Abstract
The quantity of interest in the classical Cram\'er-Rao theory of unbiasedestimation (e.g., the Cram\'er-Rao lower bound, its exact attainment forexponential families, and asymptotic efficiency of maximum likelihoodestimation) is the variance, which represents the instability of an estimatorwhen its value is compared to the value for an independently-sampled data setfrom the same distribution. In this paper we are interested in a quantity whichrepresents the instability of an estimator when its value is compared to thevalue for an infinitesimal additive perturbation of the original data set; werefer to this as the "sensitivity" of an estimator. The resulting theory ofsensitivity is based on the Wasserstein geometry in the same way that theclassical theory of variance is based on the Fisher-Rao (equivalently,Hellinger) geometry, and this insight allows us to determine a collection ofresults which are analogous to the classical case: a Wasserstein-Cram\'er-Raolower bound for the sensitivity of any unbiased estimator, a characterizationof models in which there exist unbiased estimators achieving the lower boundexactly, and some concrete results that show that the Wasserstein projectionestimator achieves the lower bound asymptotically. We use these results totreat many statistical examples, sometimes revealing new optimality propertiesfor existing estimators and other times revealing entirely new estimators.