acceptodds
Under review as a conference paper at ICLR 2027

Width Controls Wasserstein Robustness in Learned Spectral Projectors

Abstract

A spectral projector may vary continuously even when every narrow continuous factor incurs order-one error somewhere. To the best of our knowledge, we establish the first sharp joint characterization of output-factor width, and Wasserstein robustness for spectral projector prediction on real Grassmann manifolds. For predictors on , let have output columns, , and . For , set . At fixed dimensions, in the large-, small- regime, the optimal in-distribution squared operator-norm risk under Haar sampling is , while the optimal risk over a ball of radius is for . Imposing a density-ratio cap yields . Matching lower bounds and explicit continuous constructions turn a topological existence obstruction into quantitative approximation and distribution-shift exponents. The lower bounds apply to every neural architecture satisfying the stated constraints; matching neural upper rates remain architecture-dependent. The exponents persist under input densities bounded above and away from zero, bi-Lipschitz coordinate changes, and uniformly gapped spectral deformations. Consequently, at fixed subthreshold output width, vanishing in-distribution error can coexist with an unavoidable robust-error floor at a fixed positive shift radius, while the density cap determines whether this floor persists. Frozen-backbone Fourier neural operator experiments on elliptic, Darcy, and plate problems support this distinction: widening only the output factor head substantially reduces targeted shifted error, whereas a sixteenfold increase in the size of a one-column plate model leaves its shifted-error floor essentially unchanged. Analytic rate tests and structural-mode experiments provide additional checks across ranks and discretizations.

Primary area:
learning theory
Keywords:
spectral projector, matrix factorization, distributional robustness, topological obstruction, Grassmann manifold

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_C4jAOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies approximation of spectral projectors P(ξ)P(\xi) associated with parameter-dependent operators A(ξ)A(\xi), using predictors of the form W(ξ)W(ξ)⊤W(\xi)W(\xi)^\top. Although spectral projectors can vary continuously, it is known that topological obstructions can prevent their exact representation by continuous factors with too few columns. The paper aims to quantify the resulting approximation error and its sensitivity to distribution shift.

In the canonical setting X=Gr(k,N)X=\mathrm{Gr}(k,N) with Haar measure μ\mu, it considers Lipschitz predictors W:X→RN×mW:X\to\mathbb R^{N\times m} approximating the orthogonal projector P(x)P(x). For k≤m<Nk\le m<N, with r=m−k+1r=m-k+1, the authors establish sharp rates:

\mathbb E_\mu\|WW^\top-P\|_{\mathrm{op}}^2 \asymp L^{-r},$$ and $$\inf_{\operatorname{Lip}(W)\le L, \sup_{x\in X}\|W(x)\|_F\le\sqrt{k}} \sup_{W_p(\mu,\nu)\le\varepsilon} \mathbb E_\nu\|WW^\top-P\|_{\mathrm{op}}^2 \asymp L^{-r}+\varepsilon^{pr/(p+r)}.$$ The results combine topological lower bounds with matching constructive upper bounds, and are supplemented by extensions to density-constrained shifts and numerical experiments on neural-operator approximations of spectral problems.
Critical strengths
  1. The paper provides a rigorous quantitative analysis of a known topological obstruction to continuous spectral-projector factorization.

  2. The matching upper and lower bounds establish sharp approximation rates, with an explicit dependence on the output width.

  3. The extension to Wasserstein distribution shift is mathematically interesting, yielding sharp robustness rates and additional results under density-ratio constraints.

Critical weaknesses
  1. Restrictive distributional assumptions. The sharp rates rely on Haar measure or distributions with densities uniformly comparable to it. It is unclear whether similar results hold for more general distributions, particularly those supported on lower-dimensional subsets of the Grassmannian.

  2. Novelty of the techniques. The analysis builds on well-known topological obstructions and geometric arguments. The paper could better clarify which technical ingredients are genuinely new.

