acceptodds
Under review as a conference paper at ICLR 2027

Geometric Dictionary Learning for Operator Representations of Dynamical Systems

Abstract

Operator-theoretic representations provide a powerful framework for learning dynamical systems from data and interpreting their behavior: through spectral decomposition, they characterize the dynamics in terms of modes, frequencies, and time scales. Yet existing approaches typically estimate an operator independently for each system, even when data from a set of related dynamics are available, thereby failing to leverage population-level information to obtain more compact representations and more data-efficient estimators. In this work, we posit that the operator representations of related dynamical systems concentrate near a low-dimensional manifold. Based on this hypothesis, we introduce DOODL (Dynamical OperatOr Dictionary Learning), a framework that approximates this population manifold through a dictionary of characteristic operators. The resulting dictionary yields compact and interpretable representations of individual systems while preserving their dynamical properties. Beyond representation learning, DOODL enables operator estimation from short trajectories by constraining new operators to be combinations of the learned operators. Experiments on metastable Langevin dynamics and turbulent plasma simulations show that DOODL captures meaningful variations across populations of complex multiscale systems while enabling accurate operator estimation in low-data regimes, with errors one to two orders of magnitude lower than independent operator estimation methods.

Primary area:
learning on time series and dynamical systems
Keywords:
dynamical systems, operator learning, dictionary learning, spectral theory, Riemannian geometry

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Cn1gOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DOODL, a framework for learning a representation of a population of related dynamical systems using Koopman operators, under the assumption that estimated Koopman operators of systems share a low-dimensional manifold. Its key components include: (1) Spectral decompositions are treated as points on a smooth manifold, quotiented by eigenvector rescaling, and equipped with a scale-invariant Riemannian metric and a retraction. (2) Dictionary learning: atoms and simplex codes are obtained by minimizing the SGOT divergence between each operator and its reconstruction B^p, which is a convex combination of the atoms’ eigenvalues and eigenvectors followed by a bi-orthogonality projection. (3) Dictionary-constrained estimation: from a new short trajectory, only the code is fit via a one-step feature-prediction loss. Theorem 1 bounds the excess risk with a complexity term in the intrinsic dimension rather than the feature dimension.

Critical strengths

The problem is well motivated and studies an interesting problem. Population-level priors for operator learning are understudied in the literature, and the paper makes a meaningful conceptual contribution towards this goal. The idea of learning a population-level prior directly on spectral decompositions is also interesting and useful in practice. The proposed part pipeline is sound and also constitutes an innovative technical contribution.

Critical weaknesses

(1) The central hypothesis is that operators of related systems concentrate near a low-dimensional manifold, and that convex combinations of atoms remain physically meaningful. However, for a parametrized family of SDEs, how strong is this assumption? When a parameter crosses a bifurcation, for example, a single-well dynamics becomes a double well after parameter changes, a second near-zero eigenvalue separated by the spectral gap appears. In its current form of the paper, the estimator always returns a point in the convex hull of the atoms, and nothing flags a system that lies off the learned manifold. Some intuition or experimental checks would be helpful. (2) Koopman estimation is often characterized w.r.t Lπ2L_\pi^2 where π\pi is the invariant measure of the system. Appendix A.1 states its error bounds in exactly this space. However, in a population, every system has its own invariant measure πi\pi_i. It would be helpful to clarify this. (3)Theorem 1 assumes a stationary test trajectory. Section 4, however, advertises operator estimation from short non-equilibrium trajectories, and windows of 10–100 samples are unlikely to be sampled according to the invariant measure. The guarantee therefore does not cover the regime the experiments highlight. (4) The paper takes the per-system spectral decompositions as given. Spectral estimation from data, however, is a hard problem itself, as finite-rank truncations can produce spurious eigenvalues, known as spectral pollution, that do not correspond to the underlying operator. The paper cites ResDMD, it would be helpful to discuss how robust the method is. (5) The Langevin dynamics is a one-parameter, one-dimensional family. A more convincing evaluation would include a multidimensional or non-reversible system. The text also missed a dt in the description of the SDE.

Questions for authors

(1) Under what conditions on the parameter family is the manifold hypothesis expected to hold? How does DOODL behave on a system that lies off the learned manifold? (2) Theorem 1 assumes a stationary test trajectory, while the experiments target short non-equilibrium windows. Can the guarantee be extended to non-stationary trajectories? (3) How sensitive are the learned atoms to spurious eigenvalues?

6 comments

W1: Clarification on invariant measure

Koopman estimation is often characterized w.r.t where is the invariant measure of the system. Appendix A.1 states its error bounds in exactly this space. However, in a population, every system has its own invariant measure . It would be helpful to clarify this.

The per-system spaces Lπi2L^2_{\pi_i} only appear in Appendix A.1, which recalls the guarantees for estimating each training operator individually. DOODL itself does not compare operators in these spaces. By assumption (A1), all operators act on a common finite-dimensional feature space H\mathcal H (e.g. shared Random Fourier Features). The SGOT loss compares their spectral decompositions in the geometry of H\mathcal H, which does not depend on any πi\pi_i. In Theorem 1, the only measure involved is the invariant measure π\pi of the test system, through the population risk Eπ[ℓt(G)]\mathbb E_\pi[\ell_t(G)]. The training systems enter only through the learned dictionary, which is fixed and independent of the test trajectory.

W2: Langevin increase

The Langevin dynamics is a one-parameter, one-dimensional family. A more convincing evaluation would include a multidimensional or non-reversible system. The text also missed a dt in the description of the SDE.

We volontarily kept the Langevin example simple so that we can better interpret the embeddings. We would like to highlight that we did a very high dimensional and non-reversible system as asked by the reveiwer with the plasma experiments and we did recover quite well physical parameters from the DOODL embeddings.

Q1: Formalism of related dynamical systems

  • The central hypothesis is that operators of related systems concentrate near a low-dimensional manifold, and that convex combinations of atoms remain physically meaningful. However, for a parametrized family of SDEs, how strong is this assumption? When a parameter crosses a bifurcation, for example, a single-well dynamics becomes a double well after parameter changes, a second near-zero eigenvalue separated by the spectral gap appears.
  • Under what conditions on the parameter family is the manifold hypothesis expected to hold?

We thank the reviewer for raising bifurcations, which are the right stress test for the hypothesis. The new Appendix G.5 (general comment R3) gives sufficient conditions: smooth dependence on the parameters, and a uniform gap between the rr leading eigenvalues and from the rest of the spectrum (Proposition 9). We clarify below what this implies for reviewer's example, and where the hypothesis genuinely fails.

