{"id":"8a49986a-692b-4390-8c32-dd980e840547","arxiv_id":"2602.07165","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PoissonRatioUQ is an R package that turns Poisson count data into Bayesian posterior distributions for intensity ratios and derived physical quantities.","lead":"PoissonRatioUQ is a new R package for Bayesian uncertainty quantification when the data are counts and the quantity of interest is a ratio of Poisson means. It provides both spatial (permanental process) and pointwise estimation, including the case where the ratio is transformed by Z=(mT+z0)^p.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated Laplace/Gamma posterior: closed-form UQ may be miscalibrated in low-count regimes; no coverage check.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the Laplace/Gamma approximation may fail in low-count settings, making the claimed closed-form posterior unreliable. The paper's demonstration does not include a coverage or calibration check, so this concern is not addressed. The Eq. (2.7) typo is real but secondary; the gradient (2.8) suggests the implementation follows the corrected form. The central claim of uncertainty quantification stands or falls on whether the approximate intervals are calibrated, so a simulation-based coverage check is the decisive test. The verdict remains CONDITIONAL (UNCHANGED), because the concern is not a demonstrated failure, but the missing validation is significant enough to preclude unconditional acceptance.","tokens_in":10841,"tokens_out":5646,"duration_ms":51422,"concrete_test":"Run a simulation study with n=50 bins and low Poisson means (e.g., drawn uniformly from 0.5 to 5). For 500 simulated datasets, compute the 95% HPD intervals for Z from the package and record empirical coverage of the true Z. Also, on one representative dataset, sample the exact posterior of f using an MCMC method (e.g., Stan with a non-centered parameterization) and compare the resulting marginal intervals for Z to the package's Beta Prime intervals. If empirical coverage is below ~90% or MCMC intervals differ substantially from the closed-form intervals, the approximate closed-form posterior is not reliable in the low-count regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PoissonRatioUQ yields closed-form posterior distributions for the ratio Z and the quantity of interest T. This rests on two approximations: the Laplace approximation to the latent-field posterior (Eq. 2.9) and the moment-matched Gamma approximation for each Λ_i (Eq. 2.10). These are not exact closed forms, and the paper provides no evidence that the resulting HPD intervals are calibrated. In low-count regimes (Poisson means near or below 1, many zero bins), the log-posterior (2.4) contains terms a_i log(f_i^2) that are highly non-quadratic near f_i=0; the Laplace approximation can put substantial mass on negative f_i, and the Gamma approximation may poorly represent the skewness of c/2 f_i^2. The demonstration in Section 3.1 reports only CRPS and MAE of the MAP, not interval coverage or comparison to exact posterior sampling. A secondary issue is that Eq. (2.7) is internally inconsistent: with f = K̃ψ, the penalty −1/2 ⟨f, K̃^{-1}f⟩ becomes −1/2 ⟨ψ, K̃ψ⟩, not −1/2⟨ψ, K̃^{-1}ψ⟩; the gradient in Eq. (2.8) matches the corrected form, suggesting a typographical error in the paper rather than the code, but it obscures the derivation. The load-bearing concern is the lack of validation of the approximate posterior: if the intervals are miscalibrated, the package's UQ output is not trustworthy for the low-count applications that motivate it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces an R package, PoissonRatioUQ, for Bayesian estimation and uncertainty quantification of ratios of Poisson intensities from binned count data. The modeling uses a permanental process with a latent Gaussian field and the quadratic link Lambda_i = c/2 f_i^2; after a Laplace approximation for the latent field and a moment-matched Gamma approximation for each Lambda_i, the intensity ratio is modeled as a generalized Beta Prime distribution, and quantities of interest T defined by Z=(mT+z0)^p are obtained by a location-scale transformation. The paper describes the mathematical derivation, package functions for optimization, generalized Beta Prime calculations, CRPS, and HPD intervals, and gives two toy demonstrations. A timing experiment reports that a full posterior for 1000 bins takes roughly 15 