2nd Pan-European Advanced School on Statistics in High Energy Physics

28 Mar 2022, 14:00 → 30 Mar 2022, 20:25 Europe/Berlin

SR 4a/b (DESY Hamburg)

We are looking forward to welcome you to the second Pan-European Advanced School on Statistics in High Energy Physics. The virtual school is open to master and PhD students as well as to Post-docs. Participants should already have a good knowledge of statistical methods in data analysis. The school focuses on two topics:

1. Modeling of Data 
2. Data Combination.

The compact school (3 days a ~4 hours) comprises lectures by renowned statisticians and particle physicists, followed by ample time for discussion.
 

Click here to join Zoom

The biennial series of Pan-European Advanced Schools on Statistics in High Energy Physics was founded in 2019 by the INSIGHTS Marie Sklodowska-Curie ITN and DESY. More information on INSIGHTS can be found here.

Monday, 28 March

Modeling of data 1

1 Introduction

Speaker: Olaf Behnke (CMS (CMS Fachgruppe TOP))

obehnke220330.pdf

Olaf_Behnke_Introduction.mp4

2 Goodness-Of-Fit tests (Part 1)

A goodness-of-fit test is concerned with the question whether a
given data set was generated by a specific probability distribution
such as an exponential. In this seminar we will discuss a variety
of such tests. We will consider their relative merits, and how to
run several of them simultaneously. We will also discuss tests for
multivariate data and a number of special cases such as binned
data and comparing Monte Carlo to data.

Speaker: Wolfgang Rolke (UPR )

GOF_tests.pdf

GOF_tests.pptx

Wolfgang_Rolke_Lecture_1.mp4

14:55 Discussion time

15:05 Virtual coffee break

3 Goodness-Of-Fit tests (Part 2)

Speaker: Wolfgang Rolke (UPR - Mayaguez)

Wolfgang_Rolke_Lecture_2.mp4

16:00 Discussion time

16:15 Virtual coffee break

4 Statistical Modeling

In this lecture I will given an overview of various aspects of
statistical modeling. I'll start be reviewing parametric models.
Then I will discuss methods for dealing with model
misspecification. Next I will discuss nonparametric methods and
then move on to universal inference which is a new method for
handling irregular models. I'll finish with a few quick remarks on
semiparametric inference.

Speaker: Larry Wasserman (CMU )

Larry_Wasserman_modeling.mp4

Modeling.pdf

17:45 Discussion time

Tuesday, 29 March

Data combination

5 Data combination - introduction

The lecture will address the topic of combining information
from different sources in an analysis of Particle Physics data.
The general formalism by which this is done in both the
Bayesian and Frequentist approaches will first be reviewed.
Combination of results relies fundamentally on constructing
a likelihood that reflects all of the available data. Often this
requires some approximations and assumptions as the
detailed information needed to write down the full likelihood
may not be available. An important aspect of combined (and
individual) data analyses is the assignment of uncertainties to estimates
to nuisance parameters. A method will be described by which
uncertainties on the assigned uncertainties
themselves can be incorporated, and the impact of this type of
a model on combinations will be shown.

Link to the python code used for the example of a+bx fit:
https://www.pp.rhul.ac.uk/~cowan/stat/fitCombo.py
https://www.pp.rhul.ac.uk/~cowan/stat/fitCombo.ipynb

Speaker: Glen Cowan (RHUL)

cowan_desy_adv_school_29mar22.pdf

video1544467930.mp4

14:45 Discussion time

15:00 Virtual coffee break

6 Data combination - in practice

The lecture will address practical aspects and possible pitfalls
when combining particle physics measurements or limits and will
give pointers to methods and tools that can be used for that
purpose. One particular focus will be the combination of single
valued and multiple valued (differential) measurements with
complex correlations between their nuisance parameters. The
lecture will be accompanied by hands-on examples. A CERN
account with the possibility to log in to lxplus is recommended, but
all examples can also be followed without running the software.

For use of convino code at lxplus.cern.ch:

After login do:

bash

cd /afs/cern.ch/user/j/jkiesele/public/Convino/latest

source lxplus_env.sh

cd

mkdir convino_tutorial

cd convino_tutorial

convino /afs/cern.ch/user/j/jkiesele/public/Convino/latest/examples/exampleconfig.txt

cp -r /afs/cern.ch/user/j/jkiesele/public/Convino/tutorial/* .

Speaker: Jan Kieseler (CERN )

20220229_kieseler.pdf

Jan_Kiesler.mp4

16:15 Discussion time

16:30 Virtual coffee break

7 What to publish

The statistical model is the a unique summary of a physics
analysis from which many of the key results can be derived
from. Preserving it allows not only the reproduction of key
results but also its reuse in statistical combinations and
reinterpretations. In this talk we will cover how to best
publish statistical model data to enable these use-cases.

A different way of making measurements available to physicists
outside of the experimental collaboration is by reversing
the smearing effects of the detector
and reconstruction. Various different methods
of unfolding allow to publish data in a way
that is independent of a specific experimental setup.

Speakers: Carsten Burgard (ATLAS (ATLAS Dark Matter with Higgs)), Lukas Heinrich (CERN)

Carsten_Burgard.mp4

cburgard-unfolding

DESY_Stats_2022_PubLhood.pdf

Lukas_Heinrich.mp4

17:45 Discussion time

Wednesday, 30 March

Modeling of data 2

8 Introduction to Optimal Transport

Optimal transport (OT) is a method for mapping one probability
distribution into another. OT also leads to a method for defining a
geodesic between distributions which allows us to morph one
distribution into another. I will introduce the basics of optimal
transport and I will explain how optimal transport maps and
morphings can be estimated from data.

Speaker: Larry Wasserman (CMU )

Transport.pdf

video1518103067.mp4

14:45 Discussion time

15:00 Virtual coffee break

9 Gaussian Processes

In this lecture, I will provide an introduction to Gaussian
processes (GPs), with a view toward applications in high-energy
physics. I will start with the basic definition of a GP and explain
how to perform inference with these models. I will then describe
the choice and estimation of the mean and the covariance
function and demonstrate these ideas with simple examples.
I will close with a brief overview of applications of GPs in
high-energy physics.

Speaker: Mikael Kuusela (Carnegie Mellon University)

Kuusela_GPs_DESY_Mar_2022.pdf

Mikael_Kuusela_gp.mp4

16:15 Discussion time

16:30 Virtual coffee break

10 EFT Lagrangian Morphing

In this lecture I will discuss a method of morphing distributions
that is useful to measure the parameters of an Effective Field
Theory (EFT). I will introduce EFT which is a powerful
theoretical framework that is used to systematically extend
known physics lagrangians. I will then talk about the idea
behind the morphing between distributions given the
predictions at some point in the parameter space which allow
to obtain a continuous prediction in terms of EFT parameters.
I will finally show a couple of examples of the implementation
of this technique as the RooLagrangianMorphFunc class
within RooFit toolkit that is available with the ROOT software.

Speaker: Rahul Balasubramanian (Nikhef and University of Amsterdam)

desy_stat_school.pdf

video2748463886.mp4

17:45 Discussion time

11 Closing of School

Speaker: Olaf Behnke (CMS (CMS Fachgruppe TOP))

obehnke220330_finale.pdf

video3748463886.mp4