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

2: 

3: \usepackage{graphicx}% Include figure files

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

5: \usepackage{bm}% bold math

6: 

7: %\nofiles

8: 

9: \begin{document}

10: 

11: \title{Detrended fluctuation analysis on the correlations of complex networks \\

12: under attack and repair strategy}

13: 

14: \author{L.P. Chi$^\dag$, C.B. Yang, K. Ma, and X. Cai}

15: \affiliation{Institute of Particle Physics,

16: HuaZhong (Central China) Normal University, Wuhan 430079, China \\

17: $^\dag$Email of the Corresponding Author: chilp@iopp.ccnu.edu.cn}

18: 

19: \date{\today}

20: 

21: \begin{abstract}

22: 

23: We analyze the correlation properties of the Erd\"{o}s-R\'{e}nyi

24: random graph (RG) and the Barab\'{a}si-Albert scale-free network

25: (SF) under the attack and repair strategy with detrended

26: fluctuation analysis (DFA). The maximum degree $k_{max}$,

27: representing the local property of the system, shows similar

28: scaling behaviors for random graphs and scale-free networks. The

29: fluctuations are quite random at short time scales but display

30: strong anticorrelation at longer time scales under the same system

31: size $N$ and different repair probability $p_{re}$. The average

32: degree $\langle k \rangle$, revealing the statistical property of

33: the system, exhibits completely different scaling behaviors for

34: random graphs and scale-free networks. Random graphs display

35: long-range power-law correlations. Scale-free networks are

36: uncorrelated at short time scales; while anticorrelated at longer

37: time scales and the anticorrelation becoming stronger with the

38: increase of $p_{re}$.

39: 

40: \end{abstract}

41: 

42: \pacs{05.40.-a, 05.45.Tp, 89.75.-k}

43: %05.45.Tp  Time series analysis

44: %05.40.-a  Fluctuation phenomena, random processes, noise, and Brownian motion

45: %89.75.-k  Complex systems

46: 

47: \keywords{Correlations, Detrended fluctuation analysis, Complex

48: networks}

49: 

50: \maketitle

51: 

52: \section{Introduction}

53: 

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

55: social, biological, transportation and communication systems can

56: be cast into the form of complex networks, a set of nodes joined

57: together by links indicating interactions. \cite{s1,s2,s3,s4}

58: Recently enormous interest has been devoted to the study of

59: complex networks. Among various existing models of complex

60: networks, two different model topologies, the Erd\"{o}s-R\'{e}nyi

61: model \cite{s5} of the random graph (RG) and the

62: Barab\'{a}si-Albert model \cite{s6} of the scale-free network

63: (SF), have been widely studied. \cite{s7,s8,s9}

64: 

65: The random graph is constructed starting from an initial condition

66: of $N$ nodes and no edges. Then edges are added between pairs of

67: randomly selected nodes with connection probability $p_{rg}$. The

68: number of edges connected to any particular node is called the

69: degree $k$ of that node. The average degree of the random graph is

70: $\langle k \rangle =  Np_{rg}$. The scale-free network puts the

71: emphasis on the network dynamics and is constructed with the

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

73: According to the scale-free model, the network grows over time by

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

75: network randomly selects $m$ other nodes to establish new links,

76: with a selection probability that increases with the number of

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

78: power-law degree distribution $P(k)$, defined as the probability

79: that a randomly chosen node has degree $k$, $P(k) \sim

80: k^{-\gamma}$.

81: 

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

83: or intentional attacks has attracted a great deal of attention.

84: \cite{s10,s11,s12,s13,s14,s15} The random failure is simulated as

85: the deletion of network nodes or links chosen at random, while

86: intentional attack as the targeted removal of a specific class of

87: nodes or links. In order to protect existing networks and design

88: robust networks, we introduced the attack and repair strategy in

89: Ref. [16], which will be described briefly in the following

90: section. In this paper, we will study the correlations for random

91: graphs and scale-free networks under the attack and repair

92: strategy with the detrended fluctuation analysis (DFA)

93: \cite{s17,s18} and investigate the intrinsic dynamics in both

94: networks. The advantages of DFA over many methods are that it

95: permits the detection of long-range correlations embedded in

96: seemingly nonstationary time series, and also avoids the spurious

97: detection of apparent long-range correlations that are an artifact

98: of nonstationary. \cite{s19,s20,s21,s22}

99: 

100: The paper is organized as follows. In Section II we review the

101: algorithm of the DFA method. In Section III, we apply the DFA

102: method to investigate the correlations of the maximum degree

103: $k_{max}$ and the average degree $\langle k \rangle$ for random

104: graphs and scale-free networks. In last section we present our

105: brief conclusions.

