Add positivity support#132
Conversation
|
Just a spare comment: I have no better proposal, but personally I find |
Agreed. If you have a better suggestion, I'd be happy to adopt it! |
alecandido
left a comment
alecandido
left a comment
There was a problem hiding this comment.
I'd only split the run function, everything else is perfect.
I still have to run and check the result, but (just reading) it looks alright.
Do we need this name tag? it was not present in the old name and I wonder how quickly it will change (the xgrid isn't part of the name either and is at least equally important ...) |
|
@felixhekhorn good point. Do you know where the old positivity grids are and how they are named? Maybe we can add the NNPDF version, when they were first added, to the positivity dataset name. So something like |
|
Another point: does the |
I believe it's not required: indeed Tommaso is about to propose to raise a bit the scale at which positivity is imposed. |
|
I see your point, but I've no good suggestion ... |
We could add some metadata that specifies the 'positivity scale`. The fit can then check for potential pitfalls. |
Most Valuable Positivity
Why is I would not add the value of
exactly |
Actually, grid level this information is present (because is the scale contained in the subgrid). We just need to make sure the information is propagated down to FkTables, where it would be flushed by evolution. |
we can just add this here: https://github.com/NNPDF/runcards/blob/ecd40ef995b3f2f06257a448a4f5ec7fafe21539/runcardsrunner/external/positivity.py#L73 |
For the easiness of parsing, I would add a separate key with a single float. |
This scale is implicitly contained in |
|
... on the other hand it will be contained in the |
|
Commit 2b9de22 writes the complete runcard as a JSON string into the metadata just like yadism, and therefore automatically contains the positivity scale |
|
What's finally left is to implement all the positivity datasets themselves, so
How can I find out what the other datasets are? What are the |
While I understand that we may want to change the |
What I meant was the |
|
@enocera where can I get the positivity grids for each of the flavours? Do we have them in a form before evolution? |
|
However, consider that for 4.0 replacement we want to force whole PDFs positivity, but later we might want to impose positivity only at large x. For About FTDY of course we can do nothing, and as for all the other FTDY observables, we'll keep using old FkTables, up to FTDY provider replacement. |
|
By the way, how does positivity work when scales are varied? Are they also varied in the positivity grids? Would that make sense? |
|
Just for the record: |
NNPDF never fits scale varied theories, so they are only used to construct the theory covmat (and that one has no contribution from pseudo-observables, I guess). |
Do you mean FK tables? These are available in each theory along with all other FK tables. |
I believe @cschwan really meant grids in the sense of PineAPPL grids (or in this case APPLgrids). The problem is that they were all generated with APFEL, meaning that they were never dumped in the middle, evolution was applied online. |
|
Note to myself: the theories are stored on the CERN server (look at the NNPDF wiki, 'storage servers'), at |
|
Actually, I'd say that Is it the same? |
|
@cschwan @alecandido @felixhekhorn I would like to send a fit with pineko-positivities tomorrow so that it is ready for Amsterdam. Are they ready? (if not I'll do a fit with the old positivities) |
If you want to have PDF positivity, you just need to plug yourself the For everything else, the answer is just no. P.S.: consider that half of the observables used are FTDY, so the ability of completely generating new positivity FkTables is strictly dependent on the implementation of a FTDY provider |
|
Ok, then I'll do a fit with the old positiv |
|
@scarlehoff I also have to test this, which means comparing the old FK tables with the grids generated using this generator, but for that I need NNPDF/pineappl#70, and that'll take a while ... |
|
Using the FK table importer from NNPDF/pineappl#70 it seems that the generator developed in this branch works: The first column in diff is the value of the PineAPPL grid, the second the converted FK table |
|
Here the remaining comparisons:
|
|
Commit 7d0bf2f adds all positivity runcards that I can easily generate. The only possible issue (@scarlehoff ?) left is that the
|
5 participants
Merging this PR will close #124.
Please have a look at this, that's the first Python code other than matplotlib scripts I've written in a while, every suggestion is welcome. At this point it's sort of a MVP, but a few things have to be done still: scale variation, metadata, ?. The x-grid must be specified in the runcard, let me know if that's a good idea.