{"id":"47b84396-0b6e-44a8-8302-1900e8b99ba9","arxiv_id":"2608.03687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For 1-Lipschitz multi-spin interactions with support at most r on a p-biased Boolean cube with rp ≥ 12, E exp(Σλ_iφ_i) ≠ 0 whenever every coordinate's total influence is at most 1/(10√(rp)).","lead":"This paper proves that a many-spin partition function on the Boolean cube stays free of complex zeros when the total influence of every coordinate is below 1/(10√(rp)), with r the interaction size and p the spin bias. The result implies that a linear growth in multi-spin interaction energy can be offset by only a logarithmic growth of the external field, and it yields a quasi-polynomial algorithm for approximating such partition functions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Induction step in §4.3 relies on a false claim that Ψ(t) is affine in each coordinate; the branch-of-log construction is invalid as written.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption identified as Theorem 3.1 Part (1). I agree that the proof of Theorem 3.1 is the source of the 1/√(rp) scaling, but a more immediate flaw appears in the induction step of Theorem 1.2. The paper claims in Section 4.3 that Ψ(t) is affine in each coordinate τ_j, and uses this to conclude that the argument change along an edge is less than π, which justifies the existence of a branch ψ with the Lipschitz bound (4.3.2). This claim is false: for fixed y, φ_i(y|t) is affine in τ_j, so the exponent is affine, but the expectation over y of exp(affine) is a sum of exponentials, not affine. The explicit counterexample with one coordinate shows the assertion is untrue. Without a valid branch construction, the bound on the Lipschitz constant of ψ is unsupported, and the induction step cannot be completed. I also noticed the minor padding issue: Theorem 3.1 requires pn ≥ 12, but in the induction one often works on a subcube of dimension |J| < r, so the hypothesis may fail unless the functions are extended to r coordinates; this too is fixable but not stated. The main theorem is likely true: the per-edge log-ratio has magnitude < π, and the hypercube cycles are generated by 4-cycles, so a consistent branch can be defined by integrating edge log-ratios. However, the paper's stated proof is incorrect at a load-bearing point. Therefore the verdict should remain CONDITIONAL, requiring the branch construction to be replaced or repaired.","tokens_in":17385,"tokens_out":22034,"duration_ms":202431,"concrete_test":"Compute Ψ(τ) for the one-coordinate subcube J={1}, p=q=1/2, λ=1, φ(ξ)=ξ: Ψ(τ)=E exp{λ φ(·|τ)} = (e^τ+1)/2, which is not affine in τ. This directly falsifies the affine claim in §4.3. If the claim fails, the paper's branch-of-log argument does not prove (4.3.2), and the induction step needs the alternative cycle-consistency argument to be supplied.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section 4.3, to construct ψ on the subcube {0,1}^J, the paper extends Ψ to the solid cube [0,1]^J and asserts that in each coordinate τ_j the function Ψ(t) is affine (ατ_j+β), so the argument changes by less than π and the chosen branch satisfies the desired Lipschitz bound (4.3.2). This assertion is false. Although each φ_i(·|t) is affine in τ_j when the other coordinates are fixed, the exponent Σ λ_i φ_i(·|t) is affine, but the expectation over the inner cube is a sum of exponentials: Ψ(τ_j) = Σ_y P(y) exp{A_y τ_j + B_y}, which is not affine unless all A_y = 0. For example, with J={1}, p=q=1/2, λ=1, and φ(ξ)=ξ, we get Ψ(τ)= (e^τ+1)/2, not affine. Consequently, the path τ↦Ψ(τ) is not a line segment and its argument may vary by more than π; the claimed bound |ψ(x)-ψ(y)| ≤ 2Σ|λ_i| for points differing in one coordinate is not established by the given argument. This is load-bearing because (4.3.2) provides the Lipschitz constant for λ_m φ_τ + ψ, which is needed for the final application of Theorem 3.1. The gap is repairable: for each edge the log-ratio has a representative with |α| ≤ 2Σ|λ_i| < π, and