acceptodds
Under review as a conference paper at ICLR 2027

Flow Matching under Noisy Latent Structure: Beyond Exact Low-Dimensional Support

Abstract

Flow Matching (FM) learns a velocity field whose ODE transports a simple source distribution to a target law. Existing finite-sample theory largely treats ambient-space regularity or data supported exactly on low-dimensional sets. We study linear FM under a noisy latent-generator model, where a low-dimensional H\"older map is perturbed by nondegenerate ambient Gaussian noise, so the target law is full-dimensional despite its latent structure. We construct a spatially regular ReLU velocity class and establish non-asymptotic high-probability approximation and estimation bounds whose leading sample-size exponent is governed by the latent dimension rather than the ambient dimension, with ambient and noise dependence kept explicit. Fixed positive target noise keeps the interpolation nondegenerate over the full time interval. The same spatial regularity propagates the learned velocity error through the transport ODE, yielding a corresponding Wasserstein convergence guarantee. These results show that exact low-dimensional support is not necessary for Flow Matching to retain latent-dimensional statistical behavior.

Primary area:
generative models
Keywords:
Flow Matching, Generative Modeling, Statistical Learning Theory, Low-Dimensional Structure, Latent Variable Models, Wasserstein Convergence

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rBbAOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies linear Flow Matching when the target observations follow the noisy latent-generator model. Although the target distribution has full DD-dimensional support for every fixed σ>0\sigma>0, the authors show that the nonlinear part of the population FM velocity can be represented through a posterior mean over the dd-dimensional latent variable. They construct a spatially regular ReLU approximation whose leading resolution complexity scales as ε−d\varepsilon^{-d}, prove a high-probability integrated velocity-risk bound with leading rate n−2β/(2β+d)n^{-2\beta/(2\beta+d)}, and propagate this error through the learned ODE to obtain a W2W_2 convergence rate with leading factor n−β/(2β+d)n^{-\beta/(2\beta+d)}. The paper also gives an Euler-discretization bound under an additional temporal-regularity assumption.

The result should not be interpreted as learning to remove the target noise or recover the clean law of g∗(U)g^*(U). The target law in the analysis is the noisy, full-dimensional distribution itself. The contribution is instead an intrinsic-dimensional statistical analysis of FM without requiring exact low-dimensional support or a separately learned latent representation.

Critical strengths
  1. A technically interesting extension beyond exact low-dimensional support. The paper proves that the leading approximation and estimation exponents can depend on the latent dimension dd even though the observed distribution is full-dimensional. This addresses a gap between exact-manifold theory and noisy latent-variable models and yields the rate n−2β/(2β+d)n^{-2\beta/(2\beta+d)} under the paper's assumptions.

  2. The approximation class is designed for both estimation and transport. The authors do more than approximate the population velocity in value: they carry a spatial Lipschitz profile through the same neural construction and use it to convert integrated velocity error into a non-asymptotic W2W_2 guarantee. This makes the statistical result relevant to the generated distribution rather than stopping at a regression bound.

Critical weaknesses
  1. The paper does not establish that the latent-dimensional exponent yields meaningful sample efficiency in a practically relevant regime. Theorem 4.5 chooses ε=n−1/(2β+d)\varepsilon=n^{-1/(2\beta+d)} but also requires conditions (5)--(6). For example, ε−2β>D\varepsilon^{-2\beta}>D already implies D<n2β/(2β+d)D<n^{2\beta/(2\beta+d)}, while the other conditions impose additional restrictions involving HH, DD, and σ−1\sigma^{-1}, including terms that grow polynomially with DD and deteriorate as σ\sigma decreases. Moreover, the network width and sparsity scale polynomially with DD. Thus the result has a latent-dimensional exponent but not a uniformly latent-dimensional sample or computational complexity. The paper neither translates these assumptions into a clear non-vacuous high-dimensional regime nor demonstrates that the favorable exponent dominates the ambient-dependent thresholds and constants at realistic sample sizes. This substantially weakens the significance of the central claim.

  2. The finite-sample theorem does not analyze the FM objective used in implementable training. In equation (9), the expectation over the Gaussian source and the integral over time are evaluated exactly inside each per-target loss. Consequently, the finite-sample randomness comes only from the nn target observations. Practical FM instead samples X0X_0 and tt and optimizes the resulting stochastic objective. The theorem therefore omits auxiliary Monte Carlo error, optimization error, finite minibatches, and the difficulty of finding a global ERM under the imposed Lipschitz constraint. Although the paper acknowledges finite auxiliary sampling as future work, this gap makes it difficult to interpret the result as a sample-complexity guarantee for a trainable FM procedure.

