On the spectral radius of the ratio of Girko matrices
Djalil Chafaï, David García-Zelada, Yuan Yuan Xu · 2025
Read On the spectral radius of the ratio of Girko matrices by Djalil Chafaï, David García-Zelada, Yuan Yuan Xu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-18669.txt.
Introduction
Djalil Chafaï, with David García-Zelada and Yuan Yuan Xu, studies a random matrix built as the ratio of two simpler random matrices. Each ingredient matrix has independent entries with mean zero and unit variance, but dividing one by the other creates entries that depend on each other and include occasional huge values. The authors ask how far the ratio's spectrum, its set of eigenvalues, reaches as the matrix grows.
Their main message is that the farthest eigenvalue, properly scaled, settles into a single universal heavy-tailed distribution as the dimension grows without bound. The proof leans on a symmetric Gaussian special case, whose spectrum can be pictured as charged particles on a sphere, plus tools from random matrix theory and determinantal processes. Curiously, they note, the ratio case proved more tractable than the same question for a single random matrix.
Catalog record
| CREATOR | Djalil Chafaï, David García-Zelada, Yuan Yuan Xu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 73 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-18669
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
On the spectral radius of the ratio of Girko matrices Djalil Chafaï, David García-Zelada, Yuan Yuan Xu Abstract . Girko matrices have independent and identically distributed entries of mean zero and unit variance. In this note, we consider the random matrix model formed by the ratio of two independent Girko matrices, its entries are dependent and heavy-tailed. Our main message is that divided by the square root of the dimension, the spectral radius of the ratio converges in distribution, when the dimension tends to infinity, to a universal heavy-tailed distribution. We provide a mathematical proof of this high-dimensional phenomenon, under a fourth moment matching with a Gaussian case known as the complex Ginibre ensemble. In this Gaussian case, the model is known as the spherical ensemble, and its spectrum is a determinantal planar Coulomb gas. Its image by the inverse stereographic projection is a rotationally invariant gas on the two-sphere. A crucial observation is the invariance in law of the model under inversion, related to its spherical symmetry, and that makes, in a sense, edge and bulk equivalent. Our approach involves Girko Hermitization, local law estimates for Wigner matrices, lower bound estimates on the smallest singular value, and convergence of kernels of determinantal point processes. The universality of the high-dimensional fluctuation of the spectral radius of the ratio of Girko matrices turns out to be remarkably more accessible mathematically than for a single Girko matrix! 2020 Mathematics Subject Classification 60B20 - Random matrices (probabilistic aspects); 15A60 - Norms of matrices, numerical range, applications of functional analysis to matrix theory; 60F05 - Central limit and other weak theorems. 