P1253296670aDurG

1
LCFI Vertex Package Status
Ben Jeffery
LCFI Collaboration Meeting, Liverpool, 19th
September 2006
2

• Aim C based package for the MARLIN/LCIO
  framework with
• ZVTOP vertex finder
• Neural net flavour tag
• Quark charge sign selection

3
Summary of progress since June

• SGV modified to produce LCIO output.

• Flavour tag routines coded and tested within SGV
  Framework.

• Working classes designed, implemented, tested and
  partly documented.

• All code incorporated into MARLIN framework and
  build system

• Rudimentary LCIO/Working Class interface code
  written.

• ZVTOP code interfaced to working classes

• Neural Net code ported to Marlin.

• ZVRES and first stage of ZVTOP Ghost Track tested
  within MARLIN.

ZVTOP
FLAVOUR TAG ROUTINES
SGV
WORKING CLASSES
NEURAL NET
LCIO
LCIO
MARLIN
4
Working Classes

• A set of minimal, coherent, easy to use classes
  to implement the package without the
  complications of working directly with LCIO
  objects.

WORKING CLASSES

• Includes automatic memory management.
• Objects exist for the run or for an event.
• Easily extensible to other lifetimes.

5
Working Classes

• Key concept of modular algorithm objects that
  can be handled interchangeably
• Algorithms can be plugged into modifiers such
  as track cuts.

ALGO 1 produces DecayChain
Can I have the result of ALGO 1?
Jet
ALGO 2 produces double
Can I have the result of ALGO 2 With a cut on the
track significance?
WORKING CLASSES
TrackCut Object
6
ZVRES in MARLIN framework

• Runtime performance improved no algorithmic
  tuning just code tweaks.
• Median 660ms within order 20 of FORTRAN

• Code spends 67 time fitting
• 30 evaluating gaussian tubes
• Further improvement expected from fitter tuning

ZVTOP
7
ZVRES in MARLIN framework
Performance maintained after move to MARLIN
ZVTOP
8
ZVKIN (Ghost Track)

• First stage fully tested.
• Swivel Ghost Track to find better estimate for B
  direction using Jet momentum as seed

Ghost Track gives better direction than seed
L
T of Ghost
Also detailed tests of intermediate states of
ZVKIN
ZVTOP
9
Flavour Tag Routines (Neural Net Inputs)
Neural Net input variables key part of the
package

• Inputs requiring further computation
• Momentum Corrected Mass
• Joint Probability

Coded and tested by Erik D

• Functions of most significant track
• D0 Significance Of Track
• Z0 Significance Of Track

• Simple functions of the secondary vertex found by
  ZVRES
• Momentum Of Vertex
• Probability Of Vertex
• Number Of Tracks in Vertex
• Decay Length
• Decay Length Error

Being coded in a modular transparent fashion for
low resistance to change
FLAVOUR TAG ROUTINES
10
LCIO Vertex Class
In order not to crowbar vertex information into
LCIO A dedicated vertex class was
preferable. The LCIO developers have now decided
on the design of this After considering the input
from LCFI and others.
Main feature is the ability to easily encode
decay chain information including attached tracks.
LCIO
11
Remaining work before release

• Algorithmic Code
• Remaining parts of Ghost Track (Dave J assures
  me the the hard bit is done!)
• Vertex charge track attachment routines
• Integration and testing
• Flavour tag interfaces to working classes
• Complete LCIO interface to working classes using
  new LCIO vertex class when released
• Neural net interface to LCIO
• Bad user input and other error handling
• System test
• General Usage Documentation (Independent class
  documentation mainly complete)

Much good progress, gearing up for the final push!
12
(No Transcript)
13
The ZVKIN (ghost track) algorithm

• more specialised algorithm to extend coverage to
  b-jets in which one or both
• secondary and tertiary vertex are 1-pronged
  and / or in which the B is very
• short-lived
• algorithm relies on the fact that IP, B- and
  D-decay vertex lie on an approximately
• straight line due to the boost of the B hadron

• should improve flavour tagging capabilities

14
D. Jackson, NIM A 388 (1997) 247
The ZVTOP vertex finder

• two branches ZVRES and ZVKIN (also known as
  ghost track algorithm)
• The ZVRES algorithm
• tracks approximated as Gaussian probability
  tubes
• from these, a vertex function is obtained
• 3D-space searched for maxima in the vertex
  function that satisfy
• resolubility criterion track can be
  contained in gt 1 candidate vertex
• iterative cuts on c2 of vertex fit and
  maximisation of vertex
• function results in unambiguous assignment of
  tracks to vertices
• has been shown to work in various environments
  differing in
• energy range, detectors used and physics
  extracted
• very general algorithm that can cope with
  arbitrary multi-prong decay topologies

15
Flavour tag

• Vertex package will provide flavour tag
  procedure developed by R. Hawkings et al
• (LC-PHSM-2000-021) and recently used by K.
  Desch / Th. Kuhl as default
• NN-input variables used
• if secondary vertex found MPt , momentum
• of secondary vertex, and its decay length and
• decay length significance
• if only primary vertex found momentum and
• impact parameter significance in R-f and z for
  the
• two most-significant tracks in the jet
• in both cases joint probability in R-f and z
  (estimator of
• probability for all tracks to originate
  from primary vertex)
• will be flexible enough to permit user further
  tuning of the input variables for the neural net,
• and of the NN-architecture (number and type
  of nodes) and training algorithm

16
Performance of decay length reconstruction
difference between the C and the FORTRAN
deviation from MC in decay length right
distribution of this difference, left difference
plotted vs MC decay length
Ben Jeffery
54 negative (C version better)
C finds some vertices FORTRAN misses and comes
closer to MC truth than FORTRAN for entire range
of decay lengths
17
Comparison of C FORTRAN ZVRES

• compare distributions of number of vertices and
  of tracks per vertex between
• C, FORTRAN and MC truth for 100 GeV b-jets

Ben Jeffery

• excellent agreement between the results from C
  and from FORTRAN