Gradient descent with identity initialization efficiently learns positive definite linear transformations.

Peter L. Bartlett, David P. Helmbold, Philip M. Long

Browse the full ICML paper archive.