{"id":"ab46b3aa-c285-44dc-87a4-5963f7641a38","arxiv_id":"2505.22841","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"In a simplified linear-manifold model, kernel-smoothing the empirical score lowers sampling-noise variance and improves the asymptotic KL bound between true and generated distributions.","lead":"This paper studies why diffusion models memorize training data and shows that smoothing the estimated score, in theory, reduces this memorization. It derives asymptotic bounds on how much a kernel-smoothed score improves the trade-off between bias and variance, and interprets the result as a new density estimator.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's KL transition rests on replacing the estimator by its Gaussian CLT approximation (Eq.","rationale":"The reader's weakest-assumption analysis identifies exactly the load-bearing gap I find: the CLT in Theorem 2 is a fixed-time asymptotic statement, while Theorem 4 requires a small-time uniformity that is not controlled. The proof of Theorem 2 in Appendix A.4 uses Slutsky/Skorokhod arguments at fixed t and fixed x; the delta-method remainder has constants that blow up as t -> 0 because the denominator phi^(0)(t,x) vanishes like t^{k/2}. Since the claimed critical sample size N_c = t_N^{-k/2} is precisely where N phi^(0) is O(1), the central claim is being asserted in the regime where the Gaussian approximation is least reliable. The paper is honest about this in the Limitations, which strengthens my confidence that this is a real limitation rather than a malicious omission. I do not see an internal contradiction that would force rejection: the light-cone intuition, the bias-variance decomposition, and the numerical experiments are all consistent with the claim, and the KL bounds are upper bounds that may be valid in regimes where N t_N^{k/2} -> infinity. The concern is that the phase-transition statement is stronger than what the proof supports. The reader's conditional verdict is therefore appropriate, and my stress-test does not move it.","tokens_in":29277,"tokens_out":8086,"duration_ms":90868,"concrete_test":"Derive an explicit bound on the delta-method remainder in the proof of Theorem 2 (Appendix A.4) as a function of t and N, and use it to bound ED||m_N_t - m_t - N^{-1/2}G||^2 uniformly for t in [t_N,T]. If this remainder is not o((N t^{k/2+1})^{-1}) in the regime N t_N^{k/2} = O(1), then Eq. (13) is invalid exactly at the claimed transition scale. A complementary simulation on p_* = N(0,I_k) in R^k can check the same ratio: compute the left-hand side by Monte Carlo over datasets and compare it with tr(Sigma)/N for t = t_N with N = ceil(t_N^{-k/2}) and N = ceil(t_N^{-beta k/2}); divergence of this ratio as t_N -> 0 would confirm the failure of the Gaussian approximation in the regime where the transition is claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central bound (13) is obtained by replacing the empirical estimator with its Gaussian approximation m_t + G/sqrt(N) (Eq. 9) and then integrating the resulting score error from t_N to T. Theorem 2 proves the CLT (8) only for fixed t > 0, with a delta-method proof whose remainder is controlled at fixed t rather than uniformly as t -> 0. As t -> 0, the denominator phi^(0)(t,x) = E[exp(-||x-X||^2/(2t))] is of order (2*pi*t)^(k/2) p_*(pi(x)); after division by t^2 in the score, the omitted o(N^{-1/2}) terms have coefficients that diverge like t^{-(k/2+1)}. Theorem 4 then lets t_N -> 0, and the claimed transition scale N_c = t_N^{-k/2} is precisely the scale on which N*phi^(0) is O(1), i.e. the regime where the CLT has no valid asymptotic margin. The authors concede this in the Limitations section: for a fixed N the analysis requires the time to not be too small. Therefore the upper bounds (12)-(13) are conditional on a uniformity in t that is neither stated nor proved, and the headline phase-transition claim rests on the least secure assumption in the paper. A separate issue is that (12)-(13) are upper bounds only; the 'transition' is inferred from where the bounds blow up, not from a matching lower bound, so even with the CLT fixed the existence of a sharp transition at N_c is not established by the theorem alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies denoising diffusion models driven by the empirical score built from a finite dataset, and proposes a kernel-smoothed (mollified) empirical score as an analytically tractable proxy for the regularization induced by parametric score networks. The main theoretical results are: (i) a central limit theorem for the empirical conditional expectation m_t^N, with covariance asymptotics showing a small-time blow-up and decorrelation