{"id":"c65f9d2a-e6bd-428c-9d0f-7a28ab5daa5f","arxiv_id":"2608.12309","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A new proof framework for replica-symmetric marginals is applied to the Sherrington-Kirkpatrick model, but the proof contains a fixable circular dependency.","lead":"The paper introduces a cavity-contiguity framework for proving that single-spin marginals in mean-field spin glasses converge to the one-dimensional replica-symmetric cavity measure. It applies this framework to the high-temperature Sherrington-Kirkpatrick model, where the convergence result is already known.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's proof invokes Proposition 2 without verifying its hypothesis, and no rate of convergence of E<Q12> to q is provided; the stated O(1/N) overlap concentration is therefore unsupported, leaving Theorem 1's proof incomplete.","rationale":"The reader's stated weakest assumption concerns applicability of the log-Sobolev inequality when J is not positive semidefinite. I do not press that point: the cited [9] inequality is formulated through the operator norm of the interaction matrix, and the phrase 'positive semidefinite' in Appendix B appears to be a slip rather than the operative hypothesis; the norm condition ||B||_op < 2 is what is used. The reader's rationale, however, already flags a circular dependency between Propositions 1 and 2, and my independent reading locates a sharper version of the same problem: Proposition 1 needs a rate of convergence of E<Q12> to q that is never derived, and it invokes Proposition 2 before the hypothesis of that proposition is verified. This is a genuine proof-level gap in a paper whose contribution is precisely the proof framework. The theorem itself is classical (see [25]) and the gap appears repairable through either a direct rate for the mean overlap or a reorganized proof, so the appropriate disposition remains CONDITIONAL rather than REJECT or ACCEPT.","tokens_in":16541,"tokens_out":9971,"duration_ms":90976,"concrete_test":"Formalize the last paragraph of Appendix B using only Lemmas 9, 10, and 11, with Proposition 2 and Lemma 4 removed. Specifically, attempt to prove sup_N N^{1/2} |E<Q12> - q| < ∞ from the fixed-point equation and the O(N^{-1/2}) cavity/full mean comparison. If the mean gap cannot be bounded at rate N^{-1/2}, Proposition 1's O(1/N) overlap concentration fails as stated; independently check whether Lemma 4 can be replaced by Proposition 2's own hypothesis without altering the final argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive gap is in the proof of Proposition 1 (Appendix B), which is what makes Theorem 1 non-conditional. The proof takes a subsequential limit q' of E<Q12> and tries to rule out q' ≠ q by asserting that 'over this sub-subsequence, Proposition 2 holds'. But Proposition 2 requires E[(Q12 - q')^2]_c ≤ K/N, and this condition is not established. Lemma 9 and the Poincaré bound (10) give concentration of Q12 around the random mean E<Q12> at rate 1/N, namely E<(Q12 - <Q12>)^2> ≤ C/N and Var(<Q12>) ≤ C/N; together they yield E<(Q12 - E<Q12>)^2> ≤ C/N, but they give no rate for the gap |E<Q12> - q'|, which is only known to tend to 0 along the subsequence. Lemma 11 gives only O(N^{-1/2}) bounds relating full and cavity means, not a rate for the mean gap to q'. Thus the hypothesis of Proposition 2 has not been verified at the point where it is invoked. Conversely, Lemma 4 in the proof of Proposition 2 invokes Proposition 1 to control E[(Q_ll' - q)^2]_c; in a standalone proof this could be replaced by Proposition 2's own assumption, but as written Proposition 2 is not proved independently of Proposition 1. The two results are therefore mutually dependent, and the unconditional N^{-1/8} rate claimed in Theorem 1 is not supported by the argument as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a probabilistic 'cavity-contiguity' framework for proving replica-symmetric convergence of local marginals in mean-field Gibbs systems, and implements it in the high-temperature Sherrington-Kirkpatrick model. The main result, Theorem 1, states that for beta<1/4 and h not equal to 0, the marginal law of a fixed spin converges in expected total variation to the one-dimensional cavity measure mu_{i0}(sigma)=exp{(2 beta sqrt(q) Z+h) sigma}/(2 cosh(2 beta sqrt(q) Z+h)) at rate O(N^(-1/8)), where q is the unique solution of q=E tanh^2(2 beta sqrt(q) Z+h). The proof separates concentration of the overlaps, identification of the limiting cavity field, and comparison of measures via Radon-Nikodym derivatives and Hellinger integrals. The paper also states corollaries for local functions