PaperlayerView exact revision on arXiv

Source-linked AI summary

The internal model principle in reservoir computing

Danyal Saqib, Alberto Padoan

arXiv:2610.08987v1eess.SYcs.SYmath.OC

TL;DR

Reservoir computing offers efficient temporal modeling, but empirical reservoir selection and the echo state property do not ensure stable, robust predictors. The paper models training signals as outputs of exogenous generators and derives links to control theory. Its case studies show how training can leave prediction unstable and how signal-origin information can guide reservoir design.

  • Problem

    Reservoir selection and training remain largely empirical, and the echo state property guarantees neither stability nor robustness of trained predictors.

    Source: §I · p. 1

  • Method

    The paper models training signals as outputs of exogenous generators and analyzes reservoir synchronization and readout training through the internal model principle.

    Source: §I · p. 1

  • Results

    Exact training can still yield unstable prediction because finite-horizon least-squares readouts may leave closed-loop eigenvalues unstable.

    Source: A. Fundamental limitations of linear… · p. 7

  • Takeaways & Limitations

    Information about the origin of training signals can inform reservoir design with prescribed performance.

    Source: §I · p. 1

    For finite training horizons, the least-squares readout need not satisfy the exact readout condition, and closed-loop eigenvalues can leave the imaginary axis.

    Source: A. Fundamental limitations of linear… · p. 7

Abstract

from arXiv · show

The paper studies learning in reservoir computing systems through the lens of the internal model principle. The key idea is to model training signals as outputs of an exogenous signal generator, thus connecting reservoir computing to output regulation, observer theory, and frequency-domain interpolation. For linear signal generators, we characterize the steady-state response of a linear reservoir through invariant subspaces, establishing connections with Luenberger observers and classical frequency-domain interpolation. For nonlinear signal generators, we obtain analogous results using invariant manifolds, connecting synchronization to Kazantzis-Kravaris/Luenberger observers and readout training to nonlinear moment matching. Case studies illustrate how our perspective explains possible instability of trained predictors and informs reservoir design.

I. INTRODUCTION

Reservoir computing trains a linear readout over a fixed dynamical system, but empirical reservoir selection and the echo state property alone do not ensure stable, robust predictors. The paper models training signals as outputs of a signal generator, linking reservoir learning to control theory and principled reservoir analysis.

II. RESERVOIR COMPUTING

  • Reservoir computing uses a fixed dynamical system to generate features and fits a linear readout, leaving the reservoir itself untrained.The readout is fitted by linear regression; reservoirs may be neural networks, physical systems, or biological substrates.Source: §I · p. 1, §II · p. 2

I. INTRODUCTION

  • The echo state property guarantees neither stability nor robustness of the resulting predictors, while reservoir selection and training remain largely empirical.This motivates a more principled approach to reservoir analysis and design.Source: §I · p. 1
  • Modeling training signals as outputs of an exogenous generator connects reservoir learning to the internal model principle, output regulation, observer design, and moment matching.After a transient, the reservoir state is determined by the signal-generator state through a synchronization map.Source: §I · p. 1
  • For linear and nonlinear signal generators, the paper connects steady-state reservoir responses to invariant subspaces or manifolds, observer theory, and frequency-domain or nonlinear moment matching.The stated connections include Luenberger observers for linear generators and Kazantzis–Kravaris/Luenberger observers for nonlinear generators.Source: §I · p. 1
  • Case studies show that exact prediction with linear reservoirs requires finite-dimensional linear realizations, unassigned closed-loop modes can destabilize prediction, and signal-origin information can guide reservoir design.The paper presents these as implications of its theory for reservoir computing.Source: §I · p. 1

II. RESERVOIR COMPUTING

  • The echo state property formalizes forgetting of the reservoir's initial condition, and under additional assumptions permits approximation of broad classes of input–output maps.When the input comes from a dynamical system, the reservoir state can asymptotically become a function of that system's state through a synchronization map.Source: §I · p. 1, §II · p. 2

III. LINEAR RESERVOIR COMPUTING

The linear setting studies a stable, controllable reservoir driven by a linear exosystem that generates the training and target signals. Its assumptions make the reservoir forget its initial condition and restrict the generator to bounded constant and oscillatory modes.

