Authors
Hüselitz A, Li H, Munk A
Journal
Arxiv
Citation
arXiv:2503.05270.
Abstract
We consider the problem of sequential estimation of a single change point in a piecewise linear regression model under a Gaussian setup. We demonstrate that a certain CUSUM-type statistic attains the minimax optimal rates for localizing the change point. Our minimax analysis unveils an interesting phase transition from a jump (discontinuity in values) to a kink (change in slope). Specifically, for a jump, the minimax rate is of order log(n)/n, whereas for a kink it scales as (log(n)/n)1/3, given that the sampling rate is of order 1/n. We further introduce an algorithm for the proposed online change point detector, which requires constant computational steps and constant memory per incoming sample. Finally, the empirical performance of our method is examined on both simulated and real-world data sets. An implementation is available in the R package FLOC on GitHub.
