The prediction that equivariance reduces sample complexity by a factor of |G| appears in roughly every paper on geometric deep learning and is measured as an actual scaling law in roughly none of them. This paper does the measurement. The methodology is the interesting part. Naive estimators conflat
