1: \documentclass[pre,a4paper,superscriptaddress,showpacs]{revtex4}
2: % in the preprint version, add preprint, 12pt, and delete pre
3: \usepackage{graphicx}
4: \setlength{\textwidth}{15.5cm}
5: %\setlength{\textwidth}{13.8cm}
6:
7: \begin{document}
8:
9: %\setlength{\baselineskip}{0.77cm}
10:
11: \newcommand{\EQ}{Eq.~}
12: \newcommand{\EQS}{Eqs.~}
13: \newcommand{\FIG}{Fig.~}
14: \newcommand{\FIGS}{Figs.~}
15: \newcommand{\TAB}{Tab.~}
16: \newcommand{\SEC}{Sec.~}
17: \newcommand{\SECS}{Secs.~}
18:
19: %\newtheorem{theorem}{Theorem}
20:
21: \title{Return times of random walk on generalized random graphs}
22: \author{Naoki Masuda}
23: \affiliation{Laboratory for Mathematical Neuroscience,
24: RIKEN Brain Science Institute, 2-1, Hirosawa, Wako, Saitama, 351-0198,
25: Japan}
26: \author{Norio Konno}
27: \affiliation{Faculty of Engineering,
28: Yokohama National University,
29: 79-5, Tokiwadai, Hodogaya, Yokohama, 240-8501 Japan}
30: %norio@mathlab.sci.ynu.ac.jp\\
31: %
32: \date{Received 14 November 2003}
33:
34: \begin{abstract}
35: Random walks are used for modeling various dynamics in, for
36: example, physical, biological, and social contexts. Furthermore,
37: their characteristics provide us with useful information on the phase
38: transition and critical phenomena of even broader classes of related
39: stochastic models. Abundant results are obtained for random walk on
40: simple graphs such as the regular lattices and the Cayley
41: trees. However, random walks and related processes on more complex
42: networks, which are often more relevant in the real world, are still
43: open issues, possibly yielding different characteristics. In this
44: paper, we investigate the return times of random walks on random
45: graphs with arbitrary vertex degree distributions. We analytically
46: derive the distributions of the return times. The
47: results are applied to some types of networks and compared with
48: numerical data.
49: \end{abstract}
50:
51: \pacs{02.50.Ga, 05.40.Fb, 89.75.-k}
52:
53: \maketitle
54:
55: \section{Introduction}\label{sec:introduction}
56:
57: The theory of the random walk has a long history. Random walks and their
58: extensions
59: have also been applied
60: with profound theoretical bases to modeling numerous types of
61: physical, biological, sociological, and economical dynamics
62: \cite{Spitzer}. For example, distributions of return times
63: and scaled limit distributions of the walkers' positions
64: are broadly known for simple underlying graphs. They provide
65: useful information on critical values and phase transitions
66: in regard to survival of the branching random walks and the contact
67: processes \cite{Liggettbook99,Schinazibook,Pemantle01}, survival in the voter
68: models \cite{Liggettbook99,Durrettbook}, and
69: occurrence of percolation
70: \cite{Lyons90_Lyons95},
71:
72: Indeed, a large body of theoretical results are available for random
73: walks performed on regular lattices such as ${\bf Z}^d$ and on the
74: Cayley (or regular) trees, which are defined to be trees with
75: homogeneous vertex degree. However, it has been suggested recently
76: that more complex networks as opposed to regular graphs and
77: conventional random graphs \cite{Erdos} are concerned to real worlds.
78: Particularly, important classes of random graphs such as small-world
79: networks and scale-free networks were proposed and have been examined
80: in the last several years. These networks share some important
81: properties with real networks, such as the clustering property, short
82: average path length, and the power-law of the vertex degree
83: distributions \cite{SW-SF,Callaway00,Albert02,Newman}. They have been
84: applied to the analysis of various biological, engineering, and social
85: networks including information flow in the Internet
86: \cite{Callaway00,Albert02,Newman} and epidemics
87: \cite{Newman,Eguiluz_Pastor_Masuda_SARS}. The
88: properties of spatial stochastic models, both static configurations
89: and dynamical processes, typically change as the network topology
90: varies even when other basic quantities such as the mean vertex degree
91: is conserved. For example, the analysis of percolation-based models
92: revealed that the critical parameter values for the
93: occurrence of global epidemics, or even their existence, depend on
94: network topology \cite{Callaway00,Albert02,Newman}.
95:
96: It is highly likely that the properties of random walks depend on
97: network topology \cite{Schinazibook,Pemantle01}, as numerical and
98: approximate results suggest for the quenched \cite{Pandit}
99: and annealed \cite{Jasch_Lahtinen} Watts-Strogatz-type small-world
100: networks and for quenched random graphs with homogeneous vertex
101: degree \cite{Szabo00rw}. In relation to this issue, how eigenvalues of
102: the adjacency matrices are distributed has been numerically examined
103: for scale-free and small-world networks \cite{Farkas_Goh}. The largest
104: eigenvalue of an adjacency matrix measures how the number of closed
105: paths increases as the path length tends to
106: infinity. The eigenvalues supply useful information on the return
107: times of random walks \cite{Spitzer}, serving to a wide range of
108: applications as mentioned above. However, the largest eigenvalue
109: $\rho$ has
110: been characterized only in terms of the numerical scaling law
111: for the scale-free networks
112: in an unnormalized manner, namely, $\rho\propto
113: m^{1/2}N^{1/4}$, where $N$ is the system size and $2m$ is the mean
114: vertex degree \cite{Farkas_Goh}.
115:
116: In this paper, we analyze random walks on a general class of random
117: graphs that includes random scale-free networks, the
118: Erd\"{o}s-R\'{e}nyi random graph, and the Cayley trees as special
119: cases \cite{Lyons90_Lyons95,Pemantle01,Albert02,Newman}. Explicit
120: expressions for the first return time probability and the annealed
121: approximation forms for the general
122: return time probability are derived with the use of partition of
123: integers. In \SEC\ref{sec:model}, we introduce the network model and
124: the generating functions. In
125: \SEC\ref{sec:return}, we calculate the probability distribution
126: functions of the return time of random walk. Then, in
127: \SEC\ref{sec:examples}, we confirm with some
128: examples that our theoretical estimates match
129: numerical results. Lastly, the conclusion
130: follows, and the difference in the decay
131: rate of the return time probability between regular and random
132: networks,
133: which implies the difference in
134: the possibility of percolation and the survival of contact processes,
135: are also touched upon.
