Uniqueness of the solution of the filtering equations in spaces of measures

Dan Crisan,Etienne Pardoux

2024-11-18

Abstract:Nonlinear filtering is a pivotal problem that has attracted significant attention from mathematicians, statisticians, engineers, and various other scientific disciplines. The solution to this problem is governed by the so-called filtering equations. In this paper, we investigate the uniqueness of solutions to these equations within measure spaces and introduce a novel, generalized framework for this analysis. Our approach provides new insights and extends the applicability of existing theories in the study of nonlinear filtering.

Probability,Functional Analysis

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to solve the uniqueness problem of the solutions of filtering equations in nonlinear filtering problems. Specifically, the goal of the paper is to study the uniqueness of the solutions of filtering equations in the measure space and introduce a new and more general framework to analyze these problems. The following are the main contributions and goals of the paper:

1. **Nonlinear filtering problem**:
Nonlinear filtering is a crucial problem that has attracted extensive attention from mathematicians, statisticians, engineers, and other scientific fields. The core problem is to estimate a stochastic system (signal process \(X\)) that evolves over time, and this signal process is indirectly observed through another related process (observation process \(Y\)).

2. **Filtering equations**:
The solutions of the filtering problem are governed by the so - called "filtering equations". These equations describe how the conditional distribution of the signal evolves over time given the observed data. Specifically, the filtering equations consist of two parts:
- **Nonlinear Kushner - Stratonovich equation**: Used to describe the evolution of the conditional distribution.
- **Linear Zakai equation**: Used to describe the evolution of the unnormalized conditional distribution.

3. **Uniqueness problem**:
The focus of the paper is to prove that these filtering equations have unique solutions in an appropriately chosen measure space. The author proposes a new framework that is as close to a symmetric form as possible, thus generalizing the existing theoretical framework. Under this framework, the relationship between the signal and the observation process is more complex, and the signal process is driven not only by its own Brownian motion but also by the Brownian motion of the observation process.

4. **Main contributions**:
- Derived new filtering equations satisfied by the signal and the observation process, and the coefficients of these equations depend on the signal - observation pair.
- Proved that the uniqueness of the solutions of the nonlinear Kushner - Stratonovich equation and the linear Zakai equation are equivalent.
- Established the uniqueness of the solutions of the unnormalized conditional distribution in the space of measure - valued processes.

5. **Innovations**:
- Introduced the backward stochastic partial differential equation (BSPDE) to replace the traditional deterministic backward PDE to overcome the difficulties in existing methods.
- Utilized the Sobolev embedding theorem to ensure the smoothness of the BSPDE solutions, so that they can be integrated with any measure - valued solution.

In summary, this paper provides a deeper understanding of nonlinear filtering problems by introducing new mathematical tools and frameworks and proves the uniqueness of the solutions of filtering equations, which is of great significance for both theory and applications.