{"id":"bc588e21-a9e7-4787-a710-4e2d807b08b4","arxiv_id":"2608.17920","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Every antiferromagnetic spin model, meaning a nonnegative symmetric interaction matrix with at most one positive eigenvalue, satisfies a degree-wise clique-minimizing homomorphism inequality.","lead":"A new theorem shows that any spin model whose interaction matrix has at most one positive eigenvalue makes complete graphs the minimizers of weighted homomorphism counts. The result unifies known extremal inequalities for independent sets, proper colorings, and Ising models, and proves two open conjectures.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; the central claim appears internally consistent and the proof closes.","rationale":"I focused on the reader's identified weakest assumption, Lemma 3.5, because it is the most technically substantive ingredient and the one on which the rest of the proof depends most directly. After tracing its proof, I found no error: the chain-rule reduction to count laws is justified by exchangeability, the concavity of the m-th root of the Lorentzian generating polynomial gives the needed tangent-plane bound, and the variational formula yields the contraction with the correct factor 1/m. The subsequent uses of Lemma 3.5 in Lemma 3.7, Proposition 4.2, and Lemma 5.3 are algebraically consistent; I also checked the induction, the deletion identity (P2), and the continuity reductions. The paper is not machine-checked, and the proof relies on the substantial BH20 Lorentzian framework, so confidence is not maximum, but I found no concrete gap that would move the verdict. The reader's emphasis on Lemma 3.5 is therefore reasonable, but it does not translate into an objection in my reading. I recommend leaving the ACCEPT verdict unchanged.","tokens_in":24011,"tokens_out":58789,"duration_ms":509066,"concrete_test":"Write a small program enumerating all exchangeable laws on A^m for |A|=2,3 and m≤5, all Lorentzian ν in that set (checking M-convex support, Hessian for degree 2, and derivative recursion for higher degrees), and all exchangeable µ absolutely continuous with respect to ν, and test D(µ1||ν1)≤(1/m)D(µ||ν); a counterexample would refute Lemma 3.5, while exhaustive agreement for these sizes would corroborate the central step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I could not identify a load-bearing flaw. The main risk is Lemma 3.5, the relative entropy contraction for exchangeable Lorentzian laws; it is the engine behind Lemma 3.7, Proposition 4.2, and the local membership resolution in Lemma 5.3. I checked the proof carefully: the reduction D(µ||ν)=D(πµ||πν) is valid because exchangeability makes the T-conditional laws uniform, the tangent-plane bound (3.6) follows from concavity of hν^{1/m}, and the variational argument yields exactly the 1/m factor. The closures in Lemma 3.4, the induction in Lemma 3.7, the optimized pressure identity in Lemma 4.1, and the monotonicity in Proposition 4.2 are algebraically consistent. The reduction to positive H and positive fugacities by continuity is also sound. Thus I do not see a concrete defect to attack; the paper's main theorem is as strong as its coherent technical machinery.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves Theorem 1.2: for every antiferromagnetic edge-weighted model H (entrywise nonnegative symmetric matrix with at most one positive eigenvalue), every graph G, and every collection of vertex fugacity vectors λ^(v), the vertex-inhomogeneous partition function satisfies eZ_G(λ^(v)) ≥ ∏_v Z_{d_v+1}(λ^(v))^{1/(d_v+1)}. Setting all fugacities to 1 gives Theorem 1.1, i.e., every antiferromagnetic model is clique-minimizing. The proof combines Lorentzian polynomial techniques with a new relative-entropy contraction for exchangeable Lorentzian laws (Lemma 3.5), a strengthened delete-one Han/Shearer inequality (Lemma 3.7), and a monotonicity principle for an optimized pressure (Proposition 4.2) to resolve the local membership problem in an inductive localization argument (Section 5). The results recover and generalize known inequalities for independent sets, q-colorings, and semiproper colorings, and confirm conjectures of the authors and of Davies and LeBlanc.","tokens_in":24157,"tokens_out":19698,"duration_ms":161174,"significance":"The main theorem is a clean and broad unification: a single spectral condition implies clique minimization for a large class of graph homomorphism models, including the hard-core model, q-colorings, and the antiferromagnetic Ising model. The vertex-inhomogeneous strengthening is not artificial generality; it is needed for the induction and yields new consequences such as the Davies–LeBlanc conjecture. The proof is essentially self-contained and introduces a genuinely new tool—the one-coordinate relative entropy contraction for exchangeable Lorentzian