acceptodds
Under review as a conference paper at ICLR 2027

SPEC: Exact Structural Constraints Are Cheap, Specifying Them Is Not

Abstract

Exactness is the defining property of structure-preserving dynamics models, yet the specification it rests on, which law to impose, under which metric, and in which coordinates, is never itself tested, because a prediction loss cannot tell a correct specification from a wrong one. To address this gap, we present SPEC, a staged framework that makes the specification computable from trajectory data and applies an exact constraint only where the data support one. Its foundation is a per-step identity, , which reduces the structural question to the sign of one symmetric matrix and yields, under the Euclidean block metric, a library of five exact constructions together with the boundaries beyond which none exists. Three conditional stages then decide what the identity leaves open, separating a statistical decision taken on finite trajectories from a guarantee that holds at every parameter value. Experiments on analytically labelled systems isolate what each stage contributes. Discover recovers coordinates without a reconstruction weight, at median invariant of to against to for a comparison objective. Refusing unsupported blocks costs more out-of-family error than an oracle router, and enforcing every block costs . Over the fitting horizon, exact enforcement matches a penalty of equal capacity in modal-energy error while holding the selected latent law to where the penalty reaches only . Accuracy is therefore limited by neither stage, but by what sits behind them: the metric the construction is instantiated in, for which we give the exact form in the damped case, and the decision rule, which measures agreement inside the selection window rather than validity beyond it.

Primary area:
learning on time series and dynamical systems
Keywords:
structure-preserving learning, dynamical systems, exact constraint enforcement, constraint specification, coordinate discovery, Cayley transform, model selection, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ZXvNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework (i.e., SPEC) for identifying and enforcing structural constraints in learned dynamical systems. The main motivation is that existing structure-preserving models usually assume the correct physical laws and coordinate representations are already known, while imposing an incorrect constraint may hurt prediction accuracy. Speificially, the proposed SPEC seems to have three stages: (1) Discover, which learns latent coordinates where structural laws can be expressed. (2) Diagnose, which selects a suitable constraint or rejects it based on trajectory data. and (3) Enforce, which applies the selected constraint through an exact parameterization based on the Cayley transform. Experiments on conducted on several analytical dynamical systems. The results show that exact constraints can be maintained at numerical precision, but the performance is still limited by inaccurate coordinate discovery, metric selection, and constraint diagnosis.

Critical strengths
  1. The considered problem is often overlooked in structure-preserving learning literature. I think the main motivation is interesting. Existing methods usually focus on how to enforce a certain physical constraint, while less attention is paid to whether this constraint is actually appropriate for the system. This things actually happens a lot in practice. People's understanding of the instanteous system could be inaccurate, due to outdated data, ever-evovling system dynamics. SPEC tries to address this issue by introducing a diagnosis stage before enforcement. It could be useful tool/idea for a lot of such works.

  2. The theoretical analysis seems solid, and the authors provide a good discussion of the method's limitations. The paper gives mathematical guarantees for different radial constraints and shows that they can be enforced up to numerical precision. I also appreciate that the authors explicitly investigate cases where exact enforcement may not improve prediction accuracy. This makes the analysis more convincing than only reporting successful cases.

Critical weaknesses
  1. The effectiveness of the complete SPEC framework is not sufficiently validated. My main concern is that most experiments evaluate the three stages separately, rather than testing the complete framework from learned coordinates to final trajectory prediction. In particular, Q3 and Q4 use analytical modal coordinates instead of those learned by Discover. However, Table 3 shows that Discover performs quite poorly on the damped-system benchmark, with only 2.4% of evaluations reaching the target score, compared to 86.2% using reference modal coordinates. The real-data experiment in Table 15 also shows worse results than the comparison objective. These results make me question whether the three stages can actually work together on more complicated systems.

  2. The practical benefit of exact enforcement is still unclear. Although SPEC satisfies the selected constraints exactly, this does not necessarily lead to better prediction. In Table 5, the penalty-based method achieves lower prediction error than the exact construction at every evaluated horizon. For example, at 100 times the fitting horizon, the errors are 0.206 and 0.377, respectively. In addition, the guarantees are defined in the learned latent space, which may not directly translate into physical constraints in the original system. Therefore, I am not fully convinced that the proposed exact enforcement provides a clear advantage over existing approximate or unconstrained methods.

  3. The metric selection problem is not actually solved by the proposed framework. One of the main motivations of this paper is to determine which structural constraint should be applied, under which metric, and in which coordinates. However, SPEC seems to only learn the coordinates and select the constraint, while the metric is fixed to the Euclidean metric (M=I). Although Proposition 7 derives a different metric for damped systems, this metric is obtained analytically and is not incorporated into the proposed learning framework. This raises a concern about the completeness of SPEC, since an incorrect metric may lead to incorrect constraint selection even if Discover and Diagnose work as expected. I think this limitation is quite important given that metric specification is presented as one of the central problems addressed by this paper.

  4. The Diagnose stage may make incorrect decisions under noise or slightly misspecified dynamics. The reliability of Diagnose is another concern. As shown in Figure 4(b), when the rate ratio reaches (\kappa=5), enforcing the selected constraint already becomes harmful, but 89% of these blocks are still accepted. On the other hand, Figure 5(c) shows that under observation noise (\sigma=0.3), only 0.9% of correctly specified blocks are accepted, even though enforcing them would still be beneficial. These results suggest that the fixed threshold does not reliably distinguish when enforcement is helpful. Although the authors acknowledge this limitation, they do not provide an effective solution. Since selective enforcement is the main advantage of SPEC, I think the robustness of this decision mechanism needs further justification.

Questions for authors
  1. Can the authors provide an end-to-end evaluation of SPEC? Currently, the diagnosis and enforcement experiments mostly rely on analytical coordinates. I would like to see the performance when all three stages are combined, especially compared with unconstrained and penalty-based models on the same systems.

  2. How does Discover handle structural laws with time-varying decay rates? The objective in Eq. (10) minimizes the temporal variance of the log-radius rate, which seems more directly suited to conservation or exponential decay. I wonder whether this objective introduces a bias when discovering coordinates for algebraic decay or attracting-sphere dynamics.

  3. Does SPEC actually identify the appropriate metric from data? The paper emphasizes that choosing the correct metric is an important part of constraint specification, but the implemented framework uses the Euclidean metric. The alternative metric for damped dynamics is derived analytically rather than learned. Could the authors clarify how this part of the specification is addressed by the proposed framework?

7 comments
AuthorsOpenReview ↗

Response to Reviewer ZXvN (1/4): Overview, and W1/Q1 on Table 3 and the Discover objective

We thank the reviewer for observing that the structure-preserving literature often overlooks whether a constraint is appropriate, and that a diagnosis stage before enforcement can benefit many works (S1). We also thank the reviewer for recognising the theoretical analysis and the explicit study of cases in which exactness does not improve accuracy (S2). Our response has four parts. Part 1 traces the Table 3 gap of Discover on the damped grid to an encoder offset and to optimisation from a random start, shows that the minimiser of the objective is the correct one, and closes the gap (W1, Q1). Part 2 runs the paper's own benchmarks, Tables 4 and 5, end to end with coordinates learned by Discover, extends the pipeline to the other radial laws and answers Q2 (W1, Q1, Q2). Part 3 shows why the penalty has the lower error in Table 5 and that exactness costs nothing once the rate and the metric are right (W2). Part 4 estimates the metric from data and adds a decision rule that separates the two failure modes of Diagnose (W3, Q3, W4). In short, end to end on the Table 5 grid with learned coordinates, the exact construction reaches a truth error of 0.038 at one hundred times the fitting interval, against 0.377 on the analytical coordinates of Table 5, while holding its law to 7×10−137 \times 10^{-13}.

