{"id":"1c3ae816-83ac-49c5-8f52-5183aaa35ba0","arxiv_id":"2411.13131","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A data augmentation Gibbs sampler estimates normal distribution parameters from sample mean, minimum, maximum, and size, but the core truncation-correction step is an unsolved implicit equation.","lead":"This paper proposes a Bayesian Gibbs sampler to estimate the mean and variance of a normal distribution when only the sample size, mean, minimum, and maximum are available, treating the missing intermediate values as augmented data. The method could help analysts who work with government reports that publish only these summary statistics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Data augmentation step fails to condition on the observed sample mean; the Gibbs sampler targets a different posterior, not the stated Bayesian model.","rationale":"The reader's weakest_assumption correctly flags the implicit equation for mu_trunc in Section 2.1, but the more fundamental issue is that the entire data augmentation step does not condition on the observed sample mean. Even if mu_trunc were solved exactly, the augmented data would not have sample mean equal to xbar; it would only have that mean in expectation. For a valid Gibbs sampler, the imputed values must be drawn from their conditional distribution given the observed sufficient statistics, which includes a sum constraint. The paper does not do this, so the sampler is not sampling from the target posterior. This is load-bearing because the central claim promises Bayesian estimation accuracy comparable to theoretical expectations; if the sampler does not target the posterior, the promise is not met. The simulation comparison to Eq. (1) further weakens the claim, since Eq. (1) is an approximate likelihood rather than a true benchmark. The n=3 test is a sharp, decisive check: it eliminates all modeling ambiguity and shows whether the sampler preserves the observed mean. I find no ad hominem issues; the concern is purely technical. The verdict should remain REJECT, agreeing with the reader, but for a broader reason than the one highlighted.","tokens_in":4395,"tokens_out":6001,"duration_ms":64264,"concrete_test":"Run a minimal example with n=3. Set x_(1)=0, x_(3)=10, and xbar=5; then the only missing value is forced to be x_(2)=5, so the exact posterior of (mu, sigma^2) given the full data (0,5,10) can be computed in closed form. Run the proposed Gibbs sampler with any numerically solved mu_trunc and compare the resulting posterior to this exact posterior. If they differ substantially (e.g., the sampler's sigma^2 posterior has larger variance or its mu posterior has the wrong center), the data augmentation scheme is not conditioning on xbar and the sampler is not Bayesian for the stated summaries.","verdict_should_be":"REJECT","load_bearing_attack":"The core assumption in Section 2.1 is that the observed sample mean can be imposed by centering the truncated normal and solving for mu_trunc so that E[Z]=0, then adding back xbar_adj. This only fixes the expected value of each augmented draw, not the actual sample mean of the augmented dataset. The observed summaries include xbar as a sufficient statistic, so the correct posterior conditional for the missing intermediate values is the joint density restricted to the hyperplane (x_(1)+...+x_(n))/n = xbar. Independent draws from TN(mu_trunc, sigma^2, a*, b*) + xbar_adj do not satisfy this sum constraint, so Step 2 does not sample from the true conditional distribution of the missing data given the observed summaries. Consequently, Steps 3 and 4 are not the full conditionals of any joint posterior for (mu, sigma^2) given (xbar, x_(1), x_(n), n), and the sampler's stationary distribution is not the target posterior. The implicit equation for mu_trunc is a further gap: the displayed formula still contains mu inside phi and Phi, and no existence, uniqueness, or solution procedure is given. But even a correctly solved mu_trunc only produces zero mean in expectation, not an observed sample mean equal to xbar. The validation against Eq. (1) is not reassuring because Eq. (1) is itself an approximate likelihood; agreement between two heuristics does not establish Bayesian correctness. The central claim, as stated, therefore lacks support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian data-augmentation method, implemented as a Gibbs sampler, for estimating the mean and