acceptodds
Desk rejected at ICLR 2027

Certified Frozen-Midpoint Neural Integrators: Original-Energy Dissipation and Second-Order Physical Time

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural gradient-flow solvers can decrease the correct energy while evolving on an incorrect physical clock. With state-dependent mobility, even an exactly solved, energy-dissipative step can be only first order. We develop a certified frozen-midpoint interface that separates three obligations: mobility consistency, full-state nonlinear accuracy, and signed energy work. A predicted midpoint freezes the mobility before optimization, preserving a globally strongly convex average-vector-field subproblem whenever its stated curvature and existence conditions hold. Deterministic oracle envelopes give computable objective-gap and endpoint-radius certificates, without assuming global neural training convergence. We prove variable-step second-order accuracy, finite correction for nonstationary states under explicit solver assumptions, and sharp worst-case tolerance scalings. Applications include coupled positive quantile diffusions and full-covariance Gaussian Wasserstein flows with noncommuting matrices. For smooth quantile reconstructions, a separate posterior theorem certifies error against the continuum evolution, including an explicit spatial defect. We also characterize validation under adaptive sampling and the obstruction at exact equilibrium. This is a purely theoretical study: it establishes conditional guarantees and computability constructions, not empirical speedups, unrestricted-generator convergence, or priority for discrete-gradient energy preservation itself.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural ordinary differential equations, Physics-informed machine learning, Gradient flows, Structure-preserving integration, Energy dissipation, Differentiable simulation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aZJVOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how to check approximate numerical steps for systems that evolve downhill in energy. The main point is that energy can decrease at every step while the system still evolves at the wrong rate. The proposed checks allow steps to be accepted, rejected, or corrected. This is a focused theoretical study providing mathematical guarantees under stated assumptions.

Critical strengths
  1. The distinction between energy decrease and correct temporal evolution is important and clearly motivated. The counterexamples help explain why these require separate checks.

  2. The analysis provides explicit accuracy requirements for approximate solves and establishes that valid steps exist for the quantile and Gaussian models considered. These give the framework concrete mathematical support.

Critical weaknesses
  1. My main concern is that the prose and organisation are extremely difficult to follow. The abstract introduces many specialized terms before providing an accessible explanation. Across seven main-text pages and extensive appendices, I found it difficult to track which assumptions support each result and how the results fit together. This substantially limits my ability to confidently assess the full mathematical development.

  2. The benefit of the learned predictor remains unclear. The paper assumes prediction accuracy when discussing reduced correction work, while Appendix O.3 shows that a classical predictor with a correction can already achieve the required residual order under its assumptions. I understand that this is a theoretical study, but neither a theoretical nor an empirical advantage from learning is established. This leaves the significance for a broader machine-learning audience uncertain.

Questions for authors
  1. Could the authors provide one worked example showing what the learned predictor outputs, how each check is evaluated, and how a failed step is corrected? This would substantially improve understanding of the method.

  2. What prediction accuracy would make learning useful compared with the classical predictor and correction in Appendix O.3? Including validation and correction costs, when should the learned approach require less work?

  3. The paper acknowledges increasing precision requirements near equilibrium. How demanding do these become in the stated model problems, and under what circumstances could the acceptance checks remain unresolved?

5 comments
AuthorsOpenReview ↗

Response to W1: A clearer explanation and an explicit map from assumptions to results

Thank you for explaining where the presentation made the mathematical development difficult to assess. We appreciate this detailed feedback and have reorganized the revision around the practical question: when can an approximate proposed step be trusted to advance physical time?

The abstract now opens with the distinction between energy decrease and the physical clock. The introduction illustrates it with a simple example. Both x˙=−x\dot x=-x and x˙=−2x\dot x=-2x decrease the energy x2/2x^2/2, but their states at the same time differ. Energy decrease therefore answers one question, while accuracy at a specified time answers another. Our method checks both. A predictor proposes the next state, the checks assess it, and a correction repairs it when needed. Specialized terms are introduced after this explanation.

