A Necessary Condition of Convergence for Reinforcement Learning with Function Approximation