Simultaneous fit of multiple 1D projections

[MarcelHoh]

Dear experts,

In certain cases it can be useful to simultaneously fit several 1D projections of a data sample, to avoid complicated shapes in N dimensions or issues with low sample sizes. As far as I can tell this isn't supported in pyhf (or histfactory). Are there any plans to provide future support in pyhf?

For an example see https://arxiv.org/pdf/1809.03290.pdf, which simultaneously fits 4 projections.

Cheers,
Marcel

[matthewfeickert]

From https://arxiv.org/abs/1809.03290:

We use a set of one-dimensional projections of $w$, $\cos \theta_{\ell}$, $\cos \theta_{v}$, and $\chi$.

@MarcelHoh I didn't read the paper in full, but just jumping to that part and then looking at the plots of Figure 5, e.g.

, I would think that you would just implement these as each projection being its own HistFactory "channel" as you still have the different processes contributing as "samples". You of course need to think carefully about which modifiers are appropriate, but you need to do this regardless of what analysis you're doing.

As far as I can tell this isn't supported in pyhf (or histfactory).

I'm responding quickly between meetings, so can you elaborate on this?

[alexander-held]

I assume the projections re-use the same data, so the model would need to model the statistical correlation (which the HistFactory model does not support).

[matthewfeickert]

so the model would need to model the statistical correlation (which the HistFactory model does not support).

Ah yes, correct. My bad.

@MarcelHoh As @alexander-held pointed out that is indeed not supported and would need to be in some HistFactory vX for X $>1$ release in the future.

[MarcelHoh]

Thanks for the quick reply! Yes, the projections re-use the same data so the statistical correlations in the data and MC sample would need to be taken into account.