Searches for Dark Matter Production in Two Fixed Target
Experiments: HPS and LDMX
A DISSERTATION
SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL
OF THE UNIVERSITY OF MINNESOTA
BY
Thomas Brice Eichlersmith
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Jeremiah Mans
Jan, 2025
© Thomas Brice Eichlersmith 2025
ALL RIGHTS RESERVED
Acknowledgements
There are many people that have earned my gratitude for their contribution to my time
in graduate school. While I will struggle to name them all, I do wish to specifically thank
those people who have helped teach me their knowledge of physics along the way: Tim
Nelson, Cam Bravo, Matt Graham, PF Butti, Omar Moreno, Philip Schuster, Natalia
Toro, Stefania Gori, Lene Kristian Bryngemark, Einar Elén, Erik Wallin, Christian
Herwig, Nhan Tran, Cristina Mantilla Suarez, Lauren Tompkins, Valentina Dutta, Ruth
Pöttgen, and Alic Spellman.
Of particular note is my advisor Jeremy Mans. Professor Mans has been wonderfully
flexible as I have changed course while navigating the uncertain waters of graduate
school. Thank you.
Finally, there are many friends who I have met while in graduate school or with
whom I have gotten closer. Your support outside of physics has been necessary for me
over the years and I am in your debt.
i
Dedication
To my friends and family trying to comprehend anything I have done over the past six
and a half years, I hope this helps.
To a future graduate student desparately trying to understand these experiments
and their code, I also hope this helps and I wish you luck.
ii
Abstract
Astrophysical evidence strongly indicates the presence of particulate Dark Matter (DM)
within our universe; however, the specific particle nature of DM is still unknown. The
wide variety of possible DM particles produces a similar range of experiments focused on
probing these different categories of possible DM. This work describes two experiments
taking different approaches to search for Light DM residing in the 1MeV-1GeV mass
range being produced by electron interactions. The Light Dark Matter eXperiment
(LDMX) is a proposed fixed target experiment designed for a missing momentum search
with an additional, orthogonal missing energy search channel described here. The Heavy
Photon Search experiment (HPS) is another fixed target experiment designed for a
displaced vertex search with distances of O(10cm) which are not probed by longer
baseline experiments. Specifically, a search in HPS data for a specific Light DM model
with a strongly-coupled dark sector enabling a higher expected production rate while
keeping the characteristic decay length within HPS’s acceptance is also presented.
iii
Contents
Acknowledgements i
Dedication ii
Abstract iii
Contents iv
List of Tables vii
List of Figures ix
I Laying the Groundwork 1
1 Introduction 2
1.1 Standard Model of Particle Physics . . . . . . . . . . . . . . . . . . . . . 4
1.2 Standard Model Analogies . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Particle Mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.2 Displaced Particle Decays . . . . . . . . . . . . . . . . . . . . . . 9
2 Dark Matter 12
2.1 Dark Matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Evidence for Dark Matter . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Particle Nature of Dark Matter . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Invisible Signature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
iv
2.5 Visible Signature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
II LDMX 22
3 Light Dark Matter eXperiment 23
3.1 Missing Momentum Signature . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 The Beam Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Detector Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4 Mid-Shower Simulation 31
4.1 General Data Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.1 Data Processing Stages . . . . . . . . . . . . . . . . . . . . . . . 33
4.1.2 Biasing and Filtering Technique . . . . . . . . . . . . . . . . . . 34
4.2 Standard Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.1 Dominant Contributors in this Analysis . . . . . . . . . . . . . . 37
4.2.2 Expanding Production of these Processes . . . . . . . . . . . . . 39
4.3 Dark Matter Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.1 G4DarkBreM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.2 MadGraph/MadEvent . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.3 Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5 Missing Energy Search 46
5.1 Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Background Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.3 Systematics and Background Uncertainty . . . . . . . . . . . . . . . . . 55
5.4 Reach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
III HPS 61
6 The Heavy Photon Search Experiment 62
6.1 Strongly Interacting Massive Particles . . . . . . . . . . . . . . . . . . . 62
6.2 Detector Apparatus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
v
6.2.1 Silicon Vertex Tracker . . . . . . . . . . . . . . . . . . . . . . . . 65
6.2.2 Electromagnetic Calorimeter . . . . . . . . . . . . . . . . . . . . 67
7 Alignment 70
7.1 Alignment Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.2 Detector Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
8 Data Set 80
8.1 Collected Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.1.1 Pair 1 Trigger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
8.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
8.3 Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
8.4 Analysis Pre-Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
9 Displaced Vertex Search 88
9.1 Signal Yield . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9.2 Reconstruction Categories . . . . . . . . . . . . . . . . . . . . . . . . . . 93
9.3 Physics Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
9.4 Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
9.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
9.5.1 Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
9.5.2 Sensitivity and Exclusion . . . . . . . . . . . . . . . . . . . . . . 103
IV Conclusion 106
10 Conclusion 107
10.1 Simulation Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
10.2 Tracking Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
10.3 Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
References 110
vi
List of Tables
4.1 Parameters of mid-shower biased samples. In all of the samples used in
this study, we use the same value for the Sorting and Biasing Thresholds. 36
4.2 Configuration of the simulation samples used in this analysis. Ebeam is
the beam energy being studied (4 or 8 GeV in this work). mA is the mass
of the A’ in MeV, ϵ is the dark brem mixing strength, EECal Front
primary is the
energy of the primary electron at the front of the ECal, EA′ is the energy
of the generated A’, Etot nuc is the total energy transferred to nuclear
interactions during the event, and Etot µ is the total energy of produced
muons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.1 Cut-flow analysis comparing background and various signal hypotheses
for the simple cuts used in this analysis. The event yield for the back-
ground sample is calculated using the event weights and represent the
number of events out of 1013 EoT equivalent. The signal efficiency is rel-
ative to the full simulation sample. The efficiency and event yield values
on a given row are for after the analysis stage of that row. The first table
is the cutflow for the 4GeV beam, and the second table is for the 8GeV
beam. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.1 Values for cuts used within the Pair 1 Trigger during data collection of
the HPS 2016 dataset. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
9.1 Polynomial coefficients for radiative fraction and acceptance functions. 91
9.2 Parameters for Equation (9.14). . . . . . . . . . . . . . . . . . . . . . . . 100
vii
9.3 Region definitions for use in background estimation via sidebands. Region
F is the signal region in which we are searching for an excess. mVD
is
the mass point we are searching for, σm is the detector mass resolution
evaluated at mVD
, ycut0,min is the optimized cut value evaluated at mVD
,
and yfloor0,min is the maximum value of y0,min such that region C has at least
one thousand events in it. . . . . . . . . . . . . . . . . . . . . . . . . . 102
viii
List of Figures
1.1 Comic #2351 from XKCD[1]. If able to be designed from scratch, the
names of things in the standard model could be more helpful. . . . . . . 2
1.2 The Standard Model of Particle Physics showing the twelve fermions and
five bosons, their various properities (mass, charge, spin), labels (box
and circle colors), and interactions (brown loops). Credit to Cush on
wikipedia for providing this diagram freely accessible and usable for any
purpose. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Feynman diagram for the bremsstrahlung process where a charged lepton
(ℓ−) interacts with a nucleus (Z) via a photon (γ) and then emits another
photon before recoiling. . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Energy of the outgoing electron (Ee) for standard processes (dominated
by bremsstrahlung) in gray and some dark bremsstrahlung (see Sec-
tion 2.4) in colors. Figure 3 from [2]. . . . . . . . . . . . . . . . . . . . 8
1.5 Dipiction of neutral kaon and anti-kaon mixing including the underlying
Feynman diagram that results from applying the Standard Model (SM)
to this system. This figure was created by user NikNaks on Wikipedia
and is licensed under CC Attribution-Share Alike 3.0 Unported. . . . . 10
1.6 Image of CERN’s first liquid hydrogen bubble chamber from 1960 [3]
with pink annotations added. The pink circle highlights the location
described in the text where a Λ baryon decays into two particles (most
likely a proton and a negative pion). The pink dashed line traces out a
possible path the Λ traveled before its decay. . . . . . . . . . . . . . . . 11
2.1 Comic #1758 from XKCD[1]. Generally, claiming that either Einstein or
Newton was wrong will be a bad time. . . . . . . . . . . . . . . . . . . . 12
ix
https://commons.wikimedia.org/wiki/File:Kaon-box-diagram-with-bar.svg
https://creativecommons.org/licenses/by-sa/3.0/deed.en
2.2 Depiction of the velocity of stars within a galaxy as a function of their
distance from the galactic center. The dotted line is a prediction of this
relationship using General Relativity (GR) along with the mass tabulated
from the visible starts while the data points (and the solid line fitted to
them) are what are actually seen in galaxies today. . . . . . . . . . . . 15
2.3 From [4], the co-moving cosmological number density of Dark Matter
(DM) as a function of the universe temperature. As the universe cools,
the number density decreases until the DM becomes too sparse to interact
with other DM particles, “freezing” to a specific number density until
today. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Mass scale of Thermal Relic DM. The regions in red are excluded by ap-
plying the thermal relic assumption to our observations of the universe’s
early evolution. me is the electron mass and mp the proton mass. . . . 18
2.5 Feynman diagram of how a massive field Φ could allow for a standard
photon (γ) to mix with a dark photon (A′). . . . . . . . . . . . . . . . . 18
2.6 Feynman diagram for the dark brem process. . . . . . . . . . . . . . . . 19
3.1 Diagram showing relative rates of background processes within LDMX
along with how they motivate various aspects of the design. /E stands for
“missing” energy or energy that is “lost” to neutrinos that are extremely
unlikely to be detectable within LDMX. . . . . . . . . . . . . . . . . . . 24
3.2 Diagram of LDMX detector apparatus with a representation of a signal
event where a dark brem occurs within the target. Credit to Christian
Herwig for original development of diagram. . . . . . . . . . . . . . . . 28
3.3 Rendering of LDMX detector apparatus focusing on tracker, target, and
ECal. The magnet would fully encompass the tracker, target and trigger
scintillator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Diagram of LDMX Electromagnetic Calorimeter (ECal) design showing
the longitudinal segmentation (top and bottom right) and the transverse
segmentation (bottom left). Credit to Joe Muse. . . . . . . . . . . . . . 30
x
4.1 Flow chart of how data is processed in the Light Dark Matter eXperiment
(LDMX) framework. Each processor has the ability to “add” data to the
event as well as “get” data from the event. The processors are run in a
user-defined sequence. Data can also be loaded from one or more input
data files where the event data from those files is loaded into memory
before the first processor is run. After all processors are done with an
event, it is saved to the output data file. . . . . . . . . . . . . . . . . . . 32
4.2 Diagram of processing stages within LDMX showing the external sources
of data or software that are used in those stages as well as which stages
are commonly done within the centralized processing framework. . . . . 34
4.3 The total reconstructed energy within the ECal as a fraction of the beam
energy separated by the total nuclear energy within the event (the energy
going into the photon-nuclear and electron-nuclear processes). . . . . . . 39
4.4 Fraction of hits within the Hadronic Calorimeter (HCal) that were caused
by muons or anti-muons. The events lying within the zero bin still have
HCal hits caused by other types of particles. . . . . . . . . . . . . . . . . 40
4.5 Longitudinal location (z) where the µ+ was produced within the simu-
lation for a variety of biasing factors B for the 4 GeV beam (left) and
8 GeV beam (right). The rate R is the equivalent number of unbiased
Electrons on Target (EoT) divided by the CPU time necessary to pro-
duce the sample. We can observe that while increasing B also improves
the rate R, we eventually over-bias and lose access to the late tail of the
distribution (red, puple, and brown in the left plot and gray in the right
plot). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 ECal reconstructed energy EECAL as a fraction of the beam energy EBeam
comparing the unbiased and biased samples scaled to the same EoT. The
4 GeV (8 GeV) beam case is shown on the left (right). . . . . . . . . . . 42
4.7 Distributions of simulated signal events with a 4GeV beam. The energy
of the produced dark photon is shown on the left while the longitudinal
location where the production occurred is on the right. . . . . . . . . . . 44
xi
5.1 The total reconstructed energy in all layers of the ECal (EECal). The sig-
nal and background distributions are normalized such that their integral
is one. The 4GeV beam is shown on the left and the 8GeV beam is shown
on the right. The events falling into bins with EECal > 1.5GeV (3.16GeV)
for 4GeV (8GeV) are omitted from this plot but included in efficiency
calculations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2 Variables used in signal selection for the ECal as Target (EaT) analysis
channel. In each figure, all other cuts are applied except for the variable
in question. Some of the bins are empty in which case the upper Poisson
limit given the total sample size is drawn with error bars. The grey line
shows the selection cut. The background sample is the enriched nuclear
and dimuon samples. The top (bottom) row shows the 4GeV (8GeV)
beam. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3 Studies of how the tracker requirement affect the simulation samples. . . 52
5.4 The total reconstructed energy in all layers of the ECal (EECal) as a
fraction of the beam energy (EBeam) for all samples that pass the selection
criteria except the selection on ECal energy. The 4GeV beam is shown on
the left and the 8GeV beam is shown on the right. The gray lines mark
the edges of the analysis bins used to estimate the expected exclusion
limit and the black line is the upper limit on the ECal energy which also
serves as the upper limit of an analysis bin. . . . . . . . . . . . . . . . . 53
5.5 Exponential fit of the total simulated background distribution. The
shaded region is a 95% confidence band on the fit which is interpreted
as the uncertainty on the value of the fit. Some bins in the simulated
distribution are empty in which case the upper Poisson limit given the
sample size is drawn as an error bar in that bin. . . . . . . . . . . . . . 54
5.6 Background prediction within the three final analysis bins (blue) com-
pared to the unconstrained simulation prediction in gray. The uncer-
tainty on the background prediction is taken from the 95% confidence
band shown on the fit. Some bins in the simulation prediction are empty
in which case the upper Poisson limit given the sample size is drawn as
an error bar in that bin. . . . . . . . . . . . . . . . . . . . . . . . . . . 55
xii
5.7 The total reconstructed energy in the ECal EECAL comparing ten differ-
ent smeared calibrations (colors) to the original unsmeared calibrations
(black). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.8 Effect of smeared ECal calibrations on the ECal hit RMS for events that
pass the original trigger. Outside of statistical fluctuations in the lowest-
populated bins, the smearing has a less than 5% effect throughout the
range of potential cuts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.9 Reach (left) and maximum allowed yield (right) comparisons for the 4GeV
beam case as we vary the cut on maximum HCal photoelectrons (PE).
We observe minimal movement in the final reach as a function of this cut;
nevertheless, the biggest change occurs around a threshold of 30. . . . . 58
5.10 The sensitivity of the EaT analysis channel compared to other experi-
ments led by NA64[5, 6], BABAR [7], and COHERENT[8] (gray), Scalar,
Majorana, and Pseudo-Dirac theory expectations (black, top-to-bottom),
and LDMX projections (colors). The blue (orange) line corresponds to
the 4GeV (8GeV) beam studied in this work. The red line is the nomi-
nal LDMX analysis sensitivity for the Missing Momentum (MM) analysis
channel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.1 Diagram of dark brem production of a dark photon followed by its decay
back into an electron-positron pair through Strongly-Interacting Massive
Particle (SIMP) composite particle VD with the emission of a lighter
SIMP composite particle πD. The two-particle vertex VD −A′ is allowed
since they share enough quantum properties to mix. . . . . . . . . . . . 63
6.2 Feynman diagrams for the SM process of trident production. . . . . . . 64
6.3 Full rendering of Heavy Photon Search experiment (HPS). Beam would
enter the detector from the left and exit to the right if it does not interact. 66
6.4 Simplified diagram of HPS showing an example displaced decay within
the named subsystems. The blue arrow is some DM candidate particle
that has a macroscopic lifetime causing the decay vertex to be observably
displaced from the production within the target. . . . . . . . . . . . . . 66
xiii
6.5 Figure 10 of [9]. Rendering of the Silicon Vertex Tracker (SVT) with
the vacuum enclosure shown in gray, active silicon components in red,
and readout electronics in green. Similar to Figure 6.3, the beam would
traverse this rendering from left to right. . . . . . . . . . . . . . . . . . 68
6.6 Figure 27 of [9]. Depiction of an event passing the pair trigger in the
2016 data collection run. This depiction also displays the variables used
in order to evaluate events and make the trigger decision. . . . . . . . . 69
7.1 Momentum error derived from injecting a certain amount of position error
(blue and orange) compared to normal distributions (red and purple) with
known momentum error. These distributions are normalized to have unit
integrals. The trials were done in a simplified experimental model with
quantities similar to HPS: a 2GeV electron curving within a 1T magnetic
field, the sensors measuring the position were separated by 5 cm but their
separation were also allowed to deviate from the true value by the position
error given. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.2 Plotting all six global coordinates completely specifying position and ori-
entation of each of the sensors in the HPS tracking detector. Two different
versions of the in-software detector model are plotted; however, on this
scale, they overlap one another. . . . . . . . . . . . . . . . . . . . . . . . 76
7.3 Plotting all six global coordinates completely specifying position and ori-
entation of each of the sensors in the HPS tracking detector relative to
the “Original” detector from Figure 7.2. We can now observe the quan-
titative differences between the different versions. . . . . . . . . . . . . . 77
7.4 The transverse impact parameter of the tracks at the target d0 =
√
x20 + y20
which should be centered on zero to align with the known beamspot at
the target. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
7.5 The magnitude of the track momenta which should be centered on the
known beam value of 2.3GeV. . . . . . . . . . . . . . . . . . . . . . . . . 79
8.1 Relative efficiency of vertex quality pre-selection requirements. . . . . . 85
8.2 Number of vertices passing the quality pre-selection. . . . . . . . . . . . 86
xiv
8.3 Relative efficiency of pre-selection of events. Nvtx is the number of ver-
tices passing the vertex pre-selection requirements. Since other triggers
for different purposes are present in the data, the Pair 1 Trigger is applied
to data but not to simulation. . . . . . . . . . . . . . . . . . . . . . . . . 87
9.1 Depiction of dark photon yield calculation using a single run of the data
(∼ 1.6%). The gray vertical lines show the bounds of the search using
the nominal ratio mA′/mVD
= 1.66. . . . . . . . . . . . . . . . . . . . . . 91
9.2 Expected signal yieldNsig for some rudimentary example efficiencies. Red
contour lines are given at 1 and 1.6% (the approximate fraction of this
data subsample) to give a sense of scale. Values are “clipped” to the
color bar (for example, a signal yield of 2 × 104 is colored yellow which is
marked as 1 × 103) so we can focus on the important orders of magnitude. 93
9.3 Diagrams showing L1L1 (left) and L1L2 (right) vertex examples. . . . . 94
9.4 Diagrams of examples for how different types of L1L2 vertices effect the
observed values of y0. Truly displaced vertices (left) have both tracks
whose y0 is far from 0mm while fake displaced vertices (middle) and not
displaced (right) have at least one if not both tracks with y0 ≈ 0mm.
