WorksMathematics
Early warning of critical transitions: Distinguishing supercritical and subcritical bifurcations from noisy observations
Abstract
Early-warning methods can indicate that a system is approaching a loss of stability, but they do not usually reveal whether the approaching bifurcation is supercritical or subcritical. We investigate whether this distinction can be made from noisy pre-transition observations. For this investigation, we extend a regression-based stability method by including the leading cubic nonlinear term, whose sign is used to predict the bifurcation type. We test the method on a stochastic pitchfork model and a spatial stochastic Swift-Hohenberg equation. In both models, classification improves with longer observation horizons, closer proximity to the bifurcation, stronger noise, and larger nonlinear effects. However, these conditions are also linked to a higher risk of noise-induced tipping in the subcritical case. The useful early-warning window is therefore determined by a balance between obtaining enough information to identify the bifurcation type and doing so before noise-induced tipping becomes too likely. In spatially extended systems, we find that observing the full spatial field allows classification over shorter observation horizons, and performance improves when a broader range of spatial modes contributes to the observed dynamics. Altogether, we identify the conditions under which the method can reliably distinguish between supercritical and subcritical bifurcations.
Many natural and engineered systems, from ice sheets and ocean currents to ecosystems and power grids, can change abruptly when slowly changing conditions push them past a critical threshold. Existing early-warning indicators can signal that a system is losing stability, but they cannot tell whether the coming change will be gradual or abrupt. This difference depends on whether the underlying bifurcation is supercritical, where a new state grows smoothly out of the old one, or subcritical, where the system can jump to a distant state and random fluctuations can trigger that jump before the threshold itself is reached. We show that noisy observations recorded before the transition can contain enough information to tell these two cases apart. Fitting a simple model that includes the leading nonlinear term to simulated stochastic data, we find that the distinction becomes reliable for longer records, closer to the transition, and for stronger noise or nonlinearity, but these conditions also raise the risk of an early noise-induced jump. For spatially extended systems, observing the full spatial field shortens the record needed, especially when fluctuations are spread over many spatial scales.
Introduction
An important motivation for studying bifurcations is the understanding of tipping points in complex systems. A tipping point is a critical threshold at which a system undergoes a qualitative transition to a substantially different dynamical regime [1, 2, 3]. Such transitions occur in many natural and engineered systems, including climate dynamics, ecology, financial markets, infrastructure networks, and social systems [1, 4, 5]. Examples include shifts in ocean circulation [3], tipping of the Greenland and Antarctic ice sheets [6], abrupt shifts in arid ecosystems [7], cascading failures in power grids [4], financial market crashes [8], and abrupt changes in social conventions or the spreading of information through social systems [5, 9]. In many systems, the main question is therefore not only when a transition occurs, but also whether it develops gradually or through an abrupt shift between states.
Specifically, in the case that the system undergoes a pitchfork bifurcation with normal form (1), the distinction between gradual and abrupt transitions is closely related to whether the underlying bifurcation is supercritical or subcritical; the qualitative difference is illustrated in Figure 1, where denotes the unknown variable and is a parameter in the system that encodes the distance to the bifurcation point. In a supercritical bifurcation, stable finite-amplitude states emerge continuously when the critical threshold is crossed. In a subcritical bifurcation, unstable finite-amplitude states already exist before the bifurcation point [10 Sec. 3.4]. These states form a threshold that can be crossed by sufficiently large perturbations. Stochastic fluctuations can trigger such abrupt shifts before the deterministic bifurcation point is reached, a process known as noise-induced tipping [11]. The distinction between supercritical and subcritical bifurcations is therefore important both for the nature of the transition and for the sensitivity of the system to stochastic perturbations.
Figure 1. Schematic comparison of noisy trajectories near supercritical and subcritical pitchfork bifurcations. Blue solid lines denote stable equilibrium branches, red dashed lines denote unstable branches, and the black curves illustrate noisy trajectories. In the supercritical case, stable finite-amplitude branches emerge after the loss of stability, whereas in the subcritical case unstable finite-amplitude branches are already present for .
The distinction between supercritical and subcritical bifurcations also applies to spatially extended systems. In these systems, instabilities can give rise to spatial patterns, as in a Turing bifurcation [12, 13, 14]. The dynamics close to such an instability are determined by a universal amplitude equation, the (real) Ginzburg–Landau equation [13], whose structure closely resembles the normal form of the pitchfork bifurcation. The formal derivation and rigorous justification of these models in pattern-forming systems is by now a well-established theory; see, for example, [15] for the deterministic case, and [16] for the stochastic case. Therefore, as in the pitchfork case depicted in Figure 1, the pattern emerges continuously from the homogeneous state for a supercritical bifurcation, and it typically appears through an abrupt finite-amplitude transition for a subcritical bifurcation.
Traditional early-warning indicators, such as those based on the theory of critical slowing down, identify changes in the linearized dynamics near a stable equilibrium [1, 17, 18, 19]. These changes can signal a loss of stability and thereby provide an early warning of an approaching bifurcation. However, these indicators do not distinguish between supercritical and subcritical bifurcations because this distinction is determined by the nonlinear dynamics. In fact, leading nonlinear terms can already influence a system’s behavior and predictability before the bifurcation is crossed [20], suggesting that pre-transition observations may contain information beyond the linearized dynamics. This raises the question under which observational conditions these nonlinear contributions can be recovered reliably, thus enabling an early-warning indicator that is capable of distinguishing imminent supercritical bifurcations from subcritical bifurcations.
In this work, we investigate whether noisy observations of a system approaching a bifurcation can be used to determine its criticality. We extend the regression method introduced in [21], which estimated the linearized dynamics, by also estimating the leading cubic drift term from temporal and spatio-temporal data (§ 2). We consider synthetic data generated from a stochastic variant of the normal form for a pitchfork bifurcation ((19)), and a stochastic Swift–Hohenberg equation (25). The latter is a standard model for a pattern-forming system close to a Turing instability, and was first introduced as a model for temperature-driven convection in fluids [22]. For both models, we study how the prediction depends on the distance to the bifurcation and the observation horizon (§ 3–§ 4). For the Swift–Hohenberg equation, we additionally investigate how the strength of the spatial operator and the critical spatial scale affect the prediction of the bifurcation criticality (§ 4). Finally, in § 5 we discuss the feasibility and limitations of applying the method to observational data from complex systems.
Distinguishing supercritical and subcritical bifurcations
To predict the bifurcation criticality from data, we approximate the local dynamics by a minimal normal form containing the linear term and the leading nonlinear contribution. For the bifurcations considered in this paper, the leading nonlinear term is cubic. For non-spatial models, we thus consider
where denotes the deviation from the equilibrium, describes its linear stability, and is the leading nonlinear coefficient. We use and throughout to emphasize their roles as the stability parameter and the coefficient determining the bifurcation criticality, respectively. For the equilibrium , indicates linear stability, and the bifurcation occurs as crosses zero. The sign of determines the bifurcation criticality: corresponds to a supercritical bifurcation and to a subcritical bifurcation, as illustrated in Figure 1.
Equation (1) is the normal form of a pitchfork bifurcation and describes the leading-order dynamics close to the bifurcation after reduction to the critical mode [23]. A similar cubic structure arises in spatially extended systems close to pattern-forming instabilities.
For spatially extended systems, we consider the cubic Swift–Hohenberg equation [22]
where , , , and , , and . At , the homogeneous equilibrium loses stability to spatial modes with critical wavenumber , resulting in a Turing instability. Close to this instability, the dynamics are dominated by these critical modes, whose amplitude varies slowly in space and time. At leading order, this slow modulation is described by a cubic Ginzburg–Landau-type amplitude equation [13, 15]. For a spatially uniform amplitude with fixed phase, the spatial derivative terms in the amplitude equation vanish, leaving a cubic ordinary differential equation with the same normal-form structure as Equation (1) [13]. A non-zero amplitude corresponds to a spatially periodic pattern in the original Swift–Hohenberg system, and the cubic coefficient in the amplitude equation is proportional to . The sign of therefore determines whether the patterned branch emerges supercritically or subcritically, just as in the non-spatial normal form. Since such amplitude equations arise generically close to stationary finite-wavenumber instabilities, the Swift–Hohenberg equation provides a representative test case for a broader class of pattern-forming systems [13].
Throughout this paper, we use bifurcation criticality classification to refer to distinguishing between supercritical and subcritical bifurcations based on the estimated cubic coefficient . The classification performance therefore measures how well the two bifurcation types can be separated from the fitted values of this coefficient , which we denote by .
We estimate these coefficients by writing the governing dynamics as a regression problem, similar to SINDy methods [24]
where contains the selected candidate terms and contains their coefficients. Note that, although the observations are noisy, we do not explicitly include a stochastic forcing in the regression model. Instead, we fit the deterministic drift of the dynamics. For zero-mean stochastic forcing, the noise contribution is expected to average out over sufficiently many observations or realizations and is therefore treated as a residual in the regression. This allows us to focus on the deterministic coefficients that determine the qualitative properties of the transition, while keeping the fitted model as simple as possible.
We first describe the method for ordinary differential equations and then extend it to spatially extended systems.
Method for ordinary differential equations
We assume that we have scalar observations of a system at times and seek to describe the dynamics using the scalar normal form in Equation (1).
The data are divided into observation windows that are analyzed separately. Within each window, the dynamics are assumed to be approximately stationary. The linear and cubic coefficients are estimated using linear regression, after which the stability and the predicted bifurcation criticality are determined.
Step 1: Prepare the data and define the observation windows
The fluctuations around the equilibrium are defined as
where is the equilibrium state. For the models considered in this paper, the equilibrium is known to be located at . For observational data with an unknown equilibrium, it can instead be estimated from the data.
From the full dataset, we define possibly overlapping observation windows,
where denotes the number of temporal observations in window . The number of observations may therefore differ between windows, and each window is analyzed independently.
Step 2: Construct the regression model
The minimal library contains the linear and cubic terms, i.e.
Within observation window , the time derivative is estimated using forward differences,
Step 3: Estimate the coefficients
For each observation window, the coefficients are estimated by minimizing
This gives one estimate of and for each observation window.
Step 4: Determine the bifurcation type and linear stability
The estimate is an indicator for the linear stability, similar to classic early-warning signs [1, 17, 18, 19], with corresponding to a stable equilibrium and values approaching zero from below indicating an approaching loss of stability. The estimated cubic coefficient is an indicator for the bifurcation criticality. In an ideal situation, a negative value, , indicates a supercritical bifurcation, and a positive value, , a subcritical bifurcation. However, due to neglected effects (e.g. stochasticity, temporal variations of forcing within a time window), the best criticality classification threshold might be located at a different value for .
Method for spatially extended systems
For spatially extended systems, we assume scalar observations at times and spatial locations . The formulation presented here considers a single spatial variable and one spatial dimension, but can be extended to multiple variables and higher spatial dimensions.
As in the ODE case, the data are divided into temporal observation windows that are analyzed separately, and the coefficients are assumed to remain approximately constant within each window. The method builds on the spatial stability framework introduced in [21]. In the present formulation, spatial derivatives are evaluated spectrally using Fourier transforms, while the fitted dynamics are extended to include the leading cubic nonlinear term. The cubic coefficient is used to predict the bifurcation type, while the linear spatial coefficients are used to construct the dispersion relation.
Step 1: Prepare the data and define the observation windows
The fluctuations around the homogeneous equilibrium are defined as
If the equilibrium is unknown, it can be estimated, e.g. from the spatial average,
We define possibly overlapping temporal observation windows,
where denotes the number of temporal observations in window . Each temporal observation in the window contains measurements at all available spatial locations.
Step 2: Estimate temporal and spatial derivatives
The temporal derivative is approximated from consecutive observations using a forward difference,
Spatial derivatives are evaluated spectrally. Denoting the spatial discretized Fourier transform by , derivatives of order are calculated according to
This avoids additional finite-difference approximation errors in the spatial derivatives, which is particularly useful when fitting dynamics from noisy observations.
Step 3: Construct and fit the spatial regression model
The candidate library contains the linear spatial terms expected in the governing equation together with the leading cubic nonlinear term. In one spatial dimension, it can be written as
where dots represent higher order derivatives when required; only the spatial derivative terms corresponding to the assumed governing equation are included in the regression.
The coefficients are estimated separately for each observation window using all spatial locations and time steps within that window. The least-squares estimate is given by
This gives one estimate of the linear spatial coefficients and the cubic coefficient for each observation window.
Step 4: Determine the bifurcation type and construct the dispersion relation
As before the estimated cubic coefficient is an indicator for the bifurcation criticality, with negative values suggesting supercritical bifurcations and positive values subcritical bifurcations.
For spatial systems, the linear stability can be determined via the dispersion relation [21]. For example, for the Swift–Hohenberg model (§ 4), the fitted linear operator is
which gives the estimated dispersion relation
The dominant wavenumber and corresponding growth rate are then obtained from
The value of describes the stability of the dominant spatial mode, while identifies the dominant wavenumber and hence the corresponding characteristic spatial scale.
Criticality of pitchfork bifurcation
We first apply and test the bifurcation criticality classification method to the stochastic pitchfork model. This provides a controlled setting in which the linear coefficient and the sign of the cubic term are known exactly. This section is structured as follows. First, we briefly describe the model. Second, we give details for the numerical experimental setup. Third, we give the results of these experiments.
Model
We consider the stochastic pitchfork normal form with additive noise, i.e.
where is a standard Wiener process. We introduce the rescaled state and time variables
to obtain the rescaled model
with
This scaling normalizes both the noise strength and the magnitude of the cubic coefficient, leaving and as the reduced dynamical parameters. The parameter gives the linear stability relative to the noise and nonlinear scales. The bifurcation remains located at . The sign determines the bifurcation type, with corresponding to the supercritical case and to the subcritical case.
For the subcritical model, the stable equilibrium is bounded by the unstable states for . A trajectory that crosses either threshold leaves the region in which the cubic normal form provides a bounded local description. Appendix § 9 derives the mean first passage time to these thresholds using the standard backward equation [25] and compares it with the Eyring–Kramers approximation [26, 27]. Based on these calculations, Figure 2 shows how the escape time depends on . This provides the basis for the treatment of escaped trajectories in the numerical experiments described in § 3.2.

