Improving gravitational wave search sensitivity with TIER: Trigger Inference using Extended strain Representation

Ajit Kumar Mehta; Arush Pimpalkar; Barak Zackay; Benjamin Wandelt; Digvijay Wadekar; Emanuele Berti; Javier Roulet; Jonathan Mushkin; Mark Ho-Yeuk Cheung; Matias Zaldarriaga

REVIEW 4 major objections 5 minor 87 references

Improving gravitational wave search sensitivity with TIER: Trigger Inference using Extended strain Representation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A machine-learning layer that reads the ~10 seconds of strain around each candidate can raise gravitational-wave search sensitivity by up to ~20%, mainly for high-mass, unequal-ratio binaries.

desk verdict A genuinely useful, cheap sensitivity boost for GW searches, but the headline ~20% VT gain rests on a signal-class training set that needs stronger validation before the number is trusted. read the letter →

arxiv 2507.08318 v1 pith:DTXINWUQ submitted 2025-07-11 gr-qc astro-ph.IMstat.ML

Digvijay Wadekar , Arush Pimpalkar , Mark Ho-Yeuk Cheung , Benjamin Wandelt , Emanuele Berti , Ajit Kumar Mehta , Tejaswi Venumadhav , Javier Roulet

show 4 more authors

Tousif Islam Barak Zackay Jonathan Mushkin Matias Zaldarriaga

This is my paper

classification gr-qcastro-ph.IMstat.ML PACS 04.30.-w95.55.Ym07.05.Mh

keywords gravitationalwavesmachinelearningsearchpipelinesextendedstraindatasensitivevolume-timerandomforestclassifierbinaryblackholesdetectorglitches

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Search pipelines typically score a gravitational-wave candidate using only a short window of strain around the trigger. This paper argues that the surrounding seconds of data carry independent information about whether the candidate is a real signal or a noise transient, because glitches tend to cluster and live in noisy environments while real mergers do not. The authors build TIER, a machine-learning framework that compresses that extended strain into a handful of summary statistics and learns, for a given candidate, the probability $p_{\rm ext}$ that the candidate is astrophysical based only on the extended data. That probability is multiplied into the pipeline's ranking statistic, and on the IAS-HM search of LIGO-Virgo-Kagra O3 data the re-ranked statistic raises the sensitive volume-time (the standard figure of merit for surveyed space over time) by up to ~20%, concentrated in high-mass, unequal-mass-ratio binaries, while lifting several near-threshold candidates in the pair-instability mass gap and intermediate-mass black hole ranges.

What carries the argument

The object that carries the argument is $p_{\rm ext}$, the calibrated probability that a candidate is a gravitational-wave signal given only the extended strain data, estimated by a random-forest classifier and inserted into the ranking statistic through $L \approx L_{\rm pipeline} \prod_i p^i_{\rm ext}/(1-p^i_{\rm ext})$. The extended data are reduced to sparse summary statistics—separation, SNR, and template ID of the three loudest nearby triggers, the local-PSD sensitivity integral $\int |h_{\rm norm}(f)|^2/\mathrm{PSD}_{\rm local}(f)\,df$, the gap to the nearest excavated 'hole', and the number of noisy bands—so the model is cheap to train and interpretable. The noise class is built from timeslide background triggers and the signal class from simulated uniform times with data-quality cuts, and predictions are made by a three-chunk cross-validation so no candidate is scored by a model trained on its own time chunk.

What would settle it

Run a standard injection-recovery campaign on O3 or O4 data, apply TIER using the simulated-time-trained classifier, and compare the recovered volume-time gain and the $p_{\rm ext}$ calibration with the values reported here; if the gain vanishes or the calibration drifts outside the bootstrap band, the simulated-time approximation is the cause.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that the nonlocal strain environment of a candidate—the roughly ten seconds of data just outside the local matched-filter window—is statistically different for genuine signals and for noise transients, and a calibrated classifier can exploit that difference. TIER trains a random-forest model on six families of features of the extended data: the time separations, SNRs, and template identities of the loudest nearby triggers, the local sensitivity of the chunk to the template bank, the time gap to the nearest excised data 'hole', and the number of noisy frequency bands overlapping the trigger. The classifier output is calibrated to a Bayesian probability and folded into the search ranking statistic as $L \approx L_{\rm pipeline} \prod_i p^i_{\rm ext}/(1-p^i_{\rm ext})$ for each detector. On O3a and O3b data, this changes the false-alarm rates of IAS-HM candidates and increases the recovered sensitive volume-time by up to ~20%, with most of the gain at high total mass and low mass ratio.

Load-bearing premise

The load-bearing premise is that simulated uniform, data-quality-cleaned times are a faithful stand-in for the environment around real gravitational-wave signals; if real signal times differ from those simulated times in ways the two validated features do not capture, the learned probabilities and the reported volume-time gain are miscalibrated.

