{"id":"aaffab6d-482b-4b40-aafe-16ef76541079","arxiv_id":"2411.10271","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For Z_q clock models on regular trees, the A-localized Gibbs state at strong coupling is non-extremal and decomposes into uncountably many glassy states.","lead":"This paper proves that the low-temperature localized Gibbs states of clock models on regular trees are not pure: each one decomposes into uncountably many spatially inhomogeneous glassy states. The result gives a new example of how tree-structured spin systems can harbor a continuous family of extremal phases.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 4's propagation constant can exceed u/d, so the Peierls bound in Lemma 1(b) is unproved for potentials with u<1 and U<du; Theorem 2 lacks a proof in that regime.","rationale":"The paper's central claim is that low-temperature A-localized states decompose into uncountably many extremal states. The proof is a chain: Proposition 1 → Lemma 1 → Theorems 3-6 → Theorem 2. Lemma 1 is the key technical estimate: it controls the bad events B_A(γ) via an exponential Peierls bound. Lemma 4 is the engine inside Lemma 1. I checked the derivation of Lemma 4's propagation step. The constants do not match the exponents produced by the induction. In particular, the term C1 e^{-(d-2+1/d)βu + t} in (3.23) forces the prefactor P to contain e^{-βu/2 + t} for d=2. The stated c = 1/d = 1/2 is not small enough to dominate this when u < 1. The example d=2, u=1/2, U=3/4 is allowed by (2.2) and makes the required inequality fail. This is a proof gap, not a contradiction of the theorem: choosing c = min{ (d^2+1)/d u - U, u/d } would likely repair it, or the theorem could be restricted to u ≥ 1. Because the gap affects the central lemma, the paper is not yet fully rigorous for the full parameter range claimed. The reader's weakest assumption was Theorem 1; my concern is different and internal, so I marked disagreement. The reader's other noted issues (|A|=2 in Theorem 4, 'small enough' in Lemma 5) are also present and also support a conditional verdict. On balance, the idea is promising and the fix seems local, so I do not recommend rejection; the verdict should remain conditional, pending correction of the constants.","tokens_in":31277,"tokens_out":34311,"duration_ms":269134,"concrete_test":"Analytically verify the induction inequality for Lemma 4 with d=2, u=1/2, U=3/4. Compute P from the combination of (3.23) and (3.24): P-1 ≥ C1 e^{-β/4 + t}. Compare with f(β)e^t = (2C3+C1)e^{-β/2 + t}; the ratio is (C1/(2C3+C1)) e^{β/4}, which diverges, so P ≤ f e^t + 1 fails for large β. If this holds, Lemma 4's stated constants are wrong and Lemma 1(b) needs either a smaller c or an added hypothesis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main theorem rests on Lemma 1, whose key step is Lemma 4. In Lemma 4 the constant is c = min{ (d^2+1)/d u - U, 1/d } and f(β) = (2C3+C1)e^{-cβ}. In the induction step, equations (3.23)-(3.24) combine to give a prefactor P with P-1 containing the term C1 e^{-(d-2+1/d)βu + t} (and for d=2 also C3 e^{-βu/d + t}). To close the induction one needs P ≤ f(β)e^t + 1, which forces c ≤ min{ (d-2+1/d)u, u/d, (d+1/d)u - U }. For d=2 the relevant minimum is u/2. The stated c can be larger: take d=2, u=1/2, U=3/4, which satisfies (2.2) (5u = 2.5 > 2U = 1.5). Then c = min{ (5/2)(1/2)-3/4, 1/2 } = min{0.5,0.5} = 0.5, while u/2 = 0.25. Hence f(β) decays like e^{-0.5β}, faster than the e^{-0.25β + t} term in P, so P ≤ f(β)e^t + 1 fails for large β. The induction step is therefore invalid for a range of allowed potentials, and Lemma 1(b) is not established there. This flaw is internal and independent of Theorem 1. A smaller c (e.g., c = u/d) would repair the argument, but as written Theorem 2 is not proven for all models satisfying (2.2).","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Z_q-valued ferromagnetic clock models on d-regular trees with nearest-neighbor potentials satisfying the u,U,d-bounds (2.2). Building on the construction of homogeneous tree-indexed Markov chain Gibbs states μ_A localized on a subset A⊂Z_q from [1], it claims that for |A|≥2 and sufficiently large inverse temperature, μ_A is not extremal and its extremal decomposition measure is atomless, supported on uncountably many spatially inhomogeneous pure states. The proof introduces a new notion of A-irregular edges and a coarse-grained state space, proving a Peierls-type bound with errors (Lemma 1), then derives restricted single-site reconstruction (Theorem 3), non-extremality via the Edwards-Anderson parameter (Theorem 4), multi-site reconstruction on thinned branches (Theorem 5), and almost-sure singularity of the π-kernels (Theorem 6), which yields the atomless decomposition.","tokens_in":31586,"tokens_out":12213,"duration_ms":113454,"significance":"If the proof is correct, this is a valuable extension of the recent atomless decomposition results for free states of Potts and clock models to the non-symmetric A-localized states, including the Potts model as a special case. The proposed A-irregular edge decomposition and the local coarse-graining method are genuinely new technical tools that may be useful in other tree models. The paper also makes the interesting observation that reconstruction holds only for spin values in the localization set A, not for all states. The explicit exponential bounds and the clear statement of the u,U,d-regime are strengths. However, the current version contains a load-bearing technical gap in the main Peierls estimate and a separate gap in the proof of non-extremality for |A|=2, both of which need to be repaired before the main claims are established.","major_comments":[{"comment":"The constant c in Lemma 4 (and in the definition of λ(β) in Lemma 1) is specified as c = min{ (d^2+1)/d · u − U, 1/d }. However, the induction step combining equations (3.23) and (3.24) requires c to be no larger than each of (d^2+1)/d · u − U, (d−2+1/d)·u, and u/d. The stated value can violate the last two requirements. For example, with d=2, u=1/2, U=3/4, which satisfies (2.2) because 5u = 2.5 > 2U = 1.5, the stated c equals min{0.5, 0.5} = 0.5, while u/d = 0.25 and (d−2+1/d)u = 0.25. Consequently f(β) = (2C3+C1)e^{−0.5β} decays faster than the terms C1 e^{−0.25β+t} and C3 e^{−0.25β+t} appearing in the prefactor, and the closing inequality P ≤ f(β)e^t + 1 cannot hold for large β. Thus the induction in Lemma 4, and hence the Peierls-type bound (3.4) of Lemma 1(b), is not proved for all potentials satisfying (2.2). This is load-bearing for Theorems 2, 3, 5 and 6. A likely repair is to replace the second entry 1/d by u/d and to re-check the constant c̃ in Lemma 1, but as written the proof has a real gap.","section":"§3.2, Lemma 4 and Lemma 1(b)"},{"comment":"The proof of Theorem 4 chooses β large enough such that (1 − C2 e^{−βu})^{-1} < |A| − 1. For |A| = 2 this condition is impossible, because the left-hand side is strictly larger than 1 while |A|−1 = 1. Since Theorem 2 explicitly includes |A| = 2, the argument as written does not cover this case. The final bound (4.3) may still be true, for instance by using the spin-space symmetry that the authors state holds for |A|=2, but that argument is not given. This gap directly affects the proof of part i) of Theorem 2.","section":"§4.1, proof of Theorem 4"},{"comment":"The multi-site reconstruction theorem (Theorem 5) and the almost-sure singularity theorem (Theorem 6) rely on Lemmas 6, 7 and 8, but the proofs of these lemmas are only sketched and refer to [8] for the main ideas. In particular, Lemma 8's covariance bound mixes the exponential decay e^{−c1 λ(β)} with the Perron-Frobenius decay |λ2(PA)|^{c2/6}. In the low-temperature A-localized regime the second eigenvalue of PA may be close to 1, so the resulting rate may not be small uniformly in β. Since these lemmas are essential for the atomless-decomposition claim, the authors should provide a complete and self-contained proof, or at least a precise adaptation of [8] with explicit estimates for the A-localized case.","section":"§4.2, Lemmas 6–8"}],"minor_comments":[{"comment":"The statement 'If β is small enough, h possesses a unique minimizer' should read 'If β is large enough'. The minimizer formula (3.26) gives t* ≥ 0 only when f(β) is sufficiently