acceptodds
Under review as a conference paper at ICLR 2027

SALTO: Solver-Anchored Lifted Transfer Operators for Nonlinear Event Inference

Abstract

Learned dynamical models can accurately predict trajectories while producing unreliable sensitivities when differentiated or transposed, limiting their use in inverse problems and scientific decision-making. Classical adjoints provide exact local derivatives, but a single adjoint cannot represent the nonlinear response of a system to a family of finite-amplitude perturbations. We introduce SALTO (Solver-Anchored Lifted Transfer Operators), a Koopman-based framework for inferring finite-amplitude sources. In the lifted space, the nonlinear dynamics are approximated by a linear propagator, and its transpose carries each sensor backward in a single solve. The result is a reusable lifted adjoint that predicts the measurement produced by any candidate localized event, such as a source of given position and amplitude, with one encoder pass and an inner product, so that the search needs no repeated nonlinear simulation. To give the model the simulator's exact local behavior, SALTO imposes three structural constraints: it anchors the physical-state dynamics to the simulator's linearized propagator, centers the learned observables so that they vanish to second order at the reference state, and uses a block-triangular lifted propagator. These constraints guarantee that the model reduces to the simulator's tangent dynamics and classical adjoint as the amplitude vanishes, while the learned observables carry only the finite-amplitude correction. On three nonlinear PDE systems, SALTO predicts the measurements of finite-amplitude sources to within – where the classical adjoint is in error by –, recovers source position and intensity from sparse sensors nearly as accurately as inversion through the solver, and samples the joint posterior of two interacting sources, matching the solver's posterior on reaction–diffusion at a twentieth of its cost. One trained model serves new sensor layouts and measurement times with a single backward solve, and its exact local structure keeps both measurements and their gradients correct as the amplitude vanishes, where unconstrained Koopman models do not. Enforcing exact local structure thus turns a learned lift into a reliable instrument for nonlinear sensitivity and inference.

Primary area:
learning on time series and dynamical systems
Keywords:
Koopman operator, sensitivity analysis, source inference, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_6Z3nOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SALTO (Solver-Anchored Lifted Transfer Operators), a Koopman-based framework for nonlinear PDE source inference. The main idea is to combine the tangent linear propagator provided by a numerical solver with learned nonlinear observables, allowing the nonlinear dynamics to be approximated by a linear propagator in a lifted space. The authors introduce three structural constraints—quadratic centering, solver anchoring, and block triangularity—to ensure that the learned model recovers the simulator's tangent linear dynamics and classical adjoint in the infinitesimal-perturbation limit.

The method is evaluated on three nonlinear PDE systems: Burgers, reaction–diffusion, and Kolmogorov flow. The experiments demonstrate improvements in finite-amplitude measurement prediction, source localization, intensity estimation, and multi-source posterior inference compared with classical adjoint methods and several learned baselines.

Critical strengths
  1. SALTO integrates the solver's tangent linear propagator into the learned Koopman model and guarantees recovery of the classical adjoint in the infinitesimal-perturbation limit. This is a meaningful contribution compared with approaches that rely solely on training losses to improve derivative accuracy.

  2. The experiments cover three PDE systems with distinct nonlinear mechanisms and evaluate measurement accuracy, source localization, and posterior inference. In particular, the multi-source reaction–diffusion experiment demonstrates that SALTO can approximate the solver-based posterior at substantially lower inference cost.

Critical weaknesses
  1. Computational scalability and overall efficiency

Although SALTO avoids repeated nonlinear simulations during source inference, its offline costs include training-data generation, tangent linear propagator construction, and learned operator fitting. The supplementary material provides useful implementation details and introduces translation-equivariant blocks to reduce computational complexity, but a systematic analysis of runtime and memory scaling remains lacking.

Moreover, the reaction–diffusion experiment in Table 1 achieves only approximately 1.8× online speedup over the nonlinear solver. Could the authors provide a breakdown of offline and online computational costs, evaluate scaling with spatial resolution and lifted-channel count, and estimate how many inference queries are required to amortize the offline costs?