The diamonds are the reconstructed vertex, the solid circles are the re-
constructed hits, and the empty circles are “missed” hits either because
the particle did not pass through those sensors (left) or due to some
physical or electronic inefficiencies (center and right). . . . . . . . . . . . 96
9.5 Distributions of cut variables for a few example mass points and ≈ 10%
of the full data sample. The vertices are required to be L1L2, have their
momentum sum be within the signal region, and their invariant mass
within the specified mass window. . . . . . . . . . . . . . . . . . . . . . 97
9.6 y0,min distribution after the other selections (L1L2, Psum in SR, recon-
structed mass lying in mass window, Vertex Projection Significance (VPS)
< 4, and σy0,max < 0.4mm). . . . . . . . . . . . . . . . . . . . . . . . . 98
9.7 Binomail significance (left) being maximized after the VPS< 4 and σy0,max
< 0.4mm cuts leading to cuts (right, blue) which are smoothed into a
continuous function (right, orange) as described in the text. . . . . . . 99
xv
9.8 y0,min distribution for the 10% data sample with the ycut0,min function over-
layed in red. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.9 Signal efficiency as a function of z and mass scale mVD
for this analysis
(left) and L1L1 (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.10 Search results for the full 2016 data set. . . . . . . . . . . . . . . . . . . 103
9.11 Signal yield for the L1L2 reconstruction category with the described
cuts showing the theoretical expecation (left) and the maximum allowed
(right) derived using the Optimum Interval Method (OIM). . . . . . . . 104
9.12 Sensitivity of this analysis with contour drawn including systematic errors
surpressing the expected signal yield. . . . . . . . . . . . . . . . . . . . . 105
xvi
Part I
Laying the Groundwork
1
Chapter 1
Introduction
Figure 1.1: Comic #2351 from XKCD[1]. If able to be designed from scratch, the names
of things in the standard model could be more helpful.
The long and winding road of a dissertation is not always neatly packaged into a
template - this can easily be seen in my own experience. While focusing much of my
time on the technical aspects of data processing software for a proposed (yet to be built)
experiment, I found myself near the end of my journey without real data to analyze
from which I can make real conclusions about the physical world. This motivated me
to participate in another experiment – extremely similar to the original one – sharing
2
3
theoretical motivations, technological designs, and even people. This experience of
participating in two similar experiments has provided an abundant field of learning
opportunities (as well as roadblocks) for me and my journey as a new physicist.
Physics is a lazy science. I would go even further and state that this reductionist
perspective is a main attraction for many physicists. We1 want to avoid memorizing as
much as possible; therefore, motivating the idea of condensing sets of observations into
“laws” that can be represented in an even more compact mathematical form. We make
up these “laws” and the vocabulary surrounding them not to decieve but merely to make
communication about our observations and the experiments that make these observa-
tions easier. We sometimes debate the origins of these laws and their true philosophical
meaning, but often, on a day-to-day basis, the background of them is unimportant. The
interesting work comes from testing them, breaking them, and remaking them. For our
purposes here, this is what a “theory” is: a package of laws and their mathematical
forms with which we can make predictions of the observations of our experiments.
While I speak in the context of all physics, in reality, I am residing within a small
corner. Primarily concerned with individual particles and how they interact with one
another, my field can be described as investigating the foundations of the universe.
Our experiments require giving these particles comparatively large amounts of energy,
contextualizing the name High Energy Physics (HEP). Giving such small particles
such large amounts of energy requires extrodinarily large and complex apparatuses.
HEP is filled with experiments many stories tall, collaborations consisting of hundreds
of institutions and thousands of people, and observations lasting years if not decades.
This grand scale is helpful to keep in mind when I speak of the experiments in this thesis
– constructed and operated by only a few institutions (less than 10 in each) and only
dozens of collaborators. Contrary to many other sciences and experimental methods,
these experiments still last longer than a typical doctoral student career. While I report
on these experiments, I will not detail their full operation, design, or capabilities. I will
focus on the work that I was able to contribute; hopefully providing a building block for
these experiments and HEP to grow in the future, and give only the necessary context
for parts with which I am less familiar.
1Here I use the “royal we” as representing my point of view of the culture within the field. It should
in no way be construed as scientific or exact statements and does not represent every physicist’s point
of view.
4
Before partitioning this thesis according to the two experiments in which I par-
ticipated, Section 1.1 summarizes our current laws and its representative theory that
best predict our set of observations within HEP, Chapter 2 will explore the theoretical
ground on which both of the experiments rest which will also provide necessary vocab-
ulary for discussing these two experiments. Part II presents the work of my initial years
as a graduate student developing data processing and realistic simulation software for
a proposed experiment. This part includes a description of this proposed experiment
in Chapter 3, the simulation infrastructure for it in Chapter 4, and an analysis of this
simulated data showcasing this experiment’s performance with ∼ 2.5% of its target data
set in Chapter 5. Part III describes the work of my last years showing the difficulties
of working with data taken in the real world (with all the complexities that implies).
Similar to Part II, this part first describes the experimental setup in Chapter 6, how
we can tune our detector model to better match how it was constructed in Chapter 7,
the simulation and collected data used are detailed in Chapter 8, and the analysis per-
formed on this data is described in Chapter 9. We then conclude with Chapter 10 which
returns back to this high-level-view from which we can discuss what we have learned
about physics via these two experiments.
1.1 Standard Model of Particle Physics
What makes up all the stuff around us? This question has been posed, debated, dis-
cussed, and studied since the ancient times. Our best current understanding is that
stuff is made of particles that interact with one another. While “particle” is difficult to
strictly define qualitatively, we have a model for them and their interactions which we
creatively named the “Standard Model” (SM). Figure 1.2 displays the known elemen-
tary particles, some of their basic properties, and how they interact with one another.
Not all of the particles represented within this diagram are of critical importance to this
work, but some deserve further description.
The two lowest mass quarks – the up and the down – are the fundamental con-
stituents of protons and neutrons which make up the nuclei of atoms which make up all
the day-to-day stuff we interact with. These stay bound together within these nuclei via
the strong nuclear force (mediated and represented in Figure 1.2 by the gluon). These
5
Figure 1.2: The Standard Model of Particle Physics showing the twelve fermions and
five bosons, their various properities (mass, charge, spin), labels (box and circle colors),
and interactions (brown loops). Credit to Cush on wikipedia for providing this diagram
freely accessible and usable for any purpose.
6
quarks and the top row of leptons also have an electric charge enabling them to interact
via with electromagnetic force. The electromagnetic force, represented and mediated by
the photon, is the force responsible for light and magnetism. The lowest mass charged
lepton (electron) is also very abundant – most commonly residing in orbits around nu-
clei helping form atoms. Both experiments in this thesis utilize beams of electrons that
have been accelerated to high energies and then impinge upon a material of high density
to initiate interactions for further study.
The mathematical expressions to calculate how these particles interact are compli-
cated. As mentioned earlier, physicists are lazy and so we have developed a method
to represent the specifics of these interactions graphically in a form where many key
aspects of the interaction can be understood at an intuitive level without needing to
write down any of the long mathematical formulae. These representations are called
“Feynman diagrams” after the 20th century physicist Richard Feynman. Feynman dia-
grams allow us to represent interactions by defining a set of “vertices” that are allowed
by our theory and then constructing processes from these vertices that include the in-
and out- going particles that we wish to study. Such diagrams are more than visual
sketches; they can be algorithmically converted into precise mathematical forms which
can be used to directly calculate measurable predictions of the model containing this
set of vertices. The conversion of these diagrams into mathematical forms and the cal-
culation of these forms has been written into various computer programs (I say this
to emphasize that these diagrams directly represent the mathematics that can be used
to calculate them). Figure 1.3 shows an example Feynman diagram representing the
bremsstrahlung process where a charged lepton (ℓ−) interacts with a nucleus (Z) via a
photon (γ) and then emits another photon before recoiling. We can see three verticies
in this diagram: two “fundamental” vertices where the lepton and photon lines connect
and one “effective” vertex where the photon connects with the nucleus. The fundamen-
tal vertices are actually strictly defined within the Standard Model, but the effective
vertex represents a helpful approximation that is accurate at the energy scales we are
studying.2
Due to the complexity of the nucleus itself, the calculation of Figure 1.3 into any
2The distinction between fundamental and effective vertices is not a well defined one. I find it helpful
here, but it is not necessarily made elsewhere in physics literature.
7
ℓ−
Z
ℓ−
γ
γ
Figure 1.3: Feynman diagram for the bremsstrahlung process where a charged lepton
(ℓ−) interacts with a nucleus (Z) via a photon (γ) and then emits another photon before
recoiling.
observables is difficult to do without any approximations. However, one of the initial
parametrizations of bremsstrahlung offers a good glimpse of how it behaves [10]. When
doing experiments, we often count the number of particles produced with certain criteria.
These “rates” can then be translated into estimates of process cross sections (how likely
a process is to occur producing particles with the observed properties) with knowledge
of how the experiment was conducted.
These cross sections are also calculatable from diagrams like Figure 1.3 and the
initial parameterization can be summarized as
dσ
dEγ
∝ 1
Eγ
(1.1)
This shows the dominant shape of the cross section – it grows as the energy of the
produced photon (Eγ) gets smaller. The full expression prevents the cross section
from diverging as the photon energy approaches zero; nevertheless, the general inverse
relationship between the rate and the energy of the produced photon holds in both the
SM calculations and observations from experiments. Figure 1.4 shows another view of
this rate: instead of looking at the energy of the produced photon (Eγ), we can inspect
the energy of the outgoing electron after it emits the photon (Ee). This is more readily
compared to other processes that do not produce a photon and we can observe a large
difference between this SM process and other possible processes that we search for.
This inverse relationship is somewhat special and is mostly due to the fact that
photons do not have any mass; however, this special relationship opens a door towards
potential discovery. The vast majority of electrons interacting with a material will do
8
Figure 1.4: Energy of the outgoing electron (Ee) for standard processes (dominated
by bremsstrahlung) in gray and some dark bremsstrahlung (see Section 2.4) in colors.
Figure 3 from [2].
so softly, only emitting low energy photons, which distinctly separates these electrons
from ones that undergo rarer processes (or perhaps so-far undiscovered processes). Both
experiments described in this work make heavy use of this fact.
As the work carried out for this dissertation is experimental rather than the defi-
nition and calculation of a model, I intend to stay within this diagrammatic realm –
representing the theoretical calculations being tested with these diagrams and limiting
the scope of mathematical formulae presented. Occasionally, the weight of these vertices
which can be interpreted as the strength of the interaction is presented along with the
vertices in order to give a sense of scale and connect these diagrams to the formulae
that are included here.
1.2 Standard Model Analogies
The SM has been extremely successful in describing the current world of particle physics,
including some intricate behaviors of particles that were incredibly confusing when first
encountered. When attempting to describe unknown phenomena (Chapter 2), we often
draw on this previous experience to form new ideas and models. This section, while
a brief detour, will help give context for two key behaviors that the models we are
9
searching for would give their particles.
1.2.1 Particle Mixing
All of the models tested for in this work use the idea of “particle mixing” in order to
connect the realm of theoretical/undiscovered particles with the realm of SM particles.
In the quantum arena, we treat individual particles in a probabilistic manner – often
times when we observe a particle with certain properties, it can be observed again in a
different context with different properties. While this sounds odd, several experiments
have supported this random nature of the smallest particles over the course of twentieth
century physics. As a well-understood example, neutral kaons have been observed to
“swap” with their own anti-particle, which we can describe via this concept of mixing.
Kaons are particles that contain two quarks, one of which is a strange quark (or its
anti-quark). A neutral kaon (K0) and its anti-particle (K0) are nearly indistinguishable
– they have the same electric charge (0) and the same mass amongst other shared
properties – but we can distinguish them by observing the types of particles they produce
when they decay. Describing kaons as consisting of two quarks allows us to understand
the mixing ofK0 andK0, diagrammed in Figure 1.5. From a higher-level view, where we
are unaware of the constituents of the kaon, it appears like the K0 and K0 are swapping
places with some probability. Our experiments can observe this probability by counting
the frequency with which we see K0 compared to K0. We can also use our model of the
kaons (as composed of quarks) and calculate this same probability (specifically, using
the Feynman diagram shown in Figure 1.5). Comparing the two gives us information
about the SM and how the quarks interact with one another.
1.2.2 Displaced Particle Decays
Ever since particle physicists have started designing detectors, we have also observed
particles that are “missed” by our detector mechanisms. Figure 1.6 shows an early
example of this where a “bubble chamber” shows the paths of charged particles passing
through the liquid (the white lines). In the middle of this image, you can observe two
lines emerging seemingly out of nowhere (the red annotations were added later). The
mechanism that allows a bubble chamber to observe the path of particles requires the
10
W
K0
s du,c,t
W
+
d u,c,t s
K0
Figure 1.5: Dipiction of neutral kaon and anti-kaon mixing including the underlying
Feynman diagram that results from applying the SM to this system. This figure was
created by user NikNaks on Wikipedia and is licensed under CC Attribution-Share Alike
3.0 Unported.
particle to have a charge, so a neutral particle (a Λ in this case) does not leave a path
itself. The only signal that it was there is the displaced vertex from which two other
particles (probably a proton and a negative pion) emerge (emphasized with a red circle
in Figure 1.6). This is a “displaced decay” of a particle which was produced by a high
energy (16GeV) negative pion interacting with the liquid in the bubble chamber.
While particle physicists have developed other detector apparatuses, some of which
can observe neutral particles more directly, this more indirect detection mechanism
is still extremely useful especially if a model being tested has particles that have no
way of directly interacting with SM matter (the stuff we make our detectors out of).
Part III of this thesis focuses entirely on an experiment designed to search for these
displaced vertices, the displacment of the vertex coming from the presence of an as-yet
undiscovered particle being produced and then decaying back into SM particles.
https://commons.wikimedia.org/wiki/File:Kaon-box-diagram-with-bar.svg
https://creativecommons.org/licenses/by-sa/3.0/deed.en
https://creativecommons.org/licenses/by-sa/3.0/deed.en
11
π−
Figure 1.6: Image of CERN’s first liquid hydrogen bubble chamber from 1960 [3] with
pink annotations added. The pink circle highlights the location described in the text
where a Λ baryon decays into two particles (most likely a proton and a negative pion).
The pink dashed line traces out a possible path the Λ traveled before its decay.
Chapter 2
Dark Matter
Figure 2.1: Comic #1758 from XKCD[1]. Generally, claiming that either Einstein or
Newton was wrong will be a bad time.
New phenomena have puzzled physicists throughout history, but the confusion sur-
rounding dark matter – estimated to be roughly 85% of the total matter in the universe
– is surely one of the biggest puzzles. While many astronomical observations at various
scales have confirmed the existence of dark matter, we have not yet seen any observa-
tions of its particle interacting with our detectors. This has ruled out many models for
dark matter’s particle nature, but there are still many more available which can both
explain the current observations of dark matter gravitational and cosmological effects
while skirting the limitations defined by the lack of observation in particle experiments.
12
13
2.1 Dark Matter?
The term DM rose out of a brutally honest description of the present level of under-
standing. “Matter” originates from comparison to the particles that make ourselves up
and with which we are familiar – both this new phenomena and our normal “stuff”
interact gravitationally, attracting other groups of “matter” into complex cosmological
systems. The adjective “dark” refers to the literal fact that we cannot see it with our
telescopes. Other matter out in the universe can be seen via the light that it gives off (or
reflects), so the fact that this matter is not visible in this way motivates the adjective.
That is it. Dark Matter (DM) is merely a shorthand for this aspect of the universe
that is both known to exist within the cosmos and whose particle nature is completely
unknown.
2.2 Evidence for Dark Matter
The first (replicable) evidence physicists had for an unseen material floating throughout
the universe was the observation of galactic rotation curves [11, 12]. These observations
measure the speed of different stars within a galaxy and compare this speed to the
distance of that star from the center of the galaxy. We can calculate this relationship
using GR [13] and the observations differ drastically from this calculation. The stars
within galaxies we’ve observed move much faster than GR would predict (Figure 2.2)
leaving us with two explanations: either GR is not the correct theory to use in this
situation or there is more un-seen mass floating within the galaxy allowing these stars
to move faster without leaving the galactic orbit.
Other indirect measurements give us additional ways to access information about
this odd phenomena. Within the framework of GR, since energy and mass actually
warp the fabric of spacetime, we expect to see light itself follow a bent path around
massive objects - a phenomenum that is called gravitational lensing and is observed and
well modeled by GR’s predictions [14]. The accuracy of GR within this context – a
mass and distance scale similar to the rotation curve oddities also observed – put more
requirements on any modified theory of gravity that could both explain the rotation
curves and gravitational lensing. Additionally, measurements on some of the largest
scales and from the early universe display signs of a certain mass density attributed to
14
matter that does not interact in the same ways as our normal matter (called “baryonic”
matter).
In the early universe, standard matter was compressed into high densities and ex-
isted at high temperature creating a sea of plasma. Gravity attracts while pressure
from the squeezing of this plasma repulses which produces oscillations in the density
of matter in space and time. The repulsive pressure within the plasma originates from
particles interacting electromagnetically with each other, so the oscillations would be
disrupted by the presence of extra matter interacting gravitationally but not interact-
ing electromagnetically. These Baryon Acoustic Oscillations (BAO) are imprinted on
our snapshot of this early-universe plasma, the Cosmic Microwave Background (CMB),
which we can measure with a high degree of accuracy and fit to various models of what
existed at this time of the universe. The best fit of these models corresponds to only
∼ 5% of the mass density being normal matter like we see today (“baryonic”) while the
rest is composed of material that only significantly interacts gravitationally [15].
Additional astromomy observations from Type1a supernovae [16], fitting models of
big bang nucleosynthesis [17], and constraints on non-particulate theories explaining
these phenomena (for example mini black holes [18]) all allow us to conclude pretty
comfortably that DM exists as a particle in our universe.
2.3 Particle Nature of Dark Matter
The theoretical possibilities explaining DM are broad [19] even when excluding ourselves
to the assumption that the DM phenomenum is explained by the existence of a DM
particle. The CMB observations are tied to very early on in the existence of the universe,
so it is natural to assume that this DM particle has existed since the start of the universe
alongside our normal matter particles. With this in mind, we can outline a few criteria
that must be met by a proposed DM candidate.
• Dark There has been no detection of these particles via the light observed with
our telescopes; therefore, the DM has to not interact via the electromagnetic force.
• Long Living Measurements of DM’s mass density and presence agree across time
(from as early as the CMB era), so DM needs to have a long lifetime.
