Understanding Latent-Dimension Scaling in Dynamical-System Learning through Spectral Reliability
The study compared two methods for learning time evolution in chaotic systems using a Koopman autoencoder: latent-prediction loss versus spectral-residual loss. The spectral-residual model consistently reduced rollout error by a larger factor as the latent dimension increased, suggesting that using residual information from eigenpairs provides more reliable predictions than just predicting in the latent space.