on the start value problem of the general track fit
DESCRIPTION
On the start value problem of the general track fit. M. de Jong. What is the problem?. General track fit is a non-linear problem multiple solutions (local minima, saddle points, etc.) requires iterative process Probability density function non-Gaussian - PowerPoint PPT PresentationTRANSCRIPT
On the start value problemof the general track fit
M. de Jong
What is the problem?
General track fit is a non-linear problem
– multiple solutions (local minima, saddle points, etc.)
– requires iterative process Probability density function non-Gaussian
– only for small range of t (random background)
– is not negative-definite (ARS token ring)¶
( )f t
2
2
( )f t
t
( )0
f tt
¶ This could be solved by taking only first hit in each PMT (thesis R. Bruijn)
Traditional strategy
1. find start values phase space too large to scan2. apply M-estimator fit enter regime where3. apply Likelihood fit obtain ultimate angular resolution
( )0
f tt
Start values are obtained using a (linear) pre-fit For an overview of the various (linear) pre-fits, see
Karl Lyons’ talk at Colmar PAW
Angular resolution of pre-fits¶
[degrees]
num
ber o
f eve
nts
[a.u
.]
K. Lyons
¶ Atmospheric muon simulation (K. Lyons)
Alternative strategy
Scan part of the 5 dimensional phase space– grid of direction angles and • closed surface ( = 4)• 3-parameter fit (x, y, t0) is linear¶
Obtain complete set of solutions– detect (hidden) symmetries• e.g. local minima
– select subset for subsequent fit(s)• subset should contain at least 1 good solution
¶ ANTARES-SOFT-2007-001
Procedure
1. choose grid angle (e.g. 5 degrees ≡ ~800 directions)2. apply 1D clustering3. make 3-parameter fit4. remove outliers and repeat fit5. sort solutions6. limit subset to N (e.g. N = 10)7. determine space angle between true track and each
track in this subset
Angular resolution ≡
smallest space angle between true track and each solution in subset
Angular resolution¶
[degrees]
num
ber o
f eve
nts
¶ Atmospheric muon simulation (same as before)
comparison
[degrees] [degrees]
5 degrees5 degrees
Median• Aart 6 degrees• Inertia Tensor 7 degrees• Direct Walk 9 degrees
Median• new method 4 degrees
fewer eventsin tail
cumulative distribution
max [degrees]
P(≤
max
deg
rees
)
Probabilities 50% 3.6 degrees 60% 4.2 degrees 70% 5.3 degrees 80% 7.8 degrees 90% 19.0 degrees
Event classification
if space angle between best quality solution and any other -but equally good- solution
is larger than some number of degrees
then event is classified as ambiguous
Event classification (II)
1. Discard event if there is 2nd solution, with:– P(2,NDF) ≥ 0.01– #hits ≥ #hits of best solution– Angular difference with best solution ≥ 20 degrees
2. Discard event if there is 2nd solution, with:– P(2,NDF) ≥ 0.01– #hits ≥ #hits of best solution – 1¶
– Angular difference with best solution ≥ 20 degrees
¶ This means that symmetry is broken by only 1 hit
cumulative distribution (II)
all eventsclass 1. (= 75%)class 2. (= 45%)
max [degrees]
P(≤
max
deg
rees
)Probabilities 50% 3.5 degrees 60% 4.0 degrees 70% 4.9 degrees 80% 6.6 degrees 90% 12.0 degrees
Probabilities 50% 3.2 degrees 60% 3.5 degrees 70% 3.9 degrees 80% 4.7 degrees 90% 6.5 degrees
Summary
• Alternative method to obtain start values– scan of directions within solid angle
• List of solutions instead of ‘one-and-only’– there is a solution in subset of 10 elements that is
closer to true track than other available pre-fits
• Detection of (hidden) symmetries– 90% of unambiguous events within 12 degrees
from true track