Title: Monte Carlo Tuning and More In View of the LHC
1Monte Carlo Tuning (and More) In View of the LHC
- Paolo Bartalini
- EP Division, CERN
- IRES, March 28 - 2003
(Thanks to M.L.Mangano, A.Moraes, N.Brook,
C.Mesropian, V.Tano etc.)
2OUTLINE
- Monte Carlo Tools for the LHC
- Cross Section Evaluators
- Parton Level MCs
- Shower MCs
- The Underlying Event
- Multiplicity Tuning of Pythia
- Recent CDF Results
3Final States at the LHC
High Luminosity Golden Channel
4Final states at the LHC
Goal of MC development for the LHC is to provide
a description as accurate as possible of these
events (and more), as well as of the features of
new physics processes rates, distributions, fine
details of the final states (overall
multiplicities, heavy-quark content)
5Use and Abuse of MC Simulation
- Use
- benchmarks for the design of detectors, trigger
and analysis strategies - tests and measurements of SM
- study of properties of new particles (masses,
cross-sections, couplings) - Abuse claims of discoveries!
- top and SUSY discovery in UA1
- Rb at LEP
- quark compositeness at CDF
- Only the benchmarking against real data can turn
MC simulation into powerful study tools
6Example H?bb in qq ?Hqq
- bbjj bg is 102 times the signal, but can be
extracted from data (smooth behaviour under the
signal peak) - bg from multiple collisions (jjb ? jjb)
signal, but peak under the signal! Much more
sensitive to MC simulation uncertainties!
M.L.Mangano et al., Phys. Lett. B556 (2003) 50.
7Factorization Theorem
- transition from partonic final state to the
hadronic observable (hadronization, fragm.
function, jet definition, etc) - Sum over all histories with X in them
- sum over all initial state histories leading, at
the scale Q, to
8Cross-section Evaluators
- Only some component of the final state is singled
out for the measurement, all the rest being
ignored (i.e. integrated over). E.g. pp?ee- X - No events are generated, only cross-sections
are evaluated - Experimental selection criteria (e.g. jet
definition or acceptance) are applied on
parton-level quantities. Provided these are
infrared/collinear finite, it therefore doesnt
matter what F(X) is, as we assume (fact. theorem)
that - Thanks to the inclusiveness of the result, it is
straightforward to include higher-order
corrections, as well as to resum classes of
dominant and subdominant logs
9State of the Art
- NLO available for
- jet and heavy quarks production
- prompt photon production
- gauge boson pairs
- most new physics processes (e.g. SUSY)
- NNLO available for
- W/Z/DY production
- Higgs production
10Parton Level (Matrix Element) Mcs
- Parton level configurations (i.E. Sets of quarks
and gluons) are generated, with probability
proportional to the respective perturbative M.E. - Transition function between a final-state parton
and the observed object (jet, missing energy,
lepton, etc) is unity - No need to expand f(x) or F(X) in terms of
histories, since they all lead to the same
observable - Experimentally, equivalent to assuming
- Smart jet clustering (parton ? jet)
- Linear detector response
11Codes Available For
- W/Z/gamma N jets (N?6)
- W/Z/gamma Q Qbar N jets (N?4)
- Q Qbar N jets (N?4)
- Q Qbar Q Qbar N jets (N?2)
- Q Qbar H N jets (N?3)
- nW mZ kH N jets (nmkN ?8, N?2)
- N jets (N?5)
12Shower Monte Carlo
- After the generation of a given parton-level
configuration (typically LO, 2?1 or 2?2) , each
possible IS and FS parton-level history
(shower) is generated, with probability defined
by the shower algorithm (unitary evolution). - Algorithm numerical, Markov-like evolution,
implementing within a given appoximation scheme
the QCD dynamics - branching probabilities
- infrared cutoff scheme
- hadronization model
- Herwig, Pythia, Isajet
13Complementarity of the 3 Approaches
14Brief Guide to Shower MCs
- Evaluate parton-level probability, from Feynman
rules phase space. E.g.
- As a result of acceleration, q will emit
radiation - The probability that radiation will (or will
not) be emitted is evaluated as a function of the
acceleration of the colour charges
15Sudakov
- Generate ?? ?1 ??
