acceptodds
Under review as a conference paper at ICLR 2027

Generalized Moment-Informed Flow Matching for Trajectory Inference

30 views

Abstract

Trajectory inference seeks to reconstruct continuous dynamics from sparse, unpaired snapshots of evolving populations. These observations do not uniquely determine the underlying path process, so modeling assumptions are needed to select plausible dynamics. Collecting more samples per snapshot or observing additional time points can improve reconstruction, but detailed measurements are often costly or destructive. We therefore consider settings in which less costly measurements provide aggregate information in the form of expectations of test functions under intermediate marginal distributions. We build on flow matching, which learns a vector field to reproduce the marginal evolution of a prescribed stochastic interpolant through simulation-free regression. To incorporate the aggregate observations into this framework, we introduce Generalized Moment-Informed Flow Matching (GMI-FM). Our method first learns an endpoint-preserving neural interpolant under the aggregate constraints using an augmented-Lagrangian formulation, then regresses the corresponding vector field against the interpolant’s velocities, keeping both stages simulation-free. Our framework also accommodates geometric priors, such as data-dependent metrics, through the interpolant regularization, encouraging learned trajectories to follow the geometry of the data. We provide theoretical grounding for GMI-FM by relating its construction to a Schrödinger bridge formulation with the same aggregate constraints. Experiments on synthetic and real-world datasets demonstrate that GMI-FM achieves better performance on marginal reconstruction metrics than competing methods. These findings highlight the potential of integrating lower-cost aggregate measurements with costly, detailed population snapshots to improve trajectory inference.

Primary area:
generative models
Keywords:
Trajectory Inference, Flow Matching, Schrödinger Bridge, Single-cell data

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_5ycmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Generalized Moment-Informed Flow Matching (GMI-FM) to incorporate intermediate moment constraints into trajectory inference. It learns a constrained conditional interpolant and then trains a velocity field through simulation-free flow matching. The authors establish connections to moment-constrained Schrödinger bridges and demonstrate improved marginal reconstruction on synthetic and real-world datasets.

Critical strengths
  1. Incorporates intermediate moment constraints into flow matching through a simulation-free framework.
  2. Provides theoretical connections to moment-constrained Schrödinger bridges.
Critical weaknesses
  1. The fixed OT coupling does not incorporate moment constraints, creating a gap between the theoretical assumptions and the implemented algorithm.
  2. The experimental evaluation raises concerns about checkpoint selection using intermediate snapshots and lacks experiments with multiple constrained time points.
Questions for authors

1. Inconsistency between the coupling and theoretical formulation.

The method uses standard OT or entropic OT coupling without incorporating intermediate moment constraints. However, Proposition 2 assumes the optimal moment-constrained Schrödinger bridge coupling, which generally differs from the coupling used in practice. The authors should provide a theoretical justification for this discrepancy, clarify how the coupling mismatch affects the validity of Proposition 2, and conduct numerical experiments comparing standard OT with moment-informed couplings in terms of constraint satisfaction and reconstruction accuracy.

2. Potential overlap between validation and test data.

In the Multi experiments, the complete marginal distribution at the unconstrained time point is used for checkpoint selection, while reconstruction is also evaluated at that time point. The authors should clarify whether validation and test samples are independent and report results without using the unconstrained intermediate distribution for model selection.

3. Lack of experiments with multiple constrained time points.

Although the framework supports moment constraints at multiple intermediate time points, the experiments only consider one constrained time point at a time. Additional experiments with multiple constrained time points are needed to evaluate the method's generality and scalability.

5 comments
AuthorsOpenReview ↗

Influence of the coupling (1/3)

Thank you for your thoughtful comments and suggestions, which have helped us significantly improve our manuscript. Please see the revised PDF, where the changes are written in blue. We also directly respond to your points below.


I. Influence of the choice of coupling

We respond to your comment on the choice of coupling and its theoretical and empirical implications.

(A) Theoretical Results

Proposition 2 uses the endpoint coupling π⋆=P0,1⋆\pi^\star=P^\star_{0,1} of the moment-constrained Schrödinger bridge to isolate the difference between conditional path constructions. However, we can relax this assumption and account for coupling mismatch.

For couplings with finite relative entropy to π⋆\pi^\star, the same bound holds on the excess KL above the best feasible approximation under the chosen coupling. For singular population OT couplings, a finite path-space KL comparison is unavailable, but additional regularity of the conditional bridges yields a W2W_2 bound between time marginals. We summarize these two extensions below. Further details are provided below the two statements. Additionally,The full statements, assumptions, and proofs are included in Appendix H.5 of the revised manuscript.

Throughout, θπ⋆\theta_\pi^\star denotes an ideal minimizer of (6) with Rprior=RlinR_{\mathrm{prior}}=R_{\mathrm{lin}}, and rH=∣λ⋆∣Hdσ2τ(1−τ)r_H=|\lambda^\star|Hd\sigma^2\tau(1-\tau) is the remainder in Proposition 2.

  1. Couplings with finite relative entropy. Under the original regularity assumptions and the required feasibility and integrability under π\pi, if KL(π∥π⋆)<∞\mathrm{KL}(\pi\|\pi^\star) \lt \infty, then

0≤KL(Qθπ⋆σ,π∥P⋆)−εapp(π)≤rH.\boxed{ 0\le \mathrm{KL}(Q_{\theta_\pi^\star}^{\sigma,\pi}\|P^\star) -\varepsilon_{\mathrm{app}}(\pi) \le r_H. }

This includes entropic OT couplings when the stated conditions hold. Here, εapp(π)\varepsilon_{\mathrm{app}}(\pi) is the best KL approximation achievable within the feasible interpolant family using π\pi. The upper bound on selection excess remains unchanged; the approximation error itself can change because it also incorporates coupling mismatch.

  1. Population OT coupling. Suppose the conditional SB marginals satisfy a uniform transport–entropy inequality with constant CtC_t and are LtL_t-Lipschitz in their endpoints, as specified below. Then, under the remaining assumptions in Appendix H.5, the marginal ptOTp_t^{\mathrm{OT}} associated with θπOT∗\theta^*_{\pi_{\mathrm{OT}}} of the ideal Stage-A interpolant satisfies

W2(ptOT,Pt⋆)≤σdt(1−t)+LtW2(πOT,π⋆)+2Ct(εcond(πOT)+rH).\boxed{ \begin{aligned} W_2(p_t^{\mathrm{OT}},P_t^\star) \le{}& \sigma\sqrt{dt(1-t)} +L_tW_2(\pi_{\mathrm{OT}},\pi^\star)\\ &+\sqrt{2C_t\bigl( \varepsilon_{\mathrm{cond}}(\pi_{\mathrm{OT}})+r_H \bigr)}. \end{aligned} }

Here, εcond\varepsilon_{\mathrm{cond}} is the best average conditional path-space KL within the feasible family. This result controls the distance between time marginals, separating auxiliary noise, endpoint-coupling mismatch, and conditional approximation and selection errors. It remains meaningful when the endpoint-coupling KL is infinite.