W1 and Q1. End-to-end evaluation. We agree that the complete pipeline deserves a direct evaluation, and we report it below and in Part 2.

(i) Q3 and Q4 run on analytical modal coordinates so that each error is attributable to a single stage (Table 11 caption). This attribution is what located the bottleneck: on the damped grid, the learned coordinates fall more than an order of magnitude short of the modal reference (§4.2, Table 3).

(ii) The paper attributes this gap to the weaker encoder class and shorter budget of the grid (§4.2). One concrete mechanism behind it is a degenerate direction of the objective that §3.3 names: the rate term is driven towards zero by a constant offset in zz, which the whitening term does not exclude. The linear encoder has no bias, but it receives states normalised by the window mean, so it maps the equilibrium x=0x = 0 to a nonzero point. This effective offset accounts for a median of 16% of the block radius, and 52% at N=40N = 40. A large offset drives the block log rate towards zero. This is harmless for conservation, whose rate is zero and whose invariant sits in the per-trajectory level (§4.2), but it masks the rate of a decaying mode. Every exact construction satisfies U(0)=0U(0) = 0, so the equilibrium must map to the latent origin, and anchoring the encoder there, z=W(x−xeq)z = W(x - x_{\mathrm{eq}}), removes the offset and changes nothing else. On the same grid and seeds, anchoring raises the share of evaluations whose held-out score Rlog⁡2R^2_{\log} (Eq. 28) exceeds 0.8 from 2.2% in our rerun with three of the ten seeds (2.4% in Table 3) to 22.8%, and the median held-out Rlog⁡2R^2_{\log} at ten characteristic decay times from −3.30-3.30 to 0.840.84. Anchoring differs from the centring of Appendix F.3, which subtracts the latent mean and removes the per-trajectory level carrying the invariant.

(iii) The remaining gap lies in the optimisation rather than in the objective. At analytical modal coordinates rescaled by the metric of Proposition 7, the rate term of Eq. 10 is zero to rounding (4×10−304 \times 10^{-30}), and in every cell of the grid the loss there is lower than at the encoders learned from a random start (median 2.7×10−52.7 \times 10^{-5} against 0.50). For a linear encoder on linear dynamics this minimiser has a closed form, because each complex eigenvector pair of the one-step map gives a block whose radius decays exactly, and exact DMD computes these pairs (Schmid, JFM 2010, Tu et al., Journal of Computational Dynamics 2014). Initialising Discover at this minimiser raises the share of evaluations above 0.8 to 98.6%, above the 86.5% of the Euclidean modal reference in the same rerun (86.2% in Table 3), with a median held-out Rlog⁡2R^2_{\log} of 1.000 at every horizon. If training continues from there, the whitening term trades against the rate term, and returning the iterate with the lowest total loss, a rule that uses no validation data, gives 83.8%, still close to the modal reference against 2.2% from a random start. Within the linear encoder class of the grid, the gap in Table 3 is therefore one of optimisation. The closed form exists only for a linear encoder, so on nonlinear systems Discover is trained from a random start, as below and in Part 2 (vi). We will describe this initialisation in §3.3 and report the grid with it in §4.2, updating the explanation given there. On the analytic systems, with a three-layer encoder trained from a random start, Discover reaches a median invariant R2R^2 (the R2R^2 with which the block radius predicts the true invariant) of 0.728, 0.767 and 0.858, with seed distributions disjoint from those of the comparison objective (Table 2, Fig. 3d).

AuthorsOpenReview ↗

Response to Reviewer ZXvN (2/4): W1/Q1 continued, Tables 4, 5 and 15 end to end, and Q2

(iv) With these coordinates we ran the paper's own enforcement and routing benchmarks end to end on all 16 configurations of the Table 5 grid and all 6 of the Table 4 benchmark, three seeds each. With one block per mode, the learned blocks match single modes (median R2R^2 0.99986 against the modal radius), and Diagnose accepts every block in the median cell. The truth error at 100Tfit100 T_{\mathrm{fit}}, one hundred times the fitting interval TfitT_{\mathrm{fit}}, is then 0.038 and 0.040 for the exact construction in the Euclidean and the learned metric, against 0.377 on analytical coordinates (Table 5) and 11.7 for one aggregate law, while the exact law is held to 6.9×10−136.9 \times 10^{-13} against 6.4×10−46.4 \times 10^{-4} for the penalty. The learned coordinates outperform the analytical ones because each learned block is, or is close to, an eigenvector pair in which the Euclidean radius decays exactly, so the coordinates absorb the metric of Proposition 7. On the same blocks the per-block penalty and the per-block least-squares map reach comparable error (0.038 and 0.033), and an unconstrained least-squares map on the full latent reaches 0.996. On the Table 4 benchmark, whose components we mix by a random orthogonal map so that the coordinates must be learned, Diagnose makes the decision of the oracle router on 265 of 270 blocks, accepting 134 of 135 in-family blocks and refusing 131 of 135 out-of-family blocks. Refusal protects the out-of-family blocks, whose error at 10×10\times is 0.24 against 1.63 when every block is enforced.

(v) On the NL-A3 orbit the comparison objective has the higher median at every reconstruction weight (Fig. 8c), while the block radius stays flatter under the rate objective (Fig. 8b). At λrec=0\lambda_{\mathrm{rec}} = 0 the seed distributions overlap (Table 15), and a rank test over the ten seeds finds no difference (two-sided Mann-Whitney p=0.25p = 0.25, paired Wilcoxon p=0.70p = 0.70), whereas on each analytic system at the same setting the rate objective is higher on all ten seeds (p=1.1×10−5p = 1.1 \times 10^{-5}). The 20 segments are consecutive windows of one orbit and hence dependent, the standard caution for time-ordered data (Appendix F.4, Bergmeir and Benítez, Information Sciences 2012). They also lie on essentially one orbit, with a relative standard deviation of the orbital energy of 3.4×10−33.4 \times 10^{-3}, so the invariant R2R^2 is dominated by residual fluctuations, the classical effect of range restriction on correlation (Sackett and Yang, Journal of Applied Psychology 2000). Table 15 therefore cannot rank the two objectives, whereas the analytic systems do.

(vi) For the laws that Diagnose names or imposes, we also ran all three stages from raw observations, including planar systems observed through a fixed nonlinear map and the Van der Pol oscillator, whose cycle is not circular in raw coordinates, against the per-block penalty and an unconstrained block map on the same learned blocks. At 100Tfit100 T_{\mathrm{fit}} the log-radius error for algebraic decay is 0.23 to 0.25, against 35 to 63 for the penalty and 41 to 70 for the unconstrained map. For the attracting sphere it is at most 1.2×10−31.2 \times 10^{-3}, with the radius ending within 6×10−146 \times 10^{-14} of r∗r_{\ast}, against 2.8 to 6.2 and 4.9 to 6.5. For the forward-invariant region on the pendulum and Kepler it is 0.327 and 0.0028, against 0.425 and 0.0428 for the penalty and 36.1 and 4.18 for the unconstrained map, and no rollout leaves the ball, whereas the unconstrained map leaves it in 12.5% to 100% of rollouts. We will add these end-to-end results to §4.