15
Figure 2.2: Depiction of the velocity of stars within a galaxy as a function of their
distance from the galactic center. The dotted line is a prediction of this relationship
using GR along with the mass tabulated from the visible starts while the data points
(and the solid line fitted to them) are what are actually seen in galaxies today.
16
• Universal Density Since we can indirectly measure a DM mass density on cos-
mological scales, we impose the requirement that any DM model needs to allow
for this density.
• Thermal Relic Both DM and standard matter have similar cosmological densi-
ties, so we expect some interaction (even a weak one) should connect their origins
to the early universe allowing both to exist from the Big Bang. This criterion is
not as firm as the others (DM models correctly describing the current observations
while avoiding this criterion do exist); nevertheless, this criterion is well motivated
and it is satisfied by all of the models for DM studied in this work.
Even with these assumptions and the requirements they imply, there still exist a plethora
of theoretical models that can satisfy all of them.
The upside is that a thermal-relic assumption closely connects the mass of individual
DM particles to the interaction strength it has with standard matter. In this assumption,
the DM evolves along with the universe allowing its number density to follow the density
of standard matter until the standard matter does not have enough energy to produce
more DM. While the universe continues to expand, the DM continues to decay into
standard matter until it becomes too sparse to annihilate with itself and is thus “frozen”
at a specific number density. Since this “frozen” density changes depending on how easy
it is for the DM to interact with standard matter, the “frozen” number density goes
down as the interaction strength increases (Figure 2.3). The additional requirement of
the observed astronomical mass density connects the mass of individual DM particles
to their interaction strength with standard matter.
mχ ↔ observed mass density ↔ dN
dV
↔ thermal relic hypothesis ↔ ⟨σv⟩ (2.1)
This connection allows us to define strict “thermal relic targets” which can help be
measuring sticks for how well our experiments search for DM (these targets will appear
in plots later).
In addition to the connection between interaction strength and particle mass implied
by the thermal relic hypothesis, it also puts some loose bounds on mass of individual
particles (Figure 2.4). If the mass is too high (≳ 100 TeV), the DM will be too strongly
coupled to the standard matter and would be over-produced within the early universe.
17
Figure 2.3: From [4], the co-moving cosmological number density of DM as a function
of the universe temperature. As the universe cools, the number density decreases until
the DM becomes too sparse to interact with other DM particles, “freezing” to a specific
number density until today.
18
∼ 1 MeV ∼ mp ∼ 100 TeV
Too few “Light” DM “WIMPs” Too many
Figure 2.4: Mass scale of Thermal Relic DM. The regions in red are excluded by apply-
ing the thermal relic assumption to our observations of the universe’s early evolution.
me is the electron mass and mp the proton mass.
γ A′
Φ
Φ
Figure 2.5: Feynman diagram of how a massive field Φ could allow for a standard photon
(γ) to mix with a dark photon (A′).
Highly precise agreement between nuclear abundance observations and predictions from
a model of the early universe with only SM particles provides the lower limit; allowing a
thermal relic DM mass ≲ 1MeV would then break the consistency between this model
and several of its predictions [17, 20, 21].
The mass scale of thermal relic DM is further divided by the mass of the proton
(mp). Above this threshold, the DM could be interacting with standard matter through
the standard Weak Force (as described in Chapter 1); thus, they are named Weakly
Interacting Massive Particles (WIMPs). This phase space was first searched due to
its theoretical simplicitity: no new forces, just an extra particle (or two) creating the
clouds of DM we see today. Unfortunately, these searches have not found evidence of
a WIMP-like signature[22, 23, 24] and the phase space has become tighter and more
exluded after many years of searching. Below mp, the DM requires a new force which is
even weaker than the Weak Force. In comparison to the heavier WIMPs, this category
of DM is called “Light” DM.
In general, the definition of a new force of nature is not well constrained; however, we
can define a “baseline” model that can represent more concrete theories in the context
of our experiments. For this situation, we postulate the existence of a massive gauge
boson that represents an additional UD(1) symmetry of nature. Since this additional
symmetry has the same structure as the electromagnetic interaction whose gauge boson
19
ℓ−
Z
ℓ−
A′
γ
Figure 2.6: Feynman diagram for the dark brem process.
is the standard photon, we generally refer to this postulated massive boson as a dark
photon. Without any additional assumptions, this new dark photon does not interact
with any standard matter, so we also assume that some previously-unknown massive
fields are able to interact with both. These other fields need to be sufficiently massive
(or sufficiently weakly coupled) so that they only allow the standard and dark photons
to weakly mix. Figure 2.5 diagrams how the standard and dark photons mix which can
then be effectively represented by a new vertex in our model.
A′
ℓ+
ℓ−
⇒ A′
ℓ+
ℓ−
γ ∝ ϵ
(2.2)
where the mixing strength ϵ encapsulates the effect of a massive field Φ at the energies
of our experiments (presumably low enough to avoid creation of a Φ directly) and is
generally small (≪ 1).
This new vertex within our model for the universe enables a new process to occur:
so-called “Dark Bremsstrahlung” (dark brem) where a charged particle exchanges a
standard photon with a nucleus and then emits a dark photon and recoils.
But what happens to the dark photon after it interacts with standard matter via
this vertex? It cannot be the long-lived DM we view in the universe today because
this vertex allows for it to decay back into standard matter eventually (and in some
models, rather quickly). This is where we expand on the idea of a “dark sector”. In this
description of DM, we already have a dark photon representing some force and a very
20
heavy field that can interact with it. Suppose this “dark sector” also has other particles
(like the standard sector) – one (or more) of which could be long-lived and represent
the DM we observe in the universe today (so-called “DM candidates”). We can further
partition this category of models depending on what happens to the dark photon after
we produce it within an experiment.
2.4 Invisible Signature
One of the simplest options is to hypothesize another particle that can take on the role
of long-lived DM and only interact “within” this dark sector (i.e. it only interacts with
the dark photon from our point of view). Calling this particle χ (and its anti-particle
χ), we then have an additional dark sector vertex.
A′
χ
χ
∝ αD
(2.3)
where αD is a parameter representing the strength of this dark sector interaction, limited
to be below 1 but usually much larger than ϵ so that this interaction is more likely than
the one with standard matter. If the mass of the dark photon is greater than twice the
mass of the χ (mA′ > 2mχ), then this vertex can occur immediately after a dark photon
is produced within an experiment. Since the χ does not interact with our normal SM
particles, the energy they have is “lost” from our perspective. Part II focuses on a
proposed experiment designed to precisely measure all of the visible energy so that this
“invisible signature” of DM being produced could be observed.
When comparing analyses across DM search literature, it is common to parameter-
ize the DM phase space with an effective interaction strength y and the mass of the
candidate DM mχ that would constitute the astronomical DM observed today. The
parameter y is related to the parameters of our “baseline” model by
y = αDϵ
2
(
mχ
mA′
)4
21
where we make standard, benchmark choices αD = 0.5 and mA′ = 3mχ when converting
from (ϵ,mA′) to (y,mχ).
2.5 Visible Signature
There is no a priori reason for mA′ > 2mχ – the quantum nature of these fundamental
particles allows for “virtual” dark photons to have more or less energy than precisely
mA′ allowing them to mediate the interactions between χ and SM particles, so we must
accomodate the possiblity that mA′ < 2mχ. In this case, there would be a significant
probability that the dark photon, after it is produced, would convert back into SM
particles. Since these decay products are observable by our detectors, this signature is
called visible.
These visible signatures are a bit more complex to model due to the fact that both
a production process and a decay process need to occur according to the model of DM
we are testing. With this in mind, many models that provide more detail about how
particles exist within a dark sector have been created, each of which providing a specific
estimate for production and decay rates. Part III focuses on an experiment looking for
these visible signatures and my work specifically was investigating one of these more
intricate dark sector models in detail. Section 6.1 provides more detail on the dark
sector model studied in this work.
Part II
LDMX
22
Chapter 3
Light Dark Matter eXperiment
LDMX is a proposed fixed-target experiment aiming to definitively explore the thermal
relic light dark matter phase space. Even as a proposed experiment, it has a detailed
plan for construction, a beam already in construction, and well established connections
with current technologies used within HEP. While LDMX is not yet built, it has a
well formulated simulation infrastructure that can realistically model how the detector
design responds to various types of interactions happening within it.
3.1 Missing Momentum Signature
LDMX aims to search for an invisible signature – with known incident particle kinemat-
ics, measuring the outgoing particle kinematics allows for a natural deduction. All of
the incoming momentum must exit somehow and so if the detector is unable to observe
some momentum (some momentum is “missing”) frequently enough, we can conclude
that some other, previously-unknown, process is taking place and carrying momentum
away from the experiment.
Precisely understanding both the incident and outgoing momenta requires knowledge
of the currently known processes and how to detect the particles they produce. The
center of Figure 3.1 displays an ordering of known processes based on how frequently
they occur given an incident electron with 4GeV of energy. This chart can be broken
into three regions.
23
24
K± decay
in ECal
� ! µ+µ�
� ! hadrons
� ! 1n/K0
L + soft
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
� ! K± + soft
AAACDXicbVA9SwNBEN3zM8avqKXNYhQEIdyJYMqAjWATwaiQi2Fus5cs7t4eu3NKOPIHbPwrNhaK2Nrb+W/cxBSa+GDg8d4MM/OiVAqLvv/lzczOzS8sFpaKyyura+uljc1LqzPDeINpqc11BJZLkfAGCpT8OjUcVCT5VXR7MvSv7rixQicX2E95S0E3EbFggE5ql3bDLigFNDSi20MwRt/Ts5swVfTAaSq3OsZBu1T2K/4IdJoEY1ImY9Tbpc+wo1mmeIJMgrXNwE+xlYNBwSQfFMPM8hTYLXR509EEFLetfPTNgO45pUNjbVwlSEfq74kclLV9FblOBdizk95Q/M9rZhhXW7lI0gx5wn4WxZmkqOkwGtoRhjOUfUeAGeFupawHBhi6AIsuhGDy5WlyeVgJ/EpwflSuVcdxFMg22SH7JCDHpEZOSZ00CCMP5Im8kFfv0Xv23rz3n9YZbzyzRf7A+/gGGWCbiQ==AAACDXicbVA9SwNBEN3zM8avqKXNYhQEIdyJYMqAjWATwaiQi2Fus5cs7t4eu3NKOPIHbPwrNhaK2Nrb+W/cxBSa+GDg8d4MM/OiVAqLvv/lzczOzS8sFpaKyyura+uljc1LqzPDeINpqc11BJZLkfAGCpT8OjUcVCT5VXR7MvSv7rixQicX2E95S0E3EbFggE5ql3bDLigFNDSi20MwRt/Ts5swVfTAaSq3OsZBu1T2K/4IdJoEY1ImY9Tbpc+wo1mmeIJMgrXNwE+xlYNBwSQfFMPM8hTYLXR509EEFLetfPTNgO45pUNjbVwlSEfq74kclLV9FblOBdizk95Q/M9rZhhXW7lI0gx5wn4WxZmkqOkwGtoRhjOUfUeAGeFupawHBhi6AIsuhGDy5WlyeVgJ/EpwflSuVcdxFMg22SH7JCDHpEZOSZ00CCMP5Im8kFfv0Xv23rz3n9YZbzyzRf7A+/gGGWCbiQ==AAACDXicbVA9SwNBEN3zM8avqKXNYhQEIdyJYMqAjWATwaiQi2Fus5cs7t4eu3NKOPIHbPwrNhaK2Nrb+W/cxBSa+GDg8d4MM/OiVAqLvv/lzczOzS8sFpaKyyura+uljc1LqzPDeINpqc11BJZLkfAGCpT8OjUcVCT5VXR7MvSv7rixQicX2E95S0E3EbFggE5ql3bDLigFNDSi20MwRt/Ts5swVfTAaSq3OsZBu1T2K/4IdJoEY1ImY9Tbpc+wo1mmeIJMgrXNwE+xlYNBwSQfFMPM8hTYLXR509EEFLetfPTNgO45pUNjbVwlSEfq74kclLV9FblOBdizk95Q/M9rZhhXW7lI0gx5wn4WxZmkqOkwGtoRhjOUfUeAGeFupawHBhi6AIsuhGDy5WlyeVgJ/EpwflSuVcdxFMg22SH7JCDHpEZOSZ00CCMP5Im8kFfv0Xv23rz3n9YZbzyzRf7A+/gGGWCbiQ==AAACDXicbVA9SwNBEN3zM8avqKXNYhQEIdyJYMqAjWATwaiQi2Fus5cs7t4eu3NKOPIHbPwrNhaK2Nrb+W/cxBSa+GDg8d4MM/OiVAqLvv/lzczOzS8sFpaKyyura+uljc1LqzPDeINpqc11BJZLkfAGCpT8OjUcVCT5VXR7MvSv7rixQicX2E95S0E3EbFggE5ql3bDLigFNDSi20MwRt/Ts5swVfTAaSq3OsZBu1T2K/4IdJoEY1ImY9Tbpc+wo1mmeIJMgrXNwE+xlYNBwSQfFMPM8hTYLXR509EEFLetfPTNgO45pUNjbVwlSEfq74kclLV9FblOBdizk95Q/M9rZhhXW7lI0gx5wn4WxZmkqOkwGtoRhjOUfUeAGeFupawHBhi6AIsuhGDy5WlyeVgJ/EpwflSuVcdxFMg22SH7JCDHpEZOSZ00CCMP5Im8kFfv0Xv23rz3n9YZbzyzRf7A+/gGGWCbiQ==
increasingly rare
photo-nuclear
... ν (no e–)
incoming outgoing
100
10-1
10-2
10-3
10-4
10-5
10-6
10-7
10-8
10-9
10-10
10-11
10-12
10-13
10-14
10-15
10-16
…
relative
rate
(4 GeV)
irreducible
“invisible” backgrounds ≪ 10-16
“visible”
backgrounds
(EM interactions)
⌫⌫̄
⌫ (Møller + CCQE)
e� e�
EN
+µ+µ�
+hadrons
...
+e+e–
trident
Veto Handles
bremsstrahlung +hard �
Track momentum
ECal Energy
ECal BDT
ECal Track
Track multiplicity
HCal Hits
Design driver
charged-current
Gaussian energy
fluctuations
Irreducible
prompt E
Rare reactions ➝
products escape
ECal and/or
anomalous
energy deposition
/
ECal depth,
resolution, layout (no
projective cracks)
Design Drivers
HCal depth, segmen-
tation, veto energy
threshold
ECal 3D granularity,
MIP-sensitivity
SRT acceptance
Figure 3.1: Diagram showing relative rates of background processes within LDMX along
with how they motivate various aspects of the design. /E stands for “missing” energy
or energy that is “lost” to neutrinos that are extremely unlikely to be detectable within
LDMX.
The blue region at the top are the most frequent types of processes, but are simul-
taneously less complicated. The processes in this region (as the right side shows) put
requirements on the design of the ECal, making sure it can quickly and faithfully re-
construct the outgoing energy in photons and electrons. The ECal must be able to veto
the first several orders of magnitude in basic energy fluctuations including simple inter-
actions in the target like bremsstrahlung or trident production of an electron-positron
pair.
Entering the pink region is where the background processes become rarer and more
complicated. One of the first ways to partition these backgrounds is whether charged
secondary particles are produced within the target. These “prompt” backgrounds help
define the recoil tracker’s design and are thus left to be caught by it. More frequently, a
photon is produced within the target (which is not observable by the tracker) and then
undergoes complicated photon-nuclear interactions within the ECal producing particles
that are difficult for the ECal itself to observe. The HCal is able to detect the presence
of these hadrons and muons, acting as a “pure” veto in the sense that the signal process
25
should never create significant activity within it.
All of these subsystems can collaborate to help LDMX reject known SM backgrounds
down to ∼ 10−16 fraction of all incoming electrons where extremely rare known processes
that produce invisible products (neutrinos ν) begin to emerge.
Detailed study of the missing momentum search strategy[25, 2, 26] shows that LDMX
can reject all simulated backgrounds to a relative rate of ≳ 10−14. This performance is
accomplished through the basic design drivers outlined above and in Figure 3.1, but also
through the additional granularity of the ECal enabling the use of a Boosted Decision
Tree (BDT) to distinguish between background and signal events using features of the
showers within the ECal.
Cycles of detailed simulation and redesign have led to a detailed construction plan
for the LDMX detector apparatus. The subsequent sections of this chapter detail this
design as well as the beam it is expected to receive.
3.2 The Beam Line
LDMX is planned to receive electrons from the superconducting Linear Accelerator
(Linac) at SLAC National Accelerator Laboratory. The SLAC Linac can provide high
energy, high rate, and low intensity electron beams for the various experiments it hosts.
Specifically, the Linac Coherent Light Source (LCLS) is used to guide the beam (of a
certain energy) towards the experimental hall - the upgraded phase of LCLS (LCLS II
[27]) is currently under construction and is what will be used for running with LDMX.
The experiment is hosted in End Station A (ESA) at SLAC which requires an
additional upgrade to the accelerator complex in order to recieve its beam. LESA
(Linac to ESA) beamline [28] is also currently being constructed and will be ready for
test beam in early 2025. Part of the infrastructure that transfers the beam to ESA
(Sector 30 Transfer Line – S30XL) is already constructed and LDMX components are
expected to participate in a test beam runs with it.
26
3.3 Detector Design
As described above, LDMX is a missing momentum experiment and its design is focused
on measuring both the incoming and outgoing momenta of charged particles interacting
with a thin target such that any momentum given to undetectable (dark) particles can
be precisely determined. In addition, the energy of neutral particles must be measured.
This design has led to four subsystems each with specialized roles.1
1. Tracker Measure charged particle momenta both before (“Tagger”) and after
(“Recoil”) the target, using a dipole magnetic field of 1.4T.
2. Electromagnetic Calorimeter (ECal) Measure the total energy of electrons,
positrons, and photons.
3. Hadronic Calorimeter (HCal) Veto additional particles difficult for other sub-
systems to measured (muons, pions, hadrons,...).
4. Trigger Scintillator Count the number of electrons incident on the target in
time to make a trigger decision (more on what a “trigger” is below).
Figure 3.2 displays these subsystems in a diagram along with a representation of a
dark brem interaction occurrring within the target. Figure 3.3 shows a rendering of the
detector design.
The Tracker (purple in Figure 3.2) is a thin silicon strip detector modeled after the
HPS tracker. These silicon strip sensors are arranged in layer pairs where one layer
is angled slightly askew relative to the other in the pair to enable reconstruction of
three dimensional hit locations. The part of the tracker upstream of the target (to the
left in Figure 3.2) is named the “tagger” since its purpose is to measure the incident
electrons’ momenta, rejecting electrons with momentum below 30% of the expected
beam momentum. The tagger is situated within the bulk of the magnetic field enabling
highly precise measurement of the high incident momentum. The other part of the
tracker located downstream of the target (to the right in Figure 3.2) is named the
“recoil” tracker since its job is to measure the momenta of all charged particles recoiling
1These subsystems also take on additional roles when the full breadth of the LDMX physics program
is taken into account. This description just focuses on the DM search.
