Jets 1

MadGraph School
Shanghai - November 2015
Jets
Lecture 1: jet algorithms
Lecture 2: jet substructure
Matteo Cacciari
LPTHE Paris
Why jets
A jet is something that happens

in high energy events:
a collimated bunch of hadrons
flying roughly in the
same direction
We could eyeball the collimated

bunches, but it becomes impractical
with millions of events
The classification of particles into jets is best done

using a clustering algorithm
Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 2
The pervasiveness of jets
‣ ATLAS and CMS have each published 400+ papers since 2010
‣ More than half of these papers make use of jets
‣ 60% of the searches papers makes use of jets
Plot by G. Salam
(Source: INSPIRE.
Results may vary when
employing different search
keywords)
Why are jets so important?

Taming reality
Multileg + PS
??
QCD predictions Real data

Jets
One purpose of a ‘jet clustering’ algorithm is to
reduce the complexity of the final state, simplifying many hadrons
to simpler objects that one can hope to calculate
Jet clustering algorithm
A jet algorithm maps the momenta of the final state particles
into the momenta of a certain number of jets:
{pi} {jk}
jet algorithm
particles, jets
4-momenta,
calorimeter towers, ....
Most algorithms contain a resolution parameter, R,

which controls the extension of the jet

Jet Definition
A jet algorithm
+
its parameters (e.g. R)
+
a recombination scheme
=
a Jet Definition
/// define a jet definition

JetDefinition jet_def(JetAlgorithm jet_algorithm,
double R,
RecombinationScheme rec_sch = E_scheme);

Two main classes of jet algorithms
‣ Sequential recombination algorithms
Bottom-up approach: combine particles starting from closest ones
How? Choose a distance measure, iterate recombination until
few objects left, call them jets
Works because of mapping closeness QCD divergence
Examples: Jade, kt, Cambridge/Aachen, anti-kt, …..
‣ Cone algorithms
Top-down approach: find coarse regions of energy flow.
How? Find stable cones (i.e. their axis coincides with sum of momenta of particles in it)
Works because QCD only modifies energy flow on small scales
Examples: JetClu, MidPoint, ATLAS cone, CMS cone, SISCone…...

A little history
‣Cone-type jets were introduced first in QCD in the 1970s
(Sterman-Weinberg ’77)
‣In the 1980s cone-type jets were adapted for use in hadron
colliders (SppS, Tevatron...) ➙ iterative cone algorithms
‣LEP was a golden era for jets: new algorithms and many
relevant calculations during the 1990s
‣ Introduction of the ‘theory-friendly’ kt algorithm
‣ sequential recombination type algorithm, IRC safe
‣ it allows for all order resummation of jet rates
‣ Several accurate calculations in perturbative QCD of jet
properties: rates, jet mass, thrust, ....

ee
+ - kt (Durham) algorithm
[Catani, Dokshitzer, Olsson, Turnock, Webber ’91]
Distance:
In the collinear limit, the numerator reduces to the relative transverse

momentum (squared) of the two particles, hence the name of the algorithm
‣ Find the minimum ymin of all yij
‣ If ymin is below some jet resolution threshold ycut, recombine i and j

into a single new particle (‘pseudojet’), and repeat
‣ If no ymin < ycut are left, all remaining particles are jets

ee
+ - kt (Durham) algorithm in action
2-jet
Characterise events
in terms of number of jets
(as a function of ycut)
3-jet
4-jet
5-jet
Resummed calculations for distributions of ycut doable with the kt algorithm

ee
+ - kt (Durham) algorithm v. QCD
kt is a sequential recombination type algorithm
One key feature of the kt
algorithm is its relation to the
structure of QCD divergences:
The yij distance is the inverse of the emission probability
‣ The kt algorithm roughly inverts the QCD branching sequence

(the pair which is recombined first is the one with the largest
probability to have branched)
‣ The history of successive clusterings has physical meaning

Jet challenges at the LHC
The LHC environment differs from the LEP one
(and even the Tevatron) under many respects
‣ Number of final state particles much larger (order 103)
➙ needs a fast algorithm
‣ Many higher order calculations (NLO, NNLO) available

➙ needs an IRC-safe algorithm
‣ Presence of background (underlying event and pileup)

