A matrix model of a non-Hermitian $\beta$-ensemble
Extracted source text · arXiv 2305.13184
\documentclass[12pt,a4paper]{article}
\usepackage[utf8]{inputenc}
\usepackage{amsmath,amsfonts,amssymb,amsthm}
%\usepackage{showlabels}
\usepackage{graphicx}
\usepackage[export]{adjustbox}
\usepackage{subcaption}
\usepackage{verbatim}
\usepackage{authblk}
\usepackage{tikz}
\usepackage{wrapfig}
% Biblatex
% \usepackage[style=numeric,sorting=nyt]{biblatex}
% \addbibresource{Ref.bib}
% \addbibresource{complexbeta.bib}
% New Additions
\usepackage{titlesec}
\usepackage{titlecaps}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{corollary}{Corollary}[theorem]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem*{conjecture}{Conjecture}
\newtheorem*{remark}{Remark}
\graphicspath{ {./images/} }
\usepackage{hyperref}
\usetikzlibrary{positioning}
% Handy Shortcuts
\newcommand{\R}{{\mathbb R}}
\newcommand{\N}{{\mathbb N}}
\newcommand{\C}{{\mathbb C}}
\newcommand{\Q}{{\mathbb Q}}
\newcommand{\Z}{{\mathbb Z}}
\newcommand{\V}{{\mathbb V}}
\newcommand{\A}{{\mathbb A}}
\newcommand{\I}{{\mathbb I}}
\newcommand{\E}{{\mathbb E}}
\newcommand{\T}{\mathrm{T}}
\newcommand{\Tw}{\widetilde{T}}
\newcommand{\Tb}{\mathcal{T}_{\beta}}
\newcommand{\CR}{\mathcal{R}}
\newcommand{\PC}{\mathcal{P}_j[\mathbb{C}]}
\newcommand{\XL}{X_{\mathrm{L}}}
\newcommand{\XR}{X_{\mathrm{R}}}
\renewcommand{\P}{{\mathbb P}}
\newcommand{\dz}{\partial_{\overline{z}}}
%\newcommand{\ov}{{\overline{}}
\usepackage{adjustbox}
\newcommand{\e}{\mathbf{e}}
\newcommand{\rv}{\mathbf{r}}
\newcommand{\lv}{\mathbf{l}}
\newcommand{\Tr}{{\mathrm{Tr}}}
\newcommand{\Rank}{{\mathrm{Rank}}}
\newcommand{\ord}{{\mathrm{ord}}}
\newcommand{\bl}{\Big\langle}
\newcommand{\br}{\Big\rangle}
\newcommand{\x}{\boldsymbol{x}}
\newcommand{\ii}{{\mathrm i}}
\newcommand{\dd}{{\mathrm d}}
\newcommand{\blambda}{\boldsymbol{\lambda}}
\newcommand{\abs}[1]{\left| #1 \right|}
\newcommand{\av}{\mathbf{a}}
\newcommand{\bv}{\mathbf{b}}
\newcommand{\jpdf}{\emph{j.p.d.f.}}
\newcommand{\GL}{\mathrm{GL}}
\newcommand{\norm}[1]{\left \| #1 \right \|}
\DeclareMathOperator{\diag}{diag}
\DeclareMathOperator{\re}{Re}
\DeclareMathOperator{\im}{Im}
\DeclareMathOperator{\adj}{adj}
\newcommand{\supp}{\mathrm{supp}}
\renewcommand{\deg}{\mathrm{deg}}
\renewcommand{\arg}{\mathrm{arg}}
\newcommand{\sbq}{\subseteq}
\newcommand{\spq}{\supseteq}
\renewcommand{\epsilon}{\varepsilon}
\newcommand{\qandq}{\quad \mathrm{and} \quad}
\newcommand{\qasq}{\quad \mathrm{as} \quad }
\newcommand{\qforq}{\quad \mathrm{for} \quad }
\newcommand{\sgn}{\mathrm{sgn}}
\newcommand{\be}{\begin{equation}}
\newcommand{\ee}{\end{equation}}
\numberwithin{equation}{section}
%\usepackage{amsfonts}
\usepackage{bbm}
\usepackage{nicefrac}
\usepackage{mathtools}
\usepackage{soul}
\begin{document}
\title{A matrix model of a non-Hermitian $\beta$-ensemble}
\author{Francesco Mezzadri\footnote{E-mail: f.mezzadri@bristol.ac.uk}}
\author{Henry Taylor\footnote{E-mail: ht17630@bristol.ac.uk}}
\affil{\small{\emph{School of Mathematics, University of Bristol, Fry Building,
Woodland Road, Bristol BS8 1UG, UK}}}
\date{}
\maketitle
\begin{abstract}
We introduce the first random matrix model of a complex $\beta$-ensemble.
The matrices are tridiagonal and can be thought of as the non-Hermitian
analogue of the Hermite $\beta$-ensembles
discovered by Dumitriu and Edelman (\emph{J. Math. Phys.,}
\textbf{43} (2002), 5830).
The main feature of the model is that the exponent $\beta$ of the Vandermonde
determinant in the joint probability
density function (\emph{j.p.d.f.}) of the eigenvalues can take
any value in $\mathbb{R}_+$. However, when $\beta=2$, the \emph{j.p.d.f.}
does not reduce to that of the Ginibre ensemble, but it contains an extra
factor expressed as a multidimensional integral over the space of the eigenvectors.
\end{abstract}
\section{Introduction}
\label{introduction}
In 1965 Ginibre~\cite{Gin65} computed the \emph{j.p.d.f.} of the eigenvalues
of non-Hermitian matrices whose elements are independent
complex normal random variables, and showed that
\begin{equation}
\label{complex_ginibre}
P_2(\blambda) = \frac{1}{Z_2} \prod_{j=1}^n
\exp\left(-\abs{\lambda_j}^2\right)\prod_{1\le j < k \le n}\abs{\lambda_k - \lambda_j}^\beta,
\end{equation}
where $\blambda = (\lambda_1,\dotsc,\lambda_n)\in \mathbb{C}^n$ is the spectrum, $\beta=2$, $n$ is the dimension of the matrices and $Z_2$ is a normalisation constant.
This ensemble of random matrices is known in the literature
as \emph{Ginibre Unitary Ensemble,} or GinUE.
Formula~\eqref{complex_ginibre} can be interpreted as the
equilibrium distribution of a two-dimensional system of unit point charges,
which interact with logarithmic repulsion in a quadratic confining
potential at inverse temperature $\beta=2$. This is known as
\emph{Dyson's Coulomb gas model.}
The realization of a two-dimensional gas of charged particles in terms of eigenvalues of matrices in the GinUE means that
at $\beta=2$ the model is exactly integrable, and can be studied analytically in great detail. Recent advances in the theory
of non-Hermitian random matrices have prompted a large body of research into
this statistical mechanical model. Non-Hermitian random matrices
have found applications in quantum chromodynamics, growth problems, scattering theory,
open quantum systems and quantum chaos. (See the review articles~\cite{BF22,FS03} and references therein.)
Given the success of Dyson's Coulomb gas model, it is natural to ask
whether it is possible to create a \emph{complex} $\beta$-\emph{matrix model,}
analogous to the tridiagonal $\beta$-Hermite ensemble of Hermitian matrices introduced
by Dumitriu and Edelman~\cite{DE02} almost forty years after the work of Ginibre.
% \begin{equation}
% \label{complex_tri}
% P_\beta(\lambda_1,\dotsc,\lambda_n) = \frac{1}{Z_\beta} \prod_{j=1}^n
% \exp\left(-\abs{\lambda_j}^2\right)\prod_{1\le j < k \le n}\abs{\lambda_k -
% \lambda_j}^\beta,
% \end{equation}
The purpose of this article is to introduce the first non-Hermitian $\beta$-ensemble.
Consider the complex tridiagonal matrix
\begin{equation}
\label{eq:tridiagmatrix}
T = \begin{pmatrix}
a_n & 1 & & & \\
b_{n-1} & a_{n-1} & 1 & & \\
& \ddots & \ddots & \ddots & \\
& & b_2 & a_2 & 1 \\
& & & b_1 & a_1
\end{pmatrix}.
\end{equation}
The elements of $T$ are independent random variables with distributions
\begin{subequations}
\label{mat_el_dis}
\begin{gather}
\re(a_j) \sim \mathcal{N}(0,1), \quad \im(a_j) \sim \mathcal{N}(0,1)\quad j=1,\dotsc,n, \\
\abs{b_j} \sim \chi_{\beta j/2}, \quad j=1,\dotsc,n-1, \\
\arg (b_j) \sim \mathcal{U}(0,2\pi), \quad j=1,\dotsc,n-1,
\end{gather}
\end{subequations}
where $\mathcal{N}(\mu,\sigma)$ denotes
the normal distribution with mean $\mu$ and standard deviation
$\sigma$, $\mathcal{U}(a,b)$ the uniform distribution in $[a,b)$ and
$\chi_\alpha$ the chi-distribution with parameter $\alpha \in
\mathbb{R}_+$. Recall that the probability density
function of a chi-distributed random variable is
\[
f_{\alpha}(x) =\begin{cases}
\frac{2^{1-\alpha/2}}{\Gamma(\frac{\alpha}{2})}e^{-x^2/2}x^{\alpha-1}&\text{for}\hspace{2mm}x\geq0, \; \alpha >0, \\
0&\text{otherwise}.
