Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

SHARPy combines Sequential Monte Carlo with the No-U-Turn Sampler to deliver posterior samples and evidence estimates consistent with Nested Sampling in about ten minutes on a single GPU.

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

2026-08-03 12:30 UTC pith:POURTIOF

load-bearing objection SHARPy is a genuinely new NUTS+SMC combination for GW inference with a plausible speedup, but the position-dependent mass matrix in Eq. (11) breaks detailed balance and likely explains the systematic evidence bias; worth refereeing, not yet trustworthy as presented. the 4 major comments →

arxiv 2601.02336 v2 pith:POURTIOF submitted 2026-01-05 gr-qc astro-ph.COastro-ph.IM

The Sequential Monte Carlo goes NUTS: Boosting Gravitational-Wave Inference

Gabriele Demasi , Giulia Capurri , Massimo Lenti , Angelo Ricciardone , Barbara Patricelli , Adriano Frattale Mascioli , Lorenzo Piccari , Saulo Albuquerque
show 4 more authors
Gianluca M. Guidi Francesco Pannarale Giulia Stratta Walter Del Pozzo
This is my paper
classification gr-qc astro-ph.COastro-ph.IM PACS 04.80.Nn02.70.Uu
keywords gravitational-wave inferenceSequential Monte CarloNo-U-Turn SamplerBayesian evidenceGPU accelerationparameter estimationbinary black holesautomatic differentiation
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.

This paper attempts to establish that gravitational-wave parameter estimation and model comparison can be done with a likelihood-based sampler in minutes rather than hours. The authors combine Sequential Monte Carlo's parallel particle evolution with the No-U-Turn Sampler's gradient-guided exploration, and implement the combination on hardware accelerators with automatic differentiation. For the benchmark event GW150914, they report posterior marginals and evidence estimates consistent with a Nested Sampling reference, with a single-event run taking about ten minutes on one GPU. They also present a probability-probability test on 100 simulated binary black-hole signals to support the claim that the sampler is statistically unbiased. If these results hold, the main payoff is that evidence-driven model comparison, historically a costly sequential step, can become routine and fast enough for low-latency analysis.

Core claim

On its own terms, the paper demonstrates that a population-based sampler with a gradient-based mutation kernel can replace Nested Sampling for single-event gravitational-wave inference. The authors introduce SHARPy, which evolves 9000 particles through a temperature ladder from prior to posterior, mutating each particle with the No-U-Turn Sampler and using a Hessian-based mass matrix to encode local geometry. Applied to GW150914 with an 11-dimensional aligned-spin waveform, the paper reports posterior samples and evidence estimates consistent with a Nested Sampling reference, in about ten minutes on a single GPU. A probability-probability test on 100 simulated injections is presented as evid

What carries the argument

The load-bearing mechanism is the pairing of Sequential Monte Carlo with the No-U-Turn Sampler as the mutation kernel. Sequential Monte Carlo provides a population of particles evolved from prior to posterior through an adaptive temperature ladder, yielding an unbiased evidence estimate as a product of normalizing-constant ratios; the No-U-Turn Sampler replaces the usual random-walk MCMC kernel with Hamiltonian trajectories whose length is auto-tuned, suppressing random-walk behavior. Three refinements carry the efficiency claim: a Hessian-based mass matrix computed at the start of each SMC iteration to exploit local geometry while keeping the Hamiltonian dynamics separable; reflective and p

Load-bearing premise

The claim rests on the assumption that the fixed sampler configuration — 9000 particles, one No-U-Turn-Sampler move per particle per iteration, step size 0.3, and a Hessian mass matrix recomputed only at the start of each SMC iteration — mixes the tempered sequence well enough that the evidence product is unbiased and the recycled particles represent the posterior.

What would settle it

Run the same event with a larger number of No-U-Turn-Sampler moves per particle per SMC iteration or a stricter effective-sample-size target and check whether the evidence distribution shifts upward toward the Nested Sampling reference; if it does, the reported downward evidence offset is a mixing artifact. A complementary check is to tabulate effective sample sizes and acceptance rates across the 100 runs — low values for the outer parameters (declination, distance, inclination) would indicate the recycled-sample posterior approximation is suspect.