➙ needs small/known susceptibility
and/or ability to subtract background
‣ Jets often initiated by a large-momentum heavy particle
➙ needs capability to distinguish
boosted objet jet from QCD jet
IRC safety
An observable is infrared and collinear safe if,
in the limit of a collinear splitting, or the emission of an
infinitely soft particle, the observable remains unchanged:
O(X; p1 , . . . , pn , pn+1 0) O(X; p1 , . . . , pn )
O(X; p1 , . . . , pn ⇥ pn+1 ) O(X; p1 , . . . , pn + pn+1 )
This property ensures cancellation of real and virtual divergences
in higher order calculations
If we wish to be able to calculate a jet rate in perturbative QCD

the jet algorithm that we use must be IRC safe:
soft emissions and collinear splittings must not change the hard jets

Janus Jets
From Wikipedia:
In ancient Roman religion and myth, Janus (Latin: Ianus, pronounced [ˈiaː.nus])
is the god of beginnings and transitions, and thereby of gates, doors, passages,
endings and time. He is usually depicted as having two faces, since he looks to
the future and to the past. The Romans named the month of January (Ianuarius)
in his honor.
Like Janus, jets can serve two purposes:
‣ Observables
to be defined, calculated, measured
‣ Tools
to be used to extract specific properties of the final state
Different clustering algorithms have different properties and characteristics

that can make them more or less appropriate for each of these tasks

Jets as tools
Background
Mass
characterisation
reconstruction
and subtraction
Remove soft
Tag heavy objects
contamination
originating the jet
from a hard jet

(Boosted) jet studies at the LHC
Lily Asquith, summary talk at BOOST 2015
Essentially none of these tools existed

as lately as seven years ago
Glossary
What i.e. When Ref.
AKT Anti-kt algorithm 2008 0802.1189
CA Cambridge/Aachen algorithm 1999 9907280
BDRS mass-drop tagger, includes filtering 2008 0802.2470
trimmed Trimming, tagger/groomer 2009 0912.1342
pruned Pruning, tagger/groomer 2009 0903.5081
HTT HepTopTagger 2009 0910.5472
N-subjettiness jet shape function, used in tagging 2010 1011.2268
WTA Winner-Take-All (recombination scheme) 2013 1310.7584
one-pass choice of axis for N-subjettiness 2010

JVT Jet Vertex Tagger (used in pileup subtr.) 2014
ρ background density (used in pileup subtr.) 2007 0707.1378
D2 jet shape function, used in tagging 2014 1409.6298
PUPPI particle-by-particle pileup subtr. 2014 1407.6013
Soft Drop tagger/groomer 2014 1402.2657

Outline
‣Algorithms, speed
‣Infrared and collinear safety Lecture 1
‣Background (pileup)
‣Substructure Lecture 2

Outline
‣Infrared and collinear safety
‣Substructure

The common wisdom circa 2005
‣Cone algorithms are IRC unsafe
➙ because, to make them reasonably
fast, they were usually implemented via
approximate methods using seeds
‣Sequential recombination algorithms (i.e. kt)

are slow and too susceptible to background
contamination
➙ because they scale like N3
➙ because they tend to collect soft
particles up to large distances from centre
➙ because they were often run with R=1 and
compared to cones with R=0.5!
Geometry
The solution to the speed problem came from
considering the clustering problem from a
geometrical rather from a combinatorial point of view
‣ Sequential recombination algorithms could be
implemented with O(N2) or even O(NlnN)
complexity rather than O(N3)
[MC, Salam, 2006]
‣ Cone algorithms could be implemented exactly

(and therefore made IRC safe) with O(N2lnN)
rather than O(N2N) complexity
[Salam, Soyez 2007]

The kt algorithm and its siblings
Δy2 + Δφ2
2p
2p 2p
di j = min(kti , kt j ) diB = kti
R2
p=1 kt algorithm S. Catani,Y. Dokshitzer, M. Seymour and B. Webber, Nucl. Phys. B406 (1993) 187
S.D. Ellis and D.E. Soper, Phys. Rev. D48 (1993) 3160
p = 0 Cambridge/Aachen algorithm
Y. Dokshitzer, G. Leder, S.Moretti and B. Webber, JHEP 08 (1997) 001
M. Wobisch and T. Wengler, hep-ph/9907280
p = -1 anti-kt algorithm MC, G. Salam and G. Soyez, arXiv:0802.1189