  3. Dimension dependence. Rates are sharp only for fixed ambient dimension NN, and the dependence of the constants on NN is not sufficiently characterized. This could be a significant limitation given that the motivating applications involve high-dimensional operators.

  4. Presentation. Main body is overly dense and lacks pedagogical explanations. The intuition behind the results, the role of the assumptions, and the progression of the arguments are unnecessarily difficult to follow.

Questions for authors
  1. To what extent do the rates depend on the Haar measure assumption? Can analogous results be established for distributions supported on lower-dimensional subsets of the Grassmannian?

  2. What are the main new mathematical ingredients needed to derive the sharp rates, beyond the qualitative topological obstruction discussed in prior work such as Esmaeili et al. (2023)?

  3. Can the authors characterize the dependence of the constants in their upper and lower bounds on the ambient dimension NN? In particular, do the rates remain informative as NN grows?

  4. Can the theoretical insights be used to design practical architectures that avoid or mitigate the identified topological obstruction, while retaining the computational advantages of low-rank factorization?

I would like to point out that I am willing to raise the score to weak accept if the presentation is significantly improved, and possibly further pending on the good answer to questions 3 and 4 above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer C4jA

Thank you for the constructive review and for indicating what could change your assessment. We reorganized the Introduction around the learning task, sign-reversal example, mechanism, and rates. Section 3 now defines labels, loss, Error, and paired heads before the results; technical and training details are in the appendices.

Q1. To what extent do the rates depend on the Haar measure assumption? Can analogous results be established for distributions supported on lower-dimensional subsets of the Grassmannian?
A1. The lower rates require both obstructed geometry and sufficient local mass; they are not universal over all laws. If dμ~=f dμHaard\widetilde{\mu}=f\,d\mu_{\mathrm{Haar}} with 0<c≤f≤C0<c\le f\le C, the exponents are unchanged, with constants depending on c,Cc,C. For lower-dimensional families, an embedded RPd−1\mathbb{RP}^{d-1} has rank-one exponent r=mr=m for m<dm<d, while a sign-reversing loop has r=1r=1 at width one and exact width two. Conversely, Appendix C gives a full-support density for which the Haar exponent fails, and a contractible domain where width kk is exact. Thus support dimension alone is insufficient; Appendix C.1 states the positive extensions and counterexample.

Q2. What are the main new mathematical ingredients needed to derive the sharp rates....
A2. The contribution is the sharp subthreshold approximation and robustness law, beyond the qualitative obstruction. Four ingredients give the matching rates: (i) averaging unavoidable unit-error points over rotated projective rr-slices yields the sharp L−rL^{-r} mass; (ii) clipped polar factors and small-singular-value estimates attain the upper order; (iii) transporting mass trt^r over distance tt yields exponent pr/(p+r)pr/(p+r), while a capped coupling gives the MM-dependent crossover; and (iv) an equivariant partition and concentration estimates expose the N\sqrt{N} scale. The revised related work and Appendix A separate these results from prior obstruction and exact-representation results.

Q3. Can the authors characterize the dependence of the constants in their upper and lower bounds on the ambient dimension N? In particular, do the rates remain informative as N grows?
A3. Theorem 14 now characterizes this dependence. For full Haar sampling on Gr⁡(k,N)\operatorname{Gr}(k,N), fixed k,m,pk,m,p, r=m−k+1r=m-k+1, and finite pp, set a=(N/L)r,b=(Nε)pr/(p+r).a=(\sqrt{N}/L)^r,\qquad b=(\sqrt{N}\varepsilon)^{pr/(p+r)}. For every N>mN>m, the optimal reference error is of order min⁡(1,a)\min(1,a), with comparison constants depending only on k,mk,m. The optimal density-capped shifted error is of order min⁡(1,Ma,a+b)\min(1,Ma,a+b), with comparison constants depending only on k,m,pk,m,p. For M=∞M=\infty, the term MaMa is omitted. The comparison constants are independent of NN: they depend only on k,mk,m for reference error and on k,m,pk,m,p for shifted error. Therefore the reference error vanishes as N→∞N\to\infty exactly when LN/N→∞L_N/\sqrt{N}\to\infty. The unrestricted shifted error vanishes exactly when this holds and NεN→0\sqrt{N}\varepsilon_N\to0. Appendix C.2--C.3 proves the matching bounds and coefficient scale Nr/2N^{r/2}. Proposition 13 separately proves that a fixed intrinsic family embedded isometrically into increasing ambient spaces has unchanged norm-bounded optimal error. Thus the N\sqrt{N} growth is specific to full-Haar sampling, not merely to extra coordinates.

