[Paper Review] Universal Gaussian fluctuations of non-Hermitian matrix ensembles

Ivan Nourdin, Giovanni Peccati|ArXiv.org|Aug 4, 2009

Random Matrices and Applications3 citations

TL;DR

This paper establishes a multidimensional central limit theorem for spectral moments of non-Hermitian random matrices with i.i.d. real-valued entries, using universality principles from Gaussian chaos and Malliavin calculus. It proves that normalized traces of matrix powers converge jointly to independent Gaussian variables with variance $k$, and provides $O(N^{-1/4})$ bounds for smooth functionals of the normalized traces.

ABSTRACT

We prove multi-dimensional central limit theorems for the spectral moments (of arbitrary degrees) associated with random matrices with real-valued i.i.d. entries, satisfying some appropriate moment conditions. Our techniques rely on a universality principle for the Gaussian Wiener chaos, recently proved by the authors together with Gesine Reinert, as well as on some combinatorial estimates. Unlike other related results in the probabilistic literature, we do not require that the law of the entries has a density with respect to the Lebesgue measure. In particular, our results apply to the ensemble of Bernoulli random matrices.

Motivation & Objective

• To establish a multidimensional central limit theorem for spectral moments of non-Hermitian random matrices with i.i.d. real entries.
• To extend universality principles from Gaussian chaos to random matrix theory, enabling normal approximation without complex analysis.
• To derive explicit convergence rate bounds for smooth functionals of normalized spectral moments.
• To handle discrete distributions directly, avoiding the need for Gaussian approximation in the initial step.
• To demonstrate that the limiting distribution of normalized traces of matrix powers is multivariate Gaussian with variances proportional to the power index $k$.

Proposed method

• Applies universality principles from Nourdin and Reinert (2019) to replace i.i.d. entries with i.i.d. Gaussians in the context of homogeneous sums.
• Represents spectral moments $\operatorname{Tr}(X_N^k)$ as sums over index tuples $\mathbf{i} \in D_N^{(k)}$, which are finite homogeneous sums of i.i.d. random variables.
• Uses Malliavin calculus and Stein's method to analyze normal approximation of these homogeneous sums.
• Employs combinatorial estimates on pairings and chains in the moment expansions to control the variance and cumulants.
• Derives bounds on the difference between expectations of smooth functions applied to normalized traces and their Gaussian counterparts.
• Applies a key result from Corollary 3.4 to link the convergence of normalized traces to the limiting Gaussian vector.

Experimental results

Research questions

• RQ1 Do the normalized spectral moments $\operatorname{Tr}(X_N^k)$ of non-Hermitian random matrices with i.i.d. real entries converge jointly to a multivariate normal distribution?
• RQ2 Can the universality principle for Gaussian chaos be applied to derive convergence rates in random matrix spectral statistics?
• RQ3 What is the rate of convergence for smooth functionals of normalized spectral moments in non-Hermitian matrix ensembles?
• RQ4 Can the method handle discrete distributions without requiring Gaussian approximation of the original entries?
• RQ5 How do moment conditions on the entries affect the validity of the central limit theorem and the convergence rate?

Key findings

• The joint distribution of normalized traces $\left(\frac{\operatorname{Tr}(X_N^{k_1}) - \mathbb{E}[\operatorname{Tr}(X_N^{k_1})]}{\sqrt{\operatorname{Var}(\operatorname{Tr}(X_N^{k_1}))}}, \dots, \frac{\operatorname{Tr}(X_N^{k_m}) - \mathbb{E}[\operatorname{Tr}(X_N^{k_m})]}{\sqrt{\operatorname{Var}(\operatorname{Tr}(X_N^{k_m}))}}\right)$ converges in law to a multivariate Gaussian vector with independent components $Z_k \sim \mathcal{N}(0, k)$.
• For any thrice differentiable function $\varphi$ with bounded derivatives up to order three, the difference between the expectation of $\varphi$ applied to the normalized traces and its Gaussian counterpart is bounded by $C \cdot N^{-1/4}$, where $C$ depends on $\beta = \mathbb{E}|X|^3$, the bound $B$ on derivatives, and the indices $k_1, \dots, k_m$.
• The convergence rate $O(N^{-1/4})$ is uniform over classes of smooth functions with bounded third-order derivatives.
• The method applies directly to discrete distributions, as the universality principle allows replacing the original i.i.d. entries with i.i.d. Gaussians without loss of generality.
• The result holds under finite moments of all orders, and extends to cases with finite $2K$-th moments if $k_j \leq K$ for all $j$.
• The approach does not rely on complex analysis or characteristic function methods, distinguishing it from prior work such as Rider and Silverstein (2008).