The Unscented Particle Filter

Rudolph van der MerweArnaud DoucetNando de FreitasEric Wan

article2000NeurIPS1,775 citations

Introduces the unscented particle filter, an advanced sequential Monte Carlo method that uses the unscented Kalman filter to generate superior, heavy-tailed proposal distributions for highly accurate state estimation in nonlinear, non-Gaussian systems.

Listen

Real-time state estimation is essential across engineering and finance, where systems are frequently characterized by complex nonlinear behaviors, sudden shifts, and non-Gaussian noise. Standard estimation tools like the Extended Kalman Filter (EKF) linearize models using approximations that can cause the filter to diverge. Conversely, standard particle filters can handle non-Gaussian distributions but often fail when incoming observations are highly accurate or fall in the tails of prior distributions. This causes sample depletion, where only a handful of sampled particles remain useful for estimation.

The article introduces and evaluates the Unscented Particle Filter (UPF), a novel filtering method that integrates the Unscented Kalman Filter (UKF) to generate proposal distributions within a sequential Monte Carlo framework. The authors demonstrate that this combination significantly enhances estimation accuracy and algorithmic robustness by moving sampling particles toward regions of high likelihood while maintaining heavier-tailed proposal distributions.

To establish credibility and prove practical value, the authors provide both a theoretical convergence proof and a controlled simulation experiment. The simulation evaluated state tracking across a non-stationary observation model over 60 time steps, comparing the UPF against the standard EKF, UKF, generic particle filters, and EKF-based particle filters across 100 independent Monte Carlo trials using 200 particles per run.

Key findings show that the Unscented Particle Filter outperformed all competing methods by a substantial margin. The UPF achieved a mean-square error of 0.070, which is approximately four to six times lower than standard particle filters (0.424) and the EKF (0.374), and roughly four times lower than EKF-based particle filters (0.307 to 0.310). The variance of the error across trials was also the lowest (0.006), demonstrating superior stability. These empirical gains align with theoretical findings confirming that the UPF's convergence rate is independent of state-space dimensions as long as proposal distribution weights remain upper-bounded. Standalone comparisons also verified that the UKF generates more realistic covariance estimates than the EKF, preventing severe underestimation of uncertainty.

These findings imply that engineering and financial decision-makers can achieve significantly higher tracking accuracy and lower operational risk in environments with abrupt regime changes or high-precision sensors. Organizations should consider adopting the UPF framework for nonlinear estimation tasks where legacy EKF or standard particle filter implementations struggle with divergence. Although the UPF requires propagating individual filter statistics for each particle—incurring additional computational cost—the dramatic reduction in estimation error makes it a compelling choice. Future efforts should focus on validating the algorithm in live operational environments and domain-specific production pipelines.

Oregon Graduate Institute of Science and Technology University of California Berkeley University of Cambridge mcmc sequential importance sampling unscented particle filters Unscented Particle Filter sequential Monte Carlo methods nonlinear filtering

Merwe et al (2000).pdf

Paper: CONDENSATION—Conditional Density Propagation for Visual Tracking, MICHAEL ISARD et al. (1998). Introduces sequential Monte Carlo and CONDENSATION density propagation for tracking non-Gaussian dynamic systems, providing the core particle filtering paradigm that the unscented particle filter enhances.
Paper: An Introduction to the Kalman Filter, Greg Welch et al. (1995). Provides the foundational state-space filtering theory and Extended Kalman Filter mechanisms necessary for understanding the Gaussian approximations that the unscented approach improves upon.

Paper: Incremental Learning for Robust Visual Tracking, David A. Ross et al. (2008). Applies sequential Monte Carlo state-space estimation to robust visual tracking by combining particle filtering with online incremental appearance models.
Paper: Kernel-Based Object Tracking, Dorin Comaniciu et al. (2003). Extends nonlinear visual tracking concepts by combining kernel-based mode-seeking optimization with recursive Kalman-style state prediction.

Abstract

In this paper, we propose a new particle filter based on sequential importance sampling. The algorithm uses a bank of unscented filters to obtain the importance proposal distribution. This proposal has two very “nice” properties. Firstly, it makes efficient use of the latest available information and, secondly, it can have heavy tails. As a result, we find that the algorithm outperforms standard particle filtering and other nonlinear filtering methods very substantially. This experimental finding is in agreement with the theoretical convergence proof for the algorithm. The algorithm also includes resampling and (possibly) Markov chain Monte Carlo (MCMC) steps.

