On the connection between Bregman divergence and value in regularized Markov decision processes