{"id":"6584dd11-b721-4306-9062-ecd66a89ab4c","arxiv_id":"2412.12720","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A tensorized stochastic localization process with non-Gaussian tilts yields explicit spectral gap bounds and rapid mixing for tensor Ising models.","lead":"This paper builds a new version of stochastic localization that works with non-Gaussian tilts, and uses it to prove explicit bounds on how fast Glauber dynamics mixes for tensor Ising models with four-way interactions. The bounds identify a concrete high-temperature regime where mixing is rapid, and a separate tensor Curie-Weiss example where it is exponentially slow.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The bounded-drift lemma (Prop. 3.3, Lemma A.1, Eq. (A.2)) is internally inconsistent: the printed sign of v_t makes the boundary repulsive, so the a.s. bound on X_t is unproven, and Theorem 3.1 and the final spectral gap proof depend on it.","rationale":"The reader's weakest_assumption correctly identified Proposition 3.3 as the most fragile part of the technical machinery. My stress-test confirms this and locates a concrete defect: the sign inconsistency in Eq. (A.2) means the proof of Proposition 3.3, as written, does not establish the boundedness of X_t. This matters because Theorem 3.1's decomposition writes µ = E[µ_τ] with leftover R = X_τ, and Theorem 1.1's small ψ(x) term is obtained by applying Proposition 3.3. Without a valid bound ‖R‖ ≤ δ, the ψ term in the decomposed measures cannot be made small, and the final Dobrushin bound in the proof of Theorem 1.2 — the term 56n‖T‖_inj — has no basis. I nevertheless recommend CONDITIONAL rather than REJECT because the defect has the shape of a fixable sign/constant error: the confining drift v_t = +4n X_t/(δ − ‖X_t‖²) is a standard Bessel-type construction, and a corrected Lemma A.1 would likely restore all subsequent steps. The other issues noted by the reader (the unstated identity shift in applying Theorem 1.5 to a negative definite quadratic form, and the tension between positive definiteness and zero diagonal entries in Theorem 1.1/1.2) are real but secondary: they are readily repaired by the standard hypercube constant shift and do not threaten the high-level strategy. The paper's decomposition framework is novel and the numerical/structural claims about tensor Curie-Weiss are independently supported by the companion work cited as [38]. The required revision is therefore to supply a correct, internally consistent proof of Proposition 3.3, and to fix the statement of the positivity assumptions, rather than to abandon the central claim.","tokens_in":103,"tokens_out":11808,"duration_ms":410742,"concrete_test":"Independently re-derive Lemma A.1 under the two possible sign conventions. Case (i): take v_t exactly as printed in Eq. (A.2), v_t = −4n X_t/(δ − ‖X_t‖²), and compute the Itô drift of ‖X_t‖² via dX_t = C_t dB_t − C_t² v_t dt; verify that the drift is positive for ‖X_t‖ near δ, so the process can reach the boundary and the claimed a.s. bound fails. Case (ii): take the confining sign v_t = +4n X_t/(δ − ‖X_t‖²), recompute the drift of (δ − ‖X_t‖²)^{-1}, and check whether the supermartingale inequality used in the proof holds with correct constants. The sign convention that makes the supermartingale argument valid must be stated explicitly; the current text does not do so.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Proposition 3.3 is the keystone of the decomposition: it provides the adapted drift v_t so that X_t = ∫ C dW − ∫ C² v ds satisfies sup_t ‖X_t‖_HS ≤ δ. This X_t is exactly the leftover term R in the proof of Theorem 3.1, whose smallness controls the ψ(x) term in Theorem 1.1 and hence the 56n‖T‖_inj coefficient in Theorem 1.2 via the Dobrushin bound. The proof of Proposition 3.3 rests on Lemma A.1, but Eq. (A.2) as printed defines v_t := −4n X_t/(δ − ‖X_t‖²) and then asserts dX_t = C_t dB_t − 4n C_t² X_t/(δ − ‖X_t‖²) dt. Substituting the printed v_t into the defining relation dX_t = C_t dB_t − C_t² v_t