27
from interactions within the target. While it is not located within the magnet volume,
it is still situated within the fringe field enabling it to maintain track reconstruction
down to the lower-momentum products of interesting processes within the target.
In many HEP experiments, a “trigger” system is necessary in order to filter data
defined to be interesting by the experiment from the wealth of uninteresting (or normal)
data that is expected to be produced at a much higher rate. These trigger systems are
the first filter that any data goes through and are designed to help the experiment
obtain a statistically large data sample without collecting an overwhelming amount of
data. The ECal is the primary detector subsystem responsible for making this trigger
decision due to its excellent energy resolution and fast measurement capabilities. As a
search for missing momentum, the ECal requires less than 30% of the incoming beam
energy to be observed within the first twenty sensitive silicon layers.
The Trigger Scintillator (yellow-orange in Figure 3.2) is designed to help inform the
trigger decision by counting the number of electrons present within the detector. It
is made of layers of vertically segmented bars of plastic scintillator. These layers are
arranged in pairs where the layers within each pair are offset from one another to cover
any gaps between the bars. These bars are readout in time to be used within a trigger
decision combined with information from the ECal and HCal.
The HCal (green in Figure 3.2) is a sampling calorimeter made up of alternating
layers of steel absorber and plastic scintillator bars. The HCal is further subdivided
into the “side” HCal which is situated around the ECal and the “back” HCal down-
stream of the ECal. The back HCal has the orientation of the scintillator bars alternate
between vertical and horizontal so that clusters and tracks can have three-dimensional
coordinates more preceisely identified.
The ECal (blue in Figure 3.2), as a primary volume of interest within the analy-
sis discussed here, is given its own diagram Figure 3.4. The ECal is also a sampling
calorimeter; however, it uses a different absorber material and a different sensing mech-
anism to more precisely measure the energy of electrons, positrons, and photons. The
calorimeter is constructed out of seventeen layers each consisting of tungsten absorber,
service materials, and sensitive silicon sensors. Each of the layers of the detector has
two sub-layers of sensitive silicon sensors and each of these sub-layers are built up out of
the hexagonal High-Density modules designed for the CMS Phase II High Granularity
28
Figure 3.2: Diagram of LDMX detector apparatus with a representation of a signal event
where a dark brem occurs within the target. Credit to Christian Herwig for original
development of diagram.
Calorimeter upgrade[29]. These hexagons are arranged in a “flower” providing excel-
lent transverse resolution of shower location and shower shape. The layers are built
and arranged in order to space the sensitive silicon sub-layers to give good longitudinal
resolution of showers as well. In total, the designed ECal has more than one hundred
thousand channels that can each individually detect particles depositing energies from
≈ 0.1MeV up to ≈ 1GeV. This high granularity calorimeter gives LDMX excellent dis-
crimination power since it can measure the amplitude and location of several incident
particles. Due to its high performance, the ECal is a primary tool for both designing
a trigger decision (with the aid of the Trigger Scintillator’s electron count) as well as
downstream analysis separating SM background processes from potential DM signal.
29
Figure 3.3: Rendering of LDMX detector apparatus focusing on tracker, target, and
ECal. The magnet would fully encompass the tracker, target and trigger scintillator.
30
Tu
ng
ste
n
Ab
so
rb
er
Se
rv
ice
s P
CB
Co
oli
ng
P
lat
e
Se
rv
ice
s P
CB
17 ECAL Layers
Pre-shower layer
Each Layer
7 Silicon Sensors
Tu
ng
ste
n
Ab
so
rb
er
Tu
ng
ste
n
Ab
so
rb
er
Si
lic
on
S
en
so
r
Si
lic
on
S
en
so
r
432 Readout Channels
Figure 3.4: Diagram of LDMX ECal design showing the longitudinal segmentation (top
and bottom right) and the transverse segmentation (bottom left). Credit to Joe Muse.
Chapter 4
Mid-Shower Simulation
LDMX (like many other HEP experiments) uses an intricate software stack in order to
realistically and efficiently simulate particle interactions with the detector, emulate the
electronics that would be used to measure these interactions, and reconstruct the output
of these electronics into physically-understandable variables. These software tasks are
accomplished by a wide swath of different software packages, some of which custom-
written for LDMX, most of which written in C++. This chapter is focused on describing
this simulation infrastructure – focusing particularly on parts of the infrastructure I was
involved in – while also pointing out areas that are expected to remain constant in the
presence of data gathered from a real detector. After a discussion of general data
processing, I move into discussing the specific samples used to do this simulation study.
4.1 General Data Process
One of the core principles helping organize HEP data is the concept of an “event.” In
the context of LDMX, an event is the data collected within a small window of time
around the arrival of a beam electron into our detector apparatus. In essence, each
event within the software is independent from one another; however, they all share a
similar structure to the information they hold. A natural example is a data table: a
row has the same variables in each of its columns as all the other rows. Events behave
the same way; however, unlike a data table, the structure of an event can be more
intricate than simply a series of values corresponding to different column titles. While
31
32
...
Input Data File
...
...
Output Data File
Processor
Event Bus
start event
end event
get
add
Figure 4.1: Flow chart of how data is processed in the LDMX framework. Each processor
has the ability to “add” data to the event as well as “get” data from the event. The
processors are run in a user-defined sequence. Data can also be loaded from one or more
input data files where the event data from those files is loaded into memory before the
first processor is run. After all processors are done with an event, it is saved to the
output data file.
our basic “unit” of data is an event, we require many events in order to make statistical
conclusions about our data; thus, we have developed an “event processing framework”
that allows us to unify the various aspects of processing the data held within an event.
The event processing framework designed, developed, and maintained by LDMX is
designed to allow the flexibility necessary to do the wide range of tasks necessary for
the experiment. The C++ framework uses CERN’s ROOT [30] for data serialization,
Boost for logging, and Python [31] for dynamic run-time configuration. As diagramed in
Figure 4.1, the design is a sequential model: the “event bus” stops at the individual pro-
cessors in a certain sequence. The individual processors can inspect the data currently
on the bus and board more passengers onto the bus for later processors to use and which
can be eventually serialized into the output ROOT file. The processors can be built
separately from the framework and then dynamically loaded and created at run-time by
33
an abstract factory. This design choice allows for all of the computationally-intensive
software tasks necessary for LDMX (simulation, reconstruction and some analysis tasks)
to use this framework and be organized into separate modules which are only loaded
into memory when that module is being used.
The serialization portion of the framework is similarly dynamic; focusing on enabling
users’ code to add data structures from the most simple (e.g. individual booleans) to
the most complicated (e.g. containers of custom classes). This wide array of data types
is supported by ROOT’s dictionary system during the serialization stage and abstract
wrapper classes with partially-specialized template derivatives during run-time. Such
complexity within the framework is necessary in order to allow a simple interface – one
where the user interacts with simple and complex types in the same way.
Combining this highly dynamic serialization library with the sequential-processing
model configured at run-time gives a strong foundation for all of the software needs
of LDMX. Written in C++, this software framework enables high performance for all
of the major data processing tasks necessary for the experiment. Moreover, its design
focuses on flexibility and modularity so that seemingly-disparate data processing tasks
can be unified under one framework. Everything from simulation to detector emulation
to event reconstruction to analysis calculations can be done within this framework,
reading and writing files from this framework and enabling our software to be well
organized while also centralized in one location.
4.1.1 Data Processing Stages
The centralized nature of the LDMX processing framework makes it much simpler to
stay unified as a collaboration. Experts are able to work on their specialized area of the
software and share those improvements with the entire collaboration with ease. Since
the flexibility of the framework allows for arbitrary groupings of these different data
processing stages, we can choose to separate them into natural groups that correspond
to the different areas on which experts focus (diagrammed in Figure 4.2).
In this work (as the chapter title implies), we focus mainly on the detector simula-
tion stage where events are produced using random sampling of relevant physical phase
spaces in a way geared towards realistically modeling the detector and particles inter-
acting with it. LDMX like most HEP experiments use Geant4[32] to help perform this
34
SIM DIGI REC Analysis
Variables
RAW
Real Detector
Framework
Data
Software
Library
Custom
Subsystem
Tooling
Geant4
Figure 4.2: Diagram of processing stages within LDMX showing the external sources
of data or software that are used in those stages as well as which stages are commonly
done within the centralized processing framework.
complicated detector simulation. The downstream stages, namely electronic emulation
and reconstruction, are not described further here; however, it should be emphasized
that all subsystems of LDMX have their own custom implementations of these stages
in order to realistically emulate and reconstruct the data within their subsystem.
4.1.2 Biasing and Filtering Technique
Many of the processes most interesting for a DM search within LDMX are rare relative
to processes that are “easier” to reject by normal data processing. In many cases, the
rarity of these more interesting processes is a computational roadblock since waiting for
the entire detector simulation (hundreds to thousands of particles each with dozens if
not hundreds of interactions) to complete before looking for these processes of interest
wastes a lot of computer time. Typically, only one out of every ten thousand events
actually contains a process we are interested in. To make simulation of these processes
more computationally feasible, we turn to the common simulation techniques of bias-
ing (artificially increasing the probability of a certain process occurring) and filtering
(proactively ending the simulation of events if certain criteria are not met).
35
In the analysis channel studied here, all of our interesting processes occur within
the developing shower of particles in the ECal. While more complicated than checking
for processes happening to the single beam electron, we can select specific processes
occurring within this shower in a computationally efficient way by tuning the order
with which the simulation processes particles. The detector simulation is focused only on
particle-material interactions and the materials are assumed to not change significantly
during these interactions, so the order in which particles are simulated does not change
how the physics is modeled. We choose to process particles in two groups – “high” and
“low” energy – where the high-energy particles are all processed before the low-energy
particles. The border between these two (called the “Sorting Threshold”) is configurable
by the user and enables a specific Geant4 function to be called mid-shower after all
of the high-energy particles are processed (i.e. simulated until their total energy falls
below the Sorting Threshold).
Suppose we know from studying other (unbiased or low-bias) samples that showers
need to have a minimum amount of energy entering a specific process of interest before
the shower becomes “important.” We can set this minimum energy (here called the
“Filtering Threshold”) as a requirement on the ECal shower for the simulated event
to be kept within the sample and, in order to improve the speed of the filtering, we
can apply this requirment during the simulation when the event transitions from the
high-energy group to the low-energy group of particles.
Now, this infrastructure of determining a minimum simulated energy and having an
energy-based sorting of the processing order already helps improve the computational
efficiency of these simluation samples (from ∼ 6k CPU-hours to ∼ 30 CPU-hours for
the “important” events from ∼ 1B total events). Nevertheless, it is still extremely
inefficient (most simulated events do not pass our filter requirements) and so we add
biasing in order to artificially increase the rate of the process we wish the sample to
focus on. Since the processes we are interested in biasing are connected to particles that
are abundant within the electromagnetic ECal shower, we need to require only particles
above a certain energy threshold (the “Biasing Threshold”) to avoid over-sampling of
the process. In addition, we also need to define the factor that we will multiply the cross
section by when a particle is above the Biasing Threshold and in the ECal volumes (the
“Biasing Factor”).
36
Artificially increasing the rate of a process relative to its true rate measured in
nature is obviously unphysical, but we can account for this by weighting the events
depending on this Biasing Factor when counting them. The simulation engine we use
does this weight calculation for us and it accounts for both the biasing factor (beginning
the weight with a factor of 1/B) and unbiased processes happening to a particle with
biasing enabled (increasing the event weight to reflect that physical processes at the
natural rate have ocurred).
Parameter Dimension Symbol Short Description
Sorting Threshold Energy TS Minimum total energy to be pro-
cessed in first group
Biasing Threshold Energy TB Minimum total energy for a parti-
cle to be biased
Biasing Factor Dimension-less B Factor to multiply cross section by
Filtering Threshold Energy TF Minimum energy transfered into a
process for the event to be kept
Table 4.1: Parameters of mid-shower biased samples. In all of the samples used in this
study, we use the same value for the Sorting and Biasing Thresholds.
Table 4.1 summarizes the parameters related to this sample generation technique.
Practically, we choose to have the Biasing and Sorting Thresholds share the same value
for all of the samples in this work in order to reduce the number of parameters needed
to tune and to process all of the biased particles (as well as particles of similar energies
but different flavors) before making any filtering decision.
Generally, validation of this biasing and filtering technique can be done by checking
that the samples generated with this technique match samples generated without this
technique and with the filtering applied after the fact. This validation needs to be done
on a per-sample basis since the specific aspects of the simulation that need to align may
change depending on the specific physics for which we are biasing and filtering.
4.2 Standard Processes
As displayed in Figure 3.1, there are a few categories of physics processes that are
described by the standard model and we expect to happen within our experiment.
These “background” processes – while interesting for other analyses of LDMX data –
37
should be rejected by this search for DM since they are understood to not be DM by the
standard model and other experiments. In order to thoroughly and faithfully estimate
our ability to reject these backgrounds, we need to realistically simulate them within
our detector volume.
Geant4 [32] has been developed for precisely this purpose. This study uses a slightly
modified copy of Geant4 v10.2.3 which focuses on improving the accuracy of processes
relevant to a DM search within LDMX with the following needs.
1. Updating cross section and sampling of γ → µ+µ− process to align it better with
collected data and calculations of the standard model.
2. Update nuclear cascade to align with more recent CLAS data specifically focused
on the rate of low-multiplicity forward neutrons (Appendix A.A of [25]).
3. Introduction of back scattering π0 in a γp → π0p within a nuclear cascade (Ap-
pendix A.B of [25]).
These updates enabled us to more reliably study the known and expected background
processes interactions with the LDMX detector design and compare this data to simu-
lations of a model of a DM production process.
4.2.1 Dominant Contributors in this Analysis
First and foremost, the analysis detailed in Chapter 5 is focused on utilizing data col-
lected with LDMX that is not used in the nominal Missing Momentum (MM) analysis.
This design goal produces two requirements on the samples we will analyze.
• Same Trigger Requirement – while designing a separate trigger for different analy-
ses is feasible, this specific analysis is focused on working with the same data that
is collected for the primary MM analysis. The trigger requirement for the primary
MM analysis is one of missing energy within the ECal making it an appropriate
trigger for this analysis as well.
• Opposite Tracking Requirement – the primary MM analysis enforces the assump-
tion that the processes of interest occur within the thin target; thus, the electron
track leaving the target (colloquially named the “recoil”) is required to have less
38
than 30% of the beam energy. In this case, we flip this requirement and instead
require the electron to arrive at the ECal with more than 87.5% of the beam
energy. While the application of this requirement would be done after the trigger
requirement in real data, for simulation we apply this requirement first so that
the simulation samples we are studying can be more efficiently produced.
All of the simulation samples studied here require the primary beam electron to arrive
at the ECal with more than 87.5% of its original energy. Once this requirement is
imposed, we then turn our attention towards the processes that occur within the ECal.
Generally, we focus on processes that would cause the energy in the ECal to be
estimated at a significantly smaller value than the known beam energy. This “missing”
energy is used as a key piece of evidence that the event in question should be stored for
further study and thus is the main ingredient in the infrastructure deciding whether a
specific event should be stored (the “trigger”). From prior experience with LDMX (e.g.
[25, 2]), we expect certain types of standard processes to dominate the backgrounds
that pass this trigger requirement on missing energy: interactions with the nuclei of
the detector material producing hard-to-detect hadrons (so-called “nuclear” processes)
and conversion of photons into a pair of muons (which are also difficult for the ECal to
contain).
Nevertheless, we still can use a large, unbiased simulation sample in order to ver-
ify these previous conclusions and inform our decisions on how the subsequent biased
samples should be produced. Figure 4.3 is a key piece of evidence along these lines
where we separate the total reconstructed energy (the estimate from our detector) as
a fraction of the known beam energy by the amount of energy going into these nuclear
processes (so-called “nuclear” energy). We see that as more energy is given to these
nuclear interactions, the reconstructed fraction decreases. Eventually, we reach a point
where enough energy has gone to these nuclear interactions that we expect the event to
be below the trigger threshold and be kept by the trigger system at a significant rate.
Figure 4.3 shows that a significant fraction of events whose total nuclear energy is
greater than 62.5% of the beam energy fall below the threshold for the trigger. This
figure also shows that there are other types of events that do not have much (if any)
nuclear interactions (since their nuclear energy is less than 10% of the beam energy)
but still are reconstructed below the trigger threshold. Further investigation reveals
39
0.0 0.2 0.4 0.6 0.8 1.0
EECal/EBeam
100
101
102
103
104
105
106
107
Ev
en
ts
LDMX Preliminary
Below Trigger
Threshold
109 EoT (4GeV)
Nuclear Energy Fraction
< 10%
Between 10% and 62.5%
> 62.5%
All
0.0 0.2 0.4 0.6 0.8 1.0
EECal/EBeam
100
101
102
103
104
105
106
107Ev
en
ts
LDMX Preliminary
Below Trigger
Threshold
109 EoT (8GeV)
Nuclear Energy Fraction
< 10%
Between 10% and 62.5%
> 62.5%
All
Figure 4.3: The total reconstructed energy within the ECal as a fraction of the beam
energy separated by the total nuclear energy within the event (the energy going into
the photon-nuclear and electron-nuclear processes).
that these events are indeed “di-muon” events (Figure 4.4). These two categories of
processes – nuclear and di-muon – are the primary background processes since they
mimic our missing energy trigger selection at a significant rate.
4.2.2 Expanding Production of these Processes
As mentioned in Section 4.1.2, the unfiltered and unbiased simulation is not very effi-
cient. In some sense, this shows that our basic trigger selection is doing a good job of
rejecting the background processes that our analysis finds uninteresting; nevertheless,
we need to efficiently sample events that contain interesting background processes at a
large scale so we can analyze them in detail. This is where the biasing and filtering tech-
nique described above is employed – our main goal is to improve the speed with which
we can generate events originating from these background processes while maintaining
faithfulness to the unbiased and unfiltered distributions.
The process of tuning these simulation variables is largely a guess-and-check one.
We test a variety of different settings and see how they perform both in terms of their
computational speed as well as their faithfulness to the underlying distributions. For
40
0.0 0.2 0.4 0.6 0.8 1.0
Muon-Caused Fraction of HCal Hits
100
101
102
Ev
en
ts
/
1B
E
oT
Unbiased Events
Passing Trigger with
Erec < 0.5Ebeam and
Enuc < 0.25Ebeam
LDMX Internal
Beam
4GeV
8GeV
Figure 4.4: Fraction of hits within the HCal that were caused by muons or anti-muons.
The events lying within the zero bin still have HCal hits caused by other types of
particles.
our purposes, the filtering threshold TF can be estimated from the large unbiased and
unfiltered sample since that sample contains events with these processes of interest.
