Equation discovery is often declared complete when the selected support, fitted coefficients, or validation error appears stable. Those criteria describe the behavior of an estimator. They do not answer the scientific question: could another operator that is still compatible with the evidence behave differently in the regime that matters?
We define completion through the version set of all data-compatible operators. A target state envelope, declared before the unknown future response is observed, induces a target-action diameter on this set. This diameter measures the scientific ambiguity that remains after the completed experiments.
We prove an exact conditional ambiguity law and show that the realized diameter cannot increase as valid experiments are added. Under a common one-sided stability bound, we further transfer operator ambiguity to unseen-trajectory ambiguity and obtain a monotone stopping rule. Validated derivative error and bounded model discrepancy enter through explicit uncertainty inflation.
On public Reτ≈1000 turbulent-channel DNS from JHTDB, two acquisition blocks reduce normalized target-action ambiguity from 0.418 to 0.282 for an outer-layer target and from 0.434 to 0.313 for a near-wall target. Pressure and viscous multiplier coordinates nevertheless continue to span their declared prior intervals. We make no trajectory-stopping claim for this turbulent case because no useful validated global one-sided stability constant is available for the declared regime.
For an analytically controlled reaction–diffusion system, both the invariant target envelope and the stability constant are certified. The locked conservative trajectory certificate first crosses its predeclared tolerance at experiment 6. Across 100 independent noise realizations, 99 runs certify by experiment 8, the median stopping step is 6, every certificate path is monotone, and no operational false stop occurs among the 99 stopped runs.
The central distinction is therefore between finding a plausible equation and having enough evidence to declare the scientific uncertainty resolved.
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.Reviewer_eCupOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary
This paper focuses on the identifiability issue in equation discovery using limited observations. The paper provided theoretical analysis, and experimental results on real turbulent DNS as well as a simulated reaction-diffusion system.
Critical strengths
The paper investigates a very interesting and important problem -- given limited data, to which level can we identify the physics operator we have learned, and can they reliably generalise in the unseen regime we care about.
Critical weaknesses
I am deeply concerned with the quality of the manuscript. There are multiple undefined references (e.g. line 398) and not clearly referred experiments (experiment 1? experiment 6?) The manuscript does seem to be very polished to me.
The take-away as highlighted by the authors "A fitted equation may look convincing while the data still permit other operators with materially different scientific consequences." seems to be a well-known problem in machine learning, generalisation is not guaranteed unless some inductive biases we impose allow this. Why is this a new insight?
Questions for authors
What is the main purpose of this paper? is that a benchmarking paper to show generalisation failure of data-compatible operators?
This paper studies when the available observations are sufficient to stop collecting data for PDE discovery. It constructs a set of all data-compatible operators and measures their maximum disagreement over a predefined target state envelope. Under a common one-sided stability bound, this disagreement yields a certificate for future trajectory accuracy. Experiments examine operator uncertainty on turbulent-channel DNS and demonstrate certified stopping on a controlled reaction–diffusion system.
Critical strengths
1.Determining when additional data collection is no longer necessary is an important consideration in scientific discovery, particularly when experiments are expensive.
2.Assessing whether all data-compatible models agree on the intended application provides a useful perspective beyond coefficient convergence.
Critical weaknesses
The experiments assume the correct equation structure. PDE discovery involves coupled uncertainty in term selection and coefficient estimation. However, the experiments estimate only two or three coefficients within predefined structures, with all intervals excluding zero. The results therefore support parameter uncertainty certification under known structures, rather than general PDE discovery. This restricted setting limits the contribution: certifying coefficients within a prescribed equation does not establish that competing equation structures have been ruled out, which is central to deciding whether PDE discovery is complete.
The theoretical advance appears limited.
Theorem 1 follows from triangle inequalities and a midpoint construction; monotonicity follows from nested feasible sets. Proposition 2 gives standard radius–diameter bounds, while Theorem 3 uses a classical differential-inequality argument. The contribution beyond established theory is insufficiently distinguished. The paper also omits the closely related Data Informativity: A New Perspective on Data-Driven Analysis and Control (van Waarde et al., 2020).
The stability assumptions restrict practical applicability.
The trajectory certificate requires a validated stability bound covering all admissible models and guaranteed containment within the target envelope. These conditions are established only for the simple reaction–diffusion example; the turbulence experiment provides no trajectory-stopping guarantee. Positive stability constants can also make the bound prohibitively conservative over longer horizons.
Practical advantages over existing stopping rules are not demonstrated.
The complete stopping experiment uses one system, and the 100 repetitions vary only observation noise. Earlier triggering by conventional rules does not establish that their decisions are incorrect. The paper therefore does not demonstrate improved prediction risk or reduced acquisition cost, and lacks comparisons with uncertainty-aware set-based methods.
Line 398 contains an unresolved cross-reference (“??”), preventing readers from locating the referenced comparison.
The paper presents a stopping criterion in a SINDy/PDE discovery context relative to a specified region of the state space (identification only needs to be good enough in the region of state space you care about).
I have two major problems with the paper: (i) it's write-up is convoluted and hard to follow; (ii) it tests only for synthetic data and does not demonstrate usefulness outside of that regime (and the method does not give any guarantees then). I do not consider it of high enough quality for publication, so I recommend rejection.
Summary (for an ODE problem)
1. Setting. Consider
x˙=f(x)+g(t),x(0)=x0,x(t)∈Rn,
where g is known and f is unknown. Use a SINDy ansatz
fξ(x)=Θ(x),ξ,Θ(x)∈Rn×p,ξ∈Ξ⊂Rp,
where the library Θ is fixed and Ξ is a box of admissible coefficients.
2. Experiments and least squares. An experiment e is one run with initial condition x0(e) and forcing g(e), sampled at times te,1,…,te,me. It leads to a least-squares problem
where xe,k=x(e)(te,k) and x˙e,k are the (in practice numerically estimated) time derivatives.
3. Version sets. Introduce
Vk=ξ∈Ξ:∣Aeξ−ye∣2≤σe,e=1,…,k,
the set of coefficients consistent with the first k experiments. The tolerances σe are chosen such that the true coefficients satisfy every constraint with high probability (for Gaussian noise of known variance, via a χ2 quantile and a union bound). The sets are nested, Vk⊆Vk−1.
4. Target region. Declare a region U⊂Rn of interest, together with target initial conditions X0, forcings G and a horizon [0,T], such that all target trajectories stay in U for every ξ∈Ξ (e.g., by a maximum principle).
5. Criterion. Measure the worst-case disagreement of the remaining right-hand sides on U,
D(Vk)=supξ,ξ′∈Vksupv∈U∣Θ(v)(ξ−ξ′)∣.
6. Trajectory guarantee and stopping. Under a uniform one-sided Lipschitz bound with constant μ on U, a Grönwall argument gives
∣xξ(t)−xξ′(t)∣≤μeμt−1,D(Vk)for all ξ,ξ′∈Vk.
The proposed rule stops at the first k with μeμT−1,D(Vk)≤τ. With high probability, every remaining model then reproduces all target trajectories within τ of the truth.
The paper instantiates this for a (spatially discretized) reaction–diffusion equation with two unknown reaction coefficients. It applies only the right-hand-side criterion of step 5 to turbulent channel-flow data, where no stability constant μ is available.
Critical strengths
Equation discovery is a timely and important topic.
Critical weaknesses
Major problems
(i) The presentation of the paper is not good. Frequently, context is not given and symbols are not introduced. Additionally, it reads as if large parts were LLM-written and hardly edited.
(ii) The guarantees rest on critical assumptions. The data model
ye=Aeξ⋆+ηe,∣ηe∣≤σe(with high probability),
requires critical assumptions:
an exact model, f=Θξ⋆;
a known noise model.
Both are typically not given in practice, and then the guarantees break. The paper tests only on synthetic data (a simulated reaction–diffusion system with injected Gaussian noise of known variance, and DNS data of the Navier–Stokes equations), where both assumptions hold by construction, so the interesting questions are never touched.
The paper considers model discovery and the possibility of certifying the actual underlying PDE being discovered. More broadly, the method proposes the discovery as an operator learning framework.
Critical strengths
The paper provides a very nice analytical framework for more deeply understanding the model discovery architecture. This is a unique, if not niche, analysis of new methodology for "uniquely" identifying models. It does provide some very nice perspectives that should most definitely be in the literature.
Critical weaknesses
The work is quite niche and the demonstrations on reaction-diffusion models and flow physics at a modest Re around 1000 is quite limited. There is not a single figure in the paper that would recommend the work to a broader audience or beyond a limited result. I would expect much more in terms of application to models, especially given the diversity of models considered for such methods. Also, the noise studies are hard to digest to see how much is actually being added here. I think even with modest noise it would be very difficult to properly identify models and this is very difficult to assess with the current work.
Questions for authors
It would seem that identification and method proposed would be corrupt by simply increasing the signal-to-noise, is that true?
The derivative error is fundamental as both methods like pySINDy and pySR methods, for instance, really on accurate derivate estimates. Even modest noise imposes serious limitations, how is this accounted for?
From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.