So just to bump this thread I imported a 72K jpeg (linked) into a ai file and then saved it in numerous permutations. The only really small file sizes came from distilling a print to postscript file. The permutations were one image as linked or embedded or as a symbol (use_type) one instance or 6 instances of image (one/many) and saving as ai (no PDF content) or PDF (no ai content) or Distilling a postscript print file (doc_type).
So I guess that makes:
P = (use_type) . (one/many) . (doc_type)
= 3 x 2 x 3
There are 20 different versions plus 2 distiller files in the results. I did it a couple of times with slightly different scales on the jpeg object so maybe that is resulting in small size differences for same permutations done twice (difference in image compression). I kept all image compression and page thumbnails, colour profiles etc. off with exception of file compression for .ai files (it's on).
I included symbol instances because an adobe sales manager/trainer/evangelist told a group of employees I was contracted to train to use symbols for multiple instances of images (saying this was a new feature but when I said isn't that just like a brush or linked image he said yeah same result: OO embedding, guess he was trying to idiot proof the process rather than educate in a short space of time). Symbols result in bigger files than linked but are preferable to multiple embedded files.