of 21
Identification and mitigation of narrow spectral artifacts that degrade searches for
persistent gravitational waves in the first two observing runs of Advanced LIGO
P. B. Covas,
1
A. Effler,
2
E. Goetz,
3
,
4
P. M. Meyers,
5
A. Neunzert,
3
M. Oliver,
1
B. L. Pearlstone,
6
V. J. Roma,
7
R. M. S. Schofield,
7
V. B. Adya,
4
P. Astone,
8
S. Biscoveanu,
9
,
10
T. A. Callister,
11
N. Christensen,
12
,
13
A. Colla,
14
,
8
E. Coughlin,
12
M. W. Coughlin,
12
,
11
S. G. Crowder,
15
S. E. Dwyer,
16
S. Hourihane,
3
S. Kandhasamy,
2
W. Liu,
3
A. P. Lundgren,
4
A. Matas,
5
R. McCarthy,
16
J. McIver,
11
G. Mendell,
16
R. Ormiston,
5
C. Palomba,
8
O. J. Piccinni,
14
,
8
K. Rao,
3
K. Riles,
3
L. Sammut,
10
S. Schlassa,
12
D. Sigg,
16
N. Strauss,
12
D. Tao,
12
K. A. Thorne,
2
E. Thrane,
10
S. Trembath-Reichert,
3
B. P. Abbott,
11
R. Abbott,
11
T. D. Abbott,
17
C. Adams,
2
R. X. Adhikari,
11
A. Ananyeva,
11
S. Appert,
11
K. Arai,
11
S. M. Aston,
2
C. Austin,
17
S. W. Ballmer,
18
D. Barker,
16
B. Barr,
6
L. Barsotti,
19
J. Bartlett,
16
I. Bartos,
20
J. C. Batch,
16
M. Bejger,
21
A. S. Bell,
6
J. Betzwieser,
2
G. Billingsley,
11
J. Birch,
2
S. Biscans,
19
C. Biwer,
18
C. D. Blair,
2
R. M. Blair,
16
R. Bork,
11
A. F. Brooks,
11
H. Cao,
22
G. Ciani,
20
F. Clara,
16
P. Clearwater,
23
S. J. Cooper,
24
P. Corban,
2
S. T. Countryman,
25
M. J. Cowart,
2
D. C. Coyne,
11
A. Cumming,
6
L. Cunningham,
6
K. Danzmann,
26
,
4
C. F. Da Silva Costa,
20
E. J. Daw,
27
D. DeBra,
28
R. T. DeRosa,
2
R. DeSalvo,
29
K. L. Dooley,
30
S. Doravari,
4
J. C. Driggers,
16
T. B. Edo,
27
T. Etzel,
11
M. Evans,
19
T. M. Evans,
2
M. Factourovich,
25
H. Fair,
18
A. Fern ́andez Galiana,
19
E. C. Ferreira,
31
R. P. Fisher,
18
H. Fong,
32
R. Frey,
7
P. Fritschel,
19
V. V. Frolov,
2
P. Fulda,
20
M. Fyffe,
2
B. Gateley,
16
J. A. Giaime,
17
,
2
K. D. Giardina,
2
R. Goetz,
20
B. Goncharov,
10
S. Gras,
19
C. Gray,
16
H. Grote,
4
K. E. Gushwa,
11
E. K. Gustafson,
11
R. Gustafson,
3
E. D. Hall,
19
G. Hammond,
6
J. Hanks,
16
J. Hanson,
2
T. Hardwick,
17
G. M. Harry,
33
M. C. Heintze,
2
A. W. Heptonstall,
11
J. Hough,
6
R. Inta,
34
K. Izumi,
16
R. Jones,
6
S. Karki,
7
M. Kasprzack,
17
S. Kaufer,
26
K. Kawabe,
16
R. Kennedy,
27
N. Kijbunchoo,
35
W. Kim,
22
E. J. King,
22
P. J. King,
16
J. S. Kissel,
16
W. Z. Korth,
11
G. Kuehn,
4
M. Landry,
16
B. Lantz,
28
M. Laxen,
2
J. Liu,
36
N. A. Lockerbie,
37
M. Lormand,
2
M. MacInnis,
19
D. M. Macleod,
38
S. M ́arka,
25
Z. M ́arka,
25
A. S. Markosyan,
28
E. Maros,
11
P. Marsh,
39
I. W. Martin,
6
D. V. Martynov,
19
K. Mason,
19
T. J. Massinger,
11
F. Matichard,
11
,
19
N. Mavalvala,
19
D. E. McClelland,
35
S. McCormick,
2
L. McCuller,
19
G. McIntyre,
11
T. McRae,
35
E. L. Merilh,
16
J. Miller,
19
R. Mittleman,
19
G. Mo,
12
K. Mogushi,
30
D. Moraru,
16
G. Moreno,
16
G. Mueller,
20
N. Mukund,
40
A. Mullavey,
2
J. Munch,
22
T. J. N. Nelson,
2
P. Nguyen,
7
L. K. Nuttall,
38
J. Oberling,
16
O. Oktavia,
15
P. Oppermann,
4
Richard J. Oram,
2
B. O’Reilly,
2
D. J. Ottaway,
22
H. Overmier,
2
J. R. Palamos,
7
W. Parker,
2
A. Pele,
2
S. Penn,
41
C. J. Perez,
16
M. Phelps,
6
V. Pierro,
42
I. Pinto,
42
M. Principe,
42
L. G. Prokhorov,
43
O. Puncken,
4
V. Quetschke,
44
E. A. Quintero,
11
H. Radkins,
16
P. Raffai,
45
K. E. Ramirez,
44
S. Reid,
46
D. H. Reitze,
11
,
20
N. A. Robertson,
11
,
6
J. G. Rollins,
11
C. L. Romel,
16
J. H. Romie,
2
M. P. Ross,
47
S. Rowan,
6
K. Ryan,
16
T. Sadecki,
16
E. J. Sanchez,
11
L. E. Sanchez,
11
V. Sandberg,
16
R. L. Savage,
16
D. Sellers,
2
D. A. Shaddock,
35
T. J. Shaffer,
16
B. Shapiro,
28