Q4. Can the theoretical insights be used to design practical architectures ...
A4. The theory suggests changing factor-output width rather than indiscriminately increasing hidden capacity. For a verified intrinsic family, use the smallest width that removes its obstruction and allow dependent columns, so WW⊤WW^\top remains low rank. For a sign-reversing rank-one loop, two columns give W(ξ)=u(ξ)[cos⁡(πξ/T),sin⁡(πξ/T)]W(\xi)=u(\xi)[\cos(\pi\xi/T),\sin(\pi\xi/T)] and WW⊤=PWW^\top=P. This motivates a two-column pointwise head on a frozen neural-operator backbone. It retains O(Nm)O(Nm) factor storage and application cost and needs no test-time eigensolve. Corollary 3 formalizes this on Y×RPd−1Y\times\mathbb{RP}^{d-1}: every m<dm<d continuous norm-bounded head obeys the predicted lower rates independently of ambient NN and backbone size, while width dd is exact.

Experiment 5 tests the intervention. Across five paired plate seeds with identical frozen features, changing only the head from one to two columns adds 25 parameters and reduces median shifted Error from 0.1520.152 to 0.008540.00854. Increasing hidden capacity at width one leaves shifted Error at 0.1450.145--0.1470.147. Prediction-plus-application time is 3.543/3.3823.543/3.382 ms per batch of 32, with equal measured peak memory; factor storage changes from 0.56250.5625 to 1.1251.125 KiB per sample. The Darcy control also finds wide initialization best in all ten matched comparisons, with small nonzero expansion providing a recovery path. These experiments support the design prescription; the theorems provide the class-level guarantee.

AuthorsOpenReview ↗

Summary of the revised manuscript

We thank the reviewers for their detailed comments. We have uploaded a revised manuscript that preserves the original research question and the main approximation and Wasserstein-shift rates, while substantially clarifying and expanding their assumptions, scope, proofs, and practical implications.

The revised PDF has 42 pages rather than 35. The main technical text remains within 9 pages. The additional pages occur after the main text and primarily contain expanded proofs, ambient-dimension analysis, training protocols, and reproducibility documentation.

The principal revisions are:

  1. Problem and presentation. Sections 1--3 now define the parameter-to-projector learning problem, the factor predictor, the training labels and loss, the distinction between projector continuity and factor continuity, and the exact meaning of paired FNO output-head comparisons before presenting the results.

  2. Theoretical scope. Theorems 1--2 now state the quantifiers and admissible factor class explicitly. We distinguish lower bounds that hold for every admissible factor under the stated assumptions, constructive upper bounds attained by explicit factors, and architecture-specific claims about trained neural networks.

  3. Distributional assumptions. Appendix C gives extensions to comparable densities and specified lower-dimensional projective families. It also gives a counterexample showing that matching lower rates cannot hold for arbitrary input laws.

  4. Ambient-dimension dependence. Theorem 14 and Appendix C.2--C.3 give dimension-uniform bounds whose comparison constants are independent of the ambient dimension when the ranks are fixed. They identify the relevant ambient-dimension scales and state precisely when the optimal errors vanish as the ambient dimension grows.

  5. Practical architecture design. Corollary 3 connects the theory to frozen neural-operator output heads. The revised experiments compare output width with hidden capacity and report paired-seed controls, initialization and expansion controls, prediction time, memory, and factor-storage costs.

