When can unlabeled data improve the learning rate?