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

If this is right

  • Single-event parameter estimation on real data takes about ten minutes on one GPU, compared with hours-to-days for typical Nested Sampling runs.
  • The posterior samples and evidence estimates on GW150914 are consistent with Nested Sampling at the 90% level, with the paper noting a slight downward tendency in evidence.
  • A probability-probability test on 100 simulated binary black-hole injections shows the credible intervals behave as expected for an unbiased sampler.
  • Because SMC particles evolve in parallel and the No-U-Turn Sampler scales well with dimension, the paper argues the method is suited to higher-dimensional problems and to incremental inference when new data arrive.
  • The sampler recycles all tempered particles via rejection sampling, turning intermediate SMC distribution samples into posterior samples rather than discarding them.

Where Pith is reading between the lines

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

  • If the ten-minute runtime transfers to full 15-parameter precessing waveforms, alerts from a detection could carry posterior-informed sky localization and distance within the same observing run.
  • The same SMC-plus-gradient-kernel recipe applies to hierarchical population inference, where a single gradient-capable SMC run could replace thousands of per-event runs.
  • The systematic downward evidence offset is the main quantity to watch: if it reflects mixing bias, then model-comparison applications built on the evidence product would need a bias correction or a longer mutation budget before ratio-based decisions are trusted.
  • Because the mass matrix is recomputed only once per iteration, updating it a few times within an iteration on GPU could improve mixing for sharply curved posteriors at modest cost — a testable extension.

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

4 major / 4 minor

Summary. The paper presents SHARPy, a Sequential Monte Carlo (SMC) sampler for gravitational-wave parameter estimation that uses the No-U-Turn Sampler (NUTS) as its mutation kernel, with a per-particle mass matrix set to the Hessian of the target distribution. The implementation is in JAX and runs on GPUs. The authors validate the method on 100 simulated BBH injections via a probability-probability test, compare GW150914 posteriors and evidence with the Dynesty nested sampler, and report wall-clock times of roughly 10–15 minutes. The central claim is that SHARPy produces posterior samples and evidence estimates consistent with Nested Sampling while requiring only a small fraction of the runtime.

Significance. The intended contribution is timely and potentially valuable: an SMC sampler with a gradient-based mutation kernel and GPU parallelism could enable minute-scale likelihood-based inference for gravitational-wave events. The use of JAX, automatic differentiation, and publicly available code are strengths, and the PP test and Dynesty comparison are appropriate validation benchmarks. However, the validity of the sampler hinges on the mutation kernel being invariant with respect to the tempered target. As described, the state-dependent mass matrix breaks detailed balance without a correction, and the reported low-evidence bias is a plausible signature of this problem. If the kernel is corrected and the evidence discrepancy is resolved, the speed advantage would make the paper significant. In its current form, the central consistency claim is not established.

major comments (4)
  1. [Sec. V, Eq. (11), and Sec. III.3] The mutation kernel is not reversible as described. The mass matrix M is set per particle to the Hessian of p(θ|d) at that particle's current position, and one NUTS trajectory is then run with that fixed matrix. Because the proposal distribution depends on the starting point through M(θ), the forward trajectory uses M(θ) while the reverse trajectory would use M(θ′). Detailed balance with respect to the tempered target p_t therefore requires a Jacobian/preconditioning correction, as in Riemann-manifold HMC; no such correction is described. Without invariance of the mutation kernel, the SMC weight recursion in Eqs. (4)–(7) and the recycled-sample estimator in Eq. (16) are biased. The systematic downward shift of SHARPy's evidence in Fig. 5 is consistent with this bias. The authors must either use a global mass matrix per SMC iteration, implement a properly reversible state-dependent kernel
  2. [Sec. V, Eq. (11)] The Hessian of the posterior density, H_ij = ∂²p(θ|d)/∂θ_i∂θ_j, is not positive definite in general. Near a mode it is negative definite; for multimodal or heavy-tailed targets it can be indefinite. The kinetic energy term ½ rᵀM⁻¹r requires M to be positive definite, so the algorithm as written can encounter invalid mass matrices. The paper should specify the use of a positive-definite metric, e.g., the expected Fisher information or a regularized Hessian of the negative log-posterior, and describe how indefinite cases are handled.
  3. [Sec. VI, Figs. 2 and 5] The validation is insufficient to support the consistency claim. The PP test uses 100 injections and reports no effective sample size, acceptance-rate, or convergence diagnostics; it may lack power to detect a modest bias in the mutation kernel. More importantly, Fig. 5 shows a systematic evidence offset: the Dynesty log-evidence lies in the upper tail of the 100-run SHARPy distribution. The statement that the results agree 'at the 90% level' is weak; the authors should report the mean and standard deviation of the SHARPy log-evidence and the offset relative to Dynesty. The bias needs a quantitative explanation, not just attribution to KDE artifacts.
  4. [Sec. VI, Fig. 4] For declination, luminosity distance, and inclination angle, the JS divergence is systematically above the 1.5 mb threshold proposed in Ref. [47]. The assertion that this is mainly due to KDE artifacts from sharp features and boundaries is not tested. If the SHARPy and Dynesty posteriors genuinely differ, this contradicts the abstract's claim of consistency. The authors should verify with alternative density estimators or larger sample counts, or identify the actual cause of the discrepancy.
minor comments (4)
  1. [Sec. VI] The runtime comparison is incomplete: no timing for the Dynesty reference run is reported, and hardware details are limited to 'a single NVIDIA A100 GPU'. A side-by-side wall-clock comparison using the same waveform model, likelihood, and similar sample quality would strengthen the 'around ten minutes' claim.
  2. [Sec. V, Eq. (11)] The notation p(θ|d) is ambiguous: it is not clear whether the Hessian is computed for the full posterior or for the tempered target p_t at the current SMC iteration. This should be stated explicitly.
  3. [Sec. VII] The statement 'we do not expect our findings to change significantly in full scale scenarios' is an unsupported extrapolation, since only an 11-dimensional aligned-spin model is tested. This should be softened or supported by a preliminary higher-dimensional example.
  4. [Appendix A] The bimodal 11-D test uses σ = 0.01 with a known analytic evidence, but this is a very different regime from the GW posterior and is unlikely to expose the position-dependent-kernel bias. It does not substitute for evidence validation on the actual problem.

