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

Shiheng Zhang

The sampling clock decides stability near data; only stochastic noise, not geometry, charges a logarithmic cost as terminal noise vanishes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 21:35 UTC pith:KZRXZHLI

load-bearing objection Clean AP analysis of terminal-layer sampling: DDIM/rectified-flow uniqueness, log charged only to Itô, and a predictive residual audit that actually meets pre-specified gates.

arxiv 2607.04113 v2 pith:KZRXZHLI submitted 2026-07-05 cs.LG cs.NAmath.NA

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

Shiheng Zhang This is my paper

classification cs.LG cs.NAmath.NA MSC 65L1165C3068T07

keywords diffusion samplersflow matchingasymptotic-preserving methodsa posteriori residual auditsingular perturbationterminal boundary layerDDIMItô budget

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Diffusion and flow-matching models integrate a learned flow from large noise down to a tiny terminal floor where the score stiffens and a boundary layer appears. This paper treats that floor as a singular-perturbation parameter and asks which fixed-step samplers remain stable and uniformly accurate as the floor goes to zero. It shows that Euler stepping in the plain noise scale (the deterministic DDIM update) is the unique layer-exact scheme up to affine reparameterization, with rectified flow its counterpart; other common clocks either expand residuals or stall a fixed distance from the data. On solvable models, deterministic samplers stay first-order accurate with no log factor even across a symmetric posterior switch whose total budget is a universal constant; the logarithm lives entirely in the Itô term of stochastic samplers. The same residual spectra, measured once on a pretrained checkpoint, predict held-out discretization budgets across step counts, schedules, and noise levels with no refitting. A sympathetic reader cares because the analysis turns schedule folklore into rigidity theorems and gives a practical a-posteriori audit that needs only the learned network.

Core claim

As the terminal noise floor tends to zero, fixed-step samplers are asymptotic-preserving precisely when their clock keeps the normalized residual order-one; Euler in the σ-clock is the unique layer-exact frozen-field discretization (rectified flow its flow-matching twin), deterministic residual budgets remain O(1) even across a symmetric switching interface, and the log(1/σ_min) cost is charged exclusively to the Itô term of stochastic samplers.

What carries the argument

The asymptotic-preserving a-posteriori audit: four residual functionals (E1 residual amplification, E2 path budget, M_det deterministic spectrum, M1 Itô spectrum) whose coefficients stay O(1) as the floor vanishes and that are computable from checkpoint evaluations alone, without ground-truth scores or exact trajectories.

Load-bearing premise

The learned denoiser must be twice continuously differentiable with square-integrable residual, material derivative and Jacobian remainders along the flows, and the interface analysis is restricted to a single symmetric two-point mixture.

What would settle it

On any pretrained checkpoint, measure the residual spectra once from a pilot set of states; if those spectra fail to forecast held-out residual budgets across step counts, schedules and noise levels within the paper’s pre-specified relative-error gates, the predictive audit claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Practitioners can treat the σ-clock / DDIM step (and rectified-flow linear schedules) as the default layer-exact choice rather than an empirical preference.
Schedule design for deterministic sampling reduces to resolving the data’s spectral band; the terminal layer itself needs no graded mesh once the operator is fitted.
Stochastic samplers necessarily pay a horizon-extensive path-KL of order Λ²/N; any claim of log-free stochastic sampling must cancel the Itô coefficient by a non-scalar freeze.
A single set of residual spectra measured on a checkpoint can be reused to rank candidate meshes and noise levels without per-configuration refitting or ground-truth trajectories.
The logarithm that appears in sampling complexity is intrinsic to reinjected noise, not to posterior geometry or a single switching interface.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the open commitment-integral sign for asymmetric mixtures is nonzero, deterministic uniform accuracy would drop to square-root order precisely when basin misallocation occurs, giving a sharp diagnostic for multi-modal data.
A practical Rosenbrock-style freeze of the local Jacobian could remove the horizon logarithm from stochastic sampling while remaining compatible with the same audit functionals.
The same residual spectra could be monitored during training as an early-stopping or architecture signal for how much of the discretization budget is still model error rather than numerical error.
Scale-rich hierarchical data would re-introduce a deterministic logarithm through stacked O(1) interface budgets, predicting that multiscale mixtures are the regime where fitted operators alone no longer suffice.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Clean AP analysis of terminal-layer sampling: DDIM/rectified-flow uniqueness, log charged only to Itô, and a predictive residual audit that actually meets pre-specified gates.

read the letter

This is the cleanest singular-perturbation take on diffusion/flow-matching samplers I have seen. The punchline is sharp and supported: on the pure terminal layer, Euler in the σ-clock (deterministic DDIM) is the unique layer-exact frozen-field step up to affine reparameterization, with rectified flow as its flow-matching twin; the λ-clock is stable only below the Lambert-W threshold h⋆ = 1 + W(1/e); the heat clock stalls at a Wallis floor independent of σ_min. Deterministic samplers stay first-order W2-uniformly accurate with no log(1/σ_min) even across a symmetric posterior-switching interface whose budget is a universal constant (1/4 + 3π/16 ≈ 0.84, quadrature ~1.01). The logarithm lives entirely in the Itô term of stochastic samplers (path-KL ~ Λ²/N vs ODE O(Λ²/N²)). That assignment is new and useful.

