Total: 1
The paper considers the problem of detecting and localizing changepoints in a sequence of independent observations. We propose to evaluate a local test statistic on a triplet of time points, for each such triplet in a particular collection. This collection is sparse enough so that the results of the local tests can simply be combined with a weighted Bonferroni correction. This results in a simple and fast method, {\sl Lean Bonferroni Changepoint detection} (LBD), that provides finite sample guarantees for the existance of changepoints as well as simultaneous confidence intervals for their locations. LBD is free of tuning parameters, and we show that LBD allows optimal inference for the detection of changepoints. To this end, we provide a lower bound for the critical constant that measures the difficulty of the changepoint detection problem, and we show that LBD attains this critical constant. We illustrate LBD for a number of distributional settings, namely when the observations are homoscedastic normal with known or unknown variance, for observations from a natural exponential family, and in a nonparametric setting where we assume only exchangeability for segments without a changepoint.