and the free energy, and discusses extensions to Bayesian inference models.","tokens_in":1779,"tokens_out":2143,"duration_ms":128382,"significance":"The paper's main contribution is methodological: it proposes a modular route to local replica-symmetric results that avoids interpolation and instead uses concentration inequalities plus explicit change-of-measure estimates. The conceptual separation of concentration, identification, and measure comparison is appealing, and the proof is short and readable. The authors are candid that the N^(-1/8) rate is not optimal and that the framework requires strong concentration estimates. If the proof gaps described below are repaired, this cavity-contiguity approach could be a useful addition to the toolbox for mean-field spin systems and Bayesian inference. The theorem itself is classical in the SK model, so the value of the paper rests on the correctness and clarity of the new proof rather than on the novelty of the statement.","major_comments":[{"comment":"The proof of Proposition 1 is incomplete because of a circular dependency. In the final paragraph, the authors take a subsubsequence along which E<Q12> converges to q' and write 'Over this sub-subsequence, Proposition 2 holds.' However, Proposition 2 requires the hypothesis E[(Q12-q')^2]_c <= K/N, and this is not established anywhere: Lemma 9 and inequality (10) give only E[(Q12-<Q12>)^2] <= C/N and Var(<Q12>) <= C/N, which together yield E[(Q12-E<Q12>)^2] <= C/N but no rate for |E<Q12>-q'|; Lemma 11 gives only O(N^(-1/2)) relations between full and cavity means. Thus the hypothesis of Proposition 2 is not verified at the point of invocation, and the identification q'=E tanh^2(2 beta sqrt(q') Z+h) is unsupported. Conversely, the proof of Lemma 4 in Proposition 2 invokes Proposition 1 to control E[(Q_{l,l'}-q)^2]_c; this can be repaired by using the assumption of Proposition 2 together with exchangeability, but as written Proposition 2 is not proved independently of Proposition 1. Consequently, Theorem 1 is not established as written.","section":"Appendix B, proof of Proposition 1"},{"comment":"The Poincare inequality (9) is derived from the log-Sobolev inequality of [9], but the text states that [9]'s result is for 'positive semidefinite interaction matrices'. The two-replica interaction matrix B = gamma diag(J,J) is not positive semidefinite when the Wigner matrix J has negative eigenvalues, so the cited LSI does not apply in the form used. This affects the concentration bounds (10), Lemma 9, and Lemma 11 that support Proposition 1. If the LSI in [9] in fact holds without the positive-semidefiniteness assumption, the authors should state and prove the needed variant; otherwise a different concentration argument is required.","section":"Appendix B, inequality (9) and the use of [9]"}],"minor_comments":[{"comment":"The proof of Lemma 3 contains an incorrect identity: the displayed formula E<exp{t1 DeltaH(sigma_i0,Theta_N)}>_c = exp{2 t1^2 beta^2 + t1 |h_i0|} cannot be correct as written because the expectation over sigma_i0 in {+-1} should produce a factor cosh(t1 h) rather than exp(t1 |h|), and the notation h_i0 is not defined. Since only boundedness of the exponential moments is needed, this error is not load-bearing, but it should be corrected.","section":"Lemma 3"},{"comment":"The notation for Q_{l,l'} is inconsistent: in Section 3.1 the overlap is defined as a sum over all N coordinates, while in Lemma 4 and the coupling construction the property Q_{l,l}=(N-1)/N shows that the cavity overlap with i0 removed is being used. The authors should introduce separate notation, for example tilde Q_{l,l'} = N^{-1} sum_{i not equal to i0} sigma_i^{(l)} sigma_i^{(l')}, and state the elementary relation between the two overlaps.","section":"Section 3.4, Lemma 4"},{"comment":"The claim that 'as functions of Theta_N^{(1)},...,Theta_N^{(K)}, log Xbar_{N,K} is 1-Lipschitz' is imprecise: the Lipschitz constant with respect to the l1 norm is at most beta, because each derivative has magnitude at most beta. Since beta<1/4, the subsequent bound with constant C is valid, but the statement should state the constant explicitly.","section":"Proof of Proposition 2, bound for Delta3"},{"comment":"The proof of Lemma 9 invokes the 'Gauss-Poincare inequality [10]', but reference [10] is a concentration-inequality survey and does not appear to contain the Gaussian Poincare inequality; a standard textbook reference should be supplied.","section":"Lemma 9"},{"comment":"The proof of Corollary 1 is only sketched for functions of two replicas; since Theorem 1 concerns a single marginal, the extension to joint marginals