acceptodds
Under review as a conference paper at ICLR 2027

Coordinate-Specific Factor Design in Low-Rank Physics-Informed Solvers

Abstract

Low-rank physics-informed solvers use products of one-dimensional factors, but coordinates can have distinct physical structures. Suitable factors can improve accuracy, yet coupled changes in family, capacity, and assignment complicate attribution. We introduce DimSpec-LR, a coordinate-specific framework with a shared CP interaction and a unified interface for neural, dictionary-based, and hybrid factors. Parameter-matched interventions examine family selection, capacity allocation, and coordinate assignment under controlled training configurations. Experiments across four PDE families demonstrate that physically aligned factors can substantially improve accuracy under a fixed training budget. On a boundary-layer problem, the prescribed profile reduces relative error from to compared with a generic homogeneous control. However, a richer homogeneous alternative achieves an even lower error of , revealing that coordinate suitability does not necessarily require heterogeneous factor families. Additional experiments show that allocating spectral capacity to cover the required temporal frequencies improves accuracy and that target-frequency coverage alone does not determine trained performance. A separate source-specified CDR problem shows that fixed local factors remain effective without constructing an analytic target solution. Initialization and preconditioning controls clarify their contributions to complete-solver accuracy. These findings establish a controlled evaluation methodology for physically informed factor design and highlight the importance of distinguishing coordinate suitability, capacity allocation, and architectural heterogeneity.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Informed Neural Networks, Low-Rank Decomposition, Coordinate-Specific Factor Design, Scientific Machine Learning, Partial Differential Equations

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_vmxWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DimSpec-LR, a low-rank physics-informed solver that represents the solution as a CP sum of R=16R = 16 products of one-dimensional factors. Each coordinate's factor can come from a different family: fixed dictionaries with trainable coefficients, neural networks, or dictionaries with neural corrections. Profiles are designed by hand for each benchmark using the operator, coefficients, forcing, and boundary data. Parameter-matched interventions suggest that suitable factors improve accuracy, that suitability does not require heterogeneous families, and that target-mode coverage alone does not determine trained accuracy. Comparisons with five baselines on 16 cases show large gains on most of them.

Critical strengths

The experiments are carefully controlled. Checkpoints are selected by physics validation rather than reference error, matched arms share sampling streams and budgets, and inconvenient results are reported openly. Some findings seem useful, in particular that dictionaries with the same mode coverage and parameter count train to very different accuracies (H4), and that initialization and preconditioning account for much of the accuracy in some cases.

Critical weaknesses
  1. As far as I can tell, much of the gain comes from instance-specific priors, and the results are conditional on them. In the manufactured problems the forcing appears to reveal the target's structure. In A4 the dictionaries contain the exact rank-one target, S16 covers all H4 modes, and H24 contains K3's 20-radian carrier. The choice of profile does not seem to be evaluated: the interventions compare the authors' choice with a few alternatives they also built, and selection costs are excluded. The one test with profiles frozen in advance (Section E.1) looks structurally close to A4. It may be telling that DimSpec-LR loses on cases where the structure is harder to know in advance (H1, A1, S1, S3, S4).

  2. Beyond the dictionary coefficients, little about the representation seems to be learned, and it is not clear that anything transfers across problems. To me the method reads as closer to a spectral method with hand-chosen bases than to the reusable inductive biases typically published at ICLR, though the authors may see this differently. The finding that a rich shared dictionary on all axes matches or beats tailored profiles seems to point toward a less expert-dependent approach, and I wondered whether it weakens the motivation for coordinate-specific design.

  3. Eq. (5) appears to be a perturbation bound that assumes a rank-RR expansion with error δR\delta_R exists. As I read it, it does not say which functions the ansatz approximates well, and the products of factor norms may not stay bounded, since best CP approximations can fail to exist for order ≥3\geq 3 (de Silva & Lim, 2008). I could not find references to related literature that seems relevant here: separated representations (Beylkin & Mohlenkamp), tensor numerical analysis (Hackbusch), low-rank approximability of PDE solutions (Grasedyck; Dahmen, DeVore, Grasedyck & Süli; Bachmayr, Schneider & Uschmajew), approximation classes of tensor networks (Ali & Nouy; Bachmayr, Nouy & Schneider), and classical spectral methods. This literature might also help explain the H1/A1 results, since sharp features along curved sets tend to have high separation rank.

