Gravitational-Wave Astronomy

Gravitational-wave astronomy is the discipline of turning minute spacetime strain into defensible statements about sources, populations, gravity, and the Universe. The transformation is not one calculation. It is a chain: an instrument responds to a wave; calibration converts readout into a physical channel; a search assigns significance; a likelihood and prior produce parameter constraints; selection effects connect detected events to a population; and external data may add a counterpart, host, or cosmological redshift. A conclusion is only as reliable as the weakest unexamined link.

This volume develops the complete reasoning from detector measurements to astrophysical conclusions for ground-based interferometers, pulsar timing arrays, and space-based observatories. It emphasizes what the data directly constrain, which assumptions enter each inference, how systematic uncertainty propagates, and what an independent public-data analysis can actually reproduce. The observational and frontier comparisons include results through 2026-08-18; durable equations and methods are identified separately.

Helpful background. Gravitational Radiation and Waveforms supplies radiation theory and waveform construction. Probability and Statistical Inference supplies likelihoods, priors, marginalization, and decision concepts. Waves, Optics, and Radiation is useful for the detector-response chapters.

Astronomy as a chain of conditional claims

Start with a physical scale. Strain is a fractional change in separation: in a favorably aligned, ideal long-wavelength interferometer, $h\sim\Delta L/L$. A strain of $10^{-21}$ across $4\,\mathrm{km}$ corresponds to a differential length change of roughly $4\times10^{-18}\,\mathrm m$. Light measures that change through phase. An observed chirp’s increasing frequency then encodes the binary’s redshifted mass combination, while its amplitude and polarization help constrain distance and orientation. Each step adds a physical model; none requires treating the result as arbitrary.

Four questions help keep the statistics concrete:

Question	Main object	Start here
What does the instrument measure?	Response of an optical or timing channel to strain	Michelson response
Is a candidate inconsistent with the measured noise background?	Search statistic and its calibrated false-alarm rate	False alarms and trials
Which source parameters fit this event?	Likelihood, prior, and posterior	Gravitational-wave likelihood
What sources occur in the Universe?	Population model and detection efficiency	Selection from injections

An idealized detector output can be written

$$\widetilde d_I(f)=R_I^{A}(f,\boldsymbol\Omega)\,\widetilde h_A(f)+\widetilde n_I(f),$$

Here $I$ labels a detector or data channel, the repeated $A$ is summed over polarizations, $R_I^A$ is the calibrated response, and $n_I$ is noise. The frequency is $f$, $\boldsymbol\Omega$ specifies sky direction, and a tilde denotes a Fourier transform. This compact equation hides much of the work. The response may depend on arm transfer functions, optical cavities, control loops, spacecraft delays, time, sky direction, and calibration version. The noise is neither perfectly stationary nor perfectly Gaussian. A data-quality decision changes which samples enter the analysis.

For stationary Gaussian noise with a declared one-sided power spectral density $S_n(f)$, a useful inner product is

$$(a\mid b) =4\,\operatorname{Re}\int_{f_{\min}}^{f_{\max}} \frac{\widetilde a(f)\widetilde b^*(f)}{S_n(f)}\,df.$$

The star denotes complex conjugation and $\operatorname{Re}$ takes the real part. Dividing by $S_n$ gives less weight to frequencies where the instrument is noisier: the same waveform residual is more surprising in a quiet band. The strain inner product is dimensionless, since the Fourier strain has units of seconds and $S_n$ has units of strain squared per hertz.

For source parameters $\theta$ and a specified signal/noise model $\mathcal M$, the corresponding likelihood has schematic form

$$\mathcal L(d\mid\theta,\mathcal M) \propto \exp\!\left[-\frac12(d-h(\theta)\mid d-h(\theta))\right].$$

It is an analysis model, not a statement that detector noise is literally Gaussian at every time. Conditioning, segmentation, glitch treatment, calibration marginalization, and residual checks determine whether the approximation is adequate for the claim being made.

Bayes’ theorem then gives

$$p(\theta\mid d,\mathcal M) =\frac{\mathcal L(d\mid\theta,\mathcal M)\,\pi(\theta\mid\mathcal M)} {\mathcal Z_{\mathcal M}}, \qquad \mathcal Z_{\mathcal M} =\int d\theta\,\mathcal L(d\mid\theta,\mathcal M)\pi(\theta\mid\mathcal M).$$

The prior density $\pi$ describes the parameter distribution before using these data; the posterior $p$ describes it afterward. The evidence $\mathcal Z_{\mathcal M}$ normalizes that posterior and enters comparisons between models. Changing coordinates without transforming the prior changes the inference. A credible interval is not a repeated-sampling confidence interval, and a Bayes factor is not a false-alarm probability.

