For symmetric auctions, there is a close relationship between distributions of order statistics of bidders' valuations and observable bids that is often used to estimate or bound the valuation distribution, optimal reserve price, and other quantities of interest nonparametrically. However, we show that the functional mapping from distributions of order statistics to their parent distribution is, in general, not Lipschitz continuous and, therefore, introduces an irregularity into the estimation problem. More specifically, we derive the optimal rate for nonparametric point estimation of, and bounds for, the private value distribution, which is typically substantially slower than the regular root-n rate. We propose trimming rules for the nonparametric estimator that achieve that rate and derive the asymptotic distribution for a regularized estimator. We then demonstrate that policy parameters that depend on the valuation distribution, including optimal reserve price and expected revenue, are irregularly identified when bidding data are incomplete. We also give rates for nonparametric estimation of descending bid auctions and strategic equivalents.Keywords. Empirical auctions, order statistics, bounds, irregular identification, uniform consistency. JEL classification. C13, C14, D44.The order statistics approach has been very fruitful for deriving nonparametric identification results and bounds for auction models. 1 However, as we show in this paper, the central step of "inverting out" the distribution of bidders' valuations from the distribution of an order statistic introduces an irregularity into the estimation problem. We show that point estimation of, or construction of bounds for, the cumulative distribution function (c.d.f.) of valuations from order statistics is generally at a rate slower than