acceptodds
Under review as a conference paper at ICLR 2027

Beyond Gradient Flow: Identifiability and Recovery from Distribution Snapshots

Abstract

Inferring dynamics from snapshots of evolving distributions is fundamentally underdetermined: the Fokker-Planck equation constrains the drift only through its score-weighted divergence , leaving a -solenoidal gauge invisible to any single-time constraint. Time-indexed transport formulations cannot resolve this ambiguity: every admissible marginal path admits a curl-free explanation, minimum-action reconstruction selects it, and marginal fit alone cannot distinguish dynamically inequivalent explanations. Requiring one autonomous field to explain several marginals instead makes part of the hidden circulation visible through changing scores. Separating instantaneous Fokker–Planck source constraints from the snapshot experiment, we show that the source constraints identify the field modulo the kernel of a stacked score-weighted divergence operator. For generic Gaussian shape variation, source constraints at time points in intrinsic dimension eliminate every polynomial gauge direction, whereas finitely many density snapshots alone admit aliasing; we give the obstruction explicitly. Around an evolving Gaussian reference, for Sobolev smoothness and samples per time point, we derive a conditional oracle rate : the lower bound requires only a first-moment information scale, while the upper bound additionally requires a near-gauge tail condition. Strong-form fitting is non-orthogonal to score error and cannot be repaired by spectral filtering. Instead, we estimate using smooth test functions while retaining the known diffusion term, and derive a finite-sample bound that separates sampling error from fixed-grid quadrature bias. Planted-circulation experiments confirm the predicted gauge contraction and expose a design tension between cross-slice information and covariance-aware whitening.

Primary area:
learning on time series and dynamical systems
Keywords:
trajectory inference, identifiability, stochastic differential equations, snapshot data, non-gradient dynamics, Fokker-Planck equation, Stein operators, optimal transport, flow matching, minimax rates, method of moments

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EzgWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how to infer stochastic dynamics from distributional snapshots: given independent samples at several time points, without particle identities, can one recover the drift field, particularly its non-gradient (circulatory) component? The authors point out that the Fokker–Planck equation leaves an invisible circulatory degree of freedom at each time point. Consequently, time-indexed approaches such as optimal transport, flow matching, and Schrödinger bridges can perfectly fit the marginal distributions using gradient fields. The zero circulation they report is therefore a methodological choice rather than a discovery supported by the data. Under the autonomy assumption, requiring a single, time-independent drift field to explain all snapshots allows changes in distributional shape to reveal previously invisible directions. The authors prove that, for Gaussian snapshots with a common mean, K≥mK \ge m generic shapes eliminate all polynomial gauges, and that this threshold is sharp. The paper also derives conditional statistical rates in a linearized setting and proposes a weak-form estimator based on test-function moments, avoiding score estimation. All experiments use synthetic systems with deliberately introduced circulation.

Critical strengths

The problem addressed by this paper is important, and its conclusions have broad relevance. The authors place several mainstream approaches to snapshot-based dynamical inference within a unified framework and demonstrate theoretically, rather than merely empirically, why they cannot recover circulation. This has direct implications for how methods in this field should be evaluated.

The theorem showing that shape variation eliminates gauge freedom under the autonomy assumption (Theorem 8) is the paper's most valuable result. The threshold is both sufficient and necessary, and the counterexample for m=3m=3 coincides with Euler's equations for a free rigid body, making for an elegant and intuitive construction. The aliasing examples that distinguish source constraints from the information available in finite snapshots, as well as the analysis explaining why strong-form estimation is sensitive to score errors, are also clear and convincing.

Furthermore, the authors are unusually transparent about the status of their results, explicitly labeling the proof status of each claim and retaining all withdrawn claims. This level of transparency is uncommon in comparable work.

Critical weaknesses
  • The practical guidance is predominantly negative. Most of the paper's practical implications concern what cannot be done: marginal distribution fit cannot validate circulation recovery, score estimation can be unreliable, and zero circulation cannot be interpreted as a discovery. The few positive, actionable recommendations remain largely undeveloped. GLS whitening is theoretically motivated but produces no measurable improvement in the experiments. Observation 17 reveals that whitening and identifiability impose opposing requirements on snapshot spacing, and the authors explicitly acknowledge that they have not characterized when both can be satisfied. Thus, the practically important question of when snapshots should be collected remains unanswered. Proposition 15 provides a diagnostic for rejecting poor designs but offers little guidance on constructing good ones. A theory derives much of its value from its ability to guide experimental design and be tested against experimental outcomes. This feedback loop remains relatively weak in the present paper.

