Pith. sign in

REVIEW 2 major objections 5 minor 88 references

Per-event normalizing flows can keep gravitational-wave population inference trustworthy as catalogs grow past 300 mergers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 11:30 UTC pith:O2SNPOY5

load-bearing objection The variance-growth problem is real and clearly demonstrated, but the paper's headline fix relies on an unvalidated MCMC decorrelation step that could inflate the apparent gains. the 2 major comments →

arxiv 2606.14229 v2 pith:O2SNPOY5 submitted 2026-06-12 astro-ph.HE

Fortifying gravitational-wave population inference with normalizing flows

Christian Adamcewicz , Hugh McDougall , Paul D. Lasky , Eric Thrane This is my paper
classification astro-ph.HE
keywords gravitational-wave population inferencehierarchical Bayesian inferenceMonte Carlo varianceposterior samplesnormalizing flowsimportance samplingcatalog growthspin distribution
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Gravitational-wave astronomers infer the properties of black hole populations by stacking the posterior samples of individual events, but the stacking is only as good as the sample counts. This paper argues that once a catalog reaches roughly 300 mergers, the standard few tens of thousands of samples per event are too few: the Monte Carlo variance of the population likelihood grows with catalog size, and the conventional cutoff on that variance begins rejecting and biasing the very hyper-parameter space under study. The paper proves the point on a mock catalog of 300–500 binary black holes with realistic, sharply featured mass and spin distributions, then shows that training a normalizing flow on each event's existing posterior samples generates essentially unlimited cheap samples. With 100,000 samples per event the variance distribution moves well below threshold and the true injected population is recovered. The consequence, if correct, is a simple computational fix that keeps population measurements trustworthy as next-generation observatories multiply catalog size.

Core claim

The paper's central claim is that representing each gravitational-wave event with the order of 10^4 posterior samples—the current norm—renders hierarchical population inference numerically unreliable once the catalog grows past about 300 events. In a mock catalog built from an astrophysically motivated model with sharp peaks and a gap in the black-hole mass spectrum and realistic spin distributions, the authors show that the population-likelihood variance σ²_L rails against the conventional threshold σ²_L=1, and that the inferred spin hyper-parameters become biased. They propose representing each event with a per-event normalizing flow trained on the existing posterior (or nested) samples; t

What carries the argument

The central quantity is the population-likelihood variance σ²_L, the Monte Carlo variance accumulated when importance-sampling each event's posterior samples through the population model; it grows with the number of events N and falls with samples per event n, and a conventional threshold σ²_L=1 is used to veto hyper-parameter space. The solution machinery is a per-event normalizing flow—a neural spline flow trained (with nested-sampler weights and Kullback-Leibler loss) on each event's existing 13-dimensional posterior samples—which can be sampled a million times in seconds. The flow's samples are given importance weights to match the exact fiducial posterior, and a redraw-and-MCMC-walk ste

Load-bearing premise

The demonstration that 100,000 samples per event reduce the population-likelihood variance relies on a heuristic short random-walk step whose convergence to independent draws is not rigorously verified; if the final batches are effectively smaller than claimed, the variance reduction is overstated.

What would settle it

Compute the effective sample size of the final 10^5 unweighted sample batches from the paper's own 300-event mock catalog using standard autocorrelation diagnostics; if the effective size falls below roughly 5×10^4, the claimed variance reduction and the σ²_L distribution in Fig. 8 would not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • With ≳300 events, population analyses using ~10^4 samples per event will produce biased spin hyper-parameters; the paper's mock-catalog test shows the σ²_L distribution rails against the σ²_L=1 threshold.
  • Weighted nested samples, already produced by standard parameter estimation, offer a short-term accuracy improvement without rerunning analyses, but this reprieve ends as catalogs grow.
  • Per-event normalizing flows trained on existing samples can supply arbitrarily many cheap posterior draws; with 10^5 samples per event the population-likelihood variance drops well below threshold and the true population is recovered in both 300- and 500-event mock catalogs.
  • Flow-generated samples cost roughly 80% fewer likelihood evaluations per sample than nested sampling and parallelize more easily, making the fix computationally practical.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the per-event flow representation becomes standard, population inference could be reframed as a train-once-per-event process, so that future catalog updates only require drawing new samples from stored flows instead of rerunning expensive parameter estimation.
  • The variance problem also affects selection-effect integrals; the flow approach could plausibly be extended to emulate detection probabilities, removing the injection-set reweighting uncertainty that the paper notes remains even when σ²_L is controlled.
  • A direct test on real current-catalog data would quantify present bias: compare population posteriors computed with ~10^4 versus 10^5 samples per event; where they differ, the standard analysis is already unreliable.
  • The paper's brief encounter with extreme weights suggests the flows under-represent isolated high-dimensional pockets; combining flow draws with a defensive mixture that touches the prior could tame the rejection-sampling inefficiency they observed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper addresses Monte Carlo noise in hierarchical Bayesian population inference caused by representing each gravitational-wave event with a finite number of posterior samples. Section II derives the importance-sampling form of the population likelihood and defines the variance σ_L^2 that grows with the number of events N and with the spread of population-reweighting weights, and falls with the number of per-event samples n. Section III builds one mock catalog of 300 BBH mergers with LVK-like parameter estimation, showing that with ≈1×10^4–2×10^4 posterior samples per event the distribution of σ_L^2 rails against the σ_L^2=1 threshold and biases the inferred spin hyper-parameters; using only 2×10^3 samples is worse. Section IV proposes short-term improvements: keeping all available posterior samples through NaN-padded rectangular arrays, and using weighted nested samples produced by Dynesty. Section V trains per-event normalizing flows on nested samples, iteratively re-trains them with reweighted NPE samples, draws 10^6 weighted samples, converts them to 10^5 unweighted samples via a resample-with-replacement plus short MCMC walk, and reports that supplementing PE samples to 5×10^4 or 10^5 samples per event shifts the σ_L^2 distribution away from threshold. An appendix repeats the demonstration for 500 events. The paper concludes that current sample sizes are or will soon be inadequate for ≳300-event catalogs, and proposes normalizing flows as a longer-term remedy.

