{"id":"9cda141c-3b6c-4364-8255-794271579d53","arxiv_id":"2412.07710","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper constructs a Lyapunov function (sparse free energy) for the Markovian local-field equation on regular trees, proves convergence to stationary distributions, and characterizes those distributions as tree Gibbs marginals.","lead":"This paper proves an H-theorem for a family of conditional McKean-Vlasov equations that describe interacting diffusions on regular trees: a new 'sparse free energy' decreases along the dynamics until the system reaches a stationary state. It identifies stationary states with Gibbs measures on trees and, in the tree-indexed line case, shows exponential convergence via a logarithmic Sobolev inequality.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The global Lyapunov and convergence claims rest on the unpublished inequality (6.30) in Proposition 4.2; until [20] is available or a self-contained proof is supplied, the paper's strongest advertised consequences are not fully verifiable.","rationale":"The reader's weakest_assumption already flags both the linear-growth existence issue and the reliance of Proposition 4.2 on unpublished [20]. I focus on the latter because it is the only unverified external input supporting the global Lyapunov and compact-level-set claims, which are the paper's advertised consequences of the H-theorem. The H-theorem identity is a conditional result, and the paper is transparent about Open Problem 1; however, Proposition 4.2 is presented as a theorem whose proof leans on an unpublished reference. Inequality (6.30) is not immediate: the coefficient κ/2 is not the standard 1/2 of a conditional-entropy decomposition, and its nonnegativity for all ν∈Mκ,d is exactly the kind of tree-entropy estimate that needs a published proof. I did not find an internal inconsistency in the proof of Theorem 4.1 itself; the algebraic cancellations around Eqs. (6.12)-(6.28) are consistent with definition (4.3). I also noted that Lemma 5.4's displayed SDE for the edge marginal appears to have a sign typo (the mimicking SDE for (X0,X1) should have drift −γ, not +γ), but the subsequent use in Lemma 6.1 employs the correct sign, so this is not load-bearing. Thus the appropriate verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":63270,"tokens_out":28217,"duration_ms":244188,"concrete_test":"Independently derive inequality (6.30) from the chain rule, the symmetry constraints (3.1)-(3.2), and published results, without citing [20]; if the derivation requires an additional condition, state it and revise Proposition 4.2 accordingly. As a complementary check, evaluate (6.30) numerically for κ=3, d=1 with α_q=N(0,1) and trial measures ν given by Gaussian-mixture 1-MRFs of the form ν0(x0)∏_{v=1}^3 \\barν(x_v|x0) that satisfy edge symmetry; a negative value would disprove (6.30) and falsify the lower-bound/compactness argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 4.2 is load-bearing for the paper's global Lyapunov claim. Its proof (Section 6.3) invokes, at Eq. (6.30), the assertion that ν ↦ H(ν||α_q^{⊗(1+κ)}) − (κ/2)H(\\barν||α_q^{⊗2}) is a nonnegative rate function, citing the unpublished manuscript [20]. This inequality is not a routine consequence of the chain rule: it couples the full density ν to the two-point marginal \\barν, with a κ/2 factor that is delicate under the symmetry constraints (3.1)-(3.2). It is the sole source of the lower bound Hκ ≥ −log R_q in (4.9), of the containment (4.10), and of the compact level sets used in the LaSalle-type argument of Theorem 4.6 and in Remark 4.7. If (6.30) failed for some ν∈Mκ,d, the sublevel sets {Hκ≤M} need not be tight or have finite entropy, and the convergence to Sκ would lose its compactness argument. I therefore view the unresolved status of [20] as the most load-bearing concern: the H-theorem identity itself is internally plausible, but the advertised 'global Lyapunov function' and convergence-to-stationarity results are not self-contained. A secondary gap, already acknowledged by the authors, is that the H-theorem applies only to linear-growth solutions satisfying (3.10); Theorem 3.10 establishes such solutions only when ∥∇W∥L∞<∞, leaving quadratic potentials to Open Problem 1 and the companion