acceptodds
Under review as a conference paper at ICLR 2027

Approximation of Dynamical Systems by the Kuramoto Model

Abstract

Analog computing leverages physics to perform computation energy efficiently in hardware. These systems typically have restrictive dynamics, and their expressive power is poorly understood. We address this question of expressivity for the Kuramoto model, an analog compute model that leverages oscillator coupling behavior. We establish a universal dynamical system approximation theorem: An autonomous Kuramoto network can approximate the flow of any continuously differentiable vector field on a torus over a finite time interval, uniformly over initial conditions, without a learned output map or external forcing. The result extends to smooth periodic coupling functions with a nonzero first Fourier harmonic. Experiments on seven dynamical systems validate our trajectory approximation approach, and flow-matching experiments show that Kuramoto networks can approximate learned flows, suggesting a promising use of the Kuramoto model in generative modeling.

Primary area:
learning theory
Keywords:
Analog computing, Kuramoto model, Expressivity, Universal approximation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qQBNOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors establish universal approximation of autonomous dynamical systems with the Kuramoto system. The authors first establish that the broader lagged fast-slow system (LFS), which allows for a per oscillator phase lag parameter as well as a time scale separation between two sets of oscillators can approximated by the Kuramoto system. This is given in Lemma 4.1 and constructs a system of 3 non-phase lagged oscillators whose attracting equilibria in the final oscillator gives the phase lagged oscillator and a time scaling η\eta on the system establishes convergence to that equilibria, giving an error term.

The authors then show that the LFS “hidden state” system has an attracting equilibria given by a “Normalized Affine Vector” and whose convergence to this equilibria is governed by the LFS time scale ϵ\epsilon. The “visible state” system is then shown to be a projection of the complex linear combination of NAVs onto the direction tangent to the unit circle at each visible state’s phase where the NAVs are determined by the hidden state system. Writing a target dynamical system’s velocity field as the projection of a velocity field TT into the direction tangent to the unit circle at each visible state’s phase reduces the problem to showing the NAVs converge to TT.

This convergence is established with density with Prop 4.1 and exploits density of complex exponentials (fourier series). The proof shows that the fourier basis can be approximated in the span of finite NAVs and gives a bound from the fourier series error and the finite span NAV error. The authors then apply an existing lemma relating vector field convergence to trajectory convergence and use that in combination with density of NAVs (prop. 4.1) and LFS convergence to the NAV equilibria (Lemma 4.2) in to give a universal approximation theorem for LFS systems in Prop. 4.2. Then finally a universal approximation theorem for the Kuramoto model (on trajectories) is established in Theorem 3.1 with the inclusion of the kuramoto’s convergence to equilibria LFS system of Lemma 4.1. The sine specific couplings of Theorem 3.1 and Prop. 4.2 are generalized to a general periodic coupling universal approximation theorem in Theorem 4.1. Section 5 then examines some experiments.

Critical strengths
  1. Universal approximation results of Prop. 4.2, Theorem 3.1 and Theorem 4.1 seem to be very novel and quite interesting. They also seem quite relevant and fundamental if the kuramoto model can be embedded in hardware and offer a new computing framework.

  2. The proof techniques are quite different from conventional universal approximation results as they require convergence to attracting equilibria.

  3. The overall paper is well organized and the inclusion of proof sketches was useful.

Critical weaknesses
  1. Expositional clarity for theoretical results needs improvement. The proofs, both in the appendix and the main text where the sketches appear could be written with more clarity and the appearance of certain terms/constructions could be given more clarity. For example in the proof of Lemma 4.1 it is not immediately obvious why specifically three additional oscillator need to be added per added phase lag oscillator. It becomes clear in the appendix when this gives contraction of to the phase lag through a trig identity. Additionally, r0r_0 should appear in the proof sketch. The inductive argument for the qkq_k bounds is hard to follow and could be expanded to be made more clear in the appendix. The proof of Lemma 4.2 in B.2 should justify the form of Eq. 4.10 in the proof sketch. Lemma B.1 and its use to show every finite combination of complex exponentials belongs to VV was hard to follow. More exposition on details such as these could be useful to a reader.

