Cofactor canonicalisation of mpn_gcdext

[Torbjorn Granlund]

nisse at lysator.liu.se (Niels Möller) writes:

  It almost spells out the result we're talking about. Part (v) says, in
  our notation:
  
     |t| <= u / r_{\lambda-1}, |s| <= v / r_{\lambda-1}
  
  where
  
    r_{\lambda-1} = r_\lambda q_\lambda = g q_\lambda
  
  is the last-but-one nonzero remainder in the remainder sequence.
  
  So equality, |t| = u / g or |s| = v/g, is not ruled out, as far as I
  see. When does that happen, and which border is hit? When v/g is even,
  one could choose between the ranges
  
I suppose u = 0 or v = 0 both give equality.  And so does |u| = |v|.  I
suppose these are the only cases where we get equality, but I have no
proof...

-- 
Torbjörn