{"id":"a9706393-1556-4bc0-8fb8-88e6a259ce43","arxiv_id":"1908.04331","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Posterior uncertainty under a class of outer measures is asymptotically normal, yielding MAP estimators and likelihood-ratio-like tests whose limits are determined by the curvature of the possibility function.","lead":"Researchers extend statistical inference from probability measures to a more flexible class of outer measures built from supremum-based 'possibility functions', and prove analogues of the law of large numbers, central limit theorem, and Bernstein-von Mises theorem in this setting. The work offers a potential bridge between frequentist and Bayesian uncertainty quantification when the sampling model is misspecified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition B.2 (Slutsky lemma for uncertain variables) is false as stated, and Theorems 4.3 and 4.4 rely on it; the asymptotic normality and chi-squared results are therefore not established without a corrected lemma.","rationale":"The core of the paper's contribution after the BvM theorem is the claim that MAP estimators and credibility tests inherit asymptotic normality and chi-squared limits. Theorems 4.3 and 4.4 are derived by combining the CLT for possibility functions with Proposition B.2, the analogue of Slutsky's lemma. Proposition B.2 is not proved and, as stated, it is false: the counterexample with a remote spike in x_n and a contracting spike in z_n makes the product limit have an extra atom. This is not an artifact of discontinuous possibility functions; smoothing the spikes preserves the failure since the spike in x_n moves to infinity and is invisible in the pointwise limit but interacts with z_n's spike at 1/n to produce mass at s=1. Because the paper's proofs of Theorems 4.3 and 4.4 explicitly invoke Proposition B.2 with no additional hypotheses, the central frequentist-type guarantees are unsupported. The reader's CONDITIONAL verdict is appropriate: the claims may be repairable by proving a restricted Slutsky lemma under the smoothness/log-concavity and α>0 conditions that actually hold in the application, but as written the proof is incomplete. I agree with the reader's identification of this as the weakest assumption; the present stress-test makes the objection concrete by showing the lemma cannot be accepted on faith.","tokens_in":28215,"tokens_out":17117,"duration_ms":166567,"concrete_test":"Verify the counterexample analytically: for f_{x_n}(t)=max{e^{-t^2}, 1_{t=n}} and f_{z_n}(t)=1_{t=1/n}, compute f_{x_n z_n}(s) as above and check that its pointwise limit equals 1 on {0,1}, not 1_0; this disproves Proposition B.2 as stated. Then attempt to prove a restricted version of Proposition B.2 under the hypotheses actually satisfied by the sequences in Theorems 4.3 and 4.4 (α>0, strictly log-concave and smooth possibility functions); if no such proof exists, the proofs of those theorems are incomplete and must be amended.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition B.2 asserts that if x_n o.p.m.→ x and z_n c→ α, then x_n z_n o.p.m.→ αx (and sum/ratio analogues). The paper states this without proof ('beyond the scope of this work'). The statement is false as written. Take f_{x_n}(t) = max{e^{-t^2}, 1_{t=n}} and f_x(t)=e^{-t^2}; f_{x_n}(t) → f_x(t) pointwise, so x_n o.p.m.→ x. Take f_{z_n}(t)=1_{t=1/n}; then for any δ>0, sup_{|t|>δ} f_{z_n}(t)=0 once n>1/δ, so z_n c→0. The product has possibility f_{x_n z_n}(s) = sup_t f_{x_n}(t) f_{z_n}(s/t) = f_{x_n}(ns) = max{e^{-n^2s^2}, 1_{s=1}}, which converges pointwise to 1 at s=0 and s=1 and 0 elsewhere, whereas 0·x has possibility 1_{s=0}. Hence x_n z_n does not converge to 0·x, contradicting the product assertion. Theorems 4.3 and 4.4 use exactly this product/quotient form (with z_n = n/J*_n and x_n the normalized score). Even if the theorem's sequences are better behaved (α>0, log-concave possibility functions), the paper provides no such conditions or proof; a corrected, restricted Slutsky-type result is required.