A Multitaper Test For The Detection of Non-Stationary Processes Using Canonical Correlation Analysis

Abstract

A new test has been designed for detecting the presence of non-stationary component processes in a time series. The test statistic is derived from the canonical correlations between two sets of eigencoefficients which are offset in frequency. The correlation coefficient which defines the test statistic has lower estimation variance than the multitaper, linear spectral-correlation coefficient, and it reveals important evidence of non-stationarity which is missed in the detector of the latter correlation coefficient.