Single-well to double-well. A bifurcation of the potential does not, by itself, create a singularity in operator space. For one-dimensional reversible diffusions, the generator is a Sturm–Liouville operator: its eigenvalues are simple, cannot cross, and depend analytically on the parameters of the potential. The near-zero eigenvalue you describe is the first non-trivial eigenvalue λ1\lambda_1, which decreases continuously and becomes exponentially small in the barrier height (Kramers regime), rather than appearing at the bifurcation. For Ua(x)=x4/4−ax2/2U_a(x)=x^4/4-ax^2/2 with a∈[−1,2]a\in[-1,2] and σ=0.25\sigma=0.25, λ1\lambda_1 decreases smoothly from 1.431.43 to 0.0150.015, while the relative gaps among the first four eigenvalues stay above 0.300.30 (Table 7). The population therefore remains a smooth curve. Its curvature, however, concentrates around the transition, where more atoms are needed: the constants of Proposition 9 scale with the inverse gap.

Our Langevin family. For Uw(x)=U1(x/w)U_w(x)=U_1(x/w), varying ww is a pure time rescaling, Lw=w−2DwL1Dw−1\mathcal L_w=w^{-2}D_w\mathcal L_1D_w^{-1} with DwD_w the spatio-temporal rescaling (Proposition 10). Hence λk(w)=λk(1)/w2\lambda_k(w)=\lambda_k(1)/w^2 and ϕk,w=ϕk(⋅/w)\phi_{k,w}=\phi_k(\cdot/w), and the population is an exact one-dimensional curve. On your question of whether convex combinations remain physically meaningful: in generator (time-scale) coordinates, any convex combination of the atoms' eigenvalues is exactly the spectrum of another member of the family, with wα−2=∑jαjwj−2w_{\boldsymbol\alpha}^{-2}=\sum_j\alpha_jw_j^{-2}. The approximation error then only stems from the eigenfunctions, and is O(h2)O(h^2).

Where your concern applies. We agree that the hypothesis is not automatic, and Appendix G.5 now lists the cases where the gap condition fails:

  1. crossings in multi-parameter families, where eigenvectors have non-trivial monodromy and our ordering-dependent decoder needs atoms on each side;
  2. exceptional points in non-reversible families, e.g. two real eigenvalues merging into a complex pair at the onset of oscillations;
  3. an eigenvalue entering the retained rank-rr block, e.g. when a third metastable state appears.

In these regimes, DOODL returns the closest point of the learned manifold, and the remaining error does not decrease with the trajectory length (Remark 6). We discuss how to flag such systems from data in our answer to Q2 and in general comment R2.

Q2: DOODL behavior for operators outside of the learned manifold

  • In its current form of the paper, the estimator always returns a point in the convex hull of the atoms, and nothing flags a system that lies off the learned manifold. Some intuition or experimental checks would be helpful.
  • How does DOODL behave on a system that lies off the learned manifold?

We thank the reviewer for encouraging us to discuss the behavior of DOODL when estimating an operator outside of the learned manifold. In the general response R2, we provide a detailed discussion and an additional experiment on that topic.

With a dictionary of dd atoms Gˉ^\hat{\bar{\mathcal{G}}}, the DOODL spanning manifold is the set of all operators reconstructed from coefficients α\alpha on the simplex.

S(Gˉ^)={Op(Bp(α,Gˉ^))∣α∈Δd−1}.\mathcal{S}(\hat{\bar{\mathcal{G}}}) = \left\lbrace \mathrm{Op}(\mathrm{B}^{\mathrm{p}}(\alpha,\hat{\bar{\mathcal{G}}})) \mid \alpha \in \Delta^{d-1} \right\rbrace.

For a new operator A∉S(Gˉ^)A \notin \mathcal{S}(\hat{\bar{\mathcal{G}}}), DOODL returns an operator on the envelope of S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}) that is closest to AA in Hilbert–Schmidt norm. Thus, for a system outside the spanning manifold, the estimate is biased by a term proportional to the Hilbert–Schmidt distance from AA to S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}).

We illustrated DOODL's behavior when estimating new operators outside the spanning manifold in the Langevin experiment. The results in R2 show that for systems outside the spanning manifold, the coefficients α\alpha concentrate on one corner of the 1D manifold, and the Hilbert–Schmidt error increases as the operator gets further away from the spanning manifold.

Q3: Estimation from short non-stationary trajectories

Theorem 1 assumes a stationary test trajectory, while the experiments target short non-equilibrium windows. Can the guarantee be extended to non-stationary trajectories?

Theorem 1 assumes a stationary test trajectory. Section 4, however, advertises operator estimation from short non-equilibrium trajectories, and windows of 10–100 samples are unlikely to be sampled according to the invariant measure. The guarantee therefore does not cover the regime the experiments highlight.

The new Appendix G.4 extends Theorem 1 to test trajectories started from an arbitrary initial distribution μ\mu for geometrically ergodic chains satisfying

∥μPt−π∥TV≤Cμρt.\lVert \mu P^t-\pi \rVert_{\mathrm{TV}} \le C_{\mu}\rho^t.

Proposition 8 retains the stationary risk as the target and adds a term of order

(1+M)2κ2Cμ(1−ρ)n.\frac{(1+M)^2\kappa^2 C_{\mu}}{(1-\rho)n}.

See Eq. (139). The proof retains the β\beta-mixing blocking argument of Theorem 1, which also applies to non-stationary trajectories. The empirical risk is now centered around a time-averaged risk, and the block variances also change. Geometric ergodicity controls both changes. The additional term vanishes as nn grows and is most relevant for windows shorter than the mixing time 1/(1−ρ)1/(1-\rho). The results for the shortest windows in the regime-switch experiment (Figure 3) illustrate the theory empirically.

Reviewer_NuUoOpenReview ↗
Rating
3(+1)
Confidence
4(+2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DOODL, a dictionary-learning method for a population of related dynamical systems, each represented by an estimated Koopman/transfer operator or generator. Each operator is described by its spectral decomposition (eigenvalues plus bi-orthogonal left/right eigenvectors), viewed as a point on a quotient manifold with a Riemannian structure derived by the authors. A dictionary of atoms is learned by minimizing the SGOT optimal-transport divergence between each training operator and its reconstruction. A reconstruction is a convex combination of the atoms' eigenvalues and eigenvectors, followed by a projection that restores bi-orthogonality. For a new system, only the convex weights are fitted on a short trajectory, and an oracle inequality shows that the estimation complexity depends on the dimension of the learned manifold rather than on the operator size. Experiments on a 1D double-well Langevin family and 2D turbulent plasma simulations show interpretable low-dimensional embeddings (R² = 0.81 / 0.97 for regressing the plasma parameters g / κ). They also show one to two orders of magnitude lower error than RRR/GR on short trajectories, and detection of regime switches.

