Abstract. We develop a novel, fundamental, and surprisingly simple randomized iterative method for solving consistent linear systems. Our method has six different but equivalent interpretations: sketch-and-project, constrain-and-approximate, random intersect, random linear solve, random update, and random fixed point. By varying its two parameters-a positive definite matrix (defining geometry), and a random matrix (sampled in an independent and identically distributed fashion in each iteration)-we recover a comprehensive array of well-known algorithms as special cases, including the randomized Kaczmarz method, randomized Newton method, randomized coordinate descent method, and random Gaussian pursuit. We naturally also obtain variants of all these methods using blocks and importance sampling. However, our method allows for a much wider selection of these two parameters, which leads to a number of new specific methods. We prove exponential convergence of the expected norm of the error in a single theorem, from which existing complexity results for known variants can be obtained. However, we also give an exact formula for the evolution of the expected iterates, which allows us to give lower bounds on the convergence rate.Key words. linear systems, stochastic methods, iterative methods, randomized Kaczmarz, randomized Newton, randomized coordinate descent, random pursuit, randomized fixed point AMS subject classifications. 15A06, 15B52, 65F10, 68W20, 65N75, 65Y20, 68Q25, 68W40, 90C20
DOI. 10.1137/15M10254871. Introduction. The need to solve linear systems of equations is ubiquitous in essentially all quantitative areas of human endeavor, including industry and science. Linear systems are a central problem in numerical linear algebra and play an important role in computer science, mathematical computing, optimization, signal processing, engineering, numerical analysis, computer vision, machine learning, and many other fields. For instance, in the field of large scale optimization, there is a growing interest in inexact and approximate Newton-type methods for [7,11,1,40,39,13], which can benefit from fast subroutines for calculating approximate solutions of linear systems. In machine learning, applications arise for the problem of finding optimal configurations in Gaussian Markov random fields [32], in graph-based semisupervised learning and other graph-Laplacian problems [2], in least-squares SVMs, in Gaussian processes, and in others.In a large scale setting, direct methods are generally not competitive when compared to iterative approaches. While classical iterative methods are deterministic, recent breakthroughs suggest that randomization can play a powerful role in the design and analysis of efficient algorithms [38,19,22,9,41,18,21,29] which are in many situations competitive or better than existing deterministic methods.