acceptodds
Under review as a conference paper at ICLR 2027

On Parameters of Nonlinear Scalar Dynamics from Video: Invariants, Calibration, and Identifiability

Abstract

Physical parameter estimation from video aims to recover the parameters of a known family of governing dynamical equations from pixel observations. Existing identifiability theory for this setting has focused on linear time-invariant (LTI) second-order systems, leaving open what can be identified for nonlinear scalar dynamics. We develop an identifiability theory for nonlinear scalar second-order ODEs, organized by how their velocity dependence interacts with changes of the learned state coordinate. Under a shared non-collapsed state map and explicit same-state velocity-coverage conditions, we show that parameter identifiability depends on the ODE family: some parameters are uniquely identifiable, while in other families only invariant parameter combinations are identifiable or external physical calibration is required. For laws that are at most linear in velocity, compatibility forces affine coordinate alignment, yielding explicit parameter relations, invariants, and calibration conditions. This affine conclusion extends to broader finite velocity-feature families when coordinate curvature can be separated from the declared velocity dependence. For families admitting a squared-velocity term, nonlinear coordinate ambiguity can remain; a law-derived normalization instead enables affine comparison between canonical laws. Experiments on synthetic systems and real pendulum and free-fall videos support the predicted parameter relations, coverage effects, and calibration requirements.

Primary area:
learning on time series and dynamical systems
Keywords:
physical parameter estimation, parameter identifiability, nonlinear dynamical systems, system identification, learning from video

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_v8CaOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper investigates the problems and failure modes associated with physical parameter identification from videos. It focuses specifically on a class of nonlinear scalar dynamical systems, as opposed to previous work which focused on LTIs. The paper establishes a framework for determining parameter identifiability and diagnoses which ODE families present uniquely identifiable parameters, and which parameter identification questions may be ill-posed. As a solution, the authors present the coordinate ambiguity, invariants and necessary parameter calibrations for different ODE families. The theoretical predictions are tested on synthetic data spanning 6 ODE families and 3 real-world datasets.

Critical strengths

I rank the paper highly in terms of significance and timeliness. Physical parameter estimation is an important problem in many fields of science and engineering, and while there are many emerging methods to solve this problem, as described by the authors, many of the existing methods need to implement manual tricks or specific normalization schemes in order to deal with the intrinsic invariances, symmetries, coordinate ambiguities in the specific dynamical system/governing equation they are considering. This both wastes time and resources for researchers and often makes it difficult to compare different methods (as they may employ different tricks or normalizations). This paper establishes a generalized framework for identifying a priori what parameters can be uniquely identified from the data, or what calibrations/initial conditions are necessary, to start from a well-posed inverse problem. I believe Table I can already be very useful for many natural scientists as well as AI researchers who are tackling these inverse problems, by clarifying what ground truths, dataset calibrations and initial conditions are necessary. It can also make new methods more standardized and comparable.

The analysis of different ODE families in the paper seems to follow sufficient rigor. The theoretical contributions are meaningful and are confirmed by the experiments on real and synthetic data.

Critical weaknesses

Real-world datasets: I found the recording of yet new datasets for this problem slightly confusing. There are already multiple established datasets for physical parameter estimation in the literature: for example many pendulum datasets are available which could provide more baseline comparisons, and I believe the Delfys dataset also contains free-fall videos. Tests on existing datasets would strengthen the results by providing existing baselines.

Lack of baselines: The empirical methods are explained in a manner which lacks crucial details about the parameter estimation method, the specifics of the shared encoder, loss function and optimizer used, etc. in the main paper. However, my understanding is that the same encoder is used for all empirical results. The accuracy/stability/reliability of this encoder and the results gained from it are difficult to interpret or trust in the absence of baseline comparisons. Would it be possible to show how the theoretical understanding gained from this paper empirically improves upon previous baseline methods?

Presentation: The clarity of the language is good, however, I found the methods/implementation details described in Section 5 extremely difficult to follow. I also found the explanation of the empirical results and the captions in Section 5 vague and confusing. What is shown in the tables/figures (what are shown on the axes, the legend, what settings are being compared, etc.) need to be clarified in the text.

Interpretation of the results: Section 5 text lists the empirical results. However, it is a bit unclear what the take-home message is from these results. What were the hypotheses for these experiments, what did the authors expect to find? What are the conclusions, important observations? It would be nice to detail if/how/why the experimental observations confirm the theoretical predictions.

Reproducibility: Will the code for the empirical experiments be made available? Sharing the code would make the methods more clear and ease any reproducibility concerns. The reproducibility statement mentions that the code for the synthetic data will be available, but what about the real-world experiments? Will the datasets and the evaluation code be made publicly available?

Generalizability: What about other types of dynamical systems, families of ODEs? How well can the presented framework generalize to non-scalar/coupled systems? Is it possible to say a few words about it in the main text and not just Appendix K?

Questions for authors
  • Why not use existing datasets instead of collecting yet new datasets with no baseline comparisons?

  • What is the parameter estimation model which was used? Is the encoder the same one as used in the IRIS paper? If yes, can the authors state this explicitly?

