{"id":"3742cf1c-317f-43b5-8c15-de8593f5ec7d","arxiv_id":"2507.19659","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For sequences with few differences, the posterior distribution of divergence time concentrates within a logarithmic factor of the p-distance under a broad class of dependent-site evolutionary models.","lead":"This paper proves a mathematical bound showing that when two DNA sequences differ at only a few sites, the Bayesian posterior estimate of their divergence time is unlikely to be much larger than the simple p-distance. The result covers evolution models with dependent sites, where exact likelihoods are often impossible to compute, so it could make phylogenetic inference faster and more reliable.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (2) identifies the divergence-time posterior with the single-lineage transition probability p_T(y|x); for two observed sequences the ancestor must be marginalized, and even in reversible stationary models the correct likelihood is p_{2T}(y|x).","rationale":"The reader's weakest-assumption pick matches mine. The theorem itself is a plausible concentration statement for the transition semigroup p_T, and the proof is largely self-contained; I see no circularity or hidden parameter-fitting. The load-bearing gap is the identification of p_T(y|x) with the likelihood of the divergence time of two extant sequences. That identification fails unless the model is stationary and reversible, and even then only after replacing T by 2T. Since the paper advertises a large class of dependent-site models without imposing reversibility, the stated conclusion is unsupported for that class. This is a correctness risk about the object being estimated, not about the technical bounds. I would keep the reader's CONDITIONAL verdict: the mathematical core can likely be repaired by adding the reversibility/stationarity hypothesis and a factor-2 rescaling, or by rewriting the paper as a bound for one-lineage branch lengths, so rejection would be too harsh, but acceptance requires the modeling assumption to be stated and the consequence for the interval constants to be worked out. The Corollary 1 log-term issue noted by the reader is real but secondary and patchable with a larger constant in h(n); it does not change the verdict.","tokens_in":19230,"tokens_out":15481,"duration_ms":171381,"concrete_test":"Take a small non-reversible context-dependent CTMC (e.g., n=4 binary sequences with neighbor-dependent asymmetric rates), enumerate all 16 ancestor states z, and compute the exact two-lineage marginal likelihood L(T|x,y)=Σ_z π(z)p_T(x|z)p_T(y|z) by matrix exponentiation. Generate x,y from the model with r log n/n=o(1), and compare the posterior under L with the paper's posterior based on Eq. (2). If the true posterior assigns non-negligible mass above (r/n)log n c while the paper's bound predicts exponential decay, Eq. (2) is not the divergence-time likelihood and Theorem 1 does not address the stated problem; if both concentrate, the concern is mainly a missing condition and rescaling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 defines the likelihood at Eq. (2) as L(T|x,y)=p(T,Q~)(y|x), the probability that a single lineage starting at x reaches y in time T. The theorem then bounds the posterior of T generated from this L. But the quantity named in the title and abstract is the divergence time of two observed sequences. For two sequences descended from an unknown common ancestor z, the conditional likelihood is L(T|x,y)=Σ_z π(z)p_T(x|z)p_T(y|z)/π(x); under stationarity and reversibility this equals p_{2T}(y|x), not p_T(y|x), and under the general non-reversible dependent-site CTMCs covered by the theorem it is not expressible as a single transition probability. Consequently Theorem 1 is a posterior bound for a directed ancestor-to-descendant time (or for a rescaled branch length), not for divergence time as claimed. All proof steps—the Lemma 1 lower bound, the Lemma 11 upper bound, and the Theorem 2 ratio—are built on the one-lineage kernel, so the result does not establish the stated biological conclusion unless the model class is restricted and T is reinterpreted. This is load-bearing because the main contribution is exactly a guarantee about where the divergence-time posterior mass lies.