Thanks Kristofer. It's not that repeatedly transitioning between harmonic and sample-based representations itself loses equivariance but rather that it is computationally cumbersome. It is applying non-linearities in the sample-based representation that results in a loss of equivariance. Hope that clears things up.