Delocalization of Non-Mean-Field Random Matrices in Dimensions $d\ge 3$

Abstract: Consider an $N \times N$ random band matrix $H = \left(H_{xy}\right)$ with mean-zero complex Gaussian entries, where $x, y$ lie on the discrete torus $(\mathbb{Z} / \sqrt[d]{N} \mathbb{Z})^d$ in dimensions $d \ge 3$. The variance profile $\mathbb{E}\left|H_{xy}\right|^2 = S_{xy}$ vanishes when the distance between $x$ and $y$ exceeds a given bandwidth parameter $W$. We prove that if the bandwidth satisfies $W \geq N^{\mathfrak{c}}$ for some constant $\mathfrak{c} > 0$, then in the large $N$ limit, both the delocalization of bulk eigenvectors and a local semicircle law down to optimal spectral scales $\eta \ge N^{-1+\varepsilon}$ hold with high probability. Our proof is based on the primitive approximation of the loop hierarchy (arXiv:2501.01718), and builds upon methods which were developed in this series since 2021 (arXiv:1807.02447, arXiv:2104.12048, arXiv:2107.05795, arXiv:2412.15207, arXiv:2501.01718, arXiv:2503.07606).