Redundancies in Corr class about pyerrors HOT 6 CLOSED

1. It is true, that projected can do everything smearing does. But it is a bit awkward to use projected to get a single sub-correlator. I would use it in there for debugging and i plan to add an option to not just pick a single smearing but a sub-matrix
2. This is true, i left it in there for testing only. It can go, if you want. But i think it is nice to have the simple function for a quick comparison.
3. smearing symmetric can (and should) be used whenever a correlator is symmetric under flipping the source and sink. I just used it in the Eigenvalue method, because the cholesky decomposition breaks with non-symmetric matrices.
4. Yes, sum was just there for a test. It should be deleted.

from pyerrors.

Thanks for your assessment.

• I removed the sum method in commit 07ca32f as you suggested.
• I agree that we should keep smearing_symmetric. I am not 100% happy with the name smearing as the matrices used for the GEVP do not necessarily only contain different smearing levels but could contain other correlation functions of operators coupling to the same quantum numbers. Maybe we can discuss this in our next call.
• For both smearing and Eigenvalue I see two options:
  • We could keep them in the namespace of the Corr class. I would then suggest that smearing internally calls projected with the appropriate arguments to improve future maintenance. Eigenavalue is still lacking a docstring.
  • Alternatively we could remove the two functions from the module, move them to the tests and use them to independently verify their counterparts.

from pyerrors.

Okay, regarding the Eigenvalue Method. I do not know, what your needs are exactly. But it is relatively easy to get the eigenvalues back from the eigenvectors.

With the norm:

leads to

Or in code-form C.projected(C.GEVP()) gives the eigenvalue.
This is why i recently changed the norm in projected() .
There is a good reason to do it this way instead of using the eigenvalues directly.
If you (for example) use the effective mass of the (direct-)eigenvalue to get a ground state,
you might overestimate the error, because the error also includes a measure of how well the GEVP was solved.
That is something, we do not really care about. If i give you any vector, you should be able to project it and get a plateau.
If i gave you a "bad" vector, the plateau will be shorter and or noisier. This should encode all the loss in statistics.
We could of course introduce an extra method which does C.projected(C.GEVP()) for you or the GEVP() method could include the option to return the eigenvalues. I do not really think, this is needed.

from pyerrors.

That sounds very reasonable. So you would propose to change the Eigenvalue method to internally call something like C.projected(C.GEVP())? Maybe @s-kuberski can assess whether that would work also for the applications he has in mind.

from pyerrors.

Yes, i can not come up with an example, where this would not work.

from pyerrors.

I'll try to check this...

from pyerrors.