acceptodds
Withdrawn from ICLR 2027

Caustic Spectroscopy: Identifiability and Acquisition Limits of Topological Event Timing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Caustics carry the shape of a transparent surface, but their intensities are unstable under exposure and pose. We study a sparser, analytically tractable observable: sweep one knob of the illumination and record only the control values at which the caustic changes topology, a catastrophe spectrum. Isolated nondegenerate events are Morse critical points of the fold function, so their thresholds move with a closed-form differential, and the rank of the resulting event Jacobian certifies local identifiability: outright where it is full, which needs screen positions, and up to the sweep's symmetry group for thresholds alone, whose Jacobian that symmetry keeps below full rank. Paraxially the thresholds are the critical values of the surface's curvature fields. On simulated glass reliefs the differential matches finite differences to , and the singular values order a trained network's per-direction error across twelve nominal reliefs after one held-out calibration. The analysis also marks its own limits, and we measure each. Umbilic events are degenerate and leave the signature unchanged. Nine events per relief let a trained inverse recover half the shape variance up to rotation, measured against the best data-free predictor and carried mostly by their counts and types. Adaptive bisection localizes matched events eighteen times more precisely than a uniform sweep, but only on the idealized signature: injecting spurious transitions into it reproduces the end-to-end collapse, which is complete by a precision of . All evidence is simulated.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
topological descriptors, identifiability, singularity theory, catastrophe theory, Morse theory, Euler characteristic transform, quotient geometry and symmetry, Cramer-Rao bound, adaptive experimental design, inverse problems, caustics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_u7diOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes an interesting theory for estimating transparent object's relief by characterizing the caustic patterns (i.e., the catastrophe spectrum SλS_\lambda). In addition to the theory, the authors provide the detailed analysis (i.e., stability and identifiability) to characterize the applicability of the proposed theory for the object relief estimation. Simulated data is used for evaluating the proposed method.

Critical strengths

First, I have to admit I am not an expert in this particular field, so I provide my evaluation based on my reading and understanding of the paper content.

This paper asks an interesting question: whether the catastrophe spectrum, which is much less informative than the caustic image, can be used for recovering the transparent object's relief. To address this issue, the authors use the tools from dynamical systems and geometries for building the theory, which further lay a foundation for the above problem. Additionally, the authors provide detailed experimental results for evaluating the derived theory by using simulated data.

Critical weaknesses

One of the major concerns is that this method seems not competitive to the CNN baseline while the authors try to answer an interesting question.

Additionally, the theory part of this paper seems to have low readability. First, definition, lemma, proposition, and theorem are dumped without clear contextual explanation. This makes the readers "extremely" hard to following the story in the paper and lack of interesting of reading right after. It is highly recommended to provide a figure or a top-view summary for explaining the flow of the section and the relation between those theoretical components. Besides that, the background of the theoretical components should be provided (probably in Appendix). For instance, the connection of the jump and the catastrophe spectrum to the object relief should be included and the intuition for why nondegenerate cases are more important should be explained.

Overall, the structure of the paper needs some massaging. Right now, it feels like over-fried chicken—hard to digest, even though the ingredients seem promising (i.e., the paper addresses an interesting research question).

Questions for authors

What is notation EE in line 157?

In addition to the above notation, it is highly recommended to include some explanation and/or intuition for the important variables in the paper, such as β(⋅)\beta(\cdot) and gg.

Reviewer_4PxXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper represents a transparent relief through the illumination settings at which a topology signature changes. It derives event sensitivities by implicit differentiation, uses the resulting Jacobian to characterize local identifiability modulo acquisition symmetries, and relates its singular values to reconstruction error. Experiments on simulated reliefs evaluate learned inversion, adaptive event localization, and sensitivity to image nuisances.

Critical strengths
  1. I believe the work has a coherent local math framework. The envelope derivative, symmetry analysis, and rank conditions fit together naturally.
  2. Holding out nominal reliefs when calibrating the effective noise scale is stronger than fitting each nominal separately.
Critical weaknesses
  1. The signature β(λ)=(b0{gλ<0}, b0{gλ>0})\beta(\lambda)=(b_0\{g_\lambda<0\},\,b_0\{g_\lambda>0\}) describes regions in the source domain of the ray map. The visible caustic is the critical-value set F({g=0})F(\{g=0\}) on the screen but their topologies need not coincide because projection can introduce intersections and overlap without changing the source-domain component counts. Proposition 1 did not establish that the descriptor can be recovered from irradiance.
  2. Eq. 3 is only exact for the stated linear–Gaussian observation model. For the networks receive noiseless oracle thresholds, and σϵ\sigma_\epsilon is introduced afterward to represent approximation error, there is no derivation establishing that this error behaves like isotropic additive noise in event coordinates.
  3. I believe the manuscript lack some differentiation with the prior work they cited (Szep et al., 2021). Places like Eq. 10 are particularly relevant.
Questions for authors
  1. Table 5 reports NRMSE 0.293 on clean frames but 0.049 under tilt and 0.039 under zoom. Could the author give an explanation for such an anomaly?
  2. Under what conditions can the source-domain signature be recovered from screen irradiance?
  3. Can the self-calibrating baseline anomaly be resolved, and can improved image detection restore any adaptive acquisition advantage?
Reviewer_DSVsOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies catastrophe spectra: the illumination control values and event types at which a caustic undergoes topological changes. It derives a differential for isolated nondegenerate event thresholds and relates the event Jacobian to local identifiability, including ambiguities induced by acquisition symmetries. It further investigates whether the Jacobian predicts the directional errors of a learned inverse and establishes conditions for adaptive event localization. Experiments on simulated refractive and reflective reliefs evaluate these predictions using oracle spectra, image-derived events, and image-based baselines. The results support the local analysis while demonstrating substantial limitations from event crowding, detector errors, and inference across a broad shape prior.

