General Frequency Principle for Linear Approximation Classes
Abstract
In this paper, we study a general frequency principle under gradient descent training for the finite linear approximation classes. Under suitable assumptions, we prove that the learner amplitude relative to the target amplitude vanishes at high frequencies, uniformly throughout training. We also establish sufficient conditions for low-frequency priority, covering single-frequency targets trained separately and two frequency components trained in the same run. This general frequency principle does not imply that lower frequencies always converge faster. Instead, for two frequencies that satisfy certain comparison conditions, one component will take longer to learn than the other; and in common cases, the longer component has higher frequency. Numerical experiments with Legendre and trigonometric bases validate the theoretical results.
Related
- Linear Stability and a Turning Point Principle for Barotropic Liquid Stars with a Free Boundary in General Relativity
- Flag Varieties of General Quiver Representations: Characteristic Classes and Semi-Invariants
- Control, stability, and stabilization of some classes of linear time-delay systems
- Sharp approximation rates for shallow ReLU neuralnetworks on critical Besov classes
- The Sharp Constant in the Second-Order Hydrogen Uncertainty Principle: Spectral Characterization and Quantitative Approximation
BibTeX
@misc{wang2026general,
title = {{General Frequency Principle for Linear Approximation Classes}},
author = {Yujun Wang and Yiqi Gu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.41179},
eprint = {2610.41179},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.41179},
}
RIS
TY - UNPB
TI - General Frequency Principle for Linear Approximation Classes
AU - Yujun Wang
AU - Yiqi Gu
PY - 2026
DA - 2026/10/30
PB - arXiv
DO - 10.48550/arxiv.2610.41179
UR - https://arxiv.org/abs/2610.41179
AB - In this paper, we study a general frequency principle under gradient descent training for the finite linear approximation classes. Under suitable assumptions, we prove that the learner amplitude relative to the target amplitude vanishes at high frequencies, uniformly throughout training. We also establish sufficient conditions for low-frequency priority, covering single-frequency targets trained separately and two frequency components trained in the same run. This general frequency principle does not imply that lower frequencies always converge faster. Instead, for two frequencies that satisfy certain comparison conditions, one component will take longer to learn than the other; and in common cases, the longer component has higher frequency. Numerical experiments with Legendre and trigonometric bases validate the theoretical results.
ER -