{"id":"867588bf-9d8d-450a-ba0e-cb4b2fd2845a","arxiv_id":"2607.07188","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"Optimal trans-model MCMC mixing requires proposing high-posterior models and posterior-matched parameters; maximum jump probability is necessary but not always sufficient for maximum efficiency.","lead":"This paper analyzes when trans-model MCMC algorithms mix efficiently, deriving conditions under which proposing parameters from the posterior and preferentially targeting high-probability models maximizes sampling efficiency. The results give concrete design guidelines for Bayesian phylogenetic tree search and other cross-model inference problems.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Two-model theory's optimality conditions may not extend to K>2 models with parameters; the bridge to phylogenetics is an unproven analogy.","rationale":"The reader correctly identifies the most significant gap: the theory covers two models with one parameter-free, while the applications involve K>>2 models with many parameters each. I agree this is the load-bearing concern. The theoretical analysis itself is mathematically sound—the derivations for the two-model case are clean, the spectral decomposition in SI text 4 is standard, and the examples (uniform, Gaussian) are well-chosen. The concern is not about internal inconsistency but about the scope of the claims. The paper is honest about the gap in the Discussion, but the abstract and framing ('principles are applied to MCMC algorithms in phylogenetic reconstruction') suggest a stronger connection than exists. The phylogenetic empirical results are practically useful and consistent with prior work, but they rely on hand-tuned parameters (exponent -1/2 in eq. S90, scale a=0.5 in eq. S91) that are acknowledged as arbitrary, and only two datasets are tested. The K>2 extension in SI text 4 assumes posterior proposals, which is the ideal case the theory is supposed to justify, creating a mild circularity. This does not invalidate the theoretical contribution but limits its practical applicability. CONDITIONAL is appropriate: the theory is valid within its stated scope, but the extension to phylogenetics is an unproven analogy. A 3-model analytical check would clarify whether the optimality conditions generalize, which would either strengthen or weaken the bridge to applications. The lack of publicly available code (Dryad placeholder) further limits verification of the empirical results, though this is secondary to the theoretical gap.","tokens_in":45889,"tokens_out":811,"duration_ms":381678,"concrete_test":"Construct a 3-model example where H₀ has no parameters and H₁, H₂ each have one parameter with known posteriors. Discretize the parameter spaces and compute the optimal reversible Markov chain using the Frigessi et al. framework. Check whether the optimal chain satisfies the same structural conditions as the 2-model case: (a) when the target model probability π₀<1/2, is maximum P_jump necessary but not sufficient for maximum E? (b) when π₀≥1/2, is maximum P_jump sufficient for maximum E? If either condition fails, the 2-model theory does not directly generalize and the phylogenetic guidelines lack theoretical support.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central theoretical results (eqs. 28-38) characterize optimal trans-model MCMC for two models where H₀ has no parameters. The extension to K>2 models (SI text 4) explicitly assumes parameters are proposed from the posterior during cross-model moves: q(θ_{k'}|k,θ_k,k') = π(θ_{k'}|k'). This is precisely the ideal the paper aims to justify, making the K>2 extension somewhat circular as a justification for the practical guidelines. The phylogenetic applications involve K>>2 trees, each with 2n-3 branch-length parameters, where proposing from the posterior is infeasible. The paper acknowledges this gap ('General cases...are yet to be studied') but the abstract and discussion present the phylogenetic guidelines as flowing from the theory. The concern is whether the two-model optimality conditions (particularly the sufficiency of maximum P_jump for maximum E when π₀≥1/2, and the necessity of within-model moves when π₀<1/2) actually hold in the K>2, multi-parameter setting. If they do not, the empirical recommendations (direct transfer of branch lengths, branch weights, parsimony weights) are heuristics validated only by the two specific datasets tested, not by the theory.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This paper studies the mixing efficiency of trans-model MCMC algorithms, with particular attention to Bayesian phylogenetics where tree topologies constitute different models. The authors analyze two simple but illuminating examples (uniform and Gaussian target distributions) involving two models, one of which (H0) has no free parameters. They derive conditions under which the model-jump probability P_jump and the mixing efficiency E reach their maxima, and characterize when within-model moves are necessary to achieve optimal efficiency. The key theoretical results state that: (i) P_jump is bounded above by 2(1−max{π_k}) (Theorem 1); (ii) when π_0 < 1/2, maximum P_jump is necessary but not sufficient for maximum E, and within-model moves restoring parameters to their stationary distribution are needed unless the proposal g(·) already equals the posterior π_1(·); (iii) when π_0 ≥ 1/2, maximum P_jump is both necessary and sufficient for maximum E. The K > 2 discrete case is