When does third order efficiency imply fourth order efficiency

Abstract

In this article third and fourth order efficiency are studied in the framework of translation equivariant location estimators. We assume $X_1,...,X_n$ i.i.d. $f(\cdot -\theta)$. By recognizing that equality in a special form of the Cauchy-Schwarz inequality leads to a certain dependence of the cumulants of the maximum likelihood estimator (MLE) for $\theta$, it is shown that this MLE is fourth order efficient if the underlying distribution is Gumbel. Contrary to similar results which were previously published this result is not based on symmetry.