A nonuniform fast Fourier transform based on low rank approximation

TL;DR

This paper introduces a fast, stable algorithm for the nonuniform discrete Fourier transform (NUDFT) based on low-rank approximation, significantly improving computational efficiency and adaptability over existing methods.

Contribution

It presents a novel low-rank approximation approach to compute NUDFT efficiently, with a simple implementation and automatic precision adaptation.

Findings

Achieves $ ext{O}(N ext{log} N ext{log}(1/\epsilon)/\text{log}\text{log}(1/\epsilon))$ complexity

Performs comparably or better than state-of-the-art algorithms

Special case reduces to the FFT for uniform data

Abstract

By viewing the nonuniform discrete Fourier transform (NUDFT) as a perturbed version of a uniform discrete Fourier transform, we propose a fast, stable, and simple algorithm for computing the NUDFT that costs $\mathcal{O}(N\log N\log(1/\epsilon)/\log\!\log(1/\epsilon))$ operations based on the fast Fourier transform, where $N$ is the size of the transform and $0<\epsilon <1$ is a working precision. Our key observation is that a NUDFT and DFT matrix divided entry-by-entry is often well-approximated by a low rank matrix, allowing us to express a NUDFT matrix as a sum of diagonally-scaled DFT matrices. Our algorithm is simple to implement, automatically adapts to any working precision, and is competitive with state-of-the-art algorithms. In the fully uniform case, our algorithm is essentially the FFT. We also describe quasi-optimal algorithms for the inverse NUDFT and two-dimensional NUDFTs.