Pith. sign in

REVIEW 3 major objections 4 minor 25 references

Generalization bounds for score-based generative models: a synthetic proof

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Score-based generative models are minimax optimal in 1-Wasserstein distance over Hölder densities.

desk verdict A genuinely useful synthetic proof of W1 minimax rates for SGMs, but a covering-number mismatch in the closing argument and a misapplied stability lemma leave the main theorem unsupported as written. read the letter →

arxiv 2507.04794 v1 pith:PYT7RCDG submitted 2025-07-07 math.ST stat.TH

Arthur Stéphanovitch , Eddie Aamari , Clément Levrard This is my paper
classification math.STstat.TH MSC 62G0762G2060H1068T07
keywords score-basedgenerativemodelsdiffusionminimaxratesWasserstein-1distancedenoisingscorematchingneuralnetworksHölderdensitiesnonparametricdensityestimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper establishes that score-based generative models, the diffusion-style samplers behind much of modern generative AI, achieve the minimax-optimal rate for density estimation in the 1-Wasserstein distance. For any target density $p^\star$ lying in the $\beta$-Hölder model class $A^\beta_K$ with compact support or subGaussian tails, the distribution produced by an early-stopped backward diffusion driven by a tanh-neural-network score estimator is within $\mathrm{polylog}(n)\, n^{-(\beta+1)/(2\beta+d)}$ of the truth, and no estimator can do better. The argument works for every smoothness $\beta>0$ and for both stochastic and deterministic samplers. A sympathetic reader should care because the proof shows that the score function needs no bespoke smoothness assumption: shape constraints on the data density already induce the score regularity that makes score matching provably effective.

What carries the argument

The engine of the proof is the list of score-regularity properties called Properties $(P_\beta)$. The score $s^\star$ satisfies a linear-growth bound (A), the one-sided Lipschitz bound (B) (largest eigenvalue of $\nabla s^\star + \sigma^{-2} I$ bounded by $C e^{-2t}(1+t^{-(1-\beta\wedge 1)/2})$), and space-time Hölder bounds (C)-(D) on convex high-probability sets $A^\varepsilon_t$ that expand monotonically as $\varepsilon$ decreases. Lemma 9 asserts that every density in $A^\beta_K$ produces a score with exactly these properties. The one-sided Lipschitzness makes the exponential weight in the drift-stability theorem (Theorem 11) integrable, so $W_1$ error is controlled by a weighted $L^1$ score error; the high-order Hölder bounds let standard tanh-network approximation results (Propositions 15 and 16) fit the score on $A^\varepsilon_t$ with accuracy $\tau_k^{-1} \mathrm{polylog}(n)\, n^{-(\beta+1)/(2\beta+d)}$ and explicit network sizes. The whole analysis threads through these properties, so the paper is synthetic in the sense that any score satisfying $(P_\beta)$ automatically leads to the minimax rate.

What would settle it

Take a concrete density in $A^\beta_K$, for instance a smoothed uniform density on a cube with $\beta=1$, and numerically check the score's high-probability Hölder norm in bound (D) at small times: if $\|\partial_t^k(s^\star+x/\sigma^2)\|_{H^\gamma(A^\varepsilon_t)}$ exceeds $C_{k,\gamma}\log(\varepsilon^{-1})^{C_2(1+k+\gamma)}e^{-t}(1+t^{-((1/2+k+\gamma-\beta/2)\vee 0)})$ by more than a constant, the imported regularity fails and the minimax upper bound has no basis. Alternatively, check on the same example whether $\sup_x \lambda_{\max}(\nabla s^\star(t,x)+\sigma^{-2}I)$ grows faster than $C e^{-2t}(1+t^{-1/2})$ as $t\to 0$.

Watch

Extended reading notes

Core claim

The central claim is that score-based generative models are minimax optimal over the Hölder model $A^\beta_K$ in Wasserstein-1 distance. Specifically, Theorem I shows that with an empirical risk minimizer of the denoising score-matching loss over a carefully chosen class of one-sided Lipschitz tanh networks, the early-stopped backward Ornstein-Uhlenbeck process emits samples whose law $\hat{p}$ satisfies $\sup_{p^\star \in A^\beta_K} \mathbb{E}[W_1(p^\star, \hat{p})] \lesssim \mathrm{polylog}(n)\, n^{-(\beta+1)/(2\beta+d)}$; Theorem II proves a matching lower bound over the same model. The mechanism is that the true score $s^\star(t,x)$ is, at every positive time, one-sided Lipschitz and Hölder on large-probability sets, with explicit time decays. These properties feed into a two-step analysis, stability of the backward SDE and an oracle inequality for score matching, that yields the rate by a clean bias-variance trade-off.

