Probing the Edge of the Basin: Finite-Amplitude Virtual Kicks in Reservoir Twins as Early Warnings for Rate-Induced Tipping

Most early warnings of sudden collapse, in climate, ecosystems or physiology, rely on one idea. As a system nears a tipping point it recovers more slowly from small nudges. This "critical slowing down" shows up as rising autocorrelation and variance in its time series. The idea fails in an important case. In rate-induced tipping (R-tipping), the environment changes faster than the system can follow. The system falls off a moving equilibrium that stays locally stable the whole time, so small-nudge recovery never slows and the classical signals stay quiet. We propose a different warning. First, fit a reservoir computer to the pre-tipping data. This is a cheap recurrent network whose internal weights are random and fixed, and only a linear readout is trained. The fitted network then serves as a stand-in "digital twin" of the system. We then kick the twin's state with perturbations of growing size and record the smallest kick that stops it from returning. That quantity is the distance to the basin edge, the boundary past which the system goes to a different state. It shrinks before tipping whether the cause is a slowly drifting parameter or a forcing that moves too fast.

A second reason favors finite kicks. Active probing, meaning perturbing the system and watching its response, beats passive monitoring in medicine: electrical stimulation predicts seizures better than passive EEG (Lepeu et al.), and stress-recovery tests have been proposed for depression (Stapelberg et al.). Nobody can kick the Atlantic overturning circulation or a power grid. A twin can be kicked freely. It also sees large kicks, which linearization-based indicators do not.

Core Claim

Let $x(t)$ be the twin's reconstructed state at time $t$, and let $\mu(t)$ be the observed forcing, such as warming or harvesting rate. Let $\Phi_H$ be the twin's flow over a short horizon $H$ when driven by the forcing extrapolated at its current rate $\dot\mu(t)$. We define the probe margin as

$$m(t) = \min \{\, \|\delta\| : \|\Phi_H(x(t)+\delta) - \Phi_H(x(t))\| > \epsilon \,\},$$

where $\delta$ is a kick applied in the observed coordinates and $\epsilon$ is a return tolerance. In words, $m(t)$ is the size of the smallest kick after which the twin fails to rejoin its own unkicked path within $H$ steps. Let $\sigma$ be the noise level, estimated from the twin's one-step residuals.

For noise-driven escape over a barrier, the escape rate scales roughly as

$$\exp(-c\, m^2/\sigma^2)$$

near a saddle, where the local quasi-potential is approximately quadratic and $c$ is a constant of order one. The ratio $m/\sigma$ therefore yields an absolute alarm threshold, not only a relative trend.

The central claim is that $m(t)$ behaves differently under the two kinds of tipping. Under bifurcation tipping (B-tipping), where a slowly drifting parameter destroys the equilibrium, $m$ falls together with the linear recovery rate. Under R-tipping, $m$ falls while the linear recovery rate stays flat, because the growing lag behind the moving equilibrium eats into the basin. No labeled tipping events are needed. The claim is falsifiable on systems whose true basin boundary can be computed from the governing equations.

Relation to Prior Work

Three reservoir-based tipping methods motivate this proposal and set its boundaries.

Tipping Point Detection Using Reservoir Computing refits the readout in sliding windows and feeds the changes in readout weights to a random forest trained on labeled tipping events. We need no labels.

Ultra-Early Prediction of Tipping Points: Integrating Dynamical Measures with Reservoir Computing (RCDyM) extracts Jacobian eigenvalues, Floquet multipliers and Lyapunov exponents from a windowed twin and extrapolates them to their critical values. These are all infinitesimal-perturbation quantities. By construction they stay away from criticality under R-tipping, and the file reports no R-tipping evaluation for it.

Machine-learning prediction of tipping with applications to the Atlantic Meridional Overturning Circulation feeds the bifurcation parameter into the reservoir and rolls the twin forward to see when it tips. That approach supplies our forcing-channel design. Its answer, however, depends on assuming a future forcing path. Our margin asks how close the system is to the edge now, at the current forcing rate.

Two non-reservoir lines define the R-tipping competition.

Deep Learning for predicting rate-induced tipping trains a one-dimensional convolutional network on 300,000 simulated realizations per system. It confirms that classical indicators fail under R-tipping and shows that supervised learning can succeed when simulations are plentiful.

Koopman early warning signals for bifurcation and rate-induced tipping (control-aware ResKMD) measures how badly a finite linear operator, fitted with learned or random-feature embeddings, fails to represent the forced dynamics. It is the strongest unsupervised competitor. Its authors note that it lacks an absolute threshold, which $m/\sigma$ is designed to provide.

Two data-quality papers constrain the design. Uncertainties in critical slowing down indicators of observation-based fingerprints of the AMOC shows that variance in sea-surface temperature records falls artificially as observation density grows. Inconclusive early warning signals for Dansgaard-Oeschger events across Greenland ice cores shows that reported early warnings depend on preprocessing and significance testing. Both results mean $\sigma$ must be estimated locally, and apparent trends in $\sigma$ must be separated from trends in $m$.

Research Program

Twin construction. We use an echo state network, a reservoir whose recurrent matrix is scaled so that the effect of old inputs fades (Gallicchio et al.). It receives the observed state and the forcing $\mu(t)$, as in Patel et al. The ridge readout (Jaeger) is refit in sliding windows. We choose fixed recurrent poles because fixed-pole networks train with well-conditioned objectives: When Learning Hurts reports a Hessian condition number around 2.5, against roughly $10^3$ when the poles are trained.

Probing. In each window, we run the twin closed-loop from the current state and apply kicks along a set of directions: the leading singular vectors of the linearized flow plus random directions. For each direction, a bisection over kick size finds the smallest kick that breaks return. The margin $m(t)$ is the minimum over directions.

The main scientific risk. The twin learns the vector field only where noise has carried the data. Kicks that reach an unvisited region test the reservoir's extrapolation, not the physical system. We treat this as a first-class object rather than hiding it. We train an ensemble of reservoirs with different random seeds and report their disagreement at each probed kick size. Any margin reached where the ensemble disagrees is marked unreliable. We then map, on synthetic systems, how the accuracy of $m$ depends on the ratio of noise amplitude to true basin radius. If that accuracy collapses at realistic noise levels, the method is limited to noisy systems, and we will say so.

Baselines

The naive reference is rolling lag-one autocorrelation, variance and the autoregressive restoring rate. These are computed on the same windows with the same detrending and the same surrogate-based significance testing, following the stricter protocol of the ice-core study. They should fail on R-tipping. If they succeed there, our benchmark is too easy.

The strongest unsupervised competitor is control-aware ResKMD. It will be run with its random-Fourier-feature embedding, and also with its learned embedding where compute allows. It gets the same forcing channel, the same windows, and a mode-truncation search with the same number of configurations we allow for reservoir hyperparameters. The paper reports AUROC ≈ 1.0 across its tested systems and a clear separation of tipping from non-tipping GISS-E2.1-G runs around the year 2150. Matching AUROC is therefore not enough to claim a win. We count a win as either of two outcomes. The first is a longer lead time at a fixed false-alarm rate. The second is a correctly calibrated absolute threshold, meaning that predicted escape probabilities from $m/\sigma$ match empirical tipping frequencies across forcing rates, which ResKMD does not provide.

The strongest supervised reference is the convolutional network of Huang et al. It reports 95% accuracy as tipping approaches, with lead times of up to 280 steps for the saddle-node system and 1,000 steps for the Bautin system. We report our lead times on the same systems and additionally retrain that network with 1,000 rather than 300,000 realizations. Our method uses no labels, so its advantage should appear in that low-data regime.

Two reservoir baselines share our twin. The first is RCDyM's dominant Jacobian eigenvalue, extracted from the identical trained reservoir. The second is Patel-style rollout under the true future forcing, which acts as an oracle-informed upper reference.

The key ablation is an infinitesimal probe: the same twin, with the kick size fixed at a small value and the recovery rate read off. This isolates finite amplitude as the ingredient. A second ablation replaces the reservoir with a non-recurrent random-feature regression of the same width. This isolates whether the reservoir's fading memory matters for representing lagged, non-autonomous dynamics.

Why not just…?

Why not just use autocorrelation and variance? Under R-tipping the moving equilibrium stays linearly stable, so these indicators have nothing to detect. Both Huang et al. and Nathaniel et al. report that they fail or give inconsistent separation there. On B-tipping they remain a fair competitor. We expect only modest gains there and will report them honestly.

Why not just read the Jacobian or pseudospectrum off the twin, as RCDyM does? Any small-kick quantity, including transient growth from non-normal dynamics, describes the neighborhood of the current path, not the location of the basin edge. The infinitesimal-probe ablation tests this directly. If it tracks R-tipping as well as $m$ does, finite amplitude adds nothing and the core claim is refuted.

Why not just roll the twin forward under the forcing, as Patel et al. do? When the future forcing is known exactly, rollout may be as good or better, and that is a real threat. We give rollout the true future forcing in synthetic tests and expect it to be competitive. The margin answers a different question: whether the system is safe at the current forcing rate, with an escape-probability interpretation. We will show where each is preferable, most importantly when the forcing continuation is misspecified.

Why not just use the Koopman residual? It is strong, and it may match us on AUROC. Its residual measures representational failure rather than a physical distance, so it cannot be converted into an escape probability. The calibration experiment decides whether that distinction is worth anything in practice.

Proposed Experiments

The first experiment is ground-truth validation on synthetic systems with known equations. These are the moving saddle-node, the nonautonomous Bautin oscillator, the compost-bomb model, May's harvesting model and the Rosenzweig–MacArthur model, plus fold and Hopf B-tipping cases. On them we compare $m(t)$ against the true distance to the basin boundary, computed from the equations. We sweep forcing rates across the critical rate and noise levels relative to basin size. We report AUROC for tipping versus non-tipping runs, lead time at a 5% false-alarm rate, and calibration of escape probabilities predicted from $m/\sigma$.

The decisive check is a single figure. Across forcing rates that straddle the R-tipping threshold, the Jacobian eigenvalue should stay flat while $m$ falls ahead of tipping, and the ensemble-disagreement flag should correctly mark where the twin cannot be trusted.

We then move to the empirical and model datasets used by the prior work: the cyanobacteria collapse series, the NASA GISS-E2.1-G tipping and non-tipping AMOC runs, and HadSST4. On HadSST4 we propagate the 200-member ensemble through the method to test whether any trend in $m$ survives observational uncertainty and the sampling-density artifact in $\sigma$. We make no collapse-date forecast unless it does.