arXiv:1405.4342v1 [physics.soc-ph] 17 May 2014 outeso nedpnetntok.Sc conceptual cou- Such the on [9], networks. [12] interdependent pattern structures of network coupling robustness and of 11] [10, effect studies strength the deeper pling theory, on percolation focused the have random on to Based vulnerable failure. that even so are networks networks of interdependent vulnerability sig- networks the between increases dependency nificantly such al. et that of Buldyrev in showed failure ones accordingly. has collapse counterpart the to the words, network trigger other other will the network In one [4]. in to have nodes other subsystems each different on two which depend robustness in interdependent the systems Recently, explore the to of formalized 8]. been have 7, real-world networks [4, capture further systems models to networked network established coupled thus been Some with have significantly. interact other infrastructures each networked Ac- networks. various isolated of tually, framework the on based derived is distribution degree are homogenous. its example, although for robust-yet-fragile networks, degree also Small-world than rather [6]. distribution distribution heterogeneity betweenness the indicated network cascades with further of associated in al. principally property et is (RYF) Xia scenario robust-yet-fragile [5]. the both re- that to been robust random the is has whereas graph attacks, It fail- intentional to random fragile to 6]. but robust ure [5, are networks explored scale-free networks [2– that complex widely vealed issue on been cascades central thus decade, the have past undoubtedly the Over been 4]. robustness has The systems [1]. commu- of systems grids, transportation power and nication as such infrastructures networked ∗ [email protected] h frmnindcnlsos oee,aemainly are however, conclusions, aforementioned The critical many by supported society modern a in live We eateto nomto cec n lcrncEngineer Electronic and Science Information of Department fterbsns fitreedn ewrssignificantly networks interdependent of attack. intentional robustness under the fragile of but netw failure robust-ye scale-free random are BA under Interdependent graphs random attack. intentional ER rando and interdependent both under that networks found scale-free inves is BA by and question graphs this random address we ER paper, couplin this s how In l networked unexplored. of interdependent is real-world networks understanding the interdependent of Our in optimization loads of failure. failures random cascading to are they oe.Prhn t lfrhridctdta h oeinter more the that indicated further al et. Parshani model. ASnumber(s): PACS DOI: nedpnetntok aebe hw ob xrml vul extremely be to shown been have networks Interdependent .INTRODUCTION I. h outytfaientr fitreedn networks interdependent of nature robust-yet-fragile The 97.c 97.b 51.a 89.40.-a 05.10.-a, 89.75.Fb, 89.75.Hc, e a n ogin Xia Yongxiang and Tan Fei Dtd ue8 2021) 8, June (Dated: nuh nti ae,w iltyt nwrti ques- this mechanisms. answer strategies coupling to attack the try structures, and network will considering we large by always paper, tion this loads relation In dependency of the enough? cascades by against undermined robustness is network extent which the in is to loads Nevertheless, of [14]. such failures networks interdependent cascading Undoubtedly, aggravate will networks. dependency interdependent on cades coupling found. and/or be patten could connectivity coupling probability interconnected the optimal in as loads the of role networks, cascades same For the networks. within as for play networks parts across two links two 13]. interconnected between the [7, networks, links networks interdependent cou- dependency isolated the from originally capture Different two to between developed interconnected pling been interaction, also of have form networks another more networks As real-world inter- understand readily. the to towards us isolation allows action the from shift breakthrough iestontok n ihtesm ie( size same the with B and N A networks two siders opigpten r hsdsrbdbsdo oddis- load on based described three thus failures, cascading are focus to patterns im- we leads coupling As overload great networks. traffic has coupled how of established on robustness are in the way on links pact the coupling studies, two the previous between in which relationships mentioned Dif- dependency As the links networks. link. to coupling coupled, interdependent refer the fully networks, one just to interconnected only