Nonconvex penalties with analytical solutions for one-bit compressive sensing

Abstract

One-bit measurements widely exist in the real world, and they can be used to recover sparse signals. This task is known as the problem of learning halfspaces in learning theory and one-bit compressive sensing (1bit-CS) in signal processing. In this paper, we propose novel algorithms based on both convex and nonconvex sparsity-inducing penalties for robust 1bit-CS. We provide a sufficient condition to verify whether a solution is globally optimal or not. Then we show that the globally optimal solution for positive homogeneous penalties can be obtained in two steps: a proximal operator and a normalization step. For several nonconvex penalties, including minimax concave penalty (MCP), _0 norm, and sorted _1 penalty, we provide fast algorithms for finding the analytical solutions by solving the dual problem. Specifically, our algorithm is more than 200 times faster than the existing algorithm for MCP. Its efficiency is comparable to the algorithm for the _1 penalty in time, while its performance is much better. Among these penalties, the sorted _1 penalty is most robust to noise in different settings.

Tasks

Reproductions