Contraction, Criticality, and Capacity: A Dynamical-Systems Perspective on Echo-State Networks

TL;DR

This paper provides a unified dynamical-systems framework for understanding Echo-State Networks, linking stability, memory, and computational capacity with mathematical properties and neuroscientific insights.

Contribution

It introduces a comprehensive analysis connecting ESN stability, universality, and criticality, offering new design principles grounded in mathematics and neuroscience.

Findings

Proves that the Echo-State Property implies the Fading-Memory Property.

Shows polynomial reservoirs with linear read-outs are dense in the space of fading-memory filters.

Links reservoir topology and parameters to memory capacity and criticality measures.

Abstract

Echo-State Networks (ESNs) distil a key neurobiological insight: richly recurrent but fixed circuitry combined with adaptive linear read-outs can transform temporal streams with remarkable efficiency. Yet fundamental questions about stability, memory and expressive power remain fragmented across disciplines. We present a unified, dynamical-systems treatment that weaves together functional analysis, random attractor theory and recent neuroscientific findings. First, on compact multivariate input alphabets we prove that the Echo-State Property (wash-out of initial conditions) together with global Lipschitz dynamics necessarily yields the Fading-Memory Property (geometric forgetting of remote inputs). Tight algebraic tests translate activation-specific Lipschitz constants into certified spectral-norm bounds, covering both saturating and rectifying nonlinearities. Second, employing a Stone-Weierstrass strategy we give a streamlined proof that ESNs with polynomial reservoirs and linear read-outs are dense in the Banach space of causal, time-invariant fading-memory filters, extending universality to stochastic inputs. Third, we quantify computational resources via memory-capacity spectrum, show how topology and leak rate redistribute delay-specific capacities, and link these trade-offs to Lyapunov spectra at the \textit{edge of chaos}. Finally, casting ESNs as skew-product random dynamical systems, we establish existence of singleton pullback attractors and derive conditional Lyapunov bounds, providing a rigorous analogue to cortical criticality. The analysis yields concrete design rules-spectral radius, input gain, activation choice-grounded simultaneously in mathematics and neuroscience, and clarifies why modest-sized reservoirs often rival fully trained recurrent networks in practice.