arXiv Science⌕ Search

arXiv · 2611.20112

Washout-free synchronization, parameter estimation and uncertainty quantification of echo state networks

Abstract

Echo state networks (ESNs) forecast chaotic systems for several Lyapunov times. Like all recurrent networks, ESNs require a washout before forecasting, i.e., a phase in which the full state drives the network until the reservoir state synchronizes with the system. At deployment, however, sensors measure only part of the state, with noise and at a lower sampling rate than the training data. We propose a washout-free deployment of ESNs via ensemble data assimilation. In closed loop, the ESN is a Markovian dynamical system; hence, a sequential filter can correct its reservoir state whenever data are available. The filter synchronizes an ensemble of reservoir states, drawn from the training data, with the system, and quantifies the forecast uncertainty. We extend the method to parametric ESNs, to infer physical parameters, and to parallel ESNs with localization, to forecast high-dimensional systems. We test the method on the Lorenz-63 and Lorenz-96 systems. First, the filter synchronizes the ESN with a fraction of the data of a washout. Second, the filter deploys one full-state ESN from a subset of its components, which a washout cannot do. Third, the filter estimates a forcing that is not in the training data, and the estimated forcing lengthens the forecast. Fourth, localization synchronizes a parallel ESN with eight times fewer members than the global filter. Fifth, the ensemble spread signals, without the true state, when the ESN synchronizes and when the filter fails. This work enables the deployment of ESNs from partial, sparse and noisy observations.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Andrea Nóvoa. 2026-11-15. Washout-free synchronization, parameter estimation and uncertainty quantification of echo state networks. https://arxiv.science/abs/2611.20112

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

The Chaotic Dynamics of Coupled Maps with Convection and Diffusion

We explore the fundamental role of convective and diffusive coupling on the spatiotemporal chaos of a large one-dimensional lattice of coupled maps on a ring. Extensively chaotic dynamics persist over nearly the entire range of spatial coupling strengths that yield numerically stable dynamics. The spatial coupling affects the shape of the spectrum of Lyapunov exponents, for increasing convective coupling the leading Lyapunov exponents increase while the large negative Lyapunov exponents decrease resulting in a significant increase in the fractal dimension of the dynamics. Convective coupling increases the localization of the leading Lyapunov vector and adds significant power to the high-wavenumber components of the spectrum of covariant Lyapunov vectors while increasing their entanglement. The eigenvalues and eigenvectors of the diffusive coupling matrix remain insightful quantities in the presence of nonlinear convective coupling. We discuss the implications of these findings for high-dimensional chaotic dynamics of large transport systems.

nlin.CD↗

Replicating chaotic dynamics from sparse sensors with recurrent latent states

Forecasting a high-dimensional nonlinear system from a few point sensors requires an estimate of the current state and a model of its evolution. Reduced-order models usually take the full state as input, so with only sparse sensors they require an auxiliary full-state estimate whose error is inherited by the latent representation. We instead build the latent state from sensor measurements alone, using a shallow recurrent decoder (SHRED) network that encodes a history of point measurements, and train an echo state network to advance that latent state autonomously. We show that the resulting latent dynamics replicate the long-term dynamical invariants of the underlying system rather than merely tracking its short-term trajectory. For the Kuramoto-Sivashinsky equation, the autonomous latent model reproduces the leading Lyapunov exponents of the full system from 25 of 128 state components, using a latent state with 6-14% as many components as the full state. Because Lyapunov exponents are invariant under smooth coordinate changes, this is consistent with the recurrent encoder recovering an embedding of the attractor, in the spirit of Takens' delay embedding theorem. The recurrent encoding also decouples forecast skill from reconstruction accuracy. SHRED reconstructs less accurately than a convolutional autoencoder given the full state, yet it forecasts comparably on the simpler chaotic case and better on the more complex one. For Kolmogorov flow, with sensors at 6% of grid points and a latent state 0.3% the size of the vorticity field, the autonomous model remains on the learned latent manifold over long horizons.

nlin.CD↗

The Quantum Pendulum of Turbulent Flow: Decaying turbulence as a thermal state of an integrable system

We present a theory of decaying turbulence in which the statistics of a classical flow becomes quantum mechanics in loop space. The loop functional plays the role of a wave function, and the Navier-Stokes equation makes it evolve linearly, with the viscosity as Planck constant. Turbulence at infinite Reynolds number is defined by a discretized loop equation on polygons with $N$ sides and a renormalization. It has an exact manifold of decaying solutions, the Euler ensemble; each is a singular point of the deterministic equation, which does not select their weights. We interpret this instability as the source of spontaneous stochasticity: thermal noise only triggers a diffusion of the momentum loop over the Euler states, and neither its amplitude nor the temperature of the fluid enters the statistics. In a stochastic model of this diffusion equal weights are proved to be the attractor. As a fermion trace, the partition function and the vorticity correlations become Gibbs averages of two quantum pendulums with the Mathieu Hamiltonian. The energy spectrum is a chain of explicit quadratures. It falls as $k^{-7/2}$ at large $k$; at $kL=\sqrt{2}$, with $L=\sqrt{\tildeν(t+t_0)}$, its slope diverges as a double logarithm, so the decay at a fixed wave number is singular at a finite time and the vorticity correlation oscillates. At small $k$ there are two solutions, for a quenched and an annealed orientation of the plane of the loop: $k^{-1}\log(1/kL)$ and $k^{-2}/\log(1/kL)$. The energy and its dissipation depend on the size of the system logarithmically in the first and almost linearly in the second, and the energy decays with an index of 1.1-1.2 and about 1.45. The reduction to the quantum pendulum is formally verified in Lean 4 in a forthcoming paper. The theory rests on stated physical hypotheses; its comparison with simulations and experiment is left to a separate paper.

nlin.CD↗