since the hypercube's cycle space is generated by 4-cycles while any 4-cycle has total |α| sum at most 8/(10√(rp)) < 2π, these representatives are consistent on all cycles and define ψ. But this repair is absent; the paper's stated reason is incorrect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves zero-freeness of the partition-function expectation E exp{Σ λ_i φ_i} on the p-biased Boolean cube under per-coordinate influence bounds. Theorem 1.2 gives the bound Σ_{i: j∈J_i}|λ_i| ≤ 1/(10√(rp)) when each φ_i is 1-Lipschitz, depends on at most r coordinates, and rp ≥ 12. Theorem 1.3 gives an analogous linear-in-L bound when p = e^{-6L}/r. The proofs are built on Theorem 3.1, a concentration-type lower bound for |E e^f| for complex Lipschitz f, proved via martingale and Laplace-transform estimates (Lemmas 3.2–3.5), and on an inductive claim (Claim 4.1) that modifying one interaction changes the log-partition by at most a constant multiple of |λ_m|. Section 2 sketches a quasi-polynomial approximation algorithm based on the zero-free regions.","tokens_in":17623,"tokens_out":22822,"duration_ms":230764,"significance":"If the result holds, it is significant: it extends Barvinok's earlier multi-spin zero-free work to include an external field, exhibits a regime where the external field needs only grow logarithmically in the interaction strength, and achieves a scaling in r that is optimal up to a constant. The local analytic estimates are clean, self-contained, and mostly machine-checkable. However, the induction in Section 4 contains a false geometric assertion about an affine extension, and the current proof of the key Lipschitz estimate for the log-partition is incomplete. The theorem may well be true and the gap appears repairable, but the manuscript as written does not establish the advertised result.","major_comments":[{"comment":"The assertion that Ψ(t) is affine in each coordinate τ_j is false. For fixed other coordinates, each summand inside the expectation is of the form P(z) exp(A_z τ_j + B_z); a sum of such exponentials is not affine unless all A_z vanish. For example, with one inner coordinate, p=q=1/2, λ=1, φ(ξ)=ξ, Ψ(τ)=(e^τ+1)/2. Consequently, the argument of Ψ need not change by less than π along an edge, and the construction of ψ satisfying (4.3.2) is invalid as written. This is load-bearing: (4.3.2) supplies the Lipschitz bound for λ_m φ_τ + ψ used in (4.3.6) and hence in the final application of Theorem 3.1. The gap is likely repairable by defining log-ratio representatives on edges and checking consistency on 4-cycles using the smallness of the per-coordinate sums, but that argument is absent.","section":"§4.3, after Eq. (4.3.4)"},{"comment":"Theorem 3.1 Part (1) is applied to the cube {0,1}^J in the base case and in the induction step, but its hypothesis is p|J| ≥ 12. The theorem's assumption rp ≥ 12 only gives p|J| ≥ p·(something ≤ r), which can be much smaller than 12 when |J| is small. For instance, r=100, p=0.12, |J|=1 gives p|J|=0.12. Since J can be much smaller than r, the stated application is unjustified. A repair would embed J into r coordinates by adding dummy coordinates that the functions do not depend on; this is straightforward but not present.","section":"§4.2 and §4.3, applications of Theorem 3.1(1)"},{"comment":"In the paragraph after Eq. (4.3.1), the paper claims that Ψ(x)/Ψ(y)=e^α with |α| ≤ 2Σ_{i∈I_j}|λ_i| = 2Σ_{1≤i≤m-1: j∈J_i}|λ_i|. The equality is false: functions i∉I can depend on the varying coordinate j∈\\bar J and contribute a scalar factor e^{Σ λ_i(φ_i(x)-φ_i(y))} to the ratio. The final Lipschitz estimate can still be made to work using the larger bound 2Σ_{all i≤m-1: j∈J_i}|λ_i|, but the displayed equality and the subsequent conclusion are not justified as written.","section":"§4.3, ratio bound for Ψ(x)/Ψ(y)"}],"minor_comments":[{"comment":"The displayed inequality E e^{2g} ≤ exp{e^{-5L}} is incorrect: since 2g is 2L-Lipschitz, Corollary 3.4(2) gives E e^{2g} ≤ exp{e^{-6L} e^{2L}} = exp{e^{-4L}}. The subsequent bound b ≤ √2 e^{-2} still holds with this correction because e^{-4L}/2 ≤ 1/2, so the error does not affect the final estimate.","section":"§3.6, Part (2)"},{"comment":"The notation for J and its complement is very confusing, especially since overbars are lost in places. The same symbol J is used for the set of coordinates of φ_m, for the complement, and for indexing the solid cube. Please use distinct notation, e.g., J for the interaction set and K for its complement.","section":"§4.3 notation"},{"comment":"Eq. (4.1.2) writes E{λ_mφ_m + Σ...