  1. Finite-amplitude sensitivity accuracy

The theoretical guarantees concern the infinitesimal-perturbation limit and do not directly establish accurate gradients at finite amplitudes. For example, Supplementary Table 8 shows that, on reaction–diffusion, the proposed two-channel lift has a gradient error of 21% in the intensity range [1, 2), compared with 7.8% for the unconstrained lift. Beyond the training range, the corresponding errors increase to 73% and 24%.

These results suggest a trade-off between exact local consistency and finite-amplitude accuracy. Since the main experiments use larger lifts, could the authors clarify whether this trade-off persists for the four- and eight-channel models and provide a more systematic analysis of gradient accuracy across perturbation amplitudes?

Questions for authors

The same as weaknesses.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer 6Z3n

We thank the reviewer for the positive assessment.

W1. Offline and online cost, scaling and amortization. The new cost appendix (Table 13) gives every stage on one A100, every model in double precision. Offline, the training data, the propagator and the backward solves take seconds, and training takes about eight minutes for the four-channel lift. Online, a pass over the candidate grid is 6 times faster than the solver on reaction–diffusion and 70 times on Kolmogorov flow (general comment, item 2), so training is recovered after about 160 grid passes on reaction–diffusion and 12 on Kolmogorov flow. The encoder's cost per candidate grows about linearly with resolution and hardly changes with the number of channels. Memory sets the limit: the dense learned blocks grow as (cN)2(cN)^2, while the translation-equivariant blocks grow as c2Nc^2N and stay below five megabytes. For non-uniform base states, only the state block must keep the solver's resolution, and pooling the learned channels onto a coarser grid of M≪NM \ll N points would reduce the learned blocks to (cM)2(cM)^2 (Section 6).

W2. Finite-amplitude gradients and the number of channels. We retrained every ablation at four channels and ran the gradient check on them (general comment, item 1). The trade-off does not persist. At four channels the constrained lift is more accurate than the unconstrained lift in every band on both systems, including beyond the training range, and Table 8 gives every band at two and four channels. The eight-channel lift of the paper is the translation-equivariant lift for the posterior, trained on random fields rather than on the source family, so we did not include it in the gradient check.

The authors have addressed my questions. I will keep my original rating of weak acceptance.

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

This paper proposes SALTO, a Koopman-based framework for inferring finite-amplitude localized sources in nonlinear PDE systems from sparse sensors, addressing the limitation that a classical adjoint is linear in the source amplitude and cannot represent the nonlinear response of a family of sources. They lift the state with learned nonlinear observables so that the measurement becomes a linear functional of the lifted state, giving a state-independent lifted adjoint that is computed once per sensor and scores any candidate source by an encoder pass and an inner product; and they impose three structural constraints (solver anchoring, quadratic centering, block-triangular propagator) that make the model reduce exactly to the solver’s tangent dynamics and classical adjoint as the amplitude vanishes, so that learning only accounts for the finite-amplitude correction. Evaluation on viscous Burgers, two-dimensional reaction-diffusion, and Kolmogorov flow shows measurement errors of 2-8% where the classical adjoint and its quadratic correction err by tens to hundreds of percent, source inversion accuracy close to brute-force inversion through the solver, a two-source joint posterior that matches the solver’s at roughly a twentieth of the cost, and small-amplitude exactness that unconstrained Koopman lifts do not achieve.

