flav-io/flavio

Questions regarding proper use of flavio's likelihoods (SPheno - flavio)

kys-sheng opened this issue · 2 comments

kys-sheng commented

Hi,
I have been trying to get a global likelihood based on all possible relevant observables with SPheno's wcxf as an input. In this case, I can only work with WET and I am unable to use smelli for this case since smelli works with SMEFT. In other words, Im trying to build a bridge between SPheno and flavio such that I can check the global likelihood of relevant observables with flavio using SPheno outputs.

I understand that extreme attention is needed to be given into choosing appropriate observables and experiments as well as using a consistent, reasonable and reliable fast likelihood definition. However, I must admit that I am not too familiar with likelihoods and nuisance parameters other than the basics. As such, I need some help with double checking what I am doing is correct or not.

Whatever follows are based on my understanding on materials I can find online and since I have not attend any flavio talks in person yet, it is highly likely that I might misunderstood the contents. Please correct me if any of the statements I made are erroneous or just plain wrong.

So far, I have noticed a few details that needs to be taken care of to construct a good FastLikelihood with wcxf from SPheno as an input:

  1. Do not include observables that is dependent on WC from SMEFT:

    • because SPHENO output is already in WET, and WET->SMEFT cant be done

  2. Use only relevant parameters for the observables:

    • Use get_dependent_parameters_sm to check which parameters are relevant

  3. Do not use overlapping experiments. Ref[1,2]

    • Cases with combinations of experiments should be given attention (choose the combination of experiments once or the experiments separately), as shown in Ref[2]

    • as mentioned in Ref[1]

      "... can lead to inconsistent results in several cases, e.g:
      -including multiple measurements that are not independent of each other"

    • Relevant observables with this scenario: <AFB>(B0->K*mumu), <FL>(B0->K*mumu), <dBR/dq2>(B0->K*mumu), BR(Bs->mumu), BR(B0->mumu), a_mu

  4. Must consider all other observables of an experiment result if it contains multiple observables, Ref [1,2]:

    • as shown in Ref [2]
    • most experiments (though not all) contains multiple observables (if any of it is left out, a warning will come out)

  5. *Check either:
    A. *Theory uncertainties are negligible (Case 1 of ref[3]):
    - I guess one way to do this is to compare np_uncertainty() vs get_1d_errors_random() for every possible observables?
    - Although I am not sure what is a good definition of negligible.

    or :

    B. *Theoretical uncertainties in the presence of new physics are more or less equal to the ones in the SM (FastLikelihood in APIdoc and Case 2 of ref[3]):
    - I am not sure how to check this consistently

  6. *Do not use experimental results with distribution of :

    • asymmetric_normal : (Because its asymmetric, not a good approximation for normal/gaussian)**
    • general_gamma_upper_limit : (Because its not normal/gaussian)
    • gamma_upper_limit : (Because its not normal/gaussian)

*I am highly doubtful of whether what I understand/what I am doing is correct and your input is highly appreciated and crucial.
**An alternative would be to use sigma^2 vai the alternative formulations from https://en.wikipedia.org/wiki/Split_normal_distribution

Additional questions:

  1. It is mentioned in ref[1]:

    "... can lead to inconsistent results in several cases, e.g: ...
    ...including constraints on parameters that come from measurements included in
    the likelihood (e.g. the default constraint on Vub that comes from B → π`ν),"

    Is there a consistent way of double checking this?

  2. Is there anything else that I have missed out that is very important before proceeding?

Many thanks in advanced and hope to hear from you soon!

Reference:
[1] flavio paper : https://arxiv.org/pdf/1810.08132.pdf
– including multiple measurements that are not independent of each other.
[2] Peter Stangl's tutorial : https://github.com/peterstangl/flavio-lecture/blob/master/4%20Likelihoods.ipynb
[3] Peter Stangl's talk : https://indico.cern.ch/event/1011800/contributions/4245824/attachments/2224349/3767163/Stangl_21-04_EFT_fits.pdf
[4] David M. Straub's talk : https://indico.cern.ch/event/787665/contributions/3374415/attachments/1861971/3060407/straub-smefttools-2019.pdf

kys-sheng commented

I have no idea how I'd missed this!
Thanks!