Figure 2. Mean first passage time for the subcritical reduced pitchfork model in Equation (21). Starting from the stable equilibrium , the first passage time measures the average time required for a trajectory to reach either unstable threshold and leave the local stable region. The solid curve shows the mean first passage time obtained by numerically solving the backward boundary-value problem derived in Appendix § 9. The dashed curve shows the Eyring–Kramers approximation [26, 27], multiplied by an empirical factor to improve the comparison over the finite parameter range shown.
Experimental setup
We generate numerical data from the reduced stochastic pitchfork model in Equation (21) using the Euler–Maruyama method,
where are sampled independently at every time step. Each simulation starts at the stable equilibrium, i.e. .
To study how the fitting performance depends on the distance to the bifurcation point (), the bifurcation criticality (), and the available observation time horizon (), we vary these parameters as summarized in Table 1.
| Parameter | Description | Values |
| scaled linear stability | 50 log-spaced in | |
| cubic coefficient sign | ||
| observation horizon |
Table 1. Parameters varied in the stochastic pitchfork experiments. The fixed numerical settings are the time step , initial condition , unit noise strength in the reduced model, and independent realizations for each parameter combination. Simulations are performed using the Euler–Maruyama method.
For each combination of and , the different observation horizons are compared using the same underlying stochastic trajectories. A common ensemble of 100 trajectories is used for all included observation horizons, with the shorter datasets obtained by truncating the trajectories at the corresponding value of . This choice keeps the underlying noise realizations fixed, so that differences between observation horizons result from the amount of available data rather than from changes in the sampled stochastic trajectories.
In the subcritical case, some trajectories eventually leave the local regime described by the cubic normal form. Since the fitting procedure itself does not assume prior knowledge of the unstable thresholds , departures are identified using a numerical criterion: a trajectory is excluded from the analysis if non-finite values occur or if
For each parameter setting, an observation horizon is included in the analysis only if at least 10% of the trajectories generated from the first 100 random seeds are not excluded by this criterion before reaching that horizon. For each included observation horizon, a common ensemble of 100 trajectories is selected from a larger finite pool of simulated trajectories, excluding those that leave the local regime before the end of the observation window (with shorter observation horizons obtained by truncation). As approaches the bifurcation point at from below, satisfying this criterion becomes harder for long observation horizons, consistent with the decrease in the mean first passage time shown in Figure 2.
For each trajectory included in the analysis, the coefficients are estimated using the ODE fitting procedure described in § 2.1. This gives one pair of estimates for each realization and observation horizon.
The estimates are used to distinguish the two bifurcation types. The subcritical case is predicted when exceeds a classification threshold. To test this procedure and threshold, for each value of and observation horizon we consider an ROC classification ensemble that contains the 100 simulations for both and (as described above) and test how well the bifurcation criticality classification can distinguish the super- and subcritical cases. In particular, we vary the classification threshold to obtain receiver operating characteristic (ROC) curves [28, 29]. The corresponding area under the curve (AUC) measures how well the distributions of for the supercritical and subcritical ensembles can be separated. An AUC of 0.5 corresponds to chance-level discrimination between the two bifurcation types, while an AUC of 1 corresponds to complete separation.
Results
In this section we focus on the criticality classification; full details of the distributions of the fitted coefficients and for representative parameter settings are shown in Appendix § 10. In short, the linear coefficient is generally recovered well, with improved estimates for longer observation horizons and closer to the bifurcation point.
Figure 3 shows the AUC for bifurcation criticality classification as a function of the distance to the bifurcation, , and the observation horizon, . For a fixed observation horizon, the AUC generally increases as approaches zero, showing that the nonlinear coefficient becomes easier to distinguish closer to the bifurcation. At the same time, for a fixed value of , longer observation horizons lead to higher AUC values. Hence, a longer observation time allows the bifurcation criticality to be predicted farther away from the bifurcation point. Some curves, particularly for the longer observation horizons, terminate before reaching because too few subcritical trajectories remain within the local regime for the full observation horizon.