1. Introduction The asymptotic analysis in high dimension of the spectral radius of non-normal random matrices is a delicate subject, which goes back at least to Cohen and Newman 1984 ; Geman 1986 ; Bai and Yin 1986 ; Rider 2003 . There has been recent progress in the case of Girko random matrices, see for instance Bordenave et al. 2022 ; Cipolloni et al. 2024a ; Cipolloni et al. 2024b ; Byun and Forrester 2025 and references therein. Girko matrices have independent square-integrable entries. This note addresses the case of the ratio of two independent Girko matrices, a model of non-normal random matrices with dependent entries. For this model, we prove the universality of the asymptotic behavior of the spectral radius in high dimension. This model, in the special case of Ginibre matrices, known as the spherical ensemble, becomes integrable and biunitary invariant, and the law of its spectrum is a planar heavy-tailed Coulomb gas, with a determinantal structure. The spherical ensemble is a matrix analogue and extension of the ratio of independent real normal random variables, and can be then seen as a sort of matrix Cauchy law. The eigenvalues of an n\times n complex matrix A=(A_{jk}:1\leq j,k\leq n)\in\mathcal{M}_{n}(\mathbb{C}) are the roots in \mathbb{C} of its characteristic polynomial \det(A-z\mathrm{Id}_{n})\in\mathbb{C}[z] . They form a multiset of cardinal n called the spectrum of A , denoted \mathrm{spec}(A) . The empirical spectral distribution , spectral radius , and reciprocal or inner spectral radius of A are 1 1 1 Obviously when A is invertible, then \rho_{\min}(A)=\rho_{\max}(A^{-1})^{-1} , a formula that remains valid in an extended sense when A is singular. (1.1) \mu_{A}=\frac{1}{n}\sum_{\lambda\in\mathrm{spec}(A)}\delta_{\lambda},\qquad\rho_{\max}(A)=\max_{\lambda\in\mathrm{spec}(A)}|\lambda|,\quad\text{and}\quad\rho_{\min}(A)=\min_{\lambda\in\mathrm{spec}(A)}|\lambda|. 1.1. Main results We say that an n\times n random matrix A=(A_{jk}:1\leq j,k\leq n) is a Girko matrix when its entries are independent and identically distributed (i.i.d.) in \mathbb{C} with \mathbb{E}(A_{11})=0 and \mathbb{E}(|A_{11}|^{2})=1 . We say that A is a complex Ginibre matrix in the special case where A_{11}\sim\mathcal{N}_{\mathbb{C}}(0,\frac{1}{2}\mathrm{Id}_{2}) , which is equivalent to state that A has density in \mathbb{C}^{n\times n} proportional to A\in\mathcal{M}_{n}(\mathbb{C})\mapsto\exp(-\mathrm{Tr}(AA^{*})) . To distinguish, we denote such a Ginibre matrix by A^{\mathrm{Gin}} , and we always write \mathrm{Gin} as a superscript to indicate this special case. We consider the following assumptions or conditions on an n\times n Girko matrix A : (C1) the law of A_{11} has a bounded density function \varphi with respect to the Lebesgue measure on \mathbb{C} , and \|\varphi\|_{\infty}\leq D_{0}\quad\text{ for some constant $D_{0}>0$.} (C2) Fourth moment matching condition. There exists a small constant c_{0}>0 such that \mathcal{E}_{1}=\mathcal{E}_{2}=0,\quad\qquad\mathcal{E}_{3}\leq n^{-\frac{1}{2}-c_{0}},\quad\qquad\mathcal{E}_{4}\leq n^{-c_{0}}, where ( k -th moment deviation of A_{11} from standard complex Gaussian) \mathcal{E}_{k}=\max_{\begin{subarray}{c}a,b\in{\mathbb{N}}\\ a+b=k\end{subarray}}\left|\mathbb{E}[(\operatorname{Re}A_{11})^{a}(\operatorname{Im}A_{11})^{b}]-\mathbb{E}[(\operatorname{Re}A^{\mathrm{Gin}}_{11})^{a}(\operatorname{Im}A^{\mathrm{Gin}}_{11})^{b}]\right|. (C3) Finite moment condition. For each k\geq 1 , there exists constants D_{k}>0 (independent of n ) such that \mathbb{E}[|A_{11}|^{k}]<D_{k}. Condition (C1) implies that the n\times n complex matrix A is a.s. invertible for all n since the algebraic hypersurface \{\det=0\} has zero Lebesgue measure. Moreover, the uniform bound on the density allows a precise lower bound on the smallest singular value, see e.g., Jain et al. 2021 . More generally, we refer to Tikhomirov 2023 for the invertibility of random matrices beyond the density case. Condition (C2) states that the moments of A_{11} almost match those of A^{\mathrm{Gin}}_{11} up to the fourth order. Condition (C3) is assumed for technical reasons to simplify a