dt gives dX_t = C_t dB_t + 4n C_t² X_t/(δ − ‖X_t‖²) dt, an outward, repulsive drift toward the boundary ‖X_t‖ = δ. The subsequent Itô computation, which produces a confining drift −4n‖C_t X_t‖²/(δ − ‖X_t‖²) dt in d‖X_t‖², corresponds to the opposite sign, v_t := +4n X_t/(δ − ‖X_t‖²), and even then the constants in the Itô expansion do not match (a factor of 2 discrepancy in the drift of ‖X_t‖²). Thus the appendix does not actually prove the existence of the bounded process used in the main construction. If this sign error is corrected, the intended Bessel-type argument is plausible; but the current manuscript does not contain the corrected derivation, so the central theorem's proof is incomplete as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends stochastic localization to non-Gaussian tilts by introducing a tensorized process driven by Dyson Brownian motion, with the tilt applied to x^\\otimes 2 rather than to x. The main structural result is Theorem 1.1, which decomposes a quartic potential on the Boolean hypercube into a mixture of low-rank measures of the form exp(\\langle u,x\\rangle^2\\langle v,x\\rangle^2 + \\langle x,w\\rangle^2 + \\langle \\ell,x\\rangle + \\psi(x)) plus a small uniformly bounded perturbation \\psi. This is then used to prove Theorem 1.2, an explicit Poincar\\'e-constant bound for tensor Ising models with \\|T\\|_{inj} \\le 1/(56n), yielding rapid mixing for Glauber dynamics. The paper also gives a slow-mixing lower bound for the tensor Curie-Weiss model (Proposition 5.1, Lemma 1.3) and sketches extensions to tensors of higher degree.","tokens_in":31599,"tokens_out":11476,"duration_ms":103896,"significance":"If correct, the result provides the first fully explicit spectral-gap bound for quartic spin systems under Glauber dynamics, replacing the implicit constant obtained from spectral independence with the explicit constant 1/56. The proof is constructive and first-principles: the constants are not fitted, and the decomposition theorem is a genuine extension of the rank-one decomposition of [22] to quartic potentials. Corollary 1.4 gives a quantitative high-temperature bound for degree-4 spin glasses. The main technical novelty is the bounded-drift construction that controls the non-Gaussian remainder term, and this is also the most fragile part of the paper. The lower-bound example in Section 5 is a useful sanity check showing that the high-temperature regime cannot be substantially enlarged without additional ideas.","major_comments":[{"comment":"The bounded-process construction used in Proposition 3.3 is internally inconsistent. Lemma A.1 states dX_t = C_t dB_t + C_t^2 v_t dt and defines v_t := -4n X_t/(\\delta - \\|X_t\\|^2); substituting gives dX_t = C_t dB_t - 4n C_t^2 X_t/(\\delta - \\|X_t\\|^2) dt. Proposition 3.3 instead defines X_t by dX_t = C_t dW_t - C_t^2 v_t dt, so with the printed v_t the drift would be +4n C_t^2 X_t/(\\delta - \\|X_t\\|^2) dt, which is repulsive rather than confining. The displayed It\\^o computation for d\\|X_t\\|^2 also has a factor-of-two mismatch: the coefficient should be 8n, not 4n, and the later expression for df(Y_t) does not follow from the preceding formula. Since X_t is exactly the remainder R in the proof of Theorem 3.1 and controls the smallness of \\psi(x) in Theorem 1.1 and hence the 56n\\|T\\|_{inj} coefficient in Theorem 1.2, this is a load-bearing gap in the proof as written. The argument appears repairable by choosing v_t with the opposite sign and redoing the Bessel-type estimate, but the corrected derivation is not present. In addition, Proposition 3.3 concerns matrix-valued X_t, whereas Lemma A.1 is stated for vectors in R^n; the state-space dimension should be n^2 in the matrix case.","section":"Appendix A.2, Lemma A.1 and Eq. (A.2)"},{"comment":"Theorem 1.5 is invoked for the measure \\bar{\\mu}_M(x) \\propto \\exp(-2n\\sqrt{\\|T\\|_{inj}}\\langle x,\\tilde{M}x\\rangle), i.e. for a quadratic form that is negative semidefinite, whereas Theorem 1.5 requires the quadratic