  • Birth and death (proliferation and apoptosis) dynamics are not considered. The entire framework assumes conservation of mass: each snapshot is a normalized probability distribution, and the proof of Proposition 1 explicitly relies on this assumption. However, the paper's primary application motivation comes from biological systems, where state-dependent proliferation and death are common in time-series single-cell data. Indeed, Schiebinger et al. (2019), cited by the authors, adopted unbalanced optimal transport for precisely this reason and explicitly estimated growth rates. This is more than a minor technical detail. State-dependent net growth enters the governing constraint equation in the same place as the drift signature, and an unknown growth rate can, in principle, absorb changes in that signature. This affects the identifiability of even the gradient component, let alone circulation. The paper does not discuss whether its central conclusions, including Theorem 8, continue to hold in this setting.

Questions for authors
  • If the net growth rate is state-dependent but time-independent, can the autonomy assumption together with shape variation identify both the drift and the growth rate? Or does the growth rate introduce additional gauge freedoms? Would the threshold K≥mK \ge m in Theorem 8 need to be modified in this setting?

  • Many systems are observed as stationary populations containing individuals at different stages of their life cycles, such as continuously renewing tissues (e.g., hematopoietic systems) or stellar populations containing stars at different evolutionary stages. These populations are maintained by a combination of birth, death, and individual evolution, without explicit time labels. Within the framework of this paper, stationary snapshots provide no shape variation, so Theorem 8 does not apply. How does this setting relate to the population-balance analysis of Weinreb et al. (2018)? What kinds of additional information would be sufficient to break the gauge freedom in this setting, such as known birth and death distributions or an internal clock associated with each sample (e.g., pseudotime, isotopic age, or RNA velocity)?

3 comments

Thank you for the detailed review. Your questions led to a new appendix (App. AB) and a new experiment (App. AA.4). Below we separate what is now proved, what we tested, and what remains open.

1. Birth and death (W2, Q1). App. AB treats a local net rate g(x)g(x) (division minus death), not nonlocal offspring kernels. Proved, at the level of the source constraints, like Thm. 8:

  • Known growth. The results of §2–§3.1 hold unchanged; known growth joins the source.
  • Unknown, state-dependent, time-independent growth. You are right that even the gradient part is at risk: growth absorbs the single-slice footprint, so only score differences between slices constrain the drift, and source constraints at K≤mK\le m times never identify it, not even its gradient component (Cor. 41). For Gaussian slices with a common mean, K≥m+1K\ge m+1 distinct generic shapes identify drift and growth jointly, and K=mK=m leaves an infinite-dimensional gauge (Thm. 42).
  • Proportions only, as in sequencing data: the same threshold for drifts bounded near the mean; growth is identified up to an additive constant, which proportions never fix (Cor. 45).
  • Time-dependent growth. If the growth class contains the compensating source terms, any drift, including zero, reproduces the whole marginal path (Prop. 46). Growth must therefore come from outside information, as in Schiebinger et al.

In short, for Gaussian shapes an unknown time-independent growth rate costs exactly one more snapshot shape, whereas unrestricted time-dependent growth makes the drift unidentifiable.

Your question also sharpened Thm. 8 itself. With K≥m+1K\ge m+1 generic shapes the source-level gauge is trivial among all weakly differentiable fields, not only polynomial ones (Rem. 43). At K=mK=m the polynomial class of Thm. 8 is essential: Rem. 43 gives a smooth, compactly supported gauge field. The main text says so after Thm. 8 and in the abstract.

Tested (App. AB.7). A linear system with quadratic killing shows the predicted gauge in the design matrix, and ignoring growth biases both parts of the drift (symmetric-part error 0.004 → 0.11, solenoidal error 0.033 → 0.105). With exact time integrals the joint fit recovers drift and growth.

Open. A practical growth-aware estimator. With the trapezoid rule at our snapshot spacing the joint fit is not yet usable (median solenoidal error 0.64 against 0.34 with known growth, n=200n=200k); a higher-order time quadrature is the natural route.

