{"id":"44c04825-c9d1-4316-9a63-3434b2ac00f9","arxiv_id":"2608.00838","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For the spectral regularization of Dean-Kawasaki, superpolynomial weak convergence to N-particle empirical measures holds if and only if the initial density is bounded away from zero; otherwise only polynomial rates are possible.","lead":"A regularized Dean-Kawasaki equation that does not preserve positivity approximates Brownian particles with superpolynomial accuracy only when the initial density is bounded away from zero. The paper proves this with upper and lower error bounds, and numerical experiments illustrate when negative densities spoil the approximation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"Good-faith reading: the paper's main claims are proven with explicit decompositions. I checked the HJB duality with complex phi (Lemmas 5.1-5.3), the negative-part control (Section 3), the initial approximation (Proposition 4.1), and the lower bound (Section 2). The most delicate step is the exponential suppression of the negative part, which depends on Assumption 1; however the assumption is exactly the scale on which the spectral regularization is meaningful and is satisfied by the chosen epsilon. The lower bound's one-dimensional example is internally consistent: identity (25), the q_epsilon lower bound for chi=1_{[-1,1]}, and the heat-kernel estimate on B_epsilon all line up, and a smooth chi would behave the same for the single Fourier mode k=1. I do not see a missing proof or a circularity. The absence of code and the non-smooth kernel in the numerics/lower bound are disclosed and do not undermine the central mathematical results. Hence the Reader's ACCEPT verdict should remain unchanged.","tokens_in":33609,"tokens_out":37472,"duration_ms":438145,"concrete_test":"Independently verify the lower-bound construction by Monte Carlo: for d=1, rho0=2*1_[1/2,1], phi=sin(2*pi*x), t=1, epsilon=(log N)/N, compute the centered variance difference D in (24) with the spectral Galerkin scheme for N=10^4..10^6 and test that D decays polynomially and that the B_epsilon negative-mass term scales as epsilon^{5/2}|log epsilon|^{-3}; repeat with a smooth mollifier chi to confirm the non-smooth kernel is not essential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I found no load-bearing flaw in the central argument. The weakest condition is Assumption 1 (delta <= kappa <= 1), exactly as the Reader states: it guarantees the fluctuations v_t are small compared with p_t*u_0, so the negative part is exponentially suppressed in the regime rho_min > 0. This assumption is explicit and is satisfied by the choice epsilon = N^{-1/d}(log N)^{3/d} used for the superpolynomial rate. The upper bound follows from Theorem 1.3 and Corollary 3.5, and the lower bound Theorem 2.1 is built from the martingale identity (25) and Lemma 2.4; the 1D example is a valid counterexample for rho_min = 0. The only minor caveat is that (24) bounds the centered/dynamic variance difference rather than the full Laplace weak error, but for rho_min = 0 the full error already contains an O(N^{-1}) initial term, so the qualitative conclusion is unaffected. Numerical experiments lack code, but they are not load-bearing for the mathematical claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes a spectrally regularized Dean--Kawasaki SPDE, u_t = (1/2)Δu_t + N^{-1/2}∇·(K_ε*(f(u_t^+) dW_t)), with f(x) ≈ √x, as an approximation to the empirical measure of N independent Brownian particles. The main upper-bound result, Theorem 1.3, gives a uniform-in-time Laplace-transform weak error decomposed into an initial approximation error, a Fourier-cutoff error, and a negative-part error. For initial densities bounded away from zero, choosing ε = N^{-1/d}(log N)^{3/d} makes the dynamic part of the error superpolynomially small for smooth test functions. When the initial density may vanish, the upper bound worsens, and in a one-dimensional example with ρ_min = 0, Theorem 2.1 proves a lower bound of order N^{-3/2} ε^{5/2} |log ε|^{-3} for the difference of the dynamic variances, which is claimed to rule out superpolynomial convergence. The paper also contains well-posedness results, detailed Besov and