Error Propagation for Approximate Policy and Value Iteration