seconds.","tokens_in":11263,"tokens_out":9115,"duration_ms":90050,"significance":"If the approximations are well calibrated, the package would fill a practical gap in spatial count-ratio analysis: it would provide fast, analytic UQ without MCMC. The paper's strengths are the explicit formulas for the generalized Beta Prime distribution, the analytic CDF/quantile/random-generation functions, the HPD and CRPS utilities, and the public package implementation. These are genuinely useful and clearly described. However, the central claim that the package yields closed-form posterior distributions is stated strongly, while the actual posterior is only approximate, and no calibration evidence is provided. The authors should be credited for making the code and examples available, but the UQ claim is not yet supported.","major_comments":[{"comment":"There is an internal inconsistency in the penalty term. With f = Ktilde psi, the quadratic form in Eq. (2.5) becomes -(1/2)<psi, Ktilde psi>, not -(1/2)<psi, Ktilde^{-1} psi>. The gradient in Eq. (2.8) matches the corrected form, so this is likely a typographical error, but it appears in the central derivation and should be fixed.","section":"Section 2, Eq. (2.7)"},{"comment":"The abstract and Section 2 describe the method as yielding 'closed-form posterior distributions,' but Eq. (2.9) is a Laplace approximation and Eq. (2.10) is a moment-matched Gamma approximation. These are not exact. The log-posterior (2.4) contains terms a_i log(f_i^2), which are strongly non-quadratic near zero, so in low-count or zero-bin regimes the HPD intervals may be miscalibrated. Section 3.1 reports only CRPS and relative MAE of the MAP, not interval coverage or comparison to an exact posterior sampler. I recommend a simulation study reporting empirical coverage of the 95% HPD intervals over repeated datasets with Poisson means near or below 1 and with zero bins, and, if feasible, a comparison to MCMC.","section":"Section 2, Eqs. (2.9)-(2.10); Section 3.1"},{"comment":"The covariance formula defines D = diag(psi_i^2 / (2 a_i)). For zero-count bins, a_i=0, this quantity is undefined. Since zero counts are common in the motivating low-count applications, the manuscript should state how zero counts are treated in the optimization and in constructing the Laplace covariance, e.g., whether D^{-1} is set to zero. Without this, the package's behavior for the key low-count regime is ambiguous.","section":"Section 2, Eq. (2.9)"}],"minor_comments":[{"comment":"Consider replacing 'closed-form posterior distributions' with 'analytic approximations to the posterior' or 'closed-form approximate posterior distributions' to avoid overstating the Laplace/Gamma approximations.","section":"Abstract and Section 2"},{"comment":"The notation in the denominator of alpha_i and beta_i appears as '2 mu_i2 + sigma_i^2'; this should be '2 mu_i^2 + sigma_i^2'.","section":"Eq. (2.10)"},{"comment":"The text contains a typo: 'Conitnuous Rank Probability Score' should be 'Continuous Rank Probability Score.' Also report the count levels used in the toy example (e.g., average counts per bin) so readers can judge whether the low-count regime is exercised.","section":"Section 3.1"},{"comment":"The legend repeats 'True ratio function / Observed count ratios / Estimated ratio function' three times. The legend should be cleaned up.","section":"Figure 3.1(a)"},{"comment":"The second author's name is misspelled: it should be 'Raftery,' not 'Rafferty.'","section":"Reference [34]"},{"comment":"There is a typo: 'involve' should be 'involve' in 'Two immediate applications for future work invole the ability...'","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a software/application note. The package is useful and the generalized Beta Prime utilities are competently implemented. The main gap is validation: the central UQ claim rests on two approximations that are not checked against exact inference or coverage. A focused simulation study, plus a correction to Eq. (2.7), would make the contribution acceptable. The self-citation to the companion paper [7] is appropriate and not a concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a package paper for a useful R tool, and the math is mostly right, but the central UQ claim is under-validated. The package does what it says, but the paper oversells the word 'closed-form' and never checks whether the uncertainty intervals are actually calibrated.