Bernard you are in danger of sounding like one of the MFDB pundits you have so often bemoaned for claiming their images have a 'quality' not found in lesser pixel devices.... Just sayin... 
I agree with Josh, but I understand Benhard's point of having subject surface/material structure becoming almost tangible. One can e.g. see the difference between cardboard and leather. However, there is not much (some still helps, but not much) to be gained by having more pixels than the output modality needs. Enough is enough. When we have enough pixels to satisfy a 720 PPI output without upsampling, there is not much to be gained by having more (other than room to go larger and some postprocessing benefits).
Since microcontrast is lost at the verge of limiting resolution, it helps (some types of photography) to stay away
a bit from that absolute black hole which has no life in it anymore. The question then becomes, how far do we have to keep away from that dead-end by having some excess resolution?
In theory, in a perfect world, we could quantify that need as follows:
- It is generally accepted (e.g. by the ISO organisation) that a 10% modulation of the MTF curve, corresponds quite well with the perceived human visual limiting resolution. That would also leave some room to tweak that remaining low level of microcontrast in postprocessing. So everybody should be happy when we can maintain that level before we reach the Nyquist limit (with associated aliasing issues beyond that), so preferrably with some sort of low pass filter in place.
- A very good lens, at its optimum aperture, in combination with a sensor element (sensel) that takes an average aperture sample of the projected image, e.g. with a microlens, will produce a result in the optimal focus plane that resembles a Gaussian shaped Point Spread Function (PSF) with a radius of sigma=0.7. Proper deconvolution sharpening can boost that to perfect pixel resolution.
- Output resolution of 8 lp/mm (406 PPI) is generally considered to be very high quality (5 lp/mm is already good), but because of vernier acuity of human vision, we can still appreciate somewhat higher resolution as an improvement. Hence 600 or 720 PPI (=11.81 or 14.17 lp/mm) without interpolation is a practical goal and upper limit.
If we can accept those criteria, then a bit of math (you'll have to trust me on the validity of the derived formula) will tell us the following:
MTF10 is achieved at
cycles/mm = 1.072983 / (pi x senselpitch x sigma),
thus 14.17 = 1.072983 / (pi x senselpitch x 0.7), or senselpitch = 0.0344 millimetre without magnification. When we want to magnify our sensor size 10x for output (e.g. 36mm to 360mm), we then divide the 0.0344 mm, or 34.4 micron, sensel pitch by 10 (so a D800 with a 4.88 micron sensel pitch would not suffice). Or if we have a given sensel pitch, e.g. 4.88 micron, then we know we can magnify 99.98 / 14.17 = 7.06x (36mm to 254mm) without compromise.
When we need larger output with the same reading distance quality, we need more pixels (larger sensor, or by stitching). Of course we can reduce our pixel peeping requirement for more distant viewing conditions, unless we want to allow uncompromised close inspection.
This of course also assumes that one requires the MTF10 at this maximum resolution, which is not required for many subjects, but it is nice to know where to draw the line before being accused of obsessing too much ...
Cheers,
Bart
P.S. OOPS, I had made an error in simplifying the formula (a Log() vs Ln() conversion), I have corrected the relevant bits above.
