McKean--Vlasov SDEs in Nonlinear Filtering

Sahani Pathiraja, Sebastian Reich, Wilhelm Stannat

Abstract

Various particle filters have been proposed over the last couple of decades with the common feature that the update step is governed by a type of control law. This feature makes them an attractive alternative to traditional sequential Monte Carlo which scales poorly with the state dimension due to weight degeneracy. This article proposes a unifying framework that allows to systematically derive the McKean-Vlasov representations of these filters for the discrete time and continuous time observation case, taking inspiration from the smooth approximation of the data considered in Crisan & Xiong (2010) and Clark & Crisan (2005). We consider three filters that have been proposed in the literature and use this framework to derive Itô representations of their limiting forms as the approximation parameter $\delta \rightarrow 0$. All filters require the solution of a Poisson equation defined on $\mathbb{R}^{d}$, for which existence and uniqueness of solutions can be a non-trivial issue. We additionally establish conditions on the signal-observation system that ensures well-posedness of the weighted Poisson equation arising in one of the filters.

Introduction

In the field of nonlinear filtering, the primary objective is to estimate the hidden state of a dynamic system based on a stream of noisy observations. While Sequential Monte Carlo (SMC) methods—commonly known as particle filters—have been the standard for decades, they face a significant hurdle: weight degeneracy. In high-dimensional state spaces, the importance weights of particles tend to concentrate on a single individual, rendering the ensemble an ineffective representation of the posterior distribution.

To circumvent this, researchers have developed "particle flow" filters, such as the Feedback Particle Filter (FPF) and the Reich filter. Rather than reweighting particles, these algorithms utilize a control law to physically transport particles toward regions of high posterior probability. However, these methods were historically derived under disparate assumptions. This paper provides a unified mathematical framework using McKean–Vlasov stochastic differential equations (SDEs) to bridge these approaches and establish their theoretical foundations.

A Unified McKean–Vlasov Framework

The authors leverage the concept of the mean-field limit, where the number of particles $N$ tends toward infinity. In this limit, the evolution of a single particle is described by a McKean–Vlasov SDE, where the drift and diffusion coefficients depend on the particle's own state and the overall distribution of the population.

By employing a smooth approximation of the data—specifically, a piecewise linear interpolation of the observation path—the authors analyze the limit as the interpolation step size $\delta$ approaches zero. This technique allows for a rigorous transition from discrete-time observations to continuous-time filtering equations. A key contribution of this work is the "Mapping Lemma" (Lemma 3.5), which provides a systematic way to map filtering equations across these temporal domains.

The Centrality of the Poisson Equation

A defining characteristic of particle flow filters is the requirement to solve a Poisson equation on the state space $\mathbb{R}^d$. This equation determines the "velocity field" or control law necessary to move particles from the prior distribution to the posterior.

The paper demonstrates that the differences between prominent filters—the Crisan & Xiong filter, the Reich filter, and the Feedback Particle Filter—essentially arise from different assumptions regarding the form of this control law. Under this unified view, the transport of particles is fundamentally a problem of solving or approximating these underlying Poisson equations.

Theoretical Rigor and Well-Posedness

Beyond unification, the paper addresses the critical question of whether these filters are mathematically well-posed. The existence and uniqueness of solutions to the Poisson equations are not guaranteed in nonlinear settings.

The authors utilize Poincaré inequalities to establish sufficient conditions for stability. They prove that if the signal-observation system satisfies certain criteria, such as strong concavity of the potential, the resulting filtering distribution is well-behaved. Specifically, Theorems 5.2 and 6.2 provide the mathematical guarantees for the existence and uniqueness of weak solutions to the Poisson equations used in the FPF and Crisan & Xiong filters.

Furthermore, the authors demonstrate "Linear-Gaussian Consistency." They prove that in the case of linear systems with Gaussian noise, these complex nonlinear filters naturally simplify to the well-known Ensemble Kalman-Bucy Filter (EnKBF), validating the framework's consistency with established theory.

Conclusion and Future Directions

This research provides the first precise Itô SDE representations for several continuous-time filters that were previously only understood in discrete or heuristic terms. While the framework is robust, it highlights several areas for future inquiry:

  1. Relaxing Constraints: The conditions for well-posedness (like strong concavity) may be too restrictive for some real-world nonlinear systems.
  2. Numerical Scalability: Solving Poisson equations in high dimensions remains computationally expensive, often requiring further approximations like kernel-based methods or constant gain assumptions.
  3. Convergence Analysis: The rigorous justification of convergence rates for the smooth approximations used in this framework remains an open mathematical challenge.

By providing a rigorous theoretical backbone for particle flow, this work moves the field closer to reliable, high-dimensional nonlinear filtering.

Summary updated

Abstract

Various particle filters have been proposed over the last couple of decades with the common feature that the update step is governed by a type of control law. This feature makes them an attractive alternative to traditional sequential Monte Carlo which scales poorly with the state dimension due to weight degeneracy. This article proposes a unifying framework that allows to systematically derive the McKean-Vlasov representations of these filters for the discrete time and continuous time observation case, taking inspiration from the smooth approximation of the data considered in Crisan & Xiong (2010) and Clark & Crisan (2005). We consider three filters that have been proposed in the literature and use this framework to derive Itô representations of their limiting forms as the approximation parameter $δ\rightarrow 0$. All filters require the solution of a Poisson equation defined on $\mathbb{R}^{d}$, for which existence and uniqueness of solutions can be a non-trivial issue. We additionally establish conditions on the signal-observation system that ensures well-posedness of the weighted Poisson equation arising in one of the filters.
BibTeX
@article{pathiraja2021mckean,
  author = {Sahani Pathiraja and Sebastian Reich and Wilhelm Stannat},
  title = {{McKean--Vlasov SDEs in Nonlinear Filtering}},
  journal = {SIAM J. Control. Optim},
  year = {2021},
  doi = {10.1137/20m1355197},
}
Figure