Pith. sign in

REVIEW 3 major objections 6 minor 83 references

Embedding gravitational-wave orbital phase into a Transformer yields sharper, faster posteriors for eccentric black-hole binaries in pulsar timing data.

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 · grok-4.5

2026-07-11 23:07 UTC pith:RJG4TRPH

load-bearing objection Solid, carefully engineered SBI pipeline for eccentric BBHs in PTA data; gains are real on white-noise synthetics but the leap to realistic arrays is still aspirational. the 3 major comments →

arxiv 2607.03904 v1 pith:RJG4TRPH submitted 2026-07-04 cs.LG astro-ph.IMgr-qc

Transformers with Physics-Informed Encodings and Simulation-Based Inference for Robust Detection of Eccentric Binary Black Holes in Pulsar Timing Array Data

Subhajit Dandapat , Alvin J. K. Chua This is my paper
classification cs.LG astro-ph.IMgr-qc
keywords Physics-Informed MLSimulation-Based InferenceTransformersNormalizing FlowsGravitational-Wave Data AnalysisPulsar Timing ArraysEccentric Binary Black Holes
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.

Pulsar timing arrays can reveal nanohertz gravitational waves from supermassive black-hole binaries, but conventional Bayesian sampling is slow and expensive when the data are long, noisy, and high-dimensional. This paper shows that a Transformer whose positional encodings include the analytical orbital phase of an eccentric binary can learn physically meaningful features directly from multi-pulsar timing residuals. Those features then condition discrete or continuous normalizing flows that return full posterior distributions in an amortized, simulation-based framework. On synthetic white-noise data the physics-informed model recovers injected parameters more accurately and with tighter, better-calibrated posteriors than an otherwise identical phase-agnostic baseline, while remaining far cheaper than repeated likelihood evaluations. The same modular pipeline is presented as a practical route toward scalable inference for the larger, more realistic PTA data sets expected in the coming years.

Core claim

A hierarchical Transformer that injects the analytical gravitational-wave phase evolution of an eccentric binary as a gated positional encoding, when paired with conditional normalizing flows, produces amortized posterior distributions that are sharper, better calibrated, and computationally cheaper than those obtained from an otherwise identical physics-agnostic Transformer on the same synthetic multi-pulsar residuals.

What carries the argument

Physics-informed positional encoding (PIPE): the instantaneous orbital phase of the eccentric binary is computed analytically, wrapped, and added (via a learned scalar gate) to the ordinary sinusoidal positional embedding of each temporal token before the hierarchical Transformer encoder; the resulting context vector conditions a normalizing-flow density estimator.

Load-bearing premise

Every reported accuracy and calibration number is measured on synthetic residuals that contain only white Gaussian noise, with no red noise, dispersion-measure variations, or Hellings–Downs correlations.

What would settle it

Train and evaluate the identical architecture on the same synthetic sources after replacing the white-noise covariance with a realistic multi-component PTA noise model that includes red noise and Hellings–Downs correlations; if the phase-informed advantage in log posterior density and coverage disappears, the central claim fails under realistic conditions.

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

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

3 major / 6 minor

Summary. The paper proposes a hierarchical Transformer encoder with physics-informed orbital-phase encodings (PIPE), derived from the post-Newtonian GW phase of eccentric SMBHBs, combined with discrete and continuous conditional normalizing flows for amortized simulation-based inference of EBBH parameters from multi-pulsar PTA timing residuals. A separate phase-prediction network recovers a shared Earth-term phase trajectory and realization-level SNR from noisy residuals; the predicted phase is then gated into the Transformer via learnable encodings. On synthetic white-noise PTA realizations (10 pulsars, L=400), the phase-conditioned models improve log posterior density at truth (Table 4: DNF from −1.197 to 0.805), sharpen posteriors relative to phase-agnostic baselines (Figs. 5–6), and achieve near-nominal high-credibility coverage, with a limited comparison to conventional Bayesian sampling in a large-data regime.