2. Stationary populations (Q2). With no shape variation and unknown growth, the source constraint determines nothing. With known birth and death rates and known diffusion, the setting of Weinreb et al. (2018), it determines the gradient part only; they remove the remainder by assuming a potential, a modelling choice rather than a consequence of the stationary density. On the additional information you list:

  • Pseudotime computed from the same profiles is a function of the state and adds nothing.
  • An independently measured age (e.g. isotopic) turns the population into age-indexed slices. If drift and exit rate do not depend on age, a strong and untested assumption, and the age-conditioned shapes vary enough, both are identified at source level (Lem. 50).
  • Per-cell velocities and lineage are Lagrangian measurements; they break the gauge but lie outside the snapshot model (App. R) and out of scope here.

Thus, in a stationary population, circulation is identified only with information beyond the snapshots, such as an independently measured age under the assumptions above.

3. Practical guidance (W1). We agree the submission was mostly negative here. The Discussion now gives three positive rules, each tested:

  • How many shapes: K≥mK\ge m distinct covariance shapes (K≥m+1K\ge m+1 beyond polynomial fields, Rem. 43); with fewer than mm, restrict the field class to degree at most K−1K-1 (App. X.3).
  • When to sample: rank candidate snapshot times by the covariance-only audit of Prop. 15 before fitting. Over 56 candidate designs its rank correlation with the realized error is 0.84, a pilot of 2,000 samples per time picks the same design in 8 of 8 draws, and the selected design halves the default's median error (0.49 against 1.01, eight paired seeds; App. AA.4).
  • Whether nn suffices: the fit's plug-in standard error predicts the error without ground truth (medians 1.46, 0.76, 0.34, 0.15 against realized 1.34, 0.79, 0.38, 0.15 at n=n= 10k to 1M; App. AA.2).

Together, these rules say how many snapshots to collect, when to take them, and whether the sample size suffices.

Open. Two limitations remain. The audit does not capture the quadrature bias of long gaps, so one should stop lengthening the span once its score levels off. Whitening shows no gain at any sample size we tested, and the efficiency–identifiability frontier of Obs. 17 remains uncharacterized.

We are grateful for these questions, which have strengthened the paper.

I thank the authors for their detailed response and the additional appendices and experiments. Regarding birth–death dynamics, the authors have provided a fairly comprehensive theoretical characterization. However, as they acknowledge, a practical estimator that accounts for growth rates has yet to be implemented (at the current snapshot intervals, the median error is 0.64 without the source term, compared to 0.34 when the growth rate is known). Regarding practical guidance, the newly proposed sampling design rules have only been tested on synthetic benchmarks, the underlying diagnostics are valid only for linear Gaussian models, and whitening has yet to demonstrate measurable benefits. Therefore, I believe the core concerns raised in my original review remain unresolved, and I have decided to maintain my original score.

Thank you for reading the revision and for the precise summary. We agree that a growth-aware estimator remains open: App. AB.7 locates the obstacle in the time quadrature rather than in identifiability (with exact time integrals the joint fit recovers drift and growth; with the trapezoid at our spacing it does not). The design audit rests on Gaussian covariances and is tested on synthetic benchmarks only. Whitening gives no gain, but the estimator does not depend on it: unweighted and GLS fits do not differ (§5).

Within that scope, the approach delivers three things:

  • It recovers circulation that optimal-transport-type methods provably cannot (relative error 1 at every nn, Cors. 18–19); its median error falls to 0.15 at n=1n=1M.
  • Its covariance-only audit, computed before any fit, selects a design that halves the default's median error (0.49 against 1.01).
  • On a real ocean field, from tracer positions alone, it recovers an eddy's rotation at 0.90 of the value from withheld velocities and rejects every autonomous gradient explanation (App. AA.6).

Thank you for the questions that led to App. AB and to the sharper statement of Thm. 8.

Reviewer_BVZYOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper regards the problem of recovering non-gradient dynamics from multiple snapshots in time. The authors present theoretical results characterizing the shortcomings of existing methods for recovering non-gradient flow (circulation blindness), and provides conditions for when non-gradient flow is identifiable. Finally, a novel method for recovering gradient flow is proposed which is tested on an experiment with simulated data.

Critical strengths

The paper appears to present some theoretical results that may be of interest to the PDE / statistical learning community. I have little background in this specific area so I am unable to check the accuracy of all the proofs.

Critical weaknesses
  1. The paper is very hard to read. I have a background in mathematics but have no experience working on this specific problem; it took me quite a while to get through this paper and it is very dense with jargon with no explanation. It is also not clear to me if this paper is a good fit for a machine learning conference in general.
  2. The main theoretical results of the paper seem to rest on a series of technical assumptions (9-11) which by the paper's own admission (appendix A) are unproven / conjectured; in addition, the paper includes a very large number of "retracted/superseded" claims which is very strange and confusing to me as a reader (App. Y.3). In addition, from my understanding the empirical results / performance of the proposed method are quite weak (rel. err 0.751 over 12 seeds).
Questions for authors

None

1 comment

Thank you for the careful review and for raising readability; the presentation is better for it.

Core claim. A method that fits a separate field at each time can match every snapshot with zero circulation, whatever the truth (Prop. 1, Cor. 2). Requiring one time-independent field makes circulation identifiable (Thm. 8) and estimable without scores (§4).

Readability. The abstract now starts from the methods it concerns and states the problem in plain words before the results. The first paragraph of the introduction ends with an intuitive picture (a round cloud of particles rotating about its centre: every particle moves, the density does not change) and defines ρ\rho-solenoidal at first use. Four tables of terms and symbols open the appendix, and each of Assumptions 9–11 has a one-sentence plain-language reading. We would welcome pointers to passages that are still hard to follow.

Fit to a machine-learning venue. The paper characterizes what a widely used family of learned models can and cannot identify from its data: optimal-transport, flow-matching, Schrödinger-bridge and action-matching methods for trajectory inference, many of them published at ICLR, ICML and NeurIPS. It is an identifiability analysis in the sense familiar from representation learning, followed by a statistical limit and an estimator. The same structure appears in learning dynamics: the stationary state of SGD carries persistent probability currents in parameter space (Kunin et al. 2024), which the distribution of the weights does not reveal.

Assumptions 9–11. They are used in exactly one place, Thm. 12, a local rate that complements the main results. Proved without them (App. A, status P): the blindness of time-indexed methods (Prop. 1, Cor. 2); identifiability under autonomy with its sharp threshold (Props. 4 and 6, Thm. 8); the estimator results (Props. 14–16); and the estimator's finite-sample bound (Thm. 37). Thm. 12 is stated as conditional, and Rem. 13 says what it does and does not establish. The contributions list and the opening of §3 now state this. The paper's central claims therefore do not depend on Assumptions 9–11.

"Rel. err 0.751". This number has to be read against the baseline. A relative error of 1 means that none of the circulation is recovered. Optimal-transport interpolation, exact-OT flow matching and Schrödinger bridges with a gradient reference sit at exactly 1 at every sample size, including infinite data (Cors. 18–19, Fig. 1b). The value 0.75 at n=40n=40k is therefore partial recovery of a component that these methods cannot recover at all, and the error keeps falling. Over 12 paired seeds (new App. AA.2):

samples per time point10k40k200k1M
OT / exact-OT flow matching / SB, population limit1.0001.0001.0001.000
ours (GLS, median)1.340.790.380.15

The decay is consistent with n−1/2n^{-1/2}, and the fit's own standard error tracks it without ground truth (predicted medians 1.46, 0.76, 0.34, 0.15). In short, our estimator recovers circulation that these methods provably cannot, with an error that decreases steadily as nn grows.

Retracted claims (App. Y.3). We agree that the list was confusing in a submission. It recorded corrections we made to our own code and claims during development, all before submission. The section is now "Development corrections": it keeps the two items that changed reported numbers and explains each for a reader without the history; the complete record is in the supplement. The claim-status table of App. A is unchanged.

