Optimization and Control (math.OC)

Abstract: Proximal point algorithms have found numerous applications in the field of convex optimization, and their accelerated forms have also been proposed. However, the most commonly used accelerated proximal point algorithm was first introduced in 1967, and recent studies on accelerating proximal point algorithms are relatively scarce. In this paper, we propose high-resolution ODEs for the proximal point operators for both closed proper convex functions and maximally monotone operators, and present a Lyapunov function framework to demonstrate that the trajectories of our high-resolution ODEs exhibit accelerated behavior. Subsequently, by symplectically discretizing our high-resolution ODEs, we obtain new proximal point algorithms known as symplectic proximal point algorithms. By decomposing the continuous-time Lyapunov function into its elementary components, we demonstrate that symplectic proximal point algorithms possess $O(1/k^2)$ convergence rates.