Basin-of-Attraction Analysis

• Basin-of-attraction analysis defines the set of initial conditions that converge to attractors, characterizing stability, robustness, and fractal boundaries.
• Numerical methods like Monte Carlo sampling and grid-based iterative schemes robustly estimate basin stability while quantifying uncertainties in attractor convergence.
• Advanced techniques such as finite-state machine approaches and deep learning classifiers enhance the detection of intricate basin geometries, aiding control in nonlinear systems.

A basin of attraction is the set of initial conditions in the phase space of a dynamical system that asymptotically converge to a given attractor—such as a fixed point, limit cycle, or more complex invariant set—under the system's evolution. Basin-of-attraction analysis rigorously characterizes not only which states attract a system but also quantifies the geometry, fractality, and robustness of the basins, which is crucial for understanding stability, multistability, and unpredictability in nonlinear systems.

1. Mathematical Definition and Probabilistic Characterization

Given an autonomous dynamical system

$\dot{x} = f(x),\qquad x\in X\subseteq\mathbb{R}^n,$

an attractor $A$ is a minimal compact invariant set with a basin of attraction

$$B(A) = \bigl\{ x\in X : \varphi_t(x) \to A \text{ as } t\to\infty \bigr\},$$

where $\varphi_t(x)$ is the trajectory initialized at $x$. The basin of attraction quantifies the set of perturbations from which the system returns to the attractor.

Under external or modeling uncertainty, the concept of basin stability is introduced: If a probability measure $\mu$ (with density $\rho(x)$) encodes the distribution of perturbations in state space, then the basin stability of $A$ is

$$\beta = \mu\bigl(B(A)\bigr) = \int_X \mathbf{1}_{B(A)}(x)\,\rho(x)\,\mathrm{d}x\in[0,1].$$

This integral is high-dimensional and thus typically estimated numerically: $$\hat\beta = \frac{1}{N}\sum_{i=1}^N \mathbf{1}_{B(A)}(x_i)$$ via Monte Carlo sampling with initial states $x_i\sim\rho$ (Schultz et al., 2016).

2. Numerical Estimation Methodologies

Basin-of-attraction analysis relies on a variety of computational approaches, adapted to the nature of the system and the structural complexity of its basins:

• Monte Carlo Sampling: State space is sampled randomly, and trajectories are numerically evolved to determine their asymptotic fate. The sampling error in $\hat\beta$ is quantified by

$\sigma_{\rm MC} = \sqrt{\frac{\beta(1-\beta)}{N}}.$

• Grid-Based Iterative Schemes: For root-finding algorithms (e.g., Newton-Raphson) or for low-dimensional continuous systems, a grid of initial conditions is evolved, and each point is labeled by the attractor it reaches. This approach enables visualization and quantification of basin geometry. Example: the pseudo-Newtonian restricted three-body problem (Zotos, 2017).
• Finite-State Machine Approach: The phase space is partitioned into rectangular cells; cell-to-cell transitions over a fixed time interval are computed explicitly, creating a graph representation of the flow. Strongly connected components (SCCs) identify attractor regions, and graph search labels the basin of each cell. This method is highly parallelizable and enables automatic attractor discovery without prior knowledge (Datseris et al., 2021).
• Machine Learning/Clustering Techniques: In high-dimensional or network systems, one can cluster the statistics of trajectory endpoints (e.g., means/variances or histograms of node observables in oscillator networks) using unsupervised algorithms (e.g., DBSCAN). Approximate basin volumes are inferred from cluster populations (Gelbrecht et al., 2019).
• Deep Learning Classifiers: For systems where direct integration is computationally intensive or boundaries are highly intricate, supervised learning (e.g., dense or convolutional neural nets) is used. Basins are interpreted as a classification problem: a neural network is trained to predict the attractor label from an initial state, and the decision surface approximates the basin boundary (Shena et al., 2021, Valle et al., 2023).

3. Geometric and Topological Complexity

Basin boundaries can range from smooth manifolds to fractal and riddled structures with profound implications for predictability:

• Smooth boundaries: The uncertainty (or box-counting) exponent $\alpha=1$ leads to f($\varepsilon$)—the fraction of initial conditions within distance $\varepsilon$ of the boundary—scaling as $\varepsilon$. Prediction is robust except near the separatrix.
• Fractal boundaries: When $n-1<D<n$, the boundary has fractal dimension $D$. The uncertainty exponent $\alpha<1$ implies increased sensitivity to initial conditions. Empirically, even for highly fractal (e.g., Wada) boundaries, ensemble averages such as basin stability estimates remain statistically robust (Schultz et al., 2016).
• Riddled and Intermingled Basins: Riddled basins have empty interior; in any neighborhood, every other attractor's basin appears with positive measure. Intermingled basins occur when every open set intersecting one basin also intersects others with positive measure. In these scenarios, numerical and finite-precision uncertainties dominate and render basin-volume estimates unreliable (Schultz et al., 2016).
• Wada boundaries: Every boundary point is shared by at least three basins, often verified by the merging method (boundary remains invariant upon color-merging) (Valle et al., 2023, Daza et al., 2016).
• Basin entropy: Quantifies unpredictability by measuring the frequency-mixing of attractor outcomes in small spatial boxes, enabling single-number comparisons across systems and parameter regimes (Daza et al., 2016, Daza et al., 2022).

4. Applications and Illustrative Examples

Basin-of-attraction analysis appears across dynamical and computational disciplines:

• Celestial mechanics: Restricted multi-body problems analyzed via Newton-Raphson reveal intricate, multi-lobed, and highly fractal basins, whose structure and fractality are controlled by parameters such as mass ratio and oblateness (Zotos, 2017, Zotos, 2016, Zotos, 2018, Zotos et al., 2018).
• Oscillator networks: Kuramoto-type models underpinning explosive synchronization and hysteresis require basin-volume calculations to determine thresholds for incoherent/synchronous transitions, directly linking network parameters and emergent synchronization via bifurcations in basin boundaries (Zou et al., 2014, Delabays et al., 2017).
• Multistable flows and maps: E.g., the Lorenz system in bistable regimes, Hénon–Heiles Hamiltonian, and driven Duffing oscillator, exhibit sharp increases in basin entropy and reduction in classifier accuracy upon entrance to fractal or chaotic regimes (Shena et al., 2021, Daza et al., 2016).
• Chimera states and coupled oscillators: Low-dimensional reductions (Ott–Antonsen, invariant manifold tracking) plus destination maps elucidate the geometry of chimera basins and the sensitivity to parameters and initial configuration (Martens et al., 2015).
• Control and minimization: Comparative basin analysis for various optimization algorithms demonstrates trade-offs between convergence speed and boundary integrity; step-size tuning can induce fragmentation or fractality in the basins of minimizers (Asenjo et al., 2013).
• Computational intractability: There exist analytic, globally hyperbolic systems (even with explicit vector fields and sinks) whose basins of attraction are non-computable, emphasizing formal limitations in exact basin determination (Graça et al., 2014).

5. Quantitative Metrics and Characterization

Several fundamental metrics have been developed to quantify basin structure and unpredictability:

Metric	Formula/Method	Interpretation
Basin Stability $\beta$	$\int_X \mathbf{1}_{B(A)}(x)\rho(x)\,\mathrm{d}x$	Probability a perturbation returns to $A$
Uncertainty exponent $\alpha$	$f(\varepsilon)\sim\varepsilon^\alpha$	Sensitivity of boundary, fractality ($\alpha<1$)
Basin entropy $S_b$	$-\frac{1}{N}\sum_{boxes}\sum_{j} p_j \ln p_j$	Degree of unpredictability (Daza et al., 2016)
Boundary entropy $S_{bb}$	$S_{bb}=S/N_b$, boxes near boundaries	$S_{bb}>\log 2$ indicates fractal structure
Fractal dimension $D$	Box-counting or Monte Carlo scaling	Spatial complexity of boundary

Basin entropy unifies unpredictability across different forms of mixing—fractal, riddled, Wada, or lacunar (disconnected) boundaries. Riddled basins yield $S_b = S_{bb}$ and maximal unpredictability for any covering scale, while Wada sets guarantee every local box at the boundary is maximally mixed (Daza et al., 2022).

6. Practical Guidelines, Robustness, and Limits

Robust estimation of basin properties demands careful numerical protocol:

• Sampling and Integration: Use sufficient sample sizes ($N$) to shrink statistical error ($\sigma_{\rm MC}$) below application tolerance.
• Precision Testing: Compute $\hat\beta$ at high and low numerical precision; if the difference exceeds statistical error, finite-precision effects dominate.
• Visualization and Fractality Detection: Visual cross-sections, zoom-in refinement, and scaling of uncertainty fraction ($f(\varepsilon)$) help identify intricate boundaries.
• **Choose grid, ODE solver tolerance, or step-size for algorithms so that trajectories resolve boundary intricacies without overwhelming computational cost.
• High Dimensionality: For large $n$, exhaustive gridding becomes infeasible; random-IC Monte Carlo, clustering, or feature-based dimensionality reduction is necessary (Datseris et al., 2021, Gelbrecht et al., 2019).

Monte Carlo and ensemble methods are robust for basins with positive measure and non-pathological boundaries; they break down in riddled or intermingled regimes where the concept of basin volume is ill-posed at finite precision (Schultz et al., 2016).

7. Recent Developments and Future Directions

Recent research features several significant advances:

• Automated and Scalable Basin Analysis: Finite-state-machine and graph approaches enable basin mapping without prior knowledge of attractors, with substantial speed-ups over naive grid evolution (Datseris et al., 2021).
• Deep Learning Approaches: CNNs and fully-connected nets achieve rapid, accurate estimation of basin entropy, box-counting dimension, and Wada property classification in large parameter scans, with computational cost orders of magnitude lower than traditional sampling (Shena et al., 2021, Valle et al., 2023).
• Basin Entropy Frameworks: Basin entropy, with explicit classification for smooth, fractal, riddled, Wada, and intermingled basins, offers a unifying taxonomy for basin complexity, aiding both qualitative insight and quantitative prediction regarding stability and multistability (Daza et al., 2016, Daza et al., 2022).
• Scaling Laws in High Dimensions: For oscillator networks and power-grid problems, analytical and numerical work reveals that basin volumes can exhibit non-Gaussian (exponential) scaling with network size, directly impacting systemic risk and synchronization transitions (Delabays et al., 2017, Zou et al., 2014).

Ongoing open problems include:

• Efficiently quantifying basin structure in very high-dimensional systems,
• Accurately resolving extremely thin or riddled basin boundaries,
• Relating basin properties to control and stabilization strategies, and
• Extending entropy-based measures to non-Euclidean or functional spaces.

---

Key references:

(Schultz et al., 2016) (Potentials and Limits to Basin Stability Estimation) (Zotos, 2017) (Basins of convergence: pseudo-Newtonian three-body) (Daza et al., 2016) (Basin entropy) (Daza et al., 2022) (Classifying basins: basin entropy) (Shena et al., 2021, Valle et al., 2023) (Deep learning-based analysis) (Datseris et al., 2021) (Effortless estimation of basins) (Gelbrecht et al., 2019) (Monte Carlo Basin Bifurcation Analysis) (Graça et al., 2014) (Non-computable basins of attraction)