Pith. sign in

REVIEW 3 major objections 4 minor 24 references

Designing learning in high dimensional oscillator networks with low dimensional read-out

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that four oscillator populations are enough, and numerically necessary, for a Kuramoto-network reservoir to learn chaotic time series.

desk verdict The Lorenz results are the centerpiece, but the printed model only couples M=3 populations, so the P=4/6 claim doesn't follow from the equations as written. read the letter →

arxiv 2509.00848 v2 pith:MRY2UGL7 submitted 2025-08-31 nlin.CD math.DS

Thomas Geert de Jong This is my paper · ORCID
classification nlin.CDmath.DS MSC 34C1537D45
keywords reservoircomputingKuramotooscillatorsOtt-Antonsenansatztime-seriespredictionLorenzsystembifurcationanalysismean-fielddynamicsphysical
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper studies a reservoir computer made of many Kuramoto oscillators whose only trainable weights sit in the forcing term, so the oscillator network itself is never modified. The read-out is deliberately low-dimensional: for each population of oscillators, it uses the average phase rather than the full state. Working in the Ott-Antonsen continuum limit, the author maps the reservoir's performance onto bifurcations in coupling and forcing parameters, giving an exact success/failure diagram for one-dimensional linear inputs and showing how the average-phase read-out emerges naturally in finite networks. For the Lorenz system, the central result is numerical: four oscillator populations can sustain a chaotic attractor after a period-doubling bifurcation, six populations reproduce the Lorenz attractor with a leading Lyapunov exponent close to 0.906, and no configuration with two or three populations was found. If correct, this makes the number of oscillator populations a controlling design parameter for physical reservoir computers with robust, low-dimensional read-outs.

What carries the argument

The central object is the continuum-limit Kuramoto reservoir obtained through the Ott-Antonsen ansatz: instead of simulating thousands of phases, each oscillator population is reduced to a complex order parameter z = rho e^{i psi}, and the read-out takes h(psi) = [1, sin psi, sin^2 psi] (or [1, psi]). The ansatz turns reservoir design into a bifurcation problem in (F, K, omega0, P), and the argument that P is a controlling parameter is carried by the population-specific average-phase read-out.

What would settle it

Take the same CL-Kuramoto reservoir with P=2 or P=3, same read-out h(psi)=[1,sin psi,sin^2 psi], and scan (F,K,omega0,forcing assignment); if any configuration predicts the Lorenz attractor with a positive leading Lyapunov exponent near 0.9 for at least 100 time units, the 'at least 4' claim is false. A more direct numerical check on the P=4 case: locate the period-doubling bifurcation in K and verify the leading Lyapunov exponent crosses zero exactly there.

Watch

Extended reading notes

Core claim

In the continuum limit of a Kuramoto oscillator reservoir, each population is described by two mean-field variables: radius rho and average phase psi. The paper shows that a read-out using only psi, not the full phase distribution, can learn target time series. For a 1-D linear input, a complete bifurcation analysis in the (F,K) plane predicts exactly when training and testing succeed. For 2-D periodic input, success corresponds to a stable periodic orbit around the origin and failure to a stable fixed point. In a finite network with 2000 oscillators, training a read-out on individual phases produces weights that are nearly constant within each population, so the effective read-out becomes t

Load-bearing premise

The claim that at least four populations are necessary rests on the assumption that the failed searches for P=2 and P=3 were representative; the paper reports not finding successful configurations without specifying the search protocol, so a missed configuration would overturn the bound.

