Particle ID with EMCal

1
Particle ID with EMCal

• - Photons and other particles in EMCal
• - Principal Components Analysis

2
Single particles simulation

• Test on mono-energetic particles (g, p, p, n,
  antiprotons and antineutrons) with a transverse
  momentum of 2 GeV/c
• no problem with mixture of different energies
• easier to determine first cuts
• ECore lt 0.5 GeV and TwrHit lt 4 to avoid small
  clusters

3
Results (I)

• TOF the best cut in simulations

4
Results (II)

• ?2, the usual second cut used in Photon ID

5
Results (III)

• Dispersion along principal axes

6
Results with simple cuts

• TOF lt 2.9 ns and c2 lt 2.5, allow to keep 4635
  photon clusters (on 5000), 418 from p et 1 from
  antineutron.
• TOF lt 2.9 ns and l1lt 4.8 cm 4652 photon
  clusters and 447 from p
• BUT, this is on mono-energetic single particles
• TOF is more imprecise in real data
• Shower overlapping

7
Playing with more than 3 parameters

• I have to find a way for playing with more
  parameters
• Difficult task
• cant see in p-dimensional space when p gt 3
• have to eliminate correlations
• There could be a solution Principal Component
  Analysis

8
Principal Component Analysis

• Purpose
• project data points from a p-dimensional (p gt 3)
  space to a bi-dimensional space
• eliminate correlations between variables
• have to do that without distortion of the
  initial data distribution
• Means Principal Component Analysis. With this
  method one can eliminate correlation and project
  data on the best two axes

9
Methods principle

• Project cluster variables in the pattern space.
• Find D1, the straight line on which standard
  deviation of projected points is the highest.
• Do it by the same way with straight lines which
  are orthogonal to D1, and call them D2, D3 ,etc.
  These Di are the principal axes of data
  distribution.
• Directing vectors of these Di are called
  principal components of data distribution. They
  are linear combinations of pattern space base
  vectors.

10
A simple example

• PCA is already in ROOT framework, in a class
  called TPrincipal.

D2
D1
Covariance matrix
and after diagonalization
11
Results with PCA

• Only D1 and D2 are used in data analysis. Data
  points are projected on the plane (D1, D2 )

12
To be continued...

• This method has to be tested on more realistic
  simulated data
• If it works fine, estimate efficiency and purity
  in simulation
• Test on embedded events
• Apply PCA to real data