Critical strengths
  1. A coherent and interpretable analysis of the proposed measurement. The threshold differential and symmetry-aware rank conditions explain which shape variations can be recovered and which remain unobservable (Theorems 1–2). Reporting how frequently the rank conditions hold makes the analysis informative beyond a purely conditional theorem.

  2. Careful experimental controls and transparent reporting of limitations. The learned-error experiment calibrates its effective noise scale on other nominal reliefs and reports a directional Spearman correlation of 0.85 (Section 5.2). Table 1 separates timing information from event counts and types and uses an appropriate constant baseline under the quotient loss. The controlled false-positive experiment also helps explain why adaptive acquisition fails with the image detector.

Critical weaknesses
  1. The practical accessibility of the analyzed observable remains unresolved. The ideal fold-set signature requires the unknown ray map, whereas the image detector measures the topology of quantile masks (Section 3). Consequently, the approximately 18-fold localization improvement becomes only about 1.1-fold on rendered images (Section 5.4). Together with simulation-only validation, this limits the demonstrated practical significance of the acquisition strategy, although the authors clearly acknowledge the limitation.

  2. The learned-error validation does not fully isolate the fixed-event regime underlying the local surrogate. Event counts remain unchanged in only 43% of draws at the main perturbation scale, and in 56–68% of draws in the smaller-scale experiments (Section 5.2; Appendix F). The network can therefore exploit event changes absent from the Jacobian model. The reported agreement is encouraging, but its interpretation would be stronger if the contribution of such changes were separated from the local differential information.

Questions for authors
  1. Under the stated ideal optical assumptions, can the fold-set signature, or its event thresholds and types, be determined from noiseless screen irradiance alone? Please clarify the assumptions needed to distinguish limitations of the current detector from potential ambiguities in recovering the proposed observable.

  2. Does the directional-error agreement remain comparable when restricted to perturbations preserving the identities and types of the nondegenerate events? An analysis of this subset would help determine how much of the agreement specifically validates the fixed-event Jacobian model.

Reviewer_nbh3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether the shape of a transparent surface can be estimated by looking only at when its light pattern changes, instead of using the full images. The authors develop a framework to determine what shape information can be recovered from these changes and how accurately it can be estimated. They test the method mainly using simulated data and neural-network models such as DeepSets and CNNs. The results show that these light-pattern changes contain useful information about the surface shape, but methods using the full images are generally more accurate.

Critical strengths

The paper is carefully done and provides a strong theoretical analysis of the proposed sensing method.

Critical weaknesses

My main concern is that the paper is not strongly related to ICLR. The main contribution is in optics and sensing, while machine learning is mostly used for evaluation rather than as a new method. Also, all experiments are based on simulated data and a specialized optical setting, so it is unclear how broadly the approach would apply to other machine learning problems.

Questions for authors

Could the authors clarify what they view as the main machine learning contribution of the paper, beyond using neural networks to validate the proposed sensing and identifiability framework?

Comments on the paper

3 comments
Reviewer_u7diOpenReview ↗

Motivation

Thank you to the authors for addressing our concerns. The readability of the theoretical part has improved substantially. However, I still have concerns regarding the CNN experiments, particularly the motivation for considering event-based measurements.

My concerns are mainly twofold.

  1. While the paper poses an interesting question—whether the underlying geometric shape can be recovered from topological events—I am still not fully convinced by the motivation for restricting the observations to event timings. In particular, extracting these events still requires capturing images across the illumination sweep. Given that these images are already available, and that image-based reconstruction achieves substantially better accuracy, why should we discard the image information and consider this more constrained representation in the first place? What practical advantage does this restriction offer?

  2. One possible motivation, as I understand it, is the invariance and robustness of caustic events to variations in camera settings. However, I am not yet convinced that this advantage is sufficiently compelling in the context of modern vision models. With the recent advances in visual representation learning, particularly vision foundation models, it would be interesting to understand whether image-based approaches could also achieve comparable robustness to camera-dependent variations through appropriate training or adaptation. In fact, the paper's own experiments show that data augmentation substantially improves the robustness of the CNN baseline. Therefore, I would like to better understand what unique advantage event-based measurements offer over modern image-based approaches, beyond their analytical tractability.

Also, I share similar concerns with Reviewer nbh3. Based on your response, I understand that the work is fundamentally studying topological questions. However, I am still unclear about what specific ML problems could benefit from the proposed analysis and how it could contribute to the broader ML community. It might be helpful to give some concrete examples how the proposed framework can bridge to the ML applications.

Thanks to the authors for the clarification. I appreciate the additional discussion of the machine learning contributions and applications beyond optics. However, I still feel that the paper is mainly theoretical, with limited contributions to machine learning methods. The experiments are also based only on simulations. I will keep my score at 2.

Thank you for the detailed responses and additional experiments. The rebuttal has clarified some of my concerns, and I appreciate the authors' efforts to address the feedback. However, some of the limitations discussed in my original review remain. Overall, my assessment remains unchanged, and I will keep my original rating.

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

Related papers

Discussion (0)

Sign in to comment.