NNPDF Parton Distribution Functions

Maria Ubiali
(RWTH Aachen)
In collaboration with

S.Forte (U Milano), J. Rojo (CERN)

R.Ball, L.Del Debbio, N. Hartland (U Edinburgh)
V.Bertone (U Freiburg), A.Guffanti(NBI)
F.Cerutti, J.I.Latorre (U Barcelona)
[NNPDF collaboration]

17th SFB/TR9 meeting, PDFs mini workshop, KIT Karlsruhe 19.03.2012

Parton Distributions for the LHC

PDFs LHC
PDF uncertainties are crucial input for
LHC: both for standard candle processes
and for exclusion and discovery

G. Watt, JHEP 1109 (2011) 069

 Can we trust PDF uncertainties?

 How do we interpret the differences between
predictions obtained by using different parton
sets?
 PDF4LHC: huge effort in understanding
differences & improving theoretical and statistical
treatments in PDF analyses
Outline
➩ The NNPDF parton sets

 NNPDF in a nutshell
 The NNPDF2.1 parton sets
 LHC phenomenology

➩ Inclusion of LHC data

 Collider-only fit
 Inclusion of the LHC data by refitting
 Bayesian Reweighting in the NNPDF approach

➩ Conclusions and outlook

The NNPDF parton sets

 NNPDF in a nutshell
 The NNPDF2.1 parton sets
 LHC phenomenology
NNPDF partons
Monte Carlo & Neural Networks
NNPDF partons
Monte Carlo & Neural Networks
➊ Monte Carlo replicas of experimental data

 Generation through MC sampling of data

 Validation against experimental data

 No need of relying on linear propagation

of errors
 Possibility to test for non Gaussian behavior
in fitted PDFs

➋ Fit PDFs with a set of Neural Networks for each rep

 Redundant parametrization:
7 independent PDFs, 259 free parameters
 Dynamical stopping criterion: cross- validation method

 NN provide an unbiased parametrization

➌ Expectation values for observables are MC integrals

(same for errors, confidence levels, correlations...)
NNPDF partons
Timeline

NNPDF1.0 NNPDF1.2 NNPDF2.0 NNPDF2.1 NNPDF2.1 NNPDF2.2 MSTW08 CT10 HERAPDF1.5 ABM11
NLO LO and
NNLO

DIS
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
Drell-Yan data
✘ ✘ ✔ ✔ ✔ ✔ ✔ ✔ ✘ ✔
Jet data
✘ ✘ ✔ ✔ ✔ ✔ ✔ ✔ ✘ ✘
LHC data
✘ ✘ ✘ ✘ ✘ ✔ ✘ ✘ ✘ ✘
Independent
strange and
anti-strange ✘ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔

Heavy Quark
masses
✘ ✘ ✘ ✔ ✔ ✔ ✔ ✔ ✔ ✔
NNLO

✘ ✘ ✘ ✘ ✔ ✔ ✔ ✘ ✔ ✔

✔ only (s+sbar), s = sbar

The NNPDF2.1 parton set
Experimental data

 All systematic correlated uncertainties are included (when available)

 Kinematical cuts Q2 > 3 GeV2 w2 = Q2 (1-x)/x > 12.5 GeV2
 In total 3333 (LO), 3338 (NLO), 3357 (NNLO) experimental points included
The NNPDF2.1 parton set
Features of the fit
 NNPDF2.1 is an ensemble of PDF sets presently available at LO, NLO and NNLO
The NNPDF collaboration, Ball et al, Nucl.Phys. B849 (2011)
The NNPDF collaboration, Ball et al, Nucl.Phys. B855 (2012)

 Global fit: DIS + DY + JET data but available also for data subsets
(NNPDF2.2 includes also first LHC data)

 Heavy quark mass effects included using the General-Mass FONLL scheme up to NNLO
Forte, Laenen et al, Nucl.Phys. B834 (2010)

 Fast Kernel method for the inclusion of exact higher order corrections
The NNPDF collaboration, Ball et al, Nucl.Phys. B838 (2010)
 DIS up to NNLO
 DY and JET up to NLO

 NNLO correction to DY included by means of K-factors

 NNLO correction to inclusive JET observables implemented using FastNLO with approximated
NNLO corrections based on threshold resummation
Kluge, Rabbertz, Wobisch, hep-ph/0609285 (2006)
Kidonakis, Owens, Phys. Rev. D (2001)
 Available for a range of αS = 0.114,0.115,...,0.124
mc = 1.4,...,1.7
mb = 4.25,4.5,4.75,5.0
The NNPDF2.1 parton set
Results
x xg

xT3 x(db - ub)

