The et-o3-anchored-v1 glitch population¶

A registered glitch population fixes what a campaign injects: two transient classes, their morphologies, their rates, their amplitude laws, how they are shared across a network, and how their arrival times are drawn. Every arm of the campaign resolves the same name and pins the same digest, so "the arms used the same glitch model" is something the run stamps can be checked for rather than something the campaign asserts.

This page is the record of where each registered number comes from. Numbers that are measurements say what was measured, over what data, and what they were cross-checked against. Numbers that are not measurements are named as such, here and in the population's own unanchored field, which travels with it.

Everything below is reproduced by

uv run python scripts/measure_glitch_population_anchors.py

which downloads about 560 MB on a cold cache and prints every number quoted on this page. --write-summary also rewrites src/gwmock_noise/data/glitch_population/o3_reference_summary.json, the small summary the package bundles; tests/test_glitch_population.py asserts the registered constants against that summary, through the harness's own estimator and rounding rule, so the page, the printout and the assertions cannot drift apart.

The two classes, and why these two¶

They are two classes rather than ten because they span the two ways a transient population hurts a coincidence- or coherence-based search, and the pair is what the consuming campaign needs. The short class is loud, broadband, and resembles a high-mass compact-binary merger; it fires in one channel at a time, so it enters a search's background only through accidental coincidence. The longer class is low-frequency, lasts seconds, and — on interferometers sharing a site — is expected to fire in all of them at once, which is exactly the case a coincidence veto and a null stream are built on the assumption of not seeing.

The two rates are rates of different processes, and conflating them is the easiest way to be wrong by the number of interferometers. The short class runs one Poisson process per interferometer, so its rate is what each of them sees. The longer class runs one process for the whole network, so its rate is the rate of shared events; each interferometer sees that times the participation probability. At the registered participation of 1.0 the two coincide, which is one reason that value was registered.

Source data¶

Two published products, and nothing else.

The classifications. Glanzer et al., Gravity Spy Machine Learning Classifications of LIGO Glitches from Observing Runs O1, O2, O3a, and O3b, Zenodo, doi:10.5281/zenodo.5649212 (v1.0.0, 2021-11-07 — the only version). One row per Omicron trigger, carrying its SNR, duration, peak frequency, central frequency, bandwidth and classified morphology. The four O3 files are read and the selection is SNR > 7.5 — the threshold the set is built on — and machine-learning confidence > 0.9, the fiducial threshold the companion paper uses.

The livetime. The Gravitational Wave Open Science Center's O3 observing segments (H1_DATA, L1_DATA), summed. This is observing time, which is what a count of triggers must be divided by; using calendar time instead would understate every rate on this page by about a third.

Anchor on the livetime. The duty cycles implied by these sums are independently published: Davis et al., LIGO detector characterization in the second and third observing runs, Class. Quantum Grav. 38, 135014 (2021) (arXiv:2101.11673), Table 1, quotes 71 % and 76 % for H1 and L1 in O3a and 79 % for both in O3b, against 71.0 %, 75.6 %, 78.4 % and 77.2 % here. Three of the four agree to better than a percentage point; the fourth (L1 in O3b) is 1.8 points low, which is the size of the difference between the published analysis-ready definition and the GWOSC _DATA flag, and it moves the L1 rates below by under 2 %.

Where the classifications differ from the paper's own table. Glanzer et al. 2023, Data quality up to the third observing run of Advanced LIGO: Gravity Spy glitch classifications, Class. Quantum Grav. 40, 065004 (arXiv:2208.12849), Table 1 gives per-class O3 counts from a later classification pass than the data release above. The fractions agree to a fraction of a percentage point — Scattered Light 46.1 % here against "about 47 %" for Hanford, Fast Scattering 27.6 % against "approximately 27 %" and Tomte 19.2 % against "approximately 19 %" for Livingston — so the two describe the same population. The counts differ by up to 20 % (Blip at Hanford: 7,457 here, 6,020 there; Scattered Light at Hanford: 68,175 here, 57,118 there), and the rates below inherit that as a systematic. The measurement is made on the data release rather than on the table because the release is what can be re-derived.