Figure 3. Area under the ROC curve (AUC) as a function of parameter for the different observation horizons . An AUC value of 0.5 indicates the classification method is as good as flipping a coin; an AUC value of 1 indicates potential for a perfect classification. indicates closeness to the bifurcation point located at . A value of is shown only if at least 10% of the subcritical realizations remain within the local regime for the full observation horizon (§ 3.2); curves for longer horizons therefore end farther from the bifurcation. Some corresponding ROC-curves are shown in Figure 4.
The differences between the observation horizons are substantial. For the longest observation horizon (), the AUC already exceeds approximately 0.6 at and approaches perfect classification around . This means that accurate prediction is possible well before where too few trajectories remain within the local regime for the analysis, and noise-induced tipping is very probable. For intermediate observation horizons (–), perfect classification (AUC = 1) can also be approached, but only relatively close to the onset of frequent noise-induced tipping. For shorter horizons (–), the criticality classification becomes weaker, and deteriorates further for even shorter observation windows, where the AUC stays below approximately 0.7 over most of the parameter range. For the shortest observation horizon (), the AUC even falls below 0.5, to approximately 0.42, for . At this horizon the fitted is strongly negative for both bifurcation types (Appendix § 10), which indicates that the bifurcation criticality classification is unreliable when used with too little data.
This dependence on the observation horizon is also visible in the corresponding ROC curves in Figure 4, shown for two representative observation horizons ( and ) and a selection of values. For , most curves remain relatively close to the diagonal over much of the parameter range. For , a clearer separation develops as approaches zero. The parameter distributions for shown in Appendix § 10 give the same picture: the distributions of overlap strongly when little information is available and become increasingly separated as the observation horizon increases or approaches zero, which makes it possible to choose the classification threshold to separate sub- and supercritical cases.