\end{cases}
\]
The main result of this article is the following theorem.
\begin{theorem}
\label{main_theorem}
The j.p.d.f. of the eigenvalues of the random matrix~\eqref{eq:tridiagmatrix} is
\begin{equation}
\label{eq:EigDens1}
P_\beta(\blambda) = \frac{1}{Z_\beta} \prod_{j=1}^n
\exp\left(-\frac{\abs{\lambda_j}^2}{2}\right)\prod_{1\le j < k \le n}\abs{\lambda_k -
\lambda_j}^\beta f(\blambda),
\end{equation}
where $\beta\in \mathbb{R}_+$, $Z_\beta$ is a normalization constant and
\begin{equation}
\label{factor}
f(\blambda) =\int_{\mathbb{C}^{n-1}}
\exp\big(-g(\blambda,\mathbf{r})\big)\prod_{k=1}^{n}|r_k|^{\frac{\beta}{2}-1}d^2r_1\dotsm
d^2r_{n-1},
\end{equation}
$\mathbf{r} =(r_1,\dotsc,r_{n}) \in \mathbb{C}^{n}$, $r_n = -r_1 - \dotsb - r_{n-1}$ and
$ d^2 r_k = d\abs{r_k} d\phi_{r_k}$.
The exponent $g(\blambda,\mathbf{r})$ is a positive-definite rational function. Furthermore, we have
\begin{enumerate}
\item $g(\blambda,\mathbf{r}) = O\left( \abs{\lambda_j}^4\right)$ as $\lambda_j \to \infty$;
\item $g(\blambda e^{i\phi},\rv)=g(\blambda,\rv)$, $\,\phi \in [0,2\pi)$.
\end{enumerate}
\end{theorem}
A consequence of property 2 is that the density of
the eigenvalues is invariant under rotations, see Fig.~\ref{fig:image2}.
Note that formulae~\eqref{eq:EigDens1} and~\eqref{factor} imply
that the eigenvalues and eigenvectors are not statistically independent. This is also the case for the GinUE, but while in the Ginibre ensemble we can integrate
over the eigenvectors, in this model such integration is not
straightforward; therefore, the factor $f(\blambda)$ can only be
studied through its integral representation~\eqref{factor}. In general,
we would expect that for large matrix dimension
the local statistics of the eigenvalues should be governed
by the factor
\[
\prod_{1\le j < k \le n}\abs{\lambda_k -
\lambda_j}^\beta
\]
in the \emph{j.p.d.f.} of the eigenvalues~\eqref{eq:EigDens1}. Thus, at
leading order in the matrix size the local behaviour of the eigenvalues should not be affected by $f(\blambda).$
\begin{figure}[ht]
\begin{subfigure}{0.5\textwidth}
\includegraphics[width=1\linewidth, height=6cm]{beta_2_n_5000_m_100.jpg}
\caption{$\beta=2$ and $n=5000$}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\includegraphics[width=1\linewidth, height=6cm]{beta_6_n_5000_m_100.jpg}
\caption{$\beta=6$ and $n=5000$}
\end{subfigure}
\caption{Density-of-states of the complex eigenvalue spectrum of $(\ref{eq:tridiagmatrix})$ for 100 matrices of varying $\beta$ and dimension $5000\times 5000$, whose elements are distributed according to $(\ref{mat_el_dis})$.}
\label{fig:image2}
\end{figure}
The matrices~\eqref{eq:tridiagmatrix} can be thought of as the complex analogues
of the tridiagonal model discovered by Dumitriu and Edelman~\cite{DE02}, the
$\beta$-Hermite ensemble. An important difference, however, is that
an arbitrary matrix in $\GL(n,\mathbb{C})$ cannot be
reduced to the form~\eqref{eq:tridiagmatrix} by a similarity transformation,
while a Hermitian matrix can always be mapped to a Jacobi matrix by applying an appropriate sequence of Householder reflections.
The classical theory of random matrices developed in the 1950s and 1960s by
Wigner, Dyson and others~(see, \emph{e.g.}, \cite{For10,Meh04}
for a discussion) is centred around three fundamental classes of ensembles, which correspond to the three types of associative division algebra with real coefficients, namely the fields of the real, complex and
quaternion numbers. This classification is known as Dyson's
\emph{threefold way}~\cite{Dys62e}.
Typical examples are the the Gaussian Orthogonal Ensemble (GOE),
Gaussian Unitary Ensemble (GUE) and the Gaussian
Symplectic Ensemble (GSE); they are composed of real symmetric, complex Hermitian and self-dual real
quaternion matrices, respectively, whose independent elements are normal random variables. The \emph{j.p.d.f.} of the eigenvalues is given by
\begin{equation}
\label{gaussian_jpdf}
P_\beta(\blambda) = \frac{1}{Z_\beta}\prod_{j=1}^n \exp\left(-\frac{\beta}{2}
\lambda_j^2\right) \prod_{1\le j < k\le n}\abs{\lambda_k-\lambda_j}^\beta,
\end{equation}
where $n$ is the dimension of the matrices and $\lambda_j\in \R$, $j=1,\dotsc,n$.
The parameter $\beta$ is the dimension of the division algebra: $\beta=1$ for the
GOE; $\beta=2$ for the GUE; $\beta=4$ for the GSE.
Dyson~\cite{Dys62a,Dys62b,Dys62c} extended this theory to unitary matrices and
introduced the Circular Unitary Ensemble (CUE),
Circular Orthogonal Ensemble (COE) and Circular Symplectic Ensemble (CSE),
which fall into the same classification scheme.
Zirnbauer~\cite{Zir96}, Altland and Zirnbauer~\cite{AZ97} and
later Due\~{n}ez~\cite{Due01,Due04} generalised
Dyson's threefold
way to include all Cartan's symmetric spaces.
In a pioneering article, Dumitriu and Edelman~\cite{DE02} discovered a new class of
random matrix models, the $\beta$-Hermite ensemble, which are made of
tridiagonal matrices whose \emph{j.p.d.f.} of the eigenvalues is formally the same
as that in Eq.~\eqref{gaussian_jpdf}, but with the crucial feature that the exponent
$\beta$ of the Vandermonde determinant can be any positive real number and does not
identify a particular symmetric space. They also introduced the $\beta$-Laguerre
ensembles. In 2004 Killip and Nenciu~\cite{KN04} extended the theory of Edelman and Dumitriu
to matrix models of unitary matrices and generalised Dyson's circular ensembles
to any $\beta\in \mathbb{R}_+$. They also discovered the $\beta$-Jacobi
ensembles. Dumitriu and Forrester~\cite{DF10} gave a tridiagonal realisation
of the antisymmetric Gaussian $\beta$-ensemble. Forrester and Rains~\cite{FR05}
studied one-parameter interpolation of classical matrix ensembles. The impact of the
work of Dumitriu and Edelman~\cite{DE02} cannot be understated;
it opened new areas of research in random matrix theory, the theory of
stochastic operators, Anderson localisation, integrable systems, statistical mechanics
and extreme value statistics, see \emph{e.g.}
\cite{AMP22,AS21,BLY12,BLY14,BFS07,ES07,For21,GM23,GGGM23,KRV16,MPS20,NS15,Spo21,
SS12,SS15}.
The article is structured as follows: in Sec.~\ref{tridiag_mat} we introduce
the model and establish the map between the spectral parameters and
the matrix~\eqref{eq:tridiagmatrix}; Sec.~\ref{sec:Jacobian} is dedicated
to the computation of the Jacobian of this map;
in Sec.~\ref{sec:MainThm} we complete the proof of Theorem~\ref{main_theorem}.
\section{Tridiagonal complex matrices}
\label{tridiag_mat}
In order to prove Theorem~\ref{main_theorem}, we need to establish a few important
properties of non-Hermitian tridiagonal matrices.
Note that an arbitrary complex tridiagonal matrix is similar to a matrix of the form~\eqref{eq:tridiagmatrix}. Indeed, write
\be
\label{eq:gen_trid}
\widetilde{T}= \begin{pmatrix}
a_n & c_{n-1} & & \\
\widetilde{b}_{n-1} & a_{n-1} & c_{n-2} & \\
& \ddots & \ddots & \\
& & \widetilde{b}_1 & a_1
\end{pmatrix},
\ee
with $a_j,\widetilde{b}_j,c_j\in\mathbb{C}$ and $\widetilde{b}_j\neq 0$, $c_j \neq 0$. Then, $T = C \widetilde{T} C^{-1}$,
where $b_j = c_j \widetilde{b}_j$ and
\[
C = \begin{pmatrix} 1 & & & \\
& c_{n-1} & & \\
& & \ddots &\\
& & & c_{n-1}\dotsm c_1
\end{pmatrix}.
\]
The matrices $T$ and $\widetilde T$ have the same spectrum, but $T$ has a
smaller number of degrees of freedom, which makes the analysis of its spectrum
more amenable. Since we are only interested in the eigenvalues, many
properties of $T$ that we will
be discussing in this section also hold for a general tridiagonal matrix. It is
important to emphasize that the eigenvectors are not invariant under a similarity
transformation. Since the eigenvectors and eigenvalues of $T$ are not
statistically independent, the function $f(\blambda)$
in the \emph{j.p.d.f.} of the eigenvalues of $\widetilde{T}$ would be different
from that defined in Eq.~\eqref{factor}.