Load-bearing premise

The upper bound assumes that every density in the model class $A^\beta_K$ yields a score function satisfying the imported space-time Hölder bounds and set geometry of Properties $(P_\beta)$; that regularity transfer is cited from companion work rather than proved here, and Theorem I collapses if it fails.

Editorial extensions

If this is right

  • The rate $n^{-(\beta+1)/(2\beta+d)}$ in 1-Wasserstein distance is optimal up to polylog factors over $A^\beta_K$, matching the established minimax rate for nonparametric density estimation in transportation distances.
  • The result covers all $\beta>0$ and both stochastic ($b_t>0$) and deterministic ($b_t=0$) backward samplers; the regime $\beta>2$ with the probability-flow ODE was not previously covered in combination.
  • Approximating the Jacobian of the score is not needed in Wasserstein-1 distance; a one-sided Lipschitz constraint on the network suffices for stability.
  • The score estimator can use a time-dependent family of neural networks with $O(n^{2/(2\beta+d)})$ networks to achieve the tight rate, while restricting to $O(\log n)$ networks still gives near-optimal rates.
  • The proof separates diffusion stability from score estimation, so other forward processes and samplers can be plugged in whenever a denoising score trick holds.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the same Properties $(P_\beta)$ hold for other forward processes, such as variance-exploding diffusions or manifold-supported targets as suggested by the paper's conjecture, the identical two-step argument would transfer the minimax rate to those settings, making the score-regularity list the true coreset of the problem.
  • The one-sided Lipschitz constraint on the approximating networks is load-bearing for the stability estimate; a practical implication is that enforcing it, for example via spectral normalization, may matter for finite-sample guarantees even though $W_1$ does not require full Jacobian control.
  • Because only high-probability sets $A^\varepsilon_t$ enter the approximation, the result predicts that score-matching error in low-density tails is essentially irrelevant for Wasserstein-1 generation quality; this could be tested by corrupting the estimated score far from the data and measuring the $W_1$ degradation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper claims minimax optimality for score-based generative models in 1-Wasserstein distance over a Hölder class A^β_K of densities satisfying convexity/shape conditions. The proposed proof is 'synthetic': it combines imported score-regularity estimates (Stéphanovitch 2024, 2025), diffusion stability bounds, an oracle inequality for denoising score matching, and tanh neural-network approximation to obtain an upper bound n^{-(β+1)/(2β+d)} up to polylog factors, together with a matching lower bound over the same model. The main theorem applies to both stochastic and deterministic backward samplers and to all β>0. The paper is clearly written and the high-level architecture is coherent, but several load-bearing steps in the proof of Lemma 13 and Proposition 16 are not supported by the stated assumptions and estimates.

Significance. If the proof can be repaired, the result is significant: it would establish minimax Wasserstein-1 rates for score-based generative models with neural-network score estimators trained by empirical denoising score matching, covering arbitrary Hölder smoothness and unifying SDE and ODE samplers. The explicit minimax lower bound in Theorem II and the emphasis on W1 rather than TV are valuable, and the paper gives a useful modular decomposition of the argument. However, the paper currently relies on unproved companion-preprint estimates for the central score-regularity properties, and two local but essential technical gaps (the covering-number statement in Proposition 16 and the application of stability lemmas to an unbounded drift class) must be fixed before the main claims can be accepted.

