Failure and recovery in dynamical networks

Failure, damage spread and recovery crucially underlie many spatially embedded networked systems ranging from transportation structures to the human body. Here we study the interplay between spontaneous damage, induced failure and recovery in both embedded and non-embedded networks. In our model the network’s components follow three realistic processes that capture these features: (i) spontaneous failure of a component independent of the neighborhood (internal failure), (ii) failure induced by failed neighboring nodes (external failure) and (iii) spontaneous recovery of a component. We identify a metastable domain in the global network phase diagram spanned by the model’s control parameters where dramatic hysteresis effects and random switching between two coexisting states are observed. This dynamics depends on the characteristic link length of the embedded system. For the Euclidean lattice in particular, hysteresis and switching only occur in an extremely narrow region of the parameter space compared to random networks. We develop a unifying theory which links the dynamics of our model to contact processes. Our unifying framework may help to better understand controllability in spatially embedded and random networks where spontaneous recovery of components can mitigate spontaneous failure and damage spread in dynamical networks.


I. INTRODUCTION
Failure, damage spread and recovery crucially underlie many spatially embedded networked systems ranging from transportation structures to the human body [1][2][3].Advances in the study of networks have led to important progress in understanding resilience and controllability in terms of the interaction between topology and various underlying spreading dynamics [4][5][6][7][8][9].In the case of a simple contagion or contact process such as an epidemic, it is possible for the disease to spread from a single infected source to other neighboring individuals.On the other hand, many phenomena such as the diffusion of innovations [10,11], political mobilization [12], viral marketing [13] and coordination games [14] are characterized by a complex contagion where nodes need to be connected to multiple sources in order to induce a change of their state [15,16].In addition to this induced transition, individuals may spontaneously change their opinion or banks can spontaneously fail [17,18].
The consequences of the interplay between spontaneous damage, induced failure and recovery of components in spatially embedded systems are crucial for systemic risk [19], predictability and controllability but have not yet been systematically explored.Many realworld networks such as power grids, computer networks and social networks are embedded in Euclidean space [3].We here show how the process of embedding and the related characteristic link length impact the predictability of failure-recovery dynamics in networks.
Our model is based on three fundamental processes (i) spontaneous failure independent of the neighborhood (internal failure), (ii) failure induced by failed neighboring nodes if their number exceeds a threshold (external failure) and (iii) spontaneous recovery (see Fig. 1).
The interplay between these three processes results in a phase diagram with a metastable regime where hysteresis and switching between two coexisting states are observed [20].In technological systems, hysteresis effects might be potentially harmful since slight changes of the system's control parameters can entail drastic and abrupt transitions from a seemingly globally stable state to macroscopic inactivity or large-scale outage [8,[21][22][23][24][25][26][27][28][29][30][31][32][33][34].Hysteresis and spontaneous switching between coexisting states in multistable dynamical systems has received great attention for processes ranging from decision making [35], multistable perception [36][37][38] over fluid phase transitions [39], protein folding and unfolding [40] to chemical oscillations [41], magnetic systems [42] and human sleep stages [43].We therefore propose here that the extent of the metastable regime in the parameter space of the phase diagram can be regarded as a measure of predictability.The larger the metastable regime the lower is the predictability of the system.Based on this measure we find that the networks' predictability strongly increases with its regularity.In particular, for the Euclidean lattice, hysteresis only occurs in a very small range of the spontaneous failure rate -compared to random networks with the same average number of neighbors.
Our analytical approach is based on mapping the dynamics to a generalized contact process where a certain minimum number of failed neighboring nodes is necessary to activate the induced failure [44].This strongly suggests that the dynamics does not belong to the Ising universality class as conjectured earlier [20].In addition, we show that our model system is inherently linked to complex contagion phenomena [15,16] and cusp catastrophes [45][46][47].Our unifying framework for random and partially embedded networks helps to better understand predictability and controllability in systems where spontaneous recovery can mitigate spontaneous failure and damage spread.

