Change fSys datastructure

[scarrazza]

Issue by Zaharid
Sunday Jul 02, 2017 at 10:18 GMT
Originally opened as https://github.com/NNPDF/libnnpdf/issues/25

---

Experiment should get a packed array (as in vector<double> buffer of unique_ptr<double*> buffer) of correlates systematics and another packed array of uncorrelated ones at construction time. We shouldn't have to dereference a deeply nested structure in a tight loop to query basic information, which is furthermore duplicated since the systematic types are the same for each point.

[Zaharid]

I am trying to remember what exactly I wanted here and if it made sense at all.

[nhartland]

I think I know how I would do it, although it's a bit awkward to communicate. The main issue is that as we need to completely restructure fSys, the artificial replica generation is going to have to change. My feeling is that we're probably better served by first changing that to using a sampling covariance matrix, which kills two birds with one stone:

1. We want that anyway, as vp-setupfit will generate the sampling matrix
2. It decouples the artificial data generation from fSys

[Zaharid]

Note that in somewhere there has to be ComputeCovMat or something so the matrix can ve computed in the first place.

[nhartland]

Right, and the problem there is it's unclear what to do with the special case systypes (e.g THEORYUNCORR). #186 states that these should be dealt with along with scale variation errors in vp2, but I guess we aren't there yet.

[nhartland]

What I can try and do is create a ComputeCovMat with the following signature

matrix<double> ComputeCovMat(vector<CommonData> cd, vector<double*> t0, bool use_th_errors = false);

With the final bool being dropped in the final implementation of #186.

[nhartland]

That should be enough to specify the CovMat

[Zaharid]

I am not so sure of the bool thing. What happens if we omit it and it acts as if it was true?
Then at some point we would remove the THEORYUNCORR things and put them in the th covmat.

[nhartland]

We need it for the sampling matrix, at the moment the THEORY* systematics are excluded from it.