Circularity Check

0 steps flagged

No significant circularity: SHARPy's claims rest on external benchmarks and standard SMC identities.

full rationale

The paper's derivation chain is self-contained and empirically benchmarked. The evidence estimator in Eqs. (7)-(8) is the standard SMC importance-sampling identity Z_t/Z_{t-1}=E_{p_{t-1}}[L^{beta_t-beta_{t-1}}], not a fitted quantity; no parameter is fit to Dynesty outputs and then reported as a prediction. The posterior and evidence claims are validated against an independent sampler (Dynesty/Bilby) and against 100 injected BBH signals with known truth (PP test, Fig. 2), so the central comparison is external. The recycled-sample scheme in Eq. (16) does use SHARPy's own per-iteration evidence estimates to normalize the mixture components, which is an internal self-reference, but it does not reduce a predicted quantity to an input: the final posterior agreement is judged against Dynesty and injected parameters, and the evidence output is directly compared with Dynesty. References to prior SMC-GW work (e.g., [13]) are external and used to contextualize a known downward evidence tendency, not to justify the central claim. The manuscript itself flags limitations such as the KDE-based JS-divergence interpretation and the slight evidence bias (Figs. 4-5), and these are calibration concerns, not circularity. Potential issues raised by the method itself, such as the local-Hessian mass matrix in Eq. (11) lacking a positive-definiteness guarantee and the absence of ESS/acceptance diagnostics, are correctness/validation concerns, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

SHARPy adds no new physical entities; its free parameters are algorithmic hyperparameters. The central result rests on standard SMC/NUTS theory plus a nonstandard Hessian mass-matrix choice (Eq. 11) and an assumption that the fixed hand-set configuration mixes well for GW posteriors.

free parameters (4)
  • N_P (number of SMC particles) = 9000
    Chosen by hand; larger NP increases accuracy but costs compute; set once for all runs.
  • alpha (ESS target fraction) = 0.95
    Controls the adaptive temperature schedule in Eq. (6); set by hand.
  • NUTS step size epsilon = 0.3
    Fixed leapfrog step size for all runs; gradient-based sampler performance is sensitive to this value.
  • NUTS moves per particle per iteration = 1
    One NUTS trajectory per particle per SMC iteration; a tuning choice affecting mixing.
axioms (6)
  • domain assumption Detector noise is stationary and Gaussian with known PSD, so the log-likelihood is Eq. (2).
    Standard in LIGO/Virgo parameter estimation; stated in Sec. II. If violated, the likelihood is wrong.
  • standard math The SMC normalizing-constant recursion in Eqs. (7)-(8) gives an unbiased evidence estimate.
    Standard SMC sampler theory (Del Moral et al. [11]); the paper relies on it for evidence claims.
  • standard math NUTS transition kernel leaves each tempered target invariant and, combined with reweighting/resampling, yields consistent SMC estimates.
    Follows from NUTS/HMC theory [17]; used in the mutation step.
  • ad hoc to paper The Hessian H_ij in Eq. (11) is an appropriate mass matrix for the local geometry.
    As written it is the Hessian of the posterior density, not of the negative log posterior; no derivation shows this is a valid metric. If the intended object is the log-density Hessian, the text should say so.
  • standard math Recycling particles from all tempered iterations and rejection-sampling against p_tilde produces i.i.d. samples from the target.
    Rejection sampling argument assuming p_tilde has bounded ratio to the target; the paper does not quantify acceptance rate.
  • domain assumption IMRPhenomD waveform model is sufficient for aligned-spin BBH signals.
    Chosen waveform model [44,45] restricts the analysis to an 11-dimensional parameter space and ignores precession.

