My experience is that UMFPACK sometimes gives results also for singular 
matrices. 

If the matrix comes from a Navier-Stokes system, the blocks A11 
and A12 must fulfill an LBB-condition: for example velocity (A11) 
quadratic, pressure (A12) linear. The boundary conditions for the 
velocity should give a A11 invertible and for the pressure you have to 
filter away a constant.

Which is the system of equations that you want to solve?

Regards

Thomas

On Mon, 24 Nov 2008, Martin Kronbichler wrote:

> Dear Isabel,
> 
> > I am trying to solve a matricial system Ax=b, where
> > A=|A11 A12|
> >      |A21     0|
> > 
> > with four different solvers:
> > 
> > - SparseDirectUMFPACK
> > - GMRES + Precondition_Identity() as preconditioner
> > - GMRES + SparseILU<double> ilu as preconditioner
> > - GMRES + Vanka preconditioner as preconditioner
> > 
> > The only one that seems to get good results is the first one, i.e., the 
> > direct one. I decided to use the other ones because in the future this 
> > system will have a huge quantity of nodes, so direct solver won't be 
> > feasible.
> > 
> > GMRES with SparseILU as preconditioner has a Starting Value of 946,719
> > GMRES with Vanka as preconditioner has a Staring Value of 2.99566E+12
> > In those cases, the solver breaks down in the first iteration.
> > 
> > GMRES with Precondition_Identity as preconditioner has a Starting Value 
> > of 0.0129254 (i.e. the same as doing "by hand", norm L2 of(b-Ax)). In 
> > this case, the solver solves the first iteration but it breaks down just 
> > in the second one.
> 
> Since the direct solver does not break down, the matrix system seems to
> be solvable (i.e., there seems to be one unique solution, otherwise
> UMFPACK would complain/fail). In such a case, GMRES with
> Precondition_Identity() should actually work fine as well (even though
> it might need lots of iterations). The only reason that makes GMRES fail
> is when some matrix inner product y^T A x gets zero, which means that
> the matrix is actually not solvable (but singular).
> 
> Are you sure that your matrix is non-singular (maybe you were just lucky
> to get UMFPACK running)? Where does the matrix come from (I guess some
> Stokes or Navier-Stokes system) and what were the boundary conditions?
> 
> Best regards,
> Martin
> 
> 
> _______________________________________________
> 

_______________________________________________

Reply via email to