Editorial extensions

If this is right

TIER can be applied to triggers already produced by any search pipeline without running a new injection campaign, because the signal-class training set is built from simulated uniform times with standard data-quality cuts.
The sensitive volume-time gain of up to ~20% is concentrated precisely in the high-total-mass, low-mass-ratio regime where short signals are most easily mimicked by detector glitches, so the method targets the part of parameter space where current catalogs are sparsest.
Several near-threshold candidates in the pair-instability mass gap and intermediate-mass black hole range increase in $p_{\rm astro}$ and inverse false-alarm rate, which can convert tentative events into cataloged detections.
Feature-importance results show that the properties of the loudest nearby triggers—not the total number of nearby triggers—carry most of the discriminating information, guiding which environmental features future pipelines should record.
Because the ranking statistic is modified, the search selection function changes in the improved parameter region, so population studies using the new catalog must adopt the updated false-alarm estimates.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

One implicit consequence is that the same cheap training recipe could be re-run at every observing run and on every pipeline's triggers, making the noise environment a standard, continuously updated component of ranking statistics rather than a one-off calibration.
A natural next test is to check all six summary features, not just the two shown in the paper's validation figure, against real injection times; the injection catalogs already exist, so this is a low-cost way to establish whether the simulated-time approximation that the paper calls approximate is safe to inherit at higher sensitivity.
The architecture's split between fixed-length and variable-length inputs suggests auxiliary environmental channels, such as seismometer or magnetometer data, could be folded in without changing the ranking-statistic formula, extending the method beyond the strain-only information used here.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core idea is good and the paper is worth engaging with seriously. TIER takes existing pipeline triggers and reweights them using a handful of summary features of the strain environment within ~10 s of the candidate, inserted into the ranking statistic through a likelihood ratio. That is a new combination, and the paper shows it buys something: the feature-importance analysis is the clearest result, demonstrating that properties of the few loudest nearby triggers carry more information than simple trigger counts, which is a real and useful distinction from GstLAL and ARES. The VT computation itself is standard, uses the LVK injection catalog, and includes bootstrap intervals. The time-chunk cross-validation scheme is thoughtful and avoids a real leakage problem. The authors are also honest about their approximations, which makes the paper easier to trust than most ML-in-GW papers.

The soft spot is exactly where the stress-test note lands. The signal class for the classifier is populated with simulated times, not recovered injection times. The paper calls this 'a much cheaper but approximate way,' and Fig. 8 validates only two of the six feature groups, marginally. The random forest can and will use interactions among all six groups, so matching two marginals does not establish that the joint distribution of features at true signal times is the same as at the simulated times. Since the same pext model is applied to the injections used in the VT integral, any bias there propagates directly into the IFARs and into the quoted sensitivity gain. The effect is largest in exactly the high-mass, low-q bins where the 'up to ~20%' appears and where effective injection counts are smallest, so the headline number likely overstates what is robust. That is a load-bearing concern, not a minor one.

The circularity burden is low: nothing is fitted to the VT result, and the paper does not hide its assumptions. The independence assumption in Eq. (1c) is acknowledged and is reasonable as a first cut. The withheld full dataset is annoying but they do ship code and a bank-14 dataset, which is more than typical.

Who is this for? GW search practitioners who want a pipeline-agnostic way to squeeze extra sensitivity in the high-mass, unequal-q regime. It deserves a serious referee. I would send it to review with a clear request: train the signal hypothesis on actual recovered injections (or at least validate the joint feature distribution), and report VT gains only in bins with adequate injection counts. If those revisions hold up, this is a solid methods paper.

Recommendation: accept for peer review with major revision.

Referee Report

4 major / 5 minor

Summary. This paper proposes TIER, a post-processing layer for gravitational-wave search pipelines that uses extended strain data (about 0.1 s to 10 s around a candidate) to reweight the ranking statistic. The authors compress the extended data into a small set of summary features, train a random forest classifier on timeslide background triggers and on simulated uniform times (rather than recovered injection times) as signal examples, calibrate the classifier output, and insert the resulting probability pext into the ranking statistic via Eq. (1c). Using triggers from the IAS-HM O3 search and the LVK injection catalog, they report up to ~20% improvement in sensitive volume time, concentrated at high primary masses and asymmetric mass ratios, and changes in significance for near-threshold candidates, including some in the pair-instability mass gap and IMBH ranges. The paper includes public code, a time-chunk cross-validation scheme, bootstrap confidence intervals for the VT curves, and a comparison with a deep-set classifier.

Significance. If the result holds, this is a practical and useful method: it can be applied on top of existing pipeline outputs, does not require a dedicated injection campaign for training, and targets the high-mass, unequal-mass-ratio region where glitches dominate. The authors deserve credit for providing reproducible code and data products for at least one bank, using the standard Monte Carlo VT integral of Eq. (7), showing bootstrap intervals, and explicitly addressing classifier calibration. The main caveat is that the signal-hypothesis training sample is approximate, and its validation is incomplete; because pext enters the ranking statistic multiplicatively, a bias in the training distribution could directly affect the reported sensitivity gain. The significance is therefore real but contingent on the robustness of the trained pext and on the statistical strength of the VT improvement.

major comments (4)

