Information theoretic limits of learning a sparse rule