assumed from has ferent are node B each and and A networks Here A n,Zein nvriy aghu302,China 310027, Hangzhou University, Zhejiang ing, eety rffi vrodhsbe xeddt h cas- the to extended been has overload traffic Recently, o ipiiyadcaiyo h eut,ormdlcon- model our results, the of clarity and simplicity For = . ses hsqeto,hwvr slargely is and however, design question, the This for ystems. essential is but imited, rs oee,aeol robust-yet-fragile only are however, orks, hs eut dac u understanding our advance results These iaigterbsns finterdependent of robustness the tigating N smlrntok r,temr robust more the are, networks -similar B n aeaeaedegree( average same and ) -rgl ne ohrno failures random both under t-fragile alr n netoa tak It attack. intentional and failure m ∗ atrssaeadipc the impact and shape patterns g eal ae ntepercolation the on based nerable .Ntokmodel Network A. I MODELS II. h k i = h k A i = h k N B i = ). 2 tribution in individual networks [13]. When one node is removed from the network (due to a failure or attack), it will trigger two kinds of effects. • Assortative Coupling. Nodes are first sorted in net- On the one hand, the removal affects the shortest paths works A and B respectively, both in the descend- between some other nodes within the attacked network ing order of load. If different nodes share the same (e.g., network A). Then the packets from these nodes load, we sort them at random. Connect the first have to adjust their paths. In this way, loads of many node in network A with the first node in network nodes are changed. Those nodes with loads greater than B, and then connect the second node in network their capacities will fail, then it triggers further load ad- A with the second node in network B, and so on. justment over and over until loads of all remaining nodes Repeat this process until all interdependent links are less than their capacities. It is obvious that the tol- are built. erance parameter α affects the effect of cascading fail- ures. On the other hand, the removed node in network • Disassortative Coupling. Nodes are first sorted in A will cause the dependent counterpart node in network network A (B) in the descending (ascending) order B to fail. The failure of this node will cause the same of load. If different nodes share the same load, we cascades within network B. Actually, these two effects sort them at random. Connect the first node in interact with each other in cascading failures. Note that network A (with the heaviest load) with the first nodes outside the giant component are assumed to be node in network B (with the lightest load), and failed. Therefore, the dependency relationship makes the then connect the second node in network A with cascading failure process more complicated compared to the second node in network B, and so on. Repeat isolated networks. this process until all interdependent links are built. In this paper, two attack strategies will be explored, • Random Coupling. Randomly choose a node in i.e., random failure and intentional attack. We use the network A and a node in network B. If neither relative size of the giant component G to characterize of them has an interdependent link, then connect the robustness of networks. In our model, the remaining them. Repeat this process until all interdependent giant component after cascading failures in two subnet- links are built. works share the same size. Thus, G is the ratio between the remaining giant component size and original network In order to maintain tractability and facilitate com- size. parison with isolated complex networks, we presume that networks A and B share the same type of networks. In this paper, we mainly focus on the robustness of coupled III. RESULTS ER random graphs and coupled BA scale-free networks. In this section, we will carry out numerical simulations on coupled ER random graphs and coupled BA scale-free B. Traffic model networks. We mainly explore four kinds of situations, i.e., (1) coupled ER random graphs under random failure, (2) In this paper, we adopt the data-packet transport sce- coupled ER random graphs under intentional attack, (3) nario based on the shortest-path routing [5, 13]. The coupled BA scale-free networks under random failure, (4) node betweenness can thus approximate the traffic load. coupled BA scale-free networks under intentional attack. To be concrete, the betweenness of node k is denoted as We first investigate the robustness of interdependent [15] ER random graphs under random failure. As ER random

