pith. sign in

arxiv: 2604.05967 · v1 · submitted 2026-04-07 · 💻 cs.LG · math.DS· math.OC

On Dominant Manifolds in Reservoir Computing Networks

Noa Kaplan , Alberto Padoan , Anastasia Bizyaeva This is my paper

Pith reviewed 2026-05-10 20:10 UTC · model grok-4.3

classification 💻 cs.LG math.DSmath.OC
keywords reservoir computingdominant manifoldskoopman eigenfunctionsdynamic mode decompositiontime series forecastingrecurrent neural networksdynamical systemslinear reservoir model
0
0 comments X

The pith

For training data from autonomous dynamical systems, the dominant modes in a trained reservoir computing network approximate the Koopman eigenfunctions of the original system.

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

The paper shows that in a simplified linear continuous-time reservoir model, training for forecasting tasks causes low-dimensional dominant manifolds to emerge in the network state space. These manifolds are directly tied to the intrinsic dimensionality and information content of the training data. When the data comes from an autonomous dynamical system, the dominant modes relate to approximations of the Koopman eigenfunctions, creating an explicit link between reservoir computing and the Dynamic Mode Decomposition algorithm. This matters because it provides a geometric explanation for how training shapes recurrent network dynamics to capture system behavior.

Core claim

In the linear continuous-time reservoir, the eigenvalue motion during training generates dominant manifolds whose structure corresponds to the Koopman eigenfunctions of the driving dynamical system, thereby connecting reservoir computing to dynamic mode decomposition.

What carries the argument

The dominant manifolds formed by the trained reservoir modes, which approximate Koopman eigenfunctions through the reservoir's linear dynamics.

Load-bearing premise

That the simplified linear and continuous-time reservoir model captures the essential geometry and behavior of the nonlinear discrete-time reservoirs used in practice.

What would settle it

Simulating the training process on data from a known dynamical system like the Lorenz attractor and checking if the trained reservoir's dominant modes match the Koopman eigenfunctions computed via dynamic mode decomposition; mismatch would falsify the relation.

Figures

Figures reproduced from arXiv: 2604.05967 by Alberto Padoan, Anastasia Bizyaeva, Noa Kaplan.

Figure 1
Figure 1. Figure 1: Plot of the closed-loop spectrum of JT(t ′ ), for Goldbeter oscillator system. Black star (left) marks the open-loop poles of λ[J0] (at t ′ = 0), colored paths show the closed-loop shifted eigenvalues as t ′ > 0 varies (left to right in time), marked with black × at the last timestep. The reservoir is initialized with n = 100, γ = 2.4, σb = 1, and ω = 0.8 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the evolution of RC spectrum (black points) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the prediction of linear (blue dashed) and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Understanding how training shapes the geometry of recurrent network dynamics is a central problem in time-series modeling. We study the emergence of low-dimensional dominant manifolds in the training of Reservoir Computing (RC) networks for temporal forecasting tasks. For a simplified linear and continuous-time reservoir model, we link the dimensionality and structure of the dominant modes directly to the intrinsic dimensionality and information content of the training data. In particular, for training data generated by an autonomous dynamical system, we relate the dominant modes of the trained reservoir to approximations of the Koopman eigenfunctions of the original system, illuminating an explicit connection between reservoir computing and the Dynamic Mode Decomposition algorithm. We illustrate the eigenvalue motion that generates the dominant manifolds during training in simulation, and discuss generalization to nonlinear RC via tangent dynamics and differential p-dominance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper studies the emergence of low-dimensional dominant manifolds in Reservoir Computing (RC) networks for temporal forecasting. For training data generated by an autonomous dynamical system, it claims that in a simplified linear continuous-time reservoir model the dominant modes of the trained reservoir approximate the Koopman eigenfunctions of the original system, thereby establishing an explicit link between RC and the Dynamic Mode Decomposition (DMD) algorithm. The work illustrates the eigenvalue motion that produces these manifolds during training and sketches a generalization to nonlinear discrete-time RC via tangent linearization and differential p-dominance.

