Delocalization of Non-Mean-Field Random Matrices in Dimensions $d\ge 3$
Abstract
We study random band matrices with mean-zero complex Gaussian entries, where lie on the discrete torus in dimensions . The variance profile satisfies , with whenever the distance between and exceeds a bandwidth parameter . We prove that if for some constant , then in the large- limit, bulk eigenvectors are delocalized, quantum unique ergodicity (QUE) holds, and the local bulk eigenvalue statistics are universal. Our proof is based on the tree approximation of the loop hierarchy (arXiv:2501.01718) and diagrammatic techniques developed in earlier works (arXiv:1807.02447, arXiv:2104.12048, arXiv:2107.05795, arXiv:2412.15207, arXiv:2503.07606). Besides random band matrices, we also study two classical non-mean-field random matrix models: the Wegner orbital and the block Anderson models. Specifically, we consider Hermitian matrices on the same discrete torus , where is a random block potential consisting of i.i.d. complex Gaussian diagonal blocks of size , and encodes the interactions between neighboring blocks--random in the Wegner orbital model and deterministic in the block Anderson model. The parameter represents the coupling strength between blocks. Assuming again that , we establish delocalization of bulk eigenvectors, QUE, and bulk universality under the condition for any small constant . Combined with the localization results of arXiv:1608.02922 for , this identifies a localization--delocalization transition at the scale in dimensions .
Keywords
Cite
@article{arxiv.2507.20274,
title = {Delocalization of Non-Mean-Field Random Matrices in Dimensions $d\ge 3$},
author = {Sofiia Dubova and Fan Yang and Horng-Tzer Yau and Jun Yin},
journal= {arXiv preprint arXiv:2507.20274},
year = {2025}
}
Comments
95 pages. Minor updates