Finite dimensional filters with nonlinear drift. XI: Explicit solution of the generalized Kolmogorov equation in Brockett-Mitter program (Q1281885): Difference between revisions

Let a filtering problem be based on the following signal-observation model: \[ \begin{aligned} dx_t & = f(x_t)dt+ g(x_t) dv_t,\;x_0\in\mathbb{R}^n,\\ dy_t & = h(x_t)dt+ dw_t,\;y_0= 0\in\mathbb{R}^m,\end{aligned} \] where \(v\) and \(w\) are independent Brownian processes, \(f\) and \(g\) are \(C^\infty\) smooth, \(g\) takes its values in the set of orthogonal matrices. As usual, if \(\rho(t,x)\) denotes the conditional density of the state given the observation \(\{y_s, s\leq t\}\), the unnormalized filter \(\sigma(t,x)\) satisfies the well-known Duncan-Mortensen-Zakai equation: \[ d\sigma(t,x)= (L_0\sigma)(t,x)dt + \sum^m_{i=1} (h^i\sigma)(t,x) dy^i_t,\;\sigma(0,x)= x_0, \] where \(L_0={1\over 2}\Delta- \sum^n_{i=1} f_i\partial_{x_i}- \sum^n_{i=1} {\partial f_i\over\partial x_i}-{1\over 2} \sum^n_{i=1} h^2_i\). The change of function \(u(t,x)= (\exp- \langle h(x),y_t\rangle) \sigma(t,x)\) leads to the differential equation \[ {\partial u\over\partial t}= (L_0u)(t,x)+ \langle y_t,[L_0, L_i]u\rangle+{1\over 2} \sum^m_{i,j= 1} y^i_t y^j_t[[L_0, L_i]; L_j] u(t,x),\;u(0,x)= \sigma_0. \] This is near the generalized Kolmogorov equation. The main results are a formal solution on \(\mathbb{R}^n\) of the Kolmogorov equation (integral form on \(\mathbb{R}^n)\) and another solution as a convergent infinite series. So, an explicit finite sum is an approximation to the solution; the error for such an approximation is given. About finite-dimensional filters see also \textit{V. E. Beneš}, Stochastics 5, 65-92 (1981; Zbl 0458.60030).