Estimation of arbitrary order central statistical moments by the multilevel Monte Carlo method (Q507012)

The estimation of the central statistic moments is important for many reasons. The variance is one of the most important characteristics of random variables, along with the mean. Higher-order moments, particularly the third and the fourth moments, are important in statistical applications. In the present paper the authors extend the general framework of the multiple Monte Carlo method to multilevel (MLMC) estimation of the central statistical moment of arbitrary order. It is proved that under certain assumptions the total cost of an MLMC central moments estimator is asymptotically the same as the cost of the multiple sample mean estimator and thereby is asymptotically the same as the cost of a single deterministic forward solve. The general convergence theory is applied to a class of obstacle problems with rough random obstacle profiles. In Section 2 an overview of the general MLMC framework is given. Some preliminaries as two special cases of random variables -- the real-valued random variables and the Sobolev space-valued random variables are considered. An overview of the short-hand notations for Sobolev and Bochner spaces, and their inner products and norms is presented. The notion of the central statistical moments of \(r-\)th order is defined. Also, the notions of the variance and covariance are given. In Section 3 some preliminary results on the single and multilevel estimations by sampling are presented. In Subsection 3.1 the notions of the abstract single and multilevel estimation are reminded. In Subsection 3.2 the single and multilevel Monte Carlo estimators for the central statistical moments of order \(r \geq 2\) are introduced. In Section 4 the MC estimator of the central statistical moment of arbitrary order is studied. The convergence of its bias and variance is proved. Several lemmas are given, where some preliminary results are presented. In Section 5 an application of the theory of the MLMC estimators of central moments of arbitrary order is shown. In Theorem 1 an upper bound of the \(L_2\)-norm of the difference between the central moment of order \(r\) and its multilevel estimators is obtained. In Theorem 2 a relation between the magnitude of the mean-square error and the required computational cost is established. In Section 6 the developed abstract framework is applied to a class of obstacle problems with rough random obstacles. The mathematical framework is briefly introduced. In Section 7 some numerical experiments are realized. These numerical experiments confirm the theoretical findings.