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

Open in new window