PARTON DISTRIBUTIONS

FUNCTIONS

STEFANO FORTE
UNIVERSITÀ DI MILANO & INFN

GRK 1694 FREUDENSTADT, SEPT. 26-28, 2016

 THE IMPACT OF PDFS: HIGGS DISCOVERY
HIGGS PRODUCTION

(J. Campbell, HCP2012)

PDF UNCERTAINTY EITHER DOMINANT, OR VERY LARGE, OR BOTH

. . . AND NOT ONLY FOR THE HIGGS!
(W MASS DETERMINATION, NEW PHYSICS SEARCHES FOR HEAVY STATES,. . . )
 SUMMARY
LECTURE I: THE BASICS
• FACTORIZATION
– RENORMALIZATION AND FACTORIZATION IN QCD
– ELECTROPRODUCTION AND HADROPRODUCTION
– EVOLUTION EQUATIONS AND SUM RULES
• FROM DATA TO PDFS
– DATA FROM HERA TO THE LHC
– DISENTANGLING QUARK FLAVORS
– DETERMINING THE GLUON
LECTURE II: PDF DETERMINATION
• STATISTICS AND METHODOLOGY
– HESSIAN VS MONTE CARLO APPROACH
– HESSIAN UNCERTAINTY ESTIMATES AND TOLERANCE
– PARAMETRIZATION BIAS
– GENERAL PARAMETRIZATIONS AND CROSS-VALIDATION
– COMPRESSION METHODS AND MONTE CARLO ↔ HESSIAN CONVERSION
– NON-GAUSSIAN BEHAVIOUR
– CLOSURE TESTING

LECTURE III: THE STATE OF THE ART

• THEORETICAL ISSUES
– PERTURBATIVE STABILITY AND HIGHER ORDER CORRECTIONS
– RESUMMATION
– HEAVY QUARKS: RESUMMATION AND MATCHING
• PDFS NOW
– GLOBAL AND NONGLOBAL PDF DETERMINATIONS
– RECENT PROGRESS: METHODOLOGY AND LHC DATA
– THE PDF4LHC COMBINED PDF SETS
FACTORIZATION
 RENORMALIZATION
A QUICK REMINDER
g
L = − 24 φ4 φφ → φφ ELASTIC SCATTERING OF MASSIVE SCALAR FIELDS

dσ g2 1
d cos θ
= 128π s
; s = (p1 + p2 )2

dσ g2 1
d cos θ
= 128π s
F (s, t) : DIVERGES!;
t = (p1 − p3 )2 , u = (p1 − p4 )2

g M 2 (s)
F (s, t) = limΛ→∞ 1 + 32π 3+ 0
ln Λ2
+ s → t + s → u ; M 2 (s) = m2 − x(1 − x)s
 
R1

RENORMALIZATION: EXPRESS A PHYSICAL OBSERVABLE IN TERMS OF OTHER PHYSICAL

OBSERVABLES:
2
gphys
dσ 1
WHAT IS THE CHARGE g? DEFINE gphys FROM d cos θ s=4m2
= 128π m2
:
2
dσ
gphys
1 gphys M 2 (s)
d cos θ = 128π s F (s, t); F (s, t) = 1 + 32π 0
ln M 2 (4m2 )
+s→t+s→u
 
R1

UV SINGULARITY IS UNIVERSAL ⇒ REABSORBED IN DEF. OF THE COUPLING

 FACTORIZATION
IN PERTURBATIVE QCD
DEEP-INELASTIC LEPTON-HADRON SCATTERING

PROBE THE PROTON WITH A SHORT -WAVELENGTH PHOTON:

e = P → e + X: for fixed energy,

cross section depends on two variables
(cos θ, W 2 )
 FACTORIZATION
IN PERTURBATIVE QCD
DEEP-INELASTIC LEPTON-HADRON SCATTERING
PROBE THE PROTON WITH A SHORT -WAVELENGTH PHOTON:
QCD IS ASYMPTOTICALLY FREE, USE PERTURBATION THEORY:

dσ 2 dσ̂ 2
d cos θdW 2
=
2
+O M
P

Q2
i d cosθdW 2 qi (x(θ, W ))+
i e

• AT FIRST PERTURBATIVE ORDER, INCOHERENT SUM OF CONTRIBUTIONS FROM CHARGED

CONSTITUENTS (QUARKS), PROPORTIONAL TO THEIR CHARGE

• MOMENTUM OF THE CONSTITUENT PROPORTIONAL TO PROTON MOMENTUM p̂ = xp

• “MOMENTUM FRACTION” x ENTIRELY FIXED BY KINEMATICS, BY Q2 & W 2 (I.E. cos θ & W 2 )
• THE CROSS-SECTION IS PROPORTIONAL, UP TO KINEMATIC FACTORS, TO THE PROBABILITY
qi (x) OF THE PHOTON STRIKING A QUARK OF THE i-TH FLAVOR OR ANTIFLAVOR WITH
MOMENTUM p̂ = xp

PARTON MODEL; PARTON= STRUCK CONSTITUENT:

qi (x) ⇒ PARTON DISTRIBUTION (PDF)
 FACTORIZATION
IN PERTURBATIVE QCD
DEEP-INELASTIC LEPTON-HADRON SCATTERING
WHAT HAPPENS AT HIGHER ORDERS?
COLLINEAR SINGULARITIES!

dσ 2 dσ̂ Q2 M2
d cos θdW 2
= i ei d cos θdW 2
P (x1 x2 )qi (x1 )dx2 ln M2
+O Q2
dσ̂ phys 2
R  

= x, Q + O M
P

i ei d cos θdW 2 (x)qi Q2

P 2
 2

• HIGHER ORDER CORRECTIONS ARE SINGULAR;

SINGULARITY REGULATED BY THE PROTON SCALE M
• DEFINE A PHYSICAL PDF WITH SINGULARITY FACTORED IN IT ⇒ SCALE DEPENDENT
 FACTORIZATION
IN PERTURBATIVE QCD
DEEP-INELASTIC LEPTON-HADRON SCATTERING
WHAT HAPPENS AT HIGHER ORDERS?
COLLINEAR SINGULARITIES!

dσ 2 dσ̂ Q2 2
d cos θdW 2
= i ei P (x1 x2 )qi (x1 )dx2 ln
d cos θdW 2 M2
+O M 2
dσ̂ phys y 2 dy Q2
P R  

= e
i i d cos θdW 2
P (x)qi x , Q 0 y ln 2 + O M
Q2
Q0
P 2R
 2Q

• EXPRESS THE PROCESS IN TERMS OF THE PROCESS AT ANOTHER SCALE

• UP TO POWER CORRECTIONS, PROCESS FACTORIZES (NO INTERFERENCE):

PARTONIC PROCESS ⊗ SCALE DEPENDENT PDF

• SCALE DEPENDENCE CAN BE SEEN AS A BRANCHING DRIVEN BY THE SPLITTING FUNCTION

KERNEL P (x)
 DEEP-INELASTIC SCATTERING
THE STRUCTURE FUNCTIONS
p·q
Lepton fractional energy loss: y = p·k
;
k k′ gauge boson virtuality: q 2 = −Q2
Q2
Bjorken x: x = 2p·q
Q2
q lepton-nucleon CM energy: s = xy
;
virtual boson-nucleon CM energy W 2 = Q2 1−x
x
;

p X
d2 σ λp λℓ (x, y, Q2 ) G2F Q2
= −λℓ y 1 − xF3 (x, Q2 ) + (1 − y)F2 (x, Q2 )
dxdy 2π(1 xy 2
(

+ Q2 /m2W )2
h  y

+y 2 xF1 (x, Q2 ) − 2λp −λℓ y(2 − y)xg1 (x, Q2 ) − (1 − y)g4 (x, Q2 ) − y 2 xg5 (x, Q2 )
)
  

PARITY CONS. PARITY VIOL.

λl → lepton helicity UNPOL. F1 , F2 F3
λp → proton helicity POL. g1 g4 , g5
 FACTORIZATION:
STRUCTURE FUNCTIONS AND PDFS
STRUCTURE FUNCTION=HARD COEFF. (PARTONIC STRUCTURE FUNCTION) ⊗PARTON DISTN.

dy 2 x 2 2 x
F2 (x, Q2 ) =x i 1 y Ci αs (Q ), y qi (y, Q ) + q̄i (y, Q ) + Cg αs (Q2 ), y
g(y, Q2 )
P R1     

qi quark, q̄i antiquark, g gluon

 FACTORIZATION
FROM DIS TO HADRONIC PROCESSES

γ ∗, W ∗, Z ∗ p′

q̄
H
q
⇒ q W, Z

p X p
• ONE PARTON PER HADRON: p̂1 = xa p1 ; p̂2 = x2 p2
• COLLINEAR EMISSION FROM PARTON LEGS
⇒ UNIVERSAL (PROCESS-INDEPENDENT) REDEFINITION OF PDFS
• SUPPRESSION OF INTERFERENCE ⇒ FACTORIZATION
 HADRONIC PROCESSES
2 )
σX (s, MX = a,b x 2
dx1 dx2 fa/h1 (x1 )fb/h2 (x2 )σ̂qa qb →X x1 x2 s, MX
min
2 τ
L C x, α s (M 2)
R 1 THE PARTON LUMINOSITY

σX (s, M ) = σ0 a,b τ1 dx H
P

x ab x
P R

• PARTON LUMINOSITY Lab (τ ) = τ dx f

x a/h1
(x)fb/h2 (τ /x)
2
R1

2 =σ C MX 2)
• COEFFICIENT FUNCTION σ̂qa qb →X x1 x2 s, MX 0 x1 x2 s
, αs (MH
 

EXAMPLE: THE DRELL-YAN PROCESS

AT LEADING ORDER

• Hadronic c.m. energy: s = (p1 + p2 )2

p2
ŝ
• Momentum fractions x1, 2 = s
exp ±y;
q

µ+ Lead. Ord. ŝ = M 2
q̄ • Partonic c.m. energy: ŝ = x1 x2 s
2 2
s = (p1 + p2 ) ⇒ ⇒ MX
q • Invariant mass of final state X
γ∗ µ− (dilepton, Higgs,. . . ):
MW2 ⇒ scale of process

p1 MX2
• Scaling variable τ = s

dσ 4
⇒ M 2 dM 2 = σ0 L (τ ); σ0 = 9
πα 1s ;
 HADRONIC FACTORIZATION
BEYOND TOTAL CROSS-SECTIONS
WITH MORE DIFFERENTIAL KINEMATICS, MUST IMPOSE EXTRA CONSTRAINTS ⇒ MORE
INFORMATION

EXAMPLE: THE DRELL-YAN RAPIDITY DISTRIBUTION

dσX (s,MX2 )

dY dM 2
=
a b 2 1 x1
dσ̂q q →X
dx1 dx2 dy fa/h1 (x1 )fb/h2 (x2 ) x1 x2 s, MX 2 ln
a,b xmin dydM 2 x2
δ Y − −y
X
P R1
 

a b
dσ̂q q →X
LEADING ORDER:
dydM 2 1 2
= σ0 δ(y)δ(1 − x τx )
X
2
dσX (s,MX ) ±Y
⇒ dY dM 2 = σ 0 ab fa (x 1 )f b(x 2 ); x i = τ e
P

• RAPIDITY: LONGITUDINAL BOOST OF FINAL STATE WR TO CM OF HADRONIC

COLLISION

• AT LEADING ORDER, INVARIANT MASS DETERMINES τ = x1 x2 ,

HADRONIC RAPIDITY FIXES BOTH x1 , x2

FACTORIZATION HOLDS FOR A WIDE CLASS OF SUFFICIENTLY INCLUSIVE

OBSERVABLES (INCLUSIVE JETS, HIGGS AND GAUGE BOSON PRODUCTION
CHANNELS, ETC.)
FAILS FOR EXCLUSIVE OBVSERVABLES (E.G. ELASTIC SCATTERING)
 THE SCALE DEPENDENCE OF PDFS
EVOLUTION EQUATIONS
• DEFINE MELLIN OF PARTON DISTRIBUTIONS
f (N, Q2 )≡ 0 dx xN −1 f2 (x, Q2 )
NOTE LARGE/SMALL x ⇔ LARGE/SMALL N
R 1 MOMENTS

Q2
• DEFINE LOGARITHMIC SCALE t = ln :
Λ2
EVOLUTION GIVEN BY ORDINARY DIFFERENTIAL EQUATIONS (NO CONVOLUTION)

• ANOMALOUS DIMENSIONS RELATED TO DGLAP SPLITTING FUNCTIONS

1 N −1 P (x, α (t))
γ(N, αs (t)) ≡ 0 dx x s
R

d 2 αs (t) N S 2
∆qN S (N, Q ) = γqq (N, αs (t))∆qN S (N, Q ),
dt 2π
S S
d ∆Σ(N, Q2 ) αs (t) γqq (N, αs (t)) 2nf γqg (N, αs (t)) ∆Σ(N, Q2 )
= S S ,
∆g(N, Q2 ) ∆g(N, Q2 )
⊗
dt 2π γgq (N, αs (t)) γgg (N, αs (t))
  !  

