Lattice Green's functions for Poisson problems with two unbounded and one periodic direction
math.NA, cs.NA
Submitted: 2026-11-02
Updated: 2026-11-02
Code: https://github.com/vortexlab-uclouvain/flups
License: http://creativecommons.org/licenses/by/4.0/
The gist: Fourier-based Poisson solvers treat periodic directions spectrally and unbounded ones by the Hockney--Eastwood convolution with a free-space kernel.
Abstract
Fourier-based Poisson solvers treat periodic directions spectrally and unbounded ones by the Hockney--Eastwood convolution with a free-space kernel. With two unbounded and one periodic direction, the setting of vortex tubes, wakes and jets, the kernel used so far by these solvers combines the two-dimensional kernel on the zero periodic mode with the symbol of the operator sampled on the doubled grid on every other mode. We show that the second ingredient is not exact: the error is the field of the periodic images of the source and independent of the mesh size. While the period L z is not much longer than the unbounded extent L x, it is of order (-2πd/L z), with d the distance to the nearest image; beyond, it grows like (L z/L x) squared. The error is thus small on cubic domains, but larger than the solution itself on elongated ones. We remove the approximation by extending the integral representation of the exact second-order kernel of Hou and Colonius: the screened kernel of each mode is evaluated in closed form along one unbounded direction and by Gauss--Legendre quadrature along the other, for the high-order dimension-split and Mehrstellen stencils, at a one-off cost of about one solve. The kernels, implemented in the open-source library, are verified by comparison with independent quadratures, by a padding test sensitive to the boundary condition enforced, and by convergence at the order of each stencil on an axially modulated tube source, on domains up to 64 times longer than wide and up to 2.7 times10 11 unknowns.
Sources
Related papers
- Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
- A Neural-preconditioned Poisson Solver for Mixed Dirichlet and Neumann Boundary Conditions
- Advanced Linear Algebra with Applications - Part I (Numerical linear algebra for PDEs, machine learning, and data assimilation)
- Operator learning meets inverse problems: A probabilistic perspective
- Second-order consistency for learning chaotic dynamics via randomized Jacobian matching
- Windowed thinning and query complexity for the bouncy particle and Zigzag samplers