{"id":"b7cdcf98-69c1-444e-bea3-6235aac92d47","arxiv_id":"1908.01126","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Spherical spin-glass Langevin dynamics started in bands around critical points converges to explicit integro-differential equations with a new overlap coordinate, reducing to Cugliandolo-Kurchan for unbiased starts.","lead":"This paper derives the exact large-N dynamical equations for spherical spin glasses when the system starts near a critical point of the energy landscape, an initial condition chosen with knowledge of the disorder. It extends the Cugliandolo-Kurchan theory to disorder-dependent starting points and connects long-time behavior to the geometry of the Gibbs measure.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only load-bearing caveat is the disclosed, unproved concentration hypothesis (1.28) behind Theorem 1.5; Theorem 1.1 itself does not depend on it, so no verdict change.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue: the concentration property (1.28) is the bridge from single-critical-point dynamics to the empirical average over critical points, and it is open for general mixed potentials. I agree with that assessment. My own read of the proof of Theorem 1.1 found no hidden circularity or missing step that would threaten the central derivation; the hard-sphere limit in Proposition 1.6 is supported by the f_L perturbation argument, and Section 5's transfer is valid conditional on (1.28). The authors flag (1.28) explicitly and label Theorem 1.5's unconditional form accordingly. Therefore the concern is a scope limitation rather than an internal inconsistency. It does not change the ACCEPT verdict, but it should remain part of the published scope statement.","tokens_in":51770,"tokens_out":16880,"duration_ms":167719,"concrete_test":"Test (1.28) for a genuinely mixed potential such as nu(r)=b_2^2 r^2 + b_4^2 r^4 with b_2,b_4 nonzero: compute E[#C_{N,q*}(I_N,I'_N)^2]/E[#C_{N,q*}(I_N,I'_N)]^2 by the second-moment/Kac-Rice method of Subag [34] for intervals I_N,I'_N shrinking to a deep (E*,G*) with q*^2 in the 1RSB support. A ratio tending to 1 would supply the lower-tail concentration via the generalized Paley-Zygmund bound and remove the main scope caveat; a ratio bounded away from 1 would confirm that Theorem 1.5 is conditional for general mixed potentials and that the abstract-level claim must be read with this restriction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the advertised disorder-dependent initial conditions at the level of random critical points, Theorem 1.5 needs #C_{N,q*}(I_N,I'_N) to exceed a positive fraction of its mean with high probability, i.e. (1.28). The text explicitly states immediately after (1.28) that this is currently proved only for pure m-spin models [34] and small mixed perturbations [13], not for general mixed potentials of (1.2)-(1.3). The Kac-Rice transfer in Section 5 is otherwise sound: Lemmas 5.3-5.4 and the Lipschitz argument establish (1.27), and (1.29) follows from (1.28) by Markov's inequality. Thus the empirical-average-over-critical-points conclusion is conditional on an open hypothesis for general mixed models. By contrast, Theorem 1.1 concerns a single disorder-conditioned critical point and is proved without (1.28); the conditional Gaussian identities (Lemma 3.7), concentration estimates (Propositions 3.1, 3.10), martingale decomposition (Lemma 4.2), uniqueness (Proposition 3.4), and equivalence (Proposition 3.6) appear internally consistent. The FDT discussion in Section 2 contains heuristic parts, but those are peripheral to the main derivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper derives the thermodynamic limit of the empirical correlation, integrated response, overlap, and energy functions for the Langevin dynamics (1.1) on the sphere, starting from a disorder-dependent initial condition localized near a critical point of the random Hamiltonian. Theorem 1.1 treats a single critical point σ satisfying the conditioning event CP(E⋆,G⋆,σ), with an initial condition uniform on the sub-sphere of overlap q0 with σ; it proves uniform a.s. and Lp convergence of (CN,χN,qN,HN) on [0,T]^2 to the unique bounded solution of the coupled integro-differential system (1.17)–(1.21). The proof follows the framework of Ben Arous–Dembo–Guionnet: it establishes pre-compactness and self-averaging (Propositions 3.1 and 3.3), uses Gaussian conditioning identities (Lemma 3.7) to compute the modified