The nonlinear filtering problem for the unbounded case (Q1061415)

In two earlier papers [Appl. Math. Optimization 10, 159-185 (1983; Zbl 0525.93063), and ''Measure valued equations for the optimum filter in finitely additive nonlinear filtering theory.'' Z. Wahrscheinlichkeitstheor. Verw. Geb. 66, 1-18 (1984)] the authors had initiated a study of nonlinear filtering theory with Gaussian white noise (on a finitely additive probability space) replacing the differential of Brownian motion as noise in the conventional model for nonlinear filtering based on stochastic calculus. One of the advantages of the finitely additive approach is that one can work with the natural sample space, and hence does not have to enlarge the sample space of observations. Yet another of its advantages is that the equations for the optimal filter and the conditional densities turn out to be partial differential equations (PDE's) rather than stochastic PDE's. The two earlier papers of the authors had dealt with the cases when (i) the signal process takes values in a finite-dimensional Euclidean space and the measurement function is bounded, and (ii) the signal process is infinite dimensional. The present paper extends further on the former and solves, for a vector-valued signal process, the problem of existence of the unnormalized conditional density and its characterization as the unique solution to a 'Zakai' type equation without imposing any growth conditions on the measurement function. The principal tools used in the derivation of this general result are the finitely additive version of the Bayes formula and a dual Feynman-Kac type formula.