proof. Example 1.1 (Moment matching with the standard complex Gaussian distribution) . An example of a probability measure \nu on \mathbb{C}\equiv\mathbb{R}^{2} with moment matching up to order k\geq 1 with \mu=\mathcal{N}(0,\frac{1}{2}\mathrm{Id}_{2}) is (1.2) \mathrm{d}\nu=(1+f)\mathrm{d}\mu where f\in L^{2}(\mu) is such that f\neq 0 , f\geq-1 , and f\perp\mathbb{R}_{k}[X,Y] where \mathbb{R}_{k}[X,Y] is the set of bivariate polynomials with total degree less than or equal to k , the span of X^{a}Y^{b} with a,b\in\mathbb{N} , a+b\leq k . Indeed, for all P\in\mathbb{R}_{k}[X,Y] , (1.3) \int P\mathrm{d}\nu-\int P\mathrm{d}\mu=\langle P,f\rangle_{L^{2}(\mu)}=0. The special case P=1 states that \nu is a probability measure. However, this construction has no reason to produce a product measure. Also, we may prefer to take alternatively (1.4) \mathrm{d}\nu=g\mathrm{d}\mu\quad\text{where}\quad g(z)=g(x+\mathrm{i}y)=(1+f_{1}(x))(1+f_{2}(y)) where f_{1},f_{2}:\mathbb{R}\to\mathbb{R} are in L^{2}(\mathcal{N}(0,\frac{1}{2})) and satisfy f_{i}\neq 0 , f_{i}\geq-1 , and f_{i}\perp\mathbb{R}_{k}[X] where \mathbb{R}_{k}[X] is the span of univariate polynomials with total degree less than or equal to k . A basic example is given by f_{1}=f_{2}=c^{-1}H_{m} where H_{m} is the Hermite polynomial, orthogonal with respect to \mathcal{N}(0,\frac{1}{2}) , of even degree m>k , and c=|\inf_{\mathbb{R}}H_{m}|>0 . In this case \nu has a bounded density with respect to the Lebesgue measure on \mathbb{C}\equiv\mathbb{R}^{2} . The moment matching up to order k between \mu and \nu comes from \mu=\mathcal{N}(0,\frac{1}{2})^{\otimes 2} . In another direction, by using the Gauss – Hermite quadrature, it is possible to construct a finite discrete distribution that matches the moments of the Gaussian up to an arbitrary fixed order. Theorem 1.2 (From ratio of Girko matrices to infinite Ginibre) . Let M=AB^{-1} where A and B are independent n\times n Girko matrices, not necessarily of the same law, both satisfying (C1) – (C3) . Let \mathrm{Gin}_{\infty} be the infinite Ginibre determinantal point process on \mathbb{C} , with kernel K(z,w)=\sqrt{\gamma(z)\gamma(w)}\mathrm{e}^{z\overline{w}} where \gamma(z)=\frac{1}{\pi}\mathrm{e}^{-|z|^{2}} . Then, for all fixed \lambda_{0}\in\mathbb{C} , for the topology of continuous and compactly supported test functions, n\mu_{\sqrt{n}(1+|\lambda_{0}|^{2})^{-1}(M-\lambda_{0}\mathrm{Id})}=\sum_{\lambda\in\mathrm{spec}(M)}\delta_{\sqrt{n}(1+|\lambda_{0}|^{2})^{-1}(\lambda-\lambda_{0})}\xrightarrow[n\to\infty]{\mathrm{d}}\mathrm{Gin}_{\infty}. Theorem 1.3 (Spectral radius) . Let M be as in Theorem 1.2 , and {(\gamma_{k})}_{k\geq 1} independent with \gamma_{k}\sim\mathrm{Gamma}(k,1) . Then (1.5) \frac{\rho_{\max}(M)}{\sqrt{n}}\xrightarrow[n\to\infty]{\mathrm{d}}\mathrm{Law}\Bigr(\frac{1}{\sqrt{\min_{k\geq 1}\gamma_{k}}}\Bigr)\quad\text{and}\quad\sqrt{n}\rho_{\min}(M)\xrightarrow[n\to\infty]{\mathrm{d}}\mathrm{Law}\bigr(\sqrt{\min_{k\geq 1}\gamma_{k}}\bigr). Note that the laws of R_{0}=\sqrt{\min_{k\geq 1}\gamma_{k}} and R_{\infty}=\displaystyle\frac{1}{R_{0}} have respective cdf given for all x>0 by (1.6) \mathbb{P}(R_{0}\leq x)=1-\prod_{k=1}^{\infty}(1-F_{k})(x^{2})\quad\text{and}\quad\mathbb{P}(R_{\infty}\leq x)=\prod_{k=1}^{\infty}(1-F_{k})(x^{-2}) where F_{k}(x)=\mathbb{P}(\gamma_{k}<x)=\mathbb{P}(\pi_{x}<k)=\mathrm{e}^{-x}\sum_{j=0}^{k-1}\frac{x^{j}}{j!} with \pi_{x}\sim\mathrm{Poisson}(x) . In particular R_{\infty} is heavy-tailed as (1.7) \mathbb{P}(R_{\infty}>x)\underset{x\to+\infty}{\sim}\frac{1}{x^{2}}. Since {(\xi_{k})}_{k\geq 1} are not i.i.d., we are beyond classical extreme value theory : R_{0} is