Universal Local Error and Realized Amplification for the First-Order EDM Predictor

Nicolas Brosse; Arnak S. Dalalyan

stat.MLOct 7, 2026

Universal Local Error and Realized Amplification for the First-Order EDM Predictor

Authors: Nicolas Brosse, Arnak S. Dalalyan

Organizations: CREST, ENSAE Paris, Institut Polytechnique de Paris

Abstract

We analyze the first-order deterministic diffusion sampler of Karras et al. (2022), termed EDM, in 2-Wasserstein distance by separating two sources of error: local discretization error and its amplification by subsequent learned steps. We prove that local error admits a universal bound: for any data distribution with finite second moment, the one-step discretization error is quadratic in the step size, with an explicit constant that does not depend on the data distribution. Error propagation, in contrast, depends on the learned network. At high noise levels, we exploit the network parametrization of EDM to derive an explicit contraction criterion. At low noise levels, we measure propagation through the amplification realized on the distributions transported by the sampler; this realized amplification can be arbitrarily smaller than the worst-case Lipschitz constant. This analysis yields an $O(e^{Λ_K}/K)$ global discretization error for $K$ sampling steps, where $Λ_K$ is the low-noise log-amplification. Experiments on a one-dimensional Gaussian mixture show how measured amplification accounts for slower error decay on finite sampling grids. Diagnostics on a pretrained CIFAR-10 model illustrate related stability mechanisms without certifying the global assumptions.

Figures & tables

Figure 1: Local accuracy and accumulated error. For a four-Gaussian mixture with oracle denoising and exact initialization, summed local biases approach first-order decay, while final error decreases at orders $0.58$ – $0.60$ over the tested range. Propagating each bias through the measured amplification of the later steps (solid blue) recovers the slower decay of the error; see Section 7.1 .

Panel of Figure 2 — Figure 2: One-step Wasserstein error decomposition. Solid arrows correspond to pushforwards, whereas dashed links are $W_{2}$ distances.

Panel of Figure 3 — Figure 3: A member $G_{K}$ of the refinement family with $\rho=1/2$ and $K=10$ .

Figure 4: 1D mixture (a–c) and pretrained CIFAR–10 EDM model (d). (a) Oracle with exact initialization: measured $\Lambda_{K}$ and its fit $\omega\log K$ . (b) Local orders ( 38 ) of the final error $e_{K}$ and of the summed local biases, against the amplification-based order $1-(\Lambda_{K}-\Lambda_{K/2})/\log 2$ computed from (a): the biases converge at first order, the error does not. (c) Directional majorant (dashed) and finite-domain field proxy (dotted) divided by $\Lambda_{K}$ , from the canonical Gaussian initialization; trained networks as the median over three seeds. (d) Energy-weighted rates ( 41 ), maximized over nine nodes $t$ , on the coarse–fine segments of $G_{68}$ : worst direction $\lambda^{\max}$ , displacement $\lambda^{\parallel}$ and random direction $\lambda^{\mathrm{rand}}$ , in absolute value; open markers are negative rates, and shading marks where the 90th-percentile empirical margin ( 40 ) is positive.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: The split index $K_{\mathrm{hi}}$ of eq. 25 is the last level at or above the threshold $\sigma_{\mathrm{hi}}$ . The step $\sigma_{K_{\mathrm{hi}}}\to\sigma_{K_{\mathrm{hi}}+1}$ that crosses the threshold (shaded) is the first step of the low-noise block, so every high-noise step ends at or above $\sigma_{\mathrm{hi}}$ .

Panel of Figure 6 — Figure 6: Schematic, not to scale. (a) Along one direction, the local expansion rate of the learned velocity (solid) peaks between the two modes of the law. The shaded density, with one lobe per mode, is the squared-displacement occupation of the optimal-coupling segments between $\mu_{j}$ and $\widehat{\mu}_{j}$ , which is the weight in eq. 81 : the transport moves mass within each mode, not across the gap where the expansion peaks. The global one-sided rate $\mathcal{L}^{\mathrm{os}}_{j}$ is the supremum of the expansion rate (dotted); the directional rate $\mathcal{E}^{\mathrm{OT}}_{j}$ is its average under this weight (dashed), and stays small. In higher dimension the supremum also runs over all directions. (b) Majorants of $\Lambda_{K}$ ; an arrow from $A$ to $B$ means $A\leq B$ , and each arrow discards the information named beside it. The directional and map-level majorants are not ordered, and Proposition 10 shows that the chain cannot be reversed.

quantity of the analysis	one-dimensional mixture	CIFAR–10
final error $e_{K}$ , factors $\gamma_{j,K}$ , $\Lambda_{K}$ ( Definition 8 )	measured by quadrature	not accessible
one-step recursion ( 10 )	every term measured	not accessible
directional majorant $\Lambda_{K}^{\mathrm{dir}}$ eq. 30	measured	displacement rate $\lambda^{\parallel}$ on coarse–fine segments eq. 41
field-level majorant ( 33 )	finite-domain proxy	worst-direction rate $\lambda^{\max}$ eq. 41 ; empirical EDM margin ( 40 )
stability hypotheses	tested quantitatively	not certified

Appendix

Table 1: Quantities of the analysis and what the two experiments measure of them.

