I don't want to play the smart ass here, but I would want to give some precisions/limitations about the technique you describe here.
What you explain is true if your rendering is linear meaning without engine optimization like the architectural sampler optimization (like explained on the nivida
You must be registered to see the links
:
"Also don’t be disappointed if the rendering after some dozen frames still looks more noisy than the “classic” version, most of the time the architectural sampler will catch up later-on.") that optimization occures later during the render (start between 30% and 60% of convergency) the same apply with the blooming effect that attenuates itself for the part that are less affected by the effect along the iterations.
Also depending the size you want to upsample the image the benefit could be tamed down by the increase in rendering time the increase of target pixels will cause. That point is more obvious if you render on CPU only like me...
Sorry to spoil your post, it is not my intention... I like the fact that you are the first to try to help newbies...