Variational Formulas for the Spectrum of Block Wishart Matrices
Extracted source text ยท arXiv 2606.27774
%%%%%%Arxiv submission format:
\documentclass[11pt]{article}
\usepackage{ifthen}
\newboolean{arxiv}
\setboolean{arxiv}{true}
%%%%%%Colt submission format:
%\documentclass[anon,12pt]{colt2022} % Anonymized submission
%\documentclass[final,12pt]{colt2022} % Include author names
%\newboolean{arxiv}
%\setboolean{arxiv}{false}
\usepackage{ifthen}
\usepackage{mathtools}
\mathtoolsset{showonlyrefs}
%%%%Import packages according to submission
\ifthenelse{\boolean{arxiv}}{
\usepackage{arxiv_def}
}
{
\usepackage{times}
\usepackage{colt_def}
}
\usepackage[toc,page]{appendix}
\usepackage{xr}
\usepackage{mathtools}
\usepackage{yfonts}
%\usepackage{tgheros}
\newcommand{\am}[1]{\textcolor{red}{[AM: #1]}}
\newcommand{\bns}[1]{\textcolor{blue}{[BS: #1]}}
\DeclareSymbolFont{Greekletters}{OT1}{iwona}{m}{n}
\DeclareSymbolFont{greekletters}{OML}{iwona}{m}{it}
\DeclareMathSymbol{\salpha}{\mathord}{greekletters}{"0B}
\DeclareMathSymbol{\sbeta}{\mathord}{greekletters}{"0C}
\DeclareMathSymbol{\sgamma}{\mathord}{greekletters}{"0D}
\DeclareMathSymbol{\sOmega}{\mathord}{Greekletters}{"0A}
\DeclareMathSymbol{\smu}{\mathord}{greekletters}{"16}
\DeclareMathSymbol{\svarepsilon}{\mathord}{greekletters}{"22}
\DeclareMathSymbol{\svarrho}{\mathord}{greekletters}{"25}
\DeclareMathSymbol{\svarphi}{\mathord}{greekletters}{"27}
%\definecolor{cc}{RGB}{1,11,111}
% \am, \bns are defined in arxiv_def.sty (duplicate \newcommand breaks with e.g. yfonts).
\let\De\cuD
\makeatletter
\newcommand{\vast}{\bBigg@{3}}
\newcommand{\Vast}{\bBigg@{4}}
\makeatother
\begin{document}
\title{Variational Formulas for the Spectrum of\\ Block Wishart Matrices}
\author{
Andrea Montanari\thanks{Department of Statistics and Department of Mathematics, Stanford University}
%
\and
%
Basil Saeed\thanks{Department of Electrical Engineering, Stanford University}
%
}
\maketitle
%\tableofcontents
\begin{abstract}
We analyze the asymptotics of a block-Wishart random matrix ensemble of the type $\bW_k = (\bX^* \otimes \bI_k)\bT (\bX\otimes\bI_k)$
for $\bX \in\C^{n\times p}$ with i.i.d. rows
satisfying a suitable concentration-of-measure property,
and $\bT := \Diag(\bT_i)_{i\in[n]}$ a block diagonal matrix with self-adjoint blocks $\bT_i\in \C^{k\times k}$,
under the proportional asymptotics $n/p\to\alpha$
with $k$ fixed. These matrices play a prominent
role in the analysis of $k$-index models in high-dimensional statistics.
By studying the matrix Stieltjes transform of this random matrix model
and its inverse ($K$-transform), we derive variational formulas for two functionals of the asymptotic spectral density of $\bW_k$: the left (equivalently right) edge of its support, and its logarithmic potential.
\end{abstract}
\section{Introduction}
%
Let $\bX_n \equiv\bX \in \C^{n\times p}$ be a sequence of random matrices indexed by $n \in \N$,
with i.i.d. rows $\bx_i \in\C^p, i\in[n]$, and for each $n$, consider matrices $\{\bT^{(n)}_1,\dots,\bT^{(n)}_n\}$
that are deterministic or independent of $\bX$,
with $\bT_i\in\sfM_{k}(\C)_{\textrm{sa}}$ for each $i\in[n]$.
(Here, we denote by $\sfM_{k}(\C)_{\textrm{sa}}$ the set of self-adjoint $k\times k$ matrix with elements in $\C$.)
In order to lighten the notation, we will drop the superscript from $\bT^{(n)}_i$,
We will be interested in the asymptotics of Wishart-type matrices defined by
\begin{equation}
\label{eq:W_k_def}
\bW_k := \frac1n(\bX^* \otimes \bI_{k})\Diag(\bT_{i})_{i\in[n]} (\bX\otimes \bI_k)
\equiv \frac1n\sum_{i=1}^n (\bI_k\otimes \bx_i) \bT_i(\bI_k\otimes \bx_i^*)
\in\C^{pk \times pk}
\end{equation}
in the proportional regime
\begin{equation}
\frac{n}{p} \to \alpha\in(0,\infty) \quad\quad n,p\to\infty,
\end{equation}
when
\begin{equation}
\frac1n \sum_{i=1}^n \delta_{\bT_i} \Rightarrow {\nu} \in\cuP_0(\sa{k}{\C}),
\end{equation}
here and throughout,
$\cuP(\cX)$ denotes the collection of probability measures on the Polish space $\cX$, while
$\cuP_0(\cX)$ denotes those with compact support.
We will assume throughout the $\bx_i$ to be centered (i.e. $\E[\bx_i]=\bzero$)
and isotropic (i.e. $\E[\bx_i\bx_i^{*}]=\bI_p$).
Matrices of this type appear in several applications, for example, in communications $\bW_k$ is the Gram matrix for MIMO channels describing the effect of the channel on the
signal~\cite{couillet2011random}. In high-dimensional statistics and machine learning,
the matrix $\bW_k$ is the Hessian of a general
empirical risk $\hat R_n(\bTheta) := n^{-1}\sum_{i\le n} \ell(\bTheta^{\sT}\bx_i)$
where $\bTheta \in\R^{p\times k}$ is a matrix of vector parameters \cite{asgari2025local,arous2025local}.
In this setting, parameters are typically estimated by minimizing $\hR_n(\bTheta)$, and it is therefore
of interest to study the properties of the Hessian at a point $\bTheta$:
%
\begin{align}
\nabla^2\hR(\bTheta) = \frac{1}{n}\sum_{i=1}^n
(\bI_k\otimes \bx_i) \nabla^2
\ell(\bTheta^{\sT}\bx_i)(\bI_k\otimes \bx_i^*)\, .
\end{align}
%
This is the same as Eq.~\eqref{eq:W_k_def}
where we identify $\bT_i= \nabla^2\ell(\bTheta^{\sT}\bx_i)$.
Two remarks are in order with regard to the application to
empirical risk minimization. \emph{First,}
unlike in our setting $\nabla^2\ell(\bTheta^{\sT}\bx_i)$
is not independent of $\bx_i$. However this dependence can often be accounted for by a low rank deformation
\cite{mondelli2019fundamental,lu2020phase}.
\emph{Second,} when the cost function $\ell$ is non-convex, the matrices $\bT_i$ are not necessarily positive semidefinite (PSD). Moving beyond the PSD case will be a
main motivation for the present note.
For $k=1$, the ensemble $\bW_1$ is
the classical Wishart-type matrix deformed multiplicatively by the diagonal entries $t_i$, i.e.,
\begin{equation}
\label{eq:wishart_classical}
\bW_1 = \frac1n \sum_{i=1}^n t_i\bx_i \bx_i^* \in \R^{p\times p}
\end{equation}
with $\bx_i \in\R^p$ the rows of $\bX$. Under certain tail decay conditions on $\bx_i$, the asymptotic spectral density in this setting was characterized by the seminal work~\cite{marvcenko1967distribution,silverstein1995empirical} (with several refinements since then, e.g.~\cite{hachem2007deterministic})
as the measure $\mu_\up{\alpha,\nu}$ whose Stieltjes transform $s_\up{\alpha,\nu}(z)$ uniquely solves
\begin{equation}
\label{eq:ST_FP_EQ_1}
K_{\up{\nu,\alpha}}(s) = z \quad\quad z \in \bbH_+, \quad\quad K_\up{\nu,\alpha}(s) := \int \frac{t}{1 + s t} \nu(\de t) - \frac1{\alpha s}
\end{equation}
where $\bbH_+ := \{z\in \C : \Im(z) > 0\}$. Following \cite{lehner1999computing},
we will refer to $K_{\up{\nu,\alpha}}(s)$ (and its generalizations below)
as the \emph{$K$-transform}\footnote{The term ``$K$-transform'' is typically reserved for the inverse of the \emph{Cauchy transform}, but we will use it here for the inverse of the Stieltjes transform.}.
For general $k>1$, the asymptotics of $\bW_k$ are given by the following proposition, which is a direct generalization of
Eq.~\eqref{eq:ST_FP_EQ_1} and follows
from earlier work (see Remark \ref{rmk:PropoProof}
below).
We note here that this asymptotic characterization holds
under various other assumptions on the matrix $\bX$.
For concreteness, we state it here under the assumption that the rows of $\bX$ enjoy
a Hanson--Wright type concentration inequality, but this is not the only setting.
%
\begin{proposition}[Standard asymptotic characterization]
\label{prop:asymptotic_FP}
Assume the rows $\bx_i\in\C^{p}$ are i.i.d., centered (i.e., $\E[\bx_i] =\bzero$), and isotropic
(i.e., $\E[\bx_i\bx_i^*] = \bI_p$). Assume further that there exists
$\sfK \in (0,\infty)$
such that, for each $i\in[n]$, all deterministic $\bA\in \C^{p\times p}$, and all $t>0$,
%
\begin{equation}
\label{eq:HW_condition}
\P\Big( \big|\bx_i^*\bA \bx_i - \Tr(\bA)\big| \ge t \Big)
\le
2 \exp\bigg\{-c \min\bigg(\frac{t^2}{\sfK^4 \|\bA\|_\Fnorm^2},\,
\frac{t}{\sfK^2 \|\bA\|_\op}\bigg)\bigg\}\, ,
\end{equation}
for some universal constant $c>0$, where $\|\bA\|_\Fnorm$ and $\|\bA\|_\op$ denote
the Frobenius and operator norms of $\bA$, respectively.
Fix integer $k\ge 1$, $\alpha \in (0,\infty)$ and
let $\nu\in\cuP_0(\sa{k}{\C})$.
Define
\begin{equation}
\label{eq:K_transform}
\bK_\up{\alpha,\nu}(\bS) := \int \bT(\bI + \bS\bT)^{-1} \nu(\de \bT) - \frac1\alpha\bS^{-1},\quad\quad \bK_\up{\alpha,\nu}:\C^{k\times k} \to\C^{k\times k}.
\end{equation}
For any $z \in\bbH_+$, there exists a unique
solution
$\bS_\up{\nu,\alpha}(z)$
to
\begin{equation}
\label{eq:ST_FP}
\bK_\up{\alpha,\nu}(\bS) = z\bI,\quad\quad \bS\in\bbH_+^k := \{\bB\in \C^{k\times k}: \Im(\bB) \succ\bzero\}.
\end{equation}
Furthermore,
\begin{equation}
s_\up{\nu,\alpha}(z) :=
\frac{\alpha}{k} \Tr(\bS_\up{\alpha,\nu}(z))\quad\quad\textrm{for}\quad z\in\bbH_+
\end{equation}
is the Stieltjes transform of a probability measure
$\mu_\up{\alpha,\nu}\in\cuP(\R)$ such that, almost surely
\begin{equation}
\hmu_{\bW_k} \Rightarrow \mu_\up{\alpha,\nu} \quad\quad \textrm{almost surely, as} \quad\quad n\to\infty,\;\frac{n}{p} \to\alpha,
\end{equation}
where $\hmu_{\bW_k}$ is the empirical spectral distribution of $\bW_k$ and $\Rightarrow$ denotes weak convergence of probability measures.
Finally, if $\dist(\{\bT_i:i\le n\},\supp(\nu))\to 0$ (with $\dist$ denoting
Hausdorff distance between sets in $(\reals^k,\|\,\cdot\,\|$), then,
almost surely,
%
\begin{align}
\dist\big(\supp(\hmu_{\bW_k}),\supp (\mu_\up{\alpha,\nu})\big) \to 0\, .
\end{align}
\end{proposition}
\begin{remark}\label{rmk:PropoProof}
The proof of Proposition \ref{prop:asymptotic_FP} follows from several standard results
in the literature. One approach is to use the linearization trick (see~\cite{far2006spectra,speicher2020lecture}), noting that matrix $\bW_k$
takes the form (for a suitable ordering of rows and columns)
%
\begin{align}
\bW_k = \left[\begin{matrix}
\bU_{11} &\bU_{12} &\cdots & \bU_{1k}\\
\bU_{21} & \bU_{22} & \cdots & \bU_{2k}\\
\vdots & \vdots & \cdots & \vdots\\
\bU_{k1} &\bU_{k2} &\cdots & \bU_{kk}\\
\end{matrix}\right]\, ,
\;\;\;\;\;\; \bU_{ij}:=\frac{1}{n}\bX^{*}\hat{\bT}_{ij}\bX\, ,
\end{align}
%
where $\hat{\bT}_{ij}:=\Diag((\bT_{\ell})_{ij}:\ell\le n)$.
Hence, for any integer $m$, there is a non-commutative polynomial $Q_m$ such that
$\Tr(\bW^{m}_k) = \Tr(Q_m(\{\bU_{ij}\}))$, and the asymptotics of the latter
for Gaussian matrices
follows from \cite[Theorem 5.4.10]{anderson2010introduction} (for weak convergence
of the empirical spectral distribution)
and \cite[Theorem 5.5.1]{anderson2010introduction}
(for convergence of the support).
For non-Gaussian matrices, the same asymptotics holds by a Lindeberg argument
\cite{chatterjee2006generalization,montanari2022universality}.
An alternative approach is to use the Matrix Dyson equation of
\cite{alt2017local}.
For a complete proof of the exact statement of Proposition~\ref{prop:asymptotic_FP} in the Gaussian setting (in particular, existence and uniquenness of
the solution of Eq.~\eqref{eq:ST_FP}, see the Appendix of \cite{asgari2025local}.
It is not too hard to see that their proof is directly generalizable to the current setting.
\end{remark}
\section{Computing functionals of $\mu_\up{\alpha,\nu}$}
While the above proposition characterizes the asymptotic spectral measure $\mu_\up{\alpha,\nu}$ of $\bW_k$, it does so in an implicit way and studying functionals of this measure from this equation can be challenging.
We will focus here on two particular functionals of the measure
$\mu_\up{\alpha,\nu}$: the \emph{edges of its support},
\begin{equation}
\zeta_{-,\up{\alpha,\nu}} := \inf\supp(\mu_{\up{\alpha,\nu}}),\quad\quad
\zeta_{+,\up{\alpha,\nu}} := \sup\supp(\mu_{\up{\alpha,\nu}}),
\end{equation}
and its \emph{logarithmic potential}
\begin{equation}
\Phi_\up{\alpha,\nu}(u) := \int \log |u -\zeta| \mu_\up{\alpha,\nu}(\de \zeta),\quad\quad u \in \R \setminus [\zeta_-,\zeta_+].
\end{equation}
Both of these quantities are useful in applications.
The logarithmic potential coincides with the limit of
$\log\det(\bW_k)$, which
in wireless communications,
determines \emph{channel capacity} (i.e., the mutual information) of a MIMO channel modeled by $\bW_k$.
In high-dimensional statistics, $\zeta_{-}$ can be used
to compute the lower edge of the asymptotic
spectrum of the Hessian $\nabla^2\hat R_n(\bTheta)$.
Hence, it provides information about the stability of landscape at a particular point $\bTheta$. The log potential also plays an important role in the Kac-Rice approach
to the analysis of the empirical risk minimization problem
\cite{maillard2020landscape,asgari2025local}.
In the rest of this section, we will present the main ideas for the functional $\zeta_-$
to be definite, and state the characterization of the log-potential at the end of the section.
In what follows, we will often omit the subscript $\up{\alpha,\nu}$ from all quantities whenever there is no risk of confusion; for example, we write $\mu$ for $\mu_\up{\alpha,\nu}$.
In the classical setting $k=1$, \cite{silverstein1995analysis}
already gives a useful characterization of $\zeta_-$.
Namely, letting $t_{\min} := \inf \supp(\nu) \wedge 0$, and recalling the
definitions of~\eqref{eq:ST_FP_EQ_1}, we have
%
\begin{equation}
\zeta_{-}= \sup_{s\in \cM_1} K(s)\quad\quad\textrm{with}\quad\quad
\cM_1 := \left\{s \in (0, -t_{\min}^{-1}),\; \int \left(\frac{s t }{1 + st} \right)^{2} \nu(\de t)< \frac1\alpha\right\}.\label{eq:OnedCase}
\end{equation}
%
(Here, it is understood that
$-1/t_{\min}=\infty$ whenever $t_{\min}=0$.)
Namely, if $x < \inf \supp(\mu),$ then $s(x) \in (0,-t_{\min}^{-1})$
where $t_{\min} := \inf \supp(\nu) \wedge 0.$
This formula can be read off of Eq.~\eqref{eq:ST_FP_EQ_1} as follows:
By definition, $s(z)$ has an analytic continuation on a the maximal interval $(-\infty,\zeta_{-})$ (i.e. no analytic continuation exists on $(-\infty,\zeta_{-}+\eps)$
for any $\eps>0$). Further $s$ maps $(-\infty,x)$ to $(0,s(x))$,
$s(x)>0$ for every $x<\zeta_-$.
Since $K(s)$ is its analytic inverse on $(0,s(x))$, then
we must have $\zeta_{-} = K(s_0)$ when
$s_0 >0$ is the smallest value at which $K$ fails to be invertible.
This coincides with the smallest $s$ at which $K'(s)\le 0$, which
yields Eq.~\eqref{eq:OnedCase}, by noting that
%
\begin{align}
s^2K'(s) = \frac{1}{\alpha}-\int \left(\frac{s t }{1 + st} \right)^{2} \nu(\de t)\, .
\end{align}
%
In other words,
\begin{align}
\label{eq:pre_var_zmin_1}
s_\up{\alpha,\nu}( (-\infty, \zeta_{-})) = \cM_1 ,
%
\end{align}
which is equivalent to Eq.~\eqref{eq:OnedCase}.
A naive attempt to generalize this idea to
$k>1$ yields
%
\begin{equation}
\label{eq:pre_var_zmin_k}
\zeta_{-} = \sup \big\{z : \exists \bS\in \cM_k \;\;\textrm{s.t.}\;\; \bK(\bS)=z\bI_k \big\}
\end{equation}
with $\bK$ the $K$-transform defined in Proposition~\ref{prop:asymptotic_FP},
and the set $\cM_k$ defined as the (bijective) image of $(-\infty,\zeta_{\min})$
under $\bS_\up{\alpha,\nu}(x)$.
However, this description of the set $\cM_k$ is not explicit, and the resulting
characterization is circular.
In order to obtain a more useful generalization of Eq.~\eqref{eq:OnedCase}, we will consider
the matrix valued $K$-transform of Eq.~\eqref{eq:K_transform}
$\bK:\sfM_k(\C)\to\sfM_k(\C)$, and in the equation $\bK(\bS) = \bZ$ for general $\bZ\in\C^k$.
As we will see below, the inverse of $\bS\mapsto\bK_{\up{\alpha,\nu}}(\bS)$,
call it $\hat\bS_{\up{\alpha,\nu}}(\bZ)$ has the interpretation of providing the limit
%
\begin{align}
\hat\bS_{\up{\alpha,\nu}}(\bZ)= \lim_{n,p\to\infty}\frac{1}{n}\Big(\bI_k\otimes \Tr\Big)\Big((\bW_k-\bZ\otimes \bI_p)^{-1}\Big)\,
\end{align}
for more general $\bZ\in\C^k$ than multiples of the identity.
The details of this approach will be clarified in Section~\ref{sec:proofs_main}.
We can now state our main results. We denote by
$\cuD \bK(\bS)[\cdot]:\sfM_{k}(\C)\to\sfM_{k}(\C)$ the
directional derivative of $\bK$ at $\bS$,
which can be computed explicitly as
%
\begin{equation}
\De_\bS\bK(\bS)[\bDelta] =
\frac1\alpha \bS^{-1}\bDelta\bS^{-1} -
\int \bT(\bI+\bS\bT)^{-1}\bDelta (\bI+\bT\bS)^{-1}\bT\, \nu(\de\bT) \quad\quad
\textrm{for}\quad
\bDelta\in\sa{k}{\C}.
\end{equation}
%
Notice that $\cuD \bK(\bS)[\cdot]$ is a self-adjoint operator (with respect to the standard
scalar product on $\sfM_{k}(\C)$). Hence it
makes sense to ask whether $\De_\bS\bK(\bS)[\cdot ]$ is positive definite
(in formulas $\De_\bS\bK(\bS)[\cdot ]\succ 0$) or positive semidefinite
($\De_\bS\bK(\bS)[\cdot ]\succeq 0$),
which amounts respectively to $\<\bDelta,\De_\bS\bK(\bS)[\bDelta ]\> > 0$ for all $\bDelta\neq \bzero$,
or $\<\bDelta,\De_\bS\bK(\bS)[\bDelta ]\> \ge 0$ for all $\bDelta\neq \bzero$.
%
\begin{theorem}[Variational formula for left edge]
\label{thm:edge}
Fix $k \ge 1$ integer, $\alpha \in (0,\infty)$, and $\nu\in\cuP_0(\sa{k}{\C})$.
Let $\mu_{\alpha,\nu}$
be as in Proposition~\ref{prop:asymptotic_FP}.
Then we have
\begin{equation}
\inf \supp(\mu_\up{\alpha,\nu})= \sup_{\bS \in \cP_{-}} \lambda_{\min}\left(\bK_{(\alpha,\nu)}(\bS)\right).
\end{equation}
where $\cP_-$ is given by
\begin{equation}
\label{eq:P_def}
\cP_- \equiv \cP_-(\alpha,\nu) := \{\bS\succ\bzero : \bS^{-1}+ \bT\succ\bzero \;\; \forall\,\bT\in\supp(\nu),\;\;
\cuD_\bS \bK(\bS)[\,\cdot\,]\succ 0\;
\}.
\end{equation}
Furthermore, the set $\cP_-$ is equivalent to
\begin{equation}
\label{eq:P_def_with_B}
\cP_- = \{\bS\succ\bzero : \bS^{-1}+ \bT\succ\bzero \;\; \forall\,\bT\in\supp(\nu),\;\;\textrm{there exists }\bB\succ\bzero \textrm{ s.t.} \;
\cuD_\bS \bK(\bS)[\bB]\succ \bzero\;
\}.
\end{equation}
%the condition that $\cuD_\bS\bK(\bS)[\,\cdot\,] \succ 0$ is equivalent to the existence of \emph{some} $\bB\equiv\bB_\bS$ s.t. $\cuD_\bS\bK(\bS)[\bB] \succ\bzero$.
\end{theorem}
A related characterization for the left and right edges $\zeta_{\pm}$ appeared recently in~\cite{parmaksiz2025computing} applicable when $\bT_i$ are positive semi-definite,
albeit stated differently. The derivation of \cite{parmaksiz2025computing}
is based on a technique introduced in~\cite{lehner1999computing}
to compute the operator norm of free operators.
The approach of \cite{lehner1999computing} realizes the edge
as the operator norm of a free operator which can be explicitly decomposed as the sum of
operators on the Fock Space. By carefully bounding the norm of these basis operators, they obtain a tight upper bound on the operator norm\footnote{After
finishing this manuscript, Ramon van Handel communicated to us that the
technique of \cite{parmaksiz2025computing} can also be generalized to indefinite
case, yielding a characterization analogous (but identical) to
Theorem \ref{thm:edge}}. We point out that the setting of
\cite{parmaksiz2025computing} is more general in other ways, and in particular accommodating for non-isotropic\footnote{We believe our approach can be generalized to the non-isotropic case but we refrain from doing so at the moment.} $\bx_i$'s.
The next theorem gives the formula for logarithmic potential $\Phi(u)$.
\begin{theorem}[Variational formula for the logarithmic potential]
\label{thm:log}
Fix $k \ge 1$ integer,
$\alpha \in (0,\infty)$, and $\nu\in\cuP_0(\sa{k}{\C})$.
For $u<\zeta_{-,\up{\alpha,\nu}}$,
set
\begin{equation}
g_\up{\alpha,\nu}(\bS;u) := -\alpha u \Tr(\bS) + \alpha \int \log \det(\bI + \bT \bS) \nu(\de\bT) - \log|\det(\bS)| - k(\log \alpha +1 ).
\end{equation}
then we have (for $\cP_-$ defined as in Eq~\eqref{eq:P_def})
\begin{equation}
\Phi(u) := \int \log |\zeta - u| \, \mu_{\up{\alpha,\nu}}(\de \zeta)
= \frac1k\inf_{\bS\in\cP_{-}} g_\up{\alpha,\nu}(\bS;u)\, .
\end{equation}
\end{theorem}
We remark that since $\bT$ is not assumed to be positive semi-definite, the above two theorems directly lead to a formula for $\zeta_+$, and for $\Phi(u)$ for $u> \zeta_+$, respectively. We state the exact form in Remarks~\ref{remark:zeta_+} and~\ref{remark:log_+} below.
\section{Proofs of Theorem \ref{thm:edge} and \ref{thm:log}}
\label{sec:proofs_main}
In this section, we prove Theorems~\ref{thm:edge} and~\ref{thm:log}
by leveraging properties of the matrix Stieltjes and $K$-transforms.
We defer the proof of the key technical fact, Theorem
\ref{thm:K_transform_real}, to the next section.
We begin with some background on free probability. Most of these preliminaries are taken from~\cite{nica2006lectures} and~\cite{speicher2020lecture}.
\paragraph{Free probability preliminaries.}
Let $(\cA,\tau)$ be a $C^\star$-probability space (i.e., it is both tracial and faithful), with elements $m,t_{i,j} \in \cA$, $i,j\in[k]$
satisfying the following conditions. First,
$m$ is a \emph{free-Poisson} element with rate $\alpha$, i.e.,
for any integer $d>0$,
\begin{equation}
\tau(x^d) = \int \zeta^d\, \mu_{\MP,\alpha}(\de \zeta)
\end{equation}
with $\mu_{\MP,\alpha}$ denoting the Marchenko-Pastur law with aspect ratio $\alpha$.