Specifically, Figure 4.3 displays a value of TF = 0.625EBeam for the nuclear processes.
More detailed study of the events residing in the low-side tail of the blue distribution
showed that TF = 0.5EBeam was a good choice for the di-muon process; however, since
there were fewer events containing the di-muon process within this sample, other choices
of TF were surveyed in order to confirm this value.
The other threshold TB (which we also set to be TS in this work) is tightly coupled
with the biasing factor B and so we test them both together. A rather simple proxy for
checking that the events sampled during biasing and filtering matches the underlying
distribution is to inspect the longitudinal distribution of where the process ocurred.
While the biasing and filtering should improve the speed of the simulation, it should
not observably distort this distribution which roughly follows the development of the
shower (an example of this is shown in Figure 4.5).
We can confirm that our biased and filtered simulations are properly extending the
tail of the distributions in which we are interested. Figure 4.6 shows a comparison
41
200 300 400 500 600
Starting Z of + / mm
10 5
10 4
10 3
10 2
10 1
Ev
en
t F
ra
ct
io
n
LDMX Internal 4 GeV
B = 1e4 R = 4.25MHz
B = 5e4 R = 4.81MHz
B = 1e5 R = 5.49MHz
B = 5e5 R = 6.30MHz
B = 1e6 R = 6.86MHz
B = 5e6 R = 10.93MHz
200 300 400 500 600
Starting Z of + / mm
10 4
10 3
10 2
10 1
Ev
en
t F
ra
ct
io
n
LDMX Internal 8GeV
B = 1e2 R = 0.09MHz
B = 5e2 R = 0.36MHz
B = 1e3 R = 0.58MHz
B = 5e3 R = 1.23MHz
B = 1e4 R = 1.45MHz
B = 5e4 R = 1.94MHz
B = 1e5 R = 2.40MHz
B = 5e5 R = 5.90MHz
Figure 4.5: Longitudinal location (z) where the µ+ was produced within the simulation
for a variety of biasing factors B for the 4 GeV beam (left) and 8 GeV beam (right). The
rate R is the equivalent number of unbiased EoT divided by the CPU time necessary to
produce the sample. We can observe that while increasing B also improves the rate R,
we eventually over-bias and lose access to the late tail of the distribution (red, puple,
and brown in the left plot and gray in the right plot).
between a large unbiased sample that has been scaled to the same equivalent EoT as
the biased and filtered samples – there we see that the biased samples properly and
smoothly extend the low reconstructed energy fraction tail which corresponds to the
signal region in this missing energy search.
4.3 Dark Matter Signal
The particular signal process this analysis channel is looking for is the production of a
dark photon followed by an invisible decay. In this regime, what happens to the dark
photon after it is produced is irrelevant to the analysis since both it and its products
are not observable by our detector.
With this focus in mind, we developed a dark bremsstrahlung simulation method
that allows for the visible particle (the recoiling lepton) to be distributed according to
a full matrix element calculation (via MG/ME) while the incident particle can have
42
0.0 0.2 0.4 0.6 0.8 1.0
EECAL/EBeam
101
103
105
107
109
1011
Ev
en
ts
in
1
013
E
oT
4GeVLDMX Internal
Scaled Unbiased
Biased Nuclear
Biased Di-Muon
Biased Total
0.0 0.2 0.4 0.6 0.8 1.0
EECAL/EBeam
101
103
105
107
109
1011
Ev
en
ts
in
1
013
E
oT
8GeVLDMX Internal
Scaled Unbiased
Biased Nuclear
Biased Di-Muon
Biased Total
Figure 4.6: ECal reconstructed energy EECAL as a fraction of the beam energy EBeam
comparing the unbiased and biased samples scaled to the same EoT. The 4 GeV (8
GeV) beam case is shown on the left (right).
variable energy and be handled by Geant4 directly. This novel simulation technique
allows for the dark bremsstrahlung process to be treated (from Geant4’s perspective)
on the same footing as the background processes while maintaining the precision of a
matrix-calculator method.
4.3.1 G4DarkBreM
To accurately simulate the kinematics of the dark bremsstrahlung process for electrons
in thick targets, the process must be included at the level of experimental simulation
instead of using initial state event generators to account for the possibility of energy
loss through bremsstrahlung or multiple scattering prior to the dark matter interac-
tion. Accurate kinematic simulation of the outgoing electron are required for optimal
experimental sensitivity measurements and appropriate design of search strategies.
For this reason, we utilize G4DarkBreM [33] which performs this embedding of
the dark bremmstrahlung process into Geant4. G4DarkBreM calculates the cross
section using numerical integrals of the Weizsäcker-Williams approximation, and the
kinematics are simulated using a scaling technique of MadGraph/MadEvent event
libraries. The accuracy of the total cross section and kinematics is validated using
43
MadGraph/MadEvent samples at a range of incident lepton energies.
4.3.2 MadGraph/MadEvent
For this study in LDMX, the code package used to develop and validate G4DarkBreM
was also used to generate the input refernece libraries for sample generation. This code
is a custom version of MadGraph/MadEvent4 with the following updates.
• Introduction of basic dark sector particles (massive boson and spin-1/2 fermion)
which act as the representatives of the dark sector interacting with the standard
model particles.
• Definition of a nuclear particle (electrically neutral, spin-1/2 fermion) with new
couplings to the dark photon including the nuclear form factors.
• Updating the definition of the electron to include its small (but non-zero) mass to
prevent divergence of the cross section at lower energies.
These updates, along with some wrapping code, enable the generation of dark bremm-
strahlung events for a range of target nuclei, incident energies, and either incident
electrons or muons.
As suggested by the validation of G4DarkBreM [33], we generate libraries where
the incident electron’s energy drops in steps of 10% down from the beam energy. Since
the dark photon is required to be simulated with at least 50% of the beam energy, we
stop the library generation also at 50% of the beam energy. The primary nuclei that
could dark brem within the ECal are tungsten, silicon, oxygen, and copper so those are
the nuclei for which we generate libraries. The rest of the nuclei within the simulated
ECal (sodium, calcium, carbon for example) have atomic numbers close to those nuclei
already sampled and so can be faithfully simulated with these libraries. In order to avoid
duplicating the same event, a unique reference library is generated for each simulation
run even when keeping the dark photon mass and beam energy the same.
4.3.3 Characterization
These samples, as expected, show that the dark bremsstrahlung process can occur fol-
lowing normal shower development peaking at ∼ 1X0 into the ECal (with the exception
44
0 1000 2000 3000 4000
Total Energy of A' / MeV
10 3
10 2
W
ei
gh
te
d
Ev
en
t F
ra
ct
io
n
4GeVLDMX Internal
mA / MeV
1
10
100
1000
0 1 5 10 15 20 25
Depth of DB in ECal / X0
0 100 200 300 400
Z Location of DB / mm
10 4
10 3
10 2
10 1
100
W
ei
gh
te
d
Ev
en
t F
ra
ct
io
n
LDMX Internal mA / MeV
1
10
100
1000
Figure 4.7: Distributions of simulated signal events with a 4GeV beam. The energy of
the produced dark photon is shown on the left while the longitudinal location where the
production occurred is on the right.
of higher mass dark photon samples which are kinematically prevented from being gen-
erated further into the shower). Figure 4.7 shows some example distributions from these
samples. We can observe the normal shower development behavior in the location of
the dark bremsstrahlung and we can see the simulation-level cut on the dark photon
energy at half the beam energy (in this case 2GeV) which is imposed to improve com-
putational efficiency without losing events that would otherwise pass the downstream
trigger requirement.
4.4 Summary
While this chapter has been a long and winding road through the simulation infras-
tructure of LDMX, the key feature that I would like to emphasize is that it is able
to efficiently and reliably simulate both background and signal processes within this
search for DM. These samples, after being thoroughly studied and validated, were en-
larged to provide a sufficiently sized sample for the downstream analysis of Chapter 5.
45
Sample Biasing Factor Biasing Threshold
Signal m
max(log10(mA),2))
A /ϵ2 0.5EBeam
Enriched Nuclear 200 0.375EBeam
Di-Muon 105 0.5EBeam
Sample Filtering Cuts
Unbiased EECal Front
primary ≥ 0.875EBeam
Signal EECal Front
primary ≥ 0.875EBeam & EA′ ≥ 0.5EBeam
Enriched Nuclear EECal Front
primary ≥ 0.875EBeam & Etot nuc ≥ 0.625EBeam
Di-Muon EECal Front
primary ≥ 0.875EBeam & Etot µ ≥ 0.5EBeam
Table 4.2: Configuration of the simulation samples used in this analysis. Ebeam is the
beam energy being studied (4 or 8 GeV in this work). mA is the mass of the A’ in MeV,
ϵ is the dark brem mixing strength, EECal Front
primary is the energy of the primary electron at
the front of the ECal, EA′ is the energy of the generated A’, Etot nuc is the total energy
transferred to nuclear interactions during the event, and Etot µ is the total energy of
produced muons.
Specifically, the background samples were generated to be statistically equivalent to
1 × 1013 EoT and the signal samples were generated to provide ∼ 1M events for each
dark photon mass point to study selection efficiencies. Table 4.2 gives a summary of
the configuration used to generate these samples.
Chapter 5
Missing Energy Search
One of the primary strengths of the LDMX detector design is its ability to use the
tagging and recoil tracker system to reject a large number of background events by
separating a nominal beam with non-standard energy loss from a nominal beam with
standard energy loss or even low-energy beam. Moreover, the tagging and recoil tracker
system gives LDMX the potential to further suppress backgrounds or potentially study
DM properties by studying the transverse momenta of electrons recoiling from dark-
bremsstrahlung candidate events.
While this design is optimal for a large number of Electrons on Target (EoT), the
strategy has some limitations for a low EoT data run. To limit multiple scattering
which ruins momentum measurements, the baseline detector configuration requires a
thin (≈ 0.1X0) target. In the early stages of LDMX, when the total EoT will be lower,
a different analysis strategy and detector configuration may be optimal to probe the
largest amount of the y–mχ phase space. The alternative strategy ignores the dedicated
target inside the tracker volume and instead uses the ECal as an active target. The
EaT analysis channel for LDMX thus has two primary purposes which can be separated
by the timeline over which they are relevant.
1. Short Term: In early running, when the number of EoT is relatively small,
the nominal MM analysis will not have obtained significant reach into new DM
phase space (yet). The EaT channel serves here as a way to obtain world-leading
sensitivity early in the lifetime of LDMX and give the collaboration a first look at
46
47
the data the apparatus has collected.
2. Long Term: As LDMX collects data, the MM analysis enters into unexplored
phase space and serves as a better discovery mechanism due to its access to the
Tagger and Recoil trackers. The EaT channel, while struggling to suppress com-
plicated backgrounds with relatively limited analysis handles, can operate “or-
thogonally” in the collected data since its primary selection (an approximately
beam-energy electron passing through the Recoil tracker) is inverted relative to
the MM analysis (the electron passing through the Recoil tracker has significantly
less energy than the beam).
An initial study of the EaT analysis channel is the primary focus of Part II of this
thesis, focusing primarily on the first (short term) purpose. In this regard, we target
an EoT that is reasonable to accomplish early in the running of LDMX and avoids
particularly intricate backgrounds. A total of 1 × 1013 EoT fits these requirements by
avoiding the charged current production of neutrinos and represents ∼ 2.5% of the first
full LDMX dataset which is obtainable within approximately a two weeks of nominal
beam time1. Since EaT is expected to be the first physics analysis on LDMX data,
we want a simple and robust analysis that can withstand the test of time and the
complexities of real data.
With these design goals in mind, a simple “cut-and-count” analysis has been devel-
oped. The simplicity of this analysis is one of its strengths, enabling it to be applicable
despite potential surprises arising from first encounters with real data. The bulk of time
and effort on this first investigation was focused on making this investigation possible via
the introduction of midshower process filtering and a dark bremsstrahlung simulation
process described in Chapter 4.
5.1 Selection
The core goal of most search analyses is to develop a selection that avoids events known
to be standard processes (i.e. backgrounds) while keeping events containing the process
1Assuming the LDMX detector apparatus and beam delivery is operating according to specifications,
we expect the beam to be delivered on a frequency of 37.5MHz with a duty cycle of ≈0.5 and the
number of electrons within each bunch to be Poisson distributed with µ = 1. 37.5MHz× 0.5× P (µ =
1, 1)× 2week ≈ 1 × 1013 EoT.
48
0 200 400 600 800 1000 1200 1400
EECal / MeV
10 7
10 6
10 5
10 4
10 3
Ev
en
ts
(I
nt
eg
ra
l N
or
m
al
iz
ed
to
1
) 4GeVLDMX Preliminary
Background
mA = 1MeV
mA = 10MeV
mA = 100MeV
mA = 1000MeV
0 500 1000 1500 2000 2500 3000
EECal / MeV
10 8
10 7
10 6
10 5
10 4
10 3
Ev
en
ts
(I
nt
eg
ra
l N
or
m
al
iz
ed
to
1
) 8GeVLDMX Preliminary
Background
mA = 1MeV
mA = 10MeV
mA = 100MeV
mA = 1000MeV
Figure 5.1: The total reconstructed energy in all layers of the ECal (EECal). The signal
and background distributions are normalized such that their integral is one. The 4GeV
beam is shown on the left and the 8GeV beam is shown on the right. The events falling
into bins with EECal > 1.5GeV (3.16GeV) for 4GeV (8GeV) are omitted from this plot
but included in efficiency calculations.
being searched for (in this case, the production of DM via dark-bremsstrahlung). Both
the EaT analysis channel and the primary MM channel share the dark-bremsstrahlung
signature of missing energy within the ECal relative to the known incident beam energy.
This allows for the first selection made to be shared between these two channels – a
requirement that the sum of the observed energy in the first twenty layers of the ECal is
less than 1.5GeV (3.16GeV) for a 4GeV (8GeV) beam. This preliminary selection acts
as the primary trigger for both of these analyses – selecting events that resemble dark-
bremsstrahlung via their lower-than-average observed energy during data collection –
however, the EaT analysis channel in the early-running scenario may not be collecting
data with a fully calibrated energy scale at the trigger level. The potential for mis-
calibration motivates tightening the analysis missing energy requirement by 400MeV as
well as requiring the sum to be performed over all thirty-four layers of the ECal instead
of only the first twenty. Figure 5.1 shows the unit-normalized distributions of this total
reconstructed energy within the ECal after the trigger has been applied.
After this total energy requirement, there are still backgrounds left and those that
49
remain are largely events where a significant fraction of the energy is carried by particles
difficult to contain (or even observe) within the ECal (neutrons, KL, or muons for
example). To suppress these types of persistent backgrounds, the event is required to
have less than 10 PE deposited in any single bar of the HCal – max(PEHCal) < 10 –
well below the typical signal of a through-going muon of 80 PE in any single bar.
While this selection does a good job of removing events containing long-living parti-
cles, there are some events containing several low-energy particles that range out within
the ECal itself. These particles (often due to their number) distribute energy over a
range of cells, so we can suppress these events by requiring the spatial distribution of
energy within the ECal to be small. We choose to measure the spatial distribution using
the energy-weighted root-mean-squared spread of the hits within the ECal2 – the ECal
Hit RMS – which is typically larger for these background processes than for a single,
truncated electromagnetic shower present within a signal event. We therefore require
the ECal Hit RMS to be less than 20mm.
These two selection variables are shown in Figure 5.2 where the distributions are
shown after all other selections are applied (including the trigger selection and the
tighter selection on the total ECal energy including all layers). Table 5.1 shows the
tabulation of this selection for the studied simulation samples. Figure 5.4 shows the
signal and background total ECal energy distributions after these selections are applied
where we can still observe difference in the shapes of these distributions which we will
exploit in the final statistical analysis of these samples.
An additional, sneaky selection has also been applied due to how the simulation
samples were constructed. As mentioned in Section 4.2, the simulation samples have
a filter requiring the primary electron to arrive at the ECal with at least 87.5% of the
beam energy. This simulation filter is attempting to replicate a requirement by the
tracking system that would be applied after the trigger requirement when analyzing
real data. Studies of other simulation samples without this tracker requirement showed
that while the tracker is required to help confirm a near-full-energy electron arrives
at the ECal, it keeps over 90% of the events that are interesting for this analysis (i.e.
nothing interesting happens in the thin target). Figure 5.3 shows example figures from
2A “hit” is defined as a single cell within the ECal registering a signal corresponding to 50% (or
more) of the energy deposited by a typical muon during the event.
50
0 10 20 30 40 >50
max(PEHCal)
10 2
100
102
104
106
Ev
en
ts
4GeVLDMX Preliminary
N 1 D = 0.5 mA = 3m
Background
mA = 1MeV; y = 1e-13
mA = 10MeV; y = 1e-11
mA = 100MeV; y = 1e-09
mA = 1000MeV; y = 1e-04
0 10 20 30 40 >50
ECal Hit RMS / mm
10 2
10 1
100
101
102
103
104Ev
en
ts
4GeVLDMX Preliminary
N 1 D = 0.5 mA = 3m
Background
mA = 1MeV; y = 1e-13
mA = 10MeV; y = 1e-11
mA = 100MeV; y = 1e-09
mA = 1000MeV; y = 1e-04
0 10 20 30 40 >50
max(PEHCal)
10 2
10 1
100
101
102
103
104
105
106
Ev
en
ts
8GeVLDMX Preliminary
N 1 D = 0.5 mA = 3m
Background
mA = 1MeV; y = 1e-13
mA = 10MeV; y = 1e-11
mA = 100MeV; y = 1e-09
mA = 1000MeV; y = 1e-04
0 10 20 30 40 >50
ECal Hit RMS / mm
10 1
100
101
102
103
Ev
en
ts
8GeVLDMX Preliminary
N 1 D = 0.5 mA = 3m
Background
mA = 1MeV; y = 1e-13
mA = 10MeV; y = 1e-11
mA = 100MeV; y = 1e-09
mA = 1000MeV; y = 1e-04
Figure 5.2: Variables used in signal selection for the EaT analysis channel. In each
figure, all other cuts are applied except for the variable in question. Some of the bins
are empty in which case the upper Poisson limit given the total sample size is drawn
with error bars. The grey line shows the selection cut. The background sample is the
enriched nuclear and dimuon samples. The top (bottom) row shows the 4GeV (8GeV)
beam.
