{"id":"94b9d48c-c0fd-4aad-96cd-7da1bbd562eb","arxiv_id":"1908.08121","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Every b-Lipschitz Markov measure on a finite tree obeys a transportation-entropy inequality with constant Delta, giving a phase transition for normal Levy family behavior in terms of b and tree growth.","lead":"Broadcast models on trees, where information spreads from a root with possible errors, are shown to satisfy a sharp concentration inequality with a constant built from the tree's descendant structure. The result pins down when deep marginals become a normal Levy family, with a phase transition at a growth-rate threshold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the Proposition 16 application is outside the stated i.i.d. hypotheses, but the proof extends verbatim to non-identical product measures.","rationale":"I read the paper as proving a genuine concentration inequality for tree-indexed Markov measures. The main induction in Section 3 is coherent: the conditional expectation g is shown to be Hamming-Lipschitz with the correct descendant-generating weights via the recurrence for delta, and the final block is handled by a weighted McDiarmid bound. The reader's weakest assumption is the use of Proposition 16 for non-identical product measures. I checked the proof and it extends verbatim to arbitrary products, so the concern is an expositional gap rather than a load-bearing flaw. I also checked the growth-rate arguments in Theorem 3; the use of Proposition 11 and the operator-norm bounds are consistent, and the final lower bound on Delta_k^2/|V_k| is correct because the average of b^{d(u,v)} over u is at least b^{2k} for each v. The optimality arguments in Section 5 are sound under the standard (uniform-constant) reading of 'normal Levy family'; the paper's wording 'for each epsilon there exist constants' is slightly loose and would merit clarification, but it does not affect the central Theorem 1. The omitted proofs of Corollary 20 are explicitly acknowledged and do not support any main theorem. Overall, the reader's conditional verdict is appropriate: the paper is correct in substance, but a few statements need tightening before publication.","tokens_in":17724,"tokens_out":18178,"duration_ms":172754,"concrete_test":"Re-derive Proposition 16 with distinct marginals p_1,...,p_n: carry out the induction for the integral of e^{n lambda f} against tensor_i p_i using Hoeffding on the last marginal and the induction hypothesis for p_1,...,p_{n-1}. If the same bound e^{lambda^2 sum w(i)^2/8} follows, the Section 3 application to conditional products is justified. This is a short algebraic check that requires no new ideas.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The only concrete gap in the proof of Theorem 1 is that Proposition 16 is stated for p^n, while Section 3 applies it to the conditional product prod_{w in L_k} q_w(.|y_{pi(w)}), whose marginals are typically distinct and depend on y. This is a real mismatch between statement and use. However, it is not a correctness threat: the induction proof of Proposition 16 never uses identity of the marginals. The inductive step integrates out the last coordinate under p_n, applies Hoeffding's lemma to the residual function x -> n/w(n)(f(y,x)-g(y)) (whose 1-Lipschitz and mean-zero properties use only the current marginal), and applies the induction hypothesis to g/L under the first n-1 coordinates. Replacing p^n by tensor_i p_i changes no inequality. The same argument also handles y-dependent marginals because the relevant properties are checked pointwise in y. The other flagged omission, Corollary 20 in Section 6, concerns only comparison with prior work and is not used in the proofs of Theorem 1, Theorem 3, or Theorem 5.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves a concentration inequality for Markov measures indexed by finite trees. For a finite tree T with n vertices and a b-Lipschitz broadcast model ν on a Polish metric space H of diameter at most 1, Theorem 1 asserts that every 1-Lipschitz f : H^V → R with mean zero satisfies ∫ e^{nλ f} dν ≤ e^{λ²Δ²/8}, where Δ is defined explicitly from b and the tree's descendant generating function. By the Bobkov–Götze equivalence this also yields a transportation-entropy inequality. The paper then derives tail bounds, gives growth-rate criteria for sequences of depth-k marginals to form (normal) Lévy families (Theorem 3), proves a near-optimality lower bound for the Ising model (Theorem 4), and establishes a phase transition for normal Lévy behavior in the Ising case (Theorem 5). The proof of the main theorem is an induction on tree depth, using a weighted McDiarmid-type exponential moment bound (Proposition 16).","tokens_in":17933,"tokens_out":12253,"duration_ms":121134,"significance":"If correct, the result is a genuine and natural extension of Marton's transportation-entropy inequality and McDiarmid's inequality to tree-indexed Markov processes. The constant Δ is constructed explicitly from the