Universality and sharp matrix concentration inequalities (Q6647779)

Let \(Z_1,\dots,Z_n\) be independent \(d\times d\) random matrices with zero mean and \(X=\sum_{i=1}^n Z_i\). One can treat random matrices \(X\) of this form by using matrix concentration inequalities and mimicking the proofs of scalar concentration inequalities, and ignoring the fact that the summands \(Z_i\) are noncommutative. In the current paper, the authors investigate the spectrum of \(X\) by using a \textit{universality phenomena} approach, which reduces the study of the spectrum of sums of independent random matrices to that of Gaussian matrices \(G\).\N\NThe main results of the current paper are asymptotic inequalities that establish closeness both of the spectral distribution of \(X\) and \(G\), and of their spectra. Among many others, one of the main ideas is that universality arises through an operator-theoretical mechanism, which together with nonstandard concentration inequalities for spectral statistics, provide several applications which are not captured by classical random matrix models. The main results, which are of independent interest, have a range of applications in random graphs and expanders, matrix concentration inequalities for smallest singular values, sample covariance matrices, and phase transitions in spiked models.