Significance. If the central result holds, this is a timely and actionable methods contribution for LVK population analyses: it quantifies a Monte Carlo variance problem that is expected to worsen as the catalog grows, and it offers a concrete computational fix. The variance-growth argument in Sec. II is parameter-free and follows from standard importance-sampling theory; the mock-catalog study includes realistic PE, selection effects, and production-like Dynesty settings; and the authors release code for the padded GWPopulation arrays and for the normalizing-flow pipeline (Refs. [49,50,62]). The demonstration that nested samples already provide some relief, and that NPE-generated samples can be reweighted against the exact likelihood, are useful. The main caveat is that the headline remedy depends on an unvalidated decorrelation step (Sec. V.D), as discussed below; with that issue addressed, this would be a strong methods paper.

major comments (2)
  1. [§V.D and Eq. (5)] Eq. (5) assumes n independent posterior draws. In the resample-move step, the 10^5 unweighted samples are obtained by resampling with replacement from η=10^6 weighted NPE samples and then running a 50-step MCMC walk (up to ≈2000 steps for low-efficiency events). The manuscript itself states in §V.D that “there is no rigorous method that we know of to determine if the initially duplicate samples have been sufficiently decoupled.” The adopted distance-based threshold is a heuristic mixing diagnostic, and no effective sample size or autocorrelation time is reported for the final unweighted batches. Since σ_L^2 scales as 1/n_eff, the σ_L^2 distributions in Figs. 7, 8, and 10 could be artificially far from the threshold if the chains have not fully mixed. Please provide an ESS/autocorrelation analysis of the decorrelation walk, or validate the method by comparing σ_L^2 obtained from independe
  2. [§III, Figs. 1–2] The demonstration that ≳300 events with 10^4–2×10^4 samples “results in unreliable inference” is based on a single mock realization. Because the occurrence and severity of σ_L^2 railing depend on the random catalog realization, noise realization, and PE outcomes, one draw cannot establish the frequency or typical magnitude of the failure. Repeating the injection/recovery on several independent mock catalogs, or at least resampling the PE samples, would place error bars on the fraction of hyper-posterior samples near threshold and on the differences between the gray and red contours. This does not undermine the variance-scaling argument in Sec. II, but it is needed to support the quantitative strength of the demonstrated effect.
minor comments (5)
  1. [§V.C, Eq. (16)] The inverse-variance combination of the two KL estimates treats the nested-sample and NPE-sample estimators as independent. Since the NPE was trained on the nested samples, the two estimates are correlated to some degree. The combination remains unbiased if the weights sum to one, but it may not be optimally weighted. A brief comment or test of sensitivity to this choice would be useful.
  2. [Abstract and §V] The claim of “median ≈80% fewer likelihood evaluations per sample” should specify the comparison denominator: per final posterior sample, per nested-sample likelihood evaluation, and whether the likelihood evaluations used in the MCMC decorrelation walk are included. Section VI notes that the walk can dominate the total cost, so this distinction matters for the computational claim.
  3. [§IV.A] With NaN padding, the effective n in Eq. (4) differs from the rectangular array length. The text says this is handled, but it would help to write the expectation explicitly as depending on n_i for each event.
  4. [§Fig. 1] The σ_L^2=1 threshold is mentioned in the text but not labeled in the figures. Adding a vertical dashed line with a label would make the railing effect immediately visible.
  5. [Appendix A] The extended 500-event catalog reuses the same 2.5×10^7 injection set. The text asserts this still adequately mitigates selection-effect variance; a one-sentence quantitative justification (e.g., the resulting variance contribution relative to Eq. 5) would strengthen the appendix.

Circularity Check

0 steps flagged

No significant circularity: central variance-growth result follows from analytic importance sampling and is measured, not imposed; self-citations are incidental.

full rationale

The paper's derivation chain is self-contained. The variance-growth result follows from the analytic importance-sampling variance in Eq. (5), with no fitted constants; the mock catalog is generated from an independently specified population, and the population fit is a controlled recovery test, not a derivation of the conclusion from the conclusion. The NPE samples are reweighted against the exact likelihood and prior via Eq. (12), so flow approximation error does not act as a fitted 'prediction'. The success metric σ²_L is computed from actual sample batches using Eq. (5), not imposed. The short-term nested-sample weighting in Eqs. (8)-(9) is standard importance sampling. Self-citations (e.g., Refs. [25]-[28], [32]) are incidental and support choices of fiducial-prior and mock-mass models, not the paper's central conclusions. The skeptic's concern about the resample-move MCMC decorrelation in Sec. V.D is a validity/robustness limitation, explicitly acknowledged by the authors ('There is no rigorous method that we know of...'), but it does not reduce the claim to its inputs and therefore is not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No invented physical entities. The central claim rests on standard importance-sampling theory, a representative mock catalog, a heuristic decorrelation step the paper itself flags as unvalidated, and hand-tuned flow hyperparameters. The mock population models are drawn from the authors' earlier published works, which are cited but do not by themselves make the central claim circular.