tree and the contraction parameter b, with no hidden free parameters, and Theorem 3 gives concrete, falsifiable growth-rate criteria that match known reconstruction thresholds in the regular-tree case. Theorem 4 independently demonstrates near-optimality of the constant via an exact variance computation for the Ising model, and Theorem 5 shows that the qualitative phase transition is real rather than an artifact of the bound. The central induction is detailed and self-contained, and there is no circularity: the main results are derived from explicit recurrences, and the optimality statement is an independent computation. The main formal gap, concerning the scope of Proposition 16, is localized and repairable.","major_comments":[{"comment":"Proposition 16 is stated only for an i.i.d. product measure p^n, but in the induction step of Theorem 1 it is applied to the conditional product ∏_{w∈L_k} q_w(·|y_{π(w)}), whose marginals are generally distinct and depend on y. This is a genuine mismatch between statement and use, and because the exponential moment bound at each level of the induction depends on this application, the proof is incomplete as written. The gap is readily repairable: the induction proof of Proposition 16 never uses identity of the marginals, and replacing p^n by ⊗_{i=1}^n μ_i, with μ_i possibly depending on a conditioning variable, changes nothing provided the mean-zero and 1-Lipschitz properties are checked pointwise. The proposition should be restated in this generality and the application in Section 3 should cite it explicitly.","section":"Section 2.4 and Section 3 (Proof of Theorem 1)"}],"minor_comments":[{"comment":"The two displayed claims of Corollary 20 are nontrivial and are asserted without proof, with the text saying the proofs are omitted for brevity. Since Corollary 20 is used to compare Theorem 1 with the inequalities of Kontorovich–Ramanan and Chazottes et al., the proofs should be included or the claims should be moved to an appendix with full arguments.","section":"Section 6, Corollary 20"},{"comment":"There is a typo in the definition: 'postive constants' should read 'positive constants'.","section":"Section 1, definition of normal Lévy family"},{"comment":"The argument uses operator norms of Q on 𝓁²(V) for an infinite tree, but Lemma 9 defines Q on the vector space R^V. The proof should state explicitly that Q is viewed as a bounded operator on 𝓁²(V) under the bounded-degree assumption, or otherwise justify the norm computation.","section":"Section 4.2, proof of Theorem 3"},{"comment":"The statement should say explicitly that T is an infinite locally finite rooted tree, since the proof of the first part uses lim_{k→∞} |V_k| = ∞.","section":"Theorem 3 statement"}],"recommendation":"major_revision","confidential_remarks":"The Proposition 16 gap is real but localized; the needed generalization is straightforward and should not require a change in the main argument. I see no deeper correctness issue with the core proof, and I would expect the paper to be publishable after the proposition is restated and proved in the required generality and the supporting claims in Section 6 are supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper proves a real concentration theorem for Markov measures indexed by trees, not a repackaging of known results. Theorem 1 gives an exponential moment bound with a constant Delta built from descendant generating functions, and from it you get the equivalent transportation-entropy inequality and tail bounds. It recovers Marton's inequality for Markov chains and McDiarmid's inequality for product measures as special cases. That alone is a solid contribution to the concentration-of-measure literature. Second, the paper does something genuinely useful with the theorem: it derives conditions on the tree growth rate and the contraction parameter b under which depth-k marginals form a normal Levy family. For subperiodic trees, b^2 grT < 1 marks the phase transition for the growth of Delta_k, and the Ising model analysis (Theorems 4 and 5) shows the bound is close to sharp. The comparison with Kontorovich-Ramanan and Chazottes et al. in Section 6 is careful and fair, not a strawman.\n\nThe soft spot the reader flagged is real but not fatal. Proposition 16 is stated for i.i.d. product measures p^n, while the induction in Theorem 1 applies it to a product of distinct, y-dependent marginals q_w(.|y_{pi(w)}). I read the proof of Proposition 16; it never uses the fact that the marginals are identical. Hoeffding's lemma is applied to the residual function under a single marginal, and the induction hypothesis is applied to g/L under the first n-1 coordinates. Replacing p^n by a tensor product of distinct measures changes nothing, and the y-dependence is handled pointwise. So this is a genuine mismatch between statement and use, but a two-sentence remark would close it. The paper should say that the proposition extends verbatim to non-identical product measures; as written, it's a gap in presentation, not a hidden flaw in the central argument.