Asymptotically Optimal Discrete Time Nonlinear Filters From Stochastically Convergent State Process Approximations

Dionysios S. Kalogerias, Athina P. Petropulu

Abstract

We consider the problem of approximating optimal in the Minimum Mean Squared Error (MMSE) sense nonlinear filters in a discrete time setting, exploiting properties of stochastically convergent state process approximations. More specifically, we consider a class of nonlinear, partially observable stochastic systems, comprised by a (possibly nonstationary) hidden stochastic process (the state), observed through another conditionally Gaussian stochastic process (the observations). Under general assumptions, we show that, given an approximating process which, for each time step, is stochastically convergent to the state process, an approximate filtering operator can be defined, which converges to the true optimal nonlinear filter of the state in a strong and well defined sense. In particular, the convergence is compact in time and uniform in a completely characterized measurable set of probability measure almost unity, also providing a purely quantitative justification of Egoroff's Theorem for the problem at hand. The results presented in this paper can form a common basis for the analysis and characterization of a number of heuristic approaches for approximating optimal nonlinear filters, such as approximate grid based techniques, known to perform well in a variety of applications.

Introduction

In signal processing and control, "filtering" refers to the estimation of a hidden stochastic process (the state) based on a sequence of noisy measurements (the observations). While the Kalman filter provides a perfect, recursive solution for linear Gaussian systems, real-world systems are frequently nonlinear and non-Gaussian. In these cases, exact recursive filters generally do not exist, forcing practitioners to rely on approximate methods such as Extended Kalman Filters, grid-based techniques, or particle filters.

This paper addresses a fundamental theoretical question: if we construct a filter by replacing a complex, true hidden state process with a simpler, mathematically "close" approximation, under what conditions will the approximate filter converge to the true optimal Minimum Mean Squared Error (MMSE) filter? Furthermore, how can we quantify the strength and reliability of this convergence?

Methodological Framework

The authors employ a measure-theoretic approach to decouple the complexities of the state-observation relationship. The methodology follows three primary steps:

  1. Change of Probability Measure: Under standard probability measures, the hidden state and observations are stochastically coupled, complicating the derivation of optimal filters. The authors utilize a technique based on the Radon-Nikodym derivative to shift the problem into a "reference" probability measure. In this reference frame, the state and observations are independent, and the observations behave as standard Gaussian white noise.
  2. State Approximation: Once decoupled, the true hidden state process is replaced with an approximating process parameterized by a resolution variable.
  3. Stochastic Convergence Analysis: The authors prove that if the state approximation converges to the true state—either in probability or through a stricter "conditionally weak" ($\mathcal{C}$-weak) convergence—the resulting approximate filtering operator will converge to the true optimal filter.

Key Insights and Theoretical Results

The Primacy of Observation Quality

A significant takeaway from this framework is that the internal stochastic structure of the hidden state process is less critical to filter convergence than the quality of the observations. As long as the state is bounded and the observations are a Lipschitz continuous function of the state, the filter remains robust. This suggests that the fidelity of the observation model is the primary driver of filtering performance.

Theorem 3: Strong Convergence Guarantees

The paper’s main result, Theorem 3, provides a mathematical guarantee that the absolute difference between the approximate filter and the true optimal MMSE filter vanishes as the approximation resolution increases. This convergence is characterized by two critical properties:

  • Uniformity: The convergence is uniform over any finite time horizon $T$.
  • Exponential Probability Bounds: The convergence holds true on a measurable set of events with a probability of at least $1 - (T+1)^{1-CN}\exp(-CN)$, where $N$ is the dimension of the observation vector and $C \geq 1$. This implies that the reliability of the filter approaches certainty exponentially fast as the dimensionality of the observation data grows.

Justification for Grid-Based Filters

The results provide a rigorous quantitative justification for heuristic, grid-based recursive filters. By showing that snapping a continuous state to a discrete grid (marginal quantization) leads to an asymptotically optimal filter, the paper validates methods commonly used in practical applications like wireless channel tracking.

Limitations and Future Research

While the paper establishes a powerful theoretical foundation, it operates under specific constraints. The hidden state process must be confined to a compact (bounded) set, and the observation noise must be conditionally Gaussian. Additionally, the paper focuses on the existence and convergence of these filters rather than their computational efficiency or recursive implementation.

Finally, the authors note that while their mathematical representation of the filter resembles Importance Sampling, these theorems cannot yet be directly applied to Particle Filters. Particle filters rely on empirical measures rather than simple state approximations, and bridging this gap to provide similar asymptotic guarantees remains an open area for future research.

Summary updated

Abstract

We consider the problem of approximating optimal in the Minimum Mean Squared Error (MMSE) sense nonlinear filters in a discrete time setting, exploiting properties of stochastically convergent state process approximations. More specifically, we consider a class of nonlinear, partially observable stochastic systems, comprised by a (possibly nonstationary) hidden stochastic process (the state), observed through another conditionally Gaussian stochastic process (the observations). Under general assumptions, we show that, given an approximating process which, for each time step, is stochastically convergent to the state process, an approximate filtering operator can be defined, which converges to the true optimal nonlinear filter of the state in a strong and well defined sense. In particular, the convergence is compact in time and uniform in a completely characterized measurable set of probability measure almost unity, also providing a purely quantitative justification of Egoroff's Theorem for the problem at hand. The results presented in this paper can form a common basis for the analysis and characterization of a number of heuristic approaches for approximating optimal nonlinear filters, such as approximate grid based techniques, known to perform well in a variety of applications.
BibTeX
@article{kalogerias2015asymptotically,
  author = {Dionysios S. Kalogerias and Athina P. Petropulu},
  title = {{Asymptotically Optimal Discrete Time Nonlinear Filters From Stochastically Convergent State Process Approximations}},
  journal = {IEEE Trans. Signal Process.},
  year = {2015},
  url = {https://openreview.net/forum?id=wmmCR5rqSJ},
}
Figure