acceptodds
Withdrawn from ICLR 2027

RF-LIN: Linear PDE Representations via Invertible Random-Feature Coordinates

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Linear representations of nonlinear PDEs must preserve state information while capturing dynamics. We introduce RF-LIN, which learns invertible random-feature coordinates and a finite-dimensional linear evolution operator. Frozen nonlinear features make each appended additive-coupling readout a conditional least-squares problem. For fixed coordinates, a projected singular-value solve fits the operator with a stationarity certificate in the retained subspace. Exact reconstruction and multistep composition separate coordinate invertibility from dynamical accuracy. We derive a finite-horizon error bound linking latent residuals, matrix-power growth, and decoder sensitivity, and show that a disconnected complete fixed-point set precludes global homeomorphic linearization. In three-seed Burgers experiments at two viscosities, RF-LIN achieves mean spectral errors of 0.927% and 4.415%, reductions of 28.3% and 17.1% relative to RF-EDMD, the strongest tested spectral baseline. Reconstruction errors remain below 2.8 × 10^-14, while RF-EDMD yields lower mean rollout error. Allen–Cahn tests expose a spurious near-fixed bridge between attractors. A single Kuramoto–Sivashinsky validation window exhibits strong transient amplification and poor rollout accuracy despite a spectral radius below one. These results demonstrate spectral gains from learned invertible coordinates and identify geometric and finite-time limits on their predictive use.

Primary area:
learning on time series and dynamical systems
Keywords:
Koopman operators, random features, invertible neural networks, linear representations, PDE dynamics, spectral learning

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_GzhwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Paper proposes RFLIN, a framework for learning invertible random feature coordinates in which nonlinear PDE dynamics are approximated by a linear operator. It combines conditional least-squares fitting with an analysis of prediction errors and geometric obstructions to global linearization.

Critical strengths

separation of coordinate invertibility, spectral accuracy, and dynamical prediction is mathematically meaningful. The finite horizon error bound and fixed point obstruction provide useful theoretical perspectives. Experiments are transparent about the limitations, particularly the distinction between recovering operator eigenvalues and accurately forecasting PDE trajectories.

Critical weaknesses

main concern is that improved spectral recovery does not translate into better forecasting. Table 1 shows that RFEDMD achieves substantially lower rollout errors at both viscosities, while the Kuramoto Sivashinsky experiment reports approximately 130% mean prediction error. Therefore, the practical advantage of the proposed representation remains unclear. theoretical contributions are also somewhat limited. The least squares certificate establishes conditional optimization stationarity rather than dynamical accuracy, and Theorem 2 assumes acceptance conditions that the implemented normalized loss checks do not necessarily satisfy. Moreover, comparisons with existing invertible Koopman methods are missing, and the experiments lack component ablations and computational cost measurements. These omissions make it difficult to isolate the contribution of the proposed random feature parameterization.

Questions for authors

Can the authors explain why improved spectral accuracy consistently produces worse rollout performance than RFEDMD? What evidence isolates the benefit of conditional random feature fitting from invertibility alone? How do the theoretical monotonicity guarantees relate to the normalized loss acceptance criterion actually used in training?

Reviewer_kaj6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RF-LIN to model nonlinear PDE dynamics using invertible coordinates and a linear operator. Fixed random features allow each new coordinate block to be fitted through least squares. Experiments on three PDE systems examine spectral accuracy, prediction errors, and limitations of the representation.

Critical strengths
  1. The fitting method is clearly motivated. It allows coordinate blocks to be fitted through linear problems while preserving the information in the original state.

  2. The paper clearly distinguishes reconstruction accuracy from prediction accuracy. It also reports failure cases and explains why a small fitting error does not necessarily lead to accurate long-term predictions.

Critical weaknesses
  1. The technical innovation needs stronger support. Invertible Koopman models already exist, and the paper identifies its fitting method as the main contribution. However, it does not compare with the closest methods, such as FlowDMD and EDMD-IDL. What does the proposed fitting method improve in accuracy or computational cost? The paper also uses gradient refinement in one Burgers setting. Ablations are needed to show how much improvement comes from the proposed least-squares fitting.