Denote by $\T(n)$ the set of matrices of the
form~\eqref{eq:tridiagmatrix} whose spectrum
is non-degenrate\footnote{While Jacobi matrices have a simple spectrum,
the eigenvalues of the matrix ~\eqref{eq:tridiagmatrix} may be degenerate.
See Appendix~\ref{multiplicity} for a
discussion about tridiagonal complex matrices with degenerate spectra.}
and by $\Lambda(n)$ the set of $n \times n$ simple
complex diagonal matrices. Write
\[
\mathrm{R}(n) = \left\{R \in \mathrm{GL}(n,\C) : R^{-1}TR = \Lambda,\: T \in \T(n)\; \text{and} \; \Lambda \in \Lambda(n)\right\}.
\]
Let $D$ be diagonal such that $\det D \neq 0$.
If $R$ diagonalises $T$, so does $RD$; therefore, the spectral decomposition
\begin{equation}
\label{diag_fact}
T = R\Lambda R^{-1}
\end{equation}
is not unique. It is natural to introduce the quotient space
\[
\CR(n)= \mathrm{R}(n)/\mathrm{D}(n),
\]
where $\mathrm{D}(n)$ is the set of non-singular diagonal matrices.
Since the spectrum is simple, the factorization~\eqref{diag_fact} defines a
bijection
\[
\mathcal{S} : T(n) \longrightarrow \Lambda(n) \times \CR(n).
\]
Write
\begin{align*}
\av &=(a_1,\dotsc,a_n), & \bv &=(b_1,\dotsc,b_{n-1}),\\
\blambda &=(\lambda_1,\dotsc,\lambda_n) &
\rv & = (r_1,\dots,r_n).
\end{align*}
In order to prove
Theorem~\ref{main_theorem}, we need to find a one-to-one map
\begin{equation}
\label{bijection2}
(\av,\bv) \mapsto (\blambda,\rv),
\end{equation}
that preserves the spectral decomposition~\eqref{diag_fact}.\footnote{The
careful reader will have noticed that $\rv$ contains an extra component.
Indeed, the $r_j$'s are related by Eq.~\eqref{setrep}. However, $r_n$ will
appear in most of the calculations that will follow. It is simply convenient to
include it in the definition of $\rv.$} More precisely,
the transformation~\eqref{bijection2} must satisfy the following two properties:
\begin{enumerate}
\item the coordinates $\blambda$ and $\rv$
span $\Lambda(n)$ and $\CR(n)$, respectively;
\item the knowledge of $\blambda$ and $\rv$ is enough to reconstruct
the tridiagonal matrix $T$ from the right-hand side of Eq.~\eqref{diag_fact} uniquely.
\end{enumerate}
After having constructed such a map, we need to compute
its Jacobian. This will be achieved in Sec.~\ref{sec:Jacobian}. The first step
is to choose a suitable set of representatives of the equivalent classes in $\CR(n).$
\begin{lemma}
\label{parametrazion}
Let $\rv= (r_1,\dots,r_n)$ be the first row of $R\in \mathrm{R}(n)$. The subset of $\mathrm{R}(n)$ such that
\begin{equation}
\label{setrep}
r_1 + \dotsb + r_n =1
\end{equation}
spans a set of representatives in $\CR(n)$.
\end{lemma}
\begin{proof}
Take $\tilde{R}\in \mathrm{R}(n)$ and write $R=\tilde{R}D$, where $D\in \Lambda(n)$. We can always choose $D$ such that the elements of the first column of $R^{-1}$ are all $1.$ The identity
$R^{-1}R=I$, where $I=\diag(1,\dotsc,1)$, implies Eq.~\eqref{setrep}.
\end{proof}
The columns of $R$ are the right eigenvectors of $T$ and the
rows of $R^{-1}$ are its left eigenvectors. Therefore, $r_1,\dotsc,r_n$ are the
first elements of each right eigenvector of $T$.
Fixing a set of representatives in $\CR$ is tantamount to choosing the
normalization of the right and left eigenvectors. Let us set
\begin{equation}
\label{eq:newnot}
R =\begin{pmatrix}\rv_1^t\\ \vdots \\ \rv_n^t\end{pmatrix}, \quad
R^{-1}= \left(\lv_1,\dotsc,\lv_n\right),
\end{equation}
where
\[
\rv_j^t = \left( r_{j1},\dotsc, r_{jn}\right), \quad
\lv_j = \begin{pmatrix} l_{1j} \\ \vdots \\ l_{nj} \end{pmatrix}, \quad j=1,\dotsc,n.
\]
We write
\begin{equation*}
\rv^t_j \lv_{k} = r_{j1}l_{1k}+ \dotsb + r_{jn}l_{nk} = \delta_{jk}, \quad
j,k=1,\dotsc,n,
\end{equation*}
where the notation on the left-hand side stands for matrix multiplication between
the row vector $\rv_j^t$ and the column vector $\lv_k$. The set of
representatives defined in Lemma~\ref{parametrazion} is fixed by taking
$\rv=\rv_1^t$ and
\begin{equation}
\label{one_vec}
\lv_1=\begin{pmatrix} 1\\ \vdots \\ 1 \end{pmatrix}.
\end{equation}
The following theorem answers point 1 and 2.
\begin{theorem}
\label{theorem:bijection}
Let $T\in \T(n)$ and $T=R\Lambda R^{-1}$ be its spectral decomposition, where
$R$ satisfies the hypotheses of Lemma~{\rm \ref{parametrazion}}.
Denote by $\rv$ the first row of $R.$
Then, $T$ is uniquely determined by $\Lambda$ and $\rv$.
\end{theorem}
\begin{proof}
% Furthermore, the hypotheses of the theorem state that
% \be
% \label{eq:bc}
% \lv_1 = \begin{pmatrix} 1 \\ \vdots \\ 1 \end{pmatrix}.
% \ee
Let us write the matrix equations
\[
TR = R\Lambda \quad \text{and} \quad R^{-1}T = \Lambda R^{-1}
\]
as a set of vector equations:
\begin{subequations}
\label{eq:rec_r}
\begin{align}
\label{eq:rec_ra}
\rv_j^t\Lambda & = b_{n-j+1}\rv_{j-1}^t + a_{n-j+1}\rv_j^t + \rv_{j+ 1}^t, \\
\label{eq:rec_rb}
\Lambda\lv_j & = \lv_{j-1} + a_{n-j+1}\lv_j +b_{n-j}\lv_{j+1},
\end{align}
\end{subequations}
for $j=1,\dotsc,n$, where we have adopted the notation~\eqref{eq:newnot}. We want to show that given $\Lambda$ and $\rv_1^t$ with $\rv^t_1\lv_1=1$, where $\lv_1$ is defined in Eq.~\eqref{one_vec}, we can reconstruct $T$, $R$ and $R^{-1}$ uniquely from the recurrence relations~\eqref{eq:rec_r} with boundary conditions $b_n=0$ and $\lv_0=\mathbf{0}$,
$\rv_{n+1}^t=\mathbf{0}$. The vectors $\rv_0^t$ and $\lv_{n+1}$ are undetermined.
Write
\begin{subequations}
\label{eq:sol_rec_r1}
\begin{align}
\label{eq:an}
a_n & = \rv_1^t \Lambda \lv_1,\\
\label{eq:r2}
\rv_2^t & = \rv_1^t \Lambda - a_n \rv_1^t,\\
\label{eq:bn1}
b_{n-1} & = \rv_2^t \Lambda \lv_1,\\
\label{eq:l2}
\lv_2 &= \frac{1}{b_{n-1}}\left(\Lambda \lv_1 - a_n \lv_1\right).
\end{align}
\end{subequations}
The quantities $a_n$, $b_{n-1}$, $\rv_1^t$ and $\lv_1$ solve~\eqref{eq:rec_r}
for $j=1$. We need to check that they are consistent with the
condition $RR^{-1} = I$. Eqs.\eqref{eq:an} and~\eqref{eq:r2} give $\rv_2^t \lv_1=0$. Similarly, by Eqs.~\eqref{eq:an}
and~\eqref{eq:l2} we have $\rv_1^t \lv_2 =0$. Finally, we have
\begin{equation*}
\rv_2^t \lv_2 = \frac{1}{b_{n-1}} \left(\rv_2^t \Lambda
\lv_1 - \rv_2^t \lv_1\right) = 1.
\end{equation*}
Note that combining the orthogonality relations $\rv_1^t \lv_2 = 0$ and $\rv_2^t \lv_2=1$ with Eq.~\eqref{eq:r2} gives
\[
\rv_1^t \Lambda \lv_2 = 1,
\]
as it should be.
When $1<j<n$ take
\[
\rv_{j-1}^t,\: \rv_j^t,\: \lv_{j- 1},\: \lv_j
\]
subject to the conditions
\begin{subequations}
\label{eq:int_ic}
\begin{gather}
\rv^t_j \lv_j= \rv^t_{j-1}\lv_{j-1}=1, \quad \rv^t_j \lv_{j-1}= \rv^t_{j-1} \lv_{j}=0,\\
b_{n - j +1} = \rv_{j}^t \Lambda \lv_{j-1}, \quad \rv_{j-1}^t \Lambda \lv_{j}=1.