NB: in anti-kt pairs with a hard particle will cluster first: if no other
hard particles are close by, the algorithm will give perfect cones
Quite ironically, a sequential recombination algorithm is the ‘perfect’ cone algorithm

IRC safety of generalised-kt algorithms
Δy2 + Δφ2
2p
2p 2p
di j = min(kti , kt j ) diB = kti
R2
p>0
New soft particle (kt →0) means that d → 0 clustered first, no effect on jets
New collinear particle (Δy2+ΔΦ2 → 0) means that d → 0 clustered first, no effect on jets
p=0
New soft particle (kt →0) can be new jet of zero momentum no effect on hard jets
p<0
New soft particle (kt →0) means d →∞ clustered last or new zero-jet, no effect on hard jets

IRC safe algorithms
SR
Catani et al ‘91
kt dij = min(kti2,ktj2)ΔRij2/R2 Ellis, Soper ‘93 NlnN
hierarchical in rel p
t
SR
Cambridge/ dij = ΔRij2/R2 Dokshitzer et al ‘97
NlnN
Aachen hierarchical in angle
Wengler, Wobish ‘98
SR
MC, Salam, Soyez ’08
anti-kt -2 2
dij = min(kti ,ktj )ΔRij /R
-2 2
(Delsart, Loch)
gives perfectly conical hard jets
N3/2
Seedless iterative cone

SISCone with split-merge
gives ‘economical’ jets
Salam, Soyez ‘07 N2lnN
‘second-generation’ algorithms
All are available in FastJet, http://fastjet.fr
(As well as many IRC unsafe ones)
Jet clustering in FastJet
/// define a jet definition
JetDefinition jet_def(JetAlgorithm jet_algorithm,
double R,
RecombinationScheme rec_sch = E_scheme);
jet_algorithm can be any one of the four IRC safe algorithms, or also
most of the old IRC-unsafe ones, for legacy purposes
/// create a ClusterSequence, extract the jets

ClusterSequence cs(input_particles, jet_def);
vector<PseudoJet> jets = sorted_by_pt(cs.inclusive(jets));
...
// pt of hardest jet
double pt_hardest = jets[0].pt();
...
// constituents of hardest jet
vector<PseudoJet> constits = jets[0].constituents();

FastJet speed
Time needed to cluster an event with N particles
2
10
kt
i-k
t
Intel i5 760
t e n
10 FastJet 3.0.1 an ach
R=0.7 e /A
i dg
1
b r
m
Ca
ne
10-1
Co
-2
SIS PbPb
time (s)
10
collisions
-3
10
-4 hadron
10 collisions
-5 with pileup anti-kt
10 kt
-6 hadron C/A
10 collisions
SISCone
10-7
1 10 102 103 104 105 106 107
N
Outline
‣Infrared and collinear safety
‣Substructure

Hard jets and background
In a realistic set-up underlying event (UE) and pile-up (PU)

from multiple collisions produce many soft particles which
can ‘contaminate’ the hard jet

Pileup
78-vertices event
from CMS
https://cds.cern.ch/record/1479324
Pileup can deposit several tens of GeV (or even

hundreds, in a heavy ion collision) into a medium-sized jet

How are the hard jets
modified by the background?
Susceptibility Jet areas

(how much bkgd gets picked up)
Resiliency Backreaction
(how much the original jet changes)

Jet areas
A jet’s area is defined as the extent of the region where
infinitesimally soft particles get clustered into the jet
More in details, a jet’s active area is

the extent of the region where a
distribution of infinitesimally soft
particles, that can also cluster among
themselves, is clustered into the jet
A jet’s active area measures a jet’s

sensitivity to contamination from soft
particles like underlying event and pileup
Jets do not necessarily have
cone-shaped profiles

kt Cam/Aa
SISCone anti-kt
Resiliency: backreaction
“How (much) a jet changes when immersed in a background”
Without With
background background
Backreaction loss
Backreaction gain
Resiliency: backreaction
1
pT loss SISCone (f=075)
Cam/Aachen
Pythia 6.4
LHC (high lumi)
pT gain
kt
anti-kt 2 hardest jets
1/N dN/dpt (GeV-1)
pt,jet> 1 TeV
0.1
|y|<2
R=1
0.01
0.001
-20 -15 -10 -5 0 5 10 15
pt(B) (GeV)
Anti-kt jets are much more resilient to changes from background immersion
(NB. Backreaction is a minimal issue in pp background and at large pt.

