We concentrated on full-rank Gaussians because these give dramatically better