Long-Time Trajectory Approximation via SA-NODEs: Model Predictive and Floquet Strategies
arXiv:2608.10738
2026
Dynamics
2 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper identifies state resets as a principled way to prevent long-horizon neural ODE error bounds from becoming double-exponential in time. Its model-predictive construction adaptively chooses windows using the first time the learned flow exceeds a uniform tolerance, then trains a separate warm-started vector field per window; this allocates capacity where prediction is difficult while avoiding accumulated rollout error during training. For approximately periodic systems, the Floquet construction replaces repeated trajectory fitting with a learned return-map contraction, yielding controlled orbital error across many periods. These ideas transfer most naturally to neural ODEs, latent dynamical models, world models, and recurrent long-rollout predictors.
Ideas from this paper
✗ Failed on benchmark
2026
Replace one neural ODE trained over the entire rollout with a sequence of locally trained vector fields, and reset each window from the observed or teacher state during training. Choose the next window boundary at the first time the current model's supervised flow error exceeds a tolerance, so difficult portions receive shorter windows and more parameters while easy portions use longer windows.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
For systems with a repeating orbit, train a periodic neural dynamical model together with a return map whose transverse deviations contract after each period. Enforce and measure orbital contraction rather than requiring phase-aligned pointwise trajectories to remain close, allowing phase drift while suppressing divergence across many cycles.
Useful7/10
Difficulty6/10
Novelty7/10