Randomized Symplectic Model Reduction: One-Pass Basis Construction and Sketched Hamiltonian Assembly

Abstract

Symplectic model order reduction of Hamiltonian systems yields reduced dynamics that conserve the reduced Hamiltonian along exact trajectories, avoiding the artificial energy drift that can arise from conventional projection methods. For large-scale spatial discretizations, however, basis construction and reduced-operator assembly can involve substantial offline costs. Randomized numerical linear algebra can reduce these costs through low-dimensional sketches while supporting streaming and parallel implementations. We develop randomized techniques for two parts of the offline stage while preserving the Hamiltonian structure of the reduced dynamics. First, we introduce the one-pass sketched randomized cSVD (Ops-rcSVD), a one-pass randomized variant of the complex-SVD algorithm. Combining randomized range finding with an independent left sketch, Ops-rcSVD constructs an ortho-symplectic reduced basis from a single pass over the snapshot ensemble. Second, for quadratic Hamiltonian systems, we introduce a randomized reduced-Hamiltonian assembly based on sketching the reduced basis and operator actions, yielding a reduced Hamiltonian system by construction. We derive bounds for the snapshot projection error, uniform-in-parameter perturbation bounds for the sketched reduced Hamiltonian, and an energy-norm estimate for the difference between reduced and sketched reduced trajectories. Experiments on large-scale systems show that Ops-rcSVD attains projection accuracy comparable to deterministic and two-pass randomized constructions while reducing basis-construction time. They further demonstrate accurate sketched reduced-Hamiltonian assembly when storing the complete snapshot ensemble is no longer feasible.