Second, $t_{i,j}$ for $i,j\in[k]$ satisfy, for any pairs $Q = \{(a_1,b_1),\dots,(a_N,b_N)\}$, $a_i,b_i \in [k],$
\begin{equation}
\tau\left(\prod_{(i,j) \in Q} t_{i,j}\right) = \int \prod_{(i,j)\in Q} (\bT)_{ij} \; \nu(\de \bT).
\end{equation}
Define
\begin{align}
w_k := ( m^{1/2} t_{i,j} m^{1/2})_{i,j \in[k]} \in \sa{k}{\cA}.\label{eq:wk-def}
\end{align}
Here, we use $\sa{k}{\cA}$ to denote the set $k$ by $k$ self-adjoint matrices with elements in the $C^*$-algebra $\cA$.
Since $\nu\in\cuP(\sa{k}{\C}),$ it is easy to see that $w_k$ will indeed be self-adjoint. The operator $w_k$ can be seen as the \emph{free-analogue} or \emph{free deterministic equivalent} of $\bW_k$ from Eq.~\eqref{eq:W_k_def} in the following sense: for any non-commutative polynomial $f$ of $N>0$ elements, and pairs
$Q = \{(a_1,b_1),\dots,(a_N,b_N)\}$, $a_i,b_i \in [k],$
\begin{equation}
\label{eq:free_determinism}
\lim_{\substack{n\to\infty \\ n/p \to\alpha}}\E\left[\frac1n
\Tr f\left( \frac1n \bX^*\hat\bT_{a_1b_1}\bX,\dots, \frac1n \bX^*\hat\bT_{a_Nb_N}\bX \right)
\right]
=
\frac1\alpha\tau\left(
f\left( m^{1/2}t_{a_1b_1}m^{1/2},\dots, m^{1/2}t_{a_Nb_N}m^{1/2}\right)
\right)
\end{equation}
where $\hat\bT_{a,b}$ is the $n$ by $n$ matrix whose diagonal elements are given by $(\bT_{1})_{a,b}, \dots, (\bT_{n})_{a,b}$.
Given an element $a\in\cA$, its spectrum can be defined as the support of the measure $\mu$ on
$\R$ whose moments
coincide with the the moments of $a$. We denote this by $\spec_\cA(a).$
The Gelfand-Naimark-Segal construction guarantees the existence of a Hilbert space $\cH$, a $*$-representation $\Pi$ of $\cA$ on $\cH$, and
and a `tracial' element $\psi\in\cH$ such that for
any $a\in\cA,$
\begin{equation}
\tau(a) = \langle \psi, \Pi(a) \psi\rangle_{\cH}.
\end{equation}
The $*$-representation $\Pi$ satisfies $\Pi(a^*) = \Pi(a)^*$,
where the $A^*$ denotes the adjoint of operator $A$.
In particular, if $a\in\cA$ is self-adjoint, then the spectral theory of bounded operators on $\cH$ defines the spectrum of $\Pi(a)$ which we denote by $\spec_\cH(\Pi(a)).$
Faithfulness of $\tau$ then guarantees the two notions of the spectrum of $a$ coincide.
\begin{proposition}[Proposition 3.15 of~\cite{nica2006lectures}]
\label{prop:spectra_match}
For $(\cA,\tau)$ as above with $\tau$ faithful,
if $a\in\cA$ is self-adjoint, then
\begin{equation}
\spec_\cA(a)= \spec_{\cH}(\Pi(a)).
\end{equation}
\end{proposition}
Having this in mind, we will throughout use $\spec(a)$ to denote both $\spec_\cH(\Pi(a))$ and $\spec_\cA(a)$.
Define the product Hilbert space $\cH^k := \cH\times \cdots \times \cH$ ($k$ times Cartesian product).
Equivalently $\cH^k=\C^k\otimes \cH$.
Then note that
$(\bI_k\otimes \Pi)w_k$ can be seen as a element of $\cB(\cH^k).$
For a bounded self-adjoint operator $B$ on $\cH$ or $\cH^k$, we write
$B\succ 0$ if $\spec(B)\subseteq (0,\infty)$.
For a self-adjoint $a\in \cA$ (or, $a_k\in\sa{k}{\cA}$),
we will use $a\succ\bzero$ ($a_k \succ\bzero$) instead of the more cumbersome
$\Pi(a)\succ\bzero$ ($(\bI_k\otimes\Pi)a\succ\bzero$, respectively).
Finally, define the matrix Stieltjes transform by
%\am{I do not think $\bbH_+^k$ is defined}
\begin{equation}
\label{eq:free_ST_def}
\hat\bS_\up{\alpha,\nu}(\bZ) := \frac1\alpha (\bI_k \otimes \tau)\Big(
w_k - \bZ\otimes \id_\cA \Big)^{-1},\quad\quad
\bZ\in\bbH_+^k,
\end{equation}
where $\bbH_+^k$ was defined in Eq.~\eqref{eq:ST_FP}.
Note that under the conditions of Proposition~\ref{prop:asymptotic_FP},
this quantity is the almost sure limit of the random matrix
\begin{equation}
\label{eq:empirical_ST}
\hat\bS_n(\bZ) :=
(\bI_k \otimes \frac1n \Tr) \left( \bW_k - (\bZ \otimes \bI_p)\right)^{-1}
\end{equation}
as $n,p\to\infty$ with $n/p\to\alpha.$
\paragraph{The image of $\cU_-$.}
The following theorem gives the inverse of $\hat\bS$, to the `left' and `right' of the spectrum of $w_k$.
This is the key technical result to establish the characterizations of $\zeta_{\pm}$
and $\Phi(u)$
from Theorems~\ref{thm:edge} and~\ref{thm:log}.
\begin{theorem}
\label{thm:K_transform_real}
Fix integer $k \ge 1,$
let $\alpha \in (0,\infty)$, and $\nu\in\cuP_0(\sa{k}{\C})$. Define
$\cU_-\equiv \cU_{-}(\alpha,\nu)\subseteq \sa{k}{\C} $ via
\begin{equation}
\cU_{-}(\alpha,\nu) := \{\bU \in \sa{k}{\C} : w_k - \bU\otimes \id_\cA \succ 0\}.
\end{equation}
Then $\hat\bS_\up{\alpha,\nu}$ has an analytic continuation to $\cU_-$, and it maps $\cU_-$ diffeomorphically to $\cP_-$ given in Eq.~\eqref{eq:P_def} and~\eqref{eq:P_def_with_B}.
% \begin{equation}
%\cP_- \equiv \cP_-(\alpha,\nu) := \{\bS\succ\bzero : \bS^{-1}+ \bT\succ\bzero \;\; \forall\,\bT\in\supp(\nu),\;\; \exists \bB \succ\bzero \;\textrm{s.t.}\;
%\cuD_\bS \bK(\bS)[\bB]\succ\bzero\;
%\}\, .
% \end{equation}
Further, $\bK_\up{\alpha,\nu}:\cP_-\to\cU_-$ is its analytic inverse in this domain.
%\am{In the formula above: why $\exists \bB$? The condition does not involve $\bB$.}
\end{theorem}
Note that since $\nu$ is not assumed to be supported on positive semi-definite matrices, a similar statement holds regarding the set
\begin{equation}
\cU_+\equiv \cU_{+}(\alpha,\nu) := \{\bU \in \sa{k}{\C} : w_k - \bU\otimes \id_\cA \prec 0\},
\end{equation}
where its image under $\hat\bS_\up{\alpha,\nu}$ is replaced by
\begin{equation}
\label{eq:def_P_+}
\cP_+ := \{\bS\prec\bzero : \bS^{-1}+ \bT\prec\bzero \;\; \forall\,\bT\in\supp(\nu),\;\; \exists \bB \succ\bzero \;\textrm{s.t.}\;
\cuD_\bS \bK(\bS)[\bB]\succ\bzero\;
\}.
\end{equation}
The set $\cU_-$ can be thought of as a generalization of the notion of `left of the support', in particular, for $k=1$, $\cU_- = \{u\in\R : u < \zeta_{-}\}.$
This analogy will be more apparent in the proof of Theorem~\ref{thm:edge}, which we present next.
\paragraph{Deriving the variational formulas.}
\begin{proof}[Proof of Theorem~\ref{thm:edge}]
Fix $\alpha,\nu$ as in the theorem statement.
Clearly, for any $\bU\in\cU_-,$ we have
$(\bI_k\otimes \Pi)w_k \succ \bU \otimes \Pi(\id_\cA).$
So for any $\xi = (\xi_1 ,\xi_2 , \dots ,\xi_k)$ with $\|\xi\|_{\cH^{k}} = 1,$
\begin{equation}
\langle\xi,
(\bI_k\otimes \Pi) w_k\;
\xi\rangle_{\cH^{k}} \ge \langle \xi, (\bU\otimes\id_\cH) \xi \rangle_{\cH^{ k}} = \sum_{i,j\in[k]} {\inner{\xi_i,\xi_j}}_{\cH} U_{ij} = \Tr( \bC \bU ),
\end{equation}
where we defined the Gram matrix $\bC := (\inner{\xi_i,\xi_j}_{\cH})_{i,j}$. Taking the spectral decomposition of the self-adjoint complex matrix $\bU = \sum_{i=1}^k d_iu_i u_i^*$, we then see that
\begin{equation}
\Tr(\bC\bU) = \sum_{i=1}^k d_i u_i^* \bC u_i \ge \min_{i} d_i \cdot \Tr(\bC) = \lambda_{\min}(\bU),
\end{equation}
where we used that $\Tr(\bC) = \sum_i \|\xi_i\|^2 = 1.$ Combining the above two displays, taking the infimum over $\xi$ of unit norm, then the supremum over $\bU\in\cU_-$ yields
\begin{equation}
\zeta_- \ge \sup_{\bU\in\cU_-}\lambda_{\min}(\bU).
\end{equation}
On the other hand, if $x < \zeta_{-}$, then $x\bI_k \in\cU_-.$ So choosing a sequence $x_m < \zeta_-$ so that $|x_m - \zeta_-| \to 0$ as $m\to\infty$ we have
\begin{equation}
\zeta_- = \lim_{m\to\infty} \lambda_{\min}(x_m \bI_k) \le \sup_{\bU\in\cU_-} \lambda_{\min}(\bU).
\end{equation}
Since by Theorem~\ref{thm:K_transform_real} we have
$\cU_- = \{
\bU : \bU=\bK(\bS),\; \bS\in\cP_-\},$
changing variables
and combining the above two displays gives
\begin{align}
\zeta_{-} &= \sup_{\bU \in\cU_-}\lambda_{\min}(\bU)
= \sup_{\bS\in\cP_-} \lambda_{\min}\left(\bK(\bS)\right).
\end{align}
\end{proof}
\begin{remark}[The right edge]
\label{remark:zeta_+}
As remarked previously, a similar formula for the right edge $\zeta_+$ can be directly obtained from Theorem~\ref{thm:edge} by replacing $\bT$ with $-\bT$.
This gives
\begin{equation}
\zeta_+ := \sup \supp(\mu_\up{\alpha,\nu})= \inf_{\bS \in \cP_{+}} \lambda_{\max}\left(\bK_\nu(\bS)\right),
\end{equation}
where $\cP_+$ was defined in~\eqref{eq:def_P_+}.
\end{remark}
We now present the proof of Theorem~\ref{thm:log}.
\begin{proof}[Proof of Theorem~\ref{thm:log}]
Fix $u<\zeta_{-}$ as in the theorem statement.
We suppress the dependence on $\alpha$ and $\nu$ for brevity and write, for example, $\hat\bS$ instead of $\hat\bS_{\up{\alpha,\nu}}$,
$g(\cdot)$ instead of $g_\up{\alpha,\nu}(\cdot)$,
and so on.
By Theorem~\ref{thm:K_transform_real}, we can re-write the variational formula of interest as
\begin{equation}
\inf_{\bS\in\cP_-} g(\bS;u) = \inf_{\bU\in\cU} g(\hat\bS(\bU);u).
\end{equation}
Fix $\bU_0 \in\cU_-$ arbitrary. We will show that $g(\hat\bS(u\bI); u) \le g(\hat\bS(\bU_0);u)$.
Since $u <\zeta_-$, we have $u\bI\in \cU_-$, and therefore
$g(\hat\bS(u\bI); u) =\inf_{\bU\in \cU_-} g(\hat\bS(\bU);u)$. This will then conclude the proof, since
\begin{equation}
g(\hat\bS(u\bI);u ) = k\int \log | \zeta - u| \mu(\de \zeta);
\end{equation}
this identity can be seen by differentiating both sides and taking $u\to -\infty$ to match the boundary
conditions.
Define for $\eps\in(0,1],$
\begin{equation}
\bU_\eps := \bU_0 - \eps\bDelta \quad\quad
\bDelta := \bU_0 - u\bI_k
\end{equation}
and note that we have
\begin{align}
w_k - \bU_\eps\otimes \id_\cA&=
w_k - ((1-\eps)\bU_0 + \eps u \bI_k) \otimes \id_\cA
=(1-\eps) \left( w_k - \bU_0\otimes \id_\cA\right)
+ \eps \big(w_k - u\bI_k\otimes \id_\cA\big)
\succ\bzero
\end{align}
for all $\eps\in(0,1].$
Now introduce the notation
$m_\eps := w_k - (\bU_\eps\otimes\id_\cA)$
and compute the derivative.
\begin{align}
\frac{\partial}{\partial \eps}
g(\hat\bS(\bU_\eps);u)
&\stackrel{(a)}{=}-\cuD_\bU g(\hat\bS(\bU_\eps);u)[\bDelta]\\
&\stackrel{(b)}{=}
-\frac1\alpha (\Tr\otimes \tau)
\bigg(
\Big( \big[-\alpha u\bI + \alpha \bK(\hat\bS(\bU_\eps))\big] \otimes \id_\cA\Big)
m_\eps^{-1} (\bDelta \otimes \id_\cA) m_\eps^{-1}
\bigg)\\
% &\stackrel{(c)}{=}
% -(\Tr\otimes \tau)
% \big(
% \big( (\bU_\eps - x\bI) \otimes \id\big)
% m(\bU)^{-1} \big( \bDelta \otimes \id\big) m(\bU)^{-1}
% \big)\\
&\stackrel{(c)}{=}
-(1-\eps)(\Tr\otimes \tau)
\big(
\big( \bDelta \otimes \id\big)
m_\eps^{-1} \big( \bDelta \otimes \id\big) m_\eps^{-1}
\big)\\
&\stackrel{(d)}{=}
-(1-\eps)(\Tr\otimes \tau)
\Big(
\big(
m_\eps^{-1/2} \big( \bDelta\otimes \id\big) m_\eps^{-1/2} \big)^2
\Big)\\
&\le 0
\end{align}
for all $\eps\in(0,1),$
where $(a)$ is by the chain rule, $(b)$ is by direct computation, $(c)$ is by Theorem~\ref{thm:K_transform_real}, whence
\begin{equation}
\alpha\left(-u\bI + \bK(\hat\bS(\bU_\eps)) \right) = \alpha (\bU_\eps - u\bI) =\alpha (1-\eps) \bDelta,
\end{equation}
and in $(d)$ we used the fact that $\bU_\eps\in\cU$ to conclude that $m_\eps\succeq 0$.
(In fact, the inequality is strict whenever $\bDelta \neq 0$, i.e., $\bU_0\neq u\bI_k$).
Hence,
\begin{equation}
g(\hat\bS(u \bI_k);u) =
g(\hat\bS(\bU_1);u) \le
g(\hat\bS(\bU_0);u)
\end{equation}
as desired.
\end{proof}
\begin{remark}[$\Phi$ to the right of the support]
\label{remark:log_+}
Once again, we can directly derive a formula for the logarithmic potential for $u > \zeta_+$ by symmetry,
\begin{equation}
\int \log |\zeta - u| \mu_{\up{\alpha,\nu}}(\de \zeta)
= \frac1k \inf_{\bS\in\cP_{+}} g_\up{\alpha,\nu}(\bS;u).
\end{equation}
\end{remark}
\paragraph{Some useful consequences.}
As corollaries of the formulas above, we note a few direct consequences that are a priori not obvious.
First we have the following characterization of measures $\mu_\up{\alpha,\nu}$ that are positively supported.
Such a question is for instance relevant in non-convex minimization problems such as minimizing the empirical risk $\hat R_n(\bTheta)$ mentioned in the introduction, since local minima correspond to points where the Hessian is positive semi-definite.
\begin{corollary}
Fix $\alpha\in(0,\infty)$, integer $k\ge1$, $\nu\in\cuP_0(\sa{k}{\C})$.
Then $\supp(\mu_\up{\alpha,\nu})\subseteq(0,\infty)$ if and only
if there exist $\bS,\bB\succ\bzero$ satisfying all of the following
\begin{enumerate}
\item
$
\alpha^{-1} \bS^{-1}\bB\bS^{-1} -
\int \bT(\bI+\bS\bT)^{-1}\bB (\bI+\bT\bS)^{-1}\bT\, \nu(\de\bT)\succ\bzero$
\item
$\int \bT(\bI+\bS\bT)^{-1} \nu(\de\bT) - \alpha^{-1} \bS^{-1}\succ\bzero.$
\item $\bS^{-1} +\bT \succ\bzero$ for all $\bT\in\supp(\nu).$
\end{enumerate}
\end{corollary}
Another useful consequence is continuity of $\zeta_{-,\up{\alpha,\nu}}$ at $\nu$ whenever
we restrict to sequences $\nu_j$ whose supports are (approximately and eventually) contained in that of $\nu$.
\begin{corollary}
Fix $\alpha\in(0,\infty)$, integer $k\ge1$, $\nu\in\cuP_0(\sa{k}{\C})$.
%Assume $\supp(\mu_\up{\alpha,\nu})\subseteq(0,\infty).$
Let $\{\nu_j\}_{j\ge 1}$ be a sequence $\nu_j\in\cuP_0(\sa{k}{\C})$ such that
$\nu_j \stackrel{w}{\Rightarrow} \nu$ as $j\to\infty$, and,
for all $\delta>0$ there exists $j_0(\delta)$ such that
\begin{align}
j\ge j_0(\delta)\;\; \Rightarrow\;\;
\supp(\nu_j) \subseteq\bigcup_{\bT\in\supp(\nu)} \big\{\bX\in \sa{k}{\C}
:\;\; \bX\succeq \bT-\delta\bI\big\}\, .
\end{align}
Then
\begin{equation}
\lim_{j\to \infty} \zeta_{-,\up{\nu_j,\alpha}} = \zeta_{-,\up{\nu,\alpha}}.
\end{equation}
\end{corollary}
We note that in general, when $\nu_j\stackrel{w}{\Rightarrow} \nu$ without
assumptions about the support $\supp(\nu_j)$, the conclusion of the above corollary
can fail to hold. As an example, one can consider $\nu_j=a_j\delta_{-M}+\nu_j^+$,
where $\nu_j^+$ is supported on $(0,\infty)$ and $a_j\to 0$, but $M>0$ is sufficiently large.
%\am{Does that mean that both assumptions $\supp(\nu_j) \subseteq \supp(\nu)$ and $\supp(\mu_\up{\alpha,\nu})\subseteq(0,\infty)$ are necessary? Also, where is the proof?}
%I also added the proof below.
%I was only using this because I thought I needed the existence of a finite $\bS$ that achieves the supremum.
%}
\begin{proof}
We will suppress the dependence on $\alpha$ in the notation.
Fix $\eps >0$ and use Theorem~\ref{thm:edge} to find $\bS_\eps,\bB_\eps \succ\bzero$ and $\delta_\eps>0$
satisfying
%\am{Why $\cuD_\bS \bK_\up{\nu}(\bS_\eps)[\cdot] \succ \bzero$ implies $\cuD_\bS \bK_\up{\nu}(\bS_\eps)[\cdot] \succ \bzero$ ? Where is it proved?}
\begin{equation}
\lambda_{\min}(\bK_\up{\nu}(\bS_\eps)) \ge \zeta_{-,\up{\nu}} - \eps,
\qquad
\cuD_\bS \bK_\up{\nu}(\bS_\eps)[\bB_\eps] \succ \delta_\eps \bI_k,
\qquad
\bS_\eps^{-1} + \bT \succ \delta_\eps \bI_k \quad \forall \bT \in \supp(\nu),
\end{equation}
Note that we used the assumption that $\nu$ is compactly supported to ensure that $\bS_\eps^{-1} + \bT \succ\bzero$ for all $\bT\in\supp(\nu)$
implies a uniform bound on $\bS_\eps^{-1} + \bT$.
The condition on the supports of $\nu_j$ implies that there exists $J_{\eps}<\infty$ such that
$\bS_\eps^{-1} + \bT \succ \delta_\eps \bI_k/2$
for all $\bT \in \supp(\nu_j)$ provided $j\ge J_{\eps}$. The convergence $\nu_j \Rightarrow \nu$ then implies that
$\cuD_\bS \bK_\up{\nu_j}(\bS_\eps)[\bB_\eps]$ converges to $\cuD_\bS \bK_\up{\nu}(\bS_\eps)[\bB_\eps]$ in the operator norm.
Hence, by eventually increasing $J_{\eps}$, we have that, for all $j \ge J_\eps$,
\begin{equation}
\cuD_\bS \bK_\up{\nu_j}(\bS_\eps)[\bB_\eps] \succ \frac{\delta_\eps}{2} \bI_k,
\end{equation}
and therefore we conclude $\bS_\eps \in \cP_{-}(\nu_j)$ for all $j \ge J_\eps$.
Applying Theorem~\ref{thm:edge} to $\nu_j$ for $j \ge J_\eps$ now gives
\begin{equation}
\zeta_{-,\up{\nu_j}} = \sup_{\bS \in \cP_{-}(\nu_j)} \lambda_{\min}(\bK_\up{\nu_j}(\bS))
\ge \lambda_{\min}(\bK_\up{\nu_j}(\bS_\eps))
%\ge \zeta_{-,\up{\nu}} - \eps.
\end{equation}
%
Taking the limit inferior, and using the fact that $\bK_\up{\nu_j}(\bS_\eps)$ converges in operator norm to
$\bK_\up{\nu}(\bS_\eps)$, we get
\begin{equation}
\lim\inf_{j\to\infty}\zeta_{-,\up{\nu_j}} \ge \lambda_{\min}(\bK_\up{\nu}(\bS_\eps))
\ge \zeta_{-,\up{\nu}} - \eps\, .
\end{equation}
Letting $\eps\to 0$, yields the lower bound
$\lim\inf_{j\to\infty}\zeta_{-,\up{\nu_j}} \ge \zeta_{-,\up{\nu}}$.
To prove the matching upper bound, we note that the weak convergence of $\nu_j\Rightarrow\nu$
implies $\lim_{j\to\infty}\hat\bS_{\up{\nu_j}}(z\bI)=\hat\bS_{\up{\nu}}(z\bI)$ for all $\Im(z)>0$,
by a standard stability argument of the fixed point equation \eqref{eq:ST_FP}
(see e.g. \cite{asgari2025local}). As a consequence
$\mu_{(\nu_j)}\Rightarrow\mu_{(\nu)}$, which in turn implies $\lim\sup_{j\to\infty}\zeta_{-,\up{\nu_j}} \le \zeta_{-,\up{\nu}}$.
\end{proof}
%\begin{proof}[Proof of Theorem~\ref{thm:log}]
%
%Fix $x < \zeta_{\min}$ as in the theorem statement.
%First, compute the derivative of $K_x(\bS)$ with respect to $\bS$ to see that
%\begin{equation}
% \cuD_\bS K_x(\bS) = \alpha \left( \bG(\bS) - x\bI_k\right).
%\end{equation}
%So $K_x$ is strictly convex over the set
%$\bS \in \cP_-$, and hence it is sufficient to show that the minimizer $\hat\bS$ is achieved in $\cP_-$. Indeed, this would then imply that $\bG(\hat\bS) =x\bI_k,$ so that $\hat\bS = \bS_\star(x\bI)$ by Lemma~\ref{lemma:main}. Invoking Lemma~\ref{lemma:K_at_S_star} would then give the result.
%
%To do that, use Lemma~\ref{lemma:main} to rewrite
%\begin{equation}
% \inf_{\bS\in\cR\cap\cD} K_{x}(\bS) = \inf_{\bX\in\cX} K_x(\bS_\star(\bX)),
%\end{equation}
%and note by the open-mapping theorem, the above task is equivalent to showing that the infimum over $\cX$ is achieved inside $\cX$.
%
%If $\bX_m \in\cX$ is any diverging sequence,
%
%To do that,
%fix any point $\hat\bX$
%with $\bH - \bX\otimes\id$ singular. (We can exclude diverging sequences of $\bX$ since $\bS_\star(\bX) \to 0$ and $K$ blows up.)
%Define for $\eps\in(0,1),$
%\begin{equation}
% \bX_\eps := \hat\bX - \eps\bDelta \quad\quad
% \bDelta := \hat\bX - x\bI
%\end{equation}
%Noting that
%we must have $\bH - (\hat\bX\otimes \id) \succeq\bzero,$ we note that
%\begin{align}
% \bH - \bX_\eps\otimes \id
% &=
% \bH - ((1-\eps)\hat\bX + \eps x \bI) \otimes \id \\
% &=(1-\eps) \left(\bH - \hat\bX\otimes \id\right)
% + \eps (\bH - x\bI\otimes \id)\\
% &\succ\bzero
%\end{align}
%for all $\eps>0$
%since $x < \zeta_{\min}$,
%so that $\bX_\eps \in\cX$.