[Section III A (ii) and Fig. 8] The signal-hypothesis training set is built from uniformly sampled times with data-quality cuts rather than from recovered injection times, and the text explicitly calls this 'a much cheaper but approximate way.' The validation in Fig. 8 checks only two marginal distributions (the local sensitivity integral and log10(SNR^2) of the loudest nearby trigger) among the six feature groups used by the random forest. Because the RF can exploit interactions among all input features, matching these two marginal distributions does not establish that the joint distribution of the signal-hypothesis features matches that at actual recovered injection times. Since pext enters Eq. (1c) multiplicatively and the VT estimate in Eq. (7) counts injections above the IFAR>1 yr threshold, a systematic bias in the signal-hypothesis density can move injections across the threshold; this is especially concerning in the high-m1, low-q bins where the claimed improvements are largest and the number of effective injections is smallest. Please retrain or at least validate pext using the times of recovered injections from the same catalog used in Section V, or otherwise demonstrate that the joint feature distribution matches, and report the resulting VT ratios.
[Eq. (1c) and Section VI] The derivation replaces P(d_nonlocal|d_local,S)/P(d_nonlocal|d_local,N) with pext/(1-pext), where pext is the classifier's signal probability based on nonlocal data only. This step assumes conditional independence of local and nonlocal data given the signal or noise hypothesis, and it also assumes that the classifier's odds equals the appropriate likelihood ratio. The paper acknowledges the independence assumption in the Discussion but does not test it. If local SNR and nearby-trigger properties are correlated (for example, if glitches that produce high local SNR also tend to cluster in time), then Eq. (1c) will mis-rank candidates even if the marginal pext is well calibrated. Please provide a calibration check of the updated ranking statistic on the background timeslides and on injections, for instance by verifying that the empirical IFAR distribution after applying TIER is consistent with a uniform distribution, or otherwise quantify the sensitivity of the reported VT improvements to this approximation.
[Fig. 5 and Eq. (7)] The headline 'up to ~20%' improvement appears to be the maximum of the binned ratios VT_new/VT_old over the m1 and q bins, not an integrated gain. Please report the population-integrated VT improvement under a fiducial astrophysical population, and state for each bin the number of injections above and below the IFAR threshold that enter the ratio. In particular, indicate whether the bootstrap 90% intervals shown in Fig. 5 exclude VT_new/VT_old=1 in the bins where the ratio peaks; if they do not, the claim of a 20% sensitivity gain is not statistically supported and should be qualified accordingly.
[Section III B and Section IV A] Many design choices are set by trial and error: the 0.5-15 s nearby-trigger window, the restriction to the top three nearby triggers, the choice of summary features, the 30 calibration bins in Eq. (4), and the lognormal smoothing width used for the VT curves. Because the same O3 data are used to motivate these choices and to measure the gain, the reported improvement could be inflated by selection on the dataset. Please add a sensitivity analysis varying these hyperparameters and show that the VT ratios and the candidate IFAR changes are stable, or quantify the expected fluctuation from such tuning.

minor comments (5)

[Section IV C] There is a typo: 'similar to the to the k-fold cross-validation technique' should read 'similar to the k-fold cross-validation technique.'
[Section VI] The word 'Futhermore' should be 'Furthermore.'
[Fig. 3] The top panels of Fig. 3 label the signal class as 'GW injections,' but the training signal class is actually composed of simulated times, not injections; please clarify the labeling to avoid confusion, since the VT evaluation injections are distinct from the training samples.
[Section III B] The notation around Eqs. (2) and (3) is confusing: Eq. (2) introduces a ratio PSDglobal/PSDlocal, while Eq. (3) appears to define the square of a related quantity; please define all quantities consistently and state explicitly which combination is passed as the feature.
[Section V] The statement that the method works 'without requiring expensive injection campaigns' is only true for training the classifier; the VT measurement in Eq. (7) still relies on the LVK injection catalog, so the text should be worded more precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TIER's pext is trained on background triggers versus simulated times and evaluated against held-out LVK injections, so the claimed VT gain is not fitted to the target.

full rationale

The derivation chain in Eqs. (1a)-(1c) factorizes the likelihood ratio into the pipeline's local statistic and an extended-data odds factor pext/(1-pext). The ML model for pext is trained on two classes: timeslide background triggers for the noise hypothesis and uniformly sampled times with data-quality cuts for the signal hypothesis. The paper explicitly calls the latter 'a much cheaper but approximate way' (Section III A), and Fig. 8 validates the simulated signal times against recovered injection times using two marginal features. This is an approximation and a calibration risk, not a circularity: pext is never derived from, or fitted to, the VT improvement, the IFAR thresholds, or the candidate significances. The VT computation (Eq. 7) uses the external LVK injection catalog on Zenodo, and the improvements are measured after pext has been fixed. The authors' own IAS-HM pipeline and prior VT estimates (Refs. [8] and [73]) are self-referential in the sense that the demonstration is built on them, but they serve as a testbed and baseline, not as an argument that forces the result. The paper invokes no uniqueness theorem and does not smuggle in the answer via a self-citation chain. Concerns about whether the joint distribution of the six feature groups at simulated times matches true injection times are correctness/calibration concerns, not evidence that the prediction reduces to its inputs by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small set of domain assumptions: Poisson-distributed signal times, representative background triggers from timeslides, and conditional independence between local and nonlocal strain. The model also depends on hand-chosen summary features and training-window parameters. No new physical entities are introduced.

free parameters (6)

Nearby-trigger time window (0.5 s to 15 s) = 0.5 s to 15 s
Chosen to avoid waveform overlap and to define the extended-strain environment (Section III B 1).
Max number of nearby triggers used by RF (top 3) = 3
Fixed input size for the random forest; padding for candidates with fewer triggers (Section IV A).
Calibration histogram bins in Eq (4) = 30
Chosen to balance smoothness and discreteness in the calibrated pext (Section IV A 1).
VT smoothing lognormal sigma^2 = 0.1
Used to bin injection samples in m1 and q when computing the VT curves in Fig. 5 (Eq 8-9).
Summary-statistic feature set = hand-selected
Six features developed with trial-and-error rather than derived from an optimality principle (Section III B).
Time-chunk length for cross-validation = 100 s
Chosen to create independent train/test splits by GPS time (Section IV C).

