acceptodds
Under review as a conference paper at ICLR 2027

Counterfactual Generation via Flow Matching: Coupling-Sensitive End-to-End Rates

Abstract

Counterfactual generation seeks to sample outcomes under a hypothetical intervention or decision using observational data collected under the factual assignment mechanism. We develop a flow-matching approach that combines a sample-split, doubly robust training objective with a learned coupling between observed source outcomes and target outcomes drawn from a fitted conditional outcome model. To enable finite-step generation, we introduce a score-corrected stochastic sampler based on a Gaussian-smoothed interpolation. Our main theoretical contribution is a coupling-sensitive KL bound for constant-step Euler discretization: the error is controlled by moments of the source–target displacement under the chosen coupling, rather than by global uniform regularity of the velocity field, and has near-linear dependence on the ambient dimension. We also establish finite-sample non-parametric guarantees for the learned velocity and score fields when both the conditional outcome model and the source–target coupling are estimated from data. These bounds separate approximation, coupling-replacement, nuisance-estimation, generalization, and Monte Carlo errors and, combined with the sampler analysis, yield an end-to-end guarantee for counterfactual generation. Experiments on synthetic and semi-synthetic image benchmarks support the coupling-dependent theory and show that, at finite discretization budgets, the stochastic sampler can outperform the corresponding deterministic ODE sampler.

Primary area:
learning theory
Keywords:
Flow Matching, Counterfactual Generation, Optimal transport, Stochastic Analysis, Generative modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UFjoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies counterfactual generation from observational data using flow matching. Building on doubly robust training and learned source–target couplings, this work investigate how coupling quality affects finite-step sampling errors. Through two counterexamples, they show that small coupling displacement alone cannot guarantee small KL error for deterministic Euler sampling or sufficient score regularity without smoothing. To address these issues, they introduce Gaussian-smoothed interpolation and a score-corrected SDE sampler, establishing a coupling-sensitive KL bound controlled by the fifth displacement moment D5D_5. They further derive finite-sample guarantees for learned velocity and score fields, leading to an end-to-end generation guarantee.

Critical strengths
  1. The paper is well motivated, with a clearly formulated research problem and a convincing discussion of the limitations of existing approaches.

  2. The proposed method is introduced naturally, with two counterexamples providing clear theoretical motivation for Gaussian smoothing and score-corrected stochastic sampling.

  3. The theoretical analysis is comprehensive, providing coupling-sensitive sampling bounds and end-to-end guarantees that account for different sources of estimation error.

Critical weaknesses

W1. Limited experiments of D5D_5 in high-dimensional settings. D5D_5 may grow with dimensionality and its empirical estimation may be sensitive to large displacements. Although Figure 1(c) validates the relationship between D5D_5 and KL error in a two-dimensional setting, further analysis and experiments in high-dimensional learned settings are needed to understand how D5D_5 behaves and whether it remains predictive of the final generation error.

W2. Insufficient evaluation across different sampling budgets. The high-dimensional image experiments use a fixed number of 50 Euler steps. I think a systematic comparison across different sampling steps NN in high-dimensional settings would better demonstrate the proposed method.

W3. Minor presentation issues. Figure 1-3 contain relatively small axis labels, legends, and annotations, making them difficult to read.

Questions for authors

As mentioned in W1 and W2, my main concern is whether the proposed coupling-sensitive theoretical advantages translate into practical benefits in high-dimensional settings, particularly regarding the behavior of D5D_5 and generation error across different sampling step budgets. I would be willing to raise my score if the authors could address these concerns.

3 comments

Thank you for acknowledging our theoretical contribution and valuable suggestions.

sampling budgets (W2, Q1)

  • Regarding sampling budgets, on Fashion-MNIST Shoe, we evaluate ODE and SDE sampling using the same learned EOT velocity checkpoint across N=4,8,16,32,64N=4,8,16,32,64 Euler steps (Figure 7, left). As NN increases from 44 to 6464, the clean-target SW2 decreases from 0.01540.0154 to 0.01310.0131 for ODE and from 0.01560.0156 to 0.01220.0122 for SDE. The SDE achieves lower SW2 for N≥8N\geq 8, demonstrating its practical benefit across different sampling budgets.