paper [38].","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the κ-regular Markovian local-field equation (κ-MLFE), a conditional McKean-Vlasov equation on the root neighborhood of a κ-regular tree. It proves well-posedness for bounded interaction gradients via a Schauder fixed-point argument, establishes an H-theorem for linear-growth solutions with a new functional Hκ (the sparse free energy) and a modified Fisher information Iκ, and identifies the zeros of Iκ with stationary distributions, Cayley fixed points, and marginals of continuous Gibbs measures. For κ = 2 it proves a renormalized-entropy representation of H2, uniqueness of stationary distributions via the Lacker-Zhang fixed-point problem, and exponential convergence under a uniform logarithmic Sobolev inequality. The proofs are detailed and largely self-contained in the stochastic-analysis sense, with the central computation being the symmetrization chain in Section 6.2.","tokens_in":63608,"tokens_out":5626,"duration_ms":54107,"significance":"If the results hold, this is a substantial advance: it provides a Lyapunov structure for a class of conditional McKean-Vlasov equations on sparse graphs, connects the dynamics to the Gibbs-measure fixed points of Lacker and Zhang [47], and gives quantitative rates in the κ = 2 case. The H-theorem is genuinely derived from the PDE rather than assumed, and the well-posedness proof is substantial. The main caveat is that Proposition 4.2, which supplies the lower bound and compact level sets used for the global Lyapunov and convergence claims, relies on the unpublished manuscript [20] at Eq. (6.30); until that ingredient is available or proved, the strongest advertised consequences are not fully verifiable.","major_comments":[{"comment":"Equation (6.30) cites the unpublished manuscript [20] for the nonnegativity of ν ↦ H(ν|α_q^{⊗(1+κ)}) − (κ/2)H(ν̄|α_q^{⊗2}) and for the associated finite-entropy and level-set conclusions. This inequality is not a routine consequence of the chain rule: it couples the full density ν to the two-point marginal ν̄ with a delicate κ/2 factor under the symmetry constraints (3.1)-(3.2). It is the sole source of the lower bound (4.9), the containment (4.10), and the compact level sets used in Theorem 4.6 and Remark 4.7. Because [20] is listed as in preparation, the global Lyapunov and convergence-to-stationarity claims are not fully verifiable as submitted. Please supply a self-contained proof of (6.30) and of the level-set claims, or restate the affected theorems as conditional on a supplied inequality.","section":"Section 6.3, Proposition 4.2"}],"minor_comments":[{"comment":"The well-posedness theorem is proved only when ∥∇W∥L∞ < ∞, while later results are conditional on existence of a linear-growth solution satisfying (3.10). This is acknowledged in Open Problem 1, but the abstract and Section 1.1 could make clearer that quadratic potentials are not covered by the theorems of this paper.","section":"Section 1.2.1 and Theorem 3.10"},{"comment":"The phrase 'Assumption 7.6(2)' appears to refer to Assumption D(2); please correct the reference.","section":"Section 8.2, Remark 8.4"},{"comment":"The symbol 'H⋆' in Eq. (4.29) should be 'H∗_2' to match the definition in (4.24).","section":"Theorem 4.17, Eq. (4.29)"},{"comment":"The text refers to a 'κ-CMVE' in the proof of part (3); this should be 'κ-MLFE'.","section":"Section 8.2, proof of Theorem 4.17(3)"},{"comment":"The word 'RELA TED' appears with a spurious space in the title; this is a typographical issue.","section":"Title page"},{"comment":"'U grows quadratically at infinite' should be 'at infinity'.","section":"Remark 4.5(2)"},{"comment":"The Rd extension of Theorem 1.9 of [47] is asserted 'by inspection of the proof'; since this is used for uniqueness of the stationary distribution, a sentence outlining the changes needed for Rd would be helpful.","section":"Theorem 4.16"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the reliance on the unpublished manuscript [20] at the exact compactness step of Proposition 4.2. I do not see circularity or citation padding: the self-citations [43]-[45] are to published, on-topic results, and the Lacker-Zhang paper [47] is used as an external benchmark. If the authors can replace [20] with