{"id":"5a2a83bf-15ad-41c0-8417-282908921fe1","arxiv_id":"2602.01641","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For lower-triangular interacting diffusions, incremental path-space relative entropy decays as 1/(i-1), yielding optimal N^{-1/2} mean-field convergence and a fluctuation SPDE with a factor-2 feedback.","lead":"A new proof shows that a sequential particle system—where each particle sees only earlier particles—matches the accuracy of the fully coupled mean-field approximation, with relative entropy decaying like 1/(i-1) per particle. It also derives the N^{-1/2} empirical convergence and a Gaussian fluctuation limit with a distinctive factor-two feedback term.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's factor-2 SPDE limit rests on the unproved assertion a_j→a for prefix fluctuation fields; without it, the identification of (1.10) is incomplete.","rationale":"The reader's conditional verdict identifies the unproved assertion a_j→a in the identification step of Theorem 3 as a genuine gap, and I agree. Theorems 1–2 appear carefully argued: the incremental entropy bound, the martingale-difference replacement, the upper-envelope closure, and the negative-Sobolev energy estimate are internally consistent and the main estimates close. The Gaussian fluctuation theorem, however, is not fully proved as written. The manuscript itself acknowledges in §5 that much of the functional-analytic framework is streamlined and imported from [31], and the convergence of the prefix fluctuation fields is asserted without proof. Since the factor-2 SPDE (1.10) is a headline contribution and depends directly on this step, the appropriate disposition is conditional: the paper should be accepted only after the convergence of a_j→a (or an equivalent prefix-field convergence) is supplied, together with a rigorous derivation of uniqueness/well-posedness for (1.10) in the weighted Sobolev space. My analysis does not change the reader's verdict, so I recommend UNCHANGED.","tokens_in":39609,"tokens_out":7656,"duration_ms":75256,"concrete_test":"Attempt to prove, under the assumptions of Theorem 3, that the prefix fluctuation fields η^j converge in law in C([0,T];H^{-β',-m}) to the same process η obtained from the global subsequence, without invoking Theorem 3's conclusion. Concretely, establish Aldous tightness for {η^j}_j and uniqueness of its limit points, or directly prove E|∫_s^t ⟨ρ_r,(K*(η^j_r−η_r))·∇φ⟩ dr|→0 using the quantitative bounds of Theorem 2 and the compact embedding H^{-β}⊂H^{-β',-m}. If this convergence cannot be proved without already assuming the target SPDE's well-posedness, then Theorem 3's identification of the factor-2 term is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Theorem 3 is the identification of the limiting SPDE (1.10), including the factor-2 feedback. The decisive step in §5.1 is the Stolz–Cesàro passage: after setting a_j = ∫_s^t ⟨ρ_r, (K*η^j_r)·∇φ⟩ dr and a = ∫_s^t ⟨ρ_r, (K*η_r)·∇φ⟩ dr, the text simply states 'We have a_j→a as j→∞.' This is not proved, and it does not follow from the preceding tightness argument. Tightness of the global sequence {η^N}_N in C([0,T];H^{-β',-m}) gives at most subsequential convergence of η^N to some limit η; the theorem's conclusion η^N⇒η is exactly what is being established. The prefix fields η^j = √j(µ^j−ρ) are not shown to converge to the same η, and no quantitative estimate controlling ∥η^j−η∥ is supplied. The linear term can be passed to the limit along the subsequence, but the Stolz–Cesàro sum requires convergence of all a_j, not of a subsequence. Without a_j→a, the factor-2 term is not rigorously identified, and the uniqueness argument in Step 3 cannot compensate because it already assumes the SPDE limit. This is the load-bearing gap in Theorem 3.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the sequential (lower-triangular) interacting diffusion system (1.4), in which particle i interacts only with the empirical measure of its predecessors. It introduces an incremental path-space relative entropy R_i(t) and proves Theorem 1: R_i(t) ≤ C_T (i−1)^{-1}(1+M_{0,i}), with the consequences S_N(T) ≲ log N and quantitative tail propagation of chaos. Theorem 2 establishes the canonical N^{-1/2} rate for the empirical measure in negative Sobolev norms, E[sup_t ||μ^N_t−ρ_t||_{H^{-β}}^2] + E∫||μ^N_t−ρ_t||_{H^{-β+1}}^2 dt ≤ C_T/N, under b≡0 and a regularity condition on K⋆ρ. Theorem 3 states a Gaussian fluctuation limit in C([0,T];H^{-β',-m}) for η^N=√N(μ^N−ρ), with the limiting SPDE (1.10) containing a factor-2 feedback term. Section 6 extends the incremental entropy method to general sequential weights and states a threshold phenomenon. The main