%So with the notation
%\begin{equation}
% \bM(\bX) := \bH - (\bX\otimes \id),
%\end{equation}
%we can compute the derivative of
%\begin{equation}
%f(\eps) := K_x(\bS_\star(\bX_\eps))
%\end{equation}
%to be
%\begin{align}
% f'(\eps) &=-D_\bX K_x(\bS_\star(\bX_\eps))[\bDelta]\\
% &=
% -\frac1\alpha (\Tr\otimes \tau)
% \bigg(
% \Big( \big(-\alpha x\bI - \alpha \bG(\bS_\star(\bX_\eps))\big) \otimes \id\Big)
% \bM(\bX_\eps)^{-1} (\bDelta \otimes \id) \bM(\bX_\eps)^{-1}
% \bigg)\\
% &=
% -(\Tr\otimes \tau)
% \big(
% \big( (\bX_\eps - x\bI) \otimes \id\big)
% \bM(\bX)^{-1} \big( \bDelta \otimes \id\big) \bM(\bX)^{-1}
% \big)\\
% &=
% -(1-\eps)(\Tr\otimes \tau)
% \big(
% \big( \bDelta \otimes \id\big)
% \bM(\bX)^{-1} \big( \bDelta \otimes \id\big) \bM(\bX)^{-1}
% \big)\\
% &=
% -(1-\eps)(\Tr\otimes \tau)
% \Big(
% \big(
% \bM(\bX_\eps)^{-1/2} \big( \bDelta\otimes \id\big) \bM(\bX_\eps)^{-1/2} \big)^2
% \Big)
%\end{align}
%for all $\eps\in(0,1).$
%
%Now note that the
%square of the operator is positive semi-definite. Furthermore, since
%$\bH - (\hat\bX \otimes\id)$, we must have
%that $\bv^\star \hat\bX \bv \ge \zeta_{\min}>x$ for some $\bv.$
%So $\bDelta - x\bI\neq 0$. That is,
%\begin{equation}
% f'(\eps) < 0
%\end{equation}
%for all $\eps \in (0,1).$
%So $\hat\bX$ cannot be a minimizer.
%
%
%\end{proof}
\section{Proof of Theorem~\ref{thm:K_transform_real}}
The remainder of the manuscript is dedicated to the proof of the
Theorem~\ref{thm:K_transform_real}. It is perhaps useful to recall that
$\bbH_+^k := \{\bB\in \C^{k\times k}: \Im(\bB) \succ\bzero\}$.
The starting point is to consider a more general form of the equation Eq.~\eqref{eq:ST_FP} for $\bZ\in\bbH_+^k.$
The next lemma establishes that the solution of this matrix equation is given by the Stieltjes transform
$\hat\bS_\up{\alpha,\nu}$ of Eq.~\eqref{eq:free_ST_def}.
%\am{Why below $\hat\bS_{\alpha,\nu}(\bZ)$ instead of
%$\hat\bS_\up{\alpha,\nu}(\bZ)$?}
%
\begin{lemma}
\label{lemma:inv_on_upper_plane}
Fix integer $k \ge 1,$
let $\alpha \in (0,\infty)$, and $\nu\in\cuP_0(\sa{k}{\C})$.
For any $\bZ \in\bbH_+^k$,
$\hat\bS_\up{\alpha,\nu}(\bZ)$
of Eq.~\eqref{eq:free_ST_def}
is the unique solution of
\begin{equation}\label{eq:Matrix-Fixed-Point}
\bK_{\up{\alpha,\nu}}(\bS) = \bZ,\quad\quad \bS\in\bbH_+^k.
\end{equation}
\end{lemma}
%
%\bns{Andrea's revision: START.}
%Results of this type are fairly standard.
%but since
%we are not aware of a reference for this precise statement,
%we will give a proof in Appendix~\ref{app:proofs_FP}.
%
%In order to prove that $\hat\bS_{\alpha,\nu}(\bZ)$
%of Eq.~\eqref{eq:free_ST_def} is a solution of
%Eq.~\eqref{eq:Matrix-Fixed-Point}, we will use a standard
%`leave-one-out' argument. \am{Writing this because I suspect this can be
%done directly `at $n,d=\infty$' using free probability.
%We can talk about this.}
%\bns{I ended up only writing the uniqueness part
% and refernecing our paper for the existence part.
% I checked the computaton in our paper goes through when replacing $z\bI$ by $\bZ$ so I thought it wasn'tworth writing it up.
% I am, nonetheless curious to see what you had in mind and happy to write it.
%}
%
%In order to prove uniqueness, we use
%the Earle-Hamilton fixed-point theorem to
%show that iterates of this equation converge to a unique point
%from any initial condition~\cite{helton2007operator}.
%\bns{Andrea's revision: END.}
%\\
Such result is fairly standard.
In order to prove that $\hat\bS_\up{\alpha,\nu}(\bZ)$
of Eq.~\eqref{eq:free_ST_def} is a solution of
Eq.~\eqref{eq:Matrix-Fixed-Point}, a standard approach is to use a
`leave-one-out' argument at finite $n,p$ to establish that the
empirical Stieltjes transform $\hat\bS_n(\bZ)$ of Eq.~\eqref{eq:empirical_ST}
satisfies the fixed-point equation in an approximate sense.
We refer to Appendix A of~\cite{asgari2025local} for such an argument in this setting.
Uniqueness is typically established by showing that the iterates of the
fixed-point equation converge to a unique point from any initial condition.
See for example~\cite{helton2007operator} for a proof in the matrix-valued semi-circular setting.
We will give a full proof of uniqueness in Appendix~\ref{app:proofs_FP} via a slightly different approach.
As the statement of Theorem~\ref{thm:K_transform_real} suggests,
we will be interested in the Stieltjes transform evaluated at
self-adjoint matrices $\bU\in \cU_-$, i.e., on the boundary of $\bbH_+^k.$
The next lemma carries out the extension to $\cU_-.$
Before stating it, let us first introduce the notation
\begin{equation}
\cD_0 := \{\bS\succ\bzero : \bS^{-1} + \bT \succ\bzero \quad \forall \bT \in\supp(\nu)\}
\end{equation}
to denote PSD matrices at which $\bK$ will be evaluated.
\begin{lemma}[Regularity of $\hat\bS_\up{\alpha,\nu}$ on $\cU_-$]
\label{lemma:regularity_free_ST}
Fix integer $k \ge 1,$
let $\alpha \in (0,\infty)$, and $\nu\in\cuP_0(\sa{k}{\C})$.
The Stieltjes transform $\hat\bS_\up{\alpha,\nu}(\bU)$ defined by
Eq.~\eqref{eq:free_ST_def}
is analytic on an open set in $\sfM_{k}(\C)$ containing $\cU_-$.
In particular, for all $\bU\in\cU_-,$
\begin{equation}
\label{eq:lim_from_imaginary_matrices}
\lim_{y\downarrow 0} \hat \bS_\up{\alpha,\nu}(\bU + i y\bI) = \hat \bS_\up{\alpha,\nu}(\bU).
\end{equation}
Consequently, if $\hat \bS_\up{\alpha,\nu}(\bU)\in\cD_0$ we have
\begin{equation}
\label{eq:K_S_U_is_U_on_D0}
\bK_\up{\alpha,\nu}(\hat\bS_\up{\alpha,\nu}(\bU)) = \bU.
\end{equation}
Furthermore,
the derivative of $\hat\bS_\up{\alpha,\nu}$ at $\bU\in\cU_-$ as an operator on
$\sa{k}{\C}$ is given by
\begin{align}
\label{eq:DhatS} \De \hat\bS_\up{\alpha,\nu}(\bU)[\bDelta] &= \frac1\alpha(\bI_k \otimes \tau)\left(r_k(\bU)(\bDelta \otimes \id_\cA)r_k(\bU)\right) , \quad\bDelta\in\sa{k}{\C}\, ,\\
r_k(\bU)&:= \big(w_k -(\bU \otimes \id_\cA)\big)^{-1}\, .\nonumber
\end{align}
Further, $\De \hat\bS_\up{\alpha,\nu}(\bU)[\cdot]$ is non-singular and
\begin{equation}
\langle \bDelta, \cuD\hat \bS_\up{\alpha,\nu}(\bU)[\bDelta]\rangle_\fro >0\quad\quad\forall
\quad \bDelta\neq\bzero.
\end{equation}
\end{lemma}
\begin{proof}
We suppress the dependence on $\alpha,\nu$ in the notation.
Fix $\bU_0 \in \cU_{-}$, and define
$m_0:= w_k-\bU_0\otimes \id_\cA\in \sa{k}{\cA}.$
Set
$c_0:=\|m_0^{-1}\|^{-1}>0.$
Then for any $\bZ\in \sfM_k(\C)$ with $\|\bZ-\bU_0\|<c_0/2$, we can write
$$
w_k-(\bZ\otimes \id_\cA)
=
m_0^{1/2}\Big(\bI - m_0^{-1/2}\big((\bZ-\bU_0)\otimes \id_\cA\big)m_0^{-1/2}\Big)m_0^{1/2}.
$$
Moreover,
$$
\big\|m_0^{-1/2}\big((\bZ-\bU_0)\otimes \id_\cA\big)m_0^{-1/2}\big\|
\le \|m_0^{-1}\|\,\|\bZ-\bU_0\|<\frac{1}{2}\, .
$$
Hence the middle factor is invertible since the associated Neumann series converges, and therefore
$$
\big(w_k-(\bZ\otimes \id_\cA)\big)^{-1}
=
m_0^{-1/2}\sum_{p=0}^\infty
\Big(m_0^{-1/2}\big((\bZ-\bU_0)\otimes \id_\cA\big)m_0^{-1/2}\Big)^p m_0^{-1/2}.
$$
The series converges absolutely in operator norm, uniformly on the
ball $B(\bU_0,c_0/2)\subset \sfM_k(\C)$. It follows that
$\bZ\mapsto \big(w_k-(\bZ\otimes \id_\cA)\big)^{-1}$
is analytic on this ball. Since $(I_k\otimes \tau)$ is continuous and
linear, we conclude that $\hat\bS(\bZ)$
extends analytically to $B(\bU_0,c_0/2)$. As $\bU_0\in \cU_-$
was arbitrary, $\hat \bS$ admits an analytic continuation to $\cU_-$.
Equation~\eqref{eq:lim_from_imaginary_matrices} follows because any analytic function is also continuous.
Finally, Eq.~\eqref{eq:K_S_U_is_U_on_D0} follows because
$\bK(\hat\bS(\bU+i y \bI)) = \bU + i y \bI$ for all $y>0$ by
Lemma~\ref{lemma:inv_on_upper_plane}, $\hat\bS(\bU+i y \bI)\to \hat\bS(\bU)$ as $y\downarrow 0$
by the previous point, and $\bS\mapsto \bK(\bS)$ is continuous at $\hat\bS(\bU)\in\cD_0$
because $\inf_{\bT\in\supp(\nu)}\lambda_{\min}(\bS^{-1}+ \bT) > 0$, $\lambda_{\min}(\bS)>0$
for any $\bS\in\cD_0$.
Finally, Eq.~\eqref{eq:DhatS} for the derivative follows by simple calculus.
To show that it is non-singular, use positivity of $r_k(\bU)$ to write,
for any $\bDelta\in \sa{k}{\C}$,
\begin{align}
\langle \bDelta, \cuD\hat \bS(\bU)[\bDelta]\rangle_\fro
&=
\frac1\alpha (\Tr\otimes \tau)\Big((\bDelta\otimes \id_\cA)\,r_k(\bU)\,(\bDelta\otimes \id_\cA)\,r_k(\bU)\Big)\\
&=
\frac1\alpha (\Tr\otimes \tau)\Big(
\big(r_k(\bU)^{1/2}(\bDelta\otimes \id_\cA)r_k(\bU)^{1/2}\big)^2
\Big)\ge 0.
\end{align}
If this quantity is zero, then by faithfulness of $(\Tr\otimes \tau)$ and positivity of $r_k$,
we must have $\bDelta = \bzero$.
\end{proof}
\paragraph{The image of $\cU_-$ is contained in $\cD_0$.}
We will now proceed to show that the image of $\cU_-$ under $\hat\bS_\up{\alpha,\nu}$ is contained in the domain of $\bK,$ specifically, in $\cD_0.$
Observe that for $\bU = u\bI$ with $u\to-\infty,$ we have $\bU \in\cU_-$ eventually when $\nu$
is compactly supported. Then since $\hat \bS_\up{\alpha,\nu}(\bU) \to \bzero,$ we will have
$\hat\bS_\up{\alpha,\nu}(\bU) \in \cD_0$ eventually. The next lemma shows that this property extends to all of $\cU_-$, at least when $\nu$ is a discrete measure.
\begin{lemma}[$\hat\bS_\up{\alpha,\nu}(\cU_-) \subseteq \cD_0$: discrete $\nu$]
\label{lemma:image_Sstar_D_discrete}
Fix $\alpha \in(0,\infty)$.
Assume $\nu = \sum_{i=1}^m \delta_{\bT_i} p_i$ for $\sum_{i=1}^m p_i = 1$ with
$p_i > 0$.
Then for any $\bU \in\cU_-$, we have
\begin{equation}
\hat\bS_\up{\alpha,\nu}(\bU) \in \cD_0.
\end{equation}
\end{lemma}
\begin{proof}
Once again, we suppress the dependence on $\alpha,\nu$ in the notation.
Since $\nu$ is compactly supported, it follows from the definition
\eqref{eq:wk-def} that $\|w_k\|\le C$, for some constant $C>0$.
Hence, there exists $u_0 <0$ sufficiently negative such that
$u_0\bI_k \in\cU_-$ and $\hat\bS(u_0 \bI) \in \cD_0$.
Let $\cC_{0}:=\{\bU \prec u_0\bI\}$.
Since the inverse is monotone with respect to Loewner ordering for positive operators, we have
$\hat\bS(\bU) \preceq\hat\bS(u_0\bI)$ for all $\bU\in\cC_0,$ so that $\hat\bS(\bU) \in \cD_0$ as well.
Now fix some $\bU_0 \in \cC_0$ and arbitrary $\bU_1\in\cU_-$ and suppose for contradiction that $\bS_1:= \hat\bS(\bU_1) \not \in \cD_0.$ That is, suppose there exists some $\bT_1\in\supp(\nu)$ such that
\begin{equation}
\lambda_{\min}\left( \bS_1^{-1}+ \bT_1 \right) \le 0.
\end{equation}
For $\delta\in[0,1]$, let $\bU_\delta := \delta \bU_1 + (1-\delta)\bU_0.$
Clearly, $\bU_\delta \in \cU_-$ for all $\delta \in [0,1],$ and
\begin{equation}
\lim_{\delta \to 0} \hat\bS(\bU_\delta) \in\cD_0,\quad\quad
\lim_{\delta \to 1} \hat\bS(\bU_\delta) = \hat\bS(\bU_1) \not \in \cD_0
\end{equation}
since $\delta \mapsto \hat\bS(\bU_\delta)$ is continuous, so there must exist some $\delta_0\in(0,1)$ such that
\begin{equation}
\hat\bS(\bU_\delta) \in \cD_0 \quad\textrm{for}\;\; \delta < \delta_0\quad\quad\textrm{and}\quad\quad \lambda_{\min}( \hat\bS(\bU_{\delta_0})^{-1} + \bT_i
) = 0 \quad\textrm{for some }i\in[m],
\end{equation}
and since $\nu(\{\bT_i\}) =: p_i > 0$ for all $i\in[m]$ by assumption, we have, using the
definition of $\bK$ in Eq.~\eqref{eq:K_transform},
\begin{equation}
\lim_{\delta\uparrow \delta_0} \|\bK(\hat\bS (\bU_\delta))\|_\op = \infty.
\end{equation}
But, this contradicts
Lemma~\ref{lemma:regularity_free_ST}
which guarantees that for $\delta<\delta_0$, we have
\begin{equation}
\bK(\hat\bS(\bU_\delta)) = \bU_\delta.
\end{equation}
\end{proof}
%Noting that $\bG(\bS)$ is analytic on $\cD$,
%we must then have for all $\bX \prec \bX_{\delta_0},$
%$$\bG(\hat\bS(\bX)) =
%\lim_{y\downarrow 0}\bG(\bS_\star(\bX + i y \bI))
%=\lim_{y\downarrow 0} \left\{ \bX- i y\bI\right\}
%=\bX.
%$$
%But
%\begin{equation}
% \bX_{\delta_0} = \lim_{\bX \to \bX_{\delta_0}} \bG(\bS_\star(\bX)) \to\infty
%\end{equation}
%, reaching a contradiction.
%
%
%By Lemma~\ref{lemma:regularity_free_ST}, $\hat \bS(\bX)$ is analytic on $\cX$. Furthermore,
To extend the applicability of the previous argument to general compactly supported $\nu$,
we state the following fact that allows us to approximate a measure $\nu$ by a discrete one.
%
\begin{lemma}[Discrete approximation lemma]
\label{lemma:discrete_approx}
Fix $\alpha \in(0,\infty)$.
For any compactly supported probability measure $\nu\in\cuP(\sa{k}{\C})$,
and all $\eps > 0$,
there exists a measure $\nu^\up{\eps}$ and a set $\cC^\up{\eps} \subseteq \C^{k\times k}$
such that the following hold:
\begin{enumerate}
\item
\label{item:1_approx}
$\cC^\up{\eps} \subseteq \supp(\nu)$, has finite cardinality, and
%\am{Inverted $\sup$ and $\inf$ below}
$$
\label{eq:closed_cover}
\sup_{\bY \in \supp(\nu)} \inf_{\bT \in \cC^\up{\eps}} \|\bT - \bY\|_\op \le \eps\, .$$
\item
\label{item:2_approx}
The measure $\nu^\up{\eps}$ is supported on $\cC^\up{\eps}$, with weights
$p_\eps(\bT) =\nu^\up{\eps}(\{\bT\})$ satisfying:
\begin{equation}
\nu^\up{\eps} :=\sum_{\bT \in \cC^\up{\eps}} p_\eps(\bT)\delta_{\bT},\quad\quad \sum_{\bT\in \cC^\up{\eps}} p_\eps(\bT) = 1 ,\quad\quad
p_\eps(\bT) > 0 \quad\forall \bT \in \cC^\up{\eps}.
\end{equation}
% \item
% \label{item:3_approx}
% Recall the definition of $w_k$ from Eq.~\eqref{eq:wk-def}.
% Let $w_k^\up{\eps}$ denote the free element
% defined by Eq.~\eqref{eq:wk-def} with $\nu$ replaced by $\nu^\up{\eps}$.
% Then we have
% \begin{equation}
% \|w_k - w_k^\up{\eps}\|_{\cB(\cH^k)} \le C(\alpha) \eps
% \end{equation}
%for some $C(\alpha) >0$ independent of $\eps$.
% In particular,
% denoting by $\mathrm{dist}$ the Hausdorff distance, we have
% \begin{equation}\label{eq:Hausdorff_dist_bound}
% \mathrm{dist}\left(\supp(\mu_\up{\alpha,\nu}),
% \supp(\mu_\up{\alpha,\nu^\up{\eps}})\right) \le C(\alpha) \eps
% \end{equation}
\item
\label{item:3_approx}
Recall the definition of $w_k$ from Eq.~\eqref{eq:wk-def},
and let, for $\eps >0$,
$w_k^\up{\eps}$ denote
the free element
defined by Eq.~\eqref{eq:wk-def} with $\nu$ replaced by $\nu^\up{\eps}$,
Then for any $\bU \in\cU_-(\alpha,\nu)$,
there exists $\eps_0(\bU) >0$ such that for all $\eps < \eps_0(\bU)$,
\begin{equation}
w_k^\up{\eps} - (\bU \otimes \id_\cA) \succ 0.
\end{equation}
\item \label{item:4_approx}
For any $\bU \in\cU_- (\alpha,\nu) \cap \cU_-(\alpha,\nu^\up{\eps})$
\begin{equation}
\|\hat \bS_\up{\alpha,\nu}(\bU) - \hat \bS_\up{\alpha,\nu^\up{\eps}}(\bU)\|_\op \to 0
\end{equation}
as $\eps\to 0$.
%\item
%\label{item:4_approx}
%Denoting by $\mathrm{dist}$ the Hausdorff distance, we have
% \begin{equation}\label{eq:Hausdorff_dist_bound}
% \mathrm{dist}\left(\supp(\mu_\up{\alpha,\nu}),
% \supp(\mu_\up{\alpha,\nu^\up{\eps}})\right) \le C(\alpha) \eps
% \end{equation}
%for some $C(\alpha) >0$ independent of $\eps$.
\end{enumerate}
\end{lemma}
\begin{proof}
For $\eps>0$,
throughout, let $B_{\eps}(\bT_i)$ be the open norm ball of radius $\eps$ around $\bT_i$. Let $\cC_0 := \supp(\nu).$
Construct a \emph{closed} $\eps-$cover $\cC^\up{\eps}$
(i.e., satisfying Eq.~\eqref{eq:closed_cover})
of the compact set $\cC_0$ with balls $\overline B_\eps(\bT_i)$
satisfying
\begin{equation}
\|\bT -\bY\|_\op > \eps\quad\quad \forall \bT,\bY\in \cC^\up{\eps},\; \bT\neq\bY.
\end{equation}
Concretely, start with $\bT_1\in \cC_0$ arbitrary, then for each $i> 1$, arbitrarily choose $\bT_i \in \cC_0 - \bigcup_{j< i}\overline B_\eps(\bT_j)$ until $\cC_0$ is covered by $\{\overline B_\eps(\bT_i)\}$.
This procedure halts after a finite number $m$ of steps by compactness (For example, see Lemma~4.2.6 of
~\cite{vershynin2025high} and the remark that follows it).
Choose $\cC^\up{\eps} =
\{\bT_i:i\le m\}$.
Item~\textit{1} holds by construction
Define
\begin{equation}
S_1 := \overline B_\eps(\bT_1),\quad\quad S_i := \overline B_\eps(\bT_i) \setminus \bigcup_{j < i} \overline B_{\eps}(\bT_j),\quad i > 1
\end{equation}
and set
\begin{equation}
\label{eq:p_defs}
p_\eps(\bT_i) := \nu(S_i).
\end{equation}
Since $\{S_i\}_i$ are disjoint, it is easy to see that
\begin{equation}
\sum_{\bT\in \cC^\up{\eps}} p_\eps(\bT) = \nu(\supp(\nu)) = 1.
\end{equation}
By construction, for each $i\in[m]$, there exists $\delta > 0$ so that
$$B_\delta(\bT_i)\cap \bigcup_{j < i} \overline B_\eps(\bT_j) = \emptyset.$$
And since $\bT_i \in \cC_0$, we must have
\begin{equation}
0 < \nu( B_\delta(\bT_i)) \le p_\eps(\bT_i),
\end{equation}
showing that Item~\textit{2} holds.
For the remaining items, for $\bA\in\C^{k\times k}$, fix $\eps>0$ arbitrarily and let
\begin{eqnarray}
\bY_\eps(\bA) := \sum_{j\in[m]} \bT_j \one_{\bA \in S_j},
\end{eqnarray}
and define i.i.d. random variables
$(\hat\bT_i, \hat\bY_i)_{i\in[n]}$ as follows:
for each $i$,
\begin{equation}
\hat\bT_t \sim \nu, \qquad \text{and} \qquad \hat\bY_i := \bY_\eps( \hat\bT_i).
\end{equation}
Define the following empirical measures:
\begin{equation}
\hat\nu_n := \frac1n\sum_{i=1}^n \delta_{ \bT^\up{i}},\qquad
\hat\nu_n^\up{\eps} := \frac1n\sum_{i=1}^n \delta_{ \bY^\up{i}}
%\hat\nu_n' := \frac1n\sum_{i=1}^n \delta_{ (\bT^\up{i}, \bY^\up{i})}.
\end{equation}
This guarantees that,
almost-surely,
$\supp(\hnu_n)\subseteq \supp(\nu), \supp(\hnu_n^\up{\eps})\subseteq \cC^\up{\eps}$ for all $n$,
and as $n\to\infty$, we have
$\hnu_n\stackrel{w}{\Rightarrow} \nu$,
\begin{equation}
\hnu_n^\up{\eps}(S_j) = \frac1n \#\{i : \hat \bY_i \in S_j\} =
\frac1n \#\{i : \hat \bT_i \in S_j\}
\to \nu(S_j) = p_\eps(\bT_j),\qquad \textrm{for all}\; j\in[m],
\end{equation}
and further, for any bounded continuous function $f:\C^{k\times k}\times \C^{k\times k}\to\R$,
\begin{equation}
\frac1n\sum_{i=1}^n f(\bT^\up{i}, \bY^\up{i}) \to \E[f(\bT, \bY_\eps(\bT))],\qquad
\textrm{where} \qquad \bT \sim \nu.
\end{equation}
Now let $(\cA,\tau)$ be a $C^*$-probability space with Marchenko-Pastur elements $m$, as in Section~\ref{sec:proofs_main},
free from the algebra generated by $(t_{i,j},t^\up{\eps}_{i,j})_{i,j\in[k]}$, which are
in turn defined by
\begin{equation}
\tau \bigg( P\big( \{t_{i,j}\}_{i,j\in[k]}, \{t^\up{\eps}_{i,j}\}_{i,j\in[k]} \big) \bigg)
= \int
P\big( \{\bT_{i,j}\}_{i,j\in[k]}, \{(\bY^\up{\eps}(\bT))_{i,j}\}_{i,j\in[k]} \big)
d\nu(\bT)
\end{equation}
for any non-commutative polynomial $P$,
so that $w_k = (m^{1/2} t_{i,j} m^{1/2})_{i,j\in[k]}$
and $w_k^\up{\eps} = (m^{1/2} t_{i,j}^\up{\eps} m^{1/2})_{i,j\in[k]}$ in the sense of tracial moments.
Now let $\bX\in\R^{n\times p}$ be a matrix with i.i.d. standard Gaussian entries so that it satisfies the
assumptions in Proposition~\ref{prop:asymptotic_FP}, and
define
%
\begin{align}
\bW_k^{\up{\eps}} := \frac1n(\bX \otimes \bI_k)^* \Diag(\hat\bY_i)_{i\le n} (\bX \otimes \bI_k).