Reviewer_g75mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the identifiability and recovery of stochastic dynamics from unpaired distribution snapshots. The authors show that recovering the drift field from marginal distributions is fundamentally underdetermined because the Fokker--Planck equation leaves a ρ\rho-solenoidal gauge that is invisible at a single time point. Consequently, time-indexed transport methods, including minimum-action and optimal-transport approaches, may perfectly match the observed marginals while missing the true rotational dynamics. To reduce this ambiguity, the paper assumes an autonomous drift shared across all time points and analyzes identifiability through stacked Stein operators. For generic Gaussian shape variation, sufficiently diverse snapshots can eliminate polynomial gauge directions. The authors further derive a conditional statistical recovery rate using a tangent snapshot model. For estimation, they propose a score-free weak-form method based on It^{o}'s formula and empirical moments, combined with covariance-aware generalized least squares. Synthetic experiments validate gauge contraction and reveal a trade-off between temporal excitation and whitening efficiency.

Critical strengths

1.The paper studies the identifiability of underlying stochastic dynamics from unpaired distribution snapshots, rather than focusing only on marginal matching.

2.The paper characterizes the ambiguity induced by the ρ\rho-solenoidal gauge and shows that accurate marginal fitting does not necessarily imply recovery of the true dynamics.

3.The paper distinguishes source-level identifiability from finite-snapshot identifiability, clarifying the difference between instantaneous source constraints and information available from discrete snapshots.

Critical weaknesses
  1. It remains unclear whether recovering dynamics different from the unique ground-truth process constitutes a practical limitation. More concrete application scenarios are needed to demonstrate when identifying the true underlying dynamics is necessary.

2.The theoretical conclusions are mainly validated on synthetic systems with known ground-truth dynamics. This makes it possible to directly measure dynamical recovery, but such ground truth is generally unavailable in real-world snapshot data.

3.The paper therefore does not provide direct real-world evidence that the proposed method recovers dynamics closer to the true process than existing methods. Additional validation strategies beyond marginal consistency or other indirect metrics would strengthen the empirical support.

Questions for authors
  1. For many OT-, flow-matching-, or generative-transport-based tasks, obtaining a dynamical field that accurately reproduces or predicts distributional evolution may already be sufficient. Could authors explain the practical necessity of recovering the ``true dynamics'' ?

2.All experiments are conducted on synthetic systems with known ground-truth dynamics. How can the proposed method be validated on real-world snapshot data where the true dynamical field is unobservable?

  1. Are there realistic validation strategies, beyond marginal consistency or indirect metrics, for assessing whether the recovered dynamics are closer to the true process than those obtained by existing methods?

4.The identifiability analysis relies on the autonomy assumption that a single dynamical field governs all observed time points. In which target applications is this assumption expected to hold?

5.How robust is the proposed approach to violations of the autonomy assumption, and is there any empirical evaluation under time-varying dynamics?

2 comments

Thank you for the review and for its direct practical questions; each led to a new experiment.

Core claim. A method that fits a separate field at each time can match every snapshot with zero circulation, whatever the truth (Prop. 1, Cor. 2). Requiring one time-independent field makes circulation identifiable (Thm. 8) and estimable without scores (§4).

1. Why recover the dynamics if the marginals are fit? (W1, Q1) Marginal fit reproduces the observed population path. It does not determine how individual particles move, or how a different or perturbed population would evolve. New §5.4 measures this by your own criterion, predicting distributions. On our benchmark the gradient field that reproduces every marginal exactly (the infinite-data limit of a minimum-action fit) mispredicts

  • the mean destination of a particle between consecutive observation times, by 13.5% of the true displacement;
  • the marginals of a population started from a different initial state: its error is 38% of that of reusing the training marginals.

More data cannot change either number. The weak-form estimate reduces them to 7.0% and 24% at n=200n=200k (12 paired seeds; App. AA.1). With strong 2D rotation the held-out error is 1.98 for the gradient field against about 0.10 for the weak form (Fig. 9).

Once a population is stationary its density no longer changes, so the marginal says nothing about motion: all motion that persists is circulation. Whether it is present is the question of equilibrium (detailed balance) versus an actively driven steady state, and its size sets the entropy production (Battle et al. 2016; Gnesotto et al. 2018). The same holds in machine learning: the stationary state of SGD is out of equilibrium, with persistent currents in parameter space (Kunin et al. 2024; Adhikari et al. 2023), so the stationary distribution of the weights does not determine the training dynamics.

Recovering circulation therefore matters whenever the goal is to predict beyond the observed path (for example, how a perturbed population evolves) or to quantify nonequilibrium behaviour.