free parameters (3)
  • NPE architecture (16 neural spline layers, MLP width 128 depth 8, learning rate 1e-4, batch size 128/2048) = 16 layers / 128×8 MLP / lr 1e-4 / batch 128, 2048
    Authors state they experimented and chose the best-performing architecture across the mock catalog (Sec. V.A); the median 24% efficiency and the '≈80% fewer likelihood evals' headline depend on these hand-tuned choices.
  • Iterative training rounds and retraining threshold = 2 iterations; extra 1–2 iterations for events with NPE efficiency <1%
    Ad hoc stopping rule for flow refinement (Sec. V.C); affects the final NPE efficiency distribution but not the variance-growth math.
  • NPE sampling batch and redraw size = η = 1e6 samples; redraw 1e5 with replacement; 50-step MCMC walk
    Computational budget choices set by hand (Secs. V.B, V.D); the claim that 5e4–1e5 samples per event reduce variance depends on these numbers being sufficient post-decorrelation.
axioms (4)
  • domain assumption The importance-sampling variance formula σ²_L in Eq. 5 is the correct error model for the population likelihood, and the σ²_L=1 threshold is an appropriate reliability diagnostic.
    Invoked in Secs. II–III to declare the 1e4–2e4-sample inference 'unreliable'; the paper itself notes the threshold 'may be overly strict in some cases' citing Ref. [20].
  • domain assumption The mock catalog (2-detector design-sensitivity noise, SNR≥10, IMRPhenomXP, analytic marginalization over distance/time/phase, strongly-parameterized spin/mass models) is representative enough that conclusions transfer to real LVK catalogs.
    The 300-event failure point (Sec. III) is derived from this one synthetic setup; selection effects are included via a found-injection set, but waveform systematics, calibration error, and higher-dimensional posteriors are absent.
  • ad hoc to paper The redraw-with-replacement plus short MCMC walk yields effectively independent samples from the fiducial posterior.
    Sec. V.D states 'There is no rigorous method that we know of to determine if the initially duplicate samples have been sufficiently decoupled'; the paper uses an ad hoc distance threshold instead.
  • ad hoc to paper The inverse-variance combination of the two KL estimates in Eq. 16 treats the nested-sample and NPE-sample estimators as independent; in reality the NPE was trained on the nested samples, so the two error estimates are correlated.
    Iterative training loss (Sec. V.C) neglects this dependence; no covariance term is included.

pith-pipeline@v1.3.0-alltime-deepseek · 22032 in / 16831 out tokens · 179080 ms · 2026-08-02T11:30:08.161209+00:00 · methodology

0 comments
read the original abstract

As the LIGO-Virgo-KAGRA collaboration's (LVK's) gravitational-wave transient catalog grows, we are learning a wealth of information from the population properties of binary black hole mergers. Events in the catalog are represented with posterior samples describing the astrophysical parameters for each event. Population studies combine these samples to measure the distribution of astrophysical parameters such as black hole masses and spins. However, the posterior-sample representation of each event is only approximate. We construct a mock population with masses drawn from an astrophysically-motivated distribution with sharp features. Using this, we demonstrate that when $\gtrsim 300$ events are combined, even with each event's posterior represented by $1 \times 10^4 {-} 2 \times 10^4$ samples, the numerical error can become large enough that the resulting population inference is unreliable. We consider two solutions. In the short term, we show that nested samples (already produced by LVK analyses) can be used to more accurately describe each event in population studies. But this will only grant a temporary reprieve until the nested-sample representation becomes inadequate. In the longer term, we propose to represent each event with a normalizing flow. In order to represent each event with sufficient accuracy, each normalizing flow can be used to generate an arbitrarily large number of new posterior samples with a significantly reduced computational cost relative to traditional sampling methods. When compared to nested sampling, our normalizing flows produce posterior draws with a median of $\approx 80\%$ fewer likelihood evaluations per sample, while also providing greater opportunity for parallelization. We believe refinement of normalizing flow architectures and training techniques in future works could further reduce this per-sample cost significantly.

Figures

Figures reproduced from arXiv: 2606.14229 by Christian Adamcewicz, Eric Thrane, Hugh McDougall, Paul D. Lasky.

Figure 1
Figure 1. Figure 1: FIG. 1: Posteriors for hyper-parameters governing the spin distribution, along with the distribution of likelihood [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Posteriors for spin distribution hyper-parameters, along with the distribution of likelihood variances [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Comparison between the distributions of [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Posteriors for spin distribution hyper-parameters, along with the distribution of variances in the population [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Diagram of our normalizing flow architecture. We start with a zero-mean, unit-variance normal base [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Bar chart of the normalizing flow performance for each of the 300 mock events. The height of each blue and [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Posteriors for spin distribution hyper-parameters, along with the distribution of associated variances in the [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Posteriors for spin distribution hyper-parameters, along with the distribution of associated variances in the [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Extended bar chart of the normalizing flow performance for each of the 500 mock events. As in Fig. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10: Posteriors for spin distribution hyper-parameters, along with the distribution of associated variances in the [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

88 extracted references · 65 linked inside Pith

  1. [1]

    Increasing the number of eventsNincreases the uncertainty,

  2. [2]

    Increasing variance between the weightsw(θ k i |Λ) increases the uncertainty, and

  3. [3]

    strongly-parameterized

    Increasing the number of fiducial samplesnde- creases the uncertainty. Thus, as the size of the catalog grows, we invariably in- crease the uncertainty due to Monte-Carlo integration in the population likelihood. In practice, we have no control over the variance in weights, which becomes more prob- lematic in models with sharp features. In theory, one cou...