Topics

mcmc sequential importance sampling unscented particle filters

Keywords

Unscented Particle Filter sequential Monte Carlo methods nonlinear filtering unscented Kalman filter extended Kalman filter importance proposal distribution importance sampling non-Gaussian distributions

1 Introduction
2 Dynamic State Space Model
3 Particle Filtering
Generic PF
4 The Unscented Particle Filter
5 Theoretical Convergence
6 Demonstration
7 Conclusions
References

Knowls

Knowl 1 — Unscented Particle Filter
model/method
The Unscented Particle Filter (UPF) is a sequential Monte Carlo algorithm that uses the Unscented Kalman Filter (UKF) to generate the importance proposal distribution for each particle in a sequential importance sampling framework.

For a dynamic state-space model defined by the transition and observation equations: $x_t = f(x_{t-1}, v_{t-1})$ $y_t = h(u_t, x_t, n_t)$ where $x_t \in \mathbb{R}^{n_x}$ is the unobserved state, $y_t \in \mathbb{R}^{n_y}$ is the observation, $u_t \in \mathbb{R}^{n_u}$ is the exogenous input, $v_t \in \mathbb{R}^{n_v}$ is process noise, and $n_t \in \mathbb{R}^{n_n}$ is measurement noise, standard particle filters commonly sample from the transition prior $p(x_t \mid x_{t-1})$ . This transition-prior proposal can fail when the measurement likelihood $p(y_t \mid x_t)$ is highly peaked or lies in the tails of the prior distribution.

The UPF updates the sufficient statistics (mean and covariance) of a separate UKF attached to each particle $i \in \{1, \dots, N\}$ upon receiving the current measurement $y_t$ . The resulting UKF estimates parameterize a local Gaussian (or heavier-tailed) proposal distribution $q(x_t \mid x_{0:t-1}^{(i)}, y_{1:t})$ for each particle. This moves particles toward regions of high likelihood and allows control over the tail behavior of the proposal.
Knowl 2 — Theoretical Convergence of Particle Filters with Bounded Importance Weights
theoretical result
Let $B(\mathbb{R}^n)$ denote the space of bounded, Borel measurable functions on $\mathbb{R}^n$ , with norm $\|f\| \triangleq \sup_{x \in \mathbb{R}^n} |f(x)|$ . In a dynamic state-space model with state trajectory $x_{0:t} \in \mathbb{R}^{n_x \times (t+1)}$ and observation sequence $y_{1:t} = \{y_1, \dots, y_t\}$ , suppose the importance weight ratio: $w_t \propto \frac{p(y_t \mid x_t) \, p(x_t \mid x_{t-1})}{q(x_t \mid x_{0:t-1}, y_{1:t})}$ is upper bounded for any $(x_{t-1}, y_t)$ . Then for all time steps $t \ge 0$ , there exists a constant $c_t$ independent of the number of particles $N$ such that for any test function $f_t \in B(\mathbb{R}^{n_x \times (t+1)})$ :
$ight) - \int f_t(x_{0:t}) \, p(dx_{0:t} \mid y_{1:t}) \right)^2 \right] \le c_t \frac{\|f_t\|^2}{N}$$ where the expectation is taken with respect to the random realizations generated by the particle filtering algorithm. This convergence rate $\mathcal{O}(1/N)$ in mean squared error is completely independent of the dimension $n_x$ of the state space.$
Knowl 3 — Bounded Importance Weights Condition for Particle Filter Proposals
assumption
For a particle filter to attain the dimension-independent $\mathcal{O}(1/N)$ theoretical convergence rate, the importance weight ratio: $w_t(x_{0:t}) \propto \frac{p(y_t \mid x_t) \, p(x_t \mid x_{t-1})}{q(x_t \mid x_{0:t-1}, y_{1:t})}$ must remain upper bounded over the support of $(x_{t-1}, y_t)$ .