136:
137: \section{Network Model and Generating Functions}\label{sec:model}
138:
139: We analyze a class of random graphs called generalized random graphs
140: in physical contexts \cite{Callaway00,Albert02,Newman} or
141: Galton-Watson trees in mathematical contexts
142: \cite{Lyons90_Lyons95,Pemantle01}. These random graphs are infinite
143: trees without loops. The degree of each vertex, or the number of
144: neighbors, is distributed according to an identical and independent
145: probability density function. As shown in \FIG\ref{fig:rw_pic},
146: each realization of the graph, which is generally inhomogeneous,
147: is taken from the random ensemble. However,
148: they are regular in a
149: statistical sense. Let us denote by $p_k$ the probability that a
150: vertex has the degree equal to $k$. We assume that $p_0=0$ without
151: losing generality. Consequently, $\sum^{\infty}_{k=1}p_k=1$.
152:
153: Let us designate an arbitrary vertex $O$ of a realized graph
154: as the root. We examine a random walk
155: starting from $O$. Since we exclusively deal with trees here, the
156: random walker can return to $O$ only when the time $n\in$ $\{0, 1, 2,
157: \ldots \}$ is even. In accordance, we denote by $q_n$ the probability
158: that the random walker returns to $O$ for the first time at time $2n$,
159: and by $r_n$ the probability that it returns to $O$ irrespective of
160: the accumulated number of returns. Here we consider only the
161: annealed random walk, confining ourselves in the analysis of return
162: times averaged over both probability space of graph and that of
163: random walk. To be contrasted with the annealed randomness
164: is the quenched randomness, which
165: is concerned to the
166: ensemble of walkers on a fixed realization of random graph
167: \cite{Zeitouni02}. Both quenched
168: \cite{Pandit,Szabo00rw} and annealed \cite{Jasch_Lahtinen} random walks have
169: been implicitly treated in the studies of
170: random dynamics on complex networks.
171: Though quenched environments are realistic,
172: the statistics based on annealed walks that we derive in the
173: following
174: can be regarded as averages of the statistics of quenched walks
175: over the ensemble of a random graph.
176:
177: The generating functions for the distributions
178: $\{q_n\}$ and $\{r_n\}$, which we respectively denote by
179: $Q(z)$ and $R(z)$ are defined by
180: %
181: \begin{equation}
182: Q(z)\equiv \sum^{\infty}_{n=0} q_n z^n,\quad
183: %
184: R(z)\equiv \sum^{\infty}_{n=0} r_n z^n.
185: \end{equation}
186: %
187: With $q_0 = 0$ and $r_0 = 1$, $Q(z)$ and $R(z)$ satisfy the following
188: relation:
189: %
190: \begin{eqnarray}
191: R(z) &=& \sum^{\infty}_{n=0}
192: \left( \sum^{n}_{m=1} q_m r_{n-m} + \delta_{n,0}\right) z^n\nonumber\\
193: %
194: &=& R(z)Q(z) + 1,
195: \label{eq:Q_R}
196: \end{eqnarray}
197: %
198: where $\delta_{i,j} = 1$ for $i=j$ and $\delta_{i,j} = 0$
199: otherwise \cite{Spitzer,Grimmettbook}.
200: Strictly speaking, \EQ(\ref{eq:Q_R}) is valid only for the quenched
201: case. Therefore, the following results for $R(z)$ should be understood
202: as an approximation by annealed statistics.
203:
204: \section{Distributions of Return Times}\label{sec:return}
205:
206: To derive explicit expressions for the return time
207: distributions, we provide the
208: approximate recursion relation below.
209: Let us resort to \FIG\ref{fig:rw_pic} for explanation.
210: Suppose that the random walker starting from $O$ returns to $O$ after $2n$
211: steps for the first time ($n=7$ in \FIG\ref{fig:rw_pic}).
212: In the first step, the random walker moves
213: to a neighbor that we denote by $O^{\prime}$. Because of the
214: statistical homogeneity of the generalized random graph, the vertex
215: degree of $O^{\prime}$ is distributed as specified by $\{ p_k\}$
216: whichever neighbor of $O$ is chosen. The random walker has to arrive
217: at $O^{\prime}$ at time $2n-1$ ($=13$) and move to $O$ subsequently at time
218: $2n$ $(=14)$.
219: The last event occurs with probability $1/k$. In the meantime,
220: the random walker travels for $2n-2$ steps without visiting
221: $O$. The walker wanders in the subtrees rooted at
222: $O^{\prime}$ to complete loops, or closed paths of random walk.
223: Any such loop cannot contain $O$, and the probability that a
224: path emanating from $O^{\prime}$ enters a subtree is $(k-1)/k$. Let us
225: denote by $a$ the number of the loops originating from
226: $O^{\prime}$. In \FIG\ref{fig:rw_pic}, $a$ is equal to 2.
227: Then, the length $2n_i$ ($1\le i\le a$) of each loop is
228: even ($n_1=5$ and $n_2=1$ in \FIG\ref{fig:rw_pic}),
229: and $2n_i$ must sum up to $2n-2$. In addition, since the vertex
230: degree is homogeneously distributed, the probability law for
231: the length of loop is assumed to be the same as that for
232: the original random walk starting and ending at $O$.