51
Analysis Stage for 4GeV Beam
Background Signal Efficiency (%)
Event Yield 1 MeV 10 MeV 100 MeV 1 GeV
ECal Trigger (E20 < 1.5 GeV) 4.60 × 107 58 67 71 83
ECal Energy (EECal < 1.1 GeV) 1.95 × 106 35 48 53 72
max(PEHCal) < 10 1.15 × 103 34 47 52 69
ECal Hit RMS < 20 mm 126 28 37 41 33
Analysis Stage for 8GeV Beam
Background Signal Efficiency (%)
Event Yield 1 MeV 10 MeV 100 MeV 1 GeV
ECal Trigger (E20 < 3.16 GeV) 6.10 × 107 66 74 79 89
ECal Energy (EECal < 2.76 GeV) 6.88 × 106 52 63 69 84
max(PEHCal) < 10 31.8 50 61 67 81
ECal Hit RMS < 20 mm 7 43 52 56 52
Table 5.1: Cut-flow analysis comparing background and various signal hypotheses for
the simple cuts used in this analysis. The event yield for the background sample is
calculated using the event weights and represent the number of events out of 1013 EoT
equivalent. The signal efficiency is relative to the full simulation sample. The efficiency
and event yield values on a given row are for after the analysis stage of that row. The
first table is the cutflow for the 4GeV beam, and the second table is for the 8GeV beam.
these studies that confirm the necessity of the tracker while also displaying that it is
highly efficient. Due to the tracker requirement’s high efficiency and the large separation
between events that are acceptable into this analysis and events that are not, we neglect
any systematic error (more on this in Section 5.3) related to this selection and instead
plan to set the tracker requirement’s value to whatever threshold is necessary to keep
the 90% efficiency when studying real data.
5.2 Background Prediction
When performing a search for new processes, a common technique is to develop a well-
motivated prediction for the background processes and then look for excesses above this
prediction. Before we go through rigourously defining what an “excess” means and how
to quantify what types of new processes we would exclude if no excess is observed, we
should more precisely define what our background prediction is.
We begin with fitting the cumulative background distribution with a simple expo-
nential function. Figure 5.5 shows this cumulative distribution, the resulting fit, and
the 95% confidence intervals surrounding the fit. This fit is a helpful prediction tool
52
50 100 150 200 250
z / mm
0
500
1000
1500
2000
2500
3000
3500x
/
m
LDMX Internal
ptarget
T = 0 for all
x Difference Between
4.0 GeV - 3.50 GeV
4.0 GeV - 1.50 GeV
8.0 GeV - 7.00 GeV
8.0 GeV - 3.16 GeV
(a) Difference in x position along the electron’s
path between different energy electrons. This
shows that the change in position is smaller
than a single ECal cell width, but it is well
within a conservaitive alignment of the track-
ing detector.
0.0 0.2 0.4 0.6 0.8 1.0
Primary e Energy Fraction at ECal
0.00
0.25
0.50
0.75
1.00
1.25
1.50
1.75
2.00
Tr
ig
ge
r S
um
F
ra
ct
io
n
107 EoT (8GeV)LDMX Internal
100
101
102
103
104
105
Ev
en
ts
(b) Comparison of trigger sum in the ECal
to the primary electron’s energy at the ECal.
Below the horizontal gray line are the events
that pass the trigger where there is large sepa-
ration between events where something inter-
esting happens in the target (left) and in the
ECal (right).
Figure 5.3: Studies of how the tracker requirement affect the simulation samples.
53
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
EECal/EBeam
10 2
10 1
100
101
102Ev
en
ts
1013 EoT (4GeV)LDMX Preliminary
D = 0.5 mA = 3m
Background
mA = 1MeV; y = 1e-13
mA = 10MeV; y = 1e-11
mA = 100MeV; y = 1e-09
mA = 1000MeV; y = 1e-04
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
EECal/EBeam
10 1
100
101
Ev
en
ts
1013 EoT (8GeV)LDMX Preliminary
D = 0.5 mA = 3m
Background
mA = 1MeV; y = 1e-13
mA = 10MeV; y = 1e-11
mA = 100MeV; y = 1e-09
mA = 1000MeV; y = 1e-04
Figure 5.4: The total reconstructed energy in all layers of the ECal (EECal) as a fraction
of the beam energy (EBeam) for all samples that pass the selection criteria except the
selection on ECal energy. The 4GeV beam is shown on the left and the 8GeV beam is
shown on the right. The gray lines mark the edges of the analysis bins used to estimate
the expected exclusion limit and the black line is the upper limit on the ECal energy
which also serves as the upper limit of an analysis bin.
54
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
Efrac = EECAL/EBeam
10 1
100
101
102
103
N
bk
gd
b
el
ow
E
fra
c
4GeVLDMX Preliminary
AeB(1 Efrac)
A = 1.82e+11 B = -28.9
Cumulative Bkgd MC
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
Efrac = EECAL/EBeam
10 1
100
101
N
bk
gd
b
el
ow
E
fra
c
8GeVLDMX Preliminary
AeB(1 Efrac)
A = 2.78e+05 B = -16.6
Cumulative Bkgd MC
Figure 5.5: Exponential fit of the total simulated background distribution. The shaded
region is a 95% confidence band on the fit which is interpreted as the uncertainty on
the value of the fit. Some bins in the simulated distribution are empty in which case
the upper Poisson limit given the sample size is drawn as an error bar in that bin.
in two keys ways. The fit produces a meaningful prediction across the entire range of
observable values which is a helpful way to “smooth” out this sample that is limited in
size after the selection criteria already applied. Moreover, the fit uses the 400MeV-wide
control region between the upper limit on the total ECal energy and the trigger thresh-
old applied during readout to help constrain its parameters. This constraint reduces
the uncertainty on the background prediction while also serving as an example of what
an analysis using real data could do – this same region could be utilized in order to help
set the scale of the background in a data-driven way.
The fit is then used to provide a background prediction within each of the three
analysis bins, shown in Figure 5.6 along with the unconstrained simulation prediction
for comparison.
55
0.00 0.05 0.10 0.15 0.20 0.25
Analysis Bin Efrac = EECAL/EBeam
100
101
102
N
bk
gd
/
bi
n
4GeVLDMX Internal
Bkgd MC
AeB(1 Efrac)
A = 1.82e+11 B = -28.9
0.00 0.05 0.10 0.15 0.20 0.25 0.30
Analysis Bin Efrac = EECAL/EBeam
10 1
100
101
N
bk
gd
/
bi
n
8GeVLDMX Internal
Bkgd MC
AeB(1 Efrac)
A = 2.78e+05 B = -16.6
Figure 5.6: Background prediction within the three final analysis bins (blue) compared
to the unconstrained simulation prediction in gray. The uncertainty on the background
prediction is taken from the 95% confidence band shown on the fit. Some bins in the
simulation prediction are empty in which case the upper Poisson limit given the sample
size is drawn as an error bar in that bin.
5.3 Systematics and Background Uncertainty
Besides the statistical uncertainty due to the finite size of the simulation sample (whose
estimation is represented by the 95% confidence intervals around the fit), there is ad-
ditional sources of uncertainty that originate from the experimental design (so-called
“systematic” uncertainty). Most prominently in this early-running analysis, we expect
mis-calibration of the ECal to be a potentially major source of systematic uncertainty
especially since the calibration of the ECal used when applying the trigger selection will
probably differ from the calibration used later on during the final analysis.3 Estimating
this systematic uncertainty is rather direct; we simply vary the calibration used at the
reconstruction and analysis stage within our data processing and see how that changes
the results.4
3This difference is expected since we can improve the calibration of the detector as we collect more
data; however, acknowledging this difference helps us put requirements on the accuracy of the intial
calibration used at the trigger level in order for this analysis to function as expected.
4The estimate for both the ECal and HCal systematics is done with the 4GeV beam sample since that
estimate is expected to be more conservative because the 4GeV beam sample allows more background
56
0 250 500 750 1000 1250 1500 1750 2000
EECAL / MeV
10 2
10 1
100
101
102
103
Ev
en
ts
in
6
×
10
9 E
oT
Signal Region
4GeVLDMX Internal
Fail Trigger (Nele = 1)
Smeared
Unsmeared
(a) Events failing the original trigger.
102
103
Ev
en
ts
in
6
×
10
9 E
oT
4GeVLDMX Internal
Pass Trigger (Nele = 1)
Smeared
Unsmeared
0 200 400 600 800 1000
EECAL / MeV
0.90
0.95
1.00
1.05
1.10
Sm
ea
re
d
/ U
ns
m
ea
re
d
(b) Events passing the original trigger and
falling within one of the three final analysis
bins over EECAL.
Figure 5.7: The total reconstructed energy in the ECal EECAL comparing ten different
smeared calibrations (colors) to the original unsmeared calibrations (black).
We generated ten different calibrations, smearing the calibration constants by 10%
in an uncorrelated fashion and by an additional, correlated 2% (5%) within the central
(outer) modules of the ECal. These smearing factors are a conservative estimate of the
accuracy of the initial calibrations used by the trigger to make data collection decisions.
These ten different calibrations were then used to re-reconstruct the simulated events
ten different times yielding Figure 5.7 and Figure 5.8. Figure 5.7a reassures us that
our signal region (below 1.1GeV) is not polluted with events that would have failed the
trigger selection even after these rather large calibration changes relative to the trigger
calibrations (i.e. we would not have overestimated our signal efficiency). In addition,
we can estimate the resulting systematic variation on the background event yield within
the three final analysis bins to be 5% using Figure 5.7b. The final selection on the ECal
Hit RMS is only affected negligibly as shown in Figure 5.8 where the distribution only
changes by ≲ 5% within the range of potential cuts.
Estimating the systematic uncertainty due to the HCal variable is not as straight
yield past its selections compared to the 8GeV beam sample.
57
10 2
10 1
100
101
102
Ev
en
ts
in
6
×
10
9 E
oT
4GeVLDMX Internal
Pass Trigger (Nele = 1)
Smeared
Unsmeared
0 5 10 15 20 25 30 35 40
ECal Hit RMS / mm
0.90
0.95
1.00
1.05
1.10
Sm
ea
re
d
/ U
ns
m
ea
re
d
Figure 5.8: Effect of smeared ECal calibrations on the ECal hit RMS for events that
pass the original trigger. Outside of statistical fluctuations in the lowest-populated bins,
the smearing has a less than 5% effect throughout the range of potential cuts.
forward since we are unsure on how best to directly alter the conditions with which the
HCal is reconstructed. We can still estimate the systematic uncertainty due to this cut
variable by testing a wide range of cut variables, deducing which ones are “acceptable”
and then extracting how the background yield varies between these different cut choices.
(What we want to be emulating is the underlying distribution shifting relative to our
analysis cut, but we are mimicing this by shifting the cut relative to the distribution.)
Figure 5.9a shows how the final reach (see Section 5.4) changes depending on this
cut. While not much movement is observed, we can see separation ocurring when the
threshold goves above ∼ 30 (more clearly seen in Figure 5.9b). With this in mind, we
estimate our “tolerance” to a systematic error on the maximum HCal PE to be 1 −
B<10/B<20 where B
0
N
vt
x
<
20.0
0.2
0.4
0.6
0.8
1.0
Ef
fic
ie
nc
y
10.7pb 1HPS Internal
20
16Event Pre-Selection
Data
Sim Bkgd
SIMP 30 MeV
SIMP 60 MeV
SIMP 90 MeV
Figure 8.3: Relative efficiency of pre-selection of events. Nvtx is the number of vertices
passing the vertex pre-selection requirements. Since other triggers for different purposes
are present in the data, the Pair 1 Trigger is applied to data but not to simulation.
Chapter 9
Displaced Vertex Search
A search for visibly-decaying DM is not novel, many experiments have been purposed-
built for such a search; however, HPS is specially focused on DM that decays within
shorter distances. While HPS has previously searched for simple dark sectors more
similar to the benchmark model used within the LDMX missing energy search [48], no
DM was found and the relatively longer decay lengths of this dark sector model harmed
the signal efficiency of the analysis and decreased the resulting sensitivity. Section 6.1
introduced the idea of SIMP DM which has additional interest for HPS. Specifically,
the increase in complexity of the DM model slightly decouples the production rate from
the decay rate meaning the number of events where DM is produced is not as tightly
connected to how displaced the particles are expected to be within the detector. The
downside of this decoupling is that some energy is lost to the production of the lighter
dark meson πD when the dark photon decays within the dark sector; however, this
also means HPS has not searched this regime since one of the key selections within its
previous displaced vertex search was requiring the total momentum of the vertex to be
near the beam energy instead of significantly below it.
With the samples presented in Chapter 8, a search for SIMPs within the collected
data has been performed. An additional, orthogonal category of data was introduced to
be combined with prior work, enabling investigation of a larger portion of the already-
collected data.
88
89
9.1 Signal Yield
The expected signal yield is a crucial part of any kind of search analysis and this one is no
exception. The prior search within LDMX benefited from its missing momentum/energy
technique in this regard since the mixing strength ϵ is only connected to the scale of
the signal yield and not how the signal events would present themselves in the data.
In a visible search, the mixing strength also affects the decay length and thus a more
intricate signal yield estimate is required to account for the fact that changing ϵ not only
changes the magnitude but also changes which events are more likely to be observed
within data. In this analysis, we modify the expected signal yield calculation done
previously by HPS [48] to accomodate the two viable decay channels within this SIMP
model.
The fundamental premise of this calculation is re-weighting events based on their
decay length so that they appear to have been drawn from a specific mixing strength’s
decay distribution. Specifically, we integrate over a decay weighting factor multiplied
by the signal efficiency as a function of the decay length in order to estimate the total
signal yield for a specific mixing strength. The details of this calculation are presented
in this section since they are relatively novel and not presented fully in another pub-
lication; however, this does force me to deviate from my previous commitment to few
mathematical equations and less jargony language. Skip to Section 9.2 if this detail
does not interest you.
The magnitude of this estimate is set by the relationship between dark photon
production cross section and the radiative trident differential cross section[49].
σA′ = ϵ2
3π
2α
mA′
dσγ∗
dm
∣∣∣∣
m=mA′
(9.1)
Multiplying both sides by the dataset’s luminosity then gives us event yields.
NA′ = ϵ2
3π
2α
mA′
dNγ∗
dm
∣∣∣∣
m=mA′
(9.2)
The complexity arises when we remember that specifically radiative tridents are not
directly observable – they are intertwined with other standard processes that produce
90
the same outgoing particles (Bethe-Heitler tridents for example). Thus, we estimate the
radiative trident differential yield dNγ∗/dm by modifying the observable trident differ-
ential yield dN3e,CR/dmreco by two simulation-derived factors. The radiative fraction
frad estimates the fraction of trident events that originate from the radiative process.
frad =
dNγ∗
dm
/
dN3e
dmreco
(9.3)
And the radiative acceptance times efficiency Arad estimates the fraction of trident
events that are within the geometric acceptance of the detector and pass the trigger
and preselection requirements.
Arad =
dN3e,CR
dmreco
/
dN3e,gen
dmreco
(9.4)
where CR stands for the Control Region in Psum (see Section 9.3 for the definition of
this variable). Thus, the expected yield of dark photons created within the detector but
not necessarily within its acceptance or passing selection requirements is
NA′ = ϵ2
3π
2α
mA′
frad
Arad
dN3e,CR
dmreco
∣∣∣∣
m=mA′
(9.5)
Both frad and Arad use the event and vertex pre-selection before any downstream anal-
ysis selections, so they are common amongst all potential analysis channels that share
this pre-selection; however, they both vary as a function of the dark photon mass. Poly-
nomials were fit to the simulation-derived values for these ratios and the parameters
for these polynomials are given in Table 9.1. These ingredients along with the resulting
NA′ is shown in Figure 9.1 using a run corresponding to ≈ 1.6% of the entire dataset.
In this signal hypothesis, we do not observe the dark photon’s production or decay,
instead, the dark photon decays to an unobservable dark pion and the neutral dark
vector meson VD which in turn is what decays back into a standard model electron-
positron pair. The above calculation of NA′ accounts for the dark photon’s production,
but we need to account for the dark photon’s decay and the dark vector meson’s decay.
The first decay has a branching ratio BR(A′ → πDVD) which is complicated by the fact
that there are actually two different dark neutral vector mesons that fit our requirements.
91
Coefficient frad Arad
c0 0.105 -0.489
c1 / MeV −1.17 × 10−3 0.0737
c2 / MeV2 7.44 × 10−6 −4.34 × 10−3
c3 / MeV3 −1.67 × 10−8 1.34 × 10−4
c4 / MeV4 0.0 −2.36 × 10−6
c5 / MeV5 0.0 2.54 × 10−8
c6 / MeV6 0.0 −1.71 × 10−10
c7 / MeV7 0.0 7.03 × 10−13
c8 / MeV8 0.0 −1.62 × 10−15
c9 / MeV9 0.0 1.60 × 10−18
Table 9.1: Polynomial coefficients for radiative fraction and acceptance functions.
101
102
103
104
N
3e
,C
R
0.1712pb 1HPS Internal
20
16
0.05
0.10f ra
d
0 50 100 150 200
Mass / MeV
0.0
0.2
0.4A r
ad
(a) Ingredients to the total dark photon
yield calculation including the observed tri-
dent yield (top), radiative fraction (middle),
and radiative acceptance (bottom).
0 50 100 150 200
mA / MeV
106
107
108
109
1010
N
A
/
2
0.1712pb 1HPS Internal
20
16
(b) Resulting total dark photon yield.
Figure 9.1: Depiction of dark photon yield calculation using a single run of the data
(∼ 1.6%). The gray vertical lines show the bounds of the search using the nominal ratio
mA′/mVD
= 1.66.
92
The second process is embedded in the decay rate of the dark vector meson to electron-
positron pairs Γ(VD → e+e−).1 Let E(z) be the signal efficiency of the analysis as a
function of the z where the VD decayed into the electron-positron pair. Then we can
sum over the possible VD and estimate the fraction of NA′ that produce a VD which
decays and passes the analysis requirements.
Nsig = NA′
∫ ∞
ztarget
∑
VD
DVD
(z)E(z)dz (9.6)
where
DVD
(z) = BR(A′ → πDVD)
e−(z−ztarget)/(γcτVD )
γcτVD
(9.7)
The branching ratio BR(A′ → πDVD) and lifetime τVD
are taken from [38] (along with
the general procedure of this estimate). What is important to remember is that the
lifetime is dependent on ϵ2, so while increasing ϵ2 increases NA′ it also makes the lifetime
shorter and thus the VD decay “looks” more like standard un-displaced background.
The VD energy (and thus the relativisitic γ) used in DVD
(z) is only distributed over
a small range (within < O(1GeV)) so we replace it with the mean ⟨γ⟩ in order to make
the calculation more practical. Additionally, we impose an upper limit on the decay z
to be 25 cm after the target since the efficiency of the reconstruction past ∼ 20 cm drops
to zero.
Nsig = NA′
∫ 25 cm
0
dz′
∑
VD
BR(A′ → πDVD)
e−z′/(⟨γ⟩cτVD )
⟨γ⟩cτVD
E(z′ + ztarget) (9.8)
where z′ = z − ztarget. Figure 9.2a shows Nsig with a uniformly perfect signal efficiency
E(z) = 1 to give a sense of scale, but a slightly more realistic example is using a step-wise
signal efficiency E(z) = 0.1 if 20mm < z < 100mm and 0 otherwise Figure 9.2b.