\n\nTwo smaller things. Corollary 20 in Section 6 is stated without proof, but it is only used for comparison with prior work and is not load-bearing. And the phase-transition range for non-subperiodic trees between (maxgrT)^{-1/2} and (grT)^{-1/2} is left open; the author says so explicitly and shows only numerical evidence for the conjectured location. That is honest.\n\nThe citation pattern is fine, and the paper is self-contained. This is a paper for people working on concentration of measure, graphical models, or information flow on trees. It deserves a serious referee: the main theorem is new, the proof is mostly checkable, and the open points are clearly stated. I would send it out.","headline":"A genuine new concentration inequality for Markov measures on trees, with a proof that is largely sound and one explicitly fixable technical gap; worth sending to a serious referee.","tokens_in":18429,"tokens_out":2230,"would_cite":true,"duration_ms":23738,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","60J05","05C05","82B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single descendant-weighted sum controls concentration on tree-indexed Markov measures.","keywords":["concentration of measure","Markov measures on trees","broadcast models","Lévy families","transportation-entropy inequalities","descendant generating function","Ising model","phase transition"],"falsifier":"Take a depth-2 tree with a root and two children, choose two distinct $b$-Lipschitz binary kernels with the same Lipschitz constant, and compute $\\int e^{n\\lambda f}\\,d\\nu$ for a 1-Lipschitz function such as the normalized density of ones. If this ever exceeds $e^{\\lambda^2\\Delta^2/8}$, Theorem 1 is false. Equivalently, one could search for a counterexample to the weighted bounded-differences bound for products of non-identical marginals, which would break the induction even if the final inequality happens to hold.","tokens_in":17533,"feed_emoji":"🌳","tokens_out":8006,"duration_ms":80099,"temperature":0.7,"pith_summary":"A broadcast model on a tree passes a value from each node to its children through random kernels, and this paper asks how strongly such a process concentrates: how close Lipschitz functions of the whole configuration must be to their mean. The paper proves that concentration is governed by one number, built from the tree's descendant counts and the kernels' Lipschitz constant, and that the resulting bound recovers the classical product and chain inequalities as special cases. This yields a concrete condition, in terms of the tree's growth rate and the kernels' contraction, for a sequence of depth-k marginals to form a normal Lévy family. The proof is an induction over the depth of the tree, with a descendant generating function supplying the weights at each level.","feed_headline":"Tree shape sets when broadcast noise concentrates","feed_subtitle":"A single descendant-count function controls concentration of Lipschitz functions on tree-indexed Markov measures.","key_machinery":"The central object is the descendant generating function $\\delta(v)=\\sum_{r\\ge0}|D_r(v)|b^r$, which satisfies the recurrence $\\delta(v)=1+b\\sum_{w:\\pi(w)=v}\\delta(w)$ and whose $\\ell^2$ norm is $\\Delta$. The proof inducts on the depth of the tree, using $\\delta$ to assign weights to the leaves of each level so that the conditional product of the one-step transition kernels becomes a weighted Hamming space; a standard bounded-differences lemma then bounds the exponential moment at each level. A classical equivalence between transportation-entropy inequalities and exponential moment bounds converts the result into the transport form, and estimates comparing sums of $b^{d(v,w)}$ with $\\Delta^2$ link the bound to the tree's growth rates.","core_discovery":"For any finite tree $T$ with $n$ vertices and any $b$-Lipschitz Markov measure $\\nu$ indexed by $T$, the paper proves that every 1-Lipschitz function $f$ with mean zero satisfies the exponential moment bound\n$$\\int $e^{{n\\lambda f}}$\\,d\\nu \\le e^{\\$lambda^{2}$ \\$\\Delta$^2 / 8},$$\nwhere $\\Delta$ is the $\\ell^2$ norm of the descendant generating function\n$$\\delta(v) = \\sum_{r\\ge 0} |D_r(v)| b^r,$$\nwith $D_r(v)$ the set of descendants of $v$ at distance $r$. Equivalently, the transportation-entropy inequality\n$$\\bar d(\\mu,\\nu) \\le \\frac{\\$\\Delta$}{n} \\sqrt{\\frac{1}{2}D(\\mu\\|\\nu)}$$\nholds for all probability measures $\\mu$. From this the paper derives tail bounds of the form $2e^{-2n^2\\varepsilon^2/\\Delta^2}$, recovering McDiarmid's inequality when $b=0$ and Marton's inequality when the tree is a path. It then uses estimates on $\\Delta$ to locate a phase transition: for subperiodic trees, a sequence of depth-$k$ marginals is a normal Lévy family exactly when $b^2 \\mathrm{gr}\\,T < 1$, and for the Ising model this threshold is shown to be intrinsic rather than an artifact of the method.","pith_inferences":["The same induction may