OK, this is how far I've got:

 (1) I've tested Math::Matrix, and see that it works.
 (2) I've tested (and modified) the get_dctm and get_idctm
     routines; the versions in blur.pl now work.

 (3) Ian's adding x0/2 back in is close but not quite right;
     I've included that in the get_idctm routine.

     (explain this to Ian)

 (4) blur.pl runs without an error; however,
     the jpeg file that is created won't display.
     So something weird is going on in dct_transform, I think.

  Next testing steps : 

        * make sure that we can output an untouched copy of the 
          the original jpeg

        * look at the actual numbers for a tiny (like 4x4) jpeg
            (a) before anything happens,
            (b) after the DCT,
            (c) after the inverse DCT
            (d) make sure (a) and (c) are the same

        * make sure we can output this as an image that
          looks like the original (except for roundoff).

  Do this all on a miniscule file, not a 100x100 thing that
  we can't really examine.