Maximum Likelihood Estimation of Models with Unstable Dynamics and Non-minimum Phase Noise Zeros
1998 (English)Report (Other academic)
Maximum likelihood estimation of single-input/single-output linear timeinvariant (LTI) dynamic models requires that the model innovations (the nonmeasurable white noise source that is assumed to be the source of the randomness of the system) can be computed from the observed data. For many model structures, the prediction errors and the model innovations coincide and the prediction errors can be used in maximum likelihood estimation. However, when the model dynamics and the noise model have unstable poles which are not shared or when the noise dynamics have unstable zeros this is not the case. One such example is an unstable output error model. In this contribution we show that in this situation the model innovations can be computed by anti-causal filtering. Different implementations of the model innovations filter are also studied.
Place, publisher, year, edition, pages
Linköping: Linköping University Electronic Press, 1998. , 14 p.
LiTH-ISY-R, ISSN 1400-3902 ; 2043
Prediction error methods, Output error methods
IdentifiersURN: urn:nbn:se:liu:diva-55650ISRN: LiTH-ISY-R-2043OAI: oai:DiVA.org:liu-55650DiVA: diva2:316424