Experiments on D5D_5 (W1, Q1)

  • For the image experiments, we added Table 1, where we observe that EOT coupling, which has a smaller estimated D5D_5 than IID coupling, achieves lower SW2 under ODE sampling. Our proposed SDE sampler further improves performance and achieves the best result among the compared methods.

  • However, conducting an experiment similar to Figure 1(c) on images is challenging. Different coupling choices lead to different values of D5D_5 but require separately trained velocity and score fields. Since their field-estimation errors may differ, it is difficult to isolate the effect of coupling displacement on sampling error. Moreover, accurately estimating D5D_5 in high-dimensional image settings can be challenging.

  • To further investigate this relationship, we constructed a higher-dimensional synthetic example with d=32d=32, where we can conduct an experiment similar to Figure 1(c). See our updated Figure 7 (right). We use exact velocity and score fields to eliminate field-estimation error. We observe that larger D5D_5 is associated with substantially greater sampling error in the high-displacement regime, while our proposed SDE sampler achieves considerably lower error than the ODE sampler in this regime.

  • In summary, we would like to emphasize that both our image experiments and higher-dimensional synthetic experiment show consistent trends, supporting the practical relevance of our coupling-sensitive theoretical results.

Figure readibility (W3)

  • We enlarged the axis labels, legends, and annotations in Figures 1-3

Contribution type

  • We emphasize that our main contributions are theoretical. Our counterexamples show that coupling displacement alone does not control the regularity needed to establish coupling-dependent sampling guarantees for deterministic sampling or without smoothing. We therefore study a Gaussian-smoothed stochastic sampler, where we prove that the required regularity can be controlled by coupling displacement, leading to a coupling-sensitive sampling KL error bound with nearly linear dimension dependence. In practice, Pooladian et al. [4] show that minibatch couplings improve ImageNet sample quality at low numbers of function evaluations (Section 6.2, Fig. 3 and Table 1). Tong et al. [5] similarly find that OT-CFM improves CIFAR-10 sample quality at small sampling budgets (Section 5.3, Fig. 3). These empirical findings, together with the motivation for learned couplings in Dance et al. [1], motivate our theoretical analysis of how coupling choice affects sampling accuracy.

  • Another key challenge is that the learned coupling depends on an imperfect prediction oracle, which introduces additional errors when training with the true target distribution. We address this by constructing an ideal coupling with the correct source and target distributions and controlling the error between these two couplings. Proposition 1 establishes finite-sample velocity and score estimation bounds involving this coupling-replacement error, the doubly robust nuisance-product term, and neural approximation and statistical estimation errors. Finally, combining these learning guarantees with our sampling analysis and smoothing error yields an end-to-end guarantee for the original counterfactual target.

[1] Hugh Dance, Johnny Xi, Peter Orbanz, and Benjamin Bloem-Reddy. Debiased Counterfactual Generation via Flow Matching from Observations. 2026. https://arxiv.org/abs/2605.07665

[4] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky T. Q. Chen. Multisample Flow Matching: Straightening Flows with Mini- batch Couplings. ICML, 2023. https://proceedings.mlr.press/v202/pooladian23a.html.

[5] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and Generalizing Flow-Based Gener- ative Models with Minibatch Optimal Transport. TMLR, 2024. https://arxiv.org/abs/2302.00482

We added experiments testing different sampling budgets (W2, Q1) and performed experiments on coupling displacement $D_5 $(W1, Q1). We also enlarged the axis labels, legends, and annotations in Figures (W3).

Given the impending discussion deadline tomorrow, could you please let us know whether these revisions address your concerns?

Thank you for your detailed responses. The additional experiments provide further empirical support for the coupling-sensitive analysis. I appreciate the authors' efforts and I am willing to raise my score.

