Sparse Recovery: The Square of ₁/₂ Norms

This paper introduces a nonconvex approach to the problem of recovering sparse signals. We propose a novel model, termed the ₂-model, which utilizes the square of ₁/₂ norms for sparse recovery. This model is an advancement over the ₀ norm, which is often computationally intractable and less effective in handling practical scenarios. Our approach is grounded in the concept of effective sparsity, which robustly measures the number of effective coordinates in a signal. We demonstrate that our model is a powerful alternative for sparse signal estimation, with the ₂-model offering computational advantages and practical applicability. The model's formulation and the accompanying algorithm, based on Dinkelbach's procedure combined with a difference of convex functions strategy, are detailed. We further explore the properties of our model, including the existence of solutions under certain conditions, and discuss the algorithm's convergence properties. Numerical experiments with various sensing matrices are conducted to validate the effectiveness of our model.