1. The rates¶

Counts over pooled observing time, per class:

The site-to-site spread is the honest uncertainty, and it is larger than anything else on this page: a factor 1.62 for Blip and 1.37 for Scattered Light between two instruments of the same design, in the same run, analysed by the same pipeline. A single pooled number is a run average over two sites, not a prediction for a third.

Anchor on the short class's rate. Cabero et al., Blip glitches in Advanced LIGO data, Class. Quantum Grav. 36, 155010 (2019) (arXiv:1901.05093) measured the same class with a completely different method — a reduced-template-bank PyCBC search with manual morphological vetting, rather than a classifier on spectrograms — over a different pair of runs. They report "approximately two blip glitches per hour of data": a median of 39/day at Hanford and 31/day at Livingston in O1, rising to 47/day and 48/day in O2. The measurement here is 30.4/day and 18.8/day in O3. The two agree in order of magnitude and in the statement that the two sites are within a factor of about two of each other; the O3 figure is lower by a factor 1.3–2.6, across a change of pipeline, a change of run and a documented improvement in the instruments. This is a cross-check, not a calibration: no number here was adjusted to match it.

Anchor on the longer class's rate. Glanzer et al. 2023 report Scattered Light as the most common O3 glitch class at Hanford at "about 47 %" of all classified glitches, and about 23 % at Livingston; the measurement here gives 46.1 % and 22.7 %. Their Table 1 counts, over the same livetime, give 2.70e-3 Hz and 2.17e-3 Hz against 3.22e-3 Hz and 2.34e-3 Hz here — the count difference described above.

2. The amplitude laws¶

Both classes' loudness is registered once, as a truncated power law in optimal SNR against the configured noise curve. The lognormal amplitude multiplier every glitch model also carries is left at mean 1.0 and standard deviation 0.0, deliberately: a second scatter on top of the SNR distribution would encode the same physical quantity in two places, and the realized SNR would then be neither the drawn target nor anything measured.

The threshold is 7.5 for both, because that is where the reference sample starts. A power law fitted above a threshold describes the tail above it and says nothing about the bulk below; the model reproduces the tail and continues the same law down to the threshold.

Tail indices¶

The registered exponent is not the Hill index, and the reason is worth reading. The Hill estimator is the maximum-likelihood exponent of an untruncated Pareto tail, and it is the natural thing to reach for — the package's own PowerLawSNRDistribution takes the survival exponent in exactly the convention Hill reports it. But the registered law is truncated, at the largest SNR the reference sample contains, and a truncated law drawn with the Hill index of a sample does not reproduce that sample: cutting the far tail off raises the Hill index of what comes out. Drawn with α = 1.3763 and cut at 344.38, the short class's own draws have a Hill index of 1.4149 — 2.8 % steeper than the population the anchor exists to reproduce, with nothing in the configuration to say so.

The exponent is therefore estimated under the model that will actually be sampled: the maximum-likelihood exponent of a power law truncated at the sample maximum, which is itself the maximum-likelihood estimate of the truncation. That gives 1.3333 and 1.4184, and the third column above is the closed-form Hill index of draws from those laws — equal to the measurement, which is the property the choice was made for. tests/test_glitch_population.py asserts it, and asserts that the naive choice would not have satisfied it.

The correction is 3.1 % for the short class and 1.2 % for the longer one, both in the direction of a heavier registered tail than the Hill index alone would have given. It is small next to the factor-1.6 site-to-site rate spread, and it is not small next to the 0.9 % statistical error on the short class's own index.