    2048

  4. [4]

    Advanced LIGO detector performance in the fourth observing run,

    E. Capoteet al., “Advanced LIGO detector performance in the fourth observing run,” Phys. Rev. D111, 062002 (2025), arXiv:2411.14607 [gr-qc]

    Pith/arXiv arXiv 2025

  5. [5]

    LIGO Detector Characteriza- tion in the first half of the fourth Observing run,

    S. Soniet al.(LIGO), “LIGO Detector Characteriza- tion in the first half of the fourth Observing run,” Class. Quant. Grav.42, 085016 (2025), arXiv:2409.02831 [astro- ph.IM]

    arXiv 2025

  6. [6]

    Sensitivity and per- formance of the Advanced LIGO detectors in the third observing run,

    Aaron Buikemaet al.(aLIGO), “Sensitivity and per- formance of the Advanced LIGO detectors in the third observing run,” Phys. Rev. D102, 062003 (2020), arXiv:2008.01301 [astro-ph.IM]

    Pith/arXiv arXiv 2020

  7. [7]

    Broad- band Quantum Enhancement of the LIGO Detectors with Frequency-Dependent Squeezing,

    D. Ganapathyet al.(LIGO O4 Detector), “Broad- band Quantum Enhancement of the LIGO Detectors with Frequency-Dependent Squeezing,” Phys. Rev. X13, 041021 (2023)

    2023

  8. [8]

    Squeezing the quantum noise of a gravitational-wave de- tector below the standard quantum limit,

    Wenxuan Jiaet al.(members of the LIGO Scientific†), “Squeezing the quantum noise of a gravitational-wave de- tector below the standard quantum limit,” Science385, 1318 (2024), arXiv:2404.14569 [gr-qc]

    Pith/arXiv arXiv 2024

  9. [9]

    Advanced LIGO,

    J. Aasiet al.(LIGO Scientific), “Advanced LIGO,” Class. 17 0 100 200 300 400 500 Event number 0.01% 0.1% 1.0% 10.0% 100.0%ϵ [%] 100 1000 10000 100000 1000000 neff FIG. 9: Extended bar chart of the normalizing flow performance for each of the 500 mock events. As in Fig. 6, the height of each blue and black bar indicates the effective sample sizen eff of a b...

    Pith/arXiv arXiv 2015

  10. [10]

    Advanced Virgo: a second- generation interferometric gravitational wave detector,

    F. Acerneseet al.(VIRGO), “Advanced Virgo: a second- generation interferometric gravitational wave detector,” Class. Quant. Grav.32, 024001 (2015), arXiv:1408.3978 [gr-qc]

    Pith/arXiv arXiv 2015

  11. [11]

    Overview of KAGRA: De- tector design and construction history,

    T. Akutsuet al.(KAGRA), “Overview of KAGRA: De- tector design and construction history,” PTEP2021, 05A101 (2021), arXiv:2005.05574 [physics.ins-det]

    arXiv 2021

  12. [12]

    GWTC-4.0: An Introduction to Version 4.0 of the Gravitational-Wave Transient Catalog,

    A. G. Abacet al.(LIGO Scientific, KAGRA, VIRGO), “GWTC-4.0: An Introduction to Version 4.0 of the Gravitational-Wave Transient Catalog,” Astrophys. J. Lett.995, L18 (2025), arXiv:2508.18080 [gr-qc]

    Pith/arXiv arXiv 2025

  13. [13]

    GWTC-4.0: Methods for Identifying and Characterizing Gravitational-wave Transients,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KA- GRA), “GWTC-4.0: Methods for Identifying and Characterizing Gravitational-wave Transients,” (2025), arXiv:2508.18081 [gr-qc]

    Pith/arXiv arXiv 2025

  14. [14]

    GWTC-4.0: Updating the Gravitational-Wave Tran- sient Catalog with Observations from the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KAGRA), “GWTC-4.0: Updating the Gravitational-Wave Tran- sient Catalog with Observations from the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run,” (2025), arXiv:2508.18082 [gr-qc]

    Pith/arXiv arXiv 2025

  15. [15]

    Open Data from LIGO, Virgo, and KAGRA through the First Part of the Fourth Observing Run,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KAGRA), “Open Data from LIGO, Virgo, and KAGRA through the First Part of the Fourth Observing Run,” (2025), arXiv:2508.18079 [gr-qc]

    Pith/arXiv arXiv 2025

  16. [16]

    GWTC-4.0: Population Properties of Merging Compact Binaries,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KAGRA), “GWTC-4.0: Population Properties of Merging Compact Binaries,” (2025), arXiv:2508.18083 [astro-ph.HE]

    Pith/arXiv arXiv 2025

  17. [17]

    Observed Gravitational-Wave Popula- tions,

    T. A. Callister, “Observed Gravitational-Wave Popula- tions,” (2024), arXiv:2410.19145 [astro-ph.HE]

    Pith/arXiv arXiv 2024

  18. [18]

    GWTC-4.0: Constraints on the Cosmic Expansion Rate and Modified Gravitational-wave Propagation,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KAGRA), “GWTC-4.0: Constraints on the Cosmic Expansion Rate and Modified Gravitational-wave Propagation,” (2025), arXiv:2509.04348 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  19. [19]

    The Hitchhiker’s Guide to the Galaxy Catalog Approach for Dark Siren Gravitational-wave Cosmology,

    Jonathan R. Gairet al., “The Hitchhiker’s Guide to the Galaxy Catalog Approach for Dark Siren Gravitational-wave Cosmology,” Astron. J.166, 22 (2023), arXiv:2212.08694 [gr-qc]

    Pith/arXiv arXiv 2023

  20. [20]

    GWTC-4.0: Tests of General Relativity. I. Overview and General Tests,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KAGRA), “GWTC-4.0: Tests of General Relativity. I. Overview and General Tests,” (2026), arXiv:2603.19019 [gr-qc]

    Pith/arXiv arXiv 2026

  21. [21]

