Intel® Integrated Performance Primitives
Community support and discussions relating to developing high-performance vision, signal, security, and storage applications.

Accuracy of Eigenvalue decomposition

michael_burisch
Beginner
167 Views
I have a 3x3 matrix for which i want to calculate eigenvalues/-vectors. Using the function "ippmEigenValuesSym_m_64f" i get an ippStsSingularErr error meaning the matrix is singular, which is not the case.
The function "ippmLUDecomp_m_64f" however performs without problems on the same matrix.

The matrix elements are in the range of 1e-5 to 1e-6. Scaling the matrix by a factor of 10 yields valid results.

Are there any requirements on the size of the elements for the eigenvalue decomposition?
0 Kudos
3 Replies
Vladimir_Dudnik
Employee
167 Views

Hello,

what is your IPP version? It seems we fixed similar issue in the past. Please note that the latest available package is IPP 5.3 update 4. (and IPP 6.0 beta 2)

Regards,
Vladimir

michael_burisch
Beginner
167 Views
Hello Vladimir,

I am using IPP 5.3 update 4.

The behaviour can be reproduced for example by the eigenvalue decomposition of the following matrix (i.e. the identity matrix scaled by 1e-6).
A=[1e-6, 0, 0;
0, 1e-6, 0;
0, 0, 1e-6 ];

Setting this to 1e-5 produces a valid result.

Thanks,
Michael


Vladimir_Dudnik
Employee
167 Views

Hi Michael,

then it seems to be a different issue. I've informed our engineers on that, they will work on investigation of this. Please submit your issue report to Intel Premier Support.

Regards,
Vladimir

Reply