Skip to main content
Back to timeline
arXivSource publication:

Reservoir computing can rebuild a hidden attractor, but ring coupling widens the range without guaranteeing it stays hidden

Synopsis

For the equilibrium-free NE6 chaotic system, a reservoir built from driven Kuramoto oscillators reconstructs the hidden attractor from time series, and nearest-neighbour ring coupling keeps accurate reconstruction over a wider range of coupling and forcing strengths than all-to-all coupling; rewriting one oscillator equation so the reservoir itself has no equilibria yields the most reconstructions meeting the statistical criteria, yet matching short-term prediction and long-term statistics does not establish that the rebuilt attractor remains hidden.

Source-provided article image: Reconstructing hidden attractors with reservoir computing
arXiv

Interpretation

A reservoir of driven Kuramoto oscillators can reconstruct the hidden attractor of the NE6 chaotic system from time series, and the coupling topology sets how wide the reconstructable parameter range is. Under all-to-all coupling, accurate short-term prediction is confined to a wrench-shaped region of the coupling-strength and forcing plane; with ring coupling, where each oscillator connects only to its two nearest neighbours, the low normalized-mean-squared-error region is substantially larger. Comparing the three reservoirs on the same parameter grid and the same NE6 data, ring coupling reduces the number of parameter pairs with normalized mean squared error below threshold from 223 to 75.

Reproducing an attractor's trajectory and statistics does not guarantee that the reconstructed attractor is still hidden in the reservoir's own state space, because the reservoir can acquire equilibria that NE6 lacks. Both the all-to-all and ring reservoirs retain equilibria across much of the accurate reconstruction region, so numerical searches from equilibrium neighbourhoods yield only candidate hidden attractors rather than proof of hiddenness. Forward integration along unstable eigendirections from detected equilibria counts an attractor as reachable only when a trajectory matches the reconstruction statistically; adding NE6 transient trajectories to training shifted the detected equilibrium rather than removing it.

Rewriting the phase equation of the last oscillator in the ring makes the reservoir free of equilibria for any trained readout, so every attracting set it realizes is hidden under the equilibrium-based definition. The ordinary ring widens reconstruction but still carries many equilibria; the modified ring excludes equilibria by construction while keeping short-term accuracy comparable to the ordinary ring. The modified ring adds only eight high-error cases relative to the ring, a difference of a few percentage points, while marked cases meeting both the Lyapunov and KL criteria rise from 14 for the ring to 47, against 1 for all-to-all coupling.

Perspective

The conclusions are limited to one target system, NE6, one family of driven Kuramoto reservoirs, and the realizations and parameter ranges examined; hiddenness is judged from finite multistart equilibrium searches and forward trajectory tests, which yield candidate hidden attractors only.

The modified ring excludes equilibria but does not establish that the remaining basin geometry matches NE6; local conditional contraction along a nonconstant signal remains possible, so a negative conditional Lyapunov exponent indicates local stability rather than the global echo state property.