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops asymptotic theory for inference with 'possibility functions', i.e. supremal outer probability measures, as an alternative to probability measures. It introduces expected value and variance for uncertain variables via the argmax and the inverse second derivative of the possibility function, proves a law of large numbers and a central limit theorem for sums of independent uncertain variables, and then states a Bernstein-von Mises theorem for posterior possibility functions. From the BvM approximation it derives asymptotic normality of the MAP estimator and a chi-squared limit for the credibility statistic of a simple hypothesis test. The paper closes with a numerical illustration for the ratio of two means and positions the framework as a bridge between frequentist and Bayesian inference.","tokens_in":28586,"tokens_out":5663,"duration_ms":63650,"significance":"If the main results held as stated, the paper would provide a useful asymptotic toolbox for possibility-theoretic Bayesian inference and connect the proposed formalism to Fisher information, observed information, and likelihood-ratio-style testing. The paper is clear in its motivation, contains several worked examples, and gives a detailed proof of the LLN and a substantial proof of the CLT. It is also transparent in flagging Proposition B.2 as unproved. However, the two headline results of Section 4 — asymptotic normality of the MAP and the chi-squared limit of the credibility statistic — rest on an unproved and, as stated, false Slutsky lemma, and the BvM statement is only an informal '≈' with uncontrolled remainders. These issues are load-bearing and currently prevent the paper from delivering its central claims.","major_comments":[{"comment":"Proposition B.2 is stated without proof ('the proof ... is beyond the scope of this work') and is false as stated. A counterexample to the product assertion: let f_{x_n}(t)=max{e^{-t^2}, 1_{t=n}} and f_x(t)=e^{-t^2}; then x_n o.p.m.→x. Let f_{z_n}(t)=1_{t=1/n}; then z_n c→0. The product has possibility f_{x_n z_n}(s)=max{e^{-n^2 s^2}, 1_{s=1}}, which converges pointwise to a function equal to 1 at s=0 and s=1 and 0 elsewhere, not to 1_{s=0}. Hence x_n z_n does not converge to 0·x. Theorems 4.3 and 4.4 use exactly this product/quotient structure, as seen in the proofs around (B.5) and (B.6). A corrected, restricted version of the lemma with explicit conditions and a full proof is required before these theorems can be accepted.","section":"Appendix B, Proposition B.2"},{"comment":"Theorem 4.1 is stated only with the symbol '≈', and its proof is a Taylor expansion without control of the remainder terms or of uniformity in ψ. In particular, dropping the third-order term in the denominator of (B.2) requires more than Assumption A.1; one needs a bound on n^{-1}(θhat_n−θ0)∂^3_θ l_n(ψ_n), and if Assumption A.5 is invoked for that purpose it should be stated. Similarly, replacing the denominator by L_n(θhat_n) and neglecting the prior f_θ in (B.4) requires sup_{|ψ|≤M} |ψ|^3/n^{3/2} ∂^3 l = o(1) and a comparable prior contribution, none of which is proved. As written, the BvM result is an informal approximation, and Theorems 4.3 and 4.4 inherit this informality.","section":"Section 4.1, Theorem 4.1 and proof around (B.2)-(B.4)"},{"comment":"The convergence statements in Theorems 4.3 and 4.4 are in outer probability measure induced by the possibility function f_{y|θ}(·|θ0) for uncertain observations, not under the true sampling distribution p_Y of the random variables Y_i. The text motivates these results as 'frequentist-type guarantees', but the o.p.m. convergence of possibility functions does not by itself imply coverage properties or distributional convergence for the actual random data-generating mechanism. If the intended claim is the usual repeated-sampling behaviour, the relationship between the o.p.m. model and the true random law must be made explicit and proved; otherwise the terminology 'frequentist' is misleading.","section":"Section 4.2, Theorems 4.3 and 4.4"},{"comment":"The variance V*(x) is defined as −1/f_x''(E*(x)), and Theorem 3.3 then returns exactly this quantity as the limiting spread of the rescaled sum. The CLT is therefore close to a tautology: it shows that under strict log-concavity the limiting possibility function is determined by the local curvature of f_x at its mode. This does not invalidate the theorem, but the paper should acknowledge explicitly that the 'asymptotic variance' is a consequence of the definition of V*, not an independently derived quantity. The same observation applies to Theorem 4.3, where σ² = V*(s_θ0(y))/I*(θ0)² is a ratio of second derivatives by construction.","section":"Section 3.1, Eq. (3.2) and Theorem 3.3"}],"minor_comments":[{"comment":"The construction of the α-credible interval [a,b] assumes unimodality of the posterior possibility function, but this assumption is only mentioned in passing; it should be stated precisely in the theorem or proposition that uses it.","section":"Section 4.2.2"},{"comment":"The caption does not explain how the 'truth' line is computed, what the averaging is over, or how the standard deviation bands are constructed; these details should be added for reproducibility.","section":"Figure 