“The estimator `(r - 1) - log(r)` for KL divergence is a form of Bregman divergence, representing the difference between a convex function and its tangent plane.”