II. MATERIALS AND METHODS
We study a modified version of the failure-recovery model proposed in Ref. [20].Specifically, we consider a fully rate based kinetic Monte Carlo model [48,49] instead of, as previously, assuming fixed recovery times τ = 0 and τ ′ = 1.The system's components (i.e.nodes) are regarded as either active (not damaged) or inactive (failed).The dynamics is based on three fundamental processes: (i) a node spontaneously fails in a time interval dt with probability pdt (internal failure), (ii) if fewer than or equal to m nearest neighbors of a certain node are active, this node fails due to external causes with probability rdt (external failure) and (iii) spontaneous recovery with probability qdt (internal recovery) or probability q ′ dt (external recovery).The threshold m, similar to threshold rules in complex contagion models [15,50,51] determines if the neighborhood is critically damaged or healthy (Fig. 1).
A low value of m describes the case where a large number of infected neighbors is required in order to sustain the spread of an innovation, opinion or damage.Hence, unlike in an epidemic, where a single infected neighbor can infect a susceptible node, in complex contagion processes spread requires more than one infected neighbor.
Let a(t) ∈ [0, 1] denote the total fraction of failed nodes and z(t) = 1 − a(t) the fraction of active ones.Thus, a(t) = u int (t) + u ext (t) with u int (t) and u ext (t) being the fractions of internal and external failure respectively.The total fraction of failed nodes in the stationary state is referred to as a st .For the derivation of the mean-field rate equations we assume perfect mixing and first concentrate on the internal failure dynamics.The rate equation of internally failed nodes is given by: where the first term accounts for the fact that active nodes internally fail with rate p and the second term corresponds to the recovery of internally failed nodes.
A node is said to be located in a critically damaged neighborhood (CDN) if its number of active neighbors is smaller than or equal to m. External failure is only acting on nodes in a CDN.The probability that a node of degree k is located in a CDN is [20].Consequently, the time evolution of the external failure is described by: with f k being the degree distribution.The first term describes the failure of active nodes in a CDN with rate r and the second term accounts for recovery of externally failed nodes with rate q ′ .(see Fig. 1), with p = 0.9, r = 0.95, q = 1.0, q ′ = 0.1 and m = 1 for a square lattice with N = 1024 × 1024 nodes.All nodes are initially active, z(0) = 1.

A. Time evolution and phase-switching in embedded systems
The coupled mean-field rate equations Eqs. ( 1) and (2) determine the time evolution of the dynamics as shown in Fig. 2. Internal failures first dominate the dynamics but after some time externally induced failures start becoming prevalent in the system.Interestingly, due to the system wide spread of the total failure after a transient phase, indicated by the small fraction of active nodes, the relative abundance of the nodes susceptible to external (active in a CDN) and to internal failure (active) saturate at the same level.However, the process with the higher spreading rate (here external failure with r/q ′ > p/q) soon dominates the dynamics.This explains the relatively small contribution of internal failure in this parameter range which could not be observed employing the mean-field theory of Ref. [20] which assumes that internal and external failure are effectively decoupled processes (case 1 ≈ τ ′ ≪ τ ).Therefore, our dynamical theory allows to analytically describe the time evolution of the model's compartments, i.e. nodes in a certain state.
For the Euclidean (square) lattice we observe phase-switching as shown in Fig. 2 (left).
The fraction of active nodes z(t) undergoes rapid transitions between a phase of high and low activity.Hysteresis only occurs for large node-to-node spreading rates r and a very narrow (left), this region is very narrow.As an inset in the lower left corner, we show a typical cusp catastrophe surface [46] (mean-field k = 4, m = 1) whose bifurcation lines (black) enclose the hysteresis region.For both plots the recovery rates are kept constant (q = 1.0, q ′ = 0.1) and r, p are varied.
range of spontaneous failures rates p.Thus, we find the hysteresis region of the lattice to be much narrower than in random graphs as shown in Fig. 3. Inside this metastable domain rapid and unpredictable phase-switching occurs.In addition, crossing this region results in abrupt and dramatic transitions and it might not be possible to go back to the previous state following the same path.As an example, one can consider a nearly healthy population with a varying spontaneous infection rate which can cause the population to undergo a catastrophic transition to a highly infected state by crossing into the hysteresis region.Going back to the healthy state might not be as easy as just retracing the path followed before.In that sense, the extent of the hysteresis region in parameter space can be regarded as a measure for the predictability of the network's dynamics.The smaller the hysteresis region the less likely it is for the system to end up in this unpredictable situation.
However, the unpredictability in the hysteresis region is manifested in two ways: (1) In finite systems, random phase switching between two unstable states is observed where the mean of the random switching times increases exponentially with system size [20].(2) In the thermodynamic limit no switching is observed but the initial configuration and small random events in the initial temporal evolution of the system determine to which of the two stationary states the system converges.The latter behavior is characteristic for non-selfaveraging spin glasses [52][53][54].As an inset in Fig. 3, we show the relation to cusp catastrophe surfaces accompanying the model's dynamics.Cusp catastrophes are a prominent example in catastrophe theory describing hysteresis and possible sudden transitions as a consequence of slightly varying control parameters with applications in population dynamics, mechanical and biological systems [45][46][47].The bifurcation lines enclose the hysteresis region and merge at the cusp point.The cusp point is a degenerate critical point where not just the first derivative, but also higher derivatives of the potential function vanish.The degeneracy of this critical point can be unfolded by expanding the potential function as a Taylor series in small perturbations of the parameters r, p, and a with a characteristic fourth-order polynomial [46].For a detailed analytical treatment, see Appendices A and B. We define the hysteresis areas enclosed by the bifurcation lines in this parameter range as A M F T H (meanfield) and A SL H (square lattice).The ratio A M F T H : A SL H ≈ 200 : 1 shows that the Euclidean lattice is substantially more predictable in the presence of failure and recovery compared to random networks, where non-local connections induce a faster damage spread.For the square lattice, in contrast to random networks, failure cascades are only sustained for a large damage spreading rate r within a narrow region of the ratio p/q.To obtain the phase diagram we studied the hysteresis behavior and fluctuations for different fixed values of r by varying p.More details about the Euclidean lattice and its critical behavior for different values of m are described in Appendix B. The model's dynamics is very rich and we show in III D that for certain values in the parameter space we encounter closed orbits.We further describe the dynamics and connections to other models, in particular, Schlögl's first (contact process) and second models and the relation to cusp catastrophes, in Appendix A.

