Canonical Latent Dynamics Representation
Background: Latent representations of dynamical systems derived from high-dimensional data are defined up to smooth invertible reparameterizations (gauge freedom).
Question / Future Work: Defining a canonical, universal form for learned latent dynamics is necessary to permit quantitative comparison of models and to confirm physical equivalence between disparate learned embeddings. Future work must determine whether simple techniques like eigen-decomposition of transition operators suffice or if complex structures such as normalizing flows are required to resolve this indeterminacy.
Why It Matters: This is a critical bottleneck for the validation, benchmarking, and interpretability of data-driven dynamical modeling across arbitrary experiments.
Evidence: A canonical form into which any equivalent representation could be transformed would resolve this. … Whether this normal-mode form is enough in general, or whether a more flexible construction such as a normalizing flow on top of the encoder is needed, remains open.
Metadata & Links
- created_at
- 2026-04-30T06:03:59Z