GEVP eigenvectors with errors about pyerrors HOT 7 CLOSED

Hi @JuanUrrea6, thanks for opening this issue. Do you have a proposal for improving the code and including the feature you are mentioning? While I don't have a use case myself, I can see that this is useful and I am happy to review any pull request that adds this functionality.

from pyerrors.

I can provide the lines of code which given a Corr object calculate the GEVP eigenvectors and eigenvalues with their corresponding errors via an intermediate Cholesky decomposition. Nonetheless these eigenpairs are sorted according to the magnitude of the eigenvalue in ascending order as per the eigh() default since I have not implemented the determinant-based sorting and until now just compared the eigenvectors that I get with the sorted ones that the GEVP() function produces to check consistency.

from pyerrors.

I am a bit unsure about how to include this in the code. One way could be to add an additional keyword argument to the CorrGEVPmethod which allows to extract the eigenvectors asObs` which (at least for now) would only work with the standard sorting method. Does this sound sensible? Maybe also @s-kuberski or @JanNeuendorf have opinions on this matter?

from pyerrors.

Hi. I agree that adding this is a good idea and that a keyword for the GEVP function would be a nice way to include this. I think I did have an implementation of this in the past, but this was not related to the Corr module.

Concerning the sorting algorithm: I don't think that you have to include error propagation in the sorting algorithms as we, in contrast to other resampling methods, only have one solution for the GEVP at each time slice. Therefore, you could just go ahead with the sorting according to the mean values and use it to sort your Obs-values eigenvectors. I think that this could easily be implemented by checking the data type at the start of the sorting and, if Obs are present, use vectors of floats to sort the vectors of Obs. Or am I missing something here?

from pyerrors.

I agree that the sorting could just be done on float versions of the vectors. Would you be willing to have a look at this and open a pull request @JuanUrrea6 ?

from pyerrors.

My only argument in favor of including errors in the sorting is that the determinants for different orderings have to be compared and while this is straightforward for the values of these determinants alone, the inclusion of the corresponding errors can change the outcome if one defines a significance criterion for the comparison in terms of these errors.

I would be happy to help, but due to current commitments I cannot say for sure when I would manage to have ready a version that uses this error-inclusive sorting and I don't know if that would interfere with your schedule of future versions. If this is a modification that you would like to have rather soon then I can gladly provide the code for the calculation of the vectors with errors as I currently have it on top of which one could implement the sorting.

from pyerrors.

At least from my side, this is not an urgent issue so for me it would be fine to push this back until either of us finds time to take a closer look.

from pyerrors.