What works: closed-form calibrations on the rank-deficient Gaussian and symmetric two-point mixture with named sharp constants, an Itô residual law with the explicit σ-cancellation, and a genuine predictive campaign on the public EDM CIFAR-10 checkpoint. Spectra measured once from 256 pilot states forecast held-out residual budgets across N, schedule, and β against pre-specified gates with no refit; M1 calibrates at 1.00 ± 0.01 by matching a static same-σ kick to a dynamical jump. The audit functionals (E1, E2, M_det, M1) are computable on any checkpoint without ground truth. Math, data, and citations look solid; the paper is honest about its scope (silent on model error δ, open sign of the commitment integral for asymmetric mixtures).

Soft spots are real but proportionate: standing assumption (A) is C² with square-integrable remainders (exact on the models), the interface result is for the symmetric case only, and the audit certifies discretization of the learned field, not FID or model error. The negative gate (ledger-optimal schedule is not FID-optimal) is a feature, not a dodge. No circularity; constants are derived, not fitted.

This is for people who care about sampler design, numerical SDEs, or a-posteriori estimation under singular limits. It reorganizes a core design axis (clock vs noise vs geometry) and supplies a practical tool. I would bring it to reading group, cite the rigidity theorems and the log assignment, and send it to peer review without hesitation.

Referee Report

0 major / 5 minor

Summary. The paper treats the terminal noise floor σ_min of diffusion and flow-matching samplers as a singular-perturbation parameter and asks which fixed-step schemes are asymptotic-preserving (stable and uniformly accurate as σ_min o 0). It reduces every sampler to a traversal of one learned residual field under a clock, gauge, noise level, and frozen object, then packages the AP criteria as an a-posteriori audit of residual functionals (E1, E2, M_det, M1) with σ_min-uniform coefficients that can be read from a pretrained checkpoint without ground-truth scores. On the pure terminal layer, Euler in the σ-clock (deterministic DDIM) is shown to be the unique layer-exact frozen-field discretization up to affine reparameterization, with rectified flow its flow-matching counterpart; the λ-clock is residual-stable only for h ≤ h⋆ = 1 + W(1/e) and the uniform-σ^{2} heat clock stalls a σ_min-independent distance from the data (Theorem 2, Proposition 4). On the rank-deficient Gaussian and a symmetric two-point mixture, deterministic samplers remain first-order W2-uniformly accurate with no log(1/σ_min) factor (Theorems 3–4, Corollaries 2–3); the logarithm is charged entirely to the Itô term of stochastic samplers, whose path-KL scales as Λ^{2}/N against the ODE’s O(Λ^{2}/N^{2}) budget (Corollary 4). Spectra measured once on the public EDM CIFAR-10 checkpoint predict held-out residual budgets across step count, schedule, and noise level against pre-specified gates with no refit, and calibrate

Significance. If the results hold, the paper supplies a clean singular-perturbation account of terminal-layer sampling that converts empirical schedule and stochasticity folklore into rigidity theorems with named sharp constants (h⋆, Wallis floor 1/√π, tangential 1/4, interface tail 3π/16). The assignment of the logarithm—absent from deterministic budgets even across a symmetric switching interface, and intrinsic to the Itô term already on rank-deficient Gaussian data—is a sharp separation of ODE from SDE sampling that sharpens existing upper bounds. The a-posteriori audit is genuinely predictive: Stage-0 spectra forecast held-out budgets against pre-specified gates with no per-configuration refitting, and the same-time kick versus dynamical-jump calibration of M1 is parameter-free. Closed-form calibrations, explicit Itô cancellations, and a public-checkpoint campaign with released predictions.csv make the claims checkable. The work is scoped honestly (silent on model error δ and perceptual quality) and therefore useful as a diagnostic rather than a schedule optimizer.

minor comments (5)

Standing assumption (A) is stated clearly and holds exactly on the solvable models, but a short remark on how far C^{2} and square-integrability of the learned residual can be expected to fail near the EDM floor would help readers judge the audit’s domain of validity on real checkpoints.
Appendix E leaves the sign of the commitment integral κ for asymmetric mixtures open; a one-line numerical quadrature (or a statement that the sign is left unresolved) would close the only remaining deterministic open point without expanding the paper’s scope.
Figure 2a caption and surrounding text correctly note that full-step increments do not isolate the Itô term; a single sentence cross-referencing the two-moment inversion (Eq. 29) earlier in the paragraph would make the control experiment easier to follow on first reading.
Table 3 (clocks and conventions) is useful; adding the explicit affine map that takes the flow-matching velocity back to η would make the “one flow, many samplers” claim fully self-contained.
A few typographical inconsistencies remain (e.g., “It ˆo” vs. “Itô”, occasional missing spaces around σ_min). A light copy-edit pass would remove them.

