for pions, while experimentally the Cronin effect for $p(\bar{p})$ is larger than that for $\pi$. That's to say the initial multiple scattering with the independent fragmentation function can't account for the Cronin effect observed. Maybe in the initial multiple parton scattering, the broadening for gluon and for quark/antiquark are not the same~\cite{xinniancommu}. Or maybe the fragmentation processes in p+A collisions are not the same as those in p+p collisions~\cite{qiucommu}. Whether the Cronin effect is initial state effect or final state effect will be discussed below. \subsection{Model comparison: initial state effect?} The initial multiple parton scattering model predicts that the Cronin effect on deuteron beam outgoing side is larger than that on Au beam outgoing side since the deuteron traverses a much larger nucleus~\cite{xinnian:dAu}. Figure~\ref{etaasymmetry} (left) shows the predictions for the Cronin effect at different rapidity range. The different curves correspond to the prediction results from different shadowings. The $y=1$ is on the deuteron beam outgoing side. The $y=-1$ is on the Au beam outgoing side. The $y=0$ is at mid-rapidity. We can see that the $R_{dAu}$ on deuteron beam side ($y=1$) increases faster than that on Au beam side ($y=-1$). If we take the ratio of $R_{dAu}$ on Au beam side over $R_{dAu}$ on deuteron beam side, it will result in a minimum value at $p_{T}\sim3.5$ GeV/c, as shown in the curves on the right plot of Figure~\ref{etaasymmetry}. The solid symbol on the right plot of Figure~\ref{etaasymmetry} represents the data points~\cite{johan}, which is the ratio of $R_{dAu}$ on Au beam side at $-1<\eta<-0.5$ over $R_{dAu}$ on deuteron beam side at $0.5<\eta<1$. We observe the $\eta$ asymmetry from experiment reaches a maximum value firstly and then decreases. This is different from the predictions. That means, the model based on initial multiple parton scattering only, can't reproduce the experimental results. Recently, Qiu and Vitev have come up with the idea of coherent multiple scattering and applied it to the RHIC experiments~\cite{coherent}. In this picture, the hard probe may interact coherently with many low x parton inside different nucleons inside the nucleus. As a result, this process will lead to the suppression of the total cross section. This coherent effect will play an important role in p+A collisions at forward rapidity. In the deuteron outgoing beam direction, the coherent effect is non-negligible since the Au nucleus is big while on the Au side, the coherent effect is not big since the deuteron is of a small size. This will result in bigger suppression on the deuteron side than on the Au side. It may qualitatively reproduce the data. This coherent multiple scattering is a final-state effect. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=14pc]{r200.eps} \end{minipage} \hspace{-2cm} \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=24pc]{wongRatio_hipt_one_panel_v4-1.eps} \end{minipage} \caption{(left) The Cronin effect at different rapidity as a function of $p_{T}$. The different curve in each panel shows the different shadowing. This figure is from ~\cite{xinnian:dAu}. (right) The $\eta$ asymmetry of the Cronin effect: the ratio of Cronin effect in Au beam outgoing direction over the Cronin effect in deuteron beam outgoing direction. This figure is from ~\cite{johan}.} \label{etaasymmetry} \end{figure} As we all known, in Au+Au collisions, the suppression at intermediate $p_{T}$ can be reproduced by the initial multiple scattering and jet quenching qualitatively~\cite{starhighpt,jetquench}. However, the model based on the initial multiple scattering, jet quenching and independent fragmentation will result in the same suppression for baryons and mesons at intermediate $p_{T}$ in Au+Au collisions. Experimentally $R_{cp}$ for baryons are larger than $R_{cp}$ for mesons at intermediate $p_{T}$. This difference can be reproduced by coalescence or recombination models~\cite{hwa,fries,ko}. Recently the recombination model~\cite{hwayang} has been applied to d+Au system to see whether it can reproduce the Cronin effect or not. With the help of Prof. C.B. Yang~\cite{yang}, I also compare our pion and proton spectra in d+Au collisions with the recombination model~\cite{hwayang}. In the following the recombination model~\cite{hwayang} will be discussed and the comparison between the data and the model will be presented in detail. \subsection{Model comparison: recombination} The inclusive distribution for the production of pions can be written in the recombination model~\cite{hwayang}, when mass effects are negligible, in the invariant form \begin{eqnarray} p{dN_{\pi} \over dp} = \int {dp_1 \over p_1}{dp_2 \over p_2}F_{q\bar{q}} (p_1, p_2) R_{\pi}(p_1, p_2, p) , \label{1} \end{eqnarray} where $F_{q\bar{q}} (p_1, p_2)$ is the joint distribution of a $q$ and $\bar q$ at $p_1$ and $p_2$, and $R_{\pi}(p_1, p_2, p)$ is the recombination function for forming a pion at $p$: $R_{\pi}(p_1, p_2, p) = (p_1p_2/p)\delta (p_1+p_2- p)$. $F_{q\bar{q}}$ depends on the colliding hadron/nuclei. In general, $F_{q\bar{q}}$ has four contributing components represented schematically by \begin{eqnarray} F_{q\bar{q}} = {\cal TT} + {\cal TS} + ({\cal SS}) _1 + ({\cal SS})_2 \end{eqnarray} where $\cal{ T}$ denotes thermal distribution and $\cal{S}$ shower distribution. $({\cal SS})_1$ signifies two shower partons in the same hard-parton jet, while $({\cal SS})_2$ stands for two shower partons from two nearby jets~\cite{hwayang}. For $p+A$ collisions it may not be appropriate to refer to any partons as thermal in the sense of a hot plasma as in heavy-ion collisions. Here in d+Au collisions, the symbol $\cal{ T}$ represents the soft parton distribution at low $k_T$. At low $p_T$ the observed pion distribution is exponential; we identify it with the contribution of the ${\cal TT}$ term~\cite{hwayang}. \begin{eqnarray} {dN^{{\cal TT}}_{\pi} \over pdp} = {C^2 \over 6} exp (-p/T) \end{eqnarray} where $T$ is the inverse slope. We shall determine $C$ and $T$ by fitting the d+Au data at low $p_T$. The pion spectra for different centralities can be calculated from thermal-thermal ($pion_{tt}$), thermal-shower ($pion_{ts}$) and shower-shower ($pion_{ss}$) contributions by using parameters $C$ and $N_{bin}$: $dN/p_Tdp_T=C\times C\times pion_{tt}+2.5 \times C \times N_{bin} \times pion_{ts}+2.5 \times N_{bin} \times pion_{ss}$, where C is determined by fitting the d+Au data at $0.4<p_T<1.0$ GeV/c, and $N_{bin}$ is the number of binary collisions. The data points of $pion_{tt}, pion_{ts}$ and $pion_{ss}$ are from Prof. C.B. Yang~\cite{yang}. The $C$ values for minimum-bias, 0-20\%, 20-40\% and 40-$\sim$100\% d+Au collisions are 8.85, 13.08, 10.96 and 6.84 individually. The $T$ value of 0.21 GeV is used in the low $p_{T}$ fit. Figure~\ref{pirecombination} shows the $\pi^{+}$ spectra in d+Au collisions as well as those from recombination model. This figure shows that the recombination model can reproduce the spectra of pion in minimum-bias and centrality selected d+Au collisions. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=18pc]{pipluscent_recombine_new.eps} \end{minipage} \hspace{0mm} \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=18pc]{piplusall_recombine.eps} \end{minipage} \caption{(left) The invariant yield for $\pi^{+}$ at 0\%-20\% d+Au collisions as a function of $p_{T}$. The open circles are our data points. The curves are the calculation results from recombination model. Sum represents the total contribution from recombination model. Thermal-thermal represents the soft contribution. The thermal-shower represents the contribution from the interplay between soft and hard components. The shower-shower represents the hard contribution. (right) The invariant yields for $\pi^{+}$ in minimum-bias and centrality selected d+Au collisions as a function of $p_{T}$. The symbols represent our data points. The curves on the top of the symbols are the corresponding calculation results from recombination model. } \label{pirecombination} \end{figure} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=18pc]{ppluscent_recombine_new.eps} \end{minipage} \hspace{0mm} \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=18pc]{pplusall_recombine.eps} \end{minipage} \caption{(left) The invariant yield for $p$ at 0\%-20\% d+Au collisions as a function of $p_{T}$. The open circles are our data points. The curves are the calculation results from recombination model. Sum represents the total contribution from recombination model. TTT represents the soft contribution. The TTS+TSS represents the contribution from the interplay between soft and hard components. The SSS represents the hard contribution. (right) The invariant yields for $p$ in minimum-bias and centrality selected d+Au collisions as a function of $p_{T}$. The symbols represent our data points. The curves on the top of the symbols are the corresponding calculation results from recombination model. } \label{precombination} \end{figure} The invariant inclusive distribution for proton formation at midrapidity in the recombination model ~\cite{hwayang} \begin{eqnarray} p^0{dN_p \over dp} = \int {dp_1 \over p_1}{dp_2 \over p_2} F (p_1, p_2, p_3) R_p(p_1, p_2, p_3, p) \end{eqnarray} where all momentum variables $p_i$ and $p$ are transverse momenta, and $p^0$ denotes the energy of the proton. $F (p_1, p_2, p_3)$ is the joint distribution of $u, u,$ and $d$ quarks at $p_1, p_2$ and $p_3$, respectively. $R_p(p_1, p_2, p_3, p)$ is the recombination function for a proton with momentum $p$. We write schematically \begin{eqnarray} F = {\cal TTT} + {\cal TTS} + {\cal TSS} + {\cal SSS} \end{eqnarray} where all the shower partons $\cal{S}$ are from one hard parton jet. Shower partons from different jets are ignored here for RHIC energies. In d+Au collisions, $\cal{ T}$ denotes the soft partons that are not associated with the shower components of a hard parton. The ${\cal SSS}$ term is regarded as the fragmentation of a hard parton into a proton. The ${\cal TTT}$ term comes entirely from the soft partons, while ${\cal TTS}$ and ${\cal TSS}$ accounts for the interplay between the soft and shower partons. The soft contribution to the proton spectrum arising from ${\cal TTT}$ recombination is \begin{eqnarray} {dN^{\rm th}_{proton} \over pdp} = {C^3 \over 6} {p^2 \over p^0} e^{-p/T} { B (\alpha + 2, \gamma +2) B (\alpha + 2,\alpha + \gamma +4) \over B (\alpha + 1, \gamma +1) B (\alpha + 1,\alpha + \gamma +2) } \end{eqnarray}. Where $C$ and $T$ are determined by fitting the proton spectra at low $p_{T}$, the $\alpha$ is equal to 1.75, $\gamma$ is equal to 1.05, $B(x,y)$ is the beta function~\cite{hwayang}. For the invariant yield of proton, there are 4 different contributions: soft-soft-soft ($proton_{ttt}$), soft-soft-shower ($proton_{tts}$), soft-shower-shower ($proton_{tss}$), and shower-shower-shower ($proton_{sss}$). The total contributions are $dN/p_Tdp_T=C \times C \times C \times proton_{ttt}+C \times C \times N_{bin} \times proton_{tts}+C \times N_{bin} \times proton_{tss}+N_{bin} \times proton_{sss}$, where $C$ is determined by fitting the d+Au data at $0.5<p_T<1.5$ GeV/c, and $N_{bin}$ is the number of binary collisions. The data points of $proton_{ttt}, proton _{tts}, proton_{tss}$ and $proton_{sss}$ are from Prof. C.B. Yang~\cite{yang}. The $C$ values for minimum-bias, 0-20\%, 20-40\% and 40-$\sim$100\% d+Au collisions are 9.67, 12.34, 10.92 and 7.91 individually. The $T$ value of 0.21 GeV is used in the low $p_{T}$ fit. Figure~\ref{precombination} shows the proton spectra in d+Au collisions as well as those from recombination model. This figure shows that the recombination model can reproduce the spectra of proton in minimum-bias and centrality selected d+Au collisions. From the comparison between our data and the calculation results from the recombination model, we know that the recombination model actually can reproduce both the proton and pion spectra in d+Au collisions, while as we have mentioned above, the initial multiple parton scattering model~\cite{accardi} with independent fragmentation can't reproduce the difference of Cronin effect between proton and pion. Besides, the initial multiple parton scattering model with independent fragmentation can't reproduce the $\eta$ asymmetry of the Cronin effect. In the recombination model~\cite{hwayang}, the number of such soft partons on the Au outgoing side is larger than that on the deuteron outgoing side. This will result in the Cronin effect on the Au side larger than that on the deuteron side~\cite{hwayang}. Qualitatively the recombination model can reproduce the $\eta$ asymmetry of the Cronin effect. As we know, the recombination model is a final-state effect model. These all seem to indicate that the Cronin effect is not initial-state effect only. The final-state effect plays an important role too. To directly confirm the Cronin effect is initial or final state effect, it's necessary for us to compare the Cronin effect of Drell-Yan process with those of pion, kaon and proton. I will come to this later. \subsection{Integral yield $R_{dAu}$: shadowing effect?} \begin{figure}[h] \centering \includegraphics[height=24pc,width=24pc]{integral_rdau.eps} \caption{Integral yield $R_{dAu}$ as a function of $dN/d\eta$ in minimum-bias and centrality selected d+Au collisions at mid-rapidity. Statistic errors and systematic uncertainties have been added in quadrature. The shadowing represents the normalization uncertainty.} \label{integralrdau} \end{figure} The initial multiple elastic scattering only changes the $p_{T}$ distribution while the total cross section should not change. Thus we can look at the integral yield $dN/dy$ $R_{dAu}$, which are measured through comparison to the integral yield $dN/dy$ in p+p collisions, scaled by the number of binary collisions $N_{bin}$. Figure~\ref{integralrdau} shows that integral yield $R_{dAu}$ of pion, kaon and proton as a function of $dN/d\eta$ in minimum-bias and centrality selected d+Au collisions at mid-rapidity. The integral yield $R_{dAu}$ of pion and kaon are less than 1 while that of proton is close to 1. This may be the indication of shadowing effect at 200 GeV. The integral yield $R_{dAu}$ for proton is larger than that for kaon and a little bit more larger than that for pion. This may be the indication that the shadowing effect is mass dependent at 200 GeV. \subsection{Initial or final state effect: Drell-Yan process} In order to see the Cronin effect is initial or final state effect, we may look into the Drell-Yan process since there is little final state effect in Drell-Yan process. If there is no enhancement at high $p_{T}$ for Drell-Yan process, the enhancement for $\pi, K, p$ is due to final-state effect. Figure~\ref{drellyan} shows the integral yield Cronin ratio as a function of atomic weight at p-A fixed target experiment~\cite{drell}. The proton incident energy is 800 GeV. We can see there is no enhancement for Drell-Yan process. However, this is the total cross section while what we want to compare is Cronin ratio as a function of $p_{T}$. It will be better if we have the $p_{T}$ dependence of Cronin ratio of Drell-Yan process. However, at the same $p_T$ range with the same proton incident energy, the Cronin ratio of Drell-Yan is not available in p+A collisions. It's hard to compare the Cronin ratio of Drell-Yan process with those of $\pi, K, p$. \begin{figure}[h] \centering \includegraphics[height=24pc,width=24pc]{fig16.eps} \caption{Integral yield Cronin ratio as a function of atomic weight in p+A fixed target experiment for Drell-Yan process, etc. This figure is from~\cite{drell}.} \label{drellyan} \end{figure} \section{Baryon excess in Au+Au collisions} Now let's come to another important physics from d+Au collisions. We know that the $(p+\bar{p})/h$ ratio from minimum-bias Au+Au collisions~\cite{phenixpid} at a similar energy is about a factor of 2 higher than that in d+Au and p+p collisions for $p_{T}{}^{>}_{\sim}2.0$ GeV/c. This enhancement is most likely due to final-state effects in Au+Au collisions. There are many models trying to explain this baryon excess in Au+Au collisions~\cite{jetquench,junction,derekhydro,pisahydro,fries,ko}. In the following baryon production mechanism will be discussed. \subsection{$\bar{p}/p$ ratio vs $p_T$} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{pbarp_pQCD.eps} \caption{$\bar{p}/p$ ratio as a function of $p_{T}$ in d+Au and p+p minimum-bias collisions. The open squared symbols are for p+p collisions and the solid circled symbols for d+Au collisions. The triangled symbols represent the result from Au+Au minimum-bias collision~\cite{ex0307022}. The curve is the pQCD calculation results from~\cite{junction} in p+p collisions. Errors are statistical.} \label{pbarpdiscussion} \end{figure} In 200 GeV Au+Au collisions, $\bar{p}/p$ ratio was observed to be flat with $p_{T}$ till intermediate $p_{T}$ range~\cite{ex0307022}, as shown in Figure~\ref{pbarpdiscussion}. The baryon junction model~\cite{junction} tried to explain it by using junction anti-junction production with jet quenching, on the basis of pQCD calculation~\cite{junction} where the $\bar{p}/p$ ratio decreases with $p_{T}$ in p+p collisions. The curve from pQCD calculation~\cite{junction} is also shown in Figure~\ref{pbarpdiscussion}. However, $\bar{p}/p$ ratios in d+Au and p+p collisions in our data show to be flat with $p_{T}$ within errors. Anyway, the precise measurement with more statistics in p+p and d+Au collisions is needed to address this issue. \subsection{Baryon production at RHIC: multi-gluon dynamics?} \begin{figure}[h] \centering \includegraphics[height=18pc,width=18pc]{pnch_ee.eps} \caption{Minimum-bias ratios of ($p+\bar{p}$) over charged hadrons at $-0.5\!<\!\eta\!<\!0.0$ from $\sqrt{s_{_{NN}}} =200$ GeV p+p (open diamonds) and d+Au (filled triangles) collisions. Also shown are the $(p+\bar{p})/h$ ratios in $e^{+}e^{-}$ collisions at ARGUS~\cite{argus}. The solid line represents the $(p+\bar{p})/h$ ratio from three gluon hadronization while the dashed line for the ratio from quark and antiquark fragmentation~\cite{argus}. Errors are statistical.} \label{pbarpnchcomparison} \end{figure} Let's compare the $(p+\bar{p})/h$ ratio in p+p, d+Au and Au+Au collisions at RHIC energy 200 GeV with the ratio in $e^{+}e^{-}$ collisions at ARGUS~\cite{argus}. Using the ARGUS detector at the $e^{+}e^{-}$ storage ring DORIS II, the inclusive production of pion, kaon and proton in multihadron events at 9.98 GeV and in direct decays of the $\Upsilon(1S)$ meson were investigated~\cite{argus}. Multihadron final states in $e^{+}e^{-}$ annihilation are produced via quark and antiquark fragmentation, and those from direct $\Upsilon(1S)$ decays originate from the hadronization of three gluons~\cite{argus}. Figure~\ref{pbarpnchcomparison} shows the $(p+\bar{p})/h$ ratio in 200 GeV p+p collisions together with the ratio in $e^{+}e^{-}$ collisions at ARGUS~\cite{argus}. The plot shows that the $(p+\bar{p})/h$ ratio from three gluon hadronization is a factor of 3 higher than that from quark and antiquark fragmentation at ARGUS. Our data from 200 GeV p+p collisions is close to $(p+\bar{p})/h$ ratio from 3 gluon hadronization. This may be the indication that in the heavy ion collisions at RHIC energy, multi-gluon hadronization plays an important role for the particle production. \chapter{Physics} \label{chp:physics} \section{Deconfinement and phase diagram} The theory which describes the interaction of the color charges of quarks and gluons is called Quantum Chromodynamics (QCD). In phenomenological quark models, mesons can be described as quark-antiquark bound states, while baryons can be considered as three quark bound states. Up to now, it's found that all the hadron states which can be observed in isolation is colorless singlet states. Experimentally, no single quark, which is described by a color-triplet state, has ever been isolated. The absence of the observation of a single quark in isolation suggests that the interaction between quarks and gluons must be strong on large distance scale. In the other extreme, much insight into the nature of the interaction between quarks and gluons on short distance scales was provides by deep inelastic scattering experiments. In these experiments, the incident electron interacts with a quark within a hadron and is accompanied by the momentum transfer from the electron to the quark. The measurement of the electron momentum before and after the interaction allows a probe of the momentum distribution of the parton inside the nucleon. It was found that with very large momentum transfer, the quarks inside the hadron behave as if they were almost free~\cite{QCD}. The strong coupling between quarks and gluons at large distances and asymptotic freedom are the two remarkable features of QCD. When the energy density is high enough either due to the high temperature or high baryon density, the quark or gluon may be deconfined from a hadron. The thermalized quark gluon system is what we called quark-gluon plasma. Lattice QCD calculations, considering two light quark flavors, predict a phase transition from a confined phase, hadronic matter, to a deconfined phase, or quark-gluon plasma (QGP), at a temperature of approximately \begin{figure}[h] \centering \includegraphics[height=28pc,width=28pc]{phase_diagram.eps} \caption{Phase diagram of hadronic and partonic matter. Figure is taken from~\cite{pbm:01}.} \end{figure} 160 MeV~\cite{harris:98}. Figure 1.1 shows the phase diagram of the hadronic and partonic matter. A phase transition from the confined hadronic matter to the deconfined QGP matter is expected to happen at either high temperature or large baryon chemical potential $\mu_B$. Recent Lattice QCD calculations show that the QGP is far from ideal below 3 $T_{c}$. The nonideal nature of this strongly coupled QGP is also seen from the deviation of the pressure, $P(T)$, and energy density $\epsilon(T)$ from the Stefan Boltzmann limit as shown in Figure from~\cite{Fodor}. \begin{figure}[h] \centering \includegraphics[height=16pc,width=28pc]{FodorKatz_P_e_mu0.eps} \caption{A recent Lattice QCD calculation \protect{\cite{Fodor}} of the pressure, $P(T)/T^4$, and a measure of the deviation from the ideal Stefan-Boltzmann limit $(\epsilon(T)-3 P(T))/T^4$.} \label{Fodorplot} \end{figure} Experiments on relativistic heavy ion collisions are designed to search for and study the deconfined QGP matter. \section{Relativistic Heavy Ion Collisions} The experimental programs in relativistic heavy ions started in 1986 using the Alternating Gradient Synchrotron (AGS) at Brookhaven National Lab (BNL) and the Super Proton Synchrotron (SPS) at European laboratory for particle physics (CERN). At BNL, ion beams of silicon and gold, accelerated to momenta of 14 and 11 GeV/c per nucleon, respectively, have been utilized in 10 fixed-target experiments. There have been 15 heavy ion experiments at CERN utilizing beams of oxygen at 60 and 200 GeV/c per nucleon, sulphur at 200 GeV/c per nucleon and Pb at 160 GeV/c per nucleon~\cite{harris:98}.\\ The Relativistic Heavy Ion Collider (RHIC) at BNL is designed for head-on Au+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV. The first RHIC run was performed in 2000 with Au+Au collisions at $\sqrt{s_{NN}}$ = 130 GeV/c in four experiments, STAR, PHENIX, PHOBOS and BRAHMS. The second RHIC run was in 2001 and 2002 with Au+Au and p+p collisions at $\sqrt{s_{NN}}$ = 200 GeV. The third RHIC run was in 2002 and 2003 with d+Au and p+p collisions at $\sqrt{s_{NN}}$ = 200 GeV.\\ The above mentioned relativistic heavy ion collision experiments are designed for the search and study of the possible deconfined high energy density matter, quark-gluon plasma. In head-on relativistic heavy ion collisions, two nuclei can be represented as two thin disks approaching each other at high speed because of the Lorentz contraction effect in the moving direction. During the initial stage of the collisions, the energy density is higher than the critical energy density from the Lattice QCD calculation, so the quarks and gluons will be de-confined from nucleons and form the quarks and gluons system. The large cross section of interaction may lead to the thermalization of the quarks and gluons system. That's what we called the formation of quark-gluon plasma (QGP). In this stage, the high transverse momentum ($p_{T}$) jets and $c\bar{c}$ pair will be produced due to the large momentum transfer. After that, the QGP will expand and cool down and enter into the mixed-phase expansion. The chemical freeze out point will be formed after the inelastic interactions stop. That means that the particle yields and ratios will not change. After the chemical freeze out, the elastic interactions between hadrons will change the $p_{T}$ distribution of particles. The particles will freeze out finally from the system after the elastic interactions stop. That's what we called the kinetic freeze out point. In the following the important results from RHIC will be addressed. \section{The experimental results at RHIC} \subsection{Flow} In non-central Au+Au collisions, the spatial space asymmetry will be transferred into the momentum space asymmetry by the azimuthal asymmetry of pressure gradients. \begin{figure}[h] \centering \includegraphics[height=16pc,width=22pc]{Ks_Lam_fig1_color_2.eps} \caption{The minimum-bias (0--80\% of the collision cross section) $v_{2}(p_T)$ for $K_{S}^{0}$, $\Lambda+\overline{\Lambda}$ and $h^{\pm}$. The error bars shown include statistical and point-to-point systematic uncertainties from the background. The additional non-flow systematic uncertainties are approximately -20\%. Hydrodynamical calculations of $v_2$ for pions, kaons, protons and lambdas are also plotted~\cite{hydroPasi01}. Figure is taken from~\cite{starv2raa}.} \label{KsLamv2} \end{figure} The azimuthal particle distributions in momentum space can be expanded in a form of Fourier series \begin{equation} E\frac{d^3N}{d^3p}=\frac{1}{2\pi}\frac{d^2N}{p_Tdp_Tdy}(1+ \sum^{\infty}_{n=1}2v_n\cos[n(\phi-\Psi_r)]) \end{equation} where $\Psi_r$ denotes the reaction plane angle. The Fourier expansion coefficient $v_n$ stands for the $n$th harmonic of the event azimuthal anisotropies. $v_1$ is so called direct flow and $v_2$ is the elliptic flow. The elliptic flow is generated mainly during the highest density phase of the evolution before the initial geometry asymmetry of the plasma disappears. Hydrodynamical calculations~\cite{derekhydro} show most of $v_2$ is produced before 3 fm/c at RHIC.\\ Figure~\ref{KsLamv2} shows that The $v_2$ of $K_{S}^{0}$, $\Lambda+\overline{\Lambda}$ and charged hadrons ($h^{\pm}$) as a function of $p_T$ for 0--80\% of the collision cross section~\cite{starv2raa}. Also shown are the $v_2$ of pions, kaons, protons and lambdas from hydrodynamical model~\cite{hydroPasi01}. The $v_2$ from hydrodynamical model shows strong mass dependence, which fits the $K_{S}^{0}$ $v_2$ up to $p_{T}\sim1$ GeV/c and fits the $\Lambda+\overline{\Lambda}$ $v_2$ up to $p_{T}\sim2.5$ GeV/c. Even though the $v_2$ from hydrodynamical model shows consistency with data at low $p_T$, however, the $v_2$ from experimental results show saturation at intermediate $p_{T}$ while hydrodynamical predictions show rising trend at the same $p_T$ range. \subsection{High $p_{T}$ suppression and di-hadron azimuthal correlation} The $v_2$ from hydrodynamical models show consistency with data at lower $p_T$ and fail to reproduce data at higher $p_{T}$. At high $p_{T}$, the suppression for charged hadron production was observed in Au+Au collisions at RHIC energy. The comparison of the spectra in Au+Au collisions through those in p+p collisions, scaled by the number of binary nucleon nucleon collisions is the nuclear modification factor $R_{AA}$. \begin{equation} R_{AA}(p_T)=\frac{d^2N^{AA}/dp_Td\eta}{T_{AA}d^2\sigma^{NN}/dp_Td\eta} \end{equation} where $T_{AA}=\langle N_{\text{bin}} \rangle /\sigma^{NN}_{\text{inel}}$ accounts for the collision geometry, averaged over the event centrality class. $\langle N_{\text{bin}} \rangle$, the equivalent number of binary $NN$ collisions, is calculated using a Glauber model. The $R_{AA}$ is an experimental variable. The high $p_T$ hadron suppression in central Au+Au collisions can also be investigated by comparing the hadron spectra in central and peripheral Au+Au collisions. That's what we called $R_{CP}$. $R_{CP}$ is defined as \begin{equation} R_{CP}=\frac{\langle N_{\text{bin}}^{\text{peripheral}} \rangle d^2N^{\text{central}}/dp_Td\eta}{\langle N_{\text{bin}}^{\text{central}} \rangle d^2N^{\text{peripheral}}/dp_Td\eta}. \end{equation}\\ \begin{figure}[h] \centering \includegraphics[height=20pc,width=26pc]{highpt_200.eps} \caption{$R_{AA}(p_T)$ of inclusive charged hadron for various centrality bins. Figure is taken from ~\cite{starhighpt}.} \label{raa200} \end{figure} Figure~\ref{raa200} shows $R_{AA}(p_T)$ of inclusive charged hadron for various centrality bins in Au+Au collisions at $\sqrt{s_{NN}}$=200 GeV. $R_{AA}(p_T)$ increases monotonically for $p_T<$ 2 GeV/c at all centralities and saturates near unity for $p_T>$ 2 GeV/c in the most peripheral bins. In contrast, $R_{AA}(p_T)$ for the central bins reaches a maximum and then decreases strongly above $p_T$ = 2 GeV/c, showing the suppression of the charged hadron yield relative the $NN$ reference~\cite{starhighpt}.\\ Suppression of high $p_{T}$ hadron production in central Au+Au collisions relative to p+p collisions ~\cite{starhighpt,phenixhighpt} has been interpreted as energy loss of the energetic partons traversing the produced hot and dense medium~\cite{jetquench}, that's so called jet quenching. If a dense partonic matter is formed during the initial stage of a heavy-ion collision with a large volume and a long life time (relative to the confinement scale $1/\Lambda_{\rm QCD}$), the produced large $E_T$ parton will interact with this dense medium and will lose its energy via induced radiation. The energy loss depends on the parton density of the medium. Therefore, the study of parton energy loss can shed light on the properties of the dense matter in the early stage of heavy-ion collisions~\cite{jetquench}. At sufficiently high beam energy, gluon saturation is also expected to result in a relative suppression of hadron yield at high $p_{T}$ in A+A collisions~\cite{cgc}. Also shown in the Figure~\ref{raa200} are the results from perturbative QCD (pQCD) calculations. The Full-pQCD calculations include the partonic energy loss, the Cronin enhancement(due to initial multiple scattering) and the nuclear shadowing effect. The suppression is predicted to be $p_T$ independent when $p_T$ is larger than 6 GeV/c, which is consistent with our data. However, the discrepancy at 2-6 GeV/c was observed between the prediction and the experimental data. This discrepancy may be due to different mechanism for particle production at intermediate $p_T$. The particle production at intermediate $p_T$ will be discussed later in this chapter.\\ \begin{figure}[tbh] \begin{minipage}{0.49\textwidth} \includegraphics[width=0.95\textwidth,angle=-90]{auau_reactionplane.eps} \end{minipage}\hfill\hspace{2.5cm} \begin{minipage}{0.49\textwidth}\vspace{-0.8cm} \includegraphics[width=0.80\textwidth]{QMproc_highlight_4_Kai.eps} \end{minipage} \caption{(a) Azimuthal distribution of particles with respect to a trigger particle for p+p collisions (solid line), and mid-central Au+Au collisions within the reaction plane (squares) and out-of-plane (circles) at 200 GeV~\cite{Aihong}. (b) Mean transverse momentum for particles around the away-side region as a function of number of charged particles~\cite{Fuqiang}. The solid line shows the mean transverse momentum of inclusive hadrons.} \label{jetplot} \end{figure} A more differential probe of parton energy loss is the measurement of high $p_T$ di-hadron azimuthal correlation relative to the reaction plane orientation. The trigger hadron is in the range $4<p_{T}<6$ GeV/c and the associated particle is at $2<p_{T}<4$ GeV/c. Figure~\ref{jetplot} (left) shows the high $p_T$ di-hadron correlation when the trigger particle is selected in the azimuthal quadrants centered either in the reaction plane (in plane) or orthogonal to it(out of plane). The near side di-hadron azimuthal correlations in both cases were observed to be the same as that in p+p collisions, while the suppression of back to back correlation shows strong dependence on the relative angle between the triggered high $p_T$ hadron and the reaction plane. This systematic dependence is consistent with the picture of parton energy loss: the path length for a dijet oriented out of plane is longer than that for a dijet oriented in plane, leading to a stronger suppression of parton energy loss in the out of plane. The dependence of parton energy loss on the path length is predicted to be substantially larger than linear~\cite{jetquench}.\\ The energy lost by away side partons traversing the collision matter must in the form of the excess of softer emerging particles due to the transverse momentum conservation. An analysis of azimuthal correlations between soft and hard particles has been performed for both 200 GeV p+p and Au+Au collisions~\cite{Fuqiang} at STAR as a first of attempt to trace the degree of the degradation on the away side. With triggered hadron still in the range $4<p_{T}^{trig}<$ 6 GeV/c, but the associated hadrons now sought over $0.15<p_{T}<4$ GeV/c, combinatorial coincidences dominate this correlation and they must be subtracted carefully by mixed-event technique and also the elliptic flow effect was also subtracted by hand~\cite{Fuqiang}. The results demonstrate that, in comparison with the p+p and peripheral Au+Au collisions, the momentum-balancing hadrons opposite to the high $p_T$ triggered particle in central Au+Au are greater in number, much more widely dispersed in azimuthal angle, and significantly softer in momentum. Figure~\ref{jetplot} (right) shows the $\langle p_{T} \rangle$ of the momentum-balancing hadrons opposite to the high $p_T$ trigger as a function of centrality. The $\langle p_{T}\rangle$ were observed to decrease from peripheral to central Au+Au collisions. Also shown in the Figure~\ref{jetplot} (right) is the $\langle p_{T}\rangle$ of the inclusive hadrons as a function of centrality. This study will be extended to higher $p_T$ trigger particle. The results may suggest that the moderately hard parton traversing a significant path length through the collision matter makes substantial progress toward equilibrium with the bulk. The rapid attainment of thermalization via multitude of softer parton-parton interactions in the earliest collision stages would then not be so surprising~\cite{starwhitepaper}. \subsection{Particle composition in Au+Au at intermediate $p_{T}$} As we have mentioned above, for $R_{AA}$, the pQCD model including the parton energy loss, Cronin enhancement and nuclear shadowing can qualitatively fit the trend of data at $2<p_{T}<6$ GeV/c, however, the quantitative discrepancy between the model and the data is also obvious. In the intermediate $p_{T}$, the mechanism for particle production may be different from that at high $p_{T}$.\\ Figure~\ref{phenixAuAuspectra} (left) shows the $\pi$, K, p spectra in 0\%-5\% and 60\%-92\% Au+Au 200 GeV collisions from~\cite{ex0307022}. It shows that the shapes of the spectra show clear mass dependence. And in central collisions, the $\pi$, K, p yields are close to each other at $p_{T}>2$ GeV/c while it's not the case in peripheral collisions. Figure~\ref{phenixAuAuspectra} (right) shows proton/pion (top) and anti-proton/pion (bottom) ratios for central 0--10\%, mid-central 20--30\% and peripheral 60--92\% in Au+Au collisions at 200 GeV~\cite{ex0307022}. It shows that the $p(\bar{p})/\pi$ ratios increase fast from peripheral to central collisions. In the 0-10\% centrality bin, the proton yield is even larger than pion yield at intermediate $p_T$. Figure~\ref{KsLamRcp} shows the ratio $R_{CP}$ for identified mesons and baryons at mid-rapidity calculated using centrality intervals, 0--5\% vs. 40--60\% of the collision cross section from STAR measurement~\cite{starhighlight}. It seems that for meson, the $R_{CP}$ follows a common trend and for baryon, the $R_{CP}$ also follows a common trend, which is different from that for mesons. The $R_{CP}$ for baryons is observed to be larger than that for mesons.\\ \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=16pc,width=16pc]{pt_spectra_all.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=16pc,width=16pc]{ppi_ratio_all.eps} \end{minipage} \caption{(left) The $\pi$, K, p spectra in 0\%-5\% and 60\%-92\% Au+Au 200 GeV collisions from~\cite{ex0307022}. (right) Proton/pion (top) and anti-proton/pion (bottom) ratios for central 0--10\%, mid-central 20--30\% and peripheral 60--92\% in Au+Au collisions at 200 GeV. Open (filled) points are for charged (neutral) pions. The data at $\sqrt{s} = 53 $~GeV p+p collisions~\cite{ISR} are also shown. The solid line is the $(\bar{p} + p)/(\pi^{+} + \pi^{-})$ ratio measured in gluon jets~\cite{DELPHI}. This figure is from~\cite{ex0307022}. } \label{phenixAuAuspectra} \end{figure} These experimental results suggest that the degree of suppression depends on particle species(baryon/meson) at intermediate $p_T$. The spectra of baryons (protons and lambdas) are less suppressed than those of mesons (pions, kaons) ~\cite{starv2raa,phenixpid} in the $p_{T}$ range $2<p_{T}<5$ GeV/c. The baryon content in the hadrons at intermediate $p_{T}$ depends strongly on the impact parameter (centrality) of the Au+Au collisions with about 40\% of the hadrons being baryons in the minimum-bias collisions and 20\% in very peripheral collisions~\cite{starv2raa,phenixpid}. Hydrodynamics~\cite{derekhydro,pisahydro}, parton coalescence at hadronization~\cite{hwa,fries,ko} and gluon junctions~\cite{junction} have been suggested as explanations for the observed particle-species dependence. \begin{figure}[tbph] \centering\mbox{ \includegraphics[width=0.9\textwidth]{plot_rcp_24jan04_1.eps}} \caption{ The ratio $R_{CP}$ for identified mesons and baryons at mid-rapidity calculated using centrality intervals, 0--5\% vs. 40--60\% of the collision cross section. The bands represent the uncertainties in the model calculations of $\mathrm{N_{bin}}$. We also show the charged hadron $R_{CP}$ measured by STAR for $\sqrt{s_{_{NN}}}=200$~GeV~\cite{starhighpt}. This figure is from~\cite{starhighlight}.} \label{KsLamRcp} \end{figure} In these models, recombination/coalescence models successfully reproduce $R_{AA}$ of baryons and mesons at intermediate $p_T$, as well as showing consistency with the $v2$ measurement in the same $p_T$ range. \subsubsection{Recombination model} The concept of quark recombination was introduced to describe hadron production at forward rapidity in p+p collisions~\cite{ref50:whitepaper}. At forward rapidity, this mechanism allows a fast quark resulting from a hard parton scattering to recombine with a slow anti-quark, which could be one in the original sea in the incident hadron, or one incited by a gluon~\cite{ref50:whitepaper}. If a QGP is formed in the relativistic heavy ion collisions, then one might expect coalescence of the abundant thermal partons to provide another important hadron production mechanism, active over a wide range of rapidity and transverse momentum~\cite{ref51:whitepaper}. In particular, at moderate $p_T$ values(above the realm of hydrodynamics applicability), the hadron production from recombination of lower $p_T$ partons from thermal bath~\cite{hwa,fries,ko} has been predicted to be competitive with the production from fragmentation of higher $p_T$ scattered partons. It has been suggested~\cite{ref53:whitepaper} that the need for substantial recombination to explain the observed hadron yield and flow may be taken as a signature of QGP formation.\\ In order to explain the features of RHIC collisions, the recombination models~\cite{ref51:whitepaper,hwa,fries,ko} make the central assumption that coalescence proceeds via constituent quarks, whose number in a hadron determines its production rate. The constituent quarks are presumed to follow a thermal (exponential) momentum spectrum and to carry a collective transverse velocity distributions. This picture leads to clear predicted effects on baryon and meson production rates, with the former depending on the spectrum of thermal constituent quarks and antiquarks at roughly one-third the baryon $p_T$, and the latter determined by the spectrum at roughly one-half the meson $p_T$. Indeed, the recombination model was recently was re-introduced at RHIC context, precisely to explain the abnormal abundance of baryon vs meson observed at intermediate $p_T$~\cite{hwa,fries,ko}. If the observed saturated elliptic flow values of hadrons in this momentum range result from coalescence of collectively flowing constituent quarks, then one expect a similarly simple baryon vs meson relationship~\cite{hwa,fries,ko}: the baryon (meson) flow would be 3 (2) times the quark flow at roughly one-third (one-half) the baryon (meson) $p_T$~\cite{starwhitepaper}. \subsection{Summary} In summary, the several important results from RHIC have been introduced. The elliptic flow $v2$ can be reproduced by hydrodynamics at low $p_T$. At intermediate $p_T$, $v2$ from data show saturation and deviate from hydrodynamical model predictions. At the same time, $v2$ from data show baryon or meson species dependence. High $p_T$ suppression can be reproduced by pQCD model and gluon saturation model. The gluon saturation model is also called color glass condensate model (CGC). The production rate dependence on baryon or meson species has been observed at intermediate $p_{T}$, which can be reproduced by the recombination model. \section{Cronin effect} \subsection{Why we need d+Au run at RHIC} In order to see the intermediate and high $p_T$ suppression is due to the final-state effect or initial state effect, the measurements from d+Au collisions will provide the essential proof. Since the initial state in d+Au collisions is similar to that in Au+Au collisions, and, it's believed that the quark-gluon plasma doesn't exist in d+Au collisions, the results from d+Au collisions will be very important for us to judge whether the quark-gluon plasma exists in Au+Au collisions or not and to understand the property of the dense matter created in Au+Au collisions. Besides, if the identified particle spectra in d+Au and p+p collisions are measured, they will not only provide the reference for those in Au+Au collisions at 200 GeV, but also provide a chance to see the mechanism of the Cronin effect itself clearly at 200 GeV. Cronin effect was observed 30 years ago experimentally and the study of this effect was only limited to lower energy fixed target experiments. Before we go to the d+Au collisions, let's look back on the p+A collisions at lower energy fixed target experiment. \subsection{Lower energy} The hadron $p_{T}$ spectra have been observed to depend on the target atomic weight ($A$) and the produced particle species in lower energy p+A collisions~\cite{cronin}. This is known as the ``Cronin Effect'', a generic term for the experimentally observed broadening of the transverse momentum distributions at intermediate $p_{T}$ in p+A collisions as compared to those in p+p collisions~\cite{cronin,petersson83,accardi}. The effect can be characterized as a dependence of the yield on the target atomic weight as $A^{\alpha}$. At energies of $\sqrt{s} \simeq$ 30 GeV, $\alpha$ depends on $p_{T}$ and is greater than unity at high $p_{T}$~\cite{cronin}, indicating an enhancement of the production cross section. As shown in Figure~\ref{poweralphaplot}, the $\alpha$ is larger than 1 in the intermediate $p_{T}$ and shows strong particle-species dependence. The $\alpha$ for proton and antiproton are larger than those for kaon and pion. And $\alpha$ for kaon is larger than that for pion. This effect has been interpreted as partonic scatterings at the initial impact~\cite{petersson83,accardi}. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=18pc]{poweralpha.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=18pc,width=18pc]{R_woverbe_energy.eps} \end{minipage} \caption{(left) The power alpha of A dependence from 300 GeV incident proton-fixed target experiment. This figure is from~\cite{cronin}. (right) The Cronin ratio $R_{W/B_{e}}$ at $p_{T}=4.61$ GeV/c versus energy. This plot is from~\cite{cronin}.} \label{poweralphaplot} \end{figure} Besides, the lower energy data suggest the power $\alpha$ decreases with energy, as shown in Figure~\ref{poweralphaplot}. However, the energy dependence study of Cronin effect is limited to fixed target experiment at lower energy. What's the extrapolation of Cronin effect at higher energy such as RHIC energy 200 GeV. At higher energies, multiple parton collisions are possible even in p+p collisions~\cite{e735kno}. This combined with the hardening of the spectra with increasing beam energy would reduce the Cronin effect~\cite{accardi}. There are several models which give different predictions of Cronin effect at 200 GeV. \subsection{Predictions: RHIC energy} One of the models is the initial multiple parton scattering model. In this model, the transverse momentum of the parton inside the proton will be broadened when the proton traverses the Au nucleus due to the multiple scattering between the proton and the nucleons inside the Au nucleus. In these models, the Cronin ratio will increases to a maximum value between 1 and 2 at 2.5$<p_T<$4.5 GeV/c and then decreases with $p_T$ increasing~\cite{accardi}. The Cronin effect is predicted to be larger in central d+Au collisions than in d+Au peripheral collisions~\cite{Vitev03}. Another model is the gluon saturation model. At sufficiently high beam energy, gluon saturation is expected to result in a relative suppression of hadron yield at high $p_{T}$ in both p+A and A+A collisions and in a substantial decrease and finally in the disappearance of the Cronin effect~\cite{cgc}. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{dAu_Fig3.eps} \caption{ $R_{AB}$ for minimum bias and central d+Au collisions, and central Au+Au collisions~\cite{starhighpt}. The minimum bias d+Au data are displaced 100 MeV/c to the right for clarity. The bands show the normalization uncertainties, which are highly correlated point-to-point and between the two d+Au distributions. This Figure is from~\cite{stardau}.} \label{FigThree} \end{figure} Figure~\ref{FigThree} shows the $R_{dAu}$ of charged hadron vs $p_{T}$ from STAR. We can see that the Cronin ratio increases to a maximum value around 1.5 at $3<p_{T}<4$ GeV/c and then decreases again~\cite{stardau}. This is consistent with the initial multiple parton scattering model~\cite{accardi}. These results on inclusive hadron production from d+Au collisions indicate that hadron suppression at intermediate and high $p_{T}$ in Au+Au collisions is due to final state interactions in a dense and dissipative medium produced during the collision and not due to the initial state wave function of the Au nucleus~\cite{stardau,otherdau}.\\ Now we know that the hadron suppression at intermediate $p_{T}$ in Au+Au collisions is due to final-state effects~\cite{stardau,otherdau}. What's the effect on particle composition at the same $p_{T}$ range in Au+Au collisions? Another question is whether there is any Cronin effect dependence on particle-species in d+Au collisions or not. In order to further understand the mechanisms responsible for the particle dependence of $p_{T}$ spectra in heavy ion collisions, and to separate the effects of initial and final partonic rescatterings, we measured the $p_{T}$ distributions of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ from 200 GeV d+Au and p+p collisions. In this thesis, we discuss the dependence of particle production on $p_{T}$, collision energy, and target atomic weight. And we compare the Cronin effect of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ with models to address the mechanism for Cronin effect in d+Au collisions at $\sqrt{s_{_{NN}}} = 200$ GeV. \chapter{{\hspace{3.5cm}How to make MRPC}} This appendix is based on the procedure of the MRPC production in USTC. I will introduce the material preparations and then the chamber installation. \section{Preparations} \subsection{Glass} (1) Check the glass very carefully by eye. The glass with scrapes is not accepted. (2) Measure the size of the glass with the digital vernier caliper. The errors of the length and width are required to be within 0.1 mm. Measure the thickness in several different places. The precision of the thickness is required to be 0.01 mm for each glass. (3) Use the micrometer to measure the flatness, which is required to be less than 0.01 mm. Use the mirror and observe the stripes of interference. (4) Grind the edge and the corner of the glass, and clean it. The size of outer glass is $78(width)\times 206(length) \times 1.1(height)$ $mm^3$ and the size of inner glass is $61\times 200\times 0.54$ $mm^3$. \subsection{Graphite Layer} (1) Stick the layer in the middle of the outer glass. Squeeze the air out. (2) Stick a small copper tape, which is for high voltage (HV) applying, on to the graphite layer, which is in the middle of the long side, and 0$\sim$0.5 mm away from the edge of the glass. The size of the graphite layer is $74\times202$ $mm^2$. The size of the copper tape (the rectangle with the round angle) is $6\times10$ $mm^2$. \subsection{Mylar layer} Cut the mylar layer, and see if there is any tiny holes or scrapes. If yes, don't use it. The size of mylar is $84\times212\times0.35$ $mm^3$. \subsection{Honeycomb board} Measure the size and flatness. The error of the length and width is required to be within 0.2 mm, the error of the thickness is required to be within 0.05 mm. The flatness is required to be within 0.1 mm. The size of honeycomb board is $84\times 208\times 4$ $mm^3$. \subsection{The printed circuit board (PCB)} (1) Check the surface of the metal which is used as read-out strips carefully, and see the position of the HV-holes is right or not. Check the size of the metal holes, whose diameters are required to be larger than 0.9 mm. (2) Use double side tape to stick the PCB board with the Mylar. The size of the double side tape is the same as the PCB board. The length of the mylar is 1 mm longer than that of the PCB board. (3) Use sealing ion to open a $\phi$ 3 mm hole, the center of the hole is in the middle of the HV holes. The size of PCB is $94\times 210\times 1.5$ $mm^3$. The size of metal holes are $\phi$ 1 mm. \subsection{Lucite cylinder} Use digital vernier caliper to measure the length of the Lucite cylinder. Clean it and stick a double side tape on one side. The size is $\phi$ 3 mm and $3.87<length<3.93$ mm. \subsection{Other stuff} Besides, we also need pins, fish line and little plastic cannula. The size of the pin is 2-2.1 cm long. The fish line is $\phi$ 0.22 mm. The plastic cannula is $\phi$ 1 mm. One kind of the cannula is 7 mm long, and the other is 5.6 mm long. Table~\ref{mrpcmaterial} lists the main materials for 1 MRPC. \section{Installation} \subsection{The outer glass and mylar and PCB} (1) stick the outer glass on to the center of the mylar. (2) Use the sealing ion to connect the HV conductive line with the copper tape. Apply the HV to measure the noise rate and dark current. (3) Between the mylar and outer glass edge, on each side, use silica gel to seal. Attention: keep the surface clean and smooth. Attention: If one side is done, wait till the silica gel becomes solid. (4) Stick pins. Seal the pins which are used for the fish line coiling, into the metal holes of the PCB board. (5) Use the inner glass to fix on the position of Lucite cylinders, and keep them away from the pins for fish line. Then stick the 8$\sim$10 Lucite cylinders onto the outer glass. \subsection{Inner glass and fish-line coiling} (1) Pre-install. Don't use fish-line. Pay attention to adjust the position of the pins. (2) This is now the real installation and fish line coiling. Clean the outer glass and inner glass carefully, coil a loop of fish line, add a piece of glass, then coil another loop of fish line, add another piece of glass, and so on and so forth. Attention: clean the fish line before it coils, and blow the surface of glass to protect it from the dirt with nitrogen jet. (3) Another time for pre-installation. Pay attention to the position of the upper and lower electrodes and adjust the position of pins. (4) Paste 3140 RTV coating onto the surface of Lucite cylinders. (5) Connect the two electrodes. Make sure all the pins connect right into the metal holes. Then lay the whole flat, and put on a block which is 4 kilogram weight. (6) After 2 hours, stick the honeycomb. (7) Measure the thickness of the whole. Make sure the precision is within 0.05 mm. (8) Connect the conductive-line for the read-out strips, and then put the whole into a bag. Attention: the conductive-line should not be broken. \begin{table*} \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|} \hline material& type & character & number & source \\ \hline outer glass& window glass & 1.1 mm thick, VR: $8.7\times 10^{12}$ ohm.cm & 2 & Shanghai \\ \hline inner glass&window glass & 0.54 mm thick, VR: $8.5\times 10^{12}$ ohm.cm & 5 & USA \\ \hline graphite layer& T9149& 0.13 mm thick, SR : 2M ohm/square & 2 & Japan \\ \hline Mylar& M0 & 0.35 mm thick& 2& Dupont Corp. \\ \hline honeycomb & & 4 mm thick & 2& Shanghai \\ \hline PCB & gold & 1.58 mm thick& 2 & Shenzhen \\ \hline copper tape & & 0.08 mm thick & 2 & 3M Comp. \\ \hline LC & Lucite& $\phi$ 3 mm,3.9 mm long & 8-10 & processing \\ \hline pins (single) &metal pin&21.5 mm long& 14 & \\ \hline pins (pair) &metal pin&21.5 mm long& 12 & \\ \hline cannula & F-plastic & $\phi$ 1.4 mm & 38 & \\ \hline fish line & top line & $\phi$ 0.22 mm & &Switzerland \\ \hline DST & 9690& 0.13 mm thick && 3M Comp. \\ \hline silica gel & CAF4 & high-voltage insulation & &Switzerland \\ \hline \end{tabular} \caption{The material for 1 MRPC model. VR is the volume resistivity and SR is the surface resistivity. LC is the Lucite cylinder. DST is the double side tape.} \label{mrpcmaterial} \end{scriptsize} \end{table*} \chapter{{\hspace{3.5cm}List of Publications}} \hspace{0.7cm}1. \emph{Pion, kaon, proton and anti-proton transverse momentum distributions from p+p and d+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0309012), submitted. 2. \emph{Open Charm Yields in 200 GeV p+p and d+Au Collisions at RHIC,} Lijuan Ruan (for the STAR Collaboration), Journal of Physics G, 30 (2004) S1197-S1200, contributed to 17th International Conference on Ultra Relativistic Nucleus-Nucleus Collisions (Quark Matter 2004). 3. \emph{A Monte Carlo Simulation of Multi-gap Resistive Plate Chamber and comparision with Experimental Results,} RUAN Li-Juan, SHAO Ming, CHEN Hong-Fang, {\it et al.}, HEP and NP, Vol. 27, No. 8 (2003) 712-715. 4. \emph{ Monte Carlo Study of the Property of Multi-gap Resistive Plate Chambers,} Shao Ming, Ruan Lijuan, Chen Hongfang, {\it et al.}, HEP and NP, Vol. 27, No. 1 (2003) 67-71, (in Chinese). 5. \emph{Study on Light Collection and its Uniformity of Long Lead Tungstate crystal by Monte Carlo Method,} Ruan Lijuan, Shao Ming, Xu Tong, {\it et al.}, Chinese Journal of Computational Physics, Vol. 19, No. 5 (2002) 453-458, (in Chinese). 6. \emph{Beam test results of two kinds of multi-gap resistive plate chambers,} M. Shao, L. J. Ruan, H. F. Chen, J. Wu, , C. Li, Z. Z. Xu, X. L. Wang, S.L. Huang, Z. M. Wang and Z. P. Zhang, Nucl. Instri. and Meth. A 492 (2002) 344-350. 7. \emph{The Study of the Resistive Property of the Electrode Material of MRPC,} Ruan Lijuan, Wang Xiaolian, Li Cheng , {\it et al.}, to be published in Journal of University of Science and Technology of China (in Chinese). 8. \emph{The Calibration Method of TOFr in the STAR Experiment,} RUAN Lijuan, WU Jian, DONG Xin , {\it et al.}, to be published in HEP and NP (in Chinese). 9. \emph{Spectra of $\pi$ K p $K^{*}$ $\phi$ from Au+Au Collisions at 62.4 GeV,} Lijuan Ruan (for the STAR Collaboration), to be published in Journal of Physics G, Contributed to 8th International Conference on Strangeness in Quark Matter (SQM 2004). 10. \emph{Pseudorapidity Asymmetry and Centrality Dependence of Charged Hadron Spectra in d+Au Collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration: e-Print Archives (nu-ex/0408016), submitted. 11. \emph{Transverse momentum correlations and minijet dissipation in Au-Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0408012), submitted. 12. \emph{Azimuthal anisotropy and correlations at large transverse momenta in p+p and Au+Au collisions at $\sqrt{s_{NN}}$= 200 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0407007), submitted. 13. \emph{Open charm yields in d+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0407006), submitted. 14. \emph{Measurements of transverse energy distributions in Au+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0407003), submitted. 15. \emph{Transverse-momentum dependent modification of dynamic texture in central Au+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0407001), submitted. 16. \emph{Hadronization geometry and charge-dependent number autocorrelations on axial momentum space in Au-Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, e-Print Archives(nu-ex/0406035), submitted. 17. \emph{Phi meson production in Au+Au and p+p collisions at sqrt(s)=200 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0406003), submitted. 18. \emph{Centrality and pseudorapidity dependence of charged hadron production at intermediate pT in Au+Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0404020), to be published in Physical Review C. 19. \emph{Production of e$+$e$-$ Pairs Accompanied by Nuclear Dissociation in Ultra-Peripheral Heavy Ion Collision,} STAR Collaboration, e-Print Archives (nu-ex/0404012), to be published in Physical Review C. 20. \emph{Photon and neutral pion production in Au+Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0401008), to be published in Physical Review C. 21. \emph{Azimuthally sensitive HBT in Au+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, Phys. Rev. Lett. 93, 012301 (2004). 22. \emph{Production of Charged Pions and Hadrons in Au+Au Collisions at $\sqrt{s_{NN}}$=130 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0311017), submitted. 23. \emph{Azimuthal anisotropy at RHIC: the first and fourth harmonics,} STAR Collaboration, Phys. Rev. Lett. 92, 062301 (2004). 24. \emph{Cross Sections and Transverse Single-Spin Asymmetries in Forward Neutral Pion Production from Proton Collisions at sqrt(s) = 200 GeV,} STAR Collaboration, Phys. Rev. Lett. 92, 171801 (2004). 25. \emph{Identified particle distributions in pp and Au+Au collisions at sqrt{snn}=200 GeV,} STAR Collaboration, Phys. Rev. Lett. 92, 112301 (2004). 26. \emph{Event-by-Event (pt) fluctuations in Au-Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0308033), submitted. 27. \emph{Multi-strange baryon production in Au-Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, Phys. Rev. Lett. 92, 182301 (2004). 28. \emph{Pion-Kaon Correlations in Central Au+Au Collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, Phys. Rev. Lett. 91, 262302 (2003). 29. \emph{rho-0 Production and Possible Modification in Au+Au and p+p Collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, Phys. Rev. Lett. 92, 092301 (2004). 30. \emph{Net charge fluctuations in Au+Au collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, Phys. Rev. C 68, 044905 (2003). 31. \emph{Rapidity and Centrality Dependence of Proton and Anti-proton Production from Au+Au Collisions at $\sqrt{s_{NN}}$ = 130 GeV,} STAR Collaboration, e-Print Archives (nu-ex/0306029), submitted. 32. \emph{Three-Pion Hanbury Brown-Twiss Correlations in Relativistic Heavy-Ion Collisions from the STAR Experiment,} STAR Collaboration, Phys. Rev. Lett. 91, 262301 (2003). 33. \emph{Evidence from d+Au measurements for final-state suppression of high pT hadrons in Au+Au collisions at RHIC,} STAR Collaboration, Phys. Rev. Lett. 91, 072304 (2003). 34. \emph{Particle-type dependence of azimuthal anisotropy and nuclear modification of particle production in Au+Au collisions at $\sqrt{s_{NN}}$ = 200 GeV,} STAR Collaboration, Phys. Rev. Lett. 92, 052302 (2004). 35. \emph{Transverse momentum and collision energy dependence of high pT hadron suppression in Au+Au collisions at ultrarelativistic energies,} STAR Collaboration, Phys. Rev. Lett. 91, 172302 (2003). \chapter{Results} \label{chp:results} \section{$\pi, K, p$ and $\bar{p}$ spectra in d+Au and p+p collisions at mid-rapidity} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{spectra_dAu_TOFr1215.eps} \caption{The invariant yields of pions (filled circles), kaons (open squares), protons (filled triangles) and their anti-particles as a function of $p_{T}$ from d+Au and NSD p+p events at 200 GeV. The rapidity range was $-0.5<y<0.0$ with the direction of the outgoing Au ions as negative rapidity. Errors are statistical.} \label{spectra} \end{figure} The invariant yields $\frac{1}{2\pi p_T}\frac{d^2N}{dydp_T}$ of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ from both NSD p+p and minimum-bias d+Au events at mid-rapidity $-0.5<y<0$ are shown in Figure\ref{spectra}, where $N$ is the corrected signal number per minimum-bias event in each $p_{T}$ bin. $N=\frac{N_{raw}\times{factor3}\times{factor4}\times{factor5}}{N_{total}\times{factor1}\times{factor2}}$, where $N_{raw}$ is the raw signal number in each $p_{T}$ bin, $N_{total}$ is the total TOFr triggered events, $factor1$ is the enhancement factor of TOFr trigger, $factor2$ is the TPC efficiency times TOFr matching efficiency, $factor3$ is the background correction factor, $factor4$ is the $\langle N_{ch} \rangle$ bias factor, and $factor5$ is the vertex efficiency times trigger efficiency and normalization factor. \subsection{Systematic uncertainty} For the invariant yield of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$, the average bin-to-bin systematic uncertainty was estimated to be of the order of 8\%. The systematic uncertainty is dominated by the uncertainty in the detector response in Monte Carlo simulations ($\pm7\%$). Additional factors contributing to the total systematic uncertainty include the background correction ($\pm3\%$), the small $\eta$ acceptance of the TOFr ($\pm2\%$), TOFr response ($\pm2\%$), the correction for energy loss in the detector (${}^{<}_{\sim}10\pm10\%$ at $p_{T}<0.6$ GeV/c for the $p$ and $\bar{p}$, much smaller for other species and negligible at higher $p_{T}$), absorption of $\bar{p}$ in the material ($\pm3\%$), and the momentum resolution correction ($\simeq5\pm2\%$). The normalization uncertainties in d+Au minimum-bias and p+p NSD collisions are $10\%$ and $14\%$, respectively~\cite{starhighpt,stardau}. The charged pion yields are consistent with $\pi^0$ yields measured by the PHENIX collaboration in the overlapping $p_{T}$ range~\cite{phenixhighpt,otherdau}. The invariant yields of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ in minimum-bias, centrality selected d+Au and minimum-bias p+p collisions, are listed in the tables in Appendix A with statistical errors and systematic uncertainties. \section{Cronin effect} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{pidratio_dAu_TOFr1216.eps} \caption{The identified particle $R_{dAu}$ for minimum-bias and top 20\% d+Au collisions. The filled triangles are for $p+\bar{p}$, the filled circles are for $\pi^{+}+\pi^{-}$ and the open squares are for $K^{+}+K^{-}$. Dashed lines are $R_{dAu}$ of inclusive charged hadrons from~\cite{stardau}. The open triangles and open circles are $R_{CP}$ of $p+\bar{p}$ and $\pi^{0}$ in Au+Au collisions measured by PHENIX~\cite{phenixpid}. Errors are statistical. The gray band represents the normalization uncertainty of 16\%.} \label{Rdau} \end{figure} Nuclear effects on hadron production in d+Au collisions are measured through comparison to the p+p spectrum, scaled by the number of underlying nucleon-nucleon inelastic collisions using the ratio \[R_{dAu}=\frac{d^{2}N/(2{\pi}p_{T}dp_{T}dy)}{T_{dAu}d^{2}\sigma^{pp}_{inel}/(2{\pi}p_{T}dp_{T}dy)} ,\] where $T_{dAu}={\langle N_{bin}\rangle}/\sigma^{pp}_{inel}$ describes the nuclear geometry, and $d^{2}\sigma^{pp}_{inel}/(2{\pi}p_{T}dp_{T}dy)$ for p+p inelastic collisions is derived from the measured p+p NSD cross section. The difference between NSD and inelastic differential cross sections at mid-rapidity, as estimated from PYTHIA~\cite{pythia}, is $5\%$ at low $p_{T}$ and negligible at $p_{T}>1.0$ GeV/c. Figure.~\ref{Rdau} shows $R_{dAu}$ of $\pi^{+}+\pi^{-}$, $K^{+}+K^{-}$ and $p+\bar{p}$ for minimum-bias and central d+Au collisions. The systematic uncertainties on $R_{dAu}$ are of the order of 16\%, dominated by the uncertainty in normalization. The $R_{dAu}$ of the same particle species are similar between minimum-bias and top 20\% d+Au collisions. In both cases, the $R_{dAu}$ of protons rise faster than $R_{dAu}$ of pions and kaons. We observe that the spectra of $\pi^{\pm}$, $K^{\pm}$, $p$ and $\bar{p}$ are considerably harder in d+Au than those in p+p collisions. The $R_{dAu}$ of the identified particles has characteristics of the Cronin effect~\cite{cronin,accardi} in particle production with $R_{dAu}$ less than unity at low $p_{T}$ and above unity at $p_{T}{}^{>}_{\sim} 1.0$ GeV/c. \section{$p+\bar{p}/h$ ratio in d+Au and p+p collisions at middle pseudo-rapidity} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{baryon_nch_cronin_TOFr0910.eps} \caption{Minimum-bias ratios of ($p+\bar{p}$) over charged hadrons at $-0.5\!<\!\eta\!<\!0.0$ from $\sqrt{s_{_{NN}}} =200$ GeV p+p (open diamonds), d+Au (filled triangles) and $\sqrt{s_{_{NN}}} =130$ GeV Au+Au~\cite{phenixpid} (asterisks) collisions. Results of $\mathrm{p+\bar{p}}$ collisions at $\sqrt{s_{_{NN}}} = 1.8$ TeV~\cite{e735} are shown as open stars. Dashed lines are results of $p/h^{+}$ ratios from $\sqrt{s_{_{NN}}} = 23.8$ GeV p+p (short-dashed lines) and p+W (dot-dashed) collisions~\cite{cronin}. Errors are statistical. } \label{bnchratio} \end{figure} Figure~\ref{bnchratio} depicts $(p+\bar{p})/h$, the ratio of $p+\bar{p}$ over inclusive charged hadrons as a function of $p_{T}$ in d+Au and p+p minimum-bias collisions at $\sqrt{s_{_{NN}}} = 200$ GeV, and $p/h^{+}$ ratios in p+p and p+W minimum-bias collisions at $\sqrt{s_{_{NN}}} = 23.8$ GeV~\cite{cronin}. Although the relative yields of particles and anti-particles are very different at $\sqrt{s}<40$ GeV due to the valence quark effects from target and projectile, the Cronin effects are similar. The systematic uncertainties on these ratios were estimated to be of the order of 10\% for $p_{T}{}^{<}_{\sim}1.0$ GeV/c, decreasing to 3\% at higher $p_{T}$. At RHIC energies, the anti-particle to particle ratios approach unity ($\bar{p}/p=0.81\pm0.02\pm0.04$ in d+Au minimum-bias collisions) and their nuclear modification factors are similar. The difference between $R_{dAu}$ at $\sqrt{s_{_{NN}}} = 200$ GeV for $p+\bar{p}$ and $h$ can be obtained from the $(p+\bar{p})/h$ ratios in d+Au and p+p collisions. Table~\ref{pbarpnchratio} shows $R_{dAu}^{p+\bar{p}}/R_{dAu}^{h}$ determined by averaging over the bins within $1.2<p_{T}<3.0$ GeV/c. At lower energy, the $\alpha$ parameter in the power law dependence on target atomic weight $A^{\alpha}$ of identified particle production falls with $\sqrt{s}$~\cite{cronin}. From the ratios of $R_{dAu}$ between $p+\bar{p}$ and $h$, we may further derive the $\alpha_{p}-\alpha_{\pi}$ for $1.2< p_{T}< 3.0$ GeV/c to be $0.041\pm0.010$(stat)$\pm0.006$(syst) under the assumptions that $\alpha_{K}\simeq\alpha_{\pi}$ and that $(p+\bar{p})/{\pi}$ and $K/{\pi}$ are between 0.1 and 0.4 in p+p collisions. This result is significantly smaller than the value $0.095\pm0.004$ in the same $p_{T}$ range found at lower energies~\cite{cronin}.\\ \begin{table}[h] \caption{\label{pbarpnchratio}$\langle N_{bin}\rangle$ from a Glauber model calculation, $(p+\bar{p})/h$ averaged over the bins within $1.2<p_{T}<2.0$ GeV/c (left column) and within $2.0<p_{T}<3.0$ GeV/c (right column) and the $R_{dAu}$ ratios between $p+\bar{p}$ and $h$ averaged over $1.2<p_{T}<3.0$ GeV/c for minimum-bias, centrality selected d+Au collisions and minimum-bias p+p collisions. A p+p inelastic cross section of $\sigma_{inel}=42$ mb was used in the calculation. For $R_{dAu}$ ratios, only statistical errors are shown and the systematic uncertainties are 0.03 for all centrality bins. } {\centering {\begin{tabular}{c|c|c|c|c} \hline \hline centrality & $\langle N_{bin}\rangle$ & \multicolumn{2}{c|} {$(p+\bar{p})/h$} & ${R_{dAu}^{p+\bar{p}}}/{R_{dAu}^h}$\\ \hline min. bias & $7.5\pm0.4$ &$0.21\pm0.01$ &$0.24\pm0.01$ & $1.19\pm0.05$\\ 0--20\% & $15.0\pm1.1$ &$0.21\pm0.01$ &$0.24\pm0.02$ & $1.18\pm0.06$\\ 20--40\% & $10.2\pm1.0$ &$0.20\pm0.01$ &$0.24\pm0.02$ & $1.16\pm0.06$\\ 40--$\sim$100\% & $4.0^{+0.8}_{-0.3}$ &$0.20\pm0.01$ &$0.23\pm0.02$ & $1.13\pm0.06$\\ \hline p+p & $1.0$ &$0.17\pm0.01$ &$0.21\pm0.02$ & --- \\ \hline \hline \end{tabular} } \par} \label{Tab:D} \end{table} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{baryon_nch_cronin_TOFr_centrality_1023_color_new.eps} \caption{Minimum-bias ratios of ($p+\bar{p}$) over charged hadrons at $-0.5\!<\!\eta\!<\!0.0$ from $\sqrt{s_{_{NN}}} =200$ GeV minimum-bias and centrality selected d+Au collisions. Errors are statistical. } \label{bnchratiocentrality} \end{figure} Also shown is $(p+\bar{p})/h$ ratio from the Au+Au minimum-bias collisions at $\sqrt{s_{_{NN}}} = 130$ GeV~\cite{phenixpid}. The $(p+\bar{p})/h$ ratio from minimum-bias Au+Au collisions~\cite{phenixpid} at a similar energy is about a factor of 2 higher than that in d+Au and p+p collisions for $p_{T}{}^{>}_{\sim}2.0$ GeV/c. This enhancement is most likely due to final-state effects in Au+Au collisions~\cite{jetquench,junction,derekhydro,pisahydro,fries,ko}. The ratios show little centrality dependence in d+Au collisions, as shown in Table~\ref{Tab:D} and Figure~\ref{bnchratiocentrality}. For $p_{T}<2.0$ GeV/c, the ratio in $\mathrm{p+\bar{p}}$ collisions at $\sqrt{s_{_{NN}}} = 1.8$ TeV~\cite{e735} is very similar to those in d+Au and p+p collisions at $\sqrt{s_{_{NN}}} = 200$ GeV. \section{$K/\pi$, $p/\pi$ and anti-particle to particle ratios} \begin{figure}[h] \begin{minipage}[t]{50mm} \includegraphics[height=11pc,width=13pc]{piratio.eps} \end{minipage} \begin{minipage}[t]{50mm} \includegraphics[height=11pc,width=13pc]{kratio.eps} \end{minipage} \begin{minipage}[t]{50mm} \includegraphics[height=11pc,width=13pc]{pratio.eps} \end{minipage} \caption{$\pi^{-}$/$\pi^{+}$, $K^{-}$/$K^{+}$ and $\bar{p}/p$ ratios as a function of $p_{T}$ in d+Au and p+p minimum-bias collisions. The open symbols are for p+p collisions and the solid symbols for d+Au collisions. Errors are statistical.} \label{antiparticleratio} \end{figure} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline Centrality Bin & $\pi^{-}$/$\pi^{+}$ & $X^{2}/ndf$ & $K^{-}$/$K^{+}$ & $X^{2}/ndf$ & $\bar{p}/p$ & $X^{2}/ndf$ \\ \hline d+Au M.B. & $1.01\pm0.01$ & $0.88$ & $0.94\pm0.02$ & $1.78$ & $0.81\pm0.02$ & $0.85$\\ \hline 0\%-20\% & $1.01\pm0.01$ & $0.80$ & $0.93\pm0.03$ & $1.43$ & $0.80\pm0.03$ & $0.70$ \\ \hline 20\%-40\% & $1.00\pm0.01$ & $0.98$ & $0.91\pm0.03$ & $1.19$ & $0.79\pm0.03$ & $1.14$ \\ \hline 40\%-100\% & $1.02\pm0.01$ & $0.81$ & $1.02\pm0.03$ & $0.45$ & $0.78\pm0.03$ & $0.70$ \\ \hline p+p & $1.00\pm0.01$ & $1.24$ & $0.98\pm0.02$ & $0.71$ & $0.79\pm0.03$ & $0.73$ \\ \hline \end{tabular} \caption{$\pi^{-}$/$\pi^{+}$, $K^{-}$/$K^{+}$ and $\bar{p}/p$ ratios in p+p and d+Au minimum-bias collisions. Also shows in the table are the ratios in centrality selected d+Au collisions. Errors are statistical. } \label{antiparticletoparticleratio} \end{scriptsize} \end{table} Figure~\ref{antiparticleratio} shows the $\pi^{-}/\pi^{+}$, $K^{-}/K^{+}$ and $\bar{p}/p$ ratios as a function of $p_{T}$ in d+Au and p+p minimum-bias collisions. It shows the anti-particle to particle ratios are flat with $p_{T}$. The zero order polynominal function was used to fit the data and get the anti-particle to particle ratios. The results are list in Table~\ref{antiparticletoparticleratio}. In centrality selected d+Au collisions, the anti-particle to particle ratios are also flat with $p_{T}$ and show little centrality dependence. The results are also shown in the Table ~\ref{antiparticletoparticleratio}. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=14pc]{kppipratio.eps} \end{minipage} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=14pc]{kmpimratio.eps} \end{minipage} \caption{$K/\pi$ ratios as a function of $p_{T}$ in d+Au and p+p minimum-bias collisions. The open symbols are for p+p collisions and the solid symbols for d+Au collisions. Errors are statistical.} \label{kpiratio} \end{figure} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=14pc]{pppipratio.eps} \end{minipage} \begin{minipage}[t]{80mm} \includegraphics[height=13pc,width=14pc]{pmpimratio.eps} \end{minipage} \caption{$p(\bar{p})/\pi$ ratios as a function of $p_{T}$ in d+Au and p+p minimum-bias collisions. The open symbols are for p+p collisions and the solid symbols for d+Au collisions. Errors are statistical.} \label{ppiratio} \end{figure} The $K/\pi$ and $p/\pi$ ratios are shown in Figure~\ref{kpiratio} and Figure~\ref{ppiratio} individually. From the plots, the $K/\pi$ ratios increase with $p_{T}$ in both d+Au and p+p collisions and the increasing trend is the same within our errors. The $p/\pi$ ratios increase with $p_{T}$ in both d+Au and p+p collisions and the increasing in d+Au collisions is faster than that in p+p collisions. The trends of the $K/\pi$ and $p/\pi$ as a function of $p_{T}$ show little centrality dependence in d+Au collisions. \section{$dN/dy$, $\langle p_T \rangle$, and model fits} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{piplusyield.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{piminusyield.eps} \end{minipage} \caption{The re-scaled $\pi^{+}$ and $\pi^{-}$ spectra in minimum-bias, centrality selected d+Au collisions and also in p+p collisions. The errors are statistical.} \label{pionspectra} \end{figure} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{kplusyield.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{kminusyield.eps} \end{minipage} \caption{The re-scaled $K^{+}$ and $K^{-}$ spectra in minimum-bias, centrality selected d+Au collisions and also in p+p collisions. The errors are statistical.} \label{kaonspectra} \end{figure} The spectra in minimum-bias and centrality selected d+Au collisions and also in p+p collisions are shown in Figure~\ref{pionspectra}, Figure~\ref{kaonspectra} and Figure~\ref{protonspectra}. The spectra show little centrality dependence for each particle in d+Au collisions but harder than those in p+p collisions. The power law function was used to fit the spectra and get the $dN/dy$ and $\langle p_T \rangle$. The power law fit function is: \begin{equation} \frac{1}{2\pi p_T}\frac{d^2N}{dydp_T}=a(1+\frac{p_T}{\langle p_T \rangle \frac{n-3}{2}})^{-n} \end{equation} Where the parameter $a$ is a constant value proportional to the mid-rapidity yield $dN/dy$, the parameter $n$ is the order of the power law and $\langle p_T \rangle$ is the mean value of the transverse momentum which is extracted from the fit. Figure~\ref{powerlawfit} shows power law fit to the spectra of minimum-bias d+Au and p+p collisions. Figure~\ref{3centralitypowerlawfit} shows power law fit to the spectra of 3 centrality selected d+Au collisions. The power law fit results are listed in Table~\ref{dndypowerlawfit} and Table~\ref{meanptpowerlawfit} individually. The thermal model~\cite{thermal} was also used to fit the spectra. The final $dN/dy$ and $\langle p_T \rangle$ are shown in Table~\ref{finaldndy} and Table~\ref{finalmeanpt} respectively, which were obtained by averaging the results from the power law fit and thermal fit. Half of the differences in them are taken as the systematic errors due to the extrapolation to low $p_{T}$ region. The errors in this table include the systematic uncertainties and statistical errors. \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline Centrality Bin & $\pi^{-}$ & $\pi^{+}$ & $K^{-}$ & $K^{+}$ & $\bar{p}$ & $p$ \\ \hline d+Au M.B. & $0.403\pm0.004$ & $0.405\pm0.004$ & $0.609\pm0.009$ & $0.629\pm0.009$ & $0.714\pm0.008$ & $0.677\pm0.010$\\ \hline 0\%-20\% & $0.421\pm0.004$ & $0.421\pm0.004$ & $0.626\pm0.018$ & $0.658\pm0.016$ & $0.727\pm0.013$ & $0.705\pm0.014$ \\ \hline 20\%-40\% & $0.408\pm0.004$ & $0.411\pm0.004$ & $0.604\pm0.015$ & $0.625\pm0.015$ & $0.725\pm0.013$ & $0.691\pm0.015$ \\ \hline 40\%-100\% & $0.387\pm0.005$ & $0.391\pm0.004$ & $0.589\pm0.016$ & $0.616\pm0.016$ & $0.667\pm0.013$ & $0.646\pm0.014$ \\ \hline p+p & $0.357\pm0.004$ & $0.361\pm0.004$ & $0.571\pm0.013$ & $0.571\pm0.013$ & $0.567\pm0.010$ & $0.569\pm0.012$ \\ \hline \end{tabular} \caption{$\langle p_T \rangle$ of $\pi^{-}$, $\pi^{+}$, $K^{-}$, $K^{+}$, $\bar{p}$ and $p$ from power law fit in minimum-bias, centrality selected d+Au collisions and also in p+p collisions. The errors are from the power law fit. The unit of $p_{T}$ is GeV/c.} \label{meanptpowerlawfit} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline Centrality Bin & $\pi^{-}$ & $\pi^{+}$ & $K^{-}$ & $K^{+}$ & $\bar{p}$ & $p$ \\ \hline d+Au M.B. & $5.078\pm0.080$ & $5.032\pm0.080$ & $0.685\pm0.013$ & $0.703\pm0.012$ & $0.466\pm0.009$ & $0.594\pm0.019$\\ \hline 0\%-20\% & $10.657\pm0.190$ & $10.521\pm0.187$ & $1.448\pm0.085$ & $1.453\pm0.035$ & $0.972\pm0.026$ & $1.222\pm0.045$ \\ \hline 20\%-40\% & $7.631\pm0.148$ & $7.515\pm0.139$ & $0.988\pm0.028$ & $1.051\pm0.027$ & $0.651\pm0.018$ & $0.842\pm0.033$ \\ \hline 40\%-100\% & $3.153\pm0.069$ & $3.024\pm0.060$ & $0.399\pm0.012$ & $0.379\pm0.011$ & $0.261\pm0.008$ & $0.338\pm0.014$ \\ \hline p+p & $1.524\pm0.027$ & $1.504\pm0.027$ & $0.166\pm0.004$ & $0.173\pm0.009$ & $0.113\pm0.003$ & $0.137\pm0.007$ \\ \hline \end{tabular} \caption{$dN/dy$ of $\pi^{-}$, $\pi^{+}$, $K^{-}$, $K^{+}$, $\bar{p}$ and $p$ from power law fit in minimum-bias, centrality selected d+Au collisions and also in p+p collisions. The errors are from the power law fit.} \label{dndypowerlawfit} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline Centrality Bin & $\pi^{-}$ & $\pi^{+}$ & $K^{-}$ & $K^{+}$ & $\bar{p}$ & $p$ \\ \hline d+Au M.B. & $0.420\pm0.019$ & $0.422\pm0.019$ & $0.613\pm0.025$ & $0.625\pm0.025$ & $0.761\pm0.056$ & $0.739\pm0.069$\\ \hline 0\%-20\% & $0.435\pm0.017$ & $0.436\pm0.017$ & $0.627\pm0.025$ & $0.646\pm0.028$ & $0.774\pm0.056$ & $0.761\pm0.063$ \\ \hline 20\%-40\% & $0.425\pm0.019$ & $0.427\pm0.018$ & $0.610\pm0.025$ & $0.622\pm0.025$ & $0.766\pm0.052$ & $0.744\pm0.061$ \\ \hline 40\%-100\% & $0.405\pm0.020$ & $0.408\pm0.019$ & $0.591\pm0.024$ & $0.608\pm0.026$ & $0.715\pm0.056$ & $0.703\pm0.063$ \\ \hline p+p & $0.377\pm0.021$ & $0.379\pm0.020$ & $0.565\pm0.023$ & $0.565\pm0.023$ & $0.627\pm0.065$ & $0.634\pm0.070$ \\ \hline \end{tabular} \caption{The final $\langle p_T \rangle$ of $\pi^{-}$, $\pi^{+}$, $K^{-}$, $K^{+}$, $\bar{p}$ and $p$ in minimum-bias, centrality selected d+Au collisions and also in p+p collisions. The unit of $p_{T}$ is GeV/c.} \label{finalmeanpt} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline Centrality Bin & $\pi^{-}$ & $\pi^{+}$ & $K^{-}$ & $K^{+}$ & $\bar{p}$ & $p$ \\ \hline d+Au M.B. & $4.731\pm0.359$ & $4.668\pm0.356$ & $0.662\pm0.040$ & $0.684\pm0.039$ & $0.425\pm0.044$ & $0.531\pm0.067$\\ \hline 0\%-20\% & $10.063\pm0.628$ & $9.932\pm0.621$ & $1.383\pm0.095$ & $1.418\pm0.079$ & $0.896\pm0.084$ & $1.114\pm0.117$ \\ \hline 20\%-40\% & $7.137\pm0.514$ & $7.074\pm0.464$ & $0.952\pm0.059$ & $1.020\pm0.059$ & $0.603\pm0.054$ & $0.765\pm0.083$ \\ \hline 40\%-100\% & $2.925\pm0.236$ & $2.829\pm0.203$ & $0.387\pm0.023$ & $0.371\pm0.020$ & $0.238\pm0.025$ & $0.304\pm0.036$ \\ \hline p+p & $1.411\pm0.116$ & $1.400\pm0.108$ & $0.163\pm0.009$ & $0.168\pm0.010$ & $0.099\pm0.015$ & $0.120\pm0.018$ \\ \hline \end{tabular} \caption{The final $dN/dy$ of $\pi^{-}$, $\pi^{+}$, $K^{-}$, $K^{+}$, $\bar{p}$ and $p$ in minimum-bias, centrality selected d+Au collisions and also in p+p collisions.} \label{finaldndy} \end{scriptsize} \end{table} \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{protonyield.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{pbaryield.eps} \end{minipage} \caption{The re-scaled $p$ and $\bar{p}$ spectra in minimum-bias, centrality selected d+Au collisions and also in p+p collisions. The errors are statistical.} \label{protonspectra} \end{figure} \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{daupppowerlawfit.eps} \caption{The spectra of $\pi^{-}$, $\pi^{+}$, $K^{-}$, $K^{+}$, $\bar{p}$ and $p$ in d+Au and p+p minimum-bias collisions. The curves are from power law fit.} \label{powerlawfit} \end{figure} \begin{figure}[h] \centering \includegraphics[height=24pc,width=24pc]{3centralitypowerlawfit.eps} \caption{The spectra of $\pi^{-}$, $\pi^{+}$, $K^{-}$, $K^{+}$, $\bar{p}$ and $p$ in three centrality selected d+Au collisions. The curves are from power law fit.} \label{3centralitypowerlawfit} \end{figure} \section{System comparison} \begin{figure}[h] \centering \includegraphics[height=24pc,width=24pc]{meanpt_new.eps} \caption{$\langle p_T \rangle$ as a function of $dN/d\eta$. The squared, circled and triangled symbols are from~\cite{olga} in p+p and Au+Au collisions. The cross, star and diamond are our data points in p+p and d+Au collisions. Statistic errors and systematic uncertainties have been added in quadrature.} \label{meanptNch} \end{figure} Figure~\ref{meanptNch} shows the $\langle p_T \rangle$ of $\pi^{-}$, $K^{-}$ and $\bar{p}$ as a function of charged particle multiplicity at mid-rapidity. From p+p to d+Au collisions, the $\langle p_T \rangle$ increase with charged particle multiplicity smoothly. We observed the $\langle p_T \rangle$ in 0\%-20\% d+Au collisions are larger than those in peripheral Au+Au collisions. \begin{figure}[h] \centering \includegraphics[height=24pc,width=24pc]{kaonprotontopionratio_new.eps} \caption{$K^{-}/\pi^{-}$ and $\bar{p}/\pi^{-}$ as a function of $dN/d\eta$. The circled and triangled symbols are from~\cite{olga} in p+p and Au+Au collisions. The star and diamond are our data points in p+p and d+Au collisions. Statistic errors and systematic uncertainties have been added in quadrature.} \label{kaonpbarpionratio} \end{figure} The $K^{-}/\pi^{-}$ and $\bar{p}/\pi^{-}$ as a function of charged particle multiplicity at mid-rapidity are shown in Figure~\ref{kaonpbarpionratio}. The $K^{-}/\pi^{-}$ and $\bar{p}/\pi^{-}$ ratios were derived by taking the ratios of the dN/dy of $K^{-}$ or $\bar{p}$ over the dN/dy of $\pi^{-}$ in table~\ref{finaldndy}. These ratios increase with charged particle multiplicity from p+p, d+Au to Au+Au collisions smoothly. The kinetic freeze out temperature $T_{kin}$ and flow velocity $\langle \beta \rangle$ from thermal fit as a function of charged particle multiplicity are shown in Figure~\ref{freezeoutT}. We can see the $T_{kin}$ is flat from p+p to d+Au and then decreases from d+Au to Au+Au collisions and the $\langle \beta \rangle$ increases from p+p, d+Au to Au+Au collisions. \begin{figure}[h] \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{temperature.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{80mm} \includegraphics[height=17pc,width=18pc]{beta_new.eps} \end{minipage} \caption{The kinetic freeze out temperature $T_{kin}$ (left) and flow velocity $\langle \beta \rangle$ (right) from thermal fit as a function of charged particle multiplicity. The circled symbols are from ~\cite{olga} in p+p and Au+Au collisions. The star are our data points in p+p and d+Au collisions. Errors are systematic.} \label{freezeoutT} \end{figure} \chapter{The STAR Experiment} \label{chp:star} \section{The RHIC Accelerator} The Relativistic Heavy Ion Collider (RHIC) at Brookhaven National Lab (BNL) is the first hadron accelerator and collider consisting of two independent ring. It is designed to operate at high collision luminosity over a wide range of beam energies and particle species ranging from polarized proton to heavy ion~\cite{rhic:01,rhic:02}, where the top energy of the colliding center-of-mass energy per nucleon-nucleon pair is $\sqrt{s_{NN}}$ = 200 GeV. The RHIC facility consists of two super-conducting magnets, each with a circumference of 3.8 km, which focus and guide the beams. \\ Figure 2.1 shows the BNL accelerator complex including the accelerators used to bring the gold ions up to RHIC injection energy. In the first, gold ions are accelerated to 15 MeV/nucleon in the Tandem Van de Graaff facility. Then the beam is transferred to the Booster Synchrotron and accelerated to 95 MeV/nucleon through the Tandem-to-Booster line. Then the gold ions are transferred to the Alternating Gradient Synchrotron (AGS) and accelerated to 10.8 GeV/nucleon. Finally they are injected to RHIC and accelerated to the collision energy 100 GeV/nucleon.\\ \begin{figure} \centering \includegraphics[height=35pc,width=32pc]{rhic.eps} \caption{A diagram of the Brookhaven National Laboratory collider complex including the accelerators that bring the nuclear ions up to RHIC injection energy (10.8 GeV/nucleon for $^{197}$Au). Figure is taken from~\cite{sorenson:01,Haibin:03}.} \end{figure} RHIC's 3.8 km ring has six intersection points where its two rings of accelerating magnets cross, allowing the particle beams to collide. The collisions produce the fleeting signals that, when captured by one of RHIC's experimental detectors, provide physicists with information about the most fundamental workings of nature. If RHIC's ring is thought of as a clock face, the four current experiments are at 6 o'clock (STAR), 8 o'clock (PHENIX), 10 o'clock (PHOBOS) and 2 o'clock (BRAHMS). There are two additional intersection points at 12 and 4 o'clock where future experiments may be placed~\cite{rhic:01}. \section{The STAR Detector} \begin{figure}[h] \centering \includegraphics[height=18pc,width=28pc]{star_1.eps} \caption{Perspective view of the STAR detector, with a cutaway for viewing inner detector systems. Figure is taken from ~\cite{detector:01}.} \label{starfigure1} \end{figure} \begin{figure}[h] \centering \includegraphics[height=22pc,width=28pc]{star_2.eps} \caption{Cutaway side view of the STAR detector as configured in 2001. Figure is taken from~\cite{detector:01}.} \label{starfigure2} \end{figure} The Solenoidal Tracker at RHIC (STAR) is one of the two large detector systems constructed at the Relativistic Heavy Ion Collider (RHIC) at Brookhaven National Laboratory. STAR was constructed to investigate the behavior of strongly interacting matter at high energy density and to search for signatures of quark-gluon plasma (QGP) formation. Key features of the nuclear environment at RHIC are a large number of produced particles (up to approximately one thousand per unit pseudo-rapidity) and high momentum particles from hard parton-parton scattering. STAR can measure many observables simultaneously to study signatures of a possible QGP phase transition and to understand the space-time evolution of the collision process in ultra-relativistic heavy ion collisions. The goal is to obtain a fundamental understanding of the microscopic structure of these hadronic interactions at high energy densities. In order to accomplish this, STAR was designed primarily for measurements of hadron production over a large solid angle, featuring detector systems for high precision tracking, momentum analysis, and particle identification at the center of mass (c.m.) rapidity. The large acceptance of STAR makes it particularly well suited for event-by-event characterizations of heavy ion collisions and for the detection of hadron jets~\cite{detector:01}.\\ The layout of the STAR experiment~\cite{STAR CDR} is shown in Figure~\ref{starfigure1}. A cutaway side view of the STAR detector as configured for the RHIC 2001 run is displayed in Figure~\ref{starfigure2}. A room temperature solenoidal magnet~\cite{brown} with a maximum magnetic field of 0.5 T provides a uniform magnetic field for charged particle momentum analysis. Charged particle tracking close to the interaction region is accomplished by a Silicon Vertex Tracker~\cite{bellwied} (SVT). The Silicon Drift Detectors~\cite{baudot} (SDD) installed after 2001 is also for the inner tracking. The silicon detectors cover a pseudo-rapidity range $\mid {\eta }\mid \leq 1$ with complete azimuthal symmetry ($\Delta \phi = 2\pi$). Silicon tracking close to the interaction allows precision localization of the primary interaction vertex and identification of secondary vertices from weak decays of, for example, $\Lambda$, $\Xi$, and $\Omega$. A large volume Time Projection Chamber~\cite{wieman,tpc} (TPC) for charged particle tracking and particle identification is located at a radial distance from 50 to 200 cm from the beam axis. The TPC is 4 meters long and it covers a pseudo-rapidity range $\mid {\eta}\mid \leq 1.8$ for tracking with complete azimuthal symmetry ($\Delta \phi = 2\pi$). Both the SVT and TPC contribute to particle identification using ionization energy loss, with an anticipated combined energy loss resolution (dE/dx) of 7 \% ($\sigma$). The momentum resolution of the SVT and TPC reach a value of $\delta $p/p = 0.02 for a majority of the tracks in the TPC. The $\delta $p/p resolution improves as the number of hit points along the track increases and as the particle's momentum decreases, as expected~\cite{detector:01}. \\ To extend the tracking to the forward region, a radial-drift TPC (FTPC)~\cite{eckardt} is installed covering $2.5<\mid{\eta }\mid < 4$, also with complete azimuthal coverage and symmetry. To extend the particle identification in STAR to larger momenta over a small solid angle for identified single-particle spectra at mid-rapidity, a ring imaging Cherenkov detector ~\cite{ALICE_HMPID} covering $\mid\eta\mid < 0.3$ and $\Delta \phi = 0.11\pi$, and a time-of-flight patch (TOFp)~\cite{pVPD} covering $-1<\eta <0$ and $\Delta\phi = 0.04\pi $ (as shown in Figure~\ref{starfigure2}) was installed at STAR in 2001~\cite{detector:01}. In 2003, a time-of-flight tray (TOFr) based on multi-gap resistive plate chamber (MRPC) technology~\cite{startof} was installed in STAR detector, covering $-1<\eta <0$ and $\Delta\phi = \pi/30 $. For the time-of-flight system, the Pseudo-Vertex Position Detectors (pVPD) was installed as the start-timing detector, which was 5.4 m away from TPC center and covers $4.4<|\eta|<4.9$ with the azimuthal coverage 19\%~\cite{pVPD} in 2003.\\ The fast detectors that provide input to the trigger system are a central trigger barrel (CTB) at $|\eta|<1$ and two zero-degree calorimeters (ZDC) located in the forward directions at $\theta<2$ mrad. The CTB surrounds the outer cylinder of the TPC, and triggers on the flux of charged particles in the mid-rapidity region. The ZDCs are used for determining the energy in neutral particles remaining in the forward directions~\cite{detector:01}. A minimum bias trigger was obtained by selecting events with a pulse height larger than that of one neutron in each of the forward ZDCs, which corresponds to 95 percent of the geometrical cross section~\cite{detector:01}. \subsection{The Time Projection Chamber} The STAR detector~\cite{STAR CDR} uses the TPC as its primary tracking device. The TPC records the tracks of particles, measures their momenta, and identifies the particles by measuring their ionization energy loss ($dE/dx$). Particles are identified over a momentum range from 100 MeV/c to greater than 1 GeV/c and momenta are measured over a range of 100 MeV/c to 30 GeV/c~\cite{tpc}.\\ The STAR TPC is shown schematically in Figure~\ref{tpcman}. It is a volume of gas in a well defined uniform electric field of $\approx$ 135 V/cm. The working gas of TPC is P10 gas (10\% methane, 90\% argon) regulated at 2 mbar above atmospheric pressure\cite{gas}. This gas has long been used in TPCs. Its primary attribute is a fast drift velocity which peaks at a low electric field. Operating on the peak of the velocity curve makes the drift velocity stable and insensitive to small variations in temperature and pressure~\cite{tpc}. The paths of primary ionizing particles passing through the gas volume are reconstructed with high precision from the released secondary electrons which drift to the readout end caps at the ends of the chamber. The drift velocity of electrons is 5.45 cm/$\mu$s. The uniform electric field which is required to drift the electrons is defined by a thin conductive Central Membrane (CM) at the center of the TPC, concentric field cage cylinders and the read out end caps~\cite{tpc}. The readout system is based on Multi Wire Proportional Chambers (MWPC) with readout pads. The drifting electrons avalanche in the high fields at the 20 $\mu$m anode wires providing an amplification of 1000 to 3000. The induced charge from an avalanche is shared over several adjacent pads, so the original track position can be reconstructed to a small fraction of a pad width. There are a total of 136,608 pads in the readout system~\cite{tpc}, which give $x$-$y$ coordinate information. The $z$ position information is provided by 512 time buckets.\\ \begin{figure}[htb] \includegraphics[width=14cm]{tpcman.eps} \caption{The STAR TPC surrounds a beam-beam interaction region at RHIC. The collisions take place near the center of the TPC.} \label{tpcman} \end{figure} At the Data Acquisition (DAQ) stage, raw events containing millions of ADC values and TDC values were recorded. Raw data were then reconstructed into hits, tracks, vertices, and the collision vertex through the reconstruction chain of TPC~\cite{starsoftware} by Kalman method. The collision vertex are called the primary vertex. The tracks are called the global tracks. If the 3-dimensional distance of closest approach (DCA/dca) of the global track to the primary vertex is less than 3 cm, this track will be chosen for a re-fit by forcing a new track helix ending at the primary vertex. These newly reconstructed helices are called primary tracks~\cite{Haibin:03}. As expected, the vertex resolution decreases as the square root of the number of tracks used in the calculation. The vertex resolution is 350 $\mu$m when there are more than 1,000 tracks~\cite{tpc}. Figure~\ref{eventshow} shows the beam's eye view of a central Au+Au collision event in the STAR TPC. \begin{figure}[h] \centering \includegraphics[height=14pc,width=18pc]{event.eps} \caption{Beam's eye view of a central Au+Au collision event in the STAR Time Projection Chamber. This event was drawn by the STAR online display. Figure is taken from~\cite{detector:01}.} \label{eventshow} \end{figure} \subsubsection{Particle Identification (PID) of TPC by dE/dx}Energy lost in the TPC gas is a valuable tool for identifying particle species. It works especially well for low momentum particles but as the particle energy rises, the energy loss becomes less mass-dependent and it is hard to separate particles with velocities $v>0.7$c~\cite{tpc}. For a particle with charge $z$ (in units of $e$) and speed $\beta=v/c$ passing through a medium with density $\rho$, the mean energy loss it suffers can be described by the Bethe-Bloch formula \begin{equation} \langle \frac{dE}{dx} \rangle = 2\pi N_0r_e^2m_ec^2\rho\frac{Zz^2}{A\beta^2} [\text{ln}\frac{2m_e\gamma^2v^2E_M}{I^2}-2\beta^2] \end{equation} where $N_0$ is Avogadro's number, $m_e$ is the electron mass, $r_e$ ($=e^2/m_e$) is the classical electron radius, $c$ is the speed of light, $Z$ is the atomic number of the absorber, $A$ is the atomic weight of the absorber, $\gamma=1/\sqrt{1-\beta^2}$, $I$ is the mean excitation energy, and $E_M$ ($=2m_ec^2\beta^2/(1-\beta^2)$) is the maximum transferable energy in a single collision~\cite{tang:01,Haibin:03}. From the above equation, we can see that different charged particles (electron, muon, pion, kaon, proton or deuteron) with the same momentum $p$ passing through the TPC gas can result in different energy loss. Figure~\ref{fdedx} shows the energy loss for particles in the TPC as a function of the particle momentum, which includes both primary and secondary particles. We can see that charged pions and kaons can be identified up to about transverse momentum 0.75 GeV/c and protons and anti-protons can be identified to 1.1 GeV/c. \begin{figure}[h] \centering \includegraphics[width=22pc]{dedxPlotAllBands.eps} \caption{The energy loss distribution for primary and secondary particles in the STAR TPC as a function of the $p_T$ of the primary particle. This figure is taken from~\cite{tpc}.} \label{fdedx} \end{figure} In order to quantitatively describe the particle identification, we define the variable $N_{\sigma\pi}$ (in the case of charged pion identification) as \begin{equation} N_{\sigma\pi}=[\frac{dE}{dx}_{meas.}-\langle\frac{dE}{dx}\rangle_\pi]/ [\frac{0.55}{\sqrt{N}}\frac{dE}{dx}_{meas.}] \end{equation} in which $N$ is the number of hits for a track in the TPC, $\frac{dE}{dx}_{meas.}$ is the measured energy loss of a track and $\langle\frac{dE}{dx}\rangle_\pi$ is the mean energy loss for charged pions. In order to identify charged kaons, protons and anti-protons, we can have similar definition of $N_{\sigma K}$ and $N_{\sigma p}$. Thus we can cut on the variables $N_{\sigma\pi}$, $N_{\sigma K}$ and $N_{\sigma p}$ to select different particle species~\cite{Haibin:03}.\\ A specific part of the particle identification is the topological identification of neutral particles, such as the $K_S^0$ and $\Lambda$. These neutral particles can be reconstructed by identifying the secondary vertex, commonly called V0 vertex, of their charged daughter decay modes, $K_S^0\rightarrow\pi^+\pi^-$ and $\Lambda\rightarrow p \pi^-$~\cite{Haibin:03}. \subsection{The time-of-flight tray based on MRPC technology} \begin{figure}[h] \centering \includegraphics[height=24pc,width=32pc]{traymodule-instar.eps} \caption{Tray structure. Figure is taken from ~\cite{tofproposal}.} \label{tofrtray} \end{figure} In 2003, the time-of-flight tray (TOFr) based on multi-gap resistive plate chamber (MRPC) technology~\cite{startof} was installed in STAR detector. It extends particle identification up to $p_{T}\sim3$ GeV/c for $p$ and $\bar{p}$. This tray was installed on the Au beam outgoing direction. MRPC technology was first developed by the CERN ALICE group~\cite{williams} to provide a cost-effective solution for large-area time-of-flight coverage. For full time-of-flight coverage at STAR, there will be 120 trays, with 60 on east side and 60 on west side. For each tray, there will be 33 MRPCs. For each MRPC, there are 6 read-out channels. Figure~\ref{tofrtray} shows the tray which indicates the position of each MRPC module. The MRPCs are tilted differently so that each MRPC is most projective to the average primary vertex location at Z=0. In 2003 d+Au and p+p run, only 28 MRPCs were installed in the tray and 12 out of 28 were instrumented with the electronics, representing 0.3\% of TPC coverage. If we number the 33 MRPCs in the tray from 1 to 33, with 1 close to TPC center and 33 far from TPC center, the numbers of 12 modules instrumented with the electronics in 2003 are 3,4,5,7,9,10,11,12,13,14,26 and 32. \subsubsection{The introduction of MRPC} \begin{figure}[h] \begin{minipage}[t]{1.0\linewidth} \includegraphics[height=16pc,width=32pc]{Augusttmp1.eps} \end{minipage} \hspace{\fill} \begin{minipage}[t]{1.0\linewidth} \includegraphics[height=11.5pc,width=32pc]{Augusttmp2.eps} \end{minipage} \caption{Two side views of MRPC. The upper (lower) is for long (short) side view. The two plots are not at the same scale. Figure is taken from ~\cite{tofproposal}.} \label{mrpcstru} \end{figure} \begin{figure}[h] \centering \includegraphics[height=12pc,width=24pc]{readout.eps} \caption{The shape of the 6 read-out strips for each MRPC.} \label{readout} \end{figure} Resistive Plate Chambers (RPCs) were developed in 1980s~\cite{mysimu:01}, and were originally operated in streamer mode. This operation mode allows us to get high detection efficiency ($>$95\%) and time resolution (~1 ns), with low fluxes of incident particles. At higher fluxes ($>$200 $Hz/cm^2$), RPCs begin to lose their efficiency. A way to overcome this problem is to operate RPCs in avalanche mode. The Multi-gap Resistive Plate Chamber (MRPC) was developed less than 10 years ago~\cite{mysimu:02}. It consists of a stack of resistive plates, spaced one from the other with equal sized spacers creating a series of gas gaps. Electrodes are connected to the outer surfaces of the stack of resistive plates while all the internal plates are left electrically floating. Initially the voltage on these internal plates is given by electrostatics, but they are kept at the correct voltage due to the flow of electrons and ions created in the avalanches. Figure~\ref{mrpcstru} shows the structure of MRPC detector. For each MRPC, there are 6 read-out strips. Figure~\ref{readout} shows the shape of the read-out strip. The detailed production process can be found at Appendix B.\\ MRPC, as a new kind of detector for time of fight system, operated in avalanche mode with a non flammable gas mixture of 90\% F134A, 5\% isobutane, 5\% SF6, can fulfill all these requirements: high efficiency ($>$95\%), excellent intrinsic time resolution ($<$100 ps)~\cite{mysimu:13,startof,mysimu:15,mysimu:16,mysimu:17}, high rate capability (~500 $Hz/cm^2$), high modularity and simplicity for construction, good uniformity of response, high granularity/low occupancy, and large acceptance. \subsubsection{Simulation: the work principle of this chamber} A detailed description of the model used in the simulation was reported in these papers~\cite{mysimu:03,mysimu:04,mysimu:05,mysimu:06}, here just the main items will be repeated. The program starts from considering an ionizing particle which crosses the gas gaps and generates a certain number of clusters of ion-electron pairs. The electrons contained in the clusters drift towards the anode and, if the electric field is sufficiently high, give rise to the avalanche processes. \\ The primary cluster numbers and the avalanche growth are assumed to follow, respectively, simple Poisson statistics and the usual exponential law. Avalanche gain fluctuations have been taken into account using a Polya distribution~\cite{mysimu:07}. After the simulation of the drifting avalanches, the program computes, by means of Ramo~\cite{mysimu:08} theorem, the charge $q_{ind}$ induced on the external pick-up electrodes (strips or pads) by the avalanche motion. Under certain approximations, this is given by the formula \begin{equation} q_{ind} = \frac{q_{e}}{\eta{d}}\triangle{V}_{w}{\sum_{j=1}^{n_{cl}}{n_{j}M(e^{\eta{(d-x_{j})}}-1)}} \end{equation} where $q_e$ is the electron charge, $\eta$ 1st effective Townsend coefficient $\eta=\alpha-\beta$, $\alpha$ is the Townsend coefficient, $\beta$ is the attachment coefficient, $x_j$ the $j_{th}$ cluster initial distances from the anode, $d$ the gap width, $n_j$ the number of initial electrons in the considered $j_{th}$ cluster, $M$ the avalanche gain fluctuations factor, and $\triangle{V}_{w}/d=E_{w}$ is the normalized weighting field. In addition to $q_{ind}$, the current $i_{ind}(t)$ induced on the same electrodes by the drifting charge $q_d(t)$ may be computed as \begin{equation} i_{ind}(t) = \triangle{V}_{w}\frac{v_{d}}{d}q_d(t)Me^{\eta{v_{d}t}} \end{equation}, where $v_d$ is the electron drift velocity. The computation of $i_{ind}$ allows us to reproduce the whole information coming out from MRPC, such as time distribution.\\ \textbf{Charge Spectrum Simulation}: The almost Gaussian charge distribution obtained with the MRPC is a key ingredient to its performance. If the avalanches grew following Townsend's formula the charge distribution would be exponential in shape. Thus the space charge effects must be considered in the simulation. \\ The input parameters for the simulation program are: the Townsend coefficient $\alpha$, the attachment coefficient $\beta$, the average distance between clusters $\lambda$ and the probability distribution of the number of electrons per cluster. These pieces of information can be obtained, for a given gas mixture and given conditions (pressure and temperature) and electric field, by the programs HEED~\cite{mysimu:09} and MAGBOLTZ~\cite{mysimu:10,mysimu:11}. In addition, a maximum number of electrons in an avalanche (cutoff value) is specified. \\ In a given gap, we generate a number of clusters with distances exponentially distributed with average distance $\lambda$. For each cluster, we then generate a certain number of electrons, according to the distribution obtained by the program HEED. Each electron from the primary cluster will give rise to a number of electrons, generated according to an exponential probability law.\\ \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{tmp_ano.eps} \caption{Simulated 1st effective Townsend coefficient curve and normalized charge distribution for a 6 and 10 gap MRPC.} \label{chargedistribution} \end{figure} For each cluster, the avalanche growth is stopped when the total charge reaches a certain cutoff value, as originally suggested in ref.~\cite{mysimu:12} to take into account space charge effects in the avalanche development. This cutoff value has been set to be $1.6\times10^7$ electrons.\\ In Figure~\ref{chargedistribution} we show the results of simulations, Figure~\ref{chargedistribution}(a) is the simulated curve of the 1st effective Townsend coefficient $\eta$ versus the electric field, which is generated by Magboltz. The curve shows that the correlation between $\eta$ and the electric field is almost linear when MRPC is operated at high electric field for the gas mixture. Figure~\ref{chargedistribution}(b) is the charge spectrum for a 6 gap chamber and (c) (d) for a 10 gap chamber compared to experimental data~\cite{mysimu:13,startof}, and the number under each plot shows the electric field $E$ in the gas gap for MRPC. In both cases the gap size is 220 $\mu{m}$. The gas mixture was 90\% F134A, 5\% isobutane and 5\% SF6 in normal conditions of pressure and temperature. The value of $\lambda$ used was 0.1 $mm$, derived from HEED program. \\ The charge distribution has an almost Gaussian form, especially for the 10 gap MRPC. The left side of the distribution (very few events at values near zero) is due to the fact that the MRPC operates at high gain $\eta \times{d}\sim30$. This means that avalanches starting in the middle of the gap width, which only avalanche over half the distance, give a detectable signal. The charge distribution is the superposition of several probability distributions which, according to the central limit theorem, will tend to a Gaussian form. The right side of the charge distribution (the fact that the tails are not very long) indicates that indeed the space charge effects stop the development of the avalanche.\\ \begin{figure}[h] \centering \includegraphics[height=18pc,width=24pc]{6gap-time-velocity.eps} \caption{Simulated results of a 6 gap MRPC.} \label{timedistribution} \end{figure} \textbf{Time Distribution Simulation:} We then proceed to simulate the time distribution of these same chambers. The electron drift velocity can be obtained from HEED. When the total induced charge signal is over threshold, the time is recorded. In this paper, the threshold is 13 fc for the 6 gap MRPC and 26 fc for the 10 gap MRPC. Fig.2 is the simulated results for a 6 gap chamber. Figure~\ref{timedistribution} (a) is the simulated curve of the electron drift velocity versus the electric field, which is generated by Magboltz. Figure~\ref{timedistribution} (b) is the time distribution of a 6 gap MRPC. The intrinsic time resolution is only 19 ps or so. If we consider other contributions, such as front-end electronics 30 ps, TDC resolution 25 ps, fanout start signal 10 ps, beam size (1cm) 15 ps, we can get the MRPC resolution is $\sqrt{20^2+30^2+25^2+10^2+15^2}=47$ ps. This value is similar to the experimental result~\cite{mysimu:13,startof}. For a 10 gap MRPC, the intrinsic time resolution is about 15 ps.\\ From the simulation, we can get the bottom line of MRPC time resolution $\sim20$ ps. And we need to keep control of all these contributions to ensure best time resolution. \subsubsection{MRPC for this tray installed in 2003} In 2003, for the MRPCs in the TOFr, the inner glass thickness is 0.54 mm, the outer glass is 1.1 mm. The gas gap is 0.22 mm. Both the volume resistivity ($10^{12-13} ohm.cm$) of the glass plates and the surface resistivity(2M ohm per square) of carbon layer at room temperature are presented in~\cite{resistivity}. It is found the volume resistivity of the plate decreases with the temperature increasing. And the radiation will decrease the volume resistivity of the plate~\cite{resistivity}. In order not to pollute the working gas of TPC, SF6 is not used as part of the working gas of TOFr. The working gas of MRPC-TOFr at STAR is 95\% freon and 5\% iso-butane at normal atmospheric pressure. The high voltage applied to the electrodes is 14.0 kV. \chapter{{\hspace{3.5cm}Tables of the Invariant Yields}} \begin{table*} \begin{scriptsize} \centering \begin{tabular} {|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & M.B. & 0\%-20\% \\ \hline 3.50e-01& 1.00e-01 & $3.14e+00\pm3.65e-02\pm2.51e-01$ & $6.70e+00\pm9.69e-02\pm5.36e-01$ \\ \hline 4.50e-01& 1.00e-01 & $1.82e+00\pm2.45e-02\pm1.45e-01$ & $3.88e+00\pm6.51e-02\pm3.10e-01$ \\ \hline 5.50e-01& 1.00e-01 & $1.08e+00\pm1.73e-02\pm8.67e-02$ & $2.38e+00\pm4.68e-02\pm1.90e-01$ \\ \hline 6.50e-01& 1.00e-01 & $6.18e-01\pm1.22e-02\pm4.94e-02$ & $1.38e+00\pm3.28e-02\pm1.10e-01$ \\ \hline 7.50e-01& 1.00e-01 & $4.13e-01\pm9.15e-03\pm3.31e-02$ & $9.53e-01\pm2.56e-02\pm7.63e-02$ \\ \hline 8.50e-01& 1.00e-01 & $2.66e-01\pm6.98e-03\pm2.13e-02$ & $6.11e-01\pm1.93e-02\pm4.88e-02$ \\ \hline 9.50e-01& 1.00e-01 & $1.70e-01\pm5.36e-03\pm1.36e-02$ & $3.79e-01\pm1.44e-02\pm3.03e-02$ \\ \hline 1.05e+00& 1.00e-01 & $1.14e-01\pm4.07e-03\pm9.09e-03$ & $2.68e-01\pm1.15e-02\pm2.14e-02$ \\ \hline 1.15e+00& 1.00e-01 & $7.70e-02\pm3.27e-03\pm6.16e-03$ & $1.71e-01\pm8.82e-03\pm1.37e-02$ \\ \hline 1.30e+00& 2.00e-01 & $4.43e-02\pm1.66e-03\pm3.55e-03$ & $1.06e-01\pm4.71e-03\pm8.45e-03$ \\ \hline 1.50e+00& 2.00e-01 & $2.10e-02\pm1.07e-03\pm1.68e-03$ & $4.65e-02\pm2.96e-03\pm3.72e-03$ \\ \hline 1.70e+00& 2.00e-01 & $1.05e-02\pm7.19e-04\pm8.40e-04$ & $2.29e-02\pm1.92e-03\pm1.83e-03$ \\ \hline \end{tabular} \caption{$\pi^{+}$ spectra in minimum-bias and 0-20\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$. } \label{pionplusspectratable1} \end{scriptsize} \end{table*} \begin{table*} \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & 20\%-40\% & 40\%-100\% \\ \hline 3.50e-01& 1.00e-01 & $4.73e+00\pm6.94e-02\pm3.78e-01$ & $1.93e+00\pm2.85e-02\pm1.55e-01$ \\ \hline 4.50e-01& 1.00e-01 & $2.80e+00\pm4.73e-02\pm2.24e-01$ & $1.10e+00\pm1.89e-02\pm8.81e-02$ \\ \hline 5.50e-01& 1.00e-01 & $1.68e+00\pm3.36e-02\pm1.34e-01$ & $6.43e-01\pm1.31e-02\pm5.14e-02$ \\ \hline 6.50e-01& 1.00e-01 & $9.50e-01\pm2.30e-02\pm7.60e-02$ & $3.68e-01\pm9.03e-03\pm2.94e-02$ \\ \hline 7.50e-01& 1.00e-01 & $6.41e-01\pm1.76e-02\pm5.13e-02$ & $2.40e-01\pm6.80e-03\pm1.92e-02$ \\ \hline 8.50e-01& 1.00e-01 & $4.20e-01\pm1.36e-02\pm3.36e-02$ & $1.52e-01\pm5.09e-03\pm1.22e-02$ \\ \hline 9.50e-01& 1.00e-01 & $2.58e-01\pm1.00e-02\pm2.07e-02$ & $1.01e-01\pm3.97e-03\pm8.05e-03$ \\ \hline 1.05e+00& 1.00e-01 & $1.80e-01\pm7.92e-03\pm1.44e-02$ & $6.10e-02\pm2.84e-03\pm4.88e-03$ \\ \hline 1.15e+00& 1.00e-01 & $1.21e-01\pm6.35e-03\pm9.72e-03$ & $4.38e-02\pm2.37e-03\pm3.50e-03$ \\ \hline 1.30e+00& 2.00e-01 & $6.49e-02\pm3.02e-03\pm5.19e-03$ & $2.40e-02\pm1.16e-03\pm1.92e-03$ \\ \hline 1.50e+00& 2.00e-01 & $3.28e-02\pm2.05e-03\pm2.63e-03$ & $1.03e-02\pm7.12e-04\pm8.27e-04$ \\ \hline 1.70e+00& 2.00e-01 & $1.57e-02\pm1.31e-03\pm1.26e-03$ & $4.92e-03\pm4.49e-04\pm3.94e-04$ \\ \hline \end{tabular} \caption{$\pi^{+}$ spectra in 20-40\% and 40-100\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pionplusspectratable2} \end{scriptsize} \end{table*} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & p+p \\ \hline 3.50e-01& 1.00e-01 & $9.71e-01\pm1.21e-02\pm7.76e-02$ \\ \hline 4.50e-01& 1.00e-01 & $5.32e-01\pm7.84e-03\pm4.26e-02$ \\ \hline 5.50e-01& 1.00e-01 & $3.14e-01\pm5.46e-03\pm2.51e-02$ \\ \hline 6.50e-01& 1.00e-01 & $1.74e-01\pm3.72e-03\pm1.40e-02$ \\ \hline 7.50e-01& 1.00e-01 & $1.08e-01\pm2.65e-03\pm8.64e-03$ \\ \hline 8.50e-01& 1.00e-01 & $6.42e-02\pm1.89e-03\pm5.14e-03$ \\ \hline 9.50e-01& 1.00e-01 & $4.03e-02\pm1.42e-03\pm3.22e-03$ \\ \hline 1.05e+00& 1.00e-01 & $2.40e-02\pm9.94e-04\pm1.92e-03$ \\ \hline 1.15e+00& 1.00e-01 & $1.55e-02\pm7.71e-04\pm1.24e-03$ \\ \hline 1.30e+00& 2.00e-01 & $8.19e-03\pm3.86e-04\pm6.55e-04$ \\ \hline 1.50e+00& 2.00e-01 & $3.77e-03\pm2.46e-04\pm3.02e-04$ \\ \hline 1.70e+00& 2.00e-01 & $1.84e-03\pm1.96e-04\pm1.47e-04$ \\ \hline \end{tabular} \caption{$\pi^{+}$ spectra in p+p collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pionplusspectratable3} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & M.B. & 0\%-20\% \\ \hline 3.50e-01& 1.00e-01 & $3.20e+00\pm3.73e-02\pm2.56e-01$ & $6.77e+00\pm9.88e-02\pm5.42e-01$ \\ \hline 4.50e-01& 1.00e-01 & $1.82e+00\pm2.46e-02\pm1.45e-01$ & $3.92e+00\pm6.60e-02\pm3.13e-01$ \\ \hline 5.50e-01& 1.00e-01 & $1.08e+00\pm1.72e-02\pm8.60e-02$ & $2.42e+00\pm4.71e-02\pm1.94e-01$ \\ \hline 6.50e-01& 1.00e-01 & $6.70e-01\pm1.30e-02\pm5.36e-02$ & $1.49e+00\pm3.51e-02\pm1.20e-01$ \\ \hline 7.50e-01& 1.00e-01 & $4.05e-01\pm9.06e-03\pm3.24e-02$ & $9.22e-01\pm2.50e-02\pm7.38e-02$ \\ \hline 8.50e-01& 1.00e-01 & $2.59e-01\pm6.89e-03\pm2.07e-02$ & $6.13e-01\pm1.96e-02\pm4.91e-02$ \\ \hline 9.50e-01& 1.00e-01 & $1.68e-01\pm5.38e-03\pm1.34e-02$ & $3.85e-01\pm1.48e-02\pm3.08e-02$ \\ \hline 1.05e+00& 1.00e-01 & $1.16e-01\pm4.23e-03\pm9.29e-03$ & $2.70e-01\pm1.18e-02\pm2.16e-02$ \\ \hline 1.15e+00& 1.00e-01 & $7.63e-02\pm3.28e-03\pm6.11e-03$ & $1.71e-01\pm8.83e-03\pm1.37e-02$ \\ \hline 1.30e+00& 2.00e-01 & $4.36e-02\pm1.66e-03\pm3.49e-03$ & $9.67e-02\pm4.45e-03\pm7.73e-03$ \\ \hline 1.50e+00& 2.00e-01 & $2.04e-02\pm1.07e-03\pm1.63e-03$ & $4.94e-02\pm3.06e-03\pm3.96e-03$ \\ \hline 1.70e+00& 2.00e-01 & $1.03e-02\pm7.23e-04\pm8.26e-04$ & $2.54e-02\pm2.07e-03\pm2.03e-03$ \\ \hline \end{tabular} \caption{$\pi^{-}$ spectra in minimum-bias and 0-20\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$. } \label{pionminusspectratable1} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & 20\%-40\% & 40\%-100\% \\ \hline 3.50e-01& 1.00e-01 & $4.81e+00\pm7.11e-02\pm3.85e-01$ & $2.00e+00\pm2.95e-02\pm1.60e-01$ \\ \hline 4.50e-01& 1.00e-01 & $2.76e+00\pm4.73e-02\pm2.21e-01$ & $1.12e+00\pm1.93e-02\pm8.95e-02$ \\ \hline 5.50e-01& 1.00e-01 & $1.65e+00\pm3.28e-02\pm1.32e-01$ & $6.43e-01\pm1.31e-02\pm5.14e-02$ \\ \hline 6.50e-01& 1.00e-01 & $1.03e+00\pm2.47e-02\pm8.23e-02$ & $4.04e-01\pm9.83e-03\pm3.23e-02$ \\ \hline 7.50e-01& 1.00e-01 & $6.15e-01\pm1.71e-02\pm4.92e-02$ & $2.40e-01\pm6.80e-03\pm1.92e-02$ \\ \hline 8.50e-01& 1.00e-01 & $3.92e-01\pm1.30e-02\pm3.13e-02$ & $1.51e-01\pm5.10e-03\pm1.21e-02$ \\ \hline 9.50e-01& 1.00e-01 & $2.61e-01\pm1.03e-02\pm2.09e-02$ & $9.62e-02\pm3.91e-03\pm7.70e-03$ \\ \hline 1.05e+00& 1.00e-01 & $1.81e-01\pm8.11e-03\pm1.45e-02$ & $6.65e-02\pm3.07e-03\pm5.32e-03$ \\ \hline 1.15e+00& 1.00e-01 & $1.20e-01\pm6.26e-03\pm9.58e-03$ & $4.23e-02\pm2.31e-03\pm3.38e-03$ \\ \hline 1.30e+00& 2.00e-01 & $6.82e-02\pm3.18e-03\pm5.46e-03$ & $2.42e-02\pm1.18e-03\pm1.93e-03$ \\ \hline 1.50e+00& 2.00e-01 & $3.24e-02\pm2.13e-03\pm2.59e-03$ & $1.05e-02\pm7.69e-04\pm8.42e-04$ \\ \hline 1.70e+00& 2.00e-01 & $1.55e-02\pm1.34e-03\pm1.24e-03$ & $5.02e-03\pm1.06e-03\pm4.01e-04$ \\ \hline \end{tabular} \caption{$\pi^{-}$ spectra in 20-40\% and 40-100\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pionminusspectratable2} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & p+p \\ \hline 3.50e-01& 1.00e-01 & $9.71e-01\pm1.22e-02\pm7.76e-02$ \\ \hline 4.50e-01& 1.00e-01 & $5.47e-01\pm8.02e-03\pm4.37e-02$ \\ \hline 5.50e-01& 1.00e-01 & $3.09e-01\pm5.38e-03\pm2.47e-02$ \\ \hline 6.50e-01& 1.00e-01 & $1.84e-01\pm3.89e-03\pm1.47e-02$ \\ \hline 7.50e-01& 1.00e-01 & $1.00e-01\pm2.50e-03\pm8.01e-03$ \\ \hline 8.50e-01& 1.00e-01 & $6.36e-02\pm1.89e-03\pm5.09e-03$ \\ \hline 9.50e-01& 1.00e-01 & $3.80e-02\pm1.38e-03\pm3.04e-03$ \\ \hline 1.05e+00& 1.00e-01 & $2.44e-02\pm1.03e-03\pm1.95e-03$ \\ \hline 1.15e+00& 1.00e-01 & $1.57e-02\pm7.87e-04\pm1.25e-03$ \\ \hline 1.30e+00& 2.00e-01 & $8.70e-03\pm4.02e-04\pm6.96e-04$ \\ \hline 1.50e+00& 2.00e-01 & $3.62e-03\pm2.45e-04\pm2.90e-04$ \\ \hline 1.70e+00& 2.00e-01 & $1.69e-03\pm1.76e-04\pm1.35e-04$ \\ \hline \end{tabular} \caption{$\pi^{-}$ spectra in p+p collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pionminusspectratable3} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & M.B. & 0\%-20\% \\ \hline 4.57e-01& 1.00e-01 & $2.41e-01\pm9.47e-03\pm1.93e-02$ & $4.83e-01\pm2.75e-02\pm3.86e-02$ \\ \hline 5.56e-01& 1.00e-01 & $1.87e-01\pm7.53e-03\pm1.49e-02$ & $3.87e-01\pm2.14e-02\pm3.10e-02$ \\ \hline 6.55e-01& 1.00e-01 & $1.33e-01\pm5.14e-03\pm1.07e-02$ & $2.81e-01\pm1.53e-02\pm2.25e-02$ \\ \hline 7.54e-01& 1.00e-01 & $1.01e-01\pm3.18e-03\pm8.07e-03$ & $2.14e-01\pm1.09e-02\pm1.72e-02$ \\ \hline 8.54e-01& 1.00e-01 & $6.97e-02\pm2.48e-03\pm5.57e-03$ & $1.58e-01\pm8.74e-03\pm1.27e-02$ \\ \hline 9.54e-01& 1.00e-01 & $5.01e-02\pm1.95e-03\pm4.01e-03$ & $9.95e-02\pm6.30e-03\pm7.96e-03$ \\ \hline 1.05e+00& 1.00e-01 & $3.77e-02\pm1.65e-03\pm3.02e-03$ & $8.92e-02\pm5.83e-03\pm7.14e-03$ \\ \hline 1.15e+00& 1.00e-01 & $2.73e-02\pm1.31e-03\pm2.18e-03$ & $5.61e-02\pm4.30e-03\pm4.49e-03$ \\ \hline 1.30e+00& 2.00e-01 & $1.78e-02\pm7.25e-04\pm1.42e-03$ & $4.00e-02\pm2.49e-03\pm3.20e-03$ \\ \hline 1.50e+00& 2.00e-01 & $9.12e-03\pm5.25e-04\pm7.30e-04$ & $1.95e-02\pm1.89e-03\pm1.56e-03$ \\ \hline 1.70e+00& 2.00e-01 & $4.68e-03\pm3.96e-04\pm3.75e-04$ & $---$ \\ \hline \end{tabular} \caption{$K^{+}$ spectra in minimum-bias and 0-20\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$. } \label{kaonplusspectratable1} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & 20\%-40\% & 40\%-100\% \\ \hline 4.57e-01& 1.00e-01 & $3.63e-01\pm2.06e-02\pm2.91e-02$ & $1.27e-01\pm7.70e-03\pm1.02e-02$ \\ \hline 5.56e-01& 1.00e-01 & $2.74e-01\pm1.54e-02\pm2.19e-02$ & $1.07e-01\pm6.16e-03\pm8.60e-03$ \\ \hline 6.55e-01& 1.00e-01 & $2.01e-01\pm1.11e-02\pm1.60e-02$ & $7.37e-02\pm4.26e-03\pm5.89e-03$ \\ \hline 7.54e-01& 1.00e-01 & $1.58e-01\pm8.18e-03\pm1.27e-02$ & $5.25e-02\pm2.99e-03\pm4.20e-03$ \\ \hline 8.54e-01& 1.00e-01 & $1.01e-01\pm6.16e-03\pm8.06e-03$ & $3.73e-02\pm2.35e-03\pm2.98e-03$ \\ \hline 9.54e-01& 1.00e-01 & $7.97e-02\pm4.89e-03\pm6.37e-03$ & $2.80e-02\pm1.86e-03\pm2.24e-03$ \\ \hline 1.05e+00& 1.00e-01 & $5.34e-02\pm3.83e-03\pm4.27e-03$ & $1.91e-02\pm1.47e-03\pm1.52e-03$ \\ \hline 1.15e+00& 1.00e-01 & $3.91e-02\pm3.12e-03\pm3.13e-03$ & $1.36e-02\pm1.18e-03\pm1.09e-03$ \\ \hline 1.30e+00& 2.00e-01 & $2.51e-02\pm1.67e-03\pm2.01e-03$ & $8.65e-03\pm6.83e-04\pm6.92e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.32e-02\pm1.19e-03\pm1.05e-03$ & $4.69e-03\pm5.07e-04\pm3.75e-04$ \\ \hline \end{tabular} \caption{$K^{+}$ spectra in 20-40\% and 40-100\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{kaonplusspectratable2} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & p+p \\ \hline 4.57e-01& 1.00e-01 & $6.47e-02\pm3.01e-03\pm5.18e-03$ \\ \hline 5.56e-01& 1.00e-01 & $4.32e-02\pm2.10e-03\pm3.45e-03$ \\ \hline 6.55e-01& 1.00e-01 & $3.18e-02\pm1.51e-03\pm2.54e-03$ \\ \hline 7.54e-01& 1.00e-01 & $2.18e-02\pm9.70e-04\pm1.74e-03$ \\ \hline 8.54e-01& 1.00e-01 & $1.57e-02\pm7.69e-04\pm1.25e-03$ \\ \hline 9.54e-01& 1.00e-01 & $9.92e-03\pm5.60e-04\pm7.93e-04$ \\ \hline 1.05e+00& 1.00e-01 & $7.17e-03\pm4.74e-04\pm5.74e-04$ \\ \hline 1.15e+00& 1.00e-01 & $5.60e-03\pm4.18e-04\pm4.48e-04$ \\ \hline 1.30e+00& 2.00e-01 & $3.73e-03\pm2.72e-04\pm2.98e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.81e-03\pm1.18e-04\pm1.45e-04$ \\ \hline \end{tabular} \caption{$K^{+}$ spectra in p+p collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{kaonplusspectratable3} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & M.B. & 0\%-20\% \\ \hline 4.57e-01& 1.00e-01 & $2.34e-01\pm9.47e-03\pm1.87e-02$ & $4.87e-01\pm2.81e-02\pm3.90e-02$ \\ \hline 5.56e-01& 1.00e-01 & $1.92e-01\pm7.98e-03\pm1.54e-02$ & $4.01e-01\pm2.28e-02\pm3.21e-02$ \\ \hline 6.55e-01& 1.00e-01 & $1.39e-01\pm5.81e-03\pm1.12e-02$ & $2.91e-01\pm1.69e-02\pm2.33e-02$ \\ \hline 7.54e-01& 1.00e-01 & $9.21e-02\pm3.10e-03\pm7.37e-03$ & $1.90e-01\pm1.04e-02\pm1.52e-02$ \\ \hline 8.54e-01& 1.00e-01 & $7.11e-02\pm2.61e-03\pm5.69e-03$ & $1.45e-01\pm8.76e-03\pm1.16e-02$ \\ \hline 9.54e-01& 1.00e-01 & $4.70e-02\pm1.94e-03\pm3.76e-03$ & $9.49e-02\pm6.34e-03\pm7.59e-03$ \\ \hline 1.05e+00& 1.00e-01 & $3.11e-02\pm1.49e-03\pm2.48e-03$ & $6.56e-02\pm5.07e-03\pm5.25e-03$ \\ \hline 1.15e+00& 1.00e-01 & $2.28e-02\pm1.21e-03\pm1.82e-03$ & $4.74e-02\pm4.00e-03\pm3.79e-03$ \\ \hline 1.30e+00& 2.00e-01 & $1.61e-02\pm7.10e-04\pm1.29e-03$ & $3.70e-02\pm2.41e-03\pm2.96e-03$ \\ \hline 1.50e+00& 2.00e-01 & $9.47e-03\pm5.54e-04\pm7.58e-04$ & $2.01e-02\pm1.78e-03\pm1.61e-03$ \\ \hline 1.70e+00& 2.00e-01 & $4.44e-03\pm4.19e-04\pm3.55e-04$ & $---$ \\ \hline \end{tabular} \caption{$K^{-}$ spectra in minimum-bias and 0-20\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$. } \label{kaonminusspectratable1} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & 20\%-40\% & 40\%-100\% \\ \hline 4.57e-01& 1.00e-01 & $3.24e-01\pm1.95e-02\pm2.59e-02$ & $1.39e-01\pm8.25e-03\pm1.11e-02$ \\ \hline 5.56e-01& 1.00e-01 & $2.66e-01\pm1.58e-02\pm2.13e-02$ & $1.13e-01\pm6.63e-03\pm9.03e-03$ \\ \hline 6.55e-01& 1.00e-01 & $1.99e-01\pm1.18e-02\pm1.59e-02$ & $7.50e-02\pm4.61e-03\pm6.00e-03$ \\ \hline 7.54e-01& 1.00e-01 & $1.36e-01\pm7.60e-03\pm1.08e-02$ & $5.62e-02\pm3.16e-03\pm4.50e-03$ \\ \hline 8.54e-01& 1.00e-01 & $1.13e-01\pm6.64e-03\pm9.04e-03$ & $3.78e-02\pm2.47e-03\pm3.03e-03$ \\ \hline 9.54e-01& 1.00e-01 & $7.35e-02\pm4.83e-03\pm5.88e-03$ & $2.55e-02\pm1.82e-03\pm2.04e-03$ \\ \hline 1.05e+00& 1.00e-01 & $4.51e-02\pm3.57e-03\pm3.61e-03$ & $1.93e-02\pm1.51e-03\pm1.54e-03$ \\ \hline 1.15e+00& 1.00e-01 & $3.03e-02\pm2.74e-03\pm2.43e-03$ & $1.36e-02\pm1.20e-03\pm1.09e-03$ \\ \hline 1.30e+00& 2.00e-01 & $2.35e-02\pm1.66e-03\pm1.88e-03$ & $7.84e-03\pm6.41e-04\pm6.27e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.05e-02\pm1.20e-03\pm8.36e-04$ & $5.10e-03\pm5.72e-04\pm4.08e-04$ \\ \hline \end{tabular} \caption{$K^{-}$ spectra in 20-40\% and 40-100\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{kaonminusspectratable2} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & p+p \\ \hline 4.57e-01& 1.00e-01 & $5.99e-02\pm2.91e-03\pm4.79e-03$ \\ \hline 5.56e-01& 1.00e-01 & $4.79e-02\pm2.35e-03\pm3.83e-03$ \\ \hline 6.55e-01& 1.00e-01 & $3.06e-02\pm1.58e-03\pm2.45e-03$ \\ \hline 7.54e-01& 1.00e-01 & $2.15e-02\pm9.81e-04\pm1.72e-03$ \\ \hline 8.54e-01& 1.00e-01 & $1.46e-02\pm7.67e-04\pm1.17e-03$ \\ \hline 9.54e-01& 1.00e-01 & $1.01e-02\pm5.86e-04\pm8.09e-04$ \\ \hline 1.05e+00& 1.00e-01 & $7.87e-03\pm5.22e-04\pm6.30e-04$ \\ \hline 1.15e+00& 1.00e-01 & $5.25e-03\pm4.27e-04\pm4.20e-04$ \\ \hline 1.30e+00& 2.00e-01 & $3.42e-03\pm2.57e-04\pm2.74e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.75e-03\pm1.61e-04\pm1.40e-04$ \\ \hline \end{tabular} \caption{$K^{-}$ spectra in p+p collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{kaonminusspectratable3} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & M.B. & 0\%-20\% \\ \hline 4.68e-01& 1.00e-01 & $1.88e-01\pm1.87e-02\pm2.44e-02$ & $3.96e-01\pm4.51e-02\pm5.15e-02$ \\ \hline 5.63e-01& 1.00e-01 & $1.35e-01\pm1.11e-02\pm1.75e-02$ & $2.80e-01\pm2.67e-02\pm3.64e-02$ \\ \hline 6.61e-01& 1.00e-01 & $1.07e-01\pm8.89e-03\pm1.39e-02$ & $2.10e-01\pm2.01e-02\pm2.73e-02$ \\ \hline 7.59e-01& 1.00e-01 & $8.02e-02\pm6.78e-03\pm1.04e-02$ & $1.74e-01\pm1.67e-02\pm2.27e-02$ \\ \hline 8.58e-01& 1.00e-01 & $5.82e-02\pm5.20e-03\pm7.57e-03$ & $1.25e-01\pm1.27e-02\pm1.63e-02$ \\ \hline 9.57e-01& 1.00e-01 & $4.45e-02\pm4.85e-03\pm5.78e-03$ & $1.05e-01\pm1.25e-02\pm1.37e-02$ \\ \hline 1.06e+00& 1.00e-01 & $3.63e-02\pm4.11e-03\pm2.90e-03$ & $8.16e-02\pm1.02e-02\pm6.53e-03$ \\ \hline 1.16e+00& 1.00e-01 & $2.82e-02\pm2.11e-03\pm2.26e-03$ & $5.08e-02\pm5.11e-03\pm4.06e-03$ \\ \hline 1.31e+00& 2.00e-01 & $1.86e-02\pm1.14e-03\pm1.49e-03$ & $4.14e-02\pm3.21e-03\pm3.31e-03$ \\ \hline 1.50e+00& 2.00e-01 & $1.02e-02\pm7.85e-04\pm8.14e-04$ & $2.46e-02\pm2.34e-03\pm1.97e-03$ \\ \hline 1.70e+00& 2.00e-01 & $5.64e-03\pm3.24e-04\pm4.51e-04$ & $1.21e-02\pm1.03e-03\pm9.66e-04$ \\ \hline 1.90e+00& 2.00e-01 & $3.14e-03\pm2.33e-04\pm2.51e-04$ & $8.08e-03\pm8.38e-04\pm6.46e-04$ \\ \hline 2.25e+00& 5.00e-01 & $1.39e-03\pm9.47e-05\pm1.12e-04$ & $3.37e-03\pm3.21e-04\pm2.70e-04$ \\ \hline 2.75e+00& 5.00e-01 & $2.75e-04\pm3.88e-05\pm2.20e-05$ & $6.19e-04\pm1.26e-04\pm4.95e-05$ \\ \hline 3.50e+00& 1.00e+00 & $8.13e-05\pm1.37e-05\pm6.50e-06$ & $---$ \\ \hline \end{tabular} \caption{$p$ spectra in minimum-bias and 0-20\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$. } \label{protonspectratable1} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & 20\%-40\% & 40\%-100\% \\ \hline 4.68e-01& 1.00e-01 & $2.78e-01\pm3.24e-02\pm3.61e-02$ & $1.21e-01\pm1.40e-02\pm1.57e-02$ \\ \hline 5.63e-01& 1.00e-01 & $1.90e-01\pm1.84e-02\pm2.48e-02$ & $7.84e-02\pm7.61e-03\pm1.02e-02$ \\ \hline 6.61e-01& 1.00e-01 & $1.56e-01\pm1.49e-02\pm2.03e-02$ & $6.18e-02\pm5.96e-03\pm8.03e-03$ \\ \hline 7.59e-01& 1.00e-01 & $1.18e-01\pm1.14e-02\pm1.53e-02$ & $4.84e-02\pm4.71e-03\pm6.29e-03$ \\ \hline 8.58e-01& 1.00e-01 & $8.54e-02\pm8.68e-03\pm1.11e-02$ & $2.98e-02\pm3.12e-03\pm3.87e-03$ \\ \hline 9.57e-01& 1.00e-01 & $6.45e-02\pm7.77e-03\pm8.38e-03$ & $2.39e-02\pm2.93e-03\pm3.11e-03$ \\ \hline 1.06e+00& 1.00e-01 & $5.48e-02\pm6.88e-03\pm4.38e-03$ & $2.12e-02\pm2.69e-03\pm1.69e-03$ \\ \hline 1.16e+00& 1.00e-01 & $4.21e-02\pm3.97e-03\pm3.37e-03$ & $1.52e-02\pm1.49e-03\pm1.22e-03$ \\ \hline 1.31e+00& 2.00e-01 & $2.68e-02\pm2.09e-03\pm2.15e-03$ & $9.80e-03\pm7.91e-04\pm7.84e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.41e-02\pm1.38e-03\pm1.13e-03$ & $5.05e-03\pm5.17e-04\pm4.04e-04$ \\ \hline 1.70e+00& 2.00e-01 & $7.69e-03\pm7.02e-04\pm6.15e-04$ & $2.54e-03\pm2.55e-04\pm2.03e-04$ \\ \hline 1.90e+00& 2.00e-01 & $4.12e-03\pm5.31e-04\pm3.30e-04$ & $1.43e-03\pm2.03e-04\pm1.15e-04$ \\ \hline 2.25e+00& 5.00e-01 & $2.30e-03\pm2.28e-04\pm1.84e-04$ & $6.31e-04\pm7.57e-05\pm5.05e-05$ \\ \hline 2.75e+00& 5.00e-01 & $4.36e-04\pm1.13e-04\pm3.49e-05$ & $1.62e-04\pm4.14e-05\pm1.29e-05$ \\ \hline \end{tabular} \caption{$p$ spectra in 20-40\% and 40-100\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{protonspectratable2} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & p+p \\ \hline 4.68e-01& 1.00e-01 & $4.51e-02\pm5.24e-03\pm5.86e-03$ \\ \hline 5.63e-01& 1.00e-01 & $3.33e-02\pm3.29e-03\pm4.33e-03$ \\ \hline 6.61e-01& 1.00e-01 & $2.60e-02\pm2.57e-03\pm3.38e-03$ \\ \hline 7.59e-01& 1.00e-01 & $2.03e-02\pm2.12e-03\pm2.64e-03$ \\ \hline 8.58e-01& 1.00e-01 & $1.14e-02\pm1.30e-03\pm1.49e-03$ \\ \hline 9.57e-01& 1.00e-01 & $8.93e-03\pm1.20e-03\pm1.16e-03$ \\ \hline 1.06e+00& 1.00e-01 & $6.98e-03\pm1.00e-03\pm5.59e-04$ \\ \hline 1.16e+00& 1.00e-01 & $4.68e-03\pm5.95e-04\pm3.75e-04$ \\ \hline 1.31e+00& 2.00e-01 & $2.90e-03\pm3.14e-04\pm2.32e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.34e-03\pm2.04e-04\pm1.07e-04$ \\ \hline 1.70e+00& 2.00e-01 & $6.61e-04\pm6.44e-05\pm5.29e-05$ \\ \hline 1.90e+00& 2.00e-01 & $4.73e-04\pm5.58e-05\pm3.78e-05$ \\ \hline 2.25e+00& 5.00e-01 & $1.84e-04\pm2.23e-05\pm1.47e-05$ \\ \hline 2.75e+00& 5.00e-01 & $3.06e-05\pm7.26e-06\pm2.45e-06$ \\ \hline \end{tabular} \caption{$p$ spectra in p+p collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{protonspectratable3} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & M.B. & 0\%-20\% \\ \hline 4.68e-01& 1.00e-01 & $1.22e-01\pm7.62e-03\pm1.59e-02$ & $2.57e-01\pm2.17e-02\pm3.34e-02$ \\ \hline 5.63e-01& 1.00e-01 & $1.08e-01\pm5.87e-03\pm1.40e-02$ & $2.26e-01\pm1.62e-02\pm2.94e-02$ \\ \hline 6.61e-01& 1.00e-01 & $8.86e-02\pm4.69e-03\pm1.15e-02$ & $1.73e-01\pm1.23e-02\pm2.25e-02$ \\ \hline 7.59e-01& 1.00e-01 & $6.40e-02\pm3.47e-03\pm8.32e-03$ & $1.39e-01\pm9.76e-03\pm1.81e-02$ \\ \hline 8.58e-01& 1.00e-01 & $5.51e-02\pm3.04e-03\pm7.16e-03$ & $1.19e-01\pm8.58e-03\pm1.54e-02$ \\ \hline 9.57e-01& 1.00e-01 & $3.80e-02\pm2.30e-03\pm4.94e-03$ & $8.99e-02\pm6.92e-03\pm1.17e-02$ \\ \hline 1.06e+00& 1.00e-01 & $3.27e-02\pm1.41e-03\pm2.62e-03$ & $7.36e-02\pm5.01e-03\pm5.89e-03$ \\ \hline 1.16e+00& 1.00e-01 & $2.36e-02\pm1.12e-03\pm1.89e-03$ & $4.24e-02\pm3.50e-03\pm3.39e-03$ \\ \hline 1.31e+00& 2.00e-01 & $1.55e-02\pm6.16e-04\pm1.24e-03$ & $3.48e-02\pm2.14e-03\pm2.78e-03$ \\ \hline 1.50e+00& 2.00e-01 & $8.00e-03\pm4.04e-04\pm6.40e-04$ & $1.93e-02\pm1.46e-03\pm1.54e-03$ \\ \hline 1.70e+00& 2.00e-01 & $4.78e-03\pm2.97e-04\pm3.82e-04$ & $1.01e-02\pm9.74e-04\pm8.11e-04$ \\ \hline 1.90e+00& 2.00e-01 & $2.56e-03\pm2.04e-04\pm2.05e-04$ & $4.99e-03\pm6.54e-04\pm3.99e-04$ \\ \hline 2.25e+00& 5.00e-01 & $1.06e-03\pm7.97e-05\pm8.49e-05$ & $2.69e-03\pm2.85e-04\pm2.15e-04$ \\ \hline 2.75e+00& 5.00e-01 & $3.32e-04\pm4.61e-05\pm2.66e-05$ & $5.86e-04\pm1.35e-04\pm4.69e-05$ \\ \hline 3.50e+00& 1.00e+00 & $7.89e-05\pm1.53e-05\pm6.31e-06$ & $---$ \\ \hline \end{tabular} \caption{$\bar{p}$ spectra in minimum-bias and 0-20\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pbarspectratable1} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & 20\%-40\% & 40\%-100\% \\ \hline 4.68e-01& 1.00e-01 & $1.63e-01\pm1.45e-02\pm2.12e-02$ & $7.50e-02\pm6.44e-03\pm9.75e-03$ \\ \hline 5.63e-01& 1.00e-01 & $1.65e-01\pm1.20e-02\pm2.14e-02$ & $6.04e-02\pm4.55e-03\pm7.85e-03$ \\ \hline 6.61e-01& 1.00e-01 & $1.24e-01\pm8.94e-03\pm1.62e-02$ & $5.51e-02\pm3.89e-03\pm7.16e-03$ \\ \hline 7.59e-01& 1.00e-01 & $9.35e-02\pm6.78e-03\pm1.22e-02$ & $3.61e-02\pm2.68e-03\pm4.69e-03$ \\ \hline 8.58e-01& 1.00e-01 & $7.98e-02\pm5.96e-03\pm1.04e-02$ & $2.85e-02\pm2.22e-03\pm3.70e-03$ \\ \hline 9.57e-01& 1.00e-01 & $5.22e-02\pm4.33e-03\pm6.78e-03$ & $1.95e-02\pm1.68e-03\pm2.54e-03$ \\ \hline 1.06e+00& 1.00e-01 & $4.14e-02\pm3.20e-03\pm3.31e-03$ & $1.82e-02\pm1.37e-03\pm1.45e-03$ \\ \hline 1.16e+00& 1.00e-01 & $3.61e-02\pm2.82e-03\pm2.89e-03$ & $1.22e-02\pm1.04e-03\pm9.73e-04$ \\ \hline 1.31e+00& 2.00e-01 & $2.32e-02\pm1.50e-03\pm1.85e-03$ & $7.64e-03\pm5.47e-04\pm6.11e-04$ \\ \hline 1.50e+00& 2.00e-01 & $1.13e-02\pm9.51e-04\pm9.04e-04$ & $3.51e-03\pm3.35e-04\pm2.81e-04$ \\ \hline 1.70e+00& 2.00e-01 & $6.41e-03\pm6.63e-04\pm5.13e-04$ & $2.21e-03\pm2.48e-04\pm1.77e-04$ \\ \hline 1.90e+00& 2.00e-01 & $4.64e-03\pm5.37e-04\pm3.71e-04$ & $9.80e-04\pm1.70e-04\pm7.84e-05$ \\ \hline 2.25e+00& 5.00e-01 & $1.46e-03\pm1.78e-04\pm1.17e-04$ & $5.66e-04\pm7.10e-05\pm4.53e-05$ \\ \hline 2.75e+00& 5.00e-01 & $4.43e-04\pm1.38e-04\pm3.54e-05$ & $1.29e-04\pm4.62e-05\pm1.03e-05$ \\ \hline \end{tabular} \caption{$\bar{p}$ spectra in 20-40\% and 40-100\% d+Au collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pbarspectratable2} \end{scriptsize} \end{table} \begin{table}[h] \begin{scriptsize} \centering \begin{tabular}{|c|c|c|} \hline $p_{T}$ & $p_{T}$ width & p+p \\ \hline 4.68e-01& 1.00e-01 & $3.74e-02\pm2.60e-03\pm4.86e-03$ \\ \hline 5.63e-01& 1.00e-01 & $2.84e-02\pm1.77e-03\pm3.70e-03$ \\ \hline 6.61e-01& 1.00e-01 & $2.27e-02\pm1.38e-03\pm2.96e-03$ \\ \hline 7.59e-01& 1.00e-01 & $1.32e-02\pm8.65e-04\pm1.71e-03$ \\ \hline 8.58e-01& 1.00e-01 & $1.05e-02\pm7.20e-04\pm1.37e-03$ \\ \hline 9.57e-01& 1.00e-01 & $7.17e-03\pm5.40e-04\pm9.32e-04$ \\ \hline 1.06e+00& 1.00e-01 & $5.60e-03\pm3.77e-04\pm4.48e-04$ \\ \hline 1.16e+00& 1.00e-01 & $3.74e-03\pm2.86e-04\pm2.99e-04$ \\ \hline 1.31e+00& 2.00e-01 & $2.31e-03\pm1.49e-04\pm1.84e-04$ \\ \hline 1.50e+00& 2.00e-01 & $9.69e-04\pm8.84e-05\pm7.75e-05$ \\ \hline 1.70e+00& 2.00e-01 & $5.95e-04\pm6.65e-05\pm4.76e-05$ \\ \hline 1.90e+00& 2.00e-01 & $3.57e-04\pm4.79e-05\pm2.86e-05$ \\ \hline 2.25e+00& 5.00e-01 & $1.12e-04\pm1.90e-05\pm8.93e-06$ \\ \hline 2.75e+00& 5.00e-01 & $3.87e-05\pm1.07e-05\pm3.10e-06$ \\ \hline \end{tabular} \caption{$\bar{p}$ spectra in p+p collisions. The unit of $p_{T}$ and $p_{T}$ width is $GeV/c$.} \label{pbarspectratable3} \end{scriptsize} \end{table}
{ "timestamp": "2005-03-23T14:47:43", "yymm": "0503", "arxiv_id": "nucl-ex/0503018", "language": "en", "url": "https://arxiv.org/abs/nucl-ex/0503018" }
\section{Introduction } During the last two centuries in peaceful rich countries, people lived on average longer and longer, while during the last few decades the number of chidren born per women during her lifetime has sunken below the replacement rate of slightly above 2. Also in many poorer countries the number of births has fallen and the life expectancy increased. Thus the fear of overpopulation of our planet Earth has to be modified by fear of old-age poverty: In the year 2030 only those goods and services can be consumed by retired people which have been produced by working-age people. A million dollars of old-age savings can be halved by a ten-percent inflation rate over seven years, if not enough young people help me to live. This Econosociobiophysics problem is one of demography, not of money. We present in an appendix details of the assumptions for our extrapolations into the future. In the next section we deal with conditions as are typical for Western Europe, to be followed by a section on the different problems of Algeria. More literature on ageing models, including one applied to our demography \cite{cebrat}, is given in \cite{vancouver}. \section{Western Europe} Around 1970, the contraceptive pill reduced in the then two German states the average number of babies born by a women during her lifetime below the replacement level of two, to about 1.4. Spain and Italy followed later but levelled at a lower plateau, while in France the number is higher, about 1.7. Life expectancy rises further though slower than during the first half of the 20th century. Thus if people retire at an age of about 62 years, and if around 2030 the strongest age cohort in Germany are the 70-year olds, problems lie ahead. Only in recent years were they discussed in general newspapers. As in science in general, we need open publications of extrapolation methods and results. Only if many different simulations are compared can we see to what extent they agree and thus may be relied upon. The top curve in Fig.1 shows what happens if nothing is done: The average retirement age is 62 years, and immigration and emigration cancel each other. Then \cite{bomsdorf,stauffer,martins} the number of old people to be supported by working-age people will increase drastically, while the total population will decrease. We added here the number of children (up to age 20) to the pensioneers since both groups are not fully ''working'' in the usual sense. For the middle curve we assumed a net immigration of 0.38 percent per year, starting now, and an increase of the retirement age by about half of the increase of the life expectancy. Thus for every year which medical progress gives us, about six month are given like a tax to the labour market, while the other six months are leasure time after retirement. Now the ratio and the population are more stable. If we do not count in the latter simulation the children (bottom curve), then the ratio of pensioneers only to working age people is lower \cite{martins}. However, the reduction of the expenses for children is mainly an effect of the past, not of the future. \begin{figure} \begin{center} \includegraphics[angle=-90,scale=0.5]{granadar2.eps} \end{center} \caption{Ratio of number of pensioneers to number of working age people (+) and ratio of number of pensioneers plus number of children to working age people (x). (Algeria) } \end{figure} \begin{figure} \begin{center} \includegraphics[angle=-90,scale=0.5]{granadar3.eps} \end{center} \caption{Ratio of number of babies died before reaching the age one year to number of total birth(+) (data of the National Office of the Statistics ONS Algeria) the fit line shows fractions increase of about 20 percent from 1980 until 1901. } \end{figure} \section{Algeria} During the first half of the previous century, the fertility was very large in North Africa compared to Europe. It reached the value 8.1 during the seventies in Algeria because of a low average age of marriage in this country. Thirty years later, the average number of births per women (during her lifetime) becomes close to 2, whereas in France the fertility needed two centuries to pass from 6 in the middle of the 17th century to 2 in the 1930's. Algerian people are thus young. Figure 2 shows that the number of children (up to age $20$) added to the pensioneers (the retirement age in Algeria is $60$ years) obliged workers to support about two times their number until the year $2000$. Sixty years afterwards the population will be older but the fractions remain constant (no fear of increasing). We assumed in Fig.2 the Gompertz slope $b$ (see appendix) to increase with time from $0.07$ in 1901 to $0.082$ in 1971 and to remain constant thereafter. Only fertility data from 1950 on is available in Algeria. The fertility is constant with a mean value of $7.3$ from $1950$ to $1980$ and then decreases abruptly til 2004 to reach a value $2.04$; it is assumed to stay constant at this value thereafter. The sixty years period necessary to reach the steady state, corresponds to the age of retirement. In figure 3, we show that the ratio of the number of babies dying in their first year to the total number of births decreased by about $20$ percent from 1901 to 1980. Thus, we made a correction on the fertility data (in fig.2) by reducing them by the number of children dying before they reach maturity. We noticed also that the greatest emigration rate of Algerian people was between 1950 and 1970 but remains weak compared to the rate of births and does not influence the population evolution. In our simulation we then neglected the emigration in such calculations. However, this simulation did not account the rate of unemployeds which was very small during the period of socialism but reaches now 17 percent of population. However, the main prediction of Fig. 2 is an increase of the social load for old age by 400 percent starting from 2020, while that for children and old age combined will stabilize at the level around the year 2000. \section{Summary} With rising life expectencies and falling births, the demographic problems of rich countries can be alleviated by controlled immigration and a moderate increase of the average retirement age. That policy requires that first the unemployment is reduced appreciably. For Algeria, on the other hand, emigration could not affect sensitively the evolution of pensioneers, but their rate should be multiplied by a factor four after 15 years from now on which would create a real economic problem were it not offset by a reduction of the number of children. \bigskip LZ thanks the DAAD for supporting a one-year part of his thesis work in Cologne. We thank W.J. Paul for suggesting to add the children to the pensioneers. \section{Appendix} According to the Azbel lectures at this seminar, in all different countries and centuries, the probability of humans to survive up to a fixed age is a universal function of the life expectancy; we do not have to apply this universality to yeast cells for the purpose of human demography. Thus we use Germany as typical Western European country, without taking into account the effects of World War II. The mortality function $\mu = - d \ln S(a)/da$, where $S(a)$ is the number of survivors from birth to age $a$, is assumed to follow a Gompertz law for adults: $m \propto b\exp[(a-X)b]$ since the deviations at young age occur at such low mortalities that they are not relevant if we want to be accurate within a few percent. The deviations at old age \cite{robine} are not yet reliably established and may also be negligible as long as the fraction of centenarians among pensioneers is very small. The Gompertz slope $b$ was assumed to increase linearly with time from 0.07 in 1821 to 0.093 in 1971 and to stay constant thereafter, in contrast to Bomsdorf \cite{bomsdorf} and Azbel \cite{azbel} but in agreement with Yashin et al \cite{yashin}; see also Wilmoth et al \cite{wilmoth}. Instead, the characteristic age $X$ was constant at 103 years until 1971 and then increased each year by 0.15 years to give a rising life expectancy. Also these deviations from universality are not yet established reliably. (Therefore we ignored the effect for Algeria, keeping $X=103$ constant there.) Babies are born by mothers of age 21 to 40 with age-independent probability. The average number of children born per women over her lifetime and reaching adult age is assumed to be $2.2 - 0.4\tanh[(t-1971)/3]$ recently. Immigrants are assumed to be 6 to 40 years old with equal probability, and their number per year equals a fraction $c = 0.38 \%$ of the population, adjusted to give a constant total population. After the year 2010, the retirement age is increased by 60 percent of the increase of life expectancy at birth to 73 in 2100 at a life expectancy then of 99 years; for the problem year 2030 these ages are 64 and 84 years. The program is available from stauffer@thp.uni-koeln.de as rente16.f.
{ "timestamp": "2005-03-11T14:36:05", "yymm": "0503", "arxiv_id": "q-bio/0503015", "language": "en", "url": "https://arxiv.org/abs/q-bio/0503015" }
\section{Introduction}\label{sec:intro} Least squares fitting is a well-known and powerful method for combining information from a set of related experimental measurements to estimate the underlying theoretical parameters (see, for instance, Reference~\cite{pdg}). We discuss a specific implementation of this method for use in high-energy physics experiments, where the free parameters, denoted by the vector $\mathbf{m}$, are extracted from event yields for signal processes. Typically, these yields are subject to corrections for background, crossfeed, and efficiency. Because the sizes of these corrections depend on the values of the free parameters, we make all yield adjustments directly in the fit. Often, the uncertainties on these corrections are ignored during the fit and are propagated to the free parameters afterwards. However, if these uncertainties modify the relative weights of the measurements, then the above two-step procedure would bias both the fitted central values and the estimated uncertainties. Therefore, we build the $\chi^2$ variable from a full description of the uncertainties, statistical and systematic, as well as their correlations, on both the yields and their corrections. Thus, the input measurements --- event yields, signal efficiencies, parameters quantifying the background processes, and background efficiencies --- and their uncertainties are all treated in a uniform fashion. In the $\chi^2$ minimization, we account for the $\mathbf{m}$ dependence of the yield corrections. \section{Formalism}\label{sec:formalism} Below, we denote matrices by upper case bold letters and one-dimensional vectors by lower case bold letters. Let $\mathbf{n}$ represent a set of $N$ event yield measurements, each for a different signal process. Each measurement may receive crossfeed contributions from other signal processes as well as backgrounds from non-signal sources. The background processes are described by $\mathbf{b}$, a vector of $B$ estimated production yields, which can be functions of experimentally measured quantities, such as branching fractions, cross sections, and luminosities. In principle, the free parameters $\mathbf{m}$ can also appear in $\mathbf{b}$, although no additional degrees of freedom are introduced by $\mathbf{b}$. The rates at which these background processes contaminate the signal yields are given by the $N\times B$ background efficiency matrix, $\mathbf{F}$. Thus, the vector $\mathbf{s}\equiv \mathbf{n} - \mathbf{Fb}$ represents the background-subtracted yields. We use an $N\times N$ signal efficiency matrix, $\mathbf{E}$, to describe simultaneously detection efficiencies (diagonal elements) and crossfeed probabilities (off-diagonal elements). The elements $E_{ij}$ are defined to be the probabilities that an event of signal process $j$ is reconstructed and counted in yield $i$. The corrected yields, denoted by $\mathbf{c}$, are obtained by acting on $\mathbf{s}$ with the inverse of $\mathbf{E}$: \begin{equation}\label{eq:correctedYields} \mathbf{c} = \mathbf{E}^{-1} \mathbf{s} = \mathbf{E}^{-1}( \mathbf{n} - \mathbf{Fb} ). \end{equation} Thus, $\mathbf{c}$ encapsulates all the experimental measurements. The variance matrix of $\mathbf{c}$, denoted by $\mathbf{V_c}$, receives contributions, both statistical and systematic, from each element of $\mathbf{n}$, $\mathbf{b}$, $\mathbf{E}$, and $\mathbf{F}$. In the least squares fit, we define $\chi^2 \equiv \left(\mathbf{c}-\mathbf{\widetilde c}\right)^T \mathbf{V}_{\mathbf{c}}^{-1} \left(\mathbf{c}-\mathbf{\widetilde c}\right)$, where $\mathbf{\widetilde c}$ is the vector of predicted yields, which are also functions of $\mathbf{m}$. Because both $\mathbf{\widetilde c}$ and $\mathbf{c}$ (through $\mathbf{b}$) depend on $\mathbf{m}$, minimizing this $\chi^2$ amounts to a nonlinear version of the total least squares method~\cite{tls}. We solve this problem by extending the conventional least squares fit to include contributions from both $\mathbf{\widetilde c}$ and $\mathbf{c}$ in $\partial\chi^2/\partial\mathbf{m}$. Given a set of seed values, $\mathbf{m}_0$, the optimized estimate, $\mathbf{\widehat m}$, and its variance matrix, $\mathbf{V_m}$, are \begin{eqnarray} \label{eq:fittedParameters} \mathbf{\widehat m} &=& \mathbf{m}_0 + \left(\mathbf{D}\mathbf{V}_{\mathbf{c}}^{-1}\mathbf{D}^T\right)^{-1} \mathbf{D}\mathbf{V}_{\mathbf{c}}^{-1} \left[\mathbf{c}(\mathbf{m}_0)-\mathbf{\widetilde c}(\mathbf{m}_0)\right]\\ \label{eq:fittedError} \mathbf{V_m} &=& \frac{1}{2}\frac{\partial^2\chi^2}{\partial\mathbf{m}\, \partial\mathbf{m}^T} = \left(\mathbf{D}\mathbf{V}_{\mathbf{c}}^{-1}\mathbf{D}^T\right)^{-1}, \end{eqnarray} where the $M\times N$ derivative matrix $\mathbf{D}$ is defined to be \begin{equation} \mathbf{D}\equiv \frac{\partial\mathbf{\widetilde c}}{\partial\mathbf{m}} - \frac{\partial\mathbf{c}}{\partial\mathbf{m}} = \frac{\partial\mathbf{\widetilde c}}{\partial\mathbf{m}} + \frac{\partial\mathbf{b}}{\partial\mathbf{m}} \mathbf{F}^T\left(\mathbf{E}^{-1}\right)^T. \end{equation} In general, $\mathbf{\widetilde c}$ and $\mathbf{c}$ are nonlinear functions of $\mathbf{m}$, so the linearized solution $\mathbf{\widehat m}$ is approximate, and the above procedure is iterated until the $\chi^2$ converges. Between iterations, all the fit inputs that depend on $\mathbf{m}$ are reevaluated with the updated values of $\mathbf{\widehat m}$. Nonlinearities also occur when $\mathbf{V_c}$ contains multiplicative or Poisson uncertainties that depend on the measurement values. With the least squares method, these nonlinearities result in biased estimators unless these variable uncertainties are evaluated using the predicted yields $\mathbf{\widetilde c}$ instead of the measured $\mathbf{c}$. Therefore, all three ingredients in the $\chi^2$ --- $\mathbf{c}$, $\mathbf{\widetilde c}$, and $\mathbf{V_c}$ --- are functions of $\mathbf{m}$. However, we do not include the derivatives $\partial\mathbf{V_c}/\partial\mathbf{m}$ in $\mathbf{D}$ because doing so would generate biases in $\mathbf{\widehat m}$. For a simple demonstration of the aforementioned biases, we consider two measured yields, $c_1$ and $c_2$, which are both estimators of a true yield $\bar c$. We assume that the uncertainties on $c_1$ and $c_2$ are uncorrelated, multiplicative, and of the same fractional size, $\lambda$. We construct an improved estimator, $\widehat c$, by minimizing $\chi^2 = (c_1-c)^2/\sigma_{c_1}^2 + (c_2-c)^2/\sigma_{c_2}^2$ with respect to $c$. If, following the prescription given above, we neglect the $\partial\sigma_{c_i}^2/\partial c$ terms in $\partial\chi^2/\partial c$ and assign (iteratively) the uncertainties $\sigma_{c_1}=\sigma_{c_2}=\lambda\widehat c$, then $c_1$ and $c_2$ are equally weighted, and $\widehat c$ is an unbiased estimate of $\bar c$: \begin{eqnarray} \widehat c_{\rm unbiased} &=& \frac{c_1+c_2}{2} \\ \chi^2_{\rm unbiased} &=& \frac{2}{\lambda^2}\left(\frac{c_1-c_2}{c_1+c_2}\right)^2 . \end{eqnarray} On the other hand, including the $\partial\sigma_{c_i}^2/\partial c$ terms in $\partial\chi^2/\partial c$ results in an upward bias: \begin{eqnarray} \widehat c_{\rm biased1} &=& \frac{c_1^2+c_2^2}{c_1+c_2} = \widehat c_{\rm unbiased}\left(1+\frac{\lambda^2\chi^2_{\rm unbiased}}{2}\right) \\ \chi^2_{\rm biased1} &=& \frac{(c_1-c_2)^2}{\lambda^2(c_1^2+c_2^2)} . \end{eqnarray} Finally, if we assign uncertainties based on the measured yields, not the predicted yields, such that $\sigma_{c_1}=\lambda c_1$, $\sigma_{c_2}=\lambda c_2$, and $\partial\sigma_{c_i}^2/\partial c=0$, then the resulting estimate is biased low: \begin{eqnarray} \widehat c_{\rm biased2} &=& \frac{c_1 c_2 (c_1+c_2)}{c_1^2+c_2^2} = \widehat c_{\rm unbiased}(1-\lambda^2\chi^2_{\rm biased1}) \\ \chi^2_{\rm biased2} &=& \chi^2_{\rm biased1}. \end{eqnarray} Thus, even though $\chi^2_{\rm biased1}$ and $\chi^2_{\rm biased2}$ are smaller than $\chi^2_{\rm unbiased}$, the corresponding estimators possess undesired properties. \section{\boldmath Input Variance Matrix} \label{sec:inputVarianceMatrix} The uncertainties on the $N$ elements of $\mathbf{n}$ and the $B$ elements of $\mathbf{b}$ are characterized by the $N\times N$ matrix $\mathbf{V_n}$ and the $B\times B$ matrix $\mathbf{V_b}$, respectively. Usually, the elements of $\mathbf{E}$ and $\mathbf{F}$ share many common correlated systematic uncertainties, so we construct a joint variance matrix from the submatrices $\mathbf{V_E}$, $\mathbf{V_F}$, and $\mathbf{C_{EF}}$, where $\mathbf{V_E}$ ($N^2\times N^2$) and $\mathbf{V_F}$ ($NB\times NB$) are the variance matrices for the elements of $\mathbf{E}$ and $\mathbf{F}$, respectively, and $\mathbf{C_{EF}}$ ($N^2\times NB$) contains the correlations between $\mathbf{E}$ and $\mathbf{F}$. Below, we label each element of $\mathbf{E}$ or $\mathbf{F}$ by two indices ($E_{ij}$ or $F_{ij}$), and the two dimensions of $\mathbf{E}$ or $\mathbf{F}$ are mapped onto one dimension of $\mathbf{V_E}$ or $\mathbf{V_F}$. We form $\mathbf{V_c}$ by propagating the statistical and systematic uncertainties on $\mathbf{n}$, $\mathbf{b}$, $\mathbf{E}$, and $\mathbf{F}$ to $\mathbf{c}$ via \begin{equation} \label{eq:errorPropagation1} \mathbf{V_c} = \frac{\partial\mathbf{c}}{\partial\mathbf{n}}^T \mathbf{V_n} \frac{\partial\mathbf{c}}{\partial\mathbf{n}} + \frac{\partial\mathbf{c}}{\partial\mathbf{b}}^T \mathbf{V_b} \frac{\partial\mathbf{c}}{\partial\mathbf{b}} + \left(\begin{array}{cc} (\partial\mathbf{c}/\partial\mathbf{E})^T & (\partial\mathbf{c}/\partial\mathbf{F})^T \end{array}\right) \left(\begin{array}{cc} \mathbf{V_E} & \mathbf{C_{EF}} \\ \mathbf{C}_{\mathbf{EF}}^T & \mathbf{V_F} \end{array}\right) \left(\begin{array}{c} \partial\mathbf{c}/\partial\mathbf{E} \\ \partial\mathbf{c}/\partial\mathbf{F} \end{array}\right). \end{equation} Where appropriate, we substitute $\mathbf{\widetilde c}$ for $\mathbf{c}$, as discussed in Section~\ref{sec:formalism}. The first term of Equation~\ref{eq:errorPropagation1} is simply $\mathbf{E}^{-1}\mathbf{V_n} (\mathbf{E}^{-1})^T$, and the second term is $\mathbf{E}^{-1}\mathbf{F}\mathbf{V_b}\mathbf{F}^T (\mathbf{E}^{-1})^T$. For the third term, we evaluate the partial derivatives and find \begin{eqnarray} \frac{\partial\mathbf{c}}{\partial\mathbf{E}} &=& \mathbf{s}^T \left(\frac{\partial\mathbf{E}^{-1}}{\partial\mathbf{E}}\right)^T = -\mathbf{s}^T \left(\mathbf{E}^{-1}\right)^T \left(\frac{\partial\mathbf{E}}{\partial\mathbf{E}}\right)^T \left(\mathbf{E}^{-1}\right)^T = -\mathbf{A}\left(\mathbf{E}^{-1}\right)^T \\ \frac{\partial\mathbf{c}}{\partial\mathbf{F}} &=& -\mathbf{B}\left(\mathbf{E}^{-1}\right)^T, \end{eqnarray} where $\mathbf{A}\equiv\mathbf{c}^T (\partial\mathbf{E}/\partial\mathbf{E})^T$ and $\mathbf{B}\equiv\mathbf{b}^T (\partial\mathbf{F}/\partial\mathbf{F})^T$, with elements given in terms of the Kronecker delta ($\delta_{ij}$): $\partial E_{kl}/\partial E_{ij}=\partial F_{kl}/\partial F_{ij}=\delta_{ik}\delta_{jl}$. The matrices $\mathbf{A}$ and $\mathbf{B}$ have rows labeled by two indices, which refer to the elements of $\mathbf{E}$ and $\mathbf{F}$, respectively, and columns labeled by one index, which refers to the elements of $\mathbf{c}$. In other words, the $ij$-th row of $\mathbf{A}$ is given by $\mathbf{c}^T (\partial\mathbf{E}/\partial E_{ij})^T$, where $(\partial\mathbf{E}/\partial E_{ij})_{kl} = \partial E_{kl}/\partial E_{ij}$. Therefore, the elements of $\mathbf{A}$ and $\mathbf{B}$ are $A_{ij,k} = \delta_{ik} \widetilde c_j$ and $B_{ij,k} = \delta_{ik} b_j$. For $N=B=2$, these matrices are \begin{equation} \label{eq:Adefinition} \mathbf{A} = \left(\begin{array}{cc} \widetilde c_1 & 0 \\ \widetilde c_2 & 0 \\ 0 & \widetilde c_1 \\ 0 & \widetilde c_2 \\ \end{array}\right) \hspace{1cm}{\rm and}\hspace{1cm} \mathbf{B} = \left(\begin{array}{cc} b_1 & 0 \\ b_2 & 0 \\ 0 & b_1 \\ 0 & b_2 \\ \end{array}\right). \end{equation} This treatment of error propagation in matrix inversion agrees with that derived in Reference~\cite{Lefebvre:1999yu}. The above relations allow us to reexpress $\mathbf{V_c}$ as \begin{equation} \label{eq:errorPropagation2} \mathbf{V_c} = \mathbf{E}^{-1}\mathbf{V_{\Delta n}} \left(\mathbf{E}^{-1}\right)^T, \end{equation} where $\mathbf{V_{\Delta n}}\equiv \mathbf{V_n} + \mathbf{F}\mathbf{V_b}\mathbf{F}^T + \mathbf{A}^T \mathbf{V_E} \mathbf{A} + \mathbf{B}^T \mathbf{V_F} \mathbf{B} + \mathbf{A}^T \mathbf{C_{EF}} \mathbf{B} + \mathbf{B}^T \mathbf{C}_{\mathbf{EF}}^T \mathbf{A}$. As a result, we have $\chi^2 = \mathbf{\Delta n}^T \mathbf{V}_{\mathbf{\Delta n}}^{-1}\mathbf{\Delta n}$, where $\mathbf{\Delta n}\equiv\mathbf{n}-\mathbf{E\widetilde c}-\mathbf{Fb}$. Thus, the $\chi^2$ minimization can be formulated equivalently in terms of $\mathbf{n}$ instead of $\mathbf{c}$: $\mathbf{V_m} = \left(\mathbf{D'}\mathbf{V}_{\mathbf{\Delta n}}^{-1}\mathbf{D'}^T\right)^{-1}$ and $\mathbf{\widehat m} = \mathbf{m}_0 + \mathbf{V_m} \mathbf{D'}\mathbf{V}_{\mathbf{\Delta n}}^{-1} \mathbf{\Delta n}$, where $\mathbf{D'}\equiv \mathbf{D}\mathbf{E}^T = (\partial\mathbf{\widetilde c}/\partial\mathbf{m})\mathbf{E}^T +(\partial\mathbf{b}/\partial\mathbf{m})\mathbf{F}^T$. Systematic uncertainties on the efficiencies are often multiplicative and belong to one of three categories: those that depend only on the reconstructed mode (row-wise), those that depend only on the generated mode (column-wise), and those that are uncorrelated among elements of $\mathbf{E}$ and $\mathbf{F}$. For row-wise efficiency uncertainties, all the elements in any given row of $\mathbf{E}$ and $\mathbf{F}$ have the same fractional uncertainty, which we denote by $\lambda_i \equiv \sigma_{E_{ij}}/E_{ij} = \sigma_{F_{ij}}/F_{ij}$. The correlation coefficients between elements of different rows are $\lambda_{ij}^2 / (\lambda_i\lambda_j)$, where $\lambda_{ij}$ characterizes the uncertainties common to $c_i$ and $c_j$. For instance, if $\lambda_{\rm track}$ is the fractional uncertainty associated with the charged particle tracking efficiency, then $\lambda_i = t_i\lambda_{\rm track}$ and $\lambda_{ij}^2 = t_i t_j \lambda_{\rm track}^2$, where $t_i$ and $t_j$ are the track multiplicities in modes $i$ and $j$, respectively. Note that $\lambda_{ii} = \lambda_i$. Similarly, for column-wise uncertainties, we define the fractional uncertainties $\mu_j\equiv \sigma_{E_{ij}}/E_{ij} = \sigma_{F_{ij}}/F_{ij}$ and correlation coefficients $\mu_{ij}^2/(\mu_i\mu_j)$. We denote the uncorrelated fractional uncertainty on any element of $\mathbf{E}$ or $\mathbf{F}$ by $\nu_{ik, jl}$. Table~\ref{tab:vefElements} gives expressions for the elements of $\mathbf{V_E}$, $\mathbf{V_F}$, and $\mathbf{C_{EF}}$, as well as their contributions to $\mathbf{V_c}$ for row-wise, column-wise, and uncorrelated uncertainties. \begin{table}[ht] \begin{center} \caption{Expressions for the elements of $\mathbf{V_E}$, $\mathbf{V_F}$, and $\mathbf{C_{EF}}$, as well as their contributions to $\mathbf{V_c}$. Repeated external indices are not summed over.} \label{tab:vefElements} \begin{tabular}{cccc} \hline\hline Quantity & Row-wise & Column-wise & Uncorrelated \\ \hline $(\mathbf{V_E})_{ik, jl}$ & $\lambda_{ij}^2 E_{ik}E_{jl}$ & $\mu_{kl}^2 E_{ik}E_{jl}$ & $\nu_{ik,jl}^2 E_{ik}E_{jl}\delta_{ij}\delta_{kl}$ \\ $(\mathbf{V_F})_{ik, jl}$ & $\lambda_{ij}^2 F_{ik}F_{jl}$ & $\mu_{kl}^2 F_{ik}F_{jl}$ & $\nu_{ik,jl}^2 F_{ik}F_{jl}\delta_{ij}\delta_{kl}$ \\ $(\mathbf{C_{EF}})_{ik, jl}$ & $\lambda_{ij}^2 E_{ik}F_{jl}$ & $\mu_{kl}^2 E_{ik}F_{jl}$ & 0 \\ \hline $(\mathbf{A}^T \mathbf{V_E} \mathbf{A})_{ij}$ & $\lambda_{ij}^2 \widetilde s_i \widetilde s_j$ & $\mu_{kl}^2 E_{ik} \widetilde c_k E_{jl} \widetilde c_l$ & $\delta_{ij} \sigma^2_{E_{jk}} \widetilde c_k^2$ \\ $(\mathbf{B}^T \mathbf{V_F} \mathbf{B})_{ij}$ & $\lambda_{ij}^2 F_{ik}b_k F_{jl}b_l$ & $\mu_{kl}^2 F_{ik}b_k F_{jl}b_l$ & $\delta_{ij} \sigma_{F_{jk}}^2 b_k^2$ \\ $(\mathbf{A}^T \mathbf{C_{EF}} \mathbf{B})_{ij}$& $\lambda_{ij}^2 \widetilde s_i F_{jk}b_k$ & $\mu_{kl}^2 E_{ik} \widetilde c_k F_{jl}b_l$ & 0 \\ \hline\hline \end{tabular} \end{center} \end{table} \section{\boldmath Example: Hadronic $D$ Meson Branching Fractions} The least squares method described in the previous sections has been employed by the CLEO-c collaboration~\cite{cleoc-dhad} to measure absolute branching fractions for hadronic $D$ meson decays. Using $D\bar D$ pairs produced through the $\psi(3770)$ resonance, the branching fraction for mode $i$, denoted by ${\cal B}_i$, is measured by comparing the number of events where a single $D\to i$ decay is reconstructed (called single tag, denoted by $x_i$) with the number of events where both $D$ and $\bar D$ are reconstructed via $D\to i$ and $\bar D\to j$ (called double tag, denoted by $y_{ij}$). These yield measurements form the vector $\mathbf{n}$. The free parameters $\mathbf{m}$ are the ${\cal B}_i$ and the numbers of $D^0\bar D^0$ and $D^+D^-$ pairs produced, denoted by ${\cal N}^{00}$ and ${\cal N}^{+-}$, respectively, and denoted generically by ${\cal N}$. Yields for charge conjugate modes are measured separately, so the predicted corrected yields $\mathbf{\widetilde c}$ are ${\cal N}{\cal B}_i$ for single tags and ${\cal N}{\cal B}_i{\cal B}_j$ for double tags. Thus, ${\cal B}_i$ and ${\cal N}$ can be extracted from various products and ratios of $x_i$, $x_j$, and $y_{ij}$: ${\cal B}_i \sim y_{ij}/x_j$, ${\cal N}\sim x_i x_j/y_{ij}$, up to corrections for efficiency, crossfeed, and background. The matrix $\mathbf{V_n}$ describes the statistical uncertainties and correlations among the $x_i$ and $y_{ij}$. The $y_{ij}$ are uncorrelated, but because any given event can contain both single tag and double tag candidates, the $x_i$ are correlated among themselves as well as with the $y_{ij}$. If the selection criteria for single and double tags are the same, then the events (signal and background) used to estimate $y_{ij}$ are a proper subset of those for $x_i$ and $x_j$. Thus, any single tag yield is a sum of exclusive single tags ($x_i^{\rm excl}$) and double tags: $x_{\{i,j\}} = x_{\{i,j\}}^{\rm excl} + y_{ij}$. Propagating the uncertainties on the independent variables, $x_i^{\rm excl}$, $x_j^{\rm excl}$, and $y_{ij}$, gives the following elements for $\mathbf{V_n}$: \begin{eqnarray} \label{eq:singleVar} \langle \Delta x_i \Delta x_j \rangle &=& \delta_{ij} \sigma_{x_i} \sigma_{x_j} + (1-\delta_{ij}) \sigma^2_{y_{ij}} \\ \label{eq:doubleVar} \langle \Delta y_{ij} \Delta y_{kl} \rangle &=& \delta_{ik}\delta_{jl}\sigma_{y_{ij}}\sigma_{y_{kl}} \\ \label{eq:singleDoubleCovar} \langle \Delta x_i \Delta y_{jk} \rangle &=& (\delta_{ij} + \delta_{ik}) \sigma^2_{y_{jk}}, \end{eqnarray} where $\Delta x_i\equiv x_i - \langle x_i\rangle$, $\Delta y_{ij} \equiv y_{ij} - \langle y_{ij}\rangle$, and $\sigma_{x_{\{i,j\}}}^2=\sigma_{x_{\{i,j\}}^{\rm excl}}^2 + \sigma_{y_{ij}}^2$. Thus, for any two single tag yields and the corresponding double tag yield, the three off-diagonal elements of $\mathbf{V_n}$ are all given by the uncertainty on the number of overlapping events. In addition to these statistical uncertainties, $\mathbf{V_n}$ can also receive contributions from additive systematic uncertainties. Some of the sources of background we consider are non-signal $D$ decays, $e^+e^-\to q\bar q$ events, and $e^+e^-\to\tau^+\tau^-$ events. If there are two non-signal $D$ backgrounds with branching fractions ${\cal C}_1$ and ${\cal C}_2$, then the vector $\mathbf{b}$ is given by \begin{equation}\label{eq:backrounds} \mathbf{b} = \left(\begin{array}{c} {\cal N}{\cal C}_1 \\ {\cal N}{\cal C}_2 \\ {\cal L} X_{q\bar q} \\ {\cal L} X_{\tau^+\tau^-} \end{array}\right), \end{equation} where $X_{q\bar q}$ and $X_{\tau^+\tau^-}$ are the cross sections for $q\bar q$ and $\tau^+\tau^-$ production, respectively, and ${\cal L}$ is the integrated luminosity of the data sample. Because of the non-signal $D$ decays, the free parameter ${\cal N}$ appears in $\mathbf{b}$ but does not contribute any additional terms to the variance matrix $\mathbf{V_b}$, which takes the following block diagonal form: \begin{equation} \mathbf{V_b} = \left(\begin{array}{cccc} {\cal N}^2\sigma_{{\cal C}_1}^2 & 0 & 0 & 0 \\ 0 & {\cal N}^2\sigma_{{\cal C}_2}^2 & 0 & 0 \\ 0 & 0 & {\cal L}^2\sigma_{X_{q\bar q}}^2 + X_{q\bar q}^2\sigma_{\cal L}^2 & X_{q\bar q} X_{\tau^+\tau^-} \sigma_{\cal L}^2 \\ 0 & 0 & X_{q\bar q} X_{\tau^+\tau^-} \sigma_{\cal L}^2 & {\cal L}^2\sigma_{X_{\tau^+\tau^-}}^2 + X_{\tau^+\tau^-}^2 \sigma_{\cal L}^2 \end{array}\right). \end{equation} Also, the matrix $\partial\mathbf{b}/\partial\mathbf{m}$ is nontrivial and is incorporated into the $\chi^2$ minimization. In the joint variance matrix for $\mathbf{E}$ and $\mathbf{F}$, uncertainties of all three types discussed in Section~\ref{sec:inputVarianceMatrix} are present. Row-wise effects arise from systematic uncertainties on simulated reconstruction efficiencies for charged tracks, $\pi^0\to\gamma\gamma$ decays, $K^0_S\to\pi^+\pi^-$ decays, and particle identification (PID) for charged pions and kaons. Column-wise uncertainties reflect the poorly known resonant substructure in multi-body final states. Uncorrelated contributions come from statistical uncertainties due to the finite Monte Carlo (MC) simulated samples used to determine $\mathbf{E}$ and $\mathbf{F}$. Thus, for example, if mode $i$ is $D^0\to K^-\pi^+\pi^0$ and mode $j$ is $D^+\to K^0_S\pi^+$, then the row-wise uncertainties are given by \begin{eqnarray} \lambda_i^2 & = & ( 2\lambda_{\rm track} )^2 + \lambda_{\pi^0}^2 + \lambda_{\pi^\pm {\rm PID}}^2 + \lambda_{K^\pm {\rm PID}}^2 \\ \lambda_j^2 & = & ( 3\lambda_{\rm track} )^2 + \lambda_{\pi^\pm {\rm PID}}^2 \\ \lambda_{ij}^2 & = & 6\lambda_{\rm track}^2 + \lambda_{\pi^\pm {\rm PID}}^2. \end{eqnarray} Because these row-wise and column-wise uncertainties are completely correlated among the yields to which they pertain, they degrade the precision of ${\cal B}_i$ but not ${\cal N}$. Furthermore, they have no effect on the central values of $\mathbf{\widehat m}$ because the relative weight of each yield is unaltered by these uncertainties. However, they can introduce large systematic correlations among the fit parameters, even between statistically independent branching fractions of different charge. \subsection{Toy Monte Carlo Study}\label{sec:toyMC} We test the method presented above using a toy MC simulation with Gaussian smearing of the fit inputs. We generate data for five decay modes, $D^0\to K^-\pi^+$, $D^0\to K^-\pi^+\pi^0$, $D^0\to K^-\pi^+\pi^-\pi^+$, $D^+\to K^-\pi^+\pi^+$, and $D^+\to K^0_S\pi^+$ (charge conjugate particles are implied), for which there are ten single tag and thirteen double tag yields. The fit determines seven free parameters: ${\cal N}^{00}$, ${\cal N}^{+-}$, and five charge-averaged branching fractions. The input branching fractions are taken to be the world-average values given in Reference~\cite{pdg}, and we use ${\cal N}^{00}=2.0\times 10^5$ and ${\cal N}^{+-}=1.5\times 10^5$. The efficiencies are mode-dependent: 30\%--70\% for single tags and 10\%--50\% for double tags, with fractional statistical uncertainties of 0.5\%--1.0\%. The yield uncertainties are specified to be close to the Poisson limit, and backgrounds correspond roughly to those expected in 60 ${\rm pb}^{-1}$ of $e^+e^-$ collisions at the $\psi(3770)$. Also, we apply correlated systematic efficiency uncertainties of 1\% for tracking, 2\% for $\pi^0$ reconstruction, 2\% for $K^0_S$ reconstruction, and 1\% for charged pion and kaon PID. The fit reproduces the input parameters well. Figure~\ref{fig:toyMCPulls} shows the pull distributions for the seven fit parameters and the fit confidence level for 10000 toy MC trials. All the pull distributions are unbiased and have widths consistent with unity. Also, the confidence level is flat. Table~\ref{tab:correlations} gives the correlation coefficients among the fit parameters. Branching fractions tend to be positively correlated with each other and negatively correlated with ${\cal N}^{00}$ and ${\cal N}^{+-}$. In particular, the $D^0$ branching fractions are correlated with those for $D^+$. In the absence of correlated efficiency uncertainties, the $D^0$ and $D^+$ free parameters would essentially be independent. \begin{figure} \includegraphics*[width=0.5\linewidth]{3950205-001.eps} \caption{Toy MC fit pull distributions for ${\cal N}^{00}$ (a), ${\cal B}(D^0\to K^-\pi^+)$ (b), ${\cal B}(D^0\to K^-\pi^+\pi^0)$ (c), ${\cal B}(D^0\to K^-\pi^+\pi^-\pi^+)$ (d), ${\cal N}^{+-}$ (e), ${\cal B}(D^+\to K^-\pi^+\pi^+)$ (f), and ${\cal B}(D^+\to K^0_S\pi^+)$ (g), overlaid with Gaussian curves with zero mean and unit width. The fit confidence level distribution (h) is overlaid with a line with zero slope.} \label{fig:toyMCPulls} \end{figure} \begin{table}[htb] \caption{Correlation coefficients, including systematic uncertainties, for the free parameters determined by the fit to toy MC samples.} \label{tab:correlations} \begin{center} \begin{tabular}{l|ccccccc} \hline\hline & ~~${\cal N}^{00}$~~ & ~~$K^-\pi^+$~~ & ~~$K^-\pi^+\pi^0$~~ & ~~$K^-\pi^+\pi^-\pi^+$~~ & ~~${\cal N}^{+-}$~~ & ~~$K^-\pi^+\pi^+$~~ & ~~$K^0_S\pi^+$~~ \\ \hline ${\cal N}^{00}$ & 1 & $-0.63$ & $-0.52$ & $-0.38$ & $-0.01$ & $-0.01$ & $-0.01$ \\ $K^-\pi^+$ & & 1 & 0.79 & 0.87 & $-0.01$ & 0.40 & 0.29 \\ $K^-\pi^+\pi^0$ & & & 1 & 0.77 & $-0.01$ & 0.37 & 0.27 \\ $K^-\pi^+\pi^-\pi^+$ & & & & 1 & $-0.01$ & 0.53 & 0.39 \\ ${\cal N}^{+-}$ & & & & & 1 & $-0.82$ & $-0.77$ \\ $K^-\pi^+\pi^+$ & & & & & & 1 & 0.87 \\ $K^0_S\pi^+$ & & & & & & & 1\\ \hline\hline \end{tabular} \end{center} \end{table} Slight asymmetries can be observed in the pull distributions, especially in those for ${\cal N}^{00}$ and ${\cal N}^{+-}$. These asymmetries are caused by the nonlinear nature of the multiplicative efficiency uncertainties and of the functions $\mathbf{\widetilde c}(\mathbf{m})$. Because the fit parameters are effectively estimated from ratios of the input yields, Gaussian fluctuations in the denominators produce non-Gaussian fluctuations in the ratios, which are most visible in ${\cal N}^{00}$ and ${\cal N}^{+-}$, where the uncertainties in the denominators are dominant. Similarly, multiplicative uncertainties, which affect only the branching fractions, scale with the fitted values and, therefore, give rise to asymmetric ${\cal B}$ pulls. In both cases, larger fractional uncertainties would heighten the asymmetries. If we form the matrix $\mathbf{A}$ in Equation~\ref{eq:Adefinition} using the measured yields $\mathbf{c}$ rather than the predicted yields $\mathbf{\widetilde c}$, then the variance matrix $\mathbf{V_c}$ need not be reevaluated after each fit iteration. However, in this case, the pull distributions become significantly biased, as shown in Figure~\ref{fig:toyMCPullsBiased}. Thus, obtaining unbiased fit results and the correct uncertainties requires proper handling of the efficiency variance matrices $\mathbf{V_E}$ and $\mathbf{V_F}$. \begin{figure} \includegraphics*[width=0.5\linewidth]{3950805-002.eps} \caption{Toy MC fit pull distributions, with $\mathbf{V_c}$ calculated using $\mathbf{c}$ instead of $\mathbf{\widetilde c}$, for ${\cal N}^{00}$ (a), ${\cal B}(D^0\to K^-\pi^+)$ (b), ${\cal B}(D^0\to K^-\pi^+\pi^0)$ (c), ${\cal B}(D^0\to K^-\pi^+\pi^-\pi^+)$ (d), ${\cal N}^{+-}$ (e), ${\cal B}(D^+\to K^-\pi^+\pi^+)$ (f), and ${\cal B}(D^+\to K^0_S\pi^+)$ (g), overlaid with Gaussian curves with zero mean and unit width. The fit confidence level distribution (h) is overlaid with a line with zero slope.} \label{fig:toyMCPullsBiased} \end{figure} \section{Summary} We have developed a least squares fit that simultaneously incorporates statistical and systematic uncertainties, as well as their correlations, on all the input experimental measurements. Biases from nonlinearities are reduced by introducing fit parameter dependence in the input variance matrix. This fitting method is used to measure absolute branching fractions of hadronic $D$ meson decays, and toy Monte Carlo studies validate the performance of the fitter. By including all known sources of measurement uncertainty in the $\chi^2$, we obtain unbiased fit parameters with correct estimated uncertainties. \begin{acknowledgments} We wish to thank Roy Briere, David Cassel, Lawrence Gibbons, Wolfgang Rolke, Anders Ryd, and Ian Shipsey for many helpful discussions. This work was supported in part by the National Science Foundation under Grant No. PHY-0202078. \end{acknowledgments}
{ "timestamp": "2005-12-20T00:30:07", "yymm": "0503", "arxiv_id": "physics/0503050", "language": "en", "url": "https://arxiv.org/abs/physics/0503050" }
\section{Introduction} \label{sec:intro} Various new challenging problems in shape matching have been appearing from different scientific areas including Bioinformatics and Image Analysis. In a class of problems in Shape Analysis, one assumes that the points in two or more configurations are labelled and these configurations are to be matched after filtering out some transformation. Usually the transformation is a rigid transformation or similarity transformation. Several new problems are appearing where the points of configuration are either not labelled or the labelling is ambiguous, and in which some points do not appear in each of the configurations. An example of ambiguous labelling arises in understanding the secondary structure of proteins, where we are given not only the 3-dimensional molecular configuration but also the type of molecules (amino acids) at each point. A generic problem is to match such two configurations, where the matching has to be invariant under some transformation group. Descriptions of such problems can be found in the review article by Mardia, Taylor and Westhead (2003). We now describe two datasets related to protein structure. One is of 2-dimensional gel data where each point is a protein itself and the transformation group is affine. In this case we have a partial matching identified already by experts, that we can use to assess our procedures. In the second example we have a 3-dimensional configuration of two active sites of two proteins which has also additional chemical information. Here the underlying transformation to be filtered out is rigid motion. In this protein structure problem, one of the main aims is to take a query active site and find matches to a given database, in some ranking order. The matches will give some idea of functions of the unknown proteins, leading to the design of new enzymes for example. There are other related examples from Image Analysis such as matching buildings when one has multiple 2-dimensional views of 3-dimensional objects (see, for example, Cross and Hancock, 1998). The problem here requires filtering out the projective transformations before matching. Other examples involve matching outlines or surfaces (see, for example, Chui and Rangarajan, 2000, and Pedersen, 2002). Here there is no labelling of points involved, and we are dealing with a continuous contour or surface rather than a finite number of points. Such problems are not addressed in this paper. In Section 2 we build a hierarchical Bayesian model for the point configurations and derive inferential procedure for its parameters. In particular, modelling hidden point locations as a Poisson process leads to a considerable simplification. We discuss in particular the problem when only a linear or affine transformation has to be filtered out. In Section 3 we discuss prior specifications, and provide an implementation of the resulting methodology by means of Markov chain Monte Carlo (MCMC) samplers. Under a broad parametric family of loss functions, an optimal Bayesian point estimate of the matching matrix can be constructed, which turns out to depend on a single parameter of the family. We also discuss a modification to the likelihood in our model to make use of partial label (`colour') information at the points. Finally here there is a note on the possibilities for an alternative computational approach using the EM algorithm. Section 4 describes application of our methods to the two examples from Bioinformatics mentioned above: matching Protein gels in 2 dimensions and aligning active sites of Proteins in 3 dimensions. The paper concludes with a Discussion of some open problems and future directions, and comparisons with other methods. The principal innovations in our approach are (a) the fully model-based approach to alignment, (b) the model formulation allowing integrating out of the hidden point locations, (c) the prior specification for the rotation matrix, and (d) the MCMC algorithm. \section{Hierarchical modelling of alignment and matching problems} \label{sec:models} We will build a hierarchical model for the observed point configurations, and derive inferential procedures for its parameters, including the unknown matching between the configurations, according to the Bayesian paradigm. \subsection{Point process model, with geometrical transformation and random thinning} Suppose we are given two point configurations in $d$-dimensional space $\mathcal{R}^d$: $\{x_j, j=1,2,\ldots,m\}$ and $\{y_k, k=1,2,\ldots,n\}$. The points are labelled for identification, but arbitrarily. Both point sets are regarded as noisy observations on subsets of a set of true locations $\{\mu_i\}$, where we do not know the mappings from $j$ and $k$ to $i$. There may be a geometrical transformation between the $x$-space and the $y$-space, which may also be unknown. The objective is to make model-based inference about these mappings, and in particular make probability statements about matching -- which pairs $(j,k)$ correspond to the same true location? The geometrical transformation between the $x$-space and the $y$-space will be denoted $\mathcal{A}$; thus $y$ in $y$-space corresponds to $x=\mathcal{A} y$ in $x$-space. The notation does not imply that the transformation $\mathcal{A}$ is necessarily linear. It may be a rotation or more general linear transformation, a translation, both of these, or some non-rigid motion. We regard the true locations $\{\mu_i\}$ as being in $x$-space. The mappings between the indexing of the $\{\mu_i\}$ and that of the data $\{x_j\}$ and $\{y_k\}$ are captured by indexing arrays $\{\xi_j\}$ and $\{\eta_k\}$; specifically we assume that \bel{likx} x_j=\mu_{\xi_j}+\varepsilon_{1j} \end{equation} for $j=1,2,\ldots,m$, where $\{\varepsilon_{1j}\}$ have probability density $f_1$, and \bel{liky} \mathcal{A} y_k=\mu_{\eta_k}+\varepsilon_{2k} \end{equation} for $k=1,2,\ldots,n$, where $\{\varepsilon_{2k}\}$ have density $f_2$. Multiple matches are excluded, thus each hidden point $\mu_i$ is observed at most once in each of the $x$ and $y$ configurations; equivalently, the $\xi_j$ are distinct, as are the $\eta_k$. All $\{\varepsilon_{1j}\}$ and $\{\varepsilon_{2k}\}$ are independent of each other, and independent of the $\{\mu_i\}$. \subsection{Formulation of Poisson process prior} \label{sec:pp} Suppose that the set of true locations $\{\mu_i\}$ forms a homogeneous Poisson process with rate $\lambda$ over a region $V\subset\mathcal{R}^d$ of volume $v$, and that there are $N$ points realised in this region. Some of these give rise to both $x$ and $y$ points, some to points of one kind and not the other, and some are not observed at all. We suppose these four possibilities occur independently for each realised point, with probabilities parameterised so that with probabilities $(1-p_{\rm x}-p_{\rm y}-\rhop_{\rm x}p_{\rm y},p_{\rm x}, p_{\rm y},\rhop_{\rm x}p_{\rm y})$ we observe neither, $x$ alone, $y$ alone, or both $x$ and $y$, respectively. The parameter $\rho$ is a certain measure of the tendency {\it a priori} for points to be matched: the random thinnings leading to the observed $x$ and $y$ configurations can be dependent, but remain independent from point to point. Given $N$, $m$ and $n$, there are $L$ matched pairs of points in our sample if and only if the numbers of these four kinds of occurrence among the $N$ points are $(N-m-n+L,m-L,n-L,L)$. Under the assumptions above these four counts will be independent Poisson distributed variables, with means $(\lambda v (1-p_{\rm x}-p_{\rm y}-\rhop_{\rm x}p_{\rm y}), \lambda v p_{\rm x},\lambda v p_{\rm y},\lambda v \rhop_{\rm x}p_{\rm y})$. The prior probability distribution of $L$ conditional on $m$ and $n$ is therefore proportional to $$ \frac{e^{-\lambda v p_{\rm x}}(\lambda v p_{\rm x})^{m-L}}{(m-L)!}\times \frac{e^{-\lambda v p_{\rm y}}(\lambda v p_{\rm y})^{n-L}}{(n-L)!}\times \frac{e^{-\lambda v \rhop_{\rm x}p_{\rm y}}(\lambda v \rhop_{\rm x}p_{\rm y})^L}{L!} $$ so that \bel{lprior} p(L) \propto \frac{(\rho/\lambda v)^L}{(m-L)!(n-L)!L!} \end{equation} for $L=0,1,\ldots,\min\{m,n\}$. The normalising constant here is the reciprocal of $H(m,n,\rho/(\lambda v))$, where $H$ can be written in terms of the confluent hypergeometric function $$ H(m,n,d) = \frac{d^m}{m!(n-m)!} \: \mbox{}_1\!F_1(-m,n-m+1,-1/d), $$ assuming without loss of generality that $n>m$; see Abramowitz and Stegun (1970, p. 504). Here and later, we use the generic $p(\cdot)$ notation for distributions and conditional distributions in our hierarchical model. The matching of the configurations is represented by the {\it matching matrix} $M$, where $M_{jk}$ indicates whether $x_j$ and $y_k$ are derived from the same $\mu_i$ point, or not, that is, $$ M_{jk} =\cases {1 & if $\xi_j=\eta_k$ \cr 0 & otherwise \cr}. $$ Note that $\sum_{j,k} M_{jk}=L$, and that, since multiple matches are ruled out, there is at most one 1 in each row and in each column of $M$: $\sum_j M_{jk}\leq 1 \forall k$, $\sum_k M_{jk}\leq 1 \forall j$. We assume for the moment that conditional on $L$, $M$ is {\it a priori} uniform: there are $L! {m \choose L} {n \choose L}$ different $M$ matrices consistent with a given value of $L$, and these are taken as equally likely. Thus $$ p(M) = p(L)p(M|L) \propto \frac{(\rho/\lambda v)^L}{(m-L)!(n-L)!L!} \left\{L! {m \choose L} {n \choose L}\right\}^{-1} \propto (\rho/\lambda v)^L, $$ (where here and later `$\propto$' means proportional to, as functions of the variable(s) to the left of the conditioning $|$, in this case, $M$). Thus \bel{pm} p(M) = \frac{(\rho/\lambda v)^L} {\sum_{\ell=0}^{\min\{m,n\}} \ell! {m \choose \ell} {n \choose \ell}(\rho/\lambda v)^\ell}. \end{equation} Note that, because of the choice of parameterisation for the probabilities that hidden points are observed, this expression does not involve $p_{\rm x}$ and $p_{\rm y}$. \begin{figure}[htbp] \centering \resizebox{3.5in}{!}{\rotatebox{0}{\includegraphics{align.eps}}} \caption{Directed acyclic graph representing our model, showing all data and parameters treated as variable. \label{fig:align}} \end{figure} \subsection{Likelihood of data} We now have to specify the likelihood of the observed configurations of points, given $M$. For simplicity, we will henceforth assume that $\mathcal{A}$ is an affine transformation: $\mathcal{A} y=Ay+\tau$. From (\ref{likx}) and (\ref{liky}), the densities of $x_j$ and $y_k$, conditional on $A$, $\tau$, $\{\mu_i\}$, $\{\xi_j\}$ and $\{\eta_k\}$ are $f_1(x_j-\mu_{\xi_j})$ and $|A|f_2(Ay_k+\tau-\mu_{\eta_k})$, respectively, $|A|$ denoting the absolute value of the determinant of $A$. The locations $\{\mu_i\}$ of the $m-L$ points that generate an $x$ observation but not a $y$ observation are independently uniformly distributed over the region $V$, so that the likelihood contribution of these $m-L$ observations, namely $\{x_j:M_{jk}=0\forall k\}$, is $$ \prod_{j: M_{jk}=0\forall k} v^{-1} \int_V f_1(x_j-\mu) d\mu $$ Similarly, the contributions from the unmatched $y$ observations, and from the matched pairs are $$ \prod_{k: M_{jk}=0\forall j} v^{-1} \int_V |A|f_2(Ay_k+\tau-\mu) d\mu \quad\mbox{and}\quad \prod_{j,k: M_{jk}=1} v^{-1} \int_V f_1(x_j-\mu) |A|f_2(Ay_k+\tau-\mu) d\mu $$ respectively. These integrals all exhibit `edge effects' from the boundary of the region $V$, which can be neglected if $V$ is large relative to the supports of $f_1$ and $f_2$. In this case these three expressions approximate to $$ v^{-(m-L)}, (|A|/v)^{n-L}, \quad\mbox{and}\quad (|A|/v)^L \prod_{j,k: M_{jk}=1}\int_{\mathcal{R}^d} f_1(x_j-\mu) f_2(Ay_k+\tau-\mu) d\mu $$ respectively. The last expression can be written $$ (|A|/v)^L \prod_{j,k: M_{jk}=1} g(x_j-Ay_k-\tau) $$ where $g(z)=\int f_1(z+u)f_2(u)du$ (the density of $\varepsilon_{1j}-\varepsilon_{2k}$). Combining these terms, the complete likelihood is \bel{lik} p(x,y|M,\mathcal{A}) = v^{-(m+n)} |A|^n \prod_{j,k: M_{jk}=1} g(x_j-Ay_k-\tau). \end{equation} Multiplying (\ref{pm}) and (\ref{lik}), we then have $$ p(M,x,y|\mathcal{A}) \propto |A|^n \prod_{j,k: M_{jk}=1} \{(\rho/\lambda)g(x_j-Ay_k-\tau)\}. $$ Note that the constant of proportionality involves $m$, $n$, $\lambda$, $\rho$, and $v$, but not $A$, $\tau$, any parameters in $f_1$ or $f_2$, or $M$ of course. If we further specialise by making assumptions of spherical normality for $f_1$ and $f_2$: $$ x_j \sim N_d(\mu_{\xi_j},\sigma_{\rm x}^2I) \qquad\mbox{and}\qquad Ay_k+\tau \sim N_d(\mu_{\eta_k},\sigma_{\rm y}^2I), $$ with $\sigma_{\rm x}=\sigma_{\rm y}=\sigma$, say, then $$ g(z)=\frac{1}{(\sigma\surd 2)^d} \phi(z/\sigma\surd 2) $$ where $\phi$ is the standard normal density in $\mathcal{R}^d$, and our final joint model is \bel{post} p(M,A,\tau,\sigma,x,y) \propto |A|^{n} p(A) p(\tau) p(\sigma)\prod_{j,k:M_{jk}=1} \left( \frac{\rho\phi(\{x_j-Ay_k-\tau\}/\sigma\surd 2)}{\lambda(\sigma\surd 2)^d}\right). \end{equation} Note that not only $p_{\rm x}$ and $p_{\rm y}$ but also $v$ does not appear in this expression, principally from our choice of parameterisation, and that only the ratio $\rho/\lambda$ is identifiable. The directed acyclic graph representing this joint probability model, including the variables ($\mu$, $\xi$ and $\eta$) that we have integrated out, is displayed in Figure \ref{fig:align}. \section{Prior distributions and computational implementation} \label{sec:mcmc} We will henceforth treat $\rho$ and $\lambda$ as fixed, and consider inference for the remaining unknowns $M$, $\tau$, $\sigma^2$ and sometimes $A$, given the data $\{x_j\}$ and $\{y_k\}$. Markov chain Monte Carlo methods must be used for the computation; several introductions and overviews of MCMC are available, for example, the primer in Green (2001). In Section \ref{sec:em}, we discuss the relevance and applicability of an EM algorithm for making inference with an approximation of our model. We suppose that prior information about $\tau$, $\sigma^2$ and $A$ will be at best weak, and so we concentrate on generic prior formulations that facilitate the posterior analysis. Prior assumptions are therefore discussed in parallel with MCMC implementation. Note that our formulation has some affinity with mixture models, the matching matrix $M$ playing a similar role to the allocation variables often used in computing with mixtures; see, for example, Richardson and Green (1997). As in that paper, the fully Bayesian analysis here aims at simultaneous joint inference about both the discrete and continuously varying unknowns, in contrast to frequentist approaches. Our model has another similarity with a mixture formulation, in that as $M$ varies, the number of hidden points needed to generate all the observed data also varies, and thus there seems to be a `variable-dimension' aspect to the model. However, here our approach of integrating out the hidden point locations eliminates the variable-dimension parameter, so that reversible jump MCMC is not needed. \subsection{Priors and MCMC updating for a rotation matrix} \label{sec:rotmat} We are interested in alignment and matching problems in which either $A$ is given, and treated as fixed, or in which it is one of the objects of inference. In the latter case, we consider in this paper only the case of rotation matrices in two and three dimensions. We therefore focus on the full conditional distribution for $A$, which from (\ref{post}) is $$ p(A|M,\tau,\sigma,x,y) \propto |A|^{n} p(A) \prod_{j,k:M_{jk}=1} \phi(\{x_j-Ay_k-\tau\}/\sigma\surd 2). $$ Viewing this as a density for $A$, we are still free to choose the dominating measure for $p(A)$, which is arbitrary: this full conditional density is then with respect to the same measure. Let us restrict attention to {\it rotations}: orthogonal matrices $A$, (those with $A^{-1}$ = $A^T$) with positive determinant, so that $|A|=1$. Expanding the expression above, we then find $$ p(A|M,\tau,\sigma,x,y) \propto p(A) \exp\left(\sum_{j,k:M_{jk}=1} -0.5(||x_j-Ay_k-\tau||/\sigma\surd 2)^2 \right) $$ $$ \propto p(A) \exp \left(\mbox{tr}\left\{ (1/2\sigma^2)\sum_{j,k:M_{jk}=1} y_k(x_j-\tau)^TA\right\} \right). $$ Note a remarkable opportunity for (conditional) conjugacy -- if $p(A)$ has the form $p(A)\propto \exp(\mbox{tr}(F_0^TA))$ for some matrix $F_0$, then the posterior has the same form with $F_0$ replaced by $$ F=F_0+(1/2\sigma^2)\sum_{j,k:M_{jk}=1} (x_j-\tau)y_k^T. $$ This form of $p(A)$ is known as the matrix Fisher distribution (Downs, 1972; Mardia and Jupp, 2000, p. 289). To the best of our knowledge, this unique role of the matrix Fisher distribution (or in the two-dimensional case, the von Mises distribution) as the prior distribution for a rotation conjugate to spherical Gaussian error distributions has not previously been noted. (Although Mardia and El-Atoum (1976) have identified the von Mises--Fisher distribution as the conjugate prior for the mean direction). This may have relevance in models for other situations, including the simpler case where there is no uncertainty in the matching. The conjugacy is presumably related to the interpretation of the matrix Fisher distribution as a conditional multivariate Gaussian (see Mardia and Jupp, 2000, p.289). \subsubsection*{Two-dimensional case} Now consider the two-dimensional case, $d=2$. An arbitrary rotation matrix $A$ can be written $$ A=\left( \begin{array}{rr} \cos \theta & -\sin \theta \\ \sin \theta& \cos \theta \\ \end{array} \right) $$ and the natural dominating measure for $\theta$ is Lebesgue on $(0,2\pi)$. Then a uniformly distributed choice of $A$ corresponds to $p(A)\propto 1$. More generally, the von Mises distribution for $\theta$ $$ p(\theta) \propto \exp(\kappa\cos(\theta-\nu))=\exp(\kappa\cos\nu\cos\theta+\kappa\sin\nu\sin\theta) $$ can indeed be expressed as $p(A)\propto \exp(\mbox{tr}(F_0^TA))$, where a (non-unique) choice for $F_0$ is $$ F_0=\kappa/2\left( \begin{array}{rr} \cos \nu & -\sin \nu \\ \sin \nu& \cos \nu \\ \end{array} \right). $$ Thus the full conditional distribution for $\theta$ is of the same von Mises form, with $\kappa\cos\nu$ updated to $(\kappa\cos\nu+S_{11}+S_{22})$, and $\kappa\sin\nu$ to $(\kappa\sin\nu-S_{12}+S_{21})$, where $S$ is the $2\times 2$ matrix $(1/2\sigma^2)\sum_{j,k:M_{jk}=1} (x_j-\tau)y_k^T$. It is therefore trivial to implement a Gibbs sampler move to allow inference about $A$, assuming a von Mises prior distribution on the rotation angle $\theta$ (including the uniform case, $\kappa=0$). We can use the Best/Fisher algorithm, an efficient rejection method (see Mardia and Jupp, 2000, p.43), to sample from the full conditional for $\theta$. \subsubsection*{Three-dimensional case} In the three-dimensional case, we can represent $A$ as the product of elementary rotations \bel{geneul} A=A_{12}(\theta_{12})A_{13}(\theta_{13})A_{23}(\theta_{23}) \end{equation} as in Raffenetti and Ruedenberg (1970), and Khatri and Mardia (1977). Here, for $i<j$, $A_{ij}(\theta_{ij})$ is the matrix with $m_{ii}=m_{jj}=\cos\theta_{ij}$, $-m_{ij}=m_{ji}=\sin\theta_{ij}$, $m_{rr}=1$ for $r\neq i,j$ and other entries 0. We can then update each of the generalised Euler angles $\theta_{ij}$ in turn, conditioning on the other two angles and the other variables ($M,\tau,\sigma,x,y$) entering the expression for $F$. The joint full conditional density of the Euler angles is $$ \propto \exp[\mbox{tr}\{F^TA\}] \cos\theta_{13} $$ for $\theta_{12},\theta_{23}\in(-\pi,\pi)$ and $\theta_{13}\in(-\pi/2,\pi/2)$. The cosine term arises since the natural dominating measure, corresponding to uniform distribution of rotation, has volume element $\cos\theta_{13} \:\d\theta_{12}\:\d\theta_{13}\:\d\theta_{23}$ in these coordinates. Substituting the representation (\ref{geneul}), and simplifying, we find that the trace can be written variously as $\mbox{tr}\{F^TA\}=a_{12}\cos\theta_{12}+b_{12}\sin\theta_{12}+c_{12} =a_{13}\cos\theta_{13}+b_{13}\sin\theta_{13}+c_{13} =a_{23}\cos\theta_{23}+b_{23}\sin\theta_{23}+c_{23}$ where \begin{eqnarray*} a_{12}& = &(F_{22}-\sin\theta_{13}F_{13})\cos\theta_{23} +(-F_{23}-\sin\theta_{13}F_{12})\sin\theta_{23} +\cos\theta_{13}F_{11} \\ b_{12}& = &(-\sin\theta_{13}F_{23}-F_{12})\cos\theta_{23} +(F_{13}-\sin\theta_{13}F_{22})\sin\theta_{23} +\cos\theta_{13}F_{21} \\ a_{13}& = &\sin\theta_{12}F_{21}+\cos\theta_{12}F_{11}+\sin\theta_{23}F_{{32}}+\cos\theta_{23}F_{33} \\ b_{13}& = & (-\sin\theta_{23}F_{12}-\cos\theta_{23}F_{13})\cos\theta_{12} +(-\sin\theta_{23}F_{22}-\cos\theta_{23}F_{23})\sin\theta_{12}+F_{31} \\ a_{23}& = &(F_{22}-\sin\theta_{13}F_{13})\cos\theta_{12}+(-\sin\theta_{13}F_{23}-F_{12})\sin\theta_{12} +\cos\theta_{13}F_{33} \\ b_{23}& = &(-F_{23}-\sin\theta_{13}F_{12})\cos\theta_{12}+(F_{13}-\sin\theta_{13}F_{22})\sin\theta_{12} +\cos\theta_{13}F_{32} \end{eqnarray*} and the $c_{ij}$ can be ignored, combined into the normalising constants. Thus the full conditionals for $\theta_{12}$ and $\theta_{23}$ are von Mises distributions, and so these two variables can be updated by Gibbs sampling. That of $\theta_{13}$ is proportional to $$ \exp[a_{13}\cos\theta_{13}+b_{13}\sin\theta_{13}]\cos\theta_{13} $$ and we use a random walk Metropolis update for this variable, with a perturbation uniformly distributed on $[-0.1,0.1]$. The latter distribution has been studied in Mardia and Gadsden (1977) but with no discussion on how to simulate from it. \subsection{Priors and updating for other parameters} \label{sec:prior} We make the standard normal/inverse gamma assumptions: $$ \tau \sim N_d(\mu_\tau,\sigma_\tau^2I) \qquad\mbox{and}\qquad \sigma^{-2} \sim \Gamma(\alpha,\beta). $$ Under the assumptions of (\ref{post}), there is conjugacy for $\tau$ and $\sigma$, and we have explicit full conditionals: $$ \tau|M,A,\sigma,x,y \sim N_d\left( \frac{\mu_\tau/\sigma_\tau^2+\sum_{j,k:M_{jk}=1} (x_j-Ay_k)/2\sigma^2} {1/\sigma_\tau^2+L/2\sigma^2} , \frac{1}{1/\sigma_\tau^2+L/2\sigma^2}I \right) $$ $$ \sigma^{-2}|M, A,\tau,x,y \sim \Gamma\left(\alpha+(d/2)L, \beta+(1/4)\sum_{j,k:M_{jk}=1} ||x_j-Ay_k-\tau||^2\right), $$ and so it is trivial to implement Gibbs sampler updates for these parameters. \subsection{Updating $M$} \label{sec:updatem} The matching matrix $M$ is updated in detailed balance using Metropolis-Hastings moves that only propose changes to a few entries: the number of matches $L=\sum_{j,k}M_{jk}$ can only increase or decrease by 1 at a time, or stay the same. The possible changes are \begin{enumerate} \item[(a)] adding a match: changing one entry $M_{jk}$ from 0 to 1 \item[(b)] deleting a match: changing one entry $M_{jk}$ from 1 to 0 \item[(c)] switching a match: simultaneously changing one entry from 0 to 1, and another {\it in the same row or column} from 1 to 0. \end{enumerate} The proposal proceeds as follows: first a uniform random choice is made from all the $m+n$ data points $x_1,x_2,\ldots,x_m,y_1,y_2,\ldots,y_n$. Suppose without loss of generality, by the symmetry of the set-up, that an $x$ is chosen, say $x_j$. There are two possibilities: either $x_j$ is currently matched ($\exists k$ such that $M_{jk}=1$) or not (there is no such $k$). If $x_j$ is matched to $y_k$, with probability $p^\star$ we propose {\it deleting} the match, and with probability $1-p^\star$ we propose {\it switching} it from $y_k$ to $y_{k'}$, where $k'$ is drawn uniformly at random from the currently unmatched $y$ points. On the other hand, if $x_j$ is not currently matched, we propose {\it adding} a match between $x_j$ and a $y_{k}$, where again $k$ is drawn uniformly at random from the currently unmatched $y$ points. The acceptance probabilities for these three possibilities are easily derived from the expression (\ref{post}) for the joint distribution, since in each case the proposed new matching matrix $M'$ is only slightly perturbed from $M$, so that the ratio $p(M',\tau,\sigma|x,y)/p(M,\tau,\sigma|x,y)$ has only a few factors. Taking into account also the proposal probabilities, whose ratio is $(1/n_{\rm u})\div p^\star$, where $n_{\rm u}=\#\{k\in 1,2,\ldots,n: M_{jk}=0\forall j\}$ is the number of unmatched $y$ points in $M$, we find that the acceptance probability for adding a match $(j,k)$ is \bel{mhadd} \min\left\{1, \frac {\rho\phi(\{x_j-Ay_k-\tau\}/\sigma\surd{2})p^\star n_{\rm u}} {\lambda(\sigma\surd{2})^d} \right\}. \end{equation} Similarly, the acceptance probability for switching the match of $x_j$ from $y_k$ to $y_{k'}$ is \bel{mhswitch} \min\left\{1, \frac{\phi(\{x_j-Ay_{k'}-\tau\}/\sigma\surd{2})} {\phi(\{x_j-Ay_k-\tau\}/\sigma\surd{2})}\right\} \end{equation} and for deleting the match $(j,k)$ it is $$ \min\left\{1, \frac{\lambda(\sigma\surd{2})^d} {\rho\phi(\{x_j-Ay_k-\tau\}/\sigma\surd{2})p^\star n_{\rm u}'}\right\}, $$ where $n_{\rm u}'=\#\{k\in 1,2,\ldots,n: M_{jk}'=0\forall j\}=n_{\rm u}+1$. Along with just one of each of the other updates, we typically make several moves updating $M$ per sweep, since the changes effected are so modest. \subsection{Loss functions} The output from the MCMC sampler derived above, once equilibrated, is a sample from the posterior distribution determined by (\ref{post}). As always with sample-based computation, this provides an extremely flexible basis for reporting aspects of the full joint posterior that are of interest. The matching matrix $M$ will often be of particular inferential interest, and for some purposes a point estimate is desirable; in this section we discuss how to obtain a Bayesian point estimate of the matching matrix $M$. The most easily understood estimator of $M$ would be its posterior mode, the {\it maximum a posteriori} (MAP) estimator. However, there are difficulties here. First, the notion is itself ambiguous -- the unknown `parameter' in our model consists of the matching matrix $M$, and some real parameters. `MAP' might refer to the $M$ component of the overall maximum, or the mode of the marginal posterior for $M$ alone. Secondly, the posterior is multi-modal, and different modes may have different `widths', appropriately measured. So there is no intrinsic attraction to the MAP estimate. We should return to basic principles. By standard theory, this requires specification of a loss function, $L(M,\widehat{M})$, giving the cost incurred in declaring the matching matrix to be $\widehat{M}$ when it is in fact $M$. The optimal estimate given data $(x,y)$ is the matching matrix $\widehat{M}$ that minimises the posterior expected loss $$ E[L(M,\widehat{M})|x,y], $$ the expectation over $M$ being taken with respect to the posterior determined by (\ref{post}). In this language, the MAP estimator is optimal for the `zero--one' loss function under which a fixed total cost is paid if there is a single error in any value $M_{jk}$; this is logically unappealing, and a further argument against using MAP. We consider instead loss functions $L(M,\widehat{M})$ that penalise different kinds of error and do so cumulatively. The simplest of these are additive over pairs $(j,k)$. Suppose that the loss when $M_{jk}=a$ and $\widehat{M}_{jk}=b$, for $a,b=0,1$ is $\ell_{ab}$; for example, $\ell_{01}$ is the loss associated with declaring a match between $x_j$ and $y_k$ when there is really none, that is, a `false positive'. Then it is readily shown that $$ E[L(M,\widehat{M})|x,y] = -(\ell_{10}+\ell_{01}-\ell_{11}-\ell_{00}) \sum_{j,k:\widehat{M}_{jk}=1} (p_{jk}-K) $$ where $$ K=(\ell_{01}-\ell_{00})/ (\ell_{10}+\ell_{01}-\ell_{11}-\ell_{00}), $$ and $p_{jk}=p(M_{jk}=1|x,y)$ is the posterior probability that $(j,k)$ is a match, which is estimated from an MCMC run by the empirical frequency of this match. Thus, provided that $\ell_{10}+\ell_{01}-\ell_{11}-\ell_{00}>0$ and $\ell_{01}-\ell_{00}>0$, as is natural, the optimal estimate is that maximising the sum of marginal posterior probabilities of the declared matches $\sum_{j,k:\widehat{M}_{jk}=1} p_{jk}$, penalised by a multiple $K$ times the number of matches. The optimal match therefore depends on the four loss function parameters only through the cost ratio $K$. If false positive and false negative matches are equally undesirable, one can simply choose $K=0.5$. Computation of the optimal match $\widehat{M}$ would be trivial but for the constraint that there can be at most one positive entry in each row and column of the array. For modest-sized problems, the optimal match can be found by informal heuristic methods. These may not even be necessary, especially if $K$ is not too small. In particular, it is immediate that if the set of all $(j,k)$ pairs for which $p_{jk}>K$ includes no duplicated $j$ or $k$ values, the optimal $\widehat{M}$ consists of precisely these pairs. We could also consider loss functions that penalise mismatches differently from the sum of the losses of the individual errors. For example, declaring $(j,k)$ to be a match when it should be $(j,k')$ might deserve a relative loss greater or lesser than $(\ell_{10}+\ell_{01}-\ell_{11}-\ell_{00})$, depending on context. Such loss functions could be handled in a broadly similar way, but this is left for future work. \subsection{Using partial labelling information} \label{modlik} When the points in each configuration are `coloured', with the interpretation that like-coloured points are more likely to be matched than unlike-coloured ones, it is appropriate to use a modified likelihood that allows us to exploit such information. Let the colours for the $x$ and $y$ points be $\{r^{\rm x}_j,j=1,2,\ldots,m\}$ and $\{r^{\rm y}_k,k=1,2,\ldots,n\}$ respectively. The hidden point model is augmented to generate the point colours, as follows. Independently for each hidden point, with probability $(1-p_{\rm x}-p_{\rm y}-\rhop_{\rm x}p_{\rm y})$ we observe neither $x$ nor $y$ point, as before. With probabilities $p_{\rm x}\pi^{\rm x}_r$ and $p_{\rm y}\pi^{\rm y}_r$, respectively, we observe only an $x$ or $y$ point, with colour $r$ from an appropriate finite set. With probability $$ \rhop_{\rm x}p_{\rm y}\pi^{\rm x}_r\pi^{\rm y}_s \exp\{\gamma I[r=s]+ \delta I[r \neq s]\}, $$ we observe an $x$ point coloured $r$ and a $y$ point coloured $s$. Our original likelihood is equivalent to the case $\gamma=\delta=0$, where colours are independent and so carry no information about matching. If $\gamma$ and $\delta$ increase, then matches are more probable, {\it a posteriori}, and if $\gamma>\delta$, matches between like-coloured points are more likely than those between unlike-coloured ones. The case $\delta\to-\infty$ allows the prohibition of matches between unlike-coloured points, a feature that might be adapted to other contexts such as the matching of shapes with given landmarks. In implementation of this modified likelihood, the MCMC acceptance ratios in Section \ref{sec:updatem} have to be modified accordingly. For example, if $r^{\rm x}_j=r^{\rm y}_k$ and $r^{\rm x}_j\neq r^{\rm y}_{k'}$, then (\ref{mhadd}) has to be multiplied by $\exp(-\gamma)$ and (\ref{mhswitch}) by $\exp(\delta-\gamma)$. Other, more complicated, colouring distributions where the log probability can be expressed linearly in entries of $M$ can be handled similarly. \subsection{Alternative approach using the EM algorithm} \label{sec:em} The interplay between matching (allocation) and parameter uncertainty has something in common with mixture estimation. This might suggest considering maximisation of the posterior by using the EM algorithm, which could of course in principle be applied either to maximum likelihood estimation based on (\ref{lik}) or to MAP estimation based on (\ref{post}). For the EM formulation, the `missing data' are the matches. In an exponential family, the EM algorithm alternates between between finding expectations of missing values given data, at current parameter values, and maximising the log-posterior, with missing values replaced by these expectations. The `expectations of missing values' are just probabilities of matching. These are only tractable if we were to drop the assumption that a point can only be matched with at most one other point -- that is, that $\sum_j M_{jk}\leq 1 \forall k$, $\sum_k M_{jk} \leq 1 \forall j$. Making this approximation, the E-step is trivial: the expectation of $I[M_{jk}=1]$ is $p_{jk}=w_{jk}/(1+w_{jk})$ where $w_{jk}$ is the $(j,k)$ factor in the joint model, i.e. $$ w_{jk}=\{(\rho/\lambda)g_\sigma(x_j-Ay_k-\tau)\} $$ The M-step then requires maximising (for given $p_{jk}$) $$ \log\left[|A|^{n} p(A) p(\tau) p(\sigma)\right] +\sum_{j,k} p_{jk} \log \{w_{jk}(A,\tau,\sigma)\} $$ over $A$, $\tau$, $\sigma$ -- note that here $w_{jk}$ is a function of all three. Although for some individual parameters this seems to be explicit, in the general case we need numerical optimisation. In summary, EM allows us to study only certain aspects of an approximate version of our model, and is not trivial numerically -- so we do not pursue this approach. Obtaining the complete posterior by MCMC sampling gives much greater freedom in inference. \section{Applications} \subsection{Matching protein gels} The objective in this example is to match two electrophoretic gels automatically, given the locations of the centres of 35 proteins on each of the two gels. The data are presented in the supplementary information on the web. The correspondence between pairs of proteins, one protein from each gel, is unknown, so our aim is to match the two gels based on these sets of unlabelled points. We suppose it is known that the transformation between the gels is affine. In this case, experts have already identified 10 points; see Horgan et al (1992). Based on these 10 matches, the linear part of the transformation is estimated {\it a priori} to be \bel{gelA} A=\left( \begin{array}{rr} 0.9731 & 0.0394 \\ -0.0231 & 0.9040 \\ \end{array} \right). \end{equation} (Dryden and Mardia, 1998, pp. 20--21, 292--296). \begin{table} \caption{The 20 marginally most probable matches in the analysis of the gel data. \label{gelmatches}} \footnotesize \vspace*{5mm}\centering\leavevmode \begin{tabular}{crrl} rank & $j$ & $k$ & $p_{jk}$ \\ \hline 1 & 15 & 21 & 1 \\ 2 & 19 & 19 & 1 \\ 3 & 8 & 8 & 1 \\ 4 & 3 & 3 & 1 \\ 5 & 2 & 2 & 1 \\ 6 & 31 & 30 & 0.9989 \\ 7 & 6 & 6 & 0.9987 \\ 8 & 4 & 4 & 0.9966 \\ 9 & 5 & 5 & 0.9946 \\ 10 & 10 & 10 & 0.9927 \\ 11 & 24 & 23 & 0.9855 \\ 12 & 7 & 7 & 0.9824 \\ 13 & 32 & 31 & 0.9776 \\ 14 & 1 & 1 & 0.9763 \\ 15 & 9 & 9 & 0.9677 \\ 16 & 26 & 32 & 0.7910 \\ 17 & 12 & 13 & 0.7552 \\ 18 & 21 & 33 & 0.3998 \\ 19 & 26 & 27 & 0.1931 \\ 20 & 35 & 35 & 0.0025 \\ \hline \end{tabular} \end{table} \myfig{gelconf}{The 17 most probable matches in the gel data, the optimal match for any $K\in(0.3998,0.7552)$; + symbols signify $x$ points, o symbols the $y$ points, linearly transformed by premultiplication by the fixed affine transformation $A$ given in (\ref{gelA}). The solid line for each of the 17 matches joins the matched points, and represents the inferred translation $\tau$ plus noise.}{5} Here, we have only to make inference on the translation $\tau$ and the unknown matching between certain of the proteins. The model (\ref{post}) will therefore be taken to apply, with $d=2$ and with $A$ held fixed at (\ref{gelA}). The MCMC sampler described in Section \ref{sec:mcmc} was run for 100 000 sweeps, after a burn-in period of 20 000 sweeps, considered on the basis of an informal visual assessment of time series traces to be adequate for convergence. Prior and hyperprior settings were: $\alpha=1$, $\beta=16$, $\mu_\tau=(0,0)^T$, $\sigma_\tau=20.0$ and $\lambda/\rho=0.0001$. The sampler parameter $p^\star$ was set to 0.5. Such a run took about 2 seconds on a 800MHz PC. Acceptance rates for the moves updating $M$ were between 0.6\% and 2.1\%. The posterior expectation and variance of $\tau$ were estimated to be $(-35.950,66.685)^T$ (to be compared with $(-36.08,66.64)^T$ obtained by Dryden and Mardia (1998)) and $$ \left( \begin{array}{rr} 0.5776 & -0.0227 \\ -0.0227 & 0.6345 \\ \end{array} \right). $$ The posterior mean and variance of $\sigma$ are 2.050 and 0.1192. The 20 most probable matches between $x$ and $y$ points are listed in Table \ref{gelmatches}; note that there is no duplication in their indices until the 19th match: $j=26$ also appears in the 16th match (recall that there is a simple rule for identifying the optimal $\widehat{M}$ if there are no duplicates among the matches with $p_{jk}$ above the threshold $K$). We can conclude that for all values of $K=(\ell_{01}-\ell_{00})/ (\ell_{10}+\ell_{01}-\ell_{11}-\ell_{00})$ from 1 down to 0.1112, the optimal Bayesian matching is given by an appropriate subset of Table \ref{gelmatches}, reading down from the top. For example if this cost ratio is 0.8 we take the first 15 rows of the table, while if the ratio is 0.6 or 0.4 we include the 16th and 17th rows as well. The 17 most probable matches are displayed graphically in Figure \ref{fig:gelconf}. It will be noted that all of the expert-identified matches, points 1 to 10 in each set, are declared to be matches with high probability in the Bayesian analysis. We also repeated the analysis with these 10 pairs held fixed. The next 9 most probable matches, together with these 10, are identical to those in the first 19 lines of Table \ref{gelmatches}, and the posterior probabilities differ by less than 0.037 in all 19 cases. \subsection{Aligning proteins in three dimensions} \label{sec:3deg} We now apply the matching method to a problem in three dimensional structural biology, previously considered by Gold et al (2002). The problem consists of finding the matches for two Active sites 1 and 2 corresponding to two Proteins A and B respectively. The corresponding coordinates $x$ and $y$ of these sites are presented in the supplementary information; these coordinates are the centres of gravity of the amino acids of the two sites. Here $m= 40$ and $n=63$. The biological details of the two proteins are as follows. Protein 1 is the human protein `17--beta hydroxysteroid dehydrogenase' and is involved in the synthesis of oestrogens. This protein binds the ligands (molecules comparatively smaller than proteins) oestradiol and NADP. Protein 2 is the mouse protein `carbonyl reductase' and is involved in metabolism of carbonyl compounds. This protein binds the ligands 2--Propanol and NADP. The common element between these two sets of ligands is NADP. From chemical properties of the sites, the relevant matching should be invariant under rigid transformation. \label{sec:3dalign} \myfig{nicolaconf42}{The optimal alignment (36 matches) when $K=0.5$ for the protein alignment analysis data, without using colouring information; + symbols signify $x$ points, o symbols the $y$ points, rotated according to the inferred $\widehat{A}$ matrix given by (\ref{meana1}). The entire joint configuration has been rotated to its first two principal axes. Solid lines represent the 36 marginally most probable matches, and indicate the inferred translation $\tau$ plus noise.}{5} \myfig{nicplall}{Time series traces and histograms of the MCMC run of Section \ref{sec:3deg}, based on a thinned sub-sample of 2000 after burn-in.}{5} There is information about the identities of the amino acids in the two configurations: we defer use of this to Section \ref{sec:amino}. The MCMC sampler described in Section \ref{sec:mcmc} was run for 1 000 000 sweeps, after a burn-in period of 200 000 sweeps, considered on the basis of an informal visual assessment of time series traces to be adequate for convergence. Prior and hyperprior settings were: $\alpha=1$, $\beta=36$, $\mu_\tau=(0,0,0)^T$, $\sigma_\tau=50.0$, $\lambda/\rho=0.003$ and the matrix $F_0$ defining the prior for $A$ set to the zero matrix. The sampler parameter $p^\star$ was set to 0.5, and we made updates to $M$ 10 times in each sweep. Such a run took about 71 seconds on a 800MHz PC. Acceptance rates for the moves updating $M$ were between 0.41\% and 5.6\%. The posterior expectation and variance of $\tau$ were estimated to be $(31.60,8.89,17.44)^T$ and $$ \left( \begin{array}{rrr} 0.227 & 0.120 & -0.044 \\ 0.120 & 0.307 & 0.176 \\ -0.044 & 0.176 & 0.428 \\ \end{array} \right) $$ The posterior mean and variance of $\sigma$ are 1.051 and 0.00996. In representing the centre of the posterior distribution for the rotation matrix $A$, we we need to use a definition of mean appropriate to the geometry. We form the mean elementwise from a thinned sample of 2000 values of $A$ from the post-burn-in MCMC run. This mean matrix $\overline{A}$ is of course not a rotation matrix, but post-multiplication by the positive definite symmetric square root of $\overline{A}^T\overline{A}$ yields a rotation matrix that is known as its polar part (see Mardia and Jupp, p. 286, 290). This is an appropriate measure of location of the posterior, and takes the value \bel{meana1} \widehat{A}=\left( \begin{array}{rrr} 0.4339 & -0.8444 & 0.3140 \\ -0.7118 & -0.5350 & -0.4550 \\ 0.5522 & -0.0261 & -0.8333 \\ \end{array} \right) \end{equation} in this case. The 40 most probable matches between $x$ and $y$ points are listed in supplementary information; there is no duplication in their indices until the 39th match: $k=12$ also appears in the 38th match. We can conclude that for all values of $K$ greater than 0.2895 (the marginal posterior probability associated with the 39th match), the optimal Bayesian matching is given by an appropriate leading subset of the matches. For example if this cost ratio is 0.5 we take the first 36 matches; these are displayed graphically in Figure \ref{fig:nicolaconf42}; in this 3-dimensional example, the axes signify the first two principle coordinates of the combined cloud of data. As would be anticipated, simultaneous inference for the rotation $A$ and the matching matrix $M$ (as well as $\tau$ and $\sigma$) is a considerably greater challenge for MCMC than is the problem of the previous section, where the rotation matrix is held fixed. It is clear that there is a possibility of severe multi-modality in the posterior, with the conditional posterior for $M$ and $\tau$ given $A$ depending strongly on $A$. This challenge is quantified empirically by a heavy-tailed distribution of times to convergence, and by `meta-stability' in the time series plots of various monitoring statistics against simulation time. We found the log-posterior to be a useful summary statistic for quality of fit, and pilot runs provided experience to choose a threshold value, exceedance of which we hypothesised diagnosed convergence to the main mode of the posterior. To investigate multimodality and convergence time, we conducted a study in which the MCMC run described was repeated -- with the same parameters -- from 100 different initial configurations, obtained by independent random rotations as initial settings for $A$. After short runs of 50 000 sweeps, we tested whether the threshold log-posterior value had been exceeded, and if not the run was abandoned. 83 out of the 100 runs passed this test, and these were allowed to run on for a further 450 000 sweeps. Every one of these 83 long runs provided exactly the same set of 36 most probable matches, and we therefore felt justified to conclude that they had not been trapped in a subsidiary mode of the posterior, and that it was safe to draw inference from the results. This conclusion is specific to the data set and parameter settings used, and it would be straightforward to contrive artificial data where multiple modes were more equal in probability content. In such cases more sophisticated MCMC samplers would be needed. \subsection{Prior settings and sensitivity} \label{sec:sensitivity} Our analysis depends of course on the settings of the hyperparameters $\lambda/\rho$ (see Section \ref{sec:pp}), $F_0$ (Section \ref{sec:rotmat}), and $\mu_\tau$, $\sigma_\tau$, $\alpha$, $\beta$ (Section \ref{sec:prior}). These allow the provision of real prior information from the experimental context, if it is available. For a default analysis in the absence of such information, we would set $F_0$ to the zero matrix (a uniform prior on $A$), $\mu_\tau$ to be the zero vector, and $\sigma_\tau$ of the order of twice the distance between the centres of gravity of the two configurations. We fix $\alpha=1$, giving an exponential prior distribution for $\sigma^{-2}$. Here we briefly discuss settings of, and sensitivity to, the remaining two parameters, the scalars $\lambda/\rho$ and $\beta$. Sensitivity to $\lambda/\rho$ is pronounced, as might be anticipated. This parameter ratio has a very direct role in determining whether an $(x_j,y_k)$ pair are noisy observations of the same hidden $\mu_i$ point or not, after transformation, since it controls the density of hidden points. In practice, we should not expect to be able to draw inference about matching without real prior knowledge about this ratio or an equivalent measure of the prior tendency of points to be matched. The prior for the number of matches $L$ is parameterised by $\lambda/\rho$: see (\ref{lprior}). This distribution is non-standard, but very well-behaved. It is clear from inspection that setting $\lambda/\rho$ equal to $(m-\overline{L})(n-\overline{L})/\overline{L} v$ yields a mode of $L$ that is within 1 of $\overline{L}$, and numerical calculation in the context of the example in Section \ref{sec:3deg}, verifies that for all possible `prior guesses' $\overline{L}$ for $L$, the prior expectation and median are also both equal to $\overline{L}$ to the nearest integer. Thus prior information about $L$ is directly informative about the parameter ratio $\lambda/\rho$. As long as $v$ is known, or at least a representative value provided, and the analyst is able to make a prior guess $\overline{L}$ at the number of matches, this suggests a reasonable way to specify $\lambda/\rho$. The posterior distribution for $L$ tracks the prior rather closely, confirming that the raw data carry little information about the number of matches. The hyperparameter $\beta$ is an inverse scale parameter for the precision of the noise terms $\varepsilon$; thus as $\beta$ increases, we expect that $\sigma^2=\mbox{var}(\varepsilon)$ increases too. The runs we have presented used $\beta=36$; reducing this by a factor of 2 makes minimal difference to the posterior inference for either $\sigma^2$ or $M$. However, increasing $\beta$ by a factor of 2 leads to a 3-fold increase in $\sigma$ and a sharp reduction in the number of matches -- the posterior expectation of $L$ goes down from around 34 to 26. The latter observation is perhaps counter-intuitive, until one realises that when $\sigma$ is larger, it becomes relatively less likely that points that are nearly coincident (after transformation) are in fact matched. Finally, it would be desirable to assess the sensitivity to the Poisson assumption for the hidden point model, but this would be extremely onerous to do directly, since alternatives would require a substantially modified formulation and implementation. There is scientific reason to doubt the Poisson assumption; for example, the minimum spacing between the centres of gravity of the amino acids in proteins is approximately 3.8 Angstroms. However, experiments reported in Mardia, Nyirongo and Westhead (2005) do at least suggest strongly that the ability of our method to detect matches is little affected by real hard-core effects. \subsection{Using information about types of amino acid} \label{sec:amino} The protein alignment data includes identifiers of the type of amino acid at each point (see supplementary information). There are 20 different types, which can be categorised into 4 groups: hydrophobic, charged, polar and glycine; we use the group identifiers as colours in defining a modified likelihood as in Section \ref{modlik}. The parameter values taken were $\gamma=1.0$ and $\delta=-0.5$, providing a strong preference for like-coloured matching ($\exp(\gamma-\delta)\approx 4.48$). The analysis was repeated with this modified model, leaving all other details unchanged. The 40 most marginally probable matches are listed in supplementary information, along with displayes of the optimal alignment. The 36 most probable matches, which together form the optimal matching whtn $K=0.4$, are identical to those found in the previous section; however, there are modest variations in the posterior probabilities attached to individual matches. The posterior expectation and variance of $\tau$ were now estimated to be $(31.94,8.94,17.61)^T$ (slightly shifted from that obtained in the analysis of the previous section) and $$ \left( \begin{array}{rrr} 1.284 & -0.763 & -0.118 \\ -0.763 & 3.534 & -0.015 \\ -0.118 & -0.015 & 1.320 \\ \end{array} \right) $$ The posterior mean and variance of $\sigma$ are 1.3122 and 0.1984. The increased estimate of $\sigma$ is perhaps anticipated. The centre of the posterior distribution of $A$ is in this case: \bel{meana2} \widehat{A}=\left( \begin{array}{rrr} 0.4240 & -0.8512 & 0.3092 \\ -0.7235 & -0.5237 & -0.4497 \\ 0.5447 & -0.0331 & -0.8379 \\ \end{array} \right). \end{equation} \myfig{comb43}{The optimal matching (36 matches), when $K=0.4$, in the protein alignment analysis data, using colouring information, with $\gamma=1.0$ and $\delta=-0.5$; matches are signified by line segments joining the sequence number of the point in the $x$ configuration to that of the matched point in the $y$ configuration. The solid lines indicate the 27 matches identified by Gold et al (2002); our method discovers all of these, together with the 9 further matches shown with broken lines. The height of the vertical bars indicate the marginal probabilities of each match. The + symbols denote points that are present in either configuration but are not matched.}{6.5} In the approach to the analysis of these data taken by Gold et al (2002), the matching between the configurations was performed in two stages, and is not driven by an explicit probability model. First, inter-point distances $d(\cdot,\cdot)$ were calculated within each configuration. These distances are invariant under the rigid body motions considered here. A maximal set of pairs of indices $\{(j_1,k_1),(j_2,k_2),\ldots\}$, with no ties among the $j$s or $k$s, is found such that $|d(x_{j_r},x_{j_s})-d(y_{k_r},y_{k_s})|$ is less than some threshold, for all $s\neq r$. This is done using graph theoretical algorithms of Bron and Kerbosch (1973) and Carraghan and Pardos (1990), applied to a product graph whose vertices are labelled with $(j,k)$ pairs. This first stage of the matching alogrithm was formulated by Kuhl et al (1984). In the second stage, the matches are scored using the amino acid information, assigning a score of 1 for identity of the amino acids, and 0.5 when the amino acids are different but fall in the same group. The initial list of matches from stage one is then permuted so as to maximise the total score. Once the matches are found the rigid body transformation is estimated by Procrustes analysis; for example, see Dryden and Mardia (1998, pp 176-178). It is interesting to compare the rotation matrix resulting from this method, namely $$ A=\left(\begin{array}{rrr} 0.441 & -0.841 & 0.312 \\ -0.678 & -0.541 & -0.498 \\ 0.588 & 0.008 & -0.809 \\ \end{array} \right) $$ with that obtained by our method. The trace of the orthogonal matrix taking $A$ to $\widehat{A}$ is approximately $1+2\cos 0.07$, so the two differ by a rotation of only 0.07 radians. Figure \ref{fig:comb43} provides a comparison between the matchings achieved by the two approaches. Of the 27 matches identified by Gold et al, 14 are among the most probable 20 that we find, and all 27 are among the first 35. A referee has raised with us the role of sequence ordering along the protein in inference about alignment and matching. The example in this section concerns ligand binding site matching, in which biologically relevant matches do not necessarily preserve sequential ordering, in contrast to the more familiar situation of aligning protein backbones; see for example Eidhammer et al (2004, pp. 333--334). Examples are trypsin-subtilisin with similar active sites and unrelated folds, and many adenine binding sites in different folds. Somewhat remarkably, although sequence ordering is not used in our analysis, the resulting matches do perfectly respect this ordering. This is visualised in Figure \ref{fig:comb43}, which also reveals that some but not all of the matches revealed by our analysis additional to those of Gold et al (2002) extend already matched segments. In this particular data set, the sites must come from very closely related folds and would probably also be alignable by sequence-preserving methods aligning full structures. Intriguingly, in this example at least, knowledge of the sequence ordering would provide no additional information beyond that extracted from the point coordinates and amino-acid groups by our approach. \section{Discussion} The main conclusion of this paper is that a probability model based approach is successful in allowing simultaneous inference about partial matching between two point configurations, and a geometrical transformation between the coordinate systems in which the configurations are measured. This seems an advance over previous more ad-hoc methods. We have only used the translation and rigid motion groups in illustrating our methodology. However, the formulation allows inference about various other group transformations such as affine transformation, and so on. The fairly straightforward MCMC implementation presented here has proved adequate for the models and data sets considered, although allowing rotations did increase the needed run lengths considerably. We anticipate that, at least for models allowing rotations, dealing with larger data sets will be much more challenging, since small rotational perturbations generate large displacements at sites far from the axis of rotation; moves that simultaneously perturb allocations and geometrical and error distribution parameters will be necessary for good performance. We also anticipate more severe difficulties from multi-modality that were exposed in Section \ref{sec:3deg}. An important task left for future work is a formulation that allows smooth nonparametric transformations between coordinate systems, setting warping into a model-based framework; this would be important in dealing more comprehensively with gel matching problems. We have only used pairwise comparisons but there is scope for taking multiple combinations such as triads. The transformations considered above are parametric but some non-parametric alternatives such as non-linear deformations may be useful in some cases, e.g. to deal with dynamic aspects of the atoms in a protein. We have considered only two configurations but a natural extension would be to take three or more point configurations simultaneously, and make joint inference about patterns of matching between the configurations and the various geometrical transformations involved. More straightforward extensions would be to allow for non-Gaussian noise, other types of prior and so on. Kent et al (2004) have treated the unlabelled case by using a different model. While matching two configurations, one of them is taken as the population and the second as a random sample from this population after an unknown transformation. This approach is different from the symmetrical model for the two configurations proposed here. Further the emphasis in Kent et al (2004) is on maximum likelihood inference using the EM-algorithm. Recent independent work by Dryden, Hirst and Melville (2005), addresses a similar problem of matching unlabelled point sets. Their approach has some substantial differences, for example there is assymmetry in comparing two configurations, one being treated as a perturbation of the other. The geometrical transformation parameters are given uniform priors and maximised out, using standard ideas from shape analysis, rather than integrated out as in our fully Bayesian approach. Neither the loss function basis for estimating matches, nor the treatment of partial labelling, appear. There is other statistical work on alignment and matching in proteins by Wu et al (1998) and Schmidler (2004), which in contrast does use sequence information. Further work is needed to clarify the relationships between all these methods and their comparative performance. Finally, in the context of using methods such as ours in database search, often the reason for assessing protein alignment, there are issues related to multiple comparisons. These are not discussed here, but the answers will depend on the size of the database as well as the number of points in the query site. \section*{Acknowledgements} We are grateful to Nicola Gold and Dave Westhead for their many helpful discussions, and in particular for the data in Example 2, and to Vysaul Nyirongo and Charles Taylor for various helpful comments. \section*{References} \begin{list}{}{\setlength{\itemindent}{-\leftmargin}} \item Abramowitz, M. and Stegun, I. A. (1970). {Handbook of Mathematical Functions}. Dover, New York. \item Bron, C. and Kerbosch, J. (1973). Algorithm 457; finding all cliques of an undirected graph. {\it Communication of the ACM}, {\bf 16}, 575--577. \item Carraghan, R. and Pardalos, P. M. (1990). Exact algorithm for the minimal clique problem. {\it Operations Research Letters}, {\bf 9}, 375. \item Chui, H. and Rangarajan, A. (2000). A new algorithm for non-rigid point matching. {\em IEEE Conference on Computer Vision and Pattern Recognition.} {\bf 2}, 44--51. \item Cross, A. D. J. and Hancock, E. R. (1998). Graph matching with dual-step {EM} algorithm. {\em IEEE transactions on pattern analysis and machine intelligence.} {\bf 20}, 1236--1253. \item Downs, T. D. (1972). Orientation statistics. {\it Biometrika}, {\bf 59}, 665--676. \item Dryden, I. L., Hirst, J. D. and Melville, J. L. (2005). Statistical analysis of unlabelled point sets: comparing molecules in chemoinformatics. Under revision for {\it Biometrics}. \item Dryden, I. L. and Mardia, K. V. (1998). {\em Statistical shape analysis}. Wiley, Chichester. \item Eidhammer, T., Jonassen, T. and Taylor, W. R. (2004). {\it Protein Bioinformatics}. Wiley, Chichester. \item Gold, N. D., Pickering, S. J., and Westhead, D. R. (2002). Protein functional site matching using graph theory techniques. In {\it Proceedings of the International Conference on Bioinformatics}, Bangkok, Thailand, page 79. \item Green, P. J. (2001). A Primer on Markov chain Monte Carlo, pp. 1--62 of {\it Complex Stochastic Systems}, Barndorff-Nielsen, O. E., Cox, D. R. and Kl\"{u}ppelberg, C. (eds.), Chapman and Hall, London. \item Horgan, G. W., Creasey, A. and Fenton, B. (1992). Superimposing two dimensional gels to study genetic variation in malaria parasites. {\it Electrophoresis}, {\bf 13},871--875. \item Kent, J. T., Mardia, K. V. and Taylor, C. C. (2004). Matching problems for unlabelled configurations. In {\it Bioinformatics, Images and Wavelets}, edited by Aykroyd, R.G., Barber, S. and Mardia, K.V. Proceedings of LASR 2004, Leeds University Press, Leeds. \item Khatri, C. G. and Mardia, K. V. (1977). The von Mises--Fisher distribution in orientation statistics. {\it Journal of the Royal Statistical Society}, B, {\bf 39}, 95--106. \item Kuhl, F. S., Crippen, G. M. and Friesen, D. K. (1984). A combinatorial algorithm for calculating ligand binding. {\it Journal of Computational Chemistry}, {\bf 5}, 24--34. \item Mardia, K. V. and El-Atoum, S. A. M. (1976). Bayesian inference for the von Mises--Fisher distribution. {\it Biometrika}, {\bf 63}, 203--205. \item Mardia, K. V. and Gadsden, R. J. (1977). A circle of best fit for spherical data and areas of vulcanism. {\it Applied Statistics},{ \bf 26}, 238--245. \item Mardia, K. V. and Jupp, P. E. (2000). {\it Directional Statistics}, Wiley, Chichester. \item Mardia K. V., Taylor, C. C, and Westhead, D. R. (2003). Structural bioinformatics revisited. In {\it LASR2003}, pp11--18. Leeds University Press. \item Mardia, K. V., Nyirongo, V., and Westhead, D.R. (2005). EM algorithm, Bayesian and distance approaches to matching active sites {\it Mathematical and Statistical Annual Meeting in Bioinformatics}, Rothamsted, March 2005, Abstracts pp13-14. \item Pedersen, L. (2002). {\em Analysis of two-dimensional electrophoresis gel images.} Ph.D thesis, IMM Technical University of Denmark. \item Raffenetti, R. C. and Ruedenberg, K. (1970). Parameterization of an orthogonal matrix in terms of generalized Eulerian angles. {\it International Journal of Quantum Chemistry}, {\bf IIIS}, 625--634. \item Richardson, S. and Green, P. J. (1997). On Bayesian analysis of mixtures with an unknown number of components (with discussion). {\it Journal of the Royal Statistical Society}, B, {\bf 59}, 731--792. \item Schmidler, S. C. (2004). {\it Bayesian shape matching and structural alignment}. Presentation at the 6th World Congress of the Bernoulli Society, Barcelona, July 2004. \item Wu, T. D., Schmidler, S. C., Hastie, T. and Brutlag, G. (1998). Regression analysis of multiple protein structures. {\it Journal of Computational Biology}, {\bf 5}, pp 585--595. \end{list} \end{document}
{ "timestamp": "2005-07-01T16:37:22", "yymm": "0503", "arxiv_id": "math/0503712", "language": "en", "url": "https://arxiv.org/abs/math/0503712" }
\section{Introduction} Random matrices play an important role in physics and mathematics \cite{Mehta, courseynard, BI, DGZ, Guhr, Moerbeke:2000, DeiftBook}. It has been observed more and more in the recent years how deeply random matrices are related to integrability ($\tau$-functions), and algebraic geometry. Here, we consider the computation of large n asymptotics for orhogonal polynomials as an example of a problem where the concepts of integrability, isomonodromy and algebraic geometry appear and combine. The method presented here below, is not, to that date, rigorous mathematicaly. It is based on the asumption that an integral with a large number of variables can be approximated by a saddle-point method. This asumption was never proven rigorously, it is mostly based on ``physical intuition''. However, the results given by that method have been rigorously proven by another method, namely the Riemann--Hilbert method \cite{BlIt, BlIt1, dkmvz, dkmvz2}. The method presented below was presented in many works \cite{eynchain, eynchaint, BEHAMS, eynbetapol, eynhabilit}. \section{Definitions} Here we consider the 1-Hermitean matrix model with polynomial potential: \begin{eqnarray}\label{defZ} Z_N&:=& \int_{H_N} dM\, {\mathbf e}^{-N{\rm tr}\, V(M)}\cr &=& \int_{{\mathbb{R}}^N} dx_1\dots dx_N\,\, \left(\Delta(x_1,\dots,x_N)\right)^2 \,\, \prod_{i=1}^N {\mathbf e}^{-NV(x_i)} \end{eqnarray} where $\Delta(x_1,\dots,x_N):=\prod_{i>j} (x_i-x_j)$, and the $x_i$'s are the eigenvalues of the matrix $M$, and $V(x)$ is a polynomial called the potential: \begin{equation}\label{defV} V(x) = \sum_{k=0}^{\deg V} g_k x^k \end{equation} \begin{remark} All the calculations which are presented below, can be extended to a more general setting, with no big fundamental changes: - one can consider $V'(x)$ any rational fraction \cite{BEHsemiclas} instead of polynomial, in particular one can add logarithmic terms to the potential $V(x)$. - one can consider arbitrary paths (or homology class of paths) of integrations $\Gamma^N$ insteaf of ${\mathbb{R}}^N$, in particular finite segments \cite{marcopath} ... - one can study non hermitean matrix models \cite{eynbetapol}, where the Vandermonde $\Delta^2$ is replaced by $\Delta^\beta$ where $\beta=1,2,4$. - one can consider multi-matrix models, in particular 2-matrix model \cite{BEHAMS, eynchain, eynchaint}. \end{remark} \section{Orthogonal polynomials} Consider the family of monic polynomials $p_n(x)=x^n + O(x^{n-1})$, defined by the orthogonality relation: \begin{equation} \int_{{\mathbb{R}}} p_n(x) p_m(x) {\mathbf e}^{-NV(x)} dx = h_n \delta_{nm} \end{equation} It is well known that the partition function is given by \cite{Mehta}: \begin{equation} Z_N = N! \prod_{n=0}^{N-1} h_n \end{equation} Such an orthogonal family always exists if the integration path is ${\mathbb{R}}$ or a subset of ${\mathbb{R}}$, and if the potential is a real polynomial. In the more general setting, the orthogonal polynomials ``nearly always'' exist (for arbitrary potentials, the set of paths for which they don't exist is enumerable). \medskip We define the kernel: \begin{equation} K(x,y):=\sum_{n=0}^{N-1} {p_n(x) p_n(y)\over h_n} \end{equation} One has the following usefull theorems: \begin{theorem} Dyson's theorem \cite{thDyson}: any correlation function of eigenvalues, can be written in terms of the kernel $K$: \begin{equation} \rho(\l_1,\dots, \l_k) = \det(K(\l_i,\l_j)) \end{equation} \end{theorem} Thus, if one knows the orthogonal polynomials, then one knows all the correlation functions. \begin{theorem} Christoffel-Darboux theorem \cite{Mehta, Szego}: The kernel $K(x,y)$ can be written: \begin{equation} K(x,y) = \gamma_N\,{p_N(x)p_{N-1}(y)-p_N(y)p_{N-1}(x)\over x-y} \end{equation} \end{theorem} Thus, if one knows the polynomials $p_N$ and $p_{N-1}$, then one knows all the correlation functions. \medskip Our goal now, is to find large $N$ ''strong'' asymptotics for $p_N$ and $p_{N-1}$, in order to have the large $N$ behaviours of any correlation functions. \medskip {\bf Notation:} we define the wave functions: \begin{equation} \psi_n(x) := {1\over \sqrt{h_n}}\, p_n(x)\, {\mathbf e}^{-{N\over 2}V(x)} \end{equation} they are orthonormal: \begin{equation} \int \psi_n(x)\psi_m(x) = \delta_{nm} \end{equation} \section{Differential equations and integrability} It can be proven that $(\psi_n,\psi_{n-1})$ obey a differential equation of the form \cite{bonan, BlIt, Mehta, TW2, BEHtauiso}: \begin{equation} -{1\over N} \,{\partial \over \partial x} \pmatrix{\psi_n(x) \cr \psi_{n-1}(x)} = {{\cal D}}_n(x)\,\pmatrix{\psi_n(x) \cr \psi_{n-1}(x)} \end{equation} where ${\cal D}_n(x)$ is a $2\times 2$ matrix, whose coefficients are polynomial in $x$, of degree at most $\deg V'$. (In case $V'$ is a rational function, then ${\cal D}$ is a rational function with the same poles). \medskip $(\psi_n,\psi_{n-1})$ also obeys differential equations with respect to the parameters of the model \cite{BlIt, BEHtauiso}, i.e. the coupling constants, i.e. the $g_k$'s defined in \ref{defV}: \begin{equation} {1\over N} \,{\partial \over \partial g_k} \pmatrix{\psi_n(x) \cr \psi_{n-1}(x)} ={\cal U}_{n,k}(x)\,\pmatrix{\psi_n(x) \cr \psi_{n-1}(x)} \end{equation} where ${\cal U}_{n,k}(x)$ is a $2\times 2$ matrix, whose coefficients are polynomial in $x$, of degree at most $k$. \medskip It is also possible to find some discrete recursion relation in $n$ (see \cite{BEHtauiso}). \medskip The compatibility of these differential systems, i.e. ${\partial\over \partial x}{\partial\over \partial g_k}={\partial\over \partial g_k}{\partial\over \partial x}$, ${\partial\over \partial g_j}{\partial\over \partial g_k}={\partial\over \partial g_k}{\partial\over \partial g_j}$, as well as compatibility with the discrete recursion, imply {\bf integrability}, and allows to define a $\tau$-function \cite{MiwaJimbo, BEHtauiso}. \bigskip We define the spectral curve as the locus of eigenvalues of ${\cal D}_n(x)$: \begin{equation} E_n(x,y):=\det(y{\bf 1} - {\cal D}_n(x)) \end{equation} \begin{remark}\rm\small In the 1-hermitean-matrix model, ${\cal D}_n$ is a $2\times 2$ matrix, and thus $\deg_y E_n(x,y)=2$, i.e. the curve $E_n(x,y)=0$ is an {\bf hyperelliptical curve}. In other matrix models, one gets algebraic curves which are not hyperelliptical. \end{remark} \begin{remark}\rm\small What we will se below, is that the curve $E_N(x,y)$ has a large $N$ limit $E(x,y)$, which is also an hyperelliptical curve. In general, the matrix ${\cal D}_N(x)$ has no large $N$ limit. \end{remark} \section{Riemann-Hilbert problems and isomonodromies} The $2\times2$ system ${\cal D}_N$ has $2$ independent solutions: \begin{equation} -{1\over N} \,{\partial \over \partial x} \pmatrix{\psi_n(x) \cr \psi_{n-1}(x)} = {{\cal D}}_n(x)\,\pmatrix{\psi_n(x) \cr \psi_{n-1}(x)} \quad , \quad -{1\over N} \,{\partial \over \partial x} \pmatrix{\phi_n(x) \cr \phi_{n-1}(x)} = {{\cal D}}_n(x)\,\pmatrix{\phi_n(x) \cr \phi_{n-1}(x)} \end{equation} where the wronskian is non-vanishing: $\det\pmatrix{\psi_n(x) & \phi_n(x) \cr \psi_{n-1}(x) & \phi_{n-1}(x)}\neq 0$. We define the matrix of fundamental solutions: \begin{equation} \Psi_n(x):=\pmatrix{\psi_n(x) & \phi_n(x) \cr \psi_{n-1}(x) & \phi_{n-1}(x)} \end{equation} it obeys the same differential equation: \begin{equation} -{1\over N} \,{\partial \over \partial x} \Psi_n(x) = {{\cal D}}_n(x)\,\Psi_n(x) \end{equation} \medskip Here, the second solution can be constructed explicitely: \begin{equation} \phi_n(x) = {\mathbf e}^{+{N\over 2}V(x)}\,\int {dx'\over x-x'}\,\psi_n(x') {\mathbf e}^{-{N\over 2}V(x')} \end{equation} Notice that $\phi_n(x)$ is discontinuous along the integration path of $x'$ (i.e. the real axis in the most simple case), the discontinuity is simply $2i\pi \psi_n(x)$. In terms of fundamental solutions, one has the jump relation: \begin{equation}\label{JumpRH} \Psi_n(x+i0) = \Psi_n(x-i0)\,\pmatrix{1 & 2i\pi \cr 0 & 1} \end{equation} Finding an invertible piecewise analytical matrix, with given large $x$ behaviours, with given jumps on the borders between analytical domains, is called a {\bf Riemann--Hilbert problem} \cite{BlIt, BlIt1, BEHRH}. It is known that the Riemann--Hilbert problem has a unique solution, and that if two R-H problems differ by $\epsilon$ (i.e. the difference between jumps and behaviours at $\infty$ is bounded by $\epsilon$), then the two solutions differ by at most $\epsilon$ (roughly speeking, harmonic functions have their extremum on the boundaries). Thus, this approach can be used \cite{BlIt, dkmvz, dkmvz2} in order to find large $N$ asymptotics of orthogonal polynomials: The authors of \cite{BlIt} considered a guess for the asymptotics, which satisfies another R-H problem, which differs from this one by $O(1/N)$. \bigskip Notice that the jump matrix in \ref{JumpRH} is independent of $x$, of $n$ and of the potential, it is a constant. The jump matrix is also called a monodromy, and the fact that the monodromy is a constant, is called {\bf isomonodromy} property \cite{MiwaJimbo}. Consider an invertible, piecewise analytical matrix $\Psi_n(x)$, with appropriate behaviours\footnote{The behaviours at $\infty$ are far beyond the scope of this short lecture. They are easily obtained by computing $\phi_n(x)$ by saddle point method at large $x$.} at $\infty$, which satisfies \ref{JumpRH}, then, it is clear that the matrix $-{1\over N} \Psi_n'(x) (\Psi_n(x))^{-1}$, has no discontinuity, and given its behaviour at $\infty$, it must be a polynomial. Thus, we can prove that $\Psi_n(x)$ must satisfy a differential system ${\cal D}_n(x)$ with polynomial coefficients. Similarly, the fact that the monodromy is independent of $g_k$ and $n$ implies the deformation equations, as well as the discrete recursion relations. Thus, the isomonodromy property, implies the existence of compatible differential systems, and integrability \cite{BI, FIK, stringIts, MiwaJimbo, TW2, BEHtauiso}. \section{WKB--like asymptotics and spectral curve} \label{secasympWKBformal} Let us look for a formal solution of the form: \begin{equation}\label{asympWKBformal} \Psi_N(x) = A_N(x) \, {\mathbf e}^{-N T(x)} B_N \end{equation} where $T(x)={\rm diag}(T_1(x),T_2(x))$ is a diagonal matrix, and $B_N$ is independent of $x$. The differential system ${\cal D}_N(x)$ is such that: \begin{eqnarray} {\cal D}_N(x) &=& -{1\over N}\Psi_N'\Psi_N^{-1} = A_N(x) T'(x) A_N^{-1}(x) - {1\over N} A'_N(x) A_N^{-1}(x) \cr & =& A_N(x) T'(x) A_N^{-1}(x) + O({1\over N}) \end{eqnarray} this means, that, under the asumption that $A_N(x)$ has a large $N$ limit $A(x)$, $T'_1(x)$ and $T'_2(x)$ are the large $N$ limits of the eigenvalues of ${\cal D}_N(x)$. With such an hypothesis, one gets for the orthogonal polynomials: \begin{equation} \psi_N(x) \sim A_{11}{\mathbf e}^{-NT_1(x)} B_{1,1} + A_{12}{\mathbf e}^{-NT_2(x)} B_{2,1} \end{equation} We are now going to show how to derive such a formula. \section{Orthogonal polynomials as matrix integrals} \subsection{Heine's formula} \begin{theorem} Heine's theorem \cite{Szego}. The orthogonal polynomials $p-n(x)$ are given by: \begin{eqnarray} p_n(\xi) &=& {\int dx_1\dots dx_N\,\, \prod_{i=1}^N (\xi-x_i)\,\, (\Delta(x_1,\dots,x_N))^2\,\, \prod_{i=1}^N {\mathbf e}^{-NV(x_i)}\over \int dx_1\dots dx_N\,\, (\Delta(x_1,\dots,x_N))^2\,\, \prod_{i=1}^N {\mathbf e}^{-NV(x_i)}} \cr &=& \left<\det(\xi{\bf 1} - M)\right> \end{eqnarray} \end{theorem} i.e. the orthogonal polynomial is the average of the characteristic polynomial of the random matrix. Thus, we can define the orthogonal polynomials as matrix integrals, similar to the partition function $Z$ define in \ref{defZ}. \subsection{Another matrix model} Define the potential: \begin{equation} V_h(x):=V(x)-h\ln{(\xi-x)} \end{equation} and the partition function: \begin{equation} Z_n(h,T):={\mathbf e}^{-{n^2\over T^2}F_n(h,T)} :=\int dx_1\dots dx_n\,\, (\Delta(x_1,\dots,x_n))^2\,\, \prod_{i=1}^n {\mathbf e}^{-{n\over T}V_h(x_i)} \end{equation} i.e. $Z_N(0,1)=Z$ is our initial partition function. Heine's formula reads: \begin{equation} p_n(\xi) = {Z_n({1\over N},{n\over N})\over Z_n(0,{n\over N})} = {\mathbf e}^{-N^2(F_n({1\over N},{n\over N})-F_n(0,{n\over N}))} \end{equation} The idea, is to perform a Taylor expansion in $h$ close to $0$ and $T$ close to $1$. \subsubsection{Taylor expansion} We are interested in $n=N$ and $n=N-1$, thus $T={n\over N}=1+{n-N\over N}=1+O(1/N)$ and $h=0$ or $h=1/N$, i.e. $h=O(1/N)$: \begin{equation} T=1+O(1/N) \quad , \quad h=O(1/N) \end{equation} Roughly speaking: \begin{eqnarray}\label{asymp1} p_n(\xi) &\sim& {\mathbf e}^{-N^2\left( h {\partial F\over \partial h}+(T-1)h{\partial^2 F\over \partial h\partial T}+{h^2\over 2}{\partial^2 F\over \partial h^2}+O(1/N^3)\right)} \cr &\sim& {\mathbf e}^{-N {\partial F\over \partial h}}\,{\mathbf e}^{-(n-N){\partial^2 F\over \partial h\partial T}}\,{\mathbf e}^{-{1\over 2}{\partial^2 F\over \partial h^2}}\,\,(1+O(1/N)) \end{eqnarray} where all the derivatives are computed at $T=1$ and $h=0$. \subsubsection{Topological expansion} Imagine that $F_n$ has a $1/n^2$ expansion of the form: \begin{equation} F = F^{(0)} + {1\over n^2} F^{(1)} + O({1\over n^3}) \end{equation} where all $F^{(0)}$ and $F^{(1)}$ are analytical functions of $T$ and $h$, than one needs only $F^{(0)}$ in order to compute the asymptotics \ref{asymp1}. \smallskip Actualy, that hypothesis is not always true. It is wrong in the so called ''mutlicut'' case. But it can be adapted in that case, we will come back to it in section \ref{sectmulticutasymp}. For the moment, let us conduct the calculation only with $F^{(0)}$. \section{Computation of derivatives of $F^{(0)}$} We have defined: \begin{equation} Z_n(h,T)={\mathbf e}^{-{n^2\over T^2}F_n(h,T)} = \int dM_{n\times n} {\mathbf e}^{-{n\over T}{\rm tr}\, V(M)}\, {\mathbf e}^{h{n\over T}\ln{(\xi-M)}} \end{equation} this implies that: \begin{equation} -{n^2\over T^2} {\partial F_n\over \partial h} = \left<{n\over T}{\rm tr}\, \ln{(\xi-M)}\right>_{V_h} \end{equation} i.e. \begin{eqnarray} {\partial F_n\over \partial h} &=& -{T\over n}\left<{\rm tr}\, \ln{(\xi-M)}\right>_{V_h} \cr \end{eqnarray} It is a primitive of $-{T\over n}\left<{\rm tr}\, \ln{(x-M)}\right>_{V_h}$, which behaves as $-{T\over n}\ln{x}+O(1/x)$ at large $x$. Therefore, we define the resolvent $W(x)$: \begin{equation}\label{defWVh} W(x):={T\over n}\left<{\rm tr}\, {1\over x-M}\right>_{V_h} \end{equation} Notice that it depends on $\xi$ through the potential $V_h$, i.e. through the average $<.>$. And we define the effective potential: \begin{equation} {V_{\rm eff}}(x)= V_h(x)-2T\ln{x}-2\int_{\infty}^x (W(x')-{T\over x'}) dx' \end{equation} which is a primitive of $V'_h(x)-2W(x)$. Thus , we have: \begin{equation} {\partial F_n\over \partial h} = {1\over 2}\left({V_{\rm eff}}(\xi)-V_h(\xi)\right) \end{equation} We also introduce: \begin{equation}\label{defOm} \Omega(x):={\partial W(x)\over \partial T} \quad , \quad \ln{\Lambda(x)}:=\ln{x}+\int_{\infty}^x(\Omega(x')-{1\over x'})dx' = -{1\over 2}{\partial \over \partial T}V_{\rm eff}(x) \end{equation} \begin{equation}\label{defH} H(x,\xi):={\partial W(x)\over \partial h} \quad , \quad \ln{H(\xi)}:=\int_{\infty}^\xi H(x',\xi)dx' \end{equation} i.e. \begin{equation} {\partial^2 F_n\over \partial h^2} = -\ln{H(\xi)} \quad , \quad {\partial^2 F_n\over \partial h\partial T} = -\ln{\Lambda(\xi)} \end{equation} With these notations, the asymptotics are: \begin{equation} \psi_n(\xi) \sim \sqrt{H(\xi)}\,\,\left(\Lambda(\xi)\right)^{n-N}\,\,{\mathbf e}^{-{N\over 2}{V_{\rm eff}}(\xi)}\,\,(1+O(1/N)) \end{equation} Now, we are going to compute $W$, $\Lambda$, $H$, etc, in terms of geometric properties of an hyperelliptical curve. \begin{remark}\rm\small This is so far only a sketch of the derivation, valid only in the 1-cut case. In general, $F_n$ has no $1/n^2$ expansion, and that case will be addressed in section \ref{sectmulticutasymp}. \end{remark} \begin{remark}\rm\small These asymptoics are of the form of \ref{asympWKBformal} in section.\ref{secasympWKBformal}, and thus, ${1\over 2}V'(x)-W(x)$ is the limit of the eigenvalues of ${\cal D}_N(x)$. \end{remark} \section{Saddle point method} There exists many ways of computing the resolvent and its derivatives with respect to $h$, $T$, or other parameters. The loop equation method is a very good method, but there is not enough time to present it here. There are several saddle-point methods, which all coincide to leading order. We are going to present one of them, very intuitive, but not very rigorous on a mathematical ground, and not very appropriate for next to leading computations. However, it gives the correct answer to leading order. \bigskip Write: \begin{equation} Z_n(h,T)={\mathbf e}^{-{n^2\over T^2}F_n(h,T)}=\int dx_1\dots dx_n {\mathbf e}^{-{n^2\over T^2}{\cal S}(x_1,\dots,x_n)} \end{equation} where the action is: \begin{equation} {\cal S}(x_1,\dots,x_n):={T\over n}\sum_{i=1}^n V_h(x_i) -2{T^2\over n^2}\sum_{i>j} \ln{(x_i-x_j)} \end{equation} The saddle point method consists in finding configurations $x_i=\overline{x}_i$ where ${\cal S}$ is extremal, i.e. \begin{equation} \forall i=1,\dots n, \qquad \left.{\partial {\cal S}\over \partial x_i}\right|_{x_j=\overline{x}_j}=0 \end{equation} i.e., we have the {\bf saddle point equation}: \begin{equation} \forall i=1,\dots n, \qquad V'_h(\overline{x}_i) = 2{T\over n}\sum_{j\neq i} {1\over \overline{x}_i-\overline{x}_j} \end{equation} The saddle point approximation\footnote{The validity of the saddle point approximation is not proven rigorously for large number of variables. But here, we have many evidences that we can trust the results it gives. The asymptotics we are going to find have been proven rigorously by other methods. Basicaly, it is expected to work because the number of variables $n$ is small compared to the large parameter $n^2$ in the action.} consists in writting: \begin{equation} Z_n(h,T) \sim {1\over \sqrt{\det\left(\partial {\cal S}\over \partial x_i\partial x_j\right)}}\,\,{\mathbf e}^{-{n^2\over T^2}{\cal S}(\overline{x}_1,\dots,\overline{x}_n)}\,\,(1+O(1/n)) \end{equation} where $(\overline{x}_1,\dots,\overline{x}_n)$ is the solution of the saddlepoint equation which minimizes $\Re {\cal S}$. \begin{remark}\rm\small The saddle point equation may have more than one minimal solution $(\overline{x})$. - in particular if $\xi\in {\mathbb{R}}$, there are two solutions, complex conjugate of each other. - in the multicut case, there are many saddlepoints with near-minimal action. In all cases, one needs to sum over all the saddle points. Let us call $\{\overline{x}\}_I$, the collection of saddle points. We have: \begin{equation} Z_n \sim \sum_I {C_I\over \sqrt{{\cal S}''(\{\overline{x}\}_I)}}\,\,{\mathbf e}^{-{n^2\over T^2}{\cal S}(\{\overline{x}\}_I)}\,\,(1+O(1/n)) \end{equation} Each saddle point $\{\overline{x}\}_I$ corresponds to a particular minimal $n$-dimensional integration path in ${\mathbb{C}}^n$,noted $\Gamma_I$, and the coefficients $C_I\in {\mathbb{Z}}$ are such that: \begin{equation} {\mathbb{R}}^n = \sum_I C_I \Gamma_I \end{equation} \end{remark} \section{Solution of the saddlepoint equation} We recall the saddle point equation: \begin{equation}\label{sadlepointxbar} \forall i=1,\dots n, \qquad V'_h(\overline{x}_i) = 2{T\over n}\sum_{j\neq i} {1\over \overline{x}_i-\overline{x}_j} \end{equation} We introduce the function: \begin{equation}\label{defom} \omega(x):={T\over n}\sum_{j=1}^n {1\over x-\overline{x}_j} \end{equation} in the large $N$ limit, $\omega(x)$ is expected to tend toward the resolvent, at least in the case there is only one minimal saddle point. Indeed, the $\overline{x}_i$'s are the position of the eigenvalues minimizing the action, i.e. the most probable positions of eigenvalues of $M$, and thus \ref{defom} should be close to ${T\over n}{\rm tr}\, {1\over x-M}$. \subsection{Algebraic method} Compute $\omega^2(x)+{T\over n}\omega'(x)$, you find: \begin{eqnarray} \omega^2(x)+{T\over n}\omega'(x) &=& {T^2\over n^2} \sum_{i=1}^n \sum_{j=1}^n {1\over (x-\overline{x}_i)(x-\overline{x}_j)} - {T^2\over n^2} \sum_{i=1}^n {1\over (x-\overline{x}_i)^2} \cr &=&{T^2\over n^2} \sum_{i\neq j}^n {1\over (x-\overline{x}_i)(x-\overline{x}_j)} \cr &=&{T^2\over n^2} \sum_{i\neq j}^n \left({1\over x-\overline{x}_i}-{1\over x-\overline{x}_j}\right)\,{1\over \overline{x}_i-\overline{x}_j} \cr &=&{2T^2\over n^2} \sum_{i=1}^n {1\over x-\overline{x}_i}\,\sum_{j\neq i}^n {1\over \overline{x}_i-\overline{x}_j} \cr &=&{T\over n} \sum_{i=1}^n {V'_h(\overline{x}_i)\over x-\overline{x}_i} \cr &=&{T\over n} \sum_{i=1}^n {V'_h(x)-(V'_h(x)-V'_h(\overline{x}_i))\over x-\overline{x}_i} \cr &=& V'_h(x)\omega(x)-{T\over n} \sum_{i=1}^n {V'_h(x)-V'_h(\overline{x}_i)\over x-\overline{x}_i} \cr &=&(V'(x)-{h\over x-\xi})\omega(x)-{T\over n} \sum_{i=1}^n {V'(x)-V'(\overline{x}_i)\over x-\overline{x}_i} + h{\omega(\xi)\over x-\xi}\cr \end{eqnarray} i.e. we get the equation: \begin{equation} \omega^2(x)+{T\over n}\omega'(x) = V'(x)\omega(x)- P(x) - h{\omega(x)-\omega(\xi)\over x-\xi} \end{equation} where $P(x):={T\over n} \sum_{i=1}^n {V'(x)-V'(\overline{x}_i)\over x-\overline{x}_i}$ is a polynomial in $x$ of degree at most $\deg V-2$. In the large $N$ limit, if we assume\footnote{It is possible to do the calculation without droping the $1/N$ term. One gets a Ricati equation, which is equivalent to a Schroedinger equation. If one is interested in a large N limit for the resolvent, the asymptotic analysis of that Schroedinger equation (Stokes phenomenon) gives, to leading order, the same thing as when one drops the $1/N$ term. If one whishes to go beyond leading order, many subtleties occur.} that we can drop the $1/N W'(x)$ term, we get an algebraic equation, which is in this case an hyperelliptical curve. In particular at $h=0$ and $T=1$: \begin{equation} \omega(x) = {1\over 2}\left(V'(x)-\sqrt{V'^2(x)-4P(x)}\right) \end{equation} The properties of this algebraic equation have been studied by many authors, and the $T$ and $h$ derivatives, as well as other derivatives were computed in various works. Here, we briefly sketch the method. See \cite{kriechever, KazMar, eynmultimat} for more details. \subsection{Linear saddle point equation} In the large $N$ limit, both the average density of eigenvalues, and the density of $\overline{x}$ tend towards a continuous compact support density $\overline{\rho}(x)$. In that limit, the resolvent is given by: \begin{equation} \omega(x) = T \, \int_{{\rm supp}\,\,\overline{\rho}}{\overline{\rho}(x')\,dx'\over x-x'} \end{equation} i.e. \begin{equation} \forall x\in{\rm supp}\,\,\overline{\rho}, \qquad \overline{\rho}(x) = -{1\over 2i\pi T}(\omega(x+i0)-\omega(x-i0)) \end{equation} and the saddle point equation \ref{sadlepointxbar}, becomes a linear functional equation: \begin{equation}\label{sadlepointequrho} \forall x\in{\rm supp}\,\,\overline{\rho}, \qquad V'_h(x) = \omega(x+i0)+\omega(x-i0) \end{equation} The advantage of that equation, is that it is linear in $\omega$, and thus in $\overline{\rho}$. The nonlinearity is hidden in ${\rm supp}\,\,\overline{\rho}$. \subsubsection{Example: One cut} If the support of $\overline{\rho}$ is a single interval: \begin{equation} {\rm supp}\,\,\overline{\rho} = [a,b]\quad , \quad a<b \end{equation} then, look for a solution of the form: \begin{equation} \omega(x) = {1\over 2}\left(V'_h(x) - M_h(x)\sqrt{(x-a)(x-b)}\right) \end{equation} The saddle point equation \ref{sadlepointequrho} implies that $M_h(x+i0)=M_h(x-i0)$, i.e. $M_h$ has no discontinuities, and because of its large $x$ behaviour, as well as its behaviours near $\xi$, it must be a rational function of $x$, with a simple pole at $x=\xi$. $M_h$, $a$ and $b$ are entirely determined by their behaviours near poles, i.e.: \begin{equation} \omega(x) \mathop\sim_{x\to\infty} {T\over x} \end{equation} \begin{equation} \omega(x) \mathop\sim_{x\to\xi} {\rm regular}\quad \longrightarrow M_h(x)\mathop\sim_{x\to\xi} -{h\over x-\xi} \end{equation} Thus, one may write: \begin{equation} \omega(x) = {1\over 2}\left(V'(x) - M(x)\sqrt{(x-a)(x-b)} - {h\over x-\xi}\left(1-{\sqrt{(x-a)(x-b)}\over \sqrt{(\xi-a)(\xi-b)}}\right)\right) \end{equation} where $M(x)$ is now a polynomial (which still depends on $h$ and $T$ and the other parameters), it is such that: \begin{equation} M(x) = \mathop{\rm Pol}_{x\to\infty}\,\, {V'(x)\over \sqrt{(x-a)(x-b)}} \end{equation} The density is thus: \begin{equation} \overline{\rho}(x) ={1\over 2\pi T}M_h(x)\sqrt{(x-a)(b-x)} \quad , \quad {\rm supp}\,\,\overline{\rho} = [a,b] \end{equation} $$\begin{array}{r} {\epsfxsize 12cm\epsffile{curve.eps}} \end{array}$$ \subsubsection{Multi-cut solution} Let us assume that the support of $\overline{\rho}$ is made of $s$ separated intervals: \begin{equation} {\rm supp}\,\,\overline{\rho} = \cup_{i=1}^s [a_i,b_i] \end{equation} then, for any sequence of integers $n_1,n_2,\dots, n_s$ such that $\sum_{i_1}^s n_i=n$, it is possible to find a solution for the saddle point equation. That solution obeys \ref{sadlepointequrho}, as well as the conditions: \begin{equation} \forall i=1,\dots,s \quad , \quad \int_{a_i}^{b_i} \rho(x) dx = T {n_i\over N} \end{equation} The solution of the saddle point equation can be described as follows: let the polynomial $\sigma(x)$ be defined as: \begin{equation} \sigma(x):=\prod_{i=1}^s (x-a_i)(x-b_i) \end{equation} The solution of the saddle point equation \ref{sadlepointequrho}, is of the form: \begin{equation} \omega(x) = {1\over 2}\left(V'_h(x) - M_h(x)\sqrt{\sigma(x)}\right) \end{equation} where $M_h(x)$ is a rational function of $x$, with a simple pole at $x=\xi$. $M_h$, and $\sigma(x)$ are entirely determined by their behaviours near poles, i.e.: \begin{equation} \omega(x) \mathop\sim_{x\to\infty} {T\over x} \end{equation} \begin{equation} \omega(x) \mathop\sim_{x\to\xi} {\rm regular}\quad \longrightarrow M_h(x)\mathop\sim_{x\to\xi} -{h\over x-\xi} \end{equation} and by the conditions that: \begin{equation} \forall i=1,\dots,s \quad , \quad \int_{a_i}^{b_{i}} M_h(x)\sqrt{\sigma(x)} dx = 2i\pi T{n_i\over n} \end{equation} \subsection{Algebraic geometry: hyperelliptical curves} Consider the curve given by: \begin{equation} \omega(x) = {1\over 2}\left(V_h'(x) - M_h(x)\sqrt{(x-a)(x-b)}\right) \end{equation} It has two sheets, i.e. for each $x$, there are two values of $\omega(x)$, depending on the choice of sign of the square-root. - In the physical sheet (choice $+\sqrt{}$), it behaves near $\infty$ like $\omega(x)\sim T/x$ - In the second sheet (choice $-\sqrt{}$), it behaves near $\infty$ like $\omega(x)\sim V'_h(x)$ Since $\omega(x)$ is a complex valued, analytical function of a cmplex variable $x$, the curve can be thought of as the embedding of a Riemann surface into ${\mathbb{C}}\times {\mathbb{C}}$. I.e. we have a Riemann surface ${\cal E}$, with two (monovalued) functions defined on it: $p\in{\cal E}\, , \,\, \to x(p)\in{\mathbb{C}}$, and $p\in{\cal E}\, , \,\, \to \omega(p)\in{\mathbb{C}}$. For each $x$, there are two $p\in{\cal E}$ such that $x(p)=x$, and this is why there are two values of $\omega(x)$. Each of the two sheets is homeomorphic to the complex plane, cut along the segments $[a_i,b_i]$, and the two sheets are glued together along the cuts. The complex plane, plus its point at infinity, is the Riemann sphere. Thus, our curve ${\cal E}$, is obtained by taking two Riemann spheres, glued together along $s$ circles. It is a genus $s-1$ surface. $$\begin{array}{r} {\epsfxsize 14cm\epsffile{sheetg.eps}} \end{array}$$ $$\begin{array}{r} {\epsfysize 4cm\epsffile{surfpatate.eps}} \end{array}$$ \subsection{Genus zero case (one cut)} If the curve as genus zero, it is homeomorphic to the Riemann sphere ${\cal E}={\mathbb{C}}$. One can always choose a rational parametrization: \begin{equation} x(p)={a+b\over 2}+\gamma(p+{1/p}) \quad , \quad \gamma={b-a\over 4} \end{equation} \begin{equation} \sqrt{(x-a)(x-b)}=\gamma(p-1/p) \end{equation} so that $\omega$ is a rational function of $p$. That representation maps the physical sheet onto the exterior of the unit circle, and the second sheet onto the interior of the unit circle. The unit circle is the image of the two sides of the cut $[a,b]$, and the branchpoints $[a,b]$ are maped to $-1$ and $+1$. Changing the sign of the square root is equivalent to changing $p\to 1/p$. The branch points are of course the solutions of $dx/dp=0$, i.e. $dx(p)=0$: \begin{equation} dx(p) = \gamma\,\left(1-{1\over p^2}\right)\, dp \quad , \quad dx(p)=0\leftrightarrow p=\pm 1 \leftrightarrow x(p)=a,b \end{equation} There are two points at $\infty$, $p=\infty$ in the physical sheet, and $p=0$ in the second sheet. $$\begin{array}{r} {\epsfxsize 9cm\epsffile{sheetgenuszero.eps}} \,\, {\epsfxsize 7cm\epsffile{Egenuszero.eps}} \end{array}$$ \bigskip Since the resolvent $\omega(p)$ is a rational function of $p$, it is then entirely determined by its behaviour near its poles. the poles are at $p=\infty$, $p=0$, $p=p_\xi$ and $p=\overline{p}_\xi$ (the two points of ${\cal E}$ such that $x(p)=\xi$, such that $p_\xi$ is in the physical sheet, and $\overline{p}_\xi$ is in the second sheet): The boundary conditions: \begin{equation}\label{bndomgzero} \left\{ \begin{array}{l} \displaystyle \omega(p) \mathop\sim_{p\to\infty} {T\over x(p)} \cr \displaystyle \omega(p) \mathop\sim_{p\to 0} V'(x(p))-{T\over x(p)}-{h\over x(p)} \cr \displaystyle \omega(p) \mathop\sim_{p\to \overline{p}_\xi} -{h\over x(p)-\xi} \cr \displaystyle \omega(p) \mathop\sim_{p\to p_\xi} {\rm regular} \cr \end{array} \right. \end{equation} \subsubsection{$T$ derivative} Now, let us compute $\partial\omega(p)/\partial T$ at $x(p)$ fixed. Eq. \ref{bndomgzero} becomes: \begin{equation} \left\{ \begin{array}{l} \displaystyle {\partial \omega(p)\over \partial T} \mathop\sim_{p\to\infty} {1\over x(p)} \cr \displaystyle {\partial \omega(p)\over \partial T} \mathop\sim_{p\to 0} -{1\over x(p)} \cr \displaystyle {\partial \omega(p)\over \partial T} \mathop\sim_{p\to \overline{p}_\xi} {\rm regular} \cr \displaystyle {\partial \omega(p)\over \partial T} \mathop\sim_{p\to p_\xi} {\rm regular} \cr \end{array} \right. \end{equation} Moreover, we know that $\omega(x)$ has a square-root behaviour near $a$ and $b$, in $\sqrt{(x-a)(x-b)}$, and $a$ and $b$ depend on $T$, thus $\partial\omega/\partial T$ may behave in $((x-a)(x-b))^{-1/2}$ near $a$ and $b$, i.e. $\partial\omega/\partial T$ may have simple poles at $p=\pm 1$. Finaly, $\partial\omega(p)/\partial T$, has simple poles at $p=1$ and $p=-1$, and vanishes at $p=0$ and $p=\infty$, the only possibility is: \begin{equation} \left.{\partial \omega(p)\over \partial T}\right|_{x(p)}={p\over \gamma (p^{2}-1)} = {1\over p}\, {dp\over dx} \end{equation} which is better written in terms of differential forms: \begin{equation} \left.{\partial \omega(p)\over \partial T}\right|_{x(p)}\, dx(p)= {dp\over p}=d\ln{p} \end{equation} the RHS is independent of the potential, it is universal. With the notation \ref{defOm}, we have: \begin{equation} \Omega(p)dx(p)={dp\over p} \quad , \quad \Lambda(p)=\gamma p \end{equation} \subsubsection{$h$ derivative} The $h$ derivative is computed in a very similar way. \begin{equation} \left\{ \begin{array}{l} \displaystyle {\partial \omega(p)\over \partial h} \mathop\sim_{p\to\infty} O(p^{-2}) \cr \displaystyle {\partial \omega(p)\over \partial h} \mathop\sim_{p\to 0} -{1\over x(p)} \cr \displaystyle {\partial \omega(p)\over \partial h} \mathop\sim_{p\to \overline{p}_\xi} -{1\over x(p)-\xi} \cr \displaystyle {\partial \omega(p)\over \partial h} \mathop\sim_{p\to p_\xi} {\rm regular} \cr \end{array} \right. \end{equation} implies that $\partial \omega/\partial h$ can have poles at $p=\pm 1$ and at $p=\overline{p}_\xi$, and vanishes at $p=0$. The only possibility is: \begin{equation} \left.{\partial \omega(p)\over \partial h}\right|_{x(p)}={-p\,\overline{p}_\xi\over \gamma (p-\overline{p}_\xi)(p^{2}-1)} \end{equation} i.e. \begin{equation} \left.{\partial \omega(p)\over \partial h}\right|_{x(p)}\, dx(p) = {dp\over p}-{dp\over p-\overline{p}_\xi} = d\ln{p\over p-\overline{p}_\xi} \end{equation} which again is universal. With the notation \ref{defH}, we have: \begin{equation} H(p,p_\xi)dx(p)={dp\over p}-{dp\over p-{1\over p_\xi}} \quad , \quad H(p_\xi) = \ln{\left({p_\xi\over p_\xi-\overline{p}_\xi}\right)} = -\ln{\left({1\over\gamma}\,{dx\over dp}(\xi)\right)} \end{equation} \subsection{Higher genus} For general genus, the curve can be parametrized by $\theta$-functions. Like rational functions for genus 0, $\theta$-functions are the building blocks of functions defined on a compact Riemann surface, and any such function is entirely determined by its behaviour near its poles, as well as by its integrals around irreducible cycles. All the previous paragraph can be extended to that case. Let $\infty_+$ and $\infty_-$ be the points at infinity, i.e. the two poles of $x(p)$, with $\infty_+$ in the physical sheet and $\infty_-$ in the second sheet. Let $p=p_\xi$ and $p=\overline{p}_\xi$ be the two points of ${\cal E}$ such that $x(p)=\xi$, and with $p_\xi$ in the physical sheet, and $\overline{p}_\xi$ in the second sheet. The differential form $\omega(p) dx(p)$ is entirely determined by: \begin{equation}\label{ompoles} \left\{ \begin{array}{ll} \displaystyle \omega(p)dx(p) \mathop\sim_{p\to \infty_+} T\,{dx(p)\over x(p)} & \displaystyle \quad , \quad \mathop{\rm Res\,}_{\infty_+} \omega(p)dx(p)=-T \cr \displaystyle \omega(p)dx(p) \mathop\sim_{p\to \infty_-} dV(x(p)) - T{dx(p)\over x(p)}-h{dx(p)\over x(p)} & \displaystyle \quad , \quad \mathop{\rm Res\,}_{\infty_-} \omega(p)dx(p)=T+h \cr \displaystyle \omega(p)dx(p) \mathop\sim_{p\to \overline{p}_\xi} -h{dx(p)\over x(p)-\xi} & \displaystyle \quad , \quad \mathop{\rm Res\,}_{\overline{p}_\xi} \omega(p)dx(p)=-h \cr \displaystyle \omega(p)dx(p) \mathop\sim_{p\to p_\xi} {\rm regular} & \displaystyle \quad , \quad \mathop{\rm Res\,}_{p_\xi} \omega(p)dx(p)=0 \cr \displaystyle \oint_{{\cal A}_i} \omega(p) dx(p) = T{n_i\over n} = {n_i\over N} & \cr \end{array} \right. \end{equation} Since $\partial\omega /\partial T,h $ can diverge at most like $(x-a_i)^{-1/2}$ near a branch point $a_i$, and $dx(p)$ has a zero at $a_i$, the differential form $\partial \omega dx/\partial T,h$ has no pole at the branch points. \subsection{Introduction to algebraic geometry} We introduce some basic concepts of algebraic geometry. We refer the reader to \cite{Farkas, Fay} for instance. \begin{theorem} Given two points $q_1$ and $q_2$ on the Riemann surface ${\cal E}$, there exists a unique differential form $dS_{q_1,q_2}(p)$, with only two simple poles, one at $p=q_1$ with residue $+1$ and one at $p=q_2$ with residue $-1$, and which is normalized on the ${\cal A}_i$ cycles, i.e. \begin{equation}\label{defdS} \left\{ \begin{array}{l} \displaystyle \mathop{\rm Res\,}_{p\to q_1} dS_{q_1,q_2}(p)=+1 \cr \displaystyle \mathop{\rm Res\,}_{p\to q_2} dS_{q_1,q_2}(p)=-1 \cr \displaystyle \oint_{{\cal A}_i} dS_{q_1,q_2}(p) = 0 \cr \end{array} \right. \end{equation} $dS$ is called an ``abelian differential of the third kind''. \end{theorem} Starting from the behaviours near poles and irreducible cycles \ref{ompoles}, we easily find: \begin{equation} \Omega(p)dx(p)=\left.{\partial\omega(p) dx(p)\over \partial T}\right|_{x(p)} = -dS_{\infty_+,\infty_-}(p) \end{equation} \begin{equation} H(p,p_\xi) dx(p) = \left.{\partial\omega(p) dx(p)\over \partial h}\right|_{x(p)} = -dS_{\overline{p}_\xi,\infty_-}(p) = dS_{p_\xi,\infty_+}(p)-d\ln{\left(x(p)-x(p_\xi)\right)} \end{equation} \begin{theorem} On an algebraic curve of genus $g$, there exist exactly $g$ linearly independent ``holomorphic differential forms'' (i.e. with no poles), $du_i(p)$, $i=1,\dots, g$. They can be chosen normalized as: \begin{equation} \oint_{{\cal A}_i} du_j(p)=\delta_{ij} \end{equation} \end{theorem} For hyperelliptical surfaces, it is easy to see that if $L(x)$ is a polynomial of degree at most $g-1=s-2$, the differential form ${L(x)\over \sqrt{\prod_{i=1}^s (x-a_i)(x-b_i)}}dx$ is regular at $\infty$, at the branch points, and thus has no poles. And there are $g$ linearly independent polynomials of degree at most $g-1$. The irreducible cycles ${\cal A}_i$ is a contour surrounding $[a_i,b_i]$ in the positive direction. \begin{definition} The matrix of periods is defined by: \begin{equation} \tau_{ij}:=\oint_{{\cal B}_i} du_j(p) \end{equation} where the irreducible cycles ${\cal B}_i$ are chosen canonicaly conjugated to the ${\cal A}_i$, i.e. ${\cal A}_i\cap{\cal B}_j=\delta_{ij}$. In our hyperelliptical case, we choose ${\cal B}_i$ as a contour crossing $[a_i,b_i]$ and $[a_s,b_s]$. The matrix of periods is symmetric $\tau_{ij}=\tau_{ji}$, and its imaginary part is positive $\Im\tau_{ij}>0$. It encodes the complex structure of the curve. \end{definition} The holomrphic forms naturaly define an embedding of the curve into ${\mathbb{C}}^g$: \begin{definition} Given a base point $q_0\in{\cal E}$, we define the Abel map: \begin{eqnarray} {\cal E} &\longrightarrow& {\mathbb{C}}^g \cr p &\longrightarrow& {\vec u}(p) = (u_1(p),\dots,u_g(p)) \quad , \quad u_i(p):=\int_{q_0}^p du_i(p) \end{eqnarray} where the integration path is chosen so that it does not cross any ${\cal A}_i$ or ${\cal B}_i$. \end{definition} \begin{definition} Given a symmetric matrix $\tau$ of dimension $g$, such that $\Im\tau_{ij}>0$, we define the $\theta$-function, from ${\mathbb{C}}^g\to {\mathbb{C}}$ by: \begin{equation}\label{deftheta} \theta(\vec{u},\tau) = \sum_{\vec{m}\in {\mathbb{Z}}^g} {\mathbf e}^{i\pi {\vec m}^t \tau\vec{m}}\,{\mathbf e}^{2i\pi {\vec m}^t\vec{u}} \end{equation} It is an even entire function. For any $\vec{m}\in {\mathbb{Z}}^g$, it satisfies: \begin{equation} \theta(\vec{u}+\vec{m})=\theta(\vec{u}) \quad , \quad \theta(\vec{u}+\tau\vec{m})={\mathbf e}^{-i\pi(2 {\vec m}^t\vec{u} + {\vec m}^t\tau \vec{m})}\, \theta(\vec{u}) \end{equation} \end{definition} \begin{definition} The theta function vanishes on a codimension $1$ submanifold of ${\mathbb{C}}^g$, in particular, it vanishes at the odd half periods: \begin{equation} \vec{z}={{\vec m}_1+\tau\, {\vec m}_2\over 2} \,\, ,\,\,\, {\vec m}_1\in {\mathbb{Z}}^g\, ,\,\, {\vec m}_2\in {\mathbb{Z}}^g \,\, , \,\,\, ({\vec m}_1^t{\vec m}_1)\in 2{\mathbb{Z}}+1 \,\,\,\longrightarrow\,\, \theta(\vec{z})=0 \end{equation} For a given such odd half-period, we define the characteristic $\vec{z}$ $\theta$-function: \begin{equation} \theta_{\vec{z}}(\vec{u}):={\mathbf e}^{i\pi m_2\vec{u}+}\,\theta(\vec{u}+\vec{z}) \end{equation} so that: \begin{equation} \theta_{\vec{z}}(\vec{u}+\vec{m})= {\mathbf e}^{i\pi \vec{m}_2^t\vec{m}}\,\theta_{\vec{z}}(\vec{u}) \quad , \quad \theta_{\vec{z}}(\vec{u}+\tau\vec{m})= {\mathbf e}^{-i\pi \vec{m}_1^t \vec{m}}\,{\mathbf e}^{-i\pi(2 {\vec m}^t\vec{u} + {\vec m}^t\tau \vec{m})}\, \theta_{\vec{z}}(\vec{u}) \end{equation} and \begin{equation} \theta_{\vec{z}}(\vec{0})=0 \end{equation} \end{definition} \begin{definition} Given two points $p,q$ in ${\cal E}$, as well as a basepoint $p_0\in{\cal E}$ and an odd half period $z$, we define the prime form $E(p,q)$: \begin{equation} E(p,q):={\theta_{\vec{z}}(\vec{u}(p)-\vec{u}(q))\over \sqrt{dh_{\vec{z}}(p) dh_{\vec{z}}(q)}} \end{equation} where $dh_{\vec{z}}(p)$ is the holomorphic form: \begin{equation} dh_{\vec{z}}(p):= \sum_{i=1}^g \left.{\partial \theta_{\vec{z}}(\vec{u})\over \partial u_i}\right|_{\vec{u}=\vec{0}}\, du_i(p) \end{equation} \end{definition} \begin{theorem} The abelian differentials can be written: \begin{equation} dS_{q_1,q_2}(p) = d\ln{E(p,q_1)\over E(p,q_2)} \end{equation} \end{theorem} With these definitions, we have: \begin{equation} \Lambda(p) = \gamma\,{\theta_{\vec{z}}(\vec{u}(p)-\vec{u}(\infty_-))\over \theta_{\vec{z}}(\vec{u}(p)-\vec{u}(\infty_+))} \quad , \quad \gamma:=\mathop{\rm lim}_{p\to\infty_+} \,{x(p)\,\theta_{\vec{z}}(\vec{u}(p)-\vec{u}(\infty_+))\over \theta_{\vec{z}}(\vec{u}(\infty_+)-\vec{u}(\infty_-))} \end{equation} \begin{equation} H(p_\xi)={\theta_{\vec{z}}(\vec{u}(p_\xi)-\vec{u}(\infty_-))\theta_{\vec{z}}(\vec{u}(\infty_+)-\vec{u}(\overline{p}_\xi))\over \theta_{\vec{z}}(\vec{u}(p_\xi)-\vec{u}(\overline{p}_\xi))\theta_{\vec{z}}(\vec{u}(\infty_+)-\vec{u}(\infty_-))} = -\gamma\,{\theta_{\vec{z}}(\vec{u}(\infty_+)-\vec{u}(\infty_-))\over \theta_{\vec{z}}(\vec{u}(p_\xi)-\vec{u}(\infty_+))^2}\,{dh_{\vec{z}}(p_\xi)\over dx(p_\xi)} \end{equation} \section{Asymptotics of orthogonal polynomials} \subsection{One-cut case} In the one-cut case, (i.e. genus zero algebraic curve), and if $V$ is a real potential, there is only one dominant saddle point if $\xi\notin [a,b]$, and two conjugated dominant saddle points if $x\in[a,b]$. More generaly, there is a saddle point corresponding to each determination of $p_\xi$ such that $x(p_\xi)=\xi$. i.e. $p_\xi$ and $\overline{p}_\xi=1/p_\xi$. The dominant saddle point is the one such that $\Re(V_{\rm eff}(p_\xi)-V(\xi))$ is minimal. The two cols have a contribution of the same order if: \begin{equation} \Re V_{\rm eff}(p_\xi) = \Re V_{\rm eff}(\overline{p}_\xi) \end{equation} i.e. if $\xi$ is such that: \begin{equation}\label{defcutsVeff} \Re \int_{\overline{p}_\xi}^{p_\xi} W(x)dx =0 \end{equation} If the potential is real, it is easy to see that the set of points which satisfy \ref{defcutsVeff} is $[a,b]$, in general, it is a curve in the complex plane, going from $a$ to $b$, we call it the cut $[a,b]$ (similar curves were studied in \cite{moore}). Then we have: \begin{itemize} \item For $x\notin[a,b]$, we write $\xi={a+b\over 2} + \gamma (p_\xi+1/p_\xi)$, $\gamma={b-a\over 4}$: \begin{equation} p_n(\xi) \sim \,\sqrt{H(p_\xi)}\,\left(\Lambda(p_\xi)\right)^{n-N}\,{\mathbf e}^{-{N\over 2}(V_{\rm eff}(p_\xi)-V(\xi))} (1+O(1/N)) \end{equation} i.e. \begin{equation} p_n(\xi) \sim \,\sqrt{\gamma\over x'(p_\xi)}\,\left(\gamma\, p_\xi\right)^{n-N}\,{\mathbf e}^{-{N\over 2}(V_{\rm eff}(p_\xi)-V(\xi))} (1+O(1/N)) \end{equation} \item For $x\in[a,b]$, i.e. $p$ is on the unit circle $p={\mathbf e}^{i\phi}$, $\xi={a+b\over 2}+2\gamma\cos\phi$: \begin{eqnarray} p_n(\xi) &\sim& \,\sqrt{H(p_\xi)}\,\left(\Lambda(p_\xi)\right)^{n-N}\,{\mathbf e}^{-{N\over 2}(V_{\rm eff}(p_\xi)-V(\xi))} (1+O(1/N)) \cr && + \,\sqrt{H(\overline{p}_\xi)}\,\left(\Lambda(\overline{p}_\xi)\right)^{n-N}\,{\mathbf e}^{-{N\over 2}(V_{\rm eff}(\overline{p}_\xi)-V(\xi))} (1+O(1/N)) \end{eqnarray} i.e. \begin{equation} p_n(\xi) \sim {\gamma^{n-N}\over \sqrt{2\sin\phi(\xi)}}\,2\cos{\left(N\pi\int_a^\xi \rho(x)dx - (n-N+{1\over 2}) \phi(\xi) + \alpha\right)} (1+O(1/N)) \end{equation} i.e. we have an oscillatory behaviour $$\begin{array}{r} {\epsfxsize 10cm\epsffile{asymp.eps}} \end{array}$$ \end{itemize} \subsection{Multi-cut case} \label{sectmulticutasymp} In the multicut case, in addition to having saddle-points corresponding to both determinantions of $p_\xi$, we have a saddle point for each filling fraction configuration $n_1,\dots, n_s$ with $\sum_{i=1}^s n_i=n$. We write: \begin{equation} \epsilon_i = {n_i\over N} \end{equation} The saddle point corresponding to filling fractions which differ by a few units, contribute to the same order, and thus cannot be neglected. One has to consider the sommation over filling fractions \cite{BDE}. Thus, one has to consider the action of a saddle point as a function of the filling fractions. We leave as an exercise for the reader to prove that the derivatives of $F$ are given by: \begin{equation} {\partial F\over \partial \epsilon_i} = -\oint_{{\cal B}_i} W(x)dx \end{equation} and: \begin{equation} {\partial^2 F\over \partial \epsilon_i\partial T} = -2i\pi (u_i(\infty_+)-u_i(\infty_-)) \end{equation} \begin{equation} {\partial^2 F\over \partial \epsilon_i\partial h} = -2i\pi (u_i(p_\xi)-u_i(\infty_+)) \end{equation} \begin{equation}\label{dFtauij} {\partial^2 F\over \partial \epsilon_i\partial \epsilon_j} = -2i\pi \tau_{ij} \end{equation} The last relation implies that $\Re F$ is a convex function of $\epsilon$, thus it has a unique minimum: \begin{equation} \vec\epsilon^* \quad , \quad \Re \left.{\partial F\over \partial \epsilon_i}\right|_{\vec\epsilon=\vec\epsilon^*} = 0 \end{equation} We write: \begin{equation} \zeta_i := -{1\over 2i\pi} \left.{\partial F\over \partial \epsilon_i}\right|_{\vec\epsilon=\vec\epsilon^*} \quad , \quad \zeta_i\in {\mathbb{R}} \end{equation} We thus have the Taylor expansion: \begin{eqnarray} F(T,h,\vec\epsilon) &\sim& F(1,0,\vec\epsilon^*) -2i\pi \vec\zeta^t (\vec\epsilon-\vec\epsilon^*) +(T-1) {\partial F\over \partial T} + {h\over 2} (V_{\rm eff}(p_\xi)-V(\xi)) \cr && +{(T-1)^2\over 2} {\partial^2 F\over \partial T^2}-(T-1)h \ln{\Lambda(p_\xi)}-{h^2\over 2} \ln{H(p_\xi)}\cr && -2i\pi (\vec\epsilon-\vec\epsilon^*)^t \tau (\vec\epsilon-\vec\epsilon^*) -2i\pi (T-1) (\vec\epsilon-\vec\epsilon^*)^t (\vec{u}(\infty_+)-\vec{u}(\infty_-)) \cr && -2i\pi h (\vec\epsilon-\vec\epsilon^*)^t (\vec{u}(p_\xi)-\vec{u}(\infty_+)) + \dots \end{eqnarray} Thus: \begin{eqnarray} Z &\sim& \sum_I C_I {\mathbf e}^{-N^2 F(\{x\}_I)} \cr &\sim& \sum_{p=p_\xi,\overline{p}_\xi} {\mathbf e}^{-N^2 F(1,0,\vec\epsilon^*)}{\mathbf e}^{N^2\left(-(T-1) {\partial F\over \partial T} - {h\over 2} (V_{\rm eff}(p)-V(\xi)) -{(T-1)^2\over 2} {\partial^2 F\over \partial T^2}+(T-1)h \ln{\Lambda(p)}+{h^2\over 2} \ln{H(p)}\right)}\cr && \sum_{\vec{n}} {\mathbf e}^{i\pi (\vec{n}-N\vec\epsilon^*)^t \tau (\vec{n}-N\vec\epsilon^*)} {\mathbf e}^{2i\pi N \vec\zeta^t (\vec{n}-N\vec\epsilon^*)} \cr && \qquad {\mathbf e}^{2i\pi N(T-1) (\vec{n}-N\vec\epsilon^*)^t (\vec{u}(\infty_+)-\vec{u}(\infty_-)) } {\mathbf e}^{2i\pi Nh (\vec{n}-N\vec\epsilon^*)^t (\vec{u}(p)-\vec{u}(\infty_+))} \cr \end{eqnarray} In that last sum, because of convexity, only values of $\vec{n}$ which don't differ from $N\vec\epsilon^*$ form more than a few units, contribute substantialy. Therefore, up to a non perturbative error (exponentialy small with $N$), one can extend the sum over the $n_i$'s to the whole ${\mathbb{Z}}^g$, and recognize a $\theta$-function (see \ref{deftheta}): \begin{eqnarray} Z &\sim& \sum_{p=p_\xi,\overline{p}_\xi} {\mathbf e}^{-N^2 F(1,0,\vec\epsilon^*)}{\mathbf e}^{N^2\left((T-1) {\partial F\over \partial T} + {h\over 2} (V_{\rm eff}(p)-V(\xi)) +{(T-1)^2\over 2} {\partial^2 F\over \partial T^2}+(T-1)h \ln{\Lambda(p)}+{h^2\over 2} \ln{H(p)}\right)}\cr && {\mathbf e}^{i\pi N^2 \vec\epsilon^{*t} \tau \vec\epsilon^*} {\mathbf e}^{-2i\pi N^2 \vec\zeta^t \vec\epsilon^*} {\mathbf e}^{-2i\pi N^2(T-1) \vec\epsilon^{*t} (\vec{u}(\infty_+)-\vec{u}(\infty_-)) } {\mathbf e}^{-2i\pi N^2 h \vec\epsilon^{*t} (\vec{u}(p)-\vec{u}(\infty_+))} \cr && \theta( N (\vec\zeta-\tau \vec\epsilon^*) +N(T-1) (\vec{u}(\infty_+)-\vec{u}(\infty_-)) +Nh (\vec{u}(p)-\vec{u}(\infty_+)) ,\tau) \cr \end{eqnarray} with $T-1={n-N\over N}$ and $h=0$ or $h=1/N$, we get the asymptotics: \begin{eqnarray}\label{asympmulticut} p_n(\xi) &\sim& \sum_{x(p)=\xi} \sqrt{H(p)}\, (\Lambda(p))^{n-N}\,{\mathbf e}^{-{N\over 2} (V_{\rm eff}(p)-V(\xi))}\, {\mathbf e}^{-2i\pi N \vec\epsilon^{*t} (\vec{u}(p)-\vec{u}(\infty_+))} \cr && {\theta(N (\vec\zeta-\tau \vec\epsilon^*) +(n-N) (\vec{u}(\infty_+)-\vec{u}(\infty_-)) +(\vec{u}(p)-\vec{u}(\infty_+)) ,\tau) \over \theta(N (\vec\zeta-\tau \vec\epsilon^*) +(n-N) (\vec{u}(\infty_+)-\vec{u}(\infty_-)) ,\tau)} \cr \end{eqnarray} Again, depending on $\xi$, we have to choose the determination of $p_\xi$ which has the minimum energy. If we are on a cut, i.e. if condition \ref{defcutsVeff} holds, both determinations contribute. To summarize, outside the cuts, the sum \ref{asympmulticut} reduces to only one term, and along the cuts, the sum \ref{asympmulticut} contains two terms. \section{Conclusion} We have shown how the asymptotics of orthogonal polynomials (a notion related to integrability) is deeply related to algebraic geometry. This calculation can easily be extended to many generalizations, for multi-matrix models \cite{eynchain, eynchaint, BEHAMS, eynhabilit}, non-hermitean matrices ($\beta=1,4$) \cite{eynbetapol}, rational potentials \cite{BEHsemiclas}, ... \subsection*{Aknowledgements} The author wants to thank the organizer of the Les Houches summer school Applications of Random Matrices in Physics June 6-25 2004.