small, which happens for large β, not small β.","section":"§3.3, Lemma 5"},{"comment":"The statement 'For small enough β' should read 'For large enough β', since the covariance decay uses λ(β) → ∞ as β → ∞.","section":"§4.2.1, Lemma 8"},{"comment":"The expression c := min{ (d^2+1)/d · u − U, 1/d } is dimensionally inconsistent: the second argument should presumably be u/d (an energy), not 1/d. This is closely related to Major Comment 1 and should be corrected in both Lemma 1 and Lemma 4.","section":"§3.2, definition of c in Lemma 4 and Lemma 1"},{"comment":"The sentence 'Choosing β large enough such that (1−C2e^{−βu})^{-1} < (|A|−1)' is also problematic for |A|=2, as noted in Major Comment 2; even for |A|≥3 the proof should explain how this condition is compatible with (iv) of Proposition 1.","section":"§4.1, proof of Theorem 4"},{"comment":"The notation 1/d in the definition of c and the phrase 'small enough β' in two lemmas suggest typographical errors that should be fixed during revision; these typos do not affect the overall strategy but are confusing to the reader.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper depends heavily on the authors' own recent work [1] (published) and on the preprint [8] (to appear). The refereeing of this manuscript would be facilitated by having the full proof of Lemma 8 or a very explicit adaptation of [8]. The main technical issue, however, is internal to this paper: the constant c in the propagation lemma appears to be mis-specified, and the proof of non-extremality fails for |A|=2 as written. Both issues seem fixable without changing the overall method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: the paper takes the free-state extremal decomposition machinery from [8] and makes it work for A-localized clock states, including Potts A-localized states, where previously only existence and non-convex-combination results were known. The A-irregular edges and the A/Ac coarse-graining are genuine technical additions, and they let the authors prove the Peierls-type bound (Lemma 1) that the whole argument hangs on. Sections 3 and 4.1 are written carefully, and the restricted single-site reconstruction plus the almost-sure singularity of the π-kernels are valuable results in their own right. I did not find a circularity problem: the paper leans on [1] and [8], but those are external results with independent proofs, and the theorem here is not baked into the inputs.\n\nThe soft spots are real but minor. The proof of Theorem 4 uses the condition (1 - C2 e^{-βu})^{-1} < |A|-1, which is impossible for |A|=2; the theorem includes |A|=2, so this case needs a slightly sharper bound on the single-site marginal or a separate argument. Lemma 5 says beta 'small enough' where the argument needs beta large; that is a typo-level fix. Theorem 2 states A⊂Zq with |A|≥2 but Lemma 1 and all the reconstruction theorems require A⊊Zq; the statement should either exclude A=Zq or point out that the A=Zq case is covered by [8].\n\nAbout the stress-test concern: I checked the exponent in Lemma 4 carefully. The second rate in c is u/d, not 1/d — the induction step splits e^{-βu} uniformly over the d children, which is exactly why the u/d rate appears in Remark 2 and in the g(1) term. So the counterexample d=2, u=1/2, U=3/4 gives c = min{0.5, 0.25} = 0.25 and the induction closes. If the published formula literally typesets 1/d, that is a typo that should be fixed, but as a mathematical objection the stress-test does not land.\n\nThe reader's CONDITIONAL verdict is about right. The main theorem is very likely correct, the proof strategy is sound, and the issues I found do not threaten the central argument. This paper deserves a serious referee: it extends known results in a nontrivial way and introduces tools that will be used for other localized states. My recommendation is to send it to review, with the referee asked to verify the |A|=2 case in Theorem 4 and to clean up the statement typos. I would cite it if I work on tree Gibbs measures, and I would bring it to a reading group on disordered systems or tree-indexed Markov chains.","headline":"Solid low-temperature