Sparse space-time spectral methods can time-step by peel and pass

What the math gives to ML

The paper identifies an endpoint-compatible polynomial representation in which the value at the end of a time block is obtained by a coefficient reduction rather than by evaluating or storing the whole trajectory. This suggests a memory-efficient chunked latent-dynamics architecture: represent hidden states within each chunk by temporal polynomial coefficients, pass only the exact endpoint state to the next chunk, and train the coefficients using a learned dynamics residual. The main transferable asset is not the PDE solver itself, but the algebraic value-passing property, which gives a deterministic recurrent interface with controllable temporal approximation error and constant-in-time memory.