  • Why are other baseline parameter estimation methods not used/compared against? Would it be possible to show how the theoretical understanding gained from this paper empirically improves upon previous baseline methods?

  • Is it possible to better explain/clarify the methods used in Section 5 in the main text (i.e. not just in the appendix, as it is currently almost impossible to understand the basics of the methods only from the main text)? The tables/figures (what are shown on the axes, the legend, what settings are being compared, etc.) also need to be clarified in the text.

  • What are the take-home messages of Section 5? What were the hypotheses for these experiments, what are the conclusions, important observations? It would be nice to detail if/how/why the experimental observations confirm the theoretical predictions.

  • Will the datasets and the evaluation code be made publicly available for the real-world experiments?

  • Is it possible to say a few words about the generalizability/possible extensions of the framework in the main text and not just Appendix K?

5 comments
AuthorsOpenReview ↗

Author Response (Part 1/3)

We sincerely thank the reviewer for the positive assessment of our paper and for recognizing its significance, theoretical rigor, and practical value. Below, we address the concerns and questions raised in turn.

S1. Significance of the problem and practical value of Table 1.

A. We appreciate the recognition of Table 1's potential value to researchers working on physical parameter estimation across different fields. This reflects our motivation: to clarify what the data can identify and what additional calibration is needed. Table 1 summarizes representative dynamical families; our analysis covers broader classes of nonlinear scalar second-order ODEs under the stated assumptions. To support applications beyond these examples, we provided a lightweight tool in the supplementary material (Appendix O, Pages 63–65). For a user-declared ODE family within its supported scope, the tool reports the required data conditions, identifiable parameter invariants, and calibration requirements.

Q1 & W1. Why did the authors record new datasets instead of using existing ones?

A. Thank you for the question. Among our three real-world experiments, the pendulum experiment directly uses the public IRIS dataset, while the side-view and overhead free-fall experiments use our own recordings. As explained at the beginning of Appendix N (Page 50), although IRIS provides similar free-fall videos, these datasets do not satisfy the three-slope divergence and state-consistency conditions required by our theory. For the side-view setting, IRIS uses different balls released from the same height. This provides insufficient variation in release conditions for the three-slope divergence condition, while differences in ball appearance mean that the same physical position may correspond to different visual observations. For the overhead setting, although IRIS uses the same ball, the camera configuration varies across recordings, as illustrated in Fig. 12 (Page 51). Consequently, the same physical state may produce different visual observations, violating the shared state-to-image mapping assumption. We therefore recorded our own free-fall datasets, keeping the ball and camera configuration fixed within each view while varying the release height. This provides a controlled setting for evaluating our theoretical assumptions and complements our experiments on the public IRIS pendulum dataset.

Q2. What encoder is used? Is it the same as the encoder in the IRIS paper?

A. We agree that the encoder architecture and its provenance should be stated more explicitly. For the synthetic and real-pendulum experiments, we adopt the encoder architecture from LTI (Wang et al., ICML 2026), rather than the IRIS encoder, retaining the same network configuration except for replacing BatchNorm with GroupNorm. GroupNorm computes normalization statistics independently for each frame, avoiding dependence on other frames in the training batch and consistent with the framewise encoder assumed by the theory. For the side-view and overhead real free-fall experiments, we use similar CNN architectures with configurations adapted to their respective input resolutions. To improve reproducibility, we have added detailed encoder architecture tables in the appendices (Page 42, 52 and 56). We have also added the following clarification at the beginning of Section 5 (Page 7) of the revised manuscript:

The synthetic experiments and real-pendulum experiments use the CNN architecture of the LTI method, replacing BatchNorm with GroupNorm to avoid dependence on other frames in the training batch. Side-view and overhead free fall use similar CNNs with configurations adapted to their inputs. Implementation details are in Appendices M and N.

Q3.1 Comparison with existing baselines on additional public datasets.

A. Thank you for this suggestion. We have added pendulum-length comparisons on Delfys75 and IRIS in Appendix P.1 (Page 65). The tables directly include published results from LTI (Wang et al, ICML 2026) and statistics computed from released IRIS result, retaining their respective evaluation protocols. We additionally compare LTI and nonlinear models using matched training configuration. The nonlinear model reduces RMSE from 0.05830.0583, 0.07620.0762, and 0.10340.1034 m to 0.04950.0495, 0.05650.0565 and 0.07570.0757 m on Delfys75, and from 0.06460.0646 to 0.01500.0150 m on IRIS. These matched comparisons support the benefit of retaining the nonlinear pendulum restoring term. We also clarify the reference-length convention. Under the point-mass approximation, g/κg/\kappa represents the pivot-to-centre-of-mass distance, which may differ from the reported string length. The measurement endpoints are not explicit in the documentation examined, so this possible source of systematic offset remains unverified. We retain the published reference lengths and apply no post-hoc correction.

AuthorsOpenReview ↗

Author Response (Part 2/3)

W2. Ablation study with IRIS baselines: what enables accurate recovery?

A. Thank you for this question. To understand which factors contribute to the recovery differences between our method and the IRIS baselines, we conduct an ablation study along three dimensions: encoder architecture, training objective, and data conditions. The complete study is reported in Appendix P.2 (Page 66).

