Functional form for the leading correction to the distribution of the largest eigenvalue in the GUE and LUE

TL;DR

This paper precisely characterizes the leading correction term in the distribution of the largest eigenvalue for GUE and LUE, providing operator and differential equation methods, and explores broader ensemble implications.

Contribution

It derives the exact functional form of the leading correction to the largest eigenvalue distribution in GUE and LUE, including new operator and differential equation representations.

Findings

Explicit correction formulas for GUE and LUE

Operator theoretic and differential equation characterizations

Simulation insights on broader matrix ensembles

Abstract

The neighbourhood of the largest eigenvalue $\lambda_{\rm max}$ in the Gaussian unitary ensemble (GUE) and Laguerre unitary ensemble (LUE) is referred to as the soft edge. It is known that there exists a particular centring and scaling such that the distribution of $\lambda_{\rm max}$ tends to a universal form, with an error term bounded by $1/N^{2/3}$. We take up the problem of computing the exact functional form of the leading error term in a large $N$ asymptotic expansion for both the GUE and LUE --- two versions of the LUE are considered, one with the parameter $a$ fixed, and the other with $a$ proportional to $N$. Both settings in the LUE case allow for an interpretation in terms of the distribution of a particular weighted path length in a model involving exponential variables on a rectangular grid, as the grid size gets large. We give operator theoretic forms of the corrections, which are corollaries of knowledge of the first two terms in the large $N$ expansion of the scaled kernel, and are readily computed using a method due to Bornemann. We also give expressions in terms of the solutions of particular systems of coupled differential equations, which provide an alternative method of computation. Both characterisations are well suited to a thinned generalisation of the original ensemble, whereby each eigenvalue is deleted independently with probability $(1 - \xi)$. In the final section, we investigate using simulation the question of whether upon an appropriate centring and scaling a wider class of complex Hermitian random matrix ensembles have their leading correction to the distribution of $\lambda_{\rm max}$ proportional to $1/N^{2/3}$.