Scaled-Gradient Hard Thresholding Algorithms for Sparse Signal Recovery

Abstract

Thresholding methods provide a simple iterative framework for sparse signal recovery by combining a search direction with a thresholding operation. Gradient-based search directions are computationally efficient because they require only first-order information for the error metric. We incorporate componentwise gradient scaling into this framework by preserving large-magnitude components and attenuating the remaining components of the gradient of error metric. Combining this strategy with hard thresholding yields scaled-gradient hard thresholding (SGHT) and its pursuit variant SGHTP. Under suitable restricted isometry conditions, we establish recovery error bounds for both methods that account for measurement noise and deviations from exact sparsity. Numerical experiments indicate that SGHTP achieves recovery performance comparable to established pursuit methods on the tested ECG signals.