Invertibility for non-Hermitian random band matrices with sublinear bandwidth and discrete entries
Yi HAN · 2025
Read Invertibility for non-Hermitian random band matrices with sublinear bandwidth and discrete entries by Yi HAN (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05417.txt.
Introduction
Yi Han proves when band shaped random matrices can be flipped in a 2025 paper. A band matrix has entries just near its diagonal spine. When the band is thin next to the size, flip proof was a known open gap.
A classic result says full plus minus one matrices flip with near sure odds. Han shows band ones do too, when the band is not too thin. The entries can sit on a shifted plus minus set as well. Past proof held just for block band shapes. The result extends to mirror band matrices too, a step in random matrix theory.
Catalog record
| CREATOR | Yi HAN |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-05417
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Invertibility for non-Hermitian random band matrices with sublinear bandwidth and discrete entries Yi HAN Abstract . A well-known result in random matrix theory, proven by Kahn, Komlós and Szemerédi in 1995, states that a square random matrix with i.i.d. uniform { ± 1 } \{\pm 1\} entries is invertible with probability 1 − exp ( − Ω ( n ) ) 1-\exp(-\Omega(n)) . As a natural generalization of the model and motivated by the Anderson localization-delocalization transition in mathematical physics, we consider the invertibility of a class of random band matrices with independent entries where the bandwidth d n d_{n} scales like n α n^{\alpha} , for some α ∈ ( 0 , 1 ) \alpha\in(0,1) . The band matrix model we consider is sufficiently general and covers existing models such as the block band matrix and periodic band matrix, allowing great flexibility in the variance profile. As the bandwidth is sublinear in the dimension, estimating the invertibility and least singular values of these matrices is a well-known open problem. We make a progress towards the invertibility problem by showing that, when α > 2 3 \alpha>\frac{2}{3} and when the random variables are i.i.d. uniformly distributed on { ± 1 } \{\pm 1\} , then the band matrix is invertible with probability 1 − exp ( − Ω ( n α / 2 ) ) 1-\exp(-\Omega(n^{\alpha/2})) . Previously, even invertibility with probability 1 − o ( 1 ) 1-o(1) was not known for these band matrix models except in the very special case of block band matrices. 1. Introduction In this paper, we investigate the invertibility for a wide class of combinatorial random matrices with an inhomogeneous variance profile. Before stating the main result on band matrices, we begin with a historical account on the invertibility of n\times n random matrices with i.i.d. entries, which is much better understood in the past thirty years. Let M_{n} be an n\times n matrix with each entry independently and uniformly distributed on \{-1,1\} , and denote by P_{n}:=\mathbb{P}(\det M_{n}=0) . The first proof of asymptotic non-singularity, that \lim_{n\to\infty}P_{n}=0 , was given by Komlós [ 25 ] . Then Kahn, Komlós and Szemerédi [ 23 ] provided the first exponential bound on P_{n} : P_{n}\leq c^{n} for n sufficiently large, and where they can take c=0.999 . Tao and Vu [ 31 ] improved the value of c to be c=3/4+o(1) , and Bourgain, Vu, Wood [ 5 ] further improved to c=\sqrt{2}/2+o(1) . Finally, Tikhomirov [ 35 ] obtained the optimal value c=\frac{1}{2}+o(1) . In parallel lines, much recent progress has been made for random matrices with an inhomogeneous variance profile, in the dense case where the matrix typically has \Theta(n^{2}) nonzero entries. See [ 29 ] for least singular value estimates in the Gaussian case and see [ 12 ] , [ 10 ] , [ 13 ] for the general case where the circular law was proven when the matrix has a doubly stochastic variance profile. However, the case of sparse structured matrices was not well understood, and proving the convergence of its empirical spectral measure was listed as an open problem in the survey [ 34 ] . For sparse structured random matrices, there are essentially only two recent papers [ 22 ] and [ 33 ] studying the invertibility and pseudospectrum. The paper [ 22 ] by Jain, Jana, Luh and O’Rourke considered a very special type of matrix model called block band matrices (see Definition 1.1 ), and proved convergence of empirical eigenvalue density for bandwidth d_{n}\gg n^{33/34} utilizing an estimate they developed for the least singular value. Notably, the work [ 22 ] covers discrete entry laws such as the uniform distribution on \{\pm 1\} . However, the special structure of block band matrices is utilized in an essential way, resisting a generalization to more general models such as the periodic band matrices in Definition 1.3 . Another recent paper by Tikhomirov [ 33 ] obtained minimal singular value estimates for (the shifted version of) a wide class of random matrices with general variance profile, but assuming the random variables have a bounded density. Circular law for periodic band matrices with Gaussian entry distribution was proven as a corollary, for bandwidth up to d_{n}\gg n^{33/34} . The range of d_{n} for circular law to hold was improved in a recent preprint [ 21 ] . However, [ 33 ] completely left open the case where the nonzero entries are discrete valued, say over \{\pm 1\} . And to our best knowledge, even asymptotic nonsingularity for general random band matrices with discrete entry distribution is completely unexplored. (The spectral outliers for random band matrices are much better studied; see [ 1 ] and [ 20 ] ). The objective of this paper is to take a step into the invertibility problem for general non-Hermitian random band matrices with discrete entry distribution. We will study a class of square random matrices A_{n}=(a_{ij})_{1\leq i,j\leq n} with independent entries, and that nonzero random variables are attached along a band close to the main diagonal, with the bandwidth d_{n} sublinearly growing in n . We can also attach arbitrary integer-valued random variables at locations away from this band. Our model covers block band and periodic band matrices, although being slightly less general than the setting of [ 33 ] where they consider any doubly stochastic variance profile. Random band matrices, at least for the Hermitian version, are good models interpolating the de-localized phase of GOE/GUE ensembles with mean-field interaction, and the localized phase where Anderson localization takes place. The threshold for localization-delocalization transition is conjectured to be d_{n}\sim\sqrt{n} . A very recent breakthrough [ 36 ] confirmed that when d_{n}\gg n^{1/2+c} , the Hermitian random band matrix exhibits complete eigenvalue delocalization, universality of local eigenvalue statistics and quantum unique ergodicity, among other things. The localized phase where d_{n}\ll\sqrt{n} is also recently studied in [ 9 ] , [ 8 ] , [ 19 ] . A more comprehensive survey on progress in (Hermitian) random band matrices can be found in [ 4 ] and [ 30 ] . Although none of these results on Hermitian band matrices are proven for the non-Hermitian model, we can make the following postulate. The fact that a square matrix with discrete \{\pm 1\} entry is invertible with very high probability stems from the fact that we are in the delocalized phase with strong mean-field interaction. This can be contrasted with random Schrödinger operators in the localized phase (see for example [ 24 ] ), where the Wegner estimate, a quantitative version of invertibility, can generally be established only when the random potentials have a bounded density. Now we turn back to the discussion of non-Hermitian band matrices. When the band matrix entries have a bounded density, it is possible (and not hard) to prove that the matrix is almost surely invertible for bandwidth d_{n} much smaller than n^{1/2} , but it becomes highly questionable whether the invertibility persists when the entries are discrete over \{\pm 1\} . To restate the observation, we can use the fact that a non-Hermitian random band matrix with discrete entries is invertible with probability 1-\exp(-\Omega(n^{c})) for