More PASTd

Hi

For those who care:

My PASTd is finally working in MATLAB. Just two observations:

The algorithm seems to get certain random eigenvectors negative, as compared
to those produced by MATLAB's eig function. I don't know if this is a
general property of eigendecomposition (i.e. decompositions may not be
unique and eigenvectors can differ by sign)....?

The eigenvalues need to be scaled by (1-Beta)/Beta.

I'm not sure if these two properties are mentioned in the original (Yang)
paper on PASTd, since I don't have the paper. Maybe someone knows of an
online copy?

Rory

"Rory" <rory@invalid.com> wrote in message news:<3fb49aa5$0$91694$a32e20b9@news.nntpservers.com>...
> Hi
> 
> For those who care:
> 
> My PASTd is finally working in MATLAB. Just two observations:
> 
> The algorithm seems to get certain random eigenvectors negative, as compared
> to those produced by MATLAB's eig function. I don't know if this is a
> general property of eigendecomposition (i.e. decompositions may not be
> unique and eigenvectors can differ by sign)....?

Off the top of my head, I believe the eigen vectors generally describes
a line in N-dimensional hyper space where the eigen vector is found. 
I can't recall to have seen that the direction plays a part. 

So if v is an eigen vector, -v should be one as well. 

> The eigenvalues need to be scaled by (1-Beta)/Beta.
> 
> I'm not sure if these two properties are mentioned in the original (Yang)
> paper on PASTd, since I don't have the paper. Maybe someone knows of an
> online copy?

You can get it from ieeexplore.ieee.org. If you are at a university, 
chances are that you have access from there. If you are in industry, 
some of your colleagues may have access and can download it for you.

Rune