B. Spreading dynamics
The dynamics can be driven by the field-like spontaneous failure term or the spread of failure can be triggered by the neighboring failed nodes.We first discuss the dynamics in the hysteresis region of a square lattice with N = 50 × 50 nodes, cf.The spontaneous infection term, analogous to an external field, enables the dynamics to form multiple seeds from where the transitions might start.The regions invaded by different seeds expand and move on the lattice.Some of them merge and form larger clusters of active or failed nodes.After some time a stable phase develops.
In the limit of a vanishing external field we expect nucleation determining the growth of a certain phase.Nucleation is exemplified for a small value of p = 0.05 in Fig. 4 be better controlled compared on a random graph.When both paths cross the hysteresis region, e.g., for r/q ′ = 10, both dynamics show a discontinuous transition (Fig. 5 (right)).This however requires parameter tuning, that is, large values of the external spreading rate r/q ′ .In order to better understand the dramatic differences between random networks and embedded lattices, we analyze here the transition from a square lattice to a regular random graph.To this end, we follow the transition model of Danziger et al. [58] and study the phase space of an embedded system with degree k = 4 where randomly chosen nearestneighbor links are replaced by longer-range links.The lengths l of the links are distributed according to an exponentially decaying distribution P (l) ∼ exp (−l/ζ), with a link length l and characteristic link length ζ.In the thermodynamic limit, a square lattice in the limit of ζ → 0 is recovered, whereas in the limit of ζ → ∞ we obtain a regular random graph (as all link lengths are equally likely).The phase diagrams of an embedded system with k = 4 and exponential link length distribution, in the presence of processes (i-iii), are shown in Fig. 6.We clearly observe the transition from a situation similar to the square lattice for ζ = 0.1 to a regular random graph for ζ = 10.This again illustrates the strong dependence of the extent of the metastable region on the topology.In other words, a variation in the