We have also placed a result and assumption guide in Table 1 of the revised main text. Its purpose is to make each conclusion readable without first navigating the appendices:

ResultMain conditionsWhat the reader can conclude
Endpoint distance and objective gap, Theorem 2.1Fixed positive mobility, global positive curvature, an interior minimizer, and a validated residualA computable bound relative to the global solution of the frozen problem
Energy decrease, Theorem 2.2An admissible connecting segment, an enclosed oracle error, and the signed energy testThe stated energy decreases
Second order in physical time, Theorem 3.2A regular finite time corridor and residual, midpoint, and mobility errors of order h2h^2Accuracy at the prescribed physical times for a fixed spatial model
Finite correction, Proposition 3.4A nonstationary state, a convergent corrector, and shrinking oracle errorsThe two endpoint checks eventually pass
Continuum error, Theorem 5.1A regular positive exact solution and a continuous positive smooth reconstruction with a bounded full PDE residualA conditional bound on the total continuum error
Reduced correction work, Appendix OSufficient prediction accuracy and a stated correction rateA conditional reduction in correction work, with runtime assessed separately

The organization follows these dependencies. Section 2 introduces the tests and a complete learned example. Section 3 explains why the midpoint and residual conditions preserve the physical clock. Section 4 supplies the quantile and Gaussian models in which the required global solution exists. Section 5 explains the additional requirements for a continuum certificate. Section 7 then evaluates actual learned proposals and their complete certified costs.

The worked example makes the distinction between the tests concrete. A real network output passes the energy check immediately, but its residual is too large. Three corrections reduce the residual below the specified tolerance, at which point the step is accepted. The main text gives the sequence of decisions, while Appendix X.2 gives all input and output coordinates. This provides a direct reading path from an actual proposed state to its mathematical guarantee.

A central feature of the presentation is that the endpoint certificate concerns the full state, even when a nonconvex neural network produced that state. The reader can therefore assess the certificate without assuming that neural training found a global optimum. Thank you again for motivating these changes. They make both the practical procedure and the scope of each theorem substantially easier to follow.

AuthorsOpenReview ↗

Response to W2: The learned predictor now has a measured role and a certified deployment interface

Thank you for identifying the need to distinguish a useful learned predictor from a prediction accuracy assumption. We have added trained models, a comparison with the classical predictor in Appendix O.3, and a second comparison that holds Newton correction fixed. These experiments assess whether learning reduces the complete cost of an accepted trajectory.

The learning task is concrete. A network takes the current quantile state, its vector field, the step size, and model parameters. It predicts a correction to the logits of the positive gaps of an Euler proposal. The network is trained using certified solutions of the same frozen implicit problem. A new group of initial states and parameters is reserved for testing. Acceptance still depends on the predicted state passing the mathematical checks, which supplies a deployment guarantee independently of whether training attained a global optimum.

We use 256 training, 64 validation, and 128 test groups, with three independently trained networks. Each network has two hidden layers with 64 tanh units. For the direct Appendix O.3 comparison, Euler and learned proposals use the same safeguarded fixed point corrector. The midpoint, mobility, acceptance tolerances, physical end time, and certification arithmetic are held fixed. Proposal inference, correction, oracle evaluation, and acceptance checks all enter the online cost.

The table reports Euler time divided by learned time, so values above one favor learning. Each interval is a paired 95% bootstrap interval over 16 test groups, with ten timing repeats per group. The physical interval is T=0.02T=0.02 and the primary step size is h=0.002h=0.002.

Correction methodTraining seedCertified online time ratio95% interval
Fixed point171.158[1.113, 1.198]
Fixed point291.188[1.131, 1.241]
Fixed point431.238[1.185, 1.285]
Newton171.020[0.990, 1.052]
Newton291.004[0.978, 1.029]
Newton431.009[0.981, 1.039]

All 2,048 primary trajectories across both tested step sizes and all eight combinations of initializer and corrector reach the same final time. Every matched pair satisfies the stated numerical error comparison criterion: terminal error ratios lie within [0.9, 1.1], and tightening the numerical reference changes it by at most 1% of either error. Thus the fixed point gains are supported by measurements of completed, certified trajectories with comparable observed accuracy.

