Recursive filtering algorithms for systems with piecewise-linear nonlinearities (Q1090303)

Consider a discrete-time filtering problem where the signal process \(\theta_ t\) and the observation process \(\xi_ t\) are solutions of \[ \theta_{t+1}=f(\theta_ t)+F(\theta_ t)\epsilon_{t+1},\quad \xi_{t+1}=h(\theta_{t+1})+H(\theta_{t+1})\eta_{t+1} \] for independent white noises \(\epsilon_ t\) and \(\eta_ t\). If f, F are piecewise constant, \(H=I\) and with Gaussian initial conditions, it was shown by \textit{G. B. Di Masi}, \textit{M. Pratelli} and \textit{W. J. Runggaldier} [Stochastics 14, 247-271 (1985; Zbl 0566.60046)] that the conditional density of \(\theta_ t\) is a linear combination of fixed Gaussian densities the coefficients of which can be computed in a recursive form. This work is concerned with a generalization to piecewise linear nonlinearities (f and h piecewise linear, F and H piecewise constant); in two cases, namely if f is piecewise constant or \(F=0\), the same conclusion is proved to hold, except that the Gaussian densities now depend on the observation; the recursive formulae involve the computation of integrals on the regions of linearity.