--------------- Sept. 6, 2006 ---------------

Speed up dense-matrix-mod-p.  Use a 1D array?  Use rem everywhere?

Get other levels going on Nicholas's machine, Connie's, abel,
dirichlet, ....

----------------------------------------

After the meeting with Avner and Paul on August 24, 2006, when
Sheafhom 2.1.9 was massively converted to Sheafhom 2.2, supporting any
Euclidean domain:

Figure out why 31991 went wrong so that we had to switch to 12379.
Fix 31991.

Run the 31991 computations for each level N = 3, 4, ..., 1000.  This
is for the Part II paper with Avner and Paul.

On the side, run Shh 2.1.9 over Z on all the levels we can reach today
in order to check that the Q-rank and F_31991-rank of H^5(..., C_triv)
coincide.  It is also forward-looking for Part III.

Generalize to PID's?  PIR's?  Dedekind domains d'apres Cohen?  ...If
homolalg is not working well mod p^2, or just so we can reach higher
levels, change "domain" to "commutative ring with 1" so we can work
over Z/(p^2) and other quotients Z/(composite).  De-emphasize
Euclidean, and emphasize PIR instead.

Get back to the modular hnf idea so that Z computations work very
well.  This will be for the Part III paper.

Get back to homolalg.  Work with arbitrary Z-modules of coefficients.
This will be for the Part IV paper.

Write up our 1996 spectral sequence for the Borel-Serre boundary.
This will be for the Part V paper.

--------------- Thoughts from June 2006 ---------------

Use series for (.) a wrapper around snfp*.txt etc.; (.) msets.

A longer-term idea: replace numthy with code that calls out to Pari!

I spent all that time in csparse specifying which variables are fixnum
or nnfixnum.  But someone on c.l.l pointed out that you can't use
(safety 0) in a library because it disables argument checking =>
should use (safety 1) => fixnum addition is checked in ACL.  Run some
disassemble tests, and timing tests, to see how much we *REALLY!!!*
lose with safety 1 vs 0.

---------- After March 15, 2005 ----------

Put colors into the movies: blue for 1 digits, red for 100+

Save the movies as jpgs and assemble them?

When using pure Markowitz in movies, do the ut sort
anyway, and compare the two algorithms on the same matrix.

Use LLL in strips to find the strip's kernel and eliminate columns
from the strip (idea 3/14).

Put in the effort to port it to CLISP and GCL, to see who performs
better on bignums.  [Note added spring 2006: CLISP uses much less
memory and performs well on bignums, but is about 4 times slower on
the routine parts.]

---------- Before June 20, 2005 ----------

Run all the single-float code [a little csparsez and gui, mostly
two-complex and minimize] through ACL's compiler flags to make sure
the single-float declarations are all there.  [Note added August 2006:
assess how much it matters.]

Check the declamations in all the .lisp files, especially those after
the submission to Lispwire.