extension of the free-state glassy decomposition to A-localized clock states; the flaws are fixable, and the Peierls-bound worry about Lemma 4 comes from a misread exponent.","tokens_in":32174,"tokens_out":5917,"would_cite":true,"duration_ms":56868,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","82B20","82B44"],"pacs":[],"model":"deepseek-v4-flash","headline":"At sufficiently low temperature, the A-localized Gibbs states of $\\mathbb{Z}_q$-clock models on regular trees are not pure: their extremal decomposition is atomless and supported on uncountably many inhomogeneous pure states.","keywords":["Gibbs measures on trees","tree-indexed Markov chains","clock models","extremal decomposition","localization","reconstruction on trees","glassy states","Potts model"],"falsifier":"Direct numerical test on the Potts model ($q \\geq 3$) on the binary tree, where explicit formulas exist: sample boundary conditions $\\omega, \\omega'$ independently from $\\mu_A$ for a two-point localization set $A$ at large $\\beta$, and estimate the distribution of the thinned branch overlap $\\varphi^{\\omega}$ under the conditional kernels $\\pi(\\cdot|\\omega)$ and $\\pi(\\cdot|\\omega')$. If the two distributions fail to separate — that is, if $\\pi(\\varphi^{\\omega}|\\omega)$ does not remain close to $1$ while $\\pi(\\varphi^{\\omega}|\\omega')$ stays below $\\sum_a \\pi_A(a)^2$ for most pairs — the multi-site reconstruction of Theorem 5 fails. A cheaper check: directly estimate $\\mu_A(B_A(\\gamma))$ for contours of increasing size at couplings near the threshold $(d^2+1)u = dU$; the bound (3.4) predicts exponential decay in $|\\gamma|$ with rate $\\lambda(\\beta)$, and its failure would break the Peierls step.","tokens_in":30999,"feed_emoji":"🌳","tokens_out":12299,"duration_ms":102636,"temperature":0.7,"pith_summary":"This paper establishes that, at sufficiently low temperature, the A-localized Gibbs states of $\\mathbb{Z}_q$-valued clock models on regular trees are not extremal, even though their single-site marginals look sharply ordered: the mass concentrates on a finite subset $A \\subset \\mathbb{Z}_q$ of spin values. The central claim is that for any localization set with $|A| \\geq 2$, such a state $\\mu_A$ is a genuine convex mixture of uncountably many spatially inhomogeneous pure states, with an extremal decomposition measure that has no atoms. This matters because it shows tree geometry makes ordered states thermodynamically unstable in a strong sense: low-temperature pure phases are hidden, boundary-condition-dependent states, so a measurement of $\\mu_A$ represents uncertainty over a continuous family of phases rather than a single equilibrium state. The same conclusion is new even for the Potts model, where previously only partial extremality statements were available for A-localized states. The proof works for all ferromagnetic nearest-neighbor clock potentials satisfying a mild cost bound (the $u, U, d$-condition), with no FKG or symmetry assumptions.","feed_headline":"Low-temperature clock states split into uncountably many pure states","feed_subtitle":"On regular trees, a state pinned to a spin subset still mixes a continuous family of glassy phases.","key_machinery":"The argument runs on a good site/bad site decomposition adapted to A-localization. Edges where the configuration changes spin value or touches $A^c$ are called A-irregular; a bad event is the presence of a contour whose boundary has too high a density of A-irregular edges. Lemma 1 is the load-bearing estimate: for $\\mu_A$-almost every boundary condition $\\omega$, the probability that the $\\pi$-kernel at the root disagrees with $\\omega$ is bounded by the indicator of the bad event plus a small error $\\epsilon_1(\\beta)$, while bad events themselves decay exponentially in contour size with rate $\\lambda(\\beta) \\to \\infty$. To prove this, the model is coarse-grained to a two-letter state space $\\{A, A^c\\}$, which destroys the Markov chain property except in the Potts case, and the coarse-grained sums are bounded by a non-stochastic matrix $M$ whose entries absorb the transition probabilities together