Large deviations for the largest eigenvalue of sub-Gaussian matrices (Q2662978)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Large deviations for the largest eigenvalue of sub-Gaussian matrices |
scientific article |
Statements
Large deviations for the largest eigenvalue of sub-Gaussian matrices (English)
0 references
15 April 2021
0 references
The article studies large deviations of the largest eigenvalue for random matrices with sub-Gaussian entries. The main object is an \(N\times N\) symmetric random matrix \(X_N\) with independent entries above the diagonal so that \(\sqrt{N}X_{ij}\) has law \(\mu\) for \(i\neq j\) and \(\sqrt{N/2} X_{ii}\) has law \(\mu\) for all \(i\), where \(\mu\) is a centered sub-Gaussian variable with unit variance. The authors study the case where \(\mu\) has heavier than Gaussian tails so that \(A = 2\sup_{x\in \mathbb R} \int_{\mathbb R} e^{xt} \mu(dt)>1\). Under some technical assumptions, they show that there exist some good rate function \(I_\mu:\mathbb R\to [0,\infty]\), \(x_\mu>2\) and a closed set \(O_\mu\supset (-\infty,2]\cup [x_\mu,\infty)\) such that the largest eigenvalue \(\lambda_{X_N}\) of \(X_N\) satisfies \[ \begin{gathered} \lim_{\delta \to 0+} \liminf_{N\to \infty} \frac1N \mathbb {P}\big(|\lambda_{X_N} - x| \le \delta\big)\\ = \lim_{\delta \to 0+} \limsup_{N\to \infty} \frac1N \mathbb {P}\big(|\lambda_{X_N} - x| \le \delta\big) = - I_\mu(x) \end{gathered} \] for all \(x\in O_\mu\). The rate function \(I_\mu\) is infinite on \((-\infty,2)\) and satisfies \(I_\mu(x) \sim x^2/(4A), x\to+\infty\), moreover, \(I_\mu(x)\le I_{\mathrm{GOE}}(x) = \frac12 \int_2^x \sqrt{y^2-4}dx\), the rate function of the Gaussian orthogonal ensemble (GOE). If \(A\in (1,2)\), then for all \(x\in [2,\sqrt{A-1} + 1/\sqrt{A-1}]\subset O_\mu\), \(I_\mu(x) = I_{\mathrm{GOE}}(x)\). As a result, in the regime of very large deviations, there exists, at least for \(A<2\), a transition between two large deviation mechanisms: universality for small \(x\) where the rate function coincides with that of GOE, and non-universality for large \(x\) where the rate function depends on \(A\).
0 references
random matrix
0 references
largest eigenvalue
0 references
sub-Gaussian tail
0 references
large deviation
0 references