Andrej Leban

For Distributional Principal Autoencoders we prove a closed-form data-score–geometry identity and that any latent coordinates beyond the data manifold are uninformative.