Kalman Filter Results



Kalman Filter Evolution: pictures step by step

The following plots show how the kalman filter algorithm works in fitting 3 tracks.

evolution 1
evolution 2

In the second plot, the error of one of the tracks was multiplied by 10.

Kalman Filter Fitting: track parameters improvement

under construction. See 5 muons Results


Kalman Filter Fitting: stability and commutativity

Stability

Since the Kalman Filter algorithm starts with an arbitrary initial vertex position and covariance matrix estimation, we must be sure the final result does not depend on the particular choice for those parameters.

A small initial covariance matrix would bias the final result towars the initial position and we want to avoid this.

The following plot shows the final vertex position as a function of the diagonal elements of the initial covariance matrix for a vertex fit of 5 muons coming from (0.3,0.2)cm. The starting vertex position was (0,0)cm.

Stability, in this example, is reached for a covariance matrix of about 0.15cm. The deafult values in t01.16.00 are 10cm in r-phi and 100 cm in z.

Commutativity: Does the fit depend on track order?

Since vertexing is a non-linear dynamic system, the measurement equations which relates the measurements (track parameters) with the state vector (vertex position and track momentum at the vertex) has to be linealized around a previous estimate. This procedure is called Extended Kalman Filter (EKF)

It means that the vertex estimations for tracks added in different orders will be different because the linealization points are different. In a global fit, on the other hand, all tracks are linealized around the same point and the solution is track-order independent.

The following plots ilustrates this fact:
This is the Kalman Filter path for the vertex estimation of 3 tracks fitted in different orders. The colors correspond to different orders and the points are the intermediate positions after adding 1,2 and 3 tracks in the fit.
This is a Zoom of the same plot.

The next plot is the histogram with the errors (in sigmas wrt a global fit) of the 6 different combinations of 3 tracks.
This plot shows that the difference is not to much significantly (0.3 sigmas corresponds to about 10um) but exists.

The way this problem is solved is thru the implementation of an Iterated Extended Kalman Filter (IEKF)
In this implementation all tracks (measurements) are always linealized around the same position (the previous vertex estimate), and the filter-smoother are re-iterated (using the previous vertex estimate as linealization point) until convergence. Normally less than 5 iterations are needed to reach an stable solution. However, vertices with very high chi-square, ussually need much more iterations to converge like the following plot of vertex chi-square vs mumber of iterations shows.


Kalman Filter Finders

In the current implementation (t01.16.00) there are 4 vertex finders based on Kalman Filter: The Build-Up algorithm starts with those 2-track vertex seeds that pass quality cuts and attachs closed tracks. The following plot shows an event display of the seeds in a top event.

The next plot is a comparison between 3 finders in a simple event.

In this particular event of 2 vertices, the Tear-down algorithm fails to find the secondary vertex due to the fact that 1 of the 2 tracks of that vertex was already associated to the PV, so the algorithm stops finding 1 vertex and 1 single track.

The Build-up finder finds the primary and secondary vertex, but it also finds a fake vertex due to allow multiple track-vertex associations.

The combined algorithm correclty reconstructs the event with no fakes.

The combined aproach was created with the idea of tagging b-quark jets because those events have PV-B-D cascade decays with different track multiplicities. In fact, so far, the combined algorithm has shown to have the best purity for b-tagging.
The aplication of these algorithms to tag b-jets is descripted in the b-jet identification page


Z -> mu+ mu- Results

Results using the tear-down kalman finder algorithm (VtxKalmanFinderTearDown)

5 muons Results

The following plots show the vertex distributions for a file of 5 muons generated at (-0.3,0.2)cm by Norman Graf (Thanks a lot Norman for your interest and help to this work!)

The track parameters are improved significantly in dca, phi and tan(lambda), whereas q/pt (curvature) is almost not modified. This is due to the fact that, a first order, the kalman filter does not change the track curvature but move the tracks around the vertex position.


Last updated: 11/20/00
Ariel Schwartzman