Recovery time after localized perturbations in complex dynamical networks

Maintaining the synchronous motion of dynamical systems interacting on complex networks is often critical to their functionality. However, real-world networked dynamical systems operating synchronously are prone to random perturbations driving the system to arbitrary states within the corresponding basin of attraction, thereby leading to epochs of desynchronized dynamics with a priori unknown durations. Thus, it is highly relevant to have an estimate of the duration of such transient phases before the system returns to synchrony, following a random perturbation to the dynamical state of any particular node of the network. We address this issue here by proposing the framework of single-node recovery time (SNRT) which provides an estimate of the relative time scales underlying the transient dynamics of the nodes of a network during its restoration to synchrony. We utilize this in differentiating the particularly slow nodes of the network from the relatively fast nodes, thus identifying the critical nodes which when perturbed lead to significantly enlarged recovery time of the system before resuming synchronized operation. Further, we reveal explicit relationships between the SNRT values of a network, and its global relaxation time when starting all the nodes from random initial conditions. Earlier work on relaxation time generally focused on investigating its dependence on macroscopic topological properties of the respective network. However, we employ the proposed concept for deducing microscopic relationships between topological features of nodes and their respective SNRT values. The framework of SNRT is further extended to a measure of resilience of the different nodes of a networked dynamical system. We demonstrate the potential of SNRT in networks of Rössler oscillators on paradigmatic topologies and a model of the power grid of the United Kingdom with second-order Kuramoto-type nodal dynamics illustrating the conceivable practical applicability of the proposed concept.