D. H. Shoemaker,
19
B. J. J. Slagmolen,
35
B. Smith,
2
J. R. Smith,
48
B. Sorazu,
6
A. P. Spencer,
6
A. Staley,
25
K. A. Strain,
6
L. Sun,
23
D. B. Tanner,
20
J. D. Tasson,
12
R. Taylor,
11
M. Thomas,
2
P. Thomas,
16
K. Toland,
6
C. I. Torrie,
11
G. Traylor,
2
M. Tse,
19
D. Tuyenbayev,
44
G. Vajente,
11
G. Valdes,
17
A. A. van Veggel,
6
A. Vecchio,
24
P. J. Veitch,
22
K. Venkateswara,
47
T. Vo,
18
C. Vorvick,
16
M. Wade,
49
M. Walker,
48
R. L. Ward,
35
J. Warner,
16
B. Weaver,
16
R. Weiss,
19
P. Weßels,
4
B. Willke,
26
,
4
C. C. Wipf,
11
J. Wofford,
50
J. Worden,
16
H. Yamamoto,
11
C. C. Yancey,
51
Hang Yu,
19
Haocun Yu,
19
L. Zhang,
11
S. Zhu,
4
M. E. Zucker,
11
,
19
and J. Zweizig
11
(LSC Instrument Authors)
1
Universitat de les Illes Balears, IAC3—IEEC, E-07122 Palma de Mallorca, Spain
2
LIGO Livingston Observatory, Livingston, LA 70754, USA
3
University of Michigan, Ann Arbor, MI 48109, USA
4
Max Planck Institute for Gravitational Physics (Albert Einstein Institute), D-30167 Hannover, Germany
5
University of Minnesota, Minneapolis, MN 55455, USA
6
SUPA, University of Glasgow, Glasgow G12 8QQ, United Kingdom
7
University of Oregon, Eugene, OR 97403, USA
8
INFN, Sezione di Roma, I-00185 Roma, Italy
9
The Pennsylvania State University, University Park, PA 16802, USA
10
OzGrav, School of Physics & Astronomy, Monash University, Clayton 3800, Victoria, Australia
11
LIGO, California Institute of Technology, Pasadena, CA 91125, USA
12
Carleton College, Northfield, MN 55057, USA
13
Artemis, Universit ́e Cˆote d’Azur, Observatoire Cˆote d’Azur,
CNRS, CS 34229, F-06304 Nice Cedex 4, France
14
Universit`a di Roma ‘La Sapienza,’ I-00185 Roma, Italy
15
Bellevue College, Bellevue, WA 98007, USA
arXiv:1801.07204v1 [astro-ph.IM] 22 Jan 2018
2
16
LIGO Hanford Observatory, Richland, WA 99352, USA
17
Louisiana State University, Baton Rouge, LA 70803, USA
18
Syracuse University, Syracuse, NY 13244, USA
19
LIGO, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
20
University of Florida, Gainesville, FL 32611, USA
21
Nicolaus Copernicus Astronomical Center, Polish Academy of Sciences, 00-716, Warsaw, Poland
22
OzGrav, University of Adelaide, Adelaide, South Australia 5005, Australia
23
OzGrav, University of Melbourne, Parkville, Victoria 3010, Australia
24
University of Birmingham, Birmingham B15 2TT, United Kingdom
25
Columbia University, New York, NY 10027, USA
26
Leibniz Universit ̈at Hannover, D-30167 Hannover, Germany
27
The University of Sheffield, Sheffield S10 2TN, United Kingdom
28
Stanford University, Stanford, CA 94305, USA
29
California State University, Los Angeles, 5151 State University Dr, Los Angeles, CA 90032, USA
30
The University of Mississippi, University, MS 38677, USA
31
Instituto Nacional de Pesquisas Espaciais, 12227-010 S ̃ao Jos ́e dos Campos, S ̃ao Paulo, Brazil
32
Canadian Institute for Theoretical Astrophysics,
University of Toronto, Toronto, Ontario M5S 3H8, Canada
33
American University, Washington, D.C. 20016, USA
34
Texas Tech University, Lubbock, TX 79409, USA
35
OzGrav, Australian National University, Canberra, Australian Capital Territory 0200, Australia
36
OzGrav, University of Western Australia, Crawley, Western Australia 6009, Australia
37
SUPA, University of Strathclyde, Glasgow G1 1XQ, United Kingdom
38
Cardiff University, Cardiff CF24 3AA, United Kingdom
39
University of Washington Bothell, 18115 Campus Way NE, Bothell, WA 98011, USA
40
Inter-University Centre for Astronomy and Astrophysics, Pune 411007, India
41
Hobart and William Smith Colleges, Geneva, NY 14456, USA
42