variance of a normal distribution when only the sample mean, minimum, maximum, and sample size are available. Intermediate order statistics are treated as missing and sampled from a truncated normal distribution whose location parameter is adjusted in an attempt to preserve the observed sample mean; then μ and σ² are updated from conjugate conditional posteriors. Simulation studies compare the sampler's RMSE with an RStan implementation of an approximate likelihood based on the minimum, maximum, and adjusted mean.","tokens_in":4688,"tokens_out":4936,"duration_ms":49493,"significance":"The problem is practically important: many survey and government reports provide only these summary statistics, and a valid Bayesian procedure would be useful for meta-analysis and secondary data analysis. The manuscript is clearly written and the simulation setup is transparent. However, the central claim—that the proposed Gibbs sampler yields estimates comparable to theoretical expectations—is not supported because the data-augmentation step does not condition on the observed sample mean in the required probabilistic sense. The sampler does not appear to target a well-defined posterior for (μ, σ²) given the observed summaries, and the validation against an approximate likelihood does not remedy this. If the augmentation were corrected, the approach could be valuable, but as presented the core contribution is not established.","major_comments":[{"comment":"The displayed formula for μ_trunc is an implicit equation, not a closed-form solution. If Z ~ TN(μ_t, σ², a*, b*), setting E[Z]=0 gives μ_t = -σ[φ((a*-μ_t)/σ) - φ((b*-μ_t)/σ)]/[Φ((b*-μ_t)/σ) - Φ((a*-μ_t)/σ)], so the unknown mean appears on both sides through φ and Φ. The paper writes the right-hand side using φ(a*|μ,σ²) and Φ(·|μ,σ²), where μ is the model mean rather than μ_t, which is at best ambiguous and at worst incorrect. The manuscript does not prove existence or uniqueness of a solution, nor does it provide a numerical procedure for solving the equation. Since Step 2 depends on this quantity, the sampler is not implementable as stated.","section":"Section 2.1, definition of μ_trunc"},{"comment":"Even if μ_trunc were computed correctly, independent draws from TN(μ_trunc, σ², a*, b*) followed by adding xbar_adj do not enforce the observed sample mean xbar. The conditional distribution of the intermediate values given the sufficient statistics (xbar, x_(1), x_(n)) is the joint truncated-normal density restricted to the hyperplane (1/n)Σx_i = xbar, i.e., Σ_{j=2}^{n-1} x_(j) = n·xbar - x_(1) - x_(n). Independent draws have expected sample mean xbar but almost surely a different realized mean. Therefore the augmented dataset x* does not have sample mean xbar, and Steps 3 and 4 are not the full conditional distributions of any joint posterior for (μ, σ²) given the observed summaries. The sampler's stationary distribution is not the stated target posterior.","section":"Section 2.1, Step 2 (data augmentation)"},{"comment":"The conditional posterior used for μ, N((μ0/τ0² + n·xbar/σ²)/(1/τ0² + n/σ²), 1/(1/τ0² + n/σ²)), is the posterior conditional on the sufficient statistic xbar as if it were the mean of n iid observations from N(μ,σ²). But in a valid data-augmentation Gibbs sampler, the complete-data sufficient statistics must be computed from the augmented data x*, and the mean of x* is not xbar. Thus the algorithm conditions on the observed sample mean twice—once in the construction of μ_trunc and once in Step 3—without a well-defined joint distribution that has both as conditional expectations. This internal inconsistency reinforces that the sampler does not target the posterior claimed in the abstract.","section":"Section 2.1, Step 3"},{"comment":"The comparison with RStan based on Eq. (1) does not validate the Bayesian correctness of the proposed method. Equation (1) is introduced as an approximation: the text states that using a truncated normal for xbar_adj 'will be more practical' because the exact variance is hard to compute due to correlations among xbar_adj, X_(1), and X_(n). Agreement between the proposed sampler and an approximate likelihood shows only that two heuristics behave similarly. The paper does not report coverage properties, nor compare with the full-data posterior or an exact likelihood, so the claim of 'estimation accuracy comparable to theoretical expectations' is not supported by the evidence presented.","section":"Section 