acceptodds
Under review as a conference paper at ICLR 2027

Asymptotic Inconsistency of Interpolation in Terminal-Projected Flows for PDE Inversion

Abstract

Zero-shot constrained generative models have been widely adopted for solving partial differential equation (PDE) inverse problems. A prominent class of methods, such as ECI and PCFM, enforces hard physical constraints through terminal projection coupled with reverse optimal-transport (OT) displacement interpolation. The validity of these methods relies on a generally accepted asymptotic consistency assumption of the flow interpolation, that the instantaneous predicted velocity asymptotically matches the interpolation secant velocity as the integration step size vanishes. However, whether the consistency remains valid after projection has not been verified. To address this, we establish a reference framework based on Minimal Step-wise Perturbation (MSP). We prove that the asymptotic consistency fails after terminal projection: the state discrepancy does not vanish under step refinement, inducing an extraneous surplus velocity field of order . This unrecognized surplus velocity causes trajectory distortions and distribution deformation, which cannot be interpreted as principled posterior guidance. Across parameter sweeps on PDE benchmarks, we find that model performance exhibits strong task dependence, with substantial variations and even performance reversals across different task configurations. These results demonstrate that observed performance gains cannot be directly attributed to genuine posterior-aware optimization, but instead be influenced by the surplus dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow matching, PDE inverse problems, hard constraints, terminal projection, dynamical consistency, zero-shot generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_Wc3GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates an issue that arises when enforcing constraints during the sampling process of flow-matching models. The algorithms used typically estimate the final, noise-free endpoint of the trajectory, modify it to satisfy the desired constraint, and then interpolate the corrected endpoint back to an intermediate state along the sampling trajectory. The authors argue that this procedure introduces non-trivial distortions to the sampling and propose an alternative correction method to address it. The paper provides a detailed theoretical analysis of the problem and the proposed correction step, which it then extensively evaluates on PDE benchmarks, demonstrating improvements when applied on top of the existing algorithms.

Critical strengths
  • The paper investigates an important issue in enforcing constraints during the sampling process of flow-matching models. A large body of work relies on algorithms that estimate the final, noise-free state at each sampling step, apply the desired constraints to this estimate, and then interpolate the corrected state back to an intermediate noisy state along the sampling trajectory. An important question is whether this constraint enforcement and subsequent interpolation can push the intermediate state too far from the trajectory expected by the model, introducing errors that propagate through the following sampling steps and affect the generated samples. The paper provides a theoretical analysis of this issue and proposes an alternative constraint optimization and interpolation procedure to address it. Importantly, the problem is not limited to the PDE inverse problems studied by the authors, but is relevant to a broader range of constrained flow-matching applications, such as generation of images and video.

  • The experimental evaluation is well-executed, covering a range of PDE benchmarks. The proposed MSP correction leads to substantial improvements in most settings, while also reducing performance differences between methods. These results provide strong evidence that the issue the authors have identified has a significant impact on the sampling process and that addressing it is essential for improving constrained sampling with flow-matching algorithms.

Critical weaknesses
  • The Asymptotic Secant–Tangent Consistency (ASTC) assumption, introduced in Definition 1, requires the instantaneous velocity predicted by the model to converge to the secant velocity between the initial noise and the current noisy state. However, it is unclear why this assumption should hold for the base learned flows. In general, flow-matching models do not guarantee to learn straight trajectories (not possible in many cases). Even with a perfectly learned velocity and infinitesimal integration steps, the instantaneous velocity may differ from the average velocity along the trajectory. The authors claim that the ASTC fails after projecting the terminal state prediction onto the constraint manifold, but it is unclear whether the observed inconsistency is introduced by the projection itself or was already present in the underlying learned flow. The authors should clarify under what conditions ASTC holds before projection; does it only hold for straight-line flows, and if so, how often do these appear in the problems they study? Then they should distinguish the effects of the terminal state projection from the effects naturally arising from the non-linear flow trajectories.

  • The authors propose Minimal Step-wise Perturbation (MSP) to fix the issues with existing terminal-projection methods. Rather than projecting the predicted terminal state and interpolating it back to the sampling trajectory, MSP formulates the correction as an optimization problem that minimizes the Euclidean distance between the constrained next state and the unconstrained Euler update. While this guarantees minimal local perturbation, it does not necessarily establish that the corrections are consistent with the desired constrained posterior. Consequently, demonstrating that MSP eliminates the surplus velocity does not necessarily imply that it improves the generated constrained samples under that posterior. This raises an important question of whether MSP provides a more faithful approximation of the true posterior under the constraints, or whether it simply introduces a different local correction rule that improves performance.

Questions for authors
  • Is there a toy setting where you can easily demonstrate the issues with ASTC and projections? I believe that it is important to establish that projection is the root of the discrepancy rather than the nonlinear trajectories already present in the learned flow. A comparison of the ASTC defect before and after projection in a setting with a known velocity field would help clarify this distinction.

  • In Table 2, ProFlow seems to benefit less from MSP than PCFM and ECI. Could this be related to ProFlow resampling the noise at each integration step? Since MSP is originally formulated using a fixed initial state, it would be useful to (empirically) clarify how this affects the stochastic resampling.

