GitLab

Summary

It appears I am getting incorrect eigenvalues for a symmetric positive-definite matrix. It results in incorrect eigenvectors which end up being linearly dependant (as if the matrix was defective). This impacts the turbulentInletDFSEM as eddy objects lose a dimension.

Steps to reproduce

Here's a minimal example, I concocted:

#include "fvCFD.H"

int main(int argc, char *argv[])
{
    Info<< "\nCalculating eigenvalues of a symmetric positive-definite matrix\n";

    symmTensor A = symmTensor(
        0.005,
        -0.0003,
        0.001,
        0.002,
        0.0001,
        0.002);
    vector lambda(eigenValues(A));
    Info<< "Eigenvalues (OpenFOAM): " << lambda << endl;
    Info<< "Eigenvalues (numpy.linalg): (0.00532289, 0.00160802, 0.0020691)" << endl;
    return 0;
}

What is the current bug behaviour?

Output I am getting:

Eigenvalues (OpenFOAM): (0.00183856 0.00183856 0.00532289)
Eigenvalues (numpy.linalg): (0.00532289, 0.00160802, 0.0020691)

What is the expected correct behaviour?

See above. I compared it with numpy.linalg.

Environment information

• OpenFOAM version : v2006
• Operating system : Arch Linux
• Hardware info : Intel(R) Core(TM) i7-4810MQ CPU @ 2.80GHz
• Compiler : gcc :

Possible fixes

Use alternatives eigensolvers. I currently went with Eigen but I am looking for another OpenFOAM native solution.