{"id":"d78a4747-f266-42b4-acce-69f4b2926d3a","arxiv_id":"2607.20014","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Non-exchangeable, non-conservative particle systems converge, as N grows, to a Vlasov-type equation with a mass source term, via extended graphons and a generalized Glivenko-Cantelli lemma.","lead":"This paper proves a mean-field limit for a large class of interacting particle systems in which agents are not identical and their mass — influence or charisma — can grow or shrink over time. The result turns a previously open non-exchangeable, non-conservative case into a theorem, using the recent extended-graphon framework plus a new weighted Glivenko-Cantelli lemma.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's hypotheses (2) are strictly weaker than the kernel assumptions used in Lemmas 9–10; the proof's L2-hierarchy estimate fails for admissible W^{1,∞} kernels.","rationale":"The paper is a careful adaptation of the extended-graphon machine to non-conservative, non-exchangeable systems. The high-level chain — Theorem 3, hierarchy stability, Lemma 10, then Theorem 1 — is coherent under the stronger kernel assumptions that are actually used. I found no internal contradiction in the main construction, and the weighted Glivenko–Cantelli lemma is a genuine new ingredient. However, the reader's weakest-assumption (the mass floor) is an explicit data condition that the proof legitimately uses; it is restrictive but not a correctness gap. The more load-bearing issue is the mismatch between Theorem 1's hypothesis (2) and the kernel integrability demanded by Proposition 8, Lemma 9, and Lemma 10. This is not merely cosmetic: the L2-hierarchy estimate at the core of Lemma 10 fails for natural W^{1,∞} kernels such as arctan, so the proof of the stated theorem has a real gap. The reader did list this mismatch as one of the reasons for CONDITIONAL, though not as the weakest assumption; hence partial agreement. A strengthened theorem with K1∈W^{1,1}∩W^{1,∞}, divK1∈L∞, K2∈L^1∩W^{1,∞} would likely go through with the existing proof, which is why I keep the CONDITIONAL verdict rather than moving to REJECT or UNVERDICTED.","tokens_in":51915,"tokens_out":21529,"duration_ms":218642,"concrete_test":"Set d=1, K1(x)=arctan(x), K2(x)=1. Both satisfy (2) but violate the L1/L2 hypotheses of Lemma 10. Take h_{T+i}(x_1,...,x_T,y)=φ(x_1,...,x_T)ψ(y) with φ∈L^2(R^T) nonzero and ψ∈L^1_c(R) with ∫ψ=1. Compute ∫ K1(x_i-y) h_{T+i}(...,y)dy = φ(...) (arctan*ψ)(x_i). Since arctan*ψ tends to π/2 at +∞, its L^2(R) norm is infinite, so the convolution bound (46) in Lemma 8 fails. Verify whether an alternative bound using only ||K1||_{W^{1,∞}} closes the hierarchy estimate; if not, Theorem 1 requires stronger kernel hypotheses or a new proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central convergence proof of Theorem 1 invokes Lemma 10 to compare the empirical auxiliary solution f̄_N with the limiting weak solution f. Lemma 10 (and also Proposition 8 and Lemma 9) requires K1 ∈ L^∞∩W^{1,1}, div K1 ∈ L^∞, and K2 ∈ L^1∩L^∞. Theorem 1's hypothesis (2) only assumes K1, K2 ∈ W^{1,∞}(R^d). Since W^{1,∞} does not imply L^1 or W^{1,1} — e.g. K1(x)=arctan(x) and K2(x)=1 are both admissible under (2) — the proof cannot apply its own stability lemma to (w, f) and (w_N, f̄_N). The gap is real and located at a load-bearing step: the L2-hierarchy estimate in Lemma 8 uses the convolution bound ||K_j * h_{T+i}||_{L2} ≤ ||K_j||_{L2} ||h_{T+i}||_{L1} (cf. (46)). For K_j = arctan, this bound is unavailable; the convolution of arctan with a nonzero L1 function is not in L2. Thus the stated Theorem 1 is not established by the supplied argument unless either (2) is strengthened to include the integrability assumptions of Lemmas 9–10 or those lemmas are reproved under only W^{1,∞} regularity. This is a proof-completeness concern, not a claim that the mean-field limit is false for such kernels. The uniform-mass floor (6) is explicitly assumed and is used correctly; it is not the weakest point.