Error bounds for quasi-Monte Carlo integration with uniform point sets (Q1863289)

The paper deals with evaluation of error bounds for numerical quasi-Monte Carlo integration of a Lebesgue-integrable function \(f\) over the \(s\)-dimensional unit cube. In this method the integral is represented by a sum over a set of deterministic points. The known error bounds are given in terms of the variance of function \(f\) and the \textit{star discrepancy} \(D^*_N(\mathcal{P})\) (or degree of non-uniformity) of the given point set \(\mathcal{P}= \{ {\mathbf x}_1, \ldots ,{\mathbf x}_N \}\). Alternatively, they are given by the \textit{modulus of continuity} of the integrand, which depends on \(f\) and \(D^*_N(\mathcal{P})\). New error bounds are established for node sets with a special kind of uniformity property such that the number of points \(A(M;\mathcal{P})\) of \(\mathcal{P}\) falling into each of subsets \(M\), corresponding to a certain partition \(\mathcal{M} = \{ M_1, \ldots ,M_k \}\) of the set \(X\), is given by \(A(M;\mathcal{P})= \mu(M)N\), where \(\mu(M)\) is the probability measure of subset \(M\). The basic proofs of the new error bounds for such \((\mathcal{M},\mu)\)-uniform point sets are formulated in abstract setting of arbitrary probability spaces \((X,\mathcal{B},\mu)\), where \(X\) is an arbitrary nonempty set, \(\mathcal{B}\) is a \(\sigma\)-algebra of subsets of \(X\), and \(\mu\) is a probability measure defined on \(\mathcal{B}\). Only the bounds in terms of the modulus of continuity require also the structure of a metric space. The error bounds are given by different quantities as, e.~g., \(S_{\mathcal H}(f) := \max\limits_{1 \leq j \leq k} \left( \sup\limits_{{\mathbf t} \in H_j} f({\mathbf t}) - \inf\limits_{{\mathbf t} \in H_j} f({\mathbf t}) \right)\), where \(\mathcal{H} = \{ H_1, \ldots , H_k \}\) is a finite collection of nonempty subsets of \(X\) with \(\cup_{j=1}^k H_j = X\), instead of the star discrepancy. Examples of \((\mathcal{M},\mu)\)-uniform point sets are considered which are relevant to applications of quasi-Monte Carlo integration. They show that the new error bounds often are significantly better than the known ones.