laws—which is likely to be useful beyond this paper. The authors are transparent about relying on prior work, and they provide full proofs of the needed Lorentzian property, the entropy inequalities, and the local membership resolution. I checked the key contraction lemma (Lemma 3.5), its consequences (Lemma 3.7 and Proposition 4.2), and the induction closing in Section 5, and found no gaps or circularity. The paper ships complete derivations with no fitted parameters and a falsifiable main statement, which are notable strengths.","major_comments":[],"minor_comments":[{"comment":"The reduction to an entrywise-positive model H^(ε) is sound, but the logical order of the two limiting arguments (perturbing H and extending to nonnegative fugacity vectors) is easy to misread as taking place inside the induction step; explicitly stating that the induction is run for H^(ε) for all graphs before letting ε→0 would clarify the exposition.","section":"Section 5 (after Corollary 2.3)"},{"comment":"In the statement of Lemma 3.5, the notation D(µ1∥ν1) relies on the earlier definition of µ_i as the i-th marginal; a parenthetical reminder that µ1 and ν1 are the one-coordinate marginals would improve readability.","section":"Lemma 3.5"},{"comment":"The proof of Lemma 5.5 uses the bijection between {x∈(R_+)^I : D_{m,s}(x)=1} and P_+(A) several times; writing the inverse map explicitly (x_σ = α(σ)s_σ^{-(m-1)}) would help the reader verify the normalization.","section":"Lemma 5.5"},{"comment":"The assertion that polarization makes both the law ρ and its complementary law ρ* log-concave is stated without proof; since this remark is not used in the main argument, either add a one-sentence justification or note that it is a sketch.","section":"Remark 3.8"},{"comment":"There are some typographical artifacts in the rendered equations (for example, missing spaces around sums in Section 3); these should be corrected in the final version.","section":"Throughout"}],"recommendation":"minor_revision","confidential_remarks":"This is a strong paper well within the scope of math.CO. The central claim is internally consistent, and the key entropy contraction lemma, which is the most delicate point, checks out. The use of the authors' own prior work is transparent and the required lemmas are proved in full. The LLM-use disclosure is thorough and does not affect my assessment of the mathematics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper proves a real theorem, and the proof appears to hold together. Every antiferromagnetic model is clique-minimizing, and the vertex-inhomogeneous inequality is new and strictly stronger than the homogeneous cases that were known before. It settles conjectures from Lee–Seo and Davies–LeBlanc, and it subsumes the independent-set, q-coloring, semiproper-coloring, and Ising results under one spectral condition. That is a genuine advance, not a repackaging.\n\nWhat the paper does well: the main induction is fully written out, the local membership reduction is clean, and the entropy machinery is developed carefully rather than cited as a black box. The key technical premise, Lemma 3.5 (relative entropy contraction for exchangeable Lorentzian laws), is proved in detail, and the proof looks correct. I checked the reduction from D(µ||ν) to D(πµ||πν), the use of concavity of h^{1/m}, and the variational argument; the 1/m factor comes out honestly. The stress-test note agrees, and on reading the paper I do not see a load-bearing flaw. The reduction to entrywise-positive H by adding ε rr^T is sound, and the continuity argument is legitimate.\n\nSoft spots, in proportion: the paper leans on the Lorentzian polynomial framework of Brändén–Huh, so a reader who does not trust that framework will have to accept a substantial external dependency. Lemma 3.5 is the engine, and while it is proved, it is intricate; that is where a subtle error would most likely hide. The paper is not formally verified, so a skeptical referee should spend time there. The self-citations to [LS26] and [LOS25] are appropriate: the localization strategy is theirs, but the decisive steps are reproduced in this paper, and Lemma 2.7 is explicitly credited to [LOS25]. The LLM-use disclosure is unusual but transparent; it does not affect the mathematics as far as I can tell.