Some new smoother implementations for discrete-time gaussian reciprocal processes (Q3979755)

This paper deals with discrete time Gaussian reciprocal processes on a finite time interval. These processes can be described by a second order linear descriptor model. For linear observations with Gaussian white noise the optimal smoothed estimates obey second order descriptor type recursions with cyclic or Dirichlet type boundary conditions. This smoother is non-causal, but it can be implemented using two causal operations, one forward and one backward in time. Various implementations are discussed in this paper: Two-filter solutions, extending the Mayne- Fraser smoother, double-sweep solutions, extending the Rauch-Tung- Striebel approach, and a multi-rate solution using the structure of cyclic block diagonal matrices. A stochastic filtering interpretation of the two-filter solution is given, and various examples are presented. Much of the material depends on characterizations of Gaussian reciprocal processes given by the second author, \textit{R. Frezza} and \textit{A. J. Krener} [IEEE Trans. Autom. Control 35, No. 9, 1013-1023 (1990; Zbl 0709.60043)] and \textit{A. J. Krener}, \textit{R. Frezza} and the second author [Stochastics Stochastics Rep. 34, No. 1/2, 29-56 (1991; Zbl 0724.60063)].