Towards a theory of good SAT representations