pith-pipeline@v1.3.0-alltime-deepseek · 12429 in / 13604 out tokens · 126285 ms · 2026-08-03T12:30:30.564613+00:00 · methodology

0 comments
read the original abstract

Sequential Monte Carlo (SMC) methods have recently been applied to gravitational-wave inference as a powerful alternative to standard sampling techniques, such as Nested Sampling. At the same time, gradient-based Markov Chain Monte Carlo algorithms, most notably the No-U-Turn Sampler (NUTS), provide an efficient way to explore high-dimensional parameter spaces. In this work we present SHARPy, a Bayesian inference framework that combines the parallelism and evidence-estimation capabilities of SMC with the state-of-the-art sampling performance of NUTS. Moreover, SHARPy exploits the local geometric structure of the posterior to further improve efficiency. Built on JAX, a high-performance computing framework that enables automatic differentiation and hardware acceleration, SHARPy performs gravitational-wave inference on binary black-hole events in around ten minutes, yielding posterior samples and Bayesian evidence estimates that are consistent with those obtained through Nested Sampling. This work sets a new milestone in Gravitational-Wave inference with likelihood-based methods and paves the way for model comparison tasks to be accomplished in minutes.

Figures

Figures reproduced from arXiv: 2601.02336 by Adriano Frattale Mascioli, Angelo Ricciardone, Barbara Patricelli, Francesco Pannarale, Gabriele Demasi, Gianluca M. Guidi, Giulia Capurri, Giulia Stratta, Lorenzo Piccari, Massimo Lenti, Saulo Albuquerque, Walter Del Pozzo.

Figure 1
Figure 1. Figure 1: Illustration of the SMC algorithm applied on a bimodal Gaussian mixture distribution. Particles are first randomly drawn from the prior. Then they are reweighed according to tempered distribution with a certain β and resampled according to these weights so that particles the lie in high likelihood regions are selected. At the end of each SMC iteration, in the mutation step, particles explore the space with… view at source ↗
Figure 2
Figure 2. Figure 2: Probability-probability test for the simulated BBH systems. For each parameter of the binary, the plot reports on the y-axis the fraction of events for which the true value lies within the credible interval (C.I.) on the x-axis. The resulting p-values for each parameters are reported in the corresponding legend entry. The shaded bands represents the 1-2-3 σ quantiles. IV No-U-Turn-Sampler As highlighted in… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between the posterior samples of GW150914 obtained with SHARPy (in red) and the posterior samples obtained with Dynesty, in blue. The corner plot on the left is limited to four intrinsic parameters (the chirp mass, the mass ratio and the two spin magnitudes), while the on the right shows four extrinsic parameters, namely the right ascension, the declination, the luminosity distance and the incli… view at source ↗
Figure 4
Figure 4. Figure 4: Jensen-Shannon divergence, expressed in mbits, between the samples obtained with SHARPy and those ob￾tained with Dynesty in the GW150914 case. The triangles and the errobars indicate respectively the median and the 90% credible intervals obtained from 100 indipendent runs with SHARPy and Dynesty. tion. JAX implementation SHARPy is entirely developed in JAX. The implemen￾tation of the algorithm is publicly … view at source ↗
Figure 6
Figure 6. Figure 6: Marginal samples from the first two dimensions of the bimodal 11-D distribution introduced in Eq. (17). to study the distribution of the evidence, comparing it against the true (and analytic) value. We report the results in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Histogram of the evidences estimated by 100 independent SHARPy runs compared to the true analytical value. References 1. LIGO Scientific Collaboration et al. Advanced LIGO. Classical and Quantum Gravity 32, 074001. arXiv: 1411.4547 [gr-qc] (Apr. 2015). 2. Acernese, F. et al. Advanced Virgo: a second￾generation interferometric gravitational wave de￾tector. Classical and Quantum Gravity 32, 024001. arXiv: 14… view at source ↗
Figure 8
Figure 8. Figure 8: Corner plot of the samples obtain in one of the injections performed in section VI. The line indicates the injection parameters. 22. De Santi, F. et al. Deep learning to detect gravi￾tational waves from binary close encounters: Fast parameter estimation using normalizing flows. Phys. Rev. D 109, 102004. arXiv: 2404.12028 [gr-qc] (2024). 23. Gabbard, H., Messenger, C., Heng, I. S., Tono￾lini, F. & Murray-Sm… view at source ↗
Figure 9
Figure 9. Figure 9: Full corner plot of the comparison between the samples of GW150914 obtained with Dynesty and SHARPy, partially showed in fig. 3 of section VI. The value of the JD divergence (JSD) between the two set of samples is reported at the top of each marginal 1D plot in the diagonal. . 28. Smith, R. et al. Fast and accurate inference on gravitational waves from precessing compact bi￾naries. Phys. Rev. D 94, 044031.… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Ab Initio Real-Time Gravitational-Wave Parameter Estimation

