1: %\documentclass[preprint,showpacs,pre]{revtex4}
2: %\documentclass[twocolumn,pre]{revtex4}
3: \documentclass[pre,floatfix,twocolumn,showpacs,a4paper]{revtex4}
4: %\documentclass[a4paper,aps,pre,floats,preprint,superscriptaddress,showpacs]{revtex4}
5:
6: \usepackage{graphicx}
7: \usepackage{dcolumn}
8: \usepackage{amsmath}
9: \usepackage{amssymb}
10: \usepackage{tabularx}
11: \usepackage{epsfig}
12: \usepackage{mathrsfs}
13: \usepackage{rotating}
14: % Line below added by JP.
15: %\usepackage[applemac]{inputenc}
16:
17: \begin{document}
18: \title{Master-equation analysis of accelerating networks}
19: \author{David M.D. Smith$^{1,2,3,4}$}
20: \email{d.smith3@physics.ox.ac.uk}
21: \author{Jukka-Pekka Onnela$^{2,3,5}$}
22: \author{Nick S. Jones$^{2,3,4}$}
23: \affiliation{$^{1}$Centre for Mathematical Biology, Oxford University, Oxford OX1 3LB, U.K.}
24: \affiliation{$^{2}$Department of Physics, Clarendon Laboratory, Oxford University, Oxford OX1 3PU, U.K.}
25: \affiliation{$^{3}$CABDyN Complexity Centre, Oxford University, Oxford OX1 1HP, U.K.}
26: \affiliation{$^{4}$Oxford Centre for Integrative Systems Biology, Department of Biochemistry,
27: Oxford University, South Parks Road, Oxford, OX1 3QU, U.K.}
28: \affiliation{$^{5}$DBEC, Helsinki, University of Technology, P.O. Box 9203, FIN-02015 HUT, Finland.}
29:
30:
31: \date{\today}
32: \begin{abstract}
33: In many real-world networks, the rates of node and link addition are time dependent. This observation motivates the definition of
34: \emph{accelerating networks}. There has been relatively little investigation of accelerating networks and previous efforts at analyzing their degree distributions have employed mean-field techniques. By contrast, we show that it is possible to apply a master-equation approach to such network development. We provide full time-dependent expressions for the evolution of the degree distributions for the canonical situations of random and preferential attachment in networks undergoing constant acceleration. These results are in excellent agreement with results obtained from simulations. We note that a growing, non-equilibrium network undergoing constant acceleration with random attachment is equivalent to a classical random graph, bridging the gap between non-equilibrium and classical equilibrium networks.
35: \end{abstract}
36: \pacs{89.75.Fb, 89.75.Hc, 05.40.-a}%check
37: \maketitle
38:
39: \section{Introduction}
40: In many real-world evolving networks the rates of node and link addition differ, a phenomenon that gives rise to the concept of \emph{network acceleration}~\cite{GM, AcceleratingNetworks, Zhang:2007,
41: Sen}. Examples include the Internet~\cite{internet, Internet2, Vazquez2}, the World Wide Web~\cite{WWW} and collaboration
42: networks~\cite{collaboration} and these have prompted some recent scrutiny~\cite{GM, AcceleratingNetworks, Zhang:2007, Sen}. However, the majority of network research to date has overlooked
43: acceleration, with many non-equilibrium network models focusing on the rather specific scenario of adding a fixed (expected)
44: number of nodes at each timestep with a fixed (expected) number of links to the existing network~\cite{NewmanReview, Dorogovtsev,
45: ABReview}.
46:
47: Accelerating networks also appear in biology. The evolution of gene regulatory networks has been interpreted as a form of network acceleration~\cite{GM, mattick:2005}. Studies have shown that the number of regulatory genes scales quadratically with genome size in prokaryotic microorganisms \cite{Nimwegen, Croftetal}. In these networks, the number of links also scales quadratically with the number of nodes (operons/genes)~\cite{mattick:2005}. Indeed, gene regulatory networks have provided significant motivation for the investigation and definition of accelerating networks~\cite{GM, mattick:2005, Croftetal, Failed, SizeConstraints}. Gagen and Mattick proposed that the apparent accelerating nature of gene regulatory networks in prokaryotic organisms imposes an inherent size constraint on the system and this is consistent with empirical observations~\cite{SizeConstraints, Ahnert}.