  6. Reproducibility. Appendix F specifies the inputs, reference solvers, architectures, losses, normalizations, seeds, training schedules, checkpoint selection, and evaluation distributions. We also provide an anonymous reproducibility archive containing code, configurations, per-seed results, diagnostic scripts, retained checkpoints, and verification instructions.

Our individual responses identify the theorem, section, figure, table, and appendix addressing each reviewer question. Copyediting, notation alignment, and proof reorganization are visible in the automatic PDF diff but are not listed individually here.

AuthorsOpenReview ↗

Response to reviewer C4jA

Thank you again for the constructive review. Since posting our original response, we have uploaded a further revised manuscript focused specifically on presentation and readability.

The new version substantially reorganizes the exposition while preserving the research question and main rate statements. The Introduction now starts from the parameter-to-projector learning task and its computational motivation, followed by a concrete sign-reversal example, the geometric obstruction, the sharp rates, and their architecture implications. Figures 1 and 2 provide visual overviews of the learning problem and the rate mechanism.

The main argument is now organized as follows: Section 2 defines the admissible factor class, reference risk, Wasserstein distance, shifted risk, and theorem assumptions before stating the canonical results; Section 3 distinguishes contractible parameter domains from obstructed mode loops and frozen neural-operator output heads; and Section 4.1 defines the labels, training loss, shifted evaluation, and paired-FNO-head protocol before presenting the experiments. Detailed proofs, dimension-dependence arguments, and training protocols are moved to Appendices A--G, leaving the main technical narrative within nine pages.

We have also aligned the results emphasized in Questions 3 and 4 with this new organization. The ambient-dimension result is now Theorem C.2, with its matching proof in Appendices C.2--C.3 and the fixed-intrinsic-family result in Proposition C.1. The practical frozen-head implication is now Corollary 3.5 and is evaluated in Sections 4.1--4.3, particularly Experiment 5.

We hope this further revision directly addresses the presentation concern as well as the requested ambient-dimension and practical-architecture questions.

I thank the authors for their careful response, which I have read in detail. I found the new dimension-dependent analysis (Theorem 14) a genuinely interesting addition. However, the main proof ingredients seem standard, and the implications for learning (estimation error, sample complexity) are left unexplored. I therefore keep my score unchanged.

Reviewer_V3UdOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper estimates a subspace from data using tools from random matrix theory. The method is applied to improve some operator learning architectures. The theory is related to the Wasserstein distance on the Stiefel manifold.

Critical strengths

The authors run numerical experiments on a wide variety of different problems.

Critical weaknesses

The paper is very difficult to follow. In particular, the problem is not clearly specified. It is not at all clear how the factor network is trained. What data and loss functions should be used here? Many paragraphs read like a sequence of apparently unrelated statements.

Questions for authors

Can you please explain the connection with operator learning? In particular, what does paired FNO head interventions mean at the beginning of the experimental section? This was apparently never defined anywhere. It only seems to be discussed in Experiment 2, where it is unclear to me what the setup is.

What does the paragraph starting with `An eigengap makes P locally stable' mean? You are trying to learn an approximate projector onto a subspace, right? What does local stability mean? What does continuous factor mean? What data are you learning this projector from?

Equations (1) and (2) cannot possibly hold for every factor, right? Under which assumptions do these bounds hold? How are the approximate factors estimated or computed?

In the experiments, how are the factors computed? Are they learned, and if so what loss function is optimized?

In Section 2.1, is μ\mu the Haar measure and ν\nu a new measure?

3 comments
AuthorsOpenReview ↗

Response to Reviewer V3Ud

Thank you for identifying these presentation issues. We agree that the original version did not define the learning task and training protocol early enough. We have clarified these issues, reorganized Sections 1--3, and added complete specifications in Appendix F.

