1: \documentclass[twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}

2: \usepackage{graphicx}% Include figure files

3: \usepackage{dcolumn}% Align table columns on decimal point

4: \usepackage{bm}% bold math

5: 

6: \begin{document}

7: 

8: \title{Stability of complex networks under the evolution of attack and repair}

9: 

10: \author{L.P. Chi}

11: \email{chilp@iopp.ccnu.edu.cn} \affiliation{Institute of Particle

12: Physics, Hua-Zhong (Central China) Normal University, Wuhan

13: 430079, P.R. China}

14: \author{C.B. Yang}

15: \affiliation{Institute of Particle Physics, Hua-Zhong (Central

16: China) Normal University, Wuhan 430079, P.R. China}

17: \author{X. Cai}

18: \affiliation{Institute of Particle Physics, Hua-Zhong (Central

19: China) Normal University, Wuhan 430079, P.R. China}

20: 

21: \date{\today}

22: 

23: \begin{abstract}

24: 

25: With a simple attack and repair evolution model, we investigate

26: and  compare the stability of the Erd\"{o}s-R\'{e}nyi random

27: graphs (RG) and Barab\'{a}si-Albert scale-free (SF) networks. We

28: introduce a new quantity, \textit{invulnerability} $I(s)$, to

29: describe the stability of the system. We find that both RG and SF

30: networks can evolve to a stationary state. The stationary value

31: $I_{c}$ has a power-law dependence on the average degree $\langle

32: k \rangle_{rg}$ for RG networks; and an exponential relationship

33: with the repair probability $p_{sf}$ for SF networks. We also

34: discuss the topological changes of RG and SF networks between the

35: initial and stationary states. We observe that the networks in the

36: stationary state have smaller average degree $\langle k \rangle$

37: but larger clustering coefficient $C$ and stronger assortativity

38: $r$.

39: 

40: \end{abstract}

41: 

42: \pacs{89.75.Hc, 87.23.Kg, 89.75.Fb}

43: %89.75.Hc Networks and genealogical trees

44: %87.23.Kg Dynamics of evolution

45: %89.75.Fb Structures and organization in complex systems

46: 

47: \maketitle

48: 

49: \section{Introduction}

50: 

51: Complex networks are an essential part of modern society. Many

52: social~\cite{wasserman}, biological~\cite{williams},

53: transportation~\cite{Li,Chi,Latora} and communication~\cite{BA1}

54: systems can be cast into the form of complex networks, a set of

55: nodes joined together by links indicating interactions. The

56: investigation of complex networks was initiated by Erd\H{o}s and

57: R\'{e}nyi~\cite{ER1} in 1950s. They introduced a kind of simplest

58: and most straightforward complex networks, described as random

59: graphs. In a random graph, $N$ nodes are connected randomly with

60: the probability $p_{rg}$. This model has guided our thinking about

61: complex networks for decades since its introduction. In 1999

62: Barab\'{a}si and Albert~\cite{BA2,BA3,BA4} began to put the

63: emphasis on the network dynamics and constructed networks with the

64: algorithm of \textit{growth} and \textit{preferential attachment}.

65: According to the BA model, the network grows over time by the

66: addition of new nodes and links. A node newly added to the network

67: randomly selects $m$ other nodes to establish new links, with a

68: selection probability that increases with the number of links of

69: the selected node. One of the most relevant is given by the

70: scale-free nature of the degree distribution $P(k)$, defined as

71: the probability that a randomly chosen node has degree $k$ (i.e.,

72: it is connected to other $k$ nodes). In mathematical terms, the

73: scale-free property translates into a power-law function of the

74: form $P(k) \sim k^{-\gamma}$.

75: 

76: Recently the security of complex networks to the random failures

77: or intentional attacks has attracted a great deal of

78: attention~\cite{Crucitti1,Crucitti2,LiXiang}. The random failure

79: is simulated as the deletion of network nodes or links chosen at

80: random, while intentional attack as the targeted removal of a

81: specific class of nodes or links. The work by Albert \textit{et

82: al}~\cite{attack1} demonstrated that scale-free networks, at

83: variance of random graphs, are robust against random failures of

84: nodes but vulnerable to intentional attacks. The analytical

85: approach~\cite{Cohen,Callaway} has also been developed for

86: investigating the error and attack tolerance.

