[Rivet] [Professor] Tuning error-band (no systematics from run-combinations yet)

[Andy Buckley]

Holger Schulz wrote:
> Hi,
> 
> I was able to correct my mistake. So here are the plots
> I will show in Edinburgh:
> 
>     http://users.hepforge.org/~holsch/edinburgh_snippet.pdf
> 
> The first plot shows the prediction of Pythia6 tune 329 for 7TeV LHC.
> The others show how the tuning error-band shrinks if we add
> 100k and 1M of pseudo-data in the tuning.
> 
> The missing labels are    x: leading track pT
>                                           y: <N_ch> density in 
> transverse region
> 
> 
> Looks good, doesn't it?

Very much: nice one!

> Comments are welcome.

Okay, here are some ;)

  * It would also be nice to see how the error band responds to a less 
Tvt-consistent LHC pseudodata point, i.e. what happens if you generate 
the pseudodata with the default Py6 PARP(90) = 0.16 (I think)?

  * I think it would be really nice to have a plot which shows the error 
band blow up in a region without data, cf. the NNPDF PDF error band 
which becomes largest for very low-x where there is no data to constrain 
the "replicas". Unfortunately, the only thing which comes to mind would 
be one bin per energy, i.e. <Nch_trans for pTlead in [20,40] GeV> vs. 
sqrt(s) which requires a lot of runs to fill... but it would be a nice plot!

  * Sampling from the covariance Gaussian is actually a very NNPDF type 
approach to building the error bands: I like it. But are you doing 
something like building the band from only the central 68% or 90% of 
your MC sampled points? If not, what does the band mean statistically?

  * Also about the sampling: the place where my suggestion of 
identifying the 2P maximally-orthogonal 1-sigma param points is still 
useful (I think) if we want to be able to suggest particular 
representative tunes cf. the CTEQ/MRST "error PDFs". I'm not sure if 
there's demand for that, but if it's easy there's nothing to stop us 
from publishing the numbers.  The alternative is to construct more 
subjective error tunes like Peter's Perugia-soft/hard... which I think 
is nice, but not statistically very meaningful. We can discuss whether 
we can come up with a "more Professor" way to produce tunes of that 
sort: they certainly have the benefit of being less in number than the 
2P (+ 1 central point) tunes!

Andy