*** Updated, 13Jul03 ***
New result related to FITPI(see section 3)
*** Updated, 15Jul03 ***
1. Plots related to NN is added.
2. Correlation between TD and KIN is measured using loosened NN cut.
1. Neural Net function ( tuning )
I tried the tuning of NN function.
Basically, I used 2-pulse events for the tuning.
Setup cuts:
background sample;
SKIM2_OR_SKIM5, ICODEL14, COS3D, ZFRF,
B4BWBH, TIMING, tpi-tk > 2.0, PVCUT,
ipiflg.eq.0, EV5, ELVETO, TDFOOL and ERBOX
signal sample;
BOXCUTS, PVCUT, inverted B4DEDX, inverted DELC,
|tpi-trs| < 5.0, TIC, CKTRS, CKTAIL, BWTRS, B4TRS,
COS3D, LAYV4, RNGMOM, ZFRF, ZUTOUT, ICODEL14, EIC,
ipiflg.eq.0, EV5, ELVETO and TDFOOL
The main difference from previous NN tuning
is that I apply only "ipiflg.eq.0" instead of
PIFLG which includes "ipiflg.eq.0" and other cuts.
See the result:
fig 1, MLP
error vs. number of iterations
fig 2, NN
output for b.g(upper left) and s.g.(upper right), R vs A(lower left) and
R vs A*R(lower right)
In creating the plots in fig 2, I used the events after
all setup cuts except for ERBOX for b.g., and all the
events after setup cuts for s.g.(not prescaled).
From this result, the rejection of NN has increased
from ~30 to ~70.
Because the other cuts in PIFLG affect on the neurons
for NN function, the separation between background and
signal becomes good by not applying PIFLG cut.
It can be seen in the upper plots of fig 2
where the peaks are sharper than before.
*** updated, 15jul03 ***
Figures below show how NN cut works for band and tail events.
fig 20, Distributions
of T_mu before(left) and after(right) NN cut for band(top),
tail(middle),
and piscat(bottom) events
fig 21, Differences
(T_mu(before)-T_mu(after); left) and ratios(T_mu(after)/T_mu(before); right)
as
a function of T_mu before and after NN cut for band(top), tail(middle),
and piscat(bottom) events
**********************
2. The rejection of TD cut
With new NN function, I checked the rejection and
the acceptance of TD cuts
setup cuts:
rejection sample(skim2);
TRBIT_1, SKIM2_OR_SKIM5, ICODEL14, COS3D, ZFRF,
UTCQUAL, PSCUT, PVCUT and ipiflg.eq.0
acceptance sample(piscat ps by 5)
BAD_RUN, all PASS1 cuts, inverted DELC, inverted B4DEDX,
|tpi-trs| < 5.0, ICBIT, TIC, TARGF, DTGTTP, RTDIF,
TGQUALT, TGZFOOL, CKTRS, CKTAIL, BWTRS, B4TRS, COS3D,
LAYV4, RNGMOM, ZFRF, ZUTOUT, PVCUT in RS only,
BOXCUTS and ICODEL14
|
Rejection
| Acceptance |
----------------------------------------------|--------------|
|
TAIL |
BAND |
|
|--------------------------------------|
|
| all
ERBOX | all ERBOX
|
|
----------------------------------------------|--------------|
SETUP | 33547 10343 |
18538 8218 | 10131
|
EV5 | 17745 5280
| 9951 4516 |
8619 |
ELVETO | 10415 3143 |
5672 2567 | 7674
|
TDFOOL | 10343 3123 |
5615 2547 | 7663
|
TDVAR | 117
43 | 67
30 | 6008 |
----------------------------------------------|--------------|
Rej. 287+-26 241+-37
277+-34 274+-50 | 0.593+-0.005 |
fig 3, muon
time distributions in acceptance sample
fig 4, muon
time distributions in rejection sample
Total TD rejection doesn't change so much compared
to previous NN tuning.
Since PIFLG wasn't applied and more events with
small tmuav1 are left before EV5, the rejection
of both EV5 and ELVETO got smaller.
