Particle Filter with Measurement Error
Sequential Monte Carlo Particle Filter with Explicit Measurement Error · Also known as: SMC with measurement noise, bootstrap filter with observation error, nonlinear particle filter, particle filter state-space model
A particle filter with explicit measurement error is a Sequential Monte Carlo algorithm that tracks the hidden state of a nonlinear, non-Gaussian dynamic system while formally modelling noise in the observations. A population of weighted random samples (particles) represents the posterior state distribution at each time step, and an observation likelihood function quantifies how much each particle is consistent with the noisy measurement received.
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 particle filter with measurement error when you have sequential data from a nonlinear or non-Gaussian dynamic system and the observations are corrupted by non-trivial noise that must be modelled explicitly — for example, radar tracking with clutter, financial state estimation with fat-tailed observation noise, or ecological population surveys with uncertain counts. It is also the right tool when a Kalman filter is inapplicable because the dynamics or the measurement model are nonlinear. Avoid it when the system is linear-Gaussian (the Kalman filter is then optimal and far cheaper), when N must be huge to cover a high-dimensional state space (the curse of dimensionality renders naive particle filters impractical), or when real-time performance is required on modest hardware without dedicated implementations.
Strengths & limitations
- Handles fully nonlinear state transitions and non-Gaussian observation noise without approximation.
- Explicitly represents the full posterior distribution of the hidden state, not just its mean and variance.
- Naturally accommodates any parametric measurement-error distribution, including heavy-tailed and multimodal likelihoods.
- Straightforward conceptually: propagate, weight by likelihood, resample.
- Naturally incorporates missing observations by skipping the weighting step for missing time points.
- Computational cost grows linearly with N and can be prohibitive for high-dimensional state spaces.
- Weight degeneracy and particle impoverishment can occur without careful resampling strategies.
- The measurement model must be specified correctly; a misspecified likelihood degrades posterior accuracy.
- Variance of Monte Carlo estimates decreases only as 1/sqrt(N), requiring large N for high accuracy.
Frequently asked
How is this different from a standard Kalman filter?
The Kalman filter is optimal only when both the dynamics and the measurement equation are linear and all noise is Gaussian. A particle filter with measurement error makes no such assumptions: it handles arbitrary nonlinearities and any noise distribution, at the cost of higher computational effort.
How many particles do I need?
There is no universal answer: it depends on the dimensionality of the state, the sharpness of the posterior, and the required accuracy. Practical guidance suggests starting with N = 500–2000 for low-dimensional problems and monitoring the effective sample size (ESS = 1 / sum(w_i^2)) — if ESS drops below N/2, increase N or switch to a more efficient proposal.
What happens if my measurement model is wrong?
A misspecified likelihood causes all particles to receive incorrect weights. If the true noise is heavier-tailed than assumed (e.g. Gaussian assumed but actually Student-t), outliers are treated as nearly impossible observations and can cause the filter to lose track of the true state. Always validate the measurement model against held-out data or residual diagnostics.
Can I apply particle filtering to non-time-series problems?
Particle filtering is specifically designed for sequential (time-ordered) inference in dynamic state-space models. For static Bayesian inference without a temporal structure, standard MCMC or variational Bayes is more appropriate.
What is weight degeneracy and how do I fix it?
After many steps, almost all weight concentrates on a single particle while the rest carry negligible weight — this is degeneracy. The standard fix is systematic or stratified resampling triggered whenever ESS falls below a threshold, combined with a sufficiently large N to maintain particle diversity.
Sources
- Gordon, N. J., Salmond, D. J., & Smith, A. F. M. (1993). Novel approach to nonlinear/non-Gaussian Bayesian state estimation. IEE Proceedings F – Radar and Signal Processing, 140(2), 107–113. DOI: 10.1049/ip-f-2.1993.0015 ↗
- Doucet, A., de Freitas, N., & Gordon, N. (Eds.). (2001). Sequential Monte Carlo Methods in Practice. Springer. ISBN: 978-0387951461
How to cite this page
ScholarGate. (2026, June 3). Sequential Monte Carlo Particle Filter with Explicit Measurement Error. ScholarGate. https://scholargate.app/en/bayesian/particle-filter-with-measurement-error
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.
- Extended Kalman FilterControl Theory↔ compare
- Kalman FilterBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare
- Unscented Kalman FilterControl Theory↔ compare