of two or more spins should either be stated as a separate lemma or proved explicitly rather than asserted in one sentence.","section":"Corollary 1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test report accurately identifies the central issue: the proof of Proposition 1 invokes Proposition 2 without verifying its hypothesis, while Proposition 2's proof cites Proposition 1 in Lemma 4, making the two results mutually dependent as written. The second serious issue is the application of [9]'s log-Sobolev inequality to a matrix that is not positive semidefinite. The theorem itself is classical, so the paper's contribution is methodological; the framework is promising and the gaps appear localizable, so I recommend major revision rather than rejection. The authors should either repair the subsequential-limit argument in Proposition 1 with a direct proof that any subsequential limit satisfies the fixed-point equation, or supply the missing rate; and they should resolve the LSI applicability question, possibly by taking Proposition 1 from the existing literature and revising the claim of a self-contained non-interpolative proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's value is methodological, not the theorem. The authors say up front that a version of Theorem 1 is already in Talagrand's book, and they aim instead for a proof that separates concentration, identification, and measure comparison. That framework is real, and the cavity-contiguity exposition is quite clean. The Cholesky coupling in Section 3.4.1 is a nice way to put the finite-N and limiting cavity fields on one space, and the decomposition of the Hellinger integral into five terms is transparent and teachable.\n\nThe soft spot is load-bearing: the proof of Theorem 1 is not self-contained because Propositions 1 and 2 are mutually dependent. Proposition 2's proof uses Lemma 4, which relies on Proposition 1 for the overlap concentration E(Q_l l' - q)^2 <= C/N. Proposition 1's proof in Appendix B takes a subsequential limit q' of E<Q12>, then tries to rule out q' != q by asserting that Proposition 2 holds over that sub-subsequence. But Proposition 2 requires E[(Q12 - q')^2]_c <= K/N, and the bounds established in Appendix B only give concentration of Q12 around its disordered mean E<Q12>, with no rate for the gap |E<Q12> - q'| along the subsequence. So the hypothesis of Proposition 2 is not verified at the point it is invoked. This is a genuine circularity, and I checked the stress-test note against the text: it is right. It is likely fixable by proving overlap concentration directly from earlier results, but as written the proof does not stand.\n\nA second, smaller issue is the use of the LSI from [9]. Appendix B states the result for positive semidefinite interaction matrices, but the two-replica matrix B = gamma diag(J,J) is indefinite. If [9]'s theorem actually only requires an operator norm bound, the text misstates the hypothesis; if it genuinely needs PSD, inequality (9) is unsupported. Either way the authors should reconcile the statement.\n\nThere are also a few unchecked small claims, like the 1-Lipschitz assertion in the proof of Proposition 2, which is likely true but unstated. That is minor by comparison.\n\nWho should read this: anyone working on alternative proofs of replica-symmetric asymptotics, especially for Bayesian inference. The framework is promising, and the paper is worth a serious referee, but only as a major revision. I would encourage engaging with it, with the expectation that the circularity must be broken and the LSI citation tightened.","headline":"A clean modular framework for replica-symmetric marginals, but a circular dependency between Propositions 1 and 2 leaves Theorem 1 unproved as written.","tokens_in":17412,"tokens_out":3337,"would_cite":false,"duration_ms":29720,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B44","60F05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A cavity-contiguity framework proves replica-symmetric single-spin marginals in the high-temperature Sherrington-Kirkpatrick model.","keywords":["cavity method","contiguity","replica symmetry","Sherrington-Kirkpatrick model","total variation convergence","log-Sobolev inequality","overlap concentration","mean-field Gibbs measures"],"falsifier":"Take a fixed $\\beta<1/4$, $h\\neq 0$ (for instance $\\beta=0.24$, $h=1$), simulate the Sherrington-Kirkpatrick Gibbs measure at moderate $N$, and estimate $E\\langle (Q_{12}-q)^2\\rangle_c$. If the variance does not decay like $C/N$—or if the Poincaré inequality (9) fails explicitly for a test function under the indefinite two-replica measure—then