In summary, the key point to make here is that the signal efficiency as a function
of true decay length E(z) is the single analysis-dependent ingredient that effects the
resulting signal yield. The rest of the ingredients are either fixed by the dataset being
searched (e.g. trident differential production, frad, Arad) or are parameters that we vary
1There are other decay processes of the VD! Namely, the so-called “three-prong” decay VD → πDe+e−
would greatly increase the potential rate at the cost of preventing the e+e− from reconstructing the
mVD mass resonance.
93
20 40 60 80 100 120
Invariant Mass / MeV
10 8
10 7
10 6
10 5
10 42
0.1712pb 1HPS Internal
20
16
10 3
10 2
10 1
100
101
102
103
N
si
g (
E(
z)
=
1)
(a) Uniformly perfect signal efficiency with
yield dominated by prompt decays.
20 40 60 80 100 120
Invariant Mass / MeV
10 8
10 7
10 6
10 5
10 42
0.1712pb 1HPS Internal
20
16
10 3
10 2
10 1
100
101
102
103
N
si
g (
E(
z)
=
0.
1
if
20
.0
<
z<
10
0.
0m
m
el
se
0)
(b) A step signal efficiency mocking a perfect
analysis cut in displaced vertex z.
Figure 9.2: Expected signal yield Nsig for some rudimentary example efficiencies. Red
contour lines are given at 1 and 1.6% (the approximate fraction of this data subsample)
to give a sense of scale. Values are “clipped” to the color bar (for example, a signal
yield of 2 × 104 is colored yellow which is marked as 1 × 103) so we can focus on the
important orders of magnitude.
for the search (ϵ2, mVD
). A moderately realistic signal efficiency shows the range of ϵ2
and mVD
in which this dataset has potential sensitivity (Figure 9.2b) which informs the
ranges these values will take below.
9.2 Reconstruction Categories
The design of HPS separates reconstructed tracks into categories that are not necessarily
associated with the underlying physics for which we are searching. Specifically, the
number of layers (and which layers) that are included within a track has a large effect
on the resulting precision of the reconstructed physics variables of that track; thus,
we categorize vertices on whether one or both of their tracks contain both sensors in
the first layer. Figure 9.3 shows examples of the two categories considered within this
work. The L1L1 category whose verticies have both tracks with good precision was
first studied for this signal search[47] when optimizing the pre-selection requirements
described earlier.
94
L1
L2
Target
L1L1
e−
e+
L1
L2
Target
L1L2
e−
e+
Figure 9.3: Diagrams showing L1L1 (left) and L1L2 (right) vertex examples.
Extending this search to the L1L2 category has a rather simple motivation. While
the vertices may degrade in precision, the additional increase of data both in terms
of total volume and the potential to observe higher-displaced vertices is expected to
improve the sensitivity of the HPS SIMP search.
9.3 Physics Variables
There are many possible variables that could be used to separate candidate signal ver-
tices from the standard process backgrounds. While this section is not an exhaustive
list of all possible variables, it does explain the ones used within this analysis as well as
motivation for their use.
The SIMP signal vertex will have significantly less energy than the amount delivered
by the beam because of the production of a light dark meson when the heavier vector
meson VD is produced. Thus, a selection on the sum of the momentum magnitudes is
applied.
Psum = |p⃗e− |+ |p⃗e+ | (9.9)
Specifically, the Signal Region (SR) used for the actual search requires 1.0GeV < Psum <
1.9GeV and the Control Region (CR) used for determining the trident differential pro-
duction rate is 1.9GeV < Psum < 2.4GeV.
Since we are searching for the dark vector boson VD via its 2-body decay into an
electron and a positron, we expect the invariant mass of the vertex mreco to be within
95
the resolution of the detector σm of the mass we are searching for mVD
.
pm =
|mreco −mVD
|
σm
(9.10)
Applying an upper limit on pm is often refered to as a “mass window” since it results
in mreco residing within a small range around mVD
. For optimizing the other selections,
pm < 2.8 is used. The search for a potential excess of SIMP-like events is using sidebands
along mreco in data so no selection on pm is imposed when searching. The estimate of
the resulting sensitivity uses the signal region selection from the search of pm < 1.5.
Each vertex can be projected back to the target using its position and total mo-
mentum, and the location of the vertex at the target can be compared to the beam
position extracted from 1% of each data-collection run. The separation between the
projected position x⃗ = (x, y) and the beam position µ⃗ = (µx′ , µy′) is then measured and
normalized by the width of the beam in that direction (σx′ , σy′). The distribution of
the beamspot is allowed to be rotated relative to our chosen axes by the angle θbeam,
meaning we need to rotate x⃗ before comparing it to the beam spot. The total “ellipti-
cal” separation between the reconstructed vertex projected back to the target and the
beam spot can be quantified by
Nσ =
√(
x cos θbeam − y sin θbeam − µx′
σx′
)2
+
(
x sin θbeam + y cos θbeam − µy′
σy′
)2
(9.11)
I refer to this as VPS since it represents how significantly a vertex deviates from origi-
nating at the beam spot.
Since the detector’s tracking modules are oriented to be most senstive in the vertical
direction, the vertical impact parameter y0 has higher precision compared to the hor-
izontal impact parameter. For truly-displaced signal vertices, both tracks making the
vertex would have y0 far from zero while background vertices would have at least one
track with y0 near zero (undisplaced vertices would have both, but mis-reconstructed
fake-displaced vertices could have one far from zero). Figure 9.4 shows some diagrams
that illuminate this effect. While there are many ways for a vertex to end up being “fake
displaced” (for example, a missing or mis-chosen hit in the first layer like the example
shown in Figure 9.4), forcing both y0 to be far from zero removes these background
96
L1
L2
Target
Truly
Displaced
e−
e+
ye
−
0
ye
+
0
L1
L2
Target
Fake Displaced
ye
−
0 ≈ 0mm
e−
e+
True e+
ye
+
0
L1
L2
Target
Not Displaced
ye
−
0 ≈ 0mm
ye
+
0 ≈ 0mm e−
e+
Figure 9.4: Diagrams of examples for how different types of L1L2 vertices effect the
observed values of y0. Truly displaced vertices (left) have both tracks whose y0 is far
from 0mm while fake displaced vertices (middle) and not displaced (right) have at least
one if not both tracks with y0 ≈ 0mm. The diamonds are the reconstructed vertex, the
solid circles are the reconstructed hits, and the empty circles are “missed” hits either
because the particle did not pass through those sensors (left) or due to some physical
or electronic inefficiencies (center and right).
processes. This motivates selecting vertices based on requiring the minimum of the two
absolute value y0 to be above a certain threshold.
y0,min = min(|y0,e− |, |y0,e+ |) (9.12)
which more sharply distinguishes truly displaced vertices compared to the vertex z often
muddled by fake-displaced vertices where one track is mis-reconstructed at high |y0|.
The track fit uncertainty of the vertical impact parameter σy0 is a helpful quality
parameter measuring how confident the track fit is in the y0 value. Placing an upper
limit on this value for both tracks within a vertex effectively requires both tracks to
have good vertical resolution, helping remove some highly-displaced vertices presumably
arising from mis-reconstructed tracks.
σy0,max = max(σy0,e− , σy0,e+) (9.13)
Figure 9.5 displays the distributions of these last three variables for a few example
mass points after the other, solidified selections (L1L2 vertex, the momentum sum is in
the Signal Region, and the invariant mass falls within the chosen mass window).
Vertex z is left for late-stage statistical analysis of the results and – being highly
correlated with y0,min– is redundant with this variable.
97
0 5 10 15 20
VPS
10 4
10 3
10 2
10 1
100
Ev
en
t D
en
si
ty
1.07pb 1HPS Internal
20
16L1L2 and Psum SR
|mreco mVD|/ m < 2.8
Data (30 MeV)
Data (60 MeV)
Data (90 MeV)
SIMP 30 MeV
SIMP 60 MeV
SIMP 90 MeV
(a) VPS
0 2 4 6 8 10
y0, min / mm
10 5
10 4
10 3
10 2
10 1
100
101
102
Ev
en
t D
en
si
ty
1.07pb 1HPS Internal
20
16L1L2 and Psum SR
|mreco mVD|/ m < 2.8
Data (30 MeV)
Data (60 MeV)
Data (90 MeV)
SIMP 30 MeV
SIMP 60 MeV
SIMP 90 MeV
(b) y0,min
0.0 0.5 1.0 1.5 2.0 2.5 3.0
y0, max / mm
10 4
10 3
10 2
10 1
100
101
102
Ev
en
t D
en
si
ty
1.07pb 1HPS Internal
20
16L1L2 and Psum SR
|mreco mVD|/ m < 2.8
Data (30 MeV)
Data (60 MeV)
Data (90 MeV)
SIMP 30 MeV
SIMP 60 MeV
SIMP 90 MeV
(c) σy0,max
Figure 9.5: Distributions of cut variables for a few example mass points and ≈ 10% of
the full data sample. The vertices are required to be L1L2, have their momentum sum
be within the signal region, and their invariant mass within the specified mass window.
98
0 2 4 6 8 10
y0, min / mm
10 5
10 4
10 3
10 2
10 1
100
101
102
Ev
en
t D
en
si
ty
1.07pb 1HPS Internal
20
16L1L2 and Psum SR
|mreco mVD|/ m < 2.8
VPS < 4 and y0, max < 0.4mm
Data (30 MeV)
Data (60 MeV)
Data (90 MeV)
SIMP 30 MeV
SIMP 60 MeV
SIMP 90 MeV
Figure 9.6: y0,min distribution after the other selections (L1L2, Psum in SR, reconstructed
mass lying in mass window, VPS < 4, and σy0,max < 0.4mm).
9.4 Selection
The selection on these variables was optimized using a subsample of the collected data
amounting to ≈ 10% of the full sample. All of the variables except the last, separating
variable (y0,min) were varied while keeping the relative signal efficiency high (> 80%),
focused on removing the highly-displaced events leading to the long tail in data observed
in Figure 9.5b. Figure 9.6 shows the cleaned y0,min distribution after these selections.
There are still a few highly-displaced events left; however, removing them incurs a higher
penalty on signal efficiency and leads to a lower resulting sensitivity estimated from this
subsample.
The final variable y0,min was chosen by maximizing the binomial significance of the
expected signal at ϵ2 = 1 × 10−6 over the observed data for this 10% subsample. For
this optimization, the expected signal was artificially scaled such that it would be on
the same order-of-magnitude as the observed data (specifically, a scaling factor of 0.1/ϵ
was found to work well although variations in this scaling factor did not vary the chosen
selections by much).
The y0,min values chosen by this maximization (Figure 9.7) are ragged due to the
99
0 2 4 6 8 10
y0, min Cut
20
40
60
80
100
120
M
as
s
/ M
eV
1.07pb 1HPS Internal
20
16
-4
-2
0
2
4
6
8
Z B
i
(a) Binomial significance (ZBi) for a range of
masses and y0,min cut choices with the points
maximizing ZBi for each mass drawn in red
circles and the linear fit with its flat continu-
ation as a red line.
20 40 60 80 100 120
Mass / MeV
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
y 0
,m
in
C
ut
1.07pb 1HPS Internal
20
16
2 = 1.0e-06
Cuts Chosen by Maximizing ZBi
Fit to 40 MeV < mVD < 120 MeV and Extend
(b) Values maximing the binomial signficance
(blue) along with the linear fit and flat exten-
sion (orange).
Figure 9.7: Binomail significance (left) being maximized after the VPS< 4 and σy0,max
< 0.4mm cuts leading to cuts (right, blue) which are smoothed into a continuous
function (right, orange) as described in the text.
finite nature of the subsample on which they are being chosen. In order to avoid over-
biasing the selection to this arbitrarily-chosen subsample, the actual cuts used in the
final analysis are smoothed by fitting a line to the area whose maximumim binomial
significance is above zero and then continuining this line with flat values outside this
region.
In summary, the following cuts developed for this analysis are
• L1L2: In the vertex, one of the particles must have a hit in both sensors in layer
1 and layer 2 while the other must not satisfy this requirement.
• Low Momentum Sum: 1.0GeV < Psum < 1.9mm
• Decay after Target: z > ztarget
• Vertex Projects Back to Beamspot: VPS < 4
• Good Vertical Resolution: σy0,max < 0.4mm
100
Parameter Value
A 1.66mm
B 1.86mm
C −5.1 × 10−3mm/MeV
D 1.25mm
Table 9.2: Parameters for Equation (9.14).
A search for a statistically-significant excess of highly-displaced events reconstructed
within resolution of a specific mass is then performed at this point. The sensitivity is
estimated using two additional cuts.
• Within Mass Window: pm < 1.5
• Truly Displaced: y0,min > ycut0,min
where
ycut0,min(mreco) =
A mreco ≤ 40MeV
B + Cmreco 40MeV < mreco < 120MeV
D mreco ≥ 120MeV
(9.14)
and the parameters of this function are given in Table 9.2.
These selections lead to 6 highly-displaced events left in 10% of the data which
can be seen in Figure 9.8. The total signal efficiency of this analysis as a function of
reconstructed z Figure 9.9a is of the same magnitude but slightly further displaced than
the L1L1 analysis Figure 9.9b.
9.5 Results
With the selection fixed, we can analyze the full sample of data without fear of statistical
bias.
9.5.1 Search
The first stage in this analysis is to perform the search for an excess of highly-displaced
events above what could be expected by the background. The signal model we are
searching for would have extra events that are both highly-displaced and centered on
101
0 50 100 150 200 250
mreco / MeV
0
1
2
3
4
5
6
y 0
,m
in
/
m
m
L1L2
VPS < 4.0
y0, max < 0.4 mm
1.07pb 1HPS Internal
20
16
100
101
102
103
Ev
en
ts
Figure 9.8: y0,min distribution for the 10% data sample with the ycut0,min function overlayed
in red.
20 40 60 80 100 120
Invariant Mass / MeV
0
50
100
150
200
Tr
ue
V
er
te
x
z
/ m
m
L1L2
SIMP SimHPS Internal
20
16
0.000
0.005
0.010
0.015
0.020
0.025
Si
gn
al
E
ffi
ci
en
cy
(a) L1L2
20 40 60 80 100 120
Invariant Mass / MeV
0
50
100
150
200
Tr
ue
V
er
te
x
z
/ m
m
L1L1
VPS < 2 and y0, min ZBi Plain Max
SIMP SimHPS Internal
20
16
0.000
0.005
0.010
0.015
0.020
0.025
Si
gn
al
E
ffi
ci
en
cy
(b) L1L1
Figure 9.9: Signal efficiency as a function of z and mass scale mVD
for this analysis
(left) and L1L1 (right).
102
Region mreco Range y0,min Range
A (mVD
− 4.5σm,mVD
− 1.5σm) (ycut0,min,∞)
B (mVD
− 4.5σm,mVD
− 1.5σm) (yfloor0,min, y
cut
0,min)
C (mVD
− 1.5σm,mVD
+ 1.5σm) (yfloor0,min, y
cut
0,min)
D (mVD
+ 1.5σm,mVD
+ 4.5σm) (yfloor0,min, y
cut
0,min)
E (mVD
+ 1.5σm,mVD
+ 4.5σm) (ycut0,min,∞)
F (mVD
− 1.5σm,mVD
+ 1.5σm) (ycut0,min,∞)
Table 9.3: Region definitions for use in background estimation via sidebands. Region
F is the signal region in which we are searching for an excess. mVD
is the mass point
we are searching for, σm is the detector mass resolution evaluated at mVD
, ycut0,min is the
optimized cut value evaluated at mVD
, and yfloor0,min is the maximum value of y0,min such
that region C has at least one thousand events in it.
a specific resonant mass; thus, the search is performed in the y0,min– mreco space. The
expected number of background events is estimated using a variation of the ABCD
technique with two sidebands in mass (one below and one above the mass window in
which we are searching) and one sideband in y0,min (one below the highly-dipsplaced
region). The sidebands in mass are defined in terms of the mass resolution of the
detector σm and the values defining these edges were optimized within the L1L1 analysis.
The sideband in y0,min is defined such that the region within the mass window but at
lower displacment (Region C) has at least one thousand events within it to avoid signal
contaminating the background estimation.
Table 9.3 gives the definition of these regions. Figure 9.10a shows the search results
comparing the expected number of events to the observed along with a corresponding
p-value estimated using ten thousand toy experiments.2 Figure 9.10b shows an example
of the search calculation using the mass point resulting in the smallest p-value observed.
A few mass points had an excess of events observed within the signal region; however,
these excesses never exceeded a global significance of 3σ and the masses that do show the
largest excess in this channel do not display corresponding excess in the L1L1 channel.
Thus these excesses are understood as rare but normal statistical fluctuations.
2Each toy experiment is constructed by sampling a toy values for regions C (normally distributed
about actual value), B+D (normally distributed about actual value), and A+E (poisson distributed
about actual value) and then re-calculating the expected number in region F. The number of toy
experiments whose expected value in F was greater than or equal to the observed number in F divided
by the total number of toy experiments is then our estimate p-value.
103
0
2
4
6
8
10
12
Ev
en
ts 1.5 InvM Window
6.5 InvM Sideband
L1L2
VPS < 4, y0, max < 0.4 mm
y0, min ZBi Max in Core
10.7pb 1HPS Internal
20
16
SR L1L2
Observed
Expected
25 50 75 100 125 150 175 200
Invariant Mass / MeV
10 3
10 1
(a) Expected and observed events (top) with
their p-values (bottom).
0 50 100 150 200 250
mreco / MeV
0
1
2
3
4
5
6
y 0
,m
in
/
m
m
A
B C D
EF
mtrue = 97 MeV
A = 13 E = 1 B = 8018 D = 611 C = 1112
Fexp = C × (max(A + E, 0.4)/(B + D)) = 1.8
Fobs = 9
P Value = 4.2e-04
10.7pb 1HPS Internal
20
16
100
101
102
103
104
Ev
en
ts
(b) Search calculation yielding the minimum
p-value (with ycut0,min in light red).
Figure 9.10: Search results for the full 2016 data set.
9.5.2 Sensitivity and Exclusion
With no statistically-significant excess observed, we can then move towards estimating
the sensitivity of the analysis and then to exclude certain regions of parameter space.
The sensitivity estimate is rather simple; just a ratio between the expected signal yield
and the maximum signal yield allowed by the observed data distribution. Section 9.1
already described the expected signal calculation used here as well as previously within
the selection optimization. Figure 9.11a shows the final expected signal yield for this
selection using the full data sample. The maximum signal yield allowed is calculated
using the OIM [50] which leverages the differing shape between signal and background
and allows for the presence of unknown background to alter its scale. The OIM result
is shown in Figure 9.11b.
