1
00:00:00,000 --> 00:00:05,970
Seems like Yes, perfect. Well, we'll get
going. Okay,

2
00:00:06,419 --> 00:00:12,119
so, thanks for having me. I'm about to
bring order from university and then fn

3
00:00:12,119 --> 00:00:16,619
Baba and today I'm going to talk to you
about the measurement of disability in

4
00:00:16,619 --> 00:00:24,629
phase. She is in the best sci fi channel
by CMS at that theme TV. So, just the one

5
00:00:24,629 --> 00:00:29,699
is light of introduction. Yes is a TV
relating phase arising from the

6
00:00:29,699 --> 00:00:35,099
interference between the SDKs proceeding
directly and through mixing to a final

7
00:00:35,099 --> 00:00:41,759
state. The Standard Model prediction is
very precise. Putting a fee is roughly

8
00:00:41,759 --> 00:00:47,639
equal to minus twice with S which is one
of the angle or one of the angle of the

9
00:00:47,639 --> 00:00:53,489
unitary triangles at minus 37 million or
the one answer that is below 1 million.

10
00:00:54,059 --> 00:00:59,129
New freezes can change these values via
new particle contributing to the DSPs

11
00:00:59,219 --> 00:01:05,699
guest parking for measuring vs. vs to zip
defies the golden challenge since there is

12
00:01:05,699 --> 00:01:11,429
no predicted direct civilization only one
evaluating phase and it's easy to

13
00:01:11,429 --> 00:01:17,579
reconstruct a high signal to background
ratio. Also several other interesting

14
00:01:17,579 --> 00:01:24,059
observable Can you measure with the same
analysis such as gamma is DMS, DMS and

15
00:01:24,089 --> 00:01:31,079
lambda where lambda is an observable which
measures the direct CP violation and is

16
00:01:31,139 --> 00:01:39,479
expected to be one for this DK. What do we
need to measure? So first, we need the CP

17
00:01:39,479 --> 00:01:45,209
Angus data of the fire state which is
extracted via an angle and analysis. Then

18
00:01:45,239 --> 00:01:51,449
we need the excellent primary solution to
see the fast USPS model solution. We need

19
00:01:51,449 --> 00:01:56,699
the highly efficient flavor tagging to
infer the initial Yes, flavor and of

20
00:01:56,699 --> 00:02:02,879
course, as all recipients we need as much
statistics as possible. As for the

21
00:02:02,879 --> 00:02:09,479
candidate selection, I think it's worth to
spend a bit of time on the trigger

22
00:02:09,479 --> 00:02:14,519
strategy for this analysis. So for the
trigger for this analysis requires our job

23
00:02:14,519 --> 00:02:18,989
site to mimic candidate plus some
additional Mian that can be used to target

24
00:02:18,989 --> 00:02:23,549
the flavor of the BS via sibenik tonic
decays over the other B in the event.

25
00:02:23,909 --> 00:02:27,959
However, the requirement for the third
commune lowers the rate of selected events

26
00:02:27,959 --> 00:02:32,369
because any events that do not present the
moon at the opposite side is rejected by

27
00:02:32,369 --> 00:02:39,179
this trigger. These lower rates also allow
us to not apply our displacement cap on

28
00:02:39,209 --> 00:02:44,159
the job site at HFT level. In summary,
this triggers improve the tagging

29
00:02:44,159 --> 00:02:48,839
efficiency that is the fraction of events
for which we can give a tagging inference

30
00:02:48,839 --> 00:02:54,329
at the cost of reducing the number of
signal events. This is a schematic and

31
00:02:54,329 --> 00:02:59,459
presentation of the events were the three
new selected at the three gate level are

32
00:02:59,459 --> 00:03:06,299
highlighted. Do we have a black sock. As
for the offline selection, we apply some

33
00:03:06,299 --> 00:03:15,239
kinematic cuts at the final state particle
and DPS and our mass windows to the to

34
00:03:15,269 --> 00:03:25,169
intermediate license and our city cap of
70 micrometer. We use a total of 96

35
00:03:25,199 --> 00:03:32,999
inverse Phantom are collected in the 2017
in 2018 at 15 Db for a number of signal

36
00:03:32,999 --> 00:03:39,509
candidates around 48 49,000. I also report
here the number of candidates collected by

