Alias Geometry and Density Compensation in Perturbed Fourier Sampling
math.NA, cs.NA
Submitted: 2026-11-02
Updated: 2026-11-02
Comments: 30 pages, 4 figures. Code reproducing all results: https://github.com/circlespie/alias-geometry-fourier-sampling
Code: https://github.com/circlespie/alias-geometry-fourier-sampling
License: http://creativecommons.org/licenses/by/4.0/
The gist: On a uniform grid with equal weights, Fourier modes that remain distinct on the grid are exactly orthogonal.
Abstract
On a uniform grid with equal weights, Fourier modes that remain distinct on the grid are exactly orthogonal. When the nodes are perturbed, density compensation readjusts the weights. We ask how much of the resulting error reweighting can remove, measuring it by the Frobenius norm of the Gram defect, whose square is proportional to the mean-square error of density-compensated adjoint reconstruction. The answer is governed by aliasing. The grid cannot distinguish certain frequency differences, so a weight correction must treat each such alias class as a single number, while a node displacement affects every frequency difference in the class separately. To first order, reweighting removes the class average but not the spread. The irreducible error is a sum of covariance matrices, one for each class of colliding frequency differences. If no differences collide, every small node perturbation can be repaired exactly with positive weights. With collisions, a repair that works to first order can fail at second order. At the two-dimensional DFT, row and column shears are each exactly repairable, but their combination is not, and we compute the obstruction. In one dimension, oversampling protects a band of low-frequency node motions, and under random jitter optimal reweighting asymptotically removes a fraction 1-κ 2(2-κ) squared of the expected squared first-order error at oversampling ratio κ in[1,2]. At critical sampling, node and weight perturbations are orthogonal: no reweighting helps, Voronoi weights can never reduce the error, and for the one-dimensional DFT the remaining error is a log-Vandermonde energy with the circular H 1/2 energy as its continuum limit.
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