233:
234: Here we make a crucial approximation of disregarding any
235: memory effects. In other words, we suppose that the $a$ subtrees
236: rooted at $O^{\prime}$ are
237: independent of each other. In fact, if the same neighbor of
238: $O^{\prime}$ is chosen for different entries into the subtree, the
239: subtrees reached by these different
240: entries coincide. As an example, the random walker shown in
241: \FIG\ref{fig:rw_pic} travels from $A$ to
242: the subtree rooted at $B$ twice, before returning to
243: $O$. In
244: this occasion, it is not qualified to regard the vertex degrees and
245: the loop lengths to be independent for
246: the two neighbors of $O^{\prime}$. However, the
247: approximation error is small unless the mean vertex
248: degree is extremely small. The accuracy of the
249: following analytical methods are
250: investigated in comparison with
251: numerical simulations in \SEC\ref{sec:examples}.
252:
253: Based on the consideration above, we have the following recursion formula:
254: %
255: \begin{eqnarray}
256: q_n &=& \sum^{\infty}_{k=1}p_k \sum^{n-1}_{a=0}
257: \sum_{\sum^a_{i=1} n_i = n-1, n_i\ge 0, 1\le i\le a}
258: \frac{k-1}{k}q_{n_1}\; \frac{k-1}{k}q_{n_2}\;
259: \ldots \; \frac{k-1}{k}q_{n_a} \; \frac{1}{k}\nonumber\\
260: %
261: &=& \sum^{\infty}_{k=1}\frac{p_k}{k} \sum^{\infty}_{a=0}
262: \left(\frac{k-1}{k}\right)^a
263: \sum_{\sum^a_{i=1} n_i = n-1, n_i\ge 0, 1\le i\le a}
264: \prod^a_{a^{\prime}=1} q_{n_{a^{\prime}}},
265: \label{eq:q_n}
266: \end{eqnarray}
267: %
268: which covers the singular case $q_0=0$ as well.
269: Using \EQ(\ref{eq:q_n}), the generating function of $q_n$ is
270: calculated as
271: %
272: \begin{eqnarray}
273: Q(z) &=& \sum^{\infty}_{n=0} \left\{
274: \sum^{\infty}_{k=1}\frac{p_k}{k} \sum^{\infty}_{a=0}
275: \left(\frac{k-1}{k}\right)^a
276: \sum_{\sum^a_{i=1} n_i = n-1, n_i\ge 0, 1\le i\le a}
277: \prod^a_{a^{\prime}=1} q_{n_{a^{\prime}}} \right\} z^n\nonumber\\
278: %
279: &=& z
280: \sum^{\infty}_{k=1}\frac{p_k}{k} \sum^{\infty}_{a=0}
281: \left(\frac{k-1}{k}\right)^a
282: \sum^{\infty}_{n=0}
283: \sum_{\sum^a_{i=1} n_i = n-1, n_i\ge 0, 1\le i\le a}
284: \prod^a_{a^{\prime}=1} q_{n_{a^{\prime}}}z^{n_a^{\prime}} \nonumber\\
285: %
286: &=&
287: z \sum^{\infty}_{k=1}\frac{p_k}{k} \sum^{\infty}_{a=0}
288: \left(\frac{k-1}{k}\right)^a
289: Q(z)^a\nonumber\\
290: %
291: &=& z \sum^{\infty}_{k=1}\frac{p_k}{k-(k-1)Q(z)}.
292: \label{eq:Q}
293: \end{eqnarray}
294: %
295: Although $Q(1)=1$ is always consistent with \EQ(\ref{eq:Q}),
296: we exclude this case because the random walk on generalized random
297: graphs including the Cayley trees
298: is transient \cite{Lyons90_Lyons95}, except for the Cayley tree
299: with vertex degree 2, which is identical to ${\bf Z}$.
300: Accordingly, we look for the solution satisfying $Q(1)<1$.
301:
302: By expanding the right-hand side of
303: \EQ(\ref{eq:Q}), we obtain
304: %
305: \begin{eqnarray}
306: Q(z) &=& z \sum^{\infty}_{k=1} \frac{p_k}{k}
307: \sum^{\infty}_{n=0} \left( \frac{k-1}{k}Q(z) \right)^n\nonumber\\
308: %
309: &=& \frac{z M\left[Q(z)\right]}{Q(z)},
310: \label{eq:Q_expand}
311: \end{eqnarray}
312: %
313: where
314: %
315: \begin{equation}
316: M(z)\equiv \sum^{\infty}_{n=1} m_n z^n
317: \end{equation}
318: %
319: is the generating function of the moment function
320: given by
321: %
322: \begin{equation}
323: m_n \equiv \sum^{\infty}_{k=1} \frac{(k-1)^{n-1}}{k^n}\; p_k.
324: \end{equation}
325: %
326: In deriving \EQ(\ref{eq:Q_expand}), the expansion is justified by the
327: fact that $Q(z)$ has the radius of convergence equal to 1 and that
328: $(k-1)/k<1$. Then, we apply the following theorem to calculate
329: $Q(z)$ and $R(z)$.
330:
331: {\it Lagrange's inversion formula} \cite{Grimmettbook}
332: %
333: Let $z=w/f(w)$ where $w/f(w)$ is an analytic function of $w$ near
334: $w=0$. If $g$ is infinitely differentiable, then
335: %
336: \begin{equation}
337: g\left(w(z)\right) = g(0) + \sum^{\infty}_{n=1}\frac{z^n}{n!}
338: \left[ \frac{d^{n-1}}{du^{n-1}}
339: \left[ g^{\prime}(u)f(u)^n\right] \right]_{u=0}.
340: \label{eq:Lagrange}
341: \end{equation}
342:
343: For our purpose, we set $w(z)=Q(z)$, $f(w) = M(w)/w$, $g(w)=w$ in
344: \EQ(\ref{eq:Lagrange}). Apparently, the fact that
345: $m_1>0$ guarantees the regularity
346: of $w/f(w)$ around $w=0$. As a result, we have
347: %
348: \begin{eqnarray}
349: Q(z) &=& g(w)\nonumber\\
350: %
351: &=& 0 +
352: \sum^{\infty}_{n=1}\frac{z^n}{n!}
353: \left\{ \frac{d^{n-1}}{du^{n-1}}
354: \left[ \left(\frac{M(u)}{u}\right)^n\right] \right\}_{u=0}.
