Improved Low-Complexity Sparse Bayesian Learning With Embedded Bayesian Threshold

Abstract

Sparse Bayesian Learning (SBL) is recognized for its efficacy in sparse signal recovery, the computational demand escalates significantly with increasing data dimensionality due to the matrix inversion at each iteration. An Inverse-Free sparse Bayesian Learning (IF-SBL) approach has been introduced to mitigate computational complexity. However, IF-SBL converges easily to a sub-optimal solution with false peaks due to the neglect of the correlation between atoms. In this paper, we analyze causes of false peaks in IF-SBL. Subsequently, a novel dynamically updated embedded Bayesian threshold is designed to mitigate the interference caused by false peaks. This innovative approach retrieves the stability and reliability without significantly increasing signal recovery complexity compared with IF-SBL. Simulation experiments validate the results.