Constrained Estimation via Projected Mirror Descent

Abstract

Constraints on parameter spaces promote various structures in statistical and machine learning tasks. However, they present methodological and computational challenges. These challenges only become more evident in non-Euclidean settings. We advocate for the use of Projected Mirror Descent for constrained online estimation problems in many cases when the parameter space is non-Euclidean. We outline convergence properties of this algorithm and show how it can be connected to other algorithms of interest.

Advisor(s)

Jason Xu