355: \label{eq:Q_Lagrange}
356: \end{eqnarray}
357: %
358: We also note that
359: %
360: \begin{eqnarray}
361: M(u)^n &=& (m_1 u + m_2 u^2 + \cdots )^n\nonumber\\
362: %
363: &=& \sum^{\infty}_{n^{\prime}=n} u^{n^{\prime}}
364: \sum_{\lambda\vdash n^{\prime},
365: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
366: \frac{n!}{\prod^{\infty}_{l=1}i_{\lambda}(l)!}
367: \prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)},
368: \label{eq:M_power_n}
369: \end{eqnarray}
370: %
371: where $\sum_{\lambda\vdash n^{\prime}}$ indicates
372: the summation over all the partitions of the integer
373: $n^{\prime}$ into integers. In general, a partition $\lambda$
374: is represented by
375: $\lambda=(1^{i_{\lambda}(1)} 2^{i_{\lambda}(2)} \cdots )$, which
376: means that 1 is included $i_{\lambda}(1)$
377: times in $\lambda$,
378: 2 is included $i_{\lambda}(2)$ times, and so on \cite{Stanleybook}.
379: By the definition of partition, $\{i_{\lambda}(1), i_{\lambda}(2), \ldots
380: \}$ ($\lambda\vdash n^{\prime}$)
381: satisfies
382: %
383: \begin{equation}
384: \sum^{\infty}_{l=1} l\; i_{\lambda}(l)
385: = \sum^{n^{\prime}}_{l=1} l\; i_{\lambda}(l) = n^{\prime}.
386: \end{equation}
387: %
388: For example,
389: %
390: \begin{eqnarray}
391: \left\{ \lambda \; | \; \lambda\vdash 5 \right\} &=&
392: \left\{(1^5), (1^3 2), (1^2 3), (1 2^2), (1 4), (2 3), (5) \right\},
393: \label{eq:partition_ex_1}\\
394: %
395: \left\{ \lambda \; | \; \lambda\vdash 7 \right\} &=&
396: \left\{(1^7), (1^5 2), (1^4 3), (1^3 2^2), (1^3 4),
397: (1^2 2 3), (1^2 5), (1 2^3), (1 2 4), (1 3^2), \right.\nonumber\\
398: %
399: && \left. (1 6), (2^2 3), (2 5),
400: (3 4), (7) \right\}.
401: \label{eq:partition_ex_2}
402: \end{eqnarray}
403: %
404: In \EQ(\ref{eq:M_power_n}), only the partitions whose numbers of parts
405: are $n$ are concerned. Corresponding to \EQS(\ref{eq:partition_ex_1})
406: and (\ref{eq:partition_ex_2}), the partitions appearing in the
407: summation of \EQ(\ref{eq:M_power_n}) for $(n,n^{\prime})=(3,5)$ and
408: $(4,7)$ are as follows:
409: %
410: \begin{eqnarray}
411: \left\{ \lambda \; \bigg| \; \lambda\vdash 5, \sum^{\infty}_{l=1}
412: i_{\lambda}(l) = 3\right\} &=& \left\{ (1^2 3), (1 2^2) \right\},\\
413: %
414: \left\{ \lambda \; \bigg| \; \lambda\vdash 7, \sum^{\infty}_{l=1}
415: i_{\lambda}(l) = 4\right\} &=&
416: \left\{ (1^3 4), (1^2 2 3), (1 2^3) \right\}.
417: \end{eqnarray}
418: %
419: With \EQ(\ref{eq:M_power_n}), \EQ(\ref{eq:Q_Lagrange}) is evaluated
420: as follows:
421: %
422: \begin{eqnarray}
423: Q(z) &=&
424: \sum^{\infty}_{n=1}\frac{z^n}{n!}
425: (n-1)!
426: \sum_{\lambda\vdash 2n-1,
427: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
428: \frac{n!}{\prod^{\infty}_{l=1}i_{\lambda}(l)!}
429: \prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)}\nonumber\\
430: %
431: &=&
432: \sum^{\infty}_{n=1} z^n
433: \sum_{\lambda\vdash 2n-1,
434: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
435: c^{(n)}_{\lambda}\prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)},
436: \label{eq:Q_z_final}
437: \end{eqnarray}
438: %
439: where
440: %
441: \begin{equation}
442: c^{(n)}_{\lambda} = \frac{(n-1)!}{\prod^{\infty}_{l=1}i_{\lambda}(l)!}.
443: \end{equation}
444: %
445: Accordingly,
446: %
447: \begin{equation}
448: q_n = \sum_{\lambda\vdash 2n-1,
449: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
450: c^{(n)}_{\lambda}\prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)}
451: \label{eq:q_n_final}
452: \end{equation}
453: %
454: when $n\ge 1$, and $q_0 = 0$.
455:
456: What is necessary for deriving $R(z)$ is just to replace $g(w)=w$ with
457: $g(w)=1/(1-w)$ when applying \EQ(\ref{eq:Lagrange}). Using
458: \EQ(\ref{eq:M_power_n}), we obtain
459: the annealed approximation form of $R(z)$:
460: %
461: \begin{eqnarray}
462: \hspace*{-0.8cm}
463: R(z) &=& \frac{1}{1-Q(z)}\nonumber\\
464: %
465: &=& g(w)\nonumber\\
466: %
467: &=& 1 +
468: \sum^{\infty}_{n=1}\frac{z^n}{n!}
469: \left[ \frac{d^{n-1}}{du^{n-1}}
470: \left( \frac{-1}{(1-u)^2}
471: \left(\frac{M(u)}{u}\right)^n\right) \right]_{u=0}\nonumber\\
472: %
473: &=& 1 +
474: \sum^{\infty}_{n=1}\frac{z^n}{n!}
475: \sum^{n-1}_{n^{\prime\prime}=0}
476: \left[ \frac{(-1)^{n-n^{\prime\prime}}(n-n^{\prime\prime})!}{(1-u)^{n-n^{\prime\prime}+1}}
477: \frac{d^{n^{\prime\prime}}}{du^{n^{\prime\prime}}}
478: \left(\frac{M(u)}{u}\right)^n \right]_{u=0}\nonumber\\
479: %
480: &=& 1 +
481: \sum^{\infty}_{n=1}\frac{z^n}{n!}
482: \sum^{n-1}_{n^{\prime\prime}=0}
483: (-1)^{n-n^{\prime\prime}}(n-n^{\prime\prime})!\nonumber\\
484: %
485: &&\times
486: \left[ \sum^{\infty}_{n^{\prime}=n+n^{\prime\prime}}
487: \frac{(n^{\prime}-n)!}{(n^{\prime}-n-n^{\prime\prime})!}
488: u^{n^{\prime}-n-n^{\prime\prime}}
489: \sum_{\lambda\vdash n^{\prime},
490: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
491: \frac{n!}{\prod^{\infty}_{l=1}i_{\lambda}(l)!}
492: \prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)}
493: \right]_{u=0}\nonumber\\
494: %
495: &=& 1 +
496: \sum^{\infty}_{n=1}\frac{z^n}{(n-1)!}
497: \sum^{n-1}_{n^{\prime\prime}=0}
498: (-1)^{n-n^{\prime\prime}}(n-n^{\prime\prime})!\; n^{\prime\prime}!