2. Validation (W2, W3, Q2, Q3) We never use marginal consistency as evidence of recovery. Prop. 1 and Cor. 2 make this a property of the problem, not of our method: a zero-circulation explanation fits any marginal path, so marginal fit cannot tell right from wrong circulation for any method. Every recovery number in the paper is instead the error of the recovered circulation against a planted truth, the only setting in which recovery can be measured directly.

On real data, the Discussion gives two routes using withheld information and one falsification test:

  • Lagrangian measurements (lineage barcodes, metabolic labelling, RNA velocity);
  • held-out conditions (another start, a perturbation), the quantity measured in §5.4;
  • falsification: if no autonomous gradient field explains all slices, non-gradient structure is required, a test time-indexed models cannot run.

The fit also reports its own uncertainty without ground truth: its plug-in standard error predicts median errors 1.46, 0.76, 0.34, 0.15 at n=n= 10k to 1M, against realized 1.34, 0.79, 0.38, 0.15 (App. AA.2). On real data it tells whether an estimated circulation exceeds sampling noise under the model; it complements the routes above. A real-data study along these routes needs such measurements and is a separate project.

Against existing methods: the exact OT displacement interpolation fits every marginal and has relative error 1.000 (no circulation recovered), as Cors. 18–19 predict at every sample size; trained OT-coupled and Schrödinger-bridge flow-matching runs stay between 0.98 and 1.03 (App. AA.5).

In summary: marginal fit cannot validate circulation for any method; we give routes that can, a reliability estimate without ground truth, and a baseline that provably recovers none of it.

3. Autonomy (Q4, Q5) Assumption 3 concerns the state representation, not time-invariance of the biology (App. X.1): if the variables that drive the dynamics are part of the state, the field is autonomous; omitting one produces an apparent time dependence. Autonomous fields are standard in landscape descriptions of cell state, reaction networks with fixed rates and the diffusion approximation of constant-step-size SGD. The assumption can also be tested from the data, which a time-indexed model cannot (§2.2). On robustness, App. W already gives two exact hidden-state cases. New App. AA.3 tests explicit time variation: a drift in the relaxation rate, which changes the marginal shapes, is detected by the fit's lack-of-fit statistic on every seed, already at one tenth of the rate. A time-varying rotation rate alone can be absorbed by a constant rotation and is not flagged.

The review rates soundness 1. If there is a specific derivation or experiment you believe is incorrect, we would be grateful to be pointed to it so that we can correct it.

Follow-up: validation on real ocean currents (W2, W3, Q2, Q3)

We have now validated the method on real data, following the strategy proposed in our first reply: withheld measurements and a falsification test (new App. AA.6; pre-registered before any fit). We use the data released with Curly-FM (Petrović et al., NeurIPS 2025): tracers advected by surface currents of the HYCOM ocean reanalysis around a Gulf of Mexico eddy, observed as 9 snapshots of 111 particles, with no particle linked across time. No velocity information enters the fit: the estimator uses particle positions alone, and the velocities of the flow are reserved for validation.

From the snapshots alone, our estimator recovers the eddy's rotation. All values below are computed by us on these data. Only the reference row uses the velocities; every other row is fitted to particle positions alone.

MethodFitted toRotation (vorticity)Fraction of reference recovered
Reference: affine fit to the withheld velocitiesvelocities−1.941
Ours: weak form, trapezoidal rule (as submitted)positions only−1.750.90
Ours: weak form, exact time integration between snapshotspositions only−1.930.99
Time-indexed: trained OT-CFM networkpositions only−1.19 to +0.87 (sign changes across snapshots)none consistent
Zero-circulation: best autonomous gradient fieldpositions only00
  • Agreement with the withheld velocities. The velocities of the recovered field match the withheld ones (median cosine 0.994), and the recovered rotation is 119 bootstrap standard errors from zero. The 10% gap of the submitted estimator is consistent with the time-discretization bias that Thm. 37 separates from sampling error, since integrating exactly between snapshots removes it.
  • Falsification. The data reject an autonomous zero-circulation explanation: the best autonomous gradient field leaves 65 times the residual of the fitted field, and its velocities align poorly with the withheld ones (median cosine 0.27, against 0.994).

This test covers one two-dimensional flow, with tracers advected in a reanalysed velocity field; App. AA.6 states the pre-registered criteria.