    GWTC-4.0: Tests of General Relativity. II. Parameter- ized Tests,

    A. G. Abacet al.(LIGO Scientific, VIRGO, KAGRA), “GWTC-4.0: Tests of General Relativity. II. Parameter- ized Tests,” (2026), arXiv:2603.19020 [gr-qc]

    Pith/arXiv arXiv 2026

  22. [22]

    An introduction to Bayesian inference in gravitational-wave astronomy: pa- rameter estimation, model selection, and hierarchical models,

    Eric Thrane and Colm Talbot, “An introduction to Bayesian inference in gravitational-wave astronomy: pa- rameter estimation, model selection, and hierarchical models,” Publ. Astron. Soc. Austral.36, e010 (2019), [Erratum: Publ.Astron.Soc.Austral. 37, e036 (2020)], arXiv:1809.02293 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  23. [23]

    When (not) to trust Monte Carlo approximations for hierarchical Bayesian in- ference,

    Jack Heinzel and Salvatore Vitale, “When (not) to trust Monte Carlo approximations for hierarchical Bayesian in- ference,” (2025), arXiv:2509.07221 [astro-ph.HE]

    Pith/arXiv arXiv 2025

  24. [24]

    Growing pains: un- derstanding the impact of likelihood uncertainty on hi- erarchical Bayesian inference for gravitational-wave as- tronomy,

    Colm Talbot and Jacob Golomb, “Growing pains: un- derstanding the impact of likelihood uncertainty on hi- erarchical Bayesian inference for gravitational-wave as- tronomy,” Mon. Not. Roy. Astron. Soc.526, 3495–3503 (2023), arXiv:2304.06138 [astro-ph.IM]

    Pith/arXiv arXiv 2023

  25. [25]

    Precision Requirements for Monte Carlo Sums within Hierarchical Bayesian Infer- ence,

    Reed Essick and Will Farr, “Precision Requirements for Monte Carlo Sums within Hierarchical Bayesian Infer- ence,” (2022), arXiv:2204.00461 [astro-ph.IM]

    Pith/arXiv arXiv 2022

  26. [26]

    Accuracy Requirements for Empirically- Measured Selection Functions,

    Will M. Farr, “Accuracy Requirements for Empirically- Measured Selection Functions,” Research Notes of the AAS3, 66 (2019), arXiv:1904.10879 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  27. [27]

    Cosmic Explorer: The U.S. Contribution to Gravitational-Wave Astronomy be- yond LIGO,

    David Reitzeet al., “Cosmic Explorer: The U.S. Contribution to Gravitational-Wave Astronomy be- yond LIGO,” Bull. Am. Astron. Soc.51, 035 (2019), arXiv:1907.04833 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  28. [28]

    Do Both Black Holes Spin in Merg- ing Binaries? Evidence from GWTC-4 and Astro- physical Implications,

    Christian Adamcewicz, Nir Guttman, Paul D. Lasky, and Eric Thrane, “Do Both Black Holes Spin in Merg- ing Binaries? Evidence from GWTC-4 and Astro- physical Implications,” Astrophys. J.994, 261 (2025), 18 0.08 0.16 0.24 σχ 0.45 0.60 0.75 0.90 ξt 0.4 0.6 0.8 µt 0.25 0.50 0.75 1.00 σt 0.1 0.2 0.3 µχ 0.4 0.6 0.8 σ2 L 0.08 0.16 0.24 σχ 0.45 0.60 0.75 0.90 ξt ...

    arXiv 2025

  29. [29]

    Which Black Hole Is Spin- ning? Probing the Origin of Black Hole Spin with Grav- itational Waves,

    Christian Adamcewicz, Shanika Galaudage, Paul D. Lasky, and Eric Thrane, “Which Black Hole Is Spin- ning? Probing the Origin of Black Hole Spin with Grav- itational Waves,” Astrophys. J. Lett.964, L6 (2024), arXiv:2311.05182 [astro-ph.HE]

    Pith/arXiv arXiv 2024

  30. [30]

    Pop- ulation properties of spinning black holes using the gravitational-wave transient catalog 3,

    Hui Tong, Shanika Galaudage, and Eric Thrane, “Pop- ulation properties of spinning black holes using the gravitational-wave transient catalog 3,” Phys. Rev. D 106, 103019 (2022), arXiv:2209.02206 [astro-ph.HE]. 19

    Pith/arXiv arXiv 2022

  31. [31]

    Building Better Spin Mod- els for Merging Binary Black Holes: Evidence for Non- spinning and Rapidly Spinning Nearly Aligned Subpop- ulations,

    Shanika Galaudageet al., “Building Better Spin Mod- els for Merging Binary Black Holes: Evidence for Non- spinning and Rapidly Spinning Nearly Aligned Subpop- ulations,” Astrophys. J. Lett.921, L15 (2021), [Erra- tum: Astrophys.J.Lett. 936, L18 (2022), Erratum: As- trophys.J. 936, L18 (2022)], arXiv:2109.02424 [gr-qc]

    Pith/arXiv arXiv 2021

  32. [32]

    Determining the popu- lation properties of spinning black holes,

    Colm Talbot and Eric Thrane, “Determining the popu- lation properties of spinning black holes,” Phys. Rev. D 96, 023012 (2017), arXiv:1704.08370 [astro-ph.HE]

    Pith/arXiv arXiv 2017

  33. [33]

    Spin it as you like: The (lack of a) measurement of the spin tilt distribution with LIGO-Virgo-KAGRA bi- nary black holes,

    Salvatore Vitale, Sylvia Biscoveanu, and Colm Talbot, “Spin it as you like: The (lack of a) measurement of the spin tilt distribution with LIGO-Virgo-KAGRA bi- nary black holes,” Astron. Astrophys.668, L2 (2022), arXiv:2209.06978 [astro-ph.HE]

    Pith/arXiv arXiv 2022

  34. [34]

    Does the Black Hole Merger Rate Evolve with Redshift?