106: 

107: 

108: \section{DFA}

109: 

110: Detrended fluctuation analysis (DFA) is a well-established method

111: for determining the scaling behavior of noisy data in the presence

112: of trends without knowing their origin and shape. Briefly, the DFA

113: method involves the following steps.

114: 

115: \begin{enumerate}

116: 

117: \item[(1)]

118: We consider a noisy time series $u(i)$ ($i=1,..,N$) and $N$ is the

119: length of the series. We integrate the time series $u(i)$ and

120: subtract the mean $\langle u \rangle$,

121: 

122: \begin{equation}

123: y(k) = \sum\limits_{i=1}^{k}[u(i) - \langle u \rangle],

124: \end{equation}

125: 

126: where

127: 

128: \begin{equation}

129: \langle u\rangle = \frac{1}{N}\sum_{i=1}^{N}u(i).

130: \end{equation}

131: 

132: \item[(2)]

133: We divide the integrated signal $y(k)$ into non-overlapping boxes

134: of equal size $n$ (scale of analysis).

135: 

136: \item[(3)]

137: In each box of size $n$, we fit the integrated time series $y(k)$

138: by using a polynomial function, $y_{fit}(k)$, which is called the

139: local trend. For order-$\ell$ DFA (DFA-1 if $\ell=1$, DFA-2 if

140: $\ell=2$ etc.), $\ell$-order polynomial function is applied for

141: the fitting.

142: 

143: \item[(4)]

144: We detrend the integrated time series $y(k)$ by subtracting the

145: local trend $y_{fit}(k)$ in each box, and we calculate the

146: detrended fluctuation function

147: 

148: \begin{equation}

149: Y(k) = y(k)-y_{fit}(k).

150: \end{equation}

151: 

152: \item[(5)]

153: For a given box size $n$, we calculate the root-mean-square (rms)

154: fluctuation

155: 

156: \begin{equation}

157: F(n) = \sqrt{\frac{1}{N}\sum_{k=1}^{N}[Y(k)]^2}

158: \end{equation}

159: 

160: and repeat the above computation for different box sizes $n$

161: (different scales) to provide a relationship between $F(n)$ and

162: $n$.

163: 

164: \end{enumerate}

165: 

166: A power-law relation $F(n) \sim n^{\alpha}$ between $F(n)$ and the

167: box size $n$ indicates the presence of scaling. The parameter

168: $\alpha$, called the scaling exponent or correlation exponent,

169: represents the correlation properties of the signal. If

170: $\alpha=0.5$, there is no correlation and the signal is an

171: uncorrelated signal (white noise); if $\alpha < 0.5$, the signal

172: is anticorrelated; if $\alpha >0.5$, there are positive

173: correlations in the signal.

174: 

175: 

176: \section{DFA on correlations of RG and SF}

177: 

178: First we will give a brief description about the attack and repair

179: strategy. We start by constructing a network according to the

180: algorithms of Erd\"{o}s-R\'{e}nyi random graph model (RG) or

181: Barab\'{a}si-Albert scale-free model (SF). The dynamics of the

182: attack and repair strategy is defined in terms of the following

183: two operations:

184: 

185: \begin{itemize}

186: 

187: \item

188: \textit{Attack}: Find a node with the maximum degree $k_{max}$ and

189: remove all its links. (If several nodes happen to have the same

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

191: 

192: \item

193: \textit{Repair}: Reconnect this node with the other nodes in the

194: network with repair probability $p_{re}$.

195: 

196: \end{itemize}

197: 

198: Then, the evolution comes into the next Monte Carlo time step. At

199: each Monte Carlo time step $s$, we record the maximum degree

200: $k_{max}(s)$ and the average degree $\langle k \rangle(s)$ of the

201: system. The reason we choose $k_{max}$ and $\langle k \rangle$ to

202: analyze is that they represent the local and statistical features

203: of the system. Since the time step should be large enough to

204: enable the system to reach stationary state, it is chosen to be

205: 1.5 million in our simulation. To reduce the effect of fluctuation

206: on calculated results, for every system with size $N$, the

207: calculated results $k_{max}$ and $\langle k \rangle$ are averaged

208: over 10 independent runs for each network realization.

209: 

210: \subsection{DFA on the maximum degree $k_{max}$}

211: 

212: According to the attack and repair strategy, it can be easily seen

213: that the network has a tendency to decrease the maximum number of

214: connections among the nodes at a long time scale, because the

215: nodes with highest connections will be damaged and reconnected

216: randomly.