major comments (3)
  1. [Section 4.2, Proposition 16] The stated covering-number bound log N(Ψ(L_k,W_k,B_k,V_k,V'_k), ||·||∞, n^{-1}) ≤ C log(n)^{C2} τ_k n^{(d-2)/(2β+d)} is inconsistent with the width W_k = C log(n)^{C2} τ_k^{-1} n^{(d-2)/(2β+d)} in (14). Proposition 15, applied with ε^{-d} = W_k, gives a covering number of order W_k log n = τ_k^{-1} n^{(d-2)/(2β+d)} log n, not τ_k n^{(d-2)/(2β+d)}. For the first interval τ_0 = (Cn)^{-2(β+1)/(2β+d)} and d<2β+4, the stated bound is o(1), which is impossible for a function class with Θ(n) parameters. The proof of Lemma 13 in Section 4.3 uses the τ_k^{-1} version (the variance terms contain (1/τ_k) log N and then 1/τ_k^2), so the displayed Proposition 16 does not support the closing argument. The proposition must be corrected to the τ_k^{-1} form, and the proof of Lemma 13 updated consistently.
  2. [Section 4.3, initial-condition step] Lemma 10 is invoked to bound W1(bp_{T-T}, eµ_{T-T}), but the drift bat(x) = x + (σ^2 + b_t^2)bs(T-t,x) is unbounded on R^d because the score class S in (13) contains the linear term -x/σ^2. This violates Lemma 10's assumption that the drift is bounded in sup norm. In addition, the proof defines eµ by d eXt = -bat(eXt)dt, so eµ does not have the same drift as bp; as written, Lemma 10 cannot compare the two processes. The claim shortly afterward that ∥bs∥∞ ≤ C log(n)^{C2} n^{(β+1)/(2β+d)} is false for the class S defined in (13). This step needs a corrected argument, for instance an explicit coupling of the two initial distributions through the OU semigroup, or a reformulation in terms of the bounded centered score φ = s + x/σ^2.
  3. [Sections 3.2 and 4.1, oracle inequality and class boundedness] Proposition 12 assumes the candidate class S is uniformly bounded by V, but the class defined in (13) is not bounded on R^d due to the -x/σ^2 term. Consequently Lemma 19's sup-norm control of the excess contrast ν(s,x) on the high-probability set does not apply to the estimator (9) over this class. The paper should either state and prove the oracle inequality for the centered class {φ = s + x/σ^2 : s ∈ S} with φ bounded, or show explicitly that the unbounded linear term cancels in the loss differences and in the clipping argument. As written, the main bound in Lemma 13 is not a formal consequence of Proposition 12.
minor comments (4)
  1. [Section 4.3, definitions of eµ and X^{k,j}] The displayed equations for eµ and X^{k,j} contain minus signs before the drift terms; if these are intended, they are inconsistent with the later application of Lemma 10 and Theorem 11, which are stated for drifts without a sign flip. These are likely typos and should be corrected to +bat and +ba^{k,j}, respectively.
  2. [Appendix B, proof of Theorem II] The sentence '∥fε∥H1 ≤ 1, and therefore fε is 1-Lipschitz' is not correct: H1 control does not imply Lipschitz regularity in dimension d≥2. The lower bound can be repaired by normalizing fε by ∥∇fε∥∞, which is O(1) uniformly in j, so the rate is unchanged; the claim as written should be fixed.
  3. [Section 2.3 and Lemma 9] The quantitative score-regularity estimates (C) and (D) are imported from Stéphanovitch (2025, Corollary 3) and Stéphanovitch (2024) and are not proved in this paper; Lemma 9 only derives the geometric parts (Pβ).1-2 from those estimates. Since these estimates are the engine of Proposition 16 and Lemma 13, the paper should state clearly that Theorem I is conditional on the companion preprints or include proofs of those estimates.
  4. [Equation (13) and surrounding text] The notation '-1/σ^2 Id×d' in the definition of S is not meaningful as written; it should be '-x/σ^2' or an explicit affine term, and the sup-norm bounds V_k should be defined only for the centered network part to avoid the false implication that bs is uniformly bounded.

Circularity Check

1 steps flagged · score 4.0 of 10

The main upper bound rests on score-regularity Properties (Pβ) imported from the first author's companion preprints (Stéphanovitch 2024, 2025) rather than proved here; this is load-bearing self-citation, though the rate argument itself is not fitted and the lower bound is external.

  1. self citation load bearing [Section 2.4, Lemma 9 and Properties (Pβ); also Propositions 7 and Theorem 8 in Section 2.3]
    "Most importantly, we will show in Lemma 13 that for the Ornstein-Uhlenbeck forward process, if the score function satisfies Properties(Pβ), then score-based generative models achieve minimax convergence rates for density estimation. This fact highlights the fundamental role of these regularity properties in guaranteeing the optimal statistical performance of SGMs."

