Reinforcement Learning to Harness Approximation Errors for Long-Time Quantum Simulation

What the math gives to ML

The paper presents a useful control principle for approximate long-horizon dynamics: do not minimize local discretization error independently, but select future step sizes so that errors accumulated earlier can be compensated later. The controller observes low-dimensional invariants, such as energy and energy variance, rather than an inaccessible exact trajectory, and controls one scalar step size at a time. This transfers most naturally to neural world-model rollouts and Hamiltonian or approximately conservative learned dynamics, where invariant drift is available as a target-free diagnostic. A practical adaptation is an RL or offline-policy controller for rollout step sizes, trained on short trajectories and evaluated by long-horizon invariant preservation and task return.