{ "timestamp": "2005-03-22T13:59:08", "yymm": "0503", "arxiv_id": "math-ph/0503052", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503052" }
\section{Introduction}\label{I} In the theory of non-equilibrium statistical mechanics, the entropy production is a crucial quantity. Typical for non-equilibrium steady states is the (strict) positivity of the entropy production which is accompanied by presence of currents and hence breakage of time-reversal symmetry. In \cite{maes} the entropy production was introduced at the level of trajectories. The idea is that even for a non-equilibrium system in the steady state, the space-time measure is still a Gibbs measure and the asymmetric part under time reversal of the Hamiltonian of the space-time Gibbs measure is the entropy production. Hence, in this formalism, the entropy production is a trajectory-valued function which measures the degree of irreversibility. The relative entropy density between the forward and the backward process is then the {\em mean entropy production}, which is strictly positive if and only if the process is reversible (i.e., in ``detailed balance'', or ``equilibrium''). See also \cite{mr} for the relation between strictly positive mean entropy production and reversibility, and \cite{JQQ} for a recent account on entropy production in a broader context. In this point of view, in order to {\em estimate} the entropy production, e.g., in order to test the reversibility of the process, one needs a way to compute it from trajectories. This is quite similar to the problem of estimating the entropy of a process. A basic approach consists in approximating the measure by its empirical version \cite{shields}. Another particularly useful and simple way of estimating entropy is via the Ornstein-Weiss theorem \cite{shields,weiss}. The entropy is approximated by the logarithm of the return time of the first $n$ symbols, divided by $n$. Similarly, relative entropy density can be estimated using waiting times, see e.g. \cite{konto}. In this paper we consider Gibbsian processes with values in a finite alphabet, and with summable modulus of continuity. We introduce an estimator of the entropy production based on a single trajectory (we call it the hitting-time estimator) and an estimator based on two independent trajectories (which we call the waiting-time estimator). For both estimators we obtain consistency and asymptotic normality, with an asymptotic variance coinciding with that of the entropy production. Moreover, for the waiting-time estimator we obtain a large deviation principle. It turns out that its large deviation function has the same symmetry as in the so-called fluctuation theorem \cite{galco,lebspo,maes}, and in fact coincides with the large deviation function of the entropy production itself in the region where it is finite. This shows that the estimator has also nice properties from the physical point of view. The basic technique we use is the exponential law with good control of the error for hitting and waiting times \cite{miguel,miguelnew}. This provides us with a precise control of the difference between the estimators and the entropy production. The rest of the paper is organized as follows. In section 2 we introduce the entropy production in the spirit of \cite{maes}, see also \cite{JQQ}. In section 3 we introduce the estimators, in section 4 we state their fluctuation properties and section 5 is devoted to proofs. \section{Context}\label{C} We will consider a stationary process $\{X_n:n\in\mathbb Z\}$ taking values in a finite set $A$. A trajectory of this process, i.e., an element of $A^{\mathbb Z}$ will be denoted by $\omega$. The space of all trajectories is denoted by $\Omega=A^{\mathbb Z}$. For $\omega\in\Omega$, and $n\in\mathbb Z$, $\theta_n \omega$ is the trajectory defined by $(\theta_n \omega)_k := \omega_{k+n}$. A function $f:\Omega\to\mathbb R$ is called local if it depends only on finitely many coordinates of the trajectory. A block of length $n$ is a sequence $x_1^n:=x_1\cdots x_n$ of elements of $A$. The cylinder $[x_1^n]$ based on $x_1^n$ is the set of $\omega\in\Omega$ such that $\omega_j=x_j$ for $j=1,\ldots,n$. The distribution ${\mathbb P}$ of the process $\{X_n:n\in\mathbb Z\}$ is supposed to be a translation invariant Gibbs measure with translation invariant potential $U$. The associated ``energy per site'' $f_U$ is defined as usual: $$ f_U(\omega):= \sum_{\Lambda\ni 0} \frac{U(\Lambda,\omega)}{|\Lambda|} $$ where the sum runs over all finite subsets of $\mathbb Z$ (containing the origin). It is well-known that under mild assumptions \cite{Geo} there exists a constant $K>0$ such that for all $x_1^n$, all $\omega\in[x_1^n]$, we have the uniform estimate \begin{equation}\label{gibbs} K^{-1} \leq \frac{{\mathbb P}([x_1^n])}{\exp(n P(f_U) + \sum_{j=0}^{n-1} f_U(\theta_j \omega))} \leq K \end{equation} where $P(f_U)$ is the ``pressure'' associated to $U$. For a block $x_1^n$, its time reverse is denoted by $x_n^1=x_n x_{n-1}\cdots x_1$. Similarly, $X_1^n$ denotes the random block $X_1\cdots X_n$ whereas $X_n^1$ denotes the random block $X_n\cdots X_1$. For the definition of the entropy production of the process $\{X_n:n\in\mathbb Z\}$, we follow \cite{maes,mrv}. We denote by ${\mathbb P}^{{\scriptscriptstyle R}}$ the distribution of the time-reversed process, i.e., the distribution of $\{X_{-n}:n\in\mathbb Z\}$. The entropy production of the process up to time $n$ is defined as \begin{equation}\label{dracula} {\dot{\mathbf S}}_n(X_1,\ldots,X_n):=\log\frac{{\mathbb P}([X_1^n])}{{\mathbb P}([X_n^1])}= \log\frac{{\mathbb P}([X_1^n])}{{\mathbb P}^{{\scriptscriptstyle R}}([X_1^n])}\,\cdot \end{equation} This random variable is a measure of the irreversibility of the process up to time $n$. We recall that the relative entropy density $h({\mathbb Q}|{\mathbb P})$ between a translation invariant probability measure ${\mathbb Q}$ on $\Omega$ and ${\mathbb P}$ is the limit $$ h({\mathbb Q}|{\mathbb P})=\lim_{n\rightarrow\infty}\frac{H_n({\mathbb Q}|{\mathbb P})}{n} $$ where $$ H_n({\mathbb Q}|{\mathbb P}):=\sum_{x_1^n\in A^n} {\mathbb Q}([x_1^n]) \log\frac{{\mathbb Q}([x_1^n])}{{\mathbb P}([x_1^n])}\,\cdot $$ We have the following well-known properties \cite{Geo}: $$ h({\mathbb Q}|{\mathbb P})=P(f_U)-\int f_U\ d{\mathbb Q} + s({\mathbb Q}) $$ where $s({\mathbb Q})$ is the entropy density of ${\mathbb Q}$. Moreover, $h({\mathbb Q}|{\mathbb P})\geq 0$, with equality if and only if ${\mathbb Q}$ is an equilibrium state for $U$ (variational principle). Using \eqref{gibbs} and the Ergodic Theorem, it follows immediately that \begin{equation}\label{chou} \lim_{n\rightarrow\infty} \frac{{\dot{\mathbf S}}_n(X_1,\ldots, X_n)}{n}= h({\mathbb P}|{\mathbb P}^{{\scriptscriptstyle R}}):={\mathbf{MEP}}\quad {\mathbb P}-\textup{almost surely}\,. \end{equation} This quantity is called the {\em mean entropy production}. It is equal to $0$ if and only if the process is reversible, i.e., the potential $U^{{\scriptscriptstyle R}}$ associated to ${\mathbb P}^{{\scriptscriptstyle R}}$ is physically equivalent to the potential $U$. We now precise the classes of potentials for which our results hold. A first restriction is to assume that $f_U$ has a summable modulus of continuity, i.e., \begin{equation}\label{avion} \sum_{n\geq 1} \textup{var}_n f_U <\infty \end{equation} where $$ \textup{var}_n f_U := \sup\{ |f_U(\omega)-f_U(\omega')|: \omega_i=\omega'_i, \forall |i|\leq n\}\,. $$ In particular this implies that ${\mathbb P}$ is the unique Gibbs measure (equilibrium state) with potential $U$. It is convenient to work with an $f_U$ which depends only on ``future'' coordinates, that is, only on $\omega_1,\omega_2,\ldots$. It is indeed proved in \cite{CQ} that if $f_U$ satisfies \eqref{avion}, then there exists a function $f_U^+(\omega):=f_U^+(\omega_1,\omega_2,\ldots)$ which is physically equivalent to $f_U$, i.e., which gives the same Gibbs measure as $f_U$, and which has also summable variations. ``Physically equivalent'' means there exists a measurable function $\kappa=\kappa_U$ and a real constant $C=C_U$ such that $f_U^+ = f_U + \kappa - \kappa\circ \theta + C$. It is easy to check that \eqref{gibbs} holds with $f_U^+$ in place of $f_U$ by suitably modifying the constant $K$. Moreover, we can simplify the notations by assuming that $P(f_U^+)=0$. If it is not the case, replace $f_U^+$ by the physically equivalent potential $f_U^+ - P(f_U^+)$. Recapitulating, we obtain that there exists a constant $K'>0$ such that for all $x_1^n$, all $\omega\in[x_1^n]$, we have the uniform estimate \begin{equation}\label{gibbsbis} K'^{-1} \leq \frac{{\mathbb P}([x_1^n])}{\exp(\sum_{j=0}^{n-1} f_U^+(\theta_j \omega))} \leq K'\,. \end{equation} Of course, the same estimate holds for ${\mathbb P}^{{\scriptscriptstyle R}}$ with the obvious modifications. This immediately gives that there exists some constant $\tilde{K}>0$ such that \begin{equation}\label{porc} -\tilde{K} \leq {\dot{\mathbf S}}_n - \sum_{j=0}^{n-1} [(f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+)\circ\theta_j] \leq \tilde{K} \end{equation} for all $n\geq 1$. Using \eqref{chou} and the Ergodic Theorem, we deduce immediately that $$ {\mathbf{MEP}} = \int (f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+)\ d{\mathbb P}\,. $$ The possibility of working with a ``one-sided'' potential physically equivalent to the ``two-sided'' one is very important because it will allow us to apply known results obtained by transfer-operator techniques. The assumption \eqref{avion} also implies a ``strong mixing'' property which is needed to prove our results. When dealing with central limit asymptotics, we will restrict ourselves to potentials having exponentially decreasing modulus of continuity, i.e., \begin{equation}\label{bateau} \exists C>0, 0\leq \eta <1\quad\textup{such that}\quad \textup{var}_n f_U \leq C \eta^n\quad\forall n\geq 1\,. \end{equation} This will allow us to use a result proved in \cite{PP}. We will precise further these points at the appropriate places. \begin{remark} If we assume that $$ \sum_{\Lambda: \min \Lambda =0} \textup{diam}(\Lambda) \textup{var}(U(\Lambda,\cdot)) <\infty $$ where $\textup{var}(U(\Lambda,\cdot)):=\max(U(\Lambda,\cdot))-\min(U(\Lambda,\cdot))$ this implies \eqref{avion}, see \cite{CQ}. \end{remark} \section{Estimators of entropy production based on hitting and return times}\label{wr} In this section we introduce two estimators based on a single trajectory or on two independent trajectories. To define them we have to introduce hitting times. The hitting time of a cylinder $[x_1^n]$ is defined as $$ \T_{x_1^n}(\omega) :=\inf\{k\geq 1: \theta_k \omega\in [x_1^n] \}\,. $$ For the sake of convenience, we introduce the notations $$ \T^+_n(\omega):=\T_{\omega_1^n}(\omega) \quad\textup{and}\quad \T^-_n(\omega):=\T_{\omega_n^1}(\omega)\,. $$ The hitting-time estimator $\dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n(\omega)$ of the entropy production is defined as $$ \dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n(\omega):= \log\frac{\T^-_n(\omega)}{\T^+_n(\omega)}\,\cdot $$ In words, this is the difference of the logarithms of the first time at which we observe the first $n$ symbols in reversed order in the trajectory and the first return time of the first $n$ symbols. It will follow from our analysis that {\em typically}, $\T^-_n\gg \T^+_n$ if the process is not reversible. Hence our hitting-time estimator of the entropy production will be typically positive. The waiting-time estimator $\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n(\omega,\omega')$ of the entropy production is based on two trajectories $\omega, \omega'$ chosen {\em independently} of one another according to ${\mathbb P}$. We introduce the following convenient notations: $$ {\mathbf W}^+_n(\omega,\omega'):=\T_{\omega_1^n}(\omega')\quad\textup{and}\quad {\mathbf W}^-_n(\omega,\omega'):=\T_{\omega_n^1}(\omega')\,. $$ The waiting-time estimator is then defined as $$ \dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n(\omega,\omega'):= \log\frac{{\mathbf W}^-_n(\omega,\omega')}{{\mathbf W}^+_n(\omega,\omega')}\,\cdot $$ The main motivation to introduce this alternative estimator is that we will obtain a better control of its large deviation properties. \begin{remark} We can define two other estimators based on the so-called matching times \cite{konto}. They are in some sense the ``duals'' of the above estimators. To introduce the ``dual" of the hitting-time estimator, consider the first $n$ symbols $x_1,\ldots x_n$ of the process and define \[ \L^+_n = \min \{ k\leq n: \mbox{the word}\ x_1^k\ \mbox{does not reappear in} \ x_1^n\} \] and \[ \L^-_n = \min \{ k\leq n: \mbox{the reversed word}\ x_k^1\ \mbox{does not reappear in} \ x_1^n\} \] Then the estimator of the entropy production dual to the hitting-time estimator is given by $\log(\L^+_n/\L^+_-)$. The advantage of these estimators is that they are based on a trajectory of finite length $n$. However, all the asymptotic fluctuation properties of these estimators can be derived from the ones of the present paper by the duality relations. So we do not study them in detail in this paper. \end{remark} \section{Convergence and fluctuations of the estimators}\label{MR} We now state our results on consistency and asymptotic normality for the estimators we just introduced, as well as large deviation properties for estimators based on two independent trajectories. Recall that ${\mathbf{MEP}}$ is the mean entropy production, see \eqref{chou}. \subsection{Almost-sure approximation and consistency} The following theorem provides an almost-sure approximation of ${\dot{\mathbf S}}_n$, the entropy production up to time $n$ (see \eqref{dracula}), by both the return-time and the waiting-time estimators. \begin{theorem}\label{thm1} Assume that \eqref{avion} holds. Then there exists a constant $C=C({\mathbb P})>0$ such that \begin{enumerate} \item Eventually ${\mathbb P}$-almost surely $$ -C\log n\leq \dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n - {\dot{\mathbf S}}_n \leq C\log n\,; $$ \item Eventually ${\mathbb P}\!\times\!{\mathbb P}$-almost surely $$ -C\log n\leq \dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n - {\dot{\mathbf S}}_n \leq C\log n\,. $$ \end{enumerate} \end{theorem} Using the previous theorem and \eqref{chou}, we immediately obtain the following corollary establishing the consistency of our entropy production estimators. \begin{corollary} We have the following almost-sure convergences: \begin{enumerate} \item ${\mathbb P}$-almost surely $$ \lim_{n\rightarrow\infty}\frac{\dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n}{n}= {\mathbf{MEP}}\,; $$ \item ${\mathbb P}\!\times\!{\mathbb P}$-almost surely $$ \lim_{n\rightarrow\infty}\frac{\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n}{n}= {\mathbf{MEP}}\, . $$ \end{enumerate} \end{corollary} \subsection{Asymptotic normality} The expectation with respect to ${\mathbb P}$ is denoted by $\mathbb E$. Let \begin{equation}\label{variance} \sigma^2:=\sum_{\ell\geq 1} \left[ \mathbb E((f_U^+ -f_{U^{{\scriptscriptstyle R}}}^+)\cdot (f_U^+ -f_{U^{{\scriptscriptstyle R}}}^+)\circ \theta_\ell)-(\mathbb E(f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+))^2\right]\,. \end{equation} It can be showed that $\sigma^2<\infty$ if \eqref{bateau} holds. It is well-known that $\sigma^2>0$ unless $U$ is physically equivalent to $U^{{\scriptscriptstyle R}}$, i.e., $f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+ $ is a co-boundary, which in turn is equivalent with ${\mathbb P}={\mathbb P}^{{\scriptscriptstyle R}}$, i.e., the process is reversible. For more details on this, we refer to \cite{PP}. \begin{theorem}\label{pouac} Assume that \eqref{bateau} holds. Then we have the following central limit asymptotics: \begin{enumerate} \item For the hitting-time estimator $$ \frac{\dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n - n {\mathbf{MEP}}}{\sqrt{n}}\to \mathcal{N}(0,\sigma^2)\,,\textup{as}\;n\to\infty $$ in ${\mathbb P}$-distribution. \item For the waiting-time estimator $$ \frac{\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n - n {\mathbf{MEP}}}{\sqrt{n}}\to \mathcal{N}(0,\sigma^2)\,,\textup{as}\;n\to\infty $$ in ${\mathbb P}\!\times\!{\mathbb P}$-distribution. \end{enumerate} Moreover, \begin{equation}\label{varvar} \lim_{n\rightarrow\infty}\frac{\textup{Var}(\dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n)}{n}=\lim_{n\rightarrow\infty}\frac{\textup{Var}(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n)}{n}= \sigma^2 \end{equation} where $\textup{Var}$ denotes the variance. \end{theorem} \begin{remark} Using the results of \cite{KMS}, we could extend the previous theorem to potentials with a modulus of continuity decreasing polynomially, i.e., like $1/n^{\alpha}$ for $\alpha>0$ large enough. \end{remark} \subsection{Large deviations} Our goal is to analyze the deviations of order one of $\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n/n$ around the mean entropy production ${\mathbf{MEP}}$. To this end, we introduce the following ``free-energy-like'' function, which is nothing but the scaled-cumulant generating function for the process $(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n)$: $$ \mathcal{W}_U(p):=\lim_{n\rightarrow\infty}\frac{1}{n}\log \mathbb E_{{\mathbb P}\!\times\!{\mathbb P}}\left(e^{p\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n}\right)\, , \; p\in\mathbb R $$ provided the limit exists. On another hand, define the scaled cumulant generating function for the process $({\dot{\mathbf S}}_n)$ as: $$ \mathcal{E}_U(p):=\lim_{n\rightarrow\infty}\frac{1}{n}\log \mathbb E_{{\mathbb P}}\left(e^{p{\dot{\mathbf S}}_n}\right)\, , p\in\mathbb R\,. $$ It is easy to deduce from \eqref{gibbsbis} that $$ \mathcal{E}_U(p)=P(-p f_{U^{{\scriptscriptstyle R}}}^+ + (1+p)f_U^+)\ ,\;\forall p\in\mathbb R\,. $$ From this formula one immediately sees that $$ \mathcal{E}_U(-1-p) = \mathcal{E}_{U^{{\scriptscriptstyle R}}}(p)\,. $$ On another hand, it is obvious from the definition of ${\dot{\mathbf S}}_n$ that $$ \mathcal{E}_U(p) = \mathcal{E}_{U^{{\scriptscriptstyle R}}}(p)\,. $$ Hence $$ \mathcal{E}_U(-1-p) = \mathcal{E}_{U}(p) $$ which is a version of the Gallavotti-Cohen fluctuation theorem, see \cite{galco}, \cite{lebspo}, \cite{maes}. Notice that $\mathcal{E}_U\equiv 0$ if $U$ is physically equivalent to $U^{{\scriptscriptstyle R}}$. We now state a large deviation result for ${\dot{\mathbf S}}_n$. Let ${\mathcal I}_U$ be the Legendre transform of ${\mathcal E}_U$, i.e., $$ {\mathcal I}_U(q)= \sup_{p\in\mathbb R} \left(pq - {\mathcal E}_U(p)\right) \,. $$ Then we have \begin{proposition}\label{macbeth} Assume that \eqref{avion} holds and that the process $(X_n)$ is not reversible (i.e., that $U$ is not physically equivalent to $U^{{\scriptscriptstyle R}}$). Then the function $p\mapsto {\mathcal E}_U(p)$ is continuously differentiable and strictly convex. Moreover, there exists an open interval $(\underline{q}, \overline{q})$ such that, for every interval $J$ with $J\cap (\underline{q}, \overline{q})\neq \emptyset$ $$ \lim_{n\rightarrow\infty} \frac{1}{n}\log {\mathbb P}\left\{\frac{{\dot{\mathbf S}}_n(X_1,\ldots,X_n)}{n} \in J \right\}= -\inf_{q\in J\cap (\underline{q}, \overline{q})} {\mathcal I}_U(q)\,. $$ \end{proposition} The interest of this result lies in its formulation adapted to our context and convenient to state the next result, the main one of this section. In essence such kind of result appears, e.g., in \cite{MV}. \begin{theorem}\label{pouic} If assumption \eqref{avion} holds then we have \begin{equation} \mathcal{W}_U(p)= \left\{ \begin{array}{l} \mathcal{E}_U(p)\quad\textup{if}\;-1<p<1\\ +\infty \quad\textup{otherwise}\,. \end{array}\label{boulgakov} \right. \end{equation} In particular, if the process $(X_n)$ is not reversible (i.e., $U$ is not physically equivalent to $U^{{\scriptscriptstyle R}}$) then $\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n$ and ${\dot{\mathbf S}}_n$ have the same large deviations in the open interval $(c_-,c_+)$, with $c_-:=\lim_{p\to -1}\mathcal{E}_U'(p)<0$ and $c_+:=\lim_{p\to 1}\mathcal{E}_U'(p)>0$: For every interval $J$ with $J\cap (c_-, c_+)\neq \emptyset$ \begin{equation}\label{pelleas} \lim_{n\rightarrow\infty} \frac{1}{n}\log {\mathbb P}\left\{\frac{\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n}{n} \in J \right\}= -\inf_{q\in J\cap (c_-, c_+)} {\mathcal I}_U(q)\,. \end{equation} \end{theorem} \bigskip It is easy to check that ${\mathbf{MEP}}\in (c_-,c_+)$. Indeed ${\mathcal E}_U'(0)={\mathbf{MEP}}$ (one uses differentiability and convexity to prove that). The next proposition highlights the symmetry properties of $\mathcal{W}$. We write explicitly the dependence of $\mathcal{W}$ on the potential $U$. \begin{proposition}\label{symmetry} Under assumption \eqref{avion} we have the following identities \begin{enumerate} \item For all $-1<p\leq 0$, we have $$ \mathcal{W}_{U}(-1-p)= \mathcal{W}_{U^{{\scriptscriptstyle R}}}(p) = \mathcal{W}_U(p)= \mathcal{W}_{U^{{\scriptscriptstyle R}}}(-1-p)\,. $$ \item For all $p\in(-1,1)$, we have $$ \mathcal{W}_{U}(p) = \mathcal{W}_{U^{{\scriptscriptstyle R}}}(p)\,. $$ \end{enumerate} \end{proposition} \begin{remark} One may ask why we did not study the large deviations of $\dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n$, the hitting-time estimator. Indeed, the analysis of the corresponding scaled cumulant generating function is made more complicated due to the effect of ``too soon'' recurrent cylinders. We shall not detail more on this. Following the approach of \cite{CGS}, we can obtain a partial counterpart of Theorem \ref{pouic} for $\dot{{\mathcal S}}^{{\scriptscriptstyle H}}_n$ : its scaled cumulant generating function coincides with $\mathcal{E}_U(p)$ but only in an {\em implicit} interval $[\tilde{c}_-,\tilde{c}_+]$, where $\tilde{c}_-<0$ and $\tilde{c}_+>0$. \end{remark} \section{Proofs}\label{proofs} \subsection{Key lemmas} The following results are the main tools to derive our results. \begin{keylemma}\label{MKL} Assume that ${\mathbb P}$ is a translation invariant Gibbs measure such that \eqref{avion} holds. Then there exist strictly positive constants $c,C,\rho_1,\rho_2$, with $\rho_1\leq \rho_2$, such that for all $n\in\mathbb N$, all cylinders $[a_1^n]$ and all $t>0$ there exists $\rho(a_1^n)\in[\rho_1,\rho_2]$ such that \begin{equation}\label{strong-approximation} \Big\vert {\mathbb P}\{\T_{a_1^n}{\mathbb P}([a_1^n])>t\}- e^{-\rho(a_1^n)t}\Big\vert \leq C e^{-c n} e^{-\rho(a_1^n)t}\,. \end{equation} \end{keylemma} \begin{proof} In \cite{miguel}, the author proved this result under the assumption that the process is $\psi$-mixing. Besides, it is proved in \cite{walters} that if $f_U^+$ has summable variations, then the process $(X_n)$ is $\psi$-mixing. (This can be read off the proof of Theorem 3.2 in \cite{walters}.) \end{proof} The next lemma will be crucial to control certain moments. This is a rewriting of Lemma 9 in \cite{miguel}. \begin{lemma}\label{tarte} For all cylinder $[a_1^n]$, all $t$ such that $t \leq 1/2$, we have $$ 1-e^{-\rho_1 t} \leq {\mathbb P}\{\T_{a_1^n}{\mathbb P}([a_1^n])\leq t\}\leq 1- e^{-\rho_2 t} $$ where $\rho_1,\rho_2$ are the constants of Key-lemma \ref{MKL}. \end{lemma} We now state the analog to Key-lemma \ref{MKL} for return times. To do so, we need to define the set of $n$-cylinders with ``internal periodicity'' $k\leq n$: $$ \mathcal{S}_k(n):=\{[a_1^n]: \min\{j\in\{1,...,n\}: [a_1^n]\cap \theta_{j}[a_1^n]\neq\emptyset\}=k \}\,. $$ Notice that the set of $n$-cylinders can be written as $\bigcup_{1\leq p\leq n} \mathcal{S}_k(n)$. \begin{keylemma}\label{MKLbis} Assume that ${\mathbb P}$ is a translation invariant Gibbs measure such that \eqref{avion} holds. Then there exist strictly positive constants $c,c',C$ such that for any $n\in\mathbb N$, any $k\in\{1,...,n\}$,\ any cylinder $[a_1^n] \in \mathcal{S}_k(n)$, one has for all $t\geq k$ \begin{equation} \Big\vert {\mathbb P}\big\{\omega:\T_{a_1^n}(\omega){\mathbb P}([a_1^n]) >t\big\vert\ [a_1^n]\big\} - \zeta(a_1^n) \exp(-\zeta(a_1^n)t)\Big\vert \leq C \ e^{-c n}\ e^{-c' t} \end{equation} where $\zeta(a_1^n)$ is such that $|\ \zeta(a_1^n)-\rho(a_1^n)|\leq D e^{-c n}$, for some $D>0$. The parameter $\rho(a_1^n)$ is defined in Key-lemma \ref{MKL}. Moreover, $$ {\mathbb P}\{\omega:\T_{a_1^n}(\omega)>t\ | \ [a_1^n]\}=1\quad \textup{for all}\quad t<k\, . $$ \end{keylemma} \begin{proof} This Key-lemma is a rewriting of \cite[Section 6]{miguelnew}. As for the previous Key-lemma, the assumption is that the process is $\psi$-mixing. \end{proof} \subsection{Proof of Theorem \ref{thm1}} Let us start with the proof of the second statement of the theorem. We shall prove that eventually ${\mathbb P}\!\times\!{\mathbb P}$-almost surely \begin{equation}\label{pizza} -C_1\log n\leq \log({\mathbf W}_n^+(\omega,\omega') {\mathbb P}([\omega_1^n])\leq \log C_1 +\log\log n \end{equation} for some $C_1>0$. It will be clear that by the same reasoning we will also have that eventually ${\mathbb P}\!\times\!{\mathbb P}$-almost surely \begin{equation} -C_2\log n\leq \log({\mathbf W}_n^{-}(\omega,\omega') {\mathbb P}([\omega_n^1])\leq \log C_2 +\log\log n \end{equation} for some $C_2>0$. Putting together these two results immediately gives the statement 2 of the theorem. We first prove the upper bound in \eqref{pizza}. We want to find a summable upper-bound to $$ {\mathbb P}\!\times\!{\mathbb P}\{\log({\mathbf W}_n^+ {\mathbb P}([x_1^n]))> \log t\}= $$ \begin{equation}\label{split} \sum_{x_1^n}{\mathbb P}([x_1^n])\ {\mathbb P}\left\{\log(\T_{x_1^n} {\mathbb P}([x_1^n]))> \log t \right\} \end{equation} where $t$ will be a suitable function of $n$. We apply Key-lemma \ref{MKL} to get for all $t>0$ $$ {\mathbb P}\!\times\!{\mathbb P}\{\log({\mathbf W}_n^+ {\mathbb P}([x_1^n]))> \log t\} \leq C e^{-cn}+ e^{-\rho_1 t}\, . $$ Take $t=t_n= \log n^{\alpha_1}$, $\alpha_1>0$, to get $$ {\mathbb P}\!\times\!{\mathbb P}\{\log({\mathbf W}_n^+ {\mathbb P}([x_1^n]))> \log\log n^\alpha\} \leq C e^{-cn} + \frac{1}{n^{\rho_1 \alpha_1}}\,. $$ By the Borel-Cantelli Lemma we get $$ \log({\mathbf W}_n^+ {\mathbb P}([a_1^n]))\leq \log\log n^{\alpha_1} $$ eventually ${\mathbb P}\!\times\!{\mathbb P}$-almost surely provided that $\alpha_1\rho_1 >1$. To obtain the lower bound in \eqref{pizza}, we have, by Key-lemma \ref{MKL} $$ {\mathbb P}\!\times\!{\mathbb P}\{\log({\mathbf W}^+_n {\mathbb P}([x_1^n]))\leq \log t\} \leq C e^{-cn}+ 1-e^{-\rho_2 t}\leq C e^{-cn}+\rho_2 t $$ for all $t>0$. Choose $t=t_n=n^{-\alpha_2}$, $\alpha_2>1$ and apply the Borel-Cantelli Lemma to get $$ \log({\mathbf W}_n^+ {\mathbb P}([x_1^n]))> -\alpha_2\log n $$ eventually ${\mathbb P}\!\times\!{\mathbb P}$-almost surely. Let us now prove the first statement of the theorem. The proof is very similar except we have to deal with ``bad'' cylinders and use Key-lemma \ref{MKLbis}. We will only establish that eventually ${\mathbb P}$-almost surely the inequality \begin{equation}\label{bof} -C_1\log n\leq \log(\T_n^+(\omega) {\mathbb P}([\omega_1^n])\leq \log C_1 +\log\log n \end{equation} for some $C_1>0$. The analogous inequality for $\T^-(\omega)$ is obtained as above (i.e., using Key-lemma \ref{MKL}). We have the decomposition $$ {\mathbb P}\{\omega:\log(\T_n^+(\omega) {\mathbb P}([\omega_1^n]))> \log t\}= $$ $$ \sum_{k=1}^n \sum_{x_1^n\in \mathcal{S}_k(n)}{\mathbb P}([x_1^n])\ {\mathbb P}\left\{\omega:\log(\T_{x_1^n}(\omega) {\mathbb P}([x_1^n])> \log t \ | \ [x_1^n]\right\} $$ where $\mathcal{S}_k(n)$ is defined just before we state Key-lemma \ref{MKLbis}. For all $t\geq k {\mathbb P}([x_1^n])$ and $n$ large enough, we get using Key-lemma \ref{MKLbis} $$ {\mathbb P}\left\{\log(\T_{x_1^n} {\mathbb P}([x_1^n])> \log t \ | \ [x_1^n])\right\}\leq (\rho_2 + D) e^{-\frac{\rho_1}{2}t} + C e^{-cn} $$ where we used the fact that if $n$ is large enough, $\rho_1/2 \leq \rho_1 - D e^{-c n} \leq \zeta(a_1^n) \leq \rho_2 + D$. We now choose $t=t_n=\log n^{\alpha_1}$, $\alpha_1>0$. If $n$ is large enough, then $t_n \geq k {\mathbb P}([a_1^n])$. This is because we have the uniform estimate ${\mathbb P}([a_1^n])\leq e^{-G n}$, for some $G>0$, since ${\mathbb P}$ is a Gibbs measure. Hence we obtain $$ {\mathbb P}\{\log(\T_n^+ {\mathbb P}([\omega_1^n]))> \log\log n^{\alpha_1}\} \leq \frac{\rho_2 + D}{n^{\alpha_1 \rho_1/2}} + C e^{-cn} $$ which is summable provided that $\alpha_1 \rho_1/2 >1$. The Borel-Cantelli Lemma then gives the upper-bound in \eqref{bof}. The lower-bound is obtained as for the waiting-time estimator but using Key-lemma \ref{MKLbis}. \hfill $\qed$ \subsection{Proof of Theorem \ref{pouac}} Let us prove the second statement of the theorem and that $\lim_{n\rightarrow\infty}\frac{\textup{Var}(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n)}{n}= \sigma^2$. For this it is enough to prove that \begin{equation}\label{bong} \lim_{n\rightarrow\infty}\frac{1}{n}\int \left(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n - {\dot{\mathbf S}}_n \right)^2 d{\mathbb P}\!\times\!{\mathbb P}=0\,. \end{equation} Indeed, proving \eqref{bong} implies, on one hand, that $$ \lim_{n\rightarrow\infty}\frac{\textup{Var}({\dot{\mathbf S}}_n)}{n}=\lim_{n\rightarrow\infty}\frac{\textup{Var}(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n)}{n}\,\cdot $$ On the other hand, it also implies that $(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n-n{\mathbf{MEP}})/\sqrt{n}$ converges in law to the normal $\mathcal{N}(0,\sigma^2)$ if, and only if, $({\dot{\mathbf S}}_n-n{\mathbf{MEP}})/\sqrt{n}$ converges in law to the same law. Now it is obvious from \eqref{porc} that $$ \frac{{\dot{\mathbf S}}_n - \sum_{j=0}^{n-1} [(f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+)\circ\theta_j]}{\sqrt{n}}\to 0\quad {\mathbb P}-\textup{almost-surely}\,. $$ By applying a result of \cite{PP}, we obtain that $$ \frac{\sum_{j=0}^{n-1} [(f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+)\circ\theta_j] - n{\mathbf{MEP}} }{\sqrt{n}}\stackrel{\textup{in law}}{\longrightarrow} \mathcal{N}(0,\sigma^2)\,. $$ Since we have the formula (see \cite{PP}) $$ \sigma^2=\lim_{n\rightarrow\infty} \frac{1}{n}\int \big( \sum_{j=0}^{n-1} [(f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+)\circ\theta_j] - n{\mathbf{MEP}}\big)^2\ d{\mathbb P} $$ it is obvious by \eqref{porc} that $$ \lim_{n\rightarrow\infty}\frac{\textup{Var}({\dot{\mathbf S}}_n)}{n}=\sigma^2\,. $$ Therefore we have reduced the statements of the theorem about $\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n$ to proving \eqref{bong}. By definition we have $$ \int \left(\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n - {\dot{\mathbf S}}_n \right)^2 d{\mathbb P}\!\times\!{\mathbb P}= \sum_{x_{1}^{n}} {\mathbb P}([x_1^n]) \int \left[ \log(\T_{n}^{-}{\mathbb P}([x_n^1]))-\log(\T_{n}^{+}{\mathbb P}([x_1^n])) \right]^2 d{\mathbb P}\ . $$ Let us now prove that the integral in the rhs is bounded above by a positive number independent of $n$, implying immediately \eqref{bong}. To prove this assertion, it is sufficient to prove that \begin{equation}\label{bing} \int \left[\log(\T_{n}^{+}{\mathbb P}([x_1^n]))\right]^2 d{\mathbb P} \leq D_1,\quad \int \left[\log(\T_{n}^{-}{\mathbb P}([x_n^1]))\right]^2 d{\mathbb P} \leq D_2\quad \end{equation} where $D_1, D_2>0$ are independent of $n$. We only prove the first inequality since the other one is proved in exactly the same way. We have the following identities: $$ \int \left[\log(\T_{n}^{+}{\mathbb P}([x_1^n]))\right]^2 d{\mathbb P} = \int_0^\infty {\mathbb P}\left([\log(\T_{n}^{+}{\mathbb P}([x_1^n]))]^2 >t \right) dt = $$ $$ 2 \int_1^\infty {\mathbb P}\left( \T_{n}^{+}{\mathbb P}([x_1^n]) >t \right) \frac{\log t}{t}\ dt + 2 \int_0^1 {\mathbb P}\left( \T_{n}^{+}{\mathbb P}([x_1^n]) <t \right) \frac{-\log t}{t}\ dt= $$ $$ 2 \int_1^\infty {\mathbb P}\left( \T_{x_1^n}{\mathbb P}([x_1^n]) >t \right) \frac{\log t}{t}\ dt + 2 \int_0^1 {\mathbb P}\left( \T_{x_1^n}{\mathbb P}([x_1^n]) <t \right) \frac{-\log t}{t}\ dt =: \textup{I}\, + \, \textup{II}\,. $$ Now we use Key-lemma \ref{MKL} and get $$ {\mathbb P}\left( \T_{x_1^n}{\mathbb P}([x_1^n]) >t \right) \leq (1+C) e^{-\rho_1 t},\;\forall n\geq 1\,. $$ Therefore $$ \textup{I}\leq 2(1+C) \int_1^\infty \frac{\log t}{t} \ e^{-\rho_1 t} \ dt=:D_1'<\infty\,. $$ For the integral II, we have the following estimates $$ \textup{II} = 2 \left(\int_0^{\frac{1}{2}} +\int_{\frac{1}{2}}^1 \right) {\mathbb P}\left( \T_{x_1^n}{\mathbb P}([x_1^n]) <t \right) \frac{-\log t}{t}\ dt \leq $$ $$ \int_0^{\frac{1}{2}}\frac{-\log t}{t}\ (1-e^{-\rho_2 t})\ dt + \int_{\frac{1}{2}}^1 \frac{-\log t}{t}\ dt \leq $$ $$ -\rho_2 \int_0^{\frac{1}{2}} \log t\ dt - \int_{\frac{1}{2}}^1 \frac{\log t}{t}\ dt := D_1''<\infty $$ where we used Lemma \ref{tarte} to bound the first integral. This finishes the proof for the waiting-time estimator. Concerning the hitting-time estimator, we leave the proof to the reader. It is very similar to the previous one except that one has to use Key-lemma \ref{MKLbis}. \hfill $\qed$ \subsection{Proof of Proposition \ref{macbeth}} The proof is an application of G\"artner-Ellis theorem \cite{ellis}. In particular we have to check that the function $p\mapsto \mathcal{E}_U(p)$ is continuously differentiable and strictly convex under assumption \eqref{avion}. The strict convexity follows from the assumption that the process is not reversible. As already mentioned above, this amounts to requiring that $U$ is not physically equivalent to $U^{{\scriptscriptstyle R}}$, i.e., that $f_U^+ - f_{U^{{\scriptscriptstyle R}}}^+$ is not a co-boundary. The open interval $(\underline{q},\overline{q})$ is defined by $\underline{q}=\inf_{q\in\mathbb R}=\lim_{p\to-\infty} {\mathcal E}_U'(p)$ and $\overline{q}=\sup_{q\in\mathbb R}=\lim_{p\to+\infty} {\mathcal E}_U'(p)$. These limits exist by convexity arguments. We refer to \cite{israel} from which one can deduce these classical facts on differentiability and convexity of the pressure function. \hfill $\qed$ \subsection{Proof of Theorem \ref{pouic}} We prove formula \eqref{boulgakov}. We first deal with $0<p<1$. The case $-1<p<0$ is obtained by a similar reasoning, so we omit the proof. The case $p=0$ is trivial. We observe that $$ \mathbb E_{{\mathbb P}\!\times\!{\mathbb P}}\left(e^{p\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n}\right)= \sum_{x_1^n} {\mathbb P}([x_1^n])^{p+1} {\mathbb P}([x_n^1])^{-p}\ \mathbb E_{\mathbb P}\left[\left(\frac{Y_n}{Z_n}\right)^p\right] $$ where $Y_n:=\T_{x_n^1} {\mathbb P}([x_n^1])$, $Z_n:=\T_{x_1^n} {\mathbb P}([x_1^n])$. We then have $$ \mathbb E_{\mathbb P}\left[\left(\frac{Y_n}{Z_n}\right)^p\right] = \int_0^\infty dy \int_0^\infty dz\ \left(\frac{y}{z}\right)^p \ {\mathbb P}\{Y_n\in dy,Z_n\in dz\} $$ \begin{equation} = \int_0^1 dy\int_0^1 dz\ \left(\frac{y}{z}\right)^p \ {\mathbb P}\{Y_n\in dy,Z_n\in dz\} + \int_1^\infty dy \int_1^\infty dz\ \left(\frac{y}{z}\right)^p \ {\mathbb P}\{Y_n\in dy,Z_n\in dz\} \label{cigare} \end{equation} We obtain the obvious upper bound \begin{eqnarray} \nonumber \eqref{cigare} & \leq & \int_0^1 dy\int_0^1 dz\ \frac{1}{z^p} \ {\mathbb P}\{Y_n\in dy,Z_n\in dz\} + \int_1^\infty dy \int_1^\infty dz\ y^p \ {\mathbb P}\{Y_n\in dy,Z_n\in dz\} \\ & \leq & \mathbb E_{{\mathbb P}}\left(\frac{1}{Z_n^p}\right) +\mathbb E_{{\mathbb P}}(Y_n^p)\,. \label{briquet} \end{eqnarray} We get easily the lower bound \begin{eqnarray} \nonumber \eqref{cigare} & \geq & \int_1^\infty dy \int_1^\infty dz\ \frac{1}{z^p} \ {\mathbb P}\{Y_n\in dy,Z_n\in dz\} \\ & \geq & \mathbb E_{{\mathbb P}}\left(\frac{1}{Z_n^p} {{\mathit 1} \!\!\>\!\! I} \{Z_n\geq 1\} \right) \label{allu} \end{eqnarray} where ${{\mathit 1} \!\!\>\!\! I} \{\cdot\}$ denotes the indicator function. Proving Theorem \ref{pouic} for $0<p<1$ is thus reduced to proving that the rhs in \eqref{briquet} is bounded above by a positive number independent of $n$, and that the rhs in \eqref{allu} is bounded below by a positive number independent of $n$. Let us start with an upper bound for $\mathbb E_{{\mathbb P}}(Y_n^p)$. We have $$ \mathbb E_{{\mathbb P}}(Y_n^p)= p \int_0^\infty y^{p-1} {\mathbb P}\{\T_{x_n^1} {\mathbb P}([x_n^1])>y\}\ dy\,. $$ By using Key-lemma \ref{MKL} with $a_1^n=x_n^1$, we obviously have ${\mathbb P}\{\T_{x_n^1} {\mathbb P}([x_n^1])>y\}<A e^{-By}$ for some $A,B>0$. Let us now upper-bound $\mathbb E_{{\mathbb P}}\left(\frac{1}{Z_n^p}\right)$. We have $$ \mathbb E_{{\mathbb P}}\left(\frac{1}{Z_n^p}\right)= |p| \left(\int_0^{\frac{1}{2}}+\int_{\frac{1}{2}}^\infty\right) z^{-|p|-1} {\mathbb P}\{\T_{x_1^n} {\mathbb P}([x_1^n])\leq z\}\ dz\,. $$ The integral from $\frac{1}{2}$ to $\infty$ is bounded above by $\int_{\frac{1}{2}}^\infty z^{-|p|-1} dz<\infty$. To bound the other integral we use Lemma \ref{tarte}: $$ \int_0^{\frac{1}{2}}z^{-|p|-1} {\mathbb P}\{\T_{x_1^n} {\mathbb P}([x_1^n])\leq z\}\ dz\leq \int_0^{\frac{1}{2}} \frac{1-e^{-\rho_2 z}}{z^{|p|+1}} dz <\infty\,. $$ We now estimate from below $\mathbb E_{{\mathbb P}}(Y_n^p {{\mathit 1} \!\!\>\!\! I} \{Y_n\geq 1\})$. We have $$ \mathbb E_{{\mathbb P}}(Y_n^p {{\mathit 1} \!\!\>\!\! I} \{Y_n\geq 1\})= |p| \int_1^\infty y^{-|p|-1} {\mathbb P}\{\T_{x_n^1} {\mathbb P}([x_n^1])\leq y\}\ dy\,. $$ By Key-lemma \ref{MKL} with $a_1^n=x_1^n$ we have $$ {\mathbb P}\{\T_{x_n^1} {\mathbb P}([x_n^1])\leq y\} \geq 1-(1+C e^{-cn}) e^{-\rho_1 y}\,. $$ Observe that $1-(1+C e^{-cn}) e^{-\rho_1 y} \geq 1-(1+C e^{-cn}) e^{-\rho_1}$ for all $y\geq 1$ and for all $n\geq 1$. Therefore $$ \mathbb E_{{\mathbb P}}(Y_n^p {{\mathit 1} \!\!\>\!\! I} \{Y_n\geq 1\}) \geq 1-(1+C e^{-cn}) e^{-\rho_1}>0 $$ provided that $n$ is large enough. Recapitulating, we proved that for all $0<p<1$ and all $n$ large enough $$ E^{-1}\leq \mathbb E_{\mathbb P}\left[\left(\frac{Y_n}{Z_n}\right)^p\right] \leq E $$ for some $E>0$ independent of $n$ and $x_1^n$. Hence, for all $0<p<1$, we get $$ \lim_{n\rightarrow\infty}\frac{1}{n}\log \mathbb E_{{\mathbb P}\!\times\!{\mathbb P}}\left(e^{p\dot{{\mathcal S}}^{{\scriptscriptstyle W}}_n}\right)= \lim_{n\rightarrow\infty}\frac{1}{n}\log \sum_{x_1^n} {\mathbb P}([x_1^n])^{p+1} {\mathbb P}([x_n^1])^{-p}= \mathcal{E}_U(p)\,. $$ The last equality follows obviously from \eqref{gibbsbis} and \eqref{porc}. \bigskip We now turn to the case $|p|\geq 1$. We only deal with the case $p\geq 1$ since the case $p\leq -1$ is obtained by the same reasoning. We have \begin{eqnarray} \mathbb E_{\mathbb P}\left[\left(\frac{Y_n}{Z_n}\right)^p\right] & \geq & p \int_0^1 \frac{1}{y^{p+1}}\ {\mathbb P}\{\T_{x_1^n} {\mathbb P}([x_1^n])\leq y\}\ dy\\ & \geq & p \int_0^1 \frac{1}{y^{p+1}}\ (1-(1+C e^{-cn}) e^{-\rho_1 y})\ dy\\ & = &+ \infty \end{eqnarray} for $n$ large enough and where we used Key-lemma \ref{MKL} to get the second inequality. To prove \eqref{pelleas}, we apply a variant of G\"artner-Ellis theorem found in \cite{PS}. To this end, we use formula \eqref{boulgakov} and the differentiability/convexity properties of the function $p\mapsto {\mathcal E}_U(p)$. We have to restrict to the interval $(c_-,c_+)$ where ${\mathcal W}_U$ is finite and coincides with ${\mathcal E}_U$. \hfill $\qed$
{ "timestamp": "2005-07-11T12:09:00", "yymm": "0503", "arxiv_id": "math-ph/0503071", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503071" }
\section{Introduction} The idea of a generalized complex structure -- a concept which interpolates between complex and symplectic structures -- seems to provide a differential geometric language in which some of the structures of current interest in string theory fit very naturally. There is an associated notion of \emph{generalized K\"ahler manifold} which essentially consists of a pair of commuting generalized complex structures. A remarkable theorem of Gualtieri \cite{Gu} shows that it has an equivalent interpretation in standard geometric terms: a manifold with two complex structures $I_+$ and $I_-$; a metric $g$, Hermitian with respect to both; and connections $\nabla_+$ and $\nabla_-$ compatible with these structures but with skew torsion $db$ and $-db$ respectively for a $2$-form $b$. This so-called \emph{bihermitian structure} appeared in the physics literature as long ago as 1984 \cite{R} as a target space for the supersymmetric $\sigma$-model and in the pure mathematics literature more recently (\cite{AGG} for example) in the context of the integrability of the canonical almost complex structures defined by the Weyl tensor of a Riemannian four-manifold. The theory has suffered from a lack of interesting examples. The first purpose of this paper is to use the generalized complex structure approach to find non-trivial explicit examples on ${\mathbf C}{\rm P}^2$ and ${\mathbf C}{\rm P}^1\times {\mathbf C}{\rm P}^1$. We use an approach to generalized K\"ahler structures of generic type which involves closed $2$-forms satisfying algebraic conditions. This is in principle much easier than trying to write down the differential-geometric data above. What we show is that every $SU(2)$-invariant K\"ahler metric on ${\mathbf C}{\rm P}^2$ or the Hirzebruch surface ${\mathbf F}_2$ generates naturally a generalized K\"ahler structure, where for ${\mathbf F}_2$ (which is diffeomorphic to $S^2\times S^2$), the complex structures $I_+,I_-$ are equivalent to ${\mathbf F}_0={\mathbf C}{\rm P}^1\times {\mathbf C}{\rm P}^1$. The second part of the paper shows that a bihermitian structure on a $4$-manifold (where $I_+$ and $I_-$ define the same orientation) defines naturally a bihermitian structure on the moduli space of solutions to the anti-self-dual Yang-Mills equations, and this gives another (less explicit) source of examples. What appears naturally in approaching these goals is the appearance of holomorphic Poisson structures, and in a way the main point of the paper is to bring this aspect into the foreground. It seems as if this type of differential geometry is related to complex Poisson manifolds in the way in which hyperk\"ahler metrics are adapted to complex symplectic manifolds. Yet our structures are more flexible -- like K\"ahler metrics they can be changed in the neighbourhood of a point. The link with Poisson geometry occurs in three interlinking ways: \begin{itemize} \item a holomorphic Poisson structure defines a particular type of generalized complex structure (see \cite{Gu}), \item the skew form $g([I_+,I_-]X,Y)$ for the bihermitian metric is of type $(2,0)+(0,2)$ and defines a holomorphic Poisson structure for either complex structure $I_+$ or $I_-$ (in the four-dimensional case this was done in \cite{AGG}), \item a generalized complex structure $J:T\oplus T^*\rightarrow T\oplus T^*$ defines by restriction a homomorphism $\pi:T^*\rightarrow T$ which is a real Poisson structure (this has been noted by several authors, see \cite{AB}). \end{itemize} We may also remark that Gualtieri's deformation theorem \cite{Gu} showed that interesting deformations of complex manifolds as generalized complex manifolds require the existence of a holomorphic Poisson structure. We address all three Poisson-related issues in the paper. The starting point for our examples is the generalized complex structure determined by a complex Poisson surface (namely, a surface with an anticanonical divisor) and we solve the equations for a second generalized complex structure which commutes with this one. When we study the moduli space of instantons we show that the holomorphic Poisson structures defined by $g([I_+,I_-]X,Y)$ are the canonical ones studied by Bottacin \cite{Bot1}. Finally we examine the symplectic leaves of the real Poisson structures $\pi_1,\pi_2$ on the moduli space. The structure of the paper is as follows. We begin by studying generalized K\"ahler manifolds as a pair $J_1,J_2$ of commuting generalized complex structures, and we focus in particular on the case where each $J_1,J_2$ is the B-field transform of a symplectic structure -- determined by a closed form $\exp (B+i\omega)$ -- giving a convenient algebraic form for the commuting property. We then implement this to find the two examples. In the next section we introduce the bihermitian interpretation and prove that $g([I_+,I_-]X,Y)$ does actually define a holomorphic Poisson structure. The following sections show how to introduce a bihermitian structure on the moduli space ${\mathcal M}$ of gauge-equivalence classes of solutions to the anti-self-dual Yang-Mills equations. At first glance this seems obvious -- we have two complex structures $I_+,I_-$ on $M$ and hence two complex structures on ${\mathcal M}$, since ${\mathcal M}$ is the moduli space of $I_+$- or $I_-$-stable bundles, and we have a natural ${\mathcal L}^2$ metric. This would be fine for a K\"ahler metric but not in the non-K\"ahler case. Here L\"ubke and Teleman \cite{LT} reveal the correct approach -- one chooses a different horizontal to the gauge orbits in order to define the metric on the quotient. In our case we have two complex structures and two horizontals and much of the manipulation and integration by parts which occurs in this paper is caused by this complication. One aspect we do not get is a natural pair of commuting generalized complex structures on ${\mathcal M}$ -- we obtain the differential geometric data above, and an exact $3$-form $db$, but not a natural choice of $b$. We get a generalized K\"ahler structure only modulo a closed B-field on ${\mathcal M}$. This suggests that ${\mathcal M}$ is not, at least directly, a moduli space of objects defined solely by one of the commuting generalized complex structures on $M$, but there is clearly more to do here. We give finally a quotient construction which also demonstrates the problem of making a generalized K\"ahler structure descend to the quotient. This procedure, analogous to the hyperk\"ahler quotient, could be adapted to yield the bihermitian metric on ${\mathcal M}$ in the case $M$ is a $K3$ or torus. Unfortunately we have not found a quotient construction for the instanton moduli space which works in full generality, but this might be possible by using framings on the anticanonical divisor. \vskip .5cm \noindent{{\bf Acknowledgements:}} The author wishes to thank M. Gualtieri, G. Cavalcanti and V. Apostolov for useful discussions. \section{Generalized K\"ahler manifolds} \subsection{Basic properties} The notion of a generalized K\"ahler structure was introduced by M. Gualtieri in \cite{Gu}, in the context of the generalized complex structures defined by the author in \cite{Hit}. Recall that ``generalized geometry" consists essentially of replacing the tangent bundle $T$ of a manifold by $T\oplus T^*$ with its natural indefinite inner product $$(X+\xi,X+\xi)=-i_X\xi,$$ and the Lie bracket on sections of $T$ by the Courant bracket $$[X+\xi, Y+\eta]= [X,Y]+\mathcal{L}_{X}\eta -\mathcal{L}_{Y}\xi -\frac{1}{2} d(i_{X}\eta -i_{Y}\xi)$$ on sections of $T\oplus T^*$. One then introduces additional structures on $T\oplus T^*$ compatible with these. A \emph{generalized complex structure} is a complex structure $J$ on $T\oplus T^*$ such that $J$ is orthogonal with respect to the inner product and with the integrability condition that if $A,B$ are sections of $(T\oplus T^*)\otimes \mathbf{C}$ with $JA=iA,JB=iB$, then $J[A,B]=i[A,B]$ (using the Courant bracket). The standard examples are a complex manifold where $$J_1=\pmatrix {I&0\cr 0& -I}$$ and a symplectic manifold where $$J_2=\pmatrix {0&-\omega^{-1}\cr \omega & 0}.$$ The $+i$ eigenspace of $J$ is spanned by $\{\dots, \partial/\partial z_j\dots, \dots, d\bar z_k,\dots\}$ in the first case and $\{\dots, \partial/\partial x_j- i \sum\omega_{jk}dx_k,\dots\}$ in the second. Another example of a generalized complex manifold is a holomorphic Poisson manifold -- a complex manifold with a holomorphic bivector field $$\sigma=\sum\sigma^{ij}\frac{\partial}{\partial z_i}\wedge\frac{\partial}{\partial z_j}$$ satisfying the condition $[\sigma,\sigma]=0$, using the Schouten bracket. This defines a generalized complex structure where the $+i$ eigenspace is $$ E=\left[\dots, \frac{\partial}{\partial z_j},\dots, d\bar z_k+\sum_{\ell}\bar\sigma^{k\ell}\frac{\partial}{\partial \bar z_{\ell}},\dots \right],$$ and if $\sigma=0$ this gives a complex structure. Gualtieri observed (see also \cite{AB}) that the real bivector defined by the upper triangular part of $J:T\oplus T^*\rightarrow T\oplus T^*$ is always a real Poisson structure. In the symplectic case this is the canonical Poisson structure and in the complex case it is zero. Both facts show that Poisson geometry plays a central role in this area, a feature we shall see more of later. \vskip .25cm The algebraic compatibility condition between $\omega$ and $I$ to give a K\"ahler manifold (i.e. that $\omega$ be of type $(1,1)$) can be expressed as $J_1J_2=J_2J_1$ and this is the basis of the definition of a \emph{generalized} K\"ahler structure: \begin{definition} \label{GKdef} A \emph{generalized K\"ahler structure} on a manifold consists of two commuting generalized complex structures $J_1,J_2$ such that the quadratic form $(J_1J_2A,A)$ on $T\oplus T^*$ is definite. \end{definition} \vskip .25cm At a point, a generalized complex structure can also be described by a form $\rho$: the $+i$ eigenspace bundle $E$ consists of the $A=X+\xi\in (T\oplus T^*)\otimes \mathbf{C}$ which satisfy $A\cdot \rho=i_X\rho+\xi\wedge\rho=0$. For the symplectic structure $\rho=\exp i\omega$, and for a complex structure with complex coordinates $z_1,\dots,z_n$ we take the $n$-form $\rho=dz_1\wedge dz_2\wedge \dots\wedge dz_n$. The structure is called even or odd according to whether $\rho$ is an even or odd form. The generic even case is the so-called B-field transform of a symplectic structure where $$\rho=\exp \beta = \exp (B+i\omega)$$ and $B$ is an arbitrary $2$-form. The generalized complex structure defined by a holomorphic Poisson structure $\sigma$ is of this type if $\sigma$ defines a non-degenerate skew form on $(T^*)^{1,0}$; then $B+i\omega$ is its inverse. If $\rho$ extends smoothly to a neighbourhood of the point, and is {\it closed}, then the integrability condition for a generalized complex structure holds. \vskip .25cm The following lemma is useful for finding generalized K\"ahler structures where both are of this generic even type (which requires the dimension of $M$ to be of the form $4k$). We shall return to this case periodically to see how the various structures emerge concretely. \begin{lem} \label{commute} Let $\rho_1=\exp \beta_1, \rho_2=\exp\beta_2$ be closed forms defining generalized complex structures $J_1,J_2$ on a manifold of dimension $4k$. Suppose that $$(\beta_1-\beta_2)^{k+1}=0=(\beta_1-\bar\beta_2)^{k+1}$$ and $(\beta_1-\beta_2)^{k}$ and $(\beta_1-\bar\beta_2)^{k}$ are non-vanishing. Then $J_1$ and $J_2$ commute. \end{lem} \begin{lemprf} Suppose that $(\beta_1-\beta_2)^{k+1}=0$ and $(\beta_1-\beta_2)^{k}$ is non-zero. Then the $2$-form $\beta_1-\beta_2$ has rank $2k$, i.e. the dimension of the space of vectors $X$ satisfying $i_X(\beta_1-\beta_2)=0$ is $2k$. Since $i_X 1+\xi\wedge 1=0$ if and only if $\xi=0$, this means that the space of solutions $A=X+\xi$ to $$A\cdot \exp (\beta_1-\beta_2)=0=A\cdot 1$$ is $2k$-dimensional. Applying the invertible map $\exp \beta_2$, the same is true of solutions to $$A\cdot \exp \beta_1=0=A\cdot \exp \beta_2.$$ This is the intersection $E_1\cap E_2$ of the two $+i$ eigenspaces. Repeating for $\beta_1-\bar\beta_2$ we get $E_1\cap \bar E_2$ to be $2k$-dimensional. These two bundles are common eigenspaces of $(J_1, J_2)$ corresponding to the eigenvalues $(i,i)$ and $(i,-i)$ respectively. Together with their conjugates they decompose $(T\oplus T^*)\otimes \mathbf{C}$ into a direct sum of common eigenspaces of $J_1,J_2$, thus $J_1J_2=J_2J_1$ on every element. \end{lemprf} We also need to address the definiteness of $(J_1J_2 A,A)$ in Definition 1. Let $V_+$ be the $-1$ eigenspace of $J_1J_2$ (the notation signifies $J_1=+J_2$ on $V_+$). This is $$E_1\cap E_2\oplus \bar E_1\cap \bar E_2.$$ If $X$ is a vector in the $2k$-dimensional space defined by $i_X(\beta_1-\beta_2)=0$ then $A=X-i_X\beta_2$ satisfies $A\cdot \exp \beta_1=0=A\cdot \exp \beta_2$, i.e. $A\in E_1\cap E_2$. But then \begin{equation} (A+\bar A,A+\bar A)=i_X\beta_2(\bar X)+i_{\bar X}\bar\beta_2(X)=(\beta_2-\bar\beta_2)(X,\bar X) \label{posit} \end{equation} so we need to have this form to be definite. Note that interchanging the roles of $\beta_1,\beta_2$, this is the same as $(\beta_1-\bar\beta_1)(X,\bar X)$ being definite. \subsection{Hyperk\"ahler examples} A hyperk\"ahler manifold $M$ of dimension $4k$ provides a simple example of a generalized K\"ahler manifold. Let $\omega_1,\omega_2,\omega_3$ be the three K\"ahler forms corresponding to the complex structures $I,J,K$ and set $$\beta_1=\omega_1+\frac{i}{2}(\omega_2-\omega_3),\quad \beta_2=\frac{i}{2}(\omega_2+\omega_3).$$ Then $\beta_1-\beta_2=\omega_1-i\omega_3$ is a $J$-holomorphic symplectic $2$-form and so clearly satisfies the conditions of Lemma \ref{commute}. Similarly $\beta_1-\bar\beta_2= \omega_1+i\omega_2$ is holomorphic symplectic for $K$. The vectors $X$ satisfying $i_X(\beta_1-\beta_2)=0$ are the $(0,1)$ vectors for $J$, and $\beta_2-\bar\beta_2=i(\omega_2+\omega_3)$ whose $(1,1)$ part with respect to $J$ is $i\omega_2$. Thus $$(\beta_2-\bar\beta_2)(X,\bar X)=i\omega_2(X,\bar X)$$ which is positive definite. Thus a hyperk\"ahler manifold satisfies all the conditions to be generalized K\"ahler. \vskip .25cm D. Joyce observed (see \cite{AGG}) that one can deform this example. Let $f$ be a smooth real function on $M$, and use the symplectic form $\omega_1$ to define a Hamiltonian vector field. Now integrate it to a one-parameter group of symplectic diffeomorphisms $F_t:M\rightarrow M$, so that $F_t^*\omega_1=\omega_1$. Define $$\beta_1=\omega_1+\frac{i}{2}(\omega_2-F_t^*\omega_3),\quad \beta_2=\frac{i}{2}(\omega_2+F_t^*\omega_3),$$ and then $$\beta_1-\beta_2=\omega_1-iF^*_t\omega_3=F^*_t(\omega_1-i\omega_3).$$ This is just the pull-back by a diffeomorphism of $\omega_1-i\omega_3$ so also satisfies the constraint of Lemma \ref{commute}. We also have $\beta_1-\bar\beta_2= \omega_1+i\omega_2$ which is just the same as the hyperk\"ahler case, so both constraints hold. If $t$ is sufficiently small this will still give a positive definite metric. This simple example at least shows the flexibility of the concept -- we can find a new structure from an arbitrary real function, somewhat analogous to the addition of $\partial\bar\partial f$ to a K\"ahler form. In the compact four-dimensional situation this type of structure restricts us to tori and K3 surfaces. We give next an explicit example on the projective plane. \subsection {Example: the projective plane} \label{cp2} The standard $SU(2)$ action on $\mathbf{C}^2$ extends to ${\mathbf C}{\rm P}^2$ and the invariant $2$-form $dz_1\wedge dz_2$ extends to a meromorphic form with a triple pole on the line at infinity. Its inverse $\partial/\partial z_1\wedge \partial/\partial z_2$ is a holomorphic Poisson structure with a triple zero on the line at infinity. We shall take the generalized complex structure $J_1$ to be defined by this, and seek an $SU(2)$-invariant generalized complex structure $J_2$ defined by $\exp (B+i\omega)$ in such a way that the pair define a generalized K\"ahler structure. On $\mathbf{C}^2$ the Poisson structure is non-degenerate, so the generalized complex structure on that open set is defined by the closed form $\rho_1=\exp dz_1dz_2$. \vskip .25cm We begin by parametrizing $\mathbf{C}^2\setminus \{0 \}$ by $\mathbf{R}^+ \times SU(2)$: $$\pmatrix {z_1\cr z_2}=\pmatrix {z_1& -\bar z_2\cr z_2 & \bar z_1}\pmatrix {1\cr 0}=rA\pmatrix {1\cr 0}.$$ Then, with the left action, the entries of $A^{-1}dA=A^*dA$ are invariant $1$-forms. We calculate $$A^*dA=-\frac{dr}{r}I+\frac{1}{r^2}\pmatrix {\bar z_1 dz_1+\bar z_2 dz_2& -\bar z_1 d\bar z_2+\bar z_2 d\bar z_1\cr z_1 d z_2- z_2 d z_1 & z_1 d\bar z_1+ z_2 d \bar z_2}=\pmatrix {i\sigma_1& -\sigma_2-i\sigma_3\cr \sigma_2+i\sigma_3 & -i\sigma_1}$$ where \begin{eqnarray*} v_1&=&r^{-1}dr+i\sigma_1=(\bar z_1 dz_1+\bar z_2 dz_2)/r^2\\ v_2&=&\sigma_2+i\sigma_3= (z_1 d z_2- z_2 d z_1)/r^2 \end{eqnarray*} and these give a basis for the $(1,0)$-forms. We see that $2dr=r(v_1+\bar v_1)$ so that $$\partial r=rv_1/2,\quad \bar\partial r=r\bar v_1/2$$ and hence $$\partial v_1=0,\quad \bar\partial v_1= dv_1=id\sigma_1=2i\sigma_2\sigma_3=-v_2\bar v_2.$$ Furthermore $$\partial v_2=v_1v_2,\quad \bar\partial v_2=-\bar v_1 v_2, \quad \bar\partial(v_1v_2)=v_1\bar v_1v_2.$$ \vskip .25cm We look for invariant solutions to the generalized K\"ahler equations where $$\rho_1=\exp \beta_1=\exp[dz_1dz_2]=\exp[r^2v_1v_2]$$ and $\rho_2=\exp \beta_2$ where $$\beta_2=\sum_{i,j} H_{ij}v_i\bar v_j+ \lambda v_1v_2+\mu \bar v_1\bar v_2$$ (with $H_{ij},\lambda$ and $\mu$ functions of $r$) is a general invariant $2$-form. The algebraic compatibility conditions from Lemma 1 are: $$(\beta_2-\beta_1)^2=0=(\beta_2-\bar\beta_1)^2$$ which gives on subtraction $$\beta_2(v_1v_2-\bar v_1\bar v_2)=0$$ or equivalently $\lambda=\mu$. We then get $$0=\beta_2^2-2\beta_2\beta_1=(\sum H_{ij}v_i\bar v_j)^2+2\lambda^2v_1v_2\bar v_1\bar v_2-2\lambda r^2v_1v_2\bar v_1\bar v_2$$ or equivalently \begin{equation} \det H=\lambda(\lambda-r^2) \label{detH} \end{equation} \vskip .5cm We also know that $d\beta_2=0$ so that \begin{eqnarray*} \bar\partial (H_{ij}v_i\bar v_j)+\partial \lambda \bar v_1\bar v_2+\lambda \partial (\bar v_1\bar v_2)&=&0\\ \partial (H_{ij}v_i\bar v_j)+\bar\partial \lambda v_1 v_2+\lambda \bar\partial (v_1v_2)&=&0 \end{eqnarray*} But $H$ and $\lambda$ are functions of $r$ and so from the first equation, expanding and collecting terms in $v_1\bar v_1\bar v_2$ we obtain \begin{equation} rH_{12}'+2H_{12}=r\lambda'-2\lambda \label{H12} \end{equation} while collecting terms in $\bar v_1 v_2\bar v_2$ yields \begin{equation} rH_{22}'=2H_{11} \label{H22} \end{equation} The second equation gives (\ref{H22}) again and also \begin{equation} rH_{21}'+2H_{21}=-r\lambda'+2\lambda \label{H21} \end{equation} We can solve these by quadratures: from (\ref{H12}) we get \begin{eqnarray*} r^2H_{12}&=&\int^r (s^2\lambda' -2s\lambda)ds=r^2\lambda-4\int_a^r s\lambda ds\\ r^2H_{21}&=&-r^2\lambda+4\int_{a'}^r s\lambda ds. \end{eqnarray*} If we set $$L(r)=\int_a^r s\lambda ds$$ then $\lambda=L'/r$ and $$r^2H_{12}=rL'-4L,\quad r^2H_{21}= -rL'+4L+b$$ and then $\det H=\lambda(\lambda-r^2)$ gives $$H_{11}H_{22}=8\frac{LL'}{r^3}-16\frac{L^2}{r^4}-L'r+b\frac{L'}{r^3}-4b\frac{L}{r^4}.$$ Substituting $rH_{22}'=2H_{11}$ from (\ref{H22}) and integrating by parts leads to \begin{equation} H_{22}^2=16\frac{L^2}{r^4}+4b\frac{L}{r^4}-4L+c \label{Hform} \end{equation} Thus an arbitrary complex function $L$ and three constants of integration $a,b,c$ give the general solution to the equations. Note for comparison that an $SU(2)$-invariant K\"ahler metric involves one \emph{real} function of $r$ -- the invariant K\"ahler potential. \vskip .25cm There is a lot of choice here but to produce an example let us take for simplicity $a=a'=0$ so that $H_{12}=-H_{21}$ and therefore $b=0$, and take $c=0$ so that \begin{equation} H_{22}^2=16\frac{L^2}{r^4}-4L \label{examp1} \end{equation} Let $L$ be real, then so is $\lambda$ and $H_{12}$. If $L$ negative then $H_{22}^2$ is positive from (\ref{examp1}). This means that $H_{22}$ is real, and hence from (\ref{H22}) so is $H_{11}$. Choose the positive square root for $H_{22}$. Now $\beta_2=\sum H_{ij}v_i\bar v_j+ \lambda v_1v_2+\mu \bar v_1\bar v_2$ and $\lambda$ and $H_{ij}$ are real and $H_{12}=-H_{21}$ so \begin{equation} \beta_2-\bar\beta_2= 2(H_{11}v_1\bar v_1+H_{22}v_2\bar v_2) \label{ibeta2} \end{equation} and for this to be symplectic $H_{11}$ and $H_{22}$ must be non-zero. To get a generalized K\"ahler metric we need from (\ref{posit}) to have $(\beta_1-\bar\beta_1)(X,\bar X)$ definite on the space of vectors $X$ with $i_X(\beta_1-\beta_2)=0.$ If $\nu_1,\nu_2,\bar\nu_1,\bar\nu_2$ is the dual basis to $v_1,v_2,\bar v_1,\bar v_2$ then $X$ must be a linear combination of \begin{equation} \lambda\nu_1-H_{12}\bar\nu_1+H_{11}\bar \nu_2,\quad \lambda\nu_2-H_{22}\bar\nu_1-H_{12}\bar\nu_2. \label{CPhol} \end{equation} Since $\beta_1-\bar\beta_1=r^2(v_1v_2-\bar v_1\bar v_2)$ this gives $(\beta_1-\bar\beta_1)(X,\bar X)$ relative to this basis as the Hermitian form $$\pmatrix{2r^2\lambda H_{11}& \cr & 2r^2\lambda H_{22}}$$ so we also need $H_{11}$ to be positive. \vskip .25cm Notice now the point we have reached: $H_{11}$ and $H_{22}$ must be positive, which means that \begin{equation} H_{11}v_1\bar v_1+H_{22}v_2\bar v_2 \label{kform} \end{equation} is a positive definite Hermitian form. Moreover $rH_{22}'=2H_{11}$, and this implies that the form is K\"ahler. In fact if $\phi(r)$ satisfies $H_{22}=r\phi'/2$, this is the K\"ahler metric $i\partial\bar\partial \phi$, with $\phi$ as a K\"ahler potential. Thus each $SU(2)$-invariant K\"ahler metric defines canonically, through the functions $L,\lambda$ and $H_{12}$ defined in terms of $H_{22}$, an $SU(2)$-invariant generalized K\"ahler metric on $\mathbf{C}^2\setminus \{0\}$. \begin{prp} If the K\"ahler metric (\ref{kform}) extends to ${\mathbf C}{\rm P}^2$, so does the generalized K\"ahler structure. \end{prp} \begin{prf} Since $\beta_1^{-1}$ is a global holomorphic Poisson structure on ${\mathbf C}{\rm P}^2$, we know that the generalized complex structure $J_1$ extends to the whole of ${\mathbf C}{\rm P}^2$, so we only need to check that $\beta_2$ also extends. We begin at $r=0$, the origin in $\mathbf{C}^2$. Clearly $r^2=z_1\bar z_1+z_2\bar z_2$ is smooth on $\mathbf{C}^2$. We shall use the fact that if $f(r)$ extends to a smooth function on a neighbourhood of the origin in $\mathbf{C}^2$ then $f(r)=f(0)+r^2f_1(r)$ where $f_1$ is also a smooth function. If $g$ is the K\"ahler metric and $X=r\partial/\partial r$ the Euler vector field, then $g(X,X)=H_{11}$ is smooth on $\mathbf{C}^2$ and vanishes at the origin so $H_{11}=r^2f_1$ for smooth $f_1>0$. The volume form of $g$ is $r^{-1}H_{11}H_{22}dr\sigma_1\sigma_2\sigma_3$ and comparing with the Euclidean volume $r^3dr\sigma_1\sigma_2\sigma_3$ we see that $H_{22}=r^2f_2$ for $f_2>0$ smooth. Equation (\ref{examp1}) gives $$L=\frac{r^4}{8}\left[1-\sqrt{1+(4H_{22}^2/r^4)}\right]=\frac{r^4}{8}\left[1-\sqrt{1+4f_2^2}\right]$$ and so $L=r^4 f_3$ for $f_3$ smooth. By definition, $\lambda=L'/r=4r^2f_3+r^3f_3'=r^2f_4$ since for any smooth $f(r)$ $$rf'=\sum x_i\frac{\partial f}{\partial x_i}$$ which is smooth. (In fact since this expression also vanishes at $0$ we have $rf'=r^2g$ for $g$ smooth.) Since $r^2v_1v_2=dz_1dz_2$, this shows that the term $\lambda (v_1v_2+ \bar v_1\bar v_2)$ is smooth. Now $r^2H_{12}=rL'-4L=r^5f_3'$ so $H_{12}=r^2(rf_3')=r^4f_5$ for smooth $f_5$, which means that $H_{12}v_1\bar v_2$ and $H_{21}v_2\bar v_1$ are smooth since $r^2v_1=\bar z_1 dz_1+\bar z_2 dz_2, r^2v_2= z_1 d z_2- z_2 d z_1$. Hence the form $\beta_2$ is smooth at the origin. From (\ref{ibeta2}) the imaginary part of $\beta_2$ is nondegenerate at the origin since the K\"ahler metric is. \vskip .25cm As $r\rightarrow \infty$ we need to take homogeneous coordinates on ${\mathbf C}{\rm P}^2$ so that $\mathbf{C}^2$ is parametrized by $[1,z_1,z_2]=[1/z_1,1,z_2/z_1]$, so we use local affine coordinates $w_1,w_2$ where for $z_1\ne 0$, $$w_1=\frac{1}{z_1},\quad w_2=\frac{z_2}{z_1}.$$ The projective line at infinity is then $w_1=0$. In these coordinates we have $$r^2=\frac{1+\vert w_2\vert^2}{\vert w_1\vert^2}$$ so $1/r^2$ is smooth and \begin{equation} v_1=\frac{\bar w_2dw_2}{1+\vert w_2\vert^2}-\frac{dw_1}{w_1},\quad v_2=\frac{\bar w_1dw_2}{w_1(1+\vert w_2\vert^2)} \label{v1v2} \end{equation} Note here that $$\frac{1}{r^2}v_1=\frac{\vert w_1\vert^2\bar w_2 dw_2}{(1+\vert w_2\vert^2)^2}-\frac{\bar w_1dw_1}{1+\vert w_2\vert^2}$$ is smooth at $r=\infty$, and similarly $v_2/r^2, v_1\bar v_2/r^2$ are smooth. The coefficient of $dw_1d\bar w_1$ in $H_{11}v_1\bar v_1+H_{22}v_2\bar v_2$ is $H_{11}/\vert w_1\vert^2$ so this is smooth and hence $r^2H_{11}=g_1$, a smooth function. Considering the coefficient of $dw_2d\bar w_2$ we see that $H_{22}$ is smooth. Now \begin{equation} L=\frac{r^4}{8}\left[1-\sqrt{1+4H_{22}^2/r^4}\right]=-\frac{1}{2}\frac{H_{22}^2}{1+\sqrt{1+4H_{22}^2/r^4}} \label{ell} \end{equation} which is smooth and $g_1=r^2H_{11}=r^3H_{22}'/2$ so that differentiating (\ref{ell}) $\lambda={L'}/{r}=g_2/r^4$ where $g_2$ is smooth. This means from (\ref{v1v2}) that $\lambda(v_1\bar v_1+v_2\bar v_2)$ is smooth. Finally $H_{12}=\lambda-{4}L/{r^2}=g_3/r^2$ where $g_3$ is smooth, and so $H_{12}v_1\bar v_2$ is smooth. Thus $\beta_2$ extends as $r\rightarrow \infty$. The argument for $z_2\ne 0$ is the similar. \end{prf} \subsection{Example: the Hirzebruch surface ${\mathbf F_2}$}\label{f2} We can apply the above formalism with different boundary conditions to the Hirzebruch surface ${\mathbf F}_2$. Recall that this is $${\mathbf F}_2=P({\mathcal O}\oplus {\mathcal O}(-2))=P({\mathcal O}\oplus K)$$ since the canonical bundle $K$ of ${\mathbf C}{\rm P}^1$ is ${\mathcal O}(-2)$. The canonical symplectic form on $K$ extends to a meromorphic form $\beta_1$ on ${\mathbf F}_2$, and its inverse, a Poisson structure, defines the generalized complex structure $J_1$. \vskip .25cm On $K$ we take local coordinates $(w,z)\mapsto wdz$ where $z$ is an affine coordinate on ${\mathbf C}{\rm P}^1$. Then for each quadratic polynomial $q(z)$ $$q(z)\frac{d}{dz}$$ is a global holomorphic vector field on ${\mathbf C}{\rm P}^1$ so that $$(w,z)\mapsto (w,wz,wz^2)$$ is a well defined map from $K$ to the cone $x_2^2=x_1x_3$ in $\mathbf{C}^3$. The map $$(z_1,z_2)\mapsto (z_1^2,z_1z_2,z_2^2)$$ maps the quotient $\mathbf{C}^2/\pm 1$ isomorphically to this cone and the Hirzebruch surface is a compactification of the surface obtained by resolving the singularity at the origin of this cone. Our ansatz above for $\mathbf{C}^2\setminus\{0\}$ extends to the quotient which is $\mathbf{R}^+\times SO(3)$ since we were using left-invariant forms. We need to adapt in a different way to extend at $r\rightarrow 0$ which is a rational curve of self-intersection $-2$ and $r\rightarrow \infty$, a rational curve of self-intersection $+2$. \vskip .25cm To proceed as $r\rightarrow 0$ we change coordinates from $z_1,z_2$ to $w,z$: $$w=z_1^2,\qquad z=z_2/z_1.$$ Then $dzdw=2dz_2dz_1$, so here we see that the standard $2$-form on $\mathbf{C}^2$ is a multiple of the canonical symplectic form on the holomorphic cotangent bundle. We find $$r^2=\vert w\vert(1+\vert z\vert^2)$$ so in particular $r^4$ is smooth. Furthermore \begin{equation} 2\frac{dr}{r}=\frac{1}{2}\left[\frac{dw}{w}+\frac{d\bar w}{\bar w}\right]+\frac{d(z\bar z)}{1+\vert z\vert^2} \label{diffr} \end{equation} We also calculate \begin{equation} v_1=\frac{dw}{2w}+\frac{\bar z dz}{1+\vert z\vert^2},\quad v_2=\frac{wdz}{\vert w\vert (1+\vert z\vert^2)} \label{v12} \end{equation} Thus $r^2v_1v_2$ and $r^4v_1\bar v_1$ are smooth, and $$v_2\bar v_2=\frac{dz d\bar z}{(1+\vert z\vert^2)^2}$$ which is smooth. \vskip .25cm Suppose in this case that $H_{11}v_1\bar v_1+H_{22}v_2\bar v_2$ extends as a K\"ahler form. Then considering the coefficient of $dwd\bar w$, $H_{11}=r^4 f_1$ where $f_1>0$ is smooth and $H_{22}$ itself is smooth and positive. The reality conditions on $H_{ij}$ are the same as the ${\mathbf C}{\rm P}^2$ case and the constants of integration $a,a',b$ vanish as before but we now take $c$ in (\ref{Hform}) to be the limiting value $H_{22}^2(0)$. Since $H_{22}'>0$, $H_{22}^2-c>0$. From (\ref{Hform}) we obtain \begin{equation} L=\frac{r^4}{8}\left[1-\sqrt{1+4(H_{22}^2-c)/r^4}\right]. \label{Lformula} \end{equation} We now use the familiar formula for a smooth function $f$ \begin{equation} f(x)-f(x_0)=\sum_i(x-x_0)_i\int_0^1\frac{\partial f}{\partial x_i}(x_0+t(x-x_0))dt \label{expand} \end{equation} where the coordinates $x_i$ are the real and imaginary parts of $z,w$ and we take $x_0=(z_0,0)$. From (\ref{diffr}) we calculate the derivatives $$\frac{\partial H_{22}}{\partial w}=\frac{r}{4w}H_{22}'=\frac{1}{2w}H_{11}=\frac{r^4 f_1}{2w}=\frac{1}{2}\bar w(1+\vert z\vert^2)^2f_1$$ (since $rH_{22}'=2H_{11}$ and $H_{11}=r^4f_1$ for smooth $f_1$) and $$\frac{\partial H_{22}}{\partial z}=\frac{r\bar z}{2(1+\vert z\vert^2)}H_{22}'=\frac{w\bar w f_1}{1+\vert z\vert^2}.$$ Putting these and their conjugates into (\ref{expand}) with $f=H_{22}$ we see that $H_{22}(x)-H_{22}(x_0)=w\bar w f_2$ for a smooth function $f_2$ and hence from the formula for $L$ above $L=r^4f_3$ where $f_3$ is smooth. This gives $$\lambda=\frac{L'}{r}=4r^2f_3+4r^2w\frac{\partial f_3}{\partial w}.$$ This is $r^2f_4$ where $f_4$ is smooth and so $\lambda(v_1v_2+\bar v_1\bar v_2)$ is smooth since $r^2v_1 v_2$ is smooth. Now \begin{equation} H_{12}=\frac{L'}{r}-4\frac{L}{r^2}=r^3f_3'=4wr^2\frac{\partial f_3}{\partial w}=4w\vert w\vert (1+\vert z\vert^2)\frac{\partial f_3}{\partial w} \label{h12} \end{equation} From (\ref{v12}) we see that $H_{12}v_1\bar v_2$ is smooth. \vskip .25cm In a neighbourhood of the curve $r=\infty$ we have coordinates $w'=1/w, z'=z$ and the calculations are very similar. In particular $1/r^4$ is smooth and $H_{22}$ is smooth and nonzero at infinity. Let $c'=\lim_{r\rightarrow\infty}H_{22}^2$. Then from (\ref{Lformula}) we have $$L=-\frac{1}{4}(c'-c)+\frac{1}{r^4}g$$ where $g$ is smooth. This gives the required behaviour of $L$ and $\lambda$ for $\beta_2$ to extend to the curve at infinity. \section{Bihermitian metrics} \subsection{Generalized K\"ahler and bihermitian structures} The generalized K\"ahler structures described above have a very concrete Riemannian description, owing to the following remarkable theorem of Gualtieri \cite{Gu}: \begin{thm} \label{bi} A generalized K\"ahler structure on a manifold $M^{2m}$ is equivalent to: \begin{itemize} \item a Riemannian metric $g$ \item two integrable complex structures $I_+,I_-$ compatible with the metric \item a $2$-form $b$ such that $d^c_-\omega_-=db=-d^c_+\omega_+$ \end{itemize} where $\omega_+,\omega_-$ are the two hermitian forms and $d^c=I^{-1}dI=i(\bar\partial-\partial)$. \end{thm} An equivalent description is to say that there are two connections $\nabla^+,\nabla^-$ which preserve the metric and the complex structures $I_+,I_-$ respectively and these are related to the Levi-Civita connection $\nabla$ by \begin{equation} \nabla^{\pm}=\nabla\pm\frac{1}{2}g^{-1}h \label{del} \end{equation} where $h=db$ is of type $(2,1)+(1,2)$ with respect to both complex structures. In the K\"ahler case $I_+=I,I_-=-I$ and $b=0$. This is the geometry introduced $20$ years ago in the physics literature \cite{R} and more recently studied by differential geometers in four dimensions as ``bihermitian metrics", as in \cite{AGG}. \vskip.25cm Following \cite{Gu}, to derive this data from the generalized K\"ahler structure one looks at the eigenspaces of $J_1J_2$. Since $J_1$ and $J_2$ commute, $(J_1J_2)^2=(-1)^2=1$. As before we choose $V_+$ to be the subbundle where $J_1=J_2$ and $V_-$ where $J_1=-J_2$. If the quadratic form $(J_1J_2A,A)$ is negative definite, the natural inner product on $T\oplus T^*$ is positive definite on $V_+$, and negative definite on the complementary eigenspace $V_-$. Since the signature of the quadratic form is $(2m,2m)$ each such space is $2m$-dimensional. Moreover since $T$ and $T^*$ are isotropic, $V_+\cap T=0=V_+\cap T^*$ and so $V_+$ is the graph of an invertible map from $T$ to $T^*$, i.e. a section $g+b$ of $T^*\otimes T^*$, where $g$ is the symmetric part and $b$ the skew-symmetric part. The bundle $V_+$ is preserved by $J_1$ and identified with $T$ by projection, and hence $J_1$ (or equivalently $J_2$) induces a complex structure $I_+$. Similarly on $V_-$, $J_1$ or $-J_2$ gives $T$ the complex structure $I_-$. Conversely, as Gualtieri shows, given the bihermitian data above, the two commuting generalized complex structures are defined by \begin{equation} J_{1/2}=\frac{1}{2}\pmatrix{1&0\cr b&1}\pmatrix{I_+\pm I_- & -(\omega_+^{-1}\pm\omega_-^{-1})\cr \omega_+\pm \omega_- & -(I_+^*\pm I_-^*)}\pmatrix{1&0\cr -b&1} \label{J12} \end{equation} \vskip .25cm Our standard examples are constructed from closed forms $\rho_1=\exp\beta_1,\rho_2=\exp\beta_2$, so we look next at how the bihermitian structure is encoded in these. The identification of $T$ with $V_+$ can be written as $X\mapsto X+(g(X,-)+b(X,-))$. If $X$ is a $(1,0)$-vector with respect to $I_+$ then this is $$X+\xi= X+i_X(b- i\omega_+)$$ where $\omega_+$ is the Hermitian form for $I_+$. If this lies in $E_1$, it annihilates $\exp \beta_1$, so $i_X\beta_1+ i_X(b- i\omega_+)=0$. Thus $\beta_1+b-i\omega_+$ is of type $(0,2)$, and similarly for $E_2$. Thus there are $(0,2)$-forms $\gamma_1,\gamma_2$ such that $$\beta_1=-b+i\omega_+ +\gamma_1,\quad \beta_2=-b+i\omega_+ +\gamma_2.$$ Since $\beta_1,\beta_2$ are closed this means that $\gamma=\bar\beta_1-\bar\beta_2=\bar\gamma_1-\bar\gamma_2$ is a holomorphic $(2,0)$-form with respect to $I_+$. The form $(\beta_1-\beta_2)$ defines the complex structure $I_+$ -- the $(1,0)$ vectors are the solutions to $i_X(\beta_1-\beta_2)=0$ and the metric on such $(1,0)$-vectors is given by $(\beta_1-\bar\beta_1)(X,\bar X)$. Changing to $V_-$, the identification with $T$ is $X\mapsto X+(-g(X,-)+b(X,-))$ and then $\beta_1=-b+i\omega_- +\delta_1,\quad \bar\beta_2=-b+i\omega_- +\delta_2.$ where $\delta_1,\delta_2$ are $(0,2)$-forms with respect to $I_-$. \vskip .25cm In four real dimensions we now give the precise relationship between the bihermitian description and the generalized K\"ahler one. First note that $\omega_-^{1,1}$ is self-dual and type $(1,1)$ so there is a real smooth function $p$ such that $$\omega_-^{1,1}=p\omega_+.$$ Moreover since $\omega_+^2=\omega_-^2$, $\vert p\vert \le 1$. From above we have \begin{eqnarray*} \beta_1&=&-b+i\omega_++\gamma_1=-b+i\omega_-+\delta_1\\ \beta_2&=&-b+i\omega_++\gamma_2=-b-i\omega_-+\bar\delta_2 \end{eqnarray*} where $\gamma_1,\gamma_2$ are $(0,2)$ with respect to $I_+$ and $\delta_1,\delta_2$ are $(0,2)$ with respect to $I_-$. We let $\bar\gamma=\gamma_1-\gamma_2$ be the closed $(0,2)$ form, non-vanishing since $\beta_1-\beta_2$ is non-zero from Lemma 1. \begin{prp} \label{4formulas}In the terminology above, \begin{itemize} \item $\beta_1=b+i\omega_+-(p-1)\bar\gamma/2$ \item $\beta_2=b+i\omega_+-(p+1)\bar\gamma/2$ \item $\omega_-=p\omega_++i(p^2-1)\bar\gamma/4-i(p^2-1)\gamma/4$ \end{itemize} \end{prp} \begin{prf} Since we are in two complex dimensions, there are functions $q_1,q_2$ such that the $(0,2)$ forms $\gamma_1,\gamma_2$ are given by $\gamma_1=q_1\bar\gamma, \gamma_2=q_2\bar\gamma$ and since $$\beta_1-\beta_2=\gamma_1-\gamma_2=\bar\gamma$$ we have $q_1-q_2=1$. Similarly $\omega_-^{0,2}=r\bar\gamma$. We have $\omega_-^2=\omega_+^2$ since this is the Riemannian volume form and $\omega_-=p\omega_++r\gamma+\bar r\bar\gamma$ since it is self-dual, hence $$\omega_+^2=\omega_-^2=(p\omega_++r\gamma+\bar r\bar\gamma)^2=p^2\omega_+^2+2\vert r\vert^2\gamma\bar\gamma$$ and so \begin{equation} (1-p^2)\omega_+^2=2\vert r\vert^2\gamma\bar\gamma. \label{eqA} \end{equation} Also $i\omega_++\gamma_1=i\omega_-+\delta_1$ and $\delta_1^2=0$ since it is of type $(0,2)$ relative to $I_-$ so $$0=(i\omega_++\gamma_1-i\omega_-)^2=(i\omega_++q_1\gamma-i[p\omega_++r\gamma+\bar r\bar\gamma])^2$$ and this gives \begin{equation} -(1-p)^2\omega_+^2=2i(q_1-ir)\bar r\gamma\bar\gamma. \label{eqB} \end{equation} The same argument for $\delta_2$ gives \begin{equation} (1+p)^2\omega_+^2=2i(q_2+ir)\bar r\gamma\bar\gamma. \label{eqC} \end{equation} From (\ref{eqA}),(\ref{eqB}),(\ref{eqC}) we obtain $$q_1=\frac{2ir}{p+1},\quad q_2=\frac{2ir}{p-1}$$ and from $q_1-q_2=1$ it follows that $r=i(p^2-1)/4$ and hence $q_1=-(p-1)/2$ and $q_2=-(p+1)/2$. \end{prf} \begin{rmk} The function $p$ (which figures prominently as the \emph{angle function} in the calculations of \cite{AGG}) can be read off from the $2$-forms $\beta_1,\beta_2$ using the above formulas. Recall that the imaginary part of $\beta$ must be symplectic to define a generalized complex structure. We calculate the two Liouville volume forms: $$(\beta_1-\bar\beta_1)^2=(p-1)\gamma\bar\gamma\qquad (\beta_2-\bar\beta_2)^2=-(p+1)\gamma\bar\gamma.$$ \end{rmk} \subsection{Examples} Because of Theorem \ref{bi}, the constructions in (\ref{cp2}) and (\ref{f2}) using generalized complex structures furnish us with bihermitian metrics. We now write these down. The complex structures $I_+,I_-$ are determined by the respective $(0,2)$ forms $\beta_1-\beta_2$ and $\beta_1-\bar\beta_2$. It is straightforward to see that \begin{eqnarray*} \lambda(\beta_1-\beta_2)&=& (H_{12}v_1+H_{22}v_2+\lambda\bar v_1)(-H_{11}v_1+H_{12}v_2+\lambda \bar v_2)\\ \lambda(\beta_1-\bar\beta_2)&=& (H_{12}\bar v_1-H_{22}\bar v_2+\lambda v_1)(H_{11}\bar v_1+H_{12}\bar v_2+\lambda v_2) \end{eqnarray*} The metric is obtained from the Hermitian form $\beta_1-\bar\beta_1$ on $(1,0)$ vectors. Using the basis of $(0,1)$ forms for $I_+$ given by the decomposition of $\beta_1-\beta_2$ above this turns out to be diagonal and the metric itself written as $$H_{11}\left[\frac{dr^2}{r^2-2\lambda + 2H_{12}}+\frac{r^2\sigma_1^2}{r^2-2\lambda-2H_{12}}\right]+H_{22}\left[\frac{r^2\sigma_2^2}{r^2-2\lambda + 2H_{12}}+\frac{r^2\sigma_3^2}{r^2-2\lambda-2H_{12}}\right].$$ \begin{rmk} If we replace the Poisson structure $\sigma=\partial/\partial z_1\wedge \partial/\partial z_2$ in our examples on ${\mathbf C}{\rm P}^2$ or $ {\mathbf F}_2$ by $t\sigma$, then as $t\rightarrow 0$ the limiting generalized complex structure $J_1$ arises from a complex structure and we should obtain simply a K\"ahler metric. This is equivalent to replacing the $2$-form $\beta_1$ by $t^{-1}\beta_1$. The differential equations for $H_{ij}$ remain the same but the algebraic constraint $\det H=\lambda(\lambda-r^2)$ becomes $\det H=\lambda(\lambda-t^{-1}r^2)$. The metric then becomes $$tH_{11}\left[\frac{dr^2}{r^2-2t\lambda + 2tH_{12}}+\frac{r^2\sigma_1^2}{r^2-2t\lambda-2tH_{12}}\right]+tH_{22}\left[\frac{r^2\sigma_2^2}{r^2-2t\lambda + 2tH_{12}}+\frac{r^2\sigma_3^2}{r^2-2t\lambda-2tH_{12}}\right]$$ and removing the overall factor of $t$ this tends to the K\"ahler metric $H_{11}v_1\bar v_1+H_{22}v_2\bar v_2$ we started our constructions with. \end{rmk} Concerning our examples of ${\mathbf C}{\rm P}^2$ and ${\mathbf F}_2$, one should be careful to distinguish the various complex structures. In each case we took a complex structure which had a holomorphic Poisson structure and used that to define a \emph{generalized} complex structure $J_1$. We then found a generalized complex structure $J_2$ commuting with it and reinterpreted the pair as a bihermitian metric with two integrable complex structures $I_+$ and $I_-$. It is well-known that ${\mathbf C}{\rm P}^2$ has a unique complex structure so that all three complex structures are equivalent by a diffeomorphism in that case. However, all the Hirzebruch surfaces ${\mathbf F}_{2m}$ are diffeomorphic to $S^2\times S^2$. For $m> 0$ there is a unique holomorphic $SL(2,\mathbf{C})$ action which has two orbits of complex dimension one: a curve of self-intersection $+2m$ and one of $-2m$. The complex structures $I_+,I_-$ that arose from our construction admit a holomorphic $SU(2)$ action and there are two spherical orbits of real dimension $2$ corresponding to $r=0$ and $r=\infty$. We shall show that the sphere $S_0$ given by $r=0$ is not holomorphic with respect to $I_+$. Note first that the $2$-form $\beta_1=r^2v_1v_2$ vanishes on $S_0$ because $\beta_1$ has type $(2,0)$ in the ${\mathbf F}_2$ complex structure and $S_0$ is holomorphic. Since $\lambda=r^2f_4$, this means that $\lambda(v_1v_2+\bar v_1 \bar v_2)$ vanishes on $S_0$. But from (\ref{h12}) $$H_{12}=4w\vert w\vert (1+\vert z\vert^2)\frac{\partial f_3}{\partial w}$$ so that $$H_{12}v_1\bar v_2=4\frac{\partial f_3}{\partial w}(\bar w dw d\bar z +\bar z\vert w\vert ^2dzd\bar z)$$ and this vanishes on $S_0$ since $w=0$ there. Thus, restricted to $S_0$, all the terms in $\beta_1-\beta_2$ except $H_{11}v_1\bar v_1+H_{22}v_2 \bar v_2$ vanish, and the latter is non-zero since it is the K\"ahler metric we started from. However $\beta_1-\beta_2$ is a $(0,2)$-form in the complex structure $I_+$ and this must vanish on $S_0$ if it is a holomorphic curve. \vskip .25cm We conclude that, with the complex structure $I_+$ this must be the Hirzebruch surface ${\mathbf F}_0={\mathbf C}{\rm P}^1\times {\mathbf C}{\rm P}^1$. \subsection{Holomorphic Poisson structures} \label {holp} Apostolov et al. in \cite{AGG} considered the four-dimensional bihermitian case where $I_+$ and $I_-$ define the same orientation and proved that the subset on which $I_+=\pm I_-$ is an anticanonical divisor with respect to both complex structures. Now an anticanonical divisor is a holomorphic section of $\Lambda^2T^{1,0}$ -- a holomorphic bivector $\sigma$. Since $[\sigma,\sigma]$ is a holomorphic section of $\Lambda^3 T^{1,0}$, in two complex dimensions this automatically vanishes and we have a Poisson structure. All compact surfaces with holomorphic Poisson structure have been listed by Bartocci and Macr\`\i \,\,using the classification of complex surfaces \cite{Bart}, so considering this list provides a basis for seeking compact bihermitian metrics in this dimension. Particular cases (overlooked in \cite{AGG}) are the projective bundle $P(1\oplus K)$ over any compact algebraic curve $C$, and the ``twisted' version $P(V)$ where $$0\rightarrow K\rightarrow V\rightarrow 1\rightarrow 0$$ is the nontrivial extension in $H^1(C,K)\cong \mathbf{C}$. When $C={\mathbf C}{\rm P}^1$ these two surfaces are are ${\mathbf F}_2$ and ${\mathbf F}_0$ respectively. We show now that the Poisson structure appears naturally in higher dimensions too. \vskip .25cm Let $M$ be a generalized K\"ahler manifold, now considered from the bihermitian point of view. Following \cite{AGG} we consider the $2$-form $$S(X,Y)=g([I_+,I_-]X,Y).$$ Since \begin{eqnarray*} S(I_+X,I_+Y)&=&g(I_+I_-I_+X,I_+Y)-g(I_-I_+^2X,I_+Y)\\ &=&g(I_-I_+X,Y)+g(I_-X,I_+Y)\\ &=&=g([I_-,I_+]X,Y)=-S(X,Y) \end{eqnarray*} this form is of type $(2,0)+(0,2)$. Pick the complex structure $I_+$. Using the antilinear isomorphism $T^{1,0}\cong (\bar T^*)^{0,1}$ provided by the hermitian metric, its $(0,2)$ part can be identified with a section $\sigma_+$ of the bundle $\Lambda^2T^{1,0}$. \begin{prp} \label{biv} The bivector $\sigma_+$ is a holomorphic Poisson structure. \end{prp} \begin{prf} We shall first show that $\sigma_+$ is holomorphic, and then that its Schouten bracket vanishes. Let $z_1,\dots,z_n$ be local holomorphic coordinates, then $$\sigma_+=\sum (I_- dz_i,dz_j)\frac{\partial}{\partial z_i}\wedge\frac{\partial}{\partial z_j}$$ where we use the inner product on $1$-forms defined by the metric and the complex structure $I_-$ on $1$-forms. We need to show that the functions $(I_- dz_i,dz_j)$ are holomorphic. Now \begin{equation} \frac{\partial}{\partial \bar z_k}(I_- dz_i,dz_j)=((\nabla^+_{\bar k}I_-) dz_i,dz_j)+(I_-\nabla^+_{\bar k}dz_i,dz_j)+(I_- dz_i, \nabla^+_{\bar k}dz_j). \label{holo} \end{equation} The Levi-Civita connection $\nabla$ has zero torsion so $$0=d(dz_i)=\sum dz_k\wedge \nabla_k dz_i+\sum d\bar z_k \wedge \nabla_{\bar k}dz_i.$$ But from (\ref{del}) $\nabla^+=\nabla +H/2$ where $H=g^{-1}db$, so \begin{equation} 0=\sum_k dz_k\wedge (\nabla^+_k-H_{k}/2) dz_i+\sum_k d\bar z_k \wedge (\nabla^+_{\bar k}-H_{\bar k}/2)dz_i. \label{dbareq} \end{equation} Now $\nabla^+$ preserves $I_+$ so that $\nabla^+_k dz_i$ and $\nabla^+_{\bar k} dz_i$ are $(1,0)$-forms. However, since $H$ is of type $(2,1)+(1,2)$, $H_k(dz_i)$ has a $(0,1)$ component. Equating the $(1,1)$ component of (\ref{dbareq}) to zero, the two contributions of $H$ give \begin{equation} \nabla^+_{\bar k}dz_i=H_{\bar k}(dz_i) \label{nab1} \end{equation} Now $I_-$ is preserved by $\nabla^-$ and from (\ref{del}) $\nabla^-=\nabla^+-H$, so $$\nabla_{\bar k}^+ I_-=[H_{\bar k},I_-].$$ Using this and (\ref{nab1}) in (\ref{holo}) we obtain $$\frac{\partial}{\partial \bar z_k}(I_- dz_i,dz_j)=([H_{\bar k},I_-] dz_i,dz_j)+(I_-H_{\bar k}(dz_i),dz_j)+(I_- dz_i, H_{\bar k}(dz_j))=0$$ and so $\sigma_+$ is holomorphic. \vskip .25cm To prove that $\sigma_+$ is Poisson we use (\ref{J12}) and the observation that the upper triangular part of $J_1$ is a real Poisson structure. This means that $$[\omega_+^{-1}+\omega_-^{-1},\omega_+^{-1}+\omega_-^{-1}]=0.$$ Now since $\omega_+$ is of type $(1,1)$, $\omega_+^{-1}+\omega_-^{-1}=h+\sigma_++\bar\sigma_+$ where $h$ is a bivector of type $(1,1)$. Because $\sigma_+$ is holomorphic, $[h,\sigma_+]$ has no $(3,0)$ component and so the $(3,0)$ component of $0=[h+\sigma_++\bar\sigma_+,h+\sigma_++\bar\sigma_+]$ is just $[\sigma_+,\sigma_+]$. Hence $[\sigma_+,\sigma_+]=0$ and we have a holomorphic Poisson structure. \end{prf} \vskip .25cm When the generalized K\"ahler structure is defined by $\rho_1=\exp \beta_1,\rho_2=\exp \beta_2$, as in Lemma 1, $\sigma_+$ has a direct interpretation. Recall that $\bar\beta_1-\bar\beta_2 =\gamma$ is a non-degenerate holomorphic $2$-form with respect to $I_+$. Then \begin{prp} \label{gprop} Let $\sigma_+:(T^{1,0})^*\rightarrow T^{1,0}$ be the holomorphic Poisson structure corresponding to the generalized K\"ahler structure given by $2$-forms $\beta_1,\beta_2$, and let $\gamma=\bar\beta_1-\bar\beta_2:T^{1,0}\rightarrow (T^{1,0})^*$ be the holomorphic $2$-form. Then $$\sigma_+=2i\gamma^{-1}.$$ \end{prp} \begin{prf} From (\ref{J12}) $\sigma_+$ is given by the upper-triangular part of $J_1$ evaluated on one-forms of type $(1,0)$ with respect to $I_+$. Since $\gamma$ is a non-degenerate $(2,0)$ form, any $(1,0)$ form can be written $i_X\gamma$ for a $(1,0)$-vector $X$. So we require to prove that if $X$ is a $(1,0)$ vector, then the $(1,0)$ component of $J_1(i_X\gamma)$ is $2iX$. Now \begin{eqnarray*} J_1(i_X\gamma)&=&J_1(i_X(\bar\beta_1-\bar\beta_2))\\ &=& J_1(i_X\bar\beta_1-X+X-i_X\bar\beta_2) \end{eqnarray*} and by the definition of $J_1$, \begin{equation} J_1(i_X\bar\beta_1-X)=-i(i_X\bar\beta_1-X) \label{first} \end{equation} The term $X-i_X\bar\beta_2$ is acted on as $-i$ by $J_2$ and we split it into components for the two $J_1$ eigenspaces: $$X-i_X\bar\beta_2=Y-i_Y\bar\beta_2+Z-i_Z\bar\beta_2.$$ Since $Z-i_Z\bar\beta_2$ is in the $-i$-eigenspace of both $J_1$ and $J_2$, $Z$ is of type $(0,1)$. Since $X=Y+Z$, $X=Y^{1,0}$. Now $$J_1(X-i_X\bar\beta_2)=i(Y-i_Y\bar\beta_2)-i(Z-i_Z\bar\beta_2)$$ and adding this to (\ref{first}), the upper triangular part of $J_1$ is given by $$J_1(i_X(\bar\beta_1-\bar\beta_2))=2iX -2iZ$$ whose $(1,0)$ part is $2iX$. \end{prf} \begin{ex} The examples of ${\mathbf C}{\rm P}^2$ and ${\mathbf F}_2$ were constructed by using $2$-forms $\beta_1,\beta_2$. Since $\beta_1$ had a pole on the curve at $r=\infty$ and $\beta_2$ was smooth everywhere, the Poisson structures $\sigma_+=2i(\bar\beta_1-\bar\beta_2)^{-1}$ and $\sigma_-=2i(\bar\beta_1-\beta_2)^{-1}$ vanish there. \end{ex} \section{Moduli spaces of instantons} \subsection{Stability} On a $4$-manifold with a Hermitian structure, the anti-self-dual (ASD) $2$-forms are the $(1,1)$-forms orthogonal to the Hermitian form. Thus on a generalized K\"ahler $4$-manifold, a connection with anti-self-dual curvature (an instanton) has curvature of type $(1,1)$ with respect to both complex structures $I_+,I_-$. In fact, where $I_+\ne \pm I_-$, anti-self-duality is equivalent to this condition. The equations $d^c_-\omega_-=db=-d^c_+\omega_+$ imply that $$dd_{\pm}^c\omega_{\pm}=0$$ which means that the metric is a Gauduchon metric with respect to both complex structures. With a Gauduchon metric one defines the \emph{degree} of a holomorphic line bundle $L$ by $$\mathop{\rm deg}\nolimits L=\frac{1}{2\pi}\int_M F\wedge\omega$$ where $F$ is the curvature form of a connection on $L$ defined by a Hermitian metric. Since a different choice of metric changes $F$ by $dd^cf$, the condition $dd^c\omega=0$ and integration by parts shows that the degree, a real number, is independent of the choice of Hermitian metric on $L$. It has the usual property of degree that if a holomorphic section of $L$ vanishes on a divisor $D$ then $$\mathop{\rm deg}\nolimits L=\int_D\omega.$$ So line bundles with sections which vanish somewhere have positive degree. \begin{rmk} Let us consider this non-K\"ahler degree for a bihermitian surface such that the Poisson structure vanishes on a divisor, like our examples of ${\mathbf C}{\rm P}^2$ and ${\mathbf C}{\rm P}^1\times {\mathbf C}{\rm P}^1$, and assume for convenience that the surface also carries a K\"ahler metric. The canonical bundle $K$ has no holomorphic sections since the product with the Poisson structure, a section of $K^*$, would give a holomorphic function with zeroes. This means $H^{2,0}(M)=0$ and so $H^2(M)$ is purely of type $(1,1)$. Now suppose that one of the generalized complex structures is defined by $\exp \beta$ where $\beta$ is closed. We saw in (\ref{4formulas}) that $\beta=-b+i\omega_++\gamma_1$ where $\gamma_1$ is of type $(0,2)$, so that the $(1,1)$ component of $\beta-\bar\beta$ is $2i\omega_+$. Thus the integral of $\omega_+$ over a holomorphic curve $C$, which is positive, is the same as the integral of the \emph{closed} form $(\beta-\bar\beta)/2i$. Let $W$ be the cohomology class of this form. Then we see that for every effective divisor $D$ on $M$, $WD>0$. Furthermore, $W$ is represented by the form $$(\beta-\bar\beta)/2i=\omega_+-i(\gamma_1-\bar\gamma_1)/2$$ which is self-dual, hence $W^2>0$. It follows from Nakai's criterion that $W$ is the cohomology class of a K\"ahler metric. Since the ample cone generates the whole of the cohomology, we see that the non-K\"ahler degree in this case agrees with the ordinary K\"ahler degree of some K\"ahler metric. Observe also that $\beta-\bar\beta$ is also equal to $2i\omega_-+\delta_1-\bar\delta_1$ so that we obtain the same degree function on cohomology for $I_+$ and $I_-$. \end{rmk} \vskip .25cm Using this definition of degree, one can define the slope of a subbundle, and from that the stability of a holomorphic bundle. The key theorem in the area, proved by Buchdahl \cite{Buch} for surfaces and Li and Yau \cite{LY} in the general case, is that a bundle is stable if and only if it has an irreducible ASD connection. A good reference for this is the book \cite{LT}. From this we already see that the moduli space ${\mathcal M}$ of ASD connections on a generalized K\"ahler manifold has two complex structures, by virtue of being the moduli space of stable bundles for both $I_+$ and $I_-$. We shall prove the following theorem: \begin{thm} \label{GKmod} Let $M^4$ be a compact even generalized K\"ahler manifold. Then the smooth points of the moduli space of ASD connections on a principal $SU(k)$-bundle over $M$ carries a natural bihermitian metric such that $d^c_-\omega_-=H=-d^c_+\omega_+$ for some exact $3$-form $H$ of type $(2,1)+(1,2)$. \end{thm} From Gualtieri's theorem this has a generalized K\"ahler interpretation once we choose a $2$-form $b$ such that $db=H$. \begin{rmk} In general, the moduli space of stable bundles may have singularities if the obstruction space $H^2(M,\mathop{\rm End}\nolimits_0 E)$ (where $\mathop{\rm End}\nolimits_0$ denotes trace-free endomorphisms) is non-vanishing. However, if the Poisson structure $s$ on $M$ is non-zero, then $$s:H^0(M,\mathop{\rm End}\nolimits_0 E\otimes K)\rightarrow H^0(M,\mathop{\rm End}\nolimits_0 E)$$ is injective. But stable bundles are simple,so $H^0(M,\mathop{\rm End}\nolimits_0 E)=0$. We deduce that $H^0(M,\mathop{\rm End}\nolimits_0 E\otimes K)$, and hence also its Serre dual $H^2(M,\mathop{\rm End}\nolimits_0 E)$, must vanish, so the moduli space is smooth (see \cite{Bot1}). This vanishing also gives us by Riemann-Roch the dimension of the $SU(k)$ moduli space $$\dim_{\mathbf{C}}{\mathcal M}=2kc_2(E)-(k^2-1)\frac{1}{12}(c_1^2+c_2)(M).$$ The simplest case would be $k=2, c_2(E)=n$ for our examples ${\mathbf C}{\rm P}^2,{\mathbf F}_2$ (or any rational surface) where $\dim_{\mathbf{C}}{\mathcal M}=4n-3.$ \end{rmk} \subsection{The metric on the moduli space} In \cite{LT} the metric structure of the moduli space of instantons on a Gauduchon manifold is discussed. It differs in general from the Riemannian or K\"ahler case. In the Riemannian situation, the space of all connections is viewed as an infinite-dimensional affine space with group of translations $\Omega^1(M,\lie{g})$ and ${\mathcal L}^2$ metric $$(a_1,a_2)=-\int_M\mathop{\rm tr}\nolimits(a_1\wedge \mathop{*\!}\nolimits a_2).$$ The solutions to the ASD equations form an infinite-dimensional submanifold with induced metric, and its quotient by the group of gauge transformations ${\mathcal G}$ is the moduli space, which acquires the quotient metric. To define this, one identifies the tangent space of the quotient at a point $[A]$ with the orthogonal complement to the tangent space of the gauge orbit at the connection $A$, with its restricted inner product. The orthogonal complement is identified with the bundle-valued $1$-forms $a\in \Omega^1(M,\lie{g})$ which satisfy the equation \begin{equation} d_A^*a\, (=-\mathop{*\!}\nolimits \mathop{d_A*\!}\nolimits a)=0 \label{horizont} \end{equation} As the authors of \cite{LT} point out, this metric in the Gauduchon case is not Hermitian with respect to the natural complex structure that the moduli space acquires through its identification with the moduli space of stable bundles. Instead of the orthogonality (\ref{horizont}), one takes a different horizontal subspace defined by \begin{equation} \omega \wedge d_A^ca=0. \label{newhorizont} \end{equation} \begin{lem} \label{hori}$\omega \wedge d^c_A a = \mathop{d_A*\!}\nolimits a- d^c\omega \wedge a.$ \end{lem} From this we see that when the metric is K\"ahler, $d^c\omega=0$, and so the two horizontality conditions coincide. \begin{lemprf} Note that for any $\psi\in \Omega^0(M,\lie{g})$, \begin{equation} d_A^c(\omega \wedge\mathop{\rm tr}\nolimits(a \psi))=d^c\omega\wedge\mathop{\rm tr}\nolimits(a\psi)+\omega\wedge\mathop{\rm tr}\nolimits(d^c_A a\psi)-\omega\wedge\mathop{\rm tr}\nolimits(a \wedge d_A^c\psi) \label{exp1} \end{equation} and $d_A^c\psi=I^{-1}d_AI\psi=-Id_A\psi$, so that $$\omega\wedge\mathop{\rm tr}\nolimits(a\wedge d_A^c\psi)=-\omega\wedge\mathop{\rm tr}\nolimits(a\wedge Id_A\psi)=(a,d_A\psi)\omega^2=\mathop{\rm tr}\nolimits(\mathop{*\!}\nolimits a \wedge d_A\psi).$$ Integrating (\ref{exp1}) and using Stokes' theorem and the relation above, we get $$\int_M[d^c\omega\wedge\mathop{\rm tr}\nolimits(a\psi) +\omega\wedge\mathop{\rm tr}\nolimits(d^c_A a\psi)-\mathop{\rm tr}\nolimits(\mathop{d_A*\!}\nolimits a \psi)]=0$$ so that \begin{equation} \omega \wedge d^c_A a = \mathop{d_A*\!}\nolimits a- d^c\omega \wedge a. \label{newhorizont1} \end{equation} \end{lemprf} With this choice of horizontal, the metric on the moduli space is Hermitian with Hermitian form $$\tilde\omega(a_1,a_2)=\int_M\omega\wedge \mathop{\rm tr}\nolimits(a_1\wedge a_2).$$ It is shown in \cite{LT} that $\tilde\omega$ satisfies $dd^c\tilde\omega=0$. The horizontal subspace (\ref{newhorizont}) defines a connection on the infinite-dimensional principal ${\mathcal G}$-bundle over the moduli space and its curvature turns out to be of type $(1,1)$ on ${\mathcal M}$ (see \cite{LT}). We shall make use of these facts later. \vskip .25cm In order to prove Theorem \ref{GKmod} we need first to show that the application of L\"ubke and Teleman's approach to the two complex structures $I_+$ and $I_-$ yields the same metric. \vskip .25cm The tangent space to the moduli space at a smooth point is the first cohomology of the complex: $$\Omega^0(M,\lie{g})\stackrel{d_A}\longrightarrow \Omega^1(M,\lie{g})\stackrel{d^+_A}\longrightarrow \Omega_+^2(M,\lie{g})$$ where here the $+$ refers to projection onto the self-dual part. The metric is the induced inner product on the subspace of $\Omega^1(M,\lie{g})$ defined by the horizontality condition $\omega \wedge d^c_A a=0$. We shall write $[a]$ for the tangent vector to the moduli space represented by $a$. In our case we have two such horizontality conditions $\omega_- \wedge d^c_-a=0$ and $\omega_+ \wedge d^c_+a=0$ (suppressing the subscript $A$ for clarity) and two representatives $a$ and $a+d_A\psi$ for the same tangent vector. We shall call these plus- and minus- horizontal respectively. We prove: \begin{lem} Let $a$ and $a+d_A\psi$ satisfy $$\omega_-\wedge d^c_-a=0,\quad \omega_+ \wedge d^c_+(a+d_A\psi)=0.$$ Then $(a,a)=(a+d_A\psi,a+d_A\psi)$. \end{lem} \begin{lemprf} Since in our case $d^c_-\omega_-=db=h=-d^c_+\omega_+$ our two horizontality conditions are, from (\ref{newhorizont1}) $$ \mathop{d_A*\!}\nolimits a- h \wedge a=0\quad \mathop{d_A*\!}\nolimits\, (a+d_A\psi)+h \wedge (a+d_A\psi)=0$$ and so, eliminating $h\wedge a$, $$2\mathop{d_A*\!}\nolimits a+d_A\mathop{*\!}\nolimits d_A\psi+h\wedge d_A\psi=0.$$ This gives on integration $$\int_M[2\mathop{\rm tr}\nolimits(\mathop{d_A*\!}\nolimits a \psi)+\mathop{\rm tr}\nolimits(\mathop{d_A*\!}\nolimits d_A\psi \psi)+h\wedge\mathop{\rm tr}\nolimits(d_A\psi \psi)]=0.$$ But $\mathop{\rm tr}\nolimits(d_A\psi\psi)=d\mathop{\rm tr}\nolimits\psi^2/2$ so the last term is $d[(\mathop{\rm tr}\nolimits\psi^2)h/2]$ as $h$ is closed. By Stokes' theorem we get $$2(a,d_A\psi)+(d_A\psi,d_A\psi)=0$$ and hence $$(a+d_A\psi,a+d_A\psi)=(a,a)$$ as required. \end{lemprf} \subsection{The bihermitian structure} So far, we have seen that ${\mathcal M}$ has two complex structures and a metric, Hermitian with respect to both. We now need to show that $d^c_+\tilde\omega_+=H=-d^c_-\tilde \omega_-$ for an exact $3$-form $H$. \vskip .25cm Denote by ${\mathcal A}$ the affine space of all connections on the principal bundle, then a tangent vector is given by $a\in \Omega^1(M,\lie{g})$ and for any $2$-form $\omega$, $$\Omega(a_1,a_2)=\int_M\omega\wedge \mathop{\rm tr}\nolimits(a_1\wedge a_2)$$ is a closed and gauge-invariant $2$-form on ${\mathcal A}$. It is closed because it is translation-invariant on $\mathcal{A}$ (has ``constant coefficients"). We defined Hermitian forms $\tilde\omega_{\pm}$ on ${\mathcal M}$ by $$\tilde\omega_{\pm}([a_1],[a_2])=\Omega_{\pm}(a_1,a_2)=\int_M\omega_{\pm}\wedge\mathop{\rm tr}\nolimits(a_1\wedge a_2)$$ where $a_1,a_2$ are plus/minus-{\it horizontal}. Now the formula for the exterior derivative of a $2$-form $\alpha$ is $$3d\alpha(a_1,a_2,a_3)=a_1\cdot\alpha(a_2,a_3)-\alpha([a_1,a_2],a_3)+\mathrm {cyclic}$$ so, since $\Omega$ is closed $$3d\tilde\omega([a_1],[a_2],[a_3])=-\int_M\omega\wedge\mathop{\rm tr}\nolimits([a_1,a_2]_V)\wedge a_3)+\mathrm {cyclic}$$ where $[a_1,a_2]_V$ is the vertical component of the Lie bracket of the two vector fields. By definition this is the curvature of the ${\mathcal G}$-connection. If $\theta(a_1,a_2)\in \Omega^0(M,\lie{g})$ is this curvature then $[a_1,a_2]_V=d_A\theta(a_1,a_2)$. Using Stokes' theorem \begin{eqnarray*} 3d\tilde\omega([a_1],[a_2],[a_3])&=&-\int_M\omega\wedge\mathop{\rm tr}\nolimits(d_A\theta(a_1,a_2)\wedge a_3)+\mathrm {cyclic}\\ &=& \int_M d\omega\wedge \mathop{\rm tr}\nolimits(\theta(a_1,a_2)a_3)+ \int_M\omega\wedge\mathop{\rm tr}\nolimits(\theta(a_1,a_2)d_Aa_3)+\mathrm {cyclic}\\ &=&\int_M d\omega\wedge \mathop{\rm tr}\nolimits(\theta(a_1,a_2)a_3) +\mathrm {cyclic} \end{eqnarray*} since $d_Aa_3$ is anti-self-dual and $\omega$ is self-dual so $\omega\wedge d_Aa_3=0$. \vskip .25cm Now $d^c\omega(a_1,a_2,a_3)=-d\omega(Ia_1,Ia_2,Ia_3)$ and from \cite{LT} the curvature of the ${\mathcal G}$-bundle is of type $(1,1)$. This means that $\theta(Ia_2,Ia_3)=\theta(a_2,a_3)$ and so, for the structure $I_-$ \begin{equation} d_-^c\tilde\omega([a_1],[a_2],[a_3])=\int_M d_-^c\omega_-\wedge \mathop{\rm tr}\nolimits(\theta_-(a_1,a_2)a_3)+\mathrm {cyclic} \label{deec} \end{equation} with a similar equation for $I_+$. \vskip .25cm To proceed further we need more information about the curvature $\theta(a_1,a_2)$. On the affine space ${\mathcal A}$ the Lie bracket of $a_1$ and $a_2$ considered as vector fields is just $a_1\cdot a_2-a_2\cdot a_1$ where $a\cdot b$ denotes the flat derivative of $b$ in the direction $a$. The horizontality condition imposes a constraint: $$ \mathop{d_A*\!}\nolimits a_2- h \wedge a_2=0.$$ Differentiating the constraint in the direction $a_1$ gives $$[a_1,\mathop{*\!}\nolimits a_2]+\mathop{d_A*\!}\nolimits a_1\cdot a_2-h\wedge a_1\cdot a_2=0.$$ The vertical component of the Lie bracket is $d_A\theta(a_1,a_2)$ which thus satisfies \begin{equation} \mathop{d_A*\!}\nolimits d_A\theta-h \wedge d_A\theta +2[a_1,\mathop{*\!}\nolimits a_2]=0. \label{thetaeq} \end{equation} Define the second order operator $\Delta:\Omega^0(M,\lie{g})\rightarrow \Omega^4(M,\lie{g})$ by $$\Delta\psi=d_A\mathop{*\!}\nolimits d_A\psi-h\wedge d_A\psi,$$ then its formal adjoint is $$\Delta^*\psi=d_A\mathop{*\!}\nolimits d_A\psi+h\wedge d_A\psi$$ and we rewrite (\ref{thetaeq}) as \begin{equation} \Delta\theta(a_1,a_2)+2[a_1,\mathop{*\!}\nolimits a_2]=0 \label{thetaeqs} \end{equation} for plus-horizontal vector fields $a_i$. Let $b_i=a_i+d_A\psi_i$ be the minus-horizontal representatives of $[a_i]$. By minus-horizontality we have $$0=d_A\mathop{*\!}\nolimits b_i+h\wedge b_i=\mathop{d_A*\!}\nolimits\, (a_i+d_A\psi_i)+h\wedge (a_i+d_A\psi_i)=\Delta^*\psi_i+\mathop{d_A*\!}\nolimits a_i +h\wedge a_i$$ and together with the plus-horizontality condition $ \mathop{d_A*\!}\nolimits a_i- h \wedge a_i=0$ we get \begin{equation} 2h\wedge a_i=-\Delta^*\psi_i. \label{hai} \end{equation} Since $d^c_-\omega_-=h$, each integrand on the right hand side of (\ref{deec}) is, from (\ref{hai}), of the form $$h\wedge \mathop{\rm tr}\nolimits(\theta(a_1,a_2)a_3)=-\mathop{\rm tr}\nolimits(\theta(a_1,a_2)\Delta^*\psi_3/2).$$ Performing the integration and using Stokes' theorem, we obtain $$-\int_M \mathop{\rm tr}\nolimits(\theta(a_1,a_2)\Delta^*\psi_3)/2=-\int_M \mathop{\rm tr}\nolimits(\Delta\theta(a_1,a_2)\psi_3)/2=\int_M \mathop{\rm tr}\nolimits([a_1,\mathop{*\!}\nolimits a_2]\psi_3)$$ from (\ref{thetaeqs}). \vskip.25cm Working with the curvature of the plus-connection we get a similar expression so that we have two formulae: \begin{eqnarray*} d_-^c\tilde\omega_-([a_1],[a_2],[a_3])&=&\int_M \mathop{\rm tr}\nolimits([a_1,\mathop{*\!}\nolimits a_2]\psi_3)+\mathrm {cyclic}\\ d_+^c\tilde\omega_+([a_1],[a_2],[a_3])&=&-\int_M \mathop{\rm tr}\nolimits([b_1,\mathop{*\!}\nolimits b_2]\psi_3)+\mathrm {cyclic} \end{eqnarray*} Thus to obtain $d_-^c\tilde\omega_- =-d_+^c\tilde\omega_+$, using $b_i=a_i+d_A\psi_i$ in the above leads to the need to prove: \begin{lem} $$\int_M [\mathop{\rm tr}\nolimits([a_1,\mathop{*\!}\nolimits d_A\psi_2]\psi_3)+\mathop{\rm tr}\nolimits([d_A\psi_1,\mathop{*\!}\nolimits a_2]\psi_3)+\mathop{\rm tr}\nolimits([d_A\psi_1,\mathop{*\!}\nolimits d_A\psi_2]\psi_3)]+\mathrm {cyclic}=0.$$ \end{lem} \begin{lemprf} Picking out the integrand involving $a_1$ in the cyclic sum we have \begin{eqnarray*} \mathop{\rm tr}\nolimits([a_1,\mathop{*\!}\nolimits d_A\psi_2]\psi_3)+\mathop{\rm tr}\nolimits([d_A\psi_3,\mathop{*\!}\nolimits a_1]\psi_2)&=&\mathop{\rm tr}\nolimits(\mathop{*\!}\nolimits a_1\wedge([\psi_3, d_A\psi_2]+ [d_A\psi_3,\psi_2]))\\ &=&-\mathop{\rm tr}\nolimits(\mathop{*\!}\nolimits a_1\wedge d_A[\psi_2,\psi_3]) \end{eqnarray*} and on integrating, this is \begin{eqnarray*} -\int_M\mathop{\rm tr}\nolimits(\mathop{*\!}\nolimits a_1\wedge d_A[\psi_2,\psi_3])&=&-\int_M\mathop{\rm tr}\nolimits(\mathop{d_A*\!}\nolimits a_1[\psi_2,\psi_3])\\ &=& -\int_M h\wedge \mathop{\rm tr}\nolimits(a_1[\psi_2,\psi_3])\\ &=& \int_M\mathop{\rm tr}\nolimits(\Delta^*\psi_1[\psi_2,\psi_3])/2 \end{eqnarray*} from (\ref{hai}). But from the definition of $\Delta^*$ this is $$\frac{1}{2}\int_M\mathop{\rm tr}\nolimits(d_A\mathop{*\!}\nolimits d_A\psi_1[\psi_2,\psi_3])-\frac{1}{2}\int_M h\wedge \mathop{\rm tr}\nolimits(d_A\psi_1[\psi_2,\psi_3]).$$ The cyclic sum of the second term vanishes since $$d\mathop{\rm tr}\nolimits(\psi_1[\psi_2,\psi_3])=\mathop{\rm tr}\nolimits(d_A\psi_1[\psi_2,\psi_3])+\mathrm {cyclic}$$ and $h$ is closed. Using Stokes' theorem on the first and expanding, the cyclic sum gives $$\frac{1}{2}\int_M\mathop{\rm tr}\nolimits(\mathop{*\!}\nolimits d_A\psi_1\wedge( [d_A\psi_2,\psi_3]+[\psi_2,d_A\psi_3])+\mathrm {cyclic}$$ which is $$-\int_M\mathop{\rm tr}\nolimits([d_A\psi_1,\mathop{*\!}\nolimits d_A\psi_2]\psi_3)]+\mathrm {cyclic}$$ and this proves the lemma. \end{lemprf} \vskip .25cm We finally need to show that $H$ is exact. One might expect that we simply define a $2$-form $\tilde b$ from the $2$-form $b$ on $M$ by \begin{equation} \tilde b([a_1],[a_2])=\int_M b\wedge \mathop{\rm tr}\nolimits(a_1\wedge a_2) \label{btilde} \end{equation} to get $d\tilde b=d_-^c\tilde\omega_-$ but this does not hold. The equation for the exterior derivative of $\tilde b$ gives $$3d\tilde b([a_1],[a_2],[a_3])=\int_M db\wedge \mathop{\rm tr}\nolimits(\theta(a_1,a_2)a_3)+ \int_M b\wedge\mathop{\rm tr}\nolimits(\theta(a_1,a_2)d_Aa_3)+\mathrm {cyclic}.$$ When we used this above with $\omega_+, \omega_-$ replacing $b$, the second term vanished because $d_Aa_3$ is anti-self-dual and $\omega_{\pm}$ are self-dual. This is not the case for a general $b$, and will only be true if $b$ is self-dual. We shall see in Section 5 a more general occurrence of this phenomenon. However we do have the following: \begin{lem} Any $2$-form $b$ on a compact oriented four-manifold $M$ is the sum of a closed form and a self-dual form. \end{lem} \begin{lemprf} Use the non-degenerate pairing on $2$-forms $$(\alpha,\beta)=\int_M\alpha\wedge \beta.$$ The annihilator of the self-dual forms $\Omega^2_+$ in this pairing is $\Omega^2_-$, and the annihilator of $\Omega^2_{closed}$ is $\Omega^2_{exact}$ so the annihilator of $\Omega^2_++\Omega^2_{closed}$ is the intersection of $\Omega^2_-$ and $\Omega^2_{exact}$. But if $\alpha$ is exact, then by Stokes' theorem $$\int_M\alpha\wedge \alpha=0$$ and if $\alpha\in \Omega^2_-$ $$\int_M\alpha\wedge \alpha=-(\alpha,\alpha)$$ so if both hold then $\alpha=0$. \end{lemprf} It follows from this that $db=db_+$ where $b_+$ is self-dual, and then (\ref{btilde}) does define a form $\tilde b_+$ on the moduli space. It follows than that $d\tilde b_+=d_-^c\tilde\omega_-=-d_+^c\tilde\omega_+$. \subsection{The Poisson structures on ${\mathcal M}$} As we saw in Proposition \ref{biv}, a generalized K\"ahler structure defines a holomorphic Poisson structure for each of the complex structures $I_+,I_-$. We shall determine these on the instanton moduli space next. On the moduli space of stable bundles over a Poisson surface $M$, there is a canonical holomorphic Poisson structure, defined by Bottacin in \cite{Bot1} as follows. The holomorphic tangent space at a bundle $E$ is the sheaf cohomology group $H^1(M,\mathop{\rm End}\nolimits E)$ and by Serre duality, the cotangent space is $H^1(M,\mathop{\rm End}\nolimits E \otimes K)$. The Poisson structure on $M$ is a holomorphic section $s$ of the anticanonical bundle $K^*$ and for $\alpha,\beta\in H^1(M,\mathop{\rm End}\nolimits E \otimes K)$, the Poisson structure $\sigma$ on the moduli space is defined by taking $\mathop{\rm tr}\nolimits(\alpha\beta)\in H^2(M,K^2)$, multiplying by $s\in H^0(M,K^*)$ to get $$\sigma(\alpha,\beta)=s\mathop{\rm tr}\nolimits(\alpha\beta)\in H^2(M,K)\cong \mathbf{C}.$$ The definition is very simple, the difficult part of \cite{Bot1} is proving the vanishing of the Schouten bracket. \begin{thm} \label{botta} Let $\sigma_+$ be the $I_+$ - Poisson structure defined by the generalized K\"ahler structure on ${\mathcal M}$. Then $\sigma_+/2$ is the canonical structure on the moduli space of $I_+$-stable bundles. \end{thm} \begin{prf} In the generalized K\"ahler setup, the Poisson structure $\sigma_+$ is defined by the $(0,2)$ part of $\omega_-$ under the antilinear identification $T^{1,0}\cong (\bar T^*)^{0,1}$ defined by the metric. A tangent vector to ${\mathcal M}$ is defined by $a\in \Omega^1(M,\lie{g})$ satisfying $d^+_Aa=0$, and this implies that $a^{0,1}\in \Omega^{0,1}(M,\mathop{\rm End}\nolimits E)$ satisfies $\bar\partial_Aa^{0,1}=0\in \Omega^{0,2}(M,\mathop{\rm End}\nolimits E)$, which is the tangent vector in the holomorphic setting -- it is a Dolbeault representative for a class in $H^1(M,\mathop{\rm End}\nolimits E)$. The conjugate $a^{1,0}=\overline{a^{0,1}}$ defines a complex cotangent vector by the linear form $$b^{0,1}\mapsto \int_M\omega_+\wedge\mathop{\rm tr}\nolimits (a^{1,0}\wedge b^{0,1})$$ and this is the antilinear identification $T^{1,0}\cong (\bar T^*)^{0,1}$ on the moduli space. However $\omega_+\wedge a^{1,0}\in \Omega^{2,1}(M,\mathop{\rm End}\nolimits E)$ is not a Dolbeault representative for the Serre dual -- it is not $\bar\partial$-closed -- so to see concretely the canonical Poisson structure we must find a good representative $(2,1)$ form. \vskip .25cm Now from $d_A^+a=0$ we have $\omega_+\wedge d_A(a^{1,0}+a^{0,1})=0$ and from the horizontality condition $\omega_+\wedge d^c_+a=0$, we obtain $\omega_+\wedge d_A(a^{1,0}-a^{0,1})=0$ so putting them together \begin{equation} \omega_+\wedge\bar\partial_A a^{1,0}=0,\quad \omega_+\wedge\partial_A a^{0,1}=0 \label{infasd} \end{equation} \vskip .25cm From Lemma \ref{hori} applied to $I_+$ and $I_-$ we have $$\omega_{\pm} \wedge d_{\pm}^c a = \mathop{d_A*\!}\nolimits a- d_{\pm}^c\omega_{\pm} \wedge a$$ so that since $d^c_-\omega_-=-d^c_+\omega_+$, $$\omega_{-} \wedge d_{-}^c a = \omega_{+} \wedge d_{+}^c a +2 d_+^c\omega_{+} \wedge a.$$ If $a=a^{1,0}+d_A\psi$ is minus-horizontal then this equation tells us that $$0=\omega_{+} \wedge d_{+}^c d_A\psi+2d_+^c\omega_{+}\wedge (a^{1,0}+d_A\psi)$$ since $a^{1,0}$ is plus-horizontal. We rewrite this as \begin{equation} 2i\omega_{+}\wedge \bar\partial_A\partial_A \psi+2i\bar\partial\omega_{+}\wedge (a^{1,0}+\partial_A\psi)-2i\partial\omega_+\wedge\bar\partial_A\psi=0 \label{goodeq} \end{equation} using the fact that $\omega_+\wedge F=0$ where $F$ is the curvature of the connection $A$. This gives, using $\bar\partial\partial\omega_+=0$ and (\ref{infasd}), \begin{equation} \bar\partial_A[\omega_+\wedge (a^{1,0}+\partial_A\psi)+\psi\partial\omega_+]=0 \label{Dolrep} \end{equation} Here, then, we have a $\bar\partial$-closed form, and it represents the dual of $[a^{0,1}]$ using the metric on ${\mathcal M}$ since, from Stokes' theorem, $$\int_M[\omega_+\wedge \mathop{\rm tr}\nolimits((a^{1,0}+\partial_A\psi)\wedge b^{0,1})+\partial\omega_+\wedge\mathop{\rm tr}\nolimits(\psi b^{0,1})]= \int_M\omega_+\wedge \mathop{\rm tr}\nolimits(a^{1,0}\wedge b^{0,1})-\int_M\omega_+\wedge\mathop{\rm tr}\nolimits(\psi\partial_Ab^{0,1})$$ and the second term on the right hand side vanishes from (\ref{infasd}). \vskip .25cm Now where the Poisson structure $s$ on $M$ is non-vanishing we have a closed $2$-form $\beta_1-\bar\beta_2$ which from Proposition \ref{4formulas} can be expressed as $2i\omega_+-(p-1)\bar\gamma/2+(p+1)\gamma/2.$ Since this is closed, and $\gamma$ is of type $(2,0)$, $4id\omega_+=\partial p\wedge \bar\gamma-\bar\partial p\wedge \gamma$ and so \begin{equation} 4i\partial\omega_+=-\bar\partial p\wedge \gamma \label{domega} \end{equation} We can therefore rewrite the Dolbeault representative as $$\omega_+\wedge (a^{1,0}+\partial_A\psi)+i\psi\bar\partial p\wedge \gamma/4.$$ The canonical Poisson structure is therefore obtained by integrating over $M$ the form \begin{equation} \mathop{\rm tr}\nolimits[s(\omega_+\wedge (a_1^{1,0}+\partial_A\psi_1)+i\psi_1\bar\partial p\wedge \gamma/4)\wedge(\omega_+\wedge (a_2^{1,0}+\partial_A\psi_2)+i\psi_2\bar\partial p\wedge \gamma/4)] \label{integrand} \end{equation} \vskip .25cm Take the product of the two expressions with an $\omega_+$ factor. For $(1,0)$ forms $a,b$, at each point $[s(\omega_+\wedge a)]\wedge\omega_+\wedge b$ is a skew form on $T^{1,0}$ with values in $\Lambda^4T^*$ depending on a Hermitian form and a $(2,0)$ form $\gamma$ (recall from Proposition \ref{gprop} that $s\gamma=2i$). By $SU(2)$ invariance this must be a multiple of $\bar\gamma\wedge a\wedge b$ and a simple calculation shows that $$[s(\omega_+\wedge a)]\wedge\omega_+\wedge b=-i\frac{\omega_+^2}{\gamma\bar\gamma}\,\bar\gamma\wedge a\wedge b.$$ However from (\ref{eqA}) and $r=i(p^2-1)/4$ we see that $$\frac{\omega_+^2}{\gamma\bar\gamma}=\frac{1}{8}(1-p^2).$$ But now from Proposition \ref{4formulas}, $\omega_-=p\omega_++i(p^2-1)\bar\gamma/4-i(p^2-1)\gamma/4$ and so $$[s(\omega_+\wedge a)]\wedge\omega_+\wedge b=\frac{i}{2}\omega_-^{0,2}\wedge a\wedge b=\frac{i}{2}\omega_-\wedge a\wedge b$$ since $a$ and $b$ are of type $(1,0)$. Thus the first two expressions contribute to the integral the term \begin{equation} \frac{1}{2}\int_M\omega_- \wedge\mathop{\rm tr}\nolimits(a_1^{1,0}+\partial_A\psi_1)\wedge (a_2^{1,0}+\partial_A\psi_2) \label{two0} \end{equation} \vskip .25cm The last two terms in (\ref{integrand}) give zero contribution because of the common $\bar\partial p$ factor. For the other terms, the relation $s\gamma=2i$ means that we are considering the integral of \begin{equation} -\mathop{\rm tr}\nolimits[\psi_1\bar\partial p\wedge\omega_+\wedge (a_2^{1,0}+\partial_A\psi_2)]/2+\mathop{\rm tr}\nolimits[\psi_2\bar\partial p\wedge\omega_+\wedge (a_1^{1,0}+\partial_A\psi_1)]/2. \label{crossterm} \end{equation} Take the first expression. This no longer contains the singular term $\gamma$ so we can integrate over the manifold and using Stokes' theorem we get \begin{equation} \frac{1}{2}\int_M p\omega_+\wedge\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\wedge (a_2^{1,0}+\partial_A\psi_2))+p\mathop{\rm tr}\nolimits[\psi_1\bar\partial_A[\omega_+\wedge (a^{1,0}+\partial_A\psi)] \label{integrate} \end{equation} Now from (\ref{Dolrep}) and (\ref{domega}) $$\bar\partial_A[\omega_+\wedge (a^{1,0}+\partial_A\psi)]=-\bar\partial_A(\psi\partial\omega_+)=\bar\partial_A\psi\wedge\bar\partial p\wedge \gamma/4i$$ Using this we can write (\ref{integrate}) as \begin{equation} \frac{1}{2}\int_M p\omega_+\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\wedge (a_2^{1,0}+\partial_A\psi_2))-\frac{i}{8}\int_M p\bar\partial p\wedge\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\psi_2)\wedge \gamma \label{integrate1} \end{equation} Now $\omega_-^{1,1}=p\omega_+$ and the first term integrates a $(1,1)$ form against $p\omega_+$ so we write this as \begin{equation} \frac{1}{2}\int_M \omega_-\wedge\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\wedge (a_2^{1,0}+\partial_A\psi_2)) \label{oneone} \end{equation} From Proposition \ref{4formulas}, we have $$\omega_-^{2,0}=-i(p^2-1)\gamma/4$$ so the last term in (\ref{integrate1}) is $$\frac{1}{4}\int_M \bar\partial \omega^{2,0}\wedge\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\psi_2)$$ which using Stokes' theorem gives $$\frac{1}{4}\int_M \omega_-^{2,0}\wedge\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\bar\partial \psi_2)$$ which we write as $$\frac{1}{4}\int_M \omega_-\wedge\mathop{\rm tr}\nolimits(\bar\partial_A\psi_1\bar\partial \psi_2).$$ In the full integral there is another contribution of this form from the second term in (\ref{crossterm}) and adding all terms in (\ref{integrand}) we obtain $$\frac{1}{2}\int_M\omega_-\wedge\mathop{\rm tr}\nolimits(a_1^{1,0}+d_A\psi_1)\wedge\mathop{\rm tr}\nolimits(a_2^{1,0}+d_A\psi_2).$$ Since $a_1^{1,0}+d_A\psi_1, a_2^{1,0}+d_A\psi_2$ are minus-horizontal representatives of $a_1^{1,0},a_2^{1,0}$ we see from the definition of $\tilde\omega_-$ that this is $\tilde\omega_-^{0,2}/2$ evaluated on those two vectors and hence is half the Poisson structure defined by the bihermitian metric. \end{prf} \subsection{The generalized K\"ahler structure} As we have seen, the bihermitian structure of $M^4$ naturally induces a similar structure on the moduli space of instantons, but we only get a pair $J_1,J_2$ of commuting generalized complex structures by \emph{choosing} a $2$-form with $db=H$. In that respect $J_1,J_2$ are defined modulo a closed B-field but we can still extract some information about them. In particular the formula (\ref{J12}) shows that the real Poisson structures defined by $J_1$ and $J_2$, namely $\omega_+^{-1}\pm \omega_-^{-1}$, are unchanged by $b\mapsto b+B$. We shall determine the \emph{symplectic foliation} on ${\mathcal M}$ determined by these Poisson structures, which relates to the ``type" of the generalized complex structure as discussed by Gualtieri. The symplectic foliation of a Poisson structure $ \pi$ is determined by the subspace of the cotangent bundle annihilated by $\pi:T^*\rightarrow T$. From (\ref{J12}), in our case $$\mathop{\rm ker}\nolimits \pi_1=\mathop{\rm ker}\nolimits (I_++I_-),\quad \mathop{\rm ker}\nolimits \pi_2=\mathop{\rm ker}\nolimits (I_+-I_-)$$ where $I_+,I_-$ act on $T^*$. Note that if $I_+a=I_-a$ then $$[I_+,I_-]a=I_+I_-a-I_-I_+a=(I_+)^2a-(I_-)^2a=-a+a=0$$ so that $\mathop{\rm ker}\nolimits (I_+-I_-)\subset \mathop{\rm ker}\nolimits[I_+,I_-]$, and similarly if $I_+a=-I_-a$. It follows that if $I_+a=I_-a$, then $I_+(I_+a)=I_-(I_+a)$ since both sides are equal to $-a$. Thus $\mathop{\rm ker}\nolimits \pi_1$ and $\mathop{\rm ker}\nolimits \pi_2$ are complex subspaces of $\mathop{\rm ker}\nolimits[I_+,I_-]$ (with respect to either structure). Now the kernel of $[I_+,I_-]$ is, from \ref{holp}, the kernel of the holomorphic Poisson structure $\sigma_+$ (or $\sigma_-$). But Theorem \ref{botta} tells us that this is the canonical Poisson structure on ${\mathcal M}$. Its kernel is easily determined (see \cite{Bot1}). Recall that the Poisson structure is defined, as a map from $(T^{1,0})^*$ to $T^{1,0}$, by the multiplication operation of the section $s$ of $K^*$: $$s:H^1(M,\mathop{\rm End}\nolimits E\otimes K)\rightarrow H^1(M,\mathop{\rm End}\nolimits E).$$ If $D$ is the anticanonical divisor of $s$ then we have an exact sequence of sheaves $$0\rightarrow{\mathcal O}_M(\mathop{\rm End}\nolimits E\otimes K)\stackrel{s}\rightarrow {\mathcal O}_M(\mathop{\rm End}\nolimits E)\rightarrow {\mathcal O}_D(\mathop{\rm End}\nolimits E)\rightarrow 0$$ and the above is part of the long exact cohomology sequence. Since a stable bundle is simple, $H^0(M,\mathop{\rm End}\nolimits E)$ is just the scalars, so the map $H^0(M,\mathop{\rm End}\nolimits E)\rightarrow H^0(D,\mathop{\rm End}\nolimits E)$ just maps to the scalars. Hence the kernel of $\sigma_+$ is isomorphic from the exact sequence to $H^0(D,\mathop{\rm End}\nolimits_0 E)$ under the connecting homomorphism: $$\delta_+: H^0(D,\mathop{\rm End}\nolimits E)\rightarrow H^1(M,\mathop{\rm End}\nolimits E\otimes K).$$ When $D$, an anticanonical divisor, is of multiplicity $1$ and smooth, it is an elliptic curve by the adjunction formula: $KD+D^2=2g-2$ implies $0=K(-K)+(-K)^2=2g-2$. Generically a holomorphic bundle on an elliptic curve is a sum of line bundles, and then the dimension of $H^0(D,\mathop{\rm End}\nolimits_0 E)$ is $k-1$ if $\mathop{\rm rk}\nolimits E=k$. Thus the real dimension of $\mathop{\rm ker}\nolimits[I_+,I_-]$ is at least $2(k-1)$. \vskip .25cm Now the divisor $D$ is, by definition, the subset of $M$ on which $I_+=\pm I_-$, say $I_+=I_-$. Thus the complex structure of the bundle $E$ determined by its ASD connection is the \emph{same} on $D$ for $I_+$ and $I_-$. So the same holomorphic section $u$ of $\mathop{\rm End}\nolimits_0 E$ on $D$ maps complex linearly in two different ways to the cotangent space of ${\mathcal M}$. To study these maps we should really say that there are real isomorphisms $$\alpha_{\pm}: H_{\pm}^1(M,\mathop{\rm End}\nolimits E\otimes K)\rightarrow T^*_{[A]}$$ such that $\alpha_{\pm}$ is $I_{\pm}$-complex linear. \begin{prp} \label{residue} $\alpha_+\delta_+=\alpha_-\delta_-$ \end{prp} \begin{prf} Recall how the connecting homomorphism is defined in Dolbeault terms, for the moment in the case where $D$ has multiplicity one: we have a holomorphic section $u$ of $\mathop{\rm End}\nolimits_0 E$ on $D$, and then extend using a partition of unity to a $C^{\infty}$ section $\tilde u$ on $M$. Then since $u$ is holomorphic on $D$, $\bar\partial \tilde u$ is divisible by $s$, the section of $K^*$ whose divisor is $D$. Then $\delta(u)$ is represented by the $(2,1)$-form $s^{-1}\bar\partial_A\tilde u$. Let $a\in T_{[A]}$ be a tangent vector to the moduli space, so $a\in\Omega^1(M,\mathop{\rm End}\nolimits E)$ and satisfies $d_A^+a=0$. So $\bar\partial_A a^{0,1}=0$ and we evaluate the cotangent vector $\delta_+(u)$ on $a$ to get $$\int_M\mathop{\rm tr}\nolimits(s^{-1}\bar\partial_A\tilde u \wedge a).$$ But $s^{-1}=\gamma/2i$ so this is $$\frac{1}{2i}\int_M\gamma\mathop{\rm tr}\nolimits(\bar\partial_A\tilde u\wedge a).$$ Away from the divisor $D$, we have $$\bar\partial(\gamma\wedge\mathop{\rm tr}\nolimits(\tilde u a))=\gamma\wedge \mathop{\rm tr}\nolimits(\bar\partial_A\tilde u \wedge a)$$ since both $\gamma$ and $a$ are $\bar\partial$-closed. By Stokes' theorem the integral is reduced to an integral around the unit circle bundle of the normal bundle of $D$ and from there to an integral over $D$. In fact, if $\gamma$ has a simple pole along $D$ then locally $$\gamma=f(z_1,z_2)\frac{dz_1\wedge dz_2}{z_1}$$ where $z_1=0$ is the equation of $D$. The holomorphic one-form $f(0,z_2)dz_2$ is then globally defined on $D$ -- the \emph{residue} $\gamma_0$ of the meromorphic $2$-form. This residue is the same for $I_+$ and $I_-$ (from Proposition \ref{4formulas} the meromorphic form for $I_-$ is $-2i\omega_+-(p-1)/2\gamma+(p+1)/2\bar\gamma$ and $p=-1$ on $D$). Thus the integral becomes $$\frac{1}{2i}\int_D\gamma_0\wedge \mathop{\rm tr}\nolimits(ua).$$ This is defined entirely in terms of the data on $D$ and so is the same for $I_+$ and $I_-$. \vskip .25cm When the divisor has multiplicity $d$, the section $u$ extends holomorphically to the $(d-1)$-fold formal neighbourhood of the curve and our $C^{\infty}$ extension must agree with this. The result remains true. (Note that the discussion of Poisson surfaces and moduli spaces via the residue is the point of view advanced in Khesin's work \cite{Kh}.) \end{prf} \vskip .25cm \begin{cor} The two real Poisson structures $\pi_1,\pi_2$ defined by the generalized complex structures $J_1,J_2$ on the moduli space ${\mathcal M}$ of $SU(k)$ instantons have kernels of dimension $0$ and $\ge 2(k-1)$. \end{cor} \begin{prf} We saw at the beginning of the Section that if $I_+a=I_-a$ then $[I_+,I_-]a=0$. Proposition \ref{residue} shows that $I_+$ and $I_-$ agree on the kernel of $[I_+,I_-]$, so that $\mathop{\rm ker}\nolimits(I_+-I_-)=\mathop{\rm ker}\nolimits[I_+,I_-]$. Now $\mathop{\rm ker}\nolimits(I_+-I_-)$ is the kernel of Poisson structure $\pi_1$ say, which is isomorphic to $H^0(D,\mathop{\rm End}\nolimits_0 E)$ and has, as we have seen, at least $2(k-1)$ real dimensions. The other Poisson structure $\pi_2$ has kernel $\mathop{\rm ker}\nolimits(I_++I_-)$. But this also lies in the kernel of $[I_+,I_-]$ so $I_+a=I_-a$. With $I_+a=-I_-a$ this means $a=0$. \end{prf} \vskip .25cm The generalized complex structure $J_2$ on ${\mathcal M}$ where the kernel of the Poisson structure is zero is therefore of the form $\exp(B+i\omega)$ and it is tempting to associate it to the generalized complex structure of symplectic type on $M^4$. However, as we have seen, there appears to be no way to naturally associate or even define these structures, since the $2$-form $b$ does not descend to the moduli space. \subsection{Examples of symplectic leaves} We saw in the previous section that the symplectic leaves of $\pi_1$ are the same as the the symplectic leaves of the canonical complex Poisson structure on ${\mathcal M}$. The simplest example is to take ${\mathbf C}{\rm P}^2$ with the anticanonical divisor defined by a triple line : $D=3L$. The moduli space of stable rank $2$ bundles with $c_2=2$ has dimension $4\times 2-3=5$ and has a very concrete description. Such a bundle $E$ is trivial on a general projective line but jumps to ${\mathcal O}(1)\oplus {\mathcal O}(-1)$ on the lines which are tangent to a nonsingular conic $C_E$. The moduli space ${\mathcal M}$ is then just the space of non-singular conics, which is a homogeneous space of $PGL(3,\mathbf{C})$. The subgroup preserving $L$ (the line at infinity say) is the affine group $A(2)$ and if it preserves the Poisson structure it fixes $dz_1\wedge dz_2$. Hence the $5$-dimensional unimodular affine group $SA(2)$ acts on ${\mathcal M}$ preserving the Poisson structure. The subgroup $G$ which fixes the conic $z_1z_2=a$ consists of the transformations $(z_1,z_2)\mapsto (\lambda z_1,\lambda^{-1}z_2)$ so for each $a$, the orbit of the conic under $SA(2)$ is isomorphic to the $4$-dimensional quotient $SA(2)/G$. These orbits are the generic symplectic leaves of the Poisson structure, and thus are homogeneous symplectic and hence isomorphic to coadjoint orbits. In fact if $z\mapsto Az+b$ is in the Lie algebra of $SA(2)$ then $G$ is the stabilizer of the linear map $f(A,b)=A_{11}$ so that $SA(2)/G$ is the orbit of $f$ in the dual of the Lie algebra. This deals with conics which meet $L$ in two points. The ones which are tangential to $L$ (i.e. the bundles for which $L$ is a jumping line) are parabolas: e.g. $z_1^2=z_2$. The identity component of the stabilizer of this is the one-dimensional group $(z_1,z_2)\mapsto (z_1+c,2cz_1+z_2+c^2)$ and this stabilizes the linear map $(A,b)\mapsto A_{21}+4b_1$, so we again have a coadjoint orbit. \vskip .25cm In general, the symplectic leaves are roughly given by the bundles $E$ on $M$ which restrict to the same bundle on the anticanonical divisor $D$. ``Roughly", because we are looking at equivalence classes and a stable bundle on $M$ may not restrict to a stable bundle on $D$, so there may not be a well-defined map from ${\mathcal M}$ to a Hausdorff moduli space. On the other hand this is the quotient space of a (singular) foliation so we don't expect that. When $D$ is the triple line $3L$ in ${\mathbf C}{\rm P}^2$ there is an alternative way of describing these leaves. On a generic line $E$ is trivial and the sections along that line define the fibre of a vector bundle $F$ on the dual plane, outside the curve $J$ of jumping lines. If we take a section of $E$ on $L$ we can try and extend it to the first order neighbourhood of $L$. Since the normal bundle to $L$ is ${\mathcal O}(1)$ there is an exact sequence of sheaves for sections on the $n$-th order neighbourhood: $$0\rightarrow {\mathcal O}(E(-n))\rightarrow {\mathcal O}^{(n)}(E)\rightarrow {\mathcal O}^{(n-1)}(E)\rightarrow 0.$$ Since $H^0({\mathbf C}{\rm P}^1, {\mathcal O}(-1))=H^1({\mathbf C}{\rm P}^1, {\mathcal O}(-1))=0$, any section has a unique extension to the first order neighbourhood: this defines a \emph{connection} on $F$. The extension to the second order neighbourhood is obstructed since $H^1({\mathbf C}{\rm P}^1, {\mathcal O}(-2))\cong \mathbf{C}$ and this obstruction is the \emph{curvature} of the connection (see \cite{Hurt} for details of this twistorial construction). What it means is that if $L$ is not a jumping line, then $E$ restricted to $3L$ is essentially the curvature of the connection on $F$ at the point $\ell$ in the dual plane defined by the line $L$, and the symplectic leaves are obtained by fixing the equivalence class of the curvature at that point. The curvature acquires a double pole on $J$. From this point of view, the case $k=2,c_1=0,c_2=2$ concerns an $SO(3,\mathbf{C})$-invariant connection on a rank $2$-bundle on the complement of a conic, and this is essentially the Levi-Civita connection of ${\mathbf R}{\rm P}^2$ complexified. This is an $O(2)$-connection which becomes an $SO(2)$ connection on $S^2$ with curvature $$\frac{dz\wedge d\bar z}{(1+\vert z\vert^2)^2}.$$ So the bundle on $D$ is equivalent to the transform of the complexification of this by a projective transformation. If the dual conic is defined by the symmetric $3\times 3$ matrix $Q_{ij}$ and $x$ is a vector representing $\ell$ then the curvature is $$\frac{(\det Q)^{2/3}}{Q(x,x)^2}.$$ The symplectic leaves are then given by the equation $\det Q=aQ(x,x)^3$ for varying $a$. \section{A quotient construction} It is well-known that the moduli space of instantons on a hyperk\"ahler $4$-manifold is hyperk\"ahler and this can be viewed as an example in infinite dimensions of a hyperk\"ahler quotient -- the quotient of the space of all connections by the action of the group of gauge transformations. One may ask if, instead of the painful integration by parts that we did in the previous sections, there is a cleaner way of viewing the definition of a generalized K\"ahler structure on ${\mathcal M}$. The problem is that such a quotient would have to encompass not only the hyperk\"ahler quotient but also the ordinary K\"ahler quotient, and in finite dimensions these are very different -- the dimension of the quotient in particular is different! We offer next an example of a generalized K\"ahler quotient which could be adapted to replace the differential geometric arguments in the previous sections for the case of a torus or K3, and at least gives another reason why the calculations should hold. It also brings out in a natural way the frustrating feature that the $2$-form $b$ does not descend in general to the quotient. \vskip .25cm We suppose the generalized K\"ahler structure is even and is given by global forms $\rho_1=\exp {\beta_1}, \rho_2=\exp{\beta_2}$ where $\beta_1,\beta_2$ are closed complex forms on a real manifold $M$ of dimension $4k$. This is the test situation we have been considering throughout this paper. From Lemma \ref{commute}, the compatibility ($J_1J_2=J_2J_1$) is equivalent to $$(\beta_1-\beta_2)^{k+1}=0,\qquad (\beta_1-\bar\beta_2)^{k+1}=0.$$ Now suppose a Lie group $G$ acts, preserving the forms $\beta_1,\beta_2$, and giving complex moment maps $\mu_1,\mu_2$. To make a quotient, we would like to take the joint zero set of $\mu_1$ and $\mu_2$ and divide by the group $G$, but these are two \emph{complex} functions so if they were generic we would get as a quotient a manifold of dimension $\dim M-5\dim G$ instead of $\dim M-4\dim G$. To avoid this, we need to assume that $\beta_1,\beta_2,\bar\beta_1,\bar\beta_2$ are linearly dependent over $\mathbf{R}$. \begin{rmk} If we were trying to set up the moduli space of instantons as a quotient of the space of all connections on a K3 surface or a torus, the following lemma links the condition of linear dependence of the moment maps to the necessity to choose a self-dual $b$. \begin{lem} If $\dim M=4$, then $\beta_1,\beta_2,\bar\beta_1,\bar\beta_2$ are linearly independent over $\mathbf{R}$ at each point if and only if $b$ is self-dual. \end{lem} \begin{lemprf} From Proposition \ref{4formulas} we have \begin{eqnarray*} \beta_1+\bar\beta_1&=&2b-(p-1)(\gamma+\bar\gamma)/2\\ \beta_2+\bar\beta_2&=&2b-(p+1)(\gamma+\bar\gamma)/2\\ -i(\beta_1-\bar\beta_1)&=&2\omega_+-(p-1)i(\gamma-\bar\gamma)/2\\ -i(\beta_2-\bar\beta_2)&=&2\omega_+-(p+1)i(\gamma-\bar\gamma)/2\\ \end{eqnarray*} We can easily solve these for $b,\omega_+,\gamma+\bar\gamma,i(\gamma-\bar\gamma)$ in terms of the $\beta_i$. If $b$ is self-dual, it is a real linear combination of $\omega_+,\gamma+\bar\gamma,i(\gamma-\bar\gamma)$ since $\gamma$ is of type $(2,0)$ relative to $I_+$, hence we get a linear relation amongst the left hand sides. Conversely, a linear relation among the left hand sides will express $b$ in terms of $\omega_+,\gamma+\bar\gamma,i(\gamma-\bar\gamma)$ unless it is of the form $$(\beta_1+\bar\beta_1)-(\beta_2+\bar\beta_2)+i\lambda(\beta_1-\bar\beta_1)+i\mu (\beta_2-\bar\beta_2)=0.$$ But the $(1,1)$ component of this is $2(\mu-\lambda)\omega_+$ so $\lambda=\mu$ and then the relation can be written $$(1+i\lambda)(\beta_1-\bar\beta_2)+(1-i\lambda)(\bar\beta_1-\beta_2)=0$$ but $(\bar\beta_1-\beta_2)$ is of type $(2,0)$ relative to $I_-$ so this is impossible. \end{lemprf} We see that the condition for $b$ to define $\tilde b$ on the moduli space ${\mathcal M}$ is related to the linear dependence issue of the moment maps. \end{rmk} Returning to the general case, for each vector field $X$ from the Lie algebra of $G$ we have $i_X\beta_i=d\mu_i$ and so $\beta_i$ restricted to $\mu_1=0=\mu_2$ is annihilated by $X$, and invariant under the group and hence is the pullback of a form $\tilde\beta_i$ on the quotient, which is also closed. In the bihermitian interpretation, $\bar\beta_1-\bar\beta_2$ is a non-degenerate $(2,0)$-form relative to $I_+$ -- a holomorphic symplectic form -- and the quotient can then be identified with the holomorphic symplectic quotient. In particular if the complex dimension of the quotient is $2m$ then $(\tilde\beta_1-\tilde\beta_2)^{m+1}=0$ and $(\tilde\beta_1-\tilde\beta_2)^{m}\ne 0$. Similarly $(\bar\beta_1-\beta_2)$ is $(2,0)$ with respect to $I_-$ and we get the same property for $(\tilde\beta_1-\bar{\tilde \beta_2})$. From Lemma \ref{commute} we have a generalized K\"ahler structure on the quotient. \vskip .25cm Note that in this generic case the Poisson structures on the quotient are non-degenerate.
{ "timestamp": "2005-03-21T18:32:38", "yymm": "0503", "arxiv_id": "math/0503432", "language": "en", "url": "https://arxiv.org/abs/math/0503432" }
\section{Introduction} \label{sec:level1} One of the most exciting subject in theoretical nuclear physics is the double beta decay, especially due to the neutrino-less ($0\nu\beta\beta$) process \cite{Hax,Fas1,Fas2,Suh}. Indeed, its discovery would answer a fundamental question whether neutrino is a Majorana or a Dirac particle. The theories devoted to the description of this process suffer of the lack of reliable tests for the nuclear matrix elements. O possibility to overcome such difficulties would be to use the matrix elements which describe realistically the rate of $2\nu\beta\beta$ decay. In this context many theoretical work have been focussed on $2\nu\beta\beta$ process. Most formalisms are based on the proton-neutron quasiparticle random phase approximation (pnQRPA) which includes the particle-particle ($pp$) channel in the two body interaction. Since such an interaction is not considered in the mean field equations the approach fails at a critical value of the interaction strength, $g_{pp}$. Before this value is reached, the Gamow-Teller transition amplitude ($M_{GT}$) is decreasing rapidly and after a short interval is becoming equal to zero. The experimental data for this amplitude is reached for a value of $g_{pp}$ close to that one which vanishes $M_{GT}$ and also close to the critical value. Along the time, the instability of the pnQRPA ground state was considered in different approaches. The first formalism devoted to this feature includes anharmonicities through the boson expansion technique \cite{Rad1,Rad2,Suh1,Grif}. Another method is the renormalized pnQRPA procedure (pnQRRPA) \cite{Toi} which keeps the harmonic picture but the actual boson is renormalized by effects coming from the terms of the commutators algebra, which are not taken into account in the standard pnQRPA approach. In a previous paper\cite{Rad3} we have proved that the pnQRRPA procedure does not include the additional effects in an consistent way. Indeed, if the commutators of two quasiparticle operators involves the average of monopole terms then these terms should be considered also in the commutators of the scattering terms. If one does so, new degrees of freedom are switched on and a new pnQRPA boson can be defined. This contains, besides the standard two quasiparticle operators, the proton-neutron quasiparticles scattering terms. If the amplitude of the scattering term is dominant comparing it to the other amplitudes, the pnQRRPA phonon describes a new nuclear state. The aim of this paper is to show that such a mode appears in a natural way within a time dependent treatment. The present approach points out new properties of the new proton-neutron collective mode. We use a schematic many-body Hamiltonian which for a single j-shell is exactly solvable. In this way the approximations might be judged by comparing the predictions of the actual model with the corresponding exact results. Since the semi-classical treatment is the proper way to determine the mean field, one expects that the present approach is suitable to account for ground state correlations in a consistent way and therefore some of the drawbacks mentioned in a previous publication \cite{Rad3}, like the breaking down of the fully renormalized RPA before the standard RPA breaks down, are removed. To understand better the virtues of the present model we compare its predictions with the results obtained in a renormalized RPA approach and a boson expansion formalism. Since the semi-classical methods have, sometimes, intuitive grounds we aim at obtaining a clear interpretation for the new proton-neutron mode. It is well known that the breaking down of the RPA approach is associated to a phase transition. In this respect the semi-classical formalism is a suitable framework to define the nuclear phases which are bridged by the Goldstone mode. Above arguments justify our option for a semi-classical treatment and also sketch a set of expectations. This project is achieved according to the following plan: In Section II, we describe the model Hamiltonian. The main features of the fully renormalized RPA approach, presented in a previous paper, are briefly reviewed. A time dependent variational principle is formulated in connection with a truncated quasiparticle Hamiltonian, in Section III. This Hamiltonian is the term of the model Hamiltonian which determines the equations of motion for the quasiparticle proton-neutron scattering terms, in the de-coupling regime. The classical equations of motion and their solutions are presented in Section IV. The new $pn$ collective mode is alternatively described through the renormalized RPA approach and boson expansion formalism in Section V. Numerical results are analyzed in Section VI while the final conclusions are given in Section VII. \section{The model Hamiltonian. Brief review of frn-RPA } \label{sec:level2} Since we are not going to describe realistically some experimental data but to stress on some specific features of a heterogeneous many nucleon system with proton-neutron interaction, we consider a schematic Hamiltonian which is very often \cite{Sam,Rad4} used to study the single and double beta Fermi transitions: \begin{eqnarray} H & = & \sum_{jm}{(\varepsilon_{pj}-\lambda_{p})c^{\dag}_{pjm}c_{pjm}} +\sum_{jm}{(\varepsilon_{nj}-\lambda_{n})c^{\dag}_{njm}c_{njm}}\nonumber\\ &&-\frac{G_{p}}{4}\sum_{jm,j'm'}{c^{\dag}_{pjm} c^{\dag}_{\widetilde{pjm}}c_{\widetilde{pj'm'}}c_{pj'm'}} -\frac{G_{n}}{4}\sum_{jm,j'm'}{c^{\dag}_{njm}c^{\dag}_{\widetilde{njm}} c_{\widetilde{nj'm'}}c_{nj'm'}}\nonumber\\ &&+\chi\sum_{jm,j'm'}{c^{\dag}_{pjm}c_{njm}c^{\dag}_{nj'm'}c_{pj'm'}} -\chi_{1}\sum_{jm,j'm'}{c^{\dag}_{pjm}c^{\dag}_{\widetilde{njm}} c_{\widetilde{nj'm'}}c_{pj'm'}}. \end{eqnarray} $c^{\dag}_{\tau jm}(c_{\tau jm})$ denotes the creation (annihilation) of a $\tau(=p,n)$ nucleon in a spherical shell model state $|\tau;nljm\rangle =|\tau jm\rangle$ with $\tau$ taking the values $p$ for protons and $n$ for neutrons, respectively. The time reversed state corresponding to $|\tau jm\rangle$ is $|\tau{\widetilde{jm}}\rangle=(-)^{j-m}|\tau j-m\rangle$ For what follows it is useful to introduce the quasiparticle ($qp$) representation, defined by the Bogoliubov-Valatin (BV) transformation: \begin{eqnarray} a^{\dag}_{pjm}& = & U_{pj}c^{\dag}_{pjm}-V_{pj}c_{\widetilde{pjm}},\;\; a_{pjm} = U_{pj}c_{pjm}-V^{*}_{pj}c^{\dag}_{\widetilde{pjm}}, \nonumber\\ a^{\dag}_{njm}& = & U_{nj}c^{\dag}_{njm}-V_{nj}c_{\widetilde{njm}},\;\; a_{njm} = U_{nj}c_{njm}-V^{*}_{nj}c^{\dag}_{\widetilde{njm}} . \end{eqnarray} which quasi-diagonalizes the first four terms, i.e in the new representation they are replaced by a set of independent quasiparticles of energies: \begin{equation} E_{\tau}=\sqrt{(\epsilon_{\tau}-\lambda_{\tau})^{2}+\Delta_{\tau}^{2}}. \end{equation} In the new $qp$ representation, the model Hamiltonian, denoted by $H_q$, describes a set of independent quasiparticles, interacting among themselves through a two body interaction determined by the images of the $\chi$ and $\chi_1$ terms through the BV transformation. Various many-body approaches have been tested by using not the $qp$ image of $H$ but another Hamiltonian derived from $H$ by ignoring the scattering $qp$ terms: \begin{eqnarray} B^{\dag}(jpn)&=&\sum_{m}a^{\dag}_{pjm}a_{njm}, \nonumber\\ B(jpn)& = &\sum_{m}a^{\dag}_{njm}a_{pjm}. \end{eqnarray} and restricting the space of single particle states to a single $j$-state. Thus, the model Hamiltonian contains, besides the terms for the $qp$ independent motion, a two body term which is quadratic in the two quasiparticle operators $A^{\dag}, A$: \begin{equation} A^{\dag}(jpn)=\sum_{m}a^{\dag}_{pjm}a^{\dag}_{\widetilde{njm}}, \;\;A(jpn)=(A^{\dag}(jpn))^{\dag}. \end{equation} In a previous publication\cite{Rad3}, we showed that going beyond the quasiparticle random phase approximation (pnQRPA) through a renormalization procedure, a new degree of freedom is switched on, which results in having a renormalized pnQRPA boson operator as a superposition of the operators $A^{\dag}(jpn), A(jpn)$ and scattering terms $B^{\dag}(jpn),B(jpn)$. This picture differs from the standard $pnQRRPA$ approach, where the boson operators involve only the operators $A^{\dag}$ and $A$, and is conventionally called as fully renormalized RPA ($frn-RPA$). Obviously, when the amplitudes of scattering terms are dominant, one deals with a new kind of collective $pn$ excitation. In order to define clearly the distinct features of the new proton-neutron ($pn$) mode revealed in the present paper a brief description of the results obtained in a previous publication \cite{Rad3} is necessary. The equations of motion associated to the many-body Hamiltonian, written in terms of quasiparticle operators, are determined by the commutators algebra of the two quasiparticle ($A^{\dag}, A$) and scattering $(B^{\dag},B)$ operators defined by eqs. (2.5) and (2.4) respectively. Within the $frn-RPA$, the exact commutators are approximated as follows: \begin{eqnarray} \left[A(jpn),A^{\dag}(jpn)\right ] &=& C^{(1)}_{jpn}, \nonumber \\ \left[B(jpn),B^{\dag}(jpn)\right ] &=& C^{(2)}_{jpn}, \nonumber\\ \left[A(jpn),B^{\dag}(jpn)\right ] &=&\left[A(jpn),B(jpn)\right ] = 0. \end{eqnarray} The terms $C^{(1)}_{jpn}, C^{(2)}_{jpn}$ appearing in the r.h.s. of the above equations are the averages of the corresponding exact commutators, on the correlated ground state $|0>$: \begin{equation} C^{(1)}_{jpn}=\langle0|1-\hat{N}_{jn}-\hat{N}_{jp}|0\rangle,\: \: C^{(2)}_{jpn}=\langle0|\hat{N}_{jn}-\hat{N}_{jp}|0\rangle. \end{equation} with $\hat{N}_{j\tau}$ standing for the $\tau$ (=p,n) quasiparticle number operator in the shell j. The normalized operators \begin{eqnarray} \bar{A}^{\dag}(jpn)=\frac{1}{\sqrt{C^{(1)}_{jpn}}}A^{\dag}(jpn),\: \bar{A}(jpn)=\left(\bar{A}^{\dag}(jpn)\right)^{\dag}, \nonumber\\ \bar{B}^{\dag}(jpn)=\frac{1}{\sqrt{|C^{(2)}_{jpn}|}}B^{\dag}(jpn),\: \bar{B}(jpn)=\left(\bar{B}^{\dag}(jpn)\right)^{\dag}, \end{eqnarray} satisfy bosonic commutation relations and thereby their equations of motion are linear: \begin{equation} \left[H_q,\left(\matrix{\bar{A}^{\dag}(jpn) \cr \bar{A}(jpn) \cr \bar{B}^{\dag}(jpn) \cr \bar{A}(jpn)}\right)\right] =\sum_{j,j^{\prime}}T^{j,j^{\prime}}\left(\matrix{\bar{A}^{\dag}(jpn) \cr \bar{A}(jpn) \cr \bar{B}^{\dag}(jpn) \cr \bar{B}(jpn)}\right). \end{equation} The matrix $T^{j,j^{\prime}}$ depends on the U and V coefficients as well as on the strengths $\chi, \chi_1$ of the two body interactions. The $frn-RPA$ approach defines a linear combination of the basic operators $\bar{A}^{\dag}(jpn), \bar{A}(jpn), \bar{B}^{\dag}(jpn), \bar{A}(jpn)$, \begin{equation} \Gamma^{\dag}=\sum_{j}\left[X(j)\bar{A}^{\dag}(jpn)+Z(j)D^{\dag}(jpn) -Y(j)\bar{A}(jpn)-W(j)D(jpn)\right], \end{equation} so that the following commutation relations with its hermitian conjugate operator and the model Hamiltonian hold: \begin{eqnarray} \left[\Gamma,\Gamma^{\dag}\right] &=& 1, \\ \left[H_q,\Gamma^{\dag}\right] &=& \omega \Gamma^{\dag}. \end{eqnarray} The operators $D^{\dag}(jpn)$ are identical with ${\bar{B}}^{\dag}(jpn)$ or $\bar{B}(jpn)$ depending on whether the sign of $C^{(2)}_{jpn}$ is plus or minus. The equation (2.12) provides a set of homogeneous equations- called the $frn-RPA$ equations- for the amplitudes $X,Y,Z,W$: \begin{eqnarray} \left(\matrix{{\cal A} & {\cal B}\cr -{\cal B} &-{\cal A}}\right)\left(\matrix{X\cr Z\cr Y\cr W}\right) =\omega \left(\matrix{X\cr Z\cr Y\cr W}\right), \end{eqnarray} while the equation (2.11) yields the normalization equation \begin{equation} \sum_{j}(X^2(j)+Z^2(j)-Y^2(j)-W^2(j))=1. \end{equation} The $frn-RPA$ matrices depend on the renormalization constants $C^{(1)}, C^{(2)}$ which, at their turn, depend on the phonon amplitudes. Therefore, the equations (2.12) and (2.7) should be self-consistently solved. In ref. \cite{Rad3} the $frn-RPA$ equations have been solved both for a proton-neutron dipole-dipole interaction, needed for the description of the double beta Gamow-Teller decay and for a proton-neutron monopole-monopole interaction used in the calculation of the rates of the double beta Fermi decay. Equations obtained in the two cases have some common features which, for what follows, are worth being enumerated. \noindent 1) The dimension of the $frn-RPA$ matrix is twice as large as that of the standard RPA and consequently new solutions show up. \noindent 2) The solutions characterized by that the largest phonon amplitude is of type Z define a new class of proton-neutron excitations. \noindent 3) Due to the attractive character of the two body interaction in the particle-particle ($pp$) channel, the lowest new state has an energy which is smaller than the minimal absolute value of the relative energy of the proton and neutron quasiparticle partner states, related by the operators $B^{\dag}(jpn), B(jpn)$. \noindent 4) For the N=Z nuclei, this minimal value is vanishing and therefore the lowest mode becomes spurious or in other words saying a new symmetry is open. The new symmetry corresponds to the restriction $C^{(2)}_{jpn}=0$, i. e. the average of the third component of the isospin operator is vanishing. This means that the system is invariant to rotations around any axes in the ($X,Y$) plane of the isospin space associated to the (jpn) orbits. \noindent 5) Important quantitative effects are expected for heavy nuclei having the proton and neutron Fermi energies lying far apart from each other. \noindent 6) The presence of the additional states influences also the structure of the states lying close to those predicted by the standard RPA. Indeed, the actual normalization condition for the phonon amplitudes implies new values for the X and Y weights. Consequently, the strengths for $\beta^-$ and $\beta^+$ transitions are shared by the "old"-lying close to the standard RPA states- and the ``new'' states- for which the amplitudes Z are dominant. \noindent 7) The standard RPA approach is based on the quasi-boson approximation and therefore it ignores some important dynamic effects (only the terms $A^{\dag}A^{\dag}, A^{\dag}A, AA$ are considered in an approximative manner) and moreover the Pauli principle is violated. By contrast, within the $frn-RPA$ all the terms of the model Hamiltonian are taken into account. Also the Pauli principle is, to a certain extent, restored. Due to this feature, large corrections to the double beta transition amplitude as well as to the Ikeda sum rule are expected by changing the RPA to the frn-RPA. \noindent 8) The equations of motion for the $A^{\dag },A$ and $B^{\dag}, B$ operators are coupled by the terms $A^{\dag}B^{\dag}, A^{\dag}B, AB^{\dag}, AB$ involved in the quasiparticle Hamiltonian. These terms are multiplied by the factors $ U_pV_nU_{p^{\prime}}U_{n^{\prime}}, U_pV_nV_{p^{\prime}}V_{n^{\prime}}, V_pU_nU_{p^{\prime}}U_{n^{\prime}}, V_pU_nV_{p^{\prime}}V_{n^{\prime}}$ in the $ph-ph$ interaction (the $\chi$ term) and by $U_pU_nU_{p^{\prime}}V_{n^{\prime}}, U_pU_nV_{p^{\prime}}U_{n^{\prime}}, V_pV_nU_{p^{\prime}}V_{n^{\prime}}, V_pV_nV_{p^{\prime}}U_{n^{\prime}}$ in the $pp-hh$ interaction (the $\chi_1$ term), respectively. Note that the coupling terms change the number of either proton or neutron quasiparticles by two units. The terms bringing the main contribution to the equations of motion for the operators $A^{\dag}, A$ commute with $ \hat{N}_{jp}-\hat{N}_{jn}$ but not with $ \hat{N}_{jp}+\hat{N}_{jn}$. By contrary the terms having the dominant contribution to the equations of motion for the operators $B^{\dag},B$ commute with $ \hat{N}_{jp}+\hat{N}_{jn}$ and not with $ \hat{N}_{jp}-\hat{N}_{jn}$. None of the two operators, $ \hat{N}_{jp}-\hat{N}_{jn}$, $ \hat{N}_{jp}+\hat{N}_{jn}$, commutes with the coupling terms. Retaining from the $\chi$-interaction the $pp-hh$ terms (those multiplied by $U_pU_nV_{p^{\prime}}V_{n^{\prime}}$) and from the $\chi_1$ interaction only the $ph-ph$ terms (those proportional to $U_pV_nU_{p^{\prime}}V_{n^\prime} $ ) the equations of motion for the operators $B^{\dag},B$ are decoupled from those for $A^{\dag}$ and $A$. One may conclude that the new mode is determined by a combined effect coming from the $pp-hh$ and $ph-ph$ terms belonging to the $\chi$ and $\chi_1$ interactions, respectively. \noindent 9) In the particle representation the $frn-RPA$ phonon operator is a linear superposition of $ph, hp, pp$ and $hh$ operators. \noindent 10) In the limit of large $pp$ and negligible $ph$ interactions, the amplitudes $Z$ can be analytically calculated. The result is that $Z$ is proportional to either $U_pV_n$ or $V_pU_n$, depending on whether the sign of $E_p-E_n$ is plus or minus, respectively. When this amplitude prevails over the other ones, the corresponding mode describes a neutron-hole proton-particle (or a proton-hole neutron-particle) excitation of the mother nucleus $(N,Z)$. Therefore the state is associated to the $(N-1,Z+1)$ (or to the $(N+1,Z-1)$) nucleus. In this case the state might be reached by exciting the ground state through the transition operator $c^{\dag}_pc_n$ (or $c^{\dag}_nc_p$), which is typical for the $\beta^-$ ( or $\beta^+$) decay. Since the double beta decay is conceived as taking place through two successive $\beta^-$ transitions, one expects that this process is also influenced by considering this new state as an intermediate state characterizing the odd-odd neighboring nucleus. \noindent 11) When the $pp$ interaction is small the amplitude Z is proportional to $U_pU_n$ if $E_p>E_n$ or to $V_pV_n$ in the case $E_p<E_n$. The new mode characterizes the nucleus $(N+1,Z+1)$ in the first case and the nucleus ($N-1,Z-1$) in the second situation. The transition operators which could excite these states are obviously of the types $c^{\dag}_pc^{\dag}_n$ and $c_nc_p$, respectively. Note that the restriction of the phonon operator to the scattering terms resembles the standard RPA boson operator written in the particle representation. This comparison has, however, only a formal value since in the quasiparticle representation there is no Fermi energy and therefore one cannot speak about quasiparticle-quasihole excitations. Similar features are met in solid state physics for the description of electron excitations in narrow energy bands, spin waves and plasma oscillations \cite{Hub}. In nuclear physics, the scattering terms have been also considered but not for proton-neutron excitations. Indeed, using the thermal response theory, Tanabe \cite{Tan} studied the charge conserving phonons in nuclear systems at a finite temperature. It seems that the contribution of the scattering terms to the charge conserving bosons, does not survive at vanishing temperature \cite{Hat}. Moreover, the dispersion relation for the mode energy cannot be obtained from a linearized set of equations as it is required by the spirit of the RPA approach. \section{Semi-classical treatment} \label{sec:level3} As we already mentioned, the scope of the present paper is to study the $pn$ mode caused by the quasiparticle scattering terms within a semi-classical approach. In this formalism the renormalization condition (2.6) is missing and therefore the harmonic motion of the new degrees of freedom hinges on a more physical ground. Moreover, we address the question whether this mode survives when the non-scattering terms are switched off. Thus, it is worth to know if such a mode appears only when the scattering terms accompany the two quasipatricle operators or it might be determined by the scattering terms alone. From the brief presentation of the previous Section it is clear that the mode does not appear within the RPA approach. Indeed, it occurred within the $frn-RPA$ after a consistent renormalization was performed (i. e. not only the operators $A^{\dag}, A$ where renormalized but also $B^{\dag}$ and $B$). If that mode is a signature of the higher RPA formalisms, then it should also appear within the semi-classical formalism as well as in the boson expansion framework. As we shall see the semi-classical approach is able to predict the mode even in the harmonic approximation, the mode being associated with the small oscillations of the system around a static correlated ground state. Moreover, the semi-classical frame is expected to allow us an intuitive interpretation of this new type of excitation. We recall that the higher order corrections to the standard RPA approach are frequently studied, with different purposes, using a single j case and ignoring the scattering terms. The procedure has the advantage that the resulting Hamiltonian is exactly solvable. Therefore the quality of the adopted approximations may be tested by comparing the predictions with the corresponding exact results. To touch the goal of the present paper we adopt a similar point of view. Indeed, if the coupling terms (mentioned at the point 8 of the previous section) are ignored, the equations of motion for the scattering operators are decoupled. Moreover the motion of these operators is determined also by an exactly solvable Hamiltonian, which reads: \begin{equation} H^{(q)}_{pn}=E^{'}_{p}{\hat N}_{p}+E^{'}_{n}{\hat N}_{n}+\lambda_{1} B^{\dag}(pn)B(pn)+\lambda_{2}\big(B^{\dag2}(pn)+B^{2}(pn)\big), \end{equation} where the following notations have been used: \begin{eqnarray} E_p^{\prime}&=&E_p+(\chi_1-\chi)V_p^2,\nonumber\\ E_n^{\prime}&=&E_n+(\chi+\chi_1)V^2_p(V_n^2-U_n^2),\nonumber\\ \lambda_1&=&\chi(U_p^2U_n^2+V_p^2V_n^2)-\chi_1(U_p^2V_n^2+V_p^2U_n^2), \nonumber\\ \lambda_2&=&-(\chi+\chi_1)U_pU_nV_pV_n,\\ {\hat N}_{\tau}&=&\sum_{m}a^{\dag}_{\tau jm}a_{\tau jm}. \end{eqnarray} Also, to simplify the notation we omitted the quantum number $j$ for the operators $B^{\dag}(jpn), B(jpn)$ as well as for the $U, V$ coefficients and quasiparticle energies. This model Hamiltonian will be studied within a time dependent variational formalism. Therefore, some static and dynamic properties will be described by solving the equations provided by the time dependent variational principle (TDVP)\footnote{Throughout this paper the units of $\hbar=1$ are used}: \begin{equation} \delta \int_0^t {\langle\Psi|H^{(q)}_{pn}-i\frac{\partial}{\partial t'}|\Psi\rangle} \,dt' =0. \end{equation} If the variational state $|\Psi\rangle$ spans the whole Hilbert space describing the many-body system, solving the equation (3.4) is equivalent to solving the time dependent Schroedinger equation, which would be a very difficult task. In the present paper, the trial function is taken as: \begin{equation} |\Psi\rangle = exp[zB^{\dag}(pn)-z^{*}B(pn)]|NT\; -T\rangle, \end{equation} where $|NTT_{3}\rangle$ denotes the common eigenfunction of the quasiparticle total number (${\hat N}$), the quasiparticle isospin squared (${\hat T}^{2})$, and its z-axis projection ($T_{z}$) operators, respectively. $z$ is a complex function of time and $z^*$, the corresponding complex conjugate function. We justify this choice by the symmetry properties of the model Hamiltonian. Indeed, let us note first that $H^{(q)}_{pn}$ commutes with the quasiparticle total number operator. Moreover, it can be written in terms of the quasiparticle total number operator and generators of the SU(2) isospin algebra \begin{eqnarray} \tau_{+1} & = & -\frac{1}{\sqrt 2}B^{\dag}(pn),\nonumber\\ \tau_{-1} & = & \frac{1}{\sqrt 2}B(pn),\nonumber\\ \tau_{0} & = & \frac{1}{2}(\hat N_{p}-\hat N_{n}). \end{eqnarray} Due to this property of $H^{(q)}_{pn}$, the function $|\Psi\rangle$, which is a coherent state for the SU(2) group, is the most suitable for a semi-classical treatment. Before closing this section we would like to write the trial function in a form which suits better the further purposes. Using the Cambel Hausdorff factorization\cite{Kir} for the exponential function, as explained in Appendix A, one obtains: \begin{eqnarray} |\Psi\rangle & = & {\cal N} e^{\alpha B^{\dag}(pn)}|NT\;-T\rangle, \nonumber\\ {\cal N}& = &(1+\alpha^{*}\alpha)^{-T}. \end{eqnarray} where $\alpha$ depends on the polar coordinates, $z=\rho e^{i\varphi}$: \begin{equation} \alpha=\tan(\rho)e^{i\varphi}. \end{equation} \section{Equations of motion} \label{sec:level4} In order to write the equations of motion provided by the TDVP (3.4), we need the matrix element of $H^{(q)}_{pn}$ as well as of the time derivative operator, $\frac{\partial}{\partial t}$. These can be evaluated by direct calculation, using the expressions (3.5) when the average of $H^{(q)}_{pn}$ is considered and (3.7) for the classical action. The result is: \begin{eqnarray} \langle\Psi|H^{(q)}_{pn}|\Psi\rangle& =& -T(E_{p}^{\prime}-E_{n}^{\prime}+2\lambda_{1}) +\frac{N}{2}(E_{p}^{\prime}+E_{n}^{\prime})+2T(E_{p}^{\prime}-E_{n}^{\prime}+2\lambda_{1}) \frac{\alpha^{*}\alpha}{1+\alpha^{*}\alpha}\nonumber\\ &&+2T(2T-1)\left[\lambda_{1}\frac{\alpha^{*}\alpha} {(1+\alpha^{*}\alpha)^{2}}+\lambda_{2}\frac{\alpha^{*2}+ \alpha^{2}}{(1+\alpha^{*}\alpha)^{2}}\right], \nonumber\\ \langle\Psi|\frac{\partial}{\partial t}|\Psi \rangle &=& T\frac{\alpha^{*}\stackrel{\bullet}{\alpha}-\stackrel{\bullet}{\alpha}^{*} \alpha}{1+\alpha^{*}\alpha}. \end{eqnarray} Considering $\alpha,\alpha^*$ as classical phase space coordinates, the TDVP equation (3.4) yields the following classical equations of motion, describing the nuclear system: \begin{eqnarray} \frac{\partial{\cal H}}{\partial\alpha}&=&-2i \frac{T\stackrel{\bullet}{\alpha}^{*}}{(1+\alpha^{*}\alpha)^{2}}, \nonumber\\ \frac{\partial{\cal H}}{\partial \alpha^{*}}&=&2i \frac{T\stackrel{\bullet}{\alpha}}{(1+\alpha^{*}\alpha)^{2}}. \end{eqnarray} Here ${\cal H}$ denotes the classical energy function: \begin{equation} {\cal H}=\langle\Psi|H^{(q)}_{pn}|\Psi\rangle. \end{equation} In order to quantize the classical trajectories satisfying the equations (4.2) as well as to have an one to one correspondence between the classical and quantal behaviors of the nucleon system, it is convenient to chose those conjugate variables which bring the equations of motion in a canonical Hamilton form. A possible choice of the coordinates with the above mentioned property is \begin{eqnarray} r&=&\frac{2T}{1+\alpha^{*}\alpha},\\ \psi&=&-\frac{1}{2i}(\ln \alpha-\ln \alpha^{*})=-\varphi. \end{eqnarray} Indeed, in the new variables the classical equations read: \begin{eqnarray} \frac{\partial{\cal H}}{\partial r}&=&-\stackrel{\bullet}{\psi},\nonumber\\ \frac{\partial{\cal H}}{\partial \psi}&=&\stackrel{\bullet}{r}. \end{eqnarray} with the classical energy: \begin{eqnarray} {\cal H}&=&T(E_{p}^{\prime}-E_{n}^{\prime}+2\lambda_{1})+\frac{N}{2} (E_{p}^{\prime}+E_{n}^{\prime}) \nonumber\\ &&-(E_{p}^{\prime}-E_{n}^{\prime}+2\lambda_{1})r+\frac{2T-1}{2T}r(2T-r) (\lambda_{1}+2\lambda_{2}\cos2\psi). \end{eqnarray} Note that $r$ has the significance of a generalized coordinate while $\psi$ that of generalized linear momentum. Due to the generalized momentum $\psi $, the equations motion are not linear and therefore analytical solutions are not obtainable. The equations can however be approximatively solved if they are linearized around the minimum point of the energy function: \begin{equation} \stackrel{\circ}{r}=T\left[1-\frac{E_{p}^{\prime}-E_{n}^{\prime}+ 2\lambda_{1}}{(2T-1)(\lambda_{1}-2\lambda_{2})}\right], \;\; \stackrel{\circ}{\psi}=\frac{\pi}{2}. \end{equation} In order that the minimum exits, it is necessary that the generalized coordinates satisfy a consistency condition, required by the definition range of $r$: \begin{equation} 0\leq\stackrel{\circ}{r}\leq2T. \end{equation} By means of (4.8), this provides a constraint for the strengths of the two body interactions. The linearized equations, written in terms of the deviations \begin{equation} q=r-\stackrel{\circ}{r},\; p=\psi-\stackrel{\circ}{\psi}, \end{equation} are of harmonic type: \begin{eqnarray} -\stackrel{\bullet}{p}=-2\frac{2T-1}{2T}(\lambda_{1}-2\lambda_{2})q, \nonumber\\ \stackrel{\bullet}{q}=4\frac{2T-1}{T}\stackrel{\circ}{r} (2T-\stackrel{\circ}{r})\lambda_{2}p. \end{eqnarray} These describe a harmonic motion for the conjugate coordinates, with the angular frequency: \begin{equation} \omega=2\frac{2T-1}{T}[-\lambda_{2}(\lambda_{1}-2\lambda_{2})\stackrel{\circ} {r}(2T-\stackrel{\circ}{r})]^{\frac{1}{2}}. \end{equation} The condition that $\omega$ is a real quantity brings an additional constraint for the strength parameters $\chi,\chi_1$: \begin{equation} \chi\geq\chi_1\left(\frac{U_pV_n-V_pU_n}{U_pU_n+V_pV_n}\right)^2. \end{equation} As we said already before, the schematic model has the advantage, over the realistic formalisms, that allows us to compare the approximative solutions with the exact one. For the particular Hamiltonian used in the present paper, the exact eigenvalues can be obtained by diagonalization in the basis {$|NTM\rangle$}. Indeed, in this basis the model Hamiltonian has the following non-vanishing matrix elements: \begin{eqnarray} \langle NTM|H^{(q)}_{pn}|NTM\rangle &=&\frac{1}{2}(E_p^{\prime}+E_n^{\prime})N +(E_p^{\prime}-E_n^{\prime})M+ \lambda_1(T+M)(T-M+1),\nonumber\\ \langle NTM+2|H^{(q)}_{pn}|NTM\rangle &=&\lambda_2\left[(T-M-1)(T-M)(T+M+1)(T+M+2) \right]^{\frac{1}{2}}, \nonumber\\ \langle NTM|H^{(q)}_{pn}|NTM+2\rangle &=&\langle NTM+2|H^{(q)}_{pn}|NTM\rangle. \end{eqnarray} \section{The renormalized RPA and boson expansion} \label{sec:level5} Within the RPA approach, the renormalization of the quasiparticle mean field due to the two quasiparticle interactions is usually ignored. Therefore the Hamiltonian considered is: \begin{equation} H_{qp}=E_{p}{\hat N}_{p}+E_{n}{\hat N}_{n}+\lambda_{1} B^{\dag}(pn)B(pn)+\lambda_{2}\big(B^{\dag2}(pn)+B^{2}(pn)\big), \end{equation} The operators $B^{\dag},B$ satisfy the commutation relation: \begin{equation} [B(pn),B^{\dag}(pn)]={\hat N}_n-{\hat N}_p. \end{equation} If the r.h. side of the above equation is replaced by its average on the ground state, \begin{equation} C=\langle 0|{\hat N}_n-{\hat N}_p|0\rangle \end{equation} which is to be determined, then the operators $B,B^{\dag}$ become bosons, after the following renormalization \begin{equation} {\widetilde {B}}^{\dag}(pn)=\frac{1}{\sqrt{C}}B^{\dag}(pn),\; {\widetilde {B}}(pn)=\frac{1}{\sqrt{C}}B(pn), \end{equation} if $C$ is positive, while for negative $C$ the renormalized operators are: \begin{equation} {\widetilde {B}}^{\dag}(pn)=\frac{1}{\sqrt{|C|}}B(pn),\; {\widetilde {B}}(pn)=\frac{1}{\sqrt{|C|}}B^{\dag}(pn). \end{equation} Suppose, for the time being, that $C>0$. If that is not the case the corresponding calculations can be worked out in a similar way. The equations of motion for the renormalized operators are: \begin{eqnarray} \left[H_{qp},{\widetilde {B}}^{\dag}(pn)\right]& = & (E_p-E_n+\lambda_1C) {\widetilde{B}}^{\dag}(pn) +2\lambda_2C{\widetilde{B}}(pn), \nonumber \\ \left[ H_{qp},{\widetilde{B}}(pn)\right]& =& -2\lambda_2C{\widetilde{B}}^{\dag}(pn) -(E_p-E_n+\lambda_1C){\widetilde{B}}(pn) . \end{eqnarray} Since the equations are linear in ${\widetilde {B}}^{\dag}(pn)$ and ${\widetilde{B}}(pn)$, one can define the phonon operator \begin{equation} \Gamma^{\dag}=X\widetilde {B}^{\dag}(pn)-Y\widetilde {B}(pn), \end{equation} with the amplitudes determined such that the following equations are fulfilled: \begin{eqnarray} \left[H_{qp},\Gamma^{\dag}\right]&=&\omega\Gamma^{\dag},\nonumber\\ \left[\Gamma,\Gamma^{\dag}\right]&=&1. \end{eqnarray} The first equation provides the dispersion equation for the mode energy \begin{equation} \omega=\left[(E_p-E_n+\lambda_1C)^2-4\lambda_2^2C^2\right]^{\frac{1}{2}}, \end{equation} while the second one the normalization relation for phonon amplitudes: \begin{equation} X^2-Y^2=1. \end{equation} The renormalized RPA vacuum is defined by \begin{equation} \Gamma|0\rangle=0. \end{equation} The solution of the above equation is: \begin{equation} |0\rangle=e^{-\frac{1}{8}(\frac{Y}{X})^2}e^{\frac{Y}{2X}{\widetilde {B}}^2} |NT,-T\rangle. \end{equation} Then the renormalization constant $C$ can be exactly evaluated: \begin{equation} C=2T-2+\frac{2}{X^2}. \end{equation} Since $T\ge1$, the constant C is always positive. The equations of motion allow us to express the amplitude Y in terms of X: \begin{equation} Y=\frac{1}{2\lambda C}[\omega-(E_p-E_n+\lambda_1C)]X, \end{equation} which together with the normalization condition (5.10) determines fully the amplitudes X and Y in terms of C and $\omega$. Inserting the result for $X$ into the equation (5.13), one obtains an equation for C as a function of $\omega$. This and eq.(5.9) form a set of two nonlinear equations for the unknowns $\omega$ and $C$. As we mentioned before, another way to improve the RPA treatment is to use the boson expansion concept. Through this procedure, the $SU(2)$ algebra, with the fermionic generators $\tau_{\pm 1}, \tau_0$ defined by eq.(3.6), is mapped to a boson $SU(2)$ algebra, generated by $\hat {T}_{\pm 1}, \hat {T_0}$. Denoting by $b^+,b$ a pair of boson operators, the $SU(2)$ algebra generators $\hat {T}_{\pm 1},\hat {T}_0$ can be constructed as function of $b^+$ and b. The resulting expressions are conventionally called as the boson expansion of the fermionic generators, respectively. There are three distinct boson mappings for the fermionic SU(2) algebra found by Holstein-Primakoff \cite{Hol}, Dyson \cite{Dy} and one of the present authors (A. A. R.)\cite{Rad5}, respectively. For the present purpose here we use the Holstein-Primakoff (HP) expansion: \begin{eqnarray} \hat{T}_{+1}&=&-\sqrt{T}b^+\left(1-\frac{b^+b}{2T}\right)^{\frac{1}{2}}, \nonumber\\ \hat{T}_{-1}&=&\sqrt{T}\left(1-\frac{b^+b}{2T}\right)^{\frac{1}{2}}b, \nonumber\\ \hat{T}_0&=&b^+b-T. \end{eqnarray} By a direct calculation it can be checked that, by this mapping, to the operator $\tau ^2$ it corresponds a C-number: \begin{equation} \hat{T}^2=T(T+1). \end{equation} The fermion Hamiltonian $H_{qp}$ commutes with the quasiparticle total number and the same is true for the generators $\tau_{\pm 1},\tau_0$. Therefore the image of the quasiparticle total number operator through the HP mapping is invariant against any rotation in the isospin space and consequently, according to the above equation, is a C-number. Apart from an additive constant, the image of $H_{qp}$ through the HP boson expansion is: \begin{equation} H^{(b)}_{qp}=(E_p-E_n)\hat{T}_0-2\lambda_1\hat{T}_{+1}\hat {T}_{-1} +2\lambda_2(\hat {T}^2_{+1}+\hat {T}^2_{-1}). \end{equation} Making use of eqs. (5.15), the boson mapping of $H_{qp}$ is a infinite series in the bosons $b^+, b$, due to the square root operators. Expanding the square root operators and truncating the result at the second order in bosons, the boson Hamiltonian becomes: \begin{equation} H^{(b)}_{qp;2}=(E_p-E_n+2\lambda_1T)b^+b+2\lambda_2T({b^+}^2+b^2). \end{equation} For a limited range of the interaction strength, this Hamiltonian can be diagonalized through a canonical transformation: \begin{eqnarray} b^+&=&UB^++VB, \nonumber\\ b&=&UB+VB^+, \nonumber\\ 1&=&U^2-V^2. \end{eqnarray} The restriction that the "dangerous" terms have a vanishing strength yields the expression for the transformation coefficients and the coefficient, $\omega_1$, of the diagonal term $B^+B$: \begin{eqnarray} \left(\matrix{U\cr V}\right)&=&\frac{1}{\sqrt{2}}\left[\mp 1+\frac{|E_p-E_n|}{\sqrt{(E_p-E_n+2\lambda_1T)^2 -16\lambda_2^2T^2}} \right]^{\frac{1}{2}}, \nonumber\\ \omega_1&=&\left[(E_p-E_n+2\lambda_1T)^2 -16\lambda_2^2T^2\right]^{\frac{1}{2}}. \end{eqnarray} Comparing the expressions of $\omega_1$ (5.20) and $\omega$ (5.9), one sees that the two energies are identical for the limiting case of $X=1$, which is met when $\lambda_2=0$ (see eqs. (5.9) and (5.14)). At this stage it is worthwhile to make the following remarks: a) When the HP boson expansion of the model Hamiltonian is truncated at the second order terms in bosons, the quasiparticle total number operator is no longer a C number. Therefore the contribution of this term should have been considered in a consistent manner. Moreover the truncation is justified only for large values of the total isospin T. b) The same inconsistency appears in the calculation of the renormalization constant C. Indeed the expression (5.14) is exact and therefore includes all contributions coming from the infinite boson series of the correlated ground state, given by (5.14). c) Since the boson mapping (5.15) is an unitary transformation, the exact eigenvalues of $H_{qp}$ are reproduced by diagonalizing the boson expanded Hamiltonian $H^{(b)}_{qp}$. For the second order truncated Hamiltonian, the canonical transformation breaks down at a critical value of the attractive interaction strength. However the diagonalization procedure is able to find the eigenvalues for any strength of the attractive interaction. The resulting energies exhibits a phase transition (the first derivative has a jump) at the critical value of the strength. If the second branch of the energy curve could also be approximated by an harmonic mode, describing small oscillations of the classical system around a stationary state, this is still an open question \cite{Rad4}. d) The HP boson representation provides for the harmonic mode the interpretation of an wobbling motion of the system around the total isospin. e) The HP boson expansion is justified (in the sense that some eigenvalues of the truncated Hamiltonian are close to the corresponding exact ones) when the rotation axis in the isospin space is close to the quantization axis (z axis), which is usually taken as the axis to which the maximum ``moment of inertia'' corresponds. If the angle between the rotation axis and z-axis is large the harmonic energy may collapse. In this case the quantization axis should be chosen as one of the X and Y axes depending on the magnitude of the strength of the $\tau_x^2$ and $\tau_y^2$ terms from the quasiparticle Hamiltonian $H^{(q)}_{pn}$. In this case the boson representation suitable for the low order description should be of Dyson type \cite{Rad6}. The harmonic approximation for the new representation describes also a wobbling motion of a frequency equal to the square root of the product of the inverse of the non-maximal moments of inertia normalized to the inverse of the maximal moment of inertia. \section{Numerical results} \label{sec:level6} The formalism described in the previous sections, has been applied to the case $j=\frac{19}{2}$. On the proton level, 6 protons are distributed while in the neutron level, 14 neutrons. Alike nucleons interact with each other through pairing forces whose strength are $G_p=0.2 {\rm MeV}$ and $G_n=0.4$ MeV. From the pairing equations it results the following expression for the quasiparticle energy: \begin{equation} E_{\tau}=\frac{1}{2}G_{\tau}\Omega,\; \Omega=\frac{2j+1}{2}. \end{equation} With the data specified above the result for the quasiparticle energy is: \begin{equation} E_p=2\; {\rm MeV},\;\;E_n=1\;\;{\rm MeV}. \end{equation} According to our previous study, the renormalized RPA ground state involves a small number of quasiparticles. For example, for a small strength of the particle-particle interaction, the quasiparticle total number is about 2 while for large values of the above mentioned strength the number may reach the value 4. Due to this behavior of the correlated ground state we considered for the isospin carried by the quasiparticles in the ground state, alternatively the values 1 and 2. Although these values vary with increasing the particle particle strength we kept them constant. The numerical analysis refers to the dependence of the energy $\omega$ of the new nuclear mode, on the strengths of the $ph$ and $pp$ monopole interactions, $\chi, \chi_1$. Aiming at showing how good is the semi-classical approach for this new type of pn excitation, we calculated also the exact eigenvalues of the model Hamiltonian, by diagonalizing the associated matrix (4.14) within the basis $|NTM\rangle$. The results are shown in Fig. 1 and Fig. 2. From Fig. 1, one notices that the harmonic mode collapses for a critical value of the attractive interaction $\chi_1$. This critical value is certainly depending on the repulsive interaction strength. The larger is that strength the larger the critical value. In Fig. 1, we have also plotted the normalized energy for the first excited state. There are intervals for $\chi_1$ where the energy of the harmonic mode approximates reasonably well the exact excitation energy. Moreover, for two values of the strength parameter, the exact solutions are precisely reproduced. For the case T=2, the two energies, exact and $\omega$, are the same for $\chi_1=0$ and $\chi_1=0.3$ for $\chi=1$ and $\chi=0.5$ respectively, but the curves are going apart for the first part of interval and then converge to an intersection point close to the critical value. The peculiar feature of $\omega$ as a function of $\chi_1$, which distinguishes it from the standard RPA modes, consists of its non-monotonic behavior with respect to the increase of the strength of the $pp$ interaction. The reason is that in the common cases the mean field is constant when the two body interaction is varied, while here by changing $\chi_1$ we change also the minimum point for energy and therefore another mean field is obtained. It is interesting to notice that although the ph interaction, the $\chi$ term, is kept constant, the change of the mean field is equivalent to an increase of the effective $ph$ interaction until $\omega$ reaches the maximum value from where the attractive component of the two body interaction prevails. In Fig. 2, the energies $\omega$ and the normalized energy of the first excited state are shown as function of $\chi$, the strength parameter of the $ph$ interaction. Both energies are monotonically increasing with the increase of the interaction strength. In contrast to what happens in the case of $\chi_1$ dependence, here the change of the mean field by changing the energy minimum does not change the repulsive character of the $\chi$ interaction. The agreement between $\omega$ and the exact energy of the first excited state is reasonable good. In Fig. 3. the energies characterizing the harmonic mode predicted by the renormalized RPA and semi-classical method are plotted as function of $\chi_1$. Also, the exact energy of the first excited state is presented. Although they have different trends, the semi-classical and renormalized RPA energies are not far from each other for $\chi<0.45$. At the critical value $\chi$=0.57 the energy yielded by the renormalized RPA is going very fast to zero. This behavior is specific to the present model where only the scattering terms are considered. Indeed, if the phonon operator includes both the two quasiparticle and scattering terms, the corresponding mode collapses for larger $\chi_1$. In the semi-classical treatment this happens only for very large $\chi_1$ since the static ground state is changed by increasing $\chi_1$. The result obtained with the truncated HP boson expanded Hamiltonian (see eq. (5.20)) is very close to the result shown in Fig. 3 for the renormalized RPA procedure. Comparing the results from Fig. 1a and Fig. 3, we remark on the following features. While the renormalized RPA energy collapses at a relatively small value of $\chi_1$, the mode energy predicted by the semi-classical formalism vanishes for a very large $\chi_1$, far beyond the realistic value, which is $\chi_1=\chi$. This feature is a consequence of changing the static ground state with $\chi_1$. The energy behavior provided by the semi-classical method is also different from that predicted by the standard renormalized pnQRPA (see for example ref. 9) where the mode energy is a monotonic function of $\chi_1$ and goes asymptotically to zero. In this context we recall that the $frn-RPA$ breaks down \cite{Rad3} before the standard RPA does, and that happened due to the fact that the lowest $frn-RPA$ energy is that associated with the new collective mode. From the present calculations one sees that {\it this is not true within the semi-classical approach and therefore including the scattering terms in the expression of the phonon operator does not prevent the treatment of the many-body system for a realistic value of the $pp$-interaction strength.} The vanishing energies for the new mode, shown in Figs. 1a and 3 suggest that a phase transition occurs according to the corresponding formalisms. As we already mentioned this is clearly revealed if one diagonalizes the Hamiltonian given be eq. (5.18) \cite{Sam,Rad4}. In the renormalized RPA procedure the new phase is determined by a new minimum of the classical energy associated to $H^{(b)}_{pn;2}$, reflecting the fact that the $\lambda_2$ term is the dominant one for these values of $\chi_1$. In the full-line and dotted-line curves of Fig. 1a, the corresponding energies also vanish at certain critical values which result in having again a phase transition. This is reflected in the curve obtained by exact calculations, by the fact that the energy is minimum for the critical strength. The increasing branch shown by the exact calculations (corresponding to the second nuclear phase) might be semi-classically described by changing the trial function, involved in the time dependent variational equations, by rotating it (in the isospin space) with an angle which corresponds to the orientation of the axis of maximum ``moment of inertia''. It is remarkable that the far intersection points of the curves obtained by semi-classical and exact calculations respectively, are lying close to the critical values of the semi-classical description. Also, the first intersection point is not far from the critical value of the renormalized RPA treatment. In the classical treatment this feature is well known \cite{Ghe}. Indeed, in the above quoted reference it is shown, for a triaxial rotor cranked on an arbitrarily oriented axis, that for certain critical values of the strength parameters, the period of the harmonic orbits is equal to the period characterizing the motion on the closed exact orbit. \section{Conclusions} \label{sec:level7} The main result of this paper refers to the existence of an harmonic mode determined by the scattering quasiparticle terms, which are usually neglected in the standard RPA approach. The new mode is described within a time dependent variational formalism with an exactly solvable many-body Hamiltonian. The variational state is a coherent state for the underlying symmetry group, which is the SU(2) group. A pair of classical canonical conjugate coordinates, which bring the equations of motion to the Hamilton form, is found. The classical energy has an interesting structure. It is quadratic in coordinate but highly non-linear in the conjugate momentum. Therefore one finds first the stationary point which minimizes the energy, and then linearizes the equations of motion around the minimum point in the classical phase space. The solution for the linearized equations is harmonic and its time period determines the energy of the new mode. Despite the fact the classical system has an harmonic motion, the mode does not exist in the standard RPA approach. In this sense one may say that the present description corresponds to a ''renormalized RPA''. However as we have seen, by comparing the corresponding predictions, the renormalization involved in the semi-classical description is completely different from the renormalization described in Section V as well as from the boson expansion method. It is known the fact that the topological structure of the energy surface depends on the strength parameters involved in the model Hamiltonian. Thus, in the parameters space one can define several regions, each of them corresponding to a distinct nuclear phase. Having this in mind, we studied the behavior of the new mode energy when the strength parameter for the $pp$ interaction ($\chi_1$) is varied. A particular feature for the semi-classical description is that the energy is not monotonic decreasing function of $\chi_1$, but it increases in the first part of the interval, reflecting that here the $ph$ two body interaction prevails, reaches a maximum value, then decreases and finally vanishes. This property is caused by that for each $\chi_1$ a new ground state is determined. This aspect is missing in both the renormalized RPA and boson expansion procedures. Since the model Hamiltonian resembles the triaxial rotor which was semi-classically studied by one of the present authors (A. A. R) in refs. \cite{Rad6,Ghe}, the interpretation of the new mode is imported from there. Thus, the new mode describes a wobbling motion around a given total isospin. The vanishing energy is a sign for a phase transition. In the first phase the rotation axis, in the isospin space lies close to the z-axis, which has the maximum moment of inertia in the region of small $\chi_1$, while for $\chi_1$ larger than the critical value (where the energy vanishes), the rotation axis lies closer to the (X,Y) plane in the isospin space. While the first phase may be described by a HP boson expansion formalism, for the second phase the Dyson boson representation is the proper one [20]. In the semi-classical approach, the new phase might be described by changing the trial function associated to the first phase, through a rotation which brings the z-axis to the actual axis of maximal ``moment of inertia''. The occurrence of the phase transition can be noticed also in the curve showing the exact first excitation energy as function of $\chi_1$. Indeed at the critical value of $\chi_1$, this curve exhibits a minimum. Another critical values of $\chi_1$ are those where the mode energy is equal to the exact excitation energy produced by the diagonalization procedure. For these values the linearization does not affect at all the period of the exact closed classical orbit. As a matter of fact, for $\chi_1$ lying close to these points the linearization are best justified. It is interesting to notice that these values of $\chi_1$ lie however close to the values where the phase transitions in the semi-classical treatment (the far intersection point) and the renormalized RPA approach (the near intersection point) take place. This observation allows us to conclude that the semi-classical approach works very well for the values of $\chi_1$ where the renormalized RPA breaks down and that the interval where the linearization procedure does not work, ending with the critical value where the semi-classical energy vanishes, is very narrow. The energy of the new mode vanishes for a value of $\chi_1$ which is far beyond the physical value ($\chi_1=\chi$). In this way the drawback of the $frn-RPA$, of breaking down earlier than the standard RPA does, is removed. How could the new state be populated? We identified the transition operators which could excite the new state from the ground state. The conclusion is that these state can be seen either in a $\beta^-$ (or $\beta^+$) decay or in a deuteron transfer reaction experiment. The coupling of this mode to other collective states will be studied in a subsequent paper using a realistic interaction and a large model space for the single particle motion. \vskip0.5cm {\bf Acknowledgement} \vskip0.5cm \noindent One of us (B. C.) wants to thank Prof. Amand Faessler for hospitality in Institute of Theoretical Physics of Tuebingen University where a part of this work was performed. A. A. R. thanks Prof. Faessler for reading the manuscript and valuable remarks concerning the structure of the variational function. \section{Appendix A} \label{sec:level8} Here to derive the factorization of the trial function $|\Psi\rangle$. To this purpose we address the following more general question. Which are the t-functions $A(t), B(t), C(t)$ satisfying the equation \begin{equation} e^{t[(z\sum_{m}a_{pm}^{\dag}a_{nm}-z^{*}\sum_{m}a_{nm}^{\dag}a_{pm})]} =e^{A(t)a_{p}^{\dag}a_{n}}e^{C(t)(\hat N_{p} -\hat N_{n})}e^{B(t)a_{n}^{\dag}a_{p}}, \end{equation} with the initial conditions \begin{equation} A(0)=B(0)=C(0)=0 \end{equation} and t a real parameter. Once we solve this problem the needed factorization is obtain from (4.1) for $t=1$. Taking the first derivative of the eq.(6.1), with respect to t, and identifying the coefficients of the similar operators one obtains the following system of differential equations for the three unknown functions, $A(t), B(t), C(t)$ : \begin{eqnarray} \stackrel{\bullet}{z} &=& \stackrel{\bullet}{A}-2A \stackrel{\bullet}{C}-\stackrel{\bullet}{B}A^{2}e^{-2C(t)},\nonumber\\ 0&=& \stackrel{\bullet}{C}+\stackrel{\bullet}{B}Ae^{-2C(t)},\nonumber\\ -\stackrel{\bullet}{z}^* &= &\stackrel{\bullet}{B}e^{-2C(t)}. \end{eqnarray} Eliminating the functions $B, C$ from these equations, one obtains the following equation for $A(t)$. \begin{equation} \stackrel{\bullet}{z}=\stackrel{\bullet}{A}-A^{2}\stackrel{\bullet}{z}^* \end{equation} which admits the solution: \begin{equation} A(t)=\tan(\rho t)e^{i\varphi}. \end{equation} Here the polar coordinates $(\rho,\varphi)$($z=\rho e^{i\varphi}$) have been used. Inserting the result for $A(t)$ in the eq. (6.3), the equations for the remaining functions can be easily integrated. The result is: \begin{eqnarray} C(t)&=&-\ln \big( \cos(\rho t) \big)\nonumber\\ B(t)&=&\tan(\rho t) e^{-i\varphi}. \end{eqnarray} For the sake of simplifying the writing, hereafter the following notation will be used: \begin{equation} \alpha=A(1) \end{equation} Using these results the trial function can be written as: \begin{eqnarray} |\Psi\rangle & = & e^{-2C(1)T}e^{A(1)B^{\dag}(pn)}|NT\;-T \rangle\equiv {\cal N} e^{A(1)B^{\dag}(pn)}|NT\;-T\rangle \end{eqnarray} where $\cal N$ denotes the normalization factor: \begin{equation} {\cal N} = e^{-2C(1)T}=e^{2 \ln(\cos \rho)T}={(1+|\alpha|^2)}^{-T}. \end{equation} \nopagebreak
{ "timestamp": "2005-03-14T12:04:03", "yymm": "0503", "arxiv_id": "nucl-th/0503041", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503041" }
\section{Introduction.} Recently, a number of analysts \cite{A-M-N, B-V, M-M} have studied various generalized notions of derivations in the context of Banach algebras. There are some applications in the other fields of study \cite{H-L-S}. Such maps have been extensively studied in pure algebra; cf. \cite{A-R, BRE, HVA}. Let, throughout the paper, $A$ denote a Banach algebra (not necessarily unital) and let $M$ be a Banach right $A$-module. A linear mapping $d : A \to A$ is called a derivation if $d(ab)= d(a)b + ad(b)\quad(a, b\in A)$. If $a\in A$ and we define $d_{a}$ by $d_{a}(x)=ax-xa \quad (x\in A)$. Then $d_{a}$ is a derivation and such derivation is called inner. A linear mapping $\delta : M\to M$ is called a generalized derivation if there exists a derivation $d : A \to A$ such that $\delta(xa)=\delta(x)a + xd(a)\quad(x\in M, a\in A)$. For convenience, we say that such a generalized derivation $\delta$ is a $d$-derivation. In general, the derivation $d : A\to A$ is not unique and it may happen that $\delta$ (resp. $d$) is bounded but $d$ (resp. $\delta$) is not bounded. For instance, assume that the action of $A$ on $M$ is trivial, i.e $MA=\{0\}$. Then every linear mapping $\delta : M\to M$ is a $d$-derivation for each derivation $d$ on $A$. Our notion is a generalization of both concepts of a generalized derivation (cf. \cite{BRE, HVA}) and of a multiplier (cf. \cite{DAL}) on an algebra (see also \cite{MOS2}). For seeing this, regard the algebra as a module over itself. The authors in \cite{A-M-N} investigated the generalized derivations on Hilbert $C^*$-modules and showed that these maps may appear as the infinitesimal generators of dynamical systems. \begin{example} Let $M$ be a right Hilbert $C^{*}$-module over a $C^*$-algebra $A$ of compact operators acting on a Hilbert space (see \cite{LAN} for more details on Hilbert $C^*$-modules). By Theorem 4 of \cite{B-G}, $M$ has an orthonormal basis so that each element $x$ of $M$ can be expressed as $x=\displaystyle{\sum_{\lambda}}v_{\lambda}<v_{\lambda},x>$. If $d$ is a derivation on $A$, then the mapping $\delta : M\to M$ defined by $\delta(x)=\displaystyle{\sum _{\lambda}}v_{\lambda}d(<v_{\lambda},x>)$ is a $d$-derivation since \begin{eqnarray*} \delta(xa)&=&\delta\left(\displaystyle{\sum_{\lambda}}v_{\lambda}<v_{\lambda},xa>\right)\\ &=&\displaystyle{\sum_{\lambda}}v_{\lambda}d(<v_{\lambda},x>a)\\ &=&\displaystyle{\sum_{\lambda}}v_{\lambda}d(<v_{\lambda},x>)a+\displaystyle{\sum_{\lambda}}v_{\lambda}<v_{\lambda},x>d(a)\\ &=&\delta(x)a + xd(a). \end{eqnarray*} \end{example} The set ${\mathcal B}(M)$ of all bounded module maps on $M$ is a Banach algebra and $M$ is a Banach ${\mathcal B}(M)-A$-bimodule equipped with $T.x=T(x)\quad(x\in M, T\in {\mathcal B}(M))$, since we have $T.(xa)=T(xa)=T(x)a=(T.x)a$ and $\|T.xa\| \leq\|T\|\;\|x\|\;\|a\|$, for all $a \in A, x\in M, T\in {\mathcal B}(M)$. We call $\delta : M\to M$ a generalized inner derivation if there exist $a\in A$ and $T\in {\mathcal B}(M)$ such that $\delta(x)=T.x - xa = T(x)- xa$. Mathieu in \cite{MAT} called a map $\delta : A\to A$ a generalized inner derivation if $\delta(x)=bx-xa$ for some $a, b\in A$. If we consider $A$ as a right $A$-module in a natural way, and take $T(x)=bx$, then our definition covers the notion of Mathieu. In this paper we deal with the derivations on the triangular Banach algebras of the form ${\mathcal T}=\left(\begin{array}{cc} {\mathcal B}(M) & M\\ 0 & A \end{array}\right)$. Such algebras were introduced by Forrest and Marcoux \cite{F-M1} that in turn are motivated by work of Gilfeather and Smith in \cite{G-S} (these algebras have been also investigated by Y. Zhang who called them module extension Banach algebras \cite{ZHA}). Among some facts on generalized derivations, we investigate the relation between generalized derivations on $M$ and derivations on ${\mathcal T}$. In particular, we show that the generalized first cohomology group of $M$ is isomorphic to the first cohomology group of ${\mathcal T}$. \section{Main Results.} If we consider $A$ as an $A$-module in a natural way then we have the following lemma about generalized derivations on $A$. \begin{lemma} A linear mapping $\delta : A \to A$ is a generalized derivation if and only if there exist a derivation $d : A\to A$ and a module map $\varphi : A\to A$ such that $\delta=d+\varphi$.\end{lemma} \begin{proof} Suppose $\delta$ be a generalized derivation on $A$, then there exists a derivation $d$ on $A$ such that $\delta$ is a $d$-derivation. Put $\varphi=\delta-d$. Then for each $a, x\in A$ we have $$\varphi(xa)=\delta(xa)-d(xa)=\delta(x)a+xd(a)-(d(x)a+xd(a))=(\delta(x)-d(x))a=\varphi(x)a$$ Thus $\varphi$ is a module map and $\delta=d+\varphi$. Conversely, let $d$ be a derivation on $A$, $\varphi$ be a module map on $A$ and put $\delta=d+\varphi$. Then clearly $\delta$ is a linear map and $$\delta(xa)=d(xa)+\varphi(xa)=d(x)a+xd(a)+\varphi(x)a=(d(x)+\varphi(x))a+xd(a)=\delta(x)a+xd(a)$$ for all $a, x\in A$. Therefore $\delta$ is a $d$-derivation.\end{proof} The next two results concern the boundedness of a generalized derivation. \begin{theorem} Let $A$ have a bounded left approximate identity $\{e_{\alpha}\}_{\alpha\in I}$ and let $\delta$ be a $d$-derivation on $A$. Then $\delta$ is bounded if and only if $d$ is bounded. \end{theorem} \begin{proof} First we show that every module map on $A$ is bounded. Suppose that $\varphi$ is a module map on $A$ and let $\{a_{n}\}$ is a sequence in $A$ converging to zero in the norm topology. By a consequence of Cohen Factorization Theorem (see Corollary 11.12 of \cite{B-D}) there exist a sequence $\{b_{n}\}$ and an element $c$ in $A$ such that $b_{n}\to 0$ and $a_{n}=cb_{n},\quad (n\in {\mathbb N})$. Then $\varphi(a_{n})=\varphi(cb_{n})=\varphi(c)b_{n}\to 0$. Thus by the closed graph theorem, $\varphi$ is bounded. Now let $\delta$ be a $d$-derivation. By Lemma 2.1, $\delta=d+\varphi$ for some module map $\varphi$ on $A$. Therefore $\delta$ is bounded if and only if $d$ is bounded.\end{proof} \begin{corollary} Every generalized derivation on a $C^{*}$-algebra is bounded.\end{corollary} \begin{proof} Every derivation on a $C^*$-algebra is automatically continuous; cf. \cite{J-S}.\end{proof} Let $\varphi : A\to A$ be a homomorphism (algebra morphism). A linear mapping $T : M\to M$ is called a $\varphi$-morphism if $T(xa)=T(x)\varphi(a)\quad(a\in A, x\in M)$. If $\varphi$ is a isomorphism and T is a bijective mapping then we say T to be a $\varphi$-isomorphism. An $id_{A}$-morphism is a module map (module morphism). Here $id_{A}$ denotes the identity operator on $A$. \begin{proposition} Suppose $\delta$ is a bounded $d$-derivation on $M$ and $d$ is bounded. Then $T=\exp(\delta)$ is a bi-continuous $\exp(d)$-isomorphism.\end{proposition} \begin{proof} Using induction one can easily show that $$\delta^{(n)}(xa)=\sum_{r=0}^{n}(_{r}^{n})\delta^{(n-r)}(x)d^{(r)}(a).$$ For each $a\in A, x\in M$ we have \begin{eqnarray*} T(xa)&=&\exp(\delta)(xa)\\ &=&\sum_{n=0}^{\infty}\frac{1}{n!}\delta^{(n)}(xa)\\ &=&\sum_{n=0}^{\infty}\frac{1}{n!}\sum_{r=0}^{n}(_{r}^{n})\delta^{(n-r)}(x)d^{(r)}(a)\\ &=&\sum_{n=0}^{\infty}\sum_{r=0}^{n}(\frac{1}{(n-r)!} \delta^{(n-r)}(x)(\frac{1}{r!}d^{(r)}(a))\\ &=&(\sum_{n=0}^{\infty}\frac{1}{n!}\delta^{(n)}(x))(\sum_{n=0}^{\infty}\frac{1}{n!}d^{(n)}(a))\\ &=&\exp(\delta)(x)\exp(d)(a) \end{eqnarray*} The operators $\exp(\delta), \exp(d)$ are invertible in the Banach algebras of bounded operators on $M$ and $A$, respectively. Hence $T$ is an $\exp(d)$-isomorphism.\end{proof} \begin{proposition} Let $\delta$ be a bounded generalized derivation on $M$. Then $\delta$ is a generalized inner derivation if and only if there exists an inner derivation $d_{a}$ on $A$ such that $\delta$ is $d_{a}$-derivation. \end{proposition} \begin{proof} Let $\delta$ be a generalized inner derivation. Then there exist $a\in A$ and $T\in {\mathcal B}(M)$ such that $\delta(xa)=T(x)-xa\quad (x\in M)$. We have $\delta(x)b+xd_{a}(b)=(T(x)-xa)b+xab-xba =T(x)b-xba=T(xb)-(xb)a=\delta(xb) \quad (b \in A, x \in M)$. Hence $\delta$ is a $d_{a}$-derivation. Conversely, suppose $\delta$ is a $d_{a}$-derivation for some $a\in A$. Define $T : M\to M$ by $T(x)=\delta(x)+xa$. Then T is linear, bounded and $T(xb)=\delta(xb)+(xb)a=(\delta(x)b+xd_{a}(b))+xba=\delta(x)b+xab-xba+xba=(\delta(x)+xa)b=T(x)b$. It follows that $T\in {\mathcal B}(M)$ and $\delta(x)=(\delta(x)+xa)-xa=T(x)-xa$. Therefore $\delta$ is a generalized inner derivation.\end{proof} The linear spaces of all bounded generalized derivations and generalized inner derivations on $M$ are denoted by $GZ^{1}(M,M)$ and $GN^{1}(M,M)$, respectively. We call the quotient space $GH^{1}(M,M)=GZ^{1}(M,M)/GN^{1}(M,M)$ the generalized first cohomology group of $M$. \begin{corollary} $GH^{1}(M,M)=0$ whenever $H^{1}(A,A)=0$ \end{corollary} \begin{proof} Let $\delta : M\to M$ be a generalized derivation. Then there exists a derivation $d : A \to A$ such that $\delta$ is a $d$-derivation. Due to $H^{1}(A,A)=0$, we deduce that $d$ is inner and, by Proposition 2.5, so is $\delta$. Hence $GH^{1}(M,M)=0$. \end{proof} Using some ideas of \cite{F-M1, MOS1}, we give the following notion: \begin{definition}{\rm ${\mathcal T}=\{\left(\begin{array}{cc}T & x \\0 & a \end{array}\right); T\in {\mathcal B}(M), x\in M, a\in A \}$ equipped with the usual $2\times 2$ matrix addition and formal multiplication and with the norm $\|\left(\begin{array}{cc} T & x \\ 0 & a \\\end{array}\right)\|=\|T\|+\|x\|+\|a\|$ is a Banach algebra. We call this algebra the triangular Banach algebra associated to M.} \end{definition} The following two theorems give some interesting relations between generalized derivations on $M$ and derivations on ${\mathcal T}$. Let $\delta$ be a bounded $d$-derivation on $M$. We define $\Delta_{\delta} : {\mathcal B}(M)\to {\mathcal B}(M)$ by $\Delta_{\delta}(T)=\delta T-T\delta$. Then $\Delta_{\delta}$ is clearly a derivation on ${\mathcal B}(M)$. \begin{theorem} Let $\delta$ be a bounded $d$-derivation on $M$ and $d$ be bounded. Then the map $D^{\delta} : {\mathcal T} \to {\mathcal T}$ defined by $D^{\delta}\left(\begin{array}{cc} T & x \\ 0 & a \end{array}\right)=\left( \begin{array}{cc} \Delta_{\delta}(T) & \delta(x) \\ 0 & d(a)\end{array}\right)$ is a bounded derivation on ${\mathcal T}$. Also $\delta$ is a generalized inner derivation if and only if $D^{\delta}$ is an inner derivation. \end{theorem} \begin{proof} It is clear that $D^{\delta}$ is linear. For any $T_{1},T_{2}\in {\mathcal B} (M), x_{1},x_{2}\in M, a_{1},a_{2}\in A$ we have \begin{eqnarray*} &&D^{\delta}(\left(\begin{array}{cc}T_{1} & x_{1} \\ 0 & a_{1}\end{array}\right)\left(\begin{array}{cc} T_{2} & x_{2} \\ 0 & a_{2} \end{array}\right))= D^{\delta}\left(\begin{array}{cc} T_{1}T_{2} & T_{1}.x_{2}+x_{1}a_{2} \\ 0 & a_{1}a_{2} \end{array}\right)\\ &=&\left(\begin{array}{cc} \Delta_{\delta}(T_{1}T_{2}) & \delta(T_{1}.x_{2}+x_{1}a_{2}) \\ 0 & d(a_{1}a_{2})\end{array} \right )\\ &=&\left(\begin{array}{cc} \Delta_{\delta}(T_{1}T_{2}) & \delta(T_{1}(x_{2}))+ \delta(x_{1})a_{2}+x_{1}d(a_{2})\\ 0 & a_{1}d(a_{2})+d(a_{1})a_{2}\end{array}\right )\\ &=&\left(\begin{array}{cc} T_{1}\Delta_{\delta}(T_{2})+\Delta_{\delta}(T_{1})T_{2} & T_{1}.\delta(x_{2})+x_{1}d(a_{2})+(\delta T_{1}-T_{1}\delta)(x_{2})+\delta(x_{1})a_{2} \\ 0 & a_{1}d(a_{2})+d(a_{1})a_{2}\end{array}\right )\\ &=&\left(\begin{array}{cc} T_{1} & x_{1} \\ 0 & a_{1} \end{array}\right)\left(\begin{array}{cc} \Delta_{\delta}(T_{2}) & \delta(x_{2}) \\ 0 & d(a_{2})\end{array}\right) + \left(\begin{array}{cc} \Delta_{\delta} (T_{1}) & \delta(x_{1}) \\ 0 & d(a_{1}) \end{array}\right)\left(\begin{array}{cc} T_{2} & x_{2} \\ 0 & a_{2}\end{array}\right)\\ &=&\left(\begin{array}{cc} T_{1} & x_{1} \\ 0 & a_{1} \end{array}\right)D^{\delta}\left(\begin{array}{cc} T_{2} & x_{2} \\ 0 & a_{2}\end{array}\right)+D^{\delta}(\left(\begin{array}{cc} T_{1} & x_{1} \\ 0 & a_{1}\end{array}\right))\left( \begin{array}{cc} T_{2} & x_{2} \\ 0 & a_{2} \end{array}\right) \end{eqnarray*} Thus $D^{\delta}$ is a derivation on ${\mathcal T}$. Due to $\|\left(\begin{array}{cc} \Delta_{\delta}(T) & \delta(x) \\ 0 & d(a) \end{array}\right)\|=\| \Delta_{\delta}(T)\|+\|\delta(x)\|+\|d(a)\|\leq \max \{\|\Delta_{\delta}\|,\|\delta\|,\|d\|\}\|\left(\begin{array}{cc} T & x \\ 0 & a \end{array}\right)\|$, we infer that $D^{\delta}$ is bounded. Now suppose that $\delta$ is a generalized inner derivation. Then there exist $a\in A$ and $T\in {\mathcal B}(M)$ such that $\delta(x)=T(x)-xa \quad(x\in M)$. For all $S\in {\mathcal B}(M), b\in A$ and $y\in M $ we have \begin{eqnarray*} D_{\left(\begin{array}{cc} T & 0 \\ 0 & a \end{array}\right)}\left(\begin{array}{cc} S & y \\ 0 & b \end{array}\right) &:=& \left(\begin{array}{cc} T & 0 \\ 0 & a \end{array} \right)\left(\begin{array}{cc} S & y \\ 0 & b \end{array}\right)-\left(\begin{array}{cc} S & y \\ 0 & b\end{array}\right)\left(\begin{array}{cc} T & 0 \\ 0 & a \end{array} \right)\\ &=&\left(\begin{array}{cc} TS-ST & T.y-ya \\ 0 & ab-ba \end{array}\right)\\ &=& \left(\begin{array}{cc} \Delta_{\delta}(S) & \delta(y) \\0 & d_{a}(b)\end{array}\right)\\ &=&D^{\delta}\left(\begin{array}{cc} S & y \\ 0 & b \end{array}\right). \end{eqnarray*} Hence $D^{\delta}=D_{\left(\begin{array}{cc} T & 0 \\ 0 & a \end{array} \right)}$ and so $D^{\delta}$ is an inner derivation. Conversely, let $\delta$ be a bounded $d$-derivation such that the associated derivation $D^{\delta}$ be an inner derivation, say $D^{\delta} = D_{\left(\begin{array}{cc} T_0 & x_0 \\ 0 & a_0 \end{array}\right)}$. Then for each $T \in {\mathcal B}(M), x \in M,a \in A$ we have \begin{eqnarray}\label{inner} \left(\begin{array}{cc} \Delta_\delta(T) & \delta(x) \\ 0 & d(a) \end{array} \right) &=& D^{\delta}(\left(\begin{array}{cc} T & x \\ 0 & a \end{array} \right))\nonumber \\ &=& D_{\left(\begin{array}{cc} T_0 & x_0 \\ 0 & a_0 \end{array}\right)}\left(\begin{array}{cc} T & x \\ 0 & a \end{array}\right)\nonumber \\ &=& \left(\begin{array}{cc} T_0T-TT_0 & T_0(x)+x_0a-T(x_0)-xa_0 \\ 0 & a_0a-aa_0 \end{array} \right)\nonumber \\ &=&\left(\begin{array}{cc} T_0T-TT_0 & T_0(x)+x_0a-T(x_0)-xa_0 \\0 & d_{a_0}(a) \end{array}\right) \end{eqnarray} Hence $d=d_{a_0}$ is inner. Putting $a=0$ and $T=0$ in (\ref{inner}) we conclude that $\delta(x) = T_0(x) - xa_0 \quad (x \in M)$. Hence $\delta$ is a generalized inner derivation. \end{proof} The converse of the above theorem is true in the unital case. \begin{theorem} Let $A$ be unital and ${\mathcal T}$ be the triangular Banach algebra associated to a unital Banach right $A$-module $M$. Assume that $D : {\mathcal T} \to {\mathcal T}$ is a bounded derivation. Then there exist $m_{0}\in M$, a bounded derivation $d : A\to A$ and a bounded $d$-derivation $\delta : M\to M$ such that $$D\left(\begin{array}{cc} T & x \\ 0 & a \end{array}\right)=\left(\begin{array}{cc} \Delta_{\delta}(T)& \delta(x)+m_{0}a-T.m_{0} \\ 0 & d(a) \end{array}\right)$$ Moreover, $D$ is inner if and only if $\delta$ is a generalized inner derivation.\end{theorem} \begin{proof} We use some ideas of Proposition 2.1 of \cite{F-M1}. By simple computation one can verify that (i) $D\left(\begin{array}{cc} 0 & 0 \\ 0 & 1_{A} \end{array}\right)=\left(\begin{array}{cc} 0 & m_{0} \\ 0 & 0 \end{array}\right)$ for some $m_{0}\in M$; (ii) $D\left(\begin{array}{cc} 0 & 0 \\ 0 & a \end{array}\right)=\left(\begin{array}{cc} 0 & m_{0}a \\ 0 & d(a) \end{array}\right)$ for some bounded derivation $d$ on $A$; (iii) $D\left(\begin{array}{cc} 0 & x \\ 0 & 0 \end{array}\right)=\left(\begin{array}{cc} 0 & \delta(x) \\ 0 & 0 \end{array}\right)$ for some linear mapping $\delta$ on $M$; (iv) $D\left(\begin{array}{cc} T & 0 \\ 0 & 0 \end{array}\right)=\left(\begin{array}{cc} \Delta_{\delta}(T) & -T.m_{0} \\ 0 & 0 \end{array}\right)$; and finally $D\left(\begin{array}{cc} T & x \\ 0 & a \end{array}\right)=\left(\begin{array}{cc} \Delta_{\delta}(T) & \delta(x)+m_{0}a-T.m_{0} \\ 0 & d(a)\end {array}\right)$. We have \begin{eqnarray*} \left(\begin{array}{cc} 0 & \delta(xa) \\ 0 & 0 \end{array}\right)&=&D(\left( \begin{array}{cc} 0 & xa \\ 0 & 0 \end{array}\right))=D(\left(\begin{array}{cc} 0 & x \\ 0 & 0 \end{array}\right)\left(\begin{array}{cc} 0 & 0 \\ 0 & a \end{array}\right))\\ &=&\left( \begin{array}{cc} 0 & x \\ 0 & 0 \end{array}\right)D(\left(\begin{array}{cc} 0 & 0 \\ 0 & a \end{array}\right))+D(\left(\begin{array}{cc} 0 & x \\ 0 & 0 \end{array}\right))\left( \begin{array}{cc} 0 & 0 \\ 0 & a \end{array}\right)\\ &=&\left(\begin{array}{cc} 0 & x \\ 0 & 0\end{array}\right)\left(\begin{array}{cc} 0& m_0a \\ 0 & d(a) \end{array}\right)+\left(\begin{array}{cc} 0 & \delta(x) \\ 0 & 0 \end{array}\right)\left( \begin{array}{cc} 0 & 0 \\ 0 & a \end{array}\right)\\ &=&\left(\begin{array}{cc} 0 & \delta(x)a+xd(a) \\ 0 & 0 \end{array}\right) \end{eqnarray*} Thus $\delta(xa)=\delta(x)a+xd(a)$ and so $\delta$ is a $d$-derivation. It is clear that $D$ is inner if and only if $d$ is inner and, using Proposition 2.5, the latter holds if and only if $\delta$ is a generalized inner derivation.\end{proof} \begin{theorem} Let $A$ be a unital Banach algebra, $M$ be a unital Banach right $A$-module and ${\mathcal T}=\left(\begin{array}{cc} {\mathcal B}(M) & M\\ 0 & A \end{array}\right)$. Then $H^{1}({\mathcal T},{\mathcal T})\cong GH^{1}(M,M)$\end{theorem} \begin{proof} Let $\Psi : GZ^{1}(M,M)\to H^{1}({\mathcal T},{\mathcal T})$ be defined by $$\Psi(\delta)=[D^{\delta}]$$ where $[D^{\delta}]$ represents the equivalence class of $D^{\delta}$ in $H^{1}({\mathcal T},{\mathcal T})$. Clearly $\Psi$ is linear. We shall show that $\Psi$ is surjective. To end this, assume that $D$ is a bounded derivation on ${\mathcal T}$. Let $\delta$, $d$, $\Delta_{\delta}$ and $m_{0}\in M$ be as in the Theorem 2.9. Then \begin{eqnarray*} (D-D^{\delta})\left(\begin{array}{cc} T & x \\ 0 & a \end{array}\right)&=&\left(\begin{array}{cc} \Delta_{\delta}(T) & \delta(x)+m_{0}a-T.m_{0}\\ 0 & d(a) \end{array}\right) -\left(\begin{array}{cc} \Delta_{\delta}(T) & \delta(x)\\ 0 & d(a) \end{array}\right)\\ &=&\left(\begin{array}{cc} 0 & m_{0}a-T.m_{0}\\ 0 & 0 \end{array}\right)\\ &=&D_{\left(\begin{array}{cc} 0 & -m_{0} \\ 0 & 0 \end{array}\right)}\left( \begin{array}{cc} T & x \\ 0 & a \end{array}\right). \end{eqnarray*} So $[D]=[D^{\delta}]=\Psi(\delta)$ and thus $\Psi$ is surjective. Therefore $H^{1}({\mathcal T},{\mathcal T})\cong GZ^{1}(M,M)/Ker(\Psi)$. Note that $\delta\in Ker(\Psi)$ if and only if $D^{\delta}$ is inner derivation on ${\mathcal T}$. Hence $Ker(\Psi)=GN^{1}(M,M)$, by Theorem 2.8. Thus $H^{1}({\mathcal T},{\mathcal T})\cong GH^{1}(M,M)$. \end{proof} \begin{example} Suppose that $A$ is unital and $M=A$. Then ${\mathcal B}(A)=A$ and so $GH^{1}(A,A)\cong H^{1}(\left(\begin{array}{cc}A&A\\0&A \end{array}\right ), \left(\begin{array}{cc}A&A\\0&A \end{array}\right ))=H^1(A,A)$, by Proposition 4.4 of \cite{F-M2}. In particular, every generalized derivation on a unital commutative semisimple Banach algebra \cite{S-W}, a unital simple $C^*$-algebra \cite{SAK}, or a von Neumann algebra \cite{KAD} is generalized inner.\end{example} We have investigated the interrelation between generalized derivations on a Banach algebra and its ordinary derivations. We also studied generalized derivations on a Banach module in virtue of derivations on its associated triangular Banach algebra. Thus, we established a link between two interesting research areas: Banach algebras and triangular algebras. \textbf{Acknowledgment.} The authors sincerely thank the referee for valuable suggestions and comments.