(1) Encoder architecture. We first examine whether recovery depends on our particular CNN. Replacing it with the IRIS MLP encoder, while retaining our nonlinear objective, preprocessing, and multi-angle fitting, yields a length estimate of 0.4717±0.06860.4717\pm0.0686 m and an RMSE of 0.06760.0676 m for the 0.500.50 m IRIS pendulum. Our CNN yields 0.5150±0.00120.5150\pm0.0012 m and an RMSE of 0.01500.0150 m. These results demonstrate compatibility with the IRIS encoder.

(2) Training objective. We retain our CNN, preprocessing, and multi-angle data, and adopt the IRIS finite-difference prediction losses and optimizer settings. With KL regularization, the Corrected and Multi-step variants yield 0.4361±0.03170.4361\pm0.0317 m and 0.4589±0.04520.4589\pm0.0452 m, respectively, for the 0.500.50 m pendulum. Replacing KL with our variance-floor regularizer gives 0.5138±0.04360.5138\pm0.0436 m and 0.5196±0.01480.5196\pm0.0148 m. Our original finite-difference ODE-residual loss with a variance floor yields 0.5150±0.00120.5150\pm0.0012 m (Appendix P.2, Table 29, Page 67). Considered alongside the original IRIS baseline results in Table 27 (Page 66), these results demonstrate that useful recovery is compatible with different dynamical losses and regularizers, although their choice still affects estimation accuracy.

(3) Data conditions. Finally, we examine the role of trajectory coverage. On individual IRIS recordings, the tested first-ten-frame segments do not provide three-slope coverage, and our method also fails under this restriction. Using 100 frames (10 s) provides richer state–velocity coverage and reduces our RMSE to 0.01750.0175, 0.02130.0213, and 0.02930.0293 m at 20∘20^\circ, 45∘45^\circ, and 90∘90^\circ, respectively (Table 30, Page 68). The adapted IRIS objectives also improve at 45∘45^\circ and 90∘90^\circ, while retaining large errors at 20∘20^\circ. One possible explanation is that weaker nonlinear excitation at the smaller angle increases sensitivity to the KL regularizer's preference for unit latent variance; this mechanism has not been independently verified.

Taken together, these ablations show that successful recovery is not specific to our encoder or loss, and support sufficient trajectory coverage as an important contributing factor. This is consistent with the theory's guidance on data collection, while the remaining performance differences show that coverage alone does not ensure state consistency or successful optimization.

Q3.2. How does the theoretical understanding guide practical parameter estimation?

A. We thank the reviewer for this question. Our theory provides practical guidance on what parameter information is identifiable and what data conditions support its recovery. As illustrated by the ablations in our W2 response, improving trajectory coverage can substantially improve adapted IRIS baseline estimates without changing their encoder architecture or training objective. Under the stated assumptions, the theory also distinguishes parameters identifiable directly from passive video from those requiring additional physical calibration. For the squared-velocity setting, Theorem 4.2 additionally provides a law-derived normalization that permits affine comparison in canonical coordinates. For further discussion, we respectfully refer the reviewer to our response to Reviewer raqQ, W2 & Q1. Together, these results guide the design and interpretation of practical estimators by specifying what data are needed, which quantities are identifiable, and when normalization or physical calibration is required.

W4. Could Section 5 explain the implementation details, experimental settings, and figures and tables more clearly?

A. We thank the reviewer for this helpful suggestion. We have expanded Section 5 to clarify the experimental settings and the interpretation of the figures and tables. The opening now describes joint encoder–dynamics fitting, explicitly references the training objective in Eq. (3), and identifies the encoder architecture. It also directs readers to Appendices M and N for architectures, data preparation, optimizers, and training settings, and to Appendix L for metric definitions. Throughout the section, we explain the comparisons and findings more explicitly and connect them to the corresponding theoretical predictions. These additions are highlighted in red in the revised manuscript.

AuthorsOpenReview ↗

Author Response (Part 3/3)

Q5. What are the take-home messages of Section 5, and how do the experiments support the theoretical predictions?

A. We thank the reviewer for this suggestion. The revised Section 5 makes the purpose and main finding of each experiment explicit:

(1) Coordinate ambiguity and parameter information (Section 5.1). Figure 1 tests the family-specific predictions of Section 3. The learned coordinates follow the predicted transformations; invariant quantities remain approximately stable, while scale-dependent coefficients follow their parameter orbits. These observations support the coordinate relations and identifiable parameter information summarized in Table 1.

(2) Velocity coverage (Section 5.2). Table 2 tests the rank condition in Section 4.1. Coordinate errors are similar across the tested designs, but parameter recovery improves substantially under full-rank velocity coverage. Accurate coordinate fitting alone therefore does not ensure accurate parameter recovery: the full-rank design separates the declared velocity features from coordinate curvature.

(3) Canonical normalization (Section 5.2). Table 3 tests Section 4.2 using the same fitted model before and after law-derived normalization. Normalization reduces coordinate error and improves recovery of the reference restoring force, supporting the predicted affine relation in canonical coordinates when raw affine alignment is insufficient.

