- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
They investigate how we can transporting a trajectory to a target one when considering both are driven by Markov processes. They propose to learn a conditional flow-matching velocity function on the interpolation of independently drawn next states to do that, conditioned on the source and target states. They use single (long) source trajectories with a single target trajectory as the training data. For the generations, they use recursively via a transition function from x'(i+1) to y(i+1). To evaluate their method, they used synthetic Gaussian and nonlinear-mixture transition recovery with an EEG-to-image retrieval study on THINGS-EEG2. As their results shows a history-conditioned flow inserted before subjects' temporal map increases the mean Top-5 accuracy ~5% across ten subjects.
Critical strengths
The idea of reduction from trajectory transport to a conditional transition problem, and its verification in a Gaussian instance, is an well defined and clear idea. Also, even though the experiments do not fully match or support the paper’s claims, the EEG experiment is interesting and represents an application that is less explored with CFMs.
Critical weaknesses
1-The main issue is EEG experiment does not clearly show that flow matching is the reason for the improvement. Specifically the flow is trained together with the retrieval loss which adds much more parameters than the baseline, and uses a long history that goes beyond the paper’s theory. Also there is no comparison against a simple adapter of the same size, against the same model without the flow-matching loss, or against other cross-subject EEG methods. The synthetic experiments only test one-step accuracy and do not test full trajectory generation.
2-As I understand, the you developed theory only guarantees the output follows the target distribution. It does not explain what the map keeps from the source trajectory while I think this is the most important reason to use transport for domain adaptation. The finite model class assumption makes the bounds of limited use for neural networks. Also several assumptions and propositions are developed in the main body of the paper and left with no testing nor the experiments are sufficient or clearly test them.
Questions for authors
1- How does the EEG model perform without the flow-matching loss? and how does a simple adapter with a similar number of parameters perform?
2- Can you report how errors build up over long generated trajectories atleast in the synthetic experiments?
3- What does the learned map keep the same from the source trajectory beyond generating the correct target distribution?
4- Can the mixing-time factor in the main rate be removed?
7 comments
3. What properties of the source trajectory are preserved?
Our theoretical objective is to reproduce the target trajectory distribution, and we acknowledge the reviewer's point that this alone does not guarantee preservation of task-specific labels or semantic information from the source.
However, the learned map does retain a precise relationship with the source trajectory. Each generated observation is obtained by transporting the corresponding source observation, conditioned on the current source and generated target states. Under our flow regularity assumptions, the local ODE maps are invertible. Consequently, given the initial source and target states, the source trajectory can be recovered by successively inverting these maps.
In one dimension, our transport preserves conditional quantiles: an observation at the p-th percentile of the source transition distribution is mapped to the p-th percentile of the target transition distribution. As discussed in our response to Reviewer LGtD, recursively applying these maps gives the Knothe–Rosenblatt rearrangement.