Critical strengths
  1. The paper identifies a precise failure mode of learned dynamical models, namely that forward accuracy does not certify the transpose, and removes it by construction rather than by training: the three constraints make the lifted model reduce exactly to the solver’s tangent dynamics and classical adjoint as the amplitude vanishes, with the learned part confined to the second-order remainder. The quiescent-Burgers study in Appendix C, where an unconstrained lift with identical forward error amplifies a backward-propagated sensor kernel by two orders of magnitude, makes this argument concrete.
  2. The evaluation is thorough and candid. Results are compared against brute-force inversion through the solver with bootstrap intervals, against classical linear and quadratic adjoints, structural ablations, and several learned alternatives, and the posterior likelihood is widened by each model’s own discrepancy so that calibration can be compared fairly; the appendix also reports the settings in which the proposed lift is not the best model.
Critical weaknesses
  1. All experiments are at small scale: 48×48 grids, low-Reynolds-number laminar base states, and dense learned blocks of size cN×cN that already exceed GPU memory for the ablation variants at four channels. The translation-equivariant variant that avoids this applies only to uniform base states on periodic domains, so it is unclear how the approach extends to the turbulent, non-uniform settings the paper cites as its motivation.
  2. The headline claims are drawn from the favorable cases. The “twentieth of the cost” posterior result holds only on reaction–diffusion; on Kolmogorov flow the two-source inference falls well short of the solver (Table 10); on reaction–diffusion the unconstrained lift is more accurate at finite amplitude; and the quadratic correction gives more accurate weak-source gradients on Kolmogorov flow (Table 8). The abstract and introduction should reflect this more balanced picture.
Questions for authors
  1. The learned blocks are dense matrices of size cN×cN, and the problems are limited to 48×48 grids with laminar base states; the translation-equivariant variant in Appendix A.3 applies only to uniform base states on periodic domains and is not used for Kolmogorov flow. Could the authors discuss how the formulation scales to larger grids and non-uniform base states, and report the offline cost of forming L_k by finite differences and of training, which are not included in Table 1?
  2. On Kolmogorov flow the two-source inference falls well short of the solver (Table 10: 20.8–37.5% of pairs resolved against 70.8–100% for brute force), and no joint posterior is reported for this system. Could the authors report the Kolmogorov posterior and comment on whether the gap stems from the lift’s finite-amplitude error under strong interaction, which Table 10 puts at 43–48%?
  3. In Tables 11 and 12 a ridge readout fitted on the lift’s own features matches or outperforms the lifted adjoint on Kolmogorov flow wherever training data exist at the target layout and horizon. Could the authors clarify what the transposed-propagator construction contributes beyond the learned features, other than extrapolation to layouts and horizons without data, and whether the single-source cost comparison in Table 1 remains meaningful given that a solver dictionary can be cached per layout, as Section 5 notes?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer KBGM

We thank the reviewer for the clear summary.

W1 / Q1. Scale and non-uniform base states. We agree that the experiments are at low resolution about laminar base states. The memory bottleneck is the dense learned feature block. For uniform base states the translation-equivariant blocks reduce it by three orders of magnitude, and for larger or non-uniform problems the learned channels need not keep the solver's resolution. Only the state block must, because it is the solver's own propagator. An encoder that pools the learned channels onto a coarser grid of M≪NM \ll N points would make the learned blocks grow as (cM)2(cM)^2 instead of (cN)2(cN)^2, and the discussion now states this as the main step toward inhomogeneous and turbulent flows. The solver enters only through the products LkqkL_k q_k and its own adjoint. Each product is a Jacobian–vector product of one solver step, which a tangent-linear code or automatic differentiation of the solver returns at a small multiple of the cost of that step, so LkL_k is never needed as a matrix. The encoder's cost per candidate grows about linearly with resolution, and on our grids forming LkL_k takes under a second and training about eight minutes (Table 13).

W2. Balance of the headline claims. We agree and have revised the abstract and introduction (general comment, item 3). At four channels the constrained lift is more accurate than the unconstrained lift in every band (item 1). The quadratic correction gives the most accurate weak-source gradients on Kolmogorov flow, and the paper now says so.