  3. The paper gives no evidence that its asymptotic phenomenon is observable, even in a controlled setting. I do not view experiments as mandatory for every theoretical paper. Here, however, the main significance question is precisely whether the favorable dependence on dd survives the large DD- and σ−1\sigma^{-1}-dependent constants and applicability thresholds. A small synthetic experiment varying (n,d,D,σ)(n,d,D,\sigma) could test whether the predicted slopes appear, when ambient dependence dominates, and how the exact-integrated objective differs from sampled training. Without either such evidence or a sharper theoretical regime analysis, the practical meaning of the main rate remains unresolved.

  4. The Euler result depends on a temporal-regularity condition that is not derived from the learned class. Assumption 5.3 requires a deterministic γt\gamma_t, independent of nn, controlling the time variation of the learned ERM. Section 3 explicitly imposes no temporal-regularity constraint on the class. The Euler proposition is valid as a conditional statement, but it is unclear whether the proposed network class and ERM satisfy this assumption with the claimed uniformity. The authors should either construct a class that enforces this property and track its complexity, or clearly present Proposition 5.4 as an additional conditional guarantee whose assumption is not established by the preceding learning theory.

Questions for authors
  1. Can the authors state a simple corollary giving sufficient growth conditions on DnD_n and σn\sigma_n under which (5)--(6) hold and the advertised rate is non-vacuous?
  2. What additional term should one expect if the time variable and Gaussian source are approximated using a finite number of Monte Carlo samples per target observation?
  3. Does the constructed neural class imply Assumption 5.3 with a constant independent of nn? If not, how should the Euler result be interpreted for the ERM analyzed in Theorem 4.5?
  4. Which parts of the posterior-network approximation are new relative to the noisy-latent score-matching machinery of Yakovlev and Puchkin (2025), beyond the clearly FM-specific velocity representation, excess-risk analysis, and ODE propagation?
5 comments
AuthorsOpenReview ↗

Response to Reviewer rBbA — 1. Dimension, sample-size regimes, and the controlled experiment

Thank you for asking how the latent-dimensional rate behaves alongside the full dimension-dependent factors and scale conditions. We have added explicit sufficient growth regimes and a controlled image experiment that examines velocity learning as the ambient dimension increases.

Changes in the revised manuscript: Corollaries E.1–E.2 and Appendix H, Table 1; the original conditions are (5)–(6).

When are the conditions non-vacuous? Let s=2β+ds=2\beta+d, α=2β/s\alpha=2\beta/s, and c=carchc=c_{\rm arch} as in Theorem 4.5. Fix d,β,Hd,\beta,H, a positive σ\sigma, and δn=n−p\delta_n=n^{-p}. Corollary E.1 verifies every part of (5)–(6) eventually if, for example,

D_n\le Cn^\kappa,\qquad 0<\kappa<\min\left\\{\frac{2}{3s},\frac{\alpha}{c+5/2}\right\\}.

The corresponding integrated risk is at most Cn−α+(c+5/2)κ(log⁡n)c+1Cn^{-\alpha+(c+5/2)\kappa}(\log n)^{c+1} and the continuous-flow W2W_2 bound tends to zero. In particular, Dn=O(log⁡n)D_n=O(\log n) also satisfies the conditions. For a joint example, Corollary E.2 takes

Dn=⌈(log⁡n)1/2⌉,σn=(log⁡n)−3/8.D_n=\lceil(\log n)^{1/2}\rceil,\qquad \sigma_n=(\log n)^{-3/8}.

It gives noisy-target W2≤n−β/s+o(1)W_2\le n^{-\beta/s+o(1)} and an additional coupling bias O((log⁡n)−1/8)O((\log n)^{-1/8}) to the clean generator law. These two corollaries give concrete sufficient regimes for continuous-flow consistency. Numerical sampling is described separately by Proposition 5.5, with explicit noise-dependent constants.

What do we observe in a controlled example? Appendix H uses a two-dimensional latent Gaussian-spot image family, σ=0.1\sigma=0.1, and a time-conditioned U-Net. Each entry is mean ±\pm sample standard deviation over three training seeds; evaluation uses 16,384 held-out tuples per seed and a numerical posterior.

DDnnRisk, 1000 updatesRisk, 3000 updates
25681920.02967±0.001240.02967\pm0.001240.01629±0.001170.01629\pm0.00117
102481920.03113±0.002820.03113\pm0.002820.01617±0.000770.01617\pm0.00077
409681920.03442±0.001830.03442\pm0.001830.01761±0.000350.01761\pm0.00035
1024327680.03103±0.001090.03103\pm0.001090.01582±0.000320.01582\pm0.00032
10241310720.03080±0.001810.03080\pm0.001810.01608±0.001060.01608\pm0.00106

At fixed n=8192n=8192, the 3000-update risk changes from 0.016290.01629 to 0.017610.01761 as DD increases sixteenfold, while the corresponding zero-posterior risks are 4.5644.564, 4.4204.420, and 4.4884.488. With clean-signal norm fixed at 0.90.9, the expected squared noise norm increases from 2.562.56 to 40.9640.96 across this comparison. The observed velocity risk thus remains stable across these resolutions in this noisy-latent image family. For d=2d=2 and β=1\beta=1, the first theoretical scale condition is n>D2n>D^2; the experiment provides a complementary finite-budget view at smaller sample sizes for this structured instance.

