Efficient convolution with the Newton potential in \(d\) dimensions (Q957929): Difference between revisions

Let the mapping of \(f\in L^{2}({\mathbb R}^d)\) onto \(u\in L^{2}({\mathbb R}^d)\) by means of \[ u(x):=\left( Kf\right)(x):=\int_{{\mathbb R}^{d}}\frac{1}{\left\| x-y\right\|}f(y)dy\,\, \text{for }x\in{\mathbb R}^{d}. \tag{1} \] The integral represents the convolution of the Newton (or Coulomb) potential \(\frac{1}{\left\|\cdot\right\|}\) with \(f\). The paper is concerned with the evaluation of the convolution integral, when \(f\) is given as piecewise polynomial of possible large degree, i.e., \(f\) may be considered as an \(hp\)-finite element function. The author describes the computation of integrals of the form \[ \int_{{\mathbb R}^d}\int_{{\mathbb R}^{d}}\frac{g(x)f(x)}{\left\| x-y\right\|}dxdy, \] which reduces to the scalar product \(<g,u>_{{L^{2}}({\mathbb R}^d)}\) with \(u\) from (1). He discusses the tensor product approximations. If \(f\) is given in tensor product form, the \(d\)-dimensional convolution can be reduced to one-dimensional convolutions. Although the details are given for the kernel \(1/\left\| x\right\|\), the basis techniques can be generalized to homogeneous kernels, e.g., the fundamental solution \(const\cdot\left\| x\right\|^{2-d}\) of the \(d\)-dimensional Poisson equation.