2.1 The rejection and the acceptance on 3-pulse events
I checked the rejection and the acceptance on 3-pulse events.
Here is the result.
--------------------------------------
Rejection
Acceptance
--------------------------------------
SETUP 6792
644
EV5 4999 (1.36)
498 (0.77)
ELVETO 4207 (1.19)
410 (0.82)
TDFOOL 4189 (1.00)
410 (1.00)
TDVAR 30 (140)
193 (0.47)
--------------------------------------
226+-41
0.30+-0.018
fig 5, muon
time distributions on 3P events in acceptance sample
fig 6, muon
time distributions on 3P events in rejection sample
Concerning the acceptance, it's quite lower than that on 2-pulse
events.
The acceptances of all cuts exept for TDFOOL are lower, especially
that of TDVAR.
The smaller acceptance of TDVAR may explains that events with small
tmuav1 happen
due to a tail fluctuation of pion pulse. Those events are actually
rejected a lot.
In terms of the rejection, that of both EV5 and ELVETO is smaller but
that of TDVAR
is higher. So total rejection doesn't change so much.
2.2 discarding ELVETO5
In order to discard ELVETO5, I first commented out few lines in elveto_new02.function,
c elv_p15 = .false.
c DO J=1,NELVET
c if(isev(j).eq.1) then
c elv_p15
= .true.
* write(6,*)
' Failed pass1.5 elveto cut '
c goto 2000
c endif
c ENDDO
and checked how the rejection changed.
Of cource I re-did the elveto optimization before the check.
|
Rejection
| Acceptance |
----------------------------------------------|--------------|
|
TAIL |
BAND |
|
|--------------------------------------|
|
| all
ERBOX | all ERBOX
|
|
----------------------------------------------|--------------|
SETUP | 35110 10880 |
19520 8658 | 10131
|
EV5 | 18622 5566
| 10517 4782 | 8619
|
ELVETO | 14138 4305 |
7855 3567 | 7972
|
TDFOOL | 14063 4287 |
7802 3550 | 7928
|
TDVAR | 566
197 | 355
154 | 6163 |
----------------------------------------------|--------------|
Rej. 62+-2.6 55+-3.9
55+-2.9 56+-4.5 | 0.608+-0.005 |
R_elveto = 29139/21993 = 1.325 +- 0.022
A_elveto = 7972/8619 = 0.925 +- 0.003
Even after optimization, the rejection is small.
fig 7, muon
time distribution in rejection sample (No ELVETO5)
(Replaced with new file.
Old
file is here.)
Looking at plots above, the rejection of ELVETO to the events with large tmuav1 isn't good.
Basically, ELVETO5 used to reject those events. So this method may not
work properly.
Further modification seems to be needed or it may be better to keep
ELVETO5.
*** updated, 10Jul03 ***
The table above was updated.
Main change is modifying ifail_tdfool.function which reject events
with ifail_tdfl.ne.0 .
Because we assign ifail_tdfl.eq.9 if ELVETO5 reject the event, the
rejection of TDFOOL
wasn't correct. Now the table shows the correct rejection values.
After the modification, the total TD rejection decreased dramatically.
That is because many events with tmuav1 > ~70 still exist even after
ELVETO(ELVETO5 is not applied)
which was rejected by ELVETO5.
***********************
3. FITPI
I checked whether the selection of muon pulse is good or not
when there are 2 post-pion pulses.
There is a variable called RFLAG which is output of FITPI4 (see page
42
of
TN-K029 for
detail).
The cases of RFLAG=0,1 are not important now.
So I concentrate on the cases of RFLAG=3,4,5.
For the confirmation, I applied these cuts to make plots below.
i) all PASS1 cuts, BAD_RUN, inverted DELC, inverted B4DEDX, TIC, COS3D,
ICODEL14
ii) 3-pulse events
iii) |dt31| < 30 ns
1) fig 8, 3P energy vs. 3P time difference in RFLAG=3,4,5
Looking at the upper left plot, some events contaminate into |dt31|
< 6 ns.
