Optimal segmentation of signals in a linear regression framework

The problem of estimating the time instants when the dynamical properties of a signal make abrupt changes is studied. This segmentation problem is usually considered as exponential in time. The author presents a specific but natural signal mode-called a changing regression model-and points out a method to compute an optimal estimate of the segmentation problem linearly in time. The linear constant is always less than one and decreases to zero as the measurement noise decreases to zero. The method is thus asymptotically efficient in the measurement noise