Matrix Element Merging

CKKW-L merging [Lon01] allows for a consistent merging of parton showers with matrix elements to include multiple well-separated partons. The algorithm implemented in PYTHIA is described in [Lon11]. To perform matrix element merging, the user has to supply LHE files [Alw07] for the hard process and the corresponding process with up to N additional jets.

The usage of the merging procedure is illustrated in a few example main programs (main81.cc, main82.cc, main83.cc and main84.cc, together with the input files main81.cmnd, main82.cmnd and main84.cmnd). These examples should of course only serve as an illustration, and as such will not make use of the merging in all possible ways. For full generality, the example programs link to LHAPDF, FastJet and HepMC. Of course the user is welcome to remove these dependencies. To remove the FastJet dependence, the functions calculating example observables have to be deleted. Removing the LHAPDF dependence requires changing the cmnd input files to choose an inbuilt PDF, as outlined in the PDF documentation. The HepMC dependence can be removed by erasing the code allowing for HepMC output.

Three very short LHE files (w+_production_lhc_0.lhe, w+_production_lhc_1.lhe, w+_production_lhc_2.lhe) are included in the distribution. These files are not intended for physics studies, but only serve as input for the example main programs. For realistic studies, the user has to supply LHE files.

In the generation of LHE files, the value of the factorisation scale used in the PDFs is not important, since the cross section will be multiplied by ratios of PDFs to adjust to the PYTHIA starting scales. The same is true for the renormalisation scale (and starting value αs(MZ)) used to evaluate αs. Coupling and scale choices by the user will be transferred to the merging routines.

LHE files should be regularised with a jet measure, and additional cuts on the partons in the LHEF generation avoided as much as possible. This means that the merging scale is always a more stringent cut than all other cuts on the partons. Of course, if the hard process itself is divergent, a cut needs to be made. However, this should be chosen in such a way as to not exclude regions that will be available to the matrix elements with additional jets. An example is QCD di-jet production with additional jets: Say the 2 -> 2 process is regularised with a pTmin cut of pTminCut = 100 GeV, and the 2 - >3 sample is regularised with a kTmin-cut of kTminCut = 50 GeV. This would mean that when reclustering the emission in the 2 -> 3 sample, we could end up with a pT = pTminNow of the 2 -> 2 configuration with pTminCut > pTminNow, which is excluded in the 2 -> 2 sample. Thus, the 2 -> 3 sample will include a Sudakov factor not included in the 2 -> 2 sample, resulting in merging scale dependencies. Such dependencies can be avoided if the cuts on the hard process are minimal.

Of course, additional cuts on electroweak particles are allowed. These should be the same for all samples with 0 <= n <= N additional partons.

If it is not possible to generate LHE files with minimal cuts, the user can choose to use the MergingHooks structures in order to decide how much influence to attribute to parton shower histories in which the reclustered lowest multiplicity process does not pass the matrix element cuts. This is described below.

When generating LHE files, please refrain from explicitly putting unstable particles (e.g. massive gauge bosons) in the final state. Rather, specify a resonance by its decay products, e.g. if Les Houches events for the pp → Z + jets → e+e- + jets process are desired, generate the matrix element events with the Z decay included. From a physical point of view, on-shell final massive gauge bosons should not be considered part of a hard process, since only the boson decay products will be detectable. Furthermore, non-narrow-width approximation contributions are not present if the ME generator only produces on-shell bosons. Interference effects between different production channels for the decay products would also be neglected. These points seem an unnecessary restriction on the accuracy of the ME calculation. In addition, there is a technical reason for this strategy. Since some matrix element generators choose to put additional information on intermediate bosons into Les Houches events, depending on if they pass a certain criterion (e.g. being close to the mass shell), without exact knowledge of this criterion, the only feasible way of bookkeeping the hard process is by identifying outgoing decay products. The code implemented in PYTHIA8 uses this strategy and may become unreliable otherwise.

For all merging purposes, processes with different charge of outgoing leptons are considered different processes. That means e.g. that e+νe+ jets and e-ν̄e + jets are considered independent processes. If the user wishes to generate distributions including effects of more than one process, merged samples for all independent processes should be generated separately and added afterwards.