37
00:03:39,509 --> 00:03:45,569
CMS in another one and you can notice that
they are quite similar even if you're five

38
00:03:45,569 --> 00:03:50,309
times integrated luminosity and twice the
reduction of session from the higher

39
00:03:50,309 --> 00:03:55,679
center of mass energy and disease, in fact
that the effect of requiring an additional

40
00:03:55,679 --> 00:04:04,379
noon at target level This is an example of
a candidate events. We use it for this

41
00:04:04,379 --> 00:04:09,899
analysis, of course, that the almost two
years ago by CMS and the free noon the

42
00:04:09,899 --> 00:04:13,709
true from the job site, they want to use
it for the tagging are clearly visible.

43
00:04:15,420 --> 00:04:20,070
Moving on to the efficiencies, the
features in selecting and reconstructing

44
00:04:20,130 --> 00:04:25,230
the SDK depends on the key length, but
also on the variables as we see, we see in

45
00:04:25,230 --> 00:04:30,270
the next slide. So, the proper fit to the
decay rate model which have not described

46
00:04:30,270 --> 00:04:35,190
here, but I have a backup but we need the
parameterization of the decay length

47
00:04:35,220 --> 00:04:42,030
feature which is done separately by 2017
and 2018. And fitted in the city range 70

48
00:04:42,030 --> 00:04:47,940
meter meter which is the selection factor
up to half a centimeter with our folding

49
00:04:47,940 --> 00:04:52,890
exponential time such a VHF polynomial
order therefore, this procedure has been

50
00:04:52,890 --> 00:05:00,000
validated by fitting the D plus two Jeep
psychic lasts a lifetime in a in a

51
00:05:00,000 --> 00:05:06,240
Different the taking field and each of
these eight data sets is roughly

52
00:05:06,240 --> 00:05:09,270
equivalent to statistics to the BS sample
used to extract

53
00:05:09,300 --> 00:05:09,870
fears.

54
00:05:11,219 --> 00:05:16,349
As for the angular efficiency, they are
evaluated again separately for the two

55
00:05:16,349 --> 00:05:23,399
year c means of positivity cos CT and 50.
And the efficiency function is

56
00:05:23,399 --> 00:05:26,459
parameterize that we've spherical
harmonics and the measure and the

57
00:05:26,459 --> 00:05:32,069
polynomial up to the order of six. Moving
on to the flavor tagging as you can

58
00:05:32,069 --> 00:05:37,409
probably imagine from the choice of the
trigger for tagging is a core piece of

59
00:05:37,439 --> 00:05:43,379
these analogies. As target we use the
opposite side of the moon that use the

60
00:05:43,379 --> 00:05:48,209
moon charged as tagging feature. The moon
sees the moon is already selected to

61
00:05:48,209 --> 00:05:52,319
collaborate we have we will have a very
high efficiency that's been optimized in

62
00:05:52,349 --> 00:05:57,059
simulated events and calibrated in that
using the plus two k plus a setback

63
00:05:57,059 --> 00:06:03,509
indicates the Figaro married is the user
Are One is the tech empower, which is an

64
00:06:03,509 --> 00:06:10,079
efficiency rescale the with the dilution
efficiency is defined as the fractional

65
00:06:10,199 --> 00:06:16,079
backing events. While the dilution takes
into account the mistake events here we

66
00:06:16,079 --> 00:06:21,719
can define both from a staggered fashion
that is the fraction of events of target

67
00:06:21,719 --> 00:06:26,759
events, which carries the wrong flavor
information but also mistake probability

68
00:06:26,759 --> 00:06:34,499
that can be evaluated on event basis. This
Moon is reconstructed selected with the

69
00:06:34,499 --> 00:06:38,279
global Munich construction. These are the
construction that uses information both

70
00:06:38,279 --> 00:06:43,109
for form the tracker, the CMS tracker
system and the mirror system. We apply

71
00:06:43,139 --> 00:06:49,739
fairly loose kinematic and fully lose
kinematic selection and we also have

72
00:06:49,739 --> 00:06:55,829
developer discriminator for soft noon to
reject the background from light our drone

73
00:06:55,859 --> 00:07:00,719
mostly charged with cams in charge of
buyers, their immune cells As you can see

