High-dimensional adaptive sparse polynomial interpolation and applications to parametric PDEs (Q404259): Difference between revisions

Motivated by the fast computation of solutions to parametric partial differential equations (PDEs), the authors study the Lagrange polynomial interpolation in high dimension \(d\). The considered polynomial spaces \({\mathbb P}_{\Lambda} ={\mathrm {span}}\,\{{\mathbf y}^{\nu};\,\nu \in \Lambda\}\) with index sets \(\Lambda\) can be sparse and anisotropic with respect to the parameter \(y=(y_1,\dots,y_d)\). The objective is to study a collocation method based on a high-dimensional interpolation process that can be coupled with an adaptive selection of the polynomial spaces \({\mathbb P}_{\Lambda}\). First, the authors build an interpolation operator \(I_{\Lambda}\) and its associated grid. For a monotone index set \(\Lambda\), \(I_{\Lambda}\) is unisolvent for the polynomial space \({\mathbb P}_{\Lambda}\). The construction of \(I_{\Lambda}\) is based on a univariate sequence of points (such as a Leja sequence) and on standard tensorisation of a 1-dimensional interpolation scheme and sparsification. Upper bounds on the Lebesgue constant for this interpolation process are given. The authors discuss an adaptive, greedy selection of a nested sequence of monotone index sets \(\Lambda_n\). For an elliptic model problem, the authors show an algebraic convergence rate for the interpolation algorithm. The performance of this method is illustrated by the interpolation of functions \(u:\,[-1,1]^d \to {\mathbb R}\) for \(d\in \{8,\,16,\,32,\, 64\}\) and by the solution of parametric elliptic PDEs.