    Maya Fishbach, Daniel E. Holz, and Will M. Farr, “Does the Black Hole Merger Rate Evolve with Redshift?” Astrophys. J. Lett.863, L41 (2018), arXiv:1805.10270 [astro-ph.HE]

    Pith/arXiv arXiv 2018

  35. [35]

    No Evidence for a Dip in the Binary Black Hole Mass Spectrum,

    Christian Adamcewicz, Paul D. Lasky, Eric Thrane, and Ilya Mandel, “No Evidence for a Dip in the Binary Black Hole Mass Spectrum,” Astrophys. J.975, 253 (2024), arXiv:2406.11111 [astro-ph.HE]

    Pith/arXiv arXiv 2024

  36. [36]

    Low-mass failed supernovae and the 10M ⊙ peak in the merging black hole mass distribution,

    Isaac Legred, Jacob Golomb, and Katerina Chatziioan- nou, “Low-mass failed supernovae and the 10M ⊙ peak in the merging black hole mass distribution,” (2026), arXiv:2604.01420 [astro-ph.HE]

    arXiv 2026

  37. [37]

    Compactness peaks: An astrophysical interpretation of the mass dis- tribution of merging binary black holes,

    Shanika Galaudage and Astrid Lamberts, “Compactness peaks: An astrophysical interpretation of the mass dis- tribution of merging binary black holes,” Astron. Astro- phys.694, A186 (2025), arXiv:2407.17561 [astro-ph.HE]

    Pith/arXiv arXiv 2025

  38. [38]

    Bimodal Black Hole Mass Distribution and Chirp Masses of Binary Black Hole Mergers,

    Fabian R. N. Schneider, Philipp Podsiadlowski, and Eva Laplace, “Bimodal Black Hole Mass Distribution and Chirp Masses of Binary Black Hole Mergers,” Astrophys. J. Lett.950, L9 (2023), arXiv:2305.02380 [astro-ph.HE]

    Pith/arXiv arXiv 2023

  39. [39]

    Pre-supernova evolution, compact object masses and ex- plosion properties of stripped binary stars,

    F. R. N. Schneider, Ph. Podsiadlowski, and B. M¨ uller, “Pre-supernova evolution, compact object masses and ex- plosion properties of stripped binary stars,” Astron. As- trophys.645, A5 (2021), arXiv:2008.08599 [astro-ph.SR]

    Pith/arXiv arXiv 2021

  40. [40]

    Bayesian inference for com- pact binary coalescences with bilby: validation and appli- cation to the first LIGO–Virgo gravitational-wave tran- sient catalogue,

    I. M. Romero-Shawet al., “Bayesian inference for com- pact binary coalescences with bilby: validation and appli- cation to the first LIGO–Virgo gravitational-wave tran- sient catalogue,” Mon. Not. Roy. Astron. Soc.499, 3295– 3319 (2020), arXiv:2006.00714 [astro-ph.IM]

    Pith/arXiv arXiv 2020

  41. [41]

    Prospects for observing and localizing gravitational- wave transients with Advanced LIGO, Advanced Virgo and KAGRA,

    B. P. Abbottet al.(KAGRA, LIGO Scientific, Virgo), “Prospects for observing and localizing gravitational- wave transients with Advanced LIGO, Advanced Virgo and KAGRA,” Living Rev. Rel.19, 1 (2016), arXiv:1304.0670 [gr-qc]

    Pith/arXiv arXiv 2016

  42. [42]

    Ex- tracting distribution parameters from multiple uncer- tain observations with selection biases,

    Ilya Mandel, Will M. Farr, and Jonathan R. Gair, “Ex- tracting distribution parameters from multiple uncer- tain observations with selection biases,” Mon. Not. Roy. Astron. Soc.486, 1086–1093 (2019), arXiv:1809.02063 [physics.data-an]

    Pith/arXiv arXiv 2019

  43. [43]

    Estimation of the Sensitive Volume for Gravitational-wave Source Populations Using Weighted Monte Carlo Integration,

    Vaibhav Tiwari, “Estimation of the Sensitive Volume for Gravitational-wave Source Populations Using Weighted Monte Carlo Integration,” Class. Quant. Grav.35, 145009 (2018), arXiv:1712.00482 [astro-ph.HE]

    Pith/arXiv arXiv 2018

  44. [44]

    Computationally efficient models for the dominant and subdominant harmonic modes of precessing binary black holes,

    Geraint Prattenet al., “Computationally efficient models for the dominant and subdominant harmonic modes of precessing binary black holes,” Phys. Rev. D103, 104056 (2021), arXiv:2004.06503 [gr-qc]

    Pith/arXiv arXiv 2021

  45. [45]

    BILBY: A user-friendly Bayesian inference library for gravitational-wave astronomy,

    Gregory Ashtonet al., “BILBY: A user-friendly Bayesian inference library for gravitational-wave astronomy,” As- trophys. J. Suppl.241, 27 (2019), arXiv:1811.02042 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  46. [46]

    Parameter estimation for compact bina- ries with ground-based gravitational-wave observations using the LALInference software library,

    J. Veitchet al., “Parameter estimation for compact bina- ries with ground-based gravitational-wave observations using the LALInference software library,” Phys. Rev. D 91, 042003 (2015), arXiv:1409.7215 [gr-qc]

    Pith/arXiv arXiv 2015

  47. [47]

    Neu- ral Importance Sampling for Rapid and Reliable Gravitational-Wave Inference,