AuthorsOpenReview ↗

Influence of the coupling (2/3)

The following paragraphs provide further details on the results.

(a) Couplings with finite relative entropy

We consider a fixed coupling π∈Π(p0,p1)\pi\in\Pi(p_0,p_1) satisfying:

KL⁡(π∣∣π∗)<∞\operatorname{KL}(\pi||\pi^*) \lt \infty

under the same regularity and duality assumptions, with the required integrability imposed under π\pi.

Using the single-observation setting of Proposition 2 and Rprior=RlinR_{\mathrm{prior}}=R_{\mathrm{lin}}, define the feasible parameter set:

\Theta_\pi = \left\\{ \theta: R_{\mathrm{lin}}(\theta) \lt \infty,\quad \mathbb{E}_{Z\sim\pi} \bigl[f(I_\theta(\tau,Z))\bigr]=a \right\\}.

Assume this set is nonempty, and let θπ⋆\theta_\pi^\star be an ideal minimizer of problem (6) under coupling π\pi. Define

ϵapp(π)=inf⁡θ∈ΘπKL(Qθσ,π∥P⋆),\epsilon_{\mathrm{app}}(\pi) = \inf_{\theta\in\Theta_\pi} \mathrm{KL}(Q_\theta^{\sigma,\pi}\|P^\star),

the best relative-entropy approximation of P⋆P^\star achievable within the feasible interpolant family using that coupling. Then

0≤KL(Qθπ⋆σ,π∥P⋆)−ϵapp(π)≤∣λ⋆∣Hdσ2τ(1−τ),\boxed{ 0 \le \mathrm{KL}(Q_{\theta_\pi^\star}^{\sigma,\pi}\|P^\star) - \epsilon_{\mathrm{app}}(\pi) \le |\lambda^\star|Hd\sigma^2\tau(1-\tau), }

where HH and λ⋆\lambda^\star are as in Proposition 2.

Indeed, the entropy chain rule gives

KL(Qθσ,π∥P⋆)=KL(π∥π⋆)+∫KL(Qθσ,z∥P⋆,z) π(dz).\mathrm{KL}(Q_\theta^{\sigma,\pi}\|P^\star) = \mathrm{KL}(\pi\|\pi^\star) + \int \mathrm{KL}(Q_\theta^{\sigma,z}\|P^{\star,z})\,\pi(dz).

The endpoint-coupling term is independent of θ\theta. Repeating the conditional entropy and noise-bias arguments from the original proof, with expectations and feasibility evaluated under π\pi, therefore yields the stated bound. Thus, the excess relative entropy of the ideal solution of (6), compared with the best approximation available within the same feasible interpolant family and coupling, satisfies the same bound as in Proposition 2. The approximation error ϵapp(π)\epsilon_{\mathrm{app}}(\pi) itself depends on the coupling and includes its mismatch with π⋆\pi^\star.


(b) OT coupling

For unregularized quadratic OT between population distributions, the preceding finite-KL extension generally does not apply. Indeed, finite relative entropy requires absolute continuity:

KL(μ∥ν)<∞⟹ν(A)=0 ⇒ μ(A)=0.\mathrm{KL}(\mu\|\nu) \lt \infty \quad\Longrightarrow\quad \nu(A)=0\ \Rightarrow\ \mu(A)=0.

When p0p_0 is absolutely continuous, the quadratic-OT coupling is induced by a deterministic transport map TT:

\pi_{\mathrm{OT}}=(\operatorname{Id},T)_\\#p_0.

Consequently, all its mass lies on the graph GT={(x,y):y=T(x)}G_T=\{(x,y):y=T(x)\}. By contrast, the endpoint coupling of a finite-entropy SB with a nondegenerate Brownian reference is absolutely continuous with respect to p0(dx) dyp_0(dx)\,dy, and therefore assigns zero mass to this graph. Hence,

πOT(GT)=1,π⋆(GT)=0⟹KL(πOT∥π⋆)=+∞.\pi_{\mathrm{OT}}(G_T)=1,\qquad \pi^\star(G_T)=0 \quad\Longrightarrow\quad \mathrm{KL}(\pi_{\mathrm{OT}}\|\pi^\star)=+\infty.

This population-level obstruction should be distinguished from discrete approximations. For OT and SB coupling matrices on the same finite endpoint supports, the KL is finite whenever the SB coupling is positive on the support of the OT coupling. For the population OT case, we instead consider a W2W_2 comparison between time marginals, under additional stability assumptions on the conditional bridges.

In the population OT setting, we seek a bound on W2W_2 since the path-space KL⁡\operatorname{KL} is infinite. Let Pt⋆,zP_t^{\star,z} denote the SB marginal at time tt, conditional on the endpoint pair z=(x0,x1)z=(x_0,x_1). We assume that finite constants Ct,LtC_t,L_t, uniform over the relevant endpoint pairs, satisfy

(i)W22(ν,Pt⋆,z)≤2Ct KL(ν∥Pt⋆,z),∀ν∈P2(Rd),(ii)W2(Pt⋆,z,Pt⋆,z′)≤Lt∥z−z′∥.\begin{aligned} \text{(i)}\quad& W_2^2(\nu,P_t^{\star,z}) \le 2C_t\,\mathrm{KL}(\nu\|P_t^{\star,z}), &&\forall \nu\in\mathcal P_2(\mathbb R^d),\\ \text{(ii)}\quad& W_2(P_t^{\star,z},P_t^{\star,z'}) \le L_t\|z-z'\|. \end{aligned}

The first assumption converts conditional relative-entropy error into Wasserstein error. The second requires that small changes in the endpoints produce controlled changes in the conditional intermediate distribution. The first condition holds automatically for several common observation families in our Brownian-reference setting. For affine or normalizable quadratic tilts, the conditional SB marginals remain Gaussian and therefore satisfy a transport–entropy inequality. Globally bounded observation functions, with finite dual multipliers, produce bounded perturbations of Gaussian laws, which also satisfy this inequality. In these cases, the constants can be chosen uniformly over endpoint pairs. Thus, for these observation families, the transport–entropy condition follows from the stated model assumptions. For general unbounded nonlinear observations, additional control of concentration is required; normalizability alone is insufficient.