Figure 9.12 shows the sensitivity for both this analysis of L1L2 and the L1L1 analysis
for comparison along with a contour drawn where the sensitivity equals one after being
suppressed by potential systematic errors. Systematic errors arising from the radiative
fraction estimate, pre-selection cuts, radiative acceptance estimate, uncertainty in the
target position, final selection cuts, mass resolution, beamspot resolution, and difference
104
20 40 60 80 100 120
Invariant Mass / MeV
10 8
10 7
10 6
10 5
10 42
L1L2
10.7pb 1HPS Internal
20
16
0
2
4
6
8
10
12
Ex
pe
ct
ed
S
ig
na
l
(a) Expected
20 40 60 80 100 120
Invariant Mass / MeV
10 8
10 7
10 6
10 5
10 42
L1L2
10.7pb 1HPS Internal
20
16
4
6
8
10
12
14
16
18
M
ax
S
ig
na
l A
llo
w
ed
E
st
im
at
e
(b) Maximum Allowed
Figure 9.11: Signal yield for the L1L2 reconstruction category with the described cuts
showing the theoretical expecation (left) and the maximum allowed (right) derived using
the OIM.
in shape of Psum distributions were all evaluated leading to total systematic error of
< 10% for all but the lowest mass points evaluated (rising up to ∼ 18%).
This contour encloses the excluded parameter space where the expected yield ex-
ceeds what is allowed. In this region of parameter space, BABAR [7] has excluded the
parameters above ϵ2 = 10−6, so this analysis does not exclude new parameter space.
Nevertheless, this analysis was able to confirm BABAR’s results and is the first of its
kind (displaced-vertex search in the low-Psum region) for HPS and thus opens the door
to further refinement and investigation with later and larger datasets.
105
20 40 60 80 100 120
Invariant Mass / MeV
10 8
10 7
10 6
10 5
10 42
L1L2
10.7pb 1HPS Internal
20
16
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
Ex
pe
ct
ed
/
M
ax
A
llo
w
ed
(a) L1L2
20 40 60 80 100 120
Invariant Mass / MeV
10 8
10 7
10 6
10 5
10 42
L1L1
10.7pb 1HPS Internal
20
16
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Ex
pe
ct
ed
/
M
ax
A
llo
w
ed
(b) L1L1
Figure 9.12: Sensitivity of this analysis with contour drawn including systematic errors
surpressing the expected signal yield.
Part IV
Conclusion
106
Chapter 10
Conclusion
Both LDMX and HPS are important and required experiments in our search for the par-
ticle nature of DM. While neither has progressed far enough to exclude new parameter
space (or potentially discover DM that was previously out of reach), both experiments
take novel approaches towards this search. These approaches are required because the
traditional tactics have not yielded discovery and are unable to reach into these specific
regions of parameter space.
The LDMX EaT analysis presented here is novel for LDMX not only due to the
missing-energy search style, but also because we took care to estimate the magnitude
of potential systematic errors, form a quantitative non-zero background prediction with
its own statistical uncertainty, and performed the statistical analysis over several bins.
Previous LDMX analyses focusing on zero-background searches, while important and
interesting in their own right, do not have the flexibility necessary to be an analysis
used with data collected form a newly built apparatus. The EaT analysis channel is
such an analysis and it is well prepared to be the first analysis of LDMX data making
new physics conclusions about our universe.
The HPS SIMP L1L2 analysis extended the SIMP search to a previously-unexplored
reconstruction category, enabling us to use a larger fraction of the data already collected
by HPS. This SIMP search, while not excluding new phase space with the current data
set, shows promise. We have been able to show that a displaced-vertex search with the
possibility of missing energy is possible. Extending such a search to unexplored phase
space, an area where no other experiments have access, is simply awaiting the necessary
107
108
reconstruction studies in order to analyze later and larger data sets already collected
by the detector.
As newer and smaller experiments compared to other HEP experiments, different
excitements and difficulties arise. While participating in the full breadth of a particle
physics experiment has taught me a lot and given me opportunity to carve my own path
of learning, the lack of larger support structures has revealed some specifically intricate
problems that LDMX and HPS share.
10.1 Simulation Validity
The design and initial studies of any experiment requires some simulation of the mea-
surements it could perform. After the experiment has been constructed and collects
data, the simulation is still incredibly useful for studying how known processes present
themselves within the observations. The collected data could have a previously-unknown
and rare process within it, so we must be careful to avoid biasing our analysis in such
a way that would ruin the validity of our results. A reasonable strategy is simply to
avoid studying all of the data at once and instead design and optimize the analysis on a
particular subset of the data. This process is called “blinding” and was used in Part III;
however, this also naturally means the analysis can easily be optimized only for the
special events within the subset and the events outside of this subset are unaccounted
for.
A valid simulation, one which correctly predicts the scale and shape of the observable
distributions, is a valuable tool that could partially or completely replace this blinding
procedure. In this case, the simulated observations could be used to help design and
optimize the analysis; avoiding bias while also allowing the analysis to explore the full
breadth of potential events that could be seen.
While the “valid simulation” I describe here is not necessarily possible, we can move
forward towards it in order to make future analyses better. One of the first studies
that will be done with new LDMX data will be comparing how our simulated events
compare to the events actually being collected by the real detector. Returning to these
comparisons for HPS and its subsequent, larger sets of data is being done and will be
beneficial for those analyses as well especially given the separation between data and
109
simulation already being observed within the pre-selection (for example, in Figure 8.1).
10.2 Tracking Reconstruction
Most of my work on tracking and the alignment of a tracking detector was done within
the context of HPS. This particular tracking detector is somewhat of a “trial by fire”
due to its delicate, two-sided nature required by the specific physics for which we are
searching. Nevertheless, this experience has given me a solid introduction to charged
particle tracking.
As discussed in Chapter 7, charged particle tracking is very complicated and it
almost certainly requires familiarity with the equally-complicated process of alignment.
Going forward, both LDMX and HPS require robust and solid tracking software in order
to make the interpretations of such intricate algorithms simpler. Without the backing
of a larger collaboration to more completely test and validate this complicated software,
using a shared tracking framework like ACTS[51] is necessary, a strategy starting to be
employed in both experiments.
10.3 Reflection
This dissertation work has shown me the full breadth of a particle physics experiment.
From initial design and simulation, to extensions and validations of said simulation, to
intricate considerations of potential systematic errors in the experiment, to optimization
of selections using various figures of merit, to intentional analysis blinding to avoid
statistical bias when developing an analysis, to final unblinding after proposals and
evaluations, to post-mortem analysis of the leftover bits of data.
While I have not documented all of these steps within this document, they all brought
with them lessons for which I am grateful.
References
[1] Randall Munroe. xkcd webcomic. Licensed under Creative-Commons Attribution-
NonCommercial 2.5.
[2] Torsten Åkesson et al. (LDMX Collaboration). A High Efficiency Photon Veto for
the Light Dark Matter eXperiment. JHEP, 04:003, 2020, 1912.05535.
[3] CERN. The decay of a lambda particle in the 32 cm hydrogen bubble chamber.
CERN-EX-11465, 1960.
[4] Gerard Jungman, Marc Kamionkowski, and Kim Griest. Supersymmetric dark
matter. Physics Reports, 267(5):195–373, 1996.
[5] D. Banerjee et al. (NA64 Collaboration). Dark Matter Search in Missing Energy
Events with NA64. Physical Review Letters, 123(12), Sep 2019.
[6] D. Banerjee et al. (NA64 Collaboration). Improved limits on a hypotheticalX(16.7)
boson and a dark photon decaying into e+e− pairs. Phys. Rev. D, 101:071101, Apr
2020.
[7] J. P. Lees et al. (BABAR Collaboration). Search for Invisible Decays of a Dark
Photon Produced in e+e- Collisions at BABAR. Physical Review Letters, 119(13),
September 2017.
[8] D. Akimov et al. (COHERENT Collaboration). First Probe of Sub-GeV Dark
Matter beyond the Cosmological Expectation with the COHERENT CsI Detector
at the SNS. Phys. Rev. Lett., 130:051803, Feb 2023.
110
111
[9] Samantha K McCarty. Search for an A’ Resonance in the Heavy Photon Search
2.3 GeV Data Set. PhD thesis, University of New Hampshire, 2020.
[10] H Bethe and W Heitler. On the stopping of fast particles and on the creation of
positive electrons. Proceedings of the Royal Society London A, 146:83–112, 1934.
[11] V. C. Rubin, Jr. Ford, W. K., and N. Thonnard. Rotational properties of 21 SC
galaxies with a large range of luminosities and radii, from NGC 4605 (R=4kpc) to
UGC 2885 (R=122kpc). APJ, 238:471–487, June 1980.
[12] Edvige Corbelli and Paolo Salucci. The extended rotation curve and
the dark matter halo of M33. Monthly Notices of the Royal Astronomi-
cal Society, 311(2):441–447, 01 2000, https://academic.oup.com/mnras/article-
pdf/311/2/441/2881340/311-2-441.pdf.
[13] Fred I. Cooperstock and S. Tieu. Galactic Dynamics via General Relativity: A
Compilation and New Developments. Int. J. Mod. Phys. A, 22:2293–2325, 2007,
astro-ph/0610370.
[14] Douglas Clowe, Anthony Gonzalez, and Maxim Markevitch. Weak-Lensing Mass
Reconstruction of the Interacting Cluster 1E 0657-558: Direct Evidence for the
Existence of Dark Matter. The Astrophysical Journal, 604(2):596, apr 2004.
[15] P. A. R. Ade et al. (Planck Collaboration). Planck 2015 results: Xiii. cosmological
parameters. Astronomy & Astrophysics, 594:A13, September 2016.
[16] R. Amanullah et al. (The Supernova Cosmology Project Collaboration). Spectra
and hubble space telescope light curves of six type ia supernovae at 0.511 < z < 1.12
and the union2 compilation. The Astrophysical Journal, 716(1):712, may 2010.
[17] David N. Schramm and Michael S. Turner. Big-bang nucleosynthesis enters the
precision era. Rev. Mod. Phys., 70:303–318, Jan 1998.
[18] Bernard Carr, Kazunori Kohri, Yuuiti Sendouda, and Jun’ichi Yokoyama. Con-
straints on primordial black holes. Reports on Progress in Physics, 84(11):116902,
nov 2021.
112
[19] Jim Alexander et al. Dark Sectors 2016 Workshop: Community Report, 2016,
1608.08632.
[20] Kenneth M. Nollett and Gary Steigman. BBN and the CMB constrain neutrino
coupled light WIMPs. Physical Review D, 91(8), April 2015.
[21] B. D. Fields, P. Molaro, and Sarkar S. Big Bang Nucleosynthesis. Phys. Rev. D,
110(3):030001, 2024. Chapter 24 of PDG Review of Particle Physics.
[22] R. Agnese et al. (SuperCDMS Collaboration). First dark matter constraints from a
SuperCDMS single-charge sensitive detector. Physical Review Letters, 121(5), aug
2018.
[23] A. Aguilar-Arevalo et al. (DAMIC Collaboration). Results on low-mass weakly in-
teracting massive particles from an 11 kg d target exposure of DAMIC at SNOLAB.
Physical Review Letters, 125(24), dec 2020.
[24] E. Aprile et al. (XENON1T Collaboration). Dark matter search results from a one
ton-year exposure of XENON1T. Physical Review Letters, 121(11), sep 2018.
[25] Torsten Åkesson et al. (LDMX Collaboration). Light Dark Matter eXperiment
(LDMX), 2018, 1808.05219.
[26] Torsten Åkesson et al. (LDMX Collaboration). Photon-rejection power of the Light
Dark Matter eXperiment in an 8 GeV beam. JHEP, 12:092, 2023, 2308.15173.
[27] Tor Raubenheimer (LCLS-II Collaboration). LCLS-II: Status of the CW X-ray
FEL Upgrade to the LCLS Facility. In 37th International Free Electron Laser
Conference, pages 618–624, 2015.
[28] T G Buekers, J W Amann, and Y M Nosochkov. A Beam Spreader System for
LCLS-II. NAPAC 2019 Proceedings, 09 2019.
[29] D. Contardo et al. The Phase-2 Upgrade of the CMS Endcap Calorimeter. Technical
Report CERN-LHCC-2017-023, CMS-TDR-019, CMS, 2017.
[30] Rene Brun et al. root-project/root: v6.18/02, August 2019.
113
[31] Guido Van Rossum and Fred L. Drake. Python 3 Reference Manual. CreateSpace,
Scotts Valley, CA, 2009.
[32] S. Agostinelli et al. Geant4 – a simulation toolkit. Nuclear Instruments and Methods
in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associ-
ated Equipment, 506(3):250–303, 2003.
[33] Tom Eichlersmith, Jeremiah Mans, Omar Moreno, Joseph Muse, Michael Revering,
and Natalia Toro. Simulation of dark bremsstrahlung in Geant4. Computer Physics
Communications, 287:108690, 2023.
[34] Aram Hayrapetyan et al. (CMS Collaboration). The CMS statistical analysis and
combination tool: Combine. Comput. Softw. Big Sci., 8:19, 2024, 2404.06614.
[35] M. Bondi, A. Celentano, R. R. Dusaev, D. V. Kirpichnikov, M. M. Kirsanov, N. V.
Krasnikov, L. Marsicano, and D. Shchukin. Fully Geant4 compatible package for the
simulation of Dark Matter in fixed target experiments. Comput. Phys. Commun.,
269:108129, 2021, 2101.12192.
[36] B. Banto Oberhauser, P. Bisio, A. Celentano, E. Depero, R. R. Dusaev, D. V. Kir-
pichnikov, M. M. Kirsanov, N. V. Krasnikov, A. Marini, L. Marsicano, L. Molina-
Bueno, M. Mongillo, D. Shchukin, H. Sieber, and I. V. Voronchikhin. Development
of the fully Geant4 compatible package for the simulation of Dark Matter in fixed
target experiments, 2024, 2401.12573.
[37] Yonit Hochberg, Eric Kuflik, Tomer Volansky, and Jay G. Wacker. Mechanism
for Thermal Relic Dark Matter of Strongly Interacting Massive Particles. Physical
Review Letters, 113(17), October 2014.
[38] Asher Berlin, Nikita Blinov, Stefania Gori, Philip Schuster, and Natalia Toro. Cos-
mology and accelerator tests of strongly interacting dark matter. Physical Review
D, 97(5), March 2018.
[39] Matthew Regan Solt. Searching for long-lived dark photons with the heavy photon
search experiment. PhD thesis, Stanford, 2020.
114
[40] Volker D. Burkert. The JLab 12 GeV upgrade and the initial science program -
Selected topics. Proc. Int. Sch. Phys. Fermi, 180:303–332, 2012.
[41] Dudek, Jozef, Ent, Rolf, Essig, Rouven, Kumar, K. S., Meyer, Curtis, McKeown,
R. D., Meziani, Zein Eddine, Miller, Gerald A., Pennington, Michael, Richards,
David, Weinstein, Larry, Young, Glenn, and Brown, Susan. Physics opportunities
with the 12 GeV upgrade at Jefferson Lab. Eur. Phys. J. A, 48(12):187, 2012.
[42] Reza Kazimi, John Hansknecht, Michael Spata, and Haipeng Wang. Source and
extraction for simultaneous four-hall beam delivery system at cebaf. Proceedings,
4th International Particle Accelerator Conference (IPAC 2013), 1(4), 2013.
[43] Claus Kleinwort. General broken lines as advanced track fitting method. Nuclear
Instruments and Methods in Physics Research Section A: Accelerators, Spectrome-
ters, Detectors and Associated Equipment, 673:107–110, 2012.
[44] Volker Blobel, Claus Kleinwort, and Frank Meier. Fast alignment of a complex
tracking detector using advanced track models. Computer Physics Communica-
tions, 182(9):1760–1763, 2011. Computer Physics Communications Special Edition
for Conference on Computational Physics Trondheim, Norway, June 23-26, 2010.
[45] Johan Alwall, Pavel Demin, Simon de Visscher, Rikkert Frederix, Michel Her-
quet, Fabio Maltoni, Tilman Plehn, David L Rainwater, and Tim Stelzer. Mad-
graph/madevent v4: the new web generation. Journal of High Energy Physics,
2007(09):028–028, September 2007.
[46] J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H.-S.
Shao, T. Stelzer, P. Torrielli, and M. Zaro. The automated computation of tree-
level and next-to-leading order differential cross sections, and their matching to
parton shower simulations. Journal of High Energy Physics, 2014(7), July 2014.
[47] Alic Spellman. Searching for Strongly-Interacting Dark Matter With the Heavy
Photon Search Experiment. PhD thesis, University of California Santa Clara, 2024.
[48] P. H. Adrian et al. (HPS Collaboration). Searching for prompt and long-lived dark
photons in electroproduced electron-positron pairs with the heavy photon search
experiment at JLab. Physical Review D, 108(1), July 2023.
115
[49] James D. Bjorken, Rouven Essig, Philip Schuster, and Natalia Toro. New fixed-
target experiments to search for dark gauge forces. Physical Review D, 80(7), oct
2009.
[50] S. Yellin. Finding an upper limit in the presence of an unknown background.
Physical Review D, 66(3), August 2002.
[51] Xiaocong Ai et al. A Common Tracking Software Project. Computing and Software
for Big Science, 6, 2022.
Acknowledgements
Dedication
Abstract
Contents
List of Tables
List of Figures
I Laying the Groundwork
Introduction
Standard Model of Particle Physics
Standard Model Analogies
Particle Mixing
Displaced Particle Decays
Dark Matter
Dark Matter?
Evidence for Dark Matter
Particle Nature of Dark Matter
Invisible Signature
Visible Signature
II LDMX
Light Dark Matter eXperiment
Missing Momentum Signature
The Beam Line
Detector Design
Mid-Shower Simulation
General Data Process
Data Processing Stages
Biasing and Filtering Technique
Standard Processes
Dominant Contributors in this Analysis
Expanding Production of these Processes
Dark Matter Signal
G4DarkBreM
MadGraph/MadEvent
Characterization
Summary
Missing Energy Search
Selection
Background Prediction
Systematics and Background Uncertainty
Reach
III HPS
The Heavy Photon Search Experiment
Strongly Interacting Massive Particles
Detector Apparatus
Silicon Vertex Tracker
Electromagnetic Calorimeter
Alignment
Alignment Procedure
Detector Visualization
Results
Data Set
Collected Data
Pair 1 Trigger
Simulation
Reconstruction
Analysis Pre-Selection
Displaced Vertex Search
Signal Yield
Reconstruction Categories
Physics Variables
Selection
Results
Search
Sensitivity and Exclusion
IV Conclusion
Conclusion
Simulation Validity
Tracking Reconstruction
Reflection
References