Operator learning struggles with periodic trajectories without phase alignment
Why Directly Learning Periodic Trajectories Can Fail
Artificial Intelligence
Summary
Predicting repeating patterns in complex systems is tricky because small timing differences add up and make the patterns look different. The authors show that simply recording a full cycle of these patterns without adjusting their starting points often leads to poor predictions. By aligning these cycles based on their phase and learning the shape and period separately, predictions become more accurate. This approach works well on both ordinary and partial differential equations, helping models better capture periodic behavior.
What this means in practice
- •For engineers simulating physical systems: Improve prediction accuracy of periodic behaviors in simulation models by learning normalized aligned waveforms and periods separately.
- •For signal processing developers: Develop better models for repeating signal reconstruction by separating period estimation from waveform learning to reduce timing errors.
Authors
Kaixin Zheng, Anita Layton
Abstract
Operator learning of periodic solutions requires deciding how simulation data should be recorded and represented. A natural choice is to integrate long enough for transients to decay and record a window wide enough to contain at least one full period of all trajectories. We find that these conservative choices can make the resulting trajectories difficult to learn, even when the underlying periodic orbits vary regularly with system parameters. Unaligned trajectories generalize poorly even within the training distribution. Phase alignment substantially improves in-distribution generalization, but models trained on a fixed physical-time window still have large errors on trajectories with periods outside the training range. We explain both failures through a common mechanism: frequency differences accumulate over time, so the target phase varies rapidly with the parameters. Predictors that cannot track this variation incur a population MSE floor in both settings; for fixed window prediction, we also derive a per-sample lower bound. We then study one of the simplest representations that escape these floors: learning an aligned, normalized waveform and its period separately. We establish regularity of the decoupled targets under ODE assumptions and show experimentally that this approach avoids both failures in ODE systems and a PDE case study.