AuthorsOpenReview ↗

Influence of the coupling (3/3)

Retaining the remaining feasibility and regularity assumptions of Proposition 2, let ptOTp_t^{\mathrm{OT}} denote the marginal of the ideal Stage-A interpolant obtained under πOT\pi_{\mathrm{OT}}. Then

W2(ptOT,Pt⋆)≤σdt(1−t)+LtW2(πOT,π⋆)+2Ct(εcond(πOT)+∣λ⋆∣Hdσ2τ(1−τ)).\boxed{ \begin{aligned} W_2(p_t^{\mathrm{OT}},P_t^\star) \le{}& \sigma\sqrt{dt(1-t)} +L_tW_2(\pi_{\mathrm{OT}},\pi^\star)\\ &+\sqrt{2C_t\left( \varepsilon_{\mathrm{cond}}(\pi_{\mathrm{OT}}) +|\lambda^\star|Hd\sigma^2\tau(1-\tau) \right)}. \end{aligned} }

Here, εcond(πOT)\varepsilon_{\mathrm{cond}}(\pi_{\mathrm{OT}}) is the smallest average conditional path-space KL achievable by a feasible interpolant under the OT coupling. The bound separates three contributions: removal of the auxiliary Brownian-bridge noise, endpoint-coupling mismatch, and conditional approximation and selection errors. It therefore provides a comparison with the constrained SB even when the endpoint couplings are singular relative to one another, while making explicit that a small right-hand side guarantees closeness to the constrained SB marginals.


(B) Empirical Results

Our use of endpoint-only OT couplings was motivated by computational simplicity and empirical observations in related work. For example, Tong et al. [1] (SF²M, Appendices A.1 and D.2) explain that, although entropic OT provides the coupling required for their SB formulation, unregularized OT can perform better in some settings. While this observation concerns a different setting, it motivated our choice of OT, particularly given the additional computational cost of constructing a moment-informed coupling.

To assess the difference in performance, we ran GMI-FM (lin) and GMI-FM (LAND) with a numerically computed moment-constrained SB coupling on all four benchmarks. Values are average reconstruction W2W_2 (lower is better), reported as mean ± population standard deviation over five seeds.

DatasetMethodSB coupling: W2W_2Endpoint-only: W2W_2SB coupling: residualEndpoint-only: residual
SyntheticGMI-FM (lin)0.1723 ± 0.00640.1710 ± 0.00630.0339 ± 0.01470.0248 ± 0.0149
SyntheticGMI-FM (LAND)0.1867 ± 0.00890.1834 ± 0.00760.0414 ± 0.00960.0359 ± 0.0126
AerosolGMI-FM (lin)0.3091 ± 0.00150.2917 ± 0.00010.0335 ± 0.02120.0013 ± 0.0005
AerosolGMI-FM (LAND)0.2980 ± 0.00180.2873 ± 0.00020.0072 ± 0.00310.0011 ± 0.0008
MultiGMI-FM (lin)10.4688 ± 0.004210.5401 ± 0.00200.0340 ± 0.00100.0419 ± 0.0028
MultiGMI-FM (LAND)9.9478 ± 0.003210.0937 ± 0.00350.0762 ± 0.00120.0822 ± 0.0008
SemrauGMI-FM (lin)1.9724 ± 0.00231.9961 ± 0.00100.8356 ± 0.64130.7197 ± 0.6165
SemrauGMI-FM (LAND)1.7197 ± 0.00201.6726 ± 0.00620.3136 ± 0.10700.6887 ± 0.3340

The results suggest that the benefit of moment-informed endpoint pairing depends on both the dataset and the geometric prior. Endpoint-only configurations achieve lower mean reconstruction errors on Synthetic and Aerosol, although the differences on Synthetic are modest relative to variability across seeds. The SB coupling improves reconstruction on Multi for both variants. On Semrau, it improves GMI-FM (lin) but worsens GMI-FM (LAND). Thus, incorporating moment information into the coupling does not yield a uniform reconstruction advantage. These findings complement the theoretical discussion: optimality for the moment-constrained SB objective does not guarantee lower reconstruction error, and these experiments assess the practical effect of changing the coupling.

Computing the moment-informed coupling also introduces additional cost. On Multi, coupling preparation took approximately 6–9 minutes, including refinement and sampling checks, compared with 5–7 seconds for entropy-matched endpoint-only coupling. This overhead arises from evaluating moment-dependent conditional expectations across four million endpoint pairs and repeatedly rebalancing the coupling during optimization.

Update: We have added a discussion of the change in the **proposition 2 when considering different coupling in appendix H.5. The empirical results are included in the appendix C.9.

[1] Alexander Y. Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. “Simulation-Free Schrödinger Bridges via Score and Flow Matching.” AISTATS 2024, Proceedings of Machine Learning Research, 238:1279–1287.

Reviewer_4gx8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work aims to introduce moment matching into the usual Schrodinger-bridge-based trajectory inference problem. They formulate and solve it via a modified Flow Matching approach. A two-stage approach is formulated, with first training of a modified interpolant to fit moment constraints and then a second training of an advecting velocity field that matches the time derivative of this modified interpolant. They then perform some theoretical characterizations of the first stage objective (a path regularizer, usually velocity-deviation-based), and test their methods on a few datasets: one synthetic, two single-cell datasets, and one atmospheric particle dataset. These experiments show favorable performance in W2W_2 error when compared to CFM, MFM, and a constructed straw man CSB-FM.

Critical strengths
  1. The paper clearly describes an interesting and relevant generalization of a flow matching approach to the trajectory inference problem, and presents a reasonable strategy for solving it. The solve strategy seems well-thought-out, and significant details on all the procedures are provided in the appendices.
Critical weaknesses
  1. The approach is perhaps relatively straightforward to formulate and one might argue that technical novelty is limited.
  2. The improvements in marginal agreement are not staggering, and it is perhaps not surprising that additional information on moments leads to slightly improved marginal W2W_2. It is unclear if the gains are worth the additional runtime of the method in a practical sense.
  3. There could be much further discussion or exploration of robustness to parameter change and tuning. For example, I was curious to know how the results vary as one changes the relative amount of training in each step.
Questions for authors
  1. Can you give further details on the CSB-FM straw man? What bridge solver was used and how was the velocity training done?
4 comments
AuthorsOpenReview ↗

Magnitude and significance of the reported improvements

Thank you for your thoughtful comments and suggestions, which have helped us significantly improve our manuscript. Please see the revised PDF, where the changes are written in blue. We also directly respond to your points below.

I. Magnitude and significance of the reported improvements

We agree that the improvements in global marginal reconstruction on the real-world datasets are modest. Their magnitude should be interpreted in relation to the information supplied by the aggregate observations. In Semrau, for example, we constrain the average expression of Sox2 and Tbx3, two genes involved in pluripotency and differentiation. Although biologically important, these measurements concern only two genes within a much larger gene-expression profile. They therefore provide relevant but narrow information about the population, leaving substantial ambiguity in the full cell-state distribution. Matching these moments need not translate into large improvements in overall marginal reconstruction. Our experiment in Section C.3 illustrates the limiting case of a redundant moment constraint: the reference interpolant already satisfies the constraint, so enforcing it leaves reconstruction performance essentially unchanged.

