acceptodds
Withdrawn from ICLR 2027

Coordinate Closure Recovers Nonlinear Coupling in Coarse PDE Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Coarse PDE prediction can lose a nonlinear contribution even when current samples identify the structure that produces it. If a coherent coordinate has a continuation beyond the grid cutoff, projecting it before multiplication deletes cross interactions that return to represented modes. We introduce coordinate closure (C0): recover the coordinate from the current field, retain the visible residual, evaluate the finite returned-band interaction from analytic coordinate coefficients, and allocate this forcing between coordinate and residual dynamics. A Fourier identity localizes the required continuation and gives an exponential tail envelope. On Michelson–Sivashinsky fields, the complete transition beats exact-product controls on all 48 compatible outer fields. A cloned-state mask isolates a smaller 2.7–3.3% tail effect in the two clean low-band cells and its resolution and rank-zero exits. An independently implemented Benjamin–Ono transition reproduces tail sensitivity and the resolution exit under a different coordinate law. In a parameter-matched FNO, the interaction feature lowers pooled current-field forecast error by 14.9%; the coordinate decomposition and history arms locate the gain in the explicit interaction token. Resolved and coordinate-free boundary fields place the learned benefit in the active, compatible regime. The results identify coordinate closure as a conditional same-observation representation for coarse nonlinear prediction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
coarse-grid PDE prediction, structure-aware representations, neural operators, subgrid-scale modeling.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LwdWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces coordinate closure for coarse PDE prediction. It recovers coherent structures from current observations and retains their unresolved Fourier contributions when computing interactions with the resolved residual. Experiments cover deterministic MS and BO predictions and a feature-based FNO comparison.

Critical strengths

The paper identifies a concrete representation defect and derives the finite frequency band needed to recover the missing interaction. The cloned-state tail-removal control provides targeted evidence beyond full-method comparisons. The authors clearly acknowledge conditional benefits and negative results.

Critical weaknesses
  1. The learned comparison has an architectural confound. The raw-input arm activates one token, disabling its query/key aggregation pathways, while coordinate arms activate multiple tokens. Equal-active-token and alternative nonlinear-feature controls are needed to isolate the proposed interaction’s value.

  2. The isolated benefit is modest and conditional. T0 shows approximately 2.7–3.3% improvement in clean low-band cells, but near-equality in noisy and overlap-band cells. The larger full-method gains cannot be attributed to the returned interaction alone.

  3. Recovery-to-closure sensitivity remains unclear. Coordinate and residual estimation errors are coupled, so accurate sample reconstruction need not imply accurate closure terms. A local perturbation analysis would strengthen this connection.

  4. The coupled update needs a clearer mathematical interpretation. The relationship among coordinate dynamics, tangent allocation, residual truncation, and re-encoding should be clarified through the induced field-space dynamics.

The current evidence does not yet fully demonstrate the value of the underlying idea, but I would be open to raising my score if the rebuttal adequately addresses the main concerns.

Questions for authors

Does the learned improvement persist when the active-token count is matched and alternative nonlinear features are provided?

How do coordinate estimation errors propagate into the returned interaction?

What consistency relation does the complete coordinate–residual update satisfy?

Reviewer_C5pSOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes coordinate closure to recover missing interactions between unresolved coherent structures and visible residuals in coarse PDE prediction. The method recovers coordinate parameters and retains the observed residual. It computes their interactions analytically and uses the resulting forcing to update both components. Validation includes Michelson–Sivashinsky and Benjamin–Ono experiments. The experiments include exact-product controls, tail-removal tests, and an FNO study with matched parameter shapes. The theory identifies the required Fourier modes and bounds the interaction strength under spectral decay assumptions.

Critical strengths

Explicit Characterization of the Missing Interaction: The paper identifies the coordinate–residual interaction removed by truncation and specifies the finite Fourier shell needed to recover it. This provides a concrete computational target rather than an unspecified closure correction.

