{"id":"fe692296-6403-48c1-81ee-41edd5ca463a","arxiv_id":"1908.01520","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For interacting oscillators on graphs whose normalized adjacency matrix is close to the all-ones matrix in the infinity-to-one norm, the empirical measure follows the McKean-Vlasov equation and stays near its stable stationary states for times up to exp(o(n)).","lead":"This paper proves that a large class of networks, with growing average degree, behave like the fully connected mean-field Kuramoto model of synchronized oscillators, even at exponentially long times. It gives a clean mathematical condition on the network structure that guarantees this behavior, which is relevant to any field that models many coupled units on a graph.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.3's representation (3.45)-(3.46) has a scaling inconsistency; the resulting exponential noise bound (3.41), which sets the exp(o(n)) time scale, is not justified as written.","rationale":"The central claim is that condition (1.16) yields a finite-time mean-field limit and long-time closeness to the stable manifold or the uniform measure up to times T_n = exp(o(n)). The proof strategy is coherent, and the examples in Appendix A are consistent with the condition. The abstract's phrase 'as fast as the exponential of the size' overstates the theorem, which proves exp(o(n)), as the Reader noted. The most load-bearing element, however, is the stochastic noise estimate (3.41), because the long-time block argument simply raises 1 - exp(-Z n ε^2) to the power N_n = exp(o(n)). That estimate is proved in Lemma 3.3, and the displayed computation (3.45)-(3.49) contains a scaling inconsistency that invalidates the estimate as written. The concern is about the manuscript's proof rather than the truth of the theorem: a direct exponential martingale bound for each Fourier mode appears capable of restoring (3.41), so the appropriate verdict is CONDITIONAL rather than REJECT. This differs from the Reader's weakest assumption, which focused on the imported bounds B.2/B.3; those bounds are plausible and would only need verification, whereas the algebra in (3.45)-(3.49) can be checked immediately from the text. No other part of the argument—the Grothendieck application modulo normalization constants, the projection Lemma 3.5, or the subcritical maximal inequality—independently threatens the central claim.","tokens_in":30356,"tokens_out":23080,"duration_ms":232206,"concrete_test":"Substitute (3.46) into the right-hand side of (3.45) and compare with the integral definition of z^n_t(f^ψ_l); this settles the algebra. If the corrected identity is z^n_t(f^ψ_l) = c^2 e^{-λ_l t} A_t / √(2λ_l), redo the probability calculation (3.49)-(3.51) with the corrected threshold 2λ_l η / c^4. For (3.41) to hold, the final exponential rate must retain a factor of n uniformly over l ≥ 1. As a second check, prove the mode concentration directly from the quadratic variation ⟨z^n(f^ψ_l)⟩_t ≤ c^2(1 - e^{-2λ_l t})/(2λ_l n) and verify that the resulting tail matches the claimed exp(-c' λ_l n η^2) rate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Lemma 3.3 is the sole place where the stochastic perturbation z^n_t is converted into the exponential estimate (3.41), and Theorems 2.3/2.4 inherit the T_n = exp(o(n)) time scale from it. The displayed computation is internally inconsistent. With g_j(s) = ∂_θ f^ψ_l(θ^{j,n}_s), the definition gives z^n_t(f^ψ_l) = e^{-λ_l t} n^{-1} Σ_j ∫_0^t e^{λ_l s} g_j(s) dB^j_s. Substituting the A_t of (3.46) into (3.45) yields z^n_t(f^ψ_l) = c^2 e^{-λ_l t} A_t / √(2λ_l), not c e^{-λ_l t} √(2λ_l n) A_t. The printed equality is off by a factor depending on n and λ_l. Consequently the event threshold in (3.49) is wrong: under the corrected identity, |z^n_t(f^ψ_l)|^2 > η is e^{-2λ_l t} A_t^2 > 2λ_l η / c^4, while (3.49) uses 2λ_l n η / c^2. The factor n in the exponent of (3.51), and hence in the global bound (3.41), is not established by the displayed argument. A direct exponential martingale bound using the mode's quadratic variation, at most c^2(1 - e^{-2λ_l t})/(2λ_l n), may repair the result, but that proof is not in the manuscript. This is more load-bearing than the unit-ball normalization in Lemma 3.2, which only changes universal constants.