Consequently, we believe that our method can substantially improve marginal reconstruction when the information supplied by the generalised moments more strongly restricts the model’s possible dynamics. This is what we observed in the synthetic experiment, where the average W2W_2 at the unconstrained times drops from 0.63160.6316 to 0.17100.1710. These results suggest that reconstruction performance can improve considerably when the supplied moments are accurately estimated and informative. We also observe improvements when combining different generalised moments at the same constrained time, as shown in Section C.1. In such scenarios, we believe that these gains can justify the additional computation time. We also observe improvements when combining different generalised moments at the same constrained time, as shown in Section C.1. In a new experiment on the aerosol dataset (Section C.2), we constrain three generalised moments at three different times, obtaining an approximately 10% reduction in W2W_2 at the constrained snapshots.

Furthermore, although supplying moment information at specific times is expected to improve reconstruction, our comparison with CSB-FM shows that the way this information is incorporated also matters. CSB-FM extends the well-studied Schrödinger bridge formulation to incorporate the same generalised-moment constraints. With the same aggregate information, our method can achieve better reconstruction and, in the high-dimensional Multi experiment, a lower runtime. In this setting, the numerical integration and conditional sampling required to approximate the constrained bridge solution are computationally demanding.

Finally, we believe that one strength of our framework is its ability to use various types of metadata associated with population datasets to improve reconstruction. One major way to improve trajectory inference methods is either to collect more data—such as additional data points or snapshots—or to gather other types of information help constrain the possible dynamics underlying the observations. Our framework allows us to use these relatively inexpensive measurements to improve reconstruction, complementing the more costly measurements needed to obtain individual data points. We have demonstrated this on synthetic and real-world datasets by incorporating existing metadata as generalised moments. We believe that our framework also has potential for new datasets containing even richer metadata.

AuthorsOpenReview ↗

Sensitivity to training budgets, stage allocation, and optimisation hyperparameters

II. Sensitivity to training budgets, stage allocation, and optimisation hyperparameters

Following your suggestion, we conducted additional experiments to assess the impact of training budgets, the allocation of training between stages, and optimisation hyperparameters.

A. Impact of the training budgets

We conducted two experiments on the arched synthetic diffusion and Semrau benchmarks. First, denoting the original Stage-A and Stage-B update budgets by (NA,NB)(N_A,N_B), we evaluated reconstruction for all nine combinations (eNA,jNB)(eN_A,jN_B), with e,j∈{0.5,1,2}e,j\in\{0.5,1,2\}, for GMI-FM (lin) and GMI-FM (LAND). This tests whether performance changes substantially when either stage receives more or fewer training updates. Second, we investigated how to divide a fixed compute allowance between the two stages, allocating {10,25,50,75,90}%\{10,25,50,75,90\}\% to Stage A and the remainder to Stage B. Compute is measured using optimiser-update CPU time, excluding validation, diagnostics and checkpoint I/O. Both experiments use the same five seeds.

The results and detailed protocol are presented in Appendix C.6. In the first experiment, reconstruction is relatively stable: changes in mean W2W_2 remain below 2%2\% relative to the original setting across all tested budget combinations. Some longer Stage-A runs show slight degradation. A possible explanation is the accumulation of Lagrange multipliers when constraint residuals persist, which can destabilise longer runs if the effective constraint strength is too large. Our calibration heuristic, described in Appendix F.4 (“Initialisation of constraint hyperparameters”), accounts for this possibility by considering both the relative gradient scales and the planned number of multiplier updates.

For the fixed-compute experiment, GMI-FM (lin) maintains relatively stable reconstruction across the tested allocations. This may reflect the weak initial influence of RpriorR_{\mathrm{prior}}, allowing moment matching with limited competition from the regulariser. For GMI-FM (LAND) on Semrau, allocating more compute to Stage A improves reconstruction and generally reduces moment constraint errors. This is consistent with requiring more interpolant training to balance moment matching against RLANDR_{\mathrm{LAND}}, with improved constraint satisfaction accompanying lower reconstruction error.

B. Hyperparameter Sensitivity

We explore sensitivity to two key hyperparameters governing Stage-A optimisation: the effective constraint strength κ=ηρ\kappa=\eta\rho, which controls the influence of moment matching relative to the path regulariser, and the Stage-A learning rate, which controls the size of the interpolant updates. For GMI-FM (lin) and GMI-FM (LAND), we multiply κ\kappa by {0.25,0.5,1,2,4}\{0.25,0.5,1,2,4\}, varying ρ\rho while keeping η\eta fixed. Separately, we multiply the Stage-A learning rate by {0.5,1,2}\{0.5,1,2\}. We evaluate both sweeps on the arched synthetic diffusion and Semrau benchmarks over the same five seeds, changing one parameter at a time and retaining the original training budgets and remaining experimental settings.

The detailed protocol and results are presented in the appendix subsection C.7 “Hyperparameter sensitivity.” Reconstruction is relatively insensitive to the tested learning rates and, for GMI-FM (lin), to the constraint strength. On Semrau, changes in mean W2W_2 remain below 0.2%0.2\% relative to the reference setting for both variants across all tested values.

The main exception is GMI-FM (LAND) on the synthetic benchmark: multiplying κ\kappa by four increases mean W2W_2 from 0.1830.183 to 0.4990.499 and mean rollout constraint error from 0.0360.036 to 0.5790.579. This shows that excessive constraint strength can degrade both reconstruction and moment matching. The reference settings perform well across both benchmarks and variants, consistent with the practical usefulness of our calibration heuristic, described in Appendix F.4 (“Initialisation of constraint hyperparameters”). We retain these reference settings throughout; the sensitivity results are not used to select new hyperparameters.

Update: We added analyses of sensitivity to training stage allocation in Appendix C.6 and to optimization hyperparameters in Appendix C.7.

AuthorsOpenReview ↗

Information on CSB-FM

III. Information on CSB-FM

We start from the optimization problem in Definition 2. Under the hypotheses specified below, its solution admits an analytical characterization involving a moment multiplier λ⋆\lambda^\star and an endpoint coupling π⋆\pi^\star. Our first task is to determine these quantities. We can then generate conditional samples from the bridge and compute their analytical probability-flow velocities. Finally, conditional flow matching uses these velocities as regression targets to learn an ODE reproducing the bridge’s marginal evolution in the ideal regression limit.

