Robust Particle Filter
Also known as: RPF, robust sequential Monte Carlo, outlier-robust particle filter, heavy-tailed particle filter
The Robust Particle Filter is a sequential Monte Carlo method that tracks hidden states in nonlinear, non-Gaussian systems while remaining resistant to outliers and model misspecification. It replaces the standard Gaussian likelihood with a heavy-tailed or bounded-influence density, so that anomalous observations receive downweighted importance and cannot derail the state estimate.
Key highlights
- Handles nonlinear, non-Gaussian dynamics without linearisation.
- Resistant to measurement outliers through heavy-tailed or bounded-influence likelihoods.
- Produces a full posterior distribution at each time step, not just a point estimate.
- Degrades gracefully under model misspecification rather than diverging.
- Applicable to any state-space model where a robust likelihood can be specified.
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 the Robust Particle Filter when tracking or state estimation must survive sporadic sensor failures, measurement outliers, or heavy-tailed noise in a nonlinear system. It is the right choice for target tracking, robot localisation, and financial state estimation where Gaussian assumptions are known to break down. Do not use it when the data and system dynamics are genuinely Gaussian and well-specified — a standard particle filter or Kalman filter is more efficient in that case. Also avoid it when the state space is very high-dimensional (particle filters scale poorly regardless of robustness) or when real-time constraints are so tight that the extra likelihood evaluations cannot be afforded.
Strengths & limitations
- Handles nonlinear, non-Gaussian dynamics without linearisation.
- Resistant to measurement outliers through heavy-tailed or bounded-influence likelihoods.
- Produces a full posterior distribution at each time step, not just a point estimate.
- Degrades gracefully under model misspecification rather than diverging.
- Applicable to any state-space model where a robust likelihood can be specified.
- Computational cost scales linearly with the number of particles, making high-dimensional state spaces expensive.
- Particle degeneracy and impoverishment can still occur if the proposal distribution is poorly chosen.
- The choice of influence function and robustness tuning constant requires domain expertise.
- Performance guarantees are harder to establish than for the standard Kalman filter in linear-Gaussian settings.
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 does the Robust Particle Filter differ from the standard Particle Filter?
The standard particle filter scores particles with a Gaussian likelihood, making it sensitive to outliers that can drive all weights to near zero. The robust variant replaces the Gaussian with a heavy-tailed or bounded-influence likelihood so that extreme observations are downweighted rather than causing filter collapse.
What influence function should I use?
Common choices include a Student-t likelihood (tunable degrees of freedom), a Huber function that transitions from quadratic to linear in the tails, and a Cauchy likelihood. The Student-t with 3-5 degrees of freedom is a practical default that allows heavier tails without extreme sensitivity to the tuning parameter.
How many particles do I need?
There is no universal rule, but tracking the effective sample size (ESS = 1 / sum of squared normalised weights) guides the choice. An ESS below N/2 typically triggers resampling. Start with a few hundred particles and increase until posterior summaries stabilise across independent runs.
When should I use a Robust Kalman Filter instead?
If the underlying system is linear (or nearly so) and only the observation noise is heavy-tailed, a robust Kalman filter variant is computationally far cheaper and analytically tractable. Use the Robust Particle Filter when both the dynamics and observations are nonlinear or strongly non-Gaussian.
Can I combine the Robust Particle Filter with the Kalman Filter?
Yes. Rao-Blackwellised or marginalised particle filters exploit the Kalman filter for the linear sub-structure of the state while applying particle approximations to the nonlinear parts. A robust likelihood can be layered on top, yielding computational savings while retaining outlier resistance.
Sources
- 1.Ristic, B., Arulampalam, S. & Gordon, N. (2004). Beyond the Kalman Filter: Particle Filters for Tracking Applications. Artech House.ISBN 978-1580536318
- 2.Hurzeler, M. & Kunsch, H. R. (1998). Monte Carlo approximations for general state-space models. Journal of Computational and Graphical Statistics, 7(2), 175-193.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust Particle Filter. ScholarGate. https://scholargate.app/bayesian/robust-particle-filter