A parallel-in-time preconditioner for Crank-Nicolson discretization of a parabolic optimal control problem

TL;DR

This paper introduces a novel parallel-in-time preconditioner for efficiently solving saddle point systems from Crank-Nicolson discretized parabolic optimal control problems, achieving convergence rates independent of problem size.

Contribution

A new symmetrization and preconditioning technique for saddle point systems from Crank-Nicolson discretization, enabling fast, parallel-in-time iterative solutions with problem-independent convergence.

Findings

Eigenvalues of the preconditioned system are bounded independently of size.

The PCG solver converges at a rate independent of matrix size and regularization.

Numerical results demonstrate the effectiveness of the proposed preconditioner.

Abstract

In this paper, a fast solver is studied for saddle point system arising from a second-order Crank-Nicolson discretization of an initial-valued parabolic PDE constrained optimal control problem, which is indefinite and ill-conditioned. Different from the saddle point system arising from the first-order Euler discretization, the saddle point system arising from Crank-Nicolson discretization has a dense and non-symmetric Schur complement, which brings challenges to fast solver designing. To remedy this, a novel symmetrization technique is applied to the saddle point system so that the new Schur complement is symmetric definite and the well-known matching-Schur-complement (MSC) preconditioner is applicable to the new Schur complement. Nevertheless, the new Schur complement is still a dense matrix and the inversion of the corresponding MSC preconditioner is not parallel-in-time (PinT) and thus time consuming. For this concern, a modified MSC preconditioner for the new Schur complement system. Our new preconditioner can be implemented in a fast and PinT way via a temporal diagonalization technique. Theoretically, the eigenvalues of the preconditioned matrix by our new preconditioner are proven to be lower and upper bounded by positive constants independent of matrix size and the regularization parameter. With such spectrum, the preconditioned conjugate gradient (PCG) solver for the Schur complement system is proven to have a convergence rate independent of matrix size and regularization parameter. To the best of my knowledge, it is the first time to have an iterative solver with problem-independent convergence rate for the saddle point system arising from Crank-Nicolson discretization of the optimal control problem. Numerical results are reported to show that the performance of the proposed preconditioner.