Generative Adversarial Learning from Deterministic Processes
Generative Adversarial Learning from Deterministic Processes
Authors: Joris C. Kühl, Hanno Gottschalk Date: 2026-05-18 Paper ID: openalex:2605.18425
Summary
This paper addresses the theoretical gap between the i.i.d.-based statistical learning theory of GANs and their empirical success in learning from deterministic, chaotic physical systems. The authors demonstrate that generative adversarial learning can recover the invariant distribution of a chaotic process using only a single deterministic trajectory. They further provide explicit convergence guarantees for this learning process, bridging the gap between dynamical systems theory and generative modeling.
Key Contributions
- Establishes a theoretical foundation for generative adversarial learning on non-i.i.d. data derived from chaotic dynamical systems.
- Proves that an infinite-dimensional model can learn the invariant distribution of a chaotic system from a single deterministic time series.
- Provides explicit convergence rates for the generative process measured in Jensen-Shannon divergence.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.18425
- paper_id
- 2605.18425
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-21T06:27:27Z