Given a "data manifold" M⊂Rn and "latent space" Rℓ, an autoencoder is a pair of continuous maps consisting of an "encoder" E:Rn→Rℓ and "decoder" D:Rℓ→Rn such that the "round trip" map D∘E is as close as possible to the identity map \mboxidM on M. We present various topological limitations and capabilites inherent to the search for an autoencoder, and describe capabilities for autoencoding dynamical systems having M as an invariant manifold.
Cite
@article{arxiv.2511.04807,
title = {Autoencoding Dynamics: Topological Limitations and Capabilities},
author = {Matthew D. Kvalheim and Eduardo D. Sontag},
journal= {arXiv preprint arXiv:2511.04807},
year = {2025}
}