along a k-dimensional data manifold (Theorem 2); (ii) a bias-variance decomposition for the mollified score (Theorem 3); and (iii) upper bounds on the KL divergence between the true noised distribution p_{t_N} and the generated distributions for the empirical and mollified scores (Theorem 4), from which the authors conclude that smoothing changes the memorization-to-generalization transition from N_c = t_N^{-k/2} to roughly N_c^β. The paper also introduces the LED-KDE interpretation of the mollified score as a two-stage smoothing procedure and presents numerical experiments on Swiss-roll, MNIST, and synthetic manifolds.","tokens_in":29561,"tokens_out":3980,"duration_ms":48585,"significance":"If the bounds in Theorem 4 were fully rigorous, the paper would provide one of the few analytic handles on the memorization-generalization trade-off in diffusion models, with the striking and falsifiable prediction that explicit score smoothing behaves like an enlarged training set. The derivation of the covariance structure as a re-weighted local PCA, the LED-KDE reformulation, and the spectral heuristic are all interesting and potentially useful starting points for future work. The paper is also commendably transparent about its own limitations, and the appendices contain substantial derivations with standard tools (CLT, delta method, Laplace's method). The central caveat is that the headline phase-transition claim rests on a small-time uniformity that is neither stated nor proved; until that gap is addressed, the results should be viewed as conditional.","major_comments":[{"comment":"The bound (13) is obtained by integrating the small-time variance asymptotic over [t_N, T], but the underlying CLT (Theorem 2) is stated for fixed t > 0 and the Laplace expansions are not accompanied by a uniformity statement in t. As t -> 0, the denominator φ^(0)(t,x) is of order (2π t)^{k/2} p_*(π(x)), so after division by t^2 the omitted o(N^{-1/2}) terms have coefficients that diverge like t^{-(k/2+1)}. The transition scale N_c = t_N^{-k/2} is exactly the regime where N φ^(0) is O(1), where the CLT has no valid asymptotic margin. The authors acknowledge this in the Limitations section, but the acknowledgment does not remove the need for either a uniform CLT or a non-asymptotic bound in the relevant regime. As stated, (12)-(13) are conditional on an unproved uniformity assumption, and the headline transition claim rests on that assumption.","section":"§5.3, Theorem 4 and Appendix A.7"},{"comment":"The 'transition' from memorization to generalization is inferred from where the right-hand sides of the upper bounds blow up, not from matching lower bounds. Even if the CLT uniformity issue were fixed, an upper bound alone does not establish that a sharp transition occurs at N_c or at N_c^β; it only shows that the bound is vacuous in the claimed regime. The text should either prove a lower bound (or a two-sided bound) or be reworded so that the phase-transition language is explicitly heuristic.","section":"§5.3, Theorem 4 (12)-(13)"},{"comment":"The proof of Theorem 4 uses E[||m_t^N - m_t||^2] ≈ (1/N) tr Σ_{(t,x),(t,x)}. Theorem 2 establishes convergence in finite-dimensional distribution, not convergence of second moments; the latter requires uniform integrability, which is not shown and is plausible to fail at small t where the variance diverges. This is another load-bearing gap in the derivation of (12)-(13) that should be addressed or explicitly assumed.","section":"§5.2, Eq. (9) and Appendix A.4"}],"minor_comments":[{"comment":"The text refers to 'Theorem 14' when discussing the covariance matrix; this should be 'Theorem 2'.","section":"§6, Discussion"},{"comment":"The notation N_c^β ≪ N_c is used without defining the exponent domain in the main text; it is clear from context that β ∈ (1/2,1), but this should be restated at the point of use.","section":"§5.3, after Eq. (13)"},{"comment":"The y-axis label 'Dataset ratio' is used in both figures; the text and captions define it as N_eff/N only in the body, but the figures themselves would benefit from the explicit definition in the axis label or caption.","section":"Figure 4 and Figure 10"},{"comment":"The sentence 'which proves the claim, up to a non-important log 1/t_N factor' is inconsistent with the statement of Theorem 4, which includes the logarithmic factor; either the theorem or the proof annotation should be aligned.","section":"Appendix A.7, proof of Theorem 4"},{"comment":"The notation f_k(x) := cos(π kx) is introduced for k ∈ N but then used as a multi-index k ∈ N^d; please clarify the multi-index convention at first use.","section":"Appendix A.10, spectral