Editorial extensions

If this is right

  • For 1-D linear inputs the paper gives an exact parameter-space map of when the reservoir predicts, so success can be engineered a priori rather than tuned.
  • The average-phase read-out arises naturally from a component-wise phase read-out in finite networks, so low-dimensional read-outs are not an artificial restriction.
  • A chaotic target like Lorenz can be predicted with as few as four oscillator populations, with fidelity improving when each Lorenz component drives two populations (P=6).
  • Since the read-out is population-averaged, the scheme is robust to noisy or partial measurements of individual oscillator phases.
  • The number of populations, not just coupling and forcing strengths, is a primary design knob for learning dynamics.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication the paper leaves implicit: the 'at least four populations' claim is only as strong as the negative search for P=2 and P=3, so a systematic scan of (F, K, omega0, read-out basis, forcing assignment) could either confirm the bound or find lower-P solutions.
  • A natural testable extension is to use the same population-count recipe for other 3-D chaotic systems such as Rossler or Chua; if the four-population threshold holds there too, it may reflect a general requirement for chaotic learning rather than a Lorenz-specific accident.
  • In a physical implementation, four or six oscillator populations keep the read-out dimension at 8-12 numbers, suggesting a hardware design where only a few aggregate signals need to be measured.
  • The success of P=6 with pairs of populations per component suggests a symmetry-based design rule: duplicate the forcing of each coordinate to enrich the reservoir's response without adding read-out complexity beyond the population count.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper investigates a reservoir computer built from a forced Kuramoto oscillator network in which trainable weights appear only in the forcing term. Using the Ott-Antonsen ansatz, the authors derive a continuum-limit (CL) ODE for population order parameters and use this reduced model to analyze prediction tasks. For 1D linear input they give a full bifurcation analysis of training and testing; for 2D input they use symmetry reduction and bifurcation diagrams; for a finite-dimensional (FD) network they show that a component-wise phase read-out approximates the average-phase read-out of the CL model. For the Lorenz system they report that P=4 and P=6 CL reservoirs can sustain chaotic prediction-like dynamics, with the P=6 case giving a leading Lyapunov exponent close to Lorenz's, and they claim numerical evidence that at least 4 oscillator populations are necessary for such chaotic learning.

Significance. If the central claim holds, the paper offers a rare analytically tractable physical reservoir model with a low-dimensional average-phase read-out, and it identifies the number of oscillator populations as a key design parameter. The Ott-Antonsen reduction, the comparison of stability boundaries with the forced-Kuramoto literature, and the quantitative Lyapunov check against the Lorenz system are concrete strengths. However, the central 'at least 4 populations' claim rests on incomplete negative evidence and on a notational/coupling ambiguity that directly affects the P=4 and P=6 experiments.

major comments (3)
  1. [§2.2, Eq. (5) and Appendix A] The printed CL equations couple M populations, not P. In Eq. (5) and in the Appendix A derivation, the mean-field coupling term is (K/2M) Σ_{k=1}^M, where M is the target dimension. In the Lorenz experiments (§3.4), M=3 while P=4 or 6. As written, populations j>M are not coupled (except by the forcing term) and the normalization is by 2M, so the P=4 and P=6 simulations do not realize an all-to-all P-population reservoir. If the code sums over P, the equations in the paper are wrong and must be corrected; if the code follows Eq. (5), the experiments do not instantiate the claimed design. This ambiguity must be resolved before the central claim can be evaluated.
  2. [§3.4.2] The necessity claim 'at least 4 oscillator populations are necessary' is supported only by the sentence 'Numerically, we couldn't succeed for P=2, 3', with no search protocol. No ranges for F, K, ω0, read-out basis, forcing assignments, or number of trials are reported. This is a load-bearing negative result. Please either provide a systematic parameter search with clear negative evidence, or soften the claim in the abstract and conclusions to 'for the configurations considered, P=4 and P=6 succeed while P=2 and P=3 did not'.
  3. [§3.4] No prediction error (e.g., NMSE) is reported for the Lorenz task. Success is inferred from visual similarity and the leading Lyapunov exponent. Since the abstract says 'learn chaotic target dynamics', the paper should either report a quantitative prediction-error measure over the testing window or explicitly reframe the criterion as chaotic-attractor reproduction. As it stands, the reader cannot distinguish 'sustains a Lorenz-like attractor' from 'predicts the target time series'.
minor comments (4)
  1. [Figure 6] The caption uses 'M=4' and 'M=6', but the text and Table 4 use P=4 and P=6. Please correct the notation for consistency.
  2. [Table 4] The K row is ambiguous: for the CL(P,M)=(4,3) reservoir the entry appears as 'NA' or '55'? Please clarify the exact K value used for the P=4 case.
  3. [Eq. (8)] The closed-form expression for F is very complex and no derivation is shown. Consider moving it to the appendix or providing a brief outline, as the current placement interrupts the reading flow.
  4. [§3.3.2] The sentence 'W out_1,i and W out_1,i are approximately constant' should likely read 'W out_1,i and W out_2,i'. Please correct the typo.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the CL derivation is self-contained against standard Kuramoto/OA benchmarks; the main self-citation is framing-only, and the 'at least 4 populations' claim is an under-supported numerical generalization, not a circular reduction.