This establishes a specific empirical benefit over the Appendix O.3 baseline: about 1.16 to 1.24 times faster online execution in this setting. The Newton intervals include one, so the benefit is associated with the correction workload that the predictor can save. Indeed, Euler with Newton already averages only 10.133 corrections per ten step trajectory, compared with 10 for the learned initializers. The Newton extension uses the same models originally selected with fixed point correction, without retraining.

A separately reported follow up at h=0.004h=0.004 gives fixed point ratios of 1.21 to 1.25, while its Newton intervals again include one. Its timing protocol was fixed before execution, after the primary results were examined, and its measurements are kept separate.

The resulting contribution to learning is twofold: a verifiable interface for accepting and repairing a learned implicit proposal, and evidence that a learned initializer can reduce the full certified workload when correction leaves enough work to save. Section 7 and Appendix X.1 now make this distinction explicit. Thank you for prompting a comparison that connects the theory to an actual trained predictor and its deployment cost.

AuthorsOpenReview ↗

Response to Q1: A complete learned proposal, its failed check, and its certified correction

Thank you for this helpful suggestion. We have added a worked example to Section 2 and its complete coordinates to Appendix X.2. It uses an actual trained network on a reserved test case, so the reader can follow the proposed state through every acceptance check.

The model has two quantile components with eight cells each and fixed endpoints at −1-1 and 11. Its 14 interior coordinates describe two increasing functions. Positive gaps between adjacent coordinates ensure admissibility. The diffusion coefficients are (0.5,1)(0.5,1), confinement coefficients are (2,1)(2,1), and coupling parameters are (b,c)=(0.25,0.5)(b,c)=(0.25,0.5). We use test group 320, network seed 17, and h=0.0005h=0.0005.

What the network outputs. The network predicts an adjustment to the Euler proposal's gap logits. A softmax converts them to positive gaps whose sum is two in each component. Cumulative sums give the proposed endpoint yθy_\theta. The table shows the actual input and proposal, rounded to six decimal places for readability.

CoordinateInput xxLearned endpoint yθy_\theta
1-0.796668-0.795496
2-0.576750-0.576169
3-0.341258-0.340918
4-0.103140-0.103076
50.1347980.134705
60.3893880.389180
70.6793880.678488
8-0.799523-0.798557
9-0.571485-0.570962
10-0.312360-0.312137
11-0.041393-0.041448
120.2223490.222140
130.4776010.477416
140.7351860.734746

How the checks are evaluated. In the exact mass metric of this quantile model, K=B=IK=B=I. We use z=x+hF(x)/2z=x+hF(x)/2, computed in rational arithmetic. The midpoint and mobility discrepancies are therefore zero. The minimum midpoint gap is 0.20096889, which verifies its admissibility. Positive endpoint gaps are checked again for every proposal and correction.

For an endpoint yy, the residual is r=(y−x)/h+Bg(x,y)r=(y-x)/h+Bg(x,y), where gg is the average energy gradient along the segment from xx to yy. It measures how accurately the proposed endpoint solves the frozen time step. The oracle is exact here, so its error bound is σ=0\sigma=0. The residual must satisfy ϵˉ≤h2=2.5×10−7\bar\epsilon\le h^2=2.5\times10^{-7}.

The separate energy test uses η=1/2\eta=1/2. Define ME=ηD~−ω~M_E=\eta\widetilde D-\widetilde\omega, where D~\widetilde D is the squared gradient norm and ω~\widetilde\omega is the signed interaction between the gradient and residual. A nonnegative MEM_E certifies energy decrease. The observed decisions are:

Corrections appliedResidual upper boundEnergy margin MEM_EDecision
01.48051×10−31.48051\times10^{-3}0.857356Residual fails
13.03163×10−53.03163\times10^{-5}0.856822Residual fails
21.06348×10−61.06348\times10^{-6}0.856817Residual fails
34.18920×10−84.18920\times10^{-8}0.856817Both tests pass

How repair works. The fixed point corrector updates the proposed endpoint toward the solution of y=x−hBg(x,y)y=x-hBg(x,y), retaining admissibility and reevaluating the checks. The initial proposal already decreases energy, but the procedure keeps physical time unchanged until the residual is also small enough. After three corrections, the minimum endpoint gap is 0.20144407 and the step is accepted.

