Random Matrices, Intrinsic Freeness, and Sharp Non-Asymptotic Inequalities
Afonso S. Bandeira
TL;DR
This work develops a non-asymptotic framework around intrinsic freeness, where a Gaussian matrix X = A0 + sum g_k A_k is shown to be spectrally close to its free counterpart X_free, with quantitative bounds that depend on parameters sigma(X), v(X), sigma_*(X) via tilde_v(X) = sqrt(v(X) sigma(X)). By coupling Gaussian interpolation with matrix-chaos methods and a universality principle, the paper provides sharp, easy-to-use spectral bounds for general random matrices, including dependent and non-isotropic models. It extends these ideas to sharp phase transitions, iterated matrix concentration, and concrete applications in Tensor PCA through Kikuchi matrices and SOS analysis, culminating in a resolution of aspects of the Matrix Spencer conjecture in a broad regime. Overall, the results translate free-probability insights into practical, non-asymptotic tools with wide implications for high-dimensional statistics and computation.
Abstract
Random matrix theory has played a major role in several areas of pure and applied mathematics, as well as statistics, physics, and computer science. This lecture aims to describe the intrinsic freeness phenomenon and how it provides new easy-to-use sharp non-asymptotic bounds on the spectrum of general random matrices. We will also present a couple of illustrative applications in high dimensional statistical inference. This article accompanies a lecture that will be given by the author at the International Congress of Mathematicians in Philadelphia in the Summer of 2026.