217: 

218: \begin{figure}

219: \begin{center}

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

221: \caption{The maximum degree $k_{max}$ for (a) random graph and (b)

222: scale-free network under $N=1000$ and $p_{re}=0.01$.}

223: \end{center}

224: \end{figure}

225: 

226: In Fig. 1, we plot the maximum degree $k_{max}$ versus time step

227: $s$ for RG and SF with $N=1000$ nodes and repair probability

228: $p_{re}=0.01$. It should be illustrated here that, in order to

229: reduce the number of parameters in the model, we set the repair

230: probability $p_{re}$ the same as the connection probability

231: $p_{rg}$ in RG. Fig. 1 shows that $k_{max}$ of RG decreases slowly

232: and fluctuates widely, while $k_{max}$ of SF decreases very

233: steeply at the beginning and then very slowly with small

234: fluctuations. The different behaviors of $k_{max}$ for RG and SF

235: are rooted in their different topological structures. RG has

236: homogeneous Poissonian degree distribution: all nodes have

237: approximately the same number of links. The removal of each node

238: causes the same amount of damage. SF is an extremely inhomogeneous

239: network: because the power-law distribution implies that the

240: minority of nodes have highly connected nodes. The removal of

241: these nodes drastically decreases the degrees of the network.

242: 

243: \begin{figure}

244: \begin{center}

245: \includegraphics[height=8cm,width=8cm]{fig2.eps}

246: \caption{Scaling behavior of the root-mean-square (rms)

247: fluctuation function $F(n)$ as a function of the scale $n$, for

248: series $k_{max}$ of random graphs (RG) and scale-free networks

249: (SF) under $N=1000$ and $p_{re}=$ 0.01, 0.02, 0.03, respectively.

250: The dashed line represents the fit of the slope.}

251: \end{center}

252: \end{figure}

253: 

254: In Fig.2 we present results from the DFA-3 analysis for the

255: maximum degree $k_{max}$ of RG and SF. The system size is $N=1000$

256: nodes and the repair probability $p_{re}$ are 0.01, 0.02, 0.03,

257: respectively. In this double logarithm graph, the root-mean-square

258: (rms) fluctuation function $F(n)$ as a function of the scale $n$,

259: exhibit two distinct scaling regimes with the crossover at about

260: $n=10^4$,

261: 

262: \begin{equation}

263: F(n)\sim \left\{\begin{array}{ll} n^{\alpha_1}, & {n < 10^4}; \\

264: n^{\alpha_2}, & {n > 10^4}.

265: \end{array}\right.

266: \end{equation}

267: 

268: \noindent For $n<10^4$, the scaling exponent $\alpha_1$ is about

269: 0.51; while for $n>10^4$, the exponent $\alpha_2$ is about 0.05.

270: Our DFA results suggest that the fluctuation is quite random

271: (close to white noise) for short time scales. As the time scale

272: becomes larger, the fluctuation displays strong anticorrelation,

273: that is, a large value (compared to the average) of $k_{max}$ is

274: almost always followed by a small value. With the increase of the

275: repair probability $p_{re}$, we find that there is a vertical

276: shift of the rms fluctuation function $F(n)$ to larger values and

277: the shift between $p_{re}=0.01$ and 0.02 is very big. In addition,

278: the similar shape between $F(n)$ and $n$ is observed for both RG

279: and SF under different $p_{re}$, which illustrates that the time

280: series of $k_{max}$ for RG and SF have the similar correlation

281: properties.

282: 

283: 

284: \subsection{DFA on the average degree $\langle k \rangle$}

285: 

286: \begin{figure}

287: \begin{center}

288: \includegraphics[height=8cm,width=8cm]{fig3.eps}

289: \caption{The average degree $\langle k \rangle$ for (a) random

290: graph and (b) scale-free network under $N=1000$ and

291: $p_{re}=0.01$.}

292: \end{center}

293: \end{figure}

294: 

295: The average degree $\langle k \rangle$ is a statistical

296: topological property of complex networks. The time series of

297: $\langle k \rangle$ reflect the general changes of RG and SF under

298: the evolution of attack and repair. We find that $\langle k

299: \rangle$ of both RG and SF decay exponentially with small

300: fluctuations as shown in Fig. 3, under the system size $N=1000$

301: and the repair probability $p_{re}=0.01$. The inset panel of Fig.

