{"id":"16b3c415-1179-4f63-b8bd-a324cb7c6ee8","arxiv_id":"2509.12734","paper_version":4,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper claims to prove consistency and asymptotic normality of MLEs in the Linkage Model and to construct an asymptotic level-alpha likelihood-ratio test for Linkage versus Admixture models, but the proofs contain significant gaps.","lead":"This paper proves consistency and asymptotic normality for maximum likelihood estimates of ancestry in the Linkage Model, a hidden Markov model used in population genetics, and uses these results to propose a statistical test that decides between the Linkage Model and the simpler Admixture Model. A scientist might read it to see whether a rigorous basis exists for choosing between these two widely used models, and whether the proposed test can be trusted on real genetic data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3 is unsupported: under H0, r=∞ lies on the boundary of Θ and the Fisher information for r is singular, so Wilks' theorem does not apply; the LR statistic likely follows a mixture, not χ²(1).","rationale":"The reader identified both non-identifiability for K>2 and the boundary issue as fragile premises, but ranked identifiability as most load-bearing. I argue the boundary issue is the single most fundamental problem: it directly invalidates the paper's advertised result (Theorem 3) even if every other theorem were correct. The Fisher information singularity at r=∞ is a concrete mathematical obstruction, not a missing proof. The paper's own simulation is too underpowered to detect the difference between χ²(1) and a mixture. Therefore, the reader's REJECT verdict is supported, and no adjustment is needed. I marked partial agreement because my primary concern is the boundary issue, not the identifiability gap.","tokens_in":16384,"tokens_out":6399,"duration_ms":64640,"concrete_test":"Run the simulation of Sec. 4.1 with M=5000 markers and 10,000 replicates under H0, simulating X from the Admixture Model (r=∞). Compute Λ and (a) compare the empirical distribution to χ²(1); (b) estimate the fraction of Λ=0. If the rejection rate at the 0.95 χ²(1) quantile is ≈0.025 rather than 0.05, or if P(Λ=0)→≈0.5, the boundary issue is confirmed and Theorem 3's χ²(1) limit is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 3) asserts Λ → χ²(1) under H0: r=∞. But H0 places the true parameter on the endpoint of Θ=[r_lb,∞], and the log-likelihood depends on r only through e^{-d_m r}, whose derivative and second derivative w.r.t. r vanish at r=∞. Hence the Fisher information J(q0,∞) is not positive definite, and Theorem 2 (which requires interior parameters and J≻0) does not apply under H0. The paper then invokes Wilks (1938) with no verification of its conditions. Reparameterizing ρ=1/r (or s=e^{-r}) puts the null at the boundary ρ=0; standard asymptotics for boundary nulls (Self & Liang 1987) give a mixture limit, typically 0.5δ_0+0.5χ²(1), not χ²(1). The paper's own simulation (Sec. 4.1, M=100, 100 reps) reports type-1 error 'below 0.05', consistent with a conservative mixture rather than exact χ²(1). The real-data application (Sec. 4.2) uses diploid individuals with C=20, violating Assumption 3.1 (haploid, C=1). Thus Theorem 3 is not merely missing a proof; it appears mathematically false.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines the Linkage Model as a time-inhomogeneous hidden Markov model with stationary distribution q, and the Admixture Model as its r=∞ limit. It claims consistency (Theorem 1) and asymptotic normality (Theorem 2) of the MLE for (q,r) under Assumption 3.1, and then, invoking Wilks (1938), claims in Theorem 3 that the likelihood-ratio test statistic Λ for H0:r=∞ versus H1:r<∞ converges in distribution to χ²(1) under H0, making the test an asymptotic level-α test. The paper also reports simulations and a real-data application to 1000 Genomes data with K=5 and C=20.","tokens_in":16660,"tokens_out":12675,"duration_ms":141279,"significance":"If the central claims were correct, the paper would fill a real gap in MLE asymptotics for an inhomogeneous HMM whose stationary distribution is constant across time, and would provide the first model-selection test between the Admixture and Linkage Models with a stated asymptotic error rate. The author provides code and attempts numerical evaluation. However, the central theorem depends on an invalid application of Wilks' theorem to a boundary null with singular Fisher information, and on an unproved identifiability assumption for K>2. As it stands, the claimed distributional result and therefore the level-α justification are not established.","major_comments":[{"comment":"The null hypothesis r=∞ lies on the boundary of Θ=[r_lb,∞], not in the interior. Moreover, the transition probabilities depend on r only through e^{-d_m r}, whose derivatives with respect