When the matrix element merging is used to produce HepMC [Dob01] files to be analysed with RIVET [Buc10], special care needs to taken in how the cross section is read by RIVET (see below).

To specify the merging conditions, additionally information on the merging scale value and the functional definition of the merging scale is needed. A few standard definitions of merging scales are available. This makes the user interface less cumbersome.

Different choices intrinsic to the CKKW-L merging procedure might be relevant for the user as well. The base class MergingHooks gives the user the opportunity to define the functional form of the merging scale. In the following, the usage of the merging machinery to consistently include LHE files with additional jets into PYTHIA will be discussed.


Merging with merging scale defined in kT

The quickest way to include processes with additional jets is to produce LHE files with one of the standard ways to define the merging scale. The following switches are provided:

flag  Merging:doKTMerging   (default = off)
If the additional jets in the LHE files have been regulated by a kT cut, the user can supply the merging scale definition by setting this flag to on. kT here and below means cutting on Durham kT for e+e- collisions, and cutting on longitudinally invariant kT for hadronic collisions.

Since currently, a few slightly different definitions of longitudinally invariant kT are considered, a specific form can be chosen by setting the switch

mode  Merging:ktType   (default = 1; minimum = 1; maximum = 3)
Precise functional definition of longitudinally invariant kT. For e+e- collisions, Durham kT is always defined by the square root of min{ 2*min[ Ei2, Ej2] * [ 1 - cosθij] }, so that this switch will have no effect.
option 1 : Longitudinally invariant kT is defined by the square root of the minimum of minimal jet kinematic pT (pTkin,min = min{ pT,i } ) and pTlon,min = min{ min[ pT,i2, pT,j2] * [ (Δyij)2 + (Δφij)2 ] / D2 } , i.e. kT = min{ √pTkin,min, √pTlon,min } for hadronic collisions. Note that the true rapidity of partons is used.
option 2 : Longitudinally invariant kT is defined by the square root of the minimum of minimal jet kinematic pT (pTkin,min = min{ pT,i } ) and pTlon,min = min{ min[ pT,i2, pT,j2] * [ (Δηij)2 + (Δφij)2 ] / D2 }, i.e. kT = min{ √pTkin,min, √pTlon,min } for hadronic collisions. Note that the pseudorapidity of partons is used.
option 3 : Longitudinally invariant kT is defined by the square root of the minimum of minimal jet kinematic pT (pTkin,min = min{ pT,i } ) and pTlon,min = min{ min[ pT,i2, pT,j2] * [ cosh(Δηij) - cos(Δφij) ] / D2 } , i.e. kT = min{ √pTkin,min, √pTlon,min } for hadronic collisions.

mode  Merging:nJetMax   (default = 0; minimum = 0)
Maximal number of additional jets in the matrix element

parm  Merging:TMS   (default = 0.0)
The value of the merging scale. The name is inspired by the scale in evolution equations, which is often called 't', and the suffix 'MS' stands for merging scale. In the particular case of kT-merging, this would be the value of the kT-cut in GeV. For any merging scale definition, this input is considered the actual value of the merging scale.

word  Merging:Process   (default = void)
The string specifying the hard core process, in MG/ME notation. If e.g. W + jets merging should be performed, set this to pp>e+ve (without white spaces or quotation marks). This string may contain resonances in the MG/ME notation, e.g. for merging pp→Z W+→q q̄ e+νe + jets, the string pp>(z>jj)(w+>e+ve) would be applicable.

flag  Merging:doMGMerging   (default = off)
Even easier, but highly non-general, is to perform the merging with MadGraph/MadEvent-produced LHE files, with a merging scale defined by a kT cut. For this, set this switch to on. The merging scale value will be read from the +1 jet LHE file by searching for the string ktdurham, and extracting the value from value = ktdurham. Also, the hard process will be read from the +0 jet LHE file, from the line containing the string @1 (the tag specifying the first process in the MadGraph process card). For this to work, PYTHIA should be initialised on LHE files called NameOfYourLesHouchesFile_0.lhe (+0 jet sample) and NameOfYourLesHouchesFile_1.lhe (+1 jet sample) and the same naming convention for LHE files with two or more additional jets.

