This paper makes progress on our theoretical understanding of function approximation in RL, a crucial and tricky topic. The paper is technically strong and proposes a highly novel recursion-based algorithm that could open the door to future innovations.