Q2. Laws with a time-varying rate in Discover. The rate term of Eq. 10 is the least-squares residual of the exponential family fitted to the block log rate, so it favours coordinates in which a law appears exponential. This does not exclude the algebraic law: by Eq. 8, rk−2p=r0−2p+2pckΔtr_k^{-2p} = r_0^{-2p} + 2pck \Delta t, so the increasing map s=exp⁡(−r−2p/(2pc))s = \exp(-r^{-2p}/(2pc)) has log⁡sk=−r0−2p/(2pc)−kΔt\log s_k = -r_0^{-2p}/(2pc) - k \Delta t, a constant log rate, and exponential decay of ss is exactly algebraic decay of rr. Which name is assigned depends on the coordinates, but the prediction does not. With the encoder anchored at the equilibrium as in Part 1 (ii), Diagnose accepts the algebraic law on 17 of the 21 best learned blocks of seven algebraic systems, and the exponential law forced in its place is two orders of magnitude worse at 100Tfit100 T_{\mathrm{fit}} (30 to 59 against 0.23 to 0.25). For the attracting sphere the native log rate −γ0(r2−r∗2)-\gamma_0(r^2 - r_{\ast}^2) vanishes on the attractor, where every sphere law coincides with conservation at r∗r_{\ast}, so the rate is identifiable only from transients and the sphere candidate requires r∗r_{\ast} (§3.3). With transients present, the sphere law is accepted on all 12 learned Stuart-Landau blocks (r˙=νr−r3\dot r = \nu r - r^3) at ν=0.05\nu = 0.05, on 2 of the 3 best blocks at ν=0.1\nu = 0.1 and on 10 of the 12 Van der Pol blocks. We will report these results in §4.3.

AuthorsOpenReview ↗

Response to Reviewer ZXvN (3/4): W2, the benefit of exact enforcement

W2. The benefit of exact enforcement. Table 5 was designed around this question, and the new experiments below sharpen its answer. Its answer, stated as contribution (iii), is that exactness holds the selected law at no cost in error over the fitting horizon, while long-horizon accuracy is set by rate estimation and the metric (lines 102 to 105).

(i) Exactness guarantees adherence to the selected law. The exact construction holds the law to 9.4×10−159.4 \times 10^{-15} at TfitT_{\mathrm{fit}} and 9.1×10−139.1 \times 10^{-13} at 100Tfit100 T_{\mathrm{fit}}, whereas the penalty deviates by 1.5×10−31.5 \times 10^{-3} and 1.5×10−11.5 \times 10^{-1} (Table 5, Fig. 9b). This holds for every value of the trainable parameters, because the radius of an accepted block is independent of them (Proposition 5). Soft constraints are satisfied only approximately and can complicate optimisation (Krishnapriyan et al., NeurIPS 2021), hard constraints hold by construction (Négiar et al., ICLR 2023), and guaranteeing a property by construction is the principle behind stable learned dynamics (Kolter and Manek, NeurIPS 2019).

(ii) The lower truth error of the penalty in Table 5 traces back to two factors identified in the paper, rather than to approximate enforcement itself. All arms receive the same fitted rates, and the penalty radius follows a single exponential at a rate that differs from the fitted one by 1.4×10−41.4 \times 10^{-4} in relative terms (lines 462 to 464). The fitted rate is in turn limited by the metric: under the Euclidean metric the radius of a damped mode oscillates about an exponential with log half-amplitude 12artanh⁡(γ/ω)\frac{1}{2}\operatorname{artanh}(\gamma/\omega), which no single exponential can follow (Proposition 7, lines 465 to 471). To isolate these factors, we reran the 80 cells of Table 5. Both rate estimates are accurate (median relative error 2.7×10−42.7 \times 10^{-4}), but the effective rate of the penalty lies closer to the truth for 74% of the modes, and over 100Tfit100 T_{\mathrm{fit}} this small difference produces the gap. Given the true rates, the exact construction has a truth error (the log modal-energy error of Table 5) no larger than any penalty weight at every horizon, and it is the only arm whose error does not grow: 0.0208 at TfitT_{\mathrm{fit}} and 0.0205 at 100Tfit100 T_{\mathrm{fit}}, against 0.0603 for the penalty at β=1\beta = 1 (0.134 at β=0.1\beta = 0.1, 0.027 at β=10\beta = 10). The remaining 0.0205 is the Euclidean ripple of Proposition 7. With the metric estimated from data (Part 4), the exact construction reaches 1.1×10−121.1 \times 10^{-12} at 100Tfit100 T_{\mathrm{fit}}, the level of the true metric (9×10−139 \times 10^{-13}), because on these noise-free trajectories the log radius is exactly linear in the correct metric and the fitted rate becomes exact, against 0.377 in the Euclidean metric and 0.206 for the penalty.

(iii) The penalty weight shows this directly. With the fitted rates of Table 5, raising β\beta from 0.1 to 1 to 10 lowers the deviation of the penalty from its law at 100Tfit100 T_{\mathrm{fit}} from 0.26 to 0.15 to 0.041, and raises its truth error from 0.128 to 0.206 to 0.343, towards the 0.377 of the exact construction. The more faithfully the law is held, the more of the fitted-rate error is passed on, while the exact construction holds the law to 9.1×10−139.1 \times 10^{-13} in every setting. Once the rate and the metric are specified correctly, exactness costs nothing in accuracy and provides a guarantee that no penalty weight provides.

(iv) Against an unconstrained map, the benefit is largest beyond the fitting interval for the nonlinear laws, where the pipeline of Part 2 extrapolates more than two orders of magnitude better and keeps every region rollout inside its ball. The largest gain in §4.4 comes from assigning one law per block, which is what Diagnose provides: one law per mode reduces truth error at 100Tfit100 T_{\mathrm{fit}} by a factor of 29.4 relative to one aggregate law (Table 5, Fig. 6). As for physical meaning, the guarantee concerns the encoded observable Ci(x)C_i(x), the radius of block ii of Φϕ(x)\Phi_\phi(x), which is a function of the physical state (Appendix C.9). Its agreement with the physical invariant is what the invariant R2R^2 measures, and on the algebraic systems the learned radius is nearly monotone in the true radius on every seed (Spearman correlation at least 0.996). The guarantee is thus exact for the encoded radius and carries over to the physical radius to the accuracy of the encoder, which the invariant R2R^2 and this rank correlation report. We will state (i) to (iv) in §4.4.

Reviewer_TGZJOpenReview ↗
Rating
3(+2)
Confidence
1
Informativeness
2
Presentation
2(+1)
Soundness
3(+1)
Read review
Paper summary

The paper tackles an interesting question. In my understanding, the paper introduces SPEC, a training approach for identifying structured dynamical systems from data and trying to identify the coordinates under which some structure holds (e.g., contraction, conservation, algebraic decay, and others). The authors leverage the Cayley transform to parametrize all orthogonal matrices and perform identification in these radial dynamics and are able to capture many different structures in one shot. The authors run some experiments to validate the theoretical results.

