{"id":"18c13410-54e6-4984-b6ce-aee77077e897","arxiv_id":"2507.08695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Applying 10-fold cross-validation to Pantheon+ supernova data selects the (2,1) Padé approximation as the optimal form for luminosity distance reconstruction, consistent with prior cosmographic results.","lead":"This paper uses 10-fold cross-validation to choose the best Padé polynomial order for fitting supernova distance data, and finds that the (2,1) form wins. The scheme is a data-driven alternative to analytic rules for order selection in cosmography.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The (2,1) optimality claim is not statistically supported: the paper itself reports AIC/BIC differences are 'not significant' and gives no uncertainty on the CV scores, so the order ranking may be noise.","rationale":"The paper proposes 10-fold cross-validation as a general scheme for choosing the Padé order in luminosity-distance reconstruction and applies it to Pantheon+, concluding that (2,1) is optimal. For that conclusion to hold, the CV-based model-selection criteria must actually discriminate among the candidate orders. The text itself undercuts this: Sec. 4 says the AIC/BIC differences are 'not significant,' and no error bars are attached to any CV score. The AIC/BIC definitions are incomplete (no L_max, ambiguous m), and the j0 inconsistency between Eq. (6), the fixed-value block, and Table 1 makes the complexity penalty undefined. These are not cosmetic gaps; they are exactly the quantities that determine whether the reported ranking is meaningful. The central claim is plausible and agrees with earlier work by Capozziello et al. (2020), which provides some external support, but the paper's internal evidence for the claimed 'remarkable ability to distinguish' is not established. The reader's CONDITIONAL verdict already captures this weakness, and my stress-test does not move it; if anything, the manuscript's own admission of non-significance strengthens the need for the conditional wording. A concrete re-analysis with a defined likelihood, consistent parameter counts, and repeated-seed CV would settle whether the (2,1) preference is real or noise.","tokens_in":9453,"tokens_out":4365,"duration_ms":54088,"concrete_test":"Recompute the 10-fold CV scores with a fully specified likelihood for the Pantheon+ distance moduli, e.g. -2 ln L = Δμ^T C^{-1} Δμ using the full covariance matrix, and count m consistently (treat j0 as free in all orders or fix it in all orders). Repeat the CV for, say, 100 random fold seeds and report the mean and standard deviation of MSE/AIC/BIC for each order. Then test whether the (2,1) mean is separated from the next-best (e.g., (3,2)) by more than the fold-to-fold scatter, using a paired test or bootstrap confidence intervals. If the separation is not significant, the optimality claim in the abstract should be weakened; if it survives, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the 10-fold CV procedure identifies the (2,1) Padé approximation as optimal—depends on the statistical meaningfulness of the reported MSE/AIC/BIC rankings. That premise is insecure. In Sec. 4 the authors state that 'the difference of average of BIC and AIC is not significant,' yet the (2,1) selection is made by visually amplifying those small differences in insets. No likelihood function L_max is specified for Eqs. (11)–(12), so the AIC/BIC values cannot be reproduced. The effective parameter count m is also ambiguous: Eq. (6) contains j0 as a free coefficient while the 'where' block fixes j0 = 1, and Table 1 reports a fitted j0 = 1.5^{+1.1}_{-1.2}; different m choices change the complexity penalty and can flip an AIC/BIC ranking. No per-fold or repeated-CV uncertainties are reported, so the 'remarkable ability to distinguish' orders is unquantified. Because the differences the paper itself calls insignificant are used to select the winner, the conclusion that (2,1) is optimal could be an artifact of the random fold partition. This is load-bearing: the paper's entire contribution is an order-selection procedure, and the evidence for its success is the contested ranking.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a data-driven procedure, based on k-fold cross-validation, to select the order of a Padé polynomial used to reconstruct the luminosity distance from type-Ia supernova data. The procedure is applied to the Pantheon+ sample, and on the basis of average MSE, AIC, and BIC scores in 10-fold CV the authors conclude that the (2,1) Padé approximation is optimal and that the scheme has a 'remarkable ability to distinguish' Padé orders. The paper also reports MCMC constraints on the (2,1) model parameters.","tokens_in":9898,"tokens_out":8047,"duration_ms":90510,"significance":"If the statistical procedure were fully specified and validated, the paper would offer a practical, data-driven alternative to the heuristic rules previously used to choose Padé orders in cosmography, and its conclusion that (2,1) is optimal agrees with earlier work by Capozziello et al. (2020). The manuscript is honest about limitations such as dependence on the Pantheon+ dataset and on the data-splitting ratio, and it provides explicit Padé expressions. However, the paper's own evidence does not yet support the strength of the claimed conclusion: the likelihood underlying AIC/BIC is undefined, the CV