  2. The numerical results do not clearly establish the practical benefit. RF-LIN improves spectral accuracy, but RF-EDMD gives lower rollout error at both Burgers viscosities. The spectral comparison only considers eigenvalue magnitudes, so it does not establish that the learned modes correctly represent the dynamics. The KS result uses one validation window and has about 130% mean rollout error. Stronger comparisons are needed to show when this representation is useful. Also, the implemented acceptance rule does not establish the conditions required by Theorem 2, so the connection between this theorem and the reported algorithm needs clarification.

Questions for authors
  1. How does RF-LIN compare with existing invertible Koopman models in accuracy and training time?

  2. How much improvement comes from least-squares fitting rather than gradient refinement?

  3. What practical task benefits from the improved spectral accuracy when the rollout error is higher?

  4. Does the implemented algorithm satisfy the conditions of Theorem 2?

Reviewer_6dHYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper concerns learning linear representations of nonlinear PDE dynamics. The paper aims to learn a coordinate map gg and a matrix AA such that g(SΔt(u))≈A g(u)g(S_{\Delta t}(u)) \approx A\,g(u), where SΔtS_{\Delta t} is the time-step map of the spatially discretized PDE.

While prior Koopman methods fit a linear operator to state snapshots, to fixed observables, or to learned observables, the representation can lose state information, and the linear dynamics can fail to predict how the retained information evolves. Invertible coordinate maps avoid the loss of state information as the inverse recovers the state exactly. This leaves open how to fit the coordinate map so that the operator solve stays explicit and its numerical error measurable. Towards this, the paper proposed RF-LIN, which builds the invertible maps from additive coupling layers with frozen random features, and trains only their linear readouts. This turns training into alternating least square solves, and the layers stay invertible for any readout so the reconstruction stays exact. An optionally, all readouts can be tuned jointly via the Adam optimizer.

Experimental results on Burgers at two viscosities show that the 20 leading eigenvalue magnitudes of RF-LIN match the analytic heat-flow decay factors with lower mean error than exact DMD, degree-3 polynomial EDMD, and RF-EDMD. However, RF-EDMD has lower mean rollout error at both viscosities. On two-basin Allen–Cahn, the reference dynamics move the states on the decoded segment between the two encoded equilibria, while RF-LIN nearly fixes them. On one Kuramoto–Sivashinsky validation window, the mean rollout error is 130% against a 1.8% one-step error, and AA shows strong transient amplification.

The paper also discusses where such a linear model can fail. A finite-horizon bound separates three sources of forecast error: (1) the one-step latent residual, (2) its accumulation over repeated applications of AA, and (3) the Lipschitz constant of the inverse map. A fixed-point result shows that a global homeomorphic linearization requires a connected fixed-point set. Finally, the paper concludes that a useful spectral representation and an accurate forecaster are related but distinct goals, and states that model selection should check spectral agreement together with decoded rollout error and fixed-point geometry.

Critical strengths
  • The paper details an explicit linear-algebraic training procesure, where each subproblem is convex and has an explicit solution.
  • The additive couplings have an analytical inverse for any readout.
  • The testbeds are principled. Burgers with a Cole-Hopf reference is an analytic test for Koopman-type spectra, Allen-Cahn probes multi-attractor geometry, and the diagnostics separate fit residual, non-normal amplification, and decode sensitivity.
  • The paper also explicitly details limitations and provides code for reproducibility checks.
Critical weaknesses
  • The paper's novelty is thin. Additive and affine couplings, frozen random features with a trained linear readout, closed-form least squares inside a learned-observable training, are all prior methods. The new element is that frozen conditioners make the readout of one appended block a linear problem. This follows directly because an additive shift is linear in WW. The paper does not show that this property gives better accuracy, lower cost, or more robustness than gradient-based training of the same architecture. The reported ν = 0.03 configuration uses Adam refinement in any case.
  • FlowDMD, EDMD-IDL, Koopmanizing Flows, and invertible Koopman neural operators are cited as closest prior work but not compared.
  • The baseline comparison seem to be unmatched. RF-EDMD uses 256 random cosine features, no state in the dictionary, and a fixed bandwidth of 1.0 in every configuration. By contrast, RF-LIN uses 6 × 512 tanh features and up to 6 × 128 × 512 = 393,216 readout coefficients. Its bandwidth is tuned over {0.25, 0.5, 1, 2, 4}. As per provided supplementary material.
  • The theory is mostly elementary or known. Consider referring to Theorem 1 as a Lemma instead.
Questions for authors

See weakness.

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

Related papers

Discussion (0)

Sign in to comment.