{"id":"6a98d0de-4940-4edf-b631-fd8a5f44dd59","arxiv_id":"2412.15581","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A momentum-corrected random batch method is proposed and claimed to reduce the error of the standard random batch method for singular interaction kernels.","lead":"This paper adds a momentum-like memory term to the Random Batch Method for particle simulations, aiming to fix its failure on singular interactions. The authors report numerical accuracy gains on several test systems, though the supporting proof has serious gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2 proves a bound for the convex-combination recurrence (3.1)-(3.2), not for Algorithm 1's K_cor-forced SDE; the two trajectories already differ in a scalar linear test, so the claimed 'mathematically proven' smaller error does not apply to the proposed method.","rationale":"The reader's truncation concern is directionally right but technically imprecise: the sum of discarded coefficients is β^2(1−β^{n−2}) = O(β^2) uniformly in n; the real problems are the missing uniform bound on u_i^p and the later dropping of O(β^2). The more fundamental defect is the mismatch between (3.1)-(3.2) and Algorithm 1: the algorithm evolves positions through an SDE with an averaged interaction force, whereas the proof analyzes a convex combination of RBM solutions. For a scalar linear system the two already differ at the first step, so Theorem 3.2 cannot support the paper's central claim about the implemented RBM-M. Numerical experiments in Section 4 implement Algorithm 1 and therefore cannot validate Theorem 3.2. Independently, the O(β^2) term in (3.4) blocks the claimed asymptotic ≲C(1−β^2)√τ when β is fixed. No code, machine-checked proof, or precise reference-solution definition is provided, so the REJECT verdict remains unchanged.","tokens_in":23024,"tokens_out":12979,"duration_ms":109101,"concrete_test":"Run the following scalar check: p=N, V=0, σ=0, K(x)=−x, x_0=1, τ=0.1, β=0.5. One Algorithm 1 step with the averaged force satisfies x_1=x_0+τ(βK(x_0)+(1−β)K(x_1)), yielding x_1=(1−τβ)/(1+τ(1−β))≈0.9048. The recurrence (3.1)-(3.2) with u_1=(1−τ)x_0 yields ũ_1=β+(1−β)(1−τ)=0.95. Since 0.9048≠0.95, Theorem 3.2's object is not Algorithm 1; a corrected proof must start from the K_cor-forced SDE and would need a different error analysis. Optionally repeat for T=0.02, τ=0.001, β=0.1, N=10,000, comparing Algorithm 1 to a simulation of the recurrence, and report the L^2 difference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 defines RBM-M by solving, on each batch, the SDE whose interaction is K_cor(X_i−X_j)=βK_{t−1}(X_i−X_j)+(1−β)K_t(X_i−X_j). Section 3.1 instead states that the RBM-M solution satisfies the recurrence ũ_i^p=βũ_{i−1}^p+(1−β)u_i^p (Eqs. 3.1-3.2), i.e., a convex combination of RBM solutions. For a nonlinear or even linear system, a weighted average of position sequences is not the sequence produced by an SDE with an averaged interaction force. Example with p=N, V=0, σ=0, K(x)=−x, τ=0.1, β=0.5: the recurrence gives ũ_1=0.95x_0, while the Algorithm 1 Euler step gives x_1=(1−τβ)/(1+τ(1−β))x_0≈0.9048x_0. Hence Theorem 3.2 bounds an auxiliary method, not the algorithm numerically tested in Section 4. Moreover, even for the auxiliary recurrence, the final step of (3.4) discards a standalone O(β^2) term; with β fixed and τ→0 that term does not vanish, so the conclusion 'error smaller than RBM' does not follow from the displayed bound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a momentum-correction variant of the Random Batch Method (RBM-M) for interacting particle systems, claiming to reduce error relative to the standard RBM when interaction kernels have large singularities. The method regularizes singular kernels and introduces a momentum parameter beta, and the paper presents theoretical error bounds, a variance analysis, and numerical experiments. The central theoretical claim is Theorem 3.2, which asserts an error bound of the form