Energy Flow Polynomials for Jet Substructure MIT Jet Workshop Cambridge, MA – January 11, 2018 Patrick T. Komiske III Center for Theoretical Physics, Massachusetts Institute of Technology PTK, E.M. Metodiev, J. Thaler – 1712.07124 Patrick T. Komiske III (MIT) EFPs for Jet Substructure 1 / 28
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Energy Flow Polynomials for Jet SubstructureMIT Jet Workshop
Cambridge, MA – January 11, 2018
Patrick T. Komiske IIICenter for Theoretical Physics, Massachusetts Institute of Technology
PTK, E.M. Metodiev, J. Thaler – 1712.07124
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 1 / 28
Start with an arbitrary IRC-safe observable S(pµ1 , . . . , pµM )
Energy expansion*: Approx. S with polynomials of zij
IR safety: S unchanged by addition of infinitesimally soft particlesC safety: S unchanged by collinear splittings of particlesRelabeling symmetry: Particle indexing is arbitrary
See also F. Tkachov hep-ph/9601308, N. Sveshnikov and F. Tkachov hep-ph/9512370
=⇒ Energy correlators linearly span IRC-safe observables
Angular expansion*: Approx. angular part of S with polynomials of θijSimplify: Identify unique analytic structures that emerge as EFPsSimilar emergent multigraphs in M. Hogervorst et al. 1409.1581 and B. Henning et al. 1706.08520
=⇒ EFPs linearly span IRC-safe observables
*These expansions generically make use of the Stone-Weierstrass Theorem
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 6 / 28
Start with an arbitrary IRC-safe observable S(pµ1 , . . . , pµM )
Energy expansion*: Approx. S with polynomials of zijIR safety: S unchanged by addition of infinitesimally soft particlesC safety: S unchanged by collinear splittings of particlesRelabeling symmetry: Particle indexing is arbitrary
See also F. Tkachov hep-ph/9601308, N. Sveshnikov and F. Tkachov hep-ph/9512370
=⇒ Energy correlators linearly span IRC-safe observables
Angular expansion*: Approx. angular part of S with polynomials of θijSimplify: Identify unique analytic structures that emerge as EFPsSimilar emergent multigraphs in M. Hogervorst et al. 1409.1581 and B. Henning et al. 1706.08520
=⇒ EFPs linearly span IRC-safe observables
*These expansions generically make use of the Stone-Weierstrass Theorem
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 6 / 28
Start with an arbitrary IRC-safe observable S(pµ1 , . . . , pµM )
Energy expansion*: Approx. S with polynomials of zijIR safety: S unchanged by addition of infinitesimally soft particlesC safety: S unchanged by collinear splittings of particlesRelabeling symmetry: Particle indexing is arbitrary
See also F. Tkachov hep-ph/9601308, N. Sveshnikov and F. Tkachov hep-ph/9512370
=⇒ Energy correlators linearly span IRC-safe observablesAngular expansion*: Approx. angular part of S with polynomials of θijSimplify: Identify unique analytic structures that emerge as EFPsSimilar emergent multigraphs in M. Hogervorst et al. 1409.1581 and B. Henning et al. 1706.08520
=⇒ EFPs linearly span IRC-safe observables
*These expansions generically make use of the Stone-Weierstrass Theorem
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 6 / 28
Machine learn {sG} with a linear modelLinear models:
Convex with few/no hyperparameters to tuneAchieve global optimum via closed form solution or convergent iterationCannot have a simpler model (1 parm./input) sensitive to all inputsMany potential methods to analyze learned modelSee Ch. 3 and 4 of C. Bishop Pattern Recognition and Machine Learning
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 12 / 28
Linear Regression
Confirming Analytic Relationships
λ(2) = 12 × , λ(4) = − 3
4 ×
λ(6) = − 32 × + 5
8 ×
Energy Flow Polynomials
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Lea
rned
Coeffi
cien
t
W Jets: Ang. α = 2
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 1, d ≤ 7
Energy Flow Polynomials
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Lea
rned
Coeffi
cien
t
W Jets: Ang. α = 4
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 1, d ≤ 7
Energy Flow Polynomials
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Lea
rned
Coeffi
cien
t
W Jets: Ang. α = 6
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 1, d ≤ 7
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 13 / 28
Linear Regression
Linear Regression and IRC-safetymJ/pT,J : IRC safe - no Taylor expansion due to square rootλ(α=1/2): IRC safe - no simple analytic relationshipτ
(β=1)2 : IRC safe - algorithmically definedτ
(β=1)21 : Sudakov safe - safe for 2-prong jets and moreτ
(β=1)32 : Sudakov safe - safe for 3-prong jets and more
Multiplicity: IRC unsafe
2 3 4 5 6 7
Max Degree of EFPs
0.0
0.2
0.4
0.6
0.8
1.0
Cor
r.C
oef
.5
th−
95th
Per
centi
le
QCD Jets
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 1
mJ/pTJ
λ(α=1/2)
τ(β=1)2
τ(β=1)21
τ(β=1)32
Mult.
