== I_m and the implied A ==
   predicted c0 = 2 (from Lichtman's f_k -> 1)
   predicted A  = -log2/4 = -0.17328679514 (from GLW's a_1 = d log2/4)

  m            I_m                  I_m - 2        (I_m-2)/(m^2 2^-m)
   8  1.9781164364257271198  -0.0218835635743  -0.0875342542971
  12  1.9967773337515367226  -0.00322266624846  -0.0916669510674
  16  1.9996202620664501565  -0.00037973793355  -0.0972129109888
  20  1.9999608253639185076  -3.91746360815e-5  -0.102693958009
  24  1.9999963030569146576  -3.69694308534e-6  -0.107681272018
  28  1.9999996725943569709  -3.27405643029e-7  -0.112101126382
  32  1.9999999723455980154  -2.76544019846e-8  -0.115990968862
  36  1.999999997747881875  -2.252118125e-9  -0.11941695918

== stability of the extrapolated A ==
A three-parameter fit through three points has ZERO residual, so no single
fit is an error bar.  What is informative is which basis is stable.

fit                                                 A    rel.err
log m/m + 1/m  (28,32,36)               -0.1737870386   0.002887
log m/m + 1/m  (24,28,32)               -0.1730071625   0.001614
log m/m + 1/m  (20,28,36)               -0.1727731077   0.002964
1/m + 1/m^2    (28,32,36)                -0.159442436    0.07989
log m/m only   (32,36)                  -0.1583382646    0.08626

The one tabulated basis omitting log m/m (1/m + 1/m^2) lands ~8.0% off and
the two-parameter log m/m fit ~8.6%, so the three-parameter fit with that
correction is doing real work and A then agrees with -log2/4 to a few parts
in 1000.  This is NOT independent corroboration of GLW's O(k log(k+1)):
the basis was chosen with that remainder already in mind.