Fast convolution solver based on far-field smooth approximation

Abstract: The convolution potential arises in a wide variety of application areas, and its efficient and accurate evaluation encounters three challenges: singularity, nonlocality and anisotropy. We introduce a fast algorithm based on a far-field smooth approximation of the kernel, where the bounded domain Fourier transform, one of the most essential difficulties, is well approximated by the whole space Fourier transform which usually admits explicit formula. The convolution is split into a regular and singular integral, and they are well resolved by trapezoidal rule and Fourier spectral method respectively. The scheme is simplified to a discrete convolution and is implemented efficiently with Fast Fourier Transform (FFT). Importantly, the tensor generation procedure is quite simple, highly efficient and independent of the anisotropy strength. It is easy to implement and achieves spectral accuracy with nearly optimal efficiency and minimum memory requirement. Rigorous error estimates and extensive numerical investigations, together with a comprehensive comparison, showcase its superiorities for different kernels.