form in the exponent to be positive definite. The application can be repaired by using the fact that \\|x\\|_2^2 = n is constant on C_n and shifting \\tilde{M} by a suitable multiple of the identity, but this step is not stated. As written, the derivation of the decomposition (3.18) is not justified.","section":"Section 3.3, proof of Theorem 1.1 before Eq. (3.18)"},{"comment":"The standing assumption of Theorem 1.1 is inconsistent. The theorem requires T to be a positive definite symmetric fourth-order tensor with zero diagonal entries. Positive definiteness is defined in Section 2.2 in the n^2 \\times n^2 matrix sense, while Section 2.1 defines a diagonal entry as a multi-index with any repetition, so T_{ijij} = 0 for all i,j. Under this definition all diagonal entries of the n^2 \\times n^2 matrix vanish, which is impossible for a positive definite matrix. Thus the theorem as stated is vacuous. The statement should either redefine \"zero diagonal\" (for example, only T_{iiii}=0) or be formulated for positive semidefinite T with an explicit recentering; the proofs of Theorem 1.2 and Corollary 1.4 already rely on such recentering steps.","section":"Theorem 1.1 and Section 2.1"}],"minor_comments":[{"comment":"The displayed convergence \"P(|n^{p/2-1}\\|T\\|_{inj} - E_0(p)| \\le \\varepsilon) \\to 0\" should be convergence to 1; as printed it contradicts the concentration statement and its use in Corollary 1.4.","section":"Theorem 2.3"},{"comment":"The line \"\\mu(x) \\propto (\\beta \\tilde{T}(x))\" is missing the exponential; it should be \\mu(x) \\propto \\exp(\\beta \\tilde{T}(x)).","section":"Proof of Corollary 1.4"},{"comment":"The bound \\|D_{\\psi}\\|_{op} \\le n\\delta is off by a constant factor: since each entry of D_{\\psi} is at most 4\\delta, a more natural bound is 4n\\delta. Because \\delta is arbitrary, this does not change the final conclusion.","section":"Proof of Theorem 1.2, Section 4.2"},{"comment":"The iteration over the decreasing sequence of image subspaces is only sketched in the final paragraph of Appendix A.2; if the base Bessel-type estimate is corrected, the stopping-time induction and the behavior of the process at degeneracy times should be spelled out in detail.","section":"Proof of Proposition 3.3"}],"recommendation":"major_revision","confidential_remarks":"The sign and constant inconsistencies in the bounded-drift lemma are the kind of issue that may be a repairable typo, but they sit at the technical core of the paper: Proposition 3.3 controls the remainder term in the decomposition, which in turn controls the main spectral-gap constant. I see no evidence of circularity or fitted constants, and the overall strategy is promising, but the manuscript cannot be accepted until the appendix construction is corrected and the inconsistent statement of Theorem 1.1 is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper extends stochastic localization to non-Gaussian tilts in a way that yields the first explicit spectral-gap constant for tensor Ising models. The central argument holds up, and the novel machinery is real. The tensorized SL process, the removal of the barycenter, and the drift that suppresses the dangerous linear term are all genuinely new constructions, and the decomposition of quartic potentials into low-rank pieces is clever. The explicit 1/56 constant is a concrete improvement over the implicit bounds from [3].\n\nIt does what it claims. The iterative decomposition with normalized variance for unnormalized measures is carefully handled, and the Dobrushin-based spectral gap argument is clean. The tensor Curie-Weiss lower bound and the spin-glass corollary are reasonable applications, and the comparison with [3] is honest.