assumptions (5)

domain assumption GW merger times follow a Poisson process with a fixed rate, so signal times are uniformly distributed over the run.
Used in Section III A to justify generating simulated signal times instead of running an injection campaign.
domain assumption The extended-strain summary statistics at simulated times match those at true GW signal times.
Validated only for two features in Fig 8; if other features or their correlations differ, pext is biased.
domain assumption Background triggers from timeslides are representative of the noise hypothesis, and P(d_nonlocal|dlocal,N) does not depend on glitch type or SNR.
Stated in Section III A; the authors note this assumption may make the method less effective, not biased.
domain assumption The local and nonlocal strain data are conditionally independent for both signal and noise hypotheses.
Assumed in Eq (1b)-(1c); the paper discusses it in Section VI A as a limitation of the approach.
domain assumption The LVK Zenodo injection catalog is an unbiased sample for estimating volume-time.
Used in Section V and Eq (7) to measure VT before and after TIER.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improving gravitational wave search sensitivity with TIER: Trigger Inference using Extended strain Representation." pith.science (2026). https://pith.science/paper/DTXINWUQ

@misc{pith2026250708318,
  author       = {Pith},
  title        = {Pith review of: Improving gravitational wave search sensitivity with TIER: Trigger Inference using Extended strain Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DTXINWUQ}},
  note         = {Machine review of arXiv:2507.08318}
}

abstract

We introduce a machine learning (ML) framework called $\texttt{TIER}$ for improving the sensitivity of gravitational wave search pipelines. Typically, search pipelines only use a small region of strain data in the vicinity of a candidate signal to construct the detection statistic. However, extended strain data ($\sim 10$ s) in the candidate's vicinity can also carry valuable complementary information. We show that this information can be efficiently captured by ML classifier models trained on sparse summary representation/features of the extended data. Our framework is easy to train and can be used with already existing candidates from any search pipeline, and without requiring expensive injection campaigns. Furthermore, the output of our model can be easily integrated into the detection statistic of a search pipeline. Using $\texttt{TIER}$ on triggers from the $\texttt{IAS-HM}$ pipeline, we find up to $\sim 20\%$ improvement in sensitive volume time in LIGO-Virgo-Kagra O3 data, with improvements concentrated in regions of high masses and unequal mass ratios. Applying our framework increases the significance of several near-threshold gravitational-wave candidates, especially in the pair-instability mass gap and intermediate-mass black hole (IMBH) ranges.

Figures

Figures reproduced from arXiv: 2507.08318 by the authors.

**Figure 2.** FIG. 2. We compress the extended strain data into a few summary statistics/features to make the ML models easier to train and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** FIG. 4. We show a few examples with relatively low ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: FIG. 5. Increase in the sensitive volume time (VT) of the search pipeline upon using [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: FIG. 6. Updated version of the new candidate events from [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: FIG. 7. Similar to Fig [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: FIG. 8. We show the distribution of two most important sum [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 9.** Figure 9: FIG. 9. We compare the change in [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: FIG. 10. We described in section [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

87 extracted references · 20 canonical work pages

[1]

We then compute the relative separation of the triggers from the reference candidate, ∆t i ≡t i −t candidate

∆t triggers (separation of nearby triggers from the candidate) – To analyze a candidate trigger at tcandidate, we make a list of nearby triggers (i.e., those falling within 0.5<|t−t candidate|<15 s) from the catalog already collected at each detector in the matched-filtering stage of the search pipeline 2. We then compute the relative separation of the tr...
[2]

SNRs of nearby triggers – We also have infor- mation about the SNR of the nearby triggers in each detector. As our search includes higher har- monics, we use the full harmonic SNR (|ρ 2 HM| ≡ |ρ22|2 +|ρ ⊥ 33|2 +|ρ ⊥ 44|2) of the triggers, whereρ ⊥ 33 andρ ⊥ 44 correspond to the SNR of the orthogo- nalized higher-harmonic templates. We use the rescaled qua...
[3]

In theIASsearches, the tem- plates do not correspond to physical parameters (e.g.,m 1, m2, s1z, s2z), but to singular vector com- ponents (c 0, c1) [67, 71]

Template ID of the nearby triggers – We use infor- mation about the template corresponding to the nearby triggers. In theIASsearches, the tem- plates do not correspond to physical parameters (e.g.,m 1, m2, s1z, s2z), but to singular vector com- ponents (c 0, c1) [67, 71]. Our template ID there- fore corresponds to a unique set of{c 0, c1}values. The logic...
[4]

The LIGO Scientific Collaboration, the Virgo Collabo- ration,et al., arXiv e-prints , arXiv:2108.01045 (2021), arXiv:2108.01045 [gr-qc]

arXiv 2021
[5]

These correspond to very loud noise tran- sients and are expected to have SNR≳20

∆t hole – Distance to nearest “hole,” i.e., the re- gions where data is excised in theIAS-HMsearch pipeline. These correspond to very loud noise tran- sients and are expected to have SNR≳20. We include this parameter separately from the nearby triggers as these data regions do not have an ac- curate SNR estimate in our pipeline (as we inpaint such loud no...
[6]

However, this can be different from the true Bayesian probability, as the RF algorithm can skew the low and high probability tails when there are few training examples

Calibrating RF output probability Once the RF is trained, the functionpredict proba() outputs its classification probability. However, this can be different from the true Bayesian probability, as the RF algorithm can skew the low and high probability tails when there are few training examples. RF consists of decision trees, which are piecewise constant mo...