Introduction
The abundance of dynamical systems involving large collections of individual entities interacting with each other on complex networks can hardly be further exaggerated [1,2,3,4,5,6].Such networked dynamical systems often exhibit a multitude of stable states, whereby sustained operation of the system in the desired state is of central importance.The desired operational state (DOS) in such systems is commonly associated with the synchronized motion of the dynamical components coupled on their networked architecture [7,8].'Permissible' and 'impermissible' random perturbations (according to the terminology used by Menck et al. [9]) often disrupt the functionality of coupled dynamical systems operating in the synchronized state, driving them away either to an arbitrary state still inside the basin of attraction of the synchronized state, or to an altogether different dynamical regime.The former situation arising on account of 'permissible' perturbations, leads to arbitrary durations of desynchronized dynamics before the system regains synchronous motion.On the other hand, 'impermissible' perturbations permanently forbid the return of the system to the synchronized state, unless again affected by an appropriate external perturbation.
The stability of the synchronized state against the aforementioned perturbations is critical in the operation of many real-world networked dynamical systems such as ecosystems, power grids, the human brain, etc. [10].Subsequently, the influence of topological features on network synchronizability and the stability of the synchronized state has been well-investigated [7,11].In this context, significant developments constitute the master stability function (MSF) [12], basin stability (BS) [9] and its extensions to single-node BS [8], multiple-node BS [10], and survivability [13].On the contrary, the issue of recovery time (RT) of complex dynamical networks following a random perturbation, which is a measure of how quickly the network relaxes back to the DOS (e.g., a synchronized state) after being perturbed from the same, has received considerably less attention and is currently under active investigation [14,15,16,17,18,19,20,21,22,23,24,25].However, this is an important problem concerning dynamical robustness of complex networks, i.e., the ability of a network to restore its dynamical activity to the DOS when its dynamical components are subject to random perturbations.For example, the loss of synchrony in engineered systems such as power grids can lead to large-scale power blackouts [8].
In biological systems such as the human brain, it can impede cognitive functions such as information transfer [26] and memory [27].Thus, quickly restoring synchrony following desynchronizing perturbations is crucial in such coupled dynamical systems.Consequently, it is highly desirable to have an estimate of the RT of the system to the desired stable regime, following a perturbation to a particular node of the network (otherwise operating in the DOS).This creates the possibility of identifying (and safeguarding) specific nodes which when perturbed lead to particularly large RT of the system.In this regard, we propose here the framework of single-node recovery time (SNRT) addressing the aforementioned issue.We reserve a formal definition of SNRT to Section 2.3.
SNRT of a node under investigation relates to the time taken by the system operating in the DOS (e.g., a synchronized state) to return to the same, following a random perturbation to the dynamical state of the respective node.The framework of SNRT provides information on the different relative time scales underlying the transient dynamics of the respective nodes of the network during its restoration to the DOS.This can be utilized in revealing the particularly slow nodes of the network in contrast to the relatively fast ones, leading to the identification of the vulnerable nodes which when perturbed significantly elevate the RT of the whole system.Further, this can provide an insight into the global relaxation time (GRT) of the network to the DOS, when starting all its nodes from arbitrary initial conditions.We provide a formal definition of GRT in Section 2.4.The GRT is referred to as the global synchronization time when the synchronized state is the DOS of the network.
Previously, the dependence of synchronization time on various macroscopic topological properties of the corresponding networks has been investigated.For example, Grabow et al. [22] have shown that, largely insensitive to the type of oscillators (phase, multi-dimensional, neural), their intrinsic dynamics (periodic, chaotic) and their coupling schemes (phase-difference, diffusive, pulse-like), networks with a fixed average path length consistently synchronize slowest in the small-world regime.This is a rather unexpected phenomenon given that small-world topology has been suggested to facilitate network synchronization at weaker coupling strengths (than for analogous, appropriately normalized globally coupled systems) [28,29,30] as well as being more robust to random perturbations [9].Also, the MSF approach [12] has been extended by Grabow et al. [23] to provide analytical predictions for the asymptotic synchronization times, which is, however, locally restrictive to small perturbations.Further, the dependence of synchronization time on various macroscopic topological features such as the average path length, global clustering coefficient etc. has been systematically studied.In this context, the framework of SNRT introduced in this work is capable of providing a microscopic view on the response to arbitrary perturbations of individual nodes as well as exploring relationships between various topological features of the nodes and their respective SNRT values.
Finally, we advance on the framework of SNRT for quantifying the resilience of networked dynamical systems [31].Resilience of a given dynamical system has been defined in at least two different ways, namely, engineering resilience and ecological resilience [32].Engineering resilience (according to Pimm [33]) of a dynamical system characterizes its resistance to disturbance and speed of return to its equilibrium, following a perturbation [32,34,35].It implicitly assumes global stability, i.e., the existence of only one equilibrium state, or, if other operating states exist, they should be avoided by applying safeguards [32,34,35].On the other hand, ecological resilience [36] presumes the existence of multiple stable states and the tolerance of the system to disturbances that facilitate transitions among the stable states [32,34,35].In this case, resilience of the system is measured by its capacity to remain in the same basin of attraction in the face of random perturbations [32,34,35].
Ecological resilience of the multiple stable states of a system relates to the volume and geometry of their respective basins of attraction [9,35].In this context, Mitra et al. [35] recently reconsidered the concept of ecological resilience and its three crucial aspects of 'latitude' (L), 'resistance' (R) and 'precariousness' (P r) [36].They redefined L, R and P r in a rigorous dynamical systems context and utilized this as a foundation for characterizing multistability and proposing the quantifier of integral stability [35].Besides its extension to quantifying multistability, the framework of ecological resilience has generated widespread interest (cf.[37] and references therein).On the other hand, the facet of engineering resilience, perhaps on account of its restrictive scope to globally stable systems has received considerably less attention.However, it is equally crucial to know how long does a system operating in its desired stable state take to retain functionality in the respective dynamical state, following a random perturbation.As mentioned earlier, networked dynamical systems often exhibit multiple stable states, such as the coexistence of synchronized and desynchronized dynamical regimes, which is a notable example of bistable behaviour [10].Thus, we extend here the traditional scope of engineering resilience to quantifying the resilience of the DOS (e.g., the synchronized state) in such multistable coupled dynamical systems.More precisely, we relate the engineering resilience of each node of a networked dynamical system (for the DOS) to the SNRT of the corresponding node such that a node with a lower value of SNRT is considered more resilient and vice versa.Thus, the proposed architecture of engineering resilience complements the existing framework of ecological resilience in characterizing the overall resilience of networked dynamical systems.
This paper is further organized as follows: In Section 2, we outline the general methodology for calculating SNRT values for a given networked dynamical system.In Section 3, we illustrate applications of SNRT to networks of Rössler oscillators and a model of the power grid of the United Kingdom with second-order Kuramoto-type nodal dynamics.Finally, we present the conclusions of our work in Section 4.

Preliminaries
Consider a network of N oscillators (nodes) where the intrinsic dynamics of the i th oscillator (represented by the d-dimensional state vector is described by ẋi = F i (x i ), where where is the overall coupling strength, A is the adjacency matrix which captures the interactions between the nodes such that A ij = 0 if node j influences node i and an arbitrary coupling function from node j to node i.For the illustrations in this paper (Sec.3), we consider identical nodal dynamics (F i = F ∀ i), symmetric adjacency matrices (A ij = A ji = 1 if nodes i and j are connected and A ij = A ji = 0 otherwise) and identical coupling functions (H ij = H ∀ i, j).
We assume the desired operational state (DOS) is an attractor of the system that we denote by A with the corresponding basin of attraction B (A).We usually denote a trajectory on A by x(t).