Why?
2) fig 9, 3P energy vs. 3P time in RFLAG=3,4,5
Looking at the upper right plot, there is a band arround 3 MeV before
3P time is 60 ns.
It seems that true muon pulse may be selected as a random pulse.
The cases of RFLAG=3, 5 look OK.
3) fig 10, muon energy vs. muon time in RFLAG=3,4,5
The cases of RFLAG=3, 5 look OK.
However, when tmuav1 is greater than 60 ns, the muon energy seems to
be incorrect.
4) fig 11, muon energy vs. 3P energy in RFLAG=3,4,5
In RFLAG=4 (upper right plot), we can see clear correlation between
muon energy
and 3P energy. It is very likely that true muon pulse is mistakenly
identified as 3rd pulse.
5) fig 12, 3P energy vs. time difference between muon and 3P in RFLAG=3,4,5
In RFLAG=4 (upper right plot), there is a band arround 3 MeV when T_mu
- T_3p > 0 ns.
6) fig 13, muon energy vs. time difference between muon and 3P in RFLAG=3,4,5
In this case, when T_mu - T_3p > 0 ns, muon looks like a random.
7) In order to confirm whether muon selection works well in the case
of RFLAG=4,
I made two plots.
fig 14, |dt1| - |dt31| (left plot) and |R_mu-R_pi| vs. |R_3P-R_pi| (right plot) in RFLAG=4
where R_mu = emuon1(1)/emuon1(2), R_pi = epion1(1)/epion1(2) and R_3P = e3p1(1)/e3p1(2) .
Time difference of muon isn't necessarily smaller than that of 3rd pulse,
but from the definition, muon area ratio mostly matches pion area ratio.
8) If I select the events with | E - 3 MeV| < 1.5 MeV and with RFLAG=4,
we can see pion life time in the both distributions.
fig 15, muon
time distribution (upper plot) and 3P time distribution (lower plot) in
RFLAG=4
From the plots above,
we may have a couple of conclutions;
a) the selection of muon is good when RFLAG=3 or 5,
b) muon may be swapped by 3rd pulse when RFLAG=4,
c) the second pulse is probably muon in most events when RFLAG=4 .
The selection criteria in RFLAG=4 may not be good.
Do we use energy as a dicriminator?
True muon has smaller energy difference from 3 MeV than random pulse.
But it may cause a bias.
I'm not sure if we should do so or not.
Otherwise, do we regard the second pulse as muon when RFLAG=4 ?
From my study, other variables aren't good for discriminator.
*** updated, 13jul03 ***
fig 18, muon
energy vs 3P energy for 3P(T_mu-T_3P > 0 ,upper left) and
for
muon(T_mu-T_3P < 0, upper right), and thier projection on x-axis(lower
left)
and
on y-axis(lower right).
In most of the case that T_mu-T_3P > 0 ns, that is, the second is muon,
it seems that muon is properly selected. On the other hand, when T_mu-T_3P
> 0,
it appears that muon is misidentified as 3rd pulse.
In order to recover the misidentification of muon as 3rd pulse,
I replaced the muon info. with the 3rd pulse info. if events are satisfied
with these conditions.
1. RFLAG=4,
2. T_mu-T_3P > 0,
3. |E_3P-E_mean| < 3*E_sigma.
where E_mean and E_sigma mean the mean muon energy and its sigma.
After this, distributions for muon look correct.
fig 19, energy
vs timing for muon(upper left) and for 3P(upper right),
E_mu
vs E_3P(lower left), and muon energy distribution(lower right)
after
the modification.
4. TD-Kinematics correlation study
For the calculation of the rejection, I apply neither RSDEDX cut nor
scattering cut.
In 1998 PNN(1) analysis , we observed the correlation between TD cut
and those kinematic cuts
in Km2 range tail events.
So I check whether the correlation still exists or not.