University of Sannio at Benevento, I-82100 Benevento,
Italy and INFN, Sezione di Napoli, I-80100 Napoli, Italy
43
Faculty of Physics, Lomonosov Moscow State University, Moscow 119991, Russia
44
The University of Texas Rio Grande Valley, Brownsville, TX 78520, USA
45
Institute of Physics, E ̈otv ̈os University, P ́azm ́any P. s. 1/A, Budapest 1117, Hungary
46
SUPA, University of the West of Scotland, Paisley PA1 2BE, United Kingdom
47
University of Washington, Seattle, WA 98195, USA
48
California State University Fullerton, Fullerton, CA 92831, USA
49
Kenyon College, Gambier, OH 43022, USA
50
Rochester Institute of Technology, Rochester, NY 14623, USA
51
University of Maryland, College Park, MD 20742, USA
Searches are under way in Advanced LIGO and Virgo data for persistent gravitational waves
from continuous sources, e.g. rapidly rotating galactic neutron stars, and stochastic sources, e.g.
relic gravitational waves from the Big Bang or superposition of distant astrophysical events such as
mergers of black holes or neutron stars. These searches can be degraded by the presence of narrow
spectral artifacts (lines) due to instrumental or environmental disturbances. We describe a variety
of methods used for finding, identifying and mitigating these artifacts, illustrated with particular
examples. Results are provided in the form of lists of line artifacts that can safely be treated as
non-astrophysical. Such lists are used to improve the efficiencies and sensitivities of continuous
and stochastic gravitational wave searches by allowing vetoes of false outliers and permitting data
cleaning.
I. INTRODUCTION
The recent detections of transient gravitational waves
(GWs) from the merger of binary black holes and of bi-
nary neutron stars opened a new field of observational
GW astronomy [1, 2]. The near future may also bring
the discovery by the LIGO and Virgo detectors of
persis-
tent
gravitational waves.
Persistent sources of long-duration GWs can be
broadly classified as continuous wave (CW) sources,
which have a deterministic phase evolution, and a
stochastic gravitational-wave background (SGWB), for
which the signal is intrinsically random. The canon-
ical sources for CWs (see [3] for a review) are non-
axisymmetric rotating neutron stars, emitting long-
lasting and nearly monochromatic waves. When observed
from Earth, these waves will be frequency-modulated due
to the Doppler effect produced by the daily rotation and
orbital motion of the Earth around the Sun. The SGWB
is a superposition of many astrophysical and cosmological
GW sources. Astrophysical sources are reviewed in [4].
Cosmological sources of the SGWB include cosmic string
networks [5–8], inflation [9–16], phase transitions [17–19],
and the pre-Big Bang scenario [20–23]. For reviews of
3
search methods for the SGWB, see [24, 25].
CW and SGWB searches look for long-duration sig-
nals, and are affected by different types of noise than
those affecting short-duration searches.
While com-
pact binary coalescence and burst searches are degraded
mainly by short-duration glitches (such as those de-
scribed in [26–28]), CW and SGWB searches are mainly
affected by long-lived peaks in the frequency spectra, es-
pecially narrow peaks, typically referred to as lines. CW
searches can be degraded because their signals are intrin-
sically highly narrow-band, while SGWB searches can be
degraded because of the tendency of a subset of instru-
mental lines in the two detectors to lie so close to each
other that they exhibit spurious coherence between the
detectors.
