Montecarlo algorithms used in final gather and path tracing produce a lot of variation in difficult-to-reach dark areas and "clustering" of radiance results instead of spreading it out in a regular pattern. Clustering is a known issue of true stochastic algorithms and it is an issue latter magnified by adaptive sampling based on a color thresold, which could be useful for brighter areas and gradients but does a poor job sampling very dark clusters that integrate little or no indirect lighting in the first passes.
In the image below, you can see this effect on the bottom of the red wall, with dark patches that get consistently ignored by adaptive sampling even with a thresold as low as 0.0008. A differential thresold based on image darkness could be a solution though I am not sure whether a color thresold system is really robust enough to take on a problem like this, since a thresold on dark areas should probably stay infinitesimal or be 0 while other areas in a normal render could use a thresold as high as 0.02. I mean the difference seems to describe a geometrical progression problem as many other stuff in montecarlo engines and this problem could be also happening in less dark areas without us noticing. Maybe there are better solutions based on samples divergence for instace or there is some paper that specifically takes on this issue. The new feature "Dark areas noise detection" does not make much of a difference at the moment regarding this issue.
You can download the file at