Low-rank approximation of integral operators by using the Green formula and quadrature (Q383524): Difference between revisions

The following Fredholm integral operator is considered: \[ \mathcal{G}[u](x)=\int_{\Omega}g(x,y)u(y)dy, \quad \Omega\subset\mathbb{R}^d. \] Discretising \(\mathcal{G}\) by a standard Galerkin approach for a basis \((\varphi_i)_{i\in I}\) yields a dense matrix \(G\) with entries \[ G_{ij}:=\int_{\Omega}\int_{\Omega}\varphi_i(x)g(x,y)\varphi_j(y) dx dy. \] To avoid the quadratic complexity it takes to compute and store a dense matrix, several approaches have been introduced including \(\mathcal{H}\)-matrices. The kernel function is approximated by a separable function, this leads to a low-rank matrix. Interpolation is a robust and popular scheme, but requires us to interpolate in each spatial dimension, which leads to a complexity of \(m^d\) for \(m\)-th order. Instead of interpolation the authors propose using quadrature on the kernel function represented with Green's formula. Due to the fact that we are integrating only over the boundary, we save one spatial dimension compared to the interpolation method and get a complexity of \(m^{d-1}\). In conclusion there is a large number of numerical experiments in paper. It allows to compare this approach to other analytical techniques.