87: 

88: However, despite the fine work of studies on network tolerance,

89: little effort has been made on the repair of the attacked

90: networks. In this paper, we propose a simple model and study the

91: evolution of complex networks under continuous attacks and

92: repairs. This effect can be pictured for the organism: the

93: organism has to experience various DNA damages and corresponding

94: repairs during its growth. The effect can be also exemplified by

95: the Internet: the development of Internet always follows the

96: contest of virus attacks and protections. Our model is aiming at

97: investigating the stability appearing in the evolution of complex

98: systems frequently attacked by internal or environmental agents.

99: The model shows that complex networks will become more stable in

100: the evolution and eventually reach the stationary states. We also

101: discuss the topological changes of complex networks before and

102: after the evolution. It will be useful for understanding the

103: dynamics of complex systems.

104: 

105: 

106: \section{Evolution model}

107: 

108: We start by constructing a network according to the

109: Erd\"{o}s-R\'{e}nyi random graph (RG) model or Barab\'{a}si-Albert

110: scale-free (SF) model. The algorithms of RG and SF networks have

111: been mentioned in the Introduction. The dynamics of our model is

112: defined in terms of the following two operations:

113: 

114: \begin{enumerate}

115: 

116: \item \textit{Attack}: Find a node with the maximum degree $k_{max}$ and remove all

117: its links. (If several nodes happen to have the same highest

118: degree of connection, we randomly choose one of them.)

119: 

120: \item \textit{Repair}: Reconnect this node

121: with the other nodes in the network with probability $p_{re}$.

122: 

123: \end{enumerate}

124: 

125: Then, the evolution comes into the next Monte Carlo time step. The

126: random repair is considered because most of the time our operation

127: to the system is more or less blinded due to incomplete

128: information. In addition, the size of the system remains unchanged

129: during the evolution since the way of simulated attacks is the

130: deletion of links. In order to reduce the number of parameters in

131: the model, we set the repair probability $p_{re}$ the same as the

132: connection probability $p_{rg}$ in RG networks. This set also

133: implies that the information does not increase from initial

134: construction to later repairing.

135: 

136: 

137: \section{Results and discussions}

138: 

139: \subsection{Stability of complex networks}

140: 

141: We have performed extensive numerical simulations to study the

142: evolution of RG and SF networks. It can be easily seen that the

143: maximum degree $k_{max}$ has a tendency to decrease in the

144: evolution. In Fig. 1, we give a snapshot of the maximum degree

145: $k_{max}$ versus time step $s$ with $N=$ 10 000 nodes and repair

146: probability $p_{re}=0.005$ for RG and SF networks. We find that

147: the fluctuation of $k_{max}$ is very large for RG networks. For SF

148: networks, $k_{max}$ decreases very steeply at first and then

149: steadily with $s$. The behavior of scale-free networks is rooted

150: in their extremely inhomogeneity in which the removal of a few

151: highly connected nodes dramatically reduces the connections of the

152: networks.

153: 

154: \begin{figure}

155: \begin{center}

156: \includegraphics[height=6cm,width=8cm]{fig1.eps}

157: \caption{Plots of $k_{max}$ versus time step $s$ with size $N=$ 10

158: 000 and $p_{re}=0.005$ for (a) random graphs; (b) scale-free

159: networks.}

160: \end{center}

161: \label{fig1}

162: \end{figure}

163: 

164: From intuition, a node with less links to others will be attacked

165: less frequently. Thus a network with smaller maximum connection

166: degree is more stable. To describe the stability of the network,

167: we introduce a new quality, \textit{Invulnerability} $I(s)$. The

168: definition of $I(s)$ is analogous to that of the \textit{Gap}

169: $G(s)$ in the Bak-Sneppen evolution model~\cite{BS1,BS2}, which is

170: used to simulate the evolution of species. Considering an

171: evolution of network with maximum degrees $k_{max}(1)$,

172: $k_{max}(2)$,..., $k_{max}(s)$, invulnerability $I(s)$ at time $s$

173: is defined

174: 