(4) Real-video recovery (Section 5.3). Figure 2 and Table 4 examine these predictions on real pendulum and free-fall recordings. The observed invariants, parameter orbits, and normalized coordinate maps provide empirical support under real imaging conditions. The calibrated length and gravity estimates further illustrate how external anchors resolve the remaining physical ambiguity.

W5 & Q6. Will the code and real-world datasets be publicly available?

A. We thank the reviewer for this question. The supplementary material already includes data-generation and training code for all synthetic experiments, together with the identifiability tool described in our response to S1. Upon acceptance, we will publicly release the complete codebase and all self-recorded video datasets.

Q7. Could the main text discuss generalizability and possible extensions beyond the scalar setting?

A. Thank you for this suggestion. We have expanded the discussion in Section 6 to clarify the scope and possible extensions of our framework. Our results cover the stated scalar autonomous ODE families. In higher dimensions, finitely many smooth trajectories leave coordinate freedom outside the observed trajectories, so the scalar argument does not extend directly. We outline two possible directions: restricting the admissible coordinate transformations and considering trajectory ensembles with richer position–velocity coverage. Appendix K.2 provides the corresponding conditional vector extension and discusses the remaining challenges.

References.

  • Wang, Y., Wang, W., Zhang, K. and Gong, M. (2026). Physics from video: Identifiability of time-invariant second-order ODEs under minimal trajectory conditions. ICML.
Reviewer_ofyCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper extends the identifiability theory for physical parameter recovery from video for the underlying dynamics which are nonlinear scalar second-order ODEs. Most of its central results are either established consequences of classical differential equation equivalence theory or straightforward extensions of existing identifiability results. However, this paper has useful contributions. It brings these ideas together in a coherent framework for estimating physical parameters from videos, with explicit assumptions about trajectory coverage, learned coordinates, parameter ambiguity, and external calibration. Experiments on synthetic systems and real pendulum free fall videos support the predicted parameter relations.

Critical strengths

The paper's strengths lies in connecting existing results to a specific machine learning problem and demonstrating their practical consequences. The connection to video frame based observational data is no more than an encoder. Hence it can be potentially applied to other modalities for the same purpose.

Critical weaknesses

This paper is largely written as a system identification paper rather than machine leanring paper. The paper assumes the existence of one shared map \hat z(t)=f(z(t)). This is way too strong. Using the same neural network encoder across multiple videos may not achieve this. A common encoder may be affected by many factors such as appearance, illumination, occlusion, motion blur, or other visual appearance that has nothing to do with the scalar physical coordinate.

There seems no objective to the encoder other than the residual term fitting the ODE. In eq (3) only has a varianc floor regulariser to avoid near constant solution. It is not clear whether these conditions sufficiently constrain the learned representation to recover the physical state, up to the coordinate transformations proved by the theory. On the other hand, will any non-degenerate zero residual solution correspond to physically valid representations and identify parameter values perfectly? A flexible encoder may potentially map finite video observations to alternative trajectories satisfying the declared dynamics without recovering the underlying physical state. The paper acknowledges this limitation in Appendices and provide empirical results demonstrating successful recovery in selected settings. However, the disconnection between the pratice (optimisation of the encoder) and the theory is very obvious.

Questions for authors

How is the three-slope level coverage condition verified empirically on the real pendulum and free-fall videos? Specifically, what fraction of states in U satisfy the condition, and how is U chosen in practice?

5 comments
AuthorsOpenReview ↗

Author Response (Part 1/2)

We thank the reviewer for the positive assessment of our paper and for the thoughtful comments and questions. Below, we address each concern and question in turn.

W1. Is the shared-map assumption too strong under cross-video visual variations?

A. We thank the reviewer for this concern. State consistency requires the learned coordinate to depend on the physical state through one common map across clips. Sharing encoder weights alone does not establish this property. The theory states it as a premise, and the visual factors listed by the reviewer can interfere with it. Following the stress-test approach of LTI (Wang et al. ICML 2026), we use cubic Duffing videos to probe the practical limits of recovery under appearance, illumination, occlusion, and motion blur. Beyond the shared control settings, the three training clips receive different perturbations, with their configuration or severity varied across levels. Each setting uses five training seeds. Appendix P.3 (Pages 68–72) reports the settings and results.

(1) Appearance. Cross-clip changes in slider size and spring width degrade state and parameter recovery, highlighting the importance of consistent appearance for maintaining state consistency.

(2) Illumination. Coordinate recovery tolerates shared lighting (L1) and modest cross-clip differences (L2–L3), although parameter errors increase. Stronger differences (L4) substantially degrade both state and parameter recovery.

(3) Occlusion. Similarly, shared-position occlusion permits accurate recovery, whereas increasing positional differences across clips degrades recovery, culminating in severe failure at L4.

(4) Motion blur. Recovery remains robust to the tested randomized intermittent blur: median coordinate NRMSE stays below 22\\% at 2525\\%, 5050\\%, and 7575\\% blurred frames.

