Nonlinear filtering of diffusion processes in correlated noise: Analysis by separation of variables (Q1401570)

The paper is motivated by the nonlinear filtering problem. Let \(X\) be an (unobservable) diffusion process in \(\mathbb R^{d}\), the observable process \(Y\) in \(\mathbb R^{m}\) being given by \(Y_{t} = h(X_{s}) ds + W_{t}\). The process \(X\) and the Wiener process \(W\) are not assumed to be independent. To compute the best mean square estimate of \(f(X_{t})\) given the trajectory \(Y_{s}\), \(s\leq t\), it is desirable to know the unnormalized filtering density \(p\). The random field \(p\) solves the Zakai equation \(dp = \mathcal L^{*}p dt + \mathcal M^{*} dY\), where \(\mathcal L\) is an elliptic differential operator (the generator of the process \(X\)). Since the noise-correlated case is considered, \(\mathcal M\) is an unbounded operator, which makes the analysis more difficult. The goal of the paper is to study an algorithm for solving the Zakai equation of the following type: Let \(\{e_{k}\}\) be an orthonormal basis in \(L^2(\mathbb R^{d})\); as the first step, let us define \(p^{K}\) as the Galerkin approximation to \(p\), \(p^{K}(t,x) = \sum^{K}_{k=1} p^{K}_{k}(t)e_{k}(x)\). The second step consists in expanding the processes \(p^{K}_{k}\), for a fixed time \(t_0\), using the Wiener chaos decomposition, which leads to \[ p^{K}(t_0,x) = \sum^{K}_{k=1}\Biggl(\sum^\infty_{m=1} \varphi^{K}_{m,k}(t_0)\xi_{m}\Biggr)e_{k}(x). \] Explicit representation for the coefficients \(\varphi^{K}_{m,k}\) is found and an upper bound on the approximation error is derived.