[Forum SIS] Seminario a LUISS :: Ted Hill

[Pierpaolo Brutti]

Giovedì 21 ottobre 2010, ore 15
Aula 202 - sede LUISS di Viale Romania

Title: Combining Data Sets: Conflations of Probability Distributions

Speaker: Ted Hill
(Georgia Institute of Technology and California
Polytechnic State University)

Suppose that n independent laboratories perform different
experiments to determine the same quantity, such as the value of one
of the fundamental physical constants. If the distributions of the
laboratories’ data are P1 ,…, Pn , how do you best summarize P1 ,…, Pn
by a single probability distribution P? After describing some of the
shortcomings of various other methods, a new simple mathematical
method called conflation will be described, along with its basic
closure properties and optimality properties. For example, the
conflation of several normal distributions is always normal, and
conflation is the unique distribution that minimizes the maximum loss
of (Shannon) information that occurs when several distributions are
combined into one distribution. An application to real data will be
shown, and extensions of conflation to truncated data and to
situations where the different experiments are not uniformly weighted,
as may happen when one laboratory is known to have better equipment.
The talk is aimed for non-specialists.

 (See http://arxiv.org/abs/0808.1808 and http://arxiv.org/abs/0901.4957 )




==============================================
Pierpaolo Brutti
LUISS, Dipartimento di Scienze Economiche e Aziendali
Viale Romania 32
Email: pbrutti at luiss.it
Web: http://docenti.luiss.it/brutti/
==============================================