Abstract
We study the self-normalized concentration of vector-valued stochasticprocesses. We focus on bounds for sub-$\psi$ processes, a tail condition thatencompasses a wide variety of well-known distributions (includingsub-exponential, sub-Gaussian, sub-gamma, and sub-Poisson distributions). Ourresults recover and generalize the influential bound of Abbasi-Yadkori et al.(2011) and fill a gap in the literature between determinant-based bounds andthose based on condition numbers. As applications we prove a Bernsteininequality for random vectors satisfying a moment condition (which is moregeneral than boundedness), and also provide the first dimension-free,self-normalized empirical Bernstein inequality. Our techniques are based on thevariational (PAC-Bayes) approach to concentration.