General Frequency Principle for Linear Approximation Classes
Organizations: School of Mathematical Sciences, University of Electronic Science and Technology of China, Chengdu, 611731, China.
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.
Figures & tables
| Work | Setting and conditions | Quantities and conclusions | Main tools |
|---|---|---|---|
| Xu et al. Xu et al. (2020) | Shallow tanh networks; small input weights; two-frequency loss for rate comparisons | Low-frequency dominance in parameter gradients and instantaneous loss decay under the stated conditions | Explicit Fourier transforms; small-weight estimates |
| Rahaman et al. Rahaman et al. (2019) | ReLU networks with piecewise linear structure | Fourier decay of network outputs and parameter derivatives; implications for high-frequency changes | Fourier analysis; geometry of linear regions |
| Luo et al. Luo et al. (2021) | Population risk; regularity and bounded trajectories; nondegenerate minimizer for final-stage estimates | Bounds on high-frequency contributions relative to total loss or output changes at different training stages | Sobolev regularity; Fourier decay; gradient flow |
| Cao et al. Cao et al. (2021) | Two-layer ReLU networks; width and sample-size conditions; uniform spherical data for frequency analysis | Residual convergence along NTK eigenspaces; frequency-dependent rates through spherical harmonics | NTK spectra; probability estimates; spherical harmonics |
| Luo et al. Luo et al. (2022) | Infinite-width two-layer NTK dynamics; initialization conditions for explicit formulas | Fourier-domain residual dynamics and frequency dependence of the evolution operator | NTK dynamics; Fourier transforms; operator calculations |
| Zhang et al. Zhang et al. (2025) | Finite-width shallow ReLU models; nonlinear bounds use an interval model with fixed input weights, a cutoff, and initialization and weight conditions | Gram spectra and finite-precision limits; time lower bounds for generalized Fourier coefficients of an auxiliary function | Gram spectra; differential operators; gradient flow |
| Basis | Case | (b) | (c) | ||||
|---|---|---|---|---|---|---|---|
| Legendre | A | Yes | Yes | 66.511 | 2166.975 | 44 | 3090 |
| B | Yes | Yes | 560.567 | 4607.374 | 369 | 6522 | |
| Weighted trig. | A | Yes | Yes | 89.091 | 577.139 | 52 | 654 |
| B | Yes | Yes | 216.035 | 1130.842 | 131 | 1240 |
| Basis | Case | ( 4.33 ) | ( 4.34 ) | |||
|---|---|---|---|---|---|---|
| Legendre | A | 0.077622 | 0.750207 | 6 | 9 | 111 |
| B | 0.098914 | 0.518867 | 29 | 39 | 181 | |
| Weighted trig. | A | 0.096216 | 0.278414 | 8 | 21 | 47 |
| B | 0.099132 | 0.361444 | 17 | 27 | 75 |
Frequency-Domain Feature Learning
Frequency-domain feature learning uses Fourier, wavelet, and related spectral representations to extract task-relevant patterns from data.
See all 201 papers