technical machinery is a Girsanov identity, a martingale-difference replacement of predecessor empirical measures by averaged conditional measures, an upper-envelope closure, and a negative Sobolev energy estimate.","tokens_in":39977,"tokens_out":9804,"duration_ms":100064,"significance":"Theorems 1 and 2 are substantial and, as far as I can verify, correctly proven. The sharp (i−1)^{-1} decay for a directed non-exchangeable system and the logarithmic global entropy are genuinely new and well motivated by the online/streaming interpretation. The negative-Sobolev N^{-1/2} bound is also an honest improvement over what global entropy arguments would give. The factor-2 feedback in the fluctuation SPDE is an interesting and potentially important structural effect of the sequential architecture. However, the identification step in Theorem 3 currently contains a load-bearing gap, and the regularity assumptions for Theorem 3 are not sufficient as stated. The paper's proofs are largely transparent and the estimates are explicit, which is a strength; with a complete proof of the fluctuation limit the contribution would be significant.","major_comments":[{"comment":"The identification of the factor-2 term rests on the assertion 'We have a_j→a as j→∞', where a_j = ∫_s^t ⟨ρ_r,(K*η^j_r)·∇φ⟩ dr and a = ∫_s^t ⟨ρ_r,(K*η_r)·∇φ⟩ dr. This is not proved and does not follow from the preceding tightness argument. Tightness gives at most subsequential convergence of the global fields η^N to some η; the prefix fields η^j = √j(μ^j−ρ) are not shown to converge to the same η, and no estimate controlling ||η^j−η|| is supplied. The Stolz–Cesàro passage requires convergence of the entire sequence a_j, not of a subsequence. This is the load-bearing step for the factor-2 term in (1.10). Without a proof of a_j→a, the martingale-problem identification of Theorem 3 is incomplete.","section":"§5.1, Step 2"},{"comment":"The claim that ρ_0∈S(R^d) and 'standard parabolic bootstrapping' imply ρ∈C([0,T];S(R^d)) is not justified for bounded measurable K. In general K⋆ρ_t is only bounded measurable; it need not lie in W^{m,∞}, so high-order Sobolev bounds on ρ_t do not follow solely from smoothness of ρ_0. Consequently the well-definedness of the product ρ_t(K⋆η_t) in (1.10) and the estimates such as ||L_r φ||_{H^β} ≲ ||φ||_{H^{β'}} are not established. Theorem 3 should either impose an explicit regularity condition on K⋆ρ_t, as Theorem 2 does in (1.8), or assume smoothness of K. This is a second load-bearing gap for Theorem 3.","section":"§5.1, first paragraph"},{"comment":"The general-weight result is stated as a theorem, but the proof is only a 'Sketch of proof.' The weighted upper-envelope closure is not written out, and the threshold phenomenon for r>1 relies on unproved assertions about θ_{i−1} and non-convergence of R_i. If Theorem 4 is to be part of the paper's formal results, it needs a complete proof; otherwise it should be presented as a conjecture or heuristic extension.","section":"§6, Theorem 4"}],"minor_comments":[{"comment":"There are frequent typos: 'Propsition', 'postive integer', 'inequlaity', 'concerntration', 'defintions', 'sup|b|<∞' in Section 6, etc. The paper would benefit from a careful proofreading pass.","section":"§2 and §3"},{"comment":"Theorem 2 and the whole negative-Sobolev section assume b≡0, while the introduction and Theorem 1 allow a bounded drift b. This restriction is fine if stated clearly at the start of Section 4, but the presentation could make it more prominent.","section":"§4, Theorem 2"},{"comment":"Uniqueness is asserted via a one-line energy estimate and then attributed to [31, Section 4]. Please state precisely which theorem or proposition of [31] is being invoked, since the factor-2 drift and the weighted spaces are not literally the exchangeable setting of [31].","section":"§5.1, Step 3"},{"comment":"The tightness argument imports the martingale modulus bound from [31, Lemmas 3.1–3.2] without checking that the sequential drift decompositions verify the same hypotheses. At least the relevant verifications should be indicated.","section":"§5.1, Step 1"},{"comment":"The comparison of the three SPDE limits is useful, but the 'add–subtract trick' displaying the decomposition of the sequential interaction term contains a typo: the index in the last sum should be 1/√(i−1)δ_{X^i}(K⋆η^{i−1}), which is later written correctly in Section 5.","section":"§1.2, Remark 1.7"}],"recommendation":"major_revision","confidential_remarks":"The paper's core entropy estimates (Theorems 1 and 2) appear sound and are a real contribution. The main obstacle is Theorem 3: the factor-2 identification contains an unjustified a_j→a step, and the regularity hypotheses are not sufficient as stated. Both issues seem fixable by adding a convergence argument for prefix fluctuations (or a different route to the factor 2) and by imposing an explicit regularity condition on K⋆ρ_t. Section 6 also needs to be either fully proved or demoted from theorem to conjecture. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real contribution on the entropy side. The incremental R_i decay ~1/(i-1), the log N global entropy, and the N^{-1/2} negative-Sobolev empirical convergence are proven in detail and I could not find a hole in the arguments. The proof strategy—Girsanov representation, martingale-difference replacement with conditional measures, upper-envelope closure—is coherent and well-executed. The extension to weighted schemes with effective sample size is a nice bonus, though Theorem 4 is only sketched.\n\nThe soft spot is Theorem 3. The identification of the limit SPDE (1.10) hinges on the claim 'We have a_j→a as j→∞' in Section 5.1, where a_j involves the prefix fluctuation η^j. That convergence is not proven and does not follow from the preceding tightness. Tightness of the global sequence η^N gives subsequential convergence, and uniqueness of the limit has not been established at that point, so you can't conclude that prefixes converge to the same η. The Stolz–Cesàro sum requires convergence of every a_j, not just along a subsequence. This is a genuine, load-bearing gap. It may be fixable—for example, by first proving convergence of the whole η^N via a separate tightness-plus-uniqueness argument, then transferring to prefixes, or by an explicit bound on ∥η^j−η^N∥—but as written the factor-2 SPDE limit is not rigorously identified. The well-posedness of the limit SPDE imported from your prior work [31] is fine in principle, but it doesn't patch this.\n\nMinor issues: Theorem 2 assumes b≡0 and the W^{m,∞} regularity (1.8); the authors flag this, so it's an honest limitation. There are also typos and a few mildly sloppy phrases, none affecting the math.\n\nWho does this help? People working on quantitative propagation of chaos, especially non-exchangeable/triangular structures, and anyone simulating McKean–Vlasov SDEs with online particle updates. The entropy results will be cited.\n\nRecommendation: Do send this to peer review. A good referee can ask for the gap in Theorem 3 to be closed or stated as a conditional result; the rest of the paper deserves publication.","headline":"Theorems 1–2 on incremental entropy and N^{-1/2} empirical convergence look solid; the Gaussian fluctuation theorem has a genuine gap at the a_j→a identification step, so the paper needs a serious referee rather than acceptance as is.","tokens_in":40422,"tokens_out":4489,"would_cite":true,"duration_ms":43927,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","60F05","60H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A sequential, online particle system approximating the McKean–Vlasov diffusion achieves the same N^{−1/2} statistical accuracy as the exchangeable system, with a factor-2 feedback in its Gaussian fluctuation limit.","keywords":["sequential interacting diffusions","McKean–Vlasov","incremental relative entropy","propagation of chaos","Gaussian fluctuations","negative Sobolev norms","Girsanov identity","non-exchangeable mean field"],"falsifier":"Compute directly, for a one-dimensional linear example with explicit solvable kernel and initial density, the limiting variance of ∫φ dη_N at a fixed time; if the prefactor differs from 2, or if the asserted convergence a_j→a fails for a non-smooth initial density, the identification of the SPDE limit collapses. Alternatively, simulate the sequential system with i.i.d. initial data and compare the empirical covariance with the prediction of the factor-2 SPDE.","tokens_in":39522,"feed_emoji":"🧮","tokens_out":3585,"duration_ms":34201,"temperature":0.7,"pith_summary":"The paper studies a lower-triangular particle system where particle i interacts only with the empirical measure of its predecessors. It proves that the incremental path-space relative entropy of particle i decays as 1/(i−1), which is the sharp sampling barrier, and that summing these increments gives a global entropy of order log N. From this, the empirical measure converges at the canonical N^{−1/2} scale in negative Sobolev norms, and the fluctuation field converges to a linear SPDE with an explicit factor-2 feedback term absent in exchangeable systems. If correct, this shows that the computationally attractive online/sequential architecture does not sacrifice Monte Carlo accuracy.","feed_headline":"Sequential particles hit N^{-1/2} accuracy with online updates","feed_subtitle":"Lower-triangular