source: git/external/fastjet/contribs/Nsubjettiness/README@ a0f5d71

--------------------------------------------------------------------------------
Nsubjettiness Package
--------------------------------------------------------------------------------

The Nsubjettiness package is based on the physics described in:

   Identifying Boosted Objects with N-subjettiness.
   Jesse Thaler and Ken Van Tilburg.
   JHEP 1103:015 (2011), arXiv:1011.2268.

   Maximizing Boosted Top Identification by Minimizing N-subjettiness. 
   Jesse Thaler and Ken Van Tilburg. 
   JHEP 1202:093 (2012), arXiv:1108.2701.

New in v2.0 is the winner-take-all axis, which is described in:

   Jet Observables Without Jet Algorithms. 
   Daniele Bertolini, Tucker Chan, and Jesse Thaler. 
   JHEP 1404:013 (2014), arXiv:1310.7584.

   Jet Shapes with the Broadening Axis. 
   Andrew J. Larkoski, Duff Neill, and Jesse Thaler. 
   JHEP 1404:017 (2014), arXiv:1401.2158.

   Unpublished work by Gavin Salam

New in v2.2 are new measures used in the XCone jet algorithm, described in:

   XCone: N-jettiness as an Exclusive Cone Jet Algorithm.
   Iain W. Stewart, Frank J. Tackmann, Jesse Thaler,
   Christopher K. Vermilion, and Thomas F. Wilkason.
   arXiv:1508.01516.

   Resolving Boosted Jets with XCone.
   Jesse Thaler and Thomas F. Wilkason.
   arXiv:1508.01518.

--------------------------------------------------------------------------------
Core Classes
--------------------------------------------------------------------------------

There are various ways to access N-(sub)jettiness variables, described
in more detail below:

Nsubjettiness       [Nsubjettiness.hh]:
   A FunctionOfPseudoJet<double> interface to measure the
   N-subjettiness jet shape
   (Recommended for most users)

NsubjettinessRatio  [Nsubjettiness.hh]:
   A FunctionOfPseudoJet<double> interface to measure ratios of
   two different N-subjettiness (i.e. tau3/tau2)
   (Recommended for most users)

XConePlugin         [XConePlugin.hh]:
   A FastJet plugin for using the XCone jet algorithm.
   (Recommended for most users)

NjettinessPlugin    [NjettinessPlugin.hh]:
   A FastJet plugin for finding jets by minimizing N-jettiness.
   Same basic philosophy as XCone, but many more options.
   (Recommended for advanced users only.)

Njettiness          [Njettiness.hh]:
   Access to the core Njettiness code.
   (Not recommended for users, since the interface might change)

The code assumes that you have FastJet 3, but does not (yet) require FastJet 3.1

--------------------------------------------------------------------------------
Basic Usage: Nsubjettiness and NsubjettinessRatio  [Nsubjettiness.hh]
--------------------------------------------------------------------------------

Most users will only need to use the Nsubjettiness class.  The basic
functionality is given by:

   Nsubjettiness nSub(N, AxesDefinition, MeasureDefinition)
   // N specifies the number of (sub) jets to measure
   // AxesDefinition is WTA_KT_Axes, OnePass_KT_Axes, etc.
   // MeasureDefinition is UnnormalizedMeasure(beta),
   //                      NormalizedMeasure(beta,R0), etc.
      
   // get tau value
   double tauN = nSub.result(PseudoJet);

Also available are ratios of N-subjettiness values
    NsubjettinessRatio nSubRatio(N, M, AxesDefinition,
                                 MeasureDefinition)
    // N and M give tau_N / tau_M, all other options the same

For example, if you just want the tau_2/tau_1 value of a jet, using recommended
parameter choices, do this:

    PseudoJet this_jet = /*from your favorite jet algorithm*/;
    double beta = 1.0;
    NsubjettinessRatio nSub21(2,1,
                              OnePass_WTA_KT_Axes(),
                              UnnormalizedMeasure(beta));
    double tau21 = nSub21(this_jet);

--------------------------------------------------------------------------------
AxesDefinition  [NjettinessDefinition.hh]
--------------------------------------------------------------------------------

N-(sub)jettiness requires choosing axes as well as a measure (see below). There
are a number of axes choices available to the user, though modes with a (*) are
recommended. Arguments in parentheses are parameters that the user must set.