$K$	$e_{K}$	order	$\sum_{j}\widetilde{\delta}_{j,K}$	order	$\Lambda_{K}$	ampl. order	$c_{K}$	measured	universal
	final error		summed biases		amplification			bound / error
17	0.113	–	0.166	–	0.325	–	1.388	1.35	134
34	0.0747	0.597	0.0934	0.826	0.548	0.679	1.469	1.40	127
68	0.0497	0.590	0.0496	0.913	0.795	0.643	1.524	1.40	122
136	0.0332	0.583	0.0255	0.959	1.065	0.612	1.555	1.39	119
272	0.0222	0.581	0.0129	0.981	1.344	0.596	1.572	1.38	118

Appendix

Table 2: Oracle denoiser with exact initialization, under refinement. Orders are the local orders eq. 38 on the pair

(K/2,K)

; the amplification-based order is

1-(\Lambda_{K}-\Lambda_{K/2})/\log 2

, and

c_{K}=Ke_{K}e^{-\Lambda_{K}}

. Bound / error divides the recursion bound ( 39 ) by

e_{K}

, with the measured local biases

\widetilde{\delta}_{j,K}

or with their universal bounds

\Delta_{j}

(

B_{1,1}=16

), propagated by the same measured factors.

Figure 7: Oracle denoiser on the four-component mixture, exact initialization, $M=65{,}536$ quantiles. (a) Final error $e_{K}$ , the recursion bound ( 39 ) built from the measured local biases and factors, and the unweighted sum of the measured biases; triangles mark slopes $1$ and $0.58$ . (b) Measured $\Lambda_{K}$ , with a logarithmic fit on $K\geq 136$ . (c) Local orders eq. 38 on each pair $(K/2,K)$ , against the amplification-based order $1-(\Lambda_{K}-\Lambda_{K/2})/\log 2$ computed from (b). The summed biases converge at first order; the final error does not, and the growth of $\Lambda_{K}$ accounts for the gap.

	$K=17$			$K=1088$
denoiser	$\Lambda_{K}$	directional	field	$\Lambda_{K}$	directional	field
oracle	0.325	1.976	143	1.921	1.022	11.1
trained, seed 0	0.324	1.970	27.5	1.328	1.020	5.1
trained, seed 1	0.334	1.934	28.0	1.875	1.017	3.8
trained, seed 2	0.320	1.990	29.6	1.879	1.018	3.8
misspecified mean	0.205	3.938	944	0.673	1.012	75.6

Appendix

Table 3: Measured

\Lambda_{K}

and its two majorants, as ratios to

\Lambda_{K}

, on the coarsest and finest grids. Directional:

\Lambda_{K}^{\mathrm{dir}}

eq. 30 ; field: the finite-domain proxy for eq. 33 . Canonical Gaussian initialization, common noise threshold.

Figure 8: Where expansion sits, and what each stability control costs. Canonical Gaussian initialization. (a) Oracle on $G_{272}$ at one low noise level: local expansion rate $\lambda_{j}(x)$ , the error weight $w_{j}$ (shaded, arbitrary scale), and the two rates each bound pays at this step, the supremum of $\lambda_{j}$ (field) and its $w_{j}$ -average $\mathcal{E}^{\mathrm{OT}}_{j}$ (directional). The error sits where the predictor contracts. (b) Measured $\Lambda_{K}$ by denoiser; trained networks as median and range over three seeds. (c) Directional bound (dashed) and finite-domain field proxy (dotted), divided by $\Lambda_{K}$ , with the colours of (b).

Figure 9: Stability diagnostics for the pretrained CIFAR–10 EDM model. Shading marks the noise levels at which the 90th-percentile empirical margin of (b) is positive. (a) Normalized denoising loss $r_{\mathrm{reg}}$ per coordinate. (b) Empirical margin ( 40 ) from the median and the 90th percentile of the local stretch; dotted lines mark where the latter crosses $0$ and the supercritical threshold $6/7$ . (c) Energy-weighted rates on the coarse–fine segments of $G_{68}$ , at the same nodes: worst direction, displacement, secant, and random direction eq. 41 . Absolute values; open markers are negative rates. (d) Accumulated synchronous rate over $\sigma\leq 2$ , energy-weighted over all 1280trajectories with its leave-one-trajectory-out range, and the same accumulation of the level-wise median of the per-trajectory rates.

Figure 10: Terms of the measured recursion ( 99 ) from the canonical Gaussian initialization: final error, recursion bound, and its propagated discretization, learning and initialization contributions. (a) Oracle denoiser. (b) Trained networks, median and range over three seeds. The contributions add up to an upper bound, not to the observed error.

$\rho_{\sf EDM}$	final error	measured biases	universal bound	ratio
2	0.0229	0.0164	4.64	282
3	0.0173	0.0095	5.13	539
7	0.0139	0.0065	7.48	1150

Appendix

Table 4: Clock control at

K=544

, oracle with exact initialization; every grid satisfies

\theta_{K}<1/2

. Ratio: universal bound over measured biases.

$\overline{\sigma}$	$K$	final error	change	$\sum_{j}\Delta_{j}$	$\sum_{j}\Pi_{j+1,K_{\mathrm{hi}}}\Delta_{j}$
19.8	209	0.02176	$-1.6\times 10^{-6}$	3.36	1.54
80	272	0.02176	–	13.77	2.81
318	348	0.02176	$+0.1\times 10^{-6}$	47.60	4.00

Appendix

Table 5: Starting-noise control,

\rho_{\sf EDM}=7

with a fixed clock step; oracle started from the exact top law. The change in final error is relative to

\overline{\sigma}=80

. The last two columns sum the universal bias bounds

\Delta_{j}

over the high-noise steps (

\sigma_{j+1}\geq\sigma_{\mathrm{hi}}=4

), without and with the weights

\Pi_{j+1,K_{\mathrm{hi}}}

of the high-noise kernel.

denoiser	$K$	synchronous	optimal coupling	ratio	$\Lambda$ on window
oracle	68	0.909	0.704	1.29	0.793
	272	1.605	1.296	1.24	1.326
misspecified	68	1.112	0.762	1.46	0.804
	272	1.870	0.687	2.72	0.699

Appendix

Table 6: Calibration of the synchronous diagnostic: accumulated linear rates

\sum_{j}\ell_{j}(\mathcal{E}^{\mathrm{sync}}_{j,K})_{+}

and

\sum_{j}\ell_{j}(\mathcal{E}^{\mathrm{OT}}_{j,K})_{+}

over the levels

\sigma_{j}\leq 2

, canonical Gaussian initialization. Ratio: synchronous over optimal coupling. Last column:

\Lambda_{K}

restricted to the same levels,

\sum_{j}\log\max\{1,\gamma_{j,K}\}

pair $(K/2,K)$	quantiles	error order	ampl.-based order	bias-sum order
$(272,544)$	65536	0.5811	0.5886	0.9908
$(272,544)$	262144	0.5779	0.5854	0.9908
$(544,1088)$	65536	0.5850	0.5882	0.9955
$(544,1088)$	262144	0.5782	0.5814	0.9955

Appendix

Table 7: Paired quadrature check with exact initialization. Local orders eq. 38 on the pair

(K/2,K)

; the amplification-based order is

1-(\Lambda_{K}-\Lambda_{K/2})/\log 2

, and the bias-sum order refers to

\sum_{j}\widetilde{\delta}_{j,K}

. Digits show numerical drift rather than certified precision.

quantity	check	criterion	outcome
One-dimensional mixture
endpoint error $e_{K,K}$	$M$ vs $4M$ quantiles on $G_{1088}$ (analytic arms, trained seed 0)	discrepancy $<e_{K,K}/10$	pass; oracle error changes by $<1\%$ . Bound/error ratio not reported for arms without a $4M$ rerun
cumulative log-amplification $\Lambda_{K}$	$M$ vs $4M$ on $G_{1088}$	small relative change	Gaussian start: about $0.1\%$ ; exact start: about $0.45\%$ . Paired orders checked separately in Table 7
one-step recursion ( 98 )	measured slack	nonnegative up to quadrature	pass on every arm and step
field proxy eq. 97	domain $(R,N_{x})=(12,4097)\to(16,8193)$	small relative change	pass; $\leqslant 0.8\%$ . The maximizer hits the domain boundary at 225 of 507 oracle levels on $G_{1088}$ (never for trained arms) but sits on a flat plateau, not clipped
noise threshold	re-summation at the four candidate thresholds of Section G.1.2	range reported	$\Lambda_{K}$ bit-identical; $\Lambda_{K}^{\rm dir}$ varies by $0.13\%$ at $K=1088$

Appendix

Table 8: Numerical reliability checks. Failures lead to omission, not to a qualified report.