302: 3 gives the series of $\langle k \rangle$ in linear coordinates,

303: showing that $\langle k \rangle$ decreases very steeply at first

304: and then becomes steadily. In fact, the curve of $\langle k

305: \rangle$, to a large degree, depends on the repair probability

306: $p_{re}$. The bigger the repair probability, the smoother the

307: curve will be. Here we choose $p_{re}=0.01$ to analyze,

308: considering to have a comparison with the results of $k_{max}$.

309: 

310: We next investigate the correlation properties of series $\langle

311: k \rangle$. Fig. 4(a) shows the results of applying DFA-3 to the

312: series $\langle k \rangle$ of RG under $N=1000$ and $p_{re}=$0.01,

313: 0.02, 0.03, respectively. In the log-log coordinates, the rms

314: fluctuation function $F(n)$ and the scale $n$ has a power-law

315: dependence with scaling exponent $\alpha = 0.64$, which indicates

316: the presence of the positive correlations and the weak long-range

317: (long-memory) power-law correlations. The results suggest that a

318: value of $\langle k \rangle$ at a given time depends on values of

319: long past in a power-law fashion. We also note that $F(n)$ shifts

320: downwards slightly with the increase of $p_{re}$, which is

321: oppositive to that of $k_{max}$ in Fig. 2. Besides, the big shift

322: between $p_{re}=0.01$ and 0.02 vanishes here.

323: 

324: Fig. 4(b) reports the DFA-3 results of time series $\langle k

325: \rangle$ for SF under $N=1000$ and $p_{re}=$0.01, 0.02, 0.03,

326: correspondingly. The two-segment power-law relationship is

327: observed with the crossover at about $n=10^3$,

328: 

329: \begin{equation}

330: F(n)\sim \left\{\begin{array}{ll} n^{\alpha_1}, & {n < 10^3}; \\

331: n^{\alpha_2}, & {n > 10^3}.

332: \end{array}\right.

333: \end{equation}

334: 

335: \noindent For short time scales ($n < 10^3$), the rms fluctuation

336: function $F(n)$ has the similar scaling exponent with

337: $\alpha_1=0.55$ for different repair probability $p_{re}$. For

338: longer time scales ($n

339: > 10^3$), the exponents $\alpha_2$ of $p_{re}=0.01$, 0.02, 0.03 correspond to 0.54, 0.36,

340: 0.23. It shows that the rms fluctuation function $F(n)$ decreases

341: and the series $\langle k \rangle$ becomes strong anticorrelated

342: with increasing $p_{re}$ for longer time scales.

343: 

344: \begin{figure}

345: \begin{center}

346: \includegraphics[height=8cm,width=8cm]{fig4.eps}

347: \caption{Scaling behavior of the root-mean-square (rms)

348: fluctuation function $F(n)$ as a function of the scale $n$, for

349: series $\langle k \rangle$ of (a) random graphs (RG) and (b)

350: scale-free networks (SF), under $N=1000$ and $p_{re}=$ 0.01, 0.02,

351: 0.03, respectively. The dashed line for RG and the solid line for

352: SF represent the fit of the slopes.}

353: \end{center}

354: \end{figure}

355: 

356: Fig. 4(a) and 4(b) demonstrate that the rms fluctuation function

357: $F(n)$ of series $\langle k \rangle$ exhibits different scaling

358: behaviors for RG and SF. The positive long-lange power-law

359: correlations of RG are probably due to the strong restriction,

360: defined as the connection probability equals the repair

361: probability, $p_{rg}=p_{re}$. For scale-free networks, the repair

362: probability $p_{re}$ has a great effect on the power-law degree

363: distribution of SF in small ranges. The randomness are dominated

364: at small time scales, characterized by uncorrelated series. While

365: at large time scales, the power-law degree distributions are

366: dominated, characterized by the anticorrelation.

367: 

368: Further, Fig. 2 and Fig. 4 represent the DFA results of series

369: $k_{max}$ and $\langle k \rangle$. We note that, with the increase

370: of $p_{re}$, the rms fluctuation function $F(n)$ increases in

371: $k_{max}$ but decreases in $\langle k \rangle$. It is the

372: consequence that more random connections are added with increasing

373: $p_{re}$. According to the local property of $k_{max}$, the

374: fluctuation of the system increases with added randomness. While

375: according to the statistical property of $\langle k \rangle$, the

376: fluctuation of the system decreases since the added connections

377: balanced the degrees of the whole network.

