Bibliografie

(eng): This paper contributes to the literature on Bayesian filtering in the case where the processes driving the states and observations are uniformly distributed on finite intervals. We introduce the class of uniform distributions on parallelotopic supports (UPS). We derive optimal local distributional projections (i.e. approximations) within this UPS class-in the sense of minimum Kullback-Leibler divergence-of the outputs of the data and time updates of filtering. We demonstrate that the UPS class provides a tighter approximation (and therefore more precise inferences) than a previously reported approximation on orthotopic supports. It does this, while still achieving bounded complexity in the resulting recursive filtering algorithm. The comparative performance of the UPS-closed filtering algorithm is explored-via both Bayesian and frequentist performance measures-as a function of signal-to-noise ratio and state dimension in a position-velocity system.