Explore similar work

May 8, 2026cs.LG

Geometry-Aware Discretization Error of Diffusion Models

Practical diffusion sampling requires simulating a reverse-time ODE or SDE with a limited number of denoising steps, making the choice of sampling parameters crucial for minimizing discretization error. Non-asymptotic convergence bounds characterize sampling complexity, but their worst-case constants can obscure target geometry and thereby limit guidance on parameter optimization. Rather than bounding the error, we derive asymptotically exact small-stepsize expansions of Euler-Maruyama weak and Frechet errors for general smooth reverse diffusions, with explicit formulas for Gaussian data. These formulas provide tractable objectives for optimizing diffusion parameters, including the noise and rescaling schedules and the stochasticity coefficient, according to the target's covariance spectrum. In particular, our theory predicts lower optimal stochasticity at smaller step budgets, shows how to adapt the rescaling coefficient to the data power spectrum, and motivates a new family of effective noise schedules. A perturbative extension to Gaussian scale mixtures (GSMs) quantifies how departures from Gaussianity shift the optimal parameters. Finally, experiments on different real image datasets show that FID-optimal parameters agree with the qualitative theoretical predictions.

Samuel Hurault, Thomas Moreau, Gabriel Peyré

Univ. Gustave Eiffel, CNRS, LIGM · Univ. Paris-Saclay, Inria, CEA · CNRS, ENS, PSL Université

Diffusion Sampling Diffusion Models Discretization Stochastic Differential Equations Geometry-Aware

Ask PDF

Jul 5, 2026cs.LG

Asymptotic Preservation and Uniform Accuracy of Diffusion and Flow-Matching Samplers

Diffusion and Gaussian-interpolant flow-matching samplers approach data through a terminal noise floor

\varepsilon

, a singular limit for manifold-supported or rank-deficient data. We study two properties of a complete sampler specification, comprising its update rule, time grid, and terminal rule. Asymptotic preservation (AP) means a stable and consistent zero-noise discretization with a step count bounded independently of

\varepsilon

. Uniform accuracy (UA) of order

p

means that, at numerical resolution

h

, the endpoint

W_2

error is

O(h^p)

with a floor-independent constant. Bounded log-noise stepping fails AP because its step count diverges. Stopping a stable base solver at a positive switching scale

a

and appending one map fitted to the analytic normal mode restores AP. On smooth compact boundaryless manifolds, the standard map has exact-input error

O(a^2-\varepsilon^2)

and sharp zero-floor error

Θ(a^2)

. A base solver with a floor-uniform order-

p

estimate on the resolved interval retains that order when

a=O(h^{p/2})

, provided the terminal transfer factor remains bounded. Along exact trajectories, the posterior-mean identity

D(x(σ),σ)=x(σ)-σx'(σ)

cancels the linear terminal defect and enables higher-order fitted maps. A three-evaluation Hermite construction is uniformly third order for exact switching-scale input over

0\le\varepsilon\le a

, and a seven-evaluation construction is fourth order at zero. We classify representative diffusion and flow-matching specifications by AP and UA. On EDM and Rectified Flow checkpoints, a paired decomposition separates base-integration from terminal-completion error and predicts held-out same-seed endpoint errors.

Shiheng Zhang

University of Washington

Accelerated Sampling Time Discretization Invertible Flows Discretization Clocks

Ask PDF

Jun 14, 2026cs.LG

Wasserstein Convergence of ODE-Based Samplers in Decentralized Diffusion Model via Velocity Field Decomposition

Diffusion models have achieved impressive empirical success in generative tasks, and their convergence theory is now relatively well understood. Motivated by privacy and scalability, recent decentralized diffusion architectures replace a single global velocity field with multiple local experts and a routing mechanism, yielding a sampling dynamics with stochastic expert switching that falls outside standard diffusion convergence analyses. In this work, We study a decentralized diffusion framework with stochastic velocity fields and ODE-based sampling. We establish a convergence guarantee in Wasserstein-2 distance, showing that the distribution of the

N

-step discretization converges to the analytical solution at rate

\mathcal{O}(N^{-1/2}+\varepsilon)

W_2

, where

\varepsilon

captures the neural approximation errors. To our knowledge, this is the first

W_2

convergence result for decentralized diffusion models with an ODE-based sampling scheme.

Chencheng Tang, Xuanyu Xue, Fangyikang Wang +2

Peking University · Shanghai Jiao Tong University · MBZUAI +2

Diffusion Sampling Accelerated Sampling Convergence Discretization Velocity Field

Ask PDF