acceptodds
Under review as a conference paper at ICLR 2027

Auditing Coefficient Identifiability in Physics-Constrained Dictionary Models

Abstract

Low data-fitting error and a small physics residual do not ensure that the coefficients of a dictionary-based PDE solution are identifiable. For a fixed linear operator and sampling design, we develop a pre-training structural audit based on the classical joint null space of the data and residual maps. Without observed response values, the audit identifies coefficient perturbations that leave both predictions and residuals unchanged, returns explicit witnesses, and reports which directions the physics penalty constrains. Tolerance scans, higher-precision calculations, and symbolic checks distinguish functional identities from sampling-induced dependencies and numerical near-dependencies. Across nineteen dictionaries, the audit detects multi-term dependencies missed by pairwise and triple screens and identifies four dictionaries with identically zero residual penalties. In a public PySINDy constrained-regression example, steady-state observations leave seven coefficient directions unresolved by the data and constraint together. On these data, the published optimizer fits the sampled derivatives with a relative error of while omitting the growth and coupling terms. Quadrupling the steady-state samples leaves all seven directions unresolved. At the original sample budget, transient observations remove these blind directions. The audit identifies which individual coefficients and coefficient combinations are determined under the specified observations and physics constraints. It does not validate the imposed operator or guarantee improved reconstruction accuracy.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Coefficient identifiability, Physics-informed learning, Structural auditing, Dictionary-based models, Null-space analysis

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

The paper proposes a pre-training structural audit of coefficient identifiability for physics-constrained dictionary models with a fixed linear operator (or linear coefficient constraint) and a fixed sampling design. Building on the classical joint null space V0=ker⁡Ad∩ker⁡ApV_0=\ker A_{\mathrm d}\cap\ker A_{\mathrm p} of the data and physics maps, the audit reports:

  • explicit witness directions,
  • the number of data-blind directions closed by the physics term (Δr\Delta r),
  • a decomposition of coefficient space into direction classes,
  • a test of whether a linear reading ℓ⊤α\ell^\top\alpha is determined.

Tolerance scans, high-precision recomputation and symbolic checks separate exact identities from numerical near-dependencies. On a corpus of dictionaries, the audit finds dependencies missed by pairwise and triple screens and identifies dictionaries whose residual penalty is identically zero. In a pre-registered PySINDy mean-field case, steady-state data leave jointly blind directions along which gg and AA vary individually, while g+mAg+mA stays determined. Equal-budget transient data remove these directions, and an exploratory refit on the transient data, stopped at the iteration limit, recovers all true terms. Further rule-selected studies on official PySINDy and DeePyMoD examples are reported in the appendix.

Critical strengths
  1. Coefficient-level reporting, demonstrated on a public pipeline. Beyond flagging rank deficiency, the audit states which coefficients or combinations the data determine and which ambiguities the physics constraint resolves, with explicit witnesses. In the mean-field case, the audit predicts before training that AA is undetermined while g+mAg+mA is determined on steady-state data (Table 2). The exploratory steady-state refit indeed misestimates AA badly while recovering g+mAg+mA accurately (Table 3). This is consequential because it shows that fit quality cannot certify coefficient readings, and that the audit tells practitioners which readings they may report.
  2. Careful numerical verification. Tolerance sweeps, high-precision spectra and symbolic checks separate exact identities from numerical near-dependencies, and the paper reports that only part of the numerically near-zero directions are exact (Table 6). The cut is fixed in advance, and null outcomes of the pre-registered studies are reported. This is consequential because it makes the reported counts trustworthy and the claims well calibrated.
Critical weaknesses
  1. The value added over a direct SVD/GSVD diagnostic is not isolated. The audit's near-null dimensions match a plain stacked SVD on every library in the corpus, and on the libraries with a nonzero near-null space the subspaces themselves coincide (Table 5; App. B.1.3). The direction classes of Table 1 are the block structure of the GSVD of (Ad,λrefAp)(A_{\mathrm d},\sqrt{\lambda_{\mathrm{ref}}}A_{\mathrm p}), with νj=sj2\nu_j=s_j^2 and V0V_0 as the zero block. The reported advantages are largely by construction:

    • coherence and triple enumeration are support-limited (Prop. 3);
    • the zero-penalty dictionaries can be identified atom by atom.

    The comparison with condition numbers (Sec. 4.1; App. B.2.1, B.2.3) is qualitative. This is consequential because, apart from the exact-versus-near precision protocol credited above, the audit's verdicts coincide in principle with those of a GSVD-based analysis with coefficient projections at the same cut and normalization. The contribution therefore rests on practical value over such an analysis, and over the diagnostic of Scheipl & Greven (2016), which the paper does not demonstrate.

  2. The presentation makes the report hard to use and verify. The paper introduces a number of paper-specific terms (gauge, physics-visible, reading, equation-implied completion) and several overlapping taxonomies (exact classes, numerical bins, report tiers). Main-text claims rest on statements given only in the appendix (Algorithm 1, Prop. 4, Thm. 1). Sec. 4.3 compresses the detector comparison into a few sentences whose quantities are defined only in a table caption. This is consequential because the paper's main selling point is a practitioner-legible report.