Finally, with curvature βh=2000.25\beta_h=2000.25, the same residual yields a certified distance bound ϵˉ/βh=2.09434×10−11\bar\epsilon/\beta_h=2.09434\times10^{-11} and objective gap bound ϵˉ2/(2βh)=4.38680×10−19\bar\epsilon^2/(2\beta_h)=4.38680\times10^{-19}. These are display rounded bounds for the frozen problem. Acceptance uses the exact stored records. Only now does physical time advance by hh.

Thank you again for suggesting this example. It makes the practical advantage of the certificate visible: a learned proposal can be inspected, repaired, and accepted with an explicit quantitative guarantee on the resulting state.

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

This paper studies the problem of developing neural integrators that preserve energy dissipation while maintaining second-order accuracy in physical time. The authors argue that energy dissipation alone does not guarantee correct temporal dynamics, especially for gradient flows with state-dependent mobility. To address this problem, they propose a Certified Frozen-Midpoint (CFM) framework that combines midpoint prediction, a frozen average-vector-field formulation, and additional certification conditions for nonlinear solver accuracy and energy dissipation. The authors provide theoretical guarantees for objective-gap bounds, energy dissipation, and second-order temporal convergence under specific assumptions. They further extend the analysis to Wasserstein gradient flows, including coupled quantile diffusion and full-covariance Gaussian models. The paper also discusses inexact oracle evaluations, continuum error certification, and the limitations of neural prediction.

Critical strengths

1: The paper addresses an interesting problem in numerical integration by distinguishing energy dissipation from physical-time accuracy. The motivation is clear and meaningful.

2: The authors provide a comprehensive theoretical analysis, including guarantees for energy dissipation, nonlinear solver accuracy, and second-order temporal convergence. The analysis also considers inexact solvers and oracle errors.

Critical weaknesses

1: The paper is only loosely connected to machine learning. The proposed framework is more closely related to numerical analysis, and most of the theoretical results do not depend on machine learning techniques. In fact, the neural component is not necessary for the proposed framework, as the authors also provide a classical numerical predictor that satisfies the required conditions. Therefore, the main contribution of this work to the machine learning community remains unclear. I have very limited knowledge on this topic

Questions for authors

1: Can the authors further clarify the connection between the proposed framework and machine learning? Since most of the theoretical guarantees do not depend on neural networks and the framework can also be implemented using classical numerical methods, what is the main contribution of this work to the machine learning community?

3 comments
AuthorsOpenReview ↗

Response to W1: Contribution to reliable learned integration

Thank you for raising this important question, and for recognizing the distinction between energy decrease and accurate temporal dynamics. The contribution to machine learning is a framework for checking and repairing the outputs of a learned integrator before those outputs are used to advance physical time. This makes the mathematical reliability of an accepted prediction independent of whether neural training found a global optimum.

The relevant learning task is repeated prediction of evolving distributions at prescribed times. Reaching a low energy state is not sufficient for that task. For example, both x˙=−x\dot x=-x and x˙=−2x\dot x=-2x decrease x2/2x^2/2, but give different predictions at the same time. A neural proposal therefore needs a check of the intended dynamics as well as a check of energy decrease. Our framework supplies these checks with explicit tolerances.

What the theory contributes to a learned solver. A network supplies a candidate state. The certificate examines that state in the original state space, rather than relying on a small training loss or a small gradient with respect to network parameters. Under the stated curvature and existence assumptions, Theorem 2.1 gives a computable bound on its distance from the unique frozen subproblem solution and on its objective gap. Theorem 2.2 verifies dissipation with inexact oracle evaluations. With the midpoint and regularity assumptions, Theorem 3.2 establishes second order accuracy at the actual physical times.