  2. No dependence on the oscillator network size discussed. The final convergence bound in the main text does not sufficiently discuss its dependence on the Kuramoto model’s system dimension. From my reading δ\delta determines the number of terms in the NAV which determines the number of number of hidden variable in the LFS system which determines the number of variables in the Kuramoto system (scaling the dimension of lagged variables by 3). I wouldn’t say that an exact dependence in the bound would be required, though that might be interesting, as it would potentially require further results, however, some additional discussion or hypotheses for how the bound of Theorem 3.1 depends on mm would be helpful.

Questions for authors
  1. Is it ever the case that a target dynamical system’s velocity field can be written as the projection of a velocity field TT into the direction tangent to the unit circle at each visible state’s phase?

  2. What are the specific training details for the dynamical systems fitting. I assume this is fitting via an ℓ2\ell_2 loss on the trajectories, as the authors mention t=0.5t=0.5, Is the Kuramoto model integrated and gradients are computed through integration?

  3. The proof strategy of Proposition 4.1 seems like it could potentially be tightened, Is an appeal to density of complex exponentials necessary? Could the result be determined directly? I ask this since it might simplify the resulting convergence term δ\delta and possibly allow that term to be parameterized by the number of NAVs.

  4. What are realistic time scale separations? Are the time scale separations that appear in the kuramoto model and LFS approximation bounds realistic?

  5. I have already mentioned this but what do the authors suspect the dependence of the bound on mm is?

2 comments

Thank you for your detailed comments. We will improve the clarity of the proofs in the next revision. We also consider the dependence of approximation rates on model size an important direction for future work. We answer your questions below:

  1. Could you please provide more details on the specific aspect you are interested in? From my current understanding, your description matches how we construct the target field. Set Ti(v)=−ieiviFi(v)T_i(v)=-ie^{iv_i}F_i(v). Then Im⁡(eiviTi(v)‾)=Fi(v)\operatorname{Im}(e^{iv_i}\overline{T_i(v)})=F_i(v) exactly, and no extra assumption on the target field is needed at this step.

  2. We train on vector field MSE, ∥F~(vr)−F(vr)∥22\|\widetilde F(v_r)-F(v_r)\|_2^2, rather than a trajectory loss. We sample states vrv_r uniformly from the domain and fit the NAV reduced field to the target velocities F(vr)F(v_r). The horizon t1=0.5t_1=0.5 is used only after training to simulate and evaluate the fitted system. Training therefore does not require gradients computed through integration. Our theoretical analysis provides a potential advantage over existing ML work using the Kuramoto model: They use BPTT to learn couplings (Wang et al., Unconventional AI) while we remove this training overhead in the cost of adding time-scale separation in hardware. Balancing these training and hardware costs is future work.

  3. & 5. We currently do not have a tighter construction to claim. We expect approximation rate arguments to require a different route than the universal approximation theorem and require restricting the target function to specific classes, similar to those in neural networks, e.g., Barron.

  4. The practical separation is task-dependent and can be potentially reduced by better training and model architecture. For example, we ran a follow-up experiment on the same distributional fitting tasks and reduced the required time-scale separation by 1000$\times$ with a two-layer NAV architecture. Please refer to CQ1 for practical hardware consideration and CQ2 for time-scale separation experiment details.

Reviewer_qQBNOpenReview ↗

Response to rebuttal comment

Thank you for the clarifications.

For question 1, I actually had a typo in my original question and is self-resolved. My main concerns involved the results regarding the tightness of prop 4.1 and the dependence of the number of oscillators, outside of exposition. I maintain my score.

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

This paper studies the universal approximation capabilities of the autonomous Kuramoto model for dynamical systems. It shows that, with hidden oscillators and appropriate initialization, the visible oscillator phases can approximate the finite-time trajectories of any continuously differentiable vector field on a torus, uniformly over initial conditions. The proof combines function approximation by linear combinations of normalized affine vectors (NAVs) with tracking of quasi-steady states by fast hidden oscillators. Phase lags are then realized using additional oscillators to obtain a standard Kuramoto model. The result is extended to C² periodic coupling functions with a nonzero first Fourier coefficient, and the construction is evaluated on seven dynamical systems and in flow-matching experiments.

Critical strengths

It is theoretically interesting that a system with such restricted interactions as the Kuramoto model can, with hidden oscillators and appropriate initialization, approximate the finite-time flow of any continuously differentiable vector field on a torus without a learned output transformation or external input.

The paper also provides an explicit construction of a standard Kuramoto model by combining the universal approximation capabilities of linear combinations of NAVs, a fast–slow construction based on tracking stable quasi-steady states, which can be viewed as a singular perturbation approximation, and approximation of phase lags using additional oscillators. Separately bounding the errors from function approximation, hidden-state tracking, and phase-lag realization is useful for understanding how the construction works and what determines its accuracy.

Critical weaknesses

Although the work is motivated by analog computing, the value of the proposed construction under realistic computational and hardware resource constraints is not sufficiently established. In particular, the generative experiments rely on extreme timescale separation, using ε = 10^-5 and η = 10^-6, corresponding to inverse-timescale coefficients of 10^5–10^6. Such a construction introduces stiffness into numerical integration and places substantial demands on hardware response speeds and coupling strengths, yet the paper provides no evaluation of execution time, integration step counts, or hardware feasibility. While the mathematical universal approximation result is interesting, the practical benefit of replacing a general-purpose function approximator with this Kuramoto network remains unclear.

Furthermore, although explicit parameter-dependent formulas are provided for Ch and Cη in the trajectory error bound's term δ + Chε + Cηη, the paper does not quantify the number of oscillators or coefficient magnitudes needed to achieve a prescribed δ, or how Ch and Cη grow as δ decreases. The authors acknowledge the limitations of an existence theorem, but it remains difficult to assess quantitatively the model size and timescale separation needed to improve overall accuracy. A gap therefore remains between the universal approximation existence result and practically attainable expressivity.

Questions for authors

Given the strong timescale separation required by this construction, how do you justify using the Kuramoto model rather than a general-purpose function approximator? In particular, please explain the benefits of the proposed method under the constraints of the intended analog hardware.

1 comment

We appreciate your comments on the practical aspects of computing with the Kuramoto model. Your questions are highly critical and central to the field and are relevant to readers interested in using the model, so we address them in the common questions (CQ1 & CQ2). A short summary here:

  1. As in CQ1, we agree that universality alone does not establish a practical advantage over existing general-purpose function approximators. Our focus is not trying to outperform other methods, but to fill the missing theoretical gap in computing with the Kuramoto model, complementing the hardware and algorithm research. Similar to the development of NNs, it was the joint effort of theory, algorithm, and hardware advances that made NNs practical. We hope our work can serve as a similar theory ground for computing with the Kuramoto model.
  2. In CQ2, we show that the time-scale separation required can be reduced depending on the task, training setup, and model architecture. We reduced the required time-scale separation by 1000$\times$ in a follow-up experiment using a two-layer NAV architecture. Both empirical aspects (better training and architecture) and theoretical aspects (approximation rate, model size, and tighter bound on time-scale separation) are important future directions.
  3. To clarify the intended hardware implementation, we target analog circuits in which oscillator phases evolve continuously, without numerical integration or time discretization. This avoids the numerical-integration stability constraints of digital simulation during hardware execution. However, we are aware that these benefits do not come for free: analog circuits have nonidealities such as noise, process variation, and random device mismatch, etc. Improving robustness to these effects will require further co-design of theory, algorithms, and circuit design.
Reviewer_dqCzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors show that the flows (solutions) of C^1 vector fields on the torus in any dimension can be approximated for a finite time by Kuramoto models, which are coupled oscillators that can be executed as physical circuits.

Critical strengths
  1. Clear presentation of the results and the context for the problem of ``physical'' circuit-based approximation of dynamical systems
  2. These ideas about analog circuits replacing traditional/deep learning methods for approximating ODE solutions seem novel, although I am not fully familiar with the literature in hardware for approximation of dynamical systems.
  3. The idea of designing slow-fast systems that a finite number of coupled oscillators can reproduce and then leveraging this to prove universality.
Critical weaknesses
  1. This was a counterexample raised by an LLM. Let v be the right hand side vector field for the Kuramoto system (all the coupled oscillators together). Then, Dv(s)1=0Dv(s) 1 = 0, and so only the vector fields with this property can be represented to arbitrary accuracy. Can the authors talk about why not? Or, if this obstruction can be removed by including an encoder (lifting to a different phase space), please include this as well.

  2. There is no mention of the intuition for why universality holds. Chaotic maps on the torus, for instance, may not look like slow-fast systems.

Questions for authors
  1. What do the circuit implementations of these circuits look like? Are there difficulties in learning the coefficients from a finite amount of data -- how does the sample complexity depend on the dimension?
  2. The paper links to a blog post by an AI company: are there implementations of this already? For what kind of dynamical systems do these implementations do better than deep neural networks?
1 comment

Thank you for your detailed feedback. We would like to address your questions and clarify the following points:

  1. With the fixed reference oscillators in our construction, the counterexample does not apply to the visible dynamics. A two-oscillator system can show this: s˙=sin⁡(s−b)\dot s=\sin(s-b) with bb being a fixed oscillator. Say we initialize b=0b=0, then Dv(s)1=cos⁡(s)≠1Dv(s)\mathbf{1}=\cos(s)\neq 1. The initialization map in Theorem 3.1 includes these fixed references.

  2. The target flow, including a chaotic flow on the torus, does not need to have a fast–slow structure. The slow visible part approximates the target flow, while the fast hidden part constructs nonlinear features. An intuitive analogy is a neural network with one hidden layer:

    \qquad \Psi_i(v)=\sum_a\alpha_{ia}u\left(\gamma_a+\sum_j c_{aj}e^{iv_j}\right).$$ The NAV nonlinearity $u$, computed with the fast hidden oscillators, is analogous to the NN activation $\sigma$. The visible couplings $\sum_a\alpha_{ia}$ form a linear readout similar to the output layer $\sum_a w_{ia}$ of a neural network. Therefore, the target itself need not be fast–slow.
  3. For circuit implementations of the Kuramoto model, academic chips have been taped out, and research is ongoing with electronic oscillator circuits and tunable coupling devices across different CMOS processes. Please refer to CQ1 for examples and their measured power and latency.

  4. To the best of our knowledge, we are not aware of any implementation that follows our specific construction. The AI company blog post reports an empirical study of image generation and does not provide general approximation guarantees.

  5. On learning difficulties and sample complexity, we have not yet established the theoretical approximation rates or sample-complexity bound. We consider these an important future direction, similar to NN approximation rate theory such as Barron. Empirically, we have observed that the training method and model architecture significantly affect the time scale separation needed for fitting. Please refer to CQ2 for more details.

  6. We acknowledge that the quantitative algorithmic benefit over NNs is not yet fully established. Please refer to CQ1 for the discussion.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to common reviewer questions

We thank reviewers for the detailed feedback and questions. We will address common questions now, and individual questions tomorrow.

CQ1 (Reviewer1&2): What are the circuit implementations of the proposed dynamical system and what is the benefit over general purpose function approximators or neural networks?

The Kuramoto model can be implemented with electronic oscillators connected through tunable coupling devices, with oscillator phases following the Kuramoto dynamics. Unlike simulation on a digital computer, the phases evolve in the continuous-time analog domain and, therefore, no numerical integration or time discretization is needed in the hardware.

Research on computing with the Kuramoto model is still in its infancy and we ackowledge the quantative algorithmic benefit over general purpose function approximators or NNs is not yet fully established. Our initial focus is not trying to outperform other methods, but to understand basic computability properties. With this paper, we now have confidence these systems can approximate functions, and have a construction that demonstrates this. We want to make an analogy to the universal approximation theorem for NNs (Cybenko). By the time the universal approximation theorem was established, NNs were not as widely used as they are today. It was the joint effort of theory, algorithm, and hardware advances that made NNs practical. We hope our work can serve as a similar theory ground for computing with the Kuramoto model.

On hardware benefits, academic oscillator chips have been taped out for solving Ising problems, which leverage steady-state behavior of Kuramoto dynamics similar to our hidden state dynamics, and have shown low power and low latency computation (Dikopoulos et al., Cılasun et al., Moy et al.). These designs span 28–65 nm CMOS processes and 50–1968 oscillators per chip. Their measurements provide motivation for our work. Moy et al. report 1968 oscillators operating at 1GHz frequency, with 42mW power, and convergence fewer than 50 cycles. Although the oscillators encodes an Ising model rather than NAVs, the measurement provides a ballpark hardware reference for our fast hidden state steup: 50 cycles gives a convergence time of 50÷150\div 1GHz$=50$ns, and energy consumption of 4242mW$\times50$ns$=2.1$nJ, demonstrating the potential for efficient circuit implementation.

On the algorithm side, works that use Kuramoto model for machine learning is emerging in the recent years (Wang et al., Unconventional AI). Our theory work complements them in two ways.

  • Prior works were done empirically. It is unclear whether a failure in one task indicates a fundamental limitation of the model or the result of a sub-optimal training setup, e.g., initialization, loss function, or hyperparameters. We provide a theoretical grounding showing the model does not have a fundamental limitation.
  • The proof yields an efficient training approach to trade off with hardware implementation cost. The existing oscillator ML works require solving ODEs and backpropagation through time (BPTT) which is computationally expensive. In constrast, the NAV training evaluate the reduce field and does not require BPTT, but requires time-scale separation. Balancing these trade-offs is a future work.

CQ2 (Reviewer2&3): Is the timescale separation realistic?

Timescale separation is required by our proof, but the amount needed in practice can be reduced substantially depending on the task, the training setup, and the model architecture.

We demonstrate this in a follow-up experiment using the same distribution benchmarks as in the paper. We train a two-layer NAV field similar to the architecture of a neural network with two hidden layers: The second hidden layer takes the first layer's NAV features as inputs; the visible dynamics read out its outputs. We further restrict the model to not have "lags" in the dynamics, which eliminates the timescale η\eta completely.

We simulate the Kuramoto model with time scale ε=10−3\varepsilon=10^{-3} and observe that relative endpoint W22W_2^2 error are 1.3%, 0.8%, 8.4%, and -0.3% on D1, D2, D3, and D4, respectively (lower is better fit). That is, the two-layer model maintains similar distributional quality using 1000$\times$ less time scale separation compared with η=10−6\eta=10^{-6} in the original submission. These results show that better architecture and training reduce the separation.

Combined with the projected 5050ns convergence time and 2.12.1nJ energy consumption of the fast hidden states in CQ1, the overall oscillator hardware execution can still be physically realizable and efficient.

Thank you for bringing up this perspective. Singular perturbation methods provide a useful framework for analyzing the fast–slow construction to approximate dynamical systems. Relevant works include the ones you mentioned, Vakulenko et al. 2015 and Vakulenko et al. 2016, follow-up works such as Kurtova et al. 2026, and works on chemical reaction networks by Anderson et al. 2021 and Dack et al. 2026.

Our work follows the same high-level construction, but making it work within the Kuramoto model is nontrivial.

On theoretical advances

We studied a different, tightly constrained class of ODEs compared with ODEs in prior works. Vakulenko et al. 2015 studied quadratic nonlinearities, while Vakulenko et al. 2016 and Kurtova et al. 2026 studied sigmoidal activations of affine mixtures. Both Anderson et al. 2021 and Dack et al. 2026 focused on polynomial mass-action dynamics. In contrast, the Kuramoto model uses a fixed sine of pairwise phase differences, so approximation results based on freely chosen affine mixtures or polynomial interactions do not transfer directly.

The restricted interactions require specific constructions: the lag network (Lemma 4.1) realizes phase shifts using ordinary Kuramoto dynamics, while the NAV network and its density result (Lemma 4.2 and Proposition 4.1) provide stable nonlinear features with the required approximation capability. Combining their error bounds yields the final trajectory approximation guarantee.

Together, we establish the first universal approximation theorem for dynamical systems using the Kuramoto model (Theorem 3.1) and generalize this result to coupling functions whose first Fourier coefficient is nonzero (Theorem 4.1). This establishes that fixed pairwise phase interactions suffice for universal approximation, providing an expressivity guarantee for Kuramoto dynamics that the cited results do not establish.

Thanks for bringing the works to our attention. We would be happy to expand the Related Work section to include the above discussion and related dynamical systems studied using singular perturbation methods. We appreciate this perspective, which would help us position our contribution more clearly within the broader literature.

Reviewer_zmsrOpenReview ↗

Asymptotic Scaling of the Construction and Comparison with Prior Work

Thank you for the explanation. I understand that, while the overall approach based on singular perturbation methods is shared with prior work, the contribution here lies in realizing the construction using lag networks and NAVs within the interaction constraints of the Kuramoto model.

Building on your explanation, I would like to understand the significance of the theoretical contribution more concretely. In terms of asymptotic scaling, how do the paper's estimates for the number of oscillators and the degree of timescale separation compare with those for existing constructions that approximate or realize dynamical systems using singular perturbation methods? Drawing on Table 2, Eq. (4.15), and Appendix B.4, could you discuss how these quantities depend on approximation accuracy and the dimension or complexity of the target system, to the extent that such comparisons are possible?