74
00:07:00,719 --> 00:07:06,359
from the table is overall quite loose for
maximum efficiency if use the new charge

75
00:07:06,389 --> 00:07:10,649
as tagging feature at this stage so
without having to miss tag we achieve a

76
00:07:10,649 --> 00:07:17,759
tagging efficiency of 50%. So, one in
every two events is tagged our Miss tag of

77
00:07:17,759 --> 00:07:24,029
around 30% for up bagging power of around
7% but the parameter means like

78
00:07:24,029 --> 00:07:28,559
probability enhance the total tagging
performance is also an input feature of

79
00:07:28,589 --> 00:07:32,279
the final feature. So, our fully connected
Deep Learning Network is used to

80
00:07:32,279 --> 00:07:35,519
distinguish mistake with event and
evaluate but haven't missed the

81
00:07:35,519 --> 00:07:39,689
probability at the same time. The
imperfect it can be divided into classes

82
00:07:39,899 --> 00:07:47,399
mean viable such as momentum or impact
parameter separation. We've the signal bs

83
00:07:47,429 --> 00:07:51,989
but also called environments which are
constructed using information from all the

84
00:07:51,989 --> 00:07:57,119
tracks that are found to be in a corner
center and around the moon direction and

85
00:07:57,119 --> 00:08:01,529
these can be for example, His relation.
The charge of The corn or the ratio of the

86
00:08:01,529 --> 00:08:05,819
energy between of the ratio between the
energy of the moons and the energy of the

87
00:08:05,819 --> 00:08:10,499
goal. The dnn is constructed in such a way
that the article score F is equal to the

88
00:08:10,499 --> 00:08:15,479
probability of studying the events
correctly. So they mistake a probability

89
00:08:15,509 --> 00:08:20,969
of that given event is simply one minus
these predictions prediction is validated

90
00:08:20,999 --> 00:08:26,699
and calibrated in data using tagging the
case. In the two the two plots that I'm

91
00:08:26,699 --> 00:08:32,309
showing here at the center of the slide
the art of the 2017 and 2018 calibration

92
00:08:32,519 --> 00:08:38,369
plots, where in the x axis, we have the
predict the mistake, we have been of the

93
00:08:38,369 --> 00:08:42,899
predicted mistake aggravated by the dnn,
while in the y axis we have the

94
00:08:42,899 --> 00:08:49,829
measurement mistake from mistake fraction,
indeed, they read the line and represent

95
00:08:49,859 --> 00:08:54,569
the linear function that we use to
calibrate the response of the neural

96
00:08:54,569 --> 00:08:59,549
network. And as you can see from the
values of the A and the parameter is

97
00:08:59,999 --> 00:09:07,379
compatible with the straight line by by
sector y equal x. So we have an excellent

98
00:09:07,379 --> 00:09:11,849
agreement within the prediction and
measurement and also like to mention that

99
00:09:11,849 --> 00:09:16,619
the dnn and these linear calibration
updates table to statistical fluctuation

100
00:09:16,679 --> 00:09:21,479
leading to various more systematic
uncertainty related to different mortality

101
00:09:22,679 --> 00:09:27,419
that in performance are evaluated in the
same setting in data. And they have

102
00:09:27,419 --> 00:09:34,649
reported here the figure of merit in
478,000. In the parameter Mr. Mobility

103
00:09:34,649 --> 00:09:40,139
does not change the tagging efficiency,
which is still around 50%. What's changed

104
00:09:40,169 --> 00:09:46,289
is the overall tagging power, which goes
from the 7% that quoted a couple of slides

105
00:09:46,289 --> 00:09:52,259
backs up to 10% and improvement of more
than 40%. They also put in here in the

106
00:09:52,259 --> 00:09:56,579
last line, the performance of around one
even if the comparison cannot be done

107
00:09:56,609 --> 00:10:01,199
directly, because the data center where
these performances are valuated are

108
00:10:01,199 --> 00:10:04,469
different because you're on one we will
have on addition, we will have a moon

109
00:10:05,069 --> 00:10:10,739
candidate for each events while in around
one when the moon and electron tangle were

110
00:10:10,799 --> 00:10:17,429
used, there were not a specific selection
on the opposite side. Anyway, we can do a

111
00:10:17,579 --> 00:10:22,559
compare his own normalizing the
performance by the event rate, which leads