Critical strengths
  • The problem under consideration is interesting and the characterization of many different structures in a unified way in equation (8) is quite interesting
Critical weaknesses
  • The paper is not clearly written at all. I understand that Theorem 1 is the key result that allows us to rewrite many different structures in one unified form in equation (8), but the actual role of the different matrices is not clear at all. More generally throughout the paper, there are lots of variables or formulas that are introduced but are not standard (at least not to me) which makes reading the paper extremely challenging.

  • The experimental section is not written in a coherent manner. By reading the main text alone, I do not understand what experiments were run, on what datasets, and what the main takeaways are. I strongly encourage the authors to re-write this section to clearly articulate the results from the paper by including more details from the appendix. I do not believe Q1 is informative as this is an algebraic relationship. Q2-Q4 need more details.

Questions for authors
  • Several properties, e.g., contraction, is invariant to diffeomorphisms. If you collect time-series data from contracting dynamics, what is the concrete advantage of using this congruence? Can the congruence represent arbitrary diffeomorphisms of the data?

  • Why do you call the relationship in Theorem 1 an "exact budget congruence?" Congruence I understand, but why budget? What is the budget here exactly?

  • Why do you defer the notation to the appendix? The paper does not read in a self-contained manner this way.

  • What is the meaning of the subscript c in the second term of the loss function (10)? Same with the subscripts kl in the last term.

  • In the paragraph following equation (8), what is η\eta? In the "sph" case in equation (8), μ\mu is not defined.

  • What is the "BLAS reduction order?"

  • In general, the paper is hard to read. There are some aspects that are not explained clearly at all and make it very challenging for me to understand what is being said. At the same time, there are some sections that are overly full of new symbols and terminology that is very dense and also makes it challenging to understand what is being said. My overall comment is that the authors should please re-write for clarity and explain as simply as possible (and in a self-contained manner) what the method is doing.

6 comments
AuthorsOpenReview ↗

Response to Reviewer TGZJ (1/2): Overview, W1 on the method and Theorem 1, and W2 on the experiments

We thank the reviewer for finding the problem interesting and for singling out the unified form of Eq. 8. The comments show clearly where the presentation was too dense. We hope that this response helps the reviewer gain a fuller understanding of our work. Our response has two parts. Part 1 restates the problem and the method on one example, explains each matrix of Theorem 1 (W1), and summarises every experiment (W2). Part 2 explains what the congruence adds when contraction is coordinate free (Q1) and answers the questions on terminology and notation (Q2 to Q7).

W1. Clarity, and the role of the matrices in Theorem 1. We agree that the method section was too dense, and restate it below in the order the revision will follow.

(i) The problem and the contribution. A structure-preserving model builds a physical law into its update, such as conservation or decay of an energy, so that the law holds for every value of the trainable parameters. This stabilises long rollouts only if the law is correct: a model that conserves energy, applied to a dissipative system, conserves it perfectly and predicts poorly (§1, Lee et al., NeurIPS 2021, Gruver et al., ICLR 2022). Which law to impose, under which energy and in which coordinates is usually fixed by hand, and a prediction loss cannot test this choice, which the paper calls the specification. Theorem 1 characterises, for a linear one-step update in Cayley form and any quadratic energy, whether that energy is conserved, decreases or increases, and yields five exact constructions with a prescribed law. SPEC selects from data the law that each part of the state supports, or refuses, and builds the selected law in exactly.

(ii) The method on one example. For a damped oscillator observed through a sensor, Discover learns latent blocks ziz_i whose length ri=∥zi∥r_i = \lVert z_i \rVert evolves simply. Diagnose fits candidate laws for this length on one part of the data, scores them on a later part, and accepts the best one or refuses. Enforce builds an update that rotates each block and rescales its length so that the selected law, here exponential decay, holds exactly.

(iii) Theorem 1. The update is U=Cay(A)=(I−A/2)−1(I+A/2)U = \mathrm{Cay}(A) = (I - A/2)^{-1}(I + A/2) with generator AA. The symmetric matrix MM selects the energy x⊤Mxx^\top M x, and M=IM = I gives the squared length. The matrix ΣM(A)=A⊤M+MA\Sigma_M(A) = A^\top M + M A is the rate of change of that energy under x˙=Ax\dot x = A x, since ddt(x⊤Mx)=x⊤ΣM(A)x\frac{d}{dt}(x^\top M x) = x^\top \Sigma_M(A) x. Finally, w=(I−A/2)−1xw = (I - A/2)^{-1} x is the midpoint of the step, because Ux=(I+A/2)wUx = (I + A/2) w and x=(I−A/2)wx = (I - A/2) w give w=(x+Ux)/2w = (x + Ux)/2. The theorem states that the change of energy over one step equals its continuous-time rate at this midpoint, with the step size absorbed into AA, (Ux)⊤M(Ux)−x⊤Mx=w⊤ΣM(A)w.(Ux)^\top M (Ux) - x^\top M x = w^\top \Sigma_M(A) w. The energy is therefore conserved exactly when ΣM(A)=0\Sigma_M(A) = 0, never increases when ΣM(A)\Sigma_M(A) is negative semidefinite, and never decreases when it is positive semidefinite. This is the mechanism by which the midpoint rule preserves quadratic invariants (Cooper, IMA J. Numer. Anal. 1987, Hairer et al., Geometric Numerical Integration, 2006), extended here to decay and growth.

(iv) Why Eq. 8 has one form. With M=IM = I and a skew-symmetric generator KiK_i, ΣI(Ki)=0\Sigma_I(K_i) = 0, so by Theorem 1 the rotation Qi=Cay(Ki)Q_i = \mathrm{Cay}(K_i) preserves the length of the block exactly. Each block is rotated by QiQ_i and scaled by a factor that depends only on its length (Eqs. 6 and 7), so the length follows r+=e−γ(r)Δtrr^{+} = e^{-\gamma(r) \Delta t} r, and the five lines of Eq. 8 are five choices of the rate function γ(r)\gamma(r). We will open §1 with (i) and (ii) and state Theorem 1 in §3.2 in the form of (iii), with each matrix defined beside it.

W2. The experimental section. We agree, and will restructure §4 so that each experiment states its question, data, comparison and result, as summarised below (§4, Table 11, Appendix F.1). We also agree that the first experiment (Q1) is not informative in the main text, since the identity is algebraic. It verifies the implementation in float32 and float64 (Table 1), and we will move it to the appendix.

