{"id":"a7c66ba7-81b4-48fb-8354-4a312283f53f","arxiv_id":"2509.07950","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Bayesian score peaking at a 10-parameter three-path model suggests liquid gallium EXAFS supports one fewer usable fit parameter than the standard Shannon-Nyquist limit of 11 allows.","lead":"Scientists propose replacing a decades-old rule for how many parameters an EXAFS measurement can support with a Bayesian model-selection score, and test it on liquid gallium data. The method could give X-ray absorption analysts a statistically grounded stop rule for fitting, and is claimed to transfer to other spectroscopies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BFI's 10-parameter limit for l-Ga is prior-driven, not data-driven: the authors changed ΔR after seeing results, and the ranking changes with prior range.","rationale":"The reader's CONDITIONAL verdict points to the prior ranges and Laplace approximation as the weakest link. My stress-test focuses on the prior ranges, which are the more tractable and internally documented issue: the paper explicitly changed ΔR after seeing results, and the ranking shifts with that choice. This directly threatens the central claim that BFI provides a data-defined information limit. A sensitivity analysis across physically reasonable priors would settle whether the 10-parameter maximum is robust. If robust, the CONDITIONAL verdict can become ACCEPT (or remain CONDITIONAL with minor caveats); if not, the claim should be reduced to 'BFI is a useful heuristic' rather than a superior measure of information content. I do not propose a different verdict because the empirical study is transparent and reproducible enough to warrant a conditional acceptance; the concern is a call for an additional sensitivity analysis, not a refutation.","tokens_in":12765,"tokens_out":8290,"duration_ms":95147,"concrete_test":"For the full l-Ga data set, recompute ln(BFI) for every model in Table I under a grid of prior ranges: ΔR ∈ {0.025, 0.05, 0.1, 0.25} Å, ΔN ∈ {1, 3.5, 7}, Δσ² ∈ {0.01, 0.02, 0.05} Ų, ΔE0 ∈ {5, 10, 20} eV, and C3 ∈ {0.05, 0.1, 0.2} ų, with all other settings as in the paper. Record the model and parameter count that maximizes ln(BFI) in each cell. If the maximum parameter count is 10 and a 3-path model wins in, say, >90% of the 3×3×3×3 = 81 combinations, the prior-sensitivity concern is largely resolved. If the winning model or parameter count varies with the prior ranges, the BFI-established 'information limit' is prior-dominated, and the headline claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that BFI (Eq. 5) measures the information content of the data is undercut by the BFI's explicit dependence on user-chosen prior ranges Δp_i. Because ∏Δp_i appears in the denominator, any change in a prior range rescales every model's BFI and can reorder the models. The paper itself reports that ΔR was initially set to ±0.25 Å but 'later found to be too large' and 'was then reduced' to ±0.025 Å 'to give a more meaningful comparison of models' (§III). This post-hoc adjustment is consequential: with the default larger ranges, the full-range maximum is at 10 parameters (model 3c), but the specific winning model changes with the range, and in the poor-background case the default ranges push the maximum to Nind=11 (model 4a) instead of 10 (§III.B). Thus the '10 parameters' result is not a stable property of the spectrum; it is an artifact of the selected (and adjusted) prior. Since no sensitivity analysis is reported and no uncertainties on ln(BFI) are given, the paper's assertion that BFI is a 'superior measure of the data information content' is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a new figure of merit, the Bayes Factor Integral (BFI, Eq. 5), as a measure of the information content of EXAFS spectra. The authors argue that the traditional Shannon–Nyquist criterion (Eq. 4) assumes independent fitting parameters, which is rarely true in EXAFS, and that BFI lifts this assumption by using the parameter covariance matrix and prior ranges. They test BFI on a liquid-gallium EXAFS spectrum by fitting a ladder of models with 1–4 scattering paths and cumulant terms, under three k-ranges and two background-subtraction conditions. They report that ln(BFI) peaks at 10 parameters, corresponding to a 3-path model, below the Shannon–Nyquist limit Nind = 11, and that the ranking changes when the data range or background quality is altered. The paper concludes that BFI provides a superior measure of EXAFS information content and can guide model selection.","tokens_in":12930,"tokens_out":3856,"duration_ms":46968,"significance":"If substantiated, the BFI would offer EXAFS practitioners a principled, correlation-aware criterion for deciding when to stop adding parameters and for comparing fitting strategies, going beyond the heuristic 'happiness' parameter currently used. The experimental test is well designed in several respects: real l-Ga data, a systematic sequence of 1-, 2-, 3-, and 4-path models, a cumulant-based alternative strategy, variation of the k-range, and a deliberately poor-background control. The authors also provide transparent fit descriptions and full result tables in the supplementary material. However, the central claim that BFI is a superior measure of data information content is not yet supported, because the BFI explicitly depends on user-chosen prior ranges and the reported conclusions change with those ranges.","major_comments":[{"comment":"The BFI formula contains ∏Δp_i in the denominator, so any change in a user-set prior range rescales every model's BFI and can reorder models with different numbers of parameters. The authors themselves changed the ΔR range from ±0.25 Å to ±0.025 Å after initial fits, stating the initial range was 'later found to be too large' and reduced 'to give a more meaningful comparison of models.' This is consequential: with default ranges and poor background, the maximum moves to Nind = 11 (model 4a), while with restricted ranges the 10-parameter model 3b is preferred (Fig. 5). No sensitivity analysis over ΔR, ΔN, Δσ², or ΔE0 is reported, and no uncertainties on ln(BFI) are given. The claimed '10-parameter limit' is therefore not established as a property of the data; it is a property of the selected priors. The authors should quantify how the ranking changes as the priors are varied within physic","section":"§III, Eq. (5)"},{"comment":"The BFI is computed via a Laplace approximation to the marginal likelihood using a covariance matrix from a single Larch fit (Eq. B14–B15). The paper does not validate that the posterior is adequately Gaussian, that the Larch covariance is a reliable estimate of the Hessian, or that the uniform prior is appropriate. This matters because the covariance enters the Occam factor directly. Furthermore, the paper states in §III.A that 'in Larch covariance is no longer calculated once Nvar > Nind (and hence BFI cannot be calculated).' Thus the method cannot even evaluate models above the Shannon–Nyquist limit, so the claim that BFI lifts the independence assumption and provides a new information-content limit is only partial. The authors should test the Laplace approximation against direct numerical marginal-likelihood integration for at least a few models, and discuss whether the Nvar ≤ Nind r","section":"§III.A, §III.B, Appendix B"},{"comment":"The claim that BFI is a 'superior measure of the data information content' is not backed by any comparison with existing information criteria. The manuscript presents ln(BFI) curves showing a peak below Nind, but does not compare with AIC, BIC, χ²ν, or the Shannon–Nyquist value on the same fits, nor with any synthetic-data benchmark where the true number of independent parameters is known. A model-selection criterion that favors a 3-path model is not the same as a measure of the information content of the spectrum. The authors should either benchmark BFI against these established FoMs or soften the claim to 'a model-selection criterion that correlates with information content' unless additional evidence is provided.","section":"Abstract; §IV"}],"minor_comments":[{"comment":"The definition of the Bayes factor is mistyped: 'ln BF = lnBFI1 / lnBFI2 = ln(BFI1−BFI2)' should read ln BF = ln(BFI1 / BFI2) = ln BFI1 − ln BFI2. The current expression is dimensionally and logically incorrect.","section":"§II"},{"comment":"The R-space range used to compute Nind = 11 is not specified in the text. To reproduce Eq. (4), the reader needs ΔR; please state the R range explicitly (e.g., the FT window used for the first-shell peak).","section":"§III.A"},{"comment":"The caption refers to 'the legend indicating choice of colour for the bars,' but no legend is visible in the figure as presented. Please add the legend or describe the color coding in the caption.","section":"Fig. 2"},{"comment":"Equation (B7) writes the prior as P({Aα}|D,I), which should presumably be P({Aα}|M,I) or the prior density for the parameters; the notation is confusing as written.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The empirical study has merit, and the poor-background control is a good idea. The main obstacle to publication is the unsupported central claim that BFI is a measure of data information content rather than a prior-dependent model-selection score. If the authors can provide a systematic sensitivity analysis of the prior ranges, validate the Laplace approximation, and compare against existing FoMs (or synthetic data with known ground truth), the paper could become publishable. I do not see grounds for rejection at this stage, but the current evidence is not sufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper reuses the authors' own Bayes Factor Integral (BFI, from their 2024 Nanoscale paper) as a parameter stop-rule for EXAFS, and on one liquid-gallium spectrum it peaks at 10 parameters when Shannon-Nyquist gives Nind=11. The specific result is plausible and the analysis is transparent. But the general claim that BFI is a superior measure of information content does not hold up, because BFI's value depends directly on user-chosen prior ranges, and the authors changed those ranges after seeing the results.\n\nWhat is genuinely new: the idea of using a Bayes factor with a covariance matrix to lift the independence assumption in the Shannon-Nyquist criterion is worth taking seriously. The paper is also honest about the mechanics: they list the models, the prior ranges, the noise estimate, and they test sensitivity to k-range and background quality. The finding that a 3-path model beats cumulant-based models for l-Ga is an interesting, physically plausible result. I give them credit for reporting the ΔR adjustment explicitly, even if it undermines the strength of the conclusion.\n\nThe soft spots are real. The stress-test note lands: BFI has ∏Δp_i in the denominator, so prior ranges directly rescale every model's BFI. The paper reports that ΔR was initially ±0.25 Å, \"later found to be too large,\" and reduced to ±0.025 Å. With the default ranges, the poor-background case pushes the maximum to Nind=11 (model 4a), not 10. So the 10-parameter limit is a property of the chosen priors, not purely of the data. Second, no uncertainties are reported on any ln(BFI) value; the model ranking is a set of point estimates, and we have no idea whether the 10-vs-11 (or 3b-vs-3c) differences are meaningful. Third, the demonstration is one spectrum with a single FT peak; that is a narrow base for the abstract's \"superior measure\" claim. Minor: the Bayes factor comparison formula in §II is miswritten as printed, and the 10-vs-11 comparison confounds parameter count with parameter type (path vs cumulant). None of these kill the specific l-Ga result, but they prevent the general claim from being accepted as stated.\n\nWho is this for? EXAFS practitioners who want a Bayesian alternative to Nind will find it useful as a starting point, but they should read it critically. It deserves serious peer review—the question is important and the empirical work is reproducible in principle. I would send it out with a request for a sensitivity analysis over prior ranges, uncertainties on BFI, and ideally a second test case. Not a desk reject.\n\nBest,\n[You]","headline":"Applies the authors' own prior-sensitive BFI to EXAFS information content; the l-Ga result is plausible, but the prior-range dependence undercuts the 'superior measure' claim.","tokens_in":13572,"tokens_out":2126,"would_cite":false,"duration_ms":26310,"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":"This paper claims that the Bayes Factor Integral, not the Shannon-Nyquist criterion, gives the true information content of EXAFS spectra, and that for liquid gallium the BFI peaks at 10 free parameters even though the conventional limit all","keywords":["EXAFS","Bayesian model comparison","information content","Shannon-Nyquist criterion","Bayes Factor Integral","parameter correlation","liquid gallium","figure of merit"],"falsifier":"Recompute the BFI for the same liquid-gallium spectrum using numerical (non-Laplace) integration of the marginal likelihood, or repeat the analysis with the prior ranges scaled by a factor of two in either direction; if the ln(BFI) maximum moves away from 10 parameters or the favored three-path model changes, the claimed information-content limit is an artifact of the prior and approximation choices rather than a fixed feature of the spectrum.","tokens_in":12512,"feed_emoji":"📊","tokens_out":3582,"duration_ms":43368,"temperature":0.7,"pith_summary":"The paper challenges the long-standing Shannon-Nyquist criterion used to set the maximum number of parameters that can be extracted from an EXAFS spectrum. It argues that this criterion assumes independent parameters, which is never true in EXAFS fitting, and introduces a Bayesian measure, the Bayes Factor Integral (BFI), that accounts for parameter correlations and prior information. Testing on a liquid gallium spectrum, the BFI reaches a maximum at 10 parameters while Shannon-Nyquist allows 11, and decisively rejects an 11-parameter model. The paper also shows that the BFI can distinguish between different fitting strategies, such as adding scattering paths versus adding cumulants, and that it is sensitive to data range and data quality. If correct, the BFI gives analysts a principled stopping rule for parameter addition and a tool for model selection.","feed_headline":"A Bayesian test says EXAFS spectra stop at 10 parameters, not 11","feed_subtitle":"Bayes Factor Integral overrides the Shannon-Nyquist limit when fitting parameters are correlated.","key_machinery":"The Bayes Factor Integral (BFI), defined as BFI = (2π)^(m/2) Lmax √det(Covp) / ∏Δp_i, is the central object. It is a Laplace-approximated marginal likelihood that combines the maximum likelihood of a fit with an Occam factor built from the parameter covariance matrix and the prior ranges. The covariance matrix injects parameter correlations into the criterion, while the prior ranges encode structural knowledge and penalize models whose parameters are allowed to roam too widely. Comparing ln(BFI) between models yields a Bayes factor with a standard evidence scale, giving a quantitative rule for when an added parameter is justified.","core_discovery":"The central claim is that the information content of an EXAFS spectrum is not fixed by the Shannon-Nyquist count Nind = 2ΔkΔR/π + 2, but should be determined by a Bayesian model-comparison quantity, the Bayes Factor Integral BFI = (2π)^(m/2) Lmax √det(Covp) / ∏Δp_i, where m is the number of fitted parameters, Lmax is the maximum likelihood, Covp is the parameter covariance matrix, and Δp_i are the user-chosen prior ranges on each parameter. Because Covp is not assumed diagonal, the BFI lifts the independence assumption and includes an Occam factor that penalizes models with broader priors or stronger correlations. Applied to the liquid gallium EXAFS spectrum, the BFI-determined parameter lim","pith_inferences":["The BFI's dependence on user-chosen prior ranges means the 'information content' it reports is conditional on prior knowledge; two labs using different reasonable priors could obtain different parameter limits from the same spectrum.","Since the paper shows the BFI ranking changes when the ΔR range is tightened, the method's power to distinguish models is partly a statement about how well the analyst knows the structure in advance, not solely about the data.","If applied to other spectroscopies, the BFI might reveal when an extra model component is merely absorbing systematic errors (such as imperfect background removal) rather than capturing real physical signal.","A natural next test is whether the BFI peak converges to a stable parameter count as prior ranges are varied continuously, or whether the peak itself moves—this would separate a property of the data from a property of the prior."],"forward_implications":["Analysts can stop adding parameters when ln(BFI) peaks, rather than trusting the Shannon-Nyquist count alone.","BFI can rank competing fitting strategies on the same data, such as whether to describe a peak with more scattering paths or with higher-order cumulants.","The criterion responds to data quality: with intentionally poor background subtraction, the BFI maximum shifts toward the Shannon-Nyquist limit and favors different models.","Reducing the k-range lowers the BFI-determined parameter limit, showing that the measure tracks the information actually available in the data.","Because BFI is computed from a single fit's covariance and likelihood, it lends itself to automated testing of many structural models."],"supporting_citations":[{"why":"Establishes the original Shannon-Nyquist-based estimate Nind≈2ΔkΔR/π that the paper seeks to replace.","marker":"[7]"},{"why":"Provides the exact Shannon-Nyquist expression Nind = 2ΔkΔR/π + 2 used as the baseline information limit throughout the paper.","marker":"[9]"},{"why":"Documents the known correlation among EXAFS fitting parameters, the problem the BFI is designed to address.","marker":"[10]"},{"why":"Introduces the Bayes Factor Integral as a figure of merit for EXAFS structural model comparison, which this paper extends to information content.","marker":"[14]"},{"why":"Supplies the analysis package used to perform all fits, obtain likelihoods, and compute the covariance matrices feeding the BFI.","marker":"[17]"},{"why":"Gives the Bayesian global-likelihood derivation that the BFI expression generalizes by allowing a non-diagonal covariance matrix.","marker":"[20]"},{"why":"Provides the Bayes-factor evidence scale used to decide when one model is decisively preferred over another.","marker":"[22]"}],"fun_headline_variants":["Bayesian overhaul: EXAFS info limit isn't Shannon-Nyquist","EXAFS data info: Bayesian criterion beats Shannon-Nyquist","Gallium EXAFS: Bayesian fit finds fewer parameters than thought","New Bayesian yardstick for EXAFS information content"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The result depends on the user-chosen uniform prior ranges (Δp_i) and on the Laplace approximation of the marginal likelihood computed from a single fit; if those ranges or the approximation are inappropriate, the BFI-determined parameter limit is not a property of the data itself.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian overhaul: EXAFS info limit isn't Shannon-Nyquist","EXAFS data info: Bayesian criterion beats Shannon-Nyquist","Gallium EXAFS: Bayesian fit finds fewer parameters than thought","New Bayesian yardstick for EXAFS information content"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000445,"raw_usage":{"total_tokens":2051,"prompt_tokens":672,"completion_tokens":1379,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":1314}},"tokens_in":416,"tokens_out":1379,"duration_ms":10828,"temperature":1.0,"reasoning_tokens":1314,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:27:36.743146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the BFI for the same liquid-gallium spectrum using numerical (non-Laplace) integration of the marginal likelihood, or repeat the analysis with the prior ranges scaled by a factor of two in either direction; if the ln(BFI) maximum moves away from 10 parameters or the favored three-path model changes, the claimed information-content limit is an artifact of the prior and approximation choices rather than a fixed feature of the spectrum.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the original Shannon-Nyquist-based estimate Nind≈2ΔkΔR/π that the paper seeks to replace."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the exact Shannon-Nyquist expression Nind = 2ΔkΔR/π + 2 used as the baseline information limit throughout the paper."},{"cited_title":"Newville, Fundamentals of xafs, Reviews in Mineralogy Geochemistry 78, 33 (2014)","cited_arxiv_id":null,"evidence_quote":"Documents the known correlation among EXAFS fitting parameters, the problem the BFI is designed to address."},{"cited_title":"Gianolio, How to start an xas experiment, in X-Ray Absorption and X-Ray Emission Spectroscopy (John Wiley Sons, Ltd, 2016) Chap","cited_arxiv_id":null,"evidence_quote":"Introduces the Bayes Factor Integral as a figure of merit for EXAFS structural model comparison, which this paper extends to information content."},{"cited_title":"Ravel, 7.2","cited_arxiv_id":null,"evidence_quote":"Gives the Bayesian global-likelihood derivation that the BFI expression generalizes by allowing a non-diagonal covariance matrix."},{"cited_title":"Easy computation of the Bayes Factor to fully quantify Occam's razor","cited_arxiv_id":"2007.09702","evidence_quote":"Provides the Bayes-factor evidence scale used to decide when one model is decisively preferred over another."}],"review_version":1}