Discussion summary
------------------

"How shall one include upcoming LHC data in global BSM parameter fits?"

Specifically, we have discussed the (mutually related) questions  

1) In what form should the experimental results be made available by
the LHC collaborations?

2) How can we test models that have not been specifically analysed 
by the experimental collaborations?


Regards Q 1), Kyle Cranmer, Max Baak et al. have proposed to publish
RooFit/RooStats workspaces containing the dataset and the full
probability model (PDFs), such that the likelihood function can be
evaluated (for a given model) for any given point in parameter
space. The likelihood function depends in general on parameters of
interest and nuisance parameters. Workspaces also provide the
necessary information to generate toy data for coverage tests, etc.

However, the likelihood function refers to a specific model and cannot
be used to re-interprete published data in terms of different new
physics models without approximations.

Thus, in addition to likelihood functions for specific models, we need 
a more model-independent way to publish experimental results. First, 
one would like to test models that may not have been analysed by the 
LHC collaborations, and secondly, the published likelihoods are not
only model-specific, but also rely on theoretical tools (Monte Carlo 
programs etc.) that may improve with time. Thus it is desirable to
present data in a way to would allow to redo entire the physics
analysis with improved tools and/or re-interpret the data in terms of 
different models. The following options have been discussed:

1) Publication of measured data and estimated background in terms of 
“detector-unfolded” fiducial cross-section (ie. corrected for 
reconstruction effects). 
Pro: would allow theorists to test any model without having to know
the specific details of LHC detector effeciencies;
Con: detector-unfolding is model-dependent due to model-dependent 
reconstruction efficiencies;
Con: selection cuts are tuned to specific model and might not be
optimal for alternative models.

2) Publication of measured data and estimated background events 
without detector unfolding.
Pro: Does not involve any model dependence; 
Con: Theorists need to rely on simplified detector simulations to 
test alternative models; 
Con: selection cuts are tuned to specific model and might not be
optimal for alternative models.

3) Publication of experimental results in terms of simplified models.
Pro: Should allow to test wider class of models with less bias; 
Con: mapping of specific new physics model onto (set of) simplified 
models requires approximations and might introduce systematic 
errors. (See http://www.lhcnewphysics.org for a list of simplified 
models.)

Approaches like QUAERO and RECAST have not been discussed as they 
cannot be done outside the experimental collaborations. 

As all three approaches 1)-3) above introduce approximations, it would 
be desirable to compare those approaches for a number of selected test
cases. 

To proceed, and to prepare a more quantitative discussion at the
upcoming LPCC meeting, we propose to start working on the inclusion 
of workspaces into global fits. Max Baak has suggested to send around 
a dummy workspace with workspaces of squark/gluino (or m0/m12) mass 
plane plus instructions on how to use them in fitter programs. 

We furthermore propose to compare directly performing a likelihood fit
for a specific model provided by the experiments with the
approximative methods 1)-3) listed above.

To proceed here we would have to consider a toy analysis, where the 
experiments generate events for e.g. a mSUGRA model and provide 
i) a workspace containing the dataset and the PDFs obtained assuming 
this specific model, and ii) a fit of the toy data to (a set of) 
simplified models. In order to use the results in terms of simplified 
models in an mSUGRA fit, we need to make additional assumptions and 
approximations; we would thus like to see how the results for the two 
different approaches compare. 

MK & PW, 6.12.2010