How hard is it to detect abrupt changes in the statistics of time series?