acceptodds
Under review as a conference paper at ICLR 2027

Flow Maps That Compose Exactly: When Consistency Costs Accuracy

Abstract

Learned dynamics are often required to compose across time: a direct forecast should agree with a sequence of shorter forecasts. Deterministic dynamics on a sufficient state and Markov transition laws obey this identity, but conditional means, the targets of squared-error training, generally do not. We quantify the price of this mismatch. Under coverage and regularity assumptions, an exactly composing predictor incurs joint excess risk at least proportional to the squared composition defect of the true means. Matching examples attain the smoothness rate and small-noise risk order. A generator criterion identifies diffusions whose means compose, including nonlinear, noisy cases. Transformed linear diffusions provide compatible kernels with explicit representation limits. Fixed-model experiments isolate the forecasting consequences. On noisy Lorenz dynamics, propagating a residual distribution through the same fitted one-step map removes most of the long-horizon excess risk of its point rollout; a single intermediate collapse to the mean restores it. Without noise, autoregression is highly accurate. On motion capture, discarding a learned kernel's intermediate uncertainty degrades its forecasts, although a deterministic map remains the best point predictor. The guiding distinction is what should compose: under uncertainty, transition laws compose even when their means do not.

Primary area:
learning on time series and dynamical systems
Keywords:
flow maps, compositional consistency, conditional expectation, mean-composition defect, Chapman-Kolmogorov equation, stochastic dynamical systems, uncertainty propagation, autoregressive forecasting, transition kernels, ensemble forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NeSvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the potential incompatibility between conditional-mean prediction under squared loss and exact pointwise composition. Under coverage and Lipschitz assumptions, it relates composition defects to excess prediction risk, studies sharpness and conditions for mean composition, and constructs Chapman–Kolmogorov-consistent kernels using affine Gaussian latent dynamics.

Critical strengths
  1. Propositions 1 and 5 establish sharp smoothness dependence and matching small-noise risk orders in specific examples, strengthening the analysis beyond a qualitative observation about non-composing means.
Critical weaknesses
  1. The paper provides a careful characterization of the incompatibility between conditional-mean prediction and exact composition. However, its practical implications beyond this characterization remain less established. In particular, the restricted kernel family in Section 4 illustrates a compatible distributional construction, but the evidence for choosing this particular construction for point forecasting is limited.

  2. The quantitative applicability of the risk bound remains unclear beyond the controlled examples. Although the paper discusses the coverage and regularity assumptions and evaluates the constants in the oracle setting, it provides limited guidance on when these conditions yield an informative bound for realistic learned predictors.

Questions for authors
  1. How do the risk bound and sharpness results extend the recursive-forecasting bias analyses? What is the main additional theoretical insight?

  2. For point forecasting, what advantages or trade-offs does the kernel construction in Section 4 offer relative to the mean-preserving target

3 comments
AuthorsOpenReview ↗

Beyond recursive bias, and when the bound is informative

Thank you for reviewing our work, for your thoughtful questions, and for highlighting Propositions 1 and 5. Below, we address each of your points.

1. The main additional insight (Q1). Recursive-bias analyses show that iterating a given one-step model yields a biased multi-step forecast. Our result is that this bias cannot be trained away while the predictor still composes: it belongs to the target, not to the estimator. When the true means do not compose, every predictor built to compose (rollouts, flow maps, consistency-trained models, conjugate architectures) must carry the defect somewhere: on the data, in its sensitivity, or at intermediate means that training never visits (Theorem 1, eq. 5). Propositions 1 and 5, which you highlighted, show that the trade-off is exact in its rates: L2R∗(L)→27/64L^2 R^\ast(L) \to 27/64 in smoothness and Θ(ϵ2)\Theta(\epsilon^2) in noise. Training a composing network through its rollout illustrates it: the lag-1.6 excess falls from 60.4 to 53.7 while the lag-0.05 excess rises 75–173-fold, and the direct network's 8.14 stays out of reach (App. L, p. 38).