k graph is robust to random failure, and BA scale-free net- nst work is fragile to intentional attack, we view them as the Bk = X , (1) gst baselines. First, we find that the dependency relationship s6=k6=t between two networks reduces the network robustness. where gst is the total number of possible shortest topo- As shown in fig. 1, under random failure, the robustness k logical paths from node s to node t and nst denotes the of interdependent networks under three coupling mecha- k number of such paths running through node k. nst/gst nisms is worse than that of the single networks. Second, k is stipulated to be zero, if nst = gst = 0. the coupling patterns influence the robustness of inter- The capacity of a node is the maximum load that the dependent networks remarkably. It is shown that the node can handle. The capacity of node k is thus set to assortative > the random coupling > the disassortative be proportional to its initial load Lk [5, 6, 16] coupling, in terms of network robustness. Note that the robustness of interdependent random graphs with assor- Ci =(1+ α)Lk, (2) tative coupling approaches that of the single ER random graph under random failure. Interdependent ER ran- where the constant α is the tolerance parameter. The dom graphs with the assortative coupling can thus be network is in the free-flow state (without overload) if α ≥ regarded as robust. Nevertheless, it is obvious that the 0. disassortative coupling is fragile to random failure. They 3

1 1

0.8 0.8

0.6 0.6 G G ER_RF ER_IA 0.4 0.4 BA_IA BA_IA AS_RF AS_IA 0.2 RD_RF 0.2 RD_IA DIS_RF DIS_IA

0 0 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 α α

FIG. 1. (Color online) Cascading failures of coupled ER FIG. 2. (Color online) Cascading failures of coupled ER random graphs under random failure. The network size random graphs under intentional attack. The network size NA = NB = 1000 and average degree hkAi = hkB i = 6. The NA = NB = 1000 and average degree hkAi = hkB i = 6. The relative size of the giant component G is shown as a function relative size of the giant component G is shown as a function of the tolerance parameter α with different types of coupling of the tolerance parameter α with different types of coupling patterns, namely, assortative coupling (), random coupling patterns, namely, assortative coupling (), random coupling (N) and disassortative coupling (∗). Cascading failures of the (N) and disassortative coupling (∗). Cascading failures of the single ER random graph under random failure () and the single ER random graph () and BA scale-free network (•) single BA scale-free network under intentional attack (•). under intentional attack.

pling can thus be regarded as robust. Nevertheless, it are even more vulnerable than the single BA scale-free is obvious that the disassortative coupling is fragile to networks under intentional attack. Consider, for exam- intentional attack. They are even more vulnerable than ple, the tolerance α = 0.25 in which the relative size of the single BA scale-free networks under intentional at- the giant component G< 0.2. When α reaches 0.4, G is tack. Consider, for example, the tolerance α = 0.2 in still less than 0.9. Altogether, when random failure oc- which coupled networks are still fragmented. When α curs, interdependent ER random graphs are robust with reaches 0.4, G is still less than 0.9. Thus, similarly to the assortative coupling but fragile with disassortative random failure, interdependent ER random graphs re- coupling. mains robust-yet-fragile under intentional attack. In situation (2), the robustness of interdependent ER Take the above two situations together, under both random graphs under intentional attack is explored. As random failure and intentional attack, interdependent ER random graph is robust to intentional attack, and BA ER random graphs are robust for the assortative cou- scale-free network is fragile to intentional attack, we view pling but fragile for the disassortative coupling. Thus, them as the baselines. First, we find that the dependency interdependent ER random graphs are robust-yet-fragile. relationship between two networks reduces the network We further explore the robustness of interdependent robustness. As shown in fig. 2, under intentional attack, BA scale-free networks under random failures. As BA the robustness of interdependent ER random graphs un- scale-free network is robust to random failure but frag- der three coupling mechanisms is worse than that of the ile to intentional attack, we view them as the baselines. single ER random graph. Second, the coupling patterns First, we find that the dependency relationship between influence the robustness of interdependent networks re- two networks reduces the network robustness. As shown markably. It is shown that the assortative > the random in fig. 3, under random failure, the robustness of inter- coupling > the disassortative coupling, in terms of net- dependent BA scale-free networks under three coupling work robustness. Note that the robustness of the inter- mechanisms is worse than that of the single BA scale-free dependent network with assortative coupling approaches network. Second, the coupling patterns influence the ro- that of the single network under intentional attack. Inter- bustness of interdependent networks remarkably. It is dependent ER random graphs with the assortative cou- shown that the assortative > the random coupling > the 4

1 1

0.8 0.8

0.6 0.6 BA_RF BA_RF G BA_IA G BA_IA 0.4 0.4 AS_RF AS_IA RD_RF RD_IA DIS_IA 0.2 DIS_RF 0.2

0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 α α

FIG. 3. (Color online) Cascading failures of coupled BA FIG. 4. (Color online) Cascading failures of coupled BA scale- scale-free networks under random failure. The network size free networks under intentional attack. The network size NA = NB = 1000 and average degree hkAi = hkB i = 6. The NA = NB = 1000 and average degree hkAi = hkB i = 6. The relative size of the giant component G is shown as a function relative size of the giant component G is shown as a function of the tolerance parameter α with different types of coupling of the tolerance parameter α with different types of coupling patterns, namely, assortative coupling (), random coupling patterns, namely, assortative coupling (), random coupling (N) and disassortative coupling (∗). Cascading failures of the (N) and disassortative coupling (∗). Cascading failures of the single BA scale-free network under random failure () and single BA scale-free network under random failure () and intentional attack (•). intentional attack (•).

