Positivity checks

[juanrojochacon]

Once the PR https://github.com/NNPDF/nnpdf/pull/2370 has been merged, we can use it as starting point to understand the impact of positivity. While we discuss the ideal positivity, with the run cards we have we can already run a few interesting tests

• [ ] Remove completely all positivity constraints, both PDFs and cross-sections

• [ ] Remove FL positity at small-x, namely replace

• dataset: NNPDF_POS_2P24GEV_FLL rule: x > 5.0e-7

by

• dataset: NNPDF_POS_2P24GEV_FLL rule: x > 0.01

in the run card, everything else kept the same

• [ ] Remove only the cross-section positivity observables

• [ ] Remove PDF positivity constraints for x > 0.28 (since for larger x one has W^2 > 12.5 GeV2)

Let me know if anything is unclear @jekoorn @kamillaurent and we can discuss in this issue.

[jekoorn]

I just want to be sure: for removing all positivity constraints, one can just set the positivity multiplier to zero so that there is no loss penalty added in any case, correct?

[juanrojochacon]

I think so, or even better just remove the positivity dataset in question from the runcard

[RoyStegeman]

To run without any positivity you can pass an empty list posdatasets: []

There is a check that expects a postdatasets for all fits, so removing it entirely will raise an error.

[juanrojochacon]

Thanks @RoyStegeman for the suggestion. So yes, we add an empty list for positivity observables, this is the neat way to go

[kamillaurent]

Hi @scarlehoff, I compared the positivity test fits to a fit I named 280925-nnpdf41, which has the same runcard as example-nnpdf41. I now uploaded the 280925-nnpdf41 fit using vp-upload and added the name of this fit into this table (https://www.wiki.ed.ac.uk/spaces/nnpdfwiki/pages/652698412/Recent+versions+of+fit+variants). Is this a fair comparison or I need to make other changes?

[scarlehoff]

Hi @kamillaurent, the name should be, in any case 250928-kl-nnpdf41.

The reason to put first the year is to make sure that when you search with ls things are automatically order by year. Also, adding your name helps disambiguate.

That said, the point of that table is not to need to run a new reference fit each time. There's no point on having two fits which are exactly the same. If the fit needs to be updated, the row for 250925-jcm-001 needs to be changed. If it doesn't, just use 250925-jcm-001.