with exponential weights; an induction called propagation of smallness over subgraphs splits the $e^{-\\beta u}$ factors among the $d$ children of each vertex and yields the exponential decay. The reconstruction consequences then follow the branch-overlap route: single-site reconstruction gives a positive Edwards-Anderson parameter, and multi-site reconstruction on suitably thinned branches shows the thinned branch-overlap $\\varphi^{\\omega}$ concentrates near $1$ under $\\pi(\\cdot|\\omega)$ but near $\\sum_a \\pi_A(a)^2$ under $\\pi(\\cdot|\\omega')$, so typical kernels are mutually singular almost surely.","core_discovery":"On the paper's own terms, the discovery is Theorem 2: for a $\\mathbb{Z}_q$-valued nearest-neighbor clock model on a $d$-regular tree whose potential satisfies the $u, U, d$-bounds, for any $A \\subset \\mathbb{Z}_q$ with $|A| \\geq 2$ and all $\\beta$ large enough, the A-localized Gibbs state $\\mu_A$ is not extremal, and its extremal decomposition measure $\\alpha_{\\mu_A}$ is supported on uncountably many inhomogeneous states, with $\\alpha_{\\mu_A}(\\{\\nu\\}) = 0$ for every extremal $\\nu$. Equivalently, $\\mu_A = \\int \\nu\\, \\alpha_{\\mu_A}(d\\nu)$ over pure states, and the mixing measure is continuous. This is obtained through a restricted reconstruction statement: only spin values $a \\in A$ can be recovered from the boundary at infinity (Theorem 3), which forces the Edwards-Anderson parameter to be strictly positive (Theorem 4); then a multi-site reconstruction on thinned branches (Theorem 5) shows that typical pairs of boundary conditions produce mutually singular pure states (Theorem 6), which is exactly the atomless property of the decomposition. The extremal states are described as low-temperature perturbations around mostly flat almost ground states with spin values in $A$.","pith_inferences":["Editorial inference: by analogy with the Potts case in reference [27], $\\mu_A$ may be extremal in an intermediate temperature window below the uniqueness threshold, with the atomless glassy decomposition appearing only at low temperature; the authors leave this open.","Editorial inference: the one-step-jump condition $(d^2+1)u > dU$ could likely be relaxed to multi-step returns from $A^c$ into $A$, which would extend the result to models with slower relaxation in spin space, such as p-SOS-type models.","Editorial inference: a testable quantitative prediction is that the information loss rate for initial spins in $A^c$ is governed by the second eigenvalue of the transition matrix restricted to $A$, while recovery of spins in $A$ is governed by the exponentially small bad-event probability."],"forward_implications":["If the main theorem is correct, every A-localized state with $|A| \\geq 2$ at low temperature is a mixture of uncountably many inhomogeneous pure states, so no single boundary condition at infinity can account for its local statistics.","The restricted reconstruction bound (Theorem 3) quantifies the information flow: initial spins in $A$ stay recoverable at infinity with probability close to 1, while spins outside $A$ are lost, a sharp asymmetry inside one Gibbs state.","The Edwards-Anderson parameter is strictly positive at low temperature (Theorem 4), giving a concrete order parameter for the non-extremality of $\\mu_A$.","Typical extremal measures are almost surely mutually singular (Theorem 6), so the decomposition is genuinely continuous rather than a countable sum of pure states.","The result is new already for Potts-model A-localized states, where previously only partial extremality statements were known."],"supporting_citations":[{"why":"Supplies Theorem 1, the existence of the homogeneous tree-indexed Markov chain states $\\mu_A$ with the quantitative localization bounds (2.9); the entire proof starts from these bounds.","marker":"[1]"},{"why":"Provides the branch-overlap and thinned-branch reconstruction strategy for proving atomless decompositions of free states, which Section 4.2 adapts to A-localized states.","marker":"[8]"},{"why":"Gives the general theory used throughout: $\\pi$-kernels, extremal decomposition into pure states, and