[ODE] ODE announcement: new iterative solver

[Russ Smith]

> Does it also mean we don't have to use a fixed timestep?  Because from
> reading Alen's email it seems like their method didn't use CFM; wasn't
> that what caused blowups when a long step followed a short one?  Or
> was that ERP...

actually the QuickStep method performs the same as the big matrix method
if enough iterations are used, so the problems regarding variable
timesteps is still there. that is a separate issue which i will get
around to fixing at some point. the QuickStep method is, as Alen noted,
less susceptable to low-CFM problems, for an interesting reason: the
problem with low CFM in the big matrix method is that it can make the
matrix near-singular in some cases, which causes major numerical
problems for direct factorizers because the exact solution involves a
delicate balance of very large quantities. but iterative solvers with a
limited number of iterations dont give you the *exact* solution, which
in this case is great because the exact solution is not what you want
... the approximate solution can actually be better in these cases.

for those who are interested in the method, the QuickStep method is the
successive over-relaxation LCP method described by Katta Murty in his
book. it is provably convergent to the big-matrix method, if enough
iterations are used. the FastStep method is a kind of wierd block-jacobi
method, i am not sure that it would always converge to the correct
answer.

russ.

-- 
Russell Smith
http://www.q12.org