D. Oscillatory behavior
We will briefly describe the possibility of encountering limit cycles in our dynamics.We investigate this behavior by studying the Lyaponuv function [47].In our case, the Lyapunov function V (a, u int ) is derived from the following equations (α, β > 0): which are equivalent to Oscillatory behavior (limit cycles) for q ′ > q. (left) We show the mean-field time evolution of different compartments, cf.Fig. 1, for p/q = 19/81, q = 0.01, r/q ′ = 3125/1296, q ′ = 1.0, k = 4 and m = 1.The values of p and r correspond to the ones of the bifurcation point as described in Appendix B. We clearly see the oscillatory behavior as a consequence of q ′ > q.
Without loss of generality, we set β = 1 and compute V (a, u int ) to: where we used the binomial theorem and set α = (q − q ′ ) /p.We find that if q > q ′ since β = 1.For q > q ′ we therefore expect no oscillatory behavior, i.e. no closed orbits.However, for q ′ > q, we show in Fig. 7 the existence of closed orbits.The purple region in Fig. 7 (right) illustrates the regime where we measured a periodic orbit analyzing the Fourier transformed time evolution of the fraction of active nodes.We also show the oscillatory dynamics of a non-embedded regular network in video 4 (oscillations) [59].As discussed in Appendix A, for q = q ′ , the differential equations Eqs. ( 1) and ( 2) can be decoupled and one obtains a single first-order differential equation which has no periodic solutions [47].
This demonstrates that the phase diagram is substantially more complex than previously believed.Specifically, limit cycles occur for q ′ /q < 1 in a narrow region in the phase diagram.
This deterministic behavior is markedly different from the stochastic switching dynamics in the hysteresis region but likewise challenges control.

IV. DISCUSSION
We have derived a unifying framework for the interplay between failure, damage spread and recovery in spatially embedded and random networks.The theoretical description links diverse phenomena such as complex contagion and phase-switching due to metastability and the occurrence of cusp catastrophes.The number of failed neighbors necessary to allow external failure to act on a node is a crucial parameter of the system.Our analysis revealed that the phase space is substantially more complex than previously known owing to the coexistence of limit cycles and random phase switching within hysteresis.
We analytically demonstrated that the mean-field description of the stochastic model systems is equivalent to cusp catastrophes with two bifurcation lines enclosing a metastable domain where two stable stationary states coexist.Inside this metastable region, large fractions of nodes suddenly fail and recover.We propose the hysteresis (metastable) area as a predictability measure for the state of the system of a given topology and dynamics.Our results show that the transition from a random regular network to an embedded network with a short characteristic link length is characterized by a dramatic shrinking of the metastable domain.This suggests that embedded systems with short characteristic link lengths whose dynamics is captured by processes (i-iii) are substantially more robust against abrupt spontaneous and cascading failures compared to non-embedded systems.
Moreover, we have also shown that our theoretical framework is able to describe essential features of the model's time evolution and that it captures spontaneous failure as an external field in analogy to magnetic systems.However, based on the connection to contact process dynamics we find that the model does not belong to the Ising universality class as conjectured earlier [20].The arguments in Appendices A and B show the similarities to the (nonequilibrium) contact process belonging to the directed percolation universality class [60].In fact, as mentioned by Grassberger [61], relating this dynamics to the Ising universality class would mean an extension of the universality hypothesis from models with detailed balance to models without it.Unpredictability observed in the hysteresis region results from two effects.For finite systems, unpredictable random phase switching between two unstable states is observed.In the thermodynamic limit, however, small fluctuations in the initial phase of the systems dynamics determine the stationary (stable) state of the system.
Our framework helps to better understand predictability and controllability in spatially embedded and random systems where spontaneous recovery, spontaneous failure and cascading failure lead to a remarkably complex dynamic interplay.second model [61], the phase space of cusp catastrophes or imperfect bifurcations [47].
with the contact process where also at least one failed neighbor is necessary to turn on the spreading dynamics.We also find the corresponding exponents β = 1 and δ h = 2 in the vicinity of r c = 1/2.In general, we expect this behavior for any regular graph with degree ) (at the critical point).That is the reason why we again find the contact process exponents in the latter example and a critical value of r c = 1/2 which is just the critical point of Eq. (A2) divided by k.
At the bifurcation point the quantity ∆a st (r, p 0 ) = a st (r, p 0 ) − a st (r 0 , p 0 ) increases with r as ∆a st (r, p 0 ) ∝ (r − r 0 ) β and with p as ∆a st (r 0 , p) = a st (r 0 , p) − a st (r 0 , p 0 ) ∝ (p − p 0 ) 1/ δh where a st (r 0 , p 0 ) = 1/3, β = 1/3 and δh = 3.For this example, it is straightforward to show that the quadratic term in the Taylor expansion of f (a st , r, p) = ra 2 st (1 − a st ) + p(1 − a st ) − a st around a st (r 0 , p 0 ) = 1/3 vanishes yielding the characteristic polynomial of the cusp catastrophe [46].In general, the spontaneous recovery model resembles the dynamics of a modified contact process where a certain minimum number of nodes is necessary to turn on the spreading dynamics [44].As already mentioned in previous studies and as discussed in the latter examples, slight modifications of the standard contact process dynamics might have dramatic effects on the system's dynamics leading to uncontrollable abrupt transitions [30,32]. in Appendix A and a non-zero stationary state a st (r) is found if r > r c (r c = 1/4 MFT).
From MFT we also find a st (r) ∝ (r − r c ) β with β = 1.At r c the order parameter grows continuously.Applying the field term in this example one finds a st (r c , p) ∝ p 1/δ h with δ h = 2.
We show the order parameter a st (r) as a function of r for the square lattice in comparison with MFT in Fig. 9 (right).We also analyzed the critical behavior in the vicinity of the critical point r c = 0.47(1) of the square lattice (see Fig. 9 (right)).The growth of the order parameter with β = 0.569(16) (Fig. 10 (left)) and δ −1 h = 0.265(1) (Fig. 10 (right)) agrees with the corresponding contact process values β = 0.586 (14) [64] and δ −1 h = 0.285(35) [65].We thus conclude that the model resembles standard contact process dynamics in this case.
For m = 2 the transition in MFT is characterized by a jump at r c = 1.226 from zero to a st (r c ) = 0.322.In the simulations on the square lattice we found strong dependence on the initial conditions.
We did not find a non-zero value of a st (r) for m < 2 having a circle-shaped seed as initial condition on the square lattice.The situations where m = 0 or 1 mean that three or four failed neighbors are needed to turn on external failure.Starting from a circle-shaped seed the dynamics will never reach a stable configuration besides the absorbing state (all nodes are active).Nevertheless, we are able to study the dynamics for m = 1 as before by introducing the field-like spontaneous failure term again (p > 0).In the mean-field situation Eq.A1 yields for the bifurcation point (r 0 , p 0 ) = (3125/1296, 19/81) ≈ (2.41, 0.23) and a st (r 0 , p 0 ) = 0.4.This point is also shown in Fig. 3 (right).Similar to the arguments in Appendix A, it is again straightforward to show that the quadratic term in the Taylor expansion of the polynomial describing the stationary states around a st (r 0 , p 0 ) = 0.4 vanishes yielding the characteristic polynomial of the cusp catastrophe [46].The black lines in the latter figure characterize the hysteresis region with two coexisting stationary states similar to Fig.  conclude that the cusp point where both spinodals meet is located around (r c , p c ) = (0.86(1), 0.117(3)).
In summary, the arguments in Appendix A and above for the case m = 3 (analytical and numerical) show the similarities between our model and the (non-equilibrium) contact process belonging to the directed percolation universality class [60].Thus, we do not expect the dynamics to belong to the Ising universality class as conjectured in Ref. [20].As already mentioned in Ref. [61] if this contact process dynamics would belong to the Ising universality class it would mean the extension of the universality hypothesis from models with detailed balance to models without it.