drift v⋆, reduces the limit to integral equations (Propositions 3.5 and 4.1), identifies them with (1.17)–(1.21) (Proposition 3.6), and proves uniqueness (Proposition 3.4). Theorem 1.5 extends the dynamics statement to an empirical average over critical points in a Kac–Rice window, conditional on the concentration hypothesis (1.28) for the number of critical points; the paper correctly states that this hypothesis is currently proved only for pure m-spin models and small mixed perturbations. Section 6 proves the hard-sphere limit fL→1 (Proposition 1.6), and Section 7 analyzes the FDT regime (Proposition 2.1), including an exact solution for the spherical SK model.","tokens_in":52011,"tokens_out":6281,"duration_ms":64583,"significance":"The conditioned single-point result is the core contribution. It gives a rigorous, closed system of non-random equations for a dynamical setting in which the initial condition and the disorder are strongly correlated, which is a qualitative step beyond [10]. The derivation contains no fitted parameters: E⋆, G⋆, q⋆, and q0 are model inputs, and v⋆ is computed explicitly from the conditional Gaussian law. The paper also provides a very useful Kac–Rice bridge linking the conditioned dynamics to an average over critical points, together with the hard-constraint limit and a detailed FDT analysis. I consider the explicit disclosure of (1.28) a strength rather than a defect: the conditional status of Theorem 1.5 is clearly marked, and Theorem 1.1 is not affected. Overall, if the proofs are correct, this is a substantial and well-structured contribution suitable for publication.","major_comments":[],"minor_comments":[{"comment":"The abstract's opening sentence, 'starting uniformly within one of the spherical bands', could be read as an unconditional statement for all mixed potentials; since the empirical-average statement of Theorem 1.5 relies on the currently unproved concentration hypothesis (1.28), I suggest adding a qualifying clause or moving the caveat more prominently into the abstract.","section":"Abstract and Section 1"},{"comment":"The verification of hypotheses (H1)–(H2) for qo≠0 and β∈[0,β1) is delegated with 'leaving the details to the reader'; a few lines showing the exponential envelope on the set A would make the FDT proposition more self-contained.","section":"Section 7, proof of Proposition 2.1"},{"comment":"The derivation of the stationary solution Γ(τ) is concise; citing the relevant place in [8] at the point of (2.26) would improve readability.","section":"Section 2.1, equations (2.24)–(2.31)"},{"comment":"The text contains several typographical and formatting artifacts (e.g., 'W e derive' in the abstract, 'Cugliandolo-Ku rchan' in the bibliography, 'goe random matrix' near page 2); a copyedit pass is needed.","section":"Throughout"}],"recommendation":"accept","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is the first rigorous derivation of spin-glass Langevin dynamics from disorder-dependent initial conditions, and the main theorem is proved carefully and honestly. The one caveat people will raise is disclosed by the authors themselves.\n\nWhat is actually new is the system (1.17)–(1.21). Starting from J conditioned on CP(E*,G*,sigma) and x0 uniform on the sub-sphere S_sigma(qo), the empirical correlation, response, overlap, and energy converge to a closed deterministic system. The new pieces are the overlap coordinate q(s) and the drift terms v* coming from the Gaussian conditioning. The square-bracket modifications inside the integrals are the footprint of the conditioning, and the whole thing collapses to the classical CK equations when q0 = 0, which is the right external consistency check.\n\nTheorem 1.1 is proved by adapting the Ben Arous–Dembo–Guionnet machinery, and I read the main lines carefully. The conditional Gaussian computation (Lemma 3.7) is the workhorse, and it is clean. Pre-compactness and self-averaging are in place, the martingale decomposition is the right tool, uniqueness comes from a Gronwall argument, and the reduction from the integral equations to the PDE system is spelled out. Quite a lot of estimates are imported from [10], but [10] is published, peer-reviewed, and independent, so this is not a circularity problem.\n\nThe soft spot is the concentration hypothesis (1.28) behind Theorem 1.5. Passing from a single conditioned critical point to an empirical average over critical points requires the critical-point count to stay within a positive fraction of its mean, and the authors state in the text that this is proved only for pure m-spin models and small mixed perturbations. So the strongest advertised version—dynamics starting uniformly at a random critical point for general mixed potentials—is conditional on an open hypothesis. That is a genuine limitation, but it is flagged explicitly and Theorem 1.1 does not depend on it.\n\nThe Section 2 large-time/FDT analysis is the weakest part. Proposition 2.1 relies on hypotheses verified for small beta, and the aging discussion is clearly heuristic. None of this affects the core derivation.\n\nBottom line: a serious, rigorous paper that makes a real advance on the CK/BDG program. Send it to a strong referee. I would take it to our reading group.","headline":"First rigorous treatment of spin-glass Langevin dynamics from disorder-dependent initial conditions; Theorem 1.1 is proved carefully and holds up, and the one real caveat (concentration for critical-point counts) is disclosed by the authors themselves.","tokens_in":52606,"tokens_out":3608,"would_cite":true,"duration_ms":34232,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82C44","82C31","60H10","60F15","60K35"],"pacs":[],"model":"deepseek-v4-flash","headline":"For spherical spin glasses, starting uniformly around a conditioned critical point yields deterministic limit equations.","keywords":["Langevin dynamics","spherical spin glasses","mixed p-spin models","critical points","Kac-Rice formula","integro-differential equations","aging and FDT regime","thermodynamic limit"],"falsifier":"A concrete test of Theorem 1.1 would be to simulate the conditioned Langevin dynamics for a fixed mixed potential (say $m=3$ or $m=4$) at finite large $N$, starting uniformly on $S_\\sigma(q_o)$ around a critical point with $G_*>2\\sqrt{\\nu''(q_*^2)}$, and compare the empirical $(C_N,\\chi_N,q_N,H_N)$ on $[0,T]^2$ with the numerical solution of (1.17)-(1.21); an $N$-independent discrepancy would refute the convergence claim. For Theorem 1.5, a direct falsifier is a mixed model in which the count concentration (1.28) fails, for instance if the random number of critical points in a shrinking band around $(E_*,G_*)$ shows non-negligible fluctuations at every $N$, because the unconditional convergence statement (1.29) would then have no basis.","tokens_in":51410,"feed_emoji":"🧲","tokens_out":11331,"duration_ms":109829,"temperature":0.7,"pith_summary":"Spherical spin glasses are systems of N continuous spins on a sphere whose random interactions create a rugged energy landscape. This paper studies their Langevin dynamics when the initial condition is not drawn independently of the disorder, but uniformly on the sub-sphere of fixed overlap with a critical point of the Hamiltonian, conditional on that critical point having prescribed energy and gradient data. The central claim is that in the thermodynamic limit the empirical correlation, response, overlap, and energy functions become non-random and are the unique solution of a closed system of integro-differential equations (1.17)-(1.21). This matters because it turns the random, high-dimensional dynamics into a deterministic, low-dimensional transport problem, and because the same equations predict distinct long-time regimes: standard fluctuation-dissipation behavior at high temperature and, at low temperature, either escape from or localization around the initial critical point. The extension from one conditioned critical point to an average over all critical points is made through the Kac-Rice formula and is unconditional only under a concentration hypothesis on critical-point counts that the paper states is currently proved for pure m-spin models and small mixed perturbations.","feed_headline":"Critical-point starts give deterministic spin-glass dynamics","feed_subtitle":"Correlation, response, overlap and energy converge almost surely to closed integro-differential equations.","key_machinery":"The load-bearing object is the conditional Gaussian field. Conditioned on $CP(E_*,G_*,\\sigma)$, the Hamiltonian splits as $H_J(x)=H_{J_o}(x)-N v(N^{-1}\\langle x,\\sigma\\rangle)$, where $J_o$ is a centered Gaussian field with the modified covariance kernel (1.24) and $v$ is the polynomial in (1.22) built from $\\nu$ and its derivatives at $q_*^2$. This deterministic drift is what produces the extra terms involving $v_*$ in the limiting integro-differential equations; with $q_o=0$ those terms disappear and the system reduces to the standard closed equations for correlation and response. The second mechanism is the Kac-Rice formula, which turns the empirical average over critical points into the conditional single-point dynamics of Theorem 1.1, with the concentration property (1.28) controlling the passage from expectations to empirical averages.","core_discovery":"The paper proves Theorem 1.1: for a spherical mixed $m$-spin Hamiltonian conditioned on $CP(E_*,G_*,\\sigma)$ and initial law $\\mu^{q_o}_\\sigma$ on the sub-sphere $S_\\sigma(q_o)$, the empirical functions $(C_N,\\chi_N,q^\\sigma_N,H_N)$ converge uniformly on $[0,T]^2$, almost surely and in $L^p$, to the unique bounded solution of the integro-differential system (1.17)-(1.21), with $q(0)=q_o$, $C(0,0)=1$, $R(s,t)=0$ for $t>s$, and $R(s,s)=1$. The proof shows that every limit point of the pre-compact empirical processes satisfies the integral equations (3.19)-(3.31), then proves these are equivalent to the claimed differential system, and uniqueness of bounded solutions completes the identification. The paper also derives the hard-sphere limit of the conditional equations (Proposition 1.6) and, for small $\\beta$, proves the large-time limit enters the FDT regime with $R_{\\rm fdt}(\\tau)=-2C'_{\\rm fdt}(\\tau)$, where $C_{\\rm fdt}$ solves the scalar equation (2.3); when $q_o=0$ the whole system reduces to the standard closed correlation-response equations. Theorem 1.5 uses the Kac-Rice formula to average over critical points in a shrinking energy window, yielding convergence in expectation unconditionally and convergence in probability under the concentration hypothesis (1.28).","pith_inferences":["If the concentration property (1.28) is proved for all mixed potentials, Theorem 1.5 would give a rigorous dynamical counterpart to the low-temperature band decomposition of the Gibbs measure: for times short compared with exponential-in-$N$ escapes, the path's observables are governed by the deterministic system around one band, and transitions between bands are rare.","The quantity $\\alpha=\\lim_{t\\to\\infty} q(t)/q_*$ in the FDT analysis offers a simulation-friendly way to distinguish escaped from localized metastable states without computing TAP free energies; the paper's equations predict localization only when $G_*$ exceeds the local-minimum threshold, which is directly testable in numerical Langevin runs.","Because $v_*$ depends explicitly on $(E_*,G_*)$, the paper suggests that conditioning on deeper energy levels with large normal derivative stabilizes the initial band; an implicit question not addressed is whether the convergence rate in Theorem 1.1 is uniform as $(E_*,G_*)$ vary over the allowed region."],"forward_implications":["For a single conditioned critical point, the empirical correlation, response, overlap, and energy functions become deterministic at $N\\to\\infty$ and obey the closed system (1.17)-(1.21).","If $q_o=0$, the overlap solution is identically zero and the dynamics reduce to the classical correlation-response equations; nonzero start overlap and nonzero $E_*,G_*$ add deterministic drift terms through $v_*$.","In the hard-sphere limit the conditional equations have a unique bounded solution, the reduced kernel $C(s,t)-q(s)q(t)/q_*^2$ is non-negative definite, and $q(s)$ stays in $[-q_*,q_*]$.","For small $\\beta$ the long-time limit is the FDT solution of the scalar equation (2.3), while the paper's consistency analysis predicts localized no-aging states only when the Hessian-shift condition $G_*>2\\sqrt{\\nu''(q_*^2)}$ holds and equations (2.16)-(2.19) admit a solution.","If the concentration hypothesis holds, the empirical average over all critical points in a shrinking energy window converges in probability to the same deterministic dynamics as a single typical conditioned critical point."],"supporting_citations":[{"why":"Supplies the base convergence method of auxiliary functions, self-averaging, and Lipschitz concentration, which the paper adapts to conditional fields and disorder-dependent initial laws.","marker":"[10]"},{"why":"Provides the hard-sphere-limit equations, uniqueness of bounded solutions, and the FDT analysis that Propositions 1.6 and 2.1 extend to the conditional setting.","marker":"[24]"},{"why":"Proves the second-moment complexity bounds used as the concentration property (1.28) for pure m-spin models, making Theorem 1.5 unconditional in that case.","marker":"[34]"},{"why":"Proves