- 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
-
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 People's understanding of the instanteous system could be inaccurate, due to outdated data, ever-evovling system dynamics.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.
-
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
-
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.
-
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.
-
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.
-
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
-
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.
-
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.
-
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
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 0.99986 against the modal radius), and Diagnose accepts every block in the median cell. The truth error at , one hundred times the fitting interval , 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 against 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 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 the seed distributions overlap (Table 15), and a rank test over the ten seeds finds no difference (two-sided Mann-Whitney , paired Wilcoxon ), whereas on each analytic system at the same setting the rate objective is higher on all ten seeds (). 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 , so the invariant 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 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 , with the radius ending within of , 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, , so the increasing map has , a constant log rate, and exponential decay of is exactly algebraic decay of . 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 (30 to 59 against 0.23 to 0.25). For the attracting sphere the native log rate vanishes on the attractor, where every sphere law coincides with conservation at , so the rate is identifiable only from transients and the sphere candidate requires (§3.3). With transients present, the sphere law is accepted on all 12 learned Stuart-Landau blocks () at , on 2 of the 3 best blocks at and on 10 of the 12 Van der Pol blocks. We will report these results in §4.3.
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 at and at , whereas the penalty deviates by and (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 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 , 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 ), but the effective rate of the penalty lies closer to the truth for 74% of the modes, and over 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 and 0.0205 at , against 0.0603 for the penalty at (0.134 at , 0.027 at ). The remaining 0.0205 is the Euclidean ripple of Proposition 7. With the metric estimated from data (Part 4), the exact construction reaches at , the level of the true metric (), 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 from 0.1 to 1 to 10 lowers the deviation of the penalty from its law at 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 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 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 , the radius of block of , which is a function of the physical state (Appendix C.9). Its agreement with the physical invariant is what the invariant 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 and this rank correlation report. We will state (i) to (iv) in §4.4.
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 7×10−13.
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 z, 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=0 to a nonzero point. This effective offset accounts for a median of 16% of the block radius, and 52% at N=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)=0, so the equilibrium must map to the latent origin, and anchoring the encoder there, z=W(x−xeq), removes the offset and changes nothing else. On the same grid and seeds, anchoring raises the share of evaluations whose held-out score Rlog2 (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 Rlog2 at ten characteristic decay times from −3.30 to 0.84. 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 (4×10−30), and in every cell of the grid the loss there is lower than at the encoders learned from a random start (median 2.7×10−5 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 Rlog2 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 R2 (the R2 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).