Basin Stability in Nonlinear Systems

• Basin stability is a global, measure-theoretic method that quantifies the probability of trajectories returning to an attractor after finite perturbations.
• Its estimation relies on Monte Carlo simulations that sample initial conditions under a defined perturbation distribution to gauge return frequencies.
• The concept applies across disciplines such as power grids, climate models, and mechanical systems where basin geometry significantly influences system dynamics.

Searching arXiv for recent and foundational papers on basin stability and related variants. Basin stability is a nonlinear, global stability measure that quantifies how robust an attractor is to finite-size perturbations by assigning the probability, under a specified perturbation distribution, that trajectories return to that attractor. In contrast to local linear stability, which characterizes infinitesimal perturbations near an equilibrium, basin stability is measure-theoretic, depends on the basin of attraction and on the chosen perturbation ensemble, and is particularly useful in multistable systems where basin geometry and basin boundaries strongly influence observed behavior (Kan et al., 2016, Schultz et al., 2016).

1. Formal definition and interpretation

For an autonomous dynamical system on state space $X$, with attractor $A\subset X$, basin of attraction $B(A)\subset X$, and perturbation density $\rho$, basin stability is the probability mass of the basin under $\rho$. In the notation used across several works, this can be written as

$$\mathrm{BS}(A)=\int_X \rho(x)\,\Theta_A(x)\,dx =\int_{B(A)} \rho(x)\,dx,$$

where $\Theta_A(x)$ is the indicator that the trajectory starting from $x$ asymptotically approaches $A$ (Kan et al., 2016, Rakshit et al., 2017).

This definition makes explicit that basin stability is not an intrinsic property of the attractor alone. It is also a property of the perturbation ensemble. Different choices of $\rho$ change the numerical value of basin stability, which is why the measure must be interpreted relative to the physically or operationally relevant perturbation distribution. In this sense, basin stability quantifies robustness against finite perturbations rather than local asymptotic stability near an attractor (Kan et al., 2016, Rakshit et al., 2017).

In networked systems the same idea is often specialized to nodal or structured perturbations. In power-grid models, for example, one studies node-wise basin stability $A\subset X$0, the probability that the grid returns to synchrony after perturbing node $A\subset X$1 at transmission strength $A\subset X$2 (Kim et al., 2016). In delayed oscillator networks, basin stability is used to compare the relative prevalence of incoherent, chimera, and coherent attractors under ensembles of initial history functions rather than point initial conditions (Rakshit et al., 2017).

Because the estimator is based on Bernoulli trials, basin stability naturally comes with binomial uncertainty. If $A\subset X$3 independent samples are drawn and $A\subset X$4 denotes the Monte Carlo estimate, then the standard error has the usual form

$A\subset X$5

with $A\subset X$6, and confidence intervals can be formed by normal approximation or exact Clopper–Pearson intervals (Kan et al., 2016, Rakshit et al., 2017).

2. Estimation methodologies

The standard computational workflow is Monte Carlo estimation. One chooses a perturbation distribution $A\subset X$7, samples initial conditions, numerically integrates the dynamics, classifies the asymptotic attractor, and estimates basin stability as the fraction of successful returns. This same scheme underlies applications to power grids, delayed oscillators, climate models, and mechanical systems (Kan et al., 2016, Kim et al., 2016, Rakshit et al., 2017).

For delay differential equations, the initial condition is an entire history on $A\subset X$8, so the initial-condition space is infinite-dimensional. In the study of chimera states in coupled time-delayed Mackey–Glass oscillators, the history space is projected to a finite-dimensional polynomial space $A\subset X$9, with histories sampled as

$B(A)\subset X$0

using coefficients $B(A)\subset X$1 drawn uniformly from $B(A)\subset X$2. Basin-stability values were found to stabilize for $B(A)\subset X$3, and the main maps used $B(A)\subset X$4, $B(A)\subset X$5 samples, and a modified Heun method with $B(A)\subset X$6 (Rakshit et al., 2017).

A different estimation strategy is to probe basin size locally around known equilibria. In the Kuramoto setting, one algorithm samples random directions on the surface of hypercubes centered at a fixed point, measures the return proportion $B(A)\subset X$7 as a function of hypercube radius $B(A)\subset X$8, and defines a thresholded linear size $B(A)\subset X$9. This approach is far more efficient than global Monte Carlo when stable fixed points are already known and the objective is to estimate the linear extent and relative volume of nearby basins (Delabays et al., 2017).

Distance-based estimation introduces a radius-indexed perturbation family

$\rho$0

and asks at what distance from the attractor the basin-stability level drops below a chosen tolerance $\rho$1. The resulting “basin stability bound” is a finite-perturbation margin that depends on both basin size and basin shape, and its uncertainty can be quantified by exact binomial confidence intervals (Alvares et al., 2023).

Basin stability can also be estimated directly from observational or simulated time series when an explicit dynamical model is unavailable. In that case, one identifies perturbation events, segments the record into post-perturbation trajectories, classifies their eventual attractors, and estimates return frequencies. This data-driven procedure requires sufficiently strong and diverse perturbations, approximate independence of perturbation events, and the timescale separation

$\rho$2

so that relaxation to an attractor is typically resolved before the next perturbation occurs (Kan et al., 2021).

3. Extensions of the classical framework

Classical basin stability only records whether trajectories asymptotically return. Several extensions refine that binary notion by incorporating transient constraints, perturbation structure, or parameter uncertainty (Kan et al., 2016, Schultz et al., 2017, Mitra et al., 2016, Alvares et al., 2023, Brzeski et al., 2016).

Extension	Core definition	Emphasis
Constrained basin stability	$\rho$3	Return with transient constraint
Finite-time basin stability	Probability of returning to a designated return set within time $\rho$4	Recovery time
Multi-node basin stability	Simultaneous perturbation of $\rho$5 nodes in a network	Structured multi-node shocks
Basin stability bound	Smallest perturbation radius where basin stability drops below tolerance $\rho$6	Distance-based robustness margin
Parameter-uncertain basin stability	Joint sampling of initial conditions and uncertain parameters	Accessible ranges of initial conditions and parameters

Constrained basin stability generalizes basin stability by requiring not only asymptotic return but also satisfaction of a user-defined transient constraint. The constraint may encode bounded return time, avoidance of unsafe regions, monotonicity, or a limited number of switches. This makes CBS sensitive to transient phenomena that standard basin stability deliberately ignores (Kan et al., 2016).

Finite-time basin stability and the related notion of independence time arise when perturbations recur. Finite-time basin stability measures the probability of returning to a designated return surface within time $\rho$7, and the independence time is the recovery horizon after which successive jump-like perturbations can be treated as approximately independent for remain-probability bounds (Schultz et al., 2017).

Multi-node basin stability extends the single-node idea to simultaneous perturbations of $\rho$8 nodes in a network. Averaging over ensembles of $\rho$9-node subsets yields $\rho$0, and one may define a critical number $\rho$1 at which the return probability to the desired state drops below an operational threshold (Mitra et al., 2016).

A further modification randomizes selected system parameters together with initial conditions. In that setting one estimates the joint probability of reaching a target attractor under physically accessible ranges of both state and parameter mismatch. This is particularly useful in nonlinear mechanical systems where robustness to parameter uncertainty is as important as robustness to state perturbations (Brzeski et al., 2016).

4. Basin geometry, boundaries, and local-to-global structure

A central theme of the literature is that basin stability depends on global basin geometry, not merely on local linearization. In mechanically stable packings of frictionless grains, the basin of attraction of a packing contains a small core region around the packing, but the core volume $\rho$2 is only very weakly correlated with the observed occurrence probability. Instead, the probabilities are governed predominantly by complex geometric features of the basin far from the packing, including branched and thread-like structures in high-dimensional configuration space (Ashwin et al., 2011).

This geometric emphasis clarifies why numerical basin-stability estimation can behave very differently depending on basin structure. For systems with fractal basin boundaries, including Wada boundaries, basin-stability estimation remains meaningful: individual trajectories can be extremely sensitive to numerical precision, yet the estimated basin measure is typically robust because the uncertainty is concentrated near a lower-measure boundary. By contrast, riddled or intermingled basins reach the limits of the method, since arbitrarily small numerical or perturbative changes can alter outcomes on sets of positive measure throughout the basin (Schultz et al., 2016).

In some settings, surprisingly strong local-to-global relations do emerge. For identical phase oscillators equally coupled on a ring, the sizes of basins of attraction of stable twisted states can be estimated from the stable eigenvalues of the Jacobian at those equilibria. In that case the basin volumes scale as

$\rho$3

contrasting with the Gaussian behavior postulated in earlier work, and indicating that global basin statistics can be encoded in local spectral data under Morse–Smale gradient-flow conditions (Delabays et al., 2017, Mihara et al., 2021).

Recent work has also separated basin structure from energetic stability. In a Kuramoto ring with moderate higher-order interactions, the relative distribution of basin volumes among twisted states remains nearly invariant, while quasipotential barriers deepen and mean first passage times increase under weak noise. This distinction shows that preserving basin structure does not preclude substantial changes in escape dynamics (Wang et al., 15 Oct 2025).

5. Applications across disciplines

In coupled delayed oscillator networks, basin stability has been used to quantify the robustness of incoherent, chimera, and coherent states. For nonlocally and globally coupled time-delayed Mackey–Glass oscillators, basin portraits reveal strongly intertwined basins, and the basin stability of chimera states is substantial in intermediate coupling regimes. In the nonlocal ring, chimera basin stability emerges around $\rho$4, becomes large near $\rho$5, and coexists with incoherent and coherent states over a broad parameter range (Rakshit et al., 2017).

In power systems, basin stability has become a standard nonlinear measure of synchronization robustness. Node-wise basin stability $\rho$6 in small transmission networks typically grows from zero to one as transmission strength increases, but often in a complex, nonmonotonous way, with suppressed, enhanced, and broad-peak transition forms that correlate strongly with betweenness and more weakly with degree (Kim et al., 2016). A related scalar, integrated basin instability,

$\rho$7

was introduced to aggregate instability across the entire coupling sweep; in the IEEE 24-bus study it correlates strongly with $\rho$8 and highlights gatekeeper or bottleneck nodes as particularly vulnerable (Kim et al., 2019). For simultaneous perturbations, multi-node basin stability identifies a critical number of perturbed nodes required to drive synchronized return probabilities below a prescribed threshold (Mitra et al., 2016).

Climate applications have used basin stability to characterize Earth-system resilience in multistable tipping networks. In a conceptual model of five interacting tipping elements, the state of four or five tipped elements has the largest basin volume for large levels of global warming beyond $\rho$9, whereas for lower warming the configurations including disintegrated Greenland and West Antarctic ice sheets occupy comparatively large basins (Wunderling et al., 2020). In related work on the AMOC, the “on” state loses stability via a subcritical Hopf rather than a simple saddle-node, and homoclinic basin bifurcations together with rate-induced thresholds reshape the effective basin even before static loss of stability (1901.10111). Constrained basin stability has further been used as a return-time–dependent resilience measure in the Earth’s carbon cycle, where it decays well before the underlying basin volume changes, thereby acting as an early-warning indicator (Kan et al., 2016).

Mechanical and locomotion problems provide another distinct usage. In multistable mechanical oscillators, basin stability has been used to select parameter ranges where desired responses dominate under accessible ranges of initial conditions and parameter mismatch (Brzeski et al., 2016). In coupled Duffing–Holmes and Lorenz oscillators with mean-field coupling, basin stability quantifies amplitude death, oscillation death, multi-cluster oscillation death, and nontrivial homogeneous steady states, complementing bifurcation analysis by measuring the prevalence of each suppressed state (Rakshit et al., 2017). In human movement, a finite-time basin-of-stability computation for the Sit-to-Stand task yields a personalized measure that differentiates less and more stable strategies through the normalized volume of the set of perturbations from which standing can still be reached (Shia et al., 2016).

6. Limitations and current directions

Basin stability inherits several methodological sensitivities. First, it depends on the perturbation measure $$\mathrm{BS}(A)=\int_X \rho(x)\,\Theta_A(x)\,dx =\int_{B(A)} \rho(x)\,dx,$$0, so reported values are only meaningful together with the sampling strategy. Second, attractor classification can be nontrivial in high-dimensional, delayed, or weakly chaotic systems. In chimera studies, for example, strength of incoherence depends on binning, thresholds, and time averaging, while mean phase velocity can be insufficient for amplitude chimera (Rakshit et al., 2017). In finite-time or observational settings, convergence detection, censoring by subsequent perturbations, and nonstationarity can bias estimates (Kan et al., 2021).

Numerical precision and integration time can also become dominant error sources. Near crises or in systems with long chaotic transients, finite-time truncation can blur discontinuities. In riddled or intermingled basins, rounding errors may exceed sampling error and make Monte Carlo estimates unreliable over broad sampling domains (Schultz et al., 2016). For delay systems, one must also verify convergence with respect to the chosen basis and polynomial degree used to represent history functions (Rakshit et al., 2017).

Current work increasingly seeks surrogates and learned approximations for basin stability. In power grids, graph neural networks have been trained to predict single-node basin stability directly from graph structure and minimal node features, with the best model achieving $$\mathrm{BS}(A)=\int_X \rho(x)\,\Theta_A(x)\,dx =\int_{B(A)} \rho(x)\,dx,$$1 and accuracy $$\mathrm{BS}(A)=\int_X \rho(x)\,\Theta_A(x)\,dx =\int_{B(A)} \rho(x)\,dx,$$2 on 100-node synthetic grids, and also showing transfer from 20-node training to 100-node inference without retraining (Nauck et al., 2021). This line of work is motivated by the high computational cost of Monte Carlo estimation, which in that study was reported as about 45 hours for a 100-node grid and about 3 hours for a 20-node grid on one CPU (Nauck et al., 2021).

Another frontier appears in stochastic optimization. For preconditioned SGD, a recent analysis establishes a preconditioner-dependent basin-stability guarantee inside a well-behaved local region, with the lower bound on staying in the basin tied to geometry in the $$\mathrm{BS}(A)=\int_X \rho(x)\,\Theta_A(x)\,dx =\int_{B(A)} \rho(x)\,dx,$$3-norm and with convergence behavior controlled by the product of an effective condition number and a preconditioned noise level (Scott et al., 24 Nov 2025). This suggests that basin-stability ideas are no longer confined to classical nonlinear dynamics, but are increasingly being reformulated for data-driven systems, learning algorithms, and Scientific Machine Learning.

Across these developments, the main conceptual point remains unchanged: basin stability is a probability of return, not a local eigenvalue test. Its strength lies in converting basin geometry into a quantitative robustness measure under finite perturbations; its main difficulty lies in the fact that the answer is only as meaningful as the perturbation ensemble, attractor classifier, and computational approximation that define it.