Reviewer_v4BQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for counterfactual distribution generation using doubly robust flow matching, learned optimal-transport couplings, and a score-corrected stochastic sampler. Its main theoretical contribution is a coupling-sensitive KL discretization bound with near-linear dimension dependence, complemented by finite-sample guarantees for the learned fields and an end-to-end Wasserstein bound. Experiments on synthetic and semi-synthetic image datasets investigate the influence of coupling quality and stochastic sampling.

Critical strengths
  • Interesting theoretical contribution: The coupling-sensitive discretization bound provides a connection between source–target displacement and finite-step sampling error, without relying on global uniform regularity assumptions.

  • Promising empirical findings: The experiments suggest that stochastic sampling can improve generation quality over deterministic ODE sampling at fixed discretization budgets.

Critical weaknesses
  • Related work. missing baselines, and methodological novelty: The doubly robust generative training objective closely resembles existing approaches, particularly Normalizing Flows for Interventional Density Estimation and GDR-learners (Melnychuk et al., 2023, 2026), neither of which is discussed or compared against. Related work on frugal parameterization and generative modeling is also missing. Moreover, Dance et al. (2026) already develop doubly robust flow matching with learned OT couplings. Thus, the methodological novelty appears limited and existing approaches should be discussed and compared against.

  • Limited double robustness: Although the field-training objectives are doubly robust, the coupling is estimated using the conditional outcome model. Consequently, the end-to-end error contains prediction-oracle terms that do not vanish even with a correctly specified propensity model.

  • Theory–implementation gap: The finite-sample analysis assumes bounded outcomes, clipped network classes, and exact empirical risk minimization. It is unclear how closely these assumptions match the implemented training procedure. The signed doubly robust loss may also create optimization stability issues.

Questions for authors
  • How does the proposed doubly robust training objective differ fundamentally from the general orthogonal generative learning frameworks in Melnychuk et al. (2023, 2026)? What is the statistical novelty beyond applying these principles to flow matching?
4 comments

Thank you for recognizing the sampling result and for identifying related work.

Novelty (Q1)

  • Our primary contribution is the theoretical analysis of coupling-dependent learning and finite-step sampling. This positioning is already stated in the abstract (lines 019--028), and the contributions (lines 101--116). The submitted introduction also explicitly credits Dance et al. [1] for debiased flow matching with learned OT (lines 074--076).

  • Our counterexamples show that coupling displacement alone does not control the regularity needed to establish coupling-dependent sampling guarantees for deterministic sampling or without smoothing. We therefore study a Gaussian-smoothed stochastic sampler, where we prove that the required regularity can be controlled by coupling displacement, leading to a coupling-sensitive sampling KL error bound with nearly linear dimension dependence. In practice, Pooladian et al. [4] show that minibatch couplings improve ImageNet sample quality at low numbers of function evaluations (Section 6.2, Fig. 3 and Table 1). Tong et al. [5] similarly find that OT-CFM improves CIFAR-10 sample quality at small sampling budgets (Section 5.3, Fig. 3). These empirical findings, together with the motivation for learned couplings in Dance et al. [1], motivate our theoretical analysis of how coupling choice affects sampling accuracy.

  • Another key challenge is that the learned coupling depends on an imperfect prediction oracle, which introduces additional errors when training with the true target distribution. We address this by constructing an ideal coupling with the correct source and target distributions and controlling the error between these two couplings. Proposition 1 establishes finite-sample velocity and score estimation bounds involving this coupling-replacement error, the doubly robust nuisance-product term, and neural approximation and statistical estimation errors. Finally, combining these learning guarantees with our sampling analysis and smoothing error yields an end-to-end guarantee for the original counterfactual target.

  • These contributions are explicitly stated under ``Technical Novelty'' in the appendix (page 20, lines 1042--1057).

Related works (W1)

  • We added citations (Melnychuk et al., 2023, 2026), see lines 66-69.