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the mean-field limit of a non-exchangeable, non-conservative particle system (1), where each agent carries a weight M_i(t) that evolves in time. Under assumptions (2)-(8), it claims that the weighted empirical measure (1/N)∑ M_i(t)δ_{X_i(t)} converges, as N→∞, to a deterministic density f(t,x,ξ) solving the non-conservative limit equation (9), with the connectivity matrices converging to an extended graphon w. The proof has two main blocks: propagation of independence (Theorem 3, using the auxiliary independent system (22) and a new weighted Glivenko-Cantelli lemma (Lemma 5)), and then the extended-graphon hierarchy stability machinery adapted from [17] (Section 4), which yields Theorem 1. The paper also contains a self-contained treatment of the simpler bounded-graphon case (Section 2).","tokens_in":52122,"tokens_out":15072,"duration_ms":128980,"significance":"If the main theorem were established, the paper would be a notable extension of the non-exchangeable mean-field theory to systems with non-conserved mass, and Lemma 5 (weighted Glivenko-Cantelli) would be a useful technical contribution. The authors carefully track all constants and there are no fitted parameters. However, the proof of Theorem 1 as written has a load-bearing kernel-regularity gap: the hypotheses of the main theorem are strictly weaker than those required by the stability lemmas used to prove it. The paper can likely be repaired by strengthening assumption (2) or by providing a W^{1,∞}-only stability argument, but this must be done before the central claim is supported.","major_comments":[{"comment":"Theorem 1 assumes only (2): K1,K2 ∈ W^{1,∞}(R^d). However, the proof of Theorem 1 invokes Lemma 10 to compare f̅_N and f, and Lemma 10's hypotheses require K1 ∈ L∞∩W^{1,1}, div K1 ∈ L∞, K2 ∈ L^1∩L∞. Moreover, Lemma 8 (used inside Lemma 10) requires K1,K2 ∈ L^2(R^d), and its estimate (46) uses ∥K_j*h∥_{L2} ≤ ∥K_j∥_{L2}∥h∥_{L1}. Since W^{1,∞}(R^d) does not imply membership in any L^p for p<∞ — e.g. K1(x)=arctan(x), K2≡1 are admissible under (2) — the L^2-hierarchy estimate is unavailable for admissible kernels. Thus the supplied argument does not establish Theorem 1 for the stated class. Please either strengthen (2) to include the integrability assumptions used by Lemmas 8–10, or replace the L^2 hierarchy by a stability estimate that works for merely W^{1,∞} kernels.","section":"§4.2–§4.3, Lemma 10, Lemma 8, Proposition 8"},{"comment":"There is an internal inconsistency in the kernel assumptions for Proposition 8. The proposition statement requires only divK1∈L^1(R^d), but the proof (e.g. estimate (75)) and the later use in Lemma 9 require divK1∈L∞(R^d). Since Proposition 8 is used to construct the solutions f^ν entering Lemma 10, the assumptions must be stated uniformly and correctly. This is separate from but intertwined with the main Theorem 1 gap.","section":"§5.4, Proposition 8"}],"minor_comments":[{"comment":"The lower bound on N in the statement is 'N > max{1, ̄m^{-1/(2+3d/2)}}', but the proof (with k=2+3d/2 and ε=N^{-1/k}) requires N^{-1/k} < ̄m, i.e. N > ̄m^{-(2+3d/2)}. The stated threshold is too low. Since the proof only needs N large, the statement should be corrected.","section":"§5.3, Lemma 5"},{"comment":"In the estimate of I3, the text says 'E[G_{i,k}(t)|̄X_i(t)] = 0' but the symbol G should be H. This is a typo but slightly confusing in a delicate conditional-independence argument.","section":"§3, Lemma 4"},{"comment":"The derivation of f^0_i = g_i dx introduces the conditional expectation h_i(x) = ... / f_{X_i^0}(x). The division is only valid a.e. on {f_{X_i^0}>0}, and the final identity is correct without division. It would be cleaner to define h_i through the relation h_i f_{X_i^0} = g_i to avoid a zero-denominator issue.","section":"§4.3, proof of Theorem 1"},{"comment":"The proof is written with a long chain of estimates; some constants are combined into C without being redefined. For reproducibility, please label the constants at the end of each inequality, especially in (60)–(61).","section":"§5.1, Lemma 2 proof"}],"recommendation":"major_revision","confidential_remarks":"The kernel-regularity gap is the main reason for the major revision. It is a proof-completeness issue rather than a claim that the mean-field limit is false for W^{1,∞} kernels, so the theorem may be salvageable. I would encourage the authors to also check whether the same mismatch exists in the cited [17] framework, since the paper explicitly says it is 'matching' that framework. If [17] assumes stronger kernel integrability, the statement 'under minimal conditions' in the introduction should be qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does two things worth your time. First, it extends the Jabin–Poyato–Soler extended-graphon machinery to systems where total mass is not conserved and particles are not exchangeable. That combination is absent from the cited literature. Second, it proves a weighted Glivenko–Cantelli lemma (Lemma 5) for empirical measures with non-constant total mass, with an explicit N^{-1/(2+3d/2)} rate. The proof chain is mostly coherent: Theorem 3's propagation of independence, the hierarchy equation (43), and the stability iteration in Lemma 8 all check out. No fitted parameters; constants depend only on problem data. The debt to [17] is stated openly.