scores have no uncertainties, and several numerical and consistency issues remain.","major_comments":[{"comment":"The AIC and BIC values cannot be reproduced because the maximum likelihood L_max is never defined. It is not stated whether L_max is constructed from the χ2 of Eq. (1) with the full Pantheon+ covariance matrix or from a Gaussian likelihood based on the MSE of Eq. (10); the two choices lead to different AIC/BIC values. Furthermore, the number of free parameters m is ambiguous: Eq. (6) contains j0 as a coefficient, the 'where' block after Eq. (9) fixes j0 = 1, and Table 1 reports a fitted j0 = 1.5^{+1.1}_{-1.2}. Since AIC and BIC depend linearly on m, this ambiguity can change the penalty and could flip the ranking of Padé orders. The authors should state L_max explicitly and give m for each candidate order under a consistent treatment of j0.","section":"Section 3.2, Eqs. (11)-(12)"},{"comment":"The conclusion that the (2,1) Padé approximation is optimal is not statistically supported by the reported scores. The authors state that 'the difference of average of BIC and AIC is not significant,' yet they select (2,1) by looking at insets that amplify those small differences. No uncertainty estimates (e.g., standard errors over the ten folds or over repeated random splits) are reported for the average MSE, AIC, or BIC, so the ranking may be an artifact of the random fold partition. Please report per-fold values, standard errors, and a paired significance test for the AIC/BIC differences, and specify the number of repeats and random seed used in Steps 1-3.","section":"Section 4"},{"comment":"The sample size used in the analysis is inconsistent. The data section states that Pantheon+ comprises '1701 light curves of 1550 unique' supernovae, while Eq. (1) and Step 5 refer to '1071 SNe Ia data' from Koussour et al. (2024). Since n enters the BIC in Eq. (12) and determines the construction of the CV folds, the paper must clarify which of these samples is used and how the number 1071 is obtained.","section":"Sections 2 and 4, Step 5"},{"comment":"The Padé expressions for the luminosity distance are written as 1/H0 times a dimensionless combination of z and the cosmographic coefficients, with no factor of c. In standard units, the leading factor for d_L should be c/H0, or an explicit convention c=1 must be stated. As written, substituting H0 in km/s/Mpc gives a quantity whose units are not Mpc, so Eq. (2) would not produce the correct distance modulus. The authors should introduce c explicitly or declare the unit convention.","section":"Equations (6)-(9) and Table 1"},{"comment":"The pre-selection of candidate Padé polynomials using the threshold 'average of MSE less than 0.05' is introduced without any statistical justification. This ad hoc cutoff can remove a model with slightly higher MSE but substantially better AIC/BIC, and it is applied to an average that carries no reported uncertainty. Please justify the threshold or omit the pre-selection step and compare all orders.","section":"Section 4, Step 2"},{"comment":"The manuscript as provided does not contain the actual figures that hold the numerical evidence: Figures 2-8 appear as placeholders such as 'Figure ??' in the text, and Figures 1 and 2 share the same caption. Without these figures and the underlying numerical values, the reported CV scores and MCMC contours cannot be inspected. The figures and the exact numbers must be supplied for the central claim to be verifiable.","section":"Figures 2-8"}],"minor_comments":[{"comment":"The reference to Akaike et al. (1973) in the reference list is incomplete; it should include the full title and publisher information.","section":"References"},{"comment":"Figure 1 and Figure 2 have identical captions; please differentiate them to indicate which shows the preliminary candidates and which shows the final fits.","section":"Figures 1 and 2"},{"comment":"The MCMC setup (priors, number of walkers, chain length, burn-in) is not described, so the reported parameter uncertainties in Table 1 cannot be reproduced.","section":"Table 1"},{"comment":"The text has several typographical and grammatical errors, e.g., 'red-shifts' in the abstract and 'Sec.4 present a comparison' in Section 1; a careful language edit is needed.","section":"Throughout"},{"comment":"The notation '1 δ, and 2 δ' in Figure 8 should be replaced by the standard '68% and 95% confidence regions'.","section":"Figure 8"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for publication in its current form. The core statistical procedure—model selection via AIC/BIC in cross-validation—is not reproducible because the likelihood is undefined, and the order ranking may be within noise. The missing figures also prevent verification. The topic is within scope and the idea is potentially useful, but the statistical framework must be completed and the paper cleaned up before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper applies k-fold cross-validation to pick the order of Padé approximants for the luminosity distance, using Pantheon+ data, and lands on the (2,1) form. That conclusion is not new — Capozziello et al. (2020) argued the same from coefficient-degeneracy reasoning — but the CV recipe is, as far as I know, not in the earlier literature. The paper is honest about this and about the fact that the result is data-dependent.