{"id":"c632d59b-ecef-464a-a29c-814e03078cc6","arxiv_id":"2506.23574","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new refitting technique maps any gravitational-wave background spectrum onto a running-power-law reference model via sensitivity-weighted chi-squared minimization, and uses the pullback of the reference posterior to approximate the model's posterior from PTA data.","lead":"The paper introduces a fast 'refitting' method that maps gravitational-wave background spectra from new-physics models onto a simple running-power-law model, using the existing pulsar timing array posterior to approximate new model posteriors. A generalist should read it because it offers a fast route to compare many theoretical models against PTA data before running expensive full Bayesian analyses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Dirac-delta identification in Eq. (12) is the load-bearing approximation: it discards the conditional width of the BSM-to-RPL map and may make the induced posterior in Eq. (15) artificially narrow where the chi-square landscape is shallow.","rationale":"I read the paper's central claim as Eq. (15): the pullback of the RPL posterior through the chi-square-defined map Phi gives a fast and accurate BSM posterior without a new MCMC. The validation on SIGWs (D_H=0.011) is impressive, but it does not establish the claim generally. The reader's weakest assumption is exactly right: Eq. (12) replaces the conditional density p(theta_RPL|theta_BSM) by a delta function at the best-fit point. This is not a harmless technical shortcut; it is the step that converts a genuine marginalization over latent RPL parameters into a point evaluation. The paper's own stable-strings example is the internal evidence that the approximation can break: for a poorly fitting model the point-map refit (D_H=0.161) is worse than the naive nCPL refit (D_H=0.064), which should not happen if the delta map were universally reliable. I do not see this as a fatal flaw, because the authors hedge their conclusion and the method is explicitly approximate, but the error budget is not quantified. A finite-width conditional test would either confirm that Eq. (15) is accurate over the prior support or reveal the regions where it is not. The reader's conditional verdict is appropriate; the recommended action is to require the finite-width check before relying on Eq. (15) as a replacement for MCMC.","tokens_in":13490,"tokens_out":4974,"duration_ms":59852,"concrete_test":"For the SIGW model, evaluate Eq. (8) with the delta in Eq. (12) replaced by a Gaussian conditional p(theta_RPL|theta_BSM) proportional to exp[-Delta-chi^2(theta_BSM, theta_RPL)/2], with covariance from the local Hessian of Delta-chi^2 at Phi(theta_BSM), and compare the resulting joint posterior with the full NG15 MCMC posterior using the same Hellinger distance as in Fig. 3. If the SIGW Hellinger distance changes by more than about 0.02, or if the stable-strings posterior widths change materially, the point-map approximation is the dominant uncontrolled error in Eq. (15).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (15) is the central claim: it turns the RPL posterior into a BSM posterior by composing P_RPL with the map Phi from the chi-square minimization. The step that makes this exact rather than heuristic is Eq. (12), p(theta_RPL|theta_BSM) = delta^(3)(theta_RPL - Phi(theta_BSM)). This point-mass conditional distribution assumes that, for fixed BSM parameters, the chi-square as a function of theta_RPL is so sharply peaked that the full likelihood integral in Eq. (8) is dominated by the single minimizer, and that P_RPL does not vary appreciably over the width of the conditional. The paper validates the approximation on three models, but the stable-strings case in Fig. 3 already shows the failure mode: D_H(RPL)=0.161 while the naive nCPL refit gives 0.064, i.e., the point-map refit is worse than the much cruder Taylor-map approach when the model is a poor fit. The reason is exactly that Phi(theta_BSM) can land in a region where the point evaluation of P_RPL is not representative of the chi-square-weighted neighborhood. For BSM parameter regions where Delta-chi^2 has a shallow valley, multiple RPL parameters have similar goodness of fit, and the delta replacement discards that spread, potentially making the induced posterior narrower than the true marginal posterior. Since the paper's stated goal is fast and accurate Bayesian inference for any BSM model, the delta assumption must be shown to be controlled over the full parameter space, not only at the validation points where the BSM model happens to fit well.