Near-Optimal Representation Leanring for Hierarchical Reinforcement Learning