to r vanish at r=∞. Hence the Fisher information J(q0,∞) is not positive definite, and Theorem 2 — which explicitly requires interior parameters and J≻0 — cannot be applied under H0. The sentence 'According to Wilks (1938)...' does not verify any of Wilks' conditions. Standard boundary asymptotics (Self and Liang 1987) give a mixture distribution with point mass at 0, not χ²(1). The simulation in Sec. 4.1 (100 replicates per cell; type-1 error 'below 0.05') is consistent with a conservative boundary mixture and does not validate the χ²(1) claim.","section":"Section 3, Theorem 3 and Definition 2.4"},{"comment":"For general K>2, identifiability of (q,r) from the marginal distribution is assumed without proof: 'we will, without a proof, just assume that the parameters are identifiable... The proof is basically the same.' This is load-bearing: identifiability is used for uniqueness of the MLE (Theorem 4), consistency (Theorem 1), and the CLT (Theorem 2), and is also needed for the real-data application with K=5. The Kruskal-type rank conditions for K>2 are not stated, and no proof is provided.","section":"Section 5.1, after Lemma 5.1"},{"comment":"The algebraic analysis of the K=2 identifiability proof is incorrect in a stated detail. For the displayed matrix B2, det(B2)=q1 p1,2 −(1−q1)p2,2, which vanishes when q1=p2,2/(p1,2+p2,2), not when q1=p1,2/(p1,2+p2,2) as written. Thus the rank condition excludes a different value of q from the one stated, and the proof does not establish full rank for all admissible q. The claim that the second factor in det(M1) has roots outside [0,1] may be true, but the surrounding rank analysis therefore remains incomplete.","section":"Section 5.1, Lemma 5.1"},{"comment":"The proof asserts that the three conditions of the martingale CLT in Hall and Heyde (2014) 'follow directly' from uniform boundedness, but conditional variance convergence and the Lindeberg condition are never verified. More seriously, the object M_n as displayed does not appear to be a martingale: it is a sum Σ N_k plus a terminal conditional-expectation correction, not a sum of martingale differences; E[M_n−M_{n−1} | F_{n−1}] is not shown to vanish. In addition, the score ∇D is vector-valued, while σ_n is treated as scalar. Hence the proof of Theorem 2 is incomplete even for finite r.","section":"Section 5.3, Proposition 5.10"},{"comment":"The real-data application violates Assumption 3.1, which fixes C=1 and haploid individuals. The 1000 Genomes analysis uses C=20 and diploid individuals, and the allele frequencies are estimated from the remaining individuals rather than being known. No extension theorem is proved. Consequently the reported real-data results are not covered by Theorems 1–3.","section":"Section 4.2 versus Assumption 3.1"}],"minor_comments":[{"comment":"'Fischer information' should be 'Fisher information' throughout (e.g., Theorem 2, Remark 3.3).","section":"General"},{"comment":"The definition of h_i is garbled: 'h_i = 1− fi−1 Eπ0(fi)−1' is not readable. Please clarify the formula and the role of f_i.","section":"Section 5.1, Lemma 5.5"},{"comment":"Only 100 simulation replicates are used per cell. At a nominal 5% level, the Monte Carlo standard error is about 2.2 percentage points, so 'type-1 error below 0.05' cannot be distinguished from 0.05 and is not evidence for the χ²(1) approximation.","section":"Section 4.1"},{"comment":"Assumption 3.1 sets C=1, but condition (A3) is written with a product over c=1,...,C. The notation is inconsistent. Also, since each λ_{c,m}<1 under Assumption 3.1, the product automatically tends to 0, so the intended condition and its role in the proofs should be clarified.","section":"Assumption 3.1 / A3"},{"comment":"The parameter space Θ includes r=∞, making it non-compact as a subset of R^{K+1}. The paper should state explicitly how maxima over this extended parameter space are defined and why they exist.","section":"Definition 2.1"},{"comment":"The test statistic maximizes over q∈S^K in the numerator but over Θ in the denominator. Since Θ includes r=∞, the numerator is a restricted version of the denominator; this is fine, but the existence and uniqueness of all maximizers should be stated.","section":"Section 2, Definition 2.4"}],"recommendation":"reject","confidential_remarks":"The manuscript has a promising topic and the author's prior work is relevant, but the central claim (Theorem 3) appears to be mathematically false as stated because the null lies on the boundary and the Fisher information in r is singular. This is not a presentation issue but a fundamental gap in the model-selection test. The proof of the CLT is also incomplete, and identifiability for K>2 is explicitly assumed. These problems cannot be fixed by local revision; the paper would need a substantial reframing, e.g., a nonstandard boundary asymptotic analysis or a different test construction, before it could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's flagship result, Theorem 3, is not supported. It invokes Wilks (1938) on the basis of Theorem 2's asymptotic normality, but under H0: r=∞ the parameter is on the boundary of Θ=[r_lb,∞]. The log-likelihood depends on r through e^{-dr}, whose derivative vanishes at r=∞, so the Fisher information in r is zero. Standard boundary asymptotics (Self and Liang 1987) give a mixture, typically 0.5δ_0+0.5χ²(1), not χ²(1). The paper's own simulations show type-1 error below 0.05, which is consistent with a conservative mixture rather than the claimed chi-square calibration.