Wiener Space Cubature for Nonlinear Gaussian Filtering

Abstract

Continuous-discrete filtering is a classical problem in signal processing and control theory, whereby the state driving a stochastic differential equation (SDE) must be inferred from noisy, partial discrete-time observations. Corresponding filtering algorithms are often formed as discrete-time methods and applied to a discretization of the SDE, an approach limited by the incurred discretization error, or model a type of Gaussian moment ordinary differential equation (ODE) flow, which is approximate for nonlinear dynamics. We introduce a new method, the Wiener-space cubature Kalman filter (WSC-KF), for Gaussian filtering in this continuous-discrete setting. The WSC-KF combines spatial sigma-point representations with Wiener space cubature integration, effectively approximating iterated integral terms of the driving Brownian motion to higher-order than standard discretization (e.g., Ito-Taylor expansions), or approximate moment-ODE flows (e.g., as in the unscented Kalman filter). Theoretically, we characterize the path-law approximation degree of WSC-KF through a tunable model order parameter. We prove that, depending on the model order chosen, WSC-KF provides weak order bounds comparable to or dominating that of a cubature Kalman filter using a first-order Ito predict step. Numerically, we confirm that rate on a globally Lipschitz Levy-area diffusion. Degree-5 Wiener cubature matches mixed iterated integrals that existing continuous-discrete Kalman predictors omit, while remaining efficiently computable; this is the mechanism enabling higher-order weak path-law approximation.