TABLE I. The robustness of different network models under TABLE III. The robustness of Interdependent scale-free net- random failure and intentional attack. works with different coupling patterns under random failure ❳❳ ❳❳ Attack and intentional attack. ❳❳ Random Failure Intentional Attack ❳❳ Network ❳❳❳ ❳❳ Attack ❳❳❳ Random Failure Intentional Attack ER Robust Robust Coupling ❳❳ BA Robust Fragile Assortative Robust Fragile Disassortative Fragile Fragile

TABLE II. The robustness of Interdependent ER networks with different coupling patterns under random failure and in- explore the robustness of interdependent scale-free net- tentional attack. ❳❳ works under intentional attack. First, we find that the ❳❳❳ Attack dependency relationship between two networks reduces ❳❳ Random Failure Intentional Attack Coupling ❳❳ the network robustness. As shown in fig. 4, under inten- Assortative Robust Robust tional attack, the robustness of interdependent BA scale- Disassortative Fragile Fragile free networks under three coupling mechanisms is worse than that of the single BA scale-free network. Second, the coupling patterns influence the robustness of inter- disassortative coupling, in terms of network robustness. dependent networks remarkably. It is shown that the Note that the robustness of interdependent BA scale-free assortative > the random coupling > the disassortative networks with assortative coupling approaches that of the coupling, in terms of network robustness. Take the toler- single BA scale-free network under random failure. The ance parameter α = 0.7 for example, the relative size of G − α curve of the disassortative coupling, however, is the giant component of the assortative coupling reaches close to that of the single BA scale-free network under 0.8, whereas it is about zero for the assortative coupling. intentional network. Interdependent BA scale-free net- Interdependent scale-free networks can thus be regarded works can thus be regarded as robust-yet-fragile. as fragile. Similarly to the baselines of situation (3), we lastly Collecting these four situations (as shown in the Ta- 5 bles I, II and III), we find that interdependent ER ran- failure and intentional attack, they are robust for the as- dom graphs are robust-yet-fragile under both random sortative coupling but fragile for the disassortative cou- failure and intentional attack. Interdependent BA scale- pling. For interdependent BA scale-free networks, under free networks are robust-yet-fragile under random failure random failure, they are robust for the assortative cou- but fragile under intentional attack. Note that these re- pling but fragile for the disassortative coupling, whereas sults are different from the extreme vulnerability of in- under intentional attack, they are fragile for both. The- terdependent networks based on the percolation theory. oretically, the robust-yet-fragile property can refresh our understanding of the robustness of interdependent net- works. Furthermore, our results can provide useful in- sights for the design and optimization of real-world in- IV. CONCLUSIONS terdependent networked systems.

In this paper, we have explored cascades of loads in in- terdependent networks. The interplay of traffic overload ACKNOWLEDGMENTS and dependency makes the robustness of interdependent networks more complicated than the single network. For This work was supported by the National Natural Sci- interdependent ER random graphs under both random ence Foundation of China under Grant No. 61174153.

[1] L. Cui, S. Kumara, and R. Albert, Circuits and Systems [10] G. Dong, J. Gao, L. Tian, R. Du, and Y. He, Phys. Rev. Magazine, IEEE 10, 10 (2010). E 85, 016112 (2012). [2] R. Albert, H. Jeong, and A.-L. Barab´asi, Nature 406, [11] R. Parshani, S. V. Buldyrev, and S. Havlin, Phys. Rev. 378 (2000). Lett. 105, 48701 (2010). [3] D. J. Watts, Proc. Natl. Acad. Sci. USA 99, 5766 (2002). [12] J. Gao, S. V. Buldyrev, S. Havlin, and H. E. Stanley, [4] S. V. Buldyrev, R. Parshani, G. Paul, H. E. Stanley, and Phys. Rev. Lett. 107, 195701 (2011). S. Havlin, Nature 464, 1025 (2010). [13] F. Tan, Y. Xia, W. Zhang, and X. Jin, EPL 102, 28009 [5] A. E. Motter and Y. C. Lai, Phys. Rev. E 66, 065102 (2013). (2002). [14] P. Zhang, B. Cheng, Z. Zhao, D. Li, G. Lu, Y. Wang, [6] Y. Xia, J. Fan, and D. Hill, Physica A 389, 1281 (2010). and J. Xiao, EPL 103, 68005 (2013). [7] C. D. Brummitt, R. M. DSouza, and E. A. Leicht, Proc. [15] L. C. Freeman, Sociometry 40, 35 (1977). Natl. Acad. Sci. USA 109, E680 (2012). [16] J. W. Wang and L. L. Rong, Safety Science 47, 1332 [8] J. Gao, S. V. Buldyrev, H. E. Stanley, and S. Havlin, (2009). Nature Physics 8, 40 (2011). [9] R. Parshani, C. Rozenblat, D. Ietri, C. Ducruet, and S. Havlin, EPL 92, 68002 (2010).