Proposition 1, and with it Theorem 1, cannot hold as stated.","tokens_in":16335,"feed_emoji":"🎲","tokens_out":11481,"duration_ms":87076,"temperature":0.7,"pith_summary":"At inverse temperature $\\beta<1/4$ and nonzero external field $h$, the Sherrington-Kirkpatrick spin glass is in its replica-symmetric phase, and the paper proves that the law of any fixed spin converges in total variation to the effective one-dimensional cavity measure predicted by the replica method, with expected squared total-variation error at most $C/N^{1/8}$. The proof works through a cavity-contiguity framework: remove one coordinate from the Hamiltonian, compare the full Gibbs measure with the cavity measure by a Radon-Nikodym derivative, and control the comparison with Hellinger-type estimates. The authors deliberately separate three ingredients—concentration of overlaps, identification of the asymptotic Gaussian cavity field, and quantitative comparison of measures—so that Gaussianity of the disorder is used only where it is actually needed. This matters because it offers a modular alternative to interpolation-based arguments and a path toward Bayesian inference models with non-Gaussian disorder.","feed_headline":"Cavity-contiguity proof pins spin-glass marginals at N^{-1/8}","feed_subtitle":"A modular proof splits concentration from measure comparison and needs only weak Gaussianity, opening a route to Bayesian inference.","key_machinery":"The load-bearing object is the Radon-Nikodym derivative of the full Gibbs measure with respect to the cavity measure, $D_N=\\exp\\{\\Delta H_N(\\sigma_{i_0},\\Theta_N)\\}/\\langle\\exp\\{\\Delta H_N(\\sigma_{i_0},\\Theta_N)\\}\\rangle_c$, obtained from the cavity decomposition $H_N=H_{N-1}+\\Delta H_N+\\delta_N$. The proof couples the finite-$N$ Gaussian cavity fields $(\\Theta_N^{(l)})$ to limiting fields $(\\theta_l)$ through Cholesky decompositions of the overlap matrix, and then converts convergence of logarithmic moments of $D_N$ into total-variation control via Hellinger integrals. Overlap concentration is supplied by a log-Sobolev inequality imported into the two-replica system, which yields the Poincaré inequality that powers Proposition 1.","core_discovery":"The central claim is Theorem 1: for $\\beta<1/4$ and $h\\neq 0$, there is a constant $C$ such that $\\mathbb{E}\\|P_{N,i_0}(\\cdot|J)-\\mu_{i_0}(\\cdot)\\|^2_{TV}\\le C/N^{1/8}$, where $\\mu_{i_0}(\\sigma)=\\exp\\{(2\\beta\\sqrt{q}\\,z+h)\\sigma\\}/(2\\cosh(2\\beta\\sqrt{q}\\,z+h))$, $z$ is a standard Gaussian, and $q$ is the unique solution of $q=\\mathbb{E}\\tanh^2(2\\beta\\sqrt{q}\\,z+h)$. The statement itself is classical; the contribution is the mechanism of proof. The marginal is exhibited as a contiguity limit of the cavity measure through a likelihood-ratio tilt, making transparent that replica-symmetric local convergence does not need interpolation identities. The theorem is derived from overlap concentration, a coupling of the finite-$N$ Gaussian cavity fields to their limiting Gaussian fields, and a five-term decomposition of the log-Hellinger integral, each term controlled by a distinct estimate.","pith_inferences":["The rate $C/N^{1/8}$ is an artifact of balancing the $K=N^{1/4}$ truncation against competing error terms; a sharper handling of the Hellinger logarithm could plausibly push the rate toward the $N^{-1/2}$ scale set by overlap concentration, though the paper does not claim this.","If the positivity gap in the log-Sobolev import can be closed, the concentration step should survive for indefinite or non-Gaussian couplings, making the framework a template for leave-one-out proofs in Bayesian inference with mismatch.","The Hellinger-based comparison is stated in a way that suggests path-space analogues: replacing Hellinger integrals by Hellinger processes could yield asymptotic scalar descriptions for Langevin dynamics associated with these Gibbs measures."],"forward_implications":["Averages of local observables over replicas converge to expectations under the effective cavity law; the overlap parameter $q$ is shown to satisfy its replica-symmetric fixed point equation (Corollary 1).","The asymptotic free energy at $\\beta<1/4$, $h\\neq 0$ equals $\\mathbb{E}\\log 2\\cosh(2\\beta\\sqrt{q}Z+h)+\\beta^2(1-q)^2$, with $Z$ standard Gaussian (Corollary 2).","Because the framework uses Gaussianity only through concentration and cavity-field identification, the same cavity-contiguity steps should extend to non-Gaussian disorder whenever comparable concentration estimates hold.","The three-component structure (concentration, identification, comparison) allows different concentration