499: \sum_{\lambda\vdash n+n^{\prime\prime},
500: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
501: c^{(n)}_{\lambda} \prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)},
502: \label{eq:R_Lagrange}
503: \end{eqnarray}
504: %
505: which results in
506: %
507: \begin{equation}
508: r_n = \frac{1}{(n-1)!}
509: \sum^{n-1}_{n^{\prime\prime}=0}
510: (-1)^{n-n^{\prime\prime}}(n-n^{\prime\prime})!\; n^{\prime\prime}!
511: \sum_{\lambda\vdash n+n^{\prime\prime},
512: \sum^{\infty}_{l=1} i_{\lambda}(l) = n}
513: c^{(n)}_{\lambda} \prod^{\infty}_{l=1} m_l^{i_{\lambda}(l)},
514: \end{equation}
515: %
516: when $n\ge 1$, and $r_0 = 1$.
517:
518: \section{Examples}\label{sec:examples}
519:
520: The analytical methods developed in
521: \SEC\ref{sec:return} can be broadly applied since the only assumptions
522: that we have made on $\{p_k \}$ are
523: $p_0=0$ and that the average vertex
524: degree is not so small. In this section, we apply our
525: theoretical estimates
526: to random walk on some classes of graphs
527: that are often relevant in real-world situations and also of
528: theoretical interest.
529:
530: \subsection{Cayley trees}
531:
532: Let us first consider the Cayley trees \cite{Albert02}
533: in which each vertex has exactly $d$
534: vertices. Substituting $p_k = \delta_{k,d}$ into \EQ(\ref{eq:Q}) yields
535: %
536: \begin{equation}
537: Q(z) = \frac{1}{d-(d-1)Q(z)}.
538: \label{eq:Q_cayley1}
539: \end{equation}
540: %
541: Although \EQ(\ref{eq:Q_cayley1}) has two different solutions of $Q(z)$,
542: the one satisfying $Q(1)=1$ is excluded
543: because of the transient nature of the random walk on the Cayley trees
544: \cite{Spitzer,Lyons90_Lyons95,Grimmettbook}.
545: Then \EQ(\ref{eq:Q_cayley1}) is led to
546: %
547: \begin{equation}
548: Q(z) = \frac{d-\sqrt{d^2-4(d-1)z}}{2(d-1)}.
549: \label{eq:Q_cayley2}
550: \end{equation}
551: %
552: In this case, $Q(z)$ is related to the generating function $S(z)$
553: of Catalan numbers $D_n \equiv {}_{2n}C_{n} / (n+1)$ \cite{Stanleybook}
554: as follows:
555: %
556: \begin{equation}
557: S(z) = \frac{1-\sqrt{1-4z}}{2z}
558: = \frac{d-1}{dz}Q\left(\frac{d^2}{d-1}z\right).
559: \label{eq:S_Q}
560: \end{equation}
561: %
562: Accordingly, we obtain
563: %
564: \begin{equation}
565: q_n = \frac{(d-1)^{n-1}}{d^{2n-1}}D_{n-1}.
566: \label{eq:q_n_cayley1}
567: \end{equation}
568: %
569: On the other hand, applying $m_l = (d-1)^{l-1} / d^l$ to
570: \EQ(\ref{eq:q_n_final}) results in
571: %
572: \begin{eqnarray}
573: q_n &=& \sum_{\lambda \vdash 2n-1, \sum^{n}_{l=1}i_{\lambda}(l) = n}
574: c^{(n)}_{\lambda} \frac{(d-1)^{\sum^{n}_{l=1}(l-1) i_{\lambda}(l) }}
575: {d^{\sum^n_{l=1}l i_{\lambda}(l)}}\nonumber\\
576: %
577: &=& \frac{(d-1)^{n-1}}{d^{2n-1}}
578: \sum_{\lambda \vdash 2n-1, \sum^{n}_{l=1}i_{\lambda}(l) = n}
579: c^{(n)}_{\lambda}.
580: \label{eq:q_n_cayley2}
581: \end{eqnarray}
582: %
583: Combining
584: \EQS(\ref{eq:q_n_cayley1}) and (\ref{eq:q_n_cayley2}) provides a
585: useful by-product:
586: %
587: \begin{equation}
588: \sum_{\lambda \vdash 2n-1, \sum^{n}_{l=1}i_{\lambda}(l) = n}
589: c^{(n)}_{\lambda} = D_{n-1},
590: \end{equation}
591: %
592: which states that
593: the sum of the coefficients in the moment expansion of
594: $q_n$ [see \EQ(\ref{eq:q_n_final})] is always equal to $D_{n-1}$
595: without regard to the distribution $\{p_k\}$.
596:
597: Similarly, \EQ(\ref{eq:R_Lagrange}) becomes
598: %
599: \begin{equation}
600: R(z) = \frac{2-d+\sqrt{d^2-4(d-1)z}}{2(1-z)}.