The NNPDF2.1 parton set
Perturbative stability
NNPDF collaboration, 1107.2652
At the starting scale (2 GeV2)

At the EW scale (1002 GeV2)

The NNPDF2.1 parton set
Perturbative stability

Momentum sum rule

Even without imposing momentum

sum rules in the fit, the fitted PDFs
tend to respect momentum sum
rules and this tendency increases
as perturbative order increases

NNPDF collaboration, 1107.2652

The NNPDF2.1 parton set
Implications for LHC
At hadron colliders, observables depend on PDFs through parton luminosities

 GG luminosity particularly stable in the standard Higgs regions

 QQ luminosity significantly larger at NNLO in the W, Z region
The NNPDF2.1 parton set
Implications for LHC

 Gluon fusion Higgs production: noticeable agreement among global sets (MSTW and
NNPDF at NNLO plus CT10 at NLO)
 Agreement increases when same value of αs(Mz)
 Sizeable difference with the ABKM09 predictions, partially accounted for different value
of αs(Mz)
The NNPDF2.1 parton set
Implications for LHC

 W and Z production: weaker

dependence on αS(MZ) but non negligible
higher order corrections
 Smaller differences among PDF sets

G. Watt, JHEP 1109 (2011) 069

Inclusion of LHC data

 Collider-only fit?
 Inclusion of the LHC data by refitting
 Bayesian Reweighting in the NNPDF approach
Collider-only fit

 No fixed target data  No low energy troubles

(nuclear corrections, higher twists...)

 PDFs from HERA + Tevatron?

Collider-only fit

Total Valence Strange Valence Sea asymmetry

 No fixed target data  No low energy troubles

(nuclear corrections, higher twists...)

 HERA + Tevatron:

✔ Good accuracy for gluon

✖ Loss of accuracy for flavor separation and strange

Gluon  What about

HERA + Tevatron + LHC ?
Constraints from LHC data
 Medium and large x gluon

 Prompt photon
 Precision jets data
 Top pairs

 Light flavors at medium

and small x

 Low-mass Drell-Yan
 Z rapidity distributions
 W asymmetries
 Polarized W

 Strangeness and heavy

flavors

 Wc for strangeness
 Zc and γc for charm
 Zb for bottom

J. Rojo, PDF4LHC November 2011

Inclusion of LHC data by refitting
 To include LHC data in parton fits, TOOLS to interface slow NLO/NNLO codes to the fast
computation needed in a fit are essential

APPLGRID [T. Carli et al, Eur.Phys.J. C66 (2010)]

FASTNLO [T. Kluge et al, hep-ph/0609285]
FASTDY [NNPDF, Nucl.Phys. B838 (2010)]

 These codes are based on

 Pre-computation of hard part of the process on a grid

 Polynomial Interpolation

 Medium-long term project [work in progress... NNPDF2.3 summer 2012]

Inclusion of LHC data by refitting
N. Hartland, Moriond QCD 2012

In the meantime, is it possible to have a tool to estimate

quickly the impact of new data without refitting?
Inclusion of LHC data by reweighting
R.D.Ball et al. ArXiv:1012.0836

 Bayesian reweighting method inspired by Giele and Keller [hep-ph/9803393]

 The IDEA:

 Nrep of a Monte Carlo fit give the probability density in the space of PDFs
 Expectation values are MC integrals. Same for errors, correlations...
 One can assess the impact of including new data in the fit by updating the probability
density distribution!

Refitting: Reweighting:

 Whenever add new data, need to do  Immediate: no need to refit

full refitting, tune parametrization and
statistic treatment  Anybody can do it just evaluating
weights with each replica of a PDF set
 Can be done only by PDF fitting and producing a new PDF set through
collaborations themselves. unweighting
Reweighting versus refitting

 Start from NNPDF2.0

DIS+DY only fit (BLUE)

 Add CDF and D0 jet data

by refitting (as a
benchmark)

 Add first CDF data by

reweighting, unweight,
then reweight D0 data
on unweighted set

 Add first D0 data by

reweighting, unweight,
then reweight CDF data
on unweighted set

 Add CDF+D0 at the same

time by reweighting

 Obtain the same results!