} where the exponential is missing; the text later uses exp. Similarly in Eq. (5.1.1). This is a typo that should be corrected.","section":"Claim 4.1, Eq. (4.1.2)"},{"comment":"The condition p = e^{-6L}/r may violate the standing assumption 0 < p ≤ q unless L is sufficiently large (e.g., L ≥ (1/6)ln(2r)). The theorem statement should either impose this or note that the argument works for p>q as well.","section":"Theorem 1.3"}],"recommendation":"major_revision","confidential_remarks":"The main theorem is likely correct and the local estimates are strong, but the branch-of-log construction in §4.3 is invalid as written, and there are several auxiliary justifications missing. I would encourage revision rather than rejection: the cycle-space repair for (4.3.2) is plausible and the rest of the proof appears solid. The erroneous e^{-5L} display in Theorem 3.1(2) is minor and does not affect the conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main result is a genuinely new zero-free regime for multi-spin partition functions: per-coordinate influence up to 1/(10√(rp)) with rp ≥ 12, and the resulting logarithmic-field scaling. The proof is mostly self-contained and the constants check out; the n=1 lemma, martingale bound, and Laplace tails are clean. The display error in Theorem 3.1 part (2) (E e^{2g} ≤ e^{-5L} instead of e^{-4L}) is harmless for the conclusion.\n\nWhat the reader missed: Section 4.3's claim that the extended Ψ(t) is affine in each coordinate of [0,1]^J is false. The expectation over the inner cube is a sum of exponentials of affine functions, not affine. For instance, with one outer coordinate and one inner coordinate, Ψ(τ) = p e^{λτ} + q is not affine. So the \"argument changes by less than π\" step is not justified as written. The gap looks repairable—edge log-ratios have small representatives (bounded by 2∑|λ_i| < π) and the cycle space is generated by 4-cycles with total angle < 2π—but the repair is not in the paper. Theorem 1.3's proof is a sketch that inherits the same issue.\n\nOverall, the new regime is real and the argument is mostly rigorous; the flaws are localized and fixable. The paper deserves a serious referee, who should insist on a correct branch-of-log construction before publication. I wouldn't cite it yet.","headline":"Solid new zero-free regime with a repairable gap in the branch-of-log step; worth refereeing.","tokens_in":18329,"tokens_out":5086,"would_cite":false,"duration_ms":48230,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B20","30C15","68R05","68W05","60C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Linear multi-spin energy growth needs only log external field growth","keywords":["zero-free region","partition function","external field","multi-spin interactions","p-biased Boolean cube","Lipschitz functions","spin systems","phase transition"],"falsifier":"Directly test Theorem 3.1(1) by random or exhaustive search: for n with pn=12, set L=1/(5√(pn)), and sample complex 1-Lipschitz functions f (e.g., f(x)=L(g(x)+i h(x)) with g,h real 1-Lipschitz and Eg=Eh=0); compute the ratio |E e^f| / E|e^f|. A single sample with ratio below 1/2 would falsify the lemma and force a smaller constant in Theorem 1.2. Alternatively, enumerate small n to find a function achieving the infimum and check whether the infimum is indeed ≥1/2.","tokens_in":17061,"feed_emoji":"⚛️","tokens_out":8398,"duration_ms":74178,"temperature":0.7,"pith_summary":"This paper proves a zero-free region for a general