48:
49: For an evolving (unweighted) network, acceleration is related to the rates of link and node addition~\footnote{The definitions we use here can be easily extended to incorporate weighted links~\cite{AcceleratingNetworks}.}. Consider a network comprising $N(t)$ nodes and $M(t)$ links at some time $t$. We now add $n(t)$ nodes and $m(t)$ links to the system such that, at the next timestep, $N(t+1)=N(t)+n(t)$ and $M(t+1) = M(t)+m(t)$.
50: Some previous investigations of accelerating networks have assumed that the rate of node addition is $n(t) = 1$ and that the
51: expected rate of link addition is given by some power-law function of time $\langle m(t) \rangle \propto
52: t^{\beta}$, where the exponent $\beta$ describes the nature of the acceleration~\cite{GM, Zhang:2007, Sen}. Indeed, Gagen and
53: Mattick defined network acceleration in terms of this parameter~\footnote{The definition of Gagen and Mattick suggested the
54: functional form $\langle m(t) \rangle \propto N(t)^{\beta}$ which, for the addition of one node per timestep, has $N(t)\sim t$
55: and this is the scenario they analyzed in Ref.~\cite{GM}.}. However, Smith \emph{et al.} proposed a more conventional definition of network
56: acceleration that is a time-dependent property of the system~\cite{AcceleratingNetworks}.
57: For an unweighted network, acceleration is simply related to the rates of node and link addition, more specifically, the time derivative of the velocity-like quantity $m(t)/n(t)$. For a deterministic system (as in an empirically observed network), network acceleration can be written~\cite{AcceleratingNetworks}
58: \begin{eqnarray}\label{eqn:acc1}
59: a(t) &=& \frac{M(t+1)-M(t)}{N(t+1)-N(t)} -\frac{M(t)-M(t-1)}{N(t)-N(t-1)}\nonumber\\
60: {}&=&\frac{m(t)}{n(t)}-\frac{m(t-1)}{n(t-1)}.
61: \end{eqnarray}
62: When considering a stochastic system, the definition is more delicate~\cite{AcceleratingNetworks}. For the scenario in which the
63: number of added nodes per timestep is constant ($n(t)\equiv n$) \emph{and} the expected number of links added per timestep is independent
64: of the particular number added at the previous timestep, we can rewrite Eq.~(\ref{eqn:acc1}) for a stochastically evolving accelerating network
65: as~\cite{AcceleratingNetworks}
66: \begin{eqnarray}\label{eqn:acc2}
67: \langle a(t) \rangle &=& \frac{\langle m(t)\rangle - \langle m(t-1)\rangle}{n}.
68: \end{eqnarray}
69: This definition of network acceleration was supported by a case study of the evolution of Wikipedia that undergoes different accelerating regimes throughout its evolution~\cite{AcceleratingNetworks} and this is the definition we shall adopt in the following.
70:
71: In Section~\ref{sec:limitations} we review the use of an existing mean-field technique to analyze accelerating networks and discuss its limitations. In Sections~\ref{sec:constaccRA} and \ref{sec:constaccBA} we consider an alternative master-equation method and demonstrate its relevance by applying it to the scenarios of constant acceleration with random and preferential
72: attachment. We derive full, time-dependent solutions for the evolution of the degree distributions of these networks and compare them
73: to simulated networks. We conclude in Section \ref{sec:conc}.
74:
75: \section{Limitations of some mean-field techniques}\label{sec:limitations}
76: For non-accelerating, non-equilibrium networks there are two conventional methods to obtain degree distributions: the \emph{mean-field} and \emph{master-equation} approaches~\cite{NewmanReview, Dorogovtsev, ABReview, BarabasiDeriv2}. The former assumes that both node degree and time are continuous and the latter uses a continuous time approximation.