\end{align}
Since $\bW_k= n^{-1}(\bX \otimes \bI_k)^* \Diag(\hat\bT_i)_{i\le n} (\bX \otimes \bI_k)$,
\begin{align}
\label{eq:bound_on_W_k}
\|\bW_k^{\up{\eps}} - \bW_k\|_\op
\le \frac{\|\bX\|^2_\op}{n} \max_{i\in[n]} \|\hat\bT_i - \hat\bY_i\|_\op
\le \frac{\|\bX\|^2_\op}{n} \max_{j\in[m]} \max_{\tilde \bT \in S_j} \|\bT_j-\tilde \bT\|_\op \\
%&\le 5\left(1 + \frac1{\alpha}\right) \eps.
\end{align}
and we can bound the difference
\begin{align}
\|w_k - w_k^\up{\eps}\|_{\cB(\cH^k)}
&\stackrel{(a)}{=} \lim_{q\to\infty}
\Big[
\Big(\frac1k \Tr \otimes \tau\Big)(w_k - w_k^\up{\eps})^{2q}
\Big]^{1/(2q)}\\
&\stackrel{(b)}{=}
\lim_{q\to\infty}
\alpha^{1/(2q)}
\lim_{n\to\infty}
\E\Big[\Big(\frac1k \Tr \otimes \frac1n\Tr\Big)\big(\bW_k - \bW_k^\up{\eps}\big)^{2q}\Big]^{1/(2q)}\\
&\stackrel{(c)}{\le}
\lim_{q\to\infty}
\lim_{n\to\infty}
\Big(\alpha\frac{p}{n}\Big)^{1/(2q)}
\E\big[\|\bW_k - \bW_k^\up{\eps}\|_\op^{2q}\big]^{1/(2q)}\\
&\stackrel{(d)}{\le}
\lim_{q\to\infty} \lim_{n\to\infty}
\Big(\alpha\frac{p}{n}\Big)^{1/(2q)}
\frac1n (\sqrt{n} + \sqrt{p} + 2\sqrt{q})^2 \eps
\\
&\le C(\alpha)\, \eps,
\label{eq:bound_on_w_k_w_k_eps}
\end{align}
where in $(a)$ we used Proposition~\ref{prop:spectra_match}, in $(b)$ we used
Eq.~\eqref{eq:free_determinism}, in $(c)$ we used monotonicity of $L^p$ norms, and
in $(d)$ we used a standard bound on the operator norm of Gaussian matrices and Eq.~\eqref{eq:bound_on_W_k}.
Since $\eps>0$ was arbitrary, this shows that Item~\textit{3} holds.
Finally, Item~\textit{4} holds by applying the resolvent identity
\begin{equation}
\hat \bS_\up{\alpha,\nu}(\bU) - \hat \bS_\up{\alpha,\nu^\up{\eps}}(\bU)
=
\frac1\alpha (\bI_k\otimes \tau)\Big((w_k^\up{\eps} - \bU \otimes \id_\cA)^{-1} (w_k- w_k^\up{\eps})(w_k - \bU \otimes \id_\cA)^{-1} \Big),
\end{equation}
using Item~\ref{item:3_approx}
to uniformly bound $\|(w_k^\up{\eps} - \bU \otimes \id_\cA)^{-1}\|, \|(w_k - \bU \otimes \id_\cA)^{-1}\|$
uniformly for $\eps$ sufficiently small along with
Eq.~\eqref{eq:bound_on_w_k_w_k_eps}.
%\newpage
%For Item~\ref{item:3_approx}, fix $\bU\in\cU_-(\alpha,\nu)$.
%We want to show that $\bU\in\cU_-(\alpha,\nu^\up{\eps})$.
%%
%Proposition~\ref{prop:spectra_match} guarantees that this is equivalent to showing
%the following statement:
%let $\mu_0,\mu_0^\up{\eps}$ denote the spectral measures
% whose moments are equal to the tracial moments of $(w_k- \bU \otimes \id_\cA), (w_k^\up{\eps}- \bU \otimes \id_\cA$), respectively.
% Then
% \begin{eqnarray}
% \supp(\mu_0) \subseteq (0,\infty) \qquad\Rightarrow \qquad
% \supp(\mu_0^\up{\eps}) \subseteq (0,\infty).
% \end{eqnarray}
%
%
%
%let $\hnu_n$ be the empirical measure of $n$ i.i.d. samples $\hat\bT_1,\dots,\hat\bT_n\sim \nu$:
%\begin{equation}
% \hnu_n := \frac1n\sum_{i=1}^n \delta_{\hat \bT_i}
%\end{equation}
%which guarantees that $\hnu_n\stackrel{w}{\Rightarrow} \nu$ as $n\to\infty$ and that $\supp(\hnu_n)\subseteq \supp(\nu)$ for all $n$.
%Further, let
%\begin{equation}
%\hat \bY_i :=
%\sum_{j\in[m]} \bT_j \one_{\hat \bT_i \in S_j},\quad\quad
%\hnu_n^\up{\eps} := \frac1n\sum_{i=1}^n \delta_{\hat \bY_i}.
%\end{equation}
%Then for all $j\in[m]$,
%\begin{equation}
% \hnu_n^\up{\eps}(S_j) = \frac1n \#\{i : \hat \bY_i \in S_j\} =
% \frac1n \#\{i : \hat \bT_i \in S_j\}
% \to \nu(S_j) = p_\eps(\bT_j).
%\end{equation}
%So $\hnu^{\up{\eps}}_n \stackrel{w}{\Rightarrow} \nu^\up{\eps}$, and further $\supp(\hnu_n^\up{\eps}) \subseteq \cC^\up{\eps}.$
%Then let $\bX\in\R^{n\times p}$ be a matrix with i.i.d. standard Gaussian entries so that it satisfies the
% assumptions in Proposition~\ref{prop:asymptotic_FP}, and
% define
%%
%\begin{align}
% \bW_k^{\up{\eps}} := \frac1n(\bX \otimes \bI_k)^* \Diag(\hat\bY_i)_{i\le n} (\bX \otimes \bI_k).
%\end{align}
%Since $\bW_k= n^{-1}(\bX \otimes \bI_k)^* \Diag(\hat\bT_i)_{i\le n} (\bX \otimes \bI_k)$,
%\begin{align}
%\label{eq:bound_on_W_k}
% \|\bW_k^{\up{\eps}} - \bW_k\|_\op
% \le \frac{\|\bX\|^2_\op}{n} \max_{i\in[n]} \|\hat\bT_i - \hat\bY_i\|_\op
%\le \frac{\|\bX\|^2_\op}{n} \max_{j\in[m]} \max_{\tilde \bT \in S_j} \|\bT_j-\tilde \bT\|_\op \\
%%&\le 5\left(1 + \frac1{\alpha}\right) \eps.
%\end{align}
%
%Now note that
%\begin{align}
%\|w_k - w_k^\up{\eps}\|_{\cB(\cH^k)}
% &\stackrel{(a)}{=} \lim_{q\to\infty}
% \Big[
%\Big(\frac1k \Tr \otimes \tau\Big)(w_k - w_k^\up{\eps})^{2q}
%\Big]^{1/(2q)}\\
%&\stackrel{(b)}{=}
%\lim_{q\to\infty}
%\alpha^{1/(2q)}
% \lim_{n\to\infty}
%\E\Big[\Big(\frac1k \Tr \otimes \frac1n\Tr\Big)\big(\bW_k - \bW_k^\up{\eps}\big)^{2q}\Big]^{1/(2q)}\\
%&\stackrel{(c)}{\le}
% \lim_{q\to\infty}
% \lim_{n\to\infty}
% \Big(\alpha\frac{p}{n}\Big)^{1/(2q)}
%\E\big[\|\bW_k - \bW_k^\up{\eps}\|_\op^{2q}\big]^{1/(2q)}\\
%&\stackrel{(d)}{\le}
% \lim_{q\to\infty} \lim_{n\to\infty}
% \Big(\alpha\frac{p}{n}\Big)^{1/(2q)}
%\frac1n (\sqrt{n} + \sqrt{p} + 2\sqrt{q})^2 \eps
%\\
%&\le C(\alpha)\, \eps
%\end{align}
%where in $(a)$ we used Proposition~\ref{prop:spectra_match}, in $(b)$ we used
%Eq.~\eqref{eq:free_determinism}, in $(c)$ we used monotonicity of $L^p$ norms, and
%in $(d)$ we used a standard bound on the operator norm of Gaussian matrices and Eq.~\eqref{eq:bound_on_W_k}.
%
%
%%By Proposition~\ref{prop:asymptotic_FP}, the empirical spectral distribution of
%%$\bW_k^{\up{\eps}}$ converges to $\mu_\up{\alpha,\nu^\up{\eps}}$ as $n\to\infty$,
%%and the one of $\bW_k$ converges to $\mu_\up{\alpha,\nu}$.
%%Hence, we obtain the Hausdorff distance bound in Eq.~\eqref{eq:Hausdorff_dist_bound}.
%
%%Then let $\bX\in\R^{n\times p}$ be a matrix with i.i.d. standard Gaussian entries so that it satisfies the distributional assumptions in Proposition~\ref{prop:asymptotic_FP}, and note the estimate
%%\begin{align}
%% \frac1n \|(\bX \otimes \bI_k)^\sT &( \Diag(\hat \bT_i)_{i\le n}-\Diag(\hat\bY_i)_{i\le n}) (\bX \otimes \bI_k)\|_\op
%% \le \frac{\|\bX\|^2_\op}{n} \max_{i\in[n]} \|\hat\bT_i - \hat\bY_i\|_\op\\
%%&\le \frac{\|\bX\|^2_\op}{n} \max_{j\in[m]} \max_{\tilde \bT \in S_j} \|\bT_j-\tilde \bT\|_\op \\
%%&\le 5\left(1 + \frac1{\alpha}\right) \eps,
%%\end{align}
%%which holds with high-probability.
\end{proof}
%\begin{proof}
% %\am{Before, we used $B(center,radius)$.}
% Throughout, let $B(\bT_i,\eps)$ be the open operator norm ball of radius
% $\eps$ around $\bT_i$. Let $\cC_0 := \supp(\nu).$
%
% By compactness of $\cC_0$, there exists a finite $\eps$-cover of $\cC^\up{\eps}$ by balls
% $\{B(\bT_i,\eps)\}_{i\in[m]}$ (hence $m := \# \cC^\up{\eps}$). This proves Item \emph{1}.
%
% Next define
% \begin{equation}
% S_1 := \overline B_\eps(\bT_1),\quad\quad S_i := \overline B_\eps(\bT_i) - \bigcup_{j < i} \overline B_{\eps}(\bT_j),\quad i > 1
% \end{equation}
% and set
% \begin{equation}
% \label{eq:p_defs}
% p_\eps(\bT_i) := \nu(S_i).
% \end{equation}
%Since $\{S_i\}_i$ are disjoint, it is easy to see that
%\begin{equation}
% \sum_{\bT\in \cC^\up{\eps}} p_\eps(\bT) = \nu(\cC_0) = 1.
%\end{equation}
%Without loss of generality we can assume thar, for each $i\in[m]$, there exists $\delta > 0$ so that
%$$B_\delta(\bT_i)\cap \bigcup_{j < i} \overline B_\eps(\bT_j) = \emptyset.$$
%Indeed, if this is not the case for some $i$, we can drop $\bT_i$ from the $\cC^\up{\eps}$ and still have a
%cover of $\cC_0$.
%\bns{Why do we still have a cover? Isn't it possible that we lose
%some points in the support $\cC_0$ when we drop $\bT_i$?
%In my old proof (which is commented above),
%I choose the balls so that no ball has a center
%that lies in another ball. Isn't this needed?
%}
%
%Since $\bT_i \in \cC_0$, we must have
%\begin{equation}
% 0 < \nu( B_\delta(\bT_i)) \le p_\eps(\bT_i),
%\end{equation}
%showing that Item~\textit{2} holds.
%For Item~\textit{3}, let $\hnu_n$ be an empirical measure on $n$ i.i.d. samples $\hat\bT_1,\dots,\hat\bT_n\sim \nu$:
%\begin{equation}
% \hnu_n := \frac1n\sum_{i=1}^n \delta_{\hat \bT_i}
%\end{equation}
%which guarantees that $\hnu_n\stackrel{w}{\Rightarrow} \nu$ as $n\to\infty$ and that $\supp(\hnu_n)\subseteq \supp(\nu)$ for all $n$.
%Further, let
%\begin{equation}
%\hat \bY_i :=
%\sum_{j\in[m]} \bT_j \one_{\hat \bT_i \in S_j},\quad\quad
%\hnu_n^\up{\eps} := \frac1n\sum_{i=1}^n \delta_{\hat \bY_i}.
%\end{equation}
%Then for all $j\in[m]$,
%\begin{equation}
% \hnu_n^\up{\eps}(S_j) = \frac1n \#\{i : \hat \bY_i \in S_j\} =
% \frac1n \#\{i : \hat \bT_i \in S_j\}
% \to \nu(S_j) = p_\eps(\bT_j).
%\end{equation}
%So $\hnu^{\up{\eps}}_n \stackrel{w}{\Rightarrow} \nu^\up{\eps}$, and further $\supp(\hnu_n^\up{\eps}) \subseteq \cC^\up{\eps}.$
%Then let $\bX\in\R^{n\times p}$ be a matrix with i.i.d. standard Gaussian entries so that it satisfies the
% assumptions in Proposition~\ref{prop:asymptotic_FP}, and
% define
%%
%\begin{align}
% \bW_k^{\up{\eps}} := \frac1n(\bX \otimes \bI_k)^\sT \Diag(\hat\bY_i)_{i\le n} (\bX \otimes \bI_k).
%\end{align}
%Since $\bW_k= n^{-1}(\bX \otimes \bI_k)^\sT \Diag(\hat\bT_i)_{i\le n} (\bX \otimes \bI_k)$,
%\begin{align}
% \frac1n \|\bW_k^{\up{\eps}} - \bW_k\|_\op
% &\le \frac{\|\bX\|^2_\op}{n} \max_{i\in[n]} \|\hat\bT_i - \hat\bY_i\|_\op\\
%&\le \frac{\|\bX\|^2_\op}{n} \max_{j\in[m]} \max_{\tilde \bT \in S_j} \|\bT_j-\tilde \bT\|_\op \\
%%&\le 5\left(1 + \frac1{\alpha}\right) \eps.
%\end{align}
%
%Now note that
%\begin{align}
%\|w_k - w_k^\up{\eps}\|_{\cB(\cH^k)}
% &\stackrel{(a)}{=} \lim_{q\to\infty}
% \Big[
%\Big(\frac1k \Tr \otimes \tau\Big)(w_k - w_k^\up{\eps})^{2q}
%\Big]^{1/(2q)}\\
%&\stackrel{(b)}{=}
%\alpha
%\lim_{q\to\infty}
% \lim_{n\to\infty}
%\E\Big[\Big(\frac1k \Tr \otimes \frac1n\Tr\Big)\big(n^{-1}\bW_k - n^{-1}\bW_k^\up{\eps}\big)^{2q}\Big]^{1/(2q)}\\
%&\stackrel{(c)}{=}
% \alpha\lim_{q\to\infty} \lim_{n\to\infty}
%\frac1n \E\big[\|\bW_k - \bW_k^\up{\eps}\|_\op^{2q}\big]^{1/(2q)}\\
%&\stackrel{(d)}{=}
% \alpha \lim_{q\to\infty} \lim_{n\to\infty}
%\frac1n (\sqrt{n} + \sqrt{p} + 2\sqrt{q})^2 \eps
%\\
%&\le C(\alpha)\, \eps
%\end{align}
%where in $(a)$ we used Proposition~\ref{prop:spectra_match}, in $(b)$ we used
%Eq.~\eqref{eq:free_determinism}, in $(c)$ we used monotonicity of $L^p$ norms, and
%in $(d)$ we used a standard bound on the operator norm of Gaussian matrices.
%
%
%%By Proposition~\ref{prop:asymptotic_FP}, the empirical spectral distribution of
%%$\bW_k^{\up{\eps}}$ converges to $\mu_\up{\alpha,\nu^\up{\eps}}$ as $n\to\infty$,
%%and the one of $\bW_k$ converges to $\mu_\up{\alpha,\nu}$.
%%Hence, we obtain the Hausdorff distance bound in Eq.~\eqref{eq:Hausdorff_dist_bound}.
%\end{proof}
\begin{lemma}[$\hat\bS_\up{\alpha,\nu}(\cU_-)\subseteq\cD_0$ for general measures]
\label{lemma:image_Sstar_D_general}
Let $\alpha \in (0,\infty)$.
Assume $\nu \in\cuP_0(\sa{k}{\C})$ (i.e., $\nu$ is a compactly supported measure).
Then for any $\bU \in\cU_-$, we have
\begin{equation}
\hat\bS_\up{\alpha,\nu}(\bU) \in \cD_0.
\end{equation}
\end{lemma}
\begin{proof}
Fix $\bU \in\cU_-$ so that $w_k - (\bU \otimes \id_\cA) \succ \bzero$.
For any $\eps >0$, denote $w_k^\up{\eps}$
the free element defined by Eq.~\eqref{eq:wk-def} with $\nu^\up{\eps}$
instead of $\nu$.
By
Item~\textit{\ref{item:3_approx}} of Lemma~\ref{lemma:discrete_approx},
we can choose some $\eps := \eps(\bU) >0$, depending on $\bU$, so that
$w_k^\up{\eps} - (\bU \otimes \id_\cA) \succ 0$.
Denote $\cU_-^\up{\eps} := \{\bU\in \sa{k}{\C}:
w_k^\up{\eps} - (\bU \otimes \id_\cA) \succ 0\}$ and
$\hat\bS^\up{\eps} (\,\cdot\,) := \hat\bS_{\up{\nu^\up{\eps},\alpha}}(\,\cdot\,)$.
Item~\ref{item:2_approx} of Lemma~\ref{lemma:discrete_approx} guarantees
that the conditions of Lemma~\ref{lemma:image_Sstar_D_discrete} hold,
from which we conclude that
\begin{equation}
\label{eq:smallest_eval_discrete}
\lambda_{\min} \big(\hat\bS^\up{\eps}(\bU)^{-1}+ \tilde\bT\big) > 0 \quad\quad\textrm{for all}\quad\quad \tilde\bT\in\supp(\nu^\up{\eps}).
\end{equation}
Therefore, for any $\bT\in\supp(\nu),$ we have by Item~\textit{\ref{item:1_approx}} of Lemma~\ref{lemma:discrete_approx},
\begin{equation}
\label{eq:smallest_eval_discrete_2}
\lambda_{\min}\big( \hat\bS^\up{\eps}(\bU)^{-1} +\bT\big) \ge -\eps.
\end{equation}
Now Item~\textit{\ref{item:4_approx}} of
Lemma~\ref{lemma:discrete_approx} along with
%guarantees that
%\begin{equation}
% \supp(\nu^\up{\eps}) \subseteq \supp(\nu) \subseteq [-C(\alpha)\eps, C(\alpha)\eps] + \supp(\nu^\up{\eps}),
%\end{equation}
%meanwhile,
the fact that $\bU \in \cU_-\cap\cU_-^\up{\eps}$ guarantees that $\hat\bS^\up{\eps}(\bU) \to \hat\bS(\bU)$ as $\eps\to 0$.
Combining with the bound
of Eq.~\eqref{eq:smallest_eval_discrete_2}, we obtain
\begin{equation}
\lambda_{\min}\big( \hat\bS(\bU)^{-1} +\bT\big) \ge 0.
\end{equation}
%\begin{equation}
% 0 \le \lim_{\eps\to 0} \inf_{\bT \in \supp(\nu^\up{\eps})} \lambda_{\min}(\bI + \hat\bS^{\up{\eps}}(\bU)^{1/2} \bT \hat\bS^{\up{\eps}}(\bU)^{1/2})
% \le
% \inf_{\bT \in \supp(\nu)} \lambda_{\min}(\bI + \bS^{\up{\eps}}(\bU)^{1/2} \bT \bS^{\up{\eps}}(\bU)^{1/2}),
%\end{equation}
Since $\bU\in\cU_-$ was arbitrary, we have shown that for all $\bU\in\cU_-$, $\hat\bS(\bU) \in \overline \cD_0.$
Then since Lemma~\ref{lemma:regularity_free_ST} guarantees that $\hat\bS$ has a non-singular
derivative at $\bU \in\cU_-$,
the inverse function theorem in turn guarantees that
$\hat\bS(\cU_-)$ is an open set, and is hence contained in $\cD_0$.
\end{proof}
\paragraph{An alternative characterization of $\cP_-$.}
Before continuing on to show that $\hat\bS_\up{\alpha,\nu}(\cU_-) \subseteq \cP_-$, we will give a characterization of $\cP_-$ that will be useful in the proof.
We associate to each $\bS \in\cD_0$ the following operator mapping matrices to matrices:
\begin{equation}
\cuT_\bS(\bDelta) := \alpha\int \bZ(\bI + \bZ)^{-1} \bDelta \bZ(\bI + \bZ) ^{-1}\nu(\de\bT)\quad\quad \bZ := \bS^{1/2}\bT\bS^{1/2}.
\end{equation}
Since $\bZ$ commutes with $(\bI +\bZ)^{-1}$, $\cuT_\bS$ is self-adjoint as a bounded linear operator on the space of self-adjoint matrices
endowed with the trace inner product. Further, it is immediate to see
that $\<\bDelta, \cuT_\bS(\bDelta)\> \ge 0$ for all $\bDelta\in\sa{k}{\C}$, i.e. $\cuT_\bS$ is
positive semidefinite.
Our interest in this operator comes from its relationship with the derivative of the function $\bK\equiv \bK_\up{\alpha,\nu}$, namely, observe that for all $\bS \in\cD_0$, we have
\begin{equation}
\label{eq:DK_to_cuT}
\bS^{1/2}\cuD_\bS \bK(\bS)[\bDelta]\bS^{1/2} = \frac1\alpha \left( \bS^{-1/2} \bDelta \bS^{-1/2} - \cuT_\bS(\bS^{-1/2}\bDelta \bS^{-1/2})\right).
\end{equation}
Hence, for any $\bDelta\in\sa{k}{\C}$,
defining $\bA := \bS^{-1/2}\bDelta\bS^{-1/2}$,
we obtain
\begin{equation}
\label{eq:DK_to_cuT_norm}
\langle
\bDelta,\cuD_\bS \bK(\bS)[\bDelta] \rangle_\fro
=
\frac1\alpha \big(
\langle \bA,\bA\rangle - \langle \bA ,\cuT_\bS(\bA) \rangle
\big).
\end{equation}
Therefore, using the fact that $\cuT_\bS$ is positive semidefinite,
$\cP_-$ can be alternatively written as
\begin{equation}
\cP_- = \{\bS\succ\bzero : \|\cuT_\bS\|_{F\to F} <1 \} \cap \cD_0\, .
\end{equation}
The following lemma gives a further characterization that turns out to
be useful.
\begin{lemma}
\label{lemma:set_characterization}
For $\alpha\in(0,\infty)$, integer $k \ge 1,$
$\nu\in\cuP_0(\sa{k}{\C}),$
define
\begin{equation}
\tilde\cP_- := \{\bS \succ\bzero : \exists \bB\succ\bzero\;\;\textrm{s.t.}\;\; \cuT_\bS(\bB)\prec \bB
\} \cap \cD_0.
\end{equation}
Then $\cP_- = \tilde\cP_-$.
\end{lemma}
\begin{proof}
%Throughout, will use the less cumbersome notation $\|\cuT_\bS\|\equiv \|\cuT_\bS\|_{F\to F}$.
%We will also drop the subscript $\bS$ and write $\cuT$ when $\bS$ is clear from context.
The inclusion $\cP_- \subseteq \tilde\cP_-$ holds more generally for linear self-adjoint operators. To see this,
fix $\bS\in\cP_-$ and any $\bC\succ\bzero$ arbitrary, then define
\begin{equation}
\bB := \left(\bI + \sum_{p=1}^\infty \cuT_\bS^p\right)(\bC)
\end{equation}
where $\cuT_\bS^p$ denotes the $p-$fold composition of $\cuT_\bS$.
Since $\|\cuT_\bS\|_{\fro\to\fro} < 1$,
the operator in the above display is bounded in $\|\cdot\|_{\fro\to\fro}$ since the sum is geometric.
Clearly, $\cuT_\bS$ maps PSD matrices to PSD matrices so $\bB\succ\bzero.$
Meanwhile, by linearity
\begin{equation}
\cuT_\bS(\bB) = \sum_{p=1}^\infty \cuT_\bS^p(\bC) = \bB - \bC \prec \bB
\end{equation}
since $\bC \succ\bzero.$ i.e., we obtain the estimate
$\bzero \preceq \cuT_\bS(\bB) \prec \bB$,
which implies $\bS\in\tilde\cP_-.$
To show the opposite inclusion, fix $\bS\in\tilde\cP_-$,
and introduce the expectation notation
\begin{equation}
\cuT_\bS(\bDelta) = \alpha \E[\bZ (\bI+\bZ)^{-1} \bDelta \bZ (\bI+\bZ)^{-1}] ,\quad\quad \bZ := \bS^{1/2} \bT\bS^{1/2}\, .
\end{equation}
Here the expectation is with respect to the pushforward of the probability measure $\nu$.
We will first obtain a bound on the norm $\|\cuT_\bS^p\|_{\fro\to\fro}$ in order to bound the spectral radius of $\cuT_\bS$.
To do so, fix $\bDelta \in\sfM_{k}(\C)_{\textrm{sa}}$ and let
\begin{equation}
\bT_1,\dots,\bT_p,\tilde\bT_1,\dots,\tilde\bT_p
\end{equation}
i.i.d. copies of $\bT$ from $\nu$, and denote
\begin{equation}
\bA_i := \bZ_i (\bI+\bZ_i)^{-1} \quad \bZ_i := \bS^{1/2}\bT_i\bS^{1/2}.