This problem presents two main detector characteri-
zation tasks for long-duration searches: first, to
identify
line artifacts that are non-astrophysical in origin, allow-
ing them to be flagged as noise; and, second, to deter-
mine the
cause
of those artifacts when possible in order to
guide efforts to remove them at the detector sites. Spec-
tral lines that affect the CW and the SGWB searches
are typically quite narrow (high Q-factor, i.e., the ratio
of peak frequency to line width) during a given coherent
integration time. This focuses investigations for noise
sources onto electronic components and mechanical com-
ponents with high Q-factor resonances, and eliminating,
for example, mechanical components with damped me-
chanical resonances.
In this report, we describe tools and methods used
for data quality investigations relevant to long-duration
searches, and provide examples of issues faced in the
first two Advanced LIGO observing runs, O1 and O2.
The paper is organized as follows: section II summarizes
the effects that noise has on the searches for persistent
GWs; section III briefly introduces LIGO data and noise
sources; section IV gives examples of different noise cou-
pling mechanisms to the GW channel; section V summa-
rizes data analysis tools used for noise characterization;
section VI presents results from noise sources that were
investigated and mitigated during O1 and O2; and sec-
tion VII describes the procedures used to generate line
lists for vetoing noise outliers.
Finally, we note that all of the methods presented here
can be applied to both LIGO and Virgo detectors. We
will focus, however, on data quality applied to the LIGO
detectors only, as, at the time of this writing, there is
significantly more Advanced LIGO observational data,
which is needed for persistent GW searches.
II. EFFECTS OF NOISE ON CW AND SGWB
SEARCHES
Spectral artifacts can degrade analyses that search for
long-duration signals in different ways. Artifacts can lead
search pipelines to return spurious outliers, which require
laborious follow-up. Furthermore, if there is a putative
GW signal at a frequency corresponding or nearby to a
spectral artifact, then the signal power is obscured. For
those analyses that rely on combining data from different
detectors (e.g. cross-correlation or coherently combining
data), then detection of signals overlapping with common
detector artifacts may be impossible. On the other hand,
some searches may be able to cope with an artifact if it
occurs in just one detector.
Continuous GWs from spinning neutron stars are
nearly monochromatic, with nearly constant signal fre-
quency in the Solar System barycenter. When projected
into the frame of a detector located on Earth, the sig-
nal is Doppler shifted into many frequency bins. Con-
versely, a narrow, stationary spectral artifact in the de-
tector frame will impact many frequency bins when data
is projected into the frame of the Solar System barycen-
ter. For searches of a signal from a known pulsar with
a given ephemeris, the impact of these artifacts is less
than the impact on an all-sky search for unknown neu-
tron stars (which may also be located in a binary sys-
tem). In extreme cases, an all-sky search may be blind
to a wide region of parameter space for a particular fre-
quency range.
Searches for a stochastic GW background rely on cross-
correlating GW strain channel data from multiple de-
tectors and looking for excess power.
Excess cross-
correlation requires a stable phase between the two chan-
nels at a given frequency, and, thus, many single-detector
artifacts are not found in the cross-correlation analysis.
Correlated noise that causes excess power in the cross-
correlation analysis, however, is excised from the analysis
entirely by setting that frequency bin to zero before inte-
gration in the case of the standard search for a broadband
SGWB. This reduces the search sensitivity by a factor
N
b
N
a
where
N
b
is the number of frequency bins before
notching and
N
a
is the number of frequency bins after
notching. In directed, narrow-band searches [42] we do
not search for GWs at frequencies of known instrumental
lines.
For both CW and SGWB searches, lists of known
instrumental artifacts are created following the end of
an observing run (further details are provided in sec-
tion VII). Then, depending on the search, these lists are
used to: 1) clean the data before analysis by removing
the affected data in the Fourier domain and replacing it
with Gaussian noise measured in the nearby frequency
bins; 2) avoid specific frequencies in analyses that are
impacted by the artifacts; or 3) reject outliers that are
clearly caused by the detector artifacts. This lets the
analysis focus computational resources on regions of pa-
rameter space that are not degraded by spectral features.
If a search pipeline returns a signal candidate which does
not coincide with any known artifact, more detailed in-
vestigations are needed in order to assert that the signal
cannot be produced by an artifact.
4
HAM1
HAM2
HAM3
BSC2
BSC1
BSC3
HAM4
HAM5
BSC4
BSC5
HAM6
FIG. 1. Locations of most auxiliary sensors at LIGO Liv-
ingston Observatory (LHO shares a similar layout). The gray
dashed lines separate the End X and End Y Stations, which
are located at the end of the 4 km arms, from the Corner Sta-
tion building, located at the vertex of the detector. All sta-
tions contain an electronics room (encased by purple points
in the diagram), where the computers that control the inter-
ferometer are housed.