Q2. Kolmogorov two-source inference. We sampled the Kolmogorov joint posterior of 8 overlapping pairs with the same sampler as on reaction–diffusion, with chains long enough that every split-R^\hat R is below 1.1, and the two-source appendix now reports it. Its 95% intervals contain the truth about as often as the solver's and are wider in position, because the lift's likelihood carries its own discrepancy under the strong interaction. On point estimates, brute force uses the same search and refinement and finds every separated pair, while the lift and the linear embedding both resolve fewer pairs as the sources become stronger and then overlap. As the reviewer suggests, the gap therefore comes from nonlinearity that the added channels do not capture, and the main text now says so.

Q3. What the transposed propagator contributes beyond the features. The reviewer is right that, at a layout and horizon with training data, a ridge readout on the lift's features is about as accurate as the lifted adjoint (Table 11), and we now say so. The propagator contributes in three ways. It is what trains those features, since an anchored and centered CNN of the same architecture trained end to end is about an order of magnitude less accurate on reaction–diffusion. It makes them data-efficient, since with 50 training trajectories the lift is up to four times more accurate than the readout. And it lets one model serve any layout or horizon with one backward solve, including 1.5T1.5T, where the readout has no data and the lift's error stays below 17%. The cached dictionary serves a fixed lattice at a fixed layout and horizon, so it competes only with single-source grid search; it gives no continuous refinement, gradients or joint posteriors, which is where we place the cost case.

Thank you for the detailed response and additional experiments.

For W1/Q1, the analysis and results partially address my concerns on scalability. However, it’s still unclear whether reducing the learned feature resolution would preserve sufficient accuracy for more complex flow dynamics. I am totally satisfied with the response to W2 and Q2 without any more doubts. For Q3, the discussion helps clarify the value of the transposed-propagator construction better, while I think it could be better if the authors could more explicitly describe the characterization of the workloads where the proposed method prefer.

Overall, the rebuttal address most of my concerns and the quality of the manuscript has been improved. I would maintain my current rating.

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

The paper infers finite-amplitude localized sources from sparse sensors in nonlinear PDEs: Burgers, 2D reaction–diffusion and Kolmogorov flow. It lifts the perturbation to [q; Φθ(q)], with features centred to O(‖q‖²) and a block-triangular propagator whose state block is the solver's linearized propagator . Its transpose gives a per-sensor "lifted adjoint" scoring each candidate with one encoder pass and two inner products . It claims small-amplitude exactness, near-solver inversion and faster two-source posteriors.

Critical strengths
  1. The construction is correct. I re-derived identities (i)–(iii) in the method section: the state blocks of propagator products and of Z_j† are exactly the tangent model and linear adjoint, with O(ε²) measurement error.

  2. Finite-amplitude gains are large on all three systems: εm is 1.8/2.4/7.7% against 72.9/164.5/41.3% for the linear adjoint (as shown in Table 1). At I_s=10⁻³ an unconstrained lift errs by 0.23 (RD) and 0.67 (KF), against 8×10⁻⁵ and 9×10⁻⁵ (Table 2).

  3. Strong physics baselines, solver-checked gradients, discrepancy-widened likelihoods , very candid cost caveats.

Critical weaknesses
  1. The lift's own contribution is not isolated. The only non-lifted learned model in Table 2, the neural surrogate , is not constrained to be exact. It uses the uncentred Φ̃θ plus a bias b, so a zero source yields q₁=WΦ̃θ(0)+b, and its errors of 124 (RD) and 9 (KF) at I_s=10⁻³ (Table 2) are expected. Centred by Eq. 5 with b=0, it would also match the tangent model up to O(ε²). C3 is never removed alone (L269–270), although it keeps identity (i) beyond one step: with a lower-left block C, the two-step state block is L₁L₀+G_qΦC. Please add the centred surrogate and a C3-only ablation at c=4/24, not c=2 (L310–312).