77:
78: The mean-field approach tracks the evolution of the degree of an individual node throughout the network's evolution. The analysis consists of two distinct stages. The first stage is to derive the expected degree of an individual node in the network at a particular time, based upon the time at which the node was added. At each timestep, a single node is added to the system. The process assumes that a node's degree is continuous \emph{and} that links are added continuously throughout the network's evolution. The expected rate of change of the degree of an individual node is related to the probability of it receiving links throughout the dutation of the network's evolution. At any given stage during the network's evolution, a node added at a particular time will have a certain expected degree. This is used to generate a continuous expected degree as a function of the time at which nodes are added to the system. The inverse of this function provides a time of addition to the system corresponding to that expected degree. The second part of the process is to derive a continuous degree distribution for the resulting network. In many growing network models, old nodes tend to have higher degrees as they have longer to accumulate new links. Consider selecting a node at random in such a network. The probability that it has expected degree greater than some value $k_{\alpha}$ is simply the probability that it was added before the time corresponding to expected degree $k_{\alpha}$. A cumulative degree distribution is thus obtained by assuming the time at which a randomly selected node was added to the network will be uniformly randomly distributed across the timescale of the network's evolution. The continuous degree distribution is then derived by differentiating the cumulative degree distribution.
79:
80: This was the technique used by Barab\'asi {\it et al.}~\cite{Barabasi,BarabasiDeriv2} and it has also been applied to accelerating networks
81: \cite{Zhang:2007, collaboration, topology, tunabledegree}. This method was also adopted by Gagen and Mattick for analyzing accelerating networks~\cite{GM}. However, this technique can be inappropriate in certain situations. To highlight the shortcomings of this approach, we consider
82: its application to the scenario of constant acceleration such that the acceleration is $a(t) \equiv a$ ~\footnote{In the investigation of Gagen and Mattick, this scenario was described as a `hyper-accelerating' network~\cite{GM}.} and that the continuous (deterministic) rate of link addition is $m(t) = at$. For an initial seed comprising $N(0)$ nodes and $M(0)$ links, the total number of nodes in the system is $N(t)=N(0) + t \sim t$, the total number of links is $M(t) = M(0)+\int_{0}^{t} m(t) \textrm{d}t \sim \frac{at^2}{2}$ and the mean degree of the network is $2 M(t) /N(t) \sim at$ (where $\sim$ means asymptotically equal to). Consider the case of random attachment in which all new (undirected) links are made between newly introduced nodes and randomly selected nodes within the existing network and no connections are established between existing nodes (leading to the constraint $a(t)< 1$ as the number of newly introduced links cannot exceed the number of existing nodes). Using the mean-field technique, the continuous rate of change of the (continuous) expected degree of some node $\alpha$ can be expressed in terms of the probability that it acquires new links~\cite{BarabasiDeriv2}:
83: \begin{eqnarray}\label{eqn:meanfieldrandom1}
84: \frac{\textrm{d}\langle k_{\alpha}(t)\rangle}{\textrm{d}t} &\sim& \frac{ m(t)}{t}.
85: \end{eqnarray}
86: Integrating from the time at which the node was added to the system $t_{\alpha}$ to time $t$ yields the expected degree of
87: this node:
88: \begin{eqnarray}\label{eqn:meanfieldrandom2}
89: \langle k_{\alpha}(t) \rangle &\sim& at.
90: \end{eqnarray}
91: The expected degree of node $\alpha$ is asymptotically independent of the time $t_{\alpha}$ at which it was added so the expected degree of all nodes is $\langle k(t) \rangle \sim at$. It is argued in Ref.~\cite{GM} that the continuous degree distribution derived from the mean-field approach for this system is, therefore, a delta function at degree $at$.
92:
93: For the scenario of preferential attachment, the ends of the new links associated with new nodes connect to existing nodes preferentially according to their degree. Again, using the mean field technique, the
94: rate equation describing the evolution of the degree of some node $\alpha$ at time $t$ for this process can be expressed as
95: \begin{eqnarray}\label{eqn:meanfieldpref1}
96: \frac{\textrm{d}\langle k_{\alpha} (t) \rangle}{\textrm{d}t} &\sim& \frac{m(t) k_{\alpha}(t)}{2M(t)}.
