The approach would be just exhaustive summation, I suppose. Would symbolic integration be equally expansive for simple enough geometry? Simplified approximations to the geometry, i.e. bounding boxes, would probably not look real enough. I get, there are all kinds of approaches.
I heard Monte Carlo method used to limit the search space to a random subset.
A lot of work has been done on using analytic integration techniques, as well as higher quality numeric approximations, for global illumination, in particular with radiosity methods. https://en.m.wikipedia.org/wiki/Radiosity_(computer_graphics...
http://www.mi.uni-koeln.de/c/mirror/www.cs.curtin.edu.au/uni...
Imperfect Ray-Bundle Tracing for Interactive Multi-Bounce Global Illumination