Abstract: Finding the best discretization of a scene is a difficult problem in radiosity. In this paper we study this problem from the point of view of information theory. In previous work, we have introduced the concept of mutual information as a measure for scene complexity. A study of the effect of some commonly used patch subdivision strategies on the mutual information shows that the strategy yielding the largest decrease of mean square error corresponds well to the one resulting in the largest increase in mutual information. This observation suggests a possible refinement oracle for radiosity computations.