For a population with hyperparameters $\Lambda$, posterior samples enter through an integral that divides out their event-level priors. Selection then normalizes the detected ensemble. Let $\nu$ be the expected number of sources before detection cuts in a declared observing exposure and source domain, and let $p(\theta\mid\Lambda)$ be normalized on that domain. A Poisson likelihood is

$$\mathcal L(\{d_i\}\mid\Lambda,\nu) \propto e^{-\nu\alpha(\Lambda)}\nu^N \prod_i \int d\theta_i\, \mathcal L(d_i\mid\theta_i)\, p(\theta_i\mid\Lambda),$$

where $N$ is the detected count and $\alpha(\Lambda)=\int p_{\rm det}(\theta)p(\theta\mid\Lambda)d\theta$. Converting $\nu$ to a physical merger-rate density additionally requires the observing time, volume measure, redshift evolution, and source-time dilation. Conditioning on $N$ instead gives a shape likelihood with $\alpha(\Lambda)^{-N}$. These are alternative treatments of the count, not two corrections to apply independently. The hierarchical-likelihood page derives their relation. Selection does not add an inverse-efficiency factor to an ordinary single-event source posterior; that distinction is made explicit in Thrane and Talbot 2019, Appendices E–F, corrected PDF v9.

These equations organize the volume. Each chapter asks which term is being constructed, measured, approximated, or marginalized—and which conclusion would fail if it were wrong.

Conventions and interpretation boundaries

The site uses metric signature $(-,+,+,+)$ and time dependence $e^{-i\omega t}$. Detector spectra are one-sided unless a page explicitly says otherwise. Fourier normalization, channel definition, polarization basis, sky direction, and detector tensor are stated where they first affect a result.

Source-frame and detector-frame quantities are not interchangeable. Cosmological redshift gives

$$m^{\rm det}=(1+z)m^{\rm src}, \qquad f^{\rm det}=\frac{f^{\rm src}}{1+z},$$

while an amplitude measurement primarily constrains luminosity distance under an inclination and polarization model. “Mass,” “distance,” “spin,” “rate,” and “sensitivity” are incomplete labels unless the frame, measure, and selection conditions are known.

This volume develops the observational likelihood chain. It uses, but does not redevelop, waveform theory, numerical-relativity production, compact-object microphysics, black-hole perturbation theory, binary dynamics, cosmological dynamics, or formation modeling. Links connect each ingredient to its detailed treatment. In particular, a detector posterior on tidal deformability is not by itself an equation of state; a fitted population component is not by itself a formation channel; and an improved ringdown fit is not by itself a uniquely identified quasinormal mode.

Instruments and calibrated data

Ground-Based Interferometers and Detector Response begins with the Michelson response and adds Fabry–Pérot storage, antenna patterns, finite-arm effects, detector networks, and documented observing configurations. The goal is to understand exactly how a tensor perturbation becomes an optical and electronic measurement—and when a frequency-independent antenna approximation fails.

Noise, Calibration, Conditioning, and Data Quality follows the readout into an analysis-ready strain channel. It distinguishes amplitude spectral density from power spectral density, propagates a complex calibration response and its uncertainty, explains whitening and filtering, and treats glitches, auxiliary witnesses, veto safety, and valid data segments. A nominal sensitivity curve is never substituted for the time-dependent data used in an inference.

Together these chapters establish the first boundary:

$$\text{spacetime strain} \longrightarrow \text{instrument response} \longrightarrow \text{versioned calibrated data}.$$

Searches, event inference, and catalogs

Detection Statistics and Signal Searches derives likelihood ratios and matched filtering in their ideal regime, then adds template-bank mismatch, coherent networks, unmodeled bursts, trials, false-alarm estimation, injections, and empirical sensitivity. A ranking statistic is not a significance until its noise-background distribution and trials procedure are defined.

Bayesian and Frequentist Inference separates likelihoods, prior measures, evidences, posterior samplers, confidence procedures, coverage, and predictive checks. It treats convergence as a scientific requirement: a stable-looking trace or one large evidence estimate is not sufficient without effective-sample, repeated-run, or injection diagnostics.

Compact-Binary Parameter Estimation maps waveform parameters into observable combinations. It develops the chirp-mass hierarchy, mass–spin and distance–inclination degeneracies, waveform-model uncertainty, calibration and noise marginalization, astrophysical and analysis priors, and injection recovery. The chapter repeatedly distinguishes a narrow conditional posterior from a robust measurement.