97: \end{eqnarray}
98: Integrating between the appropriate limits provides the same expected degree $\langle k_{\alpha}(t) \rangle \sim at$ which is
99: asymptotically independent of the time the node was added and, consequently, the same for all nodes in the system. Again, it is argued in Ref.~\cite{GM} that the mean-field technique produces a delta function for the continuous degree distribution at $at$.
100:
101: Gagen and Mattick provide a method to retrieve a discrete degree distribution from a continuous one~\cite{GM}. However,
102: because the mean-field technique provides identical continuous degree distributions, the outcome will be the same for both random
103: and preferential attachment with constant acceleration (equations (18) and (31) in \cite{GM}). In practice, however, simulations
104: of these networks generate very different degree distributions as illustrated in Fig.~\ref{fig:prefrandattach}. Concerns over the above treatments motivate the approach below.
105:
106:
107: \section{Random attachment with constant acceleration}\label{sec:constaccRA}
108: We now employ a master-equation method to derive the degree distribution of a random attachment network undergoing constant acceleration.
109: The method was introduced by Krapivsky \emph{et al.}~\cite{KrapivskyDeriv} and independently by Dorogovtsev {\it et al.}~\cite{DorogovtsevDeriv} to derive the steady-state degree distribution of the Barab\'asi-Albert preferential attachment mechanism~\cite{NewmanReview}. A similar technique based upon distribution kinetics has been applied to accelerating networks~\cite{Jeon} and, recently, a master-equation method was developed to investigate the emergence of correlations within evolving networks~\cite{LinkSpaceNetworkAnalysis}. Justifications for using such an approach are discussed in Refs.~\cite{history} and ~\cite{Boll}.
110:
111: Consider the scenario of constant acceleration, $\langle a(t)\rangle = a$ (assumed $\ll 1$), for a stochastic accelerating
112: network in which the rate of node addition per timestep is constant at $n(t)=n = 1$ and the expected number of added
113: (undirected) links per timestep is $\langle m(t) \rangle = at$. Since $N(t)\sim t$ and $a\ll 1$, it follows that the expected number of new links added per pre-existing node per timestep ($\frac{\langle m(t) \rangle}{N(t)}$) is always small.
114: We consider the example of random attachment and all the newly
115: introduced links are established between the new node and the existing network. There are no links formed between existing nodes.
116: The process is governed by the attachment probability kernel $\Theta_k(t)$, defined as the probability that a specific, newly
117: introduced link connects to an existing node of degree $k$. At some time $t$, there exist $X_k(t)$ nodes of degree $k$ and we wish
118: to compute the expected number of nodes of degree $k$ at time $t+1$. The fraction of nodes of degree $k$ is $c_k(t) =
119: X_k(t)/N(t)= X_k(t)/[N(0)+t]\sim X_k(t)/t$. It is assumed that the seed component comprising $N(0)$ nodes is small. The master equation for the evolution of $X_k$ can be expressed in terms of the attachment kernel
120: and is written
121: \begin{eqnarray}
122: \langle X_k (t+1) \rangle{}& =& X_k (t) + \langle m(t) \rangle\Theta_{k-1}(t)\nonumber\\
123: {}&{}&-\langle m(t) \rangle\Theta_{k}(t)+P\{m(t)=k\}.
124: \label{eqn:nodeXmasterACC}
125: \end{eqnarray}
126: The second term on the right-hand side reflects the expected number of connections to $k-1$ degree nodes making them nodes of
127: degree $k$. The last term is the probability that the new node itself is a node of degree $k$.
128:
129: To establish $P\{m(t)=k\}$, it is important that the microscopic mechanism governing the network's evolution be specified. In
130: simulating such a network, we proceed as follows: links between the new node and \emph{all} $N(t)$ existing nodes are
131: considered and established with probability $a$ satisfying the requirements for the expected number of links for the new node as
132: $\langle m(t) \rangle = at$. Interestingly, a network generated via this mechanism will also have the property that every
133: possible link between \emph{every pair of nodes} will exist with equal probability $a$. This is equivalent to the classical random
134: graph studied by Erd\H{o}s and R\'enyi~\cite{ER} and subsequently Bollob\'as~\cite{Boll2} with identical binomial degree
135: distribution. That is, \emph{a growing (non-equilibrium) network undergoing constant acceleration with random attachment is equivalent to the classical random graph,
136: bridging the gap between non-equilibrium and classical equilibrium networks}. This is illustrated in Fig.~\ref{fig:prefrandattach}
137: where a simulated, random attachment, stochastic, accelerating network with constant acceleration grown
138: to $N_{tot} = 10^6$ nodes is compared to the Poisson degree distribution $c_k = \frac{e^{-\lambda} \lambda^k}{k!}$ of the random
139: graph~\cite{ABReview} where $\lambda$ is the mean degree of the network. In Fig.~\ref{fig:prefrandattach}, the mean degree is
140: $\lambda=10$. The acceleration parameter is solved for by setting $\lambda = \frac{2M_{tot}}{N_{tot}}\sim \frac{at^2}{t}\sim aN_{tot}$ so the acceleration is $a = \lambda/N_{tot}$ which subsequently determines $\langle m(t)\rangle$.
141:
142: \begin{figure}
143: \begin{center}
144: %\includegraphics[width=0.75\textwidth]{test3.eps}
145: \includegraphics[width=0.45\textwidth]{test3.eps}
146: \caption{ \label{fig:prefrandattach} (Color online) Degree distributions for stochastic random attachment (RA) and preferential attachment (PA) networks undergoing constant acceleration. Each simulation comprises a network grown to $N_{tot}=10^6$ nodes with mean degree of $\lambda
147: =10 \approx a N_{tot}$. Also illustrated is the Poisson distribution with the same mean and the analytic expression of
148: Eq.~(\ref{eqn:caccfinal}) for the preferential attachment, constantly accelerating network derived in
149: Section~\ref{sec:constaccBA}. The numerical implementation simply generates the degree distribution through iterating the master equation,
150: Eq.~(\ref{eqn:caccPA1}). The non-accelerating degree distributions for random and preferential attachment (Eqs.~\ref{eqn:ranonacc} and~\ref{eqn:nonaccPA}) with the same mean ($m=5$ links added with each new node) are also illustrated to highlight the effect of accelerating behaviour.}
151: \end{center}
152: \end{figure}
153:
154: To apply a master-equation approach we start by defining the random attachment probability kernel, $\Theta_k(t) =
155: X_k(t)/N(t)=c_k(t)$ for an individual new link attaching to any node of degree $k$ within the existing network.
156: %The master equation for the expected number of nodes of degree $k$ can be written
157: %\begin{eqnarray}\label{eqn:acc5}
158: %\langle X_k(t+1)\rangle&=&X_k(t)+\frac{\langle m(t) \rangle X_{k-1}(t)}{N(t)}\nonumber\\
159: %{}&{}& -\frac{\langle m(t) \rangle X_{k}(t)}{N(t)} +P\{m(t)=k\}.
160: %\end{eqnarray}
161: For the random attachment mechanism outlined earlier, the number of new links added with the new node will be binomially distributed and, if we
162: assume that $t$ is large, we can make a Poisson approximation so that $P\{m(t)=k\} = \frac{e^{-a t}(a t)^k}{k!}$. Non-accelerating networks can have a steady-state degree distribution in the long-time limit and often it is this solution which is investigated from the master equations (for example see Ref.~\cite{LinkSpaceNetworkAnalysis}). However, for accelerating networks, the degree distribution evolves and a
163: time-dependent solution must be found. The fraction of nodes of degree $k$ is given by $c_k(t) = X_k(t)/N(t) \sim X_k(t)/t$
164: and, using the continuous time approximation $\langle X_k(t+1)\rangle-X_k(t)\approx\frac{\textrm{d}
165: \left(X_k(t)\right)}{\textrm{d} t} \sim \frac{\textrm{d} \left(tc_k(t) \right)}{\textrm{d} t}$, we can rewrite Eq.~(\ref{eqn:nodeXmasterACC})
166: as
167: \begin{eqnarray}\label{eqn:acc6}
168: \frac{\textrm{d} \left( tc_k(t) \right)}{\textrm{d} t}&=& a t c_{k-1}(t) -atc_k(t) + \frac{e^{-a t}(a t)^k}{k!}.
169: \end{eqnarray}
170: The time-dependent solution of Eq.~(\ref{eqn:acc6}) gives the degree distribution for this model and is simply
171: \begin{eqnarray}\label{eqn:poisson}
172: c_k(t)&=& \frac{e^{-a t}(a t)^k}{k!},
173: \end{eqnarray}
174: as would be expected for a random graph with mean degree of $\lambda \sim at$ in the large size limit~\cite{Boll2}. While this master-equation
175: approach does not necessarily imply that a given evolving network will converge upon the solution, in practice this is often the
176: case~\cite{Evans}. The Poisson solution of Eq.~(\ref{eqn:poisson}) is compared to a simulated network in Fig.~\ref{fig:prefrandattach}.
177:
178: The effect of the accelerating nature of the random attachment network can be appreciated by comparison with a conventional, non-accelerating scenario. The master-equation analysis, when applied to a non-accelerating random attachment network with $m(t) \equiv m$ ($m$ integer), gives a steady-state (long-time limit) degree distribution for $k\ge m$ of
179: \begin{eqnarray}\label{eqn:ranonacc}
180: c_k = \frac{1}{m+1}\left( \frac{m}{m+1} \right)^{k-m}.
181: \end{eqnarray}
182: This is illustrated in Fig.~\ref{fig:prefrandattach} for mean degree of $10$ corrsponding to $m=5$ links added with each new node.
183: The mean-field technique applied to the non-accelerating scenario produces the (less accurate) steady-state degree distribution $c_k \propto e^{1-\frac{k}{m}}/m$~\cite{BarabasiDeriv2}.
184:
185: \section{Preferential attachment with constant acceleration}\label{sec:constaccBA}
186: In this section, we apply the master-equation method to the scenario of a constantly accelerating network with linear preferential
187: attachment. Again, the network evolves through the addition of one new node per timestep, $n(t)=n = 1$ such that, for a (small) initial seed component of $N(0)$ nodes, the total number of nodes is $N(t)=N(0)+t\sim t$ and links are formed between this node and the existing network. With constant acceleration, $\langle a(t)\rangle = a$, the expected number of added links per timestep is $\langle m(t) \rangle = at$. For both simulation and
188: analysis, a clear description of the microscopic process is necessary. We assume that the desired resulting network is sparsely connected, i.e. the resultant mean degree is significantly less than the size of the network. As such, $\langle m(t)\rangle \ll N(t)$
189: (equivalent to $a\ll 1$) and the attachment process can be modeled as a series of Bernoulli trials. All links between the new node
190: and existing nodes in the network are considered. The probability of a link being formed between the new node to some existing
191: node $\varsigma$ is set to be ${\langle m(t)\rangle k_{\varsigma}}/{2 M(t)}$. By performing this random trial for all nodes within
192: the existing network, both preferential attachment and the expected number of new links added $\langle m(t)\rangle= at$ are
193: preserved. Nodes with zero links when introduced to the system will remain degree zero throughout.
194:
195: Analysis of this process starts by writing the master equation. For the purposes of analysis, it is assumed that the total number of links in the network is well approximated by its expected value, $M(t)\approx \langle M(t)\rangle$. Using the preferential attachment kernel, we have $\Theta_k(t) =
196: kX_k(t)/2\langle M(t)\rangle$ and the master equation for the evolution of the network can be written
197: \begin{eqnarray}\label{eqn:caccPA1}
198: \langle X_k(t+1)\rangle&=&X_k(t)+\frac{(k-1)\langle m(t) \rangle X_{k-1}(t)}{2\langle M(t)\rangle} \nonumber\\ {}&{}&
199: -\frac{k\langle m(t) \rangle X_{k}(t)}{2\langle M(t)\rangle} + P\{m(t)=k\}.
200: \end{eqnarray}
201: We derive the probability $P\{m(t)=k\}$ in a similar manner to the random attachment scenario. Consider the nodes of degree $j$ in the
202: existing network at time $t$. There are $X_j(t)$ of them. All $X_j(t)$ possible links are considered between the new node and
203: these existing nodes of degree $j$ and each is established with probability ${j \langle m(t)\rangle}/{2\langle M(t)\rangle}$.