\n\nWhat's new: PoissonRatioUQ bundles the permanental process estimator from prior work (Flaxman, Walder & Bishop, and the authors' own ref [7]) into an R workflow, adds a pointwise Gamma-Beta-prime route that handles unequal numbers of realizations, and provides transformation to a quantity of interest of the form Z=(mT+z0)^p. The derivations for the shifted generalized Beta Prime distribution are correct, and the CDF/quantile/random-number functions are sensible, using the Beta relationship. The timing is good - 1000 bins in about 15 seconds - and the code examples are reproducible.\n\nWhere it's soft, in proportion. First, Eq. (2.7) has a typo: with f = K̃ψ the quadratic penalty should be -1/2⟨ψ, K̃ψ⟩, not -1/2⟨ψ, K̃^{-1}ψ⟩; the gradient in (2.8) matches the corrected form, so it's a presentation error, not a substantive one. Second, the 'closed-form' language. Eq. (2.9) is a Laplace approximation and (2.10) is a moment-matched Gamma; the abstract and intro call them closed-form, while the body (Section 2) says 'well-approximated'. That is an overstatement. Third, and most important, there is no calibration or coverage check anywhere. The demonstration reports only CRPS and MAE of the MAP. In the low-count regime that motivates this work - Poisson means near or below one, many zero bins - the log-posterior is far from quadratic near f_i=0, and the Gamma approximation could be badly skewed. The authors should add a small simulation comparing their approximate intervals to exact sampling (MCMC or a dense grid) and report empirical coverage. This is not a dealbreaker; the package is likely fine for moderate counts, but UQ is the selling point, so it needs the check.\n\nWho it's for: applied statisticians and remote-sensing/x-ray astronomers who need quick Bayesian ratio intervals. It deserves a serious referee, but the revision should fix the typo and add the calibration evidence. I'd send it out.","headline":"A useful R package for ratio UQ that overstates its closed-form posteriors; the math is mostly right, but the UQ needs a calibration check.","tokens_in":11694,"tokens_out":4349,"would_cite":true,"duration_ms":45029,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Closed-form posterior for the ratio of Poisson means, delivered as an R package","keywords":["Poisson ratio","permanental process","Bayesian uncertainty quantification","generalized Beta Prime distribution","Gaussian process","count data","R package","remote sensing"],"falsifier":"Simulate binned Poisson count data with mean counts well below one per bin (generating many zero counts), then compare the package's closed-form 95% HPD intervals for the intensity ratio Z against a long-run MCMC posterior. If empirical coverage of the HPD intervals falls far below 95% (e.g., below 80%) across repeated simulations, the central approximation is falsified.","tokens_in":10756,"feed_emoji":"📊","tokens_out":2847,"duration_ms":30742,"temperature":0.7,"pith_summary":"This paper introduces PoissonRatioUQ, an R package for Bayesian uncertainty quantification of count ratios. The core idea is that the physically meaningful quantity is the ratio of Poisson means, not the ratio of noisy counts. By modeling the latent intensities with a permanental process—a squared Gaussian process driving Poisson counts—the package derives closed-form posterior distributions for intensity ratios and for quantities of interest linked to those ratios through a power-law forward model. The result is fast, analytic uncertainty quantification that avoids Markov chain Monte Carlo, making full posterior retrievals feasible for problems with thousands of spatial bins.","feed_headline":"Closed-form posteriors for Poisson count ratios in R","feed_subtitle":"A permanental-process model treats the ratio of Poisson means as the estimand, yielding analytic uncertainty without MCMC.","key_machinery":"The central object is the permanental process, where the Poisson intensity is λ(s) = (c/2) f(s)^2 with f drawn from a Gaussian process. The posterior of f is approximated as Gaussian via the Laplace method; squaring and scaling turns each bin intensity Λ_i into a Gamma distribution by moment matching. The ratio of two independent Gammas is then a generalized Beta Prime distribution, and the known algebraic relationship between Beta Prime and Beta distributions enables exact CDF, quantile, and random-number generation. This chain of exact analytic transformations is what carries the argument from count data to closed-form posteriors.","core_discovery":"The paper shows that under a permanental process model with a Laplace approximation to the latent Gaussian field and a moment-matched Gamma approximation for each bin intensity, the posterior distribution of the intensity ratio Z = Λ_a/Λ_b is a generalized Beta Prime distribution. When the forward model has the form Z = (mT + z0)^p, the quantity of interest T inherits a shifted generalized Beta Prime distribution. This yields analytic posteriors, highest posterior density sets, and scoring metrics without sampling, and the package implements the full pipeline: kernel-based spatial estimation, pointwise conjugate-Gamma estimation, ratio and T retrieval, and CRPS/HPD utilities.","pith_inferences":["Because the posterior is a standard parametric family, downstream tasks like hierarchical modeling, Bayesian updating, or importance sampling become straightforward—one can analytically marginalize or reweight without drawing samples from the raw count process.","The same Beta-Prime change-of-variable trick should extend to other monotone forward models beyond the power law (e.g., exponential or log-linear), giving closed-form T posteriors whenever the inverse transformation is tractable.","A practical diagnostic for the package would be a built-in coverage check: simulate data under known parameters, compare the claimed 95% HPD intervals to empirical coverage, and warn when the Laplace/Gamma approximation breaks down in low-count regimes.","The method is naturally portable to any photon- or particle-counting instrument with band ratios—X-ray hardness ratios, lidar ratios, isotope-ratio mass spectrometry—beyond the atmospheric remote sensing applications cited."],"forward_implications":["Users can obtain full posterior distributions, including HPD intervals, for intensity ratios and derived quantities in seconds—about 15 seconds for 1000 bins on a desktop—without MCMC.","Estimating the ratio of latent Poisson means rather than the ratio of observed counts reduces bias and properly propagates Poisson shot noise into uncertainty bounds.","The package supports both spatially correlated retrievals via the permanental process and pointwise retrievals via conjugate Gamma priors, and it can handle missing realizations using NaN placeholders.","For forward models of the form Z = (mT + z0)^p, the quantity of interest T has a closed-form shifted generalized Beta Prime posterior, enabling direct uncertainty quantification for nonlinear retrievals such as temperature from FUV radiance ratios.","Built-in CRPS and HPD functions provide model scoring and credible intervals for arbitrary parametric predictive distributions, including multimodal cases."],"fun_headline_variants":["Closed-form Bayesian posteriors for Poisson count ratios in R","No-MCMC uncertainty quantification for count ratios via R","PoissonRatioUQ: analytic ratio posterior without sampling","Gamma-Beta prime closed forms for ratio UQ in R"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the log-posterior of the latent Gaussian field is close to quadratic, so the Laplace approximation and the moment-matched Gamma distribution for bin intensities adequately describe the posterior; with very low counts, many zero bins, or a mis-specified kernel, the closed-form Beta Prime posterior is only approximate and the reported uncertainty intervals may be miscalibrated.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form Bayesian posteriors for Poisson count ratios in R","No-MCMC uncertainty quantification for count ratios via R","PoissonRatioUQ: analytic ratio posterior without sampling","Gamma-Beta prime closed forms for ratio UQ in R"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000405,"raw_usage":{"total_tokens":1869,"prompt_tokens":591,"completion_tokens":1278,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":335,"completion_tokens_details":{"reasoning_tokens":1211}},"tokens_in":335,"tokens_out":1278,"duration_ms":9691,"temperature":1.0,"reasoning_tokens":1211,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:39:29.637864+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate binned Poisson count data with mean counts well below one per bin (generating many zero counts), then compare the package's closed-form 95% HPD intervals for the intensity ratio Z against a long-run MCMC posterior. If empirical coverage of the HPD intervals falls far below 95% (e.g., below 80%) across repeated simulations, the central approximation is falsified.","supporting_citations":[],"review_version":2}