The method therefore consists of three main steps:

  1. Construct the bridge distribution: compute a multiplier λ⋆\lambda^\star and coupling π⋆\pi^\star that jointly satisfy the endpoint and generalized-moment constraints.

  2. Generate samples and conditional velocity targets: sample the endpoints and constrained-time position, then use the resulting Brownian subbridges to obtain positions and their analytical probability-flow velocities.

  3. Learn the marginal velocity field: train a neural network by conditional flow matching, averaging the conditional velocities into a field that can be integrated as an ODE.

    1. Determine the moment multiplier and endpoint coupling.

For a candidate multiplier λ\lambda, define

Zλ(x,y)=EZ∼Gxy[eλ⊤f(Z)],Gxy=N ⁣((1−τ)x+τy, σ2τ(1−τ)I).Z_\lambda(x,y) = \mathbb E_{Z\sim G_{xy}} \left[e^{\lambda^\top f(Z)}\right], \qquad G_{xy} = \mathcal N\!\left( (1-\tau)x+\tau y,\, \sigma^2\tau(1-\tau)I \right).

Here GxyG_{xy} is the reference Brownian-bridge marginal at the observation time. The corresponding endpoint coupling solves the entropic transport problem

\pi^\lambda \in\arg\min_{\pi\in\Pi(p_0,p_1)} \left\\{ \mathrm{KL}(\pi\\|R_{0,1}) - \int\log Z_\lambda(x,y)\\,\pi(dx,dy) \right\\}.

After discretizing the endpoints, we solve this problem by Sinkhorn scaling of the modified Brownian kernel

Kijλ=exp⁡(−∥xi−yj∥22σ2)Zλ(xi,yj).K^\lambda_{ij} = \exp\left(-\frac{\|x_i-y_j\|^2}{2\sigma^2}\right) Z_\lambda(x_i,y_j).

The scaling enforces the prescribed endpoint masses. We then adjust λ\lambda to satisfy the moment constraint. The relevant convex dual objective is

F(\lambda) = \sup_{P:\\,P_0=p_0,\\,P_1=p_1} \left\\{ \lambda^\top\mathbb{E}_P[f(X_\tau)] - \mathrm{KL}(P\\|R) \right\\} - \lambda^\top a.

Its gradient is precisely the moment mismatch:

∇F(λ)=∫∇λlog⁡Zλ(x,y) πλ(dx,dy)⏟m(λ)−a.\nabla F(\lambda) = \underbrace{ \int\nabla_\lambda\log Z_\lambda(x,y)\, \pi^\lambda(dx,dy) }_{m(\lambda)} -a.

Consequently, a gradient update takes the form

λk+1=λk+ηk[a−m(λk)],\lambda_{k+1} = \lambda_k+\eta_k\bigl[a-m(\lambda_k)\bigr],

with the coupling recomputed for each candidate multiplier. Scalar constraints also permit monotone root finding. This formulation directly links the convex dual problem to the prescribed constraints: the dual gradient is the moment mismatch, while the coupling solve enforces the endpoint marginals. We assess numerical accuracy through endpoint and moment residuals, together with independent checks of the conditional integration or sampling where required.

  1. Sample the conditional bridge and compute its analytical velocity.

Once λ⋆\lambda^\star and π⋆=πλ⋆\pi^\star=\pi^{\lambda^\star} are available, draw (x,y)∼π⋆(x,y)\sim\pi^\star. The conditional position at time τ\tau follows the exponentially tilted Gaussian

qλ⋆(dz∣x,y)=exp⁡((λ⋆)⊤f(z))Zλ⋆(x,y)Gxy(dz).q_{\lambda^\star}(dz\mid x,y) = \frac{\exp((\lambda^\star)^\top f(z))} {Z_{\lambda^\star}(x,y)} G_{xy}(dz).

When the tilt has a known finite upper bound

M≥sup⁡z(λ⋆)⊤f(z),M\geq\sup_z(\lambda^\star)^\top f(z),

this distribution can be sampled using standard rejection sampling: propose z∼Gxyz\sim G_{xy} and accept with probability

exp⁡((λ⋆)⊤f(z)−M).\exp\bigl((\lambda^\star)^\top f(z)-M\bigr).

Repeat rejected proposals while keeping x,yx,y fixed. This requires no evaluation of the normalizer. For normalizable quadratic tilts, the conditional law is Gaussian and can instead be sampled directly. Conditional on the accepted triple (x,z,y)(x,z,y), the path consists of two independent ordinary Brownian bridges: x→zx\to z on [0,τ][0,\tau], and z→yz\to y on [τ,1][\tau,1]. Thus the constrained solution is a mixture of concatenated Brownian bridges. On either subinterval [s,r][s,r], with endpoints A,BA,B -- Here, AA and BB are placeholders for the endpoints of the subbridge containing tt: (A,B)=(x,z)(A,B)=(x,z) when 0<t<τ0 \lt t \lt \tau, and (A,B)=(z,y)(A,B)=(z,y) when τ<t<1\tau \lt t \lt 1--, a position can be sampled as

Xt=mt+ct ϵ,ϵ∼N(0,I),X_t=m_t+\sqrt{c_t}\,\epsilon, \qquad \epsilon\sim\mathcal N(0,I),

where

mt=r−tr−sA+t−sr−sB,ct=σ2(t−s)(r−t)r−s.m_t=\frac{r-t}{r-s}A+\frac{t-s}{r-s}B, \qquad c_t=\sigma^2\frac{(t-s)(r-t)}{r-s}.

These Gaussian marginals have the analytical probability-flow velocity

vcond(t,w∣A,B)=m˙t+c˙t2ct(w−mt)=12(B−wr−t+w−At−s).v_{\mathrm{cond}}(t,w\mid A,B) = \dot m_t+\frac{\dot c_t}{2c_t}(w-m_t) = \frac12\left( \frac{B-w}{r-t} + \frac{w-A}{t-s} \right).

The term mt˙\dot{m_t} moves the Gaussian’s centre; the second accounts for its changing spread.

AuthorsOpenReview ↗

Information on CSB-FM 2

  1. Learn the marginal velocity by conditional flow matching.

We freeze the bridge and generate training examples by sampling a triple, a time tt, and a position XtX_t. The corresponding regression label is

Ut=vcond(t,Xt∣A,B).U_t=v_{\mathrm{cond}}(t,X_t\mid A,B).

We train a network vθ(t,Xt)v_\theta(t,X_t) with the weighted CFM objective

L(θ)=E[w(t) ∥vθ(t,Xt)−Ut∥2].\mathcal L(\theta) = \mathbb E\left[ w(t)\,\|v_\theta(t,X_t)-U_t\|^2 \right].