These obligations have distinct practical roles:

CheckWhat it establishes
Midpoint and mobility consistencyThe frozen mobility has the required temporal consistency
Full state residualThe implicit subproblem has been solved accurately enough
Signed energy workThe accepted state satisfies the required dissipation inequality

These conclusions are conditional on the assumptions stated in each theorem, now collected in Table 1. Their independence from the training algorithm is useful for machine learning: the same acceptance interface can evaluate imperfect outputs from different networks and repair them when necessary. The classical predictor provides a concrete fallback and a fair comparator for measuring how much work learning saves.

Evidence from actual trained networks. Section 7 and Appendix X.1 now evaluate three trained networks on the coupled quantile model. We compare Euler and learned initializers while holding the corrector, physical time steps, and exact acceptance gates fixed. The ratios below are Euler time divided by learned time, so values above one favor learning.

Corrector and step sizeSeed 17Seed 29Seed 43
Fixed point, h=0.002h=0.0021.1581.1881.238
Fixed point, h=0.004h=0.0041.2101.2241.246

All six fixed point 95% intervals lie above one. For the corresponding Newton comparisons, all six 95% intervals include one, so the measured benefit is specific to the fixed point correction regime. The timing includes proposal evaluation, correction, oracle work, and exact certification, after model loading and warmup. Each step size uses a separate timing run on 16 test groups with ten paired repeats. The added Newton comparison and the second timing setting are identified as follow up analyses, using the original models without retraining.

The main contribution is thus a certificate interface for using learned proposals in reliable transient prediction, supported by explicit model classes and a measured example of reduced certified work. We have made this connection explicit in the revised introduction, the assumptions table, and the experimental section. Thank you again for prompting us to distinguish the role of learning from the mathematical conditions that make its predictions trustworthy.

AuthorsOpenReview ↗

Response to Q1: A concrete workflow from learned proposal to certified prediction

Thank you for asking us to make the machine learning connection concrete. The intended user has a specified gradient flow and wants to predict how a distribution evolves at given physical times across many initial conditions or parameter settings. Our approach lets a trained model propose the next state, while a separate certificate decides whether that proposal is accurate enough to use. The revised manuscript now includes this workflow and an example with an actual trained network.

1. Train a predictor for repeated implicit solves. In the coupled quantile experiment, the network takes the current state, vector field, time step, and model parameters as inputs. It predicts a correction to the Euler endpoint through positive gap coordinates. Those coordinates describe an increasing quantile and therefore a valid distribution. Training targets come from the same frozen implicit problem used at deployment, and each target carries a residual and endpoint radius certificate.

Training and evaluation componentImplemented setting
State representationTwo quantile components, eight cells each
NetworkTwo hidden layers of 64 tanh units
Independent groups256 training, 64 validation, 128 test
Total certified labels across the splits7,168
Retained training seeds17, 29, 43

Whole initial state and parameter groups remain in a single split. Model selection uses validation data, and all three selected models are evaluated on the test groups. Appendix X.1 records the architecture, targets, selection procedure, and cost accounting.

2. Check the prediction before advancing time. We check domain admissibility, midpoint consistency, the full state residual, and signed energy work. These are checks of the proposed physical state, so a successful certificate does not require proving global convergence of neural training. When needed, the corrector improves the candidate in the original state space until the same gates pass.

The worked example in Section 2 and Appendix X.2 uses seed 17 and h=0.0005h=0.0005. The network output already passes the energy check. The residual check determines how much correction is needed:

CandidateResidual upper boundResidual budgetDecision
Network output1.48051×10−31.48051\times10^{-3}2.5×10−72.5\times10^{-7}Correct
After correction 13.03163×10−53.03163\times10^{-5}2.5×10−72.5\times10^{-7}Correct
After correction 21.06348×10−61.06348\times10^{-6}2.5×10−72.5\times10^{-7}Correct
After correction 34.18920×10−84.18920\times10^{-8}2.5×10−72.5\times10^{-7}Accept

These are rounded displays of exact stored certificate records. Only the accepted endpoint advances time. This shows how a real learned proposal becomes a certified implicit step with an explicit stopping rule.

