Originally posted by deejjjaaaa
What do you mean?
I only see that Gordon confirms Oleg's math. You could have asked me and I would have been able to provide the same confirmation. And DxO did run the same math as well.
One problem is that NR on raw data isn't necessarily an evil thing.
(E.g., binning the green channel is a very good idea if you think about how demosaicing works. There is a threshold in noise where binning is better than not binning. You could leave this decision to the demosaicing algorithm but maybe, the camera-internal JPEG engine just needs it done before.)
Another problem is, and nobody at DPR or DxO actually touched at this, is that noise can be correlated for other reasons than NR. Banding is a typical case. On a smaller scale, a gain amplifier serving multiple photo sites may have slightly more noise than a neighboring gain amplifier. This results in finite fall-off of the auto-correlation function. Despite the fact that no NR was done. DxO would then (mistakingly) reduce the signal-to-noise ratio.
All in all, Oleg may master the math better than anybody else at DPR. But this is not enough.