RBPF: Rao-Blackwellised Particle Filtering

• RBPF is a sequential Bayesian estimation technique that decomposes the latent state into a nonlinear, sampled component and an analytically marginalized linear-Gaussian component.
• It reduces variance by sampling only the complex parts of the state while integrating tractable elements using filters like the Kalman or HMM filter.
• RBPF is applied in GNSS positioning, SLAM, object tracking, and hybrid systems, offering efficient performance in high-dimensional and non-Gaussian scenarios.

Rao-Blackwellised Particle Filtering (RBPF) is a sequential Bayesian estimation methodology that exploits conditional structure in state-space models to jointly combine sampling-based (Monte Carlo) methods and analytic marginalization, thereby reducing computational variance and improving statistical efficiency compared to standard particle filters. RBPF is particularly suitable when the latent state can be decomposed into “difficult” (typically nonlinear/non-Gaussian) components, which are sampled via particles, and “easy” (conditioned linear-Gaussian or finite-state Markov) components, which are integrated out analytically using optimal filters such as the Kalman filter or hidden Markov model (HMM) filter. In this framework, each particle carries not only a weighted sample, but also an analytic conditional posterior over the marginalized subspace, yielding lower-variance estimators for posterior expectations and facilitating computational tractability in high-dimensional or hybrid systems.

1. Structural Principle and State Factorization

At the core of RBPF is the Rao–Blackwell theorem, which asserts that conditioning on sufficient statistics and marginalizing analytically reduces estimator variance. In sequential inference for state-space models, this is operationalized by partitioning the latent state at each time $t$ into two components:

• $x_t^n$: a low-dimensional nonlinear (often non-Gaussian) vector, tracked by Monte Carlo particles;
• $x_t^l$: a linear-Gaussian (or otherwise tractable) vector, for which an optimal recursive filter is available.

The joint filtering distribution is then factorized as

$$p(x_{0:t}^n, x_t^l | y_{0:t}) = p(x_{0:t}^n | y_{0:t}) ~ p(x_t^l | x_{0:t}^n, y_{0:t})$$

as detailed in robust position estimation (Niimi et al., 4 Jun 2025), with $x_t^n$ as position and $x_t^l$ as velocity, and in RBPFs for jump Markov nonlinear systems (Özkan et al., 2013), Markov-modulated nonlinear dynamic systems (Saha et al., 2013), and generic dynamic Bayesian networks (Doucet et al., 2013).

Conditionally, given a particle path $x_{0:t}^n$, all "easy" latent variables $x_t^l$ can be integrated exactly using e.g., a Kalman filter, HMM filter, or analytic update for model parameters. This results in a filtering approximation where only the “tough” nonlinear components are sampled, while the high-dimensional tractable components are handled analytically, leading to reduced sample degeneracy and lower Monte Carlo variance.

2. RBPF Algorithms: Recursion and Weighting

The typical RBPF iteration consists of the following steps (see (Niimi et al., 4 Jun 2025, Doucet et al., 2013, Özkan et al., 2013)):

1. Prediction:
   • For each particle $i$, propagate $x_{t-1,i}^n$ according to the model; for dependent components (e.g., position), use the most current analytic estimate (e.g., velocity from the Kalman filter).
   • For each particle, perform an analytic time update for $x_t^n$0.
2. Weight Update:
   • Compute the likelihood $x_t^n$1, often integrating analytically over $x_t^n$2, leading to

   $x_t^n$3

• When observation models are nonlinear, as in GNSS phase-based positioning (Niimi et al., 4 Jun 2025), the likelihood may involve specialized ambiguity functions or pseudorange-residual-based metrics.

1. Resampling:

   • Normalize weights and resample if the effective sample size (ESS) falls below a threshold.
2. Analytic Correction:
   • For each particle, update the optimal filter for $x_t^n$4 using the latest observation and the current particle’s $x_t^n$5.

This structure generalizes to hybrid discrete/continuous models. In Markov-switching systems (Özkan et al., 2013, Saha et al., 2013), a bank of HMMs marginalizes the discrete mode, while each particle carries an associated continuous state trajectory, leading to reduced degeneracy in the discrete state.

3. Model Classes and Applications

RBPF is broadly applicable to hybrid and partially tractable state-space models:

• Mixed linear/nonlinear systems: e.g., position-velocity models in GNSS positioning, tracking applications, or signal processing, where velocity is linear-Gaussian (filtered with KF) and position is nonlinear (Niimi et al., 4 Jun 2025, Yurdakul et al., 2024).
• Jump Markov and Markov-modulated systems: Models with discrete switching regimes (modes), in which regime trajectories are marginalized via an HMM filter and the continuous state is sampled (Özkan et al., 2013, Saha et al., 2013, Nguyen et al., 2017).
• SLAM and POMDPs: State-of-the-art localization and mapping, where a robot’s trajectory is sampled while map landmarks or features are marginalized (Doucet et al., 2013, Lee et al., 2024).
• Parameter-state models: Joint estimation for state and unknown parameters, either static (marginalized via analytic updates or grid methods (Fang et al., 2023)) or stochastic (hierarchically marginalized with RBPF—see noise-adaptive LDS (Saha, 2015)).
• Pose tracking and multi-modal estimation: PoseRBPF marginalizes rotation (discretized SO(3)), tracks translation by particles for real-time 6D detection (Deng et al., 2019).
• Filtering under constraints: Multi-body manipulation tasks with contact constraints incorporated via QP into the analytic update per particle (Li et al., 2023).