A positive time weight, such as

w(t)=6ut(1−ut),ut=t−sr−s,w(t)=6u_t(1-u_t), \qquad u_t=\frac{t-s}{r-s},

controls the singular conditional-velocity variance near pinned times. Since the weight depends only on time, the population regression minimizer is

v⋆(t,x)=E[Ut∣Xt=x].v^\star(t,x)=\mathbb E[U_t\mid X_t=x].

Averaging the conditional continuity equations shows that this field transports the marginals of P⋆P^\star. With exact regression and a well-posed flow, the ODE

Y˙t=v⋆(t,Yt),Y0∼p0,\dot Y_t=v^\star(t,Y_t), \qquad Y_0\sim p_0,

therefore satisfies Law⁡(Yt)=Pt⋆\operatorname{Law}(Y_t)=P^\star_t. A finite neural fit approximates this field. The analytical representation assumes a finite-relative-entropy solution, strong duality with attained finite multipliers and endpoint potentials, and the required normalizability and integrability. The weighted CFM construction requires finite second moments and suitable regularity for the resulting ODE.

Further details on the bridge characterization and the optimality conditions for its coupling and moment multipliers are provided in Appendix G.1. Appendix G.2 describes the conditional sampling procedure, analytical velocity targets and flow-matching objective. Network architectures and training settings are given in Appendices F.5–F.6.

For your question about the bridge solver, we used a custom numerical implementation of the entropy-dual formulation of the moment-constrained Brownian Schrödinger bridge. In the full-coupling solver, each candidate moment multiplier defines a modified Brownian kernel, and its endpoint coupling is computed using stabilized Sinkhorn scaling. The remaining moment equations are solved using SciPy’s Brent root-finding routine for the synthetic and atmospheric experiments, and SLSQP for Semrau. Conditional Gaussian integrals are evaluated analytically when possible, and otherwise using Gauss–Hermite quadrature.


Thank you again for your time and effort in reviewing our submission.

Reviewer_Uwv8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Generalized Moment-Informed Flow Matching (GMI-FM) for trajectory inference from unpaired population snapshots supplemented with intermediate aggregate observations, expressed as generalized moment constraints. The method extends learned-interpolant approaches through a two-stage procedure. First, a neural interpolant is optimized to satisfy the moment constraints while preserving the endpoint distributions, using an augmented Lagrangian and a path regularizer. Second, a velocity field is trained using conditional flow matching on the learned interpolant. Both stages avoid numerical trajectory simulation during training.

The main contributions are the incorporation of generalized moment constraints into flow matching, the flexibility to impose geometric path priors, and a theoretical connection to moment-constrained Schrödinger bridges, including variational comparisons and special-case correspondence results.

The method is evaluated on a synthetic diffusion benchmark, atmospheric particle measurements, and two single-cell datasets. Comparisons include CFM [1], Metric Flow Matching (MFM) [2], and a numerical constrained Schrödinger bridge baseline (CSB-FM), with additional multimarginal baselines in the appendix. The results demonstrate substantial improvements in intermediate marginal reconstruction on synthetic data, while improvements on real-world datasets are relatively modest. Additional experiments investigate sensitivity to noisy moment observations and reconstruction with limited intermediate samples.

References

[1] Alexander Tong, Kilian FATRAS, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, & Yoshua Bengio (2024). Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research.

[2] Kapuśniak, Kacper, et al. "Metric flow matching for smooth interpolations on the data manifold." Advances in Neural Information Processing Systems 37 (2024): 135011-135042.

Critical strengths

None

Critical weaknesses

1. Limited methodological novelty and insufficient experimental validation. The proposed method extends existing learned-interpolant approaches by introducing generalized moment constraints through an augmented Lagrangian, followed by a separate flow-matching stage. However, the method is not end-to-end, and its methodological advantages over existing approaches are not sufficiently established. In particular, the evaluation lacks comparisons with recent closely related methods, such as Multimarginal Flow Matching with Optimal Transport Potentials [1] and Curly Flow Matching [2]. While the synthetic results are promising, improvements on real-world datasets are relatively modest. Consequently, the experiments do not convincingly establish the significance of the proposed contribution.

2. Major presentation issues and insufficient theoretical justification. The manuscript requires substantial revision to improve clarity and scientific rigor. Section 2 is particularly difficult to follow: the trajectory inference problem (lines 101–106) is introduced only through deterministic dynamics, without discussing the broader stochastic formulation [3], while the explanation of flow matching (lines 108–112) is neither self-contained nor sufficiently informative for unfamiliar readers. In Section 3.2, the regularizer RpriorR_{\mathrm{prior}} is introduced without adequate motivation before Eq. (6), and the feasibility of the moment constraints (line 226) is simply assumed without sufficient discussion. The theoretical contribution in Section 3.3 also needs clearer justification, as several results appear to follow from established arguments. Citation coverage is insufficient throughout the manuscript: for example, the characterization of Schrödinger bridges in lines 270–273 lacks an explicit citation, and Section 3.3 references only [4]. Important related work is relegated to the appendix, while the CSB-FM baseline and the W2W_2 metric are insufficiently introduced in the main text. Furthermore, the Multi dataset [5] lacks a citation, several references use outdated arXiv versions, and important figures are placed in the appendix.

Taken together, these issues make it unnecessarily difficult to understand the method, verify its theoretical foundations, and assess its novelty relative to existing work.

References

[1] Kansal, Raghav, David, Crair, Nghia D., Nguyen, Scott D, Pope, Bradley R, Parry. "Multimarginal flow matching with optimal transport potentials." Proceedings of the 43rd International Conference on Machine Learning. PMLR, 2026.

[2] Petrović, Katarina, et al. "Curly flow matching for learning non-gradient field dynamics." Advances in Neural Information Processing Systems 38 (2026): 121544-121573.

[3] Lavenant, Hugo, et al. "Toward a mathematical theory of trajectory inference." The Annals of Applied Probability 34.1A (2024): 428-500.

[4] Léonard, Christian. “A Survey of the Schrödinger Problem and Some of Its Connections with Optimal Transport.” Discrete and Continuous Dynamical Systems, vol. 34, no. 4, 2014, pp. 1533–1574.

[5] Burkhardt, Daniel, et al. "Multimodal single-cell integration across time, individuals, and batches." NeurIPS Competitions 54 (2022): 114.

Questions for authors

1. Comparison with related methods (clarification). Why do the authors not compare against Curly Flow Matching [1], which also learns nonlinear interpolants? How does GMI-FM differ from this approach beyond incorporating intermediate moment constraints, and what are its advantages?

2. Feasibility of moment constraints (clarification). In Section 3.2, the authors assume that the prescribed moment constraints are feasible for the chosen endpoint coupling and interpolant family. Under what conditions is this assumption justified? How does the method behave when the constraints are incompatible or only approximately feasible?

