Source Populations and Hierarchical Inference

Scrollable formula. Use Left and Right arrow keys to reveal hidden content.

Scrollable table. Use Left and Right arrow keys to reveal additional columns.

Scrollable figure. Use Left and Right arrow keys to inspect hidden content.

A gravitational-wave catalog is not a random sample of accurately measured sources. Each event has a broad likelihood, each event-level posterior was shaped by an interim prior, and the search preferentially selects some masses, spins, orientations, distances, and observing states. Population inference solves the inverse problem: infer rates and source distributions from that uncertain, selected catalog without mistaking the analysis model for an astrophysical mechanism.

Helpful background. Bayesian and Frequentist Inference develops event likelihoods, priors, evidence, coverage, and predictive checks. Catalog Construction and Selection Effects explains how triggers, thresholds, contamination, injections, and versioned releases define the sample.

Enter the population problem

Let $\theta$ be source parameters and $\Lambda$ population hyperparameters: for example, $\theta$ may be one binary’s primary mass while $\Lambda$ gives the slope and endpoints of the primary-mass distribution. For event $i$ with data $d_i$ , the likelihood $p(d_i\mid\theta)$ is combined with the normalized population density $\pi(\theta\mid\Lambda)$ through

\mathcal L_i(\Lambda) =\int d\theta\,p(d_i\mid\theta)\pi(\theta\mid\Lambda).

This integral retains measurement uncertainty. It is not reproduced by fitting a histogram of posterior medians. If posterior samples were obtained under an interim prior $\pi_0$ , the integral is evaluated by weights $\pi(\theta\mid\Lambda)/\pi_0(\theta)$ , with explicit support and effective-sample-size checks.

Selection enters through the detection probability $p_{\rm det}(\theta)$ :

\alpha(\Lambda) =\int d\theta\,p_{\rm det}(\theta)\pi(\theta\mid\Lambda).

For a Poisson source process with expected preselection count $\nu$ , the same model predicts both the detected parameter distribution and the expected count $\nu\alpha$ . The exponential $e^{-\nu\alpha}$ and event integrals form one generative likelihood. Omitting $\alpha$ favors intrinsically loud populations; inserting $p_{\rm det}$ again inside event integrals containing the selected data double counts selection.

The route through the chapter

Hierarchical Population Likelihoods derives the selected Poisson likelihood, posterior-sample recycling, injection estimator, uncertain-membership mixture, and the support conditions that make the calculation valid. It is the methodological center of the chapter.

Merger Rates and Sensitive Spacetime Volume restores dimensions. It shows how observing time, cosmological volume, source-frame time dilation, detector state, search threshold, and a population-dependent efficiency convert counts into a rate. A detector range is not automatically a catalog’s $\langle VT\rangle$ .

Mass and Mass-Ratio Distributions compares power laws, peaks, tapers, splines, Gaussian processes, and joint pairing models. It asks which broad structures survive changes in selection and functional form, while keeping pair-instability and formation interpretations outside the detector likelihood until a forward model earns them.

Spin Magnitudes and Orientations separates the comparatively well-measured effective spin $\chi_{\rm eff}$ from weakly identified component magnitudes and tilts. It explains why the spin measure, mass dependence, waveform model, and Monte Carlo support can create or erase apparent substructure.

Redshift Evolution and Cosmic History combines luminosity distance, source/detector-frame masses, cosmology, comoving volume, time dilation, and selection. A measured low-redshift slope constrains a merger-rate history; it does not uniquely recover star formation, metallicity, or delay times.

Mixture Models and Formation Channels closes the inference ladder. A statistical component may reveal inadequacy of one smooth population, but it becomes a physical channel only after identifiable, selection-aware, joint forward predictions defeat credible alternatives.

Two useful reading routes

For a first graduate route, read the likelihood and sensitive-volume pages, then mass, spin, and redshift in order. Finish with mixtures only after the selection normalization feels routine.

For an astrophysical-interpretation route, start with the mass and spin pages, use the likelihood page whenever a claim depends on posterior recycling, then read redshift and mixtures together. The corresponding formation physics belongs in Black-Hole Astrophysics; this chapter supplies the measurement interface.

One recurring mock catalog

Imagine a population with a smooth primary-mass distribution, weak positive aligned spins, and a rate rising with redshift. Draw sources from it, inject them across real detector states, retain those crossing the catalog rule, and generate broad event likelihoods. A correct analysis should recover the input within calibrated uncertainty.

That mock supports six deliberate failures:

replacing event likelihoods by point estimates confuses intrinsic spread with measurement error and prior-induced shrinkage;
omitting selection overstates intrinsically loud systems;
using the wrong interim-prior ratio shifts boundaries;
forcing mass and redshift independence moves evolution between them;
fitting a rigid one-component family can manufacture a mixture;
insufficient injection support creates precise-looking Monte Carlo features.

These failures are more instructive than a single successful fit because they identify which step of the inference each diagnostic tests.

What current results establish

The current observational example is the official GWTC-5.0 event, posterior, search-sensitivity, and injection product family, frozen at 2026-08-18. Numerical claims about rates or population features are conditional on its catalog version, event-inclusion and contamination treatment, calibration, waveform likelihoods, cosmology, population family, and injection support LVK 2026, §§2–6.

The durable conclusions are methodological: current merging compact binaries occupy structured mass and spin distributions; detector selection is substantial; and population uncertainty includes hypermodel and numerical uncertainty, not only credible bands inside one fit. The exact rates, peaks, slopes, mixture fractions, and physical interpretations are dated evidence and must be refreshed with new releases.

Common misconceptions

“Many events wash out priors.” Pooling can instead amplify a shared misspecification. A narrow hyperposterior within the wrong population family is confidently wrong.

“Selection is only a distance correction.” Detectability depends jointly on masses, spins, orientation, sky position, redshift, detector epoch, waveform, pipeline, and threshold.

“A fitted cluster is a formation channel.” A component is a density term. Causal identification requires joint forward predictions, stable identifiability, and alternatives.

“A posterior predictive fit proves the model.” It checks chosen summaries against draws from the fitted model. Adversarial summaries and out-of-family mocks are needed to expose failures the model was built to hide.

Review the chapter

1. Where selection belongs. Why does a population-shape likelihood conditioned on $N$ contain a factor $\alpha(\Lambda)^{-N}$ ?

Solution

The observed events are drawn from the population conditioned on passing the catalog rule. Conditioning normalizes the detected density by the population-averaged efficiency $\alpha(\Lambda)$ . Each of $N$ event factors contributes one inverse normalization. In the full Poisson likelihood, the same physics appears through the expected count $\nu\alpha$ ; marginalizing $\nu$ with a scale prior can yield the conditional factor for $N\ge1$ .

2. An apparently decisive component. A high-mass mixture component has a narrow branching-fraction posterior but the injection effective sample size is tiny there. What follows?