112
00:10:22,589 --> 00:10:30,629
to a 50% high performance of these run one
of these franchise tagging algorithm

113
00:10:30,629 --> 00:10:33,299
overview with respect to grandma

114
00:10:34,590 --> 00:10:39,300
moving on to the feet and the results.
This is the fit the model that we use, I

115
00:10:39,300 --> 00:10:45,540
will not describe it in it is we have the
secret PDF, which contains the efficiency

116
00:10:45,540 --> 00:10:50,280
function, the differential decoded PDFs
and then the PDFs of the various input

117
00:10:51,390 --> 00:10:56,490
variables. We have a PDF that the mode is
the main background which is mainly

118
00:10:56,490 --> 00:11:01,170
combinatorial. And then we have a PDF that
more than the picture Background from a

119
00:11:01,170 --> 00:11:09,120
big YouTube site case study that became
background from lambda be to say a key

120
00:11:09,120 --> 00:11:14,490
proton is estimated to be negligible for
the data set of candidates that we have

121
00:11:14,490 --> 00:11:21,600
selected. This slide I'm showing you the
one dimensional projection of the feet for

122
00:11:21,600 --> 00:11:27,690
six of the input variables we have the
master the key length and its uncertainty

123
00:11:27,690 --> 00:11:37,650
and the three Angular variables. While in
this line, I painting you on a summary of

124
00:11:37,650 --> 00:11:43,200
the systematic uncertainty that we have
estimated, there are plenty that I cannot

125
00:11:43,230 --> 00:11:49,560
describe it described in detail, but I
have highlighted in bold the red that the

126
00:11:49,560 --> 00:11:53,880
leading systematic uncertainty for the
most interesting parameter and the fourth

127
00:11:53,880 --> 00:11:58,350
yes, these are the modern bias which is
evaluated through MonteCarlo experiment

128
00:11:58,350 --> 00:12:04,230
the end angle and efficiency When we
propagate the limited statistics using it

129
00:12:04,350 --> 00:12:11,340
used to evaluate the deficiency functions.
As for the results, I have simulated here

130
00:12:11,370 --> 00:12:18,420
on the left vs is measured to be equal to
minus 11 million other with statistical

131
00:12:18,420 --> 00:12:23,070
uncertainties are 15 illa ns is not a
conservative 10 million other these

132
00:12:23,100 --> 00:12:28,260
results is in agreement with the Standard
Model prediction. We have also measure

133
00:12:28,530 --> 00:12:33,450
DMS, which is again an agreement with
standard model prediction gamma s which is

134
00:12:33,450 --> 00:12:37,260
consistent with the world average and DMS,
which is consistent with a vote every

135
00:12:37,260 --> 00:12:43,140
debate as like tension lambda is also
consistent with not direct specialization

136
00:12:43,140 --> 00:12:48,420
that is lambda equal to one. I'd also like
to mention that is the first measure in

137
00:12:48,420 --> 00:12:56,100
biosimilars of E m s. And that is out of
these analogies are in agreement with the

138
00:12:56,100 --> 00:13:02,460
one presented by CMS at the at the end and
For a combined or systematic uncertainty

139
00:13:02,460 --> 00:13:08,850
are considered uncorrelated in the bottom
right plot, you can see the true the CS

140
00:13:08,880 --> 00:13:16,800
Lambda Gamma Contoso for the 30 dV results
in blue, the eight dv one in green and the

141
00:13:16,800 --> 00:13:24,270
combined in one in Atlanta and the
Standard Model prediction in in black, as

142
00:13:24,270 --> 00:13:31,650
you can notice from the sides of these
contours the new trigger strategy which

143
00:13:31,650 --> 00:13:36,720
trades the number of events for lagging
power space off for us in fact that even

144
00:13:36,720 --> 00:13:43,200
if we have the same number of events, the
uncertainty on PS is reduced by a factor

145
00:13:43,200 --> 00:13:48,360
of two basically, why does not improve the
gamma So, which do not make use of the

146
00:13:48,360 --> 00:13:53,190
flavor information. And therefore, it is
mainly driven by Statistics and since we

147
00:13:53,190 --> 00:13:58,020
have the same number of events in a TV and
13 TV, you can see that in the gamma