    gr-qc 2026-07 accept novelty 6.0

    Slice-within-Gibbs nested sampling on modern GPUs delivers well-calibrated BNS parameter estimation in ~12 minutes uncompressed and ~89 seconds with heterodyning, from cold priors.

  2. Reliable model selection in the presence of parameter non-identifiability

    stat.ME 2026-05 unverdicted novelty 6.0

    Proposes adaptive multiple importance sampling for robust Bayesian model evidence estimation under parameter non-identifiability, shown to outperform deterministic methods on ecological case studies while being cheape...

Reference graph

Works this paper leans on

54 extracted references · 24 linked inside Pith · cited by 2 Pith papers

  1. [1]

    arXiv:1411.4547 [gr-qc](Apr

    LIGO Scientific Collaborationet al.Advanced LIGO.Classical and Quantum Gravity32, 074001. arXiv:1411.4547 [gr-qc](Apr. 2015)

    Pith/arXiv arXiv 2015

  2. [2]

    arXiv:1408.3978 [gr-qc](Jan

    Acernese, F.et al.Advanced Virgo: a second- generation interferometric gravitational wave de- tector.Classical and Quantum Gravity32, 024001. arXiv:1408.3978 [gr-qc](Jan. 2015)

    Pith/arXiv arXiv 2015

  3. [3]

    Aso, Y.et al.Interferometer design of the KA- GRA gravitational wave detector.Phys. Rev. D 88,043007. arXiv:1306 . 6747 [gr-qc](Aug. 2013)

    2013

  4. [4]

    & Talbot, C

    Thrane, E. & Talbot, C. An introduction to Bayesian inference in gravitational-wave astron- omy: Parameter estimation, model selection, and hierarchical models.”Publications of the Astro- nomical Society of Australia”36,e010. arXiv: 1809.02293 [astro-ph.IM](Mar. 2019)

    Pith/arXiv arXiv 2019

  5. [5]

    arXiv:2508.18083 [astro-ph.HE](Aug

    The LIGO Scientific Collaborationet al.GWTC- 4.0: Population Properties of Merging Com- pact Binaries.arXiv e-prints,arXiv:2508.18083. arXiv:2508.18083 [astro-ph.HE](Aug. 2025)

    Pith/arXiv arXiv 2025

  6. [6]

    arXiv e-prints,arXiv:2509.04348

    The LIGO Scientific Collaborationet al.GWTC- 4.0: Constraints on the Cosmic Expansion Rate and Modified Gravitational-wave Propagation. arXiv e-prints,arXiv:2509.04348. arXiv:2509 . 04348 [astro-ph.CO](Sept. 2025)

    Pith/arXiv arXiv 2025

  7. [7]

    arXiv:2112

    The LIGO Scientific Collaborationet al.Tests of General Relativity with GWTC-3.arXiv e- prints,arXiv:2112.06861. arXiv:2112 . 06861 [gr-qc](Dec. 2021)

    Pith/arXiv arXiv 2021

  8. [8]

    Skilling, J.: Nested sampling for gen- eral Bayesian computation

    Skilling, J. Skilling, J.: Nested sampling for gen- eral Bayesian computation. Bayesian Anal. 1(4), 833-860.Bayesian Analysis1,833–860 (Dec. 2006)

    2006

  9. [9]

    Ashton, G.et al.Nested sampling for physical sci- entists.Nature2.arXiv:2205.15570 [stat.CO] (2022)