• EVOLUTION OF SINGLET Σ(x, Q2 ) = i=1 qi (x, Q2 ) + q̄i (x, Q2 ) COUPLED TO GLUON

Pnf

• ALL “NONSINGLET” QUARK COMBINATIONS q N S (x, Q2 ) = qi (x, Q2 ) − qj (x, Q2 ) EVOLVE

INDEPENDENTLY
• ANOMALOUS DIMENSIONS COMPUTED IN PERTURBATION THEORY:
(0) (1)
γi (N, αs (t)) = γi (N ) + αs (t)γi (N ) + . . .
 PERTURBATIVE EVOLUTION
THE LEADING ORDER ANOMALOUS DIMENSIONS

QUALITATIVE FEATURES

• AS Q2 INCREASES, PDFS DECREASE AT LARGE x& INCREASE AT SMALL x DUE TO RADIATION

• GLUON SECTOR SINGULAR AT N =1⇒ GLUON GROWS MORE AT SMALL x

• γqq (1) = 0 ⇒ NUMBER OF QUARKS CONSERVED

 SUM RULES
CONSTRUCT CONSERVED QUANTUM NUMBERS CARRIED BY PARTON DISTRIBUTIONS:

• BARYON NUMBER p p p ¯p
0 dx (u − ū ) = 2 = 2 0 dx d − d
R1 R1

f i
• MOMENTUM 0 dxx i=1 q (x) + q̄i (x) + g(x) =1
R1 hPN
i

CANNOT DEPEND ON SCALE

• BARYON NUMBER γqq (1) − γqq̄ (1) = 0; AT LO γqq̄ (1) = 0 SO γqq (1) = 0

• MOMENTUM γqq (2) + γqg (2) = 0, γgq (2) + γgg (2) = 0

CAN EXTRACT FROM PHYSICAL OBSERVABLES: BARYON NUMBER

1 νp
• GROSS–LLEWELLYN-SMITH SUM RULE0 dx F
2 3 (x, Q 2 ) + F νn (x, Q2 ) =
3
R1

= CGLS (Q2 ) 01 dx u(x, Q2 ) − u(x, Q2 ) + d(x, Q2 ) − d(x, Q2 )

R h i
 FACTORIZATION SUMMARY
• PHYSICAL OBSERVABLES INVOLVING THE STRONG INTERACTION ARE COMPUTABLE IN
PERTURBATIVE QCD WHEN THEY INVOLVE A LARGE SCALE (ASYMPTOTIC FREEDOM)

• INITIAL-STATE (OR FINAL-STATE) HADRONS CAN BE TREATED THANKS TO

FACTORIZATION;

– COLLINEAR SINGULARITIES CAN BE REABSORBED INTO PDFS

– THEY LEAD TO ENHANCEMENT OF FACTORIZABLE CONTRIBUTIONS (POWER
SUPPRESSION OF INTERFERENCE)

– OBSERVABLE “HADRONIC” CROSS-SECTIONS ARE A CONVOLUTION OF A

“PARTONIC” PROCESS WITH INCOMING QUARKS AND GLUONS, TIMES PDFS
– PDFS ARE UNIVERSAL, I.E. PROCESS-INDEPENDENT

• PREDICTIVITY AT A HADRON COLLIDER INVOLVES THE PERTURBATIVE COMPUTATION

OF THE PARTONIC PROCESSES, AND A DETERMINATION OF PDFS

• PDFS ARE
A WAY OF EXPRESSING A PROCESS IN TERMS OF OTHER PHYSICAL PROCESSES
FROM DATA TO PDFS
 THE PDFS

(PDG 2016)

• THE MOMENTUM PROBABILITY DENSITY xfi (x) IS SHOWN AT

TWO DIFFERENT SCALES (LEFT ⇒ LOW SCALE; RIGHT ⇒ HIGH SCALE)

• GIVEN PDFS VS x AT ONE SCALE Q20

⇒ DETERMINED FOR ALL SCALES BY EVOLUTION EQUATIONS

• AS x≥1 KINEMATIC CONSTRAINT fi (x) = 0

• “VALENCE” UP AND DOWN: PEAKED AT x ∼ 0.3; EXPECT fx (x) ∼ (1 − x)β

i
x→1

• “SEA” ANTIQUARK AND GLUON GROW AT SMALL x

• “SINGLET” AND GLUON MIX ⇒ ALL PDFS LOOK THE SAME AS x→0
 BEFORE AND AFTER THE LHC
HADRONIC CROSS-SECTIONS
9 9
10 10
KINEMATIC PLANE
8 8
10 σtot 10
7 7
10 10
Tevatron LHC
6 6
10 10
5 5
10 10
σb
4 4
10 10
-2 -1

3 3
10 10
33

jet
2 σjet(ET > √s/20) 2
10 10
1 σW 1
10 10
σZ
0 0

σ (nb)
10 jet 10
σjet(ET > 100 GeV)
-1 -1
10 10
events/sec for L = 10 cm s

-2 -2
10 10
-3 -3
10 σt 10
-4 jet -4
10 σjet(ET > √s/4) 10
-5 σHiggs(MH = 150 GeV) -5
10 10
-6 -6
10 σHiggs(MH = 500 GeV) 10
-7 -7
10 10
0.1 1 10

√ s (TeV)

• Q2 : INVARIANT MASS OF FINAL STATE ⇒ WIDENING OF AVAILABLE PROCESSES

• AS ENERGY GROWS, DROP OF CROSS-SECTION MAY BE OFFSET BY GROWTH OF SMALL x PDFS

 DISENTANGLING PDFS

A SCIENTIFIC ART

• DEEP-INELASTIC SCATTERING DATA ON PROTON ABUNDANT AND PRECISE

• CC F1 AND F3 IN PRINCIPLE PROVIDE FOUR COMBINATIONS, AND NC F1 TWO MORE

⇒ ALL LIGHT FLAVORS

• HERA DATA ONLY DETERMINE FOUR COMBINATIONS OF PDFS:

FIXED COMBINATION OF F1 F3 , SO NC AND ±CC WITH e± , PLUS SEPARATE NC γ
AND Z FROM SCALE DEPENDENCE

• W± AND Z PRODUCTION (INCLUDING DOUBLE DIFFERENTIAL: MASS AND RAPIDITY)

PROVIDE A LARGE AMOUNT OF INFORMATION

• WHEN PRODUCING ELECTROWEAK FINAL STATES, THE GLUON CAN ONLY BE

ACCESSED FROM SCALE DEPENDENCE OR HIGHER ORDERS
...EXCEPT IN HIGGS PRODUCTION!

• JET PRODUCTION GIVES A DIRECT HANDLE ON THE GLUON

 LEADING PARTON CONTENT
(up to O[αs ] corrections)

DEEP-INELASTIC SCATTERING

NC F1 γ = e2i (qi + q̄i )

i ℓ e V A
NC F1 Z, int. = i Bi (qi + q̄i ) u,c,t +2/3 +1/2
P

(+1/2 − 4/3 sin2 θW )

NC F3 Z, int. = i Di (qi + q̄i ) d,s,b -1/3 (−1/2 + 2/3 sin2 θW ) -1/2
P

+
CC F1W = ū + d + s + c̄ ν 0 +1/2 +1/2
P

+
CC −F3W /2 = ū − d − s + c̄ e,µ, τ -1 (−1/2 + 2 sin2 θW ) -1/2

Bq (Q2 ) = −2eq Vℓ Vq PZ + (Vℓ2 + A2 2 2 2 2 2 2 2 2

ℓ )(Vq + Aq )PZ ; Dq (Q ) = −2eq Aℓ Aq PZ + 4Vℓ Aℓ Vq Aq PZ ; PZ = Q /(Q + MZ )

W + → W − ⇒ u ↔ d, c ↔ s; p → n ⇒ u ↔ d

DRELL-YAN
Lij (x1 , x2 ) ≡ qi (x1 , M 2 )q̄j (x2 , M 2 )
dσ 2 4πα2
γ 2
dM dy
(M , y) = 9M s 2 i ei L (x1 , x2 )
2 √
dσ πGF MV CKM
P 2 ii

W dy = i,j |Vij |Lij (x1 , x2 )

2
3s √
dσ πGF MV 2 2
Z = V + A
2 P

dy 3s i i i L (x1 , x2 )

ij

CKM CKM
2 P

Vij → CKM MATRIX (i = u, c t, j = d, s b), Vij = 1 + O(λ); λ = sin θC ≈ 0.22

 EXPLOITING ISOSPIN:
THE ISOTRIPLET STRUCTURE FUNCTION

up (x, Q2 ) = dn (x, Q2 ); dp (x, Q2 ) = un (x, Q2 )

p p ¯ p
F2p (x, Q2) − F2d (x, Q2 ) = (u + ū ) − d + d [1 + O(αs )]
3
1 p


(NNPDF, 2005)
1
...BUT IN ORDER TO ACCESS THE NEUTRON ONE MUST ASSUME F2d = 2
(F2p + F2n )
 EXPLOITING CHARGE CONJUGATION AND ISOSPIN:
QUARKS AND ANTIQUARKS AT A pp̄ COLLIDER (TEVATRON)
BY CHARGE CONJUGATION q̄P̄ = qp

DRELL-YAN p/d ASYMMETRY (ISOSPIN)

1.3

1.2
p′ 1.1 4 up d̄p + 1 dp ūp
σ pn 9 9 d̄
1 σ pp ∼ 4 up ūp + 1 dp d̄p ≈ ū
9 9

pp
µ+ large x
q̄ 0.9

pd
σ /2σ
CTEQ5M CTEQ4M
q 0.8 MRST MRS(r2)
µ GRV98
−
_ _
γ ∗
0.7 CTEQ5M (d = u)

0.6 Less than 1% systematic

p uncertainty not shown
0.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
x2 E866 (2001)
W± ASYMMETRY (C-CONJUGATION)
p 0.25
′
-1
0.2 CDF 1992-1995 (110 pb e+µ)
CTEQ-3M
0.15 RESBOS
q̄ 0.1 CTEQ-3M pp̄
DYRAD σ
0.05
W+ up (x1 )dp (x2 )+d̄p (x1 )ūp (x2 ) up d p

Charge Asymmetry
0 MRS-R2 (DYRAD) =
MRS-R2 (DYRAD)(d/u Modified)
pp̄ d p up
-0.05 σ dp (x1 )up (x2 )+ūp (x1 )d̄p (x2 )
∼
q W± -0.1 MRST (DYRAD)
W−
-0.15 if x1 , x2 in valence region,
-0.2
0 0.5 1 1.5 2
Lepton Rapidity
neglecting HQ & Cabibbo suppr.
p CDF (1998)
 EXPLOITING KINEMATIC COVERAGE AND FINAL STATES:
LIGHT FLAVORS AND STRANGENESS AT THE LHC
W± AND Z PRODUCTION
W AND Z CROSS SECTIONS

q̄ νℓ
q ℓ+
W+

ATLAS (2012)
pp̄
. σ pp̄+ = ud¯ + cs̄; σZ = uū + dd¯ + ss̄: STRANGENESS DETERMINED
ν̄ℓ W
q̄ BY COMPARISON
W MUON ASYMMETRY
q ℓ −
W −

pp̄
σ
W+ =
u(x1 )d̄(x2 )+d̄(x1 )u(x2 )
pp̄ d(x1 )ū(x2 )+ū(x1 )d(x2 )
σ
W−
q̄ ℓ− “VALENCE” x ⇒ NEGLECT STRANGENESS
q Z ℓ+
⇒ DETERMINE ū − d¯

