acceptodds
Under review as a conference paper at ICLR 2027

Recoverability Is a Subspace Property: A Benchmark for Certified State Estimation from Partial PDE Observations

Abstract

Reconstructing hidden physical states from limited measurements is a fundamental task in scientific machine learning for state monitoring and physical analysis. In partially observed partial differential equation (PDE) systems, learned models can exploit statistical correlations in training data to accurately predict state directions that are weakly constrained by observations. Aggregate reconstruction error therefore provides an incomplete picture of how predictive performance relates to observational information. We introduce UniPDE-Bench, a direction-wise evaluation protocol that uses local observation geometry as a reference independent of the estimator. In a common directional basis, the protocol separately assesses empirical recovery, agreement between prediction claims and observation geometry, and confidence-based selection across claim coverages. It constructs a complete right singular basis of the joint observation Jacobian after noise whitening and normalization by a state metric, partitions directions using a relative sensitivity threshold, and evaluates selection through recovery and abstention curves. This geometric reference complements predictive accuracy by characterizing the local observational constraints on selected directions. Experiments across models and observation configurations show that similar aggregate reconstruction errors can correspond to substantially different directional recovery and selection outcomes. In the transport configuration studied, rankings based on conformal interval width achieve an overall geometric ranking score above chance, yet their abstention rate on below-threshold directions falls below the random baseline at some high claim coverages. These findings motivate separate assessment of aggregate predictive accuracy, overall ranking quality, and direction selection at specific coverages. By relating these assessments to local observational constraints, UniPDE-Bench reveals differences in recovery and selection that aggregate metrics obscure.

Primary area:
reinforcement learning
Keywords:
Partially observed PDE systems, Recoverability, Observation subspaces, Selective prediction, Scientific machine 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_VPaEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors are proposing UniPDE-Bench that is evaluation protocol for learned reconstruction of partially observed PDE systems. Instead of relying only on reconstruction error, the authors analyze the singular-value decomposition of a noise-whitened, state-normalized observation Jacobian. The resulting singular vectors identify joint variations of the hidden state that are relatively well or poorly constrained by observations. Using this geometric reference, the paper evaluates recovery accuracy along retained directions, agreement between confidence rankings and observation sensitivity.

Critical strengths

1- The paper highlights an important limitation of conventional reconstruction benchmarks that is low error does not imply that the measurements uniquely constrain the reconstructed state.

2- Separating recovery capability, geometric agreement, and selective prediction quality is a strength that I like (illustrated in Table 2). The framework avoids equating poor uncertainty ranking with poor state reconstruction.

3- The use of a complete right singular-vector basis is appropriate, particularly when the observation dimension is smaller than the hidden-state dimension.

4- The experiments include broad ranges. It includes controlled inverse problems, six neural estimator families, five learning configurations, five random seeds, and ID/OOD evaluation.

Critical weaknesses

1- The main geometric construction is based on a normalized Jacobian, singular-value decomposition, and threshold-based subspace partitioning. These are established tools in inverse problems and observability analysis (although usage is useful, but combination of well known existing tools).

2- The proposed sCRS automatically assigns zero to non-abstaining predictors, regardless of their reconstruction accuracy. The penalty choices and fixed operating points require stronger justification and sensitivity analysis (that I could not find in appendix).

3- The retained subspace depends on the selected threshold, state metric, and linearization point. Basically we need either ablation or sensitivity analysis (that I dont find in your main nor appendix) to establish whether its main conclusions remain stable under different choices.

4- The retained subspace dimensions vary substantially across sensor configurations, directly affecting recovery scores and therefore, differences in these scores may reflect changes in the evaluation space rather than degradation in model performance (any idea to alleviate?).

5- The framework penalizes claims along poorly observed directions even when predictions are accurate and statistically reliable. To me it remains unclear whether geometric abstention improves practical predictive reliability compared with conventional uncertainty-based selection or not.

6- The experiments focus on three simulated PDE systems with only 17 state coefficients, where rank deficiency is often guaranteed by the observation dimensions. Larger and more realistic inverse problems could strengthening applicability of your approach (I am of course not deciding based on num of experiments, nor ask you to do so in rebuttal as the time is limited. I rather want intuitive explanation about how such extension is possible and later be reflected as limitation in main body).

7- The evaluation would benefit from clearer experimental specifications and broader sensitivity studies covering recovery thresholds, numerical discretization, and confidence-selection criteria.