treated via the framework of Frigessi et al. (1992). The theory is then applied to NNI and SPR tree-proposal algorithms using two real datasets (primate ψη-globin pseudogenes and mammalian mitochondrial genes), testing strategies for branch-length proposals and tree-proposal weights.","tokens_in":46205,"tokens_out":1685,"duration_ms":224180,"significance":"The paper addresses a genuinely under-studied problem: the mixing efficiency of cross-model MCMC moves, as opposed to the well-studied within-model case. The mathematical analysis for the two-model case is clean and self-contained, with explicit derivations for both uniform and Gaussian examples (SI texts 2–3) and a correct proof of Theorem 1 (SI text 1). The discrete-state optimality results (SI text 4) correctly build on Frigessi et al. (1992). The practical guidelines—preferentially propose high-posterior models, propose parameters near the posterior mode, use direct transfer of branch lengths rather than merge-and-split—are sensible and supported by the empirical results (Tables 2–3), where NNI with direct transfer outperforms SPR baseline by factors of 16–21. The paper provides falsifiable, quantitative predictions. The distinction drawn between within-model and trans-model MCMC (e.g., that intermediate acceptance rates are optimal for within-model moves but high acceptance rates are desirable for cross-model moves) is a useful conceptual contribution.","major_comments":[{"comment":"The central theoretical results (eqs. 28–38) are derived for exactly two models where H0 has no free parameters. The extension to K > 2 models (SI text 4) explicitly assumes parameters are proposed from the posterior during cross-model moves (q(θ_{k'}|k,θ_k,k') = π(θ_{k'}|k')), which is the ideal the paper aims to justify. The phylogenetic applications involve K >> 2 trees, each with 2n−3 branch-length parameters, where proposing from the posterior is infeasible. The authors acknowledge this gap (Discussion: 'General cases where both models have free parameters, or where more than two models with parameters are under comparison, are yet to be studied'), but the abstract and discussion present the phylogenetic guidelines as flowing from the theory. The paper would benefit from explicitly stating, early in the phylogenetics section, that the connection between the two-model theory and theK","section":null},{"comment":"The empirical results in Tables 2–3 are based on only two datasets. The ψη-globin dataset has 6 species (105 possible trees) with the top three trees accounting for ~100% of the posterior, while the mt dataset has 29 species with the top two trees having posteriors 0.574 and 0.107. These are quite different scenarios, yet the relative performance of algorithms is broadly similar across both. It would strengthen the paper to discuss whether the theoretical conditions (e.g., π_0 ≥ 1/2 vs. π_0 < 1/2) are met in each dataset, and whether the empirical findings are consistent with the theory's predictions for each case. Currently, the link between the theoretical conditions and the empirical settings is not made explicit.","section":null},{"comment":"The branch weight exponent in eq. S90 (w_f ∝ t_f^{-1/2}) and the parsimony weight scale parameter a in eq. S91 (w_{k'} ∝ exp((S_k − S_{k'})/a), with a = 0.5 chosen after testing values 0.1–20) are free parameters tuned on the same datasets used for evaluation. The paper does not discuss whether this tuning constitutes overfitting or how the optimal values might transfer to other datasets. The authors note that 'our branch weight (eq. S90) is arbitrary and may suit one dataset better than another' (Results section), but the parsimony weight tuning receives no such caveat. A brief discussion of the sensitivity to these choices would be appropriate.","section":null}],"minor_comments":[{"comment":"The notation π_0 is used for the posterior probability of H0, but π_1(θ) is used for the posterior density of θ within H1. This is consistent but occasionally confusing; a clarifying note on first use would help.","section":null},{"comment":"In the caption of Figure 4, the note states 'the round flat tops in (a) and (c) are the same, the apparent differences being due to numerical issues in calculation of E and in kernel density smoothing.' If numerical issues affect the figure, consider whether the figure can be improved or whether the statement should be more prominent.","section":null},{"comment":"The reference to 'Liet al., 2000' in the Introduction appears to be 'Li et al., 2000' with a missing space. Several other citation formatting issues exist (e.g., 'Chenet al., 2014', 'Ronquistet al., 2012').","section":null},{"comment":"In the Discussion, the statement 'it is common to run an MCMC algorithm for N = 10^9 iterations, say, achieving an ESS of <100, in which case E < 10^{-7}' is presented without citation. A reference or caveat would strengthen this claim.","section":null},{"comment":"Figure 8 caption: 'The cross-tree move is B4 (NNI bw) while the within-tree move changes one randomly sampled branch length.' It would help to state the y-axis units more explicitly in the caption (e.g., 'Time(s)/100' for panel a and 'Time(h) 1000' for panel b are somewhat cryptic).","section":null},{"comment":"The paper mentions Lindley's paradox in the context of Figure 5 but does not elaborate on its relevance to the mixing efficiency problem. A sentence connecting the paradox to the practical implications for trans-model MCMC would improve clarity.","section":null},{"comment":"SI text 2 derives autocorrelation functions for the case π_0 < 1/2 with maximum P_jump but suboptimal E. The derivation is detailed but the key insight—that ρ_1 achieves its optimal value but ρ_k for k ≥ 2 do not—could be stated more prominently at the beginning of the section rather than only in the surrounding main text.