At D=1024D=1024, the 3000-update risks are 0.016170.01617, 0.015820.01582, and 0.016080.01608 as nn increases from 81928192 to 131072131072. The fixed-budget sample-size sweep is therefore approximately flat, while increasing the budget from 1000 to 3000 updates consistently lowers risk in all configurations. This experiment uses a time-conditioned U-Net with convolutional weight sharing and finite-step AdamW training; our theorems analyze exact ERM in a spatially constrained ReLU class. The fixed values of dd and σ\sigma make the table a focused comparison of the tested resolutions and training budgets, complementing the dimension- and noise-dependent regimes in Appendix E.

Reference

AuthorsOpenReview ↗

Response to Reviewer rBbA — 2. Finite auxiliary Monte Carlo samples

Thank you for pointing out the role of auxiliary sampling in an implementable objective. In Appendix F we now analyze a finite-sampling ERM over the same spatially constrained class as Theorem 4.5. The analysis follows the shared-auxiliary scheme of Zhou and Liu (2025): draw nn target observations and mm independent time–Gaussian pairs, then reuse those mm pairs for each target. Theorem 4.5 itself continues to concern the exact-average objective.

Let L^n\widehat{\mathcal L}_n and L^n,m\widehat{\mathcal L}_{n,m} denote the exact-average and sampled objectives, and set v0(x,t)=atxv_0(x,t)=a_tx. Proposition F.1 controls the uniform difference of their anchored versions:

sup⁡v∈V∣[L^n,m(v)−L^n,m(v0)]−[L^n(v)−L^n(v0)]∣≤EMC,\sup_{v\in\mathcal V}\left|[\widehat{\mathcal L}_{n,m}(v)-\widehat{\mathcal L}_{n,m}(v_0)]-[\widehat{\mathcal L}_n(v)-\widehat{\mathcal L}_n(v_0)]\right|\le E_{\rm MC},

where the leading auxiliary term is

EMC;≲;K![ρ+Hρ+log⁡(8/δ)m]+Gaussian tail terms.E_{\rm MC}\\;\lesssim\\;K\\!\left[\rho+\sqrt{\frac{\mathcal H_\rho+\log(8/\delta)}{m}}\right]+\text{Gaussian tail terms}.

Here Hρ\mathcal H_\rho is the log covering number of the posterior-network class on the localized domain; Appendix F gives KK, ρ\rho, and the tail terms explicitly. The mm in the denominator reflects the mm independent shared pairs: the nmnm cross-product evaluations reuse both targets and auxiliary variables and are not nmnm independent target observations. Subtracting v0v_0 cancels the common unbounded affine contribution and does not change the minimizer.

Combining this deviation with the original target-sample concentration gives, with high probability,

R(v~n,m)≤53inf⁡v∈VR(v)+C(Jn+EMC),R(\widetilde v_{n,m})\le \tfrac53\inf_{v\in\mathcal V}R(v)+C(J_n+E_{\rm MC}),

where JnJ_n is the original target-sample remainder and RR is the integrated squared velocity error. Thus the auxiliary error is explicit and separate from the error due to the nn target observations. For fixed model parameters, set s=2β+ds=2\beta+d, α=2β/s\alpha=2\beta/s, ε=n−1/s\varepsilon=n^{-1/s}, and m=⌈n1+α⌉m=\lceil n^{1+\alpha}\rceil. The original comparator and covering bounds then yield R(v~n,m)≲n−α(log⁡n)CR(\widetilde v_{n,m})\lesssim n^{-\alpha}(\log n)^C, under the stated scale conditions. This is a sufficient budget for this particular shared-sampling bound.

In Appendix H, fresh time and Gaussian source variables are drawn at each U-Net minibatch update. Appendix F instead analyzes exact ERM for a fixed shared set of auxiliary pairs; these two settings provide complementary practical and statistical perspectives.

Zhou and Liu use the same decomposition, Gaussian truncation, and covering strategy for bounded-support targets and a different velocity class. The anchoring and target-tail localization in Appendix F adapt the argument to the present noisy-latent model.

Reference

AuthorsOpenReview ↗

Response to Reviewer rBbA — 3&4. Euler consistency and posterior-network novelty

Thank you for asking whether the discretization uses the same estimator as the learning theorem and which approximation ingredients are new. These questions clarify two distinct interfaces.

Changes in the revised manuscript: Section 3, Proposition 3.2, Appendix B.3–B.4, Section 5.2, Proposition 5.5, and Appendix G.1–G.2.

