Title: Randomized Minimization of Eigenvalue Functions

Abstract: We propose new probabilistic approximations of several eigenvalue functions related to the maximal eigenvalue of a symmetric matrix. For each of these approximations, there exists a simple unbiased stochastic subgradient oracle which is closely related to the standard Power Method for computing the leading eigenvector. We show how to apply these new tools for solving large-scale SDP problems by the Stochastic Subgradient Method using only matrix-vector products. For the resulting algorithm, we establish very attractive efficiency estimates for solving the problem with a given relative accuracy.