These experiments identify both successful and failing regimes. The estimator tolerates some cross-clip visual differences, but stronger discrepancies can compromise recovery, with sensitivity depending on the perturbation. They provide empirical evidence about the practical limits of recovery under visual variation; they do not establish state consistency from shared weights alone.

W2. (a) Is a zero value of the training objective sufficient to recover the physical state and parameters? (b) Does the identifiability theory inform the encoder that is actually optimised?

A. (a) No, and we agree with the reviewer. The residual sees the encoder only through the (state, velocity) pairs the videos visit. A non-affine coordinate change ff adds a curvature term f′′v2f''v^2 to the fitted law (Eq. (7)). Such warps are not artificial: a pendulum seen through the projection y=sin⁡qy=\sin q acquires the term −y (q′)2-y\,(q')^2 (Section 4.2). For a law linear in velocity, whether this term can be absorbed at the visited velocities depends on how many distinct velocities each state has, not on the loss. With one or two velocities it can be absorbed, which yields non-physical zero-residual solutions (Proposition E.1, Page 23). With three it cannot (Theorem 3.1). If the declared law has its own v2v^2 channel, that channel can absorb the term; Section 4.2 therefore compares such laws after canonical normalisation. Whether a zero objective suffices is therefore a property of the data, together with a state map shared across clips, and Eq. (3) encodes neither. The alternative trajectories the reviewer describes can arise only when these conditions fail. In practice there are further departures: sampling error, a non-zero residual, and a state map that is only approximately shared. Lemma H.2 (Page 28), Theorem H.3 (Page 28) and Appendix C.2 (Page 21) bound their effect. The learned coordinate stays within O(ε/δv2)O(\varepsilon/\delta_v^2) of an affine image of the physical state, where ε\varepsilon is the residual and δv\delta_v the separation between the velocities. The premises are audited post hoc (reply to W1). We do not claim that optimisation converges.

(b) Implications for training objectives and data collection. Identifiability characterises which parameter values are compatible with the observations under the declared model. When distinct parameters admit observationally compatible representations, minimising the fitting objective alone cannot determine which is physical. A regulariser may favor one solution, but that preference supplies an additional assumption. Under the shared-map, coverage, and regularity assumptions, Theorem H.3 (Pages 28–29) bounds the coordinate deviation from an affine map in terms of the residual and velocity-design conditioning, regardless of how the fitted solution was obtained. Our analysis therefore guides the data requirements, identifiable estimation targets, and necessary normalization or calibration used with a training procedure. It does not require a particular optimizer or imply that optimization converges.

AuthorsOpenReview ↗

Author Response (Part 2/2)

Continuation of our response to W2.

Two experiments illustrate this guidance. For quadratic drag, the theory requires a full-rank design that includes both directions of motion. In Table 2 (Page 9), coordinate errors remain approximately 11\\%, while the drag-coefficient error decreases from 44.9744.97\\% under one-way motion to 1.341.34\\% under the full-rank design. The IRIS ablations provide a complementary test (Appendix P.2, Pages 66–68). Keeping the implementation fixed within each method, extending individual recordings from ten to one hundred sampled frames substantially improves the adapted IRIS estimates at 45∘45^\circ and 90∘90^\circ. Our method also fails on the short inputs but achieves length RMSE below 0.030.03 m on the longer segments (Table 30, Page 68). These results demonstrate how informative data can improve existing estimators across different objectives, while the remaining failures show that coverage alone does not ensure successful optimization. The supplementary tool (Appendix O, Pages 63–65) makes the family-specific guidance accessible by reporting required velocity coverage, parameter invariants, and calibration needs. We have added the following clarification to the opening of Section 5 to emphasise that the theory guides data selection and calibration, rather than optimisation procedure:

The theory guides data requirements and calibration; the training objective alone does not guarantee recovery. A zero objective certifies neither latent-state consistency nor velocity coverage, and a zero sampled residual need not imply a zero continuous-time residual. Section 3.4 and Appendices C.2 and H.6 analyse approximation errors; optimisation convergence is not guaranteed. Its conclusions apply under the stated assumptions without prescribing a particular encoder architecture, dynamics loss, or regulariser; Appendix P.2 empirically examines alternative choices.

Q1. How is the three-slope level coverage condition verified empirically on the real cases, what fraction of states in UU satisfy the condition, and how is UU chosen in practice?

A. We thank the reviewer for requesting clarification. The required data conditions depend on the declared dynamical family. For the semilinear setting considered here, the condition is three distinct velocities at every state in a nonempty open interval UU, under the remaining assumptions. The theory prescribes no minimum interval width. Sampling frequency, discretization error, and representation error affect practical recovery, so these results do not provide a universal minimum useful interval width or an optimization convergence rate.

There are two ways to obtain the required velocity diversity. For damped oscillatory systems, a sufficiently long recording provides repeated passages through the same positions with different velocities; roughly three half-cycles provide a useful starting point, subject to checking the shared interval and velocity separation. Alternatively, trajectories with different initial conditions can supply distinct velocities over a shared position interval. For ideal free fall, three releases from rest at different heights provide distinct velocities over their common height interval. A candidate UU should therefore lie inside the overlapping position ranges, where the required passages are available and their velocities can be resolved. If this overlap or separation is insufficient, extend the recording or vary the initial conditions to obtain additional passages through the same states, while maintaining consistent visual conditions.

