A fast on-line generalized eigendecomposition algorithm for time series segmentation
Y.N. Rao, José Carlos Príncipe · 2002
This paper presents a novel, fast converging on-line rule for generalized eigendecomposition (GED) and its application in time series segmentation. We adopt the concepts of deflation and power method to iteratively estimate the generalized eigencomponents. The algorithm is guaranteed to produce stable results. In the second half of the paper, we discuss the application of GED to segment time series. GED is tested for chaotic time series and speech. The simulation results are compared with the venerable Generalized Likelihood Ratio Test (GLR) as a benchmark to gauge performance.