\end{gather}
\end{subequations}
Define
\begin{subequations}
\label{eq:sol_rec_rn}
\begin{align}
\label{eq:anj}
a_{n-j+1} & = \rv_j^t \Lambda \lv_j,\\
\label{eq:rj}
\rv_{j+1}^t & = \rv_j^t \Lambda - a_{n - j + 1}\rv_j^t - b_{n-j+1}\rv_{j-1}^t\\
\label{eq:bnj}
b_{n-j} & = \rv_{j+1}^t \Lambda \lv_j,\\
\label{eq:lj}
\lv_{j+1} &= \frac{1}{b_{n-j}}\left(\Lambda \lv_j- a_{n - j +1}\lv_j-\lv_{j-1}\right).
\end{align}
\end{subequations}
Combining~\eqref{eq:int_ic} with~\eqref{eq:sol_rec_rn} and proceeding in a similar way as when $j=1$, we can show that
\begin{align*}
\rv_{j+1}^t \lv_{j} &= 0, & \rv_{j+1}^t \lv_{j-1}& =0,\\
\rv_{j}^t \lv_{j+1} &= 0, & \rv_{j-1}^t \lv_{j+1}&=0.
\end{align*}
Furthermore, we have
\begin{align*}
\rv_{j+1}^t \lv_{j+1}& = \frac{1}{b_{n-j}}\left(\rv_{j+1}^t
\Lambda \lv_j- a_{n - j +1}
\rv_{j+1}^t \lv_j-\rv_{j+1}^t \lv_{j-1}\right)=1\\
\rv_{j}^t \Lambda \lv_{j+1} & = b_{n-j+1}\rv_{j-1}^t \lv_{j+1}
+ a_{n-j+1}\rv_j^t \lv_{j+1}+ \rv^t_{j+ 1} \lv_{j+1} =1.
\end{align*}
The orthogonality relations
\[
\rv_{j+1}^t \lv_{k}=0, \quad \rv_k^t \lv_{j+1}=0, \quad k=1, \dots, j-2
\]
follow by induction and from the fact that by construction
$\rv^t_j \Lambda \lv_k = 0$ if $\abs{j-k}>1$.
The boundary condition $\rv^t_{n+1}=\mathbf{0}$ terminates the recurrence relations.
\end{proof}
\section{The Jacobian}
\label{sec:Jacobian}
If $z \in \C$, we write $dz = d\re(z) + id\im(z)$ and
\begin{align*}
d\mathbf{a} & = \bigwedge_{k=1}^n da_k, & d\mathbf{b} & =
\bigwedge_{k=1}^{n-1}db_k,\\
d\rv & = \bigwedge_{k=1}^{n-1}dr_{k}, &
d\blambda & = \bigwedge_{k=1}^n d\lambda_k.
\end{align*}
Furthermore, we have
\begin{align*}
d^2 a_j & = d\re(a_j)d\im(a_j) = \frac{i}{2} da_j \wedge d\bar{a}_j, \\
d^2\lambda_j & = d\re\left(\lambda_j\right)\im(\lambda_j) = \frac{i}{2}d\lambda_j \wedge d\overline{\lambda}_j,\\
d^2 b_k & = d\abs{b_k} d\phi_{b_k} = \frac{i}{2\abs{b_k}}db_k \wedge d\overline{b}_k, \\
d^2 r_k & = d\abs{r_k} d\phi_{r_k} = \frac{i}{2\abs{r_k}}dr_k \wedge d\overline{r}_k,
\end{align*}
where $j=1,\dotsc,n$, $k=1,\dotsc, n-1$. Let us set
\begin{equation*}
d\mu(\mathbf{a},\mathbf{b}) = \prod_{j=1}^n d^2a_j \prod_{k=1}^{n-1}d^2b_k, \quad
d\mu(\mathbf{r},\blambda) = \prod_{j=1}^n d^2\lambda_j \prod_{k=1}^{n-1}d^2r_k.
\end{equation*}
Let
\[
d\av \wedge d\bv = J d\rv \wedge d \blambda,
\]
where
\begin{equation*}
%\label{eq:Jac_def}
J = \det \left \{\frac{\partial\left(\mathbf{a},\mathbf{b}\right)}{\partial \left( \rv, \boldsymbol{\lambda}\right)}\right \}.
\end{equation*}
The following theorem generalises well known results for Jacobi
matrices~\cite{DE02,FR06} to the model~\eqref{eq:tridiagmatrix}.
\begin{theorem}
\label{thm:jac}
We have
\begin{align*}
J &= \frac{\prod_{j=1}^{n-1}b_j }{\prod_{k=1}^n r_k},\\
d\mu(\mathbf{a},\mathbf{b}) &= \frac{\prod_{j=1}^{n-1}\abs{b_j}}{\prod_{k=1}^n \abs{r_k}}d\mu(\mathbf{r},\blambda).
\end{align*}
\end{theorem}
In order to prove this theorem, we need a few preliminary results. Denote by $\widetilde{T}_j$ the principal submatrix of $\widetilde T$ obtained by keeping the first $j$
rows and columns. We have $\widetilde T_n = \widetilde T$. Introduce the characteristic polynomial
\begin{equation}
\label{cp_sm}
\chi_j(z) =\det\left(zI - \widetilde{T}_j\right), \quad j=1,\dotsc,n,
\end{equation}
and set $\chi_0(z)=1$. Let $\mathcal{P}_j[\mathbb{C}]$ be the set of monic
polynomials of degree $j$ over $\C$ and $\e_j$ be a canonical basis vector in $\C^n$,
\emph{i.e.} $\e_j^t
=\overset{j}{(0,\dotsc,1,\dotsc,0)}$.
\begin{proposition}
\label{prop}
Let $j=1,\dots,n-1.$ The following statements hold:
\begin{enumerate}
\item $\bigl|\widetilde{b}_{n-1}\widetilde{b}_{n-2}\dotsm \widetilde{b}_{n-j}\bigr|=
\bigl \|\chi_j(\widetilde{T})\e_1\bigr \|=\underset{\psi \in \PC}
{\min}\bigl \|\psi(\widetilde{T})\e_1\bigr \|$;
\item $\chi_j(\widetilde{T})\e_1=\e_{j+1}(\widetilde{b}_{n-1}\widetilde{b}_{n-2}\dotsm \widetilde{b}_{n-j})$;
\item $|c_{n-1}c_{n-2}\dotsm c_{n -j}|=
\bigl \|\e_1^t\chi_j(\widetilde{T})\bigr \|=\underset{\psi \in \PC}
{\min}\bigl \|\e_1^t\psi(\widetilde{T})\bigr\|$;
\item $\e_1^t\chi_j(\widetilde{T})= \e_{j+1}^t(c_{n-1}c_{n-2}\dotsm c_{n-j})$.
\end{enumerate}
\end{proposition}
\begin{proof}
One can prove by induction that
\[
\e_k^t\widetilde{T}^j\mathbf{e}_1= \begin{cases} \widetilde{b}_{n-1}\widetilde{b}_{n-2}\dotsm \widetilde{b}_{n-j}, & \text{if $k=j+1$,}\\
0, & \text{if $k>j+1$.} \end{cases}
\]
It follows that for any $\psi \in \PC$ we have
\[
\bigl \|\psi(\widetilde{T})\e_1\bigr \| \ge \bigl |\widetilde{b}_{n-1}\widetilde{b}_{n-2}\dotsm \widetilde{b}_{n-j}\bigr |.
\]
This lower bound is sharp if there exist a $\psi\in \PC$ such that
\[
\psi(\widetilde{T})\e_1 = \e_{j+1}\bigl(\widetilde{b}_{n-1}\widetilde{b}_{n-2}\dotsm \widetilde{b}_{n-j}\bigr).
\]
Note that the first $j$ components of $\widetilde{T}^j\e_1$ contain only elements
from $\widetilde{T}_j$, as do the vectors $\widetilde{T}^k\mathbf{e}_1$ for $1\le k<j$.
Let
\[
E_j=(\mathbf{e}_1,\dotsc ,\mathbf{e}_j).
\]
Then we have
\[
E_j^t\psi(\widetilde{T})\e_1 = \psi(\widetilde{T}_j)\e_1,
\]
where, with an abuse of notation, $\e_1$ on the left-hand
side belongs to $\C^n$, while on the right-hand side $\e_1\in \C^j$.
The Cayley-Hamilton theorem states that $\chi_j(\widetilde{T}_j)=0$.
It follows that
\[
\chi_j(\widetilde{T})\e_1 = \e_{j+1}\bigl(\widetilde{b}_{n-1}\widetilde{b}_{n-2}\dotsm \widetilde{b}_{n-j}\bigr).
\]
This proves statements 1. and 2. The proof of 3. and 4. is analogous.
