Counterpart Association and Chance Coincidence

A transient observed in another messenger is not associated with a gravitational-wave candidate merely because the two appeared near the same time. Association compares a common-source hypothesis with unrelated astrophysical and instrumental alternatives using time, sky position, distance, source class, and both surveys’ selection processes. The result is an odds statement conditional on rates and models—not a label supplied by positional overlap alone.

Helpful background. Gravitational-Wave Astronomy supplies detector likelihoods, sky maps, and event significance. Relativistic Cosmology supplies distance–redshift relations. Black-Hole Astrophysics supplies source populations and host environments.

Common source versus chance coincidence

Let $d_{\rm GW}$ and $d_X$ denote data from the gravitational-wave and external-messenger instruments. Compare

• $\mathcal H_C$: both data sets arise from one source with shared parameters $\lambda$;
• $\mathcal H_U$: the candidates are unrelated and their source parameters are drawn independently.

The posterior odds are

$$\mathcal O_{C/U} =\frac{p(\mathcal H_C)}{p(\mathcal H_U)} \frac{p(d_{\rm GW},d_X\mid\mathcal H_C)} {p(d_{\rm GW},d_X\mid\mathcal H_U)}.$$

Under conditional independence given the common parameters, and when the two analyses use the same normalized prior $\pi(\lambda)$ on those parameters, the common-source Bayes factor contains the posterior-overlap integral

$$\mathcal B_{C/U}^{\lambda} =\int \frac{p(\lambda\mid d_{\rm GW})p(\lambda\mid d_X)} {\pi(\lambda)}\,d\lambda.$$

The division by the prior is essential. Write each single-instrument likelihood as $\mathcal L_a=Z_a p(\lambda\mid d_a)/\pi(\lambda)$. In the common-source evidence $\int\mathcal L_{\rm GW}\mathcal L_X\pi\,d\lambda$, one prior factor cancels; dividing by the independent-source evidence $Z_{\rm GW}Z_X$ then gives the displayed overlap. This assumes the common and independent hypotheses use compatible source and nuisance priors. Restricting the common hypothesis to a special source class can add evidence prefactors; weak candidates also require signal–noise alternatives. The full derivation and its rate factor are given in Ashton et al. 2018, §§II.2–III, Eqs. (11), (16)–(17).

Typical shared parameters are geocentric time, sky direction, luminosity distance or redshift, and source class. Instrument-specific parameters—gamma-ray spectrum, neutrino energy, gravitational-wave masses and spins—are connected only through an explicit joint source model. Requiring exact equality of quantities that physics predicts only probabilistically makes the common-source hypothesis artificially narrow.

The accidental-coincidence rate

A useful frequentist control starts from the background rate. If unrelated external candidates passing the declared cuts arrive as a Poisson process with rate $r_X$, and the gravitational-wave candidate defines a two-sided time window of total width $\Delta t$, then

$$\Lambda_{\rm acc}=r_X\Delta t, \qquad P(N_{\rm acc}\ge1)=1-e^{-\Lambda_{\rm acc}}.$$

Sky, distance, and class consistency can reduce the expected accidental count, but only after averaging over the background population and detector exposure. Schematically,

$$\Lambda_{\rm acc} =\int_{\mathcal W} r_X(t,\Omega,z,\kappa) \epsilon_X(t,\Omega,z,\kappa)\, dt\,d\Omega\,dz\,d\kappa,$$

Here the differential $r_X(t,\Omega,z,\kappa)$ describes the population before the external-survey selection, and $\epsilon_X$ is its selection efficiency; their product is the selected rate density. Do not apply that efficiency again to a rate already measured after the same cuts. The region $\mathcal W$ is the complete coincidence region, and $\kappa$ labels source class or other cuts. Include all windows, pipelines, thresholds, and candidates examined. Choosing a narrow time or sky window after seeing the pair incurs an unreported trials factor.

Time shifts provide an empirical null when they preserve both instruments’ nonstationarity and observing gaps while destroying true coincidence. Their resolution is limited by the number of independent shifts. A very small extrapolated probability needs a justified tail model or a rate-based calculation; zero coincidences in a finite set of shifts is not zero false-association probability.

What sky and distance overlap mean

A gravitational-wave sky map is a probability density with respect to solid angle, not a binary contour. An external localization $\mathcal L_X(\Omega)$ contributes through

$$\mathcal I_\Omega =\int p_{\rm GW}(\Omega\mid d_{\rm GW}) \mathcal L_X(\Omega)\,d\Omega.$$

For a galaxy or optical transient with effectively known direction $\Omega_X$, the likelihood is sharply localized there, so this diagnostic is proportional to the gravitational-wave density at that direction. To obtain the dimensionless common-source Bayes factor, divide the corresponding posterior overlap by the shared sky prior as above. For a uniform sky prior and a precisely known direction this gives $4\pi p_{\rm GW}(\Omega_X\mid d_{\rm GW})$, when the density is per steradian. A source outside a displayed 90% contour is not impossible, and a source inside a large contour is not automatically persuasive.

