We study verifiable sufficient conditions and computable performance bounds for sparse recovery algorithms such as the Basis Pursuit, the Dantzig selector and the Lasso estimator, in terms of a newly defined family of quality measures for the measurement matrices. With high probability, the developed measures for subgaussian random matrices are bounded away from zero as long as the number of measurements is reasonably large. Comparing to the restricted isotropic constant based performance analysis, the arguments in this paper are much more concise and the obtained bounds are tighter.Numerical experiments are presented to illustrate our theoretical results.
Index TermsCompressive sensing; q-ratio sparsity; q-ratio constrained minimal singular values; Convex-concave procedure. z =0, z 2 1 / z 2 2 ≤s Az 2 z 2 and obtained the error 2 bounds in terms of this quality measure of the measurement matrix. Similarly, in [16], the authors definedz ∞ with · ♦ denoting a general norm, and derived the performance bounds on the ∞ norm of the recovery error vector based on this quality measure. This kind of measures has also been used in establishing results for block sparsity recovery [17] and low-rank matrix recovery [18]. In this paper we generalize these two quantities to a more general quantity called q-ratio CMSV with 1 < q ≤ ∞, and establish the performance bounds for both q norm and 1 norm of the reconstruction error.
A. ContributionsOur contribution mainly has four aspects. First, we proposed a sufficient condition based on a q-ratio sparsity level for the exact recovery using 1 minimization in the noise free case, and designed a convexconcave procedure to solve the corresponding non-convex problem, leading to an acceptable verification algorithm. Second, we introduced q-ratio CMSV and derived concise bounds on both q norm and 1 norm of the reconstruction error for the Basis Pursuit (BP) [19], the Dantzig selector (DS) [20], and the Lasso estimator [21] in terms of q-ratio CMSV. We established the corresponding stable and robust recovery results involving both sparsity defect and measurement error. Third, we demonstrated that for subgaussion random matrices, the q-ratio CMSVs are bounded away from zero with high probability, as long as the number of measurement is large enough. Finally, we presented algorithms to compute the q-ratio CMSV for an arbitrary measurement matrix, and studied the effects of different parameters on the proposed q-ratio CMSV. Moreover, we illustrated that q-ratio CMSV based bound is tighter than the RIC based one.
B. Organization and NotationsThe paper is organized as follows. In Section II, we present the definitions of q-ratio sparsity and q-ratio CMSV, and give a sufficient condition for unique noiseless recovery based on the q-ratio sparsity and an inequality for the q-ratio CMSV. In Section III, we derive performance bounds on both q norm and 1 norm of the reconstruction errors for several convex recovery algorithms in terms of q-ratio CMSVs. In Section IV, we demonstrate that the subgaussi...