Three further results make this usable:

  • How much (Corollary 1). Iterated one-step means follow the noise-free flow, so a rollout's long-horizon price can be computed before training. Fitted rollouts were within 0.83–1.09 times this value at pre-specified noise levels (Fig. 11, p. 43).
  • When it is zero (Propositions 2 and 3). The means compose exactly when a:∇2Φh≡0a:\nabla^2\Phi_h \equiv 0, with local defect −h22a:∇2b-\tfrac{h^2}{2} a:\nabla^2 b. Some non-affine noisy systems satisfy this; others do not.
  • What it means for consistency training (eq. 10, App. C.4). A target evaluated at the model's own predicted state has the composed means as its population minimizer; evaluated at an observed state, it keeps the true means as a fixed point. With one network per lag, heads trained the second way are 13 to 60 times more accurate than heads trained the first way on Lorenz-63 (1.3 times without noise), and about 15 times on Lorenz-96 (App. M, p. 40; Tables 17 and 22, pp. 42 and 48).

2. When the bound is informative (W2). Two conditions: the target defect exceeds the one-step error, and the data cover the intermediate means. The second can be checked in simulation through the effective coverage ratio, the error at the intermediate means divided by the error on data (Section 2, p. 4).

  • Where both hold, the bound is tight to a small factor. The certified bound is 1.9 times below the best composing fit in the closed-form diffusion (Table 37, p. 66). Its plug-in value, used as a diagnostic, is within 1.8 times the measured risk for trained flow maps on Lorenz-63 and 40-dimensional Lorenz-96, where the ratio is 0.76–1.43 (Apps. O and P).
  • Where they fail, it says little. At native lags the one-step error dominates (App. Q.3). In the 4096-dimensional Navier–Stokes system the ratio rises to 6.2–7.4 and the error moves to the intermediate means, as eq. (5) allows (Table 18, p. 44). A short construction in Section 2 (p. 4) shows that coverage is necessary, as Proposition 1 shows for regularity.

3. The kernel and the mean-preserving target for point forecasting (Q2, W1). You are correct that the evidence does not single out the closed form of Section 4 for point forecasting, and Section 4 now says that forecasting does not require it (p. 7). Where the dynamics have this form, its advantage is an exact likelihood at every lag, including unseen ones: on warped OU, excess error 0.001 at the held-out lag 1.5, against 0.148 for a direct network (Table 2, p. 25). Its limit is scope: with additive noise, only affine drifts admit it (Propositions 6 and 8). The two practical options meet different needs. On Lorenz-96 at lag 1.6, with the point rollout and direct heads as references:

MethodExcess risk (lower is better)UncertaintyNetwork calls per forecast
Point rollout (reference)268No16
Direct head (reference)9.25No1
Mean-preserving target, eq. (10)8.15No1
Propagated kernel0.285Yes16 × 512

The target needs observed intermediate states and one head per lag; the kernel needs a single one-step model with a noise model, and its value removes Monte Carlo noise (1.07 without this correction). The target therefore suits point forecasts at fixed horizons, while the kernel adds accuracy and uncertainty estimates at a much higher inference cost. A fitted kernel can also improve existing models without retraining, by 25–40% for U-Net and FNO on a measured cylinder wake (Figure 5, p. 9), though tuned rollout filters match or outperform it on FSI, for CNO, and in most of the broader benchmark (Apps. S.1–S.2). Figure 28 (p. 73) summarizes these trade-offs.

We hope these responses address your concerns and welcome any further discussion on these points.

Thank you for your thorough and thoughtful rebuttal. I appreciate the effort you have put into addressing my questions and concerns. I was impressed by the breadth and diversity of the experimental evaluations, as well as the detailed clarifications provided in your response. Your rebuttal has addressed most of my concerns, and I will therefore raise my score by one point.

Thank you again for your dedication and the substantial effort invested in this work.

AuthorsOpenReview ↗

Thank you for the updated assessment

Thank you for taking the time to read our response and for raising your score. Your positive feedback means a lot to us.

Your review helped us improve the paper in several ways. We are also glad the breadth of the experiments came through: it reflects the generality of the result, which applies to any predictor built to compose.