Figure 4. Receiver operating characteristic (ROC) curves for the stochastic pitchfork model in Equation (21) at observation horizons (left) and (right). Each curve corresponds to a different value of and is constructed from the fitted values of . The horizontal axis gives the false-positive rate and the vertical axis the true-positive rate, with the subcritical case () taken as the positive class. The diagonal line denotes chance-level discrimination (i.e. flipping a coin).
Criticality of bifurcation in the Swift–Hohenberg equation
Next, we apply and test the bifurcation criticality classification method to the stochastic Swift–Hohenberg model. In contrast to the stochastic pitchfork model, each temporal observation now contains the complete spatial field. This allows us to investigate how the additional spatial information affects the estimation of the nonlinear coefficient and the predicted bifurcation criticality.
Model
We consider the stochastic Swift–Hohenberg equation [22]
on a bounded, periodic domain, where denotes the spatially varying system state. Here, determines the linear stability of the homogeneous equilibrium, is the leading cubic nonlinear coefficient, and determines the noise strength. The additional spatial parameters are , which sets the strength of the spatial operator, and , the critical wavenumber at which the linear growth rate is maximal. Here, denotes the spatial Laplacian in one dimension, and is a space–time Wiener process.
We first rescale space by the critical physical wavenumber,
The space–time white noise also changes under this spatial transformation. In one spatial dimension, the transformation introduces an additional factor in the noise amplitude [16 Chapter 1]. Accounting for this additional spatial scaling of the noise, we choose the amplitude and time scales as
This choice normalizes both the magnitude of the cubic coefficient and the strength of the space–time white noise. The rescaled Swift–Hohenberg model then becomes
with
Here, denotes the correspondingly rescaled space–time Wiener process. The parameters and have the same interpretation as in the stochastic pitchfork model: describes the linear stability and determines the bifurcation criticality. The additional parameter describes the strength of the spatial operator relative to the noise and nonlinear time scales.
The effect of can be seen from the dispersion relation of the linearized reduced model,
Through the spatial scaling , the physical critical wavenumber is mapped to the rescaled wavenumber , where . Hence, for fixed , changing does not change the stability of the critical mode. Instead, it determines how strongly the modes around are damped. Larger gives a narrower band of weakly damped modes, while smaller allows a broader range of spatial modes to fluctuate, as illustrated in Figure C.16.
Expanding the spatial operator in Equation (28) gives
The corresponding coefficients of the linear spatial terms are therefore
while the cubic coefficient remains . These coefficients provide the theoretical values against which the estimates from the regression are compared.
Experimental setup
We generate numerical data from the reduced Swift–Hohenberg model in Equation (28). The simulations are performed on a one-dimensional periodic domain in the reduced spatial coordinate , the spatial derivatives are evaluated spectrally, and the system is integrated in time using the Euler–Maruyama method. Unit space–time white noise is discretized such that the stochastic increment at each space–time grid point is , where are independent.
We perform two main experiments. First, we vary the observation horizon at fixed to study how the available observation time affects the ability to distinguish the two bifurcation types. Second, we vary at fixed to study how the spatial dynamics affect this distinction. The parameter values and numerical settings used in the two experiments are summarized in Table 2. In addition, Appendix § 12 investigates the spatial effect associated with changes in the unscaled critical wavenumber . This is done by varying the reduced domain length and grid spacing according to the transformation , while keeping the reduced dynamical parameters and the number of spatial observations fixed.
| Parameter | Description | Varying | Varying |
| scaled stability | 20 log-spaced in | 20 log-spaced in | |
| cubic coefficient sign | |||
| spatial strength | 1 | ||
| observation horizon | 2 |
Table 2. Parameters used in the two stochastic Swift–Hohenberg experiments. Both experiments use a one-dimensional periodic domain of length with grid spacing , initial condition , unit spatially and temporally white noise, and independent realizations for each parameter combination. Spatial derivatives are evaluated spectrally and the simulations are integrated using the Euler–Maruyama method. The observation-horizon experiment uses , while the -sweep experiment uses to retain numerical stability for the largest values of . The value is included as a test case without the spatial operator, for which Eq. (28) reduces to uncoupled pitchfork dynamics at each spatial location, driven by independent noise, i.e., a spatial version of the pitchfork model.
For the observation-horizon experiment, the different observation horizons are compared using the same underlying stochastic realizations. Each realization is simulated up to the largest required value of , with the shorter datasets obtained by truncating the realization at the corresponding observation horizon, as before (§ 3.2).
In the subcritical case, some realizations eventually leave the local stable regime described by the cubic Swift–Hohenberg model. A realization is excluded from the analysis if the simulated field becomes non-finite or if
As in the pitchfork case (§ 3.2), an observation horizon is included in the analysis only if at least 10% of the realizations generated from the first 100 random seeds are not excluded by this criterion before reaching that horizon. For each included observation horizon, an ensemble of 100 realizations is selected from a larger finite pool, excluding those that leave the local regime before the end of the observation window.
In the second experiment, we vary to investigate the role of the spatial dynamics in the fitting procedure. Since controls the range of spatial modes that contribute strongly to the fluctuations, this experiment tests how the available spatial information affects the recovery of the nonlinear coefficient and the predicted bifurcation type.
Each realization is analyzed using the spatial fitting procedure described in § 2.6. For each observation window , the spatio-temporal data are used to estimate
The linear coefficients define the estimated dispersion relation
while is used to predict the bifurcation type. The ROC curves and AUC values are constructed from the distributions of in the same way as in § 3.2. Since each fitted dataset is considered separately in the following results, we omit the observation-window index from the fitted quantities for notational simplicity.
Results
The distributions of the fitted coefficients for representative parameter settings are shown in Appendix § 11. As in the pitchfork model, the linear part of the dynamics is generally recovered well, with improved estimates for longer observation horizons and closer to the bifurcation point. Since the linear spatial dynamics are described by multiple fitted coefficients, we additionally show a representative fitted dispersion relation in Figure 5. The fitted dispersion relation closely follows the theoretical curve and reproduces both the dominant wavenumber and its corresponding growth rate . Additional fitted dispersion relations for varying , together with the errors in the estimated maxima, are shown in Appendix § 11.§ 1, indicating that the results of [21] remain accurate when the regression is extended to include the cubic term. We therefore focus below on the additional question of when the fitted nonlinear coefficient contains sufficient information to predict the bifurcation criticality, which we assess from the distribution of using ROC curves and AUC values.

Figure 5. Representative fitted dispersion relations for the stochastic Swift–Hohenberg model at and observation horizon . Each panel corresponds to a fixed combination of and . The black curve shows the theoretical dispersion relation, the blue curve the ensemble mean of the fitted relations, and the dashed curves the 5th and 95th percentiles. Blue points denote the fitted maxima of the individual realizations, while the yellow star denotes the theoretical maximum . The text box reports the number of available realizations and summary statistics for and , following [21].
Effect of observation horizon
Figure 6 shows the AUC for bifurcation criticality classification as a function of the distance to the bifurcation, , and the observation horizon, , at fixed . For a fixed observation horizon, the AUC generally increases as approaches zero, showing that the nonlinear coefficient becomes easier to distinguish closer to the bifurcation. At the same time, for a fixed value of , longer observation horizons lead to higher AUC values. Hence, as in the stochastic pitchfork model, a longer observation time allows the bifurcation type to be predicted farther away from the bifurcation point.

Figure 6. Area under the ROC curve (AUC) as a function of for different observation horizons at fixed . Each curve corresponds to one observation horizon. The AUC and class conventions are as in Figure 3.
This dependence on the observation horizon is also visible in the corresponding ROC curves in Figure 7, shown for two representative observation horizons ( and ). The longer observation horizon gives a clearer separation between the supercritical and subcritical estimates over a larger range of . The parameter distributions shown in Appendix § 11 give the same picture: the distributions of become increasingly separated as the observation horizon increases or approaches zero.

Figure 7. Representative ROC curves for the reduced stochastic Swift–Hohenberg model at (left) and (right), with . Each curve corresponds to a different value of and is constructed from the fitted values of . The axis and class conventions are as in Figure 4.
Compared with the non-spatial pitchfork case, high AUC values are obtained over much shorter scaled observation horizons. Each temporal observation contains the complete spatial field, so the regression uses variation in both space and time rather than a single scalar observation at each time step. The additional spatial information therefore allows the nonlinear dynamics to be constrained over shorter observation horizons.
This improvement at shorter observation horizons is accompanied by a smaller range of over which subcritical realizations can be included for some of these horizons. Comparing Figure 3 and Figure 6, we find that for the shared shorter observation horizons and , the Swift–Hohenberg curves terminate farther from the bifurcation, while for both models can be analyzed up to . Thus, although accurate classification can be achieved over shorter observation times, noise-induced tipping also seems more likely. However, for and , the termination points occur at comparable values of in both models.
Effect of the spatial parameter
Figure 8 shows the AUC for bifurcation criticality classification for different values of at fixed . The two criticality types are most clearly distinguished for small , while the classification performance generally decreases as increases. At the same time, the curves for larger extend closer to the bifurcation point. Stronger damping of the modes away from the critical wavenumber reduces the fluctuations, so that fewer subcritical realizations leave the local regime within the observation horizon.

Figure 8. Area under the ROC curve (AUC) as a function of for different values of at fixed observation horizon . Each curve corresponds to one value of . The AUC and class conventions are as in Figure 3. The , curve differs slightly from the corresponding curve in the observation-horizon experiment because the two experiments use separate simulation ensembles.
The representative ROC curves in Figure 9 illustrate the same change in classification performance ( and ). For , the supercritical and subcritical estimates can be separated over a larger part of the stable parameter range, whereas for the ROC curves remain closer to the diagonal. The corresponding parameter distributions are shown in Appendix § 11. To relate this change in classification performance to the spatial dynamics, we also consider how the dispersion relation and the observed spectrum change with .

Figure 9. Representative ROC curves for (left) and (right) at fixed observation horizon . Each curve corresponds to a different value of and is constructed from the fitted values of . The axis and class conventions are as in Figure 4.
The fitted dispersion relations in Appendix § 11.§ 1 show that increasing progressively narrows the band around the critical wavenumber . Consistently, the representative spectra in Figure C.16 show that the observed fluctuations become concentrated over a narrower range of wavenumbers and that the effective number of contributing modes decreases. This suggests that the change in classification performance is related to the range of spatial modes contributing to the observed dynamics, which we discuss further in § 5.2.
Discussion
The results in § 3.3 and § 4.3 show that in some cases the bifurcation criticality can be distinguished from noisy observations of a system before it crosses the associated bifurcation point or before noise-induced tipping. However, the classification performance depends strongly on the observation horizon, the distance to the bifurcation, and, in the spatial model, the distribution of fluctuations across spatial modes. At the same time, the linear stability properties (i.e., the closeness to the bifurcation point) remain generally well recovered with the extended regression method introduced here. We first discuss the effects of the observation horizon and the distance to the bifurcation, before considering the additional effects introduced by spatial observations.
Required observation time
Bifurcation criticality classification was better for longer observation times and for closer to the bifurcation point (Figure 3 and Figure 6); this was also observed in the direct parameter estimates (see Appendices § 10 and § 11). Longer observation times provide more information for estimating the (cubic) coefficient, while closer to the bifurcation the weaker linear restoring dynamics allow the nonlinear contribution to become more pronounced in the observed fluctuations. Longer observation horizons therefore not only improve the classification close to the bifurcation, but also allow reliable classification substantially farther away from the transition. In contrast, for very short observation times the two bifurcation criticality types may not be distinguishable at all; for example, the shortest observation horizons in Figure 6 give AUC values close to 0.5 over most of the investigated range.
This distinction is particularly relevant when considering the method as an early-warning indicator. In the subcritical case, near-complete separation is generally reached only close to the bifurcation, where stochastic fluctuations also make noise-induced tipping more likely and limit how long trajectories can be observed. This trade-off is consistent with the first-passage-time analysis in Figure 2 and Appendix § 9, where the expected escape time decreases rapidly as approaches zero. Longer observation records are therefore valuable because they allow reliable classification at larger , where noise-induced tipping is less likely.
In the experiments we have used rescaled parameters. It is therefore informative to understand these results in the context of the unscaled parameters. For the pitchfork model,
and for the Swift–Hohenberg model
For fixed , both models have the same qualitative dependence on observation time, noise strength, nonlinearity, and distance from the bifurcation. In particular, increasing naturally increases the available scaled observation horizon. Increasing either or simultaneously increases the scaled observation horizon and decreases . Both changes correspond to regimes in which the bifurcation type is more easily distinguished in our numerical results. This is consistent with the increase in fluctuation amplitude under stronger stochastic forcing or weaker restoring dynamics near a critical transition [30, 31]. Larger fluctuations increase the relative contribution of the nonlinear term, making it easier to identify in the model fit. However, in the subcritical case, these fluctuations also increase the possibility of noise-induced tipping, thus resulting in the trade-off between identifiability and escape discussed above.
Using spatial information
In the spatial Swift–Hohenberg equation, each observation contained data across a spatial field. Although these observations are spatially correlated, they provide variation across space in addition to that over time, which can help constrain the linear and nonlinear coefficients in the regression. Comparing the non-spatial pitchfork model (Figure 3) and the spatial Swift–Hohenberg equation (Figure 6), this is reflected in the Swift–Hohenberg model reaching high classification performance at much shorter observation horizons. However, the spatial setting also appears more prone to noise-induced tipping than the non-spatial model, particularly close to the bifurcation point. One possible reason is that farther from the bifurcation, stronger restoring dynamics make sufficiently large fluctuations increasingly unlikely, so the presence of many possible escape locations has only a limited effect. This is consistent with the first-passage-time analysis in Figure 2, where the expected escape time eventually increases exponentially with . Closer to the bifurcation, the restoring dynamics weaken and the expected escape time decreases rapidly, making it more likely that a sufficiently large fluctuation somewhere in the spatial domain initiates departure from the local regime.
The strength of the spatial effects also influences the classification performance. Figure 8 shows that the bifurcation types become more difficult to distinguish as increases. Increasing narrows the band of weakly damped modes around the critical wavenumber. For stochastic Swift–Hohenberg dynamics, the variance of individual Fourier modes is directly related to their linear eigenvalues [32 Sec. 3.2]. Increasing therefore concentrates the observed fluctuations into a smaller range of spatial modes, leaving less spatial variation available to distinguish the different terms in the fitted dynamics. This provides a possible explanation for the reduced ability to identify the nonlinear coefficient and hence the bifurcation type.
These spatial effects can be related back to the unscaled parameters as
We see that for fixed , an increase in either the noise strength or the magnitude of the nonlinear coefficient leads to a decrease of , an increase of , and a decrease of . All three changes move the system towards regimes in which the bifurcation type was easier to distinguish. In the spatial model, stronger noise or nonlinearity therefore not only makes the nonlinear dynamics more pronounced, but also increases the range of weakly damped spatial modes that can contribute to distinguishing the linear and nonlinear dynamics.
The critical physical wavenumber introduces an additional spatial dependence that cannot be determined from the separate effects of , , and alone. We therefore investigated the direct effect of increasing separately in Appendix § 12. The results show that increasing improves the bifurcation criticality classification. A larger corresponds to a shorter critical wavelength, so that more critical wavelengths are contained within a fixed physical observation domain. Equivalently, more spatial modes lie close to the critical wavenumber and exhibit substantial variance, as also observed for the stochastic Swift–Hohenberg equation on larger domains [32 Sec. 4.1]. This additional contribution from the other modes may help constrain the fitted dynamics, and thereby improve estimation of .
Applicability to observational data and limitations
The results of this study were obtained for controlled stochastic models, for which the relevant linear and nonlinear terms are known and the data can be observed at a prescribed temporal and spatial resolution. For real systems, the underlying model structure will be less certain. In this case, the regression may need to include a larger library of candidate terms. Sparse model-discovery methods, such as SINDy and its extensions, provide one approach for identifying which terms are supported by the observations [33, 34]. However, in all cases, the relevant dynamics must be represented in the candidate library and sufficiently constrained by the available data. Based on the findings in this paper, the main prerequisite for the use of a bifurcation criticality classification method is the availability of observations that are long enough and contain sufficiently strong fluctuations to identify the nonlinear dynamics before a transition occurs.
Conclusion
In this paper, we have introduced a bifurcation criticality classification method to distinguish imminent subcritical bifurcations from supercritical bifurcations using noisy observations. This extends linear stability estimates, such as those based on critical slowing down metrics [1, 17, 18, 19], by including the leading cubic nonlinear term in a regression framework. We further tested this method on a non-spatial pitchfork bifurcation and a spatial Swift–Hohenberg equation, to identify the situations in which it seems possible to determine a bifurcation’s criticality. Our results indicate that bifurcation criticality classification is indeed possible in certain situations. In particular, the prediction improves with longer observation times, when the system is closer to the bifurcation, and in systems with stronger noise or more dominant nonlinear effects. However, in the subcritical case, stronger noise or nonlinearity also increases the risk of an earlier noise-induced transition, which limits the window of opportunity for criticality classification. Further, the inclusion of spatial information can lead to improved predictions, thus decreasing the observation time needed for good criticality prediction, although the precise usefulness depends on the range of spatial modes contributing to the observations, with broader spectral fluctuations providing better separation of the fitted nonlinear dynamics.
These results show that noisy observations of a system before a bifurcation do contain information not only about an approaching instability, but also about the nonlinear character of the transition that follows. Combining estimates of the linear and nonlinear dynamics therefore provides a route towards early-warning systems that can predict both the approach to a transition and its expected criticality. Extending this framework to other bifurcation mechanisms, more complex systems, and observational data provides a natural next step towards predicting the character of critical transitions in real systems.
Acknowledgments
B.H. was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project number 543917644. P.A.S. thanks the Centre for Complex Systems Studies (CCSS) at Utrecht University for supporting a research visit to the Technical University of Munich through the Swaantje Mondt Fund.
Author Declarations
Conflict of Interest
The authors have no conflicts to disclose.
Use of generative AI
During the preparation of this work, the authors used ChatGPT (OpenAI), Claude (Anthropic) and Gemini (Google) to assist with writing and improving MATLAB code for the simulations and analysis, to brainstorm ideas, and to improve the clarity of the text and the presentation of the results. All AI-assisted code was checked and tested by the authors, and all text and ideas were critically reviewed and revised by the authors, who take full responsibility for the content of this publication. No generative AI was used to create or alter research data or figures.
Author Contributions
Paul A. Sanders: Conceptualization (lead); Methodology (lead); Software (lead); Formal analysis (lead); Investigation (lead); Visualization (lead); Writing – original draft (lead); Writing – review & editing (equal). Christian Kuehn: Conceptualization (supporting); Methodology (supporting); Supervision (equal); Writing – review & editing (supporting). Bastian Hilder: Conceptualization (supporting); Methodology (supporting); Supervision (equal); Writing – review & editing (supporting). Robbin Bastiaansen: Conceptualization (supporting); Methodology (supporting); Formal analysis (supporting); Supervision (equal); Writing – review & editing (equal).
Data Availability
The MATLAB code used to generate and analyze the data in this study, together with the processed results and the figures of the paper runs, is openly available on GitHub at https://github.com/Paul-A-Sanders/Criticality_classification. The raw simulated trajectories are not stored but can be regenerated with this code from the recorded random seeds.
Mean first passage time for the subcritical pitchfork model
This appendix quantifies the time scale on which trajectories of the subcritical pitchfork model can leave the neighborhood of the stable equilibrium before the deterministic bifurcation is reached. This provides a reference for the noise-induced tipping discussed in § 3.2 and Figure 2. We follow the standard backward-equation approach for first-passage problems [25] and compare the resulting mean first passage time with the Eyring–Kramers approximation [26, 27].
For the subcritical case , the reduced pitchfork equation in Equation (21) has drift
For , the equilibrium is stable and is bounded by the two unstable equilibria . For an initial condition with , we define the first exit time from this interval as
The mean first passage time from the starting point is
The infinitesimal generator of the reduced stochastic process is
The mean first passage time satisfies the standard backward equation [25]
The boundary conditions represent absorption once either unstable equilibrium is reached. Since is a function of the initial state rather than time, Equation (A.5) is a boundary-value problem and does not require a separate initial condition. By symmetry, we can also change one of the boundary conditions to .
Writing with
and applying an integrating factor to solve the backward equation (A.5) gives the exact integral representation from our equilibrium at
For , the Eyring–Kramers approximation gives [26, 27]

