Tag
The paper proves a general frequency principle for finite linear approximation classes under gradient descent, showing high-frequency amplitudes vanish uniformly during training and establishing sufficient conditions for low-frequency priority, validated numerically with Legendre and trigonometric bases.