Scaled proximal gradient methods for sparse optimization problems (Q6184263)

The authors propose a scaled proximal pursuit and a modified scaled proximal gradient method for solving the sparse optimization problem by taking the scaling matrix as a varying positive diagonal matrix, which is associated with residual reduction. The involved subproblems are nonconvex and non-smooth maximization problems with respect to the diagonal scaling matrix. Their relaxed versions are solved by updating the diagonal entries of the scaling matrix one by one deterministically or randomly. The global convergence of the proposed method is established under a certain assumption that the level set is bounded without imposing the restricted isometry property on the measurement matrix. A scaled proximal pursuit and a modified scaled proximal gradient method is also presented. Their global convergence is established under the assumption that the measurement matrix satisfies the restricted isometry property bound. Some numerical experiments are reported to illustrate the efficiency of the proposed algorithms for solving the sparse optimization problem over other thresholding-based algorithms such as the iterative hard thresholding algorithm and the hard thresholding pursuit algorithm.