1"},{"comment":"The claim that the bordered-Hessian check plus the stationarity condition identifies the global maximum is sketched rather than proved; the argument that any solution of (log f(x_i))' = (log f(x_j))' must have x_i = x_j under strict log-concavity deserves a short explicit proof.","section":"Proof of Theorem 3.3"},{"comment":"The decomposition (2.2) is stated as a factorisation of possibility functions, but the notation f_{y|t,θ} and f_{y|t} is introduced only informally; a precise definition of conditional possibility functions would help.","section":"Section 2, sufficient statistics"},{"comment":"There are minor typographical issues, including inconsistent use of 'Students model' for 'Student's model', and the phrase 'where the MLE is consistent' in Assumption A.1 is not a formal condition; it should be replaced by explicit regularity conditions.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"The central issue for the editor is that Theorems 4.3 and 4.4 depend on Proposition B.2, which is both unproved and false as stated. I do not see evidence of misconduct, but the mathematical gap is substantial and must be closed by the authors, not deferred. A corrected lemma with conditions and proof, together with a fully rigorous BvM statement, is a necessary condition for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The paper deserves to be taken seriously, but the main frequentist-type theorems do not currently stand. Proposition B.2, the Slutsky lemma for uncertain variables, is stated without proof and is false as written. The stress-test counterexample is correct: take f_{x_n}(t)=max{e^{-t^2},1_{t=n}} and f_{z_n}(t)=1_{t=1/n}. Then x_n converges in o.p.m. to x with f_x(t)=e^{-t^2}, and z_n converges in credibility to 0. But the product x_n z_n has possibility converging to 1 on {0,1} and 0 elsewhere, while 0·x has possibility 1 only at 0. So the product assertion fails. The paper explicitly says the proof is beyond its scope, then uses this lemma in the proofs of Theorems 4.3 and 4.4. Without a corrected and proved lemma, the asymptotic normality of the MAP and the chi-squared limit of the credibility statistic do not follow.\n\nWhat is new and worth keeping: the posterior Bernstein-von Mises theorem for possibility functions, the treatment of observed and Fisher information, and the connection between posterior credibility and likelihood-ratio-like tests. The LLN and CLT for uncertain variables are close to known results in max-plus and possibility theory, as the authors partly acknowledge, but the outer-measure packaging and the posterior results are not mere restatements. The numerical example is honest and modest.\n\nOther soft spots, in order of importance. Theorem 4.1 is proved by Taylor expansion with an explicit 'approx' and no control of remainder terms; as a theorem statement it needs more care. The convergence notion is pointwise convergence of possibility functions, not distributional convergence of estimators, and the frequentist-type guarantees are relative to the outer-measure model rather than to a true sampling distribution. That should be said up front. I would not call the variance definition circular in a damaging sense: defining V*(x) as inverse curvature and then recovering that same quantity from the CLT is by design, though it limits what the CLT claims about the world.\n\nThis paper is for people working in possibility theory, imprecise probability, and generalized Bayesian inference. It deserves a serious referee, but the referee's main job should be to force a real proof or a properly restricted version of Slutsky's lemma. As it stands, the central asymptotic results are conditional on a false statement, so I would send it back for major revision rather than desk-reject or accept.","headline":"The possibilistic Bernstein-von Mises theorem is a real contribution, but the asymptotic normality and chi-squared results rest on an unproven Slutsky lemma that is false as stated.","tokens_in":29081,"tokens_out":3392,"would_cite":false,"duration_ms":38574,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62F12","60F05","62F03"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a Bernstein-von Mises theorem for posterior possibility functions, yielding asymptotically normal MAP estimates and chi-squared credibility tests.","keywords":["outer probability measures","possibility functions","uncertain variables","Bernstein-von Mises theorem","asymptotic normality","generalised Bayesian inference","Fisher information","hypothesis testing"],"falsifier":"Take the paper's own ratio-of-means setting, simulate many datasets from the stated normal laws, and