We describe a novel representation of the light field tailored to improve importance sampling for Monte Carlo rendering. The domain of the light field i.e., the product space of spatial positions and directions is hierarchically subdivided into subsets on which local models characterize the light transport. The data structure, that is based on double trees, approximates the exact light field and enables very efficient queries for importance sampling and easy tracing of photons in the scene. The framework is simple yet flexible, enabling the usage of any type of local model for representing the light field, provided it can be efficiently importance sampled. The method also supports progressive refinement with an arbitrary number of photons. We provide a reference open source implementation.