601: \label{eq:R_cayley}
602: \end{equation}
603: %
604: Then, it follows that
605: %
606: \begin{equation}
607: \left(1-\frac{d^2}{d-1}z\right) R\left( \frac{d^2}{d-1}z \right)
608: = 1- d z S(z)
609: \end{equation}
610: %
611: and
612: %
613: \begin{equation}
614: r_n = 1 - \sum^{n-1}_{n^{\prime}=0}
615: \left(\frac{(d-1)^{n^{\prime}-1}}{d^{2n^{\prime}-1}}\right) D_{n^{\prime}}.
616: \end{equation}
617: %
618: Owing to the entire homogeneity of the Cayley trees,
619: \EQS(\ref{eq:Q_cayley2}) and (\ref{eq:R_cayley}) are exact in this case
620: and agree with the theoretical results obtained by identifying
621: random walk on the Cayley trees with unbiased random walk on
622: ${\bf Z}$ \cite{Spitzer,Grimmettbook}.
623:
624: \subsection{Erd\"{o}s-R\'{e}nyi random graph}\label{sub:er}
625:
626: The Erd\"{o}s-R\'{e}nyi (ER) random graph is generated by independently
627: assigning an edge with probability $p$ between any possible pairs
628: of vertices
629: \cite{Erdos,Albert02,Newman}. If the number of vertices $N$ scales so
630: that $\lambda \equiv N p$ converges in the limit $N\to\infty$, the
631: vertex degree is distributed as specified by the Poiss\'{o}n
632: distribution, namely,
633: %
634: \begin{equation}
635: p_k = \frac{\lambda^k}{k!} {\rm e}^{-\lambda}.
636: \end{equation}
637:
638: Numerically calculated distributions of the first return time are
639: indicated by circles
640: in \FIGS\ref{fig:er}(a) and \ref{fig:er}(b) for
641: $\lambda = 7$ and $\lambda = 10$, respectively.
642: The return probability decreases exponentially
643: in $n$ analogous to the case of the Cayley trees indicated by
644: solid lines in \FIG\ref{fig:er}
645: \cite{Spitzer,Schinazibook,Grimmettbook}. Then many sample points
646: are required for reliable estimation of the return time probability,
647: for which reason we construct the probability distributions based on
648: $5\times 10^7$ runs. A new random graph is created in each run.
649:
650: Distributions predicted by the theory in \SEC\ref{sec:return} are
651: indicated by crosses in \FIG\ref{fig:er}. The theoretical estimates agree
652: with the numerical results better when $\lambda=10$. This is
653: because our method works better for networks
654: with a larger mean vertex degree, which is equal to $\lambda$.
655: However,
656: the error is bearable in both cases
657: for sufficiently small $n$
658: for which the numerical distributions
659: are calculated based on enough sample points.
660: In other words, the minimum positive probability
661: obtained by the simulations is $1/(5\times 10^7) = 2\times
662: 10^{-8}$, and the numerically
663: estimated probabilities are not reliable around this
664: value where statistical fluctuation counts. Related to this remark,
665: \FIG\ref{fig:er} shows that the numerical results are actually
666: available just
667: up to small values of $n$, that is, $n\le 17$ for $\lambda=7$ and
668: $n\le 12$ for $\lambda=10$. As noted before, this is due to the
669: exponential decay in the return time distribution. Furthermore, the
670: decay is faster for a larger mean vertex degree, or a larger $\lambda$,
671: which more severely constrains the practical
672: upper limit of $n$ for which the
673: distribution is obtained. Compared with the cumbersome
674: brute-force method, our method needs
675: only calculation of partition of integers,
676: which are much more numerically feasible.
677:
678: Figure~\ref{fig:er} also shows, both for $\lambda=7$ and $\lambda=10$,
679: that the decay of the first return time probability is slower for the
680: ER random graphs than for the Cayley trees
681: with the same mean vertex degree.
682: This is presumably because of the dispersion of vertex degree in the
683: ER random graph, as we
684: discuss in \SEC\ref{sec:conclusions}.
685:
686:
687: \subsection{Scale-free networks}
688:
689: The vertex degrees of real networks often have power-law
690: distributions. Barab\'{a}si and co-workers presented a network
691: growth model with preferential attachment to generate such a graph
692: \cite{SW-SF,Albert02}. In their scale-free networks, the vertex degree
693: has a lower cutoff $m$, and the degree distribution is represented by
694: $p_k = {\cal N} k^{-3}$ ($k\ge m$) and $p_k=0$ ($k<m$), where ${\cal
695: N}$ is the normalization constant. The first return time
696: probabilities of random walk on scale-free random graphs with
697: $m=4$ are shown in \FIG\ref{fig:sf}, suggesting that the theory (crosses) again
698: predicts the numerical results (circles) in a satisfactory manner. In
699: this case, the mean vertex degree is numerically calculated to be
700: 7.09. Accordingly, the results for the Cayley trees with $d=7$ (solid
701: lines) and $d=8$ (dotted lines) are also shown in \FIG\ref{fig:sf} for
702: comparison. The probability of the first return time decays slower
703: for the scale-free networks, as has also been the case for the ER
704: random graphs. Moreover, comparison of \FIGS\ref{fig:er} and
705: \ref{fig:sf} reveals that the discrepancy from the regular case, which
706: is probably caused by the heterogeneous vertex degree, is larger for
707: the scale-free networks. This is
708: presumably because the vertex degree is more heterogeneous in the
709: scale-free networks than in the ER random graphs.
710:
711: Random walk on other related graphs,
712: such as ones whose degree distributions have
713: power laws without the lower cutoff, power laws with exponential higher cutoff, or
714: simple exponential decay \cite{Callaway00,Newman}, can be analyzed
715: similarly. The only caveat is that the theory is likely
716: to fail when the vertex degree is fairly
717: small on average. Let us also mention that there is little hope for
718: obtaining more tractable analytical expressions for $Q(z)$ and $R(z)$
719: even in simpler scale-free cases, because the
720: polylogarithm functions, which can be estimated only numerically
721: \cite{Newman}, appear in the calculation of $m_l$.