ExperimentQuestionData and coordinatesComparedResult
DiscoverDoes Discover find coordinates in which a law holds, without a reconstruction weight?Pendulum, Lotka-Volterra, Kepler, learned encoderOur objective vs. a comparison objectiveMedian invariant R2R^2 0.728 to 0.858 vs. 0.095 to 0.219 (Table 2)
DiagnoseWhat does refusing unsupported blocks cost?Mixed systems, true modal coordinatesHold all, enforce all, SPEC, oracleRefusal costs 4% over the oracle, enforcing all costs 39% (Table 4)
EnforceWhat does exactness buy?Damped chains, 80 configurationsAggregate law, penalty, exact lawLaw held to 10−1510^{-15} vs. 10−310^{-3}, 29.4×29.4\times gain over one aggregate law (Table 5)
AuthorsOpenReview ↗

Response to Reviewer TGZJ (2/2): Q1 on contraction and diffeomorphisms, and Q2 to Q7 on terminology and notation

Q1. Contraction and diffeomorphisms. In Theorem 1, congruence is matrix congruence, U⊤MU−M=SA−⊤ΣM(A)SA−1U^\top M U - M = S_A^{-\top} \Sigma_M(A) S_A^{-1} with SA=I−A/2S_A = I - A/2 (Eq. 4), so by Sylvester's law of inertia the two matrices have the same inertia: one vanishes, is negative semidefinite or is positive semidefinite exactly when the other does. It relates two matrices of one update and is not a change of coordinates of the data. The congruence complements the learned coordinate map rather than representing diffeomorphisms itself: nonlinear changes of coordinates are handled by the learned encoder, and the congruence certifies the linear step that remains.

(i) Contraction is coordinate free, but its certificate requires a metric, which transforms with the coordinates. A metric M=Θ⊤ΘM = \Theta^\top \Theta arises from a differential change of coordinates δz=Θδx\delta z = \Theta \delta x, and when Θ\Theta is the Jacobian of a map z=ϕ(x)z = \phi(x), a system that contracts in MM contracts in the Euclidean metric in zz (Lohmiller and Slotine, Automatica 1998). SPEC divides the problem in the same way. Discover learns the nonlinear coordinate map, and Theorem 1 applies to the linear step that remains. The energy balance is invariant under any linear change of coordinates (Proposition 6), so the congruence needs to account only for linear transformations.

(ii) The congruence decides by the sign of one symmetric matrix whether the discrete step itself conserves, decreases or increases an energy, for every value of the trainable parameters. Discretisation can destroy a property of the continuous system. For x˙=Ax\dot x = A x with A⊤=−AA^\top = -A, which conserves ∥x∥2\lVert x \rVert^2, the explicit Euler step x+=(I+hA)xx^{+} = (I + hA) x gives ∥x+∥2=∥x∥2+hx⊤(A⊤+A)x+h2∥Ax∥2=∥x∥2+h2∥Ax∥2,\lVert x^{+} \rVert^2 = \lVert x \rVert^2 + h x^\top (A^\top + A) x + h^2 \lVert A x \rVert^2 = \lVert x \rVert^2 + h^2 \lVert A x \rVert^2, so the energy grows at every step although the system is conservative (Hairer et al., 2006). The Cayley step U=Cay(hA)U = \mathrm{Cay}(hA) has ΣI(hA)=h(A⊤+A)=0\Sigma_I(hA) = h(A^\top + A) = 0 and conserves ∥x∥2\lVert x \rVert^2 exactly for every step size. Theorem 1 extends this exactness to decay and growth.

(iii) A contraction metric that does not come from a change of coordinates (a non-flat metric) cannot be made Euclidean, so SPEC with a block-constant metric covers the flat case. Two statements in the summary can be made more precise. The Cayley map covers all rotations except a set of measure zero, rather than all orthogonal matrices (Proposition 10). SPEC assigns one law to each block or refuses (§3.3), and the laws concern the length of a latent block, of which exponential decay is the contracting case.

Q2. The term "budget". The budget is the energy balance of one step, the amount of x⊤Mxx^\top M x gained or lost. By Theorem 1 it equals w⊤ΣM(A)ww^\top \Sigma_M(A) w, whose sign determines whether the energy is conserved, decreases or increases, so ΣM(A)\Sigma_M(A) is called the budget operator. The term was defined only in Table 9, and we will define it in §3.2 at first use.

Q4 and Q5. Subscripts in Eq. 10, and the symbols η\eta and μ\mu. The index cc runs over the coordinates of the observed state, so the second term is the mean squared reconstruction error over trajectories jj, times tt and coordinates cc. The indices kk and ll run over the entries of the latent covariance matrix, so the last term is the mean squared entry of Cov^(z)−I\widehat{\mathrm{Cov}}(z) - I, which prevents the latent coordinates from collapsing to a constant. η\eta is the parameter vector of a law: γˉ\bar\gamma for exponential decay, (c,p)(c, p) for algebraic decay, (γ0,r∗)(\gamma_0, r_{\ast}) for the attracting sphere and (γ0,R)(\gamma_0, R) for the forward-invariant region (§3.1). μ=γ0r∗2Δt\mu = \gamma_0 r_{\ast}^2 \Delta t is the dimensionless step size of the sphere law, which is locally attracting for 0<μ<10 < \mu < 1 (Appendix C.6). We will state these definitions next to Eqs. 8 and 10.

Q6. BLAS reduction order. BLAS is the standard linear-algebra library behind matrix products. With several threads, a long sum is split into partial sums whose order depends on the thread count, and since floating-point addition is not associative, the last digits vary with that count. The comparison objective amplifies these differences over its 1200-step optimisation, whereas our objective is bitwise identical across thread counts (§4.2). This detail will move to the appendix.

Q3 and Q7. Notation and readability. We will move a compact notation table into §3.1 and define every symbol at its first use. The revisions in Part 1 address readability: a plain statement of the problem and the method with one running example, Theorem 1 in its midpoint form, every symbol defined at first use, and an experimental section that reports the question, data, comparison and result of each experiment.

AuthorsOpenReview ↗

Response to Reviewer TGZJ: all revisions completed in the uploaded PDF

We have revised the paper to address every point raised in the review. The revised PDF is now uploaded, with every change marked in blue, and each point is addressed as follows.

(i) Clarity and Theorem 1 (W1, Q7). §1 now states the problem in plain terms and walks through the three stages on one example. §3.2 now reads Theorem 1 through its midpoint form, with each matrix defined beside it.

(ii) The experimental section (W2). The new Table 1 summarises the question, data and comparison of each experiment. The identity experiment, formerly Q1 of §4, has moved to Appendix F.7, where its table and figure are now Table 19 and Figure 9, and the new question Q5 (§4.5) evaluates the three stages together on learned coordinates.

(iii) Terminology and notation (Q2 to Q6). §3.2 defines the budget and the budget operator at first use. The symbols η\eta and μ\mu are defined next to Eq. 8, and the indices of Eq. 10 next to that equation. §3.1 now contains a compact notation table, and the BLAS detail has moved to Appendix F.2.

(iv) Contraction and diffeomorphisms (reviewer Q1). This question concerns the scope of the congruence rather than the presentation, and our response answers it in full.

The new Table 6 shifts Tables 6 to 17 of the submission to Tables 7 to 18, so the cited Tables 9 and 11 are now Tables 10 and 12. Sections, equations, propositions and Tables 2 to 5 keep their numbers.

We hope that the revised PDF makes the method and the experiments easier to follow, and we would be glad to clarify any remaining point. We would be grateful if the reviewer would reconsider the evaluation of the paper in light of the revision.

Reviewer_2sHTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses a limitation of existing structure-preserving dynamics learning methods, which typically assume prior knowledge of the structures to be preserved. To overcome this limitation, the authors propose SPEC, a framework that discovers, diagnoses, and enforces appropriate dynamical structures directly from observational data.

