Exact cubature for a class of functions of maximum effective dimension (Q855894)

A functional class with maximum effective dimension is considered. The notion of effective dimensions is developed for the purpose of explaining why the quasi-Monte Carlo (QMC) method beats the Monte Carlo method by a wide margin for some high-dimensional integration problems in finance. In section 2 the ANOVA (analysis of variance) decomposition of a function of \(d\) variables is reminded. The terms of this decomposition are defined in a recursive way. The variance of a function of \(d\) variables is given. Two ways of the definition of the effective dimension in a truncation sense and in a superposition sense are reminded by the authors. The notion of effective dimension aims at measuring how important each subset of variables on the function is. In section 3 the functional class \({\mathcal F}_{d}\) is introduced. The elements of the ANOVA decomposition of the functions of this class are shown. Theorem 1 confirms that the effective dimension of any non-constant function \(f \in {\mathcal F}_{d}\), either in the truncation or in the superposition sense, is equal to \(d.\) It is shown that although the effective dimension of all functions in the class \({\mathcal F}_{d}\) is \(d,\) they can be integrated exactly by a cubature with two points. Theorem 2 gives that there exists an exact cubature with two points for the integration of any functions of the class \({\mathcal F}_{d}.\) The two points are in the form \((x_{1}, x_{2}, \ldots , x_{d}) \in [0,1]^{d}\) and \((1 - x_{1}, x_{2}, \ldots , x_{d}) \in [0,1]^{d}\) with deterministic point \((x_{1}, x_{2}, \ldots , x_{d}) \in [0,1]^{d}.\) In this way it is shown that both the convergence of QMC does not depend on the effective dimension and its worst case error can be zero for the functions of maximum effective dimension. The paper ends with a discussion. The problem for a restriction of the class \({\mathcal F}_{d}\) caused on the zero algorithm which is optimal is considered.