heat-kernel estimates, a Cole--Hopf treatment of the complex Hamilton--Jacobi--Bellman equation, and numerical experiments.","tokens_in":33814,"tokens_out":26304,"duration_ms":297046,"significance":"If the claims hold, this is a significant contribution: it is, to my knowledge, the first quantitative analysis of a non-positivity-preserving spectral approximation of the Dean--Kawasaki equation in the low-density regime, and it identifies positivity preservation as the key mechanism separating superpolynomial from polynomial accuracy. The proof strategy is substantial and mostly self-contained: error decompositions based on Itô calculus, the KLvR19 duality, explicit Cole--Hopf regularity estimates, and the spectral gap of the Laplacian are all used in a coherent way. A particular strength is that the rates are derived from the model equations, not fitted; the constants are explicit and the assumptions are stated. The numerical experiments are illustrative and not load-bearing for the mathematical results. However, two issues -- the precise quantity estimated by the lower bound and a false inequality in the proof of Corollary 1.7 -- need to be corrected before the claims as stated are fully supported.","major_comments":[{"comment":"The lower bound (24) is proved for E[(<u_t,φ>-<p_t*u_0,φ>)^2] - E[(<μ_t,φ>-<p_t*μ_0,φ>)^2], i.e. for the 'dynamic' variance around the random mean p_t*μ_0. But the weak error studied in Theorem 1.3 is |E[e^{<u_t,φ>}]-E[e^{<μ_t,φ>}]|, whose second-moment analogue is the raw difference E[<u_t,φ>^2] - E[<μ_t,φ>^2]. These quantities differ by B_N = E[<μ_0-ρ_0, p_t*φ>^2], which is of order N^{-1} for the test function φ = sin(2πx). Because the displayed lower bound L_N is much smaller than N^{-1} for the assumed scaling ε ≳ N^{-a}, the lower bound on the dynamic part does not by itself imply a lower bound on the raw weak error: cancellation with B_N is not excluded. The statements in Remark 2.2 and the introduction that the example 'rules out superpolynomial convergence' therefore either need a separate argument bounding the raw weak error from below (e.g. by showing B_N does not cancel the d","section":"Section 2, Theorem 2.1 and Remark 2.2"},{"comment":"The superpolynomial statement is made for the 'dynamic part of the weak error', not for the full Laplace weak error. The full error in Theorem 1.3 contains the initial approximation term N^{-1/2} ε^α ||φ||_{C^α}. For any fixed α this term decays only polynomially in N when ε = N^{-1/d}(log N)^{3/d}. To conclude superpolynomial convergence for the full Laplace error on C^∞ test functions one must let α grow with the desired polynomial degree and control the growth of ||φ||_{C^α}. This is likely true and easily made precise, but it is not stated. The abstract's phrase 'superpolynomial convergence rate for smooth test functions' should be qualified or the argument supplied.","section":"Section 1, Eq. (3) and Remark 1.5"},{"comment":"The inequality κ√ρmax m ≤ κ^2ρmax + κ^2m is false in general (e.g. κ=0.1, ρmax=2, m=3 gives 0.424 on the left and 0.05 on the right). The subsequent bound m!/R^{m-2} ≲_m ||φ||_∞^{m-2} can nevertheless be recovered: one has R^2 κ^2ρmax ||φ||_∞^2 = m^2κ^2ρmax/(1+κ√ρmax m)^2 ≤ 1, and the factor (1+κ√ρmax m)^{m-2} can be absorbed into an m-dependent constant because κ ≤ 1 and m is fixed in this corollary. So the result is likely correct, but the proof as written contains an incorrect step that should be replaced.","section":"Corollary 1.7, proof around Eq. (22)"}],"minor_comments":[{"comment":"The notation ||id - f^2||_∞ should be defined explicitly, e.g. sup_{x ≥ 0} |x - f(x)^2|, since f is only defined on R_+ and 'id' is otherwise ambiguous.","section":"Notation, Theorem 1.3"},{"comment":"The identity E[v_t(x)^-] = (1/2)E|v_t(x)| follows from E[v_t(x)] = 0, but this is stated very quickly. A one-line reminder would improve readability.","section":"Section 2, Lemma 2.4"},{"comment":"The numerical experiments do not include code, random seeds, or a precise description of the Monte Carlo error bars. This is acceptable since the experiments are not load-bearing, but the captions and text should make that status explicit.","section":"Section 6"},{"comment":"Several figure captions