\n\nThe soft spot is real and load-bearing. Theorem 1 is stated under (2), which only asks K1, K2 in W^{1,∞}. But Lemma 10 — the step that compares the auxiliary empirical density to the limit — requires K1 in W^{1,1}∩L^∞, div K1 in L^∞, and K2 in L^1∩L^∞. Lemma 8 needs K1, K2 in L^2 for the convolution bound (46). W^{1,∞} does not imply any of those: for example, K1(x)=arctan(x) is admissible under (2). So as written, the argument does not prove Theorem 1 for the full class of kernels it claims. This is not a claim that the mean-field limit is false for such kernels; it's a proof-completeness gap. The fix is either to strengthen (2) to match the lemmas or to redo the stability estimates under W^{1,∞} alone. That is nontrivial because the L2-hierarchy method appears to rely on kernel integrability.\n\nAlso worth noting: Propositions 1–3 are stated without proof ('standard arguments'), and [17, Theorem 5.1] is used as a black box. These are acceptable for a note but should be flagged. The mass floor (6) is restrictive but it is explicitly assumed and used correctly; the stress-test concern about that is not the weakest point.\n\nWho is this for? Someone working on mean-field limits of non-exchangeable particle systems, especially with adaptive or non-conservative weights. The weighted GC lemma alone is a useful contribution. It deserves a serious referee: the idea is right, the gap is fixable, and the paper is transparent about its debts. I would send it to review, but the referee needs to demand either corrected assumptions or a reproof of Lemma 10 under the stated hypotheses.","headline":"A real extension to non-exchangeable non-conservative mean-field limits with a new weighted Glivenko–Cantelli lemma, but the proof of Theorem 1 uses kernel regularity stronger than the stated assumptions.","tokens_in":52839,"tokens_out":3198,"would_cite":false,"duration_ms":31763,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q83","35Q70","35R02","35Q49","35R06"],"pacs":[],"model":"deepseek-v4-flash","headline":"The random, mass-weighted empirical measure of a non-exchangeable, non-conservative particle system converges to a deterministic density solving a Vlasov-type equation with a source term.","keywords":["mean-field limit","non-exchangeable systems","extended graphons","non-conservative systems","weighted empirical measure","Glivenko-Cantelli lemma","Vlasov hierarchy","interacting particle systems"],"falsifier":"Simulate the particle system with half of the agents initially at zero influence (M^0_i = 0) and positive interaction weights connecting them to others; if the expected flat-metric distance from the deterministic limit does not go to zero as N grows — or the limit density fails to match the empirical measure — the central claim is refuted. Alternatively, measure the rate of decay in N and check whether it matches the N^{-1/(2+3d/2)} bound.","tokens_in":51555,"feed_emoji":"⚛️","tokens_out":5703,"duration_ms":51927,"temperature":0.7,"pith_summary":"This paper aims to establish a mean-field limit for a system of N interacting agents that is both non-exchangeable — each pair of agents has its own interaction strength — and non-conservative — each agent carries an influence mass that changes over time and is not preserved. The main claim is that the random, mass-weighted empirical measure of the agent states converges, as N grows, to a deterministic density f that solves a Vlasov-type equation with a source term A - V2[f], so total mass can grow or decay in the limit. The proof extends the theory of extended graphons — measure-valued connectivity kernels that encode the asymptotic interaction structure — and introduces a weighted version of the Glivenko-Cantelli lemma to control the empirical measure when the weights are random and time-dependent. If correct, this gives a rigorous continuous description of heterogeneous interacting systems whose agents gain or lose influence, such as opinion dynamics with polarization or balance laws in biology.","feed_headline":"Swarm limit holds even when influence is not conserved","feed_subtitle":"Mass-preserving assumption dropped: random weighted empirical measures converge to a Vlasov-type density with a source term.","key_machinery":"The argument rests on three linked devices. The first is the extended graphon: a weak-* measurable, measure-valued kernel w(ξ,dζ) (a measure in the second variable) that serves as the limit object for the discrete connectivity matrices, allowing heterogeneous interactions without a classical graphon. The second is the family of tree-indexed observables τ(T, w, f), which are built by integrating f over