The estimator and Euler. Theorem 4.5 studies ERM in a class with a spatial envelope. The class does not itself establish Assumption 5.3 with a constant independent of nn. Proposition 5.4 gives a deterministic Euler guarantee when the estimator also satisfies this additional temporal condition. Proposition 5.5 connects the same estimator directly to a finite-step sampler using its existing spatial envelope. After training, draw independent τk∼Unif⁡[tk,tk+1]\tau_k\sim\operatorname{Unif}[t_k,t_{k+1}] and set Yk+1=Yk+hkv^(Yk,τk)Y_{k+1}=Y_k+h_k\widehat v(Y_k,\tau_k), where hk=tk+1−tkh_k=t_{k+1}-t_k. Writing R=∫01∣v^−v∗∣L2(πt)2dtR=\int_0^1\\|\widehat v-v^*\\|^2_{L^2(\pi_t)}dt and h=max⁡khkh=\max_kh_k, its conditional output law, averaged over the numerical times and independent initialization, satisfies

W2(Law⁡(YM∣training),π1)≤eγˉx(R+Cσ,Dh),Cσ,D≲(1+D)(1+σ−4),γˉx=∣γx∣L∞(0,1).W_2(\operatorname{Law}(Y_M\mid\text{training}),\pi_1)\le e^{\bar\gamma_x}(\sqrt R+C_{\sigma,D}h),\qquad C_{\sigma,D}\lesssim(1+\sqrt D)(1+\sigma^{-4}),\quad \bar\gamma_x=\\|\gamma_x\\|_{L^\infty(0,1)}.

The key step is ∑khk∫tktk+1!∣v^−v∗∣L2(πt)2dt≤R\sum_k\sqrt{h_k\int_{t_k}^{t_{k+1}}\\!\\|\widehat v-v^*\\|^2_{L^2(\pi_t)}dt}\le\sqrt R. Spatial regularity controls the learned field along the numerical trajectory, and the required time regularity of the true field is derived from the noisy-latent model. Thus the result closes the discretization interface without adding temporal regularity to the learning class. The shared finite-auxiliary ERM in Appendix F can be used analogously after conditioning on its training randomness, with its risk controlled by the bound there.

Which posterior-network ingredients are shared or added? We use Yakovlev and Puchkin’s latent-cell Taylor and Gaussian-posterior machinery for the same noisy-generator model. The Gaussian posterior covariance identity and general first-order ReLU approximation techniques are also known tools. Our added construction is a single comparator whose posterior-value approximation and global spatial Lipschitz envelope hold together, with the leading latent-cell count ε−d\varepsilon^{-d} retained and the full ambient/noise factors tracked. Its value approximation enters the FM risk; its spatial envelope places this same comparator in the class used for both constrained ERM and transport. Proposition 3.2 separately supplies the full-time generator-to-FM-velocity stability estimate, which connects the latent surrogate to that comparator. We have made these dependencies and the imported primitives explicit in Section 3 and Appendix B.

Related FM work by Fukumizu et al. also uses a spatially Lipschitz-constrained network class, for Besov-density targets and different interpolation assumptions. Our construction addresses the noisy latent-generator setting above.

The population score–velocity identity makes the relation to score matching transparent. Under the OU time change, full-time FM risk carries the weight e2τ−1\sqrt{e^{2\tau}-1}, while the stated YP score-risk theorem is unweighted on a truncated interval. Additional estimates are consequently needed for a transfer of the full-time risk and the spatially constrained comparator.

References

Reviewer_bAMYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing statistical theory for Flow Matching either assumes regularity of the data distribution in the full ambient space or assumes the data lies exactly on a low-dimensional set. In this article, the authors consider the setting where the target X1X_1 is obtained applying a Holder function g to a low dd-dimensional uniform variable UU and then perturbing the result with D>>dD>>d-dimensional Gaussian noise. That is, $ X_1= g(U)+\sigma\xi $

  • Approximation (Thm 3.1). The authors show that the population velocity decomposes as v∗(x,t)=atx+bt⋅mt(x)v*(x,t) = a_tx + b_t·m_t(x), where mₜ is a posterior mean over the latent variable. Sparse ReLU networks can approximate it with complexity ε−dε^{-d} and squared L2L^2 error about ε2β\varepsilon^{2\beta}.

  • Generalization (Thm 4.5). The empirical risk minimizer over this class has velocity error of order n−2β/(2β+d)n^{- 2 \beta/(2 \beta +d)} with high probability. The ambient dimension DD, noise σσ and Hölder radius HH appear only in the prefactors and log terms.

  • Wasserstein convergence (Thm 5.2). A Grönwall argument pushes the velocity error through the ODE, providing theoretical guarantees of convergence in $ W_2 $. An Euler-discretization bound follows (Prop 5.4), under an added temporal-regularity assumption.

Critical strengths
  • I believe that the assumption that X1X_1 is a noisy perturbation of a distribution exactly supported on a low dimensional manifold is a natural generalization of the manifold assumption.

  • The main result, i.e. the fact that statistical rates of convergence depend almost exclusively on the dd, appears to be quite interesting to me and in line with recent findings under the manifold assumption.