2023
[7]

B. P. Abbottet al.(LIGO Scientific Collaboration and Virgo Collaboration), Phys. Rev. X6, 041015 (2016)

2016
[8]

B. P. Abbottet al.(LIGO Scientific, Virgo), Phys. Rev. X9, 031040 (2019), arXiv:1811.12907 [astro-ph.HE]

arXiv 2019

Show all 87 references

[9]

Abbottet al., Physical Review X11, 021053 (2021), arXiv:2010.14527 [gr-qc]

R. Abbottet al., Physical Review X11, 021053 (2021), arXiv:2010.14527 [gr-qc]

2021 arXiv
[10]

Abbottet al.(LIGO Scientific, VIRGO, KAGRA), (2021), arXiv:2111.03606 [gr-qc]

R. Abbottet al.(LIGO Scientific, VIRGO, KAGRA), (2021), arXiv:2111.03606 [gr-qc]

2021 arXiv
[11]

Abbottet al.(LIGO Scientific, VIRGO, KAGRA), Astron

R. Abbottet al.(LIGO Scientific, VIRGO, KAGRA), Astron. Astrophys.659, A84 (2022), arXiv:2105.15120 [astro-ph.HE]

2022 arXiv
[12]

Kumar and T

P. Kumar and T. Dent, arXiv e-prints , arXiv:2403.10439 (2024), arXiv:2403.10439 [gr-qc]

2024 arXiv
[13]

Wadekar, J

D. Wadekar, J. Roulet, T. Venumadhav, A. K. Mehta,et al., arXiv e-prints , arXiv:2312.06631 (2023), arXiv:2312.06631 [gr-qc]

2023 arXiv
[14]

Venumadhav, B

T. Venumadhav, B. Zackay, J. Roulet, L. Dai, and M. Zaldarriaga, Phys. Rev.D100, 023011 (2019), arXiv:1902.10341 [astro-ph.IM]

2019 arXiv
[15]

Williams, Class

D. Williams, Class. Quant. Grav.42, 105012 (2025), arXiv:2401.08709 [astro-ph.HE]

2025 arXiv
[16]

Venumadhav, B

T. Venumadhav, B. Zackay, J. Roulet, L. Dai, and M. Zaldarriaga, Phys. Rev. D101, 083030 (2020), arXiv:1904.07214 [astro-ph.HE]

2020 arXiv
[17]

Olsen, T

S. Olsen, T. Venumadhav, J. Mushkin, J. Roulet, B. Za- ckay, and M. Zaldarriaga, Phys. Rev. D106, 043009 (2022)

2022
[18]

A. H. Nitz, C. Capano, A. B. Nielsen, S. Reyes, R. White, D. A. Brown, and B. Krishnan, The Astrophysical Jour- nal872, 195 (2019)

2019
[19]

A. H. Nitz, T. Dent, G. S. Davies, S. Kumar,et al., Astrophys. J.891, 123 (2020), arXiv:1910.05331 [astro- ph.HE]

2020 arXiv
[20]

A. H. Nitz, C. D. Capano, S. Kumar, Y.-F. Wang,et al., arXiv e-prints , arXiv:2105.09151 (2021), arXiv:2105.09151 [astro-ph.HE]

2021 arXiv
[21]

4-OGC: Catalog of gravitational waves from compact- binary mergers,

A. H. Nitz, S. Kumar, Y.-F. Wang, S. Kastha,et al., “4-OGC: Catalog of gravitational waves from compact- binary mergers,” (2021), arXiv:2112.06878 [astro- ph.HE]

2021 arXiv
[22]

H. S. Chia, T. D. P. Edwards, D. Wadekar, A. Zim- merman,et al., Phys. Rev. D110, 063007 (2024), arXiv:2306.00050 [gr-qc]

2024 arXiv
[23]

A. K. Mehta, S. Olsen, D. Wadekar, J. Roulet, et al., arXiv e-prints , arXiv:2311.06061 (2023), arXiv:2311.06061 [gr-qc]

2023 arXiv
[24]

A. E. Koloniari, E. C. Koursoumpa, P. Nousi, P. Lam- propoulos, N. Passalis, A. Tefas, and N. Ster- gioulas, arXiv e-prints , arXiv:2407.07820 (2024), arXiv:2407.07820 [gr-qc]

2024 arXiv
[25]

Vallisneri, J

M. Vallisneri, J. Kanner, R. Williams, A. Weinstein, and B. Stephens, inJournal of Physics Conference Series, Journal of Physics Conference Series, Vol. 610 (2015) p. 012021, arXiv:1410.4839 [gr-qc]

2015 arXiv
[26]

Abbottet al., The Astrophysical Journal900, L13 (2020)

R. Abbottet al., The Astrophysical Journal900, L13 (2020)

2020
[27]