III. LIGO DATA AND NOISE SOURCES FOR
SEARCHES OF PERSISTENT GRAVITATIONAL
WAVES
The first Advanced LIGO observing run (O1) took
place between 12 September 2015 and 12 January 2016,
while the second Advanced LIGO observing run (O2)
took place between 30 November 2016 and 25 August
2017. The Advanced LIGO detectors are located in Han-
ford, Washington (H1), and Livingston, Louisiana (L1).
The LIGO detectors are dual-recycled Michelson inter-
ferometers with Fabry-Perot arm cavities of
4 km (see
[29] for a review of the Advanced LIGO detectors config-
uration).
LIGO detector data is typically characterized as sta-
tionary, Gaussian noise, but non-Gaussian detector ar-
tifacts are also present in LIGO data, e.g., occasional,
short-duration transients (“glitches”) and long-duration
narrow lines. Searches for transient GW signals will usu-
ally avoid analyzing times when a glitch occurs, while
searches for persistent GW signals avoid analyzing data
in frequency bands where narrow lines are present. This
enables either type of search to consider the detector
noise data to be essentially Gaussian.
While most lines in detector data are stationary, some
of the lines have time-varying behavior (often called wan-
dering lines), which can degrade detector sensitivity over
a larger range of frequencies and increase the difficulty
of distinguishing these artifacts from astrophysical sig-
nals when searching for a persistent signal from differ-
ent sky locations. Some lines occur in a distinct pat-
tern known as a
comb
, with even-spacing in frequency
between each tooth (each single line) of the comb. Tooth
frequencies are given by
f
n
=
f
o
+
n
δf
, where
f
o
is the
offset (from 0 Hz) of the comb,
δf
is the spacing, and
n
is an integer. These combs are associated with linear
or non-linear coupling of non-sinusoidal sources or with
non-linear coupling of sinusoidal sources. A comb can
also be recognized by a common time-dependent behav-
ior of the teeth in the comb. The Fourier coefficients of
a comb in the frequency domain can be used to uncover
the time-domain waveform and help identify the source
of the comb.
Lines and combs can have time-dependent behavior as
the configuration of the detector changes, especially dur-
ing periods of commissioning and maintenance. Some
lines and combs have high amplitude and can be iden-
tified using only a short amount of data. Others have
low amplitude and may only become evident with long
integration time. Long integration time is also useful
to better constrain the central frequency and width of a
given line or to find the spacing of a comb.
A schematic diagram showing locations of vacuum
chambers, main interferometer optics, and most of the
Physical Environment Monitoring (PEM) sensors of the
L1 detector can be seen in figure 1 (H1 has a similar lay-
out). PEM sensors include, for example, accelerometers,
microphones, temperature sensors, magnetometers, seis-
mometers, etc. PEM sensors, particularly magnetome-
ters, are often helpful in determining the causes of nar-
row spectral artifacts because they witness local noise
sources that may couple to the main GW channel, and
the PEM sensors do not witness GW signals (except in
cases of complicated cross-coupling mechanisms, which
can be identified using signal injections). Other auxil-
iary channels may also be useful in the same way.
Some of the lines observed in an amplitude spectral
density of the detector data are well-understood: for ex-
ample, 60 Hz power mains, mechanical resonances of mir-
ror suspensions known as “violin modes” (see figure 2),
calibration lines, and simulated GW signals known as
“hardware injections”. Other lines are less understood
and require considerable investigation to determine their
nature.
The majority of instrumental lines that degrade CW
searches have Q-factors in excess of
10
3
.
This is,
in part, because the astrophysical sources targeted by
these searches have high intrinsic Q-factors, and Doppler
broadening caused the Earth’s orbital velocity does not
decrease the Q-factor to less than
10
4
.
Similarly, the instrumental lines that have produced
correlations between sites, degrading searches for SGWB,
have also had high Q-factors. This is because the correla-
tions are produced not by single sources affecting both of
the widely-separated sites, but rather by similar sources
at each site that are correlated only because they produce
5
FIG. 2. Noise-weighted averaged ASD showing the first har-
monic violin mode region for H1 (red trace) and L1 (blue
trace) for the O1 observing run.
signals at the same, or nearly the same, frequency. Some
correlated lines are due to electronic sources at each site
that are set to the same frequency, controlled by a single
clock (GPS), which also controls the timing of the data
acquisition systems. These lines have Q-factors that are,
in principle, infinite. When the frequencies are not ex-
actly the same at each of the sites, the maximum width
of the instrumental lines that can produce correlations is
associated with the duration of the data segments used in
the cross-correlations and the line amplitude. The typ-
ical length of Fourier-transformed data segments is 60 s
long and the lowest Q-factor lines that have produced
inter-site correlations are the power mains-related lines
with Q-factors of
10
3
(the LIGO sites are on different
power grids that are not synchronized).
