Hi,

I've got a dense matrix of doubles 5000x500,000, and I'd like to get its first singular value. I don't need the singular vectors.

Currently I'm using

Code: Select all

JacobiSVD<MatrixXd> s(X);
double s0 = s.singularValues().maxCoeff();


which takes something like 40min on a 2Ghz Intel Xeon (gcc 4.6, -O3, -DNDEBUG).

Are there any "shortcuts" to speed it up?

Thanks

JacobiSVD is not designed for such large matrices. Since you only want the largest one, you might compute A = X X^T and extract its largest eigenvalue implementing Power iterations (4 lines of code) or using Eigen's interface to ARPACK (in de devel branch, in unsupported/Eigen/ folder).

Thanks, I found viewtopic.php?f=74&t=108033 which implements power iterations.

The ARPACK support doesn't seem to build on the latest devel, the code looks for ArpackGeneralizedSelfAdjointEigenSolver.h, but the file is actually called ArpackSelfAdjointEigenSolver.h.