3. Measure whether learning saves certified work. We rerun every method with identical gates and compare initializers using the same corrector. All 4,096 single step cases and all 2,048 primary trajectories complete. At h=0.002h=0.002, the online time ratios and paired group bootstrap intervals are:

CorrectorSeedEuler time / learned time95% interval
Fixed point171.158[1.113, 1.198]
Fixed point291.188[1.131, 1.241]
Fixed point431.238[1.185, 1.285]
Newton171.020[0.990, 1.052]
Newton291.004[0.978, 1.029]
Newton431.009[0.981, 1.039]

The ratios include prediction, correction, oracle work, and exact certification, after model loading and warmup. They use 16 test groups with ten paired repeats. The added Newton analysis reuses models selected with fixed point correction. Its intervals include one, as do the Newton intervals in the separate h=0.004h=0.004 follow up. The positive fixed point results support a specific use case: learning can reduce repeated correction work while preserving the same acceptance requirements.

The complete offline reproduction took 87.531 seconds, corresponding to descriptive fixed point amortization estimates of 4,193 to 5,908 trajectories at h=0.002h=0.002. Appendix X.1 accounts for data generation, all training candidates, and validation selection.

Thank you again for this helpful request. The revised explanation connects the network's computational role to the certificate's reliability role, following a learned output all the way to acceptance.

I really appreciate authors clarification. However, I am still struggling to understand how is the related to machine learning works. I strongly suggest the author to refine their manusciprt for a more clear illustration.

Reviewer_z28wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a certification framework for approximate gradient-flow integration. It combines a predicted midpoint for state-dependent mobility, a residual-based certificate for the frozen average-vector-field subproblem, and a separate energy-work test.

Critical strengths
  • The integrated entropy potential remains finite at singular covariances.
  • The paper distinguishes state-space residuals from parameter gradients and deterministic certificates from statistical validation.
Critical weaknesses
  • . The paper supplies enclosure constructions, but no complete cost analysis for correction or oracle refinement.
Questions for authors
  • Can the authors establish a quantitative benefit over their own classical predictor–corrector baseline, including certification costs?
  • Can the authors perform the globally solvable quantile discretization to a smooth continuum reconstruction with a proven refinement error?
3 comments
AuthorsOpenReview ↗

Response to W1: Complete accounting of correction and certification costs

Thank you for identifying the importance of accounting for correction and oracle work. We have added an explicit cost decomposition, measured the complete certified solver, and connected its stopping requirements to the correction analysis. The new material appears in Section 7 and Appendices X.1, X.3 and X.4 of the revised manuscript.

The measured online cost starts when the initial proposal is computed and ends when the final acceptance certificate is obtained. It includes neural inference or the Euler proposal, every correction, exact oracle evaluations, the remaining validation operations, and solver bookkeeping. Oracle time is separated from the other categories to avoid counting the same work twice. Model loading and warmup precede online timing. The complete offline workflow is reported separately.

The following table gives mean milliseconds per trajectory for the fixed point corrector at h=0.002h=0.002 and T=0.02T=0.02. Each column contains 16 test groups with 10 repetitions per group. The three learned columns correspond to separately trained seeds.

Cost componentEulerLearned 17Learned 29Learned 43
Proposal0.8172.1862.1602.138
Floating point correction9.1687.3167.1926.843
Exact oracle32.10428.16427.62626.877
Other validation70.02958.53657.63055.796
Bookkeeping0.5490.5110.5060.494
Total112.66696.71395.11492.148

The component means sum to the total up to rounding. Thus the observed reduction includes the cost of proving acceptance, which is the largest component here. The exact quantile AVF needs no quadrature refinement. For the Gaussian model, the separate precision study explicitly varies validated quadrature panel budgets and records oracle refinements, correction counts, and unresolved attempts. A rejected attempt never advances physical time.

We also measured all four initializers with the same structured Newton corrector. Its time ratios have intervals containing one, so the demonstrated timing benefit in this table concerns the fixed point corrector. This additional control isolates the source of the benefit rather than attributing a change of nonlinear solver to learning.

