Discriminant and cluster analysis for Gaussian stationary processes:: Local linear fitting approach

This article is concerned with discrimination and clustering of Gaussian stationary processes. The problem of classifying a realization X-n = (X-1, ..., X-n)(t) from a linear Gaussian process X into one of two categories described by their spectral densities f(1)(lambda) and f(2)(lambda) is considered first. A discrimination rule based on a general disparity measure between every f(i)(lambda), i = 1, 2, and a nonparametric spectral density estimator (f) over cap (n)(lambda) is studied when local polynomial techniques are used to obtain f W. In particular, three different local linear smoothers are considered. The discriminant statistic proposed here provides a consistent classification criterion for all three smoothers in the sense that the misclassification probabilities tend to zero. A simulation study is performed to confirm in practice the good theoretical behavior of the discriminant rule and to compare the influence of the different smoothers. The disparity measure is also used to carry out cluster analysis of time series and some examples are presented and compared with previous works.