We illustrate this guidance through a single-video duration study on the synthetic cubic system (Appendix P.4, Figure 32, Page 73). We progressively extend the prefix of the same video while keeping the training configuration fixed. The required velocity diversity first occurs at one position at approximately 2.48s2.48\mathrm{s}. This isolated occurrence does not establish coverage over an open interval or guarantee accurate recovery. As the prefix grows and the interval supporting the required coverage expands, parameter estimates improve over 3.13.1–3.4s3.4\mathrm{s}, reach accurate recovery at approximately 3.5s3.5\mathrm{s} (less than 1.5T1.5T), and remain stable for longer prefixes, where TT is the first oscillation duration. This experiment illustrates the distinction between the theoretical coverage requirement and the observations needed for reliable recovery in practice. Although the theorem imposes no minimum width on the nonempty open interval UU, the combined effects of finite sampling, measurement noise, and representation error may require a sufficiently broad interval with well-separated velocities for accurate and stable recovery. The required extent is therefore problem-dependent.

Dear Reviewer ofyC,

Thank you again for taking the time to review our paper and for sharing your feedback. We sincerely appreciate the time and effort you have devoted to reviewing our work.

With less than 36 hours remaining in the discussion period, we would greatly appreciate the opportunity to further discuss our responses and revisions with you. We would particularly welcome any further feedback on whether our responses have addressed your concerns, or if there are any remaining questions that we could help clarify.

Thank you very much for your time and consideration!

Authors

Reviewer_raqQOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper extends identifiability theory for physical parameter estimation from video beyond the LTI setting of Wang et al. (2026) to nonlinear scalar second-order ODEs. Under a shared encoder, latent-state consistency, and velocity coverage assumptions, the authors characterize when learned coordinates must be affine, derive family-specific parameter invariants and orbits, and establish conditions under which canonical normalization handles squared-velocity terms. Experiments on synthetic and real videos support the predicted parameter relations.

Critical strengths
  1. Clean mathematical framework. The three-slope affine collapse argument (Theorem 3.1) is elegant. The proof that three roots of a degree-2 polynomial force all coefficients to zero is simple and effective.
  2. Empirical verification of coordinate ambiguity. Figure 1 directly confirms that learned coordinates are affine, with λ taking values different from 1 in some fits, and that invariant parameter ratios cluster near 1 while scale-dependent coefficients follow the predicted λ-scaling. This provides concrete visual evidence that the theory’s predictions hold in practice
Critical weaknesses

W1. Extremely narrow scope with limited generalizability. The theory covers only scalar, autonomous, second-order ODEs with prescribed velocity structures. This excludes coupled systems, systems with inputs, time-varying dynamics, and unknown equation forms, essentially all settings of broad practical interest. The extension from LTI to nonlinear is limited to a very specific class of ODEs: the core technique (polynomial root counting) is the same, and the nonlinear families studied (Duffing, Van der Pol, pendulum) are textbook examples. The paper does not provide a roadmap toward the multidimensional case beyond observing that it is hard

W2. Descriptive rather than constructive. The paper characterizes which parameters are identifiable but does not propose any mechanism to achieve or enforce identifiability. It does not introduce: • A loss function that promotes identifiability

• An encoder architecture that eliminates coordinate ambiguity

• A training procedure with convergence guarantees

• Any method to verify the core assumptions from data, specifically latent-state consistency (Definition 2.1), three-slope velocity coverage, zero ODE residual, and the encoder-class representability conditions in Proposition 3.3.

The practical takeaway is limited to “know what you can’t recover” and “collect diverse velocity data.” For the ICLR audience, a theory paper that neither enables new methods nor provides actionable algorithms has limited impact.

W3. Missing connection to observability theory. The paper does not engage with linear and nonlinear observability theory from control theory. Classical tools such as Lie derivative rank conditions and observability matrices provide established frameworks for determining whether states and parameters can be recovered from outputs(Khalil, 2002; Isidori, 1995). The relationship between the paper’s orbit-based analysis and these standard approaches is not explored

W4. Weak Experimental validation. The real-video experiments report only point estimates and error percentages, without orbit plots analogous to Figure 1. Showing that invariant quantities remain constant and scale-dependent ones follow predicted orbits across multiple fits would be far stronger evidence for identification than reporting low errors on individual estimates. The synthetic experiments validate the theory on the same simple systems used to develop it. The real-video experiments, while grounded, involve single-DOF systems where the scalar assumption holds trivially. The practical relevance of the framework for more complex physical systems is not demonstrated. Additionally, the reported errors (2–13%) in real-video parameter recovery are non-negligible and the paper provides no statistical characterization beyond repeated optimizer seeds and IC collections.