    Lemma 13 takes Properties (Pβ) as the sufficient condition for the minimax rate, but Theorem I is stated unconditionally over Aβ_K. The only bridge from Aβ_K to Pβ is Lemma 9, whose proof in Section A.3 imports Proposition 7, Theorem 8, and the explicit sets Aε_t from Stéphanovitch (2025, Theorem 1, Corollary 3, Section 3.1.1) and from Stéphanovitch (2024). These are same-author companion preprints, not re-derived or independently verified in the present paper. Since the space-time Hölder bounds (C)-(D) and the convex-set geometry (Pβ.1)-(Pβ.2) are the engine of the approximation result and the oracle inequality, the main claim is conditional on unverified self-citations.

full rationale

No definitional circularity was found: the estimator, the time grid T = (Cn)^{-2(β+1)/(2β+d)}, the widths W_k, and the covering-number and approximation bounds are chosen by the proof, not fitted to the target rate, and the minimax lower bound of Theorem II is derived from independent arguments (Uppal et al., 2019; Donoho et al., 1996). The core circularity concern is instead self-citation burden. The main upper bound, Theorem I via Lemma 13, is conditioned on Properties (Pβ), and the lemma asserting that Aβ_K densities satisfy Pβ (Lemma 9) is proved only by citing the first author's companion preprints Stéphanovitch (2025, Theorem 1 and Corollary 3) and Stéphanovitch (2024) for the score estimates and for the construction of Aε_t. Those results are not machine-checked, code-reproduced, or independently established in this manuscript, so they constitute load-bearing self-citation rather than independent support. Separately, the reviewer's noted discrepancy between the covering-number bound stated in Proposition 16 (a factor τ_k n^{(d-2)/(2β+d)}) and the reciprocal form used in the closing lines of Lemma 13 is a correctness concern, not a circularity, and does not affect this circularity score. Overall, the derivation has substantial independent content, but its central hypothesis is imported from the same author's earlier work, justifying a score of 4.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

Pure theory paper; no data, no fitted constants, no new entities. The central external input is the score regularity theory from the first author's companion papers, which serves as an unproved-in-this-paper assumption. The proof also relies on standard stochastic calculus, concentration, and neural-network approximation results.

assumptions (4)
  • domain assumption Assumption 1: p* = exp(-u+a) with a in H^β_K, u strongly convex and derivative growth controlled
    Defines the statistical model A^β_K; all rates are over this class. Section 2.2.
  • domain assumption Score regularity Properties (Pβ), including bounds (A)-(D) and set geometry, hold for p* in A^β_K (Lemma 9 via Stéphanovitch 2025)
    Imported from the author's companion papers; load-bearing for Lemmas 12, 13 and 16. Sections 2.3-2.4.
  • domain assumption Empirical risk minimizer (9) is computed exactly and SDE integrals are exact (Remark 4)
    Idealized oracle framework; the paper explicitly sets aside optimization and discretization errors.
  • standard math Standard nonparametric statistics and approximation theory results (covering numbers, Bernstein, Fano, Varshamov-Gilbert, De Ryck et al. tanh approximation)
    Used throughout Section 4 and the appendices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalization bounds for score-based generative models: a synthetic proof." pith.science (2026). https://pith.science/paper/PYT7RCDG

@misc{pith2026250704794,
  author       = {Pith},
  title        = {Pith review of: Generalization bounds for score-based generative models: a synthetic proof},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PYT7RCDG}},
  note         = {Machine review of arXiv:2507.04794}
}
abstract

We establish minimax convergence rates for score-based generative models (SGMs) under the $1$-Wasserstein distance. Assuming the target density $p^\star$ lies in a nonparametric $\beta$-smooth H\"older class with either compact support or subGaussian tails on $\mathbb{R}^d$, we prove that neural network-based score estimators trained via denoising score matching yield generative models achieving rate $n^{-(\beta+1)/(2\beta+d)}$ up to polylogarithmic factors. Our unified analysis handles arbitrary smoothness $\beta > 0$, supports both deterministic and stochastic samplers, and leverages shape constraints on $p^\star$ to induce regularity of the score. The resulting proofs are more concise, and grounded in generic stability of diffusions and standard approximation theory.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 10 canonical work pages

  1. [1]

    From the mean value theorem, we have ∥s⋆(t, x)∥ ≤ ∥∇s⋆(t, ·)∥∞∥x − x⋆∥ + ∥s⋆(t, x⋆)∥

    23 A.2 Proof of Proposition 6 Proof of Proposition 6.Let us writex⋆ := arg minx u(x). From the mean value theorem, we have ∥s⋆(t, x)∥ ≤ ∥∇s⋆(t, ·)∥∞∥x − x⋆∥ + ∥s⋆(t, x⋆)∥. (20) Now, writingφt,x fortheprobabilitydensityof N (e−tx, (1−e−2t)σ2Id×d), andtaking r(y) := p⋆(y)/φ∞,0(y), define the probability density pt,x(y) := r(y)φt,x(y)R r(z)φt,x(z)dz . From (...