\n\nWho this is for: anyone working in extremal graph homomorphisms, entropy methods in combinatorics, or Lorentzian polynomials. It is a significant paper within the field and deserves a serious referee. The referee should be asked to verify Lemma 3.5 and the reduction in Section 5.1 carefully, but I would expect the proof to survive scrutiny.","headline":"A genuinely general clique-minimization theorem with a coherent proof; the main risk is the entropy-contraction lemma, but it is proved and the induction closes.","tokens_in":24653,"tokens_out":932,"would_cite":true,"duration_ms":10980,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":["05C35","05C15","05A20","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"Every antiferromagnetic spin model is clique-minimizing: for any graph G, the weighted homomorphism count into H is at least a product of clique counts determined by the degrees of G.","keywords":["antiferromagnetic models","clique-minimizing","graph homomorphisms","Lorentzian polynomials","relative entropy contraction","optimized pressure","partition functions","vertex-inhomogeneous inequalities"],"falsifier":"Fix $A=\\{0,1\\}$, $m=3$, let $\\nu$ be the exchangeable product of three independent Bernoulli(1/2) coordinates (a Lorentzian law), and let $\\mu$ be the exchangeable law placing mass $1/3$ on each word with exactly one 1. Compute $D(\\mu_1\\|\\nu_1)$ and $(1/3)D(\\mu\\|\\nu)$; if the first exceeds the second, Lemma 3.5 is false and the proof collapses. A direct check of inequality (1.1) by exhaustive search over all graphs on at most six vertices for a specific antiferromagnetic $H$, such as the $2\\times 2$ Ising matrix with off-diagonal entry $B=1/2$, would also test the main theorem computationally.","tokens_in":23826,"feed_emoji":"🔺","tokens_out":15445,"duration_ms":122969,"temperature":0.7,"pith_summary":"This paper identifies a single spectral condition that forces complete graphs to be extremizers in counting problems on graphs. A model is called antiferromagnetic when its symmetric weight matrix is entrywise nonnegative and has at most one positive eigenvalue; the paper proves that every such model is clique-minimizing. That is, for every graph $G$, $\\operatorname{hom}(G,H)$ is at least the product, over vertices $v$ of degree $d_v$, of $\\operatorname{hom}(K_{d_v+1},H)^{1/(d_v+1)}$. The proof works with a stronger vertex-inhomogeneous inequality that permits a separate fugacity vector at each vertex, and this stronger form actually characterizes antiferromagnetism. The result matters because it unifies the known lower-bound inequalities for independent sets, proper $q$-colorings, semiproper colorings with at most two proper colors, and antiferromagnetic Ising models, and settles open conjectures about them through one mechanism: relative entropy contraction for exchangeable Lorentzian laws.","feed_headline":"Antiferromagnetic models are clique-minimizing","feed_subtitle":"One spectral condition on a spin model forces complete graphs to minimize weighted homomorphism counts.","key_machinery":"The engine is the one-coordinate relative entropy contraction for exchangeable Lorentzian laws, Lemma 3.5: if $\\mu$ and $\\nu$ are exchangeable laws on $A^m$ and the count-generating polynomial of $\\nu$ is Lorentzian, then $D(\\mu_1\\|\\nu_1) \\le \\frac{1}{m} D(\\mu\\|\\nu)$. From this the paper derives a strengthened delete-one entropy inequality, which feeds into an optimized-pressure function $F_m$ satisfying the monotonicity $mF_{m-1} \\ge (m-2)F_m$. This pressure monotonicity resolves the local membership problem that blocks the inductive proof, replacing the model-specific case checks of earlier work with a general argument. The supporting algebraic fact is that clique partition functions of entrywise-positive antiferromagnetic models are Lorentzian, so their normalized $d$-th roots are concave.","core_discovery":"The paper's central claim is Theorem 1.2: for every antiferromagnetic model $H$, every graph $G$, and every collection of fugacity vectors $\\lambda^{(v)}$ attached to the vertices of $G$, the vertex-inhomogeneous partition function satisfies $\\widetilde{Z}_G(\\lambda^{(v)} : v\\in V(G)) \\ge \\prod_{v\\in V(G)} Z_{d_v+1}(\\lambda^{(v)})^{1/(d_v+1)}$. Setting every fugacity vector to the all-ones vector turns this into Theorem 1.1, $\\operatorname{hom}(G,H) \\ge \\prod_{v\\in V(G)} \\operatorname{hom}(K_{d_v+1},H)^{1/(d_v+1)}$, which is the statement that antiferromagnetic models are clique-minimizing. The inequality is sharp: equality occurs when $G$ is a disjoint union of cliques and the fugacity vector is constant on each connected component. The vertex-inhomogeneous formulation is not merely a proof device: on the two-vertex graph it reduces to $(x^\\top H y)^2 \\ge (x^\\top H x)(y^\\top H y)$ for all nonnegative $x,y$, a condition equivalent to $H$ being antiferromagnetic, so the strong inequality precisely characterizes the class.","pith_inferences":["The relative entropy contraction is a statement about exchangeable Lorentzian laws with no reference to graphs; the same optimized-pressure monotonicity may yield extremal inequalities for other counting problems whose generating polynomials are Lorentzian, such as bases of matroids or matchings in bipartite graphs.","The complementary upper-bound direction — biclique-maximizing for antiferromagnetic models — is left open, and the same entropy toolkit, applied with a reversed inequality or a different