Investigate commute time top-k algorithms.
=================

Commute time differs from Katz in that the solution is not a single 
system solve for a particular row.  This makes it much more difficult 
to approximate in a top-k sense.  Consider
  C(i,j) = (ei-ej)^T L^+ (ei-ej)
to get all the entries in the $i$th row, we need the diagonal
entries of L^+ for ALL $j$.  Eek!  This will not do.

So there is no localized algorithm for commute time.  The major
purpose of these experiments it to investigate how various approximations
to commute time work.  For example, is
  S(i,j) = -e_i^T L^+ ej 
a good surrogate for C(i,j)?  We explore this approximation, along with
many others.

Summary
-------

This experiment is a jumbled mess.  The current best-script 
(2010-02-22) is commute_approx, which includes a few 
different schemes.  The best one is marked with "current-best".

commute_approx_push uses a push algorithm to approximate the approximate
commute time measure.  Again, this uses the current best measure; which is our
F-measure (2010-03-08).

Lots of other Lp_push functions were small tests that did not work.

Lp_push_normalized : An attempt at using the normalized Laplacian for the solve
 it didn't work.  

