I have wrong results on Mac (GCC 4.2.1) and Win64 (Mingw 4.6) using the following test code:

Code: Select all

        Eigen::MatrixXd matrix(10,6);
        matrix.setZero();
        matrix(0,0) = 1.0;
        matrix(1,1) = 2.0;
        matrix(2,2) = 3.0;
        matrix(3,3) = 4.0;
        matrix(4,4) = 5.0;
        matrix(1,0) = 1.0;
        matrix(2,1) = 1.0;

        Eigen::JacobiSVD<Eigen::MatrixXd, Eigen::FullPivHouseholderQRPreconditioner> svd( matrix , Eigen::ComputeThinU | Eigen::ComputeThinV);
//        Eigen::JacobiSVD<Eigen::MatrixXd> svd( matrix , Eigen::ComputeThinU | Eigen::ComputeThinV);
//        Eigen::JacobiSVD<Eigen::MatrixXd, Eigen::HouseholderQRPreconditioner> svd( matrix , Eigen::ComputeThinU | Eigen::ComputeThinV);

        Eigen::MatrixXd u = svd.matrixU();
        Eigen::MatrixXd v = svd.matrixV();
        Eigen::MatrixXd s = svd.singularValues();
        Eigen::MatrixXd AA = u * s.asDiagonal() * v.transpose();

        std::cout << "matrix: " << std::endl << matrix << std::endl;
        std::cout << "U: " << std::endl << u << std::endl;
        std::cout << "V: " << std::endl << v << std::endl;
        std::cout << "singular values: " << std::endl << s << std::endl;
        std::cout << "matrix (restored from SVD): " << std::endl << AA << std::endl;


The output is:

Code: Select all

matrix:
1 0 0 0 0 0
1 2 0 0 0 0
0 1 3 0 0 0
0 0 0 4 0 0
0 0 0 0 5 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
U:
-2.68156e+154  2.12467e-314  2.78656e-314   1.7644e-314  1.63239e-314  5.43472e-323
-2.68156e+154  2.12466e-314  2.71314e-314  1.22091e-314 -4.20287e-315             0
 2.12465e-314  5.43231e-312  3.66692e-312  9.85896e-312 -2.71945e-312 -2.68156e+154
 2.78134e-309             0   1.0571e-309  7.82591e-309   2.9558e-308 -2.68156e+154
 2.12466e-314  1.25987e-321 -2.52434e+154  8.92817e+153 -1.46106e+153  2.12465e-314
 4.17201e-309  2.12467e-314 -2.52434e+154  8.92817e+153 -1.46106e+153  2.78134e-309
 2.12467e-314  2.12467e-314 -8.99936e+153 -2.43375e+154  6.76556e+153  2.12467e-314
 2.12467e-314  2.12467e-314 -8.99936e+153 -2.43375e+154  6.76556e+153             0
 2.12467e-314  2.12467e-314 -9.26526e+152 -6.85923e+153 -2.59069e+154  2.12467e-314
 6.95336e-309  1.39067e-308     0.0691035      0.511585       1.93223  5.56268e-309
V:
        0         0   0.11309  0.545513  0.830437         0
        0         0  0.492679  0.695022 -0.523653         0
        0         0  0.862832 -0.468359  0.190162         0
        0         1         0         0         0         0
        1         0         0         0         0         0
        0         0         0         0         0         1
singular values:
5
4
3.27307
2.13264
0.859565
0
matrix (restored from SVD):
 4.24935e-314  6.37402e-314  6.37402e-314   8.4987e-314 -1.34078e+155             0
 2.12466e-314    6.374e-314  6.37402e-314  8.49864e-314 -1.34078e+155             0
 1.08859e-311  2.17505e-311  6.37402e-314  2.17292e-311  1.06233e-313             0
 3.05948e-308  1.97626e-323   1.4822e-323             0  1.39067e-308             0
-8.93139e+138 -2.68156e+154 -8.04468e+154  5.03947e-321  1.06233e-313             0
-8.93139e+138 -2.68156e+154 -8.04468e+154  8.49869e-314  2.08601e-308             0
-2.68156e+154 -5.36312e+154 -3.38649e+139   8.4987e-314  1.06234e-313             0
-2.68156e+154 -5.36312e+154 -3.38649e+139   8.4987e-314  1.06234e-313             0
-2.68156e+154 -8.93139e+138             0   8.4987e-314  1.06234e-313             0
            2   5.55112e-16   5.55112e-17  5.56268e-308  3.47668e-308             0


Note,
Eigen::JacobiSVD<Eigen::MatrixXd> svd( matrix , Eigen::ComputeThinU | Eigen::ComputeThinV);
Eigen::JacobiSVD<Eigen::MatrixXd, Eigen::HouseholderQRPreconditioner> svd( matrix , Eigen::ComputeThinU | Eigen::ComputeThinV);
are working. Only FullPivHouseholderQRPreconditioner is not working here. Matrix U is arbitrarily filled.

GCC flags: -O2, -DNDEBUG, -std=c++98

Best regards
Sebastian

You should try without -DNDEBUG to get the explanation:

assertion failed: !(m_computeThinU || m_computeThinV) && "JacobiSVD: can't compute thin U or thin V with the FullPivHouseholderQR preconditioner. " "Use the ColPivHouseholderQR preconditioner instead."

Ups. Thanks! It would be good to produce empty matrices in this case.

I always use NDEBUG because some of the sparse operations need too much time using the old GCC 4.2
Sorry for bugging....

Indeed, asserts are extremely slow with gcc 4.2 to workaround a compiler bug of this specific version.