Approximate conditional-mean type smoothers and interpolators

Abstract

A class of robust smoother and interpolator algorithms is introduced. The motivation for these smoothers and interpolators is a theorem concerning approximate conditional-mean smoothers for vector Markov processes in additive non-Gaussian noise. This theorem is the smoothing analog of Masreliez’s approximate non-Gaussian filter theorem (IEEE-Auto. Control, AC-20, 1975). The theorem presented here relies on the assumption that a certain conditional density is Gaussian, just as does Masreliez’s result. This assumption will rarely, if ever, be satisfied exactly. Thus a continuity theorem is also presented which lends support to the intuitive notion that the conditional density in question will be nearly Gaussian in a strong sense when the additive noise is nearly Gaussian in a comparatively weak sense. Approaches for implementing the robust smoothers and interpolators is discussed and an application to a real data set is presented.