Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Starobinsky in Stereo: SKA-CMB Synergy in SBI

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SKA 21 cm data can constrain Starobinsky inflation as tightly as Planck, and combining the two sharpens the inflationary mass bound by a factor of about 9.

desk verdict A well-executed SBI forecast showing SKA+CMB can tighten Starobinsky constraints by ~9x, but the headline number leans on an optimistic foreground wedge and the simulator isn't public yet. read the letter →

arxiv 2508.10094 v2 pith:CQIV5SIA submitted 2025-08-13 astro-ph.CO astro-ph.IMphysics.data-an

Benedikt Schosser , Caroline Heneka , Björn Malte Schäfer This is my paper
classification astro-ph.COastro-ph.IMphysics.data-an
keywords Starobinskyinflation21cmcosmologysimulation-basedinferenceconditionalflowmatchingSquareKilometreArraycosmicmicrowavebackgroundneuralsummarystatisticsEpochofReionisation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Simulation-based inference can turn the 21 cm hydrogen line, observed during the Epoch of Reionisation with the SKA, into a competitive test of inflationary physics. The paper argues, on realistic mock observations, that SKA 21 cm data alone constrains Starobinsky $R+R^2$ inflation almost as well as Planck CMB data, and that combining the two sharpens all parameter bounds. The headline result is a 95\% interval on the Starobinsky mass parameter $M$ of $[1.0772,\,1.0871]\times 10^{-5}\,M_{\mathrm{Pl}}$ at fixed $N_\star=67$, about 9.15 times tighter than Planck alone. It also shows that learned neural summaries extract more information from 21 cm maps than the conventional power spectrum, which matters for designing future SKA analyses.

What carries the argument

Conditional flow matching (CFM), a simulation-based inference estimator that learns a neural vector field transporting a simple base distribution into the posterior, with the training loss evaluated purely from simulator samples. For 21 cm light cones the data are first compressed by either a 3D convolutional network or a vision transformer, and the compressed summary is concatenated with the CMB angular power spectrum before entering the flow. The forward model couples a Starobinsky-modified CLASS (primordial and CMB spectra) to 21cmFAST (reionisation light cones) and 21cmSense (SKA-Low thermal noise and foreground wedge). The $M$--$N_\star$ degeneracy, set by $A_s \propto M^2 N_\star^2$, i

What would settle it

Re-run the forecast with the foreground wedge covering the full horizon rather than only the primary field of view and check whether the 95\% HPD interval on $M$ at $N_\star=67$ still shrinks by roughly a factor of 9.15 relative to Planck alone; if the factor drops well below that, the headline claim fails. A direct cross-check is to feed the trained pipeline simulated data from a different inflationary potential and see whether the posterior correctly excludes the Starobinsky fiducial.

Watch

Extended reading notes

Core claim

Tomographic 21 cm observations of the Epoch of Reionisation, analysed with simulation-based inference and learned summaries, can test Starobinsky's $R+R^2$ inflation at a level competitive with Planck, and a joint SKA+CMB analysis sharpens every parameter. In mock observations from CLASS and 21cmFAST, SKA alone constrains $h$ and $\Omega_{\mathrm{cdm}}$ comparably to Planck, while the inflationary parameters $M$ and $N_\star$ stay prior-dominated because of the degeneracy $A_s \propto M^2 N_\star^2$. The combination breaks part of it: at fixed $N_\star = 67$ the 95\% HPD interval for $M$ is $[1.0772,\,1.0871]\times 10^{-5}\,M_{\mathrm{Pl}}$, a factor 9.15 smaller than Planck alone. Mutual in

Load-bearing premise

The forecast is only as good as the mock pipeline: the modified CLASS plus 21cmFAST forward model and the 21cmSense SKA-Low noise model, together with the optimistic assumption that the foreground wedge covers only the instrument's primary field of view, must faithfully represent the real telescope and the 21 cm signal.

Editorial extensions

If this is right

  • SKA alone constrains $h$ and $\Omega_{\mathrm{cdm}}$ with precision comparable to Planck in a nine-parameter analysis, so 21 cm tomography can serve as an independent high-redshift cosmological probe.
  • Combining SKA and Planck reduces the 95\% credible intervals on $h$ and $\Omega_{\mathrm{cdm}}$ by a factor of about 1.5 relative to Planck alone, and slightly improves the baryon density constraint.
  • At fixed $N_\star=67$, the Starobinsky mass is bounded to $[1.0772,\,1.0871]\times10^{-5}\,M_{\mathrm{Pl}}$ at 95\%, a 9.15-fold tightening over Planck alone.
  • Learned summaries (CNN and ViT) carry about 3--4 nats more mutual information than the 1D power spectrum, implying power-spectrum-only 21 cm forecasts underestimate SKA's constraining power.
  • CMB data sharpens the Epoch-of-Reionisation astrophysical parameters, most notably $T_{\mathrm{vir}}$, through its sensitivity to the reionisation optical depth, even though only the TT spectrum was used.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the pipeline transfers to real data, SKA would give an inflation test at $z\sim5$--$35$ that does not rely on CMB assumptions about reionisation; this is a new observational handle on reheating and the duration of inflation.
  • The same joint SBI machinery can be pointed at other single-field potentials (Higgs inflation, $\alpha$-attractors), where an analogous $M$--$N_\star$ degeneracy is expected; the relative gain from adding 21 cm data may differ.
  • The optimistic foreground-wedge assumption is the main lever: with a more realistic wedge excluding more $k$-modes, the quoted factor 9.15 would shrink; the paper's own over-confidence when summary dimension grows suggests the gain is data-limited and could degrade with larger parameter spaces.
  • The large mutual-information gap between band powers and learned summaries suggests future SKA analyses should move beyond power-spectrum statistics; non-Gaussian information is not a small correction but several nats.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents a simulation-based inference (SBI) forecast for Starobinsky inflation using mock SKA 21-cm tomographic data, alone and in combination with Planck CMB data. The pipeline combines a modified CLASS code for the Starobinsky primordial power spectrum with 21cmFAST for reionization-era light cones, adds thermal noise and a foreground wedge via 21cmSense, and trains a conditional flow matching (CFM) posterior estimator on neural summaries (3D CNN, Vision Transformer) as well as on the traditional 1D power spectrum. The authors validate the SBI setup against MCMC on Planck data, perform simulation-based calibration (SBC) and TARP coverage tests, and report mutual information estimates showing that neural summaries outperform the power spectrum. The headline forecast is that SKA alone will constrain h and Omega_cdm with precision comparable to Planck, and that the joint SKA+Planck analysis reduces the conditional 95% HPD interval on the Starobinsky mass parameter M at fixed N* = 67 by a factor of 9.15 compared to Planck alone.

Significance. If the forward model is accepted, this is a useful and technically careful forecast. Its strengths are the explicit end-to-end pipeline, the validation on Planck data against MCMC, the use of SBC and TARP for posterior calibration, the parameter-recovery checks, and the comparison of learned versus handcrafted summaries within a single SBI framework. The paper also demonstrates a practical joint multi-probe SBI analysis, which is of methodological interest for upcoming 21-cm surveys. The central quantitative claims are internally consistent with the stated simulator. However, the value as a forecast depends on the fidelity of the noise/foreground model and on the public availability of the modified simulation codes; these are the least secure points. The paper is not claiming an actual detection, and the self-consistency of the mock analysis is standard for forecasts, but the abstract's phrase 'realistic mock observations' is stronger than the explicitly optimistic foreground scenario used.

major comments (4)
  1. [§4, '21 cm light cone data'] The SKA constraints and the factor 9.15 quoted in §6.4 are conditioned on 'the optimistic scenario where the foreground wedge in k-space covers only the instrument's primary field-of-view.' This is the most optimistic of commonly considered foreground wedge treatments; a wedge with a buffer or a horizon-limited wedge removes significantly more k-space volume. Since the abstract and conclusion describe the forecasts as based on 'realistic mock observations' and state that SKA alone will be competitive with Planck, this assumption is load-bearing for the headline result. I ask for a quantitative robustness test under a more conservative wedge model, or, failing that, for the claims to be explicitly restated as conditional on the most optimistic foreground scenario. Without this, the quoted constraining power and the factor 9.15 cannot be taken as a robust forecast.
  2. [§4 and Data availability] The forward model is the basis of all results, but the modifications to CLASS and 21cmFAST are described only in general terms ('we have made modifications', 'we modified the code'), and the code is not publicly released until after the refereeing process. No validation of the modified Starobinsky implementation is shown, e.g., a comparison of the generated primordial power spectrum P_R(k) and C_l^TT with the analytic slow-roll prediction for the Starobinsky model. This prevents independent verification of the mock likelihood and of the key forecast. I recommend that the code be released together with the paper or as a clearly specified patch, and that a consistency test of the modified CLASS output be added.
  3. [§6.2, Eqs. (6.3)–(6.4)] The mutual information values in Table 2 are obtained by replacing the true posterior p(θ|T(y)) in Eq. (6.4) with the learned CFM approximation q(θ|T(y)). The paper does not quantify how sensitive the reported MI values are to this approximation. The SBC and TARP tests are coverage checks on posterior samples, not accuracy checks on the density used to compute MI. Since the MI ranking is a central claimed result ('neural summaries outperform power spectra'), the authors should either provide evidence that q is sufficiently accurate for MI estimation (for example, by comparing against a case with a known sufficient statistic or by bootstrap over network training) or explicitly frame Table 2 as 'mutual information under the trained CFM approximation' rather than as the intrinsic information content of the summaries.
  4. [§6.4, Fig. 6 and abstract] The headline interval M ∈ [1.0772, 1.0871]×10^-5 M_Pl is conditioned on a fixed fiducial value N* = 67, and the paper itself notes that the marginal posteriors for M and N* are largely prior-dominated. The factor 9.15 therefore describes an improvement in a conditional, fiducial-dependent slice, not a marginal constraint on the Starobinsky model. The text is mostly transparent on this, but the abstract's phrase 'tighten bounds on both inflationary and ΛCDM parameters considerably' and the concluding 'precise constraints on both fundamental cosmological parameters' risk overstating the result. Please state explicitly in the abstract that the tight M constraint is conditional on N*.
minor comments (5)
  1. [Eq. (2.3)] The exponential in the Starobinsky potential appears garbled ('exp(-\vt2/3 φ/M_Pl)'); the intended factor is presumably sqrt(2/3) φ/M_Pl. Please correct the typesetting.
  2. [§5, 'Pre-processing and data augmentations'] The sentence 'We take a similar approach for the 1dPS. We rescale with the logarithm' is awkward; it should read 'We take the logarithm and then rescale ...'.
  3. [Table 2] The Monte Carlo errors on the mutual information values are quoted, but the number of samples used for the MC estimate and the exact estimator (e.g., nearest-neighbour or density-based) are not specified. A sentence describing the estimator and sample size would be helpful.
  4. [§6.3, Figure 4] The SBC histograms are shown for the ViT summary only in the main text; the corresponding plots for CNN and 1dPS are relegated to Appendix B. Since the calibration claim is made for all summaries, a brief discussion in the main text of any differences would strengthen the presentation.
  5. [§6.2, Table 2] For the ViT, the summary dimension is fixed to 48 'due to computational constraints', whereas the CNN is tested at dimensions 9 and 32. This asymmetry makes the comparison slightly uneven; a sentence acknowledging that the ViT dimension was not optimized to the same extent would be fair.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the forecast is self-contained and externally anchored by Planck/MCMC validation; the MI diagnostics and mock-data forecasts are internal but clearly labeled as such.

full rationale

The paper's central claims are forecasts based on mock observations generated with the same forward model used for training, which is standard SBI forecast practice and explicitly labeled as such ('Forecasts based on realistic mock observations'). The pipeline is anchored externally in Section 6.1, where SBI posteriors are validated against MCMC on real Planck data with the Plik lite likelihood, and the two are 'almost identical'. The SBC/TARP checks in Section 6.3 validate internal calibration but are not used to claim external validity beyond the forecast. The mutual information analysis in Section 6.2 uses the learned posterior to estimate information content, which is an internal diagnostic; it is not load-bearing for the headline factor 9.15 or for the SKA competitiveness claim, and it is not a construction that forces the result. The main caveats—the optimistic foreground wedge assumption in Section 4 and the not-yet-public code in the Data availability section—affect robustness and reproducibility, not circularity. No fitted parameter is renamed as a prediction, no load-bearing self-citation chain is invoked, and no uniqueness theorem is imported from the authors' prior work.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The forecast rests on imported inflation model equations, the 21cmFAST simulation, and an optimistic noise model; the posteriors are conditional on these choices.

free parameters (2)
  • Fiducial astrophysical parameters (zeta, Tvir, LX, E0) = zeta=118, log Tvir=4.2, log LX=40, E0=500 eV
    Hand-picked fiducial values from the Planck best-fit (Section 6.4); forecasted constraints are evaluated at this point and may vary with the fiducial choice.
  • Summary dimensions for CNN and ViT = 32 (CNN), 48 (ViT)
    Chosen by scanning 9-96 to maximize mutual information (Section 6.2); the ViT-vs-CNN ranking depends on this hyperparameter choice.
assumptions (5)
  • standard math Starobinsky action and Einstein-frame potential (Eqs. 2.1-2.3) are taken as given.
    The model is imported from prior literature, not derived here.
  • domain assumption 21cmFAST semi-numerical simulations with the chosen astrophysical parameters faithfully represent the 21 cm signal.
    The central forecast inherits all limitations of the simulation code (Section 4).
  • domain assumption The noise model (Plik covariance for CMB, 21cmSense thermal noise for 21 cm) and the 'optimistic' foreground wedge closure are representative of real SKA/Planck observations.
    Section 4: 'We consider the optimistic scenario where the foreground wedge in k-space covers only the instrument's primary field-of-view.'
  • domain assumption tau_reio is fully determined by the astrophysical parameters and is not an independent parameter.
    This mapping is stated in Section 4 and differs from standard Planck analyses.
  • ad hoc to paper The learnt summaries (CNN, ViT) and the CFM posterior are sufficiently accurate for the mutual information comparison.
    MI values in Table 2 are computed with the same approximate posterior that the networks learn; this is not an external benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Starobinsky in Stereo: SKA-CMB Synergy in SBI." pith.science (2026). https://pith.science/paper/CQIV5SIA

@misc{pith2026250810094,
  author       = {Pith},
  title        = {Pith review of: Starobinsky in Stereo: SKA-CMB Synergy in SBI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CQIV5SIA}},
  note         = {Machine review of arXiv:2508.10094}
}
abstract

Modern machine learning techniques can unlock the vast cosmological information encoded in forthcoming Square Kilometre Array (SKA) observations. We show that tomographic 21 cm data from the reionisation era can yield stringent tests of inflationary models - here illustrated with Starobinsky $R+R^2$ inflation. Using a simulation-based inference (SBI) framework, we compare neural summaries (convolutional network and vision transformer) with a traditional power spectrum summary and perform a fully joint SBI analysis combining 21 cm data with data of the cosmic microwave background (CMB). Forecasts based on realistic mock observations indicate that SKA alone will achieve constraints competitive with Planck, and that the combined SKA + CMB dataset will tighten bounds on both inflationary and $\Lambda\mathrm{CDM}$ parameters considerably while improving precision on key astrophysical quantities.

Figures

Figures reproduced from arXiv: 2508.10094 by the authors.

Figure 1
Figure 1. Inference pipeline to constrain Starobinsky Inflation with SKA and Planck. During training (dashed lines) model parameters are given to the simulator, a combination of CLASS and 21cmFAST, yielding a mock angular power spectrum (upper plot) and a 21 cm light cone (lower plot). The light cone is summarised by one of three methods, before it is combined with the spectrum. A neural network is optimised to learn the vect… view at source ↗
Figure 2
Figure 2. Implicit prior on τreio defined from the priors on the astrophysical parameters in Ta￾ble 1 The CMB simulations have the same true labelled parame￾ters as the 21 cm simulations, i.e. they include the astrophysical parameters but not τreio. This comes with two implications for our analysis: we have an empirical prior on τreio, which is different from the typical Planck analysis, and we infer the astrophysical paramet… view at source ↗
Figure 3
Figure 3. Comparison of SBI and MCMC for Planck data with the Plik lite TT high-ℓ likelihood. Both posteriors have ≈ 500 000 samples. MCMC requires one CLASS call for each sample. SBI used 32 000 parallelisable CLASS calls, eight minutes of training and 20 seconds for inference. We can evaluate the expectation value with Monte Carlo samples, while the posterior and prior density are accessible due to our SBI setup. Here, the … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Validation metrics for different datasets, left SBC and right TARP. The black line indicates perfect calibration and the grey area the 99 % C.I. given the finite number of samples. Another performance check is parameter recovery or calibration. Each data point of the t…
Figure 5
Figure 5. Figure 5: Parameter recovery for different datasets. The true parameter values (known from simulation) are plotted on the x-axis and the mariginalised 1d posteriors, illustrated as mean and standard deviation, are shown on the y-axis. The 1600 test data points are binned for bet…
Figure 6
Figure 6. Figure 6: Posterior for cosmological and astrophysical parameters derived from mock SKA data (teal) using SBI, Planck data with SBI (pink) and joint SBI (yellow). The zoomed-in panel shows the the inferred correlation of N⋆ and M. We show the profiled distribution for a fixed N⋆…
Figure 7
Figure 7. Figure 7: Posterior for cosmological and astrophysical parameters derived from mock SKA data. The 21 cm light cone is summarised with the 1dPS (orange), a 3D CNN (purple) or a ViT (teal). coverage test because it requires only posterior samples. Evaluating exact probabilities is…
Figure 8
Figure 8. Figure 8: Parameter recovery from different summary networks for the SKA data. The true parameter values (known from simulation) are plotted on the x-axis and the mariginalised 1d posteriors, illustrated as mean and standard deviation, are shown on the y-axis. The 1600 test data…
Figure 9
Figure 9. Figure 9: Validation metrics for different summary methods, left SBC and right TARP. The black line indicates perfect calibration and the grey area the 99 % C.I. given the finite number of samples. Algorithm 2 TARP Coverage Evaluation Require: Posterior sampler p(θ | y); test se…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Constraining reionization morphology and source properties with 21cm galaxy cross-correlation surveys

    astro-ph.CO 2026-01 conditional novelty 6.0 of 10

    Simulation-based inference on mock 21cm-galaxy cross-power spectra shows these measurements can tightly constrain neutral fraction, overdensity, and source escape fraction and star-formation efficiency, if redshifts a...

  2. Machine Learning and the SKA for Cosmic Dawn and the Epoch of Reionization

    astro-ph.IM 2026-07 accept novelty 2.5 of 10

    A multi-author overview of machine-learning algorithms proposed for instrument modelling, data analysis, simulation and inference in SKA Cosmic Dawn and Epoch of Reionization science.

  3. Cosmology with Intensity Mapping via Statistics Beyond the Power Spectrum in the SKAO Era

    astro-ph.CO 2026-06 unverdicted novelty 2.0 of 10

    Reviews multiple higher-order statistics for 21-cm intensity mapping and forecasts their detectability with SKAO, incorporating noise and foreground effects.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages · cited by 3 Pith papers

  1. [2]

    Sato,First-order phase transition of a vacuum and the expansion of the universe,Monthly Notices of the Royal Astronomical Society195(1981) 467

    K. Sato,First-order phase transition of a vacuum and the expansion of the universe,Monthly Notices of the Royal Astronomical Society195(1981) 467. [3]A.A. Starobinsky ,A New Type of Isotropic Cosmological Models Without Singularity,Phys. Lett. B91 (1980) 99

    work page 1981

  2. [4]

    Starobinsky ,The Perturbation Spectrum Evolving from a Nonsingular Initially De-Sitter Cosmology and the Microwave Background Anisotropy,Sov

    A.A. Starobinsky ,The Perturbation Spectrum Evolving from a Nonsingular Initially De-Sitter Cosmology and the Microwave Background Anisotropy,Sov. Astron. Lett.9(1983) 302. [5]Y. Akrami et al.,Planck 2018 results. X. Constraints on inflation,Astron. Astrophys.641(2020) A10. [6]G. Papamakarios and I. Murray ,Fastε-free inference of simulation models with b...

    work page 1983

  3. [8]

    Neutsch, C

    S. Neutsch, C. Heneka and M. Brüggen,Inferring astrophysics and dark matter properties from 21 cm tomography using deep learning,Mon. Not. Roy. Astron. Soc.511(2022) 3446

    work page 2022

  4. [9]

    Schosser, C

    B. Schosser, C. Heneka and T . Plehn,Optimal, fast, and robust inference of reionization-era cosmology with the 21cmPIE-INN,SciPost Phys. Core8(2025) 037. – 22 –

    work page 2025

  5. [10]

    A. Ore, C. Heneka and T . Plehn,SKATR: A Self-Supervised Summary Transformer for SKA,SciPost Phys. 18(2025) 155. [11]N. Aghanim et al.,Planck 2018 results. V . CMB power spectra and likelihoods,Astron. Astrophys.641 (2020) A5

    work page 2025

  6. [12]

    Villaescusa-Navarro, J

    F . Villaescusa-Navarro, J. Ding, S. Genel, S. Tonnesen, V . La Torre, D.N. Spergel et al.,Cosmology with One Galaxy?,apj929(2022) 132[2201.02202]. [13]M. Reza, Y. Zhang, C. Avestruz, L.E. Strigari, S. Shevchuk and F . Villaescusa-Navarro,Constraining Cosmology with Simulation-based inference and Optical Galaxy Cluster Abundance, . [14]Í. Zubeldia, B. Bol...

    arXiv 2022

  7. [20]

    X. Zhao, Y. Mao, S. Zuo and B.D. Wandelt,Simulation-based Inference of Reionization Parameters from 3D Tomographic 21 cm Light-cone Images. II. Application of Solid Harmonic Wavelet Scattering Transform,Astrophys. J.973(2024) 41. [21]X. Zhao, Y. Mao, C. Cheng and B.D. Wandelt,Simulation-based Inference of Reionization Parameters from 3D Tomographic 21 cm ...

    work page 2024

  8. [24]

    Bacon, R.A

    D.J. Bacon, R.A. Battye, P . Bull, S. Camera, P .G. Ferreira, I. Harrison et al.,Cosmology with phase 1 of the square kilometre array red book 2018: Technical specifications and performance forecasts, Publications of the Astronomical Society of Australia37(2020) e007. [25]G. Autieri, M. Berti, M. Spinelli, B.S. Haridasu and M. Viel,Weighing neutrinos with...

    work page 2018

Show all 16 references
  1. [27]

    X.-W . Liu, C. Heneka and L. Amendola,Constraining coupled quintessence with the 21cm signal,JCAP 05(2020) 038. [28]A. Liu, J.R. Pritchard, R. Allison, A.R. Parsons, U. Seljak and B.D. Sherwin,Eliminating the optical depth nuisance from the CMB with 21 cm cosmology,Phys. Rev. ...

    2020

  2. [38]

    K. Wolz, N. Krachmalnicoff and L. Pagano,Inference of the optical depth to reionization τ from Planck CMB maps with convolutional neural networks,Astron. Astrophys.676(2023) A30

    2023

  3. [39]

    Franco Abellán, N

    G. Franco Abellán, N. Anau Montel, O. Savchenko and C. Weniger,How to embed any likelihood into SBI: Application to Planck+Stage IV galaxy surveys and Dynamical Dark Energy, . [40]R. Kallosh and A. Linde,On the present status of inflationary cosmology, 2025. [41]A. De Felice a...

    2025

  4. [42]

    Baumann,Inflation, inTheoretical Advanced Study Institute in Elementary Particle Physics: Physics of the Large and the Small, pp

    D. Baumann,Inflation, inTheoretical Advanced Study Institute in Elementary Particle Physics: Physics of the Large and the Small, pp. 523–686, 2011, DOI

    2011

  5. [43]

    Bassett, S

    B.A. Bassett, S. Tsujikawa and D. Wands,Inflation dynamics and reheating,Rev. Mod. Phys.78(2006) 537. [44]M. Dax, S.R. Green, J. Gair, M. Pürrer, J. Wildberger, J.H. Macke et al.,Neural importance sampling for rapid and reliable gravitational-wave inference,Phys. Rev. Lett.130...

    2006

  6. [47]

    R.T .Q. Chen, Y. Rubanova, J. Bettencourt and D.K. Duvenaud,Neural ordinary differential equations, inAdvances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi and R. Garnett, eds., Curran Associates, Inc. [48]PLANCKco...

    2018 arXiv

  7. [55]

    Pober, A

    J.C. Pober, A. Liu, J.S. Dillon, J.E. Aguirre, J.D. Bowman, R.F . Bradley et al.,What Next-generation 21 cm Power Spectrum Measurements can Teach us About the Epoch of Reionization,apj782(2014) 66. [56]I. Loshchilov and F . Hutter,Decoupled weight decay regularization, 2019. [...

    2014

  8. [58]

    Fisher,On the mathematical foundations of theoretical statistics,Philosophical Transactions of the Royal Society of London

    R.A. Fisher,On the mathematical foundations of theoretical statistics,Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character222 (1922) 309. [59]S. Talts, M. Betancourt, D. Simpson, A. Vehtari and A. Gelman...

    1922

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.