On Learning and Covering Structured Distributions