This condition requires the importance proposal distribution $q(x_t \mid x_{0:t-1}, y_{1:t})$ to possess heavier tails than the product $p(y_t \mid x_t) p(x_t \mid x_{t-1})$ . If the proposal distribution has thinner tails than the target distribution, the importance weights can become unbounded, causing weight degeneracy and divergence of the particle filter.
Knowl 4 — Particle Filtering with Sequential Importance Sampling, Selection, and MCMC Steps
algorithm
A sequential Monte Carlo algorithm approximates the joint posterior distribution $p(x_{0:t} \mid y_{1:t})$ of state trajectories given observations $y_{1:t}$ using $N$ weighted particles.
Input: Initial particles { $x_0^{(i)}$ } for $i = 1, \dots, N$ drawn from prior $p(x_0)$ , observation sequence $y_{1:T}$
Output: Weighted particle sets approximating $p(x_{0:t} \mid y_{1:t})$ for each $t = 1, \dots, T$
for $t = 1$ to $T$ do
    // Step 1: Sequential importance sampling
    for $i = 1$ to $N$ do
        Sample new state: $\tilde{x}_t^{(i)} \sim q(x_t \mid x_{0:t-1}^{(i)}, y_{1:t})$
        Update trajectory: $\tilde{x}_{0:t}^{(i)} \leftarrow (\tilde{x}_t^{(i)}, x_{0:t-1}^{(i)})$
        Compute unnormalized weight: $w_t^{(i)} \leftarrow \frac{p(y_t \mid \tilde{x}_t^{(i)}) p(\tilde{x}_t^{(i)} \mid x_{t-1}^{(i)})}{q(\tilde{x}_t^{(i)} \mid x_{0:t-1}^{(i)}, y_{1:t})}$
    end for
    for $i = 1$ to $N$ do
        Normalize weight: $\tilde{w}_t^{(i)} \leftarrow \frac{w_t^{(i)}}{\sum_{j=1}^N w_t^{(j)}}$
    end for
    // Step 2: Selection / Resampling
    Multiply and suppress samples $(\tilde{x}_{0:t}^{(i)})$ according to normalized weights $(\tilde{w}_t^{(i)})$ to obtain $N$ equally weighted samples $(x_{0:t}^{(i)})$
    // Step 3: MCMC step (optional)
    for $i = 1$ to $N$ do
        Apply a Markov transition kernel with invariant distribution $p(x_{0:t} \mid y_{1:t})$ to $(x_{0:t}^{(i)})$
    end for
end for
The algorithm updates sample trajectories recursively, eliminates low-weight particles through selection, and optionally applies a Markov Chain Monte Carlo (MCMC) transition kernel (e.g., Metropolis-Hastings or Gibbs) to inject sample diversity without altering the target distribution.
Knowl 5 — Covariance Estimation Properties of EKF vs. UKF Proposals
model/method
When generating proposal distributions for particle filters, the Extended Kalman Filter (EKF) approximates nonlinear state and measurement models via first-order Taylor series expansions. This linearization systematically underestimates the true posterior state covariance.

In contrast, the Unscented Kalman Filter (UKF) uses deterministic sigma-point sampling to capture the posterior mean and covariance accurately to second or higher orders for arbitrary nonlinearities. Because the UKF produces larger and more realistic state covariance estimates than the EKF, proposal distributions constructed with the UKF have heavier tails. This tail behavior prevents the proposal distribution from under-covering the true posterior and satisfies the bounded importance weight condition required for particle filter convergence.
Knowl 6 — Nonlinear Non-Stationary Time-Series Benchmark
experimental setup
The state estimation benchmark evaluates filters on a one-dimensional non-stationary, nonlinear state-space system defined by: $x_{t+1} = 1 + \sin(\omega \pi t) + \phi x_t + v_t$ $y_t = \begin{cases} \phi x_t^2 + n_t, & t \le 30 \\ \phi x_t - 2 + n_t, & t > 30 \end{cases}$ where:
- $x_t \in \mathbb{R}$ is the unobserved state over time steps $t = 1, \dots, 60$ .
- $y_t \in \mathbb{R}$ is the scalar measurement.
- The scalar parameters are set to $\omega = 4 \times 10^{-2}$ and $\phi = 0.5$ .
- The process noise $v_t \sim \text{Gamma}(3, 2)$ is non-Gaussian.
- The observation noise $n_t$ is drawn from a zero-mean Gaussian distribution.
All evaluated particle filter variants use $N = 200$ particles. Performance is evaluated by computing the mean and variance of the state mean-squared error (MSE) over 100 independent Monte Carlo runs with random re-initializations.

Knowl 7 — Comparative State Estimation Performance on Nonlinear Benchmark

data/table

The state estimation mean-squared error (MSE) mean and variance over 100 independent Monte Carlo runs on the non-stationary nonlinear benchmark model are reported as follows:

Algorithm	MSE Mean	MSE Variance
Extended Kalman Filter (EKF)	0.374	0.015
Unscented Kalman Filter (UKF)	0.280	0.012
Particle Filter: generic	0.424	0.053
Particle Filter: MCMC move step	0.417	0.055
Particle Filter: EKF proposal	0.310	0.016
Particle Filter: EKF proposal and MCMC move step	0.307	0.015
Particle Filter: UKF proposal (Ünscented Particle Filter")	0.070	0.006
Particle Filter: UKF proposal and MCMC move step	0.074	0.008

The Unscented Particle Filter (PF with UKF proposal) achieves an MSE mean of 0.070, outperforming the generic particle filter (0.424), the EKF (0.374), the stand-alone UKF (0.280), and the EKF-proposal particle filter (0.310). It also exhibits the lowest variance (0.006), indicating consistent estimation accuracy.

