Fast Algorithms for lp-Regression and Other Problems


Abstract

Regression in lp-norm is a canonical problem in optimization, machine learning and theoretical computer science which asks for a vector with minimum lp-norm in a linear subspace. In this thesis, we give fast, high-accuracy algorithms for lp regression for all pgreater than 2. Our algorithms are built on novel techniques including iterative refinement for lp-norms, and a fast width-reduced multiplicative weight update routine among others. Furthermore, via a new inverse maintenance scheme, we can solve lp-regression almost as fast as linear regression for all p greater than 2. For graph instances, where the goal is to optimize lp-norm objectives (flow or voltages), our algorithms incorporate sparsification routines that preserve such objectives, to obtain near-linear-time algorithms. As a consequence, we also obtain a fast algorithm for the maximum flow problem in unweighted graphs. A popular approach to fast algorithms for lp-regression in practice is via empirical modifications of the {Iteratively Reweighted Least Squares (IRLS)} algorithm since the basic version of IRLS is known to converge only for p in (1, 3) and diverge even for moderate p (e.g., p > 3.5). We give the first IRLS algorithm that provably converges to a high accuracy solution in a few iterations. Our algorithm has shown exemplary practical performance as well beating standard MATLAB CVX implementations by 10-50x. We further extend our techniques to obtain fast algorithms for Quasi-self-concordant (QSC) functions, which includes lp-regression, softmax regression, and logistic regression among others. We show how iterative refinement and width reduction techniques extend and give fast rates of convergence for all QSC functions.

Publication
PhD Thesis, University of Toronto
Date
Links