{ "timestamp": "2006-06-17T13:38:40", "yymm": "0503", "arxiv_id": "math/0503618", "language": "en", "url": "https://arxiv.org/abs/math/0503618" }
\section{Introduction} The search for discrete quantum phase-space quasiprobability distribution functions is a subject of continuous and growing interest in the literature \cite{wooters,gapi1,cohendet,gapi2,opat,voros,gama,zhang,luis,haki,muk,wooters2,vourdas}. The possibility of representing quantum systems characterized by a finite-dimensional state space by such discrete quasidistributions lays the ground for interesting developments and fruitful applications on quantum computation and quantum information theory \cite{r1s3,r2s3,r3s3,r4s3,r5s3,r6s3,r7s3, r8s3}. It is well known that, as a well established counterpart to the discrete case, a huge variety of quasiprobabilty distribution functions can be defined upon continuous phase-space \cite{lee}. In this sense, the Cahill-Glauber (CG) approach \cite{cahill} to the subject has proved to be a powerful mapping technique that provides a general class of quasiprobability distribution functions, where the Wigner, Glauber-Sudarshan and Husimi functions appear as particular cases. Therefore, it might be considered as a wide-range phase-space approach to quantum mechanics regarding degrees of freedom with classical counterparts. The aim of this paper is to present a discrete extension of the CG approach. Such extension is not obtained from that approach but, instead, properly constructed out of the finite dimensional context. Furthermore, this {\sl ab initio} construction inherently embodies the discrete analogues of the desired properties of the CG formalism. In particular, discrete Wigner, Husimi and Glauber-Sudarshan quasiprobability distribution functions are obtained. Thus, besides the theoretical interest of its own, such extension has direct applications in quantum information processing, quantum tomography and quantum teleportation, which are explored in a following work \cite{nois}. This work is organized as follows: In the next section we briefly outline the CG approach, setting the stage for section III, where our proposal for a discrete extension of the CG mapping kernel is presented. In section IV basic properties of the mapping technique are discussed, and the continuum limit is carried out on section V. Finally, section VI contains our summary and conclusions. Also, important calculations are detailed in the Appendix. \section{The Cahill-Glauber Mapping Kernel} For the sake of clarity, in what follows we will briefly review the central ideas which constitute the core of the CG approach, and that will be properly generalized in the following sections. Basically, the cornerstone of the formalism is the mapping kernel (hereafter $\hbar = 1$) \begin{equation} \label{CG1} {\bf T}^{(s)}(q,p) = \int_{-\infty}^{\infty} \frac{dp^{\prime} dq^{\prime}}{2\pi} \exp \left[ -i p^{\prime}(q-\mathbf{Q}) \right] \exp \left[ i q^{\prime} (p-\mathbf{P}) \right] \exp \left( - \frac{i}{2} p^{\prime} q^{\prime} \right) \exp \left[ \frac{s}{4}(q^{\prime 2} + p^{\prime 2}) \right] \; , \end{equation} which is responsible for the mapping of bounded operators on the continuous phase-space, being $s$ a complex variable satisfying the condition $|s| \leq 1$. Here, the momentum and coordinate operators obey the Weyl-Heisenberg commutation relation $[ {\bf Q}, {\bf P}] = i {\bf 1}$. Since the above expression explicitly depends on $s$, this parameter labels an infinite family of mapping kernels. Each mapping kernel can be seen as the double Fourier transform of the displacement generators multiplied by a phase factor $\exp \left[ (i/2) p^{\prime} q^{\prime} \right]$ and by the folding function $\exp \left[ (s/4)(q^{\prime 2} + p^{\prime 2}) \right]$. For purposes which will become evident later, we write the mapping kernel as \begin{equation} \label{II} {\bf T}^{(s)}(q,p) = \int_{-\infty}^{\infty} \frac{dp^{\prime} dq^{\prime}}{2 \pi} \exp \left[ -i p^{\prime}(q-\mathbf{Q}) \right] \exp \left[ i q^{\prime}(p-\mathbf{P}) \right] \exp \left( - \frac{i}{2} p^{\prime} q^{\prime} \right) (\langle 0 | q^{\prime},p^{\prime} \rangle)^{-s} \; , \end{equation} where $| q^{\prime}, p^{\prime} \rangle$ is a coherent state. The mapping of a given operator is achieved by the trace operation $\mathcal{O}^{(s)}(q,p) = \mbox{${\rm Tr}$} [ {\bf T}^{(s)}(q,p) {\bf O}]$, being $\mathcal{O}^{(s)}(q,p)$ the function which represents ${\bf O}$ in the associated usual phase-space. The mapping is one-to-one, and the operator is reobtained from its associated function by \begin{displaymath} {\bf O} = \int_{-\infty}^{\infty} \frac{dpdq}{2\pi} \, \mathcal{O}^{(s)}(q,p) {\bf T}^{(-s)}(q,p) \; . \end{displaymath} It is clear that, for each operator, there is an infinite family of associated functions labeled by $s$. In particular, the phase-space representatives of the density operator are referred to as quasiprobability distributions functions and have, obviously, distinguishable importance \cite{lee}. One of the great virtues of the Cahill-Glauber approach is that three special and important types of quasidistributions, namely the Glauber-Sudarshan $(s=1)$, Wigner $(s=0)$ and Husimi functions $(s=-1)$, are particular cases. Each of these functions have been extensively explored and reviewed in the literature \cite{schleich}. A particular mapping kernel, characterized by a given parameter $s$, can be expressed in terms of another mapping kernel with a different parameter value. The same holds true to the functions associated with a given operator. In fact, the procedure in the latter case can be easily shown to be the same as in the former. That is, we may discuss only the relation between the mapping kernels, knowing that equivalent relations are observed by the associated functions. In this way, the connection between the two mapping kernels is seen to be given by the trace of the product \begin{eqnarray} \label{fold0.5} \mbox{${\rm Tr}$} [ {\bf T}^{(s_1)}(q_{1},p_{1}) {\bf T}^{(s_{2})}(q_{2},p_{2}) ] &=& \int_{-\infty}^{\infty} \frac{dq dp}{2\pi} \exp \left\{ i \left[ q (p_{1}-p_{2}) - p (q_{1}-q_{2}) \right] \right\} \stackrel{(\langle 0 |q,p \rangle)^{-(s_{1}+s_{2})}}{\overbrace{\exp \left[ \frac{s_{1}+s_{2}}{4} ( q^{2}+p^{2} ) \right]}} \\ \label{fold1} &=& \frac{-2}{s_{1}+s_{2}} \exp \left\{ \frac{2}{s_{1}+s_{2}} \left[ (p_{1}-p_{2})^{2} + (q_1-q_2)^{2} \right] \right\} \qquad \mbox{${\rm Re}$} (s_{1} + s_{2}) < 0 \; . \end{eqnarray} We immediately recognize the important role played by the last exponential function in (\ref{fold0.5}), since, if the condition $\mbox{${\rm Re}$} (s_{1}+s_{2}) < 0$ is not observed, the trace gives a divergent result. Thus, that condition imposes a constraint that defines a hierarchy. That is, on continuous phase space there is a hierarchical structure of mapping kernels allowing one to express a given phase-space function in terms of a Gaussian smoothing of another, and, as such, inverse relations do {\em not} exist. In other words, {\em the Gaussian folding hierarchical structure observed by the quasidistributions has its roots in the functional form of} $\langle 0 | q,p \rangle$. We stress this particular point as the discrete equivalent to equation (\ref{fold0.5}) {\em does not} imply a hierarchical relation. \section{The discrete mapping kernel} \subsection{Preliminaries} \subsubsection{Operator bases} Long ago Schwinger proposed the following set of operators to act as a basis on an operator space \begin{displaymath} {\bf S}(\eta,\xi) = \frac{1}{\sqrt{N}} {\bf U}^{\eta} {\bf V}^{\xi} \exp \left( \frac{\pi i}{N} \eta \xi \right) \; , \end{displaymath} where the ${\bf U}$'s and ${\bf V}$'s are the so-called Schwinger unitary operators \cite{schw}, $N$ is the dimension of the associated state space and the indices $\{ \eta,\xi \}$ run on any complete set of residues mod$(N)$; in particular we choose the closed interval $\left[ -\ell,\ell \right]$, with $\ell = (N-1)/2$. For simplicity, we shall restrict ourselves to the odd $N$ case. Even dimensionalities, for the purposes of this paper, can also be dealt with simply by working on non-symmetrized intervals. The set $\{ {\bf S}(\eta,\xi) \}_{\eta,\xi = -\ell,\ldots,\ell}$ spans a complete and orthonormal basis on the $N^{2}$ space of linear operators acting on finite complex vectorial spaces, in the sense that, as the trace operation stands as the inner product on operator spaces, any linear operator can be written as \begin{equation} \label{1} {\bf O} = \sum_{\eta,\xi = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf S}^{\dagger}(\eta,\xi) {\bf O} \right] {\bf S}(\eta,\xi) \; . \end{equation} The fundamental result \begin{displaymath} \mbox{${\rm Tr}$} \left[ {\bf S}^{\dagger}(\mu,\nu) {\bf S}(\eta,\xi) \right] = \delta_{\eta,\mu}^{[N]} \, \delta_{\xi,\nu}^{[N]} \end{displaymath} ensures that this decomposition is unique. The superscript $[N]$ on the Kroenecker deltas denotes that they are different from zero whenever their indices are mod$(N)$ congruent. The Schwinger basis elements also obey the property \cite{gama} \begin{equation} \label{2} {\bf S}^{\dagger}(\eta,\xi) = {\bf S}(-\eta,-\xi) \; . \end{equation} \subsubsection{Discrete Coherent States} The Schwinger operator bases elements also act as displacement operators on a particular reference state to form discrete coherent states as \cite{gapi2,gama} \begin{equation} \label{cohe} | \eta,\xi \rangle = \sqrt{N} {\bf S}(\eta,-\xi) |0,0 \rangle \; , \end{equation} where the reference state is written by means of the Jacobi $\vartheta_{3}$-function (whose explicit form is shown in Appendix A) as \begin{equation} \label{vacuo} |0,0 \rangle = \frac{1}{\mathcal{N}} \sum_{\gamma = - \ell}^{\ell} \vartheta_{3} \left( 2a \gamma | 2ia \right) | u_{\gamma} \rangle \; , \end{equation} where $\{ | u_{\gamma} \rangle \}_{\gamma = - \ell,\ldots,\ell}$ are the eigenstates of the unitary operator ${\bf U}$, \begin{displaymath} {\mathcal{N}}^{2} = \frac{1}{2 \sqrt{a}} \left[ \vartheta_{3}(0|i a) \vartheta_{3}(0|4ia) + \vartheta_{4}(0|ia) \vartheta_{2}(0|4ia) \right] \end{displaymath} is the normalization constant, and $a=(2N)^{-1}$. Due to the properties of the $\vartheta_{3}$-function, the reference state above is preserved under the action of the Fourier operator \cite{gama,mehta} \begin{displaymath} \mbox{\boldmath $\mathfrak{F}$} |0,0 \rangle = |0,0 \rangle \; , \end{displaymath} where \begin{displaymath} \mbox{\boldmath $\mathfrak{F}$} = \sum_{\gamma = -\ell}^{\ell} |v_{\gamma} \rangle \langle u_{\gamma} | \; , \end{displaymath} and $\{ | v_{\gamma} \rangle \}_{\gamma = - \ell,\ldots,\ell}$ are the eigenstates of ${\bf V}$, with $\langle u_{\mu} | v_{\gamma} \rangle = \exp [ (2 \pi i /N) \mu \gamma ]$. Parity of the $\vartheta_{3}$-function also ensures that $\langle u_{\kappa } | 0,0 \rangle = \langle u_{-\kappa} | 0,0 \rangle$, from which it follows \begin{equation} \label{par} \langle 0,0 | \mu,\nu \rangle = \langle 0,0 | -\mu,-\nu \rangle \; . \end{equation} There are, of course, a number different recipes of discrete coherent states, some of them also in connection with $\vartheta$-functions, for instance \cite{zhang,voros}. \subsection{The extended mapping kernel} Now let us define the extended mapping kernel as \begin{displaymath} {\bf S}^{(s)}(\eta,\xi) = {\bf S}(\eta,\xi) \left[ \mathcal{K}(\eta,\xi) \right]^{-s} \end{displaymath} where $s$ is a complex number satisfying $\left| s \right| \leq 1$, and $\mathcal{K}(\eta,\xi)= \langle 0,0| \eta,\xi \rangle$ denotes the overlap of coherent states explicitly calculated in Appendix A. The set $\{ {\bf S}(\eta,\xi) \}_{\eta,\xi = -\ell,\ldots,\ell}$ itself spans a complete and orthogonal basis on operator space. Nevertheless, we can go back to decomposition (\ref{1}), use equation (\ref{2}), and introduce convenient factors to get \begin{displaymath} {\bf O} = \sum_{\eta,\xi = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf S}(-\eta,-\xi) {\bf O} \right] {\bf S}(\eta,\xi) {\underbrace{[ \mathcal{K}(-\eta,-\xi)]^{s} [\mathcal{K}(\eta,\xi)]^{-s}}_{1}} \; , \end{displaymath} where equation (\ref{par}) has been used. Conveniently grouping the terms the new decomposition reads \begin{equation} \label{novdec} {\bf O} = \sum_{\eta,\xi = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf S}^{(-s)}(-\eta,-\xi) {\bf O} \right] {\bf S}^{(s)}(\eta,\xi) \; . \end{equation} Now, introducing the double Fourier transform of ${\bf S}^{(s)}(\eta,\xi)$, i.e. \begin{displaymath} {\bf T}^{(s)}(\eta,\xi) = \frac{1}{\sqrt{N}} \sum_{\mu,\nu = -\ell}^{\ell} {\bf S}^{(s)}(\eta,\xi) \exp \left[ - \frac{2\pi i}{N} (\eta \mu + \xi \nu ) \right] \; , \end{displaymath} and its Fourier inverse, we can, after a few steps, write equation (\ref{novdec}) as \begin{equation} \label{decomp} {\bf O} = \frac{1}{N} \sum_{\mu,\nu = -\ell}^{\ell} \mathcal{O}^{(-s)}(\mu,\nu) {\bf T}^{(s)}(\mu,\nu) \; , \end{equation} with $\mathcal{O}^{(-s)}(\mu,\nu) = \mbox{${\rm Tr}$} \left[ {\bf T}^{(-s)}(\mu,\nu) {\bf O} \right]$, defining a one-to-one mapping between operators and functions defined on a discrete phase-space $\{\mu ,\nu \}$, where explicitly \begin{equation} \label{novabase} {\bf T}^{(s)}(\mu,\nu) = \frac{1}{N} \sum_{\eta,\xi = -\ell}^{\ell} {\bf U}^{\eta} {\bf V}^{\xi} \exp \left[ -\frac{2\pi i}{N} ( \eta \mu + \xi \nu ) \right] \exp \left( \frac{\pi i}{N} \eta \xi \right) [ \mathcal{K}(\eta,\xi) ]^{-s} \; , \end{equation} and $\mathcal{K}(\eta,\xi)$ can be shown to be a sum of products of Jacobi $\vartheta$-functions (as seen in Appendix A), \begin{eqnarray} \label{fold} \mathcal{K}(\eta,\xi) &=& \frac{1}{4 \sqrt{a} \mathcal{N}^{2}} \left\{ \vartheta_{3} (a \eta | ia) \vartheta_{3} (a \xi | ia) + \vartheta_{3} (a \eta | ia) \vartheta_{4} (a \xi | ia) \exp (i \pi \eta) \right. \nonumber \\ & & + \left. \vartheta_{4} (a \eta | ia) \vartheta_{3} (a \xi | ia) \exp (i \pi \xi) + \vartheta_{4} (a \eta | ia) \vartheta_{4} (a \xi | ia) \exp \left[ i \pi (\eta + \xi + N) \right] \right\} \; . \end{eqnarray} The new kernel, written as in equation (\ref{novabase}), allows us to conclude that the above sum of products of $\vartheta$-functions plays, in the discrete phase-space, the role reserved to the Gaussians in the continuous case. \section{Properties} \subsection{Basic general properties} From the properties of the mapping kernel it is straightforward to obtain general properties of the associated functions in phase-space. We observe that {\em all} following properties correctly generalize the continuous CG ones. First we note that \begin{equation} \label{prop1} \textrm{(i)} \left[ {\bf T}^{(s)}(\mu,\nu) \right]^{\dagger} = {\bf T}^{(s^{\ast})}(\mu,\nu) \; , \end{equation} implying that the mapping kernel is Hermitian for real values of the parameter $s$. As a direct consequence, the phase-space representatives of Hermitian operators are real. Direct calculations also show that \begin{eqnarray} \label{prop2} \textrm{(ii)} &\! \! \!& \frac{1}{N} \sum_{\mu,\nu = -\ell}^{\ell} {\bf T}^{(s)}(\mu,\nu) = {\bf 1} \; , \\ \textrm{(iii)} &\! \! \!& \mbox{${\rm Tr}$} \left[ {\bf T}^{(s)}(\mu,\nu) \right] = 1 \; , \\ \textrm{(iv)} &\! \! \!& \mbox{${\rm Tr}$} \left[ {\bf T}^{(s)}(\mu,\nu) {\bf T}^{(-s)}(\mu^{\prime},\nu^{\prime}) \right] = N \delta_{\mu,\mu^{\prime}}^{[N]} \delta_{\nu,\nu^{\prime}}^{[N]} \; . \end{eqnarray} The property (iv) is a crucial one from which expression (\ref{decomp}) could be immediately obtained. From this property also follows the general result \begin{displaymath} \mbox{${\rm Tr}$} ( {\bf AB} ) = \frac{1}{N} \sum_{\mu,\nu = - \ell}^{\ell} \mathcal{A}^{(s)}(\mu,\nu) \mathcal{B}^{(-s)}(\mu,\nu) \; . \end{displaymath} In fact, property (iv) is a particular case of the general expression \begin{equation} \label{gentr} \mbox{${\rm Tr}$} \left[ {\bf T}^{(s)}(\mu,\nu) {\bf T}^{(t)}(\mu^{\prime},\nu^{\prime}) \right] = \frac{1}{N} \sum_{\eta,\xi = - \ell}^{\ell} \exp \left\{ \frac{2 \pi i}{N} \left[ \eta (\mu^{\prime} - \mu) + \xi (\nu^{\prime} - \nu) \right] \right\} [ \mathcal{K}(\eta,\xi) ]^{-(t+s)} \; , \end{equation} which is the counterpart of equation (\ref{fold0.5}). It must be stressed that this expression is {\em always} well defined, even for $\mbox{${\rm Re}$} (t+s) < 0$, as $\mathcal{K}(\eta,\xi)\neq 0$. \subsection{Particular cases} There are three important particular cases to be discussed: \begin{itemize} \item[i)\hspace{.8 em}(s=0)] In such a case it is easy to see that \begin{equation} \label{DWW} {\bf T}^{(0)}(\mu,\nu) = {\bf G}(\mu,\nu) \; , \end{equation} where ${\bf G}(\mu,\nu)$ is the mapping kernel introduced by Galetti and Piza, which is a discrete generalization of the Weyl-Wigner mapping kernel \cite{gapi1,gapi2,ruga,ruga2}. In that case, being $\mbox{\boldmath $\rho$}$ the density operator, equation (\ref{decomp}) would read, for ${\bf O} = \mbox{\boldmath $\rho$}$, \begin{displaymath} \mbox{\boldmath $\rho$} = \frac{1}{N} \sum_{\mu,\nu = -\ell}^{\ell} \mathcal{W}(\mu,\nu) {\bf G}(\mu,\nu) \; , \end{displaymath} with $\mathcal{W}(\mu,\nu) = \mbox{${\rm Tr}$} \left[ {\bf G}(\mu,\nu) \mbox{\boldmath $\rho$} \right]$ a discrete Wigner function. \item[ii)\hspace{.8 em}(s=-1)] A fundamental property of our mapping kernel is \begin{equation} \label{fund} {\bf T}^{(-1)}(\mu,\nu) = | \mu,\nu \rangle \langle \mu,\nu | \; , \end{equation} which can be proved decomposing the coherent state projector in the Schwinger operator basis as (using equations (\ref{1}) and (\ref{2})) \begin{displaymath} | \mu,\nu \rangle \langle \mu,\nu | = \frac{1}{N} \sum_{\eta,\xi = -\ell}^{\ell} {\bf U}^{\eta} {\bf V}^{\xi} \exp \left( \frac{i \pi}{N} \eta \xi \right) \mbox{${\rm Tr}$} \left[ | \mu,\nu \rangle \langle \mu,\nu | {\bf V}^{-\xi} {\bf U}^{-\eta} \exp \left( - \frac{i \pi}{N} \eta \xi \right) \right] \end{displaymath} which, by applying the definition of the coherent states, equation (\ref{cohe}), reads \begin{displaymath} | \mu,\nu \rangle \langle \mu,\nu | = \frac{1}{N} \sum_{\eta,\xi = -\ell}^{\ell} {\bf U}^{\eta} {\bf V}^{\xi} \langle 0,0 | {\bf V}^{\nu} {\bf U}^{-\mu} {\bf V}^{-\xi} {\bf U}^{-\eta} {\bf U}^{\mu} {\bf V}^{-\nu} | 0,0 \rangle \; , \end{displaymath} and using the Weyl commutation relation, ${\bf U}^{\alpha} {\bf V}^{\beta} = \exp [- (2 \pi i/N) \alpha \beta] {\bf V}^{\beta} {\bf U}^{\alpha}$, \begin{displaymath} |\mu ,\nu \rangle \langle \mu ,\nu |=\frac{1}{N}\sum_{\eta ,\xi=-\ell}^{\ell} \mathbf{U}^{\eta} \mathbf{V}^{\xi} \exp \left[ -\frac{2\pi i}{N}(\eta \mu +\xi \nu ) \right] \exp \left( \frac{\pi i}{N}\eta \xi \right) \langle 0,0|\eta ,\xi \rangle \; , \end{displaymath} where in the last step parity of $\langle 0,0 | \eta,\xi \rangle$ with respect to $\eta$ was used. This proves our assertion. As a consequence, the phase-space decomposition of the density operator, associated with this particular value of the parameter $s$, reads \begin{displaymath} \label{glauber} \mbox{\boldmath $\rho$} = \frac{1}{N} \sum_{\mu,\nu = -\ell}^{\ell} P(\mu,\nu) | \mu,\nu \rangle \langle \mu,\nu | \end{displaymath} allowing us to identify $P(\mu ,\nu )$ as a discrete Glauber-Sudarshan distribution. \item[iii)\hspace{.8 em} (s=1)] In this case we may write \begin{displaymath} \mbox{\boldmath $\rho$} = \frac{1}{N} \sum_{\mu,\nu = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf T}^{(-1)}(\mu,\nu) \mbox{\boldmath $\rho$} \right] {\bf T}^{(1)}(\mu,\nu) \; , \end{displaymath} which is simply \begin{displaymath} \mbox{\boldmath $\rho$} = \frac{1}{N} \sum_{\mu,\nu = -\ell}^{\ell} \mathcal{H}(\mu,\nu) {\bf T}^{(1)}(\mu,\nu) \; . \end{displaymath} By definition $\mathcal{H}(\mu,\nu) = \langle \mu,\nu | \mbox{\boldmath $\rho$} | \mu,\nu \rangle$ is positive definite, and it can be identified as a discrete Husimi function. \end{itemize} As any operator can be decomposed by the use of expression (\ref{decomp}), it follows that we are allowed to write \begin{displaymath} {\bf T}^{(-1)}(\mu,\nu) = \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf T}^{(-0)}(\sigma,\lambda ) {\bf T}^{(-1)} (\mu,\nu) \right] {\bf T}^{(0)}(\sigma,\lambda) \; , \end{displaymath} where the minus signal was kept only for clarity. We then use equation (\ref{gentr}) to write explicitly \begin{equation} \label{ffold} \mbox{${\rm Tr}$} \left[ {\bf T}^{(0)}(\sigma,\lambda) {\bf T}^{(-1)}(\mu,\nu) \right] = \frac{1}{N} \sum_{\eta,\xi = -\ell}^{\ell} \exp \left\{ \frac{2 \pi i}{N} [\eta (\mu - \sigma) + \xi (\nu -\lambda)] \right] \mathcal{K}(\eta,\xi) \; , \end{equation} that is, the discrete Fourier transform of the $\mathcal{K}(\eta,\xi)$ is the folding function. The above result can also be written in the compact form \begin{displaymath} \mbox{${\rm Tr}$} \left[ {\bf T}^{(0)}(\sigma,\lambda) {\bf T}^{(-1)}(\mu,\nu) \right] = \langle \mu,\nu | {\bf G}(\sigma,\lambda) | \mu,\nu \rangle \; , \end{displaymath} which is precisely the Wigner function associated with a coherent state $| \mu,\nu \rangle$. We therefore have \begin{equation} \label{hier1} {\bf T}^{(-1)}(\mu,\nu) = \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \langle \mu,\nu | {\bf G}(\sigma,\lambda) | \mu,\nu \rangle {\bf T}^{(0)}(\sigma,\lambda) \; . \end{equation} In the same form we now decompose ${\bf T}^{(0)}(\mu,\nu)$ as \begin{displaymath} {\bf T}^{(0)}(\mu,\nu) = \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf T}^{(0)}(\sigma,\lambda) {\bf T}^{(-1)}(\mu,\nu) \right] {\bf T}^{(1)}(\sigma,\lambda) \; , \end{displaymath} which allows us to use once again the above result for the trace and write \begin{equation} \label{hier2} {\bf T}^{(0)}(\mu,\nu) = \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \langle \mu,\nu | {\bf G}(\sigma,\lambda) | \mu,\nu \rangle {\bf T}^{(1)}(\sigma,\lambda) \; . \end{equation} Multiplying both equations (\ref{hier1}) and (\ref{hier2}) by the density operator $\mbox{\boldmath $\rho$}$ and taking the trace, we are led to the suggestive results \begin{eqnarray} \mathcal{H}(\mu,\nu) &=& \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \langle \mu,\nu | {\bf G}(\sigma,\lambda) | \mu,\nu \rangle \mathcal{W}(\sigma,\lambda) \; , \nonumber \\ \mathcal{W}(\mu,\nu) &=& \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \langle \mu,\nu | {\bf G}(\sigma,\lambda) | \mu,\nu \rangle P(\sigma,\lambda) \; , \nonumber \end{eqnarray} which are the discrete counterparts of the well known Gaussian smoothing that occurs in the continuous case, in agreement with the hierarchy present in that context. It must be stressed, however, that, opposed to the continuous case, it is now possible to write \begin{eqnarray} {\bf T}^{(0)}(\mu,\nu) &=& \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \Lambda (\mu - \sigma, \nu - \lambda) {\bf T}^{(-1)} (\sigma,\lambda) \nonumber \\ {\bf T}^{(1)}(\mu,\nu) &=& \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \Lambda (\mu - \sigma, \nu -\lambda) {\bf T}^{(0)} (\sigma,\lambda) \; , \nonumber \end{eqnarray} where \begin{equation} \label{invfold} \Lambda (\mu-\sigma,\nu-\lambda) = \frac{1}{N} \sum_{\eta,\xi = -\ell}^{\ell} \exp \left\{ \frac{2 \pi i}{N} [\eta (\mu-\sigma)+ \xi (\nu-\lambda)] \right\} [ \mathcal{K}(\eta,\xi) ]^{-1} \; , \end{equation} which, at least in principle, can always be calculated (we remind again that $\mathcal{K}(\eta,\xi)$ is finite and different from zero). Also very illustrative is the result that follows from the decomposition \begin{equation} \label{fold2} {\bf T}^{(-1)}(\mu,\nu) = \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} \mbox{${\rm Tr}$} \left[ {\bf T}^{(-1)}(\sigma,\lambda) {\bf T}^{(-1)} (\mu,\nu) \right] {\bf T}^{(1)}(\sigma,\lambda) \; . \end{equation} With \begin{displaymath} \mbox{${\rm Tr}$} \left[ {\bf T}^{(-1)}(\sigma,\lambda) {\bf T}^{(-1)}(\mu,\nu) \right] = \frac{1}{N} \sum_{\eta,\xi = -\ell}^{\ell} \exp \left\{ \frac{2 \pi i}{N} [\eta (\mu -\sigma) + \xi (\nu -\lambda) ] \right\} [\mathcal{K}(\eta,\xi)]^{2} \; , \end{displaymath} which can be shown to be $| \langle \mu,\nu | \sigma,\lambda \rangle |^{2}$, we have \begin{equation} \label{fold3} {\bf T}^{(-1)}(\mu,\nu) = \frac{1}{N} \sum_{\sigma,\lambda = -\ell}^{\ell} | \langle \mu,\nu | \sigma,\lambda \rangle |^{2} \, {\bf T}^{(1)} (\sigma,\lambda) \; . \end{equation} Thus $| \langle \mu,\nu | \sigma,\lambda \rangle |^{2}$ itself, which is the Husimi function associated with the discrete coherent-state $| \mu,\nu \rangle$, acts here as the smoothing function. \section{Continuum limit} Following the procedure detailed in both \cite{ruga,ruga2}, the continuum limit of the mapping kernel (\ref{novabase}) is reached as follows: we introduce the scaling parameter $\epsilon = (2 \pi /N )^{1/2}$, which will become infinitesimal as $N \rightarrow \infty$, and the two Hermitian operators \begin{equation} \label{29} {\bf P} = \sum_{\mu = -\ell}^{\ell} \mu \epsilon p_{0} | v_{\mu} \rangle \langle v_{\mu} | \qquad {\bf Q} = \sum_{\mu^{\prime} = -\ell}^{\ell} \mu^{\prime} \epsilon q_{0} | u_{\mu^{\prime}} \rangle \langle u_{\mu^{\prime}} | \; , \end{equation} constructed out of the projectors of the eigenstates of ${\bf U}$ and ${\bf V}$. The parameters $p_{0}$ and $q_{0}$, with $p_{0} q_{0} = \hbar = 1$, are chosen to be real, carrying units of momentum and position, respectively, while $\epsilon p_{0}$ and $\epsilon q_{0}$ are the distance between successive eigenvalues of the ${\bf P}$ and ${\bf Q}$ operators. Then, rewriting the Schwinger operators as \begin{equation} \label{28} {\bf V} = \exp \left( \frac{i \epsilon {\bf P}}{p_{0}} \right) \qquad {\bf U} = \exp \left( \frac{i \epsilon {\bf Q}}{q_{0}} \right) \; , \end{equation} and performing the change of variables $q^{\prime} = - q_{0} \epsilon \xi$, $p^{\prime} = p_{0} \epsilon \eta$, $p = p_{0} \epsilon \nu$ and $q = q_{0} \epsilon \mu$, we obtain \begin{displaymath} {\bf T}^{(s)}(q,p) = \sum_{q^{\prime} = - q_{0} \epsilon \ell}^{q_{0} \epsilon \ell} \sum_{p^{\prime} = - p_{0} \epsilon \ell }^{p_{0} \epsilon \ell} \frac{\Delta q^{\prime} \Delta p^{\prime}}{2 \pi} \exp \left[ - i p^{\prime} (q-{\bf Q}) \right] \exp \left[ i q^{\prime} (p-{\bf P}) \right] [\mathcal{K}(p^{\prime}/p_{0} \epsilon, - q^{\prime}/q_{0} \epsilon)]^{-s} \exp \left( - \frac{i}{2} q^{\prime} p^{\prime} \right) \; . \end{displaymath} As $N \rightarrow \infty$, it follows that $\Delta q^{\prime} \rightarrow dq^{\prime}$ and $\Delta p^{\prime} \rightarrow dp^{\prime}$. Since the continuum limit of the discrete coherent-states has been already discussed in \cite{gapi2,gama}, it is clear that the term $[ \mathcal{K}(p^{\prime}/p_{0} \epsilon, - q^{\prime}/q_{0} \epsilon)]^{-s}$, which is even, will go to $(\langle 0 |q^{\prime},p^{\prime} \rangle )^{-s}$. Therefore we end up with \begin{displaymath} {\bf T}^{(s)}(q,p) = \int_{-\infty}^{\infty} \frac{dp^{\prime} dq^{\prime}}{2 \pi} \exp \left[ - i p^{\prime}(q-{\bf Q}) \right] \exp \left[ i q^{\prime}(p-{\bf P}) \right] \exp \left( - \frac{i}{2} p^{\prime} q^{\prime} \right) (\langle 0 | q^{\prime},p^{\prime} \rangle )^{-s} \; , \end{displaymath} which is exactly the mapping kernel (\ref{II}) of Cahill and Glauber. \section{Concluding Remarks} The results obtained here show a genuine discrete mathematical structure which closely parallels the one of Cahill and Glauber. This was achieved pursuing the lines proposed in \cite{gapi1}, which makes use of the discrete Fourier transform of the Schwinger operator basis, to deal with the discrete phase-space problem. Now, we stress that expression (\ref{II}) is as simple as it is important, since it clarifies the role of the coherent states overlap within the CG approach. By its turn, the discrete coherent states proposed in \cite{gapi2,zhang,gama,voros} provide a natural path for a discrete extension of the CG formalism, while the properties of these states have played a crucial role as they led, for example to the basic equation (\ref{fund}). Thus, the coherent states overlap can be seen as the link between the discrete and continuous approaches. Furthermore, the continuum limit presented in section V ensures that the CG mapping scheme is correctly recovered through a limiting procedure which is mathematically consistent \cite{ruga2,bar1,bar2,bar3}. It is worth mentioning that Opatrn\'{y} {\em et al} \cite{opat} have pursued a goal similar to ours. Although both approaches share virtues, our formalism presents mathematical features that allow us to achieve farther reaching results. It is precisely the correct choice for the reference state, and the mathematical procedure adopted here, that lead to the obtention of such a wide set of important results. The use of Schwinger operators is crucial if one is concerned with the problem of ordering. As they are unitary shift operators, equation (\ref{novabase}) makes it clear that the associated expansion is necessarily linked to a particular ordering of ${\bf U}$ and ${\bf V}$, which can be directly connected to the ${\bf Q}$ and ${\bf P}$ ordering of the continuous case. Concerning the role of the Jacobi Theta functions in the discrete phase-space context -- they are implicit in the $\mathcal{K} (\eta,\xi)$ term --, comparison with the usual CG results makes it evident that the Gaussian (or anti-Gaussian) terms, which are present in the continuous case, are here replaced by the sum of products of $\vartheta$-functions (\ref{fold}), and its Fourier transform (\ref{ffold}); both play here the role of the smoothing functions. It is always important to emphasize the discrete case's peculiar features that do not have correspondence in the continuum. A plain example of these is expressed by the well-behaved function given by equation (\ref{invfold}), whose continuum limit clearly diverges, as the hierarchical structure presented in (\ref{fold1}) would imply. The finite character of the discrete scenario prevents such a behaviour since, even if some terms in equation (\ref{invfold}) might become large for large $N$, they remain always finite due to the behaviour of the $\vartheta$-functions. This allows one -- to give a extreme example -- to express, in the discrete scenario, the Glauber-Sudarshan function in terms of the Husimi function. Finally, it is worth mentioning that the mathematical formalism developed here opens new possibilities of investigations in quantum tomography and quantum teleportation. These considerations are under current research and will be published elsewhere (for instance, see reference \cite{nois}). \section*{Acknowledgments} This work has been supported by Funda\c{c}\~{a}o de Amparo \`{a} Pesquisa do Estado de S\~{a}o Paulo (FAPESP), Brazil, project nos. 03/13488-0 (MR), 01/11209-0 (MAM), and 00/15084-5 (MAM and MR). DG acknowledges partial financial support from the Conselho Nacional de Desenvolvimento Cient\'{\i}fico e Tecnol\'{o}gico (CNPq), Brazil.
{ "timestamp": "2005-05-30T21:01:30", "yymm": "0503", "arxiv_id": "quant-ph/0503054", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503054" }
\section{Introduction}\label{sec:intro} Often the experimentalist needs a low-noise preamplifier for the analysis of low-frequency components (below 10 Hz) from a 50 \ohm\ source. The desired amplifier chiefly exhibits low residual flicker and high thermal stability, besides low white noise. Thermal stability without need for temperature control is a desirable feature. In fact the problem with temperature control, worse than complexity, is that in a nonstabilized environment thermal gradients fluctuate, and in turn low-frequency noise is taken in. A low-noise amplifier may be regarded as an old subject, nonetheless innovation in analysis methods and in available parts provides insight and new design. The application we initially had in mind is the postdetection preamplifier for phase noise measurements~\cite{rubiola02rsi}. Yet, there resulted a versatile general-purpose scheme useful in experimental electronics and physics. \section{Design Strategy}\label{sec:strategy} The choice of the input stage determines the success of a precision amplifier. This issue involves the choice of appropriate devices and of the topology. Available low-noise devices are the junction field-effect transistor (JFET) and the bipolar transistor (BJT), either as part of an operational amplifier or as a stand-alone component. The white noise of these devices is well understood~\cite{van.der.ziel:noise-ssdc,van.der.ziel:fluctuations,netzer81pieee,erdi81jssc}. Conversely, flicker noise is still elusive and relies upon models, the most accredited of which are due to McWhorter~\cite{mcwhorter57} and Hooge~\cite{hooge69pla}, or on smart narrow-domain analyses, like~\cite{green85jpd-1,green85jpd-2,jamaldeen99jap}, rather than on a unified theory. Even worse, aging and thermal drift chiefly depend on proprietary technologies, thus scientific literature ends up to be of scarce usefulness. The JFET is appealing because of the inherently low white noise. The noise temperature can be as low as a fraction of a degree Kelvin. Unfortunately, the low noise of the JFET derives from low input current, hence a high input resistance (some M\ohm) is necessary. The JFET noise voltage is hardly lower than 5 \unit{nV/\sqrt{Hz}}, some five to six times higher than the thermal noise of a 50 \ohm\ resistor ($\sqrt{4kTR}=0.89$ \unit{nV/\sqrt{Hz}}). The JFET is therefore discarded in favor of the BJT\@. A feedback scheme, in which the gain is determined by a resistive network, is necessary for gain accuracy and flatness over frequency. Besides the well known differential stage, a single-transistor configuration is possible (Ref.~\cite{motchenbacher:low-noise:1ed}, page 123), in which the input is connected to the base and the feedback to the emitter. This configuration was popular in early audio hi-fi amplifiers. The advantage of the single-transistor scheme is that noise power is half the noise of a differential stage. On the other hand, in a dc-coupled circuit thermal effects are difficult to compensate without reintroducing noise, while thermal compensation of the differential stage is guaranteed by the symmetry of the base-emitter junctions. Hence we opt for the differential pair. \begin{table} \begin{sideways} \begin{minipage}{0.88\textheight} \caption{\label{tab:opa}% \vrule width0pt height2.5ex depth2ex Selection of some low-noise BJT amplifiers.} \centering \begin{tabular}{|c|cccc|c|c|}\hline & OP27\footnotemark[1] & LT1028\footnotemark[1] & MAT02\footnotemark[2] & MAT03\footnotemark[2] & \parbox{12ex}{unit} & \parbox{12ex}{{\footnotesize MAT03}\\ measured\footnotemark[3]% \vrule width0pt height0ex depth0.5ex} \\\hline \vrule width0pt height2.5ex depth0ex WHITE NOISE&&&&&&\\ noise voltage\footnotemark[4] $\sqrt{h_{0,v}}$ & 3 & 0.9 & 0.9 & 0.7 & \unit{nV/\sqrt{Hz}} & 0.8 \\ noise current\footnotemark[4] $\sqrt{h_{0,i}}$ & 0.4 & 1 & 0.9 & 1.4~\footnotemark[5]& \unit{pA/\sqrt{Hz}} & 1.2\\ noise power $2\sqrt{h_{0,v}h_{0,i}}$ &$2.4{\times}10^{-21}$ &$1.8{\times}10^{-21}$ &$1.6{\times}10^{-21}$ & $2.0{\times}10^{-21}$ & \unit{W/Hz} & $1.9{\times}10^{-21}$\\ noise temperature $T_w$ & 174 & 130 & 117 & 142 & K & 139 \\ optimum resistance $R_{b,w}$ & 7500 & 900 & 1000 & 500 & \ohm & 667 \\ $2{\times}50$\ohm-input noise & 3.3 & 1.55 & 1.55 & 1.5 & \unit{nV/\sqrt{Hz}} & 1.5~% \footnotemark[6]\\\hline \vrule width0pt height2.5ex depth0ex FLICKER NOISE&&&&&&\\ noise voltage\footnotemark[4] $\sqrt{h_{-1,v}}$ & 4.3 & 1.7 & 1.6 & 1.2 & \unit{nV/\sqrt{Hz}} & (~$0.4$~)% \footnotemark[7] \\ noise current\footnotemark[4] $\sqrt{h_{-1,i}}$ & 4.7 & 16 & 1.6 &n.\,a.& \unit{pA/\sqrt{Hz}} & 11 \\ noise power $2\sqrt{h_{-1,v}h_{-1,i}}$ & $4.1{\times}10^{-20}$ & $5.3{\times}10^{-20}$ & $5.1{\times}10^{-21}$ & -- & \unit{W/Hz} & (\ldots)\footnotemark[8] \\ 1-Hz noise temperature $T_f$ & 2950 & 3850 & 370 & -- & K & (\ldots)\footnotemark[8] \\ optimum resistance $R_{b,f}$ & 910 & 106 & 1000 & -- & \ohm & (\ldots)\footnotemark[8] \\ $2{\times}50$\ohm-input noise & 4.3 & 2.3 & 1.6 & -- & \unit{nV/\sqrt{Hz}} & 1.1~% \footnotemark[6] \\\hline \vrule width0pt height2.5ex depth0ex THERMAL DRIFT & 200 & 250 & 100 & 300 & nV/K & -- \\\hline \end{tabular} \footnotetext[1]{Low-noise operational amplifier.} \footnotetext[2]{Matched-transistor pair. MAT02 is \textsc{npn}, MAT03 is \textsc{pnp}. Data refer to the pair, biased at $I_C=1$ mA.} \footnotetext[3]{Some MAT03 samples measured in our laboratory. See Sec.~\protect\ref{sec:frontend}} \footnotetext[4]{Power-law model of the spectrum, voltage or current, $S(f)=h_0+h_{-1}f^{-1}+h_{-2}f^{-2}+\ldots$} \footnotetext[5]{Obtained from the total noise with 100 k\ohm\ input resistance.} \footnotetext[6]{Measured on the complete amplifier (Sec.~\protect\ref{sec:results}), independently of the measurement of the above $S_v$ and $S_i$.} \footnotetext[7]{Derives from the noise current through $r_{bb'}$. See Sec.~\protect\ref{sec:results}.} \footnotetext[8]{Can not be compared to other data because voltage and current are correlated. See Sec.~\protect\ref{sec:results}.} \end{minipage} \end{sideways} \end{table} Table~\ref{tab:opa} compares a selection of low-noise bipolar amplifiers. The first columns are based on the specifications available on the web sites~\cite{www.analog-devices,www.linear-technology}. The right-hand column derives from our measurements, discussed in Secs.~\ref{sec:frontend} and \ref{sec:results}. Noise is described in terms of a pair of random sources, voltage and current, which are assumed independent. This refers to the Rothe-Dahlke model~\cite{rothe56ire}. Nonetheless, a correlation factor arises in measurements, due to the distributed base resistance $r_{bb'}$. Whether and how $r_{bb'}$ is accounted for in the specifications is often unclear. The noise spectra are approximated with the power law $S(f)=\sum_{\alpha}h_\alpha f^\alpha$. This model, commonly used in the domain of time and frequency, fits to the observations and provides simple rules of transformation of spectra into two-sample (Allan) variance $\sigma_y(\tau)$. This variance is an effective way to describe the stability of a quantity $y$ as a function of the measurement time $\tau$, avoiding the divergence problem of the $f^\alpha$ processes in which $\alpha\le-1$. References~\cite{rutman78pieee} and \cite{rubiola01im} provide the background on this subject, and application to operational amplifiers. The noise power spectrum $2\sqrt{h_vh_i}$ is the minimum noise of the device, i.e., the noise that we expect when the input is connected to a cold (0~K) resistor of value $R_b=\sqrt{h_{v}/h_{i}}$, still under the assumption that voltage and current are uncorrelated. When the input resistance takes the optimum value $R_b$, voltage and current contributions to noise are equal. The optimum resistance is $R_{b,w}$ for white noise and $R_{b,f}$ for flicker. Denoting by $f_{c}$ the corner frequency at which flicker noise is equal to white noise, thus $f_{c,v}$ for voltage and $f_{c,i}$ for current, it holds that $R_{b,w}/R_{b,f}=\sqrt{f_{c,i}/f_{c,v}}$. Interestingly, with most bipolar operational amplifiers we find $f_{c,i}/f_{c,v}\approx50{-}80$, hence $R_{b,w}/R_{b,f}\approx7{-9}$. Whereas we have no explanation for this result, the lower value of the flicker optimum resistance is a fortunate outcome. The equivalent temperature is the noise power spectrum divided by the Boltzmann constant $k=1.38{\times}10^{-23}$ J/K\@. A crucial parameter of Table~\ref{tab:opa} is the total noise when each input is connected to a 50~\ohm\ resistor at room temperature. This calculated value includes noise voltage and current, and the thermal noise of the two resistors. In a complete amplifier two resistors are needed, at the input and in the feedback circuit. Still from Table~\ref{tab:opa}, the transistor pairs show lower noise than the operational amplifiers, although the PNP pair is only partially documented. Experience indicates that PNP transistors are not as good as NPN ones to most extents, but exhibit lower noise. In other domains, frequency multipliers and radio-frequency oscillators make use of PNP transistors for critical application because of the lower flicker noise. Encouraged by this fact, we tried a differential amplifier design based on the MAT03, after independent measurement of some samples. \section{Input Stage}\label{sec:frontend} \begin{figure}[t] \centering\includegraphics[scale=1]{measure-mat} \caption{Noise measurement of a transistor pair. For clarity, the distributed base resistance $r_{bb'}$ is extracted from the transistors.} \label{fig:measure-mat} \end{figure} The typical noise spectrum of the MAT03, reported in the data sheet, shows an anomalous slope at low frequencies (0.1--1 Hz), significantly different from $f^{-1}$. This is particularly visible at low collector current (10--100 $\mu$A), but also noticeable at $I_C=1$ mA\@. We suspect that the typical spectrum reflects the temperature fluctuation of the environment through the temperature coefficient of the offset voltage $V_{OS}$ rather than providing information on the flicker noise inherent in the transistor pair. The measurement of a spectrum from 0.1 Hz takes some 5 min. At that time scale, in a normal laboratory environment the dominant fluctuation is a drift. If the drift is linear, $v(t)=ct$ starting at $t=0$, the Fourier transform is $V(\omega)=j\pi c\delta(\omega)-c/\omega^2$. Dropping off the term $\delta(\omega)$, which is a dc term not visible in a log-log scale, the power spectrum density, i.e., the squared Fourier transform, is \begin{equation} \label{eq:f-drift} S_v(\omega)=\frac{c^2}{\omega^4} \qquad\mbox{or}\qquad S_v(f)=\frac{(2\pi)^4c^2}{f^4}~~. \end{equation} A parabolic drift---seldom encountered in practice---has a spectrum proportional to $f^{-6}$, while a smoothly walking drift tends to be of the $f^{-5}$ type. As a consequence, a thermal drift can be mistaken for a random process of slope $f^{4}$ to $f^{5}$, which may hide the inherent $f^{-1}$ noise of the device. For this reason, the test circuit (Fig.~\ref{fig:measure-mat}) must be enclosed in an appropriate environment. We used, with similar results, a Dewar flask coupled to the environment via a heat exchanger, and a metal box mounted on a heat sink that has a mass of 1 kg and a thermal resistance of 0.6 K/W\@. These odd layouts provide passive temperature stabilization through a time constant and by eliminating convection, and evacuate the small amount of heat (200 mW) dissipated by the circuit. \begin{figure}[t] \centering\includegraphics[scale=0.8,angle=0]{f695} \caption{Typical spectrum of the noise voltage.} \label{fig:f695} \end{figure} Due to the low value of $r_{bb'}$ (15--20 \ohm) the current measurement can be made independent of voltage noise, but not vice versa. Thus, we first measure the noise current setting $R_B=8$~k\ohm, which is limited by the offset current; then we measure the noise voltage setting $R_B=10$~\ohm. A technical difficulty is that at 1 Hz and below most spectrum analyzers---including our one---must be coupled in dc, hence high offset stability is needed in order to prevent saturation of the analyzer. The measured spectra are $S_i(f)=1.45{\times}10^{-24}+1.2{\times}10^{-22}f^{-1}$ \unit{A^2/Hz} (i.e., 1.2\unit{pA/\sqrt{Hz}} white, and 11\unit{pA/\sqrt{Hz}} flicker), and $S_v(f)=10^{-18}+1.8{\times}10^{-19}f^{-1}$ \unit{V^2/Hz} (i.e., 1\unit{nV/\sqrt{Hz}} white, and 425\unit{pV/\sqrt{Hz}} flicker). The current spectrum is the inherent noise current of the differential pair. Conversely, with the voltage spectrum (Fig.~\ref{fig:f695}) we must account for the effect of $R_B$ and $r_{bb'}$. With our test circuit, the expected white noise is $h_{0,v}=4kTR+2qI_BR\simeq1.7{\times}10^{-20}R$ \unit{V^2/Hz}, which is the sum of thermal noise and the shot noise of the base current $I_B$. $R=2(R_B+r_{bb'})$ is the equivalent base resistance, while the shot noise of the collector current is neglected. Assuming $r_{bb'}=16$~\ohm\ (from the data sheet), the estimated noise is $h_{0,v}\simeq9{\times}10^{19}$ \unit{V^2/Hz}. This is in agreement with the measured value of $10^{-18}$ \unit{V^2/Hz}. Then, we observe the effect of the current flickering on the test circuit is $R^2h_{-1,i}\simeq1.6{\times}10^{-19}$ \unit{V^2/Hz}. The latter is close to the measured value $1.8{\times}10^{-19}$ \unit{V^2/Hz}. Hence, the observed voltage flickering derives from the current noise through the external resistors $R_B$ and the internal distributed resistance $r_{bb'}$ of the transistors. Voltage and current are therefore highly correlated. As a further consequence, the product $2\sqrt{h_{-1,v}h_{-1,i}}$ is not the minimum noise power, and the ratio $\sqrt{h_{-1,v}/h_{-1,i}}$ is not the optimum resistance. The corresponding places in Table~\ref{tab:opa} are left blank. Due to the measurement uncertainty, we can only state that a true independent voltage flickering, if any, is not greater than $4{\times}10^{-20}$ \unit{A^2/Hz}. The same uncertainty affects the optimum resistance $R_{b,f}$, which is close to zero. The measured white noise is in agreement with the data sheet. On the other hand, our measurements of flicker noise are made in such unusual conditions that the results should not be considered in contradiction with the specifications, as the specifications reflect the the low-frequency behavior of the device in a normal environment. \section{Implementation and Results}\label{sec:results} \begin{figure}[t] \centering\includegraphics[scale=1]{scheme} \caption{Scheme of the low-noise amplifier.} \label{fig:scheme} \end{figure} Figure~\ref{fig:scheme} shows the scheme of the complete amplifier, inspired to the ``super low-noise amplifier'' proposed in Fig.~3a of the MAT03 data sheet. The NPN version is also discussed in Ref.~\cite{franco:opa} (p.~344). The original circuit makes use of three differential pairs connected in parallel, as it is designed for the lowest white noise with low impedance sources ($\ll50$~\ohm), like coil microphones. In our case, using more than one differential pair would increase the flicker because of current noise. The collector current $I_C=1.05$ mA results as a trade-off between white noise, which is lower at high $I_C$, dc stability, which is better at low dissipated power, flicker, and practical convenience. The gain of the differential pair is $g_mR_C=205$, where $g_m=I_C/V_T=41$~mA/V is the transistor transconductance, and $R_C=5$ k\ohm\ is the collector resistance. The overall gain is $1+R_G/R_B\simeq500$. Hence the gain of the OP27 is of 2.5, which guarantees the closed-loop stability (here, oscillation-free operation). If a lower gain is needed, the gain of the differential stage must be lowered by inserting $R_A$. The trick is that the midpoint of $R_A$ is a ground for the dynamic signal, hence the equivalent collector resistance that sets the gain is $R_C$ in parallel to $\frac{1}{2}R_G$. The bias current source is a cascode Wilson scheme, which includes a light emitting diode (LED) that provides some temperature compensation. The stability of the collector resistors $R_C$ is a crucial point because the voltage across them is of 5~V\@. If each of these resistors has a temperature coefficient of $10^{-6}$/K, in the worst case there results a temperature coefficient of 10 $\mu$V/K at the differential output, which is equivalent to an input thermal drift of 50~nV/K\@. This is 1/6 of the thermal coefficient of the differential pair. In addition, absolute accuracy is important in order to match the collector currents. This is necessary to take the full benefit from the symmetry of the transistor pair. \begin{figure}[t] \centering\includegraphics[width=\textwidth]{franck-ampli-small} \caption{Prototype of the low-noise amplifier.} \label{fig:prototype} \end{figure} Two equal amplifiers are assembled on a printed circuit board, and inserted in a $10{\times}10{\times}2.8$ \unit{cm^3}, 4 mm thick aluminum box (Fig.~\ref{fig:prototype}). The box provides thermal coupling to the environment with a suitable time constant, and prevents fluctuations due to convection. $LC$ filters, of the type commonly used in HF/VHF circuits, are inserted in series to the power supply, in addition to the usual bypass capacitors. For best stability, and also for mechanical compatibility with our equipment, input and output connector are of the SMA type. Input cables should not PTFE-insulated because of piezoelectricity (see the review paper~\cite{fukada00uffc}). \begin{figure}[t] \centering\includegraphics[scale=0.8]{f691} \caption{Residual noise of the complete amplifier, input terminated to a 50~\ohm\ resistor.} \label{fig:f691} \end{figure} Figure~\ref{fig:f691} shows the noise spectrum of one prototype input terminated to a 50~\ohm\ resistor. The measured noise is $\sqrt{h_0}=1.5$ \unit{nV/\sqrt{Hz}} (white) and $\sqrt{h_{-1}}=1.1$ \unit{nV/\sqrt{Hz}} (flicker). The corner frequency at which the white and flicker noise are equal is $f_c=0.5$ Hz. Converting the flicker noise into two-sample (Allan) deviation, we get $\sigma_v(\tau)=1.3$ nV, independent of the measurement time $\tau$. Finally, we made a simple experiment aimed to explain in practical terms the importance of a proper mechanical assembly. We first removed the Al cover, exposing the circuit to the air flow of the room, yet in a quiet environment, far from doors, fans, etc., and then we replaced the cover with a sheet of plain paper (80 \unit{g/m^2}). The low-frequency spectrum (Fig.~\ref{fig:f694}) is $5{\times}10^{-19}f^{-5}$ \unit{V^2/Hz} in the first case, and about $1.6{\times}10^{-19}f^{-4}$ \unit{V^2/Hz} in the second case. This indicates the presence of an irregular drift, smoothed by the paper protection. Interestingly, Hashiguchi~\cite{sikula03arw} reports on thermal effects with the same slope and similar cutoff frequencies, observed on a low-noise JFET amplifier for high impedance sources. \begin{figure}[t] \centering\includegraphics[scale=0.8]{f694} \caption{Thermal effects on the amplifier.} \label{fig:f694} \end{figure} \def\bibfile#1{/Users/rubiola/Documents?workocs/bib/#1} \bibliographystyle{amsalpha}
{ "timestamp": "2005-03-01T15:06:51", "yymm": "0503", "arxiv_id": "physics/0503012", "language": "en", "url": "https://arxiv.org/abs/physics/0503012" }
\section{Introduction} In this paper we study some class of division algebras over a Laurent series field with arbitrary residue field. Namely, we study division algebras which satisfy the following condition: there exists a section $\bar{D} \hookrightarrow D$ of the residue homomorphism $D\rightarrow \bar{D}$, where $D$ is a central division algebra over a complete discrete valued field $F=k((t))$. We say that these division algebras are splittable. If $char k=0$, all such division algebras are tame and therefore belong to the group of tame division algebras, which was carefully studied in the papers \cite{JW} and \cite{PY} even in a much more general situation of a henselian field $F$ of arbitrary characteristic. So, we consider mostly wild division algebras. An extensive analysis of the wild division algebras of degree $p$ over a field $F$ with complete discrete rank 1 valuation with $char (\bar{F})= p$ was given by Saltman in \cite{Sa} (Tignol in \cite{Ti} analyzed more general case of the defectless division algebras of degree $p$ over a fild $F$ with Henselian valuation). Here we study splittable division algebras of arbitrary index. This class (which is not a subgroup in $Br (F)$) contains a class of good splittable division algebras (see the definition in section 2), which posess several beautiful properties. In particular, we prove a decomposition theorem for such algebras. This theorem is a generalization of the decomposition theorems for tame division algebras given by Jacob and Wadsworth in \cite{JW}. For arbitrary splittable division algebras we give only several assorted results, and the study of this class is far from to be complete. Nevertheless, we investigate here technical tools, which are important for the study of such algebras, and prove a relation between the level and a higher order level for some splittable division algebras (see section 6). We hope this technique will be applied to the study of the cyclisity question for certain division algebras od degree $p^k$. As an application we get several results, which are partly well known (see proposition \ref{cyclisity}) and party not. In particular, we get the positive answer on the following conjecture: the exponent of $A$ is equal to its index for any division algebra $A$ over a $C_2$-field $F= F_1((t_2))$, where $F_1$ is a $C_1$-field. Here is a brief overview of this paper. In section 2 we give a definition of splittable and good splittable division algebras and prove that all tame division algebras over $F=k((t))$ are good splittable. Section 3 contains the most important technical tools for the study of splittable division algebras. We define a notion of $\delta$-maps and investigate a theory of $\delta$-maps for such algebras. In this section we define also the notion of a local height, which is a possible generalization of Saltman's level. In section 4 we prove the period-index conjecture metioned above. This section contains also a small history of the question known to the author. We note that the proof does not use all the results from section 3. In section 5 we study good splittable division algebras and prove the decomposition theorem. In section 6 we reprove some results of Saltman about semiramified division algebras of index $p$ over $F$ using the technique from section 3. Then we define a notion of a higher order level and prove several general properties of splittable division algebras satisfying the following condition: $Z(\bar{D})/\bar{F}$ is a simple extension. At the end of section we put several open questions. We use the notation of \cite{JW}. We always denote by $D$ a division algebra finite dimensional over its center $F=k((t))= Z(D)$. Recall that any Henselian valuation on $F$ has a unique extension to a valuation on $D$. We denote the valuation on $F$ by $v$ and its unique extension on $D$ by $w$. Given a valuation $w$ on $D$, we denote by $\Gamma_{D}$ its value group, by $V_D$ its valuation ring, by $M_D$ its maximal ideal and by $\bar{D}= V_D/M_D$ its residue division ring. By \cite{S}, p.21 one has the fundamental inequality $$ [D:F]\ge |\Gamma_D:\Gamma_E| \cdot [\bar{D}:\bar{F}]. $$ $D$ is called defectless over $F$ if equality holds and defective otherwise. It is known that $D$ is defectless if it has a discrete valuation of rank 1. Jacob and Wadsworth in \cite{JW} introduced the basic homomorphism $$ \theta_D:\Gamma_D/\Gamma_F\rightarrow Gal(Z(\bar{D})/\bar{F}) $$ induced by conjugation by elements of $D$. They showed that $\theta_D$ is surjective and $Z(\bar{D})$ is the compositum of an abelian Galois and a purely inseparable extension of $\bar{F}$. We say $D$ is tame division algebra if $char (\bar{F})= 0$ or $char (\bar{F})= q\ne 0$, $D$ is defectless over $F$, $Z(\bar{D})$ is separable over $\bar{F}$, and $q{\not |}|ker(\theta_D)|$. We say $D$ is wild division algebra if it is non tame. We call a division algebra $D$ {\it inertially split} if $Z(\bar{D})$ is separable over $\bar{F}$, the map $\theta_D$ is an isomorphism, and $D$ is defectless over $F$. \bigskip {\bf Acknowledgements} I am grateful to Professor A. N. Parshin, Professor E.-W. Zink, and M. Grabitz for useful discussions and attention to my work. I am very grateful to Professor A.Wadsworth for carefully reading my paper and for showing me a mistake in the very first version of this paper and to Professor V.I.Yanchevskii for valuable discussions during his visit in Berlin. Finally, I thank my wife Olga for her support and encouragement. \section{Cohen's theorem} Recall one definition from \cite{Zh}. \begin{defi} A division algebra $D$ is said to be splittable if there is a homomorphism $\bar{D} \hookrightarrow {\cal O}_D\subset D$ that is a section of the map ${\cal O}_D \rightarrow \bar{D}$. \end{defi} There is a natural question if there exists a generalization of Cohen's theorem, i.e. is any central division algebra splittable or not. It is not true if a division algebra is not finite dimensional over its centre, as Dubrovin's example in \cite{Zh} shows. It is not true also for some finite dimensional division algebras, as the example to theorem 2.7. in \cite{Sa} shows. But it is true for tame division algebras over complete discrete valued fields. This easily follows from results of Jacob and Wadsworth \cite{JW} (compare with \cite{Zh}, Th.1). \begin{th} \label{Cohen} Let $(F,v)$ be a valued field which is complete with respect to a discrete rank 1 valuation $v$. Suppose $char F= char \bar{F}$. Let $D$ be a tame division algebra with $Z(D)= F$ and $[D:F]< \infty$. Then there exists a section $\bar{D}\hookrightarrow D$ of the residue homomorphism $D\rightarrow \bar{D}$. \end{th} {\bf Proof.} Since $F$ is a complete field, $F$ is a Henselian field and $v$ extends uniquely to a valuation $w$ on $D$. Since $D$ is tame, $Z(\bar{D})/\overline{Z(D)}$ is a cyclic Galois extension. There exists an inertial lift $Z$ of $Z(\bar{D})$ over $F$, $Z$ is Galois over $F$, and by classical Cohen's theorem there exists a section $\tilde{Z}(\bar{D})\hookrightarrow Z$. Consider the centraliser $C= C_D(Z)$ of $Z$ in $D$. Then we have $\bar{C}= \bar{D}$. Indeed, by Double Centraliser Theorem we have $[D:F]= [C:F][Z:F]$ and $[Z:F]= |Gal (Z(\bar{D})/\bar{F})|$. By \cite{JW}, prop.1.7 a homomorphism $\theta_D: \Gamma_D/\Gamma_F\rightarrow Gal(Z(\bar{D})/\bar{F})$ is surjective, so for any parameter $z$ we have $\theta_D(w(z))= \sigma$, where $<\sigma > = Gal (Z(\bar{D})/\bar{F})$. It is clear that $z\notin C$. Now let $u_1,\ldots ,u_{[C:F]}$ be a $F$-basis of $C$. It is easy to see that the elements $u_j, zu_j, \ldots , z^{n-1}u_j $, $j= 1,\ldots ,[C:F]$, where $n= ord (\sigma )$, the order of $\sigma $, are linearly independent, so form a basis for $D$ over $F$. Since $$ w(F\langle zu_j, \ldots , z^{n-1}u_j, j= 1,\ldots ,[C:F]\rangle )\cap \Gamma_C= 0, $$ where $F\langle zu_j, \ldots , z^{n-1}u_j, j= 1,\ldots ,[C:F]\rangle$ denote a vector space in $D$ over $F$ generated by elements $u_jz^i$, this implies that for any element $x\in D$ with $w(x)= 0$ we can find elements $r_1, \ldots r_{[C:F]}\in F$ such that $x= r_1u_1+\ldots +r_{[C:F]}u_{[C:F]}\mbox{\quad mod \quad} M_D$. Hence $\bar{C}= \bar{D}$. Note that $C$ is an unramified division algebra. Indeed, by \cite{JW}, th.2.8, th.2.9 $C$ contains a copy of the inertial lift of a maximal separable subfield in $\bar C$, say $\tilde C$. Then the centralizer $C_C(\tilde{C})$ must be a totally ramified division algebra, i.e. it is trivial and $\tilde C$ is a maximal subfield. So, $C$ must be unramified. Fix an embedding $i: \bar F \hookrightarrow F$. It can be extended to the embedding $i':\bar Z \hookrightarrow Z$, $i'|_{\bar F}=i$ by Hensel lemma. Now consider the algebra $A= \bar{C}\otimes_{\bar Z}Z(C)$. It is easy to see that $A$ is an unramified division algebra with $\bar{A}= \bar{C}= \bar{D}$. Therefore by \cite{Az}, Th.31, $A\cong C$; so there exists a section $\bar{D}\hookrightarrow C$. The theorem is proved.\\ $\Box$\\ Later we will see that much more can be said about good splittable algebras: \begin{defi} \label{goodsplit} A division algebra $D$ is called good splittable if there exists a section $s:\bar D\hookrightarrow D$ compatible with an embedding $i:\overline{Z(D)}\hookrightarrow Z(D)$, i.e. $s(\overline{Z(D)})=i(\overline{Z(D)})\subset Z(D)$. \end{defi} It's easy to see that all tame division algebras are good splittable, because by Hensel lemma any embedding $\overline{Z(D)}\hookrightarrow Z(D)$ can be uniquely extended to any separable extension of $Z(D)$. It is interesting to know what kind of splittable division algebras are good splittable. By theorem 3.9. in \cite{Sa} even a splittable division algebra $D$ of degree $p=char D$ is not a good splittable algebra if the level of $D$ (the notion of level we will recall in section 3, see remark to lemma \ref{svva}) is divisible by $p$. Nevertheless, it is an open question whether it is true or not, for example, for division algebras with $\bar{D}=Z(\bar{D})$ such that $\bar{D}/\bar{F}$ is a simple extension and the local height (see the definition in the same remark) is not divisible by $p$. We will discuss this question in section 6. \section{Delta-maps of splittable algebras} In this section we develop some ideas from \cite{Zh}, where some properties of $\delta$-maps for special kind of local skew fields were studied. Technical properties of $\delta$-maps play the main role in all our results. Here we will give a list of these properties. Let $D$ be a finite dimensional division algebra over a complete valued field $F= k((t))$. Let $w$ be a unique extension of the valuation $v$ to $D$. We will denote by $z$ any parameter of $D$, i.e. any element with $\langle w(z)\rangle = \Gamma_D$. Consider the ring ${\mbox{\dbl Z}} \langle\alpha, \delta \rangle$ of noncommutative polinomials in two variables. Define the map $$ \sigma :{\mbox{\dbl Z}} \langle\alpha , \sigma \rangle\rightarrow {\mbox{\dbl Z}} \langle\alpha ,\delta , \delta_i; i\ge 1\rangle , $$ $$ \sigma (\alpha^{a_1}\delta^{b_1}\ldots \alpha^{a_n}\delta^{b_n})= \alpha^{a_1}\delta_{b_1}\ldots \delta_{b_{n-1}}\alpha^{a_n-1}\delta^{b_n}, $$ where $a_1, b_n\ge 0$, $a_i, b_j\ge 1$, $i>1$, $j<n$ for every word in ${\mbox{\dbl Z}} \langle\alpha , \delta \rangle$. Let $S_i^k\in {\mbox{\dbl Z}} \langle\alpha ,\delta \rangle$, $i\ge k$, $i\ge 1$ be polynomials given by the following formula: $$ S_i^k= \sum_{\tau\in S_i/G}\tau (\underbrace{\alpha\ldots \alpha}_{i-k}\underbrace{\delta\ldots \delta}_{k}), $$ where $S_i$ is a permutation group and $G$ is an isotropy subgroup. \begin{lemma}{(\cite{Zh}, lemma 2)} The polynomials $S_i^k$ satisfy the following property: $$ S_i^i= \delta^i, \mbox{\quad} S_i^0= \alpha^i, \mbox{\quad} S_{i+1}^{k+1}= \alpha S_i^{k+1}+\delta S_i^k $$ \end{lemma} For any splittable division algebra can be defined a notion of $\delta$-maps: \begin{prop}{(\cite{Zh}, prop. 1,2)} \label{ooo} Let $D$ be a splittable division algebra. Fix some parameter $z$ and some embedding $u: \bar{D} \hookrightarrow D$. Then $D$ is isomorphic to a division algebra $\bar{D}((z))$, which is defined to be the vector space of series with multiplication defined by the formula $$ zaz^{-1}= \alpha (a)+\delta_1(a)z+\delta_2(a)z^2+\ldots ,\mbox{\quad} a\in \bar{D}, $$ where $\alpha :\bar{D}\rightarrow \bar{D}$ is an automorphism and $\delta_i:\bar{D}\rightarrow \bar{D}$ are linear maps such that the map $\delta_i$ satisfy the identity $$\delta_i(ab)= \sum_{k= 0}^i\sigma(\delta^{i-k}\alpha) (a)\sigma(S_i^k\alpha) (b),\mbox{\quad} a,b\in \bar D $$ \end{prop} {\bf Remark} Note that the values $\sigma (S_i^k\alpha )$ and $\sigma (\delta^{i-k}\alpha )$ belong to the subring ${\mbox{\dbl Z}} \langle\alpha , \delta_i, i\ge 1\rangle$, so the formula is well defined. Note that $\delta$-maps depend on the choice of a parameter and an embedding. The automorphism $\alpha$, as it easy to see, depend only on the choice of a parameter. In the proposition we identify $\bar{D}$ with $u(\bar{D})$. \begin{corol}{(\cite{Zh}, corol. 1)} \label{formuly} Suppose $\alpha= Id$. Then $$ \delta_i(ab)= \delta_i(a)b+ \sum_{k= 1}^{i}\delta_{i-k}(a)\sum_{(j_1,\ldots ,j_l)} C_{i-k+1}^l\delta_{j_1}\ldots \delta_{j_l}(b), $$ where $\delta_0=\alpha$ and the second sum is taken over all the vectors $(j_1,\ldots ,j_l)$ such that $0< l\le min\{i-k+1, k\}$, $j_m\ge 1$, $\sum j_m= k$. \end{corol} Further we will need even more general definition. \begin{defi} \label{maps} In the situation of proposition \ref{ooo} let us define maps ${}_m^{(z,u)}\delta_i: \bar{D}\rightarrow \bar{D}$, $m\in{\mbox{\dbl Z}}$, $i\in {\mbox{\dbl N}}$ as follows. $$ z^maz^{-m}= u({}^{(z)}\alpha^m(\bar{a}))+u({}_m^{(z,u)}\delta_1(\bar{a}))z+ u({}_m^{(z,u)}\delta_2(\bar{a}))z^2+\ldots ,\mbox{\quad} a\in u(\bar{D}). $$ If $m= 0$, put ${}_m^{(z,u)}\delta_i= 0$. \end{defi} Note that ${}^{(z)}\alpha |_{Z(\bar{D})}$ does not depend on the choice of $z$. Note that if ${}^{(z)}\alpha= id$, then ${}_m^{(z,u)}\delta_i= 0$ for $m= p^k$, where $k$ is sufficiently large, $k$ depends on $i$. Moreover, ${}_m^{(z,u)}\delta_i= {}_{m+p^k}^{(z,u)}\delta_i$ for $k$ sufficiently large. We will use also the following notation: $$ {}_m^{(z,u)}\tilde{\delta_i}= {}_{-m}^{(z,u)}\delta_i, \mbox{\quad} {}_{1}^{(z,u)}\delta_i= {}^{(z,u)}\delta_i $$ Sometimes, we will write ${}_m\delta_i$ instead of ${}_m^{(z,u)}\delta_i$ and ${}_m^{(z,u)}\delta_i(a)$ instead of $u({}_m^{(z,u)}\delta_i(\bar{a}))$ whenever the context is clear. Immediately from the definition follows \begin{lemma} \label{triviall} In the situation of definition \ref{maps} we have (i) for $|m|>1$ $$ {}_m^{(z,u)}\delta_i(a)={}^{(z)}\alpha^{sign(m)} ({}_{sign(m)(|m|-1)}^{(z,u)}\delta_i(a))+ {}_{sign(m)}^{(z,u)}\delta_i({}^{(z)}\alpha^{sign(m)(|m|-1)}(a))+ $$ $$ \sum_{j=1}^{i-1}{}_{sign(m)}^{(z,u)}\delta_j({}_{sign(m)(|m|-1)}^{(z,u)}\delta_{i-j}(a)), $$ where $sign(m)=m/|m|$, $a\in \bar{D}$; (ii) for any $m\ne 0$ $$ {}^{(z)}\alpha^{-m} ({}_{m}^{(z,u)}\delta_i)+ {}_{-m}^{(z,u)}\delta_i({}^{(z)}\alpha^{m})+ \sum_{j=1}^{i-1}{}_{-m}^{(z,u)}\delta_j({}_{m}^{(z,u)}\delta_{i-j})=0 $$ \end{lemma} \begin{prop} \label{flyii} For fixed $z,u$ from proposition \ref{ooo} we have (i) The maps ${}_m^{(z,u)}\delta_i$ satisfy the following identities: $$ {}_m{\delta_i}(ab)= {}_m{\delta_i}(a)\alpha^{i+m}(b)+\alpha^m(a){}_m{\delta_i}(b)+ \sum_{k= 1}^{i-1}{}_m{\delta_{i-k}}(a) {}_{i-k+m}{\delta_{k}}(b) $$ (ii) Suppose $\alpha = id$. Then the maps ${}_m^{(z,u)}\delta_i$ satisfy the following identities: $$ {}_m{\delta_i}(ab)= {}_m{\delta_i}(a)b+a{}_m{\delta_i}(b)+ \sum_{k= 1}^{i-1}{}_m{\delta_{i-k}}(a)\sum_{(j_1,\ldots ,j_l)} C_{i-k+m}^l\delta_{j_1}\ldots {\delta_{j_l}}(b) $$ where the second sum is taken over all the vectors $(j_1,\ldots ,j_l)$ such that $0< l\le min\{i-k+m, k\}$, $j_m\ge 1$, $\sum j_m= k$; $C_j^k= 0$ if $j= 0$, and $C_j^k= C_{j+p^q}^k$ for $q>>0$ if $j\le 0$. \end{prop} {\bf Proof.} For any $a,b\in\bar{D}$ we have $$ \alpha^m(ab)z^{m}+{}_m{\delta_1}(ab)z^{m+1}+{}_m{\delta_2}(ab)z^{m+2}+\ldots= z^{m}(ab)= $$ \begin{equation} \label{(*)} (\alpha^m(a)z^{m}+{}_m{\delta_1}(a)z^{m+1}+{}_m{\delta_2}(a)z^{m+2}+\ldots )b \end{equation} If we represent the right-hand side of (\ref{(*)}) as a series with coeffitients shifted to the left and then compare the corresponding coeffitients on the left-hand side and right-hand side, we get some formulas for ${}_m\delta_i(ab)$. We have to prove that these formulas are the same as in our proposition. Let $$z^{i+m-k}b= \alpha^{i+m-k}(b)z^{i+m-k}+\ldots +x'_kz^{i+m}+\ldots $$ and $$(\alpha^m(a)z^{m}+{}_m{\delta_1}(a)z^{m+1}+{}_m{\delta_2}(a)z^{m+2}+\ldots )b = \alpha^m(ab)z^m+y_{m+1}z^{m+1}+y_{m+2}z^{m+2}+\ldots $$ Then we have $$ y_{i+m}= \alpha^m(a)x'_i+\sum_{k= 0}^{i-1}{}_m\delta_{i-k}(a)x'_k $$ In the proof of \cite{Zh}, prop.2 we have shown that $$ z^{i+1-k}b= \alpha^{i+1-k}(b)z^{i+1-k}+\ldots +\sigma (S_i^k\alpha )(b)z^{i+1}+\ldots $$ Hence $x'_k= \sigma (S_{i+m-1}^k\alpha )(b)$ for $k< i$. It is easy to see that $x'_i= {}_m\delta_i(b)$, $x'_0= \alpha^{i+m}(b)$ and $\sigma (S_{i+m-1}^k\alpha ) = {}_{i+m-k}\delta_k$, which proves (i). For $\alpha= id$, by corollary \ref{formuly}, $$ \sigma (S_{i+m-1}^k\alpha )(b)= \sum_{(j_1,\ldots ,j_l)} C_{i-k+m}^l\delta_{j_1}\ldots {\delta_{j_l}}(b), $$ where $l, j_1, \ldots ,j_l$ were defined in our proposition. This proves (ii).\\ The proposition is proved.\\ $\Box$ \begin{lemma}{(\cite{Zh}, lemma 3 )} \label{ozamene} In the situation of proposition \ref{ooo} suppose ${}^{(z,u)}_i\delta_j$ is the first map such that ${}^{(z,u)}_i\delta_j(a)\ne 0$ for given $a\in \bar{D}$, $i\in {\mbox{\dbl Z}} \backslash \{0\}$, i.e. ${}^{(z,u)}_i\delta_1(a)=\ldots= {}^{(z,u)}_i\delta_{j-1}(a)= 0$, ${}^{(z,u)}_i\delta_j(a)\ne 0$ (so we have a map $i\mapsto j(i)$). Then (i) for $z'= z+u(b)z^{q+1}$, $b\in \bar{D}$ we have ${}^{(z')}\alpha^i (a)={}^{(z)}\alpha^i (a)$, ${}^{(z',u)}_i\delta_{k}(a)={}^{(z,u)}_i\delta_{k}(a)$ for $k<q$ and $$ {}^{(z',u)}_i\delta_{q}(a)= {}^{(z,u)}_i\delta_{q}(a) +b'{}^{(z)}\alpha^{q+i}(a)- {}^{(z)}\alpha^i (a)b', $$ where $b'=\sum_{k=0}^{i-1}{}^{(z)}\alpha^k(b)$. (ii) Suppose ${}^{(z)}\alpha^n|_{Z(\bar{D})}= id$, $n\ge 1$, $a\in Z(\bar{D})$ and \\ ${}^{(z,u)}_1\delta_1({}^{(z)}\alpha^k(a))=\ldots= {}^{(z,u)}_1\delta_{j-1}({}^{(z)}\alpha^k(a))= 0$ for any $k$. Then for $z'= z+u(b)z^{q+1}$, $b\in \bar{D}$ we have ${}^{(z')}\alpha^i (a)={}^{(z)}\alpha^i (a)$, ${}^{(z',u)}_i\delta_{k}(a)= {}^{(z,u)}_i\delta_{k}(a)$ for $k<q+j$ and $$ {}^{(z',u)}_i\delta_{q+j}(a)={}^{(z,u)}_i\delta_{q+j}(a)+ b'{}^{(z)}\alpha^q({}^{(z,u)}_i\delta_{j}(a))-{}^{(z,u)}_i\delta_{j}(a){}^{(z)}\alpha^j(b')+ $$ $$ b'\sum_{k= 1}^{q}{}^{(z)}\alpha^{q-k}({}^{(z,u)}\delta_{j}({}^{(z)}\alpha^{k+i-1}(a))) - {}^{(z,u)}_i\delta_{j}(a)\sum_{k= 0}^{j-1}{}^{(z)}\alpha^{k}(b), $$ where $b'=\sum_{k=0}^{i-1}{}^{(z)}\alpha^k(b)$, if $n|q$ or ${}^{(z)}\alpha (a)=a$. \\ In particular, if ${}^{(z)}\alpha= id$ and $(i,p)=1$, then $$ {}^{(z',u)}_i\delta_{q+j}(a)={}^{(z,u)}_i\delta_{q+j}(a)+ (q-j){}^{(z,u)}_i\delta_{j}(a)b $$ (iii) for $z'= u(b)z$, $b\in Z(\bar{D})$, $b\ne 0$ we have ${}^{(z')}\alpha (a)={}^{(z)}\alpha (a)$, ${}^{(z',u)}\delta_{k}(a)={}^{(z,u)}\delta_{k}(a)$ for $k<j$ and $$ {}^{(z',u)}\delta_{j}(a)={}^{(z,u)}\delta_{j}(a){}^{(z)}\alpha (b^{-1})\cdots {}^{(z)}\alpha^j(b^{-1}) $$ if $i=1$. \end{lemma} {\bf Proof.} (i) We have $$ {z'}^ia{z'}^{-i}= (1+b'z^q+\ldots )z^iaz^{-i}(1+b'z^q+\ldots )^{-1}= (z^iaz^{-i}+b'z^qz^iaz^{-i}+\ldots )(1-b'z^q+\ldots )= $$ $$ (z^iaz^{-i}-z^iaz^{-i}b'z^q+\ldots +b'z^qz^iaz^{-i}-\ldots )= $$ $$ (z^iaz^{-i}-[{}^{(z)}\alpha^i (a)+ {}^{(z,u)}_i\delta_j(a)z^j+\ldots ]b'z^q+b'z^q[{}^{(z)}\alpha^i (a)+{}^{(z,u)}_i\delta_j(a)z^j+\ldots ]+\ldots )= $$ $$ (z^iaz^{-i}-[{}^{(z)}\alpha^i (a)b'+{}^{(z,u)}_i\delta_j(a){}^{(z)}\alpha^j(b')z^j+\ldots ]z^q+ b'{}^{(z)}\alpha^{q+i}(a)z^q+ \ldots )= $$ $$ (z^iaz^{-i}+(-{}^{(z)}\alpha^i (a)b'+b'{}^{(z)}\alpha^{q+i}(a))z^q+ \ldots )= {}^{(z)}\alpha^i (a)+\ldots + {}^{(z,u)}_i\delta_{q-1}(a)z'^{q-1}+ $$ $$ ({}^{(z,u)}_i\delta_{q}(a) + b'{}^{(z)}\alpha^{q+i}(a)-{}^{(z)}\alpha^i (a)b')z'^q+ \ldots $$ (ii) Put $c=z'^iz^{-i}-1-b'z^{q+i}$. So, $w(c)>q+i$. Note that $c{}^{(z)}\alpha^k(a)={}^{(z)}\alpha^k(a)c$, since $n|q$ or ${}^{(z)}\alpha (a)=a$ and $a\in Z(\bar{D})$. We have $$ z'^iaz'^{-i}= (1+b'z^q+c)z^iaz^{-i}(1+b'z^q+c)^{-1}= (z^iaz^{-i}+b'z^qz^iaz^{-i}+cz^iaz^{-i})(1+b'z^q+c)^{-1} = $$ $$ ({}^{(z)}\alpha^i (a)+{}^{(z,u)}_i\delta_{j}(a)z^j+\ldots + {}^{(z,u)}_i\delta_{q+j}(a)z^{q+j}+\ldots + b'z^q({}^{(z)}\alpha^i (a)+{}^{(z,u)}_i\delta_{j}(a)z^j+\ldots ))(1+b'z^q+c)^{-1}= $$ $$ ({}^{(z)}\alpha^i (a)+b'{}^{(z)}\alpha^{q+i}(a)z^q+ {}^{(z)}\alpha^{i}(a)c+ {}^{(z,u)}_i\delta_{j}(a)z^j+\ldots +{}^{(z,u)}_i\delta_{q+j}(a)z^{q+j}+\ldots + $$ $$ b'\sum_{k= 1}^{q}({}^{(z)}\alpha^{q-k}({}^{(z,u)}\delta_{j}({}^{(z)}\alpha^{k+i-1}(a) ) )) z^{q+j}+ b'({}^{(z)}\alpha^{q}({}^{(z,u)}_i\delta_{j}(a))) z^{q+j} +\ldots )(1+b'z^q+c)^{-1}= $$ $$ {}^{(z)}\alpha^i (a)+[{}^{(z,u)}_i\delta_{j}(a)z^j+\ldots + {}^{(z,u)}_i\delta_{q+j}(a)z^{q+j}+\ldots + b'\sum_{k= 1}^{q}({}^{(z)}\alpha^{q-k}({}^{(z,u)}\delta_{j}({}^{(z)}\alpha^{k+i-1}(a) ) )) z^{q+j}+ $$ $$ b'({}^{(z)}\alpha^{q}({}^{(z,u)}_i\delta_{j}(a))) z^{q+j} +\ldots )](1-b'z^q-c+\ldots )= $$ $$ {}^{(z)}\alpha^i (a)+{}^{(z,u)}_i\delta_{j}(a)z^j+\ldots + {}^{(z,u)}_i\delta_{q+j}(a)z^{q+j}+\ldots + b'\sum_{k= 1}^{q}({}^{(z)}\alpha^{q-k}({}^{(z,u)}\delta_{j}({}^{(z)}\alpha^{k+i-1}(a) ) )) z^{q+j}+ $$ $$ b'({}^{(z)}\alpha^{q}({}^{(z,u)}_i\delta_{j}(a)) z^{q+j} +\ldots -{}^{(z,u)}_i\delta_{j}(a){}^{(z)}\alpha^{j}(b')z^{q+j}+\ldots = $$ $$ {}^{(z)}\alpha^i (a)+\ldots +{}^{(z,u)}_i\delta_{q+j-1}(a)z'^{q+j-1}+ ({}^{(z,u)}_i\delta_{q+j}(a)+b'{}^{(z)}\alpha^q({}^{(z,u)}_i\delta_{j}(a))- {}^{(z,u)}_i\delta_{j}(a){}^{(z)}\alpha^j(b') $$ $$ +b'\sum_{k= 1}^{q}{}^{(z)}\alpha^{q-k}({}^{(z,u)}\delta_{j}({}^{(z)}\alpha^{k+i-1}(a))) - {}^{(z,u)}_i\delta_{j}(a)\sum_{k= 0}^{j-1}{}^{(z)}\alpha^{k}(b))z'^{q+j} +\ldots , $$ since $z'^j= z^j+\sum_{k= 0}^{j-1}{}^{(z)}\alpha^{k}(b)z^{q+j}+\ldots $. (iii) We have $$ z'az'^{-1}= bzaz^{-1}b^{-1}= {}^{(z)}\alpha (a)+ b{}^{(z,u)}\delta_{j}(a){}^{(z)}\alpha^j(b^{-1})z^j+ \ldots= $$ $$ {}^{(z)}\alpha (a)+ {}^{(z,u)}\delta_{j}(a){}^{(z)}\alpha (b^{-1})\ldots {}^{(z)}\alpha^j(b^{-1})z'^j+ \ldots , $$ since ${}^{(z')}\alpha |_{Z(\bar{D})}={}^{(z)}\alpha |_{Z(\bar{D})}$.\\ $\Box$ \begin{corol} \label{ozamene3} In the situation of lemma \ref{ozamene} we have $$ j=w(xu(a)x^{-1}-u(a)), $$ where $x\in D$ is any element with $w(x)=i$, if $a\in Z(\bar{D})$, $\alpha (a)=a$ and $(i,p)=1$, where $p=char D$. If $i=1$, we will denote $j$ by $j(u,a)$ or by $i(u,a)$. \end{corol} {\bf Proof.} Since for some parameter $z$ we have $x=b(1+x_1z+\ldots )z^i$, where $b, x_k\in u(\bar{D})$, the proof is easily follows from the proof of (ii) in lemma \ref{ozamene}.\\ $\Box$ In the sequel we will need the following definition. \begin{defi} Let $(\alpha ,\beta )$ be endomorphisms of a division algebra $D$. A map $\delta :$ $D\rightarrow D'$, where $D\subset D'$ are algebras, is called a $(\alpha ,\beta )$-derivation if it is linear and satisfy the following identity $$ \delta (ab)= \delta (a)\alpha (b)+\beta (a)\delta (b) $$ where $a,b\in D$.\\ We will say that $(\alpha ,1)$-derivation is an $\alpha$-derivation. \end{defi} \begin{lemma}{(cf. \cite{Zh}, lemma 4)} \label{lemma2} Let $\delta$ be an $(\alpha ,\beta )$-derivation of an arbitrary division algebra $D$ such that $\alpha ,\beta$ preserve $Z(D)$ and $\alpha|_{Z(D)}\ne \beta|_{Z(D)}$. Then $\delta$ is an inner derivation, i.e. there exists $d\in D$ such that $$ \delta (a)= d\alpha (a)-\beta (a)d $$ for all $a\in D$. \end{lemma} {\bf Proof.} Put $d= \delta (a)(a^{\alpha}-a^{\beta})^{-1}$, where $a\in Z(D)$ is any element such that $\alpha (a)\ne \beta (a)$. Put $\delta_{in}(x)= d\alpha (x)-\beta (x)d$. We claim that $\delta= \delta_{in}$. Indeed, consider the map $\bar{\delta}= \delta -\delta_{in}$. It is an $(\alpha ,\beta )$-derivation. Take arbitrary $b\in D$. Then $\bar{\delta}(ab)= \bar{\delta}(ba)$. But we have $$\bar{\delta}(ab)= \bar{\delta}(a)\alpha (b)+\beta (a)\bar{\delta}(b)= \beta (a)\bar{\delta}(b),$$ and $$\bar{\delta}(ba)= \bar{\delta}(b)\alpha (a)+\beta (b)\bar{\delta}(a)= \alpha (a)\bar{\delta}(b)$$ Therefore, $\bar{\delta}(b)= 0$ for any $b$.\\ $\Box$ \begin{prop}{(cf. \cite{Zh}, lemma 10)} \label{X} Let $D$ be a splittable division algebra. Let $n=Gal (Z(\bar{D})/\overline{Z(D)})$. There exists a parameter $z'$ such that $$ {}^{(z',u)}_m\delta_j=0 $$ if $n\not | j$. \end{prop} {\bf Proof.} Since for $n=1$ there is nothing to prove, we will assume that $n>1$. Let $z$ be some fixed parameter. By \cite{JW}, prop. 1.7 ${}^{(z)}\alpha |_{Z(\bar{D})}$ has order $n$. By proposition \ref{flyii}, ${}^{(z,u)}\delta_1$ is a $({}^{(z)}\alpha^2,{}^{(z)}\alpha )$-derivation. Since $n>1$, ${}^{(z)}\alpha^2|_{Z(\bar{D})}\ne {}^{(z)}\alpha |_{Z(\bar{D})}$. Therefore, by lemma \ref{lemma2}, ${}^{(z,u)}\delta_1$ is an inner derivation and ${}^{(z,u)}\delta_1(a)= d{}^{(z)}\alpha^2(a)- {}^{(z)}\alpha (a)d$, $a\in \bar{D}$. Put $z_1= z-u(d)z^2$. By lemma \ref{ozamene}, (i) we have for any $a\in \bar{D}$ ${}^{(z_1,u)}\delta_1(a)=0$ and ${}^{(z)}\alpha (a)={}^{(z_1)}\alpha (a)$. So, ${}^{(z_1,u)}\delta_1=0$ and ${}^{(z)}\alpha ={}^{(z_1)}\alpha$. By proposition \ref{flyii}, ${}^{(z_1,u)}\delta_2$ is a $({}^{(z_1)}\alpha^3,{}^{(z_1)}\alpha )$-derivation. If $n\ne 2$ then it is inner and we can apply lemma \ref{ozamene}. By induction we get that there exists a parameter $z_{n-1}$ such that ${}^{(z_{n-1},u)}\delta_j=0$ for $j< n$ and ${}^{(z)}\alpha ={}^{(z_{n-1})}\alpha$. It is easy to see that then ${}^{(z_{n-1},u)}_m\delta_j=0$ for $j< n$ and all $m\in{\mbox{\dbl Z}}$. Note that ${}^{(z_{n-1},u)}\delta_n$ is a $({}^{(z_{n-1})}\alpha^{n+1},{}^{(z_{n-1})}\alpha )= ({}^{(z_{n-1})}\alpha ,{}^{(z_{n-1})}\alpha )$-derivation, i.e. ${{}^{(z_{n-1},u)}\delta_{n}}{}^{(z_{n-1})}\alpha^{-1}$ is a derivation. Note that ${{}^{(z_{n-1},u)}\delta_{n+1}}$ is a $({}^{(z_{n-1})}\alpha^2,{}^{(z_{n-1})}\alpha )$-derivation. This follows by proposition \ref{flyii}, since ${}^{(z_{n-1},u)}_m\delta_j=0$ for $j< n$ and all $m\in{\mbox{\dbl Z}}$. So, by lemma \ref{lemma2}, ${}^{(z_{n-1},u)}\delta_{n+1}$ is an inner derivation. Using lemma \ref{ozamene}, (i) with $ z_{n+1}= z_{n-1}+bz_{n-1}^{n+2}$ for an appropriate $b$, we have ${}^{(z_{n+1},u)}\delta_j=0$ for $j< n+2$, $n\not |j$ and ${}^{(z)}\alpha ={}^{(z_{n+1})}\alpha$. Moreover, ${}^{(z_{n+1},u)}_m\delta_j=0$ for $j< n+2$, $n\not |j$ and all $m\in{\mbox{\dbl Z}}$. This easily follows from lemma \ref{triviall}. By induction we can assume that there exists a parameter $z_k$ such that ${}^{(z_{k},u)}_m\delta_j=0$ for $j< k+1$, $n\not |j$ and all $m\in{\mbox{\dbl Z}}$, and ${}^{(z)}\alpha ={}^{(z_{k})}\alpha$. So, by proposition \ref{flyii}, if $n\not | k+1$, then ${{}^{(z_{k},u)}\delta_{k+1}}$ is an inner $({}^{(z_{k})}\alpha^{k+2},{}^{(z_{k})}\alpha )$-derivation. And if $n | k+1$, we can apply the same arguments and conclude that ${}^{(z_{k},u)}\delta_{k+2}$ is a $({}^{(z_{k})}\alpha^{k+2},{}^{(z_{k})}\alpha )$-derivation. Therefore, by lemma \ref{ozamene} there exists a parameter $z_{k+1}= z_k+bz_k^{k+2}$ ($z_k+bz_k^{k+3}$ if $n | k+1$) such that ${}^{(z_{k+1},u)}_m\delta_j=0$ for $j< k+2$, $n\not |j$ and all $m\in{\mbox{\dbl Z}}$, and ${}^{(z)}\alpha ={}^{(z_{k+1})}\alpha$ (or ${}^{(z_{k+1},u)}_m\delta_j=0$ for $j< k+3$, $n\not |j$ and all $m\in{\mbox{\dbl Z}}$, and ${}^{(z)}\alpha ={}^{(z_{k+1})}\alpha$ if $n | k+1$). Since $z_{l+1}= (1+b_lz_l^{k_l})z_l$ for every $l$, the sequence ${\{ z_l\}}_{l= 1}^{\infty}$ converges in $D$, which completes the proof of the proposition.\\ $\Box$ \begin{lemma} \label{(5)} Let $D$ be a splittable division algebra as in proposition \ref{ooo}, of characteristic $p>0$. Let $t\in Z(\bar{D})$ be an element such that $\alpha (t)=t$. Let $j=i(u,t)$ be the minimal positive integer such that ${}^{(z,u)}\delta_j|_{{\mbox{\sdbl F}}_p(t)}\ne 0$ (see corollary \ref{ozamene3}), and we assume $j<\infty$. Then the maps ${}^{(z,u)}_n\delta_m$, $kj\le m<(k+1)j$, $k\in \{1,\ldots ,p-1 \}$ satisfy the following properties: i) there exist elements $c_{n,m,k}\in\bar{D}$ such that $$ {}^{(z,u)}_n\delta_m|_{{\mbox{\sdbl F}}_p(t)}=c_{n,m,1}\delta +\ldots +c_{n,m,k}\delta^k, $$ where $\delta :{\mbox{\dbl F}}_p(t)\rightarrow {\mbox{\dbl F}}_p(t)$ is a derivation such that $\delta (t)= 1$, and $$c_{n,kj,k}= (k!)^{-1}{}^{(z,u)}_n\delta_j(t){}^{(z,u)}_{n+j}\delta_j(t)\ldots {}^{(z,u)}_{n+(k-1)j}\delta_j(t). $$ ii) Let $\zeta =ord ({}^{(z)}\alpha |_{Z(\bar{D})})$. Then $\zeta |j$ and \\ $c_{n,kj,k}\ne 0$ if $(n, j )=1$ and ${}^{(z)}\alpha ({}^{(z,u)}\delta_j(t))\ne {}^{(z,u)}\delta_j(t)$;\\ $c_{n,kj,k}\ne 0$ if ${}^{(z)}\alpha ({}^{(z,u)}\delta_j(t))= {}^{(z,u)}\delta_j(t)$ and $n , (n+j) , \ldots , (n+(k-1)j) \ne 0\mbox{\quad mod\quad}p$. If ${}^{(z)}\alpha =id$, then $c_{n,kj,k}\ne 0$ iff $n , (n+j) , \ldots , (n+(k-1)j) \ne 0\mbox{\quad mod\quad}p$. \end{lemma} {\bf Proof.} i) The proof is by induction on $k$. Let $a,b\in {\mbox{\dbl F}}_p(t)$. For $k= 1$, by proposition \ref{flyii}, (ii) we have $$ {}_n\delta_m(ab)= {}_n\delta_m(a)b+a{}_n\delta_m(b) $$ because all the maps $\delta_q$, $q<j$ are equal to zero on ${\mbox{\dbl F}}_p(t)$. Hence, ${}_n\delta_m$ is a derivation on ${\mbox{\dbl F}}_p(t)$, ${}_n\delta_m|_{{\mbox{\sdbl F}}_p(t)}= c_{n,m,1}\delta$ and $c_{n,j,1}= {}_n\delta_j(t)$. For arbitrary $k$, by proposition \ref{flyii}, (i) and by the induction hypothesis we have $$ {}_n\delta_m(t^q)= q{}_n\delta_m(t)t^{q-1}+{}_n\delta_j(t)(\sum_{l= 0}^{q-2}(c_{n+j,m-j,1}\delta +\ldots +c_{n+j,m-j,k-1}\delta^{k-1})(t^{q-1-l})t^l)+ $$ \begin{equation} \label{(**)} \ldots +{}_n\delta_{m-j}(t)(\sum_{l= 0}^{q-2}(c_{m-j+n,m-s,1}\delta )(t^{q-1-l})t^l). \end{equation} Therefore, ${}_n\delta_m(t^p)= 0$, because $k\le p-1$ and $\sum_{l= 0}^{p-2}\delta^i(t^{p-1-l})t^l= 0$ for $i\le p-2$. Hence, ${}_n\delta_m|_{{\mbox{\sdbl F}}_p(t)}= c_{n,m,1}\delta +\ldots +c_{n,m,p-1}\delta^{p-1}$ and we only have to show that $c_{n,m,q}= 0$ for $q>k$. Using (\ref{(**)}) we can calculate $c_{n,m,j}$. We have $$ c_{n,m,1}= {}_n\delta_m(t); $$ $$ c_{n,m,2}= \frac{1}{2!}({}_n\delta_m(t^2)-2c_{n,m,1}t)= \frac{1}{2} ({}_n\delta_j(t)(c_{n+j,m-j,1}\delta (t))+\ldots +{}_n\delta_s(t)(c_{s+n,m-s,1}\delta (t))) $$ $$ \ldots $$ $$ c_{n,m,q}= \frac{1}{q!}({}_n\delta_j(t)(\sum_{l= 0}^{q-2}c_{n+j, m-j, q-1}\delta^{q-1}(t^{q-1-l})t^l)+ \ldots $$ $$ +{}_n\delta_{m-(q-1)j}(t)(\sum_{l= 0}^{q-2}c_{m+n-(q-1)j, (q-1)j, q-1}\delta^{q-1} (t^{q-1-l})t^l)) $$ \begin{equation} \label{recurrent} = \frac{1}{q}({}_n\delta_j(t)c_{n+j, m-j, q-1}+ \ldots +{}_n\delta_{m-(q-1)j}(t)c_{m+n-(q-1)j, (q-1)j, q-1}) \end{equation} Hence, $c_{n, m, k+1}= \ldots = c_{n, m, p-1}= 0$ and $$c_{n, kj, k}= q^{-1} {}_n\delta_j(t)c_{n+j, kj-j, k-1}= (k!)^{-1}{}^{(z,u)}_n\delta_j(t){}^{(z,u)}_{n+j}\delta_j(t)\ldots {}^{(z,u)}_{n+(k-1)j}\delta_j(t). $$ ii) Let us prove first that $\zeta$ divide $i$. For, if $i$ is not divisible by $\zeta$, we have, by proposition \ref{flyii}, $$ {}^{(z,u)}\delta_j(tx)={}^{(z,u)}\delta_j(t){}^{(z)}\alpha^{j+1}(x)+{}^{(z)}\alpha (t) {}^{(z,u)}\delta_j(x)={}^{(z,u)}\delta_j(xt)= $$ $$ {}^{(z,u)}\delta_j(x){}^{(z)}\alpha^{j+1} (t) +{}^{(z)}\alpha (x){}^{(z,u)}\delta_j(t), $$ where $x\in Z(\bar{D})$, $\alpha (x)\ne x$. But then ${}^{(z)}\alpha^{j+1}(x)={}^{(z)}\alpha (x)$, a contradiction. If ${}^{(z)}\alpha =id$, the same arguments show that ${}^{(z,u)}\delta_j(t)\in Z(\bar{D})$. If $x\in \bar{D}$ is an arbitrary element, this formulae shows ${}^{(z)}\alpha^{j}$ is an inner automorphism $ad({}^{(z,u)}\delta_j(t)^{-1})$. Therefore, ${}^{(z)}\alpha^{j}({}^{(z,u)}\delta_j(t))={}^{(z,u)}\delta_j(t)$. Assume ${}^{(z)}\alpha ({}^{(z,u)}\delta_j(t))\ne {}^{(z,u)}\delta_j(t)$. It's clear then that $$ {}^{(z,u)}_{n+qj}\delta_j(t)=\sum_{l=0}^{n+qj-1}{}^{(z)}\alpha^l ({}^{(z,u)}\delta_j(t))\ne 0 $$ if $(n, j )=1$. So, $c_{n,kj,k}\ne 0$ by (i) in this case. If ${}^{(z)}\alpha ({}^{(z,u)}\delta_j(t))= {}^{(z,u)}\delta_j(t)$, then ${}^{(z,u)}_{n+qj}\delta_j(t)=(n+qj){}^{(z,u)}\delta_j(t)\ne 0$ iff $p$ does not divide $(n+qj)$. So, by (i) $c_{n,kj,k}\ne 0$ in this case iff $n , (n+j) , \ldots , (n+(k-1)j) \ne 0\mbox{\quad mod\quad}p$. The lemma is proved.\\ $\Box$ \begin{lemma} \label{ppp} Let $D$ be a splittable division algebra as in lemma \ref{(5)}. Let $s\in Z(\bar{D})$ be an element such that $\alpha (s)=s$. Let $i=i(u,s)$ be the minimal positive integer such that ${}^{(z,u)}\delta_i(s)\ne 0$ (see corollary \ref{ozamene3}). If $p|i$, then for any positive integral $k$ there exists a map ${}^{(z,u)}\delta_{j(k)}$ such that ${}^{(z,u)}\delta_{j(k)}(s^{p^k})\ne 0$. \end{lemma} {\bf Proof.} We claim that ${}^{(z,u)}\delta_{p^qi}$ is the first map such that ${}^{(z,u)}\delta_{p^qi}|_{{\mbox{\sdbl F}}_p(s^{p^q})}\ne 0$. The proof is by induction on $q$. For $q= 0$, there is nothing to prove. For arbitrary $q$, put $t= s^{p^{q-1}}$. By proposition \ref{flyii} we have $$ \delta_{p^qi}(t^p)= \delta_{p^{q-1}i}(t)\sum_{r= 0}^{p-2}{}_{1+p^{q-1}i}\delta_{p^{q-1}i(p-1)}(t^{p-1-r})t^r+ \sum_{l= p^{q-1}i+1}^{p^qi-1}\delta_l(t)\sum_{r= 0}^{p-2}{}_{1+l}\delta_{p^qi-l}(t^{p-1-r})t^r $$ By induction and lemma \ref{(5)}, ${}_{1+l}\delta_{p^qi-l}|_{{\mbox{\sdbl F}}_p(t)}= c_{1+l,p^qi-l,1}\delta +\ldots +c_{1+l,p^qi-1,p-2}\delta^{p-2}$ for $l>p^{q-1}i$. Therefore, $\sum_{r= 0}^{p-2}{}_{1+l}\delta_{p^qi-l}(t^{p-1-r})t^r= 0$. By lemma \ref{(5)}, (ii), ${}_{1+p^{q-1}i}\delta_{p^{q-1}i(p-1)}|_{{\mbox{\sdbl F}}_p(t)}= c_{1+p^{q-1}i, p^{q-1}i(p-1), 1}\delta +\ldots +c_{1+p^{q-1}i, p^{q-1}i(p-1), p-1}\delta^{p-1}$ with $c_{1+p^{q-1}i, p^{q-1}i(p-1), p-1}\ne 0$. Hence, $\delta_{p^qi}(t^p)= -c_{1+p^{q-1}i, p^{q-1}i(p-1), p-1}\delta_{p^{q-1}i}(t)\ne 0$. The same arguments show that ${}^{(z,u)}\delta_j(t^p)= 0$ for $j<p^qi$. So, ${}^{(z,u)}\delta_{p^qi}$ is the first non-zero map on ${\mbox{\dbl F}}_p(s^{p^q})$. \\ $\Box$\\ \begin{lemma} \label{svva} Let $D$ be a splittable division algebra. Let $z$ be a fixed parameter and ${}^{(z)}\alpha =id$, let $u$ be some fixed embedding $u:\bar{D}\hookrightarrow D$. Let ${}^{(z,u)}\delta_{i}$, $i\in {\mbox{\dbl N}}\cup\infty$ be the first non-zero map on $\bar{D}$. Assume $(i,p)=1$, where $p=char D$. Let ${}^{(z,u)}\delta_j$, $j>i$, $j\in {\mbox{\dbl N}}\cup\infty$ be the first map such that ${}^{(z,u)}\delta_j\ne 0$ if $j$ is not divisible by $i$ and ${}^{(z,u)}\delta_j\ne c_{j/i}{}^{(z,u)}\delta_i^{j/i}$ for some $c_{j/i}\in \bar{D}$ otherwise. Then a) for $k< p=char D$ (arbitrary $k$ if $char D=0$) we have ${}^{(z,u)}\delta_{ki}= c_{k}{}^{(z,u)}\delta_i^{k}$, where \begin{equation} \label{(188)} c_{k}= \frac{(i+1)\ldots (i(k-1)+1)}{k!}, \end{equation} if $ki<j$. b) if condition (\ref{(188)}) is satisfied for any $k$ with $ki<j$, then ${}^{(z,u)}_{-i}\delta_{q}=0$ for $i<q<j$ and ${}^{(z,u)}_{-i}\delta_{j}$ is a derivation. \end{lemma} {\bf Remark.} We will call the number $i(u,z)=\min_{a\in \bar{D}}\{w(zu(a)z^{-1}-u(a))\}$ defined in this lemma {\it a local height}. The number $i=i(z,u)$ in lemma coinside with the level of $D$ defined in \cite{Sa} if $D$ has index $p=char D$ and $D$ is splittable. As it follows from lemmas \ref{ozamene}, \ref{ozamene2} (see below), $i(z,u)$ does not depend on $z,u$ in this case. Corollary \ref{ozamene3} completes then the proof that it coinside with the level defined by Saltman in the case $D$ is splittable. This number will play an important role in this work. It was one of the important parameters in \cite{Zh}. Recall the definition of {\it level}: $h(D)=\min \{w(ab-ba)-w(a)-w(b)\}$. {\bf Proof.} If we compare coefficients in formulae for $\delta_{ki}(ab)$ from proposition \ref{flyii} with coefficients in formulae for $\delta_i^k(ab)$ multiplied by $c_k$, we must have $$ c_kk=((k-1)i+1)c_{k-1}, $$ where from follows a). >From the other hand side, if ${}_{-i}\delta_q$, $q>i$ is the first nonzero map after ${}_{-i}\delta_i$, it must be a derivation by proposition \ref{flyii}, (i). Note that in characterictic zero case this can happens only if $q\ge j$, because a map $c\delta_i^k$ can not be a derivation if $k>1$, which proves b) in this case. Since the maps $\delta_q$ are uniquely defined, by lemma \ref{triviall}, by the maps ${\tilde{\delta}}_l$, $l\le q$, and the maps ${\tilde{\delta}}_q$ are uniquely defined by the maps ${}_{-i}{{\delta}}_l$, $l\le q$, and ${}_{-i}{{\delta}}_q$ are linear combinations of ${{\delta}}_l$, $l\le q$ with integer coefficients, we see that b) holds in arbitrary characteristic.\\ $\Box$ {\bf Remark.} So we see that the maps ${}_i\delta_q$ in this lemma satisfy the same identities as $\delta_{q/i}$. This can be thought of as a possible reduction from level $i$ to level $1$. \begin{defi} \label{lemdef} Let $D$ be a splittable division algebra. Let $u$ be some fixed embedding $u:\bar{D}\hookrightarrow D$. Let $s\in Z(\bar{D})$ be an element such that $\alpha (s)=s$. Let $i=i(u,s)$ be the minimal positive integer such that ${}^{(z,u)}\delta_i(s)\ne 0$ ( corollary \ref{ozamene3} shows that $i$ does not depend on $z$). Assume $(i,p)=1$, where $p=char D$. Define $$ d(u,s)=\max_{z}\{w(z^{-i}u(s)z^i-u(s)-u({}^{(z,u)}_{-i}\delta_i(s))z^i)\} \in {\mbox{\dbl N}}\cup\infty , $$ \end{defi} As we can see from lemma \ref{svva} b), $d(u,s)$ can be interpreted under some conditions as the number $j$ there. So, this definition was motivated by this lemma. \begin{lemma} \label{vtorinv} In the definition above for $p=char D>0$ and ${}^{(z)}\alpha |_{Z(\bar{D})}=id$ we have i) ${d(u,s)}=2i \mbox{\quad mod\quad} p$ if $d(u,s)<\infty$; ii) If ${}^{(z,u)}_{-i}\delta_i(s)\ne 0$, the map ${}^{(z,u)}_{-i}\delta_{d(u,s)+(p-1)i}$ is the first map such that ${}^{(z,u)}_{-i}\delta_{d(u,s)+(p-1)i}(s^p)\ne 0$ for any parameter $z$. In particular, if $d(u,s)=\infty$, $[u(s^p), z^i]=0$. \end{lemma} {\bf Proof.} (ii) Let ${}^{(z,u)}_{-i}\delta_{\kappa}$ be the first map such that ${}^{(z,u)}_{-i}\delta_{\kappa}(s^p)\ne 0$. By corollary \ref{ozamene3} $\kappa$ does not depend on $z$. By the same reason, ${}^{(z,u)}_{-i}\delta_i$ is the first map such that ${}^{(z,u)}_{-i}\delta_i(s)\ne 0$ for any $z$. Put $w:= d(u,s)+(p-1)i$ and fix $u,z$. By proposition \ref{flyii} we have $$ {}_{-i}{{\delta}}_w(s^p)= {}_{-i}{{\delta}}_{d(u,s)}(s)\sum_{q= 0}^{p-2} {}_{d(u,s)-i}{{\delta}}_{(p-1)i}(s^{p-1-q})s^q+ $$ $$ \sum_{k= d(u,s)+1}^{w-1} {}_{-i}{{\delta}}_k(s)\sum_{q= 0}^{p-2} {}_{k-i}{{\delta}}_{w-k}(s^{p-1-q})s^q $$ By lemma \ref{(5)}, ${}_{k-i}{{\delta}}_{w-k}|_{{\mbox{\sdbl F}}_p(s)}= c_{k-i,w-k,1}\delta +\ldots +c_{k-i,w-k,p-2}\delta^{p-2}$ for $w-k<(p-1)i$ and ${}_{d(u,s)-i}{{\delta}}_{(p-1)i}|_{{\mbox{\sdbl F}}_p(s)}= c_{d(u,s)-i,(p-1)i,1}\delta +\ldots +c_{d(u,s)-i,(p-1)i,p-1}\delta^{p-1}$ with $c_{d(u,s)-i,(p-1)i,p-1}\ne 0$ if $d(u,s)-i=i \mbox{\quad mod\quad} p$. Indeed, as we have shown in the proof of lemma \ref{(5)}, (ii), the order $n$ of the automorphism ${}^{(z)}\alpha$ on ${}^{(z,u)}\delta_i(s)$ must divide $i$, so $(n,p)=1$. Now we have two possibilities: $n{\not |}d(u,s)$ and $n|d(u,s)$. In the first case we can repeat the arguments to the first assertion in lemma \ref{(5)}, (ii) to show that $c_{d(u,s)-i,(p-1)i,p-1}\ne 0$. In the second case we have ${}_{d(u,s)-i+qi}\delta_i(s)= (d(u,s)-i+qi)/i{}_i\delta_i(s)\ne 0$ if $d(u,s)-i+qi$ is not divided by $p$. So, by lemma \ref{(5)}, (i) $c_{d(u,s)-i,(p-1)i,p-1}\ne 0$ iff $d(u,s)-i=i \mbox{\quad mod\quad} p$ in this case. Hence, $$ {}_{-i}{{\delta}}_w(s^p)= -{}_{-i}{{\delta}}_{d(u,s)}(s)c_{d(u,s)-i,(p-1)i,p-1}\ne 0 $$ if $d(u,s)-i=i \mbox{\quad mod\quad} p$. This also shows that ${}_{-i}{{\delta}}_w$ is the {\it first} map such that ${}_{-i}{{\delta}}_w|_{{\mbox{\sdbl F}}_p(s^p)}\ne 0$ if $d(u,s)-i=i \mbox{\quad mod\quad} p$. i) By Skolem-Noether theorem there exists a parameter $z'$ in $D$ such that ${}^{(z')}\alpha =id$. Put $$d'(u,z',s)= w(z'^{-j}u(s)z'^{j}-u(s)-u({}^{(z',u)}_{-i}\delta_i(s))z'^i).$$ Since ${}^{(z')}\alpha =id$, the map ${}^{(z',u)}\delta_i$ is the first map such that ${}^{(z',u)}_{-i}\delta_i(s)\ne 0$. If $d'(u,z',s)\ne 2i \mbox{\quad mod\quad} p$, we can find a parameter $z''$ such that $d'(u,z'',s)>d'(u,z',s)$ using lemma \ref{ozamene}, (ii). Continuing this procedure, we find a parameter $z$ such that $d'(u,z,s)=2i \mbox{\quad mod\quad} p$ or $d'(u,z,s)=\infty$. Using arguments from ii) we get that the map ${}^{(z,u)}_{-i}\delta_{d'(u,z,s)+(p-1)i}$ is the first map such that ${}^{(z,u)}_{-i}\delta_{d'(u,z,s)+(p-1)i}(s^p)\ne 0$ for the parameter $z$. As it was noted in the beginning of the proof, the number $\kappa =d'(u,z,s)+(p-1)i$ does not depend on the parameter. Since $d'(u,z,s)\le d(u,s)$, we get $d'(u,z,s)= d(u,s)$. For, otherwise we can repeat the arguments from (ii) and conclude that ${}^{(z,u)}_{-i}\delta_{d(u,s)+(p-1)i}(s^p)= 0$, a contradiction. The lemma is proved.\\ $\Box$ It would be interesting to know more about a behaviour of ${}^{(z,u)}_m\delta_j$ with respect to the embedding $u$. We will give an answer in one special case, namely, when $\bar{D}=Z(\bar{D})$ and $Z(\bar{D})/\overline{Z(D)}$ is a simple extension. \begin{lemma} \label{simple} Let $D$ be a division algebra such that $char D=p>0$, $\bar{D}=Z(\bar{D})$, $Z(\bar{D})$ is not perfect and $Z(\bar{D})/\overline{Z(D)}$ is a simple extension (so, $D$ is splittable). Let $\bar{u}$ be a primitive element of the extension $Z(\bar{D})/\overline{Z(D)}$ such that $\bar{u}\notin (Z(\bar{D}))^p$ and let $u$ be any lift of $\bar{u}$ in $D$. Then there exists an embedding $u:\bar{D}\hookrightarrow D$ such that $u(\bar{u})=u$ and any map ${}^{(z,u)}_m\delta_j$ is uniqely defined by the values ${}^{(z,u)}_m\delta_j(u^q)$ or, equivalently, by the values ${}^{(z,u)}_l\delta_k(u)$, $k\le j$. In particular, if ${}^{(z,u)}_m\delta_k(u)=0$ for $k\le j$, then ${}^{(z,u)}_m\delta_j=0$. \end{lemma} {\bf Proof.} Consider a field $Z(D)(u)$. It is a complete discrete valued field as a finite extension of $Z(D)$. By classical Cohen theorem, there exists an embedding $\overline{Z(D)(u)}=\bar{D}\hookrightarrow Z(D)(u)\subset D$. By \cite{Co}, lemmas 11,12 the embedding is completely defined by a $p$-basis $\Gamma$ of the field $\overline{Z(D)(u)}$. Namely, for any lift $G$ of a given $p$-basis $\Gamma$ there exists an embedding $s$ such that $G\subset s(\overline{Z(D)(u)})$. Let's show that there exists a $p$-basis $\Gamma$ of the field $\bar{D}$ such that $\bar{u}\in \Gamma$ and $\Gamma\ni \gamma\in \overline{Z(D)}$ if $\gamma\ne \bar{u}$. Consider a set of all non-void sets $\Gamma'$ of elements $\gamma_{\tau}\in \bar{D}$ satisfying the following property:\\ A) $\bar{u}\in\Gamma'$, $\Gamma' \ni \gamma\in \overline{Z(D)}$ if $\gamma\ne \bar{u}$ and $[{\bar{D}}^p(\gamma_1,\ldots ,\gamma_r):\bar{D}^p]=p^r$ for any $r$ distinct elements of $\Gamma'$. This set is not void, since it contains the set $\Gamma'=\{\bar{u}\}$. By Zorn's lemma, there exists a maximal set $\Gamma$ satisfying A). Then $\bar{D}=\bar{D}^p(\Gamma )$. Indeed, since $\overline{Z(D)}^p(\bar{u})\subset \bar{D}^p(\Gamma )$, it suffice to show that any element from $\overline{Z(D)}$ lies in $\bar{D}^p(\Gamma )$. Suppose $a\in \overline{Z(D)}$, $a\notin \bar{D}^p(\Gamma )$. Then the set $\Gamma'=\{a\cup \Gamma\}$ satisfy A), a contradiction with maximality of $\Gamma$. Now, we can take a lift of $\Gamma$ in the following way. We take $u$ as a lift of $\bar{u}$, and we take lifts of all other elements in $Z(D)$. This lift defines an embedding $u:\bar{D}\hookrightarrow D$. Let us show that any map ${}^{(z,u)}_m\delta_j$ (for some fixed $z$) is uniqely defined by the values ${}^{(z,u)}_l\delta_k(u)$, $k\le j$. We have $u(\bar{D})=u(\overline{Z(D)})(u)$ and any element $a\in u(\bar{D})$ can be represented as a polynomial in finite number of elements from $\Gamma$ with coefficients from $u(\bar{D})^{p^k}$ for any $k>0$. Note that for any $j$ there exists $k>0$ such that for any $b\in \overline{Z(D)}^{p^k}$ ${}^{(z,u)}_l\delta_q(b)=0$ for all $q\le j$ and all $l$. Indeed, assume ${}^{(z,u)}_1\delta_q(b)\ne 0$ for some $q\le j$, $b\in \overline{Z(D)}^{p^k}$ and ${}^{(z,u)}_l\delta_s(c)= 0$ for all $l$, all $c\in \overline{Z(D)}^{p^k}$ and all $s<q$. Then, since ${}^{(z)}\alpha |_{\overline{Z(D)}}=id$ and by proposition \ref{flyii}, ${}^{(z,u)}_l\delta_s(b^p)= 0$ for all $b\in \overline{Z(D)}^{p^k}$, all $l$ and all $s\le q$. Now, since $u(\bar{D})^{p^k}=u(\overline{Z(D)})^{p^k}(u^{p^k})$, any element $a\in u(\bar{D})$ can be represented as a polynomial in finite number of elements from $\Gamma$ with coefficients from $u(\overline{Z(D)})^{p^k}$. Since all elements except $u$ in $\Gamma$ belong to the center $Z(D)$, the value of ${}^{(z,u)}_m\delta_j(a)$ is uniqely determined by the values ${}^{(z,u)}_m\delta_j(u^l)$ that are uniqely defined, by proposition \ref{flyii}, by the values ${}^{(z,u)}_l\delta_k(u)$, $k\le j$. \\ $\Box$ {\bf Remark} In the case $Z(\bar{D})$ perfect field there is only one embedding $u$, which is compatible with the embedding $\overline{Z(D)}\hookrightarrow Z(D)$. So, the assertion of lemma is easy in this case. \begin{lemma}{(cf. \cite{Zh}, lemma 8)} \label{ozamene2} In the situation of lemma \ref{simple} suppose ${}^{(z,u)}_m\delta_1=\ldots = {}^{(z,u)}_m\delta_{j-1}= 0$, ${}^{(z,u)}_m\delta_j\ne 0$. Let $n$ be the order of ${}^{(z)}\alpha$. Then (i) for $u'= u+bz^q$, $b\in u(\bar{D})$, $n|q$ we have ${}^{(z,u')}_m\delta_l={}^{(z,u)}_m\delta_l$, $l<q$ and $${}^{(z,u')}_m\delta_q (\bar{u})= {}^{(z,u)}_m\delta_q (\bar{u})+{}^{(z)}\alpha^m (\bar{b}) -\frac{\partial}{\partial \bar{u}}({}^{(z)}\alpha^m (\bar{u}))\bar{b},$$ where the derivative is taken in the field $\bar{D}=\bar{D}^p(\Gamma )$. (ii) Suppose ${}^{(z)}\alpha = id$. Then for $u'= u+bz^q$, $b\in u(\bar{D})$ we have ${}^{(z,u')}_m\delta_l={}^{(z,u)}_m\delta_l$, $l<q+j$ and $${}^{(z,u')}_m\delta_{q+j}(\bar{u})= {}^{(z,u)}_m\delta_{q+j}(\bar{u})+ {}^{(z,u)}_m\delta_{j}(\bar{b}) -\frac{\partial }{\partial \bar{u}}({}^{(z,u)}_m\delta_j (\bar{u}))\bar{b}, $$ where the derivative is taken in the field $\bar{D}=\bar{D}^p(\Gamma )$. (iii) Suppose ${}^{(z)}\alpha= id$. Let $\bar{u'}\in \bar{D}$ be any primitive element of the extension $\bar{D}/\overline{Z(D)}$ satisfying the conditions of lemma \ref{simple}, and let $u'\in D$ be any lift of $\bar{u'}$. Then we have ${}^{(z,u')}_m\delta_l={}^{(z,u)}_m\delta_l$, $l<j$ and $$ {}^{(z,u')}_m\delta_{j}(\bar{u'})= {}^{(z,u)}_m\delta_{j}(\bar{u}) \frac{\partial }{\partial \bar{u}}(\bar{u'}), $$ where the derivative is taken in the field $\bar{D}=\bar{D}^p(\Gamma )$. \end{lemma} {\bf Proof.} First of all, let's note that there exists $k\in {\mbox{\dbl N}}$ such that for any $a\in \overline{Z(D)}^{p^k}$ holds $u(a)-u'(a)=0 \mbox{\quad mod\quad} M_D^{q+1}$, where $u'$ is any another embedding, $q\in {\mbox{\dbl N}}$ is any given number. Indeed, assume for any $c\in \overline{Z(D)}^{p^s}$ holds $u(c)-u'(c)=0 \mbox{\quad mod\quad} M_D^{l}$, i.e. $u(c)=u'(c)+c_lz^l+\ldots$, where $c_l\in u'(\bar{D})$. Then $u(c^p)=(u(c))^p= (u'(c))^p+pu'(c)^{p-1}c_lz^l+\ldots$, so $u(c^p)-u'(c^p)=0 \mbox{\quad mod\quad} M_D^{l+1}$. >From this immediately follows that $u(a)-u'(a)=0 \mbox{\quad mod\quad} M_D^{q}$ for any $a\in \bar{D}$ if $u'$ is defined by the element $u'=u+bz^q$, because $u(\bar{u})-u'(\bar{u})=bz^q$. Moreover, if we represent $a$ as some polynomial $P(\gamma_1,\ldots ,\gamma_r, \bar{u})$ with coefficients from $\overline{Z(D)}^{p^k}$, then it is clear that $$[u(a)-u'(a)]z^{-q}=-\frac{\partial}{\partial \bar{u}}(P(\gamma_1,\ldots ,\gamma_r, \bar{u})) \bar{b} \mbox{\quad mod\quad} M_D $$ if $n|q$, since $u(\gamma_l)=u'(\gamma_l)$ for any $l$ and $z^quz^{-q}=u \mbox{\quad mod\quad} M_D$. It is also clear that the derivative can be taken even in the field $\bar{D}^p(\Gamma )$. So, we have \\ (i) $$ z^mu'z^{-m}= z^m(u+bz^q)z^{-m}= u({}^{(z)}\alpha^m (\bar{u}))+u({}^{(z,u)}_m\delta_j(\bar{u}))z^j +\ldots + (u({}^{(z)}\alpha^m(\bar{b})) $$ $$ +u({}^{(z,u)}_m\delta_j(\bar{b}))z^j+ \ldots )z^q= u({}^{(z)}\alpha^m(\bar{u}))+\ldots +(u({}^{(z,u)}\delta_q (\bar{u}))+ u({}^{(z)}\alpha^m(\bar{b})))z^q+\ldots = $$ $$ u'({}^{(z)}\alpha^m (\bar{u}))+\ldots +(u'({}^{(z,u)}_m\delta_q (\bar{u}))+u'({}^{(z)}\alpha^m(\bar{b}))- u'(\frac{\partial}{\partial \bar{u}}{}^{(z)}\alpha^m (\bar{u})\bar{b}))z^q+\ldots , $$ (ii) We have $$ z^mu'z^{-m}= z^m(u+bz^q)z^{-m}= u(\bar{u})+u({}^{(z,u)}_m\delta_j (\bar{u}))z^j+\ldots +(u(\bar{b})+ u({}^{(z,u)}_m\delta_j(\bar{b}))z^j+ \ldots )z^q= $$ $$ u(\bar{u})+u({}^{(z,u)}_m\delta_j (\bar{u}))z^j+\ldots +(u({}^{(z,u)}_m\delta_q (\bar{u}))+u(\bar{b}))z^q+u({}^{(z,u)}_m\delta_{q+1}(\bar{u})) z^{q+1}+\ldots $$ $$ +u({}^{(z,u)}_m\delta_{q+j-1}(\bar{u}))z^{q+j-1}+(u({}^{(z,u)}_m\delta_{q+j}(\bar{u}))+ u({}^{(z,u)}_m\delta_j (\bar{b})))z^{q+j}+\ldots= $$ $$ u'(\bar{u})+u'({}^{(z,u)}_m\delta_j (\bar{u}))z^j+\ldots +u'({}^{(z,u)}_m\delta_{q+j-1}(\bar{u}))z^{q+j-1}+ (u'({}^{(z,u)}_m\delta_{q+j}(\bar{u}))+u'({}^{(z,u)}_m\delta_j (\bar{b}))- $$ $$ u'(\frac{\partial}{\partial \bar{u}}({}^{(z,u)}_m\delta_j (\bar{u}))\bar{b}))z^{q+j}+\ldots $$ (iii) Assume $u'=u(\bar{u'})+a_1z+\ldots$, where $a_i\in u(\bar{D})$. Since, by proposition \ref{flyii}, the map ${}^{(z,u)}_m\delta_j$ is a derivation, we have $$ z^mu'z^{-m}= [u(\bar{u'})+u({}^{(z,u)}_m\delta_j(\bar{u'}))z^j+\ldots ] + [a_1+u({}^{(z,u)}_m\delta_j(a_1)z^j+\ldots ]z+\ldots = $$ $$ u'+u({}^{(z,u)}_m\delta_j(\bar{u'}))z^j+\ldots =u'+u({}^{(z,u)}_m\delta_j(\bar{u}) \frac{\partial }{\partial \bar{u}}(\bar{u'}))z^j+\ldots = u'+u'({}^{(z,u)}_m\delta_j(\bar{u}) \frac{\partial }{\partial \bar{u}}(\bar{u'}))z^j+\ldots $$ $\Box$ \section{The period-index problem} In this section we will prove the following theorem. \begin{th} \label{gipoteza} The following conjecture: the exponent of $A$ is equal to its index for any division algebra $A$ over a $C_2$-field $F$ has the positive answer for $F= F_1((t))$, where $F_1$ is a $C_1$-field. \end{th} Recall that a field $F$ is called a {\it $C_i$-field } if any homogeneous form $f(x_1, \ldots ,x_n)$ of degree $d$ in $n>d^i$ variables with coefficients in $F$ has a non-trivial zero. Some basic properties of $C_i$-fields see, for example, in \cite{PY}. This conjecture was proposed by M. Artin and was solved for some another examples of the field $F$ by many authors. As it is known for me, the positive answer for all division algebras of index $ind A= 2^a3^b$ was given in \cite{PY}, for division algebras over the field $F=k((X))((Y))$, where $k$ is a perfect field of characteristic $p\ne 0$ such that $\dim_{{\mbox{\sdbl F}}_p}k/\wp (k)=1$, was given by Tignol in the Appendix in \cite{AJ} (we include this case though $F$ may not be a $C_2$-field), for division algebras of index prime to the characterictic of $F$, where $F$ is a function field of a surface, was given in \cite{DJ}. I propose, the positive answer was also known for division algebras over $F=F_1((t))$ of characteristic 0. We will give the prove of the theorem above in any characteristic. {\bf Proof.} 1) Recall that any extension of a $C_1$-field is simple. Indeed, suppose $E= \bar{F}(u_1, \ldots , u_r)$. Consider the field $K= \bar{F}(u_1^p, \ldots , u_r^p)$. By Tsen's theorem, $K$ and $E$ are $C_1$-fields. So, the form $x_1^p+x_2^pu_1+\ldots + x_p^pu_1^{p-1}+x_{p+1}^pu_2$ has a non-trivial zero in $E$. But $x_i^p\in K$ and elements $1, u_1,\ldots , u_1^{p-1}, u_2$ are linearly independent over $K$, a contradiction. 2) Assume the theorem is known in the prime exponent case. We deduce the theorem by ascending induction on $e=exp A$. If $e$ is not a prime number, then write $e=lm$. By assumption $A^{\otimes m}$ can be split by a field extension $F\subset F'$ of degree $l$. This implies that $A_{F'}$ has exponent dividing $m$. Note that $F'$ is also a Laurent series field. By the induction hypothesis applied to the pair $(F', A_{F'})$, there exists a field extension $F'\subset L$ of degree dividing $m$ splitting $A_{F'}$. Therefore $A$ is split by the extension $F\subset L$ of degree dividing $lm$ and we conclude the theorem. 3) So, let $exp A= l$ be a prime number. By the basic properties of the exponent and the index (see, e.g. \cite{PY}) we have then $ind A= l^k$ for some natural $k$. Suppose $(l,p=char F)=1$. It is known that the conjecture is true for all division algebras of index $ind A= 2^a3^b$, so we can assume $l\ne 2,3$. We can assume $F$ contains the group $\mu_l$ of $l$-roots of unity, because $[F(\mu_l):F]<l$ and we can reduce the problem to the algebra $A\otimes_F F(\mu_l)$. Then by the Merkuriev-Suslin theorem $A$ is similar to the tensor product of symbol-algebras of index $l$. To conclude the statement of the corollary it is sufficient to prove that every two symbol algebras $A_1, A_2$ contain $F$-isomorphic maximal subfields. Since every division algebra over a $C_1$-field is trivial and every field extension is simple, every symbol-algebra of index $l$ over $F$ is splittable. Since $(l,p)=1$, it is good splittable and its residue field is a cyclic Galois extension of $\bar F$. So, if $z_i$ is a parameter from proposition \ref{X} for algebra $A_i$, then $z_i$ acts on $\bar{A_i}$ as a Galois automorphism and $z_i^l\in F$. We have $v(z_i^l)=1$ ($v$ is the valuation on $F$). Let us show that $A_1$ contains a $l$-root of any element $u$ in $F$ with $v(u)\ne 0$. So, $A_1$ will contain a subfield isomorphic to $F(z_2)$. Since for any element $1+b$, $v(b)>0$ there exists a $l$-root $(1+b)^{1/l}\in F$, it is sufficient to prove that $A_1$ contains any $l$-root of elements $ct$, $c\in u(\bar{F})$, where $u$ is some fixed embedding $u:\bar{A_1}\hookrightarrow A_1$. Assume $z_1^l=c_1t$, $c_1\in u(\bar{F})$. Note that for any element $b\in u(\bar{A_1})$ we have $(bz_1)^l=u(N_{\bar{A_1}/\bar{F}}(b))z_1^l$. But the norm map $N_{\bar{A_1}/\bar{F}}$ is surjective, since $\bar F$ is a $C_1$-field (see, e.g. \cite{PY}, 3.4.2), so there exists $b$ such that $(bz)^l=ct$. 4) Suppose now $exp A=p$. Then $ind A=p^k$. By Albert's theorem (in \cite{Al}) there exists a field $F'=F(u_1^{1/p},\ldots ,u_k^{1/p})$ which splits $A$. Using the same arguments as in 1) one can show that every such a field has maximum two generators, say $F'=F(u_1^{1/p}, u_2^{1/p})$. Therefore, $ind A\le p^2$. If $ind A=p$, there is nothing to prove, so we assume $ind A=p^2$ and $F'$ is a maximal subfield in $A$. 5) Suppose $F_1$ is a perfect field. By Albert's theorem, $A\cong A_1\otimes_F A_2$, where $A_1,A_2$ are cyclic algebras of degree $p$, $A_1=(L_1/F,\sigma_1, u_1)$, $A_2=(L_2/F, \sigma_2, u_2)$. Since $F_1$ is perfect, $\bar{A_1}/\bar{F}$, $\bar{A_2}/\bar{F}$ are Galois extensions. So, $A_1, A_2$ are good splittable. Let us show that $A_1, A_2$ have common splitting field of degree $p$ over $F$. This leads to a contradiction. By proposition \ref{X} there exist parameters $z_1\in A_1$, $z_2\in A_2$ such that they act on $\bar{A_1}$, $\bar{A_2}$ as Galois automorphisms. Note that then $z_1^p, z_2^p\in F$. Let us show that $F(z_1)$ splits $A_2$. Consider the centralizer $D=C_A(F(z_1))$. Consider the element $t_1=z_2z_1^{-1}$. We have $t_1^p\in F$, $w(t_1)=0$, where $w$ denote the unique extension of the valuation $v$ on $F$. Since $\bar{D}/\overline{Z(D)}$ is a Galois extension, there exists an element $b_1\in F$ such that $w(t_1-b_1)>0$. Since $(t_1-b_1)^p\in F$, there exists natural $k_1$ such that $w((t_1-b_1)z_1^{-k_1})=0$. Denote $t_2=(t_1-b_1)z_1^{-k_1}$. We have again $t_2^p\in F$. Repeating this arguments and using the completeness of $D\subset A$ we get \\ $z_2=t_1z_1=(t_2z_1^{k_1}+b_1)z_1=\ldots =b_1z_1+b_2z_1^{k_1+1}+\ldots$,\\ so, $z_2\in F(z_1)=Z(D)$. 6) Suppose $F_1$ is not perfect. Since $F'$ is generated by two elements over $F$, it contains all $p$-roots of $F$. Then, every two elements $u,z\in F$ such that $z^{1/p}\notin F(u^{1/p})$, where $z^{1/p}, u^{1/p}\in F'$, also generate $F'$ over $F$. This follows from the same arguments as in 1), 4). Now take $u\in F_1\backslash F_1^p$, $z=u+t$. It's clear that $p$-roots of these elements generate $F'$ over $F$. Moreover, the fields $F(u^{1/p}), F(z^{1/p})$ are {\it "unramified"} over $F$, i.e. $[\overline{F(u^{1/p})}:\bar{F}]=p=[F(u^{1/p}):F]$, $[\overline{F(z^{1/p})}:\bar{F}]=p$. Denote $u_1=u^{1/p}$, $u_2=z^{1/p}$ in $F'$. Then by Albert's theorem, $A\cong A_1\otimes_F A_2$, where $A_1,A_2$ are cyclic algebras of degree $p$, $A_1=(L_1/F,\sigma_1, u)$, $A_2=(L_2/F, \sigma_2, z)$. Concider the centralizer $D=C_A(F(u_1))$. Suppose $\bar{D}/\overline{Z(D)}$ is a separable extension. Then there exist a lift $u:\bar{D}\hookrightarrow D$ of arbitrary embedding $u':\overline{F(u_1)}\hookrightarrow F(u_1)$. Consider the embedding $u'=u_1$ defined in lemma \ref{simple}. Since $F(u_1)/F$ is a purely inseparable extension, $u'$ is a good embedding, so $u$ is a good embedding of $\bar{D}=\bar{A}$ in $D\subset A$. So, we get $A$ is a good splittable algebra, and $u(\bar{A})$ contain a purely inseparable over $F$ element. But this is a contradiction with lemma \ref{ppp}. So, $\bar{A}/\bar{F}$ can not contain a separable subextension, because in this case $\bar{D}/\overline{Z(D)}$ must be a separable extension. Now we can use, for shorteness, lemmas A.4., A.6. of Tignol in Appendix to the paper \cite{AJ}. These lemmas show that a tensor product $A_1\otimes A_2$ of any two symbols $A_1, A_2$ is similar either to a single symbol in $Br(F)$ (in which case we are done) or to a product of two symbols of level zero. Recall that, by Saltman's results in \cite{Sa}, every division algebra of level zero is tame, which means in our case that the residue division algebra is a separable extension over $\bar{F}$. A notion of level was already discussed above in remark to lemma \ref{svva}. So, assume $A\sim D_1\otimes D_2$, where $D_1, D_2$ are tame division algebras of degree $p$ over $F$. We can assume $A$ and $D_1\otimes D_2$ are division algebras, so $A\cong D_1\otimes D_2$. Since $D_1, D_2$ are tame, we conclude $\bar{A}$ must contain a separable element, a contradiction. The theorem is proved. \\ $\Box$ \section{Good splittable algebras} In this section we prove a decomposition theorem for good splittable division algebras. This theorem shows how the studying of good splittable division algebras can be reduced to the studying of division algebras with simple described structure. So, good splittable algebras are the most easy and good algebras to study. \begin{lemma} \label{goodspl} Let $D$ be a good splittable division algebra, $F=Z(D)$, and let $Z(\bar{D})=\bar{F}(s)$ be a purely inseparable over $\bar{F}$ field of degree $p=char D>0$. Let $u:\bar{D}\hookrightarrow D$ be a good embedding. Then there exists a parameter $z$ such that ${}^{(z,u)}_{-i}\delta_j=0$ for $j>i$, where $i=i(z,u)$ is a local height, and $u({}^{(z,u)}\delta_i(s))=x$, where $x\in Z(D)$. Moreover, $(i,p)=1$. \end{lemma} {\bf Proof.} Since $Z(\bar{D})/\bar{F}$ is a purely inseparable extension, ${}^{(z)}\alpha |_{Z(\bar{D})}=id$ for any parameter $z$. By Skolem-Noether theorem there exists a parameter $z$ in $D$ such that ${}^{(z)}\alpha =id$. Suppose ${}^{(z,u)}\delta_i(s)=0$, where $i=i(z,u)$. Then ${}^{(z,u)}\delta_i|_{Z(\bar{D})}=0$, since $u$ is a good embedding and $Z(\bar{D})/\bar{F}$ is a simple extension. So, ${}^{(z,u)}\delta_i$ is an inner derivation by Scolem-Noether theorem, and by lemma \ref{ozamene}, (i) there exists a parameter $z'$ such that ${}^{(z',u)}\delta_i=0$, ${}^{(z')}\alpha =id$. So, we can assume ${}^{(z,u)}\delta_i(s)\ne 0$ for some parameter $z$. Since $s^p\in Z(D)$, by lemma \ref{ppp} we have $(i,p)=1$. Since ${}^{(z,u)}\delta_i$ is a derivation, ${}^{(z,u)}\delta_i(s)\in Z(\bar{D})$ (see the arguments in lemma \ref{(5)}, (ii)). Since $(i,p)=1$, there exists $k$ such that $p| (1-ki)$. So, by lemma \ref{ozamene}, (iii), for the parameter $z'=({}^{(z,u)}\delta_i(s))^k$ we have ${}^{(z')}\alpha =id$, ${}^{(z',u)}\delta_i(s)\in \bar{F}$, i.e. $u({}^{(z',u)}\delta_i(s))\in Z(D)$. Since $s^p\in Z(D)$, by lemma \ref{vtorinv} we must have $d(u,s)=\infty$. In the proof of lemma \ref{vtorinv}, (i) was shown that $d(u,s)=d'(u,z,s)$ for some parameter $z$, and the construction of this element uses lemma \ref{ozamene}, (ii), so it preserves the initial values of ${}^{(z')}\alpha$, ${}^{(z',u)}\delta_i$. So, ${}^{(z,u)}_{-i}\delta_j=0$ for $j>i$ and the lemma is proved.\\ $\Box$ \begin{prop} \label{razlozhenie} Let $D$ be a splittable division algebra. Then we have $D\cong D_1\otimes_F D_2$, where $D_1, D_2$ are splittable division algebras such that $D_1$ is an inertially split algebra. If $D$ is a good splittable division algebra, then $Z(\bar{D_2})/\bar{F}$ is a purely inseparable extension and $D_2$ is a good splittable algebra ($D_1$ or $D_2$ may be trivial). So, $D\sim A\otimes_FB\otimes_FD_2$, where $A$ is a cyclic division algebra and $B$ is an unramified division algebra. \end{prop} {\bf Proof.} If $char D=0$, the proposition is obvious, so we assume $char D>0$. By \cite{P}, p.261, $D\cong D_1\otimes_F\ldots \otimes_FD_k$, where $[D:F]= p_1^{r_1}\ldots p_k^{r_k}$ and $[D_i:F]= p_i^{r_i}$. Let $p_2= p$. Since $D_i$ are defectless over $F$, $D_1,D_3,\ldots D_k$ are inertially split. Therefore, by theorem \ref{Cohen} the algebra $B=D_1\otimes D_3\otimes \ldots \otimes D_k$ is good splittable. Assume first that $D$ is good splittable. By proposition 1.7. in \cite{JW}, if $s\in Z(\bar{D})$ is an element such that $\alpha (s)=s$, then this element is a purely inseparable element over $\bar{F}$. So, if $D$ is a good splittable division algebra, then by lemma \ref{ppp} $D_2$ is either inertially split or $Z(\bar{D_2})/\bar{F}$ is a purely inseparable extension. For, otherwise there exists an element $s\in Z(\bar{D_2})\subset Z(\bar{D})$ as above and by proposition \ref{X} $p|i(u,s)$ for any embedding $u$. If $u$ is a good embedding, then $s^{p^k}\in Z(D)$ for some $k$, a contradiction. So, we assume below $Z(\bar{D_2})/\bar{F}$ is a purely inseparable extension. Now, we have (see, e.g. th.1 in \cite{Mor}) $\bar{D}\cong \bar{D_2}\otimes_{\bar{F}}\bar{B}$ and so $u(\bar{D})\cong u(\bar{D_2})\otimes_{u(\bar{F})}u(\bar{B})$, where $u$ is a good embedding. So, $E=u(Z(\bar{D_2}))$ is a purely inseparable field over $u(\bar{F})\subset Z(D)$. Consider the field $E'=u(K)\otimes_{u(\bar{F})}F$, where $K$ is a maximal separable subfield in $\bar{B}$. This is an inertial lift of $K$ in $D$. Consider the centralizer $C=C_D(E')\cong D_2\otimes_FE'$. Let $M$ be a maximal subfield in $\bar{D_2}$. Note that $u(\bar{D_2})\subset C$, so $L\subset C$, where $L=u(M)F$ is the composit of $u(M)$ and $F$, and $E\subset L$. Note that $[L:F]=ind D_2=ind C$. The field $L$ splits $C$ by dimension arguments. So, it must split $D_2$, since $([E':F],p)=1$, and $D_2$ is a $p$-algebra. Therefore, $L$ is isomorphic to a maximal subfield in $D_2$, so $D_2$ contain a copy of purely inseparable "unramified" subfield, whose residue field is isomorphic to $Z(\bar{D_2})$. Therefore, $D_2$ is a god splittable algebra. For, the centralizer of this field is an unramified division algebra, so by theorem \ref{Cohen} is splittable. So, $D_2$ is good splittable if the purely inseparable field is good splittable. But it is good splittable since it contains a subfield isomorphic to $u(Z(\bar{D_2}))$ by the construction. (Another way to see it is to use arguments from lemma \ref{simple} to show that there exists an appropriate $p$-basis). Let $D$ be a splittable algebra. Then the same arguments as in the previous paragraph show that $L$ is isomorphic to a maximal subfield in $D_2$ (it is not important that $Z(\bar{D_2})/\bar{F}$ may be not a purely inseparable extension). Now, the composit $EF\subset L$, $EF\ne L$, since every element from $E$ commute with $u(\bar{D_2})$, where $u$ is some fixed embedding. So we must have $\overline{C_{D_2}(EF)=\bar{D_2}}$ and $C_{D_2}(EF)$ is an unramified division algebra. Therefore, $D_2$ is splittable division algebra. Decomposition theorems \cite{JW}, Thm. 5.6-5.15 complete the proof.\\ $\Box$ This proposition shows that the study of splittable division algebras can be reduced to the study of splittable $p$-algebras. So, below in this section and in the next section we will deal with $p$-algebras only. \begin{prop} \label{555} Let $D$ be a good splittable division algebra such that $Z(\bar{D})/\overline{Z(D)}$ is a purely inseparable extension. Then $D\cong D_1\otimes_{Z(D)}D_2$, where $D_1$ is an unramified division algebra and $D_2$ is a good splittable division algebra such that $\bar{D_2}$ is a field, $\bar{D_2}/\overline{Z(D)}$ is a purely inseparable extension, $[\bar{D_2}:\overline{Z(D)}]= [\Gamma_{D_2}:\Gamma_{Z(D)}]$. \end{prop} {\bf Proof.} The proof is by induction on the degree $[Z(\bar{D}):\overline{Z(D)}]$. Assume $[Z(\bar{D}):\overline{Z(D)}]=p$. Let ${}^{(z,u)}\delta_i$ be the map from lemma \ref{goodspl}. Then ${}^{(z,u)}\delta_i^p$ is a derivation trivial on the centre $Z(D)$, hence by Scolem-Noether theorem it is an inner derivation. We claim that $z^p\in Z(D)$. We have $$ z^{-i}az^i= a+{}_{-i}\delta_i(a)z^i, \mbox{\quad} a\in u(\bar{D}) $$ Therefore, $$ z^{-pi}az^{pi}= a+{}_{-i}\delta_i^p(a)z^{pi}, \mbox{\quad} a\in u(\bar{D}) $$ and $$ z^{pi}az^{-pi}= a+\delta'_1(a)z^{pi}+{\delta'_1}^2(a)z^{2pi}+\ldots , $$ where $\delta'_1= (-1){}_{-i}\delta_i^p=i^p\delta_i^p$. So, $$ z^paz^{-p}= a+\frac{1}{i}\delta'_1(a)z^{pi}+c_2\frac{1}{i^2}{\delta'_1}^2(a)z^{2pi}+\ldots , $$ where $c_k$ are given by (\ref{(188)}) in lemma \ref{svva}. So, $z^p\in Z(D)$ iff $\delta_i^p= 0$. Suppose $\delta_i^p\ne 0$. Consider an element $Y\in Z(D)$, $w(Y)>0$. Let $$ Y= a_1z^p+\ldots , \mbox{\quad } a_1\in u(\bar{D}). $$ First note that $$ Y= a_1z^p+a_2z^{2p}+a_3z^{3p}+\ldots , \mbox{\quad} a_i\in u(\bar{D}) $$ Indeed, $Y$ must satisfy $[Y,s]= 0$, where $s$ is a generator of $u(Z(\bar{D}))$ over $u(\bar{F})$. Since $s\in u(Z(\bar{D}))$ and $w([z^k,s])=k+i$ if $(k,p)=1$ and $w([z^k,s])=\infty$ otherwise, we then have $[z^{i_k},s]= 0$ for every $k$, where $$ Y= \sum_{k= 1}^{\infty}a_kz^{i_k} $$ Therefore, $p|i_k$. Then, $Y$ must satisfy $Ya= aY$ for any $a\in u(\bar{D})$. Therefore, $a_1,\ldots a_i\in u(Z(\bar{D}))$ and we must have $$ aa_{i+1}-a_{i+1}a= a_1\delta'_1(a)/i $$ and $$ aa_{2i+1}-a_{2i+1}a= a_i\delta'_1(a)+a_1c_2{\delta'_1}^2(a). $$ Since $\Delta (a)= aa_{2i+1}-a_{2i+1}a$ is an inner derivation, we get ${\delta'_1}^2= \delta$, where $\delta$ is a derivation, which is a contradiction if $\delta\ne 0$ and $char D\ne 2$. In the last case we can use the same arguments with $a_{3i+1}$. Therefore, ${\delta'_1}^2= \delta= 0$ and $\delta'_1= 0$, and $z^p\in Z(D)$. Consider the algebra $W=u(Z(\bar{D}))((z))$. Since $z^p\in Z(D)$ and $u(\bar{F})\subset Z(D)$, we have $Z(W)=u(\bar{F})((z^p))=F$. So, $D\cong W\otimes_FC_D(W)$ by Double Centralizer theorem. It is clear that $C_D(W)$ is an unramified division algebra. Now suppose the proposition is proved for $[Z(\bar{D}):\overline{Z(D)}]=p^{k-1}$. By Albert's theorem (th.13 in \cite{Al}) $D_2$ then is a cyclic algebra as a product of cyclic subalgebras $D_i$, where $\bar{D_i}/\bar{F}$ is a simple purely inseparable extension and $D_i$ is a good splittable algebra. Assume $[Z(\bar{D}):\overline{Z(D)}]=p^{k}$. For a good embedding there exists a lift $\tilde{K}$ of a subfield $\overline{Z(D)}\subset K\subset Z(\bar{D})$ such that the extension $K/\overline{Z(D)}$ has degree $p$, i.e. $\bar{\tilde{K}}= K$, $\Gamma_{\tilde{K}}= \Gamma_{Z(D_2)}$, $u(K)\subset \tilde{K}$, $\tilde{K}/Z(D)$ is a purely inseparable extension of degree $p$. By the induction hypothesis the centralizer $C_D(\tilde{K})\cong A_1\otimes_{\tilde{K}}A_2$, where $A_2$ is a cyclic division algebra and $\bar{A_2}$ is a field. Note that $\bar{A_2}=Z(\bar{D})$. By theorem 6 in \cite{Al} we can assume $A_2=(L/\tilde{K},\sigma ,a)$, where $a$ generate $\tilde{K}$ over $Z(D)$. So, $A_2$ contains a maximal purely inseparable Kummer subfield $E=\tilde{K}(y)$ with $y^{p^{k-1}}=a$, so $E=Z(D)(y)$. By theorem 3 in \cite{Al} $L=L_0\times \tilde{K}$, where $L_0$ is cyclic of degree $p^{k-1}$ over $Z(D)$ and $yx_0=\sigma (x_0)y$, where $x_0\in L_0$. Consider the centralizer $B=C_D(L_0)$. We claim $B\cong B_1\otimes_{L_0} B_2$, where $B_2$ is a cyclic division algebra of degree $p$ and $B_2$ contains $\tilde{K}$. Note that $B$ contains $Z(D)(a)=\tilde{K}$ and $A_1$. If $\tilde{K}L_0=L$ is "unramified" over $L_0$, then we apply the arguments for the first step of our induction to the algebra $B$. By construction, $B_2$ then will contain $L$, so $\tilde{K}$. Suppose $L$ is totally ramified over $L_0$ and let $z$ be a parameter of $L$, i.e. an element with the least possible positive mean of valuation on $L$. Since $L$ is purely inseparable over $L_0$, $z^p$ is a parameter of $L_0$. We have $W:=C_B(L)= C_D(L)\cong A_1\otimes_{\tilde{K}}L$ is an unramified division algebra. Consider an embedding $u':\bar{L}=\bar{L_0}\hookrightarrow L_0$. As it was shown in the proof of theorem \ref{Cohen} there is a lift $\tilde{u'}$ of $u'$, $\tilde{u'}: \bar{W}\hookrightarrow W$. Now consider the subalgebra $W'=\tilde{u'}(\bar{W})((z^p))$. We have $Z(W')=u'(\bar{L})((z^p))=L_0$, so $W'$ is an unramified subalgebra in $B$. By Double Centralizer theorem, $B\cong W'\otimes_{L_0}C_B(W')$, where $C_B(W')$ is a division algebra of degree $p$ and contains $L_0(z)=L$, so it contains $\tilde{K}$ and it is cyclic by Albert's theorem (th.12 in \cite{Al}). Now we can word by word repeat the arguments in the proof of theorem 12 in \cite{Al} to show that there exists a cyclic Galois extension $L'$ of $L_0$ which is cyclic Galois over $Z(D)$, and $y$ acts as a Galois automorphism on $L'/Z(D)$ which generates $Gal(L'/Z(D))$. So, there is the cyclic subalgebra $D_2=(L'/Z(D), ad(y), y^{p^k})$ in $D$. Note that $A_2\subset D_2$, and $A_2$ is known to be a good splittable algebra with $[\bar{A_2}:\overline{Z(A_2)}]=[\Gamma_{A_2}:\Gamma_{Z(A_2)}]$. Since $\bar{A_2}=\bar{D_2}$ and $Z(A_2)=\tilde{K}$ is a purely inseparable extension of $Z(D)$, $D_2$ is a good splittable algebra such that $\bar{D_2}$ a field and $[\bar{D_2}:\overline{Z(D)}]=[\Gamma_{D_2}:\Gamma_{Z(D)}]$. By Double Centralizer theorem $D\cong D_1\otimes_{Z(D)}D_2$, where $D_1=C_D(D_2)$ must be an unramified division algebra, which completes the proof.\\ $\Box$ Combining all results in this section, we get the following theorem. \begin{th} \label{itog} Let $D$ be a finite dimensional good splittable central division algebra over a field $F=k((t))$. If $char ({F})= p>0$, then $D\cong D_1\otimes_{F}D_2\otimes_{F}A_1\otimes_{F}\ldots \otimes_{F}A_m$, where $A_i$ are cyclic division algebras such that $[\bar{A_i}:\overline{Z(D)}]= [\Gamma_{A_i}:\Gamma_{Z(D)}]$ and $\bar{A_i}/\overline{Z(D)}$ are simple purely inseparable field extensions, $D_1$ is an inertially split division algebra, $(ind (D_1), p)= 1$, $D_2$ is an unramified division algebra ($D_1, D_2, A_i$ may be trivial). If $char F=0$, then $D$ is an inertially split division algebra. \end{th} \section{Splittability and good splittability} In this section we collect some assorted results about a relation between splittable and good splittable division algebras and about splittable division algebras. We consider here only division algebras with the following property: $Z(\bar{D})/\overline{Z(D)}$ is a simple extension. \begin{prop} \label{cyclisity} Let $D$ be a central division algebra over $F$ of $char D=p>0$ such that $Z(\bar{D})=\bar{D}$ and $[Z(\bar{D}):\bar{F}]=p$. Then $D$ is a splittable algebra and the local height $i=i(u,z)$ (in the situatuion when it is defined, i.e. when $\alpha =id$) does not depend on $u$ and $z$. It is a good splittable algebra if $(i,p)=1$. If $p|i$, then there exists a parameter $z$ such that $z^p\in Z(D)$ and any "unramified" maximal subfield is cyclic Galois. So, in both cases $D$ is a cyclic division algebra of degree $p$. \end{prop} {\bf Proof.} Since $\bar{D}/\bar{F}$ is a simple extension, we have $[\bar{D}:\bar{F}]=[\Gamma_D:\Gamma_F]$. Indeed, consider the fields $E=F(s)$ and $E'=F(z)$, where $s$ is any element such that $\bar{s}$ is a primitive element of the extension $\bar{D}/\bar{F}$ and $z$ is any parameter of $D$. Then $[\bar{D}:\bar{F}]\le [E:F]\le [D:F]^{1/2}=([\bar{D}:\bar{F}][\Gamma_D:\Gamma_F])^{1/2}$, so $[\bar{D}:\bar{F}]\le [\Gamma_D:\Gamma_F]$. From another hand side, $[\Gamma_D:\Gamma_F]\le [E':F]\le ([\bar{D}:\bar{F}][\Gamma_D:\Gamma_F])^{1/2}$, so $[\bar{D}:\bar{F}]=[\Gamma_D:\Gamma_F]$. So, $D$ is splittable division algebra of degree $p$. If $Z(\bar{D})/\bar{F}$ is a separable extension, then $D$ is a good splittable algebra by theorem \ref{Cohen}. So, we assume it is a purely inseparable extension, $Z(\bar{D})=\bar{F}(\bar{u})$. For any lift $u$ of the element $\bar{u}$ let $u$ be an embedding constructed in lemma \ref{simple}, i.e. ${}^{(z,u)}\delta_j$ is defined by the values ${}^{(z,u)}\delta_j(u^k)$ for any $j$. By corollary \ref{ozamene3} the local height $i(u,z)$ does not depend on $z$, and by lemma \ref{ozamene2} $i(u,z)$ does not depend on $u$. For arbitrary embedding $u'$, since ${}^{(z,u')}\delta_{i(u',z)}$ is a derivation and $\bar{D}/\bar{F}$ is a simple extension, ${}^{(z,u')}\delta_{i(u',z)}$ is completely defined by a value at $\bar{u}$. Therefore, $i(u',z)=w(zu'(\bar{u})z^{-1}-u'(\bar{u}))$ and $i(u',z)$ is completely defined by the lift $u'(\bar{u})$. But arbitrary lift of $\bar{u}$ defines an embedding, on which we have proved $i$ does not depend. So, $i(u,z)$ does not depend on $z$ and $u$. Now assume $p|i$. Using lemma \ref{ozamene}, we can assume without loss of generality that ${}^{(z,u)}\delta_j=0$ if $j$ is not divisible by $p$. Indeed, if ${}^{(z,u)}\delta_j\ne 0$, then we apply lemma \ref{ozamene}, (ii) to show that there exists a parameter $z_j$ such that ${}^{(z_j,u)}\delta_j(u)=0$ and ${}^{(z_j,u)}\delta_k={}^{(z,u)}\delta_k$ for $k<j$, ${}^{(z_j)}\alpha =id$. Since ${}^{(z_j,u)}\delta_j$ is a derivation by proposition \ref{flyii} and by induction (similar arguments was already used in the proof of proposition \ref{X}), and since it is defined by the values on $u^k$, so by the values on $u$, we have ${}^{(z,u)}\delta_j= 0$. Since for $j_1>j_2$ we have $w(z_{j_1}-z_{j_2})> j_1-i$, the sequence $\{z_j\}$ convereges to a parameter $z'$, which satisfies our condition. So, there exists the subalgebra $A=u(\bar{D})((z^p))$. Let's show that $Z(D)\subset A$. Note that every element $a\in D$ can be written as $a=a_0+a_1z+\ldots +a_{p-1}z^{p-1}$, where $a_i\in A$. Note that $z^kAz^{-k}\subset A$ for every $k$. So, if $a\in Z(D)$, then $za_jz^{-1}=a_j$ and $ua_jz^ju^{-1}=a_jz^j$ for every $j$. For $j>0$ we have $a_jz^j=\sum_k a_{jk}z^{kp+j}$, so by corollary \ref{ozamene3} $ua_jz^ju^{-1}\ne a_jz^j$. Therefore, $a=a_0\in A$. Since $A\ne D$, $A$ must be commutative, so $z^p\in Z(D)$. Moreover, $A/Z(D)$ is cyclic Galois. Since the arguments work for arbitrary lift $u$ of the element $\bar{u}$, arbitrary "unramified" maximal subfield in $D$ must be Galois over $F$. Now let $(i,p)=1$. Using lemma \ref{ozamene}, (iii) we can find a parameter $z$ and a primitive element $s\in \bar{D}$ such that ${}^{(z,u)}\delta_i(s)=sc$, where $c\in \bar{F}$. Indeed, since $(i,p)=1$, there exists $k$ such that $1-ki$ is divisible by $p$. So, by lemma \ref{ozamene}, (iii) for a parameter $z'=u({}^{(z,u)}\delta_i(\bar{u})^k)z$ we have ${}^{(z',u)}\delta_i(\bar{u})\in \bar{F}$, so by lemma \ref{ozamene2}, (iii) ${}^{(z',u)}\delta_i(s)=1$, where $s=\bar{u}{}^{(z',u)}\delta_i(\bar{u})^{-1}$. Now, there exists $k_1$ such that $-ik_1-1$ is divisible by $p$, so for $z''=s^{k_1}z'$ we have ${}^{(z'',u)}\delta_i(s)=sc$, where $c=s^{-ik_1-1}\in \bar{F}$. It is easy to see that, since $s=\bar{u}a$, where $a\in \bar{F}$, the map ${}^{(z,u)}\delta_j$ is uniquely defined also by ${}^{(z,u)}\delta_j(s^k)$, so by ${}^{(z,u)}_m\delta_l(s)$ for $l\le j$. So, we assume without loss of generality that $s=\bar{u}$, $z=z''$. Using lemma \ref{ozamene2}, (ii) we can find a converge sequence $\{u_j\}$, $u_j\in D$, $j\ge i$ such that $u_{j+1}=u_j+b_jz^{j+1-i}$, $u_i=u$, $b_j\in u_j(\bar{D})$ (here $u_j$ is an embedding defined by $u_j$, see lemma \ref{simple}) and ${}^{(z,u_j)}_m\delta_k(\bar{u})\bar{u}^{-1}\in \bar{F}$ for all $k\le j$ and all $m$. Indeed, suppose it is true for $j\ge i$. Let ${}^{(z,u_j)}_m\delta_{j+1}(\bar{u})=a_0+\ldots a_{p-1}\bar{u}^{p-1}$, $a_k\in \bar{F}$. Since ${}^{(z,u_j)}_m\delta_i={}^{(z,u)}_m\delta_i=m{}^{(z,u)}\delta_i$, we have $$ {}^{(z,u_j)}_m\delta_i(a_k\bar{u}^k)-\frac{\partial }{\partial \bar{u}}({}^{(z,u_j)}_m\delta_i(\bar{u}))a_k\bar{u}^k= (k-1)mca_k\bar{u}^k. $$ So, $u_{j+1}=u_j-u_j(\sum_{k, k\ne 1}(k-1)^{-1}m^{-1}c^{-1}a_k\bar{u}^k)z^{j+1-i}$ will satisfy our condition. We will denote by $u$ now a limit of the sequence $\{u_j\}$. Using induction and proposition \ref{flyii} one can easily show that ${}^{(z,u)}_m\delta_j(\bar{u}^k)\bar{u}^{-k}\in \bar{F}$ for any integer $k$. So, there is the subalgebra $A=u(\bar{F})((z))$ in $D$. Using similar arguments as in the case $p|i$, one can show that $A$ contains $Z(D)$. Since $A\ne D$, it must be commutative, so $u^p\in Z(D)$. Then $u$ is a good embedding, which completes the proof. \\ $\Box$ Let $D$ be a splittable division algebra and let $Z(\bar{D})/\overline{Z(D)}$ be a purely inseparable extension. As it was shown in the proof of lemma \ref{goodspl}, then there exists a parameter $z$ in $D$ such that ${}^{(z,u)}\delta_{i}|_{Z(\bar{D})}\ne 0$, where $i=i(u,z)$ is a local height. Though $D$ may be not a good splittable algebra, the arguments from there are valid for every splittable algebra. We will call such a parameter {\it an appropriate parameter}, and the number $i(u)=\max_zi(u,z)=i(u,z)$ for an appropriate parameter {\it a semilocal height}. Let's prove the following simple lemma. \begin{lemma} \label{simple2} Let $D$ be a splittable central division $p$-algebra over $F$, where $p=char D>0$, and let $Z(\bar{D})=\bar{F}(s)$ be a simple extension over $\bar{F}$. Then i) there exists an embedding $u$ such that ${}^{(z,u)}_l\delta_j|_{Z(\bar{D})}$ is defined by the values ${}^{(z,u)}_l\delta_j(s^k)$ for any $j,l,z$ (as in lemma \ref{simple}); ii) $[Z(\bar{D}):\bar{F}]=[\Gamma_{D}:\Gamma_{F}]$; iii) if $\alpha |_{Z(\bar{D})}\ne id$ or $i(u)$ is divisible by $p$, then there exists a subalgebra $A=u(\bar{D})((z))$ for some appropriate parameter $z$ such that $Z(D)\subset Z(A)$. Moreover, $Z(A)$ is a cyclic Galois extension over $Z(D)$. \end{lemma} {\bf Proof.} i) For arbitrary embedding $u$ consider the field $E=u(Z(\bar{D}))F\subset D$ and the centralizer $W=C_D(E)$. We have $\bar{W}=\bar{D}$ and so $Z(\bar{W})=\bar{E}$. Therefore, $W$ must be an unramified division algebra, and by theorem \ref{Cohen} there exists a lift on $\bar{W}$ of arbitrary embedding $\bar{E}\hookrightarrow E$. Now we can take an embedding defined by the element $s$ as in lemma \ref{simple}. It's lift will be desired embedding. We will denote this embedding also by $s$. ii) By proposition 1.7. in \cite{JW} the basic homomorphism $\theta_D$ (see introduction) is surjective. So, it is sufficient to prove the assertion only for the centralizer $C_D(K)$, where $K$ is a lift of a Galois part of the extension $Z(\bar{D})/\bar{F}$. So, we will assume below $Z(\bar{D})/\bar{F}$ is a purely inseparable extension. Consider a maximal separable subfield $M$ in $\bar{D}$, and let $M'$ be a separable part of the extension $M/\bar{F}$. By \cite{JW}, th.2.8, th.2.9. there exists an inertial lift of $M'$ in $D$, say $\tilde{M}$. Consider the centralizer $B=C_D(\tilde{M})$. Then $\bar{B}$ is a field. Our assertion will be proved if we show it for $B$, since $[\tilde{M}:F]=ind (\bar{D})$ and $[D:F]=ind(\bar{D})^2[Z(\bar{D}):\bar{F}][\Gamma_{D}:\Gamma_{F}]$. Since $\bar{B}/\overline{Z(B)}$ is a simple extension, we can repeat the arguments from the beginning of proposition \ref{cyclisity}. iii) If $\alpha |_{Z(\bar{D})}\ne id$, consider the parameter $z$ from proposition \ref{X}. Then, clearly, $A=u(\bar{D})((z))$ will be a subalgebra with the center $K$, which is an inertial lift of a Galois part of the extension $Z(\bar{D})/\bar{F}$. Assume $\alpha |_{Z(\bar{D})}= id$ and $i(u)$ is divisible by $p$. Let $z$ be an appropriate parameter. Using lemma \ref{ozamene}, we can prove that ${}^{(z,u)}\delta_j=0$ if $j$ is not divisible by $p$. Indeed, let ${}^{(z,u)}\delta_j\ne 0$ be the first map with this property for $(j,p)=1$. If ${}^{(z,u)}\delta_j|_{Z(\bar{D})}=0$, then we apply lemma \ref{ozamene}, (i) to show that there exists a parameter $z_j$ such that ${}^{(z_j,u)}\delta_j=0$ and ${}^{(z_j,u)}\delta_k={}^{(z,u)}\delta_k$ for $k<j$, ${}^{(z_j)}\alpha =id$, since ${}^{(z,u)}\delta_j$ is a derivation by proposition \ref{flyii} and by induction (similar arguments was already used in the proof of proposition \ref{X}) and so it is an inner derivation by Scolem-Noether theorem. If ${}^{(z,u)}\delta_j|_{Z(\bar{D})}\ne 0$, then we apply lemma \ref{ozamene}, (ii) to show that there exists a parameter $z_j$ such that ${}^{(z_j,u)}\delta_j(s)=0$ and ${}^{(z_j,u)}\delta_k={}^{(z,u)}\delta_k$ for $k<j$, ${}^{(z_j)}\alpha =id$. Since ${}^{(z_j,u)}\delta_j$ is a derivation and since its restriction on $Z(\bar{D})$ is defined by the values on $s^k$, so by the values on $s$, we have ${}^{(z,u)}\delta_j|_{Z(\bar{D})}= 0$, and we reduce the problem to the previous case. Since for $j_1>j_2$ we have $w(z_{j_1}-z_{j_2})> j_1-i$, the sequence $\{z_j\}$ convereges to a parameter $z'$, which satisfies our condition. Therefore, there exists a subalgebra $A=u(\bar{D})((z'))$ in $D$. Using the same arguments as in proposition \ref{cyclisity} one can show that $Z(D)\subset Z(A)$ Since $z'$ preserves $A$, it preserves the centre $Z(A)$ >From the other hand side, it acts nontrivially on it. So, $Z(A)$ is a cyclic Galois extension of degree $p$, and $ad(z')$ generates its Galois group. \\ $\Box$ This lemma shows that the study of splittable $p$-algebras over $F$ can be reduced to the study of splittable $p$-algebras with a purely inseparable extension $Z(\bar{D})/\bar{F}$ and $(i(u),p)=1$. \begin{defi} \label{ivariant} Let $D$ be a splittable division $p$-algebra with a purely inseparable extension $Z(\bar{D})/\bar{F}$. For any element $a\in \bar{D}$ define the number $$ d_D(a)= \max_{u,z} w(z^{-i(u,a)}u(a)z^{i(u,a)}- u(a)-u({}^{(z,u)}_{-i(u,a)}\delta_{i(u,a)}(a))z^{i(u,a)})\in {\mbox{\dbl N}}\cup\infty , $$ where parameters $z$ are taken from the set of appropriate parameters and $i(u,a)$ was defined in corollary \ref{ozamene3}. \end{defi} It seems that the number $d_D(a)$ will play the role of a higher order level in a splittable division algebra. We will see that it codes a part of information about a division algebra. \begin{lemma} \label{predvarit} Let $D$ be a splittable division $p$-algebra, $p>2$, with a purely inseparable simple extension $Z(\bar{D})/\bar{F}$, let $u$ be some fixed embedding $u:\bar{D}\hookrightarrow D$. Suppose $Z(\bar{D})=\bar{F}(a)$ and $(i(u,a),p)=1$. Suppose $d(u,a)\le 2i(u,a)$. Let $z$ be a parameter such that ${}^{(z,u)}\delta_{i(u,a)} ({}^{(z,u)}_{-i(u,a)}\delta_{i(u,a)}(a))=0$, ${}^{(z)}\alpha =id$ and ${}^{(z,u)}_{-i(u,a)}\delta_{q}|_{{\mbox{\sdbl F}}_p(a)}=0$ for $i(u,a)<q<d(u,a)$. Put $j(k):=i(u,a^{p^k})$. Suppose for every $k\ge 1$ a parameter $z_k$ such that ${}^{(z_k,u)}_{-j(k)}\delta_{r}|_{{\mbox{\sdbl F}}_p(a^{p^k})}=0$ for $j(k)<r<d(u,a^{p^k})$ satisfy a condition ${}^{(z_k,u)}\delta_{i(u,a)}={}^{(z,u)}\delta_{i(u,a)}$, ${}^{(z_k)}\alpha ={}^{(z)}\alpha$. Suppose for every $k\ge 1$ we have $d(u,a^{p^k})-j(k)=d(u,a)-j(0)$. Then the maps ${}^{(z,u)}_{w+(p-1-r)j(k)}\delta_{\zeta}$, $rj(k)<\zeta\le (r-1)j(k)+d(u,a^{p^k})$, $r\in \{1, \ldots , p-1\}$, $k\ge 0$ satisfy the following properties: $$ {}^{(z,u)}_{w+(p-1-r)j(k)}\delta_{\zeta}|_{{\mbox{\sdbl F}}_p(a^{p^k})}=c_{w+(p-1-r)j(k),\zeta ,1}\delta +\ldots + c_{w+(p-1-r)j(k),\zeta ,r}\delta^{r}, $$ where the derivation $\delta$ was defined in lemma \ref{(5)}, $c_{w+(p-1-r)j(k),\zeta ,r}\in Z(\bar{D})$, $c_{w+(p-1-r)j(k),\zeta ,r}\ne 0$ only if $\zeta = (r-1)j(k)+d(u,a^{p^k})$. Moreover, $c_{w+(p-1-r)j(k), (r-1)j(k)+d(u,a^{p^k}),r}\ne 0$ if $w=i(u,a) \mbox{\quad mod\quad}p$; $$c_{w+(p-1-r)j(k),(r-1)j(k)+d(u,a^{p^k}), r}=r!c_{w+(p-r)j(k),(r-2)j(k)+d(u,a^{p^k}), r-1}{}_{w+(p-1-r)j(k)}\delta_{j(k)}(a^{p^k}),$$ and ${}^{(z_k,u)}_{w+(p-2)j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})= {}^{(z_k,u)}_{-j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})$. \end{lemma} {\bf Proof.} The proof is similar to the proof of lemma \ref{(5)}, (i). It is by induction on $r$ simultaneously for all $k\ge 0$. For $r=1$, using lemma \ref{triviall} and induction, one can easily show that ${}^{(z_k,u)}\delta_{q}(a^{p^k})=-(j(k))^{-1}{}^{(z_k,u)}_{-j(k)}\delta_{q}(a^{p^k})$ for $j(k)\le q< d(u,a^{p^k})$ (we assume here $z_0=z$). By lemma \ref{vtorinv}, (i) we have $d(u,a)-i(u,a)=i(u,a) \mbox{\quad mod\quad}p$. So, by lemma \ref{vtorinv}, (ii) and by induction we have $j(k)=j(0) \mbox{\quad mod\quad}p$. So, ${}^{(z_k,u)}_{w+(p-2)j(k)}\delta_{q}|_{{\mbox{\sdbl F}}_p(a^{p^k})}=0$ if $j(k)< q< d(u,a^{p^k})$ and ${}^{(z_k,u)}_{w+(p-2)j(k)}\delta_{q}|_{{\mbox{\sdbl F}}_p(a^{p^k})}\ne 0$ only if $q=d(u,a^{p^k})$. Since ${}^{(z_k,u)}_{-j(k)}\delta_{j(k)}|_{{\mbox{\sdbl F}}_p(a^{p^k})}$ is a derivation and since, by proposition \ref{flyii}, (i), the map ${}^{(z_k,u)}_{-j(k)}\delta_{d(u,a^{p^k})}|_{{\mbox{\sdbl F}}_p(a^{p^k})}$ must be a derivation, we have ${}^{(z_k,u)}_{w+(p-2)j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})\in Z(\bar{D})$. For, as it was shown in the proof of lemma \ref{(5)}, (ii) for any derivation $\delta$ we have $\delta (b)\in Z(\bar{D})$ for any $b\in Z(\bar{D})$. Since ${}^{(z_k,u)}_{w+(p-2)j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})= q_1{}^{(z_k,u)}_{-j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})+ q_2{}^{(z_k,u)}_{m}\delta_{j(0)}({}^{(z_k,u)}_{-j(k)}\delta_{j(k)}(a^{p^k}))$ for some integer $q_1,q_2,m$, we have proved our assertion. So, $c_{w+(p-2)j(k),d(u,a^{p^k}), 1}\in Z(\bar{D})$. If $w=j(0) \mbox{\quad mod\quad}p$, then ${}^{(z_k,u)}_{w+(p-2)j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})={}^{(z_k,u)}_{-j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})$, since $w+(p-1)j(0)=0\mbox{\quad mod\quad}p$ and $char D> 2$. So, we have $c_{w+(p-2)j(k),d(u,a^{p^k}), 1}\ne 0$. Put now $t=a^{p^k}$. For arbitrary $r$ by proposition \ref{flyii}, (i) we have $$ {}^{(z_k,u)}_{w+(p-1-r)j(k)}\delta_{\zeta}(t^q)= q{}_{w+(p-1-r)j(k)}\delta_{\zeta}(t)t^{q-1}+ $$ $$ {}_{w+(p-1-r)j(k)}\delta_{j(k)}(t) \sum_{l= 0}^{q-2}{}_{w+(p-r)j(k)}\delta_{\zeta -j(k)}(t^{q-1-l})t^l+ $$ $$ {}_{w+(p-1-r)j(k)}\delta_{d(u,t)}(t) \sum_{l= 0}^{q-2}{}_{w+(p-1-r)j(k)+d(u,t)}\delta_{\zeta -d(u,t)}(t^{q-1-l})t^l+ $$ $$ \sum_{i=d(u,t)+1}^{\zeta -1} {}_{w+(p-1-r)j(k)}\delta_{i}(t) \sum_{l= 0}^{q-2}{}_{w+(p-1-r)j(k)+i}\delta_{\zeta -i}(t^{q-1-l})t^l. $$ Using the same arguments as in the proof of lemma \ref{(5)},(i) we see that ${}_{w+(p-1-r)j(k)}\delta_{\zeta}(t^p)=0$ and ${}_{w+(p-1-r)j(k)}\delta_{\zeta}|_{{\mbox{\sdbl F}}_p(t)}=c_{w+(p-1-r)j(k),\zeta ,1}\delta +\ldots + c_{w+(p-1-r)j(k),\zeta ,p-1}\delta^{p-1}$. To show that $c_{w+(p-1-r)j(k),\zeta ,i}=0$ for $i>r$ it suffice, by formulae (\ref{recurrent}) in lemma \ref{(5)}, to show that all the maps in the formula above are represented in the form $c_1\delta +\ldots +c_{r-1}\delta^{r-1}$. Let us show it in details. Since $\zeta -d(u,t)-1<(r-1)j(k)$, by lemma \ref{(5)}, (ii) ${}_{m}\delta_{\zeta -i}|_{{\mbox{\sdbl F}}_p(t)}=c_{m,\zeta -i,1}\delta +\ldots +c_{m,\zeta -i,r-2}\delta^{r-2}$ for any $i>d(u,t)$. If $w=j(0) \mbox{\quad mod\quad}p$, then $w+(p-1-r)j(k)+d(u,t)+(r-2)j(k)=0 \mbox{\quad mod\quad}p$. Since $\zeta -d(u,t)\le (r-1)j(k)$, by lemma \ref{(5)}, (ii) we have ${}_{w+(p-1-r)j(k)+d(u,t)}\delta_{\zeta -d(u,t)}|_{{\mbox{\sdbl F}}_p(t)}=c_{w+(p-1-r)j(k)+d(u,t),\zeta -d(u,t),1}\delta +\ldots +c_{w+(p-1-r)j(k)+d(u,t),\zeta -d(u,t),r-2}\delta^{r-2}$. If $w\ne j(0) \mbox{\quad mod\quad}p$, then by the same reason we have ${}_{w+(p-1-r)j(k)+d(u,t)}\delta_{\zeta -d(u,t)}|_{{\mbox{\sdbl F}}_p(t)}=c_{w+(p-1-r)j(k)+d(u,t),\zeta -d(u,t),1}\delta +\ldots +c_{w+(p-1-r)j(k)+d(u,t),\zeta -d(u,t),r-1}\delta^{r-1}$ and by lemma \ref{(5)}, (i) $c_{w+(p-1-r)j(k)+d(u,t),\zeta -d(u,t),r-1}\in Z(\bar{D})$ as a product of elements from $Z(\bar{D})$. At last, by the induction hypothesis ${}_{w+(p-r)j(k)}\delta_{\zeta -j(k)}|_{{\mbox{\sdbl F}}_p(t)}=c_{w+(p-r)j(k),\zeta -j(k),1}\delta +\ldots +c_{w+(p-r)j(k),\zeta -j(k),r-1}\delta^{r-1}$ and $c_{w+(p-r)j(k),\zeta -j(k),r-1}\ne 0$ only if $\zeta -j(k)=(r-2)j(k)+d(u,t)$, and $c_{w+(p-r)j(k),\zeta -j(k),r-1}\in Z(\bar{D})$. Since ${}_{w+(p-1-r)j(k)}\delta_{j(k)}(t)\in Z(\bar{D})$, by formulae (\ref{recurrent}) we get $c_{w+(p-1-r)j(k),\zeta ,r}\in Z(\bar{D})$ and if $w=j(0) \mbox{\quad mod\quad}p$, then $c_{w+(p-1-r)j(k),\zeta ,r}\ne 0$ iff $\zeta =(r-1)j(k)+d(u,t)$, $$c_{w+(p-1-r)j(k), (r-1)j(k)+d(u,t),r}=r!c_{w+(p-r)j(k),(r-2)j(k)+d(u,t), r-1}{}_{w+(p-1-r)j(k)}\delta_{j(k)}(t)\ne 0.$$ The lemma is proved.\\ $\Box$ \begin{lemma} \label{predvarit2} Let $D$ be a division algebra as in lemma \ref{predvarit}. Suppose $d(u,a)\le 2i(u,a)$ and $char D>2$. Then for every $k$ there exists a parameter $z_k$ such that ${}^{(z_k,u)}_{-j(k)}\delta_{r}|_{{\mbox{\sdbl F}}_p(a^{p^k})}=0$ for $j(k)<r<d(u,a^{p^k})$ and ${}^{(z_k)}\alpha ={}^{(z)}\alpha$, ${}^{(z_k,u)}\delta_{j(l)}={}^{(z,u)}\delta_{j(l)}$ for all $l\le k$ (we use here the notation defined in lemma \ref{predvarit}). Moreover, for every $k\ge 1$ we have $d(u,a^{p^k})-j(k)=d(u,a)-j(0)$ and $${}^{(z_{k},u)}_{-j(k)}\delta_{d(u,a^{p^k})}(a^{p^k})=-{}^{(z_{k-1},u)}_{-j(k-1)} \delta_{d(u,a^{p^{k-1}})}(a^{p^k})c_{d(u,t)-j(k-1), j(k)-j(k-1),p-1},$$ where $c_{d(u,t)-j(k-1), j(k)-j(k-1),p-1}$ is defined in lemma \ref{predvarit}. \end{lemma} {\bf Proof.} The proof is by induction on $k$. By lemma \ref{vtorinv} $d(u,a)=2j(0)\mbox{\quad mod \quad }p$ and $j(1)=d(u,a)+(p-1)j(0)$. So, by the induction hypothesis we can assume for arbitrary $k$ that $d(u,a^{p^{k-1}})=2j(0)\mbox{\quad mod \quad }p$ and $j(k-1)=j(0)\mbox{\quad mod \quad }p$, and $j(k)=d(u,a^{p^{k-1}})+(p-1)j(k-1)$. For the convinience we can start with a parameter $z=z_0$, which satisfy the conditions of lemma \ref{predvarit}. Indeed, taking an appropriate parameter $z$ and changing it by a parameter $u(c)z$ for an appropriate $c\in Z(\bar{D})$ (as in the proof of proposition \ref{cyclisity}), we can assume that ${}^{(z,u)}_{-j(0)}\delta_{j(0)}(a)\in Z(\bar{D})^p$. Now, using arguments from the proof of lemma \ref{vtorinv}, (i), we can find such a parameter $z_0$. The idea of the proof is the following. We prove first that ${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{j(k)+d(u,a)-j(0)}(a^{p^k})\ne 0$. Then we prove that there exists a parameter $z_k$ such that ${}^{(z_{k},u)}_{-j(k)}\delta_{\zeta}(a^{p^k})= 0$ for $j(k)<\zeta <j(k)+d(u,a)-j(0)$ and ${}^{(z_{k},u)}_{-j(k)}\delta_{j(k)+d(u,a)-j(0)} (a^{p^k})\ne 0$. It will be shown that $z_k$ satisfy the conditions of lemma. So, assume $j(k)\le \zeta \le j(k)+d(u,a)-j(0)=j(k)+d(u,a^{p^{k-1}})-j(k-1)$. Put $t=a^{p^{k-1}}$. By proposition \ref{flyii}, (i) we have $$ {}^{(z_{k-1},u)}_{-j(k-1)}\delta_{\zeta}(t^p)= $$ $$ {}^{(z_{k-1},u)}_{-j(k-1)}\delta_{d(u,t)}(t) \sum_{l= 0}^{p-2}{}^{(z_{k-1},u)}_{d(u,t)-j(k-1)}\delta_{\zeta -d(u,t)}(t^{p-1-l})t^l+\ldots + $$ $$ {}^{(z_{k-1},u)}_{-j(k-1)}\delta_{\zeta -(p-1)j(k-1)}(t) \sum_{l= 0}^{p-2}{}^{(z_{k-1},u)}_{\zeta -pj(k-1)}\delta_{(p-1)j(k-1)}(t^{p-1-l})t^l+ $$ $$ \sum_{i=\zeta -(p-1)j(k-1)+1}^{\zeta -1} {}^{(z_{k-1},u)}_{-j(k-1)}\delta_{i}(t) \sum_{l= 0}^{p-2}{}^{(z_{k-1},u)}_{i-j(k-1)}\delta_{\zeta -i}(t^{q-1-l})t^l. $$ By lemma \ref{(5)}, (i) in the last sum ${}^{(z_{k-1},u)}_{i-j(k-1)}\delta_{\zeta -i}|_{{\mbox{\sdbl F}}_p(t)}=c_{i-j(k-1),\zeta -i,1}\delta +\ldots +c_{i-j(k-1),\zeta -i,p-2}\delta^{p-2}$, since $\zeta -i<(p-1)j(k-1)$. So, this sum is equal to zero. By lemma \ref{(5)}, (ii) we have ${}^{(z_{k-1},u)}_{\zeta -pj(k-1)} \delta_{(p-1)j(k-1)}|_{{\mbox{\sdbl F}}_p(t)}= c_{\zeta -pj(k-1),(p-1)j(k-1),1}\delta +\ldots + c_{\zeta -pj(k-1),(p-1)j(k-1),p-1}\delta^{p-1}$ and $c_{\zeta -pj(k-1),(p-1)j(k-1),p-1}\ne 0$ iff $\zeta =j(k-1)=j(0)\mbox{\quad mod\quad}p$. By lemma \ref{(5)}, (i) we have ${}^{(z_{k-1},u)}_{m}\delta_{q}|_{{\mbox{\sdbl F}}_p(t)}=c_{m,q,1}\delta +\ldots +c_{m,q,p-1}\delta^{p-1}$ for $(p-1)j(k-1)<q<(p-1)j(k-1)+d(u,a)-j(0)$, and by lemma \ref{predvarit} $c_{m,q,p-1}=0$. By lemma \ref{predvarit} we have ${}^{(z_{k-1},u)}_{d(u,t)-j(k-1)}\delta_{\zeta -d(u,t)}|_{{\mbox{\sdbl F}}_p(t)}= c_{d(u,t)-j(k-1), \zeta -d(u,t), 1}\delta +\ldots + c_{d(u,t)-j(k-1), \zeta -d(u,t), p-1}\delta^{p-1}$ with $c_{d(u,t)-j(k-1), \zeta -d(u,t), p-1}\ne 0$ if $\zeta -d(u,t)=j(0)$. So, we have the following picture: ${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{\zeta}(t^p)\ne 0$ only if $\zeta =j(0)\mbox{\quad mod \quad }p$ or if $\zeta =j(k)+d(u,a)-j(0)$. In the last case $${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{\zeta}(t^p)=-{}^{(z_{k-1},u)}_{-j(k-1)} \delta_{d(u,t)}(t^p)c_{d(u,t)-j(k-1), j(k)-j(k-1),p-1},$$ where $c_{d(u,t)-j(k-1), j(k)-j(k-1),p-1}$ can be calculated using lemma \ref{predvarit}. Let's show that there exists a parameter $z_{k}$ such that ${}^{(z_{k},u)}_{-j(k-1)}\delta_{\zeta}(t^p)=0$ for $j(k)<\zeta <j(k)+ d(u,a)-j(0)$. By lemma \ref{ozamene}, (ii) there exists a change of parameters $z_{k-1}\mapsto z'=z_{k-1}+bz_{k-1}^{p+1}$ such that ${}^{(z',u)}_{-j(k-1)}\delta_{j(k)+p}(t^p)=0$. It suffice to prove that any such a change of parameters as in lemma \ref{ozamene}, (ii) with $p|q$ changes only the values of maps ${}_{-j(k-1)}\delta_{\zeta}$ with $\zeta =j(0)\mbox{\quad mod \quad }p$. For, if it is true, we can make several changes and kill all nonzero maps ${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{\zeta}$ with $j(k)<\zeta <j(k)+ d(u,a)-j(0)$, since they are derivations and therefore are completely defined by their values at $t^p$. To prove it, we can use the calculations in the proof of lemma \ref{ozamene}, (ii). Since $d(u,a)-j(0)\le j(0)$, it is easy to see that for a change $z\mapsto z'=z+bz^{kp+1}$, $p>2$ we have there $$ z'^{-j(k-1)}t^pz'^{j(k-1)}=t^p+{}^{(z,u)}_{-j(k-1)}\delta_{j(k)}(t^p)z^{j(k)}+\ldots + {}^{(z,u)}_{-j(k-1)}\delta_{j(k)+j(0)}(t^p)z^{j(k)+j(0)}+\ldots . $$ Since $z'=z+bz^{kp+1}$, any power $z^{l}$ can be expressed as a series in $z'$, all powers of which are equal to $l$ modulo $p$. So, this change will change only maps with right indexes equal to $j(k)$ modulo $p$. Since ${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{\zeta}(t^p)\ne 0$ only if $\zeta =j(0)\mbox{\quad mod \quad }p$ for $\zeta <j(k)+d(u,a)-j(0)$, our assertion is proved. So, there exists a parameter $z_k$ we have: ${}^{(z_k,u)}_{-j(k-1)}\delta_{\zeta}(t^p)\ne 0$ only if $\zeta =j(k)+d(u,a)-j(0)$ or $\zeta =j(k)$. Since $z_k$ was constructed as a sequence of changes as in lemma \ref{ozamene}, (ii), we have ${}^{(z_k)}\alpha ={}^{(z_{k-1})}\alpha$ and ${}^{(z_k,u)}\delta_{j(q)}={}^{(z_{k-1},u)}\delta_{j(q)}$ for any $q\le k$. At last, let's prove that ${}^{(z_k,u)}_{-j(k)}\delta_{\zeta}(t^p)\ne 0$ only if $\zeta =j(k)+d(u,a)-j(0)$ or $\zeta =j(k)$. But this follows immediately from the definition of these maps, since $j(k)=j(k-1)\mbox{\quad mod \quad}p$, $d(u,a)-j(0)\le j(0)$ and $char D>2$. In particular, ${}^{(z_k,u)}_{-j(k)}\delta_{j(k)}(t^p)={}^{(z_k,u)}_{-j(k-1)}\delta_{j(k)}(t^p)$, ${}^{(z_k,u)}_{-j(k)}\delta_{j(k)+d(u,a)-j(0)}(t^p)={}^{(z_k,u)}_{-j(k-1)}\delta_{j(k)+d(u,a)-j(0)}(t^p)$. The lemma is proved.\\ $\Box$ Now we can prove the following theorem. \begin{th} \label{posledn} Let $D$ be a division $p$-algebra of $char D=p>2$ with the center $Z(D)=F$. Suppose $Z(\bar{D})=\bar{D}$ and $\bar{D}/\bar{F}$ is a simple purely inseparable extension, $\bar{D}=\bar{F}(a)$. Suppose that the semilocal height $i(u)$, which does not depend on the embedding $u$ in this case, is not divisible by $p$. Then $d_D(a)>i(u)$. \end{th} {\bf Proof.} By lemma \ref{simple2}, (ii) $[\bar{D}:\bar{F}]=[\Gamma_D:\Gamma_F]$. So, the field $F(\tilde{a})$, where $\tilde{a}$ is a lift of $a$, is a maximal "unramified" subfield and therefore $D$ is a splittable division algebra. Obviously, $\alpha =id$. Since ${}^{(z,u)}\delta_{i(u,z)}$ is a derivation and $\bar{D}/\bar{F}$ is a simple extension, ${}^{(z,u)}\delta_{i(u,z)}$ is completely defined by a value at $a$. So, by lemma \ref{ozamene} $i(u,z)$ does not depend on $z$ and $i(u,z)=i(u)$. Therefore, $i(u)=w(zu(a)z^{-1}-u(a))$ and $i(u)$ is completely defined by the lift $u(a)$. From the other hand side, any lift $\tilde{a}$ of $a$ defines, by lemma \ref{simple}, an embedding $\tilde{a}$, and by lemma \ref{ozamene2} $i(\tilde{a})$ does not depend on $\tilde{a}$. So, $i(u)$ does not depend on $u$. The idea of the proof is following. We consider linear spaces which are the images of the maps ${}^{(z,u)}\delta_{j(k)}|_{\bar{F}(a^{p^k})}$ in $\bar{D}$, where $j(k)$ were defined in lemma \ref{predvarit2} and $z,u$ are fixed. We show that every such spase has zero intersection with each other if $d_D(a)\le i(u)$. Then we show that this contradicts with the fact that $u(a)$ generate a finite dimensional space over $F$. So, assume $d_D(a)\le i(u)$. To calculate the spaces ${}^{(z,u)}\delta_{j(k)}({\bar{F}(a^{p^k})})\in \bar{D}$ we use lemmas \ref{vtorinv}, \ref{predvarit} and \ref{predvarit2}. We fix a parameter $z$ defined in lemma \ref{predvarit}. By lemmas \ref{simple}, \ref{ozamene2}, (iii) we can find a primitive element $\bar{u}\in \bar{D}$ of the extension $\bar{D}/\bar{F}$ such that ${}^{(z,u)}\delta_{j(0)}(\bar{u})=1$, where $u$ is an embedding defined in lemma \ref{simple} for some lift $u$ of the element $\bar{u}$. Using lemma \ref{ozamene}, (ii) we can find an embedding $u$ such that ${}^{(z,u)}\delta_{d(u,\bar{u})}(\bar{u})\notin {}^{(z,u)}\delta_{j(0)}(\bar{D})$. We fix this embedding. From lemmas \ref{ozamene}, \ref{ozamene2} immediately follows that $d(u,\bar{u})=d_D(\bar{u})=d_D(a)$. So, we assume without loss of generality $a=\bar{u}$. Put $J(k):={}^{(z,u)}\delta_{j(k)}(a^{p^k})$. Put $A(k):={}^{(z,u)}\delta_{j(k)}(\bar{F}(a^{p^k}))$, $A'(k):=\bar{F}(a^{p^{k+1}})\cdot a^{p^k(p-1)}J(k)$. We have $A(k)=\oplus_{q=0}^{p-2}\bar{F}(a^{p^{k+1}})\cdot a^{p^kq}J(k)$ and $\bar{D}\cdot J(k)=A(k)\oplus A'(k)$ as ${\mbox{\dbl F}}_p$-linear spaces. From lemma \ref{vtorinv} follows that $${}^{(z,u)}\delta_{j(k)}(a^{p^k})={}^{(z_k,u)}\delta_{j(k)}(a^{p^k})= q{}^{(z_{k-1},u)}_{-j(k-1)}\delta_{d(u,a^{p^{k-1}})}(a^{p^{k-1}})c_{d(u,a^{p^{k-1}})-j(k-1), (p-1)j(k-1), p-1},$$ where $q\in {\mbox{\dbl F}}_p^*$, $z_k$ were defined in lemma \ref{predvarit}, $c_{d(u,a^{p^{k-1}})-j(k-1), (p-1)j(k-1), p-1}$ is calculated in lemma \ref{(5)}, (i) and it is not equal to zero by lemma \ref{(5)}, (ii), and ${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{d(u,a^{p^{k-1}})}(a^{p^{k-1}})$ is calculated in lemma \ref{predvarit2}. By lemma \ref{predvarit2} we have ${}^{(z_{k-1},u)}_{-j(k-1)}\delta_{d(u,a^{p^{k-1}})} (a^{p^{k-1}})=-j(k-1){}^{(z,u)}\delta_{d(u,a^{p^{k-1}})}(a^{p^{k-1}})$. Combining all these calculation together and using induction, we get $J(k)=q_kJ(k-1)^pJ(1)=\tilde{q_k}J(1)^{p^{k-1}+p^{k-2}+\ldots +1}$ for $k\ge 1$, where $q_k\in {\mbox{\dbl F}}_p$. Therefore, there is the following filtration $$ \bar{F}\subset \ldots \subset \bar{F}(a^{p^{k+1}})J(k+1)\subset \bar{F}(a^{p^k})J(k)\subset \ldots \subset \bar{D}, $$ and for every $k\ge 1$ we have $\bar{F}(a^{p^k})\cdot J(k)\subset A'(k-1)$. So, $A(k)\cap A(k_1)=\{0\}$ if $k\ne k_1$. Now consider an element $b\in F$ such that $\bar{b}=a^{p^l}$ for some $l>0$. We assume $l$ is a minimal possible integer. It exists, because $D$ is a finite dimensional algebra over $F$. Let $b=u(a^{p^l})+b_1z+\ldots $, where $b_k\in u(\bar{D})$. Put $I:=\min \{w(zb_kz^{k-1}-b_kz^k)\}$ (we assume here that $b_0=u(a^{p^l})$). Note that $I<\infty$, since by lemma \ref{predvarit2} $j(l)<\infty$, i.e. ${}^{(z,u)}\delta_{j(l)}(a^{p^l})\ne 0$. Now we must have $$ zbz^{-1}=\sum_{k=0}^{\infty}zb_kz^{k-1}= b+\sum_r {}^{(z,u)}\delta_{j(r)}(b_{q_r})z^I+\ldots =b, $$ where $b_{q_r}\in \bar{F}(a^{p^r})$ and $b_{q_r}\notin \bar{F}(a^{p^{r+1}})$. So, $\sum_r {}^{(z,u)}\delta_{j(r)}(b_{q_r})=0$, but it is impossible, since $A(k)\cap A(k_1)=\{0\}$ if $k\ne k_1$, a contradiction. The theorem is proved. \\ $\Box$ {\bf Remark.} It would be interesting to know the answer on the following questions. i) Suppose $D$ is a division algebra as in the theorem \ref{posledn}. Does there exist a pair $(z,u)$ such that all nonzero maps ${}^{(z,u)}\delta_q$ satisfy the property $i(u)|q$? If it is true, there is a subalgebra $D'\subset D$ with $[D:D']<\infty$ and $D'$ has level 1 (see remark before lemma \ref{vtorinv}). So, we can reduce studying of $D$ to the algebra of level 1. ii) Is it true that $D$ is a good splittable algebra, i.e. cyclic? Probably, it is possible to apply our technique to give an answer to this question at least in the case of level 1.
{ "timestamp": "2005-03-28T16:54:44", "yymm": "0503", "arxiv_id": "math/0503637", "language": "en", "url": "https://arxiv.org/abs/math/0503637" }
\section{Introduction} For the free Bose gas with Dirichlet and Neumann boundary conditions, Bose-Einstein Condensation (BEC) is rigorously treated in \cite{LP,LW}. For the mean-field Bose gas with periodic boundary conditions BEC is rigorously proved and a detailed analysis of the thermodynamic limit is given in \cite{FV}. The proof is based on bounds on the correlation functions for equilibrium states, given in terms of the correlation inequalities \cite{FV2,FV3}. The subtle point in this proof is the analysis of the singularity around the zero-mode.\\ If one considers attractive boundary conditions instead of periodic boundary conditions, the problem changes drastically. The free Bose gas with attractive boundary conditions is extensively studied in \cite{R,LW}. Due to the gap in the one-dimensional one-body problem, one has Bose-Einstein Condensation in all dimensions $\nu\geq 1$. An important result is the fact that the condensation is a surface effect. In \cite{VVZ} it is computed that the condensate is localized at a logarithmic distance from the boundary.\\ The subject of this note is to proceed with the imperfect Bose gas with attractive boundary conditions, i.e.\ the free Bose gas with attractive boundary conditions plus a mean field term. The first problem that occurs is to express this mean field term in momentum space diagonalizing the kinetic energy (the free Bose gas part). As the spectrum of the latter one has two strictly negative eigenvalues, say $\epsilon_L(0)$ and $\epsilon_L(1)$, separated from the rest of the spectrum, then one can discuss the corresponding number operators $N_0$ and $N_1$ as being added to the total number operator in the interaction or not. In section \ref{interact_term} we argue why they should not be present. The argument is essentially based on the fact that we want a space homogeneous mean field term. The model is defined in \ref{hamiltonian}.\\ In section \ref{condensatie} we give a completely rigorous proof of the occurence of Bose-Einstein Condensation for the imperfect Bose gas with attractive boundary conditions. We perform all details only in dimension $\nu =1$. From dimension $\nu\geq 2$ on, the proof becomes technically more tedious. In particular, because of the fact that the condensate is located near the boundaries, for higher dimensions the thermodynamic limit for hypercubic boxes is not very suitable nor realistic and should in stead be taken with increasing absorbing balls. But we leave this extra exercise for a later occasion in which we consider the problem of the shape-dependence.\\ In the one-dimensional case we remark that the condensation is equally distributed over the two negative energy levels. The condensation is localized in the same area as for the free Bose gas with attractive boundary conditions, see \cite{R}. \section{The Model} \subsection{Attractive Boundary conditions} If one considers a free gas of bosons in an interval $\left[-L/2,L/2\right]$ of length $L$, then the energy levels are determined by the one-dimensional Schr\"odinger equation (with units $\frac{\hbar^2}{2m}=1$) \begin{equation*} -\Delta \phi = \epsilon_L \phi^L , \end{equation*} with boundary conditions: $$\left\{ \begin{array}{lll} \left(\displaystyle\frac{d\phi}{dx}-\sigma\phi\right)_{x=-L/2} & = & 0 ,\\ \left(\displaystyle\frac{d\phi}{dx}+\sigma\phi\right)_{x=L/2} & = & 0 , \end{array} \right.$$ where $\sigma<0$.\\ If one considers these attractive boundary conditions, the spectrum consists of two negative eigenvalues tending to the same limit $-\sigma^2$ (when $L\rightarrow \infty$) and an infinite number of positive eigenvalues (for $L|\sigma|>2$): $\epsilon_L(k)$ for $k=0,1,2,\ldots$, where $$\epsilon_L(0) < \epsilon_L(1) < 0 < \epsilon_L(2) < \epsilon_L(3) < \ldots ,$$ $$\epsilon_L(0)=-\sigma^2- O(\mathrm{e}^{-L|\sigma|}) ,$$ $$\epsilon_L(1)=-\sigma^2+O(\mathrm{e}^{-L|\sigma|}) ,$$ \begin{equation}\label{spect} k\geq 2:\ \ \left(\frac{(k-1)\pi}{L}\right)^2 < \epsilon_L(k) < \left(\frac{k\pi}{L}\right)^2 . \end{equation} The corresponding eigenfunctions $\{\phi_k^L\}_{k \in\mathbbm{N}}$ are given by \begin{eqnarray} \phi_0^L(x) & = & \sqrt{\frac{2}{L}}\left(1+\frac{\sinh(L|\sigma|)}{L|\sigma|}\right)^{-1/2} \cosh(-|\sigma|x) ,\nonumber\\ \phi_1^L(x) & = & \sqrt{\frac{2}{L}}\left(-1+\frac{\sinh(L|\sigma|)}{L|\sigma|}\right)^{-1/2} \sinh(-|\sigma|x) ,\nonumber\\ \phi_k^L(x)& = & \left\{ \begin{array}{ll} \sqrt{\displaystyle\frac{2}{L}}\left(1+\displaystyle\frac{\sin (\sqrt{\epsilon_L(k)}L)} {\sqrt{\epsilon_L(k)}L}\right)^{-1/2}\cos(\sqrt{\epsilon_L(k)}x) , \qquad & \mbox{for $k$ \ even} ,\\ \sqrt{\displaystyle\frac{2}{L}}\left(1-\displaystyle\frac{\sin (\sqrt{\epsilon_L(k)}L)} {\sqrt{\epsilon_L(k)}L}\right)^{-1/2}\sin(\sqrt{\epsilon_L(k)}x) , \qquad & \mbox{for $k$ odd} . \end{array}\right.\nonumber \end{eqnarray} \subsection{Hamiltonian}\label{hamiltonian} We consider a one-dimensional system of identical bosons on an interval $[-\frac{L}{2},\frac{L}{2}]\subset\mathbbm{R}$ with attractive boundary conditions. The model is specified by the local Hamiltonians $H_{L, MF}^{\sigma}$ on the boson Fock space $\mathcal{F}_{L, B}$: \begin{equation}\label{ham} H_{L, MF}^{\sigma} = T_{L}^{\sigma} + \frac{\l}{2}\frac{\tilde{N}_L^2}{L} \end{equation} where $T_{L}^{\sigma}$ is the kinetic energy operator with the $\epsilon_L(k)$ the eigenvalues (\ref{spect}) of the free Laplacian with attractive boundary conditions: \begin{equation*} T_{L}^{\sigma} = \sum_{k\in\mathbbm{N}}\epsilon_L(k) a_{k}^\ast a_{k} \end{equation*} The operators $a_k^\ast=a^\ast(\phi_L^k)$ and $a_k=a(\phi_L^k)$ are the Bose creation and annihilation operators with the testfunctions $\phi_L^k$ the above eigenfunctions of the free Laplacian with attractive boundary conditions. The total particle number operator is denoted by $N_L=\sum_{k\in\mathbbm{N}}N_k = \sum_{k\in\mathbbm{N}} a_{k}^\ast a_{k}$, and the particle number operator corresponding to the positive spectrum by $\tilde{N}_L$: \begin{equation*} \tilde{N}_{L} = \sum_{k=2}^{\infty}a_{k}^\ast a_{k} \end{equation*} We consider a positive coupling constant $\l \in \mathbbm{R}^+$ for the sake of thermodynamic stability, see \cite{FV}. \subsection{The Interaction Term}\label{interact_term} Remark that the interaction in the Hamiltonian $H_{L, MF}^{\sigma}$ is not of the usual form $\frac{\l}{2}\frac{N_L^2}{L}$, with \begin{eqnarray} N_L & = & \int_{-L/2}^{L/2} \, dx \, a^\ast(x)a(x)\label{number}\\ & = & \sum_{k\in\mathbbm{N}}a_k^\ast a_k,\nonumber \end{eqnarray} but of the form $\frac{\l}{2}\frac{\tilde{N}_L^2}{L}$, where \begin{equation*} \tilde{N}_L = \sum_{k=2}^\infty a_k^\ast a_k. \end{equation*} The reason for choosing $\tilde{N}_L$ in stead of $N_L$ is the breaking of the spatial translation invariance in the two terms with $k=0$ and $k=1$. Moreover the latter two terms yield also gauge symmetry breaking under the effect of space translations. Using the straightforward computation \begin{equation*} \cosh(|\sigma|(x+a)) = \cosh(|\sigma|x)\mathrm{e}^{-|\sigma|a}+\mathrm{e}^{|\sigma|x}\sinh(|\sigma|a) \end{equation*} for $a\in\mathbbm{R}$, we get (for large $L$): \begin{eqnarray*} \tau_a (a_0^\ast) & \approx & a_0^\ast \mathrm{e}^{-|\sigma|a}+(a_0^\ast - a_1)\sinh(|\sigma|a)\\ \tau_a (a_0) & \approx & a_0 \mathrm{e}^{-|\sigma|a}+(a_0 - a_1^\ast)\sinh(|\sigma|a) \end{eqnarray*} where $\tau_a$ is the translation automorphism over the distance $a\in\mathbbm{R}$. One gets a similar expression for $\tau_a (a_1^\sharp)$.\\ In particular, the terms $\tau_a (N_0)$ and $\tau_a (N_1)$ are not translation invariant and diverge exponentially for $|a|\rightarrow\infty$ as \begin{equation*} \tau_a (a_0^\ast a_0) \approx \mathrm{e}^{2|\sigma||a|}a_0^\ast a_0 \end{equation*} moreover, these terms break the gauge symmetry.\\ Remark also that on the other hand the particle number operator $\tilde{N}_L$ is a good local approximation of $N_L$ (\ref{number}) for all translation invariant states, such that the interaction term $\frac{\l}{2}\frac{\tilde{N}_L^2}{L}$ is the appropriate mean field term. \section{Bounds on the correlation function and condensation}\label{condensatie} Our aim is to find the equilibrium states of the system in the grand canonical ensemble. The equilibrium state $\omega_L$ at inverse temperature $\b$ is characterized by the following correlation inequality for all $L$, see \cite{FV2,FV3} \begin{equation}\label{corr_ineq} \b \omega_L(X^\ast [H_{L, MF}^{\sigma}-\mu_L N_L,X]) \geq \omega_L(X^\ast X)\ln \frac{\omega_L(X^\ast X)}{\omega_L(X X^\ast)} \end{equation} for all local observables $X$, where $\omega_{L}(\cdot)$ is the grand canonical equilibrium state at chemical potential $\mu_L$ and inverse temperature $\b$: \begin{equation*} \omega_{L}(X) = \frac{\Tr_{\mathcal{F}_{L, B}}X\exp\{-\b (H_{L, MF}^\sigma -\mu_L N_L)\}}{\Tr_{\mathcal{F}_{L, B}}\exp\{-\b H_{L, MF}^\sigma\}} \end{equation*} with $\mathcal{F}_{L, B}$ the boson Fock space over $\mathcal{L}^2([-L/2,L/2])$.\\ Concerning the thermodynamic limit ($L\rightarrow\infty$), we perform this limit keeping the total density $\rho$ constant. Therefore the chemical potential $\mu_L$ is now determined by the particle density $\rho$ and is the solution of the particle density equation: for each given density $\rho$ we have \begin{equation}\label{part_density} \rho = \frac{\omega_L(N_L)}{L} \end{equation} From the correlation inequality (\ref{corr_ineq}) follows immediately the inequality \begin{equation}\label{corr_ineq2} \omega_L\left(\big[X^\ast,[H_{L, MF}^\sigma -\mu_L N_L , X]\right]\big)\geq 0 . \end{equation} In this section we focus on the proof of the condensation for the model $H_{L, MF}^{\sigma}$ (\ref{ham}). In this proof we need bounds which we derive from the correlation inequality (\ref{corr_ineq}) for some specific observables $X$'s. Due to the special character of the spectrum (\ref{spect}), it is necessary to distinguish between products of creation and annihilation operators $a^\sharp_k$ in the $0$- or $1$-mode, and those in the $k$-mode (with $k\geq 2$).\\ The first Lemma is valid for all $k\in\mathbbm{N}$. \begin{lemma}\label{Nj_Nk} If $j \neq k_i$ $(i = 1,\ldots ,m)$, $k_i \neq k_{i'}$ for $i \neq i'$ and $m,n_i \in \mathbbm{N}_0$ for $i = 1,\ldots ,m$, then \begin{eqnarray} \lefteqn{\mathrm{e}^{\b (\epsilon_L(j)-\epsilon_L(k_1))}\omega_L\left(N_{j} (N_{k_1}+1)^{n_1} (N_{k_2})^{n_2} \ldots (N_{k_m})^{n_m}\right)}\nonumber\\ & = & \omega_L\left((N_{j}+1)(N_{k_1})^{n_1} (N_{k_2})^{n_2} \ldots (N_{k_m})^{n_m}\right)\label{lemma1} \end{eqnarray} \end{lemma} \textit{Proof:} The proof follows from the correlation inequality (\ref{corr_ineq}) by taking $X$ successively equal to \begin{equation*} a_{j}^\ast a_{k_1}\left((N_{k_1})^{n_1-1} (N_{k_2})^{n_2} \ldots (N_{k_m})^{n_m}\right)^{1/2} \end{equation*} and \begin{equation*} a_{k_1}^\ast a_{j}\left((N_{k_1})^{n_1-1} (N_{k_2})^{n_2} \ldots (N_{k_m})^{n_m}\right)^{1/2} \end{equation*} \hfill $\square$\\ For the chemical potential $\mu_L$, we find the same upperbound as in the case of the free Bose gas with attractive boundary conditions. \begin{lemma}\label{mu} For $k=0,1$, one has \begin{equation*} \mu_L \leq \epsilon_L(k) \end{equation*} and \begin{equation}\label{mu_expr} \mu = \lim_{L\rightarrow\infty}\mu_L \leq -\sigma^2 \end{equation} \end{lemma} \textit{Proof:} Take $X = a_{k}^\ast$ where $k=0,1$ in the inequality (\ref{corr_ineq2}).\\ The second inequality (\ref{mu_expr}) is obtained by taking the thermodynamic limit $L$ tending to infinity. \hfill $\square$ \begin{lemma}\label{Nk_B} For each $k=0,1$ and $\mu_{L}<-\sigma^2$, we have: \begin{equation} \omega_{L}(N_{k}) = \frac{1}{\mathrm{e}^{\b(\epsilon_L(k)-\mu_L)}-1} \end{equation} \end{lemma} \textit{Proof:} The result follows from the inequality (\ref{corr_ineq}) by taking $X$ successively equal to $a_{k}$ and $a_{k}^\ast$ with $k=0,1$. \hfill $\square$ \begin{lemma}\label{corr_ineq_Nk} For each $n\in\mathbbm{N}$ and $k\geq 2$, we have \begin{equation*} \b\omega_{L}\left(-\epsilon_{L}(k)N_{k}^{n+1}+\mu_{L} N_{k}^{n+1}-\l \frac{\tilde{N}_{L}}{L}N_{k}^{n+1} +\frac{\l}{2}\frac{N_{k}^{n+1}}{L}\right)\geq \omega_L(N_{k}^{n+1})\ln \frac{\omega_L(N_{k}^{n+1})}{\omega_L((N_{k}+1)^{n+1})} \end{equation*} \end{lemma} \textit{Proof:} The result follows from the correlation inequality (\ref{corr_ineq}) with $X=a_k N_k^{n/2}$. \hfill $\square$\\ \\ In order to prove condensation, we need a convenient upperbound for $\omega_{L}(N_{k})$ for all $k\in\mathbbm{N}$. This bound is derived in the following Lemma. \begin{lemma}\label{Nk} For each $k\geq 2$ we have \begin{equation*} \omega_{L}(N_{k}) \leq \frac{1}{\mathrm{e}^{c_{k}(L)}-1} \end{equation*} where \begin{equation}\label{ck} c_{k}(L) = \b\left(\epsilon_{L}(k)+\sigma^2-\frac{\l}{2L}-o(\mathrm{e}^{-L|\sigma|})\right) \end{equation} \end{lemma} \textit{Proof:} By Lemma \ref{corr_ineq_Nk} with $n=0$: \begin{equation*} \omega_{L}(N_{k})\ln\frac{\omega_{L}(N_{k})}{\omega_{L}(N_{k})+1} \leq \b\omega_L\left(-\epsilon_{L}(k)N_{k} + \mu_{L} N_{k} - \l\frac{\tilde{N}_{L}}{L}N_{k} + \frac{\l}{2}\frac{N_{k}}{L}\right) \end{equation*} From Lemma \ref{mu} we know that $\mu_L \leq -\sigma^2 - o(\mathrm{e}^{-L|\sigma|})$. It is also easy to see that $\omega_L(\tilde{N}_L N_k) \geq 0$. This leads to \begin{equation*} \omega_{L}(N_{k})\ln\frac{\omega_{L}(N_{k})}{\omega_{L}(N_{k})+1} \leq -\b\left(\epsilon_{L}(k) + \sigma^2 + o(\mathrm{e}^{-L|\sigma|}) + \frac{\l}{2L}\right)\omega_\L(N_k) \end{equation*} which gives us immediately the result. \hfill $\square$\\ \\ Using the results of the previous Lemma's, we are now ready to prove the existence of condensation in the two lowest energy levels. \begin{theorem}\label{condensation} Let $\rho_{cond}$ be equal to \begin{equation*} \rho_{cond} = \lim_{L\rightarrow\infty} \frac{1}{L}\omega_L\left(N_0 + N_1\right). \end{equation*} Then \begin{equation}\label{condensation_expr} \rho_{cond} \geq \rho - \frac{1}{\pi}\int_{0}^{\infty}\, dk \frac{1}{\mathrm{e}^{\b(k^2+\sigma^2)}-1} \end{equation} where $\rho_{cond}$ is the density of the condensate. The condensate density is localized in the $2$ lowest energy levels. \end{theorem} \textit{Proof:} From the definition of the particle density $\rho$ (\ref{part_density}), we have \begin{equation*} \frac{1}{L}\omega_L\left(N_0 + N_1\right) = \rho - \frac{1}{L} \sum_{k=2}^\infty\omega_{L}(N_k) \end{equation*} By using the estimate of Lemma \ref{Nk}, one gets \begin{equation*} \frac{1}{L}\omega_L\left(N_0 + N_1\right) \geq \rho - \frac{1}{L} \sum_{k=2}^\infty \frac{1}{\mathrm{e}^{c_{k}(L)}-1} \end{equation*} with the $c_k(L)$'s as in (\ref{ck}).\\ Taking the thermodynamic limit $L\rightarrow\infty$ gives us the result (\ref{condensation_expr}). \hfill $\square$\\ \\ Clearly (\ref{condensation_expr}) shows condensation. Indeed, remark that the integral is convergent for all $\sigma\neq 0$ and that it decreases for $\b$ increasing. Hence for $\rho$ large enough or for $\b$ large enough, it follows that the condensate density $\rho_{cond}$ is strictly positive.\\ Finally we derive a result about the type of condensation. We prove that the condensate density is realized in both the two lowest energy modes, with equal weight in the thermodynamic limit. \begin{theorem} \begin{itemize} \item[(i)]The condensate is equally distributed on the two lowest energy levels. \item[(ii)]From this, one can compute the asymptotics of the chemical potential $\mu_L$ for large $L$: \begin{equation}\label{mu_asymp} \mu_L = -\sigma^2 - \frac{2}{\b\rho_{cond} L} + o(L^{-1}) \end{equation} \end{itemize} \end{theorem} \textit{Proof:} \begin{itemize} \item[(i)] From Lemma \ref{Nj_Nk}, with $m=1$, $j=1$ and $k=0$, we get \begin{equation*} \left(\mathrm{e}^{\b(\epsilon_L(1)-\epsilon_L(0))}-1\right)\omega_L(N_1 N_0) + \mathrm{e}^{\b(\epsilon_L(1)-\epsilon_L(0))}\omega_L(N_1) = \omega_L(N_0) \end{equation*} Using the spectral properties \begin{eqnarray*} \epsilon_L(0) & = & -\sigma^2- O(\mathrm{e}^{-L|\sigma|})\\ \epsilon_L(1) & = & -\sigma^2+O(\mathrm{e}^{-L|\sigma|}) \end{eqnarray*} such that \begin{equation*} \epsilon_L(1)-\epsilon_L(0)\approx o(\mathrm{e}^{-L|\sigma|}), \end{equation*} then \begin{equation*} \frac{\omega_\L(N_0)}{L} = \frac{\omega_\L(N_1)}{L} + o(\mathrm{e}^{-L|\sigma|}). \end{equation*} Taking the thermodynamic limit $L\rightarrow\infty$, one gets (i). \item[(ii)] From Lemma \ref{Nk_B} and since we know that the condensate is equally distributed on the two lowest energy levels, we have \begin{equation*} \lim_{L\rightarrow\infty}\frac{\omega_L(N_{0})}{L} = \lim_{L\rightarrow\infty} \frac{1}{L}\frac{1}{\mathrm{e}^{\b(\epsilon_L(0)-\mu_L)}-1} = \frac{\rho_{cond}}{2} \end{equation*} Series expansion of this expression with respect to the quantities $\epsilon_L(0)-\mu_L$ and $\epsilon_L(0)=-\sigma^2- O(\mathrm{e}^{-L|\sigma|})$, gives us the asymptotics of $\mu_L$. \end{itemize} \hfill $\square$
{ "timestamp": "2005-03-29T13:17:36", "yymm": "0503", "arxiv_id": "math-ph/0503068", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503068" }
\section{Introduction} There are countless experiments which demonstrate the wave behaviour of light. Two typical experiments are the two-slit and Mach-Zehnder arrangements. That such experiments demonstrate the wave behaviour of light, even where the light is feeble\footnote{By feeble light we mean light of such low intensity that on average only one photon at a time is in the apparatus.} \cite{T09}, is not in dispute. What is questionable is the experimental evidence for the particle behaviour of light. To avoid later misunderstanding of the essential point of this article, it is necessary for me to make clear that I use the term `particle behaviour' to refer to the description prior to the final detected result but not to the character of the final detected result. This is a more restrictive usage than is usual in the literature where the term `particle behaviour' also encompasses the character of the final detected experimental result. I also use the term `particle behaviour' in two context dependent ways: In the context of Bohr's principle of complementarity I use the term `particle behaviour' to refer to the description of the experiment in terms of the complementary particle concept (understanding that according to Bohr the particle concept, along with other complementary concepts, is an abstraction to aid thought to which physical reality cannot be attached). In the context of the causal interpretation I take the term `particle behaviour' to be synonymous with `particle ontology'. Similar considerations apply to the term `wave behaviour', but the distinction here is not so crucial since a main point of this article is to demonstrate that a final detected result showing a particle character does not force a particle description or particle ontology prior to the final detected result. More recent and interesting experiments concerning particle-wave duality and complementarity have been suggested and subsequently performed. Ghose {\it et al} \cite{GHOSE91} proposed an experiment involving tunneling between two closely spaced prisms which has since been carried out by Mizobuchi {\it et al} \cite{MIZ92} (although the statistical results of the experiment have been questioned by \cite{UNNIK, GHOSE99, BRIDA04}). Later, Brida {\it et al} \cite{BRIDA04} realized an experiment suggested by Ghose \cite{GHOSE99} in which tunneling at a twin prism arrangement is replaced by birefringence. Also of interest is Afshar's experiment \cite{AFSHAR04}. All of these experiments use light and aim to disprove or generalize\footnote{Brida {\it et al} view the observation of simultaneous particle and wave behaviour as demonstrating a need to generalize complementarity in the sense of Wootters and Zurek \cite{WZ}and Greenberger and Yasin \cite{GY}. I have argued that the generalization in fact completely contradicts complementarity and is the antithesis of Bohr's teachings \cite{KPW}. See section 6 for further discussion of this point.} complementarity (whereas GRA's aim was to confirm complementarity) by claiming to have demonstrated particle and wave behaviour in the same experiment. In all of these experiments, the final detection result is attributed by the authors to which-path information and, therefore, to particle behaviour (according to the usual criteria accepted in the literature), but the experiments are so arranged that the light undergoes a process (tunneling in the case of Mizobuchi {\it et al}'s experiment, birefringence in Brida {\it et al}'s experiment, and interference in Afshar's experiment) which the authors claim necessarily represents wave behaviour. Hence, they claim to observe wave and particle behaviour in the same experiment. We do not agree with them for the same reasons that we do not agree with GRA's claim to have proved complementarity, a claim we will argue against in this article. Generally, we take the view that complementarity is so imprecise that it can neither be proved nor disproved. We will elaborate further on this in the rest of the article with regard to the GRA experiments, but we will also briefly describe and comment further on Mizobuchi {\it et al}'s, Brida {\it et al}'s and Afshar's experiments in section \ref{CGBAS}. We have chosen to focus on the GRA experiments in this article because they were the first to introduce a gating system for producing genuine single photon states and because their experiments lend themselves to illustrating important features of CIEM. Further, the detailed treatment of this experiment serves as a model that can be easily adapted to the later experiments, thereby providing arguments against the claims of observing simultaneous wave and particle behaviour in these experiments. The quantum eraser experiment of Kim {\it et al} \cite{KIM00} is a variant of the Wheeler delayed-choice idea \cite{WHR78, K05}. The use of particle-wave duality and complementarity in this experiment seems to imply that a measurement performed in the present effects the outcome of an earlier measurement. This now raises the further issue of the present effecting the past, which is surely unacceptable. We will also give a brief description and comment on this experiment in section \ref{CGBAS}. Experimental evidence for the particle behaviour of light is mainly of two forms: which-path experiments and the photoelectric effect (also the Compton effect). A closer look at each of these shows that neither unambiguously demonstrate particle behaviour. In the case of the photoelectric effect it is well known that a semiclassical description can be given in which the light is treated as a classical electromagnetic field and only the atom is treated quantum mechanically \cite{W26}. A weakness of this counter example is that semiclassical radiation theory is known not to be fully consistent with experiment and fails in those cases where light exhibits nonclassical properties (as in some experiments which involve second-order coherence). Further, it is not clear that a semiclassical model of the photoelectric effect can explain the experimental fact that the photon is absorbed in a time of the order of $10^{-9}\;\mathrm{s}$ (\cite{VW76}, p. 10). Indeed, it was just this feature of the photoelectric effect that seemed to require that a photon be a localized particle prior to absorption, and is perhaps the reason why the photoelectric effect is commonly regarded as evidence for the particle behaviour of light. A more convincing argument against the photoelectric effect as evidence of particle behaviour is the provision of a fully quantum mechanical model of the photoelectric effect based on the causal interpretation of the electromagnetic field (CIEM) \cite{K85, K87, K94}. In CIEM, light is modeled as a real vector field; there are no photon particles\footnote{From here on we will use the term `photon' very loosly to refer to a quantum of energy which may or may not be spread out over large regions of space with a value of $\hbar\omega$ for a Fock state or with a value an average around $\hbar\omega$ for a wave packet.}. The field has the property of being nonlocal, meaning that an interaction at one point in the field can change the field at points beyond $ct$. The CIEM model of the photoelectric effect is of the nonlocal absorption of a photon by a localized atom. The photon prior to absorption may be spread over large regions of space. The fact that the absorption is nonlocal explains the experimental result that the absorption of the photon takes place in a time of the order of $10^{-9}\;\mathrm{s}$. We are not forced to accept that the photon must be localized prior to absorption. We conclude that the photoelectric effect cannot be regarded as conclusive evidence for the particle behaviour of light. We note that the Compton effect, also commonly accepted as evidence for the particle behaviour of light, can also be modeled by CIEM (\cite{K94}, p. 343), so that this also cannot be taken as evidence for the particle behaviour of light. To be clear, we are not claiming that the final detected results of the photoelectric and the Compton effect do not have a particle character (they clearly do). What we claim is that a particle description prior to the final result, whether from the perspective of complementarity or from the perspective of an ontology, is not forced upon us. This is because the particle character of the experimental results can be explained in terms of a wave model. Let us now turn to which-path experiments. In a typical which-path experiment light has a choice of two paths. Determining which-path the light actually took is considered as proof of particle behaviour. As Bohr showed in response to Einstein's famous which-path two-slit experiment, if the path is determined with certainty, interference is lost \cite{BR59A}. Consider a which-path two-slit experiment in which we determine the path by closing one of the holes (obviously losing interference). Although crude, it is conceptually equivalent to Einstein's experiment. The point is, that even when we close the hole and are certain which-path the light took, this does not rule out a wave model. This argument holds even in more refined which-path two-slit experiments. We may conclude that in such experiments the which-path criteria for particle behaviour is somewhat arbitrary. There is an aspect of the two-slit experiment that seems to be universally overlooked and that we wish to draw attention to. Einstein's aim in his which-path two-slit experiment was to obtain the path of an individual photon and still retain an interference pattern, thereby experimentally detecting particle and wave behaviour in the same experiment\footnote{Actually, Einstein considered Bohr's principle of complementarity and quantum mechanics to be synonymous. By experimentally contradicting complementarity Einstein wanted to demonstrate that quantum mechanics is incomplete (\cite{JAM74}, p. 127). We have argued elsewhere that Bohr's principle of complementarity and quantum mechanics are not synonymous (\cite{K05}, p. 299).}. This is contrary to Bohr's principle of complementarity which requires mutually exclusive experimental arrangements for complementary concepts \cite{BR59A, JAM74, BR28}. As we have said, Bohr was able to show that a certain determination of the photon path would destroy the interference pattern. Bohr's response was almost universally accepted and complementarity was saved. But consider this: Forget path determination and consider a two-slit experiment in which an interference pattern is formed. This interference pattern is built up of a large number of individual photoelectric detections (or some similar process in a photographic emulsion). If the photoelectric effect is accepted as evidence of the particle behaviour of light, then is not particle and wave behaviour observed in the same experiment? We now turn to another which-path experiment which uses a beam-splitter. This will be our main focus in this article because we consider GRA's version of this experiment, which uses an atomic cascade and a gating system to produce a near ideal single photon state, as perhaps the best experimental attempt to demonstrate the particle behaviour of light \cite{G86, AG86}. In a wave model, light is split into two beams at the beam-splitter. In a particle model, each photon must choose one and only one path. Thus, using feeble light (one photon at a time) a particle model predicts perfect anticoincidence, whereas some coincidences are expected in a wave model. GRA therefore took perfect anticoincidence as the signature of particle behaviour. GRA quantified this feature in terms of the degree of second-order coherence. Semiclassical radiation theory predicts $g^{(2)}\geq 1$. As we shall see, quantum mechanical coherent or chaotic states give results in the classical regime. This is to be expected, as neither chaotic nor coherent light exhibits nonclassical behaviour. For number states on the other hand, perfect anticoincidence is expected, so that $g^{(2)}=0$. Photoelectric detectors are placed in each output arm of the beam-splitter. For a detection to take place there must be enough energy to ionize an atom in the detector. For classical light, and quantum mechanical chaotic or coherent light, there is always some probability that more than one photon is present after the beam-splitter however feeble the light, and this entails the possibility of coincidences. But, for a single photon state there is enough energy to ionize only a single atom in one and only one output arm of the beam-splitter, so that perfect anticoincidence is predicted. The novelty of the GRA experiments is the use of an atomic cascade and a gating system, which we describe below, in order to produce near ideal single photon states. Their results gave a value of $g^{(2)}$ much less than $1$ and confirmed the expected anticoincidence. GRA interpreted their results to be a conclusive demonstration of the particle behaviour of light. But, underlying the assertion that anticoincidence is a signature for particle behaviour is the assumption that the photoelectric detection process (or any other atomic absorption process) is local. This implies that the photon is a localized particle before absorption by the detecting atom. But, we saw above that the quantum theory does not rule out nonlocal absorption in the photoelectric effect (nor, more generally, in any atomic absorption process). In fact, no model of light as photon particles that is consistent with the quantum theory has ever been developed\footnote{Ghose {\it et al} have developed a particle interpretation of bosons \cite{GHOSE93, GHOSE96}, including the photon \cite{GHOSE01}, based on the Kemmer-Duffin formalism \cite{KEMMER39}. It is to be emphasized that this formalism, which allows an interpretation of bosons as particles, applies in the approximation that the energies are below the threshold for pair production. We maintain that the full theory does not allow a particle ontology. Since the particle ontology of the approximation stands in contradiction to the ontology of the full field theory (since particle and wave concepts are mutually exclusive), we maintain that the particle ontology of the approximate theory cannot have physical significance (Ghose {\it et al} do not address this issue). A further point is this: As Ghose himself points out, reference (\cite{GHOSE96}, p. 1448), for the boson particle interpretation to be consistent negative energy solutions must be interpreted as antiparticles moving backwards in time. In this case, an EPR correlated particle-antiparticle pair would exhibit the pathological feature of a nonlocal connection between the present and the past (we note that this particular criticism does not apply to the electromagnetic field). For more details on this and related approaches see reference \cite{WS2005}.}. On the other hand, CIEM models light as a nonlocal field. Atomic absorption processes, including the photoelectric effect, are modeled as the nonlocal absorption of a photon. CIEM has been shown to be fully consistent with the quantum theory \cite{K94}. Our main purpose in this article is to provide a model that explains perfect anticoincidence that does not treat photons as particles. By showing that anticoincidence experiments do not rule out a wave model we prove that GRA's experiment cannot be viewed as conclusive evidence for particle behaviour of light. The wave behaviour of light has been confirmed a countless number of times for chaotic or coherent sources. Following Einstein's 1905 explanation of the photoelectric effect \cite{E1905} in which the idea of photon particles was first invoked, the question was raised as to whether or not, in very low intensity experiments, single photons alone in the apparatus can produce interference. Numerous experiments using feeble light followed \cite{T09}. With a few exceptions the conclusion was reached that single photons can interfere with themselves. In such experiments the energy flux $\cal{E}$ is calculated and the number of photons per unit area per unit time is calculated using ${\cal E}/ \hbar \omega$. $\cal{E}$ is reduced to such low levels that it is more probable than not that only one photon is present in the apparatus at any one time. However, the probability that more than one photon is present remains, so that the single photon nature of these experiments can be questioned. By building a Mach-Zehnder interferometer around their which-path apparatus GRA were able to confirm that the near ideal single photon state produced the expected interference. Although no surprise, GRA's experiment is perhaps the first experiment to confirm the interference of single photons. The wave nature of light is not disputed and it is obvious how in CIEM interference is obtained given that light is modeled as a field (always). We will nevertheless outline the CIEM treatment of the Mach-Zehnder interferometer given in detail in reference \cite{K05}. CIEM is a hidden variable theory. There is a large literature on hidden variable theories and we direct the interested reader to the three articles cited in reference \cite{HVTHR}. Two of these, one old one new, are surveys of hidden variable theories and include a comprehensive list of references. We also refer the reader to two interesting Ph.D thesis in the area of hidden variable theories \cite{SC2005, WS2005}. In the next sections we describe GRA's two experiments focusing on theoretical derivations, and then go on to give the CIEM model of these experiments, focusing on the which-path experiment. \section{The GRA experiments} The following description of the GRA experiments is based mainly on reference \cite{G86}. The experiments use the radiative cascade of calcium $4p^2\;^1S_0\rightarrow4s4p\;^1P_1\rightarrow4s^2\;^1S_0$ described in reference \cite{AGR81}. The first cascade to the intermediate state yields a photon $\nu_1$ of wavelength $551.3\;\mathrm{nm}$. The intermediate state, with lifetime $\tau=4.7\;\mathrm{ns}$, decays according to the usual atomic decay law for the lifetime of a state (\cite{BJ89}, p. 538): \begin{equation} P(t)=1-e^{-t/\tau}, \label{DL} \end{equation} where $P(t)$ is the probability of decay in time $t$. The second cascade photon $\nu_2$ has wavelength $422.7\;\mathrm{nm}$. The $\nu_2$ photon, according to the decay law, is emitted with near certainty within the time $\omega=2\tau=9.8\;\mathrm{ns}$ of emission of the first $\nu_1$ photon. The number of $\nu_1$ photons per second, $N_1$, is counted by photomultiplier $PM_1$, and each $\nu_1$ photon triggers a gate of duration $\omega$. Because the probability of decay within gate $\omega$ is high, there is a high probability that the $\nu_2$ partner of $\nu_1$ enters the beam-splitter. For low count rates we can be nearly certain that there is only one $\nu_2$ photon in the beam-splitter arrangement within the gate time $\omega$. In this way a near ideal single photon state is produced. \begin{figure}[tb] \unitlength=1in \begin{picture}(6,3) \put(.3,0){\scalebox{3}{\includegraphics{GRAfig1new.eps}}} \put(1.7, .2){Figure 1. GRA's which-path experiment} \end{picture} \end{figure} \section{GRA's which-path experiment} Refer to figure 1. The photomultipliers $PM_t$ and $PM_r$ count the number of transmitted and reflected $\nu_2$ photons per second, and photomultiplier $PM_c$ counts the number of coincidences per second. These count rates are given by $N_t$, $N_r$ and $N_c$ respectively. The counts are taken over a large number of gates with a total run time $T$ of about 5 hours. The probabilities for single and coincidence counts are given by \begin{equation} p_t=\frac{N_t}{N_1},\;\;\;\;\;\;\;\;p_r=\frac{N_r}{N_1},\;\;\;\;\;\;\;\;p_c=\frac{N_c}{N_1}. \end{equation} The classical and quantum mechanical predictions for the coincidence counts are very different. In their experiment, GRA measured the quantity $\alpha$, which they defined as \cite{G86} \begin{equation} \alpha=\frac{\mbox{\it \small COINCIDENCE PROBABILITY}}{\mbox{\it \small ACCIDENTAL COINCIDENCE PROBABILITY}}=\frac{p_c}{p_t p_r}=\frac{N_1N_c}{N_t N_r}. \label{alpha} \end{equation} Both classically and quantum mechanically, the quantity $\alpha$ is a special case of the degree of second-order coherence. Classically, $g^{(2)}_c$ is defined by (\cite{L73}, p. 111) \begin{equation} g^{(2)}_c(\mbox{{ \boldmath{$\mit r$}}}_1t_1,\mbox{{ \boldmath{$\mit r$}}}_2 t_2; \mbox{{ \boldmath{$\mit r$}}}_2 t_2, \mbox{{ \boldmath{$\mit r$}}}_1t_1)=\frac{\langle \mbox{\boldmath $E$}^*(\mbox{{ \boldmath{$\mit r$}}}_1t_1)\mbox{\boldmath $E$}^*(\mbox{{ \boldmath{$\mit r$}}}_2t_2)\mbox{\boldmath $E$}(\mbox{{ \boldmath{$\mit r$}}}_2t_2)\mbox{\boldmath $E$}(\mbox{{ \boldmath{$\mit r$}}}_1t_1)\rangle}{\langle|\mbox{\boldmath $E$}(\mbox{{ \boldmath{$\mit r$}}}_1t_1)|^2 \rangle \langle|\mbox{\boldmath $E$}(\mbox{{ \boldmath{$\mit r$}}}_2t_2)|^2\rangle}, \end{equation} where $\mbox{\boldmath $E$}$ is the electric field vector. For $\mbox{{ \boldmath{$\mit r$}}}_1=\mbox{{ \boldmath{$\mit r$}}}_2$ and $t_1=t_2$, $g^{(2)}_c$ reduces to \begin{equation} g^{(2)}_c=\frac{\langle(\mbox{\boldmath $E$}^*\mbox{\boldmath $E$})^2\rangle}{\langle\mbox{\boldmath $E$}^*\mbox{\boldmath $E$}\rangle \langle\mbox{\boldmath $E$}^*\mbox{\boldmath $E$}\rangle}=\frac{\langle I^2\rangle}{\langle I \rangle^2}, \label{DSOC} \end{equation} where $I$ is the intensity. We will see in the next subsection that $\alpha=g^{(2)}_c$. Similar definitions apply in quantum mechanics (\cite{L73}, p. 219): \begin{equation} g^{(2)}(\mbox{{ \boldmath{$\mit r$}}}_1t_1,\mbox{{ \boldmath{$\mit r$}}}_2 t_2; \mbox{{ \boldmath{$\mit r$}}}_2 t_2, \mbox{{ \boldmath{$\mit r$}}}_1t_1)=\frac{\langle \mbox{\boldmath $\hat E$}^-(\mbox{{ \boldmath{$\mit r$}}}_1t_1) \mbox{\boldmath $\hat E$}^-(\mbox{{ \boldmath{$\mit r$}}}_2t_2)\mbox{\boldmath $\hat E$}^+(\mbox{{ \boldmath{$\mit r$}}}_2t_2)\mbox{\boldmath $\hat E$}^+(\mbox{{ \boldmath{$\mit r$}}}_1t_1)\rangle}{\langle \mbox{\boldmath $\hat E$}^-(\mbox{{ \boldmath{$\mit r$}}}_1t_1)\mbox{\boldmath $\hat E$}^+(\mbox{{ \boldmath{$\mit r$}}}_1t_1) \rangle \langle\mbox{\boldmath $\hat E$}^-(\mbox{{ \boldmath{$\mit r$}}}_2t_2)\mbox{\boldmath $\hat E$}^+(\mbox{{ \boldmath{$\mit r$}}}_2t_2) \rangle}, \label{G2} \end{equation} where the $\mbox{\boldmath $\hat E$}$'s are quantum mechanical operators defined by \begin{equation} \mbox{\boldmath $\hat E$}^+(\mbox{{ \boldmath{$\mit r$}}} t)=\frac{i}{V^{\frac{1}{2}}}\sum_{k\mu}\sqrt{\frac{\hbar k c}{2}}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\hat{a}_{k\mu} e^{i(\mbox{{\scriptsize\boldmath{$k$}}}.\mbox{{\scriptsize\boldmath{$x$}}}-\omega_k t)}, \;\;\; \mbox{\boldmath $\hat E$}^-(\mbox{{ \boldmath{$\mit r$}}} t) =-\frac{i}{V^{\frac{1}{2}}}\sum_{k\mu}\sqrt{\frac{\hbar k c}{2}}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\hat{a}_{k\mu}^{\dagger} e^{-i(\mbox{{\scriptsize\boldmath{$k$}}}.\mbox{{\scriptsize\boldmath{$x$}}}-\omega_k t)}. \label{EPEM} \end{equation} By substituting eq. (\ref{EPEM}) into eq. (\ref{G2}) with $\mbox{{ \boldmath{$\mit r$}}}_1=\mbox{{ \boldmath{$\mit r$}}}_2$ and $t_1=t_2$ and considering only a single mode and a single polarization direction, eq. (\ref{G2}) reduces to \begin{equation} g^{(2)}=\frac{\langle a_{2}^{\dagger}a_2 a_{1}^{\dagger}a_1\rangle}{\langle a_{1}^{\dagger}a_1\rangle \langle a_{2}^{\dagger}a_2\rangle}.\label{G2qm} \end{equation} For a single mode and single polarization direction, the quantum mechanical operator for the magnitude of the intensity (\cite{L73}, p. 184; \cite{K05}, p. 304) reduces to \begin{equation} \hat{I}_1=\frac{\hbar k c^2 }{V}a^{\dagger}_1a_1. \end{equation} Multiplying the numerator and the denominator of eq. (\ref{G2qm}) by $(\hbar k c^2/V)^2$, we can write $g^{(2)}$ in terms of the expectation value of the intensity operator: \begin{equation} g^{(2)}=\frac{\langle I_1 I_2\rangle}{\langle I_1 \rangle \langle I_2\rangle}.\label{G2qmI} \end{equation} Again, we will see in the next subsection that this is equivalent to GRA's $\alpha$. In the following subsections we calculate the classical prediction for $g^{(2)}$ using semiclassical radiation theory and compare this with the quantum mechanical predictions for $g^{(2)}$ for a number state, a coherent state, and a chaotic state. \subsection{$g_c^{(2)}$ for a classical field} We now calculate the classical prediction for the various probabilities. The intensity of the $n^{th}$ gate is given by the time average of the instantaneous intensity $I(t)$: \begin{equation} i_n=\frac{1}{\omega}\int^{t_n+\omega}_{t_n} I(t)\;dt. \end{equation} Although the electromagnetic field is treated classically, the photoelectric detection is treated quantum mechanically. This semiclassical radiation theory gives the probability for a detection as proportional to the intensity and to time (\cite{L73}, p. 183 and p. 185; \cite{M76} p. 31 and p. 40) (as is the case quantum mechanically). The probabilities for singles counts during the $n^{th}$ gate are, therefore, \begin{equation} p_{tn}=\alpha_t i_n\omega,\;\;\;\;\;\;\;\;\;\;\;p_{rn}=\alpha_r i_n\omega, \end{equation} where $\alpha_t$ and $\alpha_r$ are the global detection efficiencies. The intensity averaged over all the gates is \begin{equation} \langle i_n \rangle =\frac{1}{N_1 T}\sum^{N_1 T}_{n=1} i_n, \end{equation} where $N_1T$ is the total number of counts in $PM_1$, which is equal to the total number of gates. So, the overall probability for singles counts becomes \begin{equation} p_t=\alpha_t\omega\langle i_n\rangle,\;\;\;\;\;\;\;\;\;\;p_r=\alpha_r\omega\langle i_n \rangle.\label{PtPr} \end{equation} During a single gate, the probability of a detection in one arm is statistically independent of detection in the other arm. Therefore, the probability of a coincidence count during a single gate is given as the product of the probabilities of detection in each arm: \begin{equation} p_{cn}=\alpha_t\alpha_r\omega^2 i_n^2. \end{equation} The probability of a coincidence count averaged over all the gates becomes \begin{equation} p_c=\alpha_t\alpha_r\omega^2\langle i_n^2\rangle.\label{Pc} \end{equation} If the coincidences are purely accidental, then the probabilities $p_t$ and $p_r$ over the ensemble of all gates are statistically independent, so that the accidental coincidence probability is given by the product \begin{equation} p_t p_r=\alpha_t\alpha_r\omega^2\langle i_n\rangle^2. \label{Ptr} \end{equation} This represents the minimum classical probability of coincidence. These averages satisfy the inequality (\cite{BS73}, p. 185, inequality no. 4) \begin{equation} \langle i_n^2 \rangle \geq \langle i_n \rangle^2, \label{PcPtPr} \end{equation} from which it follows, by using eq.'s (\ref{Pc}) and (\ref{Ptr}), that \begin{equation} p_c \geq p_t p_r. \end{equation} In terms of $\alpha$, eq. (\ref{alpha}), we can also write the inequality (\ref{PcPtPr}) as \begin{equation} \alpha\geq 1. \label{G2c} \end{equation} Substituting eqs. (\ref{Pc}) and (\ref{Ptr}) into eq. (\ref{alpha}) gives \begin{equation} \alpha=\frac{\langle i_n^2 \rangle}{\langle i_n \rangle^2}, \end{equation} which is equal to the classical second-order coherence function $g_c^{(2)}$ given in eq. (\ref{DSOC}). \subsection{Quantum mechanical $g^{(2)}$ for a number state, a coherent state and a chaotic state} In quantum mechanics, the same reasoning as for the classical case leads to the same expressions for the probabilities $p_t$, $p_r$ and $p_c$, and for $\alpha$. The difference is that the classical averages of the intensities are replaced by quantum mechanical expectation values of the intensity operator. Thus \begin{equation} \alpha=\frac{p_c}{p_t p_r}=\frac{\alpha_t\alpha_r\omega^2\langle I_{\alpha} I_{\beta} \rangle}{\alpha_t\omega\langle I_{\alpha}\rangle\alpha_r\omega\langle I_{\beta}\rangle}=\frac{\langle I_{\alpha} I_{\beta} \rangle}{\langle I_{\alpha}\rangle\langle I_{\beta}\rangle}=\frac{\langle b^{\dag}_{\alpha} b_{\alpha} b^{\dag}_{\beta} b_{\beta} \rangle}{\langle b^{\dag}_{\alpha} b_{\alpha}\rangle\langle b^{\dag}_{\beta} b_{\beta}\rangle}. \label{alphaqm} \end{equation} The subscripts $\alpha$ and $\beta$ refer to the horizontal and vertical beams that emerge after the first beam-splitter. We see that $\alpha$ is equal to $g^{(2)}$, eq. (\ref{G2qm}) or eq. (\ref{G2qmI}), in the quantum case also. To calculate $g^{(2)}$ we first consider the theoretical treatment of a single beam-splitter. By now a two input approach to the beam-splitter is almost universally accepted even when one of the inputs is the vacuum\footnote{In passing, we mention that Caves \cite{C80} uses a two input approach in connection with the search for gravitational waves using a Michelson interferometer. He suggests, as one of two possible explanations, that vacuum fluctuations due to a vacuum input are responsible for the `standard quantum limit' which places a limit on the accuracy of any measurement of the position of a free mass.} (e.g. \cite{OHM87}), but some workers still use a single input (\cite{L73}, p. 222\footnote{Here the beam-splitter is described as part of the Hanbury-Brown and Twiss experiment.}; \cite{SZ97}, p. 494\footnote{Here the beam-splitter is used as part of an atomic interferometer.}). The two input approach leads to an elegant mathematical description of the action of a beam-splitter in terms of a unitary $2\times 2$ transformation matrix which has the form of a rotation matrix \cite{CST}. Here we will use a use a single input approach since this greatly simplifies the mathematical treatment of the GRA experiments in terms of CIEM, and since it gives the same results as the two input approach for the quantities we are interested in (expectation values of the number operator, coincidence counts, and interference terms). Further, both approaches lead to essentially the same physical model of the GRA experiments in terms of CIEM. \begin{figure}[tb] \unitlength=1in \begin{picture}(6, 2.6) \put(.8,-0.1){\scalebox{3}{\includegraphics{GRAfig2new.eps}}} \put(1.5, 0){Figure 2. Input and output destruction operators.\label{INOUTBS}} \end{picture} \end{figure} The single input and two output annihilation and creation operators are related as follows: \begin{equation} a=t^{*}_{{\alpha}{\alpha}}b_{\alpha}+r^{*}_{{\alpha}{\beta}}b_{\beta},\;\;\;\;\;\;\;\;\;\;\;a^{\dagger}_{\alpha}=t_{{\alpha}{\alpha}}b^{\dagger}_{\alpha}+r_{{\alpha}{\beta}}b^{\dagger}_{\beta}. \label{cao} \end{equation} The $b$'s satisfy the usual commutation relation $[b_{\alpha}, b^{\dagger}_{\alpha}]=[b_{\beta}, b^{\dagger}_{\beta}]=1$ while any combination of $b_{\alpha}$ and $b_{\beta}$ or their conjugates commute. To preserve the commutator $[a, a^{\dagger}]=1$, we must have \begin{equation} |t_{{\alpha}{\alpha}}|^2+|r_{{\alpha}{\beta}}|^2=t^2+r^2=1, \label{TR} \end{equation} with $|t_{{\alpha}{\alpha}}|^2=t^2$ and $|r_{{\alpha}{\beta}}|^2=r^2$. Using eq.'s (\ref{cao}) and (\ref{TR}) we may proceed to calculate $g^{(2)}$ for various quantum states. We begin with the number state $| n \rangle$, \begin{equation} | n \rangle=\frac{(a^{\dagger}_{\alpha})^n}{(n!)^{\frac{1}{2}}}| 0 \rangle=\frac{(t_{{\alpha}{\alpha}}b^{\dagger}_{\alpha}+r_{{\alpha}{\beta}}b^{\dagger}_{\beta})^n}{(n!)^{\frac{1}{2}}}| 0 \rangle. \end{equation} Use of the binomial theorem to expand the brackets gives \begin{eqnarray} | n \rangle&=&\frac{1}{(n!)^{\frac{1}{2}}}\left[ \left( \begin{array}{c}n\\0 \end{array}\right) (t_{{\alpha}{\alpha}}b^{\dagger}_{\alpha})^n +\left( \begin{array}{c}n\\1 \end{array}\right) (t_{{\alpha}{\alpha}} b^{\dagger}_{\alpha})^{(n-1)} (r_{{\alpha}{\beta}}b^{\dagger}_{\beta})^1\right. \nonumber\\ & &+ \left( \begin{array}{c}n\\2 \end{array}\right) (t_{{\alpha}{\alpha}} b^{\dagger}_{\alpha})^{(n-2)} (r_{{\alpha}{\beta}}b^{\dagger}_{\beta})^2+...... +\left( \begin{array}{c}n\\n-1 \end{array}\right) (t_{{\alpha}{\alpha}} b^{\dagger}_{\alpha})^1 (r_{{\alpha}{\beta}}b^{\dagger}_{\beta})^{(n-1)}\nonumber\\ &&\left. +\left( \begin{array}{c}n\\n \end{array}\right) (r_{{\alpha}{\beta}} b^{\dagger}_{\beta})^n \right]| 0 \rangle. \end{eqnarray} With this expression for $|n\rangle$ we can evaluate the expectation value for the number of photons in the horizontal arm, $\langle n| b^{\dagger}_{\alpha} b_{\alpha} |n\rangle$, by multiplying out the brackets, noting that cross-terms are zero, and evaluating the action of the number operator on the various number states. After a number of rearrangement steps we arrive at \begin{eqnarray} \langle b^{\dagger}_{\alpha} b_{\alpha}\rangle=\langle n| b^{\dagger}_{\alpha} b_{\alpha} |n\rangle&=&nt^2\left[ t^{2(n-1)}+ t^{2(n-2)}r^2 \frac{(n-1)!}{(n-2)!}+ t^{2(n-3)}r^4\frac{(n-1)!}{(n-3)!2!}\right.\nonumber\\ &&\left.+t^{2(n-4)}r^6\frac{(n-1)!}{(n-4)!3!}+......+r^{2(n-1)}\right]. \end{eqnarray} We recognize the series in the square brackets as the binomial expansion for $(t^2+r^2)^{n-1}=1$, and we get \begin{equation} \langle b^{\dagger}_{\alpha} b_{\alpha} \rangle=nt^2. \label{bhbh} \end{equation} By the same procedure as above we also get the expectation value for the number of photons in the vertical beam, \begin{equation} \langle b^{\dagger}_{\beta} b_{\beta}\rangle=\langle n| b^{\dagger}_{\beta} b_{\beta} |n\rangle=nr^2, \label{bVbV} \end{equation} and the expectation value for the number of coincidences, \begin{equation} \langle b^{\dagger}_{\alpha} b_{\alpha}b^{\dagger}_{\beta} b_{\beta} \rangle=\langle n| b^{\dagger}_{\alpha} b_{\alpha}b^{\dagger}_{\beta} b_{\beta} |n\rangle=n(n-1)r^2t^2.\label{bHbHbVbV} \end{equation} Substituting the above expectation values into eq. (\ref{G2qm}) gives the second-order coherence function for a number state, \begin{equation} g^{(2)}=\frac{n(n-1)r^2t^2}{nt^2nr^2}=\frac{(n-1)}{n},\;\;\;\;\;\;n\geq 2. \end{equation} For $n=0,1$ $g^{(2)}$=0. We see that a single photon input shows perfect anticorrelation, contrary to the classical result for $g^{(2)}_c$, eq. (\ref{G2c}). Next we consider the coherent state \begin{equation} |\alpha\rangle=e^{-|\alpha|^2/2}\sum_n \frac{\alpha^n}{(n!)^{\frac{1}{2}}}|n\rangle. \end{equation} The expectation value in the horizontal arm is \begin{equation} \langle b^{\dagger}_{\alpha} b_{\alpha}\rangle=\langle\alpha| b^{\dagger}_{\alpha}b_{\alpha}|\alpha\rangle=e^{-|\alpha|^2}\sum_{n=0} \frac{|\alpha|^{2n}}{n!}\langle n| b^{\dagger}_{\alpha}b_{\alpha}|n\rangle +e^{-|\alpha|^2}\sum_{n'}\sum_{\stackrel{\mbox{\scriptsize$n$}}{\!\!\!\!\!\!\!\!\!\!\!n\neq n'}} \frac{(\alpha^{*})^{n'}}{(n'!)^{\frac{1}{2}}}\frac{\alpha^{n}}{(n!)^{\frac{1}{2}}}\langle n'| b^{\dagger}_{\alpha}b_{\alpha}|n\rangle. \end{equation} The second term consisting of cross terms is zero. After substituting eq. (\ref{bhbh}) into the above, we get \begin{equation} \langle b^{\dagger}_{\alpha}b_{\alpha}\rangle=t^2 e^{-|\alpha|^2}\sum_{n=0} \frac{|\alpha|^{2n}}{n!}n=t^2 e^{-|\alpha|^2}|\alpha|^2\sum_{n=0} \frac{|\alpha|^{2n}}{n!}=t^2 e^{-|\alpha|^2}|\alpha|^2e^{|\alpha|^2} =t^2 |\alpha|^2. \end{equation} In a similar way, we calculate the expectation value of the number operator in the vertical beam to be \begin{equation} \langle b^{\dagger}_{\beta} b_{\beta} \rangle=\langle\alpha| b^{\dagger}_{\beta}b_{\beta}|\alpha\rangle=r^2 |\alpha|^2, \end{equation} and the expectation value for coincidence counts to be \begin{equation} \langle b^{\dagger}_{\alpha} b_{\alpha}b^{\dagger}_{\beta} b_{\beta} \rangle=\langle\alpha| b^{\dagger}_{\alpha}b_{\alpha}b^{\dagger}_{\beta}b_{\beta}|\alpha\rangle=t^2r^2 |\alpha|^4. \end{equation} Substituting the above expectation values into eq. (\ref{G2qm}) gives the second-order coherence function for a coherent state as \begin{equation} g^{(2)}=\frac{t^2r^2 |\alpha|^4}{t^2 |\alpha|^2r^2 |\alpha|^2}=1. \end{equation} This corresponds to the minimum classical value for $g^{(2)}$ so that measurement of the degree of second order coherence cannot distinguish between classical and coherent light. Lastly, we consider chaotic light. In quantum mechanics, chaotic light is a mixture of number states and is represented by the density operator (\cite{L73}, p. 158) \begin{equation} \rho=\sum_n P_n|n\rangle \langle n|. \end{equation} For light in thermal equilibrium, let $P_n$ be the probability of occurance of a number state $|n\rangle$ with energy $E_n=n\hbar\omega$. The probability $P_n$ is given by the Boltzmann distribution law applied to discrete quantum states (\cite{L73}, p. 8), \begin{equation} P_n=\frac{e^{-n\hbar\omega/kT}}{\sum^{\infty}_{n=0}e^{-n\hbar\omega/kT}}=(1-e^{-\hbar\omega/kT})\sum_n e^{-n\hbar\omega/kT}, \end{equation} where $k$ is Boltzmann's constant, and $T$ is the temperature in degrees Kelvin. The expectation value of the horizontal beam number operator is \begin{eqnarray} \langle b^{\dagger}_{\alpha} b_{\alpha} \rangle&=&{\mathrm Tr}(\rho b^{\dag}_{\alpha} b_{\alpha})=\sum_{n'}\langle n' |\rho b^{\dagger}_{\alpha} b_{\alpha}| n' \rangle=\sum_{n'} \sum_n(1-U)U^n \langle n'| n \rangle\langle n|b^{\dagger}_{\alpha} b_{\alpha}| n' \rangle \nonumber\\ &= &(1-U)\sum_n U^n \langle n| b^{\dagger}_{\alpha} b_{\alpha}| n \rangle, \end{eqnarray} with $U=\exp(-\hbar\omega/kT)$. Substituting the expectation value (\ref{bhbh}), and rearranging gives \begin{equation} \langle b^{\dagger}_{\alpha} b_{\alpha}\rangle=t^2 \frac{U}{1-U}. \end{equation} Using the other expectation values for the number state as above, we easily get the results \begin{equation} \langle b^{\dagger}_{\beta} b_{\beta} \rangle=r^2\frac{U}{1-U},\;\;\;\;\;\;\langle b^{\dagger}_{\alpha} b_{\alpha}b^{\dagger}_{\beta} b_{\beta} \rangle=t^2r^2 \frac{2U^2}{(1-U)^2}. \end{equation} Substituting the above into eq. (\ref{G2qm}) gives the degree of second-order coherence for a chaotic state \begin{equation} g^{(2)}=2 \end{equation} Like the result with the coherent state this value lies in the classical range. \subsection{Comparison of theoretical and experimental results} GRA's arrangement, figure 1, gives the degree of second-order coherence $g^{(2)}$ directly by measurement of $N_t$, $N_r$ and $N_c$ and use of eq. (\ref{alpha}). A value of $g^{(2)}\geq 1$ would agree with classical mechanics while a zero value would confirm quantum mechanics. In practice, experimental error prevents an exact zero value. Therefore, before comparing experimental and theoretical results, we first derive, following GRA \cite{G86}, a practical quantum mechanical prediction. \begin{figure}[tb] \unitlength=1in \begin{picture}(6,3) \put(0.7,0.5){\scalebox{2.5}{\includegraphics{GRAfig3.eps}}} \put(1.1, .1){Figure 3. Plot of the function $g^{(2)}(N\omega)$ with $f(w)=0.9$. \label{G2qmexp}} \end{picture} \end{figure} Let $N$ be the number of decays per second in the window of photomultiplier $PM_1$ of efficiency $\epsilon_1$. Then, $N_1=\epsilon_1 N$ is the number of $\nu_1$ photons detected per second by $PM_1$. From the atomic decay law (\ref{DL}), the probability $P_2$ of a $\nu_2$ photon partner of a $\nu_1$ photon entering the beam-splitter during a gate $\omega$ triggered by $\nu_1$ is $1-\exp(-\omega/\tau)$. Because of the angular correlation between $\nu_1$ and $\nu_2$, the probability $P_2$ is increased by a factor $a$ slightly greater than $1$ \cite{F73}. This probability is denoted by $f(\omega)=a[1-\exp(-\omega/\tau)]$, and is a number close to $1$ in GRA's experiment.The probability $P_2$ is also increased by accidental $\nu_2$'s. These are $\nu_2$ photons that enter the beam-splitter whose $\nu_1$ partners do not trigger a gate $\omega$. Once a $\nu_1$ photon has triggered a gate, the $\nu_1$ photons resulting from $N\omega$ decays during the gate $\omega$ cannot trigger another decay. Hence, their $N\omega$ $\nu_2$ partners are the accidental $\nu_2$ photons. Since $N\omega$ is the number of accidental $\nu_2$'s entering the beam-splitter during gate $\omega$, then $N_1N\omega$ is the number of accidental $\nu_2$'s entering the beam-splitter per second. The probability of an accidental $\nu_2$ photon entering the beam-splitter is therefore $N_1N\omega/N_1=N\omega$. Thus, \begin{equation} P_2=f(\omega)+N\omega=\frac{N_2}{N_1}, \end{equation} where \begin{equation} N_2=N_1[f(\omega)+N\omega] \end{equation} is the number of $\nu_2$ photons that enter the beam-splitter per second. Now, define $\epsilon_t$ and $\epsilon_r$ to be the efficiencies of $PM_t$ and $PM_r$, respectively. These efficiencies include the reflection and transmission coefficients, the collection solid angle, and the detector efficiency. The number $N_t$ of $v_2$ photons transmitted is $N_t=\epsilon_tN_2$, while the number reflected is $\epsilon_r N_2$. Then, the probabilities of detecting a transmitted $v_2$ photon in $PM_t$ and a reflected $v_2$ in $PM_r$ are \begin{equation} p_t=\frac{N_t}{N_1}=\frac{\epsilon_tN_2}{N_1} =\epsilon_t\left[f(\omega)+N\omega\right], \;\;\;\;\;\;\;\;\;\;\\ p_r=\frac{N_r}{N_1}=\frac{\epsilon_r N_2}{N_1} =\epsilon_r\left[ f(\omega)+N\omega\right]. \end{equation} Since $p_t$ and $p_r$ are statistically independent classically, the probability of a coincidence count becomes \begin{equation} p_c=p_tp_r=\epsilon_t\epsilon_r\left[ f(\omega)+N\omega\right]^2 =\epsilon_t\epsilon_r\left[ f(\omega)^2+2N\omega f(\omega) +N^2\omega^2\right].\label{PC} \end{equation} The term $f(\omega)^2$ suggests a repeated detection of the same photon. Since this is not possible, $f(\omega)^2$ is set equal to zero. Thus, substituting $f(\omega)^2=0$ into eq. (\ref{PC}) gives the quantum mechanical experimental expression for $p_c$. Substituting $p_t$, $p_r$ and $p_c$ into eq. (\ref{alpha}) gives: \begin{equation} g^{(2)}(N\omega)=\frac{2N\omega f(\omega) + N^2\omega^2}{[f(\omega)+N\omega]^2}. \end{equation} A plot of this function is given in figure 3. It is noticeable that as the erroneous $N\omega$ $\nu_2$ photon count increases compared to $f(\omega)$ the value of $g^{(2)}$ approaches the classical minimum value. GRA's experimental results closely agree with the plot of figure 3, and therefore confirm the quantum mechanical anticorrelation of the two beams. \section{GRA's Interference Experiment} \begin{figure}[tb] \unitlength=1in \begin{picture}(6,3.5) \put(.6,.9){\scalebox{3}{\includegraphics{GRAfig4new.eps}}} \put(0.5, .4){Figure 4. GRA's interference experiment. The experiment uses the same novel gating} \put(0.5, .2){system (not shown) to produce a near ideal single photon state as in GRA's which-path,} \put(0.5, 0){experiment.\label{GRAmz}} \end{picture} \end{figure} In the second interference experiment, GRA built a Mach-Zehnder interferometer around the first beam-splitter as shown in figure 4. Quantum mechanics predicts that each beam is oppositely modulated and that the fringe visibility of each beam as a function of path difference (or of a phase shift produced by a phase shifter) is $1$. In the experiment, interference fringes with visibility greater than 98\% were observed. Although the interference is expected, this is perhaps the first experiment to demonstrate interference for a genuine single photon state, as GRA themselves have emphasized. \section{GRA's experiments according to CIEM} GRA concluded from their results that in a which-path measurement a photon does not split at the beam-splitter and therefore chooses only one path, but, in a one-photon-at-a-time interference experiment a photon splits at the beam-splitter and interferes with itself to produce an interference pattern. They view this result as experimental confirmation of particle-wave duality, and hence, of Bohr's principle of complementarity. Without doubt, GRA's experiments with the novel and ingenious gating system constitutes an important experimental confirmation of quantum mechanics for genuine single photon states. But, by providing a detailed wave model of both experiments, we want to show that GRA's experiments cannot be regarded as confirmation of particle-wave duality, and hence, nor of Bohr's principle of complementarity. We refer the reader to reference \cite{K87}, but particularly reference \cite{K94} for details of CIEM. Before proceeding we first give an outline of CIEM as given in reference (\cite{K05}, p. 300). \subsection{Outline of CIEM} In what follows we use the radiation gauge in which the divergence of the vector potential is zero $\nabla.\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)=0$, and the scalar potential is also zero $\phi(\mbox{\boldmath $x$},t) = 0$. In this gauge the electromagnetic field has only two transverse components. Heavyside-Lorentz units are used throughout. Second quantization is effected by treating the field $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ and its conjugate momentum $\mbox{{ \boldmath{$\mit \Pi$}}}(\mbox{\boldmath $x$},t)$ as operators satisfying the equal-time commutation relations. This procedure is equivalent to introducing a field Schr\"{o}dinger equation \begin{equation} \int {\cal H} ( \mbox{\boldmath{$A$}}', \mbox{{ \boldmath{$\mit \Pi$}}}') {\mit \Phi}[\Ab,t]\; d\mbox{\boldmath $x$}'= i \hbar \frac{\partial {\mit \Phi}[\Ab,t]}{\partial t},\label{SE} \end{equation} where the Hamiltonian density operator ${\cal H}$ is obtained from the classical Hamiltonian density of the electromagnetic field, \begin{equation} {\cal H} =\frac{1}{2}(\mbox{\boldmath{$E$}}^{2}+\mbox{\boldmath{$B$}}^{2})= \frac{1}{2}[c^{2}\mbox{{ \boldmath{$\mit \Pi$}}}^{2}+(\nabla\times\mbox{\boldmath{$A$}})^2 ], \label{H} \end{equation} by the operator replacement $\mbox{{ \boldmath{$\mit \Pi$}}}\rightarrow -i\hbar\, \delta'/\delta' \mbox{\boldmath{$A$}}$. $\mbox{\boldmath{$A$}}'$ is shorthand for $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$}',t)$. In earlier articles \cite{K05, K94} $\delta'/\delta' \mbox{\boldmath{$A$}}$ (without the prime) was defined as the variational derivative \footnote{\label{FD} For a scalar function $\phi$ the variational or functional derivative is defined as $\frac{\delta'}{\delta' \phi}=\frac{\partial}{\partial\phi}-\Sigma_i\left(\frac{\partial}{\partial\left(\frac{\partial \phi}{\partial x_i}\right)}\right)$ (\cite{SHFF68}, p. 494). For a vector function $\mbox{\boldmath{$A$}}$ we have defined it to be $\frac{\delta}{\delta \mbox{{\scriptsize\boldmath{$A$}}}}=\frac{\delta}{\delta A_{x}}{\mbox{{\boldmath{$\mit i$}}}}+\frac{\delta}{\delta A_{y}}{\mbox{{\boldmath{$\mit j$}}}}+\frac{\delta}{\delta A_{z}}{\mbox{{\boldmath{$\mit k$}}}}$, where each component is defined in the same as for the scalar function.}. This definition leads to the equal-time commutation relations \[ [A_{i}(\mbox{\boldmath $x$},t), \mathit{\Pi}_{j}(\mbox{\boldmath $x$}',t)]=-\frac{1}{c}[A_{i}(\mbox{\boldmath $x$},t), E_{j}(\mbox{\boldmath $x$}',t)]= i\hbar \delta_{ij}\delta^3( \mbox{\boldmath $x$} - {\mbox{\boldmath $x$}'}). \] Unfortunately, these commutation relations are known to be inconsistent both with Gauss's law in free space, $\nabla.\mbox{\boldmath{$E$}}=0$, and the Coulomb gauge condition, $\nabla.\mbox{\boldmath{$A$}}=0$, since it follows from these that either of the two left-hand-side terms are zero, whereas the divergence of the delta function $\delta^3( \mbox{\boldmath $x$} - {\mbox{\boldmath $x$}'})$ is not zero. We noted this inconsistency in our original development of CIEM \cite{K94}, but justified this simplification by noting that it leads to the correct equations of motion. This justification, however, has recently been criticized by Struyve in reference \cite{WS2005}, p. 88\footnote{We would like to thank one of the referees for pointing out this reference and for re-emphasizing this inconsistency.}. As is well known, the commutation relations that are consistent with $\nabla.\mbox{\boldmath{$E$}}=0$ and $\nabla.\mbox{\boldmath{$A$}}=0$ are \begin{equation} [A_{i}(\mbox{\boldmath $x$},t), \mathit{\Pi}_{j}(\mbox{\boldmath $x$}',t)]=i\hbar \delta_{ij}^{tr}( \mbox{\boldmath $x$} - \mbox{\boldmath $x$}') \label{CCB} \end{equation} where $ \delta_{ij}^{tr}( \mbox{\boldmath $x$} - \mbox{\boldmath $x$}')$ is the transverse delta function defined by \cite{BJD, LHR} \[ \delta_{ij}^{tr}(\mbox{\boldmath $x$}-\mbox{\boldmath $x$}')= \frac{1}{(2 \pi)^3} \int e^{i{\mbox{{\boldmath{$\mit k$}}}}.(\mbox{\boldmath $x$}-\mbox{\boldmath $x$}')} \left ( \delta_{ij} - \frac{k_i k_j}{\mbox{{\boldmath{$\mit k$}}}^{2}} \right ) d^3{\mbox{{\boldmath{$\mit k$}}}}=\left(\delta_{ij} - \frac{\partial_i \partial_j}{\nabla^2}\right)\delta^3(\mbox{\boldmath $x$}-\mbox{\boldmath $x$}') \] We can establish consistency with the correct equal-time commutation relations, eq. (\ref{CCB}), by modifying the definition of the momentum operator as follows: \[ \mathit{\Pi}_i=-i\hbar\frac{\delta}{\delta A_i}=-i\hbar\left( \frac{\delta'}{\delta' A_i}-\sum_k\frac{\partial_i\partial_k}{\nabla^2}\frac{\delta'}{\delta' A_k}\right), \] where $\delta'/\delta' A_k$ is the usual functional derivative defined in footnote \ref{FD}. We note that the definition of the normal mode momentum operator given in the original article in which CIEM is developed \cite{K94} is consistent with the correct commutation relations, eq. (\ref{CCB}), and does not need modification. The solution of the field Schr\"{o}dinger equation is the wave functional ${\mit \Phi}[\Ab,t]$. The square of the modulus of the wave functional $|{\mit \Phi}[\Ab,t]|^2$ gives the probability density for a given field configuration $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$. This suggests that we take $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ as a beable. Thus, as we have already said, the basic ontology is that of a field; there are no photon particles. We substitute ${\mit \Phi}=R[\mbox{\boldmath{$A$}},t]\exp(iS[\mbox{\boldmath{$A$}},t]/\hbar)$, where $R[\mbox{\boldmath{$A$}},t]$ and $S[\mbox{\boldmath{$A$}},t]$ are two real functionals which codetermine one another, into the field Schr\"{o}dinger equation. Then, differentiating, rearranging and equating imaginary terms gives a continuity equation: \begin{equation} \frac{\partial R^{2}}{\partial t} + c^{2} \int \frac{\delta}{\delta \mbox{\boldmath{$A$}}'} \left(R^{2}\frac{\delta S}{\delta \mbox{\boldmath{$A$}}'} \right) \; d\mbox{\boldmath $x$}' = 0. \end{equation} The continuity equation is interpreted as expressing conservation of probability in function space. Equating real terms gives a Hamilton-Jacobi type equation: \begin{equation} \frac{\partial S}{\partial t}+\frac{1}{2}\int\left(\frac{\delta S}{\delta \mbox{\boldmath{$A$}}'}\right)^{2} c^{2}+(\nabla\times\mbox{\boldmath{$A$}}')^{2}+\left(-\frac{\hbar^2 c^{2}}{R}\frac{\delta^{2} R}{\delta \mbox{\boldmath{$A$}}'^{2}} \right) d\mbox{\boldmath $x$}'= 0. \label{HJ1} \end{equation} This Hamilton-Jacobi equation differs from its classical counterpart by the extra classical term \begin{equation} Q =-\frac{1}{2}\int\frac{\hbar^{2} c^{2}}{R} \frac{\delta^{2} R}{\delta \mbox{\boldmath{$A$}}'^{2}}\;d\mbox{\boldmath $x$}', \end{equation} which we call the field quantum potential. By analogy with classical Hamilton-Jacobi theory we define the total energy and momentum conjugate to the field as \begin{equation} E = -\frac{\partial S[\mbox{\boldmath{$A$}}]}{\partial t},\;\;\;\;\;\mbox{{ \boldmath{$\mit \Pi$}}}=\frac{\delta S[\mbox{\boldmath{$A$}}]}{\delta \mbox{\boldmath{$A$}}}. \end{equation} In addition to the beables $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ and $\mbox{{ \boldmath{$\mit \Pi$}}}(\mbox{\boldmath $x$},t)$, we can define other field beables: the electric field, the magnetic induction, the energy and energy density, the momentum and momentum density, the intensity, etc. Formulae for these beables are obtained by replacing $\mbox{{ \boldmath{$\mit \Pi$}}}$ by $\delta S/\delta \mbox{\boldmath{$A$}}$ in the classical formula. Thus, we can picture an electromagnetic field as a field in the classical sense, but with the additional property of nonlocality. That the field is inherently nonlocal, meaning that an interaction at one point in the field instantaneously influences the field at all other points, can be seen in two ways: First, by using Euler's method of finite differences a functional can be approximated as a function of infinitely many variables: ${\mit \Phi}[\Ab,t]\rightarrow{\mit \Phi}(\mbox{\boldmath{$A$}}_1,\mbox{\boldmath{$A$}}_2,\ldots,t)$. Comparison with a many-body wavefunction $\psi(\mbox{\boldmath $x$}_1,\mbox{\boldmath $x$}_2,...,t)$ reveals the nonlocality. The second way is from the equation of motion of $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$, i.e., the free field wave equation. This is obtained by taking the functional derivative of the Hamilton-Jacobi equation, (\ref{HJ1}): \begin{equation} \nabla^{2}\mbox{\boldmath{$A$}}-\frac{1}{c^{2}}\frac{\partial^{2}\mbox{\boldmath{$A$}}}{\partial t^{2}}= \frac{\delta Q}{\delta\mbox{\boldmath{$A$}}}. \end{equation} In general $\delta Q/\delta\mbox{\boldmath{$A$}}$ will involve an integral over space in which the integrand contains $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$. This means that the way that $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ changes with time at one point depends on $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ at all other points, hence the inherent nonlocality. \subsection{Normal mode coordinates\label{NMC}} To proceed it is mathematically easier to expand $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ and $\mbox{{ \boldmath{$\mit \Pi$}}}(\mbox{\boldmath $x$},t)$ as Fourier series \begin{equation} \mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)=\frac{1}{V^{\frac{1}{2}}}\sum_{k\mu}\ekq_{k\mu}(t)e^{i\kb.\xb},\;\;\;\;\;\;\;\;\;\; \mbox{{ \boldmath{$\mit \Pi$}}}(\mbox{\boldmath $x$},t)=\frac{1}{V^{\frac{1}{2}}}\sum_{k\mu}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\pi_{k\mu}(t) e^{-i\kb.\xb}, \label{AFS} \end{equation} where the field is assumed to be enclosed in a large volume $V=L^3$. The wavenumber $k$ runs from $-\infty$ to $+\infty$ and $\mu=1,2$ is the polarization index. For $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ to be a real function we must have \begin{equation} \mbox{\boldmath $\hat\varepsilon$}_{-k\mu}q_{-k\mu}=\ekq_{k\mu}^{*}.\label{QMP} \end{equation} Substituting eq.'s (\ref{H}) and (\ref{AFS}) into eq. (\ref{SE}) gives the Schr\"{o}dingier equation in terms of the normal mode coordinates $q_{k\mu}$: \begin{equation} \frac{1}{2}\sum_{k\mu}\left(-\hbar^{2}c^{2}\frac{\partial^2{\mit \Phi}}{\partial q_{k\mu}^{*}\partialq_{k\mu} }+\kappa^{2}\qksq_{k\mu}{\mit \Phi}\right)= i\hbar\frac{\partial{\mit \Phi}}{\partial t}. \label{SEN} \end{equation} The solution ${\mit \Phi}( q_{k\mu},t)$ is an ordinary function of all the normal mode coordinates and this simplifies proceedings. We substitute ${\mit \Phi}=R(q_{k\mu},t)\exp[iS(q_{k\mu},t)/\hbar]$, where $R(q_{k\mu},t)$ and $S(q_{k\mu},t)$ are real functions which codetermine one another, into eq. (\ref{SEN}). Then, differentiating, rearranging and equating real terms gives the continuity equation in terms of normal modes: \begin{equation} \frac{\partial R^2}{\partial t}+\sum_{k\mu}\left[\frac{c^2}{2}\frac{\partial}{\partial q_{k\mu}}\left(R^2\frac{\partial S}{\partialq_{k\mu}^{*}}\right)+ \frac{c^2}{2}\frac{\partial}{\partialq_{k\mu}^{*}}\left(R^2\frac{\partial S}{\partialq_{k\mu}} \right) \right]=0. \end{equation} Equating imaginary terms gives the Hamilton-Jacobi equation in terms of normal modes: \begin{equation} \frac{\partial S}{\partial t}+\sum_{k\mu}\left[\frac{c^2}{2}\frac{\partial S}{\partialq_{k\mu}^{*}}\frac{\partial S}{\partial q_{k\mu}}+\frac{\kappa^{2}}{2}\qksq_{k\mu} +\left(-\frac{\hbar^{2} c^{2}}{2R}\frac{\partial^{2}R} {\partialq_{k\mu}^{*}\partialq_{k\mu}}\right)\right]=0. \label{HJ2} \end{equation} The term \begin{equation} Q = -\sum_{k\mu}\frac{\hbar^{2} c^{2}}{2R}\frac{\partial^{2}R} {\partialq_{k\mu}^{*}\partialq_{k\mu}} \label{QP} \end{equation} is the field quantum potential. Again, by analogy with classical Hamilton-Jacobi theory we define the total energy and the conjugate momenta as \begin{equation} E=-\frac{\partial S}{\partial t},\;\;\;\;\;\pi_{k\mu}=\frac{\partial S}{\partialq_{k\mu}},\;\;\;\;\; \pi_{k\mu}^{*}=\frac{\partial S}{\partialq_{k\mu}^{*}}. \end{equation} The square of the modulus of the wave function $|{\mit \Phi}( q_{k\mu},t)|^2$ is the probability density for each $q_{k\mu}(t)$ to take a particular value at time $t$. Substituting a particular set of values of $q_{k\mu}(t)$ at time $t$ into eq. (\ref{AFS}) gives a particular field configuration at time $t$, as before. Substituting the initial values of $q_{k\mu}(t)$ gives the initial field configuration. The normalized ground state solution of the Schr\"{o}dinger equation is given by \begin{equation} {\mit \Phi}_0=N e^{-\sum_{k\mu}(\kappa/2\hbar c)q_{k\mu}^{*}q_{k\mu}}e^{-\sum_{k}i\kappa ct/2}, \end{equation} with $N= \prod_{k=1}^{\infty}(k/\hbar c \pi)^{\frac{1}{2}}$\footnote{The normalization factor $N$ is found by substituting $q_{k\mu}^{*}=f_{k\mu}+ig_{k\mu}$ and its conjugate into ${\mit \Phi}_0$ and using the normalization condition $\int_{-\infty}^{\infty} |{\mit \Phi}_0|^2df_{k\mu}dg_{k\mu}=1$, with $df_{k\mu}\equiv df_{k_11}df_{k_12}df_{k_21}\ldots$, and similarly for $dg_{k\mu}$.}. Higher excited states are obtained by the action of the creation operator $a^{\dag}_{k\mu}$: \begin{equation} {\mit \Phi}_{n_{k\mu}}=\frac{(a_{k\mu}^{\dag})^{n_{k\mu}}}{\sqrt{n_{k\mu}!}}{\mit \Phi}_{0}e^{- in_{k\mu}\kappa ct}. \end{equation} For a normalized ground state, the higher excited states remain normalized. For ease of writing we will not include the normalization factor $N$ in most expressions, but normalization of states will be assumed when calculating expectation values. Again, the formula for the field beables are obtained by replacing the conjugate momenta $\pi_{k\mu}$ and $\pi_{k\mu}^{*}$ by $\partial S/\partialq_{k\mu}$ and $\partial S/\partialq_{k\mu}^{*}$ in the corresponding classical formula. The following is a list of formulae for the beables:\\ \mbox{}\\ The vector potential $\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)$ is given in eq. (\ref{AFS}). The electric field is \begin{equation} \mbox{\boldmath $E$}(\mbox{\boldmath $x$},t)=-c\mbox{{ \boldmath{$\mit \Pi$}}}(\mbox{\boldmath $x$},t)=-\frac{1}{c}\frac{\partial \mbox{\boldmath{$A$}}}{\partial t}= - \frac{c}{V^{\frac{1}{2}}}\sum_{k\mu}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\frac{\partial S}{\partialq_{k\mu}}e^{-i\kb.\xb}. \label{PEX} \end{equation} The magnetic induction is \begin{equation} \mbox{\boldmath $B$}(\mbox{\boldmath $x$},t)=\nabla\times\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)=\frac{i}{V^{\frac{1}{2}}}\sum_{k\mu}(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})q_{k\mu}(t)e^{i\kb.\xb}. \label{BEX} \end{equation} We may also define the energy density, which includes the quantum potential density (see reference \cite{K94}), but we will not write these here as we will not need them. The total energy is found by integrating the energy density over $V$ to get \begin{equation} E=-\frac{\partial S}{\partial t}=\sum_{k\mu}\left[\frac{c^{2}}{2} \frac{\partial S}{\partialq_{k\mu}^{*}}\frac{\partial S}{\partialq_{k\mu}}+\frac{\kappa^{2}}{2}\qksq_{k\mu} +\left(-\frac{\hbar^{2}c^{2}}{2R}\frac{\partial^{2}R} {\partial q_{k\mu}^{*}\partialq_{k\mu}} \right)\right]. \end{equation} The intensity is equal to momentum density multiplied by $c^2$: \begin{equation} \mbox{\boldmath $I$}(\mbox{\boldmath $x$},t)=c^2\mbox{\boldmath${\cal G}$}= \frac{-ic^2}{V}\sum_{k\mu}\sum_{k'\mu'}\left[ \mbox{\boldmath $\hat\varepsilon$}_{k'\mu'}\times(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\frac{\partial S}{\partial q_{k'\mu'}}q_{k\mu}e^{i(\kb-\kbp).\xb} \right]. \label{I} \end{equation} We have adopted the classical definition of intensity in which the intensity is equal to the Poynting vector (in Heavyside-Lorentz units), i.e., $\mbox{\boldmath $I$}=c(\mbox{\boldmath $E$}\times\mbox{\boldmath $B$})$. The definition leads to a moderately simple formula for the intensity beable. We note that the definition above contains a zero point intensity. But, because $\mbox{\boldmath $I$}$ is a vector (whereas energy is not) the contributions to the zero point intensity from individual waves with wave vector $\mbox{\boldmath $k$}$ cancel each other because of symmetry; for each $\mbox{\boldmath $k$}$ there is another $\mbox{\boldmath $k$}$ pointing in the opposite direction. The above, however, is not the definition normally used in quantum optics. This is probably because, although it leads to a simple formula for the intensity beable, it leads to a very cumbersome expression for the intensity operator in terms of the creation and annihilation operators: \begin{eqnarray} &&\mbox{{ \boldmath{$\mbox{\boldmath $\hat I$}$}}} =\frac{-\hbar c^2}{4V}\sum_{k\mu}\sum_{k'\mu'}\left[ \frac{k}{k'}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\times(\mbox{\boldmath $k$}'\times\mbox{\boldmath $\hat\varepsilon$}_{k'\mu'})- \frac{k'}{k}(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\times\mbox{\boldmath $\hat\varepsilon$}_{k'\mu'} \right] \nonumber\\ &&\times\left[ \hat{a}_{k\mu}\hat{a}_{k'\mu'}e^{i(\mbox{{\scriptsize\boldmath{$k$}}}+\mbox{{\scriptsize\boldmath{$k'$}}}).\mbox{{\scriptsize\boldmath{$x$}}}} - \hat{a}_{k\mu}\hat{a}^\dag_{k'\mu'}e^{i(\kb-\kbp).\xb}- \hat{a}^\dag_{k\mu}\hat{a}_{k'\mu'}e^{-i(\kb-\kbp).\xb} +\hat{a}^\dag_{k\mu}\hat{a}^\dag_{k'\mu'}e^{-i(\mbox{{\scriptsize\boldmath{$k$}}}+\mbox{{\scriptsize\boldmath{$k'$}}}).\mbox{{\scriptsize\boldmath{$x$}}}} \right]. \label{Ipan} \end{eqnarray} In quantum optics the intensity operator is defined instead as $\mbox{\boldmath $\hat I$} =c(\mbox{{ \boldmath{$\hat{E}^+$}}}\times \mbox{{ \boldmath{$\hat{B}^-$}}} - \mbox{{ \boldmath{$\hat{B}^-$}}}\times \mbox{{ \boldmath{$\hat{E}^+$}}})$, and leads to a much simpler expression in terms of creation and annihilation operators \begin{equation} \mbox{\boldmath $\hat I$}= \frac{\hbar c^2}{V}\sum_{k\mu}\sum_{k'\mu'}\hat{\mbox{\boldmath $k$}}\sqrt{kk'} \hat{a}^\dag_{k\mu}\hat{a}_{k'\mu'}e^{i(\mbox{{\scriptsize\boldmath{$k'$}}}-\mbox{{\scriptsize\boldmath{$k$}}}).\mbox{{\scriptsize\boldmath{$x$}}}}. \label{Iqo} \end{equation} This definition is justified because it is proportional to the dominant term in the interaction Hamiltonian for the photoelectric effect upon which instruments that measure intensity are based. We note that the two forms of the intensity operator lead to identical expectation values and perhaps further justifies the simpler definition of the intensity operator. From the above we see that objects such as $q_{k\mu}$, $\pi_{k\mu}$, etc., regarded as time independent operators in the Schr\"{o}dinger picture of the usual interpretation, become functions of time in CIEM. For a given state ${\mit \Phi}(q_{k\mu},t)$ of the field we determine the beables by first finding $\partial S/\partialq_{k\mu}$ and its complex conjugate using the formula \begin{equation} S=\left(\frac{\hbar}{2i}\right)\ln\left(\frac{{\mit \Phi}}{{\mit \Phi}^*}\right).\label{FlaS} \end{equation} This gives the beables as functions of the $q_{k\mu}(t)$ and $q_{k\mu}^{*}(t)$. The beables can then be obtained in terms of the initial values by solving the equations of motion for $q_{k\mu}(t)$ and $q_{k\mu}^{*}(t)$. There are two alternative but equivalent forms of the equations of motion. The first follows from the classical formula \begin{equation} \pi_{k\mu} = \frac{\partial{\cal L}}{\partial\left(\frac{d q_{k\mu}}{d t} \right)} =\frac{1}{c^2}\frac{dq_{k\mu}^{*}}{d t}, \end{equation} where ${\cal L}$ is the Lagrangian density of the electromagnetic field, by replacing $\pi_{k\mu}$ by $\partial S/\partialq_{k\mu}$. This gives the equations of motion as \begin{equation} \frac{1}{c^2}\frac{dq_{k\mu}^{*}(t)}{d t} =\frac{\partial S}{\partialq_{k\mu}(t)}.\label{EQMG} \end{equation} The second form of the equations of motion for $q_{k\mu}$ is obtained by differentiating the Hamilton Jacobi equation (\ref{HJ2}) by $q_{k\mu}^{*}$. This gives the wave equations \begin{equation} \frac{1}{c^2}\frac{d^{2}q_{k\mu}^{*}}{d t^{2}}+\kappa^{2}q_{k\mu}^{*}=-\frac{\partial Q}{\partialq_{k\mu}}. \label{WEQ} \end{equation} The corresponding equations for $q_{k\mu}$ are the complex conjugates of the above. These equations of motion differ from the classical free field wave equation by the derivative of the quantum potential. From this it follows that where the quantum potential is zero or small the quantum field behaves like a classical field. In applications we will obviously choose to solve the simpler eq. ({\ref{EQMG}}). We conclude with a few words to clarify our model. The electromagnetic field beables are $\mbox{\boldmath $E$}(\mbox{\boldmath $x$},t)$ and $\mbox{\boldmath $B$}(\mbox{\boldmath $x$},t)$ and are objectively existing entities in real space. The state ${\mit \Phi}=R\exp[iS/\hbar]$ is made up of the $R$ and $S$ functionals. By thinking in terms of the approximation of a functional as a function of infinitely many variables or in term of normal mode coordinates we can picture $R$ and $S$ as connecting the field coordinates and shaping the behaviour of the field through the equations of motion (\ref{EQMG}) or (\ref{WEQ}), but the $R$ and $S$ beables (and hence the state ${\mit \Phi}$) are not the electromagnetic field itself. The $R$ and $S$ beables co-determine one another and the motion of the field can be determined from either one without reference to the other. This is reflected in the two possible forms of the equations of motion. \subsection{GRA's which-path experiment according to CIEM} Refer to figure 1. To keep the mathematics simple we assume (a) a symmetrical beam-splitter so that the reflection and transmission coefficients are equal and given by $r=t=1/\sqrt{2}$, (b) a $\pi/2$ phase shift upon reflection, and (c) no phase shift upon transmission. With this in mind, the state of the photon after the beam-splitter but before the mirrors and phase shifter is \begin{equation} {\mit \Phi}_{I}=\frac{1}{\sqrt{2}}\left( {\mit \Phi}_{\alpha}+i{\mit \Phi}_{\beta}\right), \label{PHRI} \end{equation} where ${\mit \Phi}_{\alpha}$ and ${\mit \Phi}_{\beta}$ are solutions of the normal mode Schr\"{o}dinger equation and are given by \begin{eqnarray} {\mit \Phi}_{\alpha}(q_{k\mu},t)& =& \left(\frac{2\kappa_\alpha}{\hbar c}\right)^{\frac{1}{2}} \alpha_{k_\alpha\mu_\alpha}^{*}{\mit \Phi}_{0}e^{-i\kappa_\alpha ct}, \;\;\;\;\;\;\;\;\; {\mit \Phi}_{\beta}(q_{k\mu},t) = \left(\frac{2\kappa_\beta}{\hbar c}\right)^{\frac{1}{2}} \beta_{k_\beta\mu_\beta}^{*}{\mit \Phi}_{0}e^{-i\kappa_\beta ct},\nonumber\\ {\mit \Phi}_0(q_{k\mu},t)&=&N e^{-\sum_{k\mu}(\kappa/2\hbar c)q_{k\mu}^{*}q_{k\mu}}e^{-\sum_{k}i\kappa ct/2}. \end{eqnarray} The magnitudes of the $k$-vectors are equal, i.e., $k_\alpha=k_\beta=k_0$. The $\alpha_{k_\alpha\mu_\alpha}$ normal mode coordinates represent the horizontal beam and the $\beta_{k_\beta\mu_\beta}$ coordinates represent the vertical beam. It is clear that the single photon input state ${\mit \Phi}_i(q_{k\mu},t) =(2\kappa_0/(\hbar c))^{\frac{1}{2}} q_{k_0\mu_0}^{*}(t){\mit \Phi}_{0}e^{-i\kappa_0 ct} $ is split by the beam-splitter into two beams. This remains true irrespective of whether a subsequent measurement is a which-path measurement or it is the observation of interference. The mathematical description is unique. In CIEM the normal mode coordinates are regarded as functions of time and represent an actually existing electromagnetic field. The modulus squared of the wavefunction is a probability density from which the probabilities for the normal modes to have particular values are found. The totality of these probabilities gives the probability for a particular field configuration. Thus, the ontology is that of a field; there are no photon particles. In fact, for a number state the most probable field configuration is one or more plane waves, which, in general, are nonlocal (\cite{K94}, p. 326). As we mentioned earlier, in CIEM we use the term photon to refer to a quantum of energy $\hbar\omega$ (or an average about this value for a wave packet) without in any way implying particle properties. To find the equations of motion for the normal mode coordinates we first find $S$ from ${\mit \Phi}_{I}=R(q_{k\mu},t)\exp(iS(q_{k\mu}, t))$ and then substitute into \begin{equation} \frac{1}{c^2}\frac{dq_{k\mu}^{*}(t)}{d t} =\frac{\partial S}{\partialq_{k\mu}(t)}. \end{equation} This gives the equations of motion \begin{eqnarray} \frac{d\alpha_{k_\alpha\mu_\alpha}^{*}}{dt}&=&c^2\frac{\partial S}{\partial\alpha_{k_\alpha\mu_\alpha}}=\frac{\hbar c^2}{2}\frac{i}{\left(\alpha_{k_\alpha\mu_\alpha}-i\beta_{k_\beta\mu_\beta} \right)}, \label{EQMa} \\ \frac{d\beta_{k_\beta\mu_\beta}^{*}}{dt} &=&c^2\frac{\partial S}{\partial\beta_{k_\beta\mu_\beta}} =\frac{\hbar c^2}{2}\frac{1}{\left(\alpha_{k_\alpha\mu_\alpha}-i\beta_{k_\beta\mu_\beta} \right)}, \label{EQMb}\\ \frac{dq_{k\mu}^{*}}{dt} &=&c^2\frac{\partial S}{\partialq_{k\mu}} =0,\;\;\;\; \mathrm{for}\; k\neq \pm k_{\alpha}, \pm k_{\beta}. \label{EQMq} \end{eqnarray} Eqs. (\ref{EQMa}) and (\ref{EQMb}) are coupled differential equations and the coupling indicates that the two beams are nonlocally connected. The solutions are \begin{equation} \alpha_{k_\alpha\mu_\alpha}^{*}(t)=\alpha_0 e^{i(\omega_\alpha t+\sigma_0)}, \;\;\; \beta_{k_\beta\mu_\beta}^{*}(t)=\beta_0 e^{i(\omega_\beta t+\tau_0)}, \;\;\; q_{k\mu}^{*}(t)= q_{k\mu 0}e^{i\zeta_{k\mu 0}}\;\mathrm{for}\;k\neq \pm k_{\alpha}, \pm k_{\beta}, \label{QKS} \end{equation} where $\sigma_0$ and $\tau_0$ are integration constants corresponding to the initial phases, and $\alpha_0$ and $\beta_0$ are constant initial amplitudes. The omega's, $\omega_\alpha=\hbar c^2/4\alpha_0^2$ and $\omega_{\beta}=\hbar c^2/4\beta_0^2$, are nonclassical frequencies which depend on the amplitudes $\alpha_0$ and $\beta_0$. The vector potential, electric intensity, magnetic induction and intensity beables are given by the formulae \begin{eqnarray} \mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)&=&\frac{1}{V^{\frac{1}{2}}}\sum_{k\mu}\ekq_{k\mu}(t)e^{i\kb.\xb},\nonumber\\ \mbox{\boldmath $E$}(\mbox{\boldmath $x$},t)&=&-c\mbox{{ \boldmath{$\mit \Pi$}}}(\mbox{\boldmath $x$},t)=-\frac{1}{c}\frac{\partial \mbox{\boldmath{$A$}}}{\partial t}= - \frac{c}{V^{\frac{1}{2}}}\sum_{k\mu}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\frac{\partial S}{\partialq_{k\mu}}e^{-i\kb.\xb}, \nonumber\\ \mbox{\boldmath $B$}(\mbox{\boldmath $x$},t)&=&\nabla\times\mbox{\boldmath{$A$}}(\mbox{\boldmath $x$},t)=\frac{i}{V^{\frac{1}{2}}}\sum_{k\mu}(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})q_{k\mu}(t)e^{i\kb.\xb}, \nonumber\\ \mbox{\boldmath $I$}(\mbox{\boldmath $x$},t)&=&c^2\mbox{\boldmath${\cal G}$}= \frac{-ic^2}{V}\sum_{k\mu}\sum_{k'\mu'}\left[ \mbox{\boldmath $\hat\varepsilon$}_{k'\mu'}\times(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\frac{\partial S}{\partial q_{k'\mu'}}q_{k\mu}e^{i(\kb-\kbp).\xb} \right]. \end{eqnarray} Substituting equations (\ref{EQMa}) to (\ref{QKS}) into the above formulae gives the field beables associated with the state ${\mit \Phi}_I$: \begin{eqnarray} \mbox{\boldmath{$A$}}_I(x,t)&=&\frac{2}{V^{\frac{1}{2}}}\left(\mbox{\boldmath $\hat\varepsilon$}_{k_\alpha\mu_\alpha}\alpha_0\cos{\mit \Theta}_{\alpha}+\mbox{\boldmath $\hat\varepsilon$}_{k_\beta\mu_\beta}\beta_0\cos{\mit \Theta}_{\beta} \right) +\frac{\mbox{\boldmath $u$}_I(\mbox{\boldmath $x$})}{V^{\frac{1}{2}}},\nonumber \\ \mbox{\boldmath $E$}_I(\mbox{\boldmath $x$},t)&=&\frac{-\hbar c}{2V^{\frac{1}{2}}}\left(\frac{\mbox{\boldmath $\hat\varepsilon$}_{k_\alpha\mu_\alpha}}{\alpha_0}\sin{\mit \Theta}_{\alpha}+\frac{\mbox{\boldmath $\hat\varepsilon$}_{k_\beta\mu_\beta}}{\beta_0}\sin{\mit \Theta}_{\beta} \right), \label{EI}\\ \mbox{\boldmath $B$}_I(\mbox{\boldmath $x$},t) &=&\frac{-2}{V^{\frac{1}{2}}}\left[(\mbox{\boldmath $k$}_\alpha\times\mbox{\boldmath $\hat\varepsilon$}_{k_\alpha\mu_\alpha})\alpha_0\sin{\mit \Theta}_{\alpha} + (\mbox{\boldmath $k$}_\beta\times\mbox{\boldmath $\hat\varepsilon$}_{k_\beta\mu_\beta})\beta_0 \sin {\mit \Theta}_{\beta}\right] + \frac{\mbox{\boldmath $v$}_I(\mbox{\boldmath $x$})}{V^{\frac{1}{2}}}, \nonumber\\ \mbox{\boldmath $I$}_I(\mbox{\boldmath $x$},t)&=&\frac{\hbar c^2}{2V}\left(\mbox{\boldmath $k$}_\alpha+\mbox{\boldmath $k$}_\beta-\mbox{\boldmath $k$}_\alpha\cos2{\mit \Theta}_{\alpha} -\mbox{\boldmath $k$}_\beta\cos 2{\mit \Theta}_{\beta}\right)-\frac{\mbox{\boldmath $f$}_I(\mbox{\boldmath $x$})\mbox{\boldmath $g$}_I(\mbox{\boldmath $x$},t)}{V}, \end{eqnarray} with ${\mit \Theta}_{\alpha}=\mbox{\boldmath $k$}_\alpha.\mbox{\boldmath $x$}-\omega_{\alpha} t-\sigma_0$ and ${\mit \Theta}_{\beta}=\mbox{\boldmath $k$}_\beta.\mbox{\boldmath $x$}-\omega_{\beta} t-\tau_0$, and \begin{eqnarray} \mbox{\boldmath $u$}_I(\mbox{\boldmath $x$})&=&\!\!\!\!\!\sum_{\stackrel{\scriptstyle{k\mu}}{k\neq \pm k_{\alpha},\pm k_{\beta}}}\!\!\!\!\!\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\qke^{i\kb.\xb}, \;\;\;\;\;\mbox{\boldmath $v$}_I(\mbox{\boldmath $x$})=\nabla\times\mbox{\boldmath $u$}_I(\mbox{\boldmath $x$})=i\!\!\!\!\!\sum_{\stackrel{\scriptstyle{k\mu}}{k\neq \pm k_{\alpha},\pm k_{\beta}}}\!\!\!\!\!(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\qke^{i\kb.\xb}, \label{VXI}\\ \mbox{\boldmath $f$}_I(\mbox{\boldmath $x$})&=&i\hbar c^2\!\!\!\!\!\sum_{\stackrel{\scriptstyle{k\mu}}{k\neq \pm k_{\alpha},\pm k_{\beta}}}\!\!\!\!\!\mbox{\boldmath $\hat\varepsilon$}_{k_\alpha\mu_\alpha}\times(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\qke^{i\kb.\xb},\;\;\;\;\; \mbox{\boldmath $g$}_I(\mbox{\boldmath $x$},t)=\sin {\mit \Theta}_{\alpha}+\sin {\mit \Theta}_{\beta}. \end{eqnarray} Complementarity is not a direct interpretation of the mathematical formalism, so that the uniqueness of the mathematical description is not reflected in the duality of complementary concepts. The ontology of CIEM, on the other hand, is a direct interpretation of the elements of the mathematical formalism. The beables above therefore reflect the splitting of the state ${\mit \Phi}_i$ into two beams. In other words, the photon always splits at the beam-splitter irrespective of the nature of any planned future measurement. Quantum mechanics predicts that in a which-path measurement a photon will be detected in only one path. Feeble light experiments of the past have confirmed this prediction indirectly, while GRA's which-path experiment provides direct confirmation. Our CIEM model must therefore explain how a photon is detected in only one path, even though the photon must split at the beam-splitter. To see how this comes about we outline the interaction of the electromagnetic field in state ${\mit \Phi}_I$ with the photomultipliers. For mathematical simplicity we model the photomultipliers $PM_t$ and $PM_r$ as hydrogen atoms. We assume that the incident photon has sufficient energy to ionize one of the hydrogen atoms. The treatment we give here is a short summary of a more detailed outline given in reference (\cite{K05}, p. 310). The initial state of the field before interaction with the hydrogen atom is given by eq. (\ref{PHRI}). The initial state of the hydrogen atom is \begin{equation} u_i(\mbox{\boldmath $x$},t)=\frac{1}{\sqrt{\pi a^3}}e^{-r/a}e^{-iE_{ei}t/\hbar}, \end{equation} where $a=4\pi\hbar^2/\mu e^2$ is the Bohr magneton. With the initial state ${\mit \Phi}_{I_{k\mu}i}(q_{k\mu},\mbox{\boldmath $x$},t)={\mit \Phi}_{I_{k\mu}}(q_{k\mu},t)u_i(\mbox{\boldmath $x$},t)$, the Schr\"{o}dinger equation \begin{equation} i\hbar\frac{\partial {\mit \Phi}}{\partial t}= (H_R+H_A+H_I){\mit \Phi} \end{equation} can be solved using standard perturbation theory. $H_R$, $H_A$ and $H_I$ are the free radiation, free atomic, and interaction Hamiltonians, respectively, and are given by \begin{equation} H_R = \sum_{k\mu}\left(a^{\dag}_{k\mu}a_{k\mu}+\frac{1}{2}\right)\hbar\omega_k, \;\;\;\;\; H_A = \frac{-\hbar^2}{2\mu}\nabla^2+V(\mbox{\boldmath $x$}), \;\;\;\;\;H_I = \frac{i\hbar e}{\mu c}\left(\frac{\hbar c}{2V}\right)^{\frac{1}{2}} \sum_{k\mu}\frac{1}{\sqrt{k}}a_{k\mu}e^{i\kb.\xb}\mbox{\boldmath $\hat\varepsilon$}_{k\mu}.\nabla, \end{equation} with $\omega_k=kc$ and $\mu=m_em_n/(m_e+m_n)$ is the reduced mass. The final solution is \begin{equation} {\mit \Phi}={\mit \Phi}_{I_{k\mu}i}(q_{k\mu},\mbox{\boldmath $x$},t)-\frac{{\mit \Phi}_0(q_{k\mu},t)}{V}\sum_n \eta_{0n}(t)\mbox{\boldmath $\hat\varepsilon$}_{k_0\mu_0}.\mbox{\boldmath $k$}_{en}\frac{1}{\sqrt{V}}e^{i\left(\mbox{{\scriptsize\boldmath{$k$}}}_{en}.\mbox{{\scriptsize\boldmath{$x$}}}-E_{en}t/\hbar\right)}, \label{FSOL} \end{equation} with \begin{equation} \eta_{0n}(t)=\left(\frac{e}{\mu c}\right)\sqrt{\frac{\hbar c}{2V}}\left[\frac{(i-e^{i\phi})}{\sqrt{2 k_0}}\right]\left[\frac{\hbar}{\sqrt{V\pi a^3}} \frac{8\pi a^3}{(1+a^2 k_{en}^2)^2} \right]\left(\frac{1- e^{iE_{0n,I_{k\mu}i}t/\hbar}}{E_{0n,I_{k\mu}i}}\right). \end{equation} $E_{0n,I_{k\mu}i}$ is given by \begin{equation} E_{0n,I_{k\mu}i}=E_0+E_{en}-E_{I_{k\mu}}-E_{ei}. \end{equation} Eq. (\ref{FSOL}) clearly shows that one entire photon is absorbed. This is further emphasized by the integral \begin{equation} \sum_{k\mu}\frac{1}{\sqrt{k}}\int{\mit \Phi}^{*}_{N_{k\mu}}a_{k\mu}{\mit \Phi}_{I_{k\mu}}\;dq_{k\mu} = \frac{1}{\sqrt{2k_0}}(i-e^{i\phi})\int {\mit \Phi}^{*}_{N_{k\mu}}{\mit \Phi}_0\;dq_{k\mu}= \frac{1}{\sqrt{2k_0}}(i-e^{i\phi})\delta_{N_{k\mu}0}\delta_{kk_0}\delta{\mu\mu_0}, \end{equation} which is part of the matrix element $H_{N_{k\mu}n,I_{k\mu}i}$ used in obtaining the final solution. This term shows that if the interaction takes place at all then an entire electromagnetic quantum must be absorbed by the hydrogen atom. The initial state ${\mit \Phi}_{I_{k\mu}}$ represents a single photon divided between the two beams, but in the interaction with an atom positioned in one of the beams, the entire photon must be absorbed. Given that the interferometer arms can be of arbitrary length such absorption must in general be nonlocal. In this way we can explain why a photon that always divides at the beam-splitter nevertheless registers in only one path. The fact that this wave model exists prevents GRA's which-path experiment from being regarded as confirmation of the particle behaviour of light. \subsection{GRA's interference experiment according to CIEM} Refer to figure 4. Using the same phase and amplitude changes as in the previous section, and tracing the development of the two beams after $BM_2$, we arrive at the wavefunction \begin{equation} {\mit \Phi}_{II}=-\frac{1}{2}{\mit \Phi}_{c}(1+e^{i\phi})+\frac{i}{2}{\mit \Phi}_d(1-e^{i\phi}). \end{equation} By following a similar procedure to that of region I, we can find the $S$ corresponding to ${\mit \Phi}_{II}$ and hence set up and solve the equations of motion. Using these solutions the beables for region II are found to be \begin{eqnarray} \mbox{\boldmath{$A$}}_{II}(\mbox{\boldmath $x$},t)&=&\frac{2}{V^{\frac{1}{2}}}\left( \mbox{\boldmath $\hat\varepsilon$}_{k_c\mu_c} c_0 \cos{\mit \Theta}_c +\mbox{\boldmath $\hat\varepsilon$}_{k_d\mu_d} d_0\cos {\mit \Theta}_d\right)+\frac{\mbox{\boldmath $u$}_{II}(\mbox{\boldmath $x$})}{V^{\frac{1}{2}}},\nonumber\\ \mbox{\boldmath $E$}_{II}(\mbox{\boldmath $x$},t)&=&\frac{-\hbar c}{2V^{\frac{1}{2}}}\left(\frac{\mbox{\boldmath $\hat\varepsilon$}_{k_\alpha\mu_\alpha}}{c_0}(1+\cos\phi)\sin{\mit \Theta}_c + \frac{\mbox{\boldmath $\hat\varepsilon$}_{k_d\mu_d}}{d_0}(1-\cos\phi)\sin{\mit \Theta}_d \right),\nonumber\\ \mbox{\boldmath $B$}_{II}(\mbox{\boldmath $x$},t) &=&\frac{-2}{V^{\frac{1}{2}}}\left[(\mbox{\boldmath $k$}_c\times\mbox{\boldmath $\hat\varepsilon$}_{k_c\mu_c})c_0\sin{\mit \Theta}_c + (\mbox{\boldmath $k$}_d\times\mbox{\boldmath $\hat\varepsilon$}_{k_d\mu_d})d_0 \sin{\mit \Theta}_d\right] +\frac{\mbox{\boldmath $v$}_{II}(\mbox{\boldmath $x$})}{V^{\frac{1}{2}}}, \nonumber\\ \mbox{\boldmath $I$}_{II}(\mbox{\boldmath $x$},t)&=&\frac{\hbar c^2}{2V}\left[\mbox{\boldmath $k$}_c(1+\cos\phi)+\mbox{\boldmath $k$}_d(1-\cos\phi) -\mbox{\boldmath $k$}_c(1+\cos\phi)\cos2{\mit \Theta}_c+ \mbox{\boldmath $k$}_d(1-\cos\phi)\cos 2{\mit \Theta}_d)\right] \nonumber\\ &&-\frac{\mbox{\boldmath $f$}_{II}(\mbox{\boldmath $x$})\mbox{\boldmath $g$}_{II}(\mbox{\boldmath $x$},t)}{V}, \label{INTII} \end{eqnarray} with \begin{eqnarray} \mbox{\boldmath $u$}_{II}(\mbox{\boldmath $x$})&=&\!\!\!\!\!\sum_{\stackrel{\scriptstyle{k\mu}}{k\neq \pm k_c,\pm k_d}}\!\!\!\!\!\mbox{\boldmath $\hat\varepsilon$}_{k\mu}\qke^{i\kb.\xb},\;\;\;\;\; \mbox{\boldmath $v$}_{II}(\mbox{\boldmath $x$})=i\!\!\!\!\!\sum_{\stackrel{\scriptstyle{k\mu}}{k\neq \pm k_c,\pm k_d}}\!\!\!\!\!(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\qke^{i\kb.\xb}=\nabla\times \mbox{\boldmath $u$}_{II}(\mbox{\boldmath $x$}), \label{VXII}\\ \mbox{\boldmath $f$}_{II}(\mbox{\boldmath $x$})&=&\frac{i\hbar c^2}{V}\!\!\!\!\!\sum_{\stackrel{\scriptstyle{k\mu}}{k\neq \pm k_c,\pm k_d}}\!\!\!\!\!\mbox{\boldmath $\hat\varepsilon$}_{k_0\mu_0}\times(\mbox{\boldmath $k$}\times\mbox{\boldmath $\hat\varepsilon$}_{k\mu})\qke^{i\kb.\xb},\; \mbox{\boldmath $g$}_{II}(\mbox{\boldmath $x$},t)=(1+\cos\phi)\sin{\mit \Theta}_c \\ \nonumber &&\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;+(1-\cos\phi)\sin{\mit \Theta}_d, \end{eqnarray} and with ${\mit \Theta}_c=\mbox{\boldmath $k$}_c.\mbox{\boldmath $x$}-\omega_c t-\chi_0$ and ${\mit \Theta}_d=\mbox{\boldmath $k$}_d.\mbox{\boldmath $x$}-\omega_d t -\xi_0$. The wavefunction and the beables clearly show interference. For example, for $\phi=0$ the $d$-beam is extinguished and for $\phi=\pi$ the $c$-beam is extinguished by interference. \section{Comments on some other recent experimental tests of complementarity\label{CGBAS}} In the proposed experiment of Ghose {\it et al} \cite{GHOSE91}, light is incident on a prism at an angle greater than the critical angle and hence undergoes total internal reflection. A second prism placed less than a wavelength from the first allows light to tunnel into the transmitted channel. Quantum mechanics predicts perfect anticoincidence. This is interpreted by Ghose {\it et al}, as is usual, as which-path information and hence as particle behaviour. Transmitted photons necessarily tunnel through the gap between the prisms, a phenomenon which the authors interpret as wave behaviour. In this way, the authors claim that wave and particle behaviour are observed in the same experiment in contradiction to Bohr's principle of complementarity. This experiment has since been performed by Mizobuchi {\it et al} \cite{MIZ92} using a GRA single photon source, but as we mentioned earlier, the statistical accuracy of their results has been questioned in references \cite{UNNIK, GHOSE99, BRIDA04}. To resolve the technical difficulties with Mizobuchi {\it et al}'s experiment, Brida {\it et al}, following a suggested experiment by Ghose \cite{GHOSE99} and also employing the GRA single photon source, used a birefringent crystal to split a light beam into two beams (the ordinary and the extraordinary beams) instead of using tunneling between two closely spaced prisms. They interpreted the birefringent splitting as wave behaviour, while the perfect anticoincidence they observed they interpreted as particle behaviour. Again, the claim is the observation of wave and particle behaviour in the same experiment in contradiction of complementarity. Afshar's experiment is of the two-slit type. He first observes interference a short distance in front of the slits and determines the position of the dark fringes. He then replaces the screen with a wire grid such that the grid wires coincide with the dark fringes. A lens is placed after the grid to form an image of the two slits. The images showed no loss of sharpness or intensity as compared to the image of the two slits without the grid in position. Afshar concluded that there was interference prior to formation of the image which he interpretes as wave behaviour. He assumes that the images of the slits are formed by photons coming from the slit on the same side as the image. He then interpretes image formation as providing path information, and hence particle behaviour. Ashar concludes that particle and wave behaviour is observed in the same experiment in contradiction of complementarity. We do not agree that these experiments either disprove Bohr's principle of complementarity, or, as argued by Brida {\it et al},that they can be viewed as a generalization of Bohr's principle of complementarity. Our reasons follow. As for the GRA experiments, all the above experiments can be explained using CIEM, i.e., they can be explained entirely in terms of a wave model. One is therefore not forced to conclude that these experiments require a generalization of Bohr's principle of complementarity (a generalization first suggested by Wootters and Zurek \cite{WZ} as mentioned in the introduction), a generalization which is severely flawed, as mentioned in the introduction. We will comment further below. Arguments from the perspective of complementarity can be put to show that these experiments do not disprove complementarity. Let us first consider the experiments of Mizobuchi {\it et al} and Brida {\it et al}. Bohr emphasized that only the final experimental result (pointer reading) has physical significance and that an experiment should be viewed as a whole, not further analyzable \cite{BR59A, BR28}. We recall the statement of Wheeler, `No phenomenon is a phenomenon until it is an observed phenomenon' (\cite{WHR78}, p 14). In these two experiments, the observed results are anticoincidence detections which the above authors and advocates of complementarity or its variants can reasonably and unambiguously attribute to particle behaviour. The wave behaviour is not detected. It is therefore perfectly consistent for a Bohrian to maintain that the experiments unambiguously define a particle model even if this is counter-intuitive. The Afshar experiment avoids this criticism because the presence of the wire grid physically detects the interference. But, the Afshar experiment still fails because of the first point above, namely that CIEM provides a wave model of image formation by a large series of single photon detections. Another point to consider is that the mutually exclusive wave and particle complementary concepts are not related to the mathematical formalism of the quantum theory. In this way they differ from complementary concepts such as position and momentum or the components of angular momentum which are not mutually exclusive classical concepts and are represented in the mathematical formalism of the quantum theory by Heisenberg uncertainty relations. In this case, what is called wave or particle behaviour in a given experiment is somewhat arbitrary. Apart from other points, this arbitrariness is an important reason why we feel complementarity can neither be proved nor disproved. We now comment on a widely accepted generalization of complementarity by Wootters and Zurek in their influential article \cite{WZ}. This generalization admits partial wave and partial particle behaviour in the same experiment. Based on this generalization Wootters and Zurek \cite{WZ}, and later Yasin and Greenberger \cite{GY}, cast particle-wave duality in mathematical form. We have argued in earlier articles \cite{KPW} that far from being a generalization of complementarity, this approach in fact contradicts complementarity. From the mathematical perspective, these mathematical relations are constructs appended to the formalism of the quantum theory but not derived from it. As a measure of coherence they can be thought of as useful heuristic rules, but for the reasons we will give, can be attributed no more fundamental significance than this. For detailed arguments against this generalization we refer the reader to reference \cite{KPW} and restrict ourselves here to briefly emphasizing aspects of complementarity which demonstrate our point of view. In his explanations of his principle of complementarity \cite{BR59A, JAM74,BR28}, Bohr repeatedly emphasized the mutual exclusiveness of complementary concepts, and the requirement of mutually exclusive experimental arrangements for their correct use or definition. He further emphasized that complementary concepts are abstractions to aid thought, and cannot be attributed physical reality. It seems to the present author that Bohr was concerned to provide a framework for the correct use of classical language or concepts. Thus, for the same physical object to be both a wave and a particle is, quite simply, a contradiction of definitions. This, the present author believes, is what led Bohr to emphasize that complementary concepts could not be attributed physical reality. By insisting on mutually exclusive experimental arrangements for the realization of complementary concepts, Bohr, in the authors view, allowed for the use of classical language/concepts in a way that avoids contradiction. It is for these reasons that we regard the Wootters and Zurek generalization of complementarity in terms of partial particle behaviour/knowledge and partial wave beaviour/knowledge as the complete antithesis of Bohr's principle of complementarity. Even apart from Bohr's teachings, what can it mean for a physical object to be partially a wave and partially a particle? Above, we made a distinction between particle and wave complementary concepts and other pairs of complementary concepts that Bohr did not make. Our arguments here need not apply to complementary concepts such as position and momentum, which classically are not mutually exclusive concepts. We note two things: First, the Wootters and Zurek generalization of complementarity is in terms of wave and particle concepts. Second, from the point of view of interpretation, particle and wave complementary concepts are the most fundamental, and lie at the heart of the interpretational issues of the quantum theory. The experiment of Kim {\it et al} concerns both complementarity and the Wheeler delayed-choice issue, but its significance goes beyond these issues. The results of this experiment appear to suggest that a present measurement affects a past measurement. The Wheeler delayed-choice experiments indicate that a present measurement either creates or changes the past history leading to a particular result (there are subtle differences between Wheeler's and Bohr's position which are discussed in reference \cite{K05} section 1). The Kim {\it et al} and Wheeler delayed-choice experiment differ in that the past history is not actually observed in Wheeler's experiment, whereas in Kim et al's experiment it is the result of an actual past measurement that is changed by a measurement in the present. We will leave a detailed discussion of this experiment for a later article, but make one observation. The experiment uses a pair of correlated photons produced by the process of spontaneous parametric down conversion. By detecting the photon partner {\it after} the first photon is detected, the earlier measured wave or particle behaviour of the first photon is determined. What seems to have been left out of the Kim {\it et al} analysis is that once the first photon is detected and the state of the EPR partner changes accordingly, thereafter, the EPR correlation is broken. Hence, any measurement performed on the second photon can have no effect on its partner. This is a firm prediction of quantum mechanics. Nevertheless, the strange result in which a present measurement appears to determine the outcome of an earlier measurement needs explanation. Other articles relating to this issue can be found in reference \cite{QE}. \section{Conclusion} Their ingenious gating system allowed GRA to test, perhaps for the first time, quantum mechanical predictions for a single photon state. Interference is confirmed in the obvious way. The which-path predictions are also confirmed; the photon is detected in only one path. What we have shown though, is that a wave model (CIEM) can explain this result. It cannot therefore be concluded that the detection of the photon on one path confirms particle behaviour. In a particle model, the photon takes one path at the beam-splitter and is detected in that path, whereas in our wave model the photon splits at the beam-splitter, is nonlocally absorbed, and is again detected in only one path. Since the which-path measurement does not confirm particle behaviour, Bohr's principle of complementarity is also not confirmed, contrary to what is claimed by GRA. We conclude then, that GRA's experiments do not confirm complementarity. We may further add that if complementary is accepted, Wheeler's delayed-choice experiments lead to very strange conclusions: either history is changed at the time of measurement, or history is created at the time of measurement \cite{K05, BDH85}. CIEM, on the other hand, explains Wheeler's delayed-choice experiments in a unique and causal way. \section{Acknowledgements} I would like to thank Dr H.V. Mweene and Mr Y. Banda for proof reading my article. I would also like to thank the Dean of Natural Sciences, Dr S.F. Banda, for granting me a short study leave to complete this article.