C(1-beta^2) sqrt(tau/(p-1)+tau^2) + C beta(1-beta) tau + O(beta^2), supposedly proving that the error of RBM-M is smaller than that of RBM. The paper also presents unbiasedness and variance results (Theorems 3.4 and 3.5) and a relative-entropy estimate (Theorem 3.7).","tokens_in":23284,"tokens_out":10766,"duration_ms":83771,"significance":"If established, the result would be a useful contribution: an O(N) method with improved accuracy for singular interaction kernels, backed by an explicit error bound and variance reduction. The paper correctly identifies a real limitation of the standard RBM for nonsmooth kernels, and the regularized-kernel idea is reasonable. However, the proof of the central theorem does not analyze the algorithm that is actually implemented, the key truncation step in the proof is unjustified, and the variance theorem rests on a false independence assumption. These issues undermine the paper's main claim, so the significance cannot be credited from the evidence presented.","major_comments":[{"comment":"The proof truncates the expansion of \\tilde{u}_n^p to its last two terms, discarding all earlier terms as O(beta^2). This truncation is not justified: the discarded terms are (1-beta) sum_{i=1}^{n-2} beta^{n-i} u_i^p, and although each coefficient beta^{n-i} with n-i >= 2 is O(beta^2) as beta -> 0, the number of terms grows as n = T/tau, and the RBM solutions u_i^p are not shown to be uniformly bounded in i and tau. Consequently, the sum of discarded terms need not be O(beta^2), and the factor (1-beta^2) in the final bound, which arises solely from this truncation, is not supported.","section":"Section 3.1, proof of Theorem 3.2 (Eq. (3.4))"},{"comment":"The method analyzed in Theorem 3.2 is not the method presented in Algorithm 1. Section 2.3 defines RBM-M by solving, over each batch interval, the SDE with interaction kernel K_cor(X_i - X_j) = beta K_{t-1}(X_i - X_j) + (1-beta) K_t(X_i - X_j), whereas Section 3.1 (Eqs. (3.1)-(3.2)) analyzes the recurrence \\tilde{u}_i^p = beta \\tilde{u}_{i-1}^p + (1-beta) u_i^p, i.e., a convex combination of previously computed RBM solutions. These two processes differ already for a scalar linear problem: with p=N, V=0, sigma=0, K(x)=-x, tau=0.1, beta=0.5, the recurrence gives \\tilde{u}_1 = 0.95 x_0, while the Algorithm 1 Euler step gives x_1 = (1 - tau beta)/(1 + tau(1-beta)) x_0 = 0.9048 x_0. Therefore Theorem 3.2 does not provide an error bound for the algorithm that is implemented and tested in Section 4.","section":"Section 2.3 and Section 3.1 (Algorithm 1 vs. Eqs. (3.1)-(3.2))"},{"comment":"The proof explicitly assumes that K(x_i^{t1} - x_j^{t1}) and K(x_i^{t2} - x_j^{t2}) are independent for t1 != t2 because the regroupings at different times are independent. This assumption is false for the particle system (1.1): the positions themselves are strongly dependent over time through the shared drift, interaction, and Brownian paths, so the kernel evaluations at different times are correlated. Only the random batch partitions are independent. The variance decomposition in the proof is therefore invalid, and the claimed variance reduction factor (1-beta)^2/(1-beta^2) is not established.","section":"Section 3.2, Theorem 3.5 (Eq. (3.18))"},{"comment":"Even if the truncation step were valid, Eq. (3.4) does not imply that the RBM-M error is smaller than the RBM error. The displayed bound contains an additional positive term C beta(1-beta) tau + O(beta^2) that is not present in the RBM bound (Proposition 2.1). For typical parameters used in Section 4 (tau=0.001, p=360, beta=0.1), the RBM-M bound is actually larger than the RBM bound unless the O(beta^2) term is negative or the first-term reduction dominates; the paper gives no condition under which the extra term is dominated by the factor (1-beta^2). Thus the sentence 'We have mathematically proven that the error of the RBM-M is smaller than that of the RBM' does not follow from the displayed inequality.","section":"Section 3.1, Eq. (3.4) and Section 4.2"}],"minor_comments":[{"comment":"The manuscript contains numerous typographical errors and inconsistent notation (e.g., 'Lipshitz', 'Tayler', 'satisfies', and the mixing of \\tilde{\\rho} with \\tilde{\\varrho} in Section 3.3); a careful proofreading pass is needed.","section":"Throughout"},{"comment":"The line 'K(X_i - X_j)=0 when time=0' is ambiguous; it presumably initializes the momentum buffer, but the algorithm does not specify how K_{t-1} is stored or updated for each particle pair across time steps.","section":"Section 2.3, Algorithm 1"},{"comment":"The regularization parameter delta in Eq. (2.1) is never reported for the numerical experiments in Section 4, and no code is provided; this makes the numerical results difficult to reproduce and prevents verification that the experiments satisfy the smoothness assumptions of the theorems.","section":"Section 2.2 and Section 4"},{"comment":"The text states that each point in Figure 6 is the average of 10 repeated calculations, but error bars or standard deviations are not shown, and Table 4 reports only single-run errors; adding measures of variability would make the comparison between RBM and RBM-M more convincing.","section":"Section 4.3, Figure 6 and Table 4"},{"comment":"The sentence 'When the function is flatter, we use a smaller beta, and when the function is more steep, we use a larger beta (the final difference has decreased because the singularity of the function is too large, and a larger beta should be selected)' is confusing and appears to contradict the previous prescription; it should be rewritten for clarity.","section":"Section 4.3"}],"recommendation":"reject","confidential_remarks":"The central claim of the paper is not supported by the proofs: the theorem analyzes a recurrence that is different from the implemented algorithm, the truncation step is unjustified, and the variance result relies on a false independence assumption. These are load-bearing issues that would require a fundamental reformulation of the analysis rather than local corrections. The numerical experiments are suggestive for the intended singular-kernel applications, but they are not reproducible from the information given. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ken,\n\nWhat you should know first: the RBM-M idea is simple — replace the batch interaction force with an exponentially weighted average of past forces, K̃_i = βK̃_{i−1} + (1−β)K_i — and it is cheap: two multiplications and one addition per interaction, so the O(N) complexity of RBM survives. The experiments on the self-constructed singular kernels (K4–K6 in Table 3) show the error roughly halved, and the trend across β is consistent. That is genuine signal. But the theory in Section 3 does not bound this algorithm, and that gap is load-bearing.\n\nWhat is actually new: the K_cor update rule; the numerical comparison on high-singularity kernels; the variance-reduction intuition that exponential averaging shrinks batch noise. The paper is honest in places, especially Remark 2.1 (β must be small) and the \"we hypothesize\" language around singular kernels.\n\nSoft spots, in proportion:\n\n1. Theorem 3.2 analyzes the wrong object. Section 3.1 defines RBM-M as ũ_i = βũ_{i−1} + (1−β)u_i, a convex combination of RBM position sequences, while Algorithm 1 solves the SDE with the averaged force. These differ already in a scalar linear test: with K(x) = −x, τ = 0.1, β = 0.5, the recurrence gives ũ_1 = 0.95x_0, while the force-averaged step gives about 0.905x_0. Averaging forces is not averaging trajectories, so the theorem bounds an auxiliary method, not the algorithm tested in Section 4. The central claim of a \"mathematically proven\" smaller error rests on this mismatch.\n\n2. Even for the auxiliary recurrence, the conclusion overreaches. The O(β²) term in bound (3.4) is discarded in the final \"≲\" step. With β fixed and τ → 0 that term does not vanish, so the bound does not imply a smaller error than RBM; comparing upper bounds would not prove that anyway.\n\n3. Theorem 3.5 assumes kernel evaluations at different times are independent. The batch groupings are independent; the positions are not, since they evolve from the same initial data. The variance-reduction conclusion may be right in spirit, but this proof does not establish it.\n\n4. The theorem's smooth-kernel assumptions exclude the singular kernels that motivate the paper. The regularization in Section 2.2 is never tracked: δ is never specified in the experiments, which also lack error bars, code, and a precise reference-solution definition.\n\nMinor flags: \"Data availability: No data was used\" contradicts the experiments in Section 4, and the acknowledgments are unfinished; this is clearly a draft.\n\nWho is this for: people building fast particle methods who want a cheap variance-reduction trick; the numerics justify a look, but I would not cite the theorem as it stands. This paper deserves a serious referee — the idea is cheap and the numerical signal is real — but it needs major revision: prove a bound for the actual K̃-forced algorithm or present Theorem 3.2 as a simplified model, and rerun the experiments with error bars, code, and explicit δ.\n\nRecommendation: send to peer review, expect heavy revision.","headline":"The momentum idea is cheap and the numerics on singular kernels show real gains, but the main theorem bounds a position-averaging recurrence that is not the force-averaging algorithm actually implemented, and the 'mathematically proven smaller error' claim does not follow from the displayed bound.","tokens_in":23846,"tokens_out":11610,"would_cite":false,"duration_ms":94781,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F10","65K05","65Y20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding a momentum-like correction to the Random Batch Method yields a provably smaller error bound than the original method, with the same O(N) cost, and numerically cuts the error roughly in half for particle…","keywords":["random batch method","interacting particle systems","momentum correction","singular kernels","stochastic differential equations","O(N) algorithm","variance reduction","error analysis"],"falsifier":"One concrete check would be to evaluate the discarded tail directly: for a small system, compute $u_i^p$ from RBM for $i=0,\\dots,n$ and compare $(1-\\beta)\\sum_{i=0}^{n-2}\\beta^{n-i}\\|u_i^p - E\\| + \\beta^n\\|u_0^p - E\\|$ against the claimed $O(\\beta^2)$ term in Theorem 3.2 with, say, $\\beta=0.1$, $\\tau=10^{-4}$, and large $n$. If the tail is not bounded by a constant times $\\beta^2$ (with constants set by Propositions 2.1 and 3.1), the leading $(1-\\beta^2)$ factor is unsupported. A numerical companion test is to run RBM and RBM-M on kernel K6 from Table 3 at $\\tau=0.001$ and $\\beta=0.1$ and see whether the error ratio matches $(1-\\beta^2)$ or instead reflects an unrelated improvement.","tokens_in":22725,"feed_emoji":"⚛️","tokens_out":10549,"duration_ms":74018,"temperature":0.7,"pith_summary":"The Random Batch Method (RBM) cuts the cost of simulating N interacting particles from O($N^{2}$) to O(N) per step by evaluating forces on small random batches, but its error guarantee requires a smooth interaction kernel. This paper proposes RBM-M, a variant that replaces the current batch force with an exponentially weighted blend of the current and previous forces, i.e. a momentum correction. The main theoretical result, Theorem 3.2, bounds the RBM-M error by $C(1-\\beta^2)\\sqrt{\\tau/(p-1)+\\tau^2}$ plus smaller terms, compared with RBM's $C\\sqrt{\\tau/(p-1)+\\tau^2}$; because $(1-\\beta^2) < 1$ for $0 < \\beta < 1$, the paper concludes it has mathematically proven that RBM-M's error is smaller than RBM's. The extra computation is two multiplications and one addition per interaction, so the O(N) complexity is retained. The paper's stated practical payoff is for kernels with strong singularities, where plain RBM is inaccurate or fails to converge; numerical experiments report RBM-M errors roughly half those of RBM.","feed_headline":"Momentum correction cuts random-batch error for singular kernels","feed_subtitle":"Weighting past batch interactions into the force keeps O(N) cost while cutting error, especially for rough kernels.","key_machinery":"The load-bearing object is the momentum-corrected interaction kernel $K_{\\mathrm{cor}}(X_i-X_j) = \\beta K_{t-1}(X_i-X_j) + (1-\\beta) K_t(X_i-X_j)$, equivalently the exponential-weighted averaging of RBM iterates $\\tilde{u}_n^p = \\beta^n u_0^p + (1-\\beta)\\sum_{i=1}^n \\beta^{n-i}u_i^p$. The proof of Theorem 3.2 expands this average, keeps only the final two history terms, bounds the rest as $O(\\beta^2)$, and uses Proposition 2.1 (the RBM error bound from [1]) plus Proposition 3.1 (adjacent-step error $O(\\tau)$) on the surviving terms. The hyperparameter $\\beta$ controls correction strength: larger $\\beta$ damps the RBM baseline error more strongly but adds the momentum-error term $\\beta(1-\\beta)\\tau$.","core_discovery":"On the paper's own terms, the discovery is that a momentum-like correction to the random batch force produces a strictly smaller leading error term without changing the algorithm's complexity. Writing $u_i^p$ for the RBM solution at step i and $\\tilde{u}_i^p$ for the RBM-M solution, the correction is $\\tilde{u}_i^p = \\beta \\tilde{u}_{i-1}^p + (1-\\beta) u_i^p$, equivalently an exponential-weighted average of all past RBM iterates. Theorem 3.2 states that under the same convexity and smoothness assumptions used for RBM, $\\lVert E - \\tilde{u}_n^p \\rVert \\le C(1-\\beta^2)\\sqrt{\\tau/(p-1)+\\tau^2} + C\\beta(1-\\beta)\\tau + O(\\beta^2)$, leading the authors to write: 'We have mathematically proven that the error of the RBM-M is smaller than that of the RBM.' The paper also proves an asymptotic-unbiasedness result for the corrected estimator (Theorem 3.4), a variance reduction relative to RBM (Theorem 3.5), and a rescaled relative-entropy bound of order $C_1\\tau^2 + C_2/N$ between the RBM-M joint distribution and the reference solution (Theorem 3.7).","pith_inferences":["The theorem is proven only for bounded Lipschitz kernels, the regime where plain RBM already works; the claimed advantage for singular kernels rests on the heuristic that the momentum term smooths the batch estimator plus the numerical tables. A rigorous singular-kernel version would need an RBM error bound that grows with the kernel's singularity, which the current proof does not supply.","Balancing the two terms in Theorem 3.2 (damping $(1-\\beta^2)\\sqrt{\\tau}$ against momentum cost $\\beta(1-\\beta)\\tau$) suggests an optimal $\\beta$ of order $\\sqrt{\\tau}$ in the smooth regime, which would make the improvement over RBM of order $\\tau$; this is a testable prediction the paper does not state.","The momentum update is mathematically identical to exponential smoothing of the batch force estimator, so the variance reduction in Theorem 3.5 should carry over to other stochastic approximation schemes, suggesting a general principle: smoothing a high-variance, heavy-tailed batch estimator helps most when the estimator's distribution has large outliers, exactly the singular-kernel case.","Because the correction reuses only past forces, it can be combined with other fast summation ideas that also approximate the all-to-all force; the momentum correction is orthogonal to how the batch force is computed."],"forward_implications":["Under the theorem's ideal smooth-kernel conditions, the leading term of the RBM-M error bound is smaller than RBM's by the factor $(1-\\beta^2)$, so in the $\\tau\\to 0$ limit the method is claimed to be more accurate rather than merely comparable.","RBM-M keeps the O(N) per-step complexity of RBM, since the extra computation is only two multiplications and one addition per interaction.","The hyperparameter $\\beta$ provides a tunable trade-off: the numerical tables show small $\\beta$ (about 0.01) best for smoother systems and larger $\\beta$ (about 0.1) best for more singular systems, with RBM-M errors falling to roughly 43-59% of RBM's error in the singular cases.","The corrected batch estimator is asymptotically unbiased as $\\tau,\\beta\\to 0$, and for small $\\tau$ and $\\beta$ its variance is below RBM's by roughly the factor $(1-\\beta)^2/(1-\\beta^2)$, which the paper interprets as making RBM-M statistically more stable.","The rescaled relative entropy between the RBM-M joint distribution and the reference solution is bounded by $C_1\\tau^2 + C_2/N$, giving a mean-field-style guarantee alongside the trajectory error bound."],"supporting_citations":[{"why":"Supplies the proposition used as the baseline: the RBM error bound $C\\sqrt{\\tau/(p-1)+\\tau^2}$ that Theorem 3.2 improves, and the algorithm RBM-M modifies.","marker":"[1]"},{"why":"Provides the mean-field and Liouville-equation error-estimation machinery used in Theorem 3.7's relative-entropy bound.","marker":"[28]"},{"why":"Provides the sharp uniform-in-time estimate technique referenced for the entropy-analysis terms.","marker":"[29]"},{"why":"Supplies the uniform moment bound $\\sup_t \\mathbb{E}\\|\\tilde{X}_t^i\\|^p \\le C_p$ used to control the Taylor remainder in Theorem 3.7.","marker":"[30]"}],"fun_headline_variants":["Momentum fix tames singular kernels in random batch method","Random batch method gets momentum boost for rough kernels","Momentum correction shrinks error in random batch simulation","RBM-M: same cost, smaller error, handles singular kernels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, inside the norm in the proof of Theorem 3.2, all but the last two terms of the exponential-weighted history can be discarded with only $O(\\beta^2)$ error; that discarding step is what produces the $(1-\\beta^2)$ factor, and it requires the earlier RBM iterates to stay uniformly close to the true solution.","fun_headline_variants_meta":{"raw":{"variants":["Momentum fix tames singular kernels in random batch method","Random batch method gets momentum boost for rough kernels","Momentum correction shrinks error in random batch simulation","RBM-M: same cost, smaller error, handles singular kernels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001141,"raw_usage":{"total_tokens":4758,"prompt_tokens":987,"completion_tokens":3771,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":3703}},"tokens_in":603,"tokens_out":3771,"duration_ms":20058,"temperature":1.0,"reasoning_tokens":3703,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:18:00.442083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete check would be to evaluate the discarded tail directly: for a small system, compute $u_i^p$ from RBM for $i=0,\\dots,n$ and compare $(1-\\beta)\\sum_{i=0}^{n-2}\\beta^{n-i}\\|u_i^p - E\\| + \\beta^n\\|u_0^p - E\\|$ against the claimed $O(\\beta^2)$ term in Theorem 3.2 with, say, $\\beta=0.1$, $\\tau=10^{-4}$, and large $n$. If the tail is not bounded by a constant times $\\beta^2$ (with constants set by Propositions 2.1 and 3.1), the leading $(1-\\beta^2)$ factor is unsupported. A numerical companion test is to run RBM and RBM-M on kernel K6 from Table 3 at $\\tau=0.001$ and $\\beta=0.1$ and see whether the error ratio matches $(1-\\beta^2)$ or instead reflects an unrelated improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the proposition used as the baseline: the RBM error bound $C\\sqrt{\\tau/(p-1)+\\tau^2}$ that Theorem 3.2 improves, and the algorithm RBM-M modifies."},{"cited_title":"Mean field error estimate of the random batch method for large interacting particle system","cited_arxiv_id":"2403.08336","evidence_quote":"Provides the mean-field and Liouville-equation error-estimation machinery used in Theorem 3.7's relative-entropy bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the uniform moment bound $\\sup_t \\mathbb{E}\\|\\tilde{X}_t^i\\|^p \\le C_p$ used to control the Taylor remainder in Theorem 3.7."}],"review_version":1}