Significance. The work addresses a genuine gap: SBI for deterministic eccentric continuous-wave sources in PTA data, as opposed to the SGWB-focused literature. Embedding analytical PN phase evolution into Transformer positional encodings is a concrete, reusable inductive bias, and the modular pipeline (phase provider + hierarchical attention + DNF/CNF) is clearly engineered for amortization. The controlled synthetic experiments—phase MAE, LPD at truth, coverage, and a Bayesian comparison on 4×10^5 realizations—are carefully reported and support the claimed gains under the stated noise model. If the approach extends robustly to red noise and Hellings–Downs structure, it would be a useful complement to MCMC pipelines for next-generation PTA continuous-wave searches. The present evidence, however, is confined to white-noise synthetics, so the significance for realistic PTA analyses remains prospective rather than demonstrated.

major comments (3)
  1. [Abstract; Sec. 3 Eqs. 18–21; Sec. 7; Sec. 8] Abstract and Sec. 8 claim the framework as a scalable alternative for next-generation PTA datasets and that it “readily generalizes” to red noise and additional components, yet every quantitative result (Secs. 3, 7; Eqs. 18–21; Table 4; Figs. 5–7) uses a diagonal white-noise covariance C=σ_r²I with no red noise, DM variations, or Hellings–Downs correlations. The large-data experiment (Sec. 7.4) already shows that the benefit of explicit phase conditioning shrinks once the network can learn structure from data alone; structured noise would further weaken the shared Earth-term phase signal that PIPE injects. The central claim of improved accuracy/sharpness is supported only under the white-noise axiom. Either temper the abstract/conclusion claims to the demonstrated regime, or add at least one controlled experiment with red noise (or a simple HD component) to show that the LPD and calibrat
  2. [Sec. 7.4; Fig. 6] The Bayesian comparison in Sec. 7.4 and Fig. 6 is limited to a single representative validation sample in the large-data regime and does not report wall-clock cost, effective sample size, or a systematic multi-realization comparison of posterior means, widths, or KL divergence. Without a broader head-to-head (e.g., median LPD or coverage relative to the same likelihood), the claim of “faster inference compared to physics-agnostic baselines” and the positioning against conventional Bayesian pipelines remain only partially substantiated. A quantitative multi-sample comparison table would make this load-bearing claim falsifiable.
  3. [Sec. 7.5; Fig. 7] Fig. 7 shows substantial under/over-coverage at the 68% level for several parameters even with predicted phase, while 99.7% coverage is near nominal. The paper reports this but does not diagnose whether the miscalibration is due to flow capacity, phase-prediction error, or the restricted SNR training range. Because the abstract and Sec. 7 emphasize “sharper posteriors” and improved calibration, the 68% discrepancy needs either a fix (e.g., temperature scaling, more expressive flow, or recalibration) or an explicit caveat that sharpness gains come with imperfect frequentist coverage at 1σ.
minor comments (6)
  1. [Title; Abstract] Title and abstract use “Robust Detection,” but the experiments are almost entirely parameter estimation (posterior sharpness, LPD, coverage); no ROC/detection-threshold analysis is presented. Align title/abstract language with the actual evaluation.
  2. [Table 2; Sec. 5] Table 2 lists targets as (n0, e0, M, S) while Sec. 5 and Table 4 use (log10 n, e0, log10 M, log10 S); keep notation consistent throughout.
  3. [Sec. 4.1.1 Eq. (26); Sec. 8] Eq. (26) for gate gradients is standard backprop; a brief note that ω_ϕ is learned end-to-end and can down-weight bad phase predictions would help readers interpret the adaptive-safeguard claim in Sec. 8.
  4. [Sec. 2.3; Fig. 1] Fig. 1 caption and Sec. 2.3 fix γ0=l0=0; state whether this is also true for the training prior or only for the illustrative figure.
  5. [Sec. 5; Sec. 7.1] Clarify whether the phase predictor is trained on the same 4×10^5 set used for the large-data posterior experiment or a disjoint split, to avoid any train–test leakage concern for the phase channel.
  6. [Fig. 3; throughout] Minor typos: “F eature extraction” in Fig. 3 caption; “realisation”/“realization” spelling is mixed; “an nPN correction” footnote formatting.

