samsara: a continuous-time Markov chain Monte Carlo sampler for trans-dimensional Bayesian analysis
samsara: a continuous-time Markov chain Monte Carlo sampler for trans-dimensional Bayesian analysis
Authors: Gabriele Astorino, Lorenzo Valbusa Dall’Armi, R. Buscicchio, Joachim Pomper, Angelo Ricciardone, W. Del Pozzo Date: 2026-04-20 Paper ID: openalex:2511.07385
Summary
SAMSARA is a continuous-time MCMC framework designed for Bayesian inference where the parameter dimensionality is unknown. By framing parameter evolution as a sequence of Poisson-driven birth, death, and mutation processes, the method achieves automatic acceptance for trans-dimensional moves, bypassing the traditional bottlenecks of RJMCMC. The approach is validated on multi-signal time series and mixture model problems, demonstrating superior sampling efficiency and robust agreement with nested sampling benchmarks.
Key Contributions
- Introduces SAMSARA, a CTMCMC framework for variable-dimensional Bayesian inference, utilizing Poisson-driven birth-death-mutation processes.
- Achieves automatic acceptance of trans-dimensional moves through adaptive rate definitions, enhancing convergence efficiency compared to standard RJMCMC.
- Demonstrates performance parity with nested sampling on complex tasks including Gaussian mixture models and multi-signal time series decomposition.
Open Questions & Future Work
Key Concepts
- ctmcmc: A trans-dimensional sampling framework that uses Poisson-driven continuous-time dynamics to avoid traditional RJMCMC acceptance bottlenecks.
Archivist Review
I approved the CTMCMC concept as it provides a distinct, reusable alternative to traditional RJMCMC for trans-dimensional Bayesian problems, particularly relevant for time-series modeling where dimensionality may be unknown. I approved the open question on proposal optimization because it addresses a fundamental bottleneck in the scalability of this class of samplers. I rejected the tool name ‘SAMSARA’ in favor of the mechanism ‘CTMCMC’ to avoid capturing paper-local artifacts.
Approved Concepts
- Continuous-time Markov chain Monte Carlo (CTMCMC): Represents a foundational algorithmic shift from RJMCMC for trans-dimensional Bayesian inference.
Approved Open Questions
- Optimizing Mutation Proposals and Dynamics: Improving proposal schemes is critical for the practical scalability of CTMCMC to high-dimensional parameter spaces with unknown dimensionality.
Rejected Candidates
- [concept] SAMSARA (
samsara) - subcomponent_of_broader_mechanism: The paper is the primary introduction of the sampler, but the broader mechanism (CTMCMC) is already approved as a concept, rendering the specific implementation name redundant.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2511.07385
- paper_id
- 2511.07385
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- ctmcmc
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-23T05:48:38Z