Alongside this paper, the archive holds studies that also reconstruct black hole populations from gravitational waves: some use parametric models with peaks and breaks, others build nonparametric reconstructions or apply machine learning to speed up the analysis. The shared thrust of these neighbors is to extract from the growing LIGO–Virgo–KAGRA catalogs the shape of the mass distribution and its possible dependence on redshift. This article sits in the same stream but tackles the problem differently: instead of a fixed functional form, it trains a neural density directly through hierarchical likelihood, which lets it capture structure that was not put in by hand.
The population density in redshift, primary mass and secondary mass coordinates is modeled by an autoregressive neural spline flow. The flow parameters are optimized through hierarchical likelihood: for each event the integral over hidden parameters is estimated from its posterior samples, and the selection factor is computed from detected injections from the campaign. Training proceeds by stochastic gradient descent over mini-batches of events, with repeated runs and cross-validation. The final density emerges as a weighted ensemble of checkpoints from the converging part of the trajectories. All conclusions are restricted to the region supported by detected injections, which is constructed by conformal prediction.
On two simulated catalogs that differ only in whether the Gaussian peak of primary masses evolves with redshift, the reconstruction reproduces the injected mass structure within the supported region. The mutual information diagnostic for the stationary model gives a median of −0.0021 with a spread from −0.0065 to +0.0078, while for the evolving model it gives +0.0122 with a spread from +0.0070 to +0.0097, clearly separating the two scenarios. Applied to 231 events from GWTC-5.0, it reveals features near 10 and 35 solar masses, consistent with the collaboration's official analysis. The diagnostic value for the real data is +0.0026 with a spread from −0.0043 to +0.0049, showing no positive shift characteristic of evolution.
The work stands apart from neighboring studies in that it does not fix the functional form of the mass distribution in advance, but lets the data determine it through a flexible neural density. Unlike earlier attempts to use flows without selection correction, here selection effects are accounted for explicitly and training remains stable. This is the first test of such an approach on GWTC-5.0 with simultaneous control over redshift dependence, which provides a more reliable foundation for future cosmological inferences.
A natural extension would be to include cosmological parameters directly in the inference, so as to jointly constrain the population and the distance-redshift relation. The method can be transferred to other problems where objects are observed with noise and selection, such as the demographics of exoplanets or supernovae. Further improvements to the flow architecture and training strategy could boost reconstruction accuracy as catalogs grow.
The results will affect gravitational-wave astrophysics and cosmology, especially dark siren methods, where the shape of the mass distribution feeds directly into constraints on cosmological parameters.
The next step will be to apply the method to future LIGO–Virgo–KAGRA catalogs with more events, and to adapt it for joint inference of the population and cosmology.
Gravitational-wave astrophysics still faces an open question about the origin of the peaks in the black hole mass distribution and their possible evolution with redshift. This work offers a tool that can test such hypotheses without prior assumptions about the shape of the distribution.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors train a normalizing flow for the population density of binary black holes directly through hierarchical likelihood, accounting for event uncertainties via posterior samples and for selection effects via injections. The method is validated on simulated catalogs with and without mass-peak evolution, then applied to GWTC-5.0, where it reproduces features near 10 and 35 solar masses and finds no evidence for evolution of the primary masses.
Where its strength is
A flexible neural density with no fixed functional form, trained with explicit selection correction, can capture distributional structure in the mass distribution that was not built in beforehand, and can robustly estimate the dependence on redshift.
Where it could go next
Worth trying
Why this may matter
This approach could become a standard tool for analyzing growing gravitational-wave catalogs, making it possible to identify subtle features of the black hole population and their evolution without prior assumptions, which matters for understanding the formation of binary systems and for cosmology.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.56, against a usual 0.64. It looks like few have drilled nearby.
The injection campaign for GWTC-5.0 generated 1.57 billion signals, of which only 1.34 million passed the detection threshold, corresponding to roughly one survivor per thousand generated.