New Variance-Reduced Zero-Order Hard-Thresholding Algorithm for Sparse Optimization

A new paper on arXiv (2605.18035) addresses limitations of the SZOHT algorithm for ℓ0 constrained optimization with zeroth-order gradients. The authors identify a conflict between gradient deviation and hard-thresholding operator expansivity. They propose a generalized variance-reduced ZO hard-thresholding algorithm to mitigate this conflict, improving convergence. The work provides new theoretical insights into variance reduction for sparse optimization.