Fast time-invariant implementations for linear least-squares smoothing filters

TLDR: In this article, a new solution for the fixed interval linear least-squares smoothing of a random signal, finite dimensional or not, inadditive white noise is presented, where the smoothed estimate for stationary processes is expressed entirely in terms of time-invariant causal and anticausal filtering operations.

Abstract: We present a new solution for the fixed interval linear least-squares smoothing of a random signal, finite dimensional or not, inadditive white noise. By using the so-called Sobolev identity of radiative transfer theory, the smoothed estimate for stationary processes is expressed entirely in terms of time-invariant causal and anticausal filtering operations; these are interpreted from a stochastic point of view as giving certain constrained (time-invariant) filtered estimates of the signal. Then by using a recently introduced notion of processes close to stationary, these results are extended in a natural way to general nonstationary processes. From a computational point of view, the representations presented here are particularly convenient, not only because time-invariant filters can be used to find the smoothed estimate, but also because a fast algorithm based on the so-called generalized Krein-Levinson recursions can be used to compute the time-invariant filters themselves.