Soniet al.(LIGO), Class

S. Soniet al.(LIGO), Class. Quant. Grav.42, 085016 (2025), arXiv:2409.02831 [astro-ph.IM]

2025
[28]

Cabero, A

M. Cabero, A. Lundgren, A. H. Nitz, T. Dent, et al., arXiv e-prints , arXiv:1901.05093 (2019), arXiv:1901.05093 [physics.ins-det]

2019 arXiv
[29]

Soni and LIGO Scientific Collaboration, Classical and Quantum Gravity38, 025016 (2021), arXiv:2007.14876 [astro-ph.IM]

S. Soni and LIGO Scientific Collaboration, Classical and Quantum Gravity38, 025016 (2021), arXiv:2007.14876 [astro-ph.IM]

2021 arXiv
[30]

Ho-Yeuk Cheung, D

M. Ho-Yeuk Cheung, D. Wadekar, A. K. Mehta, T. Is- lam,et al., arXiv e-prints , arXiv:2507.01083 (2025), arXiv:2507.01083 [gr-qc]

2025 arXiv
[31]

Marchant and T

P. Marchant and T. Moriya, Astron. Astrophys.640, L18 (2020), arXiv:2007.06220 [astro-ph.HE]

2020 arXiv
[32]

Farmer, M

R. Farmer, M. Renzo, S. E. de Mink, P. Marchant, and S. Justham, (2019), 10.3847/1538-4357/ab518b, arXiv:1910.12874 [astro-ph.SR]

2019 arXiv
[33]

L. A. C. van Son, S. E. De Mink, F. S. Broekgaarden, M. Renzo,et al., ApJ897, 100 (2020), arXiv:2004.05187 [astro-ph.HE]

2020 arXiv
[34]

A. K. Mehta, A. Buonanno, J. Gair, M. C. Miller,et al., Astrophys. J.924, 39 (2022), arXiv:2105.06366 [gr-qc]

2022 arXiv
[35]

D. D. Hendriks, L. A. C. van Son, M. Renzo, R. G. Izzard, and R. Farmer, MNRAS (2023), 10.1093/mn- ras/stad2857, arXiv:2309.09339 [astro-ph.HE]

2023 arXiv
[36]

Golomb, M

J. Golomb, M. Isi, and W. Farr, arXiv e-prints , arXiv:2312.03973 (2023), arXiv:2312.03973 [astro- ph.HE]

2023 arXiv
[37]

Gerosa, N

D. Gerosa, N. Giacobbo, and A. Vecchio, Astrophys. J. 915, 56 (2021), arXiv:2104.11247 [astro-ph.HE]

2021 arXiv
[38]

Fragione, I

G. Fragione, I. Ginsburg, and B. Kocsis, ApJ856, 92 (2018), arXiv:1711.00483 [astro-ph.GA]

2018 arXiv
[39]

Kritos, E

K. Kritos, E. Berti, and J. Silk, Phys. Rev. D108, 083012 (2023), arXiv:2212.06845 [astro-ph.HE]

2023 arXiv
[40]

Kritos, L

K. Kritos, L. Reali, D. Gerosa, and E. Berti, Phys. Rev. D110, 123017 (2024), arXiv:2409.15439 [astro-ph.HE]

2024 arXiv
[41]

Kritos, V

K. Kritos, V. Strokov, V. Baibhav, and E. Berti, Phys. Rev. D110, 043023 (2024), arXiv:2210.10055 [astro-ph.HE]

2024 arXiv
[42]

Y. Yang, I. Bartos, V. Gayathri, K. E. S. Ford,et al., Phys. Rev. Lett.123, 181101 (2019)

2019
[43]

McKernan, K

B. McKernan, K. E. S. Ford, R. O’Shaugnessy, and D. Wysocki, Monthly Notices of the Royal Astronomical Society494, 1203 (2020), https://academic.oup.com/mnras/article- pdf/494/1/1203/33029648/staa740.pdf

2020
[44]

Tagawa, Z

H. Tagawa, Z. Haiman, I. Bartos, and B. Kocsis, The Astrophysical Journal899, 26 (2020)

2020
[45]

and Gr¨ obner, M., A&A639, A108 (2020)

Ishibashi, W. and Gr¨ obner, M., A&A639, A108 (2020)

2020
[46]

Secunda, J

A. Secunda, J. Bellovary, M.-M. Mac Low, K. E. S. Ford, et al., Astrophys. J.903, 133 (2020), arXiv:2004.11936 [astro-ph.HE]

2020 arXiv
[47]

Tagawa, B

H. Tagawa, B. Kocsis, Z. Haiman, I. Bartos, K. Omukai, and J. Samsing, Astrophys. J.908, 194 (2021), arXiv:2012.00011 [astro-ph.HE]

2021 arXiv
[48]

G. Fabj, S. S. Nasim, F. Caban, K. E. S. Ford, B. McK- ernan, and J. M. Bellovary, Mon. Not. Roy. Astron. Soc. 499, 2608 (2020), arXiv:2006.11229 [astro-ph.GA]

2020 arXiv
[49]

Tagawa, Z

H. Tagawa, Z. Haiman, I. Bartos, and B. Kocsis, ApJ 899, 26 (2020), arXiv:2004.11914 [astro-ph.HE]

2020 arXiv
[50]

Samsing, I

J. Samsing, I. Bartos, D. J. D’Orazio, Z. Haiman,et al., Nature603, 237 (2022), arXiv:2010.09765 [astro-ph.HE]. 14

2022 arXiv
[51]