The analysis explains the relevant accuracy dependence. Under the conditions of Appendix G, if a corrector contracts the objective gap by a factor 1−ν1-\nu per iteration, reducing an initial gap Δ0\Delta_0 to an acceptance target Δtar\Delta_{\mathrm{tar}} requires at most

max⁡(0,⌈log⁡(Δ0/Δtar)−log⁡(1−ν)⌉)\max\left(0,\left\lceil \frac{\log(\Delta_0/\Delta_{\mathrm{tar}})}{-\log(1-\nu)} \right\rceil\right)

iterations. The target depends on the time step, curvature, and distance from stationarity. Appendix X.3 makes those dependencies explicit. In particular, oracle accuracy must resolve both the time accuracy budget and the energy work margin. The measured cost decomposition complements this conditional iteration bound by charging the arithmetic and validation work actually performed.

An isolated repeat of the complete offline workflow took 87.531 seconds, including all 7168 labels, all six training candidates, validation selection, and process initialization. With the measured fixed point savings, the three amortization point estimates are approximately 5908, 5097 and 4193 trajectories. These are estimates for repeated use in this setting. The separation of offline and online costs makes the intended use case explicit.

We appreciate this suggestion. It has allowed us to give a concrete account of both the work needed to construct a certificate and the work saved by a learned initializer.

AuthorsOpenReview ↗

Response to Q1: Quantitative benefit over the classical predictor and corrector

Thank you for asking for a comparison that includes certification costs. Yes, the new controlled experiment establishes a quantitative benefit over the classical Euler predictor with the safeguarded fixed point correction associated with Appendix O.3. We compare initializers while keeping the corrector, problem, and acceptance checks fixed.

The experiment uses a coupled quantile flow with two components and eight cells per component. We train three networks using seeds 17, 29 and 43. Complete trajectory and parameter groups are separated into 256 training, 64 validation and 128 test groups. Checkpoints are selected using validation data. The learned initializer and Euler initializer use identical midpoint construction, mobility, error tolerances, energy tests, and correction budgets. Floating point proposals are accepted only after exact rational validation.

At h=0.002h=0.002 and final time T=0.02T=0.02, the ratios below divide Euler solver time by learned solver time. A value above one therefore means a faster learned initializer. Every timing includes the proposal, all corrections, oracle work, and final certification.

Training seedSame fixed point correctorSame Newton corrector
171.158 [1.113, 1.198]1.020 [0.990, 1.052]
291.188 [1.131, 1.241]1.004 [0.978, 1.029]
431.238 [1.185, 1.285]1.009 [0.981, 1.039]

Brackets show 95% intervals from 2000 bootstrap resamples of trajectory groups. We use 16 fixed test groups with 10 repetitions each, randomized paired order, one CPU core, and one BLAS thread. Each group's repeated times are reduced to a median before calculating the paired geometric mean ratio. These intervals quantify this experiment, rather than variability across arbitrary machines.

The fixed point comparison shows a complete online speedup of approximately 1.16 to 1.24 across the three seeds. Its mechanism is visible in the full test set. The average number of corrections per ten step trajectory decreases from 39.80 for Euler to 30.75, 30.35 and 30.04 for the learned initializers. The network saves repeated validation and correction work despite having a more expensive proposal.

The Newton columns provide an additional control. This follow up reuses the models originally selected with fixed point correction, without retraining. Their intervals include one. Euler with Newton already averages 10.13 corrections per ten step trajectory, compared with 10 for each learned initializer. Accordingly, our quantitative advantage is stated for the classical fixed point baseline, rather than for every corrector.

All 4096 single step cases and all 2048 primary trajectories finish successfully. We also check physical task error against a tightened numerical ODE reference. Every matched comparison satisfies the disclosed descriptive error ratio range of 0.9 to 1.1, with reference refinement disagreement below 1% of the smaller error. Thus the timing benefit is accompanied by comparable measured accuracy and identical acceptance requirements.

A separate follow up timing at h=0.004h=0.004 gives fixed point ratios of 1.210, 1.224 and 1.246. All three intervals are above one. This analysis uses the same models and is reported separately because it was prompted by the primary results. Newton intervals again include one. Section 7 and Appendix X.1 give both experiments and the complete accounting, including the separately measured offline cost of 87.531 seconds.

