I saw this recently, from: http://www.fxguide.com/featured/the-state-of-rendering-part-2/#arnold
Would this be possible to implement in cycles?
“If you look at the ‘random’ samples below on the left you see a seemingly sensible distribution across a square patch which represents an area light. The problem is when an area light is seen from the computer’s point of view via the ‘soild angle’ maths that is the area light as ‘projected’ along sight line with the computational ‘dome’ used in computer graphics, it is now easy to see just how much the samples are all collecting along the edges. So much stronger is this bias or effect than one might imagine - it is worth checking any square for yourself (count in from the left and bottom and you can see that both of the shapes in (A) marked Area sampling are exactly the same. What is needed is to start from a distribution. If one starts with the scattering on the left in (b) Spherical Rectangle sampling, when the ‘projection’ is taken into account the sample is now more evenly spread. This directed sampling is just a refinement that falls under improved Importance Sampling.”