Figure A.1. (a) Mean first passage time and quantile intervals obtained from the two-moment Gamma approximation, shown together with the Eyring–Kramers scaling. (b) Ratios between the third and fourth moments of the exit time and the corresponding moments of the fitted Gamma distribution; the reference value corresponds to equal moments.
The comparison in Figure 2 uses with . This empirical factor improves the prefactor over the finite parameter range shown in the figure.
Gamma approximation of the exit-time distribution
The moments satisfy the backward hierarchy [25]
with . We approximate the exit-time distribution by a Gamma distribution, choosing its shape and scale parameters to match the mean and variance of the exit time at . This follows the standard moment-matching procedure for the Gamma distribution [35]. Similar Gamma approximations have been used for stochastic exit-time distributions [36].
The corresponding quantiles of the exit-time distribution are then approximated by
where is the inverse cumulative distribution function of the fitted Gamma distribution. To test this approximation, we compare the higher moments with the Gamma moments and its ratio
Values of close to one indicate that the two-moment Gamma approximation also describes the higher moments. In Figure A.1 we show the ratios and .
Parameter distributions for the pitchfork model
To provide additional insight into the fitting procedure and the ROC analysis presented in § 3.3, we show the distributions of the estimated linear and nonlinear coefficients obtained from the stochastic simulations. The distributions are visualized using kernel density estimates computed with the MATLAB ksdensity function using a Gaussian kernel. The estimates of and are shown separately to make the accuracy and spread of both fitted coefficients easier to assess.
The distributions of for representative observation horizons are shown in Figures B.1–B.4. In general, the linear coefficient is recovered more accurately for longer observation horizons and closer to the bifurcation point.
The distributions of , shown in Figures B.5–B.8, show how well the two bifurcation types can be distinguished directly from the fitted nonlinear coefficient. The true values and correspond to the supercritical and subcritical models, respectively. Their separation therefore determines the ROC and AUC performance discussed in § 3.

Figure B.1. Distributions of the estimated linear coefficient for an observation horizon . Each panel corresponds to a fixed combination of the true parameters and , based on 100 independent stochastic realizations when sufficiently many realizations can be included in the analysis. The filled curves show the estimated probability densities, with the fill colors distinguishing the supercritical () and subcritical () models. The black line outlines each density and the vertical dashed line marks the sample mean of the fitted estimates. The text box gives the number of available realizations, the sample mean, and the sample standard deviation. Parameter combinations for which too few realizations can be included in the analysis are left empty.

Figure B.2. Distributions of the estimated linear coefficient for . Plotting conventions are as in Figure B.1.

Figure B.3. Distributions of the estimated linear coefficient for . Plotting conventions are as in Figure B.1.

Figure B.4. Distributions of the estimated linear coefficient for . Plotting conventions are as in Figure B.1.

Figure B.5. Distributions of the estimated nonlinear coefficient for . The true values and correspond to the supercritical and subcritical models, respectively. Plotting conventions are as in Figure B.1.

Figure B.6. Distributions of the estimated nonlinear coefficient for . Plotting conventions are as in Figure B.1.

Figure B.7. Distributions of the estimated nonlinear coefficient for . Plotting conventions are as in Figure B.1.

Figure B.8. Distributions of the estimated nonlinear coefficient for . Plotting conventions are as in Figure B.1.
Validation of the Swift–Hohenberg fits
To provide additional insight into the fitting procedure and the ROC analyses presented in § 4, we first assess the recovery of the complete linear dispersion relation and subsequently examine the fitted linear and nonlinear coefficients separately. For the varying- experiment, we additionally show how the spatial spectrum changes as the linear spatial filtering becomes stronger. Together, these results provide additional context for both the accuracy of the recovered linear stability and the ability to distinguish the supercritical and subcritical models.
Fitted dispersion relations
We first assess the combined recovery of the linear spatial dynamics by comparing the fitted dispersion relations with the theoretical relation in Equation (30). Representative results are shown for , , and at the fixed observation horizon (Figures C.1–C.3). These cases span the transition from a broad, relatively flat maximum to a narrow and steep maximum around the critical mode .
For each realization, the fitted coefficients define through Equation (35). The errors in the estimated maximum are summarized by
with and . Across the three values of , the mean fitted dispersion relations remain close to the theoretical curves, showing that the linear spatial dynamics are generally recovered well alongside the nonlinear coefficient.
The two error measures reflect different aspects of this reconstruction. For small , the dispersion relation is relatively flat near its maximum, so uncertainty in the location produces only a small change in the corresponding growth rate. As increases, the maximum becomes increasingly sharp, constraining more tightly and reducing the spread in . At the same time, deviations in the fitted dispersion curve can produce larger errors in . Hence, improved localization of the dominant wavenumber does not necessarily imply a more accurate estimate of its growth rate.
Fitting of the linear coefficient
We next examine the fitted coefficients individually. For the linear fit, the displayed coefficient is , whose theoretical value is
Varying the observation horizon
Figures C.4–C.7 show the distributions of the fitted linear coefficient for varying observation horizons at fixed , for which the theoretical value is .
Varying the spatial parameter
Figures C.8–C.11 show the corresponding distributions when is varied at fixed . In this case the theoretical linear coefficient changes with both and according to .

Figure C.1. Fitted dispersion relations for and . Each panel corresponds to a fixed combination of and . Plotting conventions are as in Figure 5; the text boxes report the number of available realizations and the mean and standard deviation of and .

Figure C.2. Fitted dispersion relations for and . Plotting conventions and error definitions are as in Figure C.1.

Figure C.3. Fitted dispersion relations for and . Plotting conventions and error definitions are as in Figure C.1.

Figure C.4. Distributions of the estimated linear coefficient for and , for which the theoretical value is . The panel layout and plotting conventions are as in Figure B.1.

Figure C.5. Distributions of the estimated linear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.6. Distributions of the estimated linear coefficient for and . Plotting conventions are as in Figure C.4.
Fitting of the nonlinear coefficient
The nonlinear coefficient determines the predicted bifurcation type, with corresponding to the supercritical model and to the subcritical model. The overlap between the fitted distributions of therefore determines how well the supercritical and subcritical models can be distinguished, as quantified by the ROC and AUC analyses in § 4.
Varying the observation horizon
Figures C.12–C.15 show the distributions of for representative observation horizons at fixed . With increasing observation horizon, the fitted distributions generally become narrower and the supercritical and subcritical cases become more clearly separated, consistent with the improved classification performance reported in § 4.3.

Figure C.7. Distributions of the estimated linear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.8. Distributions of the estimated linear coefficient for and , with theoretical value . Plotting conventions are as in Figure C.4.
Varying the spatial parameter
Figure C.16 shows representative power spectra over a wider illustrative range of . Increasing concentrates the fluctuations more strongly around the critical wavenumber and decreases the effective number of substantially contributing modes. The same spectral narrowing is visible for both bifurcation criticality types. Its possible implications for the identifiability of the cubic term are discussed in § 5.2. The corresponding distributions of are shown in Figures C.17–C.20.

Figure C.9. Distributions of the estimated linear coefficient for and , with theoretical value . Plotting conventions are as in Figure C.4.

Figure C.10. Distributions of the estimated linear coefficient for and , with theoretical value . Plotting conventions are as in Figure C.4.

Figure C.11. Distributions of the estimated linear coefficient for and , with theoretical value . Plotting conventions are as in Figure C.4.

Figure C.12. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.13. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.14. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.15. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.16. Time-averaged one-sided spatial power spectra for the reduced stochastic Swift–Hohenberg model at and . Columns correspond to , , and , and rows to the supercritical () and subcritical () models. Each spectrum is normalized by its maximum value for visualization, and the vertical dashed line marks the critical wavenumber . The displayed effective spectral mode count is , where is the fraction of the total spectral power in mode . The first 25% of each simulation is discarded before the time average is computed.

Figure C.17. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.18. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.19. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.

Figure C.20. Distributions of the estimated nonlinear coefficient for and . Plotting conventions are as in Figure C.4.
Effect of the critical spatial scale
The physical critical wavenumber enters the reduced spatial coordinate through
Consequently, for a fixed physical observation domain of length and a fixed physical spatial grid, increasing increases both the reduced domain length and the reduced grid spacing ,
while the number of spatial observations remains unchanged. Since the critical physical wavelength is , increasing also means that more critical wavelengths are contained within the same physical observation domain. To isolate this direct spatial effect from the simultaneous dependence of , , and on , we vary and proportionally while keeping the reduced dynamical parameters and the number of spatial observations fixed.
We use , , , and independent realizations for each parameter combination. The parameter is varied over 20 logarithmically spaced values in , with , and the numerical time step is . The reduced domain length is varied as
Increasing and in this way corresponds directly to the spatial change produced by increasing for a fixed physical observation setup.
Figure D.1 shows the resulting AUC as a function of . Over the investigated range, the AUC generally increases with , showing that the direct spatial effect associated with increasing improves the distinction between the supercritical and subcritical dynamics.

Figure D.1. AUC for bifurcation criticality classification as a function of for different spatial scalings corresponding to increasing . The number of spatial observations is fixed at , with , while and are fixed. The AUC and class conventions are as in Figure 3.
References
- [1]Marten Scheffer, Jordi Bascompte, William A. Brock, Victor Brovkin, Stephen R. Carpenter, Vasilis Dakos, Hermann Held, Egbert H. van Nes, Max Rietkerk, and George Sugihara. Early-warning signals for critical transitions. Nature, 461:53–59, 2009. doi:10.1038/nature08227.DOI
- [2]Egbert H. van Nes, Babak MS Arani, Arie Staal, Bregje van der Bolt, Bernardo M Flores, Sebastian Bathiany, and Marten Scheffer. What do you mean, ‘tipping point’? Trends in Ecology & Evolution, 31(12):902–904, 2016. doi:10.1016/j.tree.2016.09.011.DOI
- [3]Timothy M. Lenton, Hermann Held, Elmar Kriegler, Jim W. Hall, Wolfgang Lucht, Stefan Rahmstorf, and Hans Joachim Schellnhuber. Tipping elements in the Earth’s climate system. Proceedings of the National Academy of Sciences, 105(6):1786–1793, 2008. doi:10.1073/pnas.0705414105.DOI
- [4]Oriol Artime, Marco Grassia, Manlio De Domenico, James P. Gleeson, Hernán A. Makse, Giuseppe Mangioni, Matjaž Perc, and Filippo Radicchi. Robustness and resilience of complex networks. Nature Reviews Physics, 6:114–131, 2024. doi:10.1038/s42254-023-00676-y.DOI
- [5]Alessandro Vespignani. Modelling dynamical processes in complex socio-technical systems. Nature Physics, 8:32–39, 2012. doi:10.1038/nphys2160.DOI
- [6]Frank Pattyn, Catherine Ritz, Edward Hanna, Xylar Asay-Davis, Rob DeConto, Gaël Durand, Lionel Favier, Xavier Fettweis, Heiko Goelzer, Nicholas R. Golledge, Peter Kuipers Munneke, Jan T. M. Lenaerts, Sophie Nowicki, Antony J. Payne, Alexander Robinson, Hélène Seroussi, Luke D. Trusel, and Michiel van den Broeke. The Greenland and Antarctic ice sheets under 1.5 °C global warming. Nature Climate Change, 8:1053–1061, 2018. doi:10.1038/s41558-018-0305-8.
- [7]Marten Scheffer, Steve Carpenter, Jonathan A. Foley, Carl Folke, and Brian Walker. Catastrophic shifts in ecosystems. Nature, 413:591–596, 2001. doi:10.1038/35098000.DOI
- [8]Didier Sornette. Critical market crashes. Physics Reports, 378(1):1–98, 2003. doi:10.1016/S0370-1573(02)00634-8.DOI
- [9]Damon Centola, Joshua Becker, Devon Brackbill, and Andrea Baronchelli. Experimental evidence for tipping points in social convention. Science, 360(6393):1116–1119, 2018. doi:10.1126/science.aas8827.DOI
- [10]Steven H. Strogatz. Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering. Westview Press, Boulder, CO, 2nd edition, 2015.
- [11]Peter Ashwin, Sebastian Wieczorek, Renato Vitolo, and Peter Cox. Tipping points in open systems: Bifurcation, noise-induced and rate-dependent examples in the climate system. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 370(1962):1166–1184, 2012. doi:10.1098/rsta.2011.0306.DOI
- [12]Alan Mathison Turing. The chemical basis of morphogenesis. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 237(641):37–72, 1952. doi:10.1098/rstb.1952.0012.DOI
- [13]Michael C. Cross and Henry Greenside. Pattern Formation and Dynamics in Nonequilibrium Systems. Cambridge University Press, 2009. doi:10.1017/CBO9780511627200.DOI
- [14]Max Rietkerk, Robbin Bastiaansen, Swarnendu Banerjee, Johan van de Koppel, Mara Baudena, and Arjen Doelman. Evasion of tipping in complex systems through spatial pattern formation. Science, 374(6564):eabj0359, 2021. doi:10.1126/science.abj0359.
- [15]Guido Schneider and Hannes Uecker. Nonlinear PDEs: a dynamical systems approach, volume 182 of Graduate Studies in Mathematics. American Mathematical Society, Providence, Rhode Island, 2017. doi:10.1090/gsm/182.DOI
- [16]Dirk Blömker. Amplitude Equations for Stochastic Partial Differential Equations, volume 3 of Interdisciplinary Mathematical Sciences. World Scientific, 2007. doi:10.1142/6440.DOI
- [17]Egbert H. van Nes and Marten Scheffer. Slow recovery from perturbations as a generic indicator of a nearby catastrophic shift. The American Naturalist, 169(6):738–747, 2007. doi:10.1086/516845.DOI
- [18]Christian Wissel. A universal law of the characteristic return time near thresholds. Oecologia, 65(1):101–107, 1984. doi:10.1007/BF00384470.DOI
- [19]Vasilis Dakos, Sonia Kéfi, Max Rietkerk, Egbert H. van Nes, and Marten Scheffer. Slowing down in spatially patterned ecosystems at the brink of collapse. The American Naturalist, 177(6):E153–E166, 2011. doi:10.1086/659945.DOI
- [20]Francesco Romano and Christian Kuehn. Analysis and predictability of tipping points with leading-order nonlinear term. International Journal of Bifurcation and Chaos, 28(8):1850103, 2018. doi:10.1142/S0218127418501031.DOI
- [21]Paul A. Sanders and Robbin Bastiaansen. Early warning of critical transitions: Distinguishing tipping points from Turing destabilizations. Physica D: Nonlinear Phenomena, 495:135300, 2026. doi:10.1016/j.physd.2026.135300.DOI
- [22]Jack Swift and Pierre C Hohenberg. Hydrodynamic fluctuations at the convective instability. Physical Review A, 15(1):319–328, 1977. doi:10.1103/PhysRevA.15.319.DOI
- [23]Stephen Wiggins. Introduction to Applied Nonlinear Dynamical Systems and Chaos. Number 2 in Texts in Applied Mathematics. Springer, New York Berlin Heidelberg, 2nd edition, 2003. doi:10.1007/b97481.DOI
- [24]Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016. doi:10.1073/pnas.1517384113.DOI
- [25]Crispin W. Gardiner. Handbook of Stochastic Methods: For Physics, Chemistry and the Natural Sciences, volume 13 of Springer Series in Synergetics. Springer-Verlag, Berlin, Heidelberg, 2nd edition, 1985. doi:10.1007/978–3–662–02452–2.
- [26]H. A. Kramers. Brownian motion in a field of force and the diffusion model of chemical reactions. Physica, 7(4):284–304, 1940. doi:10.1016/S0031–8914(40)90098–2.DOI
- [27]Nils Berglund. Kramers’ law: Validity, derivations and generalisations. Markov Processes and Related Fields, 19(3):459–490, 2013. arXiv:1106.5799.arxiv.org/abs/1106.5799
- [28]Tom Fawcett. An introduction to ROC analysis. Pattern Recognition Letters, 27(8):861–874, 2006. doi:10.1016/j.patrec.2005.10.010.DOI
- [29]Xiaozhu Zhang, Christian Kuehn, and Sarah Hallerberg. Predictability of critical transitions. Physical Review E, 92(5):052905, 2015. doi:10.1103/PhysRevE.92.052905.DOI
- [30]Peter D. Ditlevsen and Sigfus J. Johnsen. Tipping points: Early warning and wishful thinking. Geophysical Research Letters, 37:L19703, 2010. doi:10.1029/2010GL044486.DOI
- [31]Christian Kuehn. A mathematical framework for critical transitions: Normal forms, variance and applications. Journal of Nonlinear Science, 23(3):457–510, 2013. doi:10.1007/s00332-012-9158-x.DOI
- [32]Karna Gowda and Christian Kuehn. Early-warning signs for pattern-formation in stochastic partial differential equations. Communications in Nonlinear Science and Numerical Simulation, 22(1–3):55–69, 2015. doi:10.1016/j.cnsns.2014.09.019.DOI
- [33]Steven L. Brunton, Nicholas Zolman, J. Nathan Kutz, and Urban Fasel. Machine learning for sparse nonlinear modeling and control. Annual Review of Control, Robotics, and Autonomous Systems, 8:127–152, 2025. doi:10.1146/annurev-control-030123-015238.DOI
- [34]Wenxiang Song, Shijie Jiang, Gustau Camps-Valls, Mathew Williams, Lu Zhang, Markus Reichstein, Harry Vereecken, Leilei He, Xiaolong Hu, and Liangsheng Shi. Towards data-driven discovery of governing equations in geosciences. Communications Earth & Environment, 5:589, 2024. doi:10.1038/s43247-024-01760-6.DOI
- [35]Kyle Siegrist. Probability, Mathematical Statistics, and Stochastic Processes. LibreTexts, 2017. Section 7.2, The Method of Moments. URL: https://stats.libretexts.org/Bookshelves/Probability_Theory/Probability_Mathematical_Statistics_and_Stochastic_Processes_(Siegrist).
- [36]Matthew J. Simpson, Alexander P. Browning, Christopher Drovandi, Elliot J. Carr, Oliver J. Maclaren, and Ruth E. Baker. Profile likelihood analysis for a stochastic model of diffusion in heterogeneous media. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 477(2250):20210214, 2021. doi:10.1098/rspa.2021.0214.DOI
