Sparse recovery with partial support knowledge

Abstract

The goal of sparse recovery is to recover the (approximately) best k-sparse approximation [ˆ over x] of an n-dimensional vector x from linear measurements Ax of x. We consider a variant of the problem which takes into account partial knowledge about the signal. In particular, we focus on the scenario where, after the measurements are taken, we are given a set S of size s that is supposed to contain most of the “large” coefficients of x. The goal is then to find [ˆ over x] such that [ ||x-[ˆ over x]|| [subscript p] ≤ C min ||x-x'||[subscript q]. [over] k-sparse x' [over] supp (x') [c over _] S] We refer to this formulation as the sparse recovery with partial support knowledge problem ( SRPSK ). We show that SRPSK can be solved, up to an approximation factor of C = 1 + ε, using O( (k/ε) log(s/k)) measurements, for p = q = 2. Moreover, this bound is tight as long as s = O(εn / log(n/ε)). This completely resolves the asymptotic measurement complexity of the problem except for a very small range of the parameter s. To the best of our knowledge, this is the first variant of (1 + ε)-approximate sparse recovery for which the asymptotic measurement complexity has been determined.