Title:Dual-primal balanced augmented Lagrangian method for linearly constrained convex programming

Abstract:The augmented Lagrangian method (ALM) is a fundamental tool for solving the canonical convex minimization problem with linear constraints, and efficiently and easily how to implement the original ALM is affirmatively significant. Recently, He and Yuan have proposed a balanced version of ALM [B.S. He and X.M. Yuan, arXiv:2108.08554, 2021], which reshapes the original ALM by balancing its subproblems and makes the benchmark ALM easier to implement without any additional condition. In practice, the balanced ALM updates the new iterate by a primal-dual order. In this note, exploiting the variational inequality structure of the most recent balanced ALM, we propose a dual-primal version of the balanced ALM for linearly constrained convex minimization problems. The novel proposed method generates the new iterate by a dual-primal order and enjoys the same computational difficulty with the original primal-dual balanced ALM. Furthermore, under the lens of the proximal point algorithm, we conduct the convergence analysis of the novel introduced method in the context of variational inequalities. Numerical tests on the basic pursuit problem demonstrate that the introduced method enjoys the same high efficiency with the prototype balanced ALM.