full rationale

The paper's derivation chain is not circular. Equations (5) are obtained from the standard forced Kuramoto model via the Ott-Antonsen ansatz and are checked against external results (Childs–Strogatz stability diagram, Antonsen et al. driven-oscillator analysis, the Lorenz attractor). The read-out weights W_out are trained by ridge regression on training segments and then evaluated on the autonomous testing phase; success/failure is measured by NMSE and Lyapunov exponents, not by the fitted weights themselves, so there is no fitted-input-called-prediction step. The only self-citation is to the authors' 'omnipresent computing' framework [dJNN25], which supplies the setup (trainable forcing weights) and is not the source of the quantitative bifurcation or chaos results. Two non-circular weaknesses are worth flagging: (1) Section 3.4.2 states 'Numerically, we couldn't succeed for P = 2, 3' without giving the search protocol, so the abstract's 'at least 4 oscillator populations are necessary' is an overclaim from absence of counterexamples rather than a demonstrated lower bound; and (2) Eq. (5) and Appendix A define the coupling sum over M (the target dimension), not P, while Section 3.4 uses M=3 with P=4 or 6 — so the printed equations do not implement an all-to-all reservoir over P populations, and the extra populations may only be enlarging the read-out. These are evidence/correctness concerns, not circular reductions, and do not raise the circularity score beyond a minor framing self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central results depend on the OA ansatz, Cauchy distributions, the symmetric P=M=2 ansatz, and hand-chosen reservoir parameters (F, K, omega0, Delta1). No new physical entities are introduced. The 'at least 4 populations' claim additionally depends on an unreported search over configurations.

free parameters (5)
  • Coupling constant K for chaotic tests = varied; P=4 range 3.5 to 5, P=6 labeled 50/80
    Not fitted to the target, but the period-doubling and chaotic regime is found by manual continuation; the success of P=4/P=6 cases depends on K selection.
  • Forcing amplitude F = 55
    Hand-chosen for the Lorenz tasks (Table 4); success in P=4/P=6 cases was found at this value, and failure at other F is not systematically mapped.
  • Natural frequency locations omega0_i = Table 4 vectors
    Hand-specified per population; no rule is given for choosing them, and the 4/6 population results depend on these choices.
  • Input rescaling c = 0.01
    Listed in Table 4 for the Lorenz tasks; acts as a time or amplitude scaling of the forced input.
  • Cauchy width Delta1 = 1 for CL, 0.01 for FD
    Set by hand; for FD, the small Delta is chosen to reduce Cauchy-tail weights, affecting the emergence result in Section 3.3.
assumptions (4)
  • domain assumption Ott-Antonsen ansatz: phase density Fourier modes obey f_j^(n)=alpha_j^n and alpha_j analytically continues into the lower half-plane
    All CL reservoir equations (5) and the subsequent bifurcation analysis rest on this reduction; the author notes in the conclusion that it removes solutions.
  • domain assumption All-to-all coupling and Cauchy-distributed natural frequencies for each population
    The mean-field reduction in Appendix A and the symmetry analysis in Section 3.2 assume this interaction structure.
  • ad hoc to paper Symmetric reduction for P=M=2: u1=-u2, omega01=-omega02, psi1=-psi2, rho1=rho2
    Section 3.2 reduces the state space to two variables; the analysis does not cover broken-symmetry solutions.
  • domain assumption Fixed read-out families h(psi)=[1,psi] and h(rho,psi)=[1,sin(psi),sin^2(psi)] are assumed rather than derived
    The 'necessary populations' conclusion is conditional on this read-out choice; other readouts could change the threshold.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Designing learning in high dimensional oscillator networks with low dimensional read-out." pith.science (2026). https://pith.science/paper/MRY2UGL7

@misc{pith2026250900848,
  author       = {Pith},
  title        = {Pith review of: Designing learning in high dimensional oscillator networks with low dimensional read-out},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRY2UGL7}},
  note         = {Machine review of arXiv:2509.00848}
}
read the original abstract