  • Flow matching has become a popular alternative to traditional normalizing-flow training, offering simpler and more efficient training with strong generation performance [4] [5] [15]. Dance et al. [1] also demonstrate improvements over interventional normalizing flows [2]. Building on their framework, we provide theoretical guarantees for learning with estimated couplings, coupling-dependent sampling, and end-to-end generation. Our method follows the counterfactual flow-matching framework of Dance et al. [1] and adds a score-corrected stochastic sampler and the associated theoretical analysis. Its field-training objective uses a plug-in loss plus an inverse-propensity-weighted observed-minus-predicted residual (p.4, lines 180--190).

  • We added Table 1, where we compared against existing approaches on Fashion-MNIST shoes. We added an adapted GDR diffusion baseline using the same conditional prediction oracle and propensity estimator. It achieves an SW2 of 0.07930.0793, compared with 0.01600.0160 for our DR flow matching with IID coupling and 0.01220.0122 for our EOT SDE, despite using 300 sampling steps instead of 50. We remark that the iid coupling result can be viewed as applying the same doubly robust principle as GDR to flow matching.

Doubly Robustness (W2)

  • Under IID coupling, our estimator retains the standard doubly robust guarantee: the population field-loss bias vanishes when either nuisance model is correct. For learned OT coupling, Proposition 1 explicitly accounts for an additional prediction-oracle-dependent coupling-replacement error, which is not removed by the DR correction. This reflects a trade-off: IID coupling provides stronger robustness guarantees, while learned OT can yield shorter, easier-to-learn transport paths and better empirical accuracy, as also observed by [1]. We revised the submission, added Remark 2 (page 8) to make it more explicit.

  • On the other hand, the prediction-oracle error εpred,train\varepsilon_{\mathrm{pred,train}} is expected to vanish as n→∞n\to\infty when the conditional outcome model is consistently estimated. Under consistent nuisance estimation, εpred,train,εpred,gen,εα→0\varepsilon_{\mathrm{pred,train}},\varepsilon_{\mathrm{pred,gen}},\varepsilon_\alpha\to0, while the explicit statistical terms in Proposition~1 also vanish. Thus, for fixed dd, σ\sigma, and εOT\varepsilon_{\mathrm{OT}}, the overall field-estimation error bound converges to zero.

Theory–implementation gap (W3)

Our theory bounds the estimated inverse-propensity weights by 0≤α^(c)≤2/πmin⁡0\leq\widehat\alpha(c)\leq2/\pi_{\min} (Assumption~2). In ColorMNIST, we clip estimated propensities to [0.04,0.96][0.04,0.96], so the inverse-propensity weights are at most 2525. This limits the amplification of the signed correction by large weights.

Our sampling bound depends on the velocity and score estimation errors in Assumption 1 and does not require a specific network architecture. Our sparse-ReLU analysis provides one concrete finite-sample verification of these errors under the stated statistical assumptions, and the paper explicitly connects these learning bounds to the sampling assumption under ``Verification of Assumption 1'' (p. 7, lines 374--377).

Using analytically tractable network classes to establish approximation and generalization rates is standard in the theoretical literature on diffusion and flow-based generative models. For example, existing works use specific ReLU or GELU network classes to establish approximation and statistical guarantees [6, 7, 8, 9, 10, 11, 12, 13]. These theoretical parameterizations are generally simpler than the architectures used in large-scale image generation.

For the image experiments, we use convolutional U-Nets, as is standard in practical diffusion and flow-matching implementations [14, 15]. These experiments evaluate the sampling procedure with practical learned fields; we do not claim that the explicit sparse-ReLU finite-sample rates are proved for the experimental U-Nets.

[1] Hugh Dance, Johnny Xi, Peter Orbanz, and Benjamin Bloem-Reddy. Debiased Counterfactual Generation via Flow Matching from Observations. 2026. https://arxiv.org/abs/2605.07665.

[2] Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Normalizing Flows for Interventional Density Estimation. ICML, 2023. https://proceedings.mlr.press/v202/melnychuk23a.html.

[3] Valentyn Melnychuk and Stefan Feuerriegel. GDR-Learners: Orthogonal Learning of Generative Models for Potential Outcomes. ICLR, 2026.