Controlled Attribution of the Tail Contribution: The cloned-state T0 intervention preserves the recovered state and numerical schedule while removing only the coordinate tail from the cross product (Sec. 4.2). Resolution and rank-zero checks further distinguish this narrow mechanism from complete-method gains.

Critical weaknesses

Absence of Matched Controls for the Learned Representation Advantage: The learned comparison changes active token counts, leaving R’s query/key aggregation inactive while enabling those pathways in C. Without equal-active-slot or ordinary nonlinear-feature controls, the reported improvement cannot distinguish physical interaction content from architectural activation effects (Table 1). Establishing the claimed learned utility requires additional controlled experiments.

Presentation Quality Requires Further Refinement: The figures and overall layout would benefit from further refinement to meet conference publication standards. For example, the layout on page 9 could be improved by moving some of the subsequent supplementary material onto that page. Several lines contain only one or two words (e.g., lines 260, 237, 115, 077, and 039), resulting in awkward line breaks and inefficient use of space.

Questions for authors

1.Could the authors provide equal-active-token and standard nonlinear-feature controls to determine whether the reported gains arise from the physical interaction features rather than differences in active aggregation pathways?

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

This paper proposes Coordinate Closure (C0) for coarse-resolution PDE prediction. The main idea is to reconstruct analytically recoverable coherent structures from current coarse observations and retain their nonlinear interactions with resolved residual modes before spectral truncation. The method combines coordinate recovery, analytic coupling evaluation, and tangent-residual dynamics. Experiments on Michelson–Sivashinsky and Benjamin–Ono equations demonstrate improved forecasting in compatible regimes.

Critical strengths
  • The paper addresses a well-motivated problem in coarse PDE prediction and provides a clear mathematical characterization of the missing nonlinear interaction.

  • The experimental methodology is rigorous, including exact-product spectral baselines, time-refinement controls, cloned-state interventions, and resolution-dependent tests.

  • The second PDE implementation and comprehensive negative controls strengthen the empirical investigation. The authors also carefully distinguish the isolated physical mechanism from overall solver improvements.

Critical weaknesses
  • Limited theoretical novelty. Reconstructing unresolved high-frequency components and retaining their nonlinear influence is well established in nonlinear Galerkin and approximate inertial-manifold methods. The main identity follows from Fourier convolution, and the contribution appears to be a specialized construction rather than a fundamentally new closure principle.

  • Limited isolated mechanism effect. The cloned-state intervention shows only a 2.7–3.3% improvement in two clean low-band cases, with negligible or inconclusive effects in noisy and overlap-band conditions. Much larger full-method gains involve additional changes to coordinate dynamics and output representation.

  • Insufficient attribution of learned gains. Although parameter counts match, the raw-input and interaction-aware FNOs activate different numbers of tokens, changing their effective attention mechanisms. Moreover, substantial learned gains occur in regimes where the isolated physical tail effect is nearly zero. Equal-active-slot and in-band interaction controls are missing.

  • Restricted generalizability. The successful experiments rely on analytically recoverable structures in specialized one-dimensional periodic PDE families. The failed KdV/Allen–Cahn transfer tests, poor learned performance on boundary regimes, and additional computational cost limit the broader applicability.

Questions for authors
  • Could the authors provide an equal-active-slot learned baseline, replacing the returned-interaction feature with an in-band nonlinear interaction or another matched control? This would help isolate the value of the proposed physical feature.

  • Why does the learned model exhibit substantial improvements in overlap/noisy conditions where the isolated deterministic tail effect is nearly zero? Could the authors analyze this discrepancy using the existing experimental records?

  • Could the authors clarify the methodological novelty relative to nonlinear Galerkin and approximate inertial-manifold methods, particularly regarding high-mode reconstruction and feedback into resolved dynamics?

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

Related papers

Discussion (0)

Sign in to comment.