Updated positivity in n3fit fits

[juanrojochacon]

Hi @scarlehoff @scarrazza one the things we agreed to do in Varenna was to verify that the n3fit code was able to properly include the new positivity observables, namely those that were not part of NNPDF3.1 (basically because we were not able to fit them successfully)!

In principle the data files and FK tables are ready, it is just a matter of verifying that the code can deal with them (hopefully, else we are a bit stuck). The new positivity files are

DATA_POSDYC.dat
DATA_POSDYCBSB.dat
DATA_POSDYCSB.dat
DATA_POSDYUBD.dat
DATA_POSDYUD.dat
DATA_POSDYCBD.dat
DATA_POSDYCD.dat
DATA_POSDYUBDB.dat
DATA_POSDYUDB.dat
DATA_POSDYCBDB.dat
DATA_POSDYCDB.dat
DATA_POSDYUBS.dat
DATA_POSDYUS.dat
DATA_POSDYCBS.dat
DATA_POSDYCS.dat
DATA_POSDYUBSB.dat
DATA_POSDYUSB.dat

Furthermore, to avoid surprises at a later stage it would be great if someone, perhaps @felixhekhorn could check the calculation of the positivity FK tables, using some standalone code. We have found issues with the positivity observables a lot of times in the past so better safe than sorry.

I would really like to verify that n3fit can deal with the updated positivity constraints before it is too late. I am more than happy to assist with this!

[scarlehoff]

Not sure what "deal with them" actually means,

if it is just from a technical standpoint, I just tried running a 1-replica DIS fit with them and it ran ok (and the positivity did pas)

Do we want a more specific check? (and if so, what do we want this check to be exactly?) I can also leave a global running and then do the typical vp report?

[juanrojochacon]

Well, I would like to do a NNPDF3.1-like global fit with the new positivity, and check that it is satisfied for the majority of the replicas. I would also like to compare these results with the same fit but with the old positivity, to have an idea of what is their physical impact on the PDFs.

So the check is to run NNPDF3.1 global fits with the old and new positivity, and verify that the fraction of replicas that satisfy the post-selection pruning is comparable.

Note that here doing a global fit is important since the positivity tension is related to some of the LHC experiments, so it is harder than in the DIS-only case.

Let me know @scarlehoff if something is not clear.

@scarrazza do you agree with the plan? It would be really nice if this works since we have been struggling with these new positivity observables for a long time now.

[Zaharid]

@juanrojochacon do we have listed somewhere what the issues with the positivity sets were? I don't quite remember what it was and if it was solved well.

[scarlehoff]

Well, I would like to do a NNPDF3.1-like global fit with the new positivity, and check that it is satisfied for the majority of the replicas. I would also like to compare these results with the same fit but with the old positivity, to have an idea of what is their physical impact on the PDFs.

Ok, perfect. I'll run so. It should be doable.

@Zaharid if I run vp-comparefit with positivity sets that were not included in the fit, will it test them as well?

[juanrojochacon]

no, as far as I can remember (but you were actually the last in trying!!) we did not manage to get a chi2 with the new positivity which was decent, even forgetting about the loss of replicas via postselection. I can try to fish the reports that I think you produced, these are of the pre-vp2 times!

[Zaharid]

On Fri, Oct 11, 2019 at 4:23 PM Juacrumar <notifications@github.com> wrote

@Zaharid <https://github.com/Zaharid> if I run vp-comparefit with positivity sets that were not included in the fit, will it test them as well?

No, it won't. Why would it? It is however easy enough to make your own comparisons. See $ validphys --help plot_positivit

…

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#593?email_source=notifications&email_token=ABLJWUWM6M5RQ2EMQAXL67LQOCK5PA5CNFSM4I7ZMTW2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBAKYUQ#issuecomment-541109330>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABLJWUTCCZEPZWQWDM7HMNDQOCK5PANCNFSM4I7ZMTWQ> .

[juanrojochacon]

Well, I would like to do a NNPDF3.1-like global fit with the new positivity, and check that it is satisfied for the majority of the replicas. I would also like to compare these results with the same fit but with the old positivity, to have an idea of what is their physical impact on the PDFs.

Ok, perfect. I'll run so. It should be doable.

@Zaharid if I run vp-comparefit with positivity sets that were not included in the fit, will it test them as well?

Super, very nice, thanks @scarlehoff for taking care of this! And indeed it would be useful to test the new positivity with the current fits, to see how bad the problem is to begin with

[Zaharid]

On Fri, Oct 11, 2019 at 4:23 PM Juan Rojo ***@***.***> wrote: no, as far as I can remember (but you were actually the last in trying!!) we did not manage to get a chi2 with the new positivity which was decent, even forgetting about the loss of replicas via postselection. I can try to fish the reports that I think you produced, these are of the pre-vp2 times!

So there are two things: whether we could fit the positivity sets and whether the predictions made any sense (which is a prerequisite). I seem to recall that there were some problems with the predictions (something that should be symmetrical wasn't, or the other way around), which were discovered after I tried...

…

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#593?email_source=notifications&email_token=ABLJWUXGDXIQGBJLFBOJEFDQOCLALA5CNFSM4I7ZMTW2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBAK2ZQ#issuecomment-541109606>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABLJWUWRG3BQGV33HWZ7P4TQOCLALANCNFSM4I7ZMTWQ> .

[juanrojochacon]

Yes, this was fixed. The Drell-Yan positivity implementation in APFEL was bugged, but hopefully Valerio fixed it some time ago. This said it would be important if someone can cross-check it, this will save time later.

You can find our failed tests here

https://www.wiki.ed.ac.uk/display/nnpdfwiki/Towards+NNPDF3.1+fits

long ago when we still had a trunk ;) Some of the fits with new positivity and poor chi2 are here
https://www.wiki.ed.ac.uk/display/nnpdfwiki/Towards+NNPDF3.1+fits?preview=/264176483/294628460/151120-1435ca6-jr-004.pdf

[juanrojochacon]

This is why I would like to assign the task of checking that what now APFEL returns for the positivity cross-section is sensible. It is very easy, one can compare with the LO calculation that can literally be done on the back of an envelope ...

[scarlehoff]

No, it won't. Why would it? It is however easy enough to make your own comparisons. See $ validphys --help plot_positivit

I was wondering whether it would take the union of the sets of the two fits being compared.

[Zaharid]

It does the intersection as per {@with matched_positivity_from_dataspecs::dataspecs@}. Intersection is a safer default in that it is not a given that e.g. you'll have fktables for all the theories.