\end{proof}
\begin{lemma}
\label{lem:jac1}
Let $T=R\Lambda R^{-1}$ be the spectral decomposition of $T\in \T(n)$, where $R$ is defined in~\eqref{eq:newnot} and
$\Lambda = \diag(\lambda_1, \dotsc,\lambda_n)$. We have
\[
\Delta(\boldsymbol{\lambda})^2=\prod_{1\le j<k\le n}\left(\lambda_k-\lambda_j\right)^2=\frac{\prod_{j=1}^{n-1}b_j^j}{\prod_{j=1}^{n}r_{j}},
\]
where $\rv = (r_1,\dotsc,r_n)$ is the first row of $R$ and
\[
r_1 + \dotsb + r_n =1.
\]
\end{lemma}
\begin{proof}
The definition of $T$ and Proposition~\ref{prop} give
\begin{subequations}
\label{eq:minstat}
\begin{align}
\label{eq:lem231}
\chi_j(T)\mathbf{e}_1 & =\mathbf{e}_{j+1}(b_{n-1}b_{n-2}\dotsm b_{n-j}),\\
\label{eq:lem232}
\mathbf{e}_1^t\chi_j(T) &= \mathbf{e}_{j+1}^t,
\end{align}
\end{subequations}
for $j=1,2,\dotsc,n-1$. Using the spectral decomposition $T = R\Lambda R^{-1}$ and the definition of $R$ given in~\eqref{eq:newnot}, Eqs.~\eqref{eq:minstat} become
\begin{align*}
\chi_j(\Lambda)R^{-1}\mathbf{e}_1 & = \chi_j(\Lambda) \lv_1 = R^{-1}\mathbf{e}_{j+1}(b_{n-1}b_{n-2}\dotsm b_{n-j}) \nonumber \\
& = \lv_{j+1}(b_{n-1}b_{n-2} \dotsm b_{n-j}),\\
\mathbf{e}_1^tR\chi_j(\Lambda) &= \rv_1^t\chi_j(\Lambda) = \mathbf{e}_{j+1}^t R = \rv_{j+1}^t.
\end{align*}
Multiplying these equations side by side gives
\[
\rv_1^t\chi_j(\Lambda) \chi_k(\Lambda)\lv_1 =
\rv_{j+1}^t \lv_{k+1}\prod_{m=1}^k b_{n-m},
\]
which can be rewritten as
\begin{equation}
\label{eq:lem235}
\sum_{l=1}^n r_l\chi_j(\lambda_l)\chi_k(\lambda_l)=
\begin{cases} 1 & \text{if $j=k=0$,} \\
\delta_{jk}\prod_{m=1}^k b_{n-m}, & \text{if $j \neq 0$ or $k \neq 0$,}
\end{cases}
\end{equation}
for $j,k=0,\dotsc, n-1$, where we have used $\lv_1^t = (1,\dotsc,1)$ and the fact that by
definition $\chi_0(z)=1.$
Write
\begin{equation*}
W = \begin{pmatrix} 1 & \chi_1(\lambda_1) & \dots &\chi_{n-1} (\lambda_{1}) \\
1 & \chi_1(\lambda_2) & \dots &\chi_{n-1} (\lambda_{2}) \\
\hdotsfor[2]{4}\\
1 & \chi_1(\lambda_n) & \dots &\chi_{n-1} (\lambda_{n}) \\
\end{pmatrix}
\end{equation*}
and
\begin{equation*}
Q = \begin{pmatrix} r_1 & & & \\
& r_2 & & \\
& & \ddots& \\
& & & r_n
\end{pmatrix}, \quad
B = \begin{pmatrix} 1 & & & \\
& b_{n-1} & & \\
& & \ddots& \\
& & & \prod_{m=1}^{n-1} b_{n-m}
\end{pmatrix}.
\end{equation*}
Eq.~\eqref{eq:lem235} is the matrix product
\begin{equation*}
W^t Q W = B
\end{equation*}
written in terms of matrix elements. Since $\det(W) = \Delta(\boldsymbol{\lambda})$, taking the determinant of both sides of this equation gives
\[
\prod_{1\le j<k\le n}\left(\lambda_k-\lambda_j\right)^2
\prod_{j=1}^n r_j = \prod_{k=1}^{n-1} b_k^k.
\]
\end{proof}
A second proof of Lemma~\ref{lem:jac1} can be found in Appendix~\ref{sec_proof}.
\begin{lemma}
\label{lem:jac2}
We have
\begin{equation}
\label{jac_abrlb}
J = \det \left \{\frac{\partial\left(\mathbf{a},\mathbf{b}\right)}{\partial \left(\boldsymbol{\rv},\blambda\right)}\right \} =
\frac{\left(\prod_{k=1}^n r_k\right) \Delta(\blambda)^4}{\prod_{j=1}^{n-1}b_{j}^{2j-1}}.
\end{equation}
\end{lemma}
\begin{proof}
% $d\mathbf{a}=\bigwedge_{i=1}^nda_i$,\hspace{1mm} $d\mathbf{b}=\bigwedge_{i=1}^{n-1}db_i$,\hspace{1mm}
% $d\lambda=\bigwedge_{i=1}^nd\lambda_i$ and $d\mathbf{r}=\bigwedge_{i=1}^{n-1}dr_{i}.$
Write
\begin{equation*}
((I_n-zT)^{-1})_{11}=\sum_{j=1}^n\frac{r_j}{1-z\lambda_j},
\end{equation*}
and take the Taylor expansion of both sides of this equation. By equating the coefficients of the powers of $z$ up to $z^{2n-1}$ we arrive at
\begin{equation}
\label{eq:Taylor_exps}
\begin{split}
O(z^0):&\hspace{3mm}1=\sum_{j=1}^nr_j\\
O(z^1):&\hspace{3mm}a_n=\sum_{j=1}^nr_j\lambda_j\\
O(z^2):&\hspace{3mm}*\;+b_{n-1}=\sum_{j=1}^nr_j\lambda_j^2\\
O(z^3):&\hspace{3mm}*\;+a_{n-1}b_{n-1}=\sum_{j=1}^nr_j\lambda_j^3\\
O(z^4):&\hspace{3mm}*\; +b_{n-1}b_{n-2}=\sum_{j=1}^nr_j\lambda_j^4\\
O(z^5):&\hspace{3mm}*\;+a_{n-2}b_{n-1}b_{n-2}=\sum_{j=1}^nr_j\lambda_j^5\\
\vdots \hspace{.6cm} & \hspace{1.8cm}\vdots \hspace{2.5cm}\vdots\\
O(z^{2n-1}):&\hspace{3mm}*\; +a_1b_1\dots b_{n-1}=\sum_{j=1}^nr_j\lambda_j^{2n-1}.
\end{split}
\end{equation}
The symbol $*$ in Eqs.~\eqref{eq:Taylor_exps}
indicates polynomials in which only the
$a_j$'s and $b_k$'s contained in lower order terms appear.
Therefore, if we choose the ordering
\[
a_n, b_{n-1}, a_{n-1}, b_{n-2}, \dotsc, b_1,a_1,
\]
the matrix of the coefficients of the differentials of the left-hand sides of Eqs.~\eqref{eq:Taylor_exps} is triangular. Taking the wedge product line by
line gives
\begin{equation}
\label{eq:wplhs}
(-1)^{(n-1)(n-2)/2} \left(\prod_{j=1}^{n-1}b_j^{2j-1} \right)d\mathbf{a} \wedge d\mathbf{b}.
\end{equation}
Differentiating the right-hand sides of Eqs.~\eqref{eq:Taylor_exps} leads to
\begin{equation}
\label{eq:diffrhs}
d\left(\sum_{k=1}^nr_k\lambda_k^{j}\right)=\sum_{k=1}^{n-1}dr_k(\lambda_k^{j}-\lambda_n^{j})+j\sum_{k=1}^nr_k\lambda_k^{j-1}d\lambda_k,
\end{equation}
where we have used the relation
\begin{equation*}
%\label{eq:diff}
dr_n = - \sum_{j=1}^{n-1} dr_j,
\end{equation*}
which follows from the first equation in~\eqref{eq:Taylor_exps}. Taking
the wedge product of the right-hand side of Eq.~\eqref{eq:diffrhs}
for $j=1,\dotsc, 2n-1$ we arrive at
\begin{equation}
\label{eq:jac:2}
\left(\prod_{k=1}^nr_k \right)
\det\bigg(\left(\lambda_k^j-\lambda_n^j\right)_{
\begin{subarray}{l}
j=1,\dotsc,2n-1 \\ k=1,\dotsc,n-1
\end{subarray}}
\left(j\lambda_k^{j-1}\right)_{
\begin{subarray}{l}
j=1,\dotsc,2n-1 \\ k=1,\dotsc,n
\end{subarray}}
\bigg)d \rv \wedge d\blambda.
\end{equation}
The determinant in this equation can be calculated explicitly and is found as
Eq.~(1.175) in the book by Forrester~\cite{For10}.
We have
\begin{gather*}
(-1)^{(n-1)(n-2)/2}\det\bigg(\left(\lambda_k^j-\lambda_n^j\right)_{
\begin{subarray}{l}
j=1,\dotsc,2n-1 \\ k=1,\dotsc,n-1
\end{subarray}}
\left(j\lambda_k^{j-1}\right)_{
\begin{subarray}{l}
j=1,\dotsc,2n-1 \\ k=1,\dotsc,n
\end{subarray}}
\bigg) \nonumber \\
\label{eq:crazy_det}
= \prod_{1 \le j < k \le n}\left(\lambda_k - \lambda_j\right)^4 = \Delta(\blambda)^4.
