Exploring The (Metric) Space of Collider Events MIT CTP - - PowerPoint PPT Presentation

Exploring The (Metric) Space of Collider Events

MIT CTP QCD/LHC/BSM/DM Journal Club

Eric M. Metodiev

Center for Theoretical Physics Massachusetts Institute of T echnology Joint work with Patrick Komiske and Jesse Thaler

[1902.02346]

March 8, 2019

1

Exploring the (Metric) Space of Collider Events

Outline

2

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Applications Introduction

Exploring the (Metric) Space of Collider Events

Applications Introduction

3

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Exploring the (Metric) Space of Collider Events

When are two events similar?

Eric M. Metodiev, MIT 4

Exploring the (Metric) Space of Collider Events

When are two collider events similar?

Eric M. Metodiev, MIT 5

Fragmentation

partons ! " # …

Collision Detection Hadronization

hadrons $± &± … ' '

How an event gets its shape

Exploring the (Metric) Space of Collider Events

When are two collider events similar? Experimentally: very complicated

Eric M. Metodiev, MIT 6

Theoretically: very complicated

A collider event is… The energy flow (distribution of energy) is the information that is robust to: fragmentation, hadronization, detector effects, … Energy flow ó Infrared and Collinear (IRC) Safe information However:

[F.V. Tkachov, 9601308] [N.A. Sveshnikov, F.V. Tkachov, 9512370] [P.S. Cherzor, N.A. Sveshnikov, 9710349]

Exploring the (Metric) Space of Collider Events

When are two collider events similar?

Eric M. Metodiev, MIT 7

Rapidity ! A z i m u t h "

Fragmentation

partons # $ % …

Collision Detection Hadronization

hadrons &± (± … ) )

Energy flow is robust information Treat events as distributions of energy: *

+,- .

/+ 0( ̂ )+)

energy direction Ignoring particle flavor, charge…

Exploring the (Metric) Space of Collider Events

Applications Introduction

8

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

When they have similar distributions of energy

Exploring the (Metric) Space of Collider Events

Applications Introduction When they have similar distributions of energy

9

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Exploring the (Metric) Space of Collider Events

The Energy Mover’s Distance

Earth Mover’s Distance: the minimum “work” (stuff x distance) to rearrange one pile of dirt into another

Eric M. Metodiev, MIT 10

Review: The Earth Mover’s Distance Metric on the space of (normalized) distributions: symmetric, non-negative, triangle inequality Distributions are close in EMD ó their expectation values are close. Also known as the 1-Wasserstein metric.

[Y. Rubner, C. Tomasi, and L.J. Guibas] [S. Peleg, M. Werman, H. Rom]

Exploring the (Metric) Space of Collider Events

The Energy Mover’s Distance

Energy Mover’s Distance: the minimum “work” (energy x angle) to rearrange one event (pile of energy) into another

Eric M. Metodiev, MIT 11

EMD ℇ, ℇ& = min

{,} . /01 2

.

301 24

5

/3

6/3 7

8/ 8

3 &

6/3 5

/3

Difference in radiation pattern

From Earth to Energy

[P.T. Komiske, EMM, J. Thaler, 1902.02346]

Exploring the (Metric) Space of Collider Events

The Energy Mover’s Distance

Energy Mover’s Distance: the minimum “work” (energy x angle) to rearrange one event (pile of energy) into another

Eric M. Metodiev, MIT 12

EMD ℇ, ℇ& = min

{,} . /01 2

.

301 24

5

/3

6/3 7 + .

/01 2

9/ − .

301 24

9

3 &

9/ 9

3 &

6/3 5

/3

Difference in radiation pattern Difference in total energy

From Earth to Energy

[P.T. Komiske, EMM, J. Thaler, 1902.02346]

Exploring the (Metric) Space of Collider Events

The Energy Mover’s Distance

Energy Mover’s Distance: the minimum “work” (energy x angle) to rearrange one event (pile of energy) into another

Eric M. Metodiev, MIT 13

EMD ℇ, ℇ& = min

{,} . /01 2

.

