A concentration inequality and a local law for the sum of two random matrices

TL;DR

This paper improves concentration inequalities for the sum of two random matrices and establishes a local law for eigenvalues, showing convergence to free convolution density with high precision.

Contribution

It enhances existing concentration bounds to quadratic rate and proves a local law for eigenvalues in the sum of Hermitian matrices with Haar-distributed unitaries.

Findings

Improved concentration inequality with quadratic rate in N.

Established local law for eigenvalues in small intervals.

Eigenvalue distribution converges to free convolution density.

Abstract

Let H=A+UBU* where A and B are two N-by-N Hermitian matrices and U is a Haar-distributed random unitary matrix, and let \mu_H, \mu_A, and \mu_B be empirical measures of eigenvalues of matrices H, A, and B, respectively. Then, it is known (see, for example, Pastur-Vasilchuk, CMP, 2000, v.214, pp.249-286) that for large N, measure \mu_H is close to the free convolution of measures \mu_A and \mu_B, where the free convolution is a non-linear operation on probability measures. The large deviations of the cumulative distribution function of \mu_H from its expectation have been studied by Chatterjee in in JFA, 2007, v. 245, pp.379-389. In this paper we improve Chatterjee's concentration inequality and show that it holds with the rate which is quadratic in N. In addition, we prove a local law for eigenvalues of H, by showing that the normalized number of eigenvalues in an interval converges to the density of the free convolution of \mu_A and \mu_B provided that the interval has width (log N)^{-1/2}.