Regularized reaching time
For a trajectory initiated from x(0) = (x 1 (0), x 2 (0), . . ., x N (0)) T (∈ B (A)), the attractor is usually reached asymptotically.This implies that the associated reaching time is not finite, thus posing a problem in its measurement.A way to address this problem is regularization of the time variable [25].We now discuss the framework of regularized reaching time proposed by Kittel et al. [25] and then resort to the same in dealing with the above issue.
The distance of a state at time t on a trajectory initiated from x(0), to the desired attractor is given by, where x (t, x(0)) represents the state of the system after a time t has elapsed.The last-entry time for the corresponding trajectory to enter a δ-neighbourhood around the desired attractor (A) is given by where δ → 0 leads to the aforementioned divergence.
Kittel et al. [25] argued that even though the actual reaching times diverge for the respective trajectories, their differences actually converge.Subsequently, they proposed the regularized reaching time T RR (x(0)) for any trajectory (starting from x(0)) as the difference between the last-entry times along the respective trajectory and a reference trajectory (starting from x ref ), for a given δ > 0. This can be interpreted as the additional time the trajectory starting from x(0) needs to arrive in the vicinity of the desired attractor, after the reference trajectory starting from x ref has reached it.Thus, A positive or negative value of T RR (x(0)) indicates that the considered trajectory arrives by this value later or earlier than the reference trajectory, respectively.This allows the distinction between slower and faster trajectories of the system during their return to the desired attractor (cf.Kittel et al. [25] for further details on T RR ).

Single-node recovery time (SNRT)
In the following, we outline the general methodology for calculating SNRT values for all nodes of any networked dynamical system.We assume that the networked dynamical system of Eq. ( 1) is in its DOS x(t).Now, consider a 'permissible' random perturbation ∆x i to the dynamical state of the i th oscillator of the network.The system (otherwise functioning in its DOS) is pushed to a perturbed state x ∆i = (x 1 , x2 , . . ., xi + ∆x i , . . ., xN ) T .The perturbed state (on account of the perturbation being 'permissible') remains in the basin of attraction B (A) of the DOS because we chose x ∆i to be permissible, thus ensuring the system's return to the same.We then define the SNRT of the i th oscillator as, where P i is the projector into the subspace of the i th oscillator, i.e., is the density of 'permissible' perturbed states in state space that the i th oscillator may be pushed to even via large perturbations with ρ i (x ∆i ) d∆x i = 1, where this integral is performed over the subspace of the i th oscillator.The integrals in the numerator and denominator of Eq. ( 3) are performed over the basin of attraction of the DOS (i.e., x ∆i ∈ B (A)).Thus, the SNRT of the i th oscillator T 1 R (i) corresponds to the mean regularized reaching time of the system to the DOS, after a random 'permissible' perturbation hits the respective oscillator.
Equation ( 2) demands the choice of a reference initial condition x ref ∈ B (A) that needs to be kept fixed for all single-node perturbations to allow comparability between SNRT values of the different nodes of the network.However, different choices of x ref (as long as we do not choose it on A) simply lead to a shift of all T 1 R (i) values by a constant only [25].Although not posing a serious problem, this methodology of choosing x ref leaves an element of arbitrariness.As we seek to utilize the T 1 R values in estimating the duration by which a particular node of the network returns faster or slower than another, this naturally leads to the condition demanding the lowest T 1 R (i) value to be 0, min Using this equation, we can fix x ref implicitly instead of explicitly specifying it.We denote the node (or one representative if there might be more) with T We now present an algorithm for estimating the SNRT of the i th oscillator/node of a network (modelled using Eq. ( 1)): (i) Identify the DOS of the network.This state often corresponds to the synchronized dynamics of the oscillators coupled on the network.
(ii) When the attractor corresponding to the DOS A is not a fixed point, choose P (> 1) different points on the attractor.Otherwise, choose P = 1.
(iii) For a particular value of p (p = 1, 2, . . ., P ), perturb the i th oscillator by drawing I C randomly distributed (according to ρ i (x ∆i )) initial conditions x p ∆i (j = 1, 2, . . ., I C ) from inside the basin of attraction of the DOS, while each time initiating the system from the DOS corresponding to the p th point on A. For the results described in this paper, we assume a uniform distribution of ρ i (x ∆i ).
(iv) For a fixed value of δ > 0, calculate the last-entry time t L (x p ∆i (j) , δ) of the system for the j th initial condition.
(v) Estimate the SNRT of the i th oscillator (T 1 R (i, p)) for the p th point on the attractor as, and then average over p to obtain, (vi) Finally, we identify the node i ref with the minimum T 1 R value (as computed above for all nodes) T 1 R min and subtract this value from the T 1 R (i) of the i th oscillator computed above, thus yielding the SNRT value of the respective oscillator.
As mentioned earlier, this concept of SNRT can be utilized in identifying the slow and fast nodes/sub-components of networked dynamical systems.Also, the proposed machinery can be used in revealing systematic relationships between SNRT values of different nodes and their respective topological features.Further, it can be extended to a measure of (engineering) resilience of the different nodes of a networked dynamical system (see Sec. 2.6) and thereby utilized in identifying the particularly vulnerable nodes of the network as well as the more resilient ones.Subsequently, this framework of SNRT can be potentially relevant in selecting specific nodes to be safeguarded from external perturbations.We now define the global relaxation time of a network, which relates to the overall time scale of the dynamics of a network during its relaxation to the DOS.

Global relaxation time (GRT)
Starting all nodes of a networked dynamical system from random initial conditions inside the basin of attraction of the desired attractor involves a transient time before the system reaches the associated attractor.We label the duration of this transient regime as the relaxation time of the system for the respective initial state.We estimate the global relaxation time T R of a network as follows: (i) Draw I C random initial conditions from inside the basin of attraction of the DOS.The j th initial condition can be written as x (j) = x 1 , x 2 , . . ., x N T where j = 1, 2, . . ., I C .Note, that the value of I C chosen for computing the GRT can be different from the one chosen for calculating SNRT above (Sec.2.3).
(ii) For the j th initial condition, calculate the last-entry time t L (x (j) , δ) of the system with the same value of δ as chosen for computing SNRT (Sec.2.3).
(iii) Calculate the GRT of the network as, (iv) Finally, subtract the value of T 1 R min (obtained in Sec.2.3) from the T R computed above in obtaining the GRT of the network.
When the DOS of the network is a synchronized state, its GRT is referred to as the global synchronization time of the system.
The GRT of a network is useful for quantifying the expected transient time to reach the DOS, when starting the system from a random initial condition.In Section 3, we will illustrate the relationship between SNRT values and the GRT of a network for different systems.
In order to avoid terminological confusion, we explicitly distinguish between the usage of recovery, reaching and relaxation time.We use the term recovery with reference to the time taken by the system to recover from a perturbation and resume operation in the DOS.On the other hand, when initiating all the nodes of the system from arbitrary conditions, the term relaxation is used with reference to the time before the system relaxes to the DOS.It is the difference between the relaxation times of a trajectory starting from a particular initial condition and that of a reference trajectory, which is termed as the regularized reaching time for the respective initial condition.

Single-node basin stability (SNBS)
The BS of a particular attractor relates the volume of its basin of attraction to the likelihood of returning to the same attractor in the face of random perturbations [8,9,10].More precisely, the BS of a particular attractor is defined as the fraction of the volume of the state space belonging to the basin of attraction of the respective attractor [8,9,10].In practice, BS of any particular attractor is estimated using a numerical Monte-Carlo procedure by drawing random initial states from a chosen subset of the entire state space, simulating the associated trajectories, and calculating the fraction of trajectories that approach the respective attractor [8,9,10].As mentioned earlier, the ecological resilience of a stable state is (among other properties) determined by the size and shape of its basin of attraction, and is therefore closely related to its

BS.
BS has been further extended to the framework of single-node BS (SNBS) [8,10].SNBS S 1 B of a node under investigation corresponds to the probability of the network (operating in the DOS) to return to the DOS, after that particular node has been hit by a non-infinitesimal perturbation [8,10].We refer to Mitra et al. [10] for the general methodology used throughout this paper for estimating SNBS values for any networked dynamical system.