Critical weaknesses
  • Dependency on $ D $ remains in the prefactors

  • The additional assumption on the time-regularity of the velocity field is hard to check, and is not directly imposed on the data distribution.

  • I cannot assess whether the results obtained here follow from an adaptation of results obtained under the manifold assumption or if the proofs contain genuinely new elements. Unfortunately, I am not an expert of this stream of literature and this should be taken more as a limitation of the reviewer rather than of the paper.

Questions for authors
  • Recent progress in the analysis of the sampling error for diffusion flow matching models does note require time-regularity of the velocity fields. Can the authors imagine profiting from these findings to improve Proposition 5.4?

  • If the noise intensity is zero, i.e. X1=g(U)X_1=g(U) what do the findings of this article say? If they dont' cover this case, can they give some intuition on why a result covering both scenarios is difficult to achieve?

4 comments
AuthorsOpenReview ↗

Response to Reviewer bAMY — 1. Time regularity and sampling

Thank you for recognizing the noisy-latent assumption and the intrinsic-dimensional rate, and for suggesting that we consider sampling analyses without learned-field time regularity. Your suggestion led us to add a random-time scheme for the same estimator and spatially constrained class studied in Theorem 4.5. For the ambient-dimension dependence you noted, Appendix E now gives explicit sufficient regimes, retaining the full dimension- and noise-dependent factors.

Changes in the revised manuscript: Section 5.2, Proposition 5.5, and Appendix G.1–G.2. Proposition 5.4 remains a deterministic Euler result under Assumption 5.3.

The revised scheme chooses independent τk∼Unif⁡[tk,tk+1]\tau_k\sim\operatorname{Unif}[t_k,t_{k+1}] and sets

Yk+1=Yk+hkv^(Yk,τk),hk=tk+1−tk.Y_{k+1}=Y_k+h_k\widehat v(Y_k,\tau_k),\qquad h_k=t_{k+1}-t_k.

The numerical times and initialization are independent of the training data. Conditioned on those data, let π^rand\widehat\pi^{\rm rand} denote the output law averaged over initialization and numerical times, and write h=max⁡khkh=\max_kh_k and R=∫01∣v^−v∗∣L2(πt)2dtR=\int_0^1\\|\widehat v-v^*\\|^2_{L^2(\pi_t)}dt. Proposition 5.5 gives

W2(π^rand,π1)≤eγˉx(R+Cσ,Dh),Cσ,D≲(1+D)(1+σ−4),W_2(\widehat\pi^{\rm rand},\pi_1)\le e^{\bar\gamma_x}\bigl(\sqrt R+C_{\sigma,D}h\bigr),\quad C_{\sigma,D}\lesssim(1+\sqrt D)(1+\sigma^{-4}),

where γˉx=∣γx∣L∞(0,1)\bar\gamma_x=\\|\gamma_x\\|_{L^\infty(0,1)} is the uniform spatial envelope of the learned class. With Theorem 4.5, for fixed model parameters this has the leading statistical factor O~(n−β/(2β+d))\widetilde O(n^{-\beta/(2\beta+d)}) and a first-order grid term.

The reason randomization helps is its direct match to the training risk. Couple the numerical scheme with the population ODE Z˙t=v∗(Zt,t)\dot Z_t=v^*(Z_t,t); the latter has law πt\pi_t. Independence and uniform sampling give

hk∣v^(Zτk,τk)−v∗(Zτk,τk)∣L2=(hk∫tktk+1!∣v^−v∗∣L2(πt)2dt)1/2.h_k\\|\widehat v(Z_{\tau_k},\tau_k)-v^*(Z_{\tau_k},\tau_k)\\|_{L^2}=\left(h_k\int_{t_k}^{t_{k+1}}\\!\\|\widehat v-v^*\\|^2_{L^2(\pi_t)}dt\right)^{1/2}.

Summing these terms by Cauchy–Schwarz gives R\sqrt R. Spatial Lipschitz control handles the numerical trajectory, while the true velocity’s time regularity is derived from the noisy-latent model in Appendix G.1. The resulting guarantee uses the existing spatial class, with output law averaged over the independent numerical times and initialization.

Randomized Euler methods for ODEs with irregular time dependence were studied by Kruse and Wu. Beyler and Bach analyze Wasserstein error for diffusion probability-flow samplers under a different deterministic-grid setup. These works provide related numerical and diffusion-sampling context; the proof in Appendix G establishes the FM integrated-risk and model-derived regularity interfaces stated above.

References

AuthorsOpenReview ↗

# Response to Reviewer bAMY — 2. Zero data noise

Thank you for asking about the clean target. This question made us examine the endpoint separately.

Changes in the revised manuscript: Appendix G.3, including the explicit early-stopping decomposition.

