Nonlinear continuous-discrete filtering using kernel density estimatesand functional integrals (Q4409372)

Simulation studies are conducted for the continuous-discrete filtering problem related to (nonlinear) ordinary stochastic differential equations (SDE, as continuous part) \[ dy(t) = f(y(t),t,\psi) dt + g(y(t),t,\psi) dW(t) \] driven by an \(r\)-dimensional standard Wiener process \(W=(W(t))_{t_0 \leq t \leq t_T}\), where the discrete measurements \(z_i\) are taken according to the scheme \[ z_i = h(y(t_i),t_i,\psi) + \varepsilon_i \] with \(k\)-dimensional random variables \(\varepsilon_i \in {\mathcal N}(0,R(t_i,\psi))\) at the times \(\{t_0,t_1,\dots ,t_T\}\). It is assumed that at least one component of \(z_i\) can be observed at each time \(t_i\), whereas the nonobserved parts are dropped as missing data. Five filter algorithms are presented and tested with the examples of the Ginzburg-Landau equation with additive noise and a stochastic volatility model with time-dependent diffusion term. Among them are the kernel density filter, discretized continuous sampling approach, the filter based on the functional path-integral representation of the transition density, extended Kalman filter and the Gaussian sum filter. Weak Euler-Maruyama methods, driven by uniformly distributed martingale-differences with variances \(\Delta_i= t_{i+1}-t_i\) are used to approximate the conditional densities \(p(y,t\mid z(t): t \leq t_i)\) of the related SDE in the Monte Carlo simulation part.