Dynamic Sequential Monte Carlo
Also known as: Dynamic SMC, SMC for dynamic models, sequential particle filter, dynamic particle sampler
Dynamic Sequential Monte Carlo (Dynamic SMC) is a Bayesian computational method that maintains and updates a population of weighted samples — particles — as new observations arrive over time. It propagates particles through a dynamic system model, reweights them by how well they match the observed data, and periodically resamples to concentrate effort on high-probability regions, yielding online posterior inference for state-space and time-evolving models.
Key highlights
- Handles nonlinear and non-Gaussian state-space models where analytic solutions do not exist.
- Operates online: the posterior is updated incrementally as each new observation arrives, with no need to reprocess past data.
- Provides a consistent estimate of the marginal likelihood, enabling Bayes factors for model comparison.
- Naturally parallelisable — particles are independent and can be propagated simultaneously.
- Flexible proposal design allows tuning for efficiency in specific application domains.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Dynamic SMC when the state-space model is nonlinear or non-Gaussian — conditions that defeat the standard Kalman filter — and data arrive sequentially or in a streaming fashion. It is the natural choice for tracking, navigation, robotics, stochastic volatility models, and epidemiological filtering. Do not use it when the model is linear-Gaussian (use the Kalman filter instead) or when a single large batch of data must be analysed offline without a sequential structure (standard MCMC is more straightforward there). It struggles with very high-dimensional state spaces unless enhanced with MCMC rejuvenation steps.
Strengths & limitations
- Handles nonlinear and non-Gaussian state-space models where analytic solutions do not exist.
- Operates online: the posterior is updated incrementally as each new observation arrives, with no need to reprocess past data.
- Provides a consistent estimate of the marginal likelihood, enabling Bayes factors for model comparison.
- Naturally parallelisable — particles are independent and can be propagated simultaneously.
- Flexible proposal design allows tuning for efficiency in specific application domains.
- Suffers from weight degeneracy in high-dimensional state spaces: the number of particles required grows exponentially with state dimension.
- Resampling introduces Monte Carlo variance and can discard particle diversity (sample impoverishment).
- Approximation quality depends heavily on the choice of proposal distribution; a poor proposal inflates variance.
- Computationally expensive relative to the Kalman filter when the model is actually linear-Gaussian.
- Smoothing (retrospective inference over past states) requires additional algorithms such as backward simulation.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How is Dynamic SMC different from the standard Kalman filter?
The Kalman filter gives an exact analytic posterior only for linear-Gaussian models. Dynamic SMC approximates the posterior numerically with particles and handles any nonlinear or non-Gaussian model; it reduces to the Kalman filter in the linear-Gaussian limit as the particle count grows.
What is weight degeneracy and how does resampling help?
After many sequential updates, one or a few particles accumulate almost all of the weight while the rest become negligible, so the effective sample size collapses. Resampling replaces the weighted set with an equally-weighted set by drawing with replacement proportional to weights, restoring diversity at the cost of introducing resampling variance.
How many particles do I need?
There is no universal rule; the number needed scales with state-space dimension and observation informativeness. A practical approach is to start with N = 1000, monitor ESS over time, and double N until posterior summaries stabilise across repeated runs. Problems with informative likelihoods often need more particles than weakly observed ones.
Can Dynamic SMC estimate static model parameters as well as states?
Yes, but naively including static parameters as part of the particle state causes path degeneracy because parameter diversity is never refreshed. Solutions include particle MCMC (PMCMC), SMC-squared, or adding MCMC rejuvenation steps after each resampling to restore parameter diversity.
Is Dynamic SMC the same as a particle filter?
The terms overlap substantially. A particle filter traditionally refers to the bootstrap or auxiliary particle filter for online state-space filtering. Dynamic SMC is a broader framework (from Del Moral et al. 2006) that subsumes particle filters and extends them to annealed or tempered sequences of distributions, enabling use for static Bayesian inference as well.
Sources
- 1.Del Moral, P., Doucet, A. & Jasra, A. (2006). Sequential Monte Carlo samplers. Journal of the Royal Statistical Society: Series B, 68(3), 411–436.
- 2.Doucet, A., de Freitas, N. & Gordon, N. (Eds.) (2001). Sequential Monte Carlo Methods in Practice. Springer.ISBN 978-0387951461
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Dynamic Sequential Monte Carlo. ScholarGate. https://scholargate.app/bayesian/dynamic-sequential-monte-carlo