Re:Power Method for Approximating Eigenvalues [pdf]
by gcr on Wednesday December 27, 2017 @09:59AM
(#16015060)
Once you know the dominant eigenvector, I recall there was some trick you could do to get the second-dominant eigenvector, by projecting the dominant one out somehow. How can you repeat power iteration to get all the eigenvectors of a matrix?
[ Parent
| on HN ]
Re:Power Method for Approximating Eigenvalues [pdf]
by qwerty1793 on Wednesday December 27, 2017 @11:41AM
(#16015827)
The largest eigenvalue of (A - k)^{-1} is 1 / (x - k) where x is the smallest eigenvalue of A that is larger than k. So by using this technique on (A - k)^{-1} for various values of k you can find all of the eigenvalues of A.
Re:Power Method for Approximating Eigenvalues [pdf]
by acidburnNSA on Wednesday December 27, 2017 @03:40PM
(#16017843)
In practice this is numerically unstable and it's generally better to use methods that use a larger and orthogonal subspace (i.e. Krylov methods) to get multiple eigenvalue/vectors. In nuclear reactor problems you can maybe get about 5 eigenvalues by filtering the larger ones but after that it gets noisy fast. I've gotten up to 1000 good eigenvalues from a large neutron diffusion problem using Arnoldi.
Re:Power Method for Approximating Eigenvalues [pdf]
by jdonaldson on Wednesday December 27, 2017 @10:12AM
(#16015136)
You're talking about Arnoldi Iteration. It's useful in a few cases, but it doesn't scale as well due to the fact the component matrices are no longer sparse.
Re:Power Method for Approximating Eigenvalues [pdf]
by thomasahle on Wednesday December 27, 2017 @10:30AM
(#16015277)
If you know the first one, you can project everything onto its orthogonal plane.