class of partition functions on the p-biased Boolean cube: sums of 1-Lipschitz functions, each depending on at most r coordinates, with the total influence of every coordinate bounded by 1/(10√(rp)), have a nonzero expectation of exp(Σ λ_i φ_i) as long as rp≥12. In spin language, this means that when interactions involve r spins and the external field is not too weak, the strength of the external field needed to keep the partition function away from zero grows only logarithmically with the interaction energy, not linearly. A companion result covers the small-p regime, where a linear growth of the field is again required. The proof is an induction that compares the expectation before and after perturbing one interaction, controlled by a concentration lemma for complex Lipschitz functions on the biased cube. This yields, among other things, a quasi-polynomial algorithm for approximating such partition functions.","feed_headline":"Linear multi-spin energy growth needs only log external field growth","feed_subtitle":"With rp≥12, partition functions stay zero-free and pressure analytic; the proof yields quasi-polynomial approximation algorithms.","key_machinery":"The argument rests on Theorem 3.1, a concentration lemma for a single complex L-Lipschitz function f on the p-biased cube: if pn≥12 and L=1/(5√(pn)), then |E e^f| ≥ (1/2)E|e^f| (Part (1); Part (2) gives a factor 1/5 when p=e^{-6L}/n). This injects the 1/√(pn) scale into the induction. The induction (Claim 4.1) shows that replacing the last function φ_m by any 1-Lipschitz function within distance 1 that depends on the same coordinates changes the expectation by a factor e^α with |α| ≤ 2|λ_m|; the ratio bound is obtained by interpolating linearly between φ_m and the perturbed function, extending the expectation to the solid cube [0,1]^J to select a consistent branch of the logarithm, and apply","core_discovery":"Theorem 1.2 states: on the p-biased cube with p≤q and rp≥12, if φ_1,...,φ_m are complex 1-Lipschitz functions each depending on at most r coordinates, and the sum of |λ_i| over functions depending on coordinate j is at most 1/(10√(rp)) for every j, then E exp(Σ λ_i φ_i) ≠ 0. Reading λ_i as interaction strengths and p as a function of the external field α through p=e^{-βα}/(e^{βα}+e^{-βα}), this says the external field needs to grow only logarithmically with the energy of multi-spin interactions to keep the partition function zero-free. A second theorem covers the very small p regime (p=e^{-6L}/r), where the field must again grow linearly, matching cluster-expansion predictions.","pith_inferences":["The crossover between the logarithmic regime (Theorem 1.2) and the linear regime (Theorem 1.3) is governed by the product rp: the paper's threshold rp≥12 suggests the transition occurs when the external field α is of order (1/β) ln(r/12); one could test numerically in a concrete model whether the zero-free region's shape changes sharply near that value.","The key lemma Theorem 3.1(1) has the flavor of an anti-concentration inequality; the same inductive scheme should transfer to other product spaces (e.g., continuous or discrete spins with more than two values), where a similar lower bound on |E e^f| would yield analogous logarithmic-field theorems.","The factor 1/2 in Theorem 3.1(1) is not optimized; if it could be improved toward 1, the per-coordinate bound 1/(10√(rp)) could be enlarged, potentially improving the constants in the algorithmic application.","The approximation algorithm derived from the zero-free region (Section 2) does not require real-rootedness or other algebraic structure, so the same strategy might apply to partition functions of hypergraph matchings, where real-rootedness fails; this could be checked by testing the algorithm's predicted O(m^k) complexity on random hypergraphs."],"forward_implications":["For a fixed inverse temperature β and interaction range r with rp≥12, multiplying the interaction strengths by a factor c only requires adding O(ln c) to the external field α to keep Z(H,β) ≠ 