    Pith/arXiv arXiv 2022

  10. [10]

    arXiv:2508

    The LIGO Scientific Collaborationet al.GWTC- 4.0: Methods for Identifying and Characterizing Gravitational-wave Transients.arXiv e-prints, arXiv:2508.18081. arXiv:2508 . 18081 [gr-qc] (Aug. 2025)

    Pith/arXiv arXiv 2025

  11. [11]

    & Jasra, A

    Del Moral, P., Doucet, A. & Jasra, A. Sequen- tial Monte Carlo Samplers.J. R. Stat. Soc. Ser. B Stat. Methodol68,411–436.issn: 1369-7412. eprint:https : / / academic . oup . com / jrsssb / article-pdf/68/3/411/49795343/jrsssb_68_ 3 _ 411 . pdf.https : / / doi . org / 10 . 1111 / j . 1467-9868.2006.00553.x(May 2006)

    arXiv 2006

  12. [12]

    A., Nabergoj, D

    Karamanis, M., Beutler, F., Peacock, J. A., Nabergoj, D. & Seljak, U. Accelerating astro- nomical and cosmological inference with pre- conditioned Monte Carlo.Mon. Not. Roy. As- tron. Soc.516,1644–1653. arXiv:2207 . 05652 [astro-ph.IM](2022)

    2022

  13. [13]

    J., Karamanis, M., Luo, Y

    Williams, M. J., Karamanis, M., Luo, Y. & Sel- jak, U. Validating Sequential Monte Carlo for Gravitational-Wave Inference.Mon. Not. Roy. Astron. Soc.1479,1493. arXiv:2506 . 18977 [astro-ph.IM](2025)

    2025

  14. [14]

    Williams, M. J. Accelerated Sequential Posterior Inference via Reuse for Gravitational-Wave Anal- yses. arXiv:2511.04218 [hep-ex](Nov. 2025)

    Pith/arXiv arXiv 2025

  15. [15]

    F., Drovandi, C

    Salomone, R., South, L. F., Drovandi, C. C., Kroese, D. P. & Johansen, A. M. Unbiased and Consistent Nested Sampling via Sequen- tial Monte Carlo.87,1221–1238.https : / / academic . oup . com / jrsssb / article / 87 / 4 / 1221/8129577(2025)

    2025

  16. [16]

    inHandbook of Markov Chain Monte Carlo113–162 (2011)

    Neal, R. inHandbook of Markov Chain Monte Carlo113–162 (2011)

    2011

  17. [18]

    Veitch, J.et al.Parameter estimation for com- pact binaries with ground-based gravitational- wave observations using the LALInference soft- ware library.Phys. Rev. D91,042003. arXiv: 1409.7215 [gr-qc](Feb. 2015)

    Pith/arXiv arXiv 2015

  18. [20]

    & Rizzo, M

    Lange, J., O’Shaughnessy, R. & Rizzo, M. Rapid and accurate parameter inference for coalescing, precessing compact binaries. arXiv:1805.10457 [gr-qc](May 2018)

    Pith/arXiv arXiv 2018

  19. [21]

    Dax, M.et al.Real-Time Gravitational Wave Science with Neural Posterior Estimation.Phys. Rev. Lett.127,241103. arXiv:2106 . 12594 [gr-qc](2021). 9 REFERENCES REFERENCES Figure 8:Corner plot of the samples obtain in one of the injections performed in section VI. The line indicates the injection parameters

    2021

  20. [22]

    De Santi, F.et al.Deep learning to detect gravi- tational waves from binary close encounters: Fast parameter estimation using normalizing flows. Phys. Rev. D109,102004. arXiv:2404 . 12028 [gr-qc](2024)

    2024

  21. [23]

    S., Tono- lini, F

    Gabbard, H., Messenger, C., Heng, I. S., Tono- lini, F. & Murray-Smith, R. Bayesian parame- ter estimation using conditional variational au- toencoders for gravitational-wave astronomy.Na- ture Phys.18,112–117. arXiv:1909 . 06296 [astro-ph.IM](2022)

    1909

  22. [24]

    Chua, A. J. K. & Vallisneri, M. Learning Bayesian posteriors with neural networks for gravitational- wave inference.Phys. Rev. Lett.124,041102. arXiv:1909.05966 [gr-qc](2020)

    Pith/arXiv arXiv 1909

  23. [25]

    Wong, K. W. K., Isi, M. & Edwards, T. D. P. Fast Gravitational-wave Parameter Estimation with- out Compromises.Astrophys. J.958,129. arXiv: 2302.05333 [astro-ph.IM](2023)

