David A. Case a écrit :
> On Fri, Feb 02, 2007, Francois Theillet wrote:
>
>> I have a problem with the treatment of the iRED method in ptraj (AMBER 8):
>>
>> I have a system of n vectors (i.e. a nxn covariance matrix), and I
>> obtain only n-1 eigenvectors (and eigenvalues) instead of the hoped n.
>> The sum of the eigenvalues is not exactly equal to n, but around
>> n-lambda(n-1) (with lamda(n-1) the associated eigenvalue of the n-1th
>> eigenvector).
>> (I tried several n, the result is identical. If I precise that I want k
>> eigenvectors (k<n), I obtain the first k eigenvectors correctly. With
>> k>n, I obtain n-1 eigenvectors)
>>
>
> This question (I think) has come up before:
>
> http://amber.ch.ic.ac.uk/archive/200501/0154.html
>
> The ptraj code uses an arpack routine to compute eigenvalues/vectors, which
> seems to be limited to n-1 results. For PCA analysis, this is not really a
> limitation; I'm not sure whether the same is true for iRED or not. Please let
> us know if you really need that last eigenvector: we can put it on a list of
> things to look at in the next round of ptraj updates.
>
> ...thanks for the report....dac
>
>
(Sorry I did not find this archived-mail that you mentionned before my
first mail..)
In this iRED case, if the ptraj code only truncates the eigenvectors
list, and gives the n-1 firsts of them in term of eigenvalue's
amplitude, it's unfortunate, but we can work without the last eigenvectors.
If it recomposes the n-1st with the last (I do not know how it could be
possible and if it is possible, it is just an idea), in my opinion it
could be more unpleasant, if the difference between the eigenvalues is
not more than an order of magnitude and if n is small.
I remember that ptraj can give the iRED-matrix, and it might be possible
to find a tool for diagonalization of a small matrix, and verify the
ptraj's result.
I will try. (unless you say that I think wrongly on this subject...)
Thank you for your usual goodwill
François Theillet
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Feb 04 2007 - 06:08:02 PST