\n\nWhat is genuinely new: consistency and asymptotic normality of MLEs in the Linkage Model, a time-inhomogeneous HMM whose hidden chain has a common stationary distribution across markers. That is a nontrivial extension of standard HMM theory, even if the paper doesn't sharply delineate the novelty from Douc et al. (2011). The paper also ships code and applies the test to 1000 Genomes data, which is more than most theory papers do.\n\nWhere it falls apart: the identifiability assumption. Section 5.1 states that for K>2 the parameters are assumed identifiable 'without a proof,' and that the proof is 'basically the same.' That assumption is load-bearing for Theorem 4 and therefore for consistency and the CLT. The K=2 proof (Lemma 5.1) also has a determinant factorization I couldn't reproduce—quick numeric checks suggest the algebra is wrong. Proposition 5.10 asserts the Hall-Heyde conditions 'follow directly' without verification; they might hold, but 'directly' is doing a lot of work. And the real-data application uses diploid individuals with C=20 and estimated allele frequencies, which violates Assumption 3.1; Remark 2.2 gestures at the diploid extension but it's not proven.\n\nThe paper is honest about the gaps, and the boundary issue is a known subtlety, not something the author invented. But the central advertised theorem is, as stated, likely false. The fix—reparameterize to ρ=1/r and use mixture asymptotics—would salvage the test, but that's a fresh piece of work.\n\nWho this is for: someone working on HMM asymptotics or forensic ancestry inference who wants a concrete worked model with nonstationary transitions. The consistency/CLT content deserves a serious referee, but not as it stands. I'd send it out rather than desk-reject, because the mathematical core is real and the flaws are identifiable and repairable. I would not cite the test result.","headline":"The advertised level-alpha test is not proven—boundary problem at r=∞ breaks the Wilks argument—but the consistency and CLT results for the Linkage Model are real contributions worth a referee's time.","tokens_in":17176,"tokens_out":3141,"would_cite":false,"duration_ms":32393,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F03","62F12","62M05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A likelihood-ratio test for choosing between linkage and admixture genetic models is proven to be an asymptotic level-alpha test.","keywords":["Linkage Model","Admixture Model","Hidden Markov Model","Maximum Likelihood Estimator","Consistency","Asymptotic Normality","Likelihood Ratio Test","Model Selection"],"falsifier":"Simulate data under the Admixture Model (r = infinity) with K = 3 populations and many markers satisfying Assumption 3.1, and estimate the distribution of Lambda. If the empirical Type I error does not approach the nominal level alpha as the number of markers grows, Theorem 3 is false. Alternatively, exhibit two different pairs (q, r) with K = 3 that produce identical marginal distributions for the observed alleles, which would refute the identifiability assumption on which the MLE theory rests.","tokens_in":16196,"feed_emoji":"🧬","tokens_out":7228,"duration_ms":68150,"temperature":0.7,"pith_summary":"This paper establishes asymptotic guarantees for maximum likelihood estimation in the Linkage Model, a hidden Markov model used in population genetics to describe ancestry across linked genetic markers. It proves that the maximum likelihood estimators of an individual's ancestry proportions and of the recombination rate are consistent and asymptotically normal as the number of markers grows. It then proves that the likelihood-ratio test statistic for comparing the Linkage Model with its special case, the Admixture Model (independent markers), converges in distribution to a chi-square distribution with one degree of freedom under the null hypothesis, making the test an asymptotic level-alpha test. Applied to human genetic data, the test rejects the Admixture Model for roughly 87.5% of individuals, with large differences across continental groups.","feed_headline":"New test separates linked vs. unlinked genetic ancestry models","feed_subtitle":"Proven to follow a chi-square distribution, the test