\end{equation}
Similarly define $\tilde \bA_i,\tilde \bZ_i$ using $\tilde\bT_i.$ Then from the definition of $\cuT_\bS$, we see that
\begin{align}
\cuT_\bS^p(\bDelta) &= \alpha^p\E\left[\bA_p\cdots \bA_1 \bDelta \bA_1 \cdots\bA_p \right]
\end{align}
Using the fact that $2\Tr(\bDelta\bQ\bDelta\bQ^{*})\le
\Tr(\bDelta^2\bQ\bQ^{*})+\Tr(\bDelta^2\bQ\bQ^{*})$, we can bound
\begin{align}
\|\cuT_\bS^p(\bDelta)\|_\fro^2 &=
\alpha^{2p}\E\left[
\Tr\left(\bA_p\cdots \bA_1 \bDelta \bA_1 \cdots\bA_p
\tilde\bA_p\cdots \tilde\bA_1 \bDelta \tilde\bA_1 \cdots\tilde\bA_p
\right)
\right]\\
&\le
\alpha^{2p}
\E\left[
\Tr\left(\bDelta \bA_1 \cdots\bA_p
\tilde\bA_p\cdots \tilde\bA_1 \tilde\bA_1 \cdots\tilde\bA_p
\bA_p\cdots \bA_1 \bDelta
\right)
\right].
\end{align}
Since the matrix inside the trace is PSD, for any $\bB\succ\bzero,$ we have
\begin{align}
&\E\left[
\Tr\left(\bDelta \bA_1 \cdots\bA_p
\tilde\bA_p\cdots \tilde\bA_1\bB^{1/2}\bB^{-1}\bB^{1/2} \tilde\bA_1 \cdots\tilde\bA_p
\bA_p\cdots \bA_1 \bDelta
\right)
\right]\\
&\le
\|\bB^{-1}\|_{\op}
\E\left[
\Tr\left(\bDelta \bA_1 \cdots\bA_p
\tilde\bA_p\cdots \bB^{1/2} \E[\bB^{-1/2} \tilde\bA_1\bB\tilde\bA_1\bB^{-1/2}]\bB^{1/2} \cdots\tilde\bA_p
\bA_p\cdots \bA_1 \bDelta
\right)
\right]\\
&\le\|\bB^{-1}\|_\op
\|\E[\bB^{-1/2} \tilde\bA_1\bB\tilde\bA_1\bB^{-1/2}]\|_\op
\E\left[
\Tr\left(\bDelta \bA_1 \cdots\bA_p
\tilde\bA_p\cdots \tilde\bA_2\bB\tilde\bA_2 \cdots\tilde\bA_p
\bA_p\cdots \bA_1 \bDelta
\right)
\right].
\end{align}
Repeating this argument $2p-1$ many times, we obtain that, for any $\bB\succ\bzero$,
\begin{equation}
\|\cuT_\bS^p(\bDelta)\|_\fro^2 \le\alpha^{2p} \|\bB^{-1}\|_\op \|\bB\|_\op
\|\E[\bB^{-1/2}\bA\bB\bA\bB^{-1/2}]\|_\op^{2p} \|\bDelta\|_\fro^2,
\end{equation}
i.e.,
\begin{align}
\label{eq:F_to_F_bound_composition_A}
\|\cuT_\bS^p\|_{\fro\to\fro} \le \alpha^{p} \|\bB\|_\op^{1/2}\|\bB^{-1}\|_\op^{1/2}
\|\E[\bB^{-1/2}\bA\bB\bA\bB^{-1/2}]\|_\op^{p},
\end{align}
where $\bA = \bZ(\bI + \bZ)^{-1}$.
Consequently, if there is a $\bB\succ\bzero$ satisfying $\cuT_\bS(\bB)\prec\bB$, then for some $\delta\in(0,1),$
\begin{equation}
\|\E[\bB^{-1/2} \bA \bB\bA \bB^{-1/2}]\|_\op =
\frac1{\alpha}\|\bB^{-1/2} \cuT_\bS(\bB) \bB^{-1/2}\|_\op < \frac{\delta}{\alpha}.
\end{equation}
Then since $\cuT_\bS^p$ is self-adjoint, combining the last bound with
with Eq.~\eqref{eq:F_to_F_bound_composition_A}, we obtain that
\begin{equation}
\|\cuT_\bS\|_{\fro\to\fro} =\lim_{p\to\infty} \|\cuT_\bS^p\|_{\fro\to\fro}^{1/p}
\le \delta \,\lim_{p\to\infty} C^{1/p} < 1\, .
\end{equation}
\end{proof}
\paragraph{The image of $\cU_-$ is contained in $\cP_-$.}
Finally, we return to show that $\hat\bS_\up{\alpha,\nu}(\cU_-)\subseteq\cP_-.$
\begin{lemma}
\label{lemma:image_U_in_P}
Fix $k \ge 1$ integer,
$\alpha \in (0,\infty)$, and $\nu\in\cuP_0(\sa{k}{\C})$.
Then for any $\bU \in\cU_-$, we have
\begin{equation}
\hat\bS_\up{\nu;\alpha}(\bU) \in \cP_-.
\end{equation}
\end{lemma}
\begin{proof}
Fix $\bU \in\cU_-$ throughout the proof, and consider for $y \in(0,1)$,
\begin{equation}
\bZ_y := \bU + i y \bI,
\end{equation}
and let
\begin{equation}
\bS_y := \hat\bS(\bZ_y)= \bA_y + i \bB_y \in \bbH_+^k.
\end{equation}
Recall that
%\am{Changed sign in the next eq. I think it was wrong and was not consistent with what follows.}
\begin{equation}
\bK(\bS_y) = \bZ_y\, .
\end{equation}
Using Lemma~\ref{lemma:inv_on_upper_plane}
and taking the imaginary part of both sides to get
\begin{equation}
y \bI = \frac1\alpha \bS_y^{-1} \bB_y \bS_y^{*-1} - \E[(\bI + \bT \bS_y)^{-1}\bT \bB_y \bT (\bI + \bS_y^* \bT)^{-1}],
\end{equation}
which, after re-arranging, gives
\begin{equation}
\bB_y^{-1/2}\bS_y \E[(I + \bT \bS_y)^{-1}\bT \bB_y \bT (\bI + \bS_y^* \bT)^{-1}] \bS_y^*\bB_y^{-1/2}
= \frac1\alpha\bI - y\bB_y^{-1/2} \bS_y \bS_y^*\bB_y^{-1/2}.\label{eq:Rearranged}
\end{equation}
Define $r_y = \big(w_k -(\bZ_y \otimes \id_\cA)\big)^{-1}\in \mat{k}{\C}\otimes\cA$ and note that
\begin{align}
\alpha \bB_y &= \Im\circ (\bI_k \otimes \tau)\left(r_y\right)
= y\cdot(\bI_k \otimes \tau)(r_y r_y^*).
\end{align}
Noting that
\begin{equation}
\lim_{y\to 0} r_y = r_0(\bU):=
\big(w_k -(\bU \otimes \id_\cA)\big)^{-1}
\succ \bzero,\quad\quad
\lim_{y\to 0} \bS_y = \hat\bS(\bU)\succ\bzero,
\end{equation}
we have $\lim_{y\to 0}\bB_y/y=\bB_0 := \alpha^{-1}(\bI_k \otimes \tau)(r_0(\bU)^2)$ and therefore
\begin{equation}
\lim_{y\to 0}y\bB_y^{-1/2} \bS_y \bS_y^*\bB_y^{-1/2} =
\bB_0^{-1/2} \hat\bS(\bU)^2\bB_0^{-1/2} \succ\bzero\, .
\end{equation}
Meanwhile, since $\hat\bS(\bU) \in \cD_0$ by Lemma~\ref{lemma:image_Sstar_D_general}, we have,
as $y\to 0$,
\begin{equation}
\bB_y^{-1/2}\bS_y \E[(\bI + \bT \bS_y)^{-1}\bT \bB_y \bT (\bI + \bS_y^* \bT)^{-1}] \bS_y^*\bB_y^{-1/2} \to
\frac{1}{\alpha} \bB_0^{-1/2}
\hat\bS^{1/2}\cuT_{\hat\bS(\bU)}\big(\hat\bS^{-1/2}\bB_0\hat\bS^{-1/2}\big) \hat\bS^{1/2}\bB_0^{-1/2}\, ,
\end{equation}
where $\hat\bS \equiv \hat\bS(\bU).$
Using the last two displays in Eq.~\eqref{eq:Rearranged}, we obtain that
\begin{equation}
\cuT_{\hat\bS(\bU)}(\bB') \prec \bB'
\end{equation}
for the matrix (strictly) positive definite matrix $\bB' := \hat\bS^{-1/2}\bB_0 \hat\bS^{-1/2}$, so that $\hat\bS \in \tilde\cP_-$ of Lemma~\ref{lemma:set_characterization}, which is equal to $\cP_-.$
\end{proof}
\paragraph{Global injectivity of $\bK$ and connectedness of $\cP_-$.}
Note that non-degeneracy of the derivative of $\bK$ on $\cP_-$ guarantees that it is locally injective.
We will next show that $\cP_-$ is an open connected set and that $\bK$ is globally injective on it.
%\am{I do not understand the next sentence. Also, did we prove that $\bK(\cP_-)\subseteq\cU_-$?}
%Since Lemma~\ref{lemma:image_U_in_P} asserts that the composition $\hat\bS = \bK^{-1}|_{\cU_-}$,
This will allow us to conclude that
we cannot have $\hat \bS(\cU) \subsetneq \cP_-$
unless $\hat \bS$ has an analytic continuation through a boundary point of $\cU_-.$
We start off with the global injectivity in the following lemma.
\begin{lemma}[Injectivity of $\bK$ on $\cP_-$]
\label{lemma:K_injective}
Assume $\nu$ is compactly supported.
Then $\bK$ is \emph{globally injective} on $\cP_-$.
\end{lemma}
\begin{proof}
Once again, we use the expectation notation.
Fix $\bS,\tilde\bS\in\cP_-$
and define the operator
\begin{equation}
\cuG_{0}[\bDelta] := \alpha \E[\bZ(\bI+\bZ)^{-1}\bDelta \tilde\bZ(\bI+\tilde\bZ)^{-1}]
\quad\quad\textrm{where}\quad\quad
\bZ := \bS^{1/2} \bT\bS^{1/2},\;\tilde\bZ :=\tilde\bS^{1/2} \bT\tilde\bS^{1/2},
\end{equation}
for $\bDelta\in\sfM_k(\C).$
A straightforward calculation shows that this relates to the difference $\bK(\bS) -\bK(\tilde\bS)$ by
\begin{align}
\alpha \bS^{1/2} \left(\bK(\bS) - \bK(\tilde\bS)\right)\tilde\bS^{1/2} =
\left(\bI - \cuG_0\right)[\bDelta_0],\quad\quad \bDelta_0 := \bS^{-1/2}(\bS -\tilde\bS)\tilde\bS^{-1/2}.
\end{align}
Hence, to prove injectivity, since $\bS,\tilde\bS\succ\bzero,$
it is sufficient to show that $\bI -\cuG_0$ is invertible.
To do so, we will follow an argument similar to the one in Lemma~\ref{lemma:set_characterization} and bound the norm of compositions of $\cuG_0^p.$
Let $\bT_1,\dots,\bT_{2p}$ be distributed with respect to the product $\nu^{\otimes 2p}$.
Set for $i\in[2p],$
\begin{equation}
\bA_{i} := \bZ_i (\bI+\bZ_i)^{-1},\quad\bZ_i := \bS^{1/2}\bT_i\bS^{1/2},\quad\quad\textrm{and}\quad\quad
\tilde\bA_{i} := \tilde\bZ_i (\bI+\tilde\bZ_i)^{-1},\quad\tilde\bZ_i := \tilde\bS^{1/2}\bT_i\tilde\bS^{1/2}.
\end{equation}
Then for $\bDelta\in\sfM_{k}(\C)$, we have
\begin{align}
\|\cuG_0^p(\bDelta)\|_\fro^2
&= \alpha^{2p}\E[\Tr(
\tilde\bA_{p} \dots \tilde\bA_1
\bDelta^*
\bA_1\dots\bA_p
\bA_{p+1} \dots \bA_{2p}
\bDelta
\tilde\bA_{2p}
\dots
\tilde\bA_{p+1}
)]\\
&= \alpha^{2p}\E[\Tr(
\bA_1 \dots \bA_{2p}
\bDelta
\tilde\bA_{2p}
\dots
\tilde\bA_1
\bDelta^*
)]\\
&\le \alpha^{2p}
{\underbrace{\E[\Tr(
\bDelta^* \bA_{2p}\dots\bA_1 \bA_1\dots \bA_{2p} \bDelta
)]}_{T_1}}^{1/2}
{\underbrace{\E[\Tr(
\bDelta \tilde\bA_1\dots\tilde\bA_{2p}\tilde\bA_{2p} \dots \tilde\bA_1\bDelta^{*}
)]}_{T_2}}^{1/2}.
\end{align}
Now since $\bS,\tilde\bS \in\cP_-$ which is equal to $\tilde\cP_-$ by Lemma~\ref{lemma:set_characterization}, there exist $\bB,\tilde\bB\succ\bzero$ so that
$\cuT_{\bS}[\bB] \prec \bB, \cuT_{\tilde\bS}[\tilde\bB]\prec \tilde\bB$.
Then we bound $T_1$ as
\begin{align}
T_1 &\le \|\bB^{-1}\|_\op \E[ \Tr(\bDelta^{*}\bA_{2p} \dots
\bA_{2}\bB^{1/2}\E[\bB^{-1/2}\bA_{1}\bB\bA_1 \bB^{-1/2}] \bB^{1/2}\bA_2 \dots \bA_{2p}\bDelta) ]\\
&\le
\|\bB^{-1}\|_\op
\|\E[\bB^{-1/2}\bA_{1}\bB\bA_1 \bB^{-1/2}]\|_\op
\E[ \Tr(\bDelta^{*}\bA_{2p} \dots \bA_2\bB\bA_2 \dots \bA_{2p}\bDelta)]
\end{align}
which after iterating yields
\begin{equation}
T_1 \le \alpha^{-p}\|\bB^{-1}\|_\op
\|\bB\|_\op \|\bB^{-1/2}\cuT_\bS(\bB)\bB^{-1/2}\|_\op^p \Tr(\bDelta^{*}\bDelta)
\end{equation}
and a similar computation shows
\begin{equation}
T_2 \le \alpha^{-p}\|\tilde\bB^{-1}\|_\op
\|\tilde\bB\|_\op \|\tilde\bB^{-1/2}\cuT_\bS(\tilde\bB)\tilde\bB^{-1/2}\|_\op^p \Tr(\bDelta^{*}\bDelta),
\end{equation}
so that for some $C>0$ and $\delta\in(0,1),$
\begin{equation}
\|\cuG_0^{p}\|_{\fro\to\fro} \le
C \delta^p.
\end{equation}
This shows that the Nuemann series $\sum_{p=0}^\infty \cuG_{0}^p$ converges in $\|\,\cdot\,\|_{\fro\to\fro}$ norm, and thereby $\bI -\cuG_0$ is invertible
\end{proof}
Now we move on to showing that the connectedness of $\cP_-.$
\begin{lemma}
\label{lemma:connectedness}
Assume $\nu$ is compactly supported.
Then the set $\cP_-$ is a connected open set.
\end{lemma}
\begin{proof}
Recall first the characterization $\tilde\cP_-$ from
Lemma~\ref{lemma:set_characterization}.
Observe that for a given $\bS \in\cD_0$, the condition
$\cuT_\bS(\bB) \prec\bB$
is equivalent to
\begin{equation}
\alpha\E[\bT(\bS^{-1} +\bT)^{-1} \bB_0 (\bS^{-1} + \bT)^{-1} \bT ] \prec \bB_0 \quad\quad \bB_0 := \bS^{-1/2} \bB \bS^{-1/2}.
\end{equation}
%
So it is sufficient to show the connectedness of the set
%\am{Shouldn't it be $\exists \bB_0 \succ\bzero$ such that ...?}
\begin{equation}
\label{eq:alt_connectedness_set}
\cG := \{\bY \succ\bzero : \bY + \bT \succ\bzero \; \;\forall\bT\in\supp(\nu), \;\;
\exists \bB_0 \succ\bzero \;\textrm{s.t.}\;
%\E[\bT(\bY +\bT)^{-1} \bB_0 (\bY + \bT)^{-1}\bT]
\cuF_0(\bY)[\bB_0]
\prec\bB_0
\}
\end{equation}
where %\am{Changed notation from $\cuT_0$ to $\cuF_0$ to avoid confusion with the function $\cuT_\bS$.}
\begin{equation}
\cuF_0(\bY)[\bB_0] :=
\alpha\E[\bT(\bY +\bT)^{-1} \bB_0 (\bY + \bT)^{-1}\bT].
\end{equation}
First, fix
$\bY\in\cG$ and let $\bB_0\succ\bzero$ satsify $\cuF_0(\bY)[\bB_0] \prec \bB_0$.
We will show that for all $t\ge0,$ the path
\begin{equation}
\bY_t := \bY + t \bB_0
\end{equation}
remains in $\cG.$
It is clear that $\bY_t + \bT \succ\bzero$ for $\bT\in\supp(\nu)$.
For the remaining condition, set
\begin{equation}
\bQ \equiv \bQ(\bT) := \bB_0^{-1/2} (\bY + \bT) \bB_0^{-1/2}.
\end{equation}
Since $\bY\in\cG$, $\bQ(\bT) \succ\bzero$ for $\bT\in\supp(\nu)$ and so $(\bQ + t \bI)^2 = \bQ^2 + 2t\bQ + t^2 \bI \succ \bQ^2$ for all $t\ge 0$. Recalling that the inverse is matrix monotone on PSD matrices we conclude that for all
\begin{equation}
\label{eq:ordering_key}
\bT\bB_0^{-1/2}(\bQ(\bT) + t\bI)^{-2}
\bB_0^{-1/2}\bT
\preceq
\bT\bB_0^{-1/2}
\bQ(\bT)^{-2}
\bB_0^{-1/2}\bT
,
\end{equation}
almost-surely.
Now observe that for all $t\ge0$,
\begin{align}
\E[\bT\bB_0^{-1/2}(\bQ(\bT) +t\bI)^{-2} \bB_0^{-1/2}\bT]
&=
\E[\bT (\bY +\bT + t\bB_0)^{-1} \bB_0
(\bY +\bT + t\bB_0)^{-1}\bT]\\
&=
\E[\bT (\bY_t + \bT )^{-1} \bB_0
(\bY_t +\bT )^{-1}\bT]\\
&=\alpha^{-1}\cuF_0(\bY_t)[\bB_0],
\end{align}
Eq.~\eqref{eq:ordering_key} then allows us to conclude that
$\cuF_0(\bY_t)[\bB_0] \preceq \cuF_0(\bY)[\bB_0] \prec \bB_0$
so that $\bY_t \in \cG$.
Now let $C_\alpha> 0$ be chosen sufficiently large so that
\begin{eqnarray}
\alpha \E[\bT^2] \le C_\alpha^2.
\end{eqnarray}
Consider the set
\begin{equation}
\cG_0 := \{\bY \succ\bzero : \bY + \bT \succ C_\alpha \bI \;\; \forall\bT\in\supp(\nu)\}.
\end{equation}
Clearly, if $\bY \in\cG_0$, then $\bY + \bT \succ\bzero$ for $\bT\in\supp(\nu)$.
Further, for all $\bY \in\cG_0$ and $\bT\in\supp(\nu)$,
we have
\begin{eqnarray}
\cuF_0(\bY)[\bI] =
\alpha \E\big[\bT (\bY + \bT)^{-2} \bT\big]
\prec \frac{\alpha}{C_\alpha^2} \E[\bT^2] \bI \preceq\bI,
\end{eqnarray}
so that $\bT \in\cG$. Hence, $\cG_0$ is a convex subset of $\cG$.
Now if $\bY\in\cG$, with $\bB_0\succ \bzero$
satisfying $\cuF_0(\bY)[\bB_0] \prec \bB_0$,
since $\bT$ is compactly supported, there exists $t>0$ large enough so that $\bY + t\bB_0 \in\cG_0$.
%We next claim that there exists $t>0$ large enough so that
%letting $\bY_{t,\lambda}:=(1-\lambda)\bY_{t}+\lambda\tilde\bY_{t}$,
%we have $\bY_{t,\lambda}\in\cG$ for all $\lambda\in[0,1]$.
%Together with the above, this implies that $\cG$ is path-connected.
%
%In order to prove the claim, we first note the following:
%\begin{itemize}
%\item
%We have $\bY_{t,\lambda}+\bT\succ\bzero$ for all $\bT\in\supp(\nu)$
%because \am{???}
%\am{I do not understand the next sentence}
%(For example, choose $t$ so that $\bY_{t} + 10 \bT \succ\bzero, \tilde\bY_t +10 \bT \succ\bzero$ for all $\bT\in\supp(\nu).$)
%\item We have that $\cF_0[\bY_{t,\lambda}][\bB_{\lambda}] \prec \bB_{\lambda}$ for
%$\bB_{\lambda}=???$ because \am{???}
%\end{itemize}
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:K_transform_real}]
Let $\partial\cU_-$ denote the boundary
\begin{equation}
\partial\cU_- = \big\{\bU \in\sa{k}{\C} :
w_k - (\bU \otimes\id)\succeq 0,\; 0\in\spec(w_k - (\bU\otimes \id))
\big\}.
\end{equation}
Assume by contradiction that $\cU_- \neq \bK(\cP_-)$. Since
$\hat\bS(\cU_-)\subseteq\cP_-$ by Lemma~\ref{lemma:image_U_in_P},
and $\bK(\hat\bS(\bU))=\bU$ for all $\bU\in\cU_-$ by Lemma \ref{lemma:regularity_free_ST}
it follows that $\cU_-\subseteq\bK(\cP_-)$.
The set $\cP_-$ is connected by Lemma~\ref{lemma:connectedness}, and hence so is $\bK(\cP_-)$.
Furthermore,
both $\cU_-$ and $\bK(\cP_-)$ are open; the latter is open because $\cP_-$ is open and
$\bK$ is a local diffeomorphism. As a result, we conclude that there must exist a point $\bU_0\in\bK(\cP_-)\cap\partial\cU_-$.
%
%we must have some point $\bS_0\in\partial\cP_-\cap\hat\bS(\cU_-)$ with a neighborhood contained in $\hat\bS(\cU_-).$
%$\cU_-\subseteq \bK(\cP_-)$ by Lemma~\ref{lemma:image_U_in_P}, $\bK(\cP_-)$ is connected by Lemma~\ref{lemma:connectedness} and both are open, we must have some point $\bU_0\in\partial\cU_-\cap\bK(\cP_-)$ with a neighborhood contained in $\bK(\cP_-).$
Now by Lemma~\ref{lemma:K_injective}, (and the fact that $\cuD_\bS \bK(\bS)[\,\cdot\,]\succ\bzero$
for all $\bS\in\cP_-$), $\bK$ has an analytic inverse
$\bK^{-1}$ on $\bK(\cP_-)$ and by Lemma~\ref{lemma:image_U_in_P} we must have
\begin{equation}
\label{eq:S_is_K_inv}
\bU\in\cU_- \;\;\Rightarrow\;\; \hat\bS(\bU) = \bK^{-1}(\bU).
\end{equation}
Therefore, $\hat\bS$ can be continued analytically in a neighborhood of $\bU_0$.
But then the Stieltjes transform
\begin{equation}
s_0(x) := \frac1{\alpha k}\left(\Tr\otimes\tau\right) \left(w_k - ( (\bU + x\bI )\otimes \id)\right)^{-1},\quad\quad
\bU_x := \bU + x\bI,\quad\quad x < 0
\end{equation}
would have an analytic continuation through $0$, which is a point in the support of the measure defined by $w_k -(\bU \otimes \id)$ through tracial moments by Proposition~\ref{prop:spectra_match}; a contradiction.
Finally, the equivalence of the definitions in Eq.~\eqref{eq:P_def} and Eq.~\eqref{eq:P_def_with_B}
follows from Lemma~\ref{lemma:set_characterization} and
Eq.~\eqref{eq:DK_to_cuT}.
\end{proof}
\subsection*{Statement of AI use}
The proof of Lemma~\ref{lemma:connectedness} was given
by GPT 5.5 Pro, prompted by Alexandru Lopotenco with
the parameterization of $\cP_-$ appearing in
Eq.~\eqref{eq:alt_connectedness_set}. We are grateful to the
Alexandru for this.
\subsection*{Acknowledgements}
We are grateful to Ramon van Handel for his comments on this manuscript, and
clarifications regarding their \cite{parmaksiz2025computing}.
This work was partially supported by the NSF through
Award MFAI-2501597.
\bibliographystyle{amsalpha}
\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}{AGHJ25}
\bibitem[AEK17]{alt2017local}
Johannes Alt, L{\'a}szl{\'o} Erd{\"o}s, and Torben~H Kr{\"u}ger, \emph{Local
law for random gram matrices}, Electronic Journal of Probability \textbf{22}
(2017).
\bibitem[AGHJ25]{arous2025local}
Gerard~Ben Arous, Reza Gheissari, Jiaoyang Huang, and Aukosh Jagannath,
\emph{Local geometry of high-dimensional mixture models: Effective spectral
theory and dynamical transitions}, arXiv preprint arXiv:2502.15655 (2025).
\bibitem[AGZ10]{anderson2010introduction}
Greg~W Anderson, Alice Guionnet, and Ofer Zeitouni, \emph{An introduction to
random matrices}, no. 118, {Cambridge University Press}, 2010.
\bibitem[AMS25]{asgari2025local}
Kiana Asgari, Andrea Montanari, and Basil Saeed, \emph{Local minima of the
empirical risk in high dimension: General theorems and convex examples},
arXiv preprint arXiv:2502.01953 (2025).
\bibitem[CD11]{couillet2011random}
Romain Couillet and Merouane Debbah, \emph{Random matrix methods for wireless
communications}, Cambridge University Press, 2011.
\bibitem[Cha06]{chatterjee2006generalization}
Sourav Chatterjee, \emph{A generalization of the lindeberg principle}, The
Annals of Probability (2006), 2061--2076.
\bibitem[FOBS06]{far2006spectra}
Reza~Rashidi Far, Tamer Oraby, Wlodzimierz Bryc, and Roland Speicher,
\emph{Spectra of large block matrices}, arXiv preprint cs/0610045 (2006).
\bibitem[HFS07]{helton2007operator}
J~William Helton, Reza~Rashidi Far, and Roland Speicher, \emph{Operator-valued
semicircular elements: solving a quadratic matrix equation with positivity
constraints}, International Mathematics Research Notices \textbf{2007}
(2007), no.~9, rnm086--rnm086.
