Data Augmentation: A Fourier Analysis Perspective

Data augmentation is a standard technique for improving machine learning models by exploiting known symmetries. By transforming training samples according to a group of symmetries, practitioners encourage models to learn invariant features. However, when the symmetry group is large or infinite, performing full augmentation—averaging over every possible transformation—becomes computationally prohibitive. This paper investigates whether partial data augmentation, which uses only a small, randomly sampled subset of the group, can match the statistical performance of full augmentation.

The authors utilize Fourier analysis and representation theory to analyze projection-based estimators in density estimation and regression. They demonstrate that partial augmentation is surprisingly effective. Specifically, the expected excess error of a model trained with a random subset of group elements $S$ scales with the dimension of the invariant subspace ($r_{inv}$) and an additional term that vanishes as the size of $S$ increases. Crucially, the number of samples required from the group depends only on the dimension of the invariant subspace, not on the total size of the group itself. This provides a theoretical justification for why partial augmentation is often sufficient in practice to capture the benefits of symmetry.

Beyond individual tasks, the authors show that a single randomly chosen augmentation set can be reused across different learning tasks with only a minor logarithmic penalty in performance. This suggests that partial augmentation is a robust and scalable strategy. However, the paper also presents a fundamental limitation: while partial augmentation can achieve optimal statistical rates, it cannot enforce exact invariance. The authors prove that achieving exact symmetry requires averaging over the entire group; any strict subset will fail to eliminate all non-invariant components if the hypothesis space is sufficiently rich.