Distance consistency similarly uses the conditional three-dimensional map and an external-data likelihood:

$$\mathcal I_D =\int p_{\rm GW}(D_L\mid\Omega_X,d_{\rm GW}) \mathcal L_X(d_X\mid D_L,\mathcal C)\,dD_L,$$

where $d_X$ contains the measured redshift or other distance information and $\mathcal C$ denotes cosmology and peculiar-velocity uncertainty. If an external posterior is used instead, then $\mathcal L_X\propto p_X(D_L\mid d_X,\mathcal C)/\pi_X(D_L\mid\mathcal C)$: multiplying two posteriors without dividing the external prior would make the diagnostic prior- and coordinate-dependent. A two-dimensional map cannot support a distance-overlap claim. If several galaxies or transients are tested, the association calculation must include their search selection and multiplicity.

The sky and distance integrals above illustrate separate checks; they are not automatically independent factors to multiply. A sky map generally correlates distance with direction. Use the joint posterior, or the exact factorization $p(\Omega,D_L\mid d)=p(\Omega\mid d)p(D_L\mid\Omega,d)$, and retain the matching prior factors. Independence between instruments conditional on a source does not imply independence between the inferred source parameters Ashton et al. 2018, §II.3.

Results through the stated cutoff

The comparisons below cover results available through 2026-08-18.

Evidence or interpretation	What it supports	Principal limitation or alternative
A common-source model predicts joint time, sky, distance, and class consistency	Large calibrated overlap can strongly favor association	The odds retain prior source rates and selection functions
Empirical time shifts estimate unrelated coincidences	They include real background structure when shifts preserve observing state	Finite shifts, tuning, and nonstationary rates limit tail resolution
An electromagnetic host can sharpen sky position and redshift	It enables a bright-siren likelihood	Host misidentification and counterpart-search incompleteness remain alternatives
No external counterpart is found	It constrains only the searched messenger model and survey coverage	Beaming, obscuration, depth, cadence, and sky coverage can hide a real counterpart
GW170817, GRB 170817A, and the optical transient agree in time, direction, and source interpretation	This bounded event has exceptionally strong common-source evidence LVK and gamma-ray collaborations 2017	Its association strength does not set the odds for a new event with different rates and coverage

The strongest supported reading of the GW170817 evidence is a common source; treating that case as mere chance would ignore several independently consistent measurements. The conservative lesson is not that all multimessenger pairs are ambiguous, but that each new pair needs its own rate, selection, and overlap calculation.

A reproducible association study

Record the gravitational-wave alert or final-event version, sky-map filename and hash, detector network, calibration and quality state, time convention, distance parameterization, external-alert version, localization and calibration, follow-up footprint and limiting sensitivity, source-class model, population rates, all coincidence windows, and every candidate tested. Low-latency and final products must not be mixed silently.

An independent check should inject common and unrelated pairs into the actual observing windows. Verify probability normalization, recover calibrated association odds, and test false-association coverage under rate drift. Then vary the sky prior, distance model, time window, source-rate prior, counterpart completeness, and alert version. A conclusion that changes under one credible choice should be reported as model sensitive.

This page does not recompute an association probability. The GW170817 example is tied to its published analysis; current alert products and any newer counterpart claims require a fresh official record after the cutoff.

Common pitfalls

Using contour membership as a probability. A contour is a ranked set with specified enclosed mass. Association uses probability density and the external localization over the whole sphere.

Multiplying posteriors without dividing by the shared prior. That double counts prior information and can create spurious evidence from a narrow prior.

Ignoring nondetections and follow-up selection. A counterpart search covers a time-, sky-, wavelength-, and depth-dependent region. “No counterpart” has no likelihood until that detection efficiency is modeled.

Exercises

1. Accidental time coincidences. An external survey produces selected background triggers at $r_X=0.20\,\mathrm{day}^{-1}$. A prespecified coincidence window spans $1000\,\mathrm{s}$ in total. Find the probability of at least one accidental trigger.

Solution

Convert the window to days:

$$\Delta t=\frac{1000}{86400}\,\mathrm{day} \simeq0.01157\,\mathrm{day}.$$

Thus $\Lambda_{\rm acc}=0.20(0.01157)=2.31\times10^{-3}$, and

$$P(N_{\rm acc}\ge1) =1-e^{-2.31\times10^{-3}} \simeq2.31\times10^{-3}.$$

This is the probability for one prespecified window and trigger class. Searching many gravitational-wave events, messenger streams, or window widths changes the trials accounting.

2. Posterior-overlap normalization. A shared parameter has prior density $\pi=1/10$ on an interval of width 10. Both experiments return a uniform posterior on the same subinterval of width 2. Find $\mathcal B_{C/U}^{\lambda}$.

Solution

Each posterior density is $1/2$ on the common subinterval. Therefore

$$\mathcal B_{C/U}^{\lambda} =\int_{\rm overlap}\frac{(1/2)(1/2)}{1/10}\,d\lambda =2\left(\frac{10}{4}\right)=5.$$

The data concentrate both measurements into one fifth of the prior range, producing a factor of five for shared versus independently drawn $\lambda$. This factor is not the posterior odds until source-rate prior odds and all other shared parameters are included.

References

• Ashton, Gregory, et al. “Coincident Detection Significance in Multimessenger Astronomy.” Astrophysical Journal 860 (2018): 6. DOI. Open PDF.
• LIGO Scientific Collaboration, Virgo Collaboration, Fermi Gamma-Ray Burst Monitor, and INTEGRAL. “Gravitational Waves and Gamma-Rays from a Binary Neutron Star Merger: GW170817 and GRB 170817A.” Astrophysical Journal Letters 848 (2017): L13. DOI. Open PDF.