Quadratically fast IRLS for sparse signal recovery

We present a new class of iterative algorithms for sparse recovery problems that combine iterative support detection and estimation. More precisely, these methods use a two state Gaussian scale mixture as a proxy for the signal model and can be interpreted both as iteratively reweighted least squares (IRLS) and Expectation/Maximization (EM) algorithms for the constrained maximization of the log-likelihood function. Under certain conditions, these methods are proved to converge to a sparse solution and to be quadratically fast in a neighborhood of that sparse solution, outperforming classical IRLS for lp-minimization. Numerical experiments validate the theoretical derivations and show that these new reconstruction schemes outperform classical IRLS for lp-minimization with p\in(0,1] in terms of rate of convergence and sparsity-undersampling tradeoff