\bibitem[HLN07]{hachem2007deterministic}
Walid Hachem, Philippe Loubaton, and Jamal Najim, \emph{Deterministic
equivalents for certain functionals of large random matrices}, The Annals of
Applied Probability (2007), 875--930.
\bibitem[Leh99]{lehner1999computing}
Franz Lehner, \emph{Computing norms of free operators with matrix
coefficients}, American Journal of Mathematics \textbf{121} (1999), no.~3,
453--486.
\bibitem[LL20]{lu2020phase}
Yue~M Lu and Gen Li, \emph{Phase transitions of spectral initialization for
high-dimensional non-convex estimation}, Information and Inference: A Journal
of the IMA \textbf{9} (2020), no.~3, 507--541.
\bibitem[MAB20]{maillard2020landscape}
Antoine Maillard, G{\'e}rard~Ben Arous, and Giulio Biroli, \emph{Landscape
complexity for the empirical risk of generalized linear models}, Mathematical
and Scientific Machine Learning, PMLR, 2020, pp.~287--327.
\bibitem[MM19]{mondelli2019fundamental}
Marco Mondelli and Andrea Montanari, \emph{Fundamental limits of weak recovery
with applications to phase retrieval}, Foundations of Computational
Mathematics \textbf{19} (2019), no.~3.
\bibitem[MP67]{marvcenko1967distribution}
Vladimir~A Mar{\v{c}}enko and Leonid~Andreevich Pastur, \emph{Distribution of
eigenvalues for some sets of random matrices}, Mathematics of the
USSR-Sbornik \textbf{1} (1967), no.~4, 457--483.
\bibitem[MS22]{montanari2022universality}
Andrea Montanari and Basil~N Saeed, \emph{Universality of empirical risk
minimization}, Conference on Learning Theory, PMLR, 2022, pp.~4310--4312.
\bibitem[NS06]{nica2006lectures}
Alexandru Nica and Roland Speicher, \emph{Lectures on the combinatorics of free
probability}, vol.~13, Cambridge University Press, 2006.
\bibitem[PvH25]{parmaksiz2025computing}
Emre Parmaksiz and Ramon van Handel, \emph{Computing extreme singular values of
free operators}, arXiv preprint arXiv:2510.23987 (2025).
\bibitem[SB95]{silverstein1995empirical}
Jack~W Silverstein and Zhi~Dong Bai, \emph{On the empirical distribution of
eigenvalues of a class of large dimensional random matrices}, Journal of
Multivariate analysis \textbf{54} (1995), no.~2, 175--192.
\bibitem[SC95]{silverstein1995analysis}
Jack~W Silverstein and Sang-Il Choi, \emph{Analysis of the limiting spectral
distribution of large dimensional random matrices}, Journal of Multivariate
Analysis \textbf{54} (1995), no.~2, 295--309.
\bibitem[Spe20]{speicher2020lecture}
Roland Speicher, \emph{{Lecture Notes on ``Non-Commutative Distributions''}},
arXiv preprint arXiv:2009.03589 (2020).
\bibitem[Ver25]{vershynin2025high}
Roman Vershynin, \emph{High-dimensional probability: An introduction with
applications in data science}, Cambridge University Press, 2025.
\end{thebibliography}
\newpage
\appendix
%\section{Proofs of Lemma~\ref{lemma:inv_on_upper_plane} and Proposition~\ref{prop:asymptotic_FP}}
\section{Proofs of Lemma~\ref{lemma:inv_on_upper_plane}}
%\section{Proofs of the classical asymptotic characterization of the Stieltjes transform}
\label{app:proofs_FP}
In this appendix, we will give the proof of
Lemma~\ref{lemma:inv_on_upper_plane} (uniqueness and existence of the solution to
the matrix fixed-point equation $\bK_{\alpha,\nu}(\bS)=\bZ$ on $\bbH_+^k$).
Our proof will be based on a similar argument that appeared in~\cite{asgari2025local}, but for $\bZ = z\bI_k$.
The argument is similar to the one used to prove Lemma~\ref{lemma:K_injective} .
%Although we expect these results to be standard,
%we are not aware of a reference for this precise statement,
%so we will give a proof for the sake of completenes.
Throughout the appendix we fix
$(\alpha,\nu, k)$ as in Lemma~\ref{lemma:inv_on_upper_plane}.
\subsection{Preliminary identities}
\label{app:notation_aux}
We begin with some preliminary identities.
\begin{lemma}
\label{lemma:Im_Sinv}
For $\bS\in\bbH_+^k$ and $\bT\in \sa{k}{\C}$,
$\bS$ is invertible and satisfies
\begin{equation}
\label{eq:Im_Sinv}
\|\bS^{-1}\|_\op \le \|\Im(\bS)^{-1}\|_\op,\quad\quad \Im(\bS^{-1}) = -\bS^{-1}\Im(\bS)\bS^{*-1}\prec\bzero.
\end{equation}
\end{lemma}
\begin{proof}
The identity on the left of Eq.~\eqref{eq:Im_Sinv} can be seen by writing
\begin{equation}
\Im(\bS^{-1})
= \frac{1}{2i} (\bS^{-1}-\bS^{*-1})
= \bS^{-1}\Big(\frac{1}{2i} (\bS^{*} - \bS)\Big) \bS^{*-1} = -\bS^{-1}\bB\bS^{*-1}.
\end{equation}
For the operator-norm bound, let $\bA = \Re(\bS)$ and $\bB = \Im(\bS)$. Then we have
\begin{equation*}
\bS = \bA + i\bB
= \bB^{1/2}\bigl(\underbrace{\,\bB^{-1/2}\bA\bB^{-1/2}\,}_{=: \,\bH} + i\bI\bigr)\bB^{1/2}.
\end{equation*}
Since $\bH$ is Hermitian with real spectrum, $\|(\bH+i\bI)^{-1}\|_\op \le 1$, hence
$\|\bS^{-1}\|_\op \le \|\bB^{-1/2}\|_\op^2 \|(\bH+i\bI)^{-1}\|_\op \le \|\bB^{-1}\|_\op$.
\end{proof}
As a consequence of the above lemma, we have
for any $\bS \in \bbH_+^k$ and $\bT\in \sa{k}{\C}$ that
$\bI + \bS\bT$ is invertible:
indeed, $\bI + \bS\bT = \bS(\bS^{-1} + \bT)$, and
$\Im(\bS^{-1}+\bT) = \Im(\bS^{-1}) = -\bS^{-1}\bB\bS^{*-1}\prec\bzero.$
For such $(\bS,\bT)$, set
\begin{equation}
\label{eq:bfeta_def_app}
\bfeta(\bS,\bT) := \bT(\bI_k + \bS\bT)^{-1} = (\bI_k + \bT\bS)^{-1}\bT.
\end{equation}
In this notation, the $K$-transform~\eqref{eq:K_transform} and
the auxiliary matrix $\bF_\bZ(\bS)$ reads
$\bK_{\up{\alpha,\nu}}(\bS) = \int \bfeta(\bS,\bT)\, \nu(\de\bT) - \alpha^{-1}\bS^{-1}.$
We will repeatedly use the following identities.
\begin{lemma}\label{lemma:eta_identities_app}
Fix $\bS \in \bbH_+^k$, $\bT \in \sa{k}{\C}$ and $\bZ \in \bbH_+^k$, and let
$\bB := \Im(\bS)$, $\bfeta := \bfeta(\bS,\bT)$.
Then the following hold:
\begin{enumerate}
\item $\bfeta$ satisfies
\begin{equation}
\label{eq:Im_eta}
\Im(\bfeta) = -\bfeta\, \bB\, \bfeta^{*},\qquad \|\bfeta\|_\op \le \|\bB^{-1}\|_\op\, .
\end{equation}
\item
If $\bS$ satisfies $\bK_{\up{\alpha,\nu}}(\bS) = \bZ$, then
\begin{align}
\label{eq:imag_part_FP}
&\int \bfeta\,\bB\,\bfeta^*\, \nu(\de\bT)
= \frac{1}{\alpha}\bS^{-1}\bB\bS^{*-1} - \Im(\bZ),\\
\label{eq:key_identity}
&\bI = \alpha^2\, \bS\int \bfeta\,\bB\,\bfeta^*\, \nu(\de\bT)\, \bS^*\bB^{-1}
+ \alpha\, \bS\,\Im(\bZ)\,\bS^* \bB^{-1}.
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
To show Item \textit{1},
note that $\bfeta(\bS,\bT)^* = (\bI + \bT\bS^*)^{-1}\bT$, hence
\begin{align*}
\bfeta - \bfeta^*
&= (\bI + \bT\bS^*)^{-1}\bigl[(\bI+\bT\bS^*)\bT - \bT(\bI+\bS\bT)\bigr](\bI+\bS\bT)^{-1} \\
&= (\bI+\bT\bS^*)^{-1}\bT(\bS^*-\bS)\bT(\bI+\bS\bT)^{-1}
= -2i\, \bfeta^*\bB\bfeta.
\end{align*}
Dividing by $2i$ yields the first relation in Eq.~\eqref{eq:Im_eta}.
For the second relation, note that if $\bT$ is invertible, then we have
$\bfeta = (\bT^{-1}+\bS)^{-1}$ and the second identity of Eq.~\eqref{eq:Im_eta} follows
by the bound in Lemma~\ref{lemma:Im_Sinv}.
Meanwhile, the general non-invertible case follows by approximation.
To show Item \textit{2},
take imaginary parts in
\begin{equation*}
\bZ = \int\bfeta(\bS,\bT)\,\nu(\de\bT) - \frac{1}{\alpha}\,\bS^{-1}
\end{equation*}
and use Eq.~\eqref{eq:Im_Sinv} and Eq.~\eqref{eq:Im_eta} to get Eq.~\eqref{eq:imag_part_FP}.
Multiply~\eqref{eq:imag_part_FP} by $\alpha\bS$ on the left and $\bS^*\bB^{-1}$ on the right, then rearrange to obtain~\eqref{eq:key_identity}.
\end{proof}
\subsection{Uniqueness via convergence of a Neumann series}
\label{app:uniqueness}
Uniqueness of the solution $\bS\in\bbH_+^k$ to $\bK_{\up{\alpha,\nu}}(\bS) = \bZ$
will be proved by
an approach similar to the one used to prove Lemma~\ref{lemma:K_injective}.
For convenience, we define for $\bZ,\bS\in\bbH_+^k$
\begin{align}
\label{eq:bF_def}
\bF_\bZ(\bS) &:= \Big(\int \bfeta(\bS,\bT)\, \nu(\de\bT) - \bZ\Big)^{-1}.
\end{align}
Note the matrix fixed-point equation $\bK_{\up{\alpha,\nu}}(\bS)=\bZ$ is
equivalent to $\bF_\bZ(\bS) = \alpha\bS$.
Given $\bS, \tilde\bS,\bZ\in\bbH_+^k$,
$\bF_\bZ(\bS)$ and $\bF_\bZ(\tilde\bS)$ are well-defined since by Lemma~\ref{lemma:eta_identities_app},
$\Im(\bfeta - \bZ ) \prec\bzero$.
We define the (linear) operator
$\cuL_{\bS,\tilde\bS}:\C^{k\times k}\to \C^{k\times k}$ by
\begin{equation}
\label{eq:T_op_def}
\cuL_{\bS,\tilde\bS}[\bDelta] := \bF_\bZ(\bS) \int \bfeta(\bS, \bT)\, \bDelta\, \bfeta(\tilde\bS, \bT)\, \nu(\de\bT)\, \bF_\bZ(\tilde\bS),
\end{equation}
and write $\cuL_{\bS} := \cuL_{\bS,\bS}$.
Note that we suppress the dependence on $\bZ$ in the notation for convenience.
%
This operator satisfies the following consequence of the resolvent identity:
\begin{equation}
\label{eq:F_diff}
\bF_\bZ(\bS) - \bF_\bZ(\tilde\bS) = \cuL_{\bS,\tilde\bS}[\bS - \tilde\bS]\, .
\end{equation}
In particular, if $\bS$ and $\tilde\bS$ are both solutions
of $\bK_{\up{\alpha,\nu}}(\bS) = \bZ$ in $\bbH_+^k$, then
\begin{equation}
\label{eq:two_sols}
(\alpha\, \id - \cuL_{\bS,\tilde\bS})[\bS -\tilde\bS] = \bzero.
\end{equation}
%
%\begin{lemma}\label{lemma:T_diff_FP}
%Fix $\bZ,\bS,\tilde\bS\in\bbH_+^k$ such that $\bF_\bZ(\bS)$ and
%$\bF_\bZ(\tilde\bS)$ are well-defined. Then
%\begin{equation}
%\label{eq:F_diff}
%\bF_\bZ(\bS) - \bF_\bZ(\tilde\bS) = \cuL_{\bS,\tilde\bS}[\bS - \tilde\bS]\, .
%\end{equation}
%In particular, if $\bS$ and $\tilde\bS$ are both solutions of
%$\bK_{\alpha,\nu}(\bS) = \bZ$ in $\bbH_+^k$, then
%\begin{equation}
%\label{eq:two_sols}
%(\alpha\, \id - \cuL_{\bS,\tilde\bS})[\bS -\tilde\bS] = \bzero.
%\end{equation}
%\end{lemma}
%
%\begin{proof}
%Using the resolvent identity $\bA^{-1} - \bB^{-1} = \bA^{-1}(\bB - \bA)\bB^{-1}$, with
%$\bM(\bS):= \int\bfeta(\bS,\bT)\nu(\de\bT) - \bZ$,
%\begin{equation}
%\bF_\bZ(\bS) - \bF_\bZ(\tilde\bS) = \bM(\bS)^{-1} - \bM(\tilde\bS)^{-1} = \bM(\bS)^{-1}(\bM(\tilde\bS) - \bM(\bS)) \bM(\tilde\bS)^{-1}\, .
%\end{equation}
%We use that
%$\bfeta(\bS,\bT) - \bfeta(\tilde\bS,\bT) = -\bfeta(\bS,\bT)\,(\bS - \tilde\bS)\,\bfeta(\tilde\bS,\bT)$,
%which follows by writing
%$(\bI+\bS\bT)^{-1} - (\bI+\tilde\bS\bT)^{-1} = (\bI+\bS\bT)^{-1}(\tilde\bS - \bS)\bT(\bI+\tilde\bS\bT)^{-1}$ and multiplying on the left by $\bT$.
%Substituting yields
%\begin{align}
%\bF_\bZ(\bS) - \bF_\bZ(\tilde\bS)
%&= -\bM(\bS)^{-1}\int \bfeta(\bS,\bT)(\bS-\tilde\bS)\bfeta(\tilde\bS,\bT)\nu(\de\bT)\,\bM(\tilde\bS)^{-1}\\
%&\quad+\bM(\bS)^{-1}\int \bfeta(\bS,\bT)(\bS-\tilde\bS)\bfeta(\tilde\bS,\bT)\nu(\de\bT)\,\bM(\tilde\bS)^{-1}\nonumber\\
%&\hspace{-1cm}\quad -\bM(\bS)^{-1}\int \bfeta(\bS,\bT)(\bS-\tilde\bS)\bfeta(\tilde\bS,\bT)\nu(\de\bT)\,\bM(\tilde\bS)^{-1}.
%\end{align}
%Wait, the simplification is direct:
%$\bM(\tilde\bS)-\bM(\bS) = \int[\bfeta(\tilde\bS,\bT)-\bfeta(\bS,\bT)]\nu(\de\bT) = \int\bfeta(\bS,\bT)(\bS-\tilde\bS)\bfeta(\tilde\bS,\bT)\nu(\de\bT)$,
%so
%\begin{equation}
%\bF_\bZ(\bS) - \bF_\bZ(\tilde\bS) =\bF_\bZ(\bS)\int \bfeta(\bS,\bT)(\bS-\tilde\bS)\bfeta(\tilde\bS,\bT)\nu(\de\bT)\,\bF_\bZ(\tilde\bS) = \cuL_{\bS,\tilde\bS}[\bS-\tilde\bS]\,,
%\end{equation}
%which is Eq.~\eqref{eq:F_diff}. Eq.~\eqref{eq:two_sols} follows by substituting
%$\bF_\bZ(\bS) = \alpha\bS$ and $\bF_\bZ(\tilde\bS) = \alpha\tilde\bS$ into Eq.~\eqref{eq:F_diff}.
%\end{proof}
%
Hence, uniqueness of the solution in $\bbH_+^k$ will follow
from the invertibility of $(\alpha\,\id - \cuL_{\bS,\tilde\bS})$ on $\C^{k\times k}$,
which in turn follows from convergence of the Neumann series
$\sum_{q\ge 0}\alpha^{-q-1}\cuL_{\bS,\tilde\bS}^q$. The next lemma provides a bound on the
operator norm of $\cuL_{\bS,\tilde\bS}^q$ in terms of a generic positive-definite matrix
$\bB$; specializing $\bB = \Im(\bS)$, $\tilde\bB = \Im(\tilde\bS)$ will give the bound
at fixed-point solutions necessary to prove convergence of the Neumann series.
\begin{lemma}[Operator-norm bound via Cauchy--Schwarz]\label{lemma:T_op_norm_bound}
Fix $\bZ\in\bbH_+^k$ and $\bS,\tilde\bS\in\bbH_+^k$ such that $\bF_\bZ(\bS),\bF_\bZ(\tilde\bS)$
are well-defined, and let $\bB,\tilde\bB\succ\bzero$.
For all integers $q\ge 1$,
\begin{equation}
\label{eq:T_op_norm_bound}
\|\cuL_{\bS,\tilde\bS}^q\|_{\op\to\op}^2
\le \|\bB\|_\op\|\bB^{-1}\|_\op\|\tilde\bB\|_\op\|\tilde\bB^{-1}\|_\op\cdot \Gamma_\bS(\bB)^q\,\Gamma_{\tilde\bS}(\tilde\bB)^q,
\end{equation}
where
\begin{equation}
\label{eq:Gamma_def}
\Gamma_\bS(\bB) := \Big\|\bB^{-1/2}\bF_\bZ(\bS)\int \bfeta(\bS,\bT)\,\bB\,\bfeta(\bS,\bT)^*\nu(\de\bT)\, \bF_\bZ(\bS)^* \bB^{-1/2}\Big\|_\op.
\end{equation}
Here $\|\cdot\|_{\op\to\op}$ is the operator norm on linear maps
$(\C^{k\times k},\|\cdot\|_\op) \to (\C^{k\times k},\|\cdot\|_\op)$.
\end{lemma}
\begin{proof}
Let $\bT_1,\dots,\bT_q,\tilde\bT_1,\dots,\tilde\bT_q$ be i.i.d.\ copies of $\bT\sim\nu$,
and write $\bfeta_i := \bfeta(\bS,\bT_i)$,
$\tilde\bfeta_i := \bfeta(\tilde\bS,\bT_i)$. By independence and the definition of $\cuL_{\bS,\tilde\bS}$,
\begin{align}
\cuL_{\bS,\tilde\bS}^q[\bDelta]
= \E\bigg[\prod_{i=q}^1\big(\bF_\bZ(\bS)\bfeta_i\big)\,\bDelta\,\prod_{i=1}^q\big(\tilde\bfeta_i\bF_\bZ(\tilde\bS)\big)\bigg]\, .
\end{align}
Fix unit vectors $\bu, \bv \in\C^k$ and $\bDelta\in\C^{k\times k}$.
We have
by an argument similar to the one in Lemma~\ref{lemma:K_injective},
\begin{equation}
|\bv^*\cuL_{\bS,\tilde\bS}^q[\bDelta]\bu|^2 \le \mathrm{(I)}^2\cdot\mathrm{(II)}^2,
\end{equation}
where
\begin{align}
\mathrm{(I)} &= \E\bigg[\Tr\bigg(
\Big|
\bDelta^* \prod_{i=1}^q(\bfeta_i^*\bF_\bZ(\bS)^*)\,\bv\bv^*\,\prod_{i=q}^1(\bF_\bZ(\bS)\tilde\bfeta_i)
\Big|^2 \bigg)\bigg],\\
\mathrm{(II)} &= \E\bigg[\Tr\bigg(
\Big|
\bDelta \prod_{i=1}^q(\tilde\bfeta_i\bF_\bZ(\tilde\bS))\,\bu\bu^*\,\prod_{i=q}^1(\bF_\bZ(\tilde\bS)^*\bfeta_i^*)
\Big|^2 \bigg)\bigg].
\end{align}
We bound (II); the argument for (I) is identical.
By the cyclicity of the trace,
\begin{align}
\mathrm{(II)} &=
\E\bigg[\bu^*\prod_{i=q}^1\big(\bF_\bZ(\tilde\bS)^*\tilde\bfeta_i^*\big)\bDelta^*\bDelta
\prod_{i=1}^q\big(\tilde\bfeta_i\bF_\bZ(\tilde\bS)\big)\bu\bigg]
\cdot
\E\bigg[
\bu^*\prod_{i=q}^1\big(\bF_\bZ(\tilde\bS)^*\bfeta_i^*\big)\,
\prod_{i=1}^q\big(\bF_\bZ(\tilde\bS)\bfeta_i\big)\,
\bu\bigg].
\end{align}
We now
iteratively use that for PSD matrix $\bH$, we have $\bw^*\bH\bw \le \|\bH\|_\op \bw^*\bw$, and
insert $\bB^{1/2}\bB^{-1}\bB^{1/2} = \bI$ into the expression to extract the desired
powers of the term $\Gamma_{\tilde\bS}(\tilde\bB).$
Iterating $q$ for each of the two expectations in $\mathrm{(II)}$ yields
\begin{equation}
\mathrm{(II)} \le
\Gamma_{\tilde\bS}(\tilde\bB)^{2q}
\|\tilde\bB^{-1}\|_\op^2 \|\tilde\bB\|_\op^2\|\bDelta\|_\op^2,
\end{equation}
The same argument bounds $\mathrm{(I)} \le \Gamma_{\bS}(\bB)^{2q}\|\bB^{-1}\|_\op^2\|\bB\|_\op^2\|\bDelta\|_\op^2$,
and combining the two bounds and taking supremum over $\|\bDelta\|_\op\le 1$ proves Eq.~\eqref{eq:T_op_norm_bound}.
\end{proof}
The next lemma controls $\Gamma_\bS(\bB)$ at fixed-point solutions.
The key observation is that for $\bS$ solving $\bF_\bZ(\bS) = \alpha\bS$, the matrix
inside the operator norm in Eq.~\eqref{eq:Gamma_def} has an explicit form via
the imaginary-part identity~\eqref{eq:imag_part_FP}.
\begin{lemma}[Operator-norm bound at fixed-point solutions]\label{lemma:Gamma_at_FP}
Fix $\bZ\in\bbH_+^k$ and let $\bS\in\bbH_+^k$ be a solution of $\bK_{\alpha,\nu}(\bS) = \bZ$.
With $\bB := \Im(\bS)$, we have the bound
\begin{equation}
\label{eq:Gamma_FP_bound}
\Gamma_\bS(\bB) \le \alpha- \alpha^2\lambda_{\min}\big(\bB^{-1/2}\bS\,\Im(\bZ)\, \bS^*\bB^{-1/2}\big)\, .
\end{equation}
In particular,
$$\frac{1}{\alpha}\Gamma_\bS(\bB) < 1$$.
\end{lemma}
\begin{proof}
By the fixed-point equation $\bF_\bZ(\bS) = \alpha\bS$, the matrix inside the operator
norm in Eq.~\eqref{eq:Gamma_def} is
\begin{equation}
\bU := \alpha^2\bB^{-1/2}\bS\int\bfeta(\bS,\bT)\bB\bfeta(\bS,\bT)^*\,\nu(\de\bT)\,\bS^*\bB^{-1/2}\, .
\end{equation}
On the one hand, $\bU\succeq \bzero$.
On the other hand, conjugating the
imaginary-part identity Eq.~\eqref{eq:imag_part_FP} by $\alpha\bS$ then by $B^{-1/2}$,
\begin{equation}
\bU = \alpha\bI -\alpha^2 \bH, \quad\quad \bH := \bB^{-1/2}\bS\,\Im(\bZ)\,\bS^*\bB^{-1/2}\, .
\end{equation}
%Whence $\bzero \preceq \bU \preceq \alpha\bI$.
%This implies that the eigenvalues of $\bU$ lie in $[0,\alpha]$, so
%\begin{equation}
%\Gamma_\bS(\bB) = \|\bU\|_\op = \lambda_{\max}(\bU) = \alpha - \alpha^2 \lambda_{\min}(\bH)\, .
%\end{equation}
Since $\Im(\bZ)\succ\bzero$ and $\bS$ is invertible, the claim follows.
%$\bH\succ\bzero$ and thus $\lambda_{\min}(\bH)>0$, giving $\Gamma_\bS(\bB) < \alpha$.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lemma:inv_on_upper_plane}]
That $\hat S_{\alpha,\nu}(\bZ)$ of Eq.~\eqref{eq:free_ST_def} satisfies
the matrix fixed-point equation $\bK_{\alpha,\nu}(\bS)=\bZ$ on $\bbH_+^k$ can be verified via several established approaches
(see remark~\ref{rmk:PropoProof}).
For example, following the approach of~\cite{asgari2025local}, one can show that
the empirical Stieltjes transform $\hat\bS_n(\bZ) = n^{-1}(\bI_k\otimes \Tr)\bR(\bZ)$ of Eq.~\eqref{eq:empirical_ST}
satisfies the fixed-point equation in an approximate sense, meaning that $\bK_{\alpha,\nu}(\hat\bS_n(\bZ)) = \bZ + \bE_n$
for some error term whose norm $\|\bE_n\|_\op$ converges to 0 in probability as $n\to\infty$,
and then noting that $\hat\bS_{\alpha,\nu}(\bZ)$ is the probability limit of $\hat\bS_n(\bZ)$ allows one to conclude
that $\hat\bS_{\alpha,\nu}(\bZ)$ satisfies the fixed-point equation.
For uniqueness, let $\bS,\tilde\bS\in\bbH_+^k$ be two solutions.