\end{gather*}
Eqs.~\eqref{eq:wplhs} and~\eqref{eq:jac:2} give
\begin{equation*}
\label{jacobiand_eq}
\left(\prod_{j=1}^{n-1}b_j^{2j-1} \right)d\mathbf{a} \wedge d\mathbf{b}
= \left(\prod_{k=1}^nr_k \right) \Delta(\blambda)^4 d\rv \wedge d\blambda,
\end{equation*}
which leads to~\eqref{jac_abrlb}.
\end{proof}
Combining Lemmas~\ref{lem:jac1} and~\ref{lem:jac2} proves Theorem~\ref{thm:jac}.
\section{Proof of Theorem \ref{main_theorem}}
\label{sec:MainThm}
The \emph{j.p.d.f.} of the matrix elements of matrices in $\T(n)$ can be
read directly from~\eqref{mat_el_dis}:
\begin{equation}
\label{jpdf_mat_el}
p_{\beta}(\av,\bv) = \frac{1}{Z_\beta}\prod_{j=1}^{n-1}\abs{b_j}^{\frac{\beta j}{2} -1}
\exp\left[ -\frac12\left(\sum_{j=1}^n \abs{a_j}^2 + \sum_{j=1}^{n-1}
\abs{b_j}^2\right)\right],
\end{equation}
where $\av =(a_1,\dotsc,a_n)$, $\bv =(b_1,\dotsc,b_{n-1})$ and
\begin{equation*}
Z_\beta = \frac{(2\pi)^{3n/2 - 1}}{2^{n- \frac{n(n-1)\beta}{8} -1}}\prod_{k=1}^{n-1}
\Gamma\left(\frac{\beta k}{4}\right).
\end{equation*}
We are now in a position to prove Theorem~\ref{main_theorem}. By Lemma~\ref{lem:jac1} and Theorem~\ref{thm:jac} we obtain
\begin{equation}
\label{eqmes}
\prod_{j=1}^{n-1}\abs{b_j}^{\frac{\beta j}{2} -1} d\mu(\av,\bv) = \abs{\Delta(\blambda)}^\beta \prod_{j=1}^{n}\abs{r_j}^{\frac{\beta }{2}-1}
d\mu(\blambda,\rv).
\end{equation}
The Frobenius norm of a matrix $M=(m_{jk})_{j,k=1,\dotsc,n}$
is defined by
\[
\norm{M}_{\mathrm{F}} = \sqrt{\sum_{j,k=1}^n\abs{m_{jk}}^2}.
\]
Therefore, the exponent in~\eqref{jpdf_mat_el} can be written as
\begin{equation*}
%\label{Frobenius_exp}
\sum_{j=1}^n \abs{a_j}^2 + \sum_{j=1}^{n-1}\abs{b_j}^2 = \norm{T}_{\mathrm{F}}^2 - n + 1,
\end{equation*}
where $T \in \T(n)$. Recall that the Schur decomposition is defined by
\begin{equation}
\label{schur}
T = Q\left(\Lambda + N\right)Q^*,
\end{equation}
where $\Lambda \in \Lambda(n)$, $Q$ is a unitary matrix and $N$ strictly upper triangular. Combining~\eqref{schur} with Theorem~\ref{theorem:bijection},
we write
\begin{equation}
\label{exponent}
\norm{T}_{\mathrm{F}}^2 - n + 1 =
\sum_{j=1}^n \abs{\lambda_j}^2 + 2g(\blambda,\rv),
\end{equation}
where $g(\blambda,\rv)$ is a positive-definite function. The fact that
$g(\blambda,\rv)$ is a rational function of the $\lambda_j$'s and $r_j$'s follows from the
recurrence relations~\eqref{eq:sol_rec_r1} and~\eqref{eq:sol_rec_rn}.
Combining Eqs.~\eqref{jpdf_mat_el},~\eqref{eqmes} and~\eqref{exponent} gives
\begin{gather*}
p_\beta(\av,\bv)d\mu(\av,\bv) =P_\beta(\blambda,\rv)d\mu(\blambda,\rv)\nonumber\\
= \frac{1}{Z_\beta}\exp\left( -\frac12 \sum_{j=1}^n \abs{\lambda_j}^2 - g(\blambda,\rv)\right)\abs{\Delta(\blambda)}^\beta \prod_{j=1}^{n}\abs{r_j}^{\frac{\beta }{2}-1}
d\mu(\blambda,\rv).
\end{gather*}
Integrating over $\rv$ leads to Eqs~\eqref{eq:EigDens1} and~\eqref{factor}.
In order to show that
\begin{equation}
\label{order}
g(\blambda, \rv) = O(\abs{\lambda_k}^4), \quad \lambda_k \to \infty, \quad k=1,\dotsc,n,
\end{equation}
we observe that from Eq.~\eqref{eq:sol_rec_r1} and~\eqref{eq:sol_rec_rn}
we have
\begin{subequations}
\begin{align}
\label{abrl_1}
\rv_{j+1} &= O\left(\lambda_k^{j}\right), & \lv_{j+1} &=
O\left(\lambda_k^{- j }\right),\\
\label{abrl_2}
a_{n-j+1} &= O\left(\lambda_k\right), & b_{n-j} &= O\left(\lambda^2_k\right),
\end{align}
\end{subequations}
as $\lambda_k \to \infty$ and $j=1,\dots,n-1$. Substituting \eqref{abrl_2} into
Eq.~\eqref{exponent} we arrive at Eq.~\eqref{order}.
The equality
\[
g(\blambda e^{i\phi},\rv) = g(\blambda,\rv), \quad \phi \in [0,2\pi)
\]
follows from the observation that from Eqs.~\eqref{eq:anj} and~\eqref{eq:bnj} we have
\begin{equation*}
%\label{abv}
\abs{a_{n-j+1}} = \abs{ \rv_j^t\Lambda \lv_j}, \quad
\abs{b_{n-j}} = \abs{\rv_{j+1}^t \Lambda \lv_{j}}, \quad j=1,\dotsc,n.
\end{equation*}
Since only the absolute values of the $a_j$'s and $b_j$'s appear in the right-hand side of Eq.~\eqref{jpdf_mat_el},
the map $\blambda \mapsto \blambda e^{i\phi}$ leaves the \emph{j.p.d.f.} of the
eigenvalues~\eqref{eq:EigDens1} invariant.
This completes the proof of Theorem~\ref{main_theorem}.
\section*{Acknowledgements}
We are very grateful to Gernot Akemann, Guillaume Dubach, Tamara Grava
and Joseph Najnudel for many valuable discussions.
H.T. was supported by an EPSRC Doctoral Training Partnership
Scholarship no. EP/T517872/1.
\appendix
\section{The multiplicity of the eigenvalues of a tridiagonal matrix}
\label{multiplicity}
It is tempting to think of the matrices~\eqref{eq:tridiagmatrix} as the complex
analogue of Jacobi matrices. Certain properties that are linked to the
tridiagonal structure
carry through; however, this analogy must be taken with some care.
For example, the spectrum of
Jacobi matrices is simple. This is not necessarily the case for matrices in
$\T(n)$; a counterexample is
$\left (\begin{smallmatrix} 6 & 1\\ -16 & -2\end{smallmatrix} \right)$.
However, it is still
true that each eigenvalue has at most one eigenvector, even if its
algebraic multiplicity is greater than one.
\begin{theorem}
The geometric multiplicity of an eigenvalue of $T\in \T(n)$ is at most one.
\end{theorem}
\begin{proof}
The canonical vector
\[
\mathbf{e}_1 = \begin{pmatrix} 1 \\ \vdots \\ 0 \end{pmatrix}
\]
is cyclic for $T$; therefore,
\begin{equation}
\label{lin_ind}
\mathbf{e}_1, \, T\mathbf{e}_1, \dotsc,\, T^{n-1}\mathbf{e}_1
\end{equation}
are linearly independent. As a consequence, $T^n\mathbf{e}_1$ has a
unique representation as a linear combination of the vectors~\eqref{lin_ind}. Equivalently,
there exists a unique monic polynomial $\chi_n$ of degree $n$ such that $\chi_n(T)\mathbf{e}_1=0$. Furthermore, $\chi_n(T)$ is the polynomial in $T$ of
least degree that annihilates $\mathbf{e}_1$. Next, let
$\phi$ be the minimal polynomial of $T$. It satisfies the following properties:
\begin{enumerate}
\item $\phi(T)=0 \implies \phi(T)\mathbf{e}_1=0$;
\item by the Calyley-Hamilton theorem, the degree of $\phi$ cannot exceed $n$;
\item $\phi$ is a divisor of the characteristic polynomial of $T$.
\end{enumerate}
Thus, it follows that $\chi_n = \phi$ is both the minimal and
the characteristic polynomial of $T$. Thus, the spectrum of $T$ is either simple, or, if $\lambda$ is a multiple root of $\chi_n$, then it has only one block in the Jordan
normal form of $T$.
