Abstract
Filter stability is a classical problem for partially observed Markov processes (POMP). For a POMP, an incorrectly initialized non-linear filter is said to be stable if the filter eventually corrects itself with the arrival of new measurement information. In the literature, studies on the stability of non-linear filters either focus on the ergodic properties on the hidden Markov process, or the observability of the measurement channel. While notions of observability exist in the literature, they are difficult to verify for many systems and specific examples of observable systems are mostly restricted to additive noise models. In this paper, we introduce a general definition of observability for stochastic non-linear dynamical systems and compare it with related findings in the literature. Our notion involves a functional characterization which is easily computed for a variety of systems as we demonstrate. Under this observability definition we establish filter stability results for a variety of criteria including weak merging and total variation merging, both in expectation and in an almost sure sense, as well as relative entropy. We also consider the implications between these notions, which collect various results in the literature in a clear and concise manner. Our conditions, and the examples we study, complement and generalize the existing results on filter stability in the literature.
Introduction
In many engineering and science applications, we deal with Partially Observed Markov Processes (POMPs), where a hidden state sequence ($X_n$) must be estimated from noisy observations ($Y_n$). A central challenge in this context is filter stability: if an observer initializes their filter with an incorrect probability distribution (a "prior"), will the filter eventually converge to the true state distribution as data accumulates?
While filter stability is well-understood for linear systems via the Kalman filter, non-linear systems present significant difficulties. Traditional research often relies on the ergodicity of the hidden process—the property that the system "forgets" its initial state. However, many practical systems are not ergodic. This work shifts the focus toward stochastic observability, the principle that measurements themselves can provide sufficient information to override an incorrect initial guess, independent of the system's internal forgetting properties.
Stochastic Observability
The authors propose a new functional characterization of observability that is more verifiable than previous abstract definitions.
- N-Step Observability: A system is defined as observable if any continuous, bounded function of the state ($f(x)$) can be approximated with arbitrary precision by a function of the observations ($g(y)$). This essentially asks whether the measurements contain enough "signal" to reconstruct meaningful properties of the state.
- Predictor-to-Filter Mapping: The framework proves that observability ensures the stability of the predictor (the estimate of the next state). Through continuity arguments, it is shown that a stable predictor implies a stable filter (the estimate of the current state).
Stability Criteria and Hierarchy
A key contribution of this paper is the establishment of a hierarchy between different notions of stability. Using tools such as Pinsker’s Inequality and Blackwell and Dubins’ merging of opinions, the authors map the relationships between weak merging, total variation (TV) merging, and relative entropy merging.
As illustrated in the hierarchy above, the paper demonstrates how "weak merging" (convergence of averages) can lead to "total variation merging" (convergence of the entire distribution shape) and eventually "relative entropy merging" (convergence in information content) under specific continuity and observability conditions.
Key Theoretical Results
- Theorem II.7: Establishes that observability guarantees filter stability in a weak sense.
- Theorems II.18 & II.19: Show that under mild continuity conditions, weak stability implies stability in Total Variation (TV).
- Theorem II.20: Links TV stability to Relative Entropy stability, connecting filter theory with information theory and robust control.
Evaluation and Synthetic Models
The theoretical framework is validated using several synthetic mathematical models that represent common challenges in non-linear filtering:
- Finite State/Noise Spaces: Discrete systems where observability is checked via matrix rank conditions.
- Affine Observations: Systems with linear measurement functions plus noise.
- Non-linear Measurement Functions: Complex relationships, such as "telescoping" measurements.
- Non-Compact State Spaces: Extension to Euclidean spaces ($ℝ^n$) where the state may not be bounded.
Conclusion and Future Directions
This work provides a rigorous roadmap for filter stability based on observability. It clarifies that perfect invertibility of measurements is not required; rather, measurements must be sufficiently informative to diminish the influence of the initial prior.
Future work may address "active" filtering, where an agent's actions influence observability. Additionally, while the functional definition is verifiable, finding the optimal approximation functions $g(y)$ in high-dimensional spaces remains a computational challenge. Finally, the reliance on continuous noise densities for the strongest results suggests further investigation into systems with quantized or discrete sensors.
Summary updated