  2. "A twentieth of its cost" (L032–033) is not established. [Unsupported] A cached solver dictionary could serve single-source grids (L305–308), so the cost case rests on the posterior. On RD, a grid pass is only 21.3/11.9 = 1.79× faster than the solver (Table 1), but the posterior is 295/13 = 22.7× faster (L402–403). It uses a different, eight-channel lift (L374–376); the main text gives no evaluation counts, batching or offline cost. The reference under-covers: the solver's 95% intervals contain the truth in 85% of 102 cases (L400–402), 4.5 binomial standard errors low (1.8 if only the 17 pairs are independent). The lift's σ_mod widening (L216–219) may inflate its 87%.

  3. Inference comparison is weakly resolved and fails for KF pairs. For Euclidean δx in candidate spacings , continuous positions and a square grid, a nearest-candidate oracle has median √(0.5/π) ≈ 0.40. The KF lift and solver both report 0.41, and even the RD linear embedding reaches 0.63 (Table 1). For separated strong KF pairs, the lift places both sources within one cell in 38% of cases, "against … all pairs for brute force" (L419–421). "Nearly as accurately as inversion through the solver" (L030–031) needs qualification.

  4. Close structural precedents are not discussed. Truncated Carleman linearization, treated only as an alternative, already has the C1–C3 form: for x_{k+1}=Lx_k+H(x_k⊗x_k), [x; x⊗x] evolves under [[L, H],[0, L⊗L]] up to O(‖x‖³). The invariant-subspace example of Brunton et al. (PLoS ONE 2016, Eq. 25) has the same pattern: (x₁, x₂, x₁²) under [[μ,0,0],[0,λ,−λ],[0,0,2μ]]. Eqs. 2–3 are the dual of any linear lifted predictor (L188–190), and state-inclusive predictors are standard (Korda & Mezić, Automatica 2018). Transfer-operator Bayesian source identification exists for passive transport (Sharma et al., Energies 2021). Please state the delta.

  5. Exactness is guaranteed only where the linear embedding already works; on RD it costs finite-amplitude accuracy. At I_s=10⁻³ the linear embedding errs by only 6×10⁻⁴ (RD) and 2×10⁻⁴ (KF) (Table 2). On RD the constrained lift trails the unconstrained one in top-band measurement (6.0% vs 3.2%) and training-range gradient (21% vs 7.8%). On KF its training-range gradient error is 17%. "Competitive" (L423) and "reliable instrument for nonlinear sensitivity" (L035–036) overstate this. Please report gradients for the c=4/24 lifts.

Questions for authors
  1. How does the centred residual surrogate compare in accuracy and per-candidate cost? A clear advantage for the lift would raise my contribution score.

  2. For 13 s vs 295 s, what were the evaluation count, batching, hardware and offline cost? An equal-batching ~20× speedup would resolve W2's cost point.

  3. Why does the solver posterior under-cover (sampler, R̂/ESS, label handling), and what are lift-to-solver width ratios? A calibrated reference would resolve the rest of W2.

  4. Is δx Euclidean with continuous positions (is the 13,824-candidate grid 24³)? Do KF pair failures come from the functional or the pair-only gradient refinement (L199–200)? The latter would soften W3.

  5. What are the errors on new sensor layouts (none in the main text, L474–477)? Errors near Table 1's would support the reuse claim.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer JPdr, part 1/2

We thank the reviewer for re-deriving identities (i)–(iii) and for precise questions.

W1 / Q1. Centered surrogate and C3-only ablation. Together the two controls separate the lifting from the constraints: the first removes the lifting and keeps the centering, and the second keeps the lifting and removes (C3). We trained the centered surrogate the reviewer describes, qk+1=Lkqk+WΦθ(qk)q_{k+1} = L_k q_k + W\Phi_\theta(q_k) with Φθ\Phi_\theta centered by Eq. 5 and no bias, with the same encoder and trainer as the lift. Table 2 lists it as "encoder in the loop" and the original surrogate as "learned correction (uncentered)". As the reviewer predicts, centering makes it exact at small amplitude. At two channels it is more accurate than our lift for the stronger sources. At four channels our lift is more accurate in top-band measurement on both systems and in gradient on reaction–diffusion (RD), and the two are close in gradient on Kolmogorov flow (KF). On per-candidate cost, the surrogate rolls out each candidate over all time steps and applies the encoder at every step, whereas the lift needs one encoder pass and two inner products (Eq. 3), so a pass over the candidate grid takes about 32 times longer with the surrogate. For the original uncentered surrogate we chose the learning rate by validation, which changes its entries in Table 2.