\end{proof}
\section{An alternative proof of Lemma \ref{lem:jac1}}
\label{sec_proof}
The results in this appendix are generalizations to complex tridiagonal
matrices of well known properties of Jacobi matrices (see, \emph{e.g.}~\cite{DE02,Par98}).
Write
\begin{equation*}
\widetilde{T}_{i,j}=\begin{pmatrix}
a_i & c_{i-1} & & \\
\widetilde{b}_{i-1} & a_{i-1} & \ddots & \\
& \ddots & \ddots & c_j \\
& & \widetilde{b}_j & a_j
\end{pmatrix},
\end{equation*}
where $1\le j \le i \le n $. We call $\widetilde{T}_{ij}$ a
\emph{principal submatrix} of the matrix $\widetilde{T}$ defined in
Eq.~\eqref{eq:gen_trid}. Furthermore, set
\[
\chi_{i,j}(\mu)=\det(\mu I-\widetilde{T}_{i,j})
\]
with the convention $\chi_{0,1}(\mu) = \chi_{n,n+1}(\mu) =1$ and $\chi=\chi_{n,1}$.
Let $\widetilde{T}=R\Lambda R^{-1}$ be the spectral decomposition of $\widetilde{T}$,
where $\Lambda=\diag(\lambda_1,\dots,\lambda_n)$. The columns of $R$ are the right eigenvectors of $\Tw$, which
we denote by $\mathbf{R}_j$, $j=1,\dotsc,n$. Similarly, we write $\mathbf{L}_j^t$ for the
left eigenvectors of $\Tw$, which are the rows $R^{-1}$. Recall that if $M \in \GL(n,\mathbb{C})$ is
invertible, then the \emph{adjugate} of $M$ is defined by
\[
\adj(M) = \det(M)M^{-1} \quad \text{and} \quad \adj(M)_{ij} = \frac{\partial \det(M)}{\partial M_{ji}}.
\]
\begin{lemma}
\label{lem:adjoint}
We have
\[
\adj\left(\lambda_jI-\widetilde{T}\right)
=\chi'(\lambda_j)\mathbf{R}_j\mathbf{L}_j^t, \quad j=1,\dotsc,n.
\]
\end{lemma}
\begin{proof}
Let $\mu\neq\lambda_j$, for all $j=1,\dotsc,n$, so that the
resolvent $(\mu I-\widetilde{T})^{-1}$ is well defined. We have
\begin{equation}
\label{eq:appendix:limit}
\begin{split}
\adj \left(\mu I-\widetilde{T}\right)&=
\det(\mu I-\widetilde{T})(\mu I-\widetilde{T})^{-1}\\
&=\chi(\mu)(\mu I-R\Lambda R^{-1})^{-1} \\
&=R\Delta(\mu)R^{-1},
\end{split}
\end{equation}
where $\Delta(\mu)=\diag(\delta_1(\mu),\dotsc,\delta_n(\mu))$ and
\begin{equation*}
\delta_k(\mu)=\frac{\chi(\mu)}{\mu-\lambda_k}=
\prod_{\substack{i=1\\i\neq k}}^n(\mu-\lambda_i).
\end{equation*}
Taking the limit as $\mu\rightarrow\lambda_j$ of both sides of
$\eqref{eq:appendix:limit}$ yields
\[
\text{adj}\left(\lambda_j I-\widetilde{T}\right)
=R\Delta(\lambda_j)R^{-1},
\]
where
\begin{align*}
\delta_k(\lambda_j)=\begin{cases}
0&\text{for $k\neq j$},\\
\prod_{i\neq j}(\lambda_j-\lambda_i)=\chi'(\lambda_j)&\text{for $k=j$.}
\end{cases}
\end{align*}
\end{proof}
\begin{proposition}
\label{theorem:spec}
Let $R_{rj}=\mathbf{e}_r^t\mathbf{R}_j$ and $L_{js}=\mathbf{L}_j^t\mathbf{e}_s$,
where $j,r,s=1,\dotsc,n$. Then, we have
\begin{equation}
\label{devrel}
\chi'(\lambda_j)R_{rj}L_{j s} =\begin{cases}
\chi_{n-s,1}(\lambda_j)c_{n-r}\chi_{n,n-r+2}(\lambda_j)&
\text{if $s-r=1$, }\\
\chi_{n-s,1}(\lambda_j)c_{n-s+1}\dotsm c_{n-r}\chi_{n,n-r+2}(\lambda_j)&
\text{if $s-r>1$, }\\
\chi_{n-r,1}(\lambda_j) \widetilde{b}_{n-s}\chi_{n,n-s+2}(\lambda_j)&\text{if $r-s=1$,}\\
\chi_{n-r,1}(\lambda_j)\widetilde{b}_{n-r+1}\dotsm \widetilde{b}_{n-s}\chi_{n,n-s+2}(\lambda_j)&\text{if $r-s>1$,}\\
\chi_{n-r,1}(\lambda_j)\chi_{n,n-r+2}(\lambda_j) & \text{if $r=s$.}
\end{cases}
\end{equation}
\end{proposition}
\begin{proof}
The right hand side of Eq.~\eqref{devrel} is obtained by computing the matrix elements of $\adj\left(\lambda_jI-\widetilde{T}\right)$ using the Laplace expansion.
The left hand side is given by Lemma~\ref{lem:adjoint}.
\end{proof}
Some particular cases of Proposition~\ref{theorem:spec} are
\begin{equation}
\label{rel_eq}
\begin{split}
\chi'(\lambda_j)R_{1 j}L_{jn} &= c_1 \dotsm c_{n-1},\\
\chi'(\lambda_j)R_{n j}L_{j 1} & = \widetilde{b}_1\dotsm \widetilde{b}_{n-1},\\
\chi'(\lambda_j)R_{1 j}L_{j1 } & = \chi_{n-1, 1}(\lambda_j),\\
\chi'(\lambda_j)R_{n j}L_{j n }& = \chi_{n,2}(\lambda_j).
\end{split}
\end{equation}
We are now in a position to prove Lemma \ref{lem:jac1}. Let $T\in \T(n)$ be
the matrix introduced in Eq.\eqref{eq:tridiagmatrix} and let $T=R\Lambda R^{-1}$ be
its spectral decomposition. Consider the sequence of monic
polynomials constructed from the recurrence relation
\begin{equation}
\label{rec_rel}
\chi_k(\xi) = (\xi - a_k)\chi_{k-1}(\xi) -b_{k-1}\chi_{k-2}(\xi), \quad k=1,\dotsc,n,
\end{equation}
with initial conditions $\chi_{0}=1$, $\chi_{-1}=0$ and $b_0$ undetermined.
Now, $\lambda_j^{(k)}$ is a zero of
$\chi_k(\xi)$ if
\[
\left( \chi_{k-1}\left(\lambda^{(k)}_j\right), \chi_{k-2}\left(\lambda^{(k)}_j\right),
\dotsc, \chi_0\left( \lambda_j^{(k)}\right) \right)
\]
is a left eigenvector of
\begin{equation*}
T_{k,1} =\begin{pmatrix}
a_k & 1 & & \\
b_{k-1} & a_{k-1} & \ddots & \\
& \ddots & \ddots & 1 \\
& & b_1 & a_1
\end{pmatrix},
\end{equation*}
which is the principal submatrix of $T$ obtained by keeping the last $k$ rows
and columns. It follows that\footnote{The characteristic
polynomials $\chi_k$ are defined differently from those introduced in Eq.~\eqref{cp_sm}. In Sec.~\ref{sec:Jacobian} they are defined as the
characteristic polynomials of the principal submatrix obtained by keeping the
\emph{first} $k$ rows and columns.
Definition~\eqref{cp_sma} is more convenient for the calculations in
this appendix.}
\be
\label{cp_sma}
\chi_k(\xi) = \chi_{k,1}(\xi)= \det\left(\xi I - T_{k,1}\right) =
\prod_{j=1}^k\left(\xi-\lambda_j^{(k)}\right).
\ee
Write
\begin{align}
\label{eq:twoterm}
\prod_{\substack{1\leq i \leq k\\ 1\leq j\leq k-1}}
\left(\lambda_i^{(k)}-\lambda_j^{(k-1)}\right)=\prod_{i=1}^k\chi_{k-1}\left(\lambda_i^{(k)}\right)=\prod_{i=1}^{k-1}\chi_{k}\left(\lambda_i^{(k-1)}\right).
\end{align}
From the recurrence relation~\eqref{rec_rel}, we have
\[
\prod_{i=1}^{k-1}\chi_{k}\left(\lambda_i^{(k-1)}\right)=(-b_{k-1})^{k-1}
\prod_{i=1}^{k-1}\chi_{k-2}\left(\lambda_i^{(k-1)}\right).
\]
Then, the relation~\eqref{eq:twoterm} leads to
\[
\prod_{i=1}^{k-1}\chi_{k}\left(\lambda_i^{(k-1)}\right)=(-b_{k-1})^{k-1}
\prod_{i=1}^{k-2}\chi_{k-1}\left(\lambda_i^{(k-2)}\right)
\]
Therefore, by iterating the last equality we arrive at
\begin{equation}
\label{prod_bj}
\begin{split}
\prod_{i=1}^n\chi_{n-1}\left(\lambda_i^{(n)}\right) &= \prod_{i=1}^{n-1}\chi_{n}\left(\lambda_i^{(n-1)}\right)\\
& =(-b_{n-1})^{n-1} \prod_{i=1}^{n-2}
\chi_{n-1}\left(\lambda_i^{(n-2)}\right)\\
&=(-b_{n-1})^{n-1}(-b_{n-2})^{n-2}\prod_{i=1}^{n-3}
\chi_{n-2}\left(\lambda_i^{(n-3)}\right)\\
&=\dotsb = (-1)^{\frac{n(n-1)}{2}}\prod_{j=1}^{n-1}b_j^j.