175: \begin{equation}

176: I(s)=1/Min\{k_{\rm max}(i)\} \ \ \textbf{for\ \ \ } i\leq s\ ,

177: \end{equation}

178: 

179: \noindent i.e., the inverse of the minimum of all the maximum

180: degrees in the evolution before moment $s$. Initial value of

181: $I(s)$ is equal to $1/{k_{max}(1)}$. $I(s)$ reflects the attack

182: tolerance of the network. When $I(s)$ is small, the network is

183: vulnerable to attack and behaves less stability. Obviously from

184: definition, $I(s)$ is non-decreasing function of evolution, and

185: some fluctuations in $k_{max}$ have been filtered out.

186: 

187: \begin{figure}

188: \begin{center}

189: \includegraphics[width=8cm]{fig2.eps}

190: \caption{Plots of $I(s)$ versus step $s (million)$ with $N=$ 10

191: 000 and $p_{re}=0.005$.}

192: \end{center}

193: \label{fig2}

194: \end{figure}

195: 

196: In Fig. 2 we plot the evolution $I(s)$ versus $s (million)$ of RG

197: and SF networks under $N=$ 10 000 and $p_{re}=0.005$. We observe

198: that in both networks $I(s)$ increases very quickly at small $s$

199: but slowly at large $s$ and finally reaches a stationary value

200: $I_{c}$ when $s$ is large enough. The increase of $I(s)$ indicates

201: that the system is getting more and more stable in general by

202: experiencing continuous attacks and repairs. This figure is

203: similar to the envelope function of Bak-Sneppen evolution model.

204: Without interference from outside world, the network evolves to a

205: stationary state, and the process takes place over a very long

206: transient period.

207: 

208: 

209: \subsection{Stationary states of random graphs}

210: 

211: In this section we will present the properties of stationary value

212: $I_c^{rg}$ in the stationary state for RG networks. In RG

213: networks, the average degree $\langle k \rangle_{rg}$ is

214: determined by the connection probability $p_{rg}$ and the size

215: $N$, $\langle k \rangle_{rg} = p_{rg}N$. With $\langle k

216: \rangle_{rg} = 20$ fixed and $p_{re}=p_{rg}$, we find that

217: $I_{c}^{rg}$ stays unchanged at 0.2 with size $N$ ranging from 100

218: to 10 000. This result is interesting, because it shows that the

219: stationary value $I_{c}^{rg}$ depends not on the network size $N$,

220: but the average degree $\langle k \rangle_{rg}$.

221: 

222: \begin{figure}

223: \begin{center}

224: \includegraphics[width=6cm]{fig3.eps}

225: \caption{Plot of $I_{c}^{rg}$ versus $\langle k \rangle_{rg}$ with

226: size $N$ ranging from 100 to 10 000.}

227: \end{center}

228: \label{fig3}

229: \end{figure}

230: 

231: To get the relationship between $I_{c}^{rg}$ and $\langle k

232: \rangle_{rg}$, we show in Fig. 3 $I_{c}^{rg}$ as a function of

233: $\langle k \rangle_{rg}$ in a log-log plot. We find that

234: $I_{c}^{rg}$ has a power-law dependence on the average degree

235: $\langle k \rangle_{rg}$,

236: 

237: \begin{equation}

238: I_{c}^{rg}(\langle k \rangle_{rg}) \ \propto \  \langle k

239: \rangle_{rg}^{-\tau},

240: \end{equation}

241: 

242: 

243: \noindent where the exponent $\tau$ is about 1.5. Fig. 3

244: illustrates that after the network has relaxed to the stationary

245: state, the stability of the network will increase rapidly with the

246: decrease of average degree $\langle k \rangle_{rg}$. Thus, when

247: the $\langle k \rangle_{rg}$ is small, i.e., less communications

248: and interactions in the network, the system will be more stable.