\n\nThe soft spots are real but mostly cosmetic. Theorem 1.1 assumes T is positive definite with zero diagonal entries, which is impossible for a positive definite matrix; this needs a fix, likely replacing it with a shifted or semidefinite version. The application of Theorem 1.5 to a negative definite quadratic form in the proof of Theorem 1.1 is not justified in the text; a standard hypercube shift probably repairs it but should be stated. Proposition 3.3 and Lemma A.1 use opposite sign conventions for v_t; they are consistent under a sign flip, but as written it is confusing and will trip readers. There is also a factor-of-2 typo in the Itô calculation in Lemma A.1: the drift coefficient should be -8n, not -4n. The sign remains confining, so the Bessel-type argument still goes through, but the constants need correction.\n\nOn the stress-test claim: it does not hold up on reading. Substituting the printed v_t into the printed SDE in Lemma A.1 (which has a plus sign) gives a confining drift toward the origin, not a repulsive one. The bounded-drift lemma is therefore plausible, and the main decomposition theorem is not undermined.\n\nThis paper is for researchers in stochastic localization, Markov chain mixing, and spin glasses. It deserves a serious referee. I would send it out, asking for fixes to Theorem 1.1's statement, the negative-definite application, and the sign/constant conventions in the appendix. None of these look load-bearing, but they must be cleaned up before the paper is citable as a reference.","headline":"A genuinely new localization tool with explicit tensor Ising bounds; the central proof survives the stress-test, though the statement needs cleanup.","tokens_in":32102,"tokens_out":4213,"would_cite":true,"duration_ms":35594,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60J27","82B20","60H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper extends stochastic localization to non-Gaussian tilts and proves that tensor Ising measures satisfy the explicit bound $C_P(\\mu) \\le 1/(1-56n\\|T\\|_{\\mathrm{inj}})$, giving rapid mixing for Glauber dynamics whenever…","keywords":["stochastic localization","non-Gaussian tilts","tensor Ising model","spectral gap","Glauber dynamics","Poincaré inequality","mixing time","Boolean hypercube"],"falsifier":"Take the rank-one tensor $T(x) = \\beta(\\sum_i x_i)^4/n^3$ on $\\{\\pm1\\}^n$ with $\\beta = 0.016$, so that $\\|T\\|_{\\mathrm{inj}} = 0.016/n$ is below $1/(56n)$, and compute the true Poincaré constant of Glauber dynamics by exact diagonalization for $n$ up to around 20. If the constant exceeds $1/(1 - 56n\\|T\\|_{\\mathrm{inj}}) \\approx 9.6$ at any such $n$, Theorem 1.2 is false; if it stays below, the bound survives this direct test.","tokens_in":30979,"feed_emoji":"🧲","tokens_out":12339,"duration_ms":101284,"temperature":0.7,"pith_summary":"The paper asks whether stochastic localization, a technique that simplifies a complicated probability measure by tilting it with Gaussians, can work when the tilt is not Gaussian, and whether that helps for genuinely non-quadratic models. It answers yes for quartic potentials on the Boolean hypercube: a new tensorized localization process decomposes any tensor Ising measure into a mixture of simple low-rank pieces, up to an arbitrarily small error. From that decomposition it proves an explicit variance bound, $C_P(\\mu) \\le 1/(1 - 56 n \\|T\\|_{\\mathrm{inj}})$ whenever $\\|T\\|_{\\mathrm{inj}} \\le 1/(56n)$. Since the Poincaré constant is the reciprocal of the spectral gap of Glauber dynamics, this yields rapid mixing throughout a high-temperature regime with explicit constants. The interest is that explicit, checkable spectral-gap bounds for non-quadratic spin models were previously available only with implicit constants or in special cases.","feed_headline":"Quartic spin systems mix fast below an explicit threshold","feed_subtitle":"Non-Gaussian localization yields spectral gap 1 − 56n‖T‖ and rapid Glauber mixing for tensor Ising models.","key_machinery":"The load-bearing object is the tensorized stochastic localization process: a continuous-time evolution of relative densities $dF_t(x) = \\langle x^{\\otimes 2} - v_t, C_t\\, dW_t\\rangle F_t(x)$, where $W_t$ is a Brownian motion in the space of symmetric matrices, $C_t$ is a smoothed projection onto