As its structural backbone, SPEC combines the Cayley transform with radial scaling to separate the angular and radial dynamics of latent blocks. This parameterization can exactly represent five types of radial laws: conservation, exponential decay, algebraic decay, attracting sphere, and forward-invariant region.

Then, the SPEC framework consists of three sequential steps: (1) Discover: observed trajectories are decomposed into multiple latent blocks, and their coordinate representations are learned by minimizing the temporal variance of the log-radius growth rate of each block, along with additional regularizer terms. (2) Diagnose: With the learned coordinates fixed, five candidate radial laws are fitted to each block and evaluated using validation scores. A radial law is assigned to a block only when supported by sufficient evidence; otherwise, the block remains unconstrained. (3) Enforce: The dynamics are parameterized to satisfy the selected radial laws exactly, while the angular dynamics, represented by the Cayley transform, are also learned.

Experimental results demonstrate that the Cayley-based update preserves the prescribed radial laws to machine precision; the discovery step often identifies latent coordinates that effectively capture dynamical invariants; the diagnosis step can appropriately accept or reject candidate radial laws, with selective rejection of unsupported constraints improving predictive performance; some pedagogical examples illustrate that exact enforcement of structural laws does not always outperform soft-penalty approaches.

Critical strengths

S1. I find this paper well-motivated. Structure-preserving learning plays an important role in scientific machine learning. However, many existing studies introduce specific structural constraints and evaluate their effectiveness on systems where those constraints are known a priori. While such evaluations are certainly valuable, they represent a somewhat idealized setting, as the appropriate physical structures are often unknown in real-world problems. This paper addresses this important gap by proposing a practical learning framework that does not require prior knowledge of which structural constraints are appropriate. I believe this direction has considerable potential for future applications in automated scientific modeling and discovery.


S2. I also appreciate the transparency of the experimental evaluation, which examines both successful cases and failure modes. The authors demonstrate that enforcing structural constraints is not always beneficial. More interestingly, even when the chosen structural family is appropriate, imperfect specification can make exact enforcement perform worse than a softer approach. These findings provide valuable insights into the practical limitations of structure-preserving learning and make the experimental conclusions more convincing and trustworthy.

Critical weaknesses

W1. While the discover and diagnose stages are reasonably motivated, they remain largely heuristic. For example, the coordinate discovery objective minimizes the temporal variance of block log-radius growth rates, which appears to favor latent coordinates exhibiting conservation or exponential decay. It is therefore unclear why this objective should also discover coordinates suitable for the other three radial laws or more general structured coordinates Similarly, the diagnose stage relies on thresholding validation scores computed from finite trajectories, making it closer to an empirical model-selection procedure.


W2. While the experimental analysis is thorough, the benchmark systems themselves appear to focus primarily on conservative and exponentially decaying dynamics. I could not find clear demonstrations of the data-driven discovery and selection pipeline for the other three radial laws, algebraic decay, attracting sphere, and forward-invariant region. Given that the exact parameterization of five radial laws is an important component of the proposed framework, broader experimental validation would help establish its applicability beyond conservation and exponential decay.


W3. Although the five radial laws considered in this paper capture important properties of many physical systems, it is unclear how broadly the proposed Cayley-based radial-angular representation can express general nonlinear dynamics. I do not consider this restriction a weakness, since structure-preserving models naturally constrain the class of functions they can approximate. Nevertheless, a clearer discussion of the representational capacity and limitations of the proposed parameterization would strengthen the theoretical completeness of the paper. I have not checked the Appendix, so I would appreciate clarification if these points have already been addressed there.

Questions for authors

Q1. Could the authors further clarify the motivation for minimizing the temporal variance of block log-radius growth rates as the primary coordinate-discovery objective? This criterion appears naturally suited to conservation and exponential decay, for which the log-radius growth rate is constant. However, the other three radial laws generally exhibit time-varying log-radius growth rates. Why should the proposed objective be expected to discover coordinates that are also suitable for these laws?


Q2. Related to Q1, the experiments appear to focus primarily on conservative and exponentially decaying systems. I could not find clear examples demonstrating that SPEC can successfully discover algebraic decay, attracting sphere, or forward-invariant region in a data-driven manner. Given that the exact construction of these five radial laws is an important contribution of the paper, could the authors provide additional experiments demonstrating the effectiveness of the proposed SPEC for the remaining three laws?


Q3. Could the authors provide a more detailed theoretical discussion of the representational capacity of the proposed Cayley-based parameterization? I understand that the construction can exactly realize admissible radial dynamics. However, it remains unclear what classes of general nonlinear dynamical systems can be represented. While I do not expect a structure-preserving parameterization to universally approximate arbitrary dynamical systems, a clearer characterization of its representational capabilities and limitations would strengthen the theoretical contribution.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 2sHT (1/4): Overview, and W1/Q1 on the Discover objective

We thank the reviewer for recognising that structure-preserving methods are usually evaluated where the structure is known in advance, and that SPEC addresses the realistic setting in which it is not (S1), and that reporting failure modes alongside successes strengthens the conclusions (S2). Our response has four parts. Part 1 clarifies which three laws Diagnose searches, why the Discover objective serves them, and what the other two require (W1, Q1). Part 2 gives the threshold of Diagnose an exact meaning and measures its reliability on a new labelled set (W1, second part). Part 3 demonstrates the algebraic, sphere and region laws from data through all three stages (W2, Q2). Part 4 characterises what the construction can and cannot represent, with a new completeness result (W3, Q3).

One point of the summary underlies W1, W2, Q1 and Q2. Of the five laws, Diagnose searches three from data, namely conservation, exponential decay and algebraic decay (§1 lines 083 to 085, Eq. 26). The attracting sphere enters the candidate set only when its radius r∗r_{\ast} is supplied, and the forward-invariant region is assigned only when its boundary RR is supplied (Algorithm 2 lines 5 to 12, Appendix D.1).

W1 and Q1. The Discover objective. The reviewer is right that the rate term rewards a constant log rate. This covers all three searched laws, and the other two are conditional because of identifiability rather than because of the objective.

(i) The temporal variance of the block log rate δi,t\delta_{i,t} is the minimum over γ\gamma of its mean squared deviation from γ\gamma, so the rate term is the least-squares residual of the exponential family (conservation is γ=0\gamma = 0), and it vanishes exactly when every block radius evolves exponentially. Every complex Koopman eigenfunction pair is a zero of the rate term: if φ(xt+1)=λφ(xt)\varphi(x_{t+1}) = \lambda \varphi(x_t), the block z=(Reφ,Imφ)z = (\mathrm{Re} \varphi, \mathrm{Im} \varphi) satisfies ∥zt+1∥=∣λ∣∥zt∥\lVert z_{t+1} \rVert = |\lambda| \lVert z_t \rVert (Mezić, Nonlinear Dynamics 2005, Mauroy et al., Physica D 2013, Lusch et al., Nature Communications 2018). The rate term is also invariant under positive blockwise rescaling (Proposition 6), so it needs no reconstruction weight, and its median varies by at most 0.16 across three orders of magnitude of that weight, against 0.81 for the comparison objective (Table 12).