Coverage note — None was omitted; all primary contributed algorithms, theoretical convergence results, proposal mechanisms, and experimental findings from the paper are fully covered.

References

1.Anderson, B. D. and Moore, J. B. (1979). Optimal Filtering, Prentice-Hall, New Jersey.
2.Crisan, D. and Doucet, A. (2000). Convergence of generalized particle filters, Technical Report CUED/F-INFENG/TR 381, Cambridge University Engineering Department.
3.de Freitas, J. F. G. (1999). Bayesian Methods for Neural Networks, PhD thesis, Department of Engineering, Cambridge University, Cambridge, UK.
4.de Freitas, J. F. G., Niranjan, M., Gee, A. H. and Doucet, A. (2000). Sequential Monte Carlo methods to train neural network models, Neural Computation 12(4): 955-993.
5.Doucet, A., de Freitas, J. F. G. and Gordon, N. J. (eds) (2001). Sequential Monte Carlo Methods in Practice, Springer-Verlag.
6.Doucet, A., Godsill, S. and Andrieu, C. (2000). On sequential Monte Carlo sampling methods for Bayesian filtering, Statistics and Computing 10(3): 197-208.
7.Gelman, A., Carlin, J. B., Stern, H. S. and Rubin, D. B. (1995). Bayesian Data Analysis, Chapman and Hall.
8.Gordon, N. J., Salmond, D. J. and Smith, A. F. M. (1993). Novel approach to nonlinear/non-Gaussian Bayesian state estimation, IEE Proceedings-F 140(2): 107-113.
9.Julier, S. J. and Uhlmann, J. K. (1997). A new extension of the Kalman filter to nonlinear systems, Proc. of AeroSense: The 11th International Symposium on Aerospace/Defence Sensing, Simulation and Controls, Orlando, Florida., Vol. Multi Sensor Fusion, Tracking and Resource Management II.
10.Pitt, M. K. and Shephard, N. (1999). Filtering via simulation: Auxiliary particle filters, Journal of the American Statistical Association 94(446): 590-599.
11.Thrun, S. (2000). Monte Carlo POMDPs, in S. Solla, T. Leen and K.-R. Müller (eds), Advances in Neural Information Processing Systems 12, MIT Press, pp. 1064-1070.
12.van der Merwe, R., Doucet, A., de Freitas, J. F. G. and Wan, E. (2000). The unscented particle filter, Technical Report CUED/F-INFENG/TR 380, Cambridge University Engineering Department.

Citation

MLA

Merwe, R. van . der ., et al. “The Unscented Particle Filter”. Neural Information Processing Systems, vol. 13, 2000, pp. 584–90, http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.32.9011.

APA

Merwe, R. van . der ., Doucet, A., Freitas, N. de ., & Wan, E. A. (2000). The Unscented Particle Filter. Neural Information Processing Systems, 13, 584–590. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.32.9011

Chicago

Merwe, R. van . der ., A. Doucet, N. de . Freitas, and E. A. Wan. 2000. “The Unscented Particle Filter”. Neural Information Processing Systems 13: 584–90. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.32.9011.

Harvard

Merwe, R. van . der . et al. (2000) “The Unscented Particle Filter”, Neural Information Processing Systems, 13, pp. 584–590. Available at: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.32.9011.

Vancouver

1. Merwe R van der, Doucet A, Freitas N de, Wan EA (2000) The Unscented Particle Filter. Neural Information Processing Systems 13:584–590

BibTeX

@article{merwe2000the,
  title = {The Unscented Particle Filter},
  author = {Merwe, Rudolph van der and Doucet, Arnaud and Freitas, Nando de and Wan, Eric A.},
  year = {2000},
  journal = {Neural Information Processing Systems},
  volume = {13},
  pages = {584-590},
  url = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.32.9011}
}

Metadata:DOI registry

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors

The Unscented Particle Filter

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Unscented Particle Filter

Knowl 2 — Theoretical Convergence of Particle Filters with Bounded Importance Weights

Knowl 3 — Bounded Importance Weights Condition for Particle Filter Proposals

Knowl 4 — Particle Filtering with Sequential Importance Sampling, Selection, and MCMC Steps

Knowl 5 — Covariance Estimation Properties of EKF vs. UKF Proposals

Knowl 6 — Nonlinear Non-Stationary Time-Series Benchmark

Knowl 7 — Comparative State Estimation Performance on Nonlinear Benchmark

References

Citation

BibTeX

Access the Paper