{ "timestamp": "2006-08-29T17:54:33", "yymm": "0503", "arxiv_id": "quant-ph/0503201", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503201" }
\section{Background and notation} We begin by defining the class of matrix function algebras we will study in this paper. From directed graphs, they arise as the left regular representation of the directed graphs with $n$ vertices and $n$ edges connecting each successive vertex in turn, to form a single loop, or $n$-cycle. We will use the notation $\mathcal{T}^{+}(\mathcal{C}_n)$ for these algebras, where $n$ is the length of the cycle in the algebra. We can view $\mathcal{T}^{+}(\mathcal{C}_n)$ as a matrix function algebra of the form \[ \begin{bmatrix} f_{1,1}(z^n) & z f_{1,2}(z^n) & z^2 f_{1,3}(z^n) & \cdots & z^{n-1}f_{1,n}(z^n) \\ z^{n-1} f_{2,1}(z^n) & f_{2,2}(z^n) & z f_{2,3}(z^n) & \cdots & z^{n-2}f_{2,n}(z^n) \\ z^{n-2}f_{3,1}(z^n) & z^{n-1}f_{3,2}(z^n) & f_{3,3}(z^n) & \cdots & z^{n-3}f_{3,n}(z^n) \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ zf_{n,1}(z^n) & z^2 f_{n,2}(z^n) & z^{3}f_{n,3}(z^n) & \cdots & f_{n,n}(z^n) \end{bmatrix} \] where $f_{i,j} \in A(\mathbb{D})$ for all $ 1\leq i,j \leq n$. These algebras inherit a matricial norm from the matricial norm on $A(\mathbb{D})$ as $\mathcal{T}^{+}(\mathcal{C}_n)$ can be viewed as sitting inside $M_n\otimes A(\mathbb{D})$. We will denote by $A(z^n)$ the algebra $\{ f(z^n): f \in A(\mathbb{D}) \}.$ Notice that $A(z^n)$ is a subalgebra of $A(\mathbb{D})$ for all $n$. In what follows, $A$ will always denote an operator algebra, and by representation we mean a continuous representation of $A$ as an algebra of operators acting on a Hilbert space $\mathcal{H}$. We will denote the elementary matrices with a 1 in the $i$-th diagonal spot by $e_{ii}$. The notation $x = [x_{ij}]$ will denote an $n \times n$ matrix, where the $i$-$j$ entry is $x_{ij}$. We will write $\ell(i,j)$ for the formula $|i-j|(\mod{n})$. So that the above matrix form of $\mathcal{T}^{+}(\mathcal{C}_n)$ can be written as \[ \{[z^{\ell(i,j)}f_{i,j}(z^n)]: f_{i,j} \in A(\mathbb{D}) \mbox{ for all } 1 \leq i,j \leq n\} . \] Lastly, for $1 \leq i \leq n-1$ define $Z_i \in \mathcal{T}^{+}(\mathcal{C}_n)$ as the matrix with $z$ in the $i$-$(i+1)$ position and zeroes elsewhere. Define $Z_n \in \mathcal{T}^{+}(\mathcal{C}_n)$ as the matrix with $z$ in the $n$-$1$ position and zeroes everywhere else. It is not hard to see that $\mathcal{T}^{+}(\mathcal{C}_n)$ is generated by the set $\{ e_{ii}, Z_i: 1 \leq i \leq n \}$. This shorthand will be used later when dealing with specific matrices. \section{Noncommutative point derivations} Some authors take the definition that follows as the definition of derivation, see Chapter 9 in \cite{Paul:2002} for example. We use this notation since we wanted to emphasize the connection between the derivation and the particular representation. This particular definition also emphasizes the connections with point derivations from \cite{Browder:1969} which we exploit in later sections. \begin{defn} Let $\pi: A \rightarrow B(\mathcal{H})$ be a representation of $A$. We say that a continuous linear map $D: A \rightarrow B(\mathcal{H})$ is a {\em point derivation at $\pi$} if $D(ab) = D(a) \pi(b) + \pi(a) D(b)$ for all $a,b \in A$. \end{defn} Of course the function $D(a) = 0$ is a derivation. We refer to this derivation as the trivial, or zero, derivation. We begin by identifying a special class of derivations, of which the trivial derivation is a special case. \begin{defn} For $\pi:A \rightarrow B(\mathcal{H})$ a representation of the operator algebra $A$ and for $X \in B(\mathcal{H})$ we define the function $\delta_X : A \rightarrow B(\mathcal{H})$ by $\delta_X(a) = \pi(a)X-X\pi(a)$ for all $a \in A$.\end{defn} Linearity of $\delta_X$ is obvious. If we let $\{ a_n \}$ be a sequence in $A$, then $ \lim (\pi(a)X - X \pi(a)) = \pi(\lim a_n)X - X \pi(\lim a_n) = \delta_{X}(\lim a_n)$ and hence $\delta_X$ is continuous. We can also see this by noting that $\| \delta_X \| \leq 2 \|\pi \| \|X \|$. \begin{prop} If $\pi: A \rightarrow B(\mathcal{H})$ is a representation and $X \in B(\mathcal{H})$ then the function $\delta_X$ is a continuous derivation at $\pi$. \end{prop} \begin{proof} It remains to show only that $\delta_X$ is a derivation. Let $a,b \in A$, then \begin{align*} \delta_X(ab) & = \pi(ab) X -X \pi(ab) \\ & = \pi(a)\pi(b)X - \pi(a) X \pi(b) + \pi(a) X \pi(b) - X \pi(a) \pi(b) \\ & = \pi(a) \delta_X(b) + \delta_X(a) \pi(b).\end{align*}\end{proof} \begin{defn} Letting $\pi: A \rightarrow B(\mathcal{H})$ be a representation of the operator algebra $A$ we say that a derivation at $\pi$ of the form $\delta_X$ is an {\em inner derivation at $\pi$}.\end{defn} If the range of $\pi$ is $\mathbb{C}$ non-trivial inner derivations do not arise. However, in a noncommutative setting they often do. \begin{prop} Assume that $\pi: A \rightarrow B(\mathcal{H})$ is a representation such that $ \mathop{\mathrm{ran}} \pi$ is not isomorphic to $\mathbb{C}$. There exists $X \in B(\mathcal{H})$ such that $ \delta_X \not\equiv 0$. \end{prop} \begin{proof} It is well known that $B(\mathcal{H})' = \mathbb{C}$. Since, $\mathop{\mathrm{ran}} \pi$ is not isomorphic to $\mathbb{C}$ there exists $a \in A$ such that $ \pi(a) \not\in B(\mathcal{H})'$. In other words, there is $X \in B(\mathcal{H})$ such that $X\pi(a) \neq \pi(a) X$. It follows that $ \delta_X$ is nontrivial.\end{proof} What follows is a theorem that, in certain cases, will allow us to distinguish the inner derivations from other derivations. \begin{thm}\label{inner} Let $\pi: A \rightarrow B(\mathbb{C}^{nk})$ be a representation such that \[ \mathop{\mathrm{ran}} \pi \cong \oplus_{i=1}^n M_k \] where $k \geq 1$ and $n$ is finite. A derivation $D$ at $\pi$ is inner if and only if $D|_{\ker \pi} \equiv 0 $. \end{thm} \begin{proof} Assume first, that $D$ is inner at $\pi$. Then for $a \in \ker \pi$, $D(a) = \pi(a) X- X \pi(a)$ for some $X \in B(\mathcal{H})$. Now $\pi(a) = 0$ and hence $ D(a) = 0$. As $a$ was an arbitrary element of the kernel the forward direction follows. Now suppose that $\ker \pi$ is in the kernel of $D$. We define a map $\widehat{D}: \mathop{\mathrm{ran}} \pi \rightarrow B(\mathcal{H})$ by $ \widehat{D} (\pi(a)) = D(a)$. If $\pi(a) = \pi(b)$, then $ a-b \in \ker \pi$ and hence $D(a-b) = 0$. It follows, by linearity of $D$, that $D(a) = D(b)$ and hence the map $\widehat{D}$ is well defined. Next, \begin{align*} \widehat{D} (\pi(a) \pi(b)) &= \widehat{D}( \pi(ab)) \\ &= D(ab) \\ &= D(a) \pi(b) + \pi(a) D(b) \\ & = \widehat{D}(\pi(a)) \pi(b) + \pi(a) \widehat{D}(\pi(b)). \end{align*} It follows that $\widehat{D}$ defines a derivation on $\mathop{\mathrm{ran}} \pi$. Further, since $\mathop{\mathrm{ran}} \pi$ is finite dimensional it follows that $\widehat{D}$ is continuous. Notice that if $ n=1$ then as $M_k$ is simple every $M_k$ valued derivation is inner, \cite{Kadison-Ringrose:1997}. If $n > 1$ we can use exact sequences of cohomology groups, see \cite{Johnson:1972} to see that a continuous $B(\mathbb{C}^{nk})$-valued derivation on \[ \oplus_{i=1}^n M_k \] is inner. Hence there is $X \in M_{nk}$ such that $\widehat{D}(\pi(a)) = \pi(a) X - X \pi(a)$. Since $\widehat{D}(\pi(a)) = D(a)$ the result now follows. \end{proof} The next two propositions give us a short method of checking whether non-inner derivations can occur at $\pi$. For an ideal $M$, we denote by $M^2$ the algebraic ideal generated by elements of the form $bc$ such that $b, c \in M$. We will denote the norm closure of the ideal $M^2$ by $ \overline{M^2}$. \begin{prop}\label{kernelsquared} If $\ker \pi = \overline{(\ker \pi)^2}$ for a representation $\pi: A \rightarrow B(\mathcal{H})$, then for a continuous derivation $D$ at $\pi$, $D|_{\ker \pi} \equiv 0$. \end{prop} \begin{proof} Let $a = bc$ where $b,c \in \ker \pi$. Then, \begin{align*} D(a) &= D(bc) \\ &= \pi(b) D(c) + D(b) \pi(c) \\ & = 0 D(c) + D(b) 0 \\ & = 0. \end{align*} Since $(\ker \pi)^2$ is the ideal generated by elements of the form $bc$ where $b,c \in \ker \pi$ it follows that $D|_{(\ker \pi)^2} \equiv 0 $. Now, continuity of $D$ yields the result. \end{proof} \begin{prop}\label{approximateidentity} If the kernel of the representation $\pi: A \rightarrow B(\mathcal{H})$ has a bounded left (right) approximate identity then any continuous derivation $D$ at $\pi$ is identically zero on $\ker \pi$. \end{prop} \begin{proof} Let $\{ e_{\lambda} \}$ be a bounded left (right) approximate identity in $\ker \pi$. Then for any $f \in \ker \pi$ we know that $ \lim e_{\lambda} f = f$. But notice that $ e_{\lambda} f \in (\ker \pi)^2$ and hence $D(e_{\lambda} f) = 0$ for all $ \lambda$. As $D$ is continuous it follows that $D(f)= 0$. As $f$ was arbitrary the result follows.\end{proof} \begin{cor} If $A$ is a $C^*$-algebra and $\pi$ is a $*$-representation then every derivation at $\pi$ is identically zero on $\ker \pi$. \end{cor} \begin{proof} It is well known that the kernel of a $*$-representation is a $*$-ideal. Further every $*$-ideal in a $C^*$-algebra is a $C^*$-algebra and hence has an approximate identity. The result now follows. \end{proof} \section{Point derivations on $\mathcal{T}^{+}(\mathcal{C}_n)$} \begin{defn} For $\lambda \in \overline{\mathbb{D}}$ we define the representation $\varphi_{\lambda}: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow M_n$ by \[ \varphi_{\lambda} ( [z^{\ell(i,j)}f_{i,j}(z^n)]) = [ {\lambda}^{\ell(i,j)}f_{i,j}({\lambda}^n)]. \] \end{defn} Notice that for $\lambda \neq 0$ the range of $\varphi_{\lambda}$ is isomorphic to $M_n$. It follows that $\ker(\varphi_{\lambda})$ is a maximal ideal of type $\lambda$, see \cite{Alaimia:1999}. In contrast, the range of $\varphi_0$ is the diagonal matrices in $M_n$. It is not the case that the kernel of $ \varphi_0$ is a maximal ideal. The representations of the form $\varphi_{\lambda}$ are enough to ensure semisimplicity of $\mathcal{T}^{+}(\mathcal{C}_n)$ a well known result for certain graph operator algebras, see \cite{Davidson-Katsoulis:2004}, or \cite{Jury-Kribs:2004}, and semicrossed products, see \cite{Pet:1988}. We include the proof in this context for completeness, and since it is not difficult. \begin{prop} The algebras $\mathcal{T}^{+}(\mathcal{C}_n)$ are semisimple. \end{prop} \begin{proof} Let \[ a = [z^{\ell(i,j)}f_{i,j}(z^n)]. \] Assume that $ \varphi_{\lambda} (a) = 0$ for all $0 < |\lambda| < 1$. Then in particular, $f_{i,j}(\lambda^n) = 0$ for all $ 0< |\lambda| < 1$. But since $ f_{i,j}(z^n)$ is analytic in $\mathbb{D}$ and identically zero on a set containing a limit point in $\mathbb{D}$ then $f_{i,j}(z^n) \equiv 0$ for all $i,j$. Hence $a = (0)$ and the result follows. \end{proof} We now define another important class of representations. \begin{defn} For $1 \leq i \leq n$ define the representation $\varphi_{i,0}: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow \mathbb{C}$ by \[ \varphi_{i,0} ( [z^{\ell(i,j)}f_{i,j}(z^n)]) = f_{i,i}(0). \]\end{defn} For these representations, the range is $\mathbb{C}$ and hence the kernels give rise to maximal ideals which, in the notation of \cite{Alaimia:1999}, are of type $0$. Notice also, that since the range is $\mathbb{C}$ there will be no inner derivations at $\varphi_{0,i}$ for all $1 \leq i \leq n$. More is actually true. \begin{prop} For $n \geq 2$, there is no nontrivial point derivation at $\varphi_{0,i}: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow M_n$, where $1 \leq i \leq n$.\end{prop} \begin{proof} We will prove the result for $i = 1$, the general case proceeds in a similar fashion. A simple calculation tells us that \[ \ker \varphi_{0,1} = \left\{ \begin{bmatrix} z^nf_{1,1}(z^n) & z f_{1,2}(z^n) & z^2 f_{1,3}(z^n) & \cdots & z^{n-1}f_{1,n}(z^n) \\ z^{n-1} f_{2,1}(z^n) & f_{2,2}(z^n) & z f_{2,3}(z^n) & \cdots & z^{n-2}f_{2,n}(z^n) \\ z^{n-2}f_{3,1}(z^n) & z^{n-1}f_{3,2}(z^n) & f_{3,3}(z^n) & \cdots & z^{n-3}f_{3,n}(z^n) \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ zf_{n,1}(z^n) & z^2 f_{n,2}(z^n) & z^{3}f_{n,3}(z^n) & \cdots & f_{n,n}(z^n) \end{bmatrix} \right\} \] where $ f_{i,j} \in A(\mathbb{D}) $ for all $i,j$. Multiplying two general elements of $\ker \varphi_{0,1}$ together one can verify that $\ker \varphi_{0,1} = (\ker \varphi_{0,1})^2$. Using Proposition \ref{kernelsquared} together with Theorem \ref{inner} we know that every derivation at $\varphi_{0,1}$ is inner. But since $\mathop{\mathrm{ran}} \varphi_{0,i} = \mathbb{C}$ any inner derivation is the zero derivation. The result now follows.\end{proof} Unlike the previous class of representations, the representations $\varphi_{\lambda}$ give rise to derivations which are inner at $ \varphi_{\lambda}$. It is the derivations which are not inner at $ \varphi_{\lambda}$ which interest us so we now look at what values of $ \lambda$ give rise to derivations which are not inner. \begin{thm} For $|\lambda|<1$ there exist non-inner derivations at $\varphi_{\lambda}: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow M_n$. \end{thm} \begin{proof} We begin by noticing that the map $F: A(\mathbb{D}) \rightarrow \mathbb{C}$ given by $ F(f) = f'(\lambda)$ is a continuous linear functional and hence completely continuous \cite[Corollary 2.2.3]{Eff-Ruan:2000}. In particular we know that the matricial map $F_{n, \lambda}: M_n \otimes A(\mathbb{D}) \rightarrow M_n$, given by $F_{n, \lambda}([f_{ij}]) = [f'_{ij}(\lambda)]$ is continuous. Now as $\mathcal{T}^{+}(\mathcal{C}_n)$ is a subalgebra of $M_n \otimes A(\mathbb{D})$ we know that $F_{n, \lambda}$ restricted to $\mathcal{T}^{+}(\mathcal{C}_n)$ yields a continuous linear map. We need only show that $F_{n, \lambda}$ is a non-inner derivation at $\varphi_{\lambda}$. To see that it is a derivation we will look at $F_n$ applied to $M_n \otimes A(\mathbb{D})$. In particular, choose two elements $f = [f_{ij}],g = [g_{ij}] \in M_n \otimes A(\mathbb{D})$. Now notice that \begin{align*} F_{n, \lambda}(fg) & = F_{n, \lambda} \left[ \sum_{j=1}^n f_{ij}g_{jk} \right] \\ & = \sum_{j=1}^n [ f'_{i,j}( \lambda)g_{jk}(\lambda) + f_{ij}(\lambda) g'_{jk}(\lambda)] \\ & = \left( \sum_{j=1}^n [ f'_{i,j}( \lambda)g_{jk}(\lambda)]\right) + \left( \sum_{j=1}^n [f_{ij}(\lambda) g'_{jk}(\lambda)] \right) \\ & = [f'_{ij}(\lambda)][g_{ij}(\lambda)] + [f_{ij}(\lambda)][g'_{ij}(\lambda)] \\ & = F_{n, \lambda}(f)\varphi_{\lambda}(g) + \varphi_{\lambda}(f)F_{n, \lambda}(g). \end{align*} Restricting to $\mathcal{T}^{+}(\mathcal{C}_n)$ will not affect the derivation property and hence $F_{n,\lambda}$ yields a derivation at $ \varphi_{\lambda}$. Recall the definition of $Z_i$ as the matrix with a z in the $i$-$(i+1)$ position for $ 1 \leq i \leq n-1$, or the $n$-$1$ position for $i=n$ and zeroes elsewhere. For $\lambda = 0$ we see that $F_{n, 0}$ is not inner since $F_{n, 0}(Z_i) \neq 0$ for all $i$ and yet $\varphi_{0}(Z_i) = 0$, applying Theorem \ref{inner} verifies the result. For $0 < |\lambda| <1$ let $f = z- (\lambda)^n$. Notice that $f$ is an analytic function such that $f'(\lambda^n) \neq 0$ and yet $f(\lambda^n) = 0$. Now let $\tilde{f}$ be the element of $\mathcal{T}^{+}(\mathcal{C}_n)$ given by $[z^{\ell(i,j)}f(z^n)]$. Notice that $\tilde{f} \in \ker \varphi_{\lambda}$. However, $F_{n, \lambda}(\tilde{f}) = [(\lambda)^{\ell(i,j)}n \lambda^{n-1}] \neq 0$. The result now follows as in the case of $\lambda = 0$.\end{proof} In the special case of point derivations at $\varphi_0$ we are able to show more. In analogy with a description of certain homology groups for the quiver algebras corresponding to a single vertex and countable edges in \cite{Pop:1998a}, we now show a certain amount of uniqueness for derivations at $ \varphi_0$. \begin{prop} Let $D$ be a point derivation at $\varphi_0.$ Then $D$ can be written as $D_0+D_1$ where $D_0$ is inner at $ \varphi_0$, $D_1$ is a point derivation at $ \varphi_0$, and $D_1(a) \neq 0$ guarantees that $a \in \ker \varphi_0$. Further, $D_1$ is uniquely determined by the numbers $D_1(Z_i)$ with $ 1 \leq i \leq n$ \end{prop} \begin{proof} Let $D: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow M_n$ be a point derivation at $\varphi_0$. Notice that $\mathop{\mathrm{ran}} \varphi_0$ is finite dimensional and hence $\ker \varphi_0$ has a Banach space complement in $\mathcal{T}^{+}(\mathcal{C}_n)$ which we will denote by $ ( \ker \varphi_0)^c$. Further, every $ a \in \mathcal{T}^{+}(\mathcal{C}_n)$ can be written uniquely as $x_a + y_a$ where $ x_a \in (\ker \varphi_0)^c$ and $ y_a \in \ker \varphi_0$. Now there exist $ \lambda _i$ such that $ x_a = \displaystyle{ \sum_{i=1}^n \lambda_ie_{ii} }$ where $e_{ii}$ is the elementary matrix with $1$ in the $i$-$i$ position and zero everywhere else. We claim that if $ a ,b \in \mathcal{T}^{+}(\mathcal{C}_n)$ then, with respect to the decomposition above, $ \varphi_0(x_ax_b) = 0 $ if and only if $x_a x_b = 0$. Writing $x_a = \sum_{i=1}^n \lambda_ie_{ii}$ and $x_b = \sum_{i=1}^n \mu_ie_{ii}$ then, \[ x_a x_b = \sum_{i=1}^n \lambda_i \mu_i e_{ii} \] and the claim follows. Define the map $D_1: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow M_n$ by letting $D_1(x_a + y_a) = D(y_a)$ with respect to the above decomposition. We will use the claim in the previous paragraph to show that $D_1$ is a derivation at $\varphi_{0}$. Linearity, and continuity are clear. We need only establish the derivation property. Now \begin{align*} D_1(ab) & = D_1((x_a + y_a)(x_b+y_b)) \\ &= D_1(x_ax_b + y_ax_b + x_ay_b + y_ay_b) \\ & = D(y_ax_b) + D(x_ay_b) \\ &= D(y_a) \varphi_0(x_b) + \varphi_0(y_a)D(x_b) + D(x_a)\varphi_0(y_b) + \varphi_0(x_a)D(y_b) \\ & = D(y_a) \varphi_0(x_b) + \varphi_0(x_a)D(y_b) \\ &= D(y_a) \varphi_0 (x_b + y_b) + \varphi_0(x_a) D_1(x_b + y_b) \\ & = D_1(x_a + y_a) \varphi_0(x_b + y_b) + \varphi_0(x_a + y_a) D_1(x_b + y_b) \\ &= D_1(a) \varphi_0(b) + \varphi_0(a) D_1(b) \end{align*} and hence $ D_1$ is a derivation at $ \varphi_0$. Notice that $D_0 = D-D_1$ is an inner derivation since $D-D_1|_{\ker \varphi_0} = 0$. It follows that every point derivation at $ \varphi_0$ can be written as an inner derivation and a derivation which sends $ (\ker \varphi_0)^c$ to zero. Notice that each derivation of the form $D_1$ is uniquely determined by the value on $\ker \varphi_0 \setminus \overline{(\ker \varphi_0)^2}$. A technical calculation shows us that the set $\ker \varphi_0 \setminus \overline{(\ker \varphi_0)^2}$ is given by \[ \{ \lambda_i Z_i: 1 \leq i \leq n \}. \] The result now follows. \end{proof} The previous result relies on a nice decomposition of every element of $\mathcal{T}^{+}(\mathcal{C}_n)$ which is invariant under derivations. Although we expect a similar result for the point derivations at $ \varphi_{\lambda}$ for all $0 < |\lambda|< 1$ we have not been able to prove such a result. \begin{thm}\label{inneratt} For $\lambda \in \mathbb{T}$ every derivation at $\varphi_{\lambda}: A \rightarrow M_n$ is inner.\end{thm} \begin{proof} We will show that $\ker \varphi_{\lambda}$ has a bounded approximate identity and then apply Proposition \ref{approximateidentity} and Theorem \ref{inner}. We let $ \pi_n: A(\mathbb{D}) \rightarrow A(\mathbb{D})$ be the contractive homomorphism induced by sending $ z \mapsto z^n$. Denote the range of this map by $A(z^n)$ which matches our previous definition of $A(z^n)$. Further $\pi_n$ is a contractive isomorphism onto $A(z^n)$. (We are not making any claims about contractivity of the reverse map). Notice that \[ \pi_n (\{ f \in A(\mathbb{D}): f(\lambda) = 0 \}) \subseteq \{ g \in A(z^n): g(\lambda^{\frac{1}{n}}) = 0 \}.\] Further, since $ | \lambda | = 1$ we know that there is a uniformly bounded net, see \cite[Section 1.6]{Browder:1969}, \[ \{ f_{\iota} \} \subseteq \{ f \in A(\mathbb{D}): f(\lambda) = 0 \}\] such that $ f_{\iota} g \rightarrow g$ for all \[ g \in \{ f \in A(\mathbb{D}): f(\lambda) = 0 \}.\] Notice that $ \| \pi_n(f_{\iota}) \| \leq \| f_{\iota} \|$ and hence $ \{ \pi_n(f_{\iota}) \} $ is a bounded net in $\{ g \in A(z^n): g(\lambda^{\frac{1}{n}}) = 0 \}$. Now if $g(\lambda^{\frac{1}{n}}) = 0$ and $g \in A(z^n)$ then \[ h = \pi_n^{-1}(g) \in \{ f \in A(\mathbb{D}): f(\lambda) = 0 \}.\] It follows that $ f_{\iota} h \rightarrow h$. Now $ \pi_n(f_{\iota} h) \rightarrow g$ and hence the ideal \[ \{ g \in A(z^n): g(\lambda^{\frac{1}{n}}) = 0 \}\] has a bounded approximate identity. We define the net $\{ F_{\iota} \}$ to be the diagonal matrices with $\pi_n(f_{\iota})$ along the diagonals. Now $ \{ F_{\iota} \}$ is a bounded net as $\{ f_{\iota} \}$ is. Further, $F_{\iota} \in \ker \varphi_{\lambda}$ for all $\iota$. It is not hard to see that $\{ F_{\iota} \}$ is an approximate identity in $\ker \varphi_{\lambda}$.\end{proof} We will use this theorem to show the main result in this paper, that every $\mathcal{T}^{+}(\mathcal{C}_n)$-valued derivation is inner. \section{Derivations on $\mathcal{T}^{+}(\mathcal{C}_n)$} We begin with an elementary lemma relating derivations and point derivations. \begin{lem} Let $D: A \rightarrow A$ be a continuous derivation on the operator algebra $A$. For a representation $\pi: A \rightarrow B(\mathcal{H})$, the map $\pi \circ D : A \rightarrow B(\mathcal{H})$ is a continuous derivation at $\pi$.\end{lem} \begin{proof} Since $ \pi \circ D$ is a composition of continuous linear maps, it follows that $ \pi \circ D$ is a continuous linear map. Now let $a, b \in A$. Then \begin{align*} \pi \circ D (ab) & = \pi ( D(a)b + aD(b)) \\ & = \pi(D(a))\pi(b) + \pi(a) \pi(D(b)) \\ &= \pi \circ D (a) \pi(b) + \pi(a) \pi \circ D (b). \end{align*} It follows that $\pi \circ D$ is a derivation at $\pi$.\end{proof} \begin{defn} Let $\pi: A \rightarrow B(\mathcal{H})$ be a representation and $D: A \rightarrow A$ be a continuous derivation. We say that $D$ is {\em locally inner at $\pi$} if $\pi \circ D$ is inner at $\pi$.\end{defn} We are now in a position to tackle the main theorem of this paper. Showing that every $\mathcal{T}^{+}(\mathcal{C}_n)$-valued derivation on $\mathcal{T}^{+}(\mathcal{C}_n)$ is inner will be a simple corollary. \begin{thm}\label{locallyinner} Let $D: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow \mathcal{T}^{+}(\mathcal{C}_n)$ be a continuous derivation which is locally inner at $\varphi_{\lambda}$ for all $\lambda \in \mathbb{T}$, then $D$ is inner.\end{thm} \begin{proof} For $\lambda \in \mathbb{T}$ let $D_{\lambda}:= \varphi_{\lambda} \circ D$. Then, by hypothesis, $D_{\lambda}(a)$ can be written as $ X_{\lambda} \varphi_{\lambda}(a) - \varphi_{\lambda}(a) X_{\lambda}$ for all $a \in \mathcal{T}^{+}(\mathcal{C}_n)$. Notice that as $e_{ii} \in \mathcal{T}^{+}(\mathcal{C}_n)$ it follows that $D(e_{ii}) \in \mathcal{T}^{+}(\mathcal{C}_n)$. By hypothesis $D_{\lambda}(e_{ii}) = X_{\lambda}e_{ii} - e_{ii}X_{\lambda}$. Now $X_{\lambda}e_{ii} - e_{ii}X_{\lambda}$ is the matrix with $(-X_{\lambda})_{ij} $ in the $i$-$j$ position, for $i\neq 0$, $(X_{\lambda})_{ji}$ in the $j$-$i$ position for $i \neq j$ and 0 elsewhere. In particular, the matrix $Y$ with $0$ on the diagonal such that $\varphi_{\lambda}(Y_{ij}) = (X_{\lambda})_{ij}$ off the diagonal for all $\lambda \in \mathbb{T}$ is an element of $ \mathcal{T}^{+}(\mathcal{C}_n)$. Recall the definition of the matrices $Z_i$. Now $D_{\lambda}(Z_i) = \lambda(X_{\lambda} e_{i,i+1} - e_{i,i+1}X_{\lambda})$ for all $ \lambda \in \mathbb{T}$ where we define $e_{n, n+1}$ to mean $e_{n,1}$. Now the $i$-$(i+1)$ entry of $D_{\lambda}(Z_i)$ is $ \lambda (-X_{i+1,i+1}(\lambda) + X_{i,i}(\lambda))$, and hence for all $i$, $X_{i+1,i+1}(\lambda) - X_{i,i}(\lambda)$ defines an element of $\mathcal{T}^{+}(\mathcal{C}_n)$, call it $X_i$. Now define a diagonal matrix $X'$ by letting the $i$-$i$ entry be $X_i - X_1$. Define $X = Y + X'$ which is in $\mathcal{T}^{+}(\mathcal{C}_n)$. Further, $D_{\lambda}(e_{ii}) = \varphi_{\lambda}(e_iiX-Xe_ii)$ and $D_{\lambda}(Z_i) = \varphi_{\lambda}(Z_iX-XZ_i)$. It follows that for any $ a \in \mathcal{T}^{+}(\mathcal{C}_n)$, $ D_{\lambda}(a) = \varphi_{\lambda}(aX-Xa)$ for all $ \lambda \in \mathbb{T}$. Now, every element of $a \in \mathcal{T}^{+}(\mathcal{C}_n)$ is uniquely determined by the values of $ \varphi_{\lambda}(a) $ for $ \lambda \in \mathbb{T}$. It follows that $D(a) = aX-Xa$ and the result is established. \end{proof} \begin{cor}\label{graphinner} Every derivation $D: \mathcal{T}^{+}(\mathcal{C}_n) \rightarrow \mathcal{T}^{+}(\mathcal{C}_n)$ is inner. \end{cor} \begin{proof} First, since $\mathcal{T}^{+}(\mathcal{C}_n)$ is semisimple we know, \cite{Johnson-Sinclair:1968} that every derivation is automatically continuous. We know from Theorem \ref{inneratt} that every continuous derivation at $ \varphi_{\lambda}$ is inner. In particular, $D_{\lambda}$ is continuous and locally inner on $\mathbb{T}$. \end{proof} We remark that, as $A(\mathbb{D})$ is a special case of $\mathcal{T}^{+}(\mathcal{C}_n)$, the above proof is an alternate approach to the fact there are no nontrivial derivations on $A(\mathbb{D})$. It would be interesting to know if the above result can be extended to $\mathcal{L}_{\mathcal{C}_n}$ which is the matrix function algebra as $\mathcal{T}^{+}(\mathcal{C}_n)$ with $A(z^n)$ replaced by $H^{\infty}(z^n)$.
{ "timestamp": "2006-08-22T18:33:58", "yymm": "0503", "arxiv_id": "math/0503643", "language": "en", "url": "https://arxiv.org/abs/math/0503643" }
\section{Introduction} The detection of gravitational waves (GW) from astrophysical sources is one of the most outstanding problems in experimental gravitation today. Large laser interferometric gravitational wave detectors like the LIGO, VIRGO, LISA, TAMA 300, GEO 600 and AIGO are potentially opening a new window for the study of a vast and rich variety of nonlinear curvature phenomena. In recent works \cite{JVD96} we have analyzed the Fourier transform (FT) of the Doppler shifted GW signal from a pulsar with the use of the Plane Wave Expansion in Spherical Harmonics (PWESH). Spherical-harmonic multipole expansions are used throughout theoretical physics. The expansion of a plane wave in spherical harmonics has a variety of applications not only in quantum mechanics and electromagnetic theory \cite{MWIEEE}, but also in many other areas. A number of researchers have used spherical-harmonic expansions for a variety of problems in general relativity, including problems where nonlinearity shows up\cite{KThorne80}. The basis states in the PWESH expansion form a complete set and facilitate such a study. It also turns out that the consequent analysis of the Fourier Transform (FT) of the GW signal from a pulsar has a very interesting and convenient development in terms of the resulting spherical Bessel, generalized hypergeometric function, the Gamma functions and the Legendre functions. Both rotational and orbital motions of the Earth and spindown of the pulsar can be considered in this analysis which happens to have a nice analytic representation for the GW signal in terms of the above special functions. The signal can then be studied as a function of a variety of different parameters associated with both the GW pulsar signal as well as the orbital and rotational parameters. The numerical analysis of this analytical expression for the signal offers a challenge for fast and high performance parallel computation. The plane wave expansion approach was also used by Bruce Allen and Adrian C. Ottewill \cite{AO96} in their study of the correlation of GW signals from ground-based GW detectors. They use the correlation to search for anisotropies from stochastic background in terms of the $l, m$ multipole moments. Our PWESH formalism enables a similar study. Recent studies of the Cosmic Microwave Background Explorer have raised the interesting question of the study of very large multipole moments with angular momentum $l$ and its projection $m$ going up to very large values of $l\sim1000$. Such problems warrant an intensive analytic study supplemented by numerical and parallel computation. Since our FT depends on the Bessel function, a computational issue arises due to large values of the index or order $n$ of the function. In the GW form of the pulsar, the Doppler shifted orbiting motion gives rise to Bessel functions $J_{n}(\frac{2 \pi f_0 A \sin \theta}{c})$, where $\frac{2 \pi f_0 A \sin\theta}{c}$ is large for non-negligible angle $\theta$ as is shown in the following section. Even for $\sin{\theta}\sim\frac{1}{1000}$, the argument is large necessitating the consideration of large values of $n$. The motivation of this work, is to extend the analysis in Watson \cite{Watson} for large index, argument and overlapping situations. Meissel \cite{M1} has made derivations for large order Bessel functions both when the argument is smaller than the order and vice versa. The asymptotics of these large order Bessel functions are tricky in the sense that one runs into so-called ``transition" regions where such expansions fail. These regions are values of the function when the argument is close to the given order. As an application, we will address the phenomenological situation of GW signal analysis of large order $n$ (which does arise with combinations of $l$ and $m$) and supplement the related computations with the presently derived results in a forthcoming paper. Captures of stellar-mass compact objects (CO) by massive black holes are important capture sources for the Laser Interferometer Space Antenna (LISA), the space based GW detector due to be launched in about a decade\cite{PM}. Higher Harmonics of the orbital frequency of the COs arise in the post Newtonian (PN) capture GW model forms and contribute considerably to the total signal to noise (S/N) ratio of the waveform. The GW form can be decomposed into gravitational multipole moments which are treated in the Fourier analysis of Keplerian eccentric orbits. The radiation depends strongly on the orbital eccentricity $e$, and Bessel functions $J_{n}(ne)$ are a natural consequence of the analysis. The calculation of partial derivatives of the potential scattering phase shifts which often contain Bessel and Legendre functions of large order angular momentum $l$, with respect to angular momentum arise in a variety of scattering problems in atomic, molecular and nuclear physics. In particular, large values of $l$ can arise in rainbow, glory and orbit scattering. The analysis in our paper should help provide suitable approximations for large order and/or argument for the Bessel functions that arise in such problems. \section{Fourier Transform of the GW signal} The FT for the GW Doppler shifted pulsar signal \cite{JVD96} is given as follows: \begin{eqnarray} \widetilde{h}(f)=S_{n l m}(\omega _{0},\omega _{orb},T_{rE},n,l,m,A,R,k,\alpha ,\theta ,\phi)= \nonumber \end{eqnarray} \\ \begin{eqnarray} {\sum_{n=-\infty}^{\infty}\sum_{l=0}^{\infty}\sum_{m=-l}^{l}\psi_0 \psi_1 \psi_2 \psi_3 \psi_4} \end{eqnarray} where \begin{equation} \psi_0(n,l,m,\alpha,\theta,\phi)= 4\pi i^{l}Y_{l m}(\theta ,\phi )N_{l m}P_{l }^{m}(\cos \alpha ) \end{equation} \begin{eqnarray} \psi_1(n,\theta, \phi, T_{rE}, f_0, A)=T_{rE}\sqrt{\frac{\pi}{2}}e^{-i\frac{2\pi f_{0}A}{c}\sin \theta \cos \phi } i^{n}e^{-in\phi }\nonumber \end{eqnarray} \begin{eqnarray} \times J_{n}\left(\frac{2\pi f_{0}A\sin \theta }{c}\right) \end{eqnarray} \begin{equation} \psi_2(l,\omega_{orb}, \omega_{r}, n, m, R)=\left\{\frac{1-e^{i\pi (l -B_{orb})R}}{1-e^{i\pi (l -B_{orb})}} \right\} \frac{e^{-iB_{orb}\frac{\pi}{2}}}{2^{2l}} \end{equation} \begin{eqnarray} \psi_3(k,l,m,n,\omega_{orb}, \omega_r)=k^{l+\frac{1}{2}}\nonumber\\ \times \frac{\Gamma \left(l +1\right) }{\Gamma \left(l +\frac{3}{2}\right)\Gamma \left(\frac{l +B_{orb}+2}{2}\right)\Gamma \left(\frac{l -B_{orb}+2}{2}\right)} \end{eqnarray} \begin{eqnarray} \psi_4(k,l,m,n,\omega_{orb}, \omega_r)=_1F_{3}(l +1;l+\frac{3}{2}, \nonumber \\ \frac{l+B_{orb}+2}{2},\frac{l-B_{orb}+2}{2};\frac{-k^{2}}{16}) \end{eqnarray} The angle $\alpha$ is the co-latitude detector angle and angles $\theta$, $\phi$ are associated with the pulsar source. Here $\omega_0=2\pi f_0$, $\omega_{orb}=\frac{2\pi}{T_{orb}}$ ($T_{orb}=365$ days, $T_{rE} = 1$ day), $B_{orb}=2\left(\frac{\omega-\omega_0}{\omega_r}+\frac{m}{2}+\frac{n \omega_{orb}}{\omega_{rot}}\right)$, $k=\frac{4\pi f_0 R_E \sin(\alpha)}{c}$ ($R_E$ is the radius of Earth, $c$ is the velocity of light) and $A=1.5 \times 10^{11}$ meters is the sun-earth distance. \section{Extensions of Meissel's and Steepest Descent Expansions} The Bessel function, of the type, $J_{\nu}(x)$ obeys the following differential equation \cite{Watson}, \begin{equation} z^2\frac{d^2J_{\nu}(\nu z)}{d z^2}+z\frac{dJ_{\nu}(\nu z)}{d z}+\nu^2(1-z^2)J_{\nu}(\nu z)=0 \end{equation} where the argument $x$ is parameterized by $\nu z$. If a function $u(z)$ is introduced such that \begin{equation} J_{\nu}(\nu z)=\frac{\nu^{\nu}}{\Gamma (\nu+1)} \exp \left\{\int_{}^{z} u(z) dz \right \} \end{equation} where $u(z)$ is a series in descending powers of $\nu$, \begin{eqnarray} u(z)=\nu u_0 + u_1 + \frac{u_2}{\nu} + \frac{u_3}{\nu^2}+ \frac{u_4}{\nu^3}+\frac{u_5}{\nu^4}+\frac{u_6}{\nu^5}+\frac{u_7}{\nu^6}\nonumber\\ +\frac{u_8}{\nu^7}+\frac{u_9}{\nu^8}+... \end{eqnarray} Substitution of this series and equation (8) in the differential equation (7) yields the following expressions for $u_i(z)$, $i=0...5$, \begin{eqnarray} u_0=\frac{\sqrt{1-z^2}}{z}, u_1=\frac{z}{2(1-z^2)}, u_2=-\frac{4z+z^2}{8(1-z^2)^{5/2}} \nonumber\\ u_3=\frac{4z+10z^3+z^5}{8(1-z^2)^{4}}, u_4=-\frac{64z+560z^3+456z^5+25z^7}{128(1-z^2)^{11/2}} \nonumber\\ u_5=\frac{16z+368z^3+924z^5+347z^7+13z^9}{32(1-z^2)^{7}} \nonumber \end{eqnarray} Hence, by integrating $u_i$, and substituting in Equation (8) we arrive at Meissel's \textit{First} expansion \cite{M1}, which is valid for the case when the argument is less than the order $\nu$. We do not list $u_6,u_7,u_8$ and $u_9$ as one can obtain these straightforwardly from their respective integrals shown below. These results are expressed as, \begin{equation} J_{\nu}(\nu z)=\frac{(\nu z)^{\nu} \exp (\nu \sqrt{1-z^2})\exp(-V_{\nu})}{e^{\nu}\Gamma (\nu+1) (1-z^2)^{1/4}[1+\sqrt{1-z^2}]^{\nu}} \end{equation} where, \begin{equation} V_{\nu}=V_1+V_2+V_3+V_4+V_5+V_6+V_7+V_8+... \end{equation} and, \begin{eqnarray} V_1=\frac{1}{24\nu}\left(\frac{2+3z^2}{(1-z^2)^{3/2}}-2\right),V_2=-\frac{4z^2+z^4}{16\nu^2(1-z^2)^3}\nonumber \end{eqnarray} \begin{eqnarray} V_3=-\frac{1}{5760\nu^3}\left(\frac{16-1512z^2-3654z^4-375z^6}{(1-z^2)^{9/2}}-16\right)\nonumber \end{eqnarray} \begin{eqnarray} V_4=-\frac{32z^2+288z^4+232z^6+13z^8}{128\nu^{4}(1-z^2)^6}\nonumber \end{eqnarray} \begin{eqnarray} V_5=-\frac{1}{322560\nu^5(1-z^2)^{15/2}}(67599\,{z}^{10}+1914210\,{z}^{8}\nonumber\\ +4744640\,{z}^{6}+1891200\,{z}^{4}+78720\,{z}^{2}+256)+\frac{1}{1260\nu^5}\nonumber \end{eqnarray} \begin{eqnarray} V_6=\frac{z^2}{192(1-{z}^{2})^{9}{\nu}^{6}}(48+2580{z}^{2}+14884{z}^{4}\nonumber\\ +17493{z}^{6}+4242{z}^{8}+103{z}^{10})\nonumber \end{eqnarray} \begin{eqnarray} V_7=-\frac{(1-z^2)^{-21/2}}{3440640\nu^7}(881664{z}^{2}+99783936{z}^{4}\nonumber \end{eqnarray} \begin{eqnarray} +1135145088{z}^{6}+2884531440{z}^{8}+1965889800{z}^{10}\nonumber\\ +318291750{z}^{12}+5635995{z}^{14}-2048)-\frac{1}{1680\nu^7}\nonumber \end{eqnarray} \begin{eqnarray} V_8={\frac{z^2}{4096(1-{z}^{2})^{12}{\nu}^{8}}}(1024+248320{z}^{2}+5095936{z}^{4}\nonumber\\ +24059968{z}^{6}+34280896{z}^{8}+15252048{z}^{10}\nonumber\\ +1765936{z}^{12}+23797{z}^{14})\nonumber \end{eqnarray} Hence we have actually increased Meissel's analysis by two orders. Using symbolic packages these orders were computed and higher terms should pose no problem if the application requires higher accuracy. For the case when the argument is larger than the index, Meissel used the parametrization $z=\sec{\beta}$ \cite{M1}, and we shall term it as his \textit{Second} expansion. Hence, \begin{equation} J_{\nu}(\nu \sec{\beta})=\sqrt{\frac{2 \cot{\beta}}{\nu \pi}} e^{-P_{\nu}}\cos\left(Q_{\nu}-\frac{1}{4}\pi \right) \end{equation} where $P_{\nu}$ is given as, \begin{equation} P_{\nu}=P_1+P_2+P_3+P_4+... \end{equation} where \begin{eqnarray} P_{1}=\frac{\cot^{6}\beta}{16\nu^2}\left(4\sec^2{\beta}+\sec^4{\beta}\right)\nonumber \end{eqnarray} \begin{eqnarray} P_{2}=-\frac{\cot^{12}\beta}{128\nu^4}(32\sec^2{\beta}+288\sec^4{\beta}+232\sec^6{\beta}\nonumber\\ +13\sec^8{\beta})\nonumber \end{eqnarray} \begin{eqnarray} P_{3}=\frac{\cot^{18}\beta}{192\nu^6}(48\sec^2{\beta}+2580\sec^4{\beta}+14884\sec^6{\beta}\nonumber\\ +17493\sec^8{\beta}+4242\sec^{10}{\beta}+103\sec^{12}{\beta})\nonumber \end{eqnarray} \begin{eqnarray} P_{4}=\frac{\cot^{24}\beta\sec^2{\beta}}{4096\nu^8}(1024+248320\sec^2{\beta}+5095936\sec^4{\beta}\nonumber \end{eqnarray} \begin{eqnarray} +24059968\sec^6{\beta}+34280896\sec^8{\beta}+15252048\sec^{10}{\beta}\nonumber\\ +1765936\sec^{12}{\beta}+23797\sec^{14}{\beta})\nonumber \end{eqnarray} and $Q_{\nu}$ is given as, \begin{equation} Q_{\nu}=Q_1+Q_2+Q_3+Q_4+... \end{equation} and, \begin{eqnarray} Q_1=\nu(\tan{\beta}-\beta)-\frac{\cot^{3}\beta}{24\nu}\left(2+3\sec^2{\beta}\right)\nonumber \end{eqnarray} \begin{eqnarray} Q_2=-\frac{\cot^{9}\beta}{5760\nu^3}(16-1512\sec^2{\beta}-3654\sec^4{\beta}\nonumber\\ -375\sec^6{\beta})\nonumber \end{eqnarray} \begin{eqnarray} Q_3=-\frac{\cot^{15}\beta}{322560\nu^5}(256+78720\sec^2{\beta}+1891200\sec^4{\beta}\nonumber\\ +4744640\sec^6{\beta}+1914210\sec^{8}{\beta}+67599\sec^{10}{\beta})\nonumber \end{eqnarray} \begin{eqnarray} Q_4=-\frac{\cot^{21}\beta}{3440640\nu^7}(881664\sec^2{\beta}+99783936\sec^4{\beta}\nonumber \end{eqnarray} \begin{eqnarray} +1135145088\sec^6{\beta}+2884531440\sec^8{\beta}+1965889800\sec^{10}{\beta}\nonumber\\ +318291750\sec^{12}{\beta}+5635995\sec^{14}{\beta}-2048)\nonumber \end{eqnarray} It should be remarked that we disagree with Meissel's result for $P_3$ in the last four terms. However, we obtain perfect agreement with the rest of his results \cite{M1}. We have improved on his result by using $V_7$, $V_8$ to obtain $P_4$ and $Q_4$. Hence, we have increased the accuracy of this expansion by at least one order from Meissel's earlier result. Again, higher order results are easily obtainable and are available if needed. In Figures 1 and 2 we have plotted these expansions in the regions they are expected to fail. These are the so called ``transition" regions, where each expansion approaches a singularity (as the order equals the argument). For the computationally motivated (we can compute exact values of Bessel functions with ease) case of the $\nu=300$, we note the following. Fig.1 indicates the onset of breakdown in the \textit{First} expansion for argument values around and larger than 290. Similarly, Figure 2, indicates a similar breakdown starting around the values 300 and persisting till 310. Hence, the values outside these regions of breakdown or transition regions are well covered by Meissel's expansions. However, the issue as to deal with these regions need to be addressed via separate methods, which will be addressed in more detail in Section IV. The CPU time for these approximations was less than 0.01 seconds per value on a 2.4 GHz Pentium IV processor running MAPLE version 9. The ``exact" MAPLE solver took somewhere between 0.03 to 0.08 seconds to compute each value. Clearly, there is a lot more computational speed in using a few terms present in these expansions. As an application, it should be noted that values of this order are applicable to the Peters-Mathews model of gravitational radiation from binary inspiralling stars \cite{PM}. \begin{figure} \centering \epsfig{file=Fig1_Meissel_First.eps,width= 2.5 in, height= 2.5 in, angle=270} \caption{Meissel's \textit{First} expansion and actual Bessel function graphed for argument $x$ and order $\nu=300$ near the transition region. Solid line indicates actual Bessel function values and circles indicate values given by the expansion.} \label{fig_three} \end{figure} \begin{figure} \centering \epsfig{file=Fig2_Meissel_Second.eps,width= 2.5 in, height= 2.5 in, angle=270} \caption{Meissel's \textit{Second} expansion and actual Bessel function graphed for argument $x$ and order $\nu=300$ near the transition region. Solid line indicates actual Bessel function values and circles indicate values given by the expansion.} \label{fig_two} \end{figure} For the case when the argument equals the index, we extend Meissel's \textit{Third} expansion \cite{M1} by two orders as follows: \begin{equation} J_{n}(n)\sim\frac{1}{\pi}\sum_{m=0}^{\infty} \lambda_{m} \Gamma\left(\frac{2m}{3} + \frac{4}{3} \right)\left(\frac{6}{n}\right)^{\frac{2}{3}m + \frac{1}{3}}\cos\pi (\frac{m}{3} + \frac{1}{6}) \end{equation} where the terms, $\lambda_m$ ($m=0,1,2,..7$), are given by, \begin{eqnarray} \lambda_0=1,\lambda_1=\frac{1}{60},\lambda_2=\frac{1}{1400},\lambda_3=\frac{1}{25200},\nonumber \end{eqnarray} \begin{eqnarray} \lambda_4=\frac{43}{17248000},\lambda_5=\frac{1213}{7207200000},\lambda_6=\frac{681563}{5721073600000},\nonumber \end{eqnarray} \begin{eqnarray} \lambda_7=\frac{63319}{726485760000000} \end{eqnarray} We observe that inclusion of the higher order terms leads to 10 decimal accuracy compared to actual values of large order Bessel functions. The method of steepest descents was employed by Debye in \cite{DBY}. For the case when the argument is less than the order, he obtained, \begin{equation} J_{\nu}(\nu sech(\alpha))\sim \frac{e^{\nu(\tanh{\alpha}-\alpha)}}{\sqrt{2 \pi \nu\tanh{\alpha}}} \sum_{m=0}^{\infty}\frac{\Gamma(m+\frac{1}{2})}{\Gamma(\frac{1}{2})}\frac{A_m}{(\frac{1}{2}\nu \tanh{\alpha})^{m}} \end{equation} where, \begin{eqnarray} A_0=1, A_1=\frac{1}{8}-\frac{5}{24}\coth^2{\alpha} \nonumber \end{eqnarray} \begin{eqnarray} A_2=\frac{3}{128}-\frac{77}{576}\coth^2{\alpha}+\frac{385}{3456}\coth^4{\alpha}\nonumber \end{eqnarray} \begin{eqnarray} A_3=\frac{5}{1024}-\frac{1521}{25600}\coth^2{\alpha}+\frac{17017}{138240}\coth^4{\alpha}\nonumber\\ -\frac{17017}{248832}\coth^6{\alpha}\nonumber \end{eqnarray} \begin{eqnarray} A_4=\frac{11513}{92897280}-\frac{21023}{9953280}\coth^2{\alpha}+\frac{138919}{19906560}\coth^4{\alpha}\nonumber\\ -\frac{49049}{5971968}\coth^6{\alpha}+\frac{230945}{71663616}\coth^8{\alpha}\nonumber \end{eqnarray} Following this method, we have computed two higher orders $A_3$ and $A_4$, using symbolic computation. For the case when the argument is larger than the order, Debye obtains the following expansion: \begin{eqnarray} J_{\nu}(\nu\sec{\beta})\sim\sqrt{\frac{2}{\pi\nu\tan{\beta}}}[\cos\left(\nu\tan{\beta}-\nu\beta-\frac{1}{4}\beta\right)\nonumber \end{eqnarray} \begin{eqnarray} \times\sum_{m=0}^{\infty}(-1)^{m}\frac{\Gamma(m+\frac{1}{2})}{\Gamma(\frac{1}{2})}\frac{A_{2m}}{(\frac{1}{2}\nu\tanh{\alpha})^{2m}}+\sin(\nu\tan{\beta}\nonumber \end{eqnarray} \begin{eqnarray} -\nu\beta-\frac{1}{4}\beta)\sum_{m=0}^{\infty}(-1)^{m}\frac{\Gamma(2m+\frac{3}{2})}{\Gamma(\frac{1}{2})}\frac{A_{2m+1}}{(\frac{1}{2}\nu \tanh{\alpha})^{2m+1}}] \end{eqnarray} where, \begin{eqnarray} A_0=1, A_1=\frac{1}{8}+\frac{5}{24}\cot^2{\beta}\nonumber \end{eqnarray} \begin{eqnarray} A_2=\frac{3}{128}+\frac{77}{576}\cot^2{\beta}+\frac{385}{3456}\cot^4{\beta}\nonumber \end{eqnarray} \begin{eqnarray} A_3=\frac{5}{1024}+\frac{1521}{25600}\cot^2{\beta}+\frac{17017}{138240}\cot^4{\beta}\nonumber\\ +\frac{17017}{248832}\cot^6{\beta}\nonumber \end{eqnarray} \begin{eqnarray} A_4=\frac{11513}{92897280}+\frac{21023}{9953280}\cot^2{\beta}+\frac{138919}{19906560}\cot^4{\beta}\nonumber\\ +\frac{49049}{5971968}\cot^6{\beta}+\frac{230945}{71663616}\cot^8{\beta}\nonumber \end{eqnarray} Again, we have extended Debye's result by two higher orders by obtaining $A_3$ and $A_4$. However, due to the nature of this method we could not obtain reliable results that spanned in a generally predictable direction. Accuracy was limited to the region of the stationary phase as expected and hence, we recommend Meissel's expansions to be more reliable (except of course in the ``transition" region) than the method of steepest descent. \section{Transitional regions: Contour Integration and extension of $\epsilon$ expansion} To address the issues related to computation for large order Bessel functions in the transition regions we present two methods that are geared to work in such domains. Firstly, we present the results by Watson, \cite{Watson}. For the case of the argument being less than the order, he obtained via use of contour integration, \begin{eqnarray} J_{\nu}(\nu sech(\alpha))= \frac{\tanh{\alpha}}{\pi \sqrt{3}}\exp\left[\nu\left(\tanh{\alpha}+\frac{1}{3}\tanh^3{\alpha}-\alpha\right)\right]\nonumber\\ \times K_{\frac{1}{3}}\left(\frac{1}{3}\nu\tanh^3{\alpha}\right)\nonumber \end{eqnarray} \begin{eqnarray} +3\theta_{1}\nu^{-1}\exp[\nu(\tanh{\alpha}-\alpha)] \end{eqnarray} where $\theta_1<1$. Similarly, for the case when the argument is greater than the order, he derived the following: \begin{eqnarray} J_{\nu}(\nu \sec{\beta})=\frac{1}{3}\tan{\beta}\cos\left[\nu\left(\tan{\beta}-\frac{1}{3}\tan^3{\beta}-\beta\right)\right]\times\nonumber \end{eqnarray} \begin{eqnarray} \left(J_{-\frac{1}{3}}+J_{\frac{1}{3}}\right)+3^{-\frac{1}{2}}\tan{\beta}\sin\left[\nu\left(\tan{\beta}-\frac{1}{3}\tan^3{\beta}-\beta\right)\right]\times\nonumber\\ \left(J_{-\frac{1}{3}}-J_{\frac{1}{3}}\right)\nonumber \end{eqnarray} \begin{eqnarray} +24\theta_{2}\nu^{-1} \end{eqnarray} where $\theta_2<1$ and the argument for the Bessel functions $J_{\pm\frac{1}{3}}$ is $\frac{1}{3}\tan^{3}{\beta}$. The great advantage of these formulae is that they have error bounds given. However, these extensions are not trivial as this involves solving extensions to Airy-type integrals, for which we do not presently have closed form answers. The other issue with these formulae is that they are themselves given in fractional Bessel function form which would pose computational problems once the arguments involved are large. On the other hand, Debye \cite{DBY}, introduced, what we will term as ``$\epsilon$ expansion". The idea is motivated by introducing a small parameter $\epsilon$, such that $\nu=z(1-\epsilon)$, where $\nu$ denotes the order and $z$ is the argument of the Bessel function. \begin{equation} J_{\nu}(z)\sim\frac{1}{3\pi}\sum_{m=0}^{\infty} B_{m}(\epsilon z) \sin\frac{1}{3}(m+1)\pi\cdot \frac{\Gamma(\frac{1}{3}m+\frac{1}{3})}{(\frac{1}{6}z)^{\frac{1}{3}(m+1)}} \end{equation} We have extended this analysis by 5 orders and the terms $B_m(\epsilon z), m=0,1,2,..15$, are given as, \begin{eqnarray} B_0(\epsilon z)=1, B_1(\epsilon z)=\epsilon z, B_3(\epsilon z)=\frac{1}{6}\epsilon^3 z^3 -\frac{1}{15} \epsilon z \nonumber \end{eqnarray} \begin{eqnarray} B_4(\epsilon z)=\frac{1}{24}\epsilon^4 z^4 -\frac{1}{24} \epsilon^2 z^2 + \frac{1}{280}\nonumber \end{eqnarray} \begin{eqnarray} B_6(\epsilon z)={\frac {1}{720}}\,{z}^{6}{\epsilon}^{6}-{\frac {7}{1440}}\,{z}^{4}{ \epsilon}^{4}+{\frac {1}{288}}\,{z}^{2}{\epsilon}^{2}-{\frac {1}{3600} } \end{eqnarray} \begin{eqnarray} B_7(\epsilon z)={\frac {1}{5040}}\,{z}^{7}{\epsilon}^{7}-{\frac {1}{900}}\,{z}^{5}{ \epsilon}^{5}+{\frac {19}{12600}}\,{z}^{3}{\epsilon}^{3}-{\frac {13}{ 31500}}\,z\epsilon \nonumber \end{eqnarray} \begin{eqnarray} B_{9}(\epsilon z)={\frac {1}{362880}}\,{z}^{9}{\epsilon}^{9}-{\frac {1}{30240}}\,{z}^{7} {\epsilon}^{7}+{\frac {71}{604800}}\,{z}^{5}{\epsilon}^{5}\nonumber\\ -{\frac{121}{907200}}\,{z}^{3}{\epsilon}^{3}+{\frac{7939}{232848000}}\,z\epsilon\nonumber \end{eqnarray} \begin{eqnarray} B_{10}(\epsilon z)={\frac {1}{3628800}}\,{z}^{10}{\epsilon}^{10}-{\frac {11}{2419200}}\,{ z}^{8}{\epsilon}^{8}+{\frac {143}{6048000}}\,{z}^{6}{\epsilon}^{6}\nonumber\\ -{ \frac{803}{18144000}}\,{z}^{4}{\epsilon}^{4}+{\frac {43}{1728000}}\,{ z}^{2}{\epsilon}^{2}-{\frac {1213}{655200000}}\nonumber \end{eqnarray} \begin{eqnarray} B_{12}(\epsilon z)={\frac {1}{479001600}}\,{z}^{12}{\epsilon}^{12}-{\frac{13}{217728000}}\,{z}^{10}{\epsilon}^{10}+\nonumber\\ {\frac {299}{508032000}}\,{z}^{8}{\epsilon}^{8}-{\frac{377}{155520000}}\,{z}^{6}{\epsilon}^{6}+{\frac{337207}{83825280000}}\,{z}^{4}{\epsilon}^{4}\nonumber\\ -{\frac{59503}{27941760000}}\,{z}^{2}{\epsilon}^{2}+{\frac{151439}{977961600000}}\nonumber \end{eqnarray} \begin{eqnarray} B_{13}(\epsilon z)={\frac {1}{6227020800}}\,{z}^{13}{\epsilon}^{13}-{\frac{1}{171072000}}\,{z}^{11}{\epsilon}^{11}+\nonumber\\ {\frac{11}{145152000}}\,{z}^{9}{\epsilon}^{9}-{\frac{47}{108864000}}\,{z}^{7}{\epsilon}^{7}+{\frac{25853}{23950080000}}\,{z}^{5}{\epsilon}^{5}\nonumber\\ -{\frac{266303}{259459200000}}\,{z}^{3}{\epsilon}^{3}+\frac{169039}{698544000000}\,z\epsilon\nonumber \end{eqnarray} \begin{eqnarray} B_{15}(\epsilon z)={\frac {1}{1307674368000}}\,{z}^{15}{\epsilon}^{15}-{\frac{1}{23351328000}}\,{z}^{13}{\epsilon}^{13}\nonumber\\ +{\frac{113}{125737920000}}\,{z}^{11}{\epsilon}^{11}-{\frac{17}{1905120000}}\,{z}^{9}{\epsilon}^{9}\nonumber\\ +{\frac{76841}{1760330880000}}\,{z}^{7}{\epsilon}^{7}-{\frac{37021}{371498400000}}\,{z}^{5}{\epsilon}^{5}\nonumber\\ +{\frac{5141933}{57210753600000}}\,{z}^{3}{\epsilon}^{3}-{\frac{16720141}{810485676000000}}\,z\epsilon\nonumber \end{eqnarray} Terms $B_{3m-1}$, $m=1,2...$ do not contribute in eqn. (21) due to the periodicity of the sine function. With symbolic computation, we are able to generate higher orders if needed. To illustrate the applicability and issues of both these methods to the transition region, we present Figures 3 and 4, which are plotted for the problematic regions (when the order is $\nu=300$) in Figures 1 and 2. Both methods show remarkable ability in capturing the functions in the domains of interest. In Figure 3, the $\epsilon$ expansion starts working at values at 286 and Watson's formula works to even a larger domain. Similarly, in Figure 4, both the methods indicate success in regions where Meissel's expansions fail. This starts at values of the argument, and works up to $x=316$ for the $\epsilon$ expansion whereas, again, the domain of Watson's formula is much greater. The reasons for lesser range of the $\epsilon$ expansion can be attributed to the fact that it is a power series compared to Watson's formula which actually depends on fractional Bessel functions themselves. Further, the $\epsilon$ expansion depends crucially on the size of the parameter, which is connected with the order one is working with. However, the reason why we will persist with this method is that it will be more applicable when the argument of the Bessel function is quite large. \begin{figure} \centering \epsfig{file=Fig3_eps_Watson.eps,width= 2.5 in, height= 2.5 in, angle=270} \caption{Comparison of $\epsilon$ expansion and Watson's formulae for argument $x<300$ and order $\nu=300$.Solid line indicates exact Bessel function values, diamonds represent $\epsilon$ expansion and circles indicate values given by Watson's formula.} \label{fig_three} \end{figure} \begin{figure} \centering \epsfig{file=Fig4_eps_Watson.eps,width= 2.5 in, height= 2.5 in, angle=270} \caption{Comparison of $\epsilon$ expansion and Watson's formula in the transition region for argument $x>300$ and order $\nu=300$. Solid line indicates exact Bessel function values, diamonds represent $\epsilon$ expansion and circles indicate values given by Watson's formula.} \label{fig_four} \end{figure} To illustrate the type of values a GW pulsar FT would require, we present Figures 5 and 6. Here, we choose a very large order (yet realistic phenomenologically) for the Bessel function, which is 1 million. Also, in such a scenario, we would be looking at values greater than one million, hence Meissel's second expansion along with the appropriate Watson's formula (eq. 20) will be put to use. We were not able to make exact comparison, obviously due to massive computer times required. In this regard, the problem of ``exact" Bessel functions presents a genuine challenge to SHARCNET (Shared Hierarchical Academic Research Cluster Network) and HPC in general. In Figure 5, we observe strong evidence that the proposed asymptotic expansions are appropriate for GW signal analysis. Here, we note the transition region starting at values of the argument at 1,000,000 and going up to 1,000,200. In this region, both the $\epsilon$ expansion and Watson's formula almost coincide with each other. As usual, the $\epsilon$ expansion breaks down earlier, however, all three methods coincide in a certain region indicating that we have consistent methods that work for values relevant to GW analysis. Meissel's expansion is fairly easy to implement computationally and indicates good stability for rather large values of the argument. This is illustrated in Figure 6, where we plot this expansion for values ranging from 1,000,200 to 32,500,000, which are relevant for GW phenomenology. This appears as a black band and is a continuous function which indicates oscillations tightly bunched together. It is noteworthy that the method is stable and shows consistent behaviour over an extreme range of values for the argument. The CPU time consumed by each of the points, on the average took less than 0.01 seconds on MAPLE. The Bessel utility in MAPLE crashed repeatedly after 15-30 minutes on the same system described above. It should be remarked that Watson's formula lacks in this capacity as it depends on fractional Bessel functions itself, which will provide computational challenge for such values. A detailed analysis regarding computational advantage over exact computation will be addressed in a later work. It is aimed to not only address the question of GW analysis but will deal with general computational issues regarding large order Bessel functions. \begin{figure} \centering \epsfig{file=Fig5_million_eps_fail.eps,width= 2.5 in, height= 2.5 in, angle=270} \caption{Comparison of Meissel's \textit{Second} expansion, $\epsilon$ expansion and Watson's formulae for argument $x>1,000,000$ and order $\nu=1,000,000$. Solid line indicates Meissel's \textit{Second} expansion values, diamonds represent $\epsilon$ expansion and circles indicate values given by Watson's formula.} \label{fig_five} \end{figure} \begin{figure} \centering \epsfig{file=Fig6_million_Messel_success.eps,width= 2.5 in, height= 2.5 in, angle=270} \caption{Plot of Meissel's \textit{Second} expansion for argument, $x$ ranging from 1,000,200 to 32,500,000 for order 1,000,000.} \label{fig_six} \end{figure} \section{Conclusion} In this present work, we have given an extended asymptotic analysis for the large order and argument Bessel functions. This analytically improves the earlier pioneering works of Meissel, Airey, Debye and Watson. These extensions should be of possible use not only in GW signal analysis, but also in a variety of problems in Engineering and the Sciences where the ubiquitous Bessel functions are encountered. \section*{Acknowledgments} We are deeply grateful to SHARCNET for valuable grant support that made this study feasible. We are also indebted to Drs. Nico Temme (CWI, Amsterdam), Walter Gautschi (Purdue U.), D.G.C. McKeon (U. Western Ontario), Tom Prince (JPL, Pasadena) and the referee for valuable suggestions.
{ "timestamp": "2005-03-15T07:47:17", "yymm": "0503", "arxiv_id": "math-ph/0503037", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503037" }
\section*{} The concept of financial log-periodicity is based on the appealing assumption that the financial dynamics is governed by phenomena analogous to criticality in the statistical physics sense (Sornette et al. 1996, Feigenbaum and Freund 1996). Criticality implies a scale invariance which, for a properly defined function $F(x)$ characterizing the system, means that \begin{equation} F(\lambda x) = \gamma F(x). \label{eq:F} \end{equation} A constant $\gamma$ describes how the properties of the system change when it is rescaled by the factor $\lambda$. The general solution to this equation reads: \begin{equation} F(x) = x^{\alpha} P(\ln(x)/\ln(\lambda)), \label{eq:FP} \end{equation} where the first term represents a standard power-law that is characteristic of continuous scale-invariance with the critical exponent $\alpha = \ln(\gamma) / \ln(\lambda)$ and $P$ denotes a periodic function of period one. This general solution can be interpreted in terms of discrete scale invariance. It is due to the second term that the conventional dominating scaling acquires a correction that is periodic in $\ln(x)$ and may account for the zig-zag character of financial dynamics. It demands however that if $x = \vert T - T_c \vert$, where $T$ denotes the ordinary time labeling the original price time series, represents a distance to the critical point $T_c$, the resulting spacing between the corresponding consecutive repeatable structures at $x_n$ seen in the linear scale follow a geometric contraction according to the relation $(x_{n+1}-x_n) / (x_{n+2}-x_{n+1}) = \lambda$. The critical points correspond to the accumulation of such oscillations and, in the context of the financial dynamics, it is this effect that potentially can be used for prediction. An extremely important related element, for a proper interpretation and handling of the financial patterns as well as for consistency of the theory, is that such log-periodic oscillations manifest their action self-similarly through various time scales (Dro\.zd\.z et al. 1999). This applies both to the log-periodically accelerating bubble market phase as well as to the log-periodically decelerating anti-bubble phase. Furthermore, more and more evidence is collected that the preferred scaling factor $\lambda \approx 2$ and is common to all the scales and markets (Dro\.zd\.z et al. 2003). These two elements, self-similarity and universality of the $\lambda$, set very valuable and in fact crucial constraints on possible forms of the analytic representations of the market trends and oscillation patterns, including the future ones as well. A specific form of the periodic function $P$ in Eq.~\ref{eq:FP} is as yet not provided by any first principles which opens room for certain, seemingly mathematically unrigorous assignments of patterns. This, on the other hand, allows to correct for frequent market 'imprecisions' when relating its real behavior versus the theory. Very helpful in this respect is the requirement of self-similarity which greatly clarifies the significance of a given pattern and allows to determine on what time scale it operates. Since in the corresponding methodology the oscillation structure carries the most relevant information about the market dynamics, for transparency of this presentation, we use the first term of its Fourier expansion, \begin{equation} P(\ln(x)/\ln(\lambda)) = A + B \cos({\omega \over 2\pi} \ln(x) + \phi). \label{eq:FPE} \end{equation} This implies that $\omega = 2\pi / \ln(\lambda)$. Already such a simple parametrization allows to properly reflect the contraction of oscillations, especially on the larger time scales. On the smaller time scales just replacing the {\it cosine} by its modulus often, even quantitatively in addition, describes departures of the market amplitude from its average trend. \begin{figure}[ht] \begin{center} \includegraphics[width=6cm, angle=270]{Figure1a.eps} \includegraphics[width=6cm, angle=270]{Figure1b.eps} \caption{(a) Logarithm of the Standard $\&$ Poor's 500 index since 1800 (http://www.globalfindata.com). The thick solid line displays its optimal log-periodic representation with $\lambda = 2$. The thin solid line represents the inflation corrected S$\&$P500 expressed in 2004 US$\$$. It significantly shifts the third minimum to the early 1980s and improves agreement with the theoretical representation. (b) Logarithm of the S$\&$P500 from 1997 till the end of 2002, which corresponds to the magnification of the small rectangle in (a). The solid lines illustrate the log-periodic accelerating and decelerating representations with $\lambda = 2$, modulus of the cosine used in Eq.~(3), and a common $T_c = 1.9.2000$.} \end{center} \end{figure} One particularly relevant and special, for several reasons, example is shown in Fig.~1. The upper panel (a) illustrates a nearly optimal log-periodic representation of the S$\&$P500 data over the most extended time-period of the recorded stock market activity as dated since 1800. As already pointed out (Dro\.zd\.z et al. 2003) this development signals in around 2025 a transition of the S$\&$P500 to a globally declining phase as measured in the contemporary terms. The magnification of the small rectangle covering the period 1997-2002 is displayed in the lower panel (b) of the same Fig.~1. It thus illustrates the nature of the stock market evolution on a much smaller time scale of resolution. An impressive log-periodicity with the same $\lambda=2$ on both sides of the transition date (September 1, 2000) can be seen. The next stock market top from the perspective of the largest time scale (Fig.~1a) can be estimated to occur in around the years 2010-2011. In the spirit of the log-periodicity its neighborhood is to be accompanied by the smaller time scale oscillations - similar in character to those in Fig.~1b. Of course, when going far away from those large scale transition points such pure log-periodic structures - representative to the one level lower time scale - must get dissolved. A particularly interesting related question then is what characteristics are to govern the stock market dynamics in the transition period when going from 2000 to 2010. The most natural and straightforward way is to view this process as schematically is indicated in Fig.~2. This whole period is thus covered by the two main components represented by the thin lines and the market dynamics is driven by the superposition of of these two components whose phases, slopes and weights are adjusted such that the overall global market trend up to now is reproduced. \begin{figure}[ht] \begin{center} \includegraphics[width=6cm, angle=270]{Figure2.eps} \caption{A hypothetical log-periodic scenario, represented by the thick solid line, for the S$\&$P500 development until 2010. This solid line is obtained by summing up the two $\lambda = 2$ components (thin lines): log-periodically decelerating since 1.9.2000 and log-periodically accelerating toward 1.9.2010.} \end{center} \end{figure} In this scenario, close to the two large-scale transition points (September 2000 and, as provisionally estimated here based on Fig.~1a, September 2010) the market is driven, as needed, essentially by the single log-periodic components, decelerating and accelerating one, correspondingly. More complicated is the situation in the middle of this time interval where the two components contribute comparably. Most interestingly, it indicates that the period of the stock market stagnation may extend even into the year 2008, before it seriously starts rising. It also demonstrates a possible mechanism that generates modulation structures responsible for the apparent higher order corrections (Johansen and Sornette 1999) to Eq.~(\ref{eq:FPE}). The changes in the frequency relations observed in the transition period between the bear and the bull market phases originate here from the interference between the two components, both of the simple form as prescribed by Eq.~(\ref{eq:FPE}) and with the same $\lambda = 2$. Of course, similar effects of interference may occur on the whole hierarchy of different time scales. There is one more element that from time to time takes place in the financial dynamics and whose identification appears relevant for a proper interpretation of the financial patterns with the same universal value of the preferred scaling factor $\lambda$. This is the phenomenon of a "super-bubble" (Dro\.zd\.z et al. 2003) which is a local bubble, itself evolving log-periodically, superimposed on top of a long-term bubble. Two such spectacular examples are provided by the Nasdaq in the first quarter of 2000 and by the gold price in the beginning of 1981 (Dro\.zd\.z et al. 2003). \begin{figure}[ht] \begin{center} \includegraphics[width=6cm, angle=270]{Figure3.eps} \caption{The New York traded oil futures since 1998 and the corresponding log-periodic $\lambda = 2$ representation in terms of Eq.~(3).} \end{center} \end{figure} In connection with this second case it is important to remember that the same value of $\lambda$ as for the stock market turns out appropriate. That such its value may be characteristic to the whole commodities market as well, is shown in Fig.~3 which displays the New York traded oil futures versus the best log-periodic $(\lambda = 2)$ representation. In fact, this scenario has been drawn by the authors on September 15, 2004, insisting on using $\lambda = 2$, even though one local minimum (in the beginning of 2004) in the corresponding sequence did not look very convincing. Designed this way it was indicating a continuation of the increase until the end of October and then a more serious reverse of the trend. Subsequent development of the oil futures provides further arguments in favor of this way of handling the financial log-periodicity. \section*{References} \begin{description} \item[]Dro\.zd\.z S, Ruf F, Speth J, W\'ojcik M (1999)~ Imprints of log-periodic self-similarity in the stock market. Eur. Phys. J. B 10:589-593 \item[]Dro\.zd\.z S, Gr\"ummer F, Ruf F, Speth J (2003)~ Log-periodic self-similarity: an emerging financial law? Physica A 324:174-182 \item[]Feigenbaum JA, Freund PGO (1996)~ Discrete scale invariance in stock markets before crashes. Int. J. Mod. Phys. B 10:3737-3745 \item[]Johansen A, Sornette D (1999)~ Financial "anti-bubbles": Log-periodicity in gold and Nikkei collapses. Int. J. Mod. Phys. C 10:563-575 \item[]Sornette D, Johansen A, Bouchaud J.-P (1996)~ Stock market crashes, precursors and replicas. J. Physique (France) 6:167-175 \end{description} \end{document}
{ "timestamp": "2005-03-01T21:03:07", "yymm": "0503", "arxiv_id": "physics/0503006", "language": "en", "url": "https://arxiv.org/abs/physics/0503006" }
\section{\label{sec:intro}Introduction} Fluid flow with interfaces and free surfaces is common in nature and in many engineering applications. Such interfacial flows which typically involve multiple scales remain a formidable non-linear problem rich in physics and continue to pose challenges to experimentalists and theoreticians alike~\cite{eggers97}. Numerical simulation of multiphase flows is challenging as the shape and location of the interfaces must be computed in conjunction with the solution of the flow field~\cite{hyman84,scardovelli99}. Computational methods based on the lattice Boltzmann equation (LBE) for simulating complex emergent physical phenomena have attracted much attention in recent years~\cite{chen98,succi02}. The LBE simulates multiphase flows by incorporating interfacial physics at scales smaller than macroscopic scales. Phase segregation and interfacial fluid dynamics can be simulated by incorporating inter-particle potentials~\cite{shan93,shan94}, concepts based on free energy~\cite{swift95,swift96} or kinetic theory of dense fluids~\cite{he98,he99,he02}. The formulation of the standard LBE is based on the Cartesian coordinate system and does not take into account axial symmetry that may exist. Numerous multiphase flow situations exist where the fluid dynamics can be approximated as axisymmetric~\cite{sussman96,eggers97}. Examples include head-on collision of drops, normal drop impingement on solid surfaces and Rayleigh instability of cylindrical liquid columns. Currently, full three-dimensional (3D) calculations have to be carried out for problems which may be approximated as axisymmetric~\cite{he99a,inamuro03,premnath04}. In 3D computations, computational considerations restrict the numerical resolution that may be employed and the physics may not be well resolved. For example, in breakup of drops into satellite droplets the size of the droplets may be such that the 3D grids may not resolve them. To improve the computational efficiency of the LBE for axisymmetric multiphase flows, we propose an axisymmetric LB model in this paper. The approach consists of adding source terms to the two-dimensional (2D) Cartesian LBE model based on the kinetic theory of dense fluids for multiphase flows~\cite{he98,he99}. This approach is similar in spirit to the idea proposed in~\cite{halliday01} to solve single-phase axisymmetric flows. However, multiphase flow problems involve additional complexity as a result of interfacial physics involved, i.e. the surface tension forces and the need to track the interfaces. In this case, the accuracy of the numerical discretization of the source terms representing interfacial physics also becomes an important consideration. This paper is organized as follows. In Section \ref{sec:axismodel}, the axisymmetric LBE multiphase model is described. Then, in Section \ref{sec:axismodelc}, its extension to simulate axisymmetric multiphase flows with reduced compressibility effects is described. The computational methodology adopted is also discussed in this section. In Section \ref{sec:results}, the axisymmetric model is applied to benchmark problems to evaluate its accuracy. Finally, the paper closes with summary in Section \ref{sec:summary}. \section{\label{sec:axismodel}Axisymmetric LBE Multiphase Flow Model} To simulate axisymmetric multiphase flows, axisymmetric contributions of the order parameter, and inertial, viscous and surface tension forces may be introduced to the standard 2D LBE. The source terms, which will be shown to be spatially and temporally dependent, are determined by performing a Chapman-Enskog multiscale analysis in such a way that the macroscopic mass and momentum equations for multiphase flows are recovered self-consistently. The introduction of source terms makes it necessary to calculate additional spatial gradients when compared to those in the standard LBE. While this approach is developed for a specific LBE multiphase flow model based on kinetic theory of dense fluids~\cite{he98,he99}, it can be readily extended to other LBE multiphase flow models. The governing continuum equations of isothermal multiphase flow~\cite{nadiga96,zou99} in the cylindrical coordinate system when the axisymmetric assumption is employed are \begin{equation} \partial_t \rho + \frac{1}{r} \partial_r \left( \rho r u_r \right) + \partial_z \left( \rho u_z \right) = 0, \label{eq:axiscont} \end{equation} \begin{equation} \rho\left( \partial_t u_r + u_r \partial_r u_r + u_z \partial_z u_r \right)= -\partial_r P + F_{s,r}+F_{ext,r}+\frac{1}{r}\partial_r \left( r \Pi_{rr} \right)+ \partial_z \left( \Pi_{rz} \right), \label{eq:axismomr} \end{equation} \begin{equation} \rho\left( \partial_t u_z + u_r \partial_r u_z + u_z \partial_z u_z \right)= -\partial_z P + F_{s,z}+F_{ext,z}+\frac{1}{r}\partial_r \left( r \Pi_{zr} \right)+ \partial_z \left( \Pi_{zz} \right), \label{eq:axismomz} \end{equation} where $\rho$ is the density and $u_r$ and $u_z$ are the radial and axial components of velocity. These equations are derived from kinetic theory that incorporates intermolecular interactions forces which are modeled as a function of density following the work of van der Waals~\cite{rowlinson}. The exclusion volume effect of Enskog~\cite{chapman} is also incorporated to account for increase in collision probability due to the increase in the density of non-ideal fluids. These features naturally give rise to surface tension and phase segregation effects. The other variables which appear in the above equations will now be described. $\Pi_{rr}$, $\Pi_{rz}$, $\Pi_{zz}$ are the components of the viscous stress tensor and are given by \begin{eqnarray} \Pi_{rr}&=&2\mu \partial_r u_r, \\ \Pi_{rz}&=& \Pi_{zr}=\mu \left( \partial_z u_r + \partial_r u_z \right),\\ \Pi_{zz}&=&2\mu \partial_z u_z, \end{eqnarray} where $\mu$ is the dynamic viscosity. $F_{s,r}$ and $F_{s,z}$ are the axial and radial components respectively of the surface tension force, which are given by~\cite{zou99} \begin{eqnarray} F_{s,r}&=& \kappa \rho \partial_r \left[\frac{1}{r} \partial_r (r\partial_r\rho)+\partial_z(\partial_z\rho) \right], \label{eq:surfr}\\ F_{s,z}&=& \kappa \rho \partial_z \left[\frac{1}{r} \partial_r (r\partial_r\rho)+\partial_z(\partial_z\rho) \right], \label{eq:surfz} \end{eqnarray} where $\kappa$ controls the strength of the surface tension force. This parameter is related to the surface tension of the fluid, $\sigma$, through the density gradient across the interface by the equation~\cite{evans79} \begin{equation} \sigma = \kappa \int \left( \frac{\partial \rho}{\partial n} \right)^2 dn. \label{eq:sigmakappa} \end{equation} Thus, the surface tension is a function of both the parameter $\kappa$ and the density profile across the interface. The terms $F_{ext,r}$ and $F_{ext,z}$ in Eqs. (\ref{eq:axismomr}) and (\ref{eq:axismomz}) respectively are the radial and axial components of external forces such as gravity. The pressure, $P$, is related to density through the Carnahan-Starling-van der Waals equation of state (EOS)~\cite{carnahan69} \begin{equation} P=\rho R T \left\{ \frac{1+\gamma+\gamma^2-\gamma^3}{(1-\gamma)^3} \right\} - a\rho^2, \label{eq:axiseos} \end{equation} where $\gamma=b\rho/4$. The parameter $a$ is related to the intermolecular pair-wise potential and $b$ to the effective diameter of the molecule, $d$, and the mass of a single molecule, $m$, by $b=2\pi d^3/3m$. $R$ is a gas constant and $T$ is the temperature. The Carnahan-Starling EOS has a \emph{supernodal} $P-1/\rho-T$ curve, i.e., $dP/d\rho<0$, for certain range of values of $\rho$, when the state fluid temperature is below its critical value. This unstable part of the curve is the driving mechanism responsible for keeping the phases of fluids segregated and for maintaining a self-generated sharp interface. We now modify the standard LBE in such a way that it effectively yields the axisymmetric multiphase flow equations, Eqs. (\ref{eq:axiscont})- (\ref{eq:axiseos}), in a self-consistent way. To facilitate this, we employ the following coordinate transformation, illustrated in Fig.~\ref{fig:schemaxis}, which allows the governing equations to be represented in a Cartesian-like coordinate system, i.e. $(x,y)$: \begin{figure*} \includegraphics{fig1 \caption{\label{fig:schemaxis} Schematic of arrangement of coordinate system in axisymmetric multiphase flow ($(r,z)$ and $(y,x)$ coordinate directions are shown).} \end{figure*} \begin{equation} (r,z) \rightarrow (y,x), \end{equation} \begin{equation} (u_r,u_z) \rightarrow (u_y,u_x). \end{equation} Assuming summation convention for repeated subscript indices, Eqs. (\ref{eq:axiscont})-(\ref{eq:surfz}) may be transformed to \begin{equation} \partial_t \rho + \partial_k \left( \rho u_k \right)=-\frac{\rho u_y}{y}, \label{eq:axiscont1} \end{equation} \begin{equation} \rho \left( \partial_t u_i+ u_k\partial_k u_i \right)=-\partial_i P + F_{s,i}+F_{ext,i}+ \partial_k \left[ \mu\left( \partial_k u_i+\partial_i u_k \right) \right]+F_{ax,i}, \label{eq:axismom1} \end{equation} where \begin{equation} F_{s,i}=\kappa \rho \partial_i \nabla^2 \rho \label{eq:forcemp} \end{equation} and $i,j,k\in \left\{ x,y \right\}$. The right hand side (RHS) in Eq. (\ref{eq:axiscont1}), $-\rho u_y/y$, is the additional term in the continuity equation that arises from axisymmetry. The corresponding term for the momentum equation, Eq. (\ref{eq:axismom1}), is \begin{equation} F_{ax,i}=\frac{\mu}{y}\left[\partial_y u_i+\partial_i u_y \right]+ \kappa \rho \partial_i \left( \frac{1}{y}\partial_y \rho \right). \end{equation} To recover Eqs. (\ref{eq:axiscont1}) and (\ref{eq:axismom1}), we introduce two additional source terms, $S_{\alpha}^{'}$ and $S_{\alpha}^{''}$, to the standard 2D Cartesian LBE which has $\Omega_{\alpha}$ as its collision term and a source term for the internal and external forces, $S_{\alpha}$. These unknown additional terms, representing the axisymmetric mass and momentum contributions respectively, are to be determined so that the macroscopic behavior of the proposed LBE corresponds to axisymmetric multiphase flow. Thus, we propose the following LBE \begin{eqnarray} f_{\alpha}( \mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha}\delta_t,t+\delta_t )-f_{\alpha}( \mbox{\boldmath$x$},t )&=& \frac{1}{2}\left[\Omega_{\alpha}|_{(x,t)}+ \Omega_{\alpha}|_{(x+e_{\alpha}\delta_t,t+\delta_t)} \right]+\nonumber\\ & & \frac{1}{2}\left[S_{\alpha}|_{(x,t)}+ S_{\alpha}|_{(x+e_{\alpha}\delta_t,t+\delta_t)} \right]\delta_t+\nonumber\\ & & \frac{1}{2}\left[S_{\alpha}^{'}|_{(x,t)}+ S_{\alpha}^{'}|_{(x+e_{\alpha}\delta_t,t+\delta_t)} \right]\delta_t+\nonumber\\ & & \frac{1}{2}\left[S_{\alpha}^{''}|_{(x,t)}+ S_{\alpha}^{''}|_{(x+e_{\alpha}\delta_t,t+\delta_t)} \right]\delta_t, \label{eq:axislbe} \end{eqnarray} where $f_{\alpha}$ is the discrete single-particle distribution function, corresponding to the particle velocity, $\mbox{\boldmath$e$}_{\alpha}$, where $\alpha$ is the velocity direction. The Cartesian component of the particle velocity, $c$, is given by $c=\delta_x/\delta_t$, where $\delta_x$ is the lattice spacing and $\delta_t$ is the time step corresponding to the two-dimensional, nine-velocity model(D2Q9)~\cite{qian92} shown in Fig.~\ref{fig:schemaxis}. Here, the collision term is given by the BGK approximation~\cite{bhatnagar54} \begin{equation} \Omega_{\alpha}=-\frac{f_{\alpha}-f_{\alpha}^{eq}}{\tau}, \quad \tau=\frac{\lambda}{\delta_t}, \end{equation} where $\lambda$ is the relaxation time due to collisions, $\delta_t$ is the time step and $f_{\alpha}^{eq}$ is the truncated discrete form of the Maxwellian \begin{equation} f_{\alpha}^{eq}\equiv f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$})= \omega_{\alpha}\left\{ 1+\frac{\mbox{\boldmath$e$}_{\alpha} \cdotp \mbox{\boldmath$u$}}{RT}+ \frac{\left( \mbox{\boldmath$e$}_{\alpha} \cdotp \mbox{\boldmath$u$} \right)^2}{2(RT)^2}- \frac{1}{2}\frac{\mbox{\boldmath$u$} \cdotp \mbox{\boldmath$u$}}{RT} \right\}, \label{eq:trunceq} \end{equation} where $R$ is the gas constant, $T$ is the temperature and $w_{\alpha}$ is the weighting coefficients in the Gauss-Hermite quadrature to represent the kinetic moment integrals of the distribution functions exactly~\cite{he97}. For isothermal flows, the factor $RT$ is related to the particle speed $c$ as $RT=1/3c^2$. The term in Eq. (\ref{eq:axislbe}) \begin{equation} S_{\alpha}=\frac{(e_{\alpha j}-u_j)(F_j+F_{ext,j})}{\rho RT}f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$}) \label{eq:sourcemp} \end{equation} represents the effect of internal and external forcing terms on the change in the distribution function. The internal force term gives rise to surface tension and phase segregation effects which are given by \begin{equation} F_j=-\partial_j \psi + F_{s,j}, \label{eq:forceint} \end{equation} where the function $\psi=P-\rho RT$ is the non-ideal part of the equation of state given in Eq. (\ref{eq:axiseos}). The first two terms on the RHS of Eq. (\ref{eq:axislbe}) corresponds to those presented by He \emph{et al.} (1998). As mentioned above, the last two terms, $S_{\alpha}^{'}$ and $S_{\alpha}^{''}$, in this equation is to be selected such that its behavior in the continuum limit would simulate the influence of the non-Cartesian-like terms in Eqs. (\ref{eq:axiscont1}) and (\ref{eq:axismom1}) in a self-consistent way. Since the zeroth kinetic moment of the term $f_{\alpha}^{eq,M}(\rho,0)$ is involved in the derivation of the macroscopic mass conservation equation from the LBE, the source term $S_{\alpha}^{'}$ in Eq. (\ref{eq:axislbe}) is proposed to be equal to $f_{\alpha}^{eq,M}(\rho,0)$ multiplied by an unknown $m^{'}$ and normalized by the density $\rho$. The other source term $S_{\alpha}^{''}$ is proposed analogous to the source term in Eq.(\ref{eq:sourcemp}). Thus, we propose \begin{eqnarray} S_{\alpha}^{'}&=&\frac{f_{\alpha}^{eq,M}(\rho,0)}{\rho}m^{'}\label{eq:sourcea1},\\ S_{\alpha}^{''}&=&\frac{(e_{\alpha j}-u_j)F_j^{''}}{\rho RT}f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$}). \label{eq:sourcea2} \end{eqnarray} Here the unknowns, $m^{'}$ and $F_j^{''}$, in the above two equations can be determined through Chapman-Enskog analysis as will be shown later. It must be stressed that all terms, including the collision term, on the RHS are discretized by the application of the trapezoidal rule, since it has been argued that at least a second-order treatment of the source terms is necessary for simulation of multiphase flow~\cite{he98,he99}. The macroscopic fields are given by \begin{eqnarray} \rho&=&\sum_{\alpha} f_{\alpha},\\ \rho u_i&=&\sum_{\alpha} f_{\alpha} e_{\alpha i}. \label{eq:amacrofields} \end{eqnarray} In this model, the order parameter is the density, $\rho$, which distinguishes the different phases in the flow. Equation (\ref{eq:axislbe}) is implicit in time. To remove implicitness in this equation we introduce a transformation following the procedure described by He and others~\cite{he98,he98a}, whereby \begin{equation} \bar{f}_{\alpha}=f_{\alpha}-\frac{1}{2}\Omega_{\alpha}- \frac{1}{2}\left(S_{\alpha}+S_{\alpha}^{'}+S_{\alpha}^{''}\right)\delta_t \label{eq:implicittr} \end{equation} in Eq. (\ref{eq:axislbe}), so that we obtain \begin{equation} \bar{f}_{\alpha}( \mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha}\delta_t,t+\delta_t )- \bar{f}_{\alpha}( \mbox{\boldmath$x$},t )= \bar{\Omega}_{\alpha}|_{(x,t)}+\frac{\tau}{\tau+1/2}\left[ S_{\alpha}+S_{\alpha}^{'}+ S_{\alpha}^{''}\right]|_{(x,t)}\delta_t, \label{eq:axislbee} \end{equation} where \begin{equation} \bar{\Omega}_{\alpha}=-\frac{\bar{f}_{\alpha}-f_{\alpha}^{eq}}{\tau+1/2}. \label{eq:axisbgk} \end{equation} Thus, $\bar{f}_{\alpha}$ is the transformed distribution function that removes implicitness in the proposed LBE, Eq. (\ref{eq:axislbe}), which describes the evolution of the $f_{\alpha}$ distribution function. The following constraints on the equilibrium distribution and the various source terms~\cite{luo00,guo02} are imposed from their definition: \begin{equation} \sum_{\alpha} f_{\alpha}^{eq}=\rho, \quad \sum_{\alpha} f_{\alpha}^{eq} e_{\alpha i}=\rho u_i, \quad \sum_{\alpha} f_{\alpha}^{eq} e_{\alpha i} e_{\alpha j}=\rho RT \delta_{ij}+\rho u_i u_j, \nonumber \end{equation} \begin{equation} \sum_{\alpha} f_{\alpha}^{eq} e_{\alpha i} e_{\alpha j} e_{\alpha k}=\rho (RT)^2 \left( u_i \delta_{jk}+u_j \delta_{ki}+u_k \delta_{ij}\right), \end{equation} \begin{equation} \sum_{\alpha} S_{\alpha}=0, \quad \sum_{\alpha} S_{\alpha} e_{\alpha i} = F_i, \quad \sum_{\alpha} S_{\alpha} e_{\alpha i} e_{\alpha j}=(F_i+F_{ext,i}) u_j+(F_j+F_{ext,j}) u_i, \end{equation} \begin{equation} \sum_{\alpha} S_{\alpha}^{'}=m^{'}, \quad \sum_{\alpha} S_{\alpha}^{'} e_{\alpha i} = 0, \quad \sum_{\alpha} S_{\alpha}^{'} e_{\alpha i} e_{\alpha j}=m^{'}RT\delta_{ij}, \end{equation} \begin{equation} \sum_{\alpha} S_{\alpha}^{''}=0, \quad \sum_{\alpha} S_{\alpha}^{''} e_{\alpha i} = F_i^{''}, \quad \sum_{\alpha} S_{\alpha}^{''} e_{\alpha i} e_{\alpha j}=(F_i^{''} u_j+F_j^{''} u_i). \end{equation} Then the following relationships are obtained between the transformed distribution function and the macroscopic fields, which also include the curvature effects resulting from axial symmetry: \begin{eqnarray} \rho&=&\sum_{\alpha} \bar{f}_{\alpha}+\frac{1}{2}m^{'}\delta_t \label{eq:dens1},\\ \rho u_i&=&\sum_{\alpha} \bar{f}_{\alpha} e_{\alpha i}+\frac{1}{2}(F_i+F_{ext,i}+F_i^{''})\delta_t. \label{eq:densvel1} \end{eqnarray} Now, to establish the unknowns $m^{'}$ and $F_j^{''}$ in the above formulation, the Chapman-Enskog multiscale analysis is performed~\cite{chapman}. Introducing the expansions~\cite{he97a} \begin{eqnarray} \bar{f}_{\alpha}( \mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha}\delta_t,t+\delta_t )&=& \sum_{\alpha=0}^{\infty} D_{t_n}\bar{f}_{\alpha}(\mbox{\boldmath$x$},t),\\ D_{t_n} &\equiv& \partial_{t_n}+e_{\alpha k} \partial_k,\\ f_{\alpha}&=&\sum_{\alpha=0}^{\infty}\epsilon^n f_{\alpha}^{(n)},\\ \partial_t&=&\sum_{\alpha=0}^{\infty}\epsilon^n \partial_{t_n}, \end{eqnarray} where $\epsilon=\delta_t$ in Eq. (\ref{eq:axislbee}) and using Eq. (\ref{eq:implicittr}) to transform $\bar{f}_\alpha$ back to $f_\alpha$, the following equations are obtained in the consecutive order of the parameter $\epsilon$: \begin{eqnarray} O(\epsilon^0): f_{\alpha}^{(0)}&=&f_{\alpha}^{eq}\label{eq:order0},\\ O(\epsilon^1): D_{t_0} f_{\alpha}^{(0)}&=&-\frac{1}{\tau} f_{\alpha}^{(1)}+S_{\alpha}+ S_{\alpha}^{'}+S_{\alpha}^{''}\label{eq:order1},\\ O(\epsilon^2): \partial_{t_1} f_{\alpha}^{(0)}+ D_{t_0} f_{\alpha}^{(1)}&=&-\frac{1}{\tau} f_{\alpha}^{(2)}. \label{eq:order2} \end{eqnarray} Now, invoking the Chapman-Enskog ansatz \begin{equation} \sum_{\alpha} \left( \begin{array}{c} 1 \\ e_{\alpha i} \end{array} \right)f_{\alpha}^{(0)}= \left( \begin{array}{c} \rho \\ \rho u_i \end{array} \right), \sum_{\alpha} \left( \begin{array}{c} 1 \\ e_{\alpha i} \end{array} \right)f_{\alpha}^{(n)}= \left( \begin{array}{c} 0 \\ 0 \end{array} \right),n \geq 1 \end{equation} and performing $\sum_{\alpha}(\cdotp)$ on Eqs. (\ref{eq:order1}) and (\ref{eq:order2}), we obtain \begin{eqnarray} \partial_{t_0} \rho + \partial_k (\rho u_k)&=& m^{'} \label{eq:axisc1},\\ \partial_{t_1} \rho &=& 0,\label{eq:axisc2} \end{eqnarray} respectively. Combining the first- and second- order results given by Eqs. (\ref{eq:axisc1}) and (\ref{eq:axisc2}) and considering $\partial_t=\partial_{t_0}+\epsilon \partial_{t_1}$, we get \begin{equation} \partial_t \rho + \partial_k (\rho u_k)= m^{'} \label{eq:axisc}.\\ \end{equation} Comparing this equation and Eq. (\ref{eq:axiscont1}), the unknown $m^{'}$ is obtained as \begin{equation} m^{'}=-\frac{\rho u_y}{y}. \label{eq:mvalue}\\ \end{equation} This is the axisymmetric contribution to the Cartesian form of the equation for the order parameter, i.e., density characterizing the different phases of the flow. Taking the first kinetic moment, $\sum_{\alpha}e_{\alpha i}(\cdotp)$, of Eqs. (\ref{eq:order1}) and (\ref{eq:order2}), respectively, we get \begin{eqnarray} \partial_{t_0} (\rho u_i) + \partial_k (\rho u_i u_k)&=& -\partial_i (\rho RT)+F_i+F_{ext,i}+F_i^{''}, \label{eq:axism1}\\ \partial_{t_1} (\rho u_i) + \partial_k \Pi_{ij}^{(1)}&=& 0, \label{eq:axism2} \end{eqnarray} where \begin{equation} \Pi_{ij}^{(1)}=\sum_{\alpha} f_{\alpha}^{(1)}e_{\alpha i} e_{\alpha j}. \label{eq:visct1} \end{equation} Employing the expression for $f_{\alpha}^{(1)}$ from Eq. (\ref{eq:order1}) in Eq. (\ref{eq:visct1}), together with the summational constraints given above, and neglecting terms of the order $O(Ma^3)$ or higher, we get \begin{equation} \Pi_{ij}^{(1)}=-\tau RT \rho (\partial_j u_i+\partial_i u_j). \label{eq:visct2} \end{equation} Equation (\ref{eq:axism2}) then simplifies to \begin{equation} \partial_{t_1} (\rho u_i) = \partial_j \left( \tau RT \rho (\partial_j u_i+\partial_i u_j) \right) \label{eq:axism2n}. \end{equation} Combining Eqs. (\ref{eq:axism1}) and (\ref{eq:axism2n}), we get \begin{equation} \partial_t (\rho u_i) + \partial_k (\rho u_i u_k)= -\partial_i (\rho RT)+F_i+F_{ext,i}+F_i^{''}+ \partial_j \left( \tau \delta_t RT \rho (\partial_j u_i+\partial_i u_j) \right), \end{equation} or substituting for $F_i$ from Eq. (\ref{eq:forceint}), we obtain \begin{equation} \partial_t (\rho u_i) + \partial_k (\rho u_i u_k)= -\partial_i P +F_{s,i}+F_{ext,i}+F_i^{''}+ \partial_j \left( \tau \delta_t RT \rho (\partial_j u_i+\partial_i u_j) \right). \label{eq:axism3} \end{equation} Using Eqs. (\ref{eq:axisc}) and (\ref{eq:axism3}), this can be simplified to \begin{eqnarray} \rho \left( \partial_t u_i + u_k \partial_k u_i \right)-\frac{\rho u_i u_y}{y}&=& -\partial_i P +F_{s,i}+F_{ext,i}+F_i^{''}+\nonumber\\ & & \partial_j \left( \tau \delta_t RT \rho (\partial_j u_i+\partial_i u_j) \right). \label{eq:axism4} \end{eqnarray} Comparing Eqs. (\ref{eq:axismom1}) and (\ref{eq:axism4}), we obtain the other unknown $F_i^{''}$ where \begin{equation} F_i^{''}=F_{ax,i}-\frac{\rho u_i u_y}{y}=\frac{\mu}{y}\left[ \partial_y u_i + \partial_i u_y \right]+ \kappa \rho \partial_i \left(\frac{1}{y}\partial_y \rho \right)- \frac{\rho u_i u_y}{y}. \label{eq:fvalue} \end{equation} This is the axisymmetric contribution to the Cartesian form of the equation for the momentum, where the first, second and the third terms on the RHS correspond to the viscous, surface tension and inertial force contributions, respectively. The dynamic viscosity is related to the relaxation time for collisions by $\mu=\rho \tau \delta_t RT = \rho \lambda c_s^2$, where $c_s^2=1/3c^2$. The set of equations corresponding to the axisymmetric LBE multiphase flow model is given by Eqs. (\ref{eq:axislbee}) and (\ref{eq:axisbgk}) together with Eqs. (\ref{eq:sourcemp}), (\ref{eq:sourcea1}) and (\ref{eq:sourcea2}), (\ref{eq:dens1}) and (\ref{eq:densvel1}), and (\ref{eq:mvalue}) and (\ref{eq:fvalue}). In general, this multiphase model and that proposed by He and others~\cite{he98} face difficulties for fluids far from the critical point and/or in the presence of external forces. This difficulty is related to the calculation of the intermolecular force in Eq.(\ref{eq:forceint}), involving the computation of $\partial_j\psi$ which can become quite large across interfaces. Unless this term is accurately computed, the model may become unstable because of numerical errors~\cite{he99a,he04}. Hence, an improved treatment of this term is necessary. This will now be described. \section{\label{sec:axismodelc}Axisymmetric LBE Multiphase Flow Model with Reduced Compressibility Effects} He and co-workers~\cite{he99} have proposed that through a suitable transformation of the distribution function, $f_{\alpha}$, which involves invoking the incompressibility condition of the fluid, and employing a new distribution function for capturing the interface, the difficulty with handling the intermolecular force term, $\partial_j\psi$, can be reduced. We apply this idea to the axisymmetric model developed in the previous section. We replace the distribution function $f_{\alpha}$ by another distribution function $g_{\alpha}$ through the transformation~\cite{he99} \begin{equation} g_{\alpha}=f_{\alpha}RT+\psi(\rho)\frac{f_{\alpha}^{eq,M}(\rho,0)}{\rho}. \label{eq:transg} \end{equation} The effect of this transformation will be discussed in greater detail below. By considering the fluid to be incompressible, i.e. \begin{equation} \frac{d}{dt}\psi(\rho)=\left( \partial_t+u_k \partial_k \right)\psi (\rho)=0, \end{equation} and using the transformation Eqs. (\ref{eq:transg}) and (\ref{eq:implicittr}), Eq. (\ref{eq:axislbee}) is replaced by \begin{equation} \bar{g}_{\alpha}( \mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha}\delta_t,t+\delta_t )- \bar{g}_{\alpha}( \mbox{\boldmath$x$},t )= \bar{\Omega}_{g\alpha}|_{(x,t)}+\frac{\tau}{\tau+1/2}\left[ S_{g\alpha}+S_{g\alpha}^{'}+ S_{g\alpha}^{''}\right]|_{(x,t)}\delta_t, \label{eq:axislbeg} \end{equation} where \begin{equation} \bar{\Omega}_{g\alpha}=-\frac{\bar{g}_{\alpha}-g_{\alpha}^{eq}}{\tau+1/2}, \label{eq:axisbgkg} \end{equation} and \begin{equation} g_{\alpha}^{eq}=f_{\alpha}^{eq}RT+\psi(\rho)\frac{f_{\alpha}^{eq,M}(\rho,0)}{\rho}. \end{equation} The corresponding source terms become \begin{eqnarray} S_{g\alpha}&=&(e_{\alpha j}-u_j)\times \nonumber\\ & & \left[ (F_j+F_{ext,j})\frac{f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$})}{\rho}- \left( \frac{f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$})}{\rho}-\frac{f_{\alpha}^{eq,M}(\rho,0)}{\rho} \right)\partial_j \psi(\rho) \right], \label{eq:srcrefined} \end{eqnarray} \begin{equation} S_{g\alpha}^{'}=S_{\alpha}^{'}RT=\frac{f_{\alpha}^{eq,M}(\rho,0)}{\rho}\left( -\frac{\rho u_y}{y} \right)RT, \end{equation} \begin{equation} S_{g\alpha}^{''}=S_{\alpha}^{''}RT=(e_j-u_j)F_j^{''}\frac{f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$})}{\rho}. \label{eq:sourceg2} \end{equation} The term $\partial_j \psi$ in Eq. (\ref{eq:srcrefined}) is multiplied by the factor $\left( f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$})/\rho-f_{\alpha}^{eq,M}(\rho,0)/\rho \right)$. This factor, from the definition of the equilibrium distribution function, $f_{\alpha}^{eq}$, in Eq. (\ref{eq:trunceq}) is proportional to the Mach number and thus becomes smaller in the incompressible limit. Hence, it alleviates the difficulties associated with the calculation of the $\partial_j \psi$, a major source of numerical instability with the original model~\cite{he98}. Thus, Eqs. (\ref{eq:axislbeg})-(\ref{eq:sourceg2}) are found to be numerically more stable compared to Eq. (\ref{eq:axislbee}) supplemented with Eqs. (\ref{eq:sourcemp}),(\ref{eq:sourcea1}) and (\ref{eq:sourcea2}). In this new framework, we still need to introduce an order parameter to capture interfaces. Here, we employ a function, $\phi$, referred to henceforth as the index function, in place of the density, as the order parameter to distinguish the phases in the flow. The evolution equation of the distribution function whose emergent dynamics govern the index function has to be able to maintain phase segregation and mass conservation. To do this, we employ Eq. (\ref{eq:axislbee}) together with Eqs. (\ref{eq:sourcemp}),(\ref{eq:sourcea1}) and (\ref{eq:sourcea2}) by keeping the term involving $\partial_j \psi$ and $m^{'}$, while the rest of the terms may be dropped as they play no role in mass conservation. In addition, the density is replaced by the index function in these equations. Hence, the evolution of the distribution function for the index function is given by \begin{equation} \bar{f}_{\alpha}( \mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha}\delta_t,t+\delta_t )- \bar{f}_{\alpha}( \mbox{\boldmath$x$},t )= \bar{\Omega}_{f\alpha}|_{(x,t)}+\frac{\tau}{\tau+1/2}\left[ S_{f\alpha}+ S_{f\alpha}^{'}\right]|_{(x,t)}\delta_t, \label{eq:axislbef} \end{equation} where the collision and the source terms are given by \begin{equation} \bar{\Omega}_{f\alpha}=-\frac{\bar{f}_{\alpha}-\frac{\phi}{\rho}f_{\alpha}^{eq}}{\tau+1/2}, \label{eq:axisbgkf} \end{equation} \begin{equation} S_{f\alpha}=\frac{(e_j-u_j)(-\partial_j \psi(\phi))}{\rho RT}f_{\alpha}^{eq,M}(\rho,\mbox{\boldmath$u$}), \label{eq:srciref} \end{equation} \begin{equation} S_{f\alpha}^{'}=\frac{\phi}{\rho}S_{\alpha}^{'}= \frac{f_{\alpha}^{eq,M}(\rho,0)}{\rho}\left( -\frac{\phi u_y}{y} \right). \end{equation} The hydrodynamic variables such as pressure and fluid velocity can be obtained by taking appropriate kinetic moments of the distribution function $g_{\alpha}$, i.e. \begin{eqnarray} P&=&\sum_{\alpha} \bar{g}_{\alpha}-\frac{1}{2}u_j\partial_j \psi(\rho)+\frac{1}{2}m^{'}RT\delta_t,\label{eq:axpres}\\ \rho RT u_i &=& \sum_{\alpha} \bar{g}_{\alpha} e_{\alpha i}+\frac{1}{2}\left( F_{s,i}+F_{ext,i} \right)\delta_t+ \frac{1}{2} F_i^{''}\delta_t. \end{eqnarray} This follows from the definition of $\bar{g}_{\alpha}$ given in Eq. (\ref{eq:transg}) and also includes curvature effects. The index function is obtained from the distribution function $\bar{f}_{\alpha}$ by taking the zeroth kinetic moment, i.e. \begin{equation} \phi=\sum_{\alpha}\bar{f}_{\alpha}+\frac{1}{2}\frac{\phi}{\rho}m^{'}\delta_t. \end{equation} The terms $m^{'}$ and $F_i^{''}$ are given in Eqs. (\ref{eq:mvalue}) and (\ref{eq:fvalue}), respectively. The density is obtained from the index function through linear interpolation, i.e. \begin{equation} \rho(\phi)= \rho_L+\frac{\phi-\phi_L}{\phi_H-\phi_L}(\rho_H-\rho_L), \label{eq:rintp} \end{equation} where $\rho_L$ and $\rho_H$ are the densities of the light and heavy fluids, respectively, and $\phi_L$ and $\phi_H$ refer to the minimum and maximum values of the index function, respectively. These limits of the index function are determined from Maxwell's equal area construction~\cite{rowlinson} applied to the function $\psi(\phi)+\phi RT$. Thus, the axisymmetric LBE multiphase flow model with reduced compressibility effects corresponds to Eqs. (\ref{eq:axislbeg})-(\ref{eq:rintp}). The relaxation time for collisions is related to the viscosity of the fluid using the same expression as derived in the previous section. If the kinematic viscosity of the light fluid, $\nu_L$, is different from that of the heavy fluid, $\nu_H$, its value at any point in the fluid is obtained from the index function through linear interpolation, i.e. \begin{equation} \nu(\phi)= \nu+\frac{\phi-\phi_L}{\phi_H-\phi_L}(\nu-\nu_L). \label{eq:vintp} \end{equation} It may be seen that the model requires the calculation of spatial gradients in Eqs. (\ref{eq:srcrefined}) and (\ref{eq:srciref}) and of the Laplacian in Eq. (\ref{eq:forcemp}). Since maintaining accuracy as well as isotropy is important for the surface tension terms, they are calculated by employing a fourth-order finite-difference scheme for the gradient and a second-order scheme for the Laplacian, given respectively by \begin{equation} \partial_i \varpi=\frac{1}{36\delta_x}\sum_{\alpha=1}^{8}\left[ 8\varpi(\mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha i}\delta_t)- \varpi(\mbox{\boldmath$x$}+2\mbox{\boldmath$e$}_{\alpha i}\delta_t) \right] \left( \frac{e_{\alpha i}}{c} \right)+O(\delta_t^4), \end{equation} and \begin{equation} \nabla^2 \varpi \equiv \partial_i \partial_i \varpi = \frac{1}{3\delta_x^2} \sum_{\alpha=1}^8\left[ \varpi(\mbox{\boldmath$x$}+\mbox{\boldmath$e$}_{\alpha i}\delta_t)- \varpi(\mbox{\boldmath$x$}) \right]+O(\delta_x^2), \end{equation} for any function $\varpi$. Notice that these discretizations are both based on the lattice based stencil, instead of the standard stencil based on the coordinate directions. In addition, in the application of this model, the implementation of boundary conditions plays an important role. In particular, along the axisymmetric line, i.e. $y=0$, specular reflection boundary conditions are employed for the distribution functions. For the two-dimensional, nine velocity (D2Q9) model shown in the inset of Fig.~\ref{fig:schemaxis}, we set $\bar{f}_2=\bar{f}_4$, $\bar{f}_5=\bar{f}_8$, $\bar{f}_6=\bar{f}_7$ and $\bar{g}_2=\bar{g}_4$, $\bar{g}_5=\bar{g}_8$ and $\bar{g}_6=\bar{g}_7$ for the distribution functions after the streaming step. For macroscopic conditions, along this line, $u_y=\partial_y(\cdotp)=0$, through which the singular source terms of type $1/y(\cdotp)$ in the model can be appropriately treated. On the other hand, boundary conditions along the other lines are similar to those for the standard LBE. \section{\label{sec:results}Results and Discussion} In the rest of this paper, unless otherwise specified, the results are presented in lattice units, i.e. the velocities are scaled by the particle velocity $c$, the distance by the minimum lattice spacing $\delta_x$ and time by $c/\delta_x$. All other quantities are scaled as appropriate combinations of these basic units. First, the axisymmetric LBE multiphase flow models are applied to verify the well-known Laplace-Young relation for an axisymmetric drop. According to this relation, $\Delta P=2 \sigma/R_d$, where $\Delta P$ is the difference between the pressure inside and outside of a drop, $\sigma$ is the surface tension and $R_d$ is the drop radius. For different choices of the surface tension parameter, $\kappa$, the surface tension values are obtained from Eq. (\ref{eq:sigmakappa}) by the replacing density in Eq. (\ref{eq:surfr}) and (\ref{eq:surfz}) by the index function. To obtain the normal gradient used in Eq. (\ref{eq:sigmakappa}), a physical configuration consisting of a liquid and a gas layer is set up. Once equilibrium is reached, the density gradient may be computed and hence the surface tension. Having obtained the relationship between the surface tension $\sigma$, and the parameter $\kappa$, axisymmetric drops of four different radii, $R_d=40, 50, 60$ and $70$, are set up in a domain discretized by $201\times 101$ lattice sites. Periodic boundaries are considered in the $x$ direction and an open boundary condition is considered along the boundary that is parallel to the axisymmetric boundary. By considering three different values of $\kappa$, $0.05, 1.0$ and $0.15$, the pressure difference across the drops is determined. Figure~\ref{fig:laplaceyoung1} shows a comparison of the pressure difference across the interface of the drops computed using the axisymmetric model developed in Section \ref{sec:axismodelc} and that predicted by the Laplace-Young relation. It is found that the computed results are in good agreement with the theoretical values, with \begin{figure} \begin{center} \includegraphics{fig2.eps} \caption{\label{fig:laplaceyoung1}Pressure difference across axisymmetric drops as a function of radius for different values of the surface tension parameter $\kappa$; Comparison of computed results using the axisymmetric LBE model versus theoretical prediction based on the Laplace-Young relation. Quantities are in lattice units.} \end{center} \end{figure} a maximum relative error of about $3\%$. Another important test problem is that of an oscillating axisymmetric drop immersed in a gas. Since current versions of the LBE simulate a relatively viscous fluid, it is appropriate to compare the oscillation frequency with that of Miller and Scriven (1968)~\nocite{miller68}. In contrast to earlier analytical solutions on drop oscillations, this work considers viscous dissipation effects in the boundary layer at the interface. According to~\cite{miller68}, the frequency for the $n^{th}$ mode of oscillation for a drop is given by \begin{equation} \omega_{n}=\omega_{n}^{*}-\frac{1}{2} \alpha \omega_{n}^{*\frac{1}{2}}+\frac{1}{4}\alpha^2, \label{eq:msperiod} \end{equation} where $\omega_{n}$ is the angular response frequency, and $\omega_{n}^{*}$ is Lamb's natural resonance frequency expressed as~\cite{lamb} \begin{equation} \left(\omega_{n}^{*}\right)^{2}= \frac{n(n+1)(n-1)(n+2)}{R_d^3 \left[ n\rho_{g}+(n+1)\rho_{l} \right]} \sigma. \end{equation} $R_d$ is the equilibrium radius of the drop, $\sigma$ is the interfacial surface tension, and $\rho_{l}$ and $\rho_{g}$ are the densities of the two fluids. The parameter $\alpha$ is given by \begin{equation} \alpha = \frac{(2n+1)^2 (\mu_{l} \mu_{g} \rho_{l} \rho_{g})^{\frac{1}{2}}} {2^{\frac{1}{2}} R_d \left[ n\rho_{g}+(n+1)\rho_{l} \right] \left[ (\mu_{l} \rho_{l})^{\frac{1}{2}}+ (\mu_{g} \rho_{g})^{\frac{1}{2}} \right]}, \end{equation} where $\mu_{l}$ and $\mu_{g}$ are the dynamic viscosity of the two liquids. The subscripts $g$ and $l$ refer to the ambient gas and liquid phases, respectively. We consider the second mode of oscillation and analytical expressions for the time period are presented in Eq. (\ref{eq:msperiod}). The initial computational setup consists of a prolate spheroid of minimum ($R_{min}$) and maximum ($R_{max}$) radii of $40$ and $55$, respectively, placed in the center of the domain discretized by $201 \times 101$ lattice sites. We consider the surface tension parameters: $\kappa=0.2$, and the density of the gas and the drop to be $\rho_g=0.1$ and $\rho_l=0.4$, respectively. The kinematic viscosity of both the gas and the drop are considered to be the same and given by $\nu_g=\nu_l=1.6667\times 10^{-2}$. Figure~\ref{fig:oscchem} shows the configurations of an oscillating drop at different times computed using the standard axisymmetric model with these conditions. \begin{figure} \begin{center} \includegraphics{fig3.EPS} \caption{\label{fig:oscchem}Configurations of an oscillating drop as a function of time; $R_{min}=40$, $R_{max}=55$, $\rho_g=0.1$, $\rho_l=0.4$, $\nu_l=\nu_g=1.6667\times 10^{-2}$. Quantities are in lattice units.} \end{center} \end{figure} The drop changes from a prolate shape at $t=2000$ to oblate shape at $t=16000$. Such shape changes continue till the drop reaches its equilibrium spherical shape. Figure \ref{fig:dropint1} shows the temporal evolution of the interface locations of the oscillating drop with the conditions above for two different surface tension parameter: $\kappa=0.02$ and $0.08$. \begin{figure} \begin{center} \includegraphics{fig4.eps} \caption{\label{fig:dropint1}Interface location of an oscillating drop as a function of time for two values of the surface tension parameter $\kappa$; $R_{min}=40$, $R_{max}=55$, $\rho_g=0.1$, $\rho_l=0.4$, $\nu_l=\nu_g=1.6667\times 10^{-2}$. Quantities are in lattice units.} \end{center} \end{figure} It is expected that increasing the surface tension will reduce the time period of oscillations. The computed ($T_{LBE}$) and analytical ($T_{anal}$) time periods, where $T_{anal}=2\pi/\omega_2$, when $\kappa=0.02$ are $29483$ and $29448$ respectively. As $\kappa$ is increased to $0.08$, $T_{LBE}$ and $T_{anal}$ become $14388$ and $14313$ respectively. It may be seen that the computed and analytical values agree well, the difference being less than $1\%$. Also, the time period decreases as $\kappa$ is increased, which is consistent with expectations. Consider next the effect of changing the drop size on the time period of oscillations. Figure \ref{fig:dropint2} shows the interface locations of an oscillating drop as a function of time for the following two initial sizes: $R_{min}=30$ and $R_{max}=45$; $R_{min}=40$, $R_{max}=55$. Reducing the drop size reduces its time period. \begin{figure} \begin{center} \includegraphics{fig5.eps} \caption{\label{fig:dropint2}Interface location of an oscillating drop as a function of time for two drop sizes; $\rho_g=0.1$, $\rho_l=0.4$, $\nu_l=\nu_g=1.6667\times 10^{-2}$, $\kappa=0.02$. Quantities are in lattice units.} \end{center} \end{figure} The computed time period of the larger drop is equal to $29483$, while that for the smaller drop is $20118$. Comparison of the computed time periods with the analytical solution shows that they agree within $1\%$ for these cases. Next, consider three different kinematic viscosities of the liquid: $\nu_l=1.6667\times 10^{-2}, 3.3333\times 10^{-2}$ and $5.0\times 10^{-2}$. Figure \ref{fig:dropint3} shows the effect of drop viscosity on the temporal evolution of the interface locations of the drop. \begin{figure} \begin{center} \includegraphics{fig6.eps} \caption{\label{fig:dropint3}Interface location of an oscillating drop as a function of time for different kinematic viscosities $\nu_l$; $R_{min}=40$, $R_{max}=55$, $\rho_g=0.1$, $\rho_l=0.4$, $\kappa=0.02$. Quantities are in lattice units.} \end{center} \end{figure} It is found that as the kinematic viscosity is increased the time period increases moderately which is consistent with the analytical solution. The computed time periods at these viscosities are $29483$, $31030$ and $32925$, while the analytical values are $29448$, $30597$ and $31318$, respectively, with a maximum error within $5.1\%$. The third test problem considered here is that of the break-up of a cylindrical liquid column into drops, a fascinating problem of long standing theoretical and practical interest. In a seminal work, Rayleigh (1878)~\nocite{rayleigh78} showed through a linear stability analysis of an inviscid column of cylindrical liquid of radius $R_c$ that the column will be unstable if the axisymmetric wavelength of any disturbance $\lambda_d$ is longer than its circumference, i.e. the wave number $k^{*}=2\pi R_c/\lambda_d$ should be less than one. Later, the theoretical analysis was extended to more realistic conditions by including viscosity. In the last three decades, several experimental and numerical investigations have also been performed. To evaluate the axisymmetric LBE model, we study the Rayleigh capillary instability for different wavenumbers. Initial studies carried out with $k^{*}>1$ showed that the liquid does not break-up. We will now present results of cases with break-up. Consider a cylindrical liquid column of radius $R_c=45$ subject to an axisymmetric co-sinusoidal wavelength $\lambda_d=320$, i.e. $k^{*}=0.88$. To simulate the dynamics of instability for this wavenumber, we consider a domain discretized by $321\times 151$ lattice sites with $\rho_g=0.1$, $\rho_l=0.4$, $\nu_g=\nu_l=6.6667\times 10^{-2}$ and $\kappa=0.1$. Since $k^{*}<1$, it is expected that the liquid column would eventually breakup. Figure \ref{fig:rayleigh1} shows the configurations of the liquid column at different times. As time progresses, the imposed interfacial disturbances on the \begin{figure} \begin{center} \includegraphics[height=6.50in,clip=]{fig7.EPS} \caption{\label{fig:rayleigh1}Configurations of a cylindrical liquid column at different times undergoing Rayleigh breakup and satellite droplet formation; $k^{*}=0.88$, $\rho_g=0.1$, $\rho_l=0.4$, $\nu_l=\nu_g=6.6667\times 10^{-2}$. Quantities are in lattice units.} \end{center} \end{figure} liquid column grow. At $t=28000, 46000$ $52000$, the cross-section of the column becomes progressively thinner in the center, and by mass conservation, the ends becomes larger. At $t=60000$, notice that a bead-type structure is formed at the ends and with a thin ligament between them. Such a structure has been observed in experiments~\cite{eggers97} and in other numerical simulations~\cite{ashgriz95}. Eventually, the column breaks up forming a thin ligament in the middle, which then becomes a satellite droplet. Let us now increase the wavelength of the disturbance to $\lambda_d=600$, keeping the physical parameters the same as before. We consider a domain represented by $601\times 151$ lattice sites. Since, $R_c=45$ as before, the wavenumber is $0.47$. Figure \ref{fig:rayleigh2} shows the temporal evolution of the configurations of the liquid column at this reduced wavenumber. The axisymmetric disturbance grows with time. \begin{figure} \begin{center} \includegraphics{fig8.EPS} \caption{\label{fig:rayleigh2}Configurations of a cylindrical liquid column at different times undergoing Rayleigh breakup and satellite droplet formation; $k^{*}=0.47$, $\rho_g=0.1$, $\rho_l=0.4$, $\nu_l=\nu_g=6.6667\times 10^{-2}$. Quantities are in lattice units.} \end{center} \end{figure} Since the wavelength is longer, it can be noticed that the ligament that is formed during the Rayleigh instability is also longer. As a result, after the column breaks up, a larger satellite droplet is formed. To express the drop size distribution with wave numbers more quantitatively, we plot the non-dimensional size of the main and satellite drops, $r^{*}=R/R_c$, as a function of wave number, $k^{*}$ in Fig. \ref{fig:rayleigh3}. It may be noted that Rayleigh's \begin{figure} \begin{center} \includegraphics{fig9.eps} \caption{\label{fig:rayleigh3}Drop sizes resulting from Rayleigh breakup of liquid cylindrical column as a function of wave number $k^{*}$. Quantities are dimensionless.} \end{center} \end{figure} original analysis predicts only the onset of breakup and not the formation of satellite droplets. To predict analytically satellite droplet formation, it has been shown that at least a third-order perturbation analysis of the Navier-Stokes equations (NSE) is needed~\cite{lafrance75}. Computations based on direct solutions of the NSE also predict the formation of the satellite droplets. To evaluate the drop size distribution computed using the axisymmetric LBE model, we consider the experimental data of Rutland and Jameson (1971),~\nocite{rutland71} the experimental data and analytical solution based on a third-order perturbation analysis of the NSE by Lafrance (1975),~\nocite{lafrance75} a boundary integral solution of the NSE by Mansour and Lundgren (1990)~\nocite{mansour90} and a finite element solution of the NSE by Ashgriz and Mashayek (1995).~\nocite{ashgriz95} It can be seen in the figure that as long as the wavenumber is less than one, as expected there will be a satellite droplet formation. As the wavenumber is reduced, the sizes of both the main drop and satellite droplet increases. The rate of increase of the size of the satellite droplet is greater than that of the main drop. Notice that there is considerable scatter in the available data in the figure. The computed results from the axisymmetric LBE model are presented for wavenumbers greater than or equal to $0.47$. Ignoring the two experimental data points of Lafrance (1975) for the satellite drop sizes that deviate considerably from the others, we find that the axisymmetric model is able to reproduce the drop size distribution quantitatively within $12\%$. The axisymmetric model has been employed to study head-on collisions of drops of radii $R_1$ and $R_2$ approaching each other with a relative velocity $U$. The dynamics and outcome of colliding drops is characterized mainly by the Weber number, $We$ defined by $We=\rho_{l} (R_{1}+R_{2}) U^{2} / \sigma$~\cite{qian97}. Additional parameters that may have an influence are the Ohnesorge number, $Oh$, defined by $Oh=16\mu_{l}/\sqrt{\rho_{l}R_{1}\sigma}$ and ratios of liquid and gas densities($r$) and dynamic viscosities ($\lambda$). According to experiments~\cite{qian97}, it is expected that lower $We$ collisions lead to coalescence while higher $We$ to separation by reflexive action. Figures \ref{fig:collWe20} and \ref{fig:collWe100} present drop configurations at $We=20$ and $We=100$ respectively. Notice that at $We=20$, the drops coalesce, while at $We=100$, they eventually separate \begin{figure} \begin{center} \includegraphics{fig10.EPS} \caption{\label{fig:collWe20} Colliding drops at different times $T$; $We=20$, $Oh=0.589$, $r=4$, $\lambda=1$. Time is normalized by the relative velocity between the drops and their diameter. Axes are in lattice units.} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics{fig11.EPS} \caption{\label{fig:collWe100} Colliding drops at different times $T$; $We=100$, $Oh=0.589$, $r=4$, $\lambda=1$. Time is normalized by the relative velocity between the drops and their diameter. Axes are in lattice units.} \end{center} \end{figure} with the formation of a satellite droplet, which are consistent with experimental observations. Also notice that for the latter case, the temporarily coalesced drop undergoes various stages of deformation which are consistent with a recent theoretical analysis~\cite{roisman04}. Additional details of these and other studies of drop collisions are given in Ref.~\cite{premnath04a}. \section{\label{sec:summary}Summary} In this paper, a LB model for axisymmetric multiphase flows is developed. The axisymmetric model is developed by adding source terms to the standard Cartesian BGK LBE. The source terms, which are temporally and spatially dependent, represent the axisymmetric contributions of the order parameter, which distinguish the different phases, as well as inertial, viscous and surface tension forces. Consistency of the model in achieving the desired axisymmetric flow multiphase behavior is established through the Chapman-Enskog multiscale analysis. The analysis shows that the axisymmetric macroscopic conservation equations are recovered in the continuum limit. An axisymmetric model with reduced compressibility effects is then developed to improve its computational stability. In this version, a transformation is introduced to the distribution function in the LBE such that it reduces the compressibility effects. Comparisons of computed axisymmetric equilibrium drop formation and oscillations, Rayleigh capillary instability, breakup and formation of satellite drops liquid cylindrical liquid columns and the outcomes of head-on drop collisions with available data show satisfactory agreement. The maximum error for the frequency of drop oscillations is less than $5.1\%$ and that for drop sizes as a result of Rayleigh breakup is $12\%$. \begin{acknowledgments} The authors thank Dr.\ X. He for helpful discussions and the Purdue University Computing Center (PUCC) and National Center for Supercomputing Applications (NCSA) for providing access to computing resources. \end{acknowledgments}
{ "timestamp": "2005-03-18T21:14:18", "yymm": "0503", "arxiv_id": "physics/0503160", "language": "en", "url": "https://arxiv.org/abs/physics/0503160" }
\section{Introduction} \subsection{Metric stability for random walks} In the study of a dynamical system, some of the most important questions concerns the stability of their dynamical properties under (most of the) perturbations: how robust are they? Here we are mainly interested in the stability of metric (measure-theoretical) properties of dynamical systems. A well-known example is given by ($C^2$) Markov expanding maps on the circle: this is a class stable under perturbations and all of them have an absolutely continuous and ergodic invariant probability satisfying certain decay of correlations estimatives. In particular, in the measure theoretical sense, most of the orbits are dense in the phase space. Now let us study a slightly more complicated situation: consider a $C^2$ Markov almost onto expanding map of the interval $f\colon I \rightarrow I$ with bounded distortion and large images (see Section 2 for details) and let $\psi \colon I \rightarrow \mathbb{Z}$ be a function which is constant in each interval of the Markov partition of $f$. We can define $F\colon I \times \mathbb{Z} \rightarrow I \times \mathbb{Z}$ as $$F(x,n):= (f(x), \psi(x) + n).$$ The second entry of $(x,n)$ will be called its {\bf state}. We also assume that \begin{equation}\label{cond1} \inf \psi > -\infty \end{equation} and that $F$ is topologically mixing. \begin{figure} \centering \psfrag{f}{$f$} \psfrag{Fx}[][][0.8]{$F$} \psfrag{Fy}[][][0.8]{$F$} \psfrag{x}[][][0.8]{$x$} \psfrag{y}[][][0.8]{$y$} \psfrag{fx}[][][0.8]{$f(x)$} \psfrag{fy}[][][0.8]{$f(y)$} \psfrag{a}[][][0.8]{$i$} \psfrag{b}[][][0.8]{$i+1$} \psfrag{c}[][][0.8]{$i+2$} \psfrag{d}[][][0.8]{$i-1$} \psfrag{e}[][][0.8]{$i-2$} \psfrag{psix}[][][0.8]{$\psi(x)=-1$} \psfrag{psiy}[][][0.8]{$\psi(y)=1$} \includegraphics[width=1.0\textwidth]{figure1.eps} \caption{A deterministic random walk} \end{figure} The map $F$ is refereed to in literature in many ways: as a "skew-product between $f$ and the translation on the group $\mathbb{Z}$", a "group extension of $f$", or even a "deterministic random walk generated by $f$", and its metric behavior is very well studied: for instance, are most the orbits recurrent? Everything depends on the {\bf mean drift } $$M = \int \psi d\mu,$$ where $\mu$ is the absolutely continuous invariant probability of $f$ (the function $\psi$ will be called {\bf drift function}). Indeed, note that $$F^n(x,i)= (\ f^n(x)\ ,\ i + \sum_{k=0}^{n-1}\psi(f^k(x))\ ).$$ By the Birkhoff Ergodic Theorem $$ \lim_{n \rightarrow \infty}\ \frac{\pi_2(F^n(x,i))- \pi_2(x,i)}{n} = \lim_{n \rightarrow \infty}\ \frac{1}{n} \sum_{k=0}^{n-1}\psi(f^k(x))= M.$$ for almost every $x \in I$ (here $\pi_2(x,n):=n$). In particular if $M \neq 0$ then almost every point $(x,i) \in I \times \mathbb{Z}$ is {\bf transient}: in other words we have $$\lim_{n\rightarrow \infty} |\pi_2(F^n(x,i))|= \infty.$$ So most of the points are not recurrent. On the other hand, if $M=0$, most of points are recurrent (see Guivarc'h \cite{guivarch}): by the Central Limit Theorem for expanding maps (here we need to assume that $\psi$ in no constant and $f \in aO$: see Section 2) of the interval $$sup_{\epsilon \in \mathbb{R}} \ | \mu( x \in I\colon \frac{\sum_{k=0}^{n-1}\psi(f^k(x))}{\sigma \sqrt{n}} \leq \ \epsilon) - \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\epsilon} e^{-\frac{u^2}{2}} \ du | \leq \frac{C}{\sqrt{n}},$$ Given $\delta >0$ we can easily obtain, taking $\epsilon = n^{-1/4}$ and applying Borel-Cantelli Lemma, that $$\mu(A_{+}):=\mu( x \in I \colon \ \limsup_{n\rightarrow \infty} \frac{\sum_{k=0}^{n-1}\psi(f^k(x))}{ \sqrt[2 + \delta]{n}}=\infty)\geq \frac{1}{2},$$ $$\mu(A_{-}):=\mu( x \in I \colon \ \liminf_{n\rightarrow \infty} \frac{\sum_{k=0}^{n-1}\psi(f^k(x))}{ \sqrt[2 + \delta]{n}}=-\infty)\geq \frac{1}{2}.$$ Clearly $A_{+}$ and $A_{-}$ are invariant sets: the ergodicity of $f$ implies that $$\mu(A_{+}~\cap~A_{-})~=~1.$$ By the conditions on $\psi$ in Eq. (\ref{cond1}), that $f$ is expanding with distortion control and that $F$ is transitive, we can easily conclude that almost every point in $I\times \mathbb{Z}$ is a $F$-recurrent point. Note that the random walk $F$ is a dynamical system quite similar to expanding circle maps: $F$ is an expanding map, with good bounded distortion properties; but the lack of compactness of the phase space allows the non-existence of an invatiant probability absolutely continuous with respect to the Lebesgue measure on $I\times \mathbb{Z}$. Moreover, in general the random walk is not even recurrent and the recurrence property lost its stability: given a recurrent random walk $(f,\psi)$, it is possible to obtain a transient random walk by changing a little bit $f$ and $\psi$. Since the non compactness of the phase space seems to be the origin of the lack of stability of recurrence and transience properties, a natural question is to ask if such properties are stable by compact perturbations. The answer is yes. Indeed, as we are going to see in Theorems \ref{sttr}-\ref{strec}, the transience and recurrence are preserved even by non-compact perturbations which decreases fast away from state $0$. For instance,we can choose perturbations like $$\tilde{F}(x,n)=(f_n(x),\psi(x)+n),$$ where, for some $\lambda \in [0, 1)$, \begin{equation}\label{decay}|f_n-f|_{C^3} \leq \lambda^{|n|}.\end{equation} The notations and conventions are more or less obvious: we postponed the rigorous definitions to the next section. With respect to the stability of transience and recurrence, there is a previous quite elegant result by R. L. Tweedie \cite{tw}: if $p_{ij}$ are the transition probabilities of a Markov chain on $\mathbb{Z}$, then any perturbation $\tilde{p}_{ij}$ so that $$ (1+\epsilon_i)^{-1}p_{ij} \leq \tilde{p}_{ij} \leq p_{ij} (1+\epsilon_i), \ j \neq i,$$ and $$\prod_{i=0}^{\infty} (1+\epsilon_i) < \infty$$ preserves the recurrence or transience of the original Markov chain. But Tweedie argument does not seem to work in our setting. Our result coincides with Tweedie result in the very special case where $f$ and $f_n$ are linear Markov maps and $\epsilon_i \sim C\lambda^{|i|}$. In the transient case we can tell a little more: there will be a conjugacy between the original random walk $f$ and its perturbation which is a martingale strongly quasisymmetric map (for short, mSQS-map) with respect to certain dynamically defined set of partitions. Unlike the usual class of one-dimensional quasisymmetric functions, which does not share many of most interesting properties of higher dimensional quasisymmetric maps, the one-dimensional mSQS-maps are much closer to their high-dimensional cousins, as quasiconformal maps in dimension $2$. For instance, they are absolutely continuous. We also study the behavior of the Hausdorff dimension of dynamically defined sets: Denote by $\Omega_+(F)$ the set of points which have non-negative states along the positive orbit by $F$. We prove that $\Omega_+(F)$ has Hausdorff dimension strictly smaller than one if and only if $\Omega_+(\tilde{F})$ has dimension less than one for all perturbation satisfying Eq. (\ref{decay}). Furthermore we give a variational characterization for the Hausdorff dimension $HD(\Omega_+(F))$ as the minimum of $HD(\Omega_+(\tilde{F}))$, where $\tilde{F}$ runs on the set of such perturbations. For these results we study of the stability of the multifractal spectrum of the random walk $F$ under those perturbations. \subsection{Applications to (generalized) renormalization theory}An unimodal map is a map with an unique critical point. Under reasonable conditions (real-analytic maps with negative Schwarzian derivative and non-flat critical point) two non renormalizable unimodal maps with the same topological entropy are indeed topologically conjugated. A key question in one-dimensional dynamics is about the regularity of the conjugacy: is it H\"older? Is it absolutely continuous? Since Dennis Sullivan work in the 80's the quasisymmetry of the conjugacy became a very useful tool to obtain deep results in one-dimensional dynamics. Lyubich proved that under the reasonable condition above the conjugacy between two non renormalizable unimodal maps is quasisymmetric. Later on, the density of the hyperbolic maps in the real quadratic family was proved verifying the quasisymmetry of the conjugacies for all combinatorics, including infinitely renormalizable ones. Note that quasisymmetric maps are not, in general, absolutely continuous: they do not even preserve (in general) sets of Hausdorff dimension one. Are the conjugacy between unimodal maps absolutely continuous? The answer is no: M.~Martens and W.~de~Melo \cite{mm} proved that under the reasonable conditions above an absolutely continuous conjugacy is actually $C^\infty$, provided the unimodal maps \begin{itemize} \item[ ] \ \item[{\it i.}] {\it do not have a periodic attractor,}\\ \item[{\it ii.}]{\it are not infinitely renormalizable, }\\ \item[{\it iii.}] {\it do not have a wild attractor (the topological and measure-theoretical attractor must coincide).} \\ \end{itemize} Since we can change the eigenvalues of the periodic points of maps preserving its topological class, and the eigenvalues are preserved by $C^1$ conjugacies, we conclude that in general a conjugacy between unimodal maps is not absolutely continuous. Condition i. is clearly necessary. This work (Theorem \ref{apl1}) shows that the Condition ii. is necessary proving that the conjugacy between two arbitrary Feigenbaum unimodal maps with same critical order is {\em always } absolutely continuous . Actually the conjugacy is martingale strongly quasisymmetric with respect to a set of dynamically defined partitions. Condition iii. is never violated when the critical point is quadratic. But for certain topological classes of unimodal maps wild attractor appears when the order of the critical point increases: Fibonacci maps are the simplest kind of such maps. We are going to prove (Theorem \ref{apl2}) that a Fibonacci map with even order has a wild attractor if and only if all Fibonacci maps with the same even order are conjugated to each other by an absolutely continuous mapping (in particular all these Fibonacci maps have a wild attractor). So Condition iii. is necessary. In both examples above the study about perturbations of transient and recurrent random walks are going to be crucial, as the (generalized) renormalization theory for unimodal maps: for these maps it is possible to construct an induced map which is essentially a perturbation of a deterministic random walk. In the Fibonacci case the transience of this random walk is equivalent to the existence of a wild attractor. The random walk associated to the Feigenbaum map will always be transient. For both Feigenbaum and Fibonacci maps there are infinitely many periodic points (indeed in the Fibonacci case the periodic points are also dense in the maximal invariant set). It is well known that the conjugacy between critical circle maps with same irrational rotation number and satisfying certain Diophantine condition is absolutely continuous, but we think that these are the first interesting examples of a similar phenomena for maps with many periodic points. \section{Expanding Markov maps, random walks and its perturbations} In this article we will deal with maps $$F\colon I \times \mathbb{Z} \rightarrow I \times \mathbb{Z}$$ which are piecewise $C^2$ diffeomorphisms, which means that there is a partition $\mathcal{P}^0$ of $I \times \mathbb{Z}$ so that each element $J \in \mathcal{P}^0$ is an open interval where $F|_{\overline{J}}$ is a $C^2$ diffeomorphism. Denote $I_n=I\times \{ n \}$. Denote by $m$ the Lebesgue measure in the in $I\times\mathbb{Z}$, that is, if $A\subset I\times\mathbb{Z}$ is a Borelian set then $$m(A)=\sum_n m_I(\pi(A\cap I_n)),$$ where $m_I$ is the Lebesgue measure in the interval $I$ and $\pi(n,x)=x$. If $A_J$ denotes the unique affine transformation which maps the interval $J$ to $[0,1]$ and preserves orientation, then define, for each $J \in \mathcal{P}^0$, $$\tau_J^F:= A_{J}\circ F^{-1} \circ A_{F(J)}^{-1}.$$ Throughout this article we will assume that $F$ satisfies some of the following properties: \begin{itemize} \item[\ ] \ \\ \item {\bf Markovian (Mk)}: For each $J \in \mathcal{P}^0$, $F(J)$ is a connected union of elements in $\mathcal{P}^0$. In particular we can write $F(x,n)=(f_n(x), n + \psi(x,n))$, where $f_n\colon I \rightarrow I$ is a piecewise $C^2$ diffeomorphism relative to the partition $\mathcal{P}^0_n:=\{J \in \mathcal{P}^0\colon \ J \subset I_n \}$ and $\psi\colon I \times \mathbb{Z} \rightarrow \mathbb{Z}$, called the {\bf drift function}, is constant on each element of $\mathcal{P}^0$.\\ \item {\bf Lower Bounded Drift (LBD)} $F$ is Markovian and $\min \psi > -\infty$.\\ \item {\bf Large Image (LI)}: $F$ is Markovian and there exists $\delta > 0$ so that for each $J \in \mathcal{P}^0$ we have $|F(J)|\geq \delta$.\\ \item {\bf Onto (On)}: $F$ is Markovian and for each $J \in \mathcal{P}^0$ we have $F(J)=I^n$, for some $n \in \mathbb{Z}$.\\ \item {\bf Bounded Distortion (BD)}: There exists $C > 0$ so that every $J \in \mathcal{P}^0_n$ and map $\tau_J$ is a $C^2$ function satisfying $$\sup_{J} \Big| \frac{D^2\tau_J}{(D\tau_J)^2} \Big| \leq C.$$ \\ \item {\bf Strong Bounded Distortion (sBD)}: There exists $C > 0$ so that every $J \in \mathcal{P}^0_n$ and map $\tau_J$ is a $C^2$ function satisfying $$\sup_{J} \Big| \frac{D^2\tau_J}{(D\tau_J)^2} \Big| \leq C|J|.$$ \\ \item {\bf Expansivity (Ex)}: If $J \in \mathcal{P}^0_n:=\{ J \in \mathcal{P}^0\colon \ J \subset I_n\}$, denote $\phi_J:=f_n^{-1}|_{f_n(J)}$. Then either $\phi_J$ can be extended to a function in a $\delta$-neighborhood of $J$ so that $$ S\phi_J > 0,$$ where $S\phi_J$ denotes the Schwarzian derivative of $\phi_J$, or there exists $\theta \in (0,1)$ so that $$|\phi'_J| < \theta$$ on $I$. \\ \item {\bf Regularity a (Ra)}: There exists $N \in \mathbb{N}$, $\delta > 0$ and $C > 0$ with the following properties: the intervals in $\mathcal{P}^0_n$ are positioned in $I_n$ in such way that the complement of $$\bigcup_{J \in \mathcal{P}^0_n}\ int \ J$$ contains at most $N$ accumulation points $$c_1^n < c_2^n < \dots < c^{n}_{i_n},$$ with $i_n\leq N$, which is in the interior of $I_n$. Furthermore $|c^n_{i+1}-c^n_i|\geq \delta$. Moreover, given $P$ and $Q \in \mathcal{P}^0_n$ so that $\overline{P} \cap \overline{Q} \neq \phi$ then $$\frac{1}{C} \leq \frac{|P|}{|Q|} \leq C.$$ \item {\bf Regularity b (Rb)}: Assume $Ra$. There exists $C > 0$, $\lambda \in (0,1)$, $\delta > 0$ so that for each $1< i<i_n$ we can find a point $$d_i^n \in (c^n_i,c^n_{i+1}),$$ which does not belong to any $P \in \mathcal{P}^0_n$, and $$\min \{|c^n_{i+1}-d_i^n|, |d_i^n- c^n_{i}| \} \geq \delta$$ with the following property: If $J$ is a connected component of $$I_n \setminus \{d_i^n, c^n_j\}_{i,j}$$ then we can enumerate the set $$\{P\}_{P \in \mathcal{P}^0_n,\ P \subset J}=\{J_i\}_{i \in \mathbb{N}}$$ in such way that $\partial J_{i} \cap \partial J_{i+1} \neq \phi$ for each $i$ and $$\frac{|J_{i+j}|}{|J_{i}|}\leq C\lambda^j$$ for $i \geq 0$, $j > 0$.\\ \item {\bf Good Drift (GD)}: , if $\psi$ is the drift function of the random walk then there exists $\gamma \in (0,1)$ and $C > 0$ so that $$m(\{(x,n) \ s.t. \ \psi(x,n)\geq k \}) \leq C\gamma^k.$$ \\ \item {\bf Transitive (T)}: $F$ has a dense orbit. \end{itemize} For convenience of the notation if for instance $F$ is Markovian and it has Bounded Distortion, we will write $F \in Mk+BD$. A {\bf deterministic random walk} (or simply random walk) is a map $$F \in Mk+LBD+LI+Ex+BD+GD.$$ It is generated by the pair $(\{ f_n\},\psi)$ if $$F(x,n):= (f_n(x),\psi(x,n)+n).$$ When $f_n=f \in Mk$ and $\psi(x,n)=\psi(x)$, we say that $F$ is the {\bf spatially homogeneous deterministic random walk} generated by the pair $(f,\psi)$. There is a large literature about such random walks. We will sometimes assume the following property: \begin{itemize} \item[\ ] \ \\ \item {\bf Almost Onto (aO)}: For every $i, j \in \Lambda$ there exists a finite sequence $i=i_0,i_1,i_2,\dots,i_{n-1},i_n=j \in \Lambda$ so that $$f(I_{i_k})\cap f(I_{i_{k+1}})\neq \emptyset$$ for each $k < j$. \\ \end{itemize} Denote $\pi(x,n):= \pi_2(x,n):=n$. A random walk is called {\bf transient} if for almost every $(x,n) \in I \times \mathbb{Z}$ $$\lim_{k\rightarrow \infty} |\pi_2(F^k(x,n))| = \infty,$$ and it is {\bf recurrent} if for almost every $(x,n) \in I \times \mathbb{Z}$ $$\# \{ k \colon \pi_2(F^k(x,n))=n \} = \infty.$$ Making use of usual bounded distortion tricks it is easy to show that every $F \in Mk+LI+Ex+BD+T$ is either recurrent or transient. A (topological) {\bf perturbation} of a random walk is a random walk $\tilde{F}$, generated by a pair $(\{\tilde{f}_i\},\tilde{\psi})$, so that $F\circ H = H \circ \tilde{F}$ for some homeomorphism $$H\colon I \times \mathbb{Z} \rightarrow I \times \mathbb{Z}$$ which preserves states: $\pi_2(H(x,i))=i$. Define $\mathcal{P}^n(F) := \vee_{i=0}^{n-1} F^{-i}\mathcal{P}^0(F)$. If $F$ and $\tilde{F}$ are random walks and $h$ is a topological conjugacy that preserves states between $F$ and $\tilde{F}$, then for each interval $L$ such that $L\subset J \in \mathcal{P}^n(F)$, define $$dist_n(L):= \sup_{x \in L} \Big|\ln \frac{DF^n(x)}{DF^n(y)} \Big|,$$ Similarly, if $x \in J \in \mathcal{P}^n(F)$ define $$dist_n(x):= dist_n(J)$$ and $$dist_\infty(x):= \sup_n dist_n(x).$$ Another kind of random walk which will have a central role in our results are those which are {\bf asymptotically small} perturbations: these are perturbations $(\{\tilde{f}_i\},\tilde{\psi})$ of a homogeneous random walk $(\{f_i\},\psi)$ such that there exists $\lambda \in (0,1)$ and $C >0$ satisfying either \begin{equation}\label{asymp} |\log \frac{DF(H(p))}{D\tilde{F}(p)}| \leq C\lambda^{|\pi_2(p)|},\end{equation} if $\psi$ is bounded, or \begin{equation}\label{asymp2} |\log \frac{DF(H(p))}{D\tilde{F}(p)}| \leq C\lambda^{\pi_2(p)},\end{equation} for $\pi_2(p)\geq 0$ and $DF(H(p))=D\tilde{F}(p)$ otherwise, if $\psi$ has only a lower bound. It is easy to see that properties $Ra$, $Rb$ and $GD$ are invariant by asymptotically small perturbations (if we allow to change the constants described in these properties). Let $F=(\{f_i\}_i,\psi)$ be a random walk, where $\psi$ is Lebesgue integrable on compact subsets of $I \times \mathbb{Z}$. We say that $F$ is {\bf strongly transient} if $K > 0$ and $$\mathbb{E}(\psi\circ F^n | \mathcal{P}^{n-1}(F)) > K$$ for every $n\geq 1$. We will also say that $F$ is $K$-strongly transient. Here we are considering conditional expectations relative to the Lebesgue measure. As the notation suggest, every strongly transient random walk is transient. Moreover we have the following large deviations result: \begin{prop}\label{largedeviationsst}\label{bru} Every $K$-strongly transient random walk $F\in Ra+Rb$ is transient. Furthermore for every small $\epsilon > 0$ there exist $\lambda \in [0,1)$ and $C > 0$ so that for each $P \in \mathcal{P}^0$ we have $$m( p \in P \colon \ \pi_2(F^n(p))-\pi_2(p) < (K - \epsilon) n )\leq C\lambda^n |P|.$$ \end{prop} We will postpone the proof of this result to Section \ref{sectiontransience}. \begin{rem}{\rm By the Birkhoff Ergodic Theorem it is easy to see that a sufficiently high iteration of a homogeneous random walk with positive mean drift is strongly transient (see the proof of Proposition \ref{homstr} for details). } \end{rem} \section{Statements of results} \subsection{Stability of transience} \begin{thm}[Stability of Transience I]\label{sttr} Assume that the random walk $F$ defined by the pair $(\{f_i\}_i,\psi)$ is strongly transient. Then every asymptotically small perturbation $G$ of $F$ is also transient. Indeed there is a topological conjugacy between $F$ and $G$ which is an absolutely continuous map and preserves the states. \end{thm} We have a similar theorem for all transient homogeneous random walks: \begin{thm}[Stability of Transience II]\label{sttrho}\label{abscont} Suppose that the homogeneous random walk $F$ defined by the pair $(f,\psi)$ is transient. Then every asymptotically small perturbation of $F$ is topologically conjugated to $F$ by an absolutely continuous map which preserves the states. \end{thm} We can be more precise regarding the regularity of the conjugacy if the drift is non-negative: Let $\mathcal{A}_0,\ \mathcal{A}_1, \ \cdots , \mathcal{A}_n, \ \mathcal{A}_{n+1}, \cdots $ be a succession of partitions by intervals of $I \times \mathbb{Z}$, such that $\mathcal{A}_{n+1}$ refines $\mathcal{A}_{n}$ and whose union generates the Borelian algebra of $\sqcup_n I_n$. We say that $h\colon \sqcup_n I_n \rightarrow \sqcup_n I_n$ is a {\bf martingale strongly quasisymmetric (mSQS)} map with respect to the {\bf stochastic basis } $\cup_n \mathcal{A}_n$ if there exist $C > 0$ and $\alpha \in (0,1]$ so that $$ \frac{m(h(B))}{|h(J)|} \leq C \left( \frac{m(B)}{|J|} \right )^\alpha$$ for all Borelian $B \subset J \in \cup_n \mathcal{A}_n$, and the same inequality holds replacing $h$ by $h^{-1}$ and $\cup_n \mathcal{A}_n$ by $\cup_n h(\mathcal{A}_n)$. \begin{thm}[Strongly quasisymmetric rigidity]\label{sqr} Let $F$ be either a strongly transient random walk or a transient homogeneous random walk with positive mean drift. Moreover assume in both cases that $\psi\geq 0$. Then every asymptotically small perturbation $G$ of $F$ is topologically conjugated to $F$ by an absolutely continuous map $h$ which preserves the states. Furthermore $h$ on $\cup_{i\geq 0} I_i$ is a martingale strongly quasisymmetric mapping with respect to the stochastic basis $\cup_i \mathcal{P}^i.$ \end{thm} \subsection{Stability of recurrence} In the recurrent case, we are going to restrict ourselves to the stability of the metric properties of homogeneous random walks under asymptotically small perturbations: it is easy to see that the recurrence is not stable by perturbations which are not asymptotically small. Nevertheless \begin{thm}[Stability of Recurrence]\label{strec} Suppose that $F \in aO+T$ is a recurrent homogeneous random walk generated by the pair $(f,\psi)$. Then every asymptotically small perturbation of $F$ is also recurrent. \end{thm} If $p$ is a periodic point with prime period $n$ then $DF^n(p)$ is called the spectrum of the periodic point $p$. Note that we can not expect, as in the transient case, an absolutely continuous conjugacy which preserves states between $F$ and $G$, once asymptotic small perturbations do not preserve (in general) the spectrum of the periodic points and: \begin{prop}[Rigidity]\label{rigidity} Suppose that the random walk $F \in On$ generated by a pair $(\{f_i \}_i,\psi)$ is recurrent. If there is an absolutely continuous conjugacy which preserves states $H$ between $F$ and a random walk $G$, then $H$ is $C^1$ in each state. In particular the spectrum of the corresponding periodic points of $F$ and $G$ are the same. \end{prop} The reader should compare this result with similar results by Shub and Sullivan \cite{ss} for expanding maps on the circle and de Melo and Martens \cite{mm} for unimodal maps. \subsection{Stability of the multifractal spectrum} Let $F$ be a random walk and denote $$\Omega_+(F):=\{p \colon \ \pi_2(F^jp)\geq 0, \ for \ j\geq 0 \},$$ $$\Omega_+^k(F):= \{(x,k) \colon \ \pi_2(F^j(x,k))\geq 0, \ for \ j\geq 0 \}$$ and $$\Omega_{+\beta}^k(F):= \{ (x,k) \in \Omega_+^k\ s.t\ \ \underline{\lim}_{\ n} \ \frac{\pi_2(F^n(x,k))}{n}\geq\beta\}$$ \begin{thm}\label{multi} Let $F\in Ra + Rb+ On$ be a random walk. Then, for all $k \in \mathbb{Z}$ and $\beta > 0$ the Hausdorff dimension $HD(\Omega_{+\beta}^k)$ is invariant by asymptotically small perturbations. \end{thm} Besides its inner interest, the previous result will be useful by other reason: \begin{prop}\label{ms}\label{inv} Let $F \in Ra+Rb+On$ be a homogeneous random walk. Then $$HD(\Omega_+^k(F))= \lim_{\beta\rightarrow 0^+} HD(\Omega_{+\beta}^k(F)).$$ \end{prop} and as a consequence of Theorem \ref{multi} and Proposition \ref{ms}: \begin{thm}\label{omega} Let $F\in Ra+Rb+On$ be a homogeneous random walk. If $G$ is an asymptotically small perturbation of $F$ then \begin{equation}\label{ineq} HD(\Omega_{+}^k(G)) \geq HD(\Omega_{+}^k(F)).\end{equation} \end{thm} We can not replace the inequality in Eq. (\ref{ineq}) by an equality. Indeed, even if $HD(\Omega_{+}^k (F)) < 1$, we have that $sup \ HD(\Omega_{+}^k(G)) =1$, where the supremum is taken on all asymptotically small perturbations $G$ of $F$. Nevertheless: \begin{thm}\label{menor} Let $F \in Ra+Rb+On$ be the homogeneous random walk generated by the pair $(f,\psi)$. Consider $M=\int \psi d\mu$, where $\mu$ is the unique absolutely continuous invariant measure of $f$. \begin{itemize} \item[-] If $M >0$ then for all asymptotically small perturbations $G$ of $F$ we have $m (\Omega_{+}(G))> 0$.\\ \item[-] If $M =0$ then for all asymptotically small perturbations $G$ of $F$ we have $HD (\Omega_{+}(G))=1$ but $m (\Omega_{+}(G))= 0$.\\ \item[-] If $M < 0$ then for all asymptotically small perturbations $G$ of $F$ we have $HD (\Omega_{+}(G))< 1$. \end{itemize} \end{thm} \begin{rem}{\rm Since the authors are more familiar with deterministic rather than stochastic terminology, we stated and proved the results in this work for determinist random walks. However we believe that the above results could be easily translated to the theory of chains with complete connections (g-measures, chains of infinite order) and one-side shifts on an infinite alphabet. } \end{rem} \subsection{Applications to renormalization theory of one-dimensional maps} \begin{thm}\label{apl1} Let $f$ and $g$ be unimodal maps which are infinitely renormalizable with the same bounded combinatorial type and even critical order. Then the continuous conjugacy $h$ between $f$ and $g$ is a strongly quasisymmetric mapping with respect to a certain stochastic basis of intervals $\mathcal{P}$.\end{thm} The set of intervals $\mathcal{P}$ is defined using a map induced by $f$. See the details in Section \ref{apl}. Let $\mathcal{F}_d$ be the class of analytic maps with schwarzian negative derivative which are infinitely renormalizable in the Fibonacci sense with even critical order $d$ (see Section \ref{aplf} for definitions). If $f$ is a Fibonacci map, denote by $J_{\mathbb{R}}(f)$ the maximal invariant set of $f$. Let $\mathcal{F}_d^{uni}$ be the class of Fibonacci {\it unimodal} maps with negative Schwarzian derivative. \begin{thm}[Metric Universality]\label{juliathm} For each even critical order $d$, one of the following statements holds: \begin{itemize} \item $HD(J_{\mathbb{R}}(f)) < 1$, for all $f \in \mathcal{F}_d$. \item $HD(J_{\mathbb{R}}(f))= 1$ and $m(J_{\mathbb{R}})=0$ for all $f \in \mathcal{F}_d$. \item $HD(J_{\mathbb{R}}(f))= 1$ and $f$ has a wild attractor (in particular, $m(J_{\mathbb{R}}(f))> 0$) for all $f \in \mathcal{F}_d$ \end{itemize} \end{thm} \begin{thm}[Measurable Deep Point]\label{deep} Let $f \in \mathcal{F}_d$, and assume that $0$ is its critical point. If $J_{\mathbb{R}}(f)$ has positive Lebesgue measure then there exists $\alpha > 0$ and $C > 0$ so that $$m(x \in (-\delta,\delta)\colon \ x \not\in J_{\mathbb{R}}(f))\leq C\delta^{1+\alpha}.$$ \end{thm} \begin{rem} {\rm Indeed $\alpha$ can be taken depending only on $d$. }\end{rem} \begin{thm}\label{apl2}For each even critical order $d$, the following statements are equivalent: \begin{enumerate} \item There exists $f \in \mathcal{F}_d$ such that $m(J_{\mathbb{R}}(F)) > 0$. \item There exists $f \in \mathcal{F}_d$ with a wild attractor. \item There exist maps $f,g \in \mathcal{F}_d^{uni}$ which are conjugated by a continuous absolutely continuous maps $h$, but $f$ has a periodic point $p$ whose eigenvalue is different from the eigenvalue of the periodic point $h(p)$ of $g$. \item All maps in $\mathcal{F}_d$ have wild attractors. \item All maps in $\mathcal{F}_d^{uni}$ can be conjugated with each other by an absolutely continuous conjugacy. \end{enumerate} \end{thm} \section{Preliminaries} \subsection{Probabilistic tools.} We are going to collect here a handful of probabilistic tools which are going to be useful along the article. A good reference for these results is \cite{broise}. Most of the probabilistic results in dynamical systems (large deviation, central limit theorem) assumes the observable $\psi$ is quite regular: usual regularity assumptions are either Holder continuity or bounded variation. Fix $f \in Mk+BD$. We are interested in $\mathcal{P}^0$-measurable observables with integer values which do not have such properties. Fortunally this is almost true: Denote by $ \mathcal{O}(f)$ the class of $\mathcal{P}^0$-measurable functions $\psi\colon I \rightarrow \mathbb{Z}$ so that \begin{itemize} \item[ ] \ \item[-] $\psi \in L^2(\mu)$,\\ \item[-] If $P$ denotes the Perron-Frobenius-Ruelle operator of $f$, then $P\psi$ has bounded variation.\\ \end{itemize} For instance, if $(f,\psi) \in Mk+ sBD+Ra+Rb+GD$ then $\psi \in \mathcal{O}(f)$. Let $\mu$ be the absolutely continuous invariant measure of a Markov map $f$ and let $\psi\colon I \rightarrow \mathbb{R}$ be a measurable function. \begin{prop}[Large Deviations Theorem \cite{broise}]\label{ldt} For every $\psi \in \mathcal{O}(f)$ and $\epsilon > 0$ there exists $\gamma \in (0,1)$ so that $$\mu(\{x \in I \colon |\frac{1}{n} \sum_{i=0}^{n-1}\psi(f^i(x)) - \int \psi d\mu|\geq \epsilon \}) \leq \gamma^n$$ \end{prop} Up to simple modifications in the proofs in \cite{broise}, we have \begin{prop}[Proposition 6.1 of \cite{broise}] For every $\psi \in \mathcal{O}(f)$ the limit $$\sigma^2 := \lim_{n\rightarrow \infty} \int \left( \frac{1}{\sqrt{n}} \sum_{k=0}^{n-1}\psi(f^k(x)) \right) ^2 d\mu$$ exists. Furthermore $\sigma^2=0$ if and only if there exists a function $\alpha \in L^2(\mu)$ so that $$\psi = \alpha\circ f - \alpha. $$ \end{prop} and \begin{prop}[Central Limit Theorem: Theorem 8.1 in \cite{broise}]\label{clt}For every $\psi \in \mathcal{O}(f)$ so that $\sigma^2\neq 0$ we have \begin{equation}\label{cltt} sup_{\epsilon \in \mathbb{R}} \ | \mu( x \in I\colon \frac{\sum_{k=0}^{n-1}\psi(f^k(x))}{\sigma \sqrt{n}} \leq \ \epsilon) - \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\epsilon} e^{-\frac{u^2}{2}} \ du | \leq \frac{C}{\sqrt{n}},\end{equation} \end{prop} Indeed we are going to see that the assumption $\sigma^2 \neq 0$ is very weak: to this end we need the following result: \begin{prop}[Theorem 3.1 in \cite{ad}]\label{cocycle} Let $f\colon \cup_i I_i \rightarrow I$ be a map in Mk + BD + Ex + Ra + Rb. Let $\psi \colon \cup_i I_i \rightarrow \mathbb{S}^1$ be a $\mathcal{P}_0$-measurable function. If $$\psi = \frac{\alpha\circ f}{\alpha},$$ where $\alpha$ is measurable, then $\alpha$ is $\mathcal{P}^\star$-measurable, where $\mathcal{P}^\star$ is the finest partition of $I$ so that $f(I_i)$ is included in an atom of $\mathcal{P}^\star$ for each $i \in \Lambda$. \end{prop} \begin{prop}\label{cco} Let $\psi \colon \cup_i I_i \rightarrow \mathbb{Z}$ be a $\mathcal{P}^0$-measurable function. If $\psi = \alpha\circ f - \alpha$, where $\alpha$ is measurable, then $\alpha$ is constant on $f(I_i)$, for each $i \in \Lambda$. \end{prop} \begin{proof} Note that we can assume that $\alpha(x) \in \mathbb{Z}$, for every $x$. Indeed, the relation $\psi = \alpha\circ f - \alpha$ implies that the function $\beta(x)=\alpha(x) \mod 1$ is $f$-invariant, so we can replace $\alpha$ by $\alpha -\beta$, if necessary. Fix an irrational number $\gamma$. Then $$e^{2\pi \gamma \psi(x)i} = \frac{e^{2\pi \gamma \alpha(f(x))i}}{e^{2\pi \gamma \alpha(x)i}},$$ so by Proposition \ref{cocycle} we have that $e^{2\pi \gamma \alpha(x)i}$ is a $\mathcal{P}^\star$-measurable function. Since $j \in \mathbb{Z} \rightarrow e^{2\pi \gamma j i} \in \mathbb{S}^1$ is one-to-one, we get that $\alpha$ is $\mathcal{P}^\star$-measurable. \end{proof} A Markov map $f$ is almost onto if and only if $\mathcal{P}_0^\star= \{ I\}$, so \begin{cor}On the conditions of Proposition \ref{cco}, if $f$ is almost onto then $\alpha$ is constant.\end{cor} \begin{cor} For every nonconstant $\psi \in \mathcal{O}(f)$ we have that $\sigma^2 \neq 0$. In particular the Central Limit Theorem as given in Eq. (\ref{cltt}) holds for every non-constant $\psi$.\end{cor} Let $\mathcal{A}_0 \subset \mathcal{A}_1 \subset \mathcal{A}_2 \subset \dots $ be an increasing sequence of $\sigma$-subalgebras of a probability space $(\Omega,\mathcal{A},\mu)$. A {\bf martingale difference sequence } is a sequence of functions $\psi_n\colon \Omega \rightarrow \mathbb{R}$, where $\psi_n$ is $\mathcal{A}_n$-measurable for $n\geq 1$, so that $$\mathbb{E}(\psi_n | \mathcal{A}_{n-1}) =0$$ for every $n$. Here $\mathbb{E}(\psi | \mathcal{B})$ denotes de conditional expectation of $\psi$ relative to the sub-algebra $\mathcal{B}$. When $\mathcal{B}$ is generated by atoms $\{J_i \}_i$ then $\mathbb{E}(\psi | \mathcal{B})$ is the function defined as $$\mathbb{E}(\psi | \mathcal{B})(x)= \frac{1}{\mu(J_i)} \int_{J_i} \psi \ d\mu$$ for every $x \in J_i$. The following Proposition is the classic Azuma-Hoeffding inequality: see, for instance Exercise E14.2 in \cite{williams}: \begin{prop}[Azuma-Hoeffding inequality]\label{azuma} Let $\psi_n$ as above and furthermore assume that $$|| \psi_i ||_{\infty} = c_i < \infty.$$ Define $$\psi:= \sum_{i=1}^{n} \psi_i.$$ Then $$\mu(x \in \Omega\colon \ |\psi - \mathbb{E}(\psi)| > t ) \leq 2\exp ({-\frac{t^2}{2\sum_{i=1}^{n} c_i^2}}).$$ \end{prop} \subsection{How to construct asymptotically small perturbations.} As we will see in the next Proposition, it is easy to construct asymptotically small perturbations of a random walk: \begin{prop}\label{identity} \label{how}Let $F$ and $G$ be random walks satisfying the properties $LI$, $Ex$, $sBD$, $Ra$ and $Rb$, where $G$ is a topological perturbation of $F$. Assume that there exist $C >0$ and $\lambda \in (0,1)$ with the following properties: if $I^n_j$ is as in properties $Ra$ and $Rb$, then \begin{itemize} \item[ ] \ \item[i.] For every $I^n_j \in \mathcal{P}^0_n$ we have $$| log \frac{|I^n_{j+1}|}{|I^n_j|}\frac{|H(I^n_j)|}{|H(I^n_{j+1})|}| \leq C\lambda^{|n|+|j|}.$$\\ \item[ii.] For every $J \in \mathcal{P}^0_n$ we have $$ |\tau_{J}^F - \tau_{H(J)}^G|_{C^2} \leq C\lambda^{|n|}.$$\\ \item[iii.] If $I_i^n=[a_i^n,b_i^n]$ then $$\max_i \max \{|a_i^n - H(a_i^n)|, |b_i^n - H(b_i^n)| \} \leq C\lambda^{|n|}.$$\\ \item[iv.] Either $\psi$ is a bounded funtion or $\psi$ has a lower bound and $F = G$ on $\cup_{n<0}I_n$. \end{itemize} Then $G$ is an asymptotically small perturbation of $F$. Furthermore there exist $\beta \in [0,1)$ and $C > 0$ so that $$|H(p)-p|\leq C\beta^{|\pi_2(p)|}.$$ \end{prop} \begin{proof} We will assume that $\psi$ is bounded: the other case is analogous. Consider $(x,n) \in I\times \mathbb{Z}$ and $(y,n)=H(x,n)$. Denote $(x_i,n_i):=F^i(x,n)$, $(y_i,n_i):=G^i(y,n)$. Denote $\delta_i = |y_i-x_i|$ and $\tilde{\delta}_i=|A_{G(H(J_i))}(y_i)-A_{F(J_i)}(x_i)|$. Here $(x_i,n_i)~\in~J_i \in~\mathcal{P}^0$. It is easy to conclude, using iii. and property $LI$, that \begin{equation}\label{til} \tilde{\delta}_i \leq \frac{\delta_i}{|F(J)|} + C\lambda^{|n_i|}\end{equation} and making use of ii. to get $$|\tau_{H(J)}^G(A_{G(H(J_i))}(y_i))-\tau_J^F(A_{F(J_i)}(x_i))| \leq D\tau^F_J(z_i) \frac{\delta_i}{|F(J)|} +C\lambda^{|n_i|}. $$ Here $z_i \in [0,1]$. Since $D\tau^F_J(z_i)|F(J)|/|J|\leq \lambda$ (property $Ex$), we get, using again $iii.$ \begin{equation} \label{rec} \delta_{i-1} \leq \lambda \delta_i + C\lambda^{|n_i|}.\end{equation} Because $\psi$ is bounded, $|n_{i+1}-n_i|\leq B=\max |\psi|$. So if $i< n/2B$ then $|n_i| > |n_0|/2$. Since $\delta_{[\frac{n}{2B}]} \leq 1$, Eq. (\ref{rec}) implies $$|H(x,n)-(x,n)|=|y_0-x_0|\leq C\lambda^{\frac{|n|}{2}}.$$ In particular, by Eq. (\ref{til}) and property ii., we have \begin{equation} \label{normalizado} |D\tau_{H(J)}^G(A_{G(H(J_0))}(y_1))-D\tau_J^F(A_{F(J_0)}(x_1))|\leq C\lambda^{\frac{|n|}{2}}.\end{equation} By $Ra+Rb$ there exists $\theta \in (0,1)$ so that \begin{equation}\label{lowerb} \theta^{|i|}\leq |I^n_i|.\end{equation} Let $i$ so that $J=I^n_i$. {\it Case A.} $|i|\geq |n/2|(\log \lambda/\log \theta)$: Due i. and iii. and property $Ra$, there exists $C > 0$ so that $$ |\log \frac{|H(I^n_i)|}{|I^n_i|}| \leq C \lambda^n.$$ Together with $sBD+LI$ and $iii.$, this implies that for every $p \in I^n_i$, with $|i|\geq |n/2|(\log \lambda/\log \theta)$, we have $$|\log \frac{DG(H(p))}{DF(p)}| \leq C\lambda^{\frac{|n|}{2}\frac{\log \lambda}{\log \theta}}.$$ {\it Case B.} $|i| < |n/2|(\log \lambda/\log \theta)$: In this case, by iii. and Eq. (\ref{lowerb}) we have $$\log \frac{|H(I^n_i)|}{|I^n_i|} \leq C\frac{|H(b^n_i)-b^n_i| + |H(a^n_i)-a^n_i|}{|b^n_i - a^n_i|} \leq C\lambda^{\frac{|n|}{2}}. $$ Now using Eq. (\ref{normalizado}) we can easilly obtain $$|\log \frac{DG(H(p))}{DF(p)}|\leq C\lambda^{\frac{|n|}{2}}.$$ \end{proof} \section{Stability of transience}\label{sectiontransience} We will begin this section with the large deviations result to strongly transient random walks: \begin{proof}[\bf Proof of Proposition \ref{largedeviationsst}] Fix $\epsilon > 0$ small. We intend to apply the Azuma-Hoeffding inequality, but since $\psi$ is not necessarily bounded, we need to make some adjustments first: Fix $P \in \mathcal{P}^0(F)$ and define $\mathcal{F}_0:=\{P\}$ and $\mathcal{F}_n:=\{Q\}_{Q\subset P,\ Q \in \mathcal{P}^n(F)}$. Since $F\in GD$, by the usual distortion control tricks for $F$, we can find $M > \min \psi$ such that $\alpha(x):= \min \{\psi(x),M \}$ satisfies \begin{equation}\label{perturbacao} \mathbb{E}(\alpha\circ F^n | \mathcal{F}_{n-1}) \geq K-\epsilon/4 \end{equation} for every $n\geq 1$. Here we are considering conditional expectations relative to the probability $$\mu_P(A):= \frac{m(A)}{|P|},$$ where $m$ is the Lebesgue measure. Define the martingale difference sequence $$\Psi_n:= \alpha\circ F^n - \mathbb{E}(\alpha\circ F^n | \mathcal{F}_{n-1}).$$ Of course $||\Psi_n||_\infty \leq M$, if $M$ is large enough. By the Azuma-Hoeffding inequality we have $$m( p \in P \colon \ |\sum_{i=1}^{n} \Psi_i(p) | > t ) \leq 2 \exp( -\frac{t^2}{2nM^2})|P|.$$ Taking $t=\epsilon n/4$ we obtain \begin{equation}\label{dcorr} m( p \in P \colon \ |\sum_{i=1}^{n} \Psi_i(p) | > \frac{\epsilon}{4} \ n ) \leq 2 \exp( -\frac{\epsilon^2 n}{32M^2})|P|.\end{equation} Since $$\pi_2(F^{n+1}p)-\pi_2(F(p))= \sum_{i=1}^n \psi(F^i(p)) \geq \sum_{i=1}^{n} \alpha(F^i(p)) = \sum_{i=1}^n \Psi_i(p) + \sum_{i=1}^n \mathbb{E}(\alpha\circ F^i|\mathcal{F}_{i-1})(x)$$ $$\geq \sum_{i=1}^n \Psi_i(p) + (K-\epsilon/4) n.$$ Due Eq. (\ref{dcorr}), this implies that $$m( p \in P \colon \ \pi_2(F^{n}p)-\pi_2(F(p))=\sum_{i=1}^{n-1} \psi(F^i(p)) < (K -\epsilon/2 ) \ (n -1)) \leq C_1 \exp( -\frac{\epsilon^2 n}{32M^2}) |P|.$$ Let $n_0$ be such that $\min \psi > -\epsilon (n_0-1)/2 -\epsilon + K$. Then for $n\geq n_0$ we have that $$\pi_2(F^{n}p)-\pi_2(p) < (K-\epsilon) n$$ implies $$\pi_2(F^{n}p)-\pi_2(F(p)) < (K -\epsilon/2 ) \ (n -1).$$ So $$m( p \in P \colon \ \pi_2(F^{n}p)-\pi_2(p) < (K -\epsilon ) \ n) \leq C_2 \exp( -\frac{\epsilon^2 n}{32M^2}) |P|$$ for every $n$. This completes the proof. \end{proof} \begin{prop}\label{prws}\label{homstr} Let $F$ be either strongly transient or a homogeneous random walk with positive mean drift. Then any asymptotically small perturbation $G$ of $F$ has the following property: there exists $\lambda \in [0,1)$, $C > 0$ and $\tilde{K} > 0$ so that for every $P \in \mathcal{P}^0(G)$ $$m(p \in P \colon \ \sum_{i=0}^{n-1}\psi(G^i(p)) < \tilde{K} n )\leq C\lambda^n |P|.$$ In particular $G$ is also transient. \end{prop} \begin{proof} We will carry out the proof assuming the strongly transience: the homogeneous case is analogous: Fix $\epsilon > 0$. Let $\tilde{\delta}_1 > 0$ be small enough such that $$(1-\tilde{\delta_1})(K-\epsilon) + \tilde{\delta_1} \min \psi> K-2\epsilon.$$ Due the bounded distortion of $G$, there exists $\delta_1 > 0$ such that for every $n\geq 1$ and every $P \in \mathcal{P}^{n-1}(G)$, interval $Q \subset G^n(P)$, and set $A\subset Q$ satisfying $$\frac{m(A)}{m(Q)} \geq 1 -\delta_1$$ we have \begin{equation}\label{distg}\frac{m(P\cap G^{-n}A)}{m(P\cap G^{-n}Q)} \geq 1 -\tilde{\delta}_1.\end{equation} By Proposition \ref{largedeviationsst} we have \begin{equation}\label{unpum} m( p \in P \colon \ \sum_{i=0}^{n-1} \psi(F^i(p)) < (K -\epsilon) n\ for \ some \ n\geq n_0) \leq C_1 \exp( -C_2n_0)|P|.\end{equation} Since $G$ is an asymptotically small perturbation, Eq. (\ref{asymp}) implies that \begin{equation}\label{uv} m( p \in P \colon \ \sum_{i=0}^{n-1} \psi(G^i(p)) < (K -\epsilon) n\ for \ some \ n\geq n_0) \leq C_3 \exp( -C_4n_0) |P|\end{equation} provided that $P \in \mathcal{P}^0_j$, $j \geq 2\ |\min \psi|\ n_0$. In particular there exists $n_0=n_0(\delta_1)$ such that for every $P \in \mathcal{P}^0_j$, $j \geq 2\ |\min \psi|\ n_0$, we have \begin{equation} \label{rato} m(\tilde{\Omega}_P)\geq (1-\delta_1)|P|,\end{equation} where $$\tilde{\Omega}_P:= \{ p \in P \colon \ \pi_2(G^n(p))\geq |\min \psi|\ n_0 \ for \ all \ n\geq 0 \ and \ \pi_2(G^n(p))-\pi_2(p)\geq (K -\epsilon)n \ for \ n\geq n_0 \}.$$ By the $GD$ condition, there exists $n_1$ such that for $n \geq n_1$ we have $$m(p \in P\colon \ there \ exists \ i\leq n \ s.t. \ \psi(F^i(p))\geq n)\leq \frac{\delta_1}{4} $$ By Eq (\ref{unpum}) there exists $n_2>n_1$ such that \begin{equation}\label{unpdois} m( p \in P \colon \ \sum_{i=0}^{n_2-1} \psi(F^i(p)) > (K -\epsilon) n_2) \geq (1-\frac{\delta_1}{4})|P|.\end{equation} So \begin{equation}\label{unptres} m( p \in P \colon \ \sum_{i=0}^{n_2-1} \psi(F^i(p)) > (K -\epsilon) n_2 \ and \ \psi(F^i(p))< n_2 \ for \ every \ i\leq n_2) \geq (1-\frac{\delta_1}{2})|P|.\end{equation} Note that for $p$ in the set in Eq (\ref{unptres}) we have $\pi_2(G^i(p))-\pi_2(p)\leq (n_2)^2$ for every $i\leq n_2$. Since $G$ is an asymptotically small perturbation of $F$, this observation and Eq. (\ref{unptres}) implies that there exists $n_3 >> (n_2)^2$ such that for $P \in \mathcal{P}^0_j$, with $j\leq - n_3$, we have \begin{equation} m( p \in P \colon \ \sum_{i=0}^{n_2-1} \psi(G^i(p)) > (K -\epsilon) n_2 \ and \ \psi(G^i(p))< n_2 \ for \ every \ i\leq n_2) \geq (1-\delta_1)|P|.\end{equation} So \begin{equation}\label{pum} m( p \in P \colon \ \sum_{i=0}^{n_2-1} \psi(G^i(p)) > (K -\epsilon) n_2) \geq (1-\delta_1)|P|.\end{equation} {\it Claim $A$:} Almost every point $x \in I\times \{j \}$, $j\leq -n_3$, visits at least once (and consequently infinitely many times) the set \begin{equation}\label{tend} \bigcup_{j\geq-n_3} I\times \{j \}\end{equation} Indeed, define a new random walk $\tilde{G}\colon I\times \mathbb{Z}\rightarrow I\times \mathbb{Z}$ $$\tilde{G}(x,n):=(\tilde{g}_n(x),n+\tilde{\psi}(x,n))$$ in the following way. Let $T$ be an integer larger than $n_2(K-\epsilon)$. If $n \geq -n_3$ then define $\tilde{g}_n\colon I \rightarrow I$ as an affine expanding map, onto on each element of $\mathcal{P}_n^{n_2}$, and $\tilde{\psi}(x,n)=T$. For $(x,n)$, with $n< -n_3$, define $\tilde{G}(x,n)=G^{n_2}(x,n)$. In this case $$\tilde{\psi}(x,n)=\sum_{i=0}^{n_2-1} \psi(G^i(x,n)).$$ It is not difficult to see that the $\tilde{G}$-orbit of a point $(x,n)$, with $n< -n_3$, visits the set in Eq. (\ref{tend}) at least once then the $G$-orbit of $(x,n)$ visits the same set at least once. To prove the claim, it is enough to show that $\tilde{G}$ is strongly transient. Indeed, let $P$ be an element of the Markov partition $\mathcal{P}^{k-1}_j(\tilde{G})$. If $\pi_2(\tilde{G}^i(P))\geq -n_3$, for some $i\leq k$ then $\pi_2(\tilde{G}^{k}(P))\geq -n_3$, so \begin{equation} \label{stest1}\frac{1}{|P|} \int_P \tilde{\psi}\circ \tilde{G}^k \ dm= \frac{1}{|P|} \int_P T \ dm \geq (K-\epsilon)n_2.\end{equation} Otherwise $\pi_2(\tilde{G}^i(P))< -n_3$ for every $i\leq k$. In particular $\tilde{G}^i=G^{i n_2}$ on $P$, for every $i\leq k$. Note that $$\tilde{G}^{k}P= \bigcup_i Q_i,$$ where $Q_i \in \mathcal{P}^0_j(G)$ (this is a consequence of the Markovian property of $G$), with $j < -n_3$. By Eq. (\ref{pum}) we have $$m(q \in Q_i\colon \tilde{\psi}(q) \geq (K-\epsilon)n_2)\geq(1-\delta_1)|Q_i|,$$ so by the distortion control in Eq. (\ref{distg}) we obtain $$m(p \in P\cap \tilde{G}^{-k} Q_i\colon \tilde{\psi}(\tilde{G}^kp) \geq (K-\epsilon)n_2)\geq(1-\tilde{\delta}_1)|P\cap \tilde{G}^{-k} Q_i|,$$ consequently \begin{equation}\label{stest2} \int_P \tilde{\psi}\circ \tilde{G}^k \ dm=\sum_i \int_{P\cap \tilde{G}^{-k} Q_i} \tilde{\psi}\circ \tilde{G}^k \ dm \end{equation} $$ \geq \sum_i ((1-\tilde{\delta}_1)(K-\epsilon)n_2 + \tilde{\delta}_1 n_2 \min \psi) |P\cap \tilde{G}^{-k} Q_i| $$ $$\geq \sum_i (K -2\epsilon)n_2 |P\cap \tilde{G}^{-k} Q_i|= (K -2\epsilon)n_2 |P|$$ Eq. (\ref{stest1}) and (\ref{stest2}) imply that $\tilde{G}$ is strongly transient, so by Proposition \ref{largedeviationsst}, $\tilde{G}$ is transient. This concludes the proof of the claim. {\it Claim $B$:} The $G$-orbit of almost every point of $I\times \mathbb{Z}$ eventually arrives at $\tilde{\Omega}_P$, for some $P \in \mathcal{P}^0_j$, with $j> 2|\min \psi|n_0$. Since $F$ is transient and $G$ is topologically conjugate to $F$ the set $$\Omega:= \{ p \colon \ -n_3\leq \pi_2(p) \leq 2|\min \psi| n_0 \ and \ \lim_n \pi_2(G^n(p))=+\infty\}$$ is dense on $$\bigcup_{j=-n_3}^{2|\min \psi| n_0 } I\times \{j\}.$$ This implies that for every non-empty open set $O \subset I_j$, with $-n_3\leq j \leq 2|\min \psi| n_0$ we have \begin{equation}\label{meio} m((x,j) \in O \colon \exists \ k\geq 0 \ s.t. \ G^k(x,j) \in \tilde{\Omega}_P, \ with \ P \in \mathcal{P}^0_q(G), \ q > 2|\min \psi|n_0)> 0. \end{equation} Indeed, pick a point $p \in O\cap \Omega$. By property $Ex$ and the definition of $\Omega$, there exists $k$ and $Q \in \mathcal{P}_j^k(G)$ such that $Q \subset O$, $P=G^k(Q)\in \mathcal{P}^0_q$, with $q> 2|\min \psi|n_0$. By Eq. (\ref{rato}) we have $m(\tilde{\Omega}_P)> 0$, so $$m(O\cap G^{-k}\tilde{\Omega}_P)\geq m(Q\cap G^{-k}\tilde{\Omega}_P) > 0.$$ By the property $LI$, Eq. (\ref{meio}) and (\ref{rato}), and the bounded distortion of the iterations of $G$, it follows that there exists $\delta_3 > 0$ such that for every $i$ and every $Q \in \mathcal{P}^{i-1}(G)$ such that $\pi_2(G^iQ)\geq -n_3$ we have that \begin{equation}\label{ubc} m(p \in Q\colon \ \exists k\geq 0 \ s.t. \ G^kp \in \tilde{\Omega}_P, \ with \ P \in \mathcal{P}^0_q(G), \ q > 2|\min \psi|n_0)\geq \delta_3 |Q| \end{equation} We will show Claim $B$ by contradiction. Suppose that it does not hold. Then there is a set $W$ of positive measure whose $G$-orbit of its elements never hits $\tilde{\Omega}_P$ for any $P\in \mathcal{P}^0_j$, with $j\geq 2 |\min \psi| n_0$. Pick a Lebesgue density point $p$ of $W$ whose $G$-orbit visits $$\bigcap_{j=-n_3}^{2|\min \psi|n_0} I\times\{j\}$$ infinitely many times, which is possible due Claim A. In particular there exists a sequence $Q_k \in \mathcal{P}^{n_k-1}(G)$ such that $|Q_k|\rightarrow_n 0$, $p \in Q_k$, $\pi_2(G^{n_k}Q_k)\geq -n_3$ and $$\lim_k \frac{m(Q_k\cap W)}{|Q_k|}=1.$$ That contradicts Eq. (\ref{ubc}). This concludes the proof of Claim $B$. Note that Claim $B$ implies the following: almost every point in $I \times \{j\}$ belongs to the set $$\Lambda_j := \bigcup_{k\geq 0} \Lambda_j^k,$$ where $$\Lambda_j^k:= \{p \in I \times \{j\}\colon \pi_2(G^n(p))-\pi_2(G^k(p))\geq (K-\epsilon)(n-k), \ for \ every \ n\geq k +n_0 \}.$$ Let $k_0$ be large enough such that for every $-n_3\leq j\leq 2|\min \psi|n_0$ we have $$m(A\cap \bigcup_{k\leq k_0} \Lambda_j^k) \geq (1-\delta_1) |A|$$ for every interval $A\subset I\times \{j\}$ satisfying $|A|\geq \delta$, where $\delta >0$ is as in the property $LI$. Pick $n_4$ satisfying $n_4\geq k_0+n_0$ and $$n_4 > \frac{-k_0\min \psi}{\epsilon}-k_0.$$ It is easy to see that if $p \in \bigcup_{k\leq k_0} \Lambda_j^k$ then $$\pi_2(G^{n_4}p)-\pi_2(p)=\sum_{i=0}^{n_4-1}\psi(G^ip)\geq (K-2\epsilon)n_4.$$ In a argument similar to the proof of Claim $A$, consider the random walk $\hat{G}$ defined in the following way: if $\pi_2(p)\leq -n_3$ define $\hat{G}(p)=G^{n_2}$. If $\pi_2(p)\geq 2|\min \psi|n_0$ define $\hat{G}(p)=G^{n_0}$. Finaly if $-n_3< \pi_2(p)< 2|\min \psi|n_0$ define $\hat{G}(p)=G^{n_4}$. The random walk $\hat{G}$ is $3\hat{K}$-strongly transient, for some $\hat{K} > 0$. The proof is quite similar to the proof of the strong transience of $\tilde{G}$, so we let it to the reader. So $\hat{G}$ is transient. It is easy to see that this implies that $G$ is transient. Finally Proposition \ref{bru} implies that $$m( p \in P \colon \ \pi_2(\hat{G}^n(p))-\pi_2(p) < 2\hat{K} n )\leq C\hat{\lambda}^n |P|,$$ for some $\hat{\lambda} \in (0,1)$, which implies $$m(Y^n_P)\leq C\hat{\lambda}^n |P|,$$ where $$Y^n_P:=\{p \in P \colon \ \exists \ m \geq n \ s.t. \ \pi_2(\hat{G}^m(p))-\pi_2(p) < 2\hat{K} m\}.$$ Let $n_5=\max\{n_0,n_4,n_2\}$. Let $p \in P$ be such that $$ \pi_2(G^i(p))-\pi_2(p) < \frac{\hat{K}}{n_5} i.$$ There exists $m$ such that $\hat{G}^m(p)=G^j(p)$, with $i \geq j$, $|i-j|\leq n_5$. Note that $$m\leq i\leq j + n_5\leq (m+1)n_5,$$ so we can find $i_0$ such that for every $i\geq i_0$ we have $$ \frac{-n_5 \min \psi}{m} + \hat{K}\frac{m+1}{m}< 2\hat{K}.$$ So $$\pi_2(\hat{G}^m(p))-\pi_2(p)= \pi_2(G^j(p))- \pi_2(G^i(p))+ \pi_2(G^i(p))- \pi_2(p)$$ $$\leq - n_5 \min \psi + \frac{\hat{K}}{n_5} i\leq- n_5 \min \psi + \hat{K}(m+1)< 2\hat{K} m, $$ where $$m\geq \frac{i}{n_5}-1.$$ This implies $$\{ p \in P\colon \pi_2(G^i(p))-\pi_2(p) < \frac{\hat{K}}{n_5} i \}\subset Y^{ \frac{i}{n_5}-1}_P,$$ so $$m( p \in P\colon \pi_2(G^i(p))-\pi_2(p) < \frac{\hat{K}}{n_5} i )\leq C\hat{\lambda}^{i/n_5}|P|$$ This completes the proof. \end{proof} Let $n > 0$ and $j$ be integers and $F$ be a deterministic random walk. Then any connected component $C$ of $F^{-n} \ int \ I_j$ is called a {\bf cylinder}. The {\bf lenght} $\ell(C)$ of the cylinder $C$ is $n$. If $C$ is a cylinder of lenght $n$ so that $F^i(C) \subset I_{j_i}$, for $i<n$, we will denote $C=C(j_0,j_1,\dots,j_n)$. \begin{prop}\label{comeco} Let $F$ be a random walk induced by the pair $(\{f_i\},\psi)$. Assume that there exists $\epsilon > 0$ so that for $K > 0$, we have $$m(\{ p \in I_n \colon \psi(p) < -K\}) \leq \frac{1}{K^{2+\epsilon}},$$ provided $n \geq n_0$. Then $$\lim_k m( \{p \in I_{n_k}\colon \text{ there exists } i \leq k^2 \text{ so that } \psi(F^i(p)) < -k\})=0,$$ uniformly for all sequence satisfying $n_k > k^3 + n_0$. \end{prop} \begin{rem} For a homogeneous random walk, the condition on $\psi$ is equivalent to $1_{I_0}\cdot\psi \in L^{2+\epsilon}(m)$.\end{rem} Let $F$ and $G$ be random walks which are topologically conjugated by a homeomorphism $h$ that preserves states. For any $p \in I \times \mathbb{Z}$ define $$dist_i(p):= \big|\log \frac{DG^i(h(p))}{ DF^i(p)}\big|$$ and $$C_p:= \sup_{i\geq 0} dist_i(p).$$ For each $n_0 \in \mathbb{Z}\cup\{-\infty\}$ define $$\Omega_{n_0+}(F):= \{p \colon \pi_2(F^n(p))\geq n_0, \text{ for all } n \geq n_0\}.$$ In particular $\Omega_{-\infty+}(F)=I\times \mathbb{Z}$. \begin{prop}\label{abs} Let $F$ and $G$ be random walks which are conjugated by a homeomorphism $h$ which preserves states. Suppose that there exists a $F$-forward invariant set $\Lambda$ so that \begin{itemize} \item[ ] \ \item[\it -H1:] $C_p:=\sup_{i\geq 0} \ dist_i(p) < \infty$, for each $p \in \Lambda$.\\ \end{itemize} then $h$ is absolutely continuous on $\cup_i F^{-i}\Lambda$ and $h^{-1}$ is absolutely continuous on $\cup_i G^{-i}h(\Lambda)$. Furthermore, if \begin{itemize} \item[ ] \ \item[\it -H2:] There exists $C > 0$, $M > 0$ and $n_0 \in \mathbb{Z}\cup\{-\infty\}$ so that for every $n\geq n_0$ with $n \in \mathbb{Z}$ and $P \in \mathcal{P}^0_n$, $$m(p \in P \cap \Lambda \colon \ C_p \leq C ) \geq M |P|.$$ \end{itemize} then $h$ is absolutely continuous on $\cup_i F^{-i}(\Omega_{n_0+}(F))$ and $h^{-1}$ is absolutely continuous on $\cup_i G^{-i}(\Omega_{n_0+}(G))$. In particular when $n_0=-\infty$ we have that $h$ and $h^{-1}$ are absolutely continuous on $I\times \mathbb{Z}$. \end{prop} \begin{proof}For each $j \in \mathbb{N}$ denote $$\Lambda_j:= \{ p \in \Lambda \colon \sup_i \ dist_i(p) \leq j\}.$$ Note that $\Lambda_i$ is forward invariant. We claim that $h$ is absolutely continuous on $\Lambda_j$ and $h^{-1}$ is absolutely continuous on $h(\Lambda_j)$. Indeed, for each $p \in \Lambda_j$ and $k \in \mathbb{N}$, denote $F^ip=(x_i,n_i)$. Denote by $J_k(x) \in \mathcal{P}^k$ the unique interval which contains $x$ so that $F^k$ maps $J_k(x)$ diffeomorphically onto $Q_k \subset I_{n_k}$. There is some ambiguity here if $x$ is in the boundary of $J_k(x)$, but these points are countable, so they are irrelevant for us. If we use the analogous notation to $h(x)$ and $G$, we have $h(J_k(x))=J_k(h(x))$ and, due the bounded distortion property of the random walks $F$ and $G$, there exist $C_1, C_2 > 0$ such that $$C_1 e^{-dist_k(p)} \leq \frac{|h(J_k(x))|}{|J_k(x)|} \leq C_2 e^{dist_k(p)}.$$ So, if $p \in \Lambda_j$ then \begin{equation}\label{dis} C_1e^{-j} \leq \frac{|h(J_k(x))|}{|J_k(x)|} \leq C_2 e^j, \ \text{ for all } k \in \mathbb{N}.\end{equation} Let $A \subset \Lambda_j$ be a set with positive Lebesgue measure. We claim that $h(A)$ also has positive Lebesgue measure. Indeed, choose a compact set $K \subset A$ with positive Lebesgue measure. Denote $U_k:= \cup_{x \in K} J_k(x)$. Since $|J_k(x)|\leq \lambda^k$, we have that $\lim_k m(U_k)=m(K)$ and $\lim_k m(h(U_k))=m(h(K))$. Since $U_k$ is a countable disjoint union of intervals of the type $J_k(x)$, by Eq. (\ref{dis}) \begin{equation}\label{dist2} C_1e^{-j} \leq \frac{m(h(U_k))}{m(U_k)} \leq C_2 e^j, \ so \ C_1e^{-j}\leq \frac{m(h(K))}{m(K)} \leq C_2 e^j,\end{equation} and we conclude that $h(K)$ also has positive Lebesgue measure. An identical argument shows that, if $A \in \Lambda_j$ has positive Lebesgue measure, then $h^{-1}A$ also has positive Lebesgue measure. The proof of the claim is finished and so $h$ and $h^{-1}$ are absolutely continuous on $\Lambda=\cup_j \Lambda_j$ and $h(\Lambda)=\cup_j h(\Lambda_j)$. Now it is easy to conclude that $h$ and $h^{-1}$ are absolutely continuous on $\cup_i F^{-i}\Lambda$ and $\cup_i G^{-i}h(\Lambda)$. Now assume $H2$. We claim that $\cup_i F^{-i}\Lambda$ has full Lebesgue measure on $\Omega_{n_0+}(F)$. Indeed, Assume that $m( \Omega_{n_0+}(F)\setminus \cup_i F^{-i}\Lambda) > 0$ and choose a Lebesgue density point $p$ of this set. Then $$ \lim_k \frac{m(J_k(p)\cap \Omega_{n_0+}(F)\setminus \cup_i F^{-i}\Lambda )}{|J_k(x)|} =1.$$ Due the bounded distortion of $F$, if $F^k(p)=(x_k,n_k)$ and $F^k(J_k(x))=Q_k \subset I_{n_k}$, with $n_k\geq n_0$, where $Q_k$ is a union of intervals in $\mathcal{P}^0_{n_k}$, then $$ \limsup_k \frac{m(Q_k\cap \Lambda)}{|Q_k|} \leq C(1- \liminf_k \frac{m(J_k(x)\cap \Omega_{n_0+}(F)\setminus \cup_i F^{-i}\Lambda)}{|J_k(x)|}) =0,$$ which contradicts H2. Since $dist_k(p)$ is uniformly bounded with respect to $k$ and $p$ on the set $\{p \in P\cap \Lambda \colon \ C_p \leq C \}$, we can use an argument identical to the proof of Eq. (\ref{dist2}) to conclude that $$\frac{m(p \in P\cap \Lambda \colon \ C_p \leq C )} {m(h(p) \in h(P)\cap h(\Lambda) \colon \ C_p \leq C )}\leq C_1,$$ so $m(h(P\cap \Lambda\colon C_p\leq C)) \geq \tilde{C}M|h(P)|$, for all $P \in \mathcal{P}^0_n$, $n\geq n_o$ and using an argument as above, we conclude that $\cup_i G^{-i}h(\Lambda)$ has full Lebesgue measure on $\Omega_{n_0+}(G)$. Since $h$ ($h^{-1}$) is absolutely continuous on $\cup_i F^{-i}\Lambda$ ($\cup_i G^{-i}h(\Lambda)$) and $$m(\Omega_{n_0+}(F)\setminus \cup_i F^{-i}\Lambda)=m(h(\Omega_{n_0+}(F)\setminus \cup_i F^{-i}\Lambda))=m(\Omega_{n_0+}(G)\setminus \cup_i G^{-i}h(\Lambda))=0,$$ we have that $h$ and $h^{-1}$ are absolutely continuous on $\Omega_{n_0+}(F)$ and $\Omega_{n_0+}(G)$. Now it is easy to prove that $h$ is absolutely continuous on $\cup_i F^{-i}\Omega_{n_0+}(F)$ and $h^{-1}$ is absolutely continuous on $\cup_i G^{-i}\Omega_{n_0+}(G)$. \end{proof} \begin{proof}[{\bf Proof of Theorem \ref{sttr}}] By Proposition \ref{homstr}, $G$ is transient. In particular for all $n_0 \in \mathbb{Z}$ the sets $$\cup_i F^{-i}\Omega_{n_0+}(F) \text{ and } \cup_i G^{-i}\Omega_{n_0+}(G)$$ have full Lebesgue measure. So by Proposition \ref{abs}, to prove that $h$ and $h^{-1}$ are absolutelly continuous, it is enough to find a forward invariant set satisfying the assumptions H1 and H2 for some $n_0 \in \mathbb{Z}$. Indeed, fix $\delta > 0$ (we will choose $\delta$ latter). Consider the $F$-forward invariant set $$\Lambda =\Lambda_\delta := \{ p \ \colon \ \liminf_k \frac{\pi_2(F^k(p))-\pi_2(p)}{k} \geq \frac{\delta}{3} \}.$$ We claim that $\Lambda$ satisfies H1. Indeed take $x \in \Lambda$. Then, for $k \geq k_0(x)$ we have $n_k:=\pi_2(F^k(p)) \geq k \delta/4$. So \begin{equation} \label{distcont} dist_k(x)\leq \sum_{i=0}^{k-1} |\log \frac{DF(F^{i+1}(p))}{ DG(h(F^{i+1}(p)))}| \end{equation} $$\leq \sum_{i=0}^{k_0-1} |\log \frac{ DF(F^{i+1}(p))}{ DG(h(F^{i+1}(p)))}| +\sum_{i=k_0}^{k-1} |\log \frac{ DF(F^{i+1}(p))}{ DG(h(F^{i+1}(p)))}|$$ $$\leq \sum_{i=0}^{k_0-1} |\log \frac{ DF(F^{i+1}(p))}{ DG(h(F^{i+1}(p)))}| + \sum_{i=k_0}^{k-1} \lambda^{n_i}$$ $$\leq \sum_{i=0}^{k_0-1} |\log \frac{ DF(F^{i+1}(p))}{ DG(h(F^{i+1}(p)))}| + \sum_{i=k_0}^{\infty} \lambda^{i\delta/4 }$$ $$ \leq K_p + C(\delta).$$ To prove that $\Lambda$ satisfies H2, By Proposition \ref{largedeviationsst} for each $P \in \mathcal{P}^0_i$ we have \begin{equation}\label{estexp} m( p \in P \colon \ \pi_2(F^k(p))-\pi_2(p) < \delta k )\leq C\lambda^k |P|,\end{equation} provided $\delta$ is small enough. From Eq. (\ref{estexp}) we obtain \begin{equation}\label{estr2} \mu( p \in P \colon \ \pi_2(F^n(p))-\pi_2(p) \geq \delta n \text{ for all } n\geq n_0 )\geq (1-C\lambda^{n_0}) |P|.\end{equation} In particular, we have that, for every $n$, \begin{equation}\label{estest} \pi_2(F^n(p))\geq \delta (n-n_0) + \pi_2(p)+ n_0 \min \psi.\end{equation} in the set in Eq. (\ref{estr2}). Using the same argument as in Eq. (\ref{distcont}) we can easily obtain $H2$ from Eq. (\ref{estest}) and Eq. (\ref{estr2}), choosing $n_0$ large enough. \end{proof} \begin{proof}[{\bf Proof of Theorem \ref{sttrho}}] Observe that using the argument in the proof of Proposition \ref{homstr}, an induced map of a homogeneous random walk with positive drift is strongly transient. From this the proof of Theorem \ref{sttrho} goes exactly as the Theorem \ref{sttr}. \end{proof} \begin{proof}[{\bf Proof of Theorem \ref{sqr}}] By Proposition \ref{prws}, for every $i$ we have $$m( p \in I_i \ \colon \ \frac{\pi_2(F^k(p))-\pi_2(p)}{k} \leq \delta ) \leq C\theta^{k}.$$ and furthermore $\theta:=\theta(\delta)$ tends to $0$ when $\delta$ tends to zero. Using an argument as in the proof of Theorem \ref{sttr} we can conclude that \begin{equation} \label{equsei} m( p \in I_i \ \colon \ \frac{\pi_2(F^k(p))-\pi_2(p)}{k} \geq \delta \ for \ k \geq k_0) \geq 1- C\theta^{k_0}\end{equation} In particular we can use the argument in the proof of Theorem \ref{sttrho} to conclude that the conjugacy $h$ is absolutely continuous. Indeed, Eq. (\ref{equsei}) implies \begin{equation} \label{bdinf} m( p \in I_i \ \colon \ dist_{k}(x) \geq \delta n+C \ for \ some \ k)\leq C\theta^n.\end{equation} where $\delta= \sup_p dist_1(p)$. Firstly we will prove Theorem \ref{sqr} when $\delta$ is small. Denote $\Lambda_1:= \{ p \in I_i \ \colon h'(x) \leq 1\}$ and, for $n \geq 1$ $$\Lambda_n:= \{ p \in I_i \ \colon \ {e}^{\delta(n-1)} < h'(x) \leq {e}^{\delta n}\}.$$ By Eq. (\ref{bdinf}) we have $m(\Lambda_n)\leq C\theta^n$. Let $B \subset I_i$ be an arbitrary Lebesgue measurable set. Let $k_1$ be so that $$ \theta^{k_1+1}< |B|\leq \theta^{k_1}.$$ Since $h$ is absolutely continuous we have $$|h(B)|= \int_B h'\ dm$$ $$ = \sum_{n=0}^{k_1} \int_{B\cap \Lambda_n} h' \ dm + \sum_{n=k_1+1}^{\infty} \int_{B\cap \Lambda_n} h' \ dm$$ $$ \leq \sum_{n=0}^{k_1} C\theta^{k_1}e^{\delta n} + \sum_{n=k_1+1}^{\infty} C(e^{\delta}\theta)^n$$ $$ \leq C (e^{\delta}\theta)^{k_1}\leq C |B|^{1+ \frac{\delta}{\ln \theta}}.$$ Now if $B \subset J \in \mathcal{P}^n$ and $F^n(J)=Q \subset I_i$, with $|Q|\geq C$ (due Property LI), then due the bounded distortion of $F$ $$\frac{|h(B)|}{|h(J)|}\leq C \frac{|h(F^n(B)|}{|h(Q)|} \leq C \Big( \frac{ |F^n(B)|}{|Q|} \Big)^{1+ \frac{\delta}{\ln \theta}}\leq C\Big( \frac{ |B|}{|J|} \Big)^{1+ \frac{\delta}{\ln \theta}}.$$ To prove a similar inequality to $h^{-1}$, define $$\tilde{\Lambda}_n:= \{ p \in I_i \ \colon \ {e}^{\delta(n-1)} < (h^{-1})'(x) \leq {e}^{\delta n}\}.$$ of course $$h^{-1}\tilde{\Lambda}_n = \{ p \in I_i \ \colon \ {e}^{-\delta(n)} < h'(x) \leq {e}^{-\delta (n-1)}\},$$ so by Eq. (\ref{bdinf}) we obtain $$m(h^{-1}\tilde{\Lambda}_n)\leq \theta^n.$$ In particular $$m(\tilde{\Lambda}_n) = \int_{h^{-1}\tilde{\Lambda}_n} h'(x) \ dm \leq (e^{-\delta}\theta)^n$$ Note that this argument gives us an exponential upper bound even if $\delta$ is large. Now we can switch the roles of $F$ and $G$ to obtain the inequality to $h^{-1}$, which shows that $h$ is a mSQS-homeomorphism relative to the stochastic basis $\cup_n \mathcal{P}^n$. To complete the proof when $\delta$ is not small do the following: find a continuous path of random walks $F_t$ so that $F_0=F$ and $F_1=G$, so that for every $t \in [0,1]$ we have that $F_t$ is a asymptotically small perturbation of $F$. By the argument above for every $t \in [0,1]$ there exists $\epsilon_t$ so that $F_{\tilde{t}}$ is mSQS-conjugated to $F_t$, provided $|\tilde{t}-t|\leq \epsilon_t$. Using the compactness of $[0,1]$ we can find a finite sequence of random walks $F_{t_0}=F, F_{t_1}, F_{t_2}, \dots F_{t_n}=G$ so that $F_{t_i}$ and $F_{t_{i+1}}$ are conjugated by a map $h_i$ which is mSQS with respect some dynamically defined stochastic basis. Composing these conjugacies we find a mSQS-conjugacy between $F$ and $G$. \end{proof} \section{Stability of recurrence} To avoid a cumbersome notation, in this section we make the convention that all inequalities holds only for large $n$. Moreover in this section we assume that $\psi$ is unbounded. Recall that in this case we assume that asymptotically small perturbations $G$ coincides with $F$ on negative states. The case where $\psi$ is bounded is similar. The following is a easy consequence of the Central Limit Theorem for Birkhoff sums (Proposition \ref{clt}) \begin{cor}\label{co}Let $a_n$ be a positive increasing sequence. Then $$\mu( \frac{|S_n|}{\sqrt{n}} > a_n) \leq Ce^{-\frac{a_n^2}{2}} + C\frac{1}{\sqrt{n}}.$$ \end{cor} \begin{proof} Use Proposition \ref{clt} and and note that the estimative $$\int_{-\infty}^v e^{-\frac{u^2}{2}} \ du \leq C e^{-\frac{v^2}{2} }$$ holds for $v << 0$. \end{proof} Given $n \in \mathbb{N}$, split $[0,2n]\cap \mathbb{N}$ in $\sqrt{\log n}$ blocks (called main blocks) , denoted $B_j$, with length $$\frac{n}{\log^{8j}n }, \ j=1,\dots, \sqrt{\log n},$$ and between the main blocks we put little blocks $H_j$, called holes, of length $\log^4n$. These holes will warranty the independence between the events in distinct main blocks. Put these blocks in the following order: $$\dots < B_{j+1} < H_{j+1} < B_j < H_{j}< \dots,$$ with $min \ B_{\sqrt{\log n}} = 0.$ Note that we let most of the second half of the interval $[0,2n]\cap \mathbb{N}$ uncovered. Define $$S(j) = \sum_{i \in B_j} \psi\circ f^i$$ $$H(j) = \sum_{i \in H_j} \psi\circ f^i$$ Denote $|B_j|:= \max B_j - \min B_j$. \begin{lem}\label{muitogrande} We have $$\mu(\sum_{i=0}^{|B_j|} \psi\circ f^i \geq \frac{\sqrt{n}}{\log^{4j} n} \log^3n) \leq C \frac{\log^{4j} n}{\sqrt{n}}.$$ \end{lem} \begin{proof} This follows from Corollary \ref{co}. \end{proof} \begin{prop}\label{aux1} For every $\epsilon > 0$ we have $$\mu(S(j) > \frac{\sqrt{n}}{\log^{4j} n} \log^3n, \ for \ some \ j \leq \sqrt{\log n}) \leq C\frac{1}{\sqrt[2+\epsilon]{n}},$$ provided $n$ is large enough. \end{prop} \begin{proof} For $j \leq \sqrt{\log n}$ define $$ \Lambda_j:=\{x \in I\colon \ S(j)(x) > \frac{\sqrt{n}}{\log^{4j} n} \log^3n \}$$ $$ = \{x \in I\colon \sum_{i < |B_j|} \psi\circ f^{i + \min B_j }(x) > \frac{\sqrt{n}}{\log^{4j} n} \log^3n \} $$and for each $P \in \mathcal{P}^{\min B_j}$ denote $\Lambda_j(P):= \Lambda_j\cap P$. Due Lemma \ref{muitogrande} and the bounded distortion of $f^{\min B_j}$ on $P$ we have $$m(\Lambda_j(P)) \leq C \frac{\log^{4j} n}{\sqrt{n}}|P|.$$ Summing on $j$ and $P$ $$m(\bigcup_{{ j }}\bigcup_{{ P }} \Lambda_j(P)) \leq \sqrt{\log n} \ \frac{\log^{4j} n}{\sqrt{n}} << C\frac{1}{\sqrt[2+\epsilon]{n}}.$$ \end{proof} \begin{prop}\label{aux2} For every $\epsilon > 0$ and $d >0$ we have \begin{equation}\label{mais} \mu(|\sum_{i \in H_j} \psi(f^i(x))| > \log^8 n, \ for \ some \ j \leq \sqrt{\log n}) \leq C\frac{1}{n^d},\end{equation} provided $n$ is large enough. \end{prop} \begin{proof} For $i \in H_j-1$, with $j \leq \sqrt{\log n}$, define $$\Lambda_{i}:=\{x \in I\colon |\psi(f^i(x))| > \log^4 n. \}.$$ By expanding and bounded distortion properties of $f$ and condition $GD$ we have that $$\mu(\Lambda_i)\leq C\lambda^{\log^4 n}.$$ Since $|H_j|= \log^4 n$, if $x$ belongs to the set in Eq. (\ref{mais}) then $x \in \Lambda_{i}$, for some $i \in H_j-1$, with $j \leq \sqrt{\log n}$. So $$\mu(|\sum_{i \in H_j} \psi(f^i(x))| > \log^8 n, \ for \ some \ j \leq \sqrt{\log n})$$ $$ \leq \mu(\bigcup_{{ j \leq \sqrt{\log n}}} \ \bigcup_{{ i \in H_j-1}} \Lambda_j)$$ $$\leq \sqrt{\log n}\ \log^4 n \ n^{\log \lambda \log^3n}$$ $$<< \frac{1}{n^d}, $$ where the least inequality holds for $n$ large enough. \end{proof} \begin{prop}[Independence between distant events] \label{inde}There exists $\lambda < 1$ so that the following holds: For all cylinders $C_1$ and $C_2$, we have $$\mu(C_1\cap f^{-(n+d)}C_2) = \mu(C_1 )\mu(C_2)(1+ O(\lambda^{d})).$$ Here $n=|C_1|$. \end{prop} \begin{proof} Let $J$ be an interval in $C_1$ so that $f^n(J)=I$. Define the measure $\rho(A):= \mu(f^{-n}A\cap J)/\mu(J)$. Note that by the bounded distortion property of $f$, we have that $\log d\rho/dm$ is $\alpha$-Holder, where $\alpha$ does not depend on $n$. Furthermore it is bounded by above by a constant which does not depend on $n$. By the well-know theory of Perron-Frobenius-Ruelle operators for Markov expanding maps, if $P$ is the Perron-Frobenius-Ruelle operator of $f$, then there exists $\lambda < 1$ so that $$ P^d\frac{d\rho}{dm} = (1+ O(\lambda^d))\frac{d\mu}{dm}.$$ So $$\frac{\mu(J\cap f^{-(n+d)}C_2)}{\mu(J)}$$ $$=\rho(f^{-d}C_2)= \int 1_{C_2}\circ f^d \ \frac{d\rho}{dm} dm $$ $$=\int 1_{C_2} \ P^d \frac{d\rho}{dm} dm$$ $$=(1 + O(\lambda^d))\int 1_{C_2} \ \frac{d\mu}{dm} dm$$ $$=(1 + O(\lambda^d))\mu(C_2).$$ Since $C_1$ is a disjoint union of intervals $J$ so that $f^{n}J=I$, we finished the proof. \end{proof} \begin{cor}\label{aux3} There exists $M > 0$ so that $$\mu( S_j < \frac{\sqrt{n}}{\log^{4j} n}\ M \text{ for all } j \leq \sqrt{\log n}) \leq C\big( \frac{2}{3} \big)^{\sqrt{\log n}}$$ \end{cor} \begin{proof}Choose $M >0 $ so that $$\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{M} e^{-\frac{u^2}{2}} \ du < \frac{2}{3}$$ Consider the disjoint union of cylinders $$C_j:= \{ x \ s.t. \ \sum_0^{|B_j|} \psi\circ f^i(x) < \frac{\sqrt{n}}{\log^{4j} n}\ M\}.$$ The Central Limit Theorem tells us that if $n$ is large enough then $$\mu(C_j) < \frac{2}{3}$$ for every $j \leq \sqrt{\log n}$. Recall that between $B_j$ and $B_{j+1}$ there is a hole with length $\log^4 n$. Applying $\sqrt{\log n}$ times Proposition \ref{inde} , we obtain $$\mu( S_j < \frac{\sqrt{n}}{\log^{4j} n}\ M \text{ for all } j \leq \sqrt{\log n}) \leq \big( \frac{2}{3} \big)^{\sqrt{\log n}} (1 + O(\lambda^{\log^4 n}))^{\sqrt{\log n}} \leq C\big( \frac{2}{3} \big)^{\sqrt{\log n}}$$ \end{proof} \begin{prop}\label{sei} There exists $C > 0$ so that for every $k$, $$\mu( x \in I_k \colon \text{ there exists } i < \ell^3 \text{ so that } \sum_{k=0}^{i} \psi\circ f^k(x) > \frac{\ell}{2}) \geq 1- C\big( \frac{2}{3} \big)^{\sqrt{3\log \ell}}$$ \end{prop} \begin{proof} Let $M$ be as in Proposition . Denote $n = \ell^3$ and define $$A_\ell:= \{x \colon \text{ there exists } i < \ell^3 \text{ so that } \sum_{k=0}^{i} \psi\circ f^k(x) > \frac{\ell}{2}\},$$ $$B_\ell:= \{ x \colon \ |S_j(x)| < \frac{\sqrt{n}}{\log^{4j} n} \log^3n, \text{ for all } j \leq \sqrt{\log n}\},$$ $$C_\ell:= \{x \colon S_j(x) \geq \frac{\sqrt{n}}{\log^{4j} n} \ M, \text{ for some } j \leq \sqrt{\log n} \},$$ $$D_\ell:= \{x \colon |H_j(x)| \leq \log^8 n, \text{ for all } j \leq \sqrt{\log n} \}.$$ We claim that if $\ell$ is large then $B_\ell\cap C_\ell \cap D_\ell \subset A_\ell$. Indeed, let $x \in B_\ell\cap C_\ell \cap D_\ell$. Then for some $j_0 \leq \sqrt{\log n}$, $$S(j_0) \geq \frac{\sqrt{n}}{\log^{4j_0} n} \ M.$$ We claim that, if $m = max \ B_{j_0}$, then $$\sum_{0}^m \psi\circ f^i(x) > \frac{\ell}{2}.$$ Indeed, since $x \in D_\ell$, $$| \sum_{i\in H_{j}, \ j > j_0} \psi\circ f^i(x)| \leq \sqrt{\log n} \log^8 n = o(\ell).$$ Moreover, since $ x \in B_\ell$, $$| \sum_{i \in B_{j}, \ j > j_0} \psi\circ f^i(x)| \leq \sum_{j > j_0} \frac{\sqrt{n}}{\log^{4j} n } \log^3 n \leq C \frac{\sqrt{n}}{\log^{4j_0 +4} n}. $$ So $$\sum_{0}^m \psi\circ f^i(x) = \sum_{i \in B_{j_0}} \psi\circ f^i(x) + \sum_{i \in B_{j}, \ j > j_0} \psi\circ f^i(x) + \sum_{i \in H_{j}, \ j > j_0} \psi\circ f^i(x)$$ $$ \geq \big(M- \frac{C}{\log^4 n}\big) \frac{\sqrt{n}}{\log^{4j_0} n} + o(\ell) > C\ell - o(\ell)> \frac{\ell}{2},$$ and we finished the proof of the claim. To finish the proof, note that by Proposition \ref{aux1}, Corollary \ref{aux3} and Proposition \ref{aux2} $$\mu(A_\ell) \geq \mu(B_\ell\cap C_\ell \cap D_\ell) \geq 1- C\frac{1}{\sqrt[2+\epsilon]{n}}-C\big( \frac{2}{3} \big)^{\sqrt{\log n}} - C\frac{1}{n^d} \geq 1- C\big( \frac{2}{3} \big)^{\sqrt{\log n}}.$$ \end{proof} \begin{prop} There exist $\epsilon$ and $D$ so that for every $\ell \geq 0$, $$\mu (\{ x \in I_\ell \colon \text{ there exists } i \text{ so that } F^i(p) \in \bigcup_{t \in [\min \psi, -\min \psi]} I_t \text { and } dist_i(p) \leq D \}) \geq \epsilon$$ \end{prop} \begin{proof} Define, for $p \in C(i_0,i_1,\dots,i_{n-1})$, $$Dist_n(p):= sup_{q \in C(i_0,i_1,\dots,i_{n-1})} \ dist_n(q).$$ We are going to prove by induction on $k$ that there is $C > 0$ so that, if we define $$B_k^{\ell}:= \{ p \in I_\ell \colon \text{ there exists } j \leq \sum_{i=0}^{k-1} \frac{\ell^3}{2^{3i}}\text{ such that } \pi_2(F^j(p)) \leq \frac{\ell}{2^k} \text{ and } Dist_j(p) \leq \sum_{i=0}^{k-1} \frac{\ell^3}{2^{3i}}\theta^{\frac{\ell}{2^i}}\},$$ then \begin{equation} \label{mest} \mu(B_k^{\ell}) \geq \prod_{i=0}^{k-1} \left( 1- C\big( \frac{2}{3} \big)^{\sqrt{\log \frac{\ell}{2^i}}} \right). \end{equation} Indeed, take $p \in B_k^\ell$. Let $p \in L=C(i_0,i_1,\dots,i_{j-1})$, where $j$ is as in the definition of $B_k^\ell$. Note that $L \subset B_k^\ell$ and $F^n(L)=I_r$, for some $r < \ell/2^k$. By Proposition \ref{sei}, \begin{equation}\label{oest} \mu( x \in I_r \colon \text{ there exists } i < \frac{\ell^3}{2^{3k}} \text{ so that } \sum_{k=0}^{i} \psi\circ f^k(x) > \frac{\ell}{2^{k+1}})\end{equation} $$ \geq 1- C\big( \frac{2}{3} \big)^{\sqrt{\log \frac{\ell}{2^k}}}.$$ Denote $$D_L: = \{ x \in I_\ell\cap L \colon \text{ there exists } i < \frac{\ell^3}{2^{3k}} \text{ so that } \sum_{k=0}^{i} \psi\circ f^k(f^j(x)) > \frac{\ell}{2^{k+1}} \}$$ Due the bounded distortion property for $F$, the estimative in Eq. (\ref{oest}) implies \begin{equation} \label{estimative} \frac{\mu(D_L)}{|L|} \geq 1- C\big( \frac{2}{3} \big)^{\sqrt{\log \frac{\ell}{2^k}}}.\end{equation} For $x \in D_L$ take the smallest $i$ so that $$\sum_{k=0}^{i} \psi\circ f^k(f^j(x)) > \frac{\ell}{2^{k+1}}.$$ Then $\pi_2(F^{j+h}(p)) \geq \frac{\ell}{2^{k+1}}$, for every $0 \leq h < i$, so $$Dist_i(F^{j}(p)) \leq \sum_{h=0}^{i} \theta^{\pi(F^{j+h}(p))} \leq \frac{\ell^3}{2^{3k}} \theta^{\frac{\ell}{2^{k+1}}}.$$ So $D_L \subset B^{\ell}_{k+1}$. Since $B^{\ell}_{k}$ is a disjoint union of cylinders $L$, the estimative in Eq. (\ref{estimative}) implies Eq. (\ref{mest}). Define $$D:= \sum_{i=0}^{\infty} \frac{\ell^3}{2^{3i}}\theta^{\frac{\ell}{2^i}} < \infty.$$ Let $k$ be so that $ 2^k \leq \ell \leq 2^{k+1}$. Now it is easy to check that $$\mu (\{ x \in I_\ell \colon \text{ there exists } i \text{ so that } F^i(p) \in I_0 \text { and } dist_i(p) \leq D \})$$ $$ \geq C\mu(B^{\ell}_k) \geq \prod_{i=0}^{k-1} \left( 1- C\big(\frac{2}{3} \big)^{\sqrt{\log \frac{\ell}{2^i}}} \right) \geq C \prod_{i=0}^{k-1} \left( 1- C\big(\frac{2}{3} \big)^{\sqrt{\log \frac{2^k}{2^i}}} \right)$$ $$ \geq \exp( -C\sum_{i=1}^{\infty} \big(\frac{2}{3} \big)^{\sqrt{i\log 2}}) > \tilde{C} > 0,$$ which finishes the proof. \end{proof} \begin{proof}{\bf Proof of the Stability of Recurrence (Theorem \ref{strec})} Because $G$ coincides with $F$ on negative states, and $F$ is recurrent, of course the orbit by $G$ of almost every point $p$ so that $\pi_2(p) < 0$ will entry $$\cup_{i\geq 0} I^i.$$ So it is enough to prove that the orbit by $G$ of almost every point $p \in \cup_{i\geq 0} I^i$ hit $I^0$. Let $\ell \geq 0$. By the previous Proposition, there exist $D~>~0$ and $\epsilon > 0$ so that $$A_\ell:= \{ p \in I_\ell \colon \text{ there exists } i \text{ so that } F^i(p) \in \bigcup_{t =\min \psi}^{ -\min \psi} I_t \text{ and } Dist_i(p) < D \}$$ satisfies $\mu(A_\ell) > \epsilon$, for all $\ell$. Consider a cylinder $C_F=C_F(\ell,k_1,\dots,k_{i-1},0) \subset A_\ell$, satisfying $k_j \neq 0$ for $0< j < i$ and $Dist_i(x) < D$, for every $x \in C_F$. We claim that that corresponding cylinder $C_G=C_G(\ell,k_1,\dots,k_{i-1},0)$ for the perturbed random walk $G$ satisfies $$\frac{1}{C} \leq \frac{|C_G|}{|C_F|} \leq C,$$ where $C$ depends only on $D$. Since $A_\ell$ is a disjoint union of cylinders of this type, we obtain that $B_\ell = H(A_\ell)$ satisfies $m(B_\ell) > C\epsilon > 0$, for all $\ell$. To prove that the set of points whose orbits returns infinitely many times to $$\bigcup_{t = \min \psi}^{ -\min \psi} I_t$$ has full Lebesgue measure, it is enough to prove that $\Lambda:=\cup_{j>0, \ell } G^{-j}B_\ell$ has full Lebesgue measure. Indeed, assume by contradiction that $\Lambda$ is not full. Choose a Lebesgue density point $p$ of the complement of $\Lambda$. Then there exist a sequence of cylinders $C_k~=~C_G(\ell_0,\ell_1,\dots,\ell_k)$ so that $p \in C_k$ and \begin{equation} \label{conv} \frac{m(C_k\setminus \Lambda)}{|C_k|} \rightarrow_k 1. \end{equation} But $G^k(C_k)=I_{\ell_k}$, and $m(I_{\ell_k} \cap B_{\ell_k}) \geq C\epsilon |I_{\ell_k}|$. By the bounded distortion property $$\frac{m(\Lambda \cap C_k)}{|C_k|} > \frac{m(G^{-k}B_{\ell_k} \cap C_k)}{|C_k|} > \tilde{C} \epsilon,$$ which contradicts Eq. (\ref{conv}). Now we can use that $G$ is transitive and has bounded distortion to prove that $G$ is recurrent. \end{proof} \begin{proof}[\bf Proof of Proposition \ref{rigidity}] Since $F$ is recurrent, almost every point of $I^0$ returns to $I^0$ at least once. So the first return map $R_F\colon I^0 \rightarrow I^0$ is defined almost everywhere is $I^0$ and the same can be said about $R_G$. Of course, the absolutely continuous conjucagy $H$ also cojugates the expanding Markovian maps $R_F$ and $R_G$. Using the same argument used in Shub and Sullivan \cite{ss} and Martens and de Melo \cite{mm}, we can prove that $H$ is actually $C^1$ on $I^0$. Using the dynamics, it is easy to prove that $H$ is $C^1$ everywhere. \end{proof} \section{Stability of the multifractal spectrum} \subsection{Dynamical defined intervals and root cylinders} When we are dealing with Markov expanding maps with {\em finite} Markov partitions, for each arbitrary interval $J$ we can find an element of $\cup_j \mathcal{P}^j$ which covers $J$ and has more or less the same size that $J$. Note that this is no longer true when the Markov partitions is infinite. Since coverings by intervals are crucial in the study of the Hausdorff dimension of an one-dimensional set, this trick is very useful to estimate the dimension of dynamically defined sets, once we can replace an arbitrary covering by intervals by another one with essentially the same metric properties but whose elements are themselves {\em dynamically defined} sets (cylinders). Consider $j\geq 0$ and let $\{C_i\}_i \subset \mathcal{P}^j$ be a finite or countable family of cylinders $\{C_i\}_{i \in \Theta} \subset \mathcal{P}^j$ such that $W:=\bigcup_i \overline{C_i}$ is connected and $int \ W$ does not contain any point $d^n_i$ (as defined in property $Rb$). Then $W$ is called a dynamically defined interval (dd-interval, for short). Define the root cylinder of $W$ as the unique cylinder $C_{i_0}$ with the following property: if $\sharp \Theta =\infty$ then $W$ is a semi-open interval and $C_{i_0}$ will be the cylinder so that $\partial C_{i_0}\cap \partial W \neq \emptyset$. Otherwise $W$ is closed and let $C_{i_0}$ be the unique cylinder such that $F=\partial C_{i_0}\cap \partial W$ is the boundary of a semi-open dd-interval which contains $W$. The following Lemmas are an easy consequence of the regularity properties $Ra+Rb$ and it will be useful to recover the trick described above for (certain) infinite Markov partitions. The proof is very simple. \begin{lem}\label{rcia} For every $d \in (0,1)$ there exists $K > 1$ so that for every dd-interval $W:=\cup_i \overline{C_i}$ with root cylinder $C_{i_0}$ we have $$\frac{1}{K}\leq \frac{|W|^\alpha}{\sum_{i}|C_i|^\alpha} \leq K$$ $$\frac{1}{K}\leq \frac{|C_{i_0}|^\alpha}{\sum_{i}|C_i|^\alpha} \leq K$$ for every $\alpha \geq d$. Indeed the constant $K$ depends only on $d$ and constants in the properties $Ra+Rb+Ex+BD$. \end{lem} \begin{lem}\label{rci} Let $N$ be as in Properties $Ra+Rb$. For every $d \in (0,1)$ there exists $K > 1$ so that the following holds: For every interval $J \subset I\times \mathbb{Z}$ there exists $m$ dd-intervals $W_j$, all of same level, with $m \leq 2N$, satisfying the following properties: \begin{itemize} \item[-] The interior of these dd-intervals are pairwise disjoint. \item[-] The closure of the union of $W_j$ covers $J$: $$J \subset \overline{\bigcup_j W_j}. $$ \item[-] We have $$ \frac{1}{K} \leq \frac{\sum_{i=1}^m |W_i|^{\alpha}}{|J|^\alpha} \leq K$$ for every $\alpha > d$. \end{itemize} Indeed the constant $K$ depends only on $d$ and constants in the properties $Ra+Rb+Ex+BD$.\end{lem} \subsection{Dimension of dynamically defined sets} Let $f \in Mk+ BD + Ex$ and denote by $\mathcal{P}^0$ its Markov partition. Let $$\mathcal{I}:= \{ C_i\}_i \subset \cup_i \mathcal{P}^n$$ be a finite or countable family of disjoint cylinders. Define the induced Markov map $f_{\mathcal{I}}\colon \cup_i C_i \rightarrow I$ by $$f_{\mathcal{I}}(x) = f^{\ell(C_i)-1}(x), \ if \ x \in C_i.$$ We can also define an induced drift function $\Psi\colon \cup_{i} C_i \rightarrow \mathbb{Z}$ in the following way: Define, for $x \in C \in \mathcal{P}^n_0$, $$\Psi_{\mathcal{I}}(x):= \sum_{i=0}^{n-1} \psi(f^i(x)).$$ On the same conditions on $x$, define $N_{\mathcal{I}}(x)=n$. The maximal invariant set of $f_{\mathcal{I}}$ is $$\Lambda(\mathcal{I}):= \{x \in I \colon \ f^j(x) \in \bigcup_i C_i, \ for \ all \ j\geq 0 \}.$$ Denote by $HD(\mathcal{I})$ the Hausdorff dimension of the maximal invariant set of $f_{\mathcal{I}}$. We are going to use the following result \begin{prop}[Theorem 1.1 in \cite{mu2}] We have $$HD(\mathcal{J})= \sup \{HD(\mathcal{I})\colon \mathcal{I} \subset \mathcal{J}, \ \mathcal{I} \ finite \}.$$ \end{prop} The following result was proved to Markov maps with finite Markov partition, however the proof can be adapted to our case. Before to give the proof of Proposition \ref{ms} we need to introduce some tools which are useful to estimate the Hausdorff dimension. Let $\mathcal{J}$ as above. If there exists $\beta$ such that $$\sum_{C \in \mathcal{J}} |C|^\beta =1,$$ we will call $\beta$ the {\bf virtual Hausdorff dimension } of $f_{\mathcal{I}}$, denoted $VHD(\mathcal{I})$. The virtual Hausdorff dimension is a nice way to estimate $HD(\mathcal{I})$: indeed if $f_{\mathcal{I}}$ is linear on each interval of the Markov partition then these values coincide. When the distortion is positive, these values remain related, as expressed in the following result (which is included, for instance, in the proof of Theorem 3, Section 4.2 of \cite{pt}): \begin{prop}\label{vhd} Let $\mathcal{I}$ be a finite family of disjoint cylinders. Then $$|HD(\mathcal{I})-VHD(\mathcal{I})| \leq \frac{d}{\log \lambda - d},$$ where $$d := \sup_{C \in \mathcal{I}} \sup_{x, y \in C} \log \frac{Df_{\mathcal{I}}(y)}{Df_{\mathcal{I}}(x)} \text{ and } \lambda := \inf_{C \in \mathcal{I}} \inf_{x \in C} |Df_{\mathcal{I}}|.$$ \end{prop} Recall that if $\mathcal{I}$ is finite then $f_{\mathcal{I}}$ has an invariant probability measure $\mu_{\mathcal{I}}$ supported on its maximal invariant set $\Lambda(\mathcal{I})$ such that for any subset $S \subset \Lambda(\mathcal{I})$ satisfying $\mu_{\mathcal{I}}(S)=1$ we have $HD(S)=HD(\mathcal{I})$. Note that for a homogeneous random walk $F$ $$\Omega_+^k(F)= \{k\}\times\{x \in I\ s.t. \sum_{i=0}^{j} \psi(f^j(x))+ k\geq 0, \ for \ j\geq 0 \}$$ and $$\Omega_{+\beta}^k(F)=$$ $$\{k\}\times \{ x \in I\ s.t. \ \sum_{j=0}^{n-1}\psi(f^j(x)) + k \geq 0\, \ for \ all \ n\geq 0 \ and \ \underline{\lim}_{\ n} \frac{1}{n}\sum_{j=0}^{n-1}\psi(f^j(x)) \geq\beta\}.$$ Define $\pi_1(x,n):=x$. The following is an easy consequence of this observation: \begin{lem}\label{aux} If $F$ is a homogeneous random walk then $\pi_1(\Omega_{+}^0(F)) \subset \pi_1(\Omega_{+}^k(F))$ and $\pi_1(\Omega_{+\beta}^0(F)) \subset \pi_1(\Omega_{+\beta}^k(F))$, for all $k\geq 0$. Furthermore $$HD(\Omega_{+}^0(F)) =HD(\Omega_{+}^k(F))$$ and $$HD(\Omega_{+\beta}^0(F)) =HD(\Omega_{+\beta}^k(F)).$$ \end{lem} \begin{prop}\label{sim} Let $F$ be a homogeneous random walk. Then there exists a sequence of finite families of cylinders $$\mathcal{\mathcal{F}}_s \subset \cup_i \mathcal{P}^i_0$$ so that \begin{itemize} \item[ ] \ \item[-]$\Lambda(\mathcal{F}_s) \subset \Omega_+^0(F),$\\ \item[-] Denote $\beta_n:= \int \Psi_{\mathcal{F}_s}\ d\mu_{\mathcal{F}_s }$. Then $\beta_n > 0$. \\ \item[-] $\lim_{s\rightarrow \infty} HD(\mathcal{F}_s)=HD(\Omega_+^0(F)).$ \\ \end{itemize} \end{prop} \begin{proof} Denote $d=HD \ \Omega_{+}^0(F)$. Given any $s \in \mathbb{N}^\star$, $m_{d_s} \ \Omega_{+}(F)=\infty$, where $d_s:=d(1-1/s)$. Here $m_{D}$ denotes the $D$-dimensional Haussdorf measure. By Theorem 5.4 in \cite{f}, for each positive number $M$ we can find a compact subset $\Lambda_s \subset \Omega_{+}^0(F)$ satisfying $m_{d_s} \ \Lambda_s = M$. We may assume that $\Lambda_s$ does not have isolated points. We will specify $M$ later. In particular, for each $\epsilon$ small enough the following holds: \begin{itemize} \item[]\ \\ \item[i.]{\em For every} family of intervals $\{ J_i \}_i$ which covers $\Lambda_s$, with $|J_i| < \epsilon$ we have $$\frac{M}{2}\leq \sum_i |J_i|^{d_s}.$$ \item[ii.] {\em There exists } a family of intervals $\{ J_i \}_i$, with $|J_i|\leq \epsilon$, which covers $\Lambda_s$ and $$\sum_i |J_i|^{d_s} \leq 2M.$$ Furthermore we can assume that $\partial J_i \subset \Lambda_s$.\\ \end{itemize} Assume that $d_s\geq d/2$. By Lemma \ref{rci}, there exists some $K$ such that we can replace the special covering $\{ J_i\}$ in ii. by a new covering by dd-intervals $\{W_i^\ell\}_{i, \ \ell}$, with root cylinders $R_i^\ell$, where \begin{equation}\label{p1} J_i\cap \Lambda_s \subset \overline{ \bigcup_\ell W_i^\ell},\end{equation} \begin{equation}\label{p2} W_i^\ell:=\bigcup_k \overline{C^{i\ell}_k}, \ for \ each \ \ell \leq m_{i\ell} \leq 2N,\end{equation} \begin{equation}\label{p3} \frac{1}{K} \leq \frac{\sum_\ell |R^{\ell}_i|^{d_s}}{|J_i|^{d_s}}\leq K,\end{equation} \begin{equation}\label{p4} \frac{1}{K} \leq \frac{\sum_k |C^{i\ell}_k|^{d_s}}{|R^{\ell}_i|^{d_s}}\leq K,\end{equation} Indeed we can replace $W_i^\ell$ by a dd-subinterval of it, if necessary, in such way that $R_i^\ell \cap \Lambda_s \neq \phi$ and Eq. (\ref{p1}), Eq. (\ref{p2}), Eq. (\ref{p3}) and Eq. (\ref{p4}) hold, except perhaps the lower bound in Eq. (\ref{p3}). The above estimates, together to the fact that $\{ W^\ell_i \}$ covers $\Lambda_s$ (up to a countable set) gives \begin{equation} \label{eqsb}\frac{M}{2K^2}\leq \sum_{i,\ell,k} |C^{i\ell}_k|^{d_s} \leq 2K^2M.\end{equation} Since these intervals are cylinders, if necessary we can replace this family of cylinders by a subfamily of disjoint cylinders which covers $\Lambda_s$ up to a countable number of points and such that each cylinder intersects $\Lambda_s$. Indeed we can choose a finite subfamily $\mathcal{F}_s:=\{C_r\}_r$ satisfying \begin{equation} \label{eqsb1}\frac{M}{3K^2}\leq \sum_r |C_r|^{d_s} \leq 2K^2M.\end{equation} Let's call this finite subfamily $\mathcal{F}_s$. Note that, since $C_r\cap \Lambda_s\neq \emptyset$ we have that $$\sum_{t=0}^{\ell} \psi(f^t(x)) \geq 0$$ for every $x \in C_r$ and $\ell \leq \ell(C_r)$. Choose a very small cylinder $\tilde{C}$ such that $$\sum_{t=0}^{\ell} \psi(f^t(x)) \geq 0$$ for every $x \in \tilde{C}$ and $\ell < \ell(\tilde{C})$, and moreover satisfying $$\sum_{t=0}^{\ell(\tilde{C})} \psi(f^t(x)) > 0$$ on $\tilde{C}$, and \begin{equation} \label{neqsb}\frac{M}{3K^2}\leq |\tilde{C}|^{d_s} + \sum_r |C_r|^{d_s} \leq 3K^2M.\end{equation} Add $\tilde{C}$ to the family $\mathcal{F}_s$. Then, if $\mu_s$ is the geometric invariant measure of $f_{\mathcal{F}_s}$, we have $$\int \Psi_{\mathcal{F}_s} \ d\mu_s > 0.$$ And by Lemma \ref{vhd} and Eq. (\ref{neqsb}) $$| HD(\Lambda(f_{\mathcal{F}_s}))- d_s|\leq -\frac{C}{\log \epsilon}.$$ Since $\epsilon$ can be taken arbitrary, we can choose $\mathcal{F}_s$ such that $$HD(\Lambda(f_{\mathcal{F}_s}))\rightarrow_s d.$$ \end{proof} \begin{cor} If $F$ is a homogeneous random walk we have that $$HD(\Omega_+(F))= \lim_{\beta\rightarrow 0^+} HD(\Omega_{+\beta}(F))=\sup_{\beta > 0} HD(\Omega_{+\beta}(F)).$$ \end{cor} \begin{proof} Due Lemma \ref{aux}, it is enough to prove the Corollary for $k=0$. Of course $\Omega_{+\beta}^0(F)\subset \Omega_{+}^0(F)$ and $\beta_0 \leq \beta_1$ implies $\Omega_{+\beta_1}^0(F) \subset \Omega_{+\beta_0}^0(F)$, so $$\lim_{\beta\rightarrow 0^+} HD(\Omega_{+\beta}^0(F))=\sup_{\beta > 0} HD(\Omega_{+\beta}^0(F)) \leq HD(\Omega_{+}^0(F)).$$ To obtain the opposite inequality, let $\mathcal{F}_s$ be as in Proposition \ref{sim}. Denote $$\gamma_s := \int \Psi_{\mathcal{F}_s} \ d\mu_{\mathcal{F}_s}, \text{ and } W_n := \int N_{\mathcal{F}_s} \ d\mu_{\mathcal{I}}$$ and $\beta_s:= \gamma_s/W_s$. Then by the Birkhoff Ergodic Theorem there is subset $T_s \subset \Lambda(\mathcal{I}_n)$ such that $\mu_{\mathcal{F}_s}(T_s)=1$ and $$\lim_k \frac{1}{k} \sum_{i=0}^{k-1} \psi(f^i(x)) = \lim_k \frac{ \sum_{j=0}^{k-1} \Psi_{\mathcal{I}_n}(f^j_{\mathcal{F}_s}(x))}{ \sum_{j=0}^{k-1} N_{\mathcal{I}_n}(f^j_{\mathcal{F}_s}(x))}=\frac{\gamma_s}{W_s}=\beta_s > 0.$$ for every $x \in T_s$. Since the Hausdorff dimension of $\mu_{\mathcal{F}_s}$ is equal to $HD(\mathcal{F}_s)$, we have that $HD(T_s)=HD(\mathcal{F}_s)$. Note also that $$T_s \subset \Omega_{+\beta_s}^0,$$ which implies $HD(\mathcal{F}_s)\leq HD(\Omega_{+\beta_s}^0)$, so by the choice of $\mathcal{F}_s$, we conclude that $$ HD(\Omega_{+}^0)= \lim_s \ HD(\mathcal{F}_s) \leq \overline{\lim}_s \ HD(\Omega_{+\beta_s}^0)\leq \sup_{\beta > 0} \ HD(\Omega_{+\beta}^0).$$ \end{proof} \begin{proof}[{\bf Proof of Theorem \ref{multi}.}] Define $$\Gamma_n(F) := \{ x \in \Omega_{+\beta}^k(F) \ s.t. \ \pi_2(F^i(x,k))\geq \frac{\beta}{2} i, \ for \ all \ i\geq n\}.$$ Of course $$\Omega_{+\beta}^k(F) = \bigcup_n \Gamma_n(F).$$ To prove the Theorem, it is enough to verify that $HD (\Gamma_n(F))=HD (\Gamma_n(G))$. Indeed, for every $\epsilon > 0$ and $\alpha \in (HD(\Gamma_n(F)),1)$ there exists a covering of $\Gamma_n(F)$ by intervals $A_i$ so that $$\sum_j |A_j|^\alpha \leq \epsilon.$$ Note that we can assume that $\partial A_j \subset \Gamma_n(F)$. Since $G$ is an asymptotically small perturbation of $F$, it is easy to see that $G$ also satisfies the properties $Ra+Rb$, replacing the points $c^n_i$ and $d^n_i$ by $h(c^n_i)$ and $h(d^n_i)$, and modifying the constant . Indeed can choose constants in the definitions of the properties $Ex+BD+Ra+Rb$ which works for both random walks, so we can take $K > 0$ in the statements of Lemma \ref{rci} and Lemma \ref{rcia} in such way that it works for both random walks. In particular (as in the proof of Proposition \ref{sim}) for each $A_j$ we can find at most $2N$ dd-intervals $$W_j^\ell:=\bigcup_k \overline{C^{j\ell}_k}, \ with \ \ell \leq m_j \leq 2N$$ which satisfy $$A_i\cap \Gamma_n(F) \subset \overline{\bigcup_{\ell} W^\ell_i},$$ and $$\sum_{k, \ell} |C^{j \ell}_k|^\alpha\leq K |A_j|^\alpha.$$ Furthermore we can assume that the root $R_{j}^\ell$ of $W_j^\ell$ satisfies \begin{equation}\label{abc}\frac{1}{K} \leq \frac{|R_{j}^\ell|^\alpha}{\sum_{k} |C^{j \ell}_k|^\alpha}\leq K\end{equation} and $R^{j}_\ell \cap \Gamma_n(F)\neq \emptyset$. The constant $K$ does not depend on $\alpha$, $j$ or $\ell$. In particular the union of all cylinders $C^{j\ell}_k$ covers $\Gamma_n(F)$ up to a countable set and \begin{equation}\label{hdq} \sum_{j,k,\ell} |C^{j\ell}_k|^\alpha \leq K\epsilon.\end{equation} Note that if $x \in \Gamma_n(F)$ then $$dist_i(x)\leq r_n:=Cn + C\lambda^n$$ for every $i \in \mathbb{N}$. So $$e^{-r_n} \leq \frac{|\mathcal{P}^i_F(x)|}{|\mathcal{P}^i_G(h(x))|} \leq e^{r_n}.$$ There is a point in the cylinder $R_j^\ell$ which belongs to $\Gamma_n(F)$, so \begin{equation}\label{hdd} e^{-\alpha r_n}\leq \frac{|R_j^\ell|^\alpha}{|h(R_j^\ell)|^\alpha} \leq e^{\alpha r_n}.\end{equation} Note that $h(W_j^\ell)=\bigcup_k \overline{h(C^{j\ell}_k)}$ is a dd-interval for $G$ and $h(R_j^\ell)$ is its root cylinder. So \begin{equation}\label{fgh} \frac{1}{K}\leq \frac{|h(R_j^\ell)|^\alpha}{\sum_{i}|h(C^{j\ell}_i)|^\alpha} \leq K\end{equation} But the union of the cylinders $h(C^{j\ell}_k)$ covers $\Gamma_n(G)$ up to a countable set and Eq. (\ref{abc}), Eq. (\ref{hdq}), Eq. (\ref{hdd}) and Eq. (\ref{fgh}) gives $$\sum_{j,k,\ell} |h(C^{j\ell}_k)|^\alpha \leq K^3 e^{\alpha r_n}\epsilon.$$ Since $\alpha > HD(\Gamma_n(F))$ and $\epsilon$ is arbitrary we obtain that $HD(\Gamma_n(G)) \leq HD(\Gamma_n(F))$. Switching the roles of $F$ and $G$ in the above argument gives the opposite inequality. \end{proof} \begin{lem}\label{ol} Let $G \in On+Ra+Rb$ be a random walk. For every $\alpha > 0$ there exist $\epsilon$ and $C$ so that \begin{equation}\label{best} \sum_{P \in \mathcal{P}^n, \ P \subset I_k} |P|^{1-\epsilon} \leq C (1+ \alpha)^n,\end{equation} for all $n$ and $k$. \end{lem} \begin{proof}For a random walk $G$, denote by $\mathcal{P}^n:=\{P^n_i\}_i$ the Markov partition of $G^n$. Since $G \in BD+On+Ex$, for each $\delta > 0$, we can choose $n_0$ large enough so that for every inverse branch $\phi$ of an iteration of $G$ and an element $P \in \mathcal{P}^{n_0}$, we have \begin{equation}\label{peq} 1-\delta \leq \frac{|D\phi(x)|}{|D\phi(y)|} \leq 1+ \delta.\end{equation}for every $x,y \in P$. F Moreover note that for every $\epsilon < 1$ there exists a constant $K=K_ {\epsilon} >1$ so that \begin{equation}\label{bound} \sum_i |P^n_i|^{1-\epsilon} \leq K^n\end{equation} for every $n$. Denote $\mathcal{P}^{n_0}=\{ Q^j \}_j$ and $\mathcal{P}^{n_0+1}=\{ Q^j_k \}_{j,k}$, in such way that $Q_k^j \subset Q_j$. Indeed, since $G \in BD+Rb$, it is possible to order $Q_k^j$ so that there exists $C$ satisfying $$\frac{|Q^j_k|}{|Q^j|}\leq C\lambda^k,$$ for every $j,k$. As a consequence the set of functions $$h_j(\epsilon)= \sum_k \frac{|Q^j_k|^{1-\epsilon}}{|Q^j|^{1-\epsilon}}$$ is a equicontinuous set of functions in a small neighborhood of $0$. In particular, since $h_j(0)=1$, there exists $\epsilon$ so that, for every $j$, \begin{equation} \label{uni} \sum_k \frac{|Q^j_k|^{1-\epsilon}}{|Q^j|^{1-\epsilon}} \leq 1+\delta.\end{equation} For $n \leq n_0$, it follows from Eq. (\ref{bound}) that there exists $C$ so that, for $n\leq n_0$, we have $\sum_{P \in \mathcal{P}^n} |P|^{1-\epsilon} \leq C$. Assume by induction that we have proved Eq. (\ref{best}) until some $n\geq n_0$. Denote by $\{ \phi_j \}$ the inverse branches of $G^{n-n_0}$, with $Im \ \phi_i = P^{n-n_0}_i$. Then $\mathcal{P}^{n+1}= \{\phi_i(Q^j_k) \}_{i,j,k}$ and $\mathcal{P}^n =\{\phi_i(Q^j) \}_{i,j}$. By the distortion control in Eq. (\ref{peq}) and the estimative in Eq. (\ref{uni}), for each $i, j$ we have $$\frac{\sum_k |\phi_i(Q^j_k)|^{1-\epsilon}}{|\phi_i(Q^j)|^{1-\epsilon} } \leq \frac{1+\delta}{1-\delta} \frac{\sum_k |Q^j_k|^{1-\epsilon}}{|Q^j|^{1-\epsilon} } \leq \frac{(1+\delta)^2}{1-\delta}. $$ So $$\sum_{P \in \mathcal{P}^{n+1}} |P|^{1-\epsilon}= \sum_{i,j,k} |\phi_i(Q^j_k)|^{1-\epsilon} \leq \sum_{i,j} |\phi_i(Q^j)|^{1-\epsilon} \sum_k \frac{|\phi_i(Q^j_k)|^{1-\epsilon}}{|\phi_i(Q^j)|^{1-\epsilon}}$$ $$ \leq \frac{(1+\delta)^2}{1-\delta}\sum_{i,j} |\phi_i(Q^j)|^{1-\epsilon} = \frac{(1+\delta)^2}{1-\delta} \sum_{p \in \mathcal{P}^n} |P|^{1-\epsilon}.$$ We finish the proof choosing $\delta$ so that $(1+\delta)^2/(1-\delta) \leq (1+\alpha)$.\end{proof} From now on we are going to assume that the mean drift is negative: $\int \psi \ d\mu < 0$. \begin{lem}\label{bom} Let $G \in On+Ra+Rb$ be a random walk with negative mean drift. For every $\alpha > \int \psi \ d\mu$, there exists $\sigma < 1$ so that for any $n_1 \geq n_0$, with $n_0$ large enough, \begin{equation}\label{nzero} m\{p \in I_ {n_1}\colon \ \pi_2(G^k(p))\geq n_0, \text{ for } k\leq n, \text{ and } \pi_2(G^n(p)) - n_1 \geq \alpha n \}\leq \sigma^{n}.\end{equation} \end{lem} \begin{proof} Denote $$\Lambda_{n_0,n_1}^n(G):=\{p \in I_ {n_1}\colon \ \pi_2(G^k(p))\geq n_0 \text{ for all } k \leq n \text{ and } \pi_2(G^n(p)) - n_1 \geq \alpha n \}.$$ The statement for $F$ is consequence of the large deviations estimative (see, for instance \cite{broise}) $$m\{ p \in I\colon | \frac{\sum_{k=0}^{n-1} \psi(f^k(p))}{n} - \int \psi \ d\mu | \geq K \}\leq C_K\sigma^n,$$ which holds for every $K > 0$. In particular choosing $K= \alpha-\int \psi \ d\mu$ we get, for any $n_0$, and $n_1\geq n_0$, $$m\{p \in I_ {n_1}\colon \pi_2(F^n(p)) -n_1 \geq \alpha n \}\leq \sigma^{n}, $$ which implies (of course) \begin{equation}\label{unp} m(\Lambda_{n_0,n_1}^n(F))\leq \sigma^{n}. \end{equation} We are going to use this estimative to obtain Eq. (\ref{nzero}) for the perturbation of $F$. Indeed, for every $\delta > 0$, there is $n_0$ so that if $\pi_2(x)\geq n_0$ then \begin{equation}\label{distor} 1-\delta \leq \frac{|DF(x)|}{|DG(H(x))|} \leq 1+ \delta,\end{equation} Here $H$ is the topological conjugacy between $F$ and $G$ which preserves states. Note that $\Lambda_{n_0,n_1}^n(F)$ is a disjoint union of elements $Q_i \in \mathcal{P}^n(F)$, so $\Lambda_{n_0,n_1}^n(G)$ is a disjoint union of the intervals $H(Q_i)$. Due Eq. (\ref{unp}) and Eq. (\ref{distor}), we have \begin{equation} \sum_i |H(Q_i)|\leq \sum_i (1+\delta)^n |Q_i| \leq (1+\delta)^n \sigma^n.\end{equation} Take $n_0$ large enough so that $(1+\delta)\sigma < 1$. \end{proof} We would like to replace $n_0$ by an arbitrary state in Eq. (\ref{nzero}). The following Lemma will be useful for this task: \begin{lem}\label{absa} Let $p_n$ and $q_n$ sequences of non-negative real numbers such that \begin{enumerate} \item $p_0+ q_0 \leq 1$, \item There exists $\epsilon > 0$ and $\ell \in \mathbb{N}$ such that $s_n:= p_n + q_n \leq (1-\epsilon)^{\ell} p_{n-\ell} + q_{n-\ell}$ and $q_n \leq C\sigma^n + \sum_{k=1}^{n} (1-\epsilon)^kp_{n-k}$ , for every $n\geq 1$. \end{enumerate} Then there exists $\delta > 0$ such that $s_n \leq (1-\delta)^n$, for every $n \in \mathbb{N}$. \end{lem} \begin{proof} If $n \geq \ell$, we have $s_n \leq (1-\epsilon)p_{n-\ell}+ q_{n-\ell}= (1-\epsilon)s_{n-\ell} + \epsilon q_{n-\ell}$. It follows by induction that if $n= i\ell + r$, with $r < \ell$, then $$s_n \leq (1-\epsilon)^i s_r + \sum_{k=0}^{i-1}\epsilon(1-\epsilon)^{k\ell} q_{n-(k+1)\ell}$$ $$\leq C(1-\epsilon)^{n/\ell} s_0 + \sum_{k=0}^{n-1}\epsilon(1-\epsilon)^{k} q_{n-\ell -k}$$ Since $q_{n-\ell} \leq C(1-\epsilon)^n + \sum_{k=1}^{n-1} (1-\epsilon)^kp_{n-\ell-k}$, we obtain $$s_n\leq (1-\epsilon)^{n/\ell} s_0 + \epsilon (1-\epsilon)^{n/\ell} + \sum_{k=1}^{n-1} \epsilon(1-\epsilon)^{k}(p_{n-\ell-k} + q_{n-\ell-k})$$ $$\leq (1-\epsilon)^{n/\ell}C(s_0 + \epsilon)+ \sum_{k=1}^{n-1} \epsilon(1-\epsilon)^{k} s_{n-\ell-k},$$ for every $n\geq \ell$. We claim that there exists $\delta < 1$ and $K$ so that $s_n \leq K(1-\delta)^n$, for every $n$. Indeed, fix $\delta < 1$, For each $n$, define $K_n := s_n/(1-\delta)^n$. Note that \begin{equation} \label{final} s_n \leq (1-\epsilon)^{n/\ell}C(s_0 + \epsilon)+ \sum_{k=1}^{n-1} \epsilon(1-\epsilon)^{k} s_{n-\ell-k}$$ $$ \leq (1-\epsilon)^{n/\ell}C(s_0 + \epsilon)+ \sum_{k=1}^{n-1} \epsilon(1-\epsilon)^{k} K_{n-\ell-k}(1-\delta)^{n-\ell-k}\end{equation} $$\leq \big[ \big( \frac{(1-\epsilon)^{1/\ell}}{1-\delta} \big)^n C(s_0 + \epsilon)+ \max_{i< \ n-\ell}K_i \ \frac{\epsilon}{(1-\delta)^\ell} \sum_{k=1}^{n-1} \big(\frac{1-\epsilon}{1-\delta}\big)^k \big](1-\delta)^n$$ Choose $\delta$ close enough to $1$ so that $$\sigma_1:= \frac{(1-\epsilon)^{1/\ell}}{1-\delta} < 1, \ and $$ $$\sigma_2:= \frac{\epsilon}{(1-\delta)^\ell} \sum_{k=1}^{\infty} \big(\frac{1-\epsilon}{1-\delta}\big)^k < 1.$$ Then by Eq. (\ref{final}) we have $K_{n} \leq \sigma_2\max_{i< \ n-\ell}K_i + C\sigma_1^n$, for every $n > \ell$, which easily implies that $\max_i K_i < \infty$. \end{proof} Define $$\Omega_{+}^{n_1,n}:= \{p \in I_{n_1}\colon \pi_2(G^{k}(p)) \geq 0, \text{ for } 0 \leq k \leq n \}.$$ \begin{lem}\label{estimativa} There exists $\delta < 1$ so that for every $n_1\geq 0$ there exists $C=C(n_1)$ satisfying $$m(\Omega_{+}^{n_1,n}(G))\leq C(1-\delta)^n.$$ \end{lem} \begin{proof}Take $n_0$ as in Lemma \ref{bom} and fix $n_1 \geq 0$. Define the sets and sequences $$s_n:= m(\Omega_{+}^{n_1,n}) $$ $$p_n:= m(B^n), \text{ where } B^n:=\{ p \in \Omega_{+}^{n_1,n}\colon \ \pi_2(G^{n}(p))\in [0,n_0] \}, \ and $$ $$q_n:= m(C^n), \text{ where } C^n:=\{p \in \Omega_{+}^{n_1,n}\colon \ \pi_2(G^{n}(p)) > n_0 \}.$$ To prove Lemma \ref{estimativa}, it is enough to verify that these sequences satisfy the assumptions of Lemma \ref{absa}. Indeed, of course $p_0 + q_0 \leq 1$. To prove the other assumptions, take $i \in [0,n_0]$. Since $G$ is topologically transitive, there are $\ell_i \in \mathbb{N}$ and intervals $J_i \subset I_i$ so that $\pi_2(G^{\ell_i}(J_i)) < 0$. Denote $\ell=max_{\ 0\leq i\leq n_0} \ell_i$ and $r= min_{\ 0\leq i\leq n_0} |J_i|/|I_i|$. Clearly $\Omega^{n_1,n}_{+}(G)=B^n\cup C^n \subset B^{n-\ell}\cup C^{n-\ell} $. Let $J \subset B^{n-\ell}$ be an interval so that $G^{n-\ell}(J)=I_i$, with $0\leq i\leq n_0$. Note that $B^{n-\ell}$ is a disjoint union of such intervals. By the bounded distortion control for $G$, \begin{equation}\label{ind} \frac{m(J\cap \Omega_{+}^{n_1,n}(G))}{m(J)} \leq 1- \frac{m(J\cap G^{-(n-\ell)}J_i)}{m(J)}\leq (1-\frac{r}{c})\end{equation} Choose $\epsilon_0$ satisfying $(1-r/c)\leq (1-\epsilon_0)^\ell$. Then Eq. (\ref{ind}) implies $$m(B^{n-\ell}\cap \Omega_{+}^{n_1,n}(G))\leq (1-\epsilon_0)^\ell m(B^{n-\ell})$$ and we obtain $$s_n = m(B^{n-\ell}\cap \Omega_{+}^{n_1,n}(G)) + m(C^{n-\ell}\cap \Omega_{+}^{n_1,n}(G)) \leq (1-\epsilon_0)^\ell p_{n-\ell} + q_{n-\ell}.$$ It remains to prove that $q_n \leq \sum_{k=1}^{n} (1-\epsilon)^kp_{n-k}$. There are two kind of points $p$ in $C^n$: {\em Type 1.} For every $j\leq n$ we have $\pi_2(G^j(p))\geq n_0$ (in particular $n_1\geq n_0$). We are going to estimate the measure of the set of these points, denoted $\Theta_1^n$. It follows from Lemma \ref{bom}, choosing, for instance, $\alpha=\int \psi \ d\mu/2$, that \begin{equation}\label{acima} m(\{p \in I_{n_1}\colon \ \pi_2(G^k(p))\geq n_0, \ for \ k\leq n \text{ and } \pi_2(G^n(p)) \geq n_1 + \alpha n \}) \leq C\sigma^n.\end{equation} But the set in the r.h.s. of Eq. (\ref{acima}) coincides with $\Theta^{n}_{1}$ provided $n\geq~(n_0-n_1)/\alpha$. So $$m(\Theta_1^n)\leq C_{n_1}\sigma^n,$$ for some $\sigma < 1$ which does not depend on $n_1$. {\em Type 2.} For some $j < n$ we have $\pi_2(G^j(p)) \leq n_0$. Denote the set of these points by $\Theta^n_2$. Denote by $\Theta_{2,k}^n$ the set of points $p$ so that $k\geq 1$ is the smallest natural satisfying $\pi_2(G^{n-k}p)\leq n_0$. Clearly $\Theta^n_2$ is a disjoint union of these sets. We are going to estimate their measure. Note that $\Theta_{2,k}^n \subset B^{n-k}$. The set $B^{n-k}$ is a disjoint union of intervals $L$ so that $\pi_2(G^{n-k}L)=I_i$, for some $i \leq n_0$. To estimate $$\frac{m(\Theta_{2,k}^n\cap L)}{|L|} $$ note that $L \subset B^{n-k},$ and $\Theta_{2,k}^n\cap L$ is the set of points $p \in L$ so that $\pi_2(G^{n-k+j}p) > n_0$, for every $0< j \leq k$. Define $$L_y := \{ p \in L\colon \psi(G^{n-k}p)=y\}.$$ Firstly note that for $y \leq n_0 -i$ we have \begin{equation} \label{estum} |L_y\cap \Theta_{2,k}^n|=0,\end{equation} since $p \in L_y\cap \Theta_{2,k}^n$ satisfies $\pi_2(G^{n-k+1}p)= i+ \psi(G^{n-k}p)= i + y > n_0$. In particular for $y < 0$ we have $|L_y\cap \Theta_{2,k}^n|=0$, which implies, due the bounded distortion control $$\frac{m(L\cap \Theta_{2,k}^n)}{|L|}\leq \frac{\sum_{y \geq 0} |L_y|}{|L|} \leq (1-\delta),$$ for some $\delta < 1$ which does not depends on $k$, $L$ or $n_1$, which implies \begin{equation}\label{estquatro} m(\Theta^n_{2,k}) \leq (1-\delta)m(B^{n-k})= (1-\delta)p_{n-k}.\end{equation} Furthermore, using again the distortion control and the regularity condition $GD$(big jumps are rare) we have \begin{equation}\label{estdois} \frac{\sum_{y > -\alpha (k-1)} |L_y \cap \Theta_{2,k}^n|}{|L|}\leq \frac{\sum_{y > -\alpha (k-1)} |L_y|}{|L|}\leq C\gamma^{k},\end{equation} for some $C\geq 0$ and $\gamma < 1$. To estimate $|L_y\cap \Theta_{2,k}^n|/|L_y|$, in the case $n_0-i \leq y \leq -\alpha (k-1)$, recall that $G^{n-k+1}L_y=I_{i+y}$, with $i+y > n_0$. By Lemma \ref{bom}, we have $$ m\{p \in I_ {i+y}\colon \ \pi_2(G^m(p))\geq n_0, \text{ for } m\leq k-1, \text{ and } \pi_2(G^{k-1}(p)) \geq i+y + \alpha (k-1) \}\leq C\sigma^{k}.$$ Since $i+y+ \alpha (k-1) \leq n_0$, this implies that $$ m\{p \in I_ {i+y}\colon \ \pi_2(G^m(p))\geq n_0, \text{ for every } m\leq k-1 \}\leq C\sigma^{k}.$$ The points in $L_y\cap\Theta^n_{2,k}$ are exactly the points whose $(n-k+1)$th-iteration belongs to the set in the estimative above. Using the bound distortion control we have $$ \frac{|L_y\cap \Theta^n_{2,k}| }{|L_y|} \leq C\sigma^k,$$ so \begin{equation}\label{esttres} \frac{|\sum_{n_0-i \leq y \leq -\alpha (k-1)}L_y\cap \Theta^n_{2,k}| }{|L|} \leq C \frac{|\sum_{n_0-i \leq y \leq -\alpha (k-1)}L_y\cap \Theta^n_{2,k}| }{\sum_{n_0-i \leq y \leq -\alpha (k-1)} |L_y|} \leq C\sigma^k. \end{equation} Choose $\epsilon < \epsilon_0$ so that $ min\{ max\{ C\sigma^k, C\gamma^k\}, 1-\delta\} \leq (1-\epsilon)^k$, for every $k \geq 0$, and put together Eq. (\ref{estum}), Eq. (\ref{estquatro}), Eq. (\ref{estdois}) and Eq. (\ref{esttres}), to get $m(L\cap \Theta_{2,k}^n)\leq (1-\epsilon)^k |L|$. Since $B^{n-k}$ is a disjoint union of such intervals $L$, we obtain $$m(\Theta_{2,k}^n)\leq (1-\epsilon)^k m(B^{n-k})= (1-\epsilon)^kp_{n-k}$$ and now we can conclude with $$q_n = m(\Theta^n_{1}) + \sum_{k} m(\Theta^n_{2,k}) \leq C_{n_1}\sigma^n + \sum_k (1-\epsilon)^kp_{n-k}.$$ \end{proof} Now we are ready to prove Theorem \ref{menor}: \begin{proof}[{\bf Proof of Theorem \ref{menor}.}] There are three cases: {\bf $F$ is transient with $ M > 0$.} If $M > 0$ then the random walk $F$ is transient and it is easy to see that $m(\Omega_+(F)) > 0$. Since the conjugacy with an asymptotically small perturbation $G$ is absolutely continuous (Theorem \ref{abscont}), we conclude that $m(\Omega_+(G)) > 0$. {\bf $F$ is recurrent ($ M = 0$).} if $M=0$ then $F$ and its asymptotically small perturbations are recurrent by Theorem \ref{strec}. In particular almost every point visits negative states infinitely many times, so $m(\Omega_+(G)) = 0$. It remains to prove that $HD \ \Omega_+(G)=1$. By Theorem \ref{omega} it is enough to verify that $HD \ \Omega_+(F)=1$. Indeed, it is easy to show using the Central Limit Theorem that if $$\int \psi \ d\mu=0$$ then there exist $C > 0$ and and for each $n$, subsets $\mathcal{A}_n \subset \mathcal{P}^n$ so that $$\sum_{i=0}^{n-1}\psi(f^i(x)) > 0$$ for all $x \in J \in \mathcal{A}_n$ and $$m(\bigcup_{J \in \mathcal{A}_n} J) \geq C >0.$$ here $C$ does not depend on $n$. Replacing $\mathcal{A}_n$ by a finite subfamily, if necessary, we can apply Proposition \ref{vhd} to obtain $$HD \ \Lambda(\mathcal{A}_n) = 1 - O(\frac{1}{n}).$$ If $\mu_{\mathcal{A}_n }$ is the geometric invariant measure of $f_{\mathcal{A}_n}$ then $$\int \psi_{\mathcal{A}_n} \ d\mu_{\mathcal{A}_n} > 0$$ So by the Birkhoff Ergodic Theorem \begin{equation} \label{conv1} \lim_{n\rightarrow \infty} \sum_{i=0}^{n-1}\psi(f^i(x)) = + \infty \end{equation} in a set $S_n \subset \Lambda(\mathcal{A}_n)$ satisfying $\mu_{\mathcal{A}_n}(S_n)=1$, so $HD \ S_n = 1 - O(1/n)$. In particular the set $S$ of points satisfying Eq.(\ref{conv1}) has Hausdorff dimension $1$. We can decompose $S$ in subsets $B_j$ defined by $$B_j :=\{ x \in S\colon {\min}_{n} \sum_{i=0}^{n-1} \psi(f^i(x)) \geq -j \}.$$ Clearly $\sup_j HD \ B_j=1$. For each $j$ choose $k_j$ and $J_j \in \mathcal{P}^{k_j}$ so that for all $x \in J_j$ we have $$\sum_{i=0}^{\ell-1}\psi(f^i(x)) \geq 0$$ for every $\ell\leq k_j$ and $$\sum_{i=0}^{k_j} \psi(f^i(x)) \geq j.$$ Then $$(J_j \cap f^{-k_j}B_j)\times \{0\}$$ belongs to $\Omega_+(F)$, for every $j$. This implies $HD \ \Omega_+(F) \geq HD \ B_j$ so $$ HD \ \Omega_+(F) \geq \sup_j HD \ B_j =1.$$ {\bf $F$ is transient with $ M < 0$.} By Lemma \ref{estimativa}, there is some $\delta \in (0,1)$, which does not depend on $n_1$, so that \begin{equation} \label{useum} m(\Omega^{n_1,n})\leq C(1-\delta)^n.\end{equation} By Lemma \ref{ol}, there exists $\epsilon$ so that \begin{equation}\label{usedois} \sum_{P \in \mathcal{P}^n, \ P \subset I_k} |P|^{1-\epsilon} \leq C(1-\delta)^{-n/2}.\end{equation} Denote by $\{J_i^n\}_i \subset \mathcal{P}^n$ the family of disjoint intervals so that $\Omega^{n_1,n} = \cup_i J_i^n$. We claim that there exists $C > 0$ satisfying \begin{equation} \sum_i |J_i^n|^{1-\epsilon/4} \leq C(1-\delta)^n.\end{equation} Since $sup_i \ |J^n_i|\rightarrow_n 0$, this proves that $HD \ \Omega_+^{n_1,\infty} \leq 1-\epsilon/4$. Indeed, $$\sum_i |J_i^n|^{1-\epsilon/4} = \sum_{ |J_i|> (1-\delta)^{2n/\epsilon}} |J_i^n|^{1-\epsilon/4} + \sum_{ |J_i|\leq (1-\delta)^{2n/\epsilon}} |J_i^n|^{1-\epsilon/4}$$ $$\leq (1-\delta)^{-n/2} \sum_{i} |J_i^n| + (1-\delta)^{3n/2} \sum_{i} |J_i^n|^{1-\epsilon}$$ $$\leq C(1-\delta)^{n/2},$$ where in the last line we made use of Eq. (\ref{useum}) and Eq. (\ref{usedois}). The proof is complete. \end{proof} \section{Applications to one-dimensional renormalization theory} \subsection{(Classic) infinitely renormalizable maps}\label{apl} Consider a real analytic unimodal maps $f\colon I \rightarrow I$, with negative Schwarzian derivative and even order critical point. The map $f$ is called infinitely renormalizable if there exists an sequence of natural numbers $n_0 < n_1 < n_2 < \dots$ and a nested sequence of intervals $$I=I_0 \supset I_1 \supset I_2 \supset \cdots $$ so that \begin{itemize} \item$f^{n_k}\partial I_k \subset \partial I_k$, \item $f^{n_k}I_k \subset I_k$, \item $f^{n_k}\colon I_k \rightarrow I_k$ is a unimodal map. \end{itemize} We say that $f$ has bounded combinatorics if there exists $C >0$ so that $n_{k+1}/n_k \leq C$, for all $k$. Two infinitely renormalizable maps $f$ and $g$ have the same combinatorics if there exists a homeomorphism $h\colon I \rightarrow I$ such that $f\circ h = h \circ g$. The following result is a deep result in renormalization theory: \begin{prop}[\cite{mc2}]\label{convren} Let $f$ and $g$ be two infinitely renormalizable unimodal maps with the same bounded combinatorics and same even order. Then for every $r > 0$ there exists $C > 0$ and $\lambda < 1$ so that $$||\frac{1}{|I_k^f|}\ f^{n_k}(|I_k^f| \cdot ) - \frac{1}{|I_k^g|}\ g^{n_k}(|I_k^g| \cdot )||_{C^r} \leq C\lambda^k. $$ \end{prop} Here $|I^f_k|$ denotes the length of $I_k^f$. \begin{figure} \centering \psfrag{f}{$f$} \psfrag{g}{$f^2$} \psfrag{h}{$f^4$} \psfrag{p1}[][]{$p_1$} \psfrag{p1l}[][]{$p_1'$} \psfrag{p2}[][]{$p_{2}$} \psfrag{p2l}[][]{$p_{2}'$} \psfrag{p3}[][]{$p_{3}$} \psfrag{p3l}[][]{$p_{3}'$} \includegraphics[width=0.70\textwidth]{figure2.eps} \caption{The "Bat" map: the induced map $F$ for a Feigenbaum unimodal map} \end{figure} \begin{proof}[{\bf Proof of Theorem \ref{apl1}.}] Let $f$ be an infinitely renormalizable map with bounded combinatorics. We are going to define an induced map $F\colon I\rightarrow I$, following Y. Jiang (see \cite{jianga}, \cite{jiangb}): Let $p_k$ be the periodic point in $\partial I_k$. Define $E$ as the set $$\{1,-1,-p_k, p_k,f(p_k), -f(p_k),\dots, f^{n_k-1}(p_k),-f^{n_k-1}(p_k)\}-\{f(p_k),-f(p_k) \}.$$ The set $E$ cuts $I_{k-1}\setminus I_k$ in $m_k$ intervals. Denote these intervals $M_{k-1,i}$, with $i=1,\dots,m_k$. For each $x \in M_{k-1,i}$, define $n(x)\geq 1$ as the minimal positive integer so that $$I_k\subset f^{n(x)n_{k-1}} M_{k-1,1}.$$ Note that $f^{n(x)n_{k-1}}$ does not have critical points on $ M_{k-1,i}$. Define the induced map $F$, which is defined everywhere in $I$, except for a countable set of points: $$F(x):=f^{n(x)}(x), \ for \ x \in I_k\setminus I_{k+1}.$$ See in Fig. 2 the induced map for an infinitely renormalizable maps satisfying $n_{i+1}=2n_i$ for all $i$ (the so called Feigenbaum maps). The map $F$ is Markovian with respect to the partition $$\mathcal{P}:=\{ M_{k,i} \}_{k \in \mathbb{N},i \leq m_k}.$$ Furthermore, if $f$ and $g$ have the same bounded combinatorics and even order, then by Proposition \ref{convren}, the corresponding induced maps $F$ and $G$ satisfies $$||\ \frac{1}{|I_k^f|}\ F(|M_{k,i}^f| \cdot + |I_k^f|-|M_{k,i}^f|)- \frac{1}{|I_k^g|}\ G(|M_{k,i}^g| \cdot + |I_k^g|-|M_{k,i}^g|)\ ||_{C^r([0,1])} \leq C\lambda^k.$$ Define $L_k$ as, say, the right component of $I_k\setminus I_{k+1}$ and $\gamma_k \colon I \rightarrow L_k$ as the unique bijective order preserving affine map between this two intervals. We are going to define a random walk $\mathcal{F} \colon I \times \mathbb{N} \rightarrow I \times \mathbb{N}$ from the map $F$ in the following way: $$ \mathcal{F}(x,k):= \begin{cases} (\gamma^{-1}_i \circ F\circ \gamma_k (x),i) & \text{if $F\circ \gamma_k (x) \in L_i$;}\\ (\gamma^{-1}_i \circ (-F)\circ \gamma_k (x),i) & \text{if $F\circ \gamma_k (x) \in -L_i$.}\\ \end{cases} $$ It is easy to see that we can extend $\mathcal{F}\colon I \times \mathbb{Z} \rightarrow I \times \mathbb{Z}$ to a strongly transient deterministic random walk with non-negative drift. Furthermore if $g$ is another infinitely renormalizable map with the same bounded combinatorics that $f$ then by Proposition \ref{convren} and Proposition \ref{how} the corresponding random walk $\mathcal{G}$ is an asymptotically small perturbation of $\mathcal{F}$. So we can apply Theorem \ref{sqr} to conclude that there is a conjugacy between $F$ and $G$ which is strongly quasisymmetric with respect to the nested sequence of partitions defined by the random walk $\mathcal{F}$. We can now easily translate this result in terms of the original unimodal maps $f$ and $g$ saying that the continuous conjugacy $h$ between $f$ and $g$ is a strongly quasisymmetric mapping with respect to $\mathcal{P}$. \end{proof} \begin{rem}\label{quotient}{\rm An interesting case is when the unimodal map $f$ is a periodic point to the renormalization operator: there exists $n_0$ and $\lambda$, with $|\lambda|< 1$ so that $$\frac{1}{\lambda}f^{n_0}(\lambda x)=f(x).$$ In this case, if we take $n_k=kn_0$, then the induced map $F$ will satisfy the functional equation \begin{equation}\label{fe} F(\lambda x)=\lambda F(x).\end{equation} Define the relation $\sim$ in the following way: $$x\sim y \text{ iff there exists $i \in \mathbb{Z}$ so that } x=\pm \lambda^i y.$$ By Eq. (\ref{fe}), $F$ preserves this relation, so we can take the quotient of $F$ by the relation $\sim$. Note that $$L_0 = \mathbb{R}^\star/\sim.$$ It is easy to see that if $q=F/\sim\colon L_0\rightarrow L_0$ is a Markov expanding map. Now define $\psi\colon L_0 \rightarrow \mathbb{Z}$ as $\psi(x)= k$, if $f(x) \in I_{k}\setminus I_{k+1}$. Then $\mathcal{F}$ is exactly the homogeneous random walk defined by the pair $(q,\psi)$. } \end{rem} \subsection{Fibonacci maps} \label{aplf} The Fibonacci renormalization is the simplest way to generalize the concept of classical renormalization as described in Section \ref{apl}. Actually we could prove all the results stated for Fibonacci maps to a wider class of maps: maps which are infinitely renormalizable in the generalized sense and with periodic combinatorics and bounded geometry, but we will keep ourselves in the simplest case to avoid more technical definitions and auxiliary results with its long proofs. Consider the class of real analytic maps $f$ with $Sf < 0$ and defined in a disjoint union of intervals $I^0_1 \sqcup I^1_1$, where $-I_1^0=I_1^0$, so that \begin{itemize} \item[ ]\ \item[{\it -}] The map $f\colon I^1_1 \rightarrow I^0_0:=f(I^1_1)$ is a diffeomorphism. Furthermore $I^1_1$ is compactly contained in $I^0_0$.\\ \item[{\it -}] The map $f\colon I^0_1 \rightarrow I^0_0$ is an even map which has as $0$ as its unique critical point of even order.\\ \end{itemize} We say that $f$ is {\bf Fibonacci renormalizable} if $$f(0) \in I_1^1, \ f^2(0) \in I^1_0 \ and \ f^3(0) \in I^1_0.$$ In this case, the Fibonacci renormalization of $f$ is defined as the first return map to the interval $I_1^0$ restricted to the connected components of its domain which contain the points $f(0)$ and $f^2(0)$. This new map is denoted $\mathcal{R}f$: it could be Fibonacci renormalizable again and so on, obtaining an infinite sequence of renormalizations $\mathcal{R}f$, $\mathcal{R}^2f$, $\mathcal{R}^3f$, $\dots$. We will denote the set of infinitely renormalizable maps in the Fibonacci sense with a critical point of order $d$ by $\mathcal{F}_d$. A map $f \in \mathcal{F}_d$ will be called a {\bf Fibonacci map}. As in the original map $f$, the $n$-th renormalization $f_n:= \mathcal{R}^nf$ of $f$ is a map defined in two disjoint intervals, denoted $I^n_0$ and $I_n^1$, where $-I^n_0=I^n_0$. Indeed $f_n$ on $I_n^0$ is a unimodal restriction of the $S_n$-th iteration of $f$, where $\{ S_n \}$ is the Fibonacci sequence $$S_0 = 1, \ S_1=2, \ S_2=3, \ S_3= 5, \ \dots \ , S_{k+2} = S_{k+1} + S_k, \dots$$ and $f_n$ on $I_n^1$ is the restriction of the $S_{n-1}$-th iteration of $f$. \begin{figure} \centering \psfrag{f}{$f$} \psfrag{un}[][][0.8]{$u_n$} \psfrag{unl}[][][0.8]{$u_n'$} \psfrag{un1}[][][0.8]{$u_{n+1}$} \psfrag{un1l}[][][0.8]{$u_{n+1}'$} \psfrag{unm1}[][][0.8]{$u_{n-1}$} \psfrag{unm1l}[][][0.8]{$u_{n-1}'$} \psfrag{unm2}[][][0.8]{$u_{n-2}$} \psfrag{pn}[][][0.8]{$p_n$} \psfrag{pnl}[][][0.8]{$p_n'$} \psfrag{pn1}[][][0.8]{$p_{n+1}$} \psfrag{pn1l}[][][0.8]{$p_{n+1}'$} \includegraphics[width=\textwidth]{figure3.eps} \caption{On the left figure the (green) solid curves represents the part of the $f^{S_n}$ used in the definition of the induced map. On the right figure the (red) solid curve is the part of $f^{S_n}$ which coincides with the $n$-th Fibonacci renormalization on its central domain. } \end{figure} \begin{figure} \centering \psfrag{f}{$f$} \psfrag{un}[][][0.8]{$u_n$} \psfrag{unl}[][][0.8]{$u_n'$} \psfrag{un1}[][][0.8]{$u_{n+1}$} \psfrag{un1l}[][][0.8]{$u_{n+1}'$} \psfrag{unm1}[][][0.8]{$u_{n-1}$} \psfrag{unm1l}[][][0.8]{$u_{n-1}'$} \psfrag{unm2}[][][0.8]{$u_{n-2}$} \psfrag{pn}[][][0.8]{$p_n$} \psfrag{pnl}[][][0.8]{$p_n'$} \psfrag{pn1}[][][0.8]{$p_{n+1}$} \psfrag{pn1l}[][][0.8]{$p_{n+1}'$} \psfrag{pnm1}[][][0.8]{$p_{n-1}$} \includegraphics[width=\textwidth]{figure4.eps} \caption{The (red) curves inside the medium square is the graph of the $n$-th Fibonacci renormalization $f_n$. The (red and blue) curves inside the largest square is the graph of an extension of $f_n$ which has the same maximal invariant set. } \label{figure:extension} \end{figure} Denote by $p_k$ the sequence of points $p_k \in \partial I^k_0$ so that $$f_k(p_{k+1})=p_k$$ and denote $I^k_0=[p_k, p_k']$. It is possible to define a sequence $u_k$ of points satisfying \begin{itemize} \item[ ] \ \item[{\it 1.}] $\dots < \ p_{k+1} < u_k < p_k < \dots < p_0, $ \\ \item[{\it 2.}] $f^{S_k}$ is monotone on $[0,u_k]$, \\ \item[{\it 3.}] $f^{S_k}(u_{k+1})=u_k$, \\ \item[{\it 4.}] $f^{S_k}(u_k)=u_{k-2}$. \\ \end{itemize} We are going to define an induced map for an infinitely renormalizable map in the Fibonacci sense in the following way: Firstly, define $f_{-1}\colon I^0_0 \setminus I^1_0$ as an $C^3$ monotone extension of $f_0$ on $I^1_1$ which has negative Schwarzian derivative and bounded distortion. Define $F\colon I^0_0 \rightarrow \mathbb{R}$ as $$F(x) := f^{S_i}(x) \ \ if \ \ x \in [u_i, -u_i]\setminus [u_{i+1},-u_{i+1}]$$ for each $i\geq 0$. Define $L_i$ as, say, the right component of $[u_i, -u_i]\setminus [u_{i+1},-u_{i+1}]$ and $\gamma_i \colon I \rightarrow L_i$ as the unique bijective order preserving affine map between these two intervals. We are ready to define the random walk $\mathcal{F}\colon I \times \mathbb{Z} \rightarrow I \times \mathbb{Z}$ as $$ \mathcal{F}(x,k):= \begin{cases} (\gamma^{-1}_i \circ F\circ \gamma_k (x),i) & \text{if $F\circ \gamma_k (x) \in L_i$,}\\ (\gamma^{-1}_i \circ (-F)\circ \gamma_k (x),i) & \text{if $F\circ \gamma_k (x) \in -L_i$.}\\ \end{cases} $$ There is a very special Fibonacci map $f^\star$, called the Fibonacci fixed point (see, for instance \cite{smfib}), whose induced map $F^\star$ satisfies (choosing a good $u_0$) $$F^\star(\lambda x) = \pm \lambda F^\star(x)$$ for some $\lambda \in (0,1)$. In this case we can use the argument in Remark \ref{quotient} to conclude that $\mathcal{F}^\star$ is a homogeneous random walk. For an arbitrary Fibonacci map $f$, $\mathcal{F}$ is not homogeneous, however due Proposition \ref{how} and the following result $\mathcal{F}$ is an asymptotically small perturbation of $\mathcal{F}^\star$: \begin{prop}[see \cite{smfib}] For each even integer larger than two the following holds: for every Fibonacci map $f$, denote $$g_i = \alpha_{i}^{-1} \circ f^{S_i}\circ \alpha_{i+1}\colon I \rightarrow I,$$ where $\alpha_i\colon I \rightarrow [u_i^{f},-u_i^{f}]$ is an bijective affine map so that $\alpha_i^{-1}(f_{i+1}(0)) > 0$ and consider the correspondent maps $g_i^\star$ for $f^\star$. Then $$|| g_i - g_i^{\star}||_{C^r} \leq K_r\rho^{i}$$ for some $\rho < 1$ and every $r \in \mathbb{N}$. \end{prop} The {\bf real Julia set} of $f$, denoted $J_{\mathbb{R}}(f)$, is the maximal invariant of the map $$f\colon I_0^1 \sqcup I_1^1 \rightarrow I_0^0,$$ in other words, $$J_{\mathbb{R}}(f_j):= \cap_i f^{-i}_j I^j_0. $$ Denote $$\Omega_{+}^j(F):= \{ (x,i) \ s.t. \ \pi_2(F^n(x,i)) \geq j\, \ for \ all \ n\geq 0\}.$$ \begin{prop}\label{ji} There exists some $k_0$ so that $$ \Omega_{+}^{j+1}(F) \subset J_{\mathbb{R}}(f_j) \subset \Omega_{+}^{j-1}(F).$$ In particular \begin{equation}\label{esthaus} HD \ \Omega_{+}^{j+1}(F) \leq HD \ J_{\mathbb{R}}(f_j) \leq HD \ \Omega_{+}^{j-1}(F), \end{equation} and, for the Fibonacci fixed point, since $\Omega_{+}^{j+1}(F)$ is an affine copy of $\Omega_{+}^{j-1}(F)$ we have \begin{equation} HD \ \Omega_{+}^{j}(F) = HD \ J_{\mathbb{R}}(f). \end{equation} for all $j\geq 0$. \end{prop} \begin{proof} Denote by $F_\ell$ the restriction of $F$ to $\cup_{i\geq \ell}L_i$. Then the maximal invariant set of $F_\ell$ $$\Lambda(F_\ell):= \cap_{i \in \mathbb{N}} F^{-i}\mathbb{R}$$ is $\Omega_{+}^{\ell}(F)$. Consider the extension of $f_j$ described in Fig. (\ref{figure:extension}). Let's call this extension $\tilde{f}_j$. An easy analysis of its graph shows that $f_j$ and $\tilde{f}_j$ have the same maximal invariant set. We claim that $\tilde{f}_{j+1}$ is just a map induced by $\tilde{f}_{j}$. Indeed, the restriction of $\tilde{f}_{j+1}$ to $[u_{j+1},u_{j+1}']$ coincides with $\tilde{f}_j^2$ on the same interval. On the rest of $\tilde{f}_{j+1}$-domain $\tilde{f}_{j+1}$ coincides with $\tilde{f}_{j}$. \begin{figure} \centering \psfrag{f}{$f$} \psfrag{f2}[][][0.6]{$f_n$} \psfrag{f1}[][][0.6]{$f_{n+1}$} \psfrag{f0}[][][0.6]{$f_{n+2}$} \psfrag{f1p}[][][0.6]{$f_{n+3}$} \psfrag{f2p}[][][0.6]{$f_{n+4}$} \psfrag{2}[][][0.6]{$u_n$} \psfrag{1}[][][0.6]{$u_{n+1}$} \psfrag{0}[][][0.6]{$u_{n+2}$} \psfrag{1p}[][][0.6]{$u_{n+3}$} \psfrag{2p}[][][0.6]{$u_{n+4}$} \psfrag{2t}[][][0.6]{$u_n'$} \psfrag{1t}[][][0.6]{$u_{n+1}'$} \psfrag{0t}[][][0.6]{$u_{n+2}'$} \psfrag{1pt}[][][0.6]{$u_{n+3}'$} \psfrag{2pt}[][][0.6]{$u_{n+4}'$} \includegraphics[width=0.730\textwidth]{figure5.eps} \caption{Induced map $F$ for a Fibonacci map} \end{figure} By consequence, for $i \geq j$ the map $\tilde{f}_i$ is induced by $\tilde{f}_j$ and, since $F_{j+1}$ restricted to $L_i$ is equal to $\tilde{f}_i$, we obtain that $F_{j+1}$ is a map induced by $\tilde{f}_{j}$. In particular $$\Lambda(F_{j+1}) \subset \Lambda(\tilde{f}_j)=J_{\mathbb{R}}(f_j).$$ To prove that $\Lambda(\tilde{f}_j) \subset \Lambda(F_{j-1})$, we are going to prove that \begin{equation}\label{contido} x \in \Lambda(\tilde{f}_j) \ implies \ F_{j-1}(x) \in \Lambda(\tilde{f}_j).\end{equation} If $x$ belongs to the interval $I^j_1 \subset L_{j-1}$, where $\tilde{f}_j$ coincides with $F_{j-1}$, then $F_{j-1}(x) \in \Lambda(\tilde{f}_j)$. Otherwise $x \in I^j_0 \subset \cup_{i\geq j}L_i$, so $x \in \Lambda(\tilde{f}_j)\cap \ L_i$, for some $i\geq j$, then $F_{j-1}$ is an iteration of $\tilde{f}_j$ on $L_i$, so $F_{j-1}(x) \in \Lambda(\tilde{f}_j)$. This finishes the proof of Eq. (\ref{contido}). Since $\Lambda(\tilde{f}_j)$ is invariant by the action of $F_{j-1}$ we have $\Lambda(\tilde{f}_j) \subset \Lambda(F_{j-1})$.\end{proof} \begin{proof}[{\bf Proof of Theorem \ref{juliathm}}] Consider the homogeneous random walk $F^\star= (g,\psi)$ induced by $f^\star$. Denote $$M = \int \psi \ d\mu,$$ where $\mu$ is the absolutely continuous invariant measure of $g$. Using Thorem \ref{menor}, there are three cases: \vspace{4mm} {\bf 1. ${ \mathbf M < 0}$.} In this case $\mathbf F^\star$ is transient and we have that $HD \ \Omega_+(F) < 1$ for every asymptotically small perturbation of $F^\star$, in particular when $F$ is a random walk induced by a Fibonacci map $f$. By Proposition \ref{ji}, $HD \ J_{\mathbb{R}}(f) < 1$. \vspace{4mm} {\bf 2. $\mathbf M =0$.} In this case every asymptotically small perturbation $G$ of $F^\star$ is recurrent and $m(\Omega_+(G))=0$ but $HD \ \Omega_+(G)=1$. By Proposition \ref{ji} we obtain $m(J_\mathbb{R}(f)) =0$ and $HD \ J_\mathbb{R}(f)=1$. \vspace{4mm} {\bf 3. $\mathbf M > 0$.} In this case $\mathbf F^\star$ is transient with $m(\Omega_+(F^\star)) > 0$ and the conjugacy between $F^\star$ and any asymptotically small perturbation of it is absolutely continuous on $\Omega_+^i(F^\star)$. In particular $m(\Omega_+(F)) > 0$ for every random walk $F$ induced by a Fibonacci map $f$ so $m(J_\mathbb{R}(f)) > 0$ by Proposition \ref{ji}.\end{proof} A map $f\colon I \rightarrow I$ is called a unimodal map if $f$ has a unique critical point, with even order $d$, which is a maximum, and $f(\partial I) \subset \partial I$. We will assume that $f$ is real analytic, symmetric with respect the critical point and $Sf < 0$. If the critical value is high enough, then $f$ has a reversing fixed point $p$. Let $I_0^0:=[-p,p]$. Consider the map of first return $R$ to $f$: if $x \in I_0^0$ and $f^r(x) \in I_0^0$, but $f^n(x) \not\in I_0^0$ for $i < r$, define $$R(x):=f^r(x).$$ If there exists exactly two connected components $I_1^0$ and $I_1^1$ of the domain of $R$ containing points in the orbit of the critical point, and furthermore the map $$R\colon I_1^0 \cup I_1^1 \rightarrow I_0^0$$ is a Fibonacci map, then we will called $f$ an {\bf unimodal Fibonacci map}. The class of all unimodal Fibonacci maps will be denoted $\mathcal{F}^{uni}_d$. \begin{proof}[{\bf Proof of Theorem \ref{deep}}]We will use the notation in the proof of Theorem \ref{juliathm}. Since $m(J_{\mathbb{R}}(f)) > 0$, we conclude that the mean drift $M$ is positive. by Proposition \ref{prws} any asymptotically small perturbation $G$ of $\mathcal{F}^\star$ has the following property: there exists $\lambda \in [0,1)$, $C >0$ and $K > 0$ so that for every $P \in \mathcal{P}^0(G)$ $$m(p \in P \colon \ \sum_{i=0}^{n-1}\psi(G^i(p)) < K n )\leq C\lambda^n |P|.$$ This implies that $$m(p \in I_j \colon \ \sum_{i=0}^{\ell}\psi(G^i(p)) \geq K \ell \text{ for every } \ell\geq n )\geq (1-C\lambda^n).$$ so if $j= n|min \psi|$ we obtain $$m(I_j\cap \Omega_+^j(G))\geq 1-C\lambda^{C_1 j}.$$ here $c_1 > 0$. If $G$ is a random walk induced by a Fibonacci map $g$ then this implies that for $j$ large $$m(L_j\setminus J_{\mathbb{R}}(g))= m((-L_j)\setminus J_{\mathbb{R}}(g)) \leq C\lambda^{C_1 j}|L_j|.$$ Since $$[-u_{j+1},u_{j+1}]=\bigcup_{i\geq j} L_i\cup(-L_i),$$ we conclude that \begin{equation}\label{soalguns} m([u_{j+1},-u_{j+1}]\setminus J_{\mathbb{R}}(g)) \leq C\lambda^{c_1 j}|u_{j+1}|.\end{equation} For every $\delta$, choose $j$ so that $|u_{j+2}|\leq \delta \leq |u_{j+1}|$. Because $|u_{j+2}|> \theta |u_{j+1}| $, where $\theta \in (0,1)$ does not depend on $j$, we have that $|u_{j}|\geq C\theta^j$. Together with Eq. (\ref{soalguns}) this implies $$m([-\delta,\delta]\setminus J_{\mathbb{R}}(g)) \leq C\lambda^{C_1 j}|u_{j+1}|\leq C|u_{j+1}|^{1+ \alpha}\leq C|\delta|^{1+ \alpha}.$$\end{proof} \begin{proof}[\bf Proof of Theorem \ref{apl2}]We will prove each one of the following implications: {\bf (1) implies (2):} From the proof of Theorem \ref{juliathm}, if $m(J_{\mathbb{R}}(f))> 0$ for some $f \in \mathcal{F}_d$ the mean drift $M$ of the homogeneous random walk $\mathcal{F}^\star$ of $f^\star$ is positive. So $\mathcal{F}^\star$ (and all its asymptotically small perturbations) is transient (to $+\infty$). In terms of the original Fibonacci map $f$, this means that almost every orbit in $J_{\mathbb{R}}(f)$ accumulates in the post-critical set: So $f$ has a wild attractor. {\bf (2) implies (3):} if there exists a wild attractor for $f$ then $m(J_{\mathbb{R}}(f))> 0$. From the proof of Theorem \ref{juliathm} we obtain that the mean drift $M$ of $\mathcal{F}^\star$ is positive. So there exists a absolutely continuous conjugacy between $\mathcal{F}^\star$ and any asymptotically small perturbation of $\mathcal{F}^\star$. This implies that any two maps $f_1, f_2 \in \mathcal{F}_d$ admits a continuous and absolutely continuous conjugacy $$h\colon J_{\mathbb{R}}(f_1) \rightarrow J_{\mathbb{R}}(f_2).$$ Now consider two arbitrary maps $g_1, g_2 \in \mathcal{F}_d^{uni}$. Then we already know that there exists an absolutely continuous conjugacy $$h\colon J_{\mathbb{R}}(R_{g_1}) \rightarrow J_{\mathbb{R}}(R_{g_2})$$ between the induced Fibonacci maps $R_{g_1}$ and $R_{g_2}$ associated to $g_1$ and $g_2$. Of course $h$ is just the restriction of a topological conjugacy between $g_1$ and $g_2$. By a Block and Lyubich result (see, for instance, page 332 in \cite{ms}), every map of $\mathcal{F}_d^{uni}$ is ergodic with respect the Lebesgue measure. Since $g_1$ and $g_2$ have wild attractors, this implies that the orbit of almost every point $x \in I$ hits $J_{\mathbb{R}}(R_{g_1}$ at least once. Let $n(x)$ be a time when this happens. So consider a arbitrary measurable set $B \subset I$ so that $m(B)>0$. Then for at least one $n_0 \in \mathbb{N}$ the set $$B_{n_0}:=\{x \in B\colon \ n(x)=n_0 \}$$ has positive Lebesgue measure. This implies that $f^{n_0}B_{n_0}$ has positive Lebesgue measure, so $m(h(f^{n_0}B_{n_0})) > 0$. Now it is easy to conclude that $m(h(B_{n_0})$ and $h(B) > 0$. Switching the places of $g_1$ and $g_2$ in this argument we can conclude that $h$ is absolutely continuous on $I$. Finally note that the eigenvalues of the periodic points are not constant on the class $\mathcal{F}_d^{uni}$. {\bf (3) implies (4):} By the argument in Martens and de Melo \cite{mm}, if a Fibonacci map does not have a wild attractor then any continuous absolutely continuous conjugacy with other Fibonacci map is $C^1$: in particular the conjugacy preserves the eigenvalues of the periodic points. So if (3) holds then we can use the same argument in the proof of the previous implication to conclude that every Fibonacci map has a wild attractor. {\bf (4) implies (5):} The proof goes exactly as the proof of (2)$\Rightarrow$ (3). {\bf (5) implies (1):} The proof goes exactly as the proof of (3)$\Rightarrow$ (4). \end{proof}