Thank you once more for your time and for providing a constructive review.

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

This paper studies a mismatch between temporal consistency and point prediction under stochastic dynamics. Markov transition laws compose, but MSE-optimal conditional means generally do not, so an exactly composing point predictor can incur extra prediction error under the paper's coverage and regularity assumptions. The paper gives a risk-consistency bound, characterizes some cases where conditional means do compose, and constructs a family of stochastic transition kernels with exact Chapman-Kolmogorov consistency. Experiments on synthetic systems, noisy Lorenz dynamics, and CMU motion capture study both exact composition and the effect of propagating uncertainty instead of repeatedly collapsing it to a mean.

Critical strengths

The paper makes a useful distinction between composing point predictions and composing transition laws. The theory around it is good. I checked the three-state chain bounds, the 27/64 constant, and the second-order local defect and they are valid to me. The experiments and fixed-model interventions are also well designed, which isolate the propagation rule from refitting and capacity differences.

Critical weaknesses

The gap between autoregression and stochastic AR(which corresponding with Monte Carlo prediction) is a classical topic in control/dynamics. Theorem 1 is the new component author introduced, which is a four-term triangle inequality where 2 constants carry most but the constants can be hard to know and judge, not really sure what I can learn from this bound. Also the experiments seem also not strong enough to show this cost actually shows up in practice.

Questions for authors

In CMU motion experiments, the exactly composing deterministic map gives the best MSE. And this should be the most important experiment since it is practical problem but not strong enough. Even author say "On motion capture, discarding a learned kernel's intermediate uncertainty degrades its forecasts, although a deterministic map remains the best point predictor." and "The sweep therefore does not exhibit the adverse risk–consistency frontier of the analytic example.......this dataset does not establish its nonzero-defect or coverage assumptions", sort of writing, defensively admit. I would suggest author consider other benchmark or pratical problem like autonomous driving or multi-agent trajectory forecasting.

Theorem 1 can become quite weak when C and L is large, and Proposition 1 shows that the excess risk can approach zero if the regularity constraint is removed. I understand you can control or bound L through model architecture while C is quite tough even impossible in high dimensional space, do authors know any tricks on this or say how should I understand this bound correctly with a uncontrollable constant C?

I think continuous time model like Neural ODE is a more appropriate subject to analysis under the theory. i.e. train a neural ODE with MSE on noisy Lorenz, compute Δ by simulation, and check whether the model’s long-horizon excess risk tracks as (22) suggests. If it does, the theorem explains a failure mode people actually run into. If it doesn’t, that tells us the RHS terms dominate in practice.

Seems quite a lot defensive writing in the paper, will make readers confused.

3 comments
AuthorsOpenReview ↗

Added suggested experiments showing the cost of composing means

Thank you for reviewing our proofs and suggesting two specific experiments, which we carried out. In the neural-ODE test, the long-horizon excess follows the defect, matching the scenario you described as “a failure mode people actually run into.” Corollary 1, not the less precise right side of (22), explains this correspondence.

1. Neural-ODE test (Q3): the long-horizon excess tracks the defect. We trained neural ODE and direct networks of the same size, each with about 34,000 parameters, on three independently generated noisy Lorenz datasets. At lag 1.6, the neural ODE’s root rollout excess differs by an average of 9% from the true defect Δ=8.06\Delta = 8.06 (Table 15, p. 40). Its excess risk, 60.4 ± 8.8, matches the value from Corollary 1, 60.3 ± 6.0, without fitted parameters (Fig. 3, p. 8). The direct network achieves 8.14 ± 1.83, but without noise, the order reverses (neural ODE 0.64, direct 39.1). The right side of (22) overestimates this gap because of the largest sampled Lipschitz ratio and cannot certify the match. Corollary 1 explains it: an accurate one-step mean model follows the noise-free flow, whose distance from the true mean is the price. Fitted rollouts stayed within 0.83–1.09 times that price across the pre-specified noise levels (Fig. 11, p. 43). When propagated as a law, with residual noise and 512 particles, the same network reduces its excess to 0.458, and its total squared error, 112.1, is lower than the direct network’s 119.7.