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The stress-test concern about the gap between two-model theory and K >> 2 phylogenetic applications is valid and is the main limitation of the paper. However, the authors are transparent about this gap in the Discussion, and the paper does not claim to have proven the extension. The theoretical results for the two-model case are correct and non-trivial, and the empirical results, while limited to two datasets, are informative. The paper is suitable for publication with minor revisions that make the heuristic nature of the bridge to phylogenetics more explicit. I would not recommend major revision because the core theoretical contribution is sound and clearly presented; the gap is a limitation of scope rather than a correctness issue."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The core of this paper is a clean theoretical result about when maximum model-jump probability implies maximum mixing efficiency in trans-model MCMC. For two models where H₀ has no parameters, the authors show that when π₀ ≥ 1/2, maximum P_jump is both necessary and sufficient for maximum E, and a range of proposals achieve the optimum because the acceptance step thins the proposal to match the target. When π₀ < 1/2, maximum P_jump is necessary but not sufficient — you also need the entrance distribution to match the posterior, which requires either proposing from the posterior directly or adding a within-model move. The entrance/exit distribution framework connecting proposal design to mixing efficiency is genuinely new, and the uniform and Gaussian worked examples make the theory concrete. Theorem 1 (P_jump ≤ 2(1−max π_k)) is correct and the proof is straightforward. SI text 4 properly credits Frigessi et al. (1992) and re-derives the discrete-state optimality results cleanly. This is real theoretical work, not a repackaging. The phylogenetic empirical results are useful as a practical matter — direct transfer of branch lengths beats merge-split by a factor of 16-20, branch weights and parsimony weights help — but they are heuristics motivated by the theory, not derived from it. The paper is mostly honest about this: the Discussion explicitly says general cases with multiple parameters are 'yet to be studied.' The abstract's phrasing ('confirms the intuition,' 'guidelines') is measured enough. What does concern me is that the K > 2 extension in SI text 4 assumes parameters are proposed from the posterior during cross-model moves, which is exactly the ideal the paper argues for — so using it to justify the practical recommendations is somewhat circular. The two phylogenetic datasets are real but small in number, and the weight parameters (exponent −1/2, parsimony scale a = 0.5) are hand-tuned and acknowledged as arbitrary. No code or data is available — the Dryad link is a placeholder. For a paper with this much simulation, that's a real gap. Overall: the theory is sound and new, the applications are suggestive but not load-bearing for the theoretical claims. Worth a serious referee who can verify the derivations and push hard on reproducibility. I'd send it out for review.","headline":"Letter to colleague on trans-model MCMC mixing paper","tokens_in":46528,"tokens_out":1590,"would_cite":false,"duration_ms":100651,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Entrance distribution governs trans-model MCMC mixing efficiency","keywords":[],"falsifier":"If, in a two-model setting where the dominant model has posterior probability below one-half, one could exhibit a proposal density g(·) ≠ π₁(·) that achieves both maximum jump rate and maximum efficiency without any within-model move, the claimed necessity of the entrance-distribution matching condition would be refuted.","tokens_in":46031,"feed_emoji":"🔀","tokens_out":1056,"duration_ms":178564,"temperature":0.7,"pith_summary":"The paper studies when and why trans-model MCMC algorithms—the kind that jump between distinct statistical models, such as different phylogenetic trees—mix well or poorly. The central object is the entrance distribution: the distribution of parameter values immediately after the chain arrives in a new model. The authors prove that for two-model comparisons, maximum mixing efficiency is achieved if and only if this entrance distribution matches the within-model posterior of the new model. The analysis reveals an asymmetry. When the simpler model has posterior probability below one-half, achieving the maximum jump rate is necessary but not sufficient for optimal mixing: the proposal density must either already equal the posterior, or a within-model move must be added to restore the parameter to its stationary distribution. When the simpler model has posterior probability at or above one-half, the acceptance step itself thins the proposal to match the target, so maximum jump rate becomes both necessary and sufficient, and a range of proposal densities can achieve optimal efficiency without any within-model move. The paper also shows that for more than two models, the optimal chain does not maximize the jump rate—proposing too many model switches can hurt