the concentration property for small mixed perturbations and supplies the low-temperature band geometry used to identify the initial bands.","marker":"[13]"},{"why":"Supplies the Kac-Rice formula underlying Proposition 5.1 and the bridge from conditional single-point dynamics to averages over critical points.","marker":"[1]"},{"why":"Computes expected critical-point counts for general mixed models, used in the normalization in Theorem 1.5.","marker":"[4]"},{"why":"Computes expected critical-point counts for pure m-spin models, used in the pure-case application of Theorem 1.5.","marker":"[5]"},{"why":"Gives the original physics derivation of the closed correlation-response equations that (1.17)-(1.21) generalize.","marker":"[23]"}],"fun_headline_variants":["Critical-point starts yield deterministic spin-glass flow","Spin-glass dynamics from band starts are deterministic","From critical points to closed equations for spin glasses","Disorder-dependent starts: deterministic limit for spin glasses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the unconditional Theorem 1.5, the load-bearing premise is the concentration property (1.28): with high probability the number of critical points in the selected energy and normal-derivative window is at least a fixed positive fraction of its expectation, and the paper explicitly says this is currently proved only for pure m-spin models and small mixed perturbations.","fun_headline_variants_meta":{"raw":{"variants":["Critical-point starts yield deterministic spin-glass flow","Spin-glass dynamics from band starts are deterministic","From critical points to closed equations for spin glasses","Disorder-dependent starts: deterministic limit for spin glasses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1252,"prompt_tokens":939,"completion_tokens":313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":555,"tokens_out":313,"duration_ms":3943,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:24:20.007548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test of Theorem 1.1 would be to simulate the conditioned Langevin dynamics for a fixed mixed potential (say $m=3$ or $m=4$) at finite large $N$, starting uniformly on $S_\\sigma(q_o)$ around a critical point with $G_*>2\\sqrt{\\nu''(q_*^2)}$, and compare the empirical $(C_N,\\chi_N,q_N,H_N)$ on $[0,T]^2$ with the numerical solution of (1.17)-(1.21); an $N$-independent discrepancy would refute the convergence claim. For Theorem 1.5, a direct falsifier is a mixed model in which the count concentration (1.28) fails, for instance if the random number of critical points in a shrinking band around $(E_*,G_*)$ shows non-negligible fluctuations at every $N$, because the unconditional convergence statement (1.29) would then have no basis.","supporting_citations":[{"cited_title":"DEMBO, A","cited_arxiv_id":null,"evidence_quote":"Supplies the base convergence method of auxiliary functions, self-averaging, and Lipschitz concentration, which the paper adapts to conditional fields and disorder-dependent initial laws."},{"cited_title":"GUIONNET, A","cited_arxiv_id":null,"evidence_quote":"Provides the hard-sphere-limit equations, uniqueness of bounded solutions, and the FDT analysis that Propositions 1.6 and 2.1 extend to the conditional setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves the second-moment complexity bounds used as the concentration property (1.28) for pure m-spin models, making Theorem 1.5 unconditional in that case."},{"cited_title":"SUBAG, E","cited_arxiv_id":null,"evidence_quote":"Proves the concentration property for small mixed perturbations and supplies the low-temperature band geometry used to identify the initial bands."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Kac-Rice formula underlying Proposition 5.1 and the bridge from conditional single-point dynamics to averages over critical points."},{"cited_title":"BEN AROUS, G","cited_arxiv_id":null,"evidence_quote":"Computes expected critical-point counts for general mixed models, used in the normalization in Theorem 1.5."},{"cited_title":"BEN AROUS, G","cited_arxiv_id":null,"evidence_quote":"Computes expected critical-point counts for pure m-spin models, used in the pure-case application of Theorem 1.5."},{"cited_title":"KURCHAN, J","cited_arxiv_id":null,"evidence_quote":"Gives the original physics derivation of the closed correlation-response equations that (1.17)-(1.21) generalize."}],"review_version":1}