The circular law for non-Hermitian random band matrices up to bandwidth $N^{1/2+c}$

TL;DR

This paper proves that the eigenvalue distribution of certain inhomogeneous random band matrices converges to the circular law when the bandwidth exceeds a threshold of approximately the square root of the matrix size, extending previous results.

Contribution

It establishes the circular law for non-Hermitian band matrices with bandwidth above $N^{1/2+c}$, relaxing previous moment and density assumptions.

Findings

Eigenvalue distribution converges to the circular law for $W \,\geq\, N^{1/2+c}$

New bounds on small singular values via Green function estimates

Extended circular law results to broader bandwidth regimes

Abstract

We consider inhomogeneous square random matrices of size $N$ with independent entries of mean 0 and finite variance. We assume that the variance profile of this matrix is doubly stochastic and has a band-like structure with an appropriately defined bandwidth $W$. We prove that when the entries have a bounded density and a subgaussian tail, then the empirical spectral measure for the eigenvalues of the matrix converges to the circular law as $N$ tends to infinity whenever $W\geq N^{1/2+c}$ for any $c>0$. In the special case of block band matrices the density assumption is not needed and the moment condition is relaxed. This establishes the circular law limit throughout the entire delocalization regime in 1-d: $W\geq N^{1/2+c}$ and extends the previous thresholds for the circular law limit with exponent $\frac{5}{6},\frac{8}{9},\frac{33}{34}$ in $N$. The main technical input is a new lower bound on the small-ish singular values via Green function estimates and a new lower bound on the least singular value with fewer moment conditions.