Catalog Construction and Selection Effects explains how triggers from heterogeneous searches become versioned catalog events. Thresholds, false alarms, astrophysical membership, parameter-estimation availability, data-quality choices, and injection support are all part of a catalog definition. Catalog versions are evidence objects: a later release may change membership or interpretation without making the older analysis irrational.

These four chapters form the central data-analysis spine:

$$\text{calibrated data} \longrightarrow \text{search candidates} \longrightarrow \text{event likelihoods} \longrightarrow \text{versioned catalog}.$$

Populations, source classes, and strong-field tests

Source Populations and Hierarchical Inference derives posterior recycling, rate terms, sensitive spacetime volume, and selection-normalized population likelihoods before interpreting mass, spin, redshift, or mixture structure. It uses mock-catalog recovery and alternate model families to separate a resolved feature from a flexible model absorbing misspecification.

Binary Black-Hole Observations asks what individual signals and catalogs establish about masses, spins, precession, eccentricity, higher modes, and inspiral–merger–ringdown consistency. Pair-instability and formation interpretations are presented as population- and stellar-model-dependent comparisons, not direct labels attached to events.

Neutron-Star–Black-Hole and Binary-Neutron-Star Observations adds matter response: inspiral tides, possible disruption, postmerger emission, and shared equation-of-state inference. High-frequency waveform accuracy, calibration, detector sensitivity, population priors, and electromagnetic coverage set the evidence ceiling.

Ringdown Tests and Strong-Field Spectroscopy connects full-signal remnant prediction to finite-window ringdown likelihoods. Start time, noise covariance, overtones, angular modes, resolvability, parametrized deviations, and echo searches are treated as coupled inference choices. Quasinormal-mode theory lives in the perturbation and waveform volumes; this chapter asks what the data identify.

At the 2026-08-18 cutoff, the official GWTC-5.0 products are the current ground-based catalog interface used for dated examples. Every such example names the release and remains conditional on its search, waveform, calibration, threshold, and selection products LVK 2026, GWTC-5.0 methods.

Long-lived, nanohertz, and space-based observations

Continuous Waves and Stochastic Backgrounds develops coherent and semicoherent searches for long-lived sources, then constructs stochastic cross-correlation, anisotropic maps, polarization and non-Gaussian extensions, and foreground-confusion models. Computing cost and template trials are part of all-sky continuous-wave sensitivity; overlap functions and correlated noise are part of stochastic significance.

Pulsar Timing Arrays turns pulse times of arrival into timing residuals, separates Earth and pulsar terms, derives Hellings–Downs correlations, and compares tensor structure with clock, ephemeris, intrinsic, and chromatic processes. The NANOGrav 15-year result and contemporaneous PTA releases provide dated evidence for a nanohertz gravitational-wave background, while source composition and individual-source identification remain less settled Agazie et al. 2023, §§3–5.

Space-Based Gravitational-Wave Astronomy builds a moving constellation from one-way phase or frequency links. Time-delay interferometry cancels laser noise with ordered delay operators; unequal, flexing arms make the ordering consequential. The chapter then treats instrumental noise, Galactic confusion, massive-black-hole binaries, extreme-mass-ratio inspirals, and multiband sources under an explicitly versioned mission configuration.

These regimes share likelihood logic but not interchangeable response functions. An interferometer cross-correlation, a pulsar angular correlation, and a TDI channel can all probe stochastic radiation, yet their timestamps, transfer functions, noise covariances, and foregrounds differ fundamentally.

Multimessenger inference and open-data closure

Multimessenger Events, Standard Sirens, and Open Data completes the chain. It distinguishes physical counterpart association from chance coincidence, relates detector networks to low-latency sky maps, derives bright- and dark-siren likelihoods, and propagates host-catalog incompleteness, peculiar velocity, inclination, selection, and weak lensing.

The final pages specify public strain, event, posterior, calibration, and data-quality products and then assemble a bounded event-analysis reproduction. “Reproduce” does not mean recreating a collaboration’s private production system. It means pinning public input versions and software, recovering declared intermediate diagnostics, explaining stochastic and waveform differences, and testing the result with an injection or analytic control. The Gravitational Wave Open Science Center is the principal public interface for ground-based data and tutorials.