204: Consequently, the number of new links between the new node and what were $j$-degree nodes will be binomially distributed with mean
205: ${j \langle m(t)\rangle X_j(t)}/{2\langle M(t)\rangle}$. The probability distribution for the degree of the new node will be a
206: convolution of all these binomial distributions corresponding to each value of $j$. However, if we assume that $X_j(t)$ is large
207: then we can make a Poisson approximation for each of these binomial distributions. The convolution of two Poisson distributions is a Poisson
208: distribution with a mean equal to the sum of the means of the two. As such, the probability distribution for
209: the degree of the new node can be approximated by a Poisson distribution. Of course, the mean of this is
210: simply $\langle m(t) \rangle$ and, consequently, $P\{m(t)=k\}$ is the same for both preferential and random attachment. Recalling that
211: $\langle M(t)\rangle \sim \frac{at^2}{2}$ and that the fraction of nodes of degree $k$ is given by $c_k(t) = X_k(t)/N(t) \sim X_k(t)/t$, the master equation of Eq.~(\ref{eqn:caccPA1}) for the evolution of a constantly accelerating network with preferential attachment can be rewritten as
212: \begin{eqnarray}
213: \frac{\textrm{d} \left( tc_k(t) \right)}{\textrm{d} t}&=& (k-1)c_{k-1}(t) -kc_k(t) + \frac{e^{-a t}(a t)^k}{k!}.
214: \end{eqnarray}
215: Multiplying both sides by the integrating factor $t^{k}$, this can be further simplified to
216: \begin{eqnarray}\label{eqn:caccME}
217: \frac{\textrm{d} \left( t^{k+1}c_k(t) \right)}{\textrm{d} t}&=& (k-1)t^{k}c_{k-1}(t) + \frac{e^{-a t}(a t^2)^k}{k!}.\nonumber\\
218: {}&{}&{}
219: \end{eqnarray}
220: The recursive nature of Eq.~(\ref{eqn:caccME}) means that the solution can be expressed in terms of repeated definite integrals
221: back to the expression for $c_1(t)$ . Each of these integrals equals zero when evaluated at the lower limit allowing the time-dependent degree distribution to be written (see Appendix~\ref{sec:derivation} for details) as
222: \begin{small}
223: \begin{eqnarray}\label{eqn:caccfinal}
224: c_k(t)&=&\frac{1}{t^{k+1}}\sum_{x=1}^{k}\frac{(k-1)!}{x!(x-1)!(k-x)!}\int_0^t (t-y)^{k-x}(ay^2)^x e^{-ay} \textrm{d}y \nonumber\\
225: {}&=&\sum_{x=1}^k \frac{(k-1)!(2x)!(at)^x}{x!(x-1)!(k+x+1)!} {_1F_1}[1+2x,2+k+x,-at],\nonumber\\
226: {}&{}&{}
227: \end{eqnarray}
228: \end{small}
229: where $_1F_1[\kappa,\rho,\phi]$ is the confluent hypergeometric function of the first kind, or Kummer's function, defined (for
230: integer $\kappa$ and $\rho$) as \cite{Abramowitz}
231: \begin{small}
232: \begin{eqnarray}
233: _1F_1[\kappa,\rho,\phi]&=&\frac{(\rho-1)!}{(\rho-\kappa-1)!(\kappa-1)!}\int_0^1 e^{\phi
234: z}z^{\kappa-1}(1-z)^{\rho-\kappa-1}\textrm{d}z.\nonumber\\
235: {}&{}&{}
236: \end{eqnarray}
237: \end{small}
238:
239: Comparison of the analytic solution Eq.~(\ref{eqn:caccfinal}) is made to a simulated network in Fig.~\ref{fig:prefrandattach}. The simulated network was seeded with two connected nodes and has acceleration $a = 10^{-5}$.