2. Multi-agent driving benchmark (Q1). Using the released QCNet on 2,000 held-out Argoverse 2 scenarios, composing the model’s own mean forecasts over 1-second segments increases its 6-second squared error by 13.6 m² compared to its direct forecast (95% interval: 10.5–16.9). Composing its law instead recovers 3.1 m² after Monte Carlo correction (App. T, pp. 58–61). This recovery is only partial, since generated histories still change the model’s inputs. On the CMU dataset, the deterministic map remains the most accurate point predictor, so we use it to show the cost of collapse rather than composition. Simulated systems show the cost at scale: enforcing exact composition on a trained flow map increases risk by 2.1 times on Lorenz-63 and by 13.4 times on 40-dimensional Lorenz-96, against 0.92 and 0.26 times with the noise switched off (Apps. O and P).

3. Reading the bound when CC cannot be controlled (Q2). Theorem 1 is most useful through its first step, equation (5): a composing predictor cannot remove the defect, only determine where it appears (on the data, in its sensitivity LL, or at the intermediate means, which are not seen during training). As you noted, LL can be controlled by architecture, but CC cannot be certified in high-dimensional settings, and CC decides if the final term reaches the data. A practical metric is the effective coverage ratio: the error at the intermediate means divided by the error on data, which can be measured whenever the system is simulatable. For Lorenz-63 and Lorenz-96, this ratio stays between 0.76 and 1.43; the cost shows up on the data (and on Lorenz-96, also at the means), and the bound’s plug-in value, used as a diagnostic rather than a certificate, is within 1.8 times the measured risk. In the 4096-dimensional Navier–Stokes system, the ratio rises to 6.2–7.4: the data-state risk drops from 0.301 to 0.247, while the map becomes 12–19 times less accurate than its unconstrained counterpart on smooth, mean-like inputs (Table 18; App. O.1). Coverage is also needed, much as Proposition 1 shows for regularity: if the intermediate means fall outside the data distribution, and different states have different means, a composing predictor can map each intermediate mean directly to its two-step mean, making all three excess risks vanish while Δ>0\Delta > 0. We have added this clarification to Section 2 (p. 4), as it shows how, in high dimensions, consistency can shift error onto inputs not found in the data.

4. Classical background. We concur that the distinction between deterministic and Monte Carlo prediction is well established, and the paper acknowledges this. Previous analyses compare forecasts from a given model, whereas our results restrict every predictor made to compose, either by architecture or training, and introduce a quantifiable price that can be determined before training. Your neural-ODE test shows the restriction at work: training the same network through its rollout lowers its lag-1.6 excess from 60.4 to 53.7 while raising its lag-0.05 excess 75–173-fold, and the direct network’s 8.14 stays out of reach (App. L, p. 38).

5. Writing (Q4). Thank you for your honest feedback. We have made the changes you suggested. The introduction now separates classical intuition from new contributions, each result is stated clearly in technical language, and the limitations are now gathered in Section 6.

We hope these responses address your concerns and are happy to discuss further.

Reviewer_WKrvOpenReview ↗

Reviewer Followup

Thank you for the response and experiments. My remaining reservation is that the quantitative usefulness of the bound in high-dimensional real systems. I will raise my score to 3.

AuthorsOpenReview ↗

Thank you, and what the bound says in high-dimensional systems

Thank you for reading our response and for raising your score. We deeply appreciate it.

Your review changed this paper for the better. The neural-ODE test you proposed led us to Corollary 1, and your suggestion of a driving benchmark became the QCNet study.

On your remaining reservation, we agree. In high-dimensional real systems the coverage constant CC of Theorem 1 cannot be certified, so there the bound only says that a composing predictor must place the defect somewhere: on the data, in its sensitivity, or at the intermediate means (eq. 5). Two quantities need no constants and remain directly useful: the price of Corollary 1, computable before training wherever a simulator exists (within 3% of the fitted rollout on 40-dimensional Lorenz-96; Table 22, p. 48), and the cost measured on a trained model by composing it with itself (13.6 m² on QCNet; all 24 cases of our main RealPDEBench study; Apps. S and T). Section 6 already notes that large constants weaken the bound, and we will state this more explicitly in the final version.