McKean–Vlasov approximation matches exchangeable Monte Carlo and adds a factor-2 feedback to its fluctuation limit.","key_machinery":"The argument is carried by the incremental path-space relative entropy R_i(t), computed through a Girsanov identity as the quadratic energy of the drift mismatch between the conditional law of particle i and an independent limit copy. The predecessor empirical measure is replaced by an average of conditional laws, producing a martingale-difference term (handled by sub-Gaussian concentration under a decoupled reference law) and a predictable bias term (handled by Pinsker's inequality). An upper-envelope closure and a discrete Hardy inequality yield the sharp decay; a negative Sobolev energy estimate then converts the entropy control into the N^{−1/2} empirical-measure rate.","core_discovery":"The central claim is that the incremental relative entropy, which measures how far particle i's path law deviates from an independent McKean–Vlasov copy given the predecessors, decays as (i−1)^{−1}. This rate is optimal: even with i.i.d. inputs drawn from the limit law, the sampling variance of the interaction field forces exactly this scale. Summing the increments yields a global path-space entropy of order log N, and the same mechanism yields N^{−1/2} convergence of the empirical measure in H^{−β} and a Gaussian limit whose linear feedback term carries a factor 2, originating from the weighted sum ∑_{i=2}^N 1/√(i−1) ~ 2√N.","pith_inferences":["The factor-2 accumulation mechanism likely persists in streaming and random-batch variants of sequential samplers, wherever predecessor averages contribute at the 1/√(i−1) fluctuation scale; a testable extension is to measure the feedback coefficient in such variants.","The sharp incremental entropy bound suggests an analogous 1/(i−1) decay for time-marginal Fisher information, which could open a route to uniform-in-time and singular-kernel sequential propagation of chaos, a direction the paper sketches but does not fully develop.","The Gaussian limit is the least settled part of the paper: the identification step relies on an unproved convergence a_j→a and on imported well-posedness of the limit SPDE, so a direct proof of uniqueness in the weighted Sobolev space would close the main remaining gap."],"forward_implications":["Sequential particle simulation with online refinement achieves the same N^{−1/2} empirical-measure accuracy as fully-coupled mean-field Monte Carlo, despite its O(N) marginal cost for adding particles.","Tail blocks of m particles are chaotic with total-variation error O(√(m/N)); in particular, the terminal particle's path law is within O(N^{−1/2}) of the McKean–Vlasov law.","The factor-2 feedback in the limiting SPDE is a testable signature of the directed architecture, distinguishing it from exchangeable and independent-copy systems.","For general step-size sequences, the method predicts incremental entropy decay governed by the effective sample size θ_{i−1}^{−1}, with a phase transition at the divergence of ∑α_i."],"fun_headline_variants":["Sequential diffusions reach N^{-1/2} convergence","Lower-triangular particles match exchangeable rate","Incremental entropy decay is sharp in causal systems","Factor-2 feedback in McKean-Vlasov fluctuation limit"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The Gaussian fluctuation limit depends on the limit SPDE being well-posed in the weighted Sobolev space and on the asserted convergence a_j→a in the identification step; neither is proved in this paper.","fun_headline_variants_meta":{"raw":{"variants":["Sequential diffusions reach N^{-1/2} convergence","Lower-triangular particles match exchangeable rate","Incremental entropy decay is sharp in causal systems","Factor-2 feedback in McKean-Vlasov fluctuation limit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":2884,"prompt_tokens":741,"completion_tokens":2143,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":2077}},"tokens_in":485,"tokens_out":2143,"duration_ms":16601,"temperature":1.0,"reasoning_tokens":2077,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:36:10.390693+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute directly, for a one-dimensional linear example with explicit solvable kernel and initial density, the limiting variance of ∫φ dη_N at a fixed time; if the prefactor differs from 2, or if the asserted convergence a_j→a fails for a non-smooth initial density, the identification of the SPDE limit collapses. Alternatively, simulate the sequential system with i.i.d. initial data and compare the empirical covariance with the prediction of the factor-2 SPDE.","supporting_citations":[],"review_version":1}