Bounds are given for the empirical and expected Rademacher complexity of classes of linear transformations from a Hilbert space H to a ...nite dimensional space. The results imply generalization guarantees for graph regularization and multi-task subspace learning.