Circularity Check

0 steps flagged

No significant circularity: physics-informed encodings and phase prediction are supervised features from the known PN waveform model, trained and evaluated on independent synthetic draws under standard SBI.

full rationale

The paper's derivation chain is self-contained and non-circular. The orbital phase φ(t) (Eq. 14) and residual model R(t) (Eqs. 7–9) are taken from the established post-Newtonian literature and used only to generate synthetic training data and to supervise a separate phase-prediction network (Sec. 4.2). That network is frozen before posterior training; the main Transformer+flow never sees the true phase (Sec. 4.3, Sec. 5: "the true phase is never exposed"). Posteriors are obtained by amortized conditional density estimation on held-out realisations whose parameters and noise realisations are drawn independently of the training set (Sec. 3, Sec. 7). Log-posterior-density gains (Table 4) and coverage (Fig. 7) are therefore empirical comparisons between two feature sets (predicted-phase vs. no-phase), not tautologies forced by construction or by self-citation. No free parameter is fitted to a subset of data and then re-labelled a prediction; no uniqueness theorem is imported from the authors' prior work; and the white-noise covariance (Eq. 19) is an explicit modelling choice, not a circular definition of the claimed accuracy. The framework is ordinary physics-informed SBI.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 1 invented entities

The central empirical claim rests on standard post-Newtonian waveform physics, a white-noise PTA likelihood, and a collection of neural-network design choices and loss weights that are free parameters of the method. No new physical entities are postulated; the only invented construct is the gated physics-informed phase encoding itself.

free parameters (4)
  • phase-loss weights (κ_von, α_smooth, α_spec, λ_SNR) = 8 / 0.10 / 0.05 / 0.05
    Fixed by hand to 8, 0.10, 0.05, 0.05 (Sec. 4.2.1) without systematic search; they control how strongly the phase predictor is driven.
  • Transformer latent dimension, heads, layers, patch size, memory size
    Architectural hyper-parameters (d=128, 8 heads, 4 layers, patch=20, memory=40) chosen without ablation and required for the reported performance.
  • learnable gates ω_pos, ω_ϕ
    Scalar multipliers that decide how much positional versus phase encoding is used; optimized jointly with the network and therefore free parameters of the model.
  • SNR training range and noise realization procedure = SNR ∈ [10,100] for phase net; [20,30] for main results
    log-uniform SNR draws and the exact white-noise scaling σ_r = ||s_r||₂ / SNR_r define the training distribution on which all metrics are computed.
axioms (3)
  • domain assumption 3PN quasi-Keplerian orbital dynamics and quadrupole GW polarizations correctly describe the Earth-term PTA residual for eccentric SMBHBs
    Secs. 2.1–2.3 import the standard PN waveform model of Damour, Gopakumar et al. without re-derivation; all synthetic data and phase encodings rest on it.
  • ad hoc to paper White-noise covariance C = σ²I is an adequate first approximation for testing the inference pipeline
    Explicitly adopted in Sec. 3 and Eq. (19); the authors acknowledge that red noise and HD correlations are omitted.
  • domain assumption Self-attention plus external-attention memory can capture the long-range temporal and inter-pulsar correlations present in PTA residuals
    Architectural premise of Sec. 4.1; standard in sequence modeling but unproven a priori for PTA signals.
invented entities (1)
  • Physics-informed orbital-phase encoding (PIPE) no independent evidence
    purpose: Injects the analytical GW phase ϕ(t) into the Transformer’s positional stream so that attention operates in a physically meaningful coordinate.
    Defined in Sec. 4.1.1, Eqs. (24)–(26); the only novel architectural primitive introduced by the paper.

pith-pipeline@v1.1.0-grok45 · 26442 in / 2768 out tokens · 25810 ms · 2026-07-11T23:07:06.158685+00:00 · methodology

