Normalized Gradients for All

Abstract

In this short note, I show how to adapt to Hölder smoothness using normalized gradients in a black-box way. Moreover, the bound will depend on a novel notion of local Hölder smoothness. The main idea directly comes from Levy [2017].