Glitches, Auxiliary Channels, and Vetoes

A glitch is a transient departure from the noise model, not a physical explanation. Its morphology can resemble a gravitational-wave signal, and its cause may lie in optics, controls, electronics, the environment, or an unidentified nonlinear coupling. Auxiliary channels and controlled injections can establish instrumental associations; vetoes can then remove contaminated data only after their efficiency, statistical validity, and safety for real signals have been demonstrated.

Required background. Probability and Inference for Gravity supplies hypothesis tests, coincidence statistics, and selection effects.

Helpful background. Gravitational Radiation and Waveforms supplies expected signal morphology and detector response. Computational Relativity supplies time–frequency transforms, injection tests, and reproducible workflows.

Transient morphology is a clue, not a cause

Glitches can be localized in time, frequency, or both. A short impulse produces broadband power; a resonant ringdown is narrow in frequency and extended in time; scattered light can trace curved time–frequency tracks; control saturation can create abrupt, state-dependent structure. The same visible morphology can arise from different mechanisms, and one mechanism can produce several morphologies as detector state changes.

A useful characterization records:

• central time, duration, frequency support, bandwidth, and amplitude or search statistic;
• morphology in raw and conditioned representations;
• detector and operating state, nearby state transitions, and data gaps;
• coincident activity in physical-environment and control channels;
• robustness to PSD choice, whitening, filter padding, and gate parameters;
• whether a modeled gravitational-wave signal gives a coherent network explanation.

Time–frequency images are diagnostics, not calibrated likelihoods unless their pixel normalization and trials are specified. A feature that looks dramatic after aggressive whitening may carry little calibrated strain power, while a broad low-amplitude disturbance can bias a long likelihood interval.

The distinction between non-Gaussian transients, environmental monitoring, and data-quality actions is developed in LIGO Scientific Collaboration and Virgo Collaboration 2020, §6.2, PDF.

Auxiliary witnesses and coupling functions

An auxiliary channel $x(t)$ monitors a degree of freedom other than the main strain readout: ground motion, magnetometers, microphones, beam alignment, laser controls, suspension controls, or internal digital states. Correlation with strain $h(t)$ is quantified spectrally by the magnitude-squared coherence

$$\gamma_{hx}^2(f) =\frac{\left\lvert S_{hx}(f)\right\rvert^2} {S_{hh}(f)S_{xx}(f)}, \qquad 0\le\gamma_{hx}^2\le1.$$

High coherence says that a linear relation explains some covariance over the averaged interval. It does not establish direction of causation: both channels may respond to a third disturbance, share digital cross-talk, or be contaminated by the same state transition.

When $x$ is an adequate witness and the coupling is linear and stationary, the least-squares transfer estimate is

$$\widehat T_{hx}(f)=\frac{\widehat S_{hx}(f)}{\widehat S_{xx}(f)}.$$

The corresponding ideal residual spectrum after subtracting $\widehat T_{hx}x$ is

$$S_{rr}(f)=S_{hh}(f)\bigl[1-\gamma_{hx}^2(f)\bigr].$$

This bound is useful precisely because real subtraction often performs worse: finite witness signal-to-noise ratio, transfer drift, nonlinear coupling, and estimation error limit it. A deliberate environmental or control injection tests amplitude and phase scaling and should be performed at several strengths to identify nonlinearity or saturation.

From an auxiliary trigger to a veto

Suppose an auxiliary algorithm marks a set of time windows $V$ within analyzed live time $T_{\rm live}$. For a population of $N$ strain-channel triggers, define

$$\text{deadtime} =D=\frac{\lvert V\rvert}{T_{\rm live}}, \qquad \text{efficiency} =E=\frac{N_{\rm vetoed}}{N}.$$

If trigger times are uniform over the live time and independent of $V$, the expected efficiency is the deadtime. For a varying trigger rate $\lambda(t)$, even an independent veto instead has expected efficiency $\int_V\lambda(t)dt/\int_{T_{\rm live}}\lambda(t)dt$. The ratio $E/D$ is therefore a useful enrichment diagnostic, but not a complete significance measure. Windows can cluster, trigger rate can drift, thresholds can be tuned after seeing the target population, and many auxiliary channels and window sizes create a trials factor.

A credible veto study declares the auxiliary channel, trigger algorithm, threshold, window construction, tuning and evaluation intervals, target trigger population, live-time denominator, and statistical test. It estimates accidental coincidences with time shifts or an appropriate inhomogeneous null model. Performance is then evaluated on data not used to tune the veto.

Veto application is analysis dependent. A broad transient search, compact-binary search, stochastic analysis, and continuous-wave search use different durations, frequency regions, and statistics. A segment unsuitable for one can remain useful for another.

Safety: do not veto the sky

An auxiliary channel used as a veto must be insensitive, within a quantified bound, to a real differential-arm gravitational-wave signal. Otherwise a loud astrophysical event could trigger both strain and the “veto” channel and remove itself.

