gsl/randist vs ranlib

[Brian Gough]

James Theiler writes:
 >   This is something we are going to look into a little closer, but in
 > the meantime, we are soliciting opinions on this topic: maybe others
 > have already noticed similar behavior, and have some wisdom to impart.
 > Also, it is to suggest a new item in gsl's TODO list, which is a more
 > comprehensive examination of the performance of gsl's randist
 > algorithms, with an eye to the ranlib implementations either as a
 > benchmark, or as some code to possibly assimilate.

Thanks for looking at this.  At the time I didn't do any
optimisation for the routines that I wrote, I just used the
algorithms from Knuth.

One point: I think it is best to keep the routines stateless,
as they are now, for simplicity.  Your discrete method with
the Walker preprocessing is the easiest way to handle repeated
sampling for the finite distributions.

Brian