Towards Theoretical Understanding of Inverse Reinforcement Learning