  • Something missing from the paper was the additional overhead added by the MSP. I would assume that it is negligible compared to forward passes and the projection step, but could the authors mention it?

Reviewer_Mbh7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the use of pretrained flow matching models for the solution of PDE inverse problems. They study projection-based solvers in which they first predict a clean target, then project the target into the constraint set, and then reproject it back onto the flow path. They argue that this procedure introduces an additional surplus velocity because the interpolation direction does not necessarily agree with the pretrained velocity field. They propose Minimal Step-wise Perturbation (MSP), which instead finds the closest feasible next state to the standard Euler update by modifying the projection center. Experiments on synthetic PDE benchmarks show substantial improvements in some configurations, but also significant performance degradation in others.

Critical strengths

The paper is well-written, and the topic and the mentioned problems are interesting.

The optimization-based intermediate state update is a good direction to explore.

The experimental results provide interesting insights into the task-dependent behavior of existing methods.

Critical weaknesses

The theoretical interpretation is questionable, as the identified velocity mismatch can exist even without projection (i.e., in practice, the flow matching models do not necessarily admit a straight path in inference, and the paths are curved).

It is unclear why minimizing the deviation from the unconditional Euler step should improve posterior sampling, particularly for stochastic methods (i.e., when we sample fresh noise instead of keeping the initial noise).

Questions for authors
  1. The identified velocity mismatch can exist even without projection, particularly for curved flow trajectories. Could the authors clarify how they distinguish the inconsistency introduced by projection from the mismatch already present in the unconstrained flow?

  2. Why should MSP be useful for stochastic methods where fresh noise is drawn from the source instead of the initial noise?

Comments on the paper

4 comments
Area_Chair_h59KOpenReview ↗

Reviewer Mbh7 & Dc2G] Urgent Reminder: Discussion Period Ends November 18

Dear Reviewers Mbh7 & Dc2G,

The discussion period will end on November 18, 2026. Please carefully review the authors' responses and actively engage in discussions.

If you have not yet initiated a discussion, please do so immediately. Reviewers are expected to address the authors' rebuttals, provide constructive follow-up comments, and update their evaluations when appropriate.

Please also ensure that your reviews reflect independent human judgment and that any AI assistance is properly documented in accordance with the review guidelines.

Active participation is an essential reviewer responsibility and is critical to ensuring a rigorous and fair review process.

Thank you for your cooperation.

Thank you for your response. Most of my concerns are solved.

Dear Reviewer Dc2G:

Thank you very much for your follow-up and for taking the time to consider our responses. We are glad that our clarifications have addressed most of your concerns.

Given the clarifications and additional explanations provided, we would sincerely appreciate it if you could reconsider your overall evaluation of the paper. We hope our responses have strengthened your confidence in the contributions of this work.

Thank you again for your thoughtful feedback and consideration.

Best regards, The Authors

With only about one hour remaining before the discussion phase closes, we would first like to sincerely thank all reviewers for their valuable feedback, as well as the AC, SAC, and PC for their coordination and support throughout the review process. Over the past days, we have carefully considered every review comment and provided detailed responses to all the concerns raised. Here, we would like to briefly summarize our main efforts during the rebuttal.

  1. Reviewers Mbh7, Wc3G, and Dc2G all raised questions about the origin of curvature, particularly regarding the relationship between projection and the ASTC assumption. We addressed these concerns by clarifying the scope of our theoretical claims, supporting our explanations with relevant statements from the original manuscript, and emphasizing the mathematical compatibility between the involved operations.

  2. Reviewers Mbh7 and Wc3G raised questions about differences in model posterior distributions. We provided a detailed analysis of the rationale for choosing the Euclidean posterior, as well as the potential harm that redundant velocity fields may cause to model reliability and the future development of this research direction.

  3. Reviewer Mbh7 questioned the applicability of MSP to stochastic methods. We provided a new explanation based on the local determinism of stochastic trajectories over the current time interval and the locally optimal functional underlying our method.

  4. Reviewer Wc3G raised questions about computational overhead and extensions to special cases. We addressed these points individually through theoretical analysis and intuitive explanations.

  5. Reviewer Dc2G questioned the theoretical foundation of the MSP pushforward law. We provided a rigorous mathematical proof, addressed the numerical assumptions regarding its alleged inapplicability to specific scenarios, and further established the invariance of our conclusions from the perspective of the full normal fiber.

Throughout this discussion, we have addressed each specific concern raised by the reviewers and provided corresponding theoretical explanations, mathematical proofs, and applicability analyses. We are particularly grateful that the reviewers' subsequent feedback has also acknowledged that most of their original concerns have been resolved.

Once again, we sincerely thank all reviewers, as well as the AC, SAC, and PC, for their time and efforts throughout the review process. If any reviewer has further questions or would like to discuss any details, we remain fully available to provide additional clarification. Meanwhile, if you consider that our additional theoretical analyses and clarifications have adequately addressed the original concerns, we would sincerely appreciate your reconsideration of the scores.

Finally, we sincerely hope that the AC will consider the issues resolved during this discussion, the reviewers' subsequent feedback, and the overall contributions of our work, and support its acceptance. We believe that the theoretical perspective offered by MSP can bring new insights to the community and help overcome a key obstacle in the development of PDE inverse problems.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.