    work page 2025

  2. [2]

    of Properties(Pβ). As given in (Stéphanovitch, 2025, Section 3.1.1), we have Aε t := ( Aε 0 = {y | u(y) ≤ C(log(ε−1) + K + 1)K} if t ≤ C−1 ⋆ log(ε−1)−C⋆ Aε ∞ = {y | ∥y − y⋆∥ ≤C log(ε−1)(1 + K)} if t > C−1 ⋆ log(ε−1)−C⋆, corresponding to the set given by Theorem

    work page 2025

  3. [4]

    doi: 10.1093/acprof:oso/9780199535255.001.0001

    ISBN 978-0-19-953525-5. doi: 10.1093/acprof:oso/9780199535255.001.0001. URL https: //doi.org/10.1093/acprof:oso/9780199535255.001.0001. A nonasymptotic theory of independence, With a foreword by Michel Ledoux. Changxiao Cai and Gen Li. Minimax optimality of the probability flow ode for diffusion models.arXiv preprint arXiv:2503.09583,

    arXiv

  4. [8]

    Beyond log-concavity and score regularity: Improved convergence bounds for score-based generative models in w2-distance.arXiv preprint arXiv:2501.02298,

    Marta Gentiloni-Silveri and Antonio Ocello. Beyond log-concavity and score regularity: Improved convergence bounds for score-based generative models in w2-distance.arXiv preprint arXiv:2501.02298,

    arXiv

  5. [10]

    Diffwave: A versatile diffusion model for audio synthesis

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761,

    arXiv 2009

  6. [12]

    Nonparametric estimation of a factorizable density using diffusion models.arXiv preprint arXiv:2501.01783,

    Hyeok Kyu Kwon, Dongha Kim, Ilsang Ohn, and Minwoo Chae. Nonparametric estimation of a factorizable density using diffusion models.arXiv preprint arXiv:2501.01783,

    work page arXiv

  7. [14]

    Score- based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score- based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    arXiv 2011

  8. [15]

    Smooth transport map via diffusion process.arXiv preprint arXiv:2411.10235,

    Arthur Stéphanovitch. Smooth transport map via diffusion process.arXiv preprint arXiv:2411.10235,

    arXiv

Show all 25 references
  1. [16]

    Regularity of the score function in generative models.arXiv preprint arXiv:2506.19559,

    Arthur Stéphanovitch. Regularity of the score function in generative models.arXiv preprint arXiv:2506.19559,

    arXiv

  2. [17]

    Generalization error bound for denoising score matching under relaxed manifold assumption

    Konstantin Yakovlev and Nikita Puchkin. Generalization error bound for denoising score matching under relaxed manifold assumption. arXiv preprint arXiv:2502.13662,

  3. [18]

    Then it holds EX∼p⋆ bf (X) = (EX∼p⋆ − En) bf (X) + En bf (X) ≤ (EX∼p⋆ − En) bf (X) + En ¯f (X) = EX∼p⋆ ¯f (X) + (EX∼p⋆ − En) bf (X) + (En − EX∼p⋆) ¯f (X)

    Proof of Proposition 18.Let bf ∈ arg minf ∈F En(f ) and ¯f ∈ arg minf ∈F EX∼p⋆f (X). Then it holds EX∼p⋆ bf (X) = (EX∼p⋆ − En) bf (X) + En bf (X) ≤ (EX∼p⋆ − En) bf (X) + En ¯f (X) = EX∼p⋆ ¯f (X) + (EX∼p⋆ − En) bf (X) + (En − EX∼p⋆) ¯f (X). (25) We letC∞(F , ε) ⊂ Fbe a minimalε...

    2013

  4. [19]

    For allx, t, z, it holds ∥s(t, etx + σtz) − z/σt∥2 − ∥s⋆(t, e−tx + σtz) − z/σt∥2 ≤ ∥(s − s⋆)((t, e−tx + σtz)∥ ∥s(t, e−tx + σtz) − z/σt∥ + ∥s⋆(t, e−tx + σtz) − z/σt∥ . Using Cauchy-Schwaz inequality entails that |ν(s, x)| = EZ Z τ τ ∥s(t, e−tx + σtZ) − Z/σt∥2dt − EZ Z τ τ ∥s⋆(t...

    2021

  5. [20]

    First of all, the caset > C−1 ⋆ log(ε−1)−C⋆ is immediate

    Now by Assumption 1.2 we have thatPX∼p⋆(Aε ∞) ≥ 1 − ε, so we only need to check Property(Pβ).2. First of all, the caset > C−1 ⋆ log(ε−1)−C⋆ is immediate. Now, in the caset ≤ C−1 ⋆ log(ε−1)−C⋆, for x ∈ Aε t and y ∈ Rd such that ∥x − y∥ ≤(KC (log(ε−1) + K + 1)K)−K, we have from ...