the tail-triviality criterion for extremality.","marker":"[22]"},{"why":"The Potts-model benchmark where A-localized states were first constructed via explicit fuzzy-transformation computations, with partial extremality statements that the present paper strengthens.","marker":"[27]"},{"why":"Establishes the first atomless extremal decomposition, for the free Ising state on a tree, the phenomenon this paper extends to A-localized clock states.","marker":"[19]"},{"why":"Introduces the contour formalism relative to a fixed reference configuration that underlies the Peierls-type estimates with errors in Lemma 1.","marker":"[9]"},{"why":"Supplies the Perron-Frobenius convergence of the transition matrix used in Lemma 8 to control covariance decay along thinned branches.","marker":"[5]"}],"fun_headline_variants":["Clock model states on trees shatter into a continuum of pure phases","A-localized Gibbs states are never pure: uncountably many phases","Tree clock states decompose into atomless glassy mixtures","Each pinned clock state hides a continuum of extremal spin phases","Spin subset states on trees: uncountably many pure states inside"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes as its starting point that the A-localized state from reference [1] exists with the stated quantitative bounds — single-site mass essentially all on $A$, transitions out of $A$ exponentially rare at low temperature — and builds every subsequent estimate on those bounds; if those bounds failed or were unavailable for a given clock model, the atomless decomposition result would not follow from this paper's proof.","fun_headline_variants_meta":{"raw":{"variants":["Clock model states on trees shatter into a continuum of pure phases","A-localized Gibbs states are never pure: uncountably many phases","Tree clock states decompose into atomless glassy mixtures","Each pinned clock state hides a continuum of extremal spin phases","Spin subset states on trees: uncountably many pure states inside"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1567,"prompt_tokens":1063,"completion_tokens":504,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":679,"tokens_out":504,"duration_ms":5026,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:47:42.056454+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Direct numerical test on the Potts model ($q \\geq 3$) on the binary tree, where explicit formulas exist: sample boundary conditions $\\omega, \\omega'$ independently from $\\mu_A$ for a two-point localization set $A$ at large $\\beta$, and estimate the distribution of the thinned branch overlap $\\varphi^{\\omega}$ under the conditional kernels $\\pi(\\cdot|\\omega)$ and $\\pi(\\cdot|\\omega')$. If the two distributions fail to separate — that is, if $\\pi(\\varphi^{\\omega}|\\omega)$ does not remain close to $1$ while $\\pi(\\varphi^{\\omega}|\\omega')$ stays below $\\sum_a \\pi_A(a)^2$ for most pairs — the multi-site reconstruction of Theorem 5 fails. A cheaper check: directly estimate $\\mu_A(B_A(\\gamma))$ for contours of increasing size at couplings near the threshold $(d^2+1)u = dU$; the bound (3.4) predicts exponential decay in $|\\gamma|$ with rate $\\lambda(\\beta)$, and its failure would break the Peierls step.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Theorem 1, the existence of the homogeneous tree-indexed Markov chain states $\\mu_A$ with the quantitative localization bounds (2.9); the entire proof starts from these bounds."},{"cited_title":"Random Struct","cited_arxiv_id":null,"evidence_quote":"The Potts-model benchmark where A-localized states were first constructed via explicit fuzzy-transformation computations, with partial extremality statements that the present paper strengthens."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the first atomless extremal decomposition, for the free Ising state on a tree, the phenomenon this paper extends to A-localized clock states."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the contour formalism relative to a fixed reference configuration that underlies the Peierls-type estimates with errors in Lemma 1."}],"review_version":1}