Questions for authors
  1. Can the framework be extended to produce a loss function or regularizer that enforces identifiability, rather than merely characterizing it?
  2. What is the relationship between the orbit structure derived here and Lie derivative methods used to construct the observability matrix for nonlinear systems (Khalil, 2002; Isidori, 1995)? Could these frameworks be bridged to provide a more principled identifiability analysis?
14 comments
AuthorsOpenReview ↗

Author Response (Part 1/7)

We thank the reviewer for the careful reading, and for recognizing the clean framework, the elegant affine-collapse argument, and the empirical evidence for the predicted parameter relations. Below, we clarify several potential misunderstandings and address each concern and question in turn.

W1. Extremely narrow scope with limited generalizability: scalar, autonomous, second-order ODEs with prescribed velocity structures; the extension from LTI uses the same technique (polynomial root counting) on textbook examples; no roadmap toward the multidimensional case.

A. We thank the reviewer for the comment.

What the theorems cover. The semilinear results hold for every scalar autonomous second-order law that is at most linear in velocity, z′′+aθ(z) z′+bθ(z)=0z'' + a_\theta(z)\, z' + b_\theta(z) = 0, with arbitrary continuous aθa_\theta and bθb_\theta. Theorem 4.1 extends affine alignment to any finite, dilation-stable space of continuous velocity features that excludes v2v^2, quadratic drag for example. Theorem 4.2 covers polynomial dependence on velocity including a v2v^2 term, with general continuous state-dependent coefficients. The pendulum, Duffing and Van der Pol laws illustrate these theorems; they do not delimit their scope. The identifiability tool in the supplementary material takes a declared scalar family and, within the supported theorem routes, reports its invariant parameter combinations, calibration requirements, and required velocity coverage. Its scope is not limited to the examples in Table 1. The screenshots in Appendix O (page 64) illustrate its use on Duffing, quadratic drag, and overhead free fall.

Why scalar, and why now. Our task is to recover physical parameter information from passive video when the governing family is known but the state coordinate is learned. The identifiability theory for this task is at the scalar level: the existing result is scalar LTI coefficient identification (Wang et al., 2026), and the real-video benchmarks are largely there as well (five of the eight IRIS systems are single-body laws; all five Delfys75 systems are single-coordinate laws). The nonlinear scalar case is the next step, and it is where the new question appears: affine alignment can now coexist with different physical coefficients, so one has to determine which quantities stay identifiable and what calibration resolves the rest. This paper answers that question for the classes above. The scalar nonlinear setting already covers a large share of quantitative video measurement: pendulum motion and swinging loads, drop and free-fall tests, vibration of a dominant mode in structures and machines (the displacement and bridge-vibration measurement work cited in Appendix B), nonlinear restoring forces modeled by Duffing laws, self-excited oscillation modeled by Van der Pol laws, sliding with friction, and motion observed through a nonlinear projection, such as toward or away from the camera, where the chain rule produces the v2v^2 term that Theorem 4.2 handles. The stakes are measurable: on real pendulum video with known gravity, the nonlinear model gives 2.92% median length error against 12.92% for the matched LTI model. Scalar nonlinear analysis changes physical inference in an existing video task.

The settings the reviewer lists. Systems with more than one configuration coordinate, coupled or not, meet the dimensional obstruction of Appendix K.1; see the last paragraph (Page 36). Known inputs add a known forcing term to the chain-rule identity in Eq.(7), with its own matching identity and its own coverage requirement, and the input regime needs its own analysis. Explicit time dependence can be declared in the family, since recorded time is preserved, but coverage must then hold at a common state and time across clips, a different requirement from ours. Unknown equation forms are the complementary problem of equation discovery, which IRIS evaluates separately from parameter estimation; selecting the family and identifying parameters within it are complementary questions, and Table 1 answers the second for each listed family, including what calibration recovers. Identifiability theory is exact: a result states which quantities are determined by which data under which coordinate freedom. No single theorem does this for all of these settings at once, because the admissible transformations and the data requirements differ in each. Any one of them, carried to the standard of the present results, is a good contribution of its own.

AuthorsOpenReview ↗

Author Response (Part 2/7)

Continuation of our response to W1.

What is new beyond LTI. In the LTI case, affine alignment ends the problem: the remaining affine freedom leaves every coefficient unchanged. In the nonlinear case it does not, and three things are needed for which the LTI analysis has no counterpart. First, the affine freedom now acts on the parameters, and the action depends on the family. Characterizing it required the translation lock (Lemma H.1, Page 26), the coefficient-weight rule (Proposition 3.5), the family-compatible set OU(θ)\mathcal{O}_U(\theta), and the proof that every point of that set is realized by an encoder (Proposition 3.3), which makes the criterion of Corollary 3.4 necessary and sufficient; the anchor-rank criterion (Proposition G.1, Page 25) then gives a local criterion for how many physical measurements remove the rest. Second, velocity dependence beyond linear needs a different argument: Theorem 4.1 separates the coordinate-curvature column from a dilation-stable feature library by a rank condition, and this is where the guarantee turns out to depend on the direction of motion and not only on the states visited (Table 2). Third, a v2v^2 term can absorb coordinate curvature, so raw affine alignment no longer follows; Theorem 4.2 compares laws after a law-derived normalization, and the real overhead-fall law produced by perspective projection (Appendix N.3, Page 60) is this case. The tools differ accordingly: Theorem 3.1 matches the coefficients of a degree-two polynomial in the velocity, which three velocities fix; Theorem 4.1 is a rank condition on a feature library; Theorem 4.2 first computes the law-derived normalization and then matches coefficients of a degree-max⁡{2,p}\max\{2,p\} polynomial, which needs correspondingly more velocities. We chose the pendulum, Duffing and Van der Pol laws deliberately: each is simple and standard, and each exhibits one of the structures above (a non-polynomial restoring force; polynomial forces whose higher-order coefficients depend on the latent scale; nonlinear state-dependent damping), so every predicted transformation, invariant and anchor requirement can be checked by hand and then tested on learned video representations. They illustrate the theory; its scope is the classes in the first paragraph.