Neither class is a pure power law from 7.5 upward, and the page says so rather than hiding it behind one number. Re-estimating the Hill index at higher thresholds:

The Blip index steepens monotonically — the class has a lighter tail than a single power law from 7.5 would give, so the registered model over-produces its loudest events relative to the measurement. Scattered Light is flat to about SNR 20 and steepens beyond. A campaign whose conclusion turns on the far tail of the short class should read the registered index as an upper bound on how heavy that tail is; this is much the larger of the two effects on this page, and unlike the truncation correction it cannot be removed without replacing the functional form.

The same statement from the other side: the registered law's median is 12.56 for Blip against a measured median of 13.41 (6 % low), and 12.21 for Scattered Light against a measured 11.97 (2 % high). The laws are fitted to the tail, and land within 10 % of the bulk as a by-product, not by construction.

The loud-tail truncation¶

Each is the largest SNR the reference sample contains.

A truncation is required, not cosmetic. Both indices are below 2, so the untruncated distributions have infinite variance, and the Blip index is close enough to 1 that the mean is barely finite. Left untruncated, a long enough run draws an SNR no instrument could produce, and a background estimate built on it is dominated by an event that cannot happen.

What the truncation is, and is not. It is a property of a finite observation: the largest of ~1.2e4 and ~1.2e5 draws from a heavy tail, which grows with observing time roughly as T^(1/α). It is not an instrumental ceiling, and a campaign running far longer than 497 days of pooled observing time is extrapolating past where this number was measured.

An alternative published truncation exists for the short class, and it was not adopted. Cabero et al. imposed ρ ≤ 150 on their blip search. That is a search selection — imposed "to avoid loud noise events that are clearly not blips but contain similar morphology near loud, long-duration noise transients" — and not a statement that the class never exceeds it; the reference sample contains 8 events above it, classified as Blip with confidence above 0.9. Adopting 150 would discard 0.066 % of the class by number and a larger fraction of its contribution to a loud-tail background. The table above carries the number so the choice can be revisited with the cost in hand.

3. The frequency supports¶

The short class: 75–780 Hz, the 1^st and 99^th percentiles of the measured Omicron peak frequency, rounded outward to two significant figures (75.26 and 779.06). The model is a Gaussian-windowed white-noise burst whose spectrum is flat across its configured band, so the band is the support.

That band is 705 Hz wide and centred on 427 Hz. The Omicron columns describing the same quantity, which are not what the edges were set from, give a median bandwidth of 913 Hz and a median central frequency of 536 Hz. The model is therefore about 25 % narrower and 20 % lower than the measured tiles, which is the size of the mismatch between a flat-in-band burst and a heavy-tailed distribution of tile shapes.

Cabero et al. describe the class as having "a large frequency bandwidth, O(100) Hz", and their example time–frequency representation spans 10 Hz to 500 Hz. Both are consistent with the band above; neither is precise enough to have set it.

The longer class: 10–120 Hz, taken from Soni et al., Reducing scattered light in LIGO's third observing run, Class. Quantum Grav. 38, 025016 (2021) (arXiv:2007.14876), which states that slow scattering "creates 'scatter shelves' in the frequency band 10 Hz to 120 Hz in h(t) spectra" and that scattered light "impacts the gravitational strain in 10 Hz to 120 Hz frequency band".

This is the one frequency edge on the page that is published rather than measured, and the reason is that the Omicron columns cannot supply it for this class. The peak frequency can — it is the apex of the arch, measured median 25.97 Hz, registered as 26 Hz, with 1^st and 99^th percentiles at 18.50 and 40.28 Hz — but the central frequency and bandwidth columns give medians of 2021 Hz and 4008 Hz for a class whose apex sits at 26 Hz. Those columns describe the whitened tiling rather than the arch, and rounding them would produce a band with no relation to the morphology. The published shelf band comfortably contains the measured apex distribution, which is the consistency check between the two kinds of anchor, and tests/test_glitch_population.py asserts it.

4. The duration¶

The longer class: 1.75 s, the median measured Omicron duration (quartiles 1.44 s and 2.50 s; 99^th percentile 10.8 s).