a self-contained proof, or if [20] becomes available and its claimed inequality is correct, I would support acceptance. As submitted, the global Lyapunov and convergence theorems are not fully verifiable, so major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kevin Hu and Kavita Ramanan have a real result here: a new Lyapunov functional for a conditional McKean-Vlasov equation, with a proof of the energy-dissipation identity that is long but careful. The sparse free energy (1.6) and modified Fisher information (4.3) are genuinely new, and the connection to splitting Gibbs measures on trees is a nice payoff. The H-theorem computation in Section 6.2 is a sustained symmetrization argument using Lemma 6.2, and it checks out as far as I can see. The well-posedness theorem under bounded ∇W is a standard Schauder fixed-point argument, but solid.\n\nThe soft spots are real but local. Proposition 4.2, which gives the lower bound and compact level sets for Hκ, invokes the unpublished manuscript [20] for the rate-function inequality (6.30). That inequality is load-bearing: without it, the containment (4.10), the compactness argument in Theorem 4.6, and hence the 'global Lyapunov' convergence claim are not self-contained. The authors say [20] is in preparation; until it is available or a proof is included, the strongest advertised consequences rest on an external promise. The same goes for the κ=2 exponential convergence, which repeatedly defers to companion paper [38] 'in preparation.' These are not flaws in the H-theorem itself—Theorem 4.1 is proven directly from the PDE—but they are gaps in the paper's own narrative.\n\nAlso, well-posedness is only established when ∇W is bounded. The authors are upfront about this (Open Problem 1), and the H-theorem is stated conditional on linear growth solutions, so it is not a hidden assumption. The numerical figure is not reproducible from the text, which is minor but annoying.\n\nWho is this for? Anyone working on local-field equations, conditional McKean-Vlasov equations, or Gibbs measures on trees. The identification of stationary distributions with marginals of splitting Gibbs measures is a genuine bridge, and the renormalized entropy representation for κ=2 is a useful tool. I would send it to referees. The main request should be: either supply the proof of (6.30) or clearly mark Theorem 4.6 and Remark 4.7 as conditional on [20]. The H-theorem itself deserves to be in the literature.","headline":"A genuinely new Lyapunov functional for a conditional McKean–Vlasov equation, with a careful H-theorem proof; the global convergence claims, however, lean on an unpublished inequality and are not yet self-contained.","tokens_in":64164,"tokens_out":2057,"would_cite":true,"duration_ms":20088,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60J60","60J70","82C22","35Q84","82C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes an H-theorem for the κ-MLFE: the sparse free energy Hκ decreases along every linear-growth solution, with dissipation given by a modified Fisher information Iκ that vanishes exactly on stationary distributions, which…","keywords":["conditional McKean-Vlasov equation","kappa-regular tree","local-field equation","sparse free energy","H-theorem","modified Fisher information","splitting Gibbs measures","logarithmic Sobolev inequality"],"falsifier":"Take the paper's quadratic example U(x) = $7x^{2}$/4, W(x) = −$3x^{2}$/8 for κ = 2, d = 1, choose a smooth admissible initial condition that is not a 1-MRF, and solve the 2-MLFE numerically; if H2(μt) + ∫_0^t I2(μs) ds is not constant to within numerical error, or if H2 increases on any interval, the H-theorem is false. Alternatively, exhibit a measure with Iκ = 0 that is not a Cayley fixed point, which would break Theorem 4.10.","tokens_in":63038,"feed_emoji":"🌳","tokens_out":11770,"duration_ms":98366,"temperature":0.7,"pith_summary":"The paper studies the long-time behavior of a conditional McKean-Vlasov equation, the κ-MLFE, which arises as the Markovian version of the local-field equation describing the root neighborhood of interacting diffusions on a κ-regular tree. It identifies a functional it calls the sparse free energy Hκ and proves an H-theorem: along every solution whose conditional drift grows at