the image of the remaining tensor potential $T_t = T - \\tfrac12\\int_0^t C_s^2\\, ds$, and $v_t$ is an adapted drift chosen so that the process $X_t$ solving $dX_t = C_t\\, dW_t - C_t^2 v_t\\, dt$ stays inside a ball of radius $\\delta$. This drift cancels the unwanted quadratic tilt that would otherwise destroy the spectral gap, and the projection keeps the normalized variance nearly constant along the flow. Stopping when the remaining tensor has rank one converts the measure into a mixture of low-rank quartic measures, and the error $\\delta$ is then absorbed into the final bound.","core_discovery":"The central claim, stated as Theorem 1.2, is that for a measure $\\mu(x) \\propto \\exp(T(x))$ on $\\{\\pm 1\\}^n$ with $T$ a positive definite symmetric fourth-order tensor, the Poincaré constant obeys $C_P(\\mu) \\le 1/(1 - 56 n \\|T\\|_{\\mathrm{inj}})$ provided $\\|T\\|_{\\mathrm{inj}} \\le 1/(56 n)$. The argument runs through a decomposition theorem: the measure is split, up to an arbitrarily small error, into a mixture of non-negative measures of the form $\\exp(\\langle u,x\\rangle^2 \\langle v,x\\rangle^2 + \\langle w,x\\rangle^2 + \\langle \\ell,x\\rangle + \\psi(x))$, with the vectors bounded by $2\\sqrt{\\|T\\|_{\\mathrm{inj}}}$ or $4n\\|T\\|_{\\mathrm{inj}}$ and $|\\psi| \\le \\delta$ almost surely. Each low-rank piece is shown to satisfy a variance bound by a derivative-matrix version of the influence-matrix criterion, and concavity of the Glauber Dirichlet form reassembles the pieces. The same machinery yields an explicit high-temperature bound for Gaussian degree-4 spin glasses and a transition statement for the rank-one tensor Curie-Weiss model, where mixing becomes exponentially slow above $\\beta \\approx 0.504$.","pith_inferences":["The explicit constant $1/56$ is likely improvable: the rank-one tensor Curie-Weiss example suggests the true threshold is about 28 times larger, so a sharper influence-matrix bound for quartic forms would immediately improve Theorem 1.2.","The bounded-drift construction is the fragile step; if a more robust or simpler way to keep the localization path inside a small ball were found, the method would probably extend to other non-quadratic potentials beyond quartic tensors.","The decomposition is not tied to spectral gaps: it supplies a generic transfer principle, so inequalities of any kind proved on the low-rank pieces lift to the original high-temperature measure.","For higher degrees, the rank of the decomposition grows so fast that the method's guarantees deteriorate; a different stopping rule or a coarser decomposition would be needed to keep spectral-gap bounds competitive in that regime."],"forward_implications":["For every positive definite degree-4 tensor with $\\|T\\|_{\\mathrm{inj}} \\le 1/(56n)$, Glauber dynamics on $\\{\\pm1\\}^n$ has spectral gap at least $1 - 56n\\|T\\|_{\\mathrm{inj}}$ and therefore polynomial mixing time.","For Gaussian degree-4 spin glasses with $N(0,1/n^3)$ entries, the same route gives $C_P(\\mu) \\le 1/(1 - 200.928\\,\\beta)$ at inverse temperatures up to about $1/200.928$.","The rank-one tensor Curie-Weiss model is covered on both sides: rapidly mixing for $\\beta \\le 1/56$ and exponentially slow above $\\beta \\approx 0.504$, so the explicit constant is within roughly a factor of 28 of the true threshold.","The decomposition transfers to the sphere and to higher-degree tensors, with the spectral-gap guarantee degrading rapidly in the degree.","Because the decomposition reduces arbitrary tensors to low-rank pieces, any future modified log-Sobolev or mixing result proved for the low-rank pieces would automatically extend to the full tensor Ising model."],"supporting_citations":[{"why":"supplies the smoothed projection lemma and the rank-one decomposition scheme that the new tensorized process extends.","marker":"[22]"},{"why":"gives the influence-matrix condition that converts derivative-matrix bounds into the explicit spectral gap estimate.","marker":"[12]"},{"why":"provides the