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies posterior concentration of the divergence time T between two length-n DNA sequences under continuous-time Markov models with site dependence. The main result (Theorem 1) states that when the Hamming distance r satisfies r log n / n = o(1), the posterior probability that T lies outside (0, (r/n) log n c(epsilon)) is at most e^{-c'(epsilon) r log n}, provided the prior gives non-exponentially small mass to a small interval around r/n. A corollary claims an improved bound T = O(r/n) with probability 1 - e^{-O(r)} for symmetric constant-rate models. The proofs use uniformization, a stochastic dominance coupling of the Hamming-distance process to a birth-death chain, hitting-time and spectral-gap estimates, and explicit tail bounds.","tokens_in":19446,"tokens_out":18712,"duration_ms":186025,"significance":"If correct, the result would provide a rigorous, model-rich generalization of Mihaescu and Steel's logarithmic posterior bound, connecting the simple p-distance to high-posterior credible regions for divergence times under dependent-site models. The paper is largely self-contained: the supporting lemmas are stated with explicit constants, the coupling argument is detailed, and the spectral-gap bound is constructive. The main concern is that the likelihood used in Eq. (2) is a single-lineage transition probability rather than the two-sequence divergence likelihood, so the theorem as stated does not apply to the biological target named in the title. The proof of Corollary 1 also leaves a 10 log n term that invalidates the e^{-O(r)} claim for bounded r.","major_comments":[{"comment":"The likelihood is defined as the single-lineage transition probability p(T,\\tilde Q)(y|x), but the paper's stated target is the divergence time of two sequences descended from an unknown common ancestor. For a stationary, reversible model the pairwise likelihood obtained by integrating over the ancestor is proportional to p_{2T}(y|x), not p_T(y|x); for general non-reversible dependent-site models it is not expressible as a single transition probability at all. Consequently Theorem 1 and Corollary 1 bound the posterior of a different quantity (an ancestor-to-descendant time, or a rescaled half-branch length) unless the model class and the meaning of T are restricted and the statements are revised. This is load-bearing because the abstract and introduction claim concentration for the divergence time of two sequences.","section":"Section 2, Eq. (2)"},{"comment":"The upper bound in Lemma 11 carries a +10 log n term. In the proof of Theorem 2, after choosing h(n) = 2r log(n c2 gamma_min)/(n c0), the ratio bound has exponent -(r-10) log n + O(r), which is positive for r < 10; since the assumption r log n / n = o(1) allows r constant, the conclusion exp(-O(r log n)) is not justified for small r. In Corollary 1, substituting t1 = O(r/n) into the displayed ratio bound leaves an exponent -O(r) + 10 log n + O(1), so the claimed e^{-O(r)} bound fails for bounded r. Both gaps can likely be repaired by choosing h(n) with a larger constant (or, for Corollary 1, by accepting a slower rate), but as written the proofs do not establish the stated rates.","section":"Section 4 (Theorem 2 proof) and Section 5.6 (Corollary 1 proof)"}],"minor_comments":[{"comment":"The definition of \\tilde gamma_max is identical to \\tilde gamma_min (both written as min over b and \\tilde x); it should be max over b and \\tilde x, since later arguments use \\tilde gamma_max as the maximum rate.","section":"Section 2, notation"},{"comment":"The constant c'(epsilon) appears in the probability bound but is not introduced in the statement; please define it alongside c(epsilon).","section":"Theorem 1 statement"},{"comment":"The final Hoeffding step is omitted; as written, the bound exp(-m* gamma_min^2 / 8) appears to require lambda = 1 or a missing lambda^2 factor, and the inequality \"2 delta / (lambda t q) - rho <= -gamma_min / 2\" is dimensionally inconsistent and should instead read -gamma_min / (2 lambda).","section":"Lemma 4 proof"},{"comment":"The phrase \"\\eta_0(J) does not decay exponentially in r\" should be formalized (for example, as \\eta_0(J) >= e^{-o(r)}), since the proof uses this condition to control \\eta_0(I^c)/\\eta_0(J).","section":"Theorem 1 and Section 3"},{"comment":"The interval notation \"I := (0, r c' / n log(n))\" is ambiguous; it should read I := (0, (r/n) c' log n).","section":"Theorem 2 statement"}],"recommendation":"major_revision","confidential_remarks":"The likelihood issue in Eq. (2) is the main obstacle to the paper's stated contribution. If the authors restrict to reversible stationary models and reinterpret T as a single-branch length (or rescale by a factor of 2), the technical results may