The paper investigates how training shapes the geometry of recurrent network dynamics in Reservoir Computing (RC) networks used for temporal forecasting. It demonstrates that, in a linear continuous-time reservoir with infinite data, training data create an invariant subspace whose dimension matches the number of dominant modes. For a simplified diagonal linear reservoir, the study connects dominant eigenvalues and eigenvectors to the spectrum of a backward Dynamic Mode Decomposition matrix, providing a finite-dimensional approximation of the backward-time Koopman operator, and illustrates these phenomena through simulation while suggesting extensions to nonlinear RC.
arXiv:2607. 17858v1 Announce Type: cross Abstract: Reservoir computing exploits nonlinear dynamical systems to encode temporal inputs into high-dimensional state space representations.
arXiv:2607. 17909v1 Announce Type: cross Abstract: The edge-of-chaos heuristic has long served as a guiding principle for designing reservoir computers, yet its relevance to machine performance remains elusive.
arXiv:2606. 19984v1 Announce Type: new Abstract: Reservoir computing offers a lightweight framework for forecasting dynamical systems but may struggle to capture long-range dependencies due to limited representational capacity.
arXiv:2607. 24420v1 Announce Type: cross Abstract: Reservoir computing has emerged as an efficient machine learning framework for predicting time series generated by dynamical systems.
arXiv:2608. 04593v1 Announce Type: cross Abstract: Echo State Networks (ESNs) offer an efficient framework for temporal prediction, but their randomly initialized reservoirs are often over-parameterized and dynamically redundant.
Echo State Networks (ESNs) offer an efficient framework for temporal prediction, but their randomly initialized reservoirs are often over-parameterized and dynamically redundant. Existing pruning methods largely rely on static connectivity or activation statistics, which may overlook neurons that shape input-driven state transitions.
The paper introduces Functional Dynamic Mode Decomposition (fDMD), extending traditional DMD to infinite-dimensional systems by learning finite-rank operators from functional data such as observables, densities, or wavefunctions. It demonstrates that conventional DMD algorithms are special cases of fDMD and illustrates the approach with examples involving Koopman, Perron‑Frobenius, and Koopman‑von Neumann operators for graphons, ODEs, and SDEs.
This paper introduces a data‑driven neural power‑iteration algorithm for approximating the dominant eigenfunctions (modes) of the Koopman operator in nonlinear dynamical systems. By avoiding explicit construction of the operator’s projection, the method sidesteps the curse of dimensionality that plagues expressive neural templates. The authors provide theoretical convergence guarantees tied to sample size and network width, and demonstrate through numerical experiments that the approach yields accurate, smooth mode approximations without the drawbacks of traditional techniques such as extended dynamic mode decomposition.
arXiv:2601. 19019v3 Announce Type: replace-cross Abstract: Neural population activity in sensory cortex is organized on low-dimensional manifolds, but why such manifolds arise and what determines their geometry remain unclear.
arXiv:2604. 19465v3 Announce Type: replace-cross Abstract: Understanding how complex systems respond to perturbations, such as whether they will remain stable or what their most sensitive patterns are, is a fundamental challenge across science and engineering.
arXiv:2404. 17429v4 Announce Type: replace-cross Abstract: A natural hypothesis for the success of reservoir computing in generic tasks is the ability of the untrained reservoir to map distinct input time series to separable reservoir states, a property we term separation capacity.
The paper introduces DOODL, a framework that learns a dictionary of spectral dynamics to represent related dynamical systems as points on a low‑dimensional manifold in operator space. By constraining operator estimation to this learned manifold, DOODL provides compact, interpretable embeddings and enables fast, accurate operator estimation from short, partially observed trajectories. Experiments on metastable Langevin dynamics and turbulent plasma simulations show that DOODL achieves one to two orders of magnitude lower errors than independent estimation methods, especially in low‑data regimes.