[4] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky T. Q. Chen. Multisample Flow Matching: Straightening Flows with Minibatch Couplings. ICML, 2023. https://proceedings.mlr.press/v202/pooladian23a.html.

[5] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR, 2024. https://arxiv.org/abs/2302.00482.

[6] Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion Models are Minimax Optimal Distribution Estimators. ICML, 2023. https://arxiv.org/abs/2303.01861.

[7] Minshuo Chen, Kaixuan Huang, Tuo Zhao, and Mengdi Wang. Score Approximation, Estimation and Distribution Recovery of Diffusion Models on Low-Dimensional Data. ICML, 2023. https://arxiv.org/abs/2302.07194.

[8] Yinbin Han, Meisam Razaviyayn, and Renyuan Xu. Neural Network-Based Score Estimation in Diffusion Models: Optimization and Generalization. ICLR, 2024. https://arxiv.org/abs/2401.15604.

[9] Zhengyu Zhou and Weiwei Liu. An Error Analysis of Flow Matching for Deep Generative Modeling. ICML, 2025. https://proceedings.mlr.press/v267/zhou25l.html.

[10] Yihan He, Qishuo Yin, Yuan Cao, Jianqing Fan, and Han Liu. A Theory on Flow Matching with Neural Networks. 2026. https://arxiv.org/abs/2606.10089.

[11] Konstantin Yakovlev and Nikita Puchkin. Generalization error bound for denoising score matching under relaxed manifold assumption. COLT, 2025. https://proceedings.mlr.press/v291/yakovlev25a.html.

[12] Konstantin Yakovlev, Anna Markovich, and Nikita Puchkin. Implicit score matching meets denoising score matching: improved rates of convergence and log-density Hessian estimation. arXiv preprint arXiv:2512.24378, 2025. https://arxiv.org/abs/2512.24378.

[13] Konstantin Yakovlev and Nikita Puchkin. Simultaneous approximation of the score function and its derivatives by deep neural networks. Journal of Complexity, 97:102070, 2026. https://doi.org/10.1016/j.jco.2026.102070.

[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. NeurIPS, 2020. https://arxiv.org/abs/2006.11239.

[15] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow Matching for Generative Modeling. ICLR, 2023. https://arxiv.org/abs/2210.02747.

Thank you for the detailed rebuttal and revisions. I particularly appreciate the clearer positioning of the theoretical contributions, the expanded discussion of related work, and the implementation of additional baselines, including GDR-learners.

Some concerns remain, particularly regarding the limited double robustness under learned OT couplings, the surprisingly weak performance of the GDR baseline and whether this comparison is sufficiently well justified, and the practical optimization stability of the signed doubly robust loss.

Nevertheless, I appreciate the authors' efforts to address my feedback. I believe the paper has improved, and I will increase my score accordingly.

Thank you for acknowledging our clarifications and for your helpful feedback.

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

Generating interventional outcome distributions from observational data under certain assumptions is of great importance in modern ML, especially in areas where counterfactual outcomes are difficult to observe, such as healthcare. This paper first presents counterexamples showing limitations of deterministic flow-matching samplers and the necessity of smoothing for its analysis. The authors then develop a theoretical framework for coupling-sensitive sampling error bounds, combining doubly robust training with a score-corrected stochastic sampler. The paper provides a substantial theoretical analysis.

Critical strengths
  1. The counter examples motivate the use of smoothing and stochastic sampling and show previous deterministic Euler sampling can still have infinite KL error, and removing Gaussian smoothing can lead to unbounded score derivatives.

  2. The paper provides a solid theoretical framework with rigorous counterexamples and coupling-sensitive discretization bounds. The analysis is well motivated and provides a clear justification for the proposed stochastic sampler.

  3. The paper is easy to follow despite being theoretically dense.

Critical weaknesses
  1. Though theretical solid, in the empricial verfication, the method first requires training a full conditional generative model, which is potentially a harder problem than learning the target marginal distribution itself. Since this model can already generate counterfactual samples directly, why do we still need to use additional EOT and flow-matching steps provide enough benefit to justify the added complexity is unclear to me.

  2. The experiments are relatively simple, focusing mainly on synthetic data and semi-synthetic MNIST benchmarks. While they support the theoretical findings, more challenging and realistic datasets would help further demonstrate the practical value of the proposed method.

Questions for authors
  1. The main theoretical contribution appears to be a general result on coupling-sensitive flow matching, which does not rely on the counterfactual setting. Any specific reason why the paper focuses on counterfactual generation setting?

  2. Can the authors provide an ablation comparing doubly robust training with standard flow matching, especially when the conditional prediction oracle is inaccurate?

5 comments

Thank you for recognizing the theoretical contribution.

Additional experiments (W2 and Q2)

  • We added comparisons with direct prediction-oracle sampling (i.e., using conditional generative model) and standard FM without doubly robust correction. On Fashion-MNIST shoes (see our newly added Table 1), bias correction improves standard FM (under iid coupling) from 0.0207420.020742 to 0.0159660.015966 in SW2, while EOT coupling and stochastic sampling further improve performance to 0.0122440.012244, outperforming direct oracle sampling (0.0143370.014337).

  • We further evaluate on CelebA (see our newly added Table 2) with 64×6464\times64 RGB images (d=12,288d=12{,}288), where our score-based SDE with EOT coupling achieves an SW2 of 0.0306150.030615, compared with 0.0380180.038018 for direct conditional-flow sampling and 0.0330010.033001 for IID coupling ODE.

  • These experiments demonstrate the benefits of incorporating bias correction into flow matching, as well as the advantages of our proposed SDE sampler.

Cost of conditional flow model (W1)

  • We also remark that the conditional flow model requires additional training, but conditioning may make the transformation easier to learn. In ColorMNIST, for example, the conditional model is given the desired color, which provides explicit information about how to change the color of an input image. The final unconditional model receives no desired color and must learn how to transform input images so that the generated colors follow the correct population distribution, which may be more challenging.

Why focus on counterfactual generation? (Q1)

  • We totally agree with the reviewer that the sampling theory applies more broadly. This broadens its usefulness to other generation problems using the same smoothed interpolation and stochastic sampler.
  • We highlighted in page 3 (line 112) that our sampling theorem can be applied to general flow matching.
  • Counterfactual generation provides a concrete motivating application to illustrate our theory, following [1]. Indeed, observed outcomes provide a natural source because they share conditional outcome laws with the counterfactual target (lines 036--060). Hence, we present our reuslts in this setting and highlight how the coupling choice in this framework affects learning and finite-step generation.

[1] Hugh Dance, Johnny Xi, Peter Orbanz, and Benjamin Bloem-Reddy. Debiased Counterfactual Generation via Flow Matching from Observations. 2026. https://arxiv.org/abs/2605.07665.

Why a conditional prediction oracle alone may be insufficient. (W1)

We illustrate how DR flow matching can improve upon directly sampling an imperfect conditional prediction oracle.

Take a=1a=1, C∈{0,1}C\in\{0,1\}, and PC({0})=PC({1})=0.5P_C(\{0\})=P_C(\{1\})=0.5. Let Gj=N(j,0.01)G_j=\mathcal N(j,0.01) for j=0,1j=0,1. Consider \begin{align*} P(A=1\mid C=0)&=0.8,& P_{Y\mid C=0,A=1}&=0.8G_0+0.2G_1,\\ P(A=1\mid C=1)&=0.2,& P_{Y\mid C=1,A=1}&=0.2G_0+0.8G_1. \end{align*}

The observed source and counterfactual target distributions are \begin{align*} P_{Y\mid A=1}&=\int \mathbb P_{Y\mid C=c,A=1} d\mathbb P_{C\mid A=1}(c)=0.68G_0+0.32G_1,\\ \int P_{Y\mid C=c,A=1},dP_C(c)&=0.5G_0+0.5G_1. \end{align*} Suppose the conditional prediction oracle is accurate at C=0C=0 but gives P^Y∣C=1,A=1=0.4G0+0.6G1\widehat P_{Y\mid C=1,A=1}=0.4G_0+0.6G_1. Direct oracle sampling then produces \begin{align*} \int\widehat P_{Y\mid C=c,A=1},dP_C(c) =0.5(0.8G_0+0.2G_1)+0.5(0.4G_0+0.6G_1)=0.6G_0+0.4G_1, \end{align*} whose mean differs from the true target mean by 0.10.1.