    Pith/arXiv arXiv 2023

  24. [26]

    Wouters, T., Pang, P. T. H., Dietrich, T. & Van Den Broeck, C. Robust parameter estima- tion within minutes on gravitational wave signals from binary neutron star inspirals.Phys. Rev. D 110,083033. arXiv:2404.11397 [astro-ph.IM] (2024)

    arXiv 2024

  25. [27]

    & Raymond, V

    Morisaki, S. & Raymond, V. Rapid Parameter Estimation of Gravitational Waves from Binary Neutron Star Coalescence using Focused Reduced Order Quadrature.Phys. Rev. D102,104020. arXiv:2007.09108 [gr-qc](2020). 10 REFERENCES REFERENCES Figure 9:Full corner plot of the comparison between the samples of GW150914 obtained with Dynesty andSHARPy, partially show...

    Pith/arXiv arXiv 2007

  26. [28]

    Smith, R.et al.Fast and accurate inference on gravitational waves from precessing compact bi- naries.Phys. Rev. D94,044031. arXiv:1604 . 08253 [gr-qc](2016)

    2016

  27. [29]

    Accelerating parameter estimation of gravitational waves from compact binary coales- cence using adaptive frequency resolutions.Phys

    Morisaki, S. Accelerating parameter estimation of gravitational waves from compact binary coales- cence using adaptive frequency resolutions.Phys. Rev. D104,044062. arXiv:2104.07813 [gr-qc] (2021)

    Pith/arXiv arXiv 2021

  28. [30]

    arXiv:2312

    Krishna, K.et al.Accelerated parameter estima- tion in Bilby with relative binning. arXiv:2312. 06009 [gr-qc](Dec. 2023)

    2023

  29. [31]

    Cornish, N. J. Fast Fisher Matrices and Lazy Likelihoods. arXiv:1007 . 4820 [gr-qc](July 2010)

    2010

  30. [32]

    J., Veitch, J

    Williams, M. J., Veitch, J. & Messenger, C. Nested sampling with normalizing flows for gravitational-wave inference.Phys. Rev. D103, 103006. arXiv:2102.11056 [gr-qc](2021)

    Pith/arXiv arXiv 2021

  31. [33]

    & Porter, E

    Perret, J., Ar´ ene, M. & Porter, E. K. DeepHMC : a deep-neural-network acclerated Hamilto- nian Monte Carlo algorithm for binary neutron star parameter estimation. arXiv:2505 . 02589 [gr-qc](May 2025)

    2025

  32. [34]

    Roulet, J.et al.Removing degeneracy and mul- timodality in gravitational wave source param- eters.Phys. Rev. D106,123015. arXiv:2207. 03508 [gr-qc](2022)

    2022

  33. [35]

    Hinne, M. An introduction to Sequential Monte Carlo for Bayesian inference and model compar- 11 REFERENCES REFERENCES ison—with examples for psychology and behav- ioral science.Behavior Research Methods57,125 (2025)

    2025

  34. [36]

    & Gordon, N

    Doucet, A., De Freitas, N. & Gordon, N. inSe- quential Monte Carlo methods in practice3–14 (Springer, 2001)

    2001

  35. [37]

    Dai, C., Heng, J., Jacob, P. E. & Whiteley, N. An invitation to sequential Monte Carlo samplers. Journal of the American Statistical Association 117,1587–1600 (2022)

    2022

  36. [38]

    Hoffman, M. D. & Gelman, A. The No-U- Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo.arXiv e-prints, arXiv:1111.4246. arXiv:1111 . 4246 [stat.CO] (Nov. 2011)

    Pith/arXiv arXiv 2011

  37. [39]

    & Calderhead, B

    Girolami, M. & Calderhead, B. Riemann mani- fold langevin and hamiltonian monte carlo meth- ods.J. R. Stat. Soc. Ser. B Stat. Methodol73, 123–214 (2011)

    2011

  38. [40]

    arXiv:2402

    Cabezas, A.et al.BlackJAX: Composable Bayesian inference in JAX.arXiv e-prints, arXiv:2402.10797. arXiv:2402 . 10797 [cs.MS] (Feb. 2024)

    Pith/arXiv arXiv 2024

  39. [41]

    Edwards, T. D. P.et al.Differentiable and hardware-accelerated waveforms for gravitational wave data analysis.Phys. Rev. D110,064028. arXiv:2302.05329 [astro-ph.IM](2024)