Problem. This is not eigenspace estimation from a noisy covariance matrix. Given an equation parameter or coefficient field ξ\xi, an offline solver provides the projector P(ξ)P(\xi) onto a selected gapped eigenspace. The neural operator learns Wθ(ξ)∈RN×mW_\theta(\xi)\in\mathbb{R}^{N\times m} such that Wθ(ξ)Wθ(ξ)⊤≈P(ξ)W_\theta(\xi)W_\theta(\xi)^\top\approx P(\xi). Our theoretical question is how the output width mm limits approximation by continuous, norm-bounded, Lipschitz factors.

Q1. Can you please explain the connection with operator learning? In particular, what does paired FNO head interventions mean at the beginning of the experimental section? This was apparently never defined anywhere. It only seems to be discussed in Experiment 2, where it is unclear to me what the setup is.
A1. The input is an equation parameter or coefficient field, and the output is a low-rank factor of the associated projector. At inference the model applies Wθ(ξ)(Wθ(ξ)⊤v)W_\theta(\xi)(W_\theta(\xi)^\top v) without eigensolving.

For a paired comparison, we pretrain one wide FNO, freeze its input projection and all Fourier layers, and train two pointwise output layers with different mm using identical features, targets, minibatches, and schedules. Equality of the frozen parameters is checked. Thus only the factor-output design changes. This is now defined in Section 3, Eq. (8), and Appendix F.4.

Q2. What does the paragraph starting with `An eigengap makes P locally stable' mean? You are trying to learn an approximate projector onto a subspace, right? What does local stability mean? What does continuous factor mean? What data are you learning this projector from?
A2. For eigengap g>0g>0, spectral perturbation theory gives ∥P(A+E)−P(A)∥op≤C∥E∥op/g\lVert P(A+E)-P(A)\rVert_{\mathrm{op}}\le C\lVert E\rVert_{\mathrm{op}}/g. Thus nearby operators have nearby projectors. This local stability does not guarantee a globally continuous eigenvector or narrow factor. A continuous factor means that W(ξ)W(\xi) depends continuously on ξ\xi.

Offline PDE/eigensolvers provide Pi=UiUi⊤P_i=U_iU_i^\top, where UiU_i contains the selected modes. The network receives only ξi\xi_i at inference. Appendix F.1 lists the input and label for every experiment.

Q3. Equations (1) and (2) cannot possibly hold for every factor, right? Under which assumptions do these bounds hold? How are the approximate factors estimated or computed?
A3. The reviewer is correct that the two-sided rate is not an equality for every factor. The lower bound holds for every admissible factor, whereas the two-sided rate describes the best error over the admissible class.

An admissible factor has Frobenius Lipschitz constant at most LL and Frobenius norm at most the square root of kk at every input. The canonical result uses Haar probability on the Grassmannian, the principal-angle metric, widths between kk and NN, sufficiently large LL, and a sufficiently small shift radius.

For the constructive upper bound, we choose an orthonormal m-frame, multiply it by the target projector, rescale it using a threshold proportional to the reciprocal of LL, and clip its singular values at one. Neural upper rates additionally require the architecture to approximate this construction. We do not claim that every trained network attains the upper rate. These quantifiers and assumptions are now explicit in Theorems 1--2 and Appendix A.

Q4. In the experiments, how are the factors computed? Are they learned, and if so what loss function is optimized?
A4. Experiment 1 uses explicit analytic factors. Learned models minimize

\lVert W_\theta(\xi_i)W_\theta(\xi_i)^\top-P_i\rVert_F^2 .$$ This loss is invariant to eigenvector sign and basis. Evaluation instead uses $\lVert W_\theta W_\theta^\top-P\rVert_{\mathrm{op}}^2$. Validation selects checkpoints before held-out and shifted evaluation, and attack samples never select checkpoints. Section 3 and Appendix F.2--F.4 give the optimizer, seeds, schedules, normalization, and selection rules. **Q5. Meaning of $\mu$ and $\nu$.** **A5.** $\mu$ is the reference input law, which is Haar in the canonical theorem. After fixing $W$, $\nu$ is a possibly different probability law on the same space satisfying $\mathsf{W}_p(\mu,\nu)\le\varepsilon$. It need not be Haar. The optional density cap additionally requires $\nu\ll\mu$ and $d\nu/d\mu\le M$. Targets are recomputed at transported inputs.