still be correct and interesting, but the title, abstract, and biological framing need to be adjusted accordingly. The Corollary 1 gap is technical and likely repairable, but it needs explicit attention. The paper is otherwise self-contained and the coupling/spectral-gap approach is sound in broad outline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious look. The main theorem is a genuine extension of Mihaescu–Steel: posterior concentration for the divergence parameter under general finite-state CTMCs with context-dependent rates, with tail e^{-c r log n} around an interval of width p log n, and the log factor removed for constant-rate models. The proof is real work — Poisson tails, a coupling to a birth-death chain, hitting-time and spectral-gap arguments — and it is self-contained, with mostly explicit constants and no fitted parameters. Credit also for stating honestly that the log factor in the general bound is probably removable but they could not remove it.\n\nBut the paper quietly changes the quantity being bounded. Eq. (2) defines the likelihood as L(T | x, y) = p_T(y | x), the single-lineage transition probability. For two extant sequences descended from an unknown ancestor, the correct likelihood marginalizes the ancestor; under a reversible stationary model that is p_{2T}(y | x), not p_T(y | x), and for the general non-reversible dependent-site models the abstract advertises, it is not a single transition probability at all. Strictly, Theorem 1 bounds the posterior of a directed ancestor-to-descendant time, or of half the divergence time under reversibility — not the divergence time of two sequences as claimed. The fix is cheap: restrict to reversible stationary models and reparameterize, or say plainly that the result is for the rooted one-lineage model. As written, the abstract's claim about dependent-site divergence times is not established.\n\nThe second issue is Corollary 1. Its proof leaves a 10 log n term in the exponent while claiming e^{-O(r)}; that only goes through when r is at least of order log n, which is not stated. This looks fixable, but as written the corollary is not fully proved.\n\nMinor: the stochastic-dominance casework in Lemma 3 is loose in one displayed inequality (a stay probability is dropped), and the proof of Corollary 1 redefines h(n) mid-argument. Both are cleanup, not structural.\n\nBottom line: the core theorem is likely correct and the machinery deserves a real referee. The referee should push hard on the model class behind Eq. (2) and on the Corollary 1 proof. I would accept conditionally, not as-is.","headline":"Real concentration result for a single-lineage branch time under dependent-site CTMCs, but Eq. (2) is not the two-sequence divergence-time likelihood in general, and Corollary 1's proof has an unabsorbed log n.","tokens_in":19983,"tokens_out":14048,"would_cite":true,"duration_ms":156158,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J27","92D15","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new theorem places the posterior distribution of divergence time within a logarithmic factor of the p-distance, even under dependent-site evolutionary models.","keywords":["divergence time estimation","posterior concentration","p-distance","site-dependent substitution models","context-dependent rates","continuous-time Markov chain","birth-death coupling","phylogenetics"],"falsifier":"Simulate a reversible, stationary dependent-site continuous-time Markov chain with known divergence time $T^*$ and a non-pathological prior, for several large $n$ with $r \\approx T^* n$ and $r\\log(n)/n$ small; if the observed posterior repeatedly puts non-negligible mass above $(r/n)\\log(n)\\, c(\\epsilon)$ for fixed $c(\\epsilon)$, the bound fails. Equivalently, for a constant-rate model, check numerically whether the probability that $T$ exceeds $c r/n$ decays like $e^{-c''r}$; a polynomial or flat tail would contradict Corollary 1.","tokens_in":18998,"feed_emoji":"🧬","tokens_out":9514,"duration_ms":104980,"temperature":0.7,"pith_summary":"For two observed DNA sequences with Hamming distance $r$, the paper proves that when $r\\log(n)/n$ is small, the posterior distribution of their divergence time $T$ concentrates below $(r/n)\\log(n)$ times a constant, with the probability of being outside this interval decaying like $\\exp(-O(r\\log n))$. The result holds for a broad class of continuous-time Markov models in which sites evolve dependently, provided the prior does not place exponentially small mass near $r/n$. For models with