The C3-only ablation ("without (C3)" in Table 2) keeps (C1) and (C2) and fits the lower-left block CC. As the reviewer notes, identity (i) then fails from the second step on, and the experiment confirms it. Its small-amplitude error levels off instead of vanishing as it does with (C3), and its weak-source gradient error is about ten times larger on both systems. The 24-channel Burgers lift shows the same in both ablations (Appendix B.2), where the encoder-in-the-loop model is close to the lift in accuracy and 51 times slower.

W2 / Q2. "A twentieth of its cost". Both models run the same differential-evolution MCMC sampler with 96 chains, 800 burn-in and 1,600 kept steps, so 230,400 likelihood evaluations per posterior, and the 96 proposals of each step are evaluated as one batch. On one A100 with both models in double precision, a posterior takes 17 s with the lift against 282 s with the solver. The paper now says "a sixteenth" and states the counts, the hardware and that the posterior uses the eight-channel equivariant lift, whose offline cost is in Table 13. The grid-pass ratio of 1.8 in Table 1 came from evaluating our encoder in double precision on a consumer GPU (general comment, item 2).

W2 / Q3. Calibration. We reran the posteriors of all 17 pairs for both models with chains ten times longer. Every split-R^\hat R is now below 1.1 and the smallest effective sample size per pair is above 700. With these chains the lift and the solver cover the truth equally often, in 85% of the 102 cases, the misses fall on largely the same pairs, and the posterior means differ by a median of 0.6 solver standard deviations. The lift's widened likelihood broadens its posteriors without raising its coverage above that of the well-mixed reference. Since the sampler is no longer the cause, the shortfall from 95% is a property of the test rather than of either model. The 17 pairs are a fixed set of overlapping sources rather than draws from the prior, and the six intervals of a pair share one noise realization and tend to miss together. On labelling, chains are relabelled at every step to agree with the best chain, and coverage aligns every sample to the true pair before taking quantiles. The appendix reports R^\hat R and effective sample sizes per pair.

AuthorsOpenReview ↗

Reply to Reviewer JPdr, part 2/2

W3 / Q4. Resolution of the inference comparison. Yes, δx\delta_x is the Euclidean distance on the periodic domain in units of the candidate spacing, true positions are drawn continuously, and the 13,824 candidates are 24×24 positions times 24 intensities. We agree that the Kolmogorov lift and solver both sit at the nearest-candidate floor. Table 1 now states the floor, and the abstract qualifies the claim as accuracy at the resolution of the candidate lattice.

The Kolmogorov pair failures come mainly from the functional, not from the search. Brute force uses the same search and refinement and finds every separated pair, while the lift and the linear embedding both resolve fewer pairs as the sources become stronger and then overlap (Table 10). The lift's gap to brute force therefore comes from nonlinearity that its added channels do not capture, and the two-source appendix now says so.

W4. Structural precedents. We agree that the block-triangular form is classical. Truncated Carleman linearization has this structure for polynomial systems, as does the invariant-subspace example of Brunton et al. (2016), and the related work now cites both and describes SALTO as a generalization of truncated Carleman linearization. SALTO makes the structure learnable for non-polynomial and solver-defined dynamics, with closed-form centering for any smooth network and closed-form fits of the learned blocks. It anchors the state block to the solver's own propagator, including a time-varying one, and uses the transpose as a reusable inference functional that reduces to the classical adjoint. State-inclusive predictors (Korda & Mezić, 2018) serve prediction and control and impose no such reduction. Sharma et al. (Energies, 2021) identify sources with the Perron–Frobenius operator of passive transport, which is linear in the source, whereas SALTO targets responses that are nonlinear in the source amplitude; the related work now cites it.