Figure 1 .
Figure 1.Model.(a) Spontaneous failure (A → X) and spontaneous recovery (X → A) takes place with rates p and q, respectively.(b) A node may also fail (become inactive) dependent on its neighborhood, if too few active nodes n ≤ m sustain the node's activity (A → Y with rate r).In addition, a failed node Y recovers (Y → A) with rate q ′ .Illustration for m = 1.Active nodes are purple.

Figure 2 .
Figure 2. Model dynamics on a square lattice.(left) Phase switching for p = 0.1065, r = 0.95, q = 1.0, q ′ = 0.1 and m = 1 on a square lattice with N = 50 × 50 nodes (z the fraction of active nodes).(right) Time evolution of different model compartments, i.e. nodes in a certain state

Figure 3 .
Figure 3. Phase diagrams of a regular random graph with k = 10 and a square lattice.(left) The phase space of a regular random graph with N = 100, 000 nodes and k = 10, m = 4 (red dots) compared to the mean-field prediction (black lines).(right) The phase diagram of the square lattice (red spinodals) for m = 1 obtained through simulations on a system with N = 2048 × 2048 nodes.Inset shows a blow-up of the hysteresis region.Compared to the regular random network Fig. 2. The two mechanisms are illustrated in video 1 (transition down) [55] and video 2 (transition up) [56].

