Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

- Intel Community
- Software
- Software Development SDKs and Libraries
- Intel® oneAPI Math Kernel Library
- Eigenvalue decomposition with CSR sparse matrix

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Mute
- Printer Friendly Page

Michael_W_1

Beginner

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

03-04-2014
02:42 AM

96 Views

Eigenvalue decomposition with CSR sparse matrix

Hi,

at the moment I use dsyevd to compute the eigenvalues and eigenvectors of a large matrix A (n = 22000). This takes about half an hour. I know that they are a lot of zeros in matrix A (90% are zeros). Matrix A is stored as CSR sparse matrix.

- Is there a function to compute the eigenvalues and eigenvectors of a CSR sparse matrix?

- Is there a function to convert a CSR sparse matrix to a band matrix? Then I could use dsbevd.

Regards Michael

Link Copied

4 Replies

Gennady_F_Intel

Moderator

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

03-04-2014
07:10 AM

96 Views

yes, since the version 11.0 mkl contains the Extended Eigensolver Routines -- please see reference manual for more details. These routines support CSR format too.

there are no routines convert CSR->Band format, ut there are a number of routines conversion csr<->dense

mecej4

Black Belt

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

03-04-2014
07:25 AM

96 Views

Michael W. wrote:

... I use dsyevd to compute the eigenvalues and eigenvectors of a large matrix A (n = 22000). Matrix A is stored as CSR sparse matrix.

There is something amiss here.

- The ?syevd routines operate on a matrix kept in dense-storage form.
- You cannot pass a matrix stored in CSR form to such a routine.
- A banded matrix is sparse, but not vice versa. If you know that the matrix is banded, look for an eigenvalue/vector routine that is tailored to such matrices, rather than to the more general sparse matrix type.

mecej4

Black Belt

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

03-06-2014
11:11 AM

96 Views

Michael W. wrote:

A = B x B'

That is very useful information. The eigenvalues of A = B.B' are obtainable from the non-zero **singular values** of B. There are several routines available to compute the SVD (**S**ingular **V**alue **D**ecomposition) of a dense matrix; see, for example, ?gesvd() in Lapack/MKL. You are probably in need of only the singular values and may look for a routine that allows you to specify that the singular vectors are unwanted.

Michael_W_1

Beginner

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

03-10-2014
03:22 AM

96 Views

Hi,

1. What I have as Input is CSR sparse Matrix B (number of rows: 20000, number of columns 100000)

2. Intermediate result A = B x B'

3. Intermediate result V, D = dsyevd(A) where V are the eigenvectors and D are the eigenvalues

4. Intermediate result E: diagonal Matrix. The elements on the diagonal are the inverse values of D.

5. Final Result W = V x E

So if you know a faster to compute W from B, please let me know.

Regards

Michael

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

For more complete information about compiler optimizations, see our Optimization Notice.