AuthorsOpenReview ↗

Author Response (Part 3/7)

Continuation of our response to W1.

Roadmap to d>1d > 1. The revised Section 6 outlines the roadmap and Appendix K.2 (Page 36) gives the conditional derivation; we summarize it here. Consider physical and fitted vector semilinear laws:

z′′+Aθ(z),z′+bθ(z)=0,\mathbf{z}'' + A_{\theta}(\mathbf{z})\\,\mathbf{z}' + \mathbf{b}_{\theta}(\mathbf{z}) = 0,

z^′′+Aη(z^),z^′+bη(z^)=0,\hat{\mathbf{z}}'' + A_{\eta}(\hat{\mathbf{z}})\\,\hat{\mathbf{z}}' + \mathbf{b}_{\eta}(\hat{\mathbf{z}}) = 0,

where z∈Rd\mathbf{z} \in \mathbb{R}^d is the physical configuration (bold for the vector case; zz is the scalar case), ′' is the time derivative, AA is matrix-valued and b\mathbf{b} is vector-valued, related by a shared C2C^2 diffeomorphism z^=f(z)\hat{\mathbf{z}} = f(\mathbf{z}) on a connected open region UU. Write v=z′\mathbf{v} = \mathbf{z}' for the velocity, Df(z)Df(\mathbf{z}) for the Jacobian of ff (the d×dd \times d matrix of first partial derivatives), and D2f(z)[v,v]D^2 f(\mathbf{z})[\mathbf{v}, \mathbf{v}] for the Hessian of ff evaluated on the velocity twice, the vector with kk-th component ∑i,j∂2fk/∂zi∂zj vivj\sum_{i,j} \partial^2 f_k / \partial z_i \partial z_j \, v_i v_j; these play the roles of f′f' and of f′′(z)(z′)2f''(z)(z')^2 in the derivation of (7). Differentiating z^=f(z)\hat{\mathbf{z}} = f(\mathbf{z}) twice and eliminating z′′\mathbf{z}'' with the physical law gives the compatibility identity

0=D2f(z)[v,v]+[Aη(f(z)),Df(z)−Df(z),Aθ(z)]v+bη(f(z))−Df(z),bθ(z),0 = D^2 f(\mathbf{z})[\mathbf{v}, \mathbf{v}] + \bigl[A_{\eta}(f(\mathbf{z}))\\,Df(\mathbf{z}) - Df(\mathbf{z})\\,A_{\theta}(\mathbf{z})\bigr]\mathbf{v} + \mathbf{b}_{\eta}(f(\mathbf{z})) - Df(\mathbf{z})\\,\mathbf{b}_{\theta}(\mathbf{z}),

the vector counterpart of (7): constant, linear and quadratic terms in the velocity, with the quadratic coefficients determined by the Hessian. A full-column-rank same-position velocity design for 1,vi,vivj1, v_i, v_i v_j (i≤ji \le j) at every z∈U\mathbf{z} \in U would therefore force D2f=0D^2 f = 0, hence f(z)=Lz+cf(\mathbf{z}) = L\mathbf{z} + \mathbf{c} with LL invertible, exactly as three distinct velocities force f′′=0f'' = 0 in the scalar case. This identifies the coverage premise for a direct extension. Proposition K.1 (Page 36) explains why passive clips cannot supply this premise when d>1d > 1: finitely many trajectories are curves in Rd\mathbb{R}^d, their union has empty interior, so there is no open region UU on which the velocity design can hold, and adding clips does not change this. Moreover, a non-affine coordinate change can agree with the identity map, together with its first and second derivatives, along all observed curves, so positions, velocities and accelerations on the data cannot detect it. The scalar case is different: a single nonconstant trajectory already sweeps an interval, so an open UU is available and only the same-position velocity variation of Theorem 3.1 has to be supplied. What blocks the extension is thus the absence of an open covered region, not the algebra, which goes through unchanged. Coupling is already allowed in the compatibility identity through AA and b\mathbf{b} and plays no role in this step; it matters for parameter recovery, which requires analysis of the remaining matrix-valued compatibility equations. Two directions follow: restricted coordinate classes, such as finite-dimensional parameterizations or coordinate-wise maps, that finite trajectories may constrain; and ideal ensembles of initial positions and velocities that provide the required coverage, with finite-data approximations under additional assumptions. Each direction requires its own coverage and family-compatibility analysis; we state them as directions, not results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.