On the concentration of eigenvalues of random symmetric matrices (Q1852724): Difference between revisions

The eigenvalues of random \(n\)-by-\(n\) matrices \(A\) whose diagonal and upper diagonal entries are independent real random variables of absolute value at most \(1\) are considered and the concentration of the largest eigenvalues is studied. It is proved for every positive integer \(1 \leq s \leq n\) that the probability that the \(s\)-th largest eigenvalue \(\lambda_s(A)\) deviates from its median by more than \(t\) is at most \(4 e^{-t^{2}/32 s^{2}}\). The same estimate of probability is valid for \(\lambda_{n-s+1}(A)\). This proof is based on the inequality of \textit{M. Talagrand} [Publ. Math., Inst. Hautes Étud. Sci. 81, 73-205 (1995; Zbl 0864.60013)] giving general estimates of measure concentration in probability product spaces. A possible generalization of the theorem about the concentration of the largest eigenvalues \(\lambda(A)\) is suggested.