Abstract
The paper is concerned with non-linear Gaussian filtering and smoothing in continuous-discrete state-space models, where the dynamic model is formulated as an Itô stochastic differential equation (SDE), and the measurements are obtained at discrete time instants. We propose novel Taylor moment expansion (TME) Gaussian filter and smoother which approximate the moments of the SDE with a temporal Taylor expansion. Differently from classical linearisation or Itô--Taylor approaches, the Taylor expansion is formed for the moment functions directly and in time variable, not by using a Taylor expansion on the non-linear functions in the model. We analyse the theoretical properties, including the positive definiteness of the covariance estimate and stability of the TME Gaussian filter and smoother. By numerical experiments, we demonstrate that the proposed TME Gaussian filter and smoother significantly outperform the state-of-the-art methods in terms of estimation accuracy and numerical stability.
Introduction
In many real-world applications like target tracking or signal processing, we deal with continuous-discrete state-space models. Here, the system's "truth" evolves continuously (modeled by Stochastic Differential Equations or SDEs), but we only get noisy observations at specific, discrete time intervals.
The primary challenge in this field is the transition density—the probability of moving from one state to another over a time step. For non-linear systems, this density is mathematically intractable. Existing solutions usually fall into two categories:
- Itô–Taylor expansions: These discretize the SDE solution itself (e.g., Euler-Maruyama). They are simple but struggle with accuracy unless the time step is very small.
- ODE approaches: These try to solve ordinary differential equations for the mean and covariance. However, these ODEs involve expectations that cannot be solved analytically, forcing the use of approximations (like linearization in the EKF) that often fail in highly non-linear systems.
This work introduces Taylor Moment Expansion (TME) to bridge this gap with a method that is both high-order and numerically robust.
Methodology
The core innovation is the Taylor Moment Expansion (TME). Instead of approximating the non-linear functions in the model or the SDE solution, the authors apply a Taylor expansion to the moment functions themselves (the mean and covariance) with respect to the time variable.
The Infinitesimal Generator
The method utilizes the "infinitesimal generator" ($\mathcal{A}$) associated with the SDE. This operator allows for the calculation of time-derivatives of expectations of any smooth function. By repeatedly applying this generator, an $M$-th order Taylor expansion of the moments is formed in terms of the time step $\Delta t$.
Direct Moment Estimation
Unlike Itô–Taylor methods that require complex "iterated integrals," TME provides a direct polynomial approximation for the transition moments. The paper also includes a formal analysis of the positive definiteness of the estimated covariance and the mean-square stability of the resulting filter.
Experimental Evaluation
The method was evaluated using complex synthetic benchmarks:
- 1D Non-linear SDEs: Models with
tanhdrift and state-dependent dispersion to test basic moment estimation. - 3D Coordinate Turn Model: A 7-dimensional model used in radar tracking involving highly non-linear coordinate transformations (Cartesian to polar) and turning dynamics. This serves as the primary benchmark for comparing RMSE and numerical stability.
Key Results
The TME approach demonstrates superior accuracy and stability compared to traditional methods.
- Superior Accuracy: In the 3D coordinate turn model, TME-3 and TME-4 filters significantly outperformed standard Extended Kalman Filters (EKF) and Unscented Kalman Filters (UKF) variants, especially as the measurement interval ($\Delta t$) increased. As shown in the figure, TME tracks the Monte Carlo ground truth for both mean and variance much more accurately than Itô-1.5 or RK4-based methods, which tend to diverge as the time interval increases.
- Numerical Stability: TME methods exhibited far fewer "divergences" compared to traditional ODE-based solvers. While standard Gauss-ODE filters often failed when the time step was large, TME remained robust.
- Efficiency: Despite being higher-order, the computational cost of TME is comparable to standard sigma-point methods (like the UKF), making it viable for real-time applications.
Discussion and Future Work
A key insight is that targeting moments directly via the generator is more efficient than approximating the SDE solution. While TME-3 often provides a "sweet spot" for stability and accuracy, the TME covariance is a truncated polynomial and is not guaranteed to be positive definite for arbitrarily large time steps.
Future research may explore adaptive expansion orders or combining TME with more sophisticated numerical integration rules to further enhance robustness in extremely non-linear environments.
Summary updated