301 24

5

/3

6/3 7 + .

/01 2

9/ − .

301 24

9

3 &

9/ 9

3 &

6/3 5

/3

Difference in radiation pattern Difference in total energy

EMD has dimensions of energy True metric as long as 7 ≥ 1

< 6=>?

Solvable via Optimal Transport problem.

~1ms to compute EMD for two jets with 100 particles. 7 ≥ the jet radius, for conical jets

From Earth to Energy

[P.T. Komiske, EMM, J. Thaler, 1902.02346]

Exploring the (Metric) Space of Collider Events

Applications Introduction When they have similar distributions of energy

14

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Work to rearrange one event into another.

Exploring the (Metric) Space of Collider Events

Applications Introduction

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

When they have similar distributions of energy Work to rearrange one event into another.

15

Part I Part II

Eric M. Metodiev, MIT

Outline

Exploring the (Metric) Space of Collider Events

Movie Time: Visualizing the EMD

Eric M. Metodiev, MIT 16

EMD is the cost of an optimal transport problem. We also get the shortest path between the events. Interpolate along path to visualize! Taking a walk in the space of events

Exploring the (Metric) Space of Collider Events

Movie Time: Visualizing Jet Formation

Eric M. Metodiev, MIT 17 ! !

QCD Jets W Jets T

• p Jets

Pythia 8, + = 1.0 jets, !3 ∈ 500,550 GeV

Fragmentation Collision Hadronization

Exploring the (Metric) Space of Collider Events

Movie Time: Visualizing QCD Jet Formation

Eric M. Metodiev, MIT 18

Quark Fragmentation Hadronization

EMD: 111.6 GeV

fragmentation

EMD: 18.1 GeV

hadronization

Exploring the (Metric) Space of Collider Events

Movie Time: Visualizing W Jet Formation

Eric M. Metodiev, MIT 19

Decay Quarks Fragmentation Hadronization W

EMD: 78.3 GeV

decay

EMD: 26.3 GeV

fragmentation

EMD: 12.9 GeV

hadronization

Exploring the (Metric) Space of Collider Events

Movie Time: Visualizing Top Jet Formation

Eric M. Metodiev, MIT 20

EMD: 161.1 GeV

decay

EMD: 47.1 GeV

fragmentation

EMD: 27.0 GeV

hadronization Decay Quarks Fragmentation Hadronization Top

Exploring the (Metric) Space of Collider Events

Applications Introduction When they have similar distributions of energy Work to rearrange one event into another.

21

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Visualize energy movement and jet formation.

Exploring the (Metric) Space of Collider Events

When they have similar distributions of energy Work to rearrange one event into another. Visualize energy movement and jet formation. Applications Introduction

22

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Exploring the (Metric) Space of Collider Events

Observables

Eric M. Metodiev, MIT 23

!-(sub)jettiness: "#

$ = & '() *

+' min

# /012{4),6 $ , 47,6 $ , … , 4#,6 $ }

")/+ ≫ 0 ")/+ > "7/+ ≫ 0 ">/+ ≃ 0 measures how well jet energy is aligned into N (sub)jets

[J. Thaler, K. Van Tilburg, 1011.2268] [J. Thaler, K. Van Tilburg, 1108.2701] [I. Stewart, F. Tackmann, W. Waalewijn, 1004.2489]

Exploring the (Metric) Space of Collider Events

Observables

Eric M. Metodiev, MIT 24

!-subjettiness: "#

$ = & '() *

+' min

# /012{4),6 $ , 47,6 $ , … , 4#,6 $ }

")/+ ≫ 0 ")/+ > "7/+ ≫ 0 ">/+ ≃ 0 measures how well jet energy is aligned into N subjets

!-subjettiness is the EMD between the event and the closest !-particle event.

@ ≠ 1 corresponds to other p-Wasserstein distances with p = @.