Significance. If the central claims hold, the manuscript would supply a concrete geometric explanation for how training aligns RC dynamics with the intrinsic structure of the data-generating system, directly connecting RC to Koopman theory and DMD. This could clarify why RC succeeds on dynamical forecasting tasks and inform reservoir design. The limitation to a linear continuous-time model, however, restricts immediate applicability to the nonlinear discrete-time reservoirs used in practice.

major comments (2)
  1. [Abstract; eigenvalue motion and DMD equivalence sections] Abstract and sections on eigenvalue motion and DMD equivalence: the explicit relation between trained-reservoir dominant modes and Koopman eigenfunctions is derived only for the linear continuous-time reservoir; the paper provides no error bounds showing that the tangent-linearization argument preserves this relation under training of nonlinear discrete-time reservoirs.
  2. [Generalization to nonlinear RC] Generalization discussion: the appeal to differential p-dominance and tangent dynamics supplies no quantitative demonstration that higher-order nonlinearities (e.g., saturation or strong attractor curvature) leave the dominant-manifold geometry unchanged, leaving the transfer of the Koopman/DMD connection to practical RC unproven.
minor comments (2)
  1. [Abstract] The abstract states that the results are illustrated 'in simulation' but does not specify the reservoir dimension, integration method, or quantitative metrics used to verify the eigenvalue motion; adding these details would improve reproducibility.
  2. [Preliminaries] Notation for the continuous-time linear reservoir (e.g., the precise definition of the state matrix and input coupling) should be introduced once in a dedicated preliminary section rather than piecemeal.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which correctly identify the scope of our derivations. We address each major point below by clarifying the manuscript's claims and indicating targeted revisions to improve precision without altering the core contributions.

read point-by-point responses

  1. Referee: Abstract and sections on eigenvalue motion and DMD equivalence: the explicit relation between trained-reservoir dominant modes and Koopman eigenfunctions is derived only for the linear continuous-time reservoir; the paper provides no error bounds showing that the tangent-linearization argument preserves this relation under training of nonlinear discrete-time reservoirs.

    Authors: The referee is correct that the explicit equivalence between the dominant modes of the trained reservoir and approximations of the Koopman eigenfunctions is derived rigorously only for the linear continuous-time reservoir model. The tangent-linearization argument for nonlinear discrete-time reservoirs is presented as a sketch for generalization and does not include error bounds or convergence guarantees. We will revise the abstract and the eigenvalue motion and DMD equivalence sections to state explicitly that the Koopman/DMD connection is proven for the linear case, while the nonlinear extension remains a conceptual outline without quantitative error analysis. revision: partial

  2. Referee: Generalization discussion: the appeal to differential p-dominance and tangent dynamics supplies no quantitative demonstration that higher-order nonlinearities (e.g., saturation or strong attractor curvature) leave the dominant-manifold geometry unchanged, leaving the transfer of the Koopman/DMD connection to practical RC unproven.

    Authors: We agree that the generalization section relies on differential p-dominance and tangent dynamics without providing quantitative evidence, such as bounds or numerical experiments, that higher-order nonlinear effects like saturation or strong curvature preserve the dominant-manifold structure. The manuscript does not claim a full transfer to practical nonlinear reservoirs. We will expand the discussion to more explicitly acknowledge this limitation and frame the linear-case analysis as a foundational step, while noting the need for future work on robustness to nonlinearities. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation is analytical for linear model

full rationale

The paper derives the link between dominant reservoir modes and Koopman eigenfunctions explicitly from the equations of the simplified linear continuous-time reservoir, relating them to DMD via eigenvalue analysis rather than any fitted quantity redefined as a prediction. Generalization to nonlinear discrete-time RC is discussed separately via tangent dynamics and p-dominance without claiming it reduces to the linear case by construction. No self-definitional steps, load-bearing self-citations, or ansatz smuggling appear in the derivation chain. The central claim follows from the model's dynamics and is self-contained against external benchmarks like DMD.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract does not introduce or rely on any free parameters, new axioms, or invented entities beyond standard concepts already present in reservoir computing and Koopman theory.

pith-pipeline@v0.9.0 · 5431 in / 1089 out tokens · 41685 ms · 2026-05-10T20:10:35.864109+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Koopman Identification of Nonlinear Systems via Reservoir Liftings

    cs.LG 2026-05 unverdicted novelty 6.0

    RC-Koopman uses reservoir computing as a stateful Koopman dictionary with spectral radius controlling temporal memory to achieve accurate and stable identification of nonlinear systems.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication,

    H. Jaeger and H. Haas, “Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication,”Science, vol. 304, no. 5667, pp. 78–80, 2004

    work page 2004

  2. [2]

    Model-free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach,

    J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott, “Model-free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach,”Phys. Rev. Lett., vol. 120, p. 024102, Jan 2018

    work page 2018

  3. [3]

    Forecasting chaotic systems with very low connectivity reservoir computers,

    A. Griffith, A. Pomerance, and D. Gauthier, “Forecasting chaotic systems with very low connectivity reservoir computers,”Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 29, p. 123108, 12 2019

    work page 2019

  4. [4]

    Emerging opportunities and challenges for the future of reservoir computing,

    M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, and J. Sun, “Emerging opportunities and challenges for the future of reservoir computing,”Nature Communications, vol. 15, p. 2056, 03 2024

    work page 2056

  5. [5]

    The “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, vol. 148, 01 2001

    work page 2001

  6. [6]

    Real-time computing without stable states: A new framework for neural computation based on perturbations,

    W. Maass, T. Natschl ¨ager, and H. Markram, “Real-time computing without stable states: A new framework for neural computation based on perturbations,”Neural Computation, vol. 14, no. 11, pp. 2531– 2560, 11 2002

    work page 2002

  7. [7]

    Searching for principles of brain computation,

    W. Maass, “Searching for principles of brain computation,”Curr. Opin. Behav. Sci., vol. 11, pp. 81–92, 2016

    work page 2016

  8. [8]

    Invertible generalized synchronization: A putative mechanism for implicit learning in neural systems,

    Z. Lu and D. S. Bassett, “Invertible generalized synchronization: A putative mechanism for implicit learning in neural systems,”Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 30, no. 6, p. 063133, 06 2020

    work page 2020

  9. [9]

    A neural manifold view of the brain,

    M. Perich, D. Narain, and J. Gallego, “A neural manifold view of the brain,”Nature Neuroscience, vol. 28, no. 8, pp. 1582–1597, Aug. 2025

    work page 2025

  10. [10]

    A dissipativity theorem for p-dominant systems,

    F. Forni and R. Sepulchre, “A dissipativity theorem for p-dominant systems,” inIEEE Conference on Decision and Control, 2017, pp. 3467–3472

    work page 2017

  11. [11]

    A geometric characteri- zation of the persistence of excitation condition for the solutions of autonomous systems,

    A. Padoan, G. Scarciotti, and A. Astolfi, “A geometric characteri- zation of the persistence of excitation condition for the solutions of autonomous systems,”IEEE Trans. Automat. Contr., vol. 62, no. 11, pp. 5666–5677, 2017

    work page 2017

  12. [12]

    Chen and C.-T

    C.-T. Chen and C.-T. Chen,Linear system theory and design. Holt, Rinehart and Winston New York, 1984, vol. 301

    work page 1984

  13. [13]

    On explaining the surprising success of reservoir computing forecaster of chaos? the universal machine learning dynamical system with contrast to var and dmd,

    E. Bollt, “On explaining the surprising success of reservoir computing forecaster of chaos? the universal machine learning dynamical system with contrast to var and dmd,”Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 31, no. 1, 2021

    work page 2021

  14. [14]

    On dynamic mode decomposition: Theory and applications,

    J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N. Kutz, “On dynamic mode decomposition: Theory and applications,” Journal of Computational Dynamics, vol. 1, no. 2, pp. 391–421, 2014

    work page 2014

  15. [15]

    Online dynamic mode decomposition for time-varying systems,

    H. Zhang, C. W. Rowley, E. A. Deem, and L. N. Cattafesta, “Online dynamic mode decomposition for time-varying systems,”SIAM Jour- nal on Applied Dynamical Systems, vol. 18, no. 3, pp. 1586–1609, 2019

    work page 2019

  16. [16]

    Fading memory echo state networks are universal,

    L. Gonon and J.-P. Ortega, “Fading memory echo state networks are universal,”Neural Networks, vol. 138, pp. 10–13, 2021

    work page 2021

  17. [17]

    A model for circadian oscillations in the drosophila period protein (per),

    A. Goldbeter, “A model for circadian oscillations in the drosophila period protein (per),”Proceedings of the Royal Society B: Biological Sciences, vol. 261, no. 1362, pp. 319–324, 09 1995

    work page 1995