SCO Seminar - Zepeng Zhang, University of Groningen

Title: How to accelerate a fast gradient method?

Abstract:

For strongly convex functions, a convergence speedup upon Nesterov's fast gradient method has been shown possible. But an intuitive explanation is still missing. In this talk, we will show, from the perspective of inertial dynamics, why this is possible. To validate our argument, we apply our insights to the inertial algorithms, where accelerated convergence rates are obtained. We also extend our analysis to a slightly different inertial system with asymptotic vanishing damping, and proximal gradient algorithms. In both cases, we derive faster provable convergence results.