This reply expands on the discussion in the comments.

The averaging idea turns out to be the right one, as ably explained by Douglas Kerr in a nice little online paper. The basic ideas are two:

1. The lens "resolution" is most fully described by considering the mathematical relationship between the light leaving the subject and what reaches the sensor. This relationship, the "modulation transfer function," can be deduced from the simplest of all possible targets: a perfectly dark half-plane on a perfectly bright homogeneous background. Obviously the image on the sensor should be a region of light abruptly terminating along a perfect line. It never is perfect, though, and the imperfections affect the resolution. Ultimately the MTF is determined by looking at how the light intensity varies as we move straight out from the boundary (in both directions, into the dark and into the light) across the sensor.

2. It is a statistical fact that averages can be more precise than the measurements of which they are constituted. For typical measurement error, the precision follows an inverse square root law: to double the precision, you need four times as many measurements. In principle you can get as precise as you want by averaging enough independently repeated measurements of the same thing.

   This idea can be exploited (and is) in two ways. One is actual repetition, achieved by taking multiple images of the same scene. This is time consuming. The slanted-edge MTF analysis creates repetition within a single image. It does this by slanting the line slightly. This does not change the MTF in any material way and it guarantees that the patterns of the lens's response do not align perfectly with the sensor's pixels.

   Imagine the line being nearly vertical. Each row of pixels serves (almost) as an independent set of measurements of the MTF. The rows march outward from the line, almost perpendicularly. The pixels are registered with respect to the (ideal) line location in varying ways, producing slightly different patterns of response. Averaging these patterns over many rows has almost the same effect as taking multiple images of the line. The result can be adjusted for the fact that the pixels aren't quite perpendicular to the line.

In this way, the slanted-edge method can detect frequencies in the MTF that exceed the limiting frequency of a single image. It works due to the simplicity and regularity of the test pattern.

I have left out many details, such as checking that the line really is straight (and adjusting for slight deviations from linearity). Kerr's article is accessible--there's almost no math there--and well illustrated, so check it out if you would like to know more.