I understand that differences in state spaces and in the assumptions underlying approximation or realization may preclude a rigorous comparison under identical conditions. Taking these differences into account, could you clarify whether the existing estimates indicate improved asymptotic scaling, comparable scaling while satisfying the Kuramoto interaction constraints, or an additional asymptotic cost of imposing those constraints?

Where a comparison is difficult, an explanation of the reasons would also help clarify the significance of the construction and its results.

In light of this discussion, it would be helpful to incorporate these comparisons into the discussion of related work in the main text, clarifying how this work builds on or differs from prior research.

We thank the reviewer for the detailed feedback and suggestion. We hope the following clarifies that because of the scope and assumption differences with prior works, a fair and rigorous comparison with the other constructions is difficult.

Model size (number of oscillators)

Considering the fast dynamics under the singular perturbation methods, the model size can be decomposed as:

=\text{number of nonlinear features to approximate the target} \times\text{cost per nonlinear feature},$$ Using prior works as examples, Vakulenko et al. (2016) and Kurtova et al. (2026) use a sigmoid nonlinearity, with the cost of one hidden state per sigmoid. Chemical reaction network constructions from Anderson et al. (2021) and Dack et al. (2026) use a smoothed-ReLU nonlinearity, with the cost of one hidden state per smoothed-ReLU. **Cost per nonlinear feature.** The prior works that use singular perturbation methods count states in their dynamical models, but do not provide a hardware or circuit cost model to implement the differential equations. It is unclear the cost of implemeting the differential equation for one state variable. In our setting, we quantify the model size with hardware costs in terms of the oscillators count. These counts depend on the choice of hardware implementation. Given $n$ visible state and $m$ nonlinear features, the original Kuramoto requires $2+6n$ oscillators per feature. If one can implement LFS model directly in hardware (e.g., Sakaguchi Kuramoto model), the cost would be lower. The LFS model requires $2$ oscillators per feature. If the constant reference phases can be supplied directly as fixed biases, the LFS construction requires only $1$ oscillators per feature, which is the same rate as the prior works' cost of one hidden state per feature. We use the original Kuramoto model as our *hardware-oriented reference* and count constant reference phases as oscillators to give a conservative cost estimate. Reporting the different counts helps practitioners assess the trade-offs under their implementation constraints. A fair comparison of physical resources per nonlinear feature would require hardware implementation models from prior works, which are not the focus of their papers. **Number of nonlinear features to approximate the target.** This is a separate approximation rate question. Prior singular perturbation works do not need to establish this step themselves, as their nonlinear features are well-studied in the neural approximation literature. For example, the approximation steps in Vakulenko et al. (2016) and Kurtova et al. (2026) cite [Barron (1993)](https://doi.org/10.1109/18.256500). Because NAVs are a new nonlinear feature family we derived, there is no prior approximation rate theory that we can directly apply. Approximation rates typically require quantitative assumptions on the target class beyond the conditions sufficient for universality. For example, Cybenko (1989) establishes universality for general continuous targets, whereas Barron (1993) requires a bounded first Fourier moment assumption to obtain an approximation rate. As the first to introduce NAVs, we believe generality is of the most importance. Therefore, our construction is designed to characterize the expressiveness *as broadly as possible*, and we choose the target class as continuously differentiable functions without further restrictions. A fair comparison of approximation rates would require independent works specifying common target-class assumptions and deriving a corresponding NAV rate under those assumptions. We consider this an important next step. ## Timescale separation A similar difficulty applies when comparing timescale requirements. The error terms $C_h\varepsilon+C_\eta\eta$ depend on both the implementation and the chosen NAV approximation. - The lag approximation term $C_\eta\eta$ arises when realizing the LFS model using standard Kuramoto interactions and can be removed if one can directly implement the LFS model. - The constants $C_h$ and $C_\eta$ depend on the feature count and NAV parameterization used to approximate the target. Thus, we believe that establishing linear dependence on the timescale parameters (which matches the linear dependence on $\kappa$ in Vakulenko et al. (2016) up to model-dependent constants), is sufficient bound given the possible implementation choices and the general target class covered by our universality result. We include this comparison to the related work and discussion in the main text of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.