Rich Ulrich wrote:
> 
> On Thu, 09 Jan 2003 18:10:20 +0000, Jonathan Bailleul
> <[EMAIL PROTECTED]> wrote:
> 
> > Hello,
> >
> > I have low background in statistics, and I *just* want to find a piece
> > of C/C++ code that would compute eigenvalues and eigenvectors of
> > training data (preferently), or of a covariance matrix I might compute
> > myself form the data (hence square and symetric).
> >
> > I managed to find from Statlib a program under the form I wanted
> > (http://lib.stat.cmu.edu/multi/pca.c), but it unfortunately showed
> > "strange" results, different from those given by matlab and another stat
> > package (signs of eigenvector coordinates changed, all eigenvalues
> > multiplied by same random factor). Maybe I'm wrong, but it seems like
> > "it doesn't work".
> 
> I bet you are wrong.
>  - it sounds to me like you need to read the documentation.
> 
> Signs are changed?
> Whenever A  is important statistically, and
> A  is an arbitrary (positive or negative) number,
> (-A) has most of the same properties.

I agree, but that wasn't the main purpose of my request. And since I
have absolutely no idea of the properties differences that justify your
use of "most of", maybe that question was not so trivial (can you give
details?)

> Random multiplying factors? - read up on "eigenvalue."

I already did it, but found nothing helping the current problem (just a
basic description and nothing about the way it is computed).

It looks like the solution seems obvious from your point of view: can
you explain further? Please, remember I'm not a statistician.

In advance, thank you for your help.

-- 
--------------------------
Jonathan BAILLEUL
Doctorant au GREYC Image
ISMRA, Universit� de Caen
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to