constant mutation rates, a corollary removes the log factor and shows that $T$ exceeds a constant multiple of the p-distance with vanishingly small posterior probability. This matters because under site-dependent models the likelihood is expensive or intractable, so knowing in advance that high-posterior $T$ values live near the cheap p-distance makes MCMC and maximum-likelihood searches much more efficient.","feed_headline":"Bayesian divergence time rarely exceeds p-distance log n","feed_subtitle":"For dependent-site models, likelihood searches can be confined to a small interval around the simple mutation count.","key_machinery":"The proof works on the uniformized jump-chain representation of the continuous-time Markov chain, writing the likelihood as $e^{-\\lambda T} \\sum_m (\\lambda T)^m/m!\\, \\tilde{R}^m_{x,y}$, a Poisson mixture over the number of mutations $m$. The sum is split into three regimes: few jumps, an intermediate range, and very many jumps. The few-jump terms are controlled by a Poisson tail bound; the intermediate and many-jump terms are bounded by coupling the Hamming distance from $x$ to a stochastically minorizing birth-death chain on $\\{0,\\dots,n\\}$, whose stationary distribution is binomial. Hitting-time estimates control the intermediate regime, and a spectral-gap bound via canonical paths controls the many-jump regime once the chain is near stationarity. A crude lower bound, Lemma 1, $p(T,\\tilde{Q})(y|x) \\ge \\exp(-\\lambda T + r\\log(T\\tilde{\\gamma}_{\\min}))$, combines with the upper bounds to give the exponential likelihood ratio.","core_discovery":"At the center is Theorem 1: under the low-divergence condition $r\\log(n)/n = o(1)$, there is a constant $c(\\epsilon)$ such that, for all sufficiently large $n$, the posterior probability that $T$ lies in $(0, (r/n)\\log(n)\\, c(\\epsilon))$ is at least $1 - \\exp(-c'(\\epsilon)\\, r\\log n)$, whenever the prior gives $J = ((1-\\epsilon)r/n, (1+\\epsilon)r/n)$ mass that does not decay exponentially in $r$. The proof obtains a sharper two-sided likelihood statement, Theorem 2: for any $T^*$ in $J$ and any $T$ outside the longer interval, the likelihood ratio $p(T,\\tilde{Q})(y|x)/p(T^*,\\tilde{Q})(y|x)$ is at most $\\exp(-O(r\\log n))$. For symmetric constant-rate models, Corollary 1 strengthens this to an interval $(0, c r/n)$ with posterior tail $\\exp(-c'' r)$, removing the log factor. The paper's interpretation is that the simple p-distance is, up to a logarithmic factor, a genuine upper limit on where the posterior of divergence time can sit under a large class of dependent-site models.","pith_inferences":["The authors suspect that the $\\log n$ factor is an artifact of ignoring the conditional distribution over sequences at distance $r$; a natural test is whether the $c r/n$ bound holds under dependent-site models once $P(X_m = y \\mid d_H = r)$ is accounted for, as done in the constant-rate corollary.","The theorem is stated for the single-lineage transition probability $p(T)(y|x)$; if the true two-sequence likelihood under a non-reversible model differs, the bound should be rechecked against a symmetrized likelihood before being used in practice.","For finite $n$ the constants $c$ and $c'$ are implicit; simulations could calibrate how large $n$ must be for the exponential tail to dominate, and whether the bound is practically useful for typical sequence lengths.","A related testable extension is that the same coupling and spectral-gap machinery might yield concentration bounds on estimated branch lengths for three or more sequences, or on whole phylogenetic trees, rather than only pairwise divergence times."],"forward_implications":["For low-to-moderately diverged sequences, posterior sampling and maximum-likelihood routines only need to evaluate the likelihood for $T$ in $(0, O((r/n)\\log n))$, avoiding expensive likelihood approximations at large $T$ under site-dependent models.","For constant-rate models, the posterior credible region shrinks to $(0, c r/n)$, recovering and refining the earlier logarithmic bound for the two-state symmetric model.","The condition $r\\log(n)/n = o(1)$ marks the limits of the result: once the sequences are too diverged, the likelihood flattens and a bound of this type cannot hold, consistent with the JC69 maximum-likelihood estimator failing for $r/n \\ge 3/4$.","The posterior tail bound is uniform over a broad class of dependent-site models, so the same interval can be used to initialize or constrain iterative optimization in models where exact