Thank you again for the careful and constructive review.

Reviewer_ZP2COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper looks at learned maps that should compose over time. A direct forecast equals a chain of shorter ones. The point is that with MSE training the target is the conditional mean, and conditional means of a noisy nonlinear process usually do not compose, while the transition distributions do. The main result is a lower bound: a predictor that composes exactly must have extra error proportional to how much the true means fail to compose, assuming a coverage condition and a Lipschitz bound. There are toy examples showing the bound is tight, a condition for when diffusion means do compose, and a transformed linear diffusion model whose distributions compose exactly. Experiments are an oracle toy, noisy Lorenz, where running an ensemble through the same one step map is much better than a point rollout, and CMU motion capture, where collapsing the ensemble to its mean hurts but the deterministic map is still best on MSE.

Critical strengths
  • clear and correct main message: compose distributions not means
  • theory is careful and the toy examples show the bound is not vacuous
  • nice examples that noise plus nonlinearity does not always break mean composition
  • clean experiments that keep the fitted model fixed and only change how uncertainty is passed forward
  • honest about mixed results
Critical weaknesses
  • the main idea is known, direct vs recursive forecasting bias and ensemble forecasting are classical
  • Theorem 1 is basically a triangle inequality, the coverage constant can grow exponentially with dimension
  • the title says consistency costs accuracy, but this only shows on toy examples
  • framed around flow maps and consistency models but none are tested, and for deterministic transport the bound gives no cost anyway
  • Lorenz result is expected
  • the proposed kernel is not strong, loses on some tasks and is close to some existing models
  • small experiments
Questions for authors
  • can you show a real setting where exact composition clearly hurts compared to a matched model that does not compose?
  • does this problem show up in an actual flow map or consistency model?
  • what should a practitioner actually do, should we train the compatible kernel always?
2 comments
AuthorsOpenReview ↗

Composing means costs accuracy in released forecasters and flow maps, and how to avoid it

Thank you for your careful review and for recognizing our main distinction (composing laws instead of means) and the value of fixed-model interventions. Your three questions guided our new experiments, and we address them below.

1. Real setting with a matched model that does not compose (Q1). The released QCNet offers a real-data comparison: one network forecasts 6 seconds directly, or composes its own mean forecasts over 1-second segments. On 2,000 held-out Argoverse 2 scenarios, composition increases squared error from 36 to 49.6 m², a cost of 13.6 m² (95% interval 10.5–16.9). Composing the law instead recovers 3.1 m² after Monte Carlo correction (App. T, pp. 58–61). For measured flows, composing each of the twelve measurement-trained RealPDEBench checkpoints in our main study with itself raises the error in all 24 checkpoint–stride cases, by 1.7–44%, and in 110 of 118 cases across the wider benchmark, short of the 113 we had pre-specified (App. S). Real data mix this cost with generated-history and model errors, so the clean isolation is in point 2.