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a fast approximate method for obtaining Bayesian posteriors of gravitational-wave background (GWB) spectral models from new physics, using the existing running-power-law (RPL) posterior from the NANOGrav 15-year data set as a reference. The method constructs a map Phi from beyond-Standard-Model (BSM) parameters to RPL parameters via a matched-filter chi-square minimization that accounts for the frequency dependence of the PTA sensitivity curve, and then uses the pullback of the RPL posterior as an induced likelihood on the BSM parameter space. The central result is Eq. (15), which expresses the BSM posterior as the product of the pullback posterior and a prior. The authors validate the method on three models (stable cosmic strings, metastable cosmic strings, and scalar-induced GWs) by comparing Hellinger distances against full MCMC fits from the NANOGrav new-physics analysis.","tokens_in":13819,"tokens_out":5398,"duration_ms":60844,"significance":"If the central approximation is controlled, this method would be a valuable tool: it would allow rapid refitting of many BSM spectral models to existing PTA data without running new MCMC chains. The mathematical chain from Eq. (5) to Eq. (15) is clearly laid out, and the comparison with naive pivot-based Taylor maps (nRPL/nCPL) is an informative benchmark. The validation against full MCMC results is a genuine strength: the Hellinger distances for SIGWs and metastable strings are encouragingly small (e.g., D_H=0.011 for the SIGW parameter Delta). The paper is also commendably transparent about the approximate nature of spectral refits. However, the load-bearing Dirac-delta approximation in Eq. (12) is not quantitatively controlled, and the stable-strings example already shows a failure mode where the new RPL method performs worse than the naive CPL method. This limits the generality of the claims and needs to be addressed before the method can be recommended as a routine substitute for direct MCMC analyses.","major_comments":[{"comment":"The identification p(theta_RPL|theta_BSM) = delta^(3)(theta_RPL - Phi(theta_BSM)) is load-bearing, but the paper does not demonstrate that the conditional distribution is narrow compared with the scale over which P_RPL varies. The stable-strings row in Fig. 3 is precisely the regime where the approximation is least controlled: D_H(RPL)=0.161 is larger than D_H(nCPL)=0.064. Please either (i) provide evidence for the narrowness of the conditional, for example by comparing against a Gaussian conditional whose width is set by the curvature of Delta chi^2 at Phi, or (ii) explicitly restrict the claimed domain of validity and give a diagnostic for identifying models for which the method is unreliable. As written, Eq. (15) may understate posterior uncertainties in shallow chi-square valleys.","section":"Induced likelihood, Eq. (12)"},{"comment":"The posterior P_RPL is reconstructed from MCMC samples via kernel density estimation, but the bandwidth (and kernel choice) is not reported. Because Eq. (13) evaluates P_RPL pointwise, the induced posterior and the Hellinger distances in Fig. 3 depend on the smoothing scale. Please specify the bandwidth selection procedure and show that the reported results are robust to reasonable variations in this free parameter.","section":"Running power law (RPL), KDE reconstruction"},{"comment":"The abstract states that the techniques 'provide the basis for fast and accurate Bayesian inference', but validation covers only three models, and one of them (stable strings) gives a hierarchy D_H(nCPL) < D_H(CPL) < D_H(RPL) < D_H(nRPL), i.e., the new RPL refit is the second-worst of the four methods for that model. The paper's own conclusion is more careful ('at least in the case of BSM models that yield a good fit'), but the abstract and Results section should carry this qualification explicitly, or the authors should propose a quantitative criterion for when the method is reliable.","section":"Results, Fig. 3 and Conclusions"}],"minor_comments":[{"comment":"There is a typo: 'plausable' should be 'plausible'.","section":"Introduction, first paragraph"},{"comment":"There is a typo: 'auch as' should be 'such as'.","section":"Introduction, second paragraph"},{"comment":"The notation 'log10( f∗/Hz)' and 'log10( Gμ)' is slightly ambiguous; please use consistent subscript and superscript formatting, e.g., log10(f_*/Hz) and log10(Gμ).","section":"Fig. 2 caption"},{"comment":"The conditional density p(theta_RPL|theta_BSM) is introduced as a probability density, but its normalization and support are never discussed; please clarify the measure with respect to which this density is defined.","section":"Eq. (8)"},{"comment":"The Hellinger distances are computed for one-dimensional marginalized posteriors; please state explicitly whether multivariate (joint) agreement was also checked, since the method is used for joint parameter inference.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is well written and the central idea is publishable after revision. The main technical weakness is the uncontrolled Dirac-delta approximation in Eq. (12), and the stable-strings example demonstrates a real failure mode. I do not view this as a reason for rejection, because the validation framework gives a clear path toward quantifying the approximation. No concerns about citation practices or novelty disclosure; the reliance on external MCMC results from Refs. [16] and [76] is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take: this is a genuinely new and useful refitting method for PTA spectral analysis, with a coherent derivation and a real validation exercise, but the delta-function assumption at its core is the soft spot, and the authors oversell it as always best when their own stable-strings example shows otherwise.\n\nWhat's new: the combination of an RPL reference model, a sensitivity-weighted chi-square map (Eq. 7), and the pullback of the RPL posterior to induce a BSM likelihood (Eq. 15). Earlier tools like ceffyl refit amplitude posteriors; this goes further. The validation against full MCMC on three models is a genuine strength—for SIGWs and metastable strings, Hellinger distances around 0.01–0.05, which is impressive for an approximate method.\n\nThe soft spot, as you'd expect, is Eq. (12): identifying p(theta_RPL|theta_BSM) as a delta function at the best-fit map. That discards the width of the chi-square landscape. If that landscape is shallow, the induced posterior gets artificially narrow. The paper's own stable-strings example shows the failure: DH(RPL)=0.161 versus DH(nCPL)=0.064, so the sophisticated map is worse than the naive local expansion. The authors attribute this to the poor fit of stable strings, but that's exactly when you need the conditional width to be handled properly. They should either fix the approximation (e.g., keep a Gaussian conditional) or add a diagnostic that tells you when the delta assumption is safe.\n\nThere's also a minor overstatement: the sentence 'RPL refits always yield the best results' is qualified two paragraphs later by the stable-strings exception, so it's not a true contradiction, but the general conclusion 'more accurate than nRPL, CPL, and nCPL' is not supported by that example. Soften it.\n\nOther soft spots: KDE bandwidth and grid resolutions aren't documented, which matters because the posterior is reconstructed from a KDE. And the paper doesn't discuss the computational cost of building the map Phi for high-dimensional BSM models; nested sampling on the pullback is mentioned, but the map itself requires a chi-square minimization per parameter point.\n\nOverall: the method is a solid contribution, the derivation is clean, and the validation is honest. The delta assumption needs to be either justified more thoroughly or relaxed, and the claims should be tempered. This deserves a serious referee; I'd accept it with major revisions, but it's not desk-rejectable.","headline":"Genuinely new refitting method for PTA spectra, validated on real MCMC chains, but the delta-function approximation in Eq. (12) is the load-bearing weak spot and the authors overstate its universality.","tokens_in":14379,"tokens_out":3551,"would_cite":true,"duration_ms":37884,"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":"A single PTA spectral posterior can refit any gravitational-wave model without new MCMC runs.","keywords":["gravitational-wave background","pulsar timing arrays","running power law","Bayesian inference","spectral refitting","matched filtering","cosmic strings","scalar-induced gravitational waves"],"falsifier":"Take any GWB model already fitted to the 15-year PTA data by full MCMC, and from the same chain compute the conditional distribution of the RPL parameters at fixed BSM parameters, for example by reweighting or by holding the BSM spectrum fixed at its best fit. If that conditional spread is comparable to the width of the RPL posterior itself, Eq. (15) cannot reproduce the full posterior, and the Hellinger distances would grow correspondingly; the stable-strings example, where the refit is worse than a naive pivot-based fit, already shows the regime in which the delta assumption fails.","tokens_in":13263,"feed_emoji":"📡","tokens_out":5818,"duration_ms":60722,"temperature":0.7,"pith_summary":"This paper claims that the expensive part of fitting exotic gravitational-wave background (GWB) models to pulsar timing array data—running a full Markov-chain Monte Carlo on the timing residuals—only needs to be done once. Starting from the posterior density of a running-power-law (RPL) reference model, the authors construct a map from any new-physics GWB spectrum to the three RPL parameters by minimizing a matched-filter chi-squared that weights frequencies by the array's sensitivity. Pulling the RPL posterior back along this map yields an induced likelihood on the new-physics parameters, so posterior inference reduces to evaluating the reference posterior at the best-fit RPL parameters and multiplying by the model's prior. They show the resulting posteriors closely reproduce full MCMC fits for stable cosmic strings, metastable strings, and scalar-induced gravitational waves, with Hellinger distances as small as 0.011 when the model actually fits the data. The payoff is fast, physically intuitive model comparison and parameter inference for the growing catalog of proposed explanations of the nanohertz signal.","feed_headline":"One PTA posterior can refit any GWB model without new MCMC","feed_subtitle":"Pull back the running-power-law posterior along a sensitivity-weighted map and every new-physics spectrum gets a Bayesian fit for free.","key_machinery":"The load-bearing object is the map $\\Phi$ defined in Eq. (7) as the minimizer of the chi-squared function in Eq. (5), which compares the BSM spectrum to an RPL template weighted by the PTA sensitivity curve across the full observing band. Composing this map with the RPL posterior—the pullback $\\Phi^* P_{\\mathrm{RPL}}$—converts a posterior on the reference model into a likelihood on the new-physics parameter space. That pullback, made explicit in Eq. (15), is what allows parameter inference and model comparison without rerunning the timing-residual analysis.","core_discovery":"The central result is Eq. (15): the posterior for a beyond-the-Standard-Model GWB spectrum factorizes as $P(\\theta_{\\mathrm{BSM}} | D) \\propto (P_{\\mathrm{RPL}} \\circ \\Phi)(\\theta_{\\mathrm{BSM}})\\, \\pi(\\theta_{\\mathrm{BSM}})$, up to the constant evidence ratio. Here $P_{\\mathrm{RPL}}$ is the three-parameter RPL