In this paper we investigate a oscillator network based reservoir computer with a large number of oscillators and a low dimensional read-out. The read-out is a function on the average phases with respect to each oscillator population. Hence, this read-out provides a robust measurement of the oscillator states. We consider a low number of populations which leads to a low-dimensional read-out. Here, the task is time-series prediction. The input time-series is introduced via a forcing term. After a training phase the input is learned. Importantly, the training weights are introduced in the forcing term meaning that the oscillator network is left untouched. Hence, we can apply classical methods for oscillator networks. Here, we consider the continuum limit for Kuramoto oscillators by using the Ott-Antonsen Ansatz. Consequently, a mean field reservoir computer arises. The success and failure of the reservoir computer is then studied by bifurcations in the coupling and forcing parameter space. We will also show that the average phase read-out can naturally arise when considering the read-out on the phase states. Finally, we give numerical evidence that at least 4 oscillator populations are necessary to learn chaotic target dynamics.

Figures

Figures reproduced from arXiv: 2509.00848 by the authors.

Figure 1
Figure 1. Overview deployment autonomous reservoir computer: The reservoir evolution is described by an u-driven ODE. Here the input is given by a solution of Lorenz system. i) During wipe￾out transients are removed by evolving the u-driven reservoir. ii) During training the u-driven reservoir is also evolved but a linear weight matrix, Wout is trained so that the read out of the states, h(r), is fitted to u. iii) During test… view at source ↗
Figure 2
Figure 2. Overview example for relation between FD oscillator network and CL oscillator network: i) For the FD oscillator network we consider M = 3 and P = 4 with ℓ = (1, 2, 3, 3). Note that the populations driven by u3 are not identical since g3, g4 are different. ii) In an infinite dimensional setting we can describe the oscillator populations by density functions over the phases and natural frequencies. iii) Using the Ott-… view at source ↗
Figure 3
Figure 3. Bifurcation diagrams for Ω = 1: i) In the training phase the Hopf, Saddle-Node, SNIPER bifurcation divide the parameter domain into the subdomains A,B,C,D. In A,B,D training is successful. ii) In the testing phase the green and the magenta curve subdivide the parameter domain into the subdomains E,F,G. In E,F testing is successful if training is successful. Hence, the prediction is successful for (A ∪ B ∪ D) ∩ (E ∪ … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: CL-reservoir for symmetric linear u: i) The intersection of the boundary between success and failure with the line F = 0 coincides with a saddle node bifurcation. ii) For successful parameter choices the prediction exhibits small oscillations around the target. The lim…
Figure 5
Figure 5. Figure 5: FD-RC for periodic motion on a circle in the plane: i,ii) The prediction is successful but exhibits small oscillations which appear to be bounded over time. iii) Oscillator populations have the same frequency but are translated in time differently in correspondence wit…
Figure 6
Figure 6. Figure 6: CL-RC for chaotic input: i) We consider 4 oscillator populations and graph the evolution of the oscillator’s leading Lyapunov exponent versus K with fixed F. We note that we only consider the first two decimals of the Lyapunov exponent. Between K = 3.5 and K = 5 a peri…
Figure 7
Figure 7. Figure 7: Supporting bifurcation diagrams 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

  1. [1]

    Antonsen, R.T

    T.M. Antonsen, R.T. Faghih, M. Girvan, E. Ott, and J. Platig. External periodic driving of large systems of globally coupled phase oscillators. Chaos: An Interdisciplinary Journal of Nonlinear Science , 18(3), 2008

    work page 2008

  2. [2]

    Energy and AI

    International Energy Agency. Energy and AI . Technical report, IEA, 2025. https://www.iea.org/reports/energy-and-ai

    work page 2025

  3. [3]

    Baltussen, T.J

    M.G. Baltussen, T.J. de Jong, Q. Duez, W.E. Robinson, and W.T.S. Huck. Chemical reservoir computation in a self-organizing reaction network. Nature , 631(8021):549--555, 2024

    work page 2024

  4. [4]

    Childs and S.H

    L.M. Childs and S.H. Strogatz. Stability diagram for the forced kuramoto model. Chaos: An Interdisciplinary Journal of Nonlinear Science , 18(4):043128, 2008

    work page 2008

  5. [5]

    de Castro

    L.N. de Castro. Fundamentals of natural computing: an overview. Physics of Life Reviews , 4(1):1--36, 2007

    work page 2007

  6. [6]

    de Jong, N

    T.G. de Jong, N. Akashi, T. Taniguchi, H. Notsu, and K. Nakajima. Virtual reservoir acceleration for cpu and gpu: Case study for coupled spin-torque oscillator reservoir. In NeurIPS: ML with New Compute Paradigms Workshop . NeurIPS, 2023

    work page 2023

  7. [7]