\n\nWhat is genuinely good: the authors try to replace analytic selection rules with a data-driven procedure, they check several k values, and they report MCMC constraints on the (2,1) parameters. The limitations they state (depends on Pantheon+, split ratio matters) show a fair sense of their own scope.\n\nThe soft spots are in the statistical core. The likelihood that enters AIC/BIC is never defined — you cannot reproduce those numbers. The effective number of free parameters is ambiguous: Sec. 3.1 sets j0 = 1, but Table 1 fits it and gets 1.5^{+1.1}_{-1.2}. That affects the penalty term and could flip a ranking. The CV scores are shown as point averages with no error bars or repeated-seed variation, yet the abstract claims \"remarkable ability to distinguish\" orders. The text itself says the AIC/BIC differences are \"not significant\" — and then the (2,1) selection is made by amplifying those small differences in insets. That is a load-bearing contradiction: the paper's entire contribution is an order-selection procedure, and the evidence that it works is a ranking that may be within noise. There are also placeholder figure references (Figure ??) and an ad hoc pre-filter (MSE < 0.05).\n\nNone of this makes the paper obviously wrong — the (2,1) result agrees with earlier analytic arguments — but it does mean the new contribution, the CV procedure, is not validated. A reader cannot tell whether the method would reliably select a sensible order on a different dataset, which is the stated future application.\n\nWho is this for? Someone working on Padé cosmography or model selection in cosmology would want to know it exists, but they should not treat the selection result as settled. It deserves a serious referee — the question is worth asking, and the flaws are fixable — but it needs major revision: define the likelihood, pin down the parameter count, report CV uncertainties, and temper the abstract.\n\nI'd send it out for peer review, expecting heavy revision.","headline":"A well-intentioned but statistically underspecified attempt to make Padé order selection data-driven; the conclusion is plausible, but the procedure's reliability is not demonstrated.","tokens_in":10296,"tokens_out":2167,"would_cite":false,"duration_ms":23799,"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":"The paper claims that a 10-fold cross-validation scheme selects the (2,1) Padé rational polynomial as the optimal reconstructor of luminosity distance from Pantheon+ supernova data.","keywords":["luminosity distance","Padé approximation","cross-validation","Pantheon+","model selection","Akaike Information Criterion","Bayesian Information Criterion","cosmography"],"falsifier":"Recompute the 10-fold cross-validation scores for the four candidate Padé orders with a fully specified Gaussian likelihood using the Pantheon+ covariance matrix, repeat over many random data splits, and check whether the AIC/BIC gap between (2,1) and (2,2) or (3,2) exceeds the spread of scores across splits; if it does not, the claim that (2,1) is optimal is not supported by the cross-validation evidence.","tokens_in":9210,"feed_emoji":"🔭","tokens_out":5475,"duration_ms":56748,"temperature":0.7,"pith_summary":"The paper proposes a general data-driven recipe for choosing the order of a Padé rational polynomial used to reconstruct luminosity distance: partition the observed sample into folds, fit each candidate order on training folds, score held-out predictions by MSE, AIC, and BIC, average over folds, and pick the order with the best average score. Applied to the 1,701 Pantheon+ supernovae, the recipe separates four candidate orders and selects the (2,1) Padé approximation as the best explanation of the data at both low and high redshift. The central claim is that (2,1) is the optimal order among those tested, and that this cross-validation scheme is a reliable way to make such a choice for any given cosmological dataset. If the paper is right, observers gain a purely data-based alternative to the usual heuristic that numerator order should be one higher than denominator order.","feed_headline":"10-fold CV picks (2,1) Padé as best cosmic distance fit","feed_subtitle":"A data-only scheme ranks rational approximations against Pantheon+ supernovae and lands on order (2,1).","key_machinery":"The machinery is the Padé rational approximation $P_{m,n}(z) = (a_0+a_1 z+\\dots+a_m z^m)/(1+b_1 z+\\dots+b_n z^n)$ for luminosity distance as a function of redshift, with coefficients matched to the Taylor expansion of the cosmographic distance series. Around that object the paper builds a 10-fold cross-validation model-selection pipeline: random splitting, training each candidate order on $k-1$ folds, scoring held-out predictions by MSE, AIC, and BIC, and averaging over folds. AIC and BIC penalize the number of free coefficients, which is what lets the procedure trade fit quality against polynomial complexity. The Pantheon+ supernova sample supplies the observed distance moduli that the candidate rational forms must predict.","core_discovery":"The central claim is that the (2,1) Padé rational polynomial is the optimal functional form for reconstructing luminosity distance from Pantheon+ data, and that this can be established by cross-validation rather than by theoretical priors. Using 2-, 5-, 10-, 20-, and 50-fold validation, the authors first screen Padé orders by mean squared error below 0.05, retain orders (2,1), (2,2), (3,1), and (3,2), and then compare fold-averaged MSE, AIC, and BIC. In 10-fold cross-validation the (2,1) order scores best on all three criteria, and MCMC fitting of its free parameters gives $q_0 = -0.52^{+0.15}_{-0.12}$, $j_0 = 1.5^{+1.1}_{-1.2}$, and $H_0 = 71.84^{+0.25}_{-0.19}$ km/s/Mpc. The paper concludes that the (2,1) Padé approximation explains Pantheon+ data at low and high redshifts and agrees with earlier theoretical arguments favoring numerator order one higher than denominator order.","pith_inferences":["A natural next test is whether (2,1) still wins under repeated 10-fold cross-validation with many random seeds; the paper reports no error bars on the AIC/BIC differences, so the stability of the ranking is an open question.","Applying the same data-driven scheme to independent distance catalogs would reveal whether (2,1) is a feature of the Pantheon+ sample specifically or of the underlying distance-redshift relation more generally.","If the AIC/BIC gaps are within noise, the pipeline could be extended to score candidate orders by predictive calibration, such as the coverage of validation residuals, rather than by point estimates of information criteria."],"forward_implications":["If the conclusion is correct, the (2,1) Padé form should be adopted as the default rational cosmographic reconstruction for Pantheon+ and similar supernova distance data, replacing ad hoc order choices.","The cross-validation procedure gives a template for order selection on other cosmological observables, such as Hubble parameter or BAO data, without assuming a specific cosmological model.","The fitted parameter values from the (2,1) fit provide a model-independent reference point against which parameter estimates from full cosmological model fits can be checked.","The result reinforces the rule that numerator order one above denominator order is preferred, because that pattern now emerges from data alone rather than from theoretical convenience."],"supporting_citations":[{"why":"Supplies the Pantheon+ dataset and its covariance matrix, the observational data that all candidate Padé orders are fitted against.","marker":"(Scolnic et al. 2022)"},{"why":"Provides the explicit Padé approximation formulas for the luminosity distance that the paper reconstructs, and the prior theoretical argument that (2,1) is statistically optimal.","marker":"(Capozziello et al. 2020)"},{"why":"Justifies the use of 10-fold cross-validation as a widely accepted model-selection procedure.","marker":"(Kohavi 1995)"},{"why":"Defines the Akaike Information Criterion used as one of the three order-selection scores.","marker":"(Akaike et al. 1973)"},{"why":"Defines the Bayesian Information Criterion used as the third order-selection score.","marker":"(Schwarz 1978)"},{"why":"Provides the MCMC sampler used to estimate the parameter uncertainties of the selected (2,1) Padé fit.","marker":"(Foreman-Mackey et al. 2013)"},{"why":"Supplies the cross-validation model-selection framework that the paper adapts to Padé order choice.","marker":"(Zhang & Yang 2015)"}],"fun_headline_variants":["10-fold CV ranks Padé (2,1) best for luminosity distance","Data-driven Padé choice: (2,1) wins on Pantheon+","Cross-validation selects Padé (2,1) for cosmic distances","Optimal Padé order (2,1) from 10-fold CV on Pantheon+","Padé (2,1) emerges as top fit via 10-fold CV"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ranking that selects (2,1) depends on AIC/BIC differences that the paper itself calls not significant, and the likelihood function and the effective number of free parameters used in those criteria are never specified.","fun_headline_variants_meta":{"raw":{"variants":["10-fold CV ranks Padé (2,1) best for luminosity distance","Data-driven Padé choice: (2,1) wins on Pantheon+","Cross-validation selects Padé (2,1) for cosmic distances","Optimal Padé order (2,1) from 10-fold CV on Pantheon+","Padé (2,1) emerges as top fit via 10-fold CV"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3381,"prompt_tokens":960,"completion_tokens":2421,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":2316}},"tokens_in":576,"tokens_out":2421,"duration_ms":18887,"temperature":1.0,"reasoning_tokens":2316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:11:24.771295+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the 10-fold cross-validation scores for the four candidate Padé orders with a fully specified Gaussian likelihood using the Pantheon+ covariance matrix, repeat over many random data splits, and check whether the AIC/BIC gap between (2,1) and (2,2) or (3,2) exceeds the spread of scores across splits; if it does not, the claim that (2,1) is optimal is not supported by the cross-validation evidence.","supporting_citations":[{"cited_title":"1995, in International joint conference on Artificial intelligence","cited_arxiv_id":null,"evidence_quote":"Justifies the use of 10-fold cross-validation as a widely accepted model-selection procedure."},{"cited_title":"N., & Czaki, F","cited_arxiv_id":null,"evidence_quote":"Defines the Akaike Information Criterion used as one of the three order-selection scores."},{"cited_title":"2015, Journal of Econometrics, 187, 95, https://doi.org/10.1016/j.jeconom.2015.02.006","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-validation model-selection framework that the paper adapts to Padé order choice."}],"review_version":1}