Baseline-referenced spatial early warning signals for tipping points on heterogeneous networks

Abstract: Anticipating tipping points in complex systems is difficult because many early warning signals require long time series, which are often unavailable in practice. Spatial early warning signals offer an alternative by using a single snapshot across many interacting elements, or nodes. However, their performance in heterogeneous systems is often inconsistent because raw node states reflect both dynamical changes associated with an approaching transition and static heterogeneity induced by network structure. Here, we propose a baseline-referenced framework for spatial early warning signals. The method compares each node's state with its own baseline far from the tipping point before computing a spatial statistic, thus reducing network-structure-induced variation. We evaluate baseline-referenced variants of five classical spatial early warning signals across diverse tipping scenarios and networks, and find that baseline referencing markedly improves variance-based spatial signals. The best variants increase consistently and progressively toward tipping points across different scenarios, outperform a single-node temporal variance that requires long time series, and retain high performance even when up to 80% of nodes are omitted from observation. These results provide a practical route for using spatial early warning signals in heterogeneous networked systems when dense temporal monitoring or complete network-wide observation is infeasible, as is often the case in real applications.

• The paper introduces per-node baseline referencing and finds that additive variance, VΔ, reliably detects approaching tipping points across four stochastic models, 40 networks, and 10 simulation conditions.
• Baseline-referenced VΔ achieves sequential-test success rates of at least 0.85 in eight conditions and at least 0.60 in all conditions, while raw variance often fails under declining coupling.
• A single spatial snapshot outperforms up to 200 temporal samples, and monitoring about 20% of nodes preserves performance, making adjacency-free variance monitoring practical for data-limited systems.

Spatial early warning signals (EWSs) promise tipping-point anticipation from a single cross-sectional snapshot, but on heterogeneous networks their behavior has been shown to be inconsistent across dynamical systems and network structures. The paper by Bandara, Yu, and Masuda (2608.06608) addresses this problem with a simple transformation: before computing a spatial statistic, each node's state is referenced to its own baseline measured far from the transition. Across four stochastic dynamical systems, 40 networks, and ten simulation conditions, this baseline referencing markedly rehabilitates variance-based spatial EWSs, while leaving skewness-, kurtosis-, and autocorrelation-based signals unreliable.

Motivation and problem statement

Classical temporal EWSs such as variance and lagged autocorrelation exploit critical slowing down but require long time series under slowly varying environmental conditions. Spatial EWSs replace repeated temporal sampling with sampling across nodes at a single control-parameter value. Prior benchmarking work, including the authors' earlier study of spatial EWS applicability to complex network dynamics, established that no classical spatial EWS is universally reliable on heterogeneous networks: performance depends on the model, the control parameter, its direction of variation, and the specific network.

The paper first reproduces this failure using stochastic coupled double-well dynamics on networks, driven toward tipping either by decreasing coupling strength $D$ or by decreasing stress $u$. Under descending $D$, the raw spatial variance $V$ and coefficient of variation (CV) systematically decrease toward the tipping point—because reducing coupling shrinks the coupling-induced spread of equilibrium node states—so these signals fail outright in exactly the scenario where connectivity loss drives collapse. Sign-adjusted skewness $g_1'$, kurtosis $g_2$, and Moran's $I$ behave inconsistently across networks under both driving protocols. The diagnosis is that a snapshot $\{x_i^*\}$ is dominated by static, degree-heterogeneity-induced baseline differences that vary from network to network, contaminating any statistic computed directly on raw states.

Baseline-referenced variants

The proposed fix defines a per-node baseline $b_i$ as the mean of $x_i^*$ over the first $u$0 control-parameter values of the home range, far from the tipping point. Each spatial EWS is then computed either on differences $u$1 (the "difference" or additive variant) or on ratios $u$2 (the "ratio" or multiplicative variant). For example, the variance becomes $u$3 or $u$4. Both transformations remove the static network-specific component—additively or multiplicatively respectively—and retain only the dynamical broadening of the node-state distribution as the bifurcation approaches.

The evaluation spans ten simulation conditions: coupled double-well, mutualistic species-interaction, SIS epidemic, and gene-regulatory dynamics, crossed with control parameter ($u$5 or $u$6; SIS uses only $u$7) and direction (ascending/descending; mutualistic and gene-regulatory models use descending only), over 40 networks (34 empirical, 6 synthetic; sizes from $u$8 to $u$9).

Performance results

Three complementary performance measures are used: sign-adjusted Kendall's $D$0 between the EWS and the control parameter over the home range; a non-sequential sup-$D$1 test for significant positive steepening (structural change with unknown breakpoint); and a sequential, online version of the sup-$D$2 test calibrated by Monte Carlo for family-wise false-alarm probability of 0.05.

The main findings are:

• Baseline referencing helps only variance-based signals. $D$3, $D$4, and $D$5 show positive average $D$6 across all ten simulation conditions (with one near-zero exception for $D$7), whereas all original EWSs and the transformed $D$8, $D$9, and Moran's $V$0 remain unreliable, with sign-flipping $V$1 across conditions.
• Steepening is detectable online. In the sequential sup-$V$2 test, $V$3 achieves success fractions of at least 0.85 in eight of ten simulation conditions, with a minimum of 0.60 across all conditions. By contrast, the original variance detects steepening with success fraction $V$4 for all four dynamical systems under descending $V$5. Alarms typically fire with large lead times (fraction of home range remaining).
• A single snapshot beats 200 temporal samples. Comparing against a single-node temporal variance computed from up to $V$6 approximately independent samples, the temporal signal never reaches the performance of $V$7 within $V$8 in any of the ten simulation conditions, for any of the three measures. This is a strong claim: one spatial snapshot carries more early-warning information than a long single-node record, even though the comparison is conducted in a noisier regime ($V$9) than prior work.
• Sentinel-node subsampling is nearly free. Computing the EWS from roughly 20% of nodes retains almost the same performance as using all nodes, down from the full network, and no sentinel-selection rule (largest/smallest baseline, equidistant ranks, extreme CV, etc.) outperforms random selection. No interior optimum beats the full network, so more nodes never hurt—but monitoring a modest fraction suffices. Notably, $g_1'$0 requires no knowledge of the adjacency matrix at all.

The authors recommend $g_1'$1 as the final choice, noting it is more robust than $g_1'$2 because ratio-based variants become unstable when baselines are near zero—a common situation since three of the four models have $g_1'$3 as an equilibrium. Conversely, the multiplicative variant is invariant to node-wise unit changes, which matters when node observables are genuinely incomparable (e.g., species biomasses or heterogeneous questionnaire items); the paper recommends testing both variants empirically.

Limitations and open questions

The study is entirely computational, using stylized models; empirical validation is deferred. Networks are static, undirected, and unweighted, and extension to temporal, multilayer, directed, weighted, or higher-order structures remains open. The sup-$g_1'$4 framework was chosen deliberately over bifurcation-specific Bayesian model comparison because it is unclear whether a cross-sectional spatial statistic on a heterogeneous network obeys the same scaling laws (e.g., power-law divergence) as ensemble statistics near a bifurcation; establishing such theory for networked spatial EWSs is explicitly left open. Finally, whether predictive sentinel nodes can be identified systematically, beyond random selection, remains unresolved.

Conclusion

This paper converts a documented failure mode of spatial EWSs—confounding of dynamical trends with static network heterogeneity—into a controlled correction via per-node baseline referencing. The resulting additive baseline-referenced variance $g_1'$5 is consistent across ten tipping scenarios and 40 networks, exhibits statistically detectable acceleration well before the transition, dominates a single-node temporal variance even with 200 samples, and tolerates omission of up to 80% of nodes. Its independence from adjacency-matrix knowledge makes it a practical candidate for data-limited monitoring, pending empirical validation.