    2020

  6. [23]

    Proof of Theorem II.As a base density, let us choose the standard gaussian g0(x) := (2π)−d/2 exp(−∥x∥2/2)

    and (Uppal et al., 2019, Theorem 4), using standard bayesian-like lower bound arguments comparing perturbed distributions. Proof of Theorem II.As a base density, let us choose the standard gaussian g0(x) := (2π)−d/2 exp(−∥x∥2/2). To perturb it, letψ : Rd → R denote a kernel-li...

    2019

  7. [24]

    Its integral is trivially equal to one

    Given ε ∈ {−1, 1}m, consider the function gε := g0 + c1 X k∈{0,...,2j −1}d εkψk, for some constantc1 > 0 to be chosen later. Its integral is trivially equal to one. Furthermore, the function gε g0 − 1 = c1 g0 X k∈{0,...,2j −1}d εkψk, has support included in[−1, 1]d, and g0 ≥ c...

    2008

  8. [104]

    provides the existence of a subset Ω(j) ∗ ⊂ Ω(j) such that • M∗ := |Ω(j) ∗ | ≥2m/8; • for all distinctε, ε′ ∈ Ω(j) ∗ , ω(ε, ε′) ≥ m/8. Next, according to (Uppal et al., 2019, Lemma 24 - Supplementary Material), it holds KL(g⊗n ε | g⊗n 0 ) = nKL(gε | g0) ≤ cn∥g0 − gε∥2 L2/2 ≤ n...

    2019

  9. [1982]

    Convergenceofdiffusionmodelsunderthemanifold hypothesis in high-dimensions.arXiv preprint arXiv:2409.18804,

    IskanderAzangulov, GeorgeDeligiannidis, andJudithRousseau. Convergenceofdiffusionmodelsunderthemanifold hypothesis in high-dimensions.arXiv preprint arXiv:2409.18804,

    arXiv

  10. [1996]

    doi: 10.1214/aos/1032894451

    ISSN 0090-5364. doi: 10.1214/aos/1032894451. Zehao Dou, Subhodh Kotekal, Zhehao Xu, and Harrison H Zhou. From optimal score matching to optimal sampling. arXiv preprint arXiv:2409.07032,

  11. [2013]

    Error bounds for flow matching methods.arXiv preprint arXiv:2305.16860,

    Joe Benton, George Deligiannidis, and Arnaud Doucet. Error bounds for flow matching methods.arXiv preprint arXiv:2305.16860,

    arXiv

  12. [2019]

    Flow matching achieves almost minimax optimal convergence.arXiv preprint arXiv:2405.20879,

    Kenji Fukumizu, Taiji Suzuki, Noboru Isobe, Kazusato Oko, and Masanori Koyama. Flow matching achieves almost minimax optimal convergence.arXiv preprint arXiv:2405.20879,

    arXiv

  13. [2020]

    On the minimax optimality of flow matching through the connection to kernel density estimation

    Lea Kunkel and Mathias Trabs. On the minimax optimality of flow matching through the connection to kernel density estimation. arXiv preprint arXiv:2504.13336,

  14. [2022]

    Convergence analysis of probability flow ode for score-based generative models.arXiv preprint arXiv:2404.09730,

    Daniel Zhengyu Huang, Jiaoyang Huang, and Zhengjiang Lin. Convergence analysis of probability flow ode for score-based generative models.arXiv preprint arXiv:2404.09730,

    arXiv

  15. [2023]

    A sharp convergence theory for the probability flow odes of diffusion models

    Gen Li, Yuting Wei, Yuejie Chi, and Yuxin Chen. A sharp convergence theory for the probability flow odes of diffusion models. arXiv preprint arXiv:2408.02320,

    arXiv

  16. [2024]

    Adapted wasserstein distance between the laws of sdes.arXiv preprint arXiv:2209.03243,

    Julio Backhoff-Veraguas, Sigrid Källblad, and Benjamin A Robinson. Adapted wasserstein distance between the laws of sdes.arXiv preprint arXiv:2209.03243,

  17. [2025]

    Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions

    Hongrui Chen, Holden Lee, and Jianfeng Lu. Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions. InProceedings of the 40th International Conference on Machine Learning, volume 202, pages 4735–4763. PMLR, 23–29 Jul 2023...

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.