{"id":"429d9b1f-70a7-46d2-a901-85c761803385","arxiv_id":"1908.00906","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":10,"one_line_summary":"Flexible Bayesian reconstructions of Planck 2018 data confirm a power-law primordial spectrum on 50<ell<2000 and leave low-ell oscillation hints statistically insignificant.","lead":"This paper reconstructs the primordial power spectrum and the inflationary potential from Planck 2018 data using flexible Bayesian splines and top-hat features. It finds no evidence for departures from a simple power law over most observable scales, while low-multipole hints of an oscillation remain statistically weak.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract and conclusion disagree about whether the low-ell oscillation survives marginalization, and no Bayes factor is given; the no-deviation claim should be conditional on that quantification.","rationale":"The reader's formal weakest assumption is the fixed flat Lambda-CDM late-time cosmology and foreground model. I regard that as a genuine but explicit and standard limitation, not the place where the central claim is least secure. The sharper, internally checkable problem is that the manuscript's own abstract and conclusions make contradictory assertions about the low-multipole oscillation after marginalization, and no quantitative significance statement is supplied. This is not a manufacturing of a doubt: the text itself contains the discrepancy, and the central no-deviation claim is framed differently depending on which passage one takes as authoritative. The paper has independent support in its use of public Planck likelihoods, a standard Boltzmann code, nested sampling, and the historical comparison across CMB datasets, and I do not question the broad conclusion that Planck 2018 does not require primordial deviations for 50<ell<2000. But until the marginalization survival is quantified, the abstract overstates the low-ell result and the reader's CONDITIONAL verdict remains the right one. My concrete test would settle the contradiction directly: a feature-vs-no-feature Bayes factor with the feature prior placed where the abstract claims a preserved hint. If the Bayes factor is small, the abstract should be revised; if large, the no-deviation claim must be qualified. Either way, the paper should report this number explicitly.","tokens_in":24952,"tokens_out":13058,"duration_ms":157798,"concrete_test":"Using the same Planck 2018 TT,TE,EE+lowE+lensing likelihood and priors, compute the Bayes factor between the sharp-feature model of Sec. V with a single top-hat whose position prior is restricted to the k-range corresponding to 20<ell<50 and the N=0 Lambda-CDM model. Report log Z for both and the difference with an uncertainty estimate from at least three independent PolyChord runs (different live points and seeds). If ln B<1, the conclusion-version is correct and the abstract should be revised; if ln B>2, the abstract is correct and the no-deviation claim is incomplete because a feature is preferred after marginalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is an internal inconsistency in the paper's central framing. The abstract says the ell~20-50 oscillation is 'to some extent preserved upon marginalization'; Sec. VII says 'the oscillations do not survive marginalization over N, indicating that the Bayesian evidence is not strong enough.' Sections III and V likewise claim the low-k feature 'survives this marginalization' and 'still comes through clearly in the fully marginalized plot.' These claims are mutually incompatible in any quantitative sense. The headline no-deviation result depends on the Sec. VII reading; the abstract advertises the stronger reading. No numerical Bayes factor or posterior probability for a feature at ell~20-50 is reported anywhere, and the evidence panels (Figs. 4, 7, 12) show only relative evidence as a function of knot number N, not feature-vs-no-feature odds. Sampling uncertainties on log Z are also not quoted. This matters because the strongest claim in the abstract is precisely this marginalization survival; without a number, a reader cannot tell whether the hint is a weak fluctuation or a preferred model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper performs three Bayesian non-parametric reconstructions of the primordial scalar power spectrum from Planck 2018 TT,TE,EE+lowE+lensing data: a linear spline in (log k, log P), a cubic-spline reconstruction of the inflationary potential, and a top-hat 'sharp features' parameterization. For each reconstruction, models with different numbers of knots/features N are compared via nested-sampling evidences with PolyChord, and the posteriors are marginalized over N. The main claims are: (i) all three methods find no evidence for deviations from a tilted power law over 50<ell<2000; (ii) the data prefer parameterizations that can reproduce the lack of constraining power at low and high ell; (iii) there are conditional hints of an oscillation at 20<ell<50 that the abstract says are 'to some extent preserved upon marginalization' but the conclusions say do not survive marginalization; and (iv) late-time cosmological parameters are stable across all reconstructions.","tokens_in":25257,"tokens_out":6980,"duration_ms":68526,"significance":"If the quantitative issues identified below are addressed, the paper would provide a useful methodological reference and a consistent Bayesian update of free-form reconstructions to Planck 2018, confirming with an independent pipeline the Planck Collaboration's conclusion that the power spectrum is consistent with a power law over the well-measured multipole range. The paper's strengths include the use of the full