extend to Bayesian networks with multiple sources and to concentration of the marginal on the leaves, as the paper itself suggests.","A direct testable extension is to compute the quantity $G(T)$ defined in Section 1.1.1 for non-subperiodic trees; if $G(T)=\\max\\mathrm{gr}\\,T$ for such trees, the phase transition threshold $b^2\\max\\mathrm{gr}\\,T=1$ would be universal.","Because the transportation-entropy form carries explicit constants, it could be converted into modified logarithmic Sobolev inequalities for tree-indexed Gibbs measures, giving a route to functional inequalities not explored in the paper."],"forward_implications":["Any $b$-Lipschitz broadcast model on a finite tree satisfies the tail bound $\\nu\\{|f-\\int f\\,d\\nu|>\\varepsilon\\}\\le 2e^{-2n^2\\varepsilon^2/\\Delta^2}$, so Lipschitz observables concentrate at speed $n/\\Delta$.","For a sequence of finite trees, $\\Delta_k=o(|V_k|)$ makes the measures a Lévy family and $\\Delta_k=O(\\sqrt{|V_k|})$ makes them a normal Lévy family; bounded-degree infinite trees enter the latter regime when $b^2\\max\\mathrm{gr}\\,T<1$ and leave it when $b^2\\,\\mathrm{gr}\\,T>1$.","For the Ising broadcast model with flip probability $p$, where $b=1-2p$, the depth-$k$ marginals fail to form a normal Lévy family exactly when $\\Delta_k$ is not $O(\\sqrt{|V_k|})$, so the phase transition in concentration quality is genuine.","The universal constant in the theorem cannot be smaller than $\\Delta\\sqrt{(1-b^2)/2}$, so the bound is close to optimal in the worst case."],"supporting_citations":[{"why":"Supplies the Markov-chain transportation-entropy inequality that Theorem 1 generalizes to trees.","marker":"[15]"},{"why":"Provides the bounded-differences lemma and the weighted McDiarmid-type bound used in the induction.","marker":"[16]"},{"why":"Establishes the equivalence between transportation-entropy inequalities and exponential moment bounds used to state and apply Theorem 1.","marker":"[2]"},{"why":"Defines the branching factor and the descendant-counting operator $Q$ used in the estimates of $\\Delta$.","marker":"[13]"},{"why":"Supplies the notions of growth rate and subperiodicity used in Theorem 3 and Proposition 7.","marker":"[14]"}],"fun_headline_variants":["Tree branch sums decide when noise concentrates","One descendant count tames tree broadcast noise","Concentration on trees: growth rate is the switch","Broadcast noise on trees: a single number rules","Tree shape gates concentration of Markov measures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The induction step applies a weighted bounded-differences inequality to a product of one-step transition kernels whose marginals are not identically distributed and depend on the conditioning parent value, while the lemma as stated covers only identical product measures; no separate proof of that extension is provided.","fun_headline_variants_meta":{"raw":{"variants":["Tree branch sums decide when noise concentrates","One descendant count tames tree broadcast noise","Concentration on trees: growth rate is the switch","Broadcast noise on trees: a single number rules","Tree shape gates concentration of Markov measures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1517,"prompt_tokens":908,"completion_tokens":609,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":541}},"tokens_in":524,"tokens_out":609,"duration_ms":6960,"temperature":1.0,"reasoning_tokens":541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:48:47.434338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a depth-2 tree with a root and two children, choose two distinct $b$-Lipschitz binary kernels with the same Lipschitz constant, and compute $\\int e^{n\\lambda f}\\,d\\nu$ for a 1-Lipschitz function such as the normalized density of ones. If this ever exceeds $e^{\\lambda^2\\Delta^2/8}$, Theorem 1 is false. Equivalently, one could search for a counterexample to the weighted bounded-differences bound for products of non-identical marginals, which would break the induction even if the final inequality happens to hold.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Markov-chain transportation-entropy inequality that Theorem 1 generalizes to trees."},{"cited_title":"Concentration","cited_arxiv_id":null,"evidence_quote":"Provides the bounded-differences lemma and the weighted McDiarmid-type bound used in the induction."},{"cited_title":"Exponential Integrability and Transportation Cost Related to Logarithmic Sobolev Inequalities","cited_arxiv_id":null,"evidence_quote":"Establishes the equivalence between transportation-entropy inequalities and exponential moment bounds used to state and apply Theorem 1."},{"cited_title":"Random Walks and Percolation on Trees","cited_arxiv_id":null,"evidence_quote":"Defines the branching factor and the descendant-counting operator $Q$ used in the estimates of $\\Delta$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the notions of growth rate and subperiodicity used in Theorem 3 and Proposition 7."}],"review_version":1}