Jack de Valpine or John Mardaljevic may know more about this than I do...
Actually, I can't really expand much on what is in the book (p387):
Having created the ambient file with the “overture” calculation, you can, with
caution, relax some of the ambient parameters for the larger renderings. The
parameter revisions could be one or both of the following:
• Reduce -ad and -as by about 50%
• Slightly increase -aa (i.e., by 0.05 or 0.10)
The other ambient parameter settings should not be changed. If you do decide
to change any of the -ad, -as, or -aa settings after the “overture” calculation, you
should be aware that the modifications will not be reflected in the header of the
ambient file. Thus, you need to track both the picture and the ambient file headers
to obtain a complete record of the parameter settings for an image.
I seem to recall that Greg reckoned it was not a good idea to reduce the ambient resolution (ar) parameter when reusing the ambient cache. But I can't quite remember the reason why. I was going to end my contribution at this point. But then I did some tests. Load up:
What we have here is a test scene: a ridged box by a window and illuminated by a diffuse (i.e. glow) sky. The ar parameter limits the extent to which ambient sampling can convincingly shade the grooves. I populated the ambient file artest.af in creating the image at the top which only sees half the box. I then reused the ambient file with lower resolution settings to create wide views of the box -- I thought that seeing the "boundary" between the already populated hi-res ambient settings and the low-res might be interesting. It was. Note that for each of the three low ambient resolution images I reused the *original* artest.af (i.e. I kept a copy).
In the left and bottom images I relaxed the aa and ad parameters respectively. The results are what most of us would probably expect. For the image on the right I relaxed the ar parameter. Now, for this image, the "new" part on the left-hand-side is what would be expected from a fresh run without any pre-existing ambient file -- ar 64 doesn't allow close-enough sampling and the grooves get some "flat" ambient shading . The interesting effect is of course on the right -- the pre-existing ambient samples on this side have somehow produced this lovely frog-spawn pattern. Why? 
PS. Dontcha just love this "I'll never quite master it" side to Radiance?
 When av is zero, as it was here, how is the ar-limited shading value arrived at?
 I have this vague feeling that I sort of might know the reason why. But it is all too hazy for me to attempt to put it into words.
Dr. John Mardaljevic
Senior Research Fellow
Institute of Energy and Sustainable Development
De Montfort University
LE1 9BH, UK
+44 (0) 116 257 7972
+44 (0) 116 257 7981 (fax)