Planck likelihood with 21 nuisance parameters, nested sampling with PolyChord, marginalization over the number of knots, conditional KL divergences, functional posterior plots (with the fgivenx code made publicly available), and a historical comparison across CMB datasets. The parameter-stability result in Fig. 15 is a valuable cross-check. However, the absence of numerical evidence values and the internal inconsistency about the low-ell oscillation prevent the current version from being fully assessable.","major_comments":[{"comment":"The manuscript is internally inconsistent about the status of the low-multipole oscillation after marginalization over the knot number N. The abstract says the feature is 'to some extent preserved upon marginalization'; Sec. III (Results) says 'hints of the low-k features survive this marginalization'; Sec. V (Results) says the 'low-k oscillation still comes through clearly in the fully marginalized plot'; but Sec. VII (Conclusions) says 'the oscillations do not survive marginalization over N, indicating that the Bayesian evidence is not strong enough.' These statements are not compatible without a quantitative definition of 'survive' and a reported Bayes factor or posterior probability for a feature at 20<ell<50; as written, the reader cannot tell whether the abstract is advertising a detection or the conclusion is retracting it.","section":"Abstract; Sec. III; Sec. V; Sec. VII"},{"comment":"No numerical evidence values are reported. The evidence panels show only relative evidence normalized to the best N, with no log Z values or sampling uncertainties, so statements such as 'N=3 is greater than N=2' (Sec. III), 'N=1 is preferred over N=0' (Sec. IV), and 'little Bayesian evidence to support the introduction of more than two features' (Sec. V) cannot be checked. Please provide a table of log Z (or delta log Z with nested-sampling errors) for all models, and ideally a quantitative Bayes factor for any model containing a low-ell feature versus the featureless tilted power law, so that the central no-deviation claim is supported by numbers rather than by visual inspection of the marginalized plots.","section":"Figs. 4, 7, 12; Secs. III, IV, V"},{"comment":"The central no-deviation conclusion is conditional on the assumption of a flat Lambda-CDM late-time cosmology and on the Planck 2018 foreground/nuisance model, as stated in Sec. II E. Because all residual features in the CMB spectra are attributed to the primordial power spectrum, a mis-specified late-time or foreground model could be absorbed into the reconstructed PPS and bias the conclusion. A simple robustness test (e.g., varying the late-time equation of state, or using an alternative foreground parametrization) would make the claim considerably more secure; without it, the conclusion should be explicitly labeled as conditional on the assumed late-time model.","section":"Sec. II E; Sec. VI"}],"minor_comments":[{"comment":"The sentence 'In the same manner as Sec. III, Fig. 8 and is consequently a form of exponential potential.' appears to be corrupted; it should be rephrased to describe the N=0 potential as an exponential potential and to refer to the appropriate figure.","section":"Sec. IV, Results"},{"comment":"The paper uses 'non-parametric' in the title and abstract while Sec. II D itself notes the terminology is misleading; consider qualifying 'non-parametric' as 'free-form' or 'flexible' in the abstract for consistency.","section":"Sec. II D and Abstract"},{"comment":"The prior on dlnV*/dphi is tabulated as log-uniform over [10^-3,10^-0.3] but the text says it is 'negatively log-uniform'; please clarify the sign convention so the table and text agree.","section":"Tab. III"},{"comment":"The statement that 'for large N, in a fraction of the samples there is a visibly clear oscillation' would be more informative if the fraction were quantified or the sentence reworded to describe the conditional posterior contours.","section":"Sec. III, Results"},{"comment":"The use of 'Planck 2018 polarization data baseline, referred to as TT,TE,EE+lowE+lensing' is potentially confusing because this baseline includes temperature; please rephrase or explain that 'polarization data baseline' is the standard Planck naming.","section":"Sec. II E"}],"recommendation":"major_revision","confidential_remarks":"The paper is by the developers of PolyChord and CosmoChord and cites several of their own tools; this is standard in the field and not a concern by itself. The main issue is that the abstract and conclusions disagree on the status of the low-l oscillation, and no evidence numbers are given; this must be fixed before publication. The paper is within scope for a cosmology journal and is likely to be useful once revised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, careful paper that deserves refereeing, but the abstract and conclusions genuinely disagree about whether the low-ell oscillation survives marginalization, and no number is given for its significance. That internal inconsistency is the main thing to fix before the framing can be taken at face value.\n\nWhat is actually new: the spline and top-hat reconstructions are not new in themselves, but this is the first time they are run against the Planck 2018 TT,TE,EE+lowE+lensing likelihood. That produces a concrete result: across the broad observable window, all three methods recover a featureless tilted power law consistent with (As, ns), and the late-time cosmological parameters are unperturbed by the extra primordial degrees of freedom. The Bayesian treatment of the knot number N, the use of PolyChord for evidence computation, and the KL-divergence-based compression plots are all done cleanly and with appropriate care. The historical reconstruction sequence (COBE through Planck 2018) is a nice touch.\n\nThe central null result is well supported. If I only took away one sentence from this paper, it would be: flexible reconstructions of the primordial spectrum do not disturb the standard tilted power-law conclusion or the late-time parameters over the Planck window. That claim is backed by the figures and by the standard pipeline.\n\nThe soft spot is the one your reader flagged. The abstract says the low-multipole oscillation is “to some extent preserved upon marginalization,” but Sec. VII says the oscillations “do not survive marginalization over N.” Those are not the same statement, and the difference matters. The evidence panels give relative evidence as a function of N, but not a feature-versus-no-feature Bayes factor, and sampling uncertainties on log Z are not quoted. So a reader cannot tell whether the hint at ell 20–50 is a weak fluctuation or a preferred model. The paper's own conclusion is the more cautious reading, and I think that is the right one, but the abstract is advertising something stronger. That should be reconciled.\n\nMinor: fixing flat Lambda-CDM is a normal modeling choice, not a flaw, though it does mean the reconstructions only probe primordial deviations within that late-time cosmology. The paper acknowledges this. The high degree of self-citation is also not a problem here because the methods genuinely come from the cited lines of work.\n\nWho this is for: cosmologists working on primordial spectrum reconstruction and inflation phenomenology. It is a useful reference, not a paradigm-shifter. A serious referee should engage with it; the main revision is to make the abstract and conclusions say the same thing and, ideally, to quote an actual Bayes factor for the low-ell feature. I would send it to review.","headline":"A solid, careful null-result paper that deserves refereeing, but the abstract and conclusions tell different stories about the low-ell oscillation, and the lack of a quantitative Bayes factor means the headline should be read as conditional.","tokens_in":25743,"tokens_out":1186,"would_cite":true,"duration_ms":13891,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Planck 2018 spectra need no primordial features, three methods agree","keywords":["primordial power spectrum reconstruction","Bayesian evidence","Planck 2018","cosmic microwave background","inflation","non-parametric inference","nested sampling","CMB features"],"falsifier":"Re-run the same three reconstructions on the same Planck 2018 likelihood while freeing late-time physics beyond flat $\\Lambda$CDM---for example, allowing curvature, a varying dark-energy equation of state, or a more flexible reionization history---and check whether the recovered spectrum remains a pure power law for $50\\lesssim\\ell\\lesssim2000$ and whether the $\\ell\\sim20$--$50$ oscillation persists. If the features move or vanish, the primordial attribution is not robust.","tokens_in":1771,"feed_emoji":"📡","tokens_out":2845,"duration_ms":99267,"temperature":0.7,"pith_summary":"The paper tries to establish that Planck 2018 temperature and polarization data, analyzed without assuming a specific inflationary model, point to a featureless tilted power law for primordial fluctuations over most of the observable window ($50\\lesssim\\ell\\lesssim2000$). Three independent free-form reconstructions---a linear spline for the primordial spectrum, a cubic spline for the inflationary potential, and a top-hat feature search---agree on this conclusion. They also agree that the Bayesian evidence rules out a scale-invariant spectrum by enormous odds and that the late-time cosmological parameters are barely affected by the added primordial freedom. The only recurring hint of new physics is a weak oscillatory feature around $\\ell\\sim20$--$50$, which appears conditionally in all three methods but does not survive marginalization over model complexity. A sympathetic reader should care because the result sets the bar for inflation models: Planck 2018 does not require primordial deviations from a power law, while leaving a specific low-multipole target for future data.","feed_headline":"Planck 2018 spectra need no primordial features, three methods agree","feed_subtitle":"Spline and feature-based Bayesian analyses find only a weak hint of an oscillation at multipoles 20-50; the rest is a simple power law.","key_machinery":"The central machinery is Bayesian free-form reconstruction with evidence-based model selection. The primordial spectrum is represented as a linear spline in the $(\\log k,\\log P)$ plane; the inflationary potential is represented by integrating twice a linear spline for the second derivative of $\\log V(\\phi)$, which guarantees a smooth potential; and sharp features are modeled as top-hat additions to the standard $(A_s,n_s)$ spectrum. Knot positions are sorted via an identifiability prior, the number of knots $N$ is treated as a discrete model parameter, and the Bayesian evidence weights each $N$ when producing marginalized functional posteriors. Conditional Kullback-Leibler divergences quantify where the data actually constrain the reconstructed function, and a nested-sampling engine provides the posterior samples and evidences needed to navigate the high-dimensional, multimodal parameter space.","core_discovery":"All three reconstructions describe the same data with the same outcome. Conditioned on a fixed number of knots, the linear-spline primordial power spectrum, the cubic-spline inflationary potential, and the sharp-feature parameterization each show a preference for extra flexibility at the edges of the observed window---where cosmic variance and instrument noise degrade the signal---and each conditionally produces a dip at $20<\\ell<30$ followed by a rise near $\\ell\\sim50$. When the number of parameters is treated as a model choice and marginalized over, that oscillation is too weakly supported to count as a detection, and the recovered primordial spectrum in the window $50\\lesssim\\ell\\lesssim2000$ is consistent with a simple tilted power law. The paper also establishes two auxiliary results: the scale-invariant spectrum is excluded with odds around a quintillion to one, and the late-time cosmological parameters remain stable when the primordial sector is given additional degrees of freedom.","pith_inferences":["Inference: The preference for an intermediate number of knots may partly reflect a prior-volume effect, where extra knots absorb cosmic-variance and noise-limited edges, so the preferred model need not correspond to a physical feature.","Inference: If the low-multipole oscillation is real, future large-scale polarization measurements with independent foreground handling should recover the same dip and peak; the paper's historical reconstructions show this feature persisting across earlier data releases.","Inference: The parameter-stability result suggests that joint analyses with non-CMB probes that fix late-time cosmology are unlikely to be biased by primordial spectrum flexibility, but this should be checked out of sample.","Inference: Applying the same evidence-marginalized spline machinery to next-generation CMB surveys would sharpen the test: if the broad-window power law and the low-$\\ell$ oscillation persist, the case for a simple primordial spectrum grows; if they shift, the attribution to primordial physics would need revision."],"forward_implications":["If the central claim is correct, Planck 2018 data do not require any primordial deviation from a tilted power law, so inflation models predicting only smooth spectra remain viable.","The Bayesian evidence against scale invariance means a Harrison-Zeldovich spectrum is effectively excluded by the data used here.","Because the late-time cosmological parameters remain stable when the primordial sector is made more flexible, cosmological parameter constraints from Planck are unlikely to depend strongly on the choice of primordial spectrum parameterization.","The recurring conditional hint of an oscillation at $20<\\ell<30$ and a peak near $\\ell\\sim50$ gives a specific target for future CMB data and for inflationary models that predict such features."],"supporting_citations":[{"why":"Supplies the Planck 2018 data release and overview that define the dataset analyzed.","marker":"[1]"},{"why":"Supplies the Planck 2018 CMB power spectra and likelihood used for all reconstructions.","marker":"[2]"},{"why":"Supplies the Planck 2018 cosmological parameter baseline that the late-time parameter stability comparison uses.","marker":"[3]"},{"why":"Supplies the Planck 2018 constraints on inflation, including the tensor bound and tilt, used for comparison and for ruling out scale invariance.","marker":"[4]"},{"why":"Establishes the spline-based Bayesian primordial power spectrum reconstruction approach the paper extends.","marker":"[5]"},{"why":"Introduces the conditional Kullback-Leibler divergence used to quantify where the data constrain the reconstructed function.","marker":"[10]"},{"why":"Provides the Boltzmann solver that turns each reconstructed primordial spectrum into CMB power spectra.","marker":"[33]"},{"why":"Provides the nested-sampling algorithm used to compute posteriors and Bayesian evidences in high-dimensional parameter spaces.","marker":"[39]"},{"why":"Supplies the nested-sampling implementation details and the identifiability prior used for knot ordering.","marker":"[40]"}],"fun_headline_variants":["Three Bayesian methods, one quiet verdict for Planck spectra","Planck 2018: no primordial features, just a gentle oscillation hint","Bayesian trio finds Planck spectra consistent with simple power law","Oscillation at low multipoles too weak to matter, say three models","Scale-invariant ruled out, but no new physics needed for Planck"],"cache_read_input_tokens":27904,"weakest_assumption_plain":"The load-bearing premise is that a flat $\\Lambda$CDM late-time model with the Planck 2018 likelihood and its 21 nuisance parameters is correct, so that any residual CMB features must come from the primordial spectrum.","fun_headline_variants_meta":{"raw":{"variants":["Three Bayesian methods, one quiet verdict for Planck spectra","Planck 2018: no primordial features, just a gentle oscillation hint","Bayesian trio finds Planck spectra consistent with simple power law","Oscillation at low multipoles too weak to matter, say three models","Scale-invariant ruled out, but no new physics needed for Planck"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1831,"prompt_tokens":914,"completion_tokens":917,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":827}},"tokens_in":530,"tokens_out":917,"duration_ms":8959,"temperature":1.0,"reasoning_tokens":827,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:29:01.474146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same three reconstructions on the same Planck 2018 likelihood while freeing late-time physics beyond flat $\\Lambda$CDM---for example, allowing curvature, a varying dark-energy equation of state, or a more flexible reionization history---and check whether the recovered spectrum remains a pure power law for $50\\lesssim\\ell\\lesssim2000$ and whether the $\\ell\\sim20$--$50$ oscillation persists. If the features move or vanish, the primordial attribution is not robust.","supporting_citations":[],"review_version":1}