Our full-time results assume σ>0\sigma>0. When σ=0\sigma=0, the interpolation Xt=tG+(1−t)ZX_t=tG+(1-t)Z has conditional Gaussian variance qt=(1−t)2q_t=(1-t)^2: it is positive for every t<1t<1 but vanishes at the target endpoint. With mt(x)=E[G∣Xt=x]m_t(x)=\mathbb E[G\mid X_t=x],

v∗(x,t)=mt(x)−x1−t,∇xmt(x)=t(1−t)2Cov⁡(G∣Xt=x).v^*(x,t)=\frac{m_t(x)-x}{1-t},\qquad \nabla_xm_t(x)=\frac{t}{(1-t)^2}\operatorname{Cov}(G\mid X_t=x).

The spatial envelopes used in the positive-noise proof can therefore diverge as t↑1t\uparrow1. For intuition, even the constant generator G=cG=c gives v∗(x,t)=(c−x)/(1−t)v^*(x,t)=(c-x)/(1-t): its ordinary spatial Lipschitz constant diverges, although the true trajectories contract to cc. A useful endpoint argument may therefore need to capture contraction more sharply than a bound based only on an absolute Lipschitz constant.

A possible extension uses early stopping. For T=1−ϵ<1T=1-\epsilon<1, coupling XTX_T and GG gives

W2(πT0,Law⁡(G))≤ϵD+1.W_2(\pi_T^0,\operatorname{Law}(G))\le\epsilon\sqrt{D+1}.

If the learned flow on [0,T][0,T] is well posed with integrable spatial envelope γx,T\gamma_{x,T} and truncated risk Rn,TR_{n,T}, the continuous-flow coupling yields

W2(π^T,Law⁡(G))≤eT/2+∫0Tγx,T(t)dtRn,T+ϵD+1.W_2(\widehat\pi_T,\operatorname{Law}(G))\le e^{T/2+\int_0^T\gamma_{x,T}(t)dt}\sqrt{R_{n,T}}+\epsilon\sqrt{D+1}.

This separates the learning error on a positive-variance interval from the stopping bias. The decomposition provides a starting point for deriving a clean-target rate by establishing the TT-dependence of the truncated learning bound and balancing it with the stopping bias.

Yakovlev and Puchkin allow zero target noise in their model and analyze scores after a positive OU smoothing cutoff, whose time convention differs from our FM endpoint. Zhou and Liu study early stopping for bounded-support targets, and Fukumizu et al. obtain Wasserstein rates with early stopping under Besov-density and interpolation-schedule assumptions. These results provide useful context for a cutoff-dependent analysis of our latent-generator setting.

References

Thank you for the clarifications. I mantain my score.

AuthorsOpenReview ↗

Thank You for Your Feedback

Thank you for your time and for considering our clarifications and revisions. We appreciate your constructive feedback throughout the review process.

Reviewer_JepKOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how many samples Flow Matching needs when data come from a low-dimensional latent variable, pass through a smooth map, and then receive Gaussian noise of a fixed level. Because of the noise, the data fill the whole ambient space instead of lying on a manifold. Lemma 2.1 splits the true velocity into an exact linear term and a bounded posterior-mean term. The authors use this split to build a ReLU velocity class with a built-in spatial Lipschitz bound.

Critical strengths

1. Clean decomposition and a useful stability tool. This work separates the exact linear part from a bounded posterior mean means the network only learns a bounded function. This makes the unbounded FM loss tractable. Proposition 3.2 uses the slow growth of the KL divergence near the source endpoint to cancel the singular time weight there. The result is a clean link from generator error to velocity error.

2. It build a complete pipeline. This paper connects approximation, estimation, and sampling error in one chain. It builds spatial regularity into the class, so one comparator serves both the risk bound and the transport argument.

Critical weaknesses

1. Contribution is limited. Previous work Yakovlev & Puchkin (2025) already proves latent-dimensional bounds for score matching under the identical noisy latent-generator model. This paper’s eq. (32) shows that the FM velocity is a fixed linear transform of the score. Key approximation tools seems come directly from Yakovlev & Puchkin (2025).

Questions for authors

Please correct me if my understanding is wrong.

  1. Which results in the paper cannot be obtained by combining the score bounds of Yakovlev & Puchkin (2025) with eq. (32)?

  2. How do the final rate, the dependence on the ambient dimension and the noise level, and the required sample size compare with those in Yakovlev & Puchkin (2025)?

4 comments
AuthorsOpenReview ↗

Response to Reviewer JepK — 1. Relation to score matching

Thank you for asking us to clarify exactly what follows from the score–velocity identity and how our results relate to Yakovlev and Puchkin (YP). This comparison helped us make the mathematical interfaces more explicit.

Changes in the revised manuscript: Section 3, Proposition 3.2, Theorem 3.1, and Appendix B.3–B.4; the subsequent learning and transport results are Theorems 4.5 and 5.2.

1. The identity and the risk it induces. We share the noisy latent-generator model and build on YP’s latent-cell Taylor and Gaussian-posterior approximation machinery. For linear FM, Eq. (32) relates the population velocity and score differences for the true and surrogate paths. A corresponding score can also be associated with a candidate velocity by transforming it consistently, and its risk carries a different time weight.