likelihoods are intractable.","The bound also implies that, under a non-pathological prior, the posterior cannot be dominated by large divergence times even though the likelihood ratio favors small $T$ exponentially strongly."],"supporting_citations":[{"why":"Supplies the earlier logarithmic posterior bound for the two-state symmetric model that this paper extends and, in the constant-rate case, improves.","marker":"[30]"},{"why":"Provides the uniformization representation of the continuous-time Markov chain and the coupling and stochastic dominance facts used throughout the proof.","marker":"[32]"},{"why":"Gives the reversible-Markov-chain bound used to control the probability of being near $r$ after hitting a larger state.","marker":"[33]"},{"why":"Supplies the spectral-gap mixing bound $B^m_{0,k} \\le \\nu(k)+e^{-\\mathrm{SpecGap}(B)m}/\\min \\nu$ used for the many-jump regime.","marker":"[34]"},{"why":"Provides the canonical-path method used to lower-bound the spectral gap of the birth-death chain.","marker":"[35]"},{"why":"The constant-rate nucleotide model (JC69) serves as the concrete setting for the corollary and for the discussion of maximum-likelihood existence.","marker":"[10]"}],"fun_headline_variants":["Posterior divergence time concentrates within log factor of p-distance","For dependent-site models, posterior on T is within log factor of p","Logarithmic upper bound on posterior divergence time from p-distance","Posterior T within log n factor of p-distance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof treats the probability that one observed sequence evolves into the other along a single lineage as the likelihood of their divergence time, which is the full likelihood only when the model is reversible and stationary; for general non-reversible dependent-site models, the object being bounded is a different quantity.","fun_headline_variants_meta":{"raw":{"variants":["Posterior divergence time concentrates within log factor of p-distance","For dependent-site models, posterior on T is within log factor of p","Logarithmic upper bound on posterior divergence time from p-distance","Posterior T within log n factor of p-distance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00138,"raw_usage":{"total_tokens":5607,"prompt_tokens":979,"completion_tokens":4628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":4558}},"tokens_in":595,"tokens_out":4628,"duration_ms":35759,"temperature":1.0,"reasoning_tokens":4558,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:13:08.568146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a reversible, stationary dependent-site continuous-time Markov chain with known divergence time $T^*$ and a non-pathological prior, for several large $n$ with $r \\approx T^* n$ and $r\\log(n)/n$ small; if the observed posterior repeatedly puts non-negligible mass above $(r/n)\\log(n)\\, c(\\epsilon)$ for fixed $c(\\epsilon)$, the bound fails. Equivalently, for a constant-rate model, check numerically whether the probability that $T$ exceeds $c r/n$ decays like $e^{-c''r}$; a polynomial or flat tail would contradict Corollary 1.","supporting_citations":[{"cited_title":"Applied Mathematics Letters 23(9), 975–979 (2010)","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier logarithmic posterior bound for the two-state symmetric model that this paper extends and, in the constant-rate case, improves."},{"cited_title":"John Wiley & Sons, Hoboken, New Jersey (1995)","cited_arxiv_id":null,"evidence_quote":"Provides the uniformization representation of the continuous-time Markov chain and the coupling and stochastic dominance facts used throughout the proof."},{"cited_title":"Unfinished monograph (2002)","cited_arxiv_id":null,"evidence_quote":"Gives the reversible-Markov-chain bound used to control the probability of being near $r$ after hitting a larger state."},{"cited_title":"American Mathematical Society, Providence, RI (2009)","cited_arxiv_id":null,"evidence_quote":"Supplies the spectral-gap mixing bound $B^m_{0,k} \\le \\nu(k)+e^{-\\mathrm{SpecGap}(B)m}/\\min \\nu$ used for the many-jump regime."},{"cited_title":"Annals of Applied Probability 1(1), 36–61 (1991)","cited_arxiv_id":null,"evidence_quote":"Provides the canonical-path method used to lower-bound the spectral gap of the birth-death chain."},{"cited_title":"In: Munro, H.N","cited_arxiv_id":null,"evidence_quote":"The constant-rate nucleotide model (JC69) serves as the concrete setting for the corollary and for the discussion of maximum-likelihood existence."}],"review_version":1}