We appreciate this question because it leads to a more precise statement of the benefit: learning can reduce the work of reaching the same certified endpoint accuracy, and the gain is measurable after paying for certification.

AuthorsOpenReview ↗

Response to Q2: Constructive smooth reconstruction and certified continuum error

Thank you for highlighting the connection between the globally solvable discrete model and the continuum certificate. We have added an explicit construction in Section 5 and Appendix X.5. It starts from the accepted piecewise affine quantile trajectory, preserves its positivity and endpoints, and supplies both a proved reconstruction error bound and a computable bound relative to the continuum solution.

Let QN(t,a)Q_N(t,a) be the piecewise affine spatial representation of the quadratic time reconstruction, where aa is the mass coordinate. We certify its slope throughout each time interval, rather than only at its endpoints. Write these bounds as 0<s−≤∂aQN≤s+0<s_-\leq\partial_a Q_N\leq s_+. For a fixed degree mm, define the Bernstein reconstruction

Um(t,a)=∑k=0mQN(t,k/m)(mk)ak(1−a)m−k.U_m(t,a)=\sum_{k=0}^{m}Q_N(t,k/m)\binom{m}{k}a^k(1-a)^{m-k}.

This is a smooth polynomial in the mass coordinate. It preserves the boundary values. Its derivative is a convex combination of secant slopes of QNQ_N, so s−≤∂aUm≤s+s_-\leq\partial_a U_m\leq s_+. Because the same linear reconstruction is applied on every time interval, the reconstructed trajectory is continuous at their interfaces.

The slope bound also proves the quantitative approximation estimate

∥Um−QN∥L2(0,1)≤s+2m.\lVert U_m-Q_N\rVert_{L^2(0,1)} \leq \frac{s_+}{2\sqrt{m}}.

To see why, the Bernstein polynomial is the expectation of QN(t,K/m)Q_N(t,K/m) for a binomial random variable KK with parameters m,am,a. The Lipschitz bound gives an error at most s+E∣K/m−a∣s_+\mathbb{E}|K/m-a|, which is bounded by s+a(1−a)/ms_+\sqrt{a(1-a)/m}. This directly connects the discrete quantile to an admissible smooth representation.

For the diffusion test with κ=1/4\kappa=1/4 and zero confinement, we then enclose the full continuum residual

R=∂U∂t−2κ∂2U/∂a2(∂U/∂a)3.R=\frac{\partial U}{\partial t}-2\kappa\frac{\partial^2 U/\partial a^2}{(\partial U/\partial a)^3}.

Assuming the positive regular exact solution XX specified in Theorem 5.1, the distance of the original discrete endpoint to that solution is bounded by the sum of the final reconstruction error, the initial reconstruction error relative to X(0)X(0), and the time integral of ∥Rm∥L2\|R_m\|_{L^2}. In one dimension this quantile distance equals the Wasserstein distance between the represented distributions.

We have executed the construction. All 105 steps from nine discrete trajectories pass the exact acceptance tests and positivity checks throughout their time intervals. For the strict enclosure subset N=8N=8, h=0.04h=0.04, T=0.2T=0.2, the computed bounds are:

Degree mmInitial error boundFinal reconstruction boundIntegrated residual boundTotal endpoint bound
80.0071930.0057630.0018900.014845
160.0041560.0030810.0013500.008586
320.0024860.0016110.0010400.005137

Each displayed bound is rounded upward. The total is rounded from the sum of the unrounded bounds. The residual enclosure covers the complete mass and time domain using rational Bernstein range bounds and interval division with a certified positive denominator. These bounds do not use a numerical reference trajectory.

The proved refinement statement here is the reconstruction estimate above, accompanied by a validated continuum posterior bound under Theorem 5.1. A general rate uniform in the spatial mesh would additionally require uniform bounds on spatial consistency and regularity. We therefore keep the full residual in the certificate, so improving the reconstruction alone cannot silently stand in for controlling the PDE error.

We appreciate this suggestion. The new construction makes the passage from the solvable quantile model to the continuum certificate explicit and executable, with all three contributions to the endpoint error separately available.

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

Related papers

Discussion (0)

Sign in to comment.