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the stochastic Kuramoto model on a sequence of graphs with interaction strength normalized by n p_n. It proposes a deterministic condition on the graph sequence, namely that the ℓ∞→ℓ1 norm of P^(n) - 1^(n) is o(n^2) (condition (1.16)), and proves three main results: Theorem 2.1 gives finite-time closeness of the empirical measure to the McKean-Vlasov limit, without requiring independence between initial conditions and the graph; Theorem 2.3 shows that in the supercritical regime K>1 the empirical measure stays close to the stable manifold M of stationary solutions up to times T_n = exp(o(n)); Theorem 2.4 shows that in the subcritical regime K<1 the empirical measure stays close to the uniform measure on the same exponential time scale. The proofs are based on a mild formulation of the empirical measure, control of the graph-induced perturbation through Grothendieck's inequality, and control of the noise through exponential and maximal inequalities for martingales. Appendix A shows that Erdős-Rényi graphs with diverging average degree and sequences of Ramanujan graphs satisfy condition (1.16).","tokens_in":30690,"tokens_out":15191,"duration_ms":141234,"significance":"If the technical gap identified below is repaired, this is a substantial contribution. The paper gives a clean deterministic graph condition, expressed in terms of a well-studied norm, that is sufficient for mean-field behavior without any independence assumption between the initial empirical measure and the graph, and it goes beyond the existing literature by pushing the synchronization statement to nearly exponential time scales in both the subcritical and supercritical regimes. Strengths of the paper include the explicit nature of condition (1.16), the concrete families of graphs covered (ER and Ramanujan), the connection to the cut-norm and graphon framework, and the self-contained treatment of the weighted Sobolev spaces and semigroup estimates in the appendices. The proof architecture is clearly presented and the central claims are plausible, provided the stochastic estimate in Lemma 3.3 is corrected.","major_comments":[{"comment":"The scaling of the stochastic coefficient is inconsistent. From (3.44), z_t^n(f_l^ψ) = e^{-λ_l t} n^{-1} Σ_j ∫_0^t e^{λ_l s} [∂_θ f_l^ψ](θ^{j,n}_s) dB^j_s. Substituting the definition of A_t in (3.46) into (3.45) gives z_t^n(f_l^ψ) = c^2 e^{-λ_l t} A_t / √(2λ_l), not c e^{-λ_l t} √(2λ_l n) A_t as printed. Consequently the event on the right-hand side of (3.49) should be e^{-2λ_l t} A_t^2 > 2λ_l η / c^4, not 2λ_l n η / c^2, and the factor n in the exponent of (3.51), and hence in the global exponential bound (3.41), is not established by the displayed computation. Since (3.41) is used in (3.73)-(3.74) to obtain convergence for N_n = exp(o(n)), this is a load-bearing gap in the proof of Theorem 2.3. The gap appears repairable: each z_t^n(f_l^ψ) is a continuous martingale with quadratic variation bounded by c^2(1-e^{-2λ_l t})/(2λ_l n), so the standard exponential martingale inequality yields a tail of the same form with the factor n; the authors should replace the self-normalized computation by this direct estimate, or correct the normalization consistently. Additionally, with A_t defined as in (3.46), the coefficient in the quadratic variation formula (3.47) should be 2λ_l/(c^4 n^2), not 2λ_l/(c^2 n).","section":"Section 3.2, Lemma 3.3"},{"comment":"Grothendieck's inequality is invoked without verifying that the vectors S_i and T_j lie in the unit ball of the Hilbert space H^{-1}. In particular, for T_j = √(t-s)/C (J*δ_{θ_j}) ∂_θ e^{(t-s)L*_ψ} h / ||h||_1, the required uniform bound ||T_j||_{-1} ≤ 1 is not proved; the same issue occurs in Lemma 4.3 with the T_j defined in (4.13). A