Complex Autoregressive Model and Its Properties
Iwao Sekita, Takio Kurita, Nobuyuki Otsu · 1999
This paper shows the following three properties of the CAR (complex autoregressive) model which were proposed by the authors for shape description and recognition. The CAR coe#cients estimated by the least squares fit are shown to be the maximum likelihood estimates and also the estimates obtained by maximizing the mutual information between boundary point z j and the previous m boundary points {z j-k } m k=1 on the assumption that each prediction error is a complex Gaussian random variable. And determining the (m + 1)-th complex autocorrelation coe#cient r m+1 so as to maximize the entropy of the consecutive (m+2) boundary points on the assumption that the points are a (m+2)-dimensional complex Gaussian random variable, is equivalent to determining the coe#cient r m+1 by the CAR model of order m on the assumption of no correlation between the prediction error # j and the boundary point z j-m-1 . # Mathematical Informatics Section, Information Science Division + Currently, Director of Machine Understanding Division 1.