Critical strengths
  • Learning a shared representation for a population of operators directly on their spectral decompositions, and reusing it as a prior for short-trajectory estimation, is a natural idea that per-system estimators miss. The Riemannian treatment of low-rank, non-normal spectral decompositions also seems new and reusable.
  • Beyond the toy Langevin family, the method works on 2D turbulent plasma simulations. There, the learned coordinates recover the (g, κ) structure and improve early operator estimation over direct GR.
Critical weaknesses
  • Eq. (6) combines the atoms' eigenvalues and eigenvectors index by index, so the result depends on two arbitrary choices. One is the ordering of spectral components. The other is the scale/phase of each atom's eigenvectors, which is exactly the rescaling ambiguity used to define the quotient manifold (e.g., combining v and −v with equal weights gives 0). Hence the reconstruction is not well defined on that manifold and is not permutation invariant, while the SGOT loss is. "Use a larger d" is only a heuristic. The problem should get worse with eigenvalue crossings, near-degenerate eigenvalues, or complex-conjugate pairs, which are likely for rank-40 plasma operators. This matters because geometric fidelity is the main claimed advantage over linear dictionary learning.
  • RRR/GR use one short trajectory, while DOODL is trained on 256 (Langevin) or 320 (plasma) related systems, so any use of population data should win at short lengths. Population-aware baselines are missing, e.g., the training operator with the lowest risk (7), RRR/GR shrunk toward the population mean, or a meta-/multi-task Koopman method such as Iwata & Kawahara, 2021. Linear DL, the worst method in Fig. 2, looks like a strawman. Errors are measured in SGOT, which is also the training loss, against a reference that is itself an estimate from the full trajectory (at least for plasma). Regime detection and parameter regression have no baselines, although a rolling variance of the trajectory should already detect the width switches. So the headline "one to two orders of magnitude" cannot yet be credited to DOODL's geometry.
Questions for authors
  • Before the combination in Eq. (6), how are spectral components matched across atoms, and how are eigenvectors normalized (scale and phase)? Is the decoder invariant to the rescaling that defines the quotient manifold? If not, what role does the quotient play, given that optimization is done on N^d? How are complex-conjugate pairs and eigenvalue crossings handled?

  • Please add population-aware baselines: the training operator with the lowest risk (7), RRR/GR shrunk toward the population mean, and a meta-/multi-task Koopman method such as Iwata & Kawahara, 2021. Please also add simple baselines for regime detection and parameter regression. How much of the short-trajectory gain is already obtained by the constant "DOODL avg" model? How was linear DL fitted (e.g., is the combined operator truncated back to rank r)?

  • What is the reference operator in the Langevin experiment? In 1D, the true spectrum can be computed accurately (e.g., Fokker–Planck eigenproblem on a fine grid). Can you report errors against it, plus a metric not used for training, such as held-out forecasting error?

  • As written, the left-hand side is relative to the best α for the learned dictionary, so the bias term is not needed. Did you mean the optimal dictionary of (5)? Why is the bias nonnegative, given that the optimal dictionary minimizes the SGOT objective (5), not the risk R? How is the intrinsic dimension m defined? Also, the theorem assumes a stationary test trajectory, while Section 4 uses short non-equilibrium ones. It also omits the approximation error of the manifold relative to the true operator, which drives the crossovers in Figs. 2 and 6.

  • Was the latent encoder (POSEIDON-B plus MLP heads) trained only on training simulations? Which regressor and protocol give R² = 0.81 / 0.97? How does a simple baseline (e.g., regression on time-averaged latent features) compare?

10 comments

Q1: Reconstruction by projection after linear combinations.

Before the combination in Eq. (6), how are spectral components matched across atoms, and how are eigenvectors normalized (scale and phase)? Is the decoder invariant to the rescaling that defines the quotient manifold? If not, what role does the quotient play, given that optimization is done on N^d? How are complex-conjugate pairs and eigenvalue crossings handled?

We thank the author for encouraging us to improve the description of data handling, the reconstruction map, and the dictionary training.

Data handling. Our current data-handling protocol keeps all pairs of eigenvalues and eigenfunctions, but for complex-conjugate systems, restricting to pairs on one side of the complex plane is entirely feasible and does not interfere with the method or its training. Eigenvalue crossings fall outside of our simple-spectrum assumption. We did not explicitly handle them in our current implementation. Handling crossings rigorously would require additional tracking/grouping algorithms for eigenvalues and eigenfunctions.

Reconstruction map. The reconstruction map (Eq. 6) reconstructs the spectral decomposition by projecting a linear combination of atoms' spectral decompositions back onto the manifold of spectral decompositions. In the general comment R1, we discuss the reconstruction map and its implications for the dictionary training. Essentially, the reconstruction map is not invariant to permutation but the training learns the atoms and the optimal permutations of atoms' spectral decompositions to optimize operator reconstructions. Therefore, we do not preprocess the spectral components before reconstruction; the training process itself determines this.

Training of the dictionary. The reconstruction map is not invariant to phase/scaling of the atoms' spectral components. Atoms are indeed represented by elements of N\mathcal{N}, however atoms are elements of the quotient space N/(C∗)r\mathcal{N}/(\mathbb{C}^*)^r and the optimization takes that difference into account by considering Riemannian optimization on the quotient space. The Riemannian manifold is detailed in Appendix C.2 (See Propositions 3-6), and the optimization algorithm is described in Appendix D.2.

Q2: Improvement of the Langevin experiment

We thank the reviewer for encouraging us to improve and clarify the Langevin experiment. In what follows we adress reviewer's questions.

  • What is the reference operator in the Langevin experiment? In 1D, the true spectrum can be computed accurately (e.g., Fokker–Planck eigenproblem on a fine grid).
  • Can you report errors against it, plus a metric not used for training, such as held-out forecasting error? Please add population-aware baselines: the training operator with the lowest risk (7), RRR/GR shrunk toward the population mean, and a meta-/multi-task Koopman method such as Iwata & Kawahara, 2021.

In this experiment there is no reference operator; we considered 256/256 training/transfer dynamical systems, each with a potential Uw(x)=w−4(x2−w2)2U_w(x) = w^{-4}(x^2-w^2)^2 where w∼U[0.5,1.2]w \sim U[0.5,1.2]. Upon reviewer suggestion we included the multi-task baseline MetaKoopman [a], and compared to the the ground truch operators with the relative error of the 1st/2nd eigenvalues and operator Hilbert-schmidt norm. We set DOODL with a dictionary of 4 atoms and also consider the RRR estimator [b] as baseline. Tables below describes the transfer results for trajectory length ranging from 100 to 10,000 samples. Table 1 & 2 account for the relative absolute error of eigenvalues and Table 3 for the relative Hilbert-Scmhidt error. For eigenvalues DOODL performs better than other baselines accross all trajectory lengths. For Hilbert-Schmidt norm DOODL performs better or equivalently to RRR estimator up to 5,000 samples. In the short-trajectory regime (100 samples), DOODL has a gain of 1 to 2 order of magnitude, an observation congruent with results in the main paper. This additional results are included in Appendix G.5 of the revised paper.

