Summary
This paper investigates a divide-and-conquer algorithm for estimating the extreme value index when data are stored in multiple machines. The oracle property of such an algorithm based on extreme value methods is not guaranteed by the general theory of distributed inference. We propose a distributed Hill estimator and establish its asymptotic theories. We consider various cases where the number of observations involved in each machine can be either homogeneous or heterogeneous, and either fixed or varying according to the total sample size. In each case, we provide sufficient, sometimes also necessary, condition, under which the oracle property holds. Some key words: Extreme value index, Distributed inference, Distributed Hill estimator