Engineering resilience
SNBS is a measure related to the ecological resilience of a node subjected to a random perturbation (when the entire network was functioning in the DOS prior to the disturbance).The time elapsed before the network returns to its DOS, following a 'permissible' random perturbation to a particular node determines the engineering resilience of the respective node.We recommend incorporating the engineering resilience of a node (besides its ecological resilience as characterized by its SNBS value) quantified as being inversely related to its SNRT value, in measuring the overall resilience of the respective node.For example, it may be possible that two nodes of a networked dynamical system have very similar values of SNBS.However, the SNRT values of the respective nodes may differ significantly (as we shall illustrate using examples in Sec. 3).In such a situation, the new framework of SNRT should complement that of SNBS in appropriately assessing the resilience of the respective nodes of a network.

Examples
We shall now illustrate applications of SNRT to various networked dynamical systems.Here, we specifically apply the framework to networks of oscillators with continuous time dynamics (Eq.( 1)) exhibiting bistability on account of coexisting synchronized and desynchronized regimes, where the former is considered as the DOS of the system.However, the framework is generally applicable to (continuous or discrete time) networked dynamical systems with multiple coexisting states as well.

Deterministic scale-free network of Rössler oscillators
We first consider a network of N identical Rössler oscillators [38], with diffusive coupling in the y-variable between two coupled nodes such that the full dynamical equations of node i (in analogy with Eq. ( 1)) read We use the parameter values of a = b = 0.2 and c = 7.0 for which the intrinsic dynamics of each uncoupled Rössler oscillator is chaotic.As a specific network topology, we use an undirected deterministic scale-free network proposed by Barabási, Ravasz and Vicsek [39].For the simulations carried out in this section, we generate a deterministic scale-free network developed over 3 generations and hence, comprising N = 81 nodes (Fig. 1).
We consider the DOS of the network as the completely synchronized state, which corresponds to all oscillators following the same trajectory.Further, we choose = 0.8 from the stability interval (of the completely synchronized state) predicted by the MSF [12] and set δ = 10 −4 [40] for estimating the SNRT ( T 1 R ) values, using the procedure described in Section 2.3.
We calculate and present the individual T 1 R (on log 10 scale) values of the nodes in Fig. 2(a).Interestingly, the 3 generations of nodes split into three classes in terms of their T 1 R values such that the lower the generation in the hierarchy, the higher is the SNRT of the individual nodes comprising it (as evident from the histogram in Fig. 2(b)).We next compare these findings with two key topological features of DSF network.The connectivity of a node i (for i = 1, 2, . . ., 81) is described by its degree k i = j A ij (where A is again the adjacency matrix of the respective network [6]).The betweenness centrality bc i of a node i is related to the fraction of shortest paths between all pairs of nodes that pass through node i [6].For an N -node network, the betweenness centrality of each node may further be normalized by dividing by the number of node pairs excluding N i.e., N   2   , obtaining a value between 0 and 1.Thus, bc i = 2 N (N −1) j =k =i σ j,k where σ j,k is the total number of shortest paths from node j to node k and σ i j,k is the number of such shortest paths which pass through node i [6]. Figure 2(c, d) shows the relationship of the log 10 ( T 1 R ) values with the topological features of degree k and betweenness centrality bc of the nodes, respectively.The T 1 R values do not exhibit any marked relationship with these two characteristics.This is further illustrated by the correlation coefficient of -0.040 (-0.085) between T 1 R and k (bc).We summarize our results in Fig. 1, which displays the network topology where the size of each node is proportional to the degree and the color corresponds to the T 1 R value of the respective node.The nodes in the 3 rd generation of the deterministic scale-free network comprise its slow nodes.It is expected that the overall time scale of synchronization of a network should be governed by the node with the highest SNRT, i.e., the 'slowest' node of the system.The 'slowest' node of the deterministic scale-free network has T 1 R ≈ 749.8.We also computed the GRT T R of the deterministic scale-free network using the methodology described in Section 2.4.We find T R ≈ 750.04 being very close to the maximum T 1 R value of the network.Thus, we conclude that the 'slowest' nodes of the deterministic scale-free network indeed govern its overall time scale of synchronization.However, this result cannot be generalized to any arbitrary topology as we will demonstrate in the following.