Table 1: Relative absolute error on 1st eigenvalue

Method1005001,0001,5002,0003,0005,0007,00010,000
DOODL0.304 ± 0.4990.0939 ± 0.06940.0694 ± 0.05040.0607 ± 0.04540.0575 ± 0.04130.0523 ± 0.03940.0482 ± 0.03440.0468 ± 0.03240.0452 ± 0.0319
MetaKoopman29.7 ± 25.211.5 ± 14.65.14 ± 9.222.5 ± 6.071.94 ± 5.31.01 ± 3.830.402 ± 1.20.248 ± 0.3640.171 ± 0.167
RRR35.1 ± 30.512.1 ± 15.65.33 ± 9.342.56 ± 6.052.05 ± 5.391.08 ± 3.890.451 ± 1.350.279 ± 0.5490.179 ± 0.186

Table 2: Relative absolute error on 2nd eigenvalue

Method1005001,0001,5002,0003,0005,0007,00010,000
DOODL0.302 ± 0.4870.101 ± 0.07680.0806 ± 0.06190.0722 ± 0.05760.0675 ± 0.0540.0629 ± 0.05170.0582 ± 0.04850.0569 ± 0.04630.0558 ± 0.0458
MetaKoopman3.84 ± 2.931.56 ± 2.020.82 ± 1.450.379 ± 0.6820.311 ± 0.6110.21 ± 0.4580.122 ± 0.1270.1 ± 0.08670.0864 ± 0.0668
RRR4.57 ± 3.631.73 ± 2.450.889 ± 1.680.408 ± 0.7050.35 ± 0.7140.233 ± 0.480.14 ± 0.150.111 ± 0.1010.0937 ± 0.073

Table 3: Relative HS norm error

Method1005001,0001,5002,0003,0005,0007,00010,000
DOODL0.181 ± 0.1420.126 ± 0.08160.125 ± 0.08260.125 ± 0.07740.125 ± 0.07730.125 ± 0.0760.124 ± 0.07490.125 ± 0.07650.125 ± 0.0748
MetaKoopman1.95 ± 1.60.717 ± 0.6470.434 ± 0.4660.293 ± 0.3080.24 ± 0.210.183 ± 0.1350.135 ± 0.06870.115 ± 0.04160.0974 ± 0.0315
RRR2.12 ± 1.920.914 ± 1.220.448 ± 0.5540.268 ± 0.2690.221 ± 0.1920.167 ± 0.1450.121 ± 0.06420.103 ± 0.03750.0892 ± 0.0273
  • How much of the short-trajectory gain is already obtained by the constant "DOODL avg" model?

Average DOODL is the reconstructed operator for the ponderation α=(1/d)i=1d\boldsymbol{\alpha} = (1/d)_{i=1}^d. It does not depend on the trajectory length, and according to Figure 2 of the main paper, the average performs better than RRR by one to two orders of magnitude in short trajectory regimes.

  • How was linear DL fitted (e.g., is the combined operator truncated back to rank r)?

For the linear DL baseline, the operators are represented by their matrix representation rather than their spectral decomposition. Atoms of the dictionary are matrices (possibly full rank); the reconstruction map is the linear combination, and the loss function is the Hilbert-Schmidt square norm. It yields a standard linear dictionary in the Hilbert space of operators. For a fair evaluation with the SGOT metric, spectral decompositions are re-estimated from the Linear DL operator according to the RRR procedure [b].

Q2: End

  • Please also add simple baselines for regime detection and parameter regression.

Switching regime detection is a proof of concept; we have no claim that it is a state-of-the-art method for change point detection. For the Langevin experiment, looking at the variance on short sliding windows is certainly enough to detect regime shifts. Our claim is that, through dictionary learning, operator representation becomes usable on short trajectories to visualize and identify changes in dynamics.

[a] Iwata, Tomoharu, and Yoshinobu Kawahara. "Meta-learning for koopman spectral analysis with short time-series." arXiv preprint arXiv:2102.04683 (2021).

[b] Kostic, Vladimir, et al. "Learning dynamical systems via Koopman operator regression in reproducing kernel Hilbert spaces." Advances in Neural Information Processing Systems 35 (2022): 4017-4031.

Q3: Clarification on the statistical guarantee

  • As written, the left-hand side is relative to the best α for the learned dictionary, so the bias term is not needed. Did you mean the optimal dictionary of (5)? Why is the bias nonnegative, given that the optimal dictionary minimizes the SGOT objective (5), not the risk R? How is the intrinsic dimension m defined?

In the infimum on the left-hand side of (9), the dictionary is fixed to the oracle dictionary Gˉ∗\bar{\mathcal{G}}^{\ast} of (5). The bias can have either sign. Proposition 7 and Remark 5 bound its absolute value.

The intrinsic dimension mm is defined through the covering numbers of the learned operator manifold. For every ε>0\varepsilon > 0, we require

log⁡N(G(Gˉ^),∥⋅∥op,ε)≤mlog⁡(1+8Ldecε).\log \mathcal{N}\left(\mathcal{G}(\hat{\bar{\mathcal{G}}}), \lVert \cdot \rVert_{\mathrm{op}}, \varepsilon\right) \le m \log\left(1 + \frac{8L_{\mathrm{dec}}}{\varepsilon}\right).

Since this set is the image of the simplex under a Lipschitz map (Lemma 3), m=d−1m=d-1 is always a valid choice (Lemma 4). A smaller mm requires an additional assumption on the learned dictionary. The definition is given in Appendix G.1.

  • Also, the theorem assumes a stationary test trajectory, while Section 4 uses short non-equilibrium ones. It also omits the approximation error of the manifold relative to the true operator, which drives the crossovers in Figs. 2 and 6.

We also thank the referee for raising the question of stationarity in Theorem 1. The new Appendix G.4 extends Theorem 1 to test trajectories started from an arbitrary initial distribution μ\mu for geometrically ergodic chains satisfying

∥μPt−π∥TV≤Cμρt.\lVert \mu P^t-\pi \rVert_{\mathrm{TV}} \le C_{\mu}\rho^t.