Questions for authors

I tried to embed my questions in the weaknesses. Please see them above.

Reviewer_xCZEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces UniPDE-Bench, an evaluation framework for partially observed PDE systems. Motivated by the distinction between observation-supported recovery and prior-supported prediction, the authors propose to analyze the local observation Jacobian using SVD and partition the state space into retained and below-threshold directions based on relative singular values. The framework then evaluates directional reconstruction errors and confidence-based selective prediction using recovery and abstention metrics. Experiments across three PDE systems compare the proposed metrics with conventional reconstruction errors and demonstrate discrepancies between prediction accuracy and sensitivity-based selection.

Critical strengths
  • Meaningful motivation. The paper highlights the distinction between reconstruction accuracy and observation-supported recovery.
  • Extensive experiments. The evaluation covers multiple PDE systems, model architectures, and confidence-based selection methods under ID and OOD settings.
Critical weaknesses
  1. Gap between the motivating problem and the proposed evaluation framework.

The Introduction motivates the need to distinguish observation-supported recovery from prior-supported prediction. However, the proposed framework does not directly distinguish these two sources of predictive information. Instead, it partitions state directions according to relative singular values of the observation Jacobian, which characterize local observation sensitivity rather than actual recoverability or the contribution of data priors. Although the authors acknowledge this distinction, the experiments primarily evaluate agreement with the sensitivity-based partition rather than demonstrating that the proposed framework addresses the original motivating problem.

  1. Limited methodological novelty.

The core framework relies on established SVD-based sensitivity analysis to identify directions with relatively weak observational responses. The subsequent evaluation combines this decomposition with directional reconstruction errors and existing selective prediction metrics. While these components are organized into a unified protocol, the paper does not introduce a substantially new method for analyzing observation sensitivity or distinguishing observation-supported recovery from prior-supported prediction. The experimental findings largely demonstrate properties of the predefined sensitivity partition and the associated evaluation metrics.

Questions for authors
  1. How does the proposed framework distinguish observation-supported recovery from prior-supported prediction, as motivated in the Introduction? Since the SVD-based partition depends only on the local observation Jacobian and does not account for the learned data prior, what additional information does the framework provide about the source of a model's predictive accuracy?

  2. What is the key methodological contribution beyond applying established SVD-based sensitivity analysis and selective prediction metrics? In particular, which findings could not be obtained through conventional singular-value analysis combined with directional reconstruction errors?

Reviewer_TAE6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a protocol for evaluating recoverability in partially observed PDE systems. In particular, it focuses on parameter/coordinate-specific evaluation.

Critical strengths

Due to the paper's poor clarity, I was not able to properly assess the strengths of the proposed approach.

Critical weaknesses

The paper is not clearly written, especially for people not familiar with this specific area. This manifests in various ways, making it difficult for me to assess the novelty and soundness of the paper.

Awkward terminology. Some words feel out of place in this particular context. Either because they are synonyms of more suitable words, or the technical meaning is not properly introduced. l034: "state variations" l040: "joint ambiguity" l053: "...state-metric normalization define the comparison scales" l058: "which predictions are claimed and which are left unclaimed" l107: "claim order" l131: "interface parameters" l132: "working point"

Poor motivation. For a person not familiar with the presented problem, it is not clear whether the described issues are important and need to be addressed. Why are they important? What would happen if they remain unaddressed? The toy example does not sufficiently motivate the practical consequences in the PDE setting. A clear case study/example showing the difference between the current approaches and the proposed approach early in the paper would make the motivation much clearer. Right now, the paper does not do a good enough job convincing me that I should care about this problem and hence the solution.

Unclear problem formulation. The problem set up in Section 3.1 does not contain enough details to clarify the task of interest. A clear problem formulation should state what our inputs are - things that we assume are known or things to which we have access, like data. It should also clearly state what the objective is and ideally use the mathematical formalism to highlight the gap that the paper is addressing. How does success look like? These things are not currently present.

*Insufficiently defined notation". Some terms are not well defined in the paper. For instance, η\eta in l133. Some terms are described, but the description leaves out a lot of details. For instance, although we know that GG is a shape metric, it is not described where it comes from. Is it problem dependent? Although H\mathcal{H} is defined as an observation map, we do not have any details about the kind of map it can be. Is this a solution operator of some underlying PDE?

Questions for authors

Could you please clarify your motivation, problem set-up, and contributions to facilitate better understanding by people with limited familiarity with those specific problems?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.