Random cubatures and quasi-Monte Carlo methods (Q500376)

The conceptual similarity between stratified Monte Carlo and quasi-Monte Carlo integration has been stated and analyzed by many authors. In the present paper, a formal connection between these two methods using the random cubature theory presented in terms of the generalized Haar system is established. The Fourier-based variance analysis of the derived stratified formula is arrived. Several equivalent expressions for the variance within the standard quasi-Monte Carlo setting is proposed. The theory of random cubature methods is supplemented with two refined versions of known results and completely new facts. In the introduction, a formal application of mathematical statistics methods to deterministic objects is considered. The problem to estimate the variance \[ \tilde{E}\left({1 \over N} \sum_{j=1}^{N} f({\mathbf X}_j)\right)^2 - \left( \tilde{E}\left({1 \over N} \sum_{j=1}^{N} f({\mathbf X}_j)\right)\right)^2, \] where \({\mathbf X}_j\) are quasi-random vectors and \(\tilde{E}\) is a formal mathematical expectation, is considered. In Section 2, an arbitrary space \(({\mathcal X}, {\mathcal B},\mu)\) is considered, where \({\mathcal X}\) is a non-empty set, \({\mathcal B}\) is a \(\sigma\)-algebra for subsets of \({\mathcal X}\) with measure \(\mu\). For a \(\mu\)-integrable function \(f: {\mathcal X} \to R\), the problem of numerical calculation of the integral \(\displaystyle I = \int_{\mathcal X} f({\mathbf X}) \mu(dx)\) is considered. For an arbitrary system of functions and a set of points of \({\mathcal X}\), the determinant \(\Delta(Q)\) of the alternant matrix, the determinant \(\Delta(f;Q)\) and the quantity \(\displaystyle S_N = {\Delta(f;Q) \over \Delta(Q)}\) are introduced. Two previous results of the second author are presented. The orthonormal generalized Haar system based on the space \({\mathcal X}\) which here is the \(s\)-dimensional unit hypercube is introduced. In Theorem 2.3 is provided that an \(N\)-point quadrature formula \(\displaystyle S_N = {1 \over N} \sum_{j=1}^N f({\mathbf X}_j)\) with a joint density of this nodes is a unibiased estimate of the integral \(I\) and this formula is exact for the generalized Haar system. In Theorem 2.4, an exact formula for the variance \(\mathrm{Var}(S_N)\) is obtained. The concept of the generalized indicator system is given. The form of the transformation matrix from the generalized Haar system and the generalized indicator system is shown. In Theorem 2.6, it is proved that the cubature formula \(S_N\) is exact for both the generalized Haar system and the generalized indicator system. The variance of this formula is expressed in explicit form. In Section 3, the existing theory of random cubature formulas by observing the relation between an arbitrary non-regular function system and a variance reduction is considered. Theorem 3.1 is a generalization of Theorems 2.2 and 2.4. Here an arbitrary orthonormal system with sliding support is used. Theorem 3.2 is a refined version of Theorem 2.2. The reduction of the variance is shown. In Section 4, the conceptual similarity between stratified Monte Carlo and quasi-Monte Carlo methods are discussed.