Nonlinear Filtering Revisited: A Spectral Approach (Q4337413)

It is known that the optimal filter \(\widehat f(x(t))\) for \(f(x(t))\) based on observations \(y(s)\), \(s\leq t\), is given by the formula \[ \widehat f(x(t))=\int_{\mathbb{R}^d}f(x)u(t,x)dx\Biggl/ \int_{\mathbb{R}^d}u(t,x)dx, \] where the unnormalized filtering density \(u(t,x)\) is a solution of the Zakai equation. The authors prove the formula \[ u(t,x)=\sum_\alpha{1\over \sqrt{\alpha!}}\varphi_\alpha(t,x)\xi_\alpha(y(\cdot)), \] where \(\xi_\alpha(y(\cdot))\) are Wick polynomials and \(\varphi_\alpha(t,x)\) are deterministic coefficients satisfying a recursive system of Kolmogorov-like equations. The formula separates the observations and parameters of the considered filtering scheme: \(\xi_\alpha(y(\cdot))\) are completely defined by the observation process \(y(t)\) and \(\varphi_\alpha(t,x)\) are determined by coefficients of equations for \(x(t)\), \(y(t)\) and by the initial distribution of the signal process \(x(t)\). The authors develop a numerical approximation scheme for \(u(t,x)\). Some results of numerical simulations are presented.