posterior obtained from a single Bayesian fit to the 15-year PTA data, and $\\Phi$ maps BSM parameters to the RPL parameters that minimize the matched-filter chi-squared in Eq. (5). The paper thereby replaces a full MCMC fit for each new model with one map evaluation; for the SIGW example, the refitted posterior for the peak-width parameter $\\Delta$ has Hellinger distance 0.011 from the full MCMC result. The authors emphasize that $\\Phi$ need not be invertible, so the method extends to BSM models with more parameters than the reference model.","pith_inferences":["The same pullback trick could transfer to CMB spectral analysis: instead of Taylor-expanding slow-roll predictions at a pivot scale, one could map them to the posterior on the spectral index and its running, and pull that posterior back—an exact analogue the paper hints at but does not develop.","If the delta approximation is validated across a wider model zoo, the method turns each new PTA data release into a single RPL posterior that all model builders can refit at negligible cost, effectively decoupling data analysis from theory scanning.","The stable-string example suggests a useful diagnostic: before trusting a refit, check that the projected best-fit RPL point lies inside the high-probability region of the RPL posterior; if it lies far outside, the induced posterior is unreliable.","One could turn the chi-squared map itself into a goodness-of-fit statistic: the minimum chi-squared in Eq. (6) measures how far a BSM spectrum is from the best RPL description, providing a cheap model-comparison score without full inference."],"forward_implications":["Any GWB spectral model can be refit to the 15-year PTA data by evaluating the pullback of the RPL posterior; no new MCMC over timing residuals is required.","Model comparisons can be visualized in a spectral-index–amplitude plane analogous to the CMB $n_s$–$r$ plane, with global best-fit projections instead of pivot-frequency Taylor expansions.","RPL refits beat both CPL refits and naive pivot-based refits by Hellinger distance, for models that fit the NG15 data well.","The method works without constructing the inverse map, so it applies to BSM models with more than three parameters.","Because only one full Bayesian fit is needed, the RPL posterior can be built once and reused as a universal starting point for later refits and future data releases."],"supporting_citations":[{"why":"Supplies the RPL posterior density $P_{\\mathrm{RPL}}$ that is pulled back along the map; reconstructed by KDE from its MCMC chains.","marker":"[76]"},{"why":"Provides the full MCMC posterior fits of stable strings, metastable strings, and SIGWs to the NG15 data, used as the ground truth for Hellinger-distance comparisons.","marker":"[16]"},{"why":"Introduces the chi-squared spectral-comparison approach that the paper's matched-filter map $\\Phi$ is built on.","marker":"[79]"},{"why":"Supplies the NG15 sensitivity curve used to evaluate the chi-squared in Eq. (5).","marker":"[75]"},{"why":"Provides the optimal-SNR formula in Eq. (4) that motivates interpreting the chi-squared as an SNR for the differential spectrum.","marker":"[77,78]"},{"why":"Earlier rapid-refitting technique for PTA spectra that the paper positions as similar in spirit but less rigorous.","marker":"[69]"}],"fun_headline_variants":["Pullback trick refits any GWB model without new MCMC","One PTA posterior maps to any GWB model via chi-square","Refit GWB spectra by pulling back the RPL posterior","No MCMC per model: use a chi-square map to refit GWB","Pullback of RPL posterior refits any BSM GWB model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that for each set of new-physics parameters there is a single best-fitting running-power-law spectrum, and that all other RPL spectra at that best fit are irrelevant—formally, the conditional distribution $p(\\theta_{\\mathrm{RPL}} | \\theta_{\\mathrm{BSM}})$ is a Dirac delta at the map $\\Phi(\\theta_{\\mathrm{BSM}})$; if the true conditional distribution has appreciable width, the induced posterior will be artificially narrow.","fun_headline_variants_meta":{"raw":{"variants":["Pullback trick refits any GWB model without new MCMC","One PTA posterior maps to any GWB model via chi-square","Refit GWB spectra by pulling back the RPL posterior","No MCMC per model: use a chi-square map to refit GWB","Pullback of RPL posterior refits any BSM GWB model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2909,"prompt_tokens":971,"completion_tokens":1938,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":1844}},"tokens_in":587,"tokens_out":1938,"duration_ms":13432,"temperature":1.0,"reasoning_tokens":1844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:37:38.554834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any GWB model already fitted to the 15-year PTA data by full MCMC, and from the same chain compute the conditional distribution of the RPL parameters at fixed BSM parameters, for example by reweighting or by holding the BSM spectrum fixed at its best fit. If that conditional spread is comparable to the width of the RPL posterior itself, Eq. (15) cannot reproduce the full posterior, and the Hellinger distances would grow correspondingly; the stable-strings example, where the refit is worse than a naive pivot-based fit, already shows the regime in which the delta assumption fails.","supporting_citations":[],"review_version":1}