Title: Fast convergence of Newton-type methods on high-dimensional problems
Authors: Yuekai Sun - University of Michigan (United States) [presenting]
Abstract: The convergence rate of Newton-type methods on high-dimensional problems is studied. The high-dimensional nature of the problem precludes the usual global strong convexity and smoothness that underlie the classical analysis of such methods. We find that restricted version of these conditions which typically arise in the study of the statistical properties of the solutions are also enough to ensure good computational properties of Newton-type methods. We explore the algorithmic consequences in distributed and online settings.