References

[1] Petrović, Katarina, et al. "Curly flow matching for learning non-gradient field dynamics." Advances in Neural Information Processing Systems 38 (2026): 121544-121573.

5 comments
AuthorsOpenReview ↗

Methodological Contribution and Experimental Validation

Thank you for your thoughtful comments and suggestions, which have helped us significantly improve our manuscript. Please see the revised PDF, where the changes are written in blue. We also directly respond to your points below.


I. Methodological contribution and experimental validation

We agree that the improvements in global marginal reconstruction on the real-world datasets are modest. Their magnitude should be interpreted in relation to the information supplied by the aggregate observations. In Semrau, for example, we constrain the average expression of Sox2 and Tbx3, two genes involved in pluripotency and differentiation. Although biologically relevant, these measurements concern only two genes within a much larger gene-expression profile, leaving substantial ambiguity in the full cell-state distribution. Matching these moments therefore need not translate into large improvements in overall marginal reconstruction. Our redundant-moment experiment in appendix B.3 illustrates the limiting case: the reference interpolant already satisfies the supplied constraint, so enforcing it leaves reconstruction essentially unchanged.

Nevertheless, informative generalised moments can yield substantial improvements. In the synthetic experiment, GMI-FM (lin) reduces average W2W_2 at unconstrained times from 0.6316 for CFM to 0.1710. The moment-complementarity experiment in appendix C.1 further shows that combining observations of the population’s location and spread improves reconstruction over using either observation alone. These results support the usefulness of aggregate information when it meaningfully restricts the possible marginal evolution.

Our new aerosol experiment in appendix B.2 provides further evidence on real data. Supplying three generalised moments at three intermediate times, instead of supplying the same moments only at the midpoint, reduces average W2W_2 across those three snapshots by approximately 9.5%9.5\% for GMI-FM (lin) and 9.2%9.2\% for GMI-FM (LAND), over ten seeds. These targets are derived from measured histograms and supplied to trajectory learning as aggregate quantities.

Our methodological contribution is a framework that makes such partial observations usable within flow matching. Collecting additional detailed population snapshots can be costly or destructive, whereas aggregate measurements can provide complementary information at lower cost. Building on learned-interpolant methods and augmented-Lagrangian optimisation, GMI-FM learns an endpoint-preserving probability path under generalised-moment constraints, accommodates geometric priors, and subsequently learns its marginal velocity field. Both training objectives can be evaluated without numerical trajectory simulation. Our theoretical analysis also characterises the relationship between this construction and moment-constrained Schrödinger bridges.

The comparison with CSB-FM assesses how the incorporation of aggregate information affects reconstruction. CSB-FM numerically approximates the constrained Schrödinger bridge problem in Definition 2 and receives the same observation functions and targets as GMI-FM. On the synthetic benchmark, its average W2W_2 at unconstrained times is 0.24600.2460, compared with 0.17100.1710 for GMI-FM (lin). GMI-FM also achieves better reconstruction and lower reported fitting time in the high-dimensional Multi experiment, where the numerical constrained-bridge approximation is computationally demanding. These comparisons provide empirical evidence that the modelling and numerical formulation matter even when the supplied aggregate information is fixed.

Regarding multimarginal methods, the suggested Multimarginal Flow Matching with Optimal Transport Potentials addresses the multimarginal setting, where samples from several population snapshots, including intermediate marginals, are available for trajectory learning. We focus here on reconstruction from the two endpoint marginals p0p_0 and p1p_1, supplemented by generalised-moment constraints at intermediate times, rather than fitting several fully observed marginals. These constraints specify population averages and generally do not determine the full intermediate distributions. This distinction motivates our comparisons with methods using endpoint snapshots, together with CSB-FM, which receives the same aggregate inputs. We acknowledge an important qualification for Multi: its fixed observation classifier is trained on annotated constraint-day cells. This additional information forms part of that experiment’s observation model and must be included when assessing its information budget.

AuthorsOpenReview ↗

Methodological Contribution and Experimental Validation 2

To broaden the experimental comparison, we have added SF²M, LightSB-M, and IPMF across all four benchmarks and discussed them in the Related Work section. The updated main-results table shows that GMI-FM (lin) achieves the lowest mean reconstruction errors on the synthetic benchmark, while GMI-FM (LAND) achieves the lowest means on the real-world benchmarks. The expanded comparison therefore preserves the distinction between substantial synthetic gains and more modest real-world improvements.

Regarding end-to-end training, we agree that GMI-FM uses two separate stages. This structure is also used by MFM and Curly-FM, which learn conditional paths before fitting marginal dynamics. In GMI-FM, this separation allows generalised-moment constraints to act directly on the interpolant while keeping both training objectives simulation-free.

More specifically, Curly-FM is designed to exploit velocity measurements or estimates associated with the observed snapshots. These inform a reference drift btb_t, defining a reference process

dXt=bt(Xt) dt+σ dWt\mathrm dX_t=b_t(X_t)\,\mathrm dt+\sigma\,\mathrm dW_t

with path law QbQ_b. Curly-FM then approximates the Schrödinger bridge problem

min⁡P: P0=p0,  P1=p1KL(P∥Qb),\min_{P:\,P_0=p_0,\;P_1=p_1} \mathrm{KL}(P\|Q_b),

using learned conditional bridges, a coupling constructed from their costs, and subsequent flow-matching regression. Our datasets do not provide the auxiliary velocity information used to construct this reference drift. GMI-FM addresses a different optimisation problem: for a fixed endpoint coupling π\pi, we minimise a path regulariser subject to the supplied aggregate observations,

min⁡θRprior(θ)subject toE(X0,X1)∼π[fk(Iθ(τk,X0,X1))]=ak,k=1,…,K,\min_\theta R_{\mathrm{prior}}(\theta) \quad\text{subject to}\quad \mathbb E_{(X_0,X_1)\sim\pi} \left[f_k\left(I_\theta(\tau_k,X_0,X_1)\right)\right] =a_k,\qquad k=1,\ldots,K,

with endpoint preservation built into IθI_\theta. Here, the regulariser selects among interpolants compatible with the population averages; our training objective does not require solving a prescribed-drift Schrödinger bridge. The connection established in Section 3.3 characterises the relationship between these modelling formulations. GMI-FM therefore enables the use of aggregate observations without requiring the velocity information that motivates Curly-FM, which is why we focus our experimental comparisons on methods applicable to the observations available in our setting.

Update: Update: We have expanded the Experiments section with comparisons against SF²M [1], LightSB-M [2], and IPMF [3] across all four datasets, and added descriptions of these methods to the Related Work section. We have also added an aerosol experiment in the appendix examining generalised-moment constraints at multiple intermediate times in appendix B.2.

[1] Tong, A. Y., et al. “Simulation-Free Schrödinger Bridges via Score and Flow Matching.” International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.

