running the following code snipped on a MatrixXd A produces a matrix U that has NaN's in it.

Code: Select all

Eigen::JacobiSVD< MatrixXd> svd = A.jacobiSvd(Eigen::ComputeThinU | Eigen::ComputeThinV);
Eigen::MatrixXd U = svd.matrixU();
Eigen::MatrixXd V = svd.matrixV();
Eigen::VectorXd S = svd.singularValues();

I'm curious (1) how is that possible and (2) how can I avoid it? This part of my application is not performance critical at all, so I'd be happy to use a different SVD algorithm if (for example) Jacobi simply sometimes runs into numerical issues.

Thanks,
Rafi

This is probably due to overflow issues, perhaps you could apply a scaling prior to factorization. We already had such a discussion in this forum a few month ago.