Safety evidence can include:

1. a physical argument that the sensor does not couple to differential arm motion at relevant amplitude;
2. hardware injections that move the test masses as a simulated gravitational wave would and test for auxiliary triggers;
3. software injections passed through the trigger and veto logic to test the analysis downstream of their insertion point;
4. checks for digital cross-coupling, saturation, shared timing, and state-vector contamination;
5. explicit handling of channels known to sense the gravitational-wave control path.

Adding a waveform only to stored strain cannot test whether a real signal couples into an auxiliary sensor: that requires a physical coupling argument, hardware injections, or a validated model that also injects the auxiliary channels. Hardware injections likewise test their actual actuation path and frequency range, not every possible coupling. Absence of an auxiliary trigger during a few tests is not proof of zero coupling. Quote the tested amplitude and frequency region, statistical power, and any excluded operating state. Veto safety is a property of a channel, configuration, algorithm, and epoch together.

Candidate-level assessment and network coherence

For a candidate near a glitch, ask two different questions. First, is there positive evidence for an instrumental disturbance: auxiliary coincidence, known morphology, nonstationary residual, or a failed state? Second, does a coherent astrophysical model explain the calibrated network data better than glitch alternatives? Neither question should be replaced by the other.

A disturbance in one detector can coexist with a real signal. Gating or joint signal–glitch modeling may recover information from the other detectors and uncontaminated frequencies. Conversely, no auxiliary witness does not prove an event is astrophysical; many coupling paths are unmonitored or nonlinear. Current catalog methods use candidate-specific data-quality investigations alongside search and inference results, as described in LVK Collaboration 2026, §§3–5, PDF.

Reproducible veto evidence

Begin with synthetic colored noise containing a known chirp and two glitch populations: one linearly coupled to a witness and one unwitnessed. Tune a veto on one realization and evaluate it on another. Refine sample rate, time–frequency resolution, PSD window, trigger threshold, veto window, and gate taper. Report efficiency, deadtime, accidental-coincidence distribution, injection survival, and recovered signal mismatch.

For a public-data case, pin dataset and minor version, detector, GPS interval, sample rate, calibration caveat, data-quality mask, strain-trigger definition, available auxiliary or derived flags, software environment, and all tuned choices through the GWOSC documentation available on 2026-08-18. Public releases do not necessarily expose every internal auxiliary channel, so unavailable causal tests must be labeled as such. Repeat under reasonable PSD, conditioning, and calibration samples. Do not describe a public flag as proving a mechanism unless its official documentation does so.

This workflow states what should be executed and reported; it does not claim reproduction of an internal collaboration detector-characterization pipeline.

Common pitfalls

Using a high $E/D$ as the only validation. Optimization over many channels and windows can produce large ratios by chance. Use held-out data and a trials-aware null distribution.

Assuming an auxiliary coincidence identifies the cause. Coincidence narrows hypotheses. A measured coupling, injection scaling, and intervention are stronger causal evidence.

Making a permanent veto from an epoch-specific coupling. Instrument state and transfer functions drift. Revalidate performance and safety for the configuration and interval in use.

Exercises

1. Evaluate a veto. In $10^5\,\mathrm{s}$ of live time, veto windows occupy $500\,\mathrm{s}$ and remove 80 of 2000 strain triggers. Find $D$, $E$, and $E/D$. Under a uniform independent-trigger null, how many vetoed triggers are expected?

Solution

The deadtime and efficiency are

$$D=\frac{500}{10^5}=0.005, \qquad E=\frac{80}{2000}=0.040.$$

Thus $E/D=8$. The simple null predicts $ND=2000(0.005)=10$ coincidences. Observing 80 is strong enrichment under that null, but a real analysis must account for nonuniform trigger rates, clustering, tuning, and trials before assigning significance.

2. Interpret coherence. Over a band, $\gamma_{hx}^2=0.36$. What fraction of strain power can an ideal stationary linear subtraction remove, and what residual remains?

Solution

The ideal projection removes the fraction $\gamma_{hx}^2=0.36$ of the strain PSD, leaving

$$\frac{S_{rr}}{S_{hh}}=1-0.36=0.64.$$

This is an ensemble linear result, not a guarantee for each glitch. Nonstationarity, transfer-estimation error, witness noise, and nonlinear coupling can leave a larger residual.

References

• Gravitational Wave Open Science Center. “O4 Technical Details.” Version available 2026-08-18. Official documentation.
• LIGO Scientific Collaboration and Virgo Collaboration. “A Guide to LIGO–Virgo Detector Noise and Extraction of Transient Gravitational-Wave Signals.” Classical and Quantum Gravity 37 (2020): 055002. DOI. Open PDF.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration. “GWTC-5.0: Methods for Identifying and Characterizing Gravitational-Wave Transients.” arXiv:2605.27224 (2026). Open PDF.