Axes can be found using standard recursive clustering procedures.  New in v2 is
the option to use the "winner-take-all" recombination scheme:
(*) KT_Axes          // exclusive kt axes
    CA_Axes          // exclusive ca axes
    AntiKT_Axes(R0)  // inclusive hardest axes with antikt, R0 = radius
(*) WTA_KT_Axes      // exclusive kt with winner-take-all recombination
    WTA_CA_Axes      // exclusive ca with winner-take-all recombination

New in v2.2 are generalized recombination/clustering schemes:
    GenET_GenKT_Axes(delta, p, R0 = inf)
    WTA_GenKT_Axes(p, R0 = inf)
    GenKT_Axes(p, R0 = inf)
Here, delta > 0 labels the generalized ET recombination scheme (delta = 1 for
standard ET scheme, delta = 2 for ET^2 scheme, delta = infinity for WTA scheme)
p >= 0 labels the generalized KT clustering metric (p = 0 for ca, p = 1 for kt),
R0 is the radius parameter, and the clustering is run in exclusive mode.  The
GenKT_Axes mode uses standard E-scheme recombination.  By default the value of
R0 is set to "infinity", namely fastjet::JetDefinition::max_allowable_R.

Also new in v2.2 is option of identifying nExtra axes through exclusive
clustering and then looking at all (N + nExtra) choose N axes and finding the 
one that gives the smallest N-(sub)jettiness value:
    Comb_GenET_GenKT_Axes(nExtra, delta, p, R0 = inf)
    Comb_WTA_GenKT_Axes(nExtra, p, R0 = inf)
    Comb_GenKT_Axes(nExtra, p, R0 = inf)
These modes are not recommended for reasons of speed.

Starting from any set of seed axes, one can run a minimization routine to find
a (local) minimum of N-(sub)jettiness.  Note that the one-pass minimization
routine is tied to the choice of MeasureDefinition.
(*) OnePass_KT_Axes          // one-pass minimization from kt starting point
    OnePass_CA_Axes          // one-pass min. from ca starting point
    OnePass_AntiKT(R0)       // one-pass min. from antikt starting point,R0=rad
(*) OnePass_WTA_KT_Axes      // one-pass min. from wta_kt starting point
    OnePass_WTA_CA_Axes      // one-pass min. from wta_ca starting point
    OnePass_GenET_GenKT_Axes(delta, p, R0 = inf) // one-pass min. from GenET/KT
    OnePass_WTA_GenKT_Axes(p, R0 = inf)          // one-pass min from WTA/GenKT
    OnePass_GenKT_Axes(p, R0 = inf)              // one-pass min from GenKT

For one-pass minimization, OnePass_CA_Axes and OnePass_WTA_CA_Axes are not
recommended as they provide a poor choice of seed axes.

In general, it is difficult to find the global minimum, but this mode attempts
to do so:
    MultiPass_Axes(NPass) // axes that (attempt to) minimize N-subjettiness
                          // (NPass = 100 is typical)
This does multi-pass minimization from KT_Axes starting points.

Finally, one can set manual axes:
    Manual_Axes              // set your own axes with setAxes()
    OnePass_Manual_Axes      // one-pass minimization from manual starting point
    MultiPass_Manual_Axes(Npass) // multi-pass min. from manual

If one wants to change the number of passes used by any of the axes finders, one
can call the function
    setNPass(NPass,nAttempts,accuracy,noise_range)
where NPass = 0 only uses the seed axes, NPass = 1 is one-pass minimization, and
NPass = 100 is the default multi-pass.  nAttempts is the number of iterations to
use in each pass, accuracy is how close to the minimum one tries to get, and
noise_range is how much in rapidity/azimuth the random axes are jiggled.

For most cases, running with OnePass_KT_Axes or OnePass_WTA_KT_Axes gives
reasonable results (and the results are IRC safe).  Because it uses random
number seeds, MultiPass_Axes is not IRC safe (and the code is rather slow).
Note that for the minimization routines, beta = 1.1 is faster than beta = 1,
with comparable performance.

--------------------------------------------------------------------------------
MeasureDefinition  [NjettinessDefinition.hh]
--------------------------------------------------------------------------------

The value of N-(sub)jettiness depends crucially on the choice of measure.  Each
measure has a different number of parameters, so one has to be careful when
switching between measures  The one indicated by (*) is the one recommended for
use by users new to Nsubjettiness.