support family, is a natural candidate route to settle it.","Because the vertex-inhomogeneous inequality characterizes antiferromagnetism exactly, any future clique-minimizing model outside this class must violate the two-vertex inequality with unequal fugacities; this gives a concrete test for broader conjectures."],"forward_implications":["For every $d$-regular graph $G$ and every antiferromagnetic Ising model with edge activity in $[0,1]$, the normalized homomorphism count $\\operatorname{hom}(G,H)^{1/|V(G)|}$ is at least the clique value $\\operatorname{hom}(K_{d+1},H)^{1/(d+1)}$, with equality for disjoint unions of $K_{d+1}$.","The lower-bound inequalities for independent sets, proper $q$-colorings, and semiproper colorings with at most two proper colors, including their irregular-degree versions, follow from a single theorem with no model-specific arguments.","The vertex-inhomogeneous inequality yields bounds for weighted list-coloring and multivariate independence-polynomial problems with arbitrary vertex fugacities, and its fugacity derivatives control one-vertex marginal probabilities (occupancy fractions).","A model satisfying the vertex-inhomogeneous inequality for all graphs and all fugacity vectors must be antiferromagnetic, so the strong inequality gives a complete structural characterization; the homogeneous inequality alone is strictly weaker, since tensor products of antiferromagnetic models are clique-minimizing without being antiferromagnetic."],"supporting_citations":[{"why":"Defines clique-minimizing and biclique-maximizing models and supplies the ferromagnetic upper-bound counterpart and the conjectures that frame the paper.","marker":"[SSSZ20]"},{"why":"Develops the inductive localization and dual-set membership reduction (its Lemma 3.4 and Equation (2.2)) which the general proof adapts, and states Conjecture 1.3 confirmed here.","marker":"[LS26]"},{"why":"Introduces Lorentzian polynomials and provides the concavity and closure properties used to show clique partition functions are Lorentzian and to prove the entropy contraction.","marker":"[BH20]"},{"why":"Supplies the variational formulas for relative entropy and the Gibbs variational principle used in the entropy-contraction and optimized-pressure arguments.","marker":"[DE97]"},{"why":"Prior work establishing that clique partition functions of antiferromagnetic models are Lorentzian in full generality and connecting this to the biclique-maximizing conjecture.","marker":"[LOS25]"},{"why":"States the antiferromagnetic-Ising occupancy conjecture that Theorem 1.2 confirms in the regular case.","marker":"[DL24]"},{"why":"Proves the irregular-graph independent-set inequality that Theorem 1.1 generalizes, and introduces the vertex-dependent fugacity treatment that the present theorem strengthens.","marker":"[SSSZ19]"}],"fun_headline_variants":["Antiferromagnetic models hit minimum on cliques","One spectral condition pins minima to cliques","Spectral twist: cliques minimize antiferromagnetic homomorphisms","Cliques are the minimizers for antiferromagnetic models","Spectral condition forces cliques to be minimal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"If the one-coordinate relative entropy contraction for exchangeable Lorentzian laws — $D(\\mu_1\\|\\nu_1) \\le \\frac{1}{m}D(\\mu\\|\\nu)$ whenever $\\nu$ is Lorentzian — fails for some pair of laws, then the monotonicity of the optimized pressure breaks and the induction in the proof of Theorem 1.2 cannot close.","fun_headline_variants_meta":{"raw":{"variants":["Antiferromagnetic models hit minimum on cliques","One spectral condition pins minima to cliques","Spectral twist: cliques minimize antiferromagnetic homomorphisms","Cliques are the minimizers for antiferromagnetic models","Spectral condition forces cliques to be minimal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000986,"raw_usage":{"total_tokens":4226,"prompt_tokens":1031,"completion_tokens":3195,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":3133}},"tokens_in":647,"tokens_out":3195,"duration_ms":22224,"temperature":1.0,"reasoning_tokens":3133,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-27T20:09:45.452261+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix $A=\\{0,1\\}$, $m=3$, let $\\nu$ be the exchangeable product of three independent Bernoulli(1/2) coordinates (a Lorentzian law), and let $\\mu$ be the exchangeable law placing mass $1/3$ on each word with exactly one 1. Compute $D(\\mu_1\\|\\nu_1)$ and $(1/3)D(\\mu\\|\\nu)$; if the first exceeds the second, Lemma 3.5 is false and the proof collapses. A direct check of inequality (1.1) by exhaustive search over all graphs on at most six vertices for a specific antiferromagnetic $H$, such as the $2\\times 2$ Ising matrix with off-diagonal entry $B=1/2$, would also test the main theorem computationally.","supporting_citations":[],"review_version":1}