What is Rao-Blackwellized Particle Filter?
Rao-Blackwellized Particle Filter is a Sequential Monte Carlo method that samples one subset of latent states while analytically integrating a conditionally tractable subset within every particle.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Factor the posterior before choosing what to sample
Suppose the latent state separates into r_k and z_k, and p(z_0:k | r_0:k, y_1:k) is tractable. An RBPF samples trajectories for r_0:k and stores the conditional distribution of z_0:k or z_k in each particle. For conditionally linear Gaussian systems, that inner calculation is commonly a Kalman Filter.
Doucet, de Freitas, Murphy, and Russell formulate RBPF for Dynamic Bayesian Networks. Rao-Blackwellization is beneficial only when the chosen partition exposes a valid conditional computation; forcing a Gaussian inner filter onto a non-Gaussian conditional state changes the target.
Weight and resample the sampled state without losing its history
Each particle proposes a new intractable state, runs its conditional estimator, and multiplies its weight by the appropriate predictive likelihood and proposal correction. Normalize in log space, monitor Effective Sample Size, and resample only under a declared policy. After resampling, every copied particle must retain the matching conditional mean, covariance, parameters, and ancestry.
Sampling from the transition is simple but can waste particles when observations are informative. Better proposals may use the current observation, but their density must appear in the weight. Marginalizing more variables can lower variance while increasing per-particle algebra and memory.
Measure variance reduction and path degeneration separately
Doucet's SMC resource guide summarizes RBPF as the principle of computing analytically whenever possible and sampling only the intractable part. That principle appears in switching state-space models, tracking, and FastSLAM, but each application has a different factorization and conditional solver.
Compare RBPF with a plain Particle Filter under equal wall time and repeated seeds. Report state RMSE, log-likelihood estimates where available, ESS, resampling count, unique ancestors, mode or map diversity, conditional covariance coverage, runtime, and memory. A large instantaneous ESS does not prove that old trajectories remain diverse.
Key Characteristics
- Samples only a selected intractable subset of latent variables
- Marginalizes a conditionally tractable subset inside every particle
- Often embeds a Kalman Filter for conditional linear Gaussian states
- Reduces Monte Carlo variance when the factorization is valid
- Retains importance-weight and ancestry-degeneration risks
- Trades fewer sampled dimensions for per-particle analytical state
Common Use Cases
- Switching state-space models with discrete modes and continuous states
- FastSLAM-style trajectory sampling with conditional map estimation
- Tracking with nonlinear intent and conditionally linear kinematics
- Inference in structured Dynamic Bayesian Networks
- Mixed nonlinear and linear Gaussian sensor-fusion models
Example
Loading code...Frequently Asked Questions
What does Rao-Blackwellization mean in a Particle Filter?
It means integrating a conditionally tractable part of the latent state analytically instead of sampling it. Each particle samples the remaining state and carries the conditional distribution, sufficient statistics, or estimator for the marginalized variables.
Why can an RBPF need fewer particles than a standard Particle Filter?
Analytical integration removes sampling noise for the marginalized variables and can reduce the sampled dimension. The gain depends on the model partition, proposal, observations, and conditional solver; it is not a universal particle-count ratio.
Is a Rao-Blackwellized Particle Filter exact?
Only the conditional calculation may be exact under its stated model. The outer particle approximation remains finite-sample Monte Carlo, and the overall result can suffer from proposal mismatch, weight degeneration, path collapse, model error, or an approximate inner filter.
How is RBPF different from IMM?
IMM keeps a fixed bank of mode-conditioned filters and moment-matches their mixture each step. RBPF samples mode or nonlinear histories, so different particles can retain different paths, while analytically filtering the conditional state. RBPF is more flexible but can degenerate.
Which diagnostics matter for an RBPF?
Track state error and coverage, Effective Sample Size, normalized-weight concentration, resampling frequency, unique ancestors, path or map diversity, repeated-seed variability, conditional-filter consistency, runtime, and memory. One high ESS value is insufficient.