Proposition 8 retains the stationary risk as the target and adds a term of order

(1+M)2κ2Cμ(1−ρ)n.\frac{(1+M)^2\kappa^2 C_{\mu}}{(1-\rho)n}.

See Eq. (139). The proof retains the β\beta-mixing blocking argument of Theorem 1, which also applies to nonstationary trajectories. Only the centering of the empirical risk and the block variances change. The additional term vanishes as nn grows and is most relevant for windows shorter than the mixing time.

Q4: Plasma experiment desciption and additional baseline

Was the latent encoder (POSEIDON-B plus MLP heads) trained only on training simulations? Which regressor and protocol give R² = 0.81 / 0.97?

We synthesize the plasma experimental settings, as fully described in Appendix F. POSEIDON-B is a pre-trained foundation model that we fully describe in Appendix F.2. We then train from the frozen POSEIDON-B embedding, new MLP heads for multi-task dynamical system representation on the plasma dataset (see Appendix F.3-4). We trained a DOODL dictionary of 12 atoms on operators estimated from the training datasets (Appendix F.7). For the estimation of parameters, we considered degree-2 polynomial ridge regressors taking operator coordinates αi∈Δd−1\boldsymbol{\alpha}_i \in \Delta^{d-1} as inputs (Appendix F.7-8).

How does a simple baseline (e.g., regression on time-averaged latent features) compare?

Following the reviewer's suggestion, we included a regression on time-averaged POSEIDON-B features. Results performance reported as R2R^2 for DOODL and the baseline per trajectory length are depicted in the Table below. DOODL performs clearly better across all trajectory lengths. κ\kappa is rather well captured by the baseline for large trajectory length; however, unlike DOODL, the baseline fails to estimate gg.

methodparameter6508181029129516312052258332524092
DOODLg0.080.170.20.20.330.310.520.650.81
DOODLκ\kappa0.590.660.560.540.710.770.880.950.97
Avg. POSEIDONg-0.73-0.44-0.120.160.290.290.40.440.49
Avg. POSEIDONκ\kappa-0.39-0.160.130.410.580.720.780.840.86
Reviewer_bt13OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DOODL, which views spectral decompositions as elements of a quotient space and endows it with a Riemannian manifold struture. Dictionary learning then proceeds via Riemannian gradient methods by identifying as objective function the OT loss between each training operator and its reconstruction. Theorem 1 then bounds excess risk relative to the best operator within the dictionary manifold. Validation on 1D double-well Langevin dynamics and Tokam2D plasma turbulence show early operator recovery by DOODL better than per-system estimation.

Critical strengths

A well-motivated and carefully engineered pipeline. Working directly on spectral decompositions, with an OT loss that is invariant to eigenvector scaling and permutation, is a sensible way to preserve modes and time scales. The comparison with linear dictionary learning shows that respecting this geometry matters.

Clear practical benefit in the short-trajectory regime. Large gains over independent estimation at short lengths, together with the expected crossover at long lengths. The plasma experiment shows the pipeline scales to a nontrivial high-dimensional system.

Critical weaknesses

"Related dynamics" is never formalized. The low-dimensional-manifold hypothesis is assumed, without any condition on the family of generators (e.g., smooth parametric dependence, continuity of the spectral map, eigenvalue gaps) under which the operators actually concentrate near such a manifold. Theorem 1 measures how far the learned dictionary is from the oracle dictionary on the training population, but does not capture how far the new system’s operator is from the dictionary manifold.

Another important point is that the reconstruction map is acknowledged to depend on how spectral components are ordered across atoms -- and the proposed solution (increase the dictionary size) feels adhoc -- particularly given the danger of blindly increasing the dictionary size when dealing with Koopman-based approximations.

Questions for authors

Define "related.” Can you state conditions on a parametric family of generators, e.g., smoothness, spectral gaps, under which the operators concentrate near a low-dimensional manifold? Can you then bound the approximation error between a new system’s operator and the dictionary manifold?

Out-of-range systems. How does DOODL behave on test systems outside the training parameter range?

In Remark 5, how is the Lipschitz dependence of the dictionary on the training operators justified, given that (5) is nonconvex? If it is an assumption, could you state it explicitly?

Make sure to revise the list of references: e.g., J. Grizzle is not an author of the book "Hassan K Khalil and Jessy W Grizzle. Nonlinear systems, volume 3. Prentice hall Upper Saddle River, NJ, 2002."

4 comments

W1: Reconstruction by projection of linear combination of atoms

The reconstruction map is acknowledged to depend on how spectral components are ordered across atoms -- and the proposed solution (increase the dictionary size) feels adhoc.

The reconstruction map (Eq. 6) reconstructs the spectral decomposition by projecting a linear combination of atoms' spectral decompositions back onto the manifold of spectral decompositions. In the general comment R1, we discuss the reconstruction map and its implications for the dictionary training. Essentially, the reconstruction map is not invariant to permutations, but training learns the atoms and the permutations of their spectral decompositions to optimize operator reconstructions. When learning becomes difficult, it is either a lack of expressivity in the dictionary or difficulties in learning the permutations. In both cases, increasing the dictionary size helps similarly to training overcomplete disctionaries have been shown to help for data on manifolds .

Q1: Formalizing "related" systems and the approximation error

Can you state conditions on a parametric family of generators, e.g., smoothness, spectral gaps, under which the operators concentrate near a low-dimensional manifold? Can you then bound the approximation error between a new system’s operator and the dictionary manifold?

We thank the reviewer for this question, which led to the new Appendix G.5 (general comment R3). We define "related" as a smooth parametric family θ↦Gθ\theta\mapsto G_\theta, θ∈Θ⊂Rk\theta\in\Theta\subset\mathbb R^k, with a uniform spectral gap γ>0\gamma>0 among the rr leading eigenvalues. Under these conditions, the spectral decompositions form a kk-dimensional manifold (Proposition 9(i)), and a dictionary with d≍h−kd\asymp h^{-k} atoms approximates it within O(h2)=O(d−2/k)O(h^2)=O(d^{-2/k}) in operator norm (Proposition 9(ii)). Through Remark 6, this gives the requested bound on the distance of a new system to the dictionary manifold:

inf⁡αEπ∥(Gα−A)∗zt∥2≤κ2(12L0CSh2+Ldec′∥Gˉ^−Gˉh∥N+Lθdist(θnew,Θ))2.\inf_{\alpha} \mathbb{E}_{\pi} \lVert (G_{\alpha}-A)^{\ast} z_t \rVert^2 \le \kappa^2 \left( \frac{1}{2} L_0 C_S h^2 + L_{\mathrm{dec}}^{\prime} \lVert \hat{\bar{\mathcal{G}}}-\bar{\mathcal{G}}^h \rVert_{\mathcal{N}} + L_{\theta} \mathrm{dist}(\theta_{\mathrm{new}},\Theta) \right)^2.

