Testing Sparsity over Known and Unknown Bases

Sparsity is a basic property of real vectors that is exploited in a wide variety of applications. In this work, we describe property testing algorithms for sparsity that observe a low-dimensional projection of the input. We consider two settings. In the first setting, for a given design matrix A in R^{d x m}, we test whether an input vector y in R^d equals Ax for some k-sparse unit vector x. Our algorithm projects the input onto O(k \eps^{-2} log m) dimensions, accepts if the property holds, rejects if ||y - Ax|| > \eps for any O(k/\eps^2)-sparse vector x, and runs in time polynomial in m. Our algorithm is based on the approximate Caratheodory's theorem. Previously known algorithms that solve the problem for arbitrary A with qualitatively similar guarantees run in exponential time. In the second setting, the design matrix A is unknown. Given input vectors y_1, y_2,...,y_p in R^d whose concatenation as columns forms Y in R^{d x p} , the goal is to decide whether Y=AX for matrices A in R^{d x m} and X in R^{m x p} such that each column of X is k-sparse, or whether Y is "far" from having such a decomposition. We give such a testing algorithm which projects the input vectors to O(log p/\eps^2) dimensions and assumes that the unknown A satisfies k-restricted isometry. Our analysis gives a new robust characterization of gaussian width in terms of sparsity.