Two published cross-checks, and they disagree with each other by more than either disagrees with this:

• Glanzer et al. 2023 describe Scattered Light as "longer-duration (∼2.0–2.5 s) arches" — 14 % to 43 % above the median here.
• Soni et al. Table 1 gives a median duration of 3.2 s for slow scattering, 83 % above. Their selection is different: Livingston only, O3a only, SNR above 10 rather than 7.5, and restricted to the slow population by arch frequency (below 0.2 Hz), which is 40 % of all scattering at that site. Restricting to a louder, longer sub-population raises the median, which is the direction the difference goes.

The short class: 0.01 s, taken from Cabero et al.'s definition of the class as "a very short duration transient, O(10) ms". It is not the measured Omicron duration, whose median is 0.125 s, and the reason is that the Omicron duration is a tile length set by the search's Q at the trigger's frequency, not the transient's own width — for this class the 1^st and 5^th percentiles are 7.8 ms and 11.7 ms, which is where the intrinsic scale shows through. The registered width is the FWHM of the model's Gaussian envelope, which is a statement about the morphology, so the morphological source is the right one.

5. What is not anchored¶

Eight things. Each is carried in the population's unanchored field so that a consumer reading a pinned population reads them with it, and each is a statement the consumer is making on its own authority.

1. That any glitch class is coherent across a network at all. This is the population's central extrapolation. The one direct measurement of the question says the opposite for separated sites: Cabero et al. Table 1 finds zero coincidences between Hanford and Livingston blips within the ±15 ms window an astrophysical signal can occupy, over both O1 and O2, with the coincidences at wider windows matching the accidental expectation exactly (19.7 expected, 24 found at ±10 s in O1; 83.8 expected, 65 found in O2). The argument for coherence in a co-located array is that its interferometers share a site, a vacuum system and a seismic environment, and that the scattering mechanism is environmental — Soni et al. observe "similar harmonic series of arches at both LHO and LLO", which says the mechanism is common to both sites but says nothing about whether two instruments at one site glitch together. No measurement of transient coincidence between co-located interferometers was found.
2. The participation probability of 1.0, and therefore the network event rate, which is the per-channel rate divided by it. One is the strongest coherence the model can express, the hardest case for a coincidence veto or a null stream, and the only value that does not introduce a second unmeasured number — at one, the network rate and the measured per-channel rate coincide. It is a choice, not a measurement, and a campaign should sweep it.
3. The zero spread of the per-interferometer amplitude ratio — that a common-mode transient couples equally into every interferometer of a site.
4. The zero inter-interferometer arrival-time offset. Co-location makes it small for an environmental source; nothing measured here bounds it.
5. Transferring a rate measured on Advanced LIGO to a different instrument, site and band. The rates are what those two detectors did, not a forecast of what another one will do — and the factor-1.6 spread between two instruments of the same design, in the same run, is the lower bound on how wrong that transfer can be.
6. Identifying the target SNR across noise curves and bands. The measured SNRs are Omicron's, defined against the O3 Advanced LIGO spectra over the band Omicron searched; the model calibrates an optimal SNR against the configured psd_file over the configured band. These are different inner products. The sampler reproduces whatever distribution it is given and cannot tell whether that identification is the one intended; read the result as "a population as loud, relative to this instrument's own noise, as the O3 population was relative to O3 noise".
7. The homogeneous Poisson arrival process. Both measured rates are strongly non-stationary. Soni et al. show slow scattering tracking microseismic ground motion, particularly on named days of high ground motion; Glanzer et al. show the hourly rate of Scattered Light varying by more than an order of magnitude across O3 and dropping after reaction-chain tracking was introduced, and Fast Scattering varying with the day of the week. A constant rate reproduces the run average and not the clustering, and a search's background is sensitive to clustering.
8. The truncation of each class's loud tail at the largest observed SNR, as discussed in §2.

