Multigrid-in-time preconditioners for KKT systems (2405.04808v1)

Abstract: We develop multigrid-in-time preconditioners for Karush-Kuhn-Tucker (KKT) systems that arise in the solution of time-dependent optimization problems. We focus on a specific instance of KKT systems, known as augmented systems, which underpin the composite-step sequential quadratic programming framework [1]. To enable time-domain decomposition, our approach introduces virtual state variables and continuity constraints at each discrete time interval. The virtual state variables not only facilitate a decoupling in time but also give rise to fixed-point iterations that aid the solution of KKT systems. These fixed-point schemes can be used either as preconditioners for Krylov subspace methods or as smoothers for multigrid-in-time schemes. For the latter, we develop a block-Jacobi scheme that parallelizes trivially in the time domain. To complete the multigrid construction, we use simple prolongation and restriction operators based on geometric multigrid ideas, and a coarse-grid solver based on a GMRES iteration preconditioned with the symmetric block Gauss-Seidel scheme. We present two optimal control examples, involving the viscous Burgers' and van der Pol oscillator equations, respectively, and demonstrate algorithmic scalability.