"#(ℇ) = min

ℇG (# EMD ℇ, ℇ′ .

[J. Thaler, K. Van Tilburg, 1011.2268] [J. Thaler, K. Van Tilburg, 1108.2701]

Exploring the (Metric) Space of Collider Events

Observables

Eric M. Metodiev, MIT 25

! ℇ = $

%&' (

)% Φ ̂ ,%

Take any additive IRC-safe observable:

• (/) = $

%&' (

)% 1%

/

e.g. jet angularities: 1% )% Getting quantitative

[C. Berger, T. Kucs, and G. Sterman, 0303051] [A. Larkoski, J. Thaler, and W. Waalewijn, 1408.3122]

Exploring the (Metric) Space of Collider Events

Observables

Via the Kantorovich-Rubinstein dual formulation of EMD:

Eric M. Metodiev, MIT 26

EMD ℇ, ℇ& ≥ 1 )* +

,-. /

0,Φ ̂ 3, − +

5-. /6 5 & Φ

̂ 35′ = 1 )* 9 ℇ − 9 ℇ& 9 ℇ = +

,-. /

0, Φ ̂ 3,

Take any additive IRC-safe observable:

Difference in

• bservable values

Earth Mover’s Distance

:(<) = +

,-. /

0, >,

<

e.g. jet angularities:

“Lipschitz constant” of Φ i.e. bound on its derivative

>, 0,

[C. Berger, T. Kucs, and G. Sterman, 0303051] [A. Larkoski, J. Thaler, and W. Waalewijn, 1408.3122]

Getting quantitative

Exploring the (Metric) Space of Collider Events

Observables

Via the Kantorovich-Rubinstein dual formulation of EMD:

Eric M. Metodiev, MIT 27

EMD ℇ, ℇ& ≥ 1 )* +

,-. /

0,Φ ̂ 3, − +

5-. /6 5 & Φ

̂ 35′ = 1 )* 9 ℇ − 9 ℇ& 9 ℇ = +

,-. /

0, Φ ̂ 3,

Take any additive IRC-safe observable:

Difference in

• bservable values

Earth Mover’s Distance

:(<) = +

,-. /

0, >,

<

e.g. jet angularities:

“Lipschitz constant” of Φ i.e. bound on its derivative

For ? ≥ 1 jet angularities, * = ?/) over the jet cone, so:

:(<) ℇ − :(<) ℇ& ≤ ? EMD ℇ, ℇ&

The EMD provides a robust upper bound to any modifications of these observables. >, 0,

[C. Berger, T. Kucs, and G. Sterman, 0303051] [A. Larkoski, J. Thaler, and W. Waalewijn, 1408.3122]

Getting quantitative

Exploring the (Metric) Space of Collider Events

Observables

Eric M. Metodiev, MIT 28

Theorem: Any infrared and collinear safe observable ! can be approximated arbitrarily well as: ! "#, … , "& = ( )

*+# &

,* Φ ̂ "* for some Φ: ℝ1 → ℝℓ and F: ℝℓ → ℝ and sufficiently large ℓ. Key idea: Energy-weighted angular structures contain all the IRC-safe information.

1 67 )

*+# &

,*Φ ̂ "* − )

9+# &:

,

9 ; Φ

̂ "9′ ≤ EMD ℇ, ℇ;

Events close in EMD are close in all infrared and collinear safe information!

[M. Zaheer, et al., 1703.06114] [P.T. Komiske, EMM, J. Thaler, 1810.05165]

Exploring the (Metric) Space of Collider Events

Applications Introduction When they have similar distributions of energy Work to rearrange one event into another. Visualize energy movement and jet formation.

29

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Conceptually rich connections to EMD.

Exploring the (Metric) Space of Collider Events

When they have similar distributions of energy Work to rearrange one event into another. Visualize energy movement and jet formation. Conceptually rich connections to EMD. Applications Introduction

30

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Hadronization

Eric M. Metodiev, MIT 31

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Hadronization

Eric M. Metodiev, MIT 32

!(#$%) ℇ − !(#$%) ℇ) ≤ EMD ℇ, ℇ) ℇ = ℇ0123456 ℇ) = ℇ7182456

partons hadrons

!(#$%) = 111.1GeV !(#$%) = 111.6GeV

!(#$%) = ?