Furthermore, in our EEG experiments, source and target recordings corresponding to the same image are paired during training, and the supervised retrieval objective encourages preservation of image-relevant information. Importantly, the flow network transports EEG signals without explicitly conditioning on image labels, since inferring the label from the signal is the task itself. In our case, letting C denote the viewed image, we use a coupling in which and are independent conditional on C, but generally dependent marginally. Suppressing temporal conditioning for simplicity, the flow-matching objective is
\mathbb E_{X,Y\sim P_X(\cdot\mid C)\times P_Y(\cdot\mid C)} \mathbb E_t \left[||v_t(Z_t)-(Y-X)||^2\right], \qquad Z_t=(1-t)X+tY.$$ If, for example, the conditional interpolation distributions $P(Z_t\mid C=c)$ have disjoint supports across image labels for almost every t, then, under suitable flow regularity, the population flow transports $P_X(\cdot\mid C=c)$ to $P_Y(\cdot\mid C=c)$ for each c. Although we do not assume this condition holds for EEG, same-image pairing and the supervised retrieval loss encourage preservation of task-relevant information. Thus, our experimental objective incorporates image-level correspondence, while our theoretical guarantees concern distributional transport under the stated assumptions. We will clarify this distinction in the revised paper. ### 4. Can the mixing-time factor be removed? We thank the reviewer for raising this important question. Proposition 3 establishes that dependence on the mixing time is unavoidable in the sample complexity under our assumptions. Specifically, we construct two instances with Gaussian conditional transitions and a common class of only two candidate velocities. If the training trajectory is too short relative to the mixing time, it may fail to visit the region where the two models differ, making them statistically indistinguishable. This construction follows the lower-bound approach of Nagaraj et al. [6]. For an estimator selecting from this class to achieve the stated risk threshold on both instances with failure probability $(0,1/2\pi)$, Proposition 3 requires $$m\geq \frac{\tau_{\mathrm{mix}}}{3}\log\left(\frac{1}{2\pi\delta}\right),$$ where $\tau_{\mathrm{mix}}$ is the mixing time at tolerance $1/4$. Thus, even for regular Gaussian transitions and a class containing only two velocities, the required sample size must scale at least linearly with the mixing time for fixed confidence. It may be possible to improve the mixing-time dependence in the estimation error bound of Theorem 1. However, Proposition 3 shows that no such improvement can eliminate the mixing-time dependence of the sample-size requirement in general. We would welcome any concrete suggestions the reviewer may have for sharpening the upper bound.5. Finite model class and assumptions
We appreciate the reviewer's concern about the applicability of our assumptions to neural-network estimators. Our main contribution is an algorithm for learning transport between time-series distributions from limited, temporally dependent observations, together with theoretical guarantees under explicit assumptions. We clarify these assumptions below.
Finite model class. Finite hypothesis classes are a classical setting in statistical learning theory, including the analysis of neural networks (Anthony and Bartlett [11]). Covering-number arguments provide a standard approach to extending finite-class bounds to infinite function classes, under suitable approximation and regularity conditions.
Moreover, as explicitly noted by Anthony and Bartlett [11, Section 2.2], neural networks with finitely many possible parameter values form finite function classes. In particular, for a fixed neural-network architecture with p parameters, each represented using at most q possible values, the corresponding function class satisfies
This finite-precision interpretation is also used in the analysis of score matching for diffusion models by Kumar, Nagaraj and Sarkar [7]. Gupta et al. [8] similarly establish finite-class score-estimation bounds and extend them to neural networks using covering-number arguments. Our guarantees can therefore naturally extend to neural-network based function classes through such techniques.
Other assumptions. The assumptions in our analysis serve distinct purposes and have precedents in statistical learning and Markov-process analysis.
The conditional sub-Gaussian noise assumption (Assumption 2, lines 298–307) controls regression fluctuations. Appendix C verifies it for bounded displacements, bounded additive noise, and Gaussian transition kernels with uniformly bounded covariance.
The fourth-to-second moment comparison (Assumption 3, lines 324–328) is also used in score-estimation analysis [7]. In Appendix B, we explicitly verify realizability and the moment comparison for our Gaussian example
For finite-sample rollout guarantees, the coverage condition (Eq. (41), lines 1193–1195) controls the difference between the training distribution and the distribution of generated state pairs, analogous to concentrability conditions in offline reinforcement learning [9].
The contraction assumption (Theorem 2, lines 1195–1198; Assumption 5, lines 1242–1248) controls how local transition errors accumulate over time, as in perturbation analyses of Markov processes [10]. Our additional trajectory experiments evaluate the behavior of recursive generation, while the theoretical guarantees hold under the stated assumptions. We will clarify their role and applicability in the revised paper.
References
[1] Kneeland et al., ENIGMA: EEG-to-Image in 15 Minutes Using Less Than 1% of the Parameters, 2026.
[2] Huang and Zhu, SATTC: Structure-Aware Label-Free Test-Time Calibration for Cross-Subject EEG-to-Image Retrieval, CVPR 2026.
[3] Cui et al., SCORE: Subject Coordinate Recovery for Label-Free Cross-Subject EEG-to-Image Retrieval, 2026.