Random scale-free networks of Rössler oscillators
Next, we consider an ensemble of 100 random scale-free networks (generated using the classical Barabási-Albert (BA) model of growth and preferential attachment [41]) of N = 81 Rössler oscillators each, with the same parameter values as for the deterministic scale-free network.While generating the random scale-free networks, we explicitly model the growing character of the network by starting with a small number of vertices and at every time step introducing a new vertex and linking it to 2 vertices already present in the system (until the network comprises 81 nodes).Preferential attachment is incorporated by assuming that the probability Π i that a new node will be connected to node i depends on the degree k i of node i, such that Π i = k i j k j .The deterministic scale-free network of N = 81 Rössler oscillators studied in Section 3.1 has 130 edges, equivalently, an edge density of 130 ( ≈ 0.04.The random scale-free networks generated using the classical BA model have edge densities (similar to that of the deterministic scale-free network) of 0.049, i.e., 158 edges in each realization.Therefore, the results obtained for both topologies are not directly comparable quantitatively.The distribution of SNBS S 1 B values of the N = 81 nodes of the considered ensemble is presented in Fig. 3(a).Surprisingly, all nodes have similar and very high S 1 B values.Similar results have been observed in a recent study on SNBS values in the deterministic scale-free network of Rössler oscillators [10].These observations lead to two important conclusions.Firstly, the similar and rather high S 1  B values indicate that the synchronized state in scale-free networks is generally very robust to perturbations affecting a single node of the system.Secondly, we observe that the presence or lack of such a specific macroscopic (hierarchical) structure in the respective scale-free network does not affect the distribution of its S 1 B values markedly.In contrast to the latter finding, we have already observed an influence of the hierarchical structure on T 1 R for the deterministic scale-free network (Fig. 2(a)).On this note, we shall further unfold dependences of T 1 R values on different topological features of random scale-free networks.
The corresponding distribution of T 1 R (for δ = 10 −6 [40]) of all nodes of the considered ensemble of random scale-free networks is shown in Fig. 3(b).As in the case of the deterministic scale-free network, we next consider the mutual dependence between SNRT and the local topological characteristics of the network.For this purpose, we study the distribution of T 1 R values of all nodes of the ensemble with respect to their degree and betweenness centrality.We collect all nodes of the ensemble having a particular degree k and calculate the mean over the T 1 R values of all these nodes which corresponds to the conditional mean T 1 R | k .Similarly, we bin the bc values of all nodes of the ensemble and calculate the conditional mean R values of all nodes belonging to the respective bin.Interestingly, the conditional mean values exhibit a strong linear dependency with respect to k and bc as illustrated in Fig. 3(c, d).This is further underlined by correlation coefficients of 0.987 (0.991) of the conditional means with k (bc).Thus, nodes with high k and bc, namely the hubs in the random scale-free network, can be classified as its slow nodes.Perturbations to a more central node of a scale-free network (operating in the synchronized state) can easily spread to other nodes of the network driving them further away from the synchronized state.As a result, a scale-free network operating in synchrony may take longer to resynchronize when its more central nodes are perturbed as opposed to less central ones.This rationale is supported by the positive correlation between the conditional mean T 1 R | bc and bc.Further, given the strong linear relationship of the conditional mean SNRT with bc, a similar dependence for k is to be expected (and vice versa) since random scale-free networks generally exhibit a strong correlation between k and bc of their nodes [42].However, the relationship of the conditional mean SNRT with k and bc being specifically linear is surprising and revealing the underlying reason requires further investigation.
We now calculate and present the GRT T R of all members of the considered ensemble of random scale-free networks (Fig. 4(a), black circles).Interestingly, we observe that unlike for the deterministic scale-free network, the overall time scale of synchronization in the different network realizations of its random counterpart differs markedly from the maximum SNRT (red crosses) of the respective realization.To further study this finding, for each network realization we compute the average of the T 1 R values of all its N = 81 nodes and denote it by T 1 R .Notably, the T R value of every network realization appears closely related to T 1 R (blue crosses) as illustrated in Fig. 4(a).This is also corroborated by a correlation coefficient of 0.991 between T R and T 1 R .We further calculate and present in Fig. 4(b) the maximum betweenness centrality bc max of all nodes of each network realization and investigate its relationship with the GRT T R of the respective realization.As mentioned earlier, perturbing the node with bc max in a scale-free network (operating in the synchronized state) may lead to a particularly large relaxation time to the synchronized state.Thus, the higher the maximum betweenness centrality of a scale-free network, the higher is the GRT of the system, which is underlined by the positive correlation coefficient of 0.882 between T R and bc max in Fig. 4(b).
The average path length L of a network is defined as the mean value of the shortest path length between all possible pairs of vertices [6].Thus, L = where (i, j) is the length of the shortest path between nodes i and j of the N -node network [6].The dependence of the GRT T R of each network realization on its average path length (L) is presented in Fig. 4(c).We observe that T R exhibits a negative correlation coefficient of -0.658 with respect to L, i.e., random scale-free networks with shorter characteristic path lengths synchronize slower.This result is compatible with the fact that random scale-free networks with longer characteristic path lengths have been previously shown to promote synchronizability and vice versa [28].The underlying heuristic picture is that a small L in such networks corresponds to a large amount of traffic passing through the few 'central' nodes connected to each other which facilitate communication between the much larger population of the other oscillators.This may lead to destructive interference of the different signals passing through such nodes.Subsequently, there may not be significant overall communication between the different oscillators of the network, thereby culminating in its reduced synchronizability [28].

