Sum covmats from various point prescriptions

[RoyStegeman]

We now have various sources of uncertainty as the default option (IHOU DIS, IHOU DGLAP, MHOU 7pt) and various other sources of uncertainty that we regularly use. This number will only grow as we use the theory covariance method for the determination/implementation of higher twist, alphas, mtop, etc.

To avoid an exponentially growing pointprescriptions.yaml as a result of having to implement a new point prescription for each combination of uncertainties that we sum, this PR enables to put a list of point_prescriptions in the n3fit runcard instead of a single point_prescription. These point prescriptions each have their own associated covmat where then the sum of them is stored.

We also store the individual covmats because the theory covariance method involves recomputing the theory covmat of the relevant parameter, so it's a good consistency check to see if the same is reproduced e.g. in case a different commit is used between fit and post-fit analysis using the theory covariance method.

As a proof of concept here is a N3LO QED+MHOU fit report comparing the current master branch to this branch: https://vp.nnpdf.science/B8uVm8CQQPK1LDNJoKdvJg==/

To see that the covmats are the same for both fits, you can use the following script

Code snippet

import numpy as np
from validphys.api import API
import pandas as pd

one_covmat_fit = API.fit(fit="241104-02-rs-n3lo-mhou-qed") # done with current master
multiple_covmat_fit = API.fit(fit="241104-01-rs-n3lo-mhou-qed-sumcovmat") # done with update_th_covmat branch

one_covmat = pd.read_csv(
  one_covmat_fit.path / "tables/datacuts_theory_theorycovmatconfig_theory_covmat_custom.csv",
  index_col=[0, 1, 2],
  header=[0, 1, 2],
  sep="\t|,",
  engine="python",
).fillna(0)

summed_covmat = pd.read_csv(
  multiple_covmat_fit.path / "tables/datacuts_theory_theorycovmatconfig_theory_covmat_custom.csv",
  index_col=[0, 1, 2],
  header=[0, 1, 2],
  sep="\t|,",
  engine="python",
).fillna(0)

print(np.allclose(one_covmat, summed_covmat, rtol=1e-2))

It does break backward compatibility and old runcards can not be used to run n3fit, though old fits can still be analyzed in the same way since the name of the theory covmat has not changed.

In particular the following changes impact how to write an n3fit runcard:

1. the point_prescription key is no longer supported and only point_prescriptions is, which takes a list of point prescriptions. Of course, that list can contain only a single point prescription.
2. theoryids are always read from scalevariationtheoryids.yaml so there is no need to explicitly declare it each time, i.e. no more

theoryids:
    from_: scale_variation_theories

3. some point prescriptions that are sums of various sources of uncertainty have been removed and are now only available as the sum of other point prescriptions, for a translation table from previously implemented point prescriptions in terms of the new see the bottom of pointprescriptions.yaml.

[giacomomagni]

Thanks @RoyStegeman, it looks much nicer now!

This prescription could also be removed as is the same as covmat_3pt, no?

nnpdf/validphys2/src/validphys/theorycovariance/construction.py

Lines 97 to 105 in 15d26d4

	def covmat_3rpt(name1, name2, deltas1, deltas2):
	"""Returns theory covariance sub-matrix for 3pt renormalisation
	scale variation *only*, given two dataset names and collections
	of scale variation shifts"""
	if name1 == name2:
	s = 0.5 * (np.outer(deltas1[0], deltas2[0]) + np.outer(deltas1[1], deltas2[1]))
	else:
	s = 0.25 * (np.outer((deltas1[0] + deltas1[1]), (deltas2[0] + deltas2[1])))
	return s

[scarlehoff]

Thanks @RoyStegeman I think this is really good and very much needed.

Since it will be easy to make mistakes here... could you add some tests (using the scale variations theories already used in the tests). Perhaps you can just use a few different versions of 3pt to make them run faster.

Checking things like ["3pt", "7pt"] produces the same result as ["7pt", "3pt"] for instance.

[scarlehoff]

Other than the tests this looks good. I'm submitting the bot just in case there is an expected change in behaviour somewhere.

It does break backward compatibility and old runcards can not be used to run n3fit, though old fits can still be analyzed in the same way since the name of the theory covmat has not changed.

This is important, would it be possible to catch early the error in the old fit and print a message saying what to change? Since some students may want to reproduce eg NNPDF40_nnlo_as_mhou and will notice it doesn't work and stop there. We don't need to catch every possible variation, but it should be able to tell you what to do for the few fits we have released.

[github-actions]

Greetings from your nice fit 🤖 !
I have good news for you, I just finished my tasks:

• Fit Name: NNBOT-e8f663353-2024-11-16
• Fit Report wrt master: https://vp.nnpdf.science/BkyK-b9JT_-rMhmCMmHg3g==
• Fit Report wrt latest stable reference: https://vp.nnpdf.science/Nbc5lBBVRdu7J3BwLxf0_A==
• Fit Data: https://data.nnpdf.science/fits/NNBOT-e8f663353-2024-11-16.tar.gz

Check the report carefully, and please buy me a ☕ , or better, a GPU 😉!

[RoyStegeman]

This is important, would it be possible to catch early the error in the old fit and print a message saying what to change?

Yes I wasn't sure it would be necessary, but it's also quick to add so I'll do that

[RoyStegeman]

@scarlehoff I added some tests and the error if people use an old runcard, if you're happy I think this can be merged