@$% A

B@ C@

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Hadronization

Eric M. Metodiev, MIT 33

ℇ = ℇ#$%&'() ℇ* = ℇ+$,%'()

partons hadrons

• (/01) = 111.1GeV
• (/01) = 111.6GeV
• (/01) = 9

:01 ;

<: =:

• (/01) ℇ − -(/01) ℇ*

≤ EMD ℇ, ℇ*

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Pileup

Eric M. Metodiev, MIT 34

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Pileup

Eric M. Metodiev, MIT 35

Leading Vertex Jet + Pileup CHS P U P P I SoftKiller How can we quantify pileup mitigation? P U M M L

[M. Cacciari, G.P. Salam, G. Soyez, 1407.0408] [D. Bertolini, P. Harris, M. Low, N. Tran, 1407.6013] [P.T. Komiske, EMM, B. Nachman, M.D. Schwartz, 1707.08600]

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Pileup

Eric M. Metodiev, MIT 36

Discontinuous under physically-sensible single-pixel perturbations. Undesirable behavior with increasing resolution. Compare calorimeter images pixel by pixel? Requires ad hoc choices of observables. Compare on a collection of observables?

Exploring the (Metric) Space of Collider Events

Quantifying event modifications: Pileup

Eric M. Metodiev, MIT 37

+ Pileup CHS PUPPI SoftKiller Measure pileup mitigation performance with EMD! PUMML Leading Vertex Jet Guarantees performance on IRC safe observables. Stable under physically-sensible perturbations. Train to optimize EMD with machine learning?

Exploring the (Metric) Space of Collider Events

Applications Introduction When they have similar distributions of energy Work to rearrange one event into another. Visualize energy movement and jet formation. Conceptually rich connections to EMD.

38

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Outline

Hadronization, pileup, detector effects

Exploring the (Metric) Space of Collider Events

Applications Introduction

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

When they have similar distributions of energy Work to rearrange one event into another. Visualize energy movement and jet formation. Conceptually rich connections to EMD. Hadronization, pileup, detector effects

39

Part I Part II

Eric M. Metodiev, MIT

Outline

Exploring the (Metric) Space of Collider Events

Exploring the Space of Events: W jets

Eric M. Metodiev, MIT 40

W !

1−!

# Visualize the space of events with t-Distributed Stochastic Neighbor Embedding (t-SNE). Finds an embedding into a low-dimensional manifold that respects distances. ! 1 − ! #% = '()

%

'*

% = +, %

'*

%

W jets are 2-pronged: !: Energy Sharing of Prongs #: Angle between Prongs /: Azimuthal orientation Constrained by W mass: Hence we expect a two-dimensional space of W jets. After / rotation: one-dimensional

[L. van der Maaten, G. Hinton]

Exploring the (Metric) Space of Collider Events

Exploring the Space of Events: W jets

Eric M. Metodiev, MIT 41

W !

1−!

# 2x zoom “bottom heavy” “top heavy” “one pronged” “balanced” ?

W jets, $ = 1.0 )* ∈ 500,510 GeV

! 1 − ! #1 = )23

1

)*

1 = 45 1

)*

1

W jets are 2-pronged: !: Energy Sharing of Prongs #: Angle between Prongs 8: Azimuthal orientation Constrained by W mass: Hence we expect a two-dimensional space of W jets. After 8 rotation: one-dimensional

Exploring the (Metric) Space of Collider Events

Exploring the Space of Jets: Correlation Dimension

Eric M. Metodiev, MIT 42

dim $ = $ & &$ ln )

*+,

• )

.+,

• Θ[EMD ℇ*, ℇ. < $]

Energy scale $ dependence Count neighbors in ball of radius $

89:;<=>?@;9<

A?;9BC

D ∝ DF;G dim(D) = r & &D ln 89:;<=>?@C D Intuition: Correlation dimension:

Exploring the (Metric) Space of Collider Events

Exploring the Space of Jets: Correlation Dimension

Eric M. Metodiev, MIT 43

QCD jets are simplest. W jets are more complicated. T

• p jets are most complex.