\end{split}
\end{equation}
Consider the derivative
\[
\chi'_{n}(\xi)=\sum_{i=1}^n
\prod_{\substack{j=1\\j\neq i}}^n(\xi-\lambda_j)
\]
and take the product
\be
\label{prod_der}
\prod_{i=1}^n\chi_{n}'(\lambda_i)=(-1)^\frac{n(n-1)}{2}\Delta(\blambda)^2.
\ee
If the matrix $R$ satisfies Lemma~\ref{parametrazion}, then
$R_{1j}=r_j$ and $L_{j1}=1$ for all $j=1,\dotsc,n$. Therefore,
the third relation in Eq.~\eqref{rel_eq} gives
\[
r_{j}=\frac{\chi_{n-1}(\lambda_j)}{\chi_{n}'(\lambda_j)},
\]
which implies
\be
\label{prod_rj}
\prod_{j=1}^nr_{j}=\prod_{j=1}^n
\frac{\chi_{n-1}(\lambda_j)}{\chi_{n}'(\lambda_j)}.
\ee
Since $\lambda_j^{(n)}=\lambda_j$, combining Eqs.~\eqref{prod_bj}, \eqref{prod_der} and \eqref{prod_rj} gives
\[
\Delta(\blambda)^2=\frac{\prod_{k=1}^{n-1}b_k^k}{\prod_{k=1}^nr_{k}}.
\]
\bibliographystyle{amsplain}
\bibliography{complexbeta}
%\printbibliography
\end{document}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
% \MRhref is called by the amsart/book/proc definition of \MR.
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
\begin{thebibliography}{10}
\bibitem{AMP22}
G.~Akemann, A.~Mielke, and P.~P\"{a}{\ss}ler, \emph{Spacing distribution in the
two-dimensional {C}oulomb gas: {S}urmise and symmetry classes of
non-hermitian random matrices at noninteger $\beta$}, Phys. Rev. E
\textbf{106} (2022), 014146.
\bibitem{AZ97}
A.~Altland and M.~R. Zirnbauer, \emph{Nonstandard symmetry classes in
mesoscopic normal-superconducting hybrid structures}, Phys. Rev. B
\textbf{55} (1997), 1142--1161.
\bibitem{AS21}
S.~Armstrong and S.~Serfaty, \emph{Local laws and rigidity for {C}oulomb gases
at any temperature}, Ann. Prob. \textbf{49} (2021), 46--121.
\bibitem{BLY12}
P.~Bourgade, L.~Erd\H{o}s, and H.~T. Yau, \emph{Bulk universality of general
$\beta$-ensembles with non-convex potential}, J. Math. Phys. \textbf{53}
(2012), 095221.
\bibitem{BLY14}
\bysame, \emph{Universality of general $\beta$-ensembles}, Duke Math. J.
\textbf{163} (2014), 1127--1190.
\bibitem{BFS07}
J.~Breuer, P.~J. Forrester, and U.~Smilansky, \emph{Random discrete
{S}chrödinger operators from random matrix theory}, J. Phys. A: Math. Theo.
\textbf{40} (2007), F161--F168.
\bibitem{BF22}
S.~S. Byun and P.~J. Forrester, \emph{Progress on the study of the {G}inibre
ensembles {I}: {G}in{UE}}, \texttt{arXiv:2211.16223v3} (2022).
\bibitem{Due01}
E.~Due{\~{n}}ez, \emph{Random matrix ensembles associated to compact symmetric
spaces}, Ph.D. thesis, Princeton University, 2001.
\bibitem{Due04}
\bysame, \emph{Random matrix ensembles associated to compact symmetric spaces},
Commun. Math. Phys. \textbf{244} (2004), 29--61.
\bibitem{DE02}
I.~Dumitriu and A.~Edelman, \emph{Matrix models for beta ensembles}, J. Math.
Phys. \textbf{43} (2002), 5830--5847.
\bibitem{DF10}
I.~Dumitriu and P.~J. Forrester, \emph{Tridiagonal realization of the
antisymmetric {G}aussian $\beta$-{E}nsemble}, J. Math. Phys. \textbf{51}
(2010), 093302.
\bibitem{Dys62a}
F.~J. Dyson, \emph{Statistical theory of the energy levels of complex systems.
{I}}, J. Math. Phys. \textbf{3} (1962), 140--156.
\bibitem{Dys62b}
\bysame, \emph{Statistical theory of the energy levels of complex systems.
{II}}, J. Math. Phys. \textbf{3} (1962), 157--165.
\bibitem{Dys62c}
\bysame, \emph{Statistical theory of the energy levels of complex systems.
{III}}, J. Math. Phys. \textbf{3} (1962), 166--175.
\bibitem{Dys62e}
\bysame, \emph{The threefold way. algebraic structure of symmetry groups and
ensembles in quantum mechanics}, J. Math. Phys. \textbf{3} (1962),
1199--1215.
\bibitem{ES07}
A.~Edelman and B.~D. Sutton, \emph{From random matrices to stochastic
operators}, J. Stat. Phys. \textbf{127} (2007), 1121--1165.
\bibitem{For10}
P.~J. Forrester, \emph{Log-gases and random matrices}, Princeton University
Press, 2010.
\bibitem{For21}
\bysame, \emph{Dyson’s disordered linear chain from a random matrix theory
viewpoint}, J. Math. Phys. \textbf{62} (2021), 103302.
\bibitem{FR05}
P.~J. Forrester and E.~M. Rains, \emph{Interpretations of some parameter
dependent generalizations of classical matrix ensembles}, Probab. Theory
Relat. Fields \textbf{131} (2005), 1--61.
\bibitem{FR06}
\bysame, \emph{Jacobians and rank 1 perturbations relating to unitary
{H}essenberg matrices}, IMRN \textbf{2006} (2006), 4836.
\bibitem{FS03}
Y.~V. Fyodorov and H.-J. Sommers, \emph{Random matrices close to {H}ermitian or
unitary: overview of methods and results}, J. Phys. A: Math. Gen. \textbf{36}
(2003), 3303--3347.
\bibitem{Gin65}
J.~Ginibre, \emph{Statistical ensembles of complex, quaternion, and real
matrices}, J. Math. Phys. \textbf{6} (1965), 440--449.
\bibitem{GGGM23}
T.~Grava, M.~Gisonni, G.~Gubbiotti, and G.~Mazzuca, \emph{Discrete integrable
systems and random {L}ax matrices}, J. Stat. Phys \textbf{190} (2023), 1--35.
\bibitem{GM23}
T.~Grava and G.~Mazzuca, \emph{Generalized {G}ibbs ensemble of the
{A}blowitz–{L}adik lattice, {C}ircular $\beta$-{E}nsemble and double
confluent {H}eun equation}, Commun. Math. Phys. \textbf{399} (2023),
1689--1729.
\bibitem{KN04}
R.~Killip and I.~Nenciu, \emph{Matrix models for circular ensembles}, IMRN
\textbf{2004} (2004), 2665--2701.
\bibitem{KRV16}
M.~Krishnapur, B.~Rider, and B.~Virág, \emph{Universality of the stochastic
airy operator}, Commun Pure Appl. Math. \textbf{69} (2016), 145--199.
\bibitem{MPS20}
S.~N. Majumdar, A.~Pal, and G.~Schehr, \emph{Extreme value statistics of
correlated random variables: A pedagogical review}, Phys. Rep. \textbf{840}
(2020), 1--32.
\bibitem{Meh04}
M.~L. Mehta, \emph{Random matrices}, third ed., Elsevier Inc., 2004.
\bibitem{NS15}
S.~R. Nodari and S.~Serfaty, \emph{Renormalized energy equidistribution and
local charge balance in {2D} {C}oulomb systems}, IMRN \textbf{2015} (2015),
3035--3093.
\bibitem{Par98}
B.~N. Parlett, \emph{The symmetric eigenvalue problem}, SIAM, 1998.
\bibitem{SS12}
E.~Sandier and S.~Serfaty, \emph{From the {G}inzburg-{L}andau model to vortex
lattice problems}, Commun. Math. Phys. \textbf{313} (2012), 635--743.
\bibitem{SS15}
\bysame, \emph{{2D} {C}oulomb gases and the renormalized energy}, Ann. Probab.
\textbf{43} (2015), 2026--2083.
\bibitem{Spo21}
H.~Spohn, \emph{Hydrodynamic equations for the {T}oda lattice},
\texttt{arXiv:2101.06528v1} (2021).
\bibitem{Zir96}
M.~R. Zirnbauer, \emph{Riemannian symmetric superspaces and their origin in
random‐matrix theory}, J. Math. Phys. \textbf{37} (1996), 4986--5018.
\end{thebibliography}