3, validation against Eq. (1)"}],"minor_comments":[{"comment":"The number of intermediate values is inconsistent: the text writes 'z*_(j), j = 1, . . . , n−1' after defining z* for j = 2, . . . , n−1. Since x_(1) and x_(n) are observed, there are n−2 intermediate values, so the index set should be j = 2, . . . , n−1 consistently.","section":"Section 2.1, Step 2"},{"comment":"The notation φ(a*|μ,σ²) and Φ(·|μ,σ²) is nonstandard for the truncated-normal expectation formula; the usual expression uses standardized bounds (a*−μ)/σ and (b*−μ)/σ. Clarifying this would also expose the implicit nature of the equation for μ_trunc.","section":"Section 2.1, truncated normal formula"},{"comment":"The prior hyperparameters (μ0, τ², α0, β0) are described only as 'detailed settings are shown in the figures below'; they should be listed explicitly in the text or in a table for reproducibility.","section":"Section 3"},{"comment":"The phrase 'comparable to theoretical expectations' overstates the evidence, because the 'theoretical expectations' are themselves based on the approximate likelihood (1), not on an exact posterior or known theoretical accuracy.","section":"Abstract and Conclusion"},{"comment":"The caption refers to 'theoretical distribution using RStan,' which is misleading; the bottom panel is an approximation based on Eq. (1). The caption should state that explicitly.","section":"Figure 2 caption"}],"recommendation":"reject","confidential_remarks":"The reader's report identifies a fundamental flaw in the data-augmentation step: the sampler does not condition on the observed sample mean in the sense required for a valid posterior. I concur that this is load-bearing and cannot be fixed by local edits. The authors would need to redesign the augmentation, for example by sampling from the joint truncated-normal distribution restricted to the sum constraint, and then re-validate the sampler. There are no citation or novelty concerns beyond the technical issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Tomoki, quick take on Matsumoto 2411.13131.\n\nThe new idea is real: mean/min/max/n summaries appear in government surveys, and the existing tools are direct formulas or max-only estimators. Treating the middle order statistics as missing data and imputing them from a truncated normal inside a Gibbs sampler is a sensible Bayesian setup that hasn't been done in the cited prior work. The paper is written clearly, and the conjugate-prior steps are each simple.\n\nThe problem is that the central data augmentation step doesn't do what it claims. The paper wants the augmented dataset's mean to equal the observed xbar. What it actually does is pick mu_trunc so that each truncated-normal draw has expected value zero, then adds xbar_adj. That fixes each draw's expected value, not the realized mean of the n-2 imputed observations. Independent draws fluctuate around zero, so the augmented sample mean will not equal xbar; the observed mean is never conditioned on. Steps 3 and 4 are therefore not the full conditionals of the stated posterior, and the sampler's stationary distribution is not the target. The implicit equation for mu_trunc is a secondary issue -- mu appears inside phi and Phi -- but the expectation-only centering would be fatal even with a perfectly solved mu_trunc.\n\nThe validation does not compensate. Equation (1) is an approximate likelihood the authors constructed themselves, so agreement between their Gibbs sampler and a Stan fit of (1) shows two heuristics roughly agree, not that either matches the true posterior. The simulation reporting is also thin: priors are not fully specified in the text, there are no error bars on the RMSE, and no code or data are provided.\n\nThere is a salvageable kernel. The problem is well-defined, and a correct version would impute the intermediate values jointly conditional on their sum constraint (a constrained truncated normal, or HMC within Gibbs). That is more involved than the current paper. I would not accept this version; the central claim fails as written. But the idea deserves more than a desk reject -- a serious referee should pin down the exact flaw and perhaps point toward the solution.