2. Actual flow-map and consistency models (Q2). Yes for forecasting flow maps under noise; no for generative ones.

  • A trained flow map, forced to compose. We take a trained lag-conditioned flow map F(x,k)F(x, k), impose F(x,2k)=F(F(x,k),k)F(x, 2k) = F(F(x, k), k) exactly, and fine-tune it and an unconstrained copy for the same number of updates. With noise, the constraint multiplies the three-pair risk by 2.1 on Lorenz-63 and by 13.4 on 40-dimensional Lorenz-96. With the noise switched off, the same constraint gives 0.92 and 0.26: no cost (Apps. O and P; Table 21, p. 47). The cost therefore comes with the noise, not from optimizing under a constraint. All eight pre-specified Lorenz-96 predictions held; one Lorenz-63 prediction did not (a four-times-wider network: 4.70 with noise, 1.23 without). In the 4096-dimensional Navier–Stokes system, the constraint instead shifts the error to the intermediate means, as eq. (5) allows (Table 18, p. 44).
  • A flow-matching forecaster. Along its physical axis, its composed one-step mean incurs 58.9 ± 6.6 against the computed price of 60.3 ± 6.0, while its propagated law incurs only 0.670 (App. X).
  • Self-consistency penalties. These are a weaker test: no penalty brought the defect of our Navier–Stokes flow map below that of the true means. A penalty that bootstraps from the model's own predicted state has the composed means as its population target (App. C.4); at weight 100 it raised excess risk from 0.27 to 3.39 at t=1.6t = 1.6 without lowering the defect (Table 25, p. 52). Bootstrapping from observed states keeps the true means as a fixed point: with one network per lag, such heads match or beat direct heads and are 13 to 60 times more accurate than heads bootstrapped from predictions on Lorenz-63, and about 15 times on Lorenz-96 (Tables 17 and 22, pp. 42 and 48).
  • Generative transport. As you noted, it incurs no cost at the target level, and a released LSUN-cat consistency model confirms it (App. Y).

3. Practitioner guidance (Q3). You do not need to fit the closed form from Section 4; its purpose is to show that exact compatibility is possible. In practice:

  • When the dynamics are deterministic on the modeled state, or their means compose as stated in Proposition 2, the cost disappears, so you can enforce composition as needed (Apps. W–Y).
  • If there is noise, for point forecasts at fixed horizons, fit a separate head for each horizon, either directly or using observed intermediate states. On Lorenz-96 at lag 1.6, these reach 9.25 and 8.15, compared to 268 for the rollout.
  • If there is noise and the model must be rolled out, provide it with a noise model, propagate samples, and average only at the end. On Lorenz-96, this removed 99.9% of the rollout's excess, though at 512 times the inference cost. Do not collapse to the mean at intermediate steps.
  • If you can simulate the system, compute the price before training. Fitted rollouts were within 0.83 to 1.09 times this value.

Figure 28 (p. 73) gives a summary of these options.

4. Additional points. We agree that recursive bias and ensembling are classical, and that propagating a fitted kernel is standard practice. We report clearly where this approach falls short. Our main contribution is analyzing when and why composing means fails, with sharp constants and a price that can be computed before training. Theorem 1's proof is a triangle inequality, and the constant CC can increase with dimension; our response to Reviewer WKrv explains what the bound still shows in that case. The Lorenz gap is expected, but now its size can be predicted in advance (60.3 versus a measured 60.4). Our experiments now cover 40- and 4096-dimensional systems and released models up to 1.3B parameters. If the title seems too broad, we are open to narrowing it to forecasting in the final version.

We hope these responses address your concerns and welcome further discussion.

AuthorsOpenReview ↗

With thanks, before the discussion closes

As the discussion comes to a close, we want to thank you again for your review.

Your three questions were clear and practical, and the paper is better for them: the experiments on released models and on trained flow maps exist largely because you asked. If you have a moment to look at our response, we would be very grateful to know whether it answers your concerns. Either way, thank you for the time you gave to our work.

Comments on the paper

2 comments

We sincerely thank the AC, SAC, and reviewers for their valuable time and constructive feedback.

In response to the reviews, we revised the main manuscript and supplementary material, utilizing the additional page permitted for extended discussion. As requested, the newly included experiments demonstrate that the cost arises in trained flow maps and released forecasters.

The reviews also enabled us to articulate more precisely the novelty of our work in relation to classical results on iterated forecasts and ensemble averaging. For predictors designed to compose, either by architecture or by training, our paper makes the following contributions:

  1. When the true means do not compose, no such predictor can achieve accuracy at every horizon simultaneously under coverage and smoothness conditions, with sharp rates (Thm. 1; Props. 1, 5).
  2. For rollouts, this cost can be determined prior to training from the noise-free dynamics (Cor. 1), and trained models closely match this predicted cost.
  3. An exact criterion on drift and noise characterizes when the conditional means of a diffusion compose, that is, when consistency incurs no target-level cost (Prop. 2). This criterion holds for deterministic transport, as learned by generative flow maps, and for certain nonlinear noisy systems, indicating that noise and nonlinearity alone do not necessarily impose a cost.
  4. For consistency training under noise, a self-consistency penalty that bootstraps from the model's own predictions targets the composed means rather than the true means, resulting in a conflict with squared-error training. Bootstrapping from observed intermediate states resolves this conflict (Sec. 2; App. C.4; Tables 17 and 22).