{ "timestamp": "2010-01-12T17:48:24", "yymm": "0503", "arxiv_id": "math/0503736", "language": "en", "url": "https://arxiv.org/abs/math/0503736" }
\section{Introduction to Non-Commutative Worlds} Aspects of gauge theory, Hamiltonian mechanics and quantum mechanics arise naturally in the mathematics of a non-commutative framework for calculus and differential geometry. This paper consists in two sections. This first section sketches our results in this domain in general. The second section gives a derivation of a generalization of the Feynman-Dyson derivation of electromagnetism using our non-commutative context and using diagrammatic techniques. The first section is based on the paper \cite{NCW}. The second section is a new approach to issues in \cite{NCW}. \bigbreak Constructions are performed in a Lie algebra $\cal A.$ One may take $\cal A$ to be a specific matrix Lie algebra, or abstract Lie algebra. If $\cal A$ is taken to be an abstract Lie algebra, then it is convenient to use the universal enveloping algebra so that the Lie product can be expressed as a commutator. In making general constructions of operators satisfying certain relations, it is understood that one can always begin with a free algebra and make a quotient algebra where the relations are satisfied. \bigbreak On $\cal A,$ a variant of calculus is built by defining derivations as commutators (or more generally as Lie products). For a fixed $N$ in $\cal A$ one defines $$\nabla_N : \cal A \longrightarrow \cal A$$ by the formula $$\nabla_{N} F = [F, N] = FN - NF.$$ $\nabla_N$ is a derivation satisfying the Leibniz rule. $$\nabla_{N}(FG) = \nabla_{N}(F)G + F\nabla_{N}(G).$$ \bigbreak There are many motivations for replacing derivatives by commutators. If $f(x)$ denotes (say) a function of a real variable $x,$ and $\tilde{f}(x) = f(x+h)$ for a fixed increment $h,$ define the {\em discrete derivative} $Df$ by the formula $Df = (\tilde{f} - f)/h,$ and find that the Leibniz rule is not satisfied. One has the basic formula for the discrete derivative of a product: $$D(fg) = D(f)g + \tilde{f}D(g).$$ Correct this deviation from the Leibniz rule by introducing a new non-commutative operator $J$ with the property that $$fJ = J\tilde{f}.$$ Define a new discrete derivative in an extended non-commutative algebra by the formula $$\nabla(f) = JD(f).$$ It follows at once that $$\nabla(fg) = JD(f)g + J\tilde{f}D(g) = JD(f)g + fJD(g) = \nabla(f)g + f\nabla(g).$$ Note that $$\nabla(f) = (J\tilde{f} - Jf)/h = (fJ-Jf)/h = [f, J/h].$$ In the extended algebra, discrete derivatives are represented by commutators, and satisfy the Leibniz rule. One can regard discrete calculus as a subset of non-commutative calculus based on commutators. \bigbreak In $\cal A$ there are as many derivations as there are elements of the algebra, and these derivations behave quite wildly with respect to one another. If one takes the concept of {\em curvature} as the non-commutation of derivations, then $\cal A$ is a highly curved world indeed. Within $\cal A$ one can build a tame world of derivations that mimics the behaviour of flat coordinates in Euclidean space. The description of the structure of $\cal A$ with respect to these flat coordinates contains many of the equations and patterns of mathematical physics. \bigbreak \noindent The flat coordinates $X_i$ satisfy the equations below with the $P_j$ chosen to represent differentiation with respect to $X_j.$: $$[X_{i}, X_{j}] = 0$$ $$[P_{i},P_{j}]=0$$ $$[X_{i},P_{j}] = \delta_{ij}.$$ Derivatives are represented by commutators. $$\partial_{i}F = \partial F/\partial X_{i} = [F, P_{i}],$$ $$\hat{\partial_{i}}F = \partial F/\partial P_{i} = [X_{i},F].$$ Temporal derivative is represented by commutation with a special (Hamiltonian) element $H$ of the algebra: $$dF/dt = [F, H].$$ (For quantum mechanics, take $i\hbar dA/dt = [A, H].$) These non-commutative coordinates are the simplest flat set of coordinates for description of temporal phenomena in a non-commutative world. Note: \noindent {\bf Hamilton's Equations.} $$dP_{i}/dt = [P_{i}, H] = -[H, P_{i}] = -\partial H/\partial X_{i}$$ $$dX_{i}/dt = [X_{i}, H] = \partial H/\partial P_{i}.$$ These are exactly Hamilton's equations of motion. The pattern of Hamilton's equations is built into the system. \bigbreak \noindent {\bf Discrete Measurement.} Consider a time series $\{X, X', X'', \cdots \}$ with commuting scalar values. Let $$\dot{X} = \nabla X = JDX = J(X'-X)/\tau$$ where $\tau$ is an elementary time step (If $X$ denotes a times series value at time $t$, then $X'$ denotes the value of the series at time $t + \tau.$). The shift operator $J$ is defined by the equation $XJ = JX'$ where this refers to any point in the time series so that $X^{(n)}J = JX^{(n+1)}$ for any non-negative integer $n.$ Moving $J$ across a variable from left to right, corresponds to one tick of the clock. This discrete, non-commutative time derivative satisfies the Leibniz rule. \bigbreak This derivative $\nabla$ also fits a significant pattern of discrete observation. Consider the act of observing $X$ at a given time and the act of observing (or obtaining) $DX$ at a given time. Since $X$ and $X'$ are ingredients in computing $(X'-X)/\tau,$ the numerical value associated with $DX,$ it is necessary to let the clock tick once, Thus, if one first observe $X$ and then obtains $DX,$ the result is different (for the $X$ measurement) if one first obtains $DX,$ and then observes $X.$ In the second case, one finds the value $X'$ instead of the value $X,$ due to the tick of the clock. \bigbreak \begin{enumerate} \item Let $\dot{X}X$ denote the sequence: observe $X$, then obtain $\dot{X}.$ \item Let $X\dot{X}$ denote the sequence: obtain $\dot{X}$, then observe $X.$ \end{enumerate} \bigbreak The commutator $[X, \dot{X}]$ expresses the difference between these two orders of discrete measurement. In the simplest case, where the elements of the time series are commuting scalars, one has $$[X,\dot{X}] = X\dot{X} - \dot{X}X =J(X'-X)^{2}/\tau.$$ Thus one can interpret the equation $$[X,\dot{X}] = Jk$$ ($k$ a constant scalar) as $$(X'-X)^{2}/\tau = k.$$ This means that the process is a walk with spatial step $$\Delta = \pm \sqrt{k\tau}$$ where $k$ is a constant. In other words, one has the equation $$k = \Delta^{2}/\tau.$$ This is the diffusion constant for a Brownian walk. A walk with spatial step size $\Delta$ and time step $\tau$ will satisfy the commutator equation above exactly when the square of the spatial step divided by the time step remains constant. This shows that the diffusion constant of a Brownian process is a structural property of that process, independent of considerations of probability and continuum limits. \bigbreak \noindent {\bf Heisenberg/Schr\"{o}dinger Equation.} Here is how the Heisenberg form of Schr\"{o}dinger's equation fits in this context. Let the time shift operator be given by the equation $J=(1 + H\Delta t/i \hbar).$ Then the non-commutative version of the discrete time derivative is expressed by the commutator $$\nabla\psi = [\psi, J/\Delta t],$$ and we calculate $$\nabla \psi = \psi[(1 + H \Delta t/i \hbar)/\Delta t] - [(1 + H\Delta t/i \hbar)/\Delta t] \psi = [\psi, H]/i \hbar,$$ $$i \hbar \nabla \psi = [\psi, H].$$ This is exactly the Heisenberg version of the Schr\"{o}dinger equation. \bigbreak \noindent {\bf Dynamics and Gauge Theory.} One can take the general dynamical equation in the form $$dX_{i}/dt = {\cal G}_{i}$$ where $\{ {\cal G}_{1},\cdots, {\cal G}_{d} \}$ is a collection of elements of $\cal A.$ Write ${\cal G}_{i}$ relative to the flat coordinates via ${\cal G}_{i} = P_{i} - A_{i}.$ This is a definition of $A_{i}$ and $\partial F/\partial X_{i} = [F,P_{i}].$ The formalism of gauge theory appears naturally. In particular, if $$\nabla_{i}(F) = [F, {\cal G}_{i}],$$ then one has the curvature $$[\nabla_{i}, \nabla_{j}]F = [R_{ij}, F]$$ and $$R_{ij} = \partial_{i} A_{j} - \partial_{j} A_{i} + [A_{i}, A_{j}].$$ This is the well-known formula for the curvature of a gauge connection. Aspects of geometry arise naturally in this context, including the Levi-Civita connection (which is seen as a consequence of the Jacobi identity in an appropriate non-commutative world). \bigbreak One can consider the consequences of the commutator $[X_{i}, \dot{X_{j}}] = g_{ij}$, deriving that $$\ddot{X_{r}} = G_{r} + F_{rs}\dot{X^{s}} + \Gamma_{rst}\dot{X^{s}}\dot{X^{t}},$$ where $G_{r}$ is the analogue of a scalar field, $F_{rs}$ is the analogue of a gauge field and $\Gamma_{rst}$ is the Levi-Civita connection associated with $g_{ij}.$ This decompositon of the acceleration is uniquely determined by the given framework. \bigbreak One can use this context to revisit the Feynman-Dyson derivation of electromagnetism from commutator equations, showing that most of the derivation is independent of any choice of commutators, but highly dependent upon the choice of definitions of the derivatives involved. Without any assumptions about initial commutator equations, but taking the right (in some sense simplest) definitions of the derivatives one obtains a significant generalization of the result of Feynman-Dyson. \bigbreak \noindent {\bf Electromagnetic Theorem.} (See Section 2.) With the appropriate [see below] definitions of the operators, and taking $$\nabla^{2} = \partial_{1}^{2} + \partial_{2}^{2} + \partial_{3}^{2}, \,\,\, B = \dot{X} \times \dot{X} \,\,\, \mbox{and} \,\,\, E = \partial_{t}\dot{X}, \,\,\, \mbox{one has}$$ \begin{enumerate} \item $\ddot{X} = E + \dot{X} \times B$ \item $\nabla \bullet B = 0$ \item $\partial_{t}B + \nabla \times E = B \times B$ \item $\partial_{t}E - \nabla \times B = (\partial_{t}^{2} - \nabla^{2})\dot{X}$ \end{enumerate} \bigbreak The key to the proof of this Theorem is the definition of the time derivative. This definition is as follows $$\partial_{t}F = \dot{F} - \Sigma_{i}\dot{X_{i}}\partial_{i}(F) = \dot{F} - \Sigma_{i} \dot{X_{i}}[F, \dot{X_{i}}]$$ for all elements or vectors of elements $F.$ The definition creates a distinction between space and time in the non-commutative world. A calculation ( done diagrammatically in Figure 3) reveals that $$\ddot{X} = \partial_{t}\dot{X} + \dot{X} \times (\dot{X} \times \dot{X}).$$ This suggests taking $E = \partial_{t}\dot{X}$ as the electric field, and $B = \dot{X} \times \dot{X}$ as the magnetic field so that the Lorentz force law $$\ddot{X} = E + \dot{X} \times B$$ is satisfied. \bigbreak \noindent This result is applied to produce many discrete models of the Theorem. These models show that, just as the commutator $[X, \dot{X}] = Jk$ describes Brownian motion in one dimension, a generalization of electromagnetism describes the interaction of triples of time series in three dimensions. \bigbreak \noindent {\bf Remark.} While there is a large literature on non-commutative geometry, emanating from the idea of replacing a space by its ring of functions, work discussed herein is not written in that tradition. Non-commutative geometry does occur here, in the sense of geometry occuring in the context of non-commutative algebra. Derivations are represented by commutators. There are relationships between the present work and the traditional non-commutative geometry, but that is a subject for further exploration. In no way is this paper intended to be an introduction to that subject. The present summary is based on \cite{Kauff:KP,KN:QEM,KN:Dirac,KN:DG,Twist,NonCom,ST,Aspects,Boundaries,NCW} and the references cited therein. \bigbreak The following references in relation to non-commutative calculus are useful in comparing with the present approach \cite{Connes, Dimakis, Forgy, MH}. Much of the present work is the fruit of a long series of discussions with Pierre Noyes, influenced at critical points by Tom Etter and Keith Bowden. Paper \cite{Mont} also works with minimal coupling for the Feynman-Dyson derivation. The first remark about the minimal coupling occurs in the original paper by Dyson \cite{Dyson}, in the context of Poisson brackets. The paper \cite{Hughes} is worth reading as a companion to Dyson. It is the purpose of this summary to indicate how non-commutative calculus can be used in foundations. \bigbreak \section{Generalized Feynman Dyson Derivation} In this section we assume that specific time-varying coordinate elements $X_{1},X_{2},X_{3}$ of the algebra $\cal{A}$ are given. {\it We do not assume any commutation relations about $X_{1},X_{2},X_{3}.$} \bigbreak In this section we no longer avail ourselves of the commutation relations that are in back of the original Feynman-Dyson derivation. We do take the definitions of the derivations from that previous context. Surprisingly, the result is very similar to the one of Feynman and Dyson, as we shall see. \bigbreak Here $A \times B$ is the non-commutative vector cross product: $$(A \times B)_{k} = \Sigma_{i,j = 1}^{3} \epsilon_{ijk}A_{i}B_{j}.$$ (We will drop this summation sign for vector cross products from now on.) Then, with $B = \dot{X} \times \dot{X},$ we have $$B_{k} = \epsilon_{ijk}\dot{X_{i}}\dot{X_{j}} = (1/2)\epsilon_{ijk}[\dot{X_{i}},\dot{X_{j}}].$$ The epsilon tensor $\epsilon_{ijk}$ is defined for the indices $\{ i,j,k \}$ ranging from $1$ to $3,$ and is equal to $0$ if there is a repeated index and is ortherwise equal to the sign of the permutation of $123$ given by $ijk.$ We represent dot products and cross products in diagrammatic tensor notation as indicated in Figure 1 and Figure 2. In Figure 1 we indicate the epsilon tensor by a trivalent vertex. The indices of the tensor correspond to labels for the three edges that impinge on the vertex. The diagram is drawn in the plane, and is well-defined since the epsilon tensor is invariant under cyclic permutation of its indices. \bigbreak We will define the fields $E$ and $B$ by the equations $$B = \dot{X} \times \dot{X} \,\,\, \mbox{and} \,\,\, E = \partial_{t}\dot{X}.$$ We will see that $E$ and $B$ obey a generalization of the Maxwell Equations, and that this generalization describes specific discrete models. The reader should note that this means that a significant part of the {\it form} of electromagnetism is the consequence of choosing three coordinates of space, and the definitions of spatial and temporal derivatives with respect to them. The background process that is being described is otherwise aribitrary, and yet appears to obey physical laws once these choices are made. \bigbreak In this section we will use diagrammatic matrix methods to carry out the mathematics. In general, in a diagram for matrix or tensor composition, we sum over all indices labeling any edge in the diagram that has no free ends. Thus matrix multiplication corresponds to the connecting of edges between diagrams, and to the summation over common indices. With this interpretation of compositions, view the first identity in Figure 1. This is a fundmental identity about the epsilon, and corresponds to the following lemma. \bigbreak \begin{center} $$ \picill4inby4.2in(EpsilonIdentity) $$ { \bf Figure 1 - Epsilon Identity} \end{center} \bigbreak \noindent {\bf Lemma.} (View Figure 1) Let $\epsilon_{ijk}$ be the epsilon tensor taking values $0$, $1$ and $-1$ as follows: When $ijk$ is a permuation of $123$, then $\epsilon_{ijk}$ is equal to the sign of the permutation. When $ijk$ contains a repetition from $\{1,2,3 \},$ then the value of epsilon is zero. Then $\epsilon$ satisfies the following identity in terms of the Kronecker delta. \begin{center} $$ \picill4inby1in(LabeledEpsilonIdentity) $$ \end{center} \bigbreak $$\Sigma_{i} \,\epsilon_{abi}\epsilon_{cdi} = -\delta_{ad}\delta_{bc} + \delta_{ac}\delta_{bd}.$$ \bigbreak \noindent The proof of this identity is left to the reader. The identity itself will be referred to as the {\em epsilon identity}. The epsilon identity is a key structure in the work of this section, and indeed in all formulas involving the vector cross product. \bigbreak The reader should compare the formula in this Lemma with the diagrams in Figure 1. The first two diagram are two versions of the Lemma. In the third diagram the labels are capitalized and refer to vectors $A,B$ and $C.$ We then see that the epsilon identity becomes the formula $$A \times (B \times C) = (A \bullet C)B - (A \bullet B)C$$ for vectors in three-dimensional space (with commuting coordinates, and a generalization of this identity to our non-commutative context. Refer to Figure 2 for the diagrammatic definitions of dot and cross product of vectors. We take these definitions (with implicit order of multiplication) in the non-commutative context. \bigbreak \begin{center} $$ \picill4inby5.2in(DefiningDiff) $$ { \bf Figure 2 - Defining Derivatives} \end{center} \bigbreak \noindent {\bf Remarks on the Derivatives.} \begin{enumerate} \item Since we do not assume that $[X_{i}, \dot{X_{j}}] = \delta_{ij},$ nor do we assume $[X_{i},X_{j}]=0,$ it will not follow that $E$ and $B$ commute with the $X_{i}.$ \item We define $$\partial_{i}(F) = [F, \dot{X_{i}}],$$ and the reader should note that, these spatial derivations are no longer flat in the sense of section 1 (nor were they in the original Feynman-Dyson derivation). See Figure 2 for the diagrammatic version of this definition. \item We define $\partial_{t} = \partial/\partial t$ by the equation $$\partial_{t}F = \dot{F} - \Sigma_{i}\dot{X_{i}}\partial_{i}(F) = \dot{F} - \Sigma_{i} \dot{X_{i}}[F, \dot{X_{i}}]$$ for all elements or vectors of elements $F.$ We take this equation as the global definition of the temporal partial derivative, even for elements that are not commuting with the $X_{i}.$ This notion of temporal partial derivative $\partial_{t}$ is a least relation that we can write to describe the temporal relationship of an arbitrary non-commutative vector $F$ and the non-commutative coordinate vector $X.$ See Figure 2 for the diagrammatic version of this definition. \item In defining $$\partial_{t}F = \dot{F} - \Sigma_{i}\dot{X_{i}}\partial_{i}(F),$$ we are using the definition itself to obtain a notion of the variation of $F$ with respect to time. The definition itself creates a distinction between space and time in the non-commutative world. \item The reader will have no difficulty verifying the following formula: $$\partial_{t}(FG) = \partial_{t}(F)G + F\partial_{t}(G) + \Sigma_{i}\partial_{i}(F)\partial_{i}(G).$$ This formula shows that $\partial_{t}$ does not satisfy the Leibniz rule in our non-commutative context. This is true for the original Feynman-Dyson context, and for our generalization of it. All derivations in this theory that are defined directly as commutators do satisfy the Leibniz rule. Thus $\partial_{t}$ is an operator in our theory that does not have a representation as a commutator. \item We define divergence and curl by the equations $$\nabla \bullet B = \Sigma_{i=1}^{3} \partial_{i}(B_{i})$$ and $$(\nabla \times E)_{k} = \epsilon_{ijk}\partial_{i}(E_{j}).$$ See Figure 2 and Figure 4 for the diagrammatic versions of curl and divergence. \end{enumerate} \bigbreak Now view Figure 3. We see from this Figure that it follows directly from the definition of the time derivatives (as discussed above) that $$\ddot{X} = \partial_{t}\dot{X} + \dot{X} \times (\dot{X} \times \dot{X}).$$ This is our motivation for defining $$E = \partial_{t}\dot{X}$$ and $$B = \dot{X} \times \dot{X}.$$ With these definition in place we have $$\ddot{X} = E + \dot{X} \times B,$$ giving an analog of the Lorentz force law for this theory. \bigbreak Just for the record, look at the following algebraic calculation for this derivative: $$ \dot{F} = \partial_{t}F + \Sigma_{i} \dot{X_{i}}[F, \dot{X_{i}}]$$ $$ = \partial_{t}F + \Sigma_{i} (\dot{X_{i}}F \dot{X_{i}} - \dot{X_{i}} \dot{X_{i}} F)$$ $$ = \partial_{t}F + \Sigma_{i} (\dot{X_{i}}F \dot{X_{i}} - \dot{X_{i}} F_{i} \dot{X}) + \dot{X_{i}} F_{i} \dot{X} - \dot{X_{i}} \dot{X_{i}} F$$ Hence $$ \dot{F} = \partial_{t}F + \dot{X} \times F + (\dot{X} \bullet F) \dot{X} - (\dot{X} \bullet \dot{X}) F$$ (using the epsilon identity). Thus we have $$\ddot{X} = \partial_{t} \dot{X} + \dot{X} \times (\dot{X} \times \dot{X}) + (\dot{X} \bullet \dot{X}) \dot{X} - (\dot{X} \bullet \dot{X})\dot{X},$$ whence $$\ddot{X} = \partial_{t}\dot{X} + \dot{X} \times (\dot{X} \times \dot{X}).$$ \bigbreak In Figure 4, we give the derivation that $B$ has zero divergence. \begin{center} $$ \picill4inby6in(Xdoubledot) $$ { \bf Figure 3 - The Formula for Acceleration} \end{center} \bigbreak \begin{center} $$ \picill4inby5in(DivB) $$ { \bf Figure 4 - Divergence of $B$ } \end{center} \bigbreak Figures 5 and 6 compute derivatives of $B$ and the Curl of $E,$ culminating in the formula $$\partial_{t}B + \nabla \times E = B \times B.$$ In classical electromagnetism, there is no term $B \times B.$ This term is an artifact of our non-commutative context. In discrete models, as we shall see at the end of this section, there is no escaping the effects of this term. \bigbreak \begin{center} $$ \picill4inby5in(Bdot) $$ { \bf Figure 5 - Computing $\dot{B}$} \end{center} \bigbreak \begin{center} $$ \picill4inby6in(CurlE) $$ { \bf Figure 6 - Curl of $E$} \end{center} \bigbreak \begin{center} $$ \picill4inby6in(CurlB) $$ { \bf Figure 7 - Curl of $B$} \end{center} \bigbreak Finally, Figure 7 gives the diagrammatic proof that $$\partial_{t}E - \nabla \times B = (\partial_{t}^{2} - \nabla^{2})\dot{X}.$$ This completes the proof of the Theorem below. \bigbreak \noindent {\bf Electromagnetic Theorem} With the above definitions of the operators, and taking $$\nabla^{2} = \partial_{1}^{2} + \partial_{2}^{2} + \partial_{3}^{2}, \,\,\, B = \dot{X} \times \dot{X} \,\,\, \mbox{and} \,\,\, E = \partial_{t}\dot{X} \,\,\, \mbox{we have}$$ \begin{enumerate} \item $\ddot{X} = E + \dot{X} \times B$ \item $\nabla \bullet B = 0$ \item $\partial_{t}B + \nabla \times E = B \times B$ \item $\partial_{t}E - \nabla \times B = (\partial_{t}^{2} - \nabla^{2})\dot{X}$ \end{enumerate} \bigbreak \noindent {\bf Remark.} Note that this Theorem is a non-trivial generalization of the Feynman-Dyson derivation of electromagnetic equations. In the Feynman-Dyson case, one assumes that the commutation relations $$[X_{i}, X_{j}] = 0$$ and $$[X_{i}, \dot{X_{j}}] = \delta_{ij}$$ are given, {\em and} that the principle of commutativity is assumed, so that if $A$ and $B$ commute with the $X_{i}$ then $A$ and $B$ commute with each other. One then can interpret $\partial_{i}$ as a standard derivative with $\partial_{i}(X_{j}) = \delta_{ij}.$ Furthermore, one can verify that $E_{j}$ and $B_{j}$ both commute with the $X_{i}.$ From this it follows that $\partial_{t}(E)$ and $\partial_{t}(B)$ have standard intepretations and that $B \times B = 0.$ The above formulation of the Theorem adds the description of $E$ as $\partial_{t}(\dot{X}),$ a non-standard use of $\partial_{t}$ in the original context of Feyman-Dyson, where $\partial_{t}$ would only be defined for those $A$ that commute with $X_{i}.$ In the same vein, the last formula $\partial_{t}E - \nabla \times B = (\partial_{t}^{2} - \nabla^{2})\dot{X}$ gives a way to express the remaining Maxwell Equation in the Feynman-Dyson context. \bigbreak \noindent {\bf Remark.} Note the role played by the epsilon tensor $\epsilon_{ijk}$ throughout the construction of generalized electromagnetism in this section. The epsilon tensor is the structure constant for the Lie algebra of the rotation group $SO(3).$ If we replace the epsilon tensor by a structure constant $f_{ijk}$ for a Lie algebra ${\cal G}$of dimension $d$ such that the tensor is invariant under cyclic permutation ($f_{ijk} = f_{kij}$), then most of the work in this section will go over to that context. We would then have $d$ operator/variables $X_1, \cdots X_d$ and a generalized cross product defined on vectors of length $d$ by the equation $$(A \times B)_{k} = f_{ijk}A_{i}B_{j}.$$ The Jacobi identity for the Lie algebra ${\cal G}$ implies that this cross product will satisfy $$A \times (B \times C) = (A \times B) \times C + [B \times (A ] \times C)$$ where $$([B \times (A ] \times C)_{r} = f_{klr}f_{ijk}A_{i}B_{k}C_{j}.$$ This extension of the Jacobi identity holds as well for the case of non-commutative cross product defined by the epsilon tensor. It is therefore of interest to explore the structure of generalized non-commutative electromagnetism over other Lie algebras (in the above sense). This will be the subject of another paper. \bigbreak \subsection{Discrete Thoughts} In the hypotheses of the Electromagnetic Theorem, we are free to take any non-commutative world, and the Electromagnetic Theorem will satisfied in that world. For example, we can take each $X_{i}$ to be an arbitary time series of real or complex numbers, or bitstrings of zeroes and ones. The global time derivative is defined by $$\dot{F} = J(F' - F) = [F, J],$$ where $FJ = JF'.$ This is the non-commutative discrete context discussed in sections 1. We will write $$\dot{F} = J\Delta(F)$$ where $\Delta(F)$ denotes the classical discrete derivative $$\Delta(F) = F' -F.$$ With this interpretation $X$ is a vector with three real or complex coordinates at each time, and $$B = \dot{X} \times \dot{X} = J^{2}\Delta(X') \times \Delta(X)$$ while $$E = \ddot{X} - \dot{X} \times (\dot{X} \times \dot{X}) = J^{2}\Delta^{2}(X) - J^{3} \Delta(X'') \times ( \Delta(X') \times \Delta(X)).$$ Note how the non-commutative vector cross products are composed through time shifts in this context of temporal sequences of scalars. The advantage of the generalization now becomes apparent. We can create very simple models of generalized electromagnetism with only the simplest of discrete materials. In the case of the model in terms of triples of time series, the generalized electromagnetic theory is a theory of measurements of the time series whose key quantities are $$\Delta(X') \times \Delta(X)$$ and $$\Delta(X'') \times (\Delta(X') \times \Delta(X)).$$ \bigbreak It is worth noting the forms of the basic derivations in this model. We have, assuming that $F$ is a commuting scalar (or vector of scalars) and taking $\Delta_{i} = X_{i}' - X_{i},$ $$\partial_{i}(F) = [F, \dot{X_{i}}] =[F, J\Delta_{i}] = FJ\Delta_{i} - J\Delta_{i}F = J(F'\Delta_{i} - \Delta_{i}F) = \dot{F}\Delta_{i}$$ and for the temporal derivative we have $$\partial_{t}F = J[1 - J \Delta' \bullet \Delta]\Delta(F)$$ where $\Delta = (\Delta_{1}, \Delta_{2}, \Delta_{3}).$ \bigbreak \noindent {\bf Acknowledgement.} Most of this effort was sponsored by the Defense Advanced Research Projects Agency (DARPA) and Air Force Research Laboratory, Air Force Materiel Command, USAF, under agreement F30602-01-2-05022. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright annotations thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the Defense Advanced Research Projects Agency, the Air Force Research Laboratory, or the U.S. Government. (Copyright 2005.) It gives the author great pleasure to acknowledge support from NSF Grant DMS-0245588 and to thank Pierre Noyes and Keith Bowden for continuing conversations related to the contents of this paper. \bigbreak
{ "timestamp": "2005-04-03T20:31:16", "yymm": "0503", "arxiv_id": "quant-ph/0503198", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503198" }
\section{Introduction} Recently, numerous efforts are being made to describe the large-scale evolution of the Earth ecosystem. Due to the ecosystem's complexity, however, such a task is extremely difficult~(\cyt{pimm}). That is why researchers in this field have to turn to very simplified and abstract models that hopefully still contain relevant factors. Such an approach proved to be successful, e.g., in modelling of some aspects of extinction dynamics (\cyt{NEWMAN}). Indeed, in very simple models of ecosystems certain properties of extinctions as, e.g., the distribution of sizes or durations of extinctions, seem to agree, at least qualitatively, with palaeontological data (\cyt{BAKSNEPP}; \cyt{SOLE}). In these models, the dynamics spontaneously drives the ecosystem toward the scale-invariant state with extinctions described by some power-law characteristics. However, since the accuracy of fossil data is rather limited, especially with respect to events on a large timescale, the applicability of such models should be considered with care. The suggestion that the extinction dynamics is not scale invariant but it has a characteristic timescale was made by Raup and Sepkoski (\cyt{raupsep}). While analysing fossil data, they noticed that during the last 250 My mass extinctions on Earth appeared more or less cyclically with a period of approximately 26My. Although their analysis was initially questioned (\cyt{patterson}), some other works confirmed Raup and Sepkoski's hypothesis (\cyt{fox}; \cyt{prokoph}; \cyt{plotnick}). The suggested large periodicity of mass extinctions turned out to be very difficult to explain. Indeed, 26My does not seem to match any of known Earth cycles and some researchers have been looking for more exotic explanations involving astronomical effects (\cyt{theories1}; \cyt{theories2}), increased volcanic activity (\cyt{stot1}), or the Earth's magnetic field reversal (\cyt{stot2}). So far, however, none of these proposals has been confirmed. One should also note that the most recent analysis of palaeontological data that span last 542My strongly supports the periodicity of mass extinctions albeit with a larger cycle of about 62My (\cyt{rohde}). Lacking a firm evidence of an exogenous cause, one can ask whether the periodicity of extinctions can be explained without referring to such a factor. It is already well known that a periodic behaviour of a system is not necessarily the result of periodic driving. In particular, since the seminal works of Lotka and Volterra, it is known that spontaneous oscillations of the population size might appear in various prey-predator systems~ (\cyt{LV}). However, the period of oscillations in such systems is determined by the growth and death rate coefficients of interacting species and is of the order of a few years rather than tens of millions. Consequently, if the periodicity of mass extinctions is to be explained within a model of interacting species, a different mechanism that generates long-period oscillations must be at work. Recently, a multi-species prey-predator model has been introduced, where long-term oscillatory behaviour is observed (\cyt{lipowski}). Only some preliminary studies of basic properties of this model have been made, and the objective of the present paper is to provide its more detailed analysis. In this model the period of oscillations is determined by the inverse of the mutation rate and as we argue, it should be several orders of magnitude longer than in the Lotka-Volterra oscillations. The mechanism that generates oscillations in our model can be briefly described as follows: A coevolution of predator species induced by the competition for food and space causes a gradual increase of their size. However, such an increase leads to the overpopulation of large predators and a shortage of preys. It is then followed by a depletion of large species and a subsequent return to the multi-species stage with mainly small species that again gradually increase their size and the cycle repeats. Numerical calculations for our model show that the longevity of a species depends on the evolutionary stage at which the species is created. A similar pattern has been observed in some palaeontological data (\cyt{aimiller}) and, to our knowledge, the presented model is the first one that reproduces such a dependence. Let us notice that the oscillatory behaviour in a prey-predator system that was also attributed to the coevolution has been already examined by Dieckmann et al.~(\cyt{DIECKMANN}). In their model, however, the number of species is kept constant and it cannot be applied to study extinctions. Moreover, the idea that an internal ecosystem dynamics might be partially responsible for the long-term periodicity in the fossil records was suggested by Stanley~(\cyt{STANLEY}) and later examined by Plotnick and McKinney (\cyt{plotnick1993}). However, in his approach mass extinctions are triggered by external impacts. Their approximately equidistant separation is the result of a delayed recovery of the ecosystem. In our approach no external factor is needed to trigger such extinctions and sustain their approximate periodicity. \section{Model} Numerical simulations and models of various levels of description have been frequently used to study extinctions of species (\cyt{NEWMAN}). In the simplest cases, the dynamics of models was formulated at the level of species and had to refer to the notion of fitness that is not commonly accepted. In more recent approaches, an individual-oriented dynamics has often been used and although computationally more demanding, such models are considered as more adequate (\cyt{higgs}; \cyt{stauffer}; \cyt{fdl}). Our model uses the individual-oriented dynamics but in addition it is spatially extended. Such a feature increases the computational complexity even more but it also takes into account, e.g., dynamically generated spatial inhomogeneities that sometimes are known to play an important role. Our model can be also considered as a multi-species generalization of the already studied spatially extended prey-predator model (\cyt{lip1999}; \cyt{lip2000}). Some other multi-species lattice models were also studied in various contexts (\cyt{pekal}; \cyt{SATO}; \cyt{DIECKMANN2000}). Our model describes a multi-species prey-predator system defined on a square lattice of linear size $N$ (\cyt{lipowski}). At each site of a lattice $i$ there is an operator $x_i$ that specifies whether this site is occupied by a prey ($x_i=1$), by a predator ($x_i=2$), by both of them ($x_i=3$), or is empty ($x_i=0$). Each predator is characterized by its size $m \ (0<m<1)$ that determines its consumption rate and at the same time its strength when it competes with other predators. Only approximately the size $m$ can be considered as related with physical size. Predators and preys evolve according to rules typical to such systems (e.g., predators must eat preys to survive, preys and predators can breed provided that there is an empty site nearby, etc.). In addition, the relative update rate for preys and predators is specified by the parameter $r \ (0<r<1)$ and during breeding mutations are taking place with the probability $p$. More detailed definition of the model dynamics is given below:\\ (a) Choose a site at random (the chosen site is denoted by $i$).\\ (b) Provided that $i$ is occupied by a prey (i.e., if $x_i=1$ or $x_i=3$) update the prey with the probability $r$. If at least one neighbor (say $j$) of the chosen site is not occupied by a prey (i.e., $x_j=0$ or $x_j=2$), the prey at the site $i$ produces an offspring and places it on an empty neighboring site (if there are more empty sites, one of them is chosen randomly). Otherwise (i.e., if there are no empty sites) the prey does not breed.\\ (c) Provided that $i$ is occupied by a predator (i.e., $x_i=2$ or $x_i=3$) update the predator with the probability $(1-r)m_i$, where $m_i$ is the size of the predator at site $i$. If the chosen site $i$ is occupied by a predator only ($x_i=2$), it dies, i.e., the site becomes empty ($x_i=0$). If there is also a prey there ($x_i=3$), the predator consumes the prey (i.e., $x_i$ is set to 2) and if possible, it places an offspring at an empty neighboring site. For a predator of the size $m_i$ it is possible to place an offspring at the site $j$ provided that $j$ is not occupied by a predator ($x_j=0$ or $x_j=1$) or is occupied by a predator ($x_j=2$ or $x_j=3$) but of a smaller size than $m_i$ (in such a case the smaller-size predator is replaced by an offspring of the larger-size predator). The offspring inherits its parent's size with the probability $1-p$ and with the probability $p$ it gets a new size that is drawn from a uniform distribution.\\ At first sight one can think that such a model describes an ecosystem with two trophic levels (preys and predators) and only with predators being equipped with evolutionary abilities, which would be of course highly unrealistic. Let us notice, however, that expansion of predators sometimes proceeds at the expense of smaller-size predators. Thus, predators themselves are involved in prey-predator-like interactions. Perhaps it would be more appropriate to consider unmutable preys as a renewable (at a finite rate) source of, e.g., energy, and predators as actual species involved in various prey-predator interactions and equipped with evolutionary abilities. \section{Results} To examine the behaviour of this model we used numerical simulations. Our results, shown in Figs.\ref{okna}-\ref{spsize}, are obtained for $r=0.2$ but we expect (\cyt{lipowski}) that the behaviour of the model should be qualitatively the same for any $r<0.27$ (a brief discussion of the behaviour of the model for $r>0.27$ is given at the end of this section). \subsection{Oscillatory behaviour} In Fig.\ref{okna}B one can see that, indeed, the number of species $s$ exhibits pronounced irregular oscillations. These oscillations are coupled with more regular oscillations of the averaged (over all predators) size $m_a$ (Fig.\ref{okna}A) and maxima of $s$ correspond approximately to minima of $m_a$ and vice versa. To have a better understanding of the behaviour of the model we also calculated the size $m_d$ of the dominant species (i.e., the predator species with the largest number of individuals) and the results are shown in Fig.\ref{okna}A. \begin{figure} \vspace{-0cm} \centerline{ \epsfxsize=9cm \epsfbox{okna_jtb.eps} } \caption{The results of numerical simulations ($N=500$,\ p=0.00001). Data on both panels are obtained from the same run. In our simulations a unit of time is defined as a single on average update of each site. (A) The time dependence of the average size $m_a$ (dashed line) and the size of the dominant species $m_d$ (short, horizontal intervals). (B) The number of species $s$ (continuous line) and the averaged lifetime of species (+). After extinction the number of species drops 3-4 times. To reduce stochastic noise in the calculation of the average lifetime data are collected in time windows of the width $\Delta t = 3000$.} \label{okna} \end{figure} These results indicate that the behaviour of our model can be described as follows: In a species-rich interval the size $m_a$ is typically quite low and there is an abundance of preys. In such a case predators of a large size are in a more favorable position (because a larger predator can replace a smaller predator) and as a result $m_a$ and $m_d$ increase. The process of increasing the size is gradual and involves a large number of species and is not related to a creation of a single (very-efficient) species, as suggested previously (\cyt{lipowski}). The increased size $m$ implies a higher consumption rate and due to a finite recovery rate of preys the large-size species, that at this stage dominate the system, are running out of food. At first sight one might expect that further evolution will gradually reduce $m_a$ and $m_d$. Numerical results show, however (Fig.\ref{okna}A), that after reaching a local maximum, $m_d$ jumps to a very low value. This indicates that abrupt changes take place in the model after which large-size species are no longer dominant and vast majority of them become extinct. At the same time, however, a lot of new, mainly small-size species is created and that increases the diversity $s$ (although we do not suggest that this was really the cause, a succession of small mammals after large dinosaurs could be a vivid example of such a change). In such a way the system returns to the initial species-rich state. Such a cycle is also illustrated in Fig.\ref{sizes} that shows the distribution of size $m$ at various stages of the evolution\footnote{ Dynamics of the model is also illustrated with a Java applet available at: http:// spin.amu.edu.pl/lipowski/prey\_pred.html}. \begin{figure} \vspace{0cm} \centerline{ \epsfxsize=9cm \epsfbox{sizes_jtb.eps} } \caption{Four panels show distribution of sizes of species at various stages of evolutionary cycle ($N=1000$,\ p=0.0001). The upper panel shows the time dependence of the number of species $s$.} \label{sizes} \end{figure} Gradual increase of size of species recalls the Cope's rule that states that species tend to increase body size over geological time. This rule is not commonly accepted among paleontologists and evolutionists and was questioned on various grounds (\cyt{STANLEY1973}). However, recent studies of fossil records of mammal species are consistent with this rule (\cyt{ALROY}; \cyt{VAN}). Perhaps our model could suggests a way to obtain a theoretical justification of this rule. From the above description, it is expected that the periodicity of such a cycle increases when the mutation rate $p$ decreases, and such a behaviour is confirmed with more detailed calculations (\cyt{lipowski}). In particular, already for $p=10^{-5}$ the estimated (\cyt{lipowski}) periodicity of oscillations in our model is approximately 1000 times larger than that of the Lotka-Volterra oscillations in the corresponding single-predator system. It shows that the oscillations in our model are indeed long-period and, perhaps for smaller $p$, on a timescale close to 26My. Although very complicated, in principle, it should be possible to estimate the value of the mutation probability $p$ from the mutational properties of living species. Let us notice that in our model mutations produce an individual that might be substantially different from its parent. In Nature, this is typically the result of many cumulative mutations and thus we expect that $p$ is indeed a very small quantity. Actually, $p$ should be considered rather as a parameter related with the speed of morphological and speciation processes that are known to be typically very slow (\cyt{gingerich}). Perhaps a modification of the mutation mechanism where a new species will be only a small modification of its parental species could be more suitable for comparison with living species, but it might require longer calculations. Alternatively, one can try to estimate the parameters $p$, $N$, and $r$ (or at least their ratios) by matching the behaviour of our model with some characteristics of the ecosystem such as the period of oscillations (26My), fraction of extinct species during a mass extinction or the average lifetime of species as compared with the periodicity of mass extinctions. The oscillatory behaviour sets probably the largest timescale in our model. However, on the shorter timescale some characteristics, such as, e.g., the number of species, exhibit strong fluctuations (Fig.~\ref{okna}B). On such a timescale some distributions might be very broad and resemble power-law distributions. Indeed, such a behaviour was demonstrated for the distribution of lifetimes of species in our model (\cyt{lipowski}). The increase of the size of species in our model resembles the fitness-increasing evolution in the real ecosystem. It is tempting to consider present-day large mammals as highly adapted dominant species and, in the context of our model, located perhaps close to the local maximum in the fitness space (as in Fig.\ref{okna}A). If so, then according to our model, the next dominant species most likely will be a small-size species that at the moment might not even exist. Its dominance will be possible due to drastic and inevitable changes of our ecosystem. Putting aside the validity of our model, such a scenario does not seem unlikely. \subsection{Longevity of species} An analysis of palaeontological data (\cyt{aimiller}) shows that the longevity is larger for species created after mass extinctions than for other species. To compare such a result with the predictions of our model, we calculated the average lifetime of species. It turns out, however, that important contributions to this quantity are coming from short-living species and their lifetime is essentially independent on the evolutionary phase at which they are born. To reduce this effect we took into account only the species that lived longer than a given threshold, which we set equal to 30. Fossils of species of short lifetime are rather scarce and palaeontological data also reflect a similar bias toward long-lifetime species. The obtained results are shown in Fig.\ref{okna}B. Although still strongly fluctuating, they clearly show that the lifetime is correlated with the global evolution of the ecosystem and they qualitatively agree with palaeontological data. In particular, the maximum lifetime appears for species born shortly after a large and abrupt decrease of the size of the dominant species (crash). Apparently, species created at this time find most favourable conditions while the worst conditions exist shortly before a crash. Again using the analogy with the real ecosystem and humans, the model predicts (not counter-intuitively) that species created during our dominance will have a rather short lifetime. \begin{figure} \vspace{0cm} \centerline{ \epsfxsize=9cm\epsfbox{rate_tau_jtb.eps} } \caption{The average lifetime of species $\tau$ as a function of the size $m$ ($N=500$).} \label{rate_tau.eps} \end{figure} For a species to have a very small size $m$ is usually a disadvantage since such a species will loose in competition with other species. On the other hand, a large size implies a high consumption rate and such a species might suffer from lack of food. It means that a lifetime of a species as a function of $m$ should have a maximum at a certain intermediate value and numerical calculations confirm such a behaviour (see Fig~\ref{rate_tau.eps}). Some data on distribution of sizes in Pleistocene and Recent molluscan faunas do show some maximum (\cyt{jablonsky}) but a more detailed comparison cannot be done yet. As our last result, we present the calculation of the average population size of species of a given lifetime (Fig.\ref{spsize}). Although all the curves look qualitatively similar, one can notice a small difference between short- and long-lifetime species. This difference is better seen on the rescaled plot (Fig.\ref{spsize}B). This data suggest that population sizes for species of a lifetime much shorter than the periodicity of extinctions (which in this case (\cyt{lipowski}) is around 3000) after rescaling fall into a single curve. For species of a lifetime comparable or larger than the periodicity of extinctions the data will deviate from such a universal curve. Although we are not aware of any palaeontological data of this kind, a comparison could provide an interesting test of our model. \begin{figure} \vspace{0cm} \centerline{ \epsfxsize=9cm \epsfbox{spsize_jtb.eps}} \caption{The analysis of the time dependence of the population size. (A) The average population size of species with a given lifetime ($p=0.001$, $N=500$). (B) Some data from panel (A) rescaled (i.e., multiplied by some factors in both directions) in such a way that the lifetime and the maximal population size overlap. For species with the lifetime equal to 100, 150, and 200, the rescaled population sizes nearly overlap. Some deviations from the overlapping data can be seen for the lifetime 1000 and 2000.} \label{spsize} \end{figure} \subsection{Unique code and the emergence of a multi-species ecosystem} All living cells use the same code that is responsible for the transcription of information from DNA to proteins (\cyt{orgel}; \cyt{szathmary}). It suggests that at a certain point of evolution of life on Earth a replicator that invented this apparently effective mechanism was able to eliminate replicators of all other species (if they existed) and establish, at least for a short time, a single-species ecosystem. Although this process is still to a large extent mysterious, one expects that subsequent evolution of these successful replicators leads to their differentiation and proliferation of species. In such a way the ecosystem shifted from a single- to multi-species one (\cyt{lipowski2000}). It seems to us that the present model might provide some insight into this problem. As we have already mentioned, the oscillatory behaviour appears in our model only for the relative update rate $r<0.27$. When preys reproduce faster ($r>0.27$), a different behaviour can be seen (\cyt{lipowski}) and the model reaches a steady state with almost all predators belonging to the same species with the size $m$ close to 1. Only from time to time a new species is created with even larger $m$ and a change of the dominant species might take place. In our opinion, it is possible that at the very early period of evolution of life on Earth, the ecosystem resembled the case $r>0.27$. This is because at that time substrates ('preys') were renewable faster than primitive replicators ('predators') could use them. If so, every invention of the increase of the efficiency ('size') could invade the entire system. In particular, the invention of the coding mechanism could spread over the entire system. A further evolution increased the efficiency of predators and that effectively shifted the (single-species) ecosystem toward the $r<0.27$ (multi-species, oscillatory) regime.\\ \par \section{Conclusions} In the present paper we examined a spatially extended multi-species prey-predator model. In a certain regime in this model densities of preys and predators as well as the number of species show long-term oscillations, even though the dynamics of the model is not exposed to any external periodic forcing. It suggests that the oscillatory behaviour of the Earth ecosystem predicted by Raup and Sepkoski could be simply a natural feature of its dynamics and not the result of an external factor. Some predictions of our model such as the lifetime of species or the time dependence of their population sizes might be testable against palaeontological data. The prediction that a lifetime of species depends on the evolutionary stage at which it was created, that qualitatively agrees with fossil data ~(\cyt{aimiller}), suggests that a further study of this model would be desirable. Certainly, our model is based on some restrictive assumptions that drastically simplifies the complexity of the real ecosystem. We hope, however, that it includes some of its important ingredients: replication, mutation, and competition for resources (food and space). As an outcome, the model shows that typically there is no equilibrium-like solution and the ecosystem remains in an evolutionary cycle. The model does not include geographical barriers but let us notice that palaeontological data that suggest the periodicity of mass extinctions are based only on marine fossils (\cyt{rohde}). More realistic versions should take into account additional trophic levels, gradual mutations, or sexual reproduction. One should also notice that the palaeontological data are mainly at a genus, and not species level. It would be desirable to check whether the behaviour of our model is in some sense generic or it is merely a consequence of its specific assumptions. An interesting possibility in this respect could be to recast our model in terms of Lotka-Volterra like equations and use the methodology of adaptive dynamics developed by Dieckmann et al. (\cyt{DIECKMANN}). Of course, the real ecosystem was and is exposed to a number of external factors such impacts of astronomical objects, volcanism or climate changes. Certainly, they affect the dynamics of an ecosystem and contribute to the stochasticity of fossil data. Filtering out these factors and checking whether the main evolutionary rhythm is indeed set by the ecosystem itself, as suggested in the present paper, is certainly a difficult task but maybe worth an effort. \vspace{5mm} \noindent {\bf Acknowledgements}\\ The research grant 1 P03B 014 27 from KBN is gratefully acknowledged. We thank Department of Physics of the University of Aveiro (Portugal) for giving us access to computing facilities.\\ \vspace{5mm} \begin{center} REFERENCES \end{center}
{ "timestamp": "2006-01-06T21:14:37", "yymm": "0503", "arxiv_id": "q-bio/0503020", "language": "en", "url": "https://arxiv.org/abs/q-bio/0503020" }
\section{Quarkonia and heavy flavors: what is different at the LHC} \label{widatl} With a nucleus-nucleus center-of-mass energy nearly 30 times larger than the one reached at RHIC, the LHC will open a new era for studying the properties of strongly interacting matter at extreme energy densities~\cite{Carminati:2004fp}. One of the most exciting aspects of this new regime is the abundant production rate of hard probes which can be used, for the first time, as high statistics probes of the medium~\cite{Bedjidian:2003gd}. Futhermore, heavy flavor measurements at the LHC should provide a comprehensive understanding of open and hidden heavy flavor production at very low $x$ values, where strong nuclear gluon shadowing is expected. The heavy flavor sector at LHC energies is subject to other significant differences with respect to SPS and RHIC energies. First, the large production rate offers the possibility to use a large variety of observables. Then, the magnitude of most of the in-medium effects is dramatically enhanced. Some of these aspects are discussed hereafter. \subsection{New observables} The Table~\ref{qqbar} shows the number of $c\bar{c}$ and $b\bar{b}$ pairs produced in central A-A collisions at SPS, RHIC and LHC. From RHIC to LHC, there are 10 times more $c\bar{c}$ pairs and 100 times more $b\bar{b}$ pairs produced. Therefore, while at SPS only charmonium states are experimentally accessible and at RHIC it remains to be seen how much of the bottom sector can be explored, at the LHC both charmonia and bottomonia can be used, thus providing powerful probes for Quark Gluon Plasma (QGP) studies. In fact, since the $\Upsilon(1S)$ state only dissolves significantly above the critical temperature~\cite{Digal:2001ue}, at a value which might only be reachable above that of RHIC, the spectroscopy of the $\Upsilon$ family at the LHC should reveal unique characteristics of the QGP~\cite{Gunion:1996qc}. In addition to the centrality dependence of the $\Upsilon$ yield, the study of the $\Upsilon^\prime/\Upsilon$ ratio versus transverse momentum ($p_{\rm T}$) is believed to be of crucial interest~\cite{Gunion:1996qc} (see below). \begin{table}[ht] \centering \caption{Number of $c\bar{c}$ and $b\bar{b}$ pairs produced in central heavy-ion collisions ($b=0$) at SPS (Pb-Pb), RHIC (Au-Au), and LHC (Pb-Pb) energies. $b\bar{b}$ production is negligible at the SPS.} \label{qqbar} \begin{tabular}{lccc} \hline\noalign{\smallskip} & SPS & RHIC & LHC \\ \noalign{\smallskip}\hline\noalign{\smallskip} N($c\bar{c}$) & 0.2 & 10 & 130 \\ N($b\bar{b}$) & -- & 0.05 & 5 \\ \noalign{\smallskip}\hline \end{tabular} \end{table} On the other hand, studies with open heavy flavors also benefit from high statistics measurements. In particular, as shown in the following, the reconstruction of the $p_{\rm T}$ distribution of $D^0$ mesons in the hadronic channel should provide valuable information on in-medium induced $c$ quark energy loss. \subsection{Large quarkonium nuclear absorption} Charmonium measurements at the SPS have shown that a detailed understanding of the normal nuclear absorption is mandatory in order to reveal any anomalous suppression behavior~\cite{louis}. According to Ref.~\cite{Bedjidian:2003gd}, the following observations can be made: \begin{itemize} \item the J/$\psi$ nuclear absorption in central Pb-Pb collisions is two times larger at the LHC than at the SPS; \item the J/$\psi$ nuclear absorption in central Ar-Ar collisions at the LHC is similar to the one in central Pb-Pb collisions at the SPS; \item the $\Upsilon$ nuclear absorption in central Pb-Pb collisions at the LHC is similar to the J/$\psi$ nuclear absorption in central Pb-Pb collisions at the SPS. \end{itemize} \subsection{Large resonance dissociation rate} It has been realized that, in addition to the normal nuclear absorption, the interactions with comoving hadrons and the melting by color screening, quarkonia can also be significantly destroyed by gluon ionization~\cite{Xu:1995eb}. Since this mechanism results from the presence of quasi-free gluons, it starts being effective for temperatures above the critical temperature but not necessarily above the resonance dissociation temperature by color screening. Recent estimates~\cite{Bedjidian:2003gd} (see Ref.~\cite{Blaschke:2004dv} for an update) of the quarkonium dissociation cross-sections show that none of the J/$\psi$ mesons survives the deconfined phase at the LHC and that about 80\,\% of the $\Upsilon$ are destroyed. Significant information about the initial temperature and lifetime of the QGP should be extracted from the $\Upsilon$ suppression pattern. \subsection{Large charmonium secondary production} An important yield of secondary charmonia is expected from $B$ meson decays~\cite{Eidelman:2004wy}, $D\overline{D}$ annihilation~\cite{Ko:1998fs}, statistical hadronization~\cite{Braun-Munzinger:2000px} and kinetic recombination~\cite{Thews:2000rj}. Contrary to the two first processes, the two last ones explicitly assume the formation of a deconfined medium. The underlying picture is that charmonium resonances form by coalescence of free $c$ and $\bar{c}$ quarks in the QGP~\cite{Thews:2000rj} or at the hadronization stage~\cite{Braun-Munzinger:2000px}. According to these models, the QGP should lead to an increase of the J/$\psi$ yield versus centrality, roughly proportional to ${\rm N}^2(c\bar{c})$, instead of a suppression. Due to the large number of $c\bar{c}$ pairs produced in central heavy ion collisions at the LHC, these models predict a spectacular enhancement of the J/$\psi$ yield; up to a factor 100 relative to the primary production yield~\cite{Bedjidian:2003gd,Andronic:2003zv}. Although the statistical accuracy of the present RHIC data cannot confirm or rule out such mechanisms~\cite{robert}, it is interesting to extrapolate from secondary charmonium production at RHIC to secondary bottomonium production at the LHC. Indeed, the expected multiplicity of $b\bar{b}$ pairs at the LHC is roughly equal to the expected multiplicity of $c\bar{c}$ pairs at RHIC (Table~\ref{qqbar}). Therefore, if secondary production of charmonia is observed at RHIC, it is conceivable to expect the same formation mechanism for bottomonium states at the LHC. \subsection{Complex structure of dilepton yield} The dilepton mass spectrum at the LHC exhibits new features, illustrated in Fig.~\ref{smbat2}. It can be seen that, with a low $p_{\rm T}$ threshold of around 2~GeV/$c$ on the decay leptons, unlike-sign dileptons from bottom decay dominate the dilepton correlated component over all the mass range. These dileptons have two different origins. In the high invariant mass region, each lepton comes from the direct decay of a $B$ meson (the so-called $BB$-diff channel). In the low invariant mass region, both leptons come from the decay of a single $B$ meson via a $D$ meson (the so-called $B$-chain channel). Next to leading order processes, such as gluon splitting, also populate significantly the low mass dilepton spectrum due to their particular kinematics. Then, as discussed in more detail below, a substantial fraction of the J/$\psi$ yield arises from bottom decays. Finally, a sizeable yield of like-sign correlated dileptons from bottom decays is present. This contribution arises from the peculiar decay chain of $B$ mesons and from $B$ meson oscillations (see below). Its yield could be even larger than the yield of unlike-sign correlated dileptons from charm. \begin{figure} \begin{center} \resizebox{0.35\textwidth}{!}{\includegraphics{smbat2.epsi}} \end{center} \caption{Invariant mass spectra of dimuons produced in central ($b<3$~fm) Pb-Pb collisions in the ALICE forward muon spectrometer~\cite{smbat}, with a $p_{\rm T}$ cut of 2~GeV/$c$ applied to each single muon. The lines correspond to: like-sign correlated dimuons from bottom (dotted); unlike-sign correlated dimuons from charm (dash-dotted) and from bottom (dashed); unlike-sign correlated and unlike-sign non-correlated pairs (solid).} \label{smbat2} \end{figure} \section{The LHC heavy ion program} The LHC will be operated several months per year in pp mode and several weeks in heavy-ion mode. The corresponding effective time for rate estimates is $10^7$~s for pp and $10^6$~s for heavy-ion operation. As described in Ref.~\cite{Carminati:2004fp}, the ``heavy-ion runs'' include, during the first five years of operation, one Pb-Pb run at low luminosity, two Pb-Pb runs at high luminosity, one p-A run and one light-ion run. In the following years different options will be considered, depending on the first results. Three of the four LHC experiments are expected to take heavy-ion data. \subsection{ALICE} ALICE (A Large Ion Collider Experiment) is the only LHC experiment dedicated to the study of nucleus-nucleus collisions~\cite{ALICEWEB}. The detector is designed to cope with large charged particle multiplicities which, in central Pb-Pb collisions, are expected to be between 2000 and 8000 per unit rapidity at mid rapidity. The detector consists of a central barrel ($|\eta|<0.9$), a forward muon spectrometer $(2.5<\eta<4$) and several forward/backward and central small acceptance detectors. Heavy flavors will be measured in ALICE through the electron channel and the hadron channel in the central barrel as well as through the muon channel in the forward region. Note that, contrary to the other LHC experiments, ALICE will be able to access most of the signals down to very low $p_{\rm T}$. \subsection{CMS} CMS (Compact Muon Solenoid)~\cite{CMSWEB} is designed for high $p_{\rm T}$ physics in pp collisions but has a strong heavy ion program~\cite{cms}. This program includes jet reconstruction, quarkonia measurements (in the dimuon channel) and high mass dimuon measurements. The detector acceptance, for quarkonia measurements, ranges from $-2.5$ to 2.5 in $\eta$, with a $p_{\rm T}$ threshold of 3~GeV/$c$ on single muons. Such a $p_{\rm T}$ cut still allows the reconstruction of $\Upsilon$ states down to $p_{\rm T} = 0$ but limits J/$\psi$ measurements to high $p_{\rm T}$. \subsection{ATLAS} Like CMS, ATLAS (A Toroidal LHC ApparatuS)~\cite{ATLASWEB} is designed for pp physics. The detector capabilities for heavy ion physics have been investigated recently~\cite{atlas}. As far as heavy flavors are concerned, the physics program will focus on measurements of $b$-jets and $\Upsilon$. The detector acceptance for muon measurements is large in $\eta$ ($|\eta|<2.4$) but, like CMS, is limited to high $p_{\rm T}$. \section{Selected physics channels} \subsection{Quarkonia} \subsubsection{Centrality dependence of resonance yields} The centrality dependence of the quarkonium yield, in the $\mu\mu$ channel, has been simulated in the ALICE detector. From the results, displayed in Table~\ref{smbatTable}, the following comments can be made. The statistics of J/$\psi$ events is large and should allow for narrower centrality bins. The $\psi^\prime$ measurement is rather uncertain, because of the small signal to background ratio (S/B). The $\Upsilon$ and $\Upsilon^\prime$ statistics and significance are quite good and the corresponding S/B ratios are almost always greater than~1. On the other hand, the $\Upsilon^{\prime\prime}$ suffers from limited statistics. The resonances will also be measured in the dielectron channel in ALICE~\cite{TRDTP}, and in the dimuon channel in CMS~\cite{cms} and ATLAS~\cite{atlas}, providing consistency cross-checks and a nice complementarity in acceptance. A recent study~\cite{sudhir} demonstrated the capabilities of ALICE to measure the resonance azimuthal emission angle with respect to the reaction plane. Such measurements are of particular importance given the latest RHIC results on open charm elliptic flow~\cite{Kelly:2004qw}. \begin{table}[ht] \caption{Preliminary yield (S), signal over background (S/B) and significance (${\rm S}/\sqrt{\rm S+B}$) for quarkonium resonances measured versus centrality in the ALICE forward muon spectrometer~\cite{smbat}. The input cross-sections are taken from Ref.~\cite{Bedjidian:2003gd}. Shadowing is taken into account. Any other suppression or enhancement effects are not included. The numbers correspond to one month of Pb-Pb data taking and are extracted with a 2$\sigma$ mass cut.} \label{smbatTable} \begin{tabular}{lllllll} \hline\noalign{\smallskip} & $b$ (fm) & 0-3 & 3-6 & 6-9 & 9-12 & 12-16 \\ \noalign{\smallskip}\hline\noalign{\smallskip} & S $(\times 10^3)$ & 86.48 & 184.6 & 153.3 & 67.68 & 10.46 \\ J/$\psi$ & S/B & 0.167 & 0.214 & 0.425 & 1.237 & 6.243 \\ & ${\rm S}/\sqrt{\rm S+B}$ & 111.3 & 180.4 & 213.8 & 193.4 & 94.95 \\ \noalign{\smallskip}\hline\noalign{\smallskip} & S $(\times 10^3)$ & 1.989 & 4.229 & 3.547 & 1.565 & 0.24 \\ $\psi^\prime$ & S/B & 0.009 & 0.011 & 0.021 & 0.063 & 0.273 \\ & ${\rm S}/\sqrt{\rm S+B}$ & 4.185 & 6.902 & 8.604 & 9.641 & 7.171 \\ \noalign{\smallskip}\hline\noalign{\smallskip} & S $(\times 10^3)$ & 1.11 & 2.376 & 1.974 & 0.83 & 0.118 \\ $\Upsilon$ & S/B & 2.084 & 2.732 & 4.31 & 7.977 & 12.01 \\ & ${\rm S}/\sqrt{\rm S+B}$ & 27.39 & 41.71 & 40.03 & 27.16 & 10.42\\ \noalign{\smallskip}\hline\noalign{\smallskip} & S $(\times 10^3)$ & 0.305 & 0.653 & 0.547 & 0.229 & 0.032 \\ $\Upsilon^\prime$ & S/B & 0.807 & 1.043 & 1.661 & 2.871 & 4.319 \\ & ${\rm S}/\sqrt{\rm S+B}$ & 11.68 & 18.26 & 18.48 & 13.02 & 5.077 \\ \noalign{\smallskip}\hline\noalign{\smallskip} & S $(\times 10^3)$ & 0.175 & 0.376 & 0.312 & 0.13 & 0.019 \\ $\Upsilon^{\prime\prime}$ & S/B & 0.566 & 0.722 & 1.18 & 1.936 & 3.024 \\ & ${\rm S}/\sqrt{\rm S+B}$ & 7.951 & 12.55 & 13 & 9.274 & 3.73 \\ \noalign{\smallskip}\hline \end{tabular} \end{table} \subsubsection{$\Upsilon^\prime/\Upsilon$ ratio versus $p_{\rm T}$} The $p_{\rm T}$ dependence of resonance suppression was recognized very early as a relevant observable to probe the characteristics of the deconfined medium~\cite{Blaizot:1987ha}. Indeed, the $p_{\rm T}$ suppression pattern of a resonance is a consequence of the competition between the resonance formation time and the QGP temperature, lifetime and spatial extent. However, quarkonium suppression is known to result not only from deconfinement but also from nuclear effects like shadowing and absorption. In order to isolate pure QGP effects, it has been proposed to study the $p_{\rm T}$ dependence of quarkonium ratios instead of single quarkonium $p_{\rm T}$ distributions. By doing so, nuclear effects are washed out, at least in the $p_{\rm T}$ variation of the ratio\footnote{Using ratios has the additional advantage that systematical detection inefficiencies cancel out to some extent.}. Following the arguments of Ref.~\cite{Gunion:1996qc}, the capabilities of the ALICE muon spectrometer to measure the $p_{\rm T}$ dependence of the $\Upsilon^\prime/\Upsilon$ ratio in central (10\,\%) Pb-Pb collisions have been recently investigated~\cite{ericTHESIS}. Two different QGP models with different system sizes were considered. The results of the simulations (Fig.~\ref{ericFIG}) show that, with the statistics collected in one month of data taking, the measured $\Upsilon^\prime/\Upsilon$ ratio exhibit a strong sensitivity to the characteristics of the QGP. Note that in the scenario of the upper right panel of Fig.~\ref{ericFIG} the expected suppression is too large for any measurement beyond the $p_{\rm T}$ integrated one. \begin{figure*} \begin{center} \resizebox{0.74\textwidth}{!}{\includegraphics{results.eps}} \end{center} \caption{$\Upsilon^\prime/\Upsilon$ ratio versus $p_{\rm T}$ for two different QGP models with different system sizes~\cite{ericTHESIS}. The solid curves correspond to the ``theoretical ratios''. The triangles show the expected measurements with the ALICE forward muon spectrometer in one month of central (10\,\%) Pb-Pb data taking (the open triangles correspond to the $p_{\rm T}$ integrated ratios). Error bars are of statistical origin only. The horizontal solid lines show the expected value of the ratio in pp collisions. More details on the ingredients used in the different scenarios can be found in Ref~\cite{Gunion:1996qc}.} \label{ericFIG} \end{figure*} \subsubsection{Secondary J/$\psi$ from bottom decay} A large fraction of the J/$\psi$ yield arises from the decay of $B$ mesons. The ratio ${\rm N}(b\bar{b}\rightarrow J/\psi)/{\rm N}({\rm direct}~J/\psi)$ can be determined as follows. The number of directly produced J/$\psi$ in central (5\,\%) Pb-Pb collisions is 0.31~\cite{Bedjidian:2003gd}\footnote{Including shadowing and no feed-down from higher states.}. The corresponding number of $b\bar{b}$ pairs (with shadowing) amounts to 4.56~\cite{Bedjidian:2003gd}. The $b\rightarrow {\rm J}/\psi$ branching ratio is $1.16\pm0.10\,\%$~\cite{Eidelman:2004wy}. Therefore ${\rm N}(b\bar{b}\rightarrow {\rm J}/\psi)/{\rm N}({\rm direct}~{\rm J}/\psi) = 34\,\%$ in $4\pi$. These secondary J/$\psi$ mesons from $b$ decays, which are not QGP suppressed, must be subtracted from the measured J/$\psi$ yield prior to J/$\psi$ suppression studies\footnote{In addition, 1.5\,\% of $B$ mesons decay into $\chi_{c1}(1P)$ which subsequently decay into $\gamma$J$/\psi$ with a 31\,\% branching ratio~\cite{Eidelman:2004wy}.}. They can further be used in order to measure the $b$ cross-section in pp collisions~\cite{Acosta:2004yw}, to estimate shadowing in p-A collisions and to probe the medium induced $b$ quark energy loss in A-A collisions. Indeed, it has been shown~\cite{Lokhtin:2001nh} that the $p_{\rm T}$ and $\eta$ distributions of those J/$\psi$ exhibit pronounced sensitivity to $b$ quark energy loss. In addition, a comparison between high mass dileptons and secondary J/$\psi$ distributions could clarify the nature of the energy loss~\cite{Lokhtin:2001nh}. Due to the large life-time of $B$ mesons, J/$\psi$ from bottom decay is the only source of J/$\psi$ not coming from the primary vertex\footnote{J/$\psi$ from statistical hadronization, kinetic recombination and $D\overline{D}$ annihilation are usually quoted as secondary J/$\psi$ but they originate from the primary vertex.}. The best way to identify them is, therefore, to reconstruct the invariant mass of dileptons with displaced vertices i.e.\ with impact parameter, $d0$, above some threshold. Simulations have shown that such measurements can successfully be performed with dielectrons measured in the central part of ALICE using the ITS, the TPC and the TRD~\cite{TRDTP} and with dimuons in CMS~\cite{Lokhtin:2001nh}, thanks to the excellent spatial resolution of the inner tracking devices of these experiments. It should also be possible to disentangle the two sources of J/$\psi$ from the slopes of the overall measured J/$\psi$ $p_{\rm T}$ distributions since primary J/$\psi$ have a harder spectrum~\cite{TRDTP}. Finally, a recent study~\cite{andreas} has demonstrated the possibility to isolate J/$\psi$ from bottom decay in pp collisions, without secondary vertex reconstruction, by triggering on three muon events in the ALICE forward muon spectrometer. Indeed, in standard (dimuon) pp events, the J/$\psi$ peak contains 85\,\% of primary J/$\psi$ and 15\,\% of J/$\psi$ from $B$ meson decays. The situation is totally inverted in tri-muon events because a $B\overline{B}$ pair can easily produce many decay leptons. In such events the J/$\psi$ peak contains 85\,\% of secondary J/$\psi$ from bottom decay and 15\,\% of direct J/$\psi$~\cite{andreas}. It is obvious that this analysis technique becomes less and less efficient as the track multiplicity increases. Nevertheless, it could still be performed for light-ion systems. \subsection{Open heavy flavors} \subsubsection{Open bottom from single leptons with displaced vertices} As mentionned above, the $d0$ distributions of leptons from heavy meson decays exhibit a significantly large tail because heavy mesons have a larger life-time than other particles decaying into leptons. Therefore, inclusive measurements of open heavy flavors can be achieved from the identification of the semi-leptonic decay of heavy mesons~\cite{TRDTP}. Recent simulation studies~\cite{rosario} performed with the ALICE central detectors show that with $d0>180~\mu{\rm m}$ and $p_{\rm T}>2$~GeV/$c$, the monthly expected statistics of electrons from $B$ decays in central Pb-Pb collisions is $5\cdot 10^4$ with a contamination of only 10\,\%, mainly coming from charm decays. The deconvolution of $d0$ distributions by imposing different $p_{\rm T}$ cuts should allow charm measurements as well. Furthermore, such analyses should give access to the $p_{\rm T}$ distribution of $D$ and $B$ mesons by exploiting the correlation between the $p_{\rm T}$ of the decay lepton and that of its parent~\cite{TRDTP}. \subsubsection{Open bottom from single muons and unlike-sign dimuons} The possibility to measure the differential $B$ hadron inclusive production cross-section in central Pb-Pb collisions at the LHC has recently been investigated by means of analyses similar to the ones performed in p$\bar{\rm p}$ collisions at the Tevatron. This study is based on unlike-sign dimuon mass and single muon $p_{\rm T}$ distributions measured with the ALICE forward muon spectrometer~\cite{rachid}. The principle is first to apply a low $p_{\rm T}$ threshold on single muons in order to reject background muons (mainly coming from charm decays) and therefore to maximize the $b$ signal significance. Then, fits are performed to the total (di)muon yield with fixed shapes for the different contributing sources and the bottom amplitude as the only free parameter. The $B$ hadron production cross-section is then obtained after corrections for decay kinematics and branching ratios and muon detection acceptance and efficiencies. This allows to extract the signal over a broad range in $p_{\rm T}$ (Fig.~\ref{rachidFIG}). A large statistics is expected~\cite{rachid} thus allowing detailed investigations on $b$ quark production mechanisms and in-medium energy loss. On the other hand, such a measurement, which can be performed for different centrality classes, provides the most natural normalization for $\Upsilon$ suppression studies. \begin{figure}[ht] \begin{center} \resizebox{0.40\textwidth}{!}{\includegraphics{rachid4.eps}} \end{center} \caption{Differential $B$ hadron inclusive production cross-section in the most central (5\,\%) Pb-Pb collisions~\cite{rachid}. Measurements from unlike-sign dimuons at low and high mass and from single muons (symbols) are compared to the input distribution (curve). Statistical errors (not shown) are negligible.} \label{rachidFIG} \end{figure} \subsubsection{Open bottom from like-sign dileptons} As shown in Fig.~\ref{smbat2}, a sizable fraction of like-sign correlated dileptons arise from the decay of $B$ mesons. These dileptons have two different origins: \begin{itemize} \item{The first decay generation of $B$ mesons contains $\sim 10\,\%$ of primary leptons and a large fraction of $D$ mesons which decay semi-leptonically with a branching ratio of $\sim 12\,\%$. Therefore a $B\overline{B}$ pair is a source of like-sign correlated dileptons via channels like:\\ \hspace*{0.5cm}$B^+$ $\rightarrow$ $\overline{D}^0$ $e^+$ $\nu_e$, $\overline{D}^0$ $\rightarrow$ $e^-$ anything\\ \hspace*{0.5cm}$B^-$ $\rightarrow$ $D^0$ $\pi^-$, $D^0$ $\rightarrow$ $e^+$ anything\\ where the $B^+B^-$ pair produces a correlated $e^+e^+$ pair in addition to the two correlated $e^+e^-$ pairs;} \item{The two neutral $B^0\overline{B}^0$ meson systems $B^0_d\overline{B}^0_d$ and $B^0_s\overline{B}^0_s$ undergo the phenomenon of particle-antiparticle mixing (or oscillation). The mixing parameters\footnote{Time-integrated probability that a produced $B^0_d$ ($B^0_s$) decays as a $\overline{B}^0_d$ ($\overline{B}^0_s$) and vice versa.} are $\chi_d~= 0.17$ and $\chi_s\ge 0.49$~\cite{Eidelman:2004wy}. Therefore, a $B^0_{d}\overline{B}^0_d$ ($B^0_s\overline{B}^0_s$) pair produces, in the first generation of decay leptons, $70\,\%$ ($50\,\%$) of unlike-sign correlated lepton pairs and $30\,\%$ ($50\,\%$) of like-sign correlated lepton pairs.} \end{itemize} This component is accessible experimentally from the subtraction of so-called event-mixing spectrum from the like-sign spectrum~\cite{Crochet:2001qd}. The corresponding signal is a reliable measurement of the bottom cross-section since i) $D$ mesons do not oscillate~\cite{Eidelman:2004wy} and ii) most (if not all) leptons from the second generation of $D$ meson decay can be removed by a low $p_{\rm T}$ threshold of about 2~GeV/$c$. \subsubsection{Hadronic charm} In the central part of ALICE, heavy mesons can be fully reconstructed from their charged particle decay products in the ITS, TPC and TOF~\cite{Dainese:2003zu}. Not only the integrated yield, but also the $p_{\rm T}$ distribution can be measured. The most promising decay channel for open charm detection is the $D^0 \rightarrow K^-\pi^+$ decay (and its charge conjugate) which has a branching ratio of 3.8\,\% and $c\tau=124~\mu{\rm m}$. The expected rates (per unit of rapidity at mid rapidity) for $D^0$ (and $\overline{D}^0$) mesons, decaying in a $K^\mp\pi^\pm$ pair, in central (5\,\%) Pb-Pb at $\sqrt{s}=5.5~{\rm TeV}$ and in pp collisions at $\sqrt{s}=14~{\rm TeV}$, are $5.3\cdot 10^{-1}$ and $7.5\cdot 10^{-4}$ per event, respectively. The selection of this decay channel allows the direct identification of the $D^0$ particles by computing the invariant mass of fully-reconstructed topologies originating from displaced secondary vertices. The expected statistics are $\sim 13\,000$ reconstructed $D^0$ in $10^7$ central Pb-Pb collisions and $\sim 20\,000$ in $10^9$ pp collisions. The significance is larger than 10 for up to about $p_{\rm T}=10$~GeV/$c$ both in Pb-Pb and in pp collisions. The cross section can be measured down to $p_{\rm T} = 1$~GeV/$c$ in Pb-Pb collisions and down to almost $p_{\rm T} = 0$ in pp collisions. \begin{figure}[ht] \begin{center} \resizebox{0.46\textwidth}{!}{\includegraphics{andrea2.epsi}} \end{center} \caption{Ratio of the nuclear modification factors for $D^0$ mesons and for charged (non-charm) hadrons with and without energy loss and dead cone effect~\cite{Dainese:2003wq}. Errors corresponding to the case ``no energy loss'' are reported. Vertical bars and shaded areas correspond to statistical and systematic errors, respectively.} \label{andrea2} \end{figure} The reconstructed $D^0$ $p_{\rm T}$ distributions can be used to investigate the energy loss of $c$ quarks by means of the nuclear modification factor $R_{\rm AA}^{D^0}$~\cite{Dainese:2003zu,Dainese:2003wq}. Even more interesting is the ratio of the nuclear modification factors of $D^0$ mesons and of charged (non-charm) hadrons ($R_{D/h}$) as a function of $p_{\rm T}$. Apart from the fact that many systematic uncertainties on $R_{\rm AA}^{D^0}$ cancel out with the double ratio, $R_{D/h}$ offers a powerful tool to investigate and quantify the so-called dead cone effect (Fig.~\ref{andrea2}). \subsubsection{Electron-muon coincidences} The semi-leptonic decay of heavy mesons involves either a muon or an electron. Therefore, the correlated $c\bar{c}$ and $b\bar{b}$ cross-sections can be measured in ALICE from unlike-sign electron-muon pairs where the electron is identified in the central part and the muon is detected in the forward muon spectrometer. The $e\mu$ channel is the only leptonic channel which gives a direct access to the correlated component of the $c\bar{c}$ and $b\bar{b}$ pairs. Indeed, in contrast to $e^+e^-$ and $\mu^+\mu^-$ channels, neither a resonance, nor direct dilepton production, nor thermal production can produce correlated $e\mu$ pairs. Within ALICE, the $e\mu$ channel has the additional advantage that the rapidity distribution of the corresponding signal extends from $\sim 1$ to $\sim 3$, therefore bridging the acceptances of the central and the forward parts of the detector~\cite{Lin:1998bd}. Electron-muon coincidences have already been successfully measured in pp collisions at $\sqrt{s}=60~{\rm GeV}$~\cite{Chilingarov:1979ur} and in p-nucleus collisions at $\sqrt{s}=29~{\rm GeV}$~\cite{Akesson:1996wf}. Preliminary simulations have shown the possibility, with ALICE, to measure the correlated $e\mu$ signal after appropriate background subtraction~\cite{MUONTDR}. \section{Summary} The heavy flavor sector will bring fantastic opportunities for systematic explorations of the dense partonic system formed in heavy ion collisions at the LHC through a wide variety of physics channels. In addition to the channels discussed here, further exciting possibilities should be opened with, for example, charmed baryons, high mass dileptons, quarkonia polarization and dilepton correlations. \section*{Acknowledgments} I am grateful to A.~Andronic, M.~Bedjidian, A.~Dainese, S.~Grigoryan, R.~Guernane, G.~Martinez and A.~Morsch for their help in preparing this paper.
{ "timestamp": "2005-04-05T09:10:37", "yymm": "0503", "arxiv_id": "nucl-ex/0503008", "language": "en", "url": "https://arxiv.org/abs/nucl-ex/0503008" }
\section{Introduction} It is interesting for theoretical and practical reasons to study coherent and squeezed states associated to the quantum Hopf algebras \cite{kn:Dri86,kn:Reshe,kn:Ma}. The Hopf algebra structure of a quantum algebra provides us with useful technical elements such as the coproduct, for exemple. In the case of boson quantum algebras, the special coproduct properties are useful to characterize multi-particle Hamiltonians \cite{kn:TsoPaJa}. For example, in the case of the Poincar\'{e} quantum algebra, the coproduct have been brought to bear to the study the fusion of phonons \cite{kn:Cele2}. In general, the concept of deformed quantum Lie algebras found various applications in quantum optics, quantum field theory, quantum statistical mechanics, supersymmetric quantum mechanics and some purely mathematical problems. For instance, in the case of the $su_q (2)$ algebra, it has been found that the $su_q (2)$ effective Hamiltonians reproduce accurately the physical properties of the $su(2) \oplus h(2)$ models \cite{kn:BaCiHeRe}. On the other side, there are some works showing that quantum algebras are connected with paragrassmann algebras \cite{kn:Spiri,kn:Fi}. Paragrassmann algebras are relevant in the studies of theories that show the necessity of unusual statistic \cite{kn:Ru}, for instance, the studies of anyons and topological field theories \cite{kn:MaWi,kn:AnBa}. Now, to associate coherent and squeezed states to a quantum deformed Lie algebra one can use the algebra eigentates (AES) technique. The AES associated to a real Lie algebra have been defined as the set of eigenstates of an arbitrary complex linear combination of generators of the considered algebra \cite{kn:Brif}. The AES associated to a quantum real deformed Lie algebra can be defined in a similar way. Indeed, if $A_k (q), \, k=1,2, \ldots, n$ denote the generators of this deformed algebra in a given representation, parametrized by the set of deformation parameters $q,$ then the AES associated to this deformed algebra will be given by the set of solutions of the eigenvalue equation \begin{equation} \sum_{k=1}^n \alpha_k A_k (q) |\psi \rangle = \lambda |\psi \rangle, \qquad \alpha_k, \lambda \in {\mathbb C}. \end{equation} The purpose of this work is to compute the AES of the deformed quantum Heisenberg Lie algebras \cite{kn:HL94}, obtained by applying the R-matrix methods \cite{kn:Dri86}, and find new classes of deformed harmonic oscillator coherent and squeezed states. We will see that these states will be new deformations of the standard coherent and squeezed states of the harmonic oscillator system and we will recover them in the limit when the deformation parameters go to zero. The approach of AES also gives us the possibility to construct, starting from a deformed algebra, some Hamiltonians, of physical systems to which these deformed coherent and squeezed states are associated, similarly as for algebras and superalgebras \cite{kn:NaVh1,kn:NaVh3}. It is important to mention that the deformed coherent states obtained by this method differ from the $q$--deformed coherent states associated to a $q$-deformed oscillator algebra, which is not a Hopf algebra, constructed by considering either deformed exponential functions, eigenstates of a given deformed annihilation operator, a generalization of the usual form of the standard coherent states, a resolution of the identity technique or a generalized group theoretical techniques \cite{kn:Dellinas,kn:Cquesne,kn:BjuPSt}. The paper is organized as follows. In section \ref{sec-two}, a Fock space representation of deformed quantum algebras associated to the Heisenberg algebra $h(2)$ is given. In section \ref{sec-three}, we compute the AES associated to these algebras and obtain new classes of deformed coherent and squeezed states that are true deformations of the standard coherent and squeezed states associated to the harmonic oscillator system. These states are parametrized by the deformation parameters which will be considered as real numbers and also as real paragrassmann numbers. In section \ref{sec-cuatro}, we compute the product of the dispersions of the position and linear momentum operators of a particle in these states when the parameters of deformation are small. We compare them with the corresponding results obtained in the minimum uncertainty states \cite{kn:NaVh1}. Some details of calculations are presented in the Appendices \ref{sec-appa} and \ref{sec-appb}. We also give general expressions of these dispersions, in the case where a non trivial one parameter algebra deformation family is concerned, for all values of the deformation parameter. Finally, we construct a class of $\eta$--pseudo Hermitian Hamiltonians \cite{kn:AMostafazadeh} to which a subset of these deformed states are the associated coherent states. \section{Deformed quantum Heisenberg algebras in the Fock representation space} \label{sec-two} We are considering in this work, the deformed Heisenberg quantum algebras obtained by V. Hussin and A. Lauzon \cite{kn:HL94}. They have been obtained using the well-known $R$--matrix method \cite{kn:Dri86} and are mainly of two types. The first one is formed by the generators $A,B,C$ which satisfy \begin{equation} \left[A,B\right] = 0, \qquad \left[B,C\right] = - { 2 z \over p^2 } ( \cosh( pB ) - 1 ), \qquad \left[A,C\right] = {1 \over p} \sinh (pB) . \label{com-he1} \end{equation} It is denoted by $ {\cal U}_{z,p} \, (h(2)),$ where $p$ and $z$ are different from zero. Let us mention that the invertible change of basis \begin{eqnarray} {\tilde A} = A, \qquad {\tilde B} = { 2 \over p} \sinh \left( {p B \over 2} \right), \qquad {\tilde C} = { 1 \over \cosh \left( {p B \over 2} \right) } \ C , \label{optilde} \end{eqnarray} leads to the new deformed algebra $ \ {\tilde {\cal U}}_{z,0} \, (h(2)):$ \begin{equation} [{\tilde A} , {\tilde B}] =0, \qquad [{\tilde B} , {\tilde C}] = - z {\tilde B}^2, \qquad [{\tilde A} , {\tilde C}] = {\tilde B}. \label{com-he1tilde} \end{equation} This means that we get the same commutation relations as in \eqref{com-he1} when $p$ goes to zero. As it has been pointed out by Ballesteros et al. \cite{kn:BaHePr}, here the $p$ parameter is superfluous and the families of bialgebras $ {\cal U}_{z,p} \, (h(2))$ and $ {\cal U}_{z,0} \, (h(2))$ are isomorphic (these families are identified there as of type $I_+ $) on the condition that the coproduct form stands invariant \cite{kn:BaCeOl}. The second quantum deformation of $h(2)$ is given by \begin{equation} \label{altype2} \left[A,B\right] = \left[B, C\right] =0, \qquad \left[A,C\right] = {e^{pB} - e^{- qB} \over p+q} . \end{equation} and is denoted by ${\cal U}_{p,q} \, (h(2)),$ where $p,q \ne 0.$ It corresponds to so-called so called type $II$ bialgebras in \cite{kn:BaHePr}. When $p=q,$ we find the quantum Heisenberg algebra obtained in Celeghini et al. \cite{kn:Cele} (see also \cite{kn:BaCeOl}), i.e., \begin{equation} \left[A,B\right] = \left[B,C\right] =0, \qquad \left[A,C\right] = {1 \over p} \sinh (p B). \label{cel-et-al} \end{equation} Let us now give a boson realization of these deformed Lie algebras, in terms of the usual creation operator, $a^\dagger, $ and annihilation operator, $a,$ associated to the standard quantum harmonic oscillator system. For ${\tilde {\cal U}}_{z,0} \, (h(2))$ given in \eqref{com-he1tilde}, it is given by \begin{equation} \label{cas1} \tilde A = - a^\dagger, \qquad \tilde B = e^{z a^\dagger}, \qquad \tilde C = e^{z a^\dagger} \ a . \end{equation} From \eqref{optilde} and \eqref{cas1}, we thus get a realization of ${\cal U}_{z,p} \, (h(2))$ as \begin{equation} \label{op-def-one-zp} A = - a^\dagger, \qquad B = {2\over p} \sinh^{-1} \left( {p \over 2} e^{z a^\dagger} \right), \qquad C = e^{ z a^\dagger} \sqrt{1+ {\left({p \over 2} e^{z a^\dagger}\right)}^2 } a. \end{equation} Another realization of ${\tilde {\cal U}}_{z,0} \, (h(2))$ is \begin{equation} \label{cas2} \tilde A = a, \qquad \tilde B = e^{- z a}, \qquad \tilde C = a^\dagger \ e^{-za}. \end{equation} We thus get another realization of ${\cal U}_{z,p} \, (h(2))$ as \begin{equation} \label{op-def-two-zp} A = a, \qquad \qquad B = {2\over p} \sinh^{-1} \left( {p \over 2} e^{- z a} \right), \qquad C = a^\dagger e^{- z a} \sqrt{1+ {\left({p \over 2} e^{-z a}\right)}^2 }. \end{equation} When $z$ goes to zero, the operators \eqref{op-def-one-zp} become \begin{equation} \label{def1} A=- a^\dagger , \qquad B= {2 \over p} \sinh^{-1} \left({p\over 2} \right) I, \qquad C = \sqrt{1+ {p^2 \over 4}} a, \end{equation} while the operators \eqref{op-def-two-zp} become \begin{equation} \label{def2} A = a, \qquad B= {2 \over p} \sinh^{-1} \left({p\over 2} \right) I, \qquad C = \sqrt{1+ {p^2 \over 4}} a^\dagger. \end{equation} The operators \eqref{def1} or \eqref{def2} thus constitute a realization of deformed Heisenberg algebra \eqref{cel-et-al}. When $p$ goes to zero, we regain $h(2).$ The algebra \eqref{altype2} is clearly isomorphic to $h(2)$ if we introduce \begin{equation} \tilde A =A, \qquad \tilde C = C, \qquad \tilde B = {e^{pB} - e^{-qB} \over p+q}. \end{equation} So to obtain new class of deformed coherent and squeezed states using the AES method we will deal in the following with ${\tilde {\cal U}}_{z,0} \, (h(2))$ and ${\cal U}_{z,p} \, (h(2)).$ \section{AES and deformed coherent and squeezed states} \label{sec-three} In this section, we compute the AES associated to ${\tilde {\cal U}}_{z,0} \, (h(2))$ and ${\cal U}_{z,p} \, (h(2)), $ using the representations obtained in the preceding section. We thus get new classes of deformed coherent and squeezed states associated to the harmonic oscillator system. \subsection{Deformed algebra eigenstates for \mathversion{bold} ${\tilde {\cal U}}_{z,0} \,( h(2))$} \label{sec-aes-He} We start with ${\tilde {\cal U}}_{z,0} \,( h(2))$ as given by \eqref{com-he1tilde} using the realizations \eqref{cas1} and \eqref{cas2}. The AES are thus defined as the set of solutions of the eigenvalue equation \begin{equation} \label{aes-10} [\alpha_+ {\tilde A} + \alpha_0 {\tilde B} + \alpha_- {\tilde C} ] |\psi\rangle = \alpha |\psi\rangle, \qquad \alpha_- , \alpha_0 , \alpha_+ , \alpha \in {\mathbb C}. \end{equation} \subsubsection{Deformed harmonic oscillator coherent and squeezed states}\label{sub-sec-coh-squee} Let us take first the realization \eqref{cas1}. Thus, if $\alpha_- \ne 0, $ equation \eqref{aes-10} can be written in the form \begin{equation} \label{eigen-10}[ e^{z a^\dagger} a + \mu a^\dagger + \nu e^{z a^\dagger}] |\psi \rangle = \lambda |\psi \rangle, \qquad \mu,\nu, \lambda \, \in {\mathbb C}. \end{equation} By defining \begin{equation} \label{psi-varphi} |\psi\rangle = e^{- \nu a^\dagger} |\varphi\rangle \end{equation} and using $ e^{-\nu a^\dagger} \, a \,e^{\nu a^\dagger}= a + \nu, $ equation \eqref{eigen-10} can be reduced to \begin{equation} \label{reduce-001}[ e^{z a^\dagger} a + \mu a^\dagger ] |\varphi \rangle = \lambda |\varphi \rangle, \qquad \mu, \lambda \, \in {\mathbb C}. \end{equation} To solve this eigenvalue equation, let us consider the Bargmann space ${\cal F}$ of analytic functions $f(\xi)$ ($ \xi \, \in {\mathbb C}),$ provided with the scalar product \begin{equation} (f_1, f_2) = \int_{{\mathbb C}} \overline{f_1 (\xi) } f_2 (\xi) e^{-{\bar \xi} \xi} {d {\bar \xi} d\xi \over 2 \pi i}, \qquad \forall \, f_1,f_2 {\in \cal F}. \end{equation} It is well-know that any function $f \in {\cal F}$ can be expressed as a linear combination of orthonormalized functions $u_n (\xi)= {\xi^n \over \sqrt{n!}}, \, n=0,1,2, \ldots,$ verifying \begin{equation} \label{ortho-umn} (u_m, u_n) = \int_{{\mathbb C}} \overline{u_m (\xi)} u_n (\xi) e^{-{\bar \xi} \xi} {d {\bar \xi} d\xi \over 2 \pi i} = \delta_{mn}, \end{equation} that is \begin{equation} f(\xi) = \sum_{n=0}^\infty c_n u_{n} (\xi), \end{equation} with \begin{equation} c_n = \int_{{\mathbb C}} \overline{u_n (\xi)} f (\xi) e^{-{\bar \xi} \xi} {d {\bar \xi} d\xi \over 2 \pi i}. \end{equation} Let us assume a solution of \eqref{reduce-001} of the type \begin{equation} \label{type-sol} | \varphi \rangle = \sum_{n=0}^{\infty} c_n |n\rangle, \end{equation} where the set of states $\{ |n\rangle \}_{n=0}^{\infty}$ form the basis of the standard Fock oscillator space, verifying the orthogonality relation \begin{equation} \langle m |n\rangle =\delta_{mn}. \label{st-mn-ortho} \end{equation} As usually, the action of the operators $a$ and $a^\dagger $ on these states is given by \begin{equation} a | n \rangle = \sqrt{n} |n-1\rangle, \qquad a^\dagger | n \rangle = \sqrt{n+1} |n+1\rangle. \end{equation} Let us take $| {\bar \xi} \rangle $ to be the standard coherent states associated to the harmonic oscillator system, that is \begin{equation} | {\bar \xi} \rangle = e^{{\bar \xi} a^\dagger} |0 \rangle = \sum_{n=0}^{\infty} {{(\bar \xi )}^n \over \sqrt{n!}} |n\rangle. \end{equation} Then, according to the orthogonality property \eqref{st-mn-ortho}, the projection of $|\varphi \rangle$ on the coherent state $| {\bar \xi} \rangle $ is given by the analytic function \begin{equation} \label{var-xi} \varphi (\xi) = \langle {\bar \xi} | \varphi \rangle =\sum_{n=0}^{\infty} c_n u_{n} (\xi). \end{equation} The action of the operators $a^\dagger $ and $a$ in this representation corresponds to \begin{equation} \label{act-xi}\langle {\bar \xi} | a^\dagger |\varphi \rangle = \xi \varphi (\xi), \qquad \; \langle {\bar \xi} | a |\varphi \rangle= {d \varphi \over d\xi} (\xi). \end{equation} respectively. Thus, by projecting both sides of the eigenvalue equation \eqref{reduce-001} on the coherent states $| {\bar \xi} \rangle$ and then using \eqref{act-xi}, we can write it as \begin{equation} \label{eigen-fock} \left( e^{z\xi} {d\over d\xi} + \mu \xi\right) \varphi (\xi) = \lambda \varphi (\xi). \end{equation} The general solution of this differential equation is given by \begin{equation} \varphi (\xi) = C_0 (\lambda,\mu, z ) \, \exp\left(\sum_{k=0}^{\infty} {{(-z\xi)}^k \over (k+1)!} \left( \lambda \xi - {k+1 \over k+2}\mu \xi^2 \right) \right), \label{solgen-varphi} \end{equation} where $ C_0 $ is an arbitrary constant which can be fixed from the normalization condition \begin{equation} \label{con-nor-var}( \varphi , \varphi) = \int_{{\mathbb C}} \overline{\varphi (\xi )} \varphi (\xi) e^{-{\bar \xi} \xi} {d {\bar \xi} d\xi \over 2 \pi i} =1.\end{equation} Let us notice that in the particular limit when $z$ goes to zero, the solution \eqref{solgen-varphi}, becomes the symbol for the squeezed states \cite{kn:Dodo} associated to the standard harmonic oscillator, that is \begin{equation} \label{sym-squee} \varphi (\xi) = C_0 (\lambda,\mu,0) \, \exp\left( \lambda \xi - {\mu \over 2} \xi^2 \right) . \end{equation} This quantity is normalizable only if $|\mu| <1 $ \cite{kn:NORu}. When $z\ne 0,$ the solution \eqref{solgen-varphi} can be written in the form \begin{equation} \varphi (\xi) = C_0 (\lambda, \mu, z ) \, \exp\left( {\lambda \over z} - {\mu \over z^2} \right) \, \exp\left( e^{-z \xi} {(\mu - \lambda z + \mu z \xi) \over z^2} \right). \label{solgen-varphi2} \end{equation} Going back to the expression \eqref{var-xi}, we get the coefficients $c_n, \, n=0,1,\ldots, $ as \begin{eqnarray} c_n = \int_{{\mathbb C}} \overline{u_n (\xi )} \varphi (\xi) e^{-{\bar \xi} \xi} {d {\bar \xi} d\xi \over 2 \pi i} &=& C_0 (\lambda, \mu, z ) \, \exp\left( {\lambda \over z} - {\mu \over z^2} \right) \nonumber \\ & & \int_{{\mathbb C}} {{\bar \xi}^n \over \sqrt{n!}} \exp\left( e^{-z \xi} {(\mu - \lambda z + \mu z \xi) \over z^2} \right) e^{-{\bar \xi} \xi} {d {\bar \xi} d\xi \over 2 \pi i} . \end{eqnarray} By using the polar change of variables $\xi = \rho e^{i \vartheta},$ this last equation can be written in the form \begin{eqnarray} c_n &=& C_0 (\lambda, \mu, z ) \, \exp\left( {\lambda \over z} - {\mu \over z^2} \right) \nonumber \\ & & \int_{0}^\infty \int_{0}^{2 \pi} { \rho^{n+1} e^{- \rho^2 } \over \sqrt{n!}} e^{-i n \vartheta} \exp\left( {e^{-z \rho e^{i \vartheta}} \over z^2} ( \mu - \lambda z + \mu z \rho e^{i \vartheta})\right) { d\rho d\vartheta \over \pi} . \end{eqnarray} Let us write the exponential factor in the form \begin{eqnarray} && \nonumber \exp\left( {e^{-z \rho e^{i \vartheta}} \over z^2} ( \mu - \lambda z + \mu z \rho e^{i \vartheta})\right) \\ \nonumber &=& \sum_{k=0}^{\infty} { \exp\left(-z k \rho e^{i \vartheta}\right) \over k! } {\left( \mu - \lambda z + u z \rho e^{i \vartheta} \over z^2\right) }^{k} \nonumber \\ &=& \sum_{k,l=0}^{\infty} \sum_{m=0}^{k} {k \choose m} \rho^{l+m} e^{i (l+m) \vartheta} {{( - z k )}^l {(\mu z)}^m {(\mu - \lambda z)}^{k-m}\over k! \, l! \, z^{2k}} \end{eqnarray} to get \begin{eqnarray} c_n &=& C_0 (\lambda, \mu, z ) \, \exp\left( {\lambda \over z} - {\mu \over z^2} \right) \, \sum_{k,l=0}^{\infty} \sum_{m=0}^{k} {k \choose m} {{( - z k )}^l {(\mu z)}^m {(\mu - \lambda z)}^{k-m}\over \sqrt{n!} \; k! \, l! \, z^{2k}} \nonumber \\ && \left(\int_{0}^{\infty} \rho^{m+l+n+1} e^{- \rho^2} d\rho\right) \left(\int_{0}^{2 \pi} e^{i (l+m-n)\vartheta} {d\vartheta \over \pi} \right). \end{eqnarray} Using the known results \begin{equation} \int_{0}^{2 \pi} e^{i (l+m-n)\vartheta} {d\vartheta \over \pi} = 2 \delta_{l+m-n,0}, \end{equation} \begin{equation}\int_{0}^{\infty} \rho^{m+l+n+1} e^{- \rho^2} d\rho = {1\over 2} \Gamma\left({m+l+n \over 2} +1 \right),\end{equation} and performing the sum over the index $l,$ the expression for the coefficients $c_n$ reduces to \begin{equation} \label{coeff-cn} c_n = C_0 (\lambda, \mu, z ) \, \exp\left( {\lambda \over z} - {\mu \over z^2} \right) {z^n \over \sqrt{n!}} \, \sum_{k=0}^{\infty} \sum_{m=0}^{k_{<}} {n \choose m} {{( -k )}^{n-m} \over (k-m)!} {\left({\mu \over z^2}\right)}^m {\left({\mu\over z^2} - {\lambda \over z} \right)}^{k-m} ,\end{equation} where $k_{<}$ denotes the minimum between $k$ and $n.$ This last expression can be written in the form \begin{equation} c_n = C_0 (\lambda, \mu, z ) \, {z^n \over \sqrt{n!}} \, \sum_{m=0}^{n} \sum_{j=0}^{n-m} {n \choose m} {(-1)}^{n-m} \upsilon_{mj} {\left({\mu \over z^2}\right)}^m {\left({\mu\over z^2} - {\lambda \over z} \right)}^j , \label{mejorcn} \end{equation} where the coefficients $\upsilon_{mj}$ are obtained from \begin{equation} {k^{n-m} \over (k-m)! } = \sum_{j=0}^{n-m} {\upsilon_{mj} \over (k-m-j)!}. \end{equation} Thus the coefficients $c_n, $ $n=1,2,\ldots, $ represent polynomials of degree $n-1$ in the z variable. For example, $ c_1 = \lambda C_0, $ \begin{equation} c_2 = C_0 \sqrt{2!} \left[ \left({\lambda^2 \over 2!} - {\mu \over 2}\right) - {\lambda \over 2 } z \right], \quad \nonumber c_3 = C_0 \sqrt{3!} \left[ \left( {\lambda^3 \over 3!} - {\mu \lambda \over 2} \right) + \left({\mu \over 3} - {\lambda^2 \over 2}\right) z + {\lambda \over 6} z^2 \right]. \end{equation} The normalization constant $C_0$ can be now computed. Indeed, inserting \eqref{mejorcn} into \eqref{var-xi} and the resulting expression into the normalization condition \eqref{con-nor-var}, using the orthogonality relation \eqref{ortho-umn}, we get \begin{eqnarray} \nonumber C_0 \, ( \lambda, \mu, z ) &=& \Biggl[ \sum_{n=0}^{\infty} \, {z^{2n} \over n!} \, \sum_{m=0}^{n} \sum_{r=0}^{n} \sum_{j=0}^{n-m} \sum_{l=0}^{n-r} {n \choose m}{n \choose r} {(-1)}^{m+r} \upsilon_{mj} \upsilon_{rl} \\ & & {\left( {\mu \over z^2 } \right)}^m {\left( {{\bar \mu} \over z^2 } \right)}^r {\left({\mu\over z^2} - {\lambda \over z} \right)}^j {\left({{\bar \mu} \over z^2} - {{\bar \lambda} \over z} \right)}^l \Biggr]^{-{1 \over 2 } } , \label{nor-fac}\end{eqnarray} which has been chosen real. The convergence of these series it not easy to determine. In the case where $z=0,$ as we have already mentioned, the series $\sum_{n=0}^\infty {|c_n|}^2$ converges for all $\lambda$ provided that $|\mu| < 1.$ In the case $\mu=0,$ this series becomes \begin{equation} \sum_{n=0}^\infty {|c_n|}^2 = {|C_0 (\lambda,z)|}^2 \exp\left({\lambda \over z}\right) \sum_{n=0}^\infty {{\biggl(- { \lambda \over z}\biggr)}^{n} \over n!} \exp\left( -{{\bar \lambda} \over z} \sum_{k=1}^\infty {{(z^2 n)}^k \over k!}\right). \end{equation} It converges for all $z > 0$ provided that the phase $\theta$ in $\lambda= \beta e^{i \theta}$ satisfies $ - {\pi \over 2} \le \theta \le {\pi \over 2}, $ whereas for all $z<0, $ it converges if $ {\pi \over 2} \le \theta \le {3 \pi \over 2}. $ Finally, we can show that the normalized algebra eigenstates $|\varphi \rangle,$ solving \eqref{reduce-001} , can be expressed in terms of a deformed squeezed operator acting on the ground state of the standard harmonic oscillator, that is \begin{equation} |\varphi \rangle = C_0 \, ( \lambda, \mu, z )\exp \left( \sum_{k=0}^{\infty} {{(-z a^\dagger )}^k \over (k+1)!} \left( \lambda a^\dagger - { k + 1 \over k+2 } \mu {(a^\dagger)}^2 \right) \right) |0 \rangle. \label{re-norm-squee} \end{equation} Also, combining this last equation with equation \eqref{psi-varphi}, we get the algebra eigenstates solving \eqref{eigen-10} to be the deformed coherent states \begin{equation} |\psi \rangle = N_0 ( \lambda, \mu, \nu, z ) \,\exp \left( \sum_{k=0}^{\infty} {{(-z a^\dagger )}^k \over (k+1)!} \left( \lambda a^\dagger - { k + 1 \over k+2 } \mu {(a^\dagger)}^2 \right) \right) e^{-\nu a^\dagger}|0 \rangle, \label{eigen-10-aes} \end{equation} where $N_0 \, ( \lambda, \mu, \nu, z )$ is a normalization constant which can computed in the same way as $ C_0 \, ( \lambda, \mu, z ).$ \subsubsection{Perturbed squeezed states} \label{sec-perturba-z-real} Let us now assume that $z$ is a small perturbation parameter of order $k_0 -1$, where $k_0$ is an integer greater or equal to $2$. From \eqref{re-norm-squee}, neglecting the terms containing the power of $ z $ greater than $k_0 -1$, we can write \begin{eqnarray} \nonumber |\varphi \rangle & \thickapprox & C_0 (\lambda, \mu, z, k_0) \Biggl[ 1 + \sum_{k=1}^{k_0 - 1} {{(-z a^\dagger )}^k \over (k+1)!} \left( \lambda a^\dagger - { k + 1 \over k+2 } \mu {(a^\dagger)}^2 \right) \\ &+& \cdots + { 1 \over (k_0 -1)!} {\Biggl( {-z a^\dagger \over 2!} \left(\lambda a^\dagger - {2\over 3} \mu {(a^\dagger)}^2 \right) \Biggr)}^{k_0 -1} \Biggr] \exp\left( \lambda a^\dagger - {\mu \over 2} {(a^\dagger)}^2 \right) |0 \rangle. \label{def-squee-st}\end{eqnarray} These states can be normalized in the standard form. For instance, when $k_0 = 2, $ $\mu = \delta e^{i \phi},$ $\lambda = \beta e^{i \theta}, $ where $\phi$ and $\theta $ are real phases, $ 0 \le \delta < 1, $ and $\beta \ge 0,$ a normalized version of the deformed squeezed states \eqref{def-squee-st}, is given by \begin{eqnarray} \nonumber |\varphi \rangle &\thickapprox& \Omega (\delta, \phi, \beta, \theta) \left[ 1 + z \left( {\delta e^{i \phi} \over 3} {(a^\dagger)}^3 - {\beta e^{i \theta} \over 2} {(a^\dagger)}^2 \right) \right] \\ & & S \left( - \arctan (\delta )e^{i \phi} \right) D \left( {\beta e^{i\theta} \over \sqrt{1 - \delta^2}} \right) |0 \rangle, \label{nor-z-def} \end{eqnarray} where \begin{eqnarray} \Omega (\delta, \phi, \beta, \theta) &=& 1 + {z \beta \over 2 {(1-\delta^2)}^2} \Biggl[ \left( 2 \delta^2 + \beta^2 \left( {1 + \delta^2 \over 1-\delta^2} \right) \right) \cos\theta \nonumber \\ &-& \delta \left( 1 + \delta^2 + { 2 \beta^2 \over 1-\delta^2 } \right) \cos(\phi - \theta) \nonumber \\ &+& \delta^2 \beta^2 \left( 1 + {2 \delta^2 \over 3(1-\delta^2)} \right) \cos(2 \phi - 3 \theta) - {2\delta \beta^2 \over 3(1-\delta^2)} \cos(\phi - 3 \theta) \Biggr]. \end{eqnarray} Here $ S(\chi) = \exp\left[ - \left( \chi {{(a^\dagger)}^2 \over 2} - {\bar \chi} {a^2 \over 2} \right)\right] $ is the standard unitary squeezed operator \cite{kn:ruso} and $ D (\lambda) = \exp\left( \lambda a^\dagger - {\bar \lambda} a \right) $ the standard displacement operator \cite{kn:pere}. \subsubsection{Deformed squeezed and coherent states parametrized by paragrassmann numbers} \label{sec-paragra} Let us now use the realization \eqref{cas2} of ${\tilde {\cal U}}_{z,0} \, (h(2))$. In the case $\alpha_+ \ne 0, $ equation \eqref{aes-10} can be now written in the form \begin{equation} \label{eigen-100}[a + \mu a^\dagger e^{- z a} + \nu e^{-z a}] |\psi \rangle = \lambda |\psi \rangle, \qquad \mu,\nu, \lambda \, \in {\mathbb C}. \end{equation} There are two types of equations to solve. The first type is obtained when $\mu \ne 0 $ and $\nu \ne 0.$ We can take \begin{equation} |\psi \rangle = \exp\left({\nu \over \mu} a \right) |\varphi\rangle \end{equation} and use the relation, $ \exp\left(- {\nu \over \mu} a \right) a^\dagger \exp\left({\nu \over \mu} a \right) = a^\dagger - {\nu \over \mu}, $ to reduce \eqref{eigen-100} to the form \begin{equation} \label{eigen-200} [a + \mu a^\dagger e^{- z a} ] |\varphi \rangle = \lambda |\varphi \rangle, \qquad \mu, \lambda \, \in {\mathbb C}. \end{equation} If $\nu =0$ and $\mu \ne 0,$ we see from \eqref{eigen-100} that the same type of eigenvalue equation must be solved. The second type is obtained when $\mu =0. $ The eigenvalue equation is \begin{equation} \label{eigen-30} [a + \nu e^{- z a} ] |\psi \rangle = \lambda |\psi \rangle, \qquad \nu, \lambda \, \in {\mathbb C}.\end{equation} We begin with the resolution of Equation \eqref{eigen-200}. Let us assume $| \varphi \rangle $ to be again a solution of the type \eqref{type-sol}. Thus, proceeding as in the preceding section, the eigenvalue equation satisfied by the symbol $\varphi (\xi),$ in the Bargmann representation, is given by \begin{equation} \label{fock-para} \left({d \over d \xi} + \mu \, \xi \, e^{- z {d \over d \xi}} \right) \varphi (\xi) = \lambda \varphi (\xi ), \qquad \mu, \lambda \, \in {\mathbb C}. \end{equation} To solve this equation, let us assume that $z$ is a real paragrassmann number \cite{kn:Fi,kn:Ru}, that is $z^{k_0}=0,$ for some integer $k_0 \ge 1.$ A detailed procedure of resolution of this equation is given in the Appendix \ref{sec-appa}. Let us notice that the case $k_0=1,$ i.e., $z=0,$ is somewhat trivial since the eigenfunctions $\varphi (\xi)$ solving \eqref{fock-para}, are given by the standard squeezed symbol \eqref{sym-squee}. When $k_0=2,$ or $z^2=0,$ i.e., when $z$ is a odd Grassmann number \cite{kn:Dewit,kn:Corn}, the eigenvalue equation \eqref{fock-para} becomes \begin{equation} \left( (1- \mu z \xi) {d \over d \xi} + \mu \, \xi \right) \varphi (\xi) = \lambda \varphi (\xi ), \qquad \mu, \lambda \in {\mathbb C}. \end{equation} There are two independent solutions (see Appendix \ref{sec-appa}). The normalizable solution of this eigenvalue equation, is given by the deformed squeezed symbol \begin{equation} \varphi (\lambda, \mu, z) (\xi) = C_0 (\lambda,\mu,z) \left[1 + z \mu \left( \lambda {\xi^2 \over 2} - \mu {\xi^3 \over 3} \right)\right] \exp\left(\lambda \xi - {\mu \over 2} \xi^2 \right). \label{nor-sol-uno}\end{equation} A normalized version of these states, in the Fock space representation, is given by \begin{eqnarray} \nonumber |\varphi \rangle &=& {\tilde \Omega} (\delta, \phi, \beta, \theta) \left[1 + z \delta \left( {\delta e^{2i \phi} \over 3} {(a^\dagger)}^3 - {\beta e^{i ( \theta + \phi )} \over 2} {(a^\dagger)}^2 \right) \right] \\ & & S \left( - \arctan (\delta )e^{i \phi} \right) D \left( {\beta e^{i\theta} \over \sqrt{1 - \delta^2}} \right) |0 \rangle, \end{eqnarray} where $\lambda$ and $\mu$ have been chosen as in the preceding subsection and \begin{eqnarray} {\tilde \Omega} (\delta, \phi, \beta, \theta) &=& 1 - {z \delta \beta \over 2 {(1-\delta^2)}^2} \Biggl[ \left( 2 \delta^2 + \beta^2 \left( {1 + \delta^2 \over 1-\delta^2} \right) \right) \cos(\theta- \phi) \nonumber \\ &-& \delta \left( 1 + \delta^2 + { 2 \beta^2 \over 1-\delta^2 } \right) \cos \theta \nonumber \\ &+& \delta^2 \beta^2 \left( 1 + {2 \delta^2 \over 3(1-\delta^2)} \right) \cos(\phi - 3 \theta) - {2\delta \beta^2 \over 3(1-\delta^2)} \cos(2 \phi - 3 \theta) \Biggr].\end{eqnarray} When $k_0 = 3, $ or $z^3=0,$ the eigenvalue equation \eqref{fock-para} becomes the second order differential equation \begin{equation} \label{sec-order-eq}\left( {1\over 2} \mu z^2 \xi {d^2 \over d\xi^2} + (1- \mu z \xi) {d \over d \xi} \right) \varphi (\xi) = (\lambda - \mu \, \xi ) \varphi (\xi ), \qquad \mu, \lambda, \, \in {\mathbb C}. \end{equation} According to the results obtained in Appendix \ref{sec-appa}, the general solution of this equation can be expanded in the form \begin{equation} \label{sol-type-3} \varphi (\xi) = \varphi_0 (\xi) + z \varphi_1 (\xi) + z^2 \varphi_2 (\xi), \end{equation} with \begin{eqnarray} \varphi_0 (\xi) &=& C_0 \exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right) , \\ \varphi_1 (\xi) &=& \left[ \mu \left( \lambda {\xi^2 \over 2} - \mu {\xi^3 \over 3} \right) C_0 + C_1 \right] \exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right) , \\ \nonumber \varphi_2 (\xi) &=& \Biggl[ \left( \mu (\mu - \lambda^2 ) {\xi^2 \over 4} + {2\over 3} \mu^2 \lambda \xi^3 + \mu^2 (\lambda^2 - 3 \mu ) {\xi^4 \over 8} - \lambda \mu^3 {\xi^5 \over 6} + \mu^4 {\xi^6 \over 18} \right) C_0 \\ &+& \mu \left( \lambda {\xi^2 \over 2} - \mu {\xi^3 \over 3} \right) C_1 + C_2 \Biggr] \exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right), \end{eqnarray} where $C_0, $ $C_1$ and $C_2$ are arbitrary integration constants. Three independent solutions may thus be obtained. The first one is obtained by taking $C_1 = C_2 =0.$ We get \begin{eqnarray} \nonumber \varphi (\xi) &=& C_0 \Biggl[ 1 + z \mu \left( \lambda {\xi^2 \over 2} - \mu {\xi^3 \over 3} \right) + z^2 \Biggl( \mu (\mu - \lambda^2 ) {\xi^2 \over 4} + {2\over 3} \mu^2 \lambda \xi^3 \\ \nonumber &+& \mu^2 (\lambda^2 - 3 \mu ) {\xi^4 \over 8} - \lambda \mu^3 {\xi^5 \over 6} + \mu^4 {\xi^6 \over 18} \Biggr) \Biggr] \exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right)\\ &=& C_0 \exp\left[ z \mu \left( \lambda {\xi^2 \over 2} - \mu {\xi^3 \over 3} \right) + z^2 f(\xi)\right]\exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right), \end{eqnarray} where \begin{equation} f(\xi) = \Biggl( \mu (\mu - \lambda^2 ) {\xi^2 \over 4} + {2\over 3} \mu^2 \lambda \xi^3 - 3 \mu^3 {\xi^4 \over 8} \Biggr). \end{equation} This solution can be normalized and represents a second order paragrassmann deformation of squeezed states associated to the standard harmonic oscillator. The other independent solutions are given respectively by \begin{equation} \varphi (\xi) = C_1 \, z \Biggl[ 1 + z \mu \left( \lambda {\xi^2 \over 2} - \mu {\xi^3 \over 3} \right) \Biggr] \exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right) \end{equation} and \begin{equation} \varphi (\xi) = C_2 \, z^2 \exp\left(\lambda \xi - \mu {\xi^2 \over 2}\right). \end{equation} These solutions can not be normalized since $z^k, \, k=1,2, $ are not invertible paragrassmann numbers and $ z^k = 0, \, k= 3,4, \ldots. $ The higher order paragrassmann deformations of the squeezed states associated to the standard harmonic oscillator can be obtained following a similar procedure (see Appendix \ref{sec-appa}). In the case of eigenvalue equation \eqref{eigen-30}, the differential equation to solve is given by \begin{equation} \label{eigen-30-diff} \left( {d\over d\xi} + \nu e^{- z {d\over d\xi}} \right) \varphi (\xi) = \lambda \varphi (\xi), \qquad \nu, \lambda, \, \in {\mathbb C}.\end{equation} Proceedings as before and considering the results of Appendix \ref{sec-appa}, the normalizable solutions of this last equation, when $k_0= 1,2,3,$ are given respectively by the deformed coherent symbols \begin{equation} \varphi^{(1)}(\xi) = C_0 \exp\biggl( (\lambda - \nu) \xi \biggr), \end{equation} \begin{equation} \varphi^{(2)}(\xi)= C_0 \left[ 1 + z (\lambda - \nu) \nu \xi \right] \exp\biggl( (\lambda -\nu) \xi\biggr)\end{equation} and \begin{eqnarray} \varphi^{(3)} (\xi)&=& C_0 \biggl\{1 + z (\lambda - \nu) \nu \xi + z^2 \biggl[\left( \frac{{\lambda }^2\,\nu }{2} \nonumber + 2\,\lambda \,{\nu }^2 - \frac{3\,{\nu }^3}{2} \right) \xi \\ &+& \left( \frac{{\lambda }^2\,{\nu }^2 }{2} - \lambda \,{\nu }^3 + \frac{{\nu }^4}{2}\right) \, {\xi }^2 \biggr] \biggr\} \exp\biggl( (\lambda -\nu) \xi \biggr). \end{eqnarray} Theses solutions can be normalized and represent zero, first and second order paragrassmann deformations, respectively, of coherent states associated to the standard harmonic oscillator. For higher values of $ k_0,$ we must proceed as in Appendix \ref{sec-appa}. \subsection{Deformed algebra eigenstates for \mathversion{bold} ${\cal U}_{z,p} (h(2))$} It is interesting to compute the AES associated to ${\cal U}_{z,p} (h(2)), \ z,p \ne 0, $ and compare it with the ones associated to ${\cal U}_{z,0} (h(2)).$ As we have noticed in section \ref{sec-two}, these quantum algebras are isomorphic in the sense that there is a nonlinear change of basis transforming one to the other. In general, the existence of this isomorphism does not imply the existence of an internal homomorphism at the AES level. Indeed, by definition, the eigenvalue equation determining the set of AES deals with an arbitrary linear combination of the deformed algebra generators, then from the inverses of transformations \eqref{optilde} and the solvable structure of the commutation relations \eqref{com-he1tilde}, it is impossible to find an internal homomorphism, at the AES level, transforming the eigenvalue equation with $z,p \ \ne 0 $ to the eigenvalue equation with $z \ne 0, p=0.$ To see that, in this section, we consider the two parameters deformed algebra ${\cal U}_{z,p} (h(2))$ as given by \eqref{com-he1}, and compute the AES using the particular realization \eqref{op-def-one-zp}. More precisely, we have to solve the eigenvalue equation \begin{equation} \left[ e^{ z a^\dagger} \sqrt{1+ {\left({p \over 2} e^{z a^\dagger}\right)}^2 } a + \mu a^\dagger + {2 \nu \over p} \sinh^{-1} \left( {p\over 2} e^{z a^\dagger} \right) \right] |\psi\rangle = \lambda |\psi\rangle, \qquad \mu,\nu,\lambda \in {\mathbb C}. \end{equation} In the Bargmann representation, this equation becomes the first order differential equation \begin{equation}\left[ e^{ z \xi} \sqrt{1+ {\left({p \over 2} e^{z \xi}\right)}^2 } {d\over d\xi} + \mu \xi + {2 \nu \over p} \sinh^{-1} \left( {p\over 2} e^{z \xi} \right) \right] \psi(\xi) = \lambda \psi (\xi), \qquad \mu,\nu,\lambda \in {\mathbb C}. \end{equation} When $z=0,$ we easily get the standard squeezed symbols \begin{equation} \psi_{o,p} (\xi) = C_0 (p,\lambda , \mu ,\nu) \exp\left[\left(\lambda - {2\nu \over p} \sinh^{-1} (p/2)\right) \xi - \mu {\xi^2 \over 2} \right]. \end{equation} These symbols correspond to the Bargmann representation of the AES associated to the deformed quantum Heisenberg algebra realization \eqref{def1}. Moreover, when $p$ goes to zero, these symbols becomes the standard squeezed symbols associated to $h(2).$ When $z \ne 0,$ making the change of variable $\zeta = e^{z\xi}, $ rearranging the terms and using the method of characteristics curves to separate the differentials, we get \begin{equation} {d\psi \over \psi} (\zeta) = {\left[ \lambda - {\mu \over z} \ln \zeta - {2\nu \over p} \sinh^{-1} {\left( p \zeta \over 2\right)} \right] \over z \, \zeta^2\, \sqrt{1 + {p^2 \zeta^2 \over 4} }} d\zeta.\end{equation} Integrating both sides of this equation and then exponentiating, we get \begin{eqnarray} \psi_{z,p} (\zeta) &=& C_0 (\lambda, \mu,\nu ; z, p ) \, \exp\Biggl[ {\sqrt{1 + {p^2 \zeta^2 \over 4}} \over z^2 \zeta} \biggl( (1+ \ln \zeta) \mu - \lambda z + {2 \nu z \over p} \sinh^{-1}(\frac{p \,\zeta}{2}) \biggr) \nonumber \\ &-& {\mu p \over 2 z^2} \sinh^{-1}(\frac{p \,\zeta}{2}) - {\nu \over z} \ln \zeta \Biggr]. \label{so--gen-zp}\end{eqnarray} This result includes the ones obtained for \eqref{eigen-10} when $p$ goes to zero. Moreover, when we set also $\nu=0,$ we regain \eqref{solgen-varphi2}. \subsubsection{Perturbed two parameters deformation coherent and squeezed states} Up to first order of approximation in $z$ and $p^2,$ the deformed symbol \eqref{so--gen-zp} writes \begin{eqnarray} \psi_{z,p} (\xi) &\approx& \tilde C_0 (\lambda, \mu,\nu ; z, p ) \biggl[1 + z \left( {\mu \xi^3 \over 3} - {\lambda \xi^2 \over 2} \right) \nonumber \\&+& { p^2 \over 4} \left( {\mu \xi^2 \over 4} - \left({\lambda \over 2} - {\nu \over 3} \right)\xi \right) \biggr] \, \exp\left( (\lambda - \nu) \xi - {1\over 2} \mu \xi^2 \right). \end{eqnarray} In the case $\mu = \delta e^{i \phi},$ $\lambda = \beta e^{i \theta}$ and $\nu = - \gamma e^{i \eta},$ where $\gamma \ge 0,$ a normalized version of these states, in the Fock representation, is given by \begin{eqnarray} \nonumber |\psi \rangle &\approx & {\tilde \Omega} (\delta, \phi, \beta, \theta, \gamma, \eta) \biggl\{1 + \left[ z \left( {\delta e^{i \phi} \over 3} {(a^\dagger)}^3 - {\beta e^{i \theta} \over 2} {(a^\dagger)}^2 \right) \right] \\ \nonumber &+& {p^2 \over 4} \left[{\delta e^{i \phi} \over 4} {(a^\dagger)}^2 - \left({\beta e^{i \theta} \over 2} + {\gamma e^{i \eta} \over 3} \right) a^\dagger \right]\biggr\} \\ & & S \left( - \arctan (\delta )e^{i \phi} \right) D \left( {{\tilde \beta} e^{i{\tilde \theta}} \over \sqrt{1 - \delta^2}} \right) |0 \rangle, \label{def-squee-sta}\end{eqnarray} where \begin{eqnarray} {\tilde \Omega} (\delta, \phi, \beta, \theta, \gamma, \eta ) &=& 1 + {z \over 2 {(1-\delta^2)}^2} \Biggl\{ {\tilde \beta} \Biggl[ \left( 2 \delta^2 + {\tilde \beta}^2 \left( {1 + \delta^2 \over 1-\delta^2} \right) \right) \cos{\tilde \theta} \nonumber \\ &-& \delta \left( 1 + \delta^2 + { 2 {\tilde \beta}^2 \over 1-\delta^2 } \right) \cos(\phi - {\tilde \theta}) \nonumber \\ &+& \delta^2 {\tilde \beta}^2 \left( 1 + {2 \delta^2 \over 3(1-\delta^2)} \right) \cos(2 \phi - 3 {\tilde \theta}) - {2\delta {\tilde \beta}^2 \over 3(1-\delta^2)} \cos(\phi - 3 {\tilde \theta}) \Biggr] \nonumber \\\nonumber &-& \gamma \Biggl[ {\tilde \beta}^2 \cos(\eta - 2 {\tilde \theta}) - \delta (2{\tilde \beta}^2 + 1 - \delta^2 ) \cos(\eta - {\tilde \theta}) \\ \nonumber &+& \delta^2 {\tilde \beta}^2 \cos(2 \phi- \eta - 2 {\tilde \theta}) \Biggr]\Biggr\} - {p^2 \over 16 {(1-\delta^2)}^2} \Biggl\{ \delta {\tilde \beta}^2 (3 \cos(\phi - 2 {\tilde \theta}) \\ &+& {2\gamma\over 3} {\tilde \beta} (1 -\delta^2) \biggl( \cos(\eta - {\tilde \theta}) + \delta \cos (\phi -\eta - {\tilde \theta}\biggr) - 2 {\tilde \beta}^2 - \delta^2 + \delta^4 \Biggr\}, \nonumber \\ \end{eqnarray} where \begin{equation} {\tilde \beta} = \sqrt{\beta^2 + \gamma^2 + 2 \beta \gamma \cos(\eta-\theta)}, \qquad {\tilde \theta}= \tan^{-1} \left( { \beta \sin \theta + \gamma \sin \eta \over \beta \cos \theta + \gamma \cos \eta } \right). \end{equation} We notice that, in the case $\gamma =0$ and $p=0,$ these normalized states become the normalized states given in equation \eqref{nor-z-def}. \section{Some properties of the deformed states} \label{sec-cuatro} In this section, we will give some properties of the deformed states found in preceding section. From Fock space representation, we will deduce the physical quantities $X$ and $P,$ representing the position and linear momentum of a particle, respectively, and compute the corresponding dispersions in both the perturbed deformed states associated to ${\cal U}_{z,p} (h(2))$ and the deformed states associated to $ {\tilde {\cal U}}_{z,0} (h(2)).$ We will also connect the last states with an $\eta$-pseudo Hermitian Halmiltonian \cite{kn:AMostafazadeh}. \subsection{Squeezing properties} \label{sec-xp-squeezed} First, let us consider the squeezing properties of $X$ and $P.$ In the Fock space representation, these quantities are given by the hermitian operators (we have assumed that the mass, angular frequency and Planck's constant are all equal to 1) \begin{equation} X = {(a + a^\dagger) \over \sqrt{2}}, \qquad P= i {(a^\dagger - a )\over \sqrt{2}}. \label{xp-def} \end{equation} They verify the canonical commutation relation \begin{equation} [X,P] = i I . \end{equation} The dispersion of these quantities, computed on a specific normalized particle state $|\psi\rangle,$ is defined as \begin{equation} {(\Delta X )}^2 = \langle \psi | X^2 | \psi \rangle - {(\langle \psi | X | \psi \rangle)}^2 \label{x-disper} \end{equation} and \begin{equation} {(\Delta P )}^2 = \langle \psi | P^2 | \psi \rangle - {(\langle \psi | P | \psi \rangle)}^2. \label{p-disper} \end{equation} The product of these dispersions satisfies the Schr\"{o}dinger-Robertson uncertainty relation (SRUR) \cite{kn:SchRo,kn:Mer} \begin{equation} {(\Delta X )}^2 \, {(\Delta P )}^2 \ge {1 \over 4} \biggl( \langle I \rangle^2 + \langle F \rangle^2 \biggr) = {1 \over 4} \biggl( 1 + \langle F \rangle^2 \biggr) , \label{SchRo-prin} \end{equation} where $F$ is the anti-commutator $ F = \{ X - \langle X \rangle I, P- \langle P \rangle I\}.$ The mean value of $F$ is a correlation measure between $X$ and $P.$ When $\langle F \rangle = 0 ,$ we regain the standard Heisenberg uncertainty principle. The minimum uncertainty states (MUS) are states that satisfy the equality in \eqref{SchRo-prin}. They are called coherent states when the dispersions of both $X$ and $P$ are the same and squeezed states when these dispersions are different to each other. The states for which the dispersion of $X$ is greater than the one of $P$ are called $X$-squeezed whereas the states for which the dispersion of $P$ is greater than the one of $X$ are called $P$-squeezed. We are interested to compute the dispersions of $X$ and $P,$ in the deformed squeezed states \eqref{def-squee-sta}, when $\nu=0,$ or $\gamma=0.$ More precisely, we want to study the effect of the deformation parameters on the squeezed properties of these quantities. As we have seen, when $z$ and $p$ go to zero, the states \eqref{def-squee-sta} becomes the standard harmonic oscillator squeezed states. In such a case, we know that the dispersions of $X$ and $P$ are independent of $\lambda = \beta e^{i \theta },$ and given by \cite{kn:NaVh1} \begin{equation} {{(\Delta X )}_0 }^2 = {1 - 2 \delta \cos\phi + \delta^2 \over 2 (1-\delta^2)} \qquad {\rm and} \qquad {{(\Delta P )}_0 }^2 = {1 + 2 \delta \cos\phi + \delta^2 \over 2 (1-\delta^2)}. \end{equation} All these states are MUS, that is, they satisfy the equality in \eqref{SchRo-prin}. When $\gamma=0,$ the square of the mean value of $X,$ in the states \eqref{def-squee-sta}, to first order of approximation in $z$ and $p^2,$ is given by \begin{eqnarray} \langle \psi |X |\psi \rangle^2 & \approx & 2 \biggl(\mathrm{Re\,} \Gamma_{01}\biggr) \, \mathrm{Re\,} \Biggl\{ \biggl( 1+ 4 \epsilon (z,p) \biggr) \Gamma_{01} \nonumber \\\nonumber &+& 2 z \, \Biggl( {\delta e^{-i \phi} \over 3} \Gamma_{04} - {\beta e^{-i \theta} \over 2} \Gamma_{03} + {\delta e^{i \phi} \over 3} \Lambda_{13} - {\beta e^{i \theta} \over 2 } \Lambda_{12} \Biggr) \\ &+& {p^2 \over 2} \, \Biggl( {\delta e^{-i \phi} \over 4} \Gamma_{03} - {\beta e^{-i \theta} \over 2} \Gamma_{02} + {\delta e^{i \phi} \over 4} \Lambda_{12} - {\beta e^{i \theta} \over 2} \Lambda_{11} \Biggr) \Biggr\}, \label{moyx2} \end{eqnarray} where $\epsilon (z,p)= {\tilde \Omega}(\delta,\phi,\beta,\theta,0,0) - 1 $ and $\Gamma_{kl}$ and $\Lambda_{kl},$ $k,l=1,2,\ldots,$ are matrix elements defined in Appendix \ref{sec-appb}. According to \eqref{xp-def}, we have the same expression for the square of the mean value of $P,$ but taking the imaginary part in place of the real part. On the other hand, the mean value of $X^2$ in the states \eqref{def-squee-sta}, to first order of approximation in $z$ and $p^2,$ is given by \begin{eqnarray} \nonumber \langle \psi |X^2 |\psi \rangle & \approx & { 1 \over 2}+ \biggl( 1 + 2 \epsilon (z,p) \biggr) (\Gamma_{11} + \mathrm{Re\,} \Gamma_{02}) \\ \nonumber &+& z \, \mathrm{Re\,} \Biggl( {\delta e^{-i \phi} \over 3} \Gamma_{05} - {\beta e^{-i \theta} \over 2} \Gamma_{04} + {\delta e^{i \phi} \over 3} \Lambda_{23} - {\beta e^{i \theta} \over 2 } \Lambda_{22} \Biggr) \\\nonumber &+& {p^2 \over 4} \, \mathrm{Re\,} \Biggl( {\delta e^{-i \phi} \over 4} \Gamma_{04} - {\beta e^{-i \theta} \over 2} \Gamma_{03} + {\delta e^{i \phi} \over 4} \Lambda_{22} - {\beta e^{i \theta} \over 2} \Lambda_{21} \Biggr) \nonumber \\ & + & z \Biggl( {\delta e^{-i \phi} \over 3} (\Lambda_{41} - \Gamma_{03}) - {\beta e^{-i \theta} \over 2} (\Lambda_{31} - \Gamma_{02}) + {\delta e^{i \phi} \over 3}(\Lambda_{14} - \Lambda_{03})\nonumber \\ \nonumber &-& {\beta e^{i \theta} \over 2 } (\Lambda_{13} - \Lambda_{02})\Biggr) + {p^2 \over 4} \Biggl( {\delta e^{-i \phi} \over 4} ( \Lambda_{31} - \Gamma_{02}) - {\beta e^{-i \theta} \over 2} ( \Lambda_{21} - \Gamma_{01}) \nonumber \\&+& {\delta e^{i \phi} \over 4} (\Lambda_{13} - \Lambda_{02}) - {\beta e^{i \theta} \over 2} (\Lambda_{12} - \Lambda_{01}) \Biggr). \label{x2moy} \end{eqnarray} Again, according to \eqref{xp-def}, we have the same expression for the mean value of $P^2,$ but taking the negative of the real part in place of the real part. Combining \eqref{moyx2} with \eqref{x2moy}, according to equation \eqref{x-disper}, we get the dispersion of $X$. In the same way, we can obtain the dispersion of $P.$ Inserting the matrix elements $\Gamma_{ij}$ and $\Lambda_{ij},$ as given in the Appendix \ref{sec-appb}, we can compute these dispersions explicitly. Figure \ref{fig:varXPz0.0-0.020p=0} show the dispersions of $X$ and $P$ in the minimum uncertainty squeezed states in dashed lines, and in the deformed squeezed states in solid lines, as a function of $\phi$ for fixed valued of the parameters $\delta, \beta, \theta$ and $p$ ($\delta=0.5, \, \beta=2.0, \, \theta =0.8 \, \pi, p=0.00 $) and for special values of $z=0.0010, 0.0015, 0.0020$ (from the smaller to the greater gray level). \begin{figure}[h] \centering \begin{picture}(31.5,21) \put(0,0){\framebox(31.5,21){}} \put(1,2.1){\includegraphics[width=70mm]{nalvar1.eps}} \put(28.9,3){\scriptsize{$\phi$}} \put(27.6,5){\scriptsize{${(\Delta P)}^2$}} \put(26.9,5){\vector(-1,0){3}} \put(25.6,15.4){\scriptsize{${(\Delta X)}^2$}} \put(25.4,15.4){\vector(-1,0){3}} \end{picture} \caption{Graphs of the dispersions of $X$ and $P$ as functions of $\phi$ for $p=0$ and $z=0.000,0.0010,0.0015,0.0020.$ } \label{fig:varXPz0.0-0.020p=0} \end{figure} We observe that, as a consequence of the small deformations in the parameters $z$ the squeezing properties of $X$ and $P$ have not been essentially changed. Thus, in all the cases, we have $P$--squeezed states when $- {\pi \over 2} < \phi < {\pi \over 2}, $ and $X$--squeezed states when ${\pi \over 2} < \phi < {3 \pi \over 2}. $ Also we observe that the product of the dispersions of $X$ and $P$ in the deformed squeezed states, for a given value of $\phi,$ is always greater than the product of the dispersions in the minimum uncertainty states, as required by the SRUR. These difference is more remarkable for values of $\phi$ in the range ${\pi \over 2} \le \phi < {3\pi \over 2}. $ Let us notice that when $\phi = \pm {\pi \over 2},$ the MUS are coherent states, in the sense of the SRUR, i,e., the dispersion of $X$ and $P,$ are the same. Indeed, in all these cases, ${{(\Delta X)}_0}^2 = {{(\Delta P)}_0}^2 = 0.83. $ This value is conserved by the product of the dispersions of $X$ and $P$ in the deformed squeezed states when $\phi = - {\pi \over 2},$ but when $ \phi = {\pi \over 2},$ it grows quickly as $z$ increases. Figure \ref{fig:varXPz-3p0-0.11} show the dispersions of $X$ and $P$ in the minimum uncertainty squeezed states in dashed lines, and in the deformed squeezed states in solid lines, as a function of $\phi$ for fixed valued of the parameters $\delta, \beta, \theta$ and $z$ ($\delta=0.5, \, \beta=2.0, \, \theta =0.8 \, \pi, z=0.0030 $) and for special values of $p=0.00, 0.06, 0.11$ (from the greater to the smaller gray level). \begin{figure}[h] \centering \begin{picture}(31.5,21) \put(0,0){\framebox(31.5,21){}} \put(1,2.1){\includegraphics[width=70mm]{nalvar2.eps}} \put(28.9,3){\scriptsize{$\phi$}} \put(27.8,5.5){\scriptsize{${(\Delta P)}^2$}} \put(27.5,5.5){\vector(-1,0){2.5}} \put(25.3,15.4){\scriptsize{${(\Delta X)}^2$}} \put(25,15.4){\vector(-1,0){2.5}} \end{picture} \caption{Graphs of the dispersions of $X$ and $P$ as functions of $\phi$ for $z= 0.0030, p=0.00,0.06,0.11, \beta=2.0, \theta =0.8 \pi $ and $ \delta= {0.5}.$} \label{fig:varXPz-3p0-0.11} \end{figure} We observe that the product of dispersions of $X$ and $P$ decreases when $p$ increases. Thus the influence of the $p$ parameter on the first order in $z$ deformed states is to reduce the uncertainty product of $X$ and $P$ and to bring closer this quantity to the minimum uncertainty values. \begin{figure}[h] \centering \begin{picture}(31.5,21) \put(0,0){\framebox(31.5,21){}} \put(1,2.1){\includegraphics[width=70mm]{nalvar43.eps}} \put(28.9,3){\scriptsize{$\delta$}} \put(27.6,7){\scriptsize{${(\Delta P)}^2$}} \put(26.9,7){\vector(-1,-1){2}} \put(14.9,13.4){\scriptsize{${(\Delta X)}^2$}} \put(16.8,12.9){\vector(1,-1){2}} \end{picture} \caption{Graphs of the dispersions of $X$ and $P$ as functions of $\delta$ for $z= 0.0025, p=0.01, \beta=2.0, \theta =0.8 \pi $ and $ \phi= {\pi\over 6} .$} \label{fig:artvap01} \end{figure} Figure \ref{fig:artvap01} shows the typical behavior of the dispersions of $X$ and $P$ in the minimum uncertainty squeezed states in dashed lines, and in the deformed squeezed states in solid lines, as a function of $\delta $ for $\phi=0.5, \, \beta=2.0, \, \theta =0.8 \, \pi, z=0.0025 $ and $p=0.001.$ We observe again that, as a consequence of the small deformations in $z$ and $p,$ the squeezing properties of $X$ and $P$ have not been essentially changed. Thus, the figure shows the behavior of $P$--squeezed and $P$-deformed squeezed states. When $ 0< \delta \lesssim 0.75,$ the product of the dispersions of $X$ and $P,$ in the deformed squeezed states is always greater than the corresponding product in the minimum uncertainty squeezed states, as required by the SRUR. For higher values of $\delta,$ only the dashed lines represent the true behavior of the dispersions of $X$ and $P.$ Indeed, the approximation for the deformed squeezed states, in this region, is not valid. These states are no longer normalizable. \subsection{General formulas for the dispersions of $X$ and $P$ in the \boldmath $z$ deformed states} The mean values of $X^k, k=1,2, \ldots,$ in the states \eqref{re-norm-squee} can be expressed in the forme \begin{equation} \biggl. \langle \varphi | X^{k} | \varphi \rangle ={ {\partial^k \over \partial \tau^k} \widetilde{\langle \varphi} | e^{\tau X} | \widetilde{\varphi \rangle} \biggr|_{\tau=0} \over \widetilde{\langle \varphi } | \widetilde{ \varphi \rangle}} ={ {\partial^k \over \partial \tau^k}\biggl\{ e^{- {\tau^2 \over 4}} \ \widetilde{\langle \varphi} | e^{{\tau\over \sqrt2} a} e^{{\tau\over \sqrt2} a^\dagger}| \widetilde{\varphi \rangle} \biggr\}\biggr|_{\tau=0} \over \widetilde{\langle \varphi } | \widetilde{ \varphi \rangle}}, \end{equation} where \begin{equation} \widetilde{|\varphi \rangle} = \exp\left( e^{-z a^\dagger} {(\mu - \lambda z + \mu z a^\dagger) \over z^2} \right) |0\rangle. \end{equation} Inserting these results into \eqref{x-disper} and evaluating we get \begin{equation} \label{x-disper-tau} {(\Delta X )}^2 = - { 1 \over 2} + { {\partial^2 \over \partial \tau^2} \widetilde{\langle \varphi} | e^{{\tau\over \sqrt2} a} e^{{\tau\over \sqrt2} a^\dagger}| \widetilde{\varphi \rangle} \biggr|_{\tau=0} \over \widetilde{\langle \varphi } | \widetilde{ \varphi \rangle}} - {\Biggl( { {\partial \over \partial \tau } \widetilde{\langle \varphi} | e^{{\tau\over \sqrt2} a} e^{{\tau\over \sqrt2} a^\dagger}| \widetilde{\varphi \rangle} \biggr|_{\tau=0} \over \widetilde{\langle \varphi } | \widetilde{ \varphi \rangle}} \Biggr)}^2 . \end{equation} To compute the matrix element $\widetilde{\langle \varphi} | e^{{\tau\over \sqrt2} a} e^{{\tau\over \sqrt2} a^\dagger} | \widetilde{\varphi \rangle}, $ we can firstly write \begin{equation} e^{{\tau\over \sqrt2} a^\dagger}| \widetilde{\varphi \rangle} = \sum_{n=0}^{\infty} C_n (\tau) |n\rangle \end{equation} and then to compute the coefficients $C_n (\tau), \ n=0,1,2,\ldots , $ in the Bargman representation, in the same way as we have do it in section \eqref{sub-sec-coh-squee}. That is \begin{equation} \label{varphi-tilde-modif} \widetilde{\langle \varphi} | e^{{\tau\over \sqrt2} a} e^{{\tau\over \sqrt2} a^\dagger} | \widetilde{\varphi \rangle} = \sum_{n=0}^{\infty} {\bar C}_{n} (\tau) C_{n} (\tau), \end{equation} where \begin{equation} \label{cn-tau} C_n (\tau) = {1\over \sqrt{n!}} \sum_{r=0}^{n} {n \choose r} {\left({\tau \over \sqrt2}\right)}^r \ z^{n-r} \ \sum_{k=0}^{\infty} \sum_{m=0}^{{\tilde k}_< } {n-r \choose m} {{(- k)}^{n-r-m} \over (k-m)!} {\left({\mu \over z^2 }\right)}^m \ {\left({\mu \over z^2} - {\lambda \over z}\right)}^{k-m}, \end{equation} with ${\tilde k}_< $ the minimum between $k$ and $n-r .$ Inserting \eqref{varphi-tilde-modif} into \eqref{x-disper-tau} and evaluating again we get \begin{eqnarray} \nonumber {(\Delta X )}^2 &=& - { 1 \over 2} + {\sum_{n=0}^\infty \Biggl. \biggl[ {\bar C}_n (\tau) C_n^{\prime \prime} (\tau) + {\bar C}_n^{\prime \prime} C_n (\tau) + 2 {\bar C}_n^{\prime} C_n^{\prime} (\tau) \biggr] \Biggr|_{\tau=0} \over \sum_{n=0}^\infty {\bar C}_n (0) C_n(0) } \\&-& {\Biggl( {\sum_{n=0}^\infty \Biggl. \biggl[ {\bar C}_n (\tau) C_n^{\prime} (\tau) + {\bar C}_n^{\prime} (\tau) C_n (\tau) \biggr] \Biggl|_{\tau=0}\over \sum_{n=0}^\infty {\bar C}_n (0) C_n(0) }\Biggr)}^2 ,\label{gen-for-x} \end{eqnarray} where, for instance, $C_{n}^{\prime} (\tau) = {d C_n \over d\tau} (\tau) . $ From \eqref{cn-tau}, we obtain \begin{equation} C_{n}^{\prime} (0) ={1\over \sqrt{n!}} {n \choose 1} {1 \over \sqrt2} \ z^{n-1} \ \sum_{k=0}^{\infty} \sum_{m=0}^{{\tilde k}_1 } {n-1 \choose m} {{(- k)}^{n-1-m} \over (k-m)!} {\left({\mu \over z^2 }\right)}^m \ {\left({\mu \over z^2} - {\lambda \over z}\right)}^{k-m}, \end{equation} when $n= 1, 2, \ldots, $ with ${\tilde k}_1 $ the minimum between $k$ and $n-1,$ \begin{equation} C_{n}^{\prime \prime} (0) ={1\over \sqrt{n!}} {n \choose 2} \ z^{n-2} \ \sum_{k=0}^{\infty} \sum_{m=0}^{{\tilde k}_2 } {n-2 \choose m} {{(- k)}^{n-2-m} \over (k-m)!} {\left({\mu \over z^2 }\right)}^m \ {\left({\mu \over z^2} - {\lambda \over z}\right)}^{k-m}, \end{equation} when $n= 2, 3, \ldots, $ with ${\tilde k}_2 $ the minimum between $k$ and $n-2,$ and \begin{equation} C_0^\prime (0) = C_0^{\prime \prime} (0) = C_1^{\prime \prime} (0)=0.\end{equation} The formula to the dispersion of $P$ can be obtained from \eqref{gen-for-x} changing the $\tau $ argument of $C_n (\tau) $ by $ i \tau $ and then deriving and evaluating to $\tau = 0.$ Thus, dispersions formulas of $X$ and $P$ at all order in $z$ can be obtained. The first order perturbation formulas of these dispersions must correspond to the dispersions obtained in the preceding subsection, in the limit when $p$ goes to zero. \subsection{\boldmath $\eta$-pseudo Hermitian and Hermitian Hamiltonians} In this section we show that the subset of deformed coherent states \eqref{eigen-10-aes}, corresponding to the eigenvalue $\lambda=0,$ are the coherent states associated to an $\eta$-pseudo Hermitian Hamiltonian \cite{kn:AMostafazadeh} but also, up to a similarity transformation, the coherent states associated to a Hermitian Hamiltonian, both isospectral to the harmonic oscillator Hamiltonian. Indeed, when $\lambda=0,$ the eigenstates \eqref{eigen-10-aes} correspond to the solutions of the eigenvalue equation \begin{equation} \label{eigen-10-reduc} {\cal A} |\psi \rangle = - \nu |\psi \rangle, \qquad \nu \in {\mathbb C}, \end{equation} where ${\cal A} = a + \mu a^\dagger e^{- z a^\dagger}. $ These solutions can be written in the form \begin{equation} \label{eigenatilde} |\psi ; - \nu \rangle = {\tilde N}_0 (\mu, - \nu, z ) \ G (\mu , z ) \ e^{- \nu a^\dagger} |0\rangle, \end{equation} where \begin{equation} G (\mu, z ) = \exp \left( - \mu \ \sum_{k=0}^{\infty} {{(-z a^\dagger )}^k \over k!} {{(a^\dagger)}^2 \over (k+2)} \right) , \label{eigen-aes-reduc} \end{equation} and $ {\tilde N}_0 \, (\mu, - \nu, z )$ is a normalization constant. Let us now to define the operator \begin{equation} \label{H-pseudo} {\cal H} = G \ a^\dagger a \ G^{-1}, \end{equation} which satisfies \begin{equation} {\cal H}^\dagger = \eta {\cal H} \eta^{-1},\end{equation} where $\eta$ is the hermitian operator \begin{equation} \eta (\mu , z) = {(G^{-1})}^\dagger G^{-1}. \end{equation} Thus ${\cal H}$ is an $\eta$-pseudo Hermitian Hamiltonian \cite{kn:AMostafazadeh}. Moreover, as \begin{equation} G a^\dagger G^{-1} = a^\dagger, \qquad G a G^{-1} = {\cal A}, \end{equation} we get \begin{equation} \label{exp-H-pseudo} {\cal H} = a^\dagger {\cal A} = a^\dagger \left( a + \mu a^\dagger e^{- z a^\dagger} \right) = a^\dagger a + \mu e^{- z a^\dagger} {(a^\dagger)}^2 .\end{equation} On the other hand, by construction, it is easy to verify that \begin{equation}\label{com-h2-hab} [ {\cal H}, {\cal A}]= - {\cal A}, \qquad [{\cal H}, a^\dagger]= a^\dagger, \qquad [{\cal A}, a^\dagger]=1 \end{equation} and \begin{equation} \label{h-on-ezero}{\cal H} |E_0 \rangle = 0, \end{equation} where \begin{equation} |E_0 \rangle ={\tilde N}_0 \, (\mu, 0, z ) G (\mu, z) |0\rangle. \end{equation} This state is thus an eigenstate of ${\cal A}$ corresponding to the eigenvalue $\nu = 0.$ Thus, according to \eqref{com-h2-hab} and \eqref{h-on-ezero}, the hamiltonian ${\cal H}$ is isospectral to the harmonic oscillator Hamiltonian. ${\cal A}$ represents an annihilation operator for this system and their eigenstates \eqref{eigenatilde} are the associated coherent states of ${\cal H}.$ Let us mention that $\cal H$ verifies all the useful properties of pseudo-Hermitian operators \cite{kn:Mostafazadeh-2004}. For instance, $\cal H$ is Hermitian on the physical Hilbert space $\mathfrak{H}_{{\rm phys}}$ spanned by their corresponding eigenstates $ | \psi_n \rangle \propto {(a^\dagger)}^n G |0\rangle, \ n=0,1,2,\ldots,$ endowed with the positive-definite inner product $\langle \cdot | \eta \ \cdot \rangle.$ Also, $\cal H$ may be mapped to a Hermitian Hamiltonian ${\tilde {\cal H}}$ by a similarity transformation ${\tilde {\cal H}}= {\hat \rho} {\cal H} {\hat \rho}^{-1}, $ where ${\hat \rho}(\mu,z)= \sqrt{\eta(\mu,z)}= \sqrt{{G^{-1}}^\dagger G^{-1}},$ is a Hermitian operator on a Hilbert space $\mathfrak{H}$ formed of same vectorial space $\mathfrak{H}_{{\rm phys}} $ but endowed with the original inner product $\langle \cdot | \cdot \rangle.$ Thus, in our case, according to \eqref{H-pseudo}, the Hermitian Hamiltonian ${\tilde {\cal H}},$ is unitarily equivalent to the standard harmonic oscillator Hamiltonian and is given by \begin{equation} \label{hami-tilde} {\tilde {\cal H}} = {\hat \rho} \ G \ a^\dagger a \ G^{-1} \ {\hat \rho}^{-1}. \end{equation} Indeed, \begin{equation} {\hat \rho} G {({\hat \rho} G)}^{\dagger} = {\hat \rho} G G^{\dagger} {\hat \rho}^{\dagger} = {\hat \rho} \eta^{-1} {\hat \rho} = {\hat \rho} {({\hat \rho}^2)}^{-1} {\hat \rho} = I \end{equation} and \begin{equation} {({\hat \rho} G)}^{\dagger} {\hat \rho} G = G^{\dagger} {\hat \rho}^{\dagger} {\hat \rho} G = G^{\dagger} {\hat \rho}^2 G = G^{\dagger} {(G^{-1})}^{\dagger} G^{-1} G = I, \end{equation} that is ${({\hat \rho} G)}^{\dagger}={({\hat \rho} G)}^{-1}, $ i.e., $ {\hat \rho} G $ is an unitary operator. Let us notice that in absence of deformation ($z=0$) the operator $\hat \rho $ is given by \begin{equation} \label{ro-u-0} \hat \rho (\mu,0) = \sqrt{\exp\left({\bar \mu} {a^2 \over 2 } \right)\exp\left( \mu { {a^\dagger}^2 \over 2} \right)} = {\biggl[\exp\left( \int_{0}^{1} [{\bar \mu} K_- + \mu K_+ + \varsigma (s) K_3 ] d s \right)\biggr]}^{1\over2}, \end{equation} where $K_- = {a^2 \over 2},$ $K_+ = {{(a^\dagger)}^2 \over 2} $ and $K_3 = {1\over 4}(a a^\dagger + a^\dagger a) $ are the standard bosonic realizations of the $su(1,1)$ Lie algebra generators verifying the commutation relations \begin{equation} [K_-, K_+] = 2 K_3, \qquad [K_3, K_{\pm}] = \pm K_{\pm}\end{equation} and \begin{equation} \varsigma (s) = - 2 {d \over ds} \ln q(s) , \end{equation} where \begin{equation} q(s)= \cosh(|\mu|(1-s)) + |\mu| \sinh(|\mu|(1-s)). \end{equation} In this case, the Hamiltonian \eqref{hami-tilde} becomes \begin{equation} \label{htilde-ro-u-0} {\tilde {\cal H}} = {\hat \rho} (\mu,0) \ [ a^\dagger a + \mu {(a^\dagger)}^2 ]\ {\hat \rho}^{-1} (\mu,0), \end{equation} and represents a Hermitian Hamiltonian describing two photon processes in a single mode. To know the explicit form of this Hamiltonian we must firstly factorize the operator \eqref{ro-u-0} in the form of a product of exponential operators of each $su(1,1)$ generators and then insert it into \eqref{htilde-ro-u-0}. This process requires to solve some Ricatti type differential equations. For small values of $z,$ the Hamiltonian \eqref{hami-tilde} describes corrections to the energy of this system as a consequence of the deformation. In general, when $z\ne 0,$ the Hamiltonian \eqref{hami-tilde} represents multi-photon processes in a single mode. The generalized coherent states associated to the system described by \eqref{hami-tilde}, can be easily obtained from the coherent states associated to the standard harmonic oscillator. Indeed, they are given by \begin{equation} \label{gen-zmunu-ch} | \nu, z, \mu \rangle = {\hat \rho} (\mu, z) \ G(\mu, z) D(\nu) |0\rangle, \end{equation} where $D(\nu)$ is the standard unitary displacement operator defined at the end of subsection \ref{sec-perturba-z-real}. These coherent states correspond to the coherent states associated to the pseudo-Hermitian Hamiltonian \eqref{exp-H-pseudo}, up to the transformation $ {\hat \rho} (\mu, z),$ and are eigenstates of the annihilation operator ${\tilde {\cal A}} = {\hat \rho} (\mu, z)\ {\cal A} \ {\hat \rho}^{-1} (\mu, z)$ corresponding to the eigenvalue $\nu.$ \section{Conclusions} In this paper, we have found some realizations of the deformed quantum Heisenberg Lie algebra ${\cal U}_{z,p} (h(2)),$ in terms of the usual creation and annihilation operators associated with Fock space representation of the standard harmonic oscillator. The method used to get these realizations can be easily applied to find the realizations of other quantum Hopf algebras and super-algebras, such as the bosonic and fermionic oscillators Hopf algebras\cite{HLR96} or the quantum super-Heisenberg algebra, that can also be obtained by using the $R$-matrix approach. We have computed the AES associated to ${\cal U}_{z,p} (h(2)).$ We have seen that the set of AES contains the set of coherent and squeezed states associated to the standard harmonic oscillator system but also a new class of deformed coherent and squeezed states, parametrized by the deformation parameters. We have studied the behavior of the dispersions of the position and linear momentum operators of a particle in a class of perturbed squeezed states and we have compared them with the behavior of these dispersions in the minimum uncertainty squeezed states. Also we have computed these dispersions on the deformed states associated to ${\cal U}_{z,0} (h(2)),$ for all values of the $z$ parameter. To first order in $z,$ these last dispersions reduce to the perturbed ones obtained to ${\cal U}_{z,p} (h(2)),$ when p goes to zero. Besides, we have constructed a $\eta$-pseudo Hermitian Hamiltonian \cite{kn:AMostafazadeh} to which a subset of the set of algebra eigenstates associated to ${\cal U}_{z,0} (h(2)),$ are the coherent states. From this point of view, our deformed states are linked to Hamiltonians presenting important physical aspects \cite{kn:Mostafazadeh-2004}. Indeed, our pseudo-Hermitian Hamiltonian verifies naturally all the properties of pseudo-Hermitian Hamiltonians such as the existence of associated biorthonormal basis, resolution of the identity, positive-definite inner product, physical Hilbert space, unitary and invertible operators mapping the pseudo-Hermitian operators to the Hermitian ones, etc. Thus, with the help of pseudo-Hermitian quantum mechanics techniques we are allowed to compute, for instance, the spectrum, the eigenstates and the associated coherent states of complicated deformed Hermitian Hamiltonians describing multi-photon processes in a single mode. Also, we can compute more easily quantities such as mean values of physical observables and transition amplitudes. Moreover, it could be interesting to know, at least for small values of the deformation parameter $z,$ the explicit form of the resolution of the identity verified by the generalized coherent states \eqref{gen-zmunu-ch}. Indeed, this fact could have important consequences, for instance, in the study of corrections to the time evolution of the quantum fluctuations associated to the quadratures of the position and linear momentum of a system characterized by a Hamiltonian describing one and two photon processes in a single mode \cite{kn:Wei-Min}. This is a no trivial problem and it could be developed elsewhere. On the other hand, we have found new classes of deformed squeezed states, parametrized by a real paragrassmann number, i.e., a number $z$ such that $z^{k_0}=0, $ for some $k_0 \, \in {\mathbb N}. $ These states can be normalized, even if $z$ is considered as a complex paragrassmann number. In this last case, when $k_0 =2,$ we can should interpret $z$ as an odd complex Grassmann number and compare this new classes of deformed squeezed states with the ones associated to the $\eta$-super-pseudo-Hermitian Hamiltonians \cite{kn:NaVh3}. \section*{Acknowledgments} The author would like to thank V. Hussin for valuable discussions and suggestions. He also thanks the referees for valuable suggestions about this article. The author's research was partially supported by research grants from NSERC of Canada. \renewcommand{\theequation}{\thesection.\arabic{equation}}
{ "timestamp": "2005-03-23T17:59:48", "yymm": "0503", "arxiv_id": "math-ph/0503055", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503055" }
\section{Introduction} Freeze out (FO) is a term referring to the stage of expanding or exploding matter when its constituents (particles) loose contact, collisions cease and the local dynamical equilibrium no longer can be maintained. When the local equilibrium is significantly perturbed, the microscopic length and time scales become comparable to the characteristic macroscopic ones and the hydrodynamical approach used to describe the evolution of matter breaks down. In the absence of collisions the momentum distribution of the particles "freezes out", hence the name kinetic freeze out. \\ \indent The final break-up corresponds to a "phase-transition" from an interacting fluid to a non-interacting gas of particles, where the interactions between the constituents ceases suddenly when reaching the "critical" FO temperature of the order of the pion mass $T_{FO}\approx 140$ MeV, as first assumed by Landau \cite{Landau_1}. Consequently, FO is a discontinuity in space-time represented by a space-time boundary or FO hypersurface, taken at the critical temperature [i.e., FO isotherm]. Across such FO hypersurface the properties of matter change suddenly. We denote the two sides of the FO hypersurface as Pre FO and Post FO sides. Originally, the Post FO distribution function was assumed to be an equilibrated J\"uttner distribution function boosted with the local flow velocity on the actual side of the FO hypersurface. This approximation and method corresponds to the so-called "sudden" FO model by Cooper and Frye \cite{Cooper-Frye}. \\ \indent The Cooper-Frye type of FO process is the zero thickness limit of a more realistic, so-called "gradual" FO process, where the FO description applies over a finite space-time domain, i.e. FO layer. Inside the finite FO layer the properties of the matter change gradually trough interactions, while the frozen out particles are formed and emitted at different "temperatures" which correspond to the actual temperature of the interacting matter, gradually during the whole evolution of the matter. \\ \indent The basic philosophy of this paper is similar to the recent work \cite{article_1}, which introduces and analyzes in detail the gradual FO description for space-like FO situations. Through the paper we are going to use the notation from Ref. \cite{article_1}, recall the governing equations and the major results for comparison. We have made this paper sole and complete and widely understandable without the need to consult our previous work where the original ideas were first introduced. \\ \indent Time-like discontinuities represent the overall sudden change in a finite volume where the events happen simultaneously at causally disconnected points of the hypersurface with a time-like normal vector. For example the assumption of instantaneous or isochronous FO [i.e., happening at a constant time in the center of mass system], belongs to this category. \\ \indent The aim of this paper is to present an analyze a simple gradual kinetic FO process with time-like normal vectors. In the first part of this study we introduce and generalize the gradual kinetic FO treatment for a finite time-like FO layer in a fully covariant footing, while in the second part we analyze its outcome. \section{Freeze-out from a finite time-like layer}\label{Fid} The basis of the gradual FO method is to separate the "full" $f=f(x,p)$ distribution function into still interacting and already frozen out parts, $f=f^i + f^f$, and describe the evolution of both components in a self consistent way \cite{grassi_1, cikk_1, cikk_2, cikk_3, hama_1}. This can be achieved by introducing the so-called escape rate, $\mathcal{P}_{esc}(s,p)$, used to drain particles, which no longer collide from the interacting component, $f^i$, and to gradually build up the free component, $f^f$. For the better understanding of the model we use Fig. \ref{figure_1t}, and assume that the FO of particles starts from the inside boundary of the FO layer, $S_1$ (thick line). Within the FO layer of finite thickness, $L$, the density of interacting particles decreases and disappears once we reach the outside boundary, $S_2$ (thin line), of the FO layer. \\ \indent In general, the kinetic description of freeze out leads to a complex multidimensional problem. To clarify the basic properties of the FO process through a finite layer we may essentially reduce the number of variables, assuming that the dominant change in the distribution function happens in the direction of the FO normal vector, while it is negligible along the directions perpendicular to it, (e.g. in a spherically symmetric system the change happens in radial direction, and it is negligible in the perpendicular directions). Thus, the FO process can be effectively described as a one-dimensional process and the space-time domain where such a process takes place can be viewed as a FO layer, where the FO normal vector is tied to the direction of the density decrease arising from velocity divergence at a curved FO surface. \\ \indent If we have a space-like normal vector, $d\sigma_{\mu}=(0,1,0,0)$, the resulting equations can be transformed into a frame where the process is stationary, while in the case of a time-like normal vector, $d\sigma_{\mu}=(1,0,0,0)$, the equations can be transformed into a frame where the process is uniform and time-dependent. In this paper we only discuss FO processes inside a finite time-like layer. \\ \indent Here we recall the governing equations [i.e., Eqs. (13-14)] from Ref. \cite{article_1}, which can be used in both time-like and space-like FO cases. The equations depend on the projection, $s = x^{\mu} d\sigma_{\mu}$, in the direction of the FO normal vector, $d\sigma_{\mu}$, where the four vector $x^{\mu}$ denotes the particle coordinate, having its origin at the inner surface of the FO layer. Thus, \begin{eqnarray}\label{first} \nonumber \partial_s f^{i} (s,p) \! &=& \! - \mathcal{P}_{esc}(s,p) \, f^{i}(s,p) + \frac{f^{i}_{eq}(s,p) - f^{i}(s,p)}{\lambda_{th}} \, , \\ \partial_s f^{f} (s,p) \! &=& \! + \mathcal{P}_{esc}(s,p) \, f^{i}(s,p) \, , \end{eqnarray} where using the relaxation time approximation we ensure that the interacting component approaches the equilibrated J\"uttner distribution, $f^i_{eq}$, with $\lambda_{th}$ relaxation time (or relaxation length in the space-like case). \\ \indent The escape rate, $\mathcal{P}_{esc}(s,p)$, describes the escape of particles from the interacting component into the free component and it is defined as: \begin{equation}\label{esc2} \mathcal{P}_{esc}(s,p) = \frac{1}{\lambda(s)} \left[ \frac{L}{L - s} \, \frac {p^\mu d\sigma_\mu}{p^\mu u_\mu}\right] \Theta(p^{\mu}d\sigma_{\mu}) \, , \end{equation} where the parameter, $L$, is the "proper" thickness of the FO layer and it is an invariant scalar. The proper thickness is analogous to the proper time, that is the time measured in the rest frame of the particle. In our case the local rest frame is the rest frame attached to the FO front (RFF), (see \ref{frames}), and thus the proper "thickness" of the FO layer is the invariant proper time interval between the start of the process and its end, (see Fig. \ref{figure_1t} at point A). Furthermore, $p^{\mu}$ is the four-momentum of particles, $d\sigma_{\mu}$ is the normal in the FO direction, while $u^{\mu}$ is the flow velocity normalized to unity. The initial characteristic time is denoted by $\tau_0$, and the $\Theta(p^{\mu}d\sigma_{\mu})$ function, was first introduced by Bugaev \cite{Bugaev_1} to ensure that all particles leave to the outside. In case of time-like FO this condition is always satisfied and does not lead to any additional constraint. \begin{figure}[t!] \centering \includegraphics[width=8.5cm, height = 8.2cm]{figure_1t.eps} \caption{(Color online) The figure shows a finite FO layer with varying thickness. The normal to a surface element is $d\sigma_{\mu}$, thus between A and B the surface is time-like, [i.e., $d\sigma^{\mu} d\sigma_{\mu} = + 1$], from B down to C it is space-like, [i.e., $d\sigma^{\mu} d\sigma_{\mu} = -1$]. The change from time-like to space-like surface happens where the normal of the surface is light-like, but not necessarily has its origin at the center of the system. The momentum of particles is $p^{\mu}$ and the four vector $x^{\mu}$ denotes the particle coordinate, having its origin at the inner surface of the FO layer. On the time-like FO region all particles emerging from a point on the Pre FO side will propagate to the Post FO side, while the particles originating from the space-like part of the FO surface are divided between Pre FO and Post FO parts. Only those particles cross the surface which have their momentum enclosed by the light cone and the Post FO surface, [i.e., if $p^{\mu}d\sigma_{\mu}>0$].} \label{figure_1t} \end{figure} \\ \indent A qualitative expression of the escape rate for both time-like and space-like FO situations is based on the following simple assumptions. Particles with higher momentum in the FO direction will freeze out first. The particles closer to the outside boundary of the FO layer have a greater chance to freeze out since the probability to find another particle to collide with is smaller as the system became sparser, (given by the $L/(L-s)$ factor). In our special case the FO direction is parallel to the time-axis, thus all particles will freeze out irrespective of their momenta. However, particles emitted at later times will freeze out "faster" since they have less chance to collide with other particles in the diluted system. Please note that, although $\tau_0$ is assumed to be a constant for simplicity, the characteristic FO length is increasing with time or distance such as, $\tau_0(L - s)/L$. The detailed treatment and analysis of the escape rate can be found in \cite{article_1, ModifiedBTE_1, QM05_1}. \\ \indent Now, if we describe the time evolution of the particle FO, then $d\sigma_{\mu} = (1,0,0,0)$, $x^{\mu}d\sigma_{\mu} = t$ and $p^{\mu}d\sigma_{\mu} = p^0$, thus eq. (\ref{first}) leads to \begin{eqnarray}\label{first-rethermalized} \nonumber \partial_t f^{i} &=& - \frac{1}{\tau_0} \left( \frac{L}{L-t} \right)\! \left( \frac{p^0}{p^{\mu} u_{\mu}} \right) f^{i} + \frac{ f^{i}_{eq} - f^{i}}{\tau_{th}} \, , \\ \partial_t f^{f} &=& + \frac{1}{\tau_0} \left(\frac{L}{L-t} \right)\! \left(\frac{p^0}{p^{\mu} u_{\mu}}\right) f^{i} \, , \end{eqnarray} where the interacting component approaches the equilibrated J\"uttner distribution, $f_{eq}$, with $\tau_{th}$ relaxation time. This is a common simplification and the practical reason to use it is to calculate the quantities depending on the equilibrium distribution function. Although the present solution mathematically is achieved taking an infinitely short relaxation time, in reality, one can show, see Ref. \cite{article_1}, that the complete thermalization of the interacting component can be achieved with good accuracy if $\tau_0$ is smaller than $\tau$ by a factor of 2 or more. For a thorough analysis of this approach, see Refs. \cite{article_1, sven} and the Appendix. Here we mention that in our calculations we use only one type of particles, namely massless pions, therefore the chemical composition of our system remains unchanged during the kinetic FO. Furthermore, for simplicity we assume simultaneous chemical and thermal equilibration, thus the chemical potential of the massless pion gas is $\mu=0$ during the FO process. \\ \indent Earlier in Ref. \cite{cikk_5}, the gradual time-like FO description was modeled with equations having a similar form to eqs. (\ref{first-rethermalized}), but it was only treating the simplest case when $u^{\mu}=d\sigma_{\mu}$, while the FO was lasting infinitely long. The model was based on the idea of the boost invariant Bjorken hydrodynamical model \cite{Bjorken}, where the evolution of matter is a function of the proper time, $\tau$, only, while the flow of matter is parallel to the normal vector of the proper time hyperbolas in every point. The covariant equations, eqs. (\ref{first-rethermalized}), return those equations if we change the time variable, $dt= d\tau$, and neglect the $L/(L-t)$ factor. However, the new equations allow $u^{\mu}\neq d\sigma_{\mu}$ and also a possibility to finish the FO process within a given duration. \subsection{Reference frames} \label{frames} Before proceeding further, first we define the reference frames in which our calculations will be handled. \begin{figure}[!hbt] \centering \includegraphics[width = 8.4cm, height = 7.2cm]{figure_2t.eps} \caption{(Color online) A simple FO hypersurface in RFG with coordinates [t,x], where $u^{\mu} = (1,0,0,0)_{RFG}$. The normal vectors of the FO front, $d\sigma_{\mu}$, are time-like at points, A, B, C, while the normal vectors are space-like at points, D, E, F. At point B in RFF with coordinates [t',x'], where $d\sigma_{\mu}=(1,0,0,0)_{RFF}$ and $u^{\mu} = \gamma_{\sigma}(1,-v_{\sigma},0,0)_{RFF}$. Note that RFF moves together with the FO front, while on the figure the origin of RFF is shifted to match the origin of RFG. } \label{figure_2t} \end{figure} \\ \indent On the Pre FO side the matter is parameterized by an equilibrium distribution function, as required by the hydrodynamical description of evolution. The frame where the matter is at rest is the local Rest Frame of the Gas (RFG), where $u^{\mu} = (1,0,0,0)_{RFG}$. On the Post FO side we can use the frame which is attached to the FO front, that is the Rest Frame of the Front (RFF). In RFF the normal vector to the FO hypersurface for the time-like part is always $d\sigma_{\mu} = (1,0,0,0)_{RFF}$, while on the space-like part is always $d\sigma_{\mu} = (0,1,0,0)_{RFF}$. \\ \indent If we are in the RFG then the four-flow is always $u^{\mu} = (1,0,0,0)_{RFG}$. If we take different characteristic points, for example points, A, B and C, on the FO hypersurface then the normal vector is different at different points of the hypersurface in RFG, see Fig. \ref{figure_2t}. To calculate the parameters of the normal vector, $d\sigma_{\mu}$, for different cases in the RFF we make use of the Lorentz transformation. \\ \indent The normal vector of the time-like part of the FO hypersurface may be defined as the local $t'$-axis, while the normal vector of the space-like part may be defined as the local $x'$-axis. This defines the axes of RFF. \subsection{Conservation laws}\label{conservation} The change of conserved quantities caused by the particle transfer from the interacting matter into the free matter can be obtained in terms of distribution function of the interacting matter calculated from eqs. (\ref{first-rethermalized}) as: \begin{eqnarray} dN^{\mu}_{i} (t) \! &=& \! dt \! \! \int \frac{d^3 p}{p^0} \, p^{\mu} \, \partial_{t} f^{i} \\ \nonumber \! &=& \! - \frac{dt}{\tau_0} \frac{L}{L-t} \int \frac{d^3 p}{p^0}\, p^{\mu} \, \frac{p^{\rho} d\sigma_{\rho}}{p^{\rho}u_{\rho}} f^i_{eq}(t,p) \, , \end{eqnarray} while the change in the energy-momentum as: \begin{eqnarray} dT^{\mu\nu}_{i} (t) \! &=& \! dt \! \! \int \frac{d^3 p}{p^0} \, p^{\mu} p^{\nu} \, \partial_{t} f^{i} \\ \nonumber \! &=& \! - \frac{dt}{\tau_0} \frac{L}{L-t} \int \frac{d^3 p}{p^0}\, p^{\mu} p^{\nu} \, \frac{p^{\rho} d\sigma_{\rho}}{p^{\rho}u_{\rho}} f^i_{eq}(t,p)\, . \end{eqnarray} The equilibrium distribution function for massless baryonfree particles is: \begin{equation} f^i_{eq}(t,p) = \frac{g}{(2\pi \hbar)^3} \, \exp{\left[-\frac{{\gamma(p^0 - jup \cos{\theta}_{\vec{p}})}}{T}\right]} \, , \end{equation} where the four-momentum of particles is $p^{\mu} = (p^0,\vec{p})$, $p=|\vec{p}|$, $p^x = p \cos \theta_{\vec{p}}$, the flow velocity of the interacting matter in RFF is $u^{\mu} = \gamma(1,v,0,0)_{RFF}$, $\gamma = 1/\sqrt{1 - v^2}$, $u = |v|$, $j = \textrm{sign} (v)$, and $g$ is the degeneracy of particles. The results of the calculations in the RFF can be found in Appendix B. \\ \indent The change in energy density after a step $dt$ is \begin{equation}\label{energy_density} d e_{i}(t) = u_{\mu,i}(t) \, dT^{\mu\nu}_{i}(t) \, u_{\nu,i}(t) \, , \end{equation} from which by using a simple EoS, $e = \sigma_{SB} T^4$, for a baryonfree massless gas with $\sigma_{SB} = \frac{\pi^2}{10}$, we can calculate the change in the temperature and Landau's flow velocity similarly to \cite{article_1, cikk_1, cikk_3}. Thus, \begin{eqnarray}\label{landau} d \ln T &=& \frac{\gamma^{2}}{4\sigma_{SB} T^{4}} \bigg[ dT^{00}_i - 2vdT^{0x}_i + v^{2} dT^{xx}_i \bigg] \, ,\\ \nonumber d v &=& \frac{3}{4 \sigma_{SB} T^{4}}\bigg[ -v dT^{00}_i + (1+v^{2}) dT^{0x}_i -v dT^{xx}_i \bigg]\,. \end{eqnarray} while in the massless limit the above equations lead to: \begin{eqnarray} \label{massless_landau} d \ln T &=& - \frac{dt}{\tau_0} \left(\frac{L}{L-t} \right)\frac{3n\gamma}{4\sigma_{SB} T^{3}} \, ,\\ \nonumber d v &=& - \frac{dt}{\tau_0} \left(\frac{L}{L-t} \right)\frac{3 n v}{4 \gamma \sigma_{SB} T^{3}} \, . \end{eqnarray} \section{Results and discussions} In this section we will present the results for the Post FO distribution and the relevant quantities calculated form this model. We will present our results for two different cases, for infinite FO (I) and finite FO (F). \begin{itemize} \item [ I) \,] The system is characterized by an infinitely long FO duration, [i.e., $L/(L-t) \rightarrow 1$ and $(t = 100 \tau_0)$, where most of the matter is frozen out]. The results are shown on Figs. \ref{figure_3t}, \ref{figure_5t}, \ref{figure_7t}, \ref{figure_8t}. \item [ F) \,] A finite FO process happening in a finite FO layer, where $(L=10 \tau_0)$. The results are shown on Figs. \ref{figure_4t}, \ref{figure_6t}, \ref{figure_7t}, \ref{figure_9t}, \ref{figure_10t}. \end{itemize} \subsection{The evolution of temperature of the interacting component} The first set of figures, Fig. \ref{figure_3t} and Fig. \ref{figure_4t}, shows the gradual decrease in temperature of the interacting component calculated in RFF. \\ \indent Comparing Fig. \ref{figure_3t} with Fig. \ref{figure_4t}, we see the difference between the finite and infinite FO. The FO in a finite layer is faster than in an infinite layer, "per se". The temperature curves belonging to different initial flow velocities but with opposite sign [i.e., $v_0 = -0.5$ and $v_0=0.5$] are the same. This is so since for time-like FO we do not have any constraint on the momenta, such as the cut-off factor $\Theta(p^{\mu}d\sigma_{\mu})$, and the initial momentum distribution is symmetric over the time axis. In the case of time-like FO the gradual cooling of the matter is faster and "smoother" compared to space-like FO, since the most energetic particles freeze out unrestricted in direction thus the remaining interacting component cools down faster. \\ \indent The matter with higher initial flow velocity, $v_0$, cools faster, but for small differences between the initial flow velocities, the resulting difference in the temperature is negligible. If the interacting gas has a higher flow velocity in RFF, the escape rate is bigger for higher values of the flow velocity. This expresses the fact that if the matter flows we remove energy faster form the interacting component. \\ \begin{figure}[t!] \centering \includegraphics[width=8.5cm, height = 5.5cm]{figure_3t.eps} \caption{(Color online) The temperature of the interacting component in RFF, calculated for an infinitely lasting FO. The initial temperature is \mbox{$T_0 = 170\,$ MeV}, the parameter, $v_0$, is the initial flow velocity. This corresponds to case I.} \label{figure_3t} \end{figure} \\ \begin{figure}[hbt!] \centering \includegraphics[width=8.5cm, height = 5.5cm]{figure_4t.eps} \caption{(Color online) The temperature of the interacting component in RFF, calculated for a finite $(L=10 \tau_0)$ FO time, where the initial temperature is \mbox{$T_0 = 170\,$ MeV}, and $v_0$ is the initial flow velocity. This corresponds to case F.} \label{figure_4t} \end{figure} \subsection{The evolution of common flow velocity of the interacting component in RFF and RFG} The second set of figures, Figs. \ref{figure_5t}, \ref{figure_6t}, shows the evolution of the flow velocity of the interacting component calculated for a baryonfree massless gas in RFF. \\ \indent Again, comparing Fig. \ref{figure_5t} with Fig. \ref{figure_6t}, we can see the difference between finite and infinite FO. In the case of finite FO the velocity decrease is much faster than in the case of infinite FO. \\ \indent Furthermore, we notice that the flow velocity of the interacting component tends to zero, while here we recall to compare, that in case of space-like FO it tends to $-1$. Again, this is due to the cut-off factor which retains particles propagating with negative momenta in space-like directions. In RFF the quantities change discontinuously at the light cone and that is why we have different results for the final flow velocity comparing space-like and time-like cases. However, in RFG all quantities are continuous when crossing the light cone, see Ref. \cite{article_1,QM05_1}. \\ \begin{figure}[!t] \centering \includegraphics[width=8.5cm, height = 5.5cm]{figure_5t.eps} \caption{(Color online) The evolution of the flow velocity of the interacting component calculated for an infinitely lasting FO, corresponding to case I. The initial temperature is $T_0 = 170\,$ MeV, and $v_0$ is the initial flow velocity of the gas.} \label{figure_5t} \end{figure} \begin{figure}[!hbt] \centering \includegraphics[width=8.5cm, height = 5.5cm]{figure_6t.eps} \caption{(Color online) The evolution of the flow velocity of the interacting component calculated for a finite $(L=10 \tau_0)$ FO, corresponding to case F. The initial temperature is $T_0 = 170\, $ MeV, and $v_0$ is the initial flow velocity of the gas.} \label{figure_6t} \end{figure} \subsection{The transverse momentum and the contour plots of the Post FO distribution} The third set of figures, Figs. \ref{figure_7t}, and \ref{figure_8t}, shows the evolution of the local transverse momentum distribution and the corresponding contour plots of the Post FO momentum distribution. \\ \indent We have presented a one-dimensional model here, but we assume that it is applicable for the direction transverse to the beam in heavy ion experiments. The plots presented should be related to the transverse momentum distribution of measured particles. \\ \begin{figure}[!hbt] \centering \includegraphics[width=8.5cm, height = 5.5cm]{figure_7t.eps} \caption{(Color online) The local transverse momentum (here $p_x$) distribution for a baryonfree and massless gas at $(p_y = 0)$. The calculations were done for an infinite FO with thin lines and for a finite FO $(L=10\tau_0)$ with marker lines. The initial flow velocity and temperature are, $u^{\mu}=(1,0,0,0)_{RFG}$ and $T_0 = 170\,$MeV. The transverse momentum spectrum at the end is obviously curved due to the FO process for low momenta.} \label{figure_7t} \end{figure} \\ \indent From Fig. \ref{figure_7t} we see that the Post FO momentum distributions for the infinite and finite FO cases are qualitatively identical. At the early stages of the FO process (for values of $t \simeq \tau_0$) the distribution of particles in the two cases match. This property persists until the end of the FO process, and can be seen on Fig. \ref{figure_7t}, where the local transverse momentum distribution calculated for an infinite and finite FO are identical. The maximum is increasing with $t$ as indicated in Figs. \ref{figure_7t} and \ref{figure_8t}. Thus, the final Post FO distributions do not differ if we switch from an infinitely long to a finite layer FO description for any initial flow velocity. This means that our finite layer FO description was done correctly. These important features of the model were already discussed in detail in Refs. \cite{article_1, QM05_2}. \\ \indent The overall conclusion is that the resulting Post FO distributions are non-thermal distributions even for time-like FO processes. The distributions strongly deviate from thermal distributions in the low momentum region [i.e., $p_x < 300$ MeV]. If one decreases the duration of total FO time, the gradual FO process would still produce similar particle spectra until the duration is not less than $2\tau_0$. Below that value the spectra becomes less curved and in the limit when the duration approaches zero the final momentum spectra corresponds to a constant temperature equilibrium distribution function. For $L < 2\tau_0$ the FO process does not have enough time to significantly change the shape of the final spectrum. \\ \indent Here one may go further and intuitively say that the FO process has a maximal lifespan, even though the parameter, L, was not defined in this work in such way that it would allow us to exactly calculate its limits from first principles. Of course for such a statement to hold one would need a realistic full scale fluid dynamical simulation including chemistry, secondaries and the expansion of the system. However, our results concluded from this simple model could still hold valuable in the realistic FO modeling in complex fluid dynamical simulations. \begin{figure}[!t] \centering \includegraphics[width=8.5cm, height=3.8cm]{figure_8t.eps} \caption{(Color online) The Post FO distribution, $f_{free}(x,\vec{p})$, at point A of Fig. \ref{figure_2t}, for an infinitely long FO length. The figures correspond to different time points, {\bf $t = 1 \tau_0,\ 10 \tau_0,\ 100 \tau_0$} respectively. Contour lines are given at values represented on the figure. The initial flow velocity and temperature are; $u^{\mu}=(1,0,0,0)_{RFG}$ and $T_0 = 170\,$ MeV. The maximum is increasing with $t$ as indicated on Fig. \ref{figure_7t}. } \label{figure_8t} \end{figure} \subsection{The boosted Post FO distributions} The fourth set of figures, Fig. \ref{figure_9t} and Fig. \ref{figure_10t}, shows the final Post FO distributions calculated for different flow velocities in RFF and the boosted post FO (J\"uttner) distributions in RFG. The FO distributions corresponding to different initial flow velocities, $u^{\mu}=\gamma_{\sigma}(1,-v_{\sigma},0,0)_{RFF}$, generally lead to non-equilibrated and anisotropic Post FO distributions. However, boosting the distribution from point A to points B or C leads to a more elongated FO distribution in the direction of the boost than the calculated Post FO distributions at those points, (compare the contours with the same values given on Figs. \ref{figure_9t} and \ref{figure_10t}). \begin{figure}[!t] \centering \includegraphics[width=8.5cm, height =3.8cm]{figure_9t.eps} \caption{(Color online) The final Post FO distribution, $f_{free}(x,\vec{p})$, at points A, B, C, of Fig. \ref{figure_2t}, calculated for a finite FO time, $L = 10\tau_0$. The contour plots correspond to different initial flow velocities, $u^{\mu}=\gamma_{\sigma}(1,-v_{\sigma},0,0)_{RFF}$, with $v_{\sigma} = 0, 0.5c, 0.9c$, respectively where the initial temperature is $T_0 = 170\, $ MeV. Note that, the Post FO distributions at points B and C are not the boosted distributions of point A. The difference is that the Post FO distribution is much less elongated than the boosted J\"uttner, because it is a superposition of sources with decreasing speed in RFF as indicated in Fig. \ref{figure_6t}.} \label{figure_9t} \end{figure} \\ \indent This is also an important outcome of our analysis leading to the conclusion that assuming an isotropic equilibrated J\"uttner distribution at time-like parts the FO hypersurface, other than at point A on Fig. \ref{figure_2t} where $u^{\mu} = d\sigma_{\mu}$, in general cannot hold \cite{gorenstein}. More importantly, the common practice of boosting the J\"uttner distribution or any post FO distribution function instead of calculating it from the conservation laws, similarly as it was done here, leads to a noticeable difference in the final particle spectra. \subsection{The non-equilibrated post FO distributions revised} Here we present another important result of our study, following the approach from Ref. \cite{cikk_5}, where an infinitely long FO was studied with momentum independent escape rate. We can reproduce that earlier result by taking $u^\mu(t_0)=d\sigma^\mu=(1,0,0,0)$, and can calculate the temperature decrease using eqs. (\ref{massless_landau}), in the case of a massless baryonfree matter for an infinitely lasting FO: \begin{equation} T(t) = T(t_0) \, \exp \left(-\frac{k}{\tau_0} (t-t_0) \right)\, , \end{equation} where $k = 3 /(4 \pi^2\sigma_{SB})$. In the general case, the flow velocity is not zero, hence one has to solve the system of equations from eqs. (\ref{massless_landau}). The distribution function of interacting particles (in the fast rethermalization limit) at any time $t$ is: \begin{equation} f^i(t,p) = \frac{1}{(2\pi)^3} \, \exp \left(-\frac{p^0}{T(t_0)} \, e^{k (t-t_0)/\tau_0}\right) \, , \end{equation} Now, we can solve the equation for the free component from eqs. (\ref{first-rethermalized}), therefore the distribution of free particles at time $t$ is: \begin{eqnarray}\label{exp_int_t} f^f(t,p) &=& \frac{1}{\tau_0}\int_{t_0}^{t} f^i(t',p) \\ \nonumber &=&\frac{k^{-1}}{(2\pi^3)} \, \left[ \textrm{Ei} \left( -\frac{p^0}{T(t)} \right) - \textrm{Ei} \left( -\frac{p^0}{T(t_0)}\right) \right]\, , \end{eqnarray} which for $t\rightarrow \infty$ leads to: \begin{equation}\label{exp_int} f^f = \frac{k^{-1}}{(2\pi^3)} \, \textrm{Ei} \left( -\frac{p^{\mu} u_{\mu}(t_0)}{T(t_0)}\right) \end{equation} where $\textrm{Ei}$ is the exponential integral function defined in Appendix B eq. (\ref{exponential_integral}). Thus, we got a simple formula, similarly to the one in Ref. \cite{cikk_5}, which correctly parameterizes the non-equilibrated post FO distribution function when $u^{\mu} = (1,0,0,0)$. It was actually shown in Ref. \cite{article_1} that the post FO distribution is not sensitive to momentum dependence of the escape rate, so we assume that this simple formula is valid for any initial flow velocity $u^\mu(t_0)$. \\ \indent Here we will use the above formula to plot the post FO distribution, for finite layers, with $L > 2\tau_0$, and extend this approximation to the general case when $u^{\mu}(t_0)\neq d\sigma_{\mu}$. On Fig. \ref{figure_12t}, we have plotted the final FO distribution functions calculated using eq. (\ref{exp_int}) with lines, and the finite FO $(L=3\tau_0)$ calculation with marker lines. The results are matching, which is a remarkable result, thus we conclude that this simple approximation is applicable for the description of gradual FO thorough finite time-like layers and correctly approximates its post FO distribution functions. \begin{figure}[!t] \centering \includegraphics[width=8.5cm, height =3.8cm]{figure_10t.eps} \caption{(Color online) The post FO distribution at point A and boosted to the frames at points B and C as depicted in Fig. \ref{figure_2t}. The different figures correspond to different initial normal vectors, $d\sigma^{\mu}=\gamma_{\sigma}(1,v_{\sigma},0,0)_{RFG}$, where {\bf $v_{\sigma} = 0, 0.5c, 0.9c$}, and the initial temperature is $T_0 = 170\,$ MeV.} \label{figure_10t} \end{figure} \begin{figure}[!hbt] \centering \includegraphics[width=8.5cm, height = 5.5cm]{figure_12t.eps} \caption{(Color online) The local transverse momentum (here $p_x$) distribution for a baryonfree and massless gas at $(p_y = 0)$. The calculations were done for an infinite FO with thin lines using eq. (\ref{exp_int}) and for a finite FO $(L=3\tau_0)$ with marker lines. The initial temperature was $T_0 = 170\,$MeV. The different initial flow velocities are given on the figure legend.} \label{figure_12t} \end{figure} \section{Conclusions} In this work we have presented a simple kinetic freeze out model for a finite time-like layer. We have demonstrated that FO across time-like surfaces leads to non-equilibrated and anisotropic distributions. These distributions in general cannot be Lorentz transformed to a frame where the distribution is isotropic. The only exception is when the normal to the FO hypersurface is parallel to the local flow velocity. Our analysis shows that the usual practice of assuming a J\"uttner distribution as a Post FO distribution is in general not valid! \\ \indent We can also see that while the boosted J\"uttner distribution is elongated in the boost direction, i.e. in the direction of $d\sigma_{\mu}$, the Post FO distribution is close to a spherical and isotropic distribution at low momenta, and becomes elongated only at higher momenta, see Fig. \ref{figure_9t}. This special Post FO distribution leads to a curved "$p_{t}$ - spectrum". Here, we can also demonstrate (as in earlier works \cite{article_1, cikk_1, cikk_2, cikk_3}) that non-equilibrium processes in kinetic FO lead to observable effects. \\ \indent We observe that the J\"uttner distribution is not a good approximation for the Post FO distribution, just like in the case of a space-like FO. While in the case of space-like FO the Cancelling-J\"uttner distribution introduced in Ref. \cite{karolis} is satisfactory, in the case of time-like FO, we have found a simple formula to use. \\ \indent Now, one may ask the question whether we observe this additional low $p_t$ effect in the experimental data. This effect has several possible explanations: products of low momentum resonance decays, the transverse expansion of the system, and possibly due to the long gradual FO with rethermalization. As already discussed, during such scenario the particles are freezing out at different gradually decreasing temperatures, thus correspondingly the final FO spectrum is a superposition of thermal distributions with different temperatures. Although the low $p_t$ enhanced non-thermal spectrum of massless pions is a necessary outcome of long gradual FO with rethermalization, for the heavy particles (if these are in the mixture with pions) it is almost unobservable, see Ref. \cite{cikk_5}. \\ \indent In our simplistic study we have found that FO in layer of finite thickness below $L < 2\tau_0$ will not show a sharp peak at low momentum in the transverse momentum spectrum. Thus, naively one can conclude from a simple fit that the FO in heavy-ion collisions happens in a narrow or wide FO layer. However, such conclusion would be premature without including the expansion of the system and calculate two particle correlations. At the moment FO in a long finite layer, $L > 2\tau_0$, cannot be excluded. If one assumes gradual FO with non-thermal post FO spectra, then one may also fit the data but with different flow velocity and slope parameter, where the curvature of the pion spectra at low $p_t$ will be partly due to FO and partly due to the expansion of the system. \section{Outlook} We do not aim directly to apply the results presented here to experimental heavy ion collision data, instead our purpose was to study qualitatively the basic features of the freeze out process, and to demonstrate the applicability of this covariant formulation for FO in a finite layer. \\ \indent Here we note that our model may be applicable in CFD calculations, where one has both time-like and space-like parts of the full FO hypersurface, thus the gradual FO calculation must be done over the full FO hypersurface, with varying flow velocities and normal vectors. The method should be applied after reaching the $T_{FO}$ critical temperature and calculate the Post FO momentum distribution function starting form the inner FO hypersurface. Such a calculation should be compared to the Cooper-Frye ansatz in the first place and then to experimental results, similarly to Refs. \cite{grassi_1, hama_2, grassi_2}. \\ \indent A successful application of this model was already used to study the impact of nucleon mass shift on the freeze out process \cite{sven}. This analysis will be carried forward to calculate the impact of mass shift on other particles, such as pions and kaons, which will help us study the effect of mass shift on two-particle correlations. An even more interesting study based on our analysis will estimate the effect of expansion on the time-like freeze out process using the Bjorken model \cite{new_article}. Therefore, we believe that our model may give a better description and understanding of the final observables which are calculated using the single (and two) particle distribution functions. \section*{ACKNOWLEDGMENTS} The authors, L. P. Csernai, E. Moln\'ar, A. Ny\'iri and K. Tamosiunas thank the hospitality of the University of Cape Town, where parts of this work were done. E. Moln\'ar, also thanks the hospitality of the Babe\c s-Bolyai University of Cluj. \\ \indent Enlightening discussions with Cs. Anderlik, T. S. Bir\'o, J. Cleymans, A. Dumitru and S. Zschocke are gratefully acknowledged. \section*{APPENDIX A} Here we discuss the properties of the rethermalization term from eq. (\ref{first-rethermalized}) and its consequences on finishing the FO process in a finite layer. \begin{figure}[t!] \centering \includegraphics[width=8.6cm, height = 2.4cm]{figure_11t.eps} \caption{(Color online) A schematic view of the FO process for a linear density profile. Initially the mean free path is $\lambda_{mfp} = a_0$, the relaxation length is $\lambda_{th} = 2 \lambda_{mfp}$, the initial characteristic length is $\lambda_0$, while the length of the FO layer is $L = 4\lambda_0$.} \label{figure_11t} \end{figure} \\ \indent From kinetic theory we know that if the following conditions: \begin{equation} \tau_{mfp} < \tau_{th} < \tau_0 \quad \text{or} \quad \lambda_{mfp} < \lambda_{th} < \lambda_{0}\, , \end{equation} between the average length between the collisions, the relaxation length, and the characteristic length are satisfied, then we can use the Boltzmann Transport Equation (BTE) for the evolution of the single particle distribution function, $f(x,p)$. \\ \indent For better understanding we first assume a linear decrease of the interacting particle density during freeze out, such as, $n(x) = (L - x)/L$, where the mean free path of interacting particles is, $\lambda_{mfp}(x) \approx 1/n(x)$. Using the relaxation time approximation for the FO process, for example at $x>2\lambda_0$ the density of the interacting particles already decreased to $n(x) < n_0/2$, while the mean free path increased to $\lambda_{mfp}(x) > a =2a_0 $, see Fig. \ref{figure_11t}. Consequently, by the end of the FO process, the thermalization length becomes longer than the initial characteristic length of the system. Although, $\tau_0$ is constant (scale parameter), the characteristic FO length is actually not constant during the evolution. Thus, using the rethermalization approximation the error we introduce within is of the order of $\tau_{th}/\tau_0$. \\ \indent In our model the change in the density is generally given as: \begin{equation}\label{density} d n_{i}(x) = u_{i,\mu}(x) \, dN^{\mu}_{i}(x) \, . \end{equation} This leads to an exponentially fast decrease of particle density, therefore more than $95\%$ of the interacting matter is frozen out before $\tau_{th}\simeq \tau_0$, thus we can safely use the relaxation time approximation in our calculations. \section*{APPENDIX B} The changes in the particle four current and energy momentum tensor are: \begin{eqnarray} \nonumber d N^{0}_i (t) &=& - \frac{dt}{\tau_0} \frac{L}{L-t} \frac{n}{4ju\gamma} \Bigg\{- G_1^+(m) + \, G_1^-(m) \Bigg\} \\ \nonumber &\buildrel m=0 \over \longrightarrow & - \frac{dt}{\tau_0} \frac{L}{L-t} n \Bigg[\frac{(3+v^2)}{3}\,\gamma^2 \Bigg] \, , \end{eqnarray} \begin{eqnarray} \nonumber d N^{x}_i (t) &=& \frac{dN^{0}_i (t)}{ju} - \\ \nonumber &&\frac{dt}{\tau_0} \frac{L}{L-t} \frac{n}{4ju\gamma} \Bigg\{- 2b\Big[2 K_1(a) + aK_0(a) \Big]\Bigg\} \\ &\buildrel m=0 \over \longrightarrow & - \frac{dt}{\tau_0} \frac{L}{L-t} n \Bigg[\frac{(3+v^2)}{3v}\,\gamma^2 - \frac{1}{v}\Bigg] \, , \end{eqnarray} \begin{eqnarray} \nonumber d T^{00}_i (t) &=& - \frac{dt}{\tau_0} \frac{L}{L-t} \frac{nT}{4ju\gamma} \Bigg\{ - G_2^+(m) + \, G_2^-(m) \Bigg\} \\ \nonumber &\buildrel m=0 \over \longrightarrow & - \frac{dt}{\tau_0} \frac{L}{L-t} nT \Bigg[ 3(1+v^2)\,\gamma^3 \Bigg] \, , \end{eqnarray} \begin{eqnarray} \nonumber d T^{0x}_i (t) &=& \frac{dT^{00}_i(t) }{ju} - \frac{dt}{\tau_0} \frac{L}{L-t} \frac{nT}{4ju\gamma} \\ \nonumber &&\Bigg\{- 2b^2(3 + u^2) K_2(a) - 2ab^2 K_1(a) \Bigg\} \\ \nonumber &\buildrel m=0 \over \longrightarrow & - \frac{dt}{\tau_0} \frac{L}{L-t} nT \Bigg[ \frac{3(1+v^2)\,\gamma^3}{v} - \frac{\gamma (3+v^2)}{v}\Bigg] \, , \end{eqnarray} \begin{eqnarray} \nonumber d T^{xx}_i (t) &=& \frac{dT^{0x}_i(t)}{ju} - \frac{T}{\gamma ju} \Bigg[ dN^{x}_i(t) - \frac{dN^{0}_i(t)}{ju} \Bigg] \\ \nonumber &-& \frac{dt}{\tau_0} \frac{L}{L-t} \frac{nT}{4ju\gamma} \Bigg\{-\frac{2b^2}{ju} (1 + 3u^2) K_2(a) \\ \nonumber &-& 2juab^2 K_1(a) \Bigg\} \\ \nonumber &\buildrel m=0 \over \longrightarrow & - \frac{dt}{\tau_0} \frac{L}{L-t} nT \Bigg[ \frac{3(1+v^2)\,\gamma^3}{v^2} - \frac{\gamma (3+v^2)}{v^2} \\ \nonumber &+& \frac{1}{\gamma v^2} - \frac{(1 + 3v^2)\gamma}{v^2}\Bigg] \, , \end{eqnarray} \begin{eqnarray} \nonumber d T^{yy}_i (t)&=& - \, \frac{dT^{xx}_i(t)}{2} - \frac{dt}{\tau_0} \frac{L}{L-t} \frac{nT}{8ju\gamma} \\ \nonumber && \Bigg\{ - G^+_{3}(m) + G^+_{3}(m) \Bigg\} \\ \nonumber &\buildrel m=0 \over \longrightarrow & - \frac{dt}{\tau_0} \frac{1}{2} \Bigg[ dT^{xx}(t) + dT^{00}(t)\Bigg] \, , \end{eqnarray} and \begin{eqnarray} d T^{zz}_i (t) = d T^{yy}_i (t)\, , \end{eqnarray} where $a = \frac{m}{T}$, $b=a\gamma$ and $n = 4 \pi T^3 a^2 K_2(a)\, g \frac{e^{\mu/T}}{(2 \pi \hbar)^3}$ is the particle density, while $g$ is the degeneracy factor. Furthermore, the definition of the modified Bessel function of the second kind $K_{n}(z)$ for $n>-1$, is \begin{equation} K_{n}(z) = \frac{2^n \, n!}{(2n)!} \, z^{-n} \int_{z}^{\infty} dx \,e^{-x} \,(x^2 - z^2)^{n-\frac{1}{2}} \, . \end{equation} The analytically not integrable functions $G_n^- (m) $ and $G_n^+ (m) $, where $n>-2$, depend on other quantities such as $u$ and $T$. However, in the case when, $m \rightarrow 0$, the dependence on other quantities persists and hence we only denote the mass dependence of the functions defined as: \begin{eqnarray} G_n^{\pm}(m) &=&\frac{1}{T^{n+2}}\int_{0}^{\infty} d p \, p \, \Big( \sqrt{p^2 + m^2} \Big)^n \, \\ \nonumber &\times&\Gamma \Big(0,\frac{\gamma}{T} \sqrt{p^2 + m^2} \pm \frac{\gamma jup}{T}\Big) \, . \end{eqnarray} The exponential integral function is defined as: \begin{equation}\label{exponential_integral} \textrm{Ei} (z) = \int_{z}^{\infty} dx \, \frac{e^{-x}}{x} \, . \end{equation}
{ "timestamp": "2007-06-16T21:27:36", "yymm": "0503", "arxiv_id": "nucl-th/0503048", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503048" }
\section{The spectral sequence} By a Grothendieck spectral sequence we mean a composite functor spectral sequence as described by Cartan and Eilenberg (\cite{cartaneilenberg}, Chapter XVII \S7). The Lyndon--Hochschild--Serre spectral sequence (\cite{cartaneilenberg}, Chapter XVI \S6 (6)) can be viewed this way: see for example Rotman's account, Theorem 11.45 of \cite{rotman}. Consider a group extension $K\rightarrowtail G\twoheadrightarrow Q$. Write $\operatorname{\Mo\HYPHEN}{\mathbb Z} G$ for the category of right ${\mathbb Z} G$-modules. The zeroth cohomology functor $H^0(G,{\phantom M})$ is the fixed point functor $({\phantom M})^G$ and the LHS spectral sequence arises from the factorization of this functor through the category of ${\mathbb Z} Q$-modules as illustrated below: \[ \xymatrix{ \operatorname{\Mo\HYPHEN}{\mathbb Z} G\ar[rr]^{({\phantom M})^G}\ar[dr]_{H^0(K,{\phantom M})=({\phantom M})^K}&&\operatorname{\Mo\HYPHEN}{\mathbb Z}.\\ &\operatorname{\Mo\HYPHEN}{\mathbb Z} Q\ar[ur]_{H^0(Q,{\phantom M})=({\phantom M})^Q}\\ } \] One needs to know that the module categories are abelian categories with enough injectives and that the $K$-fixed point functor carries injective $G$-modules to injective $Q$-modules. Now suppose that $\mathcal S$ is an admissible family of subgroups of $G$. \begin{definition}\label{new cat} We write $$\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$$ for the full subcategory of $\operatorname{\Mo\HYPHEN}{\mathbb Z} G$ whose objects are those ${\mathbb Z} G$-modules $M$ for which $M=\bigcup_{H\in\mathcal S}M^H.$ \end{definition} \begin{remark}\label{rem} Although this definition makes perfect sense for any family $\mathcal{S}$ of subgroups, it behaves particularly well when $\mathcal{S}$ is admissible and we shall always assume that this is so. Given this assumption and an arbitrary ${\mathbb Z} G$-module $M$ then $\bigcup_{H\in\mathcal S}M^H$ is a ${\mathbb Z} G$-submodule of $M$ which belongs to the new category. We write $H^0(\mathcal{S},M)$ for this submodule. Notice that the definition of admissible family is designed exactly so that this works: $H^0(\mathcal{S},M)$ inherits an action of $G$ because $\mathcal{S}$ is closed under conjugation and $H^0(\mathcal{S},M)$ is an additive subgroup of $M$ because $\mathcal{S}$ is downwardly directed. Thus $H^0(\mathcal{S},M)$ is an object of the new category $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ and {\em by definition} all objects of $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ arise this way. \end{remark} It is easy to see that $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ is an abelian category: it will be the intermediary for our spectral sequence, replacing $\operatorname{\Mo\HYPHEN}{\mathbb Z} Q$. Note that in case $\mathcal S$ consists of a single (necessarily normal) subgroup $K$ then $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ is naturally equivalent to the category of right modules for the quotient group $Q=G/K$. We now have two functors. The first is mentioned already in Remark \ref{rem}: the assignment $$H^0(\mathcal S,{\phantom M}):\operatorname{\Mo\HYPHEN}{\mathbb Z} G\to\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$$ defined by $$H^0(\mathcal S,M)=\bigcup_{H\in\mathcal S}M^H$$ is functorial in $M$. Secondly, we can restrict the $G$-fixed point functor to the new category so we have a functor $$H^0(G/\mathcal S,{\phantom M}):\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S\to\operatorname{\Mo\HYPHEN}{\mathbb Z}$$ defined by $$H^0(G/\mathcal S,M)=M^G.$$ \begin{remark}\label{rem2} The $G$-fixed point functor $({\phantom M})^G:\operatorname{\Mo\HYPHEN}{\mathbb Z} G\to\operatorname{\Mo\HYPHEN}{\mathbb Z}$ now factors through $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ as the composite: $H^0(\mathcal S,{\phantom M})$ followed by $H^0(G/\mathcal S,{\phantom M})$ as illustrated in the diagram below. \[ \xymatrix{ \operatorname{\Mo\HYPHEN}{\mathbb Z} G\ar[rr]^{({\phantom M})^G}\ar[dr]_{H^0(\mathcal S,{\phantom M})}&&\operatorname{\Mo\HYPHEN}{\mathbb Z}.\\ &\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S\ar[ur]_{H^0(G/\mathcal S,{\phantom M})}\\ } \] This will provide the basis for a Grothendieck spectral sequence. \end{remark} Beware that the notation $H^0(G/\mathcal S,{\phantom M})$ is not intended to imply the construction of any kind of object $G/\mathcal S$, but is simply notation for the functor. This functor necessarily has to be distinguished from $H^0(G,{\phantom M})$ which has a different domain. The notation {\em is} intended to suggest an analogy with the classical situation when $\mathcal S$ consists of a single normal subgroup $K$. The analogy works well, and raises the interesting question whether there is any kind of natural object which deserves to be named $G/\mathcal S$. In \S6 we show that a certain completion of $G$ appears to be the object one should expect. \begin{lemma}\label{injectives}\ \begin{enumerate} \item If $I$ is an injective ${\mathbb Z} G$-module then $H^0(\mathcal S,I)$ is injective in $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$. \item $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ has enough injectives. \item The functor $H^0(\mathcal S,{\phantom M})$ has right derived functors $H^n(\mathcal S,{\phantom M})$ and there are natural isomorphisms $$H^n(\mathcal S,{\phantom M})\cong\colimf H^n(H,M).$$ \end{enumerate} \end{lemma} \begin{proof} (i) Let $I$ be an injective ${\mathbb Z} G$-module. To show that $H^0(\mathcal S,I)$ is injective we need to address the extension problem as illustrated below: \[ \xymatrix{ 0\ar[r]&M\ar[d]_\phi\ar[r]&N\ar@{-->}[dl]^{\widehat\phi{\text?}}\\ &H^0(\mathcal S,I)&\\ } \] Using the injectivity of $I$ we can find $\widehat\phi$ to make a commutative diagram \[ \xymatrix{ 0\ar[r]&M\ar[d]_\phi\ar[r]&N\ar[ddl]^{\widehat\phi}\\ &H^0(\mathcal S,I)\ar[d]&\\ &I& } \] and since $N$ belongs to the subcategory, $\widehat\phi$ has image in $H^0(\mathcal S,I)$. This proves (i). For part (ii), observe that we can embed any ${\mathbb Z} G/\mathcal S$-module $M$ into an injective ${\mathbb Z} G$-module and then we have $M=H^0(\mathcal S,M)\hookrightarrow H^0(\mathcal S,I)$: by part (i), we have now embedded $M$ into an injective object of $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$. (iii) Derived functors are defined in the standard way using injective resolutions over ${\mathbb Z} G$, applying the functor $H^0(\mathcal S,{\phantom M})$ and passing to the cohomology of the resulting cochain complex. The isomorphism is easily established. \end{proof} \begin{lemma}\label{lem:der quot} The functor $H^0(G/\mathcal S,{\phantom M})$ has right derived functors $H^n(G/\mathcal S,{\phantom M})$. \end{lemma} \begin{proof} This time we work with injective resolutions in the category $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$, apply the $G$-fixed point functor and pass to cohomology. In general, there is no simple interpretation of the derived functors. \end{proof} Remark \ref{rem2}, Lemma \ref{injectives} and Lemma \ref{lem:der quot} together provide the ingredients necessary for a Grothendieck spectral sequence and our main tool is established: \begin{theoremA}\label{A} Let $G$ be a group and let $\mathcal S$ be an admissible family of subgroups. There is a Grothendieck spectral sequence $$H^p(G/\mathcal S,H^q(\mathcal S,M))\implies H^{p+q}(G,M)$$ which is natural in the $G$-module $M$. \end{theoremA} As a routine feature of any first/third quadrant spectral sequence we have the following: (see for example \cite{rotman}, Theorem 11.43). \begin{corollaryA}\label{corollaryA} With $G$ and $\mathcal S$ as above, \begin{enumerate} \item there is a five term exact sequence analogous to the standard inflation-restriction sequence: {\small{$$0\to H^1(G/\mathcal S,H^0(\mathcal S,M))\to H^1(G,M)\to H^1(\mathcal S,M)^G\to H^2(G/\mathcal S,H^0(\mathcal S,M))\to H^2(G,M);$$}} \item the inflation map $$H^n(G/\mathcal S,M)\to H^n(G,M)$$ (which is defined for $M$ in the subcategory $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$) is an isomorphism when $n=0$ and is injective when $n=1$. \end{enumerate} \end{corollaryA} \section{Continuity of functors} We need to consider continuity issues for the new functors. \begin{definition}\label{def:cont} Let $\mathcal C$ and $\mathcal D$ be abelian categories with all small filtered colimits. Let $F:\mathcal C\to\mathcal D$ be a functor. We say that $F$ is {\em continuous at zero} if and only if $${\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } F(M_\lambda)=0$$ whenever $(M_\lambda)$ is a small filtered colimit system in $\mathcal C$ which is {\em vanishing}: i.e. $${\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } M_\lambda=0.$$ \end{definition} First we recall a basic result from Bieri's notes, essentially the content of (\cite{bieri-qmw}, Theorem 1.3 (i)$\iff$(iiib)): \begin{lemma}\label{bieri1} A group $G$ is of type $\operatorname{FP}_n$ if and only if the cohomology functors $H^i(G,{\phantom M})$ are continuous at zero for all $i\le n$. \end{lemma} \begin{lemma}\label{lem:fpinfty} Let $\mathcal S$ be an admissible family in $G$. If all members of $\mathcal S$ have type $\operatorname{FP}_\infty$ then the functors $H^n(\mathcal S,{\phantom M})$ are continuous at zero for all $n$. \end{lemma} \begin{proof} By Lemma \ref{bieri1}, the $\operatorname{FP}_\infty$ condition guarantees that the functors $H^{m}(H,{\phantom M})$ are continuous at zero for all $H\in\mathcal S$ and all $m$. The result now follows from the natural isomorphism of Lemma \ref{injectives}(iii). \end{proof} The next lemma is a version of Strebel's criterion \cite{strebel}. \begin{lemma}\label{strebel} Let $G$ be a group of finite cohomological dimension. Suppose that for all vanishing filtered colimit systems $(P_\lambda)$ of projective modules $P_\lambda$, and all $m\in{\mathbb Z}$, $${\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^m(G,P_\lambda)=0.$$ Then $G$ is of type $\operatorname{FP}$. \end{lemma} \begin{proof} We shall use the following notation: write $FM$ for the free module on the underlying set of non-zero elements of a module $M$. Then $F$ is functorial in $M$ and the inclusion $M\setminus\{0\}\hookrightarrow M$ induces a natural surjection $FM\twoheadrightarrow M$ whose kernel, $\Omega M$, is also functorial. Moreover both $F$ and $\Omega$ take vanishing filtered colimit systems to vanishing filtered colimit systems. By Lemma \ref{bieri1} it suffices to prove that $H^m(G,{\phantom M})$ is continuous at zero for all $m$. This is proved by downward induction on $m$: it is trivial for all $m$ greater than the cohomological dimension of $G$, so we fix $m$ and assume as inductive hypothesis that $H^{m+1}(G,{\phantom M})$ is continuous at zero. Let $(M_\lambda)$ be a vanishing filtered colimit system of modules. Then we have short exact sequences $$\Omega M_\lambda\rightarrowtail FM_\lambda\twoheadrightarrow M_\lambda.$$ On passing to cohomology and taking colimits we have the exact sequence $${\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^m(G,FM_\lambda)\to {\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^m(G,M_\lambda)\to {\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^{m+1}(G,\Omega M_\lambda).$$ We need to prove that the central group here is zero. The right hand group vanishes by induction and the left hand group vanishes by hypothesis. The result follows from exactness. \end{proof} \begin{lemma}\label{lem:van1} Let $L$ be a near-normal subgroup of type $\operatorname{FP}_\infty$ and infinite index in a group $G$. Let $M$ be a ${\mathbb Z} L$-module. Let $\mathcal S$ be the set of subgroups commensurable with $L$. Then $$\left(H^m(\mathcal S,M\otimes_{{\mathbb Z} L}{\mathbb Z} G)\right)^G=0$$ for all integers $m$. \end{lemma} \begin{proof} We proceed in steps. \begin{enumerate} \item[Step 1.] The case $m=0$. \end{enumerate} We use only on the fact that all members of $\mathcal S$ have infinite index in $G$. Note that the calculation simplifies because $$\left(H^0(\mathcal S,M\otimes_{{\mathbb Z} L}{\mathbb Z} G)\right)^G=(M\otimes_{{\mathbb Z} L}{\mathbb Z} G)^G,$$ the set of $G$-fixed points. Let $T$ be a right transversal to $L$ in $G$, i.e. $G$ is the disjoint union of the cosets $Lt$, for $t\in T$. Any non-zero element of $M\otimes_{{\mathbb Z} L}{\mathbb Z} G$ has a unique expression as a finite sum $$m_1\otimes t_1+\dots+m_s\otimes t_s$$ where the $m_i\in M$ are non-zero and the $t_i$ are distinct elements of $T$. Let $X=Lt_1\cup\dots\cup Lt_s$. We can choose $g\in G$ such that $Xg\cap X=\emptyset$. To see this, suppose for a contradiction that $Xg\cap X=\emptyset$ for all $g$. Then $G$ is the union of the sets $t_i^{-1}Lt_j$ over all $i,j$: this expresses $G$ as a finite union of cosets of subgroups and implies that at least one of the subgroups has finite index in $G$ which is contrary to our assumption. For a $g$ such that $Xg\cap X=\emptyset$ we have $$(m_1\otimes t_1+\dots+m_s\otimes t_s)g\ne m_1\otimes t_1+\dots+m_s\otimes t_s$$ so there are no non-zero fixed points. \begin{enumerate} \item[Step 2.] In case $M$ is injective as a ${\mathbb Z} L$-module and $m\ge1$. \end{enumerate} We prove the stronger statement that for any $H$ in $\mathcal S$, $$H^m(H,M\otimes_{{\mathbb Z} L}{\mathbb Z} G)=0.$$ This argument uses both the commensurability and the type $\operatorname{FP}_\infty$. Fix any $H$. Mackey decomposition yields $$M\otimes_{{\mathbb Z} L}{\mathbb Z} G\cong\bigoplus_tMt\otimes_{{\mathbb Z}[L^t\cap H]}{\mathbb Z} H$$ as ${\mathbb Z} H$-modules, where $t$ runs over a set of $(L,H)$ double coset representatives in $G$. Each summand $Mt\otimes_{{\mathbb Z}[L^t\cap H]}{\mathbb Z} H$ is injective as an $H$-module because $M$ is injective over $L$ and, using commensurability, $L^t\cap H$ has finite index in $H$. In positive dimensions, the cohomology of any group vanishes on injective modules. Here we can take advantage of the fact that $H$ has type $\operatorname{FP}_\infty$ to see that the cohomology also vanishes on arbitrary direct sums of injective modules. \begin{enumerate} \item[Step 3.] The general case. \end{enumerate} The general case can be deduced by dimension shifting. We use induction on $m$, the case $m=0$ being covered by step (i). Choose any short exact sequence $$M\rightarrowtail I \twoheadrightarrow M'$$ with $I$ injective over $L$. Then we have a short exact sequence of induced modules: $$M\otimes_{{\mathbb Z} L}{\mathbb Z} G\rightarrowtail I\otimes_{{\mathbb Z} L}{\mathbb Z} G\twoheadrightarrow M'\otimes_{{\mathbb Z} L}{\mathbb Z} G.$$ Using the long exact sequence of cohomology together with step (ii) reduces the $m$-dimensional matter for $M$ to the $(m-1)$-dimensional matter for $M'$ and the result follows by induction. Of course, in case $m=1$, Step 2 does not cover everything but then Step 1 can be used as well. \end{proof} \begin{proposition}\label{bieri2} Let $n$ be natural number. Let $G$ be a finitely generated group of cohomological dimension $\le n+1$. Let $K$ be near-normal subgroup of infinite index in $G$ such that: \begin{enumerate} \item $K$ is of type $\operatorname{FP}$; \item $H^m(K,P)=0$ whenever $P$ is a projective module and $m\ne n$. \end{enumerate} Then $G$ is of type $\operatorname{FP}$. \end{proposition} \begin{proof} Let $\mathcal S$ be the family of all subgroups commensurable with $K$. Let $P$ be a projective ${\mathbb Z} G$-module. All the members of $\mathcal S$ inherit properties (i) and (ii) and it follows from Lemma \ref{injectives}(iii) and Lemma \ref{bieri1} that $\mathcal S$ itself inherits the properties as well: \begin{enumerate} \item $H^m(\mathcal S,{\phantom M})$ is continuous at zero for all $m$; \item $H^m(\mathcal S,P)=0$ whenever $P$ is a projective ${\mathbb Z} G$-module and $m\ne n$. \end{enumerate} The spectral sequence of Theorem A therefore collapses to a single column and we find that $$H^{n+1}(G,P)\cong H^1(G/\mathcal S,H^n(\mathcal S,P)),$$ $$H^{n}(G,P)\cong H^0(G/\mathcal S,H^n(\mathcal S,P)),$$ and $$H^{m}(G,P)=0$$ when $m\notin\{n,n+1\}$ because of (ii) for $m<n$ and the constraint on dimension of $G$ for $m>n+1$. By definition, $H^0(G/\mathcal S,H^n(\mathcal S,{\phantom M}))=(H^n(\mathcal S,{\phantom M}))^G$ and this functor vanishes on all induced modules (i.e. modules of the form $A\otimes{\mathbb Z} G$ where $A$ is an abelian group) by Lemma \ref{lem:van1}. Since projective modules are direct summands of free modules which are in turn examples of induced modules, we have $$H^{n}(G,P)=0.$$ Now we can apply Strebel's criterion Lemma \ref{strebel}. If $(P_\lambda)$ is a vanishing filtered colimit system of projective modules then we only have to check that $${\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^{n+1}(G,P_\lambda)=0$$ and this will follow from the isomorphism $${\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^{n+1}(G,P_\lambda)={\displaystyle\lim_{\buildrel\longrightarrow\over\lambda}\ } H^1(G/\mathcal S,H^n(\mathcal S,P_\lambda))$$ together with the observation that both the functors $H^1(G/\mathcal S,{\phantom M})$ and $H^n(\mathcal S,{\phantom M})$ are continuous at zero. The second observation is part of (i) above. The first observation follows from the injectivity of the inflation maps $$H^1(G/\mathcal S,H^n(\mathcal S,P_\lambda))\to H^1(G,H^n(\mathcal S,P_\lambda))$$ at the start of the five term exact sequence, Corollary A: note that $H^1(G,{\phantom M})$ is continuous at zero because $G$ is finitely generated. \end{proof} \section{A Decomposition Theorem} The goal of this section is to prove a decomposition theorem for certain groups. We take this in two stages. The first stage, Theorem B, is easy to state. The second stage Theorem C requires a little more introduction although it is easy to prove by combining Theorem B with results \cite{bf} of Bestvina and Feighn. \begin{theoremB}\label{B} Let $n$ be a fixed natural number. Suppose that $G$ is a group with the following properties: \begin{itemize} \item[$(\alpha)$] $G$ is finitely generated; \item[$(\beta)$] $G$ has cohomological dimension $\le n+1$; \item[$(\gamma)$] $G$ has a near-normal $\operatorname{PD}^n$-subgroup $K$. \end{itemize} Then either $K$ has finite index in $G$ or $G$ splits over a subgroup commensurable with $K$. \end{theoremB} We recall some definitions. An {\em $n$-dimensional duality group} (over ${\mathbb Z}$) is a group $G$ which affords a dualizing module $D$ so that there are natural isomorphisms $$H^i(G,M)\cong \operatorname{Tor}^{{\mathbb Z} G}_{n-i}(M,D)$$ for all $i$. {\em Poincar\'e duality groups} ($\operatorname{PD}^n$-groups) are duality groups for which $D$ has underlying additive group ${\mathbb Z}$. We say that a group $G$ {\em splits} over a subgroup $H$ if and only if $G$ is isomorphic to a free product with amalgamation $K*_HL$ with $K\ne H\ne L$ or and HNN-extension $K*_H$. According to the standard theory of group actions on trees which is described by Serre \cite{serre} and Dicks--Dunwoody \cite{dicks}, $G$ splits over $H$ if and only if there is a $G$-tree with no fixed vertex, one orbit of edges and in which $H$ is the stabilizer of one of the edges. The proof of Theorem B uses Dunwoody's graph cutting methods, \cite{dunwoody}. \begin{proof}[Proof of Theorem B]\ We may as well assume that $K$ has infinite index in $G$. Let $\mathcal S$ be the family of all subgroups commensurable with $K$. Then $\mathcal S$ is an admissible family of $\operatorname{PD}^n$-subgroups of $G$ all of which have infinite index in $G$. The hypotheses of Proposition \ref{bieri2} are satisfied and therefore $G$ has type $\operatorname{FP}$. The next step is to prove \begin{itemize} \item[{\bf Claim 1.}] $G$ is an $(n+1)$-dimensional duality group over any field $k$. \end{itemize} To be an $(n+1)$-dimensional duality group over a (non-zero) commutative ring $k$ it is necessary and sufficient that the following three conditions hold. \begin{itemize} \item $G$ is of type $\operatorname{FP}$ over $k$ and of cohomological dimension $\le n+1$. \item The cohomology groups $H^i(G,kG)$ are zero for $i\le n$. \item $H^{n+1}(G,kG)$ is flat as a $k$-module. \end{itemize} When these conditions hold, $G$ has dualizing module $H^{n+1}(G,kG)$: this is a left $kG$-module via the left action of $G$ on $kG$. Since $kG$ is an instance of an induced ${\mathbb Z} G$-module, the first two conditions are already established. The third condition is automatically satisfied if $k$ is a field. We now work over the field ${\mathbb F}$ of two elements and show how to identify $H^n(\mathcal S,{\mathbb F} G)$ with a certain set of subsets of $G$. Let $\mathcal P$ denote the powerset of $G$. This is viewed as an $({\mathbb F} G,{\mathbb F} G)$-bimodule with symmetric difference of subsets providing the additive structure and left/right multiplication by elements of $G$ for the action. Let $\mathcal B$ be the set of subsets $B$ of $G$ which satisfy the following condition \begin{itemize} \item There is a subgroup $H\in\mathcal S$ and a finite subset $F$ of $G$ such that $B=HFH$. I.e. $B$ is a finite union of double cosets of some $\mathcal S$-subgroup. \end{itemize} The set $\mathcal B$ is an $({\mathbb F} G,{\mathbb F} G)$-sub-bimodule. \begin{itemize} \item[{\bf Claim 2.}] $H^n(\mathcal S,{\mathbb F} G)$ is isomorphic to $\mathcal B$ as a $({\mathbb F} G,{\mathbb F} G)$-bimodule. \end{itemize} We need to understand the connecting maps which are involved in the colimit formulation for $H^*(\mathcal S,{\mathbb F} G)$. When $H\subseteq L$ are $\operatorname{PD}^n$-groups then $H$ has finite index in $L$ and there are commutative diagrams \[ \xymatrix{ H^i(L,M)\ar[r]\ar[d]^{\text{Res}}&\operatorname{Tor}_{n-i}^{{\mathbb F} L}(M,{\mathbb F})\ar[d]^{\text{Tr}}\\ H^i(H,M)\ar[r]&\operatorname{Tor}_{n-i}^{{\mathbb F} H}(M,{\mathbb F})\\ } \] where the horizontal maps are the duality isomorphisms, the left hand vertical map is the ordinary restriction map in cohomology and the right hand map is {\em transfer}. When $i=n$ this specializes to \[ \xymatrix{ H^n(L,M)\ar[r]\ar[d]^{\text{Res}}&M\otimes_{{\mathbb F} L}{\mathbb F}\ar[d]^{\text{Tr}}\\ H^n(H,M)\ar[r]&M\otimes_{{\mathbb F} H}{\mathbb F}\\ } \] and here the transfer map is easy to describe: it is given by $$m\otimes1\mapsto\sum_{t\in T}mt\otimes1$$ where $T$ is a left transversal to $H$ in $L$ (i.e. $L$ is the disjoint union of the left cosets $tH$, $t\in T$). We are interested in taking $M:={\mathbb F} G$. Now ${\mathbb F} G\otimes_{{\mathbb F} L}{\mathbb F}$ is isomorphic as left ${\mathbb F} G$-module to the submodule of $\mathcal P$ comprising subsets which are finite unions of right cosets of $L$. Similarly $H$ is isomorphic to the left module of finite unions of cosets of $H$. From this viewpoint, the transfer map is simply induced by inclusion of sets. On passing to the colimit over all members of $\mathcal S$ we obtain Claim 2. We know that $H^{n+1}(G,{\mathbb F} G)$ is non-zero and we shall take advantage of this to construct a graph with more than one end on which $G$ acts in a useful way. We have $$0\ne H^{n+1}(G,{\mathbb F} G)\cong H^1(G/\mathcal S,H^n(\mathcal S,{\mathbb F} G)).$$ Substituting into the five term exact sequence we obtain an exact sequence $$0\to H^{n+1}(G,{\mathbb F} G)\to H^1(G,H^n(\mathcal S,{\mathbb F} G))\to H^1(\mathcal S,H^n(\mathcal S,{\mathbb F} G))^G.$$ Thus we have an exact sequence $$0\to H^{n+1}(G,{\mathbb F} G)\to H^1(G,\mathcal B)\to H^1(\mathcal S,\mathcal B).$$ This identifies the dualizing module of $G$ over ${\mathbb F}$ with the kernel of a restriction map in first cohomology. To compute the first cohomology of $G$ observe that the power set $\mathcal P$ of $G$ is a coinduced ${\mathbb F} G$-module on which cohomology vanishes and we can view $\mathcal B$ as a submodule. The short exact sequence $\mathcal B\to\mathcal P\to \mathcal P/\mathcal B$ yields the exact sequence $$0\to\mathcal B^G\to\mathcal P^G\to(\mathcal P/\mathcal B)^G\to H^1(G,\mathcal B)\to0$$ in cohomology. Note that $\mathcal P^G=\{\emptyset,G\}={\mathbb F}$ and the assumption that all members of $\mathcal S$ have infinite index in $G$ implies that $\mathcal B^G=0$. So the exact sequence simplifies to $$0\to{\mathbb F}\to(\mathcal P/\mathcal B)^G\to H^1(G,\mathcal B)\to0$$ The first cohomology group is most easily viewed as the quotient derivations modulo inner derivations. Writing $\mathcal P_{\mathcal S}$ for the preimage of $(\mathcal P/\mathcal B)^G$ under the natural map $\mathcal P\to\mathcal P/\mathcal B$ we have the commutative diagram \[ \xymatrix{ &&0\ar[d]&0\ar[d]\\ &&\mathcal B\ar@{=}[r]\ar[d]&\operatorname{Ider}(G,\mathcal B)\ar[d]\\ 0\ar[r]&{\mathbb F}\ar[r]\ar@{=}[d]&\mathcal P_{\mathcal S}\ar[r]\ar[d]&\operatorname{Der}(G,\mathcal B)\ar[r]\ar[d]&0\\ 0\ar[r]&{\mathbb F}\ar[r]&(\mathcal P/\mathcal B)^G\ar[r]\ar[d]&H^1(G,\mathcal B)\ar[r]\ar[d]&0\\ &&0&0\\ } \] with exact rows and columns. The set $\mathcal P_{\mathcal S}$ consists of those subsets $B$ of $G$ such that for all $g\in G$, the symmetric difference $B+Bg$ belongs to $\mathcal B$. Such a set gives rise to a derivation defined by $$g\mapsto B+Bg.$$ We have identified $H^{n+1}(G,{\mathbb F} G)$ with the kernel of the restriction map $H^1(G,\mathcal B)\to H^1(\mathcal S,\mathcal B)$. Let $\xi$ be a non-zero element of this kernel. Then $\xi$ is represented by a derivation $\delta:G\to\mathcal B$ and this restricts to an inner derivation on some subgroup $H$ in $\mathcal S$. Our derivation arises from a choice of $B\in\mathcal P$: $$\delta g=B+Bg$$ for all $g$. The restriction condition says that there is a set $A\in\mathcal B$ such that $$B+Bh=A+Ah$$ for all $h\in H$. We can choose a subgroup $L$ contained in $H$ which is also a member of $\mathcal S$ such that $$A=LAL.$$ On restriction to $L$ we find that $$B+B\ell=\emptyset$$ for all $\ell\in L$. This says that $B=BL$. It follows that the number of ends of the pair $G,L$ is at least $2$: $$e(G,L)\ge2.$$ In the terminology of Dunwoody and Swenson, $L$ has codimension one in $G$ and a splitting of $G$ can be found using methods closely related to theirs, \cite{ds}. Here we shall give an argument based on the earlier result \cite{dunwoody} of Dunwoody. Let $X$ be a finite set of generators for $G$. We now construct a graph $\Gamma$ with a left action of $G$. The vertex set $V$ of $\Gamma$ is the set of cosets $gL$ of $L$. The edge set $E$ of $\Gamma$ is defined to be a subset of $V\times V$: $$E=\{(gL,gxL):\ g\in G, x\in X\}.$$ A typical edge $(gL,gxL)$ has initial vertex $gL$ and terminal vertex $gxL$. Clearly $\Gamma$ admits a left action of $G$. Also $\Gamma$ is connected because $X$ generates $G$. For each vertex $gL$ in $\Gamma$ either $gL\subseteq B$ or $gL\subseteq B^*$. \begin{itemize} \item[{\bf Claim 3.}] There are only finitely many edges $e$ of $\Gamma$ having one vertex in $B$ and one vertex in $B^*$ \end{itemize} To establish the claim, consider an edge $(gL,gxL)$ with $gL\subset B$ and $gxL\subset B^*$. Then $g\in B\setminus Bx^{-1}$. Similarly, if $gL\subset B^*$ and $gxL\subset B$ then $g\in Bx^{-1}\setminus B$. Thus, if the edge $(gL,gxL)$ has exactly one of its vertices in $B$ then this reasoning shows that $$g\in \bigcup_{x\in X}B+Bx^{-1},$$ and also $$gx\in \bigcup_{x\in X}B+Bx.$$ Set $$Y:=\left(\bigcup_{x\in X}(B+Bx^{-1})\cup (B+Bx)\right)L.$$ Then $Y$ is a union of finitely many left cosets of $L$ and every edge with exactly one vertex in $B$ has both its vertices in $Y$. This proves the Claim 3. It now follows from Dunwoody's result \cite{dunwoody} that $G$ splits over a subgroup commensurable with $L$. \end{proof} For Theorem C shall need to appeal to a result about group actions on trees. The following is the main theorem of \cite{bf}. \begin{theorem}[Bestvina and Feighn (1991)]\label{bf} Let $G$ be a group of type $\operatorname{FP}_2$ over ${\mathbb F}$. Then there exists an integer $\gamma(G)$ such that the following holds. If $T$ is a reduced $G$-tree with small edge stabilizers, then the number of vertices in $T/G$ is bounded by $\gamma(G)$. \end{theorem} The precise statement in \cite{bf} assumes that $G$ is finitely presented. However, in the subsequent remark (8), the authors state that the result holds for {\em almost finitely presented groups}, i.e. groups of type $\operatorname{FP}_2$ over ${\mathbb F}$, and that their proof requires absolutely no change. The reason for this is that finite presentation is used only to manufacture a connected $2$-dimensional CW-complex $X$ on which $G$ acts freely and cocompactly in which every {\em track} separates: this last condition is guaranteed when $H^1(X,{\mathbb F})=0$ and the construction of $X$ is therefore possible for any almost finitely presented $G$. The following definitions are supplied in \cite{bf} and we restate them so that the reader can see exactly how Theorem \ref{bf} applies to our situation. \begin{definition}\label{def:min red hyp} Let $G$ be a group and let $T$ be a $G$-tree. \begin{enumerate} \item The action of $G$ on $T$ is said to be {\em minimal} if and only if there are no proper invariant subtrees. \item The $G$-tree $T$ is called {\em reduced} if and only if it is minimal and in addition every vertex of valency $2$ properly contains the stabilizers of the two incident edges. \item When $T$ is minimal, it is called {\em hyperbolic} if and only if there exist two hyperbolic elements in $G$ (these being elements which have no fixed points but do have an invariant line, called the axis) whose axes intersect in a compact set. In this case, $G$ contains a free group on $2$ generators: in fact, sufficiently high powers of the two group elements freely generate a free group. \end{enumerate} \end{definition} \begin{definition}\label{def:small} A group $G$ is called {\em small} if and only if it does not admit a hyperbolic action on any minimal $G$-tree. In particular, if $G$ has no non-cyclic free subgroups then $G$ is small. Polycyclic-by-finite groups are small. \end{definition} The following result is a considerable generalization of the main theorem of \cite{phk-commentari}. Note that the hypothesis {\em all subgroups commensurable with $K$ are small} is clearly satisfied if $K$ is polycyclic-by-finite. The main theorem of \cite{phk-commentari} deals with the very special case when $K$ is infinite cyclic and $G$ has cohomological dimension $\le2$. \begin{theoremC}\label{C} Let $G$ be a group satisfying the conditions $(\alpha),(\beta),(\gamma)$ of Theorem B. Suppose that all subgroups commensurable with $K$ are small. Then there is a $G$-tree $T$ such that all vertices and edges have stabilizers commensurable with $K$. \end{theoremC} \begin{proof} We give the general argument below, but first, for motivation, we consider a special case which indicates how the general argument must proceed. By Theorem B, $G$ splits over a subgroup commensurable with $K$. Suppose for the sake of argument that $G=U*_HV$ is a free product with amalgamation where $H$ is commensurable with $K$. Since $G$ and $H$ are both finitely generated it is necessarily the case that $U$ and $V$ are also finitely generated. Both $U$ and $V$ satisfy all the hypotheses of Theorem $B$ and we deduce that either $|U:H|$ is finite or $U$ splits over a subgroup $L$ commensurable with $H$. If the latter holds, let $T$ be the corresponding $U$-tree. The subgroup $H$ acts on $T$ and has finite orbits on the edges of $T$. Therefore there must be at least one vertex fixed by $H$. This means that the splitting of $U$ is compatible with the original splitting of $G$ and we can find a $G$-tree combining both splittings of $G$ in which there are two orbits of edges corresponding to the subgroups $H$ and $L$. The process can be continued but we can appeal to Theorem \ref{bf} to be sure that it breaks off. When it breaks off we reach a situation where all the vertex groups as well as all the edge groups are commensurable with $K$. More precisely, choose a reduced $G$-tree $T$ with finitely many orbits of edges, stabilizers all commensurable with $K$, and subject to these conditions, with the maximum possible number of orbits of vertices. The existence of such is guaranteed by Theorem \ref{bf}. This $G$-tree has finitely many orbits of edges and vertices and the fact that $G$ itself and all the edge stabilizers are finitely generated forces the vertex stabilizers to be finitely generated. We claim that every vertex stabilizer here is also commensurable with $K$. Suppose not. Then there is a vertex $v$ whose stabilizer $G_v$ is not commensurable with $K$. Let $e$ be an edge incident with $v$. Since $G_e$ and $K$ {\em are} commensurable we have that $G_v$ and $G_e$ are not commensurable. However, $G_e\subseteq G_v$ and therefore $G_e$ has infinite index in $G_v$. We may now apply Theorem B to split $G_v$ over a subgroup $H$ commensurable with $G_e$. Let $T'$ be the corresponding $G_v$-tree and let $e_*$ be an edge of $T'$ with stabilizer $H$. Let $E_0$ be the set of edges which are incident with the vertex $v$ in the original tree $T$. For each $e_0\in E_0$, let $G_{e_0}$ denote the stabilizer of $e_0$ for the action of $G$ on $T$. Since $G_{e_0}\subseteq G_v$ we have an action of $G_{e_0}$ on $T'$. Moreover, $|G_{e_0}:G_{e_0}\cap H|$ is finite and so the $G_{e_0}$-orbit of $e_*$ in $T'$ is finite. It follows that $G_{e_0}$ fixes a vertex $v(e_0)$ of $T'$. We can now build a new $G$-tree by blowing up each of the vertices in the $G$-orbit of $v$ using the tree $T'$. We remove the vertices $v\cdot G$ of the orbit of $v$ and replace them with the forest $T'\times_{G_v}G$. The loose edge $e_0$ can be joined to the vertex $v(e_0)$ in the primary copy $T'\times 1$ of $T'$ and we can repeat this for the other edges incident with $v$. The process can be carried equivariantly over the orbits of loose edges. In this way we obtain a $G$-tree with a greater number of orbits of vertices. This contradicts the assumption and Theorem C follows. \end{proof} Other approaches to constructing the simplicial actions on trees for Theorem B and to deducing Theorem C from Theorem B can be found in the work of Dunwoody and Swenson \cite{ds} and of Mosher, Sageev and Whyte \cite{msw1,msw2}. However the construction of almost invariant sets through the use of our new spectral sequence is novel and essential for our arguments: we do not know of any alternative to this line of reasoning beyond the two dimensional case considered in \cite{phk-commentari}. \section{An application to Poincar\'e duality groups} We illustrate the potential of Theorems B and C by using them to establish the following result which tidies up and extends some of the considerations in \cite{kr}. \begin{theorem}\label{pd} Let $G$ be a $\operatorname{PD}^{n+1}$-group and let $H$ be a $\operatorname{PD}^n$-subgroup. Then either $|\operatorname{Comm}_G(H):H|$ is finite or $|G:\operatorname{Comm}_G(H)|$ is finite. In the latter case there is a subgroup $K$ commensurable with $H$, normal in $\operatorname{Comm}_G(H)$ such that $\operatorname{Comm}_G(H)/H$ is either infinite cyclic or infinite dihedral. \end{theorem} This result should not be regarded as an advance in itself. Indeed it is clear that this and further results can and have been proved by other methods in the work of Scott and Swarup \cite{ss}. However, it gives an indication of how our results may prove helpful in the study of Poincar\'e duality groups. The following will be needed in our proof of \ref{pd}. \begin{lemma}\label{asc} Let $G$ be a group which is the union of a strictly ascending sequence $$B_0<B_1<B_2<\dots$$ of $\operatorname{PD}^n$-groups $B_i$. Then $G$ has cohomological dimension $n+1$. \end{lemma} \begin{proof} $H^{n+1}(G,M)$ is isomorphic to ${\displaystyle\lim_{\longleftarrow}}^1 H^n(B_i,M)$ for any $G$-module $M$. Taking $M={\mathbb F} G$ one can calculate that this ${\displaystyle\lim_{\longleftarrow}}^1 $ does not vanish. $H^n(B_i,{\mathbb F} G)$ is isomorphic to ${\mathbb F} G/B_i$ and the connecting maps in the limit system $$\dots\to {\mathbb F} G/B_i\to \dots \to {\mathbb F} G/B_1\to {\mathbb F} G/B_0$$ all strict inclusions. (See the argument for Claim 2 in the proof of Theorem B.) We can view the inverse limit system as sitting inside the constant system $({\mathbb F} G/B_0)$ in which the connecting maps are the identity maps. We thus have a short exact sequence of limit systems $$({\mathbb F} G/B_i)\rightarrowtail({\mathbb F} G/B_0)\twoheadrightarrow (U_i)$$ where the connecting maps in the quotient system $(U_i)$ are surjections with non-zero kernels. Applying the ${\displaystyle\lim_{\longleftarrow}}$-${\displaystyle\lim_{\longleftarrow}}^1 $ exact sequence we obtain the exact sequence $${\mathbb F} G/B_0\to{\displaystyle\lim_{\longleftarrow}} U_i\to{\displaystyle\lim_{\longleftarrow}}^1 {\mathbb F} G/B_i\to0.$$ The left hand map here cannot be surjective because ${\mathbb F} G/B_0$ is countable whereas ${\displaystyle\lim_{\longleftarrow}} U_i$ has cardinality $2^{\aleph_0}$. Therefore the right hand group is non trivial, completing the proof. \end{proof} We shall also need Strebel's fundamental dimension theorem \cite{strebdim} which says that all subgroups of infinite index in a $\operatorname{PD}^{n+1}$-group have cohomological dimension $\le n$. \begin{proof}[Proof of \ref{pd}] As a first step we prove the \begin{itemize} \item[{\bf Claim.}] Every finitely generated $S\le\operatorname{Comm}_G(H)$ such that $|H:H\cap S|<\infty$ is either commensurable with $H$ or of finite index in $G$. \end{itemize} Let $S$ be such a subgroup. We can apply Theorem B to the group $S$ with subgroup $H\cap S$. This shows that either $|S:H\cap S|$ is finite or $S$ splits over a subgroup commensurable with $H$. In the latter case, the proof of Theorem B shows that $S$ is a duality group of dimension $n+1$ over any field and so by Strebel's theorem it must have finite index in $G$. We consider two cases which together cover all eventualities. \begin{itemize} \item[Case 1.] There is a finitely generated subgroup $S$ of $\operatorname{Comm}_G(H)$ with $$|S:H|=\infty.$$ We show that in this case, $\operatorname{Comm}_G(H)$ has finite index in $G$ and the existence of $K$ can be established. \end{itemize} Applying Theorem B, we know that $S$ splits over a subgroup commensurable with $H$. Since the edge group in the splitting of $S$ is commensurable with $H$, it is finitely generated. Since $S$ is also finitely generated it follows that the vertex groups in the splitting of $S$ are finitely generated. The vertex groups have infinite index in $S$ and the above shows that they are commensurable with $H$. Therefore either $S=J*_KL$, a free product with amalgamation in which $J\ne K\ne L$ are all commensurable with $H$, or $S=B*_K,t$ is an HNN-extension in which the base and associated subgroups are commensurable with $H$. If $S$ is a non-ascending HNN-extension then one can use the Kurosh subgroup theorem to exhibit finitely generated subgroups of $S$ which have infinite index and which contain an infinite index subgroup commensurable with $H$. This contradicts the claim. Similarly, if $G$ is an amalgamation in which one of the indices $|J:K|$, $|L:K|$ is $\ge3$ then we can again find intermediate finitely generated subgroups which contradict the claim. If $S=B*_Bt$ is a strictly ascending HNN-extension then the chain of subgroups $$B\subset B^t\subset B^{t^2}\subset\dots$$ has union of cohomological dimension $n+1$ by Lemma \ref{asc}. But this contradicts Strebel's theorem and so cannot happen. Therefore either $S$ is an amalgamation in which $|J:K|=|L:K|=2$ in which case $K\lhd S$ and $S/K\cong D_\infty$, or $G$ is a stationary HNN-extension meaning $B=K\lhd S$ and $S/K\cong C_\infty$. Now $S$ has finite index in $G$ and normalizes $K$. Therefore $K$ has only finitely many distinct conjugates. Thus the subgroup $K_0$ defined by $$K_0:=\bigcap_{g\in\operatorname{Comm}_G(H)}K^g$$ is a finite intersection of subgroups commensurable with $H$ and is therefore itself commensurable with $H$. Also $K_0$ is normal in $\operatorname{Comm}_G(H)$ and the corresponding quotient is virtually cyclic. If we set $K_1/K_0$ equal to the largest finite normal subgroup of $\operatorname{Comm}_G(H)$ then the quotient $\operatorname{Comm}_G(H)/K_1$ is either infinite cyclic or infinite dihedral. \begin{itemize} \item[Case 2.] For every finite subset $F$ of $\operatorname{Comm}_G(H)$, $$|\langle H\cup F\rangle:H|<\infty.$$ We show that in this case, $H$ has finite index in $\operatorname{Comm}_G(H)$. \end{itemize} In this case, if $|\operatorname{Comm}_G(H):H|$ is infinite then we can choose a strictly ascending chain of finite extensions of $H$ by successively increasing the size of finite set $F$. The union of such a chain has cohomological dimension $n+1$ by Lemma \ref{asc} and therefore finite index in $G$. This is a contradiction because $G$ is finitely generated while the union of the chain is not. Thus $|\operatorname{Comm}_G(H):H|<\infty$ as claimed. \end{proof} \section{An easy application in which the spectral sequence collapses} We conclude by mentioning a very simple application of the theory motivated by the notion of complete cohomology as described in \cite{bensoncarlson,goichot,mislin}. This again is a significant generalization of one of the results in \cite{phk-commentari}. \begin{definition}\label{def:already complete} We shall say that a group $G$ {\em already has complete cohomology} if and only if the cohomology functors $H^n(G,{\phantom M})$ vanish on projective modules for all $n$. This is equivalent to asserting that the natural map $H^n(G,{\phantom M})\to\widehat H^n(G,{\phantom M})$ is always an isomorphism. For example, free abelian groups of infinite rank already have complete cohomology whereas finite groups never enjoy the property. \end{definition} \begin{theorem}\label{star} Let $G$ be a group and let $\mathcal S$ be an admissible family of subgroups which already have complete cohomology. Then $G$ already has complete cohomology. \end{theorem} \begin{proof} The new spectral sequence collapses because the hypotheses along with Lemma \ref{injectives}(iii) ensure that $$H^*(\mathcal S,P)=0$$ for all projective modules $P$. \end{proof} This provides a transparent argument for proving one of the results of \cite{browngeoghegan} that $$H^*(F,{\mathbb Z} F)=0$$ where $F$ denotes Thompson's group given by the presentation $$F=\langle x_0,x_1,x_2,\dots:\ x_i^{-1}x^{{\phantom 1}}_jx^{{\phantom 1}}_i=x^{{\phantom 1}}_{j+1}\ (i<j)\rangle.$$ Let $A$ be the subgroup of $F$ generated by $\{x^{{\phantom 1}}_{2n+1}x_{2n}^{-1}:\ n\ge0\}$ and let $F_0$ be the subgroup of index $2$ in $F$ comprising elements which can be expressed as even weight words in the $x_i$. Thus $$A\subset F_0\subset F.$$ Clearly it suffices to prove that $$H^*(F_0,{\mathbb Z} F)=0,$$ and for this we only need to observe that the admissible family $\mathcal S$ of subgroups generated by $A$ consists entirely of free abelian groups of infinite rank. We shall write $A_m$ for the subgroup of $A$ generated by $\{x^{{\phantom 1}}_{2n+1}x_{2n}^{-1}:\ n\ge m\}$. The admissible family $\mathcal S$ is the set of subgroups which are finite intersections of conjugates $A^g$ with $g\in F_0$. Here are the precise details of the argument. \begin{lemma}\label{lem:Thompson} \begin{enumerate} \item $x^{{\phantom 1}}_1x_0^{-1}$, $x^{{\phantom 1}}_3x_2^{-1}$, $x^{{\phantom 1}}_5x_4^{-1}$, $x^{{\phantom 1}}_7x_6^{-1}$, $\dots$ is a sequence of distinct elements of $F$ which freely generate the abelian group $A$. \item For any finite subset $X$ of $F_0$, there is an $m\ge0$ such that $$A_m\subseteq\bigcap_{g\in X}A^g.$$ \item Every member of $\mathcal S$ is free abelian of infinite rank. \item $H^*(F,{\mathbb Z} F)=0$. \end{enumerate} \end{lemma} \begin{proof} \begin{enumerate} \item Notice that if $0\le m<n$ then the relations in our given presentation of $F$ immediately yield $$x_{2m}^{-1}\left(x^{{\phantom 1}}_{2n+1}x_{2n}^{-1}\right)x_{2m} =x^{{\phantom 1}}_{2n+2}x_{2n+1}^{-1}$$ and $$x_{2m+1}^{-1}\left(x^{{\phantom 1}}_{2n+1}x_{2n}^{-1}\right)x_{2m+1}= x^{{\phantom 1}}_{2n+2}x_{2n+1}^{-1}.$$ Thus $x^{{\phantom 1}}_{2m+1}x_{2m}^{-1}$ commutes with $x^{{\phantom 1}}_{2n+1}x_{2n}^{-1}$. This shows that $A$ is abelian. One can see quite easily that $A$ is free abelian on the stated generators. One way is to use the known representation of $F$ as a group of piecewise linear maps of the unit interval $[0,1]$. Alternatively, let $F_m$ be the subgroup of $F$ generated by the $x_i$ with $i\ge m$. Then each $F_m$ is isomorphic to $F=F_0$ and $F_m$ is an ascending HNN-extension over $F_{m+1}$. One can now check inductively that for each $m$, the subgroup $$x^{{\phantom 1}}_1x_0^{-1},\dots,x^{{\phantom 1}}_{2m+1}x_{2m}^{-1}$$ is free abelian of rank $m+1$ and lies in the centralizer of $F_{2m+2}$. \item Let $g$ be a word in the alphabet $x_i^{\pm1}$, $i\ge0$ with exponent sum $j$. The relations defining $F$ show that for all sufficiently large $n$, $$g^{-1}x_ng=x_{n+j}.$$ If $g$ is an element of $F_0$ then it can be expressed as a word with even exponent sum and therefore $$A_m\subset A\cap A^g$$ for sufficiently large $m$. The result for an intersection of finitely many conjugates now follows from this. \item This follows at once. \item Theorem \ref{star} shows that $H^*(F_0,{\mathbb Z} F)=0$ and then we can apply the ordinary LHS spectral sequence to the group extension $F_0\rightarrowtail F\twoheadrightarrow{\mathbb Z}/2{\mathbb Z}$. \end{enumerate} \end{proof} \section{Connection with Galois Cohomology} As a concluding remark we mention that although the derived functors $$H^n(G/\mathcal S,{\phantom M})$$ are not at first sight easily related to any familiar functors, there are nevertheless close connections with Galois cohomology of profinite groups. As an example we shall see that if $G$ is a residually finite group and $\mathcal S$ is the family of all subgroups of finite index then $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ is the category of discrete modules for the profinite completion $\widehat G$ of $G$ and $H^n(G/\mathcal S,{\phantom M})$ is isomorphic to the continuous (Galois) cohomology functor $H^n_{\operatorname{Gal}}(\widehat G,{\phantom M})$. We refer the reader to \cite{serrecohom} for an introduction to the theory. In general, given a group $G$ and admissible family $\mathcal S$, it is not immediately clear that one can form a completion analogous to the profinite completion because it may happen that $\mathcal S$ contains few, or possibly no, normal subgroups. For example if $G$ is the Baumslag--Solitar group with presentation $$\langle x,y:\ y^{-1}x^2y=x^3\rangle$$ and $\mathcal S$ is the set of subgroups commensurable with the infinite cyclic subgroup $\langle x\rangle$ then no member of $\mathcal S$ is normal. Nevertheless, one can always form a completion $\widehat G_{\mathcal S}$ for this example and any other. One can endow the completion with a product which makes it into a monoid. We do not know whether this monoid is necessarily a group in all cases, but we can show that it is a group in all the applications and examples considered in this paper. Define $\widehat G_{\mathcal S}$ to be the set of all functions $f:\mathcal S\to \mathcal P(G)$ which satisfy the conditions \begin{itemize} \item $f(H)\in H\backslash G$ for all $H\in\mathcal S$ and \item $f(K)\subseteq f(H)$ whenever $K\subseteq H$ are members of $\mathcal S$. \end{itemize} In effect, we associate the coset space $H\backslash G$ to $H$ and observe that an inclusion $K\subset H$ induces a natural surjection $K\backslash G\twoheadrightarrow H\backslash G$. Then $\widehat G_{\mathcal S}$ is the inverse limit ${\displaystyle\lim_{\longleftarrow}}\ H\backslash G$. For any $f\in\widehat G_{\mathcal S}$ and any $H\in\mathcal S$ there exists $x\in G$ such that $f(H)=Hx$. It is useful to introduce the notation $H^f$ for the conjugate $H^x$ because it depends only on $f$ and $H$; not on the particular coset representative $x$. To make $\widehat G_{\mathcal S}$ into a monoid we define the product by setting: $$f\cdot f'(H)=f(H)f'(H^f)$$ for each $H\in \mathcal S$. On the right, the product is carried through using the standard convention $AB=\{ab:a\in A,b\in B\}$ for sets $A,B\subseteq G$. It is straightforward to check that $H^{f\cdot f'}=(H^f)^{f'}$. To check the associative law: \begin{eqnarray*} (f\cdot f')\cdot f''(H)&=&f\cdot f'(H)f''\big(H^{f\cdot f'}\big)\\ &=&f(H)f'(H^f)f''\big(H^{f\cdot f'}\big),\\ \end{eqnarray*} and \begin{eqnarray*} f\cdot(f'\cdot f'')(H)&=&f(H)(f'\cdot f'')(H^f)\\ &=&f(H)f'(H^f)f''\big((H^f)^{f'}\big).\\ \end{eqnarray*} The function $e$ defined by $e(H)=H$ for all $H$ is the identity element. There is a natural map $$\widehat{\phantom g}:G\to \widehat G_{\mathcal S}$$ defined by sending each $g\in G$ to the function $\widehat g$ defined by $\widehat g(H)=Hg$. This is a homomorphism which carries $1\in G$ to $e\in\widehat G_{\mathcal S}$ because $$\widehat g\cdot\widehat{g'}(H)=% \widehat g(H)\widehat{g'}(H^g)=Hgg'.$$ The question of whether or not $\widehat G_\mathcal S$ is a group appears to be subtle. If $f\in\widehat G_{\mathcal S}$ has an inverse $f^{-1}$ then expanding the definition of $f\cdot f^{-1}(H)$ shows that $$f^{-1}(H^f)=f(H)^{-1}.$$ This shows that $f^{-1}$ is uniquely determined on the subset $\{H^f:H\in\mathcal S\}$ of $\mathcal S$. In general, for fixed $f$, the map $H\mapsto H^f$ is injective and preserves the poset structure of $\mathcal S$. To see this observe that for any finite subset $\mathcal F$ of $\mathcal S$ there is a group element $x$ such that $f(H)=Hx$ and $H^f=H^x$ for all $H\in\mathcal F$. Thus the map $H\mapsto H^f$ is given locally as conjugation by a single group element. \begin{definition} We shall say that the admissible family $\mathcal S$ is {\em stable} if and only if for all $K\le H$ in $\mathcal S$ there exists $L\in\mathcal S$ such that $L\le K$ and $L$ is normal in $H$. \end{definition} \begin{lemma} If $\mathcal S$ is a stable admissible family of subgroups of $G$ then $\widehat G_{\mathcal S}$ is a group. \end{lemma} \begin{proof} Fix $f\in\widehat G_{\mathcal S}$. Fix $H\in\mathcal S$. Choose $x\in f(H)$. Choose $K$ so that $\mathcal S\ni K\le H\cap H^f$ and $K\lhd H^f$. Choose $t\in f(K^{x^{-1}})$. Define $$f^{-1}(H)=Ht^{-1}.$$ \begin{itemize} \item[{\bf Claim 1.}] $f^{-1}$ is well defined. \end{itemize} To define $f^{-1}(H)$ we have made three choices, namely $x,K,t$. Note that $f(H)=Hx$, $H^f=H^x$ and $K^{x^{-1}}\le (H^f)^{x^{-1}}=H$ so necessarily $t\in f(H)$, $f(H)=Ht$ and $xt^{-1}\in H$. Thus $t^{-1}x\in H^x=H^f$ normalizes $K$ and $K^{x^{-1}}=K^{t^{-1}}$ so that $f(K^{t^{-1}})=tK$. Consider now a different sequence $y,L,u$ of the three choices made in the same way as $x,K,t$. Then $f(H)=Hy=Hu$ and $f(L^{u^{-1}})=uL$. Downward directedness guarantees that $tK\cap uL\ne\emptyset$. Choose $v\in tK\cap uL$. Then $tK=vK$ and $uL=vL$. Since $K$ and $L$ are contained in $H$ it follows that $Ht^{-1}=Hv^{-1}=Hu^{-1}$ as required. \begin{itemize} \item[{\bf Claim 2.}] $f^{-1}$ belongs to $\widehat G_{\mathcal S}$. \end{itemize} Given subgroups $H'\le H$ in $\mathcal S$ we can make the three choices $x,K,t$ so that $x\in f(H')$, $K\subseteq H'$ and $K\lhd H$. The choices then simultaneously supply the definitions of $f^{-1}(H')$ and $f^{-1}(H)$ so that $f^{-1}(H')=H't^{-1}$ and $f^{-1}(H)=Ht^{-1}$ have the required compatibility for the inverse system. \begin{itemize} \item[{\bf Claim 3.}] $f^{-1}$ is inverse to $f$. \end{itemize} This follows: for any $H$ we can choose $t\in G$ such that $f(H)=Ht$, $f(H^{t^{-1}})=tH$, $f^{-1}(H)=Ht^{-1}$ and so $$f^{-1}\cdot f(H)=f^{-1}(H)f(H^{f^{-1}})=Ht^{-1}tH=H=e(H).$$ Thus $f^{-1}\cdot f=e$ and since $e$ is a two-sided identity element it follows from elementary group theory that $f^{-1}$ is a two-sided inverse as required. \end{proof} Every object of $\operatorname{\Mo\HYPHEN}{\mathbb Z} G/\mathcal S$ can be endowed with a natural action of $\widehat G_{\mathcal S}$ as follows. Let $M$ be a ${\mathbb Z} G/\mathcal S$-module, let $f\in \widehat G_{\mathcal S}$ and let $m\in M$. Then we set $$m\cdot f:=m.f(H)$$ where $H$ is any choice of member of $\mathcal S$ which fixes $m$. Although the subgroups of $\mathcal S$ may not be normal, it is still true that the underlying set of $\widehat G_{\mathcal S}$ can be viewed as an inverse limit: $$\widehat G_{\mathcal S}=\limf H\backslash G$$ taken over the discrete coset spaces $$H\backslash G=\{Hg:\ g\in G\}.$$ Then we may endow $\widehat G_{\mathcal S}$ with the inverse limit topology and it becomes a topological group. One could then go on to consider the category of discrete $\widehat G_{\mathcal S}$-modules: this is equivalent to the category of $\operatorname{\Mo\HYPHEN} G/\mathcal S$-modules. The cohomology functors $H^n(G/\mathcal S,{\phantom M})$ can be identified with the continuous cohomology functors of the topological group defined for example by using continuous cocycles in the standard bar resolution construction. Of course, if $\mathcal S$ consists of normal subgroups, or more generally if $$\bigcap_{g\in G}H^g$$ belongs to $\mathcal S$ for all $H\in\mathcal S$, then $\widehat G_{\mathcal S}$ is easier to define: one can think straightforwardly in terms of the inverse limit of quotient groups.
{ "timestamp": "2005-03-24T08:06:47", "yymm": "0503", "arxiv_id": "math/0503514", "language": "en", "url": "https://arxiv.org/abs/math/0503514" }
\section{Figure captions} \noindent Fig.1. a)The complex plane of frequence. The cuts of $\Pi^{0R}$ Eq.(\ref{3}) are shown. Numbers stand for the cuts (\ref{12}). b)The complex plane of $z_1$. The cut corresponds to the cut $I$ in Fig.a. c)The complex plane of frequence. The shading marks the unphysical sheets described in the text. \noindent Fig.2. Solutions $\omega_{sd}(k)$ to Eq.(\ref{5}) at $F_0<0$. The variable $\gamma$ is $\gamma = \frac{m}{kp_F}Im~\omega_{sd}$. The shading with the right slope marks the unphysical sheet $I$. The horizontal shading marks the unphysical sheet $\tilde I$. \noindent Fig.3. The comparision of the solutions in the kinetic theory and in the RPA. The solid (dashed) lines correspond to solutions obtained at $F_0=-1.2$ ($F_0=-1.1$). \noindent Fig.4. The branches of solution $\omega_{sd}(k)$ obtained in RPA at different values of $F_0$. The curve (1) corresponds to $F_0=-0.4$; (2) $F_0=-0.9$; (3) $F_0=-1.02$; (4) $F_0=-1.1$; (5) $F_0=-1.2$. The shading marks the sheet $I$. \noindent Fig.5. The solutions to Eq. (\ref{5}) $\omega_s$ at $F_0>0$. Two symmetric solutions are presented. \noindent Fig.6. The complex plane of frequence. The solutions $\omega_s(k)$ to Eq.(\ref{1}) obtained in RPA are presented. The curve $1$ is calculated at $F_0=1$ and the curve $2$ at $F_0=2$ . The wave vector $k_d$ marks the point when the Landau damping starts at this $F_0$: $\frac{k_d}{p_F}=0.13$ when $F_0=1$, and $\frac{k_d}{p_F}=0.52$ at $F_0=2$. \noindent Fig.7. The branch $\omega_s(k)$ at $F_0=2$ is shown for the different models. The solid line is for the solutions of Eq.(\ref{5}). The real $\omega_r$ and the imaginary $\omega_i$ parts of the RPA solutions are presented by the dashed curves. The dotted line stands for the $\omega_i$ calculated by Eq.(\ref{14}). \newpage \begin{figure \centering{\epsfig{figure=f1a.eps,width=7cm} \epsfig{figure=f1b.eps,width=7cm}} \vspace{1cm} \centering{\epsfig{figure=f1c.eps,width=7cm}} \caption{} \end{figure} \begin{figure \centering{\epsfig{figure=f2.eps,width=9cm}} \caption{} \end{figure} \begin{figure \centering{\epsfig{figure=f3.eps,width=9cm}} \caption{} \end{figure} \begin{figure \centering{\epsfig{figure=f4a.eps,width=9cm}} \caption{} \end{figure} \begin{figure \centering{\epsfig{figure=f5.eps,width=9cm}} \caption{} \end{figure} \begin{figure \centering{\epsfig{figure=f5r.eps,width=9cm}} \caption{} \end{figure} \begin{figure \centering{\epsfig{figure=kdep1.eps,width=9cm}} \caption{} \end{figure} \end{document}
{ "timestamp": "2005-10-25T08:28:26", "yymm": "0503", "arxiv_id": "nucl-th/0503085", "language": "en", "url": "https://arxiv.org/abs/nucl-th/0503085" }
\section{Introduction} M.-H. Schwartz in \cite {Sch1, Sch2} introduced the technique of radial extension of stratified vector fields and frames on singular varieties, and used this to construct cocycles representing classes in the cohomology $H^*(M, M\setminus V)$, where $V$ is a singular variety embedded in a complex manifold $M$; these are now called {\it the Schwartz classes} of $V$. A basic property of radial extension is that the index of the vector fields (or frames) constructed in this way is the same when measured in the strata or in the ambient space; this is called the Schwartz index of the vector field (or frame). MacPherson in \cite{MP} introduced the notion of the local Euler obstruction, an invariant defined at each point of a singular variety using an index of an appropriate radial 1-form, and used this (among other things) to construct the homology Chern classes of singular varieties. Brasselet and Schwartz in \cite {BS} proved that the Alexander isomorphism $H^*(M, M\setminus V) \cong H_*(V)$ carries the Schwartz classes into the MacPherson classes; a key ingredient for this proof is their {\it proportionality theorem} relating the Schwartz index and the local Euler obstruction. These were the first indices of vector fields and 1-forms in the literature. Later in \cite {GSV} was introduced another index for vector fields on isolated hypersurface singularities, and this definition was extended in \cite{SS} to vector fields on complete intersection germs. This is known as the GSV-index and one of its main properties is that it is invariant under perturbations of both, the vector field and the functions that define the singular variety. The definition of this index was recently extended in \cite {BSS1} for vector fields with isolated singularities on hypersurface germs with non-isolated singularities, and it was proved that this index satisfies a proportionality property analogous to the one proved in \cite{BS} for the Schwartz index and the local Euler obstruction, the proportionality factor being now the Euler-Poincar\'e characteristic of a local Milnor fiber. In \cite{EG1} Ebeling and Gusein-Zade observed that when dealing with singular varieties, 1-forms have certain advantages over vector fields, as for instance the fact that for a vector field on the ambient space the condition of being tangent to a (stratified) singular variety is very stringent, while every 1-form on the ambient space defines, by restriction, one on the singular variety. They adapted the definition of the GSV-index to 1-forms on complete intersection germs with isolated singularities, and proved a very nice formula for it in the case when the form is holomorphic, generalizing the well-known formula of L\^e-Greuel for the Milnor number of a function. This article is about 1-forms on complex analytic varieties and it is particularly relevant when the variety has non-isolated singularities. We show in section 2 how the radial extension technique of M.-H. Schwartz can be adapted to 1-forms, allowing us to define {\it the Schwartz index} of 1-forms with isolated singularities on singular varieties. Then we see (section 3) how MacPherson's local Euler obstruction, adapted to 1-forms in general, relates to the Schwartz index, thus obtaining a proportionality theorem for these indices analogous to the one in \cite{BS} for vector fields. We then extend (in section 4) the definition of the GSV-index to 1-forms with isolated singularities on (local) complete intersections with non-isolated singularities that satisfy the Thom $a_f$-condition (which is always satisfied if the variety is a hypersurface), and we prove the corresponding proportionality theorem for this index. When the form is the differential of a holomorphic function $h$, this index measures the number of critical points of a generic perturbation of $h$ on a local Milnor fiber, so it is analogous to invariants studied by a number of authors (see for instance \cite {Go, IS,STV}). Section 1 is a review of well-known facts about real and complex valued 1-forms. The radial extension of 1-forms can be made global on compact varieties, and it can also be made for frames of differential 1-forms. One gets in this way the dual Schwartz classes of singular varieties, which equal the usual ones up to sign. One also has the dual Chern-Mather classes of $V$, already envisaged in \cite{Sa}, and the proportionality formula 3.3 can be used as in \cite{BS} to express the dual Chern-Mather classes as ``weighted" dual Schwartz classes, the weights been given by the local Euler obstruction. Similarly, in analogy with Theorem 1.1 in \cite {BSS1}, the corresponding GSV-index and the proportionality Theorem 4.4 extend to frames and can be used to express the dual Fulton-Johnson classes of singular hypersurfaces embedded with trivial normal bundle in compact complex manifolds, as ``weighted" dual Schwartz classes, the weights been now given by the Euler-Poincar\'e characteristic of the local Milnor fiber. This work was done while the second and third named authors were visiting the ``Institut de Math\'ematiques de Luminy", France; they acknowledge the support of the CNRS, France and the ``Universit\'e de la M\'editerran\'ee". The authors thank J. Sch\"urmann for his comments and suggestions on the first version of the paper. In particular, he gave us an alternative proof of Theorem 3.3 in the case of the differential form associated to a Morse function, using stratified Morse theory and the micro-local index formula in \cite {Schu2}. \section{Some basic facts about 1-forms} In this section we study some basic facts about the geometry of 1-forms and the interplay between real and complex valued 1-forms on (almost) complex manifolds, which plays an important role in the sequel. The material here is all contained in the literature; we include it for completness and to set up our notation with no possible ambiguities. We give precise references when appropriate. Let $M$ be an almost complex manifold of real dimension $2m>0$. Let $TM$ be its complex tangent bundle. We denote by $T^*M$ the cotangent bundle of $M$, dual of $TM$; each fiber $(T^*M)_x$ consists of the $\mathcal C$-linear maps $TM_x \to \mathcal C$. Similarly, we denote by $T_\mathbbm{R} M$ the underlying real tangent bundle of $M$; it is a real vector bundle of fiber dimension $2m$, endowed with a canonical orientation. Its dual $T_\mathbbm{R}^*M$ has as fiber the $\mathbbm{R}$-linear maps $(T_{\mathbbm{R}}M)_x \to \mathbbm{R}$. \vskip 0.3cm \noindent {\bf 1.1 Definition.} Let $A$ be a subset of $M$. By a real (valued) 1-form $\eta$ on $A$ we mean the restriction to $A$ of a continuous section of the bundle $T_\mathbbm{R}^*M$, i.e., for each $x \in A$, $\eta_x$ is an $\mathbbm{R}$- linear map $(T_{\mathbbm{R}}M)_x \to \mathbbm{R}$. We usually drop the word ``valued" here and speak only of real 1-forms on $A$. Similarly, a complex 1-form $\omega$ on $A$ means the restriction to $A$ of a continuous section of the bundle $T^*M$, i.e., for each $x \in A$, $\omega_x$ is a $\mathcal C$-linear map $(TM)_x \to \mathcal C$. \vskip.1cm Notice that the kernel of a real form $\eta$ at a point $x$ is either the whole fiber $(T_\mathbbm{R} M)_x$ or a real hyperplane in it. In the first case we say that $x$ is a singular point (or zero) of $\eta$. In the second case the kernel $ker\, \eta_x$ splits $(T_\mathbbm{R} M)_x$ in two half spaces $(T_\mathbbm{R} M^{\pm})_x$; in one of these the form takes positive values, in the other $\eta(v)$ is negative. We recall that a vector field $v$ in $\mathbbm{R}^N$ is radial at a point $x_o$ if it is transversal to every sufficiently small sphere around $x_o$ in $\mathbbm{R}^N$. The duality between real 1-forms and vector fields assigns to each tangent vector $\partial / \partial x_i$ the form $dx_i$ (extending it by linearity to all tangent vectors). This refines the classical duality that assigns to each hyperplane in $\mathbbm{R}^{N}$ the line orthogonal to it and motivates the following definition (c. f. \cite {EG1, EG2}): \vskip 0.3cm \noindent {\bf 1.2 Definition.} A real 1-form $\eta$ on $M$ is {\bf radial} (outwards-pointing) at a point $x_o \in M$ if, locally, it is dual over $\mathbbm{R}$ to a radial outwards-pointing vector field at $x_o$. Inwards-pointing radial vector fields are defined similarly. In other words, $\eta$ is {\bf radial} at a point $x_o$ if it is everywhere positive when evaluated in some radial vector field at $x_o$. \vskip.1cm Thus, for instance, if for a fixed $x_o \in M$ we let $\rho_{x_o}(x)$ be the function $\Vert x - x_o \Vert^2$ (for some Riemmanian metric), then its differential is a radial form. \vskip 0.3cm \noindent {\bf 1.3 Remark.} The concept of radial forms was introduced in \cite{EG1}. In \cite {EG2} radial forms are defined using more relaxed conditions than we do here. However this is a concept "imported" from the corresponding notion of radial vector fields, so we use definition 1.2. \vskip.1cm A complex 1-form $\omega$ on $A \subset M$ can be written in terms of its real and imaginary parts: $$\omega \,=\, Re\,(\omega) \,+\, i\, Im \,(\omega)\,.$$ Both $ Re\,(\omega)$ and $ Im \,(\omega)$ are real 1-forms, and the linearity of $\omega$ implies that for each tangent vector one has: $$ Im \,(\omega)(v) \,=\, - Re \, (\omega)(iv) \,,$$ thus \[\omega (v) \,=\, Re\,(\omega) (v) \,-\, i\, Re \,(\omega) (iv)\,. \] In other words the form $\omega$ is determined by its real part and one has a 1-to-1 correspondence between real and complex forms, assigning to each complex form its real part, and conversely, to a real 1-form $\eta$ corresponds the complex form $\omega$ defined by: \[\omega(v) \,=\, \eta(v) - i \eta(iv)\,. \] This statement (noted in \cite {EG2},\cite{GMP}) refines the obvious fact that a complex hyperplane $P$ in $\mathcal C^m$, say defined by a linear form $H$, is the intersection of the real hyperplanes $\widehat H := \{Re\, H = 0\}$ and $\,i\, \widehat H$. This justifies the following definition: \vskip 0.3cm \noindent {\bf 1.4 Definition.} A complex 1-form $\omega$ is {\bf radial} at a point $x \in M$ if its real part is radial at $x$. \vskip.1cm Recall that the Euler class of an oriented vector bundle is the primary obstruction for constructing a non-zero section \cite {St}. In the case of the bundle $T^*_\mathbbm{R} M$, this class equals the Euler class $\hbox{Eu}(M)$ of the underlying real tangent bundle $T_\mathbbm{R} M$, since they are isomorphic. Thus, if $M$ is compact then its Euler class evaluated on the orientation cycle of $M$ gives the Euler-Poincar\'e characteristic $\chi(M)$. We can say this in different words: let $\eta$ be a real 1-form on $M$ with isolated (hence finitely many) singularities $x_1, \cdots,x_r$. At each $x_i$ this 1-form defines a map, $\mathbbm{S}_\varepsilon \buildrel{\eta/\Vert \eta \Vert}\over \longrightarrow \mathbbm{S}^{2m-1}$, from a small sphere in $M$ around $x_i$ into the unit sphere in the fiber $(T^*_\mathbbm{R} M)_x$. The degree of this map is the {\bf Poincar\'e-Hopf} local index of $\eta$ at $x_i$, that we may denote by $\hbox{Ind}_{PH}(\eta, x_i)$. Then the total index of $\eta$ in $M$ is by definition the sum of its local indices at the $x_i$ and it equals $\chi(M)$. Its Poincar\'e dual class in $H^{2m}(M)$ is the Euler class of $T^*_\mathbbm{R} M \cong T_\mathbbm{R} M$. \vskip.1cm More generally, if $M$ is a compact, $C^\infty$ manifold of real dimension $2m$ with non-empty boundary $\partial M$ and a complex structure in its tangent bundle, one can speak of real and complex valued 1-forms as above. Elementary obstruction theory (see \cite {St}) implies that one can always find real and complex 1-forms on $M$ with isolated singularities, all contained in the interior of $M$. In fact, if a real 1-form $\eta$ is defined in a neighborhood of $\partial M$ in $M$ and it is non-singular there, then we can always extend it to the interior of $M$ with finitely many singularities, and its total index in $M$ does not depend on the choice of the extension. \vskip 0.3cm \noindent {\bf 1.5 Definition.} Let $M$ be an almost complex manifold with boundary $\partial M$ and let $\omega$ be a (real or complex) 1-form on $M$, non-singular on a neighborhood of $\partial M$; let $Re \, \omega$ be its real part if $\omega$ is a complex form, otherwise $Re \, \omega = \omega$ for real forms. The form $\omega$ is {\bf radial} at the boundary if for each vector $v(x) \in TM$, $x \in \partial M$, which is normal to the boundary (for some metric), pointing outwards of $M$, one has $Re \, \omega(v(x)) > 0$. \vskip.2cm By the theorem of Poincar\'e-Hopf for manifolds with boundary, if a real 1-form $\eta$ is radial at the boundary and $M$ is compact, then the total index of $\eta$ is $\chi(M)$. \vskip.1cm We now make similar considerations for complex 1-forms. We let $M$ be a compact, $C^\infty$ manifold of real dimension $2m$ (with or without boundary $\partial M$), with a complex structure in its tangent bundle $TM$. Let $T^*M$ be as before, the cotangent bundle of $M$, i.e., the bundle of complex valued continuous 1-forms. The top Chern class $c^m(T^*M)$ is the primary obstruction for constructing a section of this bundle, i.e., if $M$ has empty boundary, then $c^m(T^*M)$ is the number of points, counted with their local indices, of the zeroes of a section $\omega$ of $T^*M$ (i.e., a complex 1-form) with isolated singularities (i.e., points where it vanishes). It is well known (see for instance \cite {Mi}) that one has: $$c^m(T^*M) \,=\, (-1)^m \,c^m(TM)\,.$$ This corresponds to the fact that at each isolated singularity $x_i$ of $\omega$ one has two local indices: one of them is the index of its real part defined as above, $\hbox{Ind}_{PH}(Re \, \omega, x_i)$; the other is the degree of the map $\mathbbm{S}_\varepsilon \buildrel{\omega/\Vert \omega \Vert}\over \longrightarrow \mathbbm{S}^{2m-1}$, that we denote by $\hbox{Ind}_{PH}(\omega, x_i)$. These two indices are related by the equality: \[\hbox{Ind}_{PH}(\omega, x_i) \,=\, (-1)^m \, \hbox{Ind}_{PH}(Re \, \omega, x_i)\,, \] and the index on the right corresponds to the local Poincar\'e-Hopf index of the vector field defined by duality near $x_i$. For instance, the form $\omega = \sum z_i dz_i$ in $\mathcal C^m$ has index $1$ at $0$, while its real part $\sum(x_i dx_i - y_i dy_i)$ has index $(-1)^m$. If we take $M$ as above, compact and with possibly non-empty boundary, and $\omega$ is a complex 1-form with isolated singularities in the interior of $M$ and radial on the boundary, then (by the previous considerations) the total index of $\omega$ in $M$ is $(-1)^m \, \chi (M)$. We summarize some of the previous discussion in the following theorem (c.f. \cite {EG1, EG2}): \vskip 0.3cm \noindent \proclaim{1.6 Theorem}{Let $M$ be a compact, $C^\infty$ manifold of real dimension $2m$ (with or without boundary $\partial M$), with a complex structure in its tangent bundle $TM$. Let $T^*_\mathbbm{R} M$ and $T^*M$ be as before, the bundles of real and complex valued continuous 1-forms on $M$, respectively. Then: \vskip.1cm \noindent {\bf i) } Every real 1-form $\eta$ on $M$ determines a complex 1-form $\omega$ by the formula $$ \omega(v) \,=\, \eta(v) - i \eta(iv) \,;$$ so the real part of $\omega$ is $\eta$. \vskip.1cm \noindent {\bf ii) } The local Poincar\'e-Hopf indices at an isolated singularity of a complex 1-form and its real part are related by: \[ {\rm{Ind}}_{PH}(\omega, x_i) \,=\, (-1)^m \, {\rm{Ind}}_{PH}(Re \, \omega, x_i)\,. \] \vskip.1cm \noindent {\bf iii) } If a real 1-form on $M$ is radial at the boundary $\partial M$, then its total Poincar\'e-Hopf index in $M$ is $\chi(M)$. In particular, a radial real 1-form has local index 1. \vskip.1cm \noindent {\bf iv) } If a complex 1-form on $M$ is radial at the boundary $\partial M$, then its total Poincar\'e-Hopf index in $M$ is $(-1)^m \chi(M)$. } \vskip 0.3cm \noindent {\bf 1.7 Remark.} One may consider frames of complex 1-forms on $M$ instead of a single 1-form. This means considering sets of $k$ complex 1-forms, whose singularities are the points where these forms become linearly dependent over $\mathcal C$. By definition (see \cite {St}) the primary obstruction for constructing such a frame is the Chern class $c^{m-k+1}(T^*M)$, so these classes also have an expression similar to 1.6 but using indices of frames of 1-forms. One always has $c^i(T^*M) = (-1)^i c^i(TM)$. Thus the Chern classes, and all the Chern numbers of $M$, can be computed using indices of either vector fields or 1-forms. \section{Radial extension and the Schwartz index} In the sequel we will be interested in considering forms defined on singular varieties in a complex manifold, so we introduce some standard notation. Let $V$ be a reduced, equidimensional complex analytic space of dimension $n$ in a complex manifold $M$ of dimension $m$, endowed with a Whitney stratification $\{V_\alpha\}$ adapted to $V$, i.e., $V$ is union of strata. The following definition is an immediate extension for 1-forms of the corresponding (standard) definition for functions on stratified spaces in terms of its differential (c.f. \cite {EG2, GMP, Le1}). \vskip 0.3cm \noindent {\bf 2.1 Definition. } Let $\omega$ be a (real or complex) 1-form on $V$, i.e., a continuous section of either $T^*_\mathbbm{R} M \vert_V$ or $T^*M \vert_V$. A singularity of $\omega$ with respect to the Whitney stratification $\{V_\alpha\}$ means a point $x$ where the kernel of $\omega$ contains the tangent space of the corresponding stratum. This means that the pull back of the form to $V_\alpha$ vanishes at $x$. In section 1 we introduced the notion of radial forms, which is dual to the "radiality" for vector fields. We now extend this notion relaxing the condition of radiality in the directions tangent to the strata. From now on, unless it is otherwise stated explicitely, by a singularity of a 1-form on $V$ we mean a singularity in the stratified sense, i.e., in the sense of 2.1. \vskip 0.3cm \noindent {\bf 2.2 Definition. } Let $\omega$ be a (real or complex) 1-form on $V$. The form is {\bf normally radial} at a point $x_o \in V_\alpha \subset V$ if it is radial when restricted to vectors which are not tangent to the stratum $V_\alpha$ that contains $x_o$. In other words, for every vector $v(x)$ tangent to $M$ at a point $x \notin V_\alpha$, $x$ sufficiently close to $x_o$ and $v(x)$ pointing outwards a tubular neighborhood of the stratum $V_\alpha$, one has $Re \, \omega(v) > 0$ (or $Re \, \omega(v) < 0$ for all such vectors; if $\omega$ is real then it equals $Re \, \omega$). \vskip.2cm Obviously a radial 1-form is also normally radial, since it is radial in all directions. For each point $x$ in a stratum $V_\alpha$, one has a neighborhood $U_x$ of $x$ in $M$ which is diffeomorphic to the product $U_\alpha \times \mathbbm{D}_\alpha,$ where $U_\alpha = U_x \cap V_\alpha$ and $\mathbbm{D}_\alpha$ is a small disc in $M$ transversal to $V_\alpha$. Let $\pi$ be the projection $\pi : U_x \to U_\alpha$ and $p$ the projection $p: U_x \to \mathbbm{D}_\alpha$. One has an isomorphism: \[\;T^* U_x \,\cong \pi^* T^*{U_\alpha} \oplus p^* T^*\mathbbm{D}_\alpha\,.\] That a (real or complex) 1-form $\omega$ be normally radial at $x$ means that up to a local change of coordinates in $M$, $\omega$ is the direct sum of the pull back of a (real or complex) form on $U_\alpha$, i.e., a section of the (real or complex) cotangent bundle $T^*U_\alpha$, and a section of the (real or complex) cotangent bundle $T^*\mathbbm{D}_\alpha$ which is a radial form in the disc. \vskip.1cm It is possible to make for 1-forms the classical construction of {\bf radial extension} introduced by M.-H. Schwartz in \cite{Sch1, Sch2} for stratified vector fields and frames. Locally, the construction can be described as follows. We consider first real 1-forms. Let $\eta$ be a 1-form on $U_\alpha$, denote by $\widehat \eta$ its pull back to a section of $\pi^* T^*_{\mathbbm{R}}{U_\alpha}$. This corresponds to the {\bf parallel extension} of stratified vector fields done by Schwartz. Now look at the function $\rho$ given by the square of the distance to the origin in $\mathbbm{D}_\alpha$. The form $p^* d\rho$ on $U_x$ vanishes on $ U_\alpha $ and away from $U_\alpha $ its kernel is transversal to the strata of $V$ by Whitney conditions. The sum $\eta' = \widehat \eta + p^*d\rho$ defines a normally radial 1-form on $U_x$ which coincides with $\eta$ on $U_\alpha $; away from $U_\alpha $ its kernel is transversal to the strata of $V$. Thus, if $\eta$ is non-singular at $x$, then $\eta'$ is non-singular everywhere on $U_x$. If $\eta$ has an isolated singularity at $x \in V_\alpha$, then $\eta'$ also has an isolated singularity there. In particular, if the dimension of the stratum $V_\alpha$ is zero then $\eta'$ is a radial form in the sense of section 1. Following the terminology of \cite{Sch1, Sch2} we say that the form $\eta'$ is obtained from $\eta$ by {\bf radial extension}. Since the index in $M$ of a normally radial form is its index in the stratum times the index of a radial form in the disc $\mathbbm{D}_\alpha$, we obtain the following important property of forms constructed by radial extension. \vskip 0.3cm \noindent \proclaim{2.3 Proposition}{Let $\eta$ be a real 1-form on the stratum $V_\alpha$ with an isolated singularity at a point $x$ with local Poincar\'e-Hopf index ${\rm{Ind}}_{PH}(\eta, V_\alpha; x)$. Let $\eta'$ the 1-form on a neighborhood of $x$ in $M$ obtained by radial extension. Then the index of $\eta$ in the stratum equals the index of $\eta'$ in $M$: $${\rm{Ind}}_{PH}(\eta, V_\alpha; x)\,=\, {\rm{Ind}}_{PH}(\eta', M; x)\,. $$} \noindent{\bf 2.4 Definition.} The {\bf Schwartz index} of the continuous real 1-form $\eta$ at an isolated singularity $x \in V_\alpha \subset V$, denoted $\hbox{Ind}_{Sch}(\eta, V; x)$, is the Poincar\'e-Hopf index of the 1-form $\eta'$ obtained from $\eta$ by radial extension; or equivalently, if the stratum of $x$ has dimension more than 0, $\hbox{Ind}_{Sch}(\eta, V; x)$ is the Poincar\'e-Hopf index at $x$ of $\eta$ in the stratum $V_\alpha$. \vskip.2cm If $x$ is an isolated singularity of $V$ then every 1-form on $V$ must be singular at $x$ since its kernel contains the ``tangent space" of the stratum. In this case the index of the form in the stratum is defined to be 1, and this is consistent with the previous definition since in this case the radial extension of $\eta$ is actually radial at $x$, so it has index 1 in the ambient space. The previous process is easily adapted to give radial extension for complex 1-forms. Let $\omega$ be such a form on $V_\alpha$; let $\eta$ be its real part. We extend $\eta$ as above, by radial extension, to obtain a real 1-form $\eta'$ which is normally radial at $x$. Then we use statement i) in Theorem 1.6 above to obtain a complex 1-form $\omega'$ on $U_x $ that extends $\omega$ and is also normally radial at $x$. If we prefer, we can make this process in a different but equivalent way: first make a parallel extension of $\omega$ to $U_x $ as above, using the projection $\pi$; denote by $\widehat \omega$ this complex 1-form. Now use 1.6.i) to define a complex 1-form $\widehat { d\rho}$ on $ U_x $ whose real part is $d\rho$, and take the direct sum of $\widehat \omega$ and $\widehat {d\rho}$ at each point to obtain the extension $\omega'$. We say that $\omega'$ is obtained from $\omega$ by {\bf radial extension}. We have the equivalent of Proposition 2.3 for complex forms, modified with the appropriate signs: $$(-1)^s \, \hbox{Ind}_{PH}(\omega, V_\alpha; x)\,=\, (-1)^m \, \hbox{Ind}_{PH}(\omega', M; x)\,, $$ where $2s$ is the real dimension of $V_\alpha$ and $2m$ that of $M$. \vskip 0.3cm \noindent {\bf 2.5 Definition.} The {\bf Schwartz index} of the continuous complex 1-form $\omega$ at an isolated singularity $x \in V_\alpha \subset V$, denoted $\hbox{Ind}_{Sch}(\omega, V; x)$, is $(-1)^n$-times the index of its real part: $$\hbox{Ind}_{Sch}(\omega, V; x) \,=\, (-1)^n \hbox{Ind}_{Sch}(Re\, \omega, V; x)\;.$$ \section[Local Euler obstruction] {Local Euler obstruction and the Proportionality Theorem} We are now concerned only with a local situation, so we take $V$ to be embedded in an open ball $\mathbbm{B} \subset \mathcal C^m$ centered at the origin $0$. On the regular part of $V$ one has the map $\sigma : V_{reg} \to G_{n,m}$ into the Grassmannian of $n (= \text{dim}\,V)$-planes in $\mathcal C^m$, that assigns to each point the corresponding tangent space of $V_{reg}$. The Nash blow up $\widetilde V \buildrel {\nu}\over \to V$ of $V$ is by definition the closure in $\mathbbm{B} \times G_{n,m}$ of the graph of the map $\sigma$. One also has the Nash bundle $\widetilde T \buildrel {p}\over \to \widetilde V $, restriction to $\widetilde V $ of the tautological bundle over $\mathbbm{B} \times G_{n,m}$. The corresponding dual bundles of complex and real 1-forms are denoted by $\widetilde T^* \buildrel {p}\over \to \widetilde V $ and $\widetilde T^*_\mathbbm{R} \buildrel {p}\over \to \widetilde V $, respectively. Observe that a point in ${\widetilde T^*}$ is a triple $(x,P,\omega)$ where $x$ is in $V$, $P$ is an $n$-plane in the tangent space $T_x \mathbbm{B}$ which is limit of a sequence $\{(T V_{reg})_{x_i}\}$, where the $x_i$ are points in the regular part of $V$ converging to $x$, and $\omega$ is a $\mathcal C$-linear map $P \to \mathcal C$. (Similarly for $\widetilde T^*_\mathbbm{R}$.) Let us denote by $\rho$ the function given by the square of the distance to $0$. We recall that MacPherson in \cite {MP} observed that the Whitney condition (a) implies that the pull-back of the differential $d\rho$ defines a never-zero section $\widetilde {d\rho}$ of $\widetilde T^*_\mathbbm{R}$ over $\nu^{-1}(\mathbbm{S}_\varepsilon \cap V) \subset \widetilde V$, where $\mathbbm{S}_\varepsilon$ is the boundary of a small ball $\mathbbm{B}_\varepsilon$ in $\mathbbm{B}$ centered at $0$. The obstruction for extending $\widetilde {d\rho}$ as a never-zero section of $\widetilde T^*_\mathbbm{R}$ over $\nu^{-1}(\mathbbm{B}_\varepsilon \cap V) \subset \widetilde V$ is a cohomology class in $H^{2n}(\nu^{-1}(\mathbbm{B}_\varepsilon \cap V), \nu^{-1}(\mathbbm{S}_\varepsilon \cap V); \mathbbm{Z})$, and MacPherson defined {\bf the local Euler obstruction} $\hbox{Eu}_V(0)$ of $V$ at $0$ to be the integer obtained by evaluating this class on the orientation cycle $[\nu^{-1}(\mathbbm{B}_\varepsilon \cap V), \nu^{-1}(\mathbbm{S}_\varepsilon \cap V)]$. More generally, given a section $\eta$ of $T^*_{\Bbb R}\Bbb B|_A$, $A\subset V$, there is a canonical way of constructing a section $\tilde\eta$ of $\widetilde T^*_\mathbbm{R}|_{\tilde A}$, $\tilde A=\nu^{-1}A$, which is described in the following. The same construction works for complex forms. First, taking the pull-back $\nu^*\eta$, we get a section of $\nu^*T^*_{\Bbb R}\Bbb B|_V$. Then $\tilde\eta$ is obtained by projecting $\nu^*\eta$ to a section of $\tilde T^*_\mathbbm{R}$ by the canonical bundle homomorphism $$ \nu^*T^*_{\Bbb R}\Bbb B|_V\longrightarrow\tilde T^*_\mathbbm{R}. $$ Thus the value of $\tilde\eta$ at a point $(x,P)$ is simply the restriction of the linear map $\eta(x): (T_{\Bbb R}\Bbb B)_x\to\Bbb R$ to $P$. We call $\tilde\eta$ the {\bf canonical lifting} of $\eta$. By the Whitney condition (a), if $a \in V_\alpha$ is the limit point of the sequence $\{ x_i \}\in V_{\rm reg}$ such that $P = \lim (TV_{\rm reg})_{x_i} $ and if the kernel of $\eta$ is transversal to $V_\alpha$, then the linear form $\widetilde \eta$ will be non-vanishing on $P$. Thus, if $\eta$ has an isolated singularity at the point $0 \in V$ (in the stratified sense), then we have a never-zero section $\widetilde \eta$ of the dual Nash bundle $\widetilde T^*_\mathbbm{R}$ over $\nu^{-1}(\mathbbm{S}_\varepsilon \cap V) \subset \widetilde V$. Let $o(\eta) \in H^{2n}(\nu^{-1}(\mathbbm{B}_\varepsilon \cap V), \nu^{-1}(\mathbbm{S}_\varepsilon \cap V); \mathbbm{Z})$ be the cohomology class of the obstruction cycle to extend this to a section of $\widetilde T^*_\mathbbm{R}$ over $\nu^{-1}(\mathbbm{B}_\varepsilon \cap V)$. Then define (c.f. \cite {BMPS, EG2}): \vskip 0.3cm \noindent {\bf 3.1 Definition.} The {\bf local Euler obstruction} of the real differential form $\eta$ at an isolated singularity is the integer $\hbox{Eu}_{V}(\eta,0)$ obtained by evaluating the obstruction cohomology class $o(\eta)$ on the orientation cycle $[\nu^{-1}(\mathbbm{B}_\varepsilon \cap V), \nu^{-1}(\mathbbm{S}_\varepsilon \cap V)]$. \vskip 0.3cm \noindent The local Euler obstruction $\hbox{Eu}_{V}(0)$ of MacPherson corresponds to taking the differential of the squared function distance to $0$. In the complex case, one can perform the same construction, using the corresponding complex bundles. If $\omega$ is a complex differential form, section of $T^*\Bbb B|_A$ with an isolated singularity, one can define the local Euler obstruction $\hbox{Eu}_{V}(\omega,0)$. Notice that it is equal to that of its real part up to sign: \[\hbox{Eu}_{V}(\omega,0) \,=\, (-1)^n \hbox{Eu}_{V}(Re\, \omega, 0) \,. \tag{3.2}\] This is an immediate consequence of the relation between the Chern classes of a complex vector bundle and those of its dual (see for instance \cite {Mi}). We note that the idea to consider the (complex) dual Nash bundle was already present in \cite{Sa}, where Sabbah introduces a local Euler obstruction ${\rm E\check u}_V (0)$ that satisfies ${\rm E\check u}_V (0) = (-1)^{n}{\rm Eu}_V(0)$. See also Sch\"urmann \cite{Schu1}, sec. 5.2. \vskip.1cm Just as for vector fields (see \cite {BS}), one has in this situation the following: \vskip 0.3cm \noindent \proclaim{3.3 Theorem} {Let $ V_\alpha \subset V$ be the stratum containing $0$, ${\rm{Eu}}_V(0)$ the local Euler obstruction of $V$ at $0$ and $\omega$ a (real or complex) 1-form on $V_\alpha$ with an isolated singularity at $0$. Then the local Euler obstruction of the radial extension $\omega'$ of $\omega$ and the Schwartz index of $\omega$ at $0$ are related by the following proportionality formula: \[{\rm{Eu}}_{V}(\omega',0) \,=\, {\rm{Eu}}_{V} (0)\cdot {\rm{Ind}}_{Sch}(\omega, V; 0) \,.\]} \vskip 0.3cm \noindent {\bf Proof} By 3.2 and Theorem 1.6 above, it is enough to prove 3.3 for either real or complex 1-forms, each case implying the other. We prove it for real forms. Let $\eta$ and $\eta'$ be as above. Also, let $\eta_{rad}$ denote a radial form at $0$. By construction and definition, we have \[ \hbox{Ind}_{PH}(\eta,V_\alpha; 0)=\hbox{Ind}_{PH}(\eta',\Bbb B;0)=\hbox{Ind}_{Sch}(\eta,V;0).\tag{3.4} \] By definition of $\hbox{Ind}_{PH}(\eta',\Bbb B;0)$, there is a homotopy $$ \Psi:[0,1]\times\Bbb S_\varepsilon\longrightarrow T^*_\mathbbm{R}\Bbb B|_{\Bbb S_\varepsilon} $$ such that its image satisfies: \[ \partial\hbox{Im}\Psi=\eta'(\Bbb S_\varepsilon)-\hbox{Ind}_{PH}(\eta',\Bbb B;0) \cdot \eta_{rad}(\Bbb S_\varepsilon)\tag{3.5} \] as chains in $T^*_{\Bbb R}\Bbb B|_{\Bbb S_\varepsilon}$. The restriction of $\Psi$ gives a homotopy $$ \psi:[0,1]\times(\Bbb S_\varepsilon\cap V)\longrightarrow T^*_{\Bbb R}\Bbb B|_{\Bbb S_\varepsilon\cap V} $$ such that (c.f. (3.4)) $$ \partial\hbox{Im}\psi=\eta'(\Bbb S_\varepsilon\cap V)-\hbox{Ind}_{Sch}(\eta,V;0) \cdot\eta_{rad}(\Bbb S_\varepsilon\cap V). $$ Now we can lift $\psi$, $\eta'$ and $\eta_{rad}$ to sections $\nu^*\psi$, $\nu^*\eta'$ and $\nu^*\eta_{rad}$ of $\nu^*T^*_{\Bbb R}\Bbb B$ to get a homotopy $$ \nu^*\psi:[0,1]\times\nu^{-1}(\Bbb S_\varepsilon\cap V) \longrightarrow\nu^*T^*_{\Bbb R}\Bbb B|_{\nu^{-1}(\Bbb S_\varepsilon\cap V)} $$ and, since $\nu$ is an isomorphism away from the singularity of $V$, we still have \[ \partial\hbox{Im}\nu^*\psi=\nu^*\eta'(\nu^{-1}(\Bbb S_\varepsilon\cap V))-\hbox{Ind}_{Sch}(\eta,V;0) \cdot\nu^*\eta_{rad}(\nu^{-1}(\Bbb S_\varepsilon\cap V)) \] as chains in $\nu^*T^*_{\Bbb R}\Bbb B|_{\nu^{-1}(\Bbb S_\varepsilon\cap V)}$. Recall that we get the canonical liftings $\tilde\psi$, $\tilde\eta'$ and $\tilde\eta_{rad}$ of $\psi$, $\eta'$ and $\eta_{rad}$ by taking the images of $\nu^*\psi$, $\nu^*\eta'$ and $\nu^*\eta_{rad}$ by the canonical bundle homomorphism $\nu^*T^*_{\Bbb R}\Bbb B \longrightarrow \tilde T^*_{\Bbb R}$. Thus we have \[ \partial\hbox{Iml}\tilde\psi=\tilde\eta'(\nu^{-1}(\Bbb S_\varepsilon\cap V))-\hbox{Ind}_{Sch}(\eta,V;0) \cdot\tilde\eta_{rad}(\nu^{-1}(\Bbb S_\varepsilon\cap V)) \] as chains in $\tilde T^*_{\Bbb R}|_{\nu^{-1}(\Bbb S_\varepsilon\cap V)}$. The forms $\tilde\eta'$ and $\tilde\eta_{rad}$ are non-vanishing on $\nu^{-1}(\Bbb S_\varepsilon\cap V)$, by the Whitney condition, and by definition of the Euler obstructions, we have the theorem. \ensuremath{\Box} \section{The GSV-index} We recall (\cite {GSV, SS}) that the GSV-index of a vector field $v$ on an isolated complete intersection germ $V$ can be defined to be the Poincar\'e-Hopf index of an extension of $v$ to a Milnor fiber $F$. Similarly, the GSV-index of a 1-form $\omega$ on $V$ can be defined to be the Poincar\'e-Hopf index of the form on $F$, i.e., the number of singularities of $\omega$ in $F$ counted with multiplicities \cite {EG1}. When $V$ has non-isolated singularities one may not have a Milnor fibration in general, but one does if $V$ has a Whitney stratification with Thom's $a_f$-condition, $f =(f_1,\cdots,f_k)$ being the functions that define $V$ (c.f. \cite {Le2, LT, BSS1}). Let $(V,0)$ be a complete intersection of complex dimension $n$ defined in a ball $\mathbbm{B}$ in $\mathcal C^{n+k}$ by functions $f =(f_1,\cdots,f_k)$, and assume $0$ is a singular point of $V$ (not necessarily an isolated singularity). As before, we endow $\mathbbm{B}$ with a Whitney stratification adapted to $V$, and we assume that we can choose $\{V_\alpha\}$ so that it satisfies the $a_f$-condition of Thom (see for instance \cite{LT}). In particular one always has such a stratification if $k = 1$, by \cite{Hi}. Let $\omega$ be as before, a (real or complex) 1-form on $\mathbbm{B}$, and assume its restriction to $V$ has an isolated singularity at $0$. This means that the kernel of $\omega(0)$ contains the tangent space of the stratum $V_\alpha$ containing $0$, but everywhere else it is transversal to each stratum $V_\alpha \subset V$. Now let $F = F_t$ be a Milnor fiber of $V$, i.e., $F = f^{-1}(t) \cap \mathbbm{B}_\varepsilon$, where $\mathbbm{B}_\varepsilon$ is a sufficiently small ball in $\mathbbm{B}$ around $0$ and $t \in \mathcal C^k$ is a regular value of $f$ with $\Vert t \Vert$ sufficiently small with respect to $\varepsilon$. Notice that the $a_f$-condition implies that for every sequence $t_n$ of regular values converging to $0$, and for every sequence $\{x_n\}$ of points in the corresponding Milnor fibers converging to a point $x_o \in V$ so that the sequence of tangent spaces $\{(TF)_{x_n}\}$ has a limit $T$, one has that $T$ contains the space $(TV_\alpha)_{x_o}$, tangent to the stratum that contains $x_o$. By transversality this implies that choosing the regular value $t$ sufficiently close to $0$ we can assure that the kernel of $\omega$ is transversal to the Milnor fiber at every point in its boundary $\partial F$. Thus its pull-back to $F$ is a 1-form on this smooth manifold, and it is never-zero on its boundary, thus $\omega$ has a well defined Poincar\'e-Hopf index in $F$ as in section 1. This index is well-defined and depends only on the restriction of $\omega$ to $V$ and the topology of the Milnor fiber $F$, which is well-defined once we fix the defining function $f$ (which is assumed to satisfy the $a_f$-condition for some Whitney stratification). \vskip 0.3cm \noindent {\bf 4.1 Definition} The GSV-index of $\omega$ at $0 \in V$ relative to $f$, $\hbox{Ind}_{GSV}(\omega,0)$, is the Poincar\'e-Hopf index of $\omega$ in $F$. \vskip.2cm In other words this index measures the number of points (counted with signs) in which a generic perturbation of $\omega$ is tangent to $F$. In fact the inclusion $F \buildrel{i}\over \to M$ pulls the form $\omega$ to a section of the (real or complex, as the case may be) cotangent bundle of $F$, which is never-zero near the boundary because $\omega$ has an isolated singularity at $0$ and, by hypothesis, the map $f$ satisfies the $a_f$-condition of Thom. If the form $\omega$ is real then \[ \hbox{Ind}_{GSV}(\omega,0) \,=\, \hbox{Eu}(F; \omega)[F] \;, \tag{4.2}\] where $\hbox{Eu}(F; \omega) \in H^{2n}(F, \partial F)$ is the Euler class of the real cotangent bundle $T^*_\mathbbm{R} F$ relative to the section defined by $\omega$ on the boundary, and $[F]$ is the orientation cycle of the pair $(F,\partial F)$. If $\omega$ is a complex form, then one has: \[ \hbox{Ind}_{GSV}(\omega,0) \,=\, c^n(T^*F; \omega)[F] \;, \tag{4.3.i}\] where $c^n(T^*F; \omega)$ is the top Chern class of the cotangent bundle of $F$ relative to the form $\omega$ on its boundary. In this case one can, alternatively, express this index as the relative Chern class: \[ \hbox{Ind}_{GSV}(\omega,0) \,=\, c^n(T^*M \vert_F; \Omega)[F] \;, \tag{4.3.ii}\] where $\Omega$ is the frame of $k+1$ complex 1-forms on the boundary of $F$ given by \[\Omega \,=\, (\omega, df_1, df_2, \cdots, df_k)\,,\] since the forms $(df_1,\cdots,df_k)$ are linearly independent everywhere on $F$. Notice that if the form $\omega$ is holomorphic, then this index is necessarily non-negative because it can be regarded as an intersection number of complex submanifolds. For every complex 1-form one has: \[ \hbox{Ind}_{GSV}(\omega,0) \,=\, (-1)^n \hbox{Ind}_{GSV}(Re\, \omega,0) \,.\] We remark that if $V$ has an isolated singularity at $0$, this is the index envisaged in \cite{EG1}, i.e., the degree of the map from the link $K$ of $V$ into the Stiefel manifold of complex (k+1)-frames in the dual $\mathcal C^{n+k}$ given by the map $(\omega, df_1,\cdots,df_k)$. Also notice that this index is somehow dual to the index defined in \cite {BSS1} for vector fields, which is related to the top Fulton-Johnson class of singular hypersurfaces. So, given the (non-isolated) complete intersection singularity $(V,0)$ and a (real or complex) 1-form $\omega$ on $V$ with an isolated singularity at $0$, one has three different indices: the Euler obstruction (section 2), the GSV-index just defined and the index of its pull back to a 1-form on the stratum of $0$. One also has the index of the form in the ambient manifold $M$. For forms obtained by radial extension, the index in the stratum equals its index in $M$, and this is by definition the Schwartz index. The following proportionality theorem is analogous to the one in \cite {BSS1} for vector fields. \vskip 0.3cm \noindent \proclaim{4.4 Theorem} {Let $\omega$ be a (real or complex) 1-form on the stratum $V_\alpha$ of $0$ with an isolated singularity at $0$. Then the GSV index of its radial extension $\omega'$ is proportional to the Schwartz index, the proportionality factor being the Euler-Poincar\'e characteristic of the Milnor fiber $F$: \[{\rm{Ind}}_{GSV}(\omega',0)\,=\, {\chi(F) \cdot \rm{Ind}}_{Sch}(\omega, V; 0) \,.\] } \vskip.1cm \noindent {\bf Proof.} It is enough to prove 4.4 either for complex forms or for real forms, each one implying the other. The proof is similar to that of 3.3. Let $\omega'$ and $\omega_{rad}$ be as in the proof of Theorem 3.3. Then 4.4 is proved by taking the retriction to $F$ of each section in (3.5) as a differential form, noting that $\hbox{Ind}_{GSV}(\omega_{rad},0)\,=\, \chi(F)$. \ensuremath{\Box} \vskip 0.3cm \noindent {\bf 4.5 Remark.} We notice that 4.2 and 3.3 can also be proved using the stability of the index under perturbations; this works for vector fields too. More precisely, one can easily show that the Euler obstruction $\hbox{Eu}_V(\omega,x)$ and the GSV-index are stable when we perturb the 1-form (or the vector field) in the stratum and then extend it radially; then the sum of the indices at the singularities of the new 1-form (vector field) give the corresponding index for the original singularity. This implies the proportionality of the indices.
{ "timestamp": "2005-05-11T10:07:43", "yymm": "0503", "arxiv_id": "math/0503428", "language": "en", "url": "https://arxiv.org/abs/math/0503428" }
\section{Introduction} This paper is an attempt to understand topological properties of Lie group actions. Its starting point was the following theorem of McDuff--Slimowitz~\cite{MSlim} concerning circle subgroups of ${\rm Symp}(M,\omega)$, the group of symplectomorphisms of a symplectic manifold $(M,{\omega})$. Recall that an effective circle action is {\bf semifree} if the stabilizer subgroup of each point in $M$ is either the circle itself or the trivial group. Also, we say that a circle subgroup ${\Lambda}$ of a topological group ${\mathcal H}$ is {\bf essential in} ${\boldsymbol {\mathcal H}}$ if it represents a nonzero element in $\pi_1({\mathcal H})$ and {\bf inessential in} ${\boldsymbol {\mathcal H}}$ otherwise. \begin{theorem}\labell{thm:sfr} Any semifree circle action on a closed symplectic manifold $(M,\omega)$ is essential in ${\rm Symp}(M,{\omega})$. \end{theorem} This is obvious if the action is not Hamiltonian since in this case the flux homomorphism $$ {\rm Flux}: \pi_1({\rm Symp}(M,{\omega})) \longrightarrow H^1(M,{\mathbb R}) $$ does not vanish on ${\Lambda}$. However, if the action is Hamiltonian with generating Hamiltonian $K:M\longrightarrow {\mathbb R}$ then the result is not so easy: the proof in~\cite{MSlim} involved studying the Hofer length of the corresponding paths $\phi_t^K, t\in [0,T],$ in ${\rm Ham}(M,{\omega})$. The first result in this paper uses the theorem above to answer a question posed by Alan Weinstein in~\cite{Wei}. Let $G$ be a semisimple Lie group with Lie algebra $\fg$. Let $M \subset \fg^*$ be a coadjoint orbit, together with the Kostant--Kirillov symplectic form $\omega$. If the coadjoint action of $G$ on $M$ is effective, then $G$ is naturally a subgroup of ${\rm Ham}(M,\omega)$, the group of Hamiltonian symplectomorphisms of $(M,\omega)$. This inclusion induces a natural map from the fundamental group of $G$ to the fundamental group of ${\rm Ham}(M,\omega)$. Weinstein asks when this map is injective. We prove that this map is injective for all compact semisimple Lie groups. In~\cite{Vina} Vina established a special case of this result by quite different methods. \begin{theorem}\label{cor:coadj} Let a compact semisimple Lie group $G$ act effectively on a coadjoint orbit $(M,\omega)$. Then the inclusion $G\longrightarrow {\rm Ham}(M, {\omega})$ induces an injection from $\pi_1(G)$ to $\pi_1({\rm Ham}(M,{\omega})).$ \end{theorem} In view of Theorem~\ref{thm:sfr}, this is an immediate consequence of the following result, which we prove in Section \ref{ss:coadj}. \begin{proposition}\label{WQ} Let a compact semisimple Lie group $G$ act effectively on a coadjoint orbit $(M,\omega)$. Then every nontrivial element in $\pi_1(G)$ may be represented by a circle that acts semifreely on $M$. \end{proposition} Theorem~\ref{thm:sfr} immediately implies that if a compact Lie group $G$ acts effectively on a closed symplectic manifold $(M,\omega)$, then any semifree circle subgroup $\Lambda \subset G$ is essential in $G$. The other results in the paper generalize this claim. Observe that Theorem~\ref{thm:sfr} does not immediately extend to the smooth (non-symplectic) category. For example, Claude LeBrun pointed out to us that the circle action on $S^4$ induced by the diagonal action of $S^1$ on ${\mathbb C}^2 = {\mathbb R}^4 \subset {\mathbb R}^5$ is semifree but gives a nullhomotopic loop since $\pi_1(SO(5)) = {\mathbb Z}/2{\mathbb Z}$. Nevertheless, the semifree condition does have consequences in the smooth category, even if the action is only semifree on a neighborhood of a component of the fixed point set; we shall say that such components are {\bf semifree}. Further, given a circle subgroup $\Lambda \subset G$ we say that $g \in G$ {\bf reverses} ${\boldsymbol \Lambda}$ {\bf in} ${\boldsymbol G}$ if $g t g^{-1} = t^{-1}$ for all $t\in {\Lambda}$. Finally, a component $F$ of the fixed point set $M^{\Lambda}$ of $\Lambda$ is {\bf symmetric in} ${\boldsymbol G}$ if there is an element $g \in G$ whose action on $M$ fixes $F$ pointwise and which reverses $\Lambda$. \begin{thm}\label{thm:main0} Let $\Lambda$ be a circle subgroup of a compact Lie group $G$ which acts effectively on a connected manifold $M$. If there is a semifree component of the fixed point set $M^\Lambda$ which is not symmetric in $G$, then ${\Lambda}$ is essential in $G$. \end{thm} \begin{example}\label{ex}\rm First, let $G = SU(2)$ act on ${\mathbb{CP }}^2$ by the defining representation on the first two copies of ${\mathbb C}$, and let $\Lambda \subset G$ be the circle subgroup given by $\lambda \cdot [z_0:z_1: z_2]\mapsto [\lambda z_0: \lambda^{-1} z_1: z_2]$. This action has a semifree fixed point, namely $[0,0,1]$. Moreover, this circle subgroup is inessential in $G$. Therefore, by the theorem above, there exists $g \in G$ which reverses the circle action and fixes $[0,0,1]$. In fact, we can take any $g$ which lies in the normalizer $N(\Lambda)$ but not in $\Lambda$ itself. Note that $g^2 = -{{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l$ for any such $g$. In contrast, consider the natural action of $G= PU(3)$ on ${\mathbb{CP }}^2$, and let $\Lambda \subset G$ be the circle subgroup given by $\lambda \cdot [z_0:z_1: z_2]\mapsto [\lambda^2 z_0: z_1: z_2]$. This action is semifree and essential, but is not reversed by any $g \in G$. To see this, note that the circle has order $3$ in $\pi_1(G)$, whereas every circle that can be reversed has order $1$ or $2$. \end{example} One can weaken the semifree hypothesis in the above theorem, at the cost of adding a global isotropy assumption and working once more in the symplectic category. We say that a circle action has {\bf at most twofold isotropy} if every point which is not either fixed or free has stabilizer ${\mathbb Z}/(2).$ Recall, also, that a symplectic action of $G$ on $(M,{\omega})$ is Hamiltonian if it is given by an equivariant moment map $\Phi:M\longrightarrow \fg^*$. \begin{thm} \labell{thmtwo} Let $\Lambda$ be a circle subgroup of a compact Lie group $G$ which acts effectively on a connected symplectic manifold $(M,\omega)$. If $\Lambda$ has at most twofold isotropy and if there is no $g \in G$ which reverses $\Lambda$, then ${\Lambda}$ is essential in $G$. \end{thm} \begin{example}\rm This theorem does not extend to circle actions which have at most threefold isotropy. For example, the action of $S^1$ on ${\mathbb{CP }}^3$ given by $\lambda \cdot [x,y,z,w] = [\lambda^2 x, \lambda^{-1}y, \lambda^{-1}z,w]$ is inessential in $PU(4)$. However, since $F_{\rm max}$ and $F_{\rm min}$ are not diffeomorphic, this action has no reversor. We also need the symplectic hypothesis. To see this, consider the obvious action of $SU(3)$ on $S^6: = {\mathbb C}^3\cup\{\infty\}$. The subgroup ${\Lambda}: = {\rm diag\,}(\lambda^2, \lambda^{-1},\lambda^{-1})$ acts with at most twofold isotropy but has no reversor. \end{example} \begin{remark} \rm If $G$ is a simple group, we do not need to assume that $(M,\omega)$ is symplectic in Theorem~\ref{thmtwo}; we only need to assume that there exists a point $p$ which is fixed by a maximal torus containing $\Lambda$ but is not fixed by all of $G$. Note that, in contrast, in the example above, the only points fixed by ${\Lambda}$ are fixed by all of $G$. \end{remark} \begin{remark}\rm If $\Lambda$ is {\em any} circle subgroup of $SO(3)$ -- or indeed a subgroup of any simple group of type $B_n$, $C_n$, or $F_4$ -- then there exists a $g \in G$ which reverses $\Lambda$. In this case, Theorem~\ref{thmtwo} is trivial and the force of Theorem~\ref{thm:main0} is that we can choose $g$ so that it also fixes $p$. \end{remark} \begin{remark}\rm In the proof of the above theorems, we pick a maximal torus $T$ which contains $\Lambda$. The reversor $g$ that we construct lies in the normalizer $N(T)$ and has the property that $g^2$ lies in $T$. However, as we saw in Example~\ref{ex}, $g^2$ may not be equal to the identity. \end{remark} Theorems \ref{thm:main0} and \ref{thmtwo} have the following easy corollaries: \begin{corollary} Consider a Hamiltonian circle action $\Lambda$ on a closed symplectic manifold $(M,\omega)$ with moment map $K: M \longrightarrow {\mathbb R}$, normalized so that $\int_M K \omega^n = 0$. If $F$ is a semifree fixed component, then $\Lambda$ is essential in every compact subgroup $G \subset {\rm Symp}(M,\omega)$ that contains it, unless there is a symplectomorphism $g$ of $M$ that fixes $F$ and reverses ${\Lambda}$. In this case, all the following hold: \begin{enumerate} \item $K(g(p)) = - K(p)$ for all $p \in M^{{\Lambda}}$. \item There is a one-to-one correspondence between the positive weights at $p$ and the negative weights at $g(p)$, and vice versa. \item $g$ induces an isomorphism on the image of the restriction map in equivariant cohomology $H^*_{S^1}(M) \longrightarrow H^*_{S^1}(M^{\Lambda})$. \item $g(F) = F$. In particular, \begin{enumerate} \item $K(F) = 0$. \item The sum of the weights at $F$ is zero. \end{enumerate} \end{enumerate} \end{corollary} \begin{corollary} Consider a Hamiltonian circle action $\Lambda$ on a closed symplectic manifold $(M,\omega)$ with moment map $K: M \longrightarrow {\mathbb R}$, normalized so that $\int_M K \omega^n = 0$. If the action has at most twofold isotropy, then $\Lambda$ is essential in every compact subgroup $G \subset {\rm Symp}(M,\omega)$ that contains it, unless there is a symplectomorphism $g$ of $M$ that reverses ${\Lambda}$. In this case, all the following hold: \begin{enumerate} \item $K(g(p)) = - K(p)$ for all $p \in M^{{\Lambda}}$. \item There is a one-to-one correspondence between the positive weights at $p$ and the negative weights at $g(p)$, and vice versa. \item $g$ induces an isomorphism on the image of the restriction map in equivariant cohomology $H^*_{S^1}(M) \longrightarrow H^*_{S^1}(M^{\Lambda})$. \end{enumerate} \end{corollary} It is unknown whether the existence of such $g$ is necessary for ${\Lambda}$ to be inessential in ${\rm Symp}(M,{\omega})$. We make partial progress towards answering this question in \cite{MT}. All the results in this paper are proved by a case by case study of the structure of semisimple Lie algebras. \section{Coadjoint orbits}\label{ss:coadj} In this section, we prove Proposition~\ref{WQ}. We begin with a brief review of a few facts about Lie groups. Each simply connected compact semisimple Lie group is a product of simple factors, and its center is the product of the centers of its simple factors. Moreover, since its Lie algebra splits into a corresponding sum, the coadjoint orbits also are products of coadjoint orbits of simple groups. Therefore, we may assume that $G$ is simple. Let $G$ be a compact simple Lie group. Let $\widetilde{G}$ denote the universal cover of $G$, and ${\widehat{G}}$ denote the quotient of $G$ by its center. Let $\fg$ denote the Lie algebra of $G$, and let $\ft$ denote the Lie algebra of a maximal torus $T \subset G$. Let $\ell \subset \ft$, $\Tilde{\ell} \subset \ft$, and $\widehat{\ell} \subset \ft$ be the lattices consisting of vectors $\xi \in \ft$ whose exponential is the identity in $G$, $\widetilde{G}$, and ${\widehat{G}}$, respectively. There is a one-to-one correspondence between $\ell$ and circle subgroup of $G$, $\Tilde{\ell}$ and circle subgroups of $\widetilde{G}$, and $\widehat{\ell}$ and circle subgroups of ${\widehat{G}}$, given by sending $\lambda$ to $t \rightarrow \exp(t\lambda)$. Note that $\Tilde{\ell} \subseteq \ell \subseteq \widehat{\ell}$. Because $\widetilde{G}$ is simply connected, $$ \pi_1(G) \;\cong \;\ell/\Tilde{\ell} \;\subseteq \;\widehat{\ell}/\Tilde{\ell} \;\cong \; \pi_1({\widehat{G}}). $$ Let $\ft^*$ denote the dual to $\ft$, and let $\Delta \subset \ft^*$ denote the set of {\bf roots} of $G$, i.e. the nonzero weights of the adjoint action $T$ on $\fg_{\mathbb C}$, where $\fg_{\mathbb C}$ is the complexification of $\fg$. The lattice $\widehat{\ell}$ is dual to the lattice in $\ft^*$ generated by the roots, i.e. ${\lambda}\in \widehat{\ell}$ precisely when $\eta({\lambda})\in {\mathbb Z}$ for all $\eta\in \Delta$. If we use the Killing form $(\cdot, \cdot)$ to identify $\ft$ and $\ft^*$, then $\Tilde{\ell}$ is generated by the set $$ \left\{ \left. \frac{2 \eta}{(\eta,\eta)}\ \right| \eta \in \Delta \right\}. $$ Further the set of weights at any fixed point $p$ for the action of $T$ on $M$ is a nonempty subset of the set of roots. Therefore the result will follow if we find a representative $\lambda$ for each nontrivial class in $\widehat{\ell}/\Tilde{\ell}$ such that $|\eta(\lambda)| \leq 1$ for every $\eta \in \Delta$. We will check this on a case by case basis; in each case we will use the Killing form to identify $\ft$ and $\ft^*$. Let $( \cdot, \cdot)$ be the standard metric on ${\mathbb R}^k$ with the standard basis $e_1,\dots,e_k$, and define $$ \epsilon_i = e_i - \frac{1}{k} \sum_{j=1}^k e_j. $$ {\medskip} {\noindent}{\bf (I)}\,\, For the group $A_n$, where $n \geq 1$, $\ft = \ft^* = \left\{ \lambda \in {\mathbb R}^{n+1} \left| \ \sum \lambda_i = 0 \right. \right\}$ and the roots are $\epsilon_i - \epsilon_j = e_i - e_j$ for $i \neq j$. Hence $$ \widehat{\ell} = \{ \lambda \in \ft \mid \lambda_i - \lambda_j \in {\mathbb Z} \ \forall \ i,j \}, \ \mbox{and} \quad \Tilde{\ell} = \ft \cap {\mathbb Z}^{n+1}. $$ As representatives for the quotient $\widehat{\ell}/\Tilde{\ell} \cong {\mathbb Z}/(n+1)$, we take $\lambda = \sum_{i=1}^k \epsilon_i$ for $0 \leq k \leq n$. {\medskip} {\noindent}{\bf (II)}\,\, For the group $B_n$, where $n \geq 2$, $\ft^* = {\mathbb R}^n$ and the roots are $\pm e_i$ and $\pm e_i \pm e_j$ for $i \neq j$. Hence $$ \widehat{\ell} = {\mathbb Z}^n,\ \mbox{and} \qquad \Tilde{\ell} = \left\{ \lambda \in {\mathbb Z}^n \left| \ \sum \lambda_i \in 2{\mathbb Z} \right. \right\}. $$ As representatives of the quotient $\widehat{\ell}/\Tilde{\ell} \cong {\mathbb Z}/(2) $, we take $0$ and $e_1$. {\medskip} {\noindent}{\bf (III)}\,\, For the group $C_n$, where $n \geq 3$, $\ft^* = {\mathbb R}^n$ and the roots are $\pm 2 e_i$ and $\pm e_i \pm e_j$ for $i \neq j$. Hence $$ \widehat{\ell} = \{ \lambda \in {\mathbb R}^n \mid \lambda_i \pm \lambda_j \in {\mathbb Z}, \ \forall \ i,j \}, \ \mbox{and} \qquad \Tilde{\ell} = {\mathbb Z}^n. $$ As representatives of the quotient $\widehat{\ell}/\Tilde{\ell} \cong {\mathbb Z}/(2)$ , we take $0$ and $\frac{1}{2} \sum_{i=1}^n e_i$. {\medskip} {\noindent}{\bf (IV)}\,\, For the group $D_n$, where $n \geq 4$, $\ft^* = {\mathbb R}^n$ and the roots are $\pm e_i \pm e_j$ for $i \neq j$. Hence $$ \widehat{\ell} = \{ \lambda \in {\mathbb R}^n \mid \lambda_i \pm \lambda_j \in {\mathbb Z}, \ \forall \ i,j \}, \ \mbox{and} \qquad \Tilde{\ell} = \left\{ \lambda \in {\mathbb Z}^n \left| \ \sum \lambda_i \in 2{\mathbb Z} \right. \right\}. $$ The quotient $\widehat{\ell}/\Tilde{\ell}$ is isomorphic to ${\mathbb Z}/(2) \oplus {\mathbb Z}/(2)$ if $n$ is even, and to ${\mathbb Z}/(4)$ if $n$ is odd. Either way, as representatives of $\widehat{\ell}/\Tilde{\ell}$, we take $0$, $e_1$, $\frac{1}{2}\sum_{i=1}^n e_i$ and $\frac{1}{2}\sum_{i=1}^n e_i - e_n$. {\medskip} {\noindent}{\bf (V, a)}\,\, For the group $E_6$, $\ft^* = {\mathbb R}^6$ and the roots are $2 \epsilon$, $\epsilon_i - \epsilon_j$, and $\epsilon_i + \epsilon_j + \epsilon_k \pm \epsilon$ for $i, j,$ and $k$ distinct, where $\epsilon = \frac{1}{2 \sqrt{3}}(1,1,1,1,1,1)$. Hence, $$ \widehat{\ell} = \Big\{ n \epsilon + (\xi_1,\ldots,\xi_6) \in {\mathbb R}^6 \Big| \sum_{i=1}^6 \xi_i = 0, n \in {\mathbb Z}, \ \frac{n}{2} + 3 \xi_i \in {\mathbb Z} \ \mbox{and} \ \xi_i - \xi_j \in {\mathbb Z} \ \forall \ i,j \Big\}, \ \mbox{and} $$ $$ \Tilde{\ell} = \Big\{ n \epsilon + (\xi_1,\ldots,\xi_6) \in {\mathbb R}^6 \Big| \sum_{i=1}^6 \xi_i = 0, n \in {\mathbb Z} \ \mbox{and} \ \frac{n}{2} + \xi_i \in {\mathbb Z} \ \forall \ i \Big\}. $$ As representatives of the quotient $\widehat{\ell}/\Tilde{\ell} \cong {\mathbb Z}/(3)$ , we take $0$, $\epsilon_1 + \epsilon_2$, and $-\epsilon_1 - \epsilon_2$. {\medskip} {\noindent}{\bf (V, b)}\,\, For the group $E_7$, $\ft = \ft^* = \left\{ \lambda \in {\mathbb R}^{8} \mid \sum \lambda_i = 0 \right\}$, and the roots are $\epsilon_i - \epsilon_j$, and $\epsilon_i + \epsilon_j + \epsilon_k + \epsilon_l$ for $i, j, k,$ and $l$ distinct. Hence $$\widehat{\ell} = \{ \lambda \in \ft \mid 4 \lambda_i \in {\mathbb Z} \ \mbox{and} \ \lambda_i - \lambda_j \in {\mathbb Z}\ \forall\ i, j \}, \ \mbox{and} \quad \Tilde{\ell} = \{ \lambda \in \ft \mid \lambda_i \pm \lambda_j \in {\mathbb Z}\ \forall \ i, j \}. $$ As representatives for the quotient $\widehat{\ell}/\Tilde{\ell} \cong {\mathbb Z}/2{\mathbb Z}$, we take $0$ and $\epsilon_1 + \epsilon_2$.{\medskip} Every group of type $E_8$, $F_4$, and $G_2$ is simply connected, so no further argument is necessary. \hfill$\Box$\medskip \section{Lie Group Actions}\label{sec:lie} This section contains proofs of Theorems \ref{thm:main0} and \ref{thmtwo}. We begin by stating a lemma about root systems, that is proved at the end. We shall always assume that the positive Weyl chamber is closed. \begin{lemma}\labell{claims} Let $G$ be a simply connected compact simple Lie group. Let $\ft$ be the Lie algebra of a maximal torus $T \subset G$. Let $\Tilde{\ell}$ be the integral lattice, let $\Delta$ denote the set of roots, and let $W$ denote the Weyl group. Use the Killing form to identify $\ft$ and $\ft^*$. Fix $\lambda \in \Tilde{\ell}$. Choose a positive Weyl chamber which contains $\lambda$. Let $\delta \in \Delta$ denote the highest root. Then the following claims hold: \smallskip {\noindent}{\rm (a)} If $(\lambda,\delta) \leq 2$, then there exist orthogonal roots $\eta_1,\ldots,\eta_k \in \Delta$ so that $\lambda = \sum a_i \eta_i$ and so that $(\lambda,\eta_i) = a_i (\eta_i,\eta_i) = 2$ for all $i$. \smallskip {\noindent}{\rm (b)} Let $L \subset \Delta$ be a set of roots which contains every root $\eta \in \Delta$ such that $\delta + \eta$ or $\delta - \eta$ is also a root. Assume also that $L$ is closed under addition, that is, it contains every root which can be written as the sum of roots in $L$. Then $L$ contains all roots. \smallskip {\noindent}{\rm (c)} If $(\lambda,\delta) > 2$ and $-{\rm id}: \ft \longrightarrow \ft$ is not an element of the Weyl group, then for every nonzero weight $\alpha \in \Tilde{\ell}^*$ there exists $\sigma \in W$ so that $|(\sigma \cdot \alpha,\lambda)| > 1$. \smallskip {\noindent}{\rm (d)} If $-{\rm id}: \ft \longrightarrow \ft$ is not an element of the Weyl group, then $\delta$ is the only root which lies in the positive Weyl chamber. \end{lemma} Using this result, we can find find elements which reverse certain circle subgroups of simply connected compact simple Lie groups. Note that because $G$ is simply connected, every circle subgroup of $G$ is inessential in $G$. \begin{lemma} \labell{le:simple} Let $\Lambda $ be a circle subgroup of a simply connected compact simple Lie group $G$. \smallskip {\noindent}{\rm (i)} Let $\rho : G \longrightarrow {\rm GL}(V)$ be a nontrivial representation of $G$. If $\Lambda$ acts semifreely on $V$ then there exists $g \in G$ that reverses $\Lambda$. \smallskip {\noindent}{\rm (ii)} Let $H \varsubsetneq G$ be a proper subgroup containing $\Lambda$. If the adjoint action of $\Lambda$ on $\fg/{\mathfrak h}$ is semifree, then there exists $h \in H$ that reverses $\Lambda$. \smallskip {\noindent}{\rm (iii)} Let $H \varsubsetneq G$ be a proper subgroup containing a maximal torus which contains $\Lambda$. If the natural action of $\Lambda$ on $G/H$ has at most twofold isotropy, then there exists $g \in G$ that reverses $\Lambda$. \end{lemma} The assumption in (ii) above is a special case of (i) since the representation $V$ is restricted; however, the conclusion is stronger since it asserts that the reversor lies in $H$. Statement (ii) and (iii) are also related: the former makes a strong assumption about the action induced by ${\Lambda}$ on the tangent space to $G/H$ at the fixed point $eH$, the latter makes a weaker assumption about the action at all the fixed points on $G/H$. We will now use the claims in Lemma \ref{claims} to prove Lemma \ref{le:simple}. Let $T$ be a maximal torus which contains $\Lambda$. Let $\Tilde{\ell} \subset \ft$ denote the integral lattice. Let ${\lambda} \in\Tilde{\ell}$ be the vector corresponding to $\Lambda$. Choose a positive Weyl chamber which contains $\lambda$. Let $\delta \in \Delta$ denote the highest root. Recall that the Weyl group $W$ is the quotient $N(T)/T$, where $N(T)$ is the the normalizer of $T$ in $G$. Every root $\eta$ gives rise to an element $w_\eta \in W$ whose action on $\ft^*$ is given by $w_\eta(\beta) = \beta - \frac{2(\eta,\beta)}{(\eta,\eta)} \eta$. {\medskip} {\noindent} {\bf Proof of Lemma~\ref{le:simple} (i).} Let $\rho : G \longrightarrow {\rm GL}(V)$ be a nontrivial representation of $G$. Assume that $\Lambda$ acts semifreely on $V$. Suppose first that $(\lambda,{\delta}) \leq 2$. By claim (a), there exist orthogonal roots $\eta_1,\ldots,\eta_k \in \Delta$ so that $\lambda = \sum a_i \eta_i$. Since the roots are orthogonal, for each $\eta_i$ the associated element of the Weyl group $w_{\eta_i}$ takes $\eta_i$ to $-\eta_i$ and leaves $\eta_j$ fixed for all $j \neq i$. Hence, their product $w = w_{\eta_1} \cdots w_{\eta_n}$ takes $\lambda$ to $-\lambda$, and so reverses ${\Lambda}$. So assume instead that $(\lambda,{\delta}) > 2$. If $-{\rm id}$ is in the Weyl group, then statement (i) is trivial. So we assume that it is not. Let $T$ act on $V$ via restriction, and pick any nonzero weight $\alpha \in \Tilde{\ell}^*$ in the weight decomposition. By claim (c), we can find some $\sigma \in W$ such that $|(\sigma \cdot \alpha, \lambda) | >1.$ Since $\sigma \cdot \alpha$ also appears in the weight decomposition, this contradicts the assumption that the action of $\Lambda$ on $V$ is semifree.\hfill$\Box$\medskip {\noindent} {\bf Proof of Lemma~\ref{le:simple} (ii).} Let $H \subseteq G$ be a proper subgroup which contains $\Lambda$. Assume that the adjoint action of $\Lambda$ on $\fg/{\mathfrak h}$ is semifree. Let $L$ be the set of roots $\eta \in \Delta$ so that the associated weight space $E_\eta \subset \fg_{\mathbb C}$ lies in ${\mathfrak h}_{\mathbb C}$. Clearly, if $|(\eta, \lambda)| > 1$, then $\eta \in L$. Suppose first that $(\lambda,{\delta}) \leq 2$. By claim (a), there exist orthogonal roots $\eta_1,\ldots,\eta_k \in \Delta$ so that $\lambda = \sum a_i \eta_i$ and so that $(\lambda,\eta_i) = 2$ for every $i$. Since $(\eta_i,\lambda) = 2$, $\eta_i$ lies in $L$ for all $i$. Hence, the associated element of the Weyl group $w_{\eta_i}$ lies in $H$ for all $i$. Thus $w = w_{\eta_1} \cdots w_{\eta_k}$ must lie in $H$. So assume instead that $(\lambda,\delta) > 2$. We see immediately that $\delta$ and $-\delta$ lie in $L$. If $\eta$, $\eta'$ and $\eta + \eta'$ are all roots, then $[E_\eta,E_\eta'] = E_{\eta + \eta'}$. Hence, since ${\mathfrak h}_{\mathbb C}$ is closed under Lie bracket, if $\eta$ and $\eta'$ are in $L$ then $\eta + \eta' \in L$ also, that is, $L$ is closed under addition. Additionally, if $\eta$ and $\eta'$ are roots such that $\delta = \eta + \eta'$, then either $(\lambda,\eta) > 1$ or $(\lambda,\eta') > 1$. If the former holds, then $\eta$ and $-\eta$ lie in $L$. Since $L$ is closed under addition, so do $\eta' $ and $-\eta'$. The other case is identical. Thus, claim (b) implies that every root lies in $L$. This contradicts the claim that $H$ is a proper subgroup.\hfill$\Box$\medskip {\noindent} {\bf Proof of Lemma~\ref{le:simple} (iii).} Let $H \subsetneq G$ be a proper subgroup which contains the maximal torus $T$, and assume that the natural action of ${\Lambda}\subset T$ on $G/H$ has at most twofold isotropy. If $(\lambda,\delta) \leq 2$, then part (iii) follows by the argument used to prove part (i). So assume that $(\lambda, \delta) > 2$. We may also assume that $-{\rm id}$ does not lie in the Weyl group, because otherwise the claim is trivial. Since $H \subset G$ is proper, there exists at least one root $\eta$ so that the associated weight space $E_\eta$ is not contained in ${\mathfrak h}_{\mathbb C}$. Then there is $\sigma \in W$ so that the root $\sigma \cdot \eta$ lies in the positive Weyl chamber. Hence by (d) $\sigma \cdot \eta = {\delta}$, and so $|(\sigma \cdot \eta,\lambda)| > 2.$ Choose $\tilde{\sigma} \in N(T)$ which descends to $\sigma$. Then $\tilde{\sigma} H$ is a fixed point for $T$, and $\sigma \cdot \eta$ is one of the weights for ${\Lambda}$ at this fixed point. This contradicts the fact that the action has at most twofold isotropy.\hfill$\Box$\medskip We are now ready to deduce Theorems~\ref{thm:main0} and~\ref{thmtwo}. In both cases, we will do this by proving the contrapositive, that is, we will assume that $\Lambda$ is an inessential circle subgroup and use this to construct a reversor. Let $\widetilde{G}$ denote the universal cover of $G$. Then $\widetilde{G}$ is the direct product of a compact simply connected semisimple Lie group and a vector space. Since $\Lambda$ is inessential, it lifts to a circle subgroup of $\widetilde{G}$. Since this lift must lie in the compact part of $\widetilde{G}$, we may assume without loss of generality that $\widetilde{G}$ is a compact simply connected semisimple Lie group. In fact, it is enough to prove these claims for the universal cover of $G$, as long as we no longer insist on an effective action but instead allow a finite number of elements of the group to act trivially on $M$. Thus we may assume that $G$ is the product of compact simple and simply connected groups $G_1 \times \cdots \times G_n$. Let $\Lambda_i$ be the projection of $\Lambda$ to $G_i$. Without loss of generality, we may assume that ${\Lambda}_i\ne \{{\rm id}\}$ for all $i$. {\medskip} {\noindent} {\bf Proof of Theorem~\ref{thm:main0}.} Let $G = G_1\times \cdots\times G_n$ as above. Choose $p\in F$ and let $H \subset G$ be the stabilizer of $p$. Then ${\Lambda}\subset H$. There exists a representation $V$ of $H$, called the {\bf isotropy representation}, so that a neighborhood of the $G$-orbit through $p$ is equivariantly diffeomorphic to a neighborhood of the zero section of $G \times_H V$. Fix some simple factor $G_i$, and let $H_i = H \cap G_i$. Assume first that $H_i$ is a proper subgroup. Note that $\fg_i$ is invariant under the action of $\Lambda$. Thus, since $\Lambda$ acts semifreely on $\fg/{\mathfrak h}$ via the adjoint action, $\Lambda_i$ acts semifreely on $\fg_i/{\mathfrak h}_i$. Thus, by Lemma~\ref{le:simple} (ii) there exists an element $h_i \in H_i$ that reverses $\Lambda_i$. So assume on the contrary that $H_i = G_i$. Let $\Lambda'$ be the projection of $\Lambda$ onto the product of all the simple factors except $G_i$. Since ${\Lambda}_i\subset G_i\subset H$ and ${\Lambda}\subset H$, we must have ${\Lambda}'\subset H$. Hence ${\Lambda}'$ acts on $V$. For any integer $k$, let $V_k$ denote the subspace of $V$ on which $\Lambda'$ acts with weight $k$. Since $\Lambda'$ commutes with $G_i$, $V_k$ is a representation of $G_i$. Since only a finite number of elements of $G$ act trivially on $M$, $G_i$ must act nontrivially on $G \times_H V$, and hence also on $V$. Therefore, there is some $k$ so that the representation of $G_i$ on $V_k$ is nontrivial. Because $G_i$ is simple, $\Lambda_i$ must act with both positive and negative weights on $V_k$. But the weights for the action of $\Lambda$ on $V_k$ are the weights for the action of $\Lambda_i$ shifted by $k$. Hence, because $F$ is a semifree fixed point component, $k = 0$ and the action of $\Lambda_i$ on $V_k$ is itself semifree. Therefore by Lemma~\ref{le:simple} (i) there exists $h_i \in G_i = H_i$ that reverses $\Lambda_i$. Since $h_i$ reverses $\Lambda_i$ for each $i$, $g = (h_1,\ldots,h_n)$ reverses $\Lambda$, as required. Moreover, since $H_1 \times \cdots \times H_n\subset H$ (in general they are not equal), $g$ lies in $H$, and hence fixes $p$. \hfill$\Box$\medskip {\noindent} {\bf Proof of Theorem~\ref{thmtwo}.} Fix some simple factor $G_i$. Let $W$ be the Weyl group of $G_i$. Let $T \subset G_i$ be a maximal torus of $G_i$ containing $\Lambda_i$. Let $\Phi: M \longrightarrow \ft^*$ be the moment map for the $T$-action. Pick any $\xi \in \ft$ so that the one parameter subgroup generated by $\xi$ is dense in $T$. Let $p$ be any point which maps to the minimum value of $\Phi^\xi$, the component of $\Phi$ in the direction $\xi$. By construction, $p$ is a fixed point for $T$. Assume first that $\Phi^\xi(p) = 0$, that is, the function $\Phi^\xi$ is nonnegative on $M$. Since the moment polytope $\Phi(M)$ is invariant under the Weyl group $W$, this implies that $\Phi^{{\sigma} \cdot \xi}$ is also nonnegative on $M$ for all ${\sigma}\in W$. Because $G_i$ is simple and $\xi$ is a generic point of $\ft$, for any nonzero $x\in \ft^*$ there exists an element ${\sigma}\in W$ such that $({\sigma}\cdot\xi,x) < 0$. Applying this to $x\in \Phi(M){\smallsetminus} \{0\}$, we see that $\Phi(M)$ must be the single point $\{0\}$, which is impossible, because the action is effective. Therefore, $\Phi(p) \neq 0$. Now let us reconsider the action of $G$ on $M$. Let $H$ be the stabilizer of $p$ in $G$, and let $H_i = H \cap G_i$. Since $\Lambda$ acts with at most twofold isotropy on $G/H\subset M$, $\Lambda_i$ acts with at most twofold isotropy on $G_i/H_i$. Since $\Phi(p)$ is not zero, the stabilizer of $\Phi(p)$ in $G_i$. is a proper subgroup of $G_i$. Since $\Phi$ is equivariant, this implies that $H_i$ is a proper subgroup of $G_i$. By Lemma \ref{le:simple} (iii), this implies that there exists $g_i \in G_i$ which reverses $\Lambda_i$. Then $(g_1,\ldots,g_n)$ reverses ${\Lambda}$. \hfill$\Box$\medskip {\noindent} {\bf Proof of Lemma \ref{claims}.} We now prove claims (a)-(d) on a case by case basis, using the classification of compact simple Lie groups. We will use the notation of \S\ref{ss:coadj}. Note, however, that here $G =\widetilde{G} $ since $G$ is simply connected. {\medskip} {\noindent}{\bf (I)}\,\, Recall that for the group $A_n$, where $n \geq 1$, $\ft = \ft^* = \{ \xi \in {\mathbb R}^n \mid \sum \xi_i = 0 \}$, the roots are $\epsilon_i - \epsilon_j = e_i-e_j$ for $i \neq j$, and the integral lattice is $\Tilde{\ell} = {\mathbb Z}^{n+1} \cap \ft$. The positive Weyl chamber is\footnote {For uniformity, we shall always use the lexigraphical order to choose the positive Weyl chamber.} $$ \{\xi \in \ft \mid \xi_1 \geq \cdots \geq \xi_{n+1}\}. $$ The highest root is $\delta = e_1-e_{n+1}$. If $(\lambda,\delta) = \lambda_1 - \lambda_{n+1} \leq 2$, then $|\lambda_i| \leq 1$ for all $i$. Since $\sum_i \lambda_i = 0$ and $\lambda_i \in {\mathbb Z}$ for all $i$, there are an equal number of $+1$'s and $-1$'s, and the rest are $0$'s. Hence, $\lambda$ is the sum of orthogonal roots of the form $\eta = e_i - e_j$. Since $(\eta,\eta) = 2$, this proves claim (a). Since $\delta = (e_1 - e_k) + (e_k - e_{n+1})$, the roots $\pm (e_1 - e_k)$ and $\pm (e_k - e_{n+1})$ lie in $L$ for all $1 < k < n+1$. If neither $i$ nor $j$ is equal equal to $1$, then $e_i - e_j = -(e_1 - e_i) + (e_1 - e_j)$ is also in $L$. This proves claim (b). We now prove (c). The weight lattice is $\Tilde{\ell}^* = \{ \alpha \in \ft \mid \alpha_i - \alpha_j \in {\mathbb Z} \ \; \forall \ i, j \}$. By permuting the coordinates of $\alpha$, we may assume $\alpha_1 \geq \cdots \ge \alpha_n$. Since $\alpha \neq 0$, there exists $k \in (1,\ldots,n)$ such that $\alpha_{k} - \alpha_{k+1} >0$; since this difference lies in ${\mathbb Z}$, it must be at least $1$. Since $\lambda_1 - \lambda_{n+1} = \lambda_1 + \sum_{i=1}^n \lambda_i > 2$ and $\lambda_i \geq \lambda_{i + n - k}$, $\sum_{i=1}^k \lambda_i + \sum_{i = 1}^{n+1 - k} \lambda_i > 2$. Therefore, either $\sum_{i=1}^k \lambda_i >1$ or $ \sum_{i =1}^{n+1-k} \lambda_i > 1$. In the former case, $$ (\alpha,\lambda) = \sum_{j=1}^n \left( (\alpha_j - \alpha_{j+1}) \sum_{i=1}^j \lambda_i \right) \geq (\alpha_{k} - \alpha_{k+1}) \sum_{i=1}^k \lambda_i > 1. $$ In the latter case, let $\alpha'$ be obtained from $\alpha$ by the permutation which reverses the coordinates, so that $\alpha'_i = \alpha_{n + 2 - i}$. Then $$ (\alpha',\lambda) = \sum_{j=1}^n \left( (\alpha_{n + 2 - j} - \alpha_{n + 1 - j}) \sum_{i=1}^j \lambda_i \right) \leq (\alpha_{k + 1} - \alpha_{k}) \sum_{i=1}^{n + 1 - k} \lambda_i < -1. $$ The only facts we have used are that $\ft = \{ \xi \in {\mathbb R}^n \mid \sum_i\xi_i = 0 \}$, that the Weyl group contains the permutation group $S_n$, and that $\alpha_i - \alpha_j \in {\mathbb Z}$ for any $\alpha \in \Tilde{\ell}^*$. Finally, $\delta$ is the only root in the positive Weyl chamber. {\medskip} {\noindent}{\bf (II)}\,\, Recall that for the group $B_n$, where $n \geq 2$, $\ft = \ft^* = {\mathbb R}^n$, the roots are $\pm e_i$ and $\pm e_i \pm e_j$ for $i \neq j$, and the integral lattice is $\Tilde{\ell} = \{ \xi \in {\mathbb Z}^n \mid \sum_i \xi_i \in 2 {\mathbb Z} \}.$ The positive Weyl chamber is $\{\xi \in \ft \mid \xi_1 \geq \cdots \geq \xi_n \geq 0 \}$. The highest root is $\delta = e_1 + e_2$. If $(\lambda,\delta) = \lambda_1 + \lambda_2 \leq 2$, then either $\lambda_1 = 2$ and $\lambda_i = 0$ for all $i \neq 1$, or $\lambda_i \leq 1$ for all $i$. Either way, since $\sum_i \lambda_i \in 2 {\mathbb Z}$, we can write $\lambda$ as the sum of orthogonal roots $\eta_i$ such that $(\eta_i,\eta_i) = 2$. Since $\delta = (e_1 - e_k) + (e_2 + e_k) = (e_1 + e_k) + (e_2 - e_k)$, the roots $\pm e_1 \pm e_k$ and $\pm e_2 \pm e_k$ lie in $L$ for $k \neq 1 $ or $2$. Since $\delta = (e_1) + (e_2)$, the roots $\pm e_1$ and $\pm e_2$ lie in $L$. Every root can be written as a sum of these roots. Since $-{\rm id}$ lies in the Weyl group, we are done. {\medskip} {\noindent}{\bf (III)}\,\, Recall that for the group $C_n$, where $n \geq 3$, $\ft = \ft^* = {\mathbb R}^n$, the roots are $\pm 2 e_i$ and $\pm e_i \pm e_j$ for $i \neq j$, and the integral lattice is $\Tilde{\ell} = {\mathbb Z}^n$. The positive Weyl chamber is $\{\xi \in \ft \mid \xi_1 \geq \cdots \geq \xi_n \geq 0\ \}$. The highest root is $\delta = 2 e_1$. If $(\lambda,\delta) = 2 \lambda_1 \leq 2$, then $\lambda_i \leq 1$ for all $i$. Since $\lambda \in {\mathbb Z}^n$, we can write $\lambda$ as half the sum of orthogonal roots of the form $ 2 e_i$. Note that $(\lambda, 2 e_i) = 2$. Since $\delta = (e_1 - e_k) + (e_1 + e_k)$, the roots $\pm e_1 \pm e_k$ lie in $L$ for $k \neq 1 $. Every root can be written as a sum of these roots. Since $-{\rm id}$ lies in the Weyl group, we are done. {\medskip} {\noindent}{\bf (IV)}\,\, Recall that for the group $D_n$, where $n \geq 4$, $\ft = \ft^* = {\mathbb R}^n$, the roots are $\pm e_i \pm e_j$ for $i \neq j$, and the integral lattice is $\Tilde{\ell} = \{ \xi \in {\mathbb Z}^n \mid \sum \xi_i \in 2{\mathbb Z} \}.$ The positive Weyl chamber is $\{ \xi \in \ft \mid \xi_1 \geq \cdots \geq \xi_{n-1} \geq |\xi_n| \}$. The highest root is $\delta = e_1 + e_2$. If $(\lambda,\delta) = \lambda_1 + \lambda_2 \leq 2$, then either $\lambda_1 = 2$ and $\lambda_i = 0$ for all $i \neq 1$, or $|\lambda_i| \leq 1$ for all $i$. Either way, since $\sum_i \lambda_i \in 2 {\mathbb Z}$, we can write $\lambda$ as the sum of orthogonal roots $\eta_i$ such that $(\eta_i,\eta_i) = 2$. Since $\delta = (e_1 - e_k) + (e_2 + e_k) = (e_1 + e_k) + (e_2 - e_k)$, the roots $\pm e_1 \pm e_k$ and $\pm e_2 \pm e_k$ lie in $L$ for $k \neq 1 $ or $2$. Every root can be written as a sum of these roots. Now assume that $(\delta,\lambda) = \lambda_1 + \lambda_2 > 2$. Consider a nonzero weight $\alpha \in \Tilde{\ell}^* = \{ \alpha \in {\mathbb R}^n \mid \alpha_i \pm \alpha_j \in {\mathbb Z} \ \forall\ i, j \}.$ By applying the Weyl group, we may assume $\alpha$ lies in the positive Weyl chamber. Since $\lambda$ also lies in the positive Weyl chamber, $\alpha_i \lambda_i \geq 0$ for all $i \neq n$. Moreover, since $\alpha_{n-1} \geq |\alpha_n|$, and $\lambda_{n-1} \geq |\lambda_n|$, $\alpha_{n-1} \lambda_{n-1} + \alpha_n \lambda_n \geq 0$. Therefore, $\alpha_3 \lambda_3 + \cdots + \alpha_n \lambda_n \geq 0.$ (Here, we have used that $n \geq 4$.) Since $\alpha$ is nonzero, either $\alpha_1 \geq 1$, or $\alpha_1 = \alpha_2 = \frac{1}{2}$. In either case, $\alpha_1 \lambda_1 + \alpha_2 \lambda_2 > 1$ . (In the first case, we use the fact that $\lambda_1 + \lambda_2 > 2$ and $\lambda_1 \geq \lambda_2$ implies that $\lambda_1 > 1$.) Therefore, $(\alpha,\lambda) \geq \alpha_1 \lambda_1 + \alpha_2 \lambda_2 > 1$, This proves claim (c). Finally, $\delta$ is the only root in the positive Weyl chamber. {\medskip} {\noindent}{\bf (V, a)}\,\, Recall that for the group $E_6$, $\ft = \ft^* = {\mathbb R}^6$ and the roots are $2 \epsilon$, $\epsilon_i - \epsilon_j$, and $\epsilon_i + \epsilon_j + \epsilon_k \pm \epsilon$ for $i, j,$ and $k$ distinct, where $\epsilon = \frac{1}{2 \sqrt{3}}(1,1,1,1,1,1)$. Therefore $$ \Tilde{\ell} = \Big\{ n\epsilon + (\xi_1,\ldots,\xi_6)\; \Big|\; \sum_{i=1}^6 \xi_i = 0, n \in {\mathbb Z}, \mbox{and} \ \frac{n}{2}+ \xi_i \in {\mathbb Z} \ \forall \ i \Big\}. $$ The positive Weyl chamber is $$ \Bigl\{ n \epsilon + (\xi_1,\ldots,\xi_6) \in \ft\; \Big| \; \sum_{i=1}^6 \xi_i = 0, \xi_2 \geq \cdots \geq \xi_6,\; \xi_1 + \xi_5 + \xi_6 \geq n/2 \geq 0\Bigr\}. $$ (Note that these conditions imply $\xi_1 \geq \xi_2$.) The highest root is $\delta = \epsilon_1 - \epsilon_6$. Write $\lambda = n\epsilon + (\xi_1,\ldots,\xi_6)$, where $\sum_i \xi_i = 0$. Assume that $(\lambda,\delta) = \xi_1 - \xi_6 \leq 2$. Combining the inequalities $\xi_1 - \xi_6 \leq 2$, $\xi_4 \geq \xi_5$, $\xi_4 \geq \xi_6$, and $\xi_1 + \xi_5 + \xi_6 \geq \frac{n}{2}$, we see that $\xi_4 \geq \frac{n-4}{6}$. Since also $\xi_2 + \xi_3 + \xi_4 \leq 0$, $\xi_2 \geq \xi_4$, and $\xi_3 \geq \xi_4$, we have $\xi_4 \leq 0$. Moreover, in both cases, if the final inequality in the sentence is an equality, so are all the preceding ones. Since $n \geq 0$, $0 \geq \xi_4 \geq -\frac{4}{6}$. Since $\lambda \in \Tilde{\ell}$, $\xi_4 = 0$ or $\xi_4 = -\frac{1}{2}$. In the former case, $\xi_2 = \xi_3 = \xi_4 = 0$, so $\xi_1 + \xi_5 + \xi_6 = 0$, so $n = 0$. Hence, $\lambda = (\epsilon_1 - \epsilon_6).$ In the latter case, $n$ is odd, so $\xi_4 = -\frac{1}{2} \geq \frac{n-4}{6}$ implies that $n = 1$. In this case, $\lambda = (\epsilon_1 - \epsilon_6) + (\epsilon + \epsilon_1 + \epsilon_2 + \epsilon_6)$. This proves claim (a). Since $\delta = (\epsilon_1 -\epsilon_i) + (\epsilon_i - \epsilon_6),$ the roots $\pm (\epsilon_1 - \epsilon_i)$ and $\pm(\epsilon_i - \epsilon_6)$ lie in $L$ for all $1 < i < 6$. Moreover, $\delta = (\epsilon + \epsilon_1 + \epsilon_i + \epsilon_j) - (\epsilon + \epsilon_i + \epsilon_j +\epsilon_6)$, so the roots $\pm(\epsilon + \epsilon_1 + \epsilon_i + \epsilon_j)$ and $\pm(\epsilon + \epsilon_i + \epsilon_j + \epsilon_6)$ lie in $L$ for all $1 < i < j < 6$. Since, for example, $\epsilon + \epsilon_1 + \epsilon_2 + \epsilon_3 = \epsilon - \epsilon_4 - \epsilon_5 - \epsilon_6$, it follows easily that $L$ contains all roots. Let $\alpha \in \Tilde{\ell}^*$ be a nonzero weight. Write $\lambda = n{\varepsilon} + \xi$ as before. By applying the Weyl group, we may assume that $\alpha = m\epsilon + (\zeta_1,\ldots,\zeta_6)$ is in the positive Weyl chamber. Since $(\alpha,\lambda) \geq (\zeta,\xi)$, it is enough to show that $(\zeta,\xi) > 1$. This fact now follows from the argument from $A_5$, since $(\delta,{\lambda}) = ({\delta},\xi)> 2$, since the Weyl group contains the permutation group $S_5$, and since $\zeta$ must satisfy $\zeta_i - \zeta_j \in {\mathbb Z}$. Finally, $\delta$ is the only root in the positive Weyl chamber. {\medskip} {\noindent}{\bf (V, b)}\,\, Recall that for the group $E_7$, $\ft = \ft^* = \{ \xi \in {\mathbb R}^8 \mid \sum \xi_i = 0 \}$, the roots are $\epsilon_i - \epsilon_j$ and $\epsilon_i + \epsilon_j + \epsilon_k + \epsilon_l$ for $i,j,k,$ and $l$ distinct, and the integral lattice is $\Tilde{\ell} = \{\xi \in \ft \mid \xi_i \pm \xi_j \in {\mathbb Z} \ \forall \ i,j \}.$ The positive Weyl chamber is $$ \{\xi \in \ft \mid \xi_2 \geq \cdots \geq \xi_8 \ \mbox{and}\ \xi_1 + \xi_6 + \xi_7 + \xi_8 \geq 0 \}. $$ (Note that this automatically implies that $\xi_1 \geq \xi_2$.) The highest root is $\delta = \epsilon_1 - \epsilon_8$. Assume that $(\delta,\lambda) = \lambda_1 - \lambda_8 \leq 2$. Combining the inequalities $\lambda_1 - \lambda_8 \leq 2$, $\lambda_1 + \lambda_6 + \lambda_7 + \lambda_8 \geq 0$, and $\lambda_5 \geq \lambda_i$ for $i = 6,7$ and $8$, we see that $\lambda_5 \geq -\frac{1}{2}$. Since also $\lambda_2 + \lambda_3 + \lambda_4 + \lambda_5 \geq 0$ and $\lambda_i \geq \lambda_5$ for $i = 2,3$ and $4$, $\lambda_5 \leq 0$. Moreover, in both cases, if the last inequality in the sentence is an equality, all the inequalities are equalities. Since $\lambda \in \ell$, the only possibilities are $\lambda_5 = 0$ or $\lambda_5 = -\frac{1}{2}.$ In the former case, we must have $\lambda = \epsilon_1 - \epsilon_8$. In the latter case, the only possibilities are $\lambda = (\epsilon_1 + \epsilon_2 + \epsilon_3 + \epsilon_4) + (\epsilon_1 - \epsilon_4)$, or $\lambda = (\epsilon_1 + \epsilon_2 + \epsilon_3 + \epsilon_4) + (\epsilon_1 - \epsilon_4) + (\epsilon_2 - \epsilon_3)$. The proves claim (a). Since $\delta = (\epsilon_1 - \epsilon_i) + (\epsilon_i - \epsilon_8)$, the roots $\pm(\epsilon_1 - \epsilon_i)$ and $\pm(\epsilon_i - \epsilon_8)$ lie in $L$ for all $1<i<8.$ Since $\delta = (\epsilon_1 + \epsilon_i + \epsilon_j + \epsilon_k) - (\epsilon_i + \epsilon_j + \epsilon_k + \epsilon_8)$, the roots $\pm (\epsilon_1 + \epsilon_i + \epsilon_j + \epsilon_k)$ and $\pm (\epsilon_i + \epsilon_j + \epsilon_k + \epsilon_8)$ also lie in $L$ for all $1<i<j<k<8$. All roots can be written as a sum of these roots. This proves claim (b). Since $\Tilde{\ell} \subset {\mathbb Z}^8 \cap \ft$, $\alpha_i - \alpha_j \in {\mathbb Z}$ for every $\alpha \in \Tilde{\ell}^* \subset \ft^*$. Hence, the argument for claim (c) follows from the argument for $A_7$. Finally, $\delta$ is the only root in the positive Weyl chamber. {\medskip} {\noindent}{\bf (V, c)}\,\, For the group $E_8$, $\ft = \ft^* = \{\xi \in {\mathbb R}^9 \mid \sum \xi_i = 0$\} and the roots are $\epsilon_i - \epsilon_j$, and $\pm( \epsilon_i + \epsilon_j + \epsilon_k)$ for $i, j$ and $k$ distinct. Hence the integral lattice is $$\Tilde{\ell} = \{\xi \in \ft \mid 3 \xi_i \in {\mathbb Z} \ \mbox{and}\ \xi_i - \xi_j \in {\mathbb Z} \ \forall \ i,j \}.$$ The positive Weyl chamber is $$ \{\xi \in \ft \mid \xi_2 \geq \cdots \geq \xi_9 \ \mbox{and} \ \xi_2 + \xi_3 + \xi_4 \leq 0 \}. $$ (Note that these conditions imply that $\xi_1 \geq \xi_2$.) The highest root is $\delta = \epsilon_1 - \epsilon_9$. Assume that $(\delta,\lambda) = \lambda_1 - \lambda_9 \leq 2$. Combining the inequalities $$ \lambda_1 - \lambda_9 \leq 2,\quad \lambda_1 + \lambda_5 + \lambda_6 + \lambda_7 + \lambda_8 + \lambda_9 \geq 0, \quad \lambda_i \leq \lambda_4,i > 4, $$ we see that $\lambda_4 \geq - \frac{1}{3}$. Since $\lambda_2 + \lambda_3 + \lambda_4 \leq 0$ and $\lambda_2 \geq {\lambda}_3\geq \lambda_4$, $\lambda_4 \leq 0$. Moreover, in both cases, if the last inequality in the sentence is an equality, all the inequalities are equalities. Since $\lambda \in \ell$, the only possibilities are $\lambda_4 = 0$ or $\lambda_4 = -\frac{1}{3}$. In the former case, $\lambda = \epsilon_1 - \epsilon_9$. In the latter case, $\lambda = (\epsilon_1 - \epsilon_9) + (\epsilon_1 + \epsilon_2 + \epsilon_9)$. Claim (a) follows. We now notice that $\delta = \epsilon_1 - \epsilon_9 = (\epsilon_1 - \epsilon_k) + (\epsilon_k - \epsilon_9) = (\epsilon_1 + \epsilon_i + \epsilon_j) - (\epsilon_i + \epsilon_j + \epsilon_9)$ for all $1 < k < 9$ and $1 < i < j < 9$. Therefore, the corresponding roots $\pm(\epsilon_1 - \epsilon_k)$, $\pm(\epsilon_k - \epsilon_9)$, $\pm(\epsilon_1 + \epsilon_i + \epsilon_j)$, and $\pm(\epsilon_i + \epsilon_j + \epsilon_9)$ all lie in $L$. Since every root can be written as a sum of these roots, claim (b) follows. Since $\Tilde{\ell} \subset {\mathbb Z}^9 \cap \ft$, $\alpha_i - \alpha_j \in {\mathbb Z}$ for every $\alpha \in \Tilde{\ell}^* \subset \ft^*$. Hence, the argument for claim (c) carries over from the argument for the group $A_8$. Finally, $\delta$ is the only root in the positive Weyl chamber. {\medskip} {\noindent}{\bf (VI)}\,\, For the group $F_4$, $\ft = \ft^* = {\mathbb R}^4$. The roots are $ \pm e_i$, $e_i \pm e_j$ for $i \neq j$, and $\frac{1}{2}(\pm e_1 \pm e_2 \pm e_3 \pm e_4)$. Hence the integral lattice is $\Tilde{\ell} = \{ \xi \in {\mathbb Z}^4 \mid \sum \xi_i \in 2 {\mathbb Z} \}$. The positive Weyl chamber is $$ \{\xi \in \ft \mid \xi_2 \geq \xi_3 \geq \xi_4 \geq 0 \ \mbox{and} \ \xi_1 \geq \xi_2 + \xi_3 + \xi_4 \}. $$ (Note that automatically $\xi_1 \geq \xi_2$.) The highest root is $\delta = e_1 + e_2$. The argument for claim (a) carries over word for word from the argument for $B_4$. Notice that if $k=3$ or $4$ \begin{eqnarray*} \delta & = & e_1 + e_2 = (e_1) + (e_2)\; =\; (e_1 - e_k) + (e_2 + e_k) \;=\; (e_1 + e_k) + (e_2 - e_k) \\ &=& \frac{1}{2}(e_1 + e_2 + e_3 + e_4) + \frac{1}{2}(e_1 + e_2 - e_3 - e_4)\\ & = & \frac{1}{2}(e_1 + e_2 - e_3 + e_4) + \frac{1}{2}(e_1 + e_2 + e_3 - e_4). \end{eqnarray*} Hence, the corresponding roots all lie in $L$. Since every root can be written as the sum of these roots, this proves claim (b). Since $-{\rm id}$ lies in the Weyl group, we are done. {\medskip} {\noindent}{\bf (VII)}\,\, For the group $G_2$, $\ft = \ft^* = \{\xi \in {\mathbb R}^3 \mid \sum \xi_i = 0$\}. The roots are $\pm \epsilon_i$ and $\epsilon_i - \epsilon_j$ for $i$ and $j$ distinct. The positive Weyl chamber is $\{ \xi \in \ft^* \mid 0 \geq \xi_2 \geq \xi_3 \}$. (Note that automatically $\xi_1 \geq \xi_2$.) The integral lattice is $\Tilde{\ell} ={\mathbb Z}^3 \cap \ft$. The highest root is $\delta = \epsilon_1 - \epsilon_3$. The argument for claim (a) follows the argument for $A_3$ word for word. Since $\delta = (\epsilon_1 - \epsilon_2) + (\epsilon_2 - \epsilon_3) = (\epsilon_1) + (-\epsilon_3)$, the roots $\pm (\epsilon_1 - \epsilon_2)$, $\pm (\epsilon_2 - \epsilon_3)$, $\pm \epsilon_1$ and $\pm \epsilon_3$ all lie in $L$. Since every root can be written as a sum of these roots, claim (b) follows. Since $-{\rm id}$ lies in the Weyl group, we are done. {\medskip}
{ "timestamp": "2005-03-22T18:36:20", "yymm": "0503", "arxiv_id": "math/0503467", "language": "en", "url": "https://arxiv.org/abs/math/0503467" }
\section{Introduction} In heavy ion collisions an extended hot and dense fireball medium is created. The properties (mass, width, momentum distribution, yield) of the produced resonances depend on the fireball conditions of temperature and pressure. During the fireball expansion the short lived resonances and their hadronic decay daughters may interact with the medium. Two freeze-out surfaces can be defined, chemical and thermal, representing the conditions when inelastic and elastic interactions cease respectively. In a dynamical evolving system produced resonances decay and may get regenerated. Hadronic decay daughters of resonances which decay inside the medium may also scatter with other particles from the medium. For SPS and RHIC energies these are mostly pions. This results in a signal loss, because the reconstructed invariant mass of the decay daughters no longer matches that of the parent. Leptonic decay daughters on the other hand are unaffected by the nuclear medium due to their small interaction cross section. The rescattering and regeneration (pseudo-elastic) processes for resonances and their decay particles depend on the individual cross sections and are dominant after chemical but before the kinetic freeze-out. These interactions can result in changes of the reconstructed resonance yields, momentum spectra, widths and mass positions. Rescattering will decrease the measured resonance yields while regeneration will increase them. Microscopic model calculations attempt to include every step in a heavy ion interaction in terms of elastic and inelastic interactions of hadrons and strings. They are therefore better able to describe the rescattering and regeneration of the resonances from fireball interactions. The prediction of a specific model (UrQMD) is a signal loss for some of the resonances due to more rescattering than regeneration in the low momentum region p$_{\rm T}<1$~GeV for the hadronic decay channels \cite{ble02,ble02b}. Comparisons between the yield and momentum spectra of the hadronic and leptonic decay channels can indicate the magnitude of the rescattering and regeneration contribution between chemical and thermal freeze-out. In order to try to understand the medium effect during the evolution and expansion of the hot and dense fireball, we compare resonance yields and spectra (width and mass) from elementary p+p and heavy ion collisions and the results from the leptonic and hadronic decay channels. An observed difference may give an indication of in-medium modification of resonance properties. \section{Resonance Reconstruction} The signal loss due to rescattering is caused by the method of measurement, the invariant mass is not properly reconstructed if one of the decay daughters rescatters with another particle of the surrounding medium. All the resonances are reconstructed by the invariant mass of the decay daughters. The decay candidates are identified by different techniques, their energy loss (dE/dx), energy or displaced vertex (V0-reconstruction). The resonance signal is obtained by the invariant mass reconstruction of each daughter combination and subtraction of the combinatorial background calculated by mixed event or like-signed techniques. The resonance ratios, spectra and yields are measured at mid-rapidity for RHIC at $\sqrt{s_{\rm NN}} = $ 200 GeV and over 4$\pi$ for SPS at $\sqrt{s_{\rm NN}} = $ 17 GeV. The central trigger selection for Au+Au collisions at RHIC takes the 5\% or 10\% and for Pb+Pb collisions at SPS the 5\% of the most central inelastic interactions. The setup for the p+p interaction is a minimum bias trigger. \section{Resonance Yields} The resonance multiplicities at mid-rapidity for p+p and peripheral to central Au+Au collisions at RHIC energies are obtained for $\phi$(1020) \cite{ma04}, $\Delta(1232)^{++}$ \cite{mar04qm}, K(892) \cite{zha04} $\Lambda$(1520) \cite{gau04,mar03} and $\Sigma$(1385) \cite{sal04}. In order to compare different collision systems we normalize the yield to the yield of the corresponding measured ground state particle. Under the assumption that the Au+Au collision system is only a superposition of p+p collisions we would expect the same resonance/non-resonance ratio. Fig.\ref{part} shows the resonance/non-resonance ratios normalized to the K(892)/K measurement in p+p. The $\Lambda$(1520)/$\Lambda$ and the K(892)/K ratio decreases from p+p to peripheral and central Au+Au collisions. \begin{figure}[htb] \centering \includegraphics[width=0.8\textwidth]{resonances_ncharge_sqm2004.eps} \caption{Resonance/non-resonance ratios of $\phi$/K$^{-}$ \cite{ma04}, $\Delta^{++}$/p \cite{mar04qm}, $\rho/\pi$ \cite{fac04}, K(892)/K$^{-}$ \cite{zha04} and $\Lambda$(1520)/$\Lambda$ \cite{gau04,mar03} for p+p and Au+Au collisions at $\sqrt{s_{\rm NN}} = $ 200 GeV at mid-rapidity. The ratios are normalized to the K(892)/K$^{-}$ ratio measurement in p+p. Statistical and systematic errors are included.} \label{part} \end{figure} The observed ratios can not be described by thermal model predictions \cite{pbm01}, most likely because rescattering of the decay daughters in the medium and regeneration are contributing to the yield. If only rescattering occurs then the shorter lifetime of the K(892) (4 fm/c) compared to the $\Lambda$(1520) (13 fm/c) would result in a larger suppression for K(892)/K than for the $\Lambda$(1520)/$\Lambda$ ratio. This implies that the regeneration cross section is larger for the K+$\pi$ channel than for the K+p channel. The $\phi$(1020)/K ratio is constant in all collision systems within errors and can be described with the thermal model, which is expected because only a small fraction of the $\phi$(1020) are decaying inside the fireball due to the long lifetime of the $\phi$(1020) (46 fm/c). The expected contribution of rescattering for the short lived $\Delta$(1232) (1.7 fm/c) is larger than that for the K(892) and the $\Lambda$(1520). However the $\Delta$(1232)/p ratio does not decrease from p+p to Au+Au collisions and is on the order of 41\% $\pm$ 22\% higher than the thermal model prediction. This indicates a large cross section for the regeneration of $\Delta$(1232) resonance in the p+$\pi$ channel. the $\Delta$(1232) can be re-created until T = 80-90 MeV close to the kinetic freeze-out \cite{ble04}. The $\Sigma$(1385)/$\Lambda$ ratio appears to follow the same trend as the $\Delta$(1232)/p \cite{sal04}. This implies that the $\Lambda$+$\pi$ regeneration cross section is nearly as high as the p+$\pi$ regeneration cross section. From this observation we can conclude that there is a ranking order of the cross section for the different regeneration processes: \\ $\sigma_{p+\pi}$ $\geq$ $\sigma_{\Lambda+\pi}$ $>$ $\sigma_{K+\pi}$ $>$ $\sigma_{K+p}$. The microscopic model calculations (UrQMD) are able to reproduce the resonance/non-resonance ratios in Au+Au collisions for most resonances \cite{ble02,ble02b}. However the UrQMD prediction for the $\Sigma$(1385)/$\Lambda$ ratio is in the order of 40\% $\pm$ 20\% too high. In this calculation the assumption was made that the $\Lambda$+$\pi$ regeneration cross section is the same than for p+$\pi$. The trend of data would suggest that the $\Lambda$+$\pi$ regeneration cross section is smaller than the p+$\pi$ cross section. \begin{figure}[htb] \vspace{0.5cm} \centering \includegraphics[width=0.5\textwidth,angle=-90]{pt_thermal_2.eps.eps} \caption{Transverse momentum distribution of $\Delta^{++}$, $\rho$, K(892) and $\phi$(1020) in central ($\rho$ peripheral) Au+Au collisions from the STAR experiment at RHIC compared to thermal model predictions \cite{flo04}.} \label{ptmodel} \end{figure} \section{Momentum Distribution} Fig.~\ref{ptmodel} shows the momentum distribution of $\Delta^{++}$, $\rho$, K(892) and $\phi$(1020) from central Au+Au collisions ($\rho$ peripheral) from the STAR experiment at RHIC compared to thermal model predictions from W. Florkowski. The measured K(892) distribution deviates from the model predictions in the low momentum region. This observation is consistent with the UrQMD prediction of a signal loss due to rescattering in the low momentum region. Based on the similarity in the trends between the $\Lambda$(1520)/$\Lambda$ and the K(892)/K in Fig.~\ref{part}, one would also expect a signal loss in the low momentum region for the $\Lambda$(1520) compared to the thermal model predictions. The good agreement of the $\Delta^{++}$ momentum distribution with the model indicats that the regeneration also takes place predominantly in the low momentum region. This low momentum signal loss of resonances due to rescattering in results in a higher inverse slope parameter and a higher $\langle$p$_{\rm T}$$\rangle$. The STAR data from p+p and Au+Au collisions at $\sqrt{s_{\rm NN}} = $ 200 GeV confirm this trend. A strong increase $\langle$p$_{\rm T}$$\rangle$ for resonances is observed from p+p to the most peripheral Au+Au measurement. The same trend is not present for the ground state particles (see Fig~\ref{resopt}) \cite{ma04,mar04qm,zha04}. \begin{figure}[h] \centering \vspace{0.5cm} \includegraphics[width=0.55\textwidth]{resonancept.eps} \caption{The $\langle$p$_{\rm T}$$\rangle$ for resonances and ground state particles in p+p and Au+Au collisions versus number of charged particles \cite{ma04,mar04qm,zha04,sal04}.} \label{resopt} \end{figure} \section{Time Scale} Depending on the length of the time interval between chemical and kinetic freeze-out, $\Delta \tau$, the magnitude of the suppression factor of the measured resonance will change due to contributions from rescattering and regeneration. A model using thermally produced particle yields at chemical freeze-out and an additional rescattering phase, including the lifetime of the resonances and decay product interactions within the expanding fireball, can yield an estimated $\Delta \tau$ \cite{tor01,tor01a,mar02}. This model does not include regeneration and therefore predicts a lower limit of the lifetime between the two freeze-out surfaces. The two ratios K(892)/K and $\Lambda$(1520)/$\Lambda$ are expected to have a larger rescattering contribution. A $\Delta\tau$ $>$ 4~fm/c results if chemical freeze-out occurs at 160 MeV. \section{Leptonic and Hadronic Decay Channels} In heavy ion collisions direct comparisons of the spectra and yields obtained from leptonic and hadronic decay channels of a single resonance may show the influence of the hadronic interaction phase after chemical freeze-out. The $\phi$(1020) is one of the resonances where we have measurements of the leptonic and hadronic decay channel. At SPS energies the reconstruction of the $\phi$(1020) in the different decay channels seemingly leads to differing $\phi$(1020) kinematics and yields ($\phi$ puzzle). \begin{figure}[htb] \vspace{0.8cm} \centering \includegraphics[width=0.6\textwidth]{kolo_phi_eps_rot_single2.eps} \caption{Transverse momentum distribution of the hadronic decay $\phi$(1020) $\rightarrow$ K$^{+}$ + K$^{-}$ from NA49 \cite{fri97} and the leptonic decay $\phi$(1020) $\rightarrow$ $\mu^{+}$ + $\mu^{-}$ from NA50 \cite{wil99}.} \label{phi} \end{figure} Fig.~\ref{phi} shows the transverse momentum distribution from the hadronic decay $\phi$~$\rightarrow$~K$^{+}$~+~K$^{-}$ (NA49) and the leptonic decay $\phi$~$\rightarrow$~$\mu^{+}$~+$\mu^{-}$ (NA50) \cite{fri97,wil99}. The inverse slope parameter from fits to the momentum spectra, indicated as lines, are T~=~305~$\pm$~15~MeV for hadronic decay and T~=~218~$\pm$~10~MeV for leptonic decay. The extracted yield from the extrapolation of the momentum spectrum of the leptonic decay is a factor of 4~$\pm$~2 higher than the one for the hadronic decay. Measurements of the $\phi$(1020) reconstructed via the hadronic and leptonic decay from CERES presented by A. Marin \cite{mari04} at this conference confirm the NA49 results ($\phi$ $\rightarrow$ K$^{+}$ + K$^{-}$) in terms yield and momentum distribution and the NA50 yield for the $\phi$ $\rightarrow$ $e^{+}$ + $e^{-}$ decay. First results from NA60 experiment show an improved invariant mass signal (significance $>$ 20) for the $\phi$ $\rightarrow$ $\mu^{+}$ + $\mu^{-}$ channel \cite{dam04}, which should result in a conclusive contribution to the $\phi$ puzzle at SPS. Microscopic calculations (UrQMD) estimate a suppression of 20-30\% of the $\phi$(1020) yield in the hadronic decay channel due to rescattering of the kaon decay daughters in the low momentum region p$_{\rm T}$~$<$~1~GeV \cite{ble02,ble02b}. The rescattering is negligible for the leptonic decay due to the very low cross section of interaction with the hadronic phase. Therefore the lower signal in the low momentum region of the hadronic decay (NA49) compared to the leptonic decay (NA50) is in agreement with the model. However the signal loss of 20-30\% from the model calculation is not sufficient to explain the factor of 4~$\pm$~2 in the measured yield of the data. This allows for possible medium effects on the resonance production which are likely to occur at an earlier stage, before chemical freeze-out. Alternative calculations to describe in medium modification of the $\phi$(1020) resonance were published recently by K. Haglin and E. Kolomeitsev \cite{hag04,hag04a,kol99}. Here the lifetime of the $\phi$(1020) resonance is modified towards smaller lifetimes due to modification of the spectral functions in the hot and dense fireball and therefore more of the $\phi$(1020) resonances decay inside the medium. This will introduce a larger signal loss due to rescattering of the hadronic decay daughters. \section{Feeddown from Resonances} Finally I would like to conclude with a small remark. If we interpret particle spectra of ground state particles we have to take into account that a large fraction of the particles are coming from resonance feeddown, as already pointed out by E. Schnedermann et al. \cite{sch93}. For the proton we have 42\% from $\Lambda$'s, 21\% from $\Delta$'s, and 11\% from $\Sigma^{0}$'s (statistical model \cite{raf}). Therefore only 26\% of the protons are primary produced protons. 35\% of the $\Lambda$'s are from $\Sigma$(1385) and 20\% from $\Sigma^{0}$'s (statistical model) decays. If we take the contribution of multiple rescattering and regeneration processes during the expansion of the fireball source into account, the number of primary particles will be further reduced, because the regeneration does not necessarily involve the actual resonance decay particles. Since the lifetimes of the $\rho$ and $\Delta$(1232) are very short compared to the lifetime of the fireball, we would expect a larger number of $\pi$'s and protons coming from a $\Delta$(1232) decay than from higher mass baryons. Therefore many $\pi$'s and protons are coming from a later stage of the evolution of the fireball source and their momentum distribution might be different from the primary produced particles. Conclusions based on the momentum distributions of particle spectra in terms of flow and freeze-out temperatures have to take the contribution from resonance decays into account. \section*{REFERENCES}
{ "timestamp": "2005-03-25T22:04:18", "yymm": "0503", "arxiv_id": "nucl-ex/0503011", "language": "en", "url": "https://arxiv.org/abs/nucl-ex/0503011" }
\part{\partial} \define\Wb{\WW^\bul} \define\Bpq{\Bumpeq} \define\rank{\text{\rm rank}} \define\doa{\dot a} \define\dua{\dot{\un a}} \define\dw{\dot w} \define\dt{\dot t} \define\ds{\dot s} \define\dx{\dot x} \define\dy{\dot y} \define\dts{\dot{\ts}} \define\ua{\un a} \define\us{\un s} \define\uD{\un D} \define\uL{\un L} \define\uP{\un P} \define\uS{\un S} \define\uX{\un X} \define\uZ{\un Z} \define\ufs{\un{\fs}} \define\ucs{\un{\cs}} \define\Stab{\text{\rm Stab}} \define\mpb{\medpagebreak} \define\Up{\Upsilon} \define\haG{\hat\G} \define\hA{\hat A} \define\hL{\hat L} \define\hG{\hat G} \define\hX{\hat X} \define\hM{\hat M} \define\hY{\hat Y} \define\hb{\hat b} \define\hf{\hat f} \define\hy{\hat y} \define\hcl{\hat{\cl}} \define\hB{\hat B} \define\hC{\hat C} \define\hT{\hat T} \define\hD{\hat D} \define\hZ{\hat Z} \define\hK{\hat K} \define\hx{\hat x} \define\htt{\hat\t} \define\hfZ{\hat{\fZ}} \define\da{\dagger} \define\dsv{\dashv} \define\Lie{\text{\rm Lie }} \define\po{\text{\rm pos}} \define\frl{\forall} \define\pe{\perp} \define\si{\sim} \define\wt{\widetilde} \define\sqc{\sqcup} \define\ovs{\overset} \define\qua{\quad} \define\baG{\bar\G} \define\bag{\bar\g} \define\bao{\bar\o} \define\bcl{\bar{\cl}} \define\bc{\bar c} \define\bg{\bar g} \define\bj{\bar j} \define\bn{\bar n} \define\bp{\bar p} \define\bs{\bar s} \define\bt{\bar t} \define\bu{\bar u} \define\baf{\bar f} \define\br{\bar r} \define\bay{\bar y} \define\bA{\bar A} \define\bC{\bar C} \define\bE{\bar E} \define\bF{\bar F} \define\bG{\bar G} \define\bK{\bar K} \define\bN{\bar N} \define\bP{\bar P} \define\bR{\bar R} \define\bS{\bar S} \define\bT{\bar T} \define\bX{\bar X} \define\bY{\bar Y} \define\bZ{\bar Z} \define\bV{\bar V} \define\bnu{\bar\nu} \define\bce{\bar\ce} \define\bvt{\bar\vt} \define\bpi{\bar\p} \define\bsi{\bar\s} \define\bSi{\bar\Si} \define\bbq{\bar{\QQ}_l} \define\baP{\bar\Pi} \define\lb{\linebreak} \define\eSb{\endSb} \define\gt{\gets} \define\bin{\binom} \define\op{\oplus} \define\h{\frac{\hphantom{aa}}{\hphantom{aa}}} \redefine\sp{\spadesuit} \define\em{\emptyset} \define\imp{\implies} \define\ra{\rangle} \define\n{\notin} \define\iy{\infty} \define\m{\mapsto} \define\do{\dots} \define\la{\langle} \define\bsl{\backslash} \define\lras{\leftrightarrows} \define\lra{\leftrightarrow} \define\Lra{\Leftrightarrow} \define\hra{\hookrightarrow} \define\sm{\smallmatrix} \define\esm{\endsmallmatrix} \define\sub{\subset} \define\bxt{\boxtimes} \define\T{\times} \define\ti{\tilde} \define\nl{\newline} \redefine\i{^{-1}} \define\fra{\frac} \define\un{\underline} \define\ov{\overline} \define\ot{\otimes} \define\bcc{\thickfracwithdelims[]\thickness0} \define\ad{\text{\rm ad}} \define\Ad{\text{\rm Ad}} \define\Hom{\text{\rm Hom}} \define\End{\text{\rm End}} \define\Aut{\text{\rm Aut}} \define\Ind{\text{\rm Ind}} \define\ind{\text{\rm ind}} \define\Res{\text{\rm Res}} \define\res{\text{\rm res}} \define\Ker{\text{\rm Ker}} \redefine\Im{\text{\rm Im}} \define\sg{\text{\rm sgn}} \define\tr{\text{\rm tr}} \define\dom{\text{\rm dom}} \define\supp{\text{\rm supp}} \define\card{\text{\rm card}} \define\bst{\bigstar} \define\he{\heartsuit} \define\clu{\clubsuit} \define\di{\diamond} \define\a{\alpha} \redefine\b{\beta} \redefine\c{\chi} \define\g{\gamma} \redefine\d{\delta} \define\e{\epsilon} \define\et{\eta} \define\io{\iota} \redefine\o{\omega} \define\p{\pi} \define\ph{\phi} \define\ps{\psi} \define\r{\rho} \define\s{\sigma} \redefine\t{\tau} \define\th{\theta} \define\k{\kappa} \redefine\l{\lambda} \define\z{\zeta} \define\x{\xi} \define\vp{\varpi} \define\vt{\vartheta} \define\vr{\varrho} \redefine\G{\Gamma} \redefine\D{\Delta} \redefine\O{\Omega} \define\Si{\Sigma} \define\Th{\Theta} \redefine\L{\Lambda} \define\Ph{\Phi} \define\Ps{\Psi} \redefine\aa{\bold a} \define\bb{\bold b} \define\boc{\bold c} \define\dd{\bold d} \define\ee{\bold e} \define\bof{\bold f} \define\hh{\bold h} \define\ii{\bold i} \define\jj{\bold j} \define\kk{\bold k} \define\mm{\bold m} \define\nn{\bold n} \define\oo{\bold o} \define\pp{\bold p} \define\qq{\bold q} \define\rr{\bold r} \redefine\ss{\bold s} \redefine\tt{\bold t} \define\uu{\bold u} \define\vv{\bold v} \define\ww{\bold w} \define\zz{\bold z} \redefine\xx{\bold x} \define\yy{\bold y} \redefine\AA{\bold A} \define\BB{\bold B} \define\CC{\bold C} \define\DD{\bold D} \define\EE{\bold E} \define\FF{\bold F} \define\GG{\bold G} \define\HH{\bold H} \define\II{\bold I} \define\JJ{\bold J} \define\KK{\bold K} \define\LL{\bold L} \define\MM{\bold M} \define\NN{\bold N} \define\OO{\bold O} \define\PP{\bold P} \define\QQ{\bold Q} \define\RR{\bold R} \define\SS{\bold S} \define\TT{\bold T} \define\UU{\bold U} \define\VV{\bold V} \define\WW{\bold W} \define\ZZ{\bold Z} \define\XX{\bold X} \define\YY{\bold Y} \define\ca{\Cal A} \define\cb{\Cal B} \define\cc{\Cal C} \define\cd{\Cal D} \define\ce{\Cal E} \define\cf{\Cal F} \define\cg{\Cal G} \define\ch{\Cal H} \define\ci{\Cal I} \define\cj{\Cal J} \define\ck{\Cal K} \define\cl{\Cal L} \define\cm{\Cal M} \define\cn{\Cal N} \define\co{\Cal O} \define\cp{\Cal P} \define\cq{\Cal Q} \define\car{\Cal R} \define\cs{\Cal S} \define\ct{\Cal T} \define\cu{\Cal U} \define\cv{\Cal V} \define\cw{\Cal W} \define\cz{\Cal Z} \define\cx{\Cal X} \define\cy{\Cal Y} \define\fa{\frak a} \define\fb{\frak b} \define\fc{\frak c} \define\fd{\frak d} \define\fe{\frak e} \define\ff{\frak f} \define\fg{\frak g} \define\fh{\frak h} \define\fii{\frak i} \define\fj{\frak j} \define\fk{\frak k} \define\fl{\frak l} \define\fm{\frak m} \define\fn{\frak n} \define\fo{\frak o} \define\fp{\frak p} \define\fq{\frak q} \define\fr{\frak r} \define\fs{\frak s} \define\ft{\frak t} \define\fu{\frak u} \define\fv{\frak v} \define\fz{\frak z} \define\fx{\frak x} \define\fy{\frak y} \define\fA{\frak A} \define\fB{\frak B} \define\fC{\frak C} \define\fD{\frak D} \define\fE{\frak E} \define\fF{\frak F} \define\fG{\frak G} \define\fH{\frak H} \define\fJ{\frak J} \define\fK{\frak K} \define\fL{\frak L} \define\fM{\frak M} \define\fN{\frak N} \define\fO{\frak O} \define\fP{\frak P} \define\fQ{\frak Q} \define\fR{\frak R} \define\fS{\frak S} \define\fT{\frak T} \define\fU{\frak U} \define\fV{\frak V} \define\fZ{\frak Z} \define\fX{\frak X} \define\fY{\frak Y} \define\ta{\ti a} \define\tb{\ti b} \define\tc{\ti c} \define\td{\ti d} \define\te{\ti e} \define\tf{\ti f} \define\tg{\ti g} \define\tih{\ti h} \define\tj{\ti j} \define\tk{\ti k} \define\tl{\ti l} \define\tm{\ti m} \define\tn{\ti n} \define\tio{\ti o} \define\tp{\ti p} \define\tq{\ti q} \define\ts{\ti s} \define\tit{\ti t} \define\tu{\ti u} \define\tv{\ti v} \define\tw{\ti w} \define\tz{\ti z} \define\tx{\ti x} \define\ty{\ti y} \define\tA{\ti A} \define\tB{\ti B} \define\tC{\ti C} \define\tD{\ti D} \define\tE{\ti E} \define\tF{\ti F} \define\tG{\ti G} \define\tH{\ti H} \define\tI{\ti I} \define\tJ{\ti J} \define\tK{\ti K} \define\tL{\ti L} \define\tM{\ti M} \define\tN{\ti N} \define\tO{\ti O} \define\tP{\ti P} \define\tQ{\ti Q} \define\tR{\ti R} \define\tS{\ti S} \define\tT{\ti T} \define\tU{\ti U} \define\tV{\ti V} \define\tW{\ti W} \define\tX{\ti X} \define\tY{\ti Y} \define\tZ{\ti Z} \define\tfL{\ti{\fL}} \define\tcl{\ti\cl} \define\tSi{\ti\Si} \define\tis{\ti\si} \define\til{\ti\l} \define\tiz{\ti\z} \define\tix{\ti\x} \define\tid{\ti\d} \define\tss{\ti\ss} \define\ttt{\ti\tt} \define\tce{\ti\ce} \define\tib{\ti\b} \define\sh{\sharp} \define\Mod{\text{\rm Mod}} \define\Ir{\text{\rm Irr}} \define\sps{\supset} \define\app{\asymp} \define\bul{\bullet} \define\che{\check} \define\cha{\che\a} \define\prim{\text{\rm prim}} \define\Bil{\text{\rm Bil}} \define\bcb{\bar{\cb}} \redefine\bcc{\bar{\cc}} \define\gr{\text{\rm gr}} \define\Symp{\text{\rm Symp}} \define\hag{\hat\g} \define\bla{\blacktriangle} \redefine\l{\vartriangle} \define\DLP{DLP} \define\DE{D1} \define\DEII{D2} \define\EN{E} \define\KA{Ka} \define\KO{Ko} \define\LC{L1} \define\LN{L2} \define\LS{LS} \define\MI{M} \define\SH{Sh} \define\SPA{S1} \define\SPAII{S2} \define\SP{Sp} \define\WA{W} \head Introduction\endhead \subhead 0.1\endsubhead Let $\kk$ be an algebraically closed field of characteristic exponent $p\ge1$. Let $G$ be a reductive connected algebraic group over $\kk$. Let $\cu$ be the variety of unipotent elements of $G$. The unipotent classes of $G$ are the orbits of the conjugation action of $G$ on $\cu$. The theory of Dynkin and Kostant \cite{\KO} provides a classification of unipotent classes of $G$ assuming that $p=1$. It is known that this classification remains valid when $p\ge2$ is assumed to be a good prime for $G$. But the analogous classification problem in the case where $p$ is a bad prime for $G$ is more complicated. In every case a classification of unipotent classes is known: see \cite{\WA} for classical groups and \cite{\EN,\SH,\MI} for exceptional groups; but from these works it is difficult to see the general features of the classification. One of the aims of this paper is to present a picture of the unipotent elements which should apply for arbitrary $p$ and is as close as possible to the picture for $p=1$. In 1.4 we observe that the set of unipotent classes in $G$ can be parametrized by a set $\cs^p(\WW)$ of irreducible representations of the Weyl group $\WW$ which can be described apriori purely in terms of the root system. This explains clearly why the classification is different for small $p$. In 1.1 we restate in a more precise form an observation of \cite{\LN} according to which $\cu$ is naturally partitioned into finitely many "unipotent pieces" which are locally closed subvarieties stable under conjugation by $G$; the classification of unipotent pieces is independent of $p$. For $p=1$ or a good prime, each unipotent piece is a single conjugacy class. When $p$ is a bad prime a unipotent piece is in general a union of several conjugacy classes. Also each unipotent piece has some topological properties which are independent of $p$ (for example, over a finite field, the number of points of a unipotent piece is given by a formula independent of the characteristic). Another aim of this paper is the study of $\cb_u$, the variety of Borel subgroups of $G$ containing a unipotent element $u$. It is known \cite{\SP} that when $p$ is a good prime, the $l$-adic cohomology spaces of $\cb_u$ are pure. We would like to prove a similar result in the case where $p$ is a bad prime. We present a method by which this can be achieved in a number of cases. Our strategy is to extend a technique from \cite{\DLP} in which (assuming that $p=1$), $\cb_u$ is analyzed by first partitioning it into finitely many smooth locally closed subvarieties using relative position of a point in $\cb_u$ with a canonical parabolic attached to $u$. Much of our effort is concerned with trying to eliminate reference to the linearization procedure of Bass-Haboush (available only for $p=1$) which was used in an essential way in \cite{\DLP}. Our approach is based on a list of properties $\fP_1-\fP_8$ of unipotent elements of which the first five (resp. last three) are expected to hold in general (resp. in many cases). All these properties are verified for general linear and symplectic groups (any $p$) in \S2, \S3. In writing \S3 (on symplectic groups mostly with $p=2$) I found that the treatment in \cite{\WA} is not sufficient for this paper's purposes; I therefore included a treatment which does not rely on \cite{\WA}. {\it Notation.} When $p>1$ we denote by $\kk_p$ an algebraic closure of the field with $p$ elements. Let $\cb$ the variety of Borel subgroups of $G$. If $\G'$ is a subgroup of a group $\G$ and $x\in\G$ let $Z_{G'}(x)=\{z\in\G';zx=xz\}$. For a finite set $Z$ let $|Z|$ be the cardinal of $Z$. Let $l$ be a prime number invertible in $\kk$. For $a,b\in\ZZ$ let $[a,b]=\{z\in\ZZ;a\le z\le b\}$. \head Contents\endhead 1. Some properties of unipotent elements. 2. General linear groups. 3. Symplectic groups. 4. The group $A^1(u)$. 5. Study of the varieties $\cb_u$. \head 1. Some properties of unipotent elements\endhead \subhead 1.1\endsubhead $G$ acts naturally, by conjugation, on $\Hom(\kk^*,G)$ (homomorphisms of algebraic groups). The set of orbits $\Hom(\kk^*,G)/G$ is naturally in bijection with the analogous set $\Hom(\CC^*,G')/G'$ where $G'$ is a connected reductive group over $\CC$ of the same type as $G$. (Both sets may be identified with the set of Weyl group orbits on the group of $1$-parameter subgroups of some maximal torus.) Let $\tD_{G'}$ be the set of all $\o\in\Hom(\CC^*,G')$ such that there exists a homomorphism of algebraic groups $\ti\o:SL_2(\CC)@>>>G'$ with $\ti\o\left(\sm t&0\\0&t\i\esm\right)=\o(t)$ for all $t\in\CC^*$. Now $\tD_{G'}$ is $G'$-stable; it has been described explicitly by Dynkin. Let $\tD_G$ be the unique $G$-stable subset of $\Hom(\kk^*,G)$ whose image in $\Hom(\kk^*,G)/G$ corresponds under the bijection $\Hom(\kk^*,G)/G\lra\Hom(\CC^*,G')/G'$ (as above) to the image of $\tD_{G'}$ in $\Hom(\CC^*,G')/G'$. Let $D_G$ be the set of sequences $\l=(G^\l_0\sps G^\l_1\sps G^\l_2\sps\do)$ of closed connected subgroups of $G$ such that for some $\o\in\tD_G$ we have (for $n\ge0$): $\Lie G^\l_n=\{x\in\Lie G;\lim_{t\in\kk^*;t\to 0}t^{1-n}\Ad\o(t)x=0\}$. \nl Now $G$ acts on $D_G$ by conjugation and the obvious map $\tD_G@>>>D_G$ induces a bijection $\tD_G/G@>\si>>D_G/G$ on the set of orbits. If $\l\in D_G$ and $g\in G$ then $G^{g\l g\i}_n=gG^\l_ng\i$ for $n\ge0$. Moreover, $G^\l_0$ is a parabolic subgroup of $G$ with unipotent radical $G^\l_1$ and $G^\l_n$ is normalized by $G^\l_0$ for any $n$. Moreover, (a) $G^\l_2/G^\l_3$ is a commutative connected unipotent group; (b) the conjugation action of $G^\l_0$ on $G^\l_2/G^\l_3$ factors through an action of $\bG^\l_0:=G^\l_0/G^\l_1$ on $G^\l_2/G^\l_3$. \nl Note also that $G^\l_n$ for $n\ne0,2$ are uniquely determined by $G^\l_0,G^\l_2$. Let $\bla$ be a $G$-orbit in $D_G$. Then $\tH^\bla:=\cup_{\l\in\bla}G^\l_2$ is a closed irreducible subset of $\cu$ (since for $\l\in\bla$, $G^\l_2$ is a closed irreducible subset of $\cu$ stable under conjugation by $G^\l_0$ and $G/G^\l_0$ is projective). Let $$H^\bla=\tH^\bla-\cup_{\bla'\in D_G/G;\tH^{\bla'}\subsetneqq\tH^\bla}\tH^{\bla'}.$$ For $\l\in D_G$ let $X^\l=G^\l_2\cap H^\bla$ where $\bla$ is the $G$-orbit of $\l$. Then $H^\bla$ is an open dense subset of $\tH^\bla$ stable under conjugation by $G$ and $X^\l$ is an open dense subset of $G^\l_2$ stable under conjugation by $G^\l_0$. (We use that $D_G/G$ is finite.) Hence $H^\bla$ is locally closed in $\cu$. The subsets $H^\bla(\bla\in D_G/G)$ are called the {\it unipotent pieces} of $G$. We state the following properties $\fP_1-\fP_5$. $\fP_1$. {\it The sets $X^\l(\l\in D_G)$ form a partition of $\cu$.} $\fP_2$. {\it Let $\bla\in D_G/G$. The sets $X^\l(\l\in\bla)$ form a partition of $H^\bla$. More precisely, $H^\bla$ is a fibration over $\bla$ with smooth fibres isomorphic to $X^\l$ ($\l\in\bla$); in particular, $H^\bla$ is smooth.} $\fP_3$. {\it The locally closed subets $H^\bla(\bla\in D_G/G)$ form a (finite) partition of $\cu$.} $\fP_4$. {\it Let $\l\in D_G$. We have $G^\l_3X^\l=X^\l G^\l_3=X^\l$.} $\fP_5$. {\it Assume that $\kk=\kk_p$. Let $F:G@>>>G$ be the Frobenius map corresponding to a split $\FF_q$-rational structure with $q-1$ sufficiently divisible. Let $\l\in D_G$ be such that $F(G^\l_n)=G^\l_n$ for all $n\ge0$ and let $\bla$ be the $G$-orbit of $\l$. Then $|H^\bla(\FF_q)|,|X^\l(\FF_q)|$ are polynomials in $q$ with integer coefficients independent of $p$.} Assume first that $p=1$ or $p\gg0$. By the theory of Dynkin-Kostant, for $\l\in D_G$ there is a unique open $G^\l_0$-orbit $X'{}^\l$ in $G^\l_2$; we then have a bijection of $D_G/G$ with the set of unipotent classes on $G$ which to the $G$-orbit $\bla$ of $\l\in D_G$ associates the unique unipotent class $H'{}^\bla$ of $G$ that contains $X'{}^\l$; moreover, if $g\in X'{}^\l$ then $Z_{G^\l_0}(g)=Z_G(g)$. As stated by Kawanaka \cite{\KA}, the same holds when $p$ is a good prime of $G$ (but his argument is rather sketchy). To show that $\fP_1-\fP_3$ holds when $p$ is a good prime it then suffices to show that $X^\l=X'{}^\l$ for any $\l$. It also suffices to show that $X'{}^\l=G^\l_2\cap H'{}^\bla$ for $\l\in\bla$ as above. (Assume that $g\in G^\l_2\cap H'{}^\bla,g\n X'{}^\l$. Let $g'\in X'{}^\l$. By the definition of $X'{}^\l$ and the irreducibility of $G^\l_2$, the dimension of the $G^\l_0$-orbit of $g$ is strictly smaller than the dimension of the $G^\l_0$-orbit of $g'$. Hence $\dim Z_{G^\l_0}(g)>\dim Z_{G^\l_0}(g')$. We have $\dim Z_G(g)\ge\dim Z_{G^\l_0}(g)$, $\dim Z_{G^\l_0}(g')=\dim Z_G(g')$ hence $\dim Z_G(g)>\dim Z_G(g')$. This contradicts the fact that $g,g'$ are $G$-conjugate.) In this case we have $H^\bla=H'{}^\bla$ and $\tH^\bla$ is the closure of $H'{}^\bla$. We expect that $\fP_1-\fP_5$ hold in general. In the case where $G=GL_n(\kk)$ (any $p$) the validity of $\fP_1-\fP_5$ follows from 2.9. In the case where $G$ is a symplectic group (any $p$) the validity of $\fP_1-\fP_5$ follows from 3.13, 3.14. If $G$ is of type $E_n$ (any $p$) then one can deduce $\fP_1-\fP_5$ from the various lemmas in \cite{\MI}, or rather from the extensive computations (largely omitted) on which those lemmas are based; it would therefore be desirable to have an independent verification of these properties. The case of special orthogonal groups will be considered elsewhere. \mpb We note the following consequence of $\fP_1$. (c) {\it If $\l\in D_G$ and $u\in X^\l$ then $Z_G(u)\sub G^\l_0$.} \nl Let $g\in G$. Then $gug\i\in X^{g\l}$. Hence if $g\in Z_G(u)$ we have $u\in X^{g\l}$. Thus, $X^{g\l}\cap X^\l\ne\em$. From $\fP_1$ we see that $g\!\l=\l$. In particular $gG^\l_0g\i=G^\l_0$ and $g\in G^\l_0$, as required. \subhead 1.2\endsubhead Let $\l\in D_G$. We assume that $\fP_1-\fP_4$ hold for $\l$. Let $\p^\l:G^\l_2@>>>G^\l_2/G^\l_3$ be the obvious homomorphism. By $\fP_4$ we have $X^\l=(\p^\l)\i(\bX^\l)$ where $\bX^\l$ is a well defined open dense subset of $G^\l_2/G^\l_3$ stable under the action of $\bG^\l_0$. We wish to consider some properties of the sets $\bX^\l$ which may or may not hold for $G$. $\fP_6$. {\it If $u\in X^\l$ then $uG^\l_3=G^\l_3u$ is contained in the $G^\l_0$-conjugacy class of $u$. Hence $\g\m(\p^\l)\i(\g)$ is a bijection between the set of $\bG^\l_0$-orbits in $\bX^\l$ and the set of $G^\l_0$-conjugacy classes in $X^\l$.} $\fP_7$. {\it Let $\g$ be a $\bG^\l_0$-orbit in $\bX^\l$. Let $\hag$ be the union of all $\bG^\l_0$-orbits in $\bX^\l$ whose closure contains $\g$. Thus, $\hag$ is an open subset of $\bX^\l$ and $\g$ is a closed subset of $\hag$. There exists a variety $\g_1$ and a morphism $\r:\hag@>>>\g_1$ such that the restriction of $\r$ to $\g$ is a finite bijective morphism $\s:\g@>>>\g_1$ and the map of sets $\s\i\r:\hag@>>>\g$ is compatible with the actions of $\bG^\l_0$.} $\fP_8$. {\it There exists a finite set $I$ and a bijection $J\m\Ph_J$ between the set of subsets of $I$ and the set of $G^\l_0$-orbits in $X^\l$ such that for any $J\sub I$, the closure of $\Ph_J$ in $X^\l$ is $\cup_{J';J\sub J'}\Ph_{J'}$. Moreover, if $\kk,q$ are as in $\fP_5$ then there exists a function $I@>>>\{2,4,6,\do\},i\m c_i$ such that $|\Ph_J(\FF_q)|=\prod_{i\in J}(q^{c_i}-1)|\Ph_\em(\FF_q)|$ for any $J\sub I$.} When $p=1$ or $p\gg0$ property $\fP_6$ can be deduced from the theory of Dynkin-Kostant; properties $\fP_7,\fP_8$ are trivial. In the case where $G=GL_n(\kk)$ (any $p$) the validity of $\fP_6$ follows from 2.9; properties $\fP_7,\fP_8$ are trivial. In the case where $G$ is a symplectic group (any $p$) the validity of $\fP_6-\fP_8$ follows from 3.14. $\fP_6$ is false for $G$ of type $G_2$, $p=3$. \subhead 1.3\endsubhead Let $\VV$ be a finite dimensional $\QQ$-vector space. Let $R\sub\VV^*=\Hom(\VV,\QQ)$ be a (reduced) root system, let $\che R\sub\VV$ be the corresponding set of coroots and let $\WW\sub GL(\VV)$ be the Weyl group of $R$. Let $\b\lra\che\b$ be the canonical bijection $R\lra\che R$. Let $\Pi$ be a set of simple roots for $R$ and let $\che\Pi=\{\che\a;\a\in\Pi\}$. Let $\Th=\{\b\in R;\b-\a\n R\qua\frl\a\in\Pi\}$, $\ti\Th=\{\b\in R;\che\b-\che\a\n\che R\qua\frl\a\in\Pi\}$, $\ti\ca=\{J\sub\Pi\cup\ti\Th;J\text{ linearly independent in }\VV^*\}$. \nl For any prime number $r$ let $\ca_r$ be the set of all $J\sub\Pi\cup\Th$ such that $J$ is linearly independent in $\VV^*$ and the torsion subgroup of $\sum_{\a\in\Pi}\ZZ\a/\sum_{\b\in J}\ZZ\b$ has order $r^k$ for some $k\in\NN$. For any $J\in\ca_r$ or $J\in\ti\ca$ let $\WW_J$ be the subgroup of $\WW$ generated by the reflections with respects to roots in $J$. For $W'=\WW$ or $\WW_J$ let $\Ir(W')$ be the set of (isomorphism classes) of irreducible representations of $W'$ over $\QQ$. For $E\in\Ir(W')$ let $b_E$ be the smallest integer $\ge0$ such that $E$ appears with non-zero multiplicity in the $b_E$-th symmetric power of $\VV$ regarded as a $W'$-module; if this multiplicity is $1$ we say that $E$ is good. If $J$ is as above and $E\in\Ir(W')$ is good then there is a unique $\tE\in\Ir(\WW)$ such that $\tE$ appears in $\Ind_{\WW_J}^\WW E$ and $b_{\tE}=b_E$; moreover, $\tE$ is good. We set $\tE=j_{\WW_J}^\WW E$. Let $\cs_\WW\sub\Ir(\WW)$ be the set of special representations of $\WW$ (see \cite{\LC}). Now any $E\in\cs_\WW$ is good. Following \cite{\LC}, let $\cs^1_\WW$ be the set of all $E\in\Ir(\WW)$ such that $E=j_{\WW_J}^\WW E_1$ for some $J\in\ti\ca$ and some $E_1\in\cs_{\WW_J}$. (Note that $\WW_J$ is like $\WW$ with the same $\VV$ and with $R$ replaced by the root system with $J$ as set of simple roots; hence $\cs_{\WW_J}$ is defined.) Now any $E\in\cs^1_\WW$ is good. For any prime number $r$ let $\cs^r_\WW$ be the set of all $E\in\Ir(\WW)$ such that $E=j_{\WW_K}^\WW E_1$ for some $K\in\ca_r$ and some $E_1\in\cs^1_{\WW_K}$. (Note that $\WW_K$ is like $\WW$ with the same $\VV$ and with $R$ replaced by the root system with $K$ as set of simple roots; hence $\cs^1_{\WW_K}$ is defined.) Now any $E\in\cs^1_\WW$ is good. We have $\cs^1(\WW)\sub\cs^r(\WW)$. We have $\cs^1(\WW)=\cs^r(\WW)$ if $r$ is a good prime for $\WW$ and also in the following cases $\WW$ of type $G_2,r=2$; $\WW$ of type $F_4$, $r=3$; $\WW$ of type $E_6$; $\WW$ of type $E_7$, $r=3$; $\WW$ of type $E_8$, $r=5$. If $\WW$ is of type $G_2$ and $r=3$ then $\cs^r(\WW)-\cs^1(\WW)$ consists of a single representation of dimension $1$ coming under $j_{\WW_J}^\WW$ from a $\WW_J$ of type $A_2$. If $\WW$ is of type $F_4$ and $r=2$ then $\cs^r(\WW)-\cs^1(\WW)$ consists of four representations of dimensions $9/4/4/2$ coming under $j_{\WW_J}^\WW$ from a $\WW_J$ of type $C_3A_1/C_3A_1/B_4/B_4$. If $\WW$ is of type $E_7$ and $r=2$ then $\cs^r(\WW)-\cs^1(\WW)$ consists of a single representation of dimensions $84$ coming under $j_{\WW_J}^\WW$ from a $\WW_J$ of type $D_6A_1$. If $\WW$ is of type $E_8$ and $r=2$ then $\cs^r(\WW)-\cs^1(\WW)$ consists of four representations of dimensions $1050/840/168/972$ coming under $j_{\WW_J}^\WW$ from a $\WW_J$ of type $E_7A_1/D_5A_3/D_8/E_7A_1$. If $\WW$ is of type $E_8$ and $r=3$ then $\cs^r(\WW)-\cs^1(\WW)$ consists of a single representation of dimensions $175$ coming under $j_{\WW_J}^\WW$ from a $\WW_J$ of type $E_6A_2$. \subhead 1.4\endsubhead Let $\WW$ be the Weyl group of $G$. Let $u$ be a unipotent element in $G$. Springer's correspondence (generalized to arbitrary characteristic) associates to $u$ and the trivial representation of $Z_G(u)/Z_G(u)^0$ a representation $\r_u\in\Ir(\WW)$. Moreover $u\m\r_u$ defines an injective map from the set of unipotent classes in $G$ to $\Ir(\WW)$. Let $\cx^p(\WW)$ be the image of this map ($p$ as in 0.1). We state: (a) {\it If $p=1$ we have $\cx^1(\WW)=\cs^1(\WW)$.} (See \cite{\LC}). (b) {\it If $p>1$ we have $\cx^p(\WW)=\cs^p(\WW)$.} \nl The proof of (b) follows from the explicit description of the Springer correspondence for small $p$ given in \cite{\LS},\cite{\SPAII}. \head 2. General linear groups\endhead \subhead 2.1\endsubhead Let $\bcc$ be the category whose objects are $\ZZ$-graded $\kk$-vector spaces $\bV=\op_{a\in\ZZ}\bV_a$ such that $\dim\bV<\iy$; the morphisms are linear maps respecting the grading. Let $\bV\in\bcc$. For $j\in\ZZ$ let $\End_j(\bV)=\{T\in\Hom(\bV,\bV);T(\bV_a)\sub\bV_{a+j}\qua\frl a\}$. Let $\End_2^0(\bV)$ be the set of all $\nu\in\End_2(\bV)$ that satisfy the {\it Lefschetz condition}: $\nu^n:\bV_{-n}@>>>\bV_n$ is an isomorphism for any $n\ge0$. Let $\nu\in\End_2^0(\bV)$. Define a graded subspace $P^\nu=\bV^{\prim}$ of $\bV$ by $P^\nu_a=\{x\in\bV_a;\nu^{1-a}x=0\}$ for $a\le0$, $P^\nu_a=0$ for $a>0$. A standard argument shows that $N^{(a-c)/2}:P^\nu_c@>>>\bV_a$ is injective if $c\in a+2\ZZ,c\le a\le-c$ and we have (a) $\op_{c\in a+2\ZZ;c\le a\le-c}P^\nu_c@>\si>>\bV_a, (z_c)\m\sum_{c\in a+2\ZZ;c\le a\le-c}N^{(a-c)/2}z_c$. \nl We show: (b) {\it Let $j\in\NN$, $R\in\End_{j+2}(\bV)$. Then $R=T\nu-\nu T$ for some $T\in\End_j(\bV)$.} \nl Let $c\le0$. Since $\nu^{1-c}:\bV_{j-c}@>>>\bV_{j+c+2}$ is surjective, the induced map $\Hom(P^\nu_{-c},\bV_{j-c})@>>>\Hom(P^\nu_{-c},\bV_{j+c+2})$ \nl is surjective. Hence there exists $\t_c\in\Hom(P^\nu_{-c},\bV_{j-c})$ such that $\nu^{1-c}\t_c=-\sum_{i+i'=-c}\nu^iR\nu^{i'}$. \nl For $k\in[0,-c]$ we define $\t_{c,k}\in\Hom(P^\nu_c,\bV_{c+2k+j})$ by $\t_{c,0}=\t_c$ and $\t_{c,k}=\nu\t_{c,k-1}+R\nu^{k-1}$ for $k\in[1,-c]$. Then $\nu\t_{c,-c}+R\nu^{-c}=0$. Let $T:\bV@>>>\bV$ be the unique linear map such that $T(\nu^kx)=\t_{c,k}(x)$ for $x\in P^\nu_c,c\le0,k\in[0,-c]$. This $T$ has the required property. \subhead 2.2\endsubhead Let $\cc$ be the category whose objects are $\kk$-vector spaces of finite dimension; morphisms are linear maps. Let $V\in\cc$. A collection of subspaces $V_*=(V_{\ge a})_{a\in\ZZ}$ of $V$ is said to be a {\it filtration} of $V$ if $V_{\ge a+1}\sub V_{\ge a}$ for all $a$, $V_{\ge a}=0$ for some $a$, $V_{\ge a}=V$ for some $a$. We say that $V$ is {\it filtered} if a filtration $V_*$ of $V$ is given. Assume that this is the case. We set $\gr V_*=\op_{a\in\ZZ}\gr_aV_*\in\bcc$ where $\gr_aV_*=V_{\ge a}/V_{\ge a+1}$. For any $j\in\ZZ$ let $E_{\ge j}V_*=\{T\in\End(V);T(V_{\ge a})\sub V_{\ge a+j}\qua\frl a\}$. Any such $T$ induces a linear map $\bT\in\End_j(\gr V_*)$. \subhead 2.3\endsubhead Let $V\in\cc$. Let $\Nil(V)=\{T\in\End(V);T\text{ nilpotent }\}$. Let $N\in\Nil(V)$. When $p=1$, the Dynkin-Kostant theory associates to $1+N$ a canonical filtration $V^N_*$ of $V$; in terms of a basis of $V$ of the form (a) $\{N^kv_r;r\in[1,t],k\in[0,e_r-1]\}$ with $v_r\in V,e_r\ge1,N^{e_r}v_r=0$ for $r\in[1,t]$, \nl $V^N_{\ge a}$ is the subspace spanned by $\{N^kv_r;r\in[1,t],k\in[0,e_r-1],2k+1\ge e_r+a\}$. This subspace makes sense for any $p$ and we denote it in general by $V^N_{\ge a}$; it is independent of the choice of basis: we have $V^N_{\ge a}=\sum_{j\ge\max(0,a)}N^j(\ker N^{2j-a+1})$. \nl The subspaces $V^N_{\ge a}$ form a filtration $V^N_*$ of $V$; thus, $V$ becomes a filtered vector space. From the definitions we see that (b) {\it$N\in E_{\ge2}V^N_*$ and $\bN\in\End_2(\gr V_*^N)$ belongs to $\End_2^0(\gr V_*^N)$.} \nl Note that for any $j\ge1$, (c) {\it$\dim P^{\bN}_{1-j}$ is the number of Jordan blocks of size $j$ of $N:V@>>>V$.} \nl From 2.1(a) we deduce that for any $n\ge0$: (d) $\dim P^{\bN}_{-n}=\dim\gr_{-n}V^N_*-\dim\gr_{-n-2}V^N_*$. \subhead 2.4\endsubhead According to \cite{\DEII, 1.6.1}, (a) {\it if $V_*$ is a filtration of $V$ and $N\in E_{\ge2}V_*$ induces an element $\nu\in\End_2^0(\gr V_*)$ then $V_*=V_*^N$.} \nl We show that $V_{\ge a}=V^N_{\ge a}$ for all $a$. Let $e$ be the smallest integer $\ge0$ such that $N^e=0$. We argue by induction on $e$. If $a\ge e$ then $\nu^a:\gr_{-a}V_*@>>>\gr_aV_*$ is both $0$ and an isomorphism hence $V_{\ge-a}=V_{\ge1-a}$ and $V_{\ge a}=V_{\ge a+1}$. Thus, $V_{\ge e}=V_{\ge e+1}=\do=0$ and $V_{\ge1-e}=V_{\ge-e}=\do=V$. Similarly, $V^N_{\ge e}=V^N_{\ge e+1}=\do=0$ and $V^N_{\ge1-e}=V^N_{\ge-e}=\do=V$. Hence $V_{\ge a}=V^N_{\ge a}$ if $a\ge e$ or if $a\le1-e$. This already suffices in the case where $e\le1$. Thus we may assume that $e\ge2$. Now $\nu^{e-1}:\gr_{1-e}V_*@>>>\gr_{e-1}V_*$ is an isomorphism that is, $N^{e-1}:V/V_{\ge2-e}@>>>V_{\ge e-1}$ is an isomorphism. We see that $V_{\ge e-1}=N^{e-1}V$ and $V_{\ge2-e}=\ker(N^{e-1})$. Hence if $2-e\le a\le e-1$ we have $N^{e-1}V\sub V_{\ge a}\sub\ker(N^{e-1})$; let $V'_{\ge a}$ be the image of $V_{\ge a}$ under the obvious map $\r:\ker(N^{e-1})@>>>V':=\ker(N^{e-1})/N^{e-1}V$. For $a\le1-e$ we set $V'_{\ge a}=V'$ and for $a\ge e$ we set $V'_{\ge a}=0$. Now $(V'_{\ge a})_{a\in\ZZ}$ is a filtration of $V'$ satisfying a property like (a) (with $N$ replaced by the map $N':V'@>>>V'$ induced by $N$). Since $N'{}^{e-1}=0$, the induction hypothesis applies to $N'$; it shows that $V'_{\ge a}=V'{}^{N'}_{\ge a}$ for all $a$. Since for $2-e\le a\le e-1$, $V_{\ge a}=\r\i(V'_{\ge a})$, it follows that $V_{\ge a}=\r\i(V'{}^{N'}_{\ge a})$; similarly, $V^N_{\ge a}=\r\i(V'{}^{N'}_{\ge a})$ hence $V_{\ge a}=V^N_{\ge a}$. This completes the proof. With notation in the proof above we have: $V^N_{\ge a}=0$ for $a\ge e$, $V^N_{\ge a}=V$ for $a\le1-e$, $V^N_{\ge a}=\r\i(V'{}^{N'}_{\ge a}),V'{}^{N'}_{\ge a}=\r(V^N_{\ge a})$ for $e\ge2$ and $2-e\le a\le e-1$, $V^N_{\ge e-1}=N^{e-1}V$ if $e\ge1$, $V^N_{\ge2-e}=\ker(N^{e-1})$ if $e\ge1$. We have $\gr_aV^N_*=0$ for $a\ge e$ and for $a\le-e$. Note also that the proof above provides an alternative (inductive) definition of $V^N_{\ge a}$ which does not use a choice of basis. \subhead 2.5\endsubhead Let $V,N$ be as in 2.3. Let $V_*=V^N_*$. Let $\nu=\bN\in\End_2(\gr V_*)$. We can find a grading $V=\op_{a\in\ZZ}V_a$ of $V$ such that (a) $NV_a\sub V_{a+2}$ and $V_{\ge a}=V_a\op V_{a+1}\op\do$ for all $a$. \nl For example, in terms of a basis of $V$ as in 2.3(a), we can take $V_a$ to be the subspace spanned by $\{N^kv_r;r\in[1,t],k\in[0,e_r-1],2k+1=e_r+a\}$. Taking direct sum of the obvious isomorphisms $V_a@>\si>>\gr_aV_*$ we obtain an isomorphism of graded vector spaces $V@>\si>>\gr V_*$ under which $N$ corresponds to $\nu$. It follows that (b) $N\in\End_2^0(V)$ (defined in terms of the grading $\op_aV_a$). \nl We note the following result. (c) {\it Let $n\ge0$ and let $x\in P^\nu_{-n}$. There exists a representative $\dx$ of $x$ in $V_{\ge-n}$ such that $N^{n+1}\dx=0$.} \nl Let $V_a$ be as above. There is a unique representative $\dx$ of $x$ in $V_{\ge-n}$ such that $\dx\in V_{-n}$. We have $N^{n+1}\dx\in V_n$ and the image of $N^{n+1}\dx$ under the canonical isomorphism $V_n@>\si>>\gr_nV_*^N$ is $0$; hence $N^{n+1}\dx=0$. Let $E_{\ge1}^NV_*=\{S\in E_{\ge1}V_*;SN=NS\}$, $\End_1^\nu(\gr V_*)=\{\s\in\End_1(\gr V_*),\s\nu=\nu\s\}$. We show: (d) {\it The obvious map $E_{\ge1}^NV_*@>>>\End_1^\nu(\gr V_*),S\m\bS$ is surjective.} \nl Let $\s\in\End_1^\nu(\gr V_*)$. Let $V_a$ be as above. In terms of these $V_a$ we define $V@>\si>>\gr V_*$ as above. Under this isomorphism, $\s$ corresponds to a linear map $S:V@>>>V$. Clearly, $S\in E_{\ge1}^NV_*$ and $\bS=\s$. \subhead 2.6\endsubhead Let $V,N$ be as in 2.3. Let $V_*=V_*^N$. Now $1+E_{\ge1}V_*$ is a subgroup of $GL(V)$ acting on $N+E_{\ge3}V_*$ by conjugation. We show that (a) {\it the conjugation action of $1+E_{\ge1}V_*$ on $N+E_{\ge3}V_*$ is transitive.} \nl We must show: if $S\in E_{\ge3}V_*$ then there exists $T\in E_{\ge1}V_*$ such that $(1+T)N=(N+S)(1+T)$ that is, $TN-NT=S+ST$. We fix subspaces $V_a$ as in 2.5. We have $S=\sum_{j\ge3}S_j$ where $S_j\in\End(V)$ satisfy $S_jV_a\sub V_{a+j}$ for all $a$. We seek a linear map $T=\sum_{j\ge1}T_j$ where $T_j\in\End(V)$ satisfy $T_jV_a\sub V_{a+j}$ for all $a$ and $\sum_{j\ge1}(T_jN-NT_j)=\sum_{j\ge3}S_j+\sum_{j'\ge3,j''\ge1}S_{j'}T_{j''}$ that is, $(*)$ $T_jN-NT_j=S_{j+2}+\sum_{j'\in[1,j-1]}S_{j+2-j'}T_{j'}$ for $j=1,2,\do$. \nl We show that this system of equations in $T_j$ has a solution. We take $T_1=0$. Assume that $T_j$ has been found for $j<j_0$ for some $j_0\ge2$ so that $(*)$ holds for $j<j_0$. We set $R=S_{j_0+2}+\sum_{j'\in[1,j_0-1]}S_{j+2-j'}T_{j'}$. Then $R(V_a)\sub V_{a+j_0+2}$ for any $a$. The equation $T_{j_0}N-NT_{j_0}=R$ can be solved by 2.1(b) (see 2.5(b)). This shows by induction that the system $(*)$ has a solution. (a) is proved. We now show: (b) {\it if $\tN\in N+E_{\ge3}V_*$ then $V^{\tN}_*=V_*$.} \nl Indeed by (a) we can find $u\in1+E_{\ge1}V_*$ such that $\tN=uNu\i$. Since $V_*^N$ is canonically attached to $N$, we have $V^{uNu\i}_{\ge a}=u(V^N_{\ge a})=V^N_{\ge a}$ and (b) follows. For example, (c) {\it if $\tN=c_1N+c_2N^2+\do+c_kN^k$ where $c_i\in\kk,c_1\ne0$ then $V^{\tN}_*=V^N_*$.} \nl We may assume that $c_1=1$. Since $c_2N^2+\do+c_kN^k\in E_{\ge4}V_*\sub E_{\ge3}V_*$, (b) is applicable and (c) follows. \subhead 2.7\endsubhead Let $V,N$ be as in 2.3. Let $V_*=V_*^N$. Let $\nu=\bN\in\End_2(\gr V_*)$. Let $r\ge2$ be such that $N^r=0$ on $V$. Let $W$ be an $N$-stable subspace of $V$ such that there exists an $N$-stable complement of $W$ in $V$, $N:W@>>>W$ has no Jordan block of size $\ne r,r-1$ and $N^{r-2}=0$ on $V/W$. Then $W_*=W_*^N$ is defined. Define a linear map $\mu:\gr V_*@>>>\gr W_*$ as follows. Let $x\in\gr_aV_*$. We have uniquely $x=\sum_{c\in a+2\ZZ;c\le a\le-c}\nu^{(a-c)/2}x_c$ where $x_c\in P^\nu_c$; we set $$\mu(x)=\sum_{c\in a+2\ZZ;c\le a\le-c,c=1-r\text{ or }2-r}\nu^{(a-c)/2}x_c.$$ Let $\cx$ be the set of $N$-stable complements of $W$ in $V$. Then $\cx\ne\em$. For $Z\in\cx$ define $\Pi_Z:V@>>>W$ by $\Pi_Z(w+z)=w$ where $w\in W,z\in Z$. Let $\baP_Z:\gr V_*@>>>\gr W_*$ be the map induced by $\Pi_Z$. We show that (a) $\Pi_Z(V_{\ge a})\sub W_{\ge a}$ for all $a$ and $\baP_Z=\mu$. \nl We have $V_{\ge a}=W_{\ge a}\op Z_{\ge a}$. If $x\in V_{\ge a},x=w+z,w\in W_{\ge a}, z\in Z_{\ge a}$, then $\Pi_Z(x)=w$. Thus $\Pi_Z(V_{\ge a})\sub W_{\ge a}$. We can find direct sum decompositions $W=\op_mW_m,Z=\op_mZ_m$ such that $NW_m\sub W_{m+2}$, $NZ_m\sub Z_{m+2}$ and $N^m:W_{-m}@>\si>>W_m$, $N^m:Z_{-m}@>\si>>Z_m$ for $m\ge0$ (see 2.5). Let $V_a=W_a\op Z_a$. Define $V_a^{\prim},W_a^{\prim},Z_a^{\prim}$ as in 2.1 in terms of $N$. We have $V_a^{\prim}=W_a^{\prim}\op Z_a^{\prim}$. We must show that $\baP_Z(x)=\mu(x)$ for $x\in\gr_aV_*$. It suffices to show: if $w\in W_a,z\in Z_a$ and $w+z=\sum_{c\in a+2\ZZ;c\le a\le-c}\nu^{(a-c)/2}x_c$ where $x_c\in V_c^{\prim}$ then $w=\sum_{c\in a+2\ZZ;c\le a\le-c,1-r\le c\le2-r}\nu^{(a-c)/2}x_c$. We have $x_c=w_c+z_c$ where $w_c\in W_c^{\prim},z_c\in Z_c^{\prim}$ and $w=\sum_{c\in a+2\ZZ;c\le a\le-c}\nu^{(a-c)/2}w_c$. Now if $W_c^{\prim}\ne0$ then $1-r\le c\le2-r$. Hence $w=\sum_{c\in a+2\ZZ;c\le a\le-c,1-r\le c\le2-r}\nu^{(a-c)/2}w_c$. Also, $Z_{1-r}^{\prim}=Z_{2-r}^{\prim}=0$ since $N:Z@>>>Z$ has no Jordan blocks of size $\ge r-1$. Thus if $c\in a+2\ZZ;c\le a\le-c,1-r\le c\le2-r$ then $z_c=0$ and $x_c=w_c$. Thus $w=\sum_{c\in a+2\ZZ;c\le a\le-c,1-r\le c\le2-r}\nu^{(a-c)/2}x_c$, as required. Let $Z,Z'\in\cx$. By the previous argument, $\Pi_Z,\Pi_{Z'}:V@>>>W$ both map $V_{\ge a}$ into $W_{\ge a}$ and induce the same map $\gr V_*@>>>\gr W_*$. It follows that $\Pi_Z-\Pi_{Z'}:V@>>>W$ maps $V_{\ge a}$ into $W_{\ge a+1}$. In other words, (b) {\it if $x\in V_{\ge a}$ and $x=w+z=w'+z'$ where $w,w'\in W,z\in Z,z'\in Z'$, then $w-w'\in W_{\ge a+1}$.} \nl Define $\Ph\in GL(V)$ by $\Ph(x)=x$ for $x\in W$, $\Ph(x)=x'$ for $x\in Z$ where $x'\in Z'$ is given by $x-x'\in W$. We show: (c) $(1-\Ph)V_{\ge a}\sub V_{\ge a+1}$ {\it for any $a$.} \nl Let $x\in V_{\ge a}$. We have $x=w+z=w'+z'$ where $w,w'\in W,z\in Z,z'\in Z'$. We have $\Ph(x)=w+z'$ hence $(1-\Ph)(x)=(w+z)-(w+z')=z-z'=w'-w$ and this belongs to $W_{\ge a+1}$ by (b). We show: (d) $\Ph N=N\Ph$. \nl Indeed, for $x=x_1+x_2,x_1\in W,x_2\in Z$ we have $Nx=Nx_1+Nx_2$ with $Nx_1\in W$, $Nx_2\in Z$ and $x_2-x'_2\in W$ with $x'_2\in Z'$. We have $Nx_2-Nx'_2\in W$ with $Nx_2\in Z$, $Nx'_2\in Z'$. Hence $\Ph(Nx)=Nx_1+Nx'_2=N(x_1+x'_2)=N\Ph(x)$, as required. \subhead 2.8\endsubhead Let $V,N$ be as in 2.3. Let $r\ge1$ be such that $N^r=0$. A subspace $W$ of $V$ is said to be {\it $r$-special} if $NW\sub W$, $N:W@>>>W$ has no Jordan blocks of size $\ne r$ and $N^{r-1}=0$ on $N/W$. We show: (a) {\it If $W,W'$ are $r$-special subspaces then there exists a subspace $X$ of $V$ such that $NX\sub X,W\op X=V,W'\op X=V$.} \nl We argue by induction on $r$. If $r=1$ the result is obvious; we have $W=W'=V$. Assume that $r\ge2$. Let $V'=\ker N^{r-1},V''=\ker N^{r-2}$. Let $E\sub W,E'\sub W'$ be such that $W=E\op NE\op\do N^{r-1}E$, $W'=E'\op NE'\op\do N^{r-1}E'$. Clearly, $E\cap V'=0$, $E'\cap V'=0$, $NE\sub V'$, $NE\cap V''=0$, $NE'\sub V'$, $NE'\cap V''=0$. Let $E''$ be a subspace of $V'$ such that $E''$ is a complement of $NE\op V''$ in $V'$ and a complement of $NE'\op V''$ in $V'$. (Such $E''$ exists since $\dim(NE\op V'')=\dim(NE'\op V')=\dim E+\dim V''=\dim E'+\dim V''$.) Then $W_1=(E''\op NE)+N(E''\op NE)+\do+N^{r-2}(E''\op NE)$, $W'_1=(E''\op NE')+N(E''\op NE')+\do+N^{r-2}(E''\op NE')$ \nl are $(r-1)$-special subspaces of $V'$. By the induction hypothesis we can find an $N$-stable subspace $X_1$ of $V'$ such that $V_1\op X_1=V',V'_1\op X_1=V'$. Then $X=(E''+N(E'')+\do+N^{r-2}(E''))+X_1$ has the required properties. (b) {\it If $W,W'$ are $r$-special subspaces then there exists $g\in1+E_{\ge1}V_*$ such that $g(W)=W'$, $gN=Ng$.} \nl Let $X$ be as in (a). Define $g\in GL(V)$ by $g(x)=x$ for $x\in X$ and $g(w)=w'$ for $w\in W$ where $w'\in W'$ is given by $w-w'\in X$. Then $g(W)=W'$, $(g-1)X=0$ and $(g-1)W\sub X$. Clearly, $gN=Ng$. We have $V_{\ge a}=W_{\ge a}\op X_{\ge a}$. It suffices to show that $(g-1)(W_{\ge a})\sub X_{\ge a+1}$. Now $X=X_{\ge2-r}$. We have $W=W_{\ge1-r},W_{\ge2-r}=W_{\ge3-r}=NW,W_{\ge4-r}=W_{\ge5-r}=N^2W,\do$. Now if $a\le1-r$ then $(g-1)W_{\ge a}=(g-1)W\sub X=X_{\ge a+1}$. If $a=2-r$ or $a=3-r$ then $(g-1)W_{\ge a}=(g-1)NW=N(g-1)W\sub NX=NX_{\ge2-r}\sub X_{\ge4-r}\sub X_{\ge a+1}$. \nl If $a=4-r$ or $a=5-r$ then $(g-1)W_{\ge a}=(g-1)N^2W=N^2(g-1)W\sub N^2X=N^2X_{\ge2-r}\sub X_{\ge6-r}\sub X_{\ge a+1}$. \nl Continuing in this way, the result follows. \subhead 2.9\endsubhead Let $V\in\bcc$. Let $G=GL(V)$. For any filtration $V_*$ of $V$ let $$\x(V_*)=\{N\in\Nil(V);V^N_*=V_*\}=\{N\in E_{\ge2}V_*;\bN\in\End_2^0(\gr V_*)\}$$ (see 2.3(b), 2.4(a)). The following three conditions are equivalent: (i) $\x(V_*)\ne\em$; (ii) $\End_2^0(\gr V_*)\ne\em$; (iii) $\dim\gr_nV_*=\dim\gr_{-n}V_*\ge\dim\gr_{-n-2}V_*$ for all $n\ge0$. \nl We have (i)$\imp$(ii) by the definition of $\x(V_*)$; we have (ii)$\imp$(iii) by 2.3(d). The fact that (iii)$\imp$(ii) is easily checked. If (ii) holds we pick for any $a$ a subspace $V_a$ of $V_{\ge a}$ complementary to $V_{\ge a+1}$ and an element in $\End_2^0(V)$ (defined in terms of the grading $\op_aV_a$). This element is in $\x(V_*)$ and (i) holds. Let $\fF_V$ be the set of all filtrations $V_*$ of $V$ that satisfy (i)-(iii). From the definitions we have a bijection (a) $\fF_V@>\si>>D_G,V_*\m\l$ \nl ($D_G$ as in 1.1) where $\l=(G^\l_0\sps G^\l_1\sps G^\l_2\sps\do)$ is defined in terms of $V_*$ by $G^\l_0=E_{\ge0}V_*\cap G$ and $G^\l_n=1+E_{\ge n}V_*$ for $n\ge1$. The sets $\x(V_*)$ (with $V_*\in\fF_V)$ form a partition of $\Nil(V)$. (If $N\in\Nil(V)$ we have $N\in\x(V_*)$ where $V_*=V^N_*$). Let $V_*\in\fF_V$. Let $\Pi=E_{\ge0}V_*\cap G$. We show that $\x(V_*)$ is a single $\Pi$-conjugacy class. Let $N,N'\in\x(V_*)$. Since $V^N_*=V^{N'}_*$ we see from 2.3(d) that $\dim P^{\bN}_{1-j}=\dim P^{\bN'}_{1-j}$ for any $j\ge0$. Using 2.3(c) we see that for any $j\ge0$, $N,N'$ have the same number of Jordan blocks of size $j$. Hence there exists $g\in G$ such that $N'=gNg\i$. For any $a$, $gV_{\ge a}^N=V_{\ge a}^{N'}=V_{\ge a}^N$ hence $gV_{\ge a}=V_{\ge a}$. We see that $g\in E_{\ge0}$ hence $g\in\Pi$, as required. Taking in the previous argument $N'=N$, we see that, if $N\in\x(V_*)$ and $g\in G$ satisfies $gNg\i=N$ then $g\in\Pi$. Now any element in $E_{\ge2}V_*-\x(V_*)$ is in the closure of $\x(V_*)$ (since $E_{\ge2}V_*$ is irreducible and $\x(V_*)$ is open in it (and non-empty) hence it is in the closure of the $G$-conjugacy class containing $\x(V_*)$. We show that it is not contained in that $G$-conjugacy class. (Assume that it is. Then we can find $N\in\x(V_*)$ and $N'\in E_{\ge2}V_*-\x(V_*)$ that are $G$-conjugate. Then the $\Pi$-orbit $\Pi(N)$ of $N$ in $E_{\ge2}V_*$ is $\x(V_*)$ hence is dense in $E_{\ge2}V_*$ while the $\Pi$-orbit $\Pi(N')$ of $N'$ is contained in the proper closed subset $E_{\ge2}V_*-\x(V_*)$ of $E_{\ge2}V_*$; hence $\dim\Pi(N)=\dim(E_{\ge2}V_*)>\dim\Pi(N')$. It follows that $a<a'$ where $a$ (resp. $a'$) is the dimension of the centralizer of $N$ (resp. $N'$) in $\Pi$. Let $\ta$ (resp. $\ta'$) be the dimension of the centralizer of $N$ (resp. $N'$) in $G$. By an earlier argument we have $a=\ta$. Obviously, $a'\le\ta'$. Since $N,N'$ are $G$-conjugate, we have $\ta=\ta'$. Thus, $\ta=a<a'\le\ta'=\ta$, contradiction.) We see that $1+\x(V^*)=X^\l$ where $V_*\m\l$ as in (a) and $X^\l$ is as in 1.1. Thus $\fP_1$ holds for $G$. From this $\fP_2,\fP_3$ follow; $H^\bla$ in $\fP_2$ is a single conjugacy class in this case. Also, $\fP_8$ is trivial since $G^\l_0$ acts transitively on $X^\l$. Now $\fP_5$ is easily verified. $\fP_6$ (hence $\fP_4$) follows from 2.6(a); $\fP_7$ is trivial in this case. \head 3. Symplectic groups\endhead \subhead 3.1\endsubhead In this section, any text marked as $\sp\do\sp$ applies only in the case $p=2$. For $V,V'\in\cc$ let $\Bil(V,V')$ be the space of all bilinear forms $V\T V'@>>>\kk$. For $b\in\Bil(V,V')$ define $b^*\in\Bil(V',V)$ by $b^*(x,y)=b(y,x)$. We write $\Bil(V)$ instead of $\Bil(V,V)$. Let $\Symp(V)$ be the set of non-degenerate symplectic forms on $V$. Let $\bV\in\bcc$. We say that $\lar_0\in\Symp(\bV)$ is {\it admissible} if $\la x,y\ra_0=0$ for $x\in\bV_a,y\in\bV_{a'},a+a'\ne0$. Assume that $\lar_0\in\Symp(\bV)$ is admissible and that $\nu\in\End_2^0(\bV)$ is {\it skew-adjoint} that is, $\la\nu(x),y\ra_0+\la x,\nu(y)\ra_0=0$ for $x,y\in\bV$. For $n\ge0$ we define a bilinear form $b_n:P^\nu_{-n}\T P^\nu_{-n}@>>>\kk$ by $b_n(x,y)=\la x,\nu^ny\ra_0$. We show: (a) $b_n(x,y)=(-1)^{n+1}b_n(y,x)$ for $x,y\in P^\nu_{-n}$. \nl Indeed, $$b_n(x,y)=\la x,\nu^ny\ra_0=(-1)^n\la\nu^nx,y\ra_0=(-1)^{n+1}\la y,\nu^nx\ra_0 =(-1)^{n+1}b_n(y,x),$$ as required. We show: (b) {\it$b_n$ is non-degenerate.} \nl Let $y\in P^\nu_{-n}$ be such that $\la x,\nu^n y\ra_0=0$ for all $x\in P^\nu_{-n}$. If $x'\in P^\nu_{-n-2k},k>0$, we have $\la\nu^kx',\nu^ny\ra_0=\pm\la x,\nu^{n+k}y\ra_0=\pm\la x,0\ra_0=0$. Since $\bV_{-n}=\sum_{k\ge0}\nu^kP^\nu_{-n-2k}$, we see that $\la x,\nu^ny\ra_0=0$ for all $x\in\bV_{-n}$. Since $\la\bV_m,\nu^ny\ra_0=0$ for $m\ne-n$ we see that $\la\bV,\nu^ny\ra_0=0$. By the non-degeneracy of $\lar_0$, it follows that $\nu^ny=0$. Since $\nu^n:\bV_{-n}@>\si>>\bV_n$, it follows that $y=0$, as required. We show: (c) {\it if $n\ge0$ is even then $b_n$ is a symplectic form. Hence $\dim P^\nu_{-n}$ is even.} \nl Indeed, for $x\in P^\nu_{-n}$ we have $\la x,\nu^nx\ra_0=\pm\la\nu^{n/2}x,\nu^{n/2}x\ra_0=0$. \subhead 3.2\endsubhead Let $V\in\cc$ and let $\lar\in\Symp(V)$. Let $Sp(\lar)=\{T\in GL(V);T\text{ preserves }\lar\}$. \nl For any subspace $W$ of $V$ we set $W^\pe=\{x\in V;\la x,W\ra=0\}$. A filtration $V_*$ of $V$ is said to be {\it self-dual} if $(V_{\ge a})^\pe=V_{\ge1-a}$ for any $a$. It follows that (a) $\la V_{\ge a},V_{\ge a'}\ra=0$ if $a+a'\ge1$. \nl It also follows that there is a unique admissible $\lar_0\in\Symp(\gr V_*)$ such that for $x\in\gr_aV_*,y\in\gr_{-a}V_*$ we have $\la x,y\ra_0=\la\dx,\dy\ra$ where $\dx\in V_{\ge a},\dy\in V_{\ge-a}$ represent $x,y$. Moreover, (b) {\it there exists a direct sum decomposition $\op_{a\in\ZZ}V_a$ of $V$ such that $V_{\ge a}=V_a\op V_{a+1}\op\do$ for all $a$ and $\la V_a,V_{a'}\ra=0$ for all $a,a'$ such that $a+a'\ne0$.} Let $\cm_{\lar}$ be the set of $N\in\Nil(V)$ such that $\la Nx,y\ra+\la x,Ny\ra+\la Nx,Ny\ra=0$ for all $x,y$ or equivalently $1+N\in Sp(\lar)$. Define an involution $N\m N^\da$ of $\cm_{\lar}$ by $\la x,Ny\ra=\la N^\da x,y\ra$ for all $x,y\in V$ or equivalently by $N^\da=(1+N)\i-1=-N+N^2-N^3+\do$. Let $N\in\cm_{\lar}$. We set $V_*=V^N_*$. By 2.6(c) we have $V^{N^\da}_*=V_*$. We show: (c) {\it the filtration $V_*$ is self-dual.} \nl We argue by induction on $e$ as in 2.4. If $a\ge e$ then $V_{\ge a}=0,V_{\ge1-a}=V$ and (c) holds. If $a\le1-e$ then $V_{\ge a}=V,V_{\ge1-a}=0$ and (c) holds. If $e\le1$ this already suffices. Hence we may assume that $e\ge2$ and $2-e\le a\le e-1$ hence $2-e\le1-a\le e-1$. Let $V'=\ker(N^{e-1})/\Im(N^{e-1})$. Let $\r:\ker(N^{e-1})@>>>V'$ be the canonical map. We have $N^{e-1}V=\ker((N^\da)^{e-1})^\pe=\ker(N^{e-1})^\pe$ since $(N^\da)^{e-1}=(-N)^{e-1}$. Hence $\lar$ induces $\lar'\in\Symp(V')$. Also $N$ induces a linear map $N':V'@>>>V'$ such that $N'\in\cm_{\lar'}$. By the induction hypothesis, $V'{}^{N'}_{\ge1-a}$ is the perpendicular in $V'$ of $V'{}^{N'}_{\ge a}$. Hence $V_{\ge a}=\r\i(V'{}^{N'}_{\ge a})$ is the perpendicular in $V$ of $V_{\ge1-a}=\r\i(V'{}^{N'}_{\ge1-a})$. This completes the proof. Let $\nu\in\End_2^0(\gr V_*)$ be the endomorphism induced by $N$. We show that (d) {\it$\nu$ is skew-adjoint (with respect to $\lar_0$ on $\gr V^N_*$).} \nl It suffices to show that, if $a+a'+2=0$ and $x\in V_{\ge a'}$, $y\in V_{\ge a}$ then $\la Nx,y\ra+\la x,Ny\ra=0$. It suffices to show that $\la Nx,Ny\ra=0$. From (a),(b) we see that $\la V_{\ge-1-a},Ny\ra=0$ hence it suffices to show that $Nx\in V_{\ge-1-a}$. We have $Nx\in V_{\ge a'+2}\sub V_{\ge-1-a}$ since $a'+2>-1-a$. This proves (c). \subhead 3.3\endsubhead $\sp$ In this subsection we assume that $p=2$. Let $V,\lar,N,\nu,\lar_0$ be as in 3.2. Let $V_*=V^N_*$. Then $b_n\in\Bil(P_{-n}^\nu)$ is defined for $n\ge0$, see 3.1. Let $\cl$ be the set of all even integers $n\ge2$ such that $b_{n-1},b_{n+1}$ are symplectic forms. Let $\cl'$ be the set of all even integers $n\ge2$ such that $b_{n-1},b_{n+1},b_{n+3},\do$ are symplectic forms or equivalently, if $\la z,\nu^{n-1}(z)\ra_0=0$ for all $z\in\gr_{1-n}V_*$. (Assume first that $b_{n-1},b_{n+1},b_{n+3},\do$ are symplectic forms. By 2.1(a), any $z\in\gr_{1-n}V_*$ is of the form $\sum_{k\ge0}\nu^kz_k$ where $z_k\in P^\nu_{1-n-2k}$. For $k\ge0$ we have $\la\nu^kz_k,\nu^{n-1}(\nu^kz_k)\ra_0=0$ since $b_{n+2k-1}$ is symplectic. Since $z'\m\la z',\nu^{n-1}(z')\ra_0$ is additive in $z'$ it follows that $\la z,\nu^{n-1}(z)\ra_0=0$. Conversely, assume that $\la z,\nu^{n-1}(z)\ra_0=0$ for any $z\in\gr_{1-n}V_*$. In particular, for $k\ge0$ and $z_k\in P^\nu_{1-n-2k}$ we have $\la \nu^kz_k,\nu^{n-1}(\nu^kz_k)\ra_0=0$ that is, $\la z_k,\nu^{n-1+2k}z_k)\ra_0=0$. We see that $b_{n+2k-1}$ is symplectic.) Clearly, $\cl'\sub\cl$. For $n\in\cl$, we define $q_n:P^\nu_{-n}@>>>\kk$ by $q_n(x)=\la\dx,N^{n-1}\dx\ra$ where $\dx\in V_{\ge-n}$ is a representative for $x\in P^\nu_{-n}$ such that $N^{n+1}\dx=0$ (see 2.5(c)). We show that $q_n(x)$ is well defined. It suffices to show that if $y\in V_{\ge1-n},N^{n+1}y=0$ then $\la\dx+y,N^{n-1}(\dx+y)\ra=\la\dx,N^{n-1}\dx\ra$ that is, $\la y,N^{n-1}(y)\ra+\la\dx,N^{n-1}(y)\ra+\la y,N^{n-1}(\dx)\ra=0$. Since $N^{n+1}(\dx)=0$, we have $$\la\dx,N^{n-1}(y)\ra+\la y,N^{n-1}(\dx)\ra=\la y,(N^{n-1}+(N^\da)^{n-1})(\dx)\ra =\la y,N^n(\dx)\ra.$$ This is zero, since $y\in V_{\ge1-n},N^n(\dx)\in V_{\ge n}$ and $1-n+n=1$. It remains to show that $\la y,N^{n-1}(y)\ra=0$. It suffices to show that $\la z,\nu^{n-1}(z)\ra_0=0$ for all $z\in\gr_{1-n}V_*$ such that $N^{n+1}z=0$. By 2.1(a) any such $z$ is of the form $z_0+\nu z_1$ where $z_0\in P^\nu_{1-n},z_1\in P^\nu_{-1-n}$. Now $z'\m\la z',\nu^{n-1}(z')\ra_0$ is additive in $z'$ hence it suffices to show that $\la z_0,\nu^{n-1}(z_0)\ra_0=0$ and $\la\nu(z_1),\nu^{n-1}(\nu(z_1))\ra_0=0$ for $z_0,z_1$ as above. This follows from our assumption that $b_{n-1}$ and $b_{n+1}$ are symplectic. We show: (a) {\it For $x,y\in P^\nu_{-n}$ we have $q_n(x+y)=q_n(x)+q_n(y)+b_n(x,y)$.} \nl Let $\dx,\dy\in V_{\ge-n}$ be representatives for $x,y$ such that $N^{n+1}\dx=0$, $N^{n+1}\dy=0$. We must show that $\la\dx+\dy,N^{n-1}(\dx+\dy)\ra=\la\dx,N^{n-1}(\dx)\ra+\la\dy,N^{n-1}(\dy)\ra +\la\dx,N^n(\dy)\ra$, \nl or that $\la\dx,N^{n-1}(\dy)\ra+\la\dy,N^{n-1}(\dx)\ra+\la\dx,N^n(\dy)\ra=0$, \nl or that $\la\dx,((N^\da)^{n-1}+N^{n-1}+N^n)\dy\ra=0$. Since $n$ is even, $(N^\da)^{n-1}+N^{n-1}+N^n$ is a linear combination of $N^{n+1},N^{n+2},\do$ and it remains to use that $N^{n+1}(\dy)=0$. For $n\in\cl'$, we define $Q_n:\gr_{-n}V_*@>>>\kk$ by $Q_n(x)=\la\dx,N^{n-1}\dx\ra$ where $\dx\in V_{\ge-n}$ is a representative for $x$. We show that $Q_n(x)$ is well defined. It suffices to show that if $y\in V_{\ge1-n}$ then $\la\dx+y,N^{n-1}(\dx+y)\ra=\la\dx,N^{n-1}\dx\ra$ that is, $\la y,N^{n-1}(y)\ra+\la\dx,N^{n-1}(y)\ra+\la y,N^{n-1}(\dx)\ra=0$. We have $\la\dx,N^{n-1}(y)\ra+\la y,N^{n-1}(\dx)\ra=\la y,(N^{n-1}+(N^\da)^{n-1})(\dx)\ra$ \nl and this is a linear combination of terms $\la y,N^{n'}(\dx)\ra$ with $n'\ge n$. Each of these terms is $0$ since $y\in V_{\ge1-n},N^{n'}(\dx)\in V_{\ge2n'-n}$ and $1-n+2n'-n\ge1$. It remains to show that $\la y,N^{n-1}(y)\ra=0$. This follows from the fact that $\la z,\nu^{n-1}(z)\ra_0=0$ for all $z\in\gr_{1-n}V_*$. For $n\in\cl'$ we show: (b) {\it if $x,y\in\gr_{-n}V_*$ then $Q_n(x+y)=Q_n(x)+Q_n(y)+\la x,\nu^ny\ra$.} \nl Let $\dx,\dy\in V_{\ge-n}$ be representatives for $x,y$. We must show that $\la\dx+\dy,N^{n-1}(\dx+\dy)\ra=\la\dx,N^{n-1}(\dx)\ra+\la\dy,N^{n-1}(\dy)\ra +\la\dx,N^n(\dy)\ra$, \nl or that $\la\dx,N^{n-1}(\dy)\ra+\la\dy,N^{n-1}(\dx)\ra+\la\dx,N^n(\dy)\ra=0$, \nl or that $\la\dx,((N^\da)^{n-1}+N^{n-1}+N^n)\dy\ra$ is $0$. Since $n$ is even, this is a linear combination of terms $\la \dx,N^{n'}(\dy)\ra$ with $n'>n$. Each of these terms is $0$ since $N^{n'}(\dy)\in V_{\ge2n'-n},\dx\in V_{\ge-n}$ and $2n'-n-n\ge1$. Now let $n\in\cl'$ and let $x\in\gr_{-n}V_*$. We can write $x=\sum_{k\ge0}\nu^kx_k$ where $x_k\in P^\nu_{-n-2k}$. We show that (c) $Q_n(x)=\sum_{k\ge0}q_{n+2k}(x_k)$. \nl Let $\dx_k$ be a representative of $x_k$ in $V_{\ge-n-2k}$ such that $N^{n+2k+1}\dx_k=0$. Then $\sum_{k\ge0}N^k\dx_k$ is a representative of $x$ in $V_{\ge-n}$ and we must show: $$\la\sum_{k\ge0}N^k\dx_k,N^{n-1}\sum_{k'\ge0}N^{k'}\dx_{k'}\ra= \sum_{k\ge0}\la\dx_k,N^{n+2k-1}\dx_k\ra.$$ The left hand side is $\sum_{k,k'\ge0}\la N^k\dx_k,N^{n-1+k'}\dx_{k'}\ra$. If $k\ge k'+2$ we have $\la N^k\dx_k,N^{n-1+k'}\dx_{k'}\ra=\la\dx_k,(N^\da)^kN^{n-1+k'}\dx_{k'}\ra$ \nl and this is zero since $N^{n+2k'+1}\dx_{k'}=0$. If $k'\ge k+2$ we have $\la N^k\dx_k,N^{n-1+k'}\dx_{k'}\ra=\la (N^\da)^{n-1+k'}N^k\dx_k,\dx_{k'}\ra$ \nl and this is zero since $N^{n+2k+1}\dx_k=0$. It suffices to show that $$\align&\sum_{k\ge0}(\la N^k\dx_k,N^{n-1+k}\dx_k\ra +\la N^{k+1}\dx_{k+1},N^{n-1+k}\dx_k\ra+\la N^k\dx_k,N^{n+k}\dx_{k+1}\ra)\\& =\sum_{k\ge0}\la \dx_k,N^{n+2k-1}\dx_k\ra.\endalign$$ We have $$\align&\la N^{k+1}\dx_{k+1},N^{n-1+k}\dx_k\ra+\la N^k\dx_k,N^{n+k}\dx_{k+1}\ra\\& =\la N^{k+1}\dx_{k+1},(N^{n-1+k}+(N^\da)^{n-1}N^k)\dx_k\ra\\&=\la N^{k+1}\dx_{k+1}, (c_1N^{n+k}+c_2N^{n+k+1}+\do)\dx_k\ra\\&=c_1\la \nu^{k+1}x_{k+1}, \nu^{n+k}x_k\ra_0=c_1\la x_{k+1},\nu^{n+2k+1}x_k\ra_0=0.\endalign$$ (Here $c_1,c_2,\do\in\kk$.) It suffices to show that $\la N^k\dx_k,N^{n-1+k}\dx_k\ra+\la\dx_k,N^{n+2k-1}\dx_k\ra$ is $0$. This equals $$\align&\la \dx_k,(N^{n+2k-1}+(N^\da)^kN^{n-1+k})\dx_k\ra=\la \dx_k,(N^{n+2k} +c'_1N^{n+2k+1}+\do)\dx_k\ra\\& =\la x_k,\nu^{n+2k}x_k\ra_0=\la \nu^{k+n/2}x_k,\nu^{k+n/2}x_k\ra_0=0.\endalign$$ (Here $c'_1,c'_2,\do\in\kk$.) This completes the proof of (c). We say that $(q_n)_{n\in\cl}$ are {\it the quadratic forms attached to $(N,\lar)$.} We say that $(Q_n)_{n\in\cl'}$ are {\it the Quadratic forms attached to $(N,\lar)$}. $\sp$ \subhead 3.4\endsubhead Let $V\in\cc$ and let $V_*$ be a filtration of $V$. We fix $\lar_0\in\Symp(\gr V_*)$ which is admissible and $\nu\in\End_2^0(\gr V_*)$ which is skew-adjoint with respect to $\lar_0$ (see 3.1). Then $P^\nu_{-n}$ are defined in terms of $\gr V_*,\nu$ and $b_n\in\Bil(P^\nu_{-n})$ are defined as in 3.1 for any $n\ge0$. Let $\cv=1+E_{\ge1}V_*$, a subgroup of $GL(V)$. $\sp$ If $p=2$, let $\nn$ be the smallest even integer $\ge2$ such that $b_{\nn-1},b_{\nn+1},b_{\nn+3},\do$ are symplectic or, equivalently, such that $\la z,\nu^{\nn-1}(z)\ra_0=0$ for all $z\in\gr_{1-\nn}V_*$. Let $Q:\gr_{-\nn}V_*@>>>\kk$ be a quadratic form such that $Q(x+y)=Q(x)+Q(y)+\la x,\nu^\nn y\ra$ for all $x,y\in\gr_{-\nn}V_*$. $\sp$. Let $\cz$ be the set of all pairs $(N,\lar)$ where $N\in\Nil(V)$, $\lar\in\Symp(V)$ are such that $V^N_*=V_*$, $\la Nx,y\ra+\la x,Ny\ra+\la Nx,Ny\ra=0$ for $x,y\in V$, $N$ induces $\nu$ on $\gr V_*$, $\lar$ induces $\lar_0$ on $\gr V_*$; $\sp$ in the case $p=2$ we require in addition that $Q_\nn$ defined in terms of $(N,\lar)$ as in 3.3 is equal to $Q$. $\sp$ The proofs of Propositions 3.5, 3.6, 3.7 below are intertwined (see 3.11). \proclaim{Proposition 3.5}In the setup of 3.4 let $\lar\in\Symp(V)$ be such that $V_*$ is self-dual with respect to $\lar$ and $\lar$ induces $\lar_0$ on $\gr V_*$. Let $Y=Y_{\lar}=\{N;(N,\lar)\in\cz\}$. Let $U'=\cv\cap Sp(\lar)$, a subgroup of $Sp(\lar)$. Then (a) $Y\ne\em$; (b) if $N\in Y$ and $z\in U'$ then $zNz\i\in Y$ (thus $U'$ acts an $Y$ by conjugation); (c) the action (b) of $U'$ on $Y$ is transitive. \endproclaim The proof of (a) is given in 3.8. Now (b) follows immediately from 3.7(a). We show that (c) is a consequence of 3.7(c). Assume that 3.7(c) holds. Let $N,N'\in Y$. We have $(N,\lar)\in\cz,(N',\lar)\in\cz$ and by 3.7(c) there exists $g\in\cv$ such that $N'=gNg\i$, $\la g\i x,g\i y\ra=\la x,y\ra$ for $x,y\in V$. Then $g\in U'$ and (c) is proved (assuming 3.7(c)). \proclaim{Proposition 3.6}In the setup of 3.4 let $N\in\Nil(V)$ be such that $V^N_*=V_*$ and $N$ induces $\nu$ on $\gr V_*$. Let $X=X_N=\{\lar;(N,\lar)\in\cz\}$. Let $U=U_N=\{T\in\cv;TN=NT\}$, a subgroup of $GL(V)$. Then: (a) $X\ne\em$; (b) if $\lar\in X$ and $u\in U$ then the symplectic form $\lar'$ on $V$ given by $\la x,y\ra'=\la u\i x,u\i y\ra$ belongs to $X$ (thus $U$ acts naturally an $X$); (c) the action (b) of $U$ on $X$ is transitive. \endproclaim We show that (a) is a consequence of 3.7(a). By 3.7(a) there exists $(N',\lar')\in\cz$. By 2.6(a) there exists $g\in\cv$ such that $N=gN'g\i$. Define $\lar\in\Symp(V)$ by $\lar=\la g\i x,g\i y\ra'$. From 3.7(a) we see that $(N,\lar)\in\cz$ hence $\lar\in X_N$. Thus $X_N\ne\em$, as required. Now (b) follows immediately from 3.7(b). The proof of (c) is given in 3.9, 3.10. \proclaim{Proposition 3.7}In the setup of 3.4, (a) $\cz\ne\em$; (b) if $(N,\lar)\in\cz$, $g\in\cv$ and $(N',\lar')$ is defined by $N'=gNg\i$, $\la x,y\ra'=\la g\i x,g\i y\ra$ then $(N',\lar')\in\cz$ (thus $\cv$ acts naturally on $\cz$); (c) the action (b) of $\cv$ on $\cz$ is transitive. \endproclaim Clearly, (a) is a consequence of 3.5(a). We prove (b). We have $V^{N'}_{\ge a}=gV^N_{\ge a}=V^N_{\ge a}=V_{\ge a}$. Next we must show that $\la gNg\i x,y\ra'+\la x,gNg\i y\ra'+\la gNg\i x,gNg\i y\ra'=0$ for $x,y\in V$ that is, $\la Ng\i x,g\i y\ra+\la g\i x,Ng\i y\ra+\la Ng\i x,Ng\i y\ra=0$ for $x,y\in V$. This follows from $\la Nx',y'\ra+\la x',Ny'\ra+\la Nx',Ny'\ra=0$ for $x',y'\in V$. Next we must show that $gNg\i,N$ induce the same map $\gr V_*@>>>\gr V_*$. (We must show: if $x\in V_{\ge a}$ then $gNg\i(x)-Nx\in V_{\ge a+3}$; this follows from $g\in\cv$.) Next we must show that for $x\in V_{\ge-a},y\in V_{\ge a}$ we have $\la x,y\ra'=\la x,y\ra$ that is $\la g\i x,g\i y\ra=\la x,y\ra$. Set $g\i=1+S$ where $S\in E_{\ge1}V_*$. We must show that $\la Sx,y\ra+\la x,Sy\ra+\la Sx,Sy\ra=0$. But $Sx\in V_{\ge1-a},y\in V_{\ge a}$ implies $\la Sx,y\ra=0$. Similarly $\la x,Sy\ra=0,\la Sx,Sy\ra=0$. $\sp$ In the case where $p=2$ we see that the number $\nn$ defined in terms of $N,\lar$ is the same as that defined in terms of $N',\lar'$ and we must check that for $x\in V_{\ge-\nn}$ we have $\la x,(gNg\i)^{\nn-1}x\ra'=\la x,N^{\nn-1}x\ra$ that is, $\la g\i x,N^{\nn-1}g\i x\ra=\la x,N^{\nn-1}x\ra$ that is, $\la Sx,N^{\nn-1}x\ra+\la x,N^{\nn-1}Sx\ra+\la Sx,N^{\nn-1}Sx\ra=0$. We have $\la Sx,N^{\nn-1}x\ra+\la x,N^{\nn-1}Sx\ra=\la x,(N^{\nn-1}+(N^\da)^{\nn-1}Sx\ra$. This is a linear combination of terms $\la x,N^{n'}Sx\ra$ where $n'\ge\nn$; each of these terms is zero since $x\in V_{\ge-\nn},N^{n'}Sx\in V_{\ge2n'-\nn+1}$ and $-\nn+2n'-\nn+1\ge1$. Next we have $\la Sx,N^{\nn-1}Sx\ra=0$ since $\la y,N^{\nn-1}y\ra=0$ for all $y\in V_{\ge1-\nn}$ by the definition of $\nn$. $\sp$ This completes the proof of (b). We show that (c) is a consequence of 3.6(c). Let $(N,\lar)\in\cz,(N',\lar')\in\cz$. By 2.6(a), since $V_*^N=V_*^{N'}$ and $N,N'$ induce the same $\nu$, we can find $S\in E_{\ge1}V_*$ such that $R=1+S$ satisfies $N'R=RN$. Define $\lar''\in\Symp(V)$ by $\la x,y\ra''=\la Rx,Ry\ra'$. From (b) we see that $(R\i N'R,\lar'')\in\cz$ that is $(N,\lar'')\in\cz$. Thus $\lar\in X_N,\lar''\in X_N$. By 3.6(c) we can find $S'\in E_{\ge1}V_*$ such that $R'=1+S'$ satisfies $R'N=NR'$ and $\la x,y\ra=\la R'x,R'y\ra''$ for all $x,y$ that is, $\la x,y\ra=\la RR'x,RR'y\ra'$. Then $RR'\in U'$ and $RR'N=RNR'=N'RR'$. Thus under the action (b), $RR'$ carries $(N,\lar)$ to $(N',\lar')$. This proves (c) (assuming 3.6(c)). \subhead 3.8. Proof of 3.5(a)\endsubhead We choose a direct sum decomposition $\op_{a\in\ZZ}V_a$ of $V$ as in 3.2(b). Define $N_2\in\End_2(V)$ by the requirement that $N_2:V_a@>>>V_{a+2}$ corresponds to $\nu:\gr_aV_*@>>>\gr_{a+2}V_*$ under the obvious isomorphisms $V_a@>\si>>\gr_aV_*$, $V_{a+2}@>\si>>\gr_{a+2}V_*$. $\sp$. If $p=2$ we regard $Q$ as a quadratic form on $V_{-\nn}$ via the obvious isomorphism $V_{-\nn}@>\si>>\gr_{-\nn}V_*$. $\sp$ We will construct a linear map $N=\sum_{j\ge1}N_{2j}$ where $N_2$ is as above and for $j\ge2$, $N_{2j}\in\End(V)$ satisfy $N_{2j}V_a\sub V_{a+2j}$ for all $a$ and $$\la\sum_{j\ge1}N_{2j}x,y\ra+\la x,\sum_{j\ge1}N_{2j}y\ra+ \la\sum_{j'\ge1}N_{2j'}x,\sum_{j''\ge1}N_{2j''}y\ra=0$$ for any $a,c$ and any $x\in V_a,y\in V_c$ that is, $$\la N_{2j}x,y\ra+\la x,N_{2j}y\ra +\sum_{j',j''\ge1;j'+j''=j}\la N_{2j'}x,N_{2j''}y\ra=0\tag a$$ for any $j\ge1$, any $a,c$ such that $a+c+2j=0$ and any $x\in V_a,y\in V_c$. $\sp$ If $p=2$, we require in addition that $\la x,N^{\nn-1}x\ra=Q(x)$ for all $x\in V_{-\nn}$ that is, $\sum_{i+i'=\nn-2}\la x,N_2^iN_4N_2^{i'}x\ra=Q(x)$ for all $x\in V_{-\nn}$. $\sp$ We shall determine $N_j$ by induction on $j$. For $j=1$ the equation (a) is just $\la N_2x,y\ra+\la x,N_2y\ra=0$ for any $a,c$ such that $a+c+2=0$ and any $x\in V_a,y\in V_c$; this holds automatically by our choice of $N_2$. For $x\in V_a$ with $a<-2$ we set $N_4(x)=0$. Then the equation (a) for $j=2$ becomes (b) $\la N_4x,y\ra+\la x,N_4y\ra=-\la N_2x,N_2y\ra$ for any $x\in V_{-2},y\in V_{-2}$, $\la N_4x,y\ra=-\la N_2x,N_2y\ra$ for any $a>-2,x\in V_a,y\in V_{-a-4}$. \nl The second equation in (b) determines uniquely $N_4(x)$ for $x\in V_a,a>-2$. Since $\la N_2x,N_2y\ra$ is a symplectic form on $V_{-2}$ we can find $[,]\in\Bil(V_{-2})$ such that $[x,y]-[y,x]=-\la N_2x,N_2y\ra$ for any $x,y\in V_{-2}$. There is a unique linear map $N_4:V_{-2}@>>>V_2$ such that $\la N_4x,y\ra=[x,y]$ for any $x,y\in V_{-2}$. Then equation (a) for $j=2$ is satisfied. $\sp$ If $p=2$ the $N_4$ just determined satisfies $\sum_{i+i'=\nn-2}\la x,N_2^iN_4N_2^{i'}x\ra=Q'(x)$ for all $x\in V_{-\nn}$, for some quadratic form $Q':V_{-\nn}@>>>\kk$ not necessarily equal to $Q$. For $x,y\in V_{-\nn}$ we have (by the choice of $N_4$): $$\align&Q'(x+y)-Q'(x)-Q'(y)=\sum_{i+i'=\nn-2}\la x,N_2^iN_4N_2^{i'}y\ra+ \sum_{i+i'=\nn-2}\la y,N_2^iN_4N_2^{i'}x\ra\\&=\sum_{i+i'=\nn-2}\la N_2^ix, N_4N_2^{i'}y\ra+\sum_{i+i'=\nn-2}\la N_4N_2^ix,N_2^{i'}y\ra =\sum_{i+i'=\nn-2}\la N_2N_2^ix,N_2N_2^{i'}y\ra\\&= \sum_{i+i'=\nn-2}\la x,N_2^\nn y\ra=\la x,N_2^\nn y\ra=Q(x+y)-Q(x)-Q(y).\endalign$$ It follows that $Q'(x)=Q(x)+\th(x)^2$ where $\th\in\Hom(V_{-\nn},\kk)$. We try to find $\z\in\End(V)$ with $\z(V_a)\sub V_{a+4}$ for all $a$ in such a way that (a) (for $j=2$) remains true when $N_4$ is replaced by $N_4+\z$ and $\sum_{i+i'=\nn-2}\la x,N_2^i(N_4+\z)N_2^{i'}x\ra=Q(x)$ for $x\in V_{-\nn}$. (Then $N_4+\z$ will be our new $N_4$.) Thus we are seeking $\z$ such that $\la\z(x),y\ra+\la x,\z(y)\ra=0$ for any $a,c$ with $a+c+4=0$ and $x\in V_a,y\in V_c$, $\sum_{i+i'=\nn-2}\la x,N_2^i\z N_2^{i'}x\ra=\th(x)^2$ for $x\in V_{-\nn}$. \nl The first of these two equations can be satisfied for $(a,c)\ne(-2,-2)$ by defining $\z(x)=0$ for $x\in V_a,a\ne-2$. Then in the second equation the terms corresponding to $i'$ such that $2i'-\nn\ne-2$ are $0$. Thus it remains to find a linear map $\z:V_{-2}@>>>V_2$ such that $\la \z(x),y\ra+\la x,\z(y)\ra=0$ for any $x,y\in V_{-2}$, $\la N_2^tx,\z N_2^tx\ra=\th(x)^2$ for $x\in V_{-\nn}$ where $t=(\nn-2)/2$. \nl Since $N_2^t:V_{-\nn}@>>>V_{-2}$ is injective (by the Lefschetz condition), there exists $\th_1\in\Hom(V_{-2},\kk)$ such that $\th_1(N_2^tx)=\th(x)$ for all $x\in V_{-\nn}$. We see that it suffices to find $\z\in\Hom(V_{-2},V_2)$ such that $\la \z(x),y\ra+\la x,\z(y)\ra=0$ for any $x,y\in V_{-2}$, $\la x',\z x'\ra=\th_1(x')^2$ for $x'\in V_{-2}$. \nl It also suffices to find $b_0\in\Bil(V_{-2})$ such that $b_0=b_0^*$ and $b_0(x,x)=\th_1(x)^2$ for $x\in V_{-2}$. Such $b_0$ clearly exists. $\sp$. This completes the determination of $N_4$. Now assume that $j\ge3$ and that $N_{2j'}$ is already determined for $j'<j$. For $x\in V_a$ with $a<-j$ we set $N_{2j}(x)=0$. Then equation (a) for our $j$ determines uniquely $N_{2j}(x)$ for $x\in V_a$ with $a>-j$. Next, we can find $[,]\in\Bil(V_{-j})$ such that $[x,y]-[y,x]=-\sum_{j',j''\ge1;j'+j''=j}\la N_{2j'}x,N_{2j''}y\ra$. \nl To see this we observe that the right hand side is a symplectic form that is, $\sum_{j',j''\ge1;j'+j''=j}\la N_{2j'}x,N_{2j''}x\ra=0$. There is a unique $N_{2j}\in\Hom(V_{-j},V_j)$ such that $\la N_{2j}x,y\ra=[x,y]$ for any $x,y\in V_{-j}$. Then equation (a) for our $j$ is satisfied. This completes the inductive construction of $N$. We have $V_*^N=V_*$ by 2.4(a). We see that $N\in Y$. This completes the proof. \subhead 3.9\endsubhead In this subsection we prove 3.6(c) in a special case. Let $n\in\ZZ_{>0}$. We have $[-n,n]=I_0\sqc I_1$ where $I_\e=\{i\in[-n,n];i=\e\mod 2\}$ for $\e\in\{0,1\}$. For $i\in[-n,n]$ define $|i|\in\{0,1\}$ by $i=|i|\mod2$ that is by $i\in I_{|i|}$. Let $F_0,F_1\in\cc$. Let $V=\op_{i\in[-n,n]}F_i$ where $F_i=F_{|i|}$. A typical element of $V$ is of the form $(x_i)_{i\in[-n,n]}$ where $x_i\in F_{|i|}$. Define $N:V@>>>V$ by $(x_i)\m(x'_i)$ where $x'_i=x_{i-2}$ for $i\in[2-n,n]$, $x'_{-n}=0,x'_{1-n}=0$. We fix $\lar_0\in\Symp(V)$ such that $\la (x_i),(y_i)\ra_0=\sum_{i\in[-n,n]}(-1)^{(i-|i|)/2}b^{|i|}(x_i,y_{-i})$ where $b^\e\in\Bil(F_\e)(\e\in\{0,1\}$ satisfy $b^{\e*}=(-1)^{1-\e}b^\e$, $b^\e$ is non-degenerate, $b^0\in\Symp(F_0)$. Note that $\la Nx,y\ra_0+\la x,Ny\ra_0=0$ for $x,y\in V$. We assume: if $p\ne2$ then either $F_0=0$ or $F_1=0$; $\sp$ if $p=2,b^1$ is symplectic and $n\ge2$, then we are given a quadratic form $Q:F_0@>>>\kk$ such that $Q(x+y)=Q(x)+Q(y)+b^0(x,y)$ for $x,y\in F_0$. $\sp$ Let $X$ be the set of all $\lar\in\Symp(V)$ such that $\la Nx,y\ra+\la x,Ny\ra+\la Nx,Ny\ra=0$ for $x,y\in V$ and $\la x,y\ra=\la x,y\ra_0$ if there exists $i$ such that $x_j=0$ for $j\ne i$ and $y_j=0$ for $j\ne-i$; $\sp$ if $p=2,b^1$ is symplectic and $n\ge2$, we require also that $\la x,Nx\ra=Q(x_{-2})$ if $x\in V$ is such that $x_j=0$ for $j\ne-2$. $\sp$ Setting $\la (x_i),(y_i)\ra=\sum_{i,j}b_{ij}(x_i,y_j)$ identifies $X$ with the set of all families $(b_{ij})_{i,j\in[-n,n]}$ where $b_{ij}\in\Bil(F_{|i|},F_{|j|})$ are such that $b_{i-2,j}+b_{i,j-2}+b_{ij}=0$ if $i,j\in[2-n,n]$, $b_{i,-i}=(-1)^{(i-|i|)/2}b^{|i|}$ for all $i\in[-n,n]$, $b_{ii}\in\Symp(F_{|i|})$ for all $i\in[-n,n]$, $b_{ij}^*=-b_{ji}$ for all $i,j\in[-n,n]$, $b_{-2,0}(x,x)=Q(x)$ for $x\in F_0$ if $p=2,F_1=0$ and $n$ is even, $\ge2$. \nl (We have automatically $b_{ij}=0$ if $i+j\ge1$.) Let $\D=\{T\in GL(V);TN=NT\}$, a subgroup of $GL(V)$; equivalently $\D$ is the set of linear maps $T:V@>>>V$ of the form (a) $T:(x_i)\m(x'_i),x'_i=\sum_{j\in[-n,i]}T_{i-j}^{|i|,|j|}x_j$ \nl where $T_r^{\e,\d}\in\Hom(F_\d,F_\e)$ $(r\in[0,2n],\e,\d\in\{0,1\},r+\d=\e\mod 2)$ are such that $T_0^{00},T_0^{11}$ are invertible and $T_{2n}^{1-|n|,1-|n|}=0$. Now $\D$ acts on $X$ by $T:\lar\m\lar'$ where $\la Tx,Ty\ra'=\la x,y\ra$, or equivalently by $T:(b_{ij})\m(b'_{ij})$ where $$b_{ij}(x,y)=\sum_{i'\in[i,n],j'\in[j,n]}b'_{i'j'}(T_{i'-i}^{|i'|,|i|}(x), T_{j'-j}^{|j'|,|j|}(y)).$$ Let $\D_u=\{T\in\D;T_0^{00}=1,T_0^{11}=1\}$, a subgroup of $\D$. We show: (b) {\it Let $k\in[1-n,0]$ and let $(\tb_{ij}),(b_{ij})$ be two points of $X$ such that $b_{ij}=\tb_{ij}$ for $i+j\ge2k$. Then there exists $T\in\D_u$ such that $T(b_{ij})=(b'_{ij})$ and $b'_{ij}=\tb_{ij}$ for $i+j\ge2k-2$.} \nl For $\e\in\{0,1\}$ we set $a^\e=\tb_{ij}$ for $i,j\in[-n,n],i+j=-1,i=\e\mod2$. Then $a^\e$ are independent of choices; they are $0$ unless $p=2$. We have $a^{\e*}=a^{1-\e}$. For $h\in\{2k-2,2k-1\}$ we set $c^\e_h=(-1)^{(i-\e)/2}(b_{ij}-\tb_{ij})$ where $i,j\in[-n,n],i+j=h,i=\e\mod2$. Then $c^\e_h$ is independent of $i,j$. We have $c^\e_{2k-1}=0$ unless $p=2$. We have $c^{\e*}_{2k-2}=(-1)^{k-\e}c_{2k-2}^\e$, $c_{2k-1}^{\e*}=c_{2k-1}^{1-\e}$. Since $b_{k-1,k-1}-\tb_{k-1,k-1}$ is symplectic, $c_{2k-2}^\e$ is symplectic where $\e=k-1\mod2$. {\it Case 1: $p\ne2$.} Let $\e\in\{0,1\}$ be such that $F_{1-\e}=0$. Since $c^{\e*}_{2k-2}=(-1)^{k-\e}c_{2k-2}^\e$, we can find $\tc\in\Bil(F_\e)$ such that $c_{2k-2}^\e=\tc+(-1)^{k-\e}\tc^*$. Since $b^\e$ is non-degenerate we can find $\t\in\End(F_\e)$ such that $\tc(x,y)=b^\e(x,\t(y))$ for $x,y\in F_\e$. For $i,j\in[-n,n],i+j=2k-2,i=\e\mod2$ and $x,y\in F_\e$ we have $$\align&b_{ij}(x,y)-\tb_{ij}(x,y)=(-1)^{(i-\e)/2}(\tc(x,y)+(-1)^{k-\e}\tc(y,x))\\&= \tb_{i,j+2-2k}(x,\t(y))-\tb_{j,i+2-2k}(y,\t(x))=\tb_{i,j+2-2k}(x,\t(y)) +\tb_{i+2-2k,j}(\t(x),y).\endalign$$ Let $T$ be as in (a) with $T_0^{00}=1,T_0^{11}=1$, $T_{2-2k}^{\e,\e}=\t$ and the other components $0$. Define $(b'_{ij})$ by $T(b_{ij})=(b'_{ij})$. Then $(b'_{ij})$ has the required properties. $\sp$ {\it Case 2: $p=2,k=0$.} Since $b^0$ is non-degenerate we can find $T_1^{0,1}\in\Hom(F_1,F_0)$ such that $c_{-1}^0(x,y)=\tb^0(x,T_1^{0,1}(y))$ for all $x\in F_0,y\in F_1$. Then $c_{-1}^1(x,y)=\tb^0(T_1^{0,1}(x),y)$ for all $x\in F_1,y\in F_0$. Thus, for $i\in I_0,j\in I_1,i+j=-1$ and $x\in F_0,y\in F_1$ we have $b_{ij}(x,y)+\tb_{ij}(x,y)=\tb_{i,-i}(x,T_1^{0,1}(y))$; for $i\in I_1,j\in I_0,i+j=-1$ and $x\in F_1,y\in F_0$ we have $$b_{ij}(x,y)+\tb_{ij}(x,y)=\tb_{-j,j}(T_1^{0,1}(x),y).$$ Since $c_{-2}^{0*}=c_{-2}^0,b^{1*}=b^1$, we have $c_{-2}^0(y,y)=\th(y)^2$ for $y\in F_0$, $b^1(x,x)=\th_1(x)^2$ for $x\in F_1$ where $\th\in\Hom(F_0,\kk)$, $\th_1\in\Hom(F_1,\kk)$. If $b^1$ is not symplectic, we have $\th_1\ne0$. Hence there exists $T_1^{1,0}\in\Hom(F_0,F_1)$ such that $\th(y)=\th_1(T_1^{1,0}(y))$ for all $y\in F_0$. Then $c_{-2}^0(y,y)+b^1(T_1^{1,0}(y),T_1^{1,0}(y))=0$ for all $y\in F_0$. Thus $c_{-2}^0+b^1(T_1^{1,0}\ot T_1^{1,0})$ is symplectic. This also holds if $b^1$ is symplectic (in that case we have $c_{-2}^0(y,y)=b_{-2,0}(y,y)-\tb_{-2,0}(y,y)=Q(y)-Q(y)=0$ for $y\in F_0$) and we take $T_1^{1,0}=0$. Now $c_{-2}^1$ is also symplectic. Since $a^{0*}=a^1$, $a^1(T_1^{1,0}\ot1)+a^0(1\ot T_1^{1,0})$ is symplectic. Similarly $a^0(T_1^{0,1}\ot1)+a^1(1\ot T_1^{0,1})$ is symplectic. Hence $c_{-2}^0+b^1(T_1^{1,0}\ot T_1^{1,0})+a^1(T_1^{1,0}\ot1)+a^0(1\ot T_1^{1,0})$ is symplectic and $c_{-2}^1+a^0(T_1^{0,1}\ot1)+a^1(1\ot T_1^{0,1})$ is symplectic. Hence we can find $\tc^0\in\Bil(F_0),\tc^1\in\Bil(F_1)$ such that $$c_{-2}^0+b^1(T_1^{1,0}\ot T_1^{1,0})+a^1(T_1^{1,0}\ot1)+a^0(1\ot T_1^{1,0})= \tc^0+\tc^{0*},$$ $$c_{-2}^1+a^0(T_1^{0,1}\ot1)+a^1(1\ot T_1^{0,1})=\tc^1+\tc^{1*}.$$ Since $b^0,b^1$ are non-degenerate we can find $T_2^{0,0}\in\End(F_0)$, $T_2^{1,1}\in\End(F_1)$ such that $\tc^0(x,y)=b^0(x,T_2^{0,0}(y))$ for $x,y\in F_0$, $\tc^1(x,y)=b^1(x,T_2^{1,1}(y))$ for $x,y\in F_1$. For $x,y\in F_0$ we have $$\align&c_{-2}^0(x,y)+b^1(T_1^{1,0}(x)\ot T_1^{1,0}(x))+a^1(T_1^{1,0}(x),y) +a^0(x,T_1^{1,0}(y))\\&=b^0(x,T_2^{0,0}(y))+b^0(T_2^{0,0}(x),y).\endalign$$ For $x,y\in F_1$ we have $$c_{-2}^1(x,y)+a^0(T_1^{0,1}(x),y)+a^1(x,T_1^{0,1}(y))=b^1(x,T_2^{1,1}(y)) +b^1(T_2^{1,1}(x),y).$$ Thus, for $i,j\in I_0,i+j=-2$ and $x,y\in F_0$ we have $$\align&b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i+1,j}(T_1^{1,0}(x),y)+ \tb_{i,j+1}(x,T_1^{1,0}(y))\\&+\tb_{i+1,j+1}(T_1^{1,0}(x),T_1^{1,0}(y))+ \tb_{i,-i}(x,T_2^{0,0}(y))+\tb_{-j,j}(T_2^{0,0}(x),y);\endalign$$ for $i,j\in I_1,i+j=-2$ and $x,y\in F_1$ we have $$\align&b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i+1,j}(T_1^{0,1}(x),y)+\tb_{i,j+1}(x,T_1^{0,1} (y))\\&+\tb_{i,-i}(x,T_2^{1,1}(y))+\tb_{-j,j}(T_2^{1,1}(x),y).\endalign$$ Let $T$ be as in (a) with $T_0^{00}=1,T_0^{11}=1,T_1^{1,0},T_1^{0,1},T_2^{1,1}$, $T_2^{0,0}$ as above and the other components $0$. Define $(b'_{ij})$ by $T(b_{ij})=(b'_{ij})$. Then $(b'_{ij})$ has the required properties. {\it Case 3: $p=2,k=-1$.} In this case we have $n\ge2$. We first show that there exists $\s\in\End(F_1)$ such that $$b^1(x,\s(y))=b^1(\s(x),y),\quad c^1_{-4}(x,x)=b^1(x,\s(x))+b^1(\s(x),\s(x))\tag *$$ for $x,y\in F_1$. The functions $F_1@>>>\kk,x\m b^1(x,x),x\m c^1_{-4}(x,x)$ are additive and homogeneous of degree $2$, hence are of the form $x\m\th(x)^2,x\m\th_1(x)^2$ where $\th,\th_1\in\Hom(F_1,\kk)$. We can find a direct sum decomposition $F_1=F'\op F''$ where $b^1(x',x'')=0$ for all $x'\in F',x''\in F''$, $\th|_{F'}=0$, $F'=F_1$ if $\th=0$, $\dim F''\in\{1,2\}$ if $\th\ne0$. Define $\s'\in\End(F')$ by $\th_1(x)\th_1(y)=b^1(x,\s'(y))$ for $x,y\in F'$. Then $b^1(x,\s'(y))=b^1(\s'(x),y)$ for $x,y\in F'$, $\th_1(x)^2=b^1(x,\s'(x))+\th(\s'(x))^2$ for $x\in F'$. If $\dim F''=1$ we have $\th|_{F''}\ne0$ and there is a unique $v\in F''$ such that $\th(v)=1$. Let $\s'':F''@>>>F''$ be multiplication by $a$ where $a\in\kk$ satisfies $a^2+a=\th_1(v)^2$. Then $\th_1(x)^2=b^1(x,\s''(x))+\th(\s''(x))^2$ for $x\in F''$ and $b(x,\s''(y))=b(\s''(x),y)$ for $x,y\in F''$. If $\dim F''=2$ we can find a basis $\{v,v'\}$ of $F''$ such that $\th(v')=0,\th(v'')=1$. We set $b(v',v'')=f$. We have $f\ne0$. Define $\s''\in\End(F'')$ by $\s''(v')=\ta f\i v'+\ta v''$, $\s''(v'')=\th_1(v'')^2f\i v'$ where $\ta\in\kk$ satisfies $\ta^2+\ta=\th_1(v')^2$. Then $b(x,\s''(y))=b(\s''(x),y)$ for $x,y\in F''$, $\th_1(x)^2=b(x,\s''(x))+\th(\s''(x))^2$ for $x\in F''$. If $F''=0$ let $\s'':F''@>>>F''$ be the $0$ map. Define $\s\in\End(F_1)$ by $\s(x)=\s'(x)$ if $x\in F'$, $\s(x)=\s''(x)$ if $x\in F''$. Then $\s$ satisfies $(*)$. Since $b^0$ is non-degenerate we can find $T_3^{0,1}\in\Hom(F_1,F_0)$ such that $$c^0_{-3}(x,y)+a^0(x,\s(y))=b^0(x,T_3^{0,1}(y))$$ for $x\in F_0,y\in F_1$. For any $i\in I_0,j\in I_1,i+j=-3$ and $x\in F_{|i|}$, $y\in F_{|j|}$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i,j+2}(x,\s(y))+\tb_{i,j+3}(x,T_3^{0,1}(y)).$$ It follows that for any $i\in I_1,j\in I_0,i+j=-3$ and $x\in F_{|i|},y\in F_{|j|}$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i+2,j}(\s(x),y)+\tb_{i+3,j}(T_3^{0,1}(x),y).$$ Define $d_1\in\Bil(F_1)$ by $d_1(x,y)=\tb_{i,j+2}(x,\s(y))+\tb_{i+2,j}(\s(x),y)$ where $i,j\in I_1,i+j=-4$. Using the first equality in $(*)$ we see that $d_1$ is independent of the choice of $i,j$. Define $d\in\Bil(F_1)$ by $$d(x,y)=c^1_{-4}(x,y)+d_1(x,y)+b^1(\s(x),\s(y))+a^0(T_3^{0,1}(x),y) +a^1(x,T_3^{0,1}(y)).$$ We have $d(x,x)=0$ for $x\in F_1$. (We use $(*)$ and the identity $\tb_{i,j+2}+\tb_{j,i+2}=b^1$ for $i,j\in I_1,i+j=-4$.) Thus, $d$ is symplectic hence we can find $d'\in\Bil(F_1)$ such that $d=d'+d'{}^*$. Since $b^1$ is non-degenerate we can find $T_4^{1,1}\in\End(F_1)$ such that $d'(x,y)=b^1(x,T_4^{1,1}(y))$ for $x,y\in F_1$. We have $$d(x,y)=b^1(x,T_4^{1,1}(y))+b^1(y,T_4^{1,1}(x))=b^1(x,T_4^{1,1}(y)) +b^1(T_4^{1,1}(x),y).$$ Hence for $i,j\in I_1,i+j=-4$ and $x,y\in F_1$ we have $$\align&b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i,j+2}(x,\s(y))+\tb_{i+2,j}(\s(x),y)\\&+ \tb_{i+2,j+2}(\s(x),\s(y))+\tb_{i+3,j}(T_3^{0,1}(x),y)+\tb_{i,j+3}(x,T_3^{0,1}(y))\\& +b_{i,j+4}(x,T_4^{1,1}(y))+b_{i+4,j}(T_4^{1,1}(x),y).\endalign$$ For $i,j\in I_1,i+j=-2$ and $x,y\in F_1$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i,j+2}(x,\s(y))+\tb_{i+2,j}(\s(x),y)=\tb_{ij}(x,y).$$ Define $f\in\Bil(F_0)$ by $f(x,y)=c^0_{-4}(x,y)$. Then $f$ is symplectic. (We use that $c_{-4}^0$ is symplectic.) Hence we can find $f'\in\Bil(F_0)$ such that $f=f'+f'{}^*$. Since $b^0$ is non-degenerate we can find $T_4^{0,0}\in\End(F_0)$ such that $f'(x,y)=b^0(x,T_4^{0,0}(y))$ for $x,y\in F_0$. We have $$f(x,y)=b^0(x,T_4^{0,0}(y))+b^0(y,T_4^{0,0}(x))=b^0(x,T_4^{0,0}(y)) +b^0(T_4^{0,0}(x),y)$$ hence for $i,j\in I_0,i+j=-4$ and $x,y\in F_0$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+b_{i,j+4}(x,T_4^{0,0}(y))+b_{i+4,j}(T_4^{0,0}(x),y).$$ For $i,j\in I_0,i+j=-2$ and $x,y\in F_0$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y).$$ Let $T$ be as in (a) with $T_0^{00}=1,T_0^{11}=1$, $T_3^{0,1},T_4^{1,1},T_4^{0,0}$, $T_2^{1,1}=\s$ as above and the other components $0$. Define $(b'_{ij})$ by $T(b_{ij})=(b'_{ij})$. Then $(b'_{ij})$ has the required properties. {\it Case 4: $p=2,k<-1$.} In this case we have $n\ge3$. Define $\e,\d\in\{0,1\}$ by $\e=k-1\mod 2,\d=1-\e$. Since $b^\d$ is non-degenerate, we have $c_{2k-2}^\d(x,y)=b^\d(x,\s(y))$ for $x,y\in F_\d$ where $\s\in\End(F_\d)$ is well defined. Since $b^{\d*}=b^\d,c_{2k-2}^{\d*}=c_{2k-2}^\d$, we have $b^\d(x,\s(y))=b^\d(\s(x),y)$. Since $b^\e$ is non-degenerate we can find $T_{1-2k}^{\e,\d}\in\Hom(F_\d,F_\e)$ such that $$c^\e_{2k-1}(x,y)+a^\e(x,\s(y))=b^\e(x,T_{1-2k}^{\e,\d}(y))$$ for $x\in F_\e,y\in F_\d$. For any $i\in I_\e,j\in I_\d,i+j=2k-1$ and $x\in F_\e$, $y\in F_\d$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i,j-2k}(x,\s(y)) +\tb_{i,j+1-2k}(x,T_{1-2k}^{\e,\d}(y)).$$ It follows that for any $i\in I_\d,j\in I_\e,i+j=2k-1$ and $x\in F_\d,y\in F_\e$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i-2k,j}(\s(x),y) +\tb_{i+1-2k,j}(T_{1-2k}^{\e,\d}(x),y).$$ Define $d_1\in\Bil(F_\d)$ by $d_1(x,y)=\tb_{i,j-2k}(x,\s(y))+\tb_{i-2k,j}(\s(x),y)$ where $i,j\in I_\d,i+j=2k-2$. Using $b^\d(1\ot\s)=b^\d(\s\ot 1)$ we see that $d_1$ is independent of the choice of $i,j$. Define $d\in\Bil(F_\d)$ by $$d(x,y)=c^\d_{2k-2}(x,y)+d_1(x,y)+a^\e(T_{1-2k}^{\e,\d}(x),y) +a^\d(x,T_{1-2k}^{\e,\d}(y)).$$ We have $d(x,x)=0$ for $x\in F_\d$. (This follows from the choice of $\s$ and the identity $\tb_{i,j-2k}+\tb_{j,i-2k}=b^\d$ for $i,j\in I_\d,i+j=2k-2$.) Thus, $d$ is symplectic hence we can find $d'\in\Bil(F_\d)$ such that $d=d'+d'{}^*$. Since $b^\d$ is non-degenerate we can find $T_{2-2k}^{\d,\d}\in\End(F_\d)$ such that $d'(x,y)=b^\d(x,T_{2-2k}^{\d,\d}(y))$ for $x,y\in F_\d$. We have $$d(x,y)=b^\d(x,T_{2-2k}^{\d,\d}(y))+b^\d(y,T_{2-2k}^{\d,\d}(x))= b^\d(x,T_{2-2k}^{\d,\d}(y))+b^\d(T_{2-2k}^{\d,\d}(x),y).$$ Hence for $i,j\in I_\d,i+j=2k-2$ and $x,y\in F_\d$ we have $$\align&b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i,j-2k}(x,\s(y))+\tb_{i-2k,j}(\s(x),y)+ \tb_{i+1-2k,j}(T_{1-2k}^{\e,\d}(x),y)\\&+\tb_{i,j+1-2k}(x,T_{1-2k}^{\e,\d}(y)) +b_{i,j+2-2k}(x,T_{2-2k}^{\d,\d}(y))+b_{i+2-2k,j}(T_{2-2k}^{\d,\d}(x),y).\endalign$$ For $i,j\in I_\d,i+j=2k$ and $x,y\in F_\d$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+\tb_{i,j-2k}(x,\s(y))+\tb_{i-2k,j}(\s(x),y)=\tb_{ij}(x,y). $$ Define $f\in\Bil(F_\e)$ by $f(x,y)=c^\e_{2k-2}(x,y)$. Then $f$ is symplectic. (We use that $c^\e_{2k-2}$ is symplectic.) Hence we can find $f'\in\Bil(F_\e)$ such that $f=f'+f'{}^*$. Since $b^\e$ is non-degenerate we can find $T_{2-2k}^{\e,\e}\in\End(F_\e)$ such that $f'(x,y)=b^\e(x,T_{2-2k}^{\e,\e}(y))$ for $x,y\in F_\e$. We have $$f(x,y)=b^\e(x,T_{2-2k}^{\e,\e}(y))+b^\e(y,T_{2-2k}^{\e,\e}(x)) =b^\e(x,T_{2-2k}^{\e,\e}(y))+b^\e(T_{2-2k}^{\e,\e}(x),y)$$ hence for $i,j\in I_\e,i+j=2k-2$ and $x,y\in F_\e$ we have $$b_{ij}(x,y)=\tb_{ij}(x,y)+b_{i,j+2-2k}(x,T_{2-2k}^{\e,\e}(y)) +b_{i+2-2k,j}(T_{2-2k}^{\d,\d}(x),y).$$ For $i,j\in I_\d,i+j=2k$ and $x,y\in F_\e$ we have $b_{ij}(x,y)=\tb_{ij}(x,y)$. Let $T$ be as in (a) with $T_0^{00}=1,T_0^{11}=1$, $T_{1-2k}^{\e,\d},T_{2-2k}^{\e,\e}, T_{2-2k}^{\d,\d}$, $T_{-2k}^{\d,\d}=\s$ as above and the other components $0$. Define $(b'_{ij})$ by $T(b_{ij})=(b'_{ij})$. Then $(b'_{ij})$ has the required properties. $\sp$ This completes the proof of (b). We now verify the following special case of 3.6(c). (c) {\it Let $(\tb_{ij}),(b_{ij})$ be two points of $X$. Then there exists $T\in\D_u$ such that $T(b_{ij})=(\tb_{ij})$.} \nl We first prove the following statement by induction on $k\in[-n,0]$. ($P_k$) {\it Assume in addition that $b_{ij}=\tb_{ij}$ for any $i,j$ with $i+j\ge2k$. Then there exists $T\in\D_u$ such that $T(b_{ij})=(\tb_{ij})$.} \nl If $k=-n$ the result is obvious. Assume now that $k\in[1-n,0]$. By (b) we can find $T'\in\D_u$ such that $T'(b_{ij})=(b'_{ij})$ and $b'_{ij}=\tb_{ij}$ for $i+j\ge2k-2$. By the induction hypothesis we can find $T''\in\D_u$ such that $T''(b'_{ij})=(\tb_{ij})$. Let $T=T''T'\in\D_u$. Then $T(b_{ij})=(\tb_{ij})$. This completes the proof of $(P_k)$ for $k\in[-n,0]$. In particular $(P_0)$ holds and (c) is proved. \subhead 3.10. Proof of 3.6(c)\endsubhead Let $\lar,\lar'$ be two elements of the set $X$ in 3.6. We must show that $\lar,\lar'$ are in the same $U$-orbit. We argue by induction on $e$, the smallest integer $\ge0$ such that $N^e=0$. If $e=0$ we have $V=0$ and the result is obvious. If $e=1$ we have $N=0$. Then $V=\gr V_*$ canonically, $U=\{1\}$ and both $\lar,\lar'$ are the same as $\lar_0$ hence the result is clear. We now assume that $e\ge2$. $\sp$. Assume first that $p=2$. For $n\in\cl$ let $q_n:P^\nu_{-n}@>>>\kk$ be the quadratic forms attached to $(N,\lar)$ in 3.3 and let $q'_n:P^\nu_{-n}@>>>\kk$ be the analogous quadratic forms defined in terms of $(N,\lar')$. We show: (a) {\it there exists $T\in U$ such that if $\lar''\in\Symp(V)$ is given by $\la x,y\ra''=\la Tx,Ty\ra$ then for $n\in\cl$ the quadratic form $q''_n$ defined as in 3.3 in terms of $(N,\lar'')$ satisfies $q''_n=q'_n$.} \nl We are seeking an $S\in E_{\ge1}V_*$ such that $SN=NS$ and $\la(1+S)\dx,(1+S)N^{n-1}\dx\ra=\la\dx,N^{n-1}\dx\ra'$ that is, $\la(1+S)\dx,N^{n-1}(1+S)\dx\ra=\la\dx,N^{n-1}\dx\ra'$ that is, $\la S\dx,N^{n-1}\dx\ra+\la \dx,N^{n-1}S\dx\ra+\la S\dx,N^{n-1}S\dx\ra =\la\dx,N^{n-1}\dx\ra'+\la\dx,N^{n-1}\dx\ra$ \nl for any $n\in\cl$ and any $\dx\in V_{\ge-n}$ such that $N^{n+1}\dx=0$. Now $\la S\dx,N^{n-1}\dx\ra+\la\dx,N^{n-1}S\dx\ra=\la S\dx,(N^{n-1}+(N^\da)^{n-1})\dx\ra$ \nl is a linear combination of terms $\la S\dx,N^{n'}\dx\ra$ with $n'\ge n$; each of these terms is $0$ since $S\dx\in V_{\ge1-n},N^{n'}\dx\in V_{\ge2n'-n}$ and $1-n+2n'-n\ge1$. Moreover, $\la S\dx,N^{n-1}S\dx\ra=\la \bS x,\nu^{n-1}\bS x\ra_0$ where $x\in P^\nu_{-n}$ is the image of $\dx$ and $\la\dx,N^{n-1}\dx\ra'+\la\dx,N^{n-1}\dx\ra=q'_n(x)+q_n(x)$. By the surjectivity of the map $S\m\bS$ in 2.5(d), we see that it suffices to show that there exists $\s\in\End_1^\nu(\gr V_*)$ (that is $\s\in\End_1(\gr V_*)$ such that $\s\nu=\nu\s$) with $\la\s x,\nu^{n-1}\s x\ra_0=q'_n(x)+q_n(x)$ for any $n\in\cl$ and any $x\in P^\nu_{-n}$. For $n\in\cl'$ the last equation is automatically satisfied for any $\s$. (The left hand side is zero by the definition of $\cl'$. The right hand side is equal by 3.3(c) to $Q'_\nn(\nu^{(n-\nn)/2}x)+Q_\nn(\nu^{(n-\nn)/2}x)$ where $Q_\nn$ is the quadratic form attached as in 3.3 to $(N,\lar)$ and $Q'_\nn$ is the analogous quadratic form defined in terms of $(N,\lar')$. The last sum is zero since $Q_\nn=Q'_\nn=Q$.) We see that it suffices to show that there exists $\s\in\End_1^\nu(\gr V_*)$ such that $\la\s x,\nu^{n-1}\s x\ra_0=q'_n(x)+q_n(x)$ for any $n\in\cl-\cl'$ and any $x\in P^\nu_{-n}$. For $n\in\cl-\cl'$, the quadratic forms $q'_n,q_n$ have the same associated symplectic form (see 3.3(a)); hence there exists $\th_n\in\Hom(P^\nu_{-n},\kk)$ such that $q'_n(x)+q_n(x)=\th_n(x)^2$ for all $x\in P^\nu_{-n}$. Hence it suffices to show that the linear map $\r:\End_1^\nu(\gr V_*)@>>>\op_{n\in\cl-\cl'}\Hom(P^\nu_{-n},\kk)$ \nl given by $\s\m(\th_n)$ where $\th_n(x)=\sqrt{\la \s x,\nu^{n-1}\s x\ra_0}$ for $x\in P^\nu_{-n}$ is surjective. Let $\ce=\op_{n\ge0}\Hom(P^\nu_{-n},\gr_{1-n}V_*)$. We have an isomorphism $\p:\End_1^\nu(\gr V_*)@>\si>>\ce$ given by $\s\m(\s_n)$ where $\s_n\in\Hom(P^\nu_{-n},\gr_{1-n}V_*)$ is the restriction of $\s$. Define a linear map $\r':\ce@>>>\op_{n\in\cl-\cl'}\Hom(P^\nu_{-n},\kk)$ \nl by $(\s_n)\m(\th_n)$ where $\th_n(x)=\sqrt{\la \s_nx,\nu^{n-1}\s_nx\ra_0}$ for $x\in P^\nu_{-n}$. We have $\r'\p=\r$. Hence it suffices to show that $\r'$ is surjective. It also suffices to show that for any $n\in\cl-\cl'$ the linear map $\r'_n:\Hom(P^\nu_{-n},\gr_{1-n}V_*)@>>>\Hom(P^\nu_{-n},\kk)$ \nl given by $f\m\th$, where $\th(x)=\sqrt{\la fx,\nu^{n-1}fx\ra_0}$ for $x\in P^\nu_{-n}$, is surjective. Define $g\in\Hom(\gr_{1-n}V_*@>>>\kk)$ by $h\m\sqrt{\la h,\nu^{n-1}h\ra_0}$. Then $\r'_n(f)=g\circ f$ for $f\in\Hom(P^\nu_{-n},\gr_{1-n}V_*)$. Hence to show that $\r'_n$ is surjective it suffices to show that $g\ne0$. Since $n\in\cl-\cl'$, there exists $m$ odd such that $m\ge n+3$ and $b_m$ is not symplectic. Hence there exists $u'\in P^\nu_{-m}$ such that $\la u',\nu^mu'\ra_0\ne0$. We have $m=(n-1)+2k$ where $k$ is an integer $\ge2$. Let $u=N^ku'\in\gr_{1-n}V_*$ and $\la u,\nu^{n-1}u\ra_0=\la \nu^ku',\nu^{n-1+k}u'\ra_0=\la u',\nu^mu'\ra_0\ne0$. \nl Thus $g(u)\ne0$. We see that $g\ne0$, as required. This proves (a). Note that $\lar''$ in (a) is in $X$ (in fact in the $U$-orbit of $\lar$). Replacing if necessary $\lar$ by $\lar''$ we see that (b) {\it we may assume that $\lar,\lar'$ are such that $q_n=q'_n$ for all $n\in\cl$.} $\sp$ We now return to a general $p$. Let $r\ge e$. Let $F$ be a complement of $V_{\ge2-r}=\ker N^{r-1}$ in $V_{\ge1-r}=V$ and let $F'$ be a complement of $V_{\ge3-r}=\ker N^{r-2}+NV$ in $V_{\ge2-r}=\ker N^{r-1}$. Consider the linear map $\a$ of $F\op F'\op F\op\do\op F'\op F$ ($2r-1$ summands) into $V$ given by $(x_{1-r},x_{2-r},\do,x_{r-2},x_{r-1})\m$ $x_{1-r}+Nx_{3-r}+\do+N^{r-1}x_{r-1}+x_{2-r}+Nx_{4-r}+\do+N^{r-2}x_{r-2}$. \nl (Here $x_i\in F$ if $i=r+1\mod2$ and $x_i\in F'$ if $i=r\mod2$.) Let $W$ be the image of $\a$. We show that (c) $\lar$ and $\lar'$ are non-degenerate on $W$. \nl We prove this only for $\lar$; the proof for $\lar'$ is the same. Assume that $w=x_{1-r}+Nx_{3-r}+\do+N^{r-1}x_{r-1}+x_{2-r}+Nx_{4-r}+\do+N^{r-2}x_{r-2}$ with $x_i$ as above satisfies $\la w,W\ra=0$. We show that each $x_i$ is $0$. We have $0=\la w,N^{r-1}F\ra=\la x_{1-r},N^{r-1}F\ra=0$. Using the non-degeneracy of $b_{r-1}$ we see that $x_{1-r}=0$ and $w=Nx_{3-r}+\do+N^{r-1}x_{r-1}+x_{2-r}+Nx_{4-r}+\do+N^{r-2}x_{r-2}$. We have $0=\la w,N^{r-2}F'\ra=\la x_{2-r},N^{r-2}F'\ra$. Using the non-degeneracy of $b_{r-2}$ we see that $x_{2-r}=0$ and $w=Nx_{3-r}+\do+N^{r-1}x_{r-1}+Nx_{4-r}+\do+N^{r-2}x_{r-2}$. We have $0=\la w,N^{r-2}F\ra=\la Nx_{3-r},N^{r-2}F\ra=-\la x_{3-r},N^{r-1}F\ra$. Using the non-degeneracy of $b_{r-1}$ we see that $x_{3-r}=0$. Continuing in this way we see that each $x_i$ is $0$. This proves (c). The proof shows also that $\a$ is injective. Let $Z=\{x\in V;\la x,W\ra=0\},Z'=\{x\in V;\la x,W\ra'=0\}$. From (c) we see that $V=W\op Z=W\op Z'$. Clearly, $W$ is $N$-stable hence $(1+N)$-stable. Since $1+N$ is an isometry of $\lar$ it follows that $Z$ is $(1+N)$-stable hence $N$-stable. Similarly, $Z'$ is $N$-stable. Define $\Ph\in GL(V)$ by $\Ph(x)=x$ for $x\in W$, $\Ph(x)=x'$ for $x\in Z$ where $x'\in Z'$ is given by $x-x'\in W$. We have $\Ph\in U$ (see 2.7(c),(d)). Define ${}'\lar\in\Symp(V)$ by ${}'\la x,y\ra=\la \Ph(x),\Ph(y)\ra'$. By 3.6(a), we have ${}'\lar\in X$. Let ${}'Z=\{x\in V;{}'\la x,W\ra=0\}$. We show that ${}'Z=Z$. Let $x=x_1+x_2$ where $x_1\in W,x_2\in Z$. We have $x_2=w+x'_2$, $w\in W,x'_2\in Z'$. For $w'\in W$ we have $\la\Ph(x),w'\ra'=\la x_1+x'_2,w'\ra'=\la x_1,w'\ra'$. The condition that $\la\Ph(x),W\ra'=0$ is that $\la x_1,W\ra'=0$ or that $x_1=0$ (using (c)) or that $x\in Z$. Thus, ${}'Z=\{x\in V;\la \Ph(x),\Ph(W)\ra'=0\}=\{x\in V;\la \Ph(x),W\ra'=0\}=Z$ as required. $\sp$ In the case where $p=2$ we show that for any $n\in\cl$ the quadratic form $q'_n$ attached to $(N,\lar')$ as in 3.3 is equal to the analogous quadratic form attached to $(N,{}'\lar)$. We must show that, if $x\in V_{\ge-n}$, $N^{n+1}x=0$ then $\la\Ph x,\Ph N^{n-1}x\ra'=\la x,N^{n-1}x\ra'$ that is, $\la\Ph x,N^{n-1}\Ph x\ra'=\la x,N^{n-1}x\ra'$. Both sides are additive in $x$. We can write $x=x_1+x_2$ where $x_1\in W,x_2\in Z$ satisfy $x_1,x_2\in V_{\ge-n}$, $N^{n+1}x_1=0,N^{n+1}x_2=0$. We may assume that $x=x_1$ or $x=x_2$. When $x=x_1$ the desired equality is obvious. Hence we may assume that $x\in Z$. Write $x=x'+w$, $x'\in Z',w\in W$. We must show that $\la x+w,N^{n-1}x+N^{n-1}w\ra'=\la x,N^{n-1}x\ra'$ that is, $\la x,N^{n-1}w\ra'+\la w,N^{n-1}x\ra'+\la w,N^{n-1}w\ra'=0$ that is, $\la x,(N^{n-1}+(N^\da)^{n-1})w\ra'+\la w,N^{n-1}w\ra'=0$ that is, $\la x,N^nw\ra'+\la w,N^{n-1}w\ra'=0$ (we use $N^{n+1}w=0$) that is, $\la x'+w,N^nw\ra'+\la w,N^{n-1}w\ra'=0$ that is, $\la w,N^nw\ra'+\la w,N^{n-1}w\ra'=0$. Now $w\in W_{\ge1-n}$ (see 2.7(b)), $N^nw\in W_{\ge n+1}$ hence $\la w,N^nw\ra'=0$. It remains to show $\la w,N^{n-1}w\ra'=0$. Since $w\in W_{\ge1-n}$, $N^{n+1}w=0$, it suffices to show $\la y,\nu^{n-1}y\ra_0=0$ for any $y\in\gr_{1-n}V_*$ such that $\nu^{n+1}y=0$. This has already been seen in the proof in 3.3 that $q_n$ is well defined. $\sp$ Replacing $\lar'$ by ${}'\lar$ (which is in the same $U$-orbit) we see that condition (b) is preserved (for $p=2$). Thus, we may assume that $\lar,\lar'$ satisfy $Z=Z'$ and that for $p=2$ condition (b) holds. Thus $V=W\op Z$ is an othogonal decomposition with respect to either $\lar$ or $\lar'$. Let $\lar_W,\lar_Z$ be the restrictions of $\lar$ to $W,Z$. Let $\lar'_W,\lar'_Z$ be the restrictions of $\lar'$ to $W,Z$. Let $U_1$ (resp. $U_2$) be the analogue of $U$ for $W$ (resp. $Z$) defined in terms of $N$ and $W^N_*$ (resp. $Z^N_*$). We have naturally $U_1\T U_2\sub U$. We consider $5$ cases. {\it Case 1: $p\ne2$.} Take $r=e+1$. (Thus, $F=0$.) By the induction hypothesis, $\lar_Z$ is carried to $\lar'_Z$ by some $u_2\in U_2$. By 3.9, $\lar_W$ is carried to $\lar'_W$ by some $u_1\in U_1$. Then $\lar$ is carried to $\lar'$ by $(u_1,u_2)\in U$. $\sp$ {\it Case 2: $p=2$, $e$ is odd and $b_{e-2}$ is symplectic.} Take $r=e+1$. (Thus, $F=0$.) We have $e-1\in\cl$. The sets $\cl$ attached to $\lar_Z,\lar'_Z$ are the same as $\cl$ for $\lar,\lar'$. The quadratic forms attached to $\lar_Z$, $\lar'_Z$ and $n\in\cl-\{e-1\}$ are the same as those attached to $\lar,\lar'$ and $n$, hence they coincide. The quadratic forms attached to $\lar_Z,\lar'_Z$ and $n=e-1$ also coincide: they are both $0$. Hence the Quadratic forms attached to $\lar_Z,\lar'_Z$ coincide (see 3.3(c)). The quadratic forms attached to $\lar_W,\lar'_W$ coincide: for $e-1$ they are the same as those attached to $\lar,\lar'$ and $e-1$ and for other $n$ they are zero. Hence the Quadratic forms attached to $\lar_W,\lar'_W$ coincide. By the induction hypothesis, $\lar_Z$ is carried to $\lar'_Z$ by some $u_2\in U_2$. By 3.9, $\lar_W$ is carried to $\lar'_W$ by some $u_1\in U_1$. Then $\lar$ is carried to $\lar'$ by $(u_1,u_2)\in U$. {\it Case 3: $p=2$, $e$ is even and $b_{e-1}$ is symplectic.} Take $r=e+1$. (Thus, $F=0$.) The sets $\cl$ attached to $\lar_Z,\lar'_Z$ are the same as $\cl$ for $\lar,\lar'$. The quadratic forms attached to $\lar_Z,\lar'_Z$ and $n\in\cl$ are the same as those attached to $\lar,\lar'$ and $n$, hence they coincide. Hence the Quadratic forms attached to $\lar_Z,\lar'_Z$ coincide. By the induction hypothesis, $\lar_Z$ is carried to $\lar'_Z$ by some $u_2\in U_2$. By 3.9, $\lar_W$ is carried to $\lar'_W$ by some $u_1\in U_1$. Then $\lar$ is carried to $\lar'$ by $(u_1,u_2)\in U$. {\it Case 4: $p=2$, $e$ is even and $b_{e-1}$ is not symplectic.} Take $r=e$. The sets $\cl$ attached to $\lar_Z,\lar'_Z$ are the same as $\cl$ for $\lar,\lar'$. The quadratic forms attached to $\lar_Z,\lar'_Z$ and $n\in\cl$ are the same as those attached to $\lar,\lar'$ and $n$, hence they coincide. Hence the Quadratic forms attached to $\lar_Z,\lar'_Z$ coincide. By the induction hypothesis, $\lar_Z$ is carried to $\lar'_Z$ by some $u_2\in U_2$. By 3.9, $\lar_W$ is carried to $\lar'_W$ by some $u_1\in U_1$. Then $\lar$ is carried to $\lar'$ by $(u_1,u_2)\in U$. {\it Case 5: $p=2$, $e$ is odd, $\ge3$ and $b_{e-2}$ is not symplectic.} Take $r=e$. By 3.9, $\lar_W$ is carried to $\lar'_W$ by some $u_1\in U_1$. Replacing $\lar'$ by a translate under $(u_1,1)\in U$ we see that we may assume in addition that $\lar_W=\lar'_W$. Let $\tW=F+NF+\do+N^{r-1}F$. Let $W'=\{w\in W;\la w,\tW\ra=0\}=\{w\in W;\la w,\tW\ra'=0\}$. Then $W=\tW\op W'$, orthogonal direct sum for both $\lar,\lar'$. Let $\tZ=W'\op Z$, orthogonal direct sum for both $\lar,\lar'$. Then $V=\tW\op\tZ$, orthogonal direct sum for both $\lar,\lar'$. Let $\lar_{\tZ},\lar'_{\tZ}$ be the restrictions of $\lar,\lar'$ to $\tZ$. Let $\tU_1$ (resp. $\tU_2$) be the analogue of $U$ for $\tW$ (resp. $\tZ$) defined in terms of $N$ and $\tW^N_*$ (resp. $\tZ^N_*$). We have naturally $\tU_1\T\tU_2\sub U$. The sets $\cl$ attached to $\lar_{\tZ}$, $\lar'_{\tZ}$ are the same as $\cl$ for $\lar,\lar'$. The quadratic forms attached to $\lar_{\tZ},\lar'_{\tZ}$ and $n\in\cl$ are the same as those attached to $\lar,\lar'$ and $n$, hence they coincide. Hence the Quadratic forms attached to $\lar_{\tZ},\lar'_{\tZ}$ coincide. By the induction hypothesis, $\lar_{\tZ}$ is carried to $\lar'_{\tZ}$ by some $\tu_2\in\tU_2$. Then $\lar$ is carried to $\lar'$ by $(1,\tu_2)\in U$. $\sp$ This completes the proof of 3.6(c) hence also that of 3.5, 3.6, 3.7. \subhead 3.11\endsubhead Here is the order of the proof of the various assertions in Propositions 3.5-3.7: 3.5(a) (see 3.8); 3.7(a) (see 3.7); 3.6(a) (see 3.6); 3.7(b) (see 3.7); 3.6(b) (see 3.6); 3.5(b) (see 3.5); 3.6(c) (see 3.9, 3.10); 3.7(c) (see 3.7); 3.5(c) (see 3.5). \subhead 3.12\endsubhead Let $V\in\cc$ and let $\lar\in\Symp(V)$. The following result can be deduced from \cite{\SPA, I, 2.10}. Let $C,C_0$ be two $GL(V)$-conjugacy classes in $\Nil(V)$ such that $C\cap\cm_{\lar}\ne\em,C_0\cap\cm_{\lar}\ne\em$ and $C$ is contained in the closure of $C_0$ in $GL(V)$. Then $C\cap\cm_{\lar}$ is contained in te closure of $C_0\cap\cm_{\lar}$ in $\cm_{\lar}$. \subhead 3.13\endsubhead Let $V\in\cc$ and let $\lar\in\Symp(V)$. Let $G=Sp(\lar)$. For any self-dual filtration $V_*$ of $V$ and for $n\ge1$ let $E^{\lar}_{\ge n}V_*=E_{\ge n}V_*\cap\cm_{\lar}$, a unipotent algebraic group with multiplication $T*T'=T+T'+TT'$. Let $$\tix(V_*)=\x(V_*)\cap\cm_{\lar}=\{N\in\cm_{\lar};V^N_*=V_*\} =\{N\in E^{\lar}_{\ge2}V_*;\bN\in\End_2^0(\gr V_*)\}$$ (see 2.9). The following three conditions are equivalent: (i) $\tix(V_*)\ne\em$; (ii) there exists $\nu\in\End_2^0(\gr V_*)$ which is skew-adjoint with respect to the symplectic form on $\gr V_*$ induced by $\lar$; (iii) $\dim\gr_nV_*=\dim\gr_{-n}V_*\ge\dim\gr_{-n-2}V_*$ for all $n\ge0$ and $\dim\gr_{-n}V_*=\dim\gr_{-n-2}V_* \mod2$ for all $n\ge0$ even. \nl We have (i)$\imp$(ii) by the definition of $\tix(V_*)$; we have (ii)$\imp$(iii) by 2.3(d) and 3.1(c). Now (iii)$\imp$(ii) is easily checked. We have (ii)$\imp$(iii) by 3.5(a). Let $\fF_{\lar}$ be the set of all self-dual filtrations $V_*$ of $V$ that satisfy (i)-(iii). From the definitions we have a bijection (a) $\fF_{\lar}@>\si>>D_G,V_*\m\l$ \nl ($D_G$ as in 1.1) where $\l=(G^\l_0\sps G^\l_1\sps G^\l_2\sps\do)$ is defined in terms of $V_*$ by $G^\l_0=E_{\ge0}V_*\cap G$ and $G^\l_n=1+E_{\ge n}^{\lar}V_*$ for $n\ge1$. The sets $\tix(V_*)$ (with $V_*\in\fF_{\lar})$ form a partition of $\cm_{\lar}$. (If $N\in\cm_{\lar}$ we have $N\in\tix(V_*)$ where $V_*=V^N_*$.) Let $V_*\in\fF_{\lar}$. Let $C_0$ be the unique $GL(V)$-conjugacy class in $\Nil(V)$ that contains $\x(V_*)$. We have $E_{\ge2}^{\lar}V_*-\tix(V_*)=(E_{\ge2}V_*-\x(V_*))\cap\cm_{\lar} =E_{\ge2}V_*\cap(\cup_CC)\cap\cm_{\lar}$ \nl (the last equality follows from 2.9; $C$ runs over all $GL(V)$-orbits in $\Nil(V)$ such that $C\sub\bC_0-C_0$). Using 3.12 we see that $E_{\ge2}^{\lar}V_*-\tix(V_*)=E_{\ge2}^{\lar}V_*\cap(\cup_C(C\cap\cm_{\lar}))$ \nl where $C$ runs over all $GL(V)$-orbits in $\Nil(V)$ such that $C\cap\cm_{\lar}\ne\em$ and $C\sub\ov{C_0\cap\cm_{\lar}}-(C_0\cap\cm_{\lar})$. We see that, if $V_*\m\l$ (as in (a)) and $\bla$ is the $G$-orbit of $\l$ in $D_G$ then (with notation of 1.1) $\tH^\bla$ is the union of $G$-conjugacy classes in $1+\cm_{\lar}$ contained in $1+\bC_0$, $H^\bla$ is the union of $G$-conjugacy classes in $1+\cm_{\lar}$ contained in $1+C_0$, $X^\l=1+\tix(V_*)=1+(E_{\ge2}^{\lar}V_*\cap C_0)$. We see that $\fP_1-\fP_3$ hold. \subhead 3.14\endsubhead We preserve the setup of 3.13. Let $V_*\in\fF_{\lar}$ and let $\l\in D_G$ be the corresponding element. Define $\lar_0\in\Symp(\gr V_*)$ as in 3.2. The map $E_{\ge2}^{\lar}V_*@>>>\End_2^0(\gr V_*),N\m\bN$ restricts to a map $\p:\tix(V_*)@>>>E:=\{\nu\in\End_2^0(\gr V_*);\nu\text{ skew-adjoint with respect to } \lar_0\}$. \nl We show: (a) {\it The group $E_{\ge3}^{\lar}V_*$ (see 3.13) acts freely on $\tix(V^*)$ by $T,N\m T*N$ (see 3.13) and the orbit space of this action may be identified with $E$ via $\p$.} \nl We show this only at the level of sets. If $T\in E_{\ge3}^{\lar}V_*,N\in\tix(V_*)$ then $T*N\in E_{\ge2}^{\lar}V_*$ and $T*N,N$ induce the same map in $\End_2(\gr V_*)$; hence $T*N\in\tix(V_*)$. Thus $T,N\m T*N$ is an action of $E_{\ge3}^{\lar}V_*$ on $\tix(V_*)$. This action is free: it is the restriction of the action of $E_{\ge3}^{\lar}V_*$ on $E_{\ge2}^{\lar}V_*$ by left multiplication for the group structure in 3.13. If $N,N'\in\tix(V_*)$ induce the same map in $\End_2^0(\gr V_*)$ then $N'-N\in E_{\ge3}V_*$. Set $T=(N'-N)(1+N)\i\in E_{\ge3}V_*$. Then $(1+T)(1+N)=1+N'$ and we have automatically $T\in E^{\lar}_{\ge3}V_*$ and $T+N=N'$. Thus the orbits of the $E_{\ge3}^{\lar}V_*$-action on $\ti\x(V^*)$ are exactly the non-empty fibres of $\p$. It remains to show that $\p$ is surjective. This follows from 3.5(a). Now let $N,N'\in\tix(V_*)$ be such that $\bN=\bN'=\nu\in\End_2^0(\gr V_*)$. We show: (b) {\it there exists $g\in E_{\ge0}V_*\cap G$ such that $N'=gNg\i$.} \nl Assume first that $p=2$. The set $\cl\sub2\NN$ defined in 3.3 in terms of $N$ is the same as that defined in terms of $N'$. Let $q_n:P^\nu_{-n}@>>>\kk$ be the quadratic form defined in terms of $N$ (for $n\in\cl$) as in 3.3 and let $q'_n:P^\nu_{-n}@>>>\kk$ be the analogous quadratic form defined in terms of $N'$. From 3.3(a) we see that for any $n\in\cl$ there exists an automorphism $h_n:P^\nu_{-n}@>>>P^\nu_{-n}$ which preserves the symplectic form $x,y\m b_n(x,y)$ (see 3.1) and satisfies $q'_n(x)=q_n(h_nx)$ for any $x\in P^\nu_{-n}$. There is a unique $h\in Sp(\lar_0)$ such that $h(x)=h_n(x)$ for $x\in P^\nu_{-n},n\in\cl$, $h(x)=x$ for $x\in P^\nu_{-n}$, $n\in\ZZ-\cl$, $h\nu=\nu h$. Let $V=\op_aV_a$ be a direct sum decomposition as in 3.2(b). Then $\End_0(V)$ is defined and we define $\tih\in\End_0(V)$ by the requirement that for any $a$, $\tih:V_a@>>>V_a$ corresponds to $h:\gr_aV_*@>>>\gr_aV_*$ under the obvious isomorphism $V_a@>\si>>\gr_aV_*$. Then $\tih\in E_{\ge0}V_*\cap G$ and $\tih N\tih\i=N''$ where $N''\in E^{\lar}_{\ge2}V_*$ satisfies $\bN''=\nu$. Moreover, the quadratic form $P^\nu_{-n}@>>>\kk$ defined as in 3.3 in terms of $N''$ (instead of $N$) for $n\in\cl$ is $x\m h_n(x)$ that is, $q'_n$. From 3.3(c) we see that the Quadratic form $Q_n$ defined for $n\in\cl'$ in terms of $N''$ is the same as that defined in terms of $N'$. From 3.5(c) we see that there exists $h'\in1+E^{\lar}_{\ge1}V_*$ such that $h'N''h'{}\i=N'$. Setting $g=h'\tih\in E_{\ge0}V_*\cap G$ we have $gNg\i=N'$. Next assume that $p\ne2$. From 3.5(c) we see that there exists $g\in1+E^{\lar}_{\ge1}V_*$ such that $gNg\i=N'$. This proves (b). We see that $\fP_6$ (hence $\fP_4$) holds. From (a) we see that the $G^\l_0$-action on $\tix(V^*)$ (conjugation) induces an action of $\bG^\l_0=G^\l_0/G^\l_1$ on $E$ and from (b) we see that this gives rise to a bijection between the set of $G^\l_0$-orbits on $\tix(V^*)$ and the set of $\bG^\l_0$-orbits on $E$. We describe this last set of orbits. We identify $\bG^\l_0$ with $\End_0(\gr V_*)\cap Sp(\lar_0)$ with the action on $E$ given by $g:\nu\m\nu'$ where $\nu'(x)=g\nu(g\i x)$ for $x\in\gr V_*$. Let $I=\{n\in2\NN+1,\dim\gr_{-n}V_*-\dim\gr_{-n-2}V_*\in\{2,4,6,\do\}\}$. For any subset $J\sub I$ let $E_J$ be the set of all $\nu\in E$ such that for any $n\in I$ we have $\{x\in\gr_{-n}V_*;\nu^{n+1}x=0,\la x,\nu^nx\ra_0\ne0\}\ne\em\lra n\in J$. \nl Let $\bE$ be the set of all direct sum decompositions $\gr V_*=\op_{n\ge0}W^n$ where $W^n\in\bcc$ (see 2.1) are such that $\la W^n,W^{n'}\ra_0=0$ for $n\ne n'$ and for $n\ge 0$, $\dim W^n_a$ is $\dim\gr_{-n}V_*-\dim\gr_{-n-2}V_*$ if $a\in[-n,n],a=n\mod2$ and is $0$ for other $a$. Define $\ph:E@>>>\bE$ by $\nu\m(W^n)$ where $W^n=\sum_{k\ge0}\nu^kP^\nu_{-n}$. Then $\ph$ is $\bG^\l_0$-equivariant where $\bG^\l_0$ acts on $\bE$ in an obvious way (transitively). Let $w=(W^n)\in\bE$. Let $G^w$ be the stabilizer of $w$ in $\bG^\l_0$. Let $E^w=\ph\i(w)$. Now $E^w$ may be identified with $\prod_{n\ge0}E^w_n$ where $E^w_n$ is the set of all skew-adjoint elements in $\End_2^0(W^n)$ with respect to $\lar_0|_{W^n}$. Moreover $G^w$ may be identified with $\prod_{n\ge0}G^w_n$ where $G^w_n=\End_0(W^n)\cap Sp(\lar_0|_{W^n})$. Furthermore, we may identify $E^w_n=E^{w1}_n\T E^{w2}_n$ where $E^{w1}_n$ consists of all sequences of isomorphisms (c) $W^n_{-n}@>\si>>W^n_{-n+2}@>\si>>W^n_{-n+4}@>\si>>\do@>\si>>W^n_{-\d}$ \nl ($\d=0$ if $n$ is even and $\d=1$ if $n$ is odd) and $E^{w2}_n$ is the set of non-degenerate symmetric bilinear forms $W^n_{-1}\T W^n_{-1}@>>>\kk$ (if $n$ is odd) and is a point if $n$ is even. (This identification is obtained by attaching to $\nu\in E^w_n$ the isomorphisms (c) induced by $\nu$ and if $n$ is odd, the bilinear form $x,x'\m\la x,\nu x'\ra_0$ on $W^n_{-1}$.) We claim that if $p=2$, the subsets $E_J$ are precisely the orbits of $\bG^\l_0$ on $E$ while if $p\ne2$, $E$ is a single orbit of $\bG^\l_0$. Using the transitivity of the $\bG^\l_0$ action on $\bE$ we see that it suffices to prove: if $p=2$, the subsets $E^w_J=E_J\cap E^w$ are precisely the $G^w$-orbits on $E$ while if $p\ne2$, $E^w$ is a single $G^w$-orbit. If $n\n I$, $G'_n$ acts transitively on $E'_n$. If $n\in I$, $pr_2:E^w_n@>>>E^{w2}_n$ induces a bijection between the set of $G^w_n$-orbits on $E^w_n$ and the set of $GL(W^n_{-1})$-orbits on the set of non-degenerate symmetric bilinear forms on $W^n_{-1}$. The last set of orbits consists of one element if $p\ne2$ and of two elements (the symplectic forms and the non-symplectic forms) if $p=2$. This verifies our claim. We see that the first assertion of $\fP_8$ holds. As above, we identify $E$ with the set of triples $(w,\a,j)$ where $w\in\bE$, $\a$ is a collection of isomorphisms as in (c) (for each $n\ge0$) and $j$ is a sequence $(j_n)_{n\in I}$ where $j_n\in\Bil(W^n_{-1})$ is symmetric non-degenerate. Assume that $p=2$. Let $J\sub J'\sub I$. From the previous discussion we see that the $\bG^\l_0$-orbits on $E$ that contain $E_J$ in their closure and are contained in the closure of $E_{J'}$ are those of the form $E_K$ where $J\sub K\sub J'$. Let $E_{J,J'}=\cup_{K;J\sub K\sub J'}E_K$. We identify $E_J$ with the set of $(w,\a,j)\in E$ such that $j_n$ is not symplectic for $n\in J$ and symplectic for $n\in I-J$. We identify $E_{J,J'}$ with the set of $(w,\a,j)\in E$ such that $j_n$ is not symplectic for $n\in J$ and symplectic for $n\in I-J'$. Let $\tE_J$ be the set of all triples $(w,\a,\tj)$ where $w,\a$ are as above and $\tj=(\tj_n)_{n\in I}$ where for $n\in J$, $\tj_n\in\Bil(W^n_{-1})$ is a symmetric non-symplectic non-degenerate form and, for $n\in I-J$, $\tj_n:W^n_{-1}\T W^n_{-1}@>>>\kk$ is the square of a symplectic non-degenerate form. Now $E_J,E_{J,J'},\tE_J$ are naturally algebraic varieties. Define a finite bijective morphism $\s:E_J@>>>\tE_J$ by $(w,\a,j)\m(w,\a,\tj)$ where $\tj_n=j_n$ for $n\in J$, $\tj_n=j_n^2$ for $n\in I-J$. Define $\r:E_{J,J'}@>>>\tE_J$ by $(w,\a,j)\m(w,\a,\tj)$ where $\tj_n=j_n$ for $n\in J$ and $\tj_n(x,x')=j_n(x,x')^2+j_n(x,x)j_n(x',x')$ for $n\in I-J$, $x,x'\in W^n_{-1}$. (To see that this is well defined, we must check that for $n\in I-J$, the symplectic form $x,x'\m j_n(x,x')+\sqrt{j_n(x,x)j_n(x',x')}$ on $W^n_{-1}$ is non-degenerate. Let $R$ be the radical of this symplectic form. Let $H=\{x\in W^n_{-1};j_n(x,x)=0\}$. If $x\in R\cap H$, then $j_n(x,x')$ for all $x'$ hence $x=0$. Thus, $R\cap H=0$. Since $H$ is either $W^n_{-1}$ or a hyperplane in $W^n_{-1}$, we see that $R\cap H$ is either $R$ or a hyperplane in $R$. It follows that $\dim R$ is $0$ or $1$. Since $R=\dim W^n_{-1}\mod2$ we see that $\dim R$ is even. Hence $R=0$, as required.) Taking here $J'=I$, we see that $\fP_7$ holds. We now return to a general $J'$. We consider the fibre $\cf$ of $\r$ at $(w,\a,\tj)\in\tE_J$. We may identify $\cf$ with the set of all collections $(j_n)_{n\in I-J}$ where $j_n\in\Bil(W^n_{-1})$ is symmetric non-degenerate for all $n$, $j_n$ is symplectic for $n\in I-J'$ and $\tj_n(x,x')=j_n(x,x')^2+j_n(x,x)j_n(x',x')$ for $n\in I-J$, $x,x'\in W^n_{-1}$. Let $\cf'$ be the set of all collections $(h_n)_{n\in I-J}$ where $h_n$ is a linear form $W^n_{-1}@>>>\kk$, zero for $n\in I-J'$. We define a map $\cf@>>>\cf'$ by $(j_n)_{n\in I-J}\m(h_n)_{n\in I-J}$ where $h_n(x)=\sqrt{j_n(x,x)}$ for $x\in W^n_{-1}$. We define a map $\cf'@>>>\cf$ by $(h_n)_{n\in I-J}\m(j_n)_{n\in I-J}$ where $j_n(x,x')=\sqrt{\tj_n(x,x')}+h_n(x)h_n(x')$ for $x,x'\in W^n_{-1}$. (We show that this is well defined. We must show that $j_n$ given by the last equality is non-degenerate. Let $R'$ be the radical of $j_n$. Define $v\in W^n_{-1}$ by $h_n(y)=\sqrt{\tj_n(v,y)}$ for all $y\in W^n_{-1}$. If $x\in R',y\in W^n_{-1}$, we have $\sqrt{\tj_n(x,y)}=h_n(x)h_n(y)=h_n(x)\sqrt{\tj_n(v,y)}$ hence $\sqrt{\tj_n(x-h_n(x)v,y)}=0$. Since $\sqrt{\tj_n}$ is non-degenerate we have $x-h_n(x)v=0$. Hence $x=h_n(x)v=h_n(h_n(x)v)v=h_n(x)h_n(v)v$. This is $0$ since $h_n(v)=\sqrt{\tj_n(v,v)}=0$. Thus $R'=0$.) Clearly, $\cf@>>>\cf',\cf'@>>>\cf$ are inverse to each other. We see that $\cf$ is in natural bijection with a vector space of dimension $\sum_{n\in J'-J}c_n$ where $c_n=\dim W^n_{-1}$. Hence if $\kk,q$ are as in $\fP_5$, we have $\sum_{K;J\sub K\sub J'}|E_K(\FF_q)|=|E_{J,J'}(\FF_q)| =\prod_{n\in J'-J}q^{c_n}|E_J(\FF_q)|$. \nl From this we see that $|E_K(\FF_q)|=\prod_{n\in K}(q^{c_n}-1)|E_\em(\FF_q)|$ for any $K\sub I$. Using this and $\fP_6$ we see that the second assertion of $\fP_8$ holds. For $\kk,q$ as in $\fP_5$ we have $|H^\bla(\FF_q)|=|X^\l(\FF_q)||G(\FF_q)/G^\l_0(\FF_q)|$, $|X^\l(\FF_q)|=q^{\dim G^\l_3}|E(\FF_q)|$. \nl Hence to verify $\fP_5$ it suffices to show that $|E(\FF_q)|$ is a polynomial in $q$ with integer coefficients independent of $p$. Using the $\bG^\l_0$-equivariant fibration $\ph:E@>>>\bE$ we see that $|E(\FF_q)|=|\bE(\FF_q)||E^w(\FF_q)|$ for any $w\in\bE$. Since $|\bE(\FF_q)|$ is a polynomial in $q$ with integer coefficients independent of $p$, it suffices to show that for any $w\in\bE(\FF_q)$, $|E^w(\FF_q)|$ is a polynomial in $q$ with integer coefficients independent of $p$, or that $|E^w_n(\FF_q)|$ is a polynomial in $q$ with integer coefficients independent of $p$ for any $w\in\bE(\FF_q)$ and any $n\ge0$. Using the identification $E^w_n=E^{w1}_n\T E^{w2}_n$ and the fact that $|E^{w1}_n(\FF_q)|$ is a polynomial in $q$ with integer coefficients independent of $p$, we see that it suffices to show that $|E^{w2}_n(\FF_q)|$ is a polynomial in $q$ with integer coefficients independent of $p$. Thus it suffices to check the following statement. {\it Let $W$ be an $\FF_q$-vector space of dimension $d$. Let $b(W)$ be the set of non-degenerate symmetric bilinear forms $W\T W@>>>\FF_q$. Then $|b(W)|$ is a polynomial in $q$ with integer coefficients independent of $p$.} \nl We argue by induction on $d$. For $d=0$ the result is obvious. Assume that $d\ge1$. We write $|b(W)|=f(d,q)$. The set of all symmetric bilinear forms $W\T W@>>>\FF_q$ has cardinal $q^{d(d+1)/2}$; it is a disjoint union $\sqc_Xb_X(W)$ where $X$ runs over the linear subspaces of $W$ and $b_X(W)$ is the set of symmetric bilinear forms $W\T W@>>>\FF_q$ with radical equal to $X$. Thus, $q^{d(d+1)/2}=\sum_X|b_X(W)|=\sum_X|b(W/X)|=\sum_{d'\in[0,d]}g(d,d',q)f(d-d',q)$ \nl where $g(d,d',q)=|\{X\sub W,\dim X=d'\}|$. We see that $f(d,q)=q^{d(d+1)/2}-\sum_{d'\in[1,d]}g(d,d',q)f(d-d',q)$. \nl Since $g(d,d',q)$ is a polynomial in $q$ with integer coefficients independent of $p$ and the same holds for $f(d-d',q)$ with $d'\in[1,d]$ (by the induction hypothesis) it follows that $f(d,q)$ is as required. We see that $\fP_5$ holds. \head 4. The group $A^1(u)$\endhead \subhead 4.1\endsubhead In this section we assume that $p\ge2$ and that $\fP_1$ holds. Let $u\in\cu$. According to $\fP_1$ there is a unique $\l\in D_G$ such that $u\in X^\l$. Let $A^1(u)=Z_{G^\l_1}(u)/Z_{G^\l_1}(u)^0$, a finite $p$-group. The image of $A^1(u)$ in $Z_G(u)/Z_G(u)^0$ is a normal subgroup (since $Z_G(u)=Z_{G^\l_0}(u)$, see 1.1(c), and $Z_{G^\l_1}(u)$ is normal in $Z_{G^\l_0}(u)$). In this section we describe the finite group $A^1(u)$ in some examples assuming that $p=2$ and $G$ is a symplectic group. Let $n\ge 1$. Let $I=\{i\in[-n,n];i=n\mod 2\}$. Let $F\in\cc,F\ne0$. Let $V=\op_{i\in I}F_i$ where $F_i=F$. Define $N:V@>>>V$ by $(x_i)\m(x'_i)$ where $x'_i=x_{i-2}$ for $i\in I-\{-n\},x'_{-n}=0$. We fix $\lar_0\in\Symp(V)$ such that $\la(x_i),(y_i)\ra_0=\sum_{i\in I}b(x_i,y_{-i})$ where $b\in\Bil(F)$ satisfies $b^*=b$, $b$ is non-degenerate and $b\in\Symp(F)$ if $n$ is even. Let $\lar\in\Symp(V)$ be such that $\la Nx,y\ra+\la x,Ny\ra+\la Nx,Ny\ra=0$ for $x,y\in V$ and $\la x,y\ra=\la x,y\ra_0$ if there exists $i$ such that $x_j=0$ for $j\ne i$ and $y_j=0$ for $j\ne-i$. We have $\la (x_i),(y_i)\ra=\sum_{i,j\in I}b_{ij}(x_i,y_j)$ where $b_{ij}\in\Bil(F)$ are such that $b_{i-2,j}+b_{i,j-2}+b_{ij}=0$ if $i,j\in I-\{-n\}$, $b_{i,-i}=b$ for all $i\in I$, $b_{ii}\in\Symp(F)$ for all $i\in I$, $b_{ij}^*=-b_{ji}$ for all $i,j\in I$, \nl (We have automatically $b_{ij}=0$ if $i+j\ge 1$.) \nl Let $\D'=\{T\in GL(V);TN=NT,\la x,y\ra=\la Tx,Ty\ra\qua\frl x,y\in V\}$, a subgroup of $Sp(\lar)$; equivalently, $\D'$ is the set of linear maps $T:V@>>>V$ of the form $T:(x_i)\m(x'_i),x'_i=\sum_{j\in I;j\le i}T_{i-j}x_j$ \nl where $T_r\in\End(F)$ $(r\in\{0,2,4,\do,2n\})$ are such that $$b_{ij}(x,y)=\sum_{i',j'\in I;i'\ge i,j'\ge j}b_{i'j'}(T_{i'-i}(x),T_{j'-j}y) \tag $E_{ij}$ $$ for $i,j\in I,i+j\le0$ and $x,y\in F$. Now $(E_{ij}),(E_{i+2,j-2})$ with $i+j=2k$ are equivalent if $(E_{ab})$ is assumed for $a+b=2k+2$ (the sum of those two equations is just $E_{i+2,j}$). Thus the conditions that $T$ must satisfy are $E_{ii}$ and $E_{i-2,i}$. Setting $x=y$ in these equations we obtain equations $(E_{ii}^0)$, $(E_{i-2,i}^0)$. Note that the equation $(E_{ii}^0)$ is $0=0$ hence can be omitted; the equation $(E_{ii})$ is a consequence of $(E_{i-2,i}^0)$ (if it is defined). Hence the equations satisfied by the components of $T$ are as follows: $$(E_{-2,0}^0),(E_{-2,0}),(E_{-4,-2}^0),(E_{-4,-2}),\do,(E_{-n,-n+2}^0), (E_{-n,-n+2}),(E_{-n,-n})\tag a$$ (for $n$ even), $$\align&(E_{-1,1}),(E_{-3,-1}^0),(E_{-3,-1}),(E_{-5,-3}^0),(E_{-5,-3}),\do, (E_{-n,-n+2}^0),\\&(E_{-n,-n+2}),(E_{-n,-n})\tag b\endalign$$ (for $n$ odd). Assume first that $n$ is even. The solutions $T_0$ of the first equation in (a) form an even orthogonal group, a variety with two connected components. For any such $T_0$ the solutions $T_2$ of the second equation in (a) form an affine space of dimension independent of $T_0$. For any $T_0,T_2$ already determined, the solutions $T_4$ of the third equation in (a) form an affine space of dimension independent of $T_0,T_2$. Continuing in this way we see that the solutions of the equations (a) form a variety with two connected components. Moreover, the solutions in which $T_0$ is specified to be $1$ form a connected variety. Assume next that $n$ is odd and $b$ is symplectic. The solutions $T_0$ of the first equation in (b) form a symplectic group (a connected variety). For any such $T_0$ the solutions $T_2$ of the second equation in (b) form an affine space of dimension independent of $T_0$. For any $T_0,T_2$ already determined, the solutions $T_4$ of the third equation in (b) form an affine space of dimension independent of $T_0,T_2$. Continuing in this way we see that the solutions of the equations (b) form a connected variety. Moreover, the solutions in which $T_0$ is specified to be $1$ form a connected variety. One can show that, if $n$ is odd, $n\ge3$ and $b$ is not symplectic, the solutions of the equations (b) form a variety with two connected components. Moreover, the solutions in which $T_0$ is specified to be $1$ form a disconnected variety. In solving the equations above we use repeatedly the statement (c) below. Let $\fQ$ be the vector space of quadratic forms $F@>>>\kk$. Define linear maps $a_1,a_2,a_3$ as follows: $a_1:\End(F)@>>>\fQ(F)$ is $\t\m q,q(x)=b(x,\t(x))$; $a_2:\{\t\in\End(F);b(\t(x),y)=b(x,\t(y)\qua\frl x,y\in F\}@>>>\Hom(F,\kk)$ is $\t\m\th,\th(x)=\sqrt{b(x,\t(x))}$; $a_3:\{b'\in\Bil(F);b'{}^*=b'\}@>>>\Hom(F,\kk)$ is $b'\m\th,\th(x)=\sqrt{b'(x,x)}$. \nl Then (c) $a_1,a_2,a_3$ are surjective. \nl For $a_3$ this is clear. Consider now $a_2$. Let $\th\in\Hom(F,\kk)$. By (c) for $a_3$ we can find $b'\in\Bil(F),b'{}^*=b'$ such that $\th(x)=\sqrt{b'(x,x)}$. We can find a unique $\t\in End(F)$ such that $b(x,\t(y))=b'(x,y)$. Then $a_2(t)=\th$. Consider now $a_1$. Let $q\in\fQ$. Let $b^0$ be a symplectic form on $F$. We can write $b^0=d+d^*$ where $d\in\Bil(F)$. We can write $d(x,y)=b(x,\s(y))$ for some $\s\in\End(F)$. Then $b(x,\s(y))+b(y,\s(x))=b^0(x,y)$. Apply this to the symplectic form $b^0(x,y)=q(x+y)+q(x)+q(y)$. Then $b(x+y,\s(x+y))+b(x,\s(x))+b(y,\s(y))=b(x,\s(y))+b(y,\s(x))=q(x+y)+q(x)+q(y)$. \nl Hence $b(x,\s(x))+q(x)=\th(x)^2$ for some $\th\in\Hom(F,\kk)$. By (c) for $a_2$ we can find $\t\in\End(F)$ such that $b(x,\t(x))=\th(x)^2$. Then $b(x,\s(x))+b(x,\t(x))=q(x)$ that is $b(x,(\s+\t)(x))=q(x)$. Thus $a_1$ is surjective. This proves (c). \subhead 4.2\endsubhead Let $V,\lar$ be as in 3.2. Let $N\in\cm_{\lar},V_*=V_*^N$. Let $e$ be as in 2.4. We show: (a) {\it If $W,W'$ are $e$-special subspaces of $V$ (see 2.8) then there exists $g\in1+E^{\lar}_{\ge1}V_*$ such that $g(W)=W'$, $gN=Ng$.} \nl By 2.8(b) we can find $g_1\in1+E_{\ge 1}V_*$ such that $g_1(W)=W'$, $g_1N=Ng_1$. Then $g_1$ carries $\lar$ to a symplectic form $\lar'$ which induces the same symplectic form as $\lar$ on $\gr V_*$ and has the same associated quadratic forms as $\lar$ (see 3.6(b)). By the proof in 3.10 (case 2 and 3) we see that there exists $g_2\in1+E_{\ge 1}V_*$ such that $g_2(W')=W'$, $g_2N=Ng_2$ and $g_2$ carries $\lar'$ to $\lar$. Then $g=g_2g_1$ has the required properties. \subhead 4.3\endsubhead Let $V,\lar,N,V_*,e$ be as in 4.2. (a) {\it If $\la x,Nx\ra=0$ for any $x\in V_{\ge-1}$, then $\cv:=\{g\in E^{\lar}_{\ge1}V_*;gN=Ng\}$ is connected. Hence $A^1(1+N)=\{1\}$.} \nl We argue by induction on $e$. Let $\cx$ be the set of all $e$-special subspaces (see 2.8) of $V$. By 2.8(b) the group $\{g\in 1+E_{\ge 1}V_*;gN=Ng\}$ acts transitively on $\cx$. This group is connected (it may be identified as variety with the vector space $\{\x\in E_{\ge 1}V_*;\x N=N\x\}$); hence $\cx$ is connected. By 4.2(a), $\cv$ acts transitively on $\cx$. Since $\cx$ is connected, it suffices to show that the stabilizer $\cv_W$ of some $W\in\cx$ in $\cv$ is connected. This stabilizer is $\cv'\T\cv''$ where $\cv',\cv''$ are defined like $\cv$ in terms of $W,W^\pe$ instead of $V$. By results in 4.2, $\cv'$ is connected. By the induction hypothesis applied to $W^\pe$, $\cv''$ is connected. Hence $\cv'\T\cv''$ is connected. Hence $\cv$ is connected. \head 5. Study of the varieties $\cb_u$\endhead \subhead 5.1\endsubhead We assume that $\kk=\kk_p$. We say that an algebraic variety $V$ over $\kk$ has the {\it purity property} if for some/any $\FF_q$-rational structure on $V$ (where $\FF_q$ is a finite subfield of $\kk$) with Frobenius map $F:V@>>>V$ and any $n\in\ZZ$, any complex absolute value of any eigenvalue of $F^*:H^n_c(V,\bbq)@>>>H^n_c(V,\bbq)$ is $q^{n/2}$. In this section we show that for certain $u\in\cu$ the varieties $\cb_u$ (see 0.1) have the purity property. We assume that properties $\fP_1-\fP_4,\fP_6,\fP_7$ hold for $G$. Let $\l\in D_G$. Let $\Pi^\l$ be the (finite) set of orbits for the conjugation action of $G^\l_0$ on $\cb$. Let $\bcb=\{B\in\cb;B\sub G^\l_0\}$. For any $\co\in\Pi^\l$ define a morphism $\x^\co:\co@>>>\bcb$ by $B\m(B\cap G^\l_0)G^\l_1$. We show: (a) {\it The fibres of $\x^\co:\co@>>>\bcb$ are exactly the orbits of $G^\l_1$ acting on $\co$ by conjugation.} \nl If $B,B'\in\co$, $\x^\co(B)=\x^\co(B')$, then $B'=g\i Bg$ with $g\in G^\l_0$, $(B'\cap G^\l_0)G^\l_1=(B\cap G^\l_0)G^\l_1=g\i(B\cap G^\l_0)G^\l_1g$. Hence $g\in(B\cap G^\l_0)G^\l_1$. Writing $g=g'g'',g'\in B\cap G^\l_0$, $g''\in G^\l_1$, we have $B'=g\i Bg=g''{}\i Bg''$. This proves (a). Let $Y^\l=\{(u,B)\in X^\l\T\cb;u\in B\}$. We have a partition $Y^\l=\cup_{\co\in\Pi^\l}Y^\l_\co$ where $Y^\l_\co=\{(u,B)\in X^\l\T\co;u\in B\}$. Let $\co\in\Pi^\l$. We show: (b) {\it$Y^\l_\co$ is smooth.} \nl Let $\tB\in\co$. Let $Y'=\{(u,g)\in X^\l\T G^\l_0;g\i ug\in\tB\cap X^\l\}$. We have a fibration $Y'@>>>Y^\l_\co$ with smooth fibres isomorphic to $G^\l_0\cap\tB$. Hence it suffices to show that $Y'$ is smooth. Let $Y''=(\tB\cap X^\l)\T G^\l_0$. Define $Y'@>\si>>Y''$ by $(u,g)\m(g\i ug,g)$. It suffices to show that $Y''$ is smooth, or that $\tB\cap X^\l$ is smooth. But $\tB\cap X^\l$ is open in $\tB\cap G^\l_2$ which is smooth, being an algebraic group. This proves (b). For any $\b\in\bcb$ let $\cg_\b^\co=((B\cap G^\l_2)G^\l_3)/G^\l_3$ where $B\in\co$ is such that $\x^\co(B)=\b$. Note that $\cg_\b^\co$ is a closed connected subgroup of $G^\l_2/G^\l_3$, independent of the choice of $B$. (To verify the last statement it suffices, by (a), to show that for $B$ as above and $v\in G^\l_1$ we have $(vBv\i\cap G^\l_2)G^\l_3=(B\cap G^\l_2)G^\l_3$. This follows from 1.1(b).) Now $G^\l_0$ acts on $\bcb$ and on $G^\l_2/G^\l_3$ by conjugation. From the definitions we see that for $g\in G^\l_0$ and $\b\in\bcb$ we have $\cg_{g\b g\i}^\co=g\cg_\b^\co g\i$. Let $\bY^\l_\co=\{(x,\b)\in\bX^\l\T\bcb;x\in\cg_\b^\co\}$. We show: (c) {\it$\bY^\l_\co$ is a closed smooth subvariety of $\bX^\l\T\bcb$.} \nl Let $\tB\in\co$. We have a fibration $X^\l\T G^\l_0@>>>\bX^\l\T\bcb$, $(u,g)\m(\p^\l(u),\x^\co(g\tB g\i))$ with smooth fibres. It suffices to show that the inverse image of $\bY^\l_\co$ under this fibration is a closed smooth subvariety of $X^\l\T G^\l_0$, or that $\{(u,g)\in X^\l\T G^\l_0;g\i ug\in X^\l\cap((\tB\cap G^\l_2)G^\l_3)\}$ \nl is a closed smooth subvariety of $X^\l\T G^\l_0$, or that $(X^\l\cap((\tB\cap G^\l_2)G^\l_3))\T G^\l_0$ is a closed smooth subvariety of $X^\l\T G^\l_0$ or that $X^\l\cap((\tB\cap G^\l_2)G^\l_3)$ is a closed smooth subvariety of $X^\l$. It is closed since $(\tB\cap G^\l_2)G^\l_3$ is closed in $G^\l_2$. It is smooth since it is an open subset of $(\tB\cap G^\l_2)G^\l_3$ which is smooth, being an algebraic group. We show: (d) {\it The morphism $a:Y^\l_\co@>>>\bY^\l_\co,(u,B)\m(\p^\l(u),\x^\co(B))$ is a fibration with fibres isomorphic to an affine space of a fixed dimension.} \nl Clearly, $a$ is surjective. Let $(u,B)\in Y^\l_\co$. Let $Z:=a\i(a(u,B)) =\{(u',B');u=u'f,B'=vBv\i\text{ for some }v\in G^\l_1,f\in G^\l_3;u'\in B'\}$. \nl We show only that $Z$ is isomorphic to an affine space of fixed dimension. Let $\tZ=\{(f,v)\in G^\l_3\T G^\l_1;v\i uf\i v\in B\}$. Then $Z=\tZ/(B\cap G^\l_1)$ where $B\cap G^\l_1$ acts freely on $\tZ$ by $b:(f,v)\m(f,vb\i)$. Since conjugation by $G^\l_1$ acts trivially on $G^\l_2/G^\l_3$, the map $(f,v)\m(f',v),f'=u\i v\i uf\i v$ is an isomorphism $$\align&\tZ@>>>\tZ'=\{(f',v)\in G^\l_3\T G^\l_1;uf'\in B\}= \{(f',v)\in G^\l_3\T G^\l_1;f'\in B\}\\&=(G^\l_3\cap B)\T G^\l_1\endalign$$ (we use $u\in B$) and we have $Z=(G^\l_3\cap B)\T G^\l_1/(B\cap G^\l_1)$. Now $G^\l_3\cap B,G^\l_1,B\cap G^\l_1$ are connected unipotent groups of dimension independent of $B$, for $B\in\co$. (The connectedness follows from the fact that these unipotent groups are normalized by a maximal torus of $G$ contained in $G^\l_0\cap B$. The fact that the dimension does not depend on $B$ follows from the fact that $G^\l_1,G^\l_3$ are normalized by $G^\l_0$.) We see that $Z$ is an affine space of constant dimension. We now fix $x\in\bX^\l$. Let $\Si=(\p^\l)\i(x)\sub X^\l$. Let $\cb_\Si=\{(u,B)\in\Si\T\cb;u\in B\}$. We have $\cb_\Si=\sqc_{\co\in\Pi^\l}\co_\Si$ where $\co_\Si=\{(u,B)\in\Si\T\co;u\in B\}$. Let $\co\in\Pi^\l$. Let $\bcb^\co_x=\{\b\in\bcb;x\in\cg^\co_\b\}$. We show: (e) {\it$\bcb^\co_x$ is a closed subvariety of $\bcb$ and $a':\co_\Si@>>>\bcb^\co_x$, $(u,B)\m\x^\co(B)$ is a fibration with fibres isomorphic to an affine space of a fixed dimension.} \nl Let $\tB\in\co,u_0\in\Si$. We have a locally trivial fibration $G^\l_0@>>>\bcb$, $g\m\x^\co(g\tB g\i)$. To show that $\bcb^\co_x$ is closed it suffices to show that its inverse image under this fibration is closed in $G^\l_0$, or that $\{g\in G^\l_0;g\i u_0g\in(\tB\cap G^\l_2)G^\l_3\}$ is closed in $G^\l_0$. This is clear since $(\tB\cap G^\l_2)G^\l_3$ is closed in $G^\l_2$. The second assertion of (e) follows from (d) using the cartesian diagram $$\CD \co_\Si@>a'>>\bcb^\co_x\\ @VVV @VVV\\ Y^\l_\co@>a>>\bY^\l_\co \endCD$$ where the left vertical map is the obvious inclusion and the right vertical map is $\b\m(x,\b)$. (f) {\it If the closure of the $G^\l_0$-orbit in $G^\l_2$ of some/any $u\in\Si$ is a subgroup $\G$ of $G^\l_2$ then $\bcb^\co_x$ is smooth.} \nl Let $\tB\in\co,u_0\in\Si$. As in the proof of (e) it suffices to show that $\{g\in G^\l_0;g\i u_0g\in(\tB\cap G^\l_2)G^\l_3\}$ is smooth. This variety is a fibration over $R=(G^\l_0-\text{conjugacy class of }u_0)\cap((\tB\cap G^\l_2)G^\l_3)$ with smooth fibres isomorphic to $Z_{G^\l_0}(u_0)$ (via $g\m g\i u_0g$). Hence it suffices to show that $R$ is smooth. From our assumption we see that $R$ is open in $\G\cap((\tB\cap G^\l_2)G^\l_3)$ which is smooth being an algebraic group. This proves (f). \mpb Note that the hypothesis of (f) holds at least in the case where the $G^\l_0$-conjugacy class of some/any $u\in\Si$ is open dense in $G^\l_2$. We show (g) {\it If the hypothesis of (f) holds then $\cb_\Si$ has the purity property.} \nl From (e),(f) we see that $\bcb^\co_x$ is a smooth projective variety of pure dimension. From \cite{\DE} it then follows that $\bcb^\co_x$ has the purity property. From this and (e) we see that for $\co\in\Pi^\l$, $\co_\Si$ has the purity property. Using this and the partition $\cb_\Si=\sqc_{\co\in\Pi^\l}\co_\Si$, we see that (g) holds. \subhead 5.2\endsubhead Let $\bZ(x)=\{\bg\in\bG^\l_0;\bg x=x\bg\}$. Let $\tZ(x)$ be the inverse image of $\bZ(x)$ under $G^\l_0@>>>\bG^\l_0$. Thus we have $G^\l_1\sub\tZ(x)$ and $\tZ(x)/G^\l_1@>\si>>\bZ(x)$. Note that the inverse image of $\bZ(x)^0$ is $\tZ(x)^0$ and we have $\tZ(x)^0/G^\l_1@>\si>>\bZ(x)^0$. Now $\tZ(x)$ acts transitively (by conjugation) on $\Si$. (Indeed, let $u,u'\in\Si$. By $\fP_6$ we can find $g\in G^\l_0$ such that $u'=gug\i$. We have automatically $g\in\tZ(x)$.) Since $\Si$ is irreducible, it follows that $\tZ(x)^0$ acts transitively (by conjugation) on $\Si$. Let $u\in\Si$. Recall that $\cb_u=\{B\in\cb;u\in B\}$. Let $Z'_G(u)=Z_G(u)\cap\tZ(x)^0$. Since $Z_G(u)\sub\tZ(x)$, see 1.1(c), we see that $Z'_G(u)$ is a normal subgroup of $Z_G(u)$ containing $Z_G(u)^0$. Let $A'(u)$ be the image of $Z'_G(u)$ in $A(u):=Z_G(u)/Z_G(u)^0$; this is a normal subgroup of $A(u)$. We have $Z'_G(u)/Z_{G^\l_1}(u)@>\si>>\bZ(x)^0$. Hence $Z'_G(u)=Z_{G^\l_1}(u)Z'_G(u)^0=Z_{G^\l_1}(u)Z_G(u)^0$. It follows that {\it$A'(u)$ is the image of the obvious homomorphism} $A^1(u)@>>>A(u)$. \nl Now $Z_G(u)$ acts by conjugation on $\cb_u$; this induces an action of $A(u)$ on $H^n(\cb_u,\bbq)$ which restricts to an $A'(u)$-action on $H^n(\cb_u,\bbq)$. Assume that the hypothesis of 5.1(f) holds and that $A'(u)$ acts trivially on $H^n_c(\cb_u,\bbq)$ for any $n$. We show: (a) {\it$\cb_u$ has the purity property.} \nl Define $f:\cb_\Si@>>>\Si$ by $(g,B)\m g$. For any $n$, $R^nf_!(\bbq)$ is an equivariant constructible sheaf for the transitive $\tZ(x)^0$ action on $\Si$; hence it is a local system on $\Si$ corresponding to a representation of $A'(u)$ (the group of components of the isotropy group of $u$ in $\tZ(x)^0$) on $H^n_c(\cb_u,\bbq)$. This representation is trivial hence $R^nf_!(\bbq)$ is a constant local system. Since $\Si$ is an affine space of dimension say $d$ we see that $H^a_c(\Si,R^nf_!(\bbq))$ is $H^n_c(\cb_u,\bbq)(-d)$ if $a=2d$ and is zero if $a\ne2d$. It follows that the standard spectral sequence $E_2^{a,n}=H^a_c(\Si,R^nf_!(\bbq))\Rightarrow H^{a+n}_c(\cb_\Si,\bbq)$ \nl is degenerate. Hence the purity property of $\cb_\Si$ (see 5.1(g)) implies that any complex absolute value of any eigenvalue of the Frobenius map on $E_2^{2d,n}=H^n_c(\cb_u,\bbq)(-d)$ \nl is $q^{d+n/2}$. Hence any complex absolute value of any eigenvalue of the Frobenius map on $H^n_c(\cb_u,\bbq)$ is $q^{n/2}$. This proves (a). \subhead 5.3\endsubhead Since the hypothesis of 5.1(f) is not satisfied in general, we seek an alternative way to prove purity. Let $\g$ be the $\bG^\l_0$-orbit of $x$ in $\bX^\l$. Let $\hag@>\r>>\g_1@<\s<<\g$ be as in $\fP_7$. Let $\Xi=\r\i(\s(x))$. Let $\bcb^\co_\Xi=\{(x',\b)\in\bY^\l_\co;x'\in\Xi\}$, a closed subvariety of $\bY^\l_\co$. We show: (a) {\it$\bcb^\co_\Xi$ is smooth of pure dimension.} \nl Let $\b_0\in\bcb$. Let $\cg_0=\cg^\co_{\b_0}$. It suffices to show that the inverse image of $\bcb^\co_\Xi$ under the fibration $\Xi\T\bG^\l_0@>>>\Xi\T\bcb$, $(x',\bg)\m(x',\bg\b_0\bg\i)$ (with smooth connected fibres) is smooth of pure dimension, or that $\fS:=\{(x',\bg)\in\Xi\T\bG^\l_0;\bg\i x'\bg\in\cg_0\}$ is smooth of pure dimension. The morphism $f:\fS@>>>\hag\cap\cg_0$, $(x',\bg)\m\bg\i x'\bg$ is smooth with fibres of pure dimension. (We show only that for any $y\in\hag\cap\cg_0$, the fibre $f\i(y)$ is isomorphic to $\{\bg\in\bG^\l_0;\bg x\bg\i=x\}$ which is smooth of pure dimension. We have $$\align&f\i(y)=\{(x',\bg)\in\Xi\T\bG^\l_0;\bg\i x'\bg=y\}\cong \{\bg\in\bG^\l_0;\bg y\bg\i\in\Xi\}\\&=\{\bg\in\bG^\l_0;\r(\bg y\bg\i)=\s(x)\}= \{\bg\in\bG^\l_0;\bg\s\i(\r(y))\bg\i=x\}\endalign$$ and it remains to use the transitivity of the $\bG^\l_0$-action on $\g$.) It suffices to show that $\hag\cap\cg_0$ is empty or smooth, connected. Now $\hag$ is open in $G^\l_2/G^\l_3$ hence $\hag\cap\cg_0$ is open in $\cg_0$ which is connected and smooth (being an algebraic group). We show: (b) {\it Assume that for any $\co\in\Pi^\l$ there is a $\kk^*$-action on $\bcb^\co_\Xi$ which is a contraction to the projective subvariety $\bcb^\co_x$. Then $\cb_\Si$ has the purity property.} \nl Consider an $\FF_q$-rational structure on $G$ such that $G^\l_a$ is defined over $\FF_q$ for any $a$ and $\co,x,\Xi$ are defined over $\FF_q$. Let $\z$ be an eigenvalue of Frobenius on $H^n(\bcb^\co_x,\bbq)$. By \cite{\DEII, 3.3.1}, any complex absolute value of $\z$ is $\le q^{n/2}$ (since $\bcb^\co_x$ is projective). Our assumption implies that the inclusion $\bcb^\co_x\sub\bcb^\co_\Xi$ induces for any $n$ an isomorphism $H^n(\bcb^\co_\Xi,\bbq)@>\si>>H^n(\bcb^\co_x,\bbq)$. Hence $\z$ is also an eigenvalue of Frobenius on $H^n(\bcb^\co_\Xi,\bbq)$. Since $\bcb^\co_\Xi$ is smooth of pure dimension say $d$, it satisfies Poincar\'e duality; hence $q^d\z\i$ is an eigenvalue of Frobenius on $H^{2d-n}_c(\bcb^\co_\Xi,\bbq)$. By \cite{\DEII, 3.3.1} applied to $\bcb^\co_\Xi$, we see that any complex absolute value of $q^d\z\i$ is $\le q^{(2d-n)/2}$ hence any complex absolute value of $\z$ is $\ge q^{n/2}$. It follows that any complex absolute value of $\z$ is $q^{n/2}$. We see that $\bcb^\co_x$ has the purity property. (This argument is similar to one of Springer in \cite{\SP}.) From this and 5.1(e) we see that for $\co\in\Pi^\l$, $\co_\Si$ has the purity property. Using this and the partition $\cb_\Si=\sqc_{\co\in\Pi^\l}\co_\Si$, we see that $\cb_\Si$ has the purity property. \mpb If we assume in addition that $A'(u)$ acts trivially on $H^n_c(\cb_u,\bbq)$ for any $n$ we see as in 5.2 that $\cb_u$ has the purity property. \subhead 5.4\endsubhead Let $V,\lar$ be as in 3.2. Assume that $p=2$ and that $G=Sp(\lar)$. Let $u\in\cu$. We set $u=1+N,V_*=V_*^N$. Assume that (a) $\la x,Nx\ra=0$ for any $x\in V_{\ge-1}$. \nl We set $$\G=1+\{N'\in E_{\ge2}^{\lar}V_*;\la x,N'x\ra=0\qua\frl x\in V_{\ge-1}\}.$$ Now $\G$ is a subgroup of $1+E_{\ge2}^{\lar}V_*$. (Assume that $1+N',1+N''\in\tG_2$. Let $x\in V_{\ge-1}\}$. We have $\la x,N'x\ra=0,\la x,N''x\ra=0$. We must show that $\la x,(N'+N''+N'N'')x\ra=0$ or that $\la x,N'N''x\ra=0$. This follows from $N'N''x\in V_{\ge3}$ and $3-1\ge1$.) Clearly, $\G$ is normal in $G^\l_0$. Since $\G$ is a closed unipotent subgroup normalized by $G^\l_0$, it must be connected. Now $\cj:=1+\{N'\in\tG_2;\bN'\in\End_2^0(\gr V_*)\}$ \nl is open in $\G$ since it is the inverse image under $\G@>>>\End_2(\gr V_*),1+N'\m\bN'$ of the open subset $\End_2^0(\gr V_*)$ of $\End_2(\gr V_*)$. Also $\cj\ne\em$ since $1+N\in\cj$. Hence $\cj$ is an open dense subset of $\G$. By results in 3.14, $\cj$ is the $G^\l_0$-conjugacy class of $1+N$. We see that the hypothesis of 5.1(f) holds. Using 5.2(a) we see that: (b) {\it$\cb_u$ has the purity property for any $u\in G$ whose conjugacy class is minimal in the unipotent piece containing it, see 1.1, and such that any Jordan block of even size appears an even number times.} \nl (For such $u$, $A'(u)$ is trivial by 4.3(a).) Alternatively, one can show that for $u$ as in (b) the method of 5.3 is applicable (the hypothesis of 5.3(b) holds) and one obtains another proof of (b). \widestnumber\key{DLP} \Refs \ref\key{\DLP}\by C.De Concini,G.Lusztig,C.Procesi\paper Homology of the zero set of a nilpotent vector field on the flag manifold\jour J.Amer.Math.Soc.\vol1\yr1988\pages 15-34\endref \ref\key{\DE}\by P.Deligne\paper La conjecture de Weil,I\jour Publ.Math.IHES\vol43\yr 1974\pages273-308\endref \ref\key{\DEII}\by P.Deligne\paper La conjecture de Weil,II\jour Publ.Math.IHES\vol52 \yr1980\pages137-252\endref \ref\key{\EN}\by H.Enomoto\paper The conjugacy classes of Chevalley groups of type $G_2$ over finite fields of characteristic $2$ or $3$\jour J. Fac. Sci. Univ. Tokyo,I \vol16\yr1970\pages497-512\endref \ref\key{\KA}\by N.Kawanaka\paper Generalized Gelfand-Graev representations of exceptional simple algebraic groups over a finite field\jour Invent.Math.\vol84\yr1986 \pages575-616\endref \ref\key{\KO}\by B.Kostant\paper The principal three dimensional subgroup and the Betti numbers of a complex simple Lie group\jour Amer.J.Math.\vol81\yr1959\pages973-1032 \endref \ref\key{\LC}\by G.Lusztig\paper A class of irreducible representations of a Weyl group \jour Proc.Kon.Nederl.Akad. (A)\vol82\yr1979\pages323-335\endref \ref\key{\LN}\by G.Lusztig\paper Notes on unipotent classes\jour Asian J.Math.\vol1\yr 1997\pages194-207\endref \ref\key{\LS}\by G.Lusztig and N.Spaltenstein\paper On the generalized Springer correspondence for classical groups\inbook Algebraic groups and related topics, Adv.Stud.Pure Math.6\publ North Holland and Kinokuniya\yr1985\pages289-316\endref \ref\key{\MI}\by K.Mizuno\paper The conjugate classes of unipotent elements of the Chevalley groups $E_7$ and $E_8$\jour Tokyo J.Math\vol3\yr1980\pages391-459\endref \ref\key{\SH}\by K.Shinoda\paper The conjugacy classes of Chevalley groups of type $F_4$ over finite fields of characteristic $2$\jour J. Fac. Sci. Univ. Tokyo,I\vol21\yr 1974\pages133-159\endref \ref\key{\SPA}\by N.Spaltenstein\book Classes unipotentes et sous-groupes de Borel, LNM 946\publ Springer Verlag\yr1980\endref \ref\key{\SPAII}\by N.Spaltenstein\paper On the generalized Springer correspondence for exceptional groups\inbook Algebraic groups and related topics, Adv.Stud.Pure Math.6 \publ North Holland and Kinokuniya\yr1985\pages317-338\endref \ref\key{\SP}\by T.A.Springer\paper A purity result for fixed point sets varieties in flag manifolds\jour J. Fac. Sci. Univ. Tokyo,IA\vol31\yr1984\pages 271-282\endref \ref\key{\WA}\by G.E.Wall\paper On the conjugacy classes in the unitary, symplectic and orthogonal groups\jour J.Austral.Math.Soc.\vol3\yr1963\pages1-62\endref \endRefs \enddocument