[4] Rowland et al., Orthogonal Estimation of Wasserstein Distances, AISTATS 2019.
[5] Courty et al., Joint Distribution Optimal Transportation for Domain Adaptation, NeurIPS 2017.
[6] Nagaraj et al., Least Squares Regression with Markovian Data: Fundamental Limits and Algorithms, NeurIPS 2020.
[7] Kumar, Nagaraj and Sarkar, Dimension-free Score Matching and Time Bootstrapping for Diffusion Models, NeurIPS 2025.
[8] Gupta et al., Improved Sample Complexity Bounds for Diffusion Model Training, NeurIPS 2024.
[9] Zanette and Wainwright, Bellman Residual Orthogonalization for Offline Reinforcement Learning, NeurIPS 2022.
[10] Rudolf and Schweizer, Perturbation Theory for Markov Chains via Wasserstein Distance, Bernoulli 2018.
[11] Anthony and Bartlett, Neural Network Learning: Theoretical Foundations, Cambridge University Press, 1999.
We thank the reviewer for their careful reading and detailed feedback. We appreciate the questions about the experimental comparisons, the properties of the learned transport map, and our theoretical guarantees. In response, we conducted additional ablation experiments and evaluated trajectory-level transport over longer horizons. We address each point below.
1. EEG experiments and ablation studies
Contribution of the flow-matching loss and comparison with a similarly sized adapter.
We thank the reviewer for suggesting these comparisons. We conducted additional experiments on THINGS-EEG2 using the same subject splits, training data, initialization, and optimization budget as our original experiments.
To isolate the contribution of the flow-matching (FM) objective, we set its loss coefficient to zero while retaining the same flow architecture and supervised retrieval objective. Both the subject-specific temporal map and velocity network remain trainable. We also evaluated a simple two-layer linear adapter with approximately the same number of trainable parameters.
Top-5 test retrieval accuracy (%), averaged over ten subjects and three seeds:
The above results indicate that adding the flow-matching objective improves Top-5 retrieval accuracy beyond what is achieved by the identical architecture without that loss or the similarly sized adapter.
Comparison with other cross-subject EEG methods.
We thank the reviewer for encouraging a broader comparison. Recent methods such as SATTC [2] and SCORE [3] address cross-subject EEG-to-image retrieval through different approaches. SATTC calibrates retrieval similarities using frozen EEG and image encoders, while SCORE learns representations designed for cross-subject coordinate alignment without labeled target-subject calibration. Our experiment instead studies whether conditional flow matching improves adaptation of a fixed pretrained ENIGMA decoder, using the same subject-specific calibration data, encoder, and training objective across all compared methods. Since the published SATTC and SCORE results use different pretrained models and adaptation protocols, their reported accuracies are not directly comparable to ours. We will expand the related-work discussion to explain these differences and clarify how Markov CFM relates to these approaches.
Use of previous EEG samples.
We thank the reviewer for raising this point. Our construction extends directly to higher-order Markov processes by treating the previous k observations as the current state. Specifically, for an order-k process, define
Y_i^{(k)}=(Y_{i-k+1},\ldots,Y_i).$$ These history processes are first-order Markov. We can therefore learn a conditional flow transporting the next-observation law $P_X^{(k)}(X_i^{(k)},)$ to $P_Y^{(k)}(Y_i^{(k)},)$, conditioned on the two histories, exactly as in our original construction. At generation time, we apply the learned flow to the next source observation and update the target history by appending the generated observation. With the initial target history drawn from the correct joint distribution independently of the source, the same conditional-transport and induction argument as in Proposition 2 gives the correct target trajectory distribution. The finite-sample analysis likewise extends to these augmented history processes under the corresponding mixing and moment assumptions. Our EEG implementation uses k=32 as a modeling choice. We do not assume that the EEG recordings themselves satisfy the Markov or mixing assumptions required by the theoretical guarantees. We will clarify this extension in the revised paper. ### 2. Error accumulation over longer trajectories Please see our response to Reviewer LGtD