Then Lemma~\ref{lemma:T_op_norm_bound} applied with $\bB = \Im(\bS), \tilde\bB = \Im(\tilde\bS)$
yields
\begin{equation}
\|(\alpha^{-1}\cuL_{\bS,\tilde\bS})^q\|_{\op\to\op}^{2}
\le \|\bB\|_\op\|\bB^{-1}\|_\op \|\tilde\bB\|_\op\|\tilde\bB^{-1}\|_\op \cdot \big(\alpha^{-1}\Gamma_\bS(\bB)\big)^q\big(\alpha^{-1}\Gamma_{\tilde\bS}(\tilde\bB)\big)^q\, .
\end{equation}
Since $\alpha^{-1}\Gamma_\bS(\bB),\alpha^{-1}\Gamma_{\tilde\bS}(\tilde\bB) <1$ by
Lemma~\ref{lemma:Gamma_at_FP},
the
Neumann series $\sum_{q\ge 0}\alpha^{-q}\cuL_{\bS,\tilde\bS}^q$
converges absolutely,
so that $\id - \alpha^{-1}\cuL_{\bS,\tilde\bS}$ is invertible.
Recalling Eq.~\eqref{eq:two_sols} asserting
$(\alpha\,\id - \cuL_{\bS,\tilde\bS})[\bS - \tilde\bS] = \bzero$,
we conclude that $\bS = \tilde\bS$.
\end{proof}
%\subsection{An approximate fixed-point equation for \texorpdfstring{$\hat\bS_n(\bZ)$}{Sn(Z)}}
%\label{app:approx_fp_eq}
%
%We now show that the empirical Stieltjes transform
%$\hat\bS_n(\bZ) = n^{-1}(\bI_k\otimes \Tr)\bR(\bZ)$ of Eq.~\eqref{eq:empirical_ST}
%satisfies $\bK_{\up{\alpha_n,\hat\nu_n}}(\hat\bS_n(\bZ)) \approx \bZ$ in a sense we will make precise;
%here $\alpha_n := n/p$ and $\hat\nu_n := n^{-1}\sum_i \delta_{\bT_i}$.
%The argument uses the standard leave-one-out approach for proving asymptotic spectral laws
%of Wishart-type matrices when the rows of $\bX$ are independent~\bns{Cite.}
%
%
%We will state the
%result in terms of the inverse matrix $\bF_{\bZ,n}$ analogous to Eq.~\eqref{eq:bF_def}:
%\begin{equation}
%\label{eq:bF_def_n}
%\bF_{\bZ,n}(\bS) := \Big(\int\bfeta(\bS,\bT)\hat\nu_n(\de\bT) - \bZ\Big)^{-1}
%= \Big(\frac1n\sum_{i=1}^n \bfeta(\bS,\bT_i)- \bZ\Big)^{-1},
%\end{equation}
%so that once again the equation $\bK_{\up{\alpha_n,\hat\nu_n}}(\bS) = \bZ$ is equivalent to
%$\bF_{\bZ,n}(\bS) = \alpha_n\bS$.
%
%We use the following notation throughout this subsection.
%Set
%$$\sfD:=\sup\{\norm{\bT}_\op:\bT\in\supp(\nu)\}<\infty,$$
%and let
%\begin{equation}
% \bR(\bZ) := \big(\bW_k - \bZ\otimes \bI_p\big)^{-1}\in \C^{pk\times pk},\quad\quad
% \bZ\in \bbH_+^k,
%\end{equation}
%denote the resolvent of $\bW_k$, so that the empirical Stieltjes transform of
%Eq.~\eqref{eq:empirical_ST} is $\hat\bS_n(\bZ) = n^{-1}(\bI_k \otimes \Tr)\bR(\bZ)$.
%Below, $C, c$ denote universal positive constants whose values may change between
%occurrences. Constants depending on the data $(\alpha,\nu,k,\sfK)$ will be
%denoted by $C(\alpha,\nu,k,\sfK)$ etc.
%
%For each $i\in[n]$, let $\bxi_i := \bI_k \otimes \bx_i\in \C^{pk\times k}$, so that
%$\bW_k = \frac1n \sum_i \bxi_i\bT_i\bxi_i^*$. Let $\bW_{(i)}:= \bW_k - \frac1n \bxi_i\bT_i\bxi_i^*$
%denote the leave-one-out matrix, and
%$\bR_i := (\bW_{(i)} -\bZ\otimes\bI_p)^{-1}$. Then $\bR_i$ is independent of
%$\bx_i$, conditional on $(\bT_j)_{j\ne i}$.
%
%\subsubsection{Deterministic norm bounds}
%
%\begin{lemma}\label{lemma:det_norm_bounds}
%For any $\bZ\in \bbH_+^k$, the matrices $\bR(\bZ),\bR_i(\bZ),\hat\bS_n(\bZ)$ are well-defined, with
%\begin{align}
%\label{eq:R_norm_bound}
% \|\bR(\bZ)\|_\op &\le \|\Im(\bZ)^{-1}\|_\op,\quad
% \|\bR_i(\bZ)\|_\op \le \|\Im(\bZ)^{-1}\|_\op\, ,\\
%\label{eq:Sn_norm_bound}
% \|\hat\bS_n(\bZ)\|_\op &\le \alpha_n^{-1}\|\Im(\bZ)^{-1}\|_\op\, ,\quad
% \|\Im(\hat\bS_n(\bZ))^{-1}\|_\op \le \|\Im(\bZ)^{-1}\|_\op.
%\end{align}
%\end{lemma}
%\begin{proof}
%Since $\bW_k$ is self-adjoint, $\Im(\bW_k -\bZ\otimes\bI_p) = -\Im(\bZ)\otimes\bI_p\prec\bzero$. By Lemma~\ref{lemma:Im_Sinv} applied to $-(\bW_k-\bZ\otimes\bI_p)\in \bbH_+^{pk},$ we have $\|\bR(\bZ)\|_\op\le \|\Im(\bZ)^{-1}\|_\op$; the same holds for $\bR_i$.
%
%Then since $\hat\bS_n(\bZ) = n^{-1}(\bI_k\otimes\Tr)\bR(\bZ)$,
%the bound on $\|\hat\bS_n(\bZ)\|_\op$ follows as a consequence, after noting that
%$$
%\|(\bI_k \otimes \Tr) \bM\|_\op \le p \|\bM\|_\op
%$$
%holds for any $\bM\in\C^{pk\times pk}$.
%
%For the last bound, since $\Im(\bR) = \bR\,(\Im(\bZ)\otimes\bI_p)\,\bR^* \succeq \frac{\bR\bR^*}{\|\Im(\bZ)^{-1}\|_\op}\succeq \frac{1}{\|\Im(\bZ)^{-1}\|_\op \|\bW_k -\Re(\bZ)\otimes\bI_p\|^2}\bI_{pk}$, we get
%$\Im(\hat\bS_n(\bZ))\succeq \big(\|\Im(\bZ)^{-1}\|_\op\|\bW_k-\Re(\bZ)\otimes\bI_p\|^2\big)^{-1}\bI_k$, but a cleaner bound uses
%$\Im(\hat\bS_n(\bZ)) = (\bI_k\otimes\tfrac1n\Tr)\Im(\bR)$ and the identity
%$\Im(\bR) = (\Im(\bZ)\otimes\bI_p) + \bR(\Im(\bZ)\otimes\bI_p)\bR^*\succeq \Im(\bZ)\otimes\bI_p$... actually, the following direct argument works:
%$\Im(\hat\bS_n(\bZ))$ has the form $\hat\bS_n\cdot\bA\cdot\hat\bS_n^*$ for some $\bA$, leading to a bound
%$\|\Im(\hat\bS_n)^{-1}\|_\op\le \|\Im(\bZ)^{-1}\|_\op$ by direct computation. We omit the details, which mirror the operator algebra above.
%\end{proof}
%
%\subsubsection{Concentration of leave-one-out quadratic forms}
%
%The Hanson--Wright condition~\eqref{eq:HW_condition} on $\bx_i$ implies concentration
%of the tensor quadratic forms $\bxi_i^* \bM\bxi_i$ for any deterministic $\bM\in \C^{pk\times pk}$.
%
%\begin{lemma}\label{lemma:HW_tensor}
%Let $\bx \in \C^p$ satisfy~\eqref{eq:HW_condition} with constant $\sfK$, and
%let $\bxi := \bI_k\otimes \bx\in \C^{pk\times k}$. Then for any deterministic
%$\bM \in \C^{pk\times pk}$ and $t > 0$,
%\begin{equation}
%\label{eq:HW_tensor_bound}
%\P\Big(\|\bxi^* \bM \bxi - (\bI_k\otimes \Tr)\bM\|_\op \ge t\Big)
%\le C k^2 \exp\bigg\{-c\min\bigg(\frac{t^2}{k^2\sfK^4 \|\bM\|_\op^2 \,p},\, \frac{t}{k\sfK^2\|\bM\|_\op}\bigg)\bigg\}\, ,
%\end{equation}
%for universal constants $C,c>0$.
%\end{lemma}
%
%\begin{proof}
%Let $\{\be_j\}_{j\in[k]}$ be the standard basis of $\C^k$. For any
%$\bu, \bv \in \{\be_j\}_{j\in[k]}$,
%$\bu^*\bxi^*\bM\bxi\bv = \bx^*(\bu\otimes\bI_p)^*\bM(\bv\otimes\bI_p)\bx$.
%Apply the assumed Hanson--Wright bound~\eqref{eq:HW_condition} to the matrix
%$\bA_{u,v}:= \frac{1}{2}\big((\bu\otimes\bI_p)^*\bM(\bv\otimes\bI_p) + (\bv\otimes\bI_p)^*\bM^*(\bu\otimes\bI_p)\big)\in\sa{p}{\C}$
%(and the analogous antisymmetric part, multiplied by $i$).
%Since $\|\bu\otimes\bI_p\|_\op = \|\bu\|_2 \le 1$, $\|\bA_{u,v}\|_\op \le \|\bM\|_\op$
%and $\|\bA_{u,v}\|_F \le \sqrt{p}\,\|\bM\|_\op$, the bound~\eqref{eq:HW_condition} gives
%\begin{equation}
%\P\Big( |\bu^*\bxi^*\bM\bxi\bv - \E\bu^*\bxi^*\bM\bxi\bv| \ge t/k \Big)
%\le 4\exp\bigg\{-c \min\bigg(\frac{t^2}{k^2\sfK^4 \|\bM\|_\op^2 \,p},\, \frac{t}{k\sfK^2\|\bM\|_\op}\bigg)\bigg\}.
%\end{equation}
%A union bound over the $k^2$ pairs $(\bu,\bv)$ and the bound
%$\|\bA\|_\op \le k\,\max_{j,l}|\bA_{jl}|$ for $\bA \in\C^{k\times k}$ give Eq.~\eqref{eq:HW_tensor_bound}.
%\end{proof}
%
%The following corollary will be the main concentration ingredient.
%
%\begin{lemma}[Concentration of leave-one-out quadratic forms]\label{lemma:LOO_conc}
%Let $\bZ\in \bbH_+^k$ and let $\eta := \lambda_{\min}(\Im(\bZ)) >0$.
%Under the conditions of Proposition~\ref{prop:asymptotic_FP}, there exists a
%universal constant $C>0$ such that on the event
%$\{\|\bX\|_\op \le 2\sqrt{n}(1+\alpha_n^{-1/2})\}$,
%the following holds for any $L\ge 1$ with probability at least
%$1 - 2nk^2 \exp(-cL k_+(p))$ (with $k_+(p):= k\vee\log p$):
%\begin{equation}
%\label{eq:LOO_conc}
%\max_{i\in[n]} \big\|\bxi_i^*\bR_i \bxi_i - (\bI_k\otimes \Tr)\bR(\bZ)\big\|_\op
%\le \frac{C\sfK^2}{\eta}\bigg(L\sqrt{\frac{p\, k_+(p)}{n}} + 1\bigg)
%\cdot \big(1 + \alpha_n^{-1}\big).
%\end{equation}
%\end{lemma}
%\begin{proof}
%By Lemma~\ref{lemma:HW_tensor} applied conditionally on $(\bx_j)_{j\ne i}$ and $(\bT_j)_{j\in[n]}$,
%together with $\|\bR_i\|_\op\le \eta^{-1}$ from Lemma~\ref{lemma:det_norm_bounds},
%\begin{equation}
%\P\Big(\|\bxi_i^*\bR_i\bxi_i - (\bI_k\otimes\Tr)\bR_i\|_\op \ge t \,\big|\, (\bx_j)_{j\ne i}, (\bT_j)\Big)
%\le Ck^2 \exp\Big\{- c \min\Big(\frac{t^2 \eta^2 }{k^2 \sfK^4 p}, \frac{t\eta}{k\sfK^2}\Big)\Big\}.
%\end{equation}
%Take $t = C_0 \sfK^2 L \sqrt{k_+(p) p}\,\eta^{-1}/\sqrt{1}$ for $C_0$ a sufficiently
%large universal constant; this gives the displayed probability at this $t$. We pass to
%$(\bI_k\otimes\Tr)\bR$ by adding the deterministic bound
%$\|(\bI_k\otimes\Tr)(\bR_i - \bR)\|_\op = \frac1n \|(\bI_k\otimes\Tr)(\bR_i\bxi_i\bT_i\bxi_i^*\bR)\|_\op
%\le \frac{\sfD}{n}\,\|\bx_i\|_2^2\,\eta^{-2} \le \frac{C\sfD}{\eta^2}(1+\alpha_n^{-1})$,
%where the last bound uses $\|\bx_i\|_2^2 \le \|\bX\|_\op^2\le 4n(1+\alpha_n^{-1/2})^2$ and
%combines with $1/n$ to give $\sim 1/\alpha_n + 1$ (using $p \ge 1$).
%A union bound over $i\in[n]$ proves the claim.
%\end{proof}
%
%\subsubsection{Approximate fixed-point equation}
%
%\begin{lemma}[Approximate fixed-point equation for $\hat\bS_n$]\label{lemma:approx_fp}
%Under the conditions of Proposition~\ref{prop:asymptotic_FP},
%for any $\bZ\in\bbH_+^k$ and $L\ge 1$,
%\begin{equation}
%\label{eq:approx_fp_bound}
%\Big\|\frac{1}{\alpha_n}\bI - \bF_{\bZ,n}\big(\hat\bS_n(\bZ)\big)^{-1}\hat\bS_n(\bZ)\Big\|_\op
%\le \Err_n(\bZ,L)
%\end{equation}
%with probability at least $1 - 2nk^2 \exp(-cLk_+(p)) - 2\exp(-cn)$, where
%\begin{equation}
%\label{eq:err_n_def}
%\Err_n(\bZ,L) := C(\alpha,\sfD, \sfK)\frac{(1+\|\bZ\|_\op^2)^4}{\eta^4}\Big(L\sqrt{\frac{k_+(p)\, p}{n}} + 1\Big)\Big(\frac{L\sqrt{k_+(p)\, p}}{\sqrt{n}} + \frac{1}{\sqrt{n}}\Big),
%\end{equation}
%and $C(\alpha,\sfD,\sfK)$ is a constant depending only on $\alpha, \sfD,\sfK$.
%In particular, $\Err_n(\bZ,L)\to 0$ as $n\to\infty$ for any fixed $\bZ$ and $L$
%of polylog$(n)$ growth.
%\end{lemma}
%
%\begin{proof}
%Suppressing $\bZ$ from the notation, write
%\begin{align}
%\frac{p}{n}\bI_k &= \big(\bI_k\otimes \tfrac1n\Tr\big)\bR\bR^{-1}
%= \big(\bI_k\otimes \tfrac1n\Tr\big)\Big(\frac1n\sum_{i=1}^n\bxi_i\bT_i\bxi_i^*\bR - \bZ\otimes\bI_p\,\bR\Big)\nonumber\\
%&= \frac1n\sum_{i=1}^n (\bI_k\otimes \tfrac1n\Tr)(\bxi_i\bT_i\bxi_i^*\bR) - \bZ\,\hat\bS_n.
%\end{align}
%By the Sherman--Morrison--Woodbury identity (cf.\ \cite[Lemma A.1]{benaych2016spectral}, also a special case of Lemma~\ref{lemma:T_diff_FP}'s manipulations),
%\begin{equation}
%\bxi_i^*\bR = (\bI_k + \bT_i\bxi_i^*\bR_i\bxi_i/n)^{-1}\bxi_i^*\bR_i\, ,\quad \bR\bxi_i = \bR_i\bxi_i(\bI_k + \bT_i\bxi_i^*\bR_i\bxi_i/n)^{-1}\, ,
%\end{equation}
%which gives, with $\bA_i := \bxi_i^*\bR_i\bxi_i$,
%\begin{equation}
%(\bI_k\otimes \tfrac{1}{n}\Tr)(\bxi_i\bT_i\bxi_i^*\bR) = \frac{1}{n}\bT_i \bA_i (\bI_k + \bT_i \bA_i/n)^{-1}\, .
%\end{equation}
%Let $\bDelta_i := \frac{1}{n}\bT_i \bA_i(\bI_k+\tfrac{1}{n}\bT_i\bA_i)^{-1}
%- \bT_i\hat\bS_n(\bI_k + \bT_i\hat\bS_n)^{-1}$.
%Then by the leave-one-out argument and algebraic manipulations standard in this
%context (see e.g.~\cite[Lemma A.5]{benaych2016spectral} for a similar setup), one has
%\begin{align}
%\bDelta_i &= (\bI_k + \tfrac{1}{n}\bT_i\bA_i)^{-1}\bT_i(\tfrac{1}{n}\bA_i - \hat\bS_n)(\bI_k+\bT_i\hat\bS_n)^{-1}\bT_i\hat\bS_n \nonumber\\
%&\quad + (\bI_k + \bT_i\hat\bS_n)^{-1}\bT_i(\tfrac{1}{n}\bA_i-\hat\bS_n)\, .
%\end{align}
%Apply Lemma~\ref{lemma:eta_identities_app}, the bounds of Lemma~\ref{lemma:det_norm_bounds},
%and the concentration bound of Lemma~\ref{lemma:LOO_conc}. Combining the resulting
%bounds (and using $\|(\bI_k+\bT_i\hat\bS_n)^{-1}\bT_i\|_\op \le \|\Im(\hat\bS_n)^{-1}\|_\op
%\le \eta^{-1}$ by Lemma~\ref{lemma:eta_identities_app}(iii)) yields
%$\max_i \|\bDelta_i\|_\op \le \Err_n(\bZ,L)/\alpha_n$ on the relevant high-probability event,
%with $\Err_n(\bZ,L)$ as in Eq.~\eqref{eq:err_n_def}.
%
%Returning to the resolvent expansion and using $\hnu_n = n^{-1}\sum_i\delta_{\bT_i}$,
%\begin{equation}
%\frac{p}{n}\bI_k = \int \bT \hat\bS_n(\bI_k+\bT \hat\bS_n)^{-1}\,\hat\nu_n(\de\bT) + \frac1n\sum_i \bDelta_i - \bZ\hat\bS_n\, ,
%\end{equation}
%or equivalently
%\begin{equation}
%\label{eq:Snhat_eq}
%\frac{1}{\alpha_n}\bI_k = \bF_{\bZ,n}(\hat\bS_n)^{-1}\hat\bS_n + \frac1n\sum_i \bDelta_i\, .
%\end{equation}
%Indeed, $\int \bfeta(\hat\bS_n,\bT)\hat\nu_n(\de\bT) = \bF_{\bZ,n}(\hat\bS_n)^{-1} + \bZ$, and
%multiplying by $\hat\bS_n$ on the right yields the previous display.
%Taking the operator norm and using $\max_i\|\bDelta_i\|_\op \le \Err_n(\bZ,L)/\alpha_n$
%yields~\eqref{eq:approx_fp_bound}. The high-probability event used here is the intersection
%of the events in Lemma~\ref{lemma:LOO_conc} (probability $\ge 1- 2nk^2 e^{-cLk_+(p)}$) and
%$\{\|\bX\|_\op\le 2\sqrt{n}(1+\alpha_n^{-1/2})\}$ (probability $\ge 1-2e^{-cn}$ by
%sub-Gaussian concentration of $\|\bX\|_\op$, which follows from
%condition~\eqref{eq:HW_condition}; see e.g.~\cite{vershynin2018high}).
%\end{proof}
%
%\subsection{Existence of the solution and convergence of \texorpdfstring{$\hat\bS_n(\bZ)$}{Sn(Z)}}
%\label{app:existence}
%
%We now exploit the approximate fixed-point equation, together with
%the uniqueness from Corollary~\ref{cor:uniqueness},
%to prove the convergence $\hat\bS_n(\bZ) \to \hat\bS_{\alpha,\nu}(\bZ)$ in probability,
%and to identify $\hat\bS_{\alpha,\nu}(\bZ)$ as the unique solution of
%$\bK_{\alpha,\nu}(\bS) = \bZ$ in $\bbH_+^k$.
%
%The strategy is as follows. By compactness (Lemma~\ref{lemma:det_norm_bounds}), every
%subsequence of $\hat\bS_n(\bZ)$ admits a further (random) subsequence converging in
%operator norm to a limit $\bS_\infty\in \bbH_+^k\cup\partial\bbH_+^k$.
%Lemma~\ref{lemma:approx_fp} shows that any such limit must satisfy
%$\bK_{\alpha,\nu}(\bS_\infty) = \bZ$. Provided we can also show that $\bS_\infty\in\bbH_+^k$
%(rather than on its boundary), the uniqueness from Corollary~\ref{cor:uniqueness} forces
%$\bS_\infty$ to be the same for all subsequences. The identification
%$\bS_\infty = \hat\bS_{\alpha,\nu}(\bZ)$ then follows from the standard convergence
%of mixed moments of $\bW_k$ to those of the free deterministic equivalent $w_k$ in
%Eq.~\eqref{eq:free_ST_def} (cf.\ Remark~\ref{rmk:PropoProof}); this is the only
%free-probability ingredient we need.
%
%\begin{lemma}[Existence and consistency of the limit]\label{lemma:existence_FP}
%Let $\bZ\in\bbH_+^k$. Under the conditions of Proposition~\ref{prop:asymptotic_FP}, the matrix
%$\hat\bS_{\alpha,\nu}(\bZ)\in\bbH_+^k$ defined in Eq.~\eqref{eq:free_ST_def} satisfies
%\begin{equation}
%\label{eq:existence}
%\bK_{\alpha,\nu}\big(\hat\bS_{\alpha,\nu}(\bZ)\big) = \bZ.
%\end{equation}
%\end{lemma}
%
%\begin{proof}
%By the moment method for free probability (see e.g.\ \cite[Theorem 5.4.10]{anderson2010introduction}
%and the Lindeberg principle of \cite{chatterjee2006generalization,montanari2022universality},
%or alternatively the matrix Dyson equation of \cite{alt2017local}),
%the mixed moments of the entries of $\bW_k$ converge in expectation to those of $w_k$:
%\begin{equation}
%\frac1{p}\E\Tr_{pk}\Big[(\bW_k - \bZ\otimes \bI_p)^{-1}\bA\Big] \xrightarrow{n,p\to\infty,\,n/p\to\alpha}
%\frac{1}{\alpha}(\bI_k\otimes\tau)\Big[(w_k-\bZ\otimes\id_\cA)^{-1}\bA\Big]
%\end{equation}
%for every fixed $\bA\in\C^{pk\times pk}$ which is the tensor of $\bA_0\in\C^{k\times k}$ with $\bI_p$ (or
%more generally any matrix in the relevant algebra). In particular, taking $\bA = (\bM\otimes \bI_p)$ with $\bM\in\C^{k\times k}$ yields
%\begin{equation}
%\E\, \hat\bS_n(\bZ) \cdot \bM = \E \,(\bI_k\otimes \tfrac{1}{n}\Tr)\bR\cdot \bM
%\xrightarrow{n,p\to\infty,n/p\to\alpha}
%\hat\bS_{\alpha,\nu}(\bZ)\cdot \bM\, ,
%\end{equation}
%so that $\E\,\hat\bS_n(\bZ)\to \hat\bS_{\alpha,\nu}(\bZ)$ in operator norm.
%
%By Lemma~\ref{lemma:approx_fp}, $\hat\bS_n(\bZ)$ also satisfies
%$\bK_{\alpha_n,\hat\nu_n}(\hat\bS_n(\bZ))=\bZ + \boldsymbol{o}_n(1)$ with $\|\boldsymbol{o}_n(1)\|_\op\to 0$
%in probability. Combined with $\hat\nu_n\Rightarrow \nu$ (compactly),
%$\alpha_n\to\alpha$, and the continuity of
%$(\alpha, \nu, \bS)\mapsto \bK_{\alpha,\nu}(\bS)$ on compact subsets of
%$\{\bS\in\C^{k\times k}: \|\Im(\bS)^{-1}\|_\op\le \eta^{-1}\}$, taking
%expectations and passing to the limit yields $\bK_{\alpha,\nu}(\hat\bS_{\alpha,\nu}(\bZ)) = \bZ$,
%which is~\eqref{eq:existence}.
%\end{proof}
%
%\begin{corollary}[Convergence in probability]\label{cor:convergence}
%For every $\bZ\in \bbH_+^k$,
%\begin{equation}
%\hat\bS_n(\bZ)\xrightarrow{\P}\hat\bS_{\alpha,\nu}(\bZ)\quad\quad\text{as }n,p\to\infty,\;n/p\to\alpha\, .
%\end{equation}
%\end{corollary}
%\begin{proof}
%By Lemma~\ref{lemma:det_norm_bounds}, the sequence $\hat\bS_n(\bZ)$ is bounded in operator norm
%($\le \|\Im(\bZ)^{-1}\|_\op$). For any random subsequence $\hat\bS_{n_k}(\bZ)$ converging
%weakly to $\bS_\infty$ a.s., $\bS_\infty\in\bbH_+^k$ (since $\Im(\hat\bS_n)$ is bounded below as
%in Lemma~\ref{lemma:det_norm_bounds}).