verification is possible: using ||δ_θ||_{-1} ≤ C, ||J*δ_θ||_∞ ≤ K, and the semigroup estimates in Propositions B.2-B.3, one obtains ||∂_θ e^{rL*_ψ} h||_2 ≤ C r^{-1/2} ||h||_1 (up to constants), so the constant C in the definition of T_j can be enlarged to ensure the unit-ball condition uniformly in j and in t-s. This step should be stated explicitly, because as written the bound (3.37) is asserted rather than derived.","section":"Section 3.2"}],"minor_comments":[{"comment":"The factor 2/3 ε in the displayed bound is not consistent with the initial bound ε/2 in the definition of A_n^1 in (3.58). Applying Lemma B.4 with δ = ε/2 would give a factor 3ε/2 before adding the perturbation terms. The constants should be adjusted, for example by taking the initial bound in A_n^1 to be ε/6, so that the final estimate (3.68) yields sup ||ν^n_t|| ≤ ε.","section":"Section 3.3"},{"comment":"In the statement of the lemma and in the proof, the right-hand side C log(1 + log(1 + ⟨X⟩_t)) depends on t, while the left-hand side is a supremum over t in [0,T]. The bound should be stated with the supremum over t on the right, or with ⟨X⟩_T, and the proof adjusted accordingly.","section":"Section 4.2"},{"comment":"The mild equation uses the Laplacian Δ/2, and the text says the results about L_{2π} will be used. This is correct because Δ/2 has the same type of semigroup estimates with spectral gap 1/2, but this should be stated explicitly to avoid confusion.","section":"Section 5"},{"comment":"There are minor typographical inconsistencies between 'longtime' and 'long time' in section headings (e.g., Sections 2.2 and 2.3), and a few instances of missing articles or slightly awkward phrasing. These do not affect the mathematics but should be cleaned up in a revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the scaling error in Lemma 3.3, which undermines the proof of the exponential noise bound (3.41) as written. However, the repair is direct and does not change the architecture of the proof: the standard exponential martingale inequality applied to each Fourier mode gives the needed factor n. I therefore do not recommend rejection, but the manuscript must be revised to correct this estimate and to add the missing unit-ball checks in Lemmas 3.2 and 4.3. The remaining parts of the proof appear sound, and the paper makes a significant contribution to the literature on mean-field systems on graphs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline first: this is a solid paper with one genuine gap. The main results are new—finite-time mean-field limit for the stochastic Kuramoto model on graphs without independence between initial data and graph, and long-time closeness to the stable manifold up to exp(o(n)) times, which is genuinely beyond the polynomial times in [4]. The abstract says 'exponential of the size,' but the theorems give exp(o(n)), which is subexponential; that's a wording slip, not a mathematical one.\n\nWhat it does well: the deterministic graph condition (1.16) in the ℓ∞→ℓ1 norm is clean, and the examples (Erdős–Rényi with diverging average degree, Ramanujan graphs) support it. The mild-formulation framework is appropriate, and the subcritical maximal inequality in Corollary 2.5 is a nice byproduct. The paper is honest about what it imports from [4].\n\nSoft spots, in proportion. The Grothendieck step in Lemmas 3.2 and 4.3 does not verify that the chosen vectors Tj lie in the unit ball of H^{-1}; that is a minor constant issue, repairable in a footnote. The bigger problem is Lemma 3.3. The rewriting of z^n_t(f^ψ_l) in (3.45)-(3.46) is algebraically wrong: substituting the displayed A_t into (3.45) does not reproduce the definition (3.44); the powers of n and λ_l are off. Consequently the threshold in (3.49) and the exponential bound (3.51) do not follow as written. Since (3.41) is what produces the exp(o(n)) time scale in Theorems 2.3 and 2.4, that is load-bearing. The result is very likely repairable—a direct exponential martingale bound using the mode's quadratic variation, which is at most c^2(1-e^{-2λt})/(2λn), yields the same nη^2 exponent—but that argument is not in the manuscript.