Image synthesis can be described as integrating the light flux across the image plane. This is characterized by the interaction between light, objects, and their reflectance. These phenomenons are very complex and can only be evaluated through numerical integration (Monte Carlo, ...). Real time rendering poses significant constraints on the sampling pattern (usually one sample per pixel), so we are exposed to under sampling artifacts. Furthermode, the rasterization approach used by GPUs provides overall complexity by the number of geometrical objects.
We want to give representations of parts of the scene that can efficiently describe light interaction on objects at multiple scales, to get accurate rendering independently of the underlying geometrical complexity.