Power grid of the United Kingdom
As a final more realistic example, we consider a conceptual model of the power transmission grid of the United Kingdom with second-order Kuramoto-type nodal dynamics [43].The network consists of N = 120 nodes and 165 transmission lines (as illustrated in Fig. 5) with topological properties much different from those of a scale-free network.The dynamical equations of the system (in analogy with Eq. ( 1)) read [10] θi = ω i , where θ i , ω i , α and P i denote the phase, frequency, electromechanical damping constant and net power input of the i th oscillator, respectively.Furthermore, we randomly choose N 2 net generators and N 2 net consumers with P i = +P 0 and P i = −P 0 , respectively [8].We use the parameter values of α = 0.1, P 0 = 1.0 and = 9.0 for obtaining the results described below.
We again consider the synchronized state, which corresponds to all oscillators having constant phases θi and frequencies ωi = 0, as the DOS of the grid.We select I C = 1000 trials for calculating the SNRT values of the network.The  6.However, we also observe 7 slow nodes that exhibit substantially higher values ( T 1 R > 200), which are marked in red in Fig. 6.Therefore, (individually or collectively) perturbing any of these 7 nodes of the network will result in dysfunction of the grid and a significantly longer time until the system retaliates to the synchronized state.In turn, it is recommended to control or safeguard these 7 specific nodes of the network to avoid long awaiting time for the system to return to the synchronized state in the face of random perturbations.The choice of the boundary at T 1 R = 200 for distinguishing between the fast and slow nodes is motivated by the fact that we observe a first substantial gap in the histogram in Fig. 6(b) around the aforementioned value.We also find similar results from a cluster analysis of the T 1 R values of the network.These 7 nodes are not found to exhibit any specific topological features leading to their relatively higher respective T 1 R values.Further investigations analyzing these results may provide potentially important insights in this regard.
We emphasize that Erdős-Rényi random networks [6] of Rössler oscillators are found to exhibit similar distributions of T 1 R values as above; the corresponding results are described in Appendix B. Figure 6(c, d) illustrates the values of T 1 R in comparison with k and bc, respectively.The correlation coefficients of T 1 R with k and bc are 0.102 and 0.061, respectively, ruling out the existence of a systematic dependence between T 1 R and k or bc. Figure 5 displays the network topology together with the individual T 1 R values in analogy with Fig. 2 for the deterministic scale-free network of Rössler oscillators.

Conclusions
Complex systems modelled as networks of interacting dynamical units are ubiquitous and often exhibit multiple stable states.Maintaining operation of such systems in the desired stable state (which often concurs with the synchronized state of the network) is vital to their functionality.Subsequently, this has generated a lot of attention in studying stability of the desired operational state (DOS) in such coupled dynamical systems.However, given that the DOS is stable in principle, it is equally important that the system relaxes back to the same as quickly as possible, following a random perturbation to a particular node of the network.We have addressed this issue here by proposing the general framework of single-node recovery time (SNRT) which relates to the time taken by the system operating in the DOS to return to the same, following a non-infinitesimal perturbation to the dynamical state of the respective node.It is important to note that we did not address the problem of driving the perturbed system to the DOS.Instead, we aimed at unveiling the different relative time scales underlying the transient dynamics of individual nodes of the network during its relaxation to the DOS, in order to identify specific nodes which when perturbed lead to significantly enlarged RT.We thus recommend taking precautionary measures of safeguarding primarily these nodes of the network from external perturbations.
Importantly, the proposed machinery can be utilized in revealing relationships between topological features of nodes and their respective SNRT values and in turn, the global relaxation time (GRT) of the overall network.Further, we have suggested the association of SNRT with the concept of engineering resilience in quantifying the resilience of such networked dynamical systems.Finally, we have applied the framework of SNRT to deterministic and random scale-free networks of Rössler oscillators and a conceptual model of the power grid of the United Kingdom with second-order Kuramototype nodal dynamics.
We have presented here the framework of SNRT (and associated illustrations) in the special context of networks of identical oscillators with continuous time dynamics (Eq.( 1)) exhibiting bistability on account of coexisting synchronized and desynchronized regimes.However, the framework is generally applicable to any networked (continuous or discrete time) dynamical system with non-identical nodes and multiple coexisting states.Thus, future work on SNRT could comprise its extension and application to networks of non-identical nodes and/or exhibiting more complex patterns of multistability.Further development on SNRT could comprise its generalization to a framework of multiple-node recovery time, similar to recent work in the context of basin stability [10].
Regarding a potential field of application, we emphasize that time-delays arise frequently in the inherent dynamics of individual oscillators and in their interactions on complex networks [44].Therefore, another interesting endeavour could constitute incorporating time-delays in networked dynamical systems and investigating their influence on SNRT and GRT of the network.Finally, complex systems comprising oscillators coupled on prototypical network types such as Watts-Strogatz, multilayer, interdependent, etc. are open to applications of SNRT.These ventures could further unravel interesting relationships between SNRT and topological features of the aforementioned networks.

Figure 1 .
Figure 1.(Color online) Network topology of the undirected deterministic scale-free network of N = 81 identical Rössler oscillators.The size of each node is proportional to its degree and the color indicates the T 1 R value of the respective node.

Figure 2 .
Figure 2. (Color online) (a) SNRT T 1 R (on log 10 scale) of the nodes of the 3 generations of the undirected deterministic scale-free network of N identical Rössler oscillators (Eq.(8)).The first 9 nodes comprise the 1 st generation, the next 18 nodes the 2 nd generation and the final 54 nodes the 3 rd generation.Node 4 having the minimum SNRT value T 1 R (4) = 0 of the network (implying divergence of log 10 T 1 R (4) ) has not been shown in the plot.(b) Histogram of log 10 T 1 R of the nodes.(c, d) Relationship of T 1 R with (c) degree (k) and (d) betweenness centrality (bc) of the nodes.

Figure 3 .
Figure 3. (Color online) (a) Histogram of SNBS S 1 B of all nodes of the considered ensemble of random scale-free networks.(b) Same for the T 1 R values.(c, d) Conditional means (blue circles) of T 1 R with respect to (c) degree T 1 R | k and (d) betweenness centrality T 1 R | bc of the nodes.The red lines indicate linear fits to the conditional means.

Figure 4 .
Figure 4. (Color online) (a) Global RT T R (x) (black circles), maximum SNRT T 1 R (x) max (red crosses) and average SNRT T 1 R (x) (blue crosses) of all network realizations from the considered ensemble of random scale-free networks.(b) Relationship between T R (x) (blue circles) and the maximum betweenness centrality (bc max ) of all nodes of the respective network realization.(c) As in (b) for T R (x) and average path length (L) of the respective network realization.The red lines in (b, c) indicate linear fits.

Figure 5 .
Figure 5. (Color online) Network topology of the power transmission grid of the United Kingdom (comprising N = 120 nodes) with second-order Kuramoto-type nodal dynamics.Circular nodes denote net generators while square nodes are net consumers.The size of each node is proportional to its degree, and its color corresponds to the T 1 R value of the respective node.The 7 nodes further encircled by blue diamonds comprise the slow nodes of the grid in our simplified model.

Figure 6 .
Figure 6.(Color online) (a) SNRT T 1 R of all the N = 120 nodes of the power grid of the United Kingdom with second-order Kuramoto-type nodal dynamics.(b) Histogram of T 1 R of all the N = 120 nodes.(c, d) Dependence of T 1 R on (c) degree (k) and (d) betweenness centrality (bc) of the nodes.The fast nodes of the grid with T 1 R ≤ 200 are shown in black while the slow nodes having T 1 R > 200 are marked in red.

T 1 R
values (for δ = 10 −4 ) of all the N = 120 nodes are shown in Fig. 6(a) and Fig. 6(b) displays a histogram of all T 1 R values.Interestingly, we observe from Fig. 6(a, b) that 113 nodes have low values of SNRT ( T 1 R ≤ 200), which are shown in black in Fig.
1 R (i) = 0 by i ref .The resulting values of T 1 R now represent differences in time by which nodes of the network return faster or slower than the reference node i ref .As opposed to arbitrarily choosing x ref , thereby resulting in negative T RR values (which is counter-intuitive when measuring time), the above choice of x ref ensures non-negativity of T 1 R (i) values, besides eliminating the arbitrariness associated with the choice of x ref .Further details on the choice of the reference trajectory are provided in Appendix A.