[eigen] Re: Custom Scalar in QR

[Pavel Holoborodko <pavel@xxxxxxxxxxxxxxx>]

It was bug on my side (in product kernel specialization). 

Sorry for disturbing.

On Tue, Sep 9, 2014 at 3:54 PM, Pavel Holoborodko <pavel@xxxxxxxxxxxxxxx> wrote:

I have noticed strange behavior when HousholderQR (default branch) is used with mpfr::mpreal.

Accuracy is lost when QR is applied to rectangular matrix, m < n.

        //typedef double RealScalar;

        typedef mpfr::mpreal RealScalar;

        typedef Eigen::Matrix< RealScalar, Dynamic, Dynamic> RealMatrix;

        int m = 2, n = 4;

        RealMatrix a(m,n);

        RealMatrix r, q;

        srand(0);

        a..setRandom();

        HouseholderQR<RealMatrix> qr(a);

        r = qr.matrixQR().triangularView<Upper>();

        q = qr.householderQ();

        RealScalar e = (a-q*r).norm();

        

        std::cout <<"A-Q*R = "<< e << std::endl; // should be of machine epsilon scale

Output for "double" (correct):

   A-Q*R = 2.88444e-016

Output for "mpreal" (accuracy is lost):

   A-Q*R = 1.40222

This affects "default" branch. Branch 3.2 (with all recent updates) produces correct result.

Similar situation happens with dense LU.

Decided to report issue in the mail-list since it is appeared very recently and probably can be fixed quickly while it is "hot".

Would appreciate any ideas/advice what (recent) changes might cause the issue.

Thanks in advance,

Pavel.