249: 

250: \subsection{Stationary states of scale-free networks}

251: 

252: The properties of stationary value $I_c^{sf}$ in the stationary

253: state for SF networks will be illustrated in this section. In Fig.

254: 4, we plot $I_{c}^{sf}$ as a function of repair probability

255: $p_{sf}$ in a semi-log coordinate with $N=$ 1000 (circles); 5000

256: (squares) and 10 000 (triangles). We find that $I_c^{sf}$

257: decreases with the increase of $N$ or $p_{sf}$. While large $N$ or

258: $p_{sf}$ means the large number of links in networks. In other

259: words, the stability of the system relies on the internal

260: interactions within the system, that is, the stronger the

261: interactions, the less stable the system. From Fig. 4 we can also

262: observe that in this semi-log plot $I_{c}^{sf}$ has an

263: approximately linear decrease with $p_{sf}$ under fixed $N$, i.e.,

264: an exponential relationship between $I_c^{sf}$ and $p_{sf}$,

265: 

266: \begin{equation}

267: I_{c}^{sf}(p_{sf}) \ \sim \ \exp({-a(N)p_{sf}+b(N)}),

268: \end{equation}

269: 

270: \noindent where $a(N)$ and $b(N)$ are two fitting parameters

271: related to the size $N$.

272: 

273: \begin{figure}

274: \begin{center}

275: \includegraphics[width=6cm]{fig4.eps}

276: \caption{Plot of $I_{c}^{sf}$ versus repair probability $p_{sf}$

277: with size $N=$ 1000 (circles); 5000 (squares); and 10 000

278: (triangles).}

279: \end{center}

280: \label{fig4}

281: \end{figure}

282: 

283: 

284: \subsection{Topological changes of complex networks}

285: 

286: In order to offer further information of complex networks in the

287: stationary states, we compare some structural properties of the

288: networks in the initial states with those in the stationary

289: states. In Table~\ref{comparison}, we display the topological

290: changes of RG and SF networks under $N=$ 10 000 and

291: $p_{re}=0.005$. We use the superscript $i$ and $s$ to distinguish

292: the initial and stationary states.

293: 

294: \begin{table}

295: \caption{\em Comparison of topological changes in initial and

296: stationary states between RG and SF networks under $N=$ 10 000 and

297: $p_{re}=0.005$.}

298: \begin{center}

299: \begin{tabular}{c|c|cc|cc|cc|cc} \hline\hline

300: 

301:      &$I_{c}$ &$\langle k \rangle^i$&$\langle k \rangle^s$&$L^i$&$L^s$&$C^i$&$C^s$&$r^i$&$r^s$ \\

302: \hline

303: RG   & 0.037 &50.16 &19.34 &2.77 &3.68 &0.005  &0.57  &0.00041 &0.35 \\

304: \hline

305: SF   & 0.042 &4.0   &3.35  &5.09 &7.17 &0.0039 &0.014 &-0.049  &0.39 \\

306: \hline

307: 

308: \end{tabular}

309: \end{center}

310: \label{comparison}

311: \end{table}

312: 

313: \textit{Average degree}. We find that the average degree in

314: stationary states $\langle k \rangle ^ s$ decreases dramatically

315: for RG networks but slightly for SF networks. This result is

316: rooted in the homogeneity of RG networks. Since in RG networks all

317: nodes have approximately the same number of links,  majority of

318: nodes in the network will be impacted by the attacks and, hence,

319: the rapid decrease of $\langle k \rangle$.

320: 

321: \textit{Characteristic path length}. The characteristic path

322: length $L$ is the average length of the shortest paths between any

323: two nodes in the network. $L^s$ in the stationary state is a

324: little larger than $L^i $, which exhibits that it will take a

325: little longer for the nodes in the network to communicate with

326: each other after the evolution.

327: 

328: \textit{Clustering coefficient}. The clustering coefficient of

329: node $i$ is defined as the existing numbers $n_{i}$ among the

330: links of node $i$ over all the possible links, that is,

331: $C_{i}=n_{i}/[k_{i}(k_{i}-1)/2]$. The clustering coefficient of

332: the whole network is $C=\frac{1}{N} \sum C_{i}$. $C^s$ for RG and

333: SF networks are increased by $99\%$ and $72\%$, compared with

334: $C^i$ in the initial states. It suggests that the nodes are highly

335: clustered in the stationary states.

