Detecting multiple change points: a PULSE criterion

Abstract

The research described herewith investigates detecting change points of means and of variances in a sequence of observations. The number of change points can be divergent at certain rate as the sample size goes to infinity. We define a MOSUM-based objective function for this purpose. Unlike all existing MOSUM-based methods, the novel objective function exhibits an useful ``PULSE" pattern near change points in the sense: at the population level, the value at any change point plus 2 times of the segment length of the moving average attains a local minimum tending to zero following by a local maximum going to infinity. This feature provides an efficient way to simultaneously identify all change points at the sample level. In theory, the number of change points can be consistently estimated and the locations can also be consistently estimated in a certain sense. Further, because of its visualization nature, in practice, the locations can be relatively more easily identified by plots than existing methods in the literature. The method can also handle the case in which the signals of some change points are very weak in the sense that those changes go to zero. Further, the computational cost is very inexpensive. The numerical studies we conduct validate its good performance.