contain stray 'L' or 'LLL' text ('L L L LLL'), which appears to be a formatting artifact and should be corrected.","section":"Figures"},{"comment":"In the bound (45), the p ∈ {2,∞} minimization is clear, but the case p = ∞ is used in the main text without commenting that ρ_max^{1/∞} := 1. This is already noted in the proposition statement, but a short reminder in Theorem 1.3 would help.","section":"Section 4, Proposition 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong contribution to the stochastic PDE/numerical analysis literature. The main theorems are plausible and the proof strategy is convincing overall; I do not see a fundamental flaw in the upper-bound argument. The revision should focus on (i) disentangling the lower bound's target quantity from the Laplace weak error, and (ii) fixing the incorrect Young-type inequality in Corollary 1.7. Both appear fixable within the scope of the manuscript, but as written they affect load-bearing statements. The numerical section is not essential to the mathematical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this when you want to know what positivity preservation actually buys you in Dean–Kawasaki approximations. The paper gives the first superpolynomial weak-error result for a non-positivity-preserving scheme in the high-density regime, and a lower bound in 1D showing that if the initial density can vanish, superpolynomial convergence is impossible. The spectral regularization (convolving the noise coefficient with a mollifier before the divergence, with a positive-part cutoff and no noise truncation) is a genuinely new construction relative to the finite-difference and positivity-preserving SPDE approaches.\n\nWhat the paper does well: the proofs are explicit and honest. The error is decomposed into initial approximation, Fourier cutoff, and negative-part terms. The negative part is controlled by exponential tail bounds on the fluctuation field, which is the right tool. The superpolynomial rate comes from choosing epsilon = N^{-1/d}(log N)^{3/d} so the negative-part term is exponentially small. The lower bound is built from the martingale identity (25) and a localization argument: on a small interval at distance O(epsilon) from the support, the heat kernel has no mass yet while the nonlocal noise creates fluctuations. Using chi = 1_{[-1,1]} is a minor irregularity, but the authors flag it and it is not load-bearing: the lower bound only needs to rule out superpolynomial rates, and it does that cleanly. There are no fitted constants; the rates are theorem consequences.\n\nSoft spots, in proportion. The weakest point is Assumption 1, the fluctuation scaling delta <= kappa <= 1 (N >= epsilon^{-d}(1 or d log(1/epsilon))). It is explicit and standard for this kind of high-density regime, so I do not call it a flaw, but it is load-bearing: if it fails, the negative part is not suppressed and the superpolynomial claim collapses. The lower bound (24) is for the variance difference, not the full Laplace transform error, but the conclusion is the same: no superpolynomial weak convergence when rho_min = 0. The numerical experiments have no code or full implementation details, but they are illustrative and the theory stands on its own. Notably, the authors honestly point out that in d=1,2 the low-density upper bound is worse than the positivity-preserving scheme of DKP24, so the hybrid-strategy conclusion is measured, not oversold.\n\nWho this is for: anyone working on SPDE approximations of particle systems, especially the Dean–Kawasaki program. It resolves a concrete open question and sets a benchmark for what non-positivity costs. A serious referee should engage with it. I would send it to peer review and recommend acceptance, with minor requests for code and a small clarification on the lower bound's domain of validity. The math looks solid.","headline":"A well-executed analysis of a spectral Dean-Kawasaki regularization: superpolynomial weak error when the density is bounded away from zero, and a lower bound showing that positivity loss genuinely caps the rate.","tokens_in":34301,"tokens_out":2102,"would_cite":true,"duration_ms":23895,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","60J60","60H35","65C30","60K35"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a spectrally regularized Dean–Kawasaki equation, weak accuracy is superpolynomial when the initial density is bounded away from zero, and only polynomial when the density can vanish.","keywords":["Dean-Kawasaki equation","positivity preservation","spectral regularization","weak error","Laplace transform","superpolynomial convergence","Brownian particles","Hamilton-Jacobi-Bellman duality"],"falsifier":"Run the spectral scheme on the one-dimensional example rho_0=2*1_{[1/2,1]} with phi(x)=sin(2pi x), take epsilon=N^{-a} for small a>0, and measure the variance error at t=epsilon^2/log(1/epsilon) for increasing N. If the error decays faster than N^{-3/2}epsilon^{5/2}|log epsilon|^{-3}, for instance superpolynomially, the lower bound (24) is false.","tokens_in":33509,"feed_emoji":"📉","tokens_out":9211,"duration_ms":100192,"temperature":0.7,"pith_summary":"The paper studies a spectrally mollified version of the Dean–Kawasaki equation, the exact SPDE for the empirical density of N independent Brownian particles, and asks how much accuracy is lost when the approximation does not preserve positivity. Its main positive result is that if the initial density has a positive lower bound, the Laplace-transform weak error decays faster than any power of N, uniformly in time, at the optimal mollification scale. Its main negative result is that when the initial density touches zero, this superpolynomial behavior disappears; in a one-dimensional example a lower bound blocks any faster-than-polynomial convergence. The paper thus identifies positivity preservation as the mechanism that decides the achievable convergence order, and it suggests hybrid schemes that use the spectral approximation where density is high and switch to positivity-preserving or particle methods where it is low.","feed_headline":"Superpolynomial Dean-Kawasaki accuracy needs strictly positive density","feed_subtitle":"When the density touches zero, a one-dimensional lower bound blocks any faster-than-polynomial weak error.","key_machinery":"The argument runs on the split u_t = p_t*u_0 + v_t with fluctuation parameters delta=epsilon^{-d/2}N^{-1/2} and kappa=delta*sqrt(1 or d log(1/epsilon)). The inequality u^-_t <= v^-_t * 1_{||v_t||_inf > u_min} turns all positivity loss into a large-deviation estimate for ||v_t||_inf, giving exponential suppression of u^- when rho_min>0. On the deterministic side, duality of the empirical measure with the complex Hamilton-Jacobi-Bellman equation and its explicit Cole-Hopf solution phi_s = N log(p_{t-s} * exp(phi/N)) split the Laplace-transform error into initial, cutoff, and negative-part terms; the spectral gap of the Laplacian makes these uniform in time.","core_discovery":"The paper's central claim is a dichotomy governed by the minimum of the initial density. Decomposing the regularized solution as u_t = p_t*u_0 + v_t, the negative part u^- is exponentially small when rho_min>0: its L1 moments are at most delta * rho_max^{1/2} exp(-c rho_min^2/(kappa^2 rho_max)). With epsilon=N^{-1/d}(log N)^{3/d} and f(x)=sqrt(x_+), the Laplace-transform weak error becomes O(N^{-1-alpha/d}(log N)^{3 alpha/d}) for every alpha>0 and phi in C^alpha, uniformly in t. When rho_min=0 the exponential factor disappears; optimizing epsilon gives O(N^{-(3alpha+d)/(2alpha+d)}), and in a one-dimensional example with rho_0=2*1_{[1/2,1]} and phi(x)=sin(2pi x) the paper proves a lower bound","pith_inferences":["The same lower-bound mechanism, nonlocal convolution feeding mass into regions the heat flow has not yet reached, should apply to any non-positivity-preserving spatial discretization with a high-order nonlocal stencil, so the polynomial barrier likely extends beyond this spectral scheme.","For interacting particle systems whose mean-field equation can create low-density regions, the maximum-principle reasoning implies the strictly positive condition is fragile: even if it holds initially, dynamics may destroy it, so accuracy guarantees should be based on the low-density regime or on positivity-preserving methods.","The Monte Carlo estimate of the probability of loss of positivity could serve as a practical diagnostic: run a few short trajectories, measure how often min_x rho_L^n(x) < 0, and use that to decide when to trust the superpolynomial regime.","A sharper matching of upper and lower bounds in the rho_min=0 case (N^{-3/2}epsilon^{-d/2} versus N^{-3/2}epsilon^{5/2}|log epsilon|^{-3}) would pin down the exact polynomial exponent; the present bounds leave a gap in d and epsilon powers."],"forward_implications":["With strictly positive initial density and smooth test functions, the spectral scheme reaches a weak error of order N^{-1-alpha/d}(log N)^{3 alpha/d} for every alpha>0, faster than any fixed polynomial order.","All moments of <u_t, phi> are controlled by the Laplace-transform error through Cauchy's integral formula, so the superpolynomial rate transfers from generating functions to moments.","When rho_min=0, even the best choice of epsilon gives a polynomial bound O(N^{-(3alpha+d)/(2alpha+d)}), and the one-dimensional lower bound shows the polynomial obstruction is genuine, not an artifact of the proof.","The error estimates are uniform in time, the first such uniformity for approximating Dean-Kawasaki dynamics, so the rates hold for arbitrarily long simulation horizons.","In low densities, positivity-preserving approximations can be more accurate than this spectral regularization, suggesting hybrid schemes combine both."],"supporting_citations":[{"why":"Derives the Dean-Kawasaki equation as an exact SPDE for the empirical measure of independent Brownian particles; sets up the object the approximation must match.","marker":"[Dea96]"},{"why":"Supplies the martingale duality between the empirical measure and the Hamilton-Jacobi-Bellman equation, whose explicit solution drives the Laplace-transform error analysis.","marker":"[KLvR19]"},{"why":"Establishes the high-order finite-difference regime and the exponential tail control of the negative part that the paper adapts for its spectral scheme.","marker":"[CF23]"},{"why":"Provides the positivity-preserving approximation whose low-density rates the paper compares against; the comparison shows the price of losing positivity.","marker":"[DKP24]"},{"why":"Provides the Besov-space conventions and Bernstein inequalities used throughout the main estimates, including frequency-localization of K_epsilon and v_t.","marker":"[BCD11]"},{"why":"Supplies the composition estimate for smooth functions of Besov functions used to prove the uniform Cole-Hopf regularity estimates in Lemma 5.3.","marker":"[RS96]"}],"fun_headline_variants":["Zero density kills superpolynomial Dean-Kawasaki error","Positivity decides Dean-Kawasaki convergence speed","Dean-Kawasaki: positive density yields exponential wins","Vanishing density caps Dean-Kawasaki weak error","New bound: Dean-Kawasaki needs positive start for fast error"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The estimates assume the fluctuation scaling N >= epsilon^{-d}(1 or d log(1/epsilon)), meaning many particles per resolved spatial cell, so noise fluctuations stay small compared with the deterministic density; without this, high-probability positivity of u_t and the exponential suppression of its negative part collapse.","fun_headline_variants_meta":{"raw":{"variants":["Zero density kills superpolynomial Dean-Kawasaki error","Positivity decides Dean-Kawasaki convergence speed","Dean-Kawasaki: positive density yields exponential wins","Vanishing density caps Dean-Kawasaki weak error","New bound: Dean-Kawasaki needs positive start for fast error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00014,"raw_usage":{"total_tokens":973,"prompt_tokens":694,"completion_tokens":279,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":200}},"tokens_in":438,"tokens_out":279,"duration_ms":3564,"temperature":1.0,"reasoning_tokens":200,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:11:41.077052+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the spectral scheme on the one-dimensional example rho_0=2*1_{[1/2,1]} with phi(x)=sin(2pi x), take epsilon=N^{-a} for small a>0, and measure the variance error at t=epsilon^2/log(1/epsilon) for increasing N. If the error decays faster than N^{-3/2}epsilon^{5/2}|log epsilon|^{-3}, for instance superpolynomially, the lower bound (24) is false.","supporting_citations":[],"review_version":1}