Improving Self-Supervised Learning by Characterizing Idealized Representations