The primary source of lines with sufficiently high Q-
factors degrading both CW and SGWB analyses are pro-
cesses controlled by electronic clocks or oscillators. This
includes digital processes, analog electronics, and me-
chanical processes controlled by electronic clocks, e.g.,
stepper motors. Most mechanical systems do not have Q-
factors above 10
3
and so do not directly contaminate the
searches by causing additional outliers, but instead de-
grade the sensitivity of these searches. The main excep-
tions are mechanical systems that are designed to have
high Q-factors in order to concentrate noise in a narrow
frequency band, like the “violin” suspension modes.
Monitoring the frequencies associated with electronic
systems is thus the main way we detect the sources of
problematic instrumental lines. Monitoring each individ-
ual electronic component in the complex electronic sys-
tem of LIGO would be difficult. Instead, we attempt to
monitor multiple electronic systems at once, using flux-
gate magnetometers (Bartington Mag-03 series, with sen-
sitivities of about 5
×
10
12
T). The magnetometers are
placed in the experimental areas and especially in im-
FIG. 3. Method of monitoring electronic components and ca-
bles for frequencies of instrumental lines found in the data.
A Bartington fluxgate magnetometer (Mag-03 MCES100) is
mounted on the horizontal white PVC pipe in the back of an
electronics rack containing electronics that control the posi-
tion of important optics. If the magnetometer detects fields
from currents varying at the same frequency as an instru-
mental line, the source of the line may be in the vicinity. In
addition to helping with searches for sources of line artifacts,
the magnetometer can indicate that a spectral line is not as-
trophysical in origin.
portant electronics racks in the electronics rooms (see
figure 3). These magnetometers can detect even low-
amplitude periodic currents controlled by oscillators and
clocks that can produce high Q-factor line artifacts (de-
tecting as low as 5
×
10
5
A at 1 m from long wires or
traces).
The process of addressing lines or combs typically pro-
ceeds in three steps: identification of noise in the GW
strain channel, data analysis to determine properties of
the noise (precise frequency, other sensors that may wit-
ness the noise, start or end times, etc.) which may sug-
gest a cause, and on-site investigations or interventions
to mitigate the noise at its source (more details are given
in section VI). This process is often iterative and exper-
imental. Work on site is limited by available time, and
also by the risk of interventions creating new problems,
so noise sources are typically prioritized for follow-up by
their strength, pervasiveness (number of bins contam-
inated), and the ease of addressing the most probable
cause of the noise. Lines which are not identified or can-
not be mitigated during an observing run are cataloged
afterward; this is not ideal, but it does aid searches in
cleaning data and rejecting outliers.
Mitigation efforts can prove challenging.
In many
cases, low-level spectral artifacts and combs are not vis-
ible in short-duration Fourier transforms. Only by per-
forming averages over many days to weeks of data, do
these features become obvious; hence it can take of order
days to weeks of new data collection to determine if a
6
mitigation attempt has improved the data or not. Unin-
tended configuration changes that lead to line generation
can also take time to appear, be tracked down and miti-
gated. As a result, significant epochs of a data run can be
badly contaminated in some spectral bands, even when
those bands are relatively clean at the start of the run.
As can be seen in figure 4, the amplitude spectral den-
sity (ASD) of L1 and H1 exhibit different line artifacts
and have somewhat different noise floors, explained in
part by different configuration choices and by different
environmental influences [30]. As a result, the couplings
and the noise sources are different, and the lines and
combs that need to be followed and eliminated differ be-
tween the detectors, although some common artifacts can
be studied jointly. This figure also shows the improve-
ment in data quality for long-duration searches from O1
data to O2 data, because of the investigation and mitiga-
tion activities described in section VI. We show the spec-
trum only between 20-2000 Hz, over which the searches
for persistent GW are typically performed.
IV. NOISE COUPLING MECHANISMS TO THE
GRAVITATIONAL WAVE CHANNEL
A. Coupling through shared power and grounds
Most of the mitigated lines in Initial and Advanced
LIGO have coupled through shared power supplies. An
electronic component draws current at a particular fre-
quency from a power supply, which results in a small
periodic drop in voltage. If a sensitive piece of electron-
ics, such as an optic actuator driver or analog-to-digital
converter, shares the power supply, the frequency can be
imprinted on a signal controlling alignment of an optic,
for example, and thus causes a coupling to the interferom-
eter light. This imprinting may happen, if, for example,
a gain or offset in the sensitive equipment varies with the
voltage from the power supply. The solution has been to
place the source of the periodic current draw on a sepa-
rate power supply. This has led us to attempt to better
regulate power, and to isolate noise-sensitive electronics
on separate power supplies, but this is sometimes difficult
to do in practice.
Coupling through shared grounds is a similar mecha-
nism. Even when the source of the periodic current draw
is on a separate power supply from the sensitive elec-
tronics, the source may affect the sensitive electronics by
producing periodic voltage variation in shared grounds.