We would like to be precise about what this does and does not show.

  • It is an expressivity result. The first term shows that the decoder class can approximate the population at rate O(d−2/k)O(d^{-2/k}). The second term is the distance between the learned dictionary and such a mesh dictionary, which the nonconvex problem (5) does not control. Proposition 7 and Remark 5 address this term only under the stability assumption, now stated explicitly (Q3).
  • The constants depend on the gap. CSC_S scales with 1/γ1/\gamma and 1/γ21/\gamma^2, so the bound is informative away from crossings and exceptional points, and degrades near them. We listed the cases where the hypothesis fails in Appendix G.5.
  • It answers W1. We agree with the reviewer that blindly increasing dd is risky: the estimation term of Theorem 1 grows as mlog⁡n/n\sqrt{m\log n/n} with m≤d−1m\le d-1, while the approximation term decays as O(d−4/k)O(d^{-4/k}). The dictionary size is thus a resolution parameter to be selected.

For the Langevin family of Section 4.1, all conditions hold exactly with k=1k=1. Varying the width is a pure spatiotemporal rescaling, λk(w)=λk(1)/w2\lambda_k(w)=\lambda_k(1)/w^2 and ϕk,w(x)=ϕk,1(x/w)\phi_{k,w}(x) = \phi_{k,1}(x/w) (Proposition 10).

Q2: Estimating systems outside of DOODL spanning manifold

How does DOODL behave on test systems outside the training parameter range?

We thank the reviewer for encouraging us to discuss the behavior of DOODL when estimating an operator lying outside the DOODL-spanning manifold. In the general response R2, we provide a detailed discussion and an additional experiment on that topic.

Essentially, with a dictionary of dd atoms Gˉ^\hat{\bar{G}}, the DOODL spanning manifold is the set of all reconstructible operators for coefficients α\boldsymbol{\alpha} on the simplex, i.e. S(Gˉ^)={Op(Bp(α,Gˉ^))∣α∈Δd−1}\mathcal{S}(\hat{\bar{\mathcal{G}}}) = \{\mathrm{Op}(\mathrm{B}^\mathrm{p}(\boldsymbol{\alpha}, \hat{\bar{\mathcal{G}}})) \mid \boldsymbol{\alpha} \in \Delta^{d-1}\}. For a new operator outside of the spanning range, i.e. A∉S(Gˉ^)A \notin \mathcal{S}(\hat{\bar{\mathcal{G}}}), DOODL returns an operator on the envelope of S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}) that is closest to AA in Hilbert-Schmidt norm. Therefore, for a system AA outside of the spanning manifold S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}), its estimation is biased by a term that is proportional to the distance of AA to S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}) in Hilbert-Schmidt norm. We illustrated DOODL's behavior when estimating new operators outside the spanning manifold in the Langevin experiment. Results in R2 show that for systems outside the spanning manifold, the coefficients α\boldsymbol{\alpha} concentrate on one corner of the 1D manifold, and the Hilbert-Schmidt error increases as the operator gets further away from the spanning manifold.

Q3: Lipschitz continuity of the dictionary learning objective

In Remark 5, how is the Lipschitz dependence of the dictionary on the training operators justified, given that (5) is nonconvex? If it is an assumption, could you state it explicitly?

We thank the reviewer for the question. Remark 5 concerns the stability of the minimizer of (5) with respect to the training operators. We now state this stability assumption explicitly. Up to the symmetries of the reconstruction map (a common reordering and rescaling of the spectral components of all atoms), we assume

∥Gˉ^−Gˉ∗∥N≲max⁡k∈[N]∥Gˉ^k−Gk∥N.\lVert \hat{\bar{\mathcal{G}}}-\bar{\mathcal{G}}^{\ast} \rVert_{\mathcal{N}} \lesssim \max_{k\in[N]} \lVert \hat{\bar{\mathcal{G}}}_k-\mathcal{G}_k \rVert_{\mathcal{N}}.

Such stability can hold locally when the oracle minimizer is non-degenerate modulo these symmetries. In general, the nonconvexity of (5) leaves this stability an assumption. Proposition 7 gives the bias bound

∣bias(Gˉ^)∣≲∥Gˉ^−Gˉ∗∥N.\lvert \mathrm{bias}(\hat{\bar{\mathcal{G}}}) \rvert \lesssim \lVert \hat{\bar{\mathcal{G}}}-\bar{\mathcal{G}}^{\ast} \rVert_{\mathcal{N}}.

The stability assumption is used only in Remark 5 to express this bound in terms of the training trajectory length. Theorem 1 and Proposition 7 hold without it.

Many thanks for the detailed response. The authors have addressed most of my concerns.

I appreciate the new formalization in Appendix G.5 and recommend moving some of it to the main text: e.g., assumptions (P1)–(P2) and the approximation bound (142). I would make it explicit that Proposition 9 shows that a good dictionary exists, not that the learned one is close to it. The O(d^{-2/k}) approximation error and the growth of the entropy dimension with d together imply a tradeoff in dictionary size, so please discuss how d should be chosen.

Finally, the out-of-range experiment would be more convincing with an RRR comparison, extrapolation in both directions, and a check of (P1)–(P2), or of the exceptional-point failure mode, on the non-reversible plasma system.

We thank the reviewer for the careful reading and for these suggestions. We have revised the paper accordingly and, due to the short remaing time for the rebuttal, we will complete the requested experiments for the camera-ready version, as elaborated below.

1. Main text. Section 3.3 now contains a paragraph stating (P1)–(P2) and the approximation bound (143), after the discussion of Theorem 1. It also discusses the choice of dd (point 3).

2. Existence versus learning. We agree, and now state this explicitly in Section 3.3 and after Proposition 9. Proposition 9 shows that a dictionary with approximation error O(d−2/k)O(d^{-2/k}) exists in the model class, not that the dictionary learned by the nonconvex problem (5) is close to it. The gap is the term Ldec′∥G‾^−G‾h∥NL'_{\mathrm{dec}}\|\widehat{\overline{\mathcal G}}-\overline{\mathcal G}^h\|_{\mathcal N} of (143), which we control only under the stability assumption of Remark 5.