Questions for authors

Both are clarification requests.

  1. (Related to W1.) Apart from the exact-versus-near precision checks, two standard tools appear to yield the same identifiability verdicts in principle: a GSVD-based diagnostic at the same cut and normalization, and a standard estimability test (i.e., checking ℓ∈row[Ad;Ap]\ell\in\mathrm{row}[A_{\mathrm d};A_{\mathrm p}]). Could the authors clarify the additional practical value of the integrated audit? For instance, on the existing cases (e.g., the mean-field and MHD studies), which steps of the audit would a practitioner using those tools be likely to omit or get wrong, and how would that affect their coefficient interpretation or sampling decision?
  2. (On the reliability of reading-level verdicts.) In App. B.2.2, for the enlarged steady-state design, the share ∥B0⊤ℓ∥/∥ℓ∥\|B_0^\top\ell\|/\|\ell\| of the g+mAg+mA reading falls between the two reading thresholds, so the audit returns no numerical verdict. The reading is instead certified by exact rational elimination on a rational parameterization of the analytically known limit cycle. Is such certification a general, automated part of the audit, or does it require problem-specific symbolic analysis? The answer bears on how reliably reading-level verdicts can be obtained in practice.

Please let me know if I have misunderstood any part of the paper; I am open to revising my assessment.

Reviewer_gc8jOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the identifiability of coefficients in dictionary-based PDE models. The authors point out that even when a model achieves very low data-fitting error and physics residual, the coefficients associated with terms in the dictionary may still not be uniquely determined. To address this issue, the authors propose a structural auditing method that analyzes the joint null space of the data map and the physics-constraint map, identifying coefficient directions that cannot be distinguished by either the observations or the imposed physics.

The experiments cover 19 different dictionaries as well as a public PySINDy constrained-regression example. The results show that the proposed audit can identify multi-term dependencies that are missed by conventional pairwise or triple screening methods, and can further reveal which individual coefficients or coefficient combinations are actually identifiable under a given observation design and physics constraint.

Critical strengths

The paper addresses a real and important issue in PDE identification. Many PDE discovery methods directly interpret the optimized coefficients as physical mechanisms, but low fitting error and a small physics residual do not necessarily mean that these coefficients are uniquely identifiable. I think it is useful that the paper makes this issue explicit and provides a clear diagnostic framework for it. The method itself is also relatively simple and interpretable.

The experiments are also reasonably thorough. The authors test the audit on different dictionary settings, compare it with simpler pairwise and triple screening methods, and check its sensitivity to tolerance and numerical precision. They also include public PySINDy and DeePyMoD examples.The steady-state versus transient-data case study clearly shows that adding more samples does not always resolve the identifiability problem, while using more informative observations can.

Critical weaknesses

I found the presentation dense and hard to follow. There are many numerical results, but the main takeaway from each experiment is not always easy to see. Instead of numbers, I would prefer to see more analysis and insight behind the number.

The treatment of nonlinear operators is stll limited. The paper uses the Frechet Jacobian at a specified u_0, which is a local first-order sensitivity diagnostic only. Therefore, I am not sure how well the framework can generalize to nonlinear settings.

Questions for authors

please refer to weakness

Reviewer_uRb9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether the coefficients in a physics-constrained dictionary model can be uniquely determined. The method checks the shared null space of the data and physics matrices. It identifies coefficient directions that cannot be distinguished and checks which coefficient combinations can still be determined.

Critical strengths
  1. The paper studies a useful problem. A small prediction error and a small physics residual do not mean that all coefficients are uniquely determined. I think this distinction is important when the coefficients are used to explain a physical system.

  2. The authors check the numerical results carefully. They use different tolerances, higher precision, and symbolic calculations to distinguish exact dependencies from approximate ones. They also report cases where the audit cannot give a clear conclusion.

Critical weaknesses
  1. My first concern is about the contribution of the mechanism. The main idea is to check the shared null space of the data and physics matrices. The rank tests, witness directions, and tests for coefficient combinations can already be obtained from standard SVD or GSVD analysis. The paper also acknowledges this. Table 5 and Appendix B.1.3 show that stacked SVD finds the same dimensions and nearly the same subspaces. I think the authors need to explain more clearly what their method adds beyond these standard tools. The comparison with pairwise and triple checks is not enough to answer this question because those methods only check limited types of dependence.

  2. My second concern is about the scope of the experimental evidence. In Table 2, the original public PySINDy example has no remaining joint-blind directions, while the main ambiguous case uses observations exactly on a limit cycle. Section 4.1 and Tables 2–3 provide useful evidence that transient observations improve identifiability and coefficient recovery. However, the benefit of transient observations was already known for this model, and the additional public cases also have no remaining joint-blind directions after applying their constraints.

Questions for authors
  1. What does the proposed audit add beyond standard SVD or GSVD analysis? Could the authors clearly separate the new parts from the existing results and show their practical benefit using the same normalization and tolerances?

  2. Could the authors show an experiment with noisy observations or estimated derivatives where the audit gives a stable coefficient conclusion? How should the tolerance be chosen without knowing the true coefficients? I would also like to see whether the audit helps decide what additional observations to collect.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.