luxqed recipe

[scarrazza]

In order to obtain a NNPDF fit using luxqed one needs the following:

Requirements

• provide QED evolution with eko.
• provide the photon PDF using fiatlux.
• provide F2, FL and F2LO contributions for a given input PDF set from yadism.

Algorithm (simplified version)

1. take an input QCD fit and generate the photon PDF using fiatlux, at large Q value.
2. down-evolve the photon obtained in 1 to Q=1GeV
3. use this photon as fixed parton in a new PDF fit, it should enters all FK tables
4. repeat 1 until the photon shape has converged.
5. care about the extra uncertainties identified by luxqed.

Missing software requirements

• fiatlux in python
• n3fit:
  • load the photon obtained in 2, e.g. using the T0 format, see here for loading and here for the convolution.
• validphys:
  • implement the T0/photon generation by integrating fiatlux/yadism, i.e. by following all the steps summarized here.

[alecandido]

I would say that things now have to move in parallel. We should use some pseudo-structure functions, and make it working also in the fit.

In the meanwhile, exact (EXA) QED evolution is working, but too slow. @niclaurenti is taking care of it.

Concerning yadism, we should just choose how many points we want, and prepare a runcard. Then we compute, and finally we plug interpolation.
2D spline interpolation is available also in scipy.interpolate, so we just need to obtain the numbers, and plug them in.

[scarlehoff]

For the n3fit part, it should be close to trivial, as the fit itself is completely agnostic to the number of flavours and it is the fktable who decides which flavours is convoluted with what (nothing to be done for the convolution for instance).

Currently the photon in the PDF is set to 0, this needs to be instead changed to a value-per-x

nnpdf/n3fit/src/n3fit/layers/rotations.py

Line 75 in 94b80c3

0 * x[0], # photon

Since the photon PDF is not going to be modified, this can be achieved by inputting the per-x value of the photon when FkRotation is instantiated:

nnpdf/n3fit/src/n3fit/model_gen.py

Line 561 in 94b80c3

layer_evln = FkRotation(input_shape=(last_layer_nodes,), output_dim=out)

Either in the same FKTRotation or in a new class that further digests that. The photon will enter as a constant to the fit so it will have no effect whatsoever in the fit (in other words, everything that needs to be computed should be done during "build" time).

Furthermore, the code that needs to be modified in n3fit should be contained to those two places (plus the whatever way in which the photon comes into the fit form the runcard, which can be by having a luxset: entry or whatever).

Also the MSR needs to be updated.

If you need help with that when you start with it let me know @niclaurenti

[felixhekhorn]

@niclaurenti I guess this can be closed?

[RoyStegeman]

Closed by #1643