On Extremal Eigenvalues of Random Matrices with Gaussian Entries
Abstract
Consider $Z_n=\xi_1A_1+\xi_2A_2+...+\xi_nA_n$ for $\xi_1,\xi_2,\hspace{0.05cm}...\hspace{0.05cm},\xi_n$ i.i.d., $\xi_1\overset{d}{=}N(0,1),$ $A_1,A_2,\hspace{0.05cm}...\hspace{0.05cm},A_n \in \mathbb{R}^{d \times d}$ deterministic and symmetric. Moment bounds on the operator norm of $Z_n$ have been obtained via a matrix version of Markov's inequality (also known as Bernstein's trick). This work approaches these quantities with the aid of Gaussian processes, namely via interpolation alongside a variational definition of extremal eigenvalues. This perspective not only recoups the aforesaid results, but also renders both bounds that reflect a more intrinsic notion of dimension for the matrices $A_1,A_2,\hspace{0.05cm}...\hspace{0.05cm},A_n$ than $d,$ and moment bounds on the smallest (in absolute value) eigenvalue of $Z_n.$