random-projection

This paper shows that the Johnson–Lindenstrauss lemma can be uninformative about retained geometry in high dimensions and derives exact limits for features like distance recovery, nearest-neighbor rankings, and covariance shape in Gaussian models.