ECME Thresholding Methods for Sparse Signal Reconstruction

Abstract

We propose a probabilistic framework for interpreting and developing hard thresholding sparse signal reconstruction methods and present several new algorithms based on this framework. The measurements follow an underdetermined linear model, where the regression-coefficient vector is the sum of an unknown deterministic sparse signal component and a zero-mean white Gaussian component with an unknown variance. We first derive an expectation-conditional maximization either (ECME) iteration that guarantees convergence to a local maximum of the likelihood function of the unknown parameters for a given signal sparsity level. To analyze the reconstruction accuracy, we introduce the minimum sparse subspace quotient (SSQ), a more flexible measure of the sampling operator than the well-established restricted isometry property (RIP). We prove that, if the minimum SSQ is sufficiently large, ECME achieves perfect or near-optimal recovery of sparse or approximately sparse signals, respectively. We also propose a double overrelaxation (DORE) thresholding scheme for accelerating the ECME iteration. If the signal sparsity level is unknown, we introduce an unconstrained sparsity selection (USS) criterion for its selection and show that, under certain conditions, applying this criterion is equivalent to finding the sparsest solution of the underlying underdetermined linear system. Finally, we present our automatic double overrelaxation (ADORE) thresholding method that utilizes the USS criterion to select the signal sparsity level. We apply the proposed schemes to reconstruct sparse and approximately sparse signals from tomographic projections and compressive samples.