topic Low Accuracy on dgeevx results, how to increase accuracy? in Intel® oneAPI Math Kernel Library

Low Accuracy on dgeevx results, how to increase accuracy?

utab — Fri, 01 Aug 2014 15:18:07 GMT

Dear all,

I am trying to solve a non-symmetric tridiagonal eigenvalue problem with dgeevx from MKL, however the accuracy of the eigenvalues are not that good. I am also using dsyevr for symmetric tridiagonal matrices and for that routine there is a tolerance setting. I was wondering if there is a way to increase the accuracy of dgeevx. At the moment, I am solving without balancing and scaling.

Best regards,

Umut

On what basis do you conclude

mecej4 — Fri, 01 Aug 2014 16:11:29 GMT

On what basis do you conclude that DGEEVX is of low accuracy? Is this experienced with any tridiagonal matrix, or can you give a specific example? What is the size of the matrix?

Have you considered using specialized software, e.g., see the link to Dario Bini's Eigensolve at the end of the page http://fibonacci.dm.unipi.it/~bini/Ricerca/ric.html ?

Quote:mecej4 wrote:

utab — Fri, 01 Aug 2014 16:35:27 GMT

mecej4 wrote:

On what basis do you conclude that DGEEVX is of low accuracy? Is this experienced with any tridiagonal matrix, or can you give a specific example? What is the size of the matrix?

Have you considered using specialized software, e.g., see the link to Dario Bini's Eigensolve at the end of the page http://fibonacci.dm.unipi.it/~bini/Ricerca/ric.html ?

I am comparing the results with the results of MATLAB, eig , for accuracy.

These are the reduced problems of a Lanczos solver and sizes range from 40-50 to a value around 500(if the block sizes are large )

Yes I can give an example, but need more time, I will soon send a matrix where the accuracy seems to be low.

Thx.

Umut