gives a controlled error rate for model selection in population genetics.","key_machinery":"The test statistic Lambda = -2 log(LR) (Definition 2.4) is the central object. Its asymptotic behavior is inherited from the asymptotic normality of the MLE, which is established via the representation of the log-likelihood as a sum of conditional log-likelihood increments and a martingale central limit theorem for the score. The transition matrix of the hidden ancestry chain (equation 2) is the core mechanism: it defines the Linkage Model, has q as its invariant measure at every marker, and is uniformly ergodic under the paper's assumptions, which is what makes the martingale arguments work.","core_discovery":"The central claim of the paper is Theorem 3: the test statistic Lambda = -2 log(LR) converges in distribution to a chi-square distribution with one degree of freedom under the null hypothesis that the recombination parameter r is infinite, so the test that rejects when Lambda exceeds the 1-alpha quantile is an asymptotic level-alpha test. This is obtained by first proving that the maximum likelihood estimator in the Linkage Model is consistent (Theorem 1) and asymptotically normal (Theorem 2), and then invoking classical likelihood-ratio asymptotics. The key technical step is a martingale central limit theorem for the score function, using the fact that the hidden Markov chain is uniformly e","pith_inferences":["The boundary issue: because the null value r = infinity lies at the boundary of the parameter space, the chi-square approximation may need modification; this is an editorial concern, not addressed in the paper.","The identifiability gap: the proof of identifiability is given only for K = 2; for K > 2 the paper assumes it, so the general-K results rest on an unproven premise.","Power limitations: for large recombination rates r or small genetic distances d, the Linkage and Admixture models become nearly indistinguishable, so the test's power is expected to drop; this is consistent with the paper's own simulation results.","A practical extension would be to use the test to select marker sets that maximize power to detect linkage, or to adapt it to unsupervised settings where allele frequencies are estimated."],"forward_implications":["The test offers a data-driven way to choose between the Admixture Model and the Linkage Model for a given individual or population, with an asymptotically controlled false-rejection rate.","The central limit theorem provides the first uncertainty quantification for MLEs of ancestry and recombination rate in the Linkage Model, enabling confidence regions for these quantities.","The consistency and uniqueness results extend known asymptotic theory for the Admixture Model to the Linkage Model, covering hidden Markov models with time-inhomogeneous transitions but a constant stationary distribution.","Applied to real human genetic data, the test indicates that the Admixture Model is sufficient for a minority of individuals, with substantial variation across continental groups, so a single global model choice is not appropriate."],"fun_headline_variants":["Chi-square test distinguishes linkage from admixture","Valid test for linkage vs. admixture in ancestry","Model choice test for linked or independent markers","Statistical test for ancestry model comparison"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes, without proof for K > 2, that the parameters (q, r) are identifiable from the marginal distribution of the observations; it also invokes classical likelihood-ratio asymptotics even though the null value r = infinity lies on the boundary of the parameter space, where that asymptotics is not automatically valid.","fun_headline_variants_meta":{"raw":{"variants":["Chi-square test distinguishes linkage from admixture","Valid test for linkage vs. admixture in ancestry","Model choice test for linked or independent markers","Statistical test for ancestry model comparison"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2384,"prompt_tokens":709,"completion_tokens":1675,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":1618}},"tokens_in":453,"tokens_out":1675,"duration_ms":17702,"temperature":1.0,"reasoning_tokens":1618,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:34:11.228069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data under the Admixture Model (r = infinity) with K = 3 populations and many markers satisfying Assumption 3.1, and estimate the distribution of Lambda. If the empirical Type I error does not approach the nominal level alpha as the number of markers grows, Theorem 3 is false. Alternatively, exhibit two different pairs (q, r) with K = 3 that produce identical marginal distributions for the observed alleles, which would refute the identifiability assumption on which the MLE theory rests.","supporting_citations":[],"review_version":1}