3. Choice of dd. The new Remark 10 makes the trade-off explicit. With m=d−1m=d-1, balancing the approximation term O(d−4/k)O(d^{-4/k}) against the estimation term mlog⁡n/n\sqrt{m\log n/n} gives d≍(n/log⁡n)k/(k+8)d\asymp(n/\log n)^{k/(k+8)}. The optimal size therefore grows with the length nn of the test trajectories, and short-trajectory applications favour small dictionaries. Since the constants are unknown, Remark 10 also gives a selection rule: hold out training systems, fit α^\hat{\boldsymbol\alpha} by (7) on a prefix of length nn, and evaluate the one-step risk (8) on the remainder of the trajectory. This estimates the risk bounded in Theorem 1 without a reference operator. For the camera-ready version, we will report this criterion for d∈{2,…,15}d\in\{2,\dots,15\} and several nn on the Langevin family. Our current range d∈{2,…,5}d\in\{2,\dots,5\} lies in the regime where the approximation term dominates.

4. Out-of-range experiment. For the camera-ready version, we will add RRR on the same trajectories, extrapolation below the training range, and the RRR error on the full 4040k-sample trajectories as an estimate of the feature-space floor εr\varepsilon_r. Proposition 10 already predicts two features of the outcome:

  • Coordinates. The natural parameter is s=w−2s=w^{-2}, in which the eigenvalues are affine, λk=λk(1) s\lambda_k=\lambda_k(1)\,s. The training range w∈[0.5,1.2]w\in[0.5,1.2] is s∈[0.69,4]s\in[0.69,4]. The current extrapolation w∈[1.2,1.7]w\in[1.2,1.7] reaches only a distance 0.350.35 in ss, whereas w=0.3w=0.3 lies at distance 7.17.1. The two sides are therefore very asymmetric, and we will report errors against the distance in ss, the quantity entering the extrapolation term of (143).
  • Crossover with RRR. Since the error of RRR does not depend on the distance to the training range, while the extrapolation term of DOODL does not decrease with nn, the trajectory length beyond which RRR overtakes DOODL should decrease with this distance. The experiment will test this prediction.

5. Plasma system. For the camera-ready version, we will add an appendix checking (P1)–(P2) on the 400 estimated plasma operators. For each system, we compute the minimal spectral gap and the eigenvalue condition numbers ∥ℓi∥∥ri∥\|\ell_i\|\|r_i\|. Along the Delaunay graph of (g,κ)(g,\kappa), we compute the Lipschitz ratio of Hungarian-matched eigenvalues and the changes in the real/complex pattern. Exceptional points are separated from ordinary crossings by the conditioning of the colliding pair. Near an exceptional point, both eigenvalues become ill-conditioned with ∥ℓi∥∥ri∥≈c/∣λi−λj∣\|\ell_i\|\|r_i\|\approx c/|\lambda_i-\lambda_j|, whereas at an ordinary crossing the condition numbers stay bounded. We will therefore report the slope of the log condition number against the log gap over close pairs, and the correlation of the DOODL reconstruction error with the gap and the conditioning.

Comments on the paper

3 comments

We thank all reviewers for their thoughtful reviews, which helped us improve our work. In what follows, we address remarks that appear across several reviews. Our rebuttal includes a revised version of the paper with additional experiments in Appendix E.5-6, statistical guarantees in Appendix G.4, and a discussion of the manifold hypothesis in Appendix G.5. Changes are colored in blue.

R1: Reconstruction by projection of linear combination of atoms

DOODL reconstructs the spectral decomposition of a dynamical system operator representation by projecting a linear combination of atoms' spectral decompositions back onto the manifold of spectral decompositions. Let GG be a dictionary of dd atoms, with spectral decomposition (Λj,Lj,Rj)(\Lambda_j, L_j, R_j) for atom jj. Given coefficients α∈Δ(d−1)\alpha \in \Delta^{(d-1)}, the reconstruction is

Bp(α;G)=PN(∑j=1dαjΛj,∑j=1dαjLj,∑j=1dαjRj).B^{\mathrm{p}}(\alpha; G) = P_{\mathcal{N}}\left( \sum_{j=1}^d \alpha_j \Lambda_j, \sum_{j=1}^d \alpha_j L_j, \sum_{j=1}^d \alpha_j R_j \right).

The projector is defined by

PN(Λ,L,R)=(Λ,arg,minL~∗R=Ir∥L~−L∥F2,R).P_{\mathcal{N}}(\Lambda,L,R) = \left( \Lambda, \underset{\tilde{L}^{\ast}R=I_r}{\mathrm{arg\\,min}} \lVert \tilde{L}-L \rVert_F^2, R\right).

It restores feasibility by projecting the left eigenvectors to satisfy the bi-orthogonality constraint.

This reconstruction map is indeed sensitive to permutations of atoms' spectral components. However, these atoms are learned, and the training strategy accommodates the rigidity of the reconstruction map. Indeed, during training, the SGOT [a] metric compares reconstructed and population spectral decompositions independently of permutations. This helps the atoms self-organize their spectral decomposition so that the reconstructed and population operators are identical, up to a permutation, by the end of training. In other words, the training also learns the optimal permutations of atoms' spectral decompositions to optimize operator reconstructions. When learning the permutations is difficult, training can be simplified by increasing the dictionary size.

In future work, we plan to investigate a more flexible reconstruction map, namely the Wasserstein barycenter associated with the SGOT metric [a]. This reconstruction map is permutation invariant but faces important scaling issues. The BpB^{\mathrm{p}} reconstruction has complexity O(dpr+r3+r2p)\mathcal{O}(dpr + r^3 + r^2p), where dd is the size of the dictionary and r<pr < p are the dimensions of the matrices. In contrast, the current implementation of SGOT-Barycenter has complexity O(Nd(r2p+r3log⁡(r)))\mathcal{O}(Nd(r^2p + r^3\log(r))), where NN is the number of optimization steps (usually large). Considering such a reconstruction map would require new ways to perform barycenter optimization.

[a] Germain, Thibaut, et al. "A Spectral-Grassmann Wasserstein metric for operator representations of dynamical systems." International Conference on Learning Representations. Vol. 2026. 2026.

R2: DOODL behavior for operators outside of the spanning manifold.

Consider a dictionary of dd atoms Gˉ^\hat{\bar{G}} estimated from training operators, and the DOODL spanning manifold being the set of all reconstructible operators for coefficients α\boldsymbol{\alpha} on the simplex, i.e. S(Gˉ^)=Op(Bp(α,Gˉ^))∣α∈Δd−1\mathcal{S}(\hat{\bar{\mathcal{G}}}) = \\{\mathrm{Op}(\mathrm{B}^\mathrm{p}(\boldsymbol{\alpha}, \hat{\bar{\mathcal{G}}})) \mid \boldsymbol{\alpha} \in \Delta^{d-1}\\}. We aim to characterize the behavior of DOODL for estimating, from a sampled trajectory, a new operator outside of the spanning range, i.e. A∉S(Gˉ^)A \notin \mathcal{S}(\hat{\bar{\mathcal{G}}}).

