On Model-Based RIP-1 Matrices

The Restricted Isometry Property (RIP) is a fundamental property of a matrix enabling sparse recovery [5]. Informally, an m ×n matrix satisfies RIP of order k in the ℓ p norm if ∥ Ax ∥ p ≈ ∥ x ∥ p for any vector x that is k-sparse, i.e., that has at most k non-zeros. The minimal number of rows m necessary for the property to hold has been extensively investigated, and tight bounds are known. Motivated by signal processing models, a recent work of Baraniuk et al [3] has generalized this notion to the case where the support of x must belong to a given model, i.e., a given family of supports. This more general notion is much less understood, especially for norms other than ℓ2. In this paper we present tight bounds for the model-based RIP property in the ℓ1 norm. Our bounds hold for the two most frequently investigated models: tree-sparsity and block-sparsity. We also show implications of our results to sparse recovery problems.National Science Foundation (U.S.) (NSF CCF-1065125 award)Danish National Research Foundation (Center for Massive Data Algorithmics (MADALGO))David & Lucile Packard FoundationAkamai Technologies, Inc. (Presidential Fellowship