[Paper Review] Eigenvectors of Wigner matrices: universality of global fluctuations

Florent Benaych-Georges|arXiv (Cornell University)|Apr 7, 2011

Random Matrices and Applications17 references7 citations

TL;DR

This paper establishes universality for the global fluctuations of eigenvectors in Wigner matrices, proving that the normalized sum of squared eigenvector entries converges to a bivariate Brownian bridge under mild moment conditions. Surprisingly, only the fourth moment of the matrix entries' distribution affects the limit, while the third moment has no influence, confirming a conjecture on universality beyond Gaussian ensembles.

ABSTRACT

Let $U_n=[u_{i,j}]$ be the eigenvectors matrix of a Wigner matrix. We prove that under some moments conditions, the bivariate random process indexed by $[0,1]^2$ with value at $(s,t)$ equal to the sum, over $1\le i \le ns$ and $1\le j \le nt$, of $|u_{i,j}|^2 - 1/n$, converges in distribution to the bivariate Brownian bridge. This result has already been proved for GOE and GUE matrices. It is conjectured here that the necessary and sufficient condition, for the result to be true for a general Wigner matrix, is the matching of the moments of orders 1, 2 and 4 of the entries of the Wigner with the ones of a GOE or GUE matrix. Surprisingly, the third moment of the entries of the Wigner matrix has no influence on the limit distribution.

Motivation & Objective

• To establish universality of global eigenvector fluctuations in non-Gaussian Wigner matrices.
• To determine the minimal moment conditions under which the eigenvector fluctuations converge to a bivariate Brownian bridge.
• To investigate the role of higher-order moments—particularly the third and fourth moments—on the limiting distribution.
• To extend known results from GOE/GUE ensembles to general Wigner matrices with i.i.d. entries.
• To confirm a conjecture that only fourth-moment matching is necessary for the universal limit, not third moments.

Proposed method

• Uses the bivariate process $ B^n_{s,t} = \sqrt{\frac{\beta}{2}} \sum_{1 \leq i \leq ns, \, 1 \leq j \leq nt} (|u_{i,j}|^2 - 1/n) $ to study global eigenvector fluctuations.
• Applies Weingarten calculus to compute joint moments of eigenvector entries for Haar-distributed orthogonal/unitary matrices.
• Employs tightness and $ C $-tightness arguments in the Skorokhod topology to establish weak convergence.
• Compares the variance of the process under general Wigner matrices to that of GOE/GUE via moment matching and error bounds.
• Uses Chebyshev's inequality and Proposition 2.10 to control maximal jumps and prove convergence in distribution.
• Relies on functional central limit theorem techniques and asymptotic analysis of matrix entries' moments.

Experimental results

Research questions

• RQ1 Does the global fluctuation of eigenvectors in general Wigner matrices converge to a bivariate Brownian bridge, as in the GOE/GUE case?
• RQ2 What is the minimal set of moment conditions on the matrix entries required for this convergence?
• RQ3 Why does the third moment of the entries not affect the limiting distribution, despite its apparent role in perturbation theory?
• RQ4 Is the convergence universal across all Wigner matrices with matching fourth moments, regardless of higher moments?
• RQ5 Can the universality result be extended to other matrix decompositions, such as SVD or Housholder, under i.i.d. entry assumptions?

Key findings

• The bivariate process $ B^n_{s,t} $ converges in distribution to a bivariate Brownian bridge under the condition that the fourth moments of the matrix entries match those of GOE or GUE matrices.
• The convergence holds in the weak topology, and under additional regularity assumptions (e.g., continuity and higher moment matching), it holds in the Skorokhod topology.
• The third moment of the entries has no influence on the limiting distribution, which is a surprising and nontrivial universality result.
• The limiting covariance structure is $ \mathbb{E}[B_{s,t}B_{s',t'}] = (\min\{s,s'\} - ss')(\min\{t,t'\} - tt') $, identical to the GOE/GUE case.
• The proof relies on moment comparison techniques and Weingarten calculus to show that the variance of the fluctuation process matches the GOE/GUE case up to a vanishing error term.
• The maximal jump of the process is shown to vanish in probability, ensuring $ C $-tightness and enabling convergence to a continuous limit.