Since for this option, the merging scale value is read from the LHEF, no merging scale value needs to be supplied by setting Merging:TMS . Also, the hard process is read from LHEF, the input Merging::Process does not have to be defined. However, the maximal number of merged jets still has to be supplied by setting Merging:nJetMax.

Histogramming the events

After the event has been processed, histograms for observables of interest need to be filled. In order to achieve good statistical accuracy for all jet multiplicities and all subprocesses contributing to one jet multiplicity, generally a fixed number of unit-weighted events is read from each Les Houches Event file. To then arrive at the correct prediction, for each of these events, histogram bins should be filled with the corresponding cross section, or weighted with unit weight and normalised at the end to the generated cross section for each jet multiplicity separately.

Still another, even more important, event weight that has to applied on an event-by-event basis is the CKKW-L-weight. This corrective weight is the main outcome of the merging procedure and includes the correct no-emission probabilities, PDF weights and αs factors. This means that the merging implementation will generate weighted events. The CKKW-L-weight can be accessed by the following function:

double Info::mergingWeight()  
Returns the CKKW-L weight for the current event.

Note that to avoid confusion, this function does not include the the weight of a phase space point (given by Info::weight()). This weight will differ from unity when reading in weighted Les Houches events. In this case, the full weight with which to fill histogram bins is Info::mergingWeight() * Info::weight().

Finally, to arrive at a correct relative normalisation of the contributions from different number of additional jets in the matrix element, each histogram should be rescaled with the accepted cross section given by Info::sigmaGen(). The accepted cross section includes the effect of vetoes generating Sudakov form factors for the matrix elements, and is in general only known after the run.

This final step can of course be skipped if the accepted cross section had been estimated before the histgramming run, and histogram bins had instead been filled with the weight Info::mergingWeight() * σest(number of additional jets in current ME sample). This is the way HepMC events should be weighted to produce correct relative weights of events (see below, and particularly examine the example program main84.cc).

Examples how to use these options are given in main81.cc (kT merging) and main84.cc (automatic MG/ME merging for RIVET usage).


Merging with user-defined merging scale function

For all other merging scale definitions, the procedure is slightly more complicated, since the user has to write a small piece of code defining the merging scale. To allow for a user defined procedure, set the input

flag  Merging:doUserMerging   (default = off)
General user defined merging on/off.

Then, set the Merging:nJetMax, Merging:TMS and Merging:Process input as before.

Since during execution, PYTHIA needs to evaluate the merging scale with the definition of the user, the user interface is designed in a way similar to the UserHooks strategy. The class controlling the merging scale definition is called MergingHooks.

Initialisation

To initialise the merging with user-defined merging scale, we should construct a class derived from MergingHooks, with a constructor and destructor

MergingHooks::MergingHooks()  

virtual MergingHooks::~MergingHooks()  
The constructor and destructor do not need to do anything.

For the class to be called during execution, a pointer to an object of the class should be handed in with the
Pythia::setMergingHooksPtr( MergingHooks*) method. An examples of this procedure are given in main82.cc.

Defining a merging scale

Then, in the spirit of the UserHooks class, the user needs to supply the process to be merged by defining a methods to evaluate the merging scale variable.

virtual double MergingHooks::tmsDefinition(const Event& event)  
This method will have to calculate the value of the merging scale defined in some variable from the input event record. An example of such a function is given in main82.cc.

The base class MergingHooks contains many functions giving information on the hard process, to make the definition of the merging scale as easy as possible:

int MergingHooks::nMaxJets()  
Return the maximum number of additional jets to be merged.

int MergingHooks::nHardOutPartons()  
Returns the number of outgoing partons in the hard core process.

int MergingHooks::nHardOutLeptons()  
Returns the number of outgoing leptons in the hard core process.

int MergingHooks::nHardInPartons()  
Returns the number of incoming partons in the hard core process.

int MergingHooks::nHardInLeptons()  
Returns the number of incoming leptons in the hard core process.

int MergingHooks::nResInCurrent()  
The number of resonances in the hard process reconstructed from the current event. If e.g. the ME configuration was pp -> (w+->e+ve)(z -> mu+mu-)jj, and the ME generator put both intermediate bosons into the LHE file, this will return 2.

double MergingHooks::tms()  
Returns the value used as the merging scale.