The original N-subjettiness measures are:
    NormalizedMeasure(beta,R0)  //default normalized measure with
                                //parameters beta and R0 (dimensionless)
(*) UnnormalizedMeasure(beta)   //default unnormalized measure with just
                                //parameter beta (dimensionful)

There are also measures that incorporate a radial cutoff:
    NormalizedCutoffMeasure(beta,R0,Rcutoff) //normalized measure with
                                             //additional Rcutoff
    UnnormalizedCutoffMeasure(beta,Rcutoff)  //unnormalized measure with
                                             //additional Rcutoff
                                             
For all of the above measures, there is an optional argument to change from the
ordinary pt_R distance measure recommended for pp collisions to an 
E_theta distance measure recommended for ee collisions.  There are also
lorentz_dot and perp_lorentz_dot distance measures recommended only for
advanced users.

New for v2.2 is a set of measures defined in arXiv:1508.01516.  First, there is
the "conical measure":

    ConicalMeasure(beta,R0) // same jets as UnnormalizedCutoffMeasure
                            // but differs in normalization and specifics
                            // of one-pass minimization

Next, there is the geometric measure (as well as a modified version to yield
more conical jet regions):

    OriginalGeometricMeasure(R) // not recommended for analysis
    ModifiedGeometricMeasure(R)

(Prior to v2.2, there was a "GeometricMeasure" which unfortunately had the wrong
definition.  These have been commented out in the code as
"DeprecatedGeometricMeasure" and "DeprecatedGeometricCutoffMeasure", but they
should not be used.)

Next, there is a "conical geometric" measure:

    ConicalGeometricMeasure(beta, gamma, Rcutoff)

This is a hybrid between the conical and geometric measures and is the basis for
the XCone jet algorithm.  Finally, setting to the gamma = 1 default gives the
XCone default measure, which is used in the XConePlugin jet finder

(*)  XConeMeasure(beta,Rcutoff)
    
where beta = 2 is the recommended default value and beta = 1 is the recoil-free
default.

--------------------------------------------------------------------------------
A note on beta dependence
--------------------------------------------------------------------------------

The angular exponent in N-subjettiness is called beta.  The original
N-subjettiness paper advocated beta = 1, but it is now understood that different
beta values can be useful in different contexts.  The two main choices are:

beta = 1:  aka broadening/girth/width measure
   the axes behave like the "median" in that they point to the hardest cluster
   wta_kt_axes are approximately the same as minimizing beta = 1 measure

beta = 2:  aka thrust/mass measure
   the axes behave like the "mean" in that they point along the jet momentum
   kt_axes are approximately the same as minimizing beta = 2 measure

N.B. The minimization routines are only valid for 1 < beta < 3.

For quark/gluon discrimination with N = 1, beta~0.2 with wta_kt_axes appears
to be a good choice. 

--------------------------------------------------------------------------------
XConePlugin  [XConePlugin.hh]
--------------------------------------------------------------------------------

The XCone FastJet plugin is an exclusive cone jet finder which yields a
fixed N number of jets which approximately conical boundaries. The algorithm
finds N axes, and jets are simply the sum of particles closest to a given axis
(or unclustered if they are closest to the beam).  Unlike the NjettinessPlugin
below, the user is restricted to using the XConeMeasure. 

   XConePlugin plugin(N,R,beta=2);
   JetDefinition def(&plugin);
   ClusterSequence cs(vector<PseudoJet>,def);
   vector<PseudoJet> jets = cs.inclusive_jets();

Note that despite being an exclusive jet algorithm, one finds the jets using the
inclusive_jets() call.

The AxesDefinition and MeasureDefinition are defaulted in this measure to 
OnePass_GenET_GenKT_Axes and XConeMeasure, respectively. The parameters chosen 
for the OnePass_GenET_GenKT_Axes are defined according to the chosen value of
beta as delta = 1/(beta - 1) and p = 1/beta. These have been shown to give the
optimal choice of seed axes. The R value for finding the axes is chosen to be
the same as the R for the jet algorithm, although in principle, these two radii
could be different. 

N.B.:  The order of the R, beta arguments is *reversed* from the XConeMeasure
itself, since this ordering is the more natural one to use for Plugins.  We
apologize in advance for any confusion this might cause.

--------------------------------------------------------------------------------
Advanced Usage:  NjettinessPlugin  [NjettinessPlugin.hh]
--------------------------------------------------------------------------------

Same as the XConePlugin, but the axes finding methods and measures are the same
as for Nsubjettiness, allowing more flexibility.