336: 

337: \textit{Assortativity}. The assortativity $r$ in the range $-1

338: \leq r \leq 1$ is another interesting feature of complex networks.

339: A network is said to show assortative, $r>0$, if the high-degree

340: nodes in the network preferentially connect to other high-degree

341: nodes. By contrast, a network is disassortative, $r<0$, if the

342: high-degree nodes tend to connect to other low-degree nodes. The

343: values of $r^i$ and $r^s$ have shown that both RG and SF networks

344: turn to strongly assortative networks in the stationary states

345: from initial weakly assortative or even disassortative ones.

346: 

347: \textit{Degree distribution}. Fig. 4(a) and 4(b) display the

348: degree distributions $P(k)$ of RG and SF networks in the initial

349: and stationary states under $N=$ 10 000 and $p_{re}=0.005$. For RG

350: networks, the degree distribution $P(k)_{rg}^s$ in the stationary

351: state is a Gaussian one with mean value $\langle k \rangle_{rg}^s

352: \simeq 19$ and standard deviation $\sigma_{rg}^s \simeq 7$, while

353: the $P(k)_{rg}^i$ in the initial state is a Poisson distribution

354: with $\langle k \rangle_{rg}^i \simeq 50$ and

355: $\sigma_{rg}^i=\sqrt{\langle k \rangle_{rg}^i} \simeq \sqrt{50}$.

356: For SF networks, $P(k)_{sf}^i$ and $P(k)_{sf}^s$ are both

357: power-laws, $P(k)_{sf} \sim k_{sf}^ {-\gamma}$, with the similar

358: exponent about 2.4. The differences between $P(k)_{sf}^i$ and

359: $P(k)_{sf}^s$ are that the fat tail in $P(k)_{sf}^i$ vanishes; and

360: the maximum degree decreases from the initial value of 241 to 50

361: in stationary state. Besides, Fig. 4(a) and 4(b) show that the

362: shape of the degree distributions is almost the same for RG and SF

363: networks in initial and stationary states. It suggests that the

364: general structures of the systems do not change even after a long

365: time evolution to the stationary states.

366: 

367: \begin{figure}

368: \begin{center}

369: \includegraphics[height=6cm,width=8cm]{fig5.eps}

370: \caption{A comparison of degree distributions under $N=$ 10 000

371: and $p_{re}=0.005$ in initial (circles) and stationary (asterisks)

372: states for (a) random graphs; and (b) scale-free networks . Solid

373: lines are the fit curves. }

374: \end{center}

375: \label{fig5}

376: \end{figure}

377: 

378: In general, the system in the stationary state has smaller average

379: degree $\langle k \rangle$ but larger clustering coefficient $ C$

380: and stronger assortativity $r$. It implies that the system after

381: undergoing a long period of evolution is composed of small highly

382: clustered clusterings, which well fits the natural world. In

383: ecosystems, agents autonomously form communities, in which

384: individuals are highly correlated. Commercial airlines, for

385: reasons of efficiency, prefer to have a small number of hubs where

386: all routes would connect.

387: 

388: 

389: \section{Conclusions}

390: 

391: In this paper we study the stability of random graphs and

392: scale-free networks with a simple attack and repair evolution

393: model. We introduce a new quality, \textit{invulnerability}

394: $I(s)$, to describe the stability of the networks. A stationary

395: state with fixed $I(s)$ is observed during the evolution of the

396: network. The stationary value of invulnerability $I_{c}$ is found

397: to have a power-law dependence on $\langle k \rangle_{rg}$ for RG

398: networks; and an exponential relationship with repair probability

399: $p_{sf}$ for SF networks. We give further information on the

400: evolution of the networks by comparing the topological changes

401: between the initial and stationary states. We conclude that the

402: networks in the stationary states are composed of small highly

403: clustered clusterings, which well fits the evolution of complex

404: systems in natural world.

