In my windows MSVC (both 32bit and 64bit) this code made an infinite loop :

Code: Select all

#include <extern/eigen/Dense>

using namespace Eigen;
int main()
{
   Matrix<double, 3, 3, RowMajor> mat;
   mat <<
      4.4331978442502944e+307,
      -5.8585363752028680e+307,
      6.4527017443412964e+307,
      3.7841695601406358e+307,
      2.4331702789740617e+306,
      -3.5235707140272905e+307,
      -8.7190887618028355e+307,
      -7.3453213709232193e+307,
      -2.4367363684472105e+307;
   Matrix<double,Dynamic,Dynamic> u,d,v;
   JacobiSVD<Matrix<double, Dynamic, Dynamic, RowMajor> > svd(mat, ComputeThinU | ComputeThinV);
   return 0;
}


The loop is the inner while of JacobiSVD.h compute method.
The documentation says :

...the computation is guaranteed to terminate in finite (and reasonable) time.

(From http://eigen.tuxfamily.org/dox/classEigen_1_1JacobiSVD.html).

It's very dangerous to work with such high values that are very close to overflow. I guess this explain why it does not converge. Before a proper fix, I suggest you to apply scaling:

Code: Select all

   double s mat.cwiseAbs().maxCoeff();
   mat /= s;
   JacobiSVD<MatrixXd> svd(mat, ComputeThinU | ComputeThinV);
   VectorXd d = svd.singularValues() * s;


The source of those numbers is a bug (uninitialized memory), so I don't really need their svd decomposition, but infinite loop is unacceptable...
tnx.

fixed.

https://bitbucket.org/eigen/eigen/commits/9034caa3d917/
Changeset: 9034caa3d917
User: ggael
Date: 2013-11-19 11:53:48
Summary: Add scaling in JacobiSVD to avoid overflows