Filling output histograms for the event then proceeds along the lines described above in "Histogramming the events".

The full procedure is outlined in main82.cc. Special care needs to be taken when the output is stored in the form of HepMC files for RIVET usage.

Defining a cut on lowest jet multiplicity events

It can sometimes happen that when generating LHE files, a fairly restrictive cut has been used when generating the lowest multiplicity matrix element configurations. Then, it can happen that states that are (in the generation of a parton shower history) constructed by reclustering from higher multiplicity configurations, do not pass this matrix element cut.

Consider as an example pure QCD dijet merging, when up to one additional jet should be merged. Three-jet matrix element configurations for which the reclustered two-jet state does not pass the cuts applied to the two-jet matrix element would never have been produced by showering the two-jet matrix element. This means that the three-jet matrix element includes regions of phase space that would never have been populated by the parton shower. Thus, since the matrix element phase space is larger than the shower phase space, merging scale dependencies are expected. A priori, this is not troublesome, since the aim of matrix element merging is to include regions of phase space outside the range of the parton shower approximation into the shower. An example is the inclusion of configurations with only unordered histories.

Clearly, if the parton shower phase space is very constrained by applying stringent cuts to the two-jet matrix element, merging scale dependencies can become sizable, as was e.g. seen in [Lon11] when forcing shower emissions to be ordered both in the evolution variable and in rapidity. To influence the effect of large phase space differences for shower emissions and matrix element configurations due to LHEF generation cuts, the user has to write a small piece of code overwriting method

virtual double MergingHooks::dampenIfFailCuts(const Event&event)  
multiplicity reclustered state as an input Event. From this input event, the user can then check if matrix element cuts are fulfilled. The return value will be internally multiplied to the CKKW-L weight of the current event. Thus, if the user wishes to suppress contributions not passing particular cuts, a number smaller than unity can be returned.

Note that this method gives the user access to the lowest multiplicity state, which ( e.g. in the case of incomplete histories) does not have to be a 2 → 2 configuration. Also, changing the weight of the current event by hand is of course a major intervention in the algorithm, and should be considered very carefully. Generally, if this facility would have to be used extensively, it is certainly preferable to be less restrictive when applying additional, non-merging-scale-related cuts to the matrix element.

An example how to force a cut on lowest multiplicity reclustered states for pure QCD matrix element configurations is given by main83.cc (to be used with e.g. main82.cmnd).


Matrix element merging and HepMC output for RIVET

An example how to produce matrix element merged events to be analysed with RIVET is given by main84.cc.

The main issue is that the output of separate RIVET runs can not in general be combined. To perform a matrix element merging, we however need to runs over different LHE files. The solution to this problem (so far) is to only perform one RIVET run for all matrix elements, i.e. print the events for all ME parton multiplicities, with the correct weights, to a single HepMC file. Since the correct weight includes the cross section of the different samples after Sudakov vetoes --- which is not a priori known --- the cross sections have to be estimated in a test run, before the actual production run is performed. Finally, the cross section of the last event in the HepMC file has to be taken as the full merged cross section sigma_merge = Sum_{i=0}^N Sum_{j=0}*^{nEvents} sigma_est(i)*wckkwl(j).

This procedure is outlined in main84.cc.


Further variables

For more advanced manipulations of the merging machinery, all parameter changes that were investigated in [Lon11] are supplied. Please check [Lon11] for a detailed discussion of the switches.

