Robust Divergence Measures for Time Series Discrimination∗

New divergence measures are introduced for change detection and discrimination of stochastic signals (time series) on the basis of parametric filtering — a technique that combines parametric linear filtering with correlation characterization. The sensitivity of these divergence measures is investigated using local curvatures under additive and multiplicative spectral departure models. It is found that when the time series contain dominant spectral components (such as sharp peaks and notches) of similar characteristics, the new divergence measures are more effective in detecting the dominated spectral deviations than the conventional spectral divergence measures such as the Kolmogorov-Smirnov spectral distance and the Kullback-Leibler spectral divergence. Simulation results are given that illustrate and confirm the theoretical findings. Key wards and phrases: Characterization; Discriminant Analysis; Distortion Measure; Filter Bank; Pattern Recognition; Signal Processing; Spectral Analysis