On Matching Pursuit and Coordinate Descent

[edit]

Francesco Locatello, Anant Raj, Sai Praneeth Karimireddy, Gunnar Raetsch, Bernhard Schölkopf, Sebastian Stich, Martin Jaggi ;
Proceedings of the 35th International Conference on Machine Learning, PMLR 80:3198-3207, 2018.

Abstract

Two popular examples of first-order optimization methods over linear spaces are coordinate descent and matching pursuit algorithms, with their randomized variants. While the former targets the optimization by moving along coordinates, the latter considers a generalized notion of directions. Exploiting the connection between the two algorithms, we present a unified analysis of both, providing affine invariant sublinear $O(1/t)$ rates on smooth objectives and linear convergence on strongly convex objectives. As a byproduct of our affine invariant analysis of matching pursuit, our rates for steepest coordinate descent are the tightest known. Furthermore, we show the first accelerated convergence rate $O(1/t^2)$ for matching pursuit and steepest coordinate descent on convex objectives.

Related Material