    Maximilian Dax, Stephen R. Green, Jonathan Gair, Michael P¨ urrer, Jonas Wildberger, Jakob H. Macke, Alessandra Buonanno, and Bernhard Sch¨ olkopf, “Neu- ral Importance Sampling for Rapid and Reliable Gravitational-Wave Inference,” Phys. Rev. Lett.130, 171403 (2023), arXiv:2210.05686 [gr-qc]

    Pith/arXiv arXiv 2023

  48. [48]

    Fast marginalization algorithm for optimiz- ing gravitational wave detection, parameter estimation, and sky localization,

    Javier Roulet, Jonathan Mushkin, Digvijay Wadekar, Tejaswi Venumadhav, Barak Zackay, and Matias Zal- darriaga, “Fast marginalization algorithm for optimiz- ing gravitational wave detection, parameter estimation, and sky localization,” Phys. Rev. D110, 044010 (2024), arXiv:2404.02435 [gr-qc]

    Pith/arXiv arXiv 2024

  49. [49]

    dynesty: a dynamic nested sam- pling package for estimating Bayesian posteriors and ev- idences,

    Joshua S. Speagle, “dynesty: a dynamic nested sam- pling package for estimating Bayesian posteriors and ev- idences,” Mon. Not. Roy. Astron. Soc.493, 3132–3158 (2020), arXiv:1904.02180 [astro-ph.IM]

    Pith/arXiv arXiv 2020

  50. [50]

    GWPopulation: Hard- ware agnostic population inference for compact binaries and beyond,

    Colm Talbot, Amanda Farah, Shanika Galaudage, Ja- cob Golomb, and Hui Tong, “GWPopulation: Hard- ware agnostic population inference for compact binaries and beyond,” J. Open Source Softw.10, 7753 (2025), arXiv:2409.14143 [astro-ph.IM]

    Pith/arXiv arXiv 2025

  51. [51]

    Parallelized Inference for Gravitational- Wave Astronomy,

    Colm Talbot, Rory Smith, Eric Thrane, and Gre- gory B. Poole, “Parallelized Inference for Gravitational- Wave Astronomy,” Phys. Rev. D100, 043030 (2019), arXiv:1904.02863 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  52. [52]

    Padded gwpopulation,

    Christian Adamcewicz, “Padded gwpopulation,” (2026)

    2026

  53. [53]

    Padded gwpopulation pipe,

    Christian Adamcewicz, “Padded gwpopulation pipe,” (2026)

    2026

  54. [54]

    gwax: Gravitational-wave astronomy in JAX,

    Matthew Mould, “gwax: Gravitational-wave astronomy in JAX,” (2026)

    2026

  55. [55]

    Nested sampling for general Bayesian computation,

    John Skilling, “Nested sampling for general Bayesian computation,” Bayesian Analysis1, 833–859 (2006)

    2006

  56. [56]

    Normalizing Flows for Probabilistic Model- ing and Inference,

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshmi- narayanan, “Normalizing Flows for Probabilistic Model- ing and Inference,” J. Machine Learning Res.22, 2617– 2680 (2021), arXiv:1912.02762 [stat.ML]

    Pith/arXiv arXiv 2021

  57. [57]

    Complete param- eter inference for GW150914 using deep learning,

    Stephen R. Green and Jonathan Gair, “Complete param- eter inference for GW150914 using deep learning,” Mach. Learn. Sci. Tech.2, 03LT01 (2021), arXiv:2008.03312 [astro-ph.IM]

    Pith/arXiv arXiv 2021

  58. [58]

    Real-Time Gravitational Wave Science with Neural Posterior Estimation,

    Maximilian Dax, Stephen R. Green, Jonathan Gair, Jakob H. Macke, Alessandra Buonanno, and Bernhard Sch¨ olkopf, “Real-Time Gravitational Wave Science with Neural Posterior Estimation,” Phys. Rev. Lett.127, 241103 (2021), arXiv:2106.12594 [gr-qc]

    Pith/arXiv arXiv 2021

  59. [59]

    Accelerated Sequential Posterior Inference via Reuse for Gravitational-Wave Analyses,

    Michael J. Williams, “Accelerated Sequential Posterior Inference via Reuse for Gravitational-Wave Analyses,” (2025), arXiv:2511.04218 [hep-ex]

    Pith/arXiv arXiv 2025

  60. [60]

    Leveraging rapid parameter estimates for effi- cient gravitational-wave Bayesian inference via posterior repartitioning,

    Metha Prathaban, Charlie Hoy, and Michael J. Williams, “Leveraging rapid parameter estimates for effi- cient gravitational-wave Bayesian inference via posterior repartitioning,” (2026), arXiv:2601.21630 [gr-qc]

    arXiv 2026

  61. [61]

    Simple parameter estimation using observable features of gravitational- wave signals,

    Stephen Fairhurst, Charlie Hoy, Rhys Green, Cameron Mills, and Samantha A. Usman, “Simple parameter estimation using observable features of gravitational- wave signals,” Phys. Rev. D108, 082006 (2023), arXiv:2304.03731 [gr-qc]

    Pith/arXiv arXiv 2023

  62. [62]

    Hierarchical Inference 20 of Binary Neutron Star Mass Distribution and Equation of State with Gravitational Waves,

    Jacob Golomb and Colm Talbot, “Hierarchical Inference 20 of Binary Neutron Star Mass Distribution and Equation of State with Gravitational Waves,” Astrophys. J.926, 79 (2022), arXiv:2106.15745 [astro-ph.HE]

    Pith/arXiv arXiv 2022

  63. [63]

