Hi all,

I have seen the comments on eigenvalues in the CG solver class and the
step-36 tutorial which solves an eigenvalue problem, so I know that
deal.II has capabilities to solve eigenvalue problems. I am wondering
whether it would be straight forward to solve a problem of this type:

>From a combination of a finite element model and the ODE boundary
conditions applied we obtain a system x'(t) = Ax(t) + Bu(t). The
eigenvalues of A are useful for determining stability of the overall
system. However, A is not symmetrical (because of the boundary terms)
and its sparsity pattern is not typical for FEM since it is a
combination of the boundary conditions (represented by state-space
models) and the inverse of a FEM matrix.

It would be nice to be able to avoid explicitly forming A, and perhaps
a trick similar to the SchurComplement in step-22 could be used to
create a vmult method that does not require A. Then CG could be used
to extract coefficients that can then be used to estimate the
eigenvalues (which also requires finding eigenvalues of another
matrix, but presumably a smaller matrix and tri-diagonal so less
computationally intensive). However the matrix is not symmetrical so
this idea fails.

Step-36 uses wrappers to SLEPc (PETSc underlying) and if A can be
formed as a sparse matrix (with a suitable sparsity pattern), then it
appears that the approach taken there would work. It appears that
SLEPc can handle non-symmetric matrices. Lapack also can find
eigenvalues of real nonsymmetric matrices and presumably also requires
the matrix up front since it first converts it to an upper Hessenberg
form. I see that LAPACKFullMatrix already provides two methods to
compute eigenvalues, I presume that one of these could be used or at
least looked at for the implementation of a suitable method.

Am I correct in thinking that the idea of avoiding assembling A is a
non-starter, but that either SLEPc or LAPACK are viable options for
finding the eigenvalues of A in deal.II?

Thanks for you help,
Michael
_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii

Reply via email to