efficiency. These principles are tested on two real phylogenetic datasets, where direct transfer of branch lengths between trees and preferential perturbation of short internal branches substantially improve cross-tree mixing.","feed_headline":"Entrance distribution governs trans-model MCMC mixing efficiency","feed_subtitle":"Why some MCMC algorithms with high acceptance rates still mix poorly, and what to do about it","key_machinery":"The analysis proceeds through two exactly solvable examples (uniform and Gaussian target distributions) for the two-model case, where one model has no free parameters. The entrance distribution is defined and shown to be determined by the proposal density and the acceptance ratio. For the general two-model case, the parameter space of the parametric model is discretized into bins, reducing the problem to a finite-state Markov chain whose optimal transition matrix is known from prior work. Taking the discretization to the continuum yields the optimal transition kernel and the conditions on the proposal density. For more than two models, the theory of super-efficient reversible chains (via the","core_discovery":"The entrance distribution—the distribution of parameters immediately upon entering a new model—must match the within-model posterior for a trans-model MCMC algorithm to achieve maximum mixing efficiency. This single condition explains why some algorithms with high acceptance rates still mix poorly, and why adding within-model parameter updates sometimes helps and sometimes does not. The condition also implies that maximum model-jump probability is sufficient for optimal efficiency only when the dominant model has posterior probability at or above one-half; otherwise it is necessary but not sufficient.","pith_inferences":["If the entrance-distribution principle extends to cases where both models have free parameters, then cross-model proposals should propose all parameters of the new model jointly from an approximation to their conditional posterior, not independently or from the prior.","The result that the acceptance step can thin a non-matching proposal to match the target (when the dominant model has probability above one-half) suggests a connection to rejection sampling that could be exploited to design proposals with lower computational cost per accepted move.","For phylogenetic datasets with multiple local peaks in tree space, the entrance-distribution criterion would need to be satisfied within each peak, suggesting that locally informed proposals conditioned on the current tree neighborhood may be more effective than global proposals."],"forward_implications":["Algorithm designers should prioritize matching the entrance distribution to the within-model posterior when constructing cross-model proposals, rather than simply maximizing the acceptance rate.","In phylogenetic MCMC, direct transfer of branch lengths between trees is preferable to merge-and-split proposals because it produces values closer to the posterior mode, confirming the entrance-distribution principle in practice.","The finding that maximum jump rate is not always optimal for more than two models suggests that aggressive model-switching proposals in software with many candidate models may be counterproductive.","The deterioration of mixing as dataset size increases—because fixed proposals drift from an increasingly concentrated posterior—points to a need for adaptive proposals that track the posterior as data accumulate.","The distinction between within-model and cross-model moves implies that overall acceptance rate is a misleading performance metric for mixed algorithms; cross-model and within-model acceptance rates should be reported separately."],"fun_headline_variants":["Entrance distribution must match posterior for optimal trans-model MCMC mixing","High acceptance rates don't guarantee good mixing in trans-model MCMC","Optimal trans-model MCMC requires matching entrance distribution to posterior","Model-jump probability sufficiency depends on dominant model posterior weight","Entrance distribution explains poor mixing despite high acceptance in MCMC"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The theoretical results are derived for the case of exactly two models where one has no free parameters; the extension to the many-model, many-parameter setting of phylogenetics relies on an unproven analogy.","fun_headline_variants_meta":{"raw":{"variants":["Entrance distribution must match posterior for optimal trans-model MCMC mixing","High acceptance rates don't guarantee good mixing in trans-model MCMC","Optimal trans-model MCMC requires matching entrance distribution to posterior","Model-jump probability sufficiency depends on dominant model posterior weight","Entrance distribution explains poor mixing despite high acceptance in MCMC"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":531,"prompt_tokens":456,"completion_tokens":75,"prompt_tokens_details":null},"tokens_in":456,"tokens_out":75,"duration_ms":21919,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T17:50:36.660364+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If, in a two-model setting where the dominant model has posterior probability below one-half, one could exhibit a proposal density g(·) ≠ π₁(·) that achieves both maximum jump rate and maximum efficiency without any within-model move, the claimed necessity of the entrance-distribution matching condition would be refuted.","supporting_citations":[],"review_version":1}