2 3 4 5 6 7
Max Degree of EFPs
0.0
0.2
0.4
0.6
0.8
1.0
Cor
r.C
oef
.5
th−
95th
Per
centi
le
W Jets
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 1
mJ/pTJ
λ(α=1/2)
τ(β=1)2
τ(β=1)21
τ(β=1)32
Mult.
2 3 4 5 6 7
Max Degree of EFPs
0.0
0.2
0.4
0.6
0.8
1.0
Cor
r.C
oef
.5
th−
95th
Per
centi
le
Top Jets
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 1
mJ/pTJ
λ(α=1/2)
τ(β=1)2
τ(β=1)21
τ(β=1)32
Mult.
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 14 / 28
Conclusions
Part I Conclusions
EFPs:Energy correlators with angular structures indexed by multigraphsLinearly span the space of IRC-safe observablesEncompass many existing classes of substructure observables
Linear regression:Linear models are the easiest and most tractable kind of model
Convex with few/no hyperparametersGlobal optimum via closed form solution or convergent iterationMany potential tools to analyze what’s learned
Works with EFPs to match onto many IRC-safe observables
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 15 / 28
Part II - Linear Jet Tagging with EFPs
Linear Classification with EFPs
Comparison with Modern Machine Learning
Fast Computation of EFPs
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 16 / 28
Linear Classification
Linear Classification Overview
Fit a decision plane, determined by a vector w
Fisher’s linear discriminant (LDA): closed-form solutionLogistic regression: Convex, iterative solution
Decision threshold t is determined by distance from the planeG is finite set of graphs corresponding to the inputs
Organization by d is natural (equivalent to the order of the expansion)Organization by N or χ also possible, (where is the information?)
Classifier =
t+
∑G∈G
wGEFPG
≥ 0, signal< 0, background
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 17 / 28
Linear Classification
Linear Classification with EFPs
W vs. QCD jet classification (quark/gluon and top tagging in backup)
0.0 0.2 0.4 0.6 0.8 1.0
W Jet Efficiency
10−1
100
101
102
103
Inve
rse
QC
DJet
Mis
tag
Rat
e
EFPs: W vs. QCD
Pythia 8.226,√s = 13 TeV
R = 0.8, pT ∈ [500, 550] GeV
EFP β = 0.5
d ≤ 3
d ≤ 6
d ≤ 7
Linear
DNN
better
300k training samplesLinear: Fisher’s linear discriminant
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 25 / 28
1
2
3
4
5Composite EFPs are products of prime EFPs
Computation
Variable Elimination (VE)
Algorithm for finding optimal parentheses placement in EFP formulaReduces EFP computational complexity to O(Mχ):
Best case (NP-hard): χ = treewidth(G) + 1Heuristics can be used which work excellently for our small graphsχ = 2 for all tree graphs, χ = 3 for single-cycle graphs, χ = N for KN
χ 2 3 4
e.g.
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 26 / 28
Computation
EnergyFlow Python Package
A convenient and simple package for efficient implementation of EFPsCurrently written in pure Python using the NumPy library
Need a fast, arbitrary dimension multi-arrayWe’re working on a C++ implementation (not simple)
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 27 / 28
Linear classification with EFPs very comparable to MML methodsLinear methods =⇒ very nice both theoretically and experimentally
EFP linear structure potentially allows for theoretical calculationFully differentiable model, uncertainty/error propagation simpleConvex, global minimum is guaranteedNo/few hyperparametersInteresting methods made possible by linearity
Lasso regression for automatic feature selectionPCA, orthogonal subspaces, etc.
Efficient computation of EFPs has been achievedEnergyFlow Python package here, stay tuned for more
EFPs potentially bridge MML performance & theory understanding
Patrick T. Komiske III (MIT) EFPs for Jet Substructure 28 / 28