    Pith/arXiv arXiv 2024

  40. [42]

    Speagle, J. S. DYNESTY: a dynamic nested sampling package for estimating Bayesian pos- teriors and evidences.”Mon. Not. Roy. As- tron. Soc.”493,3132–3158. arXiv:1904.02180 [astro-ph.IM](Apr. 2020)

    Pith/arXiv arXiv 1904

  41. [43]

    Ashton, G.et al.BILBY: A user-friendly Bayesian inference library for gravitational-wave astronomy.Astrophys. J. Suppl.241,27. arXiv: 1811.02042 [astro-ph.IM](2019)

    Pith/arXiv arXiv 2019

  42. [44]

    Husa, S.et al.Frequency-domain gravitational waves from nonprecessing black-hole binaries. I. New numerical waveforms and anatomy of the signal.Physical Review D93.issn: 2470-0029. http://dx.doi.org/10.1103/PhysRevD.93. 044006(Feb. 2016)

    doi:10.1103/physrevd.93 2016

  43. [45]

    Khan, S.et al.Frequency-domain gravitational waves from nonprecessing black-hole binaries. II. A phenomenological model for the advanced de- tector era.Physical Review D93.issn: 2470- 0029.http://dx.doi.org/10.1103/PhysRevD. 93.044007(Feb. 2016)

    doi:10.1103/physrevd 2016

  44. [46]

    P.et al.Observation of Gravita- tional Waves from a Binary Black Hole Merger

    Abbott, B. P.et al.Observation of Gravita- tional Waves from a Binary Black Hole Merger. Phys. Rev. Lett.116,061102. arXiv:1602.03837 [gr-qc](2016)

    Pith/arXiv arXiv 2016

  45. [47]

    M.et al.Bayesian inference for compact binary coalescences with bilby: vali- dation and application to the first LIGO–Virgo gravitational-wave transient catalogue.Mon

    Romero-Shaw, I. M.et al.Bayesian inference for compact binary coalescences with bilby: vali- dation and application to the first LIGO–Virgo gravitational-wave transient catalogue.Mon. Not. Roy. Astron. Soc.499,3295–3319. arXiv: 2006.00714 [astro-ph.IM](2020)

    Pith/arXiv arXiv 2006

  46. [48]

    & Gerosa, D

    Mancarella, M. & Gerosa, D. Sampling the full hierarchical population posterior distribution in gravitational-wave astronomy.Phys. Rev. D111, 103012. arXiv:2502.12156 [gr-qc](2025)

    arXiv 2025

  47. [49]

    & Seljak, U

    Karamanis, M. & Seljak, U. Persistent Sampling: Enhancing the Efficiency of Sequential Monte Carlo.arXiv e-prints,arXiv:2407.20722. arXiv: 2407.20722 [stat.ML](2024)

    Pith/arXiv arXiv 2024

  48. [50]

    2018.http : //github.com/jax-ml/jax

    Bradbury, J.JAXversion 0.3.13. 2018.http : //github.com/jax-ml/jax

    2018

  49. [51]

    Talbot, C.bilby-dev/bilby: v2.3.0version v2.3.0. Nov. 2024.https://doi.org/10.5281/zenodo. 14025488

    doi:10.5281/zenodo 2024

  50. [52]

    corner.py: Scatterplot ma- trices in Python.Journal of Open Source Soft- ware1,24.https://doi.org/10.21105/joss

    Foreman-Mackey, D. corner.py: Scatterplot ma- trices in Python.Journal of Open Source Soft- ware1,24.https://doi.org/10.21105/joss. 00024(2016)

    doi:10.21105/joss 2016

  51. [53]

    Hunter, J. D. Matplotlib: A 2D Graphics Envi- ronment.Computing in Science & Engineering9, 90–95 (2007)

    2007

  52. [54]

    Virtanen, P.et al.SciPy 1.0: fundamental algo- rithms for scientific computing in Python.Nature Medicine17,261–272 (Feb. 2020)

    2020

  53. [55]

    R.et al.Array programming with NumPy.nature585,357–362 (2020)

    Harris, C. R.et al.Array programming with NumPy.nature585,357–362 (2020)

    2020

  54. [56]

    & Raymond, V

    Hoy, C. & Raymond, V. PESummary: The code agnostic Parameter Estimation Summary page builder.SoftwareX15,100765.issn: 2352-7110. https : / / www . sciencedirect . com / science / article/pii/S2352711021000856(2021). This paper has been typeset from aT EX/LATEXfile prepared by the author. 12

    2021