Dear Peter,
Please see below:
On 23 February 2011 19:33, Neidig, Klaus-Peter
<Klaus-Peter.Neidig@xxxxxxxxxxxxxxxxx> wrote:
Dear Edward,
in short:
the first 2 fit options (by fit, by weighted fit) both use the covariance
technique. The difference is that
the first option applies an additional internal scaling by: final
chi/number of degrees of freedom
This means that when doing modelling of (NOE, T1, T2) with (M1-M5, TM1-TM5)
it cannot be used if these
triples of number are only available at one field and the model contains 3
parameters.
Could such details be given to the user? Many will know about the
covariance matrix error estimation method (those that don't really
should).
But in
in that area the PDC user does not have a selection of fit methods anyway,
we internally use a combination
of simplex and marquardt. But this is another area and should be done by
relax anyway.
For model-free, simplex is better than Marquardt. The problem with
Levenberg-Marquardt is in the fine print. This method works well for
most cases. But there is one situation where it totally fails, the
fine print says that is when the internally used matrix is singular.
At that point it cannot be inverted and the algorithm is forced to
terminate. Unfortunately this happens a lot in model-free analysis,
and it happens when a parameter is undefined. For example when S2 is
perfectly 1 or 0. Then te can be anything - it falls out of the
equations. This is the reason why there is much published data with
order parameters of exactly 1. It is not because the residue is
rigid, but because the optimisation came to a screaming halt.
I pick up your statements about the confidence stuff. Indeed you are right
that everything would be
much simpler if the PDC did just output Rx instead of Tx and corresponding
error. Whether keep the
confidence stuff at all is then a question, in the mean I think I follow
your opinion and would rather
give it up. But this I have to discuss internally here. It might be that
I'm not allowed to just give it up
then I would have to increase the output with more columns or a further
section which you could read
instead of the current section.
I think the confidence limit values have worth, some users will
appreciated it. I would simply consider adding the additional
information so that the user need not perform the conversion
themselves. It would simply be an augmentation of the PDC output.
There will be absolutely no new information in it as it is already in
the file.
I also have to say that my developments on the PDC must end by April (ENC)
and this includes
updating the manul, release procedure and some internal developments. So
there is not too much
time left. The PDC release number will then go up to 1.2. After April I can
still maintenance work
but no major changes, at least until someone decides something else.
The simple solution would be some quick code adding a few columns.
Additional columns should be no problem for parsers, as long as the
original columns remain in place. The key for downstream analysis
would be the R(1/2) value and standard deviation columns that the user
can directly feed into the other software.
Cheers,
Edward
Re: Addition of PDC support to relax 1.3.10 and an inconsistency in the error calculation.