With proliferation of half-frame 10MP cameras, a natural question is:
what is the MP count starting from which the higher resolution of the
sensor does not help to *significantly* improve the total lens+sensor
resolution. Of course, this depends on the lens, so it makes sense to
discuss the diffraction-bounded region of f-stops.

Short answer: with a f/16 diffraction-bound lens, the total resolution
should start to level out with 30MP full frame sensor
(one similar in MTF-relative-to-pixel-count to current
generation of sensors; see P.S. for other possibilities).

One way to do this in experimental fashion, is to compare shots made
with an "ideal lens" and a real-life sensor, to shots made with an
"ideal sensor" and a real-life lens. Since in practice "defects"
contributed by the lens and the sensor combine, this would compare the
relative magnitude of these defects.

One can use the result of this comparison like this: of course,
increasing the resolution of sensor would bring some improvement in
the total resolution in a very large region (keywords are `cut-off
spacial frequency', and `Nyquist limit'). However, at some point the
benefits will start to level out; our target is to find this point.
We assume that it is close to the point where "defects" contributed by
the lens are comparable to "defects" contributed by the sensor.

First experiments (lens which provides much better resolution than
real sensor): it is easy if you believe the "short answer" above: any
resolution charts shot with a decent lens would be in this ballpark.
To avoid circular reasoning, one should do more; e.g., compare
resolution charts shot with half-frame 3MP sensor to those shot with
16 MP FF sensor; one can easily see that at 100% magnification, there
is little difference, thus the contribution of lens is negligible indeed.
Judging by these shots, one can give a simple model [**] of a digital
sensor of today; to provide a different resolution, just rescale.

The second scenario (ideal sensor, real lens) is also easy to obtain:
good film is an ideal sensor media when used in 4x5in (or larger)
format (at least with strongly stepped down lens). I use (wonderful)
data in

http://www.clarkvision.com/photoinfo/large_mosaics/

as a source of "high-quality" image (digital mosaic), comparing it to
a "low-quality" (1/2 ;-) image (4x5in Velvia shot made with aperture
f/64).

The comparison shows:

A) Resolution of the film shot is very similar to the resolution of
high-quality shot scaled down to 16 MPixel JPEG[*].

B) Resolution of the film shot is very similar to a (simulated) image
taken with 30 MPixel digital camera [**]

[Of course, the possibility to redo the shot on a 30MP FF digital
camera assumes a presence of f=22mm tilt lens with good
performance at f/16. Performance at f/16 is not a big deal; but
I suspect tilt-f=22mm lens is not available; one would need a
tilt sensor to compensate).

[This assumes that the film shot Roger did has no significant out-of-focus
and blur areas. Since he says it took 6hours to set up this shot, I assume
this holds.]

===========================

Since 4x5in format is very easy to extrapolate (up to f/32 there is no
significant dependence on the quality of the lens design), this allows
to predict "digital equivalents" of other shots (of course, to do this
one needs to assume that the film shot is in ideal focus; I hope we
can trust Roger on this; he says he was waiting for 5 hours for ideal
conditions to make this shot ;-):

f/96: 15 MPixel sensor;
f/64: 30 Mpixel sensor;
f/48: 60 Mpixel sensor;
f/32: lower than 120 Mpixel sensor (abberrations start to feel);
f/22: about 120 Mpixel sensor (assuming lens resolution
saturates near f/22, and ).

===========================

One can also rescale the numbers above to 24x36mm formfactor with
typical DSLR-type fixed-focal-length lenses:

f/22: 15 MPixel sensor;
f/16: 30 Mpixel sensor;
f/11: less than 60 Mpixel sensor;
f/8: about 60 Mpixel sensor (assuming lens resolution
saturates near f/8).

With highest-quality rangefinder-type lenses (AFAIK, these are
available with f about 45mm and above):

f/22: 15 MPixel sensor;
f/16: 30 Mpixel sensor;
f/11: 60 Mpixel sensor;
f/8: less than 120 Mpixel sensor;
f/5.6: about 120 Mpixel sensor (assuming lens resolution
saturates near f/5.6).

With high-quality zoom lenses:

f/22: 15 MPixel sensor;
f/16: less than 30 Mpixel sensor;
f/11: about 30 Mpixel sensor (assuming lens resolution
saturates near f/11).

(All these number assume some "theoretically sound" behaviour of MTF
in the region N steps down the optimal resolution of the lens:

2 steps) Practically diffraction-bound resolution;
1 step) Slightly less than diffraction-bound resolution;
0 steps) About the same as diffraction-bound resolution when
closed 1 step more.

Of course, any particular lens with show some variations, but AFAIK,
the variations are negligible)

===========================

Here is some more info on how I interpret Roger's short. Velvia at
1/64 is more or less an ideal sensor; it has practically no noise, and
practically ideal MTF. So the film shot is equivalent to, e.g., f/16
24x36mm shot made with an ideal sensor.

The "high-resolution" shot is made at f/11; since we shrank it 2.5x, the
diffraction doesn't matter; likewise for focussing errors. So, essentially,
it is "an ideal image" used with the given sensor. So one can see that
an ideal (diffraction-, abberration-, and focus-error-less) shot made
with 30 MPixel sensor is very similar to f/16 shot made with an ideal sensor.

===========================
[*] I took the crop of high quality image, reduced it to 29% linear size,
then enlarged it to 345% (to get the initial size). Then I compared
it with the film quality image. [Replacing 29% with 25% or 33%
produces images which differ significantly from the film image:
worse looking and better looking correspondingly.]

[**] To simulate a shot made with an "ideal lens" and a digital sensor
similar in construction to today's sensors: take a "reasonable
quality" image; scale it down several times; you get a very high
quality image. Now blur it to reduce MTF at high spacial frequencies.

My experiments show that the following convolution matrix produces
results very similar to real shots (tested with resolution chart
shots from DPReview): [assumes fixed-width fonts; should be divided
by 13.5]:

1 1 1
1 5.5 1
1 1 1

(Convolution with this function reduces the MTF 3 times at 2/3 of
cut-off-frequency of the sensor.)

Now: this is what I did: I took the crop of high quality image,
reduced it to 40% linear size, then enlarged it to 250% (to get
the initial size). Then I compared it with the film quality image.

Enjoy,
Ilya

P.S. Of course, only the "sensors of today's construction" are
covered. And with high MP count, there is no need to
(significant!) decreases in resolution and "sensibility" implied
by anti-aliasing filters (AAF *are* needed with the minuscule MP
counts of todays's sensors). E.g., one can easily show that
about 40MP FF sensors, there would be no need in AAF (at least
when used with SLR-type lenses). This would improve the
resolution [#] similar (?) to about 1.6x increase of MP count.

So with such a sensor, one gets (for SLR fixed-focus lenses):

f/22: 10 MPixel sensor;
f/16: 20 Mpixel sensor;
f/11: less than 40 Mpixel sensor;
f/8: about 40 Mpixel sensor (assuming lens resolution
saturates near f/8).

Similarly, with rangefinder-type lenses, the beginning-of-leveling-out
point would be about 80MP (needed at f/8 and f/5.6 shots).

[#] (Actually, removing AAF would also *significantly* increase
sensitivity [measured via the exposition needed to achieve a
certain visible S/N level], but this deserves a separate message...)