To see the weighting, let r(t)2=t2+(1−t)2r(t)^2=t^2+(1-t)^2 and set e−τ=t/r(t)e^{-\tau}=t/r(t). Then Yτ=Xt/r(t)Y_\tau=X_t/r(t) has the OU marginal starting from the same target. For consistently related candidate fields, the full-time FM risk becomes

\mathbb E\\|s_\theta(Y_\tau,\tau)-s^*(Y_\tau,\tau)\\|^2\\,d\tau.$$ The factor comes from the velocity–score coefficient, spatial rescaling, and the time Jacobian. YP Theorem 3.4 instead controls an **unweighted** score risk on a truncated OU interval. The full-time FM analysis therefore calls for control of the growing weight and both endpoint intervals. Proposition 3.2 provides the corresponding stability estimate for the latent-generator approximation. **2. The FM approximation interface.** Proposition 3.2 proves the full-time stability estimate $$\int_0^1\\|v^*-v^\circ\\|^2_{L^2(\pi_t)}dt \le\frac{\pi}{4\sigma}\\|g^*-g^\circ\\|_\infty^2.$$ The proof uses the identity above together with the time derivative of the relative entropy and a Gaussian-mixture KL bound. The $O(t^2)$ behavior of the KL divergence controls the apparent singularity at the source endpoint. This is the step that transfers the latent-generator approximation to the full-time FM velocity norm. **3. One comparator for learning and transport.** Theorem 3.1 and Appendix B.4 construct a single ReLU posterior comparator that achieves the velocity approximation error and the global spatial Lipschitz envelope required by our learning class. On an enlarged localization region, the same network controls posterior values and spatial first derivatives; a global gate preserves the spatial bound. The leading number of latent cells remains $\varepsilon^{-d}$, with the displayed ambient- and noise-dependent factors. Theorem 4.5 then analyzes FM ERM in this class, and Theorem 5.2 transports its integrated risk to the continuous-flow $W_2$ error. The posterior covariance identity and general first-order ReLU approximation tools themselves are known; the construction makes the approximation, regularity, and learning/transport requirements compatible in this model. **References** - [Yakovlev and Puchkin, 2025, *Generalization error bound for denoising score matching under relaxed manifold assumption*](https://proceedings.mlr.press/v291/yakovlev25a.html), Theorems 3.3–3.4. - [Gühring, Kutyniok and Petersen, 2020, *Error bounds for approximations with deep ReLU neural networks in $W^{s,p}$ norms*](https://arxiv.org/abs/1902.07896). - [Opschoor, Schwab and Zech, 2022, *Exponential ReLU DNN expression of holomorphic maps in high dimension*](https://doi.org/10.1007/s00365-021-09542-5).
AuthorsOpenReview ↗

Response to Reviewer JepK — 2. Rates, dimension, and noise

Thank you for requesting the quantitative comparison. We write the factors and sample conditions in the two papers’ own conventions so that the different distances and time cutoffs remain visible.

Changes in the revised manuscript: Theorem 4.5, Theorem 5.2, and Corollaries E.1–E.2.

Set s=2β+ds=2\beta+d, α=2β/s\alpha=2\beta/s, ε=n−1/s\varepsilon=n^{-1/s}, and Γε=1+D+log⁡(1/ε)+log⁡(H∨e)+log⁡(1/σ)\Gamma_\varepsilon=1+D+\log(1/\varepsilon)+\log(H\vee e)+\log(1/\sigma). Our velocity class uses the specified data-noise level σ\sigma in its coefficients at,bta_t,b_t. Under (5)–(6), our exact-average FM ERM satisfies, with probability at least 1−δ1-\delta,

R(v^)≲d,βn−α[D(1+H2)σ+D2σ−2+(σ−1+D)log⁡(en)Γε+log⁡(e/δ)carch].R(\widehat v)\lesssim_{d,\beta}n^{-\alpha}\left[\frac{D(1+H^2)}{\sigma}+D^2\\{\sigma^{-2}+(\sigma^{-1}+\sqrt D)\log(en)\\}\\{\Gamma_\varepsilon+\log(e/\delta)\\}^{c_{\rm arch}}\right].

Theorem 5.2 gives W2(π^1,π1)≤e1/2+∣γx∣1R(v^)W_2(\widehat\pi_1,\pi_1)\le e^{1/2+\\|\gamma_x\\|_1}\sqrt{R(\widehat v)}, with ∣γx∣1≲d,β1+σ−2\\|\gamma_x\\|_1\lesssim_{d,\beta}1+\sigma^{-2}. Thus for fixed model parameters the leading continuous-flow W2W_2 factor is n−β/sn^{-\beta/s} up to logarithms. The first two requirements in (5) are

n>Ds/(2β),HD,n−(β∧1)/s≤cd,βσ,n>D^{s/(2\beta)},\qquad H\sqrt D\\,n^{-(\beta\wedge1)/s}\le c_{d,\beta}\sigma,

and (6) imposes an additional localization condition.