(ii) Algebraic decay is reachable through an exact change of radial variable. By Eq. 8, an algebraic trajectory satisfies rk−2p=r0−2p+2pckΔtr_k^{-2p} = r_0^{-2p} + 2pck \Delta t, so the increasing map s=exp⁡(−r−2p/(2pc))s = \exp(-r^{-2p}/(2pc)) gives log⁡sk=−r0−2p/(2pc)−kΔt\log s_k = -r_0^{-2p}/(2pc) - k \Delta t, a constant log rate for every pp and cc. So ss minimises the rate term, and exponential decay of ss is algebraic decay of rr. The name of the law depends on the coordinates, but the prediction does not: an encoder that reaches ss is assigned exponential decay, and its exact update is exactly algebraic in rr. From data, we ran Discover without reconstruction weight on the seven algebraic systems of Part 3, with the encoder anchored at the equilibrium, z=Φ(x)−Φ(xeq)z = \Phi(x) - \Phi(x_{\mathrm{eq}}), which follows from U(0)=0U(0) = 0 for every construction. The learned radius is nearly monotone in the true radius on every seed (Spearman correlation at least 0.996), and Diagnose accepts the algebraic law on 17 of the 21 best blocks. At 100Tfit100 T_{\mathrm{fit}}, one hundred times the fitting interval TfitT_{\mathrm{fit}}, the enforced law reaches a log-radius error of 0.23 to 0.25 on the Stuart-Landau and planar systems with p=1,2p = 1, 2, against 30 to 59 for the exponential law forced on the same blocks, 35 to 63 for the penalty and 41 to 70 for an unconstrained map. The learned blocks of the planar system with p=0.5p = 0.5 are refused, so no law is imposed on them.

(iii) For the attracting sphere the limit is identifiability. The native log rate is −γ0(rt2−r∗2)-\gamma_0(r_t^2 - r_{\ast}^2), so the rate term is small near the sphere and zero on it. On the attractor, every sphere law and conservation at r∗r_{\ast} give identical radius sequences, so γ0\gamma_0 is observable only in transients, which is why the sphere candidate requires r∗r_{\ast}. For the forward-invariant region the argument is stronger: for r≤Rr \le R the region law coincides with conservation, so no trajectory inside the ball can identify RR. Supplying the invariant set in learned latent coordinates follows Takeishi and Kawahara (AAAI 2021), and treating the safe set as a design input is standard in invariant-set and barrier-function design (Blanchini, Automatica 1999, Ames et al., IEEE TAC 2017). Coordinates suited to conservation therefore serve the region law, and Discover recovers them on the pendulum, Lotka-Volterra and Kepler systems (Table 2). Part 3 confirms (ii) and (iii) experimentally. We will add (i) to (iii) to §3.3, and mark in Fig. 1 which laws are searched and which are supplied.

AuthorsOpenReview ↗

Response to Reviewer 2sHT (2/4): W1 (second part), Diagnose as model selection

W1 (second part). Diagnose as model selection. We agree that Diagnose is a model-selection procedure, and it is designed as one: the selection is statistical, its threshold was fixed before evaluation (Table 7), and the paper keeps it separate from the guarantee, which holds at every parameter value.

(i) Algorithm 1 marks what is estimated from finite trajectories and what holds at every parameter value, and contribution (ii) states this separation (lines 100 to 102). Because the radius of an accepted block is independent of the trainable rotation (Proposition 5), a decision error cannot invalidate an accepted guarantee. The rule is forward validation. Parameters are fitted on the fitting interval and propagated recursively, without refitting, over a disjoint later validation interval, so the score measures extrapolation within the window. Evaluating on a later interval respects the temporal order of dependent data, a concern central to time-series predictor evaluation (Bergmeir and Benítez, Information Sciences 2012). Held-out selection is the general principle of model selection (Arlot and Celisse, Statistics Surveys 2010), and selecting among candidate models on held-out evidence is standard practice in sparse equation discovery (Mangan et al., Proc. R. Soc. A 2017). Two choices are specific to exact enforcement: the score is invariant under joint rescaling of the radii (Proposition 6), and the constructibility gate refuses a candidate whose parameters cannot be realised exactly rather than projecting it, since projection would enforce a law other than the one scored (Eq. 11).

(ii) The threshold has an exact meaning, which we will state in §3.3. The score Rlog⁡2R^2_{\log} of Eq. 28 is the coefficient of determination of the log radius on the validation interval, so Rlog⁡2≥τR^2_{\log} \ge \tau holds exactly when the held-out squared error of the candidate is at most a fraction 1−τ1 - \tau of the error of the best constant log radius. At τd=0.30\tau_d = 0.30, a law is accepted only if it removes at least 30% of the held-out error left by a constant level. The routing conclusions do not hinge on τd=0.30\tau_d = 0.30. On the benchmark of Table 4, the in-family error is identical (0.0125 at 1×1\times, 0.0997 at 10×10\times) for every τ\tau from −0.5-0.5 to 0.80.8, and for every τ\tau from 0.4 to 0.7 SPEC enforces exactly the in-family blocks, as the oracle router does with the true labels. Replacing the threshold by AICc or BIC comparisons against the refused-block fallback enforces far more unsupported blocks (227 and 198, against 6 at τd=0.30\tau_d = 0.30) and raises the out-of-family error at 10×10\times from 0.624 to 2.55 and 2.54.

(iii) To measure reliability directly, we built a labelled set of 100 blocks per class: algebraic, sphere and exponential blocks, and as unsupported blocks the instance set of Appendix F.5 (pendulum, Van der Pol, coupled nonlinear, heavy-damped with noise, 25 each). With r∗r_{\ast} supplied for sphere systems, as in the paper, Diagnose accepts and correctly names all 100 algebraic and all 100 sphere blocks, refuses 99 of the 100 unsupported blocks (false acceptance 0.01, 95% Clopper-Pearson interval [0.0003,0.054][0.0003, 0.054]), and reaches a threshold-free ROC-AUC of 1.000. All 100 exponential blocks are accepted. Since the algebraic law of Eq. 23 contains the exponential law as its limit p→0p \to 0, 51 of them are named algebraic without a tie rule, at a median score difference of 5×10−65 \times 10^{-6}. Resolving near-ties (difference below 10−310^{-3}) in favour of the law with fewer parameters names 94 exponential and leaves every algebraic block unchanged (smallest margin 9.5). Without r∗r_{\ast}, 82 of the 100 sphere blocks are refused and 18 are named algebraic, since far above the cycle the sphere rate γ0(r2−r∗2)\gamma_0(r^2 - r_{\ast}^2) approaches the algebraic rate cr2c r^2 of Eq. 8 with p=1p = 1. This is the identifiability limit of Part 1 (iii), and the reason the paper adds the sphere candidate only when r∗r_{\ast} is supplied. The Van der Pol blocks of this set are taken in raw coordinates, where the cycle is not circular and no r∗r_{\ast} is given, so refusing them is correct. In Part 3 the same system is accepted only in coordinates learned by Discover, in which the cycle is circular, with r∗r_{\ast} supplied.