Figure 4 .
Figure 4. Nucleation for vanishing spontaneous infection.Simulation of the spontaneous recovery model with p = 0.05, q = 1.0, r = 10.0, q ′ = 0.1 and m = 1 on a square lattice with N = 128 × 128 nodes.(left) Initially, multiple spreading seeds of failed lattice sites (red) form due to spontaneous failure.(right) Contact dynamics (external failure) dominate and active sites (green) are displaced by failed ones (red).For further details we refer to the video version of the dynamics: video 3 (vanishing spontaneous infection).

30 pFigure 5 .
Figure 5. Illustration of hysteresis effects.Simulation of the failure-recovery model ( z average of the fraction of active nodes) with q = 1.0, q ′ = 0.1, m = 1 for a square lattice and a regular random graph both with k = 4 and N = 512 × 512 nodes.(left) The transition is discontinuous for r = 0.7 for the random graph but continuous for the square lattice.(right) Only for large r/q ′ (shown for r = 1.0) the transition is discontinuous for both the random graph and the square lattice.The black arrows indicate the direction of the simulation loop.

Figure 6 .
Figure 6.Phase diagrams of spatially embedded networks with k = 4 in comparison to a random graph.We set q = 1.0, q ′ = 0.1, m = 1 and perform the simulation for networks with degree k = 4 and N = 250, 000 nodes.The phase space of a regular random graph with k = 4 (black squares) is compared to an embedded network with the same degree k = 4 for different values of the characteristic link length ζ (blue dots, green inverted triangles, red diamonds).For large ζ we obtain the phase space of the regular random graph and for small ζ the square lattice behavior is recovered.

Figure 7 .
Figure 7. Oscillatory behavior (limit cycles) for q ′ > q. (left) We show the mean-field time

Figure 8 .
Figure 8. Analogy to the contact process and Schlögl's second model.(left) The order parameter a st (r) as a function of the external failure rate r for k = 1 and m = 0 (mean-field).Analogous to the contact process the black solid line corresponds to the situation where the vanishing spontaneous infection term p → 0 leads to a second-order phase transition at r = r c .The dashed lines show a smeared out transition due to the non-zero spontaneous infection rates p = 0.001, 0.01 [62].(right) The phase space for k = 2 and m = 0 (mean-field).One clearly sees the hysteresis region where two states coexist (low density and high density failure phases).The spinodals (black solid lines) merge at the bifurcation point (r 0 , p 0 ) = (27/8, 1/8).The critical point, r c = 4 indicates the transition point without additional field-like term (p = 0).This situation is similar to Schlögl's

Figure 9 .
Figure 9. Growth of the order parameter on a square lattice for different m.Simulation of the spontaneous recovery model without internal failure dynamics (p = 0) and q ′ = 1.0 for different m. (left) The order parameter a st (r) as function of r for m = 4 and (right) for m = 3.The simulations have been performed on a square lattice with N = 1024 × 1024 nodes.

Figure 10 .
Figure 10.Critical exponents of the square lattice with m = 3. Simulation of the spontaneous recovery model with p = 0, q ′ = 1.0 and m = 3. (left) The order parameter a st in the vicinity of the critical point r c = 0.47(1) for different r.The exponent found indicates contact process dynamics where β = 0.586(14) [64].(right) The order parameter a st at the critical point r c = 0.47(1) for small values of p.The critical exponent measured also indicates contact process dynamics with δ −1 h = 0.285(35) [65].The simulations have been performed on a square lattice with N = 1024 × 1024 nodes (1500 samples).The insets show the PDF's of the exponent's bootstrap analysis.

( 15 pFigure 11 .
Figure 11.Coexisting states and fluctuations on a square lattice for m = 1.(left) Measuring the region of coexisting states by running through a hysteresis-like loop on a square lattice with N = 2048×2048 nodes for fixed r and varying p. (right) Fluctuations of the spontaneous recovery model for different fixed values of r and varying p.At around r = 0.86(1) and p = 0.117(3) we find the largest fluctuations corresponding to the bifurcation point.Simulations were performed for N = 128 × 128 (smaller symbols) and N = 256 × 256 (larger symbols) nodes (50 samples).