Quasi–Monte Carlo integration over \mathbb {R}^d

From MaRDI portal
Publication:4452163





The authors consider a certain class of integrals \[ \text{Int}_{\rho}(f):= \int f({\mathbf x}) \rho({\mathbf x}) \, d{\mathbf x} \] over \(\mathbb{R}^d\) with a probability weight function \(\rho\). A quasi-Monte Carlo integration algorithm is developed based on a decomposition of the domain \(\mathbb{R}^d\) and arranging low discrepancy points over hierarchial hypercubes \(Q_j\) with \(\mid x_l \mid<2^j\), \(l=1, \ldots, d\), \(j=0, 1, \ldots, m\). In the actual method, the integral is replaced by the sum \[ S_{m,{\mathbf n}}(f,\rho):=\sum\limits_{j=0}^{m} \frac{q_j}{n_j} \sum\limits_{i=1}^{n_j} f({\mathbf y}_{ij}) \rho({\mathbf y}_{ij}) \chi_{I_j}({\mathbf y}_{ij}) \;, \] where \(m+1\) is any number of cubes, \(q_j = \text{Volume}(Q_j)=2^{(j+1)d}\), and \({\mathbf n}:= \left( n_0, \ldots, n_m \right)\), where each \(n_j\) is the number of points used within cube \(Q_j\). Here \(\chi_{I_j}\) is the indicator function of the frame \(I_j\) in the decomposition \(\mathbb{R}^d = \bigcup_{j=0}^{m+1} I_j\), where \(I_j := Q_j \backslash Q_{j-1}\), \(j=1, \ldots, m\), \(I_0=Q_0\), \(I_{m+1}=Q_m^c\). The point set \(\{ {\mathbf y}_{ij} \}\) is obtained by an appropriate scaling and shifting (to cover the cube \(Q_j\)) a low-discrepancy point set \(P_j = \{ {\mathbf x}_{ij}: 1 \leq i \leq n_j \} \subset [0,1)^d\). The total number of function evaluations is \(N= \sum_{j=0}^m n_j\). It is proven that for certain class of integrable functions \(f \in \mathcal F\) with bounded variation and power/exponential decaying weights \(\rho \in {\mathcal M}_s(R)\), such that all weak derivatives \(\frac{\partial^{| I| }}{\partial x_I} \rho({\mathbf x})\) with \(I \subset \{1, \ldots, d \}\) exist and \[ \sup\limits_{\| {\mathbf x}\| >t} \left| \frac{\partial^{| I| }} {\partial x_I} \rho({\mathbf x}) \right| \leq R t^{-s} \] holds, the algorithm has an optimal error bound order \(N^{-1} \log^d N\). In particular, \({\mathcal M}_s(R)\) covers ellipticilly contoured distributions considered in some detail. Example calculations in two dimensions (\(d=2\)) are made by using the Fibonacci sequence to generate the low-discrepancy point sets. Finally, as an application to mathematical finance, evaluation of coherent risks is discussed.




Cited in
(21)








This page was built for publication: Quasi–Monte Carlo integration over $\mathbb {R}^d$

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4452163)