CMS (2013)
 TAGGING FINAL STATES
STRANGENESS IN DIS AND AT THE LHC
DIMUONS TOTAL STRANGENESS
0.3
NNPDF1.2
0.25 NNPDF1.1

NNPDF1.0 ν
0.2 AT LO F2, c = 2xc;
0.15

0
ν̄
F2, c = 2xc̄
0.1

xs+ (x, Q 2)
UP TO CABIBBO SUPPR.
0.05

0
ASSUMPTION VS. INCLUSIVE DIS
-0.05 VS. INCLUSIVE+ DIMUONS
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

NNPDF (2009)
W+ CHARM
STRANGE TO LIGH FRACTION

AT LO σ PROPORTIONAL TO THE
STRANGE-GLUON LUMI, UP TO
CABIBBO SUPPR.
ASSUMPTION VS.DIS+ TOTAL
WZ VS. DIS+ TOTAL WZ +
W+CHARM

ATLAS (2013)
 THE GLUON
SCALE DEPENDENCE OF FLAVOR SINGLET STRUCTURE FUNCTIONS

d αs (Q2 )
F s (N, Q2 )
dt 2
= 2π
γqq (N )F2s + 2 nf γqg (N )g(N, Q2 ) + O(α2s )
ANOMALOUS DIMENSIONS
 

LARGE x GLUON DIFFICULT TO DETERMINE FROM DEEP-INELASTIC SCATTERING

THE GLUON IN HADRONIC COLLISIONS
THE GLUON ONLY INTERACTS THROUGH QCD
GLUON
JETS

• ONE-JET INCLUSIVE USED TO

CONSTRAIN THE LARGE x GLUON
SINCE TEVATRON
• WIDE KINEMATIC REGION AT LHC

CMS (2014)

GLUON
TOP

• WIDE RAPIDITY RANGE:

CAN ACCESS WIDE x REGION

prosa (LHCb
data) 2015 )
HIGGS
TOTAL CROSS-SECTION • NOT YET
A STANDARD CANDLE
• EXPERIMENTAL ACCU-
RACY
ATLAS+CMS (2016) ALREADY COMPETITIVE
 PDF DETERMINATION SUMMARY
DEEP-INELASTIC SCATTERING PROVIDES THE BULK OF INFORMATION ON PDFS:
±
•
– HERA COLLIDER e p CC+NC DATA PROVIDE FOUR INDEPENDENT
COMBINATIONS IN WIDE KINEMATIC REGION ⇒ LIGHT QUARKS AND ANTIQUARKS
– FIXED-TARGET µp & µd GIVES DIRECT HANDLE ON UP-DOWN SEPARATION,
ESPECIALLY AT LARGER x
– HERA+FT GLUON FROM SCALE DEPENDENCE (“SCALING VIOLATIONS”)
– NEUTRINO (ESPECIALLY DIMUON) ⇒ STRANGENESS
• DRELL-YAN γ ∗ ON FIXED p AND d TARGET ⇒ UP-DOWN SEPARATION AT LARGE x
• W AND Z PRODUCTION AT THE TEVATRON ⇒ ANTIUP/ANTIDOWN

• LHC W , Z HIGH AND LOW MASS

– FULL FLAVOR SEPARATION IN WIDE KINEMATIC REGION
– STRANGENESS BOTH FROM TOTAL CROSS-SECTION AND TAGGED W +c FINAL
STATE
– GLUON FROM Z TRANSVERSE MOMENTUM DISTRIBUTION
• GLUON FROM COLLIDER PROCESSES:
– LARGE x FROM TEVATRON JETS
– SMALL x FROM LHC JETS
– MEDIUM x FROM LHC TOP
METHODOLOGY & STATISTICS
 PARTON FITS
DATA → PARTON DISTRIBUTIONS
NNPDF3.0 NLO dataset

FT DIS
107 HERA1
FT DY

6 TEV EW
10 TEV JET
ATLAS EW
LHCB EW
5
10 LHC JETS
HERA2
ATLAS JETS 2.76TEV
4
10 ATLAS HIGH MASS

T
ATLAS WpT
CMS W ASY
3 CMS JETS
10
CMS WC TOT

Q2 / M2 / p2 [ GeV2 ]
CMS WC RAT
LHCB Z
102
TTBAR

10

1 -6 -5 -3
10 10 10-4 10 10-2 10-1 1
x

ISSUES AND TASKS:

• FROM PHYSICAL OBSERVABLES TO PDFS: SOLVE EVOLUTION EQUATIONS,
CONVOLUTE WITH PARTON-LEVEL CROSS-SECTIONS
• DISENTANGLING PDFS: CHOOSE A BASIS OF PDFS (2Nf QUARKS +1 GLUON) & A SET OF
SUITABLE PHYSICAL PROCESSES TO DETERMINE THEM ALL
• PROBABILITY IN THE SPACE OF FUNCTIONS: CHOOSE A STATISTICAL APPROACH (HESSIAN,
MONTE CARLO, . . . )
• UNCERTAINTY ON FUNCTIONS: CHOOSE A FUNCTIONAL FORM
 THE HESSIAN APPROACH
• CHOOSE A FIXED FUNCTIONAL FORM
– SINCE 1973, PHYSICALLY MOTIVATED ANSATZ fi (x, Q20 ) = xα (1 − x)β gi (x);
√
gi (x) POLYNOMIAL IN x OR x
– MMHT 2015:
¯ S = 2(ū + d)
∗ BASIS FUNCTIONS g; uv = u − ū; dv = d − d; ¯ + s + s̄; s+ = s + s̄; ∆ = d¯ − ū;
s− = s − s̄.
∗ FOR ALL BUT ∆ s− , g ⇒ xfi (x, Q20 ) = Axα (1 − x)β 1 + 4i=1 ai Ti (y(x)) ;
√
P

Ti CHEBYSHEV POLYNOMIALS, y = 1 − 2 x ↔ MUST MAP x = [0, 1] INTO y = [−1, 1];

Ti (−1) = Ti (1) = 1
′
∗ GLUON xg(x, Q20 ) = Axα (1 − x)β 1 + 2i=1 ai Ti (y(x)) + A′ xT α′ (1 − x)β
P

∗ SEA ASYMMETRY x∆(x, Q20 ) = Axα (1 − x)β (1 + γx + ǫx2 )

∗ STRANGENESS ASYMMETRY x∆(x, Q20 ) = Axα (1 − x)β (1 − x/x0 )
∗ 41 PARAMETERS, 4 FIXED BY SUM RULES
∗ 12 PARMS FIXED AT BEST FIT, REMAINING 25 USED FOR (HESSIAN) COVARIANCE MATRIX
• EVOLVE TO DESIRED SCALE & COMPUTE PHYSICAL OBSERVABLES
• DETERMINE BEST -FIT VALUES OF PARAMETERS
• DETERMINE ERROR BY PROPAGATION OF ERROR ON PARMS. (’HESSIAN METHOD’);
PARM. SCANS ALSO POSSIBLE (’LAGR. MULTIPLIER METHOD’)
 HESSIAN ERROR ESTIMATES
GENERAL FEATURES
OBSERVABLE X DEPENDING ON PARAMETERS ~ z : (LINEAR ERROR PROPAGATION)
X(~z) ≈ X0 + zi ∂i X(~z) ASSUMING MOST LIKELY VALUE AT ~z=0
2
VARIANCE: σX = σij ∂i X∂j X,
σij ⇒ COVARIANCE MATRIX IN PARAMETER SPACE
MAXIMUM LIKELIHOOD: COVARIANCE ⇔ HESSIAN σij = ∂i ∂j χ2 EVALUATED AT MIN. OF χ2
DIAGONALIZATION: CHOOSE zi AS EIGENVECTORS OF σij WITH UNIT EIGENVALUES
2 ~ 2 (LENGTH OF GRADIENT)
σX = |∇X|
SOME INTERESTING CONSEQUENCES
THE ONE-σ CONTOUR IN PARAMETER SPACE IS
2 2
•
ELLIPSE χ = χmin + 1
• THE TOTAL UNCERTAINTY IS THE SUM IN QUADRATURE
OF UNCERTAINTIES DUE TO EACH PARAMETER (LENGTH
OF VECTOR) EVEN WHEN NOT DIAGONALIZING (Lai et al,
CTEQ 2010)
• ANY ROTATION (ORTHOGONAL TRANSF.) IN THE SPACE
OF PARMS PRESERVES THE GRADIENT → CAN DIAGONAL-
IZE A CHOSEN OBSERVABLE WITHOUT SPOILING RESULT
(Pumplin 2009)
 THE HESSIAN UNCERTAINTY
“PARADOX”

THE STANDARD DEVIATION OF FOR

√
• χ2
Ndat DATA σχ2 = 2Ndat
2 2 2 2
HYPOTESIS-TESTING RANGE: COMPARE ∆χ = χ − hχ i TO σχ2 .
IF TOO LARGE, SOMETHING WRONG WITH THEORY (OR DATA)
2 2
• BUT THE ONE-σ RANGE FOR A PARM. OF THE THEORY IS THE CURVE χ − χmin =1
PARAMETER-FITTING RANGE: UNIT DEVIATION FROM THE PARAMETRIC MINIMUM
χ2min
WHY?

• CONSIDER DEVIATIONS ∆i FROM LINEAR FIT

y = x + k; DETERMINE INTERCEPT k AS FREE
PARAMETER
• IF STANDARD DEVIATION FOR EACH ∆i IS σ∆ ,
THEN AVERAGE SQUARE DEVIATION IN UNITS OF
σ∆ FOR Ndat DATA: σχ2 = Ndat
• BEST -FIT INCTERCEPT: k = h∆i i
σ∆
• UNCERTAINTY ON IT: σk = Ndat

• IF ∆k = σk , THEN ∆χ2 = 1
 •
•

2
Distance = ∆χglobal
0
5

-20
-15
-10
-5
10
15
20

BCDMS µ p F
2
BCDMS µ d F
2
NMC µ p F
2
NMC µ d F
2
NMC µ n/µ p
E665 µ p F
2
REPRODUCED

E665 µ d F
2
RESCALE ∆χ2

SLAC ep F
Eigenvector number 13

2
SLAC ed F 2
NMC/BCDMS/SLAC FL
• (MSTW/MMHT)

E866/NuSea pp DY
=T

E866/NuSea pd/pp DY
NuTeV ν N F2
CHORUS ν N F2
NuTeV ν N xF
3
CHORUS ν N xF3
CCFR ν N→µµ X
NuTeV ν N→µµ X
H1 ep 97-00 σNC
r
ZEUS ep 95-00 σrNC
UNREALISTICALLY SMALL

H1 ep 99-00 σrCC
INDIVIDUAL EXPERIMENTS

ZEUS ep 99-00 σrCC

charm
H1/ZEUS ep F 2
H1 ep 99-00 incl. jets
ZEUS ep 96-00 incl. jets
D∅ II pp incl. jets
MSTW 2008 NLO PDF fit

CDF II pp incl. jets

D∅ II W →lν asym.
CDF II W →lν asym.
D∅ II Z rap.
CDF II Z rap.
MSTW TOLERANCE PLOT FOR 13TH EIGENVEC.

90% C.L.
68% C.L.
68% C.L.
90% C.L.

2
Tolerance T = ∆χglobal
0
5

-20
-15
-10
-5
10
15
20

1
H1 ep 97-00 σ rNC H1 ep 97-00 σ rNC
TOLERANCE

NMC µ d F

2
NuTeV ν N→µ µ X 2

3
CCFR ν N→µ µ X NuTeV ν N→µ µ X

4
E866/NuSea pd/pp DY E866/NuSea pd/pp DY

5
GLOBAL

NuTeV ν N xF 3 NuTeV ν N→µ µ X

6
NuTeV ν N→µ µ X NuTeV ν N→µ µ X
IN GLOBAL HESSIAN FITS, UNCERTAINTITES OBTAINED BY

BCDMS µ d F

7
2 D ∅ II W →lν asym.
BCDMS µ p F BCDMS µ d F

LIMIT FOR THE DISTRIBUTION OF BEST -FITS OF EACH EXPERIMENT

2 2
ZEUS ep 95-00 σ rNC H1 ep 97-00 σ rNC
SLAC ed F 2 BCDMS µ d F
2
ZEUS ep 95-00 σ rNC H1 ep 97-00 σ rNC
E866/NuSea pd/pp DY E866/NuSea pd/pp DY
∆χ2 = 1

NuTeV ν N xF 3 E866/NuSea pp DY
D ∅ II W →lν asym. NMC µ d F
2

MSTW 2008 NLO PDF fit

NuTeV ν N F 2 H1 ep 97-00 σ rNC

MSTW TOLERANCE
E866/NuSea pd/pp DY CCFR ν N→µ µ X

INTERVAL SUCH THAT CORRECT CONFIDENCE INTERVALS ARE

CCFR ν N→µ µ X NuTeV ν N→µ µ X
E866/NuSea pd/pp DY D∅ II W → lν asym.
H1 ep 97-00 σ rNC H1 ep 97-00 σ rNC
NuTeV ν N→µ µ X NuTeV ν N xF 3

9 10 11 12 13 14 15 16 17 18 19 20
UNCERTAINTIES TUNED TO DISTRIBUTION OF DEVIATIONS FROM BEST -FITS FOR

Eigenvector number

FOR EACH EIGENVECTOR IN PARAMETER SPACE DETERMINE CONFIDENCE

- 50 (MRST)
+ 50 (MRST)

- 100 (CTEQ)
+ 100 (CTEQ)
 PARAMETRIZATION BIAS?

• IF PARM. NOT GENERAL ENOUGH, GLOBAL MIN. IS NOT TRUE MIN.

• ONE-σ VARIATION ABOUT FAKE MIN CORRESP. TO LARGE χ2 VARI-
ATION
• IN DATA REGION, UNCERTAINTY CAN BE TUNED TO DATA, BUT
WHAT ABOUT INTERPORATION/ESTRAPOLATION?

(Pumplin, 2009)

UNBIASED BASES
• OLD IDEA (Parisi, Sourlas, 1978):EXPAND PDFS OVER BASIS OF ORTHOGONAL POLYNOMIALS,
WITHOUT ANY FURTHER ASSUMPTION
• DIFFICULT TO AVOID SPURIOUS FLUCTUATIONS
• MUST IMPOSE LENGTH PENALTY TO STABILIZE THE FIT (ANOTHER BIAS?)

(Glazov, Radescu,
2009)
 THE MONTE CARLO METHOD
BASIC IDEA: MONTE CARLO SAMPLING
OF THE PROBABILITY MEASURE IN THE (FUNCTION) SPACE OF PDFS

Fi i=1,...,Ndata
• GENERATE A SET OF MONTE CARLO REPLICAS Experimental Data
NMC,BCDMS,SLAC,HERA,CHORUS...
σ (k) OF THE ORIGINAL DATASET σ (data)
⇒ REPRESENTATION OF P[σ] AT DISCRETE SET
OF POINTS IN DATA SPACE
MC generation Fi(1) Fi(2) Fi(N-1) Fi(N)
• FIT A
PDF REPLICA TO A DATA REPLICA
(k)
⇒ EACH PDF REPLICA fi IS A BEST -FIT PDF
SET FOR GIVEN DATA REPLICA TRAINING

• THE SET OF NEURAL NETS IS A REPRESENTATION

EVOLUTION
OF THE PROBABILITY DENSITY:

Nrep NN parametrization q0net(1) q0net(2) q0net(N-1) q0net(N)

1 (k)
hfi i = fi
Nrep k=1 REPRESENTATION OF
PROBABILITY DENSITY
X
 MONTE CARLO ERROR ESTIMATES
EXACT ERROR PROPAGATION
OBSERVABLE X DEPENDS ON PARAMETERS ~
z
2 2 2
VARIANCE: σX = hX i − hXi
AVERAGES:
R
hXi = dd zX(~z)P (~z), WITH
P (~z) ⇒ PROBABILITY DISTRIBUTION OF PARAMETER VALUES
IMPORTANCE SAMPLING
• SPACE OF FUNCTIONS HUGE
5 BINS FOR 10 PTS× 7 FCTNS → 570 ∼ 1049 BINS
• BUT EACH OBSERVABLE DEPENDS ONLY ON ONE PARAMETER, & OBSERVABLES CORRELATED
⇒ DATA TELL US WHICH BINS ARE POPULATED
replica averages
vs.NNPDF1.2 Central values
central -values replica standard dev.
100 vs. NNPDF1.2 - Errors
uncertainties
Nrep=10 100
Nrep=100 Nrep=10
Nrep=1000 Nrep=100
10 10 Nrep=1000

1
1
0.1
0.1
0.01

Monte Carlo replicas

0.01 0.001

0.0001
0.001 0.0001 0.001 0.01 0.1 1 10
0.001 0.01 0.1 1 10 100 Experimental data
Experimental data

10 REPLICAS ENOUGH FOR CENTRAL VALS, 100 FOR UNCERTAINTIES, 1000 FOR CORRELNS
 FLEXIBLE PARAMETRIZATION
• EACH PDF REPLICA FITTED TO A DATA REPLICA
⇒ NEED BEST -FIT, BUT NOT COVARIANCE MATRIX IN PARAMETER SPACE
• CAN USE VERY LARGE PARAMETRIZATION

NEURAL NETWORKS

MULTILAYER FEED-FORWARD NETWORKS

• Each neuron receives input from neurons
in preceding layer and feeds output to neu-
rons in subsequent layer
• Activation determined by weights and
thresholds
1
ξi = g j ωij ξj − θi
0.8


0.6 • Sigmoid activation function

P

0.4 g(x) = 1+e1−βx

0.2

-10 -5 5 10

EXAMPLE: A 1-2-1 NN
1
f (x) = (2) (2)
(3) ω11 ω12
(2) (1) (2) (1)
θ1 − −
1+e 1+e
θ1 −xω11 θ2 −xω21
1+e
THANKS TO NONLINEAR BEHAVIOUR,
ANY FUNCTION CAN BE REPRESENTED BY A SUFFICIENTLY BIG NEURAL
NETWORK
 LEARNING
• ONE CAN CHOOSE A HIGHLY REDUNDANT PARAMETRIZATION
EXAMPLE: NNPDF: 2 − 5 − 3 − 1 NN FOR EACH PDF: 37 × 7 = 259 PARAMETERS

• MINIMIZATION (“LEARNING”) CAN BE PERFORMED USING GENETIC ALGORITHMS

• COMPLEXITY INCREASES AS THE FITTING PROCEEDS

• ⇒ THE BEST FIT IS NOT THE ABSOLUTE MINIMUM:

MUST LOOK FOR OPTIMAL LEARNING POINT

UNDERLEARNING
 LEARNING
• ONE CAN CHOOSE A HIGHLY REDUNDANT PARAMETRIZATION
EXAMPLE: NNPDF: 2 − 5 − 3 − 1 NN FOR EACH PDF: 37 × 7 = 259 PARAMETERS

• MINIMIZATION (“LEARNING”) CAN BE PERFORMED USING GENETIC ALGORITHMS

• COMPLEXITY INCREASES AS THE FITTING PROCEEDS

• ⇒ THE BEST FIT IS NOT THE ABSOLUTE MINIMUM:

MUST LOOK FOR OPTIMAL LEARNING POINT

PROPER LEARNING
 LEARNING
• ONE CAN CHOOSE A HIGHLY REDUNDANT PARAMETRIZATION
EXAMPLE: NNPDF: 2 − 5 − 3 − 1 NN FOR EACH PDF: 37 × 7 = 259 PARAMETERS

• MINIMIZATION (“LEARNING”) CAN BE PERFORMED USING GENETIC ALGORITHMS

• COMPLEXITY INCREASES AS THE FITTING PROCEEDS

• ⇒ THE BEST FIT IS NOT THE ABSOLUTE MINIMUM:

MUST LOOK FOR OPTIMAL LEARNING POINT

OVERLEARNING
 CROSS-VALIDATION
GENETIC MINIMIZATION:
2
AT EACH GENERATION, χ EITHER UNCHANGED OR DECREASING

• DIVIDE THE DATA IN TWO SETS: TRAINING AND VALIDATION

• MINIMIZE THE χ2 OF THE DATA IN THE TRAINING SET

• AT EACH ITERATION, COMPUTE THE χ2 FOR THE DATA IN THE VALIDATION SET
(NOT USED FOR FITTING)
• WHEN THE VALIDATION χ2 STOPS DECREASING, STOP THE FIT
 CROSS-VALIDATION
GENETIC MINIMIZATION:
2
AT EACH GENERATION, χ EITHER UNCHANGED OR DECREASING

• DIVIDE THE DATA IN TWO SETS: TRAINING AND VALIDATION

• MINIMIZE THE χ2 OF THE DATA IN THE TRAINING SET

• AT EACH ITERATION, COMPUTE THE χ2 FOR THE DATA IN THE VALIDATION SET
(NOT USED FOR FITTING)
• WHEN THE VALIDATION χ2 STOPS DECREASING, STOP THE FIT

GO!
 CROSS-VALIDATION
GENETIC MINIMIZATION:
2
AT EACH GENERATION, χ EITHER UNCHANGED OR DECREASING

• DIVIDE THE DATA IN TWO SETS: TRAINING AND VALIDATION

• MINIMIZE THE χ2 OF THE DATA IN THE TRAINING SET

• AT EACH ITERATION, COMPUTE THE χ2 FOR THE DATA IN THE VALIDATION SET
(NOT USED FOR FITTING)
• WHEN THE VALIDATION χ2 STOPS DECREASING, STOP THE FIT

STOP!
 CROSS-VALIDATION
MINIMIZE BY GENETIC ALGORITHM:
2
AT EACH GENERATION, χ EITHER UNCHANGED OR DECREASING

• DIVIDE THE DATA IN TWO SETS: TRAINING AND VALIDATION

• MINIMIZE THE χ2 OF THE DATA IN THE TRAINING SET

• AT EACH ITERATION, COMPUTE THE χ2 FOR THE DATA IN THE VALIDATION SET
(NOT USED FOR FITTING)
• WHEN THE VALIDATION χ2 STOPS DECREASING, STOP THE FIT

TOO LATE!
 MC ⇔ HESSIAN
2
xg
0
• TO CONVERT HESSIAN INTO MONTECARLO -2
GENERATE MULTIGAUSSIAN REPLICAS IN PA- -4 At input scale Q 2 = 1 GeV 2
0
RAMETER SPACE 2
MSTW 2008 NLO ( ∆ χ = 1)
-6
• ACCURATE WHEN NUMBER OF REPLICAS 40 individual MC replicas

SIMILAR TO THAT WHICH REPRODUCES DATA -8 MC average and s.d.

-5 -4 -3 -2 -1
10 10 10 10 10 1
x
(Thorne, Watt, 2012)

• TO CONVERT MONTE CARLO INTO HESSIAN, SAMPLE

THE REPLICAS fi (x) AT A DISCRETE SET OF POINTS &
CONSTRUCT THE ENSUING COVARIANCE MATRIX
• EIGENVECTORS OF THE COVARIANCE MATRIX AS A BA-
SIS IN THE VECTOR SPACE SPANNED BY THE REPLICAS
BY SINGULAR-VALUE DECOMPOSITION
• NUMBER OF DOMINANT EIGENVECTORS SIMILAR TO
NUMBER OF REPLICAS ⇒ ACCURATE REPRESENTATION

(Carrazza, SF, Kassabov, Rojo, 2015)

 COMPRESSION
Correlations for NNPDF3.0 NLO
MONTECARLO
Correlations for NNPDF3.0 NLO
Prior Nrep=1000 @ Q=100 GeV Compressed Nrep=50 @ Q=100 GeV
1.0 1.0
s̄ s̄
• CONSTRUCT A VERY LARGE REPLICA SAMPLE
0.8 0.8
ū 0.6 ū 0.6
• SELECT (BY GENETIC ALGORITHM) A SUBSET OF
REPLICAS WHOSE STATISTICAL FEATURES ARE
0.4 0.4
d̄ d̄ AS CLOSE AS POSSIBLE TO THOSE OF THE PRIOR
0.2 0.2
g g
0.0 0.0 • ⇒ FOR ALL PDFS ON A GRID OF POINTS// MIN-
d −0.2 d −0.2 IMIZE DIFFERENCE OF: FIRST FOUR MOMENTS,
−0.4 −0.4 CORRELATIONS; OUTPUT OF KOLMOGOROV-
u u
−0.6 −0.6 SMIRNOV TEST (NUMBER OF REPLICAS BETWEEN
s −0.8 s −0.8
MEAN AND σ, 2σ, INFINITY)
−1.0 −1.0
s̄ ū d̄ g d u s s̄ ū d̄ g d u s
• 50 COMPRESSED REPLICA REPRODUCE 1000
REPLICA SET TO PRECENT ACCURACY
(Carrazza, Latorre, Kassabov, Rojo, 2015)
CAN REPRODUCE NONGAUSSIAN FEATURES WITH REASONABLY SMALL REPLICA SAMPLE
HESSIAN
• SELECT SUBSET OF THE COVARIANCE MATRIX
CORRELATED TO A GIVEN SET OF PROCESSES
• PERFORM SVD ON THE REDUCED COVARI-
ANCE MATRIX, SELECT DOMINANT EIGENVEC-
TOR, PROJECT OUT ORTHOGONAL SUBSPACE
• ITERATE UNTIL DESIRED ACCURACY REACHED
• CAN ADD PROCESSES TO GIVEN SET; CAN COM-
BINE DIFFERENT OPTIMIZED SETS
• 15 EIGENVECTORS DESCRIBE ALL HIGGS
MODES + JETS + W , Z PRODUCTION
(Carrazza, SF, Kassabov, Rojo, 2016)
VERY SMALL NUMBER OF EVECS; CAN COMBINE WITH NUISANCE PARMS
 NONGAUSSIAN BEHAVIOUR

MONTE CARLO COMPARED TO HESSIAN

CMS W + c production

• DEVIATION FROM GAUSSIANITY E.G. AT

LARGE x DUES TO LARGE UNCERTAINTY
+ POSITIVITY BOUNS ⇒ RELEVANT FOR
SEARCHES
• CANNOT BE REPRODUCED IN HESSIAN
FRAMEWORK

• WELL REPRODUCED BY COMPRESSED MC

•
∞ ln P (x)
DKL = −∞
P (x) ln Q(x)
dx
BETWEEN A PRIOR P AND ITS REPRESEN-
DEFINE RKULLBACK-LEIBLER DIVERGENCE

TATION Q
• DKL BETWEEN PRIOR AND HESSIAN DE-
PENDS ON DEGREE OF GAUSSIANITY

• DKL BETWEEN PRIOR AND COMPRESSED

MC DOES NOT

CAN GAUGE WHEN MC IS MORE ADVANTAGEOUS THAN HESSIAN!

 MONTE CARLO COMBINATION
(Watt, S.F., 2010-2013)

• MAY COMBINE DIFFERENT PDF SETS,

AFTER MC CONVERSION OF HESSIAN SETS
• COMBINE MONTE CARLO REPLICAS INTO SINGLE SET
• USEFUL FOR CONSERVATIVE UNCERTAINTY ESTIMATE
• COMBINED SET APPROXIMATELY GAUSSIAN
COMBINED PDF4LHC SETS FOR ANTIDOWN & STRANGE
 THE ULTIMATE CHECK OF PDF DETERMINATION
CLOSURE TESTS

• ASSUME PDFS KNOWN: GENERATE FAKE EXPERIMENTAL DATA

• CAN DECIDE DATA UNCERTAINTY (ZERO, OR AS IN REAL DATA, OR ...)

• FIT PDFS TO FAKE DATA

• CHECK WHETHER FIT REPRODUCES UNDERLYING “TRUTH”:

– CHECK WHETHER TRUE VALUE GAUSSIANLY DISTRIBUTED ABOUT FIT

– CHECK WHETHER UNCERTAINTIES FAITHFUL

– TRACE DIFFERENT SOURCES OF UNCERTAINTY

 TRACING SOURCES OF UNCERTAINTY
• LEVEL 0: FAKE DATA GENERATED WITH NO UNCERTAINTY
→ INTERPOLATION AND EXTRAPOLATION UNCERTAINTY
• LEVEL 1-2: FAKE DATA GENERATED WITH SAME UNCERTAINTY AS REAL
DATA (INCLUDING CORRELATIONS)
• LEVEL 1: NO PSEUDODATA REPLICAS:
⇒ REPLICAS FITTED TO SAME DATA OVER AND OVER AGAIN
→ FUNCTIONAL UNCERTAINTY DUE TO INFINITY OF EQUIVALENT MINIMA
• LEVEL 2: STANDARD NNPDF METHODOLOGY
⇒ REPLICAS FITTED TO PSEUDODATA REPLICAS
→ DATA UNCERTAINTY
• THREE SOURCES OF UNCERTAINTY COMPARABLE IN DATA REGION
THE GLUON: LEVEL 0, LEVEL 1 AND LEVEL 2
Ratios of gluon at different closure test levels Ratios of gluon at different closure test levels
4 4

3.5 Lvl0 Closure Fit 3.5 Lvl0 Closure Fit

Lvl1 Closure Fit Lvl1 Closure Fit
3 3
Lvl2 Closure Fit Lvl2 Closure Fit
2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0

-0.5 -0.5

-1 -5 -3
-1
10 10-4 10 10-2 10-1 1 0 0.2 0.4 0.6 0.8 1
x x
 FITTING EFFICIENCY
LEVEL 0 THE GLUON
ASSUME VANISHING Level 0 closure test vs. MSTW
7
•
EXPERIMENTAL UNCERTAINTY
6 Level 0 Closure Test Fit
2
• MUST BE ABLE TO GET χ =0 5 MSTW2008nlo68cl

4
UNCERTAINTY AT DATA POINTS TENDS TO ZERO
3
•
(NOT NECESSARILY ON PDF!)
2

xg(x,Q2)
DEFINE φ ≡ hχ2rep i − χ2 , 1

EQUALS FIT UNCERTAINTY/DATA UNCERTAINTY ; CHECK 0

q

-1
-2
φ→0
• CAN STUDY EFFICIENCY OF MINIMIZATION 10-5 10-4 10-3 x 10-2 10-1 1

FRACTIONAL UNCERTAINTY VS TRAINING LENGTH

χ2 VS TRAINING LENGTH
Effectiveness of Genetic Algorithm in Level 0 Closure Tests
Effectiveness of Genetic Algorithms in Level 0 Closure Tests

Old (2.3) genetic algorithm 0.45

10-1 NNPDF3.0 GA settings
New genetic algorithm 0.4

0.35 NNPDF2.3 GA settings

0.3
10-2
0.25

χ2
0.2
ϕ [ T [f fit ] ]

10-3 0.15

0.1

0.05

10-4 0
103 104 105 3 5
10 104 10
Number of Generations Number of GA Generations
 TESTING THE PDF DETERMINATION
LEVEL 2: THE WORKS
THE GLUON: FITTED/”TRUE” • CENTRAL VALUES:
Ratio of Closure Test g to MSTW2008
4 COMPARE FITTED VS. “TRUE” χ2
3.5 BOTH FOR INDIVIDUAL EXPERIMENTS
3

2.5
& TOTAL DATASET

(x,Q2)
2
FOR TOTAL ∆χ2 = 0.001 ± 0.003

MSTW
1.5

1
• UNCERTAINTIES: DISTRIBUTION OF DEVIA-
0.5
TIONS BETWEEN FITTED AND “TRUE” PDFS

CT
g (x,Q2) / g
0 SAMPLED AT 20 POINTS BETWEEN 10−5 AND 1
-0.5 FIND 0.699% FOR ONE-SIGMA,
-1
0 0.2 0.4 0.6 0.8 1 0.948% FOR TWO-SIGMA C.L.
x
NORM. DISTRIBUTION OF DEVIATIONS
LEVEL-2 FITTED χ2 VS “TRUE”
Distribution of χ2 for experiments
Distribution of single replica fits in level 2 uncertainties
Closure test χ2
MSTW2008nlo χ2
2 Closure test central χ2 200 Replica distribution
MSTW2008nlo central χ2
1.8 180 Gaussian distribution

1.6 160
1.4 140
1.2 120

χ2
1 100
Entries

0.8
80
0.6
60
0.4
40
0.2
20
0 NM SLA BCD C N H Z H H D D C D0 ATL C LHC TO P
C C MS HORU TVDMNERA 1AEUSHE 1HERA ERAF2 YE886 YE605 DF AS MS B
S V RA2 2 CHA 0
RM
Experiments -5 -4 -3 -2 -1 0 1 2 3 4 5
Difference to theory (σ)
 METHODOLOGY SUMMARY
• PDF DETERMINATION: HESSIAN METHOD
– SIMPLE LINEAR ERROR PROPAGATION
– TOLERANCE REQUIRED FOR REALISTIC UNCERTAINTIES
– PARAMETRIZATION BIAS POSSIBLE
• PDF DETERMINATION: MONTE CARLO METHOD
– TWO-STEP PROCEDURE: DATA MONTE CARLO ⇒ PDF MONTE CARLO
– VERY GENERAL PARAMETRIZATION ALLOWED
– NEED OPTIMAL FIT DETERMINATION METHOD (CROSS-VALIDATION)
• PDF REPRESENTATION: HESSIAN VS MONTE CARLO
– CONVERSION POSSIBLE EITHER WAY
– COMPRESSION METHODS AVAILABLE EITHER WAY
– MONTE CARLO VERY FLEXIBLE, HESSIAN VERY EFFICIENT
• PDF VALIDATION: CLOSURE TEST
– PERFORMED IN THE MONTE CARLO APPROACH
– INTERPOLATION & FUNCTIONAL UNCERTAINTIES SIGNIFICANT
THEORY: ISSUES & PROGRESS
 PERTURBATIVE STABILITY I
LO VS. NLO VS. NNLO PDFS
GLUON SINGLET
NNPDF3.0, αS = 0.118, Q 2 = 2 GeV2 NNPDF3.0, αS = 0.118, Q 2 = 2 GeV2
7 5
LO LO
6 4.5
NLO NLO
5 4
NNLO NNLO
3.5
4
3
3
2.5
2
2

x g ( x, Q2)
x Σ ( x, Q2)
1
1.5
0 1
-1 0.5
-2 -5 0
-3 -2 -5 -3 -2
10 10-4 10 x 10 10-1 10 10-4 10 x 10 10-1

TRIPLET VALENCE
NNPDF3.0, αS = 0.118, Q 2 = 2 GeV2 NNPDF3.0, αS = 0.118, Q 2 = 2 GeV2
0.7 2
LO LO
0.6
NLO NLO
0.5 NNLO 1.5 NNLO
0.4

0.3
x V ( x, Q2) 1

x T3 ( x, Q2)
0.2
0.5
0.1

0
0
-0.1
-5 -3 -2 -5 -3 -2
10 10-4 10 x 10 10-1 10 10-4 10 x 10 10-1

• PERTURBATIVE ACCURACY OF PREDICTION LIMITED BY PERTURBATIVE ACCURACY OF PDF

• αs (Mz ) ∼ 0.1, αs (Mp ) ∼ 1/2; αs (Q21 ) = αs (Q22 )(1 + O(α2s ))
⇒ LO: QUALITATIVE; NLO: QUANTITATIVE; NNLO: PRECISION
 PERTURBATIVE STABILITY II
THEORY UNCERTAINTIES VS PDF UNCERTAINTIES
LO VS NLO VS LO-NLO SHIFT
GLUON SINGLET
NNPDF3.0, αS = 0.118, Q2 = 104 GeV2 NNPDF3.0, αS = 0.118, Q2 = 104 GeV2
0.3 0.15
LO 1-sigma PDF uncertainty LO 1-sigma PDF uncertainty
0.25
NLO 1-sigma PDF uncertainty NLO 1-sigma PDF uncertainty
0.2 0.1
( NLO - LO ) / NLO ( NLO - LO ) / NLO
0.15
0.1 0.05
0.05
0 0
−0.05

∆( Σ ( x, Q2) ) / Σ ( x, Q2)

∆( g ( x, Q2) ) / g ( x, Q2)
−0.1 −0.05
−0.15
−0.2 −0.1
−1 −1
10−2 x 10 10−2 x 10

NLO VS NNLO VS NLO-NNLO SHIFT

GLUON SINGLET
NNPDF3.0, αS = 0.118, Q2 = 104 GeV2 NNPDF3.0, αS = 0.118, Q2 = 104 GeV2
0.3 0.15
NLO 1-sigma PDF uncertainty NLO 1-sigma PDF uncertainty
0.25
NNLO 1-sigma PDF uncertainty NNLO 1-sigma PDF uncertainty
0.2 ( NNLO - NLO ) / NNLO
0.1 ( NNLO - NLO ) / NNLO
0.15
0.1 0.05
0.05
0 0
−0.05
∆( Σ ( x, Q2) ) / Σ ( x, Q2)

∆( g ( x, Q2) ) / g ( x, Q2)
−0.1 −0.05
−0.15
−0.2 −0.1
−1 −1
10−2 x 10 10−2 x 10

• “DATA” PDF UNCERTAINTY INDEP. OF PERTURBATIVE ORDER

• TH. UNCERTAINTY (MHOU) VS DATA UNCERTAINTY ⇒ LO: DOMINANT; NLO, COMPARABLE;
NNLO: SUBDDOMINANT
 THEORETICAL UNCERTAINTIES ON PDFS:
• PDFS ARE DETERMINED BY COMPARING TO DATA THEORY AT SOME FINITE ORDER

• AFFECTED BY THEORETICAL UNCERTAINTY JUST LIKE HARD CROSS-SECTIONS

• NOT INCLUDED IN CURRENT “PDF UNCERTAINTY”

(ACCOUNTS ONLY DATA & METHODOLOGY)
CAN WE ESTIMATE THEM?
• SCALE VARIATION DIFFICULT:
CORRELATED BETWEEN PROCESSES? HOW DOES IT CORRELATE WITH PROCESSES
IN WHICH PDFS ARE USED?

• AT NLO: WE KNOW THE SHIFT TO NNLO

• AT NNLO: LOOK AT THE BEHAVIOUR OF THE PERTURBATIVE EXPANSION
(CACCIARI-HOUDEAU)
 THEORETICAL UNCERTAINTIES
NLO PDF UNC. VS NLO-NNLO SHIFT VS NLO CACCIARI-HOUDEAU (NNPDF2.1)
UP DOWN STRANGE
Relative Errors, Q2 = 104 GeV2 Relative Errors, Q2 = 104 GeV2 Relative Errors, Q2 = 104 GeV2
1.25 1.25
NNPDF2.1 - NLO PDFunc NNPDF2.1 - NLO PDFunc 2 NNPDF2.1 - NLO PDFunc
1.2 1.2 1.8
NNPDF2.1 - (NNLO-NLO) NNPDF2.1 - (NNLO-NLO) NNPDF2.1 - (NNLO-NLO)
1.15 1.15 1.6
NNPDF2.1 - NLO THunc NNPDF2.1 - NLO THunc NNPDF2.1 - NLO THunc
1.4
1.1 1.1
1.2

R( xu )
R( xd )
R( xs )
1.05 1.05
1

1 1 0.8

0.6
0.95 0.95
0.4
0.9 -5 -3
0.9 -5 -3 -5 -3
10 10-4 10 10-2 10-1 10 10-4 10 10-2 10-1 10 10-4 10 10-2 10-1
x x x

ANTIUP ANTIDOWN ANTISTRANGE

Relative Errors, Q2 = 104 GeV2 Relative Errors, Q2 = 104 GeV2 Relative Errors, Q2 = 104 GeV2
1.25 1.25
NNPDF2.1 - NLO PDFunc NNPDF2.1 - NLO PDFunc 2 NNPDF2.1 - NLO PDFunc
1.2 1.2 1.8
NNPDF2.1 - (NNLO-NLO) NNPDF2.1 - (NNLO-NLO) NNPDF2.1 - (NNLO-NLO)
1.15 1.15 1.6
NNPDF2.1 - NLO THunc NNPDF2.1 - NLO THunc NNPDF2.1 - NLO THunc
1.4
1.1 1.1
1.2
1.05 1.05
R( xsbar )

R( xubar )
R( xdbar )
1

1 1 0.8

0.6
0.95 0.95
0.4
0.9 -5 -3
0.9 -5 -3 -5 -3
10 10-4 10 10-2 10-1 10 10-4 10 10-2 10-1 10 10-4 10 10-2 10-1
x x x

GLUON CHARM BOTTOM

Relative Errors, Q2 = 104 GeV2 Relative Errors, Q2 = 104 GeV2 Relative Errors, Q2 = 104 GeV2
1.25 1.25 1.25
NNPDF2.1 - NLO PDFunc NNPDF2.1 - NLO PDFunc NNPDF2.1 - NLO PDFunc
1.2 1.2 1.2
NNPDF2.1 - (NNLO-NLO) NNPDF2.1 - (NNLO-NLO) NNPDF2.1 - (NNLO-NLO)
1.15 1.15 1.15
NNPDF2.1 - NLO THunc NNPDF2.1 - NLO THunc NNPDF2.1 - NLO THunc

1.1 1.1 1.1

R( xg )
R( xc )
R( xb )

1.05 1.05 1.05

1 1 1

0.95 0.95 0.95

0.9 -5 -3
0.9 -5 -3
0.9 -5 -3
10 10-4 10 10-2 10-1 10 10-4 10 10-2 10-1 10 10-4 10 10-2 10-1
x x x

CACCIARI-HOUDEAU PROMISING?
 HIGHER ORDERS: THEORETICAL PROGRESS

(G. Salam, 2016)

• IMPRESSIVE ACCELERATION OF PERTURBATIVE COMPUTATIONS SINCE INCEPTION OF LHC

• ⇒ HANDLE ON THEORETICAL UNCERTAINTIES
 BEYOND FIXED ORDER: RESUMMATION
EIKONAL EMISSION

EMISSION OF A SOFT (q µ → 0) GAUGE PARTICLE FROM EXTERNAL LINE

epµ
σ(α → β) → σ(α → β) p·q−iǫ
(Bloch, Nordiseck, 1937; Yenni, Frautschi, Suura, 1955; Weinberg, 1964)

• SOFT EMISSION ⇒ EIKONAL FACTOR

• CROSS SECTION FOR SINGLE (DOUBLE. . . ) EMISSION
INFRARED DIVERGENT; DIVERGENCE CANCELLED BY VIRTUAL CORRECTIONS
M 2
• 1,2,. . . ,N EMISSIONS EXPONENTIATE ⇒ Γ ∼ exp − α ln2 1 − sβ
  

(Sudakov, 1956)

• LEFTOVER
M 2
σ(α → β) → σ(α → β) ln2 1 − sβ
AFTER CANCELLATION,   SOFT LOGS:
 RESUMMATION:
EXPONENTIATION
• EXPONENTIATION OF LEFTOVER LOGS
⇒
M2
THRESHOLD RESUMMATION OF αs ln2 (1 − x), x = s
• LOGS COME IN PAIRS: SOFT+COLLINEAR → ln pt WHEN INTEGRAL OVER pt NOT PERFORMED⇒
2
qT
TRANSVERSE MOMENTUM RESUMMATION OF αs ln2 M2
• IN GLUON CHANNEL SYMMETRY OF THE TRIPLE GLUON VERTEX → LARGE LOGS ALSO WHEN
s
THE EXCHANGED GLUON IS SOFT: NO COLLINEAR CONTRIBUTION, SINGLE LOGS → ln 2 ⇒
Q
1
HIGH ENERGY RESUMMATION OF αs ln x

GLUON RADIATION
x dkt2
σ(τ, M 2 ) = y1 dy
y
P y µ2 kt2
σ̂(y, M 2 )
R   R (s−M 2 )2 /s

x
THE GLUON SPLITTING FUNCTION:
x 1−x 1−x
Pgg (x) = 2CA (1−x)+
+ x
+ x(1 − x) + β0 δ(1 − x)
h i

LOGARITHMICALLY ENHANCED TERMS

1
• INFRARED LOGS: τ
dy 1−y
+
∼ ln(1 − τ )
R1

• UV LOGS: τ
dy y1 ∼ ln(τ )
R1

2
dkt Q2 (1−τ )2 Q2
• COLLINEAR LOGS:
µ2 kt2 ∼ ln µ2 τ = ln µ2
+ ln(1 − τ )2 + ln τ
 
R (s−M 2 )2 /s
 FACTORIZATION REMINDER
THE FACTORIZED CROSS SECTION

dz τ M2 M2 M2
σ(τ, M 2 ) = τ ij τ z Lij z
, µ2 1 σ̂
F z ij z, M 2 , αs (µ2R ), µ2
, µ2
τ = s
F R
 
P R1

PARTON LUMINOSITIES
dx z
Lij (z, µ2 ) = z x fi x
, µ2 fj (x, µ2 )
COEFFICIENT FUNCTIONS
R1

M2 M2 M2 M2
σ̂ij z, M 2 , αs (µ2R ), µ2
, µ2
= z σ0 M 2 , αs (µ2R ) Cij z, αs (µ2R ), µ2
, µ2
F R F R
   

(1) (2) (3)

Cij (z, αs ) = δ(1 − z)δig δjg + αs Cij (z) + α2s Cij (z) + α3s Cij (z) + O(α4s )

MELLIN-SPACE FACTORIZATION

σ(N, M 2 ) = σ0 M 2 , αs L(N )C(N, αs ),

σ(N, M 2 ) ≡ 0
dτ τ N −2 σ(τ, M 2 ); L(N ) ≡ 0
dz z N −1 L(z) C(N, αs ) ≡ 0
dz z N −1 C(z, αs )


R1 R1 R1
 THE STRUCTURE OF RESUMMED EXPRESSIONS
THRESHOLD RESUMMATION
LOG COUNTING
Cres (N, αs ) = g0 (αs ) exp α1 g1 (αs ln N ) + g2 (αs ln N ) + αs g3 (αs ln N ) + . . . ;
s
2 k FOR
h i

g0 (αs ) = 1 + αs g0,1 + αs g0,2 + O(α3s ); g1 (λ) = ∞ g

k=2 1,k λk , g (λ) =
i k=1 gi,k λ
i≥2
P P∞

LOG APPROX. XSECT ACCURACY EXP. ACCURACY: αsn Lk g0 ACCURACY: αsi

LL k = 2n k =n+1 0
NLL 2n − 2 ≤ k ≤ 2n k=n 1
NNLL 2n − 4 ≤ k ≤ 2n k =n−1 2

THE RESUMMED EXPONENT

2 M2 2 M2
S M , N2 = M2 µ2 γ̄ αs (µ ), N 2 µ2
 
R M 2 /N 2 dµ2  

M /N
= M2 µ2 −A(αs (µ2 )) ln µ2 + B[αs (µ2 )] .
R M 2 /N 2 dµ2 h  2 2 i

• A, B ARE POWER SERIES IN αs

• A IS UNIVERSAL, COEFFICIENT OF ln N IN (DIAGONAL QUARK OR GLUON)
ANOMALOUS DIMENSION AT EACH ORDER

• B CONTAINS PROCESS-DEPENDENT TERMS, STARTS AT NLL

 RESUMMED PDFS

• SO FAR NO RESUMMED PDF SETS AVAILABLE

GLUON: NLO VS NLL • PRELIMINARY STUDY: IF
THRESHOLD RESUMMATION INCLUDED IN FIT
(DIS, DY, TOP DATA), EFFECTS
NOT NEGLIGIGLE AT NLLO, LARGE x, MORE MODER-
ATE AT NNLO
• EFFECT ON PDFS COMPARABLE TO EFFECT ON MATRIX
ELEMENT, ANTICORRELATED TO IT
• RELEVANT FOR NEW PHYSICS SEARCHES

Bonvini et al., 2015

SLEPTON PAIR PRODUCTION
GLUON: NNLO VS NNLL HIGGS IN GLUON FUSION VS mH
 THE PHOTON PDF
• THE SAME FACTORIZATION ARGUMENTS APPLY TO QCD AND QED
• THE PROTON ALSO HAS A PHOTON CONTENT ⇒ PHOTON PDF
• PHOTONS RADIATED BY QUARKS;
αs → α;
• QED INDUCED CONTRIBUTIONS TO HADRON COLLIDER PROCESSES
α(Mz ) ∼ αs (Mz )/10 ⇒ NLO QED CORRECTIONS∼ NNLO QCD CORRECTIONS
EXAMPLE: PHOTON-INDUCED DRELL-YAN
γ νℓ

W+
γ ū
W−
νℓ ℓ+

W+
d¯ ℓ+ d¯ ū



 THE PHOTON PDF FROM DATA
NNPDF2.3QED/NNPDF3.0QED DATASET
Dataset Observable Ndat [ηmin , ηmax ] Mll , Mllmax
LHCb γ ∗ /Z Low Mass dσ(Z)/dMll 9 [2,4.5] [5,120] GeV
 min 

ATLAS W, Z dσ(W ± , Z)/dη 30 [-2.5,2.5] [60,120] GeV

ATLAS γ ∗ /Z High Mass dσ(Z)/dMll 13 [-2.5,2.5] [116,1500] GeV

IMPACT
CORRELATION BETWEEN DATA AND γ PDF
Correlation between photon PDF and cross sections Correlation between photon PDF and cross sections Correlation between photon PDF and cross sections

1 1 1

0.5 0.5 0.5

0 0 0

-0.5 -0.5 -0.5

Correlation coefficient
Correlation coefficient
Correlation coefficient

LHCb Low Mass DY ATLAS W,Z ATLAS High Mass DY

-1 -1 -1
-5 -4 -3 -2 -1 -5 -4 -3 -2 -1 -5 -3 -2
10 10 10 x 10 10 10 10 10 x 10 10 10 10-4 10 x 10 10-1
 THE PHOTON PDF
NNPDF2.3QED-NNPDF3.0QED

4
NLO RESULTS
2
Photon PDF comparison at 10 GeV Photon PDF comparison at 104 GeV2
0.16 0.05
MRST2004qed MRST2004qed
0.14 NNPDF2.3 QED RW average NNPDF2.3 QED RW average
NNPDF2.3 QED RW replicas 0.04 NNPDF2.3 QED RW replicas
NNPDF 1σ NNPDF 1σ
0.12
NNPDF 68% c.l. NNPDF 68% c.l.

0.1 0.03

0.08
0.02

xγ (x,Q2)
xγ (x,Q2)
0.06
0.01
0.04

0.02 0
0
-0.01
-5 -3 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
10 10-4 10 10-2 10-1 1
x x
NNLO RESULTS
Photon PDF comparison at 104 GeV2 Photon PDF comparison at 104 GeV2
0.16 0.05
MRST2004qed MRST2004qed
0.14 NNPDF2.3 QED RW average NNPDF2.3 QED RW average
NNPDF2.3 QED RW replicas 0.04 NNPDF2.3 QED RW replicas
NNPDF 1σ NNPDF 1σ
0.12
NNPDF 68% c.l. NNPDF 68% c.l.

0.1 0.03

0.08
0.02

xγ (x,Q2)
xγ (x,Q2)

0.06
0.01
0.04

0.02 0
0
-0.01
10-5 10-4 10-3 10-2 10-1 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x x
 THE PHOTON PDF BREAKTHROUGH
(Manohar, Nason, Salam, Zanderighi, 2016)
• QED IS PERTURBATIVE DOWN TO LOW SCALES ⇒ THE PHOTON PDF MUST BE
COMPUTABLE IF THE INPUT QUARK SUBSTRUCTURE IS KNOWN

• WRITE THE CROSS-SECTION FOR A CHOSEN PROCESS:

SUSY PRODUCTION IN EP COLLISION (Drees, Zeppenfeld, 1989)

• COMPUTE IT DIRECTLY, OR USING THE PHOTON PDF

• ⇒ PDF EXPRESSED IN TERMS OF THE STRUCTURE FUNCTION INTEGRATED OVER
ALL SCALES, INCLUDING ELASTIC FORM FACTORS

xfγ/p (x, µ2 ) =
µ2
1 1 dz 1−z dQ2 2 2
2
2πα(µ ) x z 2 2 2 α (Q )
x mp Q
(

1−z
R R

2x2 m2p 2
x
zpγq (z) + Q2 F2 (x/z, Q2 ) − z 2 FL z,Q
"  #

x 2
− α2 (µ2 )z 2 F2 z,µ ,
)


 HEAVY QUARKS
EXAMPLE: HIGGS IN BOTTOM QUARK FUSION
MASSIVE (DECOUPLING) SCHEME VS. MASSLESS SCHEME

b b

H
_ H
b
_
b

• THE b QUARK IS MASSIVE • b QUARK MASS EFFECTS NEGLECTED

• THE FROM QCD
b QUARK DECOUPLES
EVOLUTION AND THE RUNNING OF αs : • THE b QUARK IS JUST ANOTHER MASS-
nf = 4 LESS PARTON, nf = 5

• THE DEPENDENCE ON mb FULLY RE- m2

b
• COLLINEAR LOGS limmb →0 ln ARE
m2 m2
H
b
TAINED, INCLUDING O TERMS FACTORIZED AND RESUMMED TO ALL
m2
H ORERS THROUGH QCD EVOLUTION
 IS 

• THERE ARE NO COLLINEAR SINGULARI-

m2 • b INITIATED PROCESSES START AT
b
TIES, ln 2 ARE INCLUDED UP TO FI- LEADING ORDER
mH
NITE ORDER

• b INITIATED PROCESSES START BE-

YOND LEADING ORDER: Hbb̄ STARTS AT
O(α2s ).
ACCURACY REQUIRES MATCHING (ACOT, FONLL, SCET-BASED)
 THE FONLL METHOD
(Cacciari, Greco, Nason, 1998)

BASIC IDEA: COMBINE N i LL MASSLESS RESUMMED & N j LO MASSIVE FIXED-ORDER

(UNRESUMMED) ⇒ EXPAND OUT THE RESUMMED RESULT AND REPLACE THE FIRST j ORDERS WITH
THEIR MASSIVE COUNTERPARTS
2 (3) 2 (4) 2 (3,0) 2
F (x, Q ) = F (x, Q ) + F (x, Q ) − F (x, Q )

2
(3) 2 (3)
F (3) (x, Q2 ) = x C , (y, Q2 )
i 2 i
 

y m
Z 1 dy

x y i=g,q,q̄
h
X (3)  x Q
, αs (Q ) f

(4) 2 (4) x (4) 2 (4) 2

F (x, Q ) = x C , αs (Q ) f (y, Q )
i i
!

x y y
Z 1 dy

i=g,q,q̄,h,h̄
X

ADVANTAGES
• RELIES ON STANDARD FACTORIZATION & DECOUPLING
• THE RESUMMED AND UNRESUMMED ORDERS CAN BE CHOSEN FREELY & INDEPENDENTLY

COMPLICATIONS
• RESUMMED & FIXED-ORDER CALCULATION ARE PERFORMED IN DIFFERENT RENORMALIZATION
& FACTORIZATION SCHEMES: 3F (MASSIVE, DECOUPLING) VS. 4F (MASSLESS)

• MUST MATCH αs & PDFS

SOLUTION
RE-EXPRESS 3F-SCHEME PDFS & αs IN TERMS OF THE 4F-SCHEME ONES
 IMPACT
PROCESSES
bb̄H
4FS, F5S AND MATCHED
• FOR HIGH-SCALE PROCESSES EFFECT
OF LOG RESUMMATION IS SIZABLE
• NAIVE MATCHING OF
MASSIVE+MASSLESS COMPUTATIONS
(“SANTANDER”) FAILS
• DIFFERENT
MATCHING METHODOLOGIES AGREE
WELL
(FONLL: sf, Napoletano, Ubiali;
SCET: Bonvini, Tackmann, 2016)
HXSWG

PDFS
GLUON & UP
Ratio to NNPDF2.3 NNLO, αS = 0.119, Q 2 = 104 GeV 2 Ratio to NNPDF2.3 NNLO, αS = 0.119, Q 2 = 104 GeV 2
1.2 1.2
• IN FFN, SMALL x 1.15 1.15
PDFS MUST GROW TO 1.1 1.1

2
COMPENSATE FOR 1.05 1.05
MISSING LARGE LOGS 1 1
0.95 0.95
• PDFS IN SCHEME
FFN 0.9 0.9

2
DIFFER SIGNIFICANTLY 0.85 NNPDF2.3 VFN 0.85 NNPDF2.3 VFN

0.8 0.8

g ( x, Q ) [new] / g ( x, Q ) [ref]
u ( x, Q2) [new] / u ( x, Q2) [ref]

⇒ ACCURACY LOSS NNPDF2.3 FFN NNPDF2.3 FFN

0.75 0.75
10-5 10-4 10-3 x 10-2 10-1 10-5 10-4 10-3 x 10-2 10-1
 THEORY SUMMARY

• PDF ACCURACY STARTS BEING LIMITED BY THEORY ACCURACY:

NNLO NEEDED FOR PERCENT ACCURACY

• PDF THEORY UNCERTAINTIES NOT YET INCLUDED,

NEEDED FOR PERCENT ACCURACY

• MANY LHC PROCESSES NOW KNOWN AT NNLO, FEW INCLUDED IN PDF

DETERMINATION:
CURRENTLY ONLY DRELL-YAN RAPIDITY DISTRIBUTION

• RESUMMATION NOT INCLUDED IN PDF DETERMINATION

• MATCHED FIXED-ORDER & RESUMMATION CURRENTLY INCLUDED ONLY FOR HEAVY

QUARKS

• MATCHED COMPUTATIONS ⇒ REALISTIC DESCRIPTION OF MEASURABLE QUANTITIES

PDFS NOW
 CONTEMPORARY PDF TIMELINE
2008 2009 2010 2011 2012 2013 2014 2015
SET CT6.6 NN1.0 MSTW ABKM09 NN2.0 CT10(N) NN2.1(NN) ABM11 NN2.3 CT10(NN) ABM12 NN3.0 MMHT CT14
MONTH (02) (08) (01) (08) (02) (07) (07) (02) (07) (02) (10) (10) (12) (06)
F. T. DIS
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
ZEUS+H1-HI
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
COMB. HI
✗ ✗ ✗ ✗ ✔ ✗ ✔ ✗ ✔ ✗ ✔ ✔ ✗ ✗
ZEUS+H1-HII some some
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✗ ✗
HERA JETS
✗ ✗ ✔ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✗
F. T. DY
✔ ✗ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
TEV. W+Z
✔ ✗ ✔ ✗ ✔ ✔ ✔ ✗ ✔ ✔ ✔ ✔ ✔
TEV. JETS
✔ ✗ ✔ ✗ ✔ ✔ ✗ ✔ ✔ ✔ ✗ ✔ ✔ ✔
LHC W+Z some
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✗ ✔ ✔ ✔
LHC JETS
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✗ ✗ ✔ ✔ ✔
TOP
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✔ ✗ ✗
W+C
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✗ ✗
W pT
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✗ ✗

