ATLAS_WCHARM_13TEV

[ecole41]

This pull request is for the implementation of the ATLAS_WCHARM_13TEV dataset

[ecole41]

@enocera Do we want to have all of the channels ((W−+D+), (W++D−), (W−+D∗+), (W++D∗−)) put together into one dataset like this?

[enocera]

@ecole41 Yes we want a single data set with all the channels. Two remarks.

1. One must be cautious w.r.t. the treatment of correlations: there may be uncertainties that are correlated across all the channels and uncertainties that are not. But this is something that you surely know better than me, because you've gone through the paper. Also, when definining the theory part of the metadata, one should have four entries, with four different names, one for each channel.
2. I realise that this measurement is only particle level, which may complicate the way in which we will have to make theoretical predictions, but this is something that we can discuss later.

[ecole41]

@enocera We are given the covariance matrices for W-/W+ and for W-( *)/W+(*). I am assuming that means there are no correlations between the excited (*) and non-excited channels.
I wanted to check what should be done to construct the systematics with this.

We are given the systematics for each point but I am not sure if these include the correlations from this covmat. The covariances matrices are also combined statistical and systematic uncertainty covariance matrices so does this mean that decomposing this covmat will give the final systematics with correlations accounted for?

[enocera]

Dear @ecole41 after looking into this data set a little, I suggest to implement two variants insofar as uncertainties are concerned. This means that you have to generate two uncertainties.yaml files, with different names of your choice (e.g. uncertainties.yaml and uncertainties_covariances.yaml).
1 - The first variant, associated to the uncertainties.yaml file, is the one in which you implement the breakdown of 280 uncertainties provided, for each bin, in the last column of Tables 19-22. These uncertainties, that have the same names across all channels, can be correlated all over the place.
2 - The second variant, associated to the uncertainties_covarainces.yaml file, is the one in which you implement the information that you've got from Tables 15-18. The way I'd do this is as follows: first construct a block-diagonal covariance matrix; then generate N_dat artificial systematics from this covariance matrix. The function that generates artificial systematics from a given covariance matrix is here.

[ecole41]

Thank you, that makes a lot of sense.

I just wanted to check the function used to symmetrise the errors. The function in: nnpdf_data/nnpdf_data/filter_utils/uncertainties.py, shows that the se_delta is equal to the average of the two errors and the se_sigma is related to their difference. Should this be the other way around?

[ecole41]

I also wanted to check how we should treat the se_delta and se_sigma values. Do we need add the se_delta onto the central data values and use these in the data.yaml file?

[enocera]

I just wanted to check the function used to symmetrise the errors. The function in: nnpdf_data/nnpdf_data/filter_utils/uncertainties.py, shows that the se_delta is equal to the average of the two errors and the se_sigma is related to their difference. Should this be the other way around?

You are totally right, that function looks wrong, it should be the other way around to be consistent with Eqs. (23)-(24) and (27) of https://arxiv.org/pdf/physics/0403086. We should check how many data sets have been affected by that typo. Thanks.

[enocera]

I also wanted to check how we should treat the se_delta and se_sigma values. Do we need add the se_delta onto the central data values and use these in the data.yaml file?

You have to compute se_delta for each data point, sum them up, shift the exp central value and dump that into the data.yaml file. And likewise for the uncertainty. We are essentially using Eqs.(27)-(28) of https://arxiv.org/pdf/physics/0403086. The function above computes se_delta and se_sigma for a single data point i, but then you have to sum over all points, manipulate central values and uncertainties, and finally write the manipulated central values and uncertainties in the corresponding .yaml files. Hope that this clarifies the issue.

[ecole41]

@enocera Should these four ((W−+D+), (W++D−), (W−+D∗+), (W++D∗−))be separated into separate observables or be kept as one?

[enocera]

These should be four different observables of the same data set, as, e.g., different differential distributions for top pair production.

[ecole41]

Ok, and should these all have a separate data, uncertainties and kinematics yaml files? So four of each?

[ecole41]

I'm not sure how to put this in Latex terms

[enocera]

Dear @ecole41 , in view of the upcoming Morimondo meeting, I have revised the implementation of this data set. After a more careful look at the paper, I came to the conclusion that the best way of proceeding is to implement only two observables: one for the decay of a W in a D meson, and one for the decay of a W in a Dstar meson. In both cases, the observable incorporates a W- and a W+. The observable is the absolute distribution differential in the pseudorapidity of the lepton. This choice greatly simplifies the implementation, because we only need to take into account the full covariance matrix (Table 12 of the paper), and decompose it into the artificial systematic uncertainties. Because the full covariance matrix is not available for normalised distributions, we simply dismiss them. We dismiss the distributions differential in pT because they are theoretically trickier than the distributions in eta. Because W- and W+ are correlated, we have to incorporate the two in the same observable, for each decay mode. The two decay modes, as per the paper, are instead largely uncorrelated, except for the luminosity uncertainty. For this reason, I first compute the luminosity covariance matrix, and I subtract it from the full covariance matrix provided by the experimentalists before generating the artificial systematic uncertainties. I then attach the luminosity uncertainty to the list of uncertainties as usual. This way the luminosity uncertainty can be correlated across data sets and experiments as it should. This way the implementation is significantly more compact. The tests now pass. We can discuss further in Morimondo. Thanks.

[scarlehoff]

@ecole41 could you generate also the (NLO) grids for this dataset?

[ecole41]

@ecole41 could you generate also the (NLO) grids for this dataset?

I am happy to give it a go. I am new to making grids without the pine cards, do we have a similar dataset already to use a base?

[enocera]

Dear @ecole41 unfortunately we don't have an example for W+charm. We used MCFM to compute applgrids back in the days, but I don't think that this is of any help now. As @scarlehoff suggests, we may want to use madgraph to generate the NLO PineAPPL grids. These will then be supplemented with K-factors to incorporate the NNLO correction. In this respect, I point you out to this paper https://arxiv.org/pdf/2510.24525, in which they perform a comparison between NNLO predictions and the very same data set in this PR (see in particular Sect. 3). Maybe you've got a chance to talk to René Poncelet who sits at Cavendish and ask him if they can provide you with any useful input to compute theoretical predictions?