148
00:13:58,020 --> 00:14:04,320
direction. There certainly is Medical.
This leads to my last couple of slides for

149
00:14:04,320 --> 00:14:09,870
the slide of summary. So, the US and the
UK with different DMS are measured using

150
00:14:10,350 --> 00:14:17,010
40,000 can be selected at 15 dV
corresponding to an integrity nosy or 96

151
00:14:17,370 --> 00:14:22,380
investment in events are selected using a
non displaced trigger that requires an

152
00:14:22,380 --> 00:14:26,910
additional means which exploit which is
exploited to be fair the flavor of the

153
00:14:26,940 --> 00:14:31,020
stretch pays off in terms of lagging
performance leading to a significant

154
00:14:31,020 --> 00:14:35,880
reduction of the PS and certainly however
the limited the number of events prevents

155
00:14:35,880 --> 00:14:40,620
improvement on the gamma is the results of
this analysis are combined into Windows

156
00:14:40,650 --> 00:14:47,010
painted at a TV and are in agreement with
the standard. As for the outlook and the

157
00:14:47,010 --> 00:14:51,630
future prospects, here I have a table with
the comparison with other LSE experiment

158
00:14:51,630 --> 00:14:59,640
the for the best which upside kk Chairman
around the fee master all of these results

159
00:14:59,640 --> 00:15:02,730
are combination of round one and partial

160
00:15:03,780 --> 00:15:08,730
round two results. I would also like to
mention that the new D gamma is

161
00:15:08,760 --> 00:15:14,790
theoretical prediction with smaller
uncertainty is available the pre printed

162
00:15:14,790 --> 00:15:22,140
dates, I think December 2019. And the same
method is at both the 15. tv one and the

163
00:15:22,140 --> 00:15:27,600
combination is in agreement with this new
prediction which has almost half of the

164
00:15:27,750 --> 00:15:33,210
certainty of the previous one. It's also
easy to see that the DMS show some

165
00:15:33,210 --> 00:15:39,660
tensions between experiment and hopefully
a foreign measurement will clarify the

166
00:15:39,660 --> 00:15:43,980
situation as for the CMS food plan, CMS
plans to analyze the following two dates

167
00:15:44,070 --> 00:15:48,180
ending a complimentary figure that they're
required to displace in job site plus two

168
00:15:48,210 --> 00:15:53,700
charges tracks the electron and in that
charge a flavor tagging algorithm will be

169
00:15:53,700 --> 00:15:58,140
used in this sample that do not have a
specific selection for the opposite side

170
00:15:58,380 --> 00:16:03,390
and the effective statistics. Number of
candidate events times again, power is

171
00:16:03,390 --> 00:16:09,330
expected to improve by a factor one and a
half true with respect with these results.

172
00:16:09,930 --> 00:16:11,400
And thanks for your attention.

173
00:16:13,259 --> 00:16:18,299
Thank you very much, Alberto. And thank
you for being on time. Other comments or

174
00:16:18,299 --> 00:16:23,009
questions for the speaker? Starting from
the attendees? Please remember to raise

175
00:16:23,009 --> 00:16:34,349
your hand. If not, I had a quick question
on the dnn usage, I see that it's

176
00:16:34,349 --> 00:16:39,329
including the use of variables that are
sensitive to the overall event shapes,

177
00:16:40,379 --> 00:16:48,329
such as the isolation and so on. I wonder
if this can introduce a potential Angular

178
00:16:48,329 --> 00:16:52,679
effect on the analysis and if that can if
that is taken into context we look into or

179
00:16:52,679 --> 00:16:54,209
maybe it's just limited

180
00:16:54,540 --> 00:17:01,530
while it's taking into account because the
PT and the ADA But also the separation

181
00:17:01,530 --> 00:17:10,140
with respect to the signal our input
feature, so even if you do not input soul

182
00:17:11,970 --> 00:17:17,730
parameters so we have some variable that
take into consideration the kinematics of

183
00:17:17,730 --> 00:17:23,820
the however we did not explore in detail.
If the tagging depends on

184
00:17:26,190 --> 00:17:27,960
on the angle of the angles on

185
00:17:29,130 --> 00:17:33,990
that chord evolution will support a bias
in your angular distribution in the fit,

186
00:17:34,380 --> 00:17:34,710
right.