According to equation 7 in section 3.3, DOODL returns the operator in the spanning manifold S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}) that is closest to AA in Hilbert-Schmidt norm. Hence, it leads to an operator on the manifold's envelope. Therefore, for a system AA outside of the spanning manifold S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}), its estimation is biased by a term that is proportional to the distance of AA to S(Gˉ^)\mathcal{S}(\hat{\bar{\mathcal{G}}}) in Hilbert-Schmidt norm.

We illustrated DOODL's behavior when estimating new operators outside the spanning manifold in the Langevin experiment. In this experiment, each training Langevin has a potential Vw(x)=w−4(x−w2)2V_w(x) = w^{-4}(x-w^2)^2 with w∼U[0.5,1.2]w \sim U[0.5,1.2]. We trained DOODL with a 4-atoms dictionary on 256 operators. As depicted in Figure 1, the learned manifold is linearly organized with almost only the 4th atom activated for systems with w≈1.2w \approx 1.2. Considering the DOODL-estimator, we then estimated new operators from trajectories of 1,000 samples associated with potentials outside of the learned manifold, i.e. w∈[1.2,1.7]w \in [1.2,1.7]. We tracked the activations and measured the relative HS norm with respect to the ground-truth operator. Results are reported in the Table below; for each bin, we sampled 100 potentials. As expected, the error grows proportionally with the parameter ww, and the activations are concentrated on the 4th atom. This experiment has been included in Appendix E.6

ww0.6–0.70.7–0.80.8–0.90.9–1.01.0–1.11.1–1.21.2–1.31.3–1.41.4–1.51.5–1.61.6–1.7
StatusInInInInInInOutOutOutOutOut
Relative HS norm0.091 ± 0.0150.113 ± 0.0110.124 ± 0.0060.104 ± 0.0090.083 ± 0.0050.087 ± 0.0080.096 ± 0.0040.109 ± 0.0050.125 ± 0.0050.143 ± 0.0070.167 ± 0.013
Atom 40.293 ± 0.0270.383 ± 0.0250.474 ± 0.0320.564 ± 0.0230.622 ± 0.0380.622 ± 0.0900.880 ± 0.0970.994 ± 0.0090.996 ± 0.0230.992 ± 0.0400.927 ± 0.150
Atom 30.116 ± 0.0260.212 ± 0.0180.265 ± 0.0150.290 ± 0.0120.303 ± 0.0340.360 ± 0.1050.120 ± 0.0970.0060.0040.0080.073 ± 0.150
Atom 20.258 ± 0.0210.285 ± 0.0060.238 ± 0.0230.145 ± 0.0270.075 ± 0.0150.017 ± 0.02000000
Atom 10.333 ± 0.0730.119 ± 0.0370.022 ± 0.0240.0010.0000.00000000

R3: When does the manifold hypothesis hold?

Conditions under which the operators of related systems concentrate near a low-dimensional manifold, and how far a new system can be from the DOODL manifold are now added in a new Appendix G.5. Here we summarize them.

1. Conditions (Proposition 9(i)). Consider a family θ↦Gθ∈Sr(H)\theta\mapsto G_\theta\in S_r(\mathcal H), θ∈Θ⊂Rk\theta\in\Theta\subset\mathbb R^k compact and convex, that is (P1) smooth in operator norm and (P2) uniformly spectrally separated: the rr leading eigenvalues are pairwise separated, and separated from 00, by a gap γ>0\gamma>0. Then there is a smooth choice of spectral decompositions S:θ↦(Λθ,Lθ,Rθ)∈NS:\theta\mapsto(\boldsymbol\Lambda_\theta,\mathbf L_\theta,\mathbf R_\theta)\in\mathcal N. The population is therefore a kk-dimensional manifold whenever θ↦Gθ\theta\mapsto G_\theta is injective and immersive. Assumption (P2) is the essential condition since constants scale with 1/γ1/\gamma and blow up at crossings.

2. Approximation error (Proposition 9(ii)–(iii)). A dictionary Gˉh\bar{\mathcal{G}}^h with atoms at the vertices of a mesh of Θ\Theta of size hh (so d≍h−kd \asymp h^{-k}) satisfies

\le \frac{1}{2} L_0 C_S h^2 = O(d^{-2/k}).$$ For a learned dictionary $\hat{\bar{\mathcal{G}}}$, this bound increases by $L'_{\mathrm{dec}}$ times its distance to the mesh dictionary. With Remark 6, this bounds the approximation term of DOODL by $$\kappa^2 \left( \frac{1}{2} L_0 C_S h^2 + L_{\mathrm{dec}}^{\prime} \lVert \hat{\bar{\mathcal{G}}} - \bar{\mathcal{G}}^h \rVert_{\mathcal{N}} + L_{\theta} \mathrm{dist}(\theta_{\mathrm{new}}, \Theta) \right)^2.$$ The last term accounts for the out-of-range effect discussed in reply R2. This is an expressivity statement. The second term measures the distance from the learned dictionary to the mesh dictionary, which the nonconvex problem (5) may fail to attain. **3. Our Langevin family is an exact curve (Proposition 10).** Since $U_w(x)=U_1(x/w)$, we have $\mathcal L_w=w^{-2}D_w\mathcal L_1D_w^{-1}$ with $D_w$ the (unitary) dilation: varying $w$ is a pure spatio-temporal rescaling. Hence $\lambda_k(w)=\lambda_k(1)/w^2$, $\phi_{k,w}=\phi_k(\cdot/w)$, and the time-scale ratios are invariant. The family satisfies (P1)–(P2), since one-dimensional reversible generators have simple spectrum. The scaling also gives an exact reference spectrum from one fine-grid computation. **4. Bifurcations.** For one-dimensional reversible diffusions, a change of the potential from a single to a double well creates no singularity in operator space: eigenvalues are simple, cannot cross, and depend analytically on the parameters. For $U_a=x^4/4-ax^2/2$, $a\in[-1,2]$, the first eigenvalue decreases smoothly from $1.43$ to $0.015$ (Kramers regime), while relative gaps stay above $0.30$ (Table 11). The curvature concentrates around the transition, where more atoms are needed. To conclude, we note that the manifold hypothesis fails when (P2) fails, wich is notably in the case of: - crossings in multi-parameter families, around which eigenvectors have non-trivial monodromy and the ordering-dependent decoder needs atoms on each side; a permutation-invariant decoder (R1) would remove this; - exceptional points in non-reversible families, where eigenvectors coalesce; - an eigenvalue entering the retained rank-$r$ block, e.g. when a new metastable state appears; These limits are now stated in Appendix G.5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.