Hi Dale,
This answer confuses two independent properties of a refinement - the target function and the optimization method.
to be accurate, the answer confused model parameterization (rigid body) and refinement target (LS vs ML, computed using different subsets of reflections). The answer did not involve the optimization method (e.g. minimization in this case). Too much jargon is never good.
Low resolution data are sufficient to define the optimal rigid body parameters. With the least-squares target the presence of the high resolution data reduced the radius of convergence of the optimization making the reduction of the resolution limit mandatory. A good ML target should set all the high resolution gradients to zero making them irrelevant. As has been mentioned elsewhere, since it is just a very computationally expensive way to calculate zero one can save time by reducing the resolution limit anyway.
That is what we hoped. It turned out it's not - ML does not set all the high res terms to zero (probably, for the reason you mentioned below), or does it insufficiently compared to manual cutting high res data off. That's what we observed when preparing that "Afonine's 2009 paper".
I should emphasize "good" in good ML target. The calculation of sigma A, itself, assumes that the atomic positional errors are uncorrelated, so the currently used ML target is not a "good" ML target for models with this type of error. This is what, I believe, is the cause of the resolution limit effects reported in Afonine's 2009 paper.
Pavel