Many causal inference problems require sampling from interventional outcome distributions rather than estimating low-dimensional summaries. Existing approaches often rely on fitting generative models of the target distribution together with additional nuisance components, which can be computationally demanding for high-dimensional outcomes. We propose Interventional Maximum Mean Discrepancy Flow (IMMD-Flow), a particle-based method that instead represents the target interventional distribution through its kernel mean embedding. This kernel embedding admits a doubly robust, Neyman-orthogonal estimator with finite-sample guarantees and can be efficiently reused to sample from the interventional distribution via MMD gradient flow. IMMD-Flow accommodates both hard and soft interventions, admits theoretical guarantees for estimation and particle approximation, and extends to high-dimensional image outcomes through frozen encoders. Across image-based benchmarks, IMMD-Flow matches or improves upon state-of-the-art generative approaches while showing significant speed advantages.
Figures & tables
Figure 1: ColorMNIST DR–IMMD at n=20,000 and M=2,000 . Left: two illustrative particles initialized from PY∣A=a ; their foreground colors change during optimization while their digit identities are preserved. Right: the empirical foreground-color laws move away from the confounded observational arms and toward the oracle interventional laws. The displayed particle paths are algorithmic trajectories, not identified individual-level counterfactuals. Additional examples appear in Appendix E.5 .
Benchmark
Method
Fitting (s)
Sampling (s)
Total (s)
× DR–IMMD
ColorMNIST
DR–IMMD (ours)
2.8±0.7
35.6±0.0
38.4
1.0
DecFM Plugin
422.6±108.1
0.59±0.00
423.2
11.0
DecFM DR
1560.0±105.1
0.77±0.04
1560.8
40.6
DecFM DR–OT
1621.0±108.7
0.77±0.04
1621.8
42.2
CelebA
DR–IMMD (ours)
1.8±0.6
698.7±0.2
700.5
1.0
DecFM Plugin
8823.6±216.0
25.98±0.22
8849.5
12.6
Table 2: End-to-end wall-clock comparison at n=20,000 and M=2,000 . Within each benchmark, all methods were run on the same hardware: an NVIDIA L40S for ColorMNIST and an NVIDIA A100-SXM4-40GB for CelebA. Each entry covers both intervention arms. Fitting and sampling times are means ± sample standard deviations over three data seeds; totals are sums of the reported means. The final column reports total time relative to DR–IMMD.
Figure 2: Decoded CelebA particles during DR–IMMD optimization at n=20,000 and M=2,000 . Particles are initialized from the corresponding observational arm PY∣A=a , and each row follows one numerical particle across optimization iterations as the empirical particle law is transported toward the interventional law PY(a) . Note that these are not identified individual-level counterfactual trajectories, but rather represent the distributional change needed for these particles to represent valid samples from the interventional distribution. Additional particle trajectories appear in Appendix 5 .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Full ridge
Nyström
Kernel storage
O(n2)
O(nmL+mL2)
One-time regression solve
O(n3)
O(nmL2+mL3)
Finite- J target per intervention
O((J+1)n2)
O((J+1)nmL+mL2)
Target force per particle iteration
O(Mn)
O(Mn)
Appendix
Table 3: Computational comparison of the full-ridge and input-projected Nyström target constructions. The per-intervention costs assume that the corresponding regression system has already been factorized.
Figure 3: Initialized and final ColorMNIST particles. Within each pair of rows, columns correspond to the same numerical particles before and after optimization. The number displayed beside each row is the empirical mean of the foreground-colour statistic R/(R+B) ; the oracle mean is 0.501 . IMMD-Flow moves the colour composition toward the interventional target while preserving the digit class.
Figure 4: Example CelebA particles before and after IMMD-Flow for the female and male groups.
Figure 5: Evolution of selected CelebA particles during IMMD-Flow. The top and bottom panels correspond to the female ( a=0 ) and male ( a=1 ) groups, respectively.
Figure 6: Initial and flowed particles in the two-dimensional benchmark of Dance et al. [15] . Rows correspond to the intervention arms a∈{0,1} . Gray points show a fresh oracle draw from the common interventional law PY(a)=PY ; red points are particles initialized from PY∣A=a , and blue points are returned by DR–IMMD with the RBF–Euler update. We use n=20,000 logged observations and M=2,000 particles.
Figure 7: Paths of 80 representative particles for each intervention arm in the run from Figure 6 over 3,000 RBF–Euler updates. Path color indicates the iteration; open red markers and black markers denote initial and returned particles, respectively. Gray points show a fresh oracle draw from PY .
103×SW2
w
initial
flow
1
10.89±0.53
7.60±0.26
2
17.69±0.30
7.64±0.27
3
22.61±0.96
8.20±0.77
4
27.43±0.82
8.09±0.42
5
31.62±0.42
8.30±0.51
Appendix
Table 9: ColorMNIST ablation over confounding strength for DR–IMMD. “Initial” denotes the observational-arm particles before optimization, and “flow” denotes the particles returned by IMMD-Flow. Lower is better. Entries are means ± sample standard deviations across three data seeds.
Estimating the full counterfactual distribution under an intervention is central to treatment risk assessment and counterfactual generation tasks. Existing approaches model the counterfactual distribution as a standalone generative target, without exploiting its relationship to the observational data. In this work, we show that under standard assumptions, observational distributions and marginal counterfactual distributions over outcomes are tightly linked: they have identical support and tail behavior, remain statistically close under weak confounding, and share any outcome features which are invariant to confounders. These properties motivate learning counterfactual distributions not from scratch, but via a deconfounding flow from the observational distribution. We formulate this problem via flow-matching and derive a semiparametrically efficient estimator based on a novel efficient influence function correction. We subsequently extend our estimator to target minimal-energy flows in high-dimensions, which we show can be especially simple targets between observational and counterfactual distributions. In experiments, deconfounding flows outperform existing debiased counterfactual distribution estimators, while also mitigating known failure modes of flow-based methods.
Hugh Dance, Johnny Xi, Peter Orbanz +1
Gatsby Computational Neuroscience Unit, University College London · Department of Statistics, University of British Columbia
Flow matching models excel at generative modeling, and many downstream applications require their samples to satisfy prescribed constraints, such as observed measurements and physical laws. However, existing constrained samplers often face a trade-off: enforcing constraints can substantially displace samples from the pretrained data distribution. To address this trade-off, we introduce MintFlow, a training-free constrained sampling framework that formulates constraint enforcement as a minimal intervention on the pretrained flow trajectory. MintFlow seeks the minimal perturbation of an intermediate flow state such that its subsequent evolution under the pretrained flow field satisfies the target constraint. By minimally perturbing the flow state while keeping the pretrained flow field unchanged, MintFlow enforces the constraint while minimizing unnecessary deviation from the pretrained distribution. An adjoint formulation yields a closed-form expression for this perturbation, eliminating expensive iterative optimization. Furthermore, MintFlow adaptively selects the intervention time to balance the required perturbation magnitude with its amplification by the remaining flow. Across a range of tasks in generative vision and physical system modeling, MintFlow achieves competitive constraint satisfaction while preserving the pretrained generative distribution substantially better than state-of-the-art constrained methods.
Yesom Park, Kelvin Kan, Qifan Chen +3
Department of Mathematics, University of California, Los Angeles · Brookhaven National Laboratory
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 leverage 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.