    Density estimation with Gaussian processes for gravi- tational wave posteriors,

    Virginia D’Emilio, Rhys Green, and Vivien Raymond, “Density estimation with Gaussian processes for gravi- tational wave posteriors,” Mon. Not. Roy. Astron. Soc. 508, 2090–2097 (2021), arXiv:2104.05357 [gr-qc]

    Pith/arXiv arXiv 2090

  64. [64]

    Inferring the neutron star equation of state simultaneously with the population of merging neutron stars,

    Daniel Wysocki, Richard O’Shaughnessy, Leslie Wade, and Jacob Lange, “Inferring the neutron star equation of state simultaneously with the population of merging neutron stars,” (2020), arXiv:2001.01747 [gr-qc]

    Pith/arXiv arXiv 2020

  65. [65]

    Normalizing flows for gravitational-wave posteriors,

    Christian Adamcewicz, “Normalizing flows for gravitational-wave posteriors,” (2026)

    2026

  66. [66]

    Flowjax: Distributions and normalizing flows in jax,

    Daniel Ward, “Flowjax: Distributions and normalizing flows in jax,” (2022)

    2022

  67. [67]

    Neural Spline Flows,

    Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios, “Neural Spline Flows,” (2019) arXiv:1906.04032 [stat.ML]

    Pith/arXiv arXiv 2019

  68. [68]

    Normalizing Flows on Tori and Spheres,

    Danilo Jimenez Rezende, George Papamakarios, S´ ebastien Racani` ere, Michael S. Albergo, Gurtej Kanwar, Phiala E. Shanahan, and Kyle Cranmer, “Normalizing Flows on Tori and Spheres,” (2020), arXiv:2002.02428 [stat.ML]

    Pith/arXiv arXiv 2020

  69. [69]

    Ac- celerated nested sampling with posterior repartitioning andβ-flows for gravitational waves,

    Metha Prathaban, Harry Bevins, and Will Handley, “Ac- celerated nested sampling with posterior repartitioning andβ-flows for gravitational waves,” Mon. Not. Roy. As- tron. Soc.541, 200–213 (2025), arXiv:2411.17663 [astro- ph.IM]

    arXiv 2025

  70. [70]

    The Statistical Accuracy of Neural Posterior and Likelihood Estimation,

    David T. Frazier, Ryan Kelly, Christopher Drovandi, and David J. Warne, “The Statistical Accuracy of Neural Posterior and Likelihood Estimation,” (2024), arXiv:2411.12068 [stat.ML]

    Pith/arXiv arXiv 2024

  71. [71]

    Following a mov- ing target—monte carlo inference for dynamic bayesian models,

    Walter R. Gilks and Carlo Berzuini, “Following a mov- ing target—monte carlo inference for dynamic bayesian models,” Journal of the Royal Statistical Society Series B: Statistical Methodology63, 127–146 (2001)

    2001

  72. [72]

    A tutorial on par- ticle filtering and smoothing: Fifteen years later,

    Arnaud Doucet and Adam Johansen, “A tutorial on par- ticle filtering and smoothing: Fifteen years later,” Hand- book of Nonlinear Filtering12, 656–704 (2009)

    2009

  73. [73]

    Weighted average importance sampling and defensive mixture distributions,

    Tim Hesterberg, “Weighted average importance sampling and defensive mixture distributions,” Technometrics37, 185–194 (1995)

    1995

  74. [74]

    Safe and effective im- portance sampling,

    Art Owen and Yi Zhou Associate, “Safe and effective im- portance sampling,” Journal of the American Statistical Association95, 135–143 (2000)

    2000

  75. [75]

    Rapid inference and comparison of gravitational- wave population models with neural variational posteri- ors,

    Matthew Mould, Noah E. Wolfe, and Salvatore Vi- tale, “Rapid inference and comparison of gravitational- wave population models with neural variational posteri- ors,” Phys. Rev. D111, 123049 (2025), arXiv:2504.07197 [astro-ph.IM]

    Pith/arXiv arXiv 2025

  76. [76]

    Variational Inference: A Review for Statisticians,

    David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe, “Variational Inference: A Review for Statisticians,” Jour- nal of the American Statistical Association112, 859–877 (2017), arXiv:1601.00670 [stat.CO]

    Pith/arXiv arXiv 2017

  77. [77]

    Neural Bayesian updates to popula- tions with growing gravitational-wave catalogs,

    Noah E. Wolfe, Matthew Mould, John Veitch, and Salvatore Vitale, “Neural Bayesian updates to popula- tions with growing gravitational-wave catalogs,” (2026), arXiv:2602.20277 [astro-ph.IM]

    arXiv 2026

  78. [78]

    End-to-End Population Inference from Gravitational-Wave Strain using Transformers,

    Konstantin Leyde, Stephen R. Green, Maximilian Dax, Matthew Mould, Cecilia Maria Fabbri, and Jonathan Gair, “End-to-End Population Inference from Gravitational-Wave Strain using Transformers,” (2026), arXiv:2605.11274 [gr-qc]

    Pith/arXiv arXiv 2026

  79. [79]

    Gravitational wave populations and cosmology with neural posterior estimation,

    Konstantin Leyde, Stephen R. Green, Alexandre Toubiana, and Jonathan Gair, “Gravitational wave populations and cosmology with neural posterior estimation,” Phys. Rev. D109, 064056 (2024), arXiv:2311.12093 [gr-qc]

    Pith/arXiv arXiv 2024

  80. [80]

    Measuring the binary black hole mass spectrum with an astrophysically moti- vated parameterization,

    Colm Talbot and Eric Thrane, “Measuring the binary black hole mass spectrum with an astrophysically moti- vated parameterization,” Astrophys. J.856, 173 (2018), arXiv:1801.02699 [astro-ph.HE]

    Pith/arXiv arXiv 2018

Showing first 80 references.