compare the empirical distribution of $\\sqrt{n}(\\theta^*_n-\\theta_0)$ with the predicted normal possibility function $N(0, V_*(s_{\\theta_0}(y))/I_*(\\theta_0)^2)$; if the mismatch persists as $n$ grows, Theorem 4.3 is wrong. More directly, exhibit two sequences of uncertain variables satisfying the convergence hypotheses of Proposition B.2 for which $x_n/z_n$ does not converge in outer probability to $x/\\alpha$; since the lemma is asserted without proof, such a counterexample would invalidate the proof chain of Theorems 4.3 and 4.4.","tokens_in":28020,"feed_emoji":"📈","tokens_out":17967,"duration_ms":155401,"temperature":0.7,"pith_summary":"The paper's goal is to show that the standard asymptotic toolkit of statistics still works when uncertainty is represented not by probability densities but by outer probability measures of the form $\\bar{P}(A)=\\sup_{\\theta\\in A} f(\\theta)$, with $f$ a 'possibility function'. Its central claim is a Bernstein-von Mises theorem: for large $n$, the posterior possibility function is approximately a normal possibility function centred near the true parameter with inverse observed information as its variance. From that single result it derives asymptotic normality of maximum a posteriori estimators and a chi-squared limit for credibility-based tests, all stated without access to the true sampling distribution. A sympathetic reader would care because the framework promises Bayesian-style inference that is less conservative, can express total ignorance without improper priors, and still carries familiar normal and chi-squared guarantees.","feed_headline":"Possibility-based posterior turns normal in large samples","feed_subtitle":"Even with no true sampling distribution, MAP estimates and credibility tests keep normal and chi-squared limits.","key_machinery":"The carrying object is the possibility function, a nonnegative function with supremum $1$, and the associated uncertain variable, a mapping from a sample space without a probability measure to the parameter space with a true state of nature. Possibility functions define outer measures by $\\bar{P}(A)=\\sup_{\\theta\\in A} f(\\theta)$, and Bayes' rule takes the form $f_\\theta(\\theta\\mid y_{1:n})=L_n(\\theta)f_\\theta(\\theta)/\\sup_{\\psi}L_n(\\psi)f_\\theta(\\psi)$. Expectation is the argmax, $E_*(x)=\\arg\\max f_x$, and variance is $V_*(x)=-1/f''_x(E_*(x))$, so the normal possibility function $N(\\mu,\\sigma^2)=\\exp(-(\\theta-\\mu)^2/(2\\sigma^2))$ replaces the normal density. Two limit theorems do the work: the law of large numbers for uncertain variables drives the sample mean's possibility function to the convex hull of the argmax, and the central limit theorem sends $n^{-1/2}\\sum_i(x_i-\\mu)$ to $N(0,1/(-f''_x(\\mu)))$. These give the score and observed-information convergences that feed the Bernstein-von Mises proof.","core_discovery":"On the paper's own terms, the discovery is that asymptotic statistics is not tied to the additivity of probability. Theorem 4.1 states that under Assumptions A.1 and A.2, for large $n$, $$f_\\$\\theta$(\\$\\theta$\\mid y_{1:n})\\approx N\\left(\\$\\theta$;\\theta_0+\\frac{\\Delta_n}{\\sqrt{n}},\\frac{1}{J^*_n}\\right),$$ where $\\Delta_n=\\sqrt{n}\\,\\partial_\\theta\\ell_n(\\theta_0)/J^*_n$ and $J^*_n=-\\partial^2_\\theta\\ell_n(\\hat\\theta_n)$ is the observed information, with $\\hat\\theta_n$ the maximum likelihood estimator. Consequently (Theorem 4.3) $\\sqrt{n}(\\theta^*_n-\\theta_0)$ converges in outer probability to $N(0,\\sigma^2)$ with $\\sigma^2=V_*(s_{\\theta_0}(y))/I_*(\\theta_0)^2$, where $s_{\\theta_0}(y)=\\partial_\\theta\\ell(\\theta_0;y)$ is the score and $I_*(\\theta_0)=E_*(-\\partial^2_\\theta\\ell(\\theta_0;y)\\mid\\theta)$ the Fisher-information analogue; and (Theorem 4.4) $-2\\log f_\\theta(\\theta_0\\mid y_{1:n})$ converges to the chi-squared possibility function $\\chi^2(0,V_*(s_{\\theta_0}(y))/I_*(\\theta_0))$. The posterior forgets the prior's shape but not the prior's nature: the likelihood can be a probability or a possibility function, yet the posterior remains a possibility function.","pith_inferences":["The unproved Slutsky-style lemma (Proposition B.2) is the main identifiable gap; a natural extension is to prove it under explicit regularity conditions, or to identify the extra assumptions needed for the quotient case used in normalising by observed information.","The variance formula $V_*(s_{\\theta_0}(y))/I_*(\\theta_0)^2$ looks like a mode-based analogue of a sandwich variance, so the theory may extend to misspecified or quasi-likelihood settings where the true distribution is unknown; the paper does not explore this