Franciolini, K

G. Franciolini, K. Kritos, L. Reali, F. Broekgaarden, and E. Berti, arXiv e-prints , arXiv:2401.13038 (2024), arXiv:2401.13038 [astro-ph.HE]

2024 arXiv
[52]

Biswas, L

R. Biswas, L. Blackburn, J. Cao, R. Essick,et al., Phys- ical Review D88(2013), 10.1103/physrevd.88.062003

2013 doi
[53]

Zevin, S

M. Zevin, S. Coughlin, S. Bahaadini, E. Besler,et al., Classical and Quantum Gravity34, 064003 (2017), arXiv:1611.04596 [gr-qc]

2017 arXiv
[54]

Cuoco, M

E. Cuoco, M. Cavagli` a, I. S. Heng, D. Keitel, and C. Mes- senger, Living Rev. Rel.28, 2 (2025), arXiv:2412.15046 [gr-qc]

2025 arXiv
[55]

Fernandes, A

J. Fernandes, A. Pai, and K. Chandra, arXiv e-prints , arXiv:2505.08332 (2025), arXiv:2505.08332 [gr-qc]

2025 arXiv
[56]

Sharma, K

K. Sharma, K. Chandra, and A. Pai, arXiv e-prints , arXiv:2208.02545 (2022), arXiv:2208.02545 [astro- ph.HE]

2022 arXiv
[57]

Saleemet al., Class

M. Saleemet al., Class. Quant. Grav.41, 195024 (2024), arXiv:2306.11366 [gr-qc]

2024 arXiv
[58]

Reissel, S

C. Reissel, S. Soni, M. Saleem, M. Coughlin, P. Harris, and E. Katsavounidis, (2025), arXiv:2501.04883 [gr-qc]

2025
[59]

Alvarez-Lopez, A

S. Alvarez-Lopez, A. Liyanage, J. Ding, R. Ng, and J. McIver, Class. Quant. Grav.41, 085007 (2024), arXiv:2304.09977 [gr-qc]

2024 arXiv
[60]

Lopez, G

M. Lopez, G. C. Santoro, A. Martins, S. Schmidt,et al., Phys. Rev. D111, 103020 (2025), arXiv:2412.17169 [astro-ph.IM]

2025
[61]

Malz and J

A.-K. Malz and J. Veitch, (2025), arXiv:2505.00657 [gr- qc]

2025 arXiv
[62]

Legin, M

R. Legin, M. Isi, K. W. K. Wong, Y. Hezaveh, and L. Perreault-Levasseur, Astrophys. J. Lett.985, L46 (2025), arXiv:2410.19956 [astro-ph.IM]

2025 arXiv
[63]

N. Shah, A. M. Knee, J. McIver, and D. Stenning, arXiv e-prints , arXiv:2306.13787 (2023), arXiv:2306.13787 [gr- qc]

2023 arXiv
[64]

Marxet al., Phys

E. Marxet al., Phys. Rev. D111, 042010 (2025), arXiv:2403.18661 [gr-qc]

2025 arXiv
[65]

M. B. Sch¨ afer, O. Zelenka, A. H. Nitz, H. Wang, et al., arXiv e-prints , arXiv:2209.11146 (2022), arXiv:2209.11146 [astro-ph.IM]

2022 arXiv
[66]

McLeod, D

A. McLeod, D. Beveridge, L. Wen, and A. Wicenec, Phys. Rev. D111, 024035 (2025), arXiv:2409.06266 [gr- qc]

2025 arXiv
[67]

S. J. Kapadia, T. Dent, and T. Dal Canton, Phys. Rev. D96, 104015 (2017), arXiv:1709.02421 [astro-ph.IM]

2017 arXiv
[68]

Essick, P

R. Essick, P. Godwin, C. Hanna, L. Blackburn, and E. Katsavounidis, (2020), 10.1088/2632-2153/abab5f, arXiv:2005.12761 [astro-ph.IM]

2020 arXiv
[69]

G. S. Davies, T. Dent, M. T´ apai, I. Harry, C. McIsaac, and A. H. Nitz, Phys. Rev. D102, 022004 (2020), arXiv:2002.08291 [astro-ph.HE]

2020 arXiv
[70]

Neyman and E

J. Neyman and E. S. Pearson, Philosophical Transactions of the Royal Society of London. Series A, Containing Pa- pers of a Mathematical or Physical Character231, 289 (1933)

1933
[71]

Zackay, T

B. Zackay, T. Venumadhav, J. Roulet, L. Dai, and M. Zaldarriaga, Phys. Rev. D104, 063034 (2021)

2021
[72]

Wadekar, T

D. Wadekar, T. Venumadhav, A. K. Mehta, J. Roulet, et al., arXiv e-prints , arXiv:2310.15233 (2023), arXiv:2310.15233 [gr-qc]

2023 arXiv
[73]

Wadekar, T

D. Wadekar, T. Venumadhav, J. Roulet, A. K. Mehta, B. Zackay, J. Mushkin, and M. Zaldarriaga, Phys. Rev. D 110, 044063 (2024), arXiv:2405.17400 [gr-qc]

2024 arXiv
[74]

Abbottet al., ApJ913, L7 (2021), arXiv:2010.14533 [astro-ph.HE]