Can be much more important in Heavy Ion collisions)

Anti-kt jets and background
Anti-kt jets maximise

resiliency, and their regular
shapes makes them easier to
correct for detector-related
effects
Default choice of all LHC collaborations

The IRC safe algorithms
UE/pileup Hierarchical
Speed Regularity contamination Backreaction
substructure
kt ☺☺☺ ☂ ☂☂ ☁☁ ☺☺
Cambridge ☺☺☺ ☂ ☂ ☁☁ ☺☺☺

/Aachen
anti-kt ☺☺☺ ☺☺ ☁/☺ ☺☺ ✘
SISCone ☺ ☁ ☺☺ ☁ ✘
Array of tools with different characteristics.

Pick the right one for the job
Modifications of the hard jet
⇧
pt = A ± (⇥ A + ⇥ A + ⇤A2 ⌅ ⇤A⌅2 ) + BR
pt
Background
momentum density
(per unit area)
background back-reaction
‘susceptibility’ ‘resiliency’

Background subtraction
Two main approaches to background subtraction
‣ Jet-based
‣ Cluster the full event, determine the event-specific (ρ) and jet-
specific (A) quantities, and subtract the relevant contamination from a
given observable
‣ Pros: largely unbiased subtraction
‣ Cons: slow, potentially large(er) residual uncertainty
‣ Examples: `jet area/median’ in FastJet, GenericSubtractor for jet shapes, JetFFMoments for
fragmentation functions, ....
‣ Particle-based
‣ Produce a reduced event, by dropping some of the particles. Cluster
this reduced event, and calculate from it the observables
‣ Pros: fast, often small(er) residual uncertainty
‣ Not natively unbiased, can depend on choice of parameters
‣ Examples: ConstituentSubtractor, SoftKiller, PUPPI, ....

Background subtraction: jet-based
Correction of a jet transverse momentum
hard jet, corrected hard jet, raw
pT = pT ρ ⇥ Areahard jet
MC, Salam, 0707.1378
If ρ is measured on an event-by-event basis, and each jet subtracted individually, this procedure
will remove many fluctuations and generally improve the resolution of, say, a mass peak
⇧
pt = A ± (⇥ A + ⇥ A + ⇤A2 ⌅ ⇤A⌅2 ) + pBR
t
Irreducible fluctuations:
uncertainty of the subtraction
Needs two ingredients: ρ and Ajet

Jet areas
Definition of a specific jet area
and its calculation for each jet in FastJEt
/// specify where to place the ghosts, their area, how many times
/// to repeat the clustering
double rapmax = 2.5;
int nrepeat = 1;
double ghost_area = 0.01;
GhostedAreaSpec gas(rapmax, nrepeat, ghost_area);
/// use this configuration to define the area

/// A sensible default for gas is provided
AreaDefinition area_def(active_area, gas);
/// construct cluster sequence with areas

ClusterSequenceArea(input_particles, jet_def, area_def);
....
jets[0].area()

Background estimation and subtraction
// constructor for a background estimator
JetMedianBackgroundEstimator bge(Selector sel,
JetDefinition jet_def,
AreaDefinition area_def);
// an alternative (faster) background estimator

//GridMedianBackgroundEstimator bge(Selector sel, grid_step);
bge.set_particles(input_particles);
....
double rho = bge.rho(jet); // extract rho estimation
// define a subtractor
Subtractor sub(&bge);
// apply it to a jet (or a vector of jets)

PseudoJet subtracted_jet = sub(jet);

Numerical jet shape correction
Soyez et al. 1211.2811
A generic jet shape

(a function of the momenta of all
constituents of a jet) is modified
by the addition of pileup
Correct it by calculating numerically the derivatives that enter its Taylor

expansion and subtracting (this generalises the jet area/median subtraction for transverse mom.)
Pileup Numerical derivative
momentum density w.r.t. ghosts momenta

Shape subtraction
Pilup subtraction from jet shapes using
GenericSubtractor from fjcontrib
#include “ExampleShapes.hh”
#include “GenericSubtractor.hh”
// define a specific jet shape

contrib::Angularity ang(1.0); // angularity with alpha=1.0
// define a generic subtractor

... construct a background estimator bge_rho....
contrib::GenericSubtractor gen_sub(&bge_rho);
// compute the subtracted shape

double subtracted_ang = gen_sub(ang, jet);