   NjettinessPlugin plugin(N, AxesDefinition, MeasureDefinition);
   JetDefinition def(&plugin);
   ClusterSequence cs(vector<PseudoJet>,def);
   vector<PseudoJet> jets = cs.inclusive_jets();

--------------------------------------------------------------------------------
Very Advanced Usage:  Njettiness  [Njettiness.hh]
--------------------------------------------------------------------------------

Most users will want to use the Nsubjettiness or NjettinessPlugin classes to
access N-(sub)jettiness information.  For direct access to the Njettiness class,
one can use Njettiness.hh directly. This class is in constant evolution, so
users who wish to extend its functionality should contact the authors first.

--------------------------------------------------------------------------------
TauComponents  [MeasureDefinition.hh]
--------------------------------------------------------------------------------

For most users, they will only need the value of N-subjettiness (i.e. tau)
itself.  For advanced users, they can access individual tau components (i.e. 
the individual numerator pieces, the denominator, etc.)  

   TauComponents tauComp = nSub.component_result(jet);
   vector<double> numer = tauComp.jet_pieces_numerator(); //tau for each subjet
   double denom = tauComp.denominator();  //normalization factor

--------------------------------------------------------------------------------
Extra Recombiners  [ExtraRecombiners.hh]
--------------------------------------------------------------------------------

New in v2.0 are winner-take-all axes.  (These have now been included in 
FastJet 3.1, but we have left the code here to allow the plugin to work under
FJ 3.0).  These axes are found with the help of the WinnerTakeAllRecombiner.
This class defines a new recombination scheme for clustering particles. This
scheme recombines two PseudoJets into a PseudoJet with pT of the sum of the two
input PseudoJet pTs and direction of the harder PseudoJet.  This is a
"recoil-free" recombination scheme that guarantees that the axes is aligned with
one of the input particles. It is IRC safe.  Axes found with the standard
E-scheme recombiner at similar to the beta = 2 minimization, while
winner-take-all is similar to the beta = 1 measure.

New in v2.2 is the GeneralEtSchemeRecombiner, as defined in arxiv:1506.XXXX.
This functions similarly to the Et-scheme defined in Fastjet, but the reweighting
of the sum of rap and phi is parameterized by an exponent delta. Thus, delta = 1
is the normal Et-scheme recombination, delta = 2 is Et^2 recombination, and
delta = infinity is the winner-take-all recombination. This recombination scheme
is used in GenET_GenKT_Axes, and we find that optimal seed axes for minimization
can be found by using delta = 1/(beta - 1). 

Note that the WinnerTakeAllRecombiner can be used outside of Nsubjettiness
itself for jet finding.  For example, the direction of anti-kT jets found
with the WinnerTakeAllRecombiner is particularly robust against soft jet
contamination.  That said, this functionality is now included in FJ 3.1, so this
code is likely to be deprecated in a future version.

--------------------------------------------------------------------------------
Technical Details
--------------------------------------------------------------------------------

In general, the user will never need access to these header files.  Here is a
brief description about how they are used to help the calculation of
N-(sub)jettiness:

AxesDefinition.hh:

The AxesDefinition class (and derived classes) defines the axes used in the
calculation of N-(sub)jettiness.  These axes can be defined from the exclusive
jets from a kT or CA algorithm, the hardest jets from an anti-kT algorithm,
manually, or from minimization of N-jettiness.  In the future, the user will be
able to write their own axes finder, though currently the interface is still
evolving.  At the moment, the user should stick to the options allowed by
AxesDefinition.

MeasureDefinition.hh:

The MeasureDefinition class (and derived classes) defines the measure by which
N-(sub)jettiness is calculated. This measure is calculated between each
particle and its corresponding axis, and then summed and normalized to
produce N-(sub)jettiness. The default measure for this calculation is
pT*dR^beta, where dR is the rapidity-azimuth distance between the particle
and its axis, and beta is the angular exponent.  Again, in the future the user
will be able to write their own measures, but for the time being, only the
predefined MeasureDefinition values should be used.  Note that the one-pass
minimization algorithms are defined within MeasureDefinition, since they are
measure specific.

--------------------------------------------------------------------------------
Known Issues
--------------------------------------------------------------------------------

-- The MultiPass_Axes mode gives different answers on different runs, since
   random numbers are used.
-- For the default measures, in rare cases, one pass minimization can give a 
   larger value of Njettiness than without minimization. The reason is due 
   to the fact that axes in default measure are not defined as light-like
-- Nsubjettiness is not thread safe, since there are mutables in Njettiness.
-- If the AxesDefinition does not find N axes, then it adds zero vectors to the
   list of axes to get the total up to N.  This can lead to unpredictable
   results (including divide by zero issues), and a warning is thrown to alert
   the user.

--------------------------------------------------------------------------------
--------------------------------------------------------------------------------