TDecompSVD

ROOT
1

Hi,

we are using root version 4.00-02 on a Linux 2.4.20-31.9 (Red Hat 9).

When using the TDecompSVD class for singular value decomposition, we see the following error message for some matrices we apply the SVD to:
“Error in : no convergence after 161 steps” and 161 goes up to 175.

What does this imply?

  • Are we doing anything wrongly?
  • Do we apply the SVD to matrices outside the limitations of the SVD (If yes, which features are the matrices not supposed to have?)?
  • Is it a known problem in this version and no longer existing in newer versions?
  • Anything other?

Any help will be appreciated.
Thanks,
Kerstin

2

Eddy Offermann wil answer your questions once he will be back (July 7).

Eddy made several changes to TDecompXXX classes in the past 2 or 3 months. He is also working on the documentation of the linear algebra package.

Could you send the shorted possible test script showing your problem?

Rene

3

We located the problem somewhere else and it does not
occur anymore.

Thanks,
Kerstin

4

Hi Kerstin,

The diagonalization algorithm is an iterative process .
If it finds that the off-diagonal elements are still larger
than DBL_EPSILON after the maximum number of iterations
(a number that depends on the matrix size), it will complain .

It is hard to give specifications for the matrices that will
fail (one would hope if we knew them we would check
before proceeding :slight_smile: ) . Your matrix must have been
numerically very ill-conditioned .

The steps taken in the SVD algorithm have been described
in great detail in

Data Analysis, “Statistical and Computational Methods for Scientists and Engineers” , by Siegmund Brandt .

Eddy