AuthorsOpenReview ↗

Response to Reviewer 2sHT (3/4): W2/Q2, the algebraic, sphere and region laws from data

W2 and Q2. Algebraic decay, attracting sphere and forward-invariant region. We agree that the algebraic law lacked a ground-truth demonstration and that the sphere evidence of Table 16 is thin (its acceptance interval is [0,1][0, 1], and the labelled set of Part 2 replaces it). The constructions are exact by proof (Appendix C.6), so we tested the data-driven part. For each law we ran Diagnose on the true radius, the full pipeline from observations, and rollouts to 100Tfit100 T_{\mathrm{fit}} against the true continuation, comparing the selected law with the wrong law forced on the same blocks, the per-block penalty of §4.4 and an unconstrained block map. The algebraic candidate is fitted through the recursion of Eq. 23, propagated as in Eq. 27, which carries the time offset of each trajectory, whereas the log-log slope of Appendix D.1 assumes that the decay starts at t=0t = 0. We will use the recursive form in the revision.

(i) Algebraic decay is tested on seven systems. The Stuart-Landau oscillator at the degenerate Hopf point, z˙=iωz−∣z∣2z\dot z = i \omega z - |z|^2 z, has r˙=−r3\dot r = -r^3 (Kuramoto, Chemical Oscillations, Waves, and Turbulence, 1984), and the alg line of Eq. 8 is its exact flow map with p=c=1p = c = 1. Planar systems r˙=−cr2p+1\dot r = -c r^{2p+1} with pp in {0.5,1,2}\lbrace 0.5, 1, 2 \rbrace are observed through a fixed invertible nonlinear map, which Discover must undo. A linear oscillator with cubic damping, x¨+ω02x+εx˙3=0\ddot x + \omega_0^2 x + \varepsilon \dot x^3 = 0 with ε\varepsilon in {0.01,0.05,0.1}\lbrace 0.01, 0.05, 0.1 \rbrace, has amplitude a˙=−38εω02a3\dot a = -\frac{3}{8} \varepsilon \omega_0^2 a^3 to first order by averaging (Nayfeh and Mook, Nonlinear Oscillations, 1979), so the law holds only approximately. On the true radius Diagnose selects and accepts the algebraic law on all seven systems. Where the law is exact, pp and cc are recovered within 1.4×10−41.4 \times 10^{-4} relative error, and at 100Tfit100 T_{\mathrm{fit}} the enforced law tracks the true radius to 1.0×10−41.0 \times 10^{-4} in log radius, against 71 to 295 for the forced exponential law, 146 to 416 for the penalty and 450 to 618 for the unconstrained map. For cubic damping, cc is recovered within 0.7% at ε=0.05\varepsilon = 0.05 and 0.10.1, with a log-radius error of 0.13 to 0.14 against 36 to 67 for the forced exponential law. At ε=0.01\varepsilon = 0.01 the amplitude decays too little within the window to determine pp, yet the selected law still lowers the error of the forced exponential law from 9.34 to 3.56. The Discover results are in Part 1 (ii).

(ii) The attracting sphere is tested on the Stuart-Landau oscillator beyond the Hopf point, r˙=νr−r3\dot r = \nu r - r^3 with r∗=νr_{\ast} = \sqrt{\nu}, and on the Van der Pol oscillator, whose limit cycle is not circular in raw coordinates. With the transient extending into the validation interval (ν\nu in {0.05,0.1}\lbrace 0.05, 0.1 \rbrace), Diagnose selects the sphere law on all 48 trajectories with r∗r_{\ast} supplied, and also when r∗r_{\ast} and γ0\gamma_0 are estimated jointly, an extension beyond Appendix D.1 that recovers both within 1.2×10−31.2 \times 10^{-3}. The enforced law reaches a log-radius error at 100Tfit100 T_{\mathrm{fit}} of 5.7×10−85.7 \times 10^{-8} and 4.7×10−104.7 \times 10^{-10}, against 0.020 and 4.0×10−44.0 \times 10^{-4} for conservation forced on the same blocks, 6.1 and 2.3 for the penalty, and 31 for the unconstrained map. With Discover in the loop and r∗r_{\ast} taken as the learned radius near the cycle, the sphere law is accepted on all 12 learned blocks for ν=0.05\nu = 0.05, on 2 of the 3 best blocks for ν=0.1\nu = 0.1, and on 10 of the 12 blocks for Van der Pol. On the best learned blocks the radius ends within 6×10−146 \times 10^{-14} of r∗r_{\ast}, with an error of 1.2×10−31.2 \times 10^{-3} and 7.8×10−47.8 \times 10^{-4}, against 2.8 and 6.2 for the penalty and 6.5 and 4.9 for the unconstrained map.

(iii) The forward-invariant region is tested where it is exact, on the Stuart-Landau oscillator with R=r∗R = r_{\ast}, and where it is imposed, with RR above the largest training radius on the blocks that Discover learns for the pendulum and Kepler. No rollout started inside the ball ever left it, after training and for ten random draws of the parameters, whereas the trained least-squares block map left the ball in 12.5% to 100% of rollouts on the learned blocks. On the Stuart-Landau oscillator the region construction extrapolates the radius to 3.3×10−123.3 \times 10^{-12} at 100Tfit100 T_{\mathrm{fit}}, against 50.4 for the penalty and 70 for the least-squares map, and from outside starts it contracts onto the boundary, with any recrossing at most 4.4×10−164.4 \times 10^{-16}. On the learned blocks the log-radius error is 0.327 and 0.0028 on the pendulum and Kepler, against 0.425 and 0.0428 for the penalty and 36.1 and 4.18 for the least-squares map. These results will be added as new appendix sections and summarised in §4.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and the revision for the Program Chairs, Senior Area Chairs, Area Chair and Reviewers

We thank the Program Chairs, the Senior Area Chairs, the Area Chair and the three reviewers for their time and careful reading. As the discussion period draws to a close, we summarise our responses in one place.

All three reviewers rate the soundness as having minor or fixable concerns, and no review identifies an error in the theory or in the experiments. We have answered every weakness and question of each reviewer in detail, and the revised PDF marks every change in blue. The main additions are as follows.

(i) End to end validation (ZXvN, 2sHT). The new Q5 (§4.5, Table 6) runs the three stages on learned coordinates and reaches a truth error of 0.038 at one hundred times the fitting interval, against 0.377 on analytical coordinates.

(ii) Discover (ZXvN). Anchoring and exact DMD initialisation raise the fraction reaching the target score from 2.2% to 22.8% and 98.6% (§4.2).

(iii) Exactness and the metric (ZXvN). With the metric estimated from data, the exact construction reaches 1.1×10−121.1 \times 10^{-12}, against 0.206 for the penalty in Table 5 (§4.4).

(iv) Diagnose (ZXvN, 2sHT). On a labelled set, Diagnose refuses 99 of 100 unsupported blocks (§4.3, Appendix F.5).

(v) Representational capacity (2sHT). The new Proposition 13 (Appendix C.11) states which block maps the construction can represent.

(vi) Clarity (TGZJ). §1 states the problem in plain terms with one example, §3.1 adds a notation table, and the new Table 1 gives an overview of the experiments.

We thank Reviewer TGZJ for the comments on the presentation, and we have answered every weakness and question in this review in full. For the Area Chair's consideration, this review reports a confidence of 1 ("Insufficient expertise, has alerted the AC") and rates the soundness as having minor or fixable concerns.

Under each review, a follow-up maps every point to its location in the revised PDF. We would be very grateful for the reviewers' feedback on whether our responses address their concerns, and we remain glad to answer any further question.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.