W5. Exactness and finite-amplitude accuracy. Retrained at four channels, the constrained lift is more accurate than the unconstrained lift in every gradient band (general comment, item 1) and in top-band measurement on reaction–diffusion. The 24-channel Burgers lift also passes the gradient check, with errors far below those of the linear embedding in the strongest training band (Appendix B.2). We agree that exactness matters most where the linear embedding is already accurate, and we have replaced "competitive" and "reliable instrument" with specific statements.

Q5. New sensor layouts. On the held-out 4×4 and 25-random-probe layouts, the four-channel lift is as accurate as on the training layout on reaction–diffusion. On Kolmogorov flow it stays within a factor of two of its training-layout error, and the 4×4 layout is harder for every model. The appendix (Table 11) now gives these errors.

Thanks for the detailed reply and additional experiments. The centred surrogate and the C3-only ablation were what I mainly wanted to see, and they help a lot. The lift doesn’t win everywhere, but at four channels it’s mostly ahead and much cheaper per grid pass. The C3 ablation also behaves the way the derivation says it should.

The cost and calibration answers settle my W2 concerns. Running the same sampler and batching for both models (17 s vs 282 s) makes the comparison clean. With the longer chains, the lift and the solver also end up with the same coverage. I’m also happy with the added related work and the toned-down abstract. One request: please keep the KF pair failures clearly stated as a limitation.

I’ll keep my current score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment (to all reviewers and the AC)

We thank the reviewers for careful and constructive reviews. Three requests are shared by several reviewers, and we answer them here. The replies to each reviewer answer the remaining points. Changes in the paper are in red.

1. Ablations and gradients at four channels (all reviewers). The reviewers asked whether the trade-off in our two-channel gradient table persists for larger lifts. We retrained every ablation at four channels, the size of the two-dimensional main results, and ran the same gradient check. We also trained both new ablations for the 24-channel Burgers lift and ran the gradient check on it (Appendix B.2). At four channels the constrained lift is more accurate than the unconstrained lift in every band on both systems. Table 2 and Table 8 give every model at two and four channels.

median gradient error (%)RD [0.01,0.1)[0.1,0.5)[0.5,1)[1,2)[2,3)†KF [0.01,0.1)[0.1,0.5)[0.5,1)[1,4)[4,6)†
lifted, c=40.10.61.35.5180.51.74.89.328
unconstrained, c=4 (new)2.11.93.48.2211312162030
lifted, c=20.73.37.521730.62.67.91735
unconstrained, c=24.92.94.77.8241613182537

† beyond the training range.

2. Cost (JPdr W2, KBGM Q1, 6Z3n W1). We re-measured every time on one A100 GPU with every model in double precision at the same batch size. The ratio of 1.8 in Table 1 came from evaluating our encoder in double precision on a consumer GPU, which runs double precision at 1/64 of its single-precision rate. One pass over the candidate grid now takes 0.58 s with the lift against 3.55 s with the solver on reaction–diffusion and 40.4 s on Kolmogorov flow. A two-source posterior with the same sampler and 230,400 likelihood evaluations takes 17 s with the lift against 282 s with the solver. Offline, the training data, the propagator and the adjoint solves take seconds, and training takes about eight minutes for a four-channel lift. As Section 5 states, a cached solver dictionary serves repeated single-source grids at a fixed layout and horizon equally well, so the cost case rests on the joint posterior. All costs are in Appendix B.6, Table 13.

3. Scope of the claims (JPdr W2, W3, W5; KBGM W2). We revised the abstract and introduction. The posterior cost ratio is stated for reaction–diffusion. "Nearly as accurately as the solver" now refers to single-source inference at the resolution of the candidate lattice. The Kolmogorov two-source gap is stated in the main text. "Reliable instrument" and "competitive" are replaced by the specific properties shown.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.