Choose a route through the volume

Reader goal	Recommended route	What the route delivers
Graduate data analysis	Chapters 1–7, then chapter 14	The complete instrument → search → event → catalog → population chain
Detector characterization	Chapters 1–3, then open-data products in chapter 14	Response, calibration, noise, data quality, search sensitivity
Compact-binary inference	Chapters 3–10, then standard sirens	Event degeneracies, selection, populations, matter, and strong-field tests
Catalog and population work	Chapters 4–8, with chapters 2–3 for sensitivity	Likelihoods, catalog definitions, injections, rates, and population interpretation
Long-baseline astronomy	Chapter 4, then chapters 11–13	Continuous, stochastic, PTA, and space-based response/inference
Multimessenger cosmology	Chapters 4–6 and 14	Distance likelihoods, localization, association, hosts, and selection
Public-data reproduction	Chapters 2–5, then the final two pages of chapter 14	Versioned inputs, conditioning, likelihood checks, and rerun boundaries

The sequential route is deliberate, but the volume is also modular. Every ordinary page states the preparation it actually uses, so a reader can enter at a later chapter without pretending that an overview replaces technical prerequisites.

What a defensible result records

A gravitational-wave result should make the following objects recoverable:

• the detector, channel, observing interval, data-quality mask, calibration version, and frequency band;
• the signal and noise models, Fourier/PSD convention, likelihood, parameter coordinates, and priors;
• the search threshold, background procedure, trials accounting, and injection population where significance or sensitivity is claimed;
• the waveform family, software and data versions, convergence diagnostics, and controlled recovery test;
• the catalog inclusion rule, astrophysical-membership treatment, selection function, and cosmology where a population is inferred;
• the evidence cutoff and the strongest credible alternative interpretation for a frontier claim.

No single diagnostic substitutes for the set. A zero-residual plot does not establish sampler convergence; a high Bayes factor does not calibrate frequentist false alarms; an injection campaign drawn from the fitted population does not test broad model misspecification; and a public notebook that runs does not prove it analyzed the intended data segment.

Common pitfalls

Treating the strain file as a raw observable. Public strain has already passed through a calibrated response and data-quality process. The response version and valid-use conditions belong in the analysis.

Calling a posterior feature a measurement without qualifiers. Posterior structure may be supplied by coordinates, priors, waveform restrictions, or degeneracies. State the combination the data constrain and show relevant alternative analyses.

Ignoring selection after detection. Detected sources are weighted toward larger amplitudes and more favorable parameters. Population and cosmological likelihoods must include the same detection process that produced the catalog.

Turning consistency into unique identification. A spectrum, mode, mixture component, or counterpart may be compatible with a physical explanation without identifying it uniquely. Competing models and discriminating evidence must be stated.

Freezing mutable facts into timeless prose. Catalog membership, bounds, mission configurations, software behavior, and detection verdicts change. Version and date them; keep durable derivations separate.

Check your preparation

1. Likelihood versus selection. An event analysis uses the correct likelihood for every detected source, but a population analysis omits detection efficiency. Why can the event posteriors still be individually valid while the population distribution is biased?

Solution

Conditioning on a detected event does not require the event posterior to describe undetected systems. Population inference does: it compares the observed sample with all systems that could have occurred. Without the selection normalization, intrinsically louder or more favorably oriented systems are overrepresented and mistaken for the underlying population.

2. A chain failure. A compact-binary candidate has a small matched-filter false-alarm probability in stationary Gaussian noise, but the data contain an unmodeled glitch near the signal. Which links must be revisited before an astrophysical claim?

Solution

The ideal null distribution no longer calibrates the actual search statistic. One must revisit data quality and glitch coupling, empirical background estimation and trials, coherent behavior across detectors, injection recovery in representative data, and the event likelihood or glitch marginalization. Parameter and population inference should not proceed as if only the final posterior were affected.

References

• Abbott, B. P., et al. (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, 055002 (2020), doi:10.1088/1361-6382/ab685e.
• Agazie, G., et al. (NANOGrav Collaboration), “The NANOGrav 15 yr Data Set: Evidence for a Gravitational-wave Background,” The Astrophysical Journal Letters 951, L8 (2023), doi:10.3847/2041-8213/acdac6.
• Colpi, Monica, et al., LISA Definition Study Report, ESA-SCI-DIR-RP-002 (2024), official PDF. Preprint.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration, “GWTC-5.0: Methods for Identifying and Characterizing Gravitational-Wave Transients,” arXiv:2605.27224 (2026), open manuscript.
• Romano, J. D., and Cornish, N. J., “Detection Methods for Stochastic Gravitational-Wave Backgrounds: A Unified Treatment,” Living Reviews in Relativity 20, 2 (2017), doi:10.1007/s41114-017-0004-1.
• Thrane, E., and Talbot, C., “An Introduction to Bayesian Inference in Gravitational-Wave Astronomy: Parameter Estimation, Model Selection, and Hierarchical Models,” Publications of the Astronomical Society of Australia 36, e010 (2019), doi:10.1017/pasa.2019.2. Open PDF v9, including the selection erratum.