most linearly, Hκ decreases with time and its rate of decrease is a nonnegative modified Fisher information Iκ. The paper then shows that the zeros of Iκ coincide with the stationary distributions of the flow and with the marginals of splitting Gibbs measures on the tree. Why this matters is that the stationary states are not known in advance and may be multiple, yet Hκ is a global Lyapunov function for the flow, so one can study relaxation to equilibrium without knowing the equilibrium first.","feed_headline":"Free energy falls along tree diffusion flow","feed_subtitle":"Its rate of decrease is a modified Fisher information that vanishes exactly at equilibrium states.","key_machinery":"The load-bearing object is the sparse free energy Hκ together with its dissipation Iκ, a modified Fisher information that subtracts the edge-marginal contribution from the full relative Fisher information. The argument rides on the energy dissipation identity (Theorem 4.1), proved by splitting the evolution of Hκ into a full entropy term and an edge-marginal correction, then using the linear Fokker-Planck equation for the time marginals and the leaf-exchangeability and edge-symmetry of the admissible measures. The zero set of Iκ is then described by a fixed-point recursion on the tree, the Cayley fixed-point equation, which is what ties the stationary states to splitting Gibbs measures.","core_discovery":"The central discovery is that the correct energy functional for the κ-MLFE is the sparse free energy, defined by an integral of the log-density minus a correction for the edge marginal plus the local potential terms U and W. Theorem 4.1 shows that along every linear-growth solution μt one has the energy dissipation identity Hκ(μt) − Hκ(μr) = −∫_r^t Iκ(μs) ds, where Iκ is the nonnegative modified Fisher information of Eq. (4.3). The zeros of Iκ are characterized as Cayley fixed points, and these are exactly the stationary distributions of the κ-MLFE and the root-neighborhood marginals of automorphism-invariant splitting Gibbs measures on the κ-regular tree. Consequently Hκ − H*κ acts as a global Lyapunov function even when there are multiple stationary states.","pith_inferences":["The paper does not claim the local-field equation shares the Lyapunov function, but were that transfer to hold, the same sparse free energy would control the original sparse-graph dynamics.","The paper does not compute rates for κ ≥ 3; a natural way to try is to push the renormalized-entropy representation through a uniform log-Sobolev inequality on the truncated tree, which would be a testable extension.","The paper's well-posedness theorem requires bounded interaction gradients ∇W; for unbounded potentials, an explicit check of the linear-growth condition (3.10) would either extend the H-theorem or identify the precise place where it fails."],"forward_implications":["Stationary distributions of the κ-MLFE can be computed by solving the Cayley fixed-point equation (or equivalently the Gibbs-marginal fixed-point equation) instead of by integrating the flow.","Every linear-growth solution converges to the zero set of Iκ as t → ∞, so the sparse free energy gives quantitative control over relaxation to equilibrium even in the presence of multiple stationary states.","For κ = 2, the sparse free energy is the renormalized limit of relative entropies of lifted 2-MRF measures against finite-tree Gibbs measures, and this representation plus a uniform log-Sobolev inequality yields exponential convergence to the unique stationary distribution.","For κ ≥ 3, the same H-theorem and stationarity characterization hold, but rates of convergence are left open because the Gibbs measures on infinite trees may be non-unique."],"supporting_citations":[{"why":"Establishes the local-field equation as the limit of interacting diffusions on sparse graphs, providing the motivation and the long-time-behavior link for the κ-MLFE.","marker":"[45]"},{"why":"Introduces the fixed-point equation for stationary local equations on regular trees; the paper shows its solutions are exactly the Cayley fixed points, and uses its uniqueness conditions for κ = 2.","marker":"[47]"},{"why":"Supplies the Schauder fixed-point and Hölder-regularity technique for well-posedness of conditional