Sequential Monte Carlo with Measurement Error
Also known as: SMC with measurement error, particle filter with noisy observations, SMC state-space measurement error, sequential particle filtering with observation noise
Sequential Monte Carlo (SMC) with measurement error is a particle-based Bayesian filtering method for tracking hidden states in dynamical systems when observations are corrupted by noise. It propagates a weighted cloud of particles through time, updating weights at each step to reflect how well each particle explains the noisy measurement, and produces a full posterior distribution over the latent state at every time point.
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 SMC with measurement error when the system is nonlinear or non-Gaussian — conditions under which the standard Kalman filter is exact but particle filters shine. It is appropriate when observations are known to be noisy, when the observation model is non-Gaussian (e.g. heavy-tailed or count data), or when full posterior distributions over the hidden state are needed at each time step. Avoid it when the state-space model is linear and Gaussian (use the Kalman filter instead, which is exact and far cheaper) or when the number of state dimensions is very high (particle filters suffer from the curse of dimensionality in very high dimensions).
Strengths & limitations
- Handles nonlinear transition and observation models without linearisation approximations.
- Works with any measurement-error distribution, including heavy-tailed, multimodal, or discrete noise.
- Delivers full posterior distributions over hidden states, not just point estimates or Gaussian approximations.
- Naturally incorporates informative priors and propagates uncertainty through time.
- The bootstrap filter requires only the ability to sample from the transition density and evaluate the likelihood.
- Particle degeneracy and weight collapse can occur if the proposal distribution does not match the posterior well.
- Computational cost scales with the number of particles N and can be prohibitive for high-frequency or high-dimensional problems.
- Performance degrades rapidly as the latent state dimension increases (curse of dimensionality).
- Choosing the number of particles and an effective importance distribution requires problem-specific tuning.
Frequently asked
How does this differ from a standard particle filter?
A standard particle filter is already designed for state-space models with noisy observations. 'SMC with measurement error' emphasises that the observation model explicitly accounts for measurement noise — often with a parameterised error variance — rather than treating observations as exact. The algorithm is the same; the framing highlights the role of measurement uncertainty in the weight update.
When should I use a Kalman filter instead?
If the state transition and observation models are both linear and the noise is Gaussian, the Kalman filter is the exact Bayesian solution and should be preferred. SMC is needed only when the model is nonlinear or the noise is non-Gaussian; otherwise it is an expensive approximation to an available closed-form solution.
How do I choose the measurement-error variance?
It can be specified from instrument calibration data, estimated jointly with the hidden state using parameter estimation extensions (e.g. Liu–West filter or particle MCMC), or treated as a hyperparameter and selected via cross-validation or marginal likelihood maximisation. Sensitivity analysis with a range of plausible values is always recommended.
How many particles are enough?
There is no universal answer. A common starting point is N = 1000 for low-dimensional problems, then increasing until posterior summaries stabilise across runs. Effective sample size diagnostics and variance of weight estimates guide the choice. High-noise observation models or multimodal posteriors generally require more particles.
Can SMC handle missing observations?
Yes. When an observation is missing, the weight update step is simply skipped for that time step — particles are propagated forward through the transition model but their weights are not changed. This is one of the practical advantages of particle filters over likelihood-based methods that require imputation.
Sources
- Doucet, A., de Freitas, N., & Gordon, N. (Eds.). (2001). Sequential Monte Carlo Methods in Practice. Springer New York. ISBN: 978-0-387-95146-1
- Cappe, O., Godsill, S. J., & Moulines, E. (2007). An overview of existing methods and recent advances in sequential Monte Carlo. Proceedings of the IEEE, 95(5), 899-924. DOI: 10.1109/JPROC.2007.893250 ↗
How to cite this page
ScholarGate. (2026, June 3). Sequential Monte Carlo with Measurement Error. ScholarGate. https://scholargate.app/en/bayesian/sequential-monte-carlo-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.
- Bayesian Inference with Measurement ErrorBayesian↔ compare
- Dynamic Bayesian InferenceBayesian↔ compare
- Kalman Filter with Measurement ErrorBayesian↔ compare
- Markov Chain Monte CarloSimulation↔ compare
- Particle FilterBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare