HUGS Topical Seminar

June 16, 2021

PDF fitting:
methods and uncertainties

Wally Melnitchouk
 Parton distributions in hadrons
Parton distribution functions (PDFs) are light-cone Hig
correlation functions
Z 1
q(x) = d⇠ e(τ ixP + ⇠
= 2) τ = 2W(⇠
h P | (⇠ ) +
, 0) (0) | P i
1 (a)
k+ 0 ⇠
light cone momentum fraction x = +
P single quark k
Wilson line (gauge invariance)
⇢ Z scattering
⇠
W(⇠ , 0) = exp ig d⌘ A+ (⌘ ) P
0

In A = 0
+ τ = 2
gauge, in fast-moving frame PDF has a probabilistic
interpretation as a particle density
Z 1
single quark
dx q(x) = h P | (0) +
(0) | P i ⇡ h P | †
(0) (0)scattering
|P i
1

number
number 0
⇡ z
operator
density
2
 Parton distributions in hadrons
Inclusive high-energy particle production A B ! C X

A B
xa xb
a ˆ b
X C Collins, Soper, Sterman (1980s)

QCD factorization: separation of hard (perturbative, calculable)

from soft (nonperturbative, parametrized) physics
XZ
AB!CX (pA , pB ) = dxa dxb fa/A (xa , µ) fb/B (xb , µ)
a,b X (n)
n
⇥ s (µ) ⇥
ˆab!CX (xa pA , xb pB , Q/µ)
n

process-independent parton distribution functions fa/A

characterizing structure of bound state A
3
 Parton distributions in hadrons
Most information on PDFs obtained from lepton-hadronHigher tw
deep-inelastic scattering (DIS)
e! ✓ ◆
d 2 4↵ 2
E cos2 ✓2
02
(a) 2 ✓ F1 F2 (b)
e
0
= 2 tan +
γ∗ d⌦dE Q4 2M ⌫
Q2 Q2 = ~q 2 ⌫ 2
X xB =
2M ⌫ ⌫ = E E0
N
H
structure function given as convolution of τ
hard
(τ = 2) = 2
Wilson coefficient with PDF ⇤
(a)
Z
X 1
dx ⇣ xB ⌘
, ↵s single
q(x, Q2quark
q
F2 (xB , Q2 ) = xB e2q Cq )
q xB x x scattering
X N
xB e2q q(xB , Q ) 2

q τ =2
⇣ xB ⌘
for leading order approximation Cq 1 single quark
x scattering
4
 Parton distributions in hadrons
Precision PDFs needed to
(1) understand basic structure of QCD bound states
(2) compute backgrounds in searches for BSM physics
2
Q2 evolution feeds Q collider

low x, high Q2 (“LHC”)

from high x, low Q 2 (“JLab”)

fixed-target
x

Information on PDFs obtained from

(1) nonperturbative approaches (low-energy models, DSE, EFT)
(2) lattice QCD
(3) global QCD analysis
 Global PDF analysis
Universality of PDFs allows data from many different
processes (DIS, SIDIS, weak boson/jet production in pp, Drell-Yan …)
to be analyzed simultaneously
distributions parametrized using a specific functional form,
with parameters fitted to data
xf (x, µ) = N x↵ (1 x) P (x)
p
with polynomial e.g. P (x) = 1 + x+⇥x
or Chebyshev, neural net, …

Extraction of PDFs is challenging because usually

there exist multiple solutions — “inverse problem”
PDFs are not directly measured, but inferred from
observables involving convolutions with other functions

6
 Global PDF analysis
Modern global QCD analyses typically fit 1000s of data points
from high-energy scattering experiments
106
x = 0.005 (i = 20) SLAC F2p ⇥ 2i
x = 0.008
x = 0.013
HERMES
105
x = 0.018 BCDMS 100 µ = 70±
0.60 < x < 0.75

x = 0.026
NMC µ = 60±

F2p £ 2 i
0.52 < x < 0.74
x = 0.036
10 4
x = 0.05
CJ15 µ = 55±
0.60 < x < 0.73
µ = 45±
x = 0.07 0.55 < x < 0.70

x = 0.09 µ = 41±
10°1
103 x = 0.1
0.53 < x < 0.69

JLab
µ = 38± (i = 0)
x = 0.11 0.48 < x < 0.68 CJ15
x = 0.14 4 5 6 7

10 2
x = 0.18
Q2 (GeV2)
F2p £ 2 i

x = 0.23
Q2 = 1.7 GeV2 (i = 5)
0.44 > x > 0.24
x = 0.28 10 1

101 Q2 = 2.0 GeV2

0.49 > x > 0.28
x = 0.35
Q2 = 2.4 GeV2

F2n/F2d £ 2 i
0.53 > x > 0.34
x = 0.45

100 Q2 = 2.9 GeV2

0.58 > x > 0.40

x = 0.55 100
Q2 = 3.4 GeV2
0.62 > x > 0.46

10°1 x = 0.65 Q2 = 4.0 GeV2 (i = 0)

0.65 > x > 0.56

n/d JLab
F2 ⇥ 2i CJ15
x = 0.75 10°1
3 4 5 6 7 8
°2
10 W 2 (GeV2)

F2p ⇥ 2i x = 0.85 (i = 0)

Accardi et al. (“CJ15”)

10°3
2 5 10 20 50 100 200 500 PRD93, 114017 (2016)
Q2 (GeV2)
 Global PDF analysis
SLAC (p)
SLAC (d)
BCDMS (p)
DIS
BCDMS (d)
NMC (p)
DY
NMC (d/p)
HERA NC (1) (e+ p) SIDIS
HERA NC (2) (e+ p)
HERA NC (3) (e+ p) SIA
HERA NC (4) (e+ p)
HERA NC (e° p)
HERA CC (e+ p)
HERA CC (e° p)
E866 (pp)
E866 (pd)
COMPASS (º + )
COMPASS (º ° )
COMPASS (K + )
COMPASS (K ° )
COMPASS (h+ )
COMPASS (h° )
ARGUS (º ± , Q = 10.0)
BABAR (º ± , Q = 10.5)
BELLE (º ± , Q = 10.5)
TASSO (º ± , Q = 12.0)
TASSO (º ± , Q = 14.0)
TASSO (º ± , Q = 22.0)
TASSO (º ± , Q = 34.0)
TASSO (º ± , Q = 44.0)
TPC (º ± , Q = 29.0)
TPC (º ± , Q = 29.0)
TPC(c) (º ± , Q = 29.0)
TPC(b) (º ± , Q = 29.0)
TOPAZ (º ± , Q = 58.0)
SLD (º ± , Q = 91.3)
SLD(c) (º ± , Q = 91.3)
SLD(b) (º ± , Q = 91.3)
ALEPH (º ± , Q = 91.2)
OPAL (º ± , Q = 91.2)
OPAL(c) (º ± , Q = 91.2)
OPAL(b) (º ± , Q = 91.2)
DELPHI (º ± , Q = 91.2)
DELPHI(b) (º ± , Q = 91.2)
ARGUS (K ± , Q = 10.0)
BABAR (K ± , Q = 10.5)
BELLE (K ± , Q = 10.5)
TASSO (K ± , Q = 14.0)
TASSO (K ± , Q = 22.0)
TASSO (K ± , Q = 34.0)
TPC (K ± , Q = 29.0)
TPC (K ± , Q = 29.0)
TOPAZ (K ± , Q = 58.0)
SLD (K ± , Q = 91.3)
SLD(c) (K ± , Q = 91.3)
SLD(b) (K ± , Q = 91.3)
ALEPH (K ± , Q = 91.2)
OPAL (K ± , Q = 91.2)
OPAL(c) (K ± , Q = 91.2)
OPAL(b) (K ± , Q = 91.2)
DELPHI (K ± , Q = 91.2)
DELPHI (K ± , Q = 91.2)
DELPHI(b) (K ± , Q = 91.2)
ALEPH (h± , Q = 91.2)
DELPHI (h± , Q = 91.2)
SLD (h± , Q = 91.3)
TPC (h± , Q = 29.0)
OPAL (h± , Q = 91.2)
OPAL(b) (h± , Q = 91.2)
OPAL(c) (h± , Q = 91.2)
Moffat, WM, Rogers, Sato
TASSO (h± , Q = 35.0)
TASSO (h± , Q = 44.0)
DELPHI(b) (h± , Q = 91.2)
arXiv:2101.04664 (“JAM20”)
SLD(c) (h± , Q = 91.3)
SLD(b) (h± , Q = 91.3)

0 1 2 3 4 °4 °3 °2 °1 0 1 2 3 4

¬2red E [residual]
 Global PDF analysis
SLAC (p)
SLAC (d)
BCDMS (p)
BCDMS (d)
NMC (p)
DIS
DY
JAM20-SIDIS
NMC (d/p)
HERA NC (1) (e+ p) SIDIS
HERA NC (2) (e+ p)
HERA NC (3) (e+ p) SIA
HERA NC (4) (e+ p)
HERA NC (e° p)
HERA CC (e+ p)
HERA CC (e° p)
E866 (pp)
E866 (pd)
COMPASS (º + )
COMPASS (º ° )
COMPASS (K + )
COMPASS (K ° )
COMPASS (h+ )
COMPASS (h° )
ARGUS (º ± , Q = 10.0)
d/u
x(d¯ + ū)
BABAR (º ± , Q = 10.5)
0.8
0.6BELLE (º ± , Q = 10.5)
TASSO (º ± , Q = 12.0)
TASSO (º ± , Q = 14.0)
xuv 0.6
TASSO (º ± , Q = 22.0)
TASSO (º ± , Q = 34.0) 0.6
0.4TASSO (º ± , Q = 44.0)
TPC (º ± , Q = 29.0)
JAM 0.4
TPC (º ± , Q = 29.0)
0.4
TPC(c) (º ± , Q = 29.0)
TPC(b) (º ± , Q = 29.0) CJ15
0.2TOPAZ (º ± , Q = 58.0)
SLD (º ± , Q = 91.3) 0.2 0.2
SLD(c) (º ± , Q = 91.3)
SLD(b) (º ± , Q = 91.3) xdv NNPDF3.1
ALEPH (º ± , Q = 91.2)
OPAL (º ± , Q = 91.2)
OPAL(c) (º ± , Q = 91.2)

x(d¯ ° ū) Rs = (s + s̄)/(d¯ + ū)

OPAL(b) (º ± , Q = 91.2)
0.06
DELPHI (º ± , Q = 91.2)
µ2 = 10 GeV2
DELPHI(b) (º ± , Q = 91.2)
ARGUS (K ± , Q = 10.0) 1.5 4
BABAR (K ± , Q = 10.5)
0.04
BELLE (K ± , Q = 10.5)
TASSO (K ± , Q = 14.0) 3
TASSO (K ± , Q = 22.0)
TASSO (K ± , Q = 34.0)
TPC (K ± , Q = 29.0)
1.0
0.02 TPC (K ± , Q = 29.0)
TOPAZ (K ± , Q = 58.0)
2
SLD (K ± , Q = 91.3)
SLD(c) (K ± , Q = 91.3)
SLD(b) (K ± , Q = 91.3)
0.5 1 xg
0.00
ALEPH (K ± , Q = 91.2)
OPAL (K ± , Q = 91.2)
OPAL(c) (K ± , Q = 91.2)
OPAL(b) (K ± , Q = 91.2)
0.01
DELPHI (K ± , Q = 91.2)
DELPHI (K ± , Q = 91.2)
DELPHI(b) (K ± , Q = 91.2)
0.03 0.1 0.3 x 0.01 0.03 0.1 0.3 x 0.01 0.03 0.1 0.3 x
ALEPH (h± , Q = 91.2)
DELPHI (h± , Q = 91.2)
SLD (h± , Q = 91.3)
TPC (h± , Q = 29.0)
OPAL (h± , Q = 91.2)
OPAL(b) (h± , Q = 91.2)
OPAL(c) (h± , Q = 91.2)
Moffat, WM, Rogers, Sato
TASSO (h± , Q = 35.0)
TASSO (h± , Q = 44.0)
DELPHI(b) (h± , Q = 91.2)
arXiv:2101.04664 (“JAM20”)
SLD(c) (h± , Q = 91.3)
SLD(b) (h± , Q = 91.3)

0 1 2 3 4 °4 °3 °2 °1 0 1 2 3 4

¬2red E [residual]
 Why are PDF uncertainties important?
In searches for new physics beyond the Standard Model
a major source of uncertainty on limits/discoveries is from
calculation of QCD backgrounds PDF errors!
“The PDF and αS uncertainties were calculated using the PDF4LHC prescription
[39] with the MSTW2008 68% CL NNLO [40, 41], CT10 NNLO [42, 43], and
NNPDF2.3 5f FFN [44] PDF sets, and added in quadrature to the scale uncertainty.”
Measurements of the charge asymmetry in top-quark pair production in the
dilepton final state at √s = 8 TeV with the ATLAS detector PRD 94, 032006 (2016)

drives a large part of the global PDF community (esp. LHC)

Limits understanding of nucleon structure

e.g. momentum and spin distributions of d quarks at large x
motivation for several JLab12 experiments
(MARATHON, BONuS, SoLID, …)

10
 Why are PDF uncertainties important?
Traditionally extracted from neutron / proton
structure function ratio (where “neutron” ~ deuteron - proton),
but large nuclear uncertainties affect high-x region
cannot discriminate between predictions for d/u at x ~ 1
1
CJ12min
CJ12mid
0.8
CJ12max
2
= 100
0.6
d/u

SU(6)
0.4
DSE

0.2 helicity

0 scalar qq
0 0.2 0.4 0.6 0.8 1
x
Owens, Accardi, WM (2013)

11
 Why are PDF uncertainties important?
Different groups use different definitions of PDF uncertainties
to take into account tensions between data sets
p
multiply uncertainties by “tolerance” factor T = 2

1
CJ15
0.8 MMHT14
CT14
JR14
CJ15: 2
= 2.7
0.6
MMHT: 2
⇡ 25 100
d/u

0.4
CT14: 2
⇡ 100
0.2 JR14: 2
=1
0
0 0.2 0.4 0.6 0.8
x

… is this a meaningful comparison?

12
 Need for new technology
A major challenge has been to characterize PDF uncertainties
— in a statistically meaningful way — in the presence of
tensions among data sets

Previous attempts sought to address tensions in data sets

by introducing
“tolerance” factors (artificially inflating PDF errors)
“neural net” parametrization (instead of polynomial
parametrization), together with MC techniques

However, to address the problem in a more statistically

rigorous way, one requires going beyond the standard
2 minimization paradigm

utilize modern techniques based on Bayesian statistics!

13
 Bayesian approach to fitting
Analysis of data requires estimating expectation values E
and variances V of “observables” O (= PDFs, FFs) which are
functions of parameters ~a
Z
E[O] = dn a P(~a|data) O(~a)
Z
n
⇥ ⇤2
V [O] = d a P(~a|data) O(~a) E[O]

“Bayesian master formulas"

Using Bayes’ theorem, probability distribution P given by

1
P(~a|data) = L(data|~a) ⇡(~a)
Z

in terms of the likelihood function L

14
 Bayesian approach to fitting
Likelihood function
✓ ◆
1 2
L(data|~a) = exp (~a)
2
is a Gaussian form in the data, with 2
function
X ✓ data i theoryi (~a)
◆2
2
(~a) =
i
(data)

with priors ⇡(~a) and “evidence” Z

Z
Z= dn a L(data|~a) ⇡(~a)

Z tests if e.g. an n-parameter fit is statistically different

from (n+1)-parameter fit

15
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood Monte Carlo

( 2
minimization)

16
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood
( 2
minimization)

maximize probability distribution P by minimizing 2

for a set of best-fit parameters ~a0

E [ ~a ] = ~a0

if O is ⇡ linear in the parameters, and if probability is

symmetric in all parameters
E [ O(~a) ] ⇡ O(~a0 )

17
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood
( 2
minimization)

variance computed by expanding O(~a) about ~a0

e.g. in 1 dimension have “master formula”
1h i2
V [O] ⇡ O(a + a) O(a a)
4

where
a2 = V [a]

18
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood
( 2
minimization)

generalization to multiple dimensions via Hessian approach:

find set of (orthogonal) contours in parameter space around ~a0
such that L along each contour is parametrized by statistically
independent parameters — directions of contours given by
eigenvectors êk of Hessian matrix H, with elements
1 @ 2 2 (~a)
Hij =
2 @ai @aj ~a=~a0
êk
and contours parametrized as a (k)
=a (k)
a 0 = tk p ,
vk
with vk eigenvectors of H
19
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood
( 2
minimization)

basic assumption: P factorizes along each eigendirection

Y
P( a) ⇡ Pk (tk )
k
where h ⇣
1 2 êk ⌘i
Pk (tk ) = Nk exp a 0 + tk p
2 vk

2
note: in quadratic approximation for , this becomes a
normal distribution

20
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood
( 2
minimization)

uncertainties on O along each eigendirection

(assuming linear approximation)
 ⇣ ⌘ ⇣ êk ⌘
2
1 ê k
( Ok ) 2 ⇡ O a 0 + Tk p O a0 Tk p
4 vk vk

where Tk is finite step size in tk , with total variance

X 2
V [O] = ( Ok )
k

21
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:
Monte Carlo

in practice, generally one has E[O(~a)] 6= O(E[~a])

so the maximal likelihood method will sometimes fail
Monte Carlo approach samples parameter space and
assigns weights wk to each set of parameters ak

expectation value and variance are then weighted averages

X X 2
E[O(~a)] = wk O(~ak ), V [O(~a)] = wk O(~ak ) E[O]
k k

22
 Bayesian approach to fitting
Two methods generally used for computing Bayesian
master formulas:

Maximum Likelihood Monte Carlo

( 2
minimization)

fast slow
assumes Gaussianity does not rely on
Gaussian assumptions
no guarantee that global
minimum has been found includes all possible
solutions
errors only characterize
local geometry of accurate
2 function

23
 Incompatible data sets
Incompatible data sets can arise because of errors
in determining central values, or underestimation
of systematic experimental uncertainties
requires some sort of modification to standard statistics

Modify the master formula by introducing a “tolerance”

factor T
V [O] ! T 2 V [O]
e.g. for one dimension
T2h i2
V [O] = O(a + a) O(a a)
4

effectively modifies the likelihood function

24
 Incompatible data sets
Simple example: consider observable m , and two measurements
(m1 , m1 ), (m2 , m2 )

compute exactly the 2

function
✓ ◆2 ✓ ◆2
2 m m1 m m2
= +
m1 m2

and, from Bayesian master formula, the mean value

m1 m22 + m2 m21
E[m] =
m21 + m22
and variance does not
1 m21 m22 depend on
V [m] = H = m1 m2 !
m21 + m22

25
 Incompatible data sets
Simple example: consider observable m , and two measurements
(m1 , m1 ), (m2 , m2 )

same m2
L(m|m1, m2)

1.5 1.5 different m2

1.0 1.0

0.5 0.5

0.0 0.0
°4 °2 0 2 4 °4 °2 0 2 4
m m

total uncertainty remains independent of degree of

(in)compatibility of data
Gaussian likelihood gives unrealistic representation
of true uncertainty
26
 Incompatible data sets
Realistic example: recent CJ (CTEQ-JLab) global PDF analysis
10000
0.40 1
1

0.35
0.30
8000
eigen-direction # 3
24 parameters,
Likelihood

0.25 6000

¢¬2
0.20
0.15
12
5 8
29
27
3
4000
33 data sets
0.10 17
18
2
2000
11
13 10 12
5
3428
0.05 30
8
29
27
7 31 19 3
17
18
3332 15 24
6 2
11
20
23
25
22 26 4 9 14 21
16 28
34
13
30
10
7
0.00 0 31
4
33
15
32
26
14
20
9
25
23
22
16
6
24
21
19

°100 °50 0 50 100 °100 °50 0 50 100

100
0.40 1

data sets
0.35 80
0.30
Likelihood

60

compatible
0.25
¢¬2

0.20
40

along this
12
0.15 5
29
8

27 3
0.10 17 18
2
20

e-direction
28
0.05 34
30
31

0.00 0
°10 °5 0 5 10 °10 °5 0 5 10

0.6
8 (0) a1uv (12) a2du (0) TOTAL (17) e866pp06xf
(1) a2uv (13) a4du (1) HerF2pCut (18) H2 CC em
0.4 (2) slac p (19) d0run2cone
(2) a4uv (14) a1g
(3) d0Lasy13 (20) d0 gamjet1
(3) a1dv (15) a2g
0.2 18 (4) e866pd06xf (21) CDFrun2jet
1
(4) a2dv (16) a3g
4 9 15 20 (5) BNS F2nd (22) d0 gamjet3
0.0 (5) a3dv (17) a4g
Projection

2 6 10 11 12 13 14 16 17 19 21 (6) NmcRatCor (23) d0 gamjet2

5 22
3 23 (6) a4dv (18) a6dv (7) slac d (24) d0 gamjet4
°0.2 (7) a0ud (19) off1 (8) D0 Z (25) jl00106F2d
0
(8) a1ud (20) off2 (9) H2 NC ep 3 (26) HerF2dCut
°0.4 (9) a2ud (21) ht1 (10) H2 NC ep 2 (27) BcdF2dCor
(10) a4ud (22) ht2 (11) H2 NC ep 1 (28) CDF Z

°0.6 (11) a1du (23) ht3 (12) H2 NC ep 4 (29) D0 Wasy

(13) CDF Wasy (30) H2 NC em
(14) H2 CC ep (31) jl00106F2p
°0.8 (15) cdfLasy05 (32) d0Lasy e15
7
(16) NmcF2pCor (33) BcdF2pCor
°1.0
27
 Incompatible data sets
Realistic example: recent CJ (CTEQ-JLab) global PDF analysis
9
0.6 10000
1 18

0.5
34 8

3
8000
eigen-direction #18 13

24 parameters,
0.4
Likelihood

6000

¢¬2
0.3 32

33 data sets
5
26
30
4000 10
15
33
0.2 4
1231
16
11
2 20
19
728
11
171427
29
20 18 2000
0.1 24
6 22 19
13
10
2116 9 23 25 15
22
21
25
24
23
0.0 0
°100 °50 0 50 100 °100 °50 0 50 100

0.6 100
1

0.5

data sets not

34 8
80
3
0.4
Likelihood

60

compatible
¢¬2

0.3 32

5
26 40

along this
30
33
0.2 4
12 31

2
7 28
11
17 29 14 27
20 20
0.1 18

e-direction
6 22
13 19
10
16 23
9

0.0 0
°10 °5 0 5 10 °10 °5 0 5 10

0.8 2
(0) a1uv (12) a2du (0) TOTAL (17) e866pp06xf
(1) a2uv (13) a4du (1) HerF2pCut (18) H2 CC em
0.6 (2) slac p (19) d0run2cone
(2) a4uv (14) a1g
(3) d0Lasy13 (20) d0 gamjet1
(3) a1dv (15) a2g
0.4 (4) e866pd06xf (21) CDFrun2jet
(4) a2dv (16) a3g
22 (5) BNS F2nd (22) d0 gamjet3
0.2 4 (5) a3dv (17) a4g
Projection

6 (6) NmcRatCor (23) d0 gamjet2

16
1 3 9 11 (6) a4dv (18) a6dv (7) slac d (24) d0 gamjet4
7 8 12 15 18 20 23

0.0 (7) a0ud (19) off1 (8) D0 Z (25) jl00106F2d

13 14 19
0
5
(8) a1ud (20) off2 (9) H2 NC ep 3 (26) HerF2dCut
10 17
°0.2 (9) a2ud (21) ht1 (10) H2 NC ep 2 (27) BcdF2dCor
(10) a4ud (22) ht2 (11) H2 NC ep 1 (28) CDF Z

°0.4 (11) a1du (23) ht3 (12) H2 NC ep 4 (29) D0 Wasy

(13) CDF Wasy (30) H2 NC em
(14) H2 CC ep (31) jl00106F2p
°0.6 21 (15) cdfLasy05 (32) d0Lasy e15
(16) NmcF2pCor (33) BcdF2pCor
°0.8
28
 Incompatible data sets
Realistic example: recent CJ (CTEQ-JLab) global PDF analysis
9
0.6 10000
1 18

0.5
34 8

3
8000
eigen-direction #18 13

24 parameters,
0.4
Likelihood

6000

¢¬2
0.3 32

33 data sets
5
26
30
4000 10
15
33
0.2 4
1231
16
11
2 20
19
728
11
171427
29
20 18 2000
0.1 24
6 22 19
13
10
2116 9 23 25 15
22
21
25
24
23
0.0 0
°100 °50 0 50 100 °100 °50 0 50 100

0.6 100
1

0.5

data sets not

34 8
80
3
0.4
Likelihood

60

compatible
¢¬2

0.3 32

5
26 40

along this
30
33
0.2 4
12 31

2
7 28
11
17 29 14 27
20 20
0.1 18

e-direction
6 22
13 19
10
16 23
9

0.0 0
°10 °5 0 5 10 °10 °5 0 5 10

0.8 2
(0) a1uv (12) a2du (0) TOTAL (17) e866pp06xf
(1) a2uv (13) a4du (1) HerF2pCut (18) H2 CC em
0.6

standard Gaussian likelihood incapable of accounting for

(2) a4uv (14) a1g (2) slac p (19) d0run2cone
(3) d0Lasy13 (20) d0 gamjet1
(3) a1dv (15) a2g
0.4 (4) e866pd06xf (21) CDFrun2jet
(4) a2dv (16) a3g
22 (5) BNS F2nd (22) d0 gamjet3
0.2 4 (5) a3dv (17) a4g

underestimated individual errors (leading to incompatible data sets)

Projection

6 (6) NmcRatCor (23) d0 gamjet2

16
1 3 9 11 (6) a4dv (18) a6dv (7) slac d (24) d0 gamjet4
7 8 12 15 18 20 23

0.0 (7) a0ud (19) off1 (8) D0 Z (25) jl00106F2d

13 14 19
0
5
(8) a1ud (20) off2 (9) H2 NC ep 3 (26) HerF2dCut
10

— not designed for such scenarios!

17
°0.2 (9) a2ud (21) ht1 (10) H2 NC ep 2 (27) BcdF2dCor
(10) a4ud (22) ht2 (11) H2 NC ep 1 (28) CDF Z

°0.4 (11) a1du (23) ht3 (12) H2 NC ep 4 (29) D0 Wasy

(13) CDF Wasy (30) H2 NC em
(14) H2 CC ep (31) jl00106F2p
°0.6 21 (15) cdfLasy05 (32) d0Lasy e15
(16) NmcF2pCor (33) BcdF2pCor
°0.8
29
 Incompatible data sets
CTEQ tolerance criteria
Eigenvector 4
40

30

BCDMSp

BCDMSd

CCFR2

CCFR3

CDFjet
NMCp

CDFw
NMCr
ZEUS
20

D0jet
E605

E866
H1b
H1a
2

10
T (⇡ 10)
distance

0
p

!10

!20

!30 eigen-direction #4 CTEQ6

for each experiment, find minimum 2

along given e-direction
from 2
distribution determine 90% CL for each experiment
along each side of e-direction, determine maximum range d±
k
allowed by the most constraining experiment
T computed by averaging over all d±
k (typically T ⇠ 5 10)

30
 Incompatible data sets
CTEQ tolerance criteria
Eigenvector 4
40

30

BCDMSp

BCDMSd

CCFR2

CCFR3

CDFjet
NMCp

CDFw
NMCr
ZEUS
20

D0jet
E605

E866
H1b
H1a
2

10
T (⇡ 10)
distance

0
p

!10

!20

!30 eigen-direction #4 CTEQ6

This approach is not consistent with Gaussian likelihood

no clear Bayesian interpretation of uncertainties
(ultimately, a prescription…)

31
To summarize standard maximum likelihood method…
Gradient search (in parameter space) depends how “good” the
starting point is
for ~30 parameters trying different starting points is
impractical, if do not have some information about shape
Common to free parameters initially, then freeze those
not sensitive to data ( 2 flat locally)
introduces bias, does not guarantee that flat 2
globally

Cannot guarantee solution is unique

Error propagation characterized by quadratic 2 near minimum

no guarantee this is quadratic globally (e.g. Student t-distribution?)

Introduction of tolerance modifies Gaussian statistics

32
 Monte Carlo
Designed to faithfully compute Bayesian master formulas

Do not assume a single minimum, include all possible solutions

(with appropriate weightings)

Do not assume likelihood is Gaussian in parameters

Allows likelihood analysis to be extended to address tensions

among data sets via Bayesian inference

More computationally demanding compared with Hessian method

33
 Monte Carlo
First group to use MC for global PDF analysis was NNPDF,
using neural network to parametrize P (x) in
Forte et al. (2002)
f (x) = N x↵ (1 x) P (x)
— ↵, are fitted “preprocessing coefficients”

Iterative Monte Carlo (IMC), developed by JAM Collaboration,

variant of NNPDF, tailored to non-neutral net parametrizations

Markov Chain MC (MCMC) / Hybid MC (HMC)

— recent “proof of principle” analysis, ideas from lattice QCD
Gbedo, Mangin-Brinet (2017)

Nested sampling (NS) — computes integrals in Bayesian master

formulas (for E, V, Z) explicitly Skilling (2004)

34
 Iterative Monte Carlo (IMC)
Use traditional functional form for input distribution shape,
but sample significantly larger parameter space than possible
in single-fit analyses

Iterative Monte Carlo (IMC)

fit
sampler priors fit posteriors

fit
no assumptions for exponents

original data
cross-validation to avoid
pseudo
overfitting
prior
data

training
data
validation
data iterate until convergence
as initial
guess
fit criteria satisfied
parameters from
minimization steps validation

posterior

35
 Iterative Monte Carlo (IMC)
e.g. of convergence (for fragmentation functions) in IMC
1.4 1.4 1.4 1.4 1.4 1.4
u+ d+ s+ c+ b+ g
1.0 1.0 1.0 1.0 1.0 1.0
initial

0.6 0.6 0.6 0.6 0.6 0.6

0.2 0.2 0.2 0.2 0.2 0.2

0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z

1.4 1.4 1.4 1.4 1.4 1.4

intermediate

1.0 º+ 1.0 1.0 1.0 1.0 1.0 º+

0.6 0.6 0.6 0.6 0.6 0.6
K+ K+
0.2 0.2 0.2 0.2 0.2 0.2
0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z

1.4 1.4 1.4 1.4 1.4 1.4

intermediate

1.0 1.0 1.0 1.0 1.0 1.0

0.6 0.6 0.6 0.6 0.6 0.6

0.2 0.2 0.2 0.2 0.2 0.2

0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z

1.4 1.4 1.4 1.4 1.4 1.4

1.0 1.0 1.0 1.0 1.0 1.0 ~ 20

final

0.6 0.6 0.6 0.6 0.6 0.6 iterations

0.2 0.2 0.2 0.2 0.2 0.2
0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z 0.2 0.4 0.6 0.8 z

Sato, Ethier, WM, Hirai, Kumano, Accardi (2016)

36
 Incompatible data sets
Rigorous (Bayesian) way to address incompatible data sets
is to use generalization of Gaussian likelihood
joint vs. disjoint distributions

empirical Bayes

hierarchical Bayes
others, used in different fields

37
 Disjoint distributions
Instead of using total likelihood that is a product (“and”)
of individual likelihoods, e.g. for simple example of two
measurements
L(m1 m2 |m; m1 m2 ) = L(m1 |m; m1 ) ⇥ L(m2 |m; m2 )

use instead sum (“or”) of individual likelihoods

1h i
L(m1 m2 |m; m1 m2 ) = L(m1 |m; m1 ) + L(m2 |m; m2 )
2

gives rather different expectation value and variance

1
E[m] = (m1 + m2 )
2
✓ ◆2
1 m1 m2
V [m] = ( m21 + m22 ) +
2 2
depends on
separation!
38
 Disjoint distributions
Symmetric uncertainties m1 = m2

joint
m2 )
L(m1 m12,|m)

1.5 1.5

1.0 1.0
L(m|m

disjoint
0.5 0.5

0.0 0.0
°4 °2 0 2 4 °4 °2 0 2 4
m m N. Sato

✓ ◆2
1 m1 m2
disjoint: V [m] = ( m21 + m22 ) +
2 2
m21 m22
joint: V [m] =
m21 + m22

39
 Disjoint distributions
Asymmetric uncertainties m1 6= m2
3 3
L(m|m1, m2)
|m)

2 2
1 m 2

1 1
L(m

0 0
°4 °2 0 2 4 °4 °2 0 2 4
m m
N. Sato

disjoint likelihood gives broader overall uncertainty,

overlapping individual (discrepant) data

40
 Empirical Bayes
Shortcoming of conventional Bayesian — still assume
prior distribution follows specific form (e.g. Gaussian)

Extend approach to more fully represent prior uncertainties,

with final uncertainties that do not depend on initial choices

In generalized approach, data uncertainties modified by

distortion parameters, whose probability distributions given
in terms of “hyperparameters” (or “nuisance parameters”)

Hyperparameters determined from data

give posteriors for both PDF and hyperparameters

41
 Empirical Bayes
Simple example of EB for symmetric & asymmetric errors
10 25
8 20 experiment 1
1–æ
P(t|D)

6 15 experiment 2
Hessian
fit
4 10
Empirical
EB fit Bayes
2 5
0 0
°1.0 °0.5 0.0 0.5 1.0 °0.5 °0.4 °0.3 °0.2 °0.1 0.0 0.1

10 25
8 20
6–æ
P(t|D)

6 15
4 10
2 5
0 0
°1.0 °0.5 0.0 0.5 1.0 °0.5 °0.4 °0.3 °0.2 °0.1 0.0 0.1

10 25
8 20
12–æ
P(t|D)

6 15
4 10
2 5
0 0
°1.0 °0.5 0.0 0.5 1.0 °0.5 °0.4 °0.3 °0.2 °0.1 0.0 0.1
t t
N. Sato

42
 PDFs in lattice QCD
Recent progress in extracting x dependence of PDFs in
lattice QCD from matrix element of nonlocal operator
Mq (z, Pz ) = hN (Pz )| q (0, z) z W(z, 0) q (0, 0)|N (Pz )i
Z 1
= dy eiyPz z qe(y, Pz )
1

quasi-PDF qe related to light-cone PDF via matching kernel e

C
Z 1
dy e ⇣ x ⌘
q(x, µ) = C , µ, Pz qe(y, Pz , µ)
1 |y| y

Conflicting results on sign of d¯ ū asymmetry

2.5
6
MSTW
2.0 CJ12
Lattice 4
1.5
u-d

1.0 2

0.5
0
0

d¯ > ū -0.5 d¯ < ū -2

-0.4 -0.2 0 0.2 0.4 0.6 0.8 1.0 -1 -0.5 0 0.5 1
x
Lin et al. (2015) Alexandrou et al. (2018)
43
 PDFs in lattice QCD
Fit lattice observable directly within JAM framework

1 exp
0.4 exp+lat
Re Mu°d

x(u ° d)
0 lat (DFT)
exp
exp+lat 0.2
°1 lat Q2 = 4 GeV2
0.0
1
0.1
Im Mu°d

x(ū ° d)
¯
0
0.0
°1
°10 0 10 10°2 10°1 100
z/a x
Bringewatt, Sato, WM, Qiu, Steffens, Constantinou (2021)

reatively weak impact of present lattice data on

unpolarized PDF determination
44
 PDFs in lattice QCD
Fit lattice observable directly within JAM framework

difficult to fit current

lattice matrix element
data for unpolarized
u-d PDFs

45
 PDFs in lattice QCD
Fit lattice observable directly within JAM framework
1.5 0.75

x(¢u ° ¢d)
Re M¢u°¢d

1.0 0.50

0.5 0.25
Q2 = 4 GeV2
0.0 0.00
0.2
1.0 exp

x(¢ū ° ¢d)
Im M¢u°¢d

¯
exp+lat
0.0
0.0
exp
°0.2 exp+lat
-1.0 lat (DFT)
°10 0 10 10°2 10°1 100
z/a x

Bringewatt, Sato, WM, Qiu, Steffens, Constantinou (2021)

better agreement between lattice and experiment for

polarized PDFs (within larger uncertainties)
46
 PDFs in lattice QCD
Fit lattice observable directly within JAM framework

good fits possible to

both experimental
and lattice data for
polarized u-d PDFs

47
 Outlook
New approaches being developed for global QCD analysis
— simultaneous determination of parton distributions
using Monte Carlo sampling of parameter space
— towards synthesis of lattice QCD data with global analysis

Treatment of discrepant data sets needs attention

— Bayesian perspective has clear merits

Near-term future: “universal” QCD analysis of all observables

sensitive to collinear (unpolarized & polarized) PDFs and FFs

Longer-term: apply MC technology to global QCD analysis

of transverse momentum dependent (TMD) PDFs and FFs
 Outlook
Study of PDFs has brought together essential elements
of nuclear and high-energy physics
1
CJ15 u
0.8 d
d¯ + ū
0.6 d¯ ° ū
xf (x, Q2)

s
g/10
0.4

0.2

Q2 = 10 GeV2
0

10°4 10°3 10°2 0.1 0.3 0.5 0.7 0.9

x

BSM
nucleon physics
structure
neutrino astrophysics
oscillations
chiral
physics
quark-hadron
duality

49