First, I applied RSDEDX cut as one of setup cuts and checked the rejection.
|
Rejection
|
-------------------------------------------------|
|
TAIL |
BAND |
|-----------------------------------------|
|
all ERBOX |
all ERBOX |
-------------------------------------------------|
SETUP | 1904
979 | 7226
4274 |
EV5 | 1038
538 | 3921
2314 |
ELVETO | 584
309 | 2235
1311 |
TDFOOL | 581
308 | 2218
1300 |
TDVAR | 11
4 | 24
11 |
-------------------------------------------------|
Rej. 173+-52 245+-122
301+-61 389+-117
Compared to the rejection when RSDEDX isn't applied,
the rejection doesn't change so much.
No correlation is found.
Second, I applied PRRF cut instead of RSDEDX cut and checked the rejection.
|
Rejection
|
-----------------------------------------------|
|
TAIL |
BAND |
|---------------------------------------|
|
all ERBOX | all
ERBOX |
-----------------------------------------------|
SETUP | 7381 3166
| 10681 5240 |
EV5 | 3397
1448 | 5794 2892
|
ELVETO | 2000 864
| 3340 1641 |
TDFOOL | 1981 855
| 3309 1630 |
TDVAR | 30
13 | 35
21 |
-----------------------------------------------|
Rej. 246+-45 244+-67
305+-52 250+-54
Even in this case, any correlation isn't found in Km2 range tail events.
However, in Km2 band events in ERBOX, the rejection has changed.
This is mainly due to the change of the rejection of TDVAR.
fig 16, muon
time distribution before(left) and after(right) NN cut for band(upper)
and tail(lower) events, (RSDEDX applied)
fig 17, muon
time distribution before(left) and after(right) NN cut for band(upper)
and tail(lower) events, (PRRF applied)
But I don't know exact reason for the change.
*** updated, 10jul03 ***
If the errors are taken into accout, the total TD rejections in band
events in ERBOX seem to be consistent in each case.
*** updated, 15jul03 ***
In order to increase the statistics, I loosened the neural net cut
and then measured the rejections in the cases;
Case 1 : No RSDEDX and no PR_RF is applied.
|
Rejection
|
-----------------------------------------------|
|
TAIL |
BAND |
|---------------------------------------|
|
all ERBOX | all
ERBOX |
-----------------------------------------------|
SETUP | 33547 10343 |
18538 8218 |
EV5 | 17745
5280 | 9951 4516
|
ELVETO | 10415 3143 |
5672 2567 |
TDFOOL | 10343 3123 |
5615 2547 |
TDVAR | 249
77 | 127
58 |
-----------------------------------------------|
Rej. 135+-9 134+-15
146+-13 142+-19
Case 2 : RSDEDX is applied
|
Rejection
|
-----------------------------------------------|
|
TAIL |
BAND |
|---------------------------------------|
|
all ERBOX | all
ERBOX |
-----------------------------------------------|
SETUP | 1904
979 | 7226 4274 |
EV5 | 1038
538 | 3921 2314 |
ELVETO | 584
309 | 2235 1311 |
TDFOOL | 581
308 | 2218 1300 |
TDVAR | 14
5 | 45
27 |
-----------------------------------------------|
Rej. 136+-36 196+-87
161+-24 158+-30
Case 3 : PR_RF is applied
|
Rejection
|
-----------------------------------------------|
|
TAIL |
BAND |
|---------------------------------------|
|
all ERBOX | all
ERBOX |
-----------------------------------------------|
SETUP | 7381 3166
| 10681 5240 |
EV5 | 3397
1448 | 5794 2892
|
ELVETO | 2000 864
| 3340 1641 |
TDFOOL | 1981 855
| 3309 1630 |
TDVAR | 60
27 | 72
41 |
-----------------------------------------------|
Rej. 123+-16 117+-22
148+-17 128+-20
Even though the uncertainty in case 3 is large, there seems to be no
correlation
between TD cut and Kinematics cut. Especially the correlation between
TD and
PR_RF may be tiny.
To measure more precise rejection in case 2, checking with more loosened
NN cut
is needed. This will be done soon.
***********************