Refinement-based Christoffel sampling for least squares approximation in non-orthogonal bases

Abstract

We introduce a refinement-based Christoffel sampling (RCS) algorithm for least squares approximation in the span of a given, generally non-orthogonal set of functions $\Phi_n = \{\phi_1, \dots, \phi_n\}$. A standard sampling strategy for this problem is Christoffel sampling, which achieves near-best approximations in probability using only $\mathcal{O}(n \log(n))$ samples. However, it requires i.i.d. sampling from a distribution whose density is proportional to the inverse Christoffel function $k_n$, the computation of which requires an orthonormal basis. As a result, existing approaches for non-orthogonal bases $\Phi_n$ typically rely on costly discrete orthogonalization. We propose a new iterative algorithm, inspired by recent advances in approximate leverage score sampling, that avoids this bottleneck. Crucially, while the computational cost of discrete orthogonalization grows proportionally with $\|k_n\|_{L^\infty(X)}$, the cost of our approach increases only logarithmically in $\|k_n\|_{L^\infty(X)}$. In addition, we account for finite-precision effects by considering a numerical variant of the Christoffel function, ensuring that the algorithm relies only on computable quantities. Alongside a convergence proof, we present extensive numerical experiments demonstrating the efficiency and robustness of the proposed method.