405: 

406: Still, there are many issues to be addressed, such as the

407: correlations and the fluctuations in $k_{max}$ in the evolution,

408: especially after the stationary state. These fluctuations may tell

409: us more about the nature of the stationary state. The behavior of

410: the average degree in the evolution also may shed some light on

411: the dynamics of the networks. In addition, it is worthwhile to

412: investigate whether the evolution of networks demonstrate

413: self-organized criticality (SOC), according to the similarity

414: between the evolution of invulnerability $I(s)$ with that of the

415: envelope function $G(s)$ in Bak-Sneppen model. All these topics

416: can not be covered in this paper, and will be discussed later.

417: 

418: \section*{Acknowledgments}

419: 

420: This work was supported in part by the National Natural Science

421: Foundation of China under grant No. 70271067 and by the Ministry

422: of Education of China under grant No. 03113.

423: 

424: 

425: \begin{thebibliography}{99}

426: \bibitem{wasserman}

427: S. Wasserman and K. Faust, \textit{Social Network Analysis},

428: (Cambridge University Press, Cambridge, England, 1994).

429: 

430: \bibitem{williams}

431: R. J. Williams and N. D. Martinez, Nature \textbf{404} (2000) 180.

432: 

433: \bibitem{Li}

434: W. Li and X. Cai, Phys. Rev. E \textbf{69} (2004) 046106.

435: 

436: \bibitem{Chi}

437: L. P. Chi, R. Wang, H. Su, \textit{et al}, Chin. Phys. Lett.

438: \textbf{20} (2003) 1393.

439: 

440: \bibitem{Latora}

441: V. Latora and M. Marchiori, Physica A \textbf{314} (2002) 109.

442: 

443: \bibitem{BA1}

444: R. Albert, H. Jeong and A.-L. Barab\'{a}si, Nature \textbf{401}

445: (1999) 130.

446: 

447: \bibitem{ER1}

448: P. Erd\H{o}s and A. R\'{e}nyi, Publ. Math. Inst. Hung. Acad. Sci.

449: \textbf{5} (1960) 17.

450: 

451: \bibitem{BA2}

452: A.-L. Barab\'{a}si and R. Albert, Science \textbf{286} (1999) 509.

453: 

454: \bibitem{BA3}

455: A.-L. Barab\'{a}si, R. Albert, and H. Jeong, Physica A

456: \textbf{281} (2000) 69.

457: 

458: \bibitem{BA4}

459: R. Albert and A.-L. Barab\'{a}si, Rev. Mod. Phys. \textbf{74}

460: (2002) 47.

461: 

462: \bibitem{Crucitti1}

463: P. Crucitti, V. Latora, M. Marchiori and A. Rapisarda, Physica A

464: \textbf{340} (2004) 388.

465: 

466: \bibitem{Crucitti2}

467: P. Crucitti, V. Latora, M. Marchiori and A. Rapisarda, Physica A

468: \textbf{320} (2003) 622.

469: 

470: \bibitem{LiXiang}

471: X. Li and G. Chen, Physica A \textbf{328} (2003) 274.

472: 

473: \bibitem{attack1}

474: R. Albert, H. Jeong and A.-L. Barab\'{a}si, Nature \textbf{406}

475: (2000) 378.

476: 

477: \bibitem{Cohen}

478: R. Cohen, K. Erez, D. ben-Avraham and S. Havlin, Phys. Rev. Lett.

479: \textbf{86} (2000) 4626.

480: 

481: \bibitem{Callaway}

482: D.S. Callaway, M.E.J. Newman, S.H. Strogatz and D.J. Watts, Phys.

483: Rev. Lett. \textbf{85} (2000) 5468.

484: 

485: \bibitem{BS1}

486: P. Bak and K. Sneppen, Phys. Rev. Lett. \textbf{71} (1993) 4083.

487: 

488: \bibitem{BS2}

489: M. Paczuski, S. Maslov and P. Bak, Phys. Rev. E \textbf{53} (1996)

490: 414.

491: 

492: \end{thebibliography}

493: 

494: 

495: \end{document}

496: