---
title: Basin Separation Score Analysis
url: https://www.emergentmind.com/topics/basin-separation-score
type: topic
---

# Basin Separation Score Analysis

Taken together, the cited works suggest that **“Basin Separation Score”** is best understood as an umbrella label for quantitative summaries of how distinctly competing basins of attraction are separated, rather than as a single standardized formal object. In several of the relevant papers, the exact phrase does not appear; instead, the literature formalizes nearby quantities that answer different questions: the probability that a perturbation returns to a target attractor, the Bayes-optimal distinguishability of future-state distributions, the entropy of mixed basin boundaries, the radius of a local basin of convexity, or the probability of avoiding escape under repeated jump perturbations [1603.01844; 2605.24136; 2005.04893; 2607.09593; 1711.03857].

## 1. Terminological status and family of related quantities

The first point of clarification is negative but decisive: the cited literature repeatedly states that the exact phrase **“Basin Separation Score”** is not the formal term used in the underlying papers. In "Potentials and Limits to Basin Stability Estimation" [1603.01844], the relevant scalar quantity is **basin stability**, defined as a basin-volume or return-probability measure. In "Detecting Metastable Basins in High Dimensions via Marginal Trajectory Distribution Discrimination" [2605.24136], the closest analogue is **Bayes-optimal 0-1 classification risk** or, equivalently, **total variation distance** between future-state marginals. In "Unpredictable basin boundaries in restricted six-body problem with square configuration" [2005.04893], the closest quantitative proxies are **basin entropy** \(S_b\) and **boundary basin entropy** \(S_{bb}\). In "Characterization of the basin of convexity for multi-snapshot spike deconvolution via variable projection" [2607.09593], the relevant object is an explicit **basin of convexity** with radius \(\varrho\). In "Bounding the first exit from the basin: Independence Times and Finite-Time Basin Stability" [1711.03857], the key objects are **finite-time basin stability** and **independence time**.

| Perspective | Closest formal quantity | Representative paper |
|---|---|---|
| Basin volume / return probability | \(B(A)=\mu(B(A))\), \(BS(S)=V_s/V\), node-wise \(B_i(K)\) | [1603.01844], [1704.05301], [1602.01712] |
| Distributional separability | \(R^*_{t^*}(x_i,x_j)\), \(d_{\mathrm{TV}}\!\bigl(P_{t^*}(x_i,\cdot),P_{t^*}(x_j,\cdot)\bigr)\) | [2605.24136] |
| Boundary unpredictability | \(S_b\), \(S_{bb}\) | [2005.04893] |
| Local convexity / optimization basin | \(\varrho\), Hessian spectral bounds | [2607.09593] |
| Repeated-perturbation insulation | \(FTBS_S(T)\), \(T_{\mathrm{ind}}(\epsilon,\delta)\) | [1711.03857] |
| Basin localization in optimization | Hessian eigenspectrum, \(\Delta E\), success probability | [2605.09909] |

This suggests that the term is not best treated as a single invariant of a landscape. It is instead a family resemblance among several basin-sensitive quantities, each emphasizing a different notion of separation: **volume**, **predictability**, **statistical distinguishability**, **local convexity**, or **robustness under repeated perturbation**.

## 2. Basin volume and return-probability formulations

In the basin-stability literature, the defining idea is probabilistic rather than geometric. For a dynamical system
\[
\dot{x}=F(x,t),
\]
with attractor \(A\) and basin of attraction \(B(A)\), basin stability is
\[
B(A):=\mu(B(A))=\int_R \mathbf{1}_{B(A)}(x)\,d\mu(x)\in[0,1].
\]
The interpretation is explicit: basin stability is **“the probability that the system will return to \(A\)”** after a perturbation drawn from \(\mu\) [1603.01844]. The same paper also states that **“we are not interested in the basin of attraction’s geometry but only in its volume w.r.t. the measure \(\mu\)”**. That sentence sharply distinguishes basin stability from any geometric notion of basin separation.

Because the integral is usually intractable, basin stability is estimated by Monte Carlo sampling. If \(N\) perturbations are drawn and \(M\) trajectories converge to the target attractor, the estimator is
\[
S=\frac{M}{N},
\qquad
\hat S_B=\frac{1}{N}\sum_{i=1}^N \mathbf{1}\{x_i\in B(A)\}.
\]
The corresponding sampling standard error is
\[
\sqrt{\frac{B(A)(1-B(A))}{N}}.
\]
The same work emphasizes that sampling error is only one component of total estimation error; approximation, integration, and rounding error can dominate when basin geometry is intricate [1603.01844].

A finite-time extension makes the return notion temporal rather than asymptotic. For a return surface \(S\) and a time-tracking Lyapunov construction, finite-time basin stability is
\[
FTBS_S(T)=\int_X 1_B(x)\,\Theta\!\left(T-V_S(x)\right)\rho(x)\,dx.
\]
The associated **independence time**
\[
T_{\mathrm{ind}}(\epsilon,\delta)
=
\inf\{T>0\mid BS-FTBS_S(T)\le \delta\}
\]
quantifies the time after which a perturbed trajectory has probably returned close enough to the attractor that subsequent perturbations can be treated approximately independently. If jump perturbations are at least \(T_{\mathrm{ind}}(\epsilon,\delta)\) apart, the remain probability is bounded below by
\[
P_{\mathrm{remain}}(t,x(0))
\ge
(BS-\delta-\epsilon)^{n(t)}.
\]
In this formulation, separation is operationalized as **insulation from basin escape** under repeated shocks, rather than as a geometric distance to a boundary [1711.03857].

The same basin-volume logic appears in applications. In chimera-state analysis, basin stability is estimated empirically as
\[
BS(S)=\frac{V_s}{V},
\]
where \(V_s\) initial histories converge to a given collective state among \(V\) sampled histories. In small power-grid motifs, node-wise basin stability is treated as a function \(B_i(K)\) of coupling strength, and transition curves are studied across topologies and producer-consumer placements [1704.05301; 1602.01712].

## 3. Statistical distinguishability and local convexity

A more direct route to something that behaves like a **separation score** appears in metastable Markov processes. For a time-homogeneous Markov process with transition kernel
\[
P_t(x,A)=\mathbb P(X_t\in A\mid X_0=x),
\]
the paper on marginal trajectory distribution discrimination reduces basin identification to a two-sample discrimination problem [2605.24136]. Two initial states \(x_0,x_1\) are compared through their time-\(t^*\) endpoint marginals
\[
P_{t^*}(x_0,\cdot),
\qquad
P_{t^*}(x_1,\cdot).
\]
The central result is a **classification-risk separation theorem**: if \(x_0,x_1\) lie in the same basin core, the Bayes-optimal classifier has risk close to \(1/2\); if they lie in different basin cores, the optimal risk is close to \(0\). The decision-theoretic and probabilistic forms are equivalent through
\[
R^*(P,Q)=\frac12\bigl(1-d_{\mathrm{TV}}(P,Q)\bigr).
\]
Accordingly, the most direct induced pairwise separation score is
\[
S(x_i,x_j)
=
d_{\mathrm{TV}}\!\bigl(P_{t^*}(x_i,\cdot),P_{t^*}(x_j,\cdot)\bigr)
=
1-2R^*_{t^*}(x_i,x_j).
\]
Here **same-basin behavior** corresponds to \(R^*\approx 1/2\) and \(S\approx 0\), whereas **different-basin behavior** corresponds to \(R^*\approx 0\) and \(S\approx 1\). The practical algorithm uses held-out neural classification risk \(\widehat R_{ij}\) and merges representatives when
\[
\widehat R_{ij}>\gamma.
\]
This is the most explicit instance in the cited material where basin separation is turned into a quantitative, pairwise discriminability problem [2605.24136].

A different but related perspective arises in variable-projection spike deconvolution. There the issue is not pairwise distinguishability of basins but the existence of a certified **basin of convexity** around the true spike locations. The governing quantity is the minimum separation \(\Delta\), compared against PSF roughness and amplitude dynamic range via the threshold
\[
\Delta>\frac{2}{3}\rho \kappa^2.
\]
When this holds, the VarProSD objective has an explicit neighborhood
\[
\mathcal N(\bm\tau,\varrho)
=
\{\bm\gamma\in\mathbb R^K:\mathrm d_2(\bm\gamma,\bm\tau)\le \varrho\},
\]
with radius
\[
\varrho
=
\frac12\!\left(\Delta-\frac23\rho\kappa^2\right)
\;\wedge\;
c_1\kappa^{-2}\sqrt{\frac{E_g}{E_{g'}}\wedge\frac{E_{g'}}{E_{g''}}}.
\]
Within this basin, the Hessian satisfies
\[
\sigma_{\min}\!\left(\nabla_{\bm\gamma}^2\ell(\bm\gamma)\right)
\ge
\frac13 E_{g'} T r_{\min}^2(\bm X),
\qquad
\sigma_{\max}\!\left(\nabla_{\bm\gamma}^2\ell(\bm\gamma)\right)
\le
E_{g'} T r_{\max}^2(\bm X),
\]
and gradient descent converges linearly when initialized inside the neighborhood [2607.09593]. In this setting, a basin-separation quantity is local, strongly tied to conditioning, and explicitly dependent on sampling bandwidth, PSF smoothness, and spike separation.

## 4. Boundary complexity, entropy, and geometric pathology

A basin may have large volume and still be poorly separated geometrically. The uncertainty-fraction analysis in basin-stability theory captures this distinction. If initial conditions are known only up to numerical uncertainty \(\varepsilon\), the uncertainty fraction \(f(\varepsilon)\) scales as
\[
f(\varepsilon)\propto \varepsilon
\]
for smooth basin boundaries, but
\[
f(\varepsilon)\propto \varepsilon^\alpha
\]
for fractal basin boundaries, with \(\alpha<1\) implying final-state sensitivity [1603.01844]. More extreme pathologies are **riddled basins**, whose complement intersects every disk in a set of positive measure, and **intermingled basins**, where any open set intersecting one basin in positive measure also intersects each of the others in positive measure. In such cases, finite-precision basin assignment becomes practically non-deterministic.

The restricted six-body problem gives an explicit entropy-based quantification of poor basin separation. After partitioning the basin image into boxes, the box entropy is
\[
S_i=\sum_{j=1}^{m_i} p_{ij}\log\!\left(\frac{1}{p_{ij}}\right),
\]
with total entropy
\[
S=\sum_{i=1}^{N}S_i,
\qquad
S_b=\frac{S}{N},
\qquad
S_{bb}=\frac{S}{N_b},
\]
where \(N_b\) is the number of boxes containing more than one color [2005.04893]. The paper uses the criterion that if \(S_b\) or \(S_{bb}\) is greater than \(\log 2\), then the basin or its boundaries are fractal. The specific conclusion is that for \(\mu=0.22\) and \(0.23\), the basin of attraction is **“unpredictable throughout”**, while for **all** values of \(\mu\), \(S_{bb}>\log 2\), so the basin boundaries are highly unpredictable. The same work also reports Wada boundary evidence, implying that arbitrarily small neighborhoods of boundary points contain multiple basin colors.

A further distinction between **volume** and **geometry** appears in the study of mechanically stable particle packings. There, the rank-ordered basin-volume distribution satisfies
\[
P_N(n)\approx A_N n^{-\alpha},
\qquad
\alpha\approx 1,
\]
and may extend over \(7\) decades [2607.08094]. Yet basin geometry is not exhausted by volume alone. With a Hessian-mode hyper-rectangular compact volume \(V_{\mathrm{comp}}(n)\), the paper defines a contortion measure
\[
E_N(n)=\frac{V_N(n)}{V_{\mathrm{comp}}(n)}.
\]
Large variation in \(E_N(n)\) shows that equal-volume basins can differ greatly in shape. This suggests that any basin-separation concept based only on volume or accessibility omits potentially decisive information about boundary contortion and local exit directions.

## 5. Operational proxies in optimization and applied systems

In variational quantum eigensolvers, the language shifts from basin separation to **basin localization** and **basin targeting**. The energy landscape is
\[
E(\boldsymbol\theta;\mathbf R)=\bra{\psi(\boldsymbol\theta)}\hat H(\mathbf R)\ket{\psi(\boldsymbol\theta)},
\]
and basin membership is assessed through the local quadratic expansion
\[
E(\boldsymbol{\theta}_0+\delta\boldsymbol{\theta})
\approx
E(\boldsymbol{\theta}_0)
+
\nabla E(\boldsymbol{\theta}_0)^\top\delta\boldsymbol{\theta}
+
\frac12\delta\boldsymbol{\theta}^\top\mathcal H(\boldsymbol{\theta}_0)\delta\boldsymbol{\theta}.
\]
Negative Hessian eigenvalues indicate escape directions toward competing basins; a nonnegative spectrum up to gauge modes indicates a locally convex region associated with the target basin. The paper also uses initialization error
\[
\Delta E=E_{\mathrm{init}}-E_{\mathrm{ref}}
\]
and improvement ratios such as \(38\times\)–\(6250\times\) as proxies for whether initialization selected the correlated ground-state basin rather than a competing one [2605.09909]. Here the relevant “score” is neither global nor purely geometric; it is a practical diagnostic of correct basin targeting before local refinement.

In multi-band astronomical source separation, the score-matching prior paper explicitly states that it does **not** define or evaluate an explicit “Basin Separation Score” and does **not** directly measure distances between optimization basins, attraction-region volumes, or margins between competing source assignments [2401.07313]. Nevertheless, it offers concrete proxies for basin quality: convergence under severe initialization perturbations, variance of recovered flux and morphology metrics, dependence on blendedness
\[
\beta_k = 1 - \frac{\mathbf{S}_k \cdot \mathbf{S}_k}{\mathbf{S} \cdot \mathbf{S}_k},
\]
and prior-vs-likelihood dominance through the hallucination score
\[
\Delta \mathbf{F}
=
\mathrm{Diag}(\mathbf{F}_{\log \mathcal{P}})
-
\mathrm{Diag}(\mathbf{F}_{\log \mathcal{L}}),
\qquad
\gamma=-\Delta\mathbf F\cdot\mathbf S.
\]
These are not separation scores in the strict sense, but they operationalize how easily optimization escapes bad local regions and returns to physically plausible solutions.

In networked dynamical systems, basin-separation ideas often appear through parameterized basin-stability transitions rather than explicit distances. For power-grid swing dynamics, node-wise basin stability \(B_i(K)\) is treated as a function of coupling strength, and the transition curves are embedded in
\[
\bigl(B_i(7),B_i(14),B_i(21)\bigr).
\]
The resulting classes correlate more clearly with betweenness and flow betweenness than with degree or edge density [1602.01712]. For chimera states in delay-coupled Mackey-Glass networks, the basin shares of incoherent, chimera, and coherent states are estimated over polynomially parameterized initial history functions, with the collective-state label determined by the strength of incoherence [1704.05301]. In both examples, basin separation is read through **relative basin shares** and **transition classes**, not through a single geometric margin.

## 6. Conceptual distinctions, limitations, and synthesis

A recurring misconception is that a basin-based scalar automatically measures geometric basin separation. The literature does not support that identification. Basin stability measures basin volume under a perturbation measure, not boundary smoothness, fractality, Wada structure, riddling, or local distance to a competitor. Two attractors can therefore have identical basin stability and radically different basin geometries [1603.01844]. Conversely, a high-entropy or high-total-variation boundary diagnostic says little by itself about asymptotic return probability.

Another limitation is context dependence. Distributional separation depends on the chosen observation horizon \(t^*\) and observation space; the metastability paper explicitly notes that same-basin and different-basin distinguishability can reverse if \(t^*\) is chosen outside the ideal within-basin-mixing / pre-transition regime [2605.24136]. Convexity-basin size depends on PSF descriptors, bandwidth, and amplitude dynamic range [2607.09593]. Perturbation-based robustness depends on the perturbation distribution \(\rho\), the return surface \(S\), and the jump spacing relative to \(T_{\mathrm{ind}}(\epsilon,\delta)\) [1711.03857]. Numerical reliability can collapse under riddled or intermingled basin geometry, even when Monte Carlo sampling error appears small [1603.01844].

Taken together, these works suggest a useful taxonomy of basin-separation quantities. A **global return-probability score** is represented by \(B(A)\), \(BS\), or \(FTBS_S(T)\). A **pairwise separation score** is represented by \(d_{\mathrm{TV}}\) or \(1-2R^*\). A **boundary unpredictability score** is represented by \(S_b\) or \(S_{bb}\). A **local optimization-basin score** is represented by \(\varrho\), Hessian spectral positivity, or a curvature bound. An **accessibility proxy** is represented by initialization error \(\Delta E\), success probability, convergence robustness, or basin-volume rank distributions [1603.01844; 2605.24136; 2005.04893; 2607.09593; 1711.03857; 2605.09909; 2607.08094]. In that sense, **Basin Separation Score** is best treated not as a single established metric, but as a task-dependent label for whichever of these quantitatively answers the relevant separation question.

Source: https://www.emergentmind.com/topics/basin-separation-score