Iterative $\ell_1$ minimization for non-convex compressed sensing

Penghang Yin, Jack Xin

Published 2016-04-27Version 1

An algorithmic framework, based on the difference of convex functions algorithm, is proposed for minimizing a class of concave sparse metrics for compressed sensing problems. The resulting algorithm iterates a sequence of $\ell_1$ minimization problems. An exact sparse recovery theory is established to show that the proposed framework always improves on the basis pursuit ($\ell_1$ minimization) and inherits robustness from it. Numerical examples on success rates of sparse solution recovery illustrate further that, unlike most existing non-convex compressed sensing solvers in the literature, our method always out-performs basis pursuit, no matter how ill-conditioned the measurement matrix is. Moreover, the iterative $\ell_1$ algorithms lead by a wide margin the state-of-the-art algorithms on $\ell_{1/2}$ and logarithimic minimizations in the strongly coherent (highly ill-conditioned) regime, despite the same objective functions.