{ "timestamp": "2005-04-03T19:28:01", "yymm": "0503", "arxiv_id": "math/0503739", "language": "en", "url": "https://arxiv.org/abs/math/0503739" }
\section{Introduction} Thompson's groups $F$, $T$ and $V$ are a remarkable family of infinite, finitely-presentable groups studied for their own properties as well as for their connections with questions in logic, homotopy theory, geometric group theory and the amenability of discrete groups. Cannon, Floyd and Parry give an excellent introduction to these groups in \cite{cfp}. These three groups can be viewed either algebraically, combinatorially, or analytically. Algebraically, each has both finite and infinite presentations. Geometrically, an element in each group can be viewed as a {\em tree pair diagram}; that is, as a pair of finite binary rooted trees with the same number of leaves, with a numbering system pairing the leaves in the two trees. Analytically, an element of each group can be viewed as a piecewise-linear self map of the unit interval: \begin{itemize} \item in $F$ as a piecewise linear homeomorphism, \item in $T$ as a homeomorphism of the unit interval with the endpoints identified, and thus of $S^1$, \item in $V$ as a right-continuous bijection which is locally orientation preserving. \end{itemize} Thompson's group $F$ in particular has been studied extensively. The group $F$ has a standard infinite presentation in which every element has a unique normal form, and a standard two-generator finite presentation. Fordham \cite{blakegd} presented a method of computing the word length of $w \in F$ with respect to the standard finite generating set directly from a tree pair diagram representing $w$. Regarding $F$ as a diagram group, Guba \cite{gubagrowth} also obtained an effective geometric method for computing the word metric with respect to the standard finite generating set. Belk and Brown \cite{belkbrown} have similar results which arise from viewing elements of $F$ as forest diagrams. In this paper, we discuss analogues for $T$ of some properties of $F$, using all three of the descriptions of $T$: algebriac, geometric and analytic. We begin by desribing unique normal forms for elements which arise from their reduced tree pair descriptions. We consider metrically how $F$ is contained as a subgroup of $T$, and show that the number of carets in a reduced tree pair diagram representing $w \in T$ estimates the word length of $w$ with respect to a particular generating set. Thus $F$ is quasi-isometrically embedded in $T$. Furthermore, we show that there are families of words in $F$ which are isometrically embedded in $T$ with respect to an alternate finite generating set. The groups $T$ and $V$, unlike $F$, contain torsion elements, and we describe how to recognize these torsion elements from their tree pair diagrams. Finally, we show that every torsion element of $T$ is conjugate to a power to a generators of $T$ and that the subgroup of rotations in $T$ is quasi-isometrically embedded. \section{Background on Thompson's groups $F$ and $T$} \subsection{Presentations and tree pair diagrams} Thompson's groups $F$ and $T$ both have representations as groups of piecewise-linear homeomorphisms. The group $F$ is the group of orientation-preserving homeomorphisms of the interval $[0,1]$, where each homeomorphism is required to have only finitely many discontinuities of slope, called {\em breakpoints}, have slopes which are powers of two and have the coordinates of the breakpoints all lie in the set of dyadic rationals. Similarly, the group $T$ consists of orientation-preserving homeomorphisms of the circle $S^1$ satisfying the same conditions where we represent the circle $S^1$ as the unit interval $[0,1]$ with the two endpoints identified. Cannon, Floyd and Parry give an excellent introduction to Thompson's groups $F$, $T$ and $V$ in \cite{cfp}. We refer the reader to this paper for full details on results mentioned in this section. Since more readers have some familiarity with $F$ than with $T$, we first give a very brief review of the group $F$, and then a slightly more detailed review of $T$. Algebraically, $F$ has well known infinite and finite presentations. With respect to the infinite presentation $$ \langle x_i, i\geq 0\, |\,x_jx_i=x_ix_{j+1}, i<j\rangle $$ for $F$, group elements have simple normal forms which are unique. It is easy to see that $F$ can be generated by $x_0$ and $x_1$, which form the standard finite generating set for $F$, and yield the finite presentation $$ \langle x_0,x_1\,|\,[x_0x_1^{-1},x_0^{-1}x_1x_0],[x_0x_1^{-1},x_0^{-2}x_1x_0^2]\rangle. $$ A geometric representation for an element $w$ in $F$ is a tree pair diagram, as discussed in \cite{cfp}. A {\em tree pair diagram} is a pair of finite rooted binary trees with the same number of leaves. By convention, the leaves of each tree are thought of as being numbered from $0$ to $n$ reading from left to right. A node of the tree together with its two downward directed edges is called a {\em caret}. The {\em left side} of the tree consists of the root caret, and all carets connected to the root by a path of left edges; the {\em right side} of the tree is defined analogously. A caret is called a {\em left caret} if its left leaf lies on the left side of the tree. A caret is called a {\em right caret} if it is not the root caret and its right leaf lies on the right side of the tree. All other carets are called {\em interior}. A caret is called {\em exposed} if it contains two leaves of the tree. For $w \in F$, we write $w = (T_-,T_+)$ to express $w$ as a tree pair diagram, and refer to $T_-$ as the {\em source} tree and $T_+$ as the {\em target} tree. These trees arise naturally from the interpretation of $F$ as a group of homeomorphisms. Thinking of $w$ as a homeomorphism of the unit interval, the source tree represents a subdivision of the domain into subintervals of width $1/2^n$ for varying values of $n$, and the target tree represents another such a subdivision of the range. The homeomorphism then maps the $i^{th}$ subinterval in the domain linearly to the $i^{th}$ subinterval in the range. A tree pair diagram representing $w$ in $F$ is not unique. A new diagram can always be produced from a given tree pair diagram representing $w$ simply by adding carets to the $i$th leaf of both trees. We impose a natural reduction condition: if $w = (T_-,T_+)$ and both trees contain a caret with two exposed leaves numbered $i$ and $i+1$, then we remove these carets, thus forming a representative for $w$ with fewer carets and leaves. A tree pair diagram which admits no such reductions is called a {\em reduced tree pair diagram}, and any element of $F$ is represented by a unique reduced tree pair diagram. When we write $w = (T_-,T_+)$ below, we are assuming that the tree pair diagram is reduced unless otherwise specified. The group $T$ also has both a finite and an infinite presentation. The infinite presentation is given by two families of generators, $\{x_i,i\ge 0\}$, the same generators as in the infinite presentation of $F$, a family $\{c_i,i\ge0\}$ of torsion elements, and the following relators: \begin{enumerate} \item $x_jx_i=x_ix_{j+1}$, if $i<j$ \item $x_kc_{n+1}=c_{n}x_{k+1}$, if $k<n$ \item $c_nx_0=c_{n+1}^2$ \item $c_n=x_nc_{n+1}$ \item$c_n^{n+2}=1$. \end{enumerate} This new family of generators $c_n$ (of order $n+2$), is simple to describe. The generator $c_n$ corresponds to the homeomorphism of the circle obtained as follows. Both domain and range can be thought of as the unit interval with the endpoints identified. We subdivide the interval into $n+1$ subintervals by successively halving the rightmost subinterval; or in other words inserting endpoints at $\frac{1}{2}, \frac{3}{4}, \ldots ,\frac{2^{n+1}-1}{2^{n+1}}$. Then the homeomorphism maps $[0,1/2]$ linearly to $[\frac{2^{n}-1}{2^{n}},\frac{2^{n+1}-1}{2^{n+1}}]$, and so on around each circle. For example, the element $c$ corresponds to the homeomorphism of $S^1$ given by $$ c(t)=\left\{\begin{array} {ll} \frac12t+\frac34&\text{if }0\le t<\frac12\\ 2t-1&\text{if }\frac12\le t<\frac34\\ t-\frac14&\text{if }\frac34\le t\le1 \end{array}\right. $$ Figure \ref{fig:c1c2} shows the graphs of the homeomorphisms corresponding to $c_1$ and $c_2$. \begin{figure}[h] \includegraphics[width=5in]{c1c2}\\ \caption{ The graphs of the homeomorphisms corresponding to the elements $c_1$ and $c_2$.\label{fig:c1c2}} \end{figure} Using the first three relators, we see that only the generators $x_0$, $x_1$ and $c_1$ are required to generate the group, since the other generators can be obtained from these three. In the following, we will use $c$ to denote the generator $c_1$. The group $T$ is finitely presented using the following relators,with respect to the finite generating set $\{x_0,x_1,c\}$: \begin{enumerate} \item $[x_0x_1^{-1},x_0^{-1}x_1x_0]=1$ \item $[x_0x_1^{-1},x_0^{-2}x_1x_0^2]=1$ \item $x_1c_3=c_2x_2$, (that is $x_1(x_0^{-2}cx_1^{-2})=(x_0^{-1}cx_1)(x_0^{-1}x_1x_0)$) \item $c_1x_0=c_2^2$, (that is, $cx_0=(x_1^{-1}cx_0)^2$) \item $x_1c_2=c$, (that is, $x_1(x_0^{-1}cx_1)=c) $ \item $c^3=1$. \end{enumerate} As with Thompson's group $F$, we will frequently work with the more convenient infinite set of generators when constructing normal forms for elements and performing computations in the group. We will need to express elements with respect to a finite generating set when discussing word length. There are two natural finite generating sets for $T$, both extending the standard finite generating set for $F$. The first and the one that we use primarily below is the generating set $\{x_0, x_1, c_1\}$ used in the finite presentation above. In Section \ref{isomembed} for the purposes of counting carets carefully, we also use the generating set $\{x_0, x_1, c_0\}$, which has the advantage that the tree pair diagram for $c_0$ has only one caret, as opposed to $c_1$, which has two carets, at the expense of slightly more complicated relators. Just as for $F$, tree pair diagrams serve as efficient representations for elements of $T$. However, since elements of $T$ represent homeomorphisms of the circle rather than the interval, the tree pair diagram must also include a bijection between the leaves of the source tree and the leaves of the target tree to fully encode the homeomorphism. Since this bijection can at most cyclically shift the leaves, it is determined by the image of the leftmost leaf in the source tree. Since by convention this leaf in the source tree is already thought of as leaf $0$, this information is recorded by writing a $0$ under the image leaf in the target tree. Hence, for $w \in T$, a {\em marked tree pair diagram} representing $w$ is a pair of finite rooted binary trees with the same number of leaves, together with a mark (the numeral 0) on one leaf of the second tree. As usual, we write $w = (T_-,T_+)$ to express $w$ as a tree pair diagram, and refer to $T_-$ as the {\em source} tree and $T_+$ (the one with the mark) as the {\em target} tree. We remark that to extend this to $V$, since now the bijection of the subintervals may permute the order in any way, the marking required on the target tree to record the bijection consists of a number on every leaf of the target tree. Just as for $F$, there are many possible tree pair diagrams for each element of $T$, which can be obtained by adding carets to the corresponding leaves in the source and target trees in the diagrams. However, when adding the carets, placement is guided by the marking. The leaves of the source tree are thought of as numbered from $0$ to $n$ reading from left to right, whereas the marking of the target tree specifies where leaf number $0$ of that tree is, and other leaves are numbered from $1$ to $n$ reading from left to right cyclically wrapping back to the left once you reach the rightmost leaf. With this numbering in mind, carets can be added as before to leaf $i$ of both trees. If $i \neq 0$, the mark stays where it is. Otherwise, if $i=0$, the mark on the new target tree is placed on the left leaf of the added caret. So for $T$, we have a similar reduction condition: if $w = (T_-,T_+)$ and both trees contain a caret with two exposed leaves numbered $i$ and $i+1$, then we remove these carets and renumber the leaves, moving the mark if needed, thus forming a representative for $w$ with fewer carets and leaves. A tree pair diagram which admits no such reductions is again called a {\em reduced tree pair diagram}, and any element of $T$ is represented by a unique reduced tree pair diagram. In $T$ as well as in $F$, when we write $w = (T_-,T_+)$ below, we are assuming that the tree pair diagram is reduced unless otherwise specified. Checking whether or not a tree pair diagram is reduced is slightly more difficult in $T$ than in $F$. The process of checking for possible reductions is illustrated in Figure \ref{fig:reduction}. A marked tree pair diagram for an element of $T$ is shown on the top left of Figure \ref{fig:reduction}. In the top right tree pair diagram of Figure \ref{fig:reduction}, the underlying numbering of the leaves of both trees determined by the marking is written explicitly, revealing the reducible carets. The bottom tree pair diagram shows the resulting reduced diagram. \begin{figure} \includegraphics[width=5in]{reduction}\\ \caption{An example of a caret reduction in a tree pair diagram representing an element of $T$. The diagram on the top left is reducible; The two dotted carets on the top right are paired with each other, since the numbering is identitcal. The resulting reduced diagram is shown on the bottom.\label{fig:reduction}} \end{figure} Note that the torsion generators $c_i$ have particularly simple tree pair diagrams. In the diagram for $c_i$, both source and target trees consist of the root caret plus $i$ right carets. The mark $0$ is placed on the rightmost leaf of the target tree. \begin{figure} \includegraphics[width=5.2in]{cs}\\ \caption{Tree pair diagrams representing the elements $c_1$, $c_2$ and $c_3$ on top, plus $c_n$ on the bottom.\label{fig:cs}} \end{figure} Figure \ref{fig:cs} shows the the tree pair diagrams of the first three generators $c_1, c_2$, and $c_3$, together with a general $c_n$. The generator $c_0$ is merely a pair of single caret trees, with the mark on the rightmost leaf of the target tree. Whether $w \in F$ or $w \in T$, we denote the number of carets in either tree of a tree pair diagram representing $w$ by $N(w)$. When $p$ is a word in the generators of $F$ or $T$, then $p$ represents an element $w$ in either $F$ or $T$, and we write $N(p)$ interchangeably with $N(w)$. \subsection{Group Multiplication in $F$ and $T$}\label{multiplication} Group multiplication in $F$ and $T$ corresponds to composition of homeomorphisms, which we can interpret on the level of tree pair diagrams as well. First, we consider $u,v \in F$, where $u = (T_-,T_+)$ and $v = (S_-,S_+)$. To compute the tree pair diagram corresponding to the product $vu$, we create unreduced representatives $(T'_-,T'_+)$ and $(S'_-,S'_+)$ of the two elements in which $T'_+ = S'_-$. Then the product is represented by the possibly unreduced tree pair diagram $(T'_-,S'_+)$. The multiplication is written following the conventions on composition of homeomorphisms, so the product $vu$ has as a source diagram that of $u$, and as a target diagram that of $v$. That is, the diagram on the left is the source of $u$ and the diagram on the right is the target of $v$. To multiply tree pair diagrams representing elements of $T$ we follow a similar procedure. We let $u,v \in T$, where $u = (T_-,T_+)$ and $v = (S_-,S_+)$. To compute the tree pair diagram corresponding to the product $vu$, we create unreduced representatives $(T'_-,T'_+)$ and $(S'_-,S'_+)$ of the two elements in which $T'_+ = S'_-$ as trees. The product $vu$ will be represented by the pair $(T'_-,S'_+)$ of trees. To decide which leaf in $S'_+$ to mark with the zero, we just note that it should be the leaf which is paired with the zero leaf in $T'_-$. To identify this leaf, we find the zero leaf in $T'_+$. Since $T'_+=S'_-$ as trees, this leaf viewed as a leaf in $S'_-$ will be labelled $m$. Then the leaf labelled $m$ in $S'_+$ will be the new zero leaf in the tree pair diagram $(T'_-,S'_+)$ for $vu$. Alternately, we can follow the composition in both pairs of trees to see how the leaves are paired. This newly constructed tree pair diagram will represent $vu$ and is not necessarily reduced. For an example of this multiplication, see Figures \ref{fig:mult1}, \ref{fig:mult2} and \ref{fig:mult3}. \begin{figure} \includegraphics[width=5.2in]{mult1}\\ \caption{ The tree pair diagram for sample elements $u$ and $v$ in $T$.\label{fig:mult1}} \end{figure} \begin{figure} \includegraphics[width=5.2in]{mult2}\\ \caption{ Unreduced versions of $u$ and $v$ necessary for the multiplication $vu$ in $T$, with carets added to perform the multiplication indicated with dashes. Now the target tree of $u$ has the same shape as the source tree of $v$, allowing the composition.\label{fig:mult2}} \end{figure} \begin{figure} \includegraphics[width=3in]{newmult3}\\ \caption{ The tree pair diagram representing the product $vu$ obtained from Figure \ref{fig:mult2}. The dotted carets must be erased to find the reduced diagram.\label{fig:mult3}} \end{figure} \section{Words and diagrams} \subsection{Normal forms and tree pair diagrams in $F$} With respect to the infinite presentation for $F$ given above, every element of $F$ has a unique normal form. Any $w$ in $F$ can be written in the form $$w=x_{i_1}^{r_1} x_{i_2}^{r_2}\ldots x_{i_k}^{r_k} x_{j_l}^{-s_l} \ldots x_{j_2}^{-s_2} x_{j_1}^{-s_1} $$ where $r_i, s_i >0$, $0 \leq i_1<i_2 \ldots < i_k$ and $0 \leq j_1<j_2 \ldots < j_l$. However, this expression is not unique. Uniqueness is guaranteed by the addition of the following condition: when both $x_i$ and $x_i^{-1}$ occur in the expression, so does at least one of $x_{i+1}$ or $x_{i+1}^{-1}$, as discussed by Brown and Geoghegan \cite{bg:thomp}. When we refer to elements of $F$ in normal form, we mean this unique normal form. If the normal form for $w \in F$ contains no generators with negative exponents, we refer to $w$ as a {\em positive word} and similarly, we say a normal form represents a {\em negative word} if there are no generators with positive exponents. We call any word which has the form $$w=x_{i_1}^{r_1} x_{i_2}^{r_2}\ldots x_{i_k}^{r_k} x_{j_l}^{-s_l} \ldots x_{j_2}^{-s_2} x_{j_1}^{-s_1} $$ where $r_i, s_i >0$, $0 \leq i_1<i_2 \ldots < i_k$ and $0 \leq j_1<j_2 \ldots < j_l$, a word in {\em pq form}, where $p$ is the positive part of the normal form and $q$ the negative part. The normal form for an element of $F$ is the shortest word among all words in $pq$ form representing the given element. To any (not necessarily reduced) tree pair diagram $(T_-,T_+)$ for an element of $F$ we may associate a word in $pq$ form representing the element, using the {\em leaf exponents} in the target and source trees. When the leaves of a finite rooted binary tree are numbered from left to right, beginning with zero, the leaf exponent of leaf $k$ is the integer length of the longest path consisting only of left edges of carets which originates at leaf $k$ and does not reach the right side of the tree. A tree pair diagram then gives the word $$ x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}x_{j_m}^{-s_m}\ldots x_{j_2}^{-s_2}x_{j_1}^{-s_1} $$ precisely when leaf $i_k$ in $T_+$ has exponent $r_k$, leaf $j_k$ in $T_-$ has leaf exponent $s_k$, and generators which do not appear in the word correspond to leaves with exponent zero. We think of this word as the $pq$ factorization of the element given by the particular tree pair diagram. We call a tree an {\em all-right tree} if it consists of a root caret together with only right carets. Note that if we let $R$ be the all-right tree with the same number of carets as $T_-$ or $T_+$, then $(T_-,R)$ is a diagram for the word $q$ and $(R,T_+)$ is a diagram for word $p$. On the other hand, any word in $pq$ form can be translated into a tree pair diagram. It can be obtained by taking diagrams for $p$ (respectively $q$), which will have all right source (respectively target) trees. Then, if one diagram has fewer carets, one adds right carets to its all-right tree, and of a corresponding path of right carets to its other tree, to make both diagrams have exactly the same all right tree. Furthermore, under this correspondence for $F$, reduced tree pair diagrams correspond exactly to normal forms. Figure \ref{fig:leafexp} is an example of this correspondence, and more details can be found in \cite{cfp,ctcomb,blake:diss}. \begin{figure} \includegraphics[width=3in]{leafexp3}\\ \caption{Computing leaf exponents. The thick edges indicate edges which contribute to non-zero leaf exponents. If a leaf labelled $i$ has $r_i$ thick edges (a path of $r_i$ left edges going up without reaching the right side of the tree) then the $i$-th leaf exponent is $r_i$ and the generator appearing in the normal form is $x_i^{r_i}$. This single tree $T$ pictured above is the target tree of the tree pair diagram $(R,T)$, where $R$ is the all-right tree with 12 leaves, and has leaf exponents 1,0,3,0,1,0,0,0,2,0,0, and 0 for the leaves 0-11 in order. The tree pair diagram $(R,T)$ represents the element $x_0x_2^3x_4x_8^2$. \label{fig:leafexp}} \end{figure} If an exposed caret has leaves numbered $i$ and $i+1$, then leaf $i+1$ must have leaf exponent zero, since it is a right leaf. If both trees in a tree pair diagram have exposed carets with leaves numbered $i$ and $i+1$, then the corresponding normal form, computed via leaf exponents, contains the generators $x_i$ to both positive and negative powers, but no instances of the generator $x_{i+1}$. This is precisely the situation when the normal form can be reduced by a relator of $F$. Thus the condition that the normal form is unique is exactly the condition that the tree pair diagram is reduced. This correspondence will be extended to elements of $T$ in the next section. \subsection{Tree pair diagrams for elements of $T$} We now discuss the relationship between words in $T$ and tree pair diagrams. This relationship is more complicated in $T$ than it is in $F$. The representation of elements of $T$ by marked tree pair diagrams suggests a way to decompose an element of $T$ into a product of three elements: the positive and negative parts together with a torsion part in the middle, as described in \cite{cfp}. \begin{defn} \label{factorization} Let the marked tree pair diagram $(T_-,T_+)$ represent $g \in T$. If~~$T_-$ and $T_+$ each have $i+1$ carets, then we let $R$ be the all-right tree which has $i+1$ carets. We can write $g$ as a product $pc_i^jq$, where: \begin{enumerate} \item $p$, a positive word in the generators of $F$, is the normal form for the element of $F$ with tree pair diagram $(R,T_+)$, ignoring the marking on $T_+$. \item $c_i^j$ is a cyclic permutation of the leaves of $R$, with $1 \leq j \le i+2$, and \item $q$, a negative word, is the normal form for the element of $F$ represented by $(T_-,R)$. \end{enumerate} Then the word $g=pc_i^jq$ is called the \emph{pcq factorization} of $g$ associated to the marked tree pair diagram $(T_-,T_+)$. In the special case where $g \in F \subset T$, the $pcq$ factorization will just be the usual $pq$ factorization, as we consider the $c$ part of the word to be empty (or equivalently, we can allow the exponent $j$ in the torsion part to be zero.) \end{defn} Figure \ref{fig:threeparts} illustrates an example of an element of $T$ decomposed in this way. \begin{figure} \includegraphics[width=5.2in]{pcq2}\\ \caption{ Three tree pair diagrams representing the word $x_1 x_2 c_5^5 x_2^{-2} x_1^{-1} x_0^{-2}$ factorized as $pcq$.\label{fig:threeparts}} \end{figure} The following theorem follows from the existence of these decompositions, and an algebraic proof of this result is found in \cite{cfp}. \begin{thm}[\cite{cfp}, Theorem 5.7]\label{pcq}Any element $x\in T $ admits an expression of the form $$ x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}\,c_i^j\,x_{j_m}^{-s_m}\ldots x_{j_2}^{-s_2}x_{j_1}^{-s_1}, $$ where $0 \leq i_1<i_2< \cdots <i_n$ and $0 \leq j_1<j_2<\cdots<j_m$ and either $1 \leq j<i+2$ or $c_i^j$ is not present.\end{thm} We refer to any word satisfying the hypotheses of Theorem \ref{pcq} as a word in $pcq$ form for an element of $T$ (just as words of this form with no $c_i^j$ term are called words in $pq$ form in the group $F$). Neither proof of the existence of $pcq$ forms gives an easy explicit method for transforming a general word in the generators $x_i^{\pm 1},c_i$ into $pcq$ form without resorting to drawing tree pair diagrams, so we will outline an algebraic method below. We recall that the five types of relators we are using in $T$ are: \begin{enumerate} \item $x_jx_i=x_ix_{j+1}$, if $i<j$ \item $x_kc_{n+1}=c_{n}x_{k+1}$, if $k<n$ \item $c_nx_0=c_{n+1}^2$ \item $c_n=x_nc_{n+1}$ \item$c_n^{n+2}=1$ \end{enumerate} \begin{lemma} [Pumping Lemma] The generators $x_i$ and $c_j$ of $T$ satisfy the following identities $$ c_n^m=x_{n-m+1}c_{n+1}^m\qquad\qquad c_n^m=c_{n+1}^{m+1}x_{m-1}^{-1} $$ if $1\le m < n+2$. \end{lemma} \begin{proof} This follows immediately from the relators. For instance, for the first identity, we have that $$ c_n^m=c_n^{m-1}c_n=c_n^{m-1}x_nc_{n+1} $$ by an application of relator of type (4). Now, several repeated applications of relator (2) allow the $x_n$ to switch with the $c_n^{m-1}$ to obtain the desired result. The second identity is the first one taking inverses, and by noticing that $c_n$ has order $n+2$, we avoid negative exponents for the $c$. \end{proof} We consider a word $w \in T$ written in the generators $\{x_i,c_j\}$, and we describe explicitly an algebraic method of rewriting it in $pcq$ form. The idea is to first combine occurrences of multiple $c_i$ generators into a power of a single one, and then to move the $x_n$ generators to the appropriate side of it. Consider first a subword of the original word $w$ of type $$ c_n^m\,w(x_i)\,c_k^l, $$ where $w(x_i)$ is a word on the generators $x_n$ only, and which may possibly be empty. We will apply relators to reduce this subword to a word of the form $w_1(x_i) c_j^h w_2(x_i)$, where $w_1$ has only positive powers of $x_n$ generators and $w_2$ consists of only negative ones. By the relators of type (1), we can assume that $w$ is of the form $pq$, that is, with all positive powers of generators on the left and in increasing order of index, and all negative ones on the right and in decreasing order of index. The goal is to move all the positive powers of $x_n$ generators to the left of $c_n^m$ and all negative ones to the right of $c_k^l$. Although these moves may change the indices and powers of the $c_i$ generators, they merely change a power of a single $c_j$ generator to another power of a different single $c_k$ generator. To move all the positive powers of generators to the left of $c_n^m$, we only need to use relators of the type (2), assuming the index of $c$ is high enough. If it is not, by repeated applications of the first identity of the pumping lemma, the index can be increased arbitrarily, adding only positive powers of generators to the {\it left\/} of $c_n^m$. When the subindex is high enough, we can use relators of type (2) to move all positive powers of generators of $w$ past $c_n^m$. We note that a relator of type (3) may allow us to eliminate a occurance of $x_0$ to the immediate right of $c_n^m$. It may be necessary to combine the $c_i$ and $c_j$ generators obtained into a single term after this elimination of $x_0$, as we see in an example: $$ c_4^3x_1=c_4^2x_0c_5 $$ At this point $x_0$ cannot be moved farther, but we can use relator (3) to obtain $$ c_4c_5^3=x_5c_5^4 $$ with the last equality being an application of the pumping lemma to $c_4$. We have achieved the goal of moving a positive power of a generator to the left of $c_n^m$. Moving the negative powers of the $x_n$ generators to the left is comparable. Using the second identity in the pumping lemma, we can increase the index in $c_k^l$ as much as necessary to be able to move all negative powers of the $x_n$ generators in $w$ to the right of $c_k^l$ using the relators (2) rewritten as $c_{n+1}x_{k+1}^{-1}=x_k^{-1}c_n$. After this process, we will have a word consisting of positive powers of $x_n$ generators, two powers of $c_i$ generators, and negative powers of $x_n$ generators. We now combine the powers of the two $c_i$ generators into a power of a single generator, by increasing the smaller index to reach the larger. To do this, if the smaller is on the left, we can use the first identity in the pumping lemma, and if it is on the right, we can use the second one. This way no $x_n$ generator will be added in between the two $c_i$ generators and after they have the same index they can be combined into a power of a single generator. The positive powers of the $x_n$ generators now appear only to the left of the single power of the $c_i$ generator, and negative powers of $x_n$ generators only to the right. After repeated applications of this process to subwords of the type $cwc$, we will have all occurrences of the $c_i$ generators combined into a power of a single one. Our original word is now of the type $$ w_1(x_i)\,c_n^m\,w_2(x_i), $$ and $w_1$ and $w_2$ may again be assumed, after using relators (1), to be in $pq$ form. We only need to move the positive powers of generators in $w_2$ to the left of $c_n^m$ and the negative powers of $x_n$ generators of $w_1$ to the right of $c_n^m$, still maintaining a power of a single $c_i$ generator in the middle. We describe above as the first step in our algorithm precisely how to do this. Furthermore, if the pumping lemma is needed to move a positive power of a generator to the left, recall that new positive positive powers of generators may appear in the word, but only to the left of the power of the $c_i$ generator. Hence, after moving each positive power of a generator, all positive powers of generators in the word are to the left of $c_n^m$. We now move each negative power of a generator to the right, and notice that the only cost of this is to add more negative powers of $x_n$ generators to the right of $c_n^m$. When this is finished, the word has only positive powers of generators to the left of a power of a single $c$ and negative ones to the right. Once the positive powers of the generators are together on the left side of the single $c$ term, we can reorder them if necessary using relators of type (1), and similarly we can reorder the negative part as well. We will work an example as an illustration. Consider the word $$ x_0^{-1}c_1x_3c_3^2x_1^{-1} $$ The process starts by trying to move the $x_3$ to the left of $c_1$. Since the index of $x_3$ exceeds the index of $c_1$, we cannot apply a relator of type (2) directly. Using the pumping lemma, we write $c_1=x_1c_2=x_1x_2c_3$. Hence our word is now the following, and we can apply the relator of type to $c_3x_3$, obtaining: $$ x_0^{-1}x_1x_2c_3x_3c_3^{2}x_1^{-1}=x_0^{-1}x_1x_2^2c_4c_3^2x_1^{-1}. $$ We need to merge $c_4c_3^2$ into a single $c_i$ term. We increase the index of $c_3$ via $c_3^2=c_4^3x_1^{-1}$ to obtain $$ x_0^{-1}x_1x_2^2c_4c_4^3x_1^{-2}=x_0^{-1}x_1x_2^2c_4^4x_1^{-2}. $$ The last step is to move the initial $x_0^{-1}$ to the right side, using several relators of type (2) to obtain $x_0^{-1}c_4^4=c_5^4x_4^{-1}$. There is no need this time to increase the index of $c_4^4$. The final result is $$ x_2x_3^2c_5^4x_4^{-1}x_1^{-2} $$ which is in $pcq$ form. The relationship between words in $pq$ form and tree pair diagrams in $F$ is different than the relationship between $pcq$ forms and tree pair diagrams in $T$. In $F$, every tree pair diagram has a $pq$ factorization associated to it, and any word in $pq$ form is in fact the $pq$ factorization associated to a (not necessarily unique) tree pair diagram. Given any word in $F$ in $pq$ form, then we can form a tree pair diagram for this element as follows. We consider reduced tree pair diagrams for $p$ and $q$, and construct a tree pair diagram for the product $pq$ as described in Section \ref{multiplication}. The middle trees of the four trees involved in the product are all-right trees. The all-right trees in this decomposition may not have the same number of carets, so in forming the diagram for $pq$ we simply enlarge the smaller of the two of these all-right trees (as well as the other tree in that diagram). Since only right carets are ever added during this process, all of whose leaves have leaf exponent zero, this results in a tree pair diagram whose $pq$ factorization is precisely the word $pq$ we began with. In $T$, the correspondence between $pcq$ factorizations and general $pcq$ words is not as straightforward as in $F$. There is a difference between $pcq$ factorization and $pcq$ algebraic form. Though every element has a tree pair diagram corresponding to a $pcq$ factorization associated to it, there are words in algebraic $pcq$ form which are not the $pcq$ factorizations associated to a tree pair diagram. The difficulty arises when the tree pair diagram for $c$ does not have as many carets as those for $p$ or $q$, as adding right carets to enlarge $c$ appropriately necessitates adding generators to the normal forms for $p$ and $q$, so the tree pair diagram one obtains by multiplying as in $F$ will not necessarily have the original word as its factorization. For example, the word $x_1c_1$ is in algebraic $pcq$ form, yet it is not the $pcq$ factorization associated to any tree pair diagram. There is a different representative for this element of $T$ which is the $pcq$ factorization associated to the reduced tree pair diagram for this group element: $x_1c_2x_1^{-1}$. We prefer to work with words which are $pcq$ factorizations associated to tree pair diagrams, which will lead us to unique normal forms. We can algebraically characterize the words of type $pcq$ which are $pcq$ factorizations associated to tree pair diagrams. The important condition is that the reduced tree pair diagram for $c$ should have at least as many carets as those for $p$ and $q$. We say that words in $T$ with this property satisfy the {\em factorization condition}. \begin{thm} \label{thm:factor} For elements in $T$ which are not in $F$, the word $$ x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}\,c_i^j\,x_{j_m}^{-s_m}\ldots x_{j_2}^{-s_2}x_{j_1}^{-s_1} ,$$ where $0 \leq i_1<i_2< \cdots < i_n$, $0\leq j_1<j_2<\cdots<j_m$, and $1 \leq j < i+2$, is the $pcq$ factorization associated to a tree pair diagram if and only if the number of carets in the reduced tree pair diagram for $c_i^j$ is greater than or equal to the number of carets in the reduced tree pair diagram for both of those for the words $ x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}$ or $x_{j_m}^{-s_m}\ldots x_{j_2}^{-s_2}x_{j_1}^{-s_1} $ in $F$. \end{thm} \begin{proof} Given a tree pair diagram, by construction, the $pcq$ factorization associated to it satisfies the factorization condition. Given a word that satisfies the factorization condition, we can easily construct the corresponding tree pair diagram as described above. The factorization condition ensures that to perform the mulitiplication, $p \cdot c \cdot q$ as tree pair diagrams, it is only necessary to (possibly) add carets to the tree pair diagrams for the words $p$ and $q$. This will not alter the normal form, and thus the diagram constructed will indeed have the original word as its $pcq$ factorization. \end{proof} We can compute the number of carets of a reduced tree pair diagram for a word $w \in F$ algebraically from the normal form of $w$, as described by Burillo, Cleary and Stein in \cite{bcs}. \begin{prop} [Proposition 2 of \cite{bcs}] \label{prop:ncarets} Given a positive word in $w \in F$ in normal form $$ w=x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}, $$ then the number of carets $N(w)$ in either tree of a reduced tree diagram representing $w$ is $$ N(w)=\max\{i_k+r_k+\ldots+r_n+1\} ,\text{ for }k=1,2,\ldots,n. $$ \end{prop} We can always decide algebraically whether $w \in T$, written in $pcq$ form, corresponds to a tree pair diagram. We use Proposition \ref{prop:ncarets} to count the carets for the positive and negative parts of the word. The number of carets in a tree pair diagram for $c_i^j$ is equal to $i+1$. \section{Normal forms in $T$} In $T$, we will declare the words in $pcq$ form which are $pcq$ factorizations associated to reduced diagrams to be the normal forms for elements of $T$, similar to the approach used in $F$. However, it is no longer true that these words cannot be shortened by applying a relator. As we saw with the normal form $x_1c_2x_1^{-1}$ in $T$, a word may be the shortest word representing an element which satisfies the factorization condition, yet there may be shorter words we can obtain by applying a relator which do not satisfy the factorization condition. Thus, when algebraically characterizing the normal form for elements of $T$, we restrict ourselves to words of $pcq$ form which satisfy the factorization condition, regardless of whether or not a relator may reduce the length of the word. We next specify algebraic conditions which characterize the $pcq$ forms that correspond to normal forms, since we have given geometric conditions in Theorem \ref{thm:factor} \begin{thm}\label{reductions} Let $w$ be a $pcq$ factorization for an element $g \in T$ associated to a marked tree pair diagram in which each tree has $i+1$ carets, where the $c$ part of the word is $c_i^{j}$ with $1 \leq j < i+2$. A reduction of a pair of carets from the tree pair diagram occurs only if the word $w$ satisfies one of the following conditions: \begin{itemize} \item[(1)] The pair of generators $x_{k-j}$ and $x_k^{-1}$ appear, with $j \leq k< i$, and neither of the two generators $x_{k-j+1}$ and $x_{k+1}^{-1}$ appear. The reduction corresponds to applying the relator $$ x_{k-j}c_{i}^jx_k^{-1}=c_{i-1}^j $$ after applying relators from $F$ in the $p$ and $q$ parts of the word, if necessary, to make $x_{k+j}$ and $x_k^{-1}$ adjacent to $c_i^j$. \item[(2)] The generator $x_{i-j}$ appears, and $x_{i-j+1}$ does not. The reduction corresponds to applying $$ x_{i-j} c_{i}^j=c_{i-1}^j $$ after possibly using relators from $F$ as in (1). \item[(3)] The pair of generators $x_{k+i-j+2}$ and $x_k^{-1}$ for $0\leq k< j-2$ appear and neither one of the generators $x_{k+i-j+1}$ or $x_{k+1}^{-1}$ appear. The reduction corresponds to applying $$ x_{k+i-j+2}c_{i}^jx_k^{-1}=c_{i-1}^{j-1} $$ after possibly applying relators from $F$. \item[(4)] The generator $x_{j-2}^{-1}$ appears, and the generator $x_{j-1}^{-1}$ does not appear. The reduction corresponds to $$ c_{i}^{j} x_{j-2}^{-1}=c_{i-1}^{j-1} $$ after possibly applying relators from $F$. \end{itemize} \end{thm} \begin{figure}[b] \includegraphics[width=5.82in]{normal}\\ \caption{The four cases in Theorem \ref{reductions}, showing the two labellings on the leaves of the trees, the cyclic labelling which indicates the correspondence of the leaves, and the leaf exponent labelling which indicates the corresponding generators in the normal form. \label{fig:normal}} \end{figure} \begin{proof} Let $g \in T$ be represented by a marked tree pair diagram $(T_-,T_+)$. If both trees have an exposed caret whose leaves are identically numbered, then we call that a {\em reducible caret pair}, as it must be removed in order to obtain the reduced tree pair diagram representing $g$. We now consider algebraic conditions corresponding to a reducible caret in a tree pair diagram. In the tree pair diagram $(T_-,T_+)$ for $g \in T$, there are two ways of labelling the leaves in the target tree $T_+$. The first labelling corresponds to the order in which the intervals in the subdivisions determined by these trees are paired in the homeomorphism, and is called the cyclic labelling. The cyclic labelling gives the marked leaf in the target tree the number zero, and the other leaves are given increasing labels from left to right around the leaves of the tree. The second labelling ignores the marking and puts the leaves in increasing order from left to right, beginning with zero. The first labelling is used to determine which leaves in $T_-$ are paired with which leaves in $T_+$, and the second labelling is used in the computation of leaf exponents to determine the powers of the generators that appear in the word. Figure \ref{fig:normal} shows the labellings for the four cases of the theorem. Suppose that the tree pair diagram for $g \in T$ is not reduced. The four cases above correspond to the following four possible locations of a reducible caret relative to the marked leaf in the target tree. \begin{itemize} \item Case (1) of the thereom corresponds to the case when the left leaf of the reducible caret is to the left of the marked leaf in $T_+$, but the reducible caret is not the rightmost caret in $T_-$. \item Case (2) corresponds to the special case when the reducible caret is a right caret in $T_-$, in which case necessarily its left leaf is to the left of the marked leaf in $T_+$. Leaf exponents from leaves of right carets will always be zero and thus right carets cannot contribute generators to the normal form. They may still result in an exposed reducible caret, which occurs exactly in this case, and the reduction will only affect the $q$ part of the normal form. \item Case (3) corresponds to the case when the left leaf of the reducible caret is either to the right of or coincides with the marked leaf in $T_+$, but the reducible caret is not the rightmost caret in $T_+$. \item Case (4) corresponds to the special case when the reducible caret is a right caret in $T_+$, in which case it cannot be to the left of the marked caret in $T_+$. As in Case (2), the exposed caret in this case is a right caret and does not contribute a generator to the normal form, but may still be reduced. This cancellation affects only the $p$ part of the normal form. \end{itemize} To see that these are all the possibilities, we note that $k$, the number of the left leaf in the cyclic numbering of the reducible caret in $T_-$, achieves all possible values in the cases above: \begin{itemize} \item If $0\le k<j-2$ we are in case (3). \item If $k=j-2$ we are in case (4). \item The case $k=j-1$ is impossible because the leaves $j-1$ and $j$ are at the two ends of the tree. With a cyclic ordering the last and first leaves do not form a caret. \item If $j\le k<i$ we are in case (1). \item If $k=i$ we are in case (2). \end{itemize} Figure \ref{fig:normal} illustrates that these are all the possibilities.\end{proof} The conditions in Theorem \ref{reductions} together with the factorization condition algebraically characterize our normal forms. The normal forms for elements in $F$ have already been characterized, so we restrict to elements not in $F$ in our description. \begin{thm} Any element $g\in T$ which is not an element of $F$ admits an expression of the form $pcq$ where $$ p=x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}\qquad c=c_i^j\qquad q=x_{j_m}^{-s_m}\ldots x_{j_2}^{-s_2}x_{j_1}^{-s_1}, $$ $0 \leq i_1<i_2< \cdots < i_n$, $0 \leq j_1<j_2<\cdots<j_m$, and $1 \leq j < i+2$. Among all the words in this form representing an element, there is a unique one satisfying the following conditions, which we call the normal form. \begin{itemize} \item The word satisfies the factorization condition, which we now state as $i+1\ge\max\{N(p),N(q)\}$. \item The word does not admit any reductions, and thus its normal form satisfies the following conditions: \begin{itemize} \item If there exists a pair of generators $x_{k-j}$ and $x_k^{-1}$ simultaneously, for $j\le k< i$, then one of the generators $x_{k-j+1}$ or $x_{k+1}^{-1}$ must appear as well. \item If there is a generator $x_{i-j}$, then $x_{i-j+1}$ must exist too. \item If there exists a pair of generators $x_{k+i-j+2}$ and $x_k^{-1}$ for $0\le k< j-2$, then one of the generators $x_{k+i-j+1}$ or $x_{k+1}^{-1}$ must appear as well. \item If there exists a generator $x_{j-2}^{-1}$, then a generator $x_{j-1}^{-1}$ must also appear. \end{itemize} \end{itemize} \end{thm} \begin{proof} We claim that the conditions above precisely describe the set of unique normal forms for $T$. A $pcq$ word satisfying the factorization condition is the $pcq$ factorization associated to a marked tree pair diagram. However, if the $pcq$ word satisfies all four reduction conditions, we have just shown in the previous theorem that this diagram is in fact the unique reduced diagram, and hence the word is in fact a normal form. \end{proof} We remark that the Pumping Lemma together with the reductions in Theorem \ref{reductions} give an explicit way of algebraically transforming any word in the generators of $T$ into a normal form. Given any word, we rewrite it in $pcq$ form using the process described following the Pumping Lemma. If the resulting word does not satisfy the factorization condition, then we iterate the Pumping Lemma until we obtain a word for which the factorization condition is satisfied. The Pumping Lemma increases the number of carets for $c$ and the number of carets for one of the words $p$ and $q$. Once a word is obtained which satisfies the factorization condition, there must be a corresponding tree pair diagram for the element. Now, if the word satisfies any of the reduction conditions in Theorem \ref{reductions}, we apply them successively using the relators described there. This method thus produces the unique normal form. \section{The word metric in $T$} \subsection{Estimating the word metric} For metric questions concerning $T$, we must consider a finite generating set instead of the one used to obtain the normal form for elements. We now approximate the word length of an element of $T$ with respect to the generating set $\{x_0,x_1,c_1\}$, using information contained in the normal form and the reduced tree pair diagram. These estimates are similar to those for the estimates of word metric in $F$ with respect to the generating set $\{x_0,x_1\}$ described \cite{burillo}, \cite{bcs}. \begin{thm} \label{thm:D} Let $w \in T$ have normal form $$ w=x_{i_1}^{r_1}x_{i_2}^{r_2}\ldots x_{i_n}^{r_n}\,c_i^j\,x_{j_m}^{-s_m}\ldots x_{j_2}^{-s_2}x_{j_1}^{-s_1}. $$ We define $$ D(w)=\sum_{k=1}^nr_k+\sum_{l=1}^ms_l+i_n+j_m+i. $$ Let $|w|$ denote the word metric in $T$ with respect to the generating set $\{x_0,x_1,c_1\}$. There exists a constant $C>0$ so that for every $w \in T$, $$ \frac{D(w)}C\le|w|\le C\,D(w) $$ and similarly, for $N(w)$ the number of carets in the reduced tree pair diagram representing $w$, $$ \frac{N(w)}C\le|w|\le C\,N(w). $$ \end{thm} \begin{proof} These inequalities follow from the correspondence between the normal form and the tree pair diagram for an element $w \in T$. It is clear, from Proposition \ref{prop:ncarets}, that $ N(w)\ge \sum_{k=1}^nr_k$, $N(w)\ge \sum_{l=1}^ms_l$, $N(x)\ge i_n$, and $N(w) \ge j_m$. The inequality $ N(w)\ge i $ is clear from the fact that $c_i$ has $i+1$ carets. These inequalities prove that $$ D(w)\le 5\,N(w). $$ We rewrite the generators $x_i$ and $c_j$ in terms of $x_0$, $x_1$ and $c_1$ and look at the lengths of the resulting words to obtain the inequality $$ |w|\le C\,D(w) $$ for some constant $C>0$. Combining the two inequalities above, we have $$ |w|\le C'\,N(w). $$ To obtain lower bound on the word length, we consider the fact that the tree pair diagram for each generator has either two or three carets. If $u$ is a word in $x_0$, $x_1$ and $c$ with length $n$, then as these generators are multiplied together, each product may add at most $3$ carets to the tree pair diagram. Thus the diagram for $u$ will have at most $3n$ carets. It then follows that $$ N(w)\le 3|w|. $$ Combining this with the above inequality, we obtain the desired bounds. \end{proof} We use Theorem \ref{thm:D} to show that the inclusion of $F$ in $T$ is a quasi-isometric embedding. This means that there are constants $K>0$ and $C$ so that for any $w,z \in F$ we have $$\frac{1}{K} d_{F}(w,z) - C \leq d_T(w,z) \leq Kd_F(w,z) + C$$ where $d_F$ and $d_T$ represent the word metric in $F$ and $T$ respectively, with regard to the generating set $\{x_0,x_1\}$ of $F$ and $\{x_0,x_1,c_1\}$ of $T$. When considering whether the inclusion of a finitely generated subgroup $H$ into a finitely generated group $G$ is a quasi-isometric embedding, we can instead equivalently show that the distortion function is bounded. The distortion function is defined by $$h(r) = \frac{1}{r} \max \{|x|_H : x \in H \mbox{~and~} |x|_G \leq r\}.$$ Word length in $F$ is comparable to the number of carets in the reduced tree pair diagram representing the word, by Theorem 3 of \cite{bcs} or more directly by Fordham's method \cite{blakegd}. This, combined with Theorem \ref{thm:D} easily shows that the distortion function is bounded, and thus proves the following corollary with respect to the generating sets $\{x_0, x_1\}$ and $\{x_0,x_1,c_1\}$ and thus all pairs of finite generating sets: \begin{cor} \label{cor:qiemb} The inclusion of $F$ in $T$ is a quasi-isometric embedding. \end{cor} \subsection{Comparing word length in $F$ and $T$\label {isomembed}} Although Corollary \ref{cor:qiemb} shows that $F$ is quasi-isometrically embedded in $T$, in fact the word length of many elements of $F$ does not change at all when these elements are considered as elements of $T$, with respect to natural finite generating sets. As an example of this phenomenon, we characterize one type of element of $F$ whose word length is unchanged when viewed as an element of $T$, using the generating set $\{x_0,x_1\}$ for $F$ and $\{x_0,x_1,c_0\}$ for $T$. These are elements $w \in F$ for which $N(w)$ exceeds the word length $|w|_F$. Fordham \cite{blakegd} computes $|w|_F$ by assigning an integer weight between zero and four to each pair of carets in the tree pair diagram representing $w$. In a given word there are at most two weights of zero. Here we investigate words in which most weights are one. Such words, for example, are represented by tree pair diagrams with no interior carets having right children. \begin{thm} \label{thm:wordlength} If $w \in F$ with $N(w) \geq |w|_F + 1$ then $|w|_T = |w|_F$, where word length if computed with respect to the generating set $\{x_0,x_1\}$ for $F$ and $\{x_0,x_1,c_0\}$ for $T$. \end{thm} This theorem is proved by taking a word in the generators of $T$, and analyzing how each generator changes the intermediate tree pair diagram as one builds up the final tree pair diagram for $w$. Carefully controlling the process allows one to obtain an upper bound on $N(w)$ in terms of the length of the word. If the word is actually shorter than $|w|_F$, then this bound, considered together with the lower bound given by the hypothesis, yields a contradiction. We immediately obtain the following corollary, since $|x_0^n|_F = |x_1^n|_F = n$, while $N(x_0^n) = n+1$ and $N(x_1^n) = n+3$. \begin{corollary} The elements $x_0^n$ and $x_1^n$ have word length $n$ in both $F$ and $T$ with respect to the finite generating sets $\{x_0,x_1\}$ and $\{x_0,x_1,c_0\}$ respectively. \end{corollary} \section{Torsion elements} Although the group $F$ is torsion free, both $T$ and $V$ contain torsion elements. It is easy to construct torsion elements in $T$ or $V$ by choosing any binary tree $S$ and making any marked tree pair diagram with $S$ as both source and target tree. If the labelling of the target tree is the same as the labelling of the source tree, we get an unreduced representative of the identity; otherwise, we get a non-trivial torsion element. If this is an element of $T$, the tree pair diagram has $pcq$ factorization in which $q=p^{-1}$. In fact, any torsion element can be represented by such a tree pair diagram, though its reduced marked tree pair diagram may well not have the same source and target trees, corresponding to the fact that although it has a $pcq$ word where $q=p^{-1}$, the normal form may well not have this special balanced appearance. \begin{prop}\label{p:torsion} If $f\in F,T$ or $V$ is a torsion element, then it can be represented by a (marked) tree pair diagram with the same source and target trees. \end{prop} Before proving Proposition \ref{p:torsion}, we establish some notation which links the analytic and algebraic representations of these groups. For $f \in F$, $T$, or $V$, if $(T_-,T_+)$ is a marked tree pair diagram representing $f$, then it is sometimes convenient to denote the tree $T_+$ by $f(T_-)$. The tree $T_-$ corresponds to a certain subdivision of the circle, which maps under $f$ linearly to another subdivision of the circle. This subdivision is represented by the tree $T_+$, and the marking describes where each subinterval of the circle is mapped. The element $f$ can be thought of as mapping the leaves of $T_-$ to the leaves of $f(T_-)=T_+$, where the marking defines this mapping of the leaves. If $f$ does not have a tree pair diagram in which the tree $T$ appears as the source tree, then the symbol $f(T)$ has no meaning. Given two rooted binary trees $T$ and $T'$, we say that $T'$ is an \emph{expansion} of $T$ if $T'$ can be obtained from $T$ by attaching the roots of additional trees to some subset of the leaves of $T$. We observe that if $(T, f(T))$ is a marked tree pair diagram for $f$, and $T'$ is an expansion of $T$, then there is always a tree pair diagram $(T', f(T'))$ for $f$, and $f(T')$ is an expansion of $f(T)$. Given two rooted binary trees $S$ and $T$, by the \emph{minimal common expansion} of $S$ and $T$ we mean the smallest rooted binary tee which is an expansion of both $S$ and $T$. Using this language, if $(T,f(T))$ and $(S,g(S))$ are marked tree pair diagrams for $f$ and $g$ respectively, the process described in Section 2.3 for creating a tree pair diagram for the product $gf$ could be summarized as follows. If $E$ is the minimal common expansion of $f(T)$ and $S$, then there are tree pair diagrams $(f^{-1}(E),E)$ for $f$, $(E,g(E))$ for $g$, and $(f^{-1}(E),g(E))$ for $gf$ (with appropriate markings). \begin{proof} Suppose that $f$ is a torsion element. We begin by describing the construction of (marked) tree pair diagrams $(A_n,B_n)$ for $f^n$ for every $n \geq 1$. These tree pair diagrams are constructed inductively, viewing $f^n$ as a product $(f^{n-1})(f)$. For $n=1$, let $(A_1,B_1)$ be the reduced marked tree pair diagram for $f$. Throughout this procedure, although markings are carefully carried through in either $T$ or $V$, since our goal is merely to produce a tree pair diagram for $f$ with the same source and target trees (regardless of marking), only the trees themselves are relevant for this argument. Hence we suppress mention of any markings throughout the construction. If $k \geq 2$, suppose the marked tree pair diagram $(A_{k-1},B_{k-1})$ for $f^{k-1}$ has been constructed. Let $E_{k-1}$ be the minimal common expansion of the trees $A_1$ and $B_{k-1}$. Then $f^k$ has tree pair diagram $(f^{-(k-1)}(E_{k-1}),f(E_{k-1}))$, and we let $B_k=f(E_{k-1})$ and $A_k=f^{-(k-1)}(E_{k-1})$. By construction, $A_{k+1}$ is an expansion of $A_k$ for all $k \geq 1$. We claim also that $B_{k+1}$ is an expansion of $B_k$ for all $k \geq 1$. For $k=1$, $E_1$ is by definition an expansion of $A_1$, which implies that $B_2=f(E_1)$ is an expansion of $B_1=f(A_1)$. Suppose inductively that $B_k$ is an expansion of $B_{k-1}$. Now $E_k$ is an expansion of $B_k$ and $A_1$, so $E_k$ is an expansion of $B_{k-1}$ and $A_1$. But $E_{k-1}$ is the minimal common expansion of $B_{k-1}$ and $A_1$, so $E_k$ is an expansion of $E_{k-1}$, which implies that $B_{k+1}=f(E_k)$ is an expansion of $B_k=f(E_{k-1})$. Since there exists a positive integer $m$ such that $f^m$ is the identity, it follows that all tree pair diagrams for $f^m$ must have the same source and target trees. Hence $A_m=B_m$, and then since $A_m$ is an expansion of $A_1$, $B_m$ is an expansion of $A_1$. But since $E_{m-1}$ is the minimal common expansion of $B_{m-1}$ and $A_1$, the fact that $B_m$ is an expansion of both $B_{m-1}$ and $A_1$ implies that $B_m=f(E_{m-1})$ is an expansion of $E_{m-1}$. But they have the same number of carets, so in fact $f(E_{m-1})=E_{m-1}$. In other words, the tree pair diagram $(E_{m-1}, B_n=f(E_{m-1}))$ is the desired tree pair diagram for $f$. \end{proof} \begin{cor} An element of $T$ is torsion if and only if it is a conjugate of some $c_i^j$. \end{cor} \begin{proof} If an element is torsion, then it admits a diagram with two equal trees. The $pcq$ factorization associated with this diagram has the form $pc_i^jp^{-1}$, where $p$ is a positive element of $F$. \end{proof} A particularly natural torsion subgroup is the subgroup $R$ of pure rotations, where by a pure rotation we mean a rotation by $d=\frac{a}{2^n}$ (where $a$ is not divisible by 2). Such pure rotations were used in Section \ref{sec:rotationnum} to conjugate the fixed point of a homeomorphism to 0. This subgroup is isomorphic to the group of dyadic rational numbers modulo 1, which has a 2-adic metric as follows: if $x=\frac{p}{2^l}$, $y= \frac{q}{2^m}$, and $z= |x-y| = \frac{r}{2^k}$, where $p, q$ and $r$ are odd, then $d(x,y)=2^k$. With respect to this metric, the subgroup of rotations is quasi-isometrically embedded in $T$. \begin{prop} The subgroup $R$ of the pure rotations, with the 2-adic metric, is quasi-isometrically embedded in $T$. \end{prop} \begin{proof} We note that if $g\in T$ is the rotation by $\frac{a}{2^n}$ where $a$ is not divisible by $2$, then there are $2^{n}-1$ carets in the reduced tree pair diagram representing $g$, so $N(g)=2^{n}-1$. Since we have shown that the word length of $g$ in $T$ is bi-Lipschitz equivalent to $N(g)$, the proposition follows. \end{proof}
{ "timestamp": "2009-09-03T16:23:44", "yymm": "0503", "arxiv_id": "math/0503670", "language": "en", "url": "https://arxiv.org/abs/math/0503670" }
\section{Introduction} Growth phenomena are ubiquitous around us. They have both very practical applications and theoretical relevance. But they are rarely easy to study analytically and very few rigorous or exact results are known. In two dimensions, the description of a growing domain is often obtained indirectly through the description of a family of univalent holomorphic representations, leading quite generally to equations known under the name of Loewner chains. These techniques, based on the Riemann mapping theorem, are conceptually important but usually far from making the problem tractable. In the last few years, Loewner chains have been discovered which have a large hidden symmetry -- conformal invariance -- that makes them more amenable to an exact treatment \cite{schramm0}. These are known under the name of Stochastic or (Schramm) Loewner evolutions SLE. Their mathematical elegance and simplicity is not their sole virtue. They are also natural candidates to describe the continuum limit of an interface in two-dimensional statistical mechanics models at criticality. At the critical temperature and in the continuum limit, the system is believed to be conformally invariant and physicists have developed many powerful techniques, known under the name conformal field theory or CFT, to deal with local questions in a conformally invariant 2d system. However, nonlocal objects like interfaces posed new nontrivial problems that finally SLE could attack in a systematic way \cite{LSW,LSW:ConformalRestriction}. The connection between CFT and SLE is now well understood \cite{Bb:2002qn,Bb:2002tf,WernerFrie,Bb:2003kd,Bb:2003vu,Bb:2004ij} and the interplay between the two approaches has proved fruitful. The way SLE describes an interface deserves some comments. As a guiding example, consider the Ising model in a simply connected domain, say on the hexagonal lattice. Suppose that the boundary is split in two arcs with endpoints say $a$ and $b$ and impose that on one arc the spins are up and on the other one the spins are down. In this situation each sample exhibits an interface. It joins the two points where the boundary conditions change and splits the domain in two pieces, one with all spins up on its boundary and one with all spins down. This interface fluctuates from sample to sample. What SLE teaches us is the following. Instead of describing the interface between $a$ and $b$ at once, SLE views it as a curve starting from say $a$ and growing toward $b$. And SLE describes the distribution for the addition of an infinitesimal piece of interface when the beginning of the interface is already known. So the description is in terms of a growth process even if there was no growth process to start with. As mentioned above, the probabilistic aspects of SLE as well as its connections with conformal field theory are now fairly well understood. However, some fundamental questions remain, again directly related to natural questions in the statistical mechanics framework. The one we shall concentrate on in this note is what happens when, due to boundary conditions, the system contains several interfaces. Proposals for multiple SLEs have already been made in the literature \cite{Cardy:nSLE,D:SLEcommutation}, but our results point to a different picture. The simplest situation is in fact when there is only one interface but we want to deal with its two ends symmetrically so that two growth processes will interact with each other. Remember that standard SLE deals with the two ends of the interface asymmetrically. This has a price : time reversed SLE is an intricate object. As a guiding example for more than one interface, consider again the Ising model in a simply connected domain on the hexagonal lattice. If one changes boundary conditions from up to down to up and so on $n=2m$ times along the boundary, each sample will exhibit $m$ interfaces, starting and ending on the boundary at points where the boundary conditions change, forming a so-called arch system. However, the interfaces will fluctuate from sample to sample and so does even the topology of the arch system. This topology, for instance, is an observable that is trivial for a single SLE. Our description will again be in terms of growth processes and Loewner chains. For standard SLE, the driving parameter is a continuous martingale and the tip of the curve separates two different states of the system (up and down spins for Ising), leading to a well defined boundary changing operator in statistical mechanics. The relation between the stochastic Loewner equation and the boundary changing operator comes via a diffusion equation that they share in common. For multiple SLEs, we expect that for short time scales each curve grows under the influence of an independent martingale. At its tip stands the same boundary changing operator. But we also expect drift terms, describing interactions between the curves. The possibility of different arch topologies makes it even more natural to have a description with one curve growing at each boundary changing point so that each of them is on the same footing. So $m$ interfaces are described by $n=2m$ growth processes of ``half-interfaces'' that finally pair in a consistent way to build arches. In statistical mechanics, each arch system has a well defined probability to show up. The law governing this finite probability space is again described by a Boltzmann weight which is nothing but a partial partition function. Our starting point is the reconsideration of the role of Boltzmann weights and partition functions in statistical mechanics and their simple but crucial relationship with probabilistic martingales. This allows us to ask the question ``by what kind of stochastic differential equations can one describe multiple SLEs ?'' by imposing a martingale property and conformal invariance. This puts strong constraints on the drift terms and our main result is a description of the family of drift terms that are compatible with the basic rules of statistical mechanics. Each drift is expressed in terms of the partition function of the system. This partition function is given by a sum of Boltzmann weights for configurations that satisfy certain boundary conditions : at the starting points of the curves the boundary conditions change. The partition function depends on the position of these changes, so up to normalization, the partition function is in fact a correlation function. It satisfies a number of partial differential equations (one equation for each point) that are related to the diffusion equations for the multi SLE process. The solutions form a finite dimensional vector space. The positivity constraint satisfied by physical partition functions singles out a cone which is expected, again guided by statistical mechanics, to have the same dimension of the underlying vector space and to be the convex hull of a family of half lines, so that a generic hyperplane section of the cone is a simplex. So geometrically, the drift terms are parametrized by a cone. Extremal drifts, i.e. drifts corresponding to extremal lines in this cone, lead to processes for which the final pattern formed by the growing curves is a given arch system. Drifts inside the simplex give rise to stochastic processes where the asymptotic arch system fluctuates from sample to sample. A crucial role to construct martingales describing interesting events is played by the short distance expansion in conformal field theory because this is what tells which terms in the partition function become dominant when an arch closes, i.e. when two driving processes of the multi SLE hit each other. The vector space of solution of the differential equations for the partition functions has a famous basis indexed by Dyck paths, which are in one to one correspondence with arch systems. But the basis elements do not in general correspond to extremal partition functions. We shall give a rationale for computing the matrix elements for the change of basis and compute a number of them, but we have no closed general formula. We shall illustrate our proposal with concrete computations for $1$ to $3$ interfaces with applications to percolation and the Ising model. We shall also discuss the classical (deterministic) limit $\kappa \rightarrow 0^+$, where only extremal drifts survive. The notes also cover the case when a number of boundary changes are very close to each other but the system is conditioned so that they do not pair with each other. The details are in the main text. Our description is rather flexible in the sense that the speed of growth of each piece of interface can be tuned. Certain limiting cases lead to previously known processes which are examples of SLE$(\kappa,\underline{\rho})$ processes. It is appropriate here to stress that many of the probabilistic properties of the solutions of the stochastic differential equations that we introduce are conjectural at this point. We have made some consistency checks\footnote{For instance, Dub\'edat has derived general ``commutation criteria'' \cite{D:SLEcommutation} for multiple SLEs. The processes we study are a special class satisfying commutation. This class extends vastly the special solution found by Dub\'edat, which in our language corresponds to self avoiding SLEs moving to infinity.} and the whole pattern is elegant, but our confidence comes more from our familiarity with conformal field theory and statistical mechanics. \section{Basics of Schramm-L\"owner evolutions: Chordal SLE} Let us briefly recall what is meant by the chordal SLE --- detailed studies can be found in \cite{RS:BasicProperties} or \cite{W:Lectures}. The chordal SLE process in the upper half plane $\bH$ is defined by the ordinary differential equation \begin{equation} \label{eq: chordal SLE} \frac{\ud}{\ud t} g_t(z) = \frac{2}{g_t(z)-\xi_t} \end{equation} where the initial condition is $g_0(z) = z \in \bH$ and $\xi_t = \sqrt{\kappa} B_t$ is a Brownian motion with variance parameter $\kappa \geq 0$. Let $\tau_z \leq \infty$ denote the explosion time of (\ref{eq: chordal SLE}) with initial condition $z$ and define the hull at time $t$ by $K_t := \overline{ \{ z \in \bH | \tau_z < t \} }$. Then $(K_t)_{t \geq 0}$ is a family of growing hulls, $K_s \subset K_t$ for $s<t$. The complement $\bH \setminus K_t$ is simply connected and $g_t$ is the unique conformal mapping $\bH \setminus K_t \rightarrow \bH$ with $g_t(z) = z + \order (1)$ at $z \rightarrow \infty$. One defines the SLE trace by $\gamma_t = \lim_{\epsilon \downarrow 0} g_t^{-1}(\xi_t + i \epsilon)$. The trace is a continuous path in $\overline{\bH}$ and it generates the hulls in the sense that $\bH \setminus K_t$ is the unbounded component of $\bH \setminus \gamma_{[0,t]}$. For $\kappa \leq 4$ the trace is a non-self-intersecting path and it doesn't hit $\partial \bH = \bR$ for $t>0$ so $K_t = \gamma_{[0,t]}$. For $4 < \kappa < 8$ a typical point $z \in \bH$ is swallowed, i.e. $z \in K_t$ for large $t$ but $z \notin \gamma_{[0, \infty)}$. In the parameter range $\kappa \geq 8$ the trace is space filling, $\gamma_{[0, \infty)} = \overline{\bH}$. Let us point out that no statistical mechanics models seem to correspond to $\kappa > 8$. In the definition of chordal SLE we took the usual parametrization of time. From equation (\ref{eq: chordal SLE}) we see that $g_t (z) = z + 2 t z^{-1} + \Order(z^{-2})$, which means (this could be taken as a definition) that the capacity of $K_t$ from infinity is $2 t$. Since the capacity goes to infinity as $t \rightarrow \infty$, the hulls $K_t$ are not contained in any bounded subset of $\overline{\bH}$. If the parametrization of time is left arbitrary, Schramm's argument yields : \[ \ud g_s(z) = \frac{2\ud q_s}{g_s(z)-M_s},\] where $M_s$ is a continuous martingale with quadratic variation $\kappa q_s$ (an increasing function going to infinity with $s$). In this formula, both $q_s$ and $M_s$ are random objects. The capacity of $K_s$ is $2q_s$. But this is not really more general than eq.(\ref{eq: chordal SLE}) which is recovered by a random time change. \section{A proposal for multiple SLEs} The motivations for our proposal require a good amount of background, but the proposal and its main features themselves can be easily stated. We gather them in this section. Some of the results are conjectures. The rest of the paper will then be split into sections whose purpose will be either to motivate our proposal in general, or to prove its correctness in certain special but nontrivial cases by explicit computations. \subsection{The basic equations} We propose to describe the local growth of $n$ interfaces in CFT, labeled by an integer $i=1,\cdots,n$ and joining fixed points on the boundary by a Loewner chain. We assume that $0 \leq \kappa <8$ in the following. We list the set of necessary conditions and equations. \vspace{.3cm} \noindent \textbf{Conformal invariance}: The measure on $n$SLE is conformally invariant. Hence it is enough to give its definition when the domain $D$ is the upper half plane $\mathbb H$ in the hydrodynamical normalization. \vspace{.3cm} \noindent \textbf{Universe}: The basic probabilistic objects are $n$ (continuous, local) martingales $M^{(i)}_t$, $i=1,\cdots,n$ with quadratic variation $\kappa q^{(i)}_t$ absolutely continuous with respect to $\ud t$ and vanishing cross variation, defined on an appropriate probability space. By a time change we can and shall assume that $\sum_i q^{(i)}_t \equiv t$. \vspace{.2cm} \noindent \textbf{Driving processes}: The processes $X^{(i)}_t$ are solutions of the stochastic differential equations \begin{equation} \label{eq:drivepro} \ud X^{(i)}_t=\ud M^{(i)}_t + \kappa \ud q^{(i)}_t(\partial_{x_i}\log Z) (X^{(1)}_t,\cdots,X^{(n)}_t)+ \sum_{j \neq i} \frac{2\ud q^{(j)}_t}{X^{(i)}_t-X^{(j)}_t}. \end{equation} The initial conditions are $X_0^{(i)}=X_i$ ordered in such a way that $X_1 < X_2< \cdots < X_n$. \vspace{.2cm} \noindent \textbf{Loewner chain}: The map $f_t$ uniformizing the complement of the hulls satisfies \begin{equation} \label{eq:loewchain} \ud f_t(z)=\sum_i \frac{2\ud q^{(i)}_t}{f_t(z)-X^{(i)}_t}. \end{equation} The initial condition is $f_0(z)=z$. With our conventions, the total capacity of the growing hulls at time $t$ is $2t$. \vspace{.2cm} \noindent \textbf{Auxiliary function}: The system depends on a function $Z(x_1,\cdots,x_n)$ which has to fulfill the following requirements : $i)$ $Z(x_1,\cdots,x_n)$ is defined and positive for $x_1 < x_2< \cdots < x_n$, $ii)$ $Z(x_1,\cdots,x_n)$ is translation invariant and homogeneous. Its weight is $h_{n-2m}(\kappa)-nh_{1}(\kappa)$ for some nonnegative integer $m \leq n/2$, where\footnote{A more traditional notation for $h_{m}(\kappa)$ is $h_{1,m+1}$ in the physics literature.} $$2\kappa h_{m}(\kappa)\equiv m(2(m+2)-\kappa).$$ $iii)$ $Z(x_1,\cdots,x_n)$ is annihilated by the $n$ differential operators $${\mathcal D}_i=\frac{\kappa}{2}\partial_{x_i}^2+2\sum_{j \neq i}\left[ \frac{1}{x_j-x_i}\partial_{x_j}-\frac{h_{1}(\kappa)}{(x_j-x_i)^2}\right].$$ \vspace{.2cm} We call this system of equations the $n$SLE system for $n$ curves joining together the points $X_1,\cdots,X_n$ and possibly the point at infinity. Systems for radial and dipolar versions of $n$SLE could be defined analogously. \subsection{Arch probabilities} It is known from CFT that (relaxing the positivity constraint), the solutions to $i),\;ii),\;iii)$ form a vector space of dimension $d_{n,m}\equiv {n \choose m}-{n \choose m-1}=(n+1-2m)\frac{n!}{m!(n-m+1)!}$. The positive solutions form a cone and from the statistical mechanics interpretation, we conjecture that this cone has the same dimension and is generated by (i.e. is the convex hull of) $d_{n,m}$ half lines (extremal lines, pure states in the sense of statistical mechanics) so that a transverse section of the cone is a simplex. So each solution $Z$ can be written in a unique way as a sum of extremal states. The numbers $d_{n,m}$ have many many combinatorial interpretations, but the one relevant for us is the following. Draw $n+1$ points $X_1<X_2\cdots <X_n < \infty$ ordered cyclically on the (extended) real line bounding the upper half plane $\mathbb H$. Consider $n-m$ disjoint curves in $\mathbb H$ such that each $X_i$ is an end point of exactly $1$ curve and $\infty$ is an end point of exactly $n-2m$ curves. There are $d_{n,m}$ topologically inequivalent configurations, called arch configurations when $n-2m=0$. We keep the same name for $m\neq 0$, writing arch$_m$ configurations when precision is needed. Motivated by this, we claim the following : \vspace{.3cm} a) To each arch configuration $\alpha$ corresponds an extremal state $Z_{\alpha}$ in the following sense : the solution of the $n$SLE system with $Z \propto Z_{\alpha}$ can be defined up to a (possibly infinite) time, at which the growing curves have either paired together or joined the point at infinity and at that time the topology is that of the arch $\alpha$ with probability one. \vspace{.2cm} b) One can decompose a general solution $Z$ of $i),\;ii),\;iii)$ as a sum of $$\sum_{\alpha \,\in \, {\mathrm{arch}_m}} Z_{\alpha}.$$ \vspace{.2cm} c) The probability that a solution of the $n$SLE system with auxiliary function $Z$ ends in arch configuration $\alpha$ is the ratio $$\frac{Z_{\alpha}(X_1,\cdots,X_n)}{Z(X_1,\cdots,X_n)}$$ evaluated at the initial condition $(X_1,\cdots,X_n)$. \vspace{.3cm} The first step toward a heuristic derivation of the above results will be to explain how to construct martingales -- in particular martingales associated to interfaces -- from statistical mechanics observables in a systematic way. But we start with a few comments. \section{First comments} \subsection{Statistical mechanics interpretation} To have a specific example in mind, think again of the Ising model at the critical temperature. Let $a$ be the lattice spacing. First, put $n=2m$ changes of boundary conditions from spins up to spins down and so on along the boundary at points $x_1/a,\cdots,x_n/a$. In the continuum limit when $a \rightarrow 0$ but $x_1,\cdots,x_n$ have a finite limit, the partition function behaves like a homogeneous function $Z(x_1/a,\cdots,x_n/a)$ of weight $0$ (when both $a$ and the $x_i$'s are rescaled) and CFT teaches us that $Z(x_1,\cdots,x_n)$ satisfies $i),\;ii),\;iii)$ for $n=2m$. Then, if $\alpha$ is an arch system, $Z_{\alpha}$ should be (proportional to the continuum limit of) the partial partition function when the sum of Boltzmann weights is performed only over the interface configurations with topology $\alpha$. To make generalized arch configurations, choose $n$ and $m$ with $n\geq 2m$. Put $2n-2m$ changes of boundary conditions from spins up to spins down and so on along the boundary, $n$ at points $x_1/a,\cdots,x_n/a$ and $n-2m$ at $y_1/a,\cdots, y_{n-2m}/a$. Sum only over configurations where the interfaces do not joint two $y$-type points to each other. Take the continuum limit for the $x$'s as before, but impose that all $y$'s go to infinity and remain at a finite number of lattice spacings from each other. This is expected to lead again to a partition function $Z(x_1/a,\cdots,x_n/a)$ of weight $0$ (when both $a$ and the $x_i$'s are rescaled) and $Z(x_1,\cdots,x_n)$ satisfies $i),\;ii),\;iii)$ for the given $n$ and $m$. If $\alpha$ is an arch$_m$ configuration, $Z_{\alpha}$ should be (proportional to the continuum limit of) the partial partition function when the sum of Boltzmann weights is performed only over the interface configurations with topology $\alpha$. Note that the prefactor between the continuum limit finite part and the real partition function is a power of the lattice spacing. The power depends on $m$, so it is likely to be unphysical to use a non homogeneous $Z$ in the $n$SLE system, mixing different values of $m$ for a fixed $n$. However, we shall later treat the example $n=2$ mixing $m=0$ and $m=1$ because it is illustrative despite the fact that it breaks scale invariance. \subsection{SLE as a special case of $2$SLE} For $n=2$ the solution of $i),\;ii),\;iii)$ with $m=1$ is elementary. Writing $x_1=a$ and $x_2=b$, one finds $Z\propto (b-a)^{(\kappa -6)/\kappa}$. Taking the first martingale to be a Brownian and the second one to be $0$, one retrieves the equations for SLE growing from point $a$ to point $b$ in the hydrodynamical normalization. Let us recall briefly why. We start from SLE from $0$ to $\infty$. The basic principle of conformal invariance makes the passage from this special case to the case when SLE goes from point $a$ to point $b$ on $\bH$ a routine task. If $u$ is any linear fractional transformation (i.e. any conformal transformation) from $\bH$ to itself mapping $0$ to $a$ and $\infty$ to $b$, the image of the SLE trace or hull from $0$ to $\infty$ by $u$ is by definition an SLE trace from $a$ to $b$ and this is measure preserving. The new uniformizing map is $h_t=u\circ g_t \circ u^{-1}$ and it is readily checked that $\frac{\ud h_t}{\ud t}$ is a rational function of $h_t$ whose precise form can be easily computed but does not concern us. Let us just mention that this rational function is regular everywhere (infinity included) except for a simple pole at $h_t=u(\xi_t)$ and has a third order zero at $h_t=u(\infty)=b$. So the map $h_t$ is normalized in such a way that $h_t(b+\varepsilon)=b+\varepsilon+O(\varepsilon^3)$, which is not the hydrodynamic normalization. But if $v_t$ is any linear fractional transformation, $v_t \circ g_t \circ u^{-1}$ describes the same trace as $h_t=u \circ g_t \circ u^{-1}$. As long as the trace does not separate $b$ from $\infty$, i.e. as long as the trace has not hit the real axis in the segment $]b,\infty[$, i.e. as long as $\infty$ is not in the hull, $v_t$ can be adjusted in such a way that $\tilde{h}_t\equiv v_t \circ g_t \circ u^{-1}$ is normalized hydrodynamically. Then $\ud \tilde{h}_t/\ud t$ is a function of $\tilde{h}_t$ which is regular everywhere but for a single pole and vanishes at infinity, i.e. one can write $\ud \tilde{h}_t/\ud t=2\mu_t/(\tilde{h}_t-\alpha_t)$. The following computation is typical of the manipulations made with SLE (see e.g. \cite{LSW:ConformalRestriction}). Write $(g_t \circ u^{-1})(z)=w$ an compute from the definition $$\frac{\ud \tilde{h}_t}{\ud t}(z)=\frac{\ud v_t}{\ud t}(w)+ v_t^{'}(w)\frac{2}{w-\xi_t}.$$ Comparison gives $$\frac{\ud v_t}{\ud t}(w)=\frac{2\mu_t}{v_t(w)-\alpha_t}-\frac{2v_t^{'}(w)}{w-\xi_t}.$$ But $v_t$ is regular at $w=\xi_t$ from which one infers that $v_t(\xi_t)=\alpha_t$ (the poles in the two terms are at the same point) and $\mu_t=v_t^{'}(\alpha_t)^2$ (the two residues add to $0$). Going one step further in the expansion close to $\xi_t$ yields $\frac{\ud v_t}{\ud t}(\xi_t)= -3v_t^{''}(\xi_t).$ Ito's formula gives $\ud \alpha_t=-3v_t^{''}(\xi_t)\ud t+v_t^{'}(\xi_t)\ud \xi_t+\frac{\kappa}{2}v_t^{''}(\xi_t)\ud t.$ So the time change $\mu_t \ud t =\ud s $ together with the definition $\ud \chi_s =v_t^{'}(\xi_t)\ud \xi_t$ yields $$\ud \alpha_t(s)=\ud \chi_s+(\kappa-6)\frac{v_t^{''}(\xi_t)}{2v_t^{'}(\xi_t)^2}\ud s.$$ But $v_t^{''}(w)/v_t^{'}(w)^2=2/(v_t(w)-v_t(\infty))$ because $v_t$ is a linear fractional transformation. Finally, setting $\tilde{h}_{t(s)}\equiv f_s$, $\tilde{h}_{t(s)}(b)=v_{t(s)}(\infty)\equiv B_s$ and $v_{t(s)}(\xi_{t(s)})=\alpha_{t(s)}\equiv A_s$ we can summarize $$\frac{\ud f_s}{\ud s}=\frac{2}{f_s-A_s}\; , \quad \frac{\ud B_s}{\ud s}=\frac{2}{B_s-A_s} \; , \quad \ud A_s= \ud \chi_s +(\kappa-6)\frac{\ud s}{A_s-B_s},$$ where $\chi_s$ is a Brownian motion with quadratic variation $\kappa s$, $f_0=id$, $A_0=a$, $B_0=b$. Thus chordal SLE from $a$ to $b$ in the hydrodynamical normalization is indeed a special case of $2$SLE. The above equations are also a special case of SLE$(\kappa,\rho)$ $(\rho=\kappa-6)$, but it should be clear that our general proposal goes in a different direction. As already mentioned, the description of chordal SLE from $a$ to $b$ in the hydrodynamical normalization in fact coincides with chordal SLE from $a$ to $b$ only up to the first time $b$ is separated from $\infty$ by the trace. This time is infinite for $\kappa \leq 4$, but it is finite with probability $1$ for $4 < \kappa < 8$. The most obvious case is $\kappa=6$. The equation is nothing but the usual chordal SLE$_6$ ending at infinity, a consequence of locality (in the SLE sense, not in the quantum field theory sense used later). At that time, the real chordal SLE from $a$ to $b$ swallows $\infty$, whereas the hydrodynamically normalized version swallows $b$. The solution to this problem is to use conformal invariance and restart the process again in the correct domain at the time when $b$ and $\infty$ get separated by the trace. But this is not coded in the equations. \subsection{Making sense} The previous example should serve as a warning. Some serious mathematical work may have to be done even to make sense of our conjectures, let alone prove their correctness. The problems might be of different natures for $\kappa \leq 4$ and $4 < \kappa <8$. We content with the following naive remarks. One of the problems is that the arches do not have to close at the same time. It may even happen that one of the growing curves touches the real line or another curve in such a way that the upper half plane is split in two domains and the one which is swallowed contains some of the growing curves. Our putative description of $n$SLE processes can be valid in this form only up the realization of such an event. The first thing to check should be that the event is realized with a probability obtained by summing $Z_{\alpha}/Z$ over all $\alpha$'s corresponding to compatible configurations (see figure \ref{fig: compatible configurations}). In particular, the connected component of $\infty$ should contain at least $n_{\infty} \geq m-1$ curves for consistency, but that's not an obvious property of our proposal. \begin{figure} \includegraphics[width=1.0\textwidth]{someconfigs.eps} \caption{\emph{The probability of closing of an arch should be obtained by summing $Z_\alpha/Z$ over all $\alpha$'s corresponding to compatible configurations. Two compatible configurations are portrayed in the figure.}} \label{fig: compatible configurations} \end{figure} Consider the fate of the connected component of $\infty$. If $n_1-m$ is even, conformal invariance suggests to continue the Loewner evolution simply by suppressing the points that have been swallowed, i.e. for the $n_1$ remaining points. If $n_1-m$ is odd, the same should be done, but the image of the point at which one interface has made a bridge should be included as a starting point for the continuation of the evolution. Preferably, the function $Z$ for this new $multi$SLE system should not be adjusted by hand to make our conjectures correct, but should appear as a natural limit. We shall make comments on this and give concrete illustrations later. For the component that is swallowed, one can use conformal invariance again to change the normalization of the Loewner map in such a way that this component is the one that survives and then restart a new $multi$SLE for the appropriate number of points. This procedure may have to be iterated. Note also that our conjectures for arch probabilities do not involve any details on the martingales $M^{(i)}_t$. Indeed, we expect that there is some robustness. But the precise criteria are beyond our understanding. \subsection{A few martingales for $n$SLEs} Our heuristic derivation of the $n$SLE system will in particular show that if $\tilde{Z}$ also solves $i),\;ii),\;iii)$ (even relaxing positivity), the quotient $$\frac{\tilde{Z}(X^{(1)}_t,\cdots,X^{(n)}_t)} {Z(X^{(1)}_t,\cdots,X^{(n)}_t)}$$ is a local martingale. This can be proved directly using Ito's formula. In particular, $$\frac{Z_{\alpha}(X^{(1)}_t,\cdots,X^{(n)}_t)} {Z(X^{(1)}_t,\cdots,X^{(n)}_t)}$$ is a local martingale bounded by $1$, hence a martingale. On the other hand, a standard argument shows that if $P_{\alpha}$ is the probability that the system ends in a definite arch configuration $\alpha$ (once one has been able to make sense of it) $P_{\alpha}(X^{(1)}_t,\cdots,X^{(n)}_t)$ is a martingale. This is an encouraging sign. To get a full proof, one would need to analyze the behavior of $Z_{\alpha}(X^{(1)}_t,\cdots,X^{(n)}_t)$ when one arch closes, or when one growing curve cuts the system in two, to get recursively a formula that looks heuristically like $$\frac{Z_{\alpha}(X^{(1)}_t,\cdots,X^{(n)}_t)} {Z(X^{(1)}_t,\cdots,X^{(n)}_t)}\sim \delta_{\alpha,\alpha'}$$ if the system forms asymptotically the arch system $\alpha'$ at large $s$. Such a formula rests on properties of $Z_{\alpha}(x_1,\cdots,x_n)$ when some points come close together in a way reminiscent to the formation of arch $\alpha'$ : $Z_{\alpha'}(x_1,\cdots,x_n)$ should dominate all $Z_{\alpha}$'s, $\alpha \neq \alpha'$ in such circumstances. In section \ref{sec:sevinterf} we shall use this to expand explicitly the $Z_{\alpha}$'s in a basis of solutions to $i),\;ii),\;iii)$ which is familiar from CFT, very explicitly at least for small $n$. \subsection{Classical limit} Our proposal for $n$SLE has a non trivial classical limit at $\kappa \rightarrow 0^+$. The martingales $M_t^{(i)}$ vanish in this limit, but the $q_t^{(i)}$ remain arbitrary increasing functions. The function $Z$ does not have a limit, but the $U_i\equiv \kappa \partial_{x_i} \log Z$ do. They are kind of Ricatti variables for which the equations read $$ \frac{1}{2} \left(\partial_{x_i}U_i + \frac{U_i^2}{\kappa}\right)+2\sum_{j \neq i} \left(\frac{1}{x_j-x_i}\frac{U_i}{\kappa}- \frac{6-\kappa}{2\kappa}\frac{1}{(x_j-x_i)^2}\right)=0, $$ which have a limit when $\kappa \rightarrow 0^+$, comparable to the classical limit of a Schroedinger equation. To summarize, the classical limit is $$\ud f_t(z)=\sum_i \frac{2\ud q^{(i)}_t}{f_t(z)-X^{(i)}_t}.$$ $$ \ud X^{(i)}_t=U_i(X^{(1)}_t,\cdots,X^{(n)}_t) \ud q^{(i)}_t +\sum_{j \neq i} \frac{2\ud q^{(j)}_t}{X^{(i)}_t-X^{(j)}_t}.$$ where the auxiliary functions $U_i(x_1,\cdots,x_n)$ are homogeneous functions of degree $-1$ which satisfy $\partial_{x_i}U_j=\partial_{x_j}U_i$ and $$ \frac{1}{2} U_i^2+2\sum_{j \neq i}\left(\frac{1}{x_j-x_i}U_i- \frac{3}{(x_j-x_i)^2}\right)=0.$$ It is not too surprising that the differential equations for $Z$ have become algebraic equations for the $U_i$'s, so that the space of solutions which was a connected manifold for $\kappa \neq 0$ concentrates on a finite number of points in the classical limit. The classical system, maybe with an educated guess for the $q_t^{(i)}$'s, could be interesting for its own sake. \subsection{Relations with other work} Several processes involving several growing curves have appeared in the literature. The first proposal was made by Cardy \cite{Cardy:nSLE}. It can be formally obtained from ours by forgetting the conditions $i),\;ii),\;iii)$ and choosing a constant $Z$. The corresponding processes are interesting, but the relationship with interfaces in statistical mechanics and CFT is unclear for us. Dub\'edat \cite{D:SLEcommutation} has derived a general criterion he calls commutativity to constrain the class of processes that could possibly be related to interfaces. Our proposal satisfies commutativity so they can be viewed as a special case satisfying other relevant physical constraints. Dub\'edat also came with a special solution of commutativity. It corresponds to the case $m=0$ in our language. Then the space of solutions has dimension $d_{n,0}=1$ and the corresponding partition function is elementary: \begin{equation} \label{eq:dabdouwa}Z \propto \prod_{i <j} (x_j-x_i)^{2/\kappa}. \end{equation} \begin{figure} \center{\includegraphics[width=0.5\textwidth]{factorizable.eps}} \caption{\emph{The factorisable $Z$ leads to a very simple geometry. This case has been suggested previously with a slightly different approach.}} \label{fig: factorizable} \end{figure} A single arch topology is possible, all interfaces converge to $\infty$, see figure \ref{fig: factorizable}. Maybe this is a good reason to call this case chordal $n$SLE. \section{CFT background} There was never any doubt that SLEs are related to conformal field theories. The original approach \cite{Bb:2002tf,Bb:2003vu,Bb:2003kd,Bb:2004ij} used the operator formalism because if yields naturally martingale generating functions. Here, we use the correlator approach for a change. We restrict the presentation to a bare minimum, referring the newcomer to the many articles, reviews and books on the subject (\cite{DMS:CFT,BPZ}). The reader who knows too little or too much about CFT can profitably skip this section. Observables in CFT can be classified according to their behavior under conformal maps. Local observables in quantum field theory are called fields. For instance, in the Ising model, on an arbitrary (discrete) domain, the average value of a product of spins on different (well separated) sites can be considered. Taking the continuum limit at the critical point, we expect that on arbitrary domains $D$ there is a local observable, the spin. The product of two spins at nearest neighbor points corresponds to the energy operator. In the continuum limit, this will also lead to a local operator. In this limit, the lattice spacing has disappeared and one can expect a definite (but nontrivial) relationship between the energy operator and the product of two spin fields close to each other. As on the lattice the product of two spins at the same point is $1$, we can expect that the identity observable also appears in such a product at short distances. Local fields come in two types, bulk fields whose argument runs over $D$ and boundary fields whose argument runs over $\partial D$. In this paper, we shall not need bulk fields so we leave them aside. The simplest conformal transformations in the upper-half plane are real dilatations and boundary fields can be classified accordingly. It is customary to write $\varphi_{\delta}(x)$ to indicate that in a real dilatation by a factor $\lambda$ the field $\varphi_{\delta}(x)$ picks a factor $\lambda^{\delta}$. By a locality argument, boundary fields in a general domain $D$ (not invariant under dilatations) can still be classified by the same quantum number. The number $\delta$ is called the conformal weight of $\varphi_{\delta}$. There are interesting situations in which (due to degeneracies) the action of dilatations cannot be diagonalized, leading to so called logarithmic CFT. While this more general setting is likely to be relevant for several aspects of SLE, we shall not need it in what follows. Under general conformal transformations, the simplest objects in CFT are so called primary fields. Their behavior is dictated by the simplest generalization of what happens under dilatations. Suppose $\varphi_{\delta_1},\cdots \varphi_{\delta_n}$ are boundary primary fields of weights $\delta_1,\cdots,\delta_n$. If $f$ is a conformal map from domain $D$ to a domain $D'$, CFT postulates that $$\bra \prod_{j=1}^n \varphi_{\delta_j}(x_j) \ket^D = \bra \prod_{j=1}^n \varphi_{\delta_j}(f(x_j)) \ket^{f(D)} \prod_{j=1}^n |f'(x_j)|^{\delta_j}. $$ Symbolically, this can be written $f:\varphi_{\delta}(x) \rightarrow \varphi_{\delta}(f(x))|f'(x)|^{\delta}$. It is interesting to make a comparison of these axioms with the previous computations relating chordal SLE from $0$ to $\infty$ to chordal SLE from $a$ to $b$ in several normalizations. This also involved pure kinematics. As usual in quantum field theory, to a symmetry corresponds an observable implementing it. In CFT, this leads to the stress tensor $T(z)$ whose conservation equation reduces to holomorphicity. The fact that conformal transformations are pure kinematics translates into the fact that insertions of $T$ in known correlation functions can be carried automatically, at least recursively. The behavior of $T(z)$ under conformal transformations can be written as $f:T(z) \rightarrow T(f(z))f'(z)^{2}+c/12Sf(z)$ where $Sf\equiv (f''/f')'-1/2(f''/f')^2$ is the Schwarzian derivative and $c$ is a conformal anomaly, a number which is the most important numerical characteristic of a CFT. When $c=0$, $T$ is be a $(2,0)$ primary field i.e. an holomorphic quadratic differential. When a (smooth) boundary is present, the Schwarz reflection principle allows to extend $T$ by holomorphicity. Holomorphicity also implies that if $O$ is any local (bulk or boundary) observable at point $z \in D$ and $v$ is vector field meromorphic close to $z$, the contour integral $L_vO \equiv\oint_z dw v(w)T(w)O$ along an infinitesimal contour around $z$ oriented counterclockwise is again a local field at $z$, corresponding to the infinitesimal variation of $O$ under the map $f(w)=w+\varepsilon v(w)$. It is customary to write $L_n$ for $v(w)=w^{n+1}$. It is one of the postulates of CFT that all local fields can be obtained as descendants of primaries, i.e. by applying this construction recursively starting from primaries. The correlation functions of descendant fields are obtained in a routine way from correlations of the primaries. But descendant fields do not transform homogeneously. When $v$ is holomorphic at $x$, $L_vO$ is a familiar object. For instance, if $\varphi_{\delta}$ is a primary boundary field, one checks readily that $L_n\varphi_{\delta}=0$ for $n\geq 1$, $L_0\varphi_{\delta}=\delta \varphi_{\delta}$ and $L_{-1}\varphi_{\delta}=\Re e \; [\partial_x\varphi_{\delta}]$. The other descendants are in general more involved, but by definition the stress tensor $T=L_{-2}Id$ is the simplest descendant of the identity $Id$. It does indeed not transform homogeneously. A primary field and its descendants form what is called a conformal family. Not all linear combinations of primaries and descendants need to be independent. The simplest example is the identity observable, which is primary with weight $0$ and whose derivative along the boundary vanishes identically\footnote{For other primary fields with the same weight if any, this does not have to be true.}. By contour deformation, this leads to translation invariance of correlation functions when $D$ has translation symmetry. The next example in order of complexity is of utmost importance for the rest of this paper. If $(2h+1)c=2h(8h-5)$, the field $$-2(2h+1)L_{-2}\varphi_{h}+3L_{-1}^2\varphi_{h}$$ is again a primary, i.e. it transforms homogeneously under conformal maps. In this case, consistent CFTs can be constructed for which it vanishes identically. This puts further constraints on correlators. For example, when $D$ is the upper half plane, so that the Schwarz principle extends $T$ to the full plane, the contour for $L_{-2}$ can be deformed and shrunken at infinity. Then, for an arbitrary boundary primary correlator one has \begin{eqnarray} \label{eq:sing} \left(\frac{3}{2(2h+1)}\partial_x^2 +\sum_{\alpha=1}^{l}\left[ \frac{1}{y_{\alpha}-x}\partial_{y_{\alpha}}- \frac{\delta_{\alpha}}{(y_{\alpha}-x)^2} \right]\right) & & \nonumber \\ & & \hspace{-3cm} \bra \varphi_{\delta}(\infty) \prod_{\alpha=1}^l \varphi_{\delta_{\alpha}}(y_{\alpha}) \varphi_{h}(x) \ket=0. \end{eqnarray} It is customary to call this type of equation a null-vector equation. Note that the primary field of weight $\delta$ sitting at $\infty$ has led to no contribution in this differential equation. Working the other way round, this equation valid for an arbitrary number of boundary primary fields with arbitrary weights characterizes the field $\varphi_{h}$ and the relation between $h$ and the central charge $c$. The case of three points correlators is instructive. Global conformal invariance implies that $$\bra \varphi_{\delta}(y)\varphi_{\delta'}(y') \varphi_{h}(x) \ket \propto |y-y'|^{h-\delta-\delta'} |x-y|^{\delta'-h-\delta}|y'-x|^{\delta-\delta'-h}.$$ The proportionality constant might depend on the cyclic ordering of the three points. But if the differential equation for $\varphi_{h}$ is used, a further constraint appears. The three point function can be non vanishing only if $$ 3(\delta-\delta')^2-(2h+1)(\delta+\delta')= h(h-1).$$ This computation has a dual interpretation : consider a correlation function with any number of fields, among them a $\varphi_{\delta}(y)$ and a $\varphi_{h}(x)$. If $x$ and $y$ come very close to each other they can be replaced by an expansion in terms of local fields. This is called fusion. Several conformal families can appear in such an expansion, but within a conformal family, the most singular contribution is always from a primary. This argument applies even if $c$ and $h$ are arbitrary. But suppose they are related as above and the differential equation eq.(\ref{eq:sing}) is valid. This equation is singular at $x=y$ and at leading order the dominant balance leads to an equation where the other points are spectators. One finds that the only conformal families that can appear are the ones whose conformal weight $\delta'$ satisfies the fusion rule. This is enough CFT background for the rest of this paper. We are now in position to give the heuristic argument that leads to our main claims. \section{Martingales from statistical mechanics} \label{sec:StatMech} The purpose of this section is to emphasize the intimate connection between the basic rules of statistical mechanics and martingales. The connection is somehow tautological, because statistical mechanics works with partition functions, i.e. unnormalized probability distributions, all the time. In the discrete setting, this makes conditional expectations a totally transparent operation that one performs without thinking and even without giving it a name. But the following argument is, despite its simplicity and its abstract nonsense flavor, the crucial observation that allows us to relate interfaces in conformally invariant statistical mechanics to SLEs. \subsection{Tautological martingales} Consider a model of statistical mechanics with a finite but arbitrarily large set of possible states $S$. Usually one starts with models defined on finite grid domains so $\# S < \infty$ is natural. To each state $s \in S$ we associate a Boltzmann weight\footnote{Usually the Boltzmann weight is related to the energy $H(s)$ of the state $s$ through $w(s) = \exp ( - \beta H(s))$, where $\beta$ is the inverse temperature (a Lagrangian multiplier related to temperature, anyway).} $w(s)$. The partition function is $Z = \sum_{s \in S} w(s)$ so that it normalizes the Boltzmann weights to probabilities, $\prob \{ s \} = \frac{w(s)}{Z}$. Since $S$ is finite, we can use the power set $\sP (S) = \{ U : U \subset S \}$ as a sigma algebra. The expected value of a random variable $\Oper : S \rightarrow \bC$ is denoted by $\expect [\Oper] = \bra \Oper \ket = \frac{1}{Z} \sum_{s \in S} \Oper(s) w(s)$. Note that if $(S_\alpha)_{\alpha \in I}$ is a collection of disjoint subsets of $S$ such that $\cup_{\alpha \in I} S_\alpha = S$, then the collection of all unions $\sF = \{ \cup_{\alpha \in I'} S_\alpha : I' \subset I \}$ is a sigma algebra on $S$. Conversely, since $S$ is finite, any sigma algebra $\sF$ on $S$ is of this type. Consider a filtration, that is an increasing family $(\sF_t)_{t \geq 0}$ of sigma algebras $\{ \emptyset , S \} \subset \sF_s \subset \sF_t \subset \sP(S)$ for all $0 \leq s < t$. Denote the corresponding collections of disjoint sets by $(S^{(t)}_{\alpha})_{\alpha \in I_t}$ and define the partial partition function $Z^{(t)}_{\alpha}\equiv \sum_{s \in S^{(t)}_\alpha} w(s)$. The conditional expectation values \begin{eqnarray*} \bra \Oper \ket_t & \equiv & \expect [ \Oper | \sF_t]= \sum_{\alpha \in I_t} \frac{\sum_{s \in S^{(t)}_\alpha} \Oper(s) w(s)} {\sum_{s \in S^{(t)}_\alpha} w(s)} \; \unit_{S^{(t)}_\alpha} \\ & = & \sum_{\alpha \in I_t} \Big( \frac{1}{Z^{(t)}_{\alpha}} \sum_{s \in S^{(t)}_\alpha} \Oper(s) w(s) \Big) \; \unit_{S^{(t)}_\alpha} \end{eqnarray*} are martingales by definition: for $s < t$ we have \begin{eqnarray*} \expect \big[ \; \expect [ \Oper | \sF_t] \; \big| \sF_s \big] = \expect [ \Oper | \sF_s] \end{eqnarray*} Notice that the probability of the event $S^{(t)}_\alpha$ is conveniently $\prob [S^{(t)}_{\alpha}] = Z^{(t)}_\alpha / Z$. Suppose that the model is defined in a domain $D \subset \bC$ and that there are interfaces in the model. Parametrize portions of these interfaces touching the boundary by an arbitrary ``time'' parameter $t$ in such a way that $n$ paths $\gamma^{(i)}_t$, $i=1,\cdots,n$ (which are pieces of the random interfaces) emerge from the boundary at $t=0$ and are disjoint at least when $t$ is small enough, see figure \ref{fig:statmech}. To avoid confusion we write the time parameter $t$ now as a subscript and continue to indicate the dependence of $s \in S$ by parenthesis, so $t \mapsto \gamma^{(i)}_t (s)$ is a parametrization of the $i^{\textrm{th}}$ piece of interface if the system is at state $s$. Then we can consider the natural filtration of the interface by taking $\sF_t = \sigma ( \gamma^{(i)}_{t'} : 0 \leq t' \leq t, \; i=1,\cdots,n )$ to be the sigma algebra generated by the random variables $\gamma^{(i)}_{t'}$ up to time $t$. \begin{figure} \center{\includegraphics[width=0.6\textwidth]{statmech.eps}} \caption{\emph{A discrete statistical mechanics model with portions of interfaces specified.}} \label{fig:statmech} \end{figure} The boundary conditions of the model are often such that conditioning on the $\gamma^{(i)}_{[0,t]}$, is the same as considering the model in a smaller domain (a part of the interface removed) but with same type of boundary conditions. Of course the position at which the new interface should start is where the original interface would have continued, that is the $\gamma^{(i)}_t$'s. Let $D_t$ be the domain $D$ with the $\gamma^{(i)}_{]0,t]}$ removed. The starting point of the next section is the input of conformal invariance in this setup. \subsection{Simplifying tautological martingales} We start from the situation and notations at the end of the previous section. If in addition we are considering a model at its critical point, then the continuum limit may be described by a conformal field theory. At least for a wide class of natural observables $\Oper$, the expectation values become CFT correlation functions in the domain $D$ of the model \begin{eqnarray*} \bra \Oper \ket = \frac{\sum_{s \in S} \Oper(s) w(s)}{Z} \longrightarrow \frac{\bra \Oper \ket^{\textrm{CFT, b.c.}}_{D}} {\bra \unit \ket^{\textrm{CFT, b.c.}}_{D}} \end{eqnarray*} We need to write the correlation function of identity (proportional to $Z$) in the denominator because the boundary conditions (b.c.) of the model may already have led to insertions of boundary changing operators that we have not mentioned explicitly. The closed martingales become \begin{eqnarray*} \bra \Oper \ket_t = \sum_{\alpha \in I_t} \frac{1}{Z^{(t)}_\alpha} \sum_{s \in S^{(t)}_\alpha} \Oper(s) w(s) \; \unit_{S^{(t)}_\alpha} \longrightarrow \frac{\bra \Oper \ket^{\textrm{CFT, b.c.}}_{D_t}} {\bra \unit \ket^{\textrm{CFT, b.c.}}_{D_t}} \end{eqnarray*} where in the continuum limit $D_t$ might be $D$ with hulls (and not only traces) removed. For certain (but not all) observables, $\bra \Oper \ket$ is computing a probability, which in a conformal field theory ought to be conformally invariant. But $\bra \Oper \ket$ is written as a quotient, and this means that the numerator and denominator should transform homogeneously (and with the same factor) under conformal transformations. In particular, the denominator should transform homogeneously. This means that $\bra \unit \ket^{\textrm{CFT, b.c.}}_{D}$ -- which depends on the position of the boundary condition changes -- behaves like a product of boundary primary fields. Then, by locality, for any $\Oper$, the transformation of the numerator under conformal maps will split in two pieces: one containing the transformations of $\Oper$ and the other one canceling with the factor in the denominator. So we infer the existence in the CFT of a primary boundary field, denoted by $\psi (x)$ in what follows, which implements boundary condition changes at which interfaces anchor. Hence we may write $$\bra \unit \ket^{\textrm{CFT, b.c.}}_{D} = \bra \psi (X^{(1)}) \cdots \psi (X^{(n)})\ket^{\textrm{CFT}}_{D}$$ and $$\bra \Oper \ket^{\textrm{CFT, b.c.}}_{D} = \bra \Oper \psi (X^{(1)}) \cdots \psi (X^{(n)})\ket^{\textrm{CFT}}_{D}.$$ As will become clear later, there might also be one further boundary operator anchoring several interfaces. We do not mention it explicitly here because it will sit at a point which will not be affected by the conformal transformations that we use. Write the transformation of the observable $\Oper$ as $f:\Oper \rightarrow \; ^f \!\Oper$ under a conformal map. Denote by $f_t$ a conformal representation $f_t:D_t\rightarrow D$ and write $f(\gamma^{(i)}_{t}) \equiv X^{(i)}_{t}$. The expression for the closed martingale $\bra \Oper \ket_t$ can now be simplified further \begin{equation} \label{eq:martCFT} \bra \Oper \ket_t \longrightarrow \frac{\bra \, ^{f_t}\Oper\, \psi (X^{(1)}_t) \cdots \psi (X^{(1)}_t)\ket^{\textrm{CFT}}_{D}} {\bra \psi (X^{(1)}_t) \cdots \psi (X^{(1)}_t) \ket^{\textrm{CFT}}_{D}}. \end{equation} The Jacobians coming from the transformations of the boundary changing primary field $\psi$ have canceled in the numerator and denominator. The explicit value of the conformal weight of $\psi$ does not appear in this formula. Of course, we have cheated. For the actual map $f_t$ which is singular at the $\gamma^{(i)}_{t}$'s these Jacobians are infinite. A more proper ``derivation'' would go through a regularization but locality should ensure that the naive formula remains valid when the regularization is removed. Eq.(\ref{eq:martCFT}) is the starting point of our analysis. \section{Derivation of the proposal} The heuristics we follow is to describe a growth process of interfaces by a Loewner chain $f_t$ compatible with conformal invariance in that the right hand side of eq.(\ref{eq:martCFT}) is a martingale. \subsection{The three ingredients} \noindent \textbf{Loewner chain}: If we use the upper half plane as a domain, $D={\mathbb H}$, and impose the hydrodynamic normalization, the equation for $f_t$ has to be of the form $$\ud f_t(z)=\sum_i \frac{2\ud q^{(i)}_t}{f_t(z)-X^{(i)}_t}$$ for some processes $X^{(i)}_t$, $i=1,\ldots ,n$. The initial condition is $f_0(z)=z$. \vspace{.3cm} \noindent \textbf{Interfaces grow independently of each other on very short time scales}: Schramm's argument deals with the case of a single point. We expect that on very short time scales the growth processes do not feel each other and Schramm's argument is still valid, so that $\ud X^{(i)}_t=\ud M^{(i)}_t + F^{(i)}_t $ where the $M^{(i)}_t$'s are $n$ (continuous,local) martingales with quadratic variation $\kappa q^{(i)}_t$ and vanishing cross variation and $F^{(i)}_t$ is a drift term. \vspace{.3cm} \noindent \textbf{The martingale property fixes the drift term}: The drift term will be computed by imposing the martingale condition on $\bra \Oper \ket_t$ when $\Oper$ is a product of an arbitrary number $l$ of boundary primary fields $\Oper =\prod_{\alpha=1}^{l} \varphi_{\delta_\alpha}(Y^{(\alpha)})$. The insertion points are away from the boundary changing operators and $f_t$ is regular with positive derivative there. Substitution of $^{f_t}\Oper$ in formula (\ref{eq:martCFT}) yields \begin{equation} \label{eq:marto} \bra \prod_{\alpha=1}^{l} \varphi_{\delta_\alpha}(Y^{(\alpha)})\ket_t = \frac{\bra \prod_{\alpha=1}^{l} \varphi_{\delta_\alpha}(f_t(Y^{(\alpha)})) \prod_{i=1}^{n}\psi (X^{(i)}_t) \ket^{\textrm{CFT}}_{D}} {\bra \prod_{i=1}^{n}\psi (X^{(i)}_t) \ket^{\textrm{CFT}}_{D}} \prod_{\alpha=1}^{l} f'_t(Y^{(\alpha)})^{\delta_\alpha}. \end{equation} \subsection{Computation of the Ito derivative of $\bra \prod_{\alpha=1}^{l} \varphi_{\delta_\alpha}(Y^{(\alpha)})\ket_t$} In formula (\ref{eq:marto}), denote respectively by $Z_t^{\varphi}$, $Z_t$ and $J^{\varphi}_t$ the numerator, denominator and Jacobian factor on the right hand side. It is useful to break the computation of $\ud \bra \prod_{\alpha=1}^{l} \varphi_{\delta_\alpha}(Y^{(\alpha)})\ket_t$ in several steps. \vspace{.3cm} -- \textit{Preliminaries}. \\ Ito's formula for the $\psi$'s gives $$\ud \psi(X^{(i)}_t)=\psi'(X^{(i)}_t)(\ud M^{(i)}_t + F^{(i)}_t)+\frac{\kappa}{2} \psi''(X^{(i)}_t)\ud q^{(i)}_t. $$ \noindent Using the Loewner chain for $f_t(z)$ and its derivative with respect to $z$, one checks that $$\ud \left(\varphi_{\delta}(f_t(Y))f'_t(Y)^\delta\right)=f'_t(Y)^\delta \sum_i 2\ud q^{(i)}_t\left(\frac{\varphi'_{\delta}(f_t(Y))}{f_t(Y)-X^{(i)}_t}- \frac{\delta\varphi_{\delta}(f_t(Y))}{(f_t(Y)-X^{(i)}_t)^2}\right).$$ \vspace{.2cm} -- \textit{The Ito derivative of $Z_t^{\varphi}J^{\varphi}_t$}.\\ The time $t$ being given, we can simplify the notation. Set $x_i \equiv X^{(i)}_t$ and $y_{\alpha} \equiv f_t(Y^{(\alpha)})$ and apply the chain rule to get \begin{eqnarray*} \frac{\ud (Z_t^{\varphi}J^{\varphi}_t)}{J^{\varphi}_t} & = & \left[\sum_i \left(\ud M^{(i)}_t + F^{(i)}_t\right)\partial_{x_i}\right. \\ & & \hspace{-2cm} + \left. \sum_i \ud q^{(i)}_t\left(\frac{\kappa}{2} \partial_{x_i}^2+2\sum_{\alpha}\left[ \frac{1}{y_{\alpha}-x_i} \partial_{y_{\alpha}}-\frac{\delta_{\alpha}} {(y_{\alpha}-x_i)^2}\right]\right)\right]Z_t^{\varphi} \end{eqnarray*} \vspace{.2cm} -- \textit{First use of the null-vector equation : identification of $\psi$}. \\ Let us concentrate for a moment on the familiar chordal SLE case, for which $n=1$. The drift term $F_t^{(1)}$ is known to be zero. The boundary conditions also change at $\infty$ (the endpoint of the interface) and there is an operator there, that we have not written explicitly because the notation is heavy enough. Anyway, $Z_t$ is a two-point function with one of the fields at infinity, so it is a constant. For chordal SLE, the drift term in the Ito derivative of the putative martingale vanishes if and only if $$\left(\frac{\kappa}{2} \partial_{x}^2+2\sum_{\alpha}\left[ \frac{1}{y_\alpha-x} \partial_{y_\alpha}-\frac{\delta_{\alpha}} {(y_\alpha-x)^2}\right]\right)Z_t^{\varphi}=0,$$ where for simplicity we have written $x \equiv x_1$.\\ Comparison with eq.(\ref{eq:sing}) implies that $\psi$ has a vanishing descendant at level two and has conformal weight $h=\frac{6-\kappa}{2\kappa}\equiv h_1(\kappa)$ : $$\psi (x)\equiv \varphi_{h_1(\kappa)}(x).$$ The central charge is $c=\frac{(6-\kappa)(3\kappa-8)}{16\kappa}$. \\ This is of course nothing but the correlation function formalism version of the original argument relating SLE to CFT, which was given in the operator formalism, see \cite{Bb:2002tf}. \vspace{.2cm} -- \textit{Second use of the null-vector equation}.\\ Now that $\psi$ has been identified, we can return to the general case, with an arbitrary number $n$ of growing curves. Each growing curve has its own field $\psi$ and each field $\psi$ comes with its differential equation, which is eq.(\ref{eq:sing}) but for $l+n-1$ spectator fields, the $l$ fields $\varphi$ and the $n-1$ other $\psi$'s. So $Z_t^{\varphi}$ is annihilated by the $n$ differential operators $$ \frac{\kappa}{2} \partial_{x_i}^2+2\sum_{\alpha} \left[ \frac{1}{y_\alpha-x_i} \partial_{y_\alpha}-\frac{\delta_{\alpha}} {(y_\alpha-x_i)^2}\right] +2\sum_{j \neq i}\left[ \frac{1}{x_j-x_i} \partial_{x_j}-\frac{h_1(\kappa)} {(x_j-x_i)^2}\right]. $$ We can use this to get a simplified formula $$ \ud (Z_t^{\varphi}J^{\varphi}_t)= J^{\varphi}_t \, {\mathcal P}Z_t^{\varphi}\; , \qquad \ud Z_t={\mathcal P} Z_t$$ where ${\mathcal P}$ is the first order differential operator $$ \sum_i \left[ \left(\ud M^{(i)}_t + F^{(i)}_t\right)\partial_{x_i}- 2\ud q^{(i)}_t \left(\sum_{j \neq i} \left[\frac{1}{x_j-x_i}\partial_{x_j} -\frac{h_1(\kappa)}{(x_j-x_i)^2}\right] \right)\right]. $$ The formula for $Z_t$ is just the special case $l=0$. \vspace{.2cm} -- \textit{Final application of Ito's formula}.\\ $$ \ud \left(\frac{Z_t^{\varphi}}{Z_t}J^{\varphi}_t \right)=J^{\varphi}_t {\mathcal Q}\left(\frac{Z_t^{\varphi}}{Z_t}\right) $$ where ${\mathcal Q}$ is the first order differential operator $$\sum_i \left[\ud M^{(i)}_t + F^{(i)}_t- \kappa \ud q^{(i)}_t(\partial_{x_i}\log Z_t) - 2 \sum_{j \neq i} \frac{\ud q^{(j)}_t}{x_i-x_j}\right]\partial_{x_i}$$ The martingale property is satisfied if and only if the drift terms vanish. \subsection{Main claim} To summarize, we have shown that the system $$\ud f_t(z)=\sum_i \frac{2\ud q^{(i)}_t}{f_t(z)-X^{(i)}_t} \quad , \quad \ud X^{(i)}_t=\ud M^{(i)}_t + F^{(i)}_t$$ admits conditioned correlation functions from CFT as martingales if and only if $$F^{(i)}_t=\kappa \ud q^{(i)}_t(\partial_{x_i}\log Z_t) + 2 \sum_{j \neq i} \frac{\ud q^{(j)}_t}{x_i-x_j}.$$ where $Z_t$ is a partition function. It is under this condition that it describes the growth of $n$ interfaces in a way compatible with statistical mechanics and conformal field theory. \vspace{.3cm} In fact, we have used a special family of correlators. But the same argument applies to all operators (hence the ``if'' part). Of special interest in the sequel will be the case when $\Oper$ is a topological observable, for instance taking value $1$ if the interface forms a given arch system and $0$ otherwise. No Jacobian is involved for such observables and the numerator looks again like a partition function. \subsection{The moduli space} From the definition of $Z_t$ as a correlation of primary fields with null descendants at level $2$, it is clear that properties $i),\;ii),\;iii)$ are satisfied, except maybe for the quantization of the possible scaling dimensions of $Z_t$, to which we turn now. This is standard material from CFT and we include it here for completeness. The correlator $\bra \varphi_{h_{\infty}} (\infty)\psi(x_1) \cdots \psi(x_n)\ket$ on the real line satisfies $n$ differential equations. We shall recall why the space of simultaneous solutions which have global conformal invariance has dimension $${n \choose m}-{n \choose m-1}=(n+1-2m)\frac{n!}{m!(n-m+1)!}$$ if $h_{\infty}= h_{n-2m}(\kappa)$ for some nonnegative integer $m \leq n/2$ and has dimension $0$ otherwise. This will end the derivation of our proposal and match the counting of arches. At the end of the background on conformal invariance, we mentioned fusion rules: when $\varphi_{h_{1}(\kappa)}(x)$ and a $\varphi_{h_{j}(\kappa)}(y)$ are brought close together, they can be expanded in a basis of local operators that can be grouped in conformal families. We also recalled why the weight $h'$ of the primaries in each conformal family had to satisfy $3(h_{j}(\kappa)-h')^2-(2h_{1}(\kappa)+1)(h_{j}(\kappa)+h')= h_{1}(\kappa)(h_{1}(\kappa)-1),$ so that only two conformal families can appear in a fusion with $\varphi_{h_{1}(\kappa)}$. The two conformal weights are easily found to be $h'=h_{j\pm 1}(\kappa)$. Furthermore, $h_0(\kappa)=0$ and one can show that the corresponding field has to be the boundary identity operator. By global conformal invariance, the only local operator with a non vanishing one point correlator is the identity and boundary two point functions vanish unless the two local fields have the same conformal weight. This takes care of the counting and selection rules for the $n=0,1$ cases. One proceeds by recursion. The points are ordered $x_1< x_2 \cdots < x_n$. If $n\geq 2$ then move $x_2$ close the $x_1$ (for instance by a global conformal transformation) and fuse to get an expansion for local fields at $x_1$ say. Only the conformal families of $\varphi_{h_{1\pm 1}(\kappa)}$ appear. If $n=2$ this fixes the weight of the field at $\infty$. If $n \geq 3$, iterate. This leads immediately to the selection rules mentioned above : the field at infinity has to be a $\varphi_{h_{n-2m}(\kappa)}$. The dimension is nothing but the number of path of $n$ steps $\pm 1$ from $0$ to $n-2m$ on the nonnegative integers, a standard combinatorial problem whose answer is ${n \choose m}-{n \choose m-1}$. The efficient way to do the counting is by the reflection principle. The possible outcomes of each fusion can be encoded in a so-called Bratelli diagram: \begin{equation} \label{eq: fusion diagram} \left. \begin{array}{ccccccccccc} & & & & & & & & & & \cdots\\ & & & & & & & & & \nearrow & \\ & & & & & & & & h_{4}(\kappa) & & \\ & & & & & & & \nearrow & & \searrow & \\ & & & & & & h_{3}(\kappa) & & & & \cdots \\ & & & & & \nearrow & & \searrow & & \nearrow & \\ & & & & h_{2}(\kappa) & & & & h_{2}(\kappa) & & \\ & & & \nearrow & & \searrow & & \nearrow & & \searrow & \\ & & h_{1}(\kappa) & & & & h_{1}(\kappa) & & & & \cdots \\ & \nearrow & & \searrow & & \nearrow & & \searrow & & \nearrow & \\ h_{0}(\kappa) & & & & h_{0}(\kappa) & & & & h_{0}(\kappa) & & \\ & & \quad & & \quad & & \quad & & \quad & & \\ & & \textrm{$1$SLE}& & \textrm{$2$SLE} & & \textrm{$3$SLE} & & \textrm{$4$SLE} & & \cdots \\ \end{array} \right. \end{equation} This is totally parallel to the discussion of composition of $n$ spins $1/2$ for the representation theory of the Lie algebra of rotations. The multiplicity is exactly one when $m=0$ which corresponds to the partition function (\ref{eq:dabdouwa}) and to the insertion of the operator $\varphi_{h_{n}(\kappa)}$ at infinity, toward which the $n$ interfaces run. What is not proved here is that the different paths lead to a basis of solutions of the $n$ partial differential equations, but it is true. Each path corresponds to a succession of choices of a single conformal family, one at each fusion step. Let us mention in advance that multi SLE processes, i.e. the consideration of multiple interfaces, will lead to the definition of another basis with a topological interpretation. \section{Multiple SLEs describing several interfaces} \label{sec:sevinterf} \subsection{Double SLEs} \label{sec: double SLE} The case of double SLEs is instructive and simple to analyze. Although double SLEs is sometimes interesting for its own sake, the purpose of this section is to give easy examples to guide the study of the general case. \subsubsection{2SLEs and Bessel processes} To specify the process we have to specify the partition function $Z$. There are only two possible choices corresponding to two different type of boundary conditions, or alternatively to two different fields inserted at infinity: \begin{eqnarray*} \bra h_\infty|\psi(X_1)\psi(X_2) | 0 \ket & = & \const \times (X_1 - X_2)^{\Delta} \end{eqnarray*} where the exponent is $\Delta = h_\infty - 2 h_{1}(\kappa)$ and the constant will be fixed to $1$ from now on. According to CFT fusion rules, $h_\infty$ can only be either $h_{2}(\kappa) = \frac{8-\kappa}{\kappa}$ or $h_{0}(\kappa)=0$. The exponent becomes $\Delta = 2/\kappa$ or $\Delta = \frac{\kappa-6}{\kappa}$ respectively, so that we have two basic choices for $Z$: $$ Z_0\equiv (X_1-X_2)^{(\kappa-6)/\kappa}\quad {\rm or}\quad Z_2 \equiv (X_1-X_2)^{2/\kappa} $$ As we shall see, choosing $Z_0$ selects configurations with no curve ending at infinity -- so that we are actually describing standard chordal SLE joining to the two initial positions of $X_1$ and $X_2$ -- while choosing $Z_2$ selects configurations with two curves emerging from the initial positions of $X_1$ and $X_2$ and ending both at infinity. Up to normalizing the quadratic variation by $dq^{(i)}_t= a_i dt$ so that the martingales $M^{(i)}$ are simply $dM^{(i)}_t=\sqrt{\kappa a_i}dB_t^{(i)}$ with $dB_t^{(i)}$ two independent normalized Brownian motions, our double SLE equations become~: \begin{eqnarray*} df_t(z) & = & \frac{2 a_1\;\ud t}{f_t(z) - X^{(1)}_t} + \frac{2 a_2\;\ud t}{f_t(z) - X^{(2)}_t} \\ d X^{(1)}_t & = & \sqrt{a_1 \kappa} \; d B^{(1)}_t + \frac{2 a_2 + \kappa \Delta a_1}{X^{(1)}_t - X^{(2)}_t} \; \ud t \\ d X^{(2)}_t & = & \sqrt{a_2 \kappa} \; d B^{(2)}_t + \frac{2 a_1 + \kappa \Delta a_2}{X^{(2)}_t - X^{(1)}_t} \; \ud t \end{eqnarray*} It describes two curves emerging from points $X_1= X^{(1)}_0$ and $X_2=X^{(2)}_0$ at speeds parametrized by $a_1$ and $a_2$. Up to an irrelevant translation, the process is actually driven by the difference $Y_t=X^{(1)}_t-X^{(2)}_t$. Up to a time change, $ds = \kappa (a_1 + a_2)dt$, this is a Bessel process, \begin{eqnarray*} dY_s = \ud \tilde{B}_s + \frac{\Delta + 2/\kappa}{Y_s} \; \ud s, \end{eqnarray*} of effective dimension $d_{\rm eff}=1 + 2\Delta + 4/\kappa$. For $h_\infty = h_{2}(\kappa)$ (i.e. $\Delta=2/\kappa$) the dimension is $d_{\rm eff}=1+ 8/\kappa$ and for $h_\infty = 0$ (i.e. $\Delta=(\kappa-6)/\kappa$) it is $d_{\rm eff}=3 - 8/\kappa$. In the physically interesting parameter range $\kappa < 8$, the former is $> 2$ and the latter is $< 2$. Recall now that a Bessel process is recurrent (not recurrent) if its effective dimension is less (greater) than $2$. Thus, the driving processes $X^{(i)}_t$ hit each other almost surely in the case $h_{\infty} = 0$ and they don't hit (a.s.) in the case $h_\infty =h_{2}(\kappa)$. Since the hitting of driving processes means hitting of the SLE traces, this teaches us that case $h_\infty=0$ describes a single curve joining $X_1$ and $X_2$ while case $h_\infty=h_{2}(\kappa)$ describes two curves converging toward infinity. Notice that previous results are independent of $a_1$ and $a_2$, provided their sum does not vanishes. We also observe that setting $a_1 = 1$ and $a_2 = 0$ (or vice versa) one recovers an SLE$(\kappa; \kappa \Delta)$. Recall that if $h_\infty = 0$ then $\rho = \kappa \Delta = \kappa - 6$ corresponds to an ordinary chordal SLE from $X_1$ to $X_2$. Our double SLEs with $h_\infty = 0$ corresponds to one chordal SLE seen from both ends and the fact that the tips of the traces hit is natural. The other case, $h_\infty = h_{2}(\kappa)$ corresponds to $\rho = \kappa \Delta = 2$ and since the driving processes can not hit, the process can be defined for all $t \geq 0$. Assuming that $\int_0^\infty ( a_1 + a_2) \ud t = \infty$, the capacity of the hulls grow indefinitely and (at least one of) the SLE traces go to infinity. The two possible geometries are illustrated in figure \ref{fig: 2SLEs}. \begin{figure} \includegraphics[width=1.0\textwidth]{doubleSLEs.eps} \caption{\emph{The two geometries for 2SLE: on the left is the case $h_\infty = 0$ and on the right $h_\infty = h_2 (\kappa)$.}} \label{fig: 2SLEs} \end{figure} \subsubsection{A mixed case for 2SLE} \label{sec: mixed 2SLE} Because of its simplicity, we use double SLE as a testing ground for mixed correlation functions. So we consider the sum $$ Z = \lambda Z_0 + \mu Z_2$$ with both $\lambda$ and $\mu$ positive. As already mentioned, the interpretation of $Z$ as the continuum limit of partition functions of lattice models is unclear since $Z_0$ and $Z_2$ do not scale the same way. We nevertheless study it to illustrate ways of computing (arch or geometry) probabilities. As one may expect, we no longer have an almost sure global geometry but rather nontrivial probabilities for the two geometries: either no curve at infinity or two curves converging there. Let $\tau = \inf \{ t \geq 0 : X_t^{(1)} = X_t^{(2)} \}$ be the stopping time which indicates the hitting of the driving processes -- and thus of the two curves. We can define the driving processes as solutions of the 2SLE system on the (random) time interval $t \in [0, \tau)$. At the stopping time we define $f_\tau(z) = \lim_{s \uparrow \tau} f_s(z)$ for such $z \in \bH$ that the limit exists and stays in the half plane $\bH$. The hull $K_\tau$ is defined as the set where the limit doesn't exist or hits $\bdry \bH$. The question of geometry is answered by the knowledge of whether the two traces hit, that is whether $\tau < \infty$ or not. Thus we again consider the difference $Y_t = X^{(1)}_t - X^{(2)}_t$, whose Ito derivative is now~: \begin{eqnarray*} \ud Y_t & = & \sqrt{\kappa} \ud \tilde B_t + \frac{2}{Y_t}(a_1+a_2)\ud t + \frac{(\kappa - 6) \lambda Y_t^{\frac{\kappa-6}{\kappa}} + 2 \mu Y_t^{\frac{2}{\kappa}}}{Y_t(\lambda Y_t^{\frac{\kappa-6}{\kappa}} + \mu Y_t^{\frac{2}{\kappa}} )} \; (a_1+a_2)\ud t \end{eqnarray*} with $\tilde B_t=\sqrt{a_1}B^{(1)}_t-\sqrt{a_2}B^{(2)}_t$ is a Brownian motion, so that after a time change, $\ud s= (a_1+a_2)\ud t$, the result doesn't depend on $a_1$ or $a_2$. The last drift term comes from the derivative of $\log Z$. One might for example try to find the distribution of $\tau$ by its Laplace transform $\expect_{Y_0 = y} [ e^{-\beta \tau} ] = f_\beta(y)$. By Markov property, \begin{eqnarray*} \expect [ e^{-\beta \tau} | \sF_t] = e^{-\beta t} f_\beta (Y_t) \end{eqnarray*} is a closed martingale on $t \in [0, \tau)$ so requiring its Ito drift to vanish leads to the differential equation \begin{eqnarray*} \Big( -\frac{\beta}{a_1 + a_2} + \big( \frac{2}{y} + \frac{(\kappa-6) \lambda + 2 \mu y^{(8-\kappa)/\kappa}}{\lambda + \mu y^{(8-\kappa)/\kappa}} \big) \partial_y + \frac{\kappa}{2} \partial_y^2 \Big) f_\beta (y) = 0 \end{eqnarray*} The result depends only on $\beta / (a_1 + a_2)$. We conclude that the distribution of $(a_1 + a_2) \tau$, the capacity of the final hull $K_\tau$, is independent of the speeds of growth $a_1$ and $a_2$. Also the result depends on $\lambda$ and $\mu$ only through $\mu / \lambda$. In particular we want to take $\beta \downarrow 0$ to compute the probability that the traces hit. Constant functions solve the differential equation but another linearly independent solution has the correct boundary values $f_0 (0)=1$ and $f_0 (\infty)=0$, namely \begin{eqnarray*} \prob_{Y_0 = y} [\tau < \infty] = \lim_{\beta \downarrow 0} \expect_{Y_0 = y} [e^{-\beta \tau}] = \frac{\lambda}{\lambda+{\mu} y^{(8-\kappa)/\kappa}} \end{eqnarray*} As expected on general ground, this is the fraction of the two partition functions $\lambda Z_0$ and $Z = \lambda Z_0 + \mu Z_2$. \subsection{Triple and/or quadruple SLEs} We will give a few more of examples of multiple SLEs. Certain triple and quadruple SLEs are the scaling limits of interfaces in percolation and Ising model with rather natural boundary conditions. These models will be considered in section \ref{sec: Ising}. Here we study triple and quadruple SLEs for their own sake. We restrict ourselves to $\kappa<8$. \subsubsection{3SLE (pure) configurations} Partition functions with $n=3$ have only two possible scaling behaviors depending whether the weight $h_\infty$ of the field at infinity equals either to $h_{3}(\kappa)=\frac{3(10-\kappa)}{2\kappa}$ or to $h_{1}(\kappa)=\frac{6-\kappa}{2\kappa}$. This follows from CFT fusion rules. For reasons already explained we shall not mixed them. The case $h_\infty=h_{3}(\kappa)$ is the simplest. There is only one possible partition function with this scaling, namely $$ [(X_2-X_1)(X_3-X_1)(X_3-X_2)]^{2/\kappa}$$ It is expected to correspond to configurations with three curves starting at initial positions $X_1,\ X_2$ and $X_3$ and converging toward infinity. The case $h_\infty=h_{1}(\kappa)$ is more interesting since the space of such partition functions is of dimension two and coincides with the space of conformal block with 4 insertions of boundary operators $\psi$, with one localized at $X_4=\infty$: $$ \langle \psi(X_4)\psi(X_3)\psi(X_2)\psi(X_1)\rangle $$ We assume the points to be ordered $X_1<X_2<X_3<X_4$. The associated process should describe a family of two curves joining any pair of adjacent points without crossing. There are thus two possible topologically distinct geometries: either the curves join the pairs $[X_1X_2]$ and $[X_3X_4]$ or they join $[X_4X_1]$ and $[X_2X_3]$, see figure \ref{fig: 3SLEs}. As expected, the number of topologically distinct configuration equals that of conformal blocks, namely two. Notice that the last process is the same as a 4SLE but with the speed $a_4$ vanishing, see figure \ref{fig: 4SLEs}. \begin{figure} \includegraphics[width=1.0\textwidth]{tripleSLE.eps} \caption{\emph{For $h_\infty=h_1(\kappa)$ the curves of 3SLE join either $[X_1 X_2]$ and $[X_3 X_4]$ (on the left) or $[X_4 X_1]$ and $[X_2 X_3]$ (on the right).}} \label{fig: 3SLEs} \end{figure} \begin{figure}[b] \includegraphics[width=1.0\textwidth]{zigzig.eps} \caption{\emph{Arch configurations for four SLE processes in an arbitrary domain.}} \label{fig: 4SLEs} \end{figure} By conformal invariance we may normalize the points so that $X_1=0$, $X_2=x$, $X_3=1$ and $X_4=\infty$ with $0<x<1$. We have two distinct topological configurations and we thus have to identify the two corresponding pure partition functions. This is will be done by specifying the way the partition functions behave when points are fused together. By construction these partition functions may be written as correlation functions $$Z(x)=\langle\psi(\infty)\psi(1)\psi(x)\psi(0)\rangle$$ so that their behavior when points are fused are governed by CFT fusion rules. As a consequence, $Z(x)$ behave either as $x^{\frac{\kappa-6}{\kappa}}$ or as $x^{\frac{2}{\kappa}}$ as $x\to 0$. We select the pure partition functions $Z_I$ and $Z_{II}$ by demanding that: \begin{eqnarray} Z_I(x) &=& x^{\frac{\kappa-6}{\kappa}}\times [1+\cdots],\quad ~~~~~~~~~~~~ {\rm as}\ x\to0 \label{bdrypure}\\ &=& (1-x)^{\frac{2}{\kappa}}\times [{\rm const.}+\cdots],\quad ~ {\rm as}\ x\to 1 \nonumber \end{eqnarray} and $Z_{II}(x)=Z_I(1-x)$ so that \begin{eqnarray*} Z_{II}(x) &=& x^{\frac{2}{\kappa}}\times [{\rm const.}+\cdots],\quad ~~~~~~~~~~ {\rm as}\ x\to 0 \\ &=& (1-x)^{\frac{\kappa-6}{\kappa}}\times [1+\cdots],\quad ~~~~~~ {\rm as}\ x\to 1 \end{eqnarray*} $Z_I$ will turn out to be the pure partition function for configurations in which the curves join the pairs $[0x]$ and $[1\infty]$ while $Z_{II}$ will turn out to correspond to the configurations $[x1]$ and $[\infty 0]$. The rationale behind these conditions consists in imposing that the pure partition function possesses the leading singularity, with exponent $(6-\kappa)/\kappa$, when $x$ is approaching the point allowed by the configuration but has subleading singularity, with exponent $2/\kappa$, when $x$ is approaching the point forbidden by the configuration. This set of conditions uniquely determines the functions $Z_I$ and $Z_{II}$. These follows from CFT rules but may also be checked by explicitly solving the differential equation that these functions satisfy. Writing $Z(x)=x^{2/\kappa}(1-x)^{2/\kappa}\; G(x)$ yields, $$ \kappa^2 x(1-x) G''(x) + 8\kappa (1-2x) G'(x) - 4(12-\kappa) G(x)=0 $$ so that $G(x)$ is an hypergeometric function and $$ Z_{II}(x) = {\rm const.} x^{2/\kappa}(1-x)^{2/\kappa}\; F(\frac{4}{\kappa}, \frac{12-\kappa}{\kappa};\frac{8}{\kappa}|x) $$ with the constant chosen to normalize $Z_I$ as above. Using this explicit formula one may verify that $Z_I(x)$ is effectively a positive number for any $x\in[0;1]$ so it has all expected properties to be a pure partition function. For $\kappa=4$, $Z_I(x)=\sqrt{(1-x)/x}$ and for $\kappa=2$, $Z_I(x)=(1-x^2)/x^2$. \subsubsection{Arch probabilities} Let us now compute the probabilities for having one of the two topologically distinct configurations: either $(I)$ with curves joining either $[0x]$ and $[1\infty]$ or $(II)$ with curves joining $[x1]$ and $[\infty0]$ as we just discussed. We shall proceed blindly, but the reader should beware that there are subtleties involved. What is computed is the probability for certain $X^{(i)}_t$'s to hit each other. What happens at the level of hulls and how the process should be properly continued is not investigated, but is expected to yield the announced probability for arch configuration. We consider a generic partition function $Z$ which is the sum of the pure partition functions $Z_I$ and $Z_{II}$: $$ Z(x)= p_I Z_I(x) + p_{II}Z_{II}(x) $$ with $p_I$ and $p_{II}$ positive. To specify the 3SLE (or 4SLE) process we need the partition function $Z(X_1,X_2,X_3,X_4)$ which is recover from $Z(x)$ by conformal transformation~: $$ Z(X_1,X_2,X_3,X_4)= [(X_4-X_2)(X_3-X_1)]^\frac{\kappa-6}{\kappa}\;Z(X) $$ with $X$ the harmonic ratio of the four points $X_1$, $X_2$, $X_3$ and $X_4$~: $$X=\Big(\frac{ X_1-X_2 }{ X_1-X_3 }\Big)\Big( \frac{ X_4-X_3 }{ X_4-X_2 }\Big).$$ Let $M_I(x)$ and $M_{II}(x)=1-M_I(x)$ be defined by $$ M_I(x) \equiv p_I Z_I(x)/Z(x)\quad,\quad M_{II}(x)\equiv p_{II} Z_{II}(x) / Z(x) $$ By construction the processes $t\to M_I(X_t)$ and $t\to M_{II}(X_t)$, with $X_t$ the harmonic ratio of the four moving points, are local martingales. Since both $Z_I$ and $Z_{II}$ are positive, $M_I$ are $M_{II}$ are bounded local martingales and thus are martingales. Let $\tau$ be the stopping time given by the first instant at which a pair of points $X_t^{(i)}$ coincide. Then, in configuration $(I)$ we have $\lim_{t\nearrow \tau} X_t=0$ while $\lim_{t\nearrow \tau} X_t =1$ in configuration $(II)$. Since, for $\kappa<8$, $M_I(x)$ is such that $\lim_{x\to 0} M_I(x)= 1$ but $\lim_{x\to 1}M_I(x)=0$, we obtain that $M_I$ evaluated at the stopping time $\tau$ is the characteristic function for events with the topological configuration $(I)$, ie: \begin{eqnarray*} \lim_{t\nearrow \tau} M_I(X_t) &=& {\bf 1}_{{\rm config.} (I)} \\ \lim_{t\nearrow \tau} M_{II}(X_t) &=& {\bf 1}_{{\rm config.} (II)} \end{eqnarray*} Since $M_I$ and $M_{II}$ are martingales, we get the probability of occurrence of configurations of topological type $(I)$: \begin{eqnarray} \prob [{\rm config.} (I)]=M_I(X_{t=0})= \frac{p_I Z_I(x)}{p_IZ_I(x)+p_{II}Z_{II}(x)} \label{crossprob} \end{eqnarray} and similarly for the probabilities of having configuration $(II)$. As expected they are ratios of partition functions. \subsection{Applications to percolation and Ising model} \label{sec: Ising} We are now ready to give an application of triple (or quadruple) SLE to percolation and Ising model. Exploration processes in critical percolation are described by SLEs with $\kappa=6$, as proved in \cite{SS: percolation}. Interfaces of spin clusters in critical Ising model is believed to correspond to $\kappa = {3}$ while interfaces of Fortuin-Kasteleyn clusters -- which occur in a high temperature expansion of the Ising partition function -- are expected to correspond to the dual value $\kappa=16/3$. What we have in mind are these statistical models, defined on the upper half plane, with boundary condition changing operators at the four points $0,\ x,\ 1$ and $\infty$. They change the boundary condition from open to closed (or vice versa) in percolation $(\kappa=6)$ and from plus to minus (or vice versa) for Ising model $(\kappa=3)$. To apply previous results on 4SLE processes to these situations, we have to specify the partition functions $Z(x)$, or equivalently, we have to specify the value of $p_I$ and $p_{II}$. This is done by noticing that these models are left-right symmetric so that for $x=1/2$ there is equal probability to find configuration $(I)$ or $(II)$. Since $Z_I(1/2)=Z_{II}(1/2)$, we have $p_I=p_{II}=1$, so that the total partition function is $Z(x)=Z_I(x)+Z_{II}(x)$ and the probability of occurrence of configuration $(I)$ for any $0<x<1$ is now: $$ \prob [{\rm config.} (I)]= \frac{ Z_I(x)}{Z_I(x)+Z_{II}(x)}\quad,\quad Z_{II}(x)=Z_I(1-x) $$ \vspace{.2cm} -- Percolation corresponds to $\kappa=6$. The boundary changing operator $\psi$ has dimension $0$. The pure partition function $Z_I$ has a simple integral representation: $$ Z_I(x)_{\rm perco}= \frac{\Gamma(2/3)}{\Gamma(1/3)^2}\ \int_x^1 \ud s\ s^{-2/3}(1-s)^{-2/3}. $$ By construction $Z_{II}(x)=Z_I(1-x)$ also possesses a simple integral representation but, most importantly, it is such that the total partition function is constant, $Z(x)=Z_I(x)+Z_{II}(x)=1$, as expected for percolation. As a consequence we find: $$ \prob [{\rm config.} (I)]_{\rm perco}= \frac{\Gamma(2/3)}{\Gamma(1/3)^2}\ \int_x^1 \ud s\ s^{-2/3}(1-s)^{-2/3} $$ This is nothing but Cardy percolation crossing formula. \vspace{.2cm} -- Ising spin clusters correspond to $\kappa=3$. The boundary changing operator $\psi$ has dimension $1/2$ and may thus be identified with a fermion on the boundary. However the pure partition functions do not correspond to the free fermion conformal block. By solving the differential equation with the appropriate boundary condition we get: $$ Z_I(x)_{\rm spin\ Ising}= \mathrm{const} \frac{1-x+x^2}{x(1-x)}\int_x^1\ud y\frac{(y(1-y))^{2/3}}{(1-y+y^2)^2} $$ The total partition function $Z_I(x)+Z_I(1-x)$ is proportional to $\frac{1-x+x^2}{x(1-x)}$, which is the free fermion result. Hence, the Ising configuration probabilities are~: $$ \prob [{\rm config.} (I)]_{\rm spin\ Ising}= \int_x^1\ud y\frac{(y(1-y))^{2/3}}{(1-y+y^2)^2}\; \Big/ \int_0^1\ud y\frac{(y(1-y))^{2/3}}{(1-y+y^2)^2} $$ This is nothing but a new -- and previously unknown -- Ising crossing formula. \vspace{.2cm} -- FK Ising clusters correspond to $\kappa=16/3$. The operator $\psi$ has then dimension $1/16$. The pure partition function are given by: $$ Z_I(x)_{\rm FK\ Ising} = \frac{(1-x)^{3/8}}{x^{1/8}(1+\sqrt{x})^{1/2}} $$ and the crossing probabilities by: \begin{eqnarray*} \prob [{\rm config.} (I)]_{\rm FK\ Ising} = \frac{\sqrt{(1-x) + (1-x)^{3/2}}}{\sqrt{x + x^{3/2}} + \sqrt{(1-x) + (1-x)^{3/2}}} \end{eqnarray*} The other critical random cluster (or Potts) models with $0 \leq Q \leq 4$ have $Q = 4 \cos^2 \big( \frac{4 \pi}{\kappa} \big)$, $4 \leq \kappa \leq 8$ and it is straightforward to obtain explicit crossing formulas involving only hypergeometric functions. \subsection{nSLEs and beyond} We now comment on how to compute multiple arch probabilities for general nSLEs. This section only aims at giving some hints on how to generalize previous computations. So it shall be sketchy. It is clear that the key point is to identify the pure partition functions -- once this is done the rest is routine. As exemplified above by eq.(\ref{bdrypure}) this is linked to CFT fusions. The rules there were that, for a given arch system, fusing two points linked by an arch produces the dominant singularity which means that the two boundary operators are fused on the identity operator, whereas fusing two points not linked by an arch produces the subleading singularity which means the fusion of the two boundary fields on the identity should vanish. In general there could be a whole hierarchy of arches with arches in the interior of others, i.e. with a family of self-surrounding arches, the next encircling the previous. So we are lead to propose the following rules.\\ For a given arch configuration:\\ --- The most interior pair of adjacent pair of points, say $X_i,X_{i+1}$ in a family of self-surrounding arches fused into the identity operator, so that the pure partition function evaluated at $X_i\simeq X_{i+1}$ should be proportional to $(X_{i+1}-X_i)^\frac{\kappa-6}{\kappa}$ times the pure partition function associated to the arch system with the interior arch $[X_iX_{i+1}]$ removed. Symbolically~: $$ Z_{\rm pure}(\cdots, X_i\simeq X_{i+1},\cdots)\simeq {\rm const.}\ (X_{i+1}-X_i)^\frac{\kappa-6}{\kappa}\times Z_{{\rm pure}\setminus[X_iX_{i+1}]}(\cdots,\cdots) $$ for $X_i$ and $X_{i+1}$ linked by an arch.\\ \noindent --- The fusion on the identity of any pair of adjacent points not linked by an arch should vanish, so that the fusion of this pair of points produces the subleading singularity. Symbolically~: $$ Z_{\rm pure}(\cdots, X_i\simeq X_{i+1},\cdots)\simeq {\rm const.}\ (X_{i+1}-X_i)^\frac{2}{\kappa} +\cdots $$ for $X_i$ and $X_{i+1}$ for not linked by an arch. We do not have a complete proof that these rules fully determine the pure partition functions but we checked it on a few cases, see figure \ref{fig:rules}. \begin{figure} \includegraphics[width=1.0\textwidth]{quadrupleSLE.eps} \caption{\emph{Illustration of the fusion rules corresponding to arch configurations.}} \label{fig:rules} \end{figure} Here are a few samples. We shall give the relation between the pure partition and the CFT conformal blocks indexed by the corresponding Bratelli diagram. For $n=4$, we may have the following arch systems $[X_1X_2][X_3X_4]$ or $[X_1[X_2X_3]X_4]$. (A given geometrical configuration may correspond to different arch systems depending at which location we open the closed boundary. But they are all equivalent to these two up to an order preserving relabeling of the points. For instance $[X_4X_1][X_2X_3]$ is equivalent to $[X_1[X_2X_3]X_4]$.) Applying the previous rules we get: \begin{eqnarray*} Z_{[X_1[X_2X_3]X_4]} &=& \langle {}_{[h_0{}]}\psi(X_1) {}_{[h_1{}]}\psi(X_2) {}_{[h_2{}]}\psi(X_3) {}_{[h_1{}]}\psi(X_4) {}_{[h_0{}]}\rangle \\ Z_{[X_1X_2][X_3X_4]}&=& \langle {}_{[h_0{}]}\psi(X_1) {}_{[h_1{}]}\psi(X_2) {}_{[h_0{}]}\psi(X_3) {}_{[h_1{}]}\psi(X_4) {}_{[h_0{}]}\rangle \\ & + & \omega \; \langle {}_{[h_0{}]}\psi(X_1) {}_{[h_1{}]}\psi(X_2) {}_{[h_2{}]}\psi(X_3) {}_{[h_1{}]}\psi(X_4) {}_{[h_0{}]}\rangle, \end{eqnarray*} where the indices $h_{m}$, $m=0,1,\cdots$ refer to the corresponding points in the Bratelli diagram, i.e. to the weights $h_{m}(\kappa)$ of the intermediate Virasoro modules. The coefficient $\omega$ is fully determined, in terms of CFT fusion coefficients, by demanding that the fusion of $X_2$ and $X_3$ on the identity vanishes. One may go on and solve for the pure partition functions in few other cases. A particularly simple example with $n=6$ is given by~: \begin{eqnarray*} Z_{[X_1[X_2[X_3X_4]X_5]X_6]} & = & \\ & & \hspace{-5cm} \langle {}_{[h_0{}]}\psi(X_1) {}_{[h_1{}]}\psi(X_2) {}_{[h_2{}]}\psi(X_3) {}_{[h_3{}]}\psi(X_4) {}_{[h_2{}]}\psi(X_5) {}_{[h_1{}]}\psi(X_6) {}_{[h_0{}]}\rangle \end{eqnarray*} As can be seen on these examples, there is no simple relation between arch systems and Bratelli diagrams and the change of basis for one to the other is quite involved. The only simple rule we find is that the pure partition function for a unique family of self-surrounding arches is a pure conformal block corresponding to a unique Bratelli diagram.
{ "timestamp": "2005-03-21T17:02:45", "yymm": "0503", "arxiv_id": "math-ph/0503024", "language": "en", "url": "https://arxiv.org/abs/math-ph/0503024" }
\section{Introduction} In a series of papers, Mweene has developed a generalized description of angular momentum which contains the standard results in a certain limit. He has thereby obtained new generalized expressions for the operators and eigenvectors for spin 1/2[1-4], spin 1[5], spin 3/2[6], spin 2[7] and spin 5/2[8]. Applying this approach to angular momentum addition, he has shown how the standard results for various states that arise from combining two values of angular momentum come about from a consideration of probability amplitudes for measurements on these systems. This had led to generalized results for angular momentum addition which also reduce to the standard results in an appropriate limit[9-10]. In a further development of the approach, Mweene has shown that the usual spherical harmonics are just special forms of more generalized quantities and he has obtained the generalized spherical harmonics for the case $l=1$% [11]. In this paper, we give the generalized spherical harmonics for $l=2$. This paper is organized as follows. This Introduction is followed in Section 2 by a brief review of the theory underlying the work. Section 3 contains the derivation of the generalized spherical harmonics and their probability amplitudes. Section 4 is a discussion of some of the properties of these quantities. The Discussion and Conclusion in Section 5 closes the paper. \section{Theoretical Background} This work is inspired by the interpretation of quantum mechanics due to Land\'e[12-15]. According to Land\'e, if a quantum system possesses three sets of observables $A$, $B$ and $C$ with respective eigenvalue spectra $A_1$% , $A_2,$...,$A_N$, $B_1$, $B_2$, ...,$B_N$ and $C_1$, $C_2$, ....,$C_N$ - where $N$ is the multiplicity of each spectrum, which is necessarily the same for each observable[13] - then three sets of probability amplitudes can be defined for the system. One set, which is denoted by $\psi (A_i,C_j),$ relates to measurement of the observable $C$ if the system is in a state corresponding to eigenvalues of the observable $A$: thus $\left| \psi (A_i;C_j)\right| ^2$ gives the probability for obtaining the value $C_j$ if the initial state corresponds to the eigenvalue $A_i$ of the observable $A$. The second set $\chi (A_i;B_j)$ relates to measurement of $B$ when the system is in a state corresponding to the eigenvalue $A$. Finally, the set $% \phi (B_i;C_j)$ describes measurements of $C$ when the initial state belongs to the observable $B$. Since the three sets of probability amplitudes belong to one system, they are interdependent, and the law of interdependence is[12,13] \begin{equation} \psi (A_i;C_j)=\sum_j\chi (A_i;B_j)\phi (B_j;C_n) \label{on1} \end{equation} Another aspect of Land\'e's interpretation of quantum mechanics is that every eigenfunction or wave function of a quantum system is first and foremost a probability amplitude and that every such probability amplitude connects two well-defined states - one corresponding to the state in which the system is before a measurement, and the other to the state that comes about as a result of the measurement[12,13]. It is therefore always possible to identify an initial and a final state for any wave function or eigenfunction. Mweene has argued that for an eigenfunction resulting from solution of a differential eigenvalue equation, the initial state corresponds to the eigenvalue while the final state corresponds to the eigenvalue defined by the continuous variable in terms of which the differential operator is defined[16]. For the Schr\"odinger equation, the eigenfunctions $\psi _{E_i}(x)$ should really be written as $\psi (E_i;x)$ to emphasize that the initial state in the probability amplitude corresponds to the eigenvalue $% E_i $ while the final state corresponds to the eigenvalue $x$. Another example comes from the solution of Legendre's equation. The spherical harmonics $Y_{lm}(\theta ,\varphi )$ should really be written as $% Y(l,m;\theta ,\varphi )$ to emphasize that in this case the initial state is defined by the eigenvalues $m\hbar $ and $l(l+1)\hbar ^2$ while the final state corresponds to the angular position $(\theta ,\varphi ).$ Since $% m\hbar $ is an angular momentum projection, it must be defined with respect to some axis. Owing to the absence of another set of angles in the expressions of the spherical harmonics which could define this direction, it must be the $z$ axis[11]. But since an axis of quantization can be chosen arbitrarily, it is possible to define spherical harmonics with respect to any other direction as the axis of initial quantization. The functions resulting from this are the generalized spherical harmonics and have already been worked out for the case $l=1$. In this work, we obtain them for $l=2$. \section{Generalized Spherical Harmonics} \subsection{Probability Amplitudes} The generalized spherical harmonics connect states of angular momentum projection in the arbitrary direction $\widehat{\mathbf{a}}$ defined by the polar angles $(\theta ^{\prime },\varphi ^{\prime })$ to states of the angular position $(\theta ,\varphi ).$ We denote them by $Y(l,m^{(\widehat{% \mathbf{a}})};\theta ,\varphi )$ . To derive them, we use the probability addition law Eq. (\ref{on1}). We start off by writing \begin{equation} Y(l,m^{(\widehat{\mathbf{a}})};\theta ,\varphi )=\sum_j\chi (l,m^{(\widehat{% \mathbf{a}})};B_j)\phi (B_j;\theta ,\varphi ) \label{th34} \end{equation} If we choose the observable $B$ carefully, we should find that both the probability amplitudes $\chi (l,m^{(\widehat{\mathbf{a}})};B_j)$ and $\phi (B_j;\theta ,\varphi )$ are known. If $B$ is chosen to be the spin projection with respect to the $z$ direction, it is found that \begin{equation} \phi (B_j;\theta ,\varphi )=Y(l,m^{(\widehat{\mathbf{k}})};\theta ,\varphi ) \label{th34a} \end{equation} are the standard spherical harmonics, while $\chi (l,m_i^{(\widehat{\mathbf{a% }})};l,m_f^{(\widehat{\mathbf{k}})})$ are just spin probability amplitudes connecting states such that the initial one corresponds to the spin projection being $m_i\hbar $ in the direction $\widehat{\mathbf{a}}$ while the final state corresponds to the spin projection being $m_f\hbar $ along the $z $ axis. These have already been worked out[7] and are as given below. If the initial spin projection in the direction $\widehat{\mathbf{a}}$ is $% 2\hbar $, these probability amplitudes are \begin{equation} \chi (2,2^{(\widehat{\mathbf{a}})};2,2^{(\widehat{\mathbf{k}})})=\cos ^4% \frac{\theta ^{\prime }}2e^{-2i\varphi ^{\prime }} \label{fi53} \end{equation} \begin{equation} \chi (2,2^{(\widehat{\mathbf{a}})};2,1^{(\widehat{\mathbf{k}})})=2\sin \frac{% \theta ^{\prime }}2\cos ^3\frac{\theta ^{\prime }}2e^{-i\varphi ^{\prime }} \label{fi54} \end{equation} \begin{equation} \chi (2,2^{(\widehat{\mathbf{a}})};2,0^{(\widehat{\mathbf{k}})})=\sqrt{6}% \sin ^2\frac{\theta ^{\prime }}2\cos ^2\frac{\theta ^{\prime }}2 \label{fi55} \end{equation} \begin{equation} \chi (2,2^{(\widehat{\mathbf{a}})};2,(-1)^{(\widehat{\mathbf{k}})})=2\sin ^3% \frac{\theta ^{\prime }}2\cos \frac{\theta ^{\prime }}2e^{i\varphi ^{\prime }} \label{fi56} \end{equation} and \begin{equation} \chi (2,2^{(\widehat{\mathbf{a}})};2,(-2)^{(\widehat{\mathbf{k}})})=\sin ^4% \frac{\theta ^{\prime }}2e^{2i\varphi ^{\prime }} \label{fi57} \end{equation} If the initial spin projection in the direction $\widehat{\mathbf{a}}$ is $% \hbar ,$ the probability amplitudes are \begin{equation} \chi (2,1^{(\widehat{\mathbf{a}})};2,2^{(\widehat{\mathbf{k}})})=2\sin \frac{% \theta ^{\prime }}2\cos ^3\frac{\theta ^{\prime }}2e^{-2i\varphi ^{\prime }} \label{fi58} \end{equation} \begin{equation} \chi (2,1^{(\widehat{\mathbf{a}})};2,1^{(\widehat{\mathbf{k}})})=-(3\sin ^2% \frac{\theta ^{\prime }}2-\cos ^2\frac{\theta ^{\prime }}2)\cos ^2\frac{% \theta ^{\prime }}2e^{-i\varphi ^{\prime }} \label{fi59} \end{equation} \begin{equation} \chi (2,1^{(\widehat{\mathbf{a}})};2,0^{(\widehat{\mathbf{k}})})=-\sqrt{6}% \cos \frac{\theta ^{\prime }}2\sin \frac{\theta ^{\prime }}2\cos \theta ^{\prime } \label{si60} \end{equation} \begin{equation} \chi (2,1^{(\widehat{\mathbf{a}})};2,(-1)^{(\widehat{\mathbf{k}})})=(3\cos ^2% \frac{\theta ^{\prime }}2-\sin ^2\frac{\theta ^{\prime }}2)\sin ^2\frac{% \theta ^{\prime }}2e^{i\varphi ^{\prime }} \label{si61} \end{equation} and \begin{equation} \chi (2,1^{(\widehat{\mathbf{a}})};2,(-2)^{(\widehat{\mathbf{k}})})=-2\sin ^3% \frac{\theta ^{\prime }}2\cos \frac{\theta ^{\prime }}2e^{2i\varphi ^{\prime }} \label{si62} \end{equation} If the initial spin projection in the direction $\widehat{\mathbf{a}}$ is $% 0, $the probability amplitudes are \begin{equation} \chi (2,0^{(\widehat{\mathbf{a}})};2,2^{(\widehat{\mathbf{k}})})=\sqrt{6}% \sin ^2\frac{\theta ^{\prime }}2\cos ^2\frac{\theta ^{\prime }}% 2e^{-2i\varphi ^{\prime }} \label{si63} \end{equation} \begin{equation} \chi (2,0^{(\widehat{\mathbf{a}})};2,1^{(\widehat{\mathbf{k}})})=-\sqrt{6}% \sin \frac{\theta ^{\prime }}2\cos \frac{\theta ^{\prime }}2\cos \theta ^{\prime }e^{-i\varphi ^{\prime }} \label{si64} \end{equation} \begin{equation} \chi (2,0^{(\widehat{\mathbf{a}})};2,0^{(\widehat{\mathbf{k}})})=\frac 12(2\cos ^2\theta ^{\prime }-\sin ^2\theta ^{\prime }) \label{si65} \end{equation} \begin{equation} \chi (2,0^{(\widehat{\mathbf{a}})};2,(-1)^{(\widehat{\mathbf{k}})})=\sqrt{6}% \sin \frac{\theta ^{\prime }}2\cos \frac{\theta ^{\prime }}2\cos \theta ^{\prime }e^{i\varphi ^{\prime }} \label{si66} \end{equation} and \begin{equation} \chi (2,0^{(\widehat{\mathbf{a}})};2,(-2)^{(\widehat{\mathbf{k}})})=\sqrt{6}% \sin ^2\frac{\theta ^{\prime }}2\cos ^2\frac{\theta ^{\prime }}2e^{2i\varphi ^{\prime }} \label{si67} \end{equation} If the initial spin projection in the direction $\widehat{\mathbf{a}}$ is $% -\hbar ,$ the probability amplitudes are \begin{equation} \chi (2,(-1)^{(\widehat{\mathbf{a}})};2,2^{(\widehat{\mathbf{k}})})=2\cos \frac{\theta ^{\prime }}2\sin ^3\frac{\theta ^{\prime }}2e^{-2i\varphi ^{\prime }} \label{si68} \end{equation} \begin{equation} \chi (2,(-1)^{(\widehat{\mathbf{a}})};2,1^{(\widehat{\mathbf{k}})})=-(3\cos ^2\frac{\theta ^{\prime }}2-\sin ^2\frac{\theta ^{\prime }}2)\sin ^2\frac{% \theta ^{\prime }}2e^{-i\varphi ^{\prime }} \label{si69} \end{equation} \begin{equation} \chi (2,(-1)^{(\widehat{\mathbf{a}})};2,0^{(\widehat{\mathbf{k}})})=\sqrt{6}% \sin \frac{\theta ^{\prime }}2\cos \frac{\theta ^{\prime }}2\cos \theta ^{\prime } \label{se70} \end{equation} \begin{equation} \chi (2,(-1)^{(\widehat{\mathbf{a}})};2,(-1)^{(\widehat{\mathbf{k}}% )})=(3\sin ^2\frac{\theta ^{\prime }}2-\cos ^2\frac{\theta ^{\prime }}2)\cos ^2\frac{\theta ^{\prime }}2e^{i\varphi ^{\prime }} \label{se71} \end{equation} and \begin{equation} \chi (2,(-1)^{(\widehat{\mathbf{a}})};2,(-2)^{(\widehat{\mathbf{k}}% )})=-2\sin \frac{\theta ^{\prime }}2\cos ^3\frac{\theta ^{\prime }}% 2e^{2i\varphi ^{\prime }} \label{se72} \end{equation} Finally, if the initial spin projection in the direction $\widehat{\mathbf{a}% }$ is $-2\hbar ,$ the probability amplitudes are \begin{equation} \chi (2,(-2)^{(\widehat{\mathbf{a}})};2,2^{(\widehat{\mathbf{k}})})=\sin ^4% \frac{\theta ^{\prime }}2e^{-2i\varphi ^{\prime }} \label{se73} \end{equation} \begin{equation} \chi (2,(-2)^{(\widehat{\mathbf{a}})};2,1^{(\widehat{\mathbf{k}})})=-2\cos \frac{\theta ^{\prime }}2\sin ^3\frac{\theta ^{\prime }}2e^{-i\varphi ^{\prime }} \label{se74} \end{equation} \begin{equation} \chi (2,(-2)^{(\widehat{\mathbf{a}})};2,0^{(\widehat{\mathbf{k}})})=\sqrt{6}% \sin ^2\frac{\theta ^{\prime }}2\cos ^2\frac{\theta ^{\prime }}2 \label{se75} \end{equation} \begin{equation} \chi (2,(-2)^{(\widehat{\mathbf{a}})};2,(-1)^{(\widehat{\mathbf{k}}% )})=-2\sin \frac{\theta ^{\prime }}2\cos ^3\frac{\theta ^{\prime }}% 2e^{i\varphi ^{\prime }} \label{se76} \end{equation} and \begin{equation} \chi (2,(-2)^{(\widehat{\mathbf{a}})};2,(-2)^{(\widehat{\mathbf{k}})})=\cos ^4\frac{\theta ^{\prime }}2e^{2i\varphi ^{\prime }} \label{se77} \end{equation} The ordinary spherical harmonics $Y_{2m}(\theta ,\varphi )=Y(2,m^{(\widehat{% \mathbf{k}})};\theta ,\varphi )$ for $l=2$ are \begin{equation} Y(2,2^{(\widehat{\mathbf{k}})};\theta ,\varphi )=\sqrt{\frac{15}{32\pi }}% \sin ^2\theta e^{2i\varphi } \label{se78} \end{equation} \begin{equation} Y(2,1^{(\widehat{\mathbf{k}})};\theta ,\varphi )=-\sqrt{\frac{15}{8\pi }}% \sin \theta \cos \theta e^{i\varphi } \label{se79} \end{equation} \begin{equation} Y(2,0^{(\widehat{\mathbf{k}})};\theta ,\varphi )=\sqrt{\frac 5{16\pi }}% (3\cos ^2\theta -1) \label{ei80} \end{equation} \begin{equation} Y(2,(-1)^{(\widehat{\mathbf{k}})};\theta ,\varphi )=\sqrt{\frac{15}{8\pi }}% \sin \theta \cos \theta e^{-i\varphi } \label{ei81} \end{equation} \begin{equation} Y(2,(-2)^{(\widehat{\mathbf{k}})};\theta ,\varphi )=\sqrt{\frac{15}{32\pi }}% \sin ^2\theta e^{-2i\varphi } \label{ei82} \end{equation} Using Eq. (\ref{th34}), the generalized spherical harmonics for $l=2$ are found to be \begin{eqnarray} Y(2,2^{(\mathbf{\hat a})};\theta ,\varphi ) &=&\sqrt{\frac{15}{32\pi }}% \{\sin ^2\theta (\cos ^4\frac{\theta ^{\prime }}2e^{2i(\varphi -\varphi ^{\prime })}+\sin ^4\frac{\theta ^{\prime }}2e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \label{eq52} \\ &&+\sin 2\theta \sin \theta ^{\prime }(-\cos ^2\frac{\theta ^{\prime }}% 2e^{i(\varphi -\varphi ^{\prime })}+\sin ^2\frac{\theta ^{\prime }}% 2e^{-i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&+\frac 12\sin ^2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ei83} \end{eqnarray} \begin{eqnarray} Y(2,1^{(\mathbf{\hat a})};\theta ,\varphi ) &=&\sqrt{\frac{15}{32\pi }}% \{\sin \theta ^{\prime }\sin ^2\theta (\cos ^2\frac{\theta ^{\prime }}% 2e^{2i(\varphi -\varphi ^{\prime })}-\sin ^2\frac{\theta ^{\prime }}% 2e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ -\sin 2\theta [(3\sin ^2\frac{\theta ^{\prime }}2-\cos ^2\frac{\theta ^{\prime }}2)\cos ^2\frac{\theta ^{\prime }}2e^{i(\varphi -\varphi ^{\prime })} \nonumber \\ &&\ +(3\cos ^2\frac{\theta ^{\prime }}2-\sin ^2\frac{\theta ^{\prime }}% 2)\sin ^2\frac{\theta ^{\prime }}2e^{-i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ -\frac 12\sin 2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ei84} \end{eqnarray} \begin{eqnarray} Y(2,0^{(\mathbf{\hat a})};\theta ,\varphi ) &=&\sqrt{\frac{45}{256\pi }}% \{\sin ^2\theta ^{\prime }\sin ^2\theta (e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \label{eq54} \\ &&+\sin 2\theta \sin 2\theta ^{\prime }(e^{i(\varphi -\varphi ^{\prime })}+e^{-i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&+\frac 23(3\cos ^2\theta -1)(2\cos ^2\theta ^{\prime }-\sin ^2\theta ^{\prime })\} \label{ei85} \end{eqnarray} \begin{eqnarray} Y(2,(-1)^{(\mathbf{\hat a})};\theta ,\varphi ) &=&\sqrt{\frac{15}{32\pi }}% \{\sin \theta ^{\prime }\sin ^2\theta (\sin ^2\frac{\theta ^{\prime }}% 2e^{2i(\varphi -\varphi ^{\prime })}-\cos ^2\frac{\theta ^{\prime }}% 2e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \label{eq55} \\ &&\ +\sin 2\theta [(3\cos ^2\frac{\theta ^{\prime }}2-\sin ^2\frac{\theta ^{\prime }}2)\sin ^2\frac{\theta ^{\prime }}2e^{i(\varphi -\varphi ^{\prime })} \nonumber \\ &&\ +(3\sin ^2\frac{\theta ^{\prime }}2-\cos ^2\frac{\theta ^{\prime }}% 2)\cos ^2\frac{\theta ^{\prime }}2e^{-i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ +\frac 12\sin 2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ei86} \end{eqnarray} \begin{eqnarray} Y(2,(-2)^{(\mathbf{\hat a})};\theta ,\varphi ) &=&\sqrt{\frac{15}{32\pi }}% \{\sin ^2\theta (\sin ^4\frac{\theta ^{\prime }}2e^{2i(\varphi -\varphi ^{\prime })}+\cos ^4\frac{\theta ^{\prime }}2e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \label{eq56} \\ &&\ +\sin 2\theta \sin \theta ^{\prime }(\sin ^2\frac{\theta ^{\prime }}% 2e^{i(\varphi -\varphi ^{\prime })}-\cos ^2\frac{\theta ^{\prime }}% 2e^{-i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ +\frac 12\sin ^2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ei87} \end{eqnarray} \subsection{Probability Amplitudes for the $x^{\prime }$ Direction} The results we have presented refer to the direction $\widehat{\mathbf{a}}$ as the direction of initial quantization. We may think of the vector $% \widehat{\mathbf{a}}$ as defining a new $z$ axis, which we denote by $% z^{\prime }$, since in the limit $\theta ^{\prime }=\varphi ^{\prime }=0,$ the results corresponding to it reduce to those for the $z$ axis. This $% z^{\prime }$ axis corresponds to a new coordinate system in which the unit vector in the $x^{\prime }$ direction is $\widehat{\mathbf{u}}$ and that in the $y^{\prime }$ direction is $\widehat{\mathbf{v}}$[11]. From the results for the $\widehat{\mathbf{a}}$ or $z^{\prime }$ axis, we can obtain the probability amplitudes and probabilities densities for the $x^{\prime }$ direction by applying the transformation $\theta ^{\prime }\rightarrow \theta ^{\prime }-\pi /2$ to them[3,5]. We are justified in associating the results so obtained with the $x^{\prime }$ axis since in the limit $\theta ^{\prime }=\varphi ^{\prime }=0, $ they reduce to those for the $x$ direction. When we make these argument changes, $\widehat{\mathbf{a}}$ becomes $\widehat{\mathbf{u}}$. Applying this prescription to the generalized spherical harmonics, we obtain the results \begin{eqnarray} Y(2,2^{(\widehat{\mathbf{u}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{128\pi }}% \{\frac 12\sin ^2\theta [(1+\sin \theta ^{\prime })^2e^{2i(\varphi -\varphi ^{\prime })} \nonumber \\ &&+(1-\sin \theta ^{\prime })^2e^{-2i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ +\sin 2\theta \cos \theta ^{\prime }[(1+\sin \theta ^{\prime })e^{i(\varphi -\varphi ^{\prime })} \nonumber \\ &&-(1-\sin \theta ^{\prime })e^{-i(\varphi -\varphi ^{\prime })}]+\cos ^2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ni99a} \end{eqnarray} \begin{eqnarray} Y(2,1^{(\widehat{\mathbf{u}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{128\pi }}% \{-\sin ^2\theta \cos \theta ^{\prime }[(1+\sin \theta ^{\prime })e^{2i(\varphi -\varphi ^{\prime })} \nonumber \\ &&-(1-\sin \theta ^{\prime })e^{-2i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ -\sin 2\theta [(1-2\sin \theta ^{\prime })(1+\sin \theta ^{\prime })e^{i(\varphi -\varphi ^{\prime })} \nonumber \\ &&+(1+2\sin \theta ^{\prime })(1-\sin \theta ^{\prime })e^{-i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ +\sin 2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ni99b} \end{eqnarray} \begin{eqnarray} Y(2,0^{(\widehat{\mathbf{u}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{128\pi }}% \{\cos ^2\theta ^{\prime }\sin ^2\theta (e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ -\sin 2\theta \sin 2\theta ^{\prime }[e^{i(\varphi -\varphi ^{\prime })}+e^{-i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ +\frac 23(2\sin ^2\theta ^{\prime }-\cos ^2\theta ^{\prime })(3\cos ^2\theta -1)\} \label{ni99c} \end{eqnarray} \begin{eqnarray} Y(2,(-1)^{(\widehat{\mathbf{u}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{% 128\pi }}\{-\sin ^2\theta \cos \theta ^{\prime }[(1-\sin \theta ^{\prime })e^{2i(\varphi -\varphi ^{\prime })} \nonumber \\ &&-(1+\sin \theta ^{\prime })e^{-2i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ +\sin 2\theta [(1+2\sin \theta ^{\prime })(1-\sin \theta ^{\prime })e^{i(\varphi -\varphi ^{\prime })} \nonumber \\ &&+(1-2\sin \theta ^{\prime })(1+\sin \theta ^{\prime })e^{-i(\varphi -\varphi ^{\prime })}] \nonumber \\ &&\ -\sin 2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ni99d} \end{eqnarray} \begin{eqnarray} Y(2,(-2)^{(\widehat{\mathbf{u}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{% 128\pi }}\{\frac 12\sin ^2\theta [(1-\sin \theta ^{\prime })^2e^{2i(\varphi -\varphi ^{\prime })} \nonumber \\ &&+(1+\sin \theta ^{\prime })^2e^{-2i(\varphi -\varphi ^{\prime })} \nonumber \\ &&\ -\sin 2\theta \cos \theta ^{\prime }[(1-\sin \theta ^{\prime })e^{i(\varphi -\varphi ^{\prime })} \nonumber \\ &&-(1+\sin \theta ^{\prime })e^{-i(\varphi -\varphi ^{\prime })}]+\cos ^2\theta ^{\prime }(3\cos ^2\theta -1)\} \label{ni99e} \end{eqnarray} \subsection{Probability Amplitudes for the $y^{\prime }$ Direction} The prescription for obtaining the probability amplitudes and probability densities corresponding to $y^{\prime }$ is to set $\theta ^{\prime }=\pi /2$% , $\varphi ^{\prime }\rightarrow \varphi ^{\prime }-\pi /2$ in the expressions corresponding to the $z^{\prime }$ direction. As well as transforming the unit vector $\widehat{\mathbf{a}}$ to the unit vector $% \widehat{\mathbf{v}}$, this yields the probability amplitudes: \begin{eqnarray} Y(2,2^{(\widehat{\mathbf{v}})};\theta ,\varphi ) &=&-\sqrt{\frac{15}{128\pi }% }\{\frac 12\sin ^2\theta (e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&+i\sin 2\theta [e^{i(\varphi -\varphi ^{\prime })}+e^{-i(\varphi -\varphi ^{\prime })}]-(3\cos ^2\theta -1)\} \label{ni99k} \end{eqnarray} \begin{eqnarray} Y(2,1^{(\widehat{\mathbf{v}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{128\pi }}% \{\sin ^2\theta (-e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ +i\sin 2\theta [e^{i(\varphi -\varphi ^{\prime })}-e^{-i(\varphi -\varphi ^{\prime })}]\} \label{ni99l} \end{eqnarray} \begin{eqnarray} Y(2,0^{(\widehat{\mathbf{v}})};\theta ,\varphi ) &=&-\sqrt{\frac{45}{256\pi }% }\{\sin ^2\theta (e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ \ +\frac 23(3\cos ^2\theta -1)\} \label{ni99m} \end{eqnarray} \begin{eqnarray} Y(2,(-1)^{(\widehat{\mathbf{v}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{% 128\pi }}\{\sin ^2\theta (-e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ \ +i\sin 2\theta [e^{i(\varphi -\varphi ^{\prime })}-e^{-i(\varphi -\varphi ^{\prime })}]\} \label{ni99n} \end{eqnarray} \begin{eqnarray} Y(2,(-2)^{(\widehat{\mathbf{v}})};\theta ,\varphi ) &=&\sqrt{\frac{15}{% 128\pi }}\{-\frac 12\sin ^2\theta (e^{2i(\varphi -\varphi ^{\prime })}+e^{-2i(\varphi -\varphi ^{\prime })}) \nonumber \\ &&\ +i\sin 2\theta [e^{i(\varphi -\varphi ^{\prime })}+e^{-i(\varphi -\varphi ^{\prime })}]+3\cos ^2\theta -1\} \nonumber \\ && \label{ni99o} \end{eqnarray} We emphasize that the unit vectors $\widehat{\mathbf{u}}$, $\widehat{\mathbf{% v}}$ and $\widehat{\mathbf{a}}$ define a system of mutually orthogonal coordinate axes. \section{ General Properties of the Generalized Spherical harmonics} The generalized quantities presented here reduce to the standard quantities in the limit $\theta ^{\prime }=\varphi ^{\prime }=0$, which corresponds to the arbitrary vector $\widehat{\mathbf{a}}$ pointing in the direction of the $z$ axis. Thus, in this limit, we get \begin{equation} Y(2,m^{(\mathbf{\hat a})};\theta ,\varphi )\rightarrow Y_{2m}(\theta ,\varphi ) \label{ni99u} \end{equation} A property of special interest with regard to the ordinary spherical harmonics is their behaviour under the parity operation $\mathbf{r}% \rightarrow -\mathbf{r,}$ a reflection in the origin. Under this operation, the spherical polar coordinates $(r,\theta ,\varphi )$ transform thus: $% r\rightarrow r,$ $\theta \rightarrow \pi -\theta ,,\;\phi \rightarrow \phi +\pi $ . Thus if $\rho $ is the parity operator defined by \begin{equation} \rho \Psi (\mathbf{r})=\Psi (-\mathbf{r}). \label{hu100} \end{equation} then \begin{equation} \rho Y(l,m^{(\widehat{\mathbf{k}})};\theta ,\varphi )=Y(l,m^{(\widehat{% \mathbf{k}})};\pi -\theta ,\varphi +\pi ) \label{hu101} \end{equation} As is well-known however, \begin{equation} Y(l,m^{(\widehat{\mathbf{k}})};\pi -\theta ,\varphi +\pi )=(-1)^lY(l,m^{(% \widehat{\mathbf{k}})};\theta ,\varphi ) \label{hu102} \end{equation} so that $Y(l,m^{(\widehat{\mathbf{k}})};\theta ,\varphi )$ has even parity if $l $ is even and odd parity if $l$ is odd. The generalized spherical harmonics have the form \begin{equation} Y(l,m^{(\widehat{\mathbf{a}})};\theta ,\varphi )=\sum_jc_jY(l,m_j^{(\widehat{% \mathbf{k}})};\theta ,\varphi ) \label{hu103} \end{equation} where $c_j=\chi (l,m_i^{(\widehat{\mathbf{a}})};l,m_f^{(\widehat{\mathbf{k}}% )})$ is a constant with respect to the angles $(\theta ,\varphi ).$ Hence, \begin{equation} \rho Y(l,m^{(\widehat{\mathbf{a}})};\theta ,\varphi )=\sum_jc_j\rho Y(l,m_j^{(\widehat{\mathbf{k}})};\theta ,\varphi )=(-1)^lY(l,m^{(\widehat{% \mathbf{a}})};\theta ,\varphi ) \label{hu104} \end{equation} Thus, the generalized spherical harmonics have the same parity as the corresponding standard spherical harmonics. The generalized spherical harmonics for value of l can be shown to be orthonormal: \begin{equation} \iint Y^{*}(l,m^{\prime (\widehat{\mathbf{a}})};\theta ,\varphi )Y(l,m^{(% \widehat{\mathbf{a}})};\theta ,\varphi )d\Omega =\delta _{m^{\prime }m} \label{hu105} \end{equation} Thus, since \begin{equation} Y(l,m^{(\widehat{\mathbf{a}})};\theta ,\varphi )=\sum_j\chi (l,m;l,m_j^{(% \widehat{\mathbf{k}})})Y(m_j^{(\widehat{\mathbf{k}})};\theta ,\varphi ) \label{hu106} \end{equation} and \begin{equation} Y^{*}(l,m^{\prime (\widehat{\mathbf{a}})};\theta ,\varphi )=\sum_{j^{\prime }}\chi ^{*}(l,m^{\prime (\widehat{\mathbf{a}})};l,m_{j^{\prime }}^{(\widehat{% \mathbf{k}})})Y^{*}(m_{j^{\prime }}^{(\widehat{\mathbf{k}})};\theta ,\varphi ) \label{hu107} \end{equation} the overlap integral is \begin{eqnarray} I &=&\iint \sum_{j^{\prime }}\chi ^{*}(l,m^{\prime (\widehat{\mathbf{a}}% )};l,m_{j^{\prime }}^{(\widehat{\mathbf{k}})})Y^{*}(l,m_{j^{\prime }}^{(% \widehat{\mathbf{k}})};\theta ,\varphi ) \nonumber \\ &&\times \sum_j\chi (l,m^{(\widehat{\mathbf{a}})};l,m_j^{(\widehat{\mathbf{k}% })})Y(l,m_j^{(\widehat{\mathbf{k}})};\theta ,\varphi )d\Omega \\ &=&\sum_{j^{\prime }}\sum_j\chi ^{*}(l,m^{\prime (\widehat{\mathbf{a}}% )};l,m_{j^{\prime }}^{(\widehat{\mathbf{k}})})\chi (l,m^{(\widehat{\mathbf{a}% })};l,m_j^{(\widehat{\mathbf{k}})}) \nonumber \\ &&\times \iint Y^{*}(l,m_{j^{\prime }}^{(\widehat{\mathbf{k}})};\theta ,\varphi )Y(l,m_j^{(\widehat{\mathbf{k}})};\theta ,\varphi )d\Omega \\ &=&\sum_{j^{\prime }}\sum_j\chi ^{*}(l,m^{\prime (\widehat{\mathbf{a}}% )};l,m_{j^{\prime }}^{(\widehat{\mathbf{k}})})\chi (l,m^{(\widehat{\mathbf{a}% })};l,m_j^{(\widehat{\mathbf{k}})})\delta _{m_jm_{j^{\prime }}} \nonumber \\ &=&\sum_j\chi ^{*}(l,m^{\prime (\widehat{\mathbf{a}})};l,m_j^{(\widehat{% \mathbf{k}})})\chi (l,m;l,m_j^{(\widehat{\mathbf{k}})}) \nonumber \\ &=&\delta _{m^{\prime }m} \label{hu108} \end{eqnarray} In the proof, we have used the result \begin{equation} \sum_j\chi ^{*}(l,m^{\prime (\widehat{\mathbf{a}})};l,m_j^{(\widehat{\mathbf{% k}})})\chi (l,m;l,m_j^{(\widehat{\mathbf{k}})})=\delta _{m^{\prime }m} \label{hu109} \end{equation} which is just the orthonormality relation for the spin probability amplitudes. \section{Discussion and Conclusion} This work has extended the derivation of the new generalized spherical harmonics to the case $l=2$. The expressions for the functions have been derived, as well as the corresponding probability densities for the $% z^{\prime }$ direction. By means of simple transformations the corresponding expressions for the $x^{\prime }$ and $y^{\prime }$ directions have been obtained. Now, for the case $l=1$, it has been shown that the generalized spherical harmonics satisfy the eigenvalue equation[11] \begin{equation} L_{(\widehat{\mathbf{a}})}Y(1,m^{(\widehat{\mathbf{a}})};\theta ,\varphi )=m\hbar Y(1,m^{(\widehat{\mathbf{a}})};\theta ,\varphi ) \label{hu110} \end{equation} where \begin{equation} L_{(\widehat{\mathbf{a}})}=i\hbar \{\sin \theta ^{\prime }\sin (\varphi -\varphi ^{\prime })\frac \partial {\partial \theta }+[\sin \theta ^{\prime }\cot \theta \cos (\varphi -\varphi ^{\prime })-\cos \theta ^{\prime }]\frac \partial {\partial \varphi }\} \label{hu111} \end{equation} We note that $L_{(\widehat{\mathbf{a}})}$ can also be written as $% L_{z^{\prime }}$ since as argued in the section on probability amplitudes and probability densities, it is convenient to think of the vector $\widehat{% \mathbf{a}}$ as defining a new $z$ direction, denoted by $z^{\prime }$. It is expected that all generalized spherical harmonics satisfy the eigenvalue equation, Eq. (\ref{hu110}). This is tedious to prove in practice, and has not been done for the present case $l=2$. This will be tackled in the near future, since it is an important part of the proof of the correctness of the philosophy underlying this work. \section{References} 1. Mweene H. V., ''Derivation of Spin Vectors and Operators From First Principles'', quant-ph/9905012 2. Mweene H. V., ''Generalized Spin-1/2 Operators and Their Eigenvectors'', quant-ph/9906002 3. Mweene H. V., ''Alternative Forms of Generalized Vectors and Operators for Spin 1/2'', quant-ph/9907031 4. Mweene H. V., ''Spin Description and Calculations in the Land\'e Interpretation of Quantum Mechanics'', quant-ph/9907033 5. Mweene H. V., ''Vectors and Operators for Spin 1 Derived From First Principles'', quant-ph/9906043 6. Mweene H. V., Unposted results on spin 3/2 systems. 7. Mweene H. V., ''Generalized Probability Amplitudes for Spin Projection Measurements on Spin 2 Systems'', quant-ph/0502005 8. Mweene H. V., Unposted results on spin 5/2 systems. 9. Mweene H. V., ''New Treatment of Systems of Compounded Angular Momentum'', quant-ph/9907082. 10. Mweene H. V., ''Derivation of Standard Treatment of Spin Addition From Probability Amplitudes'', quant-ph/0003056 11. Mweene H. V., ''Generalized Spherical Harmonics'', quant-ph/0211135 12. Land\'e A., ''From Dualism To Unity in Quantum Physics'', Cambridge University Press, 1960. 13. Land\'e A., ''New Foundations of Quantum Mechanics'', Cambridge University Press, 1965. 14. Land\'e A., ''Foundations of Quantum Theory,'' Yale University Press, 1955. 15. Land\'e A., ''Quantum Mechanics in a New Key,'' Exposition Press, 1973. 16. Mweene H. V., ''Proposed Differential Equation for Spin 1/2'', \textit{% Proc. Third Int. Workshop on Contemporary Problems in Mathematical Physics, Cotonou 2003}, ed. J. Govaerts, M. N. Hounkounnou and A. Z. Msezane (World Scientific, 2004), quant-ph/0411060. \end{document}
{ "timestamp": "2005-03-05T07:21:54", "yymm": "0503", "arxiv_id": "quant-ph/0503059", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503059" }
\section{1. Introduction} Quantum algorithms provide elegant opportunities to harness available quantum resources and perform certain computational tasks more efficiently than classical devices. The idea that a quantum computer could simulate the physical behavior of a quantum system as well as perform computation, attracted immediate attention \cite{preskill,ss}. The theory of such quantum computers is now well understood and several quantum algorithms like Deutsch-Jozsa (DJ) algorithm \cite{deu}, Grover's search algorithm \cite{grover}, Shor's prime factorization algorithm \cite{shor}, Hogg's algorithm \cite{hogg},Bernstein-Vazirani problem \cite{vazi} and quantum counting \cite{count1} have been developed . All these algorithms start from a well-defined initial state and perform computation by a sequence of reversible logic gates. After computation, the final state of the system gives the output. Various methods are being examined for building a quantum information processing (QIP) device which is coherent and unitary \cite{bou}. Nuclear Magnetic Resonance has emerged as a leading candidate for implementation of various quantum computational problems on a small number of qubits \cite{cory97,chuang97,cory98,djchu,djjo,grochu,grojo,ka1,ka,jcp,nat,ranapra2,ijqi,ranapra1,ranabirtomo}. Quantum adiabatic algorithms provide an alternative method for computing \cite{ad1,ad2}. In this method the computation is done by evolving the system under a Hamiltonian for a given amount of time. Such algorithms start from a suitable input ground state and by evolution under a slowly time-varying Hamiltonian, reach the desired output state. Quantum adiabatic algorithms have been efficiently applied to solve various optimization problems \cite{ad3,ad4,ad5,ad6}. Chuang {\it et al.} have demonstrated the implementation of a quantum adiabatic algorithm by solving the MAX-CUT \cite{garey} problem on a three qubit system by NMR \cite{chu} . In these algorithms, the condition for adiabaticity is fulfilled globally by using only the minimum energy gap between the ground state and the first excited state for calculating the time of evolution. This method of evolution is not efficient in some cases such as adiabatic Grover's search algorithm and adiabatic \dj algorithm as they result in a complexity O(N) (N is the size of the data set), which is as good as their classical algorithms. However, these algorithms can be improved by application of local adiabatic evolution, where the adiabatic condition is fulfilled at each instant of time. This technique has been adopted theoretically by Roland and Cerf \cite{cerf} for the adiabatic Grover's search algorithm and by S. Das {\it et al.} for adiabatic \dj algorithm \cite{das} yielding a complexity O($\sqrt{N}$). Experimental implementation of adiabatic Grover's search algorithm based on the proposal of Roland and Cerf and adiabatic \dj algorithm of S. Das {\it et al.}, is reported here. Section 2 contains an introduction to adiabatic algorithms. Section 3 discusses the adiabatic version of the Grover's search algorithm proposed by Roland and Cerf and its NMR implementation. Section 4 discusses the adiabatic \dj algorithm and its NMR implementation. Section 5 contains the experimental results, on a 2-qubit system, for both these algorithms. To the best of our knowledge this is the first experimental implementation of adiabatic Grover's search and adiabatic \dj algorithms. \section{2. Adiabatic Algorithm} The adiabatic theorem of quantum mechanics states that when a system is evolved under a slowly time varying Hamiltonian, it stays in its instantaneous ground state \cite{me}. This fact is used in solving certain computational problems \cite{ad3,ad4,ad5,ad6}. The problem to be solved is encoded in a final Hamiltonian ($H_F$), whose ground state is not easy to find. Adiabatic algorithms start with the ground state of a beginning Hamiltonian ($H_B$) which is easy to construct and whose ground state is also easy to prepare. The ground state of $H_B$, which is a superposition of all the eigenstates of H$_F$, is evolved under a time varying Hamiltonian $H(s)$. $H(s)$ is a linear interpolation of the beginning Hamiltonian $H_B$ and the final Hamiltonian $H_F$ such that \begin{eqnarray} H(s)= (1-s)H_B + s H_F, \hspace{3cm}\mbox{where}\;\;\; 0\leq s \leq 1. \label{hs} \end{eqnarray} The parameter $s=t/T_{total}$, where $T_{total}$ is the total time of evolution and $t$ varies from 0 to $T_{total}$. After evolution under the Hamiltonian $H(s)$ for a time $T_{total}$, the system is in the ground state of $H_F$ with a probability $(1-\varepsilon^2)^2$, provided the evolution rate satisfies, \begin{eqnarray} \frac{\underset{0\leq s \leq 1}{max}\left|\left<1;s\left|\frac{dH(s)}{dt}\right| 0;s\right>\right|}{g_{min}^{2}} \leq \varepsilon, \label{r1} \label{epsilon} \end{eqnarray} and the parameters of the algorithm are chosen to make $\varepsilon \ll$1 \cite{ad1}. The numerator in Eq. 2 is the transition amplitude between the ground state and the first excited state of {\it H(s)}, and the denominator is the square of the smallest energy gap $(g_{min})$ between them. Ideally the time of evolution ($T_{total}$) must be infinite. However as long as the gap is finite, for any finite and positive $\varepsilon$, the time of evolution can be finite. The time of evolution of the algorithm is determined by the minimum energy gap between the ground state and the first excited state. In the adiabatic case the time of evolution determines the complexity of the algorithms (that is how long it takes for the task to be completed), which can then be compared to the complexity of the discrete algorithms in classical and quantum paradigms. The time of evolution is measured in units of natural time scale associated with the system, $\bar T$ which is O($\hbar /{\bar E}$) where $\bar E$ is the fundamental energy scale associated with the physical system used to construct the states. \cite{das}. \indent In the actual implementation, the Hamiltonian $H(s)$ is discretized into $M+1$ steps as $H(\frac{m}{M})$ where m goes from $0\rightarrow M$ \cite{chu,wvd}. Thus the time varying Hamiltonian $H(s)$ goes from beginning Hamiltonian to final Hamiltonian in M+1 steps. As the total number of steps increase, the evolution becomes more and more adiabatic \cite{chu}. The evolution operator for the m$^{\rm th}$ step is given by \cite{chu} \begin{eqnarray} U_m=e^{-i[(1-\frac{m}{M})H_B + \frac{m}{M}H_F]\Delta t}, \end{eqnarray} where $ \Delta t=T/(M+1)$. The total evolution is given by, \begin{eqnarray} U=\prod_{m=0}^M U_m. \end{eqnarray} Since, $H_B$ and $H_F$ do not commute in general, the evolution operator of Eq. 3 is approximated to first order in $\Delta$t, by the use of the Trotter's formula \cite{chu} as \begin{eqnarray} U_m \approx e^{-iH_{B}(1-\frac{m}{M})\frac{\Delta t}{2}}\cdot e^{-iH_{F}\frac{m}{M}\Delta t}\cdot e^{-iH_{B}(1-\frac{m}{M})\frac{\Delta t}{2}}. \label{eq:trot} \end{eqnarray} Thus in each step only a small evolution of the system from ground state of ${\rm H_B}$ towards the ground state of ${\rm H_F}$ takes place. \section{3. Grover's search algorithm} Suppose we are given an unsorted database of N items and one of those items is marked. To search for the marked item classically, it would require on an average N/2 queries. However using quantum resources, the algorithm prescribed by Grover \cite{grover} performs the same search with O($\sqrt{N}$) queries. The algorithm starts with an equal superposition of states, representing the items, repeatedly flips the amplitude of the marked state (done by the oracle) followed by the flip of the amplitudes of all the states about the mean. The number of times this process is repeated determines the complexity of the algorithm and this scales with the size of the database as O($\sqrt{N}$). \\ \indent In the adiabatic version, the system is evolved under a time dependent Hamiltonian which is a linear interpolation of H$_B$ and H$_F$. As n qubits are used to label a database of size N (=2$^n$), the resulting Hilbert space is of dimension N. The basis states in this space are $\vert i\rangle$ where i=0,$\cdots$,N. H$_B$ is chosen such that the ground state is a linear superposition of all the basis states. Therefore for a 2-qubit case, \begin{eqnarray} \vert\psi_B\rangle &=& \frac{1}{2}\left(\vert 00\rangle + \vert 01\rangle + \vert 10\rangle + \vert 11\rangle\right). \label{groinistate}\\ H_B &=& I - \vert\psi_B\rangle\langle\psi_B\vert, \cr &=& I - \frac{1}{4}\begin{pmatrix}1&1&1&1 \cr 1&1&1&1 \cr 1&1&1&1 \cr 1&1&1&1\end{pmatrix}.\label{grohb} \end{eqnarray} The Final Hamiltonian has the marked state $\vert\psi_F \rangle$ as the ground state. \begin{eqnarray} H_F &=& I - \vert\psi_F \rangle\langle\psi_F \vert. \label{grohf} \end{eqnarray} The rate at which the interpolating Hamiltonian H(s) (given by Eq. \ref{hs}) changes from $H_B$ to $H_F$ depends on the condition, \begin{eqnarray} \left|\frac{ds}{dt}\right| \leq \varepsilon \frac{g^{2}(s)}{\left|\langle\frac{dH}{ds}\rangle\right|}. \label{adcon} \end{eqnarray} Following Roland and Cerf \cite{cerf}, t is obtained as a function of s as, \begin{eqnarray} t=\frac{1}{2\varepsilon}\frac{N}{\sqrt{N-1}}\left[arctan\{\sqrt{N-1}\left(2s-1\right)\}+arctan\sqrt{N-1}\right]. \end{eqnarray} Taking t$'= \varepsilon t$ and on inverting the above function, s(t$'$) is obtained as \begin{eqnarray} s(t') = \frac{1}{2}\left[\{ \frac{1}{\sqrt{N-1}}tan\left(\frac{2\sqrt{N-1}t'}{N} - arctan\sqrt{N-1} \right)\}+1\right]. \label{stprime} \end{eqnarray} The plot of this function for N=4 (for a 2 qubit case) is given in Fig. 1. In the experiment the time of evolution is varied according to Eq. \ref{stprime}. It has been shown by Roland and Cerf \cite{cerf} that with this adiabatic evolution, the complexity of the algorithm is O($\sqrt{N}$). \section{3.1. Experimental Implementation} The NMR Hamiltonian for a weakly coupled two-spin system is : \begin{eqnarray} {\mathcal H}= -\omega_1 I_{z1} - \omega_2 I_{z2} +2\pi J_{12}I_{z1}I_{z2}. \label{nmrham} \end{eqnarray} where $\omega_1$ and $\omega_2$ are Larmour frequencies and $J_{12}$ the indirect spin-spin coupling. The beginning Hamiltonian for a 2-qubit Grover's algorithm as stated in Eq. \ref{grohb}, written in terms of spin-half operators, is \begin{eqnarray} {\mathcal H}_B = \frac{3}{4} I - \frac{1}{2} \{I_{x1} + I_{x2} + 2 I_{x1}I_{x2}\}. \end{eqnarray} The identity term does not cause any evolution of the state and so it can be omitted, yielding the beginning Hamiltonian without the negative sign and the factor half as: \begin{eqnarray} \tilde{\mathcal H}_B = I_{x1} + I_{x1} + 2I_{x1}I_{x2} \label{nmrhb} \end{eqnarray} The evolution under $\tilde{\mathcal H}_B$ can be simulated by a free evolution under the Hamiltonian ${\mathcal H}$ of Eq. \ref{nmrham} between two $\pi$/2 pulses with appropriate phases. \begin{eqnarray} e^{i\frac{\pi}{2}\left(I_{y1} + I_{y2}\right)}\cdot e^{i{\mathcal H}T}\cdot e^{-i\frac{\pi}{2}\left(I_{y1} + I_{y2}\right)} &=&e^{i\left(\omega_1 I_{x1} + \omega_2 I_{x2} + 2JI_{x1}I_{x2}\right)T}\cr &=& e^{i{\mathcal H}'T} \label{hbevol} \end{eqnarray} Let the state $\vert 00\rangle$ be the marked state. The final Hamiltonian is, \begin{eqnarray} H_{F}^{\vert 00\rangle} = I - \begin{pmatrix}1&0&0&0\cr 0&0&0&0\cr 0&0&0&0\cr 0&0&0&0\cr\end{pmatrix} \end{eqnarray} In terms of spin operators the final Hamiltonian is, \begin{eqnarray} {\mathcal H}_{F}^{\vert 00\rangle} = \frac{3}{4} I - \frac{1}{2}\left[I_{z1} + I_{z2} + 2I_{z1}I_{z2}\right] \end{eqnarray} The final Hamiltonian keeping the spin operator terms only and without the negative sign and the factor half is \begin{eqnarray} \tilde{\mathcal H}_{F}^{\vert 00\rangle} = I_{z1} + I_{z2} + 2I_{z1}I_{z2} \label{grohf00} \end{eqnarray} Similarly the final Hamiltonian for other states being marked, in terms of the spin-half operators, is \begin{eqnarray} \tilde{\mathcal H}_{F}^{\vert 01\rangle}&=&I_{z1} - I_{z2} - 2I_{z1}I_{z2} \label{grohf01} \\ \tilde{\mathcal H}_{F}^{\vert 10\rangle}&=&-I_{z1} + I_{z2} - 2I_{z1}I_{z2} \label{grohf10} \\ \tilde{\mathcal H}_{F}^{\vert 11\rangle}&=&-I_{z1} - I_{z2} + 2I_{z1}I_{z2} \label{grohf11} \end{eqnarray} \indent The schematic representation of the experiment for the adiabatic Grover's algorithm in a two qubit system [consisting of a $^1$H spin and a $^{13}$C spin] is shown in Fig. 2a. The experiment is divided into three parts. The first part (preparation part) consists of preparation of pseudo-pure state (PPS) followed by equal superposition. The second part is the adiabatic evolution, and the third part is the tomography of the resultant state. The pulse programme for the preparation of PPS and equal superposition is shown in Fig. 2b. The PPS is prepared by the method of spatial averaging \cite{du}. After preparing PPS, equal superposition of states is obtained by application of the Hadamard gate on both the qubits. The Hadamard gate is implemented by $(\pi/2)_y$ -pulses, followed by $\pi_x$ -pulses on both proton and carbon spins (Fig. 2b) \cite{grochu}. The next stage consists of adiabatic evolution which has been carried out in the present work in 60 steps. Each step of the adiabatic evolution (Figs. 2c, 2d, 2e and 2f) consists of evolution under the final Hamiltonian for a time $\tau$ sandwiched between two evolutions under the beginning Hamiltonian for a time (T-$\tau$)/2. T is the total evolution time for one step and is equal to 1/$\pi$J. The value of $\tau (= s\times \frac{1}{\pi J})$ varies from 0 to T takes place as `s' increases from 0 to 1 according to Eq. \ref{stprime}, in 60 steps. The pulse sequence for the beginning Hamiltonian is a free evolution of the system juxtaposed between two $\pi$/2 pulses with appropriate phases on each of the spins (the part marked as H$_B$ in Figs. 2c-2f). The pulse sequence for the final Hamiltonian depends on the marked state as stated in Eqs. \ref{grohf01}-\ref{grohf11}. If the state $\vert 00\rangle$ is the marked state, then the pulse sequence for the implementation of the final Hamiltonian is a free evolution of the system under the NMR Hamiltonian juxtaposed between two $\pi$ pulses on each of the spins (Fig. 2c). Similarly, if the state $\vert 01\rangle$ is marked the pulse sequence for the final Hamiltonian is a free evolution of the system between two $\pi$ pulses on the spin 1 (Fig. 2d), if the state $\vert 10\rangle$ is marked then the pulse sequence is a free evolution between two $\pi$ pulses on the spin 2 (Fig. 2e) and if the state $\vert 11\rangle$ is marked, then the pulse sequence simulating the final Hamiltonian is just a free evolution of the system under the NMR Hamiltonian (Fig. 2f). \indent The third stage of the experiment is the tomography of the final density matrix after the adiabatic evolution. The density matrix of a 2-spin system is a 4$\times$4 matrix consisting of 6 independent off-diagonal complex elements (the remaining 6 are their complex conjugates), and the four diagonal elements which are the populations of the various levels. The diagonal elements are measured by 90$^o$ pulses on each qubit preceded by a gradient pulse. The six off-diagonal elements consist of four single quantum (SQ), one double quantum (DQ) and one zero quantum (ZQ) coherences. The real and the imaginary SQ, DQ and ZQ coherences in terms of the spin operators are; \begin{eqnarray} {\rm SQ}^{real}_{i} &=& I_{ix}\pm 2(I_{ix}I_{jz}), \cr {\rm SQ}^{imag}_{i} &=& I_{iy}\pm 2(I_{iy}I_{jz}), \cr {\rm DQ}^{real} &=& 2(I_{ix}I_{jx} - I_{iy}I_{jy}), \cr {\rm DQ}^{imag} &=& 2(I_{iy}I_{jx} + I_{ix}I_{jy}), \cr {\rm ZQ}^{real} &=& 2(I_{ix}I_{jx} + I_{iy}I_{jy}), \cr {\rm ZQ}^{imag} &=& 2(I_{iy}I_{jx} - I_{ix}I_{jy}), \end{eqnarray} where i$\neq$ j = 1,2 represents the qubits. Although the single quantum terms are directly observable, for proper scaling, all the off-diagonal elements are observed by a common protocol of two experiments; \begin{eqnarray} {\rm A:}\hspace{3cm}\left(\frac{\pi}{2}\right)^{i}_{\phi_1}\left(\theta\right)^{j}_{\phi_2} \longrightarrow &G_{z}& \longrightarrow\left(\frac{\pi}{2}\right)^{i}_{y}, \\ {\rm B:}\hspace{3cm}\left(\frac{\pi}{2}\right)^{i}_{\phi_1}\left(\theta\right)^{j}_{\phi_2} \longrightarrow &G_{z}& \longrightarrow\left(\pi \right)^{j}\left(\frac{\pi}{2}\right)^{i}_{y}. \end{eqnarray} where $\theta$ denotes the pulse angle, $\phi_1$, $\phi_2$ the pulse phases and $G_z$ a gradient pulse. The first two pulses of the experiment A (depending on the pulse angle $\theta$ and the pulse phases $\phi_1$ and $\phi_2$) convert terms like $I_{i\alpha}+2I_{i\alpha}I_{j\beta}$ into diagonal terms given by $I_{iz}+2I_{iz}I_{jz}$, where $\alpha$ and $\beta$ denote the x, y, or z component of the spin operators of the first and the second qubit respectively. The gradient destroys all the transverse magnetization retaining only the longitudinal terms. The last pulse converts the retained longitudinal magnetization $I_{iz}+2I_{iz}I_{jz}$ into observable terms $I_{ix}+2I_{ix}I_{jz}$. Thus the magnitude of $I_{i\alpha}+2I_{i\alpha}I_{j\beta}$ is mapped on to $I_{ix}+2I_{ix}I_{jz}$ which is then observed. In experiment B, a $\pi$-pulse is applied on the spin `j' just before the $\pi$/2 pulse on the spin `i'. This creates the observable term $I_{ix}-2I_{ix}I_{jz}$. The sum and difference of the two experiments yields 2$I_{i\alpha}$ and 2$I_{i\alpha}I_{j\beta}$ respectively. Six different experiments are needed to be performed to map the whole density matrix (real and imaginary). The various pulse angles and phases required during the experiment, and the resultant terms that are observed due to them are given in Table I. Experiments I and II yield the SQ, and experiments III-VI yield the ZQ and DQ coherences. \section{4. Deutsch-Jozsa Algorithm} The \dj Algorithm determines whether a binary function $f(x)$, \begin{eqnarray*} f(x\vert x\in\{ 0,1\}^n) \rightarrow \{0,1\}, \end{eqnarray*} is Constant or Balanced \cite{dja}.A constant function implies that the function has the same value 0 or 1 for all $x$. A balanced function implies that the function {\it`f'} is 0 for half the values of $x$ and 1 for the other half . For a two qubit case the constant and the balanced functions are given in Table II. In the adiabatic version of the \dj algorithm, the beginning Hamiltonian and its ground state, for a two qubit system, is given by Eq. \ref{grohb} and Eq. \ref{groinistate} respectively. The final Hamiltonian is given by Eq. \ref{grohf} and the ground state of the final Hamiltonian for two qubits is of the form \cite{das}; \begin{eqnarray} \vert\psi_F\rangle &=& \alpha\vert00\rangle + \frac{\beta}{\sqrt 3}\left(\vert01\rangle + \vert10\rangle + \vert11\rangle \right), \label{djfinstate} \end{eqnarray} where \begin{eqnarray} \alpha &=& \frac{1}{4}\left|(-1)^{f(00)}+(-1)^{f(01)}+(-1)^{f(10)}+(-1)^{f(11)}\right|, \cr \beta^2 &=& 1-\alpha^2. \label{ab} \end{eqnarray} From Eq. \ref{ab} it is seen that when $\alpha=1$ the function $f$ is constant, and when $\alpha=0$ then it is balanced. Thus $\alpha$ is chosen depending on whether the function to be encoded in the final Hamiltonian is constant or balanced. \indent Using Eqs. \ref{hs},\ref{grohb},\ref{grohf},\ref{djfinstate} and \ref{ab} the matrix for the interpolating Hamiltonian($H(s)$) can be written as \cite{das}; \begin{eqnarray} H(s)=I-\frac{1-s}{4} \begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 \end{pmatrix} -\frac{s}{3} \begin{pmatrix} 3\alpha & 0 & 0 & 0 \\ 0 & \beta & \beta & \beta \\ 0 & \beta & \beta & \beta \\ 0 & \beta & \beta & \beta \end{pmatrix}. \label{hsmatrix} \end{eqnarray} S. Das {\it et al.\;} have shown that on evolution under the Hamiltonian $H(s)$ takes the initial state $\vert\psi_B\rangle$ to the solution state $\vert\psi_{F}\rangle$ \cite{das}. In the next section we describe an NMR implementation of the above algorithm. \section{4.1. NMR Implementation} The adiabatic \dj algorithm also, is implemented on the 2-qubit system. The beginning Hamiltonian in terms of the spin-half operators is the same as given in Eq. \ref{nmrhb}, and its implementation has been discussed in section 3.1. \\ \indent The final Hamiltonian, obtained from Eqs. \ref{grohb}, \ref{grohf}, \ref{djfinstate} and \ref{ab}, for constant case ($\alpha$=1) yields, \begin{align} H_{F}^{c} = I - \begin{pmatrix}1&0&0&0 \cr 0&0&0&0 \cr 0&0&0&0 \cr 0&0&0&0\end{pmatrix},\\ \intertext{and for balanced case ($\alpha$=0) yields,} H_{F}^{b} = I - \frac{1}{3}\begin{pmatrix}0&0&0&0 \cr 0&1&1&1 \cr 0&1&1&1 \cr 0&1&1&1\end{pmatrix}. \end{align} The above final Hamiltonians in terms of spin-half operators can be written respectively as, \begin{align} {\mathcal H}_F^{c} = \frac{3}{4}I - \frac{1}{2}&(I_{z1} + I_{z2} + 2I_{z1}I_{z2}), \intertext{and,} {\mathcal H}_F^{b} = \frac{3}{4}I -\frac{1}{3}&\biggl[-\frac{1}{2}(I_{z1} + I_{z2} + 2I_{z1}I_{z2}) + 2(I_{x1}I_{x2} + I_{y1}I_{y2}) \cr & + I_{x1} + I_{x2} - 2(I_{x1}I_{z2} + I_{z1}I_{x2})\biggr]. \end{align} As the identity does not cause any evolution of the state we consider only the spin operator terms. Thus the final Hamiltonian keeping only the spin operators (dropping the minus sign), for the constant case, can be written as \begin{eqnarray} \tilde{{\mathcal H}}^{c}_F &=& \frac{1}{2}\{I_{z1} + I_{z2} + 2I_{z1}I_{z2}\}, \label{nmrhcf} \end{eqnarray} and for the balanced case as \begin{eqnarray} \tilde{{\mathcal H}}^{b}_F = - \frac{1}{6}(I_{z1} &+ I_{z2} + 2I_{z1}I_{z2})+\frac{2}{3}(I_{x1}I_{x2}+I_{y1}I_{y2}) \cr &+\frac{1}{3}I_{x1}+\frac{1}{3}I_{x2}-\frac{2}{3}(I_{x1}I_{z2}+I_{z1}I_{x2}). \label{nmrhbf} \end{eqnarray} The signs of Eqs. \ref{nmrhb}, \ref{nmrhcf} and \ref{nmrhbf} are changed for consistency. Since the various terms in Eq. \ref{nmrhbf} do not commute, the evolution under this Hamiltonian would require a complex pulse sequence in NMR. However, we have found that by keeping only the diagonal terms in the Eq. \ref{nmrhbf}, the pulse sequence simplifies considerably with the information regarding the balanced nature of the problem still encoded in it. This truncated final Hamiltonian for the balanced case is given by; \begin{eqnarray} (\tilde{\mathcal H}^{b}_{F})^{trunc} = -\frac{1}{6}(I_{z1} + I_{z2} + 2I_{z1}I_{z2}) \label{nmrhbftrunc} \end{eqnarray} The opposite signs of Eq. \ref{nmrhcf} and Eq. \ref{nmrhbftrunc} distinguish the constant and the balanced case. \indent In the following we show that the balanced nature of the \dj problem is still encoded in $(\tilde{\mathcal H}_{F}^{b})^{trunc}$. Substituting $\alpha=0$ and $\beta=1$ and dropping the off-diagonal terms from the last part of Eq. \ref{hsmatrix} , we obtain \begin{eqnarray} \tilde{H}^{b}(s) =I-\frac{1-s}{4} \begin{pmatrix} 1&1&1&1\\1&1&1&1\\1&1&1&1\\1&1&1&1 \end{pmatrix} -\frac{s}{3}\begin{pmatrix} 0&0&0&0\\0&1&0&0\\0&0&1&0\\0&0&0&1 \end{pmatrix}. \end{eqnarray} The eigenvalues of this Hamiltonian are: \begin{eqnarray} \lambda_0 &=&\frac{1}{6}\left[3+2s-\sqrt{9 + s(7s-15)}\right], \\ \lambda_1 &=&\frac{1}{6}\left[3+2s+\sqrt{9 + s(7s-15)}\right], \\ \lambda_2 &=&\lambda_3 = 1-\frac{s}{3}. \end{eqnarray} The values of $\lambda_0$, $\lambda_1$, $\lambda_2$ and $\lambda_3$ as a function of `s' are plotted in Fig. 3. $\lambda_0$ is the ground state. As `s' increases from 0, $\lambda_0$ continues to be the ground state and becomes the ground state of the final Hamiltonian in the limit $s\rightarrow 1$. The eigenvectors corresponding to $\lambda_0$, $\lambda_1$, $\lambda_2$ and $\lambda_3$ are respectively obtained as; \begin{eqnarray} v_0 \!=\! \begin{pmatrix} \frac{3-s-2\sqrt{9-15s+7s^2}}{3(s-1)} \cr 1\cr 1\cr 1 \end{pmatrix},\; v_1 \!=\! \begin{pmatrix} \frac{3-s+2\sqrt{9-15s+7s^2}}{3(s-1)} \cr 1\cr 1\cr 1 \end{pmatrix},\; v_2 \!=\! \begin{pmatrix} 0 \cr -1 \cr 0 \cr 1 \end{pmatrix},\; v_3 \!=\! \begin{pmatrix} 0 \cr -1 \cr 1 \cr 0 \end{pmatrix}, \end{eqnarray} The final state to which the system converges after the evolution is \begin{eqnarray} \underset{s\rightarrow 1}{lim}\;v_0 = \begin{pmatrix}0\cr 1\cr 1\cr 1 \end{pmatrix}, \end{eqnarray} which is the desired output state. The energy gap between the ground state and the states corresponding to $\lambda_2$ and $ \lambda_3$ goes to zero as $s\rightarrow 1$ as shown in Fig. 3. However, there is no transition from $\lambda_0$ to $\lambda_2$, $\lambda_3$ as the transition amplitude given by the numerator in Eq. \ref{epsilon} is zero in these cases. Therefore the transition amplitude from the ground state $\lambda_0$ to the next excited state $\lambda_1$ is relevant for calculation of $s(t)$. The minimum energy gap between $\lambda_0$ and $\lambda_1$, needed in Eq. \ref{epsilon}, is obtained for $s \simeq 1$ as seen in Fig. 3. Since the algorithm is implemented using local adiabatic evolutions we need to change $s(t)$ such that the adiabatic condition \cite{cerf} \begin{eqnarray} \frac{ds}{dt} \leq \varepsilon \frac{\left|g(s)\right|^2}{\left|\left<\frac{dH}{ds}\right>\right|}, \end{eqnarray} is met at each time interval. Here g(s) is the energy gap between the ground state and the first excited state, given by $\frac{1}{3}\sqrt{9 -15s + 7s^2}$ and $\left|\left< dH/ds\right>\right| = H_F - H_B$. The Hamiltonian is evolved at a rate that is a solution of \begin{eqnarray} \frac{ds}{dt} = \varepsilon \frac{\left|g(s)\right|^2}{\left|H_F - H_B\right|} \label{dsdt} \end{eqnarray} On integrating Eq. \ref{dsdt}, we obtain {\it t} as a function of {\it s}. \begin{eqnarray} t=\frac{1}{\varepsilon}\frac{14s-15}{2\sqrt{3}\sqrt{7s^2 -15s + 9}} + k,\label{eq:t} \end{eqnarray} where the constant of integration $k=\frac{5}{\varepsilon 2\sqrt 3}$ to obey $s=0$ at $t=0$. Inverting this function we obtain $s(t')$ as \begin{eqnarray} s(t')=\frac{3}{14}\left[ 5 - \frac{\sqrt{225 + 24t\left( 55\sqrt{3} - 183t +60\sqrt{3}t^2 - 18t^3\right)}}{3 + 20\sqrt{3}t - 12t^2}\right] \end{eqnarray} where $t'$ is $\varepsilon t$. The plot of $s$ as a function of $t'$ is shown in Fig. 4. From Figs. 3 and 4 it is seen that the rate of change of $s$ (and hence of the Hamiltonian) is fast when the energy gap between $\lambda_0$ and $\lambda_1$ is large, and slow when the gap is small. In practice the time of evolution for $H_B$ and $H_F$ is given by $(1-s)\times T$ and $s\times T$ respectively, where $T$ is 1/$\pi$J and $s$ is varied from 0 to 1 according to Eq. \ref{eq:t}. In our implementation, the $t^{'}$ interval for which $s$ varies from 0 to 1 is divided in 80 equal steps, and the corresponding values of s for each step (calculated from Eq. 48) are substituted in the evolution time of $H_B$ and $H_F$. \\ \indent On integrating Eq. 46 from s=0 to s=1, we get the total time of evolution \begin{eqnarray} T_{total}=\frac{1}{\varepsilon}\frac{2}{\sqrt{3}}{\bar T}. \label{ttotal} \end{eqnarray} T$_{total}$ is given in the units of ${\bar T}$ which is the time scale associated with the physical system used \cite{das}. The time scale associated with evolution under the NMR Hamiltonian is $\sim 10^{-3} s$. The total time of evolution of the experiment ($T_{total}$) is given by 80$\times$T, where T is the time for one step (see Fig. 5b). For the choice $\varepsilon\sim 10^{-2}$, T $\sim 60\times 10^{-3}$s in our case. \section{4.2. Experimental Implementation} The experimental implementation of adiabatic \dj algorithm on a 2-qubit system [consisting of a $^1$H spin and a $^{13}$C spin] also consists of three parts namely preparation, adiabatic evolution and tomography of the final density matrix. The preparation of the pseudo pure state (PPS) and making of equal superposing of states as well as the tomography of the final states has already been discussed in section 3. So we only describe the method of implementation of the final Hamiltonian for the \dj algorithm.\\ \indent The pulse sequence for the implementation of the constant case final Hamiltonian ($\tilde{\mathcal H}_{F}^{c}$) is given in Fig. 5b. The beginning Hamiltonian is implemented by a free evolution juxtaposed between $\pi$/2 pulses with required phases (Fig. 5b). The implementation of the final Hamiltonian for the constant case is a free evolution under the NMR Hamiltonian of Eq. 14, juxtaposed between two $\pi$-pulses as shown in Fig. 5b. In the balanced case the implementation of the beginning Hamiltonian is same as in Fig. 5b. However, the implementation of the final Hamiltonian $(\tilde{\mathcal H}_{F}^{b})^{trunc}$ is done in two parts [Fig. 5c]. The first part is a free evolution under the Hamiltonian given in Eq. 14 [${\rm T_f}$ period in Fig. 5c]. The operator corresponding to such an evolution for time $\tau$ will be of the form; \begin{eqnarray} e^{i\pi J(-I_{z1}-I_{z2}+2I_{z1}I_{z2})\tau}. \label{tf} \end{eqnarray} In the second evolution of $2\tau$, the chemical shifts are refocused so that the system evolves only under its scalar coupling Hamiltonian $2\pi JI_{z1}I_{z2}$. Just before and after the evolution $\pi$-pulses with appropriate phases are put on each of the spins to flip the sign of the corresponding spin operator [${\rm T_j}$ period in Fig. 5c]. The operator for the sequence of two pulses with an intermediate evolution for $2\tau$ is of the form \begin{eqnarray} e^{-i(I_{x1})\pi}\cdot e^{i\pi J(2I_{z1}I_{z2})2\tau}\cdot e^{i(I_{x1})\pi}= e^{-i\pi J(2I_{z1}I_{z2})2\tau}. \label{tj} \end{eqnarray} As these two evolutions given in Eq. \ref{tf} and Eq. \ref{tj} commute, the effective evolution for the 3$\tau$ period is: \begin{eqnarray} e^{i\pi J(-I_{z1}-I_{z2}+2I_{z1}I_{z2})\tau} \cdot e^{-i\pi J(2I_{z1}I_{z2})2\tau} =e^{i\pi J(-I_{z1}-I_{z2}-2I_{z1}I_{z2})\tau}.\label{tftj} \end{eqnarray} Thus the evolution during ${\rm T_j}$ cancels the J-evolution during ${\rm T_f}$ and adds a minus sign to it, yielding the effective Hamiltonian of Eq. \ref{tftj} and an effective evolution time of $\tau$. An evolution time of $\tau$=1/$\pi$J implements the full Hamiltonian of Eq. 39 as required for adiabatic evolution. Overall the cycle time for each step for the balanced case is increased to T$+2\tau$.\\ \section{5. Experimental Results} The experiments have been carried out using carbon-13 labeled chloroform ($\rm ^{13}CHCl_3$) where the two spins $^1$H and $^{13}$C form the two qubit system. The proton spin represents the first qubit and carbon-13 the second. The sample of $\rm ^{13}CHCl_3$ was dissolved in the solvent CDCl$_3$ and the experiments were performed at room temperature in a magnetic field of 11.2 Tesla. At this field the $^{1}\rm H$ resonance frequency is 500.13 MHz and the $^{13}\rm C$ resonance frequency is 125.76 MHz. During the entire experiment, the transmitter frequencies of $^{1}\rm H$ and $^{13}\rm C$ are set at a value $J/2$ away from resonance to achieve the condition $\omega_1=\omega_2=\pi$J. The equilibrium spectra of the two qubits are shown in Fig. 6a, and the spectrum corresponding to $\vert 00\rangle$ PPS is shown in Fig. 6b. To quantify the experimental result we calculate the {\it average absolute deviation} \cite{nures} of each element of the experimentally obtained density matrix from each element of the theoretically predicted density matrix given by, \begin{equation} \Delta x=\frac{1}{N^2}\sum^N_{i,j=1}\vert x_{i,j}^{T} - x_{i,j}^{E} \vert \label{error} \end{equation} where N=$2^n$ (n being the number of qubits), $x_{i,j}^T$ is $(i,j)^{th}$ element of the theoretically predicted density matrix and $x_{i,j}^E$ is $(i,j)^{th}$ element of the experimentally obtained density matrix. \section{5.1. Grover's Search Algorithm} The experimental spectra corresponding to the implementation of Grover's search algorithm on the above two qubit system are given in Fig. 7. the spectra given in Figs. 7a(i-iv) contain the reading of populations after respectively searching states $\vert 00\rangle$, $\vert 01\rangle$,$\vert 10\rangle$ and $\vert 11\rangle$. The population spectra are obtained by application of a gradient followed by a $\pi$/2 pulse. Depending on the final state, the population spectra consist of one single spectral line for each spin. These correspond to, $\vert 00\rangle$ $\rightarrow$ $\vert 01\rangle$ and $\vert 00\rangle$ $\rightarrow$ $\vert 10\rangle$transition when the searched state is $\vert 00\rangle$ (Fig. 7a-i); $\vert 01\rangle$ $\rightarrow$ $\vert 00\rangle$ and $\vert 01\rangle$ $\rightarrow$ $\vert 11\rangle$ when the search state $\vert 01 \rangle$ (Fig. 7a-ii); $\vert 10\rangle$ $\rightarrow$ $\vert 00\rangle$ and $\vert 10\rangle$ $\rightarrow$ $\vert 11\rangle$ when the search state $\vert 10 \rangle$ (Fig. 7a-iii); $\vert 11\rangle$ $\rightarrow$ $\vert 01\rangle$ and $\vert 11\rangle$ $\rightarrow$ $\vert 10\rangle$ when the search state $\vert 11 \rangle$ (7a-iv). The coherence spectra in Fig. 7b have been obtained by observing the searched state without application of any r.f. pulses. The absence of any signal in the spectra confirms that there is no single quantum coherences after the search. To check for the absence of zero quantum and double quantum coherences as well, the entire density matrix has been tomographed. Fig 8a shows the theoretical and the experimental density matrices after the adiabatic evolution, when state $\vert 00\rangle$ has been searched. The mean deviation of the experimentally obtained density matrix from the theoretically predicted one (calculated using Eq. \ref{error}) is 2.49$\%$. Similarly Figs. 8b, 8c and 8d contain the theoretically predicted and experimentally obtained density matrices when the states $\vert 01\rangle$, $\vert 10\rangle$ and $\vert 11\rangle$ have been searched. The mean deviation of the experimental density matrices from their theoretically predicted counterparts are 1.92$\%$, 1.89$\%$ and 1.97$\%$ respectively. \section{5.2. Deutsch-Jozsa Algorithm} {\bf{5.2.1 \em Constant case}}\\ \noindent For the constant case (Eq. \ref{ab}), the state expected after the evolution (using the pulse sequence given in Fig. 5b) is $\vert00\rangle$. The density matrix consists of population in $\vert 00\rangle$ state and no coherences. The spectrum corresponding to the population for such a state, obtained by application of a gradient followed by $\pi$/2 pulses on each of the spins, consists of one single quantum coherence in each spin (`Population spectrum' in Fig. 9a). The spectrum for coherence, observed without application of any pulses on any of the spins, has a near absence of any signal (`Coherence spectrum' in Fig. 9a). Further confirmation of the final state is done by the tomography of the complete density matrix. The Fig. 10 shows the tomography of the experimental and theoretically predicted density matrices of the final state for the constant case. The mean deviation of the experimental density matrix from the theoretical one is 5.28$\%$ \\ \indent{\bf{5.2.2 \em Balanced case}}\\ \indent For the balanced case [Eq. 31, $\alpha$=0 and $\beta$=1], the state expected after the evolution (using the pulse sequence of Fig. 5c) is $\frac{1}{\sqrt{3}}\left(\vert01\rangle +\vert10\rangle +\vert11\rangle\right)$. The theoretical density matrix of the final state is given in Fig 11(a). This state theoretically has three diagonal elements, one SQ coherence of each qubit and a ZQ coherence between the two qubits, all of equal intensity. This state is confirmed by the spectra shown in Fig. 9b and the density matrix in Fig. 11(b). The mean deviation of the experimentally obtained density matrix from the theoretically predicted one is 17.2$\%$. It is seen that in the density matrix obtained from experiment, the SQ coherence of $^{13}$C (second qubit) and the ZQ coherence between $^{13}$C and $^1$H have significantly reduced intensity, compared to the theoretically expected values. There are three sources of error in adiabatic algorithms. $\varepsilon$ gives a measure of the first source of error. Theoretically the total time of evolution in adiabatic algorithms should be infinite. However, in practice the evolution is terminated once the state is supposed to have been reached with sufficiently high probability given by $(1-\varepsilon^2)^2$ which in our case (for $\varepsilon =10^{-2}$)is obtained to be 99.98$\%$. The second source of error is due to neglect of O($\Delta t^3$) terms in the Trotter's Formula (Eq. \ref{eq:trot}). The maximum error introduced due to this is $\approx$ 0.92 $\%$ which can be safely neglected. The third source of error is due to decoherence effects arising from the interaction of the spins with their surroundings. To study decoherence, the relaxation times T$_1$ and T$_2$ of $^{1}$H and $^{13}$C were measured. The T$_{2}$ for SQ coherences were measured by CPMG sequence. For the measurement of ZQ and DQ coherence decay rate, the term I$_{1x}$I$_{2x}$ was created and its relaxation rate was measured by CPMG sequence. The T$_{2}$ of SQ coherence of $^{1}$H was found to be 3.4 s and for $^{13}$C it was found to be 0.29 s. The decay rate of I$_{1x}$I$_{2x}$ term was found to be 0.19 s. The T$_{1}$ for $^{1}$H and $^{13}$C measured from the initial part of the inversion recovery experiment was found to be 21 s for $^1$H and 16s for $^{13}$C. Using these measured values of T$_{1}$ and T$_{2}$ the simulation for the balanced case was repeated including relaxation using Bloch's equations \cite{ernst}. Significant decay of the carbon coherences was observed. The mean deviation of the of the experimental density matrix form the theoretical density matrix including relaxation is found to be 8.0$\%$. \\ \indent The observed mean deviation between the theoretically expected and the experimentally obtained density matrices for the Grover's search and the constant case of the \dj are small ($<$ 2$\%$ and $<$ 6$\%$ respectively) while that for the balanced case of the \dj is large ($\sim$ 17$\%$). In the first two cases, the results are encoded in the diagonal elements of the density matrix, which are attenuated by the spin lattice relaxation, the times for which are large ($>$ 16 sec). On the other hand, in the balanced \dj case, there are off-diagonal elements as well which are attenuated by spin-spin relaxation, the times for which are small ($<$ 4 sec for $^1$H and $<$ 0.3 sec for $^{13}$C). The decoherence times thus have a large effect in this case. A correction for the decoherence has improved the mean deviation considerably (reduced to $\sim$ 8$\%$), confirming the succesful implementation of these algorithms. \section{6. Conclusion} In this paper we have demonstrated the experimental implementation of Grover's search and \dj algorithms by using local adiabatic evolution in a two-qubit quantum computer by nuclear magnetic resonance technique. We have suggested a different Hamiltonian for the adiabatic \dj algorithm which is diagonal in the computational basis and hence easier to implement by NMR. To the best of our knowledge this is the first experimental implementation of these two algorithms by adiabatic evolution.We believe that this work will provide impetus to solving other problems by adiabatic evolution.\\ \underline{Acknowledgment}: The authors thank K.V. Ramanathan for useful discussions. The use of DRX-500 NMR spectrometer funded by the Department of Science and Technology (DST), New Delhi, at the NMR Research Centre (formerly Sophisticated Instruments Facility), Indian Institute of Science, Bangalore, is gratefully acknowledged. AK acknowledges ``DAE-BRNS" for senior scientist support and DST for a research grant for ``Quantum Computing by NMR".