  4. A direct least-squares solve over products of the same dictionary atoms seems like a natural baseline for these linear problems. The paper already uses such a solve to initialize A3, and it appears to account for most of A3's accuracy (4×10−34 \times 10^{-3} vs. 0.240.24 with random initialization).

  5. Apart from the CP interaction, the pipeline seems to change from case to case: factor types, constraint treatment, loss preconditioning, initialization, learning rates, and schedules. Removing projected initialization on A3 or spectral weighting on H4 increases error by one to several orders of magnitude, while each baseline uses one fixed configuration. My reading is that Tables 6 and 7 therefore mix representation, prior knowledge, and per-case tuning, which makes them hard to interpret as evidence for the framework itself.

  6. The baselines do not appear to be on a common footing. Parameter counts differ widely (512 vs. 727,045 on H4), and the fixed 3,000-update budget may favor fast-converging coefficient models. With longer budgets, CP-PINN overtakes DimSpec-LR on K3 and S3, and PirateNet on H4 improves from 0.0300.030 to 1.25×10−31.25 \times 10^{-3}. Wall times do not show a consistent advantage (224 s vs. 3.3 s for RFM on H1), though they were measured on shared hardware and may not be reliable. I recognize a fully fair comparison may not be possible, but accuracy-versus-time curves with equal tuning budgets, run once with the same priors given to all methods and once with generic dictionaries, might be more informative.

  7. I could not find how the comparator in Tables 6 and 14 is chosen. For S2, Table 6 uses PirateNet although Vanilla PINN has lower seed-42 error (7.95×10−37.95 \times 10^{-3} vs. 0.0100.010), and S2 is the case with the smallest margin. Table 14 similarly uses MS Fourier on N2 although Vanilla PINN appears better. There may be a reason I missed, but it would help to state the rule.

  8. Matched arms share one learning rate. The S3 study runs at 5×10−45 \times 10^{-4}, where the prescribed profile reaches 0.0450.045 compared with 0.0110.011 at its main rate, so the small differences between arms may change under per-arm tuning. The permutation tests seem somewhat uninformative by construction, and three seeds may be too few for the finer comparisons.

Questions for authors
  1. How would a user choose a profile when the solution structure is unknown? How does a generic shared dictionary perform across all 16 cases?
  2. How does the method compare with a direct least-squares solve on the same dictionaries, in accuracy and wall time?
  3. How were the comparators chosen, and why do they differ from the best baselines for S2 and N2?
  4. How do the baselines perform when given the same prior information?
Reviewer_71e6OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents DimSpec-LR, a low rank physics informed that allows different coordinate directions to use neural, dictionary-based, or hybrid factors. Evaluation is performed across 4 PDE families and it is shown that physically informed factor choices can improve accuracy.

Critical strengths

Multiple Ablation studies: I appreciate the parameter-matched experiments separating factor choice, coordinate assignment, and capacity allocation. The finding that homogeneous factors can outperform heterogeneous ones seems interesting.

Broad PDE evaluation: The paper considers four PDE families, including boundary-layer and oscillatory problems. The additional non-manufactured CDR experiment is useful.

Critical weaknesses
  1. The paper was hard to follow for me, and the writing appeared to be quite LLM-y. I guess the author's have documented that they have used AI assisted writing extensively. Nevertheless, the paper was hard for me to understand.
Questions for authors

Could you please explain the main findings and contributions of this paper in very simple language?

Reviewer_E7ZBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Introduces DimSpec-LR, which assigns neural, dictionary, or hybrid coordinate factors within a shared CP low-rank physics-informed formulation. Uses parameter-matched experiments to separate factor suitability, capacity allocation, and coordinate assignment across PDE benchmarks. Finds major gains from physically aligned factors, but also finds that rich homogeneous factors can outperform heterogeneous profiles.

