A new perspective on batched quantile estimation

We study asymptotically valid confidence intervals (CIs) for steady-state quantiles computed from nonover-lapping batches. Asymptotic validity of the CIs is established under conditions that are weaker and more easily verifiable than the usual mixing assumptions. The performance of the CIs is evaluated with a preliminary experimental study. These results form the basis for developing fully sequential procedures that yield CI estimators of steady-state quantiles with user-specified absolute or relative precision.