{ "timestamp": "2005-06-30T18:51:21", "yymm": "0503", "arxiv_id": "quant-ph/0503060", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503060" }
\section{Introduction} Melting of the DNA duplex is the process by which two DNA strands unbind upon heating. The nature of this transition has been studied for decades \cite{Wartell85a,Hwa97a,Breslauer99a}. For short DNA with fewer than 12--14 base pairs, melting and hybridization can be described by a two-state model as an equilibrium between single- and double-stranded DNA \cite{Crothers00a,CantorII}. For long and heterogeneous DNA, the melting curve exhibits a multi-step behavior consisting of plateaus with different sizes separated by sharp jumps. Although much of the thermodynamic properties of the melting of free DNA are known, DNA melting in a constrained space, such as on surfaces, is still poorly understood \cite {Magnasco02a}. DNA molecules functionalized with gold nanoparticles provide a model system for such study. The sequence-specific hybridization properties of DNA have been used for self-assembly of nanostructures and for highly sensitive DNA detection \cite{Mirkin96a,Kiang03a}. Previous work relies on a linker DNA \cite{Mirkin96a,Kiang03a,Kiang05a,Kiang05b}, and it has been suggested that entropic cooperativity plays an important role in the sharp phase transition of such DNA-linked nanoparticle assembly systems. On the other hand, most simulations do not explicitly incorporate linker DNA \cite{Stroud03a}, and the results cannot be directly compared to experimental data. Here we synthesized a system that eliminated the usage of a linker DNA and found that the melting transitions of these direct-linked gold particles exhibit distinct behavior from those connected via a linker DNA. \section{Experimental Procedures} Sample was prepared according to the procedures described in \cite{Kiang03a}. Briefly, DNA-capped gold nanoparticles were prepared by conjugating gold colloidal nanoparticles with thiol-modified DNA. The configuration of the DNA used in different experiments is illustrated in Fig.~\ref{fig:fig1}. We prepared four sets of samples with different DNA lengths and sequences. In sample I, the gold particles are connected through a 24-base DNA linker; in sample II, the gold particles are directly connected via 12-base complementary DNA on gold particles; in sample III, the gold particles are directly connected via 12- and 18-base DNA; in sample IV, the gold particles are directly connected via 18-base DNA. \begin{figure}[h] \begin{center} \epsfig{file=fig1.eps,height=3.0in,clip=} \end{center} \caption{DNA sequences used to form DNA-linked gold nanoparticles. Sample I is connected through a linker DNA. The line between bases A and G in the probe DNA sequences indicates that there is no chemical bond between these two bases. Sample II-IV are directly connected through surface-attached DNA with spacings of 12, 18, and 24 DNA bases between particles.} \label{fig:fig1} \end{figure} The aggregates of DNA-linked gold colloids were allowed to stand at 4 $^\circ$C for several days for aggregation. Optical spectroscopy was used to study the phase transition of the DNA-linked gold colloids, since DNA bases have strong absorption in the UV region \cite{Crothers00a,CantorII}. We monitor the thermal melting by measuring the extinction at 260 nm while slowly heating the solution containing aggregates. The solution was heated from 25 to 75 $^\circ$C at a rate of 0.5 $^\circ$C/min. All spectra were taken with a PerkinElmer Lambda 45 spectrophotometer equipped with a peltier temperature controller, magnetic stirrer, and a temperature probe. The recorded temperature of the sample was measured by a temperature probe. \section{Results and Discussion} Fig.~\ref{fig:fig2} shows the the melting curves of sample I (a) and sample II (b). The melting curves of corresponding DNA in solution are also shown. The melting temperature of DNA duplex attached to gold particle surfaces is lower than that of free DNA, and the melting transition is much sharper. However, direct comparison of melting temperatures between these two systems is difficult, since the DNA compositions are different for these two systems. The system without linker appears to have a lower melting temperature, perhaps due to the short distance between gold particles (12 base versus 24 base). \begin{figure}[tb] \begin{center} \epsfig{file=fig2.eps,width=\columnwidth,clip=} \end{center} \caption{Melting curves of gold nanoparticles connected (a) with a DNA linker (sample I), and (b) via direct hybridization of complementary surface-attached DNA (sample II). The corresponding free DNA melting curves are also shown.} \label{fig:fig2} \end{figure} To study how the melting depends on the spacing between gold particles, we prepared gold particles capped with 18 base DNA (Sample III), which is composed of identical sequence to the 12 base DNA in Sample II plus a 6 base DNA spacer (see Fig.~\ref{fig:fig1}). The difference between sample II and III is the spacer DNA length, which alter the spacing between gold particles. The melting temperature is 62 $^\circ$C for sample IV versus 32 $^\circ$C for sample II (see Fig~\ref{fig:fig3}), which suggests that increased particle spacing leads to higher melting temperature of the assemblies. \begin{figure}[tb] \begin{center} \epsfig{file=fig3.eps,height=2.0in,clip=} \end{center} \caption{Melting curves of 12/12 (sample~II), 12/18 (sample~III), 18/18 (sample~IV), and a combination of 12/12, 12/18, and 18/18 (Sample II~+~III~+~IV). The mixed system shows multi-step melting at temperatures corresponding to the $T_m$'s, within the experimental uncertainty, for Sample~II and Sample~III.} \label{fig:fig3} \end{figure} To introduce disorder, we mixed the 12 and 18 base DNA capped gold particles in one solution. Thus, the 12 base DNA are allowed to hybridize with either 12 or 18 base DNA, and the 18 base to 18 or 12 base DNA. This combination allows three possible duplex formations: 12/12 (sample I), 12/18 (sample II), and 18/18 (sample III) hybridization in one solution and possibly in one aggregate. Note that in all three base pairing only 12 bases are complementary, and the only variable is the non-pairing DNA spacer length, which controls the inter-particle distance. Since the duplexes with higher melting temperatures are more stable, we expect to see more of those duplexes forming. Indeed, Fig.~\ref{fig:fig3} shows that the most abundant duplex is the 18/18 combination, followed by 12/18, with almost no 12/12 duplex formed. The multi-step melting is an unusual phenomenon in DNA-capped gold particle assembly. For the system connected by either 24 or 30 base DNA linker, where the 30 base linker differs from the 24 base by an extra 6 base spacer in the middle of the linker, heating the assembly results in a single melting temperature $T_m$. The $T_m$ of the system with spacer is higher (37~$^\circ$C) than that without spacer (33~$^\circ$C). When equal amounts of linkers with and without spacer are present in the solution, the system has a $T_m$ in between the high $T_m$ (37~$^\circ$C) and low $T_m$ (33~$^\circ$C) systems. The $T_m$ of the mixed system (36.5~$^\circ$C) is much closer to the more stable system (37~$^\circ$C), as illustrated in Fig.~\ref{fig:fig4}a. However, free DNA with the same sequences exhibits different trend in $T_m$ (see Fig.~\ref{fig:fig4}b), where a linker with spacer results in lower melting temperature than that without spacer. The finding suggests that the inter-particle distance plays an important role in determining the $T_m $ in the nanoparticle system. On the other hand, the multi-step melting in the system without linker DNA suggests most clusters are composed of either 12/18 connections or 18/18 connectionis and few with both connected in the same cluster, unlike the systems with linker DNA. The abundance of clusters of a given connection is related to its stability. We speculate that once the cluster nucleate with a certain type of connection (defined here by the DNA length, hence the interparticle spacing) only the same type of connection is allowed to grow. Further studies are needed to determine whether this phenomenon is kinetics or thermodynamics driven. \begin{figure}[tb] \begin{center} \epsfig{file=fig4.eps,width=\columnwidth,clip=} \end{center} \caption{Melting curves of DNA duplex containing spacer for (a) nanoparticle assembly, and (b) free DNA.} \label{fig:fig4} \end{figure} \section{Summary} In summary, we have studied the thermal denaturation of DNA strands attached to gold nanoparticle surfaces. In the DNA-capped gold nanoparticle systems, the interactions are complex, involving DNA-DNA interactions and particle-particle interactions. The DNA are constrained to a gold particle surface and often exhibit interesting behaviors not seen by DNA in free solution. The multi-step melting of systems with different spacers is unique to the systems directly linked with DNA that are attached to gold nanoparticles. $^*$To whom correspondence should be addressed, email: chkiang@rice.edu.
{ "timestamp": "2005-03-10T23:19:10", "yymm": "0503", "arxiv_id": "physics/0503090", "language": "en", "url": "https://arxiv.org/abs/physics/0503090" }
\section{Introduction} A well known theorem of Szemer\'edi \cite{SZ} states that every dense subset of integers contains long arithmetic progressions. A different, but somehow related result of Freiman \cite{FR} says that if the sumset of a finite set of numbers $A$ is small, i.e. $|A+A|\leq C|A|,$ then $A$ is the subset of a (not very large) generalized arithmetic progression. Balog and Szemer\'edi proved in \cite{BSZ} that a similar structural statement holds under weaker assumptions. (For correct statements and details, see \cite{NA}). As a corollary of their result, Freiman's theorem, and Szemer\'edi's theorem about $k$-term arithmetic progressions, Balog and Szemer\'edi proved Theorem 1 below. The goal of this paper is to present a simple, purely combinatorial proof of this assertion. Let $A$ be a set of numbers and $G$ be a graph such that the vertex set of $G$ is $A.$ The {\em sumset of $A$ along $G$} is \[ A+_GA = \{a+b: a,b \in A \text{ and } (a,b) \in E(G)\}. \] \begin{theorem} For every $c,K,k >0$ there is a threshold $n_0=n_0(c,K,k)$ such that if $|A|=n\geq n_0$, $|A+_GA|\leq K|A|$, and $|E(G)|\geq cn^2$, then $A$ contains a $k$-term arithmetic progression. \end{theorem} \section{Lines and hyperplanes} There are arrangements of $n$ lines on the Euclidean plane such that the maximum number of points incident with at least three lines is ${n^2\over 6}.$ Not much is known about the structure of arrangements where the number of such points is close to the maximum, say $cn^2$, where $c$ is a positive constant. Nevertheless, the following is true. \begin{lemma} \label{lemma:line} For every $c>0$ there is a threshold $n_0=n_0(c)$ and a positive $\delta =\delta (c)$ such that, for any set of $n\geq n_0$ lines $L$ and any set of $m\geq cn^2$ points $P$, if every point is incident to three lines, then there are at least $\delta n^3$ triangles in the arrangement. (A triangle is a set of three distinct points from $P$ such that any two are incident to a line from $L.$) \end{lemma} {\bf Proof}.\ This lemma follows from the following theorem of Ruzsa and Szemer\'edi \cite{RSZ}. \begin{theorem} \cite{RSZ} Let $G$ be a graph on $n$ vertices. If $G$ is the union of $cn^2$ edge-disjoint triangles, then $G$ contains at least $\delta n^3$ triangles, where $\delta$ depends on $c$ only. \end{theorem} To prove Lemma 1, let us construct a graph where $L$ is the vertex set, and two vertices are adjacent if and only if the corresponding lines cross at a point of $P$. This graph is the union of $cn^2$ disjoint triangles, every point of $P$ defines a unique triangle, so we can apply Theorem 2.$\square$\vspace{.8cm} The result above suffices to prove Theorem 1 for 3-term arithmetic progressions. But for larger values of $k$, we need a generalization of Lemma 1. \begin{lemma} \label{lemma:plane} For every $c>0$ and $d\geq 2$, there is a threshold $n_0=n_0(c,d)$ and a positive $\delta =\delta (c,d)$ such that, for any set of $n\geq n_0$ hyperplanes $L$ and any set of $m\geq cn^d$ points $P$, if every point is incident to $d+1$ hyperplanes, then there are at least $\delta n^{d+1}$ simplices in the arrangement. (A simplex is a set of $d+1$ distinct points from $P$ such that any $d$ are incident to a hyperplane from $L.$) \end{lemma} Lemma 2 follows from the Frankl-R\"odl conjecture \cite{FRR}, the generalization of Theorem 2. The $d=3$ case was proved in \cite{FRR} and the conjecture has been proved recently by Gowers \cite{GO} and independently by Nagle, R\"odl, Schacht, and Skokan \cite{NRS},\cite{RS}. For details, how Lemma 2 follows from the Frankl-R\"odl conjecture, see \cite{SO}. \section{The $k=3$ case} Let $A$ be a set of numbers and $G$ be a graph such that the vertex set of $G$ is $A.$ We define the {\em difference-set of $A$ along $G$} as \[ A-_GA = \{a-b: a,b \in A \text{ and } (a,b) \in E(G)\}. \] \begin{lemma} For every $\epsilon ,c,K >0$ there is a number $D=D(\epsilon ,c,K)$ such that if $|A+_GA|\leq K|A|$ and $|E(G)|\geq c|A|^2$, then there is a graph $G'\subset G$ such that $|E(G')|\geq (1-\epsilon)|E(G)|$ and $|A-_{G'}A|\leq D|A|$. \end{lemma} {\bf Proof}.\ Let us consider the arrangement of points given by a subset of the Cartesian product $A\times A$ and the lines $y=a$, $x=a$ for every $a\in A$, and $x+y=t$ for every $t\in A+_GA.$ The pointset $P$ is defined by $(a,b)\in P$ iff $(a,b)\in E(G).$ By Lemma 1, the number of triangles in this arrangement is $\delta n^3.$ The triangles here are right isosceles triangles. We say that a point in $P$ is {\em popular} if the point is the right-angle vertex of at least $\alpha n$ triangles. Selecting $\alpha={{\delta (\epsilon c)}\over {\epsilon c}}$, where $\delta (\cdot)$ is the function from Lemma 1, all but at most $\epsilon cn^2$ points of $P$ are popular. A $t\in A-A$ is {\em popular} if $|\{(a,b):a-b=t; a,b\in A\}|\geq \alpha n.$ The number of popular $t$s is at most $Dn$, where $D$ depends on $\alpha$ only. $A\times A$ is a Cartesian product, therefore every triangle can be extended to a square adding one extra point from $A\times A$. Every popular point $p$ is the right-angle vertex of at least $\alpha n$ triangles. Therefore $p$ is incident to a line $x-y=t$, where $t$ is popular, because this line contains at least $\alpha n$ ``fourth" vertices of squares with $p$. $\square$\vspace{.8cm} {\bf Proof of Theorem 1, case $k=3.$} Let us apply Lemma 1 to the pointset $P'$ defined by $(a,b)\in P'$ iff $(a,b)\in E(G')$ and the lines are $y=a$ for every $a\in A$, $x-y=t$ for every $t\in A-_{G'}A$, and $x+y=s$ for every $s\in A+_GA.$ By Theorem 2, if $|A|$ is large enough, then there are triangles in the arrangement. The vertices of such triangles are vertices from $P'\subset A\times A.$ The vertical lines through the vertices form a 3-term arithmetic progression and therefore $A$ contains $\delta n^2$ 3-term arithmetic progressions, where $\delta > 0$ depends on $c$ only. $\square$\vspace{.8cm} \section{The general, $k>3$, case} Following the steps of the proof for $k=3$, we prove the general case by induction on $k.$ We prove the following theorem, which was conjectured by Erd\H os and proved by Balog and Szemer\'edi in \cite{BSZ}. Theorem 3, together with the $k=3$ case, gives a proof of Theorem 1. \begin{theorem} For every $c>0$ and $k>3$ there is an $n_0$ such that, if $A$ contains at least $c|A|^2$ 3-term arithmetic progressions and $|A|\geq n_0$, then $A$ contains a $k$-term arithmetic progression. \end{theorem} Instead of triangles, we must consider simplices. Set $k=d$. In the $d$-dimensional space we show that $A\times \cdots \times A$, the $d$-fold Cartesian product of $A$, contains a simplex in which the vertices' first coordinates form a $(d+1)$-term arithmetic progression. The simplices we are looking for are homothetic\footnote{Here we say that two simplices are homothetic if the corresponding facets are parallel.} images of the simplex $S_d$ whose vertices are listed below: \begin{displaymath} \begin{array}{c} (0, 0, 0,0, \ldots ,0,0)\\ (1, 1,0,0, \ldots ,0,0)\\ (2, 0, 1,0, \ldots ,0,0)\\ (3, 0,0,1, \ldots ,0,0)\\ \vdots\\ (d-1, 0, \ldots ,1,0)\\ (d, 0,0,0, \ldots ,0,0). \end{array} \end{displaymath} \indent An important property of $S_d$ is that its facets can be pushed into a ``shorter" grid. The facets of $S_d$ are parallel to hyperplanes, defined by the origin $(0,0,0,0,\ldots ,0,0)$, and some $(d-1)$-tuples of the grid $$\{0,1,2,\ldots ,d-1\}\times \{-1,0,1\}\times\{0,1\}^{d-2}.$$ For example, if $d=3$, then the facets are \begin{displaymath} \begin{array}{c} \{(0,0,0),(1,1,0),(2,0,1)\}\\ \{(0,0,0),(1,1,0),(3,0,0)\}\\ \{(0,0,0),(2,0,1),(3,0,0)\}\\ \{(1,1,0),(2,0,1),(3,0,0)\}, \end{array} \end{displaymath} and the corresponding parallel planes in $$\{0,1,2\}\times \{-1,0,1\}\times\{0,1\}$$ are the planes incident to the triples \begin{displaymath} \begin{array}{c} \{(0,0,0),(1,1,0),(2,0,1)\}\\ \{(0,0,0),(1,1,0),(2,0,0)\}\\ \{(0,0,0),(2,0,1),(2,0,0)\}\\ \{(0,0,0),(1,-1,1),(2,-1,0)\}. \end{array} \end{displaymath} \indent In general, if a facet of $S_d$ contains the origin and the ``last point" $(d, 0,0,0, \ldots ,0,0),$ then if we replace the later one by $(d-1, 0,0,0, \ldots ,0,0)$, the new $d$-tuples define the same hyperplane. The remaining facet $f$, given by \begin{displaymath} \begin{array}{c} (1, 1,0,0, \ldots ,0,0)\\ (2, 0, 1,0, \ldots ,0,0)\\ (3, 0,0,1, \ldots ,0,0)\\ \vdots\\ (d-1, 0, \ldots ,1,0)\\ (d, 0,0,0, \ldots ,0,0), \end{array} \end{displaymath} is parallel to the hyperplane through the vertices of $f-(1,1,0,0, \ldots ,0,0),$ \begin{displaymath} \begin{array}{c} (0, 0,0,0, \ldots ,0,0)\\ (1, -1, 1,0, \ldots ,0,0)\\ (2, -1,0,1, \ldots ,0,0)\\ \vdots\\ (d-2, -1, \ldots ,1,0)\\ (d-1,-1,0,0, \ldots ,0,0). \end{array} \end{displaymath} \indent In a homothetic copy of the grid $$T_d=\{0,1,2,\ldots ,d-1\}\times \{-1,0,1\}\times\{0,1\}^{d-2},$$ the image of the origin is called the {\em holder} of the grid. As the induction hypothesis, let us suppose that Theorem 3 is true for a $k\geq 3$ in a stronger form, providing that the number of $k$-term arithmetic progressions in $A$ is at least $c|A|^2.$ Then the number of distinct homothetic copies of $T_d$ in $\mathbb{A}_d=\underbrace{A\times \ldots \times A}_d$ is at least $c'|A|^{d+1}$ ($c'$ depends on $c$ only). Let us say that a point $p\in \mathbb{A}_d$ is \emph{popular} if $p$ is the holder of at least $\alpha |A|$ grids. If $p$ is popular, then for any facet of $S_d$, $f$, the point $p$ is the element of at least $\alpha |A|$ $d$-tuples, similar and parallel to $f.$ If $\alpha$ is small enough, then at least $\gamma |A|^d$ points of $\mathbb{A}_d$ are popular, where $\gamma$ depends on $c$ and $\alpha$ only. A hyperplane $H$ is \emph{$\beta$-rich} if it is incident to many points, $|H\cap \mathbb{A}_d|\geq \beta |A|^{d-1}.$ For every facet of $S_d$, $f$, let us denote the set of $\beta$-rich hyperplanes which are parallel to $f$ by $\mathcal{H}_f.$ \begin{lemma} For some choice of $\beta$, at least half of the popular points are incident to $d+1$ $\beta$-rich hyperplanes, parallel to the facets of $S_d.$ \end{lemma} Suppose to the contrary that for a facet $f$, more than ${\gamma\over 2d} |A|^d$ popular points are not incident to hyperplanes of $\mathcal{H}_f.$ Then more than \begin{equation} \alpha |A|{\gamma\over 2d} |A|^d={\gamma\alpha\over 2d} |A|^{d+1} \end{equation} \noindent $d$-tuples, similar and parallel to $f$, are not covered by $\mathcal{H}_f.$ Let us denote the hyperplanes incident to the ``uncovered" $d$-tuples by $L_1,L_2,\ldots ,L_m$, and the number of points on the hyperplanes by $\mathcal{L}_1,\mathcal{L}_2,\ldots ,\mathcal{L}_m.$ A simple result of Elekes and Erd\H os \cite{EE},\cite{EL} implies that hyperplanes with few points cannot cover many $d$-tuples. \begin{theorem} \cite{EE} The number of homothetic copies of $f$ in $L_i$ is at most $c_d\mathcal{L}_i^{1+1/(d-1)}$, where $c_d$ depends on $d$ only. \end{theorem} The inequalities $$\sum_{i=1}^m\mathcal{L}_i\leq |A|^d, \text{ and } \mathcal{L}_i\leq \beta |A|^{d-1}.$$ lead us to the proof of Lemma 4. The number of $d$-tuples covered by $L_i$s is at most $$c_d\sum_{i=1}^m\mathcal{L}_i^{1+1/(d-1)}\leq c_d{{|A|^d}\over {\beta |A|^{d-1}}}(\beta |A|^{d-1})^{1+1/(d-1)}=c_d\beta^{1/(d-1)}|A|^{d+1}.$$ If we compare this bound to (1), and choose $\beta$ such that ${\gamma\alpha\over 2d}=c_d\beta^{1/(d-1)}$, then at least half of the popular points are covered by $d+1$ $\beta$-rich hyperplanes parallel to the facets of $S_d.$ $\square$\vspace{.8cm} Finally we can apply Lemma 2 with the pointset $P$ of ``well-covered" popular points of $\mathbb{A}_d$ and with the sets of hyperplanes $L=\bigcup_{f\subset S_d}\mathcal{H}_f.$ The number of points is at least ${\gamma\alpha\over 2} |A|^{d}$. For a given $f,$ $|\mathcal{H}_f|\leq {{|A|^d}\over{\beta |A|^{d-1}}}=|A|/\beta.$ The number of hyperplanes in $L$ is at most $(d+1)|A|/\beta.$ By Lemma 2, we have at least $\delta '|A|^{d+1}$ homothetic copies of $S_d$ in $\mathbb{A}_d.$ Let us project them onto $x_1$, the first coordinate axis. Every image is a $(k+1)$-term arithmetic progression, and the multiplicity of one image is at most $|A|^{d-1}.$ Therefore there are at least $\delta '|A|^2$ $(k+1)$-term arithmetic progressions in $A.$ $\square$\vspace{.8cm} \section{$G_n=K_n$} When the full sumset $A+A$ is small then it is easier to prove that $A$ contains long arithmetic progressions. We can use the following Pl\"unecke type inequality \cite{PL,RU,NA}. \begin{theorem} Let $A$ and $B$ be finite subsets of an abelian group such that $|A|=n$ and $|A+B|\leq \delta n$. Let $k\geq 1$ and $l\geq 1.$ Then $$|kB-lB|\leq \delta^{k+l}n.$$ \end{theorem} It follows from the inequality, that for any dimension $d$ and $d$-dimensional integer vector $\vec{v}=(x_1,\ldots ,x_d), x_i\in \mathbb{Z}$, there is a $c>0$ depending on $d,\delta$ and $\vec{v}$ such that the following holds: \emph{If $|A+A|\leq \delta |A|$, then $\mathbb{A}_d$ can be covered by $c|A|$ hyperplanes with the same normal vector $\vec{v}$}. Using this, we can define our hyperplane-point arrangement, with the hyperplanes parallel to the facets of $S_d$ containing at least one point of $\mathbb{A}_d$, and the pointset of the arrangement is $\mathbb{A}_d.$ Then we do not have to deal with rich planes and popular points, and we can apply Lemma 2 directly.
{ "timestamp": "2005-03-28T23:54:29", "yymm": "0503", "arxiv_id": "math/0503649", "language": "en", "url": "https://arxiv.org/abs/math/0503649" }
\section*{Introduction} Predicting the three-dimensional structure of a protein from its amino acid sequence is an essential step toward the thorough bottom-up understanding of complex biological phenomena. Recently, much progress has been made in developing so-called \emph{ab initio} or \emph{de novo} structure prediction methods\cite{BonneauANDBaker2001}. In the standard approach to such \emph{de novo} structure predictions, a protein is represented as a physical object in three-dimensional (3D) space, and the global minimum of free energy surface is sought with a given force-field or a set of scoring functions. In the minimization process, structural features predicted from the amino acid sequence may be used as restraints to limit the conformational space to be sampled. Such structural features include so-called one-dimensional (1D) structures of proteins. Protein 1D structures are 3D structural features projected onto strings of residue-wise structural assignments along the amino acid sequence\cite{Rost2003}. For example, a string of secondary structures is a 1D structure. Other 1D structures include (solvent) accessibilities\cite{LeeANDRichards1971}, contact numbers\cite{KinjoETAL2005} and recently introduced residue-wise contact orders\cite{KinjoANDNishikawa2005}. The contact number, also referred to as coordination number or Ooi number\cite{NishikawaANDOoi1980}, of a residue is the number of contacts that the residue makes with other residues in the native 3D structure, while the residue-wise contact order of a residue is the sum of sequence separations between that residue and contacting residues. We have recently shown that it is possible to reconstruct the native 3D structure of a protein from a set of three types of native 1D structures, namely secondary structures (SS), contact numbers (CN), and residue-wise contact orders (RWCO)\cite{KinjoANDNishikawa2005}. Therefore, these 1D structures contain rich information regarding the corresponding 3D structure, and their accurate prediction may be very helpful for 3D structure prediction. In our previous study\cite{KinjoETAL2005}, we have developed a simple linear method to predict contact numbers from amino acid sequence. In that method, the use of multiple sequence alignment was shown to improve the prediction accuracy, achieving an average correlation coefficient of 0.63 between predicted and observed contact numbers per protein. There, we used amino acid frequency table obtained from the HSSP\cite{HSSP} multiple sequence alignment. In this paper, we extend the previous method by introducing a new framework called critical random networks (CRNs), and apply it to the prediction of secondary structure and residue-wise contact order in addition to contact number prediction. In this framework, a state vector of a large dimension is associated with each site of a target sequence. The state vectors are connected via random nearest-neighbor interactions. The value of the state vectors are determined by solving an equation of state. Then a 1D quantity of each site is predicted as a linear function of the state vector of the site as well as the corresponding local PSSM segment. This approach was inspired by the method of echo state networks (ESNs) which has been recently developed and successfully applied to complex time series analysis\cite{Jaeger2001,JaegerANDHaas2004}. Unlike ESNs which treat infinite series of input signals in one direction (from the past to the future), CRNs treat finite systems incorporating both up- and downstream information at the same time. Also, the so-called echo state property is not imposed to a network, but the system is instead set at a critical point of the network. As the input to CRNs-based prediction, we employ position-specific scoring matrices (PSSMs) generated by PSI-BLAST\cite{AltschulETAL1997}. By the combination of PSSMs and CRNs, accurate prediction of SS, CN and RWCO have been achieved. Currently, almost all the accurate methods for one-dimensional structure predictions combine some kind of sophisticated machine-learning approaches such as neural networks and support vector machines with PSSMs. The method presented here is no exception. This trend raises a question as to what extent the machine-learning approaches are effective. In this study, we address this question by comparing the CRNs-based method with a purely linear method based on PSSMs. Although not so good as the CRNs-based method, the linear predictions are of surprisingly high quality. This result suggests that, although not insignificant, the effect of the machine-learning approaches is relatively of minor importance while the use of PSSMs is the most significant ingredient in one-dimensional structure prediction. The problem of how to effectively extract meaningful information from the amino acid sequence beyond that provided by PSSMs requires yet further studies. \section*{Materials and Methods} \subsection*{Definition of 1D structures} \paragraph{Secondary structures (SS)} Secondary structures were defined by the DSSP program\cite{DSSP}. For three-state SS prediction, the simple encoding scheme was employed. That is, $\alpha$ helices ($H$), $\beta$ strands ($E$), and other structures (``coils'') defined by DSSP were encoded as $H$, $E$, and $C$, respectively. For SS prediction, we introduce feature variables $(y_i^H, y_i^E, y_i^C)$ to represent each type of secondary structures at the $i$-th residue position, so that $H$ is represented as $(1,-1,-1)$, $E$ as $(-1,1,-1)$, and $C$ as $(-1,-1,1)$. \paragraph{Contact numbers (CN)} Let $C_{i,j}$ represent the contact map of a protein. Usually, the contact map is defined so that $C_{i,j} = 1$ if the $i$-th and $j$-th residues are in contact by some definition, or $C_{i,j} = 0$, otherwise. As in our previous study, we slightly modify the definition using a sigmoid function. That is, \begin{equation} C_{i,j} = 1/\{1+\exp[w(r_{i,j} - d)]\} \end{equation} where $r_{i,j}$ is the distance between $C_{\beta}$ ($C_{\alpha}$ for glycines) atoms of the $i$-th and $j$-th residues, $d = 12$\AA{} is a cutoff distance, and $w$ is a sharpness parameter of the sigmoid function which is set to 3\cite{KinjoETAL2005,KinjoANDNishikawa2005}. The rather generous cutoff length of 12\AA{} was shown to optimize the prediction accuracy\cite{KinjoETAL2005}. The use of the sigmoid function enables us to use the contact numbers in molecular dynamics simulations\cite{KinjoANDNishikawa2005}. Using the above definition of the contact map, the contact number of the $i$-th residue of a protein is defined as \begin{equation} n_i = \sum_{j:|i-j|>2}C_{i,j}. \label{eq:defcn} \end{equation} The feature variable $y_i$ for CN is defined as $y_i = n_i / \log L$ where $L$ is the sequence length of a target protein. The normalization factor $\log L$ is introduced because we have observed that the contact number averaged over a protein chain is roughly proportional to $\log L$, and thus division by this value removes the size-dependence of predicted contact numbers. \paragraph{Residue-wise contact orders (RWCO)} RWCOs were first introduced in Kinjo and Nishikawa\cite{KinjoANDNishikawa2005}. Using the same notation as contact numbers (see above), the RWCO of the $i$-th residue in a protein structure is defined by \begin{equation} o_i = \sum_{j:|i-j|>2}|i-j|C_{i,j}. \label{eq:defrwco} \end{equation} The feature variable $y_i$ for RWCO is defined as $y_i = o_i / L$ where $L$ is the sequence length. Due to the similar reason as CN, the normalization factor $L$ was introduced to remove the size-dependence of the predicted RWCOs (the RWCO averaged over a protein chain is roughly proportional to the chain length). \subsection*{Linear regression scheme} The input to the prediction scheme we develop in this paper is a position-specific scoring matrix (PSSM) of the amino acid sequence of a target protein. Let us denote the PSSM by $U = (\mathbf{u}_1, \cdots , \mathbf{u}_{L})$ where $L$ is the sequence length of the target protein and $\mathbf{u}_i$ is a 20-vector containing the scores of 20 types of amino acid residues at the $i$-th position: $\mathbf{u}_i = (u_{1,i}, \cdots , u_{20,i})^{t}$. When predicting a type of 1D structures, we first predict the feature variable(s) for that type of 1D structures [i.e., $y_i = y_i^H$, etc. for SS, $n_i/\log L$ for CN, and $o_i/L$ for RWCO], and then the feature variable is converted to the target 1D structure. Prediction of the feature variable $y_i$ can be considered as a mapping from a given PSSM $U$ to $y_i$. More formally, we are going to establish the functional form of the mapping $F$ in $\hat{y}_{i} = F(U,i)$ where $\hat{y}_{i}$ is the predicted value of the feature variable $y_i$. In our previous paper, we showed that CN can be predicted to a moderate accuracy by a simple linear regression scheme with a local sequence window\cite{KinjoETAL2005}. Accordingly, we assume that the function $F$ can be decomposed into linear ($F_l$) and nonlinear ($F_n$) parts: $F = F_{l} + F_{n}$. The linear part is expressed as \begin{equation} F_l(U,i) = \sum_{m=-M}^{M}\sum_{a=1}^{21}D_{m,a}u_{a,i+m} \label{eq:lin} \end{equation} where $M$ is the half window size of the local PSSM segment around the $i$-th residue, and $\{D_{m,a}\}$ are the weights to be trained. To treat N- and C-termini separately, we introduced the ``terminal residue'' as the 21st kind of amino acid residue. The value of $u_{21,i+m}$ is set to unity if $i+m<0$ or $i+m>L$, or to zero otherwise. The ``terminal residue'' for the central residue ($m=0$) serves as a bias term and is always set to unity. To establish the nonlinear part, we first introduce an $N$-dimensional ``state vector'' $\mathbf{x}_i = (x_{1,i}, \cdots , x_{N,i})^{t}$ for the $i$-th sequence position where the dimension $N$ is a free parameter. The value of $\mathbf{x}_i$ is determined by solving the equation of state which is described in the next subsection. For the moment, let us assume that the equation of state has been solved, and denote the solution by $\mathbf{x}_{i}^{*}$. The state vector can be considered as a function of the whole PSSM $U$ (i.e., $\mathbf{x}_{i}^{*} = \mathbf{x}_{i}^{*}(U)$), and implicitly incorporates nonlinear and long-range effects. Now, the nonlinear part $F_n$ is expressed as a linear projection of the state vector: \begin{equation} F_n(U,i) = \sum_{k=1}^{N}E_{k}x_{k,i}^{*}(U) \label{eq:nonlin} \end{equation} where $\{E_{k}\}$ are the weights to be trained. In summary, the prediction scheme is expressed as \begin{equation} \hat{y}_{i} = \sum_{m=-M}^{M}\sum_{a=1}^{21}D_{m,a}u_{a,i+m} + \sum_{k=1}^{N}E_{k}x_{k,i}^{*}(U) \label{eq:pred0} \end{equation} Regarding $\mathbf{u}_{i-M}, \cdots, \mathbf{u}_{i+M}$ and $\mathbf{x}_{i}^{*}$ as independent variables, Eq. \ref{eq:pred0} reduces to a simple linear regression problem for which the optimal weights $\{D_{m,a}\}$ and $\{E_k\}$ are readily determined by using a least squares method. For CN or RWCO predictions, the predicted feature variable can be easily converted to the corresponding 1D quantities by multiplying by $\log L$ or $L$, respectively. For SS prediction, the secondary structure $\hat{s}_i$ of the $i$-th residue is given by $\hat{s}_i = \mathrm{arg}\max_{s\in \{H, E, C\}}y_i^s$. \subsection*{Critical random networks and the equation of state} We now describe the equation of state for the system of state vectors. We denote $L$ state vectors along the amino acid sequence by $\mathbf{X} = (\mathbf{x}_{1}, \cdots , \mathbf{x}_{L}) \in \mathbf{R}^{N\times L}$, and define a nonlinear mapping $g_i : \mathbf{R}^{N\times L} \to \mathbf{R}^{N}$ for $i = 1, \cdots , L$ by \begin{equation} g_i(\mathbf{X}) = \tanh \left[\beta W(\mathbf{x}_{i-1}+\mathbf{x}_{i+1})+\alpha V \mathbf{u}_{i}\right] \end{equation} where $\beta$ and $\alpha$ are positive constants, $W$ is an $N\times N$ block-diagonal orthogonal random matrix, and $V$ is an $N\times 21$ random matrix (a unit bias term is assumed in $\mathbf{u}_i$). The hyperbolic tangent function ($\tanh$) is applied element-wise. We impose the boundary conditions as $\mathbf{x}_0 = \mathbf{x}_{L+1} = \mathbf{0}$. In this equation, the term containing $W$ represents nearest-neighbor interactions along the sequence. The amino acid sequence information is taken into account as an external field in the form of $\alpha{}V\mathbf{u}_{i}$. Next we define a mapping $G : \mathbf{R}^{N\times L} \to \mathbf{R}^{N\times L}$ by \begin{equation} G(\mathbf{X}) = (g_{1}(\mathbf{X}), \cdots , g_{L}(\mathbf{X})). \end{equation} Using this mapping $G$, the equation of state is defined as \begin{equation} \mathbf{X} = G(\mathbf{X}). \label{eq:fixpoint} \end{equation} That is, the state vectors are determined as a fixed point of the mapping $G$. More explicitly, Eq. \ref{eq:fixpoint} can be expressed as \begin{equation} \mathbf{x}_{i} = \tanh \left[\beta W(\mathbf{x}_{i-1}+\mathbf{x}_{i+1})+\alpha V \mathbf{u}_{i}\right], \label{eq:eos} \end{equation} for $i = 1, \cdots, L$. That is, the state vector $\mathbf{x}_i$ of the site $i$ is determined by the interaction with the state vectors of the neighboring sites $i-1$ and $i+1$ as well as with the `external field' $\mathbf{u}_i$ of the site. The information of the external field at each site is propagated throughout the whole amino acid sequence via the nearest-neighbor interactions. Therefore, solving Eq. (\ref{eq:eos}) means finding the state vectors that are consistent with the external field as well as the nearest-neighbor interactions, and each state vector in the obtained solution $\{\mathbf{x}_i\}$ self-consistently embodies the information of the whole amino acid sequence in a mean-field sense. For $\beta < 0.5$, it can be shown that $G$ is a contraction mapping in $\mathbf{R}^{N\times L}$ (with an appropriate norm defined therein). And hence, by the contraction mapping principle\cite{TakahashiNLFA}, the mapping $G$ has a unique fixed point independently of the strength $\alpha$ of the external field. When $\beta$ is sufficiently smaller than 0.5, the correlation between two state vectors, say $\mathbf{x}_{i}$ and $\mathbf{x}_{j}$, is expected to decay exponentially as a function of the sequential separation $|i-j|$. On the other hand, for $\beta > 0.5$, the number of the fixed points varies depending on the strength of the external field $\alpha$. In this regime, we cannot reliably solve the equation of state (Eq.\ref{eq:fixpoint}). In this sense, $\beta = 0.5$ can be considered as a critical point of the system $\mathbf{X}$. From an analogy with critical phenomena of physical systems\cite{Goldenfeld1992} (note the formal similarity of Eq. \ref{eq:eos} with the mean field equation of the Ising model), the correlation length between state vectors is expected to diverge, or become long when the external field is finite but small. We call the system defined by Eq. \ref{eq:eos} with $\beta = 0.5$ a critical random network (CRN). The equation of state (Eq. \ref{eq:eos}) is parameterized by two random matrices $W$ and $V$, and consequently, so is the predicted feature variables $\hat{y}_{i}$. Following a standard technique of statistical learning such as neural networks\cite{Haykin}, we may improve the prediction accuracy by averaging $\hat{y}_{i}$ obtained by multiple CRNs with different pairs of $W$ and $V$. This averaging operation reduces the prediction errors due to the random fluctuations in the estimated parameters. We employ such an ensemble prediction with 10 sets of random matrices $W$ and $V$ in the following. The use of a larger number of random matrices for ensemble predictions improved the prediction accuracies slightly, but the difference was insignificant. \subsection*{Numerics} Here we describe the value of the free parameters used, and a numerical procedure to solve the equation of state. The half window size $M$ in the linear part of Eq. \ref{eq:pred0} is set to 9 for SS and CN predictions, and to 26 for RWCO prediction. These values are found to be optimal in preliminary studies\cite{KinjoETAL2005, KinjoANDNishikawa2005b}. Regarding the dimension $N$ of the state vector, we have found that $N=2000$ gives the best result after some experimentation, and this value is used throughout. Using the state vector of a large dimension as 2000, it is expected that various properties of amino acid sequences can be extracted and memorized. If the dimension is too large, overfitting may occur, but we did not find such a case up to $N=2000$. Therefore, in principle, the state vector dimension could be even larger (but the computational cost becomes a problem). Each element in the $N\times 21$ random matrix $V$ in Eq. \ref{eq:eos} is obtained from a uniform distribution in the range [-1, 1] and the strength parameter $\alpha$ is set to 0.01. Here and in the following, all random numbers were generated by the Mersenne twister algorithm\cite{MersenneTwister}. The $N\times N$ random matrix $W$ is obtained in the following manner. First we generate a random block diagonal matrix $A$ whose block sizes are drawn from a uniform distribution of integers 2 to 20 (both inclusive), and the values of the block elements are drawn from the standard Gaussian distribution (zero mean and unit variance). By applying singular value decomposition, we have $A = U\Sigma V^{t}$ where $U$ and $V$ are orthogonal matrices and $\Sigma$ is a diagonal matrix of singular values. We set $W = UV^{t}$ which is orthogonal as well as block diagonal. To solve the equation of state (Eq. \ref{eq:eos}), we use a simple functional iteration with a Gauss-Seidel-like updating scheme. Let $\nu$ denote the stage of iteration. We set the initial value of the state vectors (with $\nu = 0$) as \begin{equation} \mathbf{x}_{i}^{(0)} = \tanh \left[\alpha V \mathbf{u}_{i}\right].\label{eq:init_eos} \end{equation} Then, for $i = 1, \cdots , L$ (in increasing order of $i$), we update the state vectors by \begin{equation} \mathbf{x}_{i}^{(2\nu+1)} \gets \tanh \left[W(\mathbf{x}_{i-1}^{(2\nu+1)}+\mathbf{x}_{i+1}^{(2\nu)})+\alpha V \mathbf{u}_{i}\right]. \label{eq:feos} \end{equation} Next, we update them in the reverse order. That is, for $i = L, \cdots , 1$ (in decreasing order of $i$), \begin{equation} \mathbf{x}_{i}^{(2\nu+2)} \gets \tanh \left[W(\mathbf{x}_{i-1}^{(2\nu+1)}+\mathbf{x}_{i+1}^{(2\nu+2)})+\alpha V \mathbf{u}_{i}\right]. \label{eq:beos} \end{equation} We then set $\nu \gets \nu + 1$, and iterate Eqs. (\ref{eq:feos}) and (\ref{eq:beos}) until $\{\mathbf{x}_{i}\}$ converges. The convergence criterion is \begin{equation} \sqrt{\sum_{i=1}^{L}\left\|\mathbf{x}_{i}^{(2\nu+2)}-\mathbf{x}_{i}^{(2\nu+1)}\right\|_{\mathbf{R}^{N}}^{2}/{NL}}<10^{-7} \end{equation} where $\left\|\cdot\right\|_{\mathbf{R}^{N}}$ denotes the Euclidean norm. Convergence is typically achieved within 100 to 200 iterations for one protein. \subsection*{Preparation of training and test sets} We use the same set of proteins as used in our preliminary study\cite{KinjoANDNishikawa2005b}. In this set, there are 680 protein domains selected from the ASTRAL database\cite{ASTRAL}, each of which represents a superfamily from one of all-$\alpha$, all-$\beta$, $\alpha/\beta$, $\alpha+\beta$ or ``multi-domain'' classes of the SCOP database (release 1.65, December 2003)\cite{SCOP}. Conversely, each SCOP superfamily is represented by only one of the protein domains in the data set. Thus, no pair of protein domains in the data set are expected to be homologous to each other. For training the parameters and testing the prediction accuracy, 15-fold cross-validation is employed. The set of 680 proteins is randomly divided into two groups: one consisting of 630 proteins (training set), and the other consisting of 50 proteins (test set). For each training set, the regression parameters $\{D_{m,a}\}$ and $\{E_{i}\}$ are determined, and using these parameters, the prediction accuracy is evaluated for the corresponding test set. This procedure was repeated for 15 times with different random divisions, leading to 15 pairs of training and test sets. In this way, there is some redundancy in the training and test sets although each pair of these sets share no proteins in common. But this raises no problem since our objective is to estimate the average accuracy of the predictions. A similar validation procedure was also employed by Petersen et al.\cite{PetersenETAL2000} In total, 750 ($= 15\times 50$) proteins were tested over which the averages of the measures of accuracy (see below) were calculated. \subsection*{Preparation of position-specific scoring matrix} To obtain the position-specific scoring matrix (PSSM) of a protein, we conducted ten iterations of PSI-BLAST\cite{AltschulETAL1997} search against a customized sequence database with the E-value cutoff of 0.0005\cite{TomiiANDAkiyama2004}. The sequence database was compiled from the DAD database provided by DNA Data Bank of Japan\cite{DDBJ2005}, from which redundancy was removed by the program CD-HIT\cite{CD-HIT} with 95\% identity cutoff. This database was subsequently filtered by the program PFILT used in the PSIPRED program\cite{Jones1999}. We use the position-specific scoring matrices (PSSM) rather than the frequency tables for the prediction. \subsection*{Measures of accuracy} For assessing the quality of SS predictions, we mainly use $Q_3$ and $SOV$ (the 1999 revision)\cite{SOV99}. The $Q_3$ measure quantifies the percentage of correctly predicted residues, while the $SOV$ measure evaluates the segment overlaps of secondary structural elements of predicted and native structures. Optionally, we use $Q_s$ and $Q_s^{pre}$ (with $s$ being $H$, $E$, or $C$) and Matthews' correlation coefficient $MC$. The $Q_s$ is defined by the percentage of correctly predicted SS type $s$ out of the native SS type $s$, and $Q_s^{pre}$ is defined by the percentage of correctly predicted SS type $s$ out of the predicted SS type $s$. For CN and RWCO predictions, we use two measures for evaluating the prediction accuracy. The first one is the correlation coefficient ($Cor$) between the observed ($n_{i}$) and predicted ($\hat{n}_{i}$) CN or RWCO\cite{KinjoETAL2005}. The second is the RMS error normalized by the standard deviation of the native CN or RWCO ($DevA$)\cite{KinjoETAL2005}. While $Cor$ measures the quality of relative values, $DevA$ measures that of absolute values of the predicted CN or RWCO. Note that the measures $Q_3$, $SOV$, $Cor$ and $DevA$ are defined for a single protein chain. In practice, we average these quantities over the proteins in the test sets to estimate the average accuracy of prediction. On the other hand, per-residue measures, $Q_s$, $Q_s^{pre}$ and $MC$, were calculated using all the residues in the test data sets, rather than on a per-protein basis. \section*{Results} We examine the prediction accuracies for SS, CN, and RWCO in turn. The main results are summarized in Table \ref{tab:summ} and Figure \ref{fig:histo}. Finally, in order to examine the effect of nonlinear terms, we verify the prediction results obtained using only linear terms (Eq. \ref{eq:lin}). \begin{table} \caption{\label{tab:summ}Summary of average prediction accuracies.} \begin{center} \begin{tabular}[h]{ll}\hline Struct. & Accuracy \\\hline SS & $Q_3$ = 77.8; $SOV$ = 77.3\\ CN & $Cor$ = 0.726; $DevA$ = 0.707\\ RWCO& $Cor$ = 0.601; $DevA$ = 0.881\\\hline \end{tabular} \end{center} \end{table} \begin{figure}[htb] \begin{center} \includegraphics[width=7cm]{./histos.eps} \end{center} \caption{\label{fig:histo}Histograms of accuracy measure obtained by ensemble predictions using 10 critical random networks. (a) $Q_3$ for secondary structure prediction; (b) $Cor$ for contact number prediction; (c) $Cor$ for residue-wise contact order prediction.} \end{figure} \subsection*{Secondary structure prediction} The average accuracy of secondary structure prediction achieved by the ensemble CRNs-based approach is $Q_3=77.8$\% and $SOV=77.3$ (Table \ref{tab:summ}). This is comparable to the current state-of-the-art predictors such as PSIPRED\cite{Jones1999}. The results in terms of per-residue accuracies ($Q_s$ and $Q_s^{pre}$) are listed in Table \ref{tab:ss}. The values of $Q_s$ suggest that the present method underestimates $\alpha$ helices ($H$) and, especially, $\beta$ strands ($E$) compared to coils $C$. However, when a residue is predicted as being $H$ or $E$, the probability of the correct prediction is rather high, especially for $E$ ($Q_E^{pre} =$ 79.9\%). The histogram of $Q_3$ (Figure \ref{fig:histo}a) shows that the peak of the histogram resides well beyond $Q_3$ = 80\%, and that only 20\% of the predictions exhibit $Q_3$ of less than 70\%. These observations demonstrate the capability of the CRNs-based prediction schemes. \begin{table} \caption{\label{tab:ss}Summary of per-residue accuracies for SS predictions.} \begin{center} \begin{tabular}[h]{lrrr}\hline measure & $H$ & $E$ & $C$ \\\hline $Q_s$ & 78.4 & 61.9 & 84.6 \\ $Q_s^{pre}$ & 81.9 & 79.9 & 74.3\\ $MC$ & 0.704 & 0.636 & 0.602 \\\hline \end{tabular} \end{center} \end{table} \subsection*{Contact number prediction} Using an ensemble of CRNs, a correlation coefficient ($Cor$) of 0.726 and normalized RMS error ($DevA$) of 0.707 was achieved for CN predictions on average (Table \ref{tab:summ}). This result is a significant improvement over the previous method\cite{KinjoETAL2005} which yielded $Cor=0.627$ and $DevA = 0.941$. The median of the distribution of $Cor$ (Figure \ref{fig:histo}b) is 0.744, indicating that the majority of the predictions are of very high accuracy. We have also examined the dependence of prediction accuracy on the structural class of target proteins (Table \ref{tab:cnhisto}). Among all the structural classes, $\alpha/\beta$ proteins are predicted most accurately with $Cor=$ 0.757 and $DevA =$ 0.668. The accuracy for other classes do not differ qualitatively although all-$\beta$ proteins are predicted slightly less accurately. \begin{table} \caption{\label{tab:cnhisto}Summary of CN predictions for each SCOP class$^a$.} \begin{center} \begin{tabular}{lrrrrr}\hline range$^b$ &\multicolumn{5}{c}{SCOP class$^c$}\\ ($Cor$) & a & b & c & d & e\\\hline (-1,0.5] & 8 & 6 & 3 & 14 & 1 \\ (0.5,0.6] & 19 & 25 & 8 & 19 & 1 \\ (0.6,0.7] & 29 & 29 & 22 & 54 & 3 \\ (0.7,0.8] & 62 & 66 & 76 & 85 & 10 \\ (0.8,0.9] & 43 & 38 & 57 & 67 & 3 \\ (0.9,1.0] & 1 & 0 & 0 & 1 & 0 \\ total & 162 & 164 & 166 & 240 & 18 \\\hline average $Cor$ & 0.721 & 0.712 & 0.757 & 0.728 & 0.722\\ average $DevA$ & 0.715 & 0.726 & 0.668 & 0.717 & 0.705\\ \hline \end{tabular} \end{center} $^a$ The number of occurrences of $Cor$ for the proteins in the test sets, classified according to the SCOP database; average values of $Cor$ and $DevA$ are also listed for each class.\\ $^b$ The range ``$(x,y]$'' denotes $x < Cor \leq y$.\\ $^c$ a: all-$\alpha$; b: all-$\beta$; c: $\alpha / \beta$; d: $\alpha + \beta$; e: multi-domain. \end{table} \subsection*{Residue-wise contact order prediction} For RWCO prediction, the average accuracy was such that $Cor$ = 0.601 and $DevA$ = 0.881. Although these figures appear to be poor compared to those of the CN prediction described above, they are yet statistically significant. The distribution of $Cor$ appears to be rather dispersed (Figure \ref{fig:histo}c), indicating that the prediction accuracy strongly depends on the characteristics of each target protein. In a similar manner as for CN, we also examined the dependence of prediction accuracy on the structural class of target proteins (Table \ref{tab:rwcohisto}). In this case, we have found a notable dependence of prediction accuracy on structural classes. The best accuracy is obtained for $\alpha+\beta$ proteins with $Cor = $ 0.629 and $DevA = $ 0.832. For these proteins, the distribution of $Cor$ also shows good tendency in that the fraction of poor predictions is relatively small (e.g., 14\% for $Cor <$ 0.5). Interestingly, all-$\beta$ proteins also show good accuracies but all-$\alpha$ proteins are particularly poorly predicted. These observations suggest that the correlation between amino acid sequence and RWCO is strongly dependent on the structural class of the target protein. However, the rather dispersed distribution of $Cor$ for each class (Table \ref{tab:rwcohisto}) also suggests that there are more detailed effects of the global context on the accuracy of RWCO prediction. \begin{table} \caption{\label{tab:rwcohisto}Summary of RWCO predictions for each SCOP class$^a$} \begin{center} \begin{tabular}{lrrrrr}\hline range &\multicolumn{5}{c}{SCOP class}\\ ($Cor$) & a & b & c & d & e\\\hline (-1,0.5] & 58 & 31 & 46 & 34 & 6 \\ (0.5,0.6] & 29 & 37 & 31 & 56 & 4 \\ (0.6,0.7] & 41 & 27 & 33 & 65 & 5 \\ (0.7,0.8] & 24 & 47 & 40 & 72 & 3 \\ (0.8,0.9] & 10 & 22 & 16 & 13 & 0 \\ total & 162 & 164 & 166 & 240 & 18 \\\hline average $Cor$ & 0.549 & 0.620 & 0.595 & 0.629 & 0.564\\ average $DevA$ & 0.981 & 0.869 & 0.857 & 0.832 & 0.957\\ \hline \end{tabular} \end{center} $^a$See Table \ref{tab:cnhisto} for notations. \end{table} \subsection*{Purely linear predictions with PSSMs} Almost all the modern methods for 1D structure prediction make use of PSSMs in combination with some kind of machine-learning techniques such as feed-forward or recurrent neural networks or support vector machines. The present study is no exception. Curiously, machine-learning approaches have become so widespread that no attempt appears to have been made to test simplest linear predictors based on PSSMs. In this subsection, we present results of 1D predictions using only the linear terms (Eq. \ref{eq:lin}) but without CRNs. In this prediction scheme, input is a local segment of a PSSM generated by PSI-BLAST, and a feature variable is predicted by a straight forward linear regression. As can be clearly seen in Table \ref{tab:lin}, the results of the linear predictions are surprisingly good although not as good as with CRNs. For example, in SS prediction, the purely linear scheme achieved $Q_3$ = 75.2\% which is lower than that of the CRNs-based scheme by only 3.6\%. Although this is of course a large difference in a statistical sense, there may not be a discernible difference when individual predictions are concerned. (However, the improvement in the $SOV$ measure by using CRNs is quite large, indicating that the nonlinear terms in CRNs are indeed able to extract cooperative features.) It is widely accepted that the upper limit of accuracy ($Q_3$) of SS prediction based on a local window of a single sequence is less than 70\%\cite{CrooksANDBrenner2004}. Therefore, more than 5\% of the increase in $Q_3$ is brought simply by the use of PSSMs. Similar observations also hold for CN and RWCO predictions (Table \ref{tab:lin}). In case of CN prediction, we have previously obtained $Cor$ = 0.555 by a simple linear method with single sequences\cite{KinjoETAL2005}. Therefore, the effect of PSSMs is even more dramatic than SS prediction. This may be due to the fact that the most conspicuous feature of amino acid sequences conserved among distant homologs (as detected by PSI-BLAST) is the hydrophobicity of amino acid residues\cite{KinjoANDNishikawa2004}, which is closely related to contact numbers. Of course, the improvement by the use of PSSMs is largely made possible by the recent increase of amino acid sequence databases\cite{PrzybylskiANDRost2002}. \begin{table} \caption{\label{tab:lin}Summary of prediction accuracies using only linear terms.} \begin{center} \begin{tabular}[h]{ll}\hline Struct. & Accuracy \\\hline SS & $Q_3$ = 75.2; $SOV$ = 72.7\\ CN & $Cor$ = 0.701; $DevA$ = 0.735\\ RWCO& $Cor$ = 0.584; $DevA$ = 0.902\\\hline \end{tabular} \end{center} \end{table} \subsection*{The significance of criticality} The condition of criticality ($\beta = 0.5$ in Eq. \ref{eq:eos}) is expected to enhance the extraction of the long-range correlations of an amino acid sequence, thus improving the prediction accuracy. To confirm this point, we tested the method by setting $\beta = 0.1$ so that the network of state vectors is not at the critical point any more (otherwise the prediction and validation schemes were the same as above). The prediction accuracies obtained by these non-critical random networks were $Q_3 = 76.7$\% and $SOV = 76.6$ for SS, $Cor = 0.716$ and $DevA = 0.719$ for CN, and $Cor = 0.589$ and $DevA = 0.897$ for RWCO. These values are inferior to those obtained by the critical random networks (Table \ref{tab:summ}), although slightly better than the purely linear predictions (Table \ref{tab:lin}). Therefore, compared to the non-critical random networks, the critical random networks can indeed extract more information from amino acid sequence and improve the prediction accuracies. \section*{Discussion} \subsection*{Comparison with other methods} Regarding the framework of 1D structure prediction, the critical random networks are most closely related to bidirectional recurrent neural networks (BRNNs)\cite{BaldiETAL1999}, in that both can treat a whole amino acid sequence rather than only a local window segment. The main differences are the following. First, network weights between input and hidden layers as well as those between hidden units are trained in BRNNs, whereas the corresponding weights in CRNs (random matrices $V$ and $W$, respectively, in Eq. \ref{eq:eos}) are fixed. Second, the output layer is nonlinear in BRNNs but linear in CRNs. Third, the network components that propagate sequence information from N-terminus to C-terminus are decoupled from those in the opposite direction in BRNNs, but they are coupled in CRNs. Regarding the accuracy of SS prediction, BRNNs\cite{PollastriETAL2002b} and CRNs exhibit comparable results of $Q_3 \approx$ 78\%. However, a standard local window-based approach using feed-forward neural networks can also achieve this level of accuracy\cite{Jones1999}. Thus, the CRNs-based method is not a single best predictor, but may serve as an addition to consensus predictions. Although BRNNs have been also applied to CN prediction\cite{PollastriETAL2002}, contact numbers are predicted as 2-state categorical data (buried or exposed) so that the results cannot be directly compared. Nevertheless, we can convert CRNs-based real-value predictions into 2-state predictions. By using the same thresholds for the 2-state discretization as Pollastri et al.\cite{PollastriETAL2002} (i.e., the average CN for each residue type), we obtained $Q_2 =$ 75.6\% per chain (75.1\% per residue), and Matthews' correlation coefficient $MC =$ 0.503 whereas those obtained by BRNNs are $Q_2 =$ 73.9\% (per residue) and $MC =$ 0.478. Therefore, for 2-state CN prediction, the present method yields more accurate results. Since the present study is the very first attempt to predict RWCOs, there are no alternative methods to compare with. However, the comparison of CRNs-based methods for SS and CN predictions with other methods suggests that the accuracy of the RWCO prediction presented here may be the best possible result using any of the statistical learning methods currently available for 1D structure predictions. \subsection*{Possibilities for further improvements} In the present study, we employed the simplest possible architecture for CRNs in which different sites are connected via nearest-neighbor interactions. A number of possibilities exist for the elaboration of the architecture. For example, we may introduce short-cuts between distant sites to treat non-local interactions more directly. Since the prediction accuracies depend on the structural context of target proteins (Tables \ref{tab:cnhisto} and \ref{tab:rwcohisto}), it may be also useful to include more global features of amino acid sequences such as the bias of amino acid composition or the average of PSSM components. These possibilities are to be pursued in future studies. \section*{Conclusion} We have developed a novel method, CRNs-based regression, for predicting 1D protein structures from amino acid sequence. When combined with position-specific scoring matrices produced by PSI-BLAST, this method yields SS predictions as accurate as the best current predictors, CN predictions far better than previous methods, and RWCO predictions significantly correlated with observed values. We also examined the effect of PSSMs on prediction accuracy, and showed that most improvement is brought by the use of PSSMs although the further improvement due to the CRNs-based method is also significant. In order to achieve a qualitatively yet better predictions, however, it seems necessary to take into account other, more global, information than is provided by PSSMs. \section*{Acknowledgments} The authors thank Motonori Ota for critical comments on an early version of the manuscript, and Kentaro Tomii for the advice on the use of PSI-BLAST. Most of the computations were carried out at the supercomputing facility of National Institute of Genetics, Japan. This work was supported in part by a grant-in-aid from the MEXT, Japan. The source code of the programs for the CRNs-based prediction as well as the lists of protein domains used in this study are available at \verb|http://maccl01.genes.nig.ac.jp/~akinjo/crnpred_suppl/|.
{ "timestamp": "2005-10-20T10:02:43", "yymm": "0503", "arxiv_id": "q-bio/0503032", "language": "en", "url": "https://arxiv.org/abs/q-bio/0503032" }
\section{Introduction} Inevitably, the entanglement of a multi-partite quantum state becomes degraded with time due to experimental and environmental noise. The influence of noise on bipartite entanglement is a problem in the theory of open systems \cite{Yu-Eberly06B}, as well as of practical importance in any application using quantum features of information \cite{CM}. The topic of evolution of quantum coherence in the presence of noise sits between two well-investigated problems. One of these is the relaxation toward steady-state of one-body coherence of a simple quantum system (spin, atom, exciton, quantum dot, etc.) in contact with a much larger reservoir \cite{Slichter78}. The other is the newer two-body problem where the evolution of the disentanglement of the system from its environment is of interest. It is generally understood that the latter decoherence occurs much more rapidly than the former. Recently, a practical problem that includes parts of both has drawn attention -- the survival of the joint entanglement of two small systems with each other while each is exposed to a local noisy environment. Their rapid disentanglement from their environments is supposed not to be observed, but their disentanglement from each other is considered interesting and potentially important. We have shown in a specific instance of such bipartite disentanglement of qubits \cite{Yu-Eberly02,Yu-Eberly03} that entanglement is lost in a very different way compared to the usual one-body decoherence measured by the decay of off-diagonal elements of the density matrix of either qubit system separately. More surprisingly, we have shown \cite{Yu-Eberly04,Yu-Eberly05} that the presence of either pure vacuum noise or even classical noise can cause entanglement to decay to zero in a finite time, an effect that is labelled ``entanglement sudden death'' (ESD). In the last few years the issue of such entanglement decoherence has been discussed in a number of distinct contexts such as qubit pairs \cite{Yu-Eberly02, Yu-Eberly03, Yu-Eberly04, Lucamarini-etal04, Jakobczyk-Jamroz04, Tolkunov-etal05, Ban06, Ban-Shibata06,Malinovsky06, Glendinning-etal, An-etal06, Liang06, Roszak-Machnikowski06, Jamroz06, Ficek-Tanas06, Solenov05}, finite spin chains \cite{Pratt-Eberly01, Kamta-Starace02, WangJ-etal05, Khveshchenko03, Grigorenko-Khveshchenko05, Pratt04a}, multipartite systems \cite{Carvalho-etal04, Carvalho-etal05}, decoherence dynamics in adiabatic entanglement \cite{Sun01}, entanglement transfer \cite{Lamata-etal06}, and open quantum systems \cite{Diosi03, Dodd-Halliwell04, Dodd04,Zyczkowski02}, to name a few. In addition, a proposal for the direct measurement of finite-time disentanglement in a cavity QED context has been made recently \cite{Santos-etal06}. In this paper we report several steps that we expect will assist further understanding of this complex and fundamental topic. We focus on the smallest and simplest non-trivial situations, in order to help expose consequences that are dynamically fundamental, as opposed to ones originating simply in one or another kind of complexity. For greatest utility, this more or less mandates that results should be analytic rather than numeric. We will treat two qubits prepared in a mixed state as an information time-evolution question in the presence of noises. For this purpose, solutions of the appropriate Born-Markov-Lindblad master equations can be obtained \cite{BMLexamples} and we will use a Kraus operator approach throughout the paper \cite{Kraus}. The focus will be maintained strictly on the way information itself evolves by considering the entanglement of two quantum systems $A$ and $B$ exposed to local noises but completely isolated from interacting with each other. We will examine evolution toward complete disentanglement in a class of commonly occurring bipartite density matrices (which we call ``X'' states) and establish: (a) that X-state character is robust, i.e., an X state remains an X state in its evolution under the most common noise influences, (b) as Werner states are a subclass of X states, we will show that there exisits a critical Werner fidelity below which termination of entanglement must occur in a finite time, and (c) that there are purely local operations that can sometimes be used to alter the survival dynamics of bipartite entanglement. We show that ESD will customarily occur, but that in some cases it can be avoided by applying appropriate local operations initially. We will illustrate all of these in the following sections. The paper is organized as follows: In Sec. \ref{modelsection}, we present two-qubit models where the bipartite system is coupled to external sources of phase-damping and amplitude-damping noises. Explicit time-dependent solutions in terms of Kraus operators are given. Sec. \ref{concurrence/standard} deals with concurrence, the chosen measure of entanglement, and the defining character of an X state. In Sec. \ref{decoherence}, the evolution of a Werner state toward decoherence is discussed, as an important example of X-state behavior under the influence of noise. We find a new fidelity boundary below which entanglement sudden death (ESD) must occur for all Werner states. In the following Sec. \ref{depolar} we derive the ESD that is encountered with depolarizing noise for the X states. In Sec. \ref{fragile}, we show that in some cases it is feasible to transform a short-lived state into a long-lived state by applying specified local operations initially, and we conclude in Sec. \ref{conc}. \section{Models} \label{modelsection} \noindent The non-interacting quantum systems $A$ and $B$ and their separate reservoirs labeled $a$ and $b$ are assumed to follow the same evolution route separately. We use the familiar Hamiltonian (for qubit $A$ say): \begin{equation}} \newcommand{\eeq}{\end{equation} \label{model} H^A_{\rm tot}= H^A_{\rm at} + H^a_{\rm res} + H^{\rm Aa}_{\rm int}, \eeq where: \begin{equation}} \newcommand{\eeq}{\end{equation} H^A_{\rm at} = \frac{1}{2}\om_A \si^A_z \quad {\rm and} \quad H^a_{\rm res} = \sum_{{k}}\om_{k}a_{ k}^\dag a_{ k} \label{eq1} \eeq and for exposure to phase and amplitude noises the interaction Hamiltonians $H_{\rm int}$ are given by \begin{equation}} \newcommand{\eeq}{\end{equation} H^{Aa}_{\rm ph-int} = \sum_{{k}} \si^A_z (f_{ k}a^\dag_{k} + f_{ k}^*a_{ k}), \label{eq3} \eeq and \begin{equation}} \newcommand{\eeq}{\end{equation} \label{intam} H^{Aa}_{\rm am-int} = \sum_{{ k}} ( g_{ k} \si^A_- a^\dag_{k} + g_{k}^* \sigma^A_+a_{ k}). \eeq Here the $a_{ k}$ are bosonic reservoir coordinates satisfying $[a_{ k}, a_{ k'}]=\delta_{ k, k'},$ the $g_{ k}$ are broadband coupling constants, and the $\si^A$s denote the usual Pauli matrices for qubit $A$. The same forms hold for $B$, with a set of bosonic reservoir coordinates $b_k$. $\{A,a\}$ and $\{B, b\}$ are completely independent, and no decoherence-free joint subspaces are available. As remarked, these total Hamiltonians provide well-known solvable qubit-reservoir interactions, but we are interested in the evolution of joint information as a consequence of the completely separate interactions. We consider qubits $A$ and $B$ prepared in a mixed state. For this purpose, solutions of the appropriate Born-Markov-Lindblad equations can be obtained via several routes, and we find the Kraus operator form \cite{Kraus} convenient for our purpose. Given an initial state $\rho$ (pure or mixed) for two qubits $A$ and $B$, its evolution can be written compactly as \begin{equation}} \newcommand{\eeq}{\end{equation} \label{Kraus} \rho(t) = \sum_\mu K_\mu(t)\rho(0) K_\mu^\dag(t), \eeq where the so-called Kraus operators $K_\mu$ satisfy $\sum_\mu K_\mu^\dag K_\mu = 1$ for all $t$. Obviously, the Kraus operators contain the complete information about the system's dynamics. In the case of dephasing noise one has following compact Kraus operators: \begin{eqnarray} \label{k1}E_1 &=&\left(\begin{array}{clcr} \gamma_A && 0\\ 0 && 1\\ \end{array} \right)\otimes \left( \begin{array}{clcr} \gamma} \newcommand{\ka}{\kappa_B & 0\\ 0 & 1\\ \end{array} \right),\\ E_2&=&\left(\begin{array}{clcr} \gamma_A && 0\\ 0 && 1\\ \end{array} \right)\otimes\left( \begin{array}{clcr} 0 & 0 \\ 0 & \om_B\\ \end{array} \right),\\ E_3&=& \left( \begin{array}{clcr} 0 & 0\\ 0 & \om_A \\ \end{array} \right) \otimes \left( \begin{array}{clcr} \gamma} \newcommand{\ka}{\kappa_B & 0\\ 0 & 1\\ \end{array} \right),\\ E_4 &=& \left( \begin{array}{clcr} 0 & 0\\ 0 & \om_A\\ \end{array} \right)\otimes \left( \begin{array}{clcr} 0 & 0 \\ 0 & \om_B\\ \end{array} \right),\label{k5} \end{eqnarray} where the time-dependent Kraus matrix elements are $$\gamma_A(t) = \exp{(-\Gamma^A_{\rm ph} t/2)} \quad {\rm and} \quad \om_A(t) = \sqrt{1-\gamma^2_A(t)},$$ where $\Gamma^A_{\rm ph}$ is the phase damping rate of qubit A. We use the similar expressions $\gamma_B(t)$ and $\om_B(t)$ for qubit B, and will take $\Gamma^A_{\rm ph} = \Gamma^B_{\rm ph}=\Gamma_{\rm ph}$ for greatest simplicity. Similarly, the Kraus operators for zero-temperature amplitude noise are given by \begin{eqnarray} \label{k10} F_1&=&\left(\begin{array}{clcr} \gamma_A & 0\\ 0 & 1\\ \end{array} \right)\otimes\left( \begin{array}{clcr} \gamma_B & 0 \\ 0 & 1\\ \end{array} \right),\label{e1}\\ F_2&=&\left( \begin{array}{clcr} \gamma_A & 0 \\ 0 & 1\\ \end{array} \right)\otimes\left( \begin{array}{clcr} 0 & 0 \\ \omega_B & 0\\ \end{array} \right),\label{e2}\\ F_3&=& \left(\begin{array}{clcr} 0 & 0 \\ \om_A & 0\\ \end{array} \right)\otimes\left( \begin{array}{clcr} \gamma_B & 0 \\ 0 & 1\\ \end{array} \right),\label{e3}\\ F_4 &=& \left(\begin{array}{clcr} 0 & 0) \\ \omega_A & 0 \\ \end{array} \right)\otimes\left( \begin{array}{clcr} 0 & 0 \\ \omega_B & 0 \\ \end{array} \right),\label{e4} \end{eqnarray} and the time-dependent Kraus matrix elements are defined similarly as in the dephasing model, e.g., $\gamma_A(t)=\exp \left(-\Gamma^A_{\rm am } t/2\right)$, etc. With the above explicit solutions of the models, we are able to compute the degree of entanglement of the two qubits in temporal evolution. \section{Decoherence Measure and X States} \label{concurrence/standard} In order to describe the dynamic evolution of quantum entanglement we use Wootters' concurrence \cite{Wootters}. Any entropy-based measure of entanglement will yield the same conclusion about bipartite separability. Concurrence varies from $C=0$ for a separable state to $C=1$ for a maximally entangled state. For any two qubits, the concurrence may be calculated explicitly from their density matrix $\rho$ for qubits $A$ and $B$: \begin{equation}} \newcommand{\eeq}{\end{equation} \label{definationc} C(\rh)=\max\{0,\sqrt{\lam_1}-\sqrt{\lam_2}-\sqrt{\lam_3}-\sqrt{\lam_4}\,\,\}, \eeq where the quantities $\lam_i$ are the eigenvalues in decreasing order of the matrix $\zeta$: \begin{equation}} \newcommand{\eeq}{\end{equation} \zeta=\rho(\sigma_y\otimes \sigma_y)\rho^*(\sigma_y\otimes \sigma_y), \label{concurrence} \eeq where $\rh^*$ denotes the complex conjugation of $\rh$ in the standard basis $|+ +\rangle, |+ -\rangle, |- +\rangle, |- -\rangle$ and $\si_y$ is the Pauli matrix expressed in the same basis as: \begin{equation} \si_y= \left(\begin{array}{clcr} 0 & -i\\ i & 0 \\ \end{array} \right). \end{equation} In the following we will examine the evolution of entanglement under noise-induced relaxation of a class of important bipartite density matrices which are defined below. Since a density matrix in this class only contains non-zero elements in an ``X" formation, along the main diagonal and anti-diagonal, we call them ``X states": \begin{equation} \label{e.oldrho} \rho^{AB} =\left( \begin{array}{clcr} a & 0 & 0 & w\\ 0 & b & z & 0 \\ 0 & z^* & c & 0\\ w^* & 0 & 0 & d \end{array} \right). \end{equation} where $a+b+c+d = 1$. Such a simple matrix is actually not unusual. Experience shows that this X mixed state arises naturally in a wide variety of physical situations (see \cite{WangJ-etal05,Pratt04a,standard2}). We particularly note that it includes pure Bell states as well as the well-known Werner mixed state \cite{Werner} as special cases. Unitary transforms of it extend its domain even more widely, as we will explain below. The mixed states defined here not only are rather common but also have the property that they often retain the X form under noise evolution. This may be expected for phase noise, which can only give time dependence to the off-diagonal matrix elements. The interaction Hamiltonian and Kraus operators for amplitude (e.g., quantum vacuum) noise evolution are different, and evolution under amplitude noise is more elaborate, affecting all six non-zero elements (see \cite{Yu-Eberly04}), but robust form-invariance during evolution is easy to check. This very simple finding applies to a wide array of realistic noise sources. For the X state defined in (\ref{e.oldrho}), concurrence \cite{Wootters} can be easily computed as $$C(\rho^{AB})=2\max\{0, |z|-\sqrt{ad}, |w|-\sqrt{bc}\}.$$ \section{Evolution to Decoherence of the Werner State} \label{decoherence} \noindent Now we examine decoherence evolution under first phase damping and then amplitude damping. Within the set of X matrices, let us focus now on a Werner state \cite{Werner,curious}: \begin{equation}} \newcommand{\eeq}{\end{equation} \label{werner} \rho_W=\frac{1-F}{3}I_4 + \frac{4F-1}{3}|\Psi^-\rangle\langle \Psi^-|, \eeq whose matrix elements can be matched to those of the X state $\rho^{AB}$ easily. $F$ is termed fidelity, and $1 \ge F \ge \frac{1}{4}$. We will begin by obtaining the time-dependence of entanglement for the Werner state. Under phase noise the only time dependence is in $z$: \begin{equation}} \newcommand{\eeq}{\end{equation} \label{zphase} z(t) = \frac{1-4F}{6} \gamma^2(t), \quad {\rm with} \quad \gamma(t) \equiv e^{-\Gamma_{\rm ph}t/2}. \eeq \begin{figure} [htbp] \vspace*{13pt} \centerline{\epsfig{file=FigW-phs2.eps, width=8.2cm}} \vspace*{13pt} \fcaption{\label{fig2} Phase noise causes $\rho_W$ to disentangle completely in finite time for all Werner states except in the limiting case of a pure Bell state. The graph shows $C(t)$ vs. $F$ and $\Gamma_{\rm ph} t$.} \end{figure} The results are shown in Fig. \ref{fig2}. In particular we note the occurrence of ESD, in which concurrence non-smoothly goes to zero at a finite time (and remains zero). This is apparent for all initial $F<1$. It has been noted already for quantum vacuum noise qubit decoherence \cite{Yu-Eberly04} and for disentanglement of continuous joint states \cite{Diosi03, Dodd-Halliwell04, Dodd04}. The analytic expressions above make it clear why this is so. Since the matrix elements $a$ and $d$ are fixed, as $z$ decays it must become less than $\sqrt{ad}$ at a specific time $\tau_{\rm ph}$, which can be easily determined to be given by \begin{equation}} \newcommand{\eeq}{\end{equation} \label{phaseratio} \frac{\tau^{\rm ph}}{\tau_0} = {\ln} \left[\frac{4F-1}{2-2F}\right],\,\,\,\, 1>F>\frac{1}{2} \eeq where $\tau_0 = 1/\Gamma_{\rm ph}$ marks the $1/e$ point in the purely exponential decay of the underlying individual qubits. Next we consider Werner state evolution under amplitude noise, and we find from the appropriate Kraus operators given in (\ref{k10}-\ref{e4}) \cite{Yu-Eberly04} that the following time dependences specify $\rho_W(t)$ at any time: \begin{equation}} \newcommand{\eeq}{\end{equation} z(t)=\frac{1-4F}{6} \gamma} \newcommand{\ka}{\kappa^2, \eeq \begin{equation}} \newcommand{\eeq}{\end{equation} a(t)=\frac{1-F}{3} \gamma} \newcommand{\ka}{\kappa^4, \eeq \begin{equation}} \newcommand{\eeq}{\end{equation} b(t)=\frac{2F+1}{6} \gamma} \newcommand{\ka}{\kappa^2 +\frac{1-F}{3}\gamma} \newcommand{\ka}{\kappa^2\om^2, \eeq \begin{equation}} \newcommand{\eeq}{\end{equation} c(t)=\frac{2F+1}{6} \gamma} \newcommand{\ka}{\kappa^2 +\frac{1-F}{3}\gamma} \newcommand{\ka}{\kappa^2\om^2, \quad {\rm and} \eeq \begin{equation}} \newcommand{\eeq}{\end{equation} d(t)=\frac{1-F}{3} +\frac{2F+1}{3}\om^2 + \frac{1-F}{3}\om^4. \eeq In principle the time-dependent $\gamma$ and $\omega$ parameters could be different for qubits $A$ and $B$, but we again take them identical and write $\gamma} \newcommand{\ka}{\kappa = \exp[-\Gamma_{\rm am} t/2]$ and $\om^2 = 1 - \gamma} \newcommand{\ka}{\kappa^2$, where we use $\Gamma_{\rm am}$ to denote the upper level decay rate of the qubits. \begin{figure} [htbp] \vspace*{13pt} \centerline{\epsfig{file=FigW-ampl2.eps, width=8.2cm}} \vspace*{13pt} \fcaption{\label{fig1} In the presence of amplitude noise there is long-lived concurrence of Werner states only for sufficiently high fidelity, $F > F_{\rm c} \simeq 0.714$. The graph shows the critical fidelity boundary in plotting $C(t)$ vs. $\Gamma_{\rm am} t$.} \end{figure} Sudden death of Werner state entanglement appears here also, but with the important added element that sudden death from amplitude noise occurs only for a low range of fidelity values. Our result shows that for all initial $F$ above the critical fidelity $F_{\rm c} \simeq 0.714$ entanglement remains finite for all time, and has an infinitely long smooth decay, faster than but similar to, the decay of single-qubit coherence. This is shown in the plot in Fig.~\ref{fig1}. \section{Bistability Decoherence} \label{depolar} \noindent In this section, we discuss the entanglement decoherence of a representative X matrix under bistability noise, by which we mean noise that induces incoherent random transfer back and forth between the two qubit states when they are energetically degenerate. Examples occur in bistable systems of all kinds, for example in semiconductor junctions or double-well electron potentials in photonic crystals. A physically different example is polarization of photons in optical fiber with indeterminately random local birefringence. In all of these cases we can speak of the effect as arising from exposure to an infinite-temperature reservoir. In that case the population-equalizing up-transfer and down-transfer rates will be denoted $\Gamma_{\rm eq}$ and taken the same for the two qubits. Given these remarks, our basic model with Hamiltonians (\ref{model}) and (\ref{intam}) still allows a useful Kraus representation and the Kraus matrices are given by: \begin{eqnarray} G_1&=& \frac{1}{\sqrt 2}\left( \begin{array}{cc} \gamma(t) & 0\\ 0 & 1\\ \end{array} \right), \label{G1}\\ G_2 &=& \frac{1}{\sqrt 2} \left( \begin{array}{cc} 0 & 0 \\ \om(t) & 0\\ \end{array} \right), \label{G2}\\ G_3 &=& \frac{1}{\sqrt 2} \left( \begin{array}{cc} 1 & 0 \\ 0 & \gamma(t)\\ \end{array} \right), \label{G3}\\ G_4 &=& \frac{1}{\sqrt 2} \left( \begin{array}{cc} 0 & \omega(t) \\ 0 & 0 \\ \end{array} \right), \label{G4} \end{eqnarray} where now $$\gamma(t) = \exp{(-\Gamma_{\rm eq} t/2)} \quad {\rm and} \quad \om(t) = \sqrt{1-\gamma^2(t)}.$$ State equalization presents the extreme opposite case from vacuum noise, in the sense that strong equalization treats both qubit states equally incoherently, whereas vacuum noise induces incoherent decay into just the energetically lower of the two states. One finds that the X form of the density matrix (\ref{e.oldrho}) is still preserved under state equalizing noise and so at time $t$ it retains the X form (we set $w=0$ for simplicity): \begin{equation} \label{sol} \rho(t) = \left( \begin{array}{clcr} a(t) & 0 & 0 & 0 \\ 0 & b(t) & z(t) & 0 \\ 0 & z(t) & c(t) & 0\\ 0 & 0 & 0 & d(t) \end{array} \right). \end{equation} We assume that the two qubits are affected by two identical local depolarization noises, and in this case the time-dependent matrix elements are given by the following: \begin{eqnarray} 4a(t) &=& \gamma} \newcommand{\ka}{\kappa^4a + a+\om^2 (b + c) + \om^4d\nonumber\\ &&+ 2\gamma} \newcommand{\ka}{\kappa^2a + \gamma} \newcommand{\ka}{\kappa^2\om^2( b+c),\\ 4b(t) &=& 2\gamma} \newcommand{\ka}{\kappa^2b +\gamma} \newcommand{\ka}{\kappa^2\om^2 (a+d) \nonumber\\ &&+ b+\gamma} \newcommand{\ka}{\kappa^4 b+\om^2(a+d) + \om^4c,\\ 4c(t) &=& 2\gamma} \newcommand{\ka}{\kappa^2c +\gamma} \newcommand{\ka}{\kappa^2\om^2 (a +d) \nonumber\\ &&+ c+\om^2(d + a) + \om^4 b+\gamma} \newcommand{\ka}{\kappa^4 c,\\ 4d(t) &=& d+\om^2(b + c) + \om^4 a + \gamma} \newcommand{\ka}{\kappa^4 d\nonumber\\ &&+ 2\gamma} \newcommand{\ka}{\kappa^2 d + \gamma} \newcommand{\ka}{\kappa^2\om^2(b+c),\\ z(t) &=& \gamma} \newcommand{\ka}{\kappa^2 z. \end{eqnarray}} \newcommand{\lam}{\lambda Some algebraic examination shows that this result also leads to ESD, i.e., bistable equalization leads all entangled X states (\ref{e.oldrho}) to become separable states in a finite time. It is easy to check that when $t \to \infty$, \begin{equation}} \newcommand{\eeq}{\end{equation} z \to z(\infty) = 0, \eeq \begin{equation}} \newcommand{\eeq}{\end{equation} \{a, b, c, d\} \to \{\frac{1}{4}, \frac{1}{4}, \frac{1}{4},\frac{1}{4}\}, \eeq and the asymptotic separability arising from the equivalence of all diagonal elements is just a special case of the general theorem by Zyczkowski, et al.\cite{Lewenstein98}. \section{Fragile and Robust Initial Entangled States} \label{fragile} \noindent It is known that entangled states evolve differently under different environmental noise influences if special symmetries exist. For example, in the case of collective dephasing noise (see, e.g., \cite{Yu-Eberly02}), there may exist decoherence-free subspaces in which the entangled states are well protected against interaction with the noise. For the models presented here the noises influence each qubit independently, so there are no decoherence-free subspaces and there is no such protection from ESD available. However, we now show that it is still possible to avoid sudden death by using appropriate local initial preparations. To illustrate this, we consider another mixed state within the category of the X matrix defined in (\ref{e.oldrho}): \begin{equation}} \newcommand{\eeq}{\end{equation} \label{tildewerner} \tilde\rho_W=\frac{1-F}{3}I_4 + \frac{4F-1}{3}|\Phi^-\rangle\langle \Phi^-|, \eeq where $|\Phi^-\rangle = (|++\rangle - |--\rangle)/\sqrt{2},$ is a Bell state. In matrix form at any $t>0$ we can write \begin{equation} \label{tilderho} \tilde\rho_W(t)= \left( \begin{array}{clcr} a(t) & 0 & 0 & w(t) \\ 0 & b(t) & 0 & 0 \\ 0 & 0 & c(t) & 0\\ w^*(t) & 0 & 0 & d(t) \end{array} \right). \end{equation} It is easy to compute the concurrence of this mixed state: $C = 2 \max\{0, |w|-\sqrt{bc}\}$. Consider the time dependences for $\tilde\rho_W$ obtained from the amplitude-noise Kraus operators as before. It is easy to check that the sudden death condition for $\tilde\rho_W$'s concurrence is now: \begin{equation}} \newcommand{\eeq}{\end{equation} \label{ConcurtildeW} \frac{4F-1}{6}\gamma^2 \ =\ \frac{1-F}{3}\gamma^2 + \frac{2F+1}{6}\gamma^2 \omega^2 , \eeq which is satisfied at a finite $t$ for any value of fidelity $F$. That is, here there is no range of ``protected" fidelity values under amplitude noise, as was the case for the other form of X state in Fig. ~\ref{fig1}. Indefinite survival is impossible in this case, similar to what the plot in Fig.~\ref{fig2} shows for phase noise. However, this result has important implications related to survival. One easily shows that $\tilde\rho_W$ is closely related to the earlier Werner state $\rho_W$, which does have a range of protected fidelity values. In fact $\rho_W$ and $\tilde\rho_W$ are unitary transforms of each other under a {\em purely local} transformation operator: $U = i\sigma_x^A \otimes I_B$. This shows that survival against noise of initial mixed state entanglement can in a wide range of situations be dramatically improved by a simple local unitary operation (here changing $\tilde\rho_W$ into $\rho_W$), even while the degree of entanglement is not changed. Intuitively, it is easy to see that the noise influence represented by the Kraus operators varies for different matrix elements of a bipartite density matrix. Although local operations cannot change the degree of entanglememt, it is possible that local operations can rearrange the matrix elements of the X states such that the resulting density matrix is more robust (or fragile) than the original one. Therefore, ESD may be manipulated by preparatory transformation that is purely local. \section{Concluding Remarks} \label{conc} \noindent In summary, we have examined quantitatively via fully analytic expressions the non-local decoherence properties of a wide range of mixed states. We have used relatively simple Kraus operators to do this. We have also shown that three well understood physical noise types (phase noise, amplitude noise, and state-equalizing noise) do not alter the form of the X mixed state during evolution, although entanglement survival may be long or short. In particular, we have established that Werner states are subject to the sudden death effect, and have specified a new critical fidelity boundary below which sudden death must occur. Surprisingly, we have found that Werner states are more robust under pure amplitude noise (e.g., spontaneous emission) than under pure phase noise even though amplitude noise is in a sense more disruptive than dephasing, as the former causes diagonal and off-diagonal relaxation and the latter off-diagonal relaxation alone. Moreover, we have shown that bistability decoherence can cause all the X states to disentangle completely in finite time. In addition, we have shown that in some cases an initial mixed state's entanglement can be preserved under subsequent noisy evolution, i.e., sudden death avoided, by an initial local unitary operation. Such local operations may offer a useful tool in entanglement control when the duration of entangled states is crucial in the processes of quantum state storage and preparation. Finally, we note that while we have found interesting and unexpected features of Werner and other X mixed states in the presence of common noise sources, our Kraus operators treated all of them as white (Markovian) noises. It will be an important theoretical challenge to extend these results to the case of non-Markovian environmental influences \cite{Glendinning-etal,Yonac-etal06,Hu-etal00-05}. \nonumsection{Acknowledgements} \noindent We acknowledge financial support from NSF Grant PHY-0456952 and ARO Grant W911NF-05-1-0543. We also thank C. Broadbent for assistance with Figs. 1 and 2. \nonumsection{References} \noindent
{ "timestamp": "2007-01-02T14:50:15", "yymm": "0503", "arxiv_id": "quant-ph/0503089", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503089" }
\section{Introduction} There has been an accrued interest on quantum plasmas, motivated by applications in ultra small electronic devices \cite{Markowich}, dense astrophysical plasmas \cite{Chabrier}-\cite{Jung} and laser plasmas \cite{Kremp}. Recent developments involves quantum corrections to Bernstein-Greene-Kruskal equilibria \cite{Luque}, quantum beam instabilities \cite{Anderson}-\cite{Haas3}, quantum ion-acoustic waves \cite{Haas4}, quantum corrections to the Zakharov equations \cite{Garcia, Haas5}, modifications on Debye screening for quantum plasmas with magnetic fields \cite{Shokri1}, quantum drift waves \cite{Shokri2}, quantum surface waves \cite{Shokri3}, quantum plasma echoes \cite{Manfredi1}, the expansion of a quantum electron gas into vacuum \cite{Mola} and the quantum Landau damping \cite{Suh}. In addition, quantum methods have been used for the treatment of classical plasma problems \cite{Fedele, Bertrand}. One possible approach to charged particle systems where quantum effects are relevant is furnished by quantum hydrodynamics models. In fact, hydrodynamic formulations have appeared in the early days of quantum mechanics \cite{Madelung}. More recently, the quantum hydrodynamics model for semiconductors has been introduced to handle questions like negative differential resistance as well as resonant tunneling phenomena in micro-electronic devices \cite{Gardner1}-\cite{Gardner3}. The derivation and application of the quantum hydrodynamics model for charged particle systems is the subject of a series of recent works \cite{Manfredi2}-\cite{Degond}. In classical plasmas physics, fluid models are ubiquitous, with their applications ranging from astrophysics to controlled nuclear fusion \cite{Nicholson, Bittencourt}. In particular, magnetohydrodynamics provides one of the most useful fluid models, focusing on the global properties of the plasma. The purpose of this work is to obtain a quantum counterpart of magnetohydrodynamics, starting from the quantum hydrodynamics model for charged particle systems. This provides another place to study the way quantum physics can modify classical plasma physics. However, it should be noted that the quantum hydrodynamic model for charged particle systems was build for non magnetized systems only. To obtain a quantum modified magnetohydrodynamics, this work also offer the appropriated extension of the quantum hydrodynamics model to the cases of non zero magnetic field. The paper is organized as follows. In Section 2, the equations of quantum hydrodynamics are obtained, now allowing for the presence of magnetic fields. The approach for this is based on a Wigner equation with non zero vector potentials. Defining macroscopic quantities like charge density and current through moments of the Wigner function, we arrive at the desired quantum fluid model. In Section 3, we repeat the well known steps for the derivation of magnetohydrodynamics, now including the quantum corrections present in the quantum hydrodynamic model. This produces a quantum magnetohydrodynamics set of equations. In Section 4, a simplified set of quantum magnetohydrodynamics is derived, yielding a quantum version of the generalized Ohm's law. In addition, the infinity conductivity case is shown to imply an ideal quantum magnetohydrodynamic model. In this ideal case, there is the presence of quantum corrections modifying the transport of momentum and the equation for the electric field. Section 5 studies the influence of the quantum terms on the equilibrium solutions. Exact solutions are found for translational invariance. Section 6 is devoted to the conclusions. \section{Quantum Hydrodynamics in the Presence of Magnetic Fields} For completeness, we begin with the derivation of the Wigner-Maxwell system providing a kinetic description for quantum plasmas in the presence of electromagnetic fields. For notational simplicity, we first consider a quantum hydrodynamics model for non zero magnetic fields in the case of a single species plasma. Extension to multi-species plasmas is then straightforward. Our starting point is a statistical mixture with $N$ states described by the wave functions $\psi_\alpha = \psi_{\alpha}({\bf r},t)$, each with probability $p_\alpha$, with $\alpha = 1 \dots N$. Of course, $p_\alpha \geq 0$ and $\sum_{\alpha=1}^{N}p_\alpha = 1$. The wave functions obey the Schr\"odinger equation, \begin{equation} \label{e1} \frac{1}{2m}(-i\hbar\nabla - q{\bf A})^{2}\,\psi_\alpha + q\phi\,\psi_\alpha = i\hbar\frac{\partial\psi_\alpha}{\partial t} \,. \end{equation} Here we consider charge carriers of mass $m$ and charge $q$, subjected to possibly self-consistent scalar and vector potentials $\phi = \phi({\bf r},t)$ and ${\bf A} = {\bf A}({\bf r},t)$ respectively. For convenience in some calculations, we assume the Coulomb gauge, $\nabla\cdot{\bf A} = 0$. From the statistical mixture, we construct the Wigner function $f = f({\bf r},{\bf p},t)$ defined as usual from \begin{equation} \label{e2} f({\bf r},{\bf p},t) = \frac{1}{(2\pi\hbar)^3}\sum_{\alpha=1}^{N}p_{\alpha}\int\,d{\bf s}\,\psi_{\alpha}^{*}({\bf r}+\frac{\bf s}{2})\,e^{\frac{i{\bf p}\cdot{\bf s}}{\hbar}}\,\psi_{\alpha}({\bf r}-\frac{\bf s}{2}) \,. \end{equation} After some long but simple calculations involving the Schr\"odinger equations for each $\psi_\alpha$ and the choice of the Coulomb gauge, we arrive at the following integro-differential equation for the Wigner function, \begin{eqnarray} \label{e3} &\strut& \frac{\partial f}{\partial t} + \frac{\bf p}{m}\cdot\nabla\,f = \\ &\strut& \frac{iq}{\hbar(2\pi\hbar)^3}\int\int d{\bf s}\,d{\bf p}'\,e^{\frac{i({\bf p}-{\bf p}')\cdot{\bf s}}{\hbar}}\,[\phi({\bf r}+\frac{\bf s}{2})-\phi({\bf r}-\frac{\bf s}{2})]\,f({\bf r},{\bf p}',t) + \nonumber \\ &\strut& \frac{iq^2}{2\hbar m(2\pi\hbar)^3}\int\int d{\bf s}\,d{\bf p}'\,e^{\frac{i({\bf p}-{\bf p}')\cdot{\bf s}}{\hbar}}\,[A^{2}({\bf r}+\frac{\bf s}{2})-A^{2}({\bf r}-\frac{\bf s}{2})]\,f({\bf r},{\bf p}',t) + \nonumber \\ &\strut& \frac{q}{2m(2\pi\hbar)^3}\,\,\nabla\cdot\int\int d{\bf s}\,d{\bf p}'\,e^{\frac{i({\bf p}-{\bf p}')\cdot{\bf s}}{\hbar}}\,[{\bf A}({\bf r}+\frac{\bf s}{2})-{\bf A} ({\bf r}-\frac{\bf s}{2})]\,f({\bf r},{\bf p}',t) \nonumber \\ &-& \frac{iq}{\hbar m(2\pi\hbar)^3}\,\,{\bf p}\cdot\int\int d{\bf s}\,d{\bf p}'\,e^{\frac{i({\bf p}-{\bf p}')\cdot{\bf s}}{\hbar}}\,[{\bf A}({\bf r}+\frac{\bf s}{2})-{\bf A}({\bf r}-\frac{\bf s}{2})]\,f({\bf r},{\bf p}',t) \,. \nonumber \end{eqnarray} All macroscopic quantities like charge and current densities can be found taking appropriated moments of the Wigner function. This is analogous to classical kinetic theory, where charge and current densities are obtained from moments of the one-particle distribution function. Alternatively, we could have started from the complete many body wave function, defined a many body Wigner function and then obtained a quantum Bogoliubov-Born-Green-Kirkwood-Yvon hierarchy. With some closure hypothesis, in this way we arrive at a integro-differential equation for the one-particle Wigner function, which has to be supplemented by Maxwell equations. This is the Wigner-Maxwell system, which plays, in quantum physics, the same role the Vlasov-Maxwell system plays in classical physics. When the vector potential is zero, it reproduces the well known Wigner-Poisson system \cite{Drummond, Klimontovich}. In addition, in the formal classical limit when $\hbar \rightarrow 0$, the Wigner equation (\ref{e3}) goes to the Vlasov equation, \begin{equation} \label{e4} \frac{\partial f}{\partial t} + {\bf v}\cdot\nabla f + \frac{q}{m}({\bf E} + {\bf v}\times{\bf B})\cdot\frac{\partial f}{\partial{\bf v}} = 0 \,, \end{equation} where ${\bf v} = ({\bf p}-q{\bf A})/m$, ${\bf E} = - \nabla\phi - \partial{\bf A}/\partial t$ and ${\bf B} = \nabla\times{\bf A}$. However, notice that a initially positive definite Wigner function can evolve in such a way it becomes negative in some regions of phase space. Hence, it can not be considered as a true probability function. Nevertheless, all macroscopic quantities like charge, mass and current densities can be obtained from the Wigner function through appropriated moments. Equation (\ref{e3}) coupled to Maxwell equations provides a self-consistent kinetic description for a quantum plasma. As long as we know, it has been first obtained, with a different notation, in the work \cite{Arnold}. It has been rediscovered in \cite{Materdey}, in the case of homogeneous magnetic fields. Wigner functions appropriated to non zero magnetic fields have also been discussed, for instance, in \cite{Carruthers}-\cite{Bialynicki}, without the derivation of an evolution equation for the Wigner function alone. More recently, a different transport equation for Wigner functions appropriated to non zeros magnetic field and spin has been obtained in \cite{Saikin}. The starting point of this latter development, however, is the Pauli and not the Schr\"odinger equation as here and \cite{Arnold}. Finally, relativistic models for self-consistent charged particle systems with spin can be found in \cite{Masmoudi}. Most of the works dealing with quantum charged particle systems prefer to work with the wave functions and not directly with the Wigner function, as in \cite{Kumar}. The impressive form of (\ref{e3}) seems to support this approach. Indeed, probably (\ref{e3}) can be directly useful only in the linear or homogeneous magnetic field cases. This justifies the introduction of alternative descriptions. At the coast of the loss of some information about kinetic phenemena like Landau damping, we can simplify our model adopting a formal hydrodynamic formulation. Define the fluid density \begin{equation} \label{e5} n = \int\,d{\bf p}\,f \,, \end{equation} the fluid velocity \begin{equation} \label{e6} {\bf u} = \frac{1}{mn}\int\,d{\bf p}\,({\bf p}-q{\bf A})\,f \end{equation} and the pressure dyad \begin{equation} \label{e7} {\bf P} = \frac{1}{m^2}\int\,d{\bf p}\,({\bf p} - q{\bf A})\otimes({\bf p} - q{\bf A})\,f - n{\bf u}\otimes{\bf u} \,. \end{equation} We could proceed to higher order moments of the Wigner function, but (\ref{e5})-(\ref{e7}) are sufficient if we do not want to offer a detailed description of energy transport. Taking the appropriated moments of the Wigner equation (\ref{e3}) and using the definitions (\ref{e5})-(\ref{e7}), we arrive at the following quantum hydrodynamic model, \begin{eqnarray} \label{e8} \frac{\partial n}{\partial t} &+& \nabla\cdot(n{\bf u}) = 0 \,, \\ \label{e9} \frac{\partial{\bf u}}{\partial t} &+& {\bf u}\cdot\nabla{\bf u} = - \frac{1}{n}\nabla\cdot{\bf P} + \frac{q}{m}({\bf E} + {\bf u}\times{\bf B}) \,. \end{eqnarray} Equations (\ref{e8})-(\ref{e9}) does not show in an obvious way any quantum effects, since $\hbar$ is not explicitly present there. To found the hidden quantum effects, we follow mainly the style of references \cite{Manfredi2}, \cite{Gasser1} and \cite{Lopez}, but now allowing for magnetic fields. In the definition (\ref{e2}) of the Wigner function, consider the decomposition \begin{equation} \label{e10} \psi_\alpha = \sqrt{n_\alpha}\,\,\,e^{iS_{\alpha}/\hbar} \,, \end{equation} for real $n_\alpha = n_{\alpha}({\bf r},t)$ and $S_\alpha = S_{\alpha}({\bf r},t)$. Evaluating, the integral for the pressure dyad, we get a decomposition in terms of ``classical" ${\bf P}^C$ and ``quantum" ${\bf P}^Q$ contributions, \begin{equation} \label{e11} {\bf P} = {\bf P}^C + {\bf P}^Q \,, \end{equation} where \begin{eqnarray} \label{a1} {\bf P}^C &=& m\sum_{\alpha=1}^{N}{p_{\alpha}n_\alpha}({\bf u}_{\alpha} - {\bf u})\otimes({\bf u}_{\alpha} - {\bf u}) + \\ &+& m\sum_{\alpha=1}^{N}{p_{\alpha}n_\alpha}({\bf u}^{o}_{\alpha} - {\bf u}^{o})\otimes({\bf u}^{o}_{\alpha} - {\bf u}^{o}) \,, \nonumber \\ \label{a2} {\bf P}^Q &=& - \frac{\hbar^{2}n}{4m}\nabla\otimes\nabla\,\ln\,n \,. \end{eqnarray} In the definitions of classical pressure dyad ${\bf P}^C$, we considered the kinetic fluid velocity associated to the wave function $\psi_\alpha$, \begin{equation} \label{e15} {\bf u}_{\alpha} = \frac{\nabla S_\alpha}{m} \,, \end{equation} and the kinetic fluid velocity associated to the statistical mixture, \begin{equation} \label{e16} {\bf u} = \sum_{\alpha=1}^{N}\,\frac{p_{\alpha}n_\alpha}{n}{\bf u}_{\alpha} \,. \end{equation} In a similar way, the second term at the right hand side of equation (\ref{a1}) is constructed in terms of ${\bf u}_{\alpha}^o$, the osmotic fluid velocity associated to the wave function $\psi_\alpha$, \begin{equation} \label{e17} {\bf u}_{\alpha}^o = \frac{\hbar}{2m}\frac{\nabla n_\alpha}{n_\alpha} \,, \end{equation} and ${\bf u}^o$, the osmotic fluid velocity associated to the statistical mixture, \begin{equation} \label{e18} {\bf u}^o = \sum_{\alpha=1}^{N}\,\frac{p_{\alpha}n_\alpha}{n}{\bf u}_{\alpha}^o \,. \end{equation} We also observe that in terms of the fluid density $n_\alpha$ of the state $\alpha$ the density $n$ of the statistical mixture is given by \begin{equation} \label{e19} n = \sum_{\alpha=1}^{N}\,p_{\alpha}n_\alpha \,. \end{equation} Notice that ${\bf P}^C$ a faithful classical pressure dyad, since it comes from dispersion of the velocities, vanishing for a pure state. Indeed, the classical pressure dyad is the sum of a kinetic part, arising from the dispersion of the kinetic velocities, and a osmotic part, arising from the dispersion of the osmotic velocities. However, ${\bf P}^C$ is not strictly classical, since it contains $\hbar$ through the osmotic velocities. In a sense, however, it is ``classical", since it comes from statistical dispersion of the velocities. In most cases, it suffices to take some equation of state for ${\bf P}^C$. For simplicity, from now on we assume a diagonal, isotropic form $P_{ij} = \delta_{ij}P$, where $P = P(n)$ is a suitable equation of state. Certainly, strong magnetic fields have to be treated more carefully, since they are associated to anisotropic pressure dyads. However, since we are mainly interested on the role of the quantum effects, we disregard such possibility here. Now inserting the preceding results for the pressure dyad into the momentum transport equation (\ref{e9}), we obtain the suggestive equation \begin{equation} \label{e20} \frac{\partial{\bf u}}{\partial t} + {\bf u}\cdot\nabla{\bf u} = - \frac{1}{mn}\nabla P + \frac{q}{m}({\bf E} + {\bf u}\times{\bf B}) + \frac{\hbar^2}{2m^2}\nabla\left(\frac{\nabla^{2}\sqrt{n}}{\sqrt{n}}\right) \,. \end{equation} The equation of continuity (\ref{e8}) and the force equation (\ref{e20}) constitute our quantum hydrodynamic model for magnetized systems. All the quantum effects are contained in the last term of the equation (\ref{e20}), the so called Bohm potential. In comparison with standard fluid models for charged particle systems, the Bohm potential is the only quantum contribution, and the rest of the paper is devoted to study its consequences for magnetohydrodynamics. \section{Quantum Magnetohydrodynamics Model} The equations from the last Section were written for a single species charged particle system. Now we generalize to a two species system. Consider electrons with fluid density $n_e$, fluid velocity ${\bf u}_e$, charge $-e$, mass $m_e$ and pressure $P_e$. In an analogous fashion, consider ions with fluid density $n_i$, fluid velocity ${\bf u}_i$, charge $e$, mass $m_i$ and pressure $P_i$. Proceeding as before, now starting from the Wigner equations for electrons and ions, we get the following bipolar quantum fluid model, \begin{eqnarray} \label{e21} \frac{\partial n_e}{\partial t} + \nabla\cdot(n_{e}{\bf u}_e) &=& 0 \,, \\ \label{e22} \frac{\partial n_i}{\partial t} + \nabla\cdot(n_{i}{\bf u}_i) &=& 0 \,, \\ \label{e23} \frac{\partial{\bf u}_e}{\partial t} + {\bf u}_{e}\cdot\nabla{\bf u}_e &=& - \frac{\nabla P_{e}}{m_{e}n_{e}} - \frac{e}{m_e}({\bf E} + {\bf u}_{e}\times{\bf B}) + \nonumber \\ &+& \frac{\hbar^2}{2m^{2}_{e}}\nabla\left(\frac{\nabla^{2}\sqrt{n_e}}{\sqrt{n_e}}\right) - \nu_{ei}({\bf u}_e - {\bf u}_{i}) \,,\\ \label{e24} \frac{\partial{\bf u}_i}{\partial t} + {\bf u}_{i}\cdot\nabla{\bf u}_i &=& - \frac{\nabla P_{i}}{m_{i}n_{i}} + \frac{e}{m_i}({\bf E} + {\bf u}_{i}\times{\bf B}) + \nonumber \\ &+& \frac{\hbar^2}{2m^{2}_{i}}\nabla\left(\frac{\nabla^{2}\sqrt{n_i}}{\sqrt{n_i}}\right) - \nu_{ie}({\bf u}_i - {\bf u}_{e})\,. \end{eqnarray} In the equations (\ref{e23}-\ref{e24}), we have added some often used phenomenological terms to take into account for the momentum transport by collisions. The coefficients $\nu_{ei}$ and $\nu_{ie}$ are called collision frequencies for momentum transfer between electrons and ions \cite{Nicholson, Bittencourt}. For quasineutral plasmas, global momentum conservation in collisions imply $m_{e}\nu_{ei} = m_{i}\nu_{ie}$, so that $\nu_{ie} \ll \nu_{ei}$ when the ions are much more massive than electrons \cite{Nicholson, Bittencourt}. Equations (\ref{e21})-(\ref{e24}) have to be supplemented by Maxwell equations, \begin{eqnarray} \label{e25} \nabla\cdot{\bf E} &=& \frac{\rho}{\varepsilon_0} \,,\\ \label{e26} \nabla\cdot{\bf B} &=& 0 \,,\\ \label{e27} \nabla\times{\bf E} &=& - \frac{\partial\bf B}{\partial t} \,,\\ \label{e28} \nabla\times{\bf B} &=& \mu_{0}{\bf J} + \mu_{0}\varepsilon_{0}\frac{\partial\bf E}{\partial t} \,, \end{eqnarray} where the charge and current densities are given respectively by \begin{equation} \label{e29} \rho = e\,(n_i - n_{e}) \,, \quad {\bf J} = e\,(n_{i}{\bf u}_i - n_{e}{\bf u}_{e}) \,. \end{equation} Equations (\ref{e21}-\ref{e29}) constitute our complete quantum hydrodynamic model, allowing for magnetic fields. When ${\bf B} \equiv 0$, it goes to the well known quantum hydrodynamic model for bipolar charged particle systems. Several possibilities of study are open starting from (\ref{e21}-\ref{e29}). Here we are interested in obtaining equations analogous to the classical magnetohydrodynamic equations. In some places, for the sake of clarity and to point exactly for the new contributions of quantum nature, we repeat some well known steps in the derivation of classical magnetohydrodynamics. To proceed in this direction, define the global mass density \begin{equation} \label{e30} \rho_m = m_{e}n_e + m_{i}n_i \end{equation} and the global fluid velocity \begin{equation} \label{e31} {\bf U} = \frac{m_{e}n_{e}{\bf u}_{e} + m_{i}n_{i}{\bf u}_{i}}{m_{e}n_{e} + m_{i}n_{i}}\,. \end{equation} With these definitions and proceeding like in any plasma physics book \cite{Nicholson, Bittencourt}, we obtain the following equations for $\rho_m$ and ${\bf U}$, \begin{eqnarray} \label{e32} \frac{\partial\rho_m}{\partial t} + \nabla\cdot(\rho_{m}{\bf U}) &=& 0 \,,\\ \rho_{m}(\frac{\partial{\bf U}}{\partial t} + {\bf U}\cdot\nabla{\bf U}) &=& - \nabla\cdot{\bf\Pi} + \rho{\bf E} + {\bf J}\times{\bf B} + \nonumber \\ \label{e33} &+& \frac{\hbar^{2}n_{e}}{2m_{e}}\nabla\left(\frac{\nabla^{2}\sqrt{n_{e}}}{\sqrt{n_{e}}}\right) + \frac{\hbar^{2}n_{i}}{2m_{i}}\nabla\left(\frac{\nabla^{2}\sqrt{n_{i}}}{\sqrt{n_{i}}}\right) \,, \end{eqnarray} for \begin{equation} \label{e34} {\bf\Pi} = P\,{\bf I} + \frac{m_{e}m_{i}n_{e}n_{i}}{\rho_m}({\bf u}_e - {\bf u}_{i})\otimes({\bf u}_e - {\bf u}_{i}) \,, \end{equation} where $P = P_{e} + P_{i}$ and where ${\bf I}$ is the identity matrix. In equations (\ref{e33}-\ref{e34}), the electronic and ionic densities are defined in terms of the mass and charge densities according to \begin{equation} \label{e35} n_e = \frac{1}{m_i + m_{e}}\,\,(\rho_m - \frac{m_{i}}{e}\rho) \,, \quad n_i = \frac{1}{m_i + m_{e}}\,\,(\rho_m + \frac{m_{e}}{e}\rho) \,. \end{equation} We can simplify (\ref{e33}) considerably assuming, as usual, quasi-neutrality ($\rho = 0$ so that $n_e = n_i$), $P_e = P_i = P/2$ and neglecting $m_e$ in comparison to $m_i$ whenever possible. In addition, disregarding the last term at the right hand side of (\ref{e34}), we obtain \begin{equation} \label{e36} \frac{\partial{\bf U}}{\partial t} + {\bf U}\cdot\nabla{\bf U} = -\frac{1}{\rho_m}\nabla P + \frac{1}{\rho_m}{\bf J}\times{\bf B} + \frac{\hbar^{2}}{2m_{e}m_{i}}\nabla\left(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}\right) \,. \end{equation} Under the same assumptions and following the standard derivation of magnetohydrodynamics \cite{Nicholson, Bittencourt}, we obtain the following equation for the current ${\bf J}$, \begin{equation} \label{e37} \frac{m_{e}m_{i}}{\rho_{m}e^2}\frac{\partial{\bf J}}{\partial t} - \frac{m_{i}\nabla P}{\rho_{m}e} = {\bf E} + {\bf U}\times{\bf B} - \frac{m_{i}}{\rho_{m}e}\,{\bf J}\times{\bf B} - \frac{\hbar^{2}}{2e m_{e}}\nabla\left(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}\right) - \frac{1}{\sigma}\,{\bf J} \,, \end{equation} where $\sigma = \rho_{m}e^{2}/(m_{e}m_{i}\nu_{ei})$ is the longitudinal electrical conductivity. Equation (\ref{e37}) is the quantum version of the generalized Ohm's law \cite{Nicholson, Bittencourt}. The continuity equation (\ref{e32}), the force equation (\ref{e36}), the quantum version of the generalized Ohm's law (\ref{e37}), an equation of state for $P$, plus Maxwell equations, provides a full system of quantum magnetohydrodynamic equations. However, it is probably still complicated and in the next section we propose some approximations in the same spirit of those of classical magnetohydrodynamics. \section{Simplified and Ideal Quantum Magnetohydrodynamic Equations} Usually \cite{Nicholson, Bittencourt}, the left-hand side of the equation (\ref{e37}) is neglected in the cases of slowly varying processes and small pressures. Also, for slowly varying and high conductivity problems , the displacement current can be neglected in Amp\`ere's law. Finally, we assume an equation of state appropriated for adiabatic processes. This provides a complete system of simplified quantum magnetohydrodynamic equations, which we collect here for convenience, \begin{eqnarray} \label{e38} \frac{\partial\rho_m}{\partial t} &+& \nabla\cdot(\rho_{m}{\bf U}) = 0 \,,\\ \label{e39} \frac{\partial{\bf U}}{\partial t} &+& {\bf U}\cdot\nabla{\bf U} = - \frac{1}{\rho_m}\nabla P + \frac{1}{\rho_m}{\bf J}\times{\bf B} + \frac{\hbar^{2}}{2m_{e}m_{i}}\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}) \,,\\ \label{e40} \nabla P &=& V_{s}^{2}\nabla\rho_m \,,\\ \label{e41} \nabla&\times&{\bf E} = - \frac{\partial{\bf B}}{\partial t} \,,\\ \label{e42} \nabla&\times&{\bf B} = \mu_{0}{\bf J} \,,\\ \label{e43} {\bf J} &=& \sigma[{\bf E} + {\bf U}\times{\bf B} - \frac{m_{i}}{\rho_{m}e}\,{\bf J}\times{\bf B} - \frac{\hbar^{2}}{2e m_{e}}\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}})] \,. \end{eqnarray} In equation (\ref{e40}), $V_s$ is the adiabatic speed of sound of the fluid. Gauss law can be regarded as the initial condition for Faraday's law. Also notice that the Hall term ${\bf J}\times{\bf B}$ at (\ref{e43}) is often neglected in magnetohydrodynamics. Inserting (\ref{e40}) into (\ref{e39}), we are left with a system of 13 equations for 13 unknowns, namely, $\rho_m$ and the components of ${\bf U}, {\bf J}, {\bf B}$ and ${\bf E}$. This is our quantum magnetohydrodynamics model. In comparison to classical magnetohydrodynamics, the difference of the present model rests on the presence of two quantum corrections, the last terms at equations (\ref{e39}) and (\ref{e43}). In the ideal magnetohydrodynamics approximation, we assume an infinite conductivity and neglect the Hall force at (\ref{e43}). This provides the following ideal quantum magnetohydrodynamics model, \begin{eqnarray} \label{e44} {\bf E} = - {\bf U}\times{\bf B} &+& \frac{\hbar^{2}}{2e m_{e}}\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}) \,, \\ \rho_{m}(\frac{\partial{\bf U}}{\partial t} + {\bf U}\cdot\nabla{\bf U}) &=& - \nabla P + \frac{1}{\mu_0}(\nabla\times{\bf B})\times{\bf B} + \nonumber \\ \label{e45} &+& \frac{\hbar^{2}\rho_m}{2m_{e}m_{i}}\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}) \,,\\ \label{e46} \frac{\partial{\bf B}}{\partial t} &=& \nabla\times({\bf U}\times{\bf B}) \,, \end{eqnarray} supplemented by the continuity equation (\ref{e38}) and the equation of state (\ref{e40}). Taking into account (\ref{e40}), equations (\ref{e45}-\ref{e46}) plus the continuity equation provides a system of 7 equations for 7 unknowns, namely, $\rho_m$ and the components of ${\bf U}$ and ${\bf B}$. This is our ideal quantum magnetohydrodynamics model. In comparison to classical ideal magnetohydrodynamics, the difference of the present model rests on the presence of a quantum correction, the last term at equation (\ref{e45}). Interestingly, taking the curl of (\ref{e44}) makes disappear one of the quantum correction terms present in the non ideal quantum magnetohydrodynamics. This leads to a dynamo equation (\ref{e46}) identical to that of classical magnetohydrodynamics. Consequently, for infinite conductivity the magnetic field lines are still frozen to the fluid, even allowing for the quantum corrections proposed here. In fact, even for finite conductivity, the diffusion of magnetic field lines is described by the same diffusion equation as that of classical magnetohydrodynamics. This comes from the fact that the quantum correction disappear after taking the curl of both sides of (\ref{e43}), neglecting the Hall term and assuming a constant $\sigma$ as usual. However, a further quantum correction on the electric field still survives through (\ref{e44}). In order to obtain a deeper understanding of the importance of quantum effects, we propose the following rescaling for our ideal quantum magnetohydrodynamic equations, \begin{eqnarray} \bar{\rho}_m &=& \rho_{m}/\rho_0 \,, \quad \bar{\bf U} = {\bf U}/V_A \,, \quad \bar{\bf B} = {\bf B}/B_0 \,, \nonumber \\ \label{e48} \bar{\bf r} &=& \Omega_{i}{\bf r}/V_A \,, \quad \bar{t} = \Omega_{i}t \,, \end{eqnarray} where $\rho_0$ and $B_0$ are the equilibrium mass density and magnetic field. In addition, $V_A = (B_{0}^{2}/(\mu_{0}\rho_{0}))^{1/2}$ is the Alfv\'en velocity and $\Omega_i = eB_{0}/m_i$ is the ion cyclotron velocity. We justify the chosen rescaling in the following way. In magnetohydrodynamics, the Alf\'en velocity provides a natural velocity scale. Also, since we deal with low frequency problems, $\Omega_{i}^{-1}$ is a reasonable candidate for a natural time scale. These velocity and time scales induces the length scale $V_{A}/\Omega_{i}$, as shown in (\ref{e48}). Applying the rescaling (\ref{e48}) to the ideal quantum magnetohydrodynamic equations, we obtain the following non dimensional model, \begin{eqnarray} \label{e49} \frac{\partial\bar{\rho}_m}{\partial t} &+& \nabla\cdot(\bar{\rho}_{m}\bar{\bf U}) = 0 \,,\\ \bar{\rho}_{m}(\frac{\partial\bar{\bf U}}{\partial t} + \bar{\bf U}\cdot\nabla\bar{\bf U}) &=& - \frac{V_{s}^2}{V_{A}^2}\nabla\bar{\rho}_m + (\nabla\times\bar{\bf B})\times\bar{\bf B} + \nonumber \\ \label{e50} &+& \frac{H^{2}\bar{\rho}_m}{2}\nabla(\frac{\nabla^{2}\sqrt{\bar{\rho}_m}}{\sqrt{\bar{\rho}_m}}) \,,\\ \label{e51} \frac{\partial\bar{\bf B}}{\partial t} &=& \nabla\times(\bar{\bf U}\times\bar{\bf B}) \,, \end{eqnarray} where \begin{equation} \label{e52} H = \frac{\hbar\Omega_i}{\sqrt{m_{e}m_{i}}\,\,V_{A}^{2}} \end{equation} is a non dimensional parameter measuring the relevance of quantum effects. Numerically, using M.K.S. units, we have $H = 3.42 \times 10^{-30} \,\,n_{0}/B_{0}$, where $n_0$ is the ambient particle density. While for ordinary plasmas $H$ is negligible, for dense astrophysical plasmas \cite{Chabrier}-\cite{Jung}, with $n_0$ about $10^{29} - 10^{34}\,\, m^{-3}$, $H$ can be of order unity or more. Hence, in dense astrophysical plasmas like the atmosphere of neutron stars or the interior of massive white dwarfs, quantum corrections to magnetohydrodynamics can be of experimental importance. Similar comments apply to our non ideal quantum magnetohydrodynamics model. However, even for moderate $H$ quantum effects can be negligible if the density is slowly varying in comparison with some typical length scale, due to the presence of a third order derivative at the Bohm potential. This is in the same spirit of the Thomas-Fermi approximation. \section{Quan\-tum Ideal Magnetostatic E\-qui\-li\-brium} There is a myriad of developments based on classical magnetohydrodynamics (linear and nonlinear waves, dynamo theory and so on) and we shall not attempt to reproduce all the quantum counterparts of these subjects in the framework of our model. We will be restricted to just one subject, namely the construction of exact equilibria for ideal quantum magnetohydrodynamics, with no attempt to study the important question of the stability of the equilibria. Assuming that ${\bf U} = 0$ and that all quantities are time-independent, the ideal quantum magnetohydrodynamic equations (\ref{e44}-\ref{e46}) becomes \begin{eqnarray} \label{e53} {\bf E} &=& \frac{\hbar^{2}}{2e m_{e}}\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}) \,, \\ \label{e54} \nabla P &=& \frac{1}{\mu_0}(\nabla\times{\bf B})\times{\bf B} + \frac{\hbar^{2}\rho_m}{2m_{e}m_{i}}\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}) \,. \end{eqnarray} According to (\ref{e53}), the equilibrium solutions of ideal quantum magnetohydrodynamics are not electric field free any longer. In addition, equation (\ref{e54}) has an quantum correction that invalidate the classical magnetic surface equation for ${\bf B}\cdot\nabla{\bf B} = 0$, namely $P + B^{2}/(2\mu_{0}) =$ cte. Equation (\ref{e54}) together with an equation of state is the key for the search of equilibrium solutions. We will try to follow, as long as possible, the strategy of reference \cite{Hamabata} for classical magnetostatic equilibria. Inspired by well known classical solutions \cite{Hamabata}, assume a translationally invariant solution of the form \begin{eqnarray} \label{e55} P &=& P(r,\varphi) \,, \quad \rho_m = \rho_{m}(r,\varphi) \,, \\ \label{e56} {\bf B} &=& \nabla A(r,\varphi)\times\hat{z} + B_{z}(r,\varphi)\hat{z} \,, \end{eqnarray} using cylindrical coordinates and where $A = A(r,\varphi)$ and $B_{z} = B_{z}(r,\varphi)$ as well as the pressure and the mass density are functions of $(r,\varphi)$ only. Substituting the proposal (\ref{e55}-\ref{e56}) into (\ref{e54}), we get, for the radial and azimuthal components of this equation, \begin{equation} \label{e57} \nabla(P + \frac{B_{z}^2}{2\mu_0}) = - \frac{1}{\mu_0}\,\nabla A\,\,\nabla^{2}A + \frac{\hbar^{2}\rho_{m}}{2m_{e}m_{i}}\,\,\nabla(\frac{\nabla^{2}\sqrt{\rho_m}}{\sqrt{\rho_m}}) \,, \end{equation} while, for the $z$ component, the result is \begin{equation} \label{e58} \frac{\partial(B_{z},A)}{\partial(r,\varphi)} = 0 \,. \end{equation} In (\ref{e58}) and in what follows, we used the definition of Jacobian, \begin{equation} \label{e59} \frac{\partial(B_{z},A)}{\partial(r,\varphi)} = \left(\matrix{\frac{\partial\,B_z}{\partial r} & \frac{\partial\,B_z}{\partial \varphi}\cr \frac{\partial\,A}{\partial r} & \frac{\partial\,A}{\partial \varphi}\cr}\right) \,. \end{equation} From (\ref{e58}), we obtain \begin{equation} \label{e60} B_z = B_{z}(A) \,. \end{equation} Taking into account (\ref{e57}) and the fact that $B_z$ is a function of $A$, it follows that \begin{equation} \label{e61} \frac{\partial(P,A)}{\partial(r,\varphi)} = \frac{\hbar^{2}\rho_{m}}{2m_{e}m_{i}}\frac{\partial(\nabla^{2}\sqrt{\rho_{m}}\,\,/\sqrt{\rho_m}\,\, ,A)}{\partial(r,\varphi)} \,. \end{equation} In the classical limit $\hbar \rightarrow 0$, the right hand of (\ref{e61}) vanishes, implying just the functional relationship $P = P(A)$. In the present work, we still postulate \begin{equation} \label{e62} P = P(A) \,, \end{equation} so that, from (\ref{e61}), we have \begin{equation} \label{e63} \frac{\nabla^{2}\sqrt{\rho_{m}}}{\sqrt{\rho_{m}}} = F(A) \,, \end{equation} where $F = F(A)$ is an arbitrary function. The last equation is a distinctive feature of ideal quantum magnetohydrodynamic equilibrium. Indeed, (\ref{e63}) would not be necessary if $\hbar = 0$ in (\ref{e61}). Hence, even if $\hbar$ is not present in (\ref{e63}), this equation has a quantum nature, with important implications in what follows. The reason why $\hbar$ does not appear in (\ref{e63}) is that it factor at the right hand side of (\ref{e61}). From (\ref{e62}) and some subjacent equation of state, $P = P(\rho_{m})$, we deduce \begin{equation} \label{e64} \sqrt{\rho_{m}} = G(A) \,, \end{equation} for some function $G = G(A)$. Plugging this into (\ref{e63}), the result is \begin{equation} \label{e65} \frac{G'}{G}\,\nabla^{2}A + \frac{G''}{G}\,(\nabla\,A)^2 = F(A) \,, \end{equation} where the prime denotes derivation with respect to $A$. Coming back to (\ref{e57}), we obtain \begin{equation} \label{e66} \nabla^{2}A = \mu_{0}[- K'(A) + \frac{\hbar^{2}}{2m_{e}m_{i}}\,\,G^{2}F'(A)] \,, \end{equation} where we have defined \begin{equation} \label{e67} K = K(A) = P(A) + \frac{B_{z}^{2}(A)}{2\mu_0} \,. \end{equation} Recapitulating, we have three four functions of $A$ to be stipulated, namely $F$, $G$, $K$ and $P$. However, $A$ satisfy two different equations, (\ref{e65}) and (\ref{e66}). Once $A$ is found, all other quantities (pressure, mass density, electromagnetic field) comes as consequences. A reasonable choice is to take $G$ as a linear function of $A$, since then (\ref{e65}) becomes linear in the derivatives. Hence, let \begin{equation} \label{e68} G = k_{1}A + k_2 \,, \quad k_1 \neq 0 \,, \end{equation} for numerical constants $k_1$ and $k_2$. We take $k_1 \neq 0$ since $k_1 = 0$ would imply $F = 0$, making disappear the quantum correction at (\ref{e66}). With the choice (\ref{e68}), the couple (\ref{e65}-\ref{e66}) becomes \begin{eqnarray} \label{e69} \nabla^{2} A &=& \frac{1}{k_1}\,(k_{1}A + k_{2})\,F(A) \,, \\ \label{e70} \nabla^{2} A &=& \mu_{0}\,[-K'(A) + \frac{\hbar^{2}}{2m_{e}m_{i}}\,(k_{1}A+k_{2})^{2}\,F'(A)] \,. \end{eqnarray} The right hand sides of (\ref{e69}) and (\ref{e70}) should coincide, implying \begin{equation} \label{e71} K'(A) = \frac{\hbar^{2}}{2m_{e}m_{i}}\,(k_{1}A + k_{2})^{2}\,F'(A) - \frac{1}{\mu_{0}k_{1}}\,(k_{1}A+k_{2})\,F(A) \,. \end{equation} The last equation define $K$ up to an unimportant numerical constant. Equation (\ref{e69}) is the key equation for our translationally invariant magnetostatic equilibria. For a given $F(A)$ and solving (\ref{e69}) for $A$, all other quantities follows for a known equation of state. Indeed, knowing $A$ we can construct the radial and azimuthal components of the magnetic field through (\ref{e56}) and the mass density from (\ref{e64}). From the mass density and the equation of state, we obtain the pressure $P$. Proceeding, equation (\ref{e71}) yields $K(A)$ and then the $z$ component of the magnetic field through (\ref{e67}). Finally, the electric field follows from (\ref{e53}) and the current density from the curl of the magnetic field. The free ingredients to be chosen to construct explicitly the exact solution are the function $F(A)$ and the equation of state, and the numerical constants $k_1$ and $k_2$. Other possibilities can be explored if we do not restrict to linear $G(A)$ functions as in (\ref{e68}), but then $A$ will not satisfy an linear in the derivatives equation. \subsection{An Explicit Exact Solution} An interesting case of explicit solution is provided by the choice \begin{equation} \label{e72} F(A) = \frac{k_{1}\,B_{0}\,(1-\varepsilon^{2}k)}{k_{1}A + k_{2}}\,\,e^{-2kA/B_{0}} \,, \end{equation} where $B_0$ is an arbitrary constant magnetic field, $k$ is an arbitrary constant with dimensions of an inverse length and $0 \leq \varepsilon < 1$. With the choice (\ref{e72}), the equation (\ref{e69}) traduces into the Liouville equation, \begin{equation} \label{e73} \nabla^{2}A = (1-\varepsilon^{2})\,B_{0}\,k\,e^{-2kA/B_{0}} \,, \end{equation} which admits the exact cat eye solution \begin{equation} \label{e74} A = \frac{B_{0}}{k}\,\,\ln[\cosh(kr\,cos\varphi) + \varepsilon\,\cos(kr\sin\varphi)] \,. \end{equation} All other relevant quantities follows from this exact solution following the recipe just stated. The mass density, from (\ref{e64}), is \begin{equation} \label{e75} \rho_m = [\frac{k_{1}B_{0}}{k}\,\,\ln(\cosh(kr\,\cos\varphi) + \varepsilon\,\cos(kr\sin\varphi)) + k_{2}]^2 \,, \end{equation} while the radial and azimuthal components of the magnetic field follows from (\ref{e56}), \begin{eqnarray} \label{e76} B_r &=& - \frac{B_{0}\,[\sin\varphi\,\sinh(kr\,\cos\varphi) + \varepsilon\cos\varphi\sin(kr\,\sin\varphi)]}{[\cosh(kr\,\cos\varphi) + \varepsilon\cos(kr\,\sin\varphi)]} \,, \\ \label{e77} B_\varphi &=& - \frac{B_{0}\,[\cos\varphi\,\sinh(kr\,\cos\varphi) - \varepsilon\sin\varphi\sin(kr\,\sin\varphi)]}{[\cosh(kr\,\cos\varphi) + \varepsilon\cos(kr\,\sin\varphi)]} \,. \end{eqnarray} Assuming an adiabatic equation of state, $P = V_{s}\rho_m$, we get, from (\ref{e67}), \begin{eqnarray} \label{e78} B_{z}^2 &=& B_{0}^2 - 2\mu_{0}\,V_{s}^{2}\,(k_{1}A+k_{2})^{2} + \\ &+&(1 - \varepsilon^{2})\,k^{2}\,e^{-2A}\,\,[1 + \mu_{0}k_{1}\,(k_{1}+k_{2})\hbar^{2}/m + \mu_{0}\,\hbar^{2}\,k_{1}^{2}\,A/m] \,, \end{eqnarray} with $A$ given by the cat eye solution (\ref{e74}). If desired, the electric field and the current density can then be calculated via (\ref{e53}) and Amp\`ere's law respectively. In figure 1, we show the contour plot of the function $A$ given by (\ref{e74}), while in figure 2 we show the corresponding mass density. The parameters chosen were $B_0 = 1$, $k = 1$, $\varepsilon = 0.9$, $k_1 = 1$ and $k_2 = 0$. These graphics shows coherent, periodic patterns resembling quantum periodic solutions arising in other quantum plasma systems \cite{Manfredi2}. Similar graphics can be easily obtained for the electromagnetic field and other macroscopic quantities derivable from the cat eye solution (\ref{e74}). \section{Conclusion} In this work, we have obtained a quantum version of magnetohydrodynamics starting from a quantum hydrodynamics model with nonzero magnetic fields. In view of its simplicity, this magnetic quantum hydrodynamics model seems to be an attractive alternative to the Wigner magnetic equation of Section 2. The infinite conductivity approximation leads to an ideal quantum magnetohydrodynamics. For very dense plasmas and not to strong magnetic fields, the quantum corrections to magnetohydrodynamics can be relevant, as apparent from the parameter $H$ derived in Section 4. Under a number of suitable assumptions, we have derived some exact translationally invariant quantum ideal magnetostatic solutions. More general quantum ideal magnetostatic equilibria can be conjectured, in particular for axially symmetric situations. In addition, we have left a full investigation of linear waves to future works. \vskip 1cm \noindent{\bf Acknowledgments}\\ We thanks the Brazilian agency Conselho Nacional de Desenvolvimento Cien\-t\'{\i}\-fi\-co e Tecn\'ologico (CNPq) for financial support.
{ "timestamp": "2005-03-02T11:18:42", "yymm": "0503", "arxiv_id": "physics/0503021", "language": "en", "url": "https://arxiv.org/abs/physics/0503021" }
\section{Introduction} Unitary quantum mechanics (that is, quantum mechanics without collapse of the wave function) has local interactions: the quantum state of a system (e.g.\,a qubit, or a spacetime region in quantum field theory) is affected only by influences which propagate via the quantum states of its immediate past light cone.\footnote{In QFT, this is a consequence of the requirement that spacelike separated observables must commute.} As conventionally presented, though, QM does not have local \emph{states}: if $S_1$ and $S_2$ are systems with quantum states $\ensuremath{\rho}_1$ and $\ensuremath{\rho}_2$, then because of entanglement the state of the composite system $S_1 \times S_2$ is not necessarily $\ensuremath{\rho}_1 \otimes \ensuremath{\rho}_2$. Deutsch and Hayden\cite{deutschhayden} argue that this `state nonlocality' is an artifact of the normal way in which we represent quantum states, and that it disappears in an alternative formalism which they propose. Their formalism is derived from the Heisenberg picture of quantum mechanics, in which the unitary time evolution is applied to the observables rather than to the state vector. In the normal understanding of that formalism, though, the state vector is still taken to express the physical state of the system (via its role in calculating expectation values) and the algebra of observable quantities is regarded as mathematical `superstructure', used to help us to calculate those observables. Deutsch and Hayden reverse this `normal understanding'. They regard the state vector \ket{0} as fixed, once and for all and independent of the physical state of the system, and they regard the state of a quantum system as literally given by the associated observables (so that the state of a qubit, for instance, is given by the triple of Heisenberg picture operators $S_x, S_y, S_z$ pertaining to the spin observables of that qubit). The dynamics of this theory are given by \begin{equation} \label{truedyn}\ensuremath{\frac{\dr{}}{\dr{t}}}\op{X}_i= \frac{-i}{\hbar}\comm{\op{H}(\op{X}_1, \ldots \op{X}_n)}{\op{X}_i}\end{equation} (where $\op{X}_1, \ldots \op{X}_n$ are the observables of the theory). It is easy to see that the theory is local in both the interaction and the state senses, apparently vindicating Deutsch and Hayden's claims. \section{Quantum gauge transformations} Suppose $\op{V}(t)$ is a function from times to unitary operators, and suppose that for each $t$, $\op{V}(t)\ket{0}=\exp(-i \theta) \ket{0}$ (for arbitrary phase factor $\theta$). Then if the state is represented, according to Deutsch and Hayden, by observables $\op{X}_1, \ldots \op{X}_n$, suppose that we make the transformation \begin{equation} \op{X}_i(t) \longrightarrow \op{X}_i'(t)=\opad{U}(t)\op{X}_i(t)\op{U}(t).\end{equation} If $\op{V}(t)$ is not a constant then this changes the dynamics to \begin{equation} \label{gaugedyn}\ensuremath{\frac{\dr{}}{\dr{t}}}\op{X}'_i = \frac{-i}{\hbar} \comm{\op{H}(\op{X}'_1, \ldots\op{X}'_n)}{\op{X}'_i} + \frac{-i}{\hbar} \comm{\opad{V}(t)\ensuremath{\frac{\dr{}}{\dr{t}}}\op{V}(t)}{\op{X}'_i}.\end{equation} It does not, however, change anything observable, since everything observable is given by the expectation values of observables with respect to \ket{0}, and clearly \begin{equation} \matel{0}{\op{X}'_i}{0}=\matel{0}{\op{X}_i}{0}.\end{equation} To understand the significance of these `quantum gauge transformations', it is useful to consider an analogous example: electromagnetism in the context of the Aharonov-Bohm effect \cite{aharonovbohm}. Recall: the electromagnetic potential \vctr{A} couples to electron wavefunctions via the rule \begin{equation} \op{P}\longrightarrow \op{P}+e\vctr{A}.\end{equation} If an electron beam is split, passed on either side of a solenoid, and recombined, there will be interference between the beams, and as the field in the solenoid is varied the interference fringes will shift by an amount proportional to the line integral of \vctr{A} around the electron's path. This occurs despite the fact that the magnetic field outside the solenoid is zero or nearly so. The A-B effect makes clear that the electromagnetic potential \vctr{A}, and not just the fields \vctr{E} and \vctr{B}, must be regarded as physically significant; however, all observable quantities (including the A-B effect itself) are invariant under gauge transformations \begin{equation} \vctr{A}\longrightarrow \vctr{A}'=\vctr{A}+ \nabla f\end{equation} for arbitrary smooth functions $f$ (along with an associated transformation of the wavefunction). It is generally accepted that the correct response to this observation is to regard gauge-equivalent \vctr{A}s as describing the same physical situation, so as not to burden our theory with massive indeterminism (caused by the possibility of arbitrary \emph{time-dependent} gauge transformations) and with an excess of unobservable properties (caused by the fact that the observable data \emph{right now} only fixes the state up to a gauge transformation). However, this does come with a price: if we identify gauge-equivalent vector potentials then our theory has non-local states in the sense described above. For while the Aharonov-Bohm vector potential cannot be gauge-transformed to zero everywhere, it can be in any region which does not completely enclose the solenoid. Since a region which \emph{does} enclose the solenoid can be decomposed into regions which do not, it follows that whether the solenoid-enclosing region induces an A-B effect is not determined by the properties of its parts. The loop representation of \vctr{A} makes this state non-locality manifest. We replace \vctr{A} with the \emph{loop phases} \begin{equation} C_\gamma = \int_\gamma \vctr{A} \cdot \dr{x}\end{equation} where $\gamma$ is any closed loop. \vctr{A} is fixed up to gauge transformations by the $C_\gamma$, and $\vctr{B}_i$ is given at a point \vctr{x} by the loop phase for an infinitesimal loop in a plane perpendicular to $\vctr{e}_i$. A loop which encloses the solenoid cannot be expressed as the sum of loops which do not enclose the solenoid, so the loop representation has nonlocal states. \section{Lessons for quantum mechanics} The same arguments which lead us to identify gauge-equivalent vector potentials should lead us to identify gauge-equivalent quantum states. Specifically: \begin{enumerate} \item The possibility of time-dependent quantum gauge transformations makes it undetermined which dynamical equations give the true dynamics for the quantum state: is it (\ref{truedyn}) or some (\ref{gaugedyn})? (\ref{truedyn}) is somewhat simpler, but it is unclear whether this is sufficient: after all, in electromagnetism \begin{equation} \Box A_\mu =0\end{equation} is a somewhat simpler choice of dynamics than those given by many other gauges, but this does not lead us to regard it as the `true' dynamics. \item Even time-independent gauge transformations make the state grossly underdetermined by observable data. Provided that $\op{V}\ket{0}=\exp(-i \theta) \ket{0}$, nothing whatever --- no observable data, no theoretical considerations --- can tell us that the physical state is given by $\op{X}_1, \ldots \op{X}_n$ rather than $\opad{V}\op{X}_1\op{V}, \ldots \opad{V}\op{X}_n\op{V}.$ \end{enumerate} (There is also a more `philosophical' concern: in a physical theory we would normally prefer that what is `observable' (\mbox{i.\,e.\,}, the expectation values derived from \ket{0}) would emerge from a physical analysis of measurement, rather than by \emph{fiat}.) This suggests that we should identify Deutsch-Hayden states which differ only by a gauge transformation. But if we do so, we return to the usual representation of quantum states! For two Deutsch-Hayden states are gauge-equivalent if and only if they have the same expectation values --- and of course the expectation values of all possible measurements on a given quantum system are encoded in that system's density operator. So if we do identify gauge-equivalent states, we are again left with a theory whose states are non-local. \section{Conclusion} Deutsch and Hayden's proposal secures locality of states only at the cost of a gauge freedom closely analogous to the gauge freedom of electromagnetism. However, in quantum mechanics as in electromagnetism, to avoid problems of indeterminism and state underdetermination it is necessary to identify gauge-equivalent states. In quantum mechanics as in electromagnetism, if we do make this identification then it leads to nonlocality of states. Deutsch and Hayden argue \cite[p.\,1772]{deutschhayden} that if a theory is local according to any formulation, then it is local period. But their version of quantum mechanics is only a new formulation if we do indeed identify gauge-equivalent states. If not, it is not a `new formulation': it is a new \emph{theory} --- with novel properties such as associating many distinct states to the same in-principle-observable data --- albeit one which has the same observational consequences as the old theory. (Deutsch has himself insisted on this distinction in his more foundational work, for instance in discussing the de Broglie-Bohm interpretation \cite{deutschlockwood}). It is a new theory which is genuinely local, but which pays an unacceptably high price for that locality. We conclude that Deutsch and Hayden's proposal is best understood as a gauge theory whose gauge-independent physical properties are given by the normal quantum formalism. As such, although it may well give important insights into quantum-information issues such as information flow(for a detailed analysis of this point see \cite{timpson}), it does not achieve the goal of showing that quantum mechanics is completely local. Rather, quantum mechanics has only local interactions, but has nonlocal states.
{ "timestamp": "2005-03-16T13:25:40", "yymm": "0503", "arxiv_id": "quant-ph/0503149", "language": "en", "url": "https://arxiv.org/abs/quant-ph/0503149" }
\section{Preliminaries}\label{secprelim} In this article, we are working in the setting of infinite-dimensional differential calculus known as Keller's $C^\infty_c$-theory, based on smooth maps in the sense of Michal-Bastiani (see \cite{BED}, \cite{Ham}, \cite{Mic}, \cite{Mil} for further information). \begin{defn} Let $E$, $F$ be locally convex spaces and $f\!:U\to F$ be a mapping, defined on an open subset~$U$ of~$E$. We say that $f$ is {\em of class~$C^0$\/} if~$f$ is continuous. If~$f$ is a continuous map such that the two-sided directional derivatives \[ df(x,v)=\lim_{t\to 0} {\textstyle \frac{1}{t}\left( f(x+tv)-f(x)\right)} \] exist for all $(x,v)\in U\times E$, and the map $df\!: U\times E\to F$ so defined is continuous, then $f$ is said to be {\em of class~$C^1$\/}. Recursively, given $k\in {\mathbb N}$ we call~$f$ a mapping of class~$C^{k+1}$ if it is of class~$C^1$ and $df$ is of class~$C^k$ on the open subset $U\times E$ of $E\times E$. We set $d^{k+1}f:=d(d^kf)=d^k(df)\!: U\times E^{2^{k+1}-1}\to F$ in this case. The function~$f$ is called {\em smooth\/} (or of class $C^\infty$) if it is of class~$C^k$ for each $k\in {\mathbb N}_0$. \end{defn} \begin{defn} Let $M$ be a finite-dimensional, $\sigma$-compact smooth manifold and~$E$ be a locally convex topological vector space. We equip the vector space $C^\infty(M,E)$ of $E$-valued smooth mappings~$\gamma$ on~$M$ with the topology of uniform convergence of $\partial^\alpha(\gamma\circ \kappa^{-1})$ on compact subsets of~$V$, for each chart $\kappa\!:M\supseteq U\to V\subseteq {\mathbb R}^d$ of~$M$ and multi-index $\alpha\in {\mathbb N}_0^d$ (where $d:=\dim(M)$). Given a compact subset $K\subseteq M$, we equip the vector subspace $C^{\,\infty}_K(M,E):=\{\gamma\in C^\infty(M,E)\!: \gamma|_{M\backslash K}=0\}$ of $C^\infty(M,E)$ with the induced topology. We give $C^\infty_c(M,E):=\bigcup_K C^{\, \infty}_K(M,E)={\displaystyle \lim_{\longrightarrow}}\, C^{\,\infty}_K(M,E)$\vspace{-.8 mm} the locally convex direct limit topology. We abbreviate $C^\infty_c(M):=C^\infty_c(M,{\mathbb R})$, $C^\infty(M):=C^\infty(M,{\mathbb R})$, and $C^{\, \infty}_K(M):= C^{\,\infty}_K(M,{\mathbb R})$. Further details can be found, e.g., in~\cite{GCX}. \end{defn} \section{Example of a discontinuous mapping on {\boldmath $C^\infty_c({\mathbb R})$}}\label{secline} We show that the map $f\!: C^\infty_c({\mathbb R})\to C^\infty_c({\mathbb R})$, $\gamma\mapsto \gamma\circ \gamma-\gamma(0)$ is discontinuous, although its restriction to $C^\infty_{[-n,n]}({\mathbb R})$ is smooth, for each $n\in {\mathbb N}$.\\[2mm] The following fact is essential for our constructions. It follows from \cite[Cor.\,3.13]{KaM} and is also a special case of~\cite[Prop.\,11.3]{ZOO}. For the convenience of the reader, we offer a direct, elementary proof as an appendix. \begin{la}\label{La1} The composition map \[ \Gamma\!: C^\infty({\mathbb R}^n,{\mathbb R}^m)\times C^\infty(M,{\mathbb R}^n)\to C^\infty(M,{\mathbb R}^m)\,,\qquad \Gamma(\gamma,\eta)\, :=\, \gamma\circ \eta \] is smooth, for each finite-dimensional, $\sigma$-compact smooth manifold~$M$ and $m,n\in {\mathbb N}_0$.\nopagebreak\hspace*{\fill}$\Box$ \end{la} For the following proof, recall that the sets \[ {\textstyle {\cal V}(k,e):=\left\{ \gamma\in C^\infty_c({\mathbb R})\!:\; (\forall n\in {\mathbb Z})\, (\forall j\in \{0,\ldots, k_n\})\, (\forall x\in [n-\frac{1}{2},n+\frac{1}{2}])\; |\gamma^{(j)}(x)|<\varepsilon_n\right\}} \] form a basis of open zero-neighbourhoods for the topology on $C^\infty_c({\mathbb R})$, where $k=(k_n)\in ({\mathbb N}_0)^{\mathbb Z}$ and $e=(\varepsilon_n)\in ({\mathbb R}^+)^{\mathbb Z}$ (cf.\ \cite[\S\,II.1]{Sch}; see \cite[Prop.\,4.8]{GCX}). \begin{prop}\label{prototype} $f\!: C^\infty_c({\mathbb R})\to C^\infty_c({\mathbb R})$, $\gamma\mapsto \gamma\circ \gamma-\gamma(0)$ has the following properties: \begin{itemize} \item[\rm (a)] The restriction of~$f$ to a map $C^\infty_{[-n,n]}({\mathbb R})\to C^\infty_c({\mathbb R})$ is smooth $($and hence continuous$)$, for each $n\in {\mathbb N}$. \item[\rm (b)] $f$ is discontinuous at $\gamma=0$. \end{itemize} \end{prop} \begin{proof} (a) Fix $n\in {\mathbb N}$; we have to show that $f|_{C^\infty_{[-n,n]}({\mathbb R})} \!: C^\infty_{[-n,n]}({\mathbb R})\to C^\infty_c({\mathbb R})$ is smooth. The image of this map being contained in the closed vector subspace $C^\infty_{[-n,n]}({\mathbb R})$ of $C^\infty_c({\mathbb R})$, which also is a closed vector subspace of $C^\infty({\mathbb R})$ (with the same induced topology), it suffices to show that the map $C^\infty_{[-n,n]}({\mathbb R})\to C^\infty({\mathbb R})$, $\gamma\mapsto \gamma\circ \gamma-\gamma(0)$ is smooth (see \cite[Prop.\,1.9]{SEC} or \cite[La.\,10.1]{BGN}). Now $\gamma\mapsto \gamma(0)$ being a continuous linear (and thus smooth) map, it suffices to show that $C^\infty_{[-n,n]}({\mathbb R})\to C^\infty({\mathbb R})$, $\gamma\mapsto \gamma\circ \gamma$ is smooth. This readily follows from Lemma~\ref{La1}. (b) Consider the zero-neighbourhood $V:={\cal V}((|n|)_{n\in {\mathbb Z}}, (1)_{n\in {\mathbb Z}})$ in $C^\infty_c({\mathbb R})$. Let $k=(k_n)\in ({\mathbb N}_0)^{\mathbb Z}$ and $e=(\varepsilon_n)\in ({\mathbb R}^+)^{\mathbb Z}$ be arbitrary. We show that $f({\cal V}(k,e))\not\subseteq V$. Since $f(0)=0$, this entails that~$f$ is discontinuous at $\gamma=0$. It is easy to construct a function $h\in C^\infty_c({\mathbb R})$ such that $\mbox{\n supp}(h)\subseteq \; ]{- \frac{1}{2}},\frac{1}{2}[$ and $h(x)=x^{k_0+1}$ for all $x\in [-\frac{1}{4},\frac{1}{4}]$. Then $rh\in {\cal V}(k,e)$ for some $r>0$. For $m\in {\mathbb N}$, we define $h_m\in C^\infty_c({\mathbb R})$ via \[ h_m(x):= \frac{r}{m^{k_0}}h(mx).\] Then $\mbox{\n supp}(h_m)\subseteq \; ]{-\frac{1}{2m}},\frac{1}{2m}[$ and thus $h_m\in {\cal V}(k,e)$ since, for all $j=0,\ldots, k_0$ and $x\in [-\frac{1}{2},\frac{1}{2}]$, we have $|h_m^{(j)}(x)|=\frac{rm^j}{m^{k_0}}|h^{(j)}(mx)|<\varepsilon_0$. We now choose $n\in {\mathbb N}$ such that $n\geq k_0+2$. It is easy to construct a function $\psi\in C^\infty_c({\mathbb R})$ such that $\psi(x)=x-n$ for $x$ in some neighbourhood of~$n$ in~${\mathbb R}$, and $\mbox{\n supp}(\psi)\subseteq \;]n-\frac{1}{2}, n+\frac{1}{2}[$. Then $\phi:=s \cdot \psi\in {\cal V}(k,e)$ for suitable $s>0$. Choosing $s$ small enough, we may assume that $\mbox{\n im}(\phi)\subseteq [-1,1]$. The supports of $\phi$ and $h_m$ being disjoint, we easily deduce from $\phi,h_m\in {\cal V}(k,e)$ that also $\gamma_m:=\phi+ h_m\in {\cal V}(k,e)$. Then $\gamma_m(0)=0$, and since $\mbox{\n im}(\phi)\subseteq [-1,1]$, we have $f(\gamma_m)(x) = (h_m\circ \phi)(x)$ for all $x\in W:=\;]n-\frac{1}{2}, n+\frac{1}{2}[$. For $x\in W$ sufficiently close to~$n$, we have $\phi(x)=s\cdot (x-n)\in [-\frac{1}{4m},\frac{1}{4m}]$ and thus $f(\gamma_m)(x)= r\cdot m\cdot s^{k_0+1}\cdot (x-n)^{k_0+1}$, whence $f(\gamma_m)^{(k_0+1)}(n)=r\cdot m\cdot s^{k_0+1} \cdot(k_0+1)!\,$. Thus $f(\gamma_m)\not\in V$ for all $m\in {\mathbb N}$ such that $r\cdot m\cdot s^{k_0+1}\cdot (k_0+1)!\geq 1$, and so $f({\cal V}(k,e))\!\not\subseteq \!V$. As $k$ and $e$ were arbitrary, (b)\,follows. \end{proof} Note that $\mbox{\n supp}(f(\gamma))\subseteq \mbox{\n supp}(\gamma)$ here, for all $\gamma\in C^\infty_c({\mathbb R})$. \begin{rem} Although the map~$f$ from Proposition~\ref{prototype} is discontinuous and thus not smooth in the Michal-Bastiani sense, it is easily seen to be smooth in the sense of convenient differential calculus (as any map~$f$ on a ``regular'' countable strict direct limit $E={\displaystyle \lim_{\longrightarrow}}\, E_n$\vspace{-1.3 mm} of complete locally convex spaces, all of whose restrictions $f|_{E_n}$ are smooth).\footnote{Regularity means that every bounded subset of~$E$ is contained and bounded in some~$E_n$.} \end{rem} \section{Discontinuous mappings on {\boldmath $C^\infty_c(M,E)$}}\label{gencase} In this section, we generalize our discussion of $C^\infty_c({\mathbb R})$ from Section~\ref{secline} to the spaces $C^\infty_c(M,E)={\displaystyle \lim_{\longrightarrow}}\, C^\infty_K(M,E)$\vspace{-.8 mm} of compactly supported smooth mappings on a $\sigma$-compact finite-dimensional smooth manifold~$M$ with values in a locally convex space~$E$. We show: \begin{prop}\label{propzero} If $E\not=\{0\}$, the manifold $M$ is non-compact, and $\dim(M)>0$, then there exists a mapping $f\!: C^\infty_c(M,E)\to C^\infty_c(M,{\mathbb R})$ such that \begin{itemize} \item[\rm (a)] The restriction of $f$ to $C^\infty_K(M,E)$ is smooth, for each compact subset $K$ of~$M$. \item[\rm (b)] $f$ is discontinuous at~$0$. \end{itemize} In particular, the locally convex direct limit topology on $C^\infty_c(M,E)={\displaystyle \lim_{\longrightarrow}}\, C^\infty_K(M,E)\vspace{-.8 mm}$ is properly coarser than the topology making $C^\infty_c(M,E)$ the direct limit of the spaces $C^\infty_K(M,E)$ in the category of topological spaces. \end{prop} Instead of proving this proposition directly, we establish an analogous result for spaces of sections in bundles of locally convex spaces, which is no harder to prove. Noting that the function space $C^\infty_c(M,E)$ is topologically isomorphic to the space $C^\infty_c(M,M\times E)$ of compactly supported smooth sections in the trivial bundle $\mbox{\rm pr}_M\!:M\times E\to M$, clearly Proposition~\ref{propzero} is covered by the ensuing discussions for vector bundles. For background material concerning bundles of locally convex spaces and the associated spaces of sections, the reader is referred to~\cite{SEC} (or also \cite[Appendix~F]{ZOO}). For the present purposes, we recall: if $\pi\!: E\to M$ is a smooth bundle of locally convex spaces over the finite-dimensional, $\sigma$-compact smooth manifold~$M$, with typical fibre the locally convex space~$F$, then one considers on the space $C^\infty(M,E)$ of all smooth sections the initial topology with respect to the family of mappings \[ \theta_\psi\!: C^\infty(M,E)\to C^\infty(U,F),\;\;\; \theta_\psi(\sigma):=\sigma_\psi:=\mbox{\rm pr}_F\circ \psi\circ \sigma|_U^{\pi^{-1}(U)} \, , \] which take a smooth section~$\sigma$ to its local representation $\sigma_\psi\!: U\to F$ with respect to the local trivialization $\psi\!: \pi^{-1}(U)\to U\times F$ of~$E$. Given a compact subset~$K\subseteq M$, the subspace $C^{\,\infty}_K(M,E)\subseteq C^\infty(M,E)$ of sections vanishing off~$K$ is equipped with the induced topology, and $C^\infty_c(M,E):=\bigcup_K C^{\, \infty}_K(M,E)={\displaystyle \lim_{\longrightarrow}} \, C^{\,\infty}_K(M,E)$\vspace{-1.3 mm} is given the locally convex direct limit topology. \begin{thm}\label{fingeruebung} Let $M$ be a $\sigma$-compact, non-compact, finite-dimensional smooth manifold of dimension $\dim(M)>0$, and $\pi\!: E\to M$ be a smooth bundle of locally convex spaces over~$M$, whose typical fibre is a locally convex topological vector space~$F\not=\{0\}$. Then there exists a discontinuous mapping $f\!: C^\infty_c(M,E)\to C^\infty_c(M,{\mathbb R})$ whose restriction to $C^{\,\infty}_K(M,E)$ is smooth, for each compact subset~$K$ of~$M$. \end{thm} \begin{proof} Let $d:=\dim(M)$. Since~$M$ is non-compact, there exists a sequence $(U_n)_{n\in {\mathbb N}_0}$ of mutually disjoint coordinate neighbourhoods $U_n\subseteq M$ diffeomorphic to ${\mathbb R}^d$ such that local trivializations $\psi_n\!:\pi^{-1}(U_n)\to U_n\times F$ of~$E$ exist, and such that every compact subset of~$M$ meets only finitely many of the sets~$U_n$. We define \[ \theta_{\psi_n}\!: C^\infty_c(M,E)\to C^\infty(U_n,F),\;\;\;\; \theta_{\psi_n}(\sigma):=\sigma_{\psi_n}:= \mbox{\rm pr}_F\circ \psi_n\circ \sigma|_{U_n}^{\pi^{-1}(U_n)}\,. \] By definition of the topology on $C^\infty_c(M,E)$, the linear maps $\theta_{\psi_n}$ are continuous. For each $n\in{\mathbb N}_0$, let $\kappa_n\!: U_n\to {\mathbb R}^d$ be a $C^\infty$-diffeomorphism; define $x_n:=\kappa_n^{-1}(0)$. We choose a function $h\in C^\infty_c({\mathbb R}^d,{\mathbb R})$ such that $h|_{[-1,1]^d}=1$; we define $h_n\in C^\infty_c(M,{\mathbb R})$ via $h_n(x):=h(\kappa_n(x))$ if $x\in U_n$, $h_n(x):=0$ if $x\in M\,\backslash\, U_n$. Let $K_n:=\mbox{\n supp}(h_n)\subseteq U_n$. We choose a continuous linear functional $0\not= \lambda\in F'$, and pick $v\in F$ such that $\lambda(v)=1$. Note that $A:=\bigcup_{n\in {\mathbb N}} K_n$ is closed in~$M$, the sequence $(K_n)_{n\in{\mathbb N}}$ of compact sets being locally finite. Let $\mu\!:{\mathbb R}\times F\to F$ be the scalar multiplication. The eventual definition of the mapping~$f$ we are looking for will involve the map $\Phi\!: E\to M\times {\mathbb R}$, defined via \begin{equation}\label{dfn1} \Phi|_{\pi^{-1}(U_n)}:= (\pi|_{\pi^{-1}(U_n)}, \lambda\circ \mu\circ ((h_n\circ \pi)|_{\pi^{-1}(U_n)}, \mbox{\rm pr}_F\circ \psi_n)) \end{equation} for $n\in {\mathbb N}$, and $\Phi|_{E\backslash \pi^{-1}(A)}:=(\pi|_{E\backslash \pi^{-1}(A)}, 0)$. Note that $\Phi$ is well-defined as the function in Equation\,(\ref{dfn1}) coincides with $(\pi,0)$ on the set $\bigcup_{n\in {\mathbb N}} \pi^{-1}(U_n \,\backslash\, A)$. Also note that $\Phi$ is a fibre-preserving mapping from~$E$ into the trivial bundle $M\times {\mathbb R}$. Furthermore, it is readily verified that~$\Phi$ is a smooth. By \cite[Thm.\,5.9]{SEC} (or \cite[Rem.\,F.25\,(a)]{ZOO}), the pushforward \[ C^\infty_c(M,\Phi)\!: C^\infty_c(M,E)\to C^\infty_c(M,M\times {\mathbb R}),\;\;\;\; \sigma\mapsto \Phi\circ \sigma\] is smooth. For later use, we introduce the continuous linear map \[ \Lambda\; :=\; \theta_{\text{id}_{M\times {\mathbb R}}}\!: C^\infty_c(M,M\times {\mathbb R})\to C^\infty(M,{\mathbb R})\, . \] Let $\iota\!:{\mathbb R}\to {\mathbb R}^d$ denote the embedding $t\mapsto (t,0,\ldots, 0)$. The mapping~$f$ to be constructed will also involve the map $\Psi\!: C^\infty_c(M,E)\to C^\infty({\mathbb R},{\mathbb R})$ defined via \[ \Psi:= C^\infty({\mathbb R},\lambda) \circ C^\infty(\kappa_0^{-1}\circ \iota ,F)\circ \theta_{\psi_0}, \] where the pullback $C^\infty(\kappa_0^{-1}\circ \iota, F)\!: C^\infty(U_n,F)\to C^\infty({\mathbb R},F)$, $\gamma\mapsto \gamma\circ \kappa_0^{-1}\circ \iota$ and the pushforward $C^\infty({\mathbb R},\lambda)\!: C^\infty({\mathbb R},F)\to C^\infty({\mathbb R}, {\mathbb R})$, $\gamma\mapsto \lambda\circ \gamma$ are continuous linear mappings and thus smooth, by \cite[La.\,3.3, La.\,3.7]{GCX}. Being a composition of smooth maps, $\Psi$ is smooth. We now define the desired map $f\!:C^\infty_c(M,E)\to C^\infty_c(M,{\mathbb R})$ via \[ f\, :=\, \Gamma\circ (\Psi,\Lambda\circ C^\infty_c(M,\Phi)) \; - \;\lambda\circ \mbox{\n ev}_{x_0}\circ \theta_{\psi_0} \] (co-restricted from $C^\infty(M,{\mathbb R})$ to $C^\infty_c(M,{\mathbb R})$), where \[ \Gamma\!: C^\infty({\mathbb R},{\mathbb R})\times C^\infty(M,{\mathbb R})\to C^\infty(M,{\mathbb R}),\;\;\;\; \Gamma(\gamma,\eta):=\gamma\circ \eta\] denotes composition, and $\mbox{\n ev}_{x_0}\!: C^\infty(U_0,F)\to F$ the evaluation map $\gamma\mapsto \gamma(x_0)$. Here $\lambda\circ \mbox{\n ev}_{x_0}\circ \theta_{\psi_0}$ is a continuous linear map and thus smooth. Explicitly, for $\sigma\in C^\infty_c(M,E)$ \begin{eqnarray*} f(\sigma)(x) &=& \Big(\lambda\circ\sigma_{\psi_0}\circ\kappa_0^{-1}\circ\iota\Big) \bigl( \lambda\bigl( h_n(x)\, \sigma_{\psi_n}(x)\bigr)\bigr)\\ &=& \lambda\Big( \sigma_{\psi_0} \bigl(\kappa_0^{-1}( h_n(x)\cdot \lambda(\sigma_{\psi_n}(x)),\; 0)\bigr)\Big) \;-\;\lambda(\sigma_{\psi_0}(x_0)) \end{eqnarray*} if $x\in U_n$ ($n\in{\mathbb N}$), whereas $f(\sigma)(x)=0$ if $x\in M\,\backslash\, A$.\vspace{1.3mm} {\em Claim\/}: {\em The restriction of~$f$ to $C^\infty_K(M,E)$ is smooth, for each compact subset~$K$ of~$M$.}\\ To see this, note that $f(C^{\,\infty}_K(M,E)\subseteq C^{\,\infty}_K(M,{\mathbb R})$, where $C^{\, \infty}_K(M,{\mathbb R})$ is a closed vector subspace of $C^\infty(M,{\mathbb R})$ and $C^\infty_c(M,{\mathbb R})$. Thus, it suffices to show that $f|_{C^\infty_K(M,E)}$ is smooth as a map into $C^\infty(M,{\mathbb R})$ (\cite[Prop.\,1.9]{SEC}, or \cite[La.\,10.1]{BGN}). But this follows from the Chain Rule, as $\Gamma$ is smooth by Lemma~\ref{La1} and also the other constituents of~$f$ are smooth.\vspace{1.3mm} {\em Claim\/}: {\em $f$ is discontinuous at the zero-section $\sigma=0$.} To see this, consider the set~$V$ of all $\gamma\in C^\infty_c(M,{\mathbb R})$ such that, for all $n\in {\mathbb N}$ and multi-indices $\alpha\in {\mathbb N}_0^d$ of order $|\alpha|\leq n$, we have $|\partial^\alpha(\gamma\circ \kappa_n^{-1})(0)|<1$. It is easily verified that $V$ is a symmetric, convex zero-neighbourhood in $C^\infty_c(M,{\mathbb R})$. Let $U$ be any convex zero-neighbourhood in $C^\infty_c(M,E)$; we claim that $f(U)\not\subseteq V$. To see this, set $L_n:= \kappa_n^{-1}([-1,1]^d)$ for $n\in {\mathbb N}_0$. Then \[ \rho_n\!: C^\infty_{L_n}(M,E)\to C^\infty_{[-1,1]^d}({\mathbb R}^d,F),\;\;\;\; \sigma\mapsto \sigma_{\psi_n}\circ \kappa_n^{-1} \] is a topological isomorphism (cf.\ \cite[La.\,3.9, La.\,3.10]{SEC} or \cite[La.\,F.9, La.\,F.15]{ZOO}) whose inverse gives rise to a topological embedding $j_n\!: C^\infty_{[-1,1]^d}({\mathbb R}^d,F) \to C^\infty_c(M,E)$. The linear mapping $\phi\!:{\mathbb R}\to F$, $t\mapsto tv$ gives rise to a continuous linear map $C^\infty_{[-1,1]^d}({\mathbb R}^d,\phi)\!:C^\infty_{[-1,1]^d}({\mathbb R}^d,{\mathbb R}) \to C^\infty_{[-1,1]^d}({\mathbb R}^d,F)$, $\gamma\mapsto \phi\circ \gamma$. Then $W_n:=(j_n\circ C^\infty_{[-1,1]^d}({\mathbb R}^d,\phi))^{-1}(\frac{1}{2}U)$ is a convex zero-neighbourhood in $C^\infty_{[-1,1]^d}({\mathbb R}^d,{\mathbb R})$. Thus, there exists $k_n\in {\mathbb N}_0$ and $\varepsilon_n>0$ such that $W_{k_n,\varepsilon_n}\subseteq W_n$, where $W_{k_n,\varepsilon_n}$ is the set of all $\gamma\in C^\infty_{[-1,1]^d}({\mathbb R}^d,{\mathbb R})$ such that $\sup\{|\partial^\alpha \gamma(x)|\!: x\in [-1,1]^d\}<\varepsilon_n$ for all $\alpha\in {\mathbb N}_0^d$ such that $|\alpha|\leq k_n$. We let $g\in C^\infty_{[-1,1]^d}({\mathbb R}^d,{\mathbb R})$ be a function such that $g(y_1,\ldots,y_d)=y_1^{{k_0}+1}$ for all $y=(y_1,\ldots, y_d)\in [-\frac{1}{2},\frac{1}{2}]^d$. Then $rg\in W_{k_0,\varepsilon_0}$ for some $r>0$. It is clear from the definition of $W_{k_0,\varepsilon_0}$ that then also $\gamma_m\in W_{k_0,\varepsilon_0}$ for all $m\in {\mathbb N}$, where \[ \gamma_m\!:{\mathbb R}^d\to{\mathbb R}\, ,\quad \gamma_m(y_1,\ldots,y_d):= \frac{r}{m^{k_0}}\,g(my_1,y_2,\ldots, y_d)\, . \] Thus $\tau_m:= j_0(\phi\circ \gamma_m)\in \frac{1}{2}U$. Let $\ell:=k_0+1$; we easily find $\eta\in W_{k_\ell,\varepsilon_\ell}$ such that, for suitable $s>0$, we have $\eta(y)=s\cdot y_1$ for $y=(y_1,\ldots, y_d)$ in some zero-neighbourhood in~${\mathbb R}^d$. We define $\tau:=j_\ell(\phi\circ \eta)\in \frac{1}{2}U$. Then $\sigma_m:=\tau_m+\tau\in U$ by convexity of~$U$. Consider $g_m:=f(\sigma_m)\circ \kappa_\ell^{-1}\!: {\mathbb R}^d\to {\mathbb R}$. For $y\in [-1,1]^d$ sufficiently close to~$0$, we have $\eta(y)=sy_1$ and $m|\eta(y)|\leq\frac{1}{2}$. Thus \[ g_m(y)=\gamma_m(\eta(y),0,\ldots,0) =r\cdot m\cdot s^{k_0+1}\cdot y_1^{k_0+1}, \] entailing that $\frac{\partial^{k_0+1} g_m}{\partial y_1^{k_0+1}}(0) =r\cdot m\cdot s^{k_0+1}\cdot (k_0+1)!\,$. Hence $f(\sigma_m)\not\in V$ for each $m\in {\mathbb N}$ such that $r\cdot m\cdot s^{k_0+1}\cdot (k_0+1)!\geq 1$. We have shown that $f(U)\not\subseteq V$ for any $0$-neighbourhood~$U$ in $C^\infty_c(M,E)$, although $f(0)=0$. Thus $f$ is discontinuous at $\sigma=0$. \end{proof} \section{Further examples}\label{secmisc} We describe various pathological bilinear mappings. \begin{prop} Let ${\mathbb K}\in\{{\mathbb R},{\mathbb C}\}$. The pointwise multiplication map \[ \mu\!:C^\infty({\mathbb R},{\mathbb K})\times C^\infty_c({\mathbb R},{\mathbb K})\to C^\infty_c({\mathbb R},{\mathbb K}),\;\;\;\; \mu(\gamma,\eta):=\gamma\cdot\eta \] is a hypocontinuous bilinear $($and thus sequentially continuous$)$ mapping on the locally convex direct limit \[ C^\infty({\mathbb R},{\mathbb K})\times C^\infty_c({\mathbb R},{\mathbb K})= {\displaystyle \lim_{\longrightarrow}}\, (C^\infty({\mathbb R},{\mathbb K})\times C^\infty_{[-n,n]}({\mathbb R},{\mathbb K}))\, ,\] whose restriction to $C^\infty({\mathbb R},{\mathbb K})\times C^\infty_{[-n,n]}({\mathbb R},{\mathbb K})$ is continuous bilinear and thus ${\mathbb K}$-analytic, for each $n\in {\mathbb N}$. However, $\mu$ is discontinuous. \end{prop} \begin{proof} Using the Leibniz Rule for the differentiation of products of functions, it is easily verified that~$\mu$ is separately continuous.\footnote{Alternatively, we can obtain the assertion as a special case of \cite[Cor.\,2.7]{SEC} or \cite[La.\,4.5\,(a) and Prop.\,4.19\,(d)]{ZOO}, combined with the locally convex direct limit property.} The spaces $C^\infty({\mathbb R},{\mathbb K})$ and $C^\infty_c({\mathbb R},{\mathbb K})$ being barrelled, this entails that~$\mu$ is hypocontinuous and thus sequentially continuous \cite[Thm.\,41.2]{Tre}. The restriction of~$\mu$ to $C^\infty({\mathbb R},{\mathbb K})\times C^\infty_{[-n,n]}({\mathbb R},{\mathbb K})$ is a sequentially continuous bilinear mapping on a product of metrizable spaces and therefore continuous. To see that $\mu$ is discontinuous, consider the zero-neighbourhood \[ W:=\{\gamma\in C^\infty_c({\mathbb R},{\mathbb K})\!: \; (\forall x\in {\mathbb R})\;\,|\gamma(x)|<1\} \] in $C^\infty_c({\mathbb R},{\mathbb K})$. If $U$ is any zero-neighbourhood in $C^\infty({\mathbb R},{\mathbb K})$ and $V$ any zero-neighbourhood in $C^\infty_c({\mathbb R},{\mathbb K})$, then there exists a compact subset $K$ of~${\mathbb R}$ such that \[ ( \forall \gamma\in C^\infty({\mathbb R},{\mathbb K}))\;\;\; \gamma|_K=0\; \Rightarrow\; \gamma\in U. \] Pick any $x_0\in {\mathbb R}\, \backslash\, K$. There is a function $\phi\in C^\infty_c({\mathbb R},{\mathbb K})$ such that $\phi(x_0)\not=0$ and $\mbox{\n supp}(\phi)\subseteq {\mathbb R}\,\backslash\, K$. Then $r\phi\in V$ for some $r>0$, and $t\phi\in U$ for all $t\in {\mathbb R}$. Choosing $t\geq \frac{1}{r\cdot|\phi(x_0)|^2}$, we have $(t\phi,r\phi)\in U\times V$ but $|\mu(r\phi,t\phi)(x_0)|=rt|\phi(x_0)|^2\geq 1$, entailing that $\mu(U\times V)\not\subseteq W$. Thus $\mu$ is discontinuous at~$(0,0)$. \end{proof} Another instructive example is the following (compare also the examples in \cite{DaW}): \begin{example} Let $E_1\subset E_2\subset \cdots$ be a strictly ascending sequence of Banach spaces, such that $E_{n+1}$ induces the given topology on~$E_n$. Set $E:={\displaystyle \lim_{\longrightarrow}}\, E_n$\vspace{-.8mm} and $F:=E'_b$. For example, we can take $E_n:=L^2[-n,n]$, in which case $E=L^2_{\mbox{\n \footnotesize comp}}({\mathbb R})$ and $F=L^2_{\mbox{\n \footnotesize loc}}({\mathbb R})={\displaystyle \lim_{\longleftarrow}}\, L^2[-n,n]$. Then $A_n:=F\times E_n\times {\mathbb K}\times {\mathbb K}$ is a Fr\'{e}chet space (and reflexive in the example $E_n=L^2[-n,n]$). The evaluation map $E_n'\times E_n\to {\mathbb R}$ being continuous as~$E_n$ is a Banach space, it is easy to see that~$A_n$ becomes a unital associative topological algebra via \begin{equation}\label{formull} (\lambda_1,x_1,z_1,c_1)\cdot(\lambda_2,x_2,z_2,c_2) \,:=\, \bigl( c_1\lambda_2+c_2\lambda_1,\, c_1x_2+c_2x_1,\,c_1z_2+ \lambda_1(x_2)+z_1c_2, \, c_1c_2 \bigr)\, . \end{equation} The multiplication can be visualized by considering $(\lambda,x,z,c)\!\in \!A_n$ as the 3-by-3 matrix {\scriptsize \[ \left( \begin{array}{ccc} c & \lambda & z\\ 0 & c & x \\ 0 & 0 & c \end{array} \right). \] }The topological algebras~$A_n$ are very well-behaved: they have open groups of units, and inversion is a ${\mathbb K}$-analytic map. We can also use Formula\,(\ref{formull}) to define a multiplication map $\mu\!: A\times A\to A$ turning the direct limit locally convex space $A:=F\times E\times {\mathbb K}\times {\mathbb K}={\displaystyle \lim_{\longrightarrow}}\, A_n$\vspace{-.8 mm} into a unital, associative algebra. However, although the restriction of~$\mu$ to $A_n\times A_n$ is a continuous bilinear map for each~$n\in{\mathbb N}$, $\mu\!: A\times A={\displaystyle \lim_{\longrightarrow}}\, (A_n\times A_n)\to A$\vspace{-1.3 mm} is discontinuous (since the evaluation map $E'_b\times E\to{\mathbb R}$ is discontinuous, the space~$E$ not being normable). We refer to \cite[Section~10]{Glo} for more details. \end{example}
{ "timestamp": "2005-03-18T20:04:28", "yymm": "0503", "arxiv_id": "math/0503387", "language": "en", "url": "https://arxiv.org/abs/math/0503387" }
\section{\bf Introduction} Invariant structures on homogeneous manifolds are traditionally one of the most important objects in differential geometry, specifically, in Hermitian geometry. Some remarkable classes of almost Hermitian structures such as K\"ahler, nearly K\"ahler, Hermitian structures etc. are well known and intensively used in geometry and a number of applications. In particular, a special role is played by a significant class of invariant nearly K\"ahler structures based on the canonical almost complex structure on homogeneous $3$-symmetric spaces (see \cite{S2}, \cite{WG}, \cite{G2}, \cite{Ki1}). It should be mentioned that the canonical almost complex structure on such spaces became an effective tool and a remarkable example in some deep constructions of differential geometry and global analysis such as homogeneous structures (\cite{TV}, \cite{Sat}, \cite{Ki4}, \cite{GV}, \cite{LV}, \cite{AG} etc.), Einstein metrics (\cite{SW}, \cite{SY}), holomorphic and minimal submanifolds (\cite{Sal1}, \cite{Sal2}), real Killing spinors (\cite{Gru}, \cite{BFGK}, \cite{Ka}). The concept of generalized Hermitian geometry created in the 1980s (see, for example, \cite{Ki2}, \cite{Ki7}) is a natural consequence of the development of Hermitian geometry and the theory of almost contact structures with many applications. One of its central objects is the metric $f$-structures of the classical type $(f^3+f=0)$, which include the class of almost Hermitian structures. Many important classes of metric $f$-structures such as K\"ahler, Killing, nearly K\"ahler, Hermitian $f$-structures and some others were introduced and intensively investigated in various aspects (see \cite{Ki2}, \cite{Ki5}, \cite{Ki7}, \cite{KL} etc.). Specifically, Killing and nearly K\"ahler $f$-structures became natural generalizations of classical nearly K\"ahler structures in Hermitian geometry. However, this theory had not provided new invariant examples of its own up to the recent period, and so the lack of these examples was becoming all the more noticeable. There has recently been a qualitative change in the situation, related to the complete solution of the problem of describing canonical structures of classical type on regular $\Phi$-spaces \cite{BS2}. A rich collection of canonical $f$-structures has been discovered (including almost complex structures) leading to the presentation of wide classes of invariant examples in generalized Hermitian geometry (see \cite{B4}-\cite{B7}, \cite{C3} and others). In particular, nearly K\"ahler $f$-structures were provided with a remarkable class of their own invariant examples (see \cite{B6}, \cite{B7}). This has ensured a continuation of the classical results of J.A.Wolf, A.Gray, V.F.Kirichenko and others. As to Killing $f$-structures, it is really an essential problem to find proper non-trivial invariant examples of these structures. Moreover, the possibilities for constructing such examples are fairly limited (see \cite{B4}). The main goals of this paper are (i) to give a brief survey on invariant structures in generalized Hermitian geometry and (ii) to characterize all invariant $f$-structures on the flag manifold $SU(3)/T_{max}$ in the sense of generalized Hermitian geometry, in particular, to present first invariant examples of Killing $f$-structures. Sections 2-4 are mostly of survey character. In Section 2, we collect some basic notions and results on homogeneous regular $\Phi$-spaces and canonical affinor structures. In particular, a precise description of all canonical structures of classical types on homogeneous $k$-symmetric spaces is included. Besides, the exact formulae for these structures and the relationship between them on 4- and 5-symmetric spaces are presented. In Section 3, we recall the main classes of almost Hermitian structures following the Gray-Hervella division of almost Hermitian manifolds into sixteen classes (see \cite{GH}). Besides, we select particular results related to invariant almost Hermitian structures. Further, in Section 4, we describe main classes of metric $f$-structures in generalized Hermitian geometry. Here we also formulate the recent results on invariant nearly K\"ahler, $G_1f$-, Hermitian, and Killing $f$-structures. In this consideration, the canonical $f$-structures on homogeneous 4- and 5-symmetric spaces are especially important. Finally, in Section 5, we examine in detail all invariant $f$-structures on the complex flag manifold $SU(3)/T_{max}$ with respect to all invariant Riemannian metrics. We discuss belonging these structures to the main classes of metric $f$-structures above mentioned. In particular, invariant non-trivial Killing $f$-structures together with the corresponding Riemannian metrics are first presented. \section{\bf Homogeneous regular $\Phi$-spaces and canonical affinor structures} Here we briefly formulate some basic definitions and results related to regular $\Phi$-spaces and canonical affinor structures on them. More detailed information can be found in \cite{BS2}, \cite{B10}, \cite{WG}, \cite{Ko}, \cite{F}, \cite{S1}, \cite{S2}. Let $G$ be a connected Lie group, $\Phi$ its (analytic) automorphism. Denote by $G^{\Phi}$ the subgroup of all fixed points of $\Phi$ and $G_o^{\Phi}$ the identity component of $G^{\Phi}$. Suppose a closed subgroup $H$ of $G$ satisfies the condition $$G_o^{\Phi}\subset{H}\subset{G^{\Phi}}.$$ Then $G/H$ is called a {\it homogeneous $\Phi$-space}. Homogeneous $\Phi$-spaces include homogeneous symmetric spaces $(\Phi^2=id)$ and, more general, {\it homogeneous $\Phi$-spaces of order $k$} $(\Phi^k=id)$ or, in the other terminology, {\it homogeneous $k$-symmetric spaces} (see \cite{Ko}). For any homogeneous $\Phi$-space $G/H$ one can define the mapping $$ S_o = D\colon\ G/H \to G/H,\ xH\to \Phi (x) H. $$ It is known \cite{S1} that $S_o$ is an analytic diffeomorphism of $G/H$. $S_o$ is usually called a "symmetry" of $G/H$ at the point $o=H$. It is evident that in view of homogeneity the "symmetry" $S_p$ can be defined at any point $p\in G/H$. More exactly, for any $p=\tau(x)o=xH,\ q=\tau(y)o=yH$ we put $$ S_p=\tau(x)\circ S_o \circ \tau(x^{-1}). $$ It is easy to show that $$ S_p(yH)=x\Phi(x^{-1})\Phi(y)H. $$ Thus any homogeneous $\Phi$-space is equipped with the set of symmetries $\{S_p\mid p\in G/H\}$. Moreover, each $S_p$ is an analytic diffeomorphism of the manifold $G/H$ (see \cite{S1}). Note that there exist homogeneous $\Phi$-spaces that are not reductive. That is why so-called regular $\Phi$-spaces first introduced by N.A.Stepanov \cite{S1} are of fundamental importance. Let $G/H$ be a homogeneous $\Phi$-space, $\mathfrak{g}$ and $\mathfrak{h}$ the corresponding Lie algebras for $G$ and $H$, $\varphi=d{\Phi}_e$ the automorphism of $\mathfrak{g}$. Consider the linear operator $A=\varphi-id$ and the Fitting decomposition $\mathfrak{g}=\mathfrak{g}_0\oplus\mathfrak{g}_1$ with respect to $A$, where $\mathfrak{g}_0$ and $\mathfrak{g}_1$ denote $0$- and $1$-component of the decomposition respectively. Further, let $\varphi=\varphi_{s}\:\varphi_{u}$ be the Jordan decomposition, where $\varphi_{s}$ and $\varphi_{u}$ is a semisimple and unipotent component of $\varphi$ respectively, $\varphi_{s}\:\varphi_{u}=\varphi_{u}\:\varphi_{s}$. Denote by $\mathfrak{g}^{\gamma}$ a subspace of all fixed points for a linear endomorphism $\gamma$ in $\mathfrak{g}$. It is clear that $\mathfrak{h}=\mathfrak{g}^{\varphi}=Ker\,A$, $\mathfrak{h}\subset\mathfrak{g}_0$, $\mathfrak{h}\subset\mathfrak{g}^{\varphi_s}$. {\bf\noindent Definition 1} {\rm (\cite{S1}, \cite{BS2}, \cite{B10}, \cite{F}). A homogeneous $\Phi$-space $G/H$ is called a {\it regular $\Phi$-space} if one of the following equivalent conditions is satisfied: \begin{enumerate} \item $\mathfrak{h}=\mathfrak{g}_0$; \item $\mathfrak{g}=\mathfrak{h}\oplus{A}\mathfrak{g}$; \item The restriction of the operator $A$ to ${A}\mathfrak{g}$ is non-singular; \item $A^2X=0\Longrightarrow{A}X=0$ for all $X\in\mathfrak{g}$. \item The matrix of the automorphism $\varphi$ can be represented in the form \\ $\left(\begin{array}{cc} E & 0 \\ 0 & B \end{array}\right),$ where the matrix $B$ does not admit the eigenvalue $1$. \item $\mathfrak{h}=\mathfrak{g}^{\varphi_s}$. \end{enumerate}} \noindent We recall two basic facts: \newpage \begin{theorem}{\rm(\cite{S1})} \begin{itemize} \item Any homogeneous $\Phi$-space of order $k$ $(\Phi^k\ = \ id)$ is a regular $\Phi$-space. \item Any regular $\Phi$-space is reductive. More exactly, the Fitting decomposition \begin{equation}\label{f1} \mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m}, \ \mathfrak{m}=A\mathfrak{g} \end{equation} is a reductive one. \end{itemize} \end{theorem} Decomposition (\ref{f1}) is the {\it canonical reductive decomposition} corresponding to a regular $\Phi$-space $G/H$, and $\frak{m}$ is the {\it canonical reductive complement}. We also note that for any regular $\Phi$-space $G/H$ each point $p=xH\in G/H$ is an isolated fixed point of the symmetry $S_p$ (see \cite{S1}). Decomposition (\ref{f1}) is obviously $\varphi$-invariant. Denote by $\theta$ the restriction of $\varphi$ to $\mathfrak{m}$. As usual, we identify $\mathfrak{m}$ with the tangent space $T_o(G/H)$ at the point $o=H$. It is important to note that the operator $\theta$ commutes with any element of the linear isotropy group $Ad(H)$ (see \cite{S1}). It also should be noted (see \cite{S1}) that $$ (dS_o)_o=\theta. $$ An {\it affinor structure} on a manifold is known to be a tensor field of type $(1,1)$ or, equivalently, a field of endomorphisms acting on its tangent bundle. Suppose $F$ is an invariant affinor structure on a homogeneous manifold $G/H$. Then $F$ is completely determined by its value $F_o$ at the point $o$, where $F_o$ is invariant with respect to $Ad(H)$. For simplicity, we will denote by the same manner both any invariant structure on $G/H$ and its value at $o$ throughout the rest of the paper. {\bf\noindent Definition 2} {\rm (\cite{BS1},\cite{BS2}). An invariant affinor structure $F$ on a regular $\Phi$-space $G/H$ is called {\em canonical} if its value at the point $o=H$ is a polynomial in $\theta$. Denote by $\mathcal A(\theta)$ the set of all canonical affinor structures on a regular $\Phi$-space $G/H$. It is easy to see that $\mathcal A(\theta)$ is a commutative subalgebra of the algebra $\mathcal A$ of all invariant affinor structures on $G/H$. Moreover, $$dim\ \mathcal A(\theta)=deg\ \nu\ \le\ dim\ G/H,$$ where $\nu$ is a minimal polynomial of the operator $\theta$. It is evident that the algebra $\mathcal A(\theta)$ for any symmetric $\Phi$-space $(\Phi^2=id)$ consists of scalar structures only, i.e. it is isomorphic to $\mathbb{R}$. As to arbitrary regular $\Phi$-space $(G/H,\Phi)$, the algebraic structure of its commutative algebra $\mathcal A(\theta)$ has been recently completely described (see \cite{B9}). It should be mentioned that all canonical structures are, in addition, invariant with respect to the "symmetries" $\{S_p\}$ of $G/H$(see \cite{S1}). Moreover, from $(dS_o)_o=\theta$ it follows that the invariant affinor structure $F_p=(dS_p)_p, p\in G/H$ generated by the symmetries $\{S_p\}$ belongs to the algebra $\mathcal A(\theta)$. The most remarkable example of canonical structures is the canonical almost complex structure $J=\frac{1}{\sqrt{3}}(\theta-\theta^2)$ on a homogeneous $3$-symmetric space (see \cite{S2}, \cite{WG}, \cite{G2}). It turns out that it is not an exception. In other words, the algebra $\mathcal A(\theta)$ contains many affinor structures of classical types. In the sequel we will concentrate on the following affinor structures of classical types: {\it almost complex structures} $J$ $(J^2=-1)$; {\it almost product structures} $P$ $(P^2=1)$; {\it $f$-structures} $(f^3+f=0)$ \cite{Y}; $f$-structures of hyperbolic type or, briefly, {\it $h$-structures} $(h^3-h=0)$ \cite{Ki2}. \\ Clearly, $f$-structures and $h$-structures are generalizations of structures $J$ and $P$ respectively. All the canonical structures of classical type on regular $\Phi$-spaces were completely described \cite{BS1},\cite{BS2},\cite{B8}. In particular, for homogeneous $k$-symmetric spaces, precise computational formulae were indicated. For future reference we select here some results. Denote by $\tilde s$ (respectively, $s$) the number of all irreducible factors (respectively, all irreducible quadratic factors) over $\mathbb{R}$ of a minimal polynomial $\nu$. \begin{theorem}(\cite{BS1},\cite{BS2},\cite{B8}) Let $G/H$ be a regular $\Phi$-space. \begin{enumerate} \item The algebra $\mathcal{A}(\theta )$ contains precisely $2^{\tilde s}$ structures $P$. \item $G/H$ admits a canonical structure $J$ if and only if $s=\tilde{ s}$. In this case $\mathcal{ A}(\theta )$ contains $2^s$ different structures $J$. \item $G/H$ admits a canonical $f$-structure if and only if $s\ne 0$. In this case ${\mathcal A} (\theta )$ contains $3^s-1$ different $f$-structures. Suppose $s={\tilde s}$. Then $2^s$ $f$-structures are almost complex and the remaining $3^s-2^s-1$ have non-trivial kernels. \item The algebra $\mathcal{A}(\theta)$ contains $3^{\tilde{s}}$ different $h$-structures. All these structures form a (commutative) semigroup in $\mathcal{A}(\theta)$ and include a subgroup of order $2^{\tilde{s}}$ of canonical structures $P$. \end{enumerate} \end{theorem} Further, let $G/H$ be a homogeneous $k$-symmetric space. Then $\tilde{s}=s+1$ if $-1\in \, spec\,\theta$, and $\tilde{s}=s$ in the opposite case. We indicate explicit formulae enabling us to compute all canonical $f$-structures and $h$-structures. We shall also use the notation $$u= \left\{\begin{array}{ccc} n & {\text if} & k=2n+1 \\ n-1 & {\text if} & k=2n \end{array}\right..$$ \begin{theorem}(\cite{BS1},\cite{BS2},\cite{B8}) Let $G/H$ be a homogeneous $\Phi$-space of order $k$. \begin{enumerate} \item All non-trivial canonical $f$-structures on $G/H$ can be given by the operators $$f=\frac{2}{k}\sum_{m=1}^u\left(\sum_{j=1}^u\zeta_j\sin{ \frac{2\pi m j}{k}}\right)\left(\theta^m-\theta^{k-m}\right),$$ where $\zeta_j\in\{-1;0 ;1\},\;j=1,2,\ldots,u$, and not all coefficients $\zeta_j$ are zero. In particular, suppose that $-1\notin \, spec\,\theta$. Then the polynomials $f$ define canonical almost complex structures $J$ iff all $\zeta_j\in\{-1;1\}$. \item All canonical $h$-structures on $G/H$ can be given by the polynomials $h=\sum\limits_{m=0}^{k-1}a_m\theta^m$, where: \begin{enumerate} \item if $k=2n+1$, then $$a_m=a_{k-m}=\frac{2}{k}\sum_{j=1}^u\xi_j \cos{\frac{2\pi m j}{k}};$$ \item if $k=2n$, then $$a_m=a_{k-m}= \frac{1}{k}\left(2\sum_{j=1}^u\xi_j \cos{\frac{2\pi m j}{k}} + (-1)^m\xi_n \right)$$ \end{enumerate} Here the numbers $\xi_j$ take their values from the set $\{-1;0 ;1\}$ and the polynomials $h$ define canonical structures $P$ iff all $\xi_j\in\{-1;1\}$. \end{enumerate} \end{theorem} We now particularize the results above mentioned for homogeneous $\Phi$-spaces of orders $3$, $4$, and $5$ only. Note that there are no fundamental obstructions to considering of higher orders $k$. \begin{corol}\label{c1}(\cite{BS2},\cite{B8}) Let $G/H$ be a homogeneous $\Phi$-space of order $3$. There are (up to sign) only the following canonical structures of classical type on $G/H$: $$J=\frac{1}{\sqrt{3}}(\theta-\theta^2),\;P=1.$$ \end{corol} We note that the existence of the structure $J$ and its properties are well known (see \cite{S2},\cite{WG},\cite{G2},\cite{Ki1}). \begin{corol}\label{c2}(\cite{BS2},\cite{B8}) On a homogeneous $\Phi$-space of order $4$ there are (up to sign) the following canonical classical structures: $$P=\theta^2,\;f=\frac12(\theta- \theta^3),\;h_1=\frac12(1-\theta^2),\;h_2=\frac12(1+\theta^2).$$ The operators $h_1$ and $h_2$ form a pair of complementary projectors: $h_1+h_2=1$, $h_1^2=h_1$, $h_2^2=h_2$. Moreover, the following conditions are equivalent: \begin{enumerate} \item $-1\notin spec\,\theta$; \item the structure $P$ is trivial $P=-1$; \item the $f$-structure is an almost complex structure; \item the structure $h_1$ is trivial $(h_1=1)$; \item the structure $h_2$ is null. \end{enumerate} \end{corol} General properties of the canonical structures $P$ and $f$ on homogeneous $4$-symmetric spaces were investigated in \cite{BD}. \begin{corol}\label{c3}(\cite{BS1},\cite{BS2},\cite{B8}) There exist (up to sign) only the following canonical structures of classical type on any homogeneous $\Phi$-space of order $5$: \begin{gather*} P=\frac{1}{\sqrt{5}}(\theta-\theta^2-\theta^3+\theta^4);\\ J_1=\alpha(\theta-\theta^4)-\beta(\theta^2-\theta^3);\quad J_2=\beta(\theta-\theta^4)+\alpha(\theta^2-\theta^3);\\ f_1=\gamma(\theta-\theta^4)+\delta(\theta^2-\theta^3);\quad f_2=\delta(\theta-\theta^4)-\gamma(\theta^2-\theta^3);\\ h_1=\frac12(1+P);\quad h_2=\frac12(1-P); \end{gather*} where $\alpha=\frac{\sqrt{5+2\sqrt5}}{5}$;\ $\beta=\frac{\sqrt{5-2\sqrt5}}{5}$;\ $\gamma=\frac{\sqrt{10+2\sqrt5}}{10}$;\ $\delta=\frac{\sqrt{10-2\sqrt5}}{10}$. Besides, the following relations are satisfied: \begin{gather*} J_1\,P=J_2;\quad f_1\,P=J_1\,h_1=J_2\,h_1=f_1;\quad h_1\,P=h_1;\quad h_2\,P=-h_2;\\ f_2\,P=J_2\,h_2=-J_1\,h_2=-f_2;\quad f_1\,f_2=h_1\,h_2=0;\quad h_1+h_2=P. \end{gather*} In addition, the following conditions are equivalent: \begin{enumerate} \item $spec\ \theta$ consists of two elements; \item the structure $P$ is trivial; \item the structures $J_1$ and $J_2$ coincide (up to sign); \item one of the structures $f_1$ and $f_2$ is null, while the other is an almost complex structure coinciding with one of the structures $J_1$ and $J_2$; \item one of the structures $h_1$ and $h_2$ is trivial, while the other is null. \end{enumerate} \end{corol} We note that for the first time the canonical structure $P$ on homogeneous $5$-symmetric spaces was introduced and studied in \cite{BC}. Other canonical structures on these spaces were later studied in \cite{C1}-\cite{C3}. It should be also mentioned that in the particular case of homogeneous $\Phi$-spaces of any odd order $k=2n+1$ the method of constructing invariant almost complex structures was described in \cite{Ko}. It can be easily seen that all these structures are canonical in the above sense. \section{\bf Almost Hermitian structures} We briefly recall some notions of Hermitian geometry including the main classes of almost Hermitian structures. Let $M$ be a smooth manifold, $\frak{X} (M)$ the Lie algebra of all smooth vector fields on $M$, $d$ the exterior differentiation operator. An {\it almost Hermitian structure} on $M$ (briefly, {\it $AH$-structure}) is a pair $(g,J)$, where $g=\langle\cdot,\cdot\rangle$ is a (pseudo)Riemannian metric on $M$, $J$ an almost complex structure such that $\langle JX, JY \rangle = \langle X, Y \rangle$ for any $X,Y\in\frak{X}(M)$. It follows immediately that the tensor field $\Omega (X,Y)=\langle X, JY \rangle$ is skew-symmetric, i.e. $(M,\Omega)$ is an almost symplectic manifold. $\Omega$ is usually called a {\it fundamental form} (the {\it K\"{a}hler form}) of an $AH$-structure $(g,J)$. Further, denote by $\nabla$ the Levi-Civita connection of the metric $g$ on $M$. We recall below some main classes of $AH$-structures together with their defining properties (see, for example, \cite {GH}): \begin{tabbing} Kill f123 \= $f$-structure of class $G_1$, or -structure \= $T(X,X)=0,$ i.e. $\frak{X}(M)$ is an anticommutative $Q$-algebra \kill {\bf K} \> {\it K\"{a}hler structure}: \> $\nabla J=0$; \\ {\bf H} \> {\it Hermitian structure}: \> $\nabla_X(J)Y-\nabla_{JX}(J)JY=0$; \\ {\bf G$_1$} \> {$AH$-structure of class $G_1$,} or \> $\nabla_X(J)X-\nabla_{JX}(J)JX=0$; \\ \> $G_1$-{\it structure}: \> \\ {\bf QK} \> {\it quasi-K\"{a}hler structure}: \> $\nabla_X(J)Y+\nabla_{JX}(J)JY=0$;\\ {\bf AK} \> {\it almost K\"{a}hler structure}: \> $d\,\Omega=0$;\\ {\bf NK} \> {\it nearly K\"{a}hler structure,} \> $\nabla_{X}(J)X=0$.\\ \> or $NK$-{\it structure}: \> \end{tabbing} It is well known (see, for example, \cite {GH}) that $$ {\bf K}\subset{\bf H}\subset{\bf G_1};\;\;{\bf K}\subset{\bf NK}\subset{\bf G_1};\;\;{\bf NK}={\bf G_1}\cap{\bf QK};\;\;{\bf K}={\bf H}\cap{\bf QK}. $$ As usual, we will denote by $N$ the Nijenhuis tensor of an almost complex structure $J$, that is, $$ N(X,Y)=\frac14([JX,JY]-J[JX,Y]-J[X,JY]-[X,Y]) $$ for any $X,Y\in\frak{X}(M)$. Then the condition $N=0$ is equivalent to the integrability of $J$ (see, for example, \cite {KN}). Moreover, an almost Hermitian structure $(g,J)$ belongs to the class {\bf H} if and only if $N=0$ (see, for example, \cite {GH}). As was already mentioned, the role of homogeneous almost Hermitian manifolds is particularly important "because they are the model spaces to which all other almost Hermitian manifolds can be compared" (see \cite {G3}). A wealth of examples for the most classes above noted, both of general and specific character, can be found in \cite {WG}, \cite {G2}, \cite {G3}, \cite {Ki1} and others. In particular, after the detailed investigation of the 6-dimensional homogeneous nearly K\"{a}hler manifolds V.F.Kirichenko proved \cite {Ki1} that naturally reductive strictly nearly K\"{a}hler manifolds $SO(5)/U(2)$ and $SU(3)/T_{max}$ are not isometric even locally to the 6-dimensional sphere $S^6$. These examples gave a negative answer to the conjecture of S.Sawaki and Y.Yamanoue (see \cite {SaYa}) claimed that any $6$-dimensional strictly $NK$-manifold was a space of constant curvature. It should be noted that the canonical almost complex structure $J=\frac1{\sqrt{3}}(\theta-\theta^2)$ on homogeneous $3$-symmetric spaces plays a key role in these and other examples of homogeneous $AH$-manifolds. Let $G$ be a connected Lie group, $H$ its closed subgroup, $g$ an invariant (pseudo-)Riemannian metric on the homogeneous space $G/H$. Denote by $\mathfrak{g}$ and $\mathfrak{h}$ the Lie algebras corresponding to $G$ and $H$ respectively. Suppose that $G/H$ is a reductive homogeneous space, $\mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m}$ the reductive decomposition of the Lie algebra $\mathfrak{g}$. As usual, we identify $\mathfrak{m}$ with the tangent space $T_o(G/H)$ at the point $o=H$. Then the invariant metric $g$ is completely defined by its value at the point $o$. For convenience we denote by the same manner both any invariant metric $g$ on $G/H$ and its value at $o$. Recall that $(G/H,g)$ is {\it naturally reductive} with respect to a reductive decomposition $\mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m}$ \cite{KN} if $$g([X,Y]_{\mathfrak{m}},Z)=g(X,[Y,Z]_{\mathfrak{m}})$$ for all $X,Y,Z\in\mathfrak{m}$. Here the subscript $\mathfrak{m}$ denotes the projection of $\mathfrak{g}$ onto $\mathfrak{m}$ with respect to the reductive decomposition. We select here some known results closely related to the main subject of our future consideration. \begin{theorem} \label{t1}(\cite{AG}) Any invariant almost Hermitian structure on a naturally reductive space $(G/H,g)$ belongs to the class {\bf G$_1$}. \end{theorem} \begin{theorem} \label{t2}(\cite{WG}, \cite{G2}) A homogeneous $3$-symmetric space $G/H$ with the canonical almost complex structure $J$ and an invariant compatible metric $g$ is a quasi-K\"{a}hler manifold. Moreover, $(G/H,J,g)$ belongs to the class {\bf NK} if and only if $g$ is naturally reductive. \end{theorem} \begin{theorem} \label{t3}(\cite{Ma}, \cite{G4}, \cite{Ki6}) A $6$-dimensional strictly nearly K\"{a}hler manifold is Einstein. \end{theorem} Note that the latter result was obtained in \cite{Ki6} as a particular case of the general approach based on investigating nearly K\"{a}hler manifolds of constant type. \section{\bf Metric $f$-structures and homogeneous manifolds} The concept of the {\it generalized Hermitian geometry} created in the 1980s (see, for example, \cite{Ki2}, \cite{Ki7}, \cite{Ki8}) was a natural consequence of the development of Hermitian geometry and the theory of almost contact metric structures with many applications. One of its central objects is the metric $f$-structures of classical type, which include the class of almost Hermitian structures. We recall here some basic notions. An {\it $f$-structure} on a manifold $M$ is known to be a field of endomorphisms $f$ acting on its tangent bundle and satisfying the condition $f^3+f=0$ (see \cite{Y}). The number $r=dim\;Im\:f$ is constant at any point of $M$ (see \cite{St}) and called a {\it rank} of the $f$-structure. Besides, the number $dim\;Ker\:f=dim\:M-r$ is usually said to be a {\it deficiency} of the $f$-structure and denoted by $def\:f$. Recall that an $f$-structure on a (pseudo)Riemannian manifold $(M,g=\langle\cdot,\cdot\rangle)$ is called a {\it metric $f$-structure}, if $\langle {fX}, Y \rangle +\langle X, fY \rangle =0$, \;$X,Y\in\frak{X}(M)$ (see \cite{Ki2}). In the case the triple $(M,g,f)$ is called a {\it metric $f$-manifold}. It is clear that the tensor field $\Omega(X,Y)=\langle X, fY \rangle$ is skew-symmetric, i.e. $\Omega$ is a $2$-form on $M$. $\Omega$ is called a {\it fundamental form} of a metric $f$-structure \cite{Ki7}, \cite{Ki2}. It is easy to see that the particular cases $def\;f=0$ and $def\; f=1$ of metric $f$-structures lead to almost Hermitian structures and almost contact metric structures respectively. Let $M$ be a metric $f$-manifold. Then $\frak{X}(M)=\mathcal{L}\oplus\mathcal{M}$, where $\mathcal{L}=Im\;f$ and $\mathcal{M}=Ker\;f$ are mutually orthogonal distributions, which are usually called the {\it first} and the {\it second fundamental distributions} of the $f$-structure respectively. Obviously, the endomorphisms $l=-f^2$ and $m=id+f^2$ are mutually complementary projections on the distributions $\mathcal{L}$ and $\mathcal{M}$ respectively. We note that in the case when the restriction of $g$ to $\mathcal{L}$ is non-degenerate the restriction $(F,g)$ of a metric $f$-structure to $\mathcal{L}$ is an almost Hermitian structure, i.e. $F^2=-id,\;\langle FX, FY \rangle=\langle X, Y \rangle,\;X,Y\in\mathcal{L}$. A fundamental role in the geometry of metric $f$-manifolds is played by the {\it composition tensor} $T$, which was explicitly evaluated in \cite{Ki7}: \begin{equation}\label{eqT} T(X,Y)=\frac14{f}(\nabla_{fX}(f){fY}-\nabla_{f^2X}(f){f^2Y}), \end{equation} where $\nabla$ is the Levi-Civita connection of a (pseudo)Riemannian manifold $(M,g)$, \; $X,Y\in\frak{X}(M)$. Using this tensor $T$, the algebraic structure of a so-called {\it adjoint $Q$-algebra} in $\frak{X}(M)$ can be defined by the formula: $$ X\ast{Y}=T(X,Y). $$ It gives the opportunity to introduce some classes of metric $f$-structures in terms of natural properties of the adjoint $Q$-algebra (see \cite{Ki2}, \cite{Ki7}). We enumerate below the main classes of metric $f$-structures together with their defining properties: \begin{tabbing} Kill f123 \= $f$-structure of class $G_1$, or -structure \= $T(X,X)=0,$ i.e. $\frak{X}(M)$ is an anticommutative $Q$-algebra \kill {\bf Kf} \> {\it K\"{a}hler $f$--structure}: \> $\nabla f=0$; \\ {\bf Hf} \> {\it Hermitian $f$--structure}: \> $T(X,Y)=0,$ i.e. $\frak{X}(M)$ is\\ \> \> an abelian $Q$-algebra;\\ {\bf G$_1$f} \> {$f$-structure of class $G_1$,} or \> $T(X,X)=0,$ i.e. $\frak{X}(M)$ is \\ \> $G_1f$-{\it structure}: \> an anticommutative $Q$-algebra;\\ {\bf QKf} \> {\it quasi-K\"{a}hler $f$--structure}: \> $\nabla_X f +T_X f=0$;\\ {\bf Kill f} \> {\it Killing $f$-structure}: \> $\nabla_X (f) X=0$;\\ {\bf NKf} \> {\it nearly K\"{a}hler $f$-structure,} \> $\nabla_{fX}(f)fX=0$.\\ \> or $NKf$-{\it structure}: \> \end{tabbing} The classes {\bf Kf}, {\bf Hf}, {\bf G$_1$f}, {\bf QKf} (in more general situation) were introduced in \cite{Ki2} (see also \cite{Si}). Killing $f$-manifolds {\bf Kill f} were defined and studied in \cite{Gr1}, \cite{Gr2}. The class {\bf NKf} was first determined in \cite{B1} (see also \cite{B6}, \cite{B7}). The following relationships between the classes mentioned are evident: $$ {\bf Kf}={\bf Hf}\cap{\bf QKf};\;\;\; {\bf Kf}\subset{\bf Hf}\subset{\bf G_1f};\;\;\; {\bf Kf}\subset{\bf Kill\;f}\subset{\bf NKf}\subset{\bf G_1f.} $$ It is important to note that in the special case $f=J$ we obtain the corresponding classes of almost Hermitian structures (see \cite{GH}). In particular, for $f=J$ the classes {\bf Kill f} and {\bf NKf} coincide with the well-known class {\bf NK} of {\it nearly K\"ahler structures}. {\bf Remark 1.} Killing $f$-manifolds are often defined by requiring the fundamental form $\Omega$ to be a Killing form, i.e. $d\Omega=\nabla\Omega$ (see \cite{Gr1}, \cite{KL}). It is not hard to prove that the definition is equivalent to the above condition $\nabla_{X}(f)X=0$. Indeed, in accordance with \cite{YB}, $\Omega$ is a Killing $2$-form if and only if $\nabla\Omega$ is a $3$-form. Further, using \cite{KN}, we have \begin{equation}\label{eq1} (\nabla\Omega)(X,Y;Z)=Z\:\Omega(X,Y)-\Omega(\nabla_{Z}X,Y)-\Omega(X,\nabla_{Z}Y). \end{equation} It follows that $\nabla\Omega$ is always skew-symmetric in arguments $X$ and $Y$. Using the fact, it is easy to see that $\nabla\Omega$ is a $3$-form if and only if $\nabla\Omega$ is skew-symmetric in $Y$ and $Z$: \begin{equation}\label{eq2} (\nabla\Omega)(X,Y;Z)=-(\nabla\Omega)(X,Z;Y). \end{equation} Taking into account formula (\ref{eq1}) and the definition of $\Omega$, condition (\ref{eq2}) can be written in the form: $$ Z\langle X, fY \rangle-\langle \nabla_{Z}X, fY \rangle-\langle X, f\nabla_{Z}Y \rangle+Y\langle X, fZ \rangle-\langle \nabla_{Y}X, fYZ \rangle-\langle X, f\nabla_{Y}Z \rangle=0. $$ Since $\nabla$ is the Levi-Civita connection, in particular, we have: $$ Z\langle X, Y \rangle=\langle \nabla_{Z}X, Y \rangle+\langle X, \nabla_{Z}Y \rangle. $$ It follows that the previous formula can be written in the form: $$ \langle X, \nabla_{Z}fY -f\nabla_{Z}Y \rangle+\langle X, \nabla_{Y}fZ -f\nabla_{Y}Z \rangle=0. $$ Using the formula $\nabla_{X}(f)Y=\nabla_{X}fY-f\nabla_{X}Y$, we get: $$ \langle X, \nabla_{Z}(f)Y + \nabla_{Y}(f)Z \rangle=0. $$ It implies that $\nabla_{Z}(f)Y + \nabla_{Y}(f)Z=0$ for any $Y,Z\in\frak{X}(M)$. This is obviously equivalent to the condition $\nabla_{X}(f)X=0,\;X\in\frak{X}(M)$. {\hfill $\square$} Now we dwell on invariant metric $f$-structures on homogeneous spaces. Any invariant metric $f$-structure on a reductive homogeneous space $G/H$ determines the orthogonal decomposition $\mathfrak{m}=\mathfrak{m}_1\oplus\mathfrak{m}_2$ such that $\mathfrak{m}_1=Im\:f$, $\mathfrak{m}_2=Ker\:f$. As it was already noted (see Section 3), the main classes of almost Hermitian structures are provided with the remarkable set of invariant examples. It turns out that there is also a wealth of invariant examples for the basic classes of metric $f$-structures. These invariant metric $f$-structures can be realized on homogeneous $k$-symmetric spaces with canonical $f$-structures. We select here some results in this direction. More detailed information can be found in \cite{B1}-\cite{B7}, \cite{C3}, \cite{Li}. \begin{theorem} \label{t4}(\cite{B5}) Any invariant metric $f$-structure on a naturally reductive space $(G/H,g)$ is a $G_{1}f$-structure. \end{theorem} As a special case $(Ker\:f=0)$, it follows Theorem \ref{t1}. \begin{theorem} \label{t5}(\cite{B5}) Let $(G/H,g,f)$ be a naturally reductive space with an invariant metric $f$-structure that satisfies the condition $[\frak{m}_1,\frak{m}_1]\subset\frak{m}_2\oplus\frak{h}$. Then $G/H$ is a Hermitian $f$-manifold. \end{theorem} We note that Theorem \ref{t5} is also valid for arbitrary invariant (pseudo)Riemanni\-an metric $g$ compatible with an invariant $f$-structure on a reductive homogeneous space $G/H$ (see \cite{BV}). Theorems \ref{t4} and \ref{t5} can be effectively provided with a large class of examples. In particular, for a semi-simple group $G$, the invariant (pseudo)Riemannian metric $g$ generated by the Killing form on any regular $\Phi$-space $G/H$ is naturally reductive with respect to the canonical reductive decomposition $\frak{g}=\frak{h}\oplus\frak{m}$ (see \cite{S1}). Moreover, all canonical structures $f$ and $J$ on homogeneous naturally reductive $k$-symmetric spaces are compatible with such a metric, i.e. $f$ is a metric $f$-structure, $J$ is an almost Hermitian structure (see \cite{B1}, \cite{B10}). In what follows, we mean by a naturally reductive decomposition the canonical reductive decomposition for a regular $\Phi$-space $G/H$. To sum up, we have \begin{theorem} \label{t6}(\cite{B5}) Let $(G/H,g)$ be a naturally reductive $k$-symmetric space. Any canonical metric $f$-structure on $G/H$ is a $G_{1}f$-structure, and any canonical almost Hermitian structure $J$ is a $G_{1}$-structure. \end{theorem} \begin{theorem} \label{t7}(\cite{B6},\cite{B7}) Let $G/H$ be a regular $\Phi$-space, $g$ a naturally reductive metric on $G/H$ with respect to the canonical reductive decomposition $\mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m}$, $f$ a metric canonical $f$-structure on $G/H$. Suppose the $f$-structure satisfies the condition $f^2=\pm\theta\:f$. Then $(G/H,g,f)$ is a nearly K\"ahler $f$-manifold. \end{theorem} \begin{corol} (\cite{B6},\cite{B7}) Let $(G/H,g)$ be a naturally reductive homogeneous $\Phi$-space of order $k=4n,\:n\ge{1}$. If $\{i,-i\}\ \subset\ spec\ \theta$, then there exists a non-trivial canonical $NKf$-structure on $G/H$. \end{corol} We stress the particular role of homogeneous $4$- and $5$-symmetric spaces. \begin{theorem} \label{t8}(\cite{B3}-\cite{B7}) The canonical $f$-structure $f=\frac12(\theta-\theta^3)$ on any naturally reductive $4$-symmetric space $(G/H,g)$ is both a Hermitian $f$-structure and a nearly K\"ahler $f$-structure. Moreover, the following conditions are equivalent: 1) $f$ is a K\"ahler $f$-structure; 2) $f$ is a Killing $f$-structure; 3) $f$ is a quasi-K\"ahler $f$-structure; 4) $f$ is an integrable $f$-structure; 5) $[\frak{m}_1,\frak{m}_1]\subset\frak{h}$; 6) $[\frak{m}_1,\frak{m}_2]=0$; 7) $G/H$ is a locally symmetric space: $[\frak{m},\frak{m}]\subset\frak{h}$. \end{theorem} \begin{theorem} \label{t9}(\cite{B2}-\cite{B5}, \cite{B7}, \cite{C3}) Let $(G/H,g)$ be a naturally reductive $5$-symmetric space, $f_1$ and $f_2$, $J_1$ and $J_2$ the canonical structures on this space. Then $f_1$ and $f_2$ belong to both classes {\bf Hf} and {\bf NKf}. Moreover, the following conditions are equivalent: 1) $f_1$ is a K\"ahler $f$-structure; 2) $f_2$ is a K\"ahler $f$-structure; 3) $f_1$ is a Killing $f$-structure; 4) $f_2$ is a Killing $f$-structure; 5) $f_1$ is a quasi-K\"ahler $f$-structure; 6) $f_2$ is a quasi-K\"ahler $f$-structure; 7) $f_1$ is an integrable $f$-structure; 8) $f_2$ is an integrable $f$-structure; 9) $J_1$ and $J_2$ are $NK$-structures; 10) $[\frak{m}_1,\frak{m}_2]=0$ (here $\frak{m}_1=Im\:f_1=Ker\:f_2, \frak{m}_2=Im\:f_2=Ker\:f_1$); 11) $G/H$ is a locally symmetric space: $[\frak{m},\frak{m}]\subset\frak{h}$. \end{theorem} It should be mentioned that Riemannian homogeneous $4$-symmetric spaces of classical compact Lie groups were classified and geometrically described in \cite{J}. The similar problem for homogeneous $5$-symmetric spaces was considered in \cite{TX}. By Theorem \ref{t8} and Theorem \ref{t9}, it presents a collection of homogeneous $f$-manifolds in the classes {\bf NKf} and {\bf Hf}. Note that the canonical $f$-structures under consideration are generally non-integrable. Besides, there are invariant $NKf$-structures and $Hf$-structures on homogeneous spaces $(G/H,g)$, where the metric $g$ is not naturally reductive. The example of such a kind can be realized on the $6$-dimensional Heisenberg group $(N,g)$. As to details related to this group, we refer to \cite{Ka1}, \cite{Ka2}, \cite{TV}. \begin{theorem} \label{t10}(\cite{B5}-\cite{B7}) The 6-dimensional generalized Heisenberg group $(N,g)$ with respect to the canonical $f$-structure $f=\frac12(\theta-\theta^3)$ of a homogeneous $\Phi$-space of order $4$ is both $Hf$- and $NKf$-manifold. This $f$-structure is neither Killing nor integrable on $(N,g)$. \end{theorem} {\bf Remark 2.} Theorems \ref{t8} and \ref{t10}, in particular, illustrate simultaneously the analogy and the difference between the canonical almost complex structure $J$ on homogeneous $3$-symmetric spaces $(G/H,g,J)$ and the canonical $f$-structure on homogeneous $4$-symmetric spaces $(G/H,g,f)$ (see Theorem \ref{t2}). Let us also remark that the 6-dimensional generalized Heisenberg group $(N,g)$ is an example of solvable type. In Section 5, we present $NKf$-structures with non-naturally reductive metrics in the case of semi-simple type. Finally, we briefly discuss the existence problem for invariant Killing $f$-structu\-res. By Theorems \ref{t8} and \ref{t9}, the canonical $f$-structures on naturally reductive $4$- and $5$-symmetric spaces are never strictly (i.e. non-K\"ahler) Killing $f$-structures. Moreover, we recall the following general result: \begin{theorem} \label{t11}(\cite{B4}) Let $(G/H,g,f)$ be a naturally reductive Killing $f$-manifold. Then the following relations hold: $$ [\mathfrak{m}_1,\mathfrak{m}_1]\subset\mathfrak{m}_1\oplus\mathfrak{h}, \quad [\mathfrak{m}_2,\mathfrak{m}_2]\subset\mathfrak{m}_2\oplus\mathfrak{h}, \quad [\mathfrak{m}_1,\mathfrak{m}_2]\subset\mathfrak{h}. $$ In particular, both the fundamental distributions of the Killing $f$-structure generate invariant totally geodesic foliations on $G/H$. \end{theorem} By the results in \cite{Gr1} and Theorem \ref{t11}, it follows \begin{corol}(\cite{B4}) There are no non-trivial (i.e. $def\:f>0$) invariant Killing $f$-structures of the so-called fundamental type (see \cite{Gr1}) on naturally reductive homogeneous spaces $(G/H,g)$. \end{corol} This fact is a wide generalization of the similar result of A.Gritsans obtained for Riemannian globally symmetric spaces. Besides, it shows a substantial difference between invariant Killing $f$-structures and invariant $NK$-structures. In Section 5, we will indicate, in particular, first examples of invariant Killing $f$-structures. \section{\bf Invariant $f$-structures on the complex \\flag manifold $\bf M=SU(3)/T_{max}$} In this Section, we will consider all invariant $f$-structures on the flag manifold $M=SU(3)/T_{max}$. Note that invariant almost complex structures (i.e. $f$-structures of maximal rank $6$) on this space were investigated in \cite{G3}, \cite{AGI1}, \cite{AGI2} and many other papers. The homogeneous manifold $SU(3)/T_{max}$ is known to be an important example in many branches of differential geometry and beyond. In particular, $M=SU(3)/T_{max}$ is a Riemannian homogeneous $3$-symmetric space not homeomorphic with the underlying manifold $M$ of any Riemannian symmetric space (see \cite{LP1}). Further, $M$ is a homogeneous $k$-symmetric space for any $k\ge3$. Moreover, $M$ is a naturally reductive Riemannian homogeneous space that is {\it non-commutative} (see \cite{J2}). It means that the algebra of invariant differential operators $\mathcal{D}(SU(3)/T_{max})$ is not commutative (see \cite{H1}). It follows that $M=SU(3)/T_{max}$ is not even a {\it weakly symmetric space} (see, for example, \cite{Vi}). Besides, $M$ is the twistor space for the projective space $\mathbb{C}P^2$ (see, for example, \cite{Be}, Chapter 13). It was a key point for constructing the first examples of $6$-dimensional Riemannian manifolds admitting a real Killing spinor (see \cite{BFGK}). More exactly, the flag manifold $M=SU(3)/T_{max}$ with the nearly K\"ahler structure $(g,J)$ just possesses a real Killing spinor (see \cite{BFGK}, \cite{Gru}). Moreover, using the duality procedure for this space $SU(3)/T_{max}$, one can effectively construct pseudo-Riemannian homogeneous manifolds with the real Killing spinors (see \cite{Ka}). Let $\Phi=I(s)$ be an inner automorphism of the Lie group $SU(3)$ defined by the element $s=diag\:(\varepsilon,\overline \varepsilon,1)$, where $\varepsilon$ is a primitive third root of unity. Then the subgroup $H=G^{\Phi}$ of all fixed points of $\Phi$ is of the form: $$ G^{\Phi}=\{ diag(e^{i\beta_1}, e^{i\beta_2},e^{i\beta_3})| \beta_1+\beta_2+\beta_3=0,\;\beta_j\in\mathbb{R}\}. $$ Obviously, $G^{\Phi}$ is isomorphic to $T^2=T_{max}$ diagonally imbedded into $SU(3)$. It means that the flag manifold $M=SU(3)/T_{max}$ is a homogeneous $3$-symmetric space defined by the automorphism $\Phi$. Consider the canonical reductive decomposition $\frak{g}=\frak{h}\oplus\frak{m}$ of the Lie algebra $\frak{g}=\frak{s}\frak{u}(3)$ for the homogeneous $\Phi$-space $M$. Using the notations in \cite{R1}, we obtain: $$\frak{g}=\frak{su}(3)=\left\{\left(\begin{array}{ccc} \alpha_1 & a & \overline{c} \\ -\overline{a} & \alpha_2 & b \\ -c & -\overline{b} & \alpha_3 \end{array}\right) \left| \begin{array}{l} \alpha_1,\alpha_2, \alpha_3\in Im\,\mathbb{C}, \\ a,b,c\in\mathbb{C}, \\ \alpha_1+\alpha_2+\alpha_3=0 \end{array}\right.\right\}=$$ $$=E(\alpha_1,\alpha_2, \alpha_3)\oplus D(a,b,c)=\frak{h}\oplus\frak{m}.$$ If we put $X=D(a,b,c), Y=D(a_1,b_1,c_1), Z=E(\alpha_1,\alpha_2,\alpha_3)$, then the Lie brackets can be briefly indicated (see \cite{R2}): $$ [X,Y]=D(\overline{bc_1-b_1c}, \overline{ca_1-c_1a}, \overline{ab_1-a_1b}) -$$ $$ 2E(Im(a\overline{a_1}+\overline{c}c_1),Im(\overline{a}a_1+b\overline{b_1}), Im(c\overline{c_1}+\overline{b}b_1) ) ,$$ $$[Z,X]=D(\alpha_1 a-a\alpha_2, \alpha_2 b-b\alpha_3, \alpha_3c-c\alpha_1).$$ Further, we put $\frak{m}=\frak{m}_1\oplus\frak{m}_2\oplus\frak{m}_3,$ where $$\frak{m}_1 = \{X\in\frak{su}(3)| X=D(a,0,0), a\in\mathbb{C}\},$$ $$\frak{m}_2 =\{X\in\frak{su}(3)| X=D(0,b,0), b\in\mathbb{C}\},$$ $$\frak{m}_3 = \{X\in\frak{su}(3)| X=D(0,0,c), c\in\mathbb{C}\}.$$ Using the Killing form of the Lie algebra $\frak{su}(3)$, we define an invariant inner product on $\frak{m}$: $$ g_o(X,Y)=\langle X, Y\rangle_o=-\frac12Re\:tr\:XY. $$ Then (see \cite{R1}) $\frak{g}=\frak{h}\oplus\frak{m}_1\oplus\frak{m}_2\oplus\frak{m}_3$ is $\langle\cdot,\cdot\rangle_o$-orthogonal decomposition satisfying the following relations: $$ [\frak{h},\frak{m}_j]\subset\frak{m}_j,\;[\frak{m}_j,\frak{m}_j]\subset\frak{h},\; [\frak{m}_j,\frak{m}_{j+1}]\subset\frak{m}_{j+2}, $$ where $j=1,2,3$ and the index $j$ should be reduced by modulo $3$. Besides, the $H$-modules $\frak{m}_j$ are pairwise non-isomorphic. Now we turn to invariant Riemannian metrics on $M$. Taking into account the well-known one-to-one correspondence between $G$-invariant Riemannian metrics on $G/H$ and $Ad(H)$-invariant inner products on $\frak{m}$ (see \cite{KN}), we will make use of the following fact: \begin{lemma}(\cite{R1}) Any $SU(3)$-invariant Riemannian metric $g=\langle\cdot,\cdot\rangle$ on the flag manifold $M=SU(3)/T_{max}$ can be written in the form $$ g=\langle\cdot,\cdot\rangle= \lambda_1\langle\cdot,\cdot\rangle_{o{|\frak{m}_1\times\frak{m}_1}} + \lambda_2\langle\cdot,\cdot\rangle_{o{|\frak{m}_2\times\frak{m}_2}} + \lambda_3\langle\cdot,\cdot\rangle_{o{|\frak{m}_3\times\frak{m}_3}}, $$ where $\lambda_j>0,\;j=1,2,3.$ \end{lemma} A triple $(\lambda_1,\lambda_2,\lambda_3)$ is called \cite{R1} a {\it characteristic collection} of a Riemannian metric $g$ above mentioned . Considering Riemannian metrics up to homothety, one can assume that $(\lambda_1,\lambda_2,\lambda_3)=(1,t,s),\;t>0,s>0.$ For convenience we will denote this correspondence in the following way: $g=(\lambda_1,\lambda_2,\lambda_3)$ or $g=(1,t,s).$ We also recall the following result: \begin{theorem} \label{t12}(\cite{ZW},\cite{AN},\cite{R1}) There are exactly (up to homothety) the following invariant Einstein metrics on the flag manifold $SU(3)/T_{max}$ $:$ $$(1,1,1), (1,2,1), (1,1,2), (2,1,1).$$ \end{theorem} Let $\alpha$ be the Nomizu function (see \cite{N}) of the Levi-Civita connection $\nabla$ for an invariant Riemannian metric $g=\langle\cdot,\cdot\rangle$ on a reductive homogeneous space $G/H$. Then \begin{equation}\label{eq4} \alpha(X,Y)=\frac12[X,Y]_{\frak{m}}+U(X,Y),\;\;\;X,Y\in\frak{m}, \end{equation} where $U:\frak{m}\times\frak{m}\to\frak{m}$ is a symmetric bilinear mapping determined by the formula (see\cite{KN}): $$ 2\langle U(X,Y),Z\rangle=\langle X,[Z,Y]_{\frak{m}}\rangle+ \langle [Z,X]_{\frak{m}},Y\rangle. $$ For our case in these notations we have \begin{lemma}(\cite{W},\cite{R2}) For the Levi-Civita connection of a Riemannian metric $g=(\lambda_1,\lambda_2,\lambda_3)$ on the flag manifold $SU(3)/T_{max}$ the following conditions are satisfied$:$ $$ U(X,Y)=0, \;\;if\;\; X,Y\in\frak{m}_j,\,j\in\{1,2,3\}; $$ $$ U(X,Y)= - (2\lambda_j)^{-1}(\lambda_{j+1}-\lambda_{j+2})[X,Y], \;\;if\;\; X\in\frak{m}_{j+1},Y\in\frak{m}_{j+2}, $$ where $j=1,2,3$ and the numbers $j$ are reduced by modulo $3$. \end{lemma} Let us now turn to invariant $f$-structures on $M=SU(3)/T_{max}$. Keeping the above notations, any invariant $f$-structure on $M$ can be expressed by the mapping \begin{equation}\label{eqf} f:D(a,b,c)\rightarrow D(\zeta_1ia,\zeta_2ib,\zeta_3ic), \end{equation} where $\zeta_{j}\in\{1,0,-1\},\;j=1,2,3$, $i$ is the imaginary unit. We will call the collection $(\zeta_1,\zeta_2,\zeta_3)$ a {\it characteristic collection} of the invariant $f$-structure and for convenience denote $f=(\zeta_1,\zeta_2,\zeta_3).$ Obviously, all invariant $f$-structures on $M$ pairwise commute. If we agree to consider $f$-structures up to sign, then there are the following invariant $f$-structures on $M=SU(3)/T_{max}$: 1) {\it invariant $f$-structures of rank $6$ (invariant almost complex structures)}: $$ J_1=(1,1,1),\;\;J_2=(1,-1,1),\;\;J_3=(1,1,-1),\;\;J_4=(1,-1,-1). $$ 2) {\it invariant $f$-structures of rank $4$}: $$ \begin{array}{ccc} f_1=(1,1,0),\;\;\;\;f_2=(1,0,1),\;\;\;\;f_3=(0,1,1),\\ f_4=(1,-1,0),\;\;\;\;f_5=(1,0,-1),\;\;\;\;f_6=(0,1,-1). \end{array} $$ 3) {\it invariant $f$-structures of rank $2$}: $$ f_7=(1,0,0),\;\;\;\;f_8=(0,1,0),\;\;\;\;f_9=(0,0,1). $$ Our description of all invariant $f$-structures and all invariant Riemannian metrics evidently implies that any invariant $f$-structure $f=(\zeta_1,\zeta_2,\zeta_3)$ is a metric $f$-structure with respect to any invariant Riemannian metric $g=(\lambda_1,\lambda_2,\lambda_3)$. In particular, $J_j,\;j=1,2,3,4$ are invariant almost Hermitian structures with respect to all invariant Riemannian metrics $g=(\lambda_1,\lambda_2,\lambda_3)$. Now we are able to investigate all invariant $f$-structures in the sense of generalized Hermitian geometry, i.e. the special classes {\bf Kf}, {\bf NKf}, {\bf Kill f}, {\bf Hf}, {\bf G$_1$f}. A key point of our consideration belongs to the expression $\nabla_X(f)Y$. Using formula (\ref{eq4}), we get: $$ \begin{array}{cc} \nabla_X(f)Y=\nabla_X fY-f\nabla_X Y=\alpha(X,fY)-f \alpha(X,Y)\\ =\frac12 ([X,fY]_{\frak{m}}-f[X,Y]_{\frak{m}})+U(X,fY)-fU(X,Y). \end{array} $$ As a result, we can obtain: \begin{multline}\label{123} \nabla_X(f)Y=\frac12 D(A,B,C), \text{where}\\ A=\overline{i((\zeta_1+\zeta_3)(1+s-t)bc_1+(\zeta_1+\zeta_2)(s-t-1)b_1c)},\\ B=\overline{i((\zeta_2+\zeta_1)(1+\frac{1-s}{t})ca_1+(\zeta_2+\zeta_3)(\frac{1-s}{t}-1)c_1a)},\\ C=\overline{i((\zeta_3+\zeta_2)(\frac{t-1}{s}+1)ab_1+(\zeta_3+\zeta_1)(\frac{t-1}{s}-1)a_1b)}. \end{multline} \subsection{\bf K\"ahler $f$-structures} K\"ahler $f$-structures are defined by the condition $\nabla_X(f)Y=0$ (see Section 4). Using formula (\ref{123}), this condition is equivalent to the following system of equations: \begin{equation}\label{syst1} \left\{\begin{array}{c} (\zeta_1+\zeta_3)(s-t+1)=0 \\ (\zeta_1+\zeta_2)(s-t-1)=0 \\ (\zeta_2+\zeta_3)(s+t-1)=0 \ \end{array}\right. \end{equation} Solving system (\ref{syst1}) for all invariant $f$-structures, we obtain the following result: \begin{propos} The flag manifold $M=SU(3)/T_{max}$ admits the following invariant K\"ahler $f$-structures with respect to the corresponding invariant Riemannian metrics only: $$ \begin{array}{lll} J_2=(1,-1,1),&g_t=(1,t,t-1),&t>1;\\ J_3=(1,1,-1),&g_t=(1,t,t+1),&t>0;\\ J_4=(1,-1,-1),&g_t=(1,t,1-t),&0<t<1. \end{array} $$ In particular, there are no invariant K\"ahler $f$-structures of rank $2$ and $4$ on $M$. \end{propos} We note that the result is known for invariant almost complex structures (see \cite{G3},\cite{AGI2}). We can also observe that for each of K\"ahler $f$-structures $J_2,J_3,J_4$ the corresponding $1$-parameter set $g_t$ of invariant Riemannian metrics contains exactly one Einstein metric excluding the naturally reductive metric $g=(1,1,1)$ (see Theorem \ref{t12}). Taking into account Theorem \ref{t2}, the latter fact implies that the structures $J_2,J_3,J_4$ cannot be realized as the canonical almost complex structures $J=\frac1{\sqrt3}(\theta-\theta^2)$ for some homogeneous $\Phi$-spaces of order $3$. In addition, Lie brackets relations for the subspaces $\frak{m}_j, \;j=1,2,3$ imply that all invariant $f$-structures of rank $2$ and $4$ are non-integrable. It immediately follows that these $f$-structures cannot be K\"ahler $f$-structures. \subsection{\bf Killing $f$-structures} The defining condition for Killing $f$-structures can be written in the form $\nabla_X(f)X=0$ (see Section 4). From (\ref{123}), it follows \begin{multline*} \nabla_X(f)X=\frac12 D(A_0,B_0,C_0), \text{where}\\ A_0=\overline{ibc((\zeta_1+\zeta_3)(1+s-t)+(\zeta_1+\zeta_2)(s-t-1))},\\ B_0=\overline{ica((\zeta_2+\zeta_1)(1+\frac{1-s}{t})+(\zeta_2+\zeta_3)(\frac{1-s}{t}-1))},\\ C_0=\overline{iab((\zeta_3+\zeta_2)(\frac{t-1}{s}+1)+(\zeta_3+\zeta_1)(\frac{t-1}{s}-1))}. \end{multline*} It easy to show that the condition $\nabla_X(f)X=0$ is equivalent to the following system of equations: $$ \left\{\begin{array}{c} (\zeta_1+\zeta_3)(s-t+1)+(\zeta_1+\zeta_2)(s-t-1)=0 \\ (\zeta_1+\zeta_2)(s-t-1)+(\zeta_2+\zeta_3)(s+t-1)=0 \ \end{array}\right. $$ Analyzing this system for all invariant $f$-structures, we obtain the following result: \begin{propos} All invariant strictly Killing (i.e. non-K\"ahler) $f$-structures on the flag manifold $M=SU(3)/T_{max}$ and the corresponding invariant Riemannian metrics (up to homothety) are indicated below: $$ \begin{array}{ll} J_1=(1,1,1),&g=(1,1,1);\\ f_1=(1,1,0),&g=(3,3,4);\\ f_2=(1,0,1),&g=(3,4,3);\\ f_3=(0,1,1),&g=(4,3,3).\ \end{array} $$ In particular, there are no invariant Killing $f$-structures of rank $2$ on $M$. \end{propos} Note the structure $J_1$ is a well-known non-integrable nearly K\"ahler structure on a naturally reductive space $M$ (see \cite{G2}, \cite{G3}, \cite{Ki1}, \cite{AGI2} and others). The structures $f_1,f_2,f_3$ present first invariant non-trivial Killing $f$-structures \cite{B11}. The important feature of these structures is that the corresponding invariant Riemannian metrics are not Einstein (see Theorem \ref{t12}). It illustrates a substantial difference between non-trivial strictly Killing $f$-structures and strictly $NK$-structures at least in the $6$-dimensional case (see Theorem \ref{t3}). {\bf Remark 3.} It is interesting to note that all strictly Killing $f$-structures above indicated are canonical $f$-structures for suitable homogeneous $\Phi$-spaces of the Lie group $SU(3)$. We already mentioned that $M=SU(3)/T_{max}$ is a homogeneous $k$-symmetric space for any $k\ge 3$. It means $M$ as an underlying manifold could be generated by various automorphisms $\Phi$ of the Lie group $SU(3)$. In particular, $J_1$ is the canonical almost complex structure $J=\frac1{\sqrt{3}}(\theta-\theta^2)$ for the homogeneous $\Phi$-space of order $3$, where $\Phi=I(s),\; s=diag\:(\varepsilon,\overline \varepsilon,1),\;\varepsilon=\sqrt[3]{1}$ (see the beginning of this Section). Further, if we consider the automorphism $\Phi_1=I(s_1),\; s_1=diag\:(i,-i,1)$, where $i=\sqrt[4]{1}$ is the imaginary unit, then $M$ is a homogeneous $\Phi_1$-space of order $4$. The corresponding canonical $f$-structure $f=\frac12(\theta_1-\theta_{1}^3)$ for this $\Phi_1$-space just coincides (up to sign) with the $f$-structure $f_3=(0,1,1)$. The structures $f_1$ and $f_2$ can be obtained in the similar way. Moreover, all the structures $f_1,f_2,f_3$ and $f_7,f_8,f_9$ can be realized as canonical $f$-structures for suitable homogeneous $\Phi$-spaces of order $5$. We also note that all $f$-structures $f_1,f_2,f_3$ are just the restrictions of the structure $J_1$ onto the corresponding distributions $\frak{m}_p\oplus\frak{m}_q,\;p,q\in\{1,2,3\}.$ \subsection{\bf Nearly K\"ahler $f$-structures} Using (\ref{123}), we can easily obtain: \begin{multline*} \nabla_{fX}(f)fX=\frac12 D(\hat A,\hat B,\hat C), \text{where}\\ \hat A=\overline{-i\zeta_2\zeta_3bc((\zeta_1+\zeta_3)(1+s-t)+(\zeta_1+\zeta_2)(s-t-1))},\\ \hat B=\overline{-i\zeta_1\zeta_3ca((\zeta_2+\zeta_1)(1+\frac{1-s}{t})+ (\zeta_2+\zeta_3)(\frac{1-s}{t}-1))},\\ \hat C=\overline{-i\zeta_1\zeta_2ab((\zeta_3+\zeta_2)(\frac{t-1}{s}+1)+ (\zeta_3+\zeta_1)(\frac{t-1}{s}-1))}. \end{multline*} It follows that the condition $\nabla_{fX}(f)fX=0$ is reduced to the following system of equations: $$ \left\{\begin{array}{c} \zeta_2\zeta_3((\zeta_1+\zeta_3)(s-t+1)+(\zeta_1+\zeta_2)(s-t-1))=0 \\ \zeta_1\zeta_3((\zeta_2+\zeta_1)(1+t-s)+(\zeta_2+\zeta_3)(1-s-t))=0 \\ \zeta_1\zeta_2((\zeta_3+\zeta_2)(t+s-1)+(\zeta_3+\zeta_1)(t-s-1))=0 \ \end{array}\right. $$ Consideration of this system implies \begin{propos}\label{p3} The only invariant strictly nearly K\"ahler (i.e. non-K\"ahler) $f$-structure of rank $6$ on the flag manifold $M=SU(3)/T_{max}$ is the nearly K\"ahler structure $J_1=(1,1,1)$ with respect to the naturally reductive metric $g=(1,1,1)$. Invariant strictly nearly K\"ahler $f$-structures of rank $4$ and the corresponding invariant Riemannian metrics (up to homothety) on $M$ are: $$ \begin{array}{lll} f_1=(1,1,0),&g_s=(1,1,s),&s>0;\\ f_2=(1,0,1),&g_t=(1,t,1),&t>0;\\ f_3=(0,1,1),&g_t=(1,t,t),&t>0.\ \end{array} $$ The invariant $f$-structures $f_7,f_8,f_9$ of rank $2$ on $M$ are strictly $NKf$-structures with respect to all invariant Riemannian metrics $g=(1,t,s),\;t,s>0$. \end{propos} First we notice that the structures $f_1,f_2,f_3$ and $f_7,f_8,f_9$ provide invariant examples of $NKf$-structures with non-naturally reductive metrics on the homogeneous space $M=SU(3)/T_{max}$, which belongs to a semi-simple type. We can also observe that for any invariant strictly $NKf$-structure on $M$ there exists at least one (up to homothety) corresponding Einstein metric. More exactly, for these $NKf$-structures of rank $6$, $4$, and $2$ there are (up to homothety) $1$, $2$, and $4$ Einstein metrics respectively (see Theorem \ref{t12}). In a certain degree, it is a particular analogy with the result of Theorem \ref{t3}. This particular fact and some related general results lead to the following conjecture, which seems to be plausible: {\bf Conjecture}. For any strictly nearly K\"ahler $f$-structure on a $6$-dimensional manifold there exists at least one corresponding Einstein metric. {\bf Remark 4.} The invariant $f$-structures $f_4,f_5,f_6$ on the flag manifold $M=SU(3)/T_{max}$ cannot be canonical $f$-structures for all homogeneous $\Phi$-spaces of orders $4$ and $5$ of the Lie group $SU(3)$. It evidently follows by comparing the results in Theorem \ref{t8}, Theorem \ref{t9}, and Proposition \ref{p3}. \subsection{\bf Hermitian $f$-structures} First let us calculate the composition tensor $T$ (see formula (\ref{eqT})) for arbitrary invariant $f$-structure on $(M=SU(3)/T_{max},g=(1,t,s))$. Combining (\ref{123}) and (\ref{eqf}), we can obtain: \begin{multline}\label{T} T(X,Y)=\frac18 D(\check A,\check B,\check C), \text{where}\\ \check A=-\zeta_1\zeta_2\zeta_3(1+\zeta_2\zeta_3)((\zeta_1+\zeta_3)(1+s-t)\overline{bc_1}+ (\zeta_1+\zeta_2)(s-t-1)\overline{b_1c}),\\ \check B=-\zeta_1\zeta_2\zeta_3(1+\zeta_1\zeta_3)((\zeta_2+\zeta_1)(1+\frac{1-s}{t})\overline{ca_1}+ (\zeta_2+\zeta_3)(\frac{1-s}{t}-1)\overline{c_1a}),\\ \check C=-\zeta_1\zeta_2\zeta_3(1+\zeta_1\zeta_2)((\zeta_3+\zeta_2)(\frac{t-1}{s}+1)\overline{ab_1}+ (\zeta_3+\zeta_1)(\frac{t-1}{s}-1)\overline{a_1b}). \end{multline} We recall that the defining property for a Hermitian $f$-structure is the condition $T(X,Y)=0$. Now from (\ref{T}), we get the following result: \begin{propos}\label{p4} The invariant $f$-structures $J_2,J_3,J_4$ and $f_1,\dots,f_9$ are Hermitian $f$-structures with respect to all invariant Riemannian metrics $g=(1,t,s)$,\newline$t,s>0$ on the flag manifold $M=SU(3)/T_{max}$. \end{propos} Notice that the almost complex structure $J_1=(1,1,1)$ is non-integrable. It agrees with the fact that $J_1$ is not a Hermitian $f$-structure for each Riemannian metric. While we stress that all $f$-structures $f_1,\dots,f_9$ of rank $4$ and $2$ are non-integrable, but they are Hermitian $f$-structures. \subsection{\bf G$_1$f-structures} Finally, we consider the condition $T(X,X)=0$, which is the defining property for $G_1f$-structures. Using (\ref{T}) and taking into account Propositions \ref{p3} and \ref{p4}, we get \begin{propos}\label{p5} The flag manifold $M=SU(3)/T_{max}$ does not admit invariant strictly $G_1f$-structures (i.e. neither $NKf$-structures nor $Hf$-structures). In particular, there are no invariant strictly $G_1$-structures $J$ (i.e. neither nearly K\"ahler nor Hermitian) on $M$. \end{propos}
{ "timestamp": "2005-03-24T12:52:20", "yymm": "0503", "arxiv_id": "math/0503533", "language": "en", "url": "https://arxiv.org/abs/math/0503533" }