High-Dimensional, Multiscale Online Changepoint Detection

Abstract We introduce a new method for high-dimensional, online changepoint detection in settings where p-variate Gaussian data stream may undergo change mean. The procedure works by performing likelihood ratio tests against simple alternatives of different scales each coordinate, and then aggregating test statistics across coordinates. algorithm is the sense that both its storage requirements worst-case computational complexity per observation are independent number previous observations; practice, it even be significantly faster than this. prove patience, or average run length under null, our at least desired nominal level, provide guarantees on response delay alternative depend sparsity vector mean change. Simulations confirm practical effectiveness proposal, which implemented R package ocd, we also demonstrate utility seismology set.