“Decays” have ~constant dimension. Fragmentation becomes more complex at lower energy scales. Hadronization becomes relevant at scales around 20 GeV. Can we understand this analytically?

Exploring the (Metric) Space of Collider Events

Exploring the Space of Jets: Correlation Dimension

Eric M. Metodiev, MIT 44

dim$/&(() = − 8-./$/& ln ( 34/2 /$ = /6 = 4 3 /& = /9 = 3 At LL:

+ 1-loop running of -. Quark jets Gluon jets

Dimension blows up at low energies. Jets are “more than fractal”?

Exploring the (Metric) Space of Collider Events

Exploring the Space of Events: k-medoids

Eric M. Metodiev, MIT 45

Exploring the (Metric) Space of Collider Events

Exploring the Space of Events: Jet Classification

Eric M. Metodiev, MIT 46

Classify W jets vs. QCD jets Look at a jet’s nearest neighbors (kNN) to predict its class. Optimal IRC-safe classifier with enough data. Nearing performance of ML.

vs. better N-subjettiness

EMD kNN ML

Exploring the (Metric) Space of Collider Events

Exploring the Space of Events

Eric M. Metodiev, MIT 47

Use EMD as a measure of event similarity Unsupervised clustering algorithms can be used to cluster events Jets are clusters of particles ???? are clusters of jets VP Tree: O(log(N)) neighbor query time Much more to explore. Vantage Point (VP) Tree

Exploring the (Metric) Space of Collider Events 48

When are two events similar? The Energy Mover’s Distance Movie Time Observables Quantifying event modifications Exploring the Space of Events

Part I Part II

Eric M. Metodiev, MIT

Summary

When they have similar distributions of energy Work to rearrange one event into another. Visualize energy movement and jet formation. Conceptually rich connections to EMD. Hadronization, pileup, detector effects Unlock new ideas and techniques with EMD Applications Introduction

Exploring the (Metric) Space of Collider Events

Going Beyond

• Model (in)dependent anomaly detection?
• Sharpen the parton-hadron duality of energy flow?
• Train ML models to optimize EMD directly?
• Include flavor information?

Eric M. Metodiev, MIT 49

Exploring the (Metric) Space of Collider Events

The End

Thank you!

Eric M. Metodiev, MIT 50

Exploring the (Metric) Space of Collider Events

Extra Slides

Eric M. Metodiev, MIT 51

Exploring the (Metric) Space of Collider Events

Exploring the Space of Jets: Correlation Dimension

Eric M. Metodiev, MIT 52

= − 8$%&'/) * ln

• .//2

&' = &1 = 4 3 &) = &4 = 3

+ 1-loop running of $%

dim'/) - = - 8 8- ln 9

:;< =

9

>;< =

Θ[EMD ℇ:, ℇ> < -] = - 8 8- ln Pr [EMD < -] = - 8 8- ln exp − 4$M&'/) * lnN

• .//2

= - 8 8- ln Pr [O P;< < -; &'/) → 2 &'/)]

[A. Larkoski, 1709.06195]

Sketch of leading log (one emission) calculation:

Exploring the (Metric) Space of Collider Events

What is a collision event?

Eric M. Metodiev, MIT 53

! "± $± %± &± &'

(

)/ ̅ ) ,/- ,

photon electron muon pion kaon K-long proton neutron

tracker ECAL HCAL

When are two collider events similar?

How an event gets its shape: Experiment

Exploring the (Metric) Space of Collider Events

Pileup Mitigation with PUMML

Eric M. Metodiev, MIT 54