722:
723: \section{Conclusions}\label{sec:conclusions}
724:
725: In this paper, we have derived for generalized random networks the
726: analytic expressions for the probability distributions of first
727: and general return times. Our methods correctly predict the
728: numerical results as far as the mean vertex degree is not extremely
729: small. They are also useful in saving the computation time and hence
730: obtaining return time probabilities on a much longer time scale than
731: with straightforward simulations. This merit stems from the fact
732: that the algorithm for calculating partition of integers is easily
733: implemented \cite{Stanleybook}, whereas brute-force methods require
734: billions of runs to obtain the distributions and the asymptotics, particularly
735: in the case of exponentially decaying tails.
736:
737: We have also found that heterogeneous graphs such as
738: the ER random graphs and the scale-free networks yield slower
739: decay of return time probabilities than the Cayley trees with
740: the corresponding vertex degrees. The decay rate is closely linked
741: to critical phenomena and phase transitions of both static
742: \cite{Lyons90_Lyons95} and dynamical
743: \cite{Liggettbook99,Schinazibook,Pemantle01,Durrettbook} particle
744: systems. In social contexts, information and diseases are actually
745: suggested to propagate in a manner different from as we imagine by the
746: analogy of regular graphs such as the Cayley trees and regular
747: lattices. For example, percolation is more likely to occur in networks with
748: heterogeneous vertex degrees \cite{Newman}. Also for
749: dynamical processes such as contact
750: processes \cite{Liggettbook99,Schinazibook,Eguiluz_Pastor_Masuda_SARS}
751: and voter models \cite{Liggettbook99,Durrettbook},
752: occurrence of global orders such as
753: epidemics or unanimity has the same
754: tendency. Mathematically, the problem of the global orders emerging in these
755: dynamics can be associated with that of the dual or related
756: processes. For example, if simple and branching random walks
757: (resp. coalescing random walks) are more likely to return to the
758: origin, the critical value for phase transition becomes smaller, and the
759: probability of a global epidemics or
760: unanimity becomes larger in contact processes
761: \cite{Liggettbook99,Schinazibook} (resp. voter models
762: \cite{Liggettbook99,Durrettbook}). Accordingly, the asymptotic behavior
763: of random walk reported in \SEC\ref{sec:examples} suggests that global
764: orders are more likely consequences in networks with heterogeneous
765: vertex degrees such as scale-free and ER random networks. This
766: evidence substrates the results for the contact processes in
767: epidemic contexts \cite{Eguiluz_Pastor_Masuda_SARS} and poses a
768: dynamical version of the exact results on percolation \cite{Newman}.
769:
770: As for exact asymptotic behavior, questions about
771: the Cayley trees with vertex degree $d$
772: is translated into ones about the unbalanced random walk on
773: ${\bf Z}$, the analysis of which easily resulting in $r_n\propto n^{-3/2}
774: \left(2\sqrt{d-1} / d\right)^{2n}$ \cite{Schinazibook}. To
775: illuminate the asymptotic behavior of $q_n$ and $r_n$ in the case of
776: generalized random walks is an important subject of future work.
777:
778: \begin{acknowledgements}
779: We thank H. Kesten and G. F. Lawler for helpful comments.
780: This study is supported by the Grant-in-Aid for Scientific Research
781: (JSPS) and the Grant-in-Aid for Scientific Research (B)
782: (Grant No. 12440024) of the Japan Society of the Promotion of Science.
783: \end{acknowledgements}
784:
785: \begin{references}
786:
787: \bibitem{Spitzer}
788: F. Spitzer, {\it Principles of Random Walk}, 2nd ed.
789: (Springer-Verlag, New York, 1976).
790:
791: \bibitem{Liggettbook99}
792: T. M. Liggett, {\it Stochastic Interacting Systems: Contact, Voter and
793: Exclusion Processes} (Springer-Verlag, Berlin, 1999).
794:
795: \bibitem{Schinazibook}
796: R. B. Schinazi, {\it Classical and Spatial Stochastic
797: Processes} (Birkha\"{u}ser, Boston, 1999).
798:
799: \bibitem{Pemantle01}
800: R. Pemantle and A. M. Stacey,
801: % The branching random walk and contact process on Galton-Watson
802: % and nonhomogeneous trees.
803: Ann. Prob. {\bf 29}, 1568 (2001).
804: %(4), --1590
805:
806: \bibitem{Durrettbook}
807: R. Durrett, {\it Lecture Notes on Particle Systems
808: and percolation} (Wadsworth, Belmont, CA, 1988).
809:
810: \bibitem{Lyons90_Lyons95}
811: R. Lyons,
812: % Random walks and percolation on trees.
813: Ann. Prob. {\bf 18}, 931 (1990);
814: %(3), --958
815: %
816: R. Lyons, R. Pemantle, and Y. Peres,
817: % Ergodic theory on Galton-Watson trees: speed of random walk
818: % and dimension of harmonic measure.
819: Ergod. Theory Dyn. Syst. {\bf 15}, 593 (1995).
820: %--619
821:
822: \bibitem{Erdos}
823: P. Erd\"{o}s and A. R\'{e}nyi,
824: % On random graphs.
825: Publ. Math. (Debrecen) {\bf 6}, 290 (1959).
826: %--297 Publicationes Mathematicae
827:
828: \bibitem{SW-SF}
829: D. J. Watts and S. H. Strogatz,
830: %Collective dynamics of `small-world' networks.
831: Nature (London) {\bf 393}, 440 (1998);
832: %--442
833: %\bibitem{Wattsbook}
834: D. J. Watts, {\it Small Worlds} (Princeton University Press,
835: Princeton, 1999);
836: %
837: A.-L. Barab\'{a}si and R. Albert,
838: %Emergence of scaling in random networks.
839: Science {\bf 286}, 509 (1999);
840: %--512
841: %
842: S. H. Strogatz,
843: % Exploring complex networks.