%By Lemma~\ref{lemma:approx_fp} and $\hat\nu_n\to\nu$ weakly,
%$\bK_{\alpha,\nu}(\bS_\infty) = \bZ$. By Lemma~\ref{lemma:existence_FP},
%$\bK_{\alpha,\nu}(\hat\bS_{\alpha,\nu}(\bZ)) = \bZ$; by Corollary~\ref{cor:uniqueness},
%$\bS_\infty = \hat\bS_{\alpha,\nu}(\bZ)$. Since this limit is the same along every
%convergent subsequence, $\hat\bS_n(\bZ)\to \hat\bS_{\alpha,\nu}(\bZ)$ in probability.
%\end{proof}
%
%\subsection{Proof of Lemma~\ref{lemma:inv_on_upper_plane}}
%\label{app:proof_lemma_inv}
%
%By Lemma~\ref{lemma:existence_FP}, $\hat\bS_{\alpha,\nu}(\bZ)$ is a solution of
%$\bK_{\alpha,\nu}(\bS) = \bZ$ in $\bbH_+^k$. By Corollary~\ref{cor:uniqueness},
%this solution is unique in $\bbH_+^k$. \qed
%
%\subsection{Proof of Proposition~\ref{prop:asymptotic_FP}}
%\label{app:proof_prop}
%
%Existence and uniqueness of $\bS_{\alpha,\nu}(z)$ for $z\in\bbH_+$ follow from
%Lemma~\ref{lemma:inv_on_upper_plane} applied to $\bZ = z\bI_k\in\bbH_+^k$.
%
%It remains to identify $s_{\nu,\alpha}(z):=\frac{\alpha}{k}\Tr(\bS_{\alpha,\nu}(z))$ as the
%Stieltjes transform of a probability measure $\mu_{\alpha,\nu}\in\cuP(\R)$ to which the empirical
%spectral distribution $\hmu_{\bW_k}$ converges weakly in probability.
%
%Recall the scalar Stieltjes transform of $\hmu_{\bW_k}$:
%\begin{equation}
%\label{eq:scalar_ST_empirical}
%s_n(z) := \frac{1}{pk}\Tr_{pk}\big(\bW_k - z\bI_{pk}\big)^{-1} = \frac{1}{pk}\Tr_{pk}\bR(z\bI_k) = \frac{1}{k}\,\frac{p}{n}\Tr_k\hat\bS_n(z\bI_k) = \frac{\alpha_n}{k}\Tr_k\hat\bS_n(z\bI_k).
%\end{equation}
%By Corollary~\ref{cor:convergence} applied at $\bZ = z\bI_k$, $\hat\bS_n(z\bI_k)\to \hat\bS_{\alpha,\nu}(z\bI_k)$ in probability. Combined with $\alpha_n\to\alpha$ and the continuity of the trace, this gives
%\begin{equation}
%s_n(z) \xrightarrow{\P} \frac{\alpha}{k}\Tr_k\hat\bS_{\alpha,\nu}(z\bI_k) = \frac{\alpha}{k}\Tr_k\bS_{\alpha,\nu}(z) = s_{\nu,\alpha}(z)\, .
%\end{equation}
%This convergence holds for every $z\in\bbH_+$.
%
%It remains to show that $s_{\nu,\alpha}$ is the Stieltjes transform of some probability
%measure $\mu_{\alpha,\nu}\in\cuP(\R)$. Equivalently~\cite[Theorem 2.4.3]{anderson2010introduction},
%we must show that
%$s_{\nu,\alpha}:\bbH_+\to\bbH_+$ is analytic, $\Im(s_{\nu,\alpha}(z))>0$ for $z\in\bbH_+$, and
%$\lim_{y\to\infty}(-iy) s_{\nu,\alpha}(iy) = 1$.
%
%\emph{Analyticity}: Since $\bS_{\alpha,\nu}:\bbH_+\to\bbH_+^k$ is the
%unique solution of an analytic implicit equation
%$\bK_{\alpha,\nu}(\bS_{\alpha,\nu}(z)) = z\bI_k$, and the derivative
%$\cuD\bK(\bS)[\cdot]$ is invertible at any $\bS\in\bbH_+^k$
%(as can be seen via the same operator $\cuL_\bS$ as in Lemma~\ref{lemma:T_diff_FP}, applied at
%$\tilde\bS = \bS$ and using Lemma~\ref{lemma:Gamma_at_FP} which implies $\|\alpha^{-1}\cuL_\bS\|_{\op\to\op}<1$),
%the implicit function theorem gives that $z\mapsto \bS_{\alpha,\nu}(z)$ is analytic on $\bbH_+$
%into $\bbH_+^k$. Hence $s_{\nu,\alpha}(z) = \frac{\alpha}{k}\Tr\bS_{\alpha,\nu}(z)$ is analytic on
%$\bbH_+$. Moreover, $\Im(\bS_{\alpha,\nu}(z))\succ\bzero$ for $z\in\bbH_+$, so
%$\Im(s_{\nu,\alpha}(z)) = \frac{\alpha}{k}\Tr_k\Im(\bS_{\alpha,\nu}(z)) > 0$.
%
%\emph{Asymptotic at infinity}: For $z = iy$ with $y\to\infty$, the fixed-point equation
%$\bK_{\alpha,\nu}(\bS) = iy\bI_k$ forces $\bS\to\bzero$. More precisely,
%$\bS \approx -\frac{1}{\alpha iy}\bI_k + O(1/y^2)$ as $y\to\infty$: the dominant term comes from
%$-\frac{1}{\alpha}\bS^{-1}=iy\bI_k$, so $\bS^{-1} = -i\alpha y(1+ O(1/y))\bI_k$, giving
%$\bS = -\frac{1}{i\alpha y}(1+O(1/y))\bI_k = \frac{i}{\alpha y}(1+O(1/y))\bI_k$.
%Then $\Tr\bS_{\alpha,\nu}(iy) = \frac{ik}{\alpha y}(1+O(1/y))$, so
%\begin{equation}
%-i y\cdot s_{\nu,\alpha}(iy) = -iy\cdot \frac{\alpha}{k}\cdot \frac{ik}{\alpha y}(1+O(1/y)) = 1 + O(1/y) \to 1\, .
%\end{equation}
%
%\emph{Convergence of measures}: Finally, since $\hmu_{\bW_k}$ has bounded support (as $\bW_k$ has
%operator norm bounded a.s.\ in the relevant regime by Lemma~\ref{lemma:det_norm_bounds} et seq.),
%and since $s_n(z) \to s_{\nu,\alpha}(z)$ in probability for every $z\in \bbH_+$, the standard
%correspondence between weak convergence of probability measures and pointwise convergence of
%Stieltjes transforms on $\bbH_+$ \cite[Theorem 2.4.4]{anderson2010introduction} implies
%$\hmu_{\bW_k}\Rightarrow \mu_{\alpha,\nu}$ in probability, where $\mu_{\alpha,\nu}$ is the
%probability measure whose Stieltjes transform is $s_{\nu,\alpha}$. \qed
%
%%%% Old draft material removed; see Appendix \ref{app:proofs_FP} above for the consolidated proof.
%\iffalse
%\begin{lemma}
%Fix $\bS,\tilde\bS\in\cD_0 \cup \bbH_+^k$.
%For any $\bB,\tilde\bB \succ\bzero,$ and integer $p>0$, we have
%the bound
%\begin{equation}
%\|\cuG_{\bS,\tilde\bS}^p\|_{\fro\to\fro}^2
%\le
% \|\bB^{-1}\|_\op \|\bB\|_\op
% \|\tilde\bB^{-1}\|_\op \|\tilde\bB\|_\op
% \|\bB^{-1/2} \cuG_{\bS}[\bB]\bB^{-1/2} \|_\op^{p}
% \|\tilde\bB^{-1/2} \cuG_{\tilde\bS}[\tilde\bB]\tilde\bB^{-1/2} \|_\op^{p}.
%\end{equation}
%Additionally, if $\bS \in\cD_0$, in particular $\bS\succ\bzero,$ we further have for $\cuT_\bS[\bDelta]:= \bS^{-1/2}\cuG_\bS[\bS^{1/2}\bDelta \bS^{1/2}]\bS^{-1/2},$
%\begin{equation}
%\|\cuT_\bS^p\|_{\fro\to\fro}
%\le
% \|\bB^{-1}\|_\op \|\bB\|_\op
% \|\bB^{-1/2} \cuT_{\bS}[\bB]\bB^{-1/2} \|_\op^{p}
%\end{equation}
%\end{lemma}
%\begin{proof}
%Let $\bT_1,\dots,\bT_{2p}$ be distributed under the product measure $\nu^{\otimes 2p}$.
%Define for $i\in[2p],$
%\begin{equation}
% \bL_i := \bS(\bI + \bT_i\bS)^{-1}\bT_i ,\quad\quad
% \bR_i := \bT_i(\bI + \bS\bT_i)^{-1} \tilde\bS.
%\end{equation}
%Fix $\bDelta\in\C^{k\times k}.$
%It will be useful to use the expectation notation and suppress the subscript in this proof so that
%\begin{equation}
% \cuG[\bDelta] := \alpha \E[\bL_1\bDelta \bR_1]
%\end{equation}
%Then the $p$-fold composition can be written as
%\begin{equation}
% \cuG^p[\bDelta] = \alpha^{p} \E[\bL_p\dots\bL_1 \bDelta \tilde\bR_1 \dots\tilde\bR_p].
%\end{equation}
%
%\begin{align}
% \|\cuG^p(\bDelta)\|_\fro^2
% &= \Tr( \cuG^p(\bDelta)^* \cuG^p(\bDelta))\\
% &= \alpha^{2p} \E\left[ \Tr\left(
% \bR_p^*\dots\bR_1^* \bDelta^* \bL_1^* \dots\bL_p^*
% \bL_{p+1} \dots\bL_{2p} \bDelta \bR_{2p}\dots\bR_{p+1}
% \right)
% \right]\\
% &= \alpha^{2p}
% \E\left[ \Tr\left(
% \bL_1^* \dots\bL_p^*
% \bL_{p+1} \dots\bL_{2p} \bDelta \bR_{2p}\dots\bR_{p+1}
% \bR_p^*\dots\bR_1^*
% \bDelta^*
% \right)
% \right]\\
% &\le\alpha^{2p}
% {\underbrace{\E[\Tr(\bDelta^*\bL_{2p}^*\dots\bL_{p+1}^* \bL_{p} \dots\bL_{1} \bL_{1}^*
% \dots \bL^*_{p} \bL_{p+1}\dots\bL_{2p}\bDelta
% )]}_{T_1}}^{1/2}\\
% &\quad\quad
%{\underbrace{\E[\Tr(\bDelta\bR_{1}\dots\bR_{p} \bR_{p+1}^* \dots\bR_{2p}^* \bR_{2p}
% \dots \bR_{p+1} \bR_{p}^*\dots\bR_{1}^*\bDelta^*
% )]}_{T_2}}^{1/2}.
%\end{align}
%Now to bound $T_1$, since $\bB\succ\bzero,$ we can write
%\begin{align}
% T_1&:=
% \E\left[\Tr\left(\bDelta^*\bA_1\dots\bA_p \bA_{p+1}^* \dots\bA_{2p}^*\bB^{1/2}\bB^{-1} \bB^{1/2} \bA_{2p}
% \dots \bA_{p+1} \bA_p^*\dots\bA_1^*\bDelta
% \right)\right]
% \\
% &\le \|\bB^{-1}\|_\op
% \E\left[\Tr\left(\bDelta^*\bA_1\dots\bA_p \bA_{p+1}^* \dots \bB^{1/2}
% \E\big[\bB^{-1/2}\bA_{2p}^*\bB\bA_{2p}\bB^{-1/2}\big]
% \bB^{1/2}
% \dots \bA_{p+1} \bA_p^*\dots\bA_1^*\bDelta
% \right)\right]\\
% &\le \|\bB^{-1}\|_\op
% \|\E\big[\bB^{-1/2}\bA_{2p}^*\bB\bA_{2p}\bB^{-1/2}\big] \|_\op
% \E\left[\Tr\left(\bDelta^*\bA_1\dots\bA_p \bA_{p+1}^* \dots \bB
% \dots \bA_{p+1} \bA_p^*\dots\bA_1^*\bDelta
% \right)\right].
%\end{align}
%Iterating this argument $2p-1$ many more times, gives
%\begin{equation}
% T_1 \le
% \|\bB^{-1}\|_\op \|\bB\|_\op
% \|\E\big[\bB^{-1/2}\bA_{1}^*\bB\bA_{1}\bB^{-1/2}\big] \|_\op^{p} \|\bDelta\|_\fro^2.
%\end{equation}
%Using $\tilde\bB\succ\bzero$ in a similar argument for $T_2$ gives the similar bound
%\begin{equation}
% T_2 \le
% \|\tilde\bB^{-1}\|_\op \|\tilde\bB\|_\op
% \|\E\big[\tilde\bB^{-1/2}\tilde\bA_{1}^*\tilde\bB\tilde\bA_{1}\tilde\bB^{-1/2}\big] \|_\op^{p} \|\bDelta\|_\fro^2.
%\end{equation}
%Combining with the above display gives the result.
%\end{proof}
%\begin{corollary}
%Let
% \begin{equation}
% \tilde \cP := \{\bS\succ\bzero : \bS^{-1} + \bT \succ\bzero, \bT\in\supp(\nu),\; \exists \bB\succ\bzero : \cuD \bK(\bS)[\bB]\succ\bzero\}.
% \end{equation}
% Then $\cP = \tilde\cP.$
%\end{corollary}
%\begin{proof}
% Assume $\bS\in\tilde\cP.$ Then by Eq.~\eqref{eq:derivative_operator_relation}, we have
% $\bB \succ \cuG_\bS[\bB]$, i.e., for some $\delta>0$, we have
% $\|\bB^{-1/2}\cuG_\bS[\bB] \bB^{-1}\|_\op < \delta,$
% which when combined with
% Lemma~\notate{ref} yields the bound
% \begin{equation}
% \|\cuG^p_\bS\|_{\fro\to\fro} \le \alpha^{p}
% \|\bB\|_\op \|\bB^{-1}\|_\op \delta^{p.}
% \end{equation}
% Since for $\bS\in\tilde\cP$, the operator $\cuG_\bS$ is self-adjoint when restricted on self-adjoint matrices, we have
% \begin{equation}
% \|\cuG_\bS\|_{\fro\to\fro} =\lim_{p\to\infty} \|\cuG_\bS^p\|_{\fro\to\fro}^{1/p} < 1.
% \end{equation}
% Then Eq.~\eqref{eq:derivative_operator_relation} once again yields for $\bA\in\sa{k}{\C}$,
% \begin{equation}
% \Tr( \bDelta\cuD \bK(\bS)[\bDelta])
% = \frac1\alpha\Tr( \bS^{}
% \end{equation}
%\end{proof}
%
%\newpage
%
%
%Similarly define $\tilde \bA_i,\tilde \bZ_i$ using $\tilde\bT_i.$ Then from the definition of $\cuT_\bS$, we see that
%\begin{align}
%\cuT_\bS^p(\bDelta) &= \alpha^p\E\left[\bA_p\cdots \bA_1 \bDelta \bA_1 \cdots\bA_p \right]
%\end{align}
%So we can bound
%\begin{align}
% \|\cuT_\bS^p(\bDelta)\|_\fro^2 &=
% \alpha^{2p}\E\left[
% \Tr\left(\bA_p\cdots \bA_1 \bDelta \bA_1 \cdots\bA_p
% \tilde\bA_p\cdots \tilde\bA_1 \bDelta \tilde\bA_1 \cdots\tilde\bA_p
% \right)
% \right]\\
% &\le
% \alpha^{2p}
% \E\left[
% \Tr\left(\bDelta \bA_1 \cdots\bA_p
% \tilde\bA_p\cdots \tilde\bA_1 \tilde\bA_1 \cdots\tilde\bA_p
% \bA_p\cdots \bA_1 \bDelta
% \right)
% \right].
%\end{align}
%Since the matrix inside the trace is PSD almost-surely, for any $\bB\succ\bzero,$ we have
%\begin{align}
% &\E\left[
% \Tr\left(\bDelta \bA_1 \cdots\bA_p
% \tilde\bA_p\cdots \tilde\bA_1\bB^{1/2}\bB^{-1}\bB^{1/2} \tilde\bA_1 \cdots\tilde\bA_p
% \bA_p\cdots \bA_1 \bDelta
% \right)
% \right]\\
% &\le
% \|\bB^{-1}\|_{\op}
% \E\left[
% \Tr\left(\bDelta \bA_1 \cdots\bA_p
% \tilde\bA_p\cdots \bB^{1/2} \E[\bB^{-1/2} \tilde\bA_1\bB\tilde\bA_1\bB^{-1/2}]\bB^{1/2} \cdots\tilde\bA_p
% \bA_p\cdots \bA_1 \bDelta
% \right)
% \right]\\
% &\le\|\bB^{-1}\|_\op
% \|\E[\bB^{-1/2} \tilde\bA_1\bB\tilde\bA_1\bB^{-1/2}]\|_\op
% \E\left[
% \Tr\left(\bDelta \bA_1 \cdots\bA_p
% \tilde\bA_p\cdots \tilde\bA_2\bB\tilde\bA_2 \cdots\tilde\bA_p
% \bA_p\cdots \bA_1 \bDelta
% \right)
% \right].
%\end{align}
%Repeatedly applying this argument $2p-1$ many times, we obtain that for any $\bB\succ\bzero$,
%\begin{equation}
% \|\cuT_\bS^p(\bDelta)\|_\fro^2 \le\alpha^{2p} \|\bB^{-1}\|_\op \|\bB\|_\op
% \|\E[\bB^{-1/2}\bA\bB\bA\bB^{-1/2}]\|_\op^{2p} \|\bDelta\|_\fro^2,
%\end{equation}
%i.e.,
%\begin{equation}
%\label{eq:F_to_F_bound_composition}
% \|\cuT_\bS^p\|_{\fro\to\fro} \le \alpha^{p} \|\bB\|_\op^{1/2}\|\bB^{-1}\|_\op^{1/2}
% \|\E[\bB^{-1/2}\bA\bB\bA\bB^{-1/2}]\|_\op^{p},
%\end{equation}
%where $\bA = \bZ(\bI + \bZ)^{-1}$.
%Consequently, if there is a $\bB\succ\bzero$ satisfying $\cuT_\bS(\bB)\prec\bB$, then for some $\delta\in(0,1),$
%\begin{equation}
%\|\E[\bB^{-1/2} \bA \bB\bA \bB^{-1/2}]\|_\op =
% \frac1{\alpha}\|\bB^{-1/2} \cuT_\bS(\bB) \bB^{-1/2}\|_\op < \frac{\delta}{\alpha}.
%\end{equation}
%Then since $\cuT_\bS^p$ is self-adjoint, we have by combining with Eq.~\eqref{eq:F_to_F_bound_composition} that
%\begin{equation}
% \|\cuT_\bS\|_{\fro\to\fro} =\lim_{p\to\infty} \|\cuT_\bS^p\|_{\fro\to\fro}^{1/p}
% \le \delta \,\lim_{p\to\infty} C^{1/p} < 1.
%\end{equation}
%
%
%
%
%\section{Left-over proofs.}
%
%\begin{lemma}
% Assume $\bX\in\cX$. Then the derivative of
% \begin{equation}
% \bG(\bS)
% \end{equation}
% as an operator on complex matrices $\bDelta$,
% has a non-singular derivative at $\bS_\star(\bX)$.
% Consequently, there exists $\eps_0$ such that for all $\eps\in(0,\eps_0)$,
% $\bS_\star(\bX) + i \eps\bI$ is the Stieltjes transform of some $\bX_\eps + i \bY_\eps$ with $\bX_\eps \in\cX$, $\bY_\eps \succ\bzero$.
% Namely, as $\eps\to0$,
% \begin{equation}
% \bX_\eps\to\bX,\quad\quad \bY_\eps \to \bzero.
% \end{equation}
%\end{lemma}
%
%
%\bns{Proof of the operator to operator norm bound.}\\
%\bns{Proof of the asymptotics of the Stieltjes transform for general $\bZ.$}\\
%\bns{Deducing that the smallest eigenvalue sticks to the bulk and the limit of the determinant}
%\fi
%
%
%\newpage
%
%
%%\section{Proof}
%%\begin{lemma}[Regularity outside of the spectrum.]
%%Let $(\cA,\tau)$ be a non-commutative probability space.
%%Fix a self-adjoint $\bH\in\cM^k(\cA).$
%%For any $\bX\in \sfS^{k}(\R)$ satisfying
%% $\bH - (\bX \otimes \id) \succ\bzero,$
%% \begin{enumerate}
%% \item We have
%% $$\bS_\star(\bX) := (\bI\otimes \tau)\odot(\bH - (\bX\otimes \id))^{-1} = \lim_{\bY\succ\bzero, \|\bY\|_\op \to 0} \bS_\star(\bX + i \bY).$$
%% \item Fix any $\bB \succ\bzero$ with $\|\bB\| < \infty$. There exists sequences $\bA_m,\bX_m \in \sfS^k$, $b_m > 0,\bY_m \succ\bzero$ with
%% \begin{equation}
%% \|\bX_m -\bX\| \to 0,\quad \|\bA_m - \bS_\star(\bX)\| \to 0\quad b_m \to 0,\quad \|\bY_m\| \to 0
%% \end{equation}
%% such that
%% \begin{equation}
%% \bA_m + i b_m \bB = \bS_\star(\bX_m + i \bY_m).
%% \end{equation}
%% \end{enumerate}
%%% \item (Zero imaginary part implies outside of support)
%%% Assume
%%%\begin{equation}
%%% \lim_{\|\bY\|_\op \to 0, \bY\succ\bzero}\Im(\bS_\star(\bX + i\bY)) = \bzero.
%%%\end{equation}
%%%Then $\bH - (\bX\otimes \id)$ is invertible.
%%\end{lemma}
%
%%\begin{proof}
%%For the first item, it is sufficient to note that $\bS_\star(\bX)$ is analytic on $\{\bH - (\bX\otimes \id) \succ\bzero\}.$
%%For the second item, note that for any $\bZ \in \{\bX + i \bY: \}$
%%\begin{equation}
%% \bZ \mapsto
%%\end{equation}
%%\end{proof}
%%
%%
%%\subsection{Proof of the first theorem.}
%%\subsubsection{Proof of the first item.}
%%This is the first item in the lemma above.
%%\subsubsection{Proof of the second item.}
%%The first item implies that
%%\begin{equation}
%% \lim_{}
%%\end{equation}
%%
%%
%%For $\bX\in \cX$, $\bX\mapsto (\bI\otimes \tau)\odot (\bH -(\bX \otimes \id))^{-1}$ is invertible.
%%Fix $\bX \in\cX$. Since $\cX$ is open, there exists some open $\cU$ with $\bx \in \cU \subseteq \cX$ such that for all $\bX' \in \cU$, we have $G(\bS') = \bX'$ by the first item.
%%\subsubsection{Proof of the third item.}
%%Fix $\bX \in\cX$, and consider for $y \in(0,1)$,
%%\begin{equation}
%% \bZ_y := \bX + i y \bI,
%%\end{equation}
%%and let
%%\begin{equation}
%% \bS_y := \bS_\star(\bZ_y)= \bA_y + i \bB_y \in \bbH_+^k.
%%\end{equation}
%%Then by definition,
%%\begin{equation}
%% \bG(\bS_y) = - \bZ_y.
%%\end{equation}
%%Taking imaginary part of both sides gives
%%\begin{equation}
%% y \bI = \frac1\alpha \bS_y^{-1} \bB_y \bS_y^{*-1} - \E[(I + \bD \bS_y)^{-1}\bD \bB_y \bD (\bI + \bS_y^* \bD)^{-1}],
%%\end{equation}
%%which gives after re-arranging
%%\begin{equation}
%%\bB_y^{-1/2}\bS_y \E[(I + \bD \bS_y)^{-1}\bD \bB_y \bD (\bI + \bS_y^* \bD)^{-1}] \bS_y^*\bB_y^{-1/2}
%% = \frac1\alpha\bI - y\bB_y^{-1/2} \bS_y \bS_y^*\bB_y^{-1/2}.
%%\end{equation}
%%
%%We claim that
%%\begin{equation}
%% \alpha y\bB_y^{-1/2} \bS_y \bS_y^*\bB_y^{-1/2} \prec q\bI
%%\end{equation}
%%for some $q\in(0,1)$ independent of $y$.
%%
%%To see this, note that
%%\begin{align}
%% \alpha \bB_y &= \Im\circ (\bI_k \otimes \tau)\circ\left(\bH - (\bZ_y\otimes \bI)\right)^{-1}\\
%% &= (\bI_k \otimes \tau)\circ \Im\circ\left(\bH - (\bZ_y\otimes \bI)\right)^{-1}\\
%% &= y\odot(\bI_k \otimes \tau) \circ (\bH - (\bZ_y\otimes\bI ))^{-1} (\bH -(\bZ_y\otimes \bI) )^{* -1}.
%%\end{align}
%%
%%Meanwhile, \bns{prove or find a reference}
%%\begin{align}
%% \left((\bI_k \otimes \tau) \circ (\bH - (\bZ_y\otimes\bI ))^{-1}\right)
%% \left((\bI_k \otimes \tau) \circ (\bH - (\bZ_y\otimes\bI ))^{-1}\right)^{*}
%% &\stackrel{(a)}{\prec}
%% (\bI_k \otimes \tau) \circ (\bH - (\bZ_y\otimes\bI ))^{-1}(\bH - (\bZ_y\otimes\bI ))^{*-1 }\\
%% &\stackrel{(b)}{\preceq}
%% q(\bI_k \otimes \tau) \circ (\bH - (\bZ_y\otimes\bI ))^{-1}(\bH - (\bZ_y\otimes\bI ))^{*-1 }
%%\end{align}
%%where $(a)$ is strict for any non-deterministic measure, and hence holds strictly for and any $y \in [0,1]$, so there exists some $q \in (0,1)$ independent of $y$ so that $(b)$ holds.
%%Using the notation above, the previous display gives
%%\begin{equation}
%% \alpha^2 \bS_y\bS_y^* \preceq \alpha\,q \frac1{y} \bB_y,
%%\end{equation}
%%implying the desired claim, allowing us to conclude that
%%
\end{document}