Particle Kalman Filtering: A Nonlinear Bayesian Framework for Ensemble Kalman Filters

This paper investigates an approximation scheme of the optimal nonlinear Bayesian filter based on the Gaussian mixture representation of the state probability distribution function. The resulting filter is similar to the particle filter, but is different from it in that, the standard weight-type correction in the particle filter is complemented by the Kalman-type correction with the associated covariance matrices in the Gaussian mixture. We show that this filter is an algorithm in between the Kalman filter and the particle filter, and therefore is referred to as the particle Kalman filter (PKF). In the PKF, the solution of a nonlinear filtering problem is expressed as the weighted average of an “ensemble of Kalman filters” operating in parallel. Running an ensemble of Kalman filters is, however, computationally prohibitive for realistic atmospheric and oceanic data assimilation problems. For this reason, we consider the construction of the PKF through an “ensemble” of ensemble Kalman filters (EnKFs) instead, and call the implementation the particle EnKF (PEnKF). We show that different types of the EnKFs can be considered as special cases of the PEnKF. Similar to the situation in the particle filter, we also introduce a re-sampling step to the PEnKF in order to reduce the risk of weights collapse and improve the performance of the filter. Numerical experiments with the strongly nonlinear Lorenz-96 model are presented and discussed.

💡 Research Summary

The paper introduces a novel Bayesian filtering framework designed for high‑dimensional, nonlinear, and non‑Gaussian data‑assimilation problems typical of atmospheric and oceanic applications. Starting from the optimal nonlinear filter (ONF), the authors note that directly propagating the full conditional probability density function (pdf) is computationally infeasible in realistic settings. To approximate the pdf, they adopt a Gaussian mixture (Mixture of Normals, MON) representation, leading to the Particle Kalman Filter (PKF). In the PKF the state pdf is expressed as a weighted sum of N Gaussian components, each characterized by a mean, covariance, and weight. When N = 1 the PKF collapses to a standard Kalman filter; when the covariances shrink to zero the mixture becomes a set of Dirac masses, i.e., a particle filter. Thus the PKF occupies a continuum between the Kalman filter and the particle filter, inheriting the Kalman‑type mean‑covariance update and the particle‑type weight update.

The prediction step propagates each component’s mean and covariance through the nonlinear model (using either an extended Kalman filter or an ensemble Kalman filter). Upon receipt of observations, Bayes’ rule updates the component weights proportionally to the likelihood evaluated with an innovation covariance Σᵢ that includes both model‑projected error and observation error. Because Σᵢ is generally larger than the observation error covariance R, components far from the observation receive relatively higher weights than in a pure particle filter, mitigating weight collapse and reducing Monte‑Carlo variance.

Running a full Kalman filter for every component would be prohibitively expensive, especially in high dimensions. To address this, the authors replace each Kalman filter with an Ensemble Kalman Filter (EnKF), yielding the Particle Ensemble Kalman Filter (PEnKF). In the PEnKF, an ensemble of EnKFs operates in parallel; each EnKF maintains its own ensemble members, performs forecast and analysis steps, and contributes a Gaussian component to the overall mixture. The final state estimate is the weighted average of all EnKF analyses. This construction preserves the computational scalability of EnKFs while adding the mixture‑based flexibility of the PKF.

Weight collapse is further controlled by a resampling procedure. An information‑theoretic criterion (e.g., entropy) monitors the effective sample size of the mixture. When the criterion signals degradation, the mixture is re‑approximated by a new set of Gaussian components that preserve the overall mean and covariance but reset the weights to be more uniform. This step reduces degeneracy without sacrificing the statistical moments of the distribution.

The methodology is evaluated on the chaotic Lorenz‑96 model (40 variables) under various observation densities, intervals, and noise levels. Comparative experiments involve a standard EnKF, a particle filter, and the proposed PEnKF (with and without resampling). Results show that PEnKF consistently yields lower root‑mean‑square errors, especially when observations are sparse or the system exhibits strong nonlinearity. The particle filter suffers from rapid weight collapse unless an impractically large number of particles is used, while the EnKF’s performance degrades as the nonlinearity increases. The resampling‑enhanced PEnKF remains stable even with modest ensemble sizes, confirming its robustness.

In conclusion, the Particle Kalman Filter and its ensemble implementation provide a practical middle ground between Kalman‑type and particle‑type filters. They combine the efficient covariance handling of EnKFs with the ability of mixture models to capture multimodal, non‑Gaussian features, and they incorporate resampling to avoid weight degeneracy. The authors suggest that the PEnKF can be extended to operational weather and ocean forecasting, possibly integrating advanced EnKF variants (e.g., ETKF, LETKF) and exploiting parallel computing architectures for further scalability.