Multiple Lattice Rules for Multivariate $L_\infty$ Approximation in the Worst-Case Setting

TL;DR

This paper introduces a new sampling scheme using multiple lattice rules for efficient multivariate $L_$ approximation, providing near-optimal error bounds and fast algorithms in the worst-case setting.

Contribution

It develops a general framework for $L_$ approximation using rank-1 lattice sampling schemes with FFT, offering simple bounds and nearly optimal sampling rates.

Findings

Sampling numbers closely match approximation numbers.

The method achieves nearly optimal sampling rates for mixed smoothness functions.

The approach improves tractability results for $L_$ approximation.

Abstract

We develop a general framework for estimating the $L_\infty(\mathbb{T}^d)$ error for the approximation of multivariate periodic functions belonging to specific reproducing kernel Hilbert spaces (RHKS) using approximants that are trigonometric polynomials computed from sampling values. The used sampling schemes are suitable sets of rank-1 lattices that can be constructed in an extremely efficient way. Furthermore, the structure of the sampling schemes allows for fast Fourier transform (FFT) algorithms. We present and discuss one FFT algorithm and analyze the worst case $L_\infty$ error for this specific approach. Using this general result we work out very weak requirements on the RHKS that allow for a simple upper bound on the sampling numbers in terms of approximation numbers, where the approximation error is measured in the $L_\infty$ norm. Tremendous advantages of this estimate are its pre-asymptotic validity as well as its simplicity and its specification in full detail. It turns out, that approximation numbers and sampling numbers differ at most slightly. The occurring multiplicative gap does not depend on the spatial dimension $d$ and depends at most logarithmically on the number of used linear information or sampling values, respectively. Moreover, we illustrate the capability of the new sampling method with the aid of specific highly popular source spaces, which yields that the suggested algorithm is nearly optimal from different points of view. For instance, we improve tractability results for the $L_\infty$ approximation for sampling methods and we achieve almost optimal sampling rates for functions of dominating mixed smoothness. Great advantages of the suggested sampling method are the constructive methods for determining sampling sets that guarantee the shown error bounds, the simplicity and extreme efficiency of all necessary algorithms.