0 and the pressure analytic.","The zero-free region of the partition function as a function of the external field extends to a disc whose radius grows at least as p^{-1/2} as p decreases (for fixed r), up to constants.","The theorem yields a quasi-polynomial algorithm approximating the partition function within relative error ε, using O(m^k) expectations with k = O_δ(ln r + ln Σ|λ_i| - ln ε) (Section 1.5, Lemma 2.1).","In the hypergraph matching / polymer model, smaller selection probability p enlarges the zero-free disc for the penalty parameter λ, so stronger penalties can be charged while still approximating the matching statistics.","The dependence of the bound on r as r^{-1/2} is optimal up to a constant, per the author's earlier work [Ba26]; the constant 1/10 could presumably be improved but not the scaling."],"supporting_citations":[{"why":"The author's earlier inductive method for multi-spin interactions without external field; this paper adapts the induction, and [Ba26] also supplies the optimality bound for the r-dependence.","marker":"[Ba26]"},{"why":"Lee and Yang's result that complex zeros of the partition function govern phase transitions—the conceptual framework the zero-free results plug into.","marker":"[LY52]"},{"why":"Yang and Lee's original introduction of complex parameters into the partition function, establishing the connection between zeros and phase transitions.","marker":"[YL52]"},{"why":"Extended the Lee-Yang theorem to multi-spin interactions; the paper's setting is the multi-spin case, and this marks the baseline result being extended.","marker":"[SF71]"},{"why":"The standard reference for cluster expansion and statistical mechanics of lattice systems; serves as the comparison baseline and the source of the Lee-Yang framework.","marker":"[FV18]"},{"why":"Supplies the martingale concentration argument used in Lemma 3.3, on which Theorem 3.1 rests.","marker":"[Mc98]"},{"why":"Provides the Borel-Carathéodory and Cauchy estimates used in Lemma 2.1 for the approximation algorithm.","marker":"[La99]"}],"fun_headline_variants":["Zero-free: linear multi-spin energy needs only log external field","Log external field keeps multi-spin partition functions zero-free","Linear energy growth, log field growth: zero-free spin systems","Spin systems: linear multi-spin energy, log external field suffices"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole argument rests on Theorem 3.1(1): a complex-valued function on the biased cube whose values never change by more than L=1/(5√(pn)) between neighboring points must have its expectation of e^f at least half as large in absolute value as the expectation of |e^f|; if that factor 1/2 or the 1/√(pn) scale fails, the per-coordinate bound 1/(10√(rp)) and the logarithmic-field conclusion would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Zero-free: linear multi-spin energy needs only log external field","Log external field keeps multi-spin partition functions zero-free","Linear energy growth, log field growth: zero-free spin systems","Spin systems: linear multi-spin energy, log external field suffices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000307,"raw_usage":{"total_tokens":1641,"prompt_tokens":839,"completion_tokens":802,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":738}},"tokens_in":583,"tokens_out":802,"duration_ms":8975,"temperature":1.0,"reasoning_tokens":738,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:50:28.516069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly test Theorem 3.1(1) by random or exhaustive search: for n with pn=12, set L=1/(5√(pn)), and sample complex 1-Lipschitz functions f (e.g., f(x)=L(g(x)+i h(x)) with g,h real 1-Lipschitz and Eg=Eh=0); compute the ratio |E e^f| / E|e^f|. A single sample with ratio below 1/2 would falsify the lemma and force a smaller constant in Theorem 1.2. Alternatively, enumerate small n to find a function achieving the infimum and check whether the infimum is indeed ≥1/2.","supporting_citations":[],"review_version":1}