Now consider population flow matching with an IID coupling, Y0∼PY∣A=1Y_0\sim P_{Y\mid A=1} independent of Y1Y_1, and the interpolation St=Y0+t(Y1−Y0)S_t=Y_0+t(Y_1-Y_0). To isolate the DR correction, we omit Gaussian smoothing in this example. For a given target endpoint distribution, the ordinary FM objective is \begin{align*} L_{\mathrm{FM}}(v) =\mathbb E\left[(v(S_t,t)-(Y_1-Y_0))^2\right]. \end{align*} The plug-in objective samples Y1Y_1 from the conditional prediction oracle. DR objective is \begin{align*} \mathbb E\left[ \widehat\ell_i^{\mathrm{aux}}(v) +\frac{\mathbf 1{A_i=1}}{\widehat P(A=1\mid C_i)} \left(\widehat\ell_i^{\mathrm{obs}}(v)-\widehat\ell_i^{\mathrm{aux}}(v)\right) \right]. \end{align*}

For fixed vv, the FM loss is linear in the target endpoint distribution under an IID coupling, since Y0Y_0 and tt are sampled independently of the target endpoint. Thus, the expected plug-in loss is

\begin{align*} \mathbb{E}[\widehat\ell_i^{\mathrm{aux}}(v)]&=\int\mathbb E_{Y_0,t}[ (v(Y_0+t(y_1-Y_0),t)-(y_1-Y_0) )^2]\cdot [0.5(0.8G_0+0.2G_1)+0.5(0.4G_0+0.6G_1) ] (d y_{1}) \\ &=\int\mathbb E_{Y_{0},t} [ (v(Y_{0}+t(y_{1} - Y_{0}),t) - (y_{1}-Y_{0} ) )^2 ] (0.6G_0+0.4G_1)(d y_{1}). \end{align*}