B. Coupling through magnetic or electrostatic
fields
Another common coupling mechanism has been direct
coupling of magnetic fields to sensitive control systems
or signals. For example, we have observed fields from
switching power supplies coupling magnetically to sig-
nals passing through analog-to-digital converters. We
have also observed 60 Hz mains magnetic fields coupling
directly to permanent magnets that are mounted on cer-
tain optics for actuation. However, in Advanced LIGO,
our main magnetic coupling is through cables and con-
nectors. Mitigation efforts have included separating ca-
bles, smaller actuation magnets, electrostatic actuation,
active cancellation, reducing stray fields, and separating
sources and coupling sites. Digital communication sys-
tems, such as those that use Ethernet, are a common
source, but it is not always easy to keep them away from
sensitive systems.
When electrostatic fields are generated inside of the
vacuum chambers, they may couple directly to the test
masses. Electrostatic fields may also couple to control
signals at locations where shielding is imperfect, like
connectors. Investigations have suggested that certain
sources couple through periodically modulated electro-
static fields, although this mechanism has not been un-
equivocally demonstrated.
C. Mechanical coupling
Thermally-excited high Q-factor resonances of the
wires suspending optics have produced problematic lines
for the CW searches by vibrating the suspended optics,
which causes modulation of interferometer light, and thus
couples optically to the GW strain channel. The precise
frequencies of secondary suspensions may not be known
in advance. Most other mechanical components are low
Q-factor by design, and the broad lines that they produce
typically only degrade search sensitivity for CW signals.
Mechanical systems that are controlled by clocks, like
stepping motors or some fans, might have Q-factors that
are high enough to be problematic, but these have not
been among the sources that we have found.
D. Data acquisition artifacts and non-linear
coupling
We have observed lines and combs produced by alias-
ing of high-frequency spectral artifacts, as well as arti-
facts from digital-to-analog converters. Additionally, we
have observed inter-modulation products between lines
of known or unknown sources during certain periods of
data collection. It is also likely that we have observed
combs produced by occasional errors in transmission of
digitized data within the data acquisition system. The
fundamental frequency of the comb is determined by the
frequency (e.g., 16 Hz) of a process associated with the
error.
7
(a)
(b)
(c)
(d)
FIG. 4. Average amplitude spectral density plots for the L1 (plots (a) and (b)) and H1 (plots (c) and (d)) detectors during
O1 (red trace) and O2 (blue trace). Each individual amplitude spectral density that contributes to the average is weighted by
the inverse square of its running median, so that those spectra with degraded sensitivity (higher amplitude spectral density)
are de-weighted (contributing less) in the final average. (a) and (c): data in the most sensitive frequency band of the LIGO
detectors 20 Hz–2 kHz. (b) and (d): data in the low frequency region from 20 Hz to 120 Hz.
V. DATA ANALYSIS TOOLS
In this section we briefly describe some data analysis
tools used to monitor and analyze the data quality for
persistent gravitational wave searches.
A. Fscan
Fscan is a tool that finds and monitors spectral
lines [48]. It uses data from the GW strain channel
and hundreds of auxiliary channels for each detector,
and it produces “Short Fourier Transforms” (SFTs) of
1800-s-long data segments. Fscan produces two differ-
ent types of graphs: it averages the daily SFTs (with
a maximum of 48 SFTs) to produce normalized power
spectra in bands of default 100-Hz width and frequency
binning of 1/1800 Hz for each channel, and it produces
spectrograms with averaging of adjacent frequency bins
(default bin resolution of 0.1 Hz). In the absence of non-
Gaussian artifacts, the normalized spectra should be flat
with random fluctuations about an expectation value of
one, where the underlying statistical distribution would
be
χ
2
with a number of degrees of freedom equal to twice
the number of SFTs used to construct the spectra. Fig-
ure 5 shows an example of these two types of plots. Thou-
8
sands of such graphs are generated automatically each
day for each observatory from the GW strain channel
and auxiliary channels, to provide a reference archive for
line investigations.
In addition, the strain channel SFTs are used to pro-
duce (unwhitened) inverse-noise-weighted spectral aver-
ages for each day and cumulative from the start of the run
through that day. The inverse noise weighting is meant
to mimic the weightings used in many CW searches [40],
which weight more heavily those time spans with better
sensitivity. Comparing such spectral averages with arith-
metic averages also allows rapid identification of non-
stationary line artifacts.
B. FineTooth
FineTooth is a set of tools to help identify combs and
monitor them over time. It is comprised of a plotting
tool, a tracker for known combs, and a comb finding
tool. The plotting tool creates interactive browser-based
plots using the Python library Bokeh, allowing the user
to overlay combs and lines and easily explore spectral
features, as shown in figure 6. The tracker accepts a list
of known combs and a list of channels, and then draws
from Fscan data to create plots showing the historical
strength of each comb in each channel. The comb find-
ing tool searches for common spacings between peaks of
comparable heights, generating a list of potential comb
candidates to be vetted by the user.
