Dynamic Particle Filter
Also known as: dynamic sequential Monte Carlo, dynamic SMC, bootstrap particle filter, dynamic SIR filter
A dynamic particle filter is a sequential Monte Carlo algorithm that tracks an evolving hidden state over time by maintaining a population of weighted random samples — particles — each representing a plausible trajectory. As new observations arrive, particle weights are updated via the likelihood and the population is resampled, keeping the representation concentrated on the most probable state regions in a fully nonlinear and non-Gaussian setting.
Key highlights
- Handles nonlinear dynamics and non-Gaussian observation and process noise without approximation.
- Produces a full posterior distribution over the hidden state at each time step, not just a point estimate.
- Naturally extends to online / real-time inference: particles are updated sequentially as data arrive.
- Flexible: any transition and observation model that can be simulated and evaluated point-wise is supported.
- Relatively straightforward to implement; the bootstrap particle filter requires only sampling and likelihood evaluation.
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 a dynamic particle filter when the state-space model is nonlinear or the noise distributions are non-Gaussian — settings where the Kalman filter is theoretically wrong. It is the method of choice for object tracking, robotics localization, financial stochastic volatility estimation, and epidemiological filtering of disease incidence. Do not use it when the model is linear-Gaussian (the Kalman filter is optimal and far cheaper), when the state dimension is very high (particle filters suffer from the curse of dimensionality above roughly 10–20 dimensions without special structure), or when computational resources are severely limited.
Strengths & limitations
- Handles nonlinear dynamics and non-Gaussian observation and process noise without approximation.
- Produces a full posterior distribution over the hidden state at each time step, not just a point estimate.
- Naturally extends to online / real-time inference: particles are updated sequentially as data arrive.
- Flexible: any transition and observation model that can be simulated and evaluated point-wise is supported.
- Relatively straightforward to implement; the bootstrap particle filter requires only sampling and likelihood evaluation.
- Performance degrades rapidly as state dimension grows (curse of dimensionality); impractical beyond ~10–20 dimensions without structural assumptions.
- Requires choosing the number of particles N: too few leads to poor approximation, too many is computationally expensive.
- Resampling introduces Monte Carlo variance and can cause particle impoverishment in nearly deterministic systems.
- Smoothed (offline) estimates of past states require additional algorithms such as particle smoothing, adding complexity.
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 many particles do I need?
There is no universal rule; it depends on the state dimension, the nonlinearity of the model, and the signal-to-noise ratio. A common starting point is N = 1000–5000 for low-dimensional problems, with validation by checking that posterior summaries stabilize as N increases. High-dimensional or multimodal posteriors may require N in the tens of thousands.
When should I use a particle filter instead of a Kalman filter?
Use a Kalman filter when the state-space model is truly linear and the noise is Gaussian — it is optimal and efficient in that setting. Switch to a particle filter when nonlinearity or non-Gaussianity is substantial, or when the posterior is multimodal (e.g. multiple plausible positions during ambiguous tracking).
What is particle degeneracy and how do I detect it?
Degeneracy occurs when almost all weight concentrates on one or a few particles, making the approximation effectively useless. Monitor the effective sample size ESS = 1 / sum(w_i^2): if ESS drops far below N (e.g. below N/2), resample. Sequential resampling triggered by an ESS threshold is standard practice.
Can I estimate model parameters alongside the state?
Not straightforwardly with a basic particle filter, because static parameters do not evolve and quickly degenerate. The preferred approaches are particle MCMC (Andrieu et al., 2010), which uses particle filters inside an MCMC loop, or SMC-squared methods that run nested SMC algorithms.
What does 'dynamic' mean in the name?
It signals that the particle filter is applied to a time-evolving (dynamic) state-space model where the hidden state changes at every time step, as opposed to a static inference problem. The dynamic framing is the standard use case for particle filters.
Sources
- 1.Doucet, A., de Freitas, N. & Gordon, N. (Eds.). (2001). Sequential Monte Carlo Methods in Practice. Springer.ISBN 978-0387951461
- 2.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.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Dynamic Particle Filter. ScholarGate. https://scholargate.app/bayesian/dynamic-particle-filter