\n\nMy call: reject with an invitation to revise if the augmentation is fixed and the benchmark is a true likelihood or the full-data posterior. A good reading-group example of a plausible but incorrect data augmentation.","headline":"The idea is squarely on target, but the sampler never actually conditions on the observed sample mean, so the central claim fails as written.","tokens_in":5185,"tokens_out":3401,"would_cite":false,"duration_ms":33919,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62G30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a data-augmented Gibbs sampler estimates the mean and variance of a normal distribution from the sample mean, minimum, maximum, and sample size, with accuracy comparable to the theoretical order-statistic likelihood.","keywords":["Bayesian estimation","Gibbs sampler","Data augmentation","Summary statistics","Extreme values","Truncated normal distribution","Normal distribution","Posterior inference"],"falsifier":"Run a simulation that, for a fixed $\\sigma$ and observed boundaries $a^*, b^*$, solves for $\\mu_{\\mathrm{trunc}}$ and checks whether the mean of the truncated normal actually equals zero; if no solution exists or the sampled intermediate values fail to average to the observed sample mean across many Gibbs iterations, the data augmentation step is invalid.","tokens_in":1458,"feed_emoji":"📊","tokens_out":7063,"duration_ms":105014,"temperature":0.7,"pith_summary":"Government and survey reports often publish only the sample mean, minimum, maximum, and sample size, leaving raw data unavailable. This paper proposes a Bayesian Gibbs sampler that treats all intermediate values as missing data, samples them from a truncated normal distribution, and then updates the normal parameters using conjugate priors. The central claim, supported by simulation studies, is that the resulting posterior estimates of the mean and variance have accuracy comparable to sampling from the theoretical likelihood based on the observed extreme values. If correct, practitioners with only these summary statistics can obtain reliable Bayesian interval estimates without access to the full dataset.","feed_headline":"Gibbs sampler recovers normal parameters from just mean, min, and max","feed_subtitle":"When raw data are unavailable, a data-augmented Gibbs sampler estimates the mean and variance as accurately as theory allows.","key_machinery":"The central object is the zero-mean truncated normal sampler used in the data augmentation step. A truncated normal distribution restricts samples to an interval $[a,b]$, and the paper defines the mean parameter $\\mu_{\\mathrm{trunc}}$ so that $E[Z]=0$ for $Z \\sim TN(\\mu_{\\mathrm{trunc}}, \\sigma^2, a^*, b^*)$, namely the negative of the bias term involving the normal density and cumulative distribution functions evaluated at the adjusted boundaries $a^* = x_{(1)} - \\bar{x}_{\\mathrm{adj}}$ and $b^* = x_{(n)} - \\bar{x}_{\\mathrm{adj}}$. This correction forces the sampled intermediate values to have mean equal to the observed sample mean after adding back $\\bar{x}_{\\mathrm{adj}}$, preserving the summary statistic while enabling standard conjugate Gibbs updates for $\\mu$ and $\\sigma^2$. The paper also justifies treating intermediate order statistics as independent conditional on the observed extremes through a bijective ordering function.","core_discovery":"The paper's central discovery is a data augmentation identity: by centering the observed boundaries with the adjusted mean of the intermediate values, defining the truncated normal's mean parameter as the bias correction $\\mu_{\\mathrm{trunc}}$ that makes the expected augmented value zero, and adding the adjusted mean back after sampling, the augmented dataset has expectation equal to the observed sample mean while every sampled value lies between the observed minimum and maximum. This makes the complete-data conjugate update valid, so Gibbs sampling over the augmented complete data yields posterior distributions for $\\mu$ and $\\sigma^2$ from only $\\bar{x}$, $x_{(1)}$, $x_{(n)}$, and $n$. Simulation-based comparison with the theoretical order-statistic likelihood shows that the two approaches produce closely aligned RMSE values across sample sizes.","pith_inferences":["The paper does not prove that the implicit equation defining $\\mu_{\\mathrm{trunc}}$ has a solution; a follow-up could analyze existence and uniqueness for fixed boundaries and variance, and propose a fixed-point iteration or closed-form approximation.","The independence of intermediate order statistics conditional on the extremes is invoked via an ordering bijection; a testable extension is to compare the sampler's posterior