Hello Okan and George!
I have updated the "Ambient Occlusion with Environment Map" paragraph on ProblemRIBs page.
The black noise vanishing but Pixie render the environment mapping with a bad conical wrapping.
I've been thinking about this a little. Basically, the environment is cylendrical (this is per RiSpec), approximating a spherical mapping. However, there's always going to be an issue of exactly what to do at the poles of the map.
Now we look up in an area centred about the sample direction. Near the edges of the map (particularly in v), we may actually go fractionally off the edge of the map - which is better than wrapping back to the bottom and ending up sampling the whole map at the blurriest level (which is what you'll do if you have a samples which cover the whole map).
That's not per-s the root cause of what you see here - though it does mean you shouldn't set the vmode to periodic when preparing the texture.
What is really happening is that when you get very close to the pole, the samples should incorporate points in the map which are 180-degrees away in u - ie they're precisely at a distance of 0.5 away in u. These are the points which are theoretically adjacent to the target sample point. But in the map, they're some distance away.
Normally a map is symetrical in that sense - the left and right edges near the top are similar to those at the center of the top edge, and so-on. This often means there's a deal of distortion in the top of the map - it's blurry. Such is the mapping that's being used.
Perhaps PRman is doing something very smart and sampling at various points in the map. That's what your PRMan render would seem to suggest, if it's the same map, same shaders etc.
I'll have a think about code to support a smarter treatment here, though I'm undecided on whether it normally should be needed - normally, like I said the map is distorted with no discontinuities at the two poles. There's a discontinutity in yours between the green and the blue on the u-edges. Perhaps that's valid though - I just don't know.