Reinforcement Learning in Non-Markov Decision Processes: Statistical Properties of Characteristic Eligibility