A theory of weight distribution-constrained learning