[2] Gushchin, N., Kholkin, S., Burnaev, E., and Korotin, A. “Light and Optimal Schrödinger Bridge Matching.” International Conference on Machine Learning (ICML), 2024.

[3] Kholkin, S., et al. “Diffusion & Adversarial Schrödinger Bridges via Iterative Proportional Markovian Fitting.” International Conference on Learning Representations (ICLR), 2026.

AuthorsOpenReview ↗

Manuscript Revision

Thank you for these detailed comments. We have substantially revised the presentation and clarified the assumptions, scope, and interpretation of our theoretical results.

We have rewritten Section 2, now entitled “Preliminaries and Setting.” Trajectory inference is formulated as inferring a probability distribution over continuous trajectories from partial observations, encompassing stochastic dynamics. We then specify our observation setting: samples from the endpoint distributions, supplemented by generalised-moment observations at intermediate times. The section now introduces the stochastic interpolant, the flow-matching regression objective, and its population-optimal velocity field. We explain that, under the stated regularity assumptions, the resulting ODE reproduces the interpolant’s time marginals, although their full path distributions may differ. We have also added references supporting this formulation, including Lavenant et al. (2024).

In Section 3.2.1, we have expanded the motivation for RpriorR_{\mathrm{prior}}. After introducing the construction of endpoint-preserving interpolants, we explain that the endpoint and moment constraints generally leave multiple compatible intermediate evolutions. The regulariser specifies a preference among these possibilities. In particular, RlinR_{\mathrm{lin}} penalises deviations from the reference velocity; at a fixed endpoint coupling, minimising this penalty is equivalent to minimising the interpolant’s expected kinetic action. We also explain how alternative regularisers, such as the LAND prior, incorporate geometric information.

We have expanded the discussion of feasibility and clarified its distinction from existence of a minimiser and convergence of the numerical optimisation. For a single mean constraint, we provide a constructive feasible example: an endpoint-preserving translation can attain any prescribed mean in the unrestricted finite-energy interpolant family, assuming finite second moments of the endpoint distributions. For general nonlinear observations, feasibility depends on target compatibility and the chosen parameterisation. Our comparison results explicitly assume that a feasible minimiser exists; the augmented-Lagrangian updates do not guarantee convergence to such a minimiser. Numerical constraint satisfaction is therefore assessed through the attained residuals. The existence and duality conditions used for the constrained Schrödinger bridge are discussed separately in Appendix F.1.

We have also revised Section 3.3 to distinguish the established arguments on which we build from the comparisons specific to our formulation. Proposition 1 follows from the Cameron–Martin theorem and the chain rule for relative entropy. Its purpose is to interpret RlinR_{\mathrm{lin}} through an auxiliary Brownian-perturbed path law. Propositions 2 and 3 characterise the relationship between ideal solutions of our constrained-interpolant problem and the constrained Schrödinger bridge in Definition 2, under the stated assumptions and a common endpoint coupling. Proposition 2 controls the excess relative entropy beyond the best approximation available within the feasible interpolant family. Proposition 3 considers a constraint on the mean at a single intermediate time. Under the stated assumptions, it identifies the optimal interpolant and bounds the W2W_2 distance between its marginal distributions and those of the constrained Schrödinger bridge.

The purpose of these comparisons is to clarify the consequences of choosing different modelling formulations in an underdetermined inference problem. The constrained Schrödinger bridge serves as a reference for comparison; these results do not establish convergence of our numerical algorithm to that bridge or guarantee that either formulation better reconstructs the unknown process. We further explain this interpretation at the beginning of Appendix G.5, where we also discuss extensions to other endpoint couplings.

Comments on the paper

1 comment

Dear AC,

Thank you very much for your efforts in overseeing the review of our work. We are writing to summarise our discussion period. We have made a significant effort to address each point raised by the reviewers, and we believe in doing so our paper has substantially improved. Over the discussion period, we have made the following key changes to the manuscript:

  • Additional baselines. We have expanded the empirical comparisons to include SF²M, LightSB-M, and IPMF in Section 4.
  • Clearer presentation in Sections 2 and 3. We have revised both sections to better explain the problem setting, make the presentation of flow matching more self-contained, and motivate and justify our method. In particular, we now discuss the broader stochastic formulation of trajectory inference, explain the role of the prior regulariser, and discuss the feasibility of the generalised-moment constraints. We also distinguish established theoretical arguments from our method-specific contributions.
  • Extended theoretical results for different endpoint couplings. In Appendix G.5, we extend the comparison in Proposition 2 beyond the reference endpoint coupling. We provide a relative-entropy comparison for couplings satisfying KL⁡(π∣∣π∗)<∞\operatorname{KL}(\pi||\pi^*) \lt \infty, and a marginal Wasserstein bound for optimal-transport couplings under additional assumptions. These results clarify the scope of the connection between our method and the moment-constrained Schrödinger bridge.
  • Additional experiments, ablations, and sensitivity analyses. We have expanded the experimental evaluation in Section 4 and Appendix B:
    1. Constraints at multiple times. We tested generalised-moment constraints at three intermediate times rather than only at the midpoint. On the aerosol data, this reduces the average $W_2$error across these three times by approximately 9.5% for GMI-FM (lin) and 9.2% for GMI-FM (LAND), relative to midpoint-only constraints. These results demonstrate that the method can incorporate constraints at multiple times and that additional aggregate information can improve reconstruction.
    2. Hyperparameter sensitivity. We investigated sensitivity to the Stage-A learning rate and the effective constraint strength κ\kappa. Reconstruction performance is relatively stable across most tested settings.
    3. Training budgets and stage allocation. We examined the effect of increasing or decreasing the Stage-A and Stage-B training budgets, observing relatively modest changes in reconstruction performance across the tested settings. We also investigated how to allocate a fixed optimisation-compute budget between the two stages. GMI-FM (lin) is relatively insensitive to this allocation, while GMI-FM (LAND) benefits from allocating more compute to Stage A on Semrau.
    4. Updated Multi experiments. We provide updated Multi results using a revised checkpoint-selection procedure based only on W2W_2 at the final endpoint. This criterion is computed using the same cells available for endpoint training, without using intermediate marginals for checkpoint selection.
  1. Endpoint-coupling ablation. We compared endpoint-only OT couplings with moment-informed Schrödinger bridge couplings. The results show that the reconstruction benefit depends on the dataset and geometric prior, complementing the theoretical discussion of coupling choice.
  • Related work, citations, and evaluation definitions. We have moved the related work section into the main paper, expanded the explanations of the CSB-FM baseline and the W2W_2 metric, added missing dataset and theoretical references, and updated references to published versions where available.

Thanks again for your time and effort in this process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.