viewpoint"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its main gap, but the central claim in Theorem 4 is not yet supported by the proofs as written. I believe the gap is fixable within the manuscript's scope—for example, by adding explicit regularity and uniformity assumptions, or by reframing the transition claims as conditional on the CLT approximation—so I am recommending major revision rather than rejection. The numerical experiments are illustrative rather than exhaustive, and the theoretical contribution is the main reason to consider the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it. The rigorous core is Theorem 2, a CLT for the empirical score whose covariance asymptotically becomes a reweighted local PCA, and Theorem 3, the bias-variance decomposition for the mollified score. These are real contributions, proved with standard tools, and the uniform-in-x Laplace estimates are a nice touch. The flashier claim—that smoothing the score moves the memorization-to-generalization transition from N_c ~ t^{-k/2} to N_c^β—is much less secure. Theorem 4 obtains the KL bound by integrating a pointwise small-time approximation over [t_N, T], but the CLT is only proved at fixed t > 0, and the manuscript does not control the remainder as t -> 0. The transition scale N_c is exactly where the Gaussian approximation has no guaranteed margin, and the authors concede this in their Limitations section. I would treat the transition as a motivated conjecture backed by a non-rigorous bound, not as a theorem.\n\nWhat else is good: The LED-KDE reformulation is a genuinely useful way to see the two-stage smoothing, and the numerical experiments on Swiss roll, Gaussian, sphere, and MNIST are honest illustrations rather than hype. The proofs of Theorems 2 and 3 look internally consistent to me; the delta method and Laplace's method are applied with care, and the assumptions are stated clearly.\n\nSoft spots, in order of severity. First, Theorem 4's uniformity issue, just described: the stress-test concern about omitted o(N^{-1/2}) terms diverging like t^{-(k/2+1)} is real and not resolved. This is a load-bearing gap for the phase-transition claim, though not for the rest of the paper. Second, the 'transition' is read off from where the upper bounds blow up; there is no matching lower bound, so it is an upper-bound transition, not a proven sharp one. Third, the spectral heuristic is clearly labeled as heuristic, which is fine, but take it as such. Fourth, the code is promised but not yet available; minor and fixable.\n\nWho this is for: theorists and methodology people working on memorization and score estimation in diffusion models. It is worth a serious referee—the core theorems are original, the writing is careful, and the gap in Theorem 4 is either repairable with a uniform-in-t argument or can be demoted to a clear conjecture. My recommendation: send it to review, with a request that the authors address the uniformity issue or soften the transition claim to match what is proven.","headline":"Solid CLT and bias-variance core, but the headline memorization-transition claim rests on an unproved uniformity in t that the authors themselves flag.","tokens_in":30109,"tokens_out":5306,"would_cite":true,"duration_ms":51126,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","62G07","62G20","60J60"],"pacs":[],"model":"deepseek-v4-flash","headline":"A kernel-smoothed score provably reduces the dataset size at which a diffusion model switches from memorizing to generalizing, by convolving the empirical score with a Gaussian of bandwidth $h_N = t_N^\\beta$.","keywords":["denoising diffusion models","score matching","memorization","kernel smoothing","bias-variance tradeoff","Kullback-Leibler divergence","manifold data","kernel density estimation"],"falsifier":"Fix a known $k$-dimensional linear manifold with density $p_*$, draw $N$ samples, and measure the KL divergence between $G_{t_N} \\star p_*$ and the law generated by the mollified reverse SDE for $t_N \\to 0$ with $h_N = t_N^\\beta$, $\\beta \\in (1/2,1)$. Theorem 4 predicts the divergence scales as $h_N^2/t_N + \\log(1/t_N)/(N h_N^{k/2})$; if the data-dependent term instead scales like the unsmoothed bound $1/(N t_N^{k/2})$, or if the crossover sample size does not drop from $N_c = t_N^{-k/2}$ to about $N_c^\\beta$, the central claim fails. A cheaper check: compare the empirical covariance of $\\sqrt{N}(m_t^N(x)-m_t(x))$ to the predicted $t^{-k/2}P_{TM(x)}/p_*(\\pi(x))$ at small times.","tokens_in":29054,"feed_emoji":"🔄","tokens_out":13046,"duration_ms":126754,"temperature":0.7,"pith_summary":"Diffusion models can memorize their training set because the empirical score, the exact minimizer of the score-matching loss, forces the reverse process back to the nearest training point at small times. This paper models that failure as sampling noise in the score and shows, under a central limit theorem, that the noise covariance is asymptotically a re-weighted principal component analysis of the data: it blows up as $t^{-k/2}$ at small times and decorrelates in space. The proposed fix is to convolve the empirical score with a Gaussian kernel of bandwidth $h_N$. The paper proves an upper bound on the KL divergence between the true noisy distribution and the generated one, $O(h_N^2/t_N + \\log(1/t_N)/(N h_N^{k/2}))$, and shows that with $h_N = t_N^\\beta$, $\\beta\\in(1/2,1)$, smoothing shifts the memorization-to-generalization transition from $N_c = t_N^{-k/2}$ to roughly $N_c^\\beta$, acting like an enlarged dataset. If true, the source of generalization in a diffusion model lies partly in the smoothing operation itself, not in the learned parametrization.","feed_headline":"Score smoothing acts like a larger diffusion dataset","feed_subtitle":"Smoothing the score moves the memorization threshold from N_c to about N_c^β, so fewer samples are needed.","key_machinery":"The central object is the mollified empirical score $\\tilde s_t^N = G_h \\star s_t^N$, the convolution of the empirical score with a Gaussian kernel of bandwidth $h$; because the Gaussian kernel is centered, this is also the score of the LED-KDE density $(G_h, G_t)\\star p_0^N$. The argument runs on three pieces: a central limit theorem for the empirical score whose noise covariance is a re-weighted data PCA aligned with the tangent space; a bias-variance decomposition of the mollified score, with the variance term controlled by a Gaussian-process computation using Laplace's method and the bias controlled by the smoothing lengthscale; and a KL-divergence bound that converts score estimation error into distribution error. The bandwidth does the work: it suppresses the high-frequency components of the score without adding bias, effectively replacing a dataset of size $N$ by one of effective size $N_{\\mathrm{eff}}\\approx N(h/t)^{k/2}$.","core_discovery":"Under the assumption that data lies on a smooth $k$-dimensional manifold (linear for the main theorems), the empirical score seen through its Gaussian approximation has covariance $\\Sigma_{(t,x),(t,x)} \\sim t^{-k/2} p_*(\\pi(x))^{-1} P_{TM(x)}$, so the sampling noise lives in tangent directions and diverges at small times. Convolving with a Gaussian kernel of bandwidth $h_N$ turns the score into the score of the Log-Exponential Double-Kernel Density Estimator (LED-KDE), $(G_h, G_t) \\star p_0^N$. The bias-variance decomposition gives variance of order $t_N/(N h_N^{k/2})$ and bias of order $h_N^2$, and the KL bound $\\mathbb{E}_D D_{KL}(p_{t_N}\\|\\tilde q^N_{t_N}) = O(h_N^2/t_N + \\log(1/t_N)/(N h_N^{k/2})) + D_{KL}(p_T\\|N(0,T I_d))$ follows. With $h_N = t_N^\\beta$, $\\beta \\in (1/2,1)$, the critical dataset size for the memorization-to-generalization transition drops from $N_c = t_N^{-k/2}$ to $\\tilde N_c = N_c^\\beta$. The paper also gives a spectral heuristic suggesting this bound can be improved with an adaptive lengthscale, and interprets the mollified reverse diffusion as a Wasserstein gradient flow attracted to a smoothed (LED-KDE) measure.","pith_inferences":["Editorial inference: the same smoothing argument should transfer to pretrained score networks, not just the empirical score: convolving a network's output with a Gaussian kernel at generation time is a cheap post-hoc memorization mitigation, which the paper sketches but does not test.","Editorial inference: the appendix's observation that mollification approximates evaluating the score at a larger time suggests that common sampling heuristics such as early stopping or coarser time steps already realize part of the smoothing benefit; a direct comparison between explicit kernel smoothing and time-step coarsening would test this.","Editorial inference: because the score-noise covariance aligns with the data tangent space, an anisotropic or manifold-adapted kernel should control variance better than the isotropic Gaussian used in the theorems; a natural experiment is to replace $G_h$ by a kernel whose covariance follows the local empirical PCA and compare realized KL divergences.","Editorial inference: the LED-KDE factorization implies the regularization effectively smooths along the intrinsic manifold, so the memorization threshold should depend on intrinsic dimension $k$ rather than ambient dimension $d$; this can be checked by embedding the same low-dimensional data in higher ambient spaces."],"forward_implications":["Convolving the score with a Gaussian kernel of appropriate bandwidth reduces the critical dataset size for the memorization-to-generalization transition from $N_c = t_N^{-k/2}$ to roughly $N_c^\\beta$, so the same model generalizes from far fewer samples.","Because smoothing suppresses the small-time divergence of the score noise, the reverse process can start from a smaller time $t_N$, limiting the initial Gaussian leakage of mass away from the data manifold.","The paper's bounds suggest an effective dataset size $N_{\\mathrm{eff}} \\approx N(h/t)^{k/2}$, giving a quantitative target: at very small times, a regularized score built from $N$ points behaves like an unregularized score from a much larger sample.","The mollified reverse SDE is the Wasserstein gradient flow of the KL divergence toward a smoothed (LED-KDE) measure, so even a fixed smoothing, with no learned model, acts as a generalization mechanism.","The paper's spectral heuristic indicates that frequency truncation by the kernel could improve the KL bound to $O(t_N/(N h_N^{1+d/2}))$, suggesting that adaptive, time-dependent bandwidths can outperform the isotropic kernel analyzed in the theorems."],"supporting_citations":[{"why":"Supplies the KL-divergence upper bound that converts score estimation error into distribution error, the backbone of Theorem 4.","marker":"[36]"},{"why":"Defines the empirical score as the exact minimizer of the score-matching loss, the memorization-prone estimator the paper regularizes.","marker":"[42]"},{"why":"Establishes the collapse and memorization regime of optimally trained diffusion models whose transition scale the paper analyzes.","marker":"[6]"},{"why":"Provides the Laplace method used to derive the small-time asymptotics of the score noise covariance and of the bias.","marker":"[5]"},{"why":"Connects the covariance eigenstructure to the geometry-adaptive view of the score's Jacobian, grounding the local PCA interpretation.","marker":"[22]"}],"fun_headline_variants":["Score smoothing reduces diffusion memorization","Smoothing scores acts like having more data","Kernel-smoothed score cuts diffusion variance","Less data needed with smoothed diffusion scores","Diffusion score smoothing mimics larger datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole analysis relies on the sampling noise in the score being Gaussian even at the tiny times where memorization happens, but the paper only proves that Gaussian behavior for fixed positive times.","fun_headline_variants_meta":{"raw":{"variants":["Score smoothing reduces diffusion memorization","Smoothing scores acts like having more data","Kernel-smoothed score cuts diffusion variance","Less data needed with smoothed diffusion scores","Diffusion score smoothing mimics larger datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1916,"prompt_tokens":1168,"completion_tokens":748,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":784,"completion_tokens_details":{"reasoning_tokens":684}},"tokens_in":784,"tokens_out":748,"duration_ms":7896,"temperature":1.0,"reasoning_tokens":684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:59:39.046241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a known $k$-dimensional linear manifold with density $p_*$, draw $N$ samples, and measure the KL divergence between $G_{t_N} \\star p_*$ and the law generated by the mollified reverse SDE for $t_N \\to 0$ with $h_N = t_N^\\beta$, $\\beta \\in (1/2,1)$. Theorem 4 predicts the divergence scales as $h_N^2/t_N + \\log(1/t_N)/(N h_N^{k/2})$; if the data-dependent term instead scales like the unsmoothed bound $1/(N t_N^{k/2})$, or if the crossover sample size does not drop from $N_c = t_N^{-k/2}$ to about $N_c^\\beta$, the central claim fails. A cheaper check: compare the empirical covariance of $\\sqrt{N}(m_t^N(x)-m_t(x))$ to the predicted $t^{-k/2}P_{TM(x)}/p_*(\\pi(x))$ at small times.","supporting_citations":[{"cited_title":"Dynamical regimes of diffusion models","cited_arxiv_id":null,"evidence_quote":"Establishes the collapse and memorization regime of optimally trained diffusion models whose transition scale the paper analyzes."},{"cited_title":"Advanced mathematical methods for scientists and engineers I: Asymptotic methods and perturbation theory","cited_arxiv_id":null,"evidence_quote":"Provides the Laplace method used to derive the small-time asymptotics of the score noise covariance and of the bias."},{"cited_title":"Generalization in diffusion models arises from geometry-adaptive harmonic representation","cited_arxiv_id":null,"evidence_quote":"Connects the covariance eigenstructure to the geometry-adaptive view of the score's Jacobian, grounding the local PCA interpretation."}],"review_version":1}