6. Decomposing an expected count¶

GlitchPopulation.expected_counts returns the factors, not only the product, so a total that comes out wrong is traceable to the factor that is wrong:

expected = rate_hz × livetime_seconds × participation × selection

For a 2048 s segment of a three-interferometer network, with a recording threshold at SNR 10:

The selection factor is the survival function of the class's own truncated power law at the threshold, read off in closed form rather than estimated from a realization — and, like every other number in the amplitude law, evaluated at the registered exponents of §2 rather than at the Hill indices they replaced. That distinction is worth a sentence here rather than only in §2, because a table is what gets copied: the naive exponents give 0.6714 and 0.6609, a 1.2 % and a 0.5 % under-count, which is small next to the site-to-site rate spread and is still exactly the error this page exists to prevent.

test_the_documented_decomposition_is_the_one_the_code_returns parses this table out of this file and asserts every cell against expected_counts(), so the page and the implementation cannot disagree about it again.

Two reminders the table is laid out to make hard to miss. The livetime is observing time; a campaign that writes 2048 s segments at less than 100 % duty must use the analysed total, not the span. And the two rates are rates of different processes, so the network row's rate_hz is not a per-channel rate and multiplying it by the number of interferometers does not give a network total.

7. A stamped example¶

from gwmock_noise import get_glitch_population

population = get_glitch_population("et-o3-anchored-v1")
realization = population.realize(
    detectors=["E1", "E2", "E3"],
    duration=4096.0,
    sampling_frequency=4096.0,
    seed=20260919,
)

realization.stamp identifies the realization, and the field a campaign manifest pins is the digest:

population              et-o3-anchored-v1
schema_version          1.0.0
digest                  sha256:58fb23aef4b4dd0d0038ebdae09da51bdaf638821aca0df97ead74dcbc1dc3b0

The stamp also carries gwmock_noise_version, seed, detectors, duration, sampling_frequency, gps_start and event_count. Record those; do not pin them. gwmock_noise_version in particular is derived from the version-control description and moves on every commit — it read 0.12.0-post.19+c2a061f while this example was first written and something later by the time you run it — so a manifest comparing it for equality would reject a rerun of the same population from a later commit. The digest is what does not move unless the population does.

The realization produces, over this hour and a bit of three-interferometer data:

— 8 distinct network events, every one of them in all three interferometers, and two independent short glitches. Against the decomposition of §6 scaled to 4096 s: 1.16 short events per channel expected against 2 seen across three channels, and 11.4 network events expected against 8. Realized SNRs run from 7.63 to 35.66 and peak strains from 4.2e-23 to 8.2e-23.

An order-of-magnitude anchor on those strains, which nothing else here supplies. For a transient of optimal SNR ρ over a band where the PSD is roughly flat at S, the one-sided inner product gives ∫h² dt = ρ² S / 2. The bundled 10 km cryogenic curve has an amplitude spectral density of 7.5e-25 /√Hz at 26 Hz, so the loudest event here — a ρ = 35.7 arch of 1.75 s with an envelope duty of about a half — should have an RMS strain near 2.0e-23 and a peak a factor of a few above it, around 6e-23. The realization's largest peak is 8.2e-23. The agreement is the check that the SNR calibration, the PSD units and the strain scale are consistent with each other; it is a back-of-envelope, and it would not catch an error smaller than a factor of two.

realization.strain is the glitch contribution with no noise under it, which is what paired arms add so that their glitch content is bit-identical by construction. realization.events is the per-event truth catalogue; rows of a network-coherent event share a network_event_id, which is how the coincident rows are recognised.

The digest is sha256 over the canonical serialization — sorted keys, no insignificant whitespace — of every registered quantity and of every prose field, so a population whose numbers are untouched but whose meaning has been rewritten digests differently. The digest travels between machines only because the noise curve is named (ET_10_full_cryo_psd) rather than pathed; a population built against an absolute path carries that path into its digest and nobody else can reproduce it.