Matrix concentration inequalities and free probability (Q6072325)

Let \(X\) be a \(d\times d\) Gaussian random matrix. The authors establish non-asymptotic bounds on the spectrum of such a random matrix \(X\) with improved behaviour in terms of the dimension \(d\) in the non-commutative case compared to the usual Khintchine inequality. These bounds reflect the extent to which the spectrum of \(X\) differs from that of a corresponding model \(X_{\text{free}}\) arising in free probability, and make explicit the cases in which these two spectra are the same to leading order. They further lead to bounds on various spectral statistics of \(X\) in terms of \(X_{\text{free}}\), and linearization allows the authors to establish strong asymptotic freeness for a large class of random matrices. A universality principle extends results on Gaussian matrices to the case of sums of independent random matrices.