This is the validation beyond marginal consistency that the review asked for: on a real flow, the circulation recovered from unlinked snapshots agrees with the withheld velocities, and an autonomous gradient explanation is rejected by the data. We hope this addresses the main concern of the review, and we would be grateful if it could be taken into account in the final assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the three reviewers for their careful reading; each review led to a concrete addition. Changes in the revised PDF are marked in teal. No theorem, assumption or appendix has changed its number; new material is in App. AA (experiments, including real data) and App. AB (birth and death).

Core claim. From snapshots alone, the circulating (non-gradient) part of the dynamics is invisible to any method that fits a separate field at each time: such a method can always match every marginal with zero circulation, and optimal-transport, minimum-action and Schrödinger-bridge formulations with a gradient reference select exactly that explanation, whatever the truth (Prop. 1, Cor. 2, Cors. 18–19). Requiring one time-independent field to explain snapshots whose shape changes makes circulation identifiable, with a sharp threshold (Thm. 8), and we give a score-free estimator with a finite-sample guarantee (Thm. 37).

What the revision adds

  1. Presentation (BVZY, g75m). The abstract now starts from the methods it concerns, states the problem in plain words, and ends with what follows for evaluating trajectory inference. The introduction adds an intuitive picture (a rotating cloud of particles: every particle moves, the density does not change), defines ρ\rho-solenoidal at first use and poses the identifiability question for learned trajectory-inference models. Four tables of terms and symbols open the appendix.
  2. Which result needs which assumption (BVZY). Thm. 8 is proved without further assumptions, as are the blindness results (Prop. 1, Cor. 2) and the estimator's finite-sample bound (Thm. 37). Assumptions 9–11 enter only the local rate of Thm. 12. The contributions list and the opening of §3 now say so.
  3. What marginal fit misses (g75m; new §5.4, App. AA.1). A review notes ‘direct implications for how methods in this field should be evaluated’; §5.4 measures them. The gradient field that reproduces every marginal exactly still mispredicts the mean destination of a particle by 13.5% of the true displacement, at any sample size, and on a held-out initial population its error is 38% of that of reusing the training marginals. The weak-form estimate reduces these to 7.0% and 24% at n=200n=200k (12 paired seeds).
  4. Sample size (BVZY; App. AA.2). Over 12 paired seeds the median error is 1.34, 0.79, 0.38, 0.15 at n=n= 10k, 40k, 200k, 1M, against exactly 1 (no circulation recovered) at every nn for optimal transport and the flow-matching and Schrödinger-bridge formulations of Cors. 18–19. The fit's own standard error, computed without ground truth, predicts these medians (1.46, 0.76, 0.34, 0.15).
  5. Design guidance (EzgW; App. AA.4). Across 56 candidate snapshot-time designs, the covariance-only audit of Prop. 15, computed before any fit, has rank correlation 0.84 with the realized error; the design it selects halves the default's median error (0.49 against 1.01, eight paired seeds).
  6. Autonomy (g75m; App. AA.3). A time-varying relaxation rate, which changes the marginal shapes, is detected by the fit's lack-of-fit statistic on every seed, already at one tenth of the rate. The appendix also characterizes the violations this test cannot flag (a time-varying rotation rate).
  7. Baselines (g75m; App. AA.5). On the benchmark, the exact optimal-transport displacement interpolation fits every marginal and recovers no circulation (relative error 1.000), as Cor. 18 predicts; trained OT-coupled and Schrödinger-bridge flow-matching runs stay between 0.98 and 1.03.
  8. Birth and death (EzgW; new App. AB). With known growth the identifiability results hold unchanged. With unknown time-independent growth the source-level Gaussian threshold of Thm. 8 becomes K≥m+1K\ge m+1, sharply. Unrestricted time-dependent growth leaves the drift unidentifiable from marginals, which is why growth has to come from outside information.
  9. Real data (g75m; new App. AA.6). On the ocean data of Curly-FM (tracers in a real HYCOM reanalysis velocity field around an eddy), the estimator, fitted to particle positions only, recovers the eddy's rotation at 0.90 of the value computed from the withheld velocities (0.99 when time is integrated exactly between snapshots), and rejects every autonomous gradient explanation. The analysis was pre-registered.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.