During observing runs, the FineTooth tracker is run
daily on a series of magnetometer channels which typ-
ically witness noise in nearby electronics, as well as on
daily and run-cumulative spectra from the GW strain
channel, providing a summary page for data quality
checks and a tool for rapid investigation of specific combs.
The comb finding and plotting tools are also used to pro-
vide an alert for new combs appearing in the cumulative
spectrum mid-run, and to aid in comb identification for
the purpose of generating vetted noise line lists.
C. NoEMi
NoEMi (Noise frequency Event Miner) is a tool used
for line monitoring and as a line database [31]. It runs
daily and weekly, using data from the GW strain channel
and several auxiliary channels, calculating FFTs with 1
mHz resolution. It creates time-frequency diagrams from
the peaks found in the spectra; the program also calcu-
lates the persistency of the lines (number of peaks in
that frequency bin divided by the number of FFTs) and
their critical ratio (difference between the peak ampli-
tude and the mean value of the spectrum, divided by
the spectrum standard deviation). The persistency helps
to identify loud stationary lines, while the critical ratio
helps to identify non-stationary lines lacking persistency.
NoEMi can provide the starting and end times of a
line in the data. It can also follow wandering lines, by al-
lowing some change in frequency between different time
periods. NoEMi looks for coincidences (lines with the
same frequency) between the GW channel and the other
channels, calculating a value between 0 and 1 to quantify
the probability of coincidence for each different auxiliary
channel. This automated coincidence monitoring is espe-
cially valuable when searching for causes of line artifacts
seen in the GW strain channel.
D. Coherence
Searches for an SGWB are done by cross-correlating
the strain channel data in two detectors in the frequency
domain [25, 44, 46, 47]. Depending on the source model
considered, SGWB searches can be
broadband
, where the
signal is spread over a range of frequencies, or
narrow-
band
, where the signal is concentrated in a narrow fre-
quency band. Additionally, SGWB searches can either
target specific sky directions using the time-delay be-
tween detectors, or integrate over a range of physical time
delays assuming the source is isotropically distributed or
otherwise extended on the sky.
If there is a source of noise that is coherent between
the two detectors then it will show up as an excess in the
cross-correlation. We must therefore cross-check our GW
data streams with local environmental channels to verify
that any excess in our cross-correlation statistic is not
due to a local source of noise. We do this by calculating
the coherence between a GW data stream and many local
environmental monitoring channels. We also monitor the
coherence between our two GW data streams with no
phase shifts.
We define the coherence as the normalized product of
the Fourier transform of two data channels, ̃
s
1
(
f
) and
̃
s
2
(
f
):
C
(
f
) =
|
̃
s
1
(
f
) ̃
s
2
(
f
)
|
2
|
̃
s
1
(
f
)
|
2
|
̃
s
2
(
f
)
|
2
.
If the detector outputs ̃
s
1
,
2
(
f
i
) are uncorrelated Gaussian
random variables, then the coherence follows an exponen-
tial distribution
P
(
C
) =
Ne
CN
,
where
N
=
Tδf
is the number of time segments used to
compute the coherence,
T
is the observation time, and
δf
is the frequency bin width. Frequency bins with a
large coherence between two detectors can be identified
by looking at outliers of a histogram of coherences.
1. Coherence between strain data of two GW detectors
The coherence spectrum is monitored between the
two spatially separated LIGO detectors, and any ex-
9
(a)
(b)
FIG. 5. Typical plots produced by Fscan: (a) shows a spectrogram of one day (23 April 2017) of Hanford strain data (with
color-coded amplitude); (b) shows the corresponding daily averaged normalized power versus the frequency.
FIG. 6. A screenshot showing the comb plotting feature of FineTooth, on a run-averaged spectrum from Hanford in O2.
cess in this spectrum at individual frequencies is fol-
lowed up. Typically, we monitor the time-integrated co-
herence spectrum on day, week, month, and “full run”
time-scales. This allows us to try to narrow down spe-
cific times when inter-site coherence between GW chan-
nels is higher. Any loud, narrow frequency lines is also
followed up. The follow-up is done by searching for a
similar excess coherence at the same frequency in the co-
herence spectrum of a GW strain data channel with a
local environmental monitor of the same detector. Any
excess coherence between a strain data channel and a
local environmental monitor that is expected to be in-
dependent of the strain data channel is enough to sug-
gest that the inter-site coherence is likely caused by a
non-astrophysical source of noise. In figure 7, we show
a coherence spectrum made from computing the coher-
ence over all of O1 between the Hanford and Livingston
strain data channels. We also show the distribution of
coherences for all 1 mHz bins in the band 20-200 Hz.