378: 

379: 

380: \section{Conclusions}

381: 

382: In summary, we use DFA method to analyze the correlation

383: properties of random graphs and scale-free networks under the

384: attack and repair strategy. According to the analysis of the

385: maximum degree $k_{max}$ representing the local property of the

386: system, we find that both random graphs and scale-free networks

387: exhibit the similar scaling behaviors, which suggest that the

388: fluctuations are quite random at short time scales but strong

389: anticorrelated at longer time scales. With the analysis of the

390: average degree $\langle k \rangle$ representing the statistical

391: property of the system, we observe that random graphs and

392: scale-free networks display completely different scaling

393: behaviors. Random graphs show long-range power-law correlations;

394: while scale-free networks are uncorrelated at short time scales

395: but display anticorrelations at longer time scales. With the

396: increase of the repair probability $p_{re}$, the apparent vertical

397: shifts of the rms fluctuation function $F(n)$ for $k_{max}$ and

398: $\langle k \rangle$ are due to the effects of the increase of

399: random connections to the networks.

400: 

401: 

402: \section*{Acknowledgments}

403: 

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

405: Foundation of China under grant Nos. 70271067, 70401020 and by the

406: Ministry of Education of China under grant No. 03113.

407: 

408: 

409: \begin{thebibliography}{ss}

410: \bibitem{s1}

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

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

413: 

414: \bibitem{s2}

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

416: 

417: \bibitem{s3}

418: W. Li and X. Cai, Phys. Rev. \textbf{E69} (2004) 046106.

419: 

420: \bibitem{s4}

421: L.P. Chi, R. Wang, H. Su, X.P. Xu, J.S. Zhao, W. Li, and X. Cai,

422: Chin. Phys. Lett. \textbf{20} (2003) 1393.

423: 

424: \bibitem{s5}

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

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

427: 

428: \bibitem{s6}

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

430: (1999) 130.

431: 

432: \bibitem{s7}

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

434: (2002) 47.

435: 

436: \bibitem{s8}

437: M.E.J. Newman, SIAM Review \textbf{45} (2003) 167.

438: 

439: \bibitem{s9}

440: K.-I. Goh, B. Kahng, and D. Kim, Phys. Rev. Letts. \textbf{87}

441: (2001) 278701.

442: 

443: \bibitem{s10}

444: A.-L. Barab\'{a}si, R. Albert, and H. Jeong, Physica \textbf{A281}

445: (2000) 69.

446: 

447: \bibitem{s11}

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

449: \textbf{A340} (2004) 388.

450: 

451: \bibitem{s12}

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

453: \textbf{A320} (2003) 622.

454: 

455: \bibitem{s13}

456: P. Holme, B.J. Kim, C.N. Yoon, and S.K. Han, Phys. Rev.

457: \textbf{E65} (2002) 056109.

458: 

459: \bibitem{s14}

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

461: Letts. \textbf{85} (2000) 4626.

462: 

463: \bibitem{s15}

464: L.K. Gallos, R. Cohen, P. Argyrakis, A. Bunde, and S. Havlin,

465: Phys. Rev. Letts. \textbf{94} (2005) 188701.

466: 

467: \bibitem{s16}

468: L.P. Chi, C.B. Yang, and X. Cai, cond-mat/0505197.

469: 

470: \bibitem{s17}

471: C.-K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E. Stanley, and

472: A.L. Goldberger, Phys. Rev. \textbf{E49} (1994) 1685.

473: 

474: \bibitem{s18}

475: C.-K. Peng, S. Halvin, H.E. Stanley, and A.L. Goldberger, Chaos

476: \textbf{5} (1995) 82.

477: 

478: \bibitem{s19}

479: Kun Hu, Plamen Ch. Ivanov, Zhi Chen, Pedro Carpena, and H.Eugene

480: Stanley, Phys. Rev. \textbf{E64} (2001) 011114.

481: 

482: \bibitem{s20}

483: Zhi Chen, Kun Hu, Pedro Carpena, Pedro Bernaola-Galvan, H. Eugene

484: Stanley, and Plamen Ch. Ivanov, Phys. Rev. \textbf{E71} (2005)

485: 011104.

486: 

487: \bibitem{s21}

488: Jan W. Kantelhardt, Eva Koscielny-Bunde, Henio H.A. Rego, Shlomo

489: Havlin, and Armin Bunde, Physica \textbf{A295} (2001) 441.

490: 

491: \bibitem{s22}

492: A. Eke, P. Herman, L. Kocsis, and L.R. Kozak, Physiol. Meas.

493: \textbf{23} (2002) R1.

494: 

495: \end{thebibliography}

496: 

497: 

498: \end{document}

499: