Algorithms and lower bounds for sparse recovery

Abstract

We consider the following k-sparse recovery problem: design a distribution of m x n matrix A, such that for any signal x, given Ax with high probability we can efficiently recover x satisfying IIx - x l, </-Cmink-sparse x' IIx - x'II. It is known that there exist such distributions with m = O(k log(n/k)) rows; in this thesis, we show that this bound is tight. We also introduce the set query algorithm, a primitive useful for solving special cases of sparse recovery using less than 8(k log(n/k)) rows. The set query algorithm estimates the values of a vector x [epsilon] Rn over a support S of size k from a randomized sparse binary linear sketch Ax of size O(k). Given Ax and S, we can recover x' with IIlx' - xSII2 </- [theta]IIx - xsII2 with probability at least 1 - k-[omega](1). The recovery takes O(k) time. While interesting in its own right, this primitive also has a number of applications. For example, we can: * Improve the sparse recovery of Zipfian distributions O(k log n) measurements from a 1 + [epsilon] approximation to a 1 + o(1) approximation, giving the first such approximation when k </- O(n1-[epsilon]). * Recover block-sparse vectors with O(k) space and a 1 + [epsilon] approximation. Previous algorithms required either w(k) space or w(1) approximation.