In practice, these results show practitioners when consistency across time is free, when it incurs an accuracy cost (and, for rollouts, by how much), and how to avoid that cost. This matters now because composition is increasingly built into learned forecasters: emulators trained through their rollouts (GraphCast), semigroup-constrained and self-consistent models (Deep-OSG; Recursive Flow Matching; Semigroup-JEPA), and released benchmarks and models that feed point predictions back or average stochastic passes at every step (RealPDEBench; Walrus), while semigroup regularization has been reported to have mixed effects (Shikhman, 2026). Figure 28 turns our results into practical guidance for these uses.

We are deeply grateful for the reviews, which have substantially strengthened the paper.


What changed in the revision. The submitted theory and main experiments are kept. New material added, by request:

  • Neural ODE against an equal-size direct network, and Corollary 1 (WKrv): Fig. 3, App. L
  • The price across noise levels (WKrv) and neural heads for the mean-preserving target (NeSv): Apps. N, M
  • When the bound is informative, where the cost appears and why coverage is needed (WKrv, NeSv): Section 2, p. 4
  • Released models: QCNet on Argoverse 2 (WKrv, ZP2C), RealPDEBench checkpoints on measured flows and the Walrus foundation model (ZP2C): Apps. S–U
  • Exact composition imposed on trained flow maps (ZP2C): Apps. O, P
  • Consistency penalties, a flow-matching forecaster and a released consistency model (ZP2C): Apps. R, X, Y
  • Also new since submission: a 4096-dimensional Navier–Stokes system and a closed-form diffusion in which every constant of the bound is known: Apps. Q, W
  • Practitioner guidance (ZP2C, NeSv): Section 4, p. 7; Fig. 28, p. 73
  • Writing: the introduction now separates the classical intuition from what is new, results are stated in plain technical terms, and the limitations, including where propagation falls short (QCNet, RealPDEBench filters), are gathered in Section 6
AuthorsOpenReview ↗

Closing summary and thanks

As the discussion closes, we warmly thank the reviewers and the AC for the time and care they gave to this paper.

Below is a short summary of the main points raised and how the revision addresses them; our overall comment lists the changes in detail.

  1. Relation to classical recursive-forecasting bias and ensembling (all reviewers). What we add: when the true means do not compose, the bias cannot be trained away while a predictor composes (Theorem 1), with sharp rates in specific examples (Propositions 1 and 5), a price computable before training (Corollary 1), and an exact criterion for when the cost vanishes (Proposition 2).
  2. Quantitative usefulness of the bound (all reviewers, and WKrv's follow-up). Section 2 now explains when the bound is informative, where the cost must appear, and why coverage is needed. Corollary 1 and direct measurement give numbers without the coverage constant.
  3. Whether the cost shows up in practice (WKrv; ZP2C's first two questions). New experiments on released models (QCNet on Argoverse 2, RealPDEBench checkpoints trained on measured flows, the Walrus foundation model), on trained flow maps forced to compose, and on consistency models. For example, the released QCNet's squared error rises by more than a third (13.6 m²) when it composes its own mean forecasts.
  4. What a practitioner should do (ZP2C's third question; NeSv). The revision adds a short guide: when composition is free, when it costs accuracy, and how to avoid the cost.

Our main message is simple: when the true conditional means do not compose, every predictor built to compose pays for it somewhere, and the paper shows when this happens, what it costs a rollout, and how to avoid it.

We are grateful to Reviewers WKrv and NeSv for engaging with our response and for revisiting their assessments, and to Reviewer ZP2C for three questions that shaped the new experiments. The paper is better because of all three reviews. Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.