To calculate the DR correction, condition on C=cC=c. Since the oracle samples are independent of treatment assignment conditional on CC, we obtain \begin{align*} &\mathbb E\left[ \frac{\mathbf 1{A=1}}{\widehat P(A=1\mid C)} \left(\widehat\ell^{\mathrm{obs}}(v)-\widehat\ell^{\mathrm{aux}}(v)\right) \middle| C=c\right]\\ &=\frac{P(A=1\mid C=c)}{\widehat P(A=1\mid C=c)} \left( \mathbb E[\widehat\ell^{\mathrm{obs}}(v)\mid C=c,A=1] -\mathbb E[\widehat\ell^{\mathrm{aux}}(v)\mid C=c] \right). \end{align*} The contribution from C=0C=0 is zero because the conditional oracle is exact. At C=1C=1, the observed outcome follows 0.2G0+0.8G10.2G_0+0.8G_1 when A=1A=1, while the oracle-generated outcome follows 0.4G0+0.6G10.4G_0+0.6G_1 regardless of the observed treatment assignment. Substituting these distributions into the corresponding FM losses gives \begin{align*} &\mathbb E[\widehat\ell_i^{\mathrm{obs}}(v)\mid C_i=1,A_i=1] -\mathbb E[\widehat\ell_i^{\mathrm{aux}}(v)\mid C_i=1]\\ &=\int\mathbb E_{Y_0,t}\left[ \left(v(Y_0+t(y_1-Y_0),t)-(y_1-Y_0)\right)^2 \right]\cdot\left[(0.2G_0+0.8G_1)-(0.4G_0+0.6G_1)\right] (dy_1)\\ &=\int\mathbb E_{Y_0,t}\left[ \left(v(Y_0+t(y_1-Y_0),t)-(y_1-Y_0)\right)^2 \right] (-0.2G_0+0.2G_1) (dy_1). \end{align*} The indicator 1\mathbf1 {Ai=1A_i=1} contributes the true propensity P(A=1∣C=1)=0.2P(A=1\mid C=1)=0.2 in expectation, while the inverse-propensity weight is 1/0.251/0.25. Since PC({1})=0.5P_C(\{1\})=0.5, the complete population DR objective becomes \begin{align*} &\mathbb E\left[ \widehat\ell_i^{\mathrm{aux}}(v) +\frac{\mathbf1{A_i=1}}{\widehat P(A=1\mid C_i)} \left( \widehat\ell_i^{\mathrm{obs}}(v)-\widehat\ell_i^{\mathrm{aux}}(v) \right) \right]\\ &=\int\mathbb E_{Y_0,t}\left[ \left(v(Y_0+t(y_1-Y_0),t)-(y_1-Y_0)\right)^2 \right]\cdot\left[0.6G_0+0.4G_1+\frac{0.5\times0.2}{0.25}(-0.2G_0+0.2G_1)\right] (dy_1) \\ &=\int\mathbb E_{Y_0,t}\left[ \left(v(Y_0+t(y_1-Y_0),t)-(y_1-Y_0) \right)^2 \right] (0.52G_0+0.48G_1) (dy_1). \end{align*} This is exactly the ordinary population FM objective with target distribution 0.52G0+0.48G10.52G_0+0.48G_1, and the equality holds for every vv. Thus, although neither nuisance model is exact at C=1C=1, DR training changes the population-optimal FM target from 0.6G0+0.4G10.6G_0+0.4G_1 to 0.52G0+0.48G10.52G_0+0.48G_1, closer to the true target 0.5G0+0.5G10.5G_0+0.5G_1.

Contribution type

  • We emphasize that our main contributions are theoretical. Our counterexamples show that coupling displacement alone does not control the regularity needed to establish coupling-dependent sampling guarantees for deterministic sampling or without smoothing. We therefore study a Gaussian-smoothed stochastic sampler, where we prove that the required regularity can be controlled by coupling displacement, leading to a coupling-sensitive sampling KL error bound with nearly linear dimension dependence. In practice, Pooladian et al. [4] show that minibatch couplings improve ImageNet sample quality at low numbers of function evaluations (Section 6.2, Fig. 3 and Table 1). Tong et al. [5] similarly find that OT-CFM improves CIFAR-10 sample quality at small sampling budgets (Section 5.3, Fig. 3). These empirical findings, together with the motivation for learned couplings in Dance et al. [1], motivate our theoretical analysis of how coupling choice affects sampling accuracy.

  • Another key challenge is that the learned coupling depends on an imperfect prediction oracle, which introduces additional errors when training with the true target distribution. We address this by constructing an ideal coupling with the correct source and target distributions and controlling the error between these two couplings. Proposition 1 establishes finite-sample velocity and score estimation bounds involving this coupling-replacement error, the doubly robust nuisance-product term, and neural approximation and statistical estimation errors. Finally, combining these learning guarantees with our sampling analysis and smoothing error yields an end-to-end guarantee for the original counterfactual target.

[4] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky T. Q. Chen. Multisample Flow Matching: Straightening Flows with Minibatch Couplings. ICML, 2023. https://proceedings.mlr.press/v202/pooladian23a.html.

[5] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR, 2024. https://arxiv.org/abs/2302.00482

Comments on the paper

1 comment

Dear Reviewers,

We updated the manuscript with additional experiments on CelebA (64×64)(64 \times 64) comparing IID and EOT couplings under both ODE and SDE sampling (Table 2). EOT achieves lower SW2 than IID, and SDE sampling further reduces SW2 relative to ODE sampling.

We hope our responses and revision address your concerns. Please let us know if any concerns remain or if further clarification would be helpful.

Thanks for your time and helpful feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.