- If ?1 lt P(Q , Q0) ? no radiation, q goes
directly on-shell at scale Q0GeV - Else
- calculate Q1 / P(Q1,Q0) ?1
- emission at scale Q1
- Go back to 1) and reiterate, until shower stops
in 2). At each step the probability of emission
gets smaller and smaller
prob. of no radiation between Q and Q0
1
?2
?1
P
Q
Q0
Q1
Q2
16Problems Quantum Coherence
?
17Solution (Angular Ordering)
?(???1?
?(???2?
Drawbacks
18Coherence is Essential to Describe CDF Multijet
Data
- 3 jet distributions in hadronic collisions
Full Coherence
No Coherence
Soft emissions know about beam line (large Y)
Partial Coherence
Pseudorapidity of Gluon Jet
F.Abe et al., PRD 50 (1994) 5562
19The Programs (Pythia/Isajet/Herwig/Ariadne)
- Isajet
- Q2 ordering with no coherence
- large range of hard processes
- Pythia
- Q2 ordering with veto of non-ordered emissions
- large range of hard processes
- Herwig
- complete color coherence NLO evolution for
large x - smaller range of hard processes
- Ariadne
- complete color dipole model (best fit to HERA
data) - interfaced to PYTHIA/LEPTO for hard processes
20Hadronization
At the end of the perturbative evolution, the
final state consists of quarks and gluons,
forming, as a result of angular-ordering,
low-mass clusters of colour-singlet pairs
21Bottom Line
- Implementation of quantum coherence in shower
MCs is possible, in the limit of large-Nc and
for soft and collinear emission. - Large-angle, hard emission cannot be described
accurately - Possible cure requires starting the shower with
seed multi-parton configurations, evaluated
using exact (possibly tree-level only) matrix
elements. - Potential problems, however, due to double
counting for extra jet emission
22What is Minimum Bias ?
- Experimental Definition depends on the Trigger !
Minimum Bias is usually associated to
non-single diffractive events (NSD) - Constitutes unavoidable background prediction of
radiation levels detector damage occupancy
etc. - Theoretical Definition the most popular models
associate Minimum Bias events to
non-diffractive inelastic interactions (n.dif) - stot selas ss.dif sd.dif sn.dif
A.Moraes, MC at Hadron Collider Workshop,
Durham, January 2003.
23The Underlying Event
Contributions from soft events need
modeling (Pythia, Phojet, Herwig)
Minimum Bias events are dominated by soft
interactions, although theres also some
contribution from hard scattering.
The Underlying Event is everything except the
hard component (Jets). ? Contributions from beam
remnant and initialfinal state radiation
The Pythia solution Multiple Parton interactions
T. Sjostrand et al. PRD 36 (1987) 2019
Multiple hard interactions observed by AFS, UA2,
CDF!
24Multiple Parton Interactions in Pythia
- Scenario needed to describe the high multiplicity
observed at hadron colliders several
parton-parton interactions within a single
hadron-hadron collision -
- Main free parameters
- Choice of multiple interaction model
- Default model with all hadron collisions
equivalent (MSTP(82)1) - Varying impact parameter between the colliding
hadrons hadronic matter described by a Gaussian
(MSTP(82)3) -
- Minimum PT of hard parton-parton scattering
- ? CONTROLS NUMBER OF INTERACTIONS AND HENCE
MULTIPLICITY
25Varying impact parameter model (3) versus default
model (1)
- Varying impact parameter model invented to
describe the shape of UA5 multiplicity
distribution PLB 138 (1984) 304
Pythia 6.134 including double diffraction
with Minimum PT tuned to reproduce the mean
multiplicity
26Mean Charged Multiplicity at ? 0
- UA5 data at ?s 53, 200, 546, 900 GeV Z. Phys.
C 33 (1986) 1 - CDF data at ?s 630, 1800 GeV
PRD 41 (1989) 2330
dN/d? measured at ? 0 versus ?s CDF fit is
superimposed
27Minimum PT versus? s
- Pythia 6.134, double diffraction included
- For each set of parton distribution functions and
for each value of ?s, the minimum PT is adjusted
to reproduce the measured multiplicity - P. Bartalini, O. Schneider
- CERN 2000-004, pgg 293-300
28The Multiple Interactions Tuning
- In the context of the Pythia Multiple
Interactions framework, due to the increasing
color screening at small x, the post-Hera pdfs
imply a running PT min
OUR FITS
Pythia defaults for M.I. Model gt 1
Older pdfs ? Regge-like x dependency xg(x,Q2)
const for x 0
New pdfs ? Lipatov-like x dependency xg(x,Q2)
x-? for x 0
29KNO Scaling Koba, Nielsen, Olesen, Nucl. Rev.
B40 (1972) 371
z nch/ltnchgt F(z) ltnchgt?n/(?n ?n)
- The rise of multiple interactions at low x can
also be used to interpretate the KNO scaling
violations observed by UA5 - A.Moraes, MC at Hadron Collider Workshop,
Durham, January 2003
30Pseudorapidity distribution of charged tracks at
the LHC
- Comparison between prdictions from
- CTEQ4L M.I. Model 3 (tuned)
- CTEQ4L M.I. Model 3 (untuned)
- Predicted value at ?0 with tuned parameters is
in agreement with phenomenological fit of lower
energy data
31Minimum Bias Event Shapes at LHCb
No large differences in PT distributions
Neutrals included
32Minimum Bias Event Shapes at LHCb (b)
b events at LHCb
Differences in ?nch? same change for
b-events More noticeble difference in PT
distributions
33Underlying Event Herwig versus Pythia
34Multiple Interactions in Herwig (JIMMY)
- In principle MI not available within HERWIG
- In practice, interface program (JIMMY
Butterworth, Forshaw Walker) allows MI
hep-ph/9601371 - Also available, ad-hoc modelling of the soft
underlying event (SUE based on UA5 model) - Parameter available for tuning in both JIMMY and
SUE options.
35Jimmy versus UA5 Data
- essentially one free parameter the min. pT of
the hard scatt.
As pTmin ? the of scatters decrease
predictions approach UA5 data. Failed to find a
setting that could describe the data. No further
study presented here.
36Herwig UA5 Minimum Bias Model
- Mean event charged multiplicity chosen according
to - 1/k in negative binomial given by
- The mass spectrum of soft clusters derived from
- Soft cluster pT spectra
37The Underlying Event Studies at CDF
Examines jet event structure from 1 GeV to 50 GeV
looking at toward, away, transverse regions in f
for central rapidities
T.Affolder et al. PRD 65 (2003) 092002
Cone 1
use charged tracks (SpTtracks) and compare
results to MC predictions
38The Charged Jet Evolution in pp Collisions at
CDF(Toward Region)
_
- Study only the charged particle components of
jets charged particle jets - Data Minimum-Bias and Jet20 data
- Examine the properties of the leading charged
particle jet - and compare with Monte-Carlo models HERWIG,
ISAJET, and PYTHIA - Study the growth and development (evolution) from
PT1 0.5 GeV/c to 50 GeV/c
Use simple, non-standard, jet definition with
R0.7
- Assign all charged particles (PTgt 0.5 GeV/c) and
hlt1) to a jet - Jets contain particles from the underlying event
as well - as from outgoing partons
- Even one charged particle can be jet
6 particles 5 jets
39The Toward Region
QCD hard scattering models agree well with
leading jet observables
Evidence for charged particle clusters in the
MinBias data apparent around PT1 of 2 GeV/c
Jet20 data connects on smoothly to Min-Bias data
The charged particle jets in the Min-Bias data
are continuation of the high transverse momentum
charged jets observed in the Jet20 data.
40The Transverse Region
Rapid growth and then constant plateau for
PT1gt5GeV/c
Comparison of data with the QCD Monte-Carlo
predictions of HERWIG, ISAJET, and PYTHIA
41Improving the Pythia Tuning in the Transverse
Region
Mean number of charged tracks in the Transverse
region vs PT of the leading jet compared to MC
results
Good agreement with tuned Pythia 6.206
42The Transverse Region (PT)
Sum of PT of charged tracks in the Transverse
region vs PT of the leading jet compared to MC
results
43Pythia - CDF Tuning
44Pythia - Charged Multiplicity at LHCb
45Pythia - Charged Multiplicity at LHCb (b)
b events at LHCb
46New Fitting/Tuning Tool
- JetWeb based on HERA HZTOOL package being
updated to include Minimum Bias data - J.M.Butterworth and S.Butterworth
hep-ph/0210404 - also submitted to Comput. Phys. Commun.
- Web page - http//jetweb.hep.ucl.ac.uk/
- Database of data, MC and comparisons
- Web interface allows access to DB and submission
of jobs to generate MC plots
47What is JetWeb for?
- Final state in (esp. hadron-hadron) collisions
poorly understood. - Hadronization not calculable in perturbative QCD.
- Monte Carlo generators (e.g. Pythia, Herwig) are
valuable, but have many free parameters. - How do we know which predictions to trust when
planning for future colliders? - Tune to existing data, but which data?Different
models (fail to) describe different measurements. - Automate procedure allow comparison of new MC
(or set of parameters) with experimental results
stored in a database.
48HzTool
- Developed in HERA Workshop to enable comparison
of data with existing and future MC generators. - Routine written in Fortran for each analysis
fills HBOOK histograms from generated events to
compare with measurements. - Range of data already included H1, ZEUS, UA5,
OPAL, CDF, D0. Contributing authors also from
ATLAS and Linear Collider. - Still need more analyses from more experiments to
be included. - Longer-term move to OO framework?
49What does JetWeb add to HZTOOL?
- Easier access via WWW interface.
- Expanding database of existing data, predictions
and comparisons. - Reduces duplication of effort and computing
resources. - Scalable design to keep up with new data and
models.
50JetWeb home page
51 Final Remarks (high-Q2)
- A lot of progress has taken place in the recent
years, but... - 30 yrs after QCD, still a lot of work to be done
to achieve a satisfactory description of all
high-Q2 processes accessible at LHC - Most of the key conceptual difficulties have been
recently, or are being, solved, and their
implementation into concrete MC schemes should be
achievable in the next 5 years - Forthcoming data from Tevatron will help
improving our tools, but the final test will need
real LHC data
52Conclusions (The Underlying Event)
- Comparisons between Pythia and experimental data
(UA5, CDF) demonstrate that Multiple Interaction
models are successful in reproducing the charged
track multiplicity spectrum. - M.I. varying impact parameter models have to be
adopted. - Running PT cut-off in Multiple Interactions is
mandatory, predictions made at larger energies
(ex. LHC) with fixed PT cut-off are most likely
to overestimate the multiplicity observables. - running PT cut-off is even more important if a
post HERA set of parton distribution functions
is used. - But cannot relay on just one model!
53Workshop on MCs for the LHC (MC4LHC), july 7 -
aug 2 2003, at CERN
- Web page http//lhc-monte-carlo.web.cern.ch/lhc-mo
nte-carlo - Authors of all main programs are expected to come
for several weeks to CERN, give tutorials,
discuss new developments, absorb requirements as
discussed with the users, etc. - Working groups on special topics, so far there
are groups defined on - W production processes
- Tuning of minimum bias events and the underlying
event - COMMON CODE REPOSITORIES
- COMMON TOOLS
- COMMON EVENT FILES
- TUNING AND VALIDATION OF EVENT GENERATORS (see
JetWeb)
54Backup (Theory)
55Use of the Fact. Theorem
- The possible histories of initial and final
state, and their relative probabilities, are in
principle independent of the hard process (they
only depend on the flavours of partons involved
and on the scales Q) - Once an algorithm is developed to describe IS and
FS evolution, it can be applied to partonic IS
and FS arising from the calculation of an
arbitrary hard process - Depending on the extent to which different
possible FS and IS histories affect the value of
the observable X, different realizations of the
factorization theorem can be used
56Power corrections
- Classes of non-perturbative effects linked to
the dominant power-like (1/Q) corrections can be
parametrised in terms of a single quantity,
formally given by
Their effect is expected to be very large even at
the Tevatron, and in general for LHC events with
jets in the few-hundred GeV energy range.
- In the case of 1st moments of shape variables,
for example - FFPT Fnon-PT
- Fnon-PTcF P, with
- and
- PP0 ?0(?)-?0(? S) ? /?S
- The impact of these effects at LEP is very
large, and their understanding is essential for
any quantitative QCD study
57M(ontecarlo) o(f) E(verything)
Matrix Element MCs
Cross-Section Evaluators
Shower MCs
- Better treatment of radiation off heavy quarks
- Full treatment of spin correlations in
production and decay - Better description of underlying event
- Better decay tables
- ..
58Backup (Herwig)
59UA5 Model versus CDF data
current ad-hoc phenomenological model from UA5
CDF data (ltnchgt at ? 0) Good description with
UA5 model
60- HERWIG Minimum bias model versus CDF
- current ad-hoc phenomenological model from UA5
W A R N I N G No single diffractive events in
HERWIG though CDF claim UA5 model doesnt
describe underlying event in hard scatter Cant
generate b-events from min bias model need to
generate hard 2 ? 2 process
61UA5 Model versus UA1 data
pmbm1 0.1 pmbm2 9.0 pmbn1 8.0 pmbn2
0.12 pmbn3 -7.0 pmbk1 0.03 pmbk2
0.12 pmbp1 6.5 pmbp2 5.0 pmbp3 7.5
UA5 data at different ?s
62HERWIG single diffraction
Implement UA5 single diffraction model into HERWIG
Implementation complete need comparisons with
data
63- HERWIG new developments
- New min bias model MI eikonal model similar to
PYTHIA - New model contain the JIMMY hard part but a new
soft component - Authors claim reproducibility of the data
- New model should be in next release of HERWIG
- Claimed a solid weeks work for it to be properly
integrated
64Backup (CDF)
65The Underlying Event Studies at CDF
Underlying Event energy beam-beam
remnants part of initial/final state
radiation must be subtracted from jet
energies for comparison with NLO QCD
predictions
Largest uncertainty for low ET
66The Underlying Event Studies at CDF
Complimentary Analyses
Examines jet event structure from 1 GeV to 50 GeV
looking at toward, away, transverse regions in f
for central rapidities
2p
Cone 1
Examines jet events from 50 GeV to 300 GeV
looking in 2 cones at same h as leading jet, and
at 90o in f away in same central h
Leading Jet
f
Cone 2
0
-1 1
h
Both analyses use charged tracks (SpTtracks) and
compare results to MC predictions
67Backup (JetWeb)
68JetWeb
- A WWW Interface and Database for Monte Carlo
Tuning and Validation - See J. M. Butterworth and S. Butterworth,hep-ph/0
210404, also submitted to Comput. Phys. Commun. - Based on HzTool (J. Bromley et al., Future
Physics at HERA, vol 1, 611-612) - Database of data, MC and comparisons
- Web interface allows access to DB and submission
of jobs to generate more MC plots - http//jetweb.hep.ucl.ac.uk
69JetWeb Search Form
70JetWeb Search Results
71A JetWeb Fit
72JetWeb Plots
73The JetWeb Server
- Java object model
- Java servlets running in Tomcat container
- Data underlying model stored in MySQL database.
- MC Model, Logparms, Logfile
- Data Paper, Plot, DataPoint
- Comparison Fit
74The JetWeb Server
75JetWeb on the Grid
- Processing power
- Currently submit jobs to separate batch farms at
Manchester and UCL. - CPU intensive as use increases, need more power.
- Grid should enable transparent access to a wider
range of resources. - Small(ish) self-contained executable run almost
anywhere. - Users could submit jobs, using their own
certificates, to any resource they are entitled
to use. - Storage
- Make database accessible as a resource in its own
right. - Use Grid mechanisms to mirror data for faster and
more reliable access.
76The Story so Far
- Writes out Grid scripts as well as PBS/NQS.
- Semi-automatic procedure
- shell script submits jobs (sometimes)
- output retrieved by hand
- Limited success
- teething troubles with scripts
- frequent failures of Grid components (RB, LB, VO
server) - difficult to configure Grid node (CE, SE)
correctly - Four jobs run so far on GridPP testbed via IC and
CERN RBs. (At UCL, Manchester, Oxford thanks!)
Many more to come.
77The Future
- More automated job submission and output
retrieval - Running jobs with user-provided proxy
certificates - Something similar done in GUIDO?
- Grid storage and database access
- Spitfire?
- OGSA-DAI?
- Combine JetWeb DB with more general DB of results
(Durham)
78Conclusions
- Gathering useful experience, but progress is
slow. - Hard work getting anything to run
- lack of documentation ? hard for non-expert to
use - failure of Grid components (but more stable now)
- Need more (wo)manpower, more powerful web and DB
servers, more expertise! - Well, it is a TESTbed, and things should become
easier as we move towards a production Grid.