Canonical correlation theory is extended so that each variate in each of the two observed sets of variates is assigned a unique information-theoretic measure of its contribution to the total amount of information that either set provides about the other.