R. Abbottet al., ApJ913, L7 (2021), arXiv:2010.14533 [astro-ph.HE]

2021 arXiv
[75]

Pimpalkaret al., in preparation

A. Pimpalkaret al., in preparation
[76]

Roulet, L

J. Roulet, L. Dai, T. Venumadhav, B. Zackay, and M. Zaldarriaga, Phys. Rev. D99, 123022 (2019)

2019
[77]

Zaheer, S

M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. Salakhutdinov, and A. Smola, arXiv e-prints , arXiv:1703.06114 (2017), arXiv:1703.06114 [cs.LG]

2017 arXiv
[78]

A. K. Mehta, D. Wadekar, J. Roulet, I. Anant- purkar,et al., arXiv e-prints , arXiv:2501.17939 (2025), arXiv:2501.17939 [gr-qc]

2025 arXiv
[79]

K. C. LIGO Scientific Collaboration, Virgo Collabora- tion, (2021), 10.5281/zenodo.5546675

2021 doi
[80]

Tiwari, Class

V. Tiwari, Class. Quant. Grav.35, 145009 (2018), arXiv:1712.00482 [astro-ph.HE]

2018 arXiv
[81]

A. K. Mehtaet al., in preparation
[82]

Godwin, R

P. Godwin, R. Essick, C. Hanna, K. Cannon,et al., arXiv e-prints , arXiv:2010.15282 (2020), arXiv:2010.15282 [gr- qc]

2020 arXiv
[83]

Davis, M

D. Davis, M. Trevor, S. Mozzon, and L. K. Nuttall, Phys. Rev. D106, 102006 (2022), arXiv:2204.03091 [gr- qc]

2022 arXiv
[84]

B. P. Abbottet al.(LIGO Scientific, Virgo), Phys. Rev. D93, 122003 (2016), arXiv:1602.03839 [gr-qc]

2016 arXiv
[85]

A. H. Nitz, T. Dent, T. Dal Canton, S. Fairhurst, and D. A. Brown, ApJ849, 118 (2017), arXiv:1705.01513 [gr- qc]

2017 arXiv
[86]

Cannon, C

K. Cannon, C. Hanna, and J. Peoples, arXiv e-prints , arXiv:1504.04632 (2015), arXiv:1504.04632 [astro- ph.IM]

2015 arXiv
[87]

Tsukadaet al., Phys

L. Tsukadaet al., Phys. Rev. D108, 043004 (2023), arXiv:2305.06286 [astro-ph.IM]

2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

We then compute the relative separation of the triggers from the reference candidate, ∆t i ≡t i −t candidate

∆t triggers (separation of nearby triggers from the candidate) – To analyze a candidate trigger at tcandidate, we make a list of nearby triggers (i.e., those falling within 0.5<|t−t candidate|<15 s) from the catalog already collected at each detector in the matched-filtering stage of the search pipeline 2. We then compute the relative separation of the tr...

[2] [2]

SNRs of nearby triggers – We also have infor- mation about the SNR of the nearby triggers in each detector. As our search includes higher har- monics, we use the full harmonic SNR (|ρ 2 HM| ≡ |ρ22|2 +|ρ ⊥ 33|2 +|ρ ⊥ 44|2) of the triggers, whereρ ⊥ 33 andρ ⊥ 44 correspond to the SNR of the orthogo- nalized higher-harmonic templates. We use the rescaled qua...

[3] [3]

In theIASsearches, the tem- plates do not correspond to physical parameters (e.g.,m 1, m2, s1z, s2z), but to singular vector com- ponents (c 0, c1) [67, 71]

Template ID of the nearby triggers – We use infor- mation about the template corresponding to the nearby triggers. In theIASsearches, the tem- plates do not correspond to physical parameters (e.g.,m 1, m2, s1z, s2z), but to singular vector com- ponents (c 0, c1) [67, 71]. Our template ID there- fore corresponds to a unique set of{c 0, c1}values. The logic...

[4] [4]

The LIGO Scientific Collaboration, the Virgo Collabo- ration,et al., arXiv e-prints , arXiv:2108.01045 (2021), arXiv:2108.01045 [gr-qc]

arXiv 2021

[5] [5]

These correspond to very loud noise tran- sients and are expected to have SNR≳20

∆t hole – Distance to nearest “hole,” i.e., the re- gions where data is excised in theIAS-HMsearch pipeline. These correspond to very loud noise tran- sients and are expected to have SNR≳20. We include this parameter separately from the nearby triggers as these data regions do not have an ac- curate SNR estimate in our pipeline (as we inpaint such loud no...

[6] [6]

However, this can be different from the true Bayesian probability, as the RF algorithm can skew the low and high probability tails when there are few training examples

Calibrating RF output probability Once the RF is trained, the functionpredict proba() outputs its classification probability. However, this can be different from the true Bayesian probability, as the RF algorithm can skew the low and high probability tails when there are few training examples. RF consists of decision trees, which are piecewise constant mo...

2023

[7] [7]

B. P. Abbottet al.(LIGO Scientific Collaboration and Virgo Collaboration), Phys. Rev. X6, 041015 (2016)

2016

[8] [8]

B. P. Abbottet al.(LIGO Scientific, Virgo), Phys. Rev. X9, 031040 (2019), arXiv:1811.12907 [astro-ph.HE]

arXiv 2019