connection.","Because the CLT for uncertain variables has a degenerate case ($V_*=\\infty$ when the second derivative vanishes), mapping which parametric families hit that boundary would delimit where the Bernstein-von Mises conclusion breaks down.","The ratio-of-means example suggests that non-integrable posterior possibility functions can still support credible-interval claims; a testable extension would check whether the predicted normal and chi-squared limits continue to hold for such heavy-tailed posteriors."],"forward_implications":["Bayesian updating with possibility functions inherits the standard asymptotic shape of frequentist inference: large-sample posterior credibility regions centred at the MAP agree with normal intervals based on the observed information.","A practitioner can report uncertainty about a parameter using only derivatives of the chosen log-likelihood, with no need to specify the true sampling distribution; the ratio $V_*(s_{\\theta_0}(y))/I_*(\\theta_0)^2$ plays the role of the asymptotic variance.","Simple hypothesis tests $H_0:\\theta=\\theta_0$ can be calibrated from the chi-squared possibility limit of $-2\\log f_\\theta(\\theta_0\\mid y_{1:n})$, avoiding bootstrap or full probabilistic modelling.","The prior's nature persists asymptotically: even after the prior's information is forgotten, a probability prior yields a probability posterior and a possibility prior yields a possibility posterior, so the choice of uncertainty representation must be made deliberately.","Because constant possibility functions represent complete ignorance and single points can have positive credibility, the framework keeps posterior-like objects proper in cases where conventional Bayesian posteriors would be improper, such as the paper's ratio-of-means example."],"supporting_citations":[{"why":"defines the uncertain variables and outer probability measures on which the paper's posterior and estimator formalism is built.","marker":"Houssineau [2018a]"},{"why":"supplies the possibility-theoretic conditioning rule that produces the posterior possibility function.","marker":"De Baets et al. [1999]"},{"why":"motivates the generalised Bayesian update with exponentiated losses that the paper interprets through possibility functions.","marker":"Bissiri et al. [2016]"},{"why":"provides the standard consistency conditions behind Assumption A.1 for a probabilistic likelihood.","marker":"Wald [1949]"},{"why":"provides the same consistency foundations for maximum likelihood invoked by Assumption A.1.","marker":"Le Cam [1990]"},{"why":"supplies the M-estimation perspective used to define the true parameter as the maximizer of the expected log-likelihood.","marker":"Godambe [1991]"}],"fun_headline_variants":["Possibility posteriors hit normal and chi-square limits","Asymptotic normality without probability additivity","Outer-probability Bayes keeps classical asymptotic laws","No sampling distribution? Posterior still normal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paper's unproved analogue of Slutsky's lemma for uncertain variables (Proposition B.2) is true under the conditions used, including when one sequence is divided by another; the paper states that proof of this lemma is beyond its scope, so Theorems 4.3 and 4.4 collapse if the lemma fails.","fun_headline_variants_meta":{"raw":{"variants":["Possibility posteriors hit normal and chi-square limits","Asymptotic normality without probability additivity","Outer-probability Bayes keeps classical asymptotic laws","No sampling distribution? Posterior still normal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00112,"raw_usage":{"total_tokens":4714,"prompt_tokens":1052,"completion_tokens":3662,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":3604}},"tokens_in":668,"tokens_out":3662,"duration_ms":34906,"temperature":1.0,"reasoning_tokens":3604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:45:16.166262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the paper's own ratio-of-means setting, simulate many datasets from the stated normal laws, and compare the empirical distribution of $\\sqrt{n}(\\theta^*_n-\\theta_0)$ with the predicted normal possibility function $N(0, V_*(s_{\\theta_0}(y))/I_*(\\theta_0)^2)$; if the mismatch persists as $n$ grows, Theorem 4.3 is wrong. More directly, exhibit two sequences of uncertain variables satisfying the convergence hypotheses of Proposition B.2 for which $x_n/z_n$ does not converge in outer probability to $x/\\alpha$; since the lemma is asserted without proof, such a counterexample would invalidate the proof chain of Theorems 4.3 and 4.4.","supporting_citations":[],"review_version":1}