\n\nBottom line: the finite-time result and the graph condition are likely to be useful regardless; the long-time theorems are the headline but need a corrected noise estimate. This deserves peer review, not a desk rejection. A competent referee should catch the scaling error, and a conditional accept after the fix is realistic. I'd hold off citing the long-time results until a corrected version appears.","headline":"New mean-field and long-time results for Kuramoto on graphs, with a real but repairable gap in the stochastic noise bound; the paper deserves serious peer review.","tokens_in":31241,"tokens_out":7132,"would_cite":false,"duration_ms":63408,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","82C20","82C31","82C44"],"pacs":[],"model":"deepseek-v4-flash","headline":"For oscillators coupled through a graph, one explicit norm condition on the adjacency matrix guarantees the classical mean-field limit and controls long-time behavior up to almost exponential times.","keywords":["interacting oscillators","Kuramoto model","mean field limit","long time dynamics","random graphs","cut norm","Grothendieck's inequality","self-normalized processes"],"falsifier":"A numerical falsifier: simulate the finite system on binomial random graphs with average degree $np_n\\to\\infty$, starting half the vertices at phase $0$ and half at phase $\\pi$ so that the initial condition depends on the graph; if for some fixed $T$ the supremum distance $\\sup_{t\\le T}\\|\\mu^n_t-\\mu_t\\|_{-1}$ does not tend to $0$ in probability, then condition (1.16) is not sufficient for Theorem 2.1.","tokens_in":30137,"feed_emoji":"🔄","tokens_out":11766,"duration_ms":116289,"temperature":0.7,"pith_summary":"This paper studies the stochastic Kuramoto model on a sequence of graphs and asks when the network behaves like the classical all-to-all mean-field model. Its central claim is that a single condition on the graph suffices: after normalizing the adjacency matrix by the average number of neighbors, the matrix must approach the all-ones matrix in the $\\ell^\\infty\\to\\ell^1$ (cut) norm, at rate $o(n^2)$. Under this condition the empirical measure of oscillator phases stays close to the solution of the McKean–Vlasov equation on any fixed finite time interval, even when the initial phases are chosen with full knowledge of the network. The paper further shows that on long time scales the same condition keeps the system near its stable stationary states for times that can grow as fast as $\\exp(o(n))$: near the uniform state when the coupling $K<1$, and near the manifold of synchronized states when $K>1$. This matters because real oscillator networks are not complete graphs, and the question of whether the mean-field picture survives on long time scales had been open.","feed_headline":"Cut-norm condition makes graph Kuramoto systems follow mean-field","feed_subtitle":"Under one network condition, empirical measures stay near the mean-field solution for exponentially long times.","key_machinery":"The central object is the normalized adjacency matrix $P^{(n)}=\\xi^{(n)}/p_n$ compared with the all-ones matrix $\\mathbf{1}^{(n)}$ through the $\\ell^\\infty\\to\\ell^1$ norm, a cut-norm-type quantity; the condition $\\|P^{(n)}-\\mathbf{1}^{(n)}\\|_{\\infty\\to 1}=o(n^2)$ is what makes graph fluctuations vanish as $n\\to\\infty$. The proofs work with the empirical measure in the dual Sobolev space $H^{-1}$, whose norm controls the bounded-Lipschitz distance between probability measures, and with the mild SPDE satisfied by $\\mu^n_t$. The two linearized operators $L_\\psi$ and $L_{2\\pi}$, around the synchronized states and the uniform state respectively, supply the contraction and spectral gap that turn noise and graph terms into Gronwall-type estimates. Grothendieck's inequality is the device that bounds the graph term by the $\\ell^\\infty\\to\\ell^1$ norm.","core_discovery":"The paper's discovery is that the relevant measure of network distance is not the degree sequence but the $\\ell^\\infty\\to\\ell^1$ norm of the normalized adjacency matrix $P^{(n)}=\\xi^{(n)}/p_n$ relative to the all-ones matrix. Writing the empirical measure in $H^{-1}$, the graph fluctuation term $g^n_t$ is bounded in norm by $D\\sqrt{t}\\,\\|P^{(n)}-\\mathbf{1}^{(n)}\\|_{\\infty\\to 1}/n^2$ in the supercritical calculation and uniformly in the subcritical case, while the Brownian noise $z^n_t$ is controlled by exponential martingale inequalities. The combination yields Theorem 2.1: with only weak convergence of initial empirical measures and no independence from the graph, $\\sup_{t\\le T}\\|\\mu^n_t-\\mu_t\\|_{-1}\\to 0$ in probability for every fixed $T$. Theorems 2.3 and 2.4 then propagate this closeness for $T_n=\\exp(o(n))$: starting near the manifold of synchronized states ($K>1$) or near the uniform state ($0\\le K<1$), the empirical measure stays within $\\varepsilon$ of the corresponding stationary manifold. The time scale is set by the Brownian noise, not by the network, and coincides with the expected large-deviation barrier.","pith_inferences":["The proof isolates the graph's role entirely through the $\\ell^\\infty\\to\\ell^1$ norm of $P^{(n)}-\\mathbf{1}^{(n)}$, which suggests that any sparse graphon sequence converging to the constant graphon in cut distance should inherit the same finite-time and long-time theorems, even if the graph is not homogeneous.","A natural next question is whether condition (1.16) is also necessary: constructing a graph sequence that violates it only on a small set of vertices and measuring whether the escape time from the synchronized manifold changes would test the sharpness of the condition.","Extrapolating from the complete-graph case, the slow motion along the manifold at times of order $n$ should be Brownian with a diffusion coefficient independent of the graph; tracking the phase of the empirical mean for different graph densities would reveal whether the network affects only the fast transient or also the slow manifold dynamics.","The paper notes its techniques adapt to quenched intrinsic frequencies; if so, one would expect a traveling-wave manifold replacing the fixed phase manifold, with the same $\\exp(o(n))$ escape barrier, and a phase velocity observable at times of order $\\sqrt{n}$."],"forward_implications":["For any graph sequence satisfying the condition $\\|P^{(n)}-\\mathbf{1}^{(n)}\\|_{\\infty\\to 1}=o(n^2)$, including binomial random graphs with diverging average degree and deterministic expanders, the empirical measure converges to the McKean–Vlasov solution on finite time intervals even when initial conditions are chosen with full knowledge of the graph.","In the subcritical regime $0\\le K<1$, the empirical measure remains within $\\varepsilon$ of the uniform law for all times up to $\\exp(o(n))$; for $K=0$ this yields a maximal inequality with growth $\\log(1+T)/n$ for the empirical measure of independent Brownian motions.","In the supercritical regime $K>1$, the empirical measure remains within $\\varepsilon$ of the manifold of synchronized stationary states for times up to $\\exp(o(n))$, and the escape time is not degraded by the network structure.","The graph condition forces a giant component of size $n-o(n)$, so the mean-field limiting behavior cannot coexist with two macroscopic disconnected communities."],"supporting_citations":[{"why":"Establishes the mean-field McKean–Vlasov equation and the spectral properties of the linearized operator for the complete-graph model.","marker":"[3]"},{"why":"Supplies the semigroup and eigenfunction estimates for $L_\\psi$ and $L_{2\\pi}$ that the long-time noise bounds rely on.","marker":"[4]"},{"why":"Provides the self-normalized martingale exponential inequality used to control the Brownian noise term around the synchronized manifold.","marker":"[12]"},{"why":"Classifies the stationary solutions and phase transition of the McKean–Vlasov equation, giving the uniform state for $K<1$ and the