McKean-Vlasov equations, adapted here to the κ-MLFE.","marker":"[22]"},{"why":"The companion paper proving well-posedness for quadratic potentials in κ = 2, d = 1, which the authors cite to show the H-theorem extends beyond bounded ∇W.","marker":"[38]"},{"why":"Shows that trajectories of interacting diffusions on tree-like graphs are second-order Markov random fields, justifying the 2-MRF lift map used for entropy renormalization.","marker":"[43]"},{"why":"Provides the mean-field renormalized-entropy and uniform log-Sobolev method that the κ = 2 results adapt.","marker":"[35]"},{"why":"Unpublished rate-function result on which the compactness of Hκ level sets and lower semicontinuity in Proposition 4.2 depend.","marker":"[20]"},{"why":"Mimicking theorem used to represent the edge marginals of the κ-MLFE as a coupled pair of SDEs in Lemma 5.4.","marker":"[12]"},{"why":"Standard Fokker-Planck well-posedness and Hölder regularity results underlying Proposition 5.3 and Appendix A.","marker":"[7]"}],"fun_headline_variants":["Tree diffusions relax via sparse free energy drop","H-theorem on regular trees: energy decay proven","Modified Fisher info sets tree diffusion equilibria","Sparse free energy falls to zero on tree flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that a 'linear growth solution' exists, meaning the conditional drift γ satisfies the uniform bound (3.10), and for unbounded interaction gradients such existence is an open problem rather than a proved fact.","fun_headline_variants_meta":{"raw":{"variants":["Tree diffusions relax via sparse free energy drop","H-theorem on regular trees: energy decay proven","Modified Fisher info sets tree diffusion equilibria","Sparse free energy falls to zero on tree flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1196,"prompt_tokens":964,"completion_tokens":232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":172}},"tokens_in":580,"tokens_out":232,"duration_ms":3320,"temperature":1.0,"reasoning_tokens":172,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:34:37.148045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the paper's quadratic example U(x) = $7x^{2}$/4, W(x) = −$3x^{2}$/8 for κ = 2, d = 1, choose a smooth admissible initial condition that is not a 1-MRF, and solve the 2-MLFE numerically; if H2(μt) + ∫_0^t I2(μs) ds is not constant to within numerical error, or if H2 increases on any interval, the H-theorem is false. Alternatively, exhibit a measure with Iκ = 0 that is not a Cayley fixed point, which would break Theorem 4.10.","supporting_citations":[{"cited_title":"Theory Related Fields 187 (2023), no","cited_arxiv_id":null,"evidence_quote":"Establishes the local-field equation as the limit of interacting diffusions on sparse graphs, providing the motivation and the long-time-behavior link for the κ-MLFE."},{"cited_title":"Lacker and J","cited_arxiv_id":null,"evidence_quote":"Introduces the fixed-point equation for stationary local equations on regular trees; the paper shows its solutions are exactly the Cayley fixed points, and uses its uniqueness conditions for κ = 2."},{"cited_title":"Hu and K","cited_arxiv_id":null,"evidence_quote":"The companion paper proving well-posedness for quadratic potentials in κ = 2, d = 1, which the authors cite to show the H-theorem extends beyond bounded ∇W."},{"cited_title":"Lacker, K","cited_arxiv_id":null,"evidence_quote":"Shows that trajectories of interacting diffusions on tree-like graphs are second-order Markov random fields, justifying the 2-MRF lift map used for entropy renormalization."},{"cited_title":"Guillin, W","cited_arxiv_id":null,"evidence_quote":"Provides the mean-field renormalized-entropy and uniform log-Sobolev method that the κ = 2 results adapt."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Unpublished rate-function result on which the compactness of Hκ level sets and lower semicontinuity in Proposition 4.2 depend."},{"cited_title":"Brunick and S","cited_arxiv_id":null,"evidence_quote":"Mimicking theorem used to represent the edge marginals of the κ-MLFE as a coupled pair of SDEs in Lemma 5.4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Standard Fokker-Planck well-posedness and Hölder regularity results underlying Proposition 5.3 and Appendix A."}],"review_version":1}