Particle-level pilup removal
Pilup removal using SoftKiller from fjcontrib
(SoftKiller progressively removes soft particles until ρ of event is zero)
#include “SoftKiller.hh”
// define SoftKiller
double grid_size = 0.4;
contrib::SoftKiller soft_killer(rapmax, grid_size);
// apply it to the full event

double pt_thresh //returns threshold for killed particles
vector<PseudoJet> soft_killed_event;
soft_killer.apply(full_event, soft_killed_event, pt_thresh);
// proceed with clustering and calculating shapes with

// the reduced soft_killed_event
ClusterSequence(soft_killed_event,.....)

Comparisons of pileup subtractions
from the CERN pileup mitigation workshop, https://indico.cern.ch/event/306155/
PRELIMINARY CHS event, jet m, pt=020
3
areasub
constitsub
safeareasub
anti-kt R=0.4
2.5
soft killer
puppi
30, 60, 100, 140 PU
safenpcsub
DISPERSION
lin cl.(Rsub=0.3)
WORSE
σΔO (GeV)
1.5
1 BETTER
0.5
0
BIAS
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
<ΔO> (GeV)
BETTER
Recap of lecture 1
‣ A number of different IRC-safe jet algorithms exist

‣ They all try to be good proxies for hard partons, but they have
different characteristics, especially with respect to soft particles
‣ Jets from all algorithms inevitably suffer from pileup contamination

‣ Techniques exist to subtract it, either at jet-level, or at particle-level
‣ Both the jet algorithms and many pileup subtraction techniques are
packaged aither in FastJet or in fjcontrib contributions
‣ Use of standard algorithms and packages (either directly or
through interfaces) should be privileged, as it ensures
reproducibility
http://fastjet.fr http://fastjet.hepforge.org/contrib/

Jets 1

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Jets 1

Загружено:

Авторское право:

Доступные форматы

MadGraph School

Shanghai - November 2015

A jet is something that happens

We could eyeball the collimated

The classification of particles into jets is best done

Why are jets so important?

QCD predictions Real data

Most algorithms contain a resolution parameter, R,

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 5

/// define a jet definition

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 6

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 7

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 8

In the collinear limit, the numerator reduces to the relative transverse

‣ Find the minimum ymin of all yij

‣ If ymin is below some jet resolution threshold ycut, recombine i and j

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 9

Resummed calculations for distributions of ycut doable with the kt algorithm

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 10

The yij distance is the inverse of the emission probability

‣ The kt algorithm roughly inverts the QCD branching sequence

‣ The history of successive clusterings has physical meaning

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 11

‣ Many higher order calculations (NLO, NNLO) available

‣ Presence of background (underlying event and pileup)

If we wish to be able to calculate a jet rate in perturbative QCD

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 13

Like Janus, jets can serve two purposes:

Different clustering algorithms have different properties and characteristics

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 14

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 15

Essentially none of these tools existed

CA Cambridge/Aachen algorithm 1999 9907280

BDRS mass-drop tagger, includes filtering 2008 0802.2470

trimmed Trimming, tagger/groomer 2009 0912.1342

pruned Pruning, tagger/groomer 2009 0903.5081

HTT HepTopTagger 2009 0910.5472

N-subjettiness jet shape function, used in tagging 2010 1011.2268

WTA Winner-Take-All (recombination scheme) 2013 1310.7584

one-pass choice of axis for N-subjettiness 2010

D2 jet shape function, used in tagging 2014 1409.6298

PUPPI particle-by-particle pileup subtr. 2014 1407.6013

Soft Drop tagger/groomer 2014 1402.2657

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 18

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 19

‣Sequential recombination algorithms (i.e. kt)

‣ Cone algorithms could be implemented exactly

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 21

p = -1 anti-kt algorithm MC, G. Salam and G. Soyez, arXiv:0802.1189

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 22

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 23

Seedless iterative cone

/// create a ClusterSequence, extract the jets

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 25

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 27

In a realistic set-up underlying event (UE) and pile-up (PU)

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 28

Pileup can deposit several tens of GeV (or even

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 29

Susceptibility Jet areas

Matteo Cacciari - LPTHE MadGraph School - Shanghai - October 2015 30

More in details, a jet’s active area is

A jet’s active area measures a jet’s