Noise, Calibration, Conditioning, and Data Quality

Gravitational-wave analysis begins only after detector output has been calibrated, its noise characterized, transient contamination assessed, and usable time defined. These operations are coupled: calibration changes strain amplitude and phase, conditioning changes both signal and noise, a PSD supplies likelihood weights, and data-quality selection defines the population from which significance is estimated. This chapter turns those dependencies into a reproducible path from detector disturbances to analysis-ready strain.

Helpful background. Probability and Inference for Gravity supplies random processes, likelihoods, and selection effects. Gravitational Radiation and Waveforms supplies strain and signal morphology. Computational Relativity supplies discretization, convergence, and reproducible numerical practice.

Enter this chapter

Two complementary routes organize the six pages.

The noise-model route asks what fluctuations enter strain and how they should be weighted:

$$\{q_a,T_{hq_a}\} \longrightarrow S_h(f;t) \longrightarrow K(f;t)d(t) \longrightarrow(d-h\mid d-h).$$

It follows physical noise couplings into a local spectral estimate, then through conditioning and whitening into a likelihood or detection statistic.

The data-validity route asks what a stored number means and whether it may be used:

$$\{d_{\rm err},d_{\rm ctrl}\} \longrightarrow h_{\rm calibrated} \longrightarrow\text{glitch and state tests} \longrightarrow R_\alpha.$$

It follows control signals through calibration, candidate-level detector characterization, and interval selection for analysis $\alpha$. Neither route is optional. A beautiful PSD cannot repair invalid calibration, and a valid state flag cannot make a nonstationary likelihood correct.

The six steps

1. Instrumental Noise Sources organizes quantum, thermal, seismic, gravity-gradient, technical, and environmental noise by measured coupling path rather than by a name on a curve.
2. Power Spectral Density and Nonstationarity defines one-sided PSD normalization, windowed estimators, uncertainty, leakage, and tests of local stationarity.
3. Calibration Response and Uncertainty reconstructs free differential-arm strain from feedback signals and propagates frequency-dependent amplitude and phase uncertainty.
4. Data Conditioning and Whitening treats resampling, filtering, line handling, gating, and whitening as documented transformations of both data and signal model.
5. Glitches, Auxiliary Channels, and Vetoes separates morphology from cause and develops coupling, veto-performance, and safety tests.
6. Data-Quality Segments and Analysis Readiness builds usable live time from explicit interval algebra and analysis-dependent requirements.

The conceptual sequence begins with physical noise and its statistical description before reconstructing the full data product. In an actual pipeline, calibration and state selection precede most spectral calculations. The chapter therefore supplies concepts in one order and a workflow in another; the equations below make the operational dependencies explicit.

The analysis product is more than a time series

A compact representation of analysis input is

$$\mathscr D_\alpha =\left( \{h_I(t)\}, \{S_{n,I}(f;t)\}, \{\Lambda_{{\rm cal},I}\}, \{R_{\alpha,I}\}, \mathcal K_\alpha, \mathcal E \right),$$

where $h_I$ is calibrated strain, $S_{n,I}$ its local noise model, $\Lambda_{{\rm cal},I}$ calibration uncertainty, $R_{\alpha,I}$ the selected segments, $\mathcal K_\alpha$ the complete conditioning operator, and $\mathcal E$ the software environment. Releasing only an array of strain samples leaves the other objects implicit and the result difficult to reproduce.

For stationary Gaussian noise, the residual in detector $I$ is weighted by

$$(r_I\mid r_I) =4\operatorname{Re}\int_{f_{\rm low}}^{f_{\rm high}} \frac{\left\lvert\widetilde d_I(f)-\widetilde h_I(f;\theta)\right\rvert^2} {S_{n,I}(f)}\,df.$$

Every symbol has an empirical boundary. The PSD belongs to selected data and an estimator; the template must undergo the same conditioning; calibration is uncertain; the frequency band and usable interval are declared. Glitches and drift are evidence that the simple likelihood needs exclusion, marginalization, or a richer model, not permission to edit data until residuals look Gaussian.

A reproducible execution path

For one detector interval, the following order is a useful default:

1. resolve the exact dataset, release, channel, GPS span, checksum, and state/quality definitions;
2. verify calibration validity and load its response-uncertainty representation;
3. construct analysis-ready segments and the valid padding around them;
4. inspect raw and time–frequency diagnostics for state changes, lines, and glitches;
5. apply the declared anti-alias, resampling, band-limit, line, gate, and edge rules;
6. estimate the PSD from an interval justified for the signal and likelihood;
7. whiten data and injections with identical normalization and verify residual statistics;
8. repeat under justified changes of sample rate, segment length, PSD window, gate, calibration sample, and veto selection.

The order is not universal. For example, a glitch may need to be gated before a long whitening filter, and some likelihoods model lines or glitches rather than removing them. What is universal is that the full transformation, its support, and its tests are recorded.

The idealized noise model, practical conditioning, calibration, and detector-characterization framework are reviewed in LIGO Scientific Collaboration and Virgo Collaboration 2020, §§3–7, PDF. Current catalog-scale candidate methods through O4b are described in LVK Collaboration 2026, §§2–5, PDF.

Evidence, versions, and the observational boundary

Each computational claim should survive a synthetic benchmark containing colored noise, an off-bin line, a chirp, and a transient. Each public-data claim should pin detector, GPS interval, data-release and minor version, sample rate, file hash, quality mask, calibration caveat, processing code and environment, and an expected diagnostic. Failure modes—leakage, filter edges, nonstationarity, calibration interpolation, and unsafe or overfit vetoes—are results to report.

Public interfaces and current-method statements in this chapter are frozen to the GWOSC documentation available on 2026-08-18. A later data release or catalog can change file versions, flags, uncertainty products, or recommended use. Stable equations do not require a new cutoff, but every current empirical or software statement does.

Review the chapter

A $16\,384\,\mathrm{Hz}$ strain record contains a loud impulse two seconds before a candidate, a narrow line at $3000\,\mathrm{Hz}$, a slow broadband drift, and a calibration epoch boundary one second after the candidate. The intended analysis downsampled to $4096\,\mathrm{Hz}$ and uses data through three seconds after the candidate.

Design a defensible analysis and name four ways a naive one can fail.

Solution

First, verify state and calibration validity on both sides of the epoch boundary. Do not interpolate the response across it; either restrict the core to one valid epoch or process the epochs with their own responses and a justified combination. Include calibration samples in the final inference.

Second, investigate the impulse using available auxiliary and state information. If gating is justified, taper-gate before any long whitening filter and pass injections through the same gate. The filter padding must prevent the impulse or epoch boundary from contaminating the retained core.

Third, low-pass with measured stopband rejection before downsampling. Without anti-alias filtering, the $3000\,\mathrm{Hz}$ line folds to $1096\,\mathrm{Hz}$. Declare whether the line is filtered, excluded, modeled, or witnessed.

Fourth, estimate a local PSD and compare neighboring intervals because the broadband floor drifts. Repeat with several segment durations and PSD windows; a single long average can hide the change. Finally, apply identical conditioning to templates and synthetic injections and verify recovered amplitude, phase, and timing.

Naive failures include downsampling before anti-alias filtering, allowing whitening ringdown from the impulse into the candidate, smoothing calibration across the boundary, estimating one stationary PSD across the drift, and selecting a gate or veto after inspecting the desired result.

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.