A Reduced Half Thresholding Algorithm

In this paper, a reduced half thresholding algorithm and its accelerated version, a reduced fast half thresholding algorithm, are proposed for solving large-scale L_1/2 -regularized problems. At each iteration, the large-scale original problem is reduced into a sequence of small-scale subproblems, and the subproblems are solved by the (fast) half thresholding algorithm. Global convergence of the reduced half thresholding algorithm is proven, and two techniques: the Newton acceleration technique and the shrinking technique are presented to improve the performance. Compared with the state-of-the-art algorithms, the numerical results show that the presented algorithms are promising. As an example, our algorithms are efficient to recover the signal recovery problem with 19,264,097 samples and signal length 29,890,095