Hierarchical Particle Filter
Also known as: nested particle filter, multilevel particle filter, hierarchical SMC, HPF
A hierarchical particle filter extends Sequential Monte Carlo to state-space models with multiple levels of latent variables. Particles are propagated at each level of the hierarchy, allowing the method to track both fine-grained state dynamics and slower-varying hyperparameters simultaneously, yielding calibrated posterior distributions across all levels of the model.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a hierarchical particle filter when the data-generating process has multiple layers of latent variables — for example, individual-level time series governed by group-level parameters — and you need sequential, online inference rather than a batch analysis. It excels when the number of observations grows over time and hyperparameters must be tracked alongside the state. Prefer it over a flat particle filter when the state space is structured into levels, or over offline MCMC when data arrive sequentially and latency matters. Avoid it when the hierarchy has more than two or three levels (computational cost scales badly), when data are not sequential, or when a Kalman filter or linear-Gaussian approximation is sufficient — both are cheaper in those settings.
Strengths & limitations
- Handles multi-level latent structure that a standard particle filter cannot represent without collapsing levels.
- Supports fully online inference: estimates update as each new observation arrives without reprocessing past data.
- Provides a complete posterior distribution over both local states and global hyperparameters.
- Adapts naturally to non-linear, non-Gaussian state-space models where Kalman-based methods fail.
- Marginal likelihood estimates obtained as a byproduct enable sequential model comparison.
- Computational cost scales as O(N_outer x N_inner) particles per time step, making deep hierarchies expensive.
- Degeneracy of the outer (hyperparameter) particle cloud can occur unless rejuvenation MCMC steps are included and well-tuned.
- Designing effective inner proposal distributions requires model-specific knowledge; poor proposals inflate variance.
- Implementation complexity is substantially higher than a standard particle filter or Kalman filter.
- Theoretical convergence guarantees are less developed than for flat SMC, especially for the outer level.
Frequently asked
How does a hierarchical particle filter differ from a standard particle filter?
A standard particle filter maintains one cloud of particles for the latent state and treats any model parameters as fixed. A hierarchical particle filter maintains a second, outer cloud for the global parameters and for each outer particle runs an inner cloud for the local state. This nesting enables joint online inference over both levels.
What is the SMC2 algorithm and is it the same as a hierarchical particle filter?
SMC2 (Chopin et al., 2013) is a specific, theoretically grounded instance of the hierarchical particle filter idea: an outer SMC over model parameters targets the marginal posterior, while inner SMC sweeps provide unbiased likelihood estimates. It is the most widely cited formal framework, but the broader family of hierarchical or nested particle filters covers similar constructions under different names.
How many particles do I need?
This depends on both the complexity of the state space and the number of outer (hyperparameter) particles. A common practice is to choose enough inner particles that the effective sample size stays above 50% of the nominal size, and enough outer particles that hyperparameter estimates stabilise across independent runs. There is no universal rule; pilot runs with different particle counts are recommended.
When should I use MCMC instead?
If data arrive as a complete batch and computation time is not a concern, offline MCMC methods such as Hamiltonian Monte Carlo with NUTS are typically easier to implement, better understood theoretically, and produce more accurate posteriors. The hierarchical particle filter is advantageous primarily when data arrive sequentially and real-time updates are needed.
Can I use the hierarchical particle filter for model selection?
Yes. Each outer particle produces an estimate of the marginal likelihood (evidence) as a byproduct of the inner SMC sweep. These estimates can be used for sequential Bayes factors or to compare competing hierarchical models as data accumulate, without rerunning the full analysis.
Sources
- Briers, M., Doucet, A. & Maskell, S. (2010). Smoothing algorithms for state-space models. Annals of the Institute of Statistical Mathematics, 62(1), 61-89. DOI: 10.1007/s10463-009-0236-2 ↗
- Chopin, N., Jacob, P. E. & Papaspiliopoulos, O. (2013). SMC2: an efficient algorithm for sequential analysis of state-space models. Journal of the Royal Statistical Society: Series B, 75(3), 397-426. DOI: 10.1111/j.1467-9868.2012.01046.x ↗
How to cite this page
ScholarGate. (2026, June 3). Hierarchical Particle Filter. ScholarGate. https://scholargate.app/en/bayesian/hierarchical-particle-filter
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Hierarchical Bayesian InferenceBayesian↔ compare
- Hierarchical Markov Chain Monte CarloBayesian↔ compare
- Kalman FilterBayesian↔ compare
- Particle FilterBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare