A theory of spatial early warning signals for tipping points on complex networks

Abstract: Spatial early warning signals (EWSs) seek evidence of an approaching tipping point from a single snapshot of many interacting elements. Existing theory largely assumes spatial homogeneity, whereas networks introduce systematic differences among nodes that may obscure fluctuation-based warning signals. We develop a mathematical framework for spatial EWSs in stochastic dynamical systems on networks. We find that the expected spatial variance, a popular spatial EWS, decomposes exactly into a structural contribution from heterogeneity in the equilibrium state and a fluctuation contribution determined by the stationary covariance. Near a simple steady-state bifurcation, the potentially divergent covariance concentrates along the critical eigendirection: the left eigenvector determines how strongly noise excites the critical fluctuation, while the right eigenvector determines its spatial pattern. Consequently, the spatial variance has a divergent fluctuation contribution when the limiting critical eigendirection is noise-excited and spatially nonuniform after centering. In contrast, the spatial coefficient of variation generally saturates, while skewness, kurtosis, and Moran's $I$ approach network-dependent limits without a universal warning direction. We also derive results for homogeneous networks, node-wise baseline subtraction as preprocessing, and Hopf bifurcations, for which the limiting distributions are qualitatively different. These results clarify when spatial EWSs provide reliable warnings and why their performance depends on network structure, noise, and preprocessing.

• The paper develops a structural–fluctuation decomposition showing that spatial variance diverges near tipping points only when noise excites a nonuniform critical eigenvector, with rates of ε⁻¹ᐟ² for saddle-node and ε⁻¹ for transcritical or pitchfork bifurcations.
• On heterogeneous networks, structural variation can reinforce, mask, or reverse variance trends, while skewness, kurtosis, Moran’s I, and often the coefficient of variation saturate without a universal warning direction.
• The analysis identifies baseline-referenced spatial variance as the strongest practical candidate for heterogeneous systems, but warns that snapshot uncertainty remains high, with relative variance error approaching √2 and limits constrained by noise-driven escape from the local basin.

Motivation and scope

Spatial early warning signals (EWSs) replace repeated temporal sampling with a single snapshot across many interacting nodes, computing statistics such as spatial variance, coefficient of variation (CV), skewness, kurtosis, and Moran's $I$ across node states. Existing analytical treatments assume spatially homogeneous domains, where all nodes share a uniform equilibrium, yet empirical systems subject to tipping are typically heterogeneous networks on which numerical studies report erratic indicator behavior: which statistic works, and whether it rises or falls, depends on the system, network, and control parameter (2608.15476). The paper develops a mathematical theory of these five standard spatial EWSs for Itô stochastic dynamics $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$ on general networks, linearized about a stable equilibrium branch that loses stability at $\alpha_{\mathrm c}$.

The central organizing result is an exact decomposition of the expected spatial variance into a structural term $S(\alpha) = Var(x^*(\alpha))$, arising from heterogeneity in equilibrium node states, and a fluctuation term $\Phi(\alpha) = Tr(PC)/(N-1)$, determined by the stationary covariance $C$ of the Ornstein–Uhlenbeck (OU) fluctuation, with $P$ the centering projector. This decomposition requires only $E[z]=0$, no linearization or proximity to bifurcation. On homogeneous domains $S \equiv 0$, which is why lattice-based theory has not had to confront the deterministic background that can dominate—or even reverse the trend of—the raw signal on heterogeneous networks.

Critical-mode structure and divergence criteria

Under a regular simple real critical eigenvalue (covering generic saddle-node, transcritical, and pitchfork bifurcations), the stationary covariance obeys a rank-one asymptotic: $$C = (\sigma_{\mathrm{eff}}^2/2\lambda_1)\,vv^\top + O(1)$$, where $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$0 is the vanishing eigenvalue and $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$1 measures how strongly noise excites the critical mode through the left eigenvector $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$2. A critical-coordinate decomposition shows that near the transition any snapshot is governed by a single diverging random scalar $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$3 multiplying the right eigenvector $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$4: $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$5 with $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$6.

The main theorem gives

$$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$7

where $$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$8 quantifies uniformity of the critical right eigenvector. The spatial variance therefore diverges if and only if the limiting critical eigendirection is both noise-excited ($$\mathrm{d}x = F(x;\alpha)\,\mathrm{d}t + B\,\mathrm{d}W$$9) and non-uniform after centering ($\alpha_{\mathrm c}$0). For cooperative irreducible dynamics—mutualistic, SIS, gene-regulatory models in their cooperative regimes—Perron–Frobenius theory guarantees entrywise-positive critical eigenvectors, so $\alpha_{\mathrm c}$1 and the noise-excitation cancellation is excluded when every node receives noise; only the uniform-eigenvector cancellation may occur. At a generic saddle-node bifurcation $\alpha_{\mathrm c}$2; at transcritical/pitchfork bifurcations, $\alpha_{\mathrm c}$3.

By contrast, the structural term remains bounded but its derivative generically diverges with a sign that can be positive or negative depending on system, network, and parameter direction—so $\alpha_{\mathrm c}$4 can reinforce, mask, or reverse the fluctuation-driven trend. This explains numerically observed reversals of indicator direction across control parameters.

Saturation of scale-invariant statistics

Because skewness, kurtosis, and Moran's $\alpha_{\mathrm c}$5 are invariant to rescaling of centered deviations, they divide out the growing amplitude $\alpha_{\mathrm c}$6. All three converge to finite, network-dependent limits determined solely by the limiting critical right eigenvector $\alpha_{\mathrm c}$7: skewness converges in distribution to a signed two-point variable taking $\alpha_{\mathrm c}$8 with probability $\alpha_{\mathrm c}$9 each (so $S(\alpha) = Var(x^*(\alpha))$0), while kurtosis and Moran's $S(\alpha) = Var(x^*(\alpha))$1 converge to deterministic constants involving moments and quadratic forms of $S(\alpha) = Var(x^*(\alpha))$2. None diverges, and none provides a universal warning direction—a converging statistic is interpretable only with a known direction of approach, which these lack.

The CV behaves differently: since both numerator and denominator share the factor $S(\alpha) = Var(x^*(\alpha))$3, it generally saturates, converging in probability to $S(\alpha) = Var(x^*(\alpha))$4 when $S(\alpha) = Var(x^*(\alpha))$5, tending to 0 when $S(\alpha) = Var(x^*(\alpha))$6, and diverging when $S(\alpha) = Var(x^*(\alpha))$7. An important caveat is stated plainly: these asymptotic limits describe the regime where the critical contribution dominates; at fixed nonzero noise the OU approximation may fail before this regime is reached because large excursions leave the local basin, so pre-asymptotic behavior can differ substantially.

Homogeneous networks recover classical behavior

For connected undirected regular networks with dynamics $S(\alpha) = Var(x^*(\alpha))$8 (and analogously for aggregate-input coupling), the uniform branch has stability matrix $S(\alpha) = Var(x^*(\alpha))$9, so the critical eigenvector is exactly $\Phi(\alpha) = Tr(PC)/(N-1)$0, $\Phi(\alpha) = Tr(PC)/(N-1)$1, and centering removes the critical mode entirely. Consequently $\Phi(\alpha) = Tr(PC)/(N-1)$2 and

$\Phi(\alpha) = Tr(PC)/(N-1)$3

a finite limit set by the Laplacian-like spectrum. This identifies precisely why indicators calibrated on lattices omit the deterministic heterogeneity background that dominates on irregular networks.

Snapshot uncertainty

Near the bifurcation, all node fluctuations share one random coordinate, so the usual $\Phi(\alpha) = Tr(PC)/(N-1)$4 precision gain from sampling many nodes fails. Using Isserlis' formula, the relative standard deviation of the spatial variance tends to $\Phi(\alpha) = Tr(PC)/(N-1)$5 regardless of $\Phi(\alpha) = Tr(PC)/(N-1)$6—the same value as for a single Gaussian sample variance. For the CV, ordinary moments do not exist because the Gaussian spatial mean has positive density at zero, so convergence in probability is the appropriate description. Skewness retains uncertainty equal to its signal magnitude (limiting variance $\Phi(\alpha) = Tr(PC)/(N-1)$7 from the random sign), while kurtosis and Moran's $\Phi(\alpha) = Tr(PC)/(N-1)$8 become asymptotically deterministic but network-dependent.

Baseline referencing

Subtractive baseline referencing—subtracting each node's mean over $\Phi(\alpha) = Tr(PC)/(N-1)$9 reference snapshots at $C$0—yields exactly

$C$1

It leaves the current fluctuation term, including its critical divergence, untouched; replaces the structural offset $C$2 by zero at the reference point; and adds a constant baseline-estimation error shrinking as $C$3. This provides theoretical support for the empirically improved performance of baselined spatial variance (Bandara et al., 6 Aug 2026), while making clear that improvement is not guaranteed over every parameter range, since $C$4 grows as $C$5 departs from $C$6. Corollaries show fixed baselines leave the CV, skewness, kurtosis, and Moran's $C$7 limits unchanged—explaining why baselining rescues variance-based indicators but not the others. Ratio referencing is analyzed under the assumption that baselines are bounded away from zero; it preserves the $C$8 divergence rate but modifies the amplitude via $C$9, and can amplify baseline-estimation error when some $P$0 is small.

Hopf bifurcations

At a Hopf bifurcation the critical subspace is two-dimensional, spanned by $P$1 and $P$2, with covariance $P$3 and $P$4. The expected spatial variance still diverges as $P$5 when the pair is noise-excited, and the structural term stays bounded with no square-root singularity because the equilibrium continues smoothly through the bifurcation. Qualitative differences arise: the CV, kurtosis, and Moran's $P$6 retain non-degenerate limiting distributions (governed by a uniformly random phase within the critical plane) rather than deterministic constants, and the relative uncertainty of the spatial variance lies between $P$7 and $P$8 rather than universally $P$9. Notably, scalar-state dynamics on symmetric homogeneous networks cannot undergo Hopf bifurcations under these hypotheses, since $E[z]=0$0 has real spectrum.

Limitations and open questions

The author concedes several restrictions. The analysis relies on linearization; as $E[z]=0$1, stochastic excursions eventually exit the local neighborhood of the equilibrium, so the OU approximation may lose accuracy before the deterministic bifurcation is reached, and observability of the predicted limits depends on noise strength, basin geometry, and rate of parameter variation. Noise is additive and white, states are scalar per node, and the network is static. The paper analyzes the statistics themselves rather than the power of specific detection algorithms, and treats only standard local bifurcations—rate-induced tipping, noise-induced tipping, and other transition types require separate treatment. Whether the ultimate asymptotic regimes derived here are reachable in practice before noise-driven escape, and how the theory extends to time-varying networks or multivariate node states, remain open.

Conclusion

This paper supplies the first systematic analytical account of why classical spatial EWSs behave inconsistently on heterogeneous networks. Its core contributions are the exact structural–fluctuation decomposition of the spatial variance, the identification of the left/right eigenvector roles (noise excitation versus spatial pattern), the divergence criterion $E[z]=0$2, the saturation results for scale-invariant statistics, the persistent $E[z]=0$3 relative uncertainty, and the precise characterization of what subtractive baseline referencing does and does not accomplish. The practical upshot is that the baseline-referenced spatial variance is a principled default candidate for heterogeneous networks, while skewness, kurtosis, Moran's $E[z]=0$4, and ultimately the CV offer no universal warning direction.