Inexact asymmetric forward-backward-adjoint splitting algorithms for saddle point problems

Adopting a suitable approximation strategy can both enhance the robustness and improve the efficiency of the numerical algorithms. In this paper, we suggest combining two approximation criteria, the absolute one and the relative one, to the asymmetric forward-backward-adjoint splitting (AFBA) algorithm for a class of convex-concave saddle point problems, resulting in two inexact AFBA variants. These two approximation criteria are low-cost, since verifying them just involves the subgradient of a certain function. For both the absolute error AFBA and the relative error AFBA, we establish the global convergence and the O(1/N) convergence rate measured by the gap function in the ergodic sense, where N is the number of iterations. For the absolute error AFBA, we show that it possesses an O(1/N2) (linear convergence) rate of convergence, under the assumption that a part of (both) the underlying functions are strongly convex. We report some numerical results which demonstrate the effectiveness of the proposed methods.