Critical strengths
  1. The matched-factor and permuted-coordinate controls provide useful evidence about what actually causes improvements, rather than crediting heterogeneity by default (Table 3).
  2. The paper openly reports stronger homogeneous controls and distinguishes representability from finite-budget trainability (Tables 3–4).
  3. The source-specified CDR test without a manufactured target is a valuable check beyond perfectly tailored analytical solutions (Table 5).
Critical weaknesses
  1. Factor families and capacities are expert-designed and benchmark-aware; the method offers no automatic or transferable profile-selection procedure.
  2. Complete-solver gains cannot be attributed solely to factor design because initialization, constraints, preconditioning, and tuning also differ (Section 4.4).
  3. Generalization is limited: the 16-case coverage audit is largely single-seed, and some transport, curved-field, and interface cases underperform baselines (Table 7).
Questions for authors
  1. Can the authors provide a fully held-out PDE/profile-selection test with no benchmark-specific expert tuning?
  2. How much of the reported complete-solver improvement remains after matching constraints, preconditioning, and configuration-search cost?
  3. Could the authors repeat the most consequential baseline-loss and baseline-win cases across multiple seeds and equal wall-clock budgets?
Reviewer_mCSSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how factor families, capacity allocation, and coordinate assignment affect low-rank physics-informed solvers. DimSpec-LR provides a shared CP formulation supporting neural, dictionary-based, and hybrid factors. Controlled experiments examine coordinate suitability, homogeneous alternatives, frequency coverage, and finite-budget training behavior. Validation includes 16 cases across four PDE families and a separate source-specified CDR problem with frozen configurations.

Critical strengths
  1. The controlled comparisons address a useful design question. In A4, swapping B64/H32 to H32/B64 increases error from 0.00159 to 0.8519, while the approximately parameter-matched B44/B44 achieves 0.000404. These results distinguish coordinate suitability from heterogeneity. The K3 coverage-preserving control shows that the original capacity ratio is unnecessary for obtaining the observed benefit, while H4 demonstrates that target-frequency coverage alone does not ensure good trained accuracy.

  2. The additional checks strengthen the empirical assessment. The separate CDR experiment freezes configurations before reference generation and fitting. B64/H32 achieves a global error of 0.00636, compared with 0.28627 for H52/H52. The paper also reports cases where baselines perform better and longer-budget results where CP-PINN overtakes the proposed short-budget solver. These results help establish the scope of the reported advantages.

Critical weaknesses
  1. The complete-solver comparisons leave important attribution and implementation questions unresolved. For the complete-solver H4 configuration, the spectral loss component uses the exact inverse eigenvalues of the PDE operator. In the stated sine basis, this component measures coefficient error against the directly computable spectral solution. The rank-16 parameterization uses 512 coefficients to cover the full 256-dimensional coefficient space. A direct same-basis spectral comparison would clarify the contribution of factorization and iterative training. Some baseline procedures also need clarification. Equation (40) contains (u2)(u^2), whereas Table 12 describes RFM as a linear physics fit. The residual remains nonlinear in the output coefficients, and the solution procedure is unspecified. The PirateNet description also leaves its initialization and training configuration unclear, although initialization is central to the original method.

  2. The reusable design guidance and robustness evidence remain limited. The framework builds closely on established separable PINNs and functional tensor decompositions. Tables 1 and 10 provide useful physical cues, but bandwidths, scale banks, and capacities still involve expert choices whose selection costs are unclear. Appendix D.1 states that the three A4 runs share initialization seed 20260728 and vary only the sampling stream. Independent initializations would help establish how consistently the relative ordering holds. Table 19 also shows substantial sensitivity on the thinnest layer: error rises from 0.00137 with the exact scale to 0.2729 with the fixed bank. Although scale effects are nonmonotonic, this single-seed study leaves uncertainty about reliability when structural information is less precise.

Presentation suggestion: The main text relies heavily on tables. Including a representative solution and error plot from the appendix would make the factor-design results easier to interpret.

Questions for authors
  1. How does RFM handle the u2u^2 term in Equation (40)? Please specify the nonlinear solver or linearization, stopping criterion, and whether the full nonlinear residual is used during fitting.

  2. For H4, how does the proposed solver compare with directly computing the solution coefficients in the same S16×S16 basis? Error, runtime, and memory measurements would clarify its practical contribution.

  3. Do the A4 comparisons, particularly B44/B44 outperforming B64/H32, persist across independent coefficient initializations?

  4. Which PirateNet initialization and training components were retained? Please clarify adaptive residual initialization, output-layer initialization, and weight parameterization.

  5. Which factor-selection rules were reused across cases, and which required case-specific adjustment? How many factor configurations were considered, and what was their selection cost, separately from learning-rate searches?

Comments on the paper

1 comment

Dear Reviewers,

Please take a moment to review the authors' responses and engage in the discussion when you have a chance, especially if the authors have addressed major concerns raised in your reviews. Thank you for your time and continued efforts!

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.