against the exact joint distribution of order statistics for small $n$.","The same bias-correction construction could be adapted to other location-scale families, such as the log-normal, by matching the conditional mean of the truncated base distribution to the observed mean, though the paper does not explore this direction.","A practical test of the method's calibration would be to check whether posterior credible intervals achieve their nominal coverage on datasets with known parameters; the reported RMSE comparison does not assess interval calibration directly."],"forward_implications":["Practitioners with only summary statistics can use direct conjugate Gibbs updates, avoiding more expensive sampling schemes such as Metropolis-Hastings or Hamiltonian Monte Carlo that are needed for the exact likelihood.","Bayesian credible intervals for the mean and variance can be produced from survey reports that publish only the sample mean, minimum, maximum, and sample size.","Estimation accuracy improves with sample size, and the simulations indicate that reasonable estimates are obtained even at moderate sample sizes when prior parameters are suitably specified.","The augmented-dataset construction makes the complete-data posterior updates valid, so the sampler produces a posterior for the original parameters rather than an ad hoc point estimate."],"supporting_citations":[{"why":"Supplies the method for simulating from a truncated normal distribution, which is the building block of the data augmentation step.","marker":"Robert (1995)"},{"why":"Provides the conjugate prior and Gibbs sampler framework that makes the conditional posterior updates for the mean and variance valid.","marker":"Gelman et al. (2013)"},{"why":"Gives the order-statistics background and explains why the exact variance of the adjusted mean is difficult to compute, motivating the truncated-normal approximation.","marker":"David and Nagaraja (2003)"},{"why":"Represents the earlier direct estimation formulas from summary statistics that the proposed Bayesian method builds on and compares against.","marker":"Hozo et al. (2005)"},{"why":"Represents the maximum-only estimation approach that the paper contrasts with the use of both minimum and maximum values.","marker":"Capaldi and Kolba (2019)"}],"fun_headline_variants":["Gibbs sampler estimates normal mean and variance from mean, min, max","Posterior recovery from mean, min, max via Gibbs sampling","Gibbs trick: infer normal laws from only sample mean, min, max","One-line posterior: Gibbs with truncated normal from extremes"],"cache_read_input_tokens":7296,"weakest_assumption_plain":"The method must solve an implicit equation for the mean of the truncated normal so that the augmented data average to the observed sample mean, and the paper does not show this equation has a unique solution or how to solve it.","fun_headline_variants_meta":{"raw":{"variants":["Gibbs sampler estimates normal mean and variance from mean, min, max","Posterior recovery from mean, min, max via Gibbs sampling","Gibbs trick: infer normal laws from only sample mean, min, max","One-line posterior: Gibbs with truncated normal from extremes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000463,"raw_usage":{"total_tokens":2232,"prompt_tokens":777,"completion_tokens":1455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":393,"completion_tokens_details":{"reasoning_tokens":1381}},"tokens_in":393,"tokens_out":1455,"duration_ms":10554,"temperature":1.0,"reasoning_tokens":1381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:47:18.599191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a simulation that, for a fixed $\\sigma$ and observed boundaries $a^*, b^*$, solves for $\\mu_{\\mathrm{trunc}}$ and checks whether the mean of the truncated normal actually equals zero; if no solution exists or the sampled intermediate values fail to average to the observed sample mean across many Gibbs iterations, the data augmentation step is invalid.","supporting_citations":[{"cited_title":"P., Djulbegovic, B., & Hozo, I","cited_arxiv_id":null,"evidence_quote":"Represents the earlier direct estimation formulas from summary statistics that the proposed Bayesian method builds on and compares against."},{"cited_title":"& Kolba, T","cited_arxiv_id":null,"evidence_quote":"Represents the maximum-only estimation approach that the paper contrasts with the use of both minimum and maximum values."}],"review_version":1}