• INCREASINGLY WIDE DATASET USED FOR PDF DETERMINATION

• HERAPDF: ONLY HERA STRUCTURE FUNCTION DATA ⇒ EXTREME CONSISTENCY

• MANY THEORETICAL AND METHODOLOGICAL IMPROVEMENTS:

– MSTW, ABKM: ALL NNLO; NNPDF NNLO SINCE 07/11 (2.1), CT SINCE 02/13 (CT10)
– MSTW, CT ALL MATCHED HEAVY QUARK SCHEMES; NNPDF GM-VFN SINCE 01/11 (2.1)
GLOBAL FITS: THE DATASET IN DETAIL
NNPDF3.0 MMHT14 CT14
SLAC P,D DIS ✔ ✔ ✗
BCDMS P,D DIS ✔ ✔ ✔
NMC P,D DIS ✔ ✔ ✔
E665 P,D DIS ✗ ✔ ✗
CDHSW NU-DIS ✗ ✗ ✔
CCFR NU-DIS ✗ ✔ ✔
CHORUS NU-DIS ✔ ✔ ✗
CCFR DIMUON ✗ ✔ ✔
NUTEV DIMUON ✔ ✔ ✔
HERA I NC,CC ✔ ✔ ✔
HERA I CHARM ✔ ✔ ✔
H1,ZEUS JETS ✗ ✔ ✗
H1 HERA II ✔ ✗ ✗
ZEUS HERA II ✔ ✗ ✗
E605 & E866 FT DY ✔ ✔ ✔
CDF & D0 W ASYM ✗ ✔ ✔
CDF & D0 Z RAP ✔ ✔ ✔
CDF RUN-II JETS ✔ ✔ ✔
D0 RUN-II JETS ✗ ✔ ✔
D0 RUN-II W ASYM ✗ ✗ ✔
ATLAS HIGH-MASS DY ✔ ✔ ✔
CMS 2D DY ✔ ✔ ✗
ATLAS W,Z RAP ✔ ✔ ✔
ATLAS W pT ✔ ✗ ✗
CMS W ASY ✔ ✔ ✔
CMS W +C ✔ ✗ ✗
LHCB W,Z RAP ✔ ✔ ✔
ATLAS JETS ✔ ✔ ✔
CMS JETS ✔ ✔ ✔
TTBAR TOT XSEC ✔ ✔ ✗
TOTAL NLO 4276 2996 3248
TOTAL NNLO 4078 2663 3045
 THE NNPDF3.0 DATASET
NNPDF3.0 NLO dataset

FT DIS
107 HERA1
FT DY

6 TEV EW
10 TEV JET
ATLAS EW
LHCB EW
5
10 LHC JETS
HERA2
ATLAS JETS 2.76TEV
4
10 ATLAS HIGH MASS

T
ATLAS WpT
CMS W ASY
3 CMS JETS
10
CMS WC TOT

Q2 / M2 / p2 [ GeV2 ]
CMS WC RAT
LHCB Z
102
TTBAR

10

1 -6 -5 -3
10 10 10-4 10 10-2 10-1 1
x
 PARTON LUMINOSITIES
QUARK-ANTIQUARK
GLOBAL REDUCED

GLUON-GLUON

• GLOBAL FITS AGREE WELL

• FITS BASED ON REDUCED DATASET HAVE EITHER LARGE UNCERTAINTIES OR SHOW
SIZABLE DEVIATIONS
 PARTON LUMINOSITIES
QUARK-QUARK
2012 2015
LHC 8 TeV - Ratio to NNPDF2.3 NNLO - αs = 0.118 Quark-Quark, luminosity
1.3 1.3
nnpdf3.0nnlo
1.25 NNPDF2.3 NNLO
1.25 mmhtnnlo
CT10 NNLO ct14nnlo
1.2 1.2
MSTW2008 NNLO S = 8.00e+03 GeV
1.15 1.15
1.1 1.1
1.05 1.05

Ratio
1 1
0.95 0.95
Generated with APFEL 2.4.0 Web

0.9 0.9

Quark - Antiquark Luminosity

0.85 0.85
0.8 0.8
102 MX 103 102 M [GeV] 103
X

GLUON-GLUON
LHC 8 TeV - Ratio to NNPDF2.3 NNLO - αs = 0.118 Gluon-Gluon, luminosity
1.3 1.3
nnpdf3.0nnlo
1.25 NNPDF2.3 NNLO 1.25 mmhtnnlo
CT10 NNLO ct14nnlo
1.2 1.2
MSTW2008 NNLO S = 8.00e+03 GeV
1.15 1.15
1.1 1.1
1.05 1.05
Ratio

1 1
0.95 0.95
Generated with APFEL 2.4.0 Web

Gluon - Gluon Luminosity

0.9 0.9
0.85 0.85
0.8 0.8
102 MX 103 102 M [GeV] 103
X

• LONGSTANDING DISCREPANCY BETWEEN GLUON LUMINOSITIES IS GONE ⇒ IMPACT ON HIGGS

• UNCERTAINTIES BLOW UP FOR LIGHT (< 10 GEV) OR HEAVY (> 1 TEV) FINAL STATES ⇒
∼ ∼
IMPACT ON SEARCHES
 PROGRESS
• Q: WHY ARE PDF UNCERTAINTIES ON GLOBAL FITS OF SIMLAR SIZE?
– SIMILAR DATASETS
– BUT DIFFERENT PROCEDURES
• A:UNCERTAINTY TUNED TO DATA THROOUGH TOLERANCE (MMHT & CT) OR
CLOSURE TESTING (NNPDF)
• Q: WHAT HAS DRIVEN THE IMPROVED AGREEMENT OF GLOBAL FITS
– SIMILAR DATASETS
– BUT DIFFERENT PROCEDURES
• A: DATA+METHODOLOGY
METHODOLOGY

NNPDF3.0 MMHT14 CT14

NO. OF FITTED PDFS 7 7 6
a b a b
PARAMETRIZATION NEURAL NETS x (1 − x) × CHEBYSCHEV x (1 − x) ×BERNSTEIN
FREE PARAMETERS 259 37 30-35
UNCERTAINTIES REPLICAS HESSIAN HESSIAN
TOLERANCE NONE DYNAMICAL DYNAMICAL
CLOSURE TEST ✔ ✗ ✗
REWEIGHTING REPLICAS EIGENVECTORS EIGENVECTORS

• MMHT, CT10 LARGER # OF PARMS., ORTHOGONAL POLYNOMIALS

• NNPDF CLOSURE TEST
 EXAMPLE OF DATA-DRIVEN PROGRESS
MSTW/MMHT: THE d/u RATIO
THE CMS W ASYMMETRY

THE d/u RATIO

Ratio to NNPDF2.0
1.4

1.3

1.2

0
1.1

d / u ( x , Q2 )
0.9

0.8 NNPDF2.0
CT10
0.7
MSTW08
0.6
0.1 0.2 0.3 0.4 0.5 0.6 0.7
x

• LONG-STANDING DISCREPANCY IN THE d/u RATIO BETWEEN MSTW AND

OTHER GLOBAL FITS
• RESOLVED BY W ASYMMETRY DATA
• EXPLAINED BY INSUFFICIENTLY FLEXIBLE PDF PARAMETRIZATION
⇒ FIXED IN MSTW08DEUT/MMHT
 THE IMPACT ON HIGGS IN GLUON FUSION
2012 2015
LHC 8 TeV - iHixs 1.3 NNLO - PDF+αS uncertainties ggH, ggHiggs NNLO, LHC 13 TeV, αs=0.118
20.5 45
α s=0.117, 0.119 α s=0.117, 0.119 α s=0.117, 0.119
NNPDF3.0
44.5 MMHT14
20
CT14p
44 CMCPDF
19.5
Envelope
43.5

19
43

σ(H) [pb]
18.5 42.5

Cross-Section (pb)
18 NNPDF2.3 42

MSTW08
41.5
17.5
CT10
41

• PDF4LHC PRESCRIPTION 2012: ENVELOPE, PDF UNCERTAINTY ∼ 6%

• PDF4LHC PRESCRIPTION 2015: STATISTICAL COMBINATION, PDF UNCERTAINTY
∼ 2%
 THE NEW PDF4LHC PRESCRIPTION
• PERFORM MONTE CARLO COMBINATION OF UNDERLYING PDF SETS
• SETS ENTERING THE COMBINATION MUST SATISFY COMMON REQUIREMENTS
• DELIVER A SINGLE COMBINED PDF SET THROUGH SUITABLE TOOLS
MONTE CARLO VS. HESSIAN DELIVERY
• MONTE CARLO: A SET OF PDF REPLICAS IS DELIVERED;
QUANTITIES COMPUTED FOR EACH REPLICA: CENTRAL VALUE IS THE MEAN,
UNCERTAINTY IS STANDARD DEVIATION
• HESSIAN: A CENTRAL SET AND ERROR SETS ARE DELIVERED;
CENTRAL SET PROVIDES CENTRAL PREDICTION, UNCERTAINTY IS THE SUM IN
QUADRATURE OF ERROR DEVIATIONS

TREATMENT OF αs
• PDFS ARE DELIVERED FOR EACH VALUE OF αs
• PDF AND αs UNCERTAINTIES TO BE KEPT SEPARATE
 CURRENT COMBINED SET
• INCLUDES CT14, MMHT, NNPDF3.0
• 900 REPLICAS (300 FOR EACH SET) ENSURE PRECENTAGE ACCURACY ON
ALL QUANTITIES
300, 900,1800 REPLICAS
(RATIO TO 900)
gluon strange
LUMINOSITY UNCERTAINTIES
THE PDF4LHC15 SET

G.P. Salam, 2016

 OUTLOOK
• “PDF UNCERTAINTIES” AS PROPAGATED FROM DATA AND METHODOLOGY
ARE BY AND LARGE RELIABLE, OF ORDER OF 5% IN A WIDE KINEMATIC
RANGE
• LACK OF EXPERIMENTAL INFORMATION CAN BE A LIMITATION IN REGIONS
RELEVANT FOR SEARCHES BUT WILL BE REMEDIED BY LHC DATA
• THEORETICAL UNCERTAINTIES MUST BE HANDLED
• LHC DATA ARE LIKELY TO CHALLENGE THE PROPER HANDLING OF
THEORETICAL UNCERTAINTIES
PDFS ARE NOT PLUMBING!
EXTRAS
 HEAVY QUARK MASSES
REMEMBER THE QUARK LUMINOSITY
Quark-Quark, luminosity
1.3
nnpdf3.0nnlo
1.25 mmhtnnlo
1.2 ct14nnlo
S = 8.00e+03 GeV
1.15
1.1
1.05

Ratio
1
0.95
Generated with APFEL 2.4.0 Web

0.9
0.85
0.8
102 M [GeV] 103
X

• PDFS DEPEND ON HEAVY QUARK MASSES

• NNPDF3.0 mc = 1.275 Gev; MMHT mc = 1.4 Gev;
CT14 mc = 1.3 Gev (POLE) [PDG: 1.47 ± 0.03]
• INDICATIONS THAT DIFFERENCE MAY BE TO mc VALUE
 THE CHARM PDF:
STABILITY
LOW SCALE
DYNAMICAL FITTED

HIGH SCALE

• DYNAMICAL: DEPENDS SIGNIFICANTLY ON THE MASS

WHICH SETS THE PHYSICAL THRESHOLD; DEPENDENCE SEEN BOTH AT LOW AND HIGH SCALE;

• FITTED: EXTREMELY STABLE AT ALL SCALES

STRUCTURE APPEARS AT LARGE x
 STABILITY:
THE LIGHT QUARKS
DOWN
DYNAMICAL FITTED

ANTIUP

• DYNAMICAL CHARM: LIGHT QUARKS DEPEND (WEAKLY) ON THE MASS WHICH SETS THE
PHYSICAL THRESHOLD FOR CHARM, BOTH AT LOW AND HIGH SCALE;

• FITTED CHARM: LIGHT QUARKS BECOME INDEPENDENT OF CHARM MASS AT ALL SCALES

• GLUON LARGELY INSENSITIVE TO CHARM MASS IN ALL CASES