manifold of synchronized states for $K>1$.","marker":"[17]"},{"why":"Supplies the maximal inequality for Ornstein–Uhlenbeck processes that the subcritical noise control extends to infinite dimensions.","marker":"[18]"},{"why":"States Grothendieck's inequality, the tool that converts the $\\ell^\\infty\\to\\ell^1$ norm of the normalized adjacency matrix into control of the graph fluctuation term.","marker":"[31]"},{"why":"Prior work applying Grothendieck's inequality and the $\\ell^\\infty\\to\\ell^1$ norm to interacting diffusions on graphs, a technical antecedent of condition (1.16).","marker":"[32]"}],"fun_headline_variants":["Graph norm condition pins Kuramoto empirical measure to mean-field for e^n time","One graph norm makes Kuramoto networks follow mean-field exponentially long","Graph norm condition keeps Kuramoto empirical measure near mean-field for e^n","Network norm condition yields exponential mean-field lock-in for Kuramoto oscillators","Graph norm condition ensures mean-field tracking for exponential times"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proofs depend on previously established spectral estimates for the linearized Kuramoto dynamics—an assumed decay rate of the semigroup and quadratic growth of eigenvalues—that the paper imports from earlier work and does not reprove; if these failed, the long-time control of the Brownian noise would break down.","fun_headline_variants_meta":{"raw":{"variants":["Graph norm condition pins Kuramoto empirical measure to mean-field for e^n time","One graph norm makes Kuramoto networks follow mean-field exponentially long","Graph norm condition keeps Kuramoto empirical measure near mean-field for e^n","Network norm condition yields exponential mean-field lock-in for Kuramoto oscillators","Graph norm condition ensures mean-field tracking for exponential times"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001159,"raw_usage":{"total_tokens":4846,"prompt_tokens":1034,"completion_tokens":3812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":3717}},"tokens_in":650,"tokens_out":3812,"duration_ms":25918,"temperature":1.0,"reasoning_tokens":3717,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:10:33.147216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A numerical falsifier: simulate the finite system on binomial random graphs with average degree $np_n\\to\\infty$, starting half the vertices at phase $0$ and half at phase $\\pi$ so that the initial condition depends on the graph; if for some fixed $T$ the supremum distance $\\sup_{t\\le T}\\|\\mu^n_t-\\mu_t\\|_{-1}$ does not tend to $0$ in probability, then condition (1.16) is not sufficient for Theorem 2.1.","supporting_citations":[{"cited_title":"Bertini, G","cited_arxiv_id":null,"evidence_quote":"Establishes the mean-field McKean–Vlasov equation and the spectral properties of the linearized operator for the complete-graph model."},{"cited_title":"Bertini, G","cited_arxiv_id":null,"evidence_quote":"Supplies the semigroup and eigenfunction estimates for $L_\\psi$ and $L_{2\\pi}$ that the long-time noise bounds rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the self-normalized martingale exponential inequality used to control the Brownian noise term around the synchronized manifold."},{"cited_title":"Giacomin, K","cited_arxiv_id":null,"evidence_quote":"Classifies the stationary solutions and phase transition of the McKean–Vlasov equation, giving the uniform state for $K<1$ and the manifold of synchronized states for $K>1$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the maximal inequality for Ornstein–Uhlenbeck processes that the subcritical noise control extends to infinite dimensions."},{"cited_title":"Pisier, Grothendieck’s Theorem, Past and Present, Bulletin of the American Mathematical Society 49 (2012), n","cited_arxiv_id":null,"evidence_quote":"States Grothendieck's inequality, the tool that converts the $\\ell^\\infty\\to\\ell^1$ norm of the normalized adjacency matrix into control of the graph fluctuation term."}],"review_version":1}