    Harnessing omnipresent oscillator networks as computational resource

    T.G. de Jong, H. Notsu, and K. Nakajima. Harnessing omnipresent oscillator networks as computational resource. https://arxiv.org/abs/2502.04818 , 2025

    work page Pith review arXiv 2025

  8. [8]

    Goralski and T.K

    M.A. Goralski and T.K. Tan. Artificial intelligence and sustainable development. The International Journal of Management Education , 18(1):100330, 2020

    work page 2020

Show all 24 references
  1. [9]

    A.G. Hart. Generalised synchronisations, embeddings, and approximations for continuous time reservoir computers. Physica D: Nonlinear Phenomena , 458:133956, 2024

    2024

  2. [10]

    Hoerl and R

    A.E. Hoerl and R. W. Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics , 12(1):55--67, 1970

    1970

  3. [11]

    echo state

    H. Jaeger. The “echo state” approach to analysing and training recurrent neural networks-with an erratum note. Bonn, Germany: German National Research Center for Information Technology GMD Technical Report , 148(34):13, 2001

    2001

  4. [12]

    Jaeger and H

    H. Jaeger and H. Haas. Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication. science , 304(5667):78--80, 2004

    2004

  5. [13]

    Keller and M

    T.A. Keller and M. Welling. Neural wave machines: Learning spatiotemporally structured representations with locally coupled oscillatory recurrent neural networks. In Proceedings of Machine Learning Research , volume 202, pages 16168--16189. PMLR, 2023

    2023

  6. [14]

    E.N. Lorenz. Deterministic nonperiodic flow. Journal of atmospheric sciences , 20(2):130--141, 1963

    1963

  7. [15]

    Miyato, S

    T. Miyato, S. L \"o we, A. Geiger, and M. Welling. Artificial kuramoto oscillatory neurons. arXiv preprint arXiv:2410.13821 , 2024

    2024 arXiv

  8. [16]

    Nakajima

    K. Nakajima. Physical reservoir computing—an introductory perspective. Japanese Journal of Applied Physics , 59(6):060501, 2020

    2020

  9. [17]

    Nakajima, K

    K. Nakajima, K. Fujii, M. Negoro, K. Mitarai, and M. Kitagawa. Boosting computational power through spatial multiplexing in quantum reservoir computing. Physical Review Applied , 11(3):034021, 2019

    2019

  10. [18]

    Ott and T.M

    E. Ott and T.M. Antonsen. Low dimensional behavior of large systems of globally coupled oscillators. Chaos: An Interdisciplinary Journal of Nonlinear Science , 18(3), 2008

    2008

  11. [19]

    Ricci, M

    M. Ricci, M. Jung, Y. Zhang, M. Chalvidal, A. Soni, and T. Serre. Kuranet: systems of coupled oscillators that learn to synchronize. arXiv preprint arXiv:2105.02838 , 2021

    2021 arXiv

  12. [20]

    Sebastian, M

    A. Sebastian, M. Le Gallo, R. Khaddam-Aljameh, and E. Eleftheriou. Memory devices and applications for in-memory computing. Nature nanotechnology , 15(7):529--544, 2020

    2020

  13. [21]

    Strogatz

    S.H. Strogatz. From kuramoto to crawford: exploring the onset of synchronization in populations of coupled oscillators. Physica D: Nonlinear Phenomena , 143(1-4):1--20, 2000

    2000

  14. [22]

    Todri-Sanial, C

    A. Todri-Sanial, C. Delacour, M. Abernot, and F. Sabo. Computing with oscillators from theoretical underpinnings to applications and demonstrators. Npj unconventional computing , 1(1):14, 2024

    2024

  15. [23]

    Vandoorne, P

    K. Vandoorne, P. Mechet, T. Van Vaerenbergh, M. Fiers, G. Morthier, D. Verstraeten, B. Schrauwen, J. Dambre, and P. Bienstman. Experimental demonstration of reservoir computing on a silicon photonics chip. Nature communications , 5(1):3541, 2014

    2014

  16. [24]

    Von Neumann

    J. Von Neumann. First draft of a report on the edvac. IEEE Annals of the History of Computing , 15(4):27--75, 1993

    1993

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.