tools—Poincaré, log-Sobolev, transportation, or related inequalities—to be substituted without altering the measure-comparison part."],"supporting_citations":[{"why":"Supplies the log-Sobolev inequality from which the Poincaré inequality (9) is derived to prove overlap concentration in Proposition 1.","marker":"[9]"},{"why":"Provides the classical high-temperature treatment whose overlap-concentration and fixed-point-uniqueness arguments the paper adapts; the paper notes Theorem 1 already appears there with optimal rates.","marker":"[25]"},{"why":"Companion paper that first used the leave-one-out Radon-Nikodym tilt in Bayesian generalized linear models, the mechanism the present paper isolates.","marker":"[23]"},{"why":"Provides the Sherrington-Kirkpatrick background and replica-symmetric structure against which the cavity prediction is framed.","marker":"[22]"},{"why":"Gives the spectral concentration of Wigner matrices used to control the operator norm of the interaction matrix in the Poincaré regime.","marker":"[27]"},{"why":"Supplies the Cholesky perturbation bound that controls the distance between finite-$N$ and limiting cavity fields in Lemma 4.","marker":"[24]"},{"why":"Supplies the contiguity and likelihood-ratio perspective underlying the Radon-Nikodym comparison of Gibbs and cavity measures.","marker":"[11]"},{"why":"Provides the Gauss-Poincaré inequality used to bound the disorder variance of the overlap in Lemma 9.","marker":"[10]"}],"fun_headline_variants":["Cavity-contiguity proof sets spin-glass marginals at N^-1/8","Non-interpolation cavity proof for SK marginals at high temp","Spin-glass marginals via cavity-contiguity: rate N^-1/8","Cavity-contiguity: modular proof for SK marginals","Weak-Gaussianity cavity method yields SK marginals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of overlap concentration assumes that the log-Sobolev inequality of [9], stated for positive-semidefinite interaction matrices, applies to the two-replica measure with interaction matrix $\\gamma\\,\\mathrm{diag}(J,J)$ built from the indefinite Wigner matrix $J$; if that transfer fails, the Poincaré inequality (9) and the concentration that supports Theorem 1 collapse.","fun_headline_variants_meta":{"raw":{"variants":["Cavity-contiguity proof sets spin-glass marginals at N^-1/8","Non-interpolation cavity proof for SK marginals at high temp","Spin-glass marginals via cavity-contiguity: rate N^-1/8","Cavity-contiguity: modular proof for SK marginals","Weak-Gaussianity cavity method yields SK marginals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2257,"prompt_tokens":992,"completion_tokens":1265,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1170}},"tokens_in":608,"tokens_out":1265,"duration_ms":10516,"temperature":1.0,"reasoning_tokens":1170,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:09:29.017097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed $\\beta<1/4$, $h\\neq 0$ (for instance $\\beta=0.24$, $h=1$), simulate the Sherrington-Kirkpatrick Gibbs measure at moderate $N$, and estimate $E\\langle (Q_{12}-q)^2\\rangle_c$. If the variance does not decay like $C/N$—or if the Poincaré inequality (9) fails explicitly for a test function under the indefinite two-replica measure—then Proposition 1, and with it Theorem 1, cannot hold as stated.","supporting_citations":[{"cited_title":"A very simple proof of the lsi for high temper- ature spin systems.Journal of Functional Analysis, 276(8):2582–2588, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the log-Sobolev inequality from which the Poincaré inequality (9) is derived to prove overlap concentration in Proposition 1."},{"cited_title":"Springer Science & Business Media, 2013","cited_arxiv_id":null,"evidence_quote":"Provides the Sherrington-Kirkpatrick background and replica-symmetric structure against which the cavity prediction is framed."},{"cited_title":"Perturbation bounds for the cholesky and qr factorizations.BIT Numerical Mathematics, 31(2):341–352, 1991","cited_arxiv_id":null,"evidence_quote":"Supplies the Cholesky perturbation bound that controls the distance between finite-$N$ and limiting cavity fields in Lemma 4."},{"cited_title":"Springer Series in Statistics","cited_arxiv_id":null,"evidence_quote":"Supplies the contiguity and likelihood-ratio perspective underlying the Radon-Nikodym comparison of Gibbs and cavity measures."},{"cited_title":"Concentration inequalities using the entropy method.The Annals of Probability, 31(3):1583–1614, 2003","cited_arxiv_id":null,"evidence_quote":"Provides the Gauss-Poincaré inequality used to bound the disorder variance of the overlap in Lemma 9."}],"review_version":1}