844: Nature (London) {\bf 410}, 268 (2001);
845: %--276
846: %
847: A.-L. Barab\'{a}si, {\it Linked: the New Science of Networks}
848: (Perseus Publishing, New York, 2002).
849:
850: \bibitem{Callaway00}
851: D. S. Callaway, M. E. J. Newman, S. H. Strogatz, and D. J. Watts,
852: % Network robustness and fragility: percolation on random graphs.
853: Phys. Rev. Lett. {\bf 85}, 5468 (2000).
854: %(25), --5471
855:
856: \bibitem{Albert02}
857: R. Albert and A.-L. Barab\'{a}si,
858: % Statistical mechanics of complex networks.
859: Rev. Mod. Phys. {\bf 74}, 47 (2002).
860: %--97
861:
862: \bibitem{Newman}
863: M. E. J. Newman, S. H. Strogatz, and D. J. Watts,
864: % Random graphs with arbitrary
865: % degree distributions and their applications.
866: Phys. Rev. E {\bf 64}, 026118 (2001);
867: %
868: M. E. J. Newman,
869: % The structure and function of complex networks.
870: SIAM Rev. {\bf 45}, 167 (2003).
871: % --256
872:
873: \bibitem{Eguiluz_Pastor_Masuda_SARS}
874: R. Pastor-Satorras and A. Vespignani,
875: %Epidemic spreading in scale-free networks.
876: Phys. Rev. Lett. {\bf 86}, 3200 (2001);
877: %(14) --3203
878: %
879: V. M. Egu\'{\i}luz and K. Klemm,
880: % Epidemic threshold in structured scale-free networks.
881: {\it ibid.} {\bf 89}, 108701 (2002);
882: %(10)
883: %
884: N. Masuda, N. Konno, and K. Aihara,
885: % Transmission of severe acute respiratory syndrome
886: % in dynamical small-world networks.
887: Phys. Rev. E {\bf 69}, 031917 (2004).
888:
889: \bibitem{Pandit}
890: S. A. Pandit and R. E. Amritkar,
891: % Random spread on the family of small-world networks.
892: Phys. Rev. E {\bf 63}, 041104 (2001).
893: %
894:
895: \bibitem{Jasch_Lahtinen}
896: F. Jasch and A. Blumen,
897: % Trapping of random walks on small-world networks.
898: Phys. Rev. E {\bf 63}, 066104 (2001);
899: %
900: J. Lahtinen, J. Kert\'{e}sz, and K. Kaski,
901: % Scaling of random spreading in small world networks.
902: {\it ibid.} {\bf 64}, 057105 (2001).
903:
904: \bibitem{Szabo00rw}
905: G. Szab\'{o},
906: % Branching annihilating random walk on random regular graphs.
907: Phys. Rev. E {\bf 62}, 7474 (2000).
908: %(5), --7477
909:
910: \bibitem{Farkas_Goh}
911: I. J. Farkas, I. Der\'{e}nyi, A.-L. Barab\'{a}si, and T. Vicsek,
912: % Spectra of ``real-world'' graphs: beyond the semicircle law.
913: Phys. Rev. E {\bf 64}, 026704 (2001);
914: %
915: %
916: K.-I. Goh, B. Kahng, and D. Kim,
917: % Spectra and eigenvectors of scale-free networks.
918: {\it ibid.} {\bf 64}, 051903 (2001).
919:
920: \bibitem{Zeitouni02}
921: O. Zeitouni,
922: % Random walks in random environments.
923: in {\it Proceedings of the International Congress of Mathematicians}
924: (Higher Education Press, Beijing, 2002), Vol. III, p. 117.
925: %--127
926:
927: \bibitem{Grimmettbook}
928: G. R. Grimmett and D. R. Stirzaker,
929: {\it Probability and Random Processes},
930: 2nd ed. (Oxford University Press, Oxford,
931: 1992).
932:
933: \bibitem{Stanleybook}
934: R. P. Stanley,
935: {\it Enumerative Combinatorics}
936: (Cambridge University Press, Cambridge, 1999), Vol. 2.
937:
938: \end{references}
939:
940: % \newpage
941:
942: \bigskip
943: \bigskip
944:
945: Figure captions
946:
947: \bigskip
948:
949: Figure 1: Schematic diagram showing random walk on a realization of
950: generalized random graph. Integers denotes the time of random walk.
951:
952: \bigskip
953:
954: Figure 2: Probability distributions of the first return times of the
955: random walk on the ER random graph with (a)
956: $\lambda = 7$ and (b) $\lambda = 10$. Numerical and theoretical
957: results are indicated by circles and crosses, respectively. The results
958: for the Cayley trees with the same
959: mean vertex degrees, namely, (a) $d=7$ and (b) $d=10$,
960: are indicated by solid lines.
961:
962: \bigskip
963:
964: Figure 3: Probability distributions of the first return times in the
965: case of scale-free networks with $m=4$. Numerical and theoretical
966: results are indicated by circles and crosses, respectively.
967: The results for the Cayley trees with
968: $d=7$ (solid lines) and $d=8$ (dashed lines) are also shown.
969:
970: % \clearpage
971:
972: \bigskip
973: \bigskip
974:
975: \begin{figure}[b]
976: \begin{center}
977: \includegraphics[height=2.25in,width=3.25in]{rw_pic.eps}
978: \caption{}
979: \label{fig:rw_pic}
980: \end{center}
981: \end{figure}
982:
983: % \clearpage
984:
985: \begin{figure}
986: \begin{center}
987: \includegraphics[height=2.25in,width=3.25in]{er_a.eps}
988: \includegraphics[height=2.25in,width=3.25in]{er_b.eps}
989: \caption{}
990: \label{fig:er}
991: \end{center}
992: \end{figure}
993:
994: % \clearpage
995:
996: \begin{figure}
997: \begin{center}
998: \includegraphics[height=2.25in,width=3.25in]{sf.eps}
999: \caption{}
1000: \label{fig:sf}
1001: \end{center}
1002: \end{figure}
1003:
1004: \end{document}
1005: