Add operation to partially convolve a grid

[felixhekhorn]

Is it possible in v1 to turn SIDIS grids to DIS grids XOR FF grids? i.e. can I close some convolution index, but still keep others open and dump this back into a grid file?

(Maybe this is said already somewhere?)

[cschwan]

You cannot change the number of convolutions of a grid, but you can change the convolution types. What's your use case?

[felixhekhorn]

You cannot change the number of convolutions of a grid, but you can change the convolution types.

Mmm that's not very useful I think ... unless it was mislabelled there is no point in changing as different types represent different physics

What's your use case?

turn a SIDIS grid into a DIS, specifically find an argument to convince these people to add a PineAPPL interface ... and maybe investigate the impact in a PDF fit

I can see this might be technically difficult and the way out would be to have a fixed distribution laying around that you have to call the moment you're eventually can close all indices

[cschwan]

I think I understand now. We can possibly extend Grid::evolve to return a Grid and let us select which convolutions we're are performing.

[felixhekhorn]

That is precisely what I mean. Just to be clear: evolve already returns a grid as FK tables are special grids

[cschwan]

Yes, but FK-tables are even more special than the grids we'd return and FK-tables don't change the number of convolutions (which I forgot previously). We probably need a new operation, which we could call partial_convolve. This operation should

• implement the partial convolution with exactly one convolution function (that's the most important case and the general case can be constructed from it) and
• return a new Grid, with one less interpolation, one less x and possibly one less Q2 parameter and so on.
• If we start with a grid that has just a single convolution, this operation must signal an error.

@t7phy: I think we talked about exactly this when we implemented the new file format, do remember anything from our discussion?

[t7phy]

@cschwan tbh I don't remember a discussion for this particular situation. the most similar discussion I can remember had to do with convolution with EKOs, wherein if for example, we have a SIDIS grid, then do we:

• first convolve the grid with space-like eko and produce an intermediate half-grid-half-fktable and then further convolve this with time-like eko? or
• do we pass both the ekos at one go to go directly from the Grid to the Fktable?

[cschwan]

@t7phy thank you! So the discussion was somewhat related, but not exactly. In any case, I think this operation partial_convolve is worthwhile to have, and shouldn't be too complicated to implement. We basically need loop over all subgrids and create a new subgrid with lower dimension and multiplied with the right convolution function values.

[t7phy]

yes indeed it would be a nice feature to have. however, I would suggest to perhaps call it something other then partial_convolve.
the reason being that convolve leads to a convolution between a grid and a PDF/FF on the fly, and therefore one could easily mistake partial_convolve as doing something similar, whereas what it is really doing is hardcoding a given PDF/FF for a given convolution. Perhaps something like pineappl fix_convolution x{1,2,3} <PDFname> or pineappl set_convolution x{1,2,3} <PDFname> ?

[cschwan]

That's a good point. I like fix_convolution, although I'd prefer a shorter name.

[t7phy]

well in that case maybe fix_convolution shouldn't be added as a separate option but rather a flag of write (as there are already a number of flags with long names), so we'd have something like pineappl write --fix-convolution x{1,2,3} <PDFset> <input_grid> <output_grid>.
given that write already has many flags that modify the grids in non-trivial ways and write new grids, that would be the optimum place to add this feature imo.

[felixhekhorn]

just to say, in my opinion --fix-convolution='1<PDFSET' (or something similar) would be a more flexible structure