The (old) W lepton asymmetry data
 D0 electron and muon asymmetry distributions
[ArXiv:0709.4254] [ArXiv:0807.3367]
 ATLAS muon charge asymmetry 36 pb-1 [ArXiv: 1103.2929]
 CMS electron and muon asymmetry 36 pb-1 [ArXiv: 1103.3470]
 LHCb muon asymmetry [LHCb-CONF-2011-039]

LHC

10-4 ≲ x ≲ 0.7
S
Cb

CM

Tevatron
LH

S,

Cb
LA

LH
AT

10-2 ≲ x ≲ 0.2
The NNPDF2.2 set

 NNPDF2.2: Added ATLAS and CMS W lepton asymmetry data and Tevatron D0
electron and muon asymmetry data at the same time
 These data did not include full covariance matrix
 Their inclusion reduces uncertainty and moves central values
 Total uncertainty reduction is already pretty significant
 d PDFs: tension between Tevatron and low-energy data?
Updated LHC data
 LHC data in NNPDF2.2 now superseded

 Full covariance matrix is available for ATLAS W lepton and Z rapidity distributions
 Higher luminosity 234 pb-1 for CMS muon asymmetry

 Additional LHC data

 36 pb-1 inclusive jet measurements (full covariance matrix ATLAS)

 36 pb-1 LHCb Z rapidity distribution, W asymmetries at high rapidity
 840 pb-1 CMS W electron asymmetry with full covariance matrix
 4.67 fb-1 CMS inclusive jet measurements

NNLO predictions obtained with DYNNLO

Impact of the LHC EW data (ATLAS)
* Full covariance matrix provided
Impact of the LHC EW data (CMS)
* Covariance matrix not yet available
Impact of the LHC EW data (LHCb)
* Covariance matrix not yet available
Impact of the LHC EW data on PDFs
Conclusions and outlook

 Reliable PDF uncertainties are crucial for the LHC analyses

 NNPDF partons are produced with a method which overcomes some drawbacks of standard
approaches: restricted parametrization, tolerances and tuning of statistical treatment to data
included in the fits
[ArXiv:1103.2369,1102.4407,1101.1300,1012.0836,1005.0397...]

 Next steps: inclusion of LHC data, MS mass for charm and bottom quarks, exploring
intrinsic charm contribution, EW corrections, study of theoretical uncertainty, inclusion of
resummations

 NNPDF2.2 has the same features and it includes already the effect of early LHC data via
reweighting

 Inclusion of new LHC data is crucial for obtaining up-to-date parton sets

 Next NNPDF releases will use as much as possible up-to-date LHC data

Thank you for your attention!

BACK-UP
Impact of the LHC inclusive jet data
The NNPDF2.1 parton set
Results
Description of SLAC data
NNPDF21_FFN_NF3 NLO predictions versus SLAC data (Fp2 and Fd2 overlapped)

Good description
of the data

χ2/dof = 1.28 (p)

χ2/dof = 1.05 (d)
Description of HERA NC data
NNPDF21 predictions versus HERA NC data
Strangeness of the proton
s+
Rs determination
The NNPDF2.1 parton set
Comparison to MSTW08

 Apart from NNPDF MSTW is presently the only

NNLO PDF set available for different values of αS
 Comparison performed at αS(MZ) = 0.119
 Reasonable agreement for most PDF
combinations
 NNPDF singlet has larger uncertainty in
extrapolation region
The NNPDF2.1 parton set
Comparison to MSTW08

 Apart from NNPDF MSTW is presently the only

NNLO PDF set available for different values of αS
 Comparison performed at αS(MZ) = 0.119
 Reasonable agreement for most PDF
combinations
 NNPDF singlet has larger uncertainty in
extrapolation region
 NNPDF exhibits a more stable gluon at small-x
 Flexible param. of strangeness
The NNPDF2.1 parton set
Comparison to ABKM09

 ABKM09 provide only one value of αS

determined from the fit

 Harder gluon due to smaller αS

 Worse agreement than with MSTW
 ABKM fit exhibits a smaller uncertainty in
extrapolation region
 To be compared with ABM11
The NNPDF αs determination
NNPDF coll, ArXiv 1110.2483 [PLB in press]

 Performed NNPDF2.1 NNLO fits at different

αS(MZ) with large statistics
 Determine bootstrap errors of each point
 Result in agreement with PDG value,
small statistical uncertainties

 Remarkable perturbative stability

 Somewhat lower but compatible results
when fit is performed on DIS data only
PDF + αs uncertainty

Exact physical parameters+PDFs uncertainty

obtained as average over the joined Monte
Carlo ensemble [ArXiv:1005.0397]
The NMC issue
The NMC issue
The NMC issue
Heavy quark schemes

 All GM-VFNS include terms included in ZM-VFN

scheme and FFNS at a given order in PT, but sub-
leading terms are different
 The latter may have a not negligible impact
 See LH benchmark [Andersen et al, 1003.1241] and HERA
studies [Placakyte’s talk, QCD@LHC]