the tree structure of the interaction graph; these solve a non-exchangeable Vlasov hierarchy that is used to propagate stability from the initial data to the limit equation. The third, and the new piece, is a weighted Glivenko-Cantelli lemma: it bounds, in expectation, the flat-metric distance","core_discovery":"On the paper's own terms, the central claim is Theorem 1: under assumptions (2)-(8), there exist an extended graphon w and a density f in L∞((0,t*)×(0,1); W^{1,1}∩W^{1,∞}(R^d)) such that f is a weak solution of the non-conservative limit equation ∂t f + div(f V1[f]) = f(A - V2[f]), with Vj[f](t,x,ξ) = ∫∫ Kj(x-y)f(t,y,ζ)dy w(ξ,dζ). The convergence statement is that, up to a subsequence, the expected flat metric between the deterministic density ∫ f(t,·,ξ)dξ and the random weighted empirical measure (1/N)Σ M_i(t)δ_{X_i(t)} tends to 0 uniformly in t ∈ [0,t*] as N → ∞. This is the first mean-field limit for non-conservative non-exchangeable systems: even though the total mass of the particle sys","pith_inferences":["A direct consequence the authors leave implicit: if a positive fraction of agents starts with zero expected influence, the mass floor assumption (6) fails, and the weighted empirical measure can lose mass in the limit; the result may then need a different normalization or may fail outright.","The non-conservative source term allows mass amplification when A dominates V2[f]; this models sharp opinion polarization or population blow-up. One testable extension is to determine whether the convergence rate N^{-1/(2+3d/2)} is optimal under the given assumptions.","Because the graphon is measure-valued in the second variable, the framework may also cover sparse or adaptive connectivity sequences for which classical graphons do not exist, suggesting a route toward mean-field limits for co-evolving networks."],"forward_implications":["If the central claim is correct, non-conservative non-exchangeable particle systems admit a deterministic mean-field description without any conservation law; the limit density satisfies a balance law with a source term f(A - V2[f]).","The weighted Glivenko-Cantelli lemma gives a concrete convergence rate, on the order of N^{-1/(2+3d/2)} in expectation, for the empirical measure with random and time-varying weights.","The limit does not require a priori knowledge of the connectivity limit: the extended graphon is extracted as a weak-* limit of the discrete matrices, so only structural assumptions on the weights are needed.","Since the exchangeable case (all weights equal to 1/N) is a special case, this recovers the classical mean-field limit with a conservative limit when A = 0 and K2 = 0.","The result implies propagation of independence: as N grows, the agents' states become asymptotically independent, which is what makes the tree-indexed hierarchy close."],"fun_headline_variants":["Mean-field limit proven for non-conservative systems","Swarm limit works even when mass is not conserved","Non-conservative mean-field: extended graphon approach","Random weighted measures converge without mass conservation","Mass loss no barrier: mean-field limit proven"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole proof hinges on the assumption that every agent's expected influence stays uniformly bounded below by a positive constant (inf_N min_i E M^0_i ≥ m > 0) and above by a finite bound; if some agents have vanishing initial influence, the weighted empirical measure can shed mass and the convergence argument collapses.","fun_headline_variants_meta":{"raw":{"variants":["Mean-field limit proven for non-conservative systems","Swarm limit works even when mass is not conserved","Non-conservative mean-field: extended graphon approach","Random weighted measures converge without mass conservation","Mass loss no barrier: mean-field limit proven"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1203,"prompt_tokens":625,"completion_tokens":578,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":369,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":369,"tokens_out":578,"duration_ms":5002,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:04:41.151907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the particle system with half of the agents initially at zero influence (M^0_i = 0) and positive interaction weights connecting them to others; if the expected flat-metric distance from the deterministic limit does not go to zero as N grows — or the limit density fails to match the empirical measure — the central claim is refuted. Alternatively, measure the rate of decay in N and check whether it matches the N^{-1/(2+3d/2)} bound.","supporting_citations":[],"review_version":1}