On the Number of Samples Needed to Learn the Correct Structure of a Bayesian Network