multiple precision arithmetic version of semidefinite programming solver

Why I developed a multiple precision version of semidefinite programming solver is
just in some case 2-RDM method fails to converge.

The gap of the Primal objective function and Dual objective function can be larger than 1e-4.
In this case we discard the result and re-run SDPA. For some molecules, it takes 2 weeks,
then discard? I was quite frustrated.

Maybe late of 2005 or 6, Yamashita-san came to the Univ. of Tokyo and we had some discussion,
and lunch. He told me that Nakata Kazuhide-san's student did some preliminary work. Their
result is somewhat incomplete as implement multiple precision arithmetic in Java and it is
not an implementation of floating point number!

So I decided to implement SDPA-GMP based on SDPA 6.

Using GMP is quite easy. just replace double to mpf_class. One large obstacle was it uses
BLAS/LAPACK routines. I implemented all routines used by SDPA 6.

This result was applied to the Hubbard model of large correlation limit. and published as JCP 2008. Also SDPA-GMP has been released in 2008, too.