I thank the authors for their responses, which clear up some of the questions that I had. However, my main objections about the unreadability of the paper remain and so I stick with my score.

AuthorsOpenReview ↗

Response to reviewer V3Ud

Thank you for the follow-up. We understand that our earlier response clarified several technical questions but did not sufficiently address your main concern about readability.

Following this comment, we prepared and uploaded a further revised manuscript that substantially restructures the exposition rather than merely adding local explanations. The research question and main rate statements are unchanged, but the motivation, definitions, assumptions, and progression of the argument are now presented in a more explicit order.

In particular:

  1. The Introduction now begins with the parameter-to-projector learning problem, explains what the network receives and predicts, and motivates the low-rank factorization through its storage and application costs.

  2. A concrete sign-reversing mode example and Figure 1 introduce the obstruction before the general theory. Figure 2 then explains geometrically how the Lipschitz budget, Wasserstein radius, and density cap produce the stated rates.

  3. Section 2 now defines the factor class, reference risk, Wasserstein distance, shifted risk, and all theorem assumptions before stating the main results. It also distinguishes universal lower bounds from constructive upper bounds.

  4. Section 3 separately explains when no obstruction exists, how the theory transfers to parametric PDEs and mode loops, and what it implies for frozen neural-operator output heads.

  5. Section 4.1 defines the training labels, projector loss, shifted evaluation, and paired-FNO-head protocol before presenting the experiments. The six experiments are then organized by the theoretical question that each tests.

  6. Lengthy proofs, ambient-dimension arguments, implementation details, and reproduction protocols have been moved to Appendices A--G. The main technical narrative remains within nine pages.

We also standardized the notation and updated all theorem, figure, and appendix references. We hope that the newly uploaded version addresses the presentation issue more directly and would be grateful if the reviewer could assess the latest PDF rather than the earlier revision.

Reviewer_ap7wOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper asks how the number of output-factor columns limits spectral projector approximation, under nominal sampling and under Wasserstein shift. The obstacle is that a continuously varying projector may admit no narrow continuous factor across the whole input domain. Under norm and Lipschitz constraints, the paper derives matching upper and lower bounds tying together output width, smoothness, transport radius, and density restrictions. The proof combines a topological obstruction with explicit factor constructions and transport arguments. Analytic calculations and neural-operator experiments test what the bounds predict.

Critical strengths
  • Turning a qualitative obstruction to continuous factorization into quantitative approximation and robustness rates is a solid contribution, and treating output width together with distribution shift keeps the scope tight.
  • The analysis cleanly separates output-factor width from hidden capacity, and separates bounds for continuous function classes from architecture-specific claims about neural networks. The reach of the results is easy to locate.
  • The experiments are broad for a theory paper: analytic checks, density restrictions, several target ranks, and physical operator examples. Frozen-backbone comparisons and hidden-capacity controls back up the claim that output width and model capacity do different things.
  • The evaluation protocol is careful. Model selection runs on independent validation, attack evaluation is kept separate from checkpoint selection, runs are paired, and transport budgets, solver residuals, and loss computation are all checked. Recomputing targets at transported inputs is also the right call for a distribution-shift setting.
Critical weaknesses

Configurations, per-seed results, and diagnostic scripts sit in a separate source package that was not in the materials I could see, so I could not check the execution records or the numerical claims. I need access before I can judge reproducibility.

Questions for authors

Could the authors provide more details about the reproducibility of this paper?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ap7w

Q1. Could the authors provide more details about the reproducibility of this paper?

A1. Thank you for identifying this access issue. We agree that the materials previously visible to the reviewer were insufficient for checking the execution records and numerical claims. We have now attached the anonymized archive WassersteinShift_experiment_reproducibility_final_20261118.zip as supplementary material.

