Alternating Proximity Mapping Method for Convex-Concave Saddle-Point Problems

Abstract

We proposed an iterate scheme for solving convex-concave saddle-point problems associated with general convex-concave functions. We demonstrated that when our iterate scheme is applied to a special class of convex-concave functions, which are constructed by a bilinear coupling term plus a difference of two convex functions, it becomes a generalization of several popular primal-dual algorithms from constant involved parameters to involved parameters as general sequences. For this specific class of convex-concave functions, we proved that the sequence of function values, taken over the averages of iterates generated by our scheme, converges to the value of the function at a saddle-point. Additionally, we provided convergence results for both the sequence of averages of our iterates and the sequence of our iterates. In our numerical experiments, we implemented our algorithm in a matrix game, a linear program in inequality form, and a least-squares problem with $\ell_{1}$ regularization. In these examples, we also compared our algorithm with other primal-dual algorithms where parameters in their iterate schemes were kept constant. Our experimental results not only validated our theoretical findings but also demonstrated that our algorithm consistently outperforms various iterate schemes with constant involved parameters.