End-to-End Identifiable and Consistent Recurrent Switching Dynamical Systems
End-to-End Identifiable and Consistent Recurrent Switching Dynamical Systems
Authors: Carles Balsells-Rodas, Zhengrui Xiang, Xavier Sumba, Yingzhen Li Date: 2026-05-07 Paper ID: openalex:2605.06315
Summary
This paper addresses the challenge of learning identifiable representations in sequential models characterized by regime-switching dynamics. The authors move beyond the common reliance on variational approximations by providing theoretical identifiability guarantees for recurrent nonlinear switching dynamical systems. They propose ΩSDS, a flow-based estimator that allows for exact likelihood optimization through expectation-maximisation, resulting in better latent structure recovery and more precise dynamical forecasting compared to standard VAE-based methods.
Key Contributions
- Establishes theoretical identifiability for a broad class of recurrent nonlinear switching dynamical systems, extending beyond stationary or restricted emission assumptions.
- Introduces ΩSDS, a flow-based estimator that facilitates exact likelihood optimization for switching systems via expectation-maximisation.
- Demonstrates superior latent disentanglement and predictive accuracy compared to traditional VAE-based approaches on both synthetic and empirical datasets.
Key Concepts
- omega-sds: A flow-based estimator for recurrent switching dynamical systems that facilitates exact likelihood estimation via expectation-maximisation.
Archivist Review
I approved the ΩSDS concept as it represents a novel methodological approach to overcoming the approximation limitations of variational inference in switching dynamical systems. I have maintained a strict stance on datasets and open questions, finding no candidates in this specific entry that meet the high threshold for long-term vault inclusion. The review prioritized foundational algorithmic contributions over application-specific metrics.
Approved Concepts
- ΩSDS: It enables exact likelihood optimization for recurrent switching dynamical systems, overcoming VAE approximation gaps.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.06315
- paper_id
- 2605.06315
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- omega-sds
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-10T06:05:26Z