Convergence of a Branching Particle Method to the Solution of the Zakai Equation (Q4210052)

A finite-dimensional diffusion process \(X_t\) is defined via its generator, an observation process \(Y_t\) satisfies an equation \(Y_t = \int_0^t h(X_s) ds + W_t\). An unnormalized conditional measure \(U(t)\) which satisfies a Zakai equation is approximated via branching process measures \(U_n(t)\) as follows. \(U_n(0)\) is the empirical measure equal to \(n^{-1}\sum_{i=1}^n\delta_{x^{i}_{n}}\) with some \(x^i_n\)'s such that \(U_n(0)\) tends weakly to the initial distribution of \(X\). All \(n\) particles behave independently as diffusions with the same generator on the interval \([0,1/n)\). At time \(1/n\) each particle with the path \(V_t\) branches so that the mean number of offspring given \(\sigma(V_t, t<1/n)\) equals \[ \exp\left(\int_0^{1/n} h^*(V_t) dY_t - (1/2)\int_0^{1/n} h^*h(V_t) dt \right). \] It is possible to construct a branching mechanism so that the variance of the number of offspring is finite. The procedure is then repeated on each interval \([i/n, (i+1)/n)\). A mean-square convergence of \((U_n(t),\varphi)\) and a convergence of \(U_n\) to \(U\) in distribution in the trajectory space \(D\) is established. A computer simulation method is described. Numerical examples and pictures are provided.