Abstract
Simultaneous state and parameter estimation arises from various applicational areas but presents a major computational challenge. Most available Markov chain or sequential Monte Carlo techniques are applicable to relatively low dimensional problems only. Alternative methods, such as the ensemble Kalman filter or other ensemble transform filters have, on the other hand, been successfully applied to high dimensional state estimation problems. In this paper, we propose an extension of these techniques to high dimensional state space models which depend on a few unknown parameters. More specifically, we combine the ensemble Kalman-Bucy filter for the continuous-time filtering problem with a generalized ensemble transform particle filter for intermittent parameter updates. We demonstrate the performance of this two stage update filter for a wave equation with unknown wave velocity parameter.
Introduction
In scientific modeling, Stochastic Differential Equations (SDEs) are frequently used to represent complex systems like atmospheric dynamics or fluid flows. A critical task in these fields is data assimilation: the real-time updating of a model's state and its internal parameters (e.g., velocity or friction) based on noisy observations. This task faces two primary hurdles: the high dimensionality of state variables and the often non-Gaussian nature of parameter distributions. While the Ensemble Kalman Filter (EnKF) handles high dimensions well, its Gaussian assumptions can fail when estimating complex parameters. Conversely, standard Particle Filters (PF) suffer from "weight collapse" in high dimensions.
Methodology: A Two-Stage Hybrid Approach
The proposed framework addresses these challenges by "disintegrating" the joint probability distribution of states and parameters, applying specialized filters to each:
- State Estimation (EnKBF): The Ensemble Kalman-Bucy Filter, a continuous-time variant of the EnKF, is used for the high-dimensional state space. It remains computationally efficient provided the state distribution is approximately Gaussian for a fixed parameter.
- Parameter Estimation (ETPF): The Ensemble Transform Particle Filter handles the parameters. Unlike traditional resampling that discards low-weight particles, the ETPF utilizes Optimal Transport (Wasserstein distances). It computes a linear transformation to shift the ensemble toward the posterior, maintaining particle diversity and preventing collapse.
- Interaction Mechanism: The algorithm maintains $L$ parallel state filters, each conditioned on a specific parameter realization. When parameter weights diverge significantly, the ETPF triggers a transform step that updates both the parameters and their associated state ensembles.
Experimental Evaluation
The method was tested on a 1D stochastic wave equation discretized into 100 spatial points, creating a 100-dimensional state space. The objective was to estimate the wave velocity parameter $c$ while tracking the wave's evolution under continuous, noisy observations. The setup utilized an ensemble of $M=100$ states for each of $L=20$ parameter realizations.
Key Results
The hybrid filter demonstrated robust performance in the benchmark test. As shown in the figure below, both the ETPF-based hybrid approach and the standard EnKF successfully converged to the true wave velocity parameter within a short assimilation window.
While the EnKF performed well in this nearly Gaussian test case, the ETPF's primary advantage lies in its theoretical robustness to highly non-Gaussian parameter distributions where the EnKF typically struggles. The use of optimal transport ensured that the filter maintained a diverse set of parameter candidates throughout the process.
Conclusion and Future Directions
This hybrid approach offers a scalable path for simultaneous estimation in high-dimensional systems. Future work aims to reduce the computational cost of the optimal transport step—currently $O(L^3 \log L)$—using the Sinkhorn algorithm to achieve $O(L^2)$ scaling. Additionally, the authors plan to explore "localization" techniques to handle spatially varying parameters, extending the method's applicability to even more complex physical models.
Summary updated