The archive supports three distinct levels of verification:

  1. Saved-result verification. From the archive root, running python validate_compact.py checks the file manifest and executable entry points, reconstructs Tables 1 and 8 from the per-seed records, checks the reference PDE solvers, verifies the retained checkpoint inventory, and reloads the TU1402 models for the reported error calculation.

  2. Checkpoint replay. The archive contains 35 retained checkpoints: 25 plate FNO models and 10 TU1402 models, together with the corresponding configurations, evaluation inputs, and replay instructions.

  3. Fresh reproduction. experiments_and_results/REPRODUCING.md maps Experiments 1--6 to their data-generation, training, shift-evaluation, and plotting commands. It also records the software requirements, hardware requirements, fixed seeds, training schedules, checkpoint-selection rules, and evaluation metrics. RESULTS_INDEX.md maps the paper's tables and figures to the supplied CSV/JSON records, while EXPERIMENT_MAP.md distinguishes the primary experiments from diagnostic controls.

The archive additionally includes source code, dependency files, configurations, per-seed outputs, generated numerical inputs, plotting utilities, solver checks, and SHA-256 file hashes. To satisfy the supplementary-file size limit, the large Darcy retest checkpoints and intermediate arrays are not included; their executable code, exact configurations, hashes, per-seed records, and deterministic regeneration procedures are included. For studies whose original weights were not retained, the documentation explicitly marks them as fresh-retraining experiments, and we do not claim exact checkpoint-level replay.

We have also expanded the Reproducibility Statement in the revised manuscript and added the complete data, architecture, loss, sampling, seed, training, and evaluation specifications in Appendix F (“Training data, losses, and reproducibility”).

We hope these materials now provide the access needed to assess the numerical claims. We would be grateful if the reviewer could let us know if any specific record or execution path remains inaccessible.

AuthorsOpenReview ↗

Summary of the revised manuscript

We thank the reviewers for their detailed comments. We have uploaded a substantially revised manuscript that preserves the original research question and main approximation and Wasserstein-shift rates, while making their assumptions, scope, proofs, and practical implications explicit.

The revised PDF has 42 pages rather than 35. The main technical text remains within 9 pages. The additional pages occur after the main text and primarily contain expanded proofs, dimension-dependence analysis, training protocols, and reproducibility documentation.

The principal revisions are:

  1. Problem and presentation. Sections 1--3 now define the parameter-to-projector learning problem, the factor predictor, the training labels and loss, the distinction between projector continuity and factor continuity, and the exact meaning of paired FNO output-head comparisons before presenting the main results.

  2. Theoretical scope. Theorems 1--2 now state the quantifiers and admissible factor class explicitly. We distinguish the universal lower bounds from constructive upper bounds and from architecture-specific neural approximation.

  3. Distributional assumptions. Appendix C gives extensions to comparable densities and specified lower-dimensional projective families, together with a counterexample showing that matching lower rates cannot hold for arbitrary input laws.

  4. Ambient-dimension dependence. Theorem 14 and Appendix C.2--C.3 give dimension-uniform bounds with constants independent of the ambient dimension for fixed ranks. They identify the relevant scales N/L\sqrt{N}/L and Nε\sqrt{N}\varepsilon and state exactly when the optimal errors vanish as NN grows.

  5. Practical architecture design. Corollary 3 connects the theory to frozen neural-operator output heads. The revised experiments compare output width with hidden capacity, report paired-seed controls, initialization and expansion controls, and prediction time, memory, and factor-storage costs.

  6. Reproducibility. Appendix F now specifies the inputs, reference solvers, architectures, losses, normalizations, seeds, schedules, checkpoint selection, and evaluation distributions. We also provide an anonymous reproducibility archive containing code, configurations, per-seed results, diagnostic scripts, retained checkpoints, and verification instructions.

Our individual responses identify the precise theorem, section, figure, table, and appendix addressing each reviewer question. Copyediting, notation alignment, and proof reorganization are visible in the automatic PDF diff but are not listed individually here.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.