Randomly shifted lattice rules for unbounded integrands (Q855892): Difference between revisions

The authors study the problem of multivariate integration over \({\mathbb R}^{d}\) with an integrand of the form \(f({\mathbf x})\rho_{d}({\mathbf x}),\) where \(\rho_{d}\) is a \(d\)-variate probability density function. The function \(f\) belongs to the weighted tensor product reproducing kernel Hilbert space of functions which are only once differentiable in each variable and whose mixed first derivatives, multiplied with a weight function \(\psi_{d}\) are bounded in the \(L_{p}\)-norm. There are transforming formulae that the integrals over \({\mathbb R}^{d}\) are reduced to integrals over \([0,1)^{d}.\) In this paper more general probability density functions than Gaussian ones are considered. In section 2 the one-dimensional functional space \(H_{1}(D)\) and the corresponding isometric space \({\mathcal H}_{1}([0,1])\) are introduced. The main details of these spaces as the inner product, the norm and the reproducing kernel are given in an explicit form. The spaces of \(d\)-variate functions \(H_{d}(D^{d})\) and \({\mathcal H}_{d}([0,1)^{d})\) are a weighted tensor product of \(d\)-copies of the spaces \(H_{1}(D)\) and \({\mathcal H}_{1}([0,1)).\) Shift invariant kernels, associated with the reproducing kernels of the spaces \(H_{d}(D^{d})\) and \({\mathcal H}_{d}([0,1]^{d})\) are presented. Section 3 analyses the worst case error of the integration in the space \({\mathcal H}_{d}([0,1]^{d}).\) For an approximation of the integrals over \([0,1]^{d}\) it is proposed to use a quasi-Monte Carlo (QMC) method with deterministically chosen point nets in \([0,1]^{d}.\) The worst case error \(e_{n,d}\) and the QMC mean worst case error \(M_{n,d}\) are given. A concrete choice of the point nets depending on a generating vector and a shift vector is given. The worst case error for randomly shifted rank-1 lattice rules \(e_{n,d}^{\text{ran-sh-lat}}\) is defined. In Algorithm 1 the component-by-component algorithm to construct the generating vector is given. In Theorem 2 the error \(e_{n,d}^{\text{ran-sh-lat}}\) based on the generating vector which is constructed in Algorithm 1, is evaluated from above with the QMC mean worst case error \(M_{n,d}.\) In section 4 the performance of the randomly shifted lattice rules is tested for solving the problem of the Asian call option. Three choices of the weighted function \(\psi\) -- a Gaussian function, a two-tailed exponential function and a logistic function are realized. An order \({\mathcal O}(n^{-{1 \over 2}})\) of the worst case error using randomly shifted lattice rules is achieved. The constant in this order is independent of the dimension \(d.\)