1. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 0 LCFIVertex: suggestions for input to DST files  Introduction  Minimal flavour tag information  Suggestions for later additions  Summary Sonja Hillert (Oxford) ILD optimisation meeting, 21 May 2008

2. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 1 Introduction  As Mark emphasised on the ILD optimisation mailing list, the DST files are intended for performing “relatively quick analyses”  “For detailed studies of PFA/Vertexing (at least at this stage) people will need to use the more complete files. This also applies to any study using lepton id.”  Mark further recommended only including the “minimal flavour tag information” for each jet as a starting point.  From this, the questions we should try to answer in today’s discussion seem to be: What does the “minimal flavour tag information” comprise? What information will initially be needed for performing physics studies?  Note that extracting DST files from “full” LCIO files is fast  iterations will be possible, if it turns out that required information is lacking  However, DST information should be complete enough so that not everyone only uses the full files from the start (if e.g. every study needs one piece of information not in DST)

3. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 2 Flavour tag – the black box approach  The absolute minimum needed for using the flavour tag in a physics study are three numbers per jet: the “b-tag”, “c-tag” and “c-tag with b-background only”  These are the output values of neural networks trained to identify these types of jets  LCIO: collection FlavourTag, containing N LCFloatVecs of length 3, with N: number of jets; stored in the same order in which the jets are stored in their RP-collection  Variables stored in LCFloatVecs can be accessed by their name: “ BTag ”, “ CTag ”, “ BCTag ”  Names and order of variables are stored in the parameter of the run header as string vector stored under the name of the LCFloatVec collection.  For example of how to access them, see Ben’s description of the LCIO interface orBen’s description of the LCIO interface Mark Grimes’ PlotProcessor line 220 onwards.Mark Grimes’ PlotProcessor  Usage: choose a cut value on the NN-output, accept jets above the cut and you are done… … but you may not know what you are doing, for example with only these three numbers you won’t be able to look at purity vs efficiency!

4. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 3 Flavour tag – purity vs efficiency  NOTE: this plot will look different for each physics process, as it depends on the flavour composition of the input sample  even in a simple analysis, you should look at this to choose a cut suited to your study! b c (b-bkgr) c efficiency purity 2-jet events @ 92 GeV, LDC01_06Sc Clare Lynch  To find purity vs efficiency need in addition: LCIO collection TrueJetFlavour: N LCFloatVecs of length 4 * (i.e. 1 vector per jet) *) NOTE: in current release, format is 1 LCFloatVec of length N, in contrast to other collections. Will be changed to N vectors for consistency in the near future & 3 other MC-truth values (true quark charge, hadron charge & PDG code) will be added as further vector components

5. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 4 Flavour tag – further useful information  Physics analyses can profit from knowing a little more about “what’s inside the black box”.  All that is needed to use this, is 1 more collection: FlavourTagInputs: N LCFloatVecs, length 16  Could think about adding this to DSTs either now or in a later iteration  Even if not included, it might be useful for physics to know what is available in the “full files”:  Only three flavour tags have been mentioned, BUT: there are actually 9 neural nets, 3 for each of the three tags  Reason: whenever available, vertexing information provides the most reliable input for the flavour tag (e.g. Pt-corrected vertex mass, decay length…)  However: even for b-jets, sometimes only the primary (IP) vertex is found. In those cases, other information, e.g. based on impact parameter significance of the tracks in the input jet, is used as input for the flavour tag neural nets  The number of vertices is available in the vectors in FlavourTagInputs as “ NumVertices ”

6. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 5 Purity vs efficiency for jets with 1, 2, >= 3 vertices b c (b-bkgr) c efficiency purity 1 vertex found b c (b-bkgr) c efficiency purity 2 vertices found b c (b-bkgr) c efficiency purity >= 3 vertices found  flavour tag performance very different for jets with 1, 2 and >= 3 vertices  by using different NN-cut values for these cases, the overall purity vs efficiency can be improved compared to result from page 3; again depends on physics process! plots: Clare Lynch

7. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 6 How to find the NN-cut value for a given efficiency  The LCFIAIDAPlotProcessor provides access to purity efficiency informationLCFIAIDAPlotProcessor and permits selecting suitable NN-cut values. So far, it assumed all LCFIVertex collections to be available – Victoria Martin updated the code with steering file switches for separate parts (so it can cope with a subset of collections); to be committed to CVS shortly  will be finalised once it’s decided which collections available  Switch MakePurityEfficiencyPlots on for plots of purity vs efficiency and flavour tag leakage / mistag rates (combined and for 1, 2, >=3 vertex cases)  Switch PrintPurityEfficiencyValues will permit printing tables corresponding to purity graphs listing 100 NN-cut values and corresponding efficiencies & purities into an output file - default file name PurityEfficiencyOutputFile (set in steering file)  will permit to directly read off NN-cut values (still needs to be implemented, currently need to use corresponding numbers of jets passing cuts instead see next page)

8. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 7  until purity efficiency tables implemented, switch PrintNeuralNetOutput prints out : total number of b-, c-, uds-jets with 1-, 2-, >=3 vertices for 1-, 2- and >= 3-vertex case and for the combined sample, number of signal and background jets passing various values of NN-cut (100 values between 0 and 1), put out in the form: cut b-tag b b-tag other c-tag c c-tag other c-tagbb c c-tagbb other  at the moment need to find cut value from these numbers of jets : use plot to select efficiency for your study calculate number of jets that should pass cuts from chosen efficiency and total number of jets look up appropriate cut value How to find the NN-cut value: intermediate workaround

9. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 8 FlavourTagInputs: input for physics analyses  The flavour tag inputs can be relevant for physics analyses, as the flavour tag is based on reconstructing the decay chain of the leading hadron.  If you’d like to look at decay lengths and masses (as Jenny suggested) it is these variables you should be using at the level of physics studies, not the raw ones from the Vertex objects.  The reason is that the core ZVRES can only identify vertices with at least 2 tracks.  If one secondary vertex is found in addition to the IP, a special procedure is used to possibly assign further tracks to recover additional one-prong vertices.  This “track attachment procedure”track attachment procedure forms part of the calculation of flavour tag inputs and quark charge determination.

10. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 9 FlavourTagInputs for jets with 1 found vertex  If only 1 vertex is found this can be because both the B- and the D-decay vertex are 1-prong (left), both vertices have short decay length and are absorbed into the IP.  In this case, the flavour tag is based on the two most significant tracks (highest impact parameter significance – which could be the two tracks from the two 1-prong vertices) and on the joint probability for all tracks to come from primary vertex (first used at DELPHI).  Names of the components of FlavourTagInputs vector for 1-vertex case: “ JointProbRPhi ”, “ JointProbZ ”: joint probability in R-  and in z “ D0Significance1 ”, “ D0Signficicance2 ”: R-  significance of most and 2 nd -most significant trk “ Z0Significance1 ”, “ Z0Significance2 ”: impact parameter significance in z of these two tracks “ Momentum1 ”, “ Momentum2 ”: momentum of these two tracks

11. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 10 FlavourTagInputs for jets with >= 2 found vertices  Vertexing information is better suited to distinguish b-, c- and uds-jets and is used if at least 2 vertices are found (primary and at least one secondary).  The flavour tag inputs for the case of 2 and of >= 3 vertices are the same, but since distributions differ, different neural nets are trained.  Names of the components of FlavourTagInputs vector for 2, >=3-vertex cases: “ JointProbRPhi ”, “ JointProbZ ”: joint probability in R-  and in z “ NumTracksInVertices ”: sum of all tracks in ZVRES-vertices (before track-attachment) “ DecayLength ”, “ DecayLengthSignificance ”: for vertex with highest decay length significance “ RawMomentum ”: momentum of vertex furthest from IP (i.e. “seed vertex”) “ PTCorrectedMass ”: M Pt : corresponds to reconstructed hadron mass; (from a set of tracks determined from vertex furthest from IP + track attachment) “ SecondaryVertexProbability ”: probability calculated from vertex-fit  2 and ndf (seed vertex)

12. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 11 Input for later DST-file versions: quark charge  Quark charge is useful e.g. for reducing combinatoric backgrounds unfolding cross sections  for charged hadrons: obtained from vertex charge  for neutral hadrons: jet charge (in preparation for LCFIVertex) for mixing study: charge dipole procedure  further algorithms in preparation  Vertex charge used to be provided as part of FlavourTagInputs vectors.  Recently it was taken out: currently separate collections for vertex charge for b- and c-jets  Two collections of LCFloatVecs with all charge variables: one for b-jets, one for c-jets  Will be useful to add to full files & DSTs when rerunning with tuned LCFIVertex configuration

13. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 12 Collections in “full files” to be omitted in DSTs  Collections available in the full files for detailed vertexing & flavour tagging studies: BCharge: vertex charge calculated assuming jet is a b-jet CCharge: vertex charge calculated assuming jet is a c-jet TrueJetPartonCharge TrueJetHadronCharge to be absorbed into TrueJetFlavour in next LCFIVertex release TrueJetPDGCode IPVertex ZVRESDecayChains: N ReconstructedParticle objects ZVRESDecayChainRPTracks: tracks selected by track attachment procedure and used for calculation of variables such as MPt, each wrapped in an ReconstructedParticle ZVRESVertices: “raw” ZVTOP results (including decay lengths for individual vertices) ZVRESSelectedJets: track selection for vertex finding FTSelectedJets: track selection for calculation of flavour tag inputs IPFitSelectedParticles: track selection for IP fit to be extended in future

14. ILD Optimisation Meeting, 21 st May 2008Sonja Hillert (Oxford)p. 13 Summary  For the first version of DST-files for physics analyses suggest to include collections FlavourTag TrueJetFlavour FlavourTagInputs: not absolutely necessary to get started with physics analyses, but analyses would profit from having these available in the DSTs; if not included in the first instance, should be added later.  For a future iteration suggest to additionally include: FlavourTagInputs, if not yet included in the first version 2 quark charge collections – further work needed (see page 10)