0 comments
read the original abstract

Pulsar timing arrays (PTAs) provide a unique window into nanohertz gravitational waves (GWs), but extracting astrophysical parameters from noisy, long-baseline timing residuals remains computationally challenging with traditional Bayesian techniques due to the high dimensionality of the parameter space, complex and correlated noise models, and the cost of repeated likelihood evaluations. We introduce a Transformer with a physics-informed positional-encoding framework for the efficient inference of eccentric binary black holes in relativistic orbits from PTA data. Our approach embeds analytical GW phase evolution directly into the model through structured positional encodings, enabling the network to learn physically meaningful representations from raw PTA timing residuals. We then use generative models, including discrete and continuous conditional normalizing flows, to infer posterior distributions within a simulation-based inference framework. Across a range of signal-to-noise ratios, the proposed method achieves improved accuracy, sharper posteriors, and faster inference compared to physics-agnostic baselines. While presented for deterministic white-noise signals, the modular framework readily generalizes to realistic PTA analyses incorporating red noise and additional components. This work highlights the potential of physics-aware deep learning models as scalable alternatives to conventional inference pipelines for next-generation PTA datasets.

Figures

Figures reproduced from arXiv: 2607.03904 by Alvin J. K. Chua, Subhajit Dandapat.

Figure 1
Figure 1. Figure 1: Left: Orbital trajectory of the secondary black hole about the primary (fixed at the origin) in the center-of-mass frame. Relativistic periastron precession due to post-Newtonian effects is visible. Spatial coordinates are shown in units of 100 GM/c2 . The arrow indicates the direction of motion. Parameters used are M = 109M⊙, q = 1, P = 4 yr, e0 = 0.6, ι = 0, ψ = 0, γ0 = 0, l0 = 0, DGW = 1 Gpc, and tref =… view at source ↗
Figure 2
Figure 2. Figure 2: Combined overview of the phase-informed simulation-based inference pipeline. (a) The hierarchical physics-informed SBI architecture uses patch embedding, positional and optional phase encoding, temporal attention blocks, token pooling, and cross-pulsar attention to form a global context vector h. This context conditions a normalizing-flow posterior estimator for θ, while the dashed branch denotes an option… view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of our physics-informed simulation-based inference pipeline for EBBH systems in PTA data. Part 1: Feature extraction (top). The PTA residual time series X is tokenized into features T, augmented with gated positional and orbital-phase encodings, and processed by External-Attention Transformer blocks to produce a compact context vector h. Part 2: Inference (bottom). A conditional normal… view at source ↗
Figure 4
Figure 4. Figure 4: Phase and SNR prediction performance of the phase-conditioned model evaluated on a representative validation sample. Top-left: Clean and noisy timing residuals for all pulsars in the array, showing the multi-pulsar realisation used as input. Top-right: Distribution of phase alignment error across the full validation set, quantified by the mean absolute wrapped phase difference, ⟨|∆ϕ|⟩, per sample, with mea… view at source ↗
Figure 5
Figure 5. Figure 5: Posterior distributions for a representative validation sample with SNR ∈ [20, 30], comparing no-phase and predicted-phase encodings for the DNF architecture (top) and CNF architecture (bottom). Posteriors are shown for {log10 n, e0, log10 M, log10 S}, with 1σ and 2σ contours. Dashed black lines indicate the injected parameter values. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Full posterior distributions over the EBBH parameter space for a representative validation sample, comparing DNF-based SBI models without phase encoding (no ϕ), with predicted phase encoding (pred ϕ), and Bayesian inference. The true parameter values are indicated by black dashed lines. The model is trained on a large PTA realisation dataset (∼ 4 × 105 samples, each comprising 10 pulsars). In this data-ric… view at source ↗
Figure 7
Figure 7. Figure 7: Marginal coverage comparison for simulation-based inference (DNFs) using no phase encoding (no ϕ) and predicted phase encoding (pred ϕ), evaluated at nominal credibility levels of 68%, 95%, and 99.7% (1σ, 2σ, and 3σ), over the validation set (10%) of the expanded PTA dataset comprising ∼ 4 × 105 realisations (10 pulsars per realisation) with SNR ∈ [20, 30]. Dashed horizontal lines indicate the ideal expect… 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

83 extracted references · 2 canonical work pages

  1. [1]

    Abbott B Pet al.(LIGO Scientific, Virgo) 2016Phys. Rev. Lett.116061102 (Preprint 1602.03837)

    Pith/arXiv arXiv

  2. [2]

    Agazie G, Anumarlapudi A, Archibald A M, Arzoumanian Z, Baker P Tet al.2023The Astrophysical Journal Letters951L8 URL https://dx.doi.org/10.3847/2041-8213/acdac6

    doi:10.3847/2041-8213/acdac6 2041

  3. [3]

    Astrophys.678A49 (Preprint2306.16225)

    Antoniadis Jet al.(EPTA, InPTA) 2023Astron. Astrophys.678A49 (Preprint2306.16225)

    Pith/arXiv arXiv

  4. [4]

    Reardon D J, Zic A, Shannon R M, Hobbs G B, Bailes Met al.2023The Astrophysical Journal Letters951L6 URLhttps://dx.doi.org/10.3847/2041-8213/acdd02

    doi:10.3847/2041-8213/acdd02 2041

  5. [5]

    Xu H, Chen S, Guo Y, Jiang J, Wang Bet al.2023Research in Astronomy and Astrophysics 23075024 URLhttps://dx.doi.org/10.1088/1674-4527/acdfa5

    doi:10.1088/1674-4527/acdfa5

  6. [6]

    Perera B, DeCesar M, Demorest P, Kerr M, Lentati L, Nice D, Os lowski S, Ransom S, Keith M, Arzoumanian Zet al.2019Monthly Notices of the Royal Astronomical Society490 4666–4687

  7. [7]

    Agazie G, Antoniadis J, Anumarlapudi A, Archibald A M, Arumugam Pet al.2023arXiv e-printsarXiv:2309.00693 (Preprint2309.00693)

    Pith/arXiv arXiv

  8. [8]

    Phinney E 2001arXiv preprint astro-ph/0108028

    Pith/arXiv arXiv

  9. [9]

    Agazie G, Anumarlapudi A, Archibald A M, Arzoumanian Z, Baker P T, Becsy B, Blecha L, Brazier A, Brook P R, Burke-Spolaor Set al.2023The Astrophysical Journal Letters951L50

  10. [10]

    Antoniadis J, Arumugam P, Arumugam S, Babak S, Bagchi M, Nielsen A S B, Bassa C, Bathula A, Berthereau A, Bonetti Met al.2024Astronomy & Astrophysics690A118

  11. [11]

    Babak S, Falxa M, Franciolini G and Pieroni M 2024Phys. Rev. D110063022 21 Dandapat & Chua

  12. [12]

    Van Haasteren R, Levin Y, McDonald P and Lu T 2009Mon. Not. R. Astron. Soc.395 1005–1014

  13. [13]

    Ellis J A, Vallisneri M, Taylor S R and Baker P T 2020 Enterprise: Enhanced numerical toolbox enabling a robust pulsar inference suite Zenodo URL https://doi.org/10.5281/zenodo.4059815

    doi:10.5281/zenodo.4059815 2020

  14. [14]

    Ellis J and van Haasteren R 2017 jellis18/ptmcmcsampler: Official release URL https://doi.org/10.5281/zenodo.1037579

    doi:10.5281/zenodo.1037579 2017

  15. [15]

    Taylor S R and Gair J R 2013Phys. Rev. D88084001

  16. [16]

    Falxa M, Antoniadis J, Champion D J, Cognard I, Desvignes G, Guillemot L, Hu H, Janssen G, Jawor J, Karuppusamy R, Keith M J, Kramer M, Lackeos K, Liu K, McKee J W, Perrodin D, Sanidas S A, Shaifullah G M and Theureau G 2024Phys. Rev. D109123010

  17. [17]

    Freedman G E, Johnson A D, Van Haasteren R and Vigeland S J 2023Phys. Rev. D107 043013

  18. [18]

    Van Haasteren R and Vallisneri M 2014Phys. Rev. D90104012

  19. [19]

    Hobbs Get al.2010Class. Quant. Grav.27084013 (Preprint0911.5206)

    Pith/arXiv arXiv

  20. [20]

    George D and Huerta E A 2018Physics Letters B77864–70 (Preprint1711.03121)

    Pith/arXiv arXiv

  21. [21]

    Gabbard H, Williams M, Hayes F and Messenger C 2018Physical review letters120141103

  22. [22]

    Nousi P, Koloniari A E, Passalis N, Iosif P, Stergioulas N and Tefas A 2023Physical Review D 108024022

  23. [23]

    Chatterjee C, Petulante A, Jani K, Spencer-Smith J, Hu Y, Lau R, Fu H, Hoang T, Zhao S C and Deshmukh S 2024arXiv preprint arXiv:2412.20789

    Pith/arXiv arXiv

  24. [24]

    Chua A J and Vallisneri M 2020Physical review letters124041102

  25. [25]

    Dax M, Green S R, Gair J, Macke J H, Buonanno A and Sch¨ olkopf B 2021Physical review letters127241103

  26. [26]

    Gabbard H, Messenger C, Heng I S, Tonolini F and Murray-Smith R 2022Nature Phys.18 112–117 (Preprint1909.06296)

    Pith/arXiv arXiv

  27. [27]

    Dax M, Green S R, Gair J, Gupte N, P¨ urrer M, Raymond V, Wildberger J, Macke J H, Buonanno A and Sch¨ olkopf B 2025Nature63949–53

  28. [28]

    Cuoco Eet al.2023Living Reviews in Relativity262 (Preprint2210.05659) URL https://arxiv.org/abs/2210.05659

    Pith/arXiv arXiv

  29. [29]

    Chen M, Zhong Y, Feng Y, Li D and Li J 2020Science China Physics, Mechanics & Astronomy63129511

  30. [30]

    Liang B and Wang H 2025Astronomical Techniques and Instruments

  31. [31]

    Kobyzev I, Prince S J and Brubaker M A 2020IEEE transactions on pattern analysis and machine intelligence433964–3979

  32. [32]

    Papamakarios G, Nalisnick E, Rezende D J, Mohamed S and Lakshminarayanan B 2021 Journal of Machine Learning Research221–64

    2021

  33. [33]

    Shih D, Freytsis M, Taylor S R, Dror J A and Smyth N 2024Physical review letters133 011402

  34. [34]

    Vallisneri M, Crisostomi M, Johnson A D and Meyers P M 2024arXiv preprint arXiv:2405.08857

    Pith/arXiv arXiv

  35. [35]

    Liang Bet al.2024 (Preprint2412.19169) URLhttps://arxiv.org/abs/2412.19169

    Pith/arXiv arXiv 2024

  36. [36]

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, Kaiser L and Polosukhin I 2017Advances in Neural Information Processing Systems (NeurIPS)30URL https://arxiv.org/abs/1706.03762 22 Dandapat & Chua

    Pith/arXiv arXiv

  37. [37]

    OpenAI 2025 Introducing GPT-5https://openai.com/index/introducing-gpt-5/ accessed: 2025-12-03

    2025

  38. [39]

    Wen Q, Zhou T, Zhang C, Chen W, Ma Z, Yan J and Sun L 2022arXiv preprint arXiv:2202.07125

    Pith/arXiv arXiv

  39. [40]

    Hwang S Y, Sabiu C G, Park I and Hong S E 2023JCAP11075 (Preprint2304.08192)

    Pith/arXiv arXiv

  40. [41]

    Allam Jr T and McEwen J D 2024RAS Techniques and Instruments3209–223

  41. [42]

    Instrum.3472–483 (Preprint 2406.12515)

    Boersma O M, Ayache E H and van Leeuwen J 2024RAS Tech. Instrum.3472–483 (Preprint 2406.12515)

    Pith/arXiv arXiv

  42. [43]

    Hochreiter S and Schmidhuber J 1997Neural Computation91735–1780 ISSN 0899-7667 URL https://doi.org/10.1162/neco.1997.9.8.1735

    doi:10.1162/neco.1997.9.8.1735 1997

  43. [44]

    Jadhav S, Wang Y T, Gabbard H and Cavagli` a M 2024Classical and Quantum Gravity (Preprint2403.12345)

    Pith/arXiv arXiv

  44. [45]

    Leyde K, Green S R, Dax M, Mould M, Fabbri C M and Gair J 2026arXiv preprint arXiv:2605.11274

    Pith/arXiv arXiv

  45. [46]

    Shen H, Chen H and Li T 2022Physical Review D105123005

  46. [47]

    Dom´ ınguez A, Udall R, Ashton G, Torres-Forn´ e A and Font J A 2024Monthly Notices of the Royal Astronomical Society(Preprint2402.06789)

    Pith/arXiv arXiv

  47. [48]

    Benedikt M and Saltas I 2025Machine Learning: Science and Technology

  48. [49]

    Karniadakis G E, Kevrekidis I G, Lu L, Perdikaris P, Wang S and Yang L 2021Nature Reviews Physics3422–440

  49. [50]

    Li H, Jung E, Chen Z, Wang Z, Wang Y, Qu H and Lau A K H 2025arXiv preprint arXiv:2506.14786

    Pith/arXiv arXiv

  50. [51]

    Blanchet L 2014Living Reviews in Relativity17(1) 2 ISSN 14338351

  51. [52]

    Jenet F A, Lommen A, Larson S L and Wen L 2004The Astrophysical Journal606(2) 799–803 ISSN 0004-637X URLhttps://doi.org/10.1086/383020

    doi:10.1086/383020

  52. [53]

    Taylor S R, Huerta E A, Gair J R and McWilliams S T 2016The Astrophysical Journal 817(1) 70 ISSN 1538-4357 URLhttp://stacks.iop.org/0004-637X/817/i=1/a=70

  53. [54]

    Susobhanan A 2023Classical and Quantum Gravity40155014 URL https://dx.doi.org/10.1088/1361-6382/ace234

    work page doi:10.1088/1361-6382/ace234

  54. [55]

    Susobhanan A, Gopakumar A, Hobbs G and Taylor S R 2020Physical Review D101(4) 043022 ISSN 2470-0010 URLhttps://link.aps.org/doi/10.1103/PhysRevD.101.043022

    doi:10.1103/physrevd.101.043022

  55. [56]

    Thorne K S, Misner C W and Wheeler J A 2000Gravitation(Freeman San Francisco)

  56. [57]

    Einstein A 1918Sitzungsberichte der Königlich Preussischen Akademie der Wissenschaften154–167

  57. [58]

    Gopakumar A and Iyer B R 2002Physical Review D65084011

  58. [59]

    Detweiler S 1979The Astrophysical Journal2341100–1104

  59. [60]

    Physique th´ eorique43(1) 107–132 ISSN 0246-0211 URLhttp://www.numdam.org/item/AIHPA_1985__43_1_107_0

    Damour T and Deruelle N 1985Annales de l’I.H.P. Physique th´ eorique43(1) 107–132 ISSN 0246-0211 URLhttp://www.numdam.org/item/AIHPA_1985__43_1_107_0

  60. [61]

    Memmesheimer R M, Gopakumar A and Sch¨ afer G 2004Physical Review D70(10) 17 ISSN 15502368 URLhttps://link.aps.org/doi/10.1103/PhysRevD.70.104011

    doi:10.1103/physrevd.70.104011

  61. [62]

    Boetzel Y, Susobhanan A, Gopakumar A, Klein A and Jetzer P 2017Physical Review D96(4) 044011 ISSN 2470-0010 URL https://link.aps.org/doi/10.1103/PhysRevD.96.044011http: //link.aps.org/doi/10.1103/PhysRevD.96.044011 23 Dandapat & Chua

    work page doi:10.1103/physrevd.96.044011http:

  62. [63]

    Damour T, Gopakumar A and Iyer B R 2004Physical Review D70(6) 064028 ISSN 1550-7998 URLhttps://link.aps.org/doi/10.1103/PhysRevD.70.064028

    doi:10.1103/physrevd.70.064028

  63. [64]

    Agazie Get al.(NANOGrav) 2023Astrophys. J. Lett.951L8 (Preprint2306.16213)

    Pith/arXiv arXiv

  64. [65]

    Petrov P, Taylor S R, Charisi M and Ma C P 2024The Astrophysical Journal976129

  65. [66]

    Taylor S R 2021arXiv preprint arXiv:2105.13270

    Pith/arXiv arXiv

  66. [67]

    van Haasteren R and Levin Y 2013Monthly Notices of the Royal Astronomical Society428 1147–1159

  67. [68]

    Guo M H, Liu Z N, Mu T J and Hu S M 2022IEEE Transactions on Pattern Analysis and Machine Intelligence455436–5447

  68. [69]

    Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, Dehghani M, Minderer M, Heigold G, Gelly Set al.2020arXiv preprint arXiv:2010.11929

    Pith/arXiv arXiv 2010

  69. [70]

    Hendrycks D and Gimpel K 2016arXiv preprint arXiv:1606.08415

    Pith/arXiv arXiv

  70. [71]

    Nair V and Hinton G E 2010 Rectified linear units improve restricted boltzmann machines Proceedings of the 27th international conference on machine learning (ICML-10)pp 807–814

    2010

  71. [72]

    265, Feb

    Hellings R and Downs G 1983Astrophysical Journal, Part 2-Letters to the Editor, vol. 265, Feb. 15, 1983, p. L39-L42.265L39–L42

    1983

  72. [73]

    Rezende D J and Mohamed S 2015Proceedings of the 32nd International Conference on Machine Learning (ICML)(Preprint1505.05770)

    Pith/arXiv arXiv

  73. [74]

    Papamakarios G, Nalisnick E, Rezende D J, Mohamed S and Lakshminarayanan B 2021 Journal of Machine Learning Research221–64 (Preprint1912.02762)

    Pith/arXiv arXiv 2021

  74. [75]

    Chen R T Q, Rubanova Y, Bettencourt J and Duvenaud D 2018Advances in Neural Information Processing Systems (NeurIPS)31(Preprint1806.07366)

    Pith/arXiv arXiv

  75. [76]

    Dinh L, Sohl-Dickstein J and Bengio S 2017International Conference on Learning Representations (ICLR)(Preprint1605.08803)

    Pith/arXiv arXiv

  76. [77]

    Kingma D P and Dhariwal P 2018Advances in Neural Information Processing Systems (NeurIPS)31(Preprint1807.03039)

    Pith/arXiv arXiv

  77. [78]

    Papamakarios G and Murray I 2017Advances in Neural Information Processing Systems (NeurIPS)29(Preprint1605.06376)

    Pith/arXiv arXiv

  78. [79]

    Hutchinson M F 1990Communications in Statistics – Simulation and Computation19 433–450

  79. [80]

    Grathwohl W, Chen R T Q, Bettencourt J, Sutskever I and Duvenaud D 2019International Conference on Learning Representations (ICLR)(Preprint1810.01367)

    Pith/arXiv arXiv

  80. [81]

    Lentati L, Shannon R M, Coles W A, Verbiest J P, van Haasteren R, Ellis J, Caballero R, Manchester R N, Arzoumanian Z, Babak Set al.2016Monthly Notices of the Royal Astronomical Society4582161–2187

Showing first 80 references.