These switches allow enthusiastic users to perform a systematic assessment of the merging prescription. Apart from this, we advise the non-expert user to keep the default values.

flag  Merging:includeMassive   (default = on)
If on, use the correct massive evolution variable and massive splitting kernels in the reconstruction and picking of parton shower histories of the matrix element. If off, reconstruct evolution scales, kinematics and splitting kernels as if all partons were massless.

flag  Merging:enforceStrongOrdering   (default = off)
If on, preferably pick parton shower histories of the matrix element which have strongly ordered consecutive splittings, i.e. paths in which consecutive reclustered evolution scales are separated by a user-defined factor.

parm  Merging:scaleSeparationFactor   (default = 1.0; minimum = 1.0; maximum = 10.0)
The factor by which scales should differ to be classified as strongly ordered.

flag  Merging:orderInRapidity   (default = off)
If on, preferably pick parton shower histories of the matrix element with consecutive splittings ordered in rapidity and pT.

flag  Merging:pickByFullP   (default = on)
If on, pick parton shower histories of the matrix element by the full shower splitting kernels, including potential ME corrections and Jacobians from joined evolution measures.

flag  Merging:pickByPoPT2   (default = off)
If on, pick parton shower histories of the matrix element by the shower splitting kernels divided by the evolution pT.

flag  Merging:pickBySumPT   (default = off)
If on, exclusively pick parton shower histories of the matrix element for which have the smallest sum of scalar evolution pT for consecutive splittings has been calculated.

flag  Merging:includeRedundant   (default = off)
If on, then also include PDF ratios and αs factors in the splitting probabilities used for picking a parton shower history of the matrix element, when picking histories by the full shower splitting probability. As argued in [Lon11], this should not be done since a reweighting with PDF ratios and αs factors will be performed. However, it can give useful insight in how sensitive the results are to the prescription on how to choose PS histories.

parm  Merging:nonJoinedNorm   (default = 1.0; minimum = 0.0; maximum = 10.0)
Normalisation factor with which to multiply splitting probability for splittings without joined evolution equation.

parm  Merging:fsrInRecNorm   (default = 1.0; minimum = 0.0; maximum = 10.0)
Normalisation factor with which to multiply splitting probability for final state splittings with an initial state recoiler.

parm  Merging:aCollFSR   (default = 1.0; minimum = 0.0; maximum = 10.0)
Factor with which to multiply the scalar pT of a final state splitting, when choosing the history by the smallest sum of scalar pT. Default value taken from Herwig++ [Tul09].

parm  Merging:aCollISR   (default = 0.9; minimum = 0.0; maximum = 10.0)
Factor with which to multiply the scalar pT of an initial state splitting, when choosing the history by the smallest sum of scalar pT. Default value taken from Herwig++ [Tul09].

mode  Merging:unorderedScalePrescrip   (default = 0; minimum = 0; maximum = 1)
When the parton shower history of the matrix element contains a sequence of splittings which are not ordered in evolution pT (called an unordered history), this sequence is interpreted as a combined emission. Then, a decision on which starting scale for trial emissions off reconstructed states in this sequence of unordered splittings has to be made. Two options are available:
option 0 : Use larger of the two reconstructed (unordered) scales as starting scale.
option 1 : Use smaller of the two reconstructed (unordered) scales as starting scale.

mode  Merging:unorderedASscalePrescrip   (default = 1; minimum = 0; maximum = 1)
Prescription which scale to use to evaluate αs weight for splittings in a sequence of splittings which are not ordered in evolution pT.
option 0 : Use the combined splitting scale as argument in αs, for both splittings.
option 1 : Use the true reconstructed scale as as argument in αs, for each splitting separately.

mode  Merging:incompleteScalePrescrip   (default = 0; minimum = 0; maximum = 2)
When no complete parton shower history (i.e. starting from a 2 → 2 process) for a matrix element with additional jets can be found, such a configuration is said to have an incomplete history. Since in incomplete histories, not all shower starting scales are determined by clusterings, a prescription for setting the starting scale of trial showers in incomplete histories is needed. Three options are provided.
option 0 : Use factorisation scale as shower starting scale for incomplete histories.
option 1 : Use sHat as shower starting scale for incomplete histories.
option 2 : Use s as shower starting scale for incomplete histories.