previous implicit-constant spectral gap for small-Hessian measures, the comparison the explicit constant improves on.","marker":"[3]"},{"why":"sets up the localization-scheme framework and the Dirichlet-form supermartingale fact used in the variance decomposition.","marker":"[10]"},{"why":"supplies concentration of the injective norm for Gaussian tensors, used for the spin-glass corollary.","marker":"[5]"},{"why":"gives the numerical value $E_0(4)\\approx 1.794$ used to compute the spin-glass constant 200.928.","marker":"[11]"},{"why":"gives the symmetry characterization of the injective norm used throughout the paper.","marker":"[6]"},{"why":"independent study of the tensor Curie-Weiss model whose lower bound the paper compares with its own slow-mixing result.","marker":"[38]"}],"fun_headline_variants":["Fast mixing in tensor Ising below explicit threshold","Non-Gaussian tilts speed up tensor Ising mixing","Explicit Poincare bound for quartic spins","Tensor Ising rapid mixing via localization","Quartic spins mix fast under 1/56n"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on the existence of an adapted drift $v_t$ that keeps the auxiliary process $X_t$ inside a ball of radius $\\delta$ forever; if that drift construction fails, the leftover term in the decomposition is no longer small and the spectral gap bound collapses.","fun_headline_variants_meta":{"raw":{"variants":["Fast mixing in tensor Ising below explicit threshold","Non-Gaussian tilts speed up tensor Ising mixing","Explicit Poincare bound for quartic spins","Tensor Ising rapid mixing via localization","Quartic spins mix fast under 1/56n"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001097,"raw_usage":{"total_tokens":4562,"prompt_tokens":914,"completion_tokens":3648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":3575}},"tokens_in":530,"tokens_out":3648,"duration_ms":25974,"temperature":1.0,"reasoning_tokens":3575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:49:41.295234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the rank-one tensor $T(x) = \\beta(\\sum_i x_i)^4/n^3$ on $\\{\\pm1\\}^n$ with $\\beta = 0.016$, so that $\\|T\\|_{\\mathrm{inj}} = 0.016/n$ is below $1/(56n)$, and compute the true Poincaré constant of Glauber dynamics by exact diagonalization for $n$ up to around 20. If the constant exceeds $1/(1 - 56n\\|T\\|_{\\mathrm{inj}}) \\approx 9.6$ at any such $n$, Theorem 1.2 is false; if it stays below, the bound survives this direct test.","supporting_citations":[{"cited_title":"A sp ectral condition for spectral gap: fast mixing in high-temperature Ising models","cited_arxiv_id":null,"evidence_quote":"supplies the smoothed projection lemma and the rank-one decomposition scheme that the new tensorized process extends."},{"cited_title":"The Description of a Random Field by M eans of Conditional Probabil- ities and Conditions of Its Regularity","cited_arxiv_id":null,"evidence_quote":"gives the influence-matrix condition that converts derivative-matrix bounds into the explicit spectral gap estimate."},{"cited_title":"Universality of Spectral Independence with Applications to Fast Mixing in Spin Glasses","cited_arxiv_id":null,"evidence_quote":"provides the previous implicit-constant spectral gap for small-Hessian measures, the comparison the explicit constant improves on."},{"cited_title":"Localization Schemes: A F ramework for Proving Mixing Bounds for Markov Chains","cited_arxiv_id":null,"evidence_quote":"sets up the localization-scheme framework and the Dirichlet-form supermartingale fact used in the variance decomposition."},{"cited_title":"Random Matrices and Complex- ity of Spin Glasses","cited_arxiv_id":null,"evidence_quote":"supplies concentration of the injective norm for Gaussian tensors, used for the spin-glass corollary."},{"cited_title":"Polynomials and multi linear mappings in topological vector-spaces","cited_arxiv_id":null,"evidence_quote":"gives the symmetry characterization of the injective norm used throughout the paper."}],"review_version":1}