A summary table of representative applications, showing model structure and marginalized subspace, appears below:

Application Domain	Sampled State	Marginalized State
GNSS robust positioning (Niimi et al., 4 Jun 2025)	3D position	3D velocity (KF)
Connected vehicle networks (Shen et al., 2017)	GNSS biases	Vehicle EKFs
Jump Markov/Hybrid SSM (Özkan et al., 2013)	Nonlinear continuous state	Discrete mode (HMM)
SLAM, POMDP planning (Lee et al., 2024)	Robot trajectory, intent	Map landmarks (KF), continuous substate
Extended object tracking (Yurdakul et al., 2024)	Nonlinear pose/shape	Shape/extent (KF)
Parameter-state SDE (Fang et al., 2023)	State trajectory	Model parameters
Object pose estimation (Deng et al., 2019)	3D translation	3D rotation (discrete)

4. Empirical Performance and Theoretical Properties

RBPF provides quantifiable performance gains due to variance reduction in the marginalization of tractable subspaces (Niimi et al., 4 Jun 2025, Doucet et al., 2013, Lee et al., 2024, Özkan et al., 2013). Notable empirical and theoretical properties include:

• Sample efficiency: Marginalizing linear-Gaussian or discrete subspaces allows reduced particle counts for equivalent estimation error (e.g., for GNSS positioning, velocity error ≤ 0.1 m/s improves from 79.6% to 83.3%; position error ≤ 0.3 m improves from 54.4% for PF to 68.5% for RBPF (Niimi et al., 4 Jun 2025)).
• Variance reduction: Rao–Blackwellization assures, via central limit theorems (CLT), that the asymptotic variance of RBPF estimates is strictly smaller than that of joint-sampling PFs (Doucet et al., 2013, Fang et al., 2023).
• High-dimensional tractability: Integrating over analytically tractable states avoids the exponential particle-count scaling with total state dimension required by vanilla particle filters.
• Robustness in the presence of non-Gaussian noise or state constraints: Hierarchical and constrained RBPFs handle stochastic volatility, heavy-tailed errors (Saha, 2015), and inequality-constrained estimation (Li et al., 2023).

In POMDPs, RBPF–quadrature hybrids further reduce the number of tree searches required for high-quality planning, outperforming SIR-PF-based planners under fixed computational budgets (Lee et al., 2024).

5. Methodological Extensions and Pseudocode

RBPF has been widely extended and adapted:

• Generalization to arbitrary tractable subspaces: Any subblock for which an optimal, finite-sufficient-statistics filter exists can be marginalized (Doucet et al., 2013, Nguyen et al., 2017).
• Look-ahead and decentralized variants: When exact analytic marginalization is impossible, decentralized filters approximate the Rao–Blackwell step via nested (local) Monte Carlo, with parallelizability and robust adaptive state-space decomposition (Ahmed et al., 2012).
• Rao–Blackwellized POMDP solvers: Combine particle-based belief tracking in the hard subspace with analytic filters/UKFs and sparse-grid quadrature for tractable parts, boosting belief accuracy and planning efficiency (Lee et al., 2024).
• Robustification and constraint handling: RBPFs have been extended to enforce contact constraints via quadratic programming in physical estimation (Li et al., 2023).

A representative high-level pseudocode, tailored to the generic RBPF, is:

$x_t^n$6

For application-specific pseudocode, see vehicular GNSS (Algorithm block, (Niimi et al., 4 Jun 2025)), connected vehicles (Shen et al., 2017), superelliptical tracking (Yurdakul et al., 2024), and object pose tracking (Deng et al., 2019).

6. Practical Guidelines, Limitations, and Future Directions

Practical deployment of RBPF requires careful state partitioning and tuning:

• Partitioning: Select the largest-possible subspace for analytic marginalization without violating model tractability or introducing significant nonlinearities (e.g., velocity in GNSS or parameters in SDE models (Fang et al., 2023)).
• Noise and model parameter tuning: Process noise and measurement noise covariance choices are critical for stable analytic filters and reliable particle diversity (Niimi et al., 4 Jun 2025).
• Constraint handling: For systems with physical or combinatorial constraints (contacts, reachability), explicit constraint enforcement per particle via optimization (e.g., QP projection) is necessary (Li et al., 2023).
• Resampling criteria: Use effective sample size (ESS)-based thresholds to prevent weight degeneracy and particle impoverishment (Doucet et al., 2013, Özkan et al., 2013).

Current limitations include:

• Requirement of analytic (finite-dimensional) filters for marginalized subspaces (not all models admit this, though extensions to nested Monte Carlo exist (Ahmed et al., 2012)).
• In models with strong nonlinearities or abrupt switching, state partitioning and filter stabilization can become challenging.
• Computational overhead per particle increases due to analytic updates but is offset by reduced sampling variance and particle count.

Future extensions highlighted in recent literature involve:

• Tightly coupled measurement handling (e.g., direct Doppler in GNSS (Niimi et al., 4 Jun 2025));
• Modularized parameter-state RBPF with grid-based updating (Fang et al., 2023);
• Generalization to arbitrary DBN and graphical model structures (Doucet et al., 2013);
• Hybrid MCMC–SMC schemes and particle rejuvenation for reduced degeneracy in smoothing (Nguyen et al., 2017, Olsson et al., 2010);
• Efficient quadrature integration and importance distribution design in high-dimensional planning (Lee et al., 2024).

RBPF thus provides a unifying, variance-reducing inference principle across numerous applications in navigation, tracking, hybrid systems, and time-series analysis, with demonstrated advantages in efficiency, robustness, and estimator precision.