Seminare de DPNC
================

Mercredi, 5. fev. 1997

17.00 - Auditoire Stuckelberg

Universite de Geneve
Departement de physique
nucleaire et corpusculaire
24, quai Ernest-Ansermet
1211 Geneva 4

Tel. 022 702 6273
Fax. 022 781 2192

******************************************************************************
Sujet:		Bayesian Reasoning in High Energy Physics

Par:		Prof. G. d'Agostini, Rome 

Resume:

    Bayesian statistics associates the idea of probability 
    - {\it the measure of the degree of belief that an event
    will occur} - to the lack of knowledge, as it is commonly 
    perceived intuitively. The Bayes' theorem becomes then the 
    basic tool to evaluate the probability, combining {\it a priori} 
    judgements and experimental information. 
    This approach allows to treat in a logically consistent
    way all kinds of uncertainties, including those originated 
    from systematic errors. This fact has been also recognized
    recently by the international metrology organizations.
    The results are compared with the standard - frequentistic - 
    methods currently used for uncertainty evaluations
    and hypothesis tests. 

******************************************************************************
Prof d'Agostini will also give two seminars at 
CERN on Thursday, Febr. 6th  

	10.00 hours bldg 160-1-009  
  - "Probabilistic treatment of measurement uncertainties due
    to systematic errors: theory and practice".

	16.00 hours bldg 160-1-009 
  - "A multidimensional unfolding method based on Bayes' theorem"


Abstracts as follows:
CERN on Thursday, Febr. 6th  10.00 bldg 160-1-009 
  - "Probabilistic treatment of measurement uncertainties due
    to systematic errors: theory and practice".
    In Bayesian statistics also the "true" value of a physics 
    quantity plays the role of a random number, in the general
    sense of "a number respect to which we are in a status of 
    uncertainty", due to lack of knowledge. In this framework
    is possible to built a theory capable to handle all sources
    of measurement uncertainty. Some detailed examples of application
    are given, and approximated methods, necessary for routine use,
    are also shown, including the evaluation of correlations arising 
    from common systematics. The approximations, limit case
    in which linearization is meaningful and one is interested only
    in the best estimates of the quantities and on their variances,
    recover the recent ISO recommendation on the subject.

CERN on Thursday, Febr. 6th  16.00 bldg 160-1-009 
  - "A multidimensional unfolding method based on Bayes' theorem"
    The distribution of measured observables differs from the 
    "true" one, due to unavoidable physics and detector distortions.
    Although very semplicistic procedures - like the "bin-to-bin
    correction" - are currently still used, many attempts have 
    been tried to solve the problem with the due care. The desiderata
    that any unfolding method should satisfy are discussed.  
    A Bayesian approach is presented, justified by the fact that
    a statistical inferential problem should be solved within the frame
    of a probabilistic theory, rather then using ad hoc procedures. 
    The resulting unfolding method is simple, it provides the covariance 
    matrix of the results and it takes into account also the uncertainty
    due to limited statistics of Monte Carlo simulation (needed to 
    study the distortion). Moreover, it allows to handle in a 
    consistent and compact way also the side problem of background 
    subtraction.