For YP Theorem 3.4, let ν0=1−e−2t0\nu_0=1-e^{-2t_0} denote the OU noise variance at their cutoff t0t_0, and M=(d+⌊β⌋d)M=\binom{d+\lfloor\beta\rfloor}{d}. Their unweighted score risk over [t0,T][t_0,T] is bounded, up to their displayed logarithms, by

T2D12+2Mν0−1(nν0)−α.T^2D^{12+2M}\nu_0^{-1}(n\nu_0)^{-\alpha}.

Their sample-size condition (14) requires

nν0≥[Dlog⁡(nν0)ν0s]s!∨[Hd⌊β⌋D⌊β⌋!ν0]2+d/β!∨[HDν0]s,n\nu_0\ge\left[\frac{D\sqrt{\log(n\nu_0)}}{\nu_0\sqrt{s}}\right]^s\\!\vee\left[\frac{H d^{\lfloor\beta\rfloor}\sqrt D}{\lfloor\beta\rfloor!\sqrt{\nu_0}}\right]^{2+d/\beta}\\!\vee\left[\frac{H\sqrt D}{\sqrt{\nu_0}}\right]^s,

in addition to the conditions of their Theorem 3.3. YP’s Theorem 3.5 selects t0=σs/(3β+d)n−β/(3β+d)t_0=\sigma^{s/(3\beta+d)}n^{-\beta/(3\beta+d)} and T≍log⁡nT\asymp\log n and gives a reverse-SDE total variation bound, with factor D6+Mσ−1/(3β+d)n−β/(6β+2d)D^{6+M}\sigma^{-1/(3\beta+d)}n^{-\beta/(6\beta+2d)} up to its stated logarithms. Together, the two results describe complementary guarantees: YP analyze a reverse SDE in TV with a positive cutoff, while we analyze a continuous FM ODE in W2W_2 over the full interpolation interval. The different metrics, cutoffs, and regularity conditions are reflected in their respective bounds. Appendix E gives explicit sufficient regimes for the complete dimension- and noise-dependent FM bound.

References

Thank you for the rebuttal and detailed explanation. I think all of my concerns have been addressed. I have increased my score.

AuthorsOpenReview ↗

Thank You for Your Reconsideration

Thank you for taking the time to reconsider our work and for your constructive feedback. We are glad that our clarifications and revisions have addressed your concerns. We sincerely appreciate your careful review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the Area Chair

We thank the reviewers for their thoughtful assessments and constructive suggestions. We appreciate their recognition of two aspects of the paper: the extension of intrinsic-dimensional analysis beyond exact low-dimensional support, and the use of a common neural class for velocity estimation and continuous-flow transport.

Our main result studies linear Flow Matching when X1=g∗(U)+σξX_1=g^*(U)+\sigma\xi with σ>0\sigma>0. Although the target has full ambient support, Theorems 3.1 and 4.5 give an integrated velocity-risk bound with leading sample factor n−2β/(2β+d)n^{-2\beta/(2\beta+d)}; Theorem 5.2 propagates that error to the learned continuous flow in W2W_2. The complete bounds retain the dependence on ambient dimension, noise, smoothness, and confidence.

The revised manuscript addresses the reviewers’ questions as follows.

  1. Relation to score matching and spatial control — Section 3 and Appendix B. We identify the posterior approximation tools inherited from Yakovlev and Puchkin, explain the score–velocity identity and the full-time generator-to-velocity estimate, and describe the spatially regular comparator used in the constrained FM estimator and transport analysis.
  2. Dimension and noise — Corollaries E.1–E.2. Two illustrative sufficient regimes verify the scale conditions and give convergent risk and continuous-flow bounds as dimension grows, including one example with positive noise decreasing to zero.
  3. Finite auxiliary sampling — Proposition F.1 and Appendix F. Following the shared-auxiliary scheme of Zhou and Liu, we give an explicit sampling error and combine it with the target-sample bound. A sufficient auxiliary budget retains the leading velocity-risk exponent.
  4. Discretization and the zero-noise endpoint — Proposition 5.5 and Appendix G. Independent randomized evaluation times give a finite-step W2W_2 bound for the same spatially constrained estimator without an additional learned-field time-regularity assumption. Appendix G.3 describes an early-stopping decomposition at zero noise.
  5. Controlled image illustration — Appendix H, Table 1. In a two-dimensional latent image family, the held-out velocity risk changes from 0.016290.01629 to 0.017610.01761 as DD increases sixteenfold at fixed n=8192n=8192. Results across three resolutions, three sample sizes, and two training budgets complement the theory with a practical time-conditioned U-Net example.

We hope these additions clarify the contribution and its mathematical and empirical scope. Full proofs and the anonymous reproducibility code accompany the revision.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.