Circularity Check

0 steps flagged

No significant circularity: rigidity theorems, named constants, and held-out budget forecasts are derived or measured independently of the quantities they certify.

full rationale

The load-bearing claims (layer-exactness of the σ-clock/DDIM and rectified flow; first-order UA of deterministic samplers with no log(1/σ_min) even across a symmetric interface; logarithm charged only to the Itô term) rest on closed-form calculations on two solvable models (Props. 2–4, Thms. 2–4, Cors. 2–4) whose every functional is explicit, plus a predictive audit whose Stage-0 spectra (M_det from centered differences, M_1 from same-σ kicks) are fixed once and then used, without refit, to forecast held-out E_2 budgets across N, schedule and β against pre-specified gates. The named constants (h⋆=1+W(1/e), Wallis 1/√π, 1/4, 3π/16) arise by direct solution of residual-amplification equations or by quadrature of the exact scale spectra; they are not fitted. M_1=1.00±0.01 is a zero-parameter match of a static Jacobian reading to a dynamical jump, not a free parameter tuned to the budgets. Standing assumption (A) holds exactly on the calibration models and is acknowledged as a scope restriction; the open sign of the asymmetric commitment integral is likewise flagged. No equation reduces a claimed prediction to its own input by construction, no uniqueness theorem is imported from overlapping authors, and no ansatz is smuggled via self-citation. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard Itô calculus, Tweedie identities and two exactly solvable models; the only non-standard ingredients are the standing regularity assumption (A) and the restriction to symmetric mixtures for the no-log interface result. No free parameters are fitted to obtain the named constants or the M1 calibration; the audit functionals are defined from the learned residual and are therefore checkpoint-dependent by design.

axioms (4)

domain assumption Standing assumption (A): D(·,σ) ∈ C² with η, D_λη, I−J_D and third-order Taylor remainders square-integrable along the flows considered.
Invoked for Theorem 1 (local residual law) and all subsequent budget statements; guarantees the O(1) coefficients that make the audit asymptotic-preserving.
domain assumption The data distribution is either rank-deficient Gaussian or a symmetric two-point mixture (asymmetric case left open).
Used for closed-form spectra and the universal-constant interface budget of Theorem 4; symmetry is essential for order-preservation and shared fixed-point arguments that convert finite budget into W2-UA.
standard math Itô calculus and Girsanov change-of-measure for the reverse SDE family (standard).
Underpins the local residual law (Theorem 1) and the path-KL conversion (Corollary 1).
standard math Tweedie formula relating denoiser, score and posterior covariance.
Used throughout to rewrite every sampler in the (D,η) chart and to identify J_D with posterior covariance.

invented entities (2)

AP residual audit functionals (E1, E2, M_det, M1) independent evidence
purpose: Provide σ_min-uniform, checkpoint-computable certificates of terminal stability and scale-local discretization cost without ground-truth scores.
Defined in Definition 2 from η-increments of the learned field; they are the operational interface that turns the singular-perturbation analysis into a practical a-posteriori tool.
Layer-exact clock classification (σ-clock / rectified flow unique up to affine reparameterization) independent evidence
purpose: Identify which fixed-step discretizations remain residual-AP as σ_min → 0.
Proved as Proposition 3 / Theorem 2 and Proposition 4; the uniqueness is a rigidity statement, not an empirical observation.

pith-pipeline@v1.1.0-grok45 · 31299 in / 2881 out tokens · 26679 ms · 2026-07-11T21:35:29.210091+00:00 · methodology

0 comments

read the original abstract

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 $\Theta(a^2)$. 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 $D(x(\sigma),\sigma)=x(\sigma)-\sigma x'(\sigma)$ 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.

Figures

Figures reproduced from arXiv: 2607.04113 by Shiheng Zhang.

**Figure 1.** Figure 1: Uniform-in-σmin behaviour on the rank-deficient Gaussian, closed-form from Theorems 2–3. (a) Terminal W2 vs. steps N (λ-uniform mesh): DDIM (σ-clock) is first order, its curves for σmin ∈ {10−2 , . . . , 10−5} nearly coincident (only σmin-dependence is Λ = log(σmax/σmin) in h = Λ/N; Corollary 3), while the uniform-σ 2 heat-clock stagnates at the Wallis floor σmax/ √ πN (Theorem 2c). (b) Residual amplifica… view at source ↗

**Figure 2.** Figure 2: Scale spectrum on the pretrained EDM CIFAR-10 checkpoint (Karras et al., 2022) (256/128 seeds, 64 steps). (a) The full-step E∥ηn+1−ηn∥ 2/(D ∆λ 2 ) for ODE and EDM churn nearly overlap: the completed step does not isolate the Ito term. ˆ (b) The injection-local churn jump, inverted via equation 29, reads M1 = 1 D E ∥I − JD∥ 2 F = 1.00 ± 0.01 across the high/mid-σ plateau for γ ∈ {0.02, 0.05, 0.10, √ 2 − 1},… view at source ↗

**Figure 3.** Figure 3: Predictive a posteriori audit on the pretrained [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance
cs.LG 2026-07 conditional novelty 7.0

Replacing CFG's w(r-1) coefficient with r^(1+w)-r removes a sigma_min-divergent residual blow-up on a Gaussian calibration model and stabilizes high-guidance diffusion sampling at zero extra NFE.