Generalized Nullspace Property for Structurally Sparse Signals

We propose a new framework for studying the exact recovery of signals with structured sparsity patterns, which generalizes the well-known nullspace property for standard sparse recovery. We show that for each dictionary there is a maximum sparsity pattern---described by a mathematical object called an "abstract simplicial complex"---that can be recovered via $\ell_1$-minimization. We provide two different characterizations of this maximum sparsity pattern, one based on extreme points and the other based on vectors of minimal support. In addition, we show how this new framework is useful in the study of sparse recovery problems when the dictionary takes the form of a graph incidence matrix or a partial discrete Fourier transform. In both cases we successfully characterize the collection of all support sets for which exact recovery via $\ell_1$-minimization is possible. As a by product, we show that when the dictionary is an incidence matrix, exact recovery can be certified in polynomial time, although this is known to be NP-hard for general matrices.