![](https://csdnimg.cn/release/download_crawler_static/12271363/bg5.jpg)
τ
lep
τ
had
(τ
had
τ
had
) SR is allowed to vary by 29% (35%)
relative to the normalization derived in the Z → μμ þ
heavy-flavor control region in order to account for accep-
tance differences between the two. An additional 20%
normalization uncertainty in the Z → ee þ light-flavor
background, related to the misidentification of electrons
as taus, is derived by comparing data and simulation in a
Z → ee control region with 0 or 1 b-tagged jets. The ZH
(ttH) background normalization is varied by 28% (30%)
based on ATLAS measurements [98,99]. The normaliza-
tions of the remaining minor backgrounds taken from
simulation are allowed to vary within their respective
cross-section uncertainties.
The uncertainty in the modeling of backgrounds due to
jets being misidentified as τ
had-vis
is estimated by varying
the fake factors and fake rates within their statistical
uncertainties and varying the amount of true-τ
had
back-
ground subtracted. Based on studies with simulated t
¯
t and
W þ jets events, a systematic uncertainty is assigned to
cover the difference in the gluon and quark flavor compo-
sition of jets misidentified as a τ
had-vis
between the signal
region and the fake-τ
had
enriched sample, parametrized as a
function of the τ
had-vis
identification BDT score. The
uncertainty in the extrapolation of FFðQCDÞ to the signal
region is estimated from the difference between the
nominal FFs and alternative ones, calculated either in
the SS region for the τ
lep
τ
had
channel or a multijet enriched
region, where Δϕðτ
had-vis
; τ
had-vis
Þ > 2.0, in the τ
had
τ
had
case. Similarly, changes in the fake-τ
had
determination
when varying the t
¯
t control region m
W
T
requirement in
simulation and data are used to estimate a systematic
uncertainty in both the fake factors and fake rates. The
overall effect of these uncertainties on the fake-τ
had
back-
ground estimate leads to an 8.4% variation of the NR signal
strength, predominantly due to the true-τ
had
subtraction in
the t
¯
t control region and the composition of the fake τ
had
.
Theory uncertainties in the signal acceptance are calcu-
lated by independently varying the renormalization and
factorization scales, the choice of PDF and each PDF set by
its uncertainties. The uncertainty in the parton shower is
taken into account by comparing the default H
ERWIG
++
with P
YTHIA
8. Uncertainties in the underlying event,
initial-state radiation and final-state radiation are accounted
for by changing the P
YTHIA
tune, but are small. The effects
of various categories of uncertainty on the measured
nonresonant signal strength corresponding to the expected
upper limit at 95% C.L. are summarized in Table II. The
individual sources of uncertainty making up the categories
listed in the table are grouped together in the final fit to
determine their correlated combined effect on the signal
strength. For all signal hypotheses, the statistical uncer-
tainties dominate.
For each signal model considered, a profile-likelihood
fit [100] is applied to the BDT score distributions
simultaneously in the three SRs to extract the signal cross
section, along with the t
¯
t and Z þ heavy-fla vor n ormal-
izations. The lattermost is constrained by including the
dedicated control region in the fit. All sources of system-
atic and s tatistical uncertainty in the signal and back-
ground models ar e implemented a s deviations from the
nominal model, scaled by nuisance parameters that are
profiled in the fit. None of th e d ominant nu isance
parameters are significantly constrained or pulled relative
to their input value by the fit. The BDT score distributions
for the nonresonant search and the G
KK
signal are shown
in Fig. 1 after performing the fit and assu min g a back-
ground-only hypothe sis . The acceptance time s efficiency
for the NR HH signal is 4.2% (2.9%) in t he combined SLT
and LTT τ
lep
τ
had
(τ
had
τ
had
) channel over the full BDT
distribution, decreasing to 3.3% (2.4%) for the two most
sensitive BDT bins. As no sig nifi cant excess over the
expected background is obser ved, upper limits are set on
nonresonant and reso nant Higgs boson pair production at
95% C.L. using the CL
s
method [101].
Table III presents the upper limits on the cross section for
nonresonant HH production times the HH → bbττ branch-
ing ratio, and comparisons with the SM prediction. The
observed (expected) limit is 30.9 fb (36.0 fb), 12.7 (14.8)
times the SM prediction. In order to compare with previous
results, the BDTs are trained and applied to the signal
sample without reweighting the m
HH
spectrum to
Refs. [9,31], giving an observed (expected) limit of
37.4 fb (33.5 fb), 15.4 (13.8) times the SM prediction.
The results of searches for resonant HH production are
presented as exclusion limits on the cross-section times the
HH → bbττ branching ratio as a function of the resonance
mass. The expected and observed limits for narrow-width
TABLE II. The percentage uncertainties on the simulated
nonresonant signal strength, i.e., the simulated NR HH yield
assuming a cross-section times branching fraction equal to the
95% C.L. expected limit of 14.8 times the SM expectation.
Source Uncertainty (%)
Total 54
Data statistics 44
Simulation statistics 16
Experimental uncertainties
Luminosity 2.4
Pileup reweighting 1.7
τ
had
16
Fake-τ estimation 8.4
b tagging 8.3
Jets and E
miss
T
3.3
Electron and muon 0.5
Theoretical and modeling uncertainties
Top 17
Signal 9.3
Z → ττ 6.8
SM Higgs 2.9
Other backgrounds 0.3
PHYSICAL REVIEW LETTERS 121, 191801 (2018)
191801-5