III. LINEAR RESERVOIR COMPUTING

  • The linear section relates synchronization to Luenberger observers and output regulation, and readout training to frequency-domain interpolation.It specializes reservoir computing to linear reservoirs driven by linear signal generators.Source: §III · p. 2

A. Problem setup

  • The reservoir matrix F is assumed Hurwitz and (F, G) controllable; the first condition ensures the echo state property, while the second ensures every reservoir mode is excited.Controllability is also needed later to characterize closed-loop modes left free by training.Source: A. Problem setup · p. 2
  • The temporal data are modeled as outputs of an autonomous linear exosystem with state w, generated input v = Lw, and target y = Qw.For prediction, the target equals the generated signal, so Q = L and p = m.Source: A. Problem setup · pp. 2–3
  • The exosystem matrix S has distinct imaginary-axis eigenvalues, and (L, S) is observable; its trajectories therefore generate bounded finite sums of constants and sinusoids.Observability excludes exosystem modes invisible in the generated input.Source: A. Problem setup · p. 3

B. Connections to output regulation and observer design

For a stable linear reservoir driven by a linear exosystem, the steady-state response lies on an invariant subspace and the reservoir synchronizes to a linear function of the generator state. An exact readout exists precisely when the target output can be represented through that synchronization map.

B. Connections to output regulation and observer design

  • The interconnected reservoir and exosystem have an invariant subspace where the reservoir state is a linear function of the exosystem state.The subspace is characterized by a Sylvester equation and its graph is invariant under the interconnected dynamics.Source: B. Connections to output regulation… · p. 3
  • An exact readout exists when Q = WoutΠ, so the target output is representable through the synchronization map and the readout error converges exponentially to zero.The criterion is equivalent to the target's row space being contained in the row space of Π.Source: B. Connections to output regulation… · p. 3
  • After a transient, the reservoir state is given by the synchronization map w ↦ Πw, which acts as a Luenberger observer when the exact readout condition holds.The reservoir-state error from Πw decays exponentially under the stated stable-reservoir assumption.Source: B. Connections to output regulation… · p. 3
  • The rank condition is satisfied when rank Π = ν, which occurs for m = 1 and n ≥ ν under assumptions (A1)–(A4).The condition is stated as equivalent to ker Π ⊆ ker Q.Source: B. Connections to output regulation… · p. 3

C. Reservoir prediction

Exact readout feedback embeds the exosystem in the closed-loop reservoir, but training does not determine whether the predictor is stable. Its stability depends on transverse eigenvalues that can vary with data transients, noise, and regularization.

C. Reservoir prediction

  • Exact readout feedback makes the exosystem subspace invariant, so the closed loop reproduces every exosystem trajectory and its target.The modes on that subspace form an internal model of the signal generator.Source: C. Reservoir prediction · p. 4
  • The closed-loop matrix is not Hurwitz because it contains the exosystem modes on the invariant subspace.The proposition establishes σ(S) ⊆ σ(F + GWout).Source: C. Reservoir prediction · p. 4
  • Predictor stability is determined by transverse eigenvalues, not by the target, and those eigenvalues depend on data transients, noise, and regularization.Under (A2), suitable readout choices can assign the transverse eigenvalues to any multiset of n − ν complex numbers closed under conjugation.Source: C. Reservoir prediction · p. 4

D. A frequency-domain interpretation

For linear reservoirs, exact readout training can be viewed as frequency-domain interpolation at the exosystem eigenvalues. In the single-input case, the trained readout coordinates correspond to order-zero moments.

D. A frequency-domain interpretation

  • The Sylvester equation also characterizes linear-system moments at signal-generator eigenvalues, linking readout training to frequency-domain interpolation.The passage situates this connection in model reduction by moment matching.Source: D. A frequency-domain interpretation · p. 4
  • Exact readout training holds if and only if η0(si)Lvi = Qvi for each exosystem eigenvalue si.These are interpolation conditions on the reservoir transfer matrix at the exosystem eigenvalues, with directions Lvi and values Qvi.Source: D. A frequency-domain interpretation · p. 4
  • Exact readout training is equivalent to tangential interpolation at the exosystem eigenvalues, using directions Lvi and values Qvi.Each Lvi is a nonzero tangential interpolation direction under the stated assumptions.Source: D. A frequency-domain interpretation · p. 4
  • For m = 1, the trained reservoir’s order-zero moments at the exosystem eigenvalues correspond one-to-one with the entries of WoutΠ.The paper interprets this as data-driven moment matching.Source: D. A frequency-domain interpretation · p. 4

IV. NONLINEAR RESERVOIR COMPUTING

The nonlinear extension uses output-regulation and observer-design tools to characterize reservoir response to signals generated by a nonlinear exosystem. Its setup seeks a readout that minimizes error, under assumptions that yield local echo-state and exosystem properties.

IV. NONLINEAR RESERVOIR COMPUTING

  • The nonlinear extension applies output-regulation and observer-design tools to nonlinear reservoirs and signal generators.The section first characterizes steady-state response through a synchronization map.Source: §IV · p. 4, A. Problem setup · p. 5

A. Problem setup

  • Training seeks a readout minimizing the error over [0, T], with zero minimum exactly when some readout regulates that error to zero on the interval.The readout is a matrix Wout in Rp×n.Source: A. Problem setup · p. 4
  • Assumptions (A5)–(A7) give local echo-state behavior for the reservoir and bounded, recurrent-near-initial-state signals from the exosystem.They require local regularity and stability conditions, including a Hurwitz reservoir linearization and a Lyapunov-stable exosystem equilibrium with imaginary-axis linearization spectrum.Source: A. Problem setup · p. 5

B. Connections to output regulation and observer design

Under the stated assumptions, the nonlinear interconnected system has a locally invariant manifold that is the graph of a synchronization map. After a transient, this map determines reservoir state from exosystem state, although the nonlinear result is local and generally lacks an explicit map.

B. Connections to output regulation and observer design

  • Under (A5)–(A7), the invariance equation has a locally unique C2 solution π whose graph is locally invariant.The solution is defined near w = 0 and satisfies π(0) = 0.Source: B. Connections to output regulation… · p. 5
  • The center manifold theorem supplies the local invariant graph, while Lyapunov and Poisson stability yield its local uniqueness.The proof uses the stable reservoir modes and neutral exosystem modes in the linearization.Source: B. Connections to output regulation… · p. 5
  • After a transient, π maps exosystem state to reservoir state; unlike the linear case, the nonlinear conclusions are local and π generally has no explicit expression.With an exact readout condition, the invariance equation plays the role of the nonlinear output-regulation equations.Source: B. Connections to output regulation… · p. 5

1) Linear reservoirs as KKL observers:

For nonlinear signal generators, the reservoir synchronizes through an invariant-manifold map that acts as a KKL observer. An exact linear readout matches the target on the signal-generator set, linking training to nonlinear moment matching and restricting exact prediction to signals with finite-dimensional linear realizations.

1) Linear reservoirs as KKL observers:

  • Under the stated assumptions, the synchronization error decays exponentially, and a readout satisfying the exact-readout condition has an error that also converges exponentially to zero.The synchronization manifold is exponentially attractive for exosystem states in K; the readout condition requires the target components to be linear combinations of the synchronization-map components on K.Source: 1) Linear reservoirs as KKL observers · pp. 5–6
  • The synchronization map is the reservoir’s response to the driving signal’s past and gives a functional KKL observer of the target when the exact-readout condition holds.If the map is injective on K, it is also a KKL observer of the exosystem state.Source: 1) Linear reservoirs as KKL observers · p. 6
  • The map π sends exosystem trajectories in K to trajectories of the closed-loop reservoir, so exact prediction requires signals with an n-dimensional autonomous linear realization.This is a finite-dimensional linear internal model restriction.Source: 1) Linear reservoirs as KKL observers · p. 6
  • The exact-readout condition makes the trained reservoir’s steady-state output match the target map, interpreting training as data-driven nonlinear moment matching.The relevant moment is the steady-state response w ↦ Woutπ(w) of the reservoir output to the signal generator.Source: 1) Linear reservoirs as KKL observers · p. 6

V. FROM ANALYSIS TO RESERVOIR DESIGN

The internal-model perspective explains how exact training can leave predictor stability and disturbance response vulnerable, while suggesting observer design when the signal origin is known. In a three-state sinusoidal example, finite-horizon fits can be unstable or Hurwitz depending on initialization, and exact prediction has unbounded response to disturbances at the reproduced frequency.

V. FROM ANALYSIS TO RESERVOIR DESIGN

  • When information about signal origins is available, the paper frames reservoir selection as an observer-design problem rather than an empirical choice.The analysis also identifies free transverse modes and internal-model disturbance amplification as limitations of exact training.Source: §V · p. 6

A. Fundamental limitations of linear reservoir computing

  • Exact readouts preserve the ±iω internal-model modes while leaving the transverse eigenvalue free, so exact fit alone does not determine predictor stability.In the example, every real transverse eigenvalue can be attained by an exact readout.Source: A. Fundamental limitations of linear… · pp. 6–7
  • Finite-horizon least-squares fits were unstable for every T from x(0)=0, but Hurwitz for T ≥ 8 from x(0)=Πw(0)+6N.As T grows, the readout approaches an exact readout with transverse eigenvalue 0 in the first initialization and approximately −2.41 in the second.Source: A. Fundamental limitations of linear… · p. 7, Figure 2 · p. 7
  • For a predictor with λtr < 0, a disturbance at ω produces an unbounded response, whereas a disturbance at 2ω remains bounded.The reproduced-frequency modes are poles of the disturbance-to-prediction transfer function.Source: A. Fundamental limitations of linear… · p. 7, Figure 2 · p. 7

B. Reservoir design as observer design

When the signal generator is known, the observer perspective suggests a parsimonious reservoir containing a copy of that generator, corrected by the measured signal. In the frictionless-pendulum example, gain selection trades synchronization speed against noise sensitivity.

B. Reservoir design as observer design

  • The observer interpretation suggests a parsimonious reservoir that contains a copy of the signal generator, corrected by the measured signal.The design is proposed when information about the signal's origin is available.Source: B. Reservoir design as observer design · p. 7
  • For the frictionless pendulum, a two-dimensional reservoir driven by the measured output has the identity synchronization map π(w)=w.The reservoir's error ε=x−w follows the dynamics set by its gains.Source: B. Reservoir design as observer design · p. 7
  • The gains k1,k2>0 assign the error-system eigenvalues, while readout Wout=[1 0] satisfies the exact readout condition.With the chosen readout fed back, u=x1 turns the reservoir into the pendulum itself, so it contains an internal model and its synchronization rate is a design variable independent of training.Source: B. Reservoir design as observer design · p. 7
  • Fig. 3 compares target and estimate, synchronization error, and noisy RMS error across gains; larger gains speed synchronization but amplify noise, and estimation error is minimized at an intermediate gain.Panel (c) uses measurement noise with standard deviation 10^-2.Source: Figure 3 · p. 8, B. Reservoir design as observer design · p. 8

VI. CONCLUSION

The paper concludes that reservoir-computing learning reflects the internal model principle, connecting linear and nonlinear systems to output regulation, observer design, and moment matching. It identifies open questions in analysis, training, and principled reservoir design.

VI. CONCLUSION

  • Learning in reservoir computing is a manifestation of the internal model principle, with connections to output regulation, observer design, and moment matching for linear and nonlinear systems.Source: §VI · p. 8
  • Open directions include linear-reservoir sensitivity limits, the interplay between training cost and closed-loop predictor dynamics, and persistence of excitation and least-squares moment matching in readout training.Source: §VI · p. 8
  • Principled reservoir design in physical substrates is identified as a natural next step.Source: §VI · p. 8
Loading 2610.08987v1…