Comment by j2kun

This may be some nuance I'm not quite getting still. You're saying that even if you use the coefficient encoding, one would still have reason to store polynomials in the CRT domain. I agree that would be more performant, but it seems not useful because then your element-wise products only ever correspond to convolutions (as I mention in the article). As far as I can tell, everyone does Double-CRT with RNS encoding, but I may be mistaken.