187
00:17:39,480 --> 00:17:46,110
I don't see why because we input all the
angles we input to every event is used for

188
00:17:46,110 --> 00:17:51,420
the analysis even the impacted one so no
event is discarded by the tagging. In

189
00:17:51,420 --> 00:17:59,160
fact, as I said before a DMS do not use
any tagging information. The tagging

190
00:17:59,160 --> 00:18:08,370
information is spread Until in all if I
show you the key rate model, the the

191
00:18:08,400 --> 00:18:15,270
tagging inferences I lighted but this key
in Switzerland only in only alpha of the

192
00:18:15,270 --> 00:18:21,510
PDF, which is sensible if you go through
the tables to vs. While bigamous is not

193
00:18:21,510 --> 00:18:28,320
sensible. So I don't see the pattern we
don't see why you should introduce an

194
00:18:28,320 --> 00:18:34,890
Angular bias sees all the information
always visible from the feet.

195
00:18:36,480 --> 00:18:41,130
All right, what are we going to take these
often in this? Yes, yes. Is there any

196
00:18:41,130 --> 00:18:42,000
other question?

197
00:18:43,140 --> 00:18:44,310
I have one.

198
00:18:44,430 --> 00:18:45,090
Yes, please.

199
00:18:46,440 --> 00:18:51,120
Okay, so it's very interesting your, your
trigger strategy.

200
00:18:52,470 --> 00:18:57,810
Besides very impressive result, I was I
was just wondering about the efficiency.

201
00:18:57,810 --> 00:19:05,100
So you it's It's clear you sacrificed
statistics to have such a good target

202
00:19:05,100 --> 00:19:13,410
power. How much do you sacrifice? Because
I'm thinking if he you could, for example,

203
00:19:13,590 --> 00:19:20,730
have another sample with the most
statistics and slightly less or quite a

204
00:19:20,730 --> 00:19:26,160
bit less taking power and still improve
your sensitivity. Maybe not to fight us

205
00:19:26,160 --> 00:19:33,270
too much but to delta mass for example.
Yes, you have completely arrived. In fact

206
00:19:33,270 --> 00:19:39,900
that is exactly the CMS plan for the
fuller untrue analysis because we have a

207
00:19:39,900 --> 00:19:45,690
trigger that as you say, that has a very
high statistics and we have no position on

208
00:19:45,690 --> 00:19:49,890
the opposite side. So we are more
obviously sat standard t good Wi Fi

209
00:19:49,890 --> 00:19:56,880
statistic and standard tanking power. And
when we add these two trigger together, we

210
00:19:56,880 --> 00:19:57,660
will have

211
00:19:59,010 --> 00:20:00,900
benefits from our Both of them

212
00:20:02,760 --> 00:20:14,520
as for the statistics that we sacrifice, I
think it's a factor of five of leverage

213
00:20:14,730 --> 00:20:20,490
and having trade should be a factor of
five or between these two trigger. So if

214
00:20:20,490 --> 00:20:26,070
you look at the tagging power, which is
around the 10%, okay? Which cheese's

215
00:20:26,700 --> 00:20:34,080
roughly a factor of between seven and 10
with respect to one on one, and then if

216
00:20:34,080 --> 00:20:39,330
you divide by five, which is their
interest, you sacrifice, you still have

217
00:20:40,680 --> 00:20:42,030
a 50% higher,

218
00:20:43,350 --> 00:20:49,530
bagging power with respect to run one. So,
it's still Worf to deployed these three

219
00:20:49,530 --> 00:20:55,380
gear to improve the tagging power, but of
course, to perform the measurement at

220
00:20:56,640 --> 00:21:04,440
best, let's see. Also they are there
Trigger should be used. So, you would have

221
00:21:04,440 --> 00:21:09,690
offered a lot of statistics. You will have
an electron and a jet that that flavor

222
00:21:09,690 --> 00:21:20,250
target but you will also have a very high
neon tagging tagging very high efficiency,

223
00:21:20,310 --> 00:21:21,780
very high performance innocent

224
00:21:25,650 --> 00:21:27,210
sounds great, thanks

225
00:21:29,280 --> 00:21:35,910
efforts, thermal other questions then
thank the speaker again and move on to the

226
00:21:35,910 --> 00:21:38,010
next door clothes