240:
241: To highlight the effects of network acceleration, the master-equation analysis, when applied to a non-accelerating preferential attachment network with $m(t) \equiv m$ ($m$ integer), provides a steady-state (long-time limit) degree distribution for $k\ge m$ of~\cite{Dorogovtsev}
242: \begin{eqnarray}\label{eqn:nonaccPA}
243: c_k = \frac{2m(m+1)}{k(k+1)(k+2)}.
244: \end{eqnarray}
245: This is illustrated in Fig.~\ref{fig:prefrandattach} for mean degree of $10$ corrsponding to $m=5$ links added with each new node.
246: The mean-field technique produces a power-law degree distribution, $c_k \propto \frac{m^2}{k^3}$, in the steady state for large $k$ ~\cite{BarabasiDeriv2}.
247: %It was grown to $N_{tot}=10^6$ nodes
248: %such that the final mean degree is $10$. Implementing the analytic expression Eq.~(\ref{eqn:caccfinal}) is intensive and
249: %while there are approximations available to aid computation~\cite{Approximations}, a numerically derived degree distribution can
250: %easily be generated simply by iterating the master equation Eq.~(\ref{eqn:caccPA1}) the required number of times the results of
251: %which are also illustrated in Fig.~\ref{fig:prefrandattach}.
252:
253:
254: \section{Conclusions}\label{sec:conc}
255: We have demonstrated that the master-equation method can be applied to accelerating networks, highlighting the importance of
256: specifying the microscopic processes taking place within the evolving network under scrutiny. We have provided full time-dependent
257: solutions of the evolving degree distributions for random and preferential attachment with constant
258: acceleration.
259: We note that the classical random graph of Erd\H{o}s and R\'enyi can be modeled as a non-equilibrium evolving network,
260: more specifically a constantly accelerating network with random attachment.
261:
262:
263: \textbf{Acknowledgments:} Smith and Jones are supported by EPSRC and BBSRC and Onnela by a Wolfson College Junior Research Fellowship (Oxford). Smith acknowledges funding from the European Union (MMCOMNET) for part of this research. We thank Felix Reed-Tsochas for comments and suggestions and one anonymous referee for bringing Ref.~\cite{Jeon} to our attention.
264:
265: %------------------------------------------------------------------
266: \appendix
267: \section{Analytic solution for constant acceleration with preferential attachment}\label{sec:derivation}
268: As detailed in Section~\ref{sec:constaccBA}, the master equation for the evolution of the degree distribution of a constantly
269: accelerating network with preferential attachment is given by Eq.~(\ref{eqn:caccME}). For clarity, it is useful to write the last term on the right hand side as a function:
270: \begin{eqnarray}
271: G_x(t) &=&\frac{e^{-a t}(a t^2)^x}{x!},
272: \end{eqnarray}
273: such that the master equation can be rewritten as
274: \begin{eqnarray}\label{eqn:appcaccME2}
275: \frac{\textrm{d} \left( t^{k+1}c_k(t) \right)}{\textrm{d} t}&=& (k-1)t^{k}c_{k-1}(t) + G_k(t).
276: %\frac{\textrm{d} \left( tc_k(t) \right)}{\textrm{d} t}&=& (k-1)c_{k-1}(t) -kc_k(t) + G(k,t).\nonumber\\
277: \end{eqnarray}
278: We can separate the variables and integrate Eq.~(\ref{eqn:appcaccME2}) to give
279: \begin{small}
280: \begin{eqnarray}
281: c_k(t)&=&\frac{1}{t^{k+1}} \Bigg\{ \int_0^t (k-1) {t_1}^{k} c_{k-1}({t_1})\textrm{d}{t_1} + \int_0^{t} G_k({t_1})\textrm{d}{t_1} \Bigg\}.\nonumber
282: \end{eqnarray}
283: \end{small}
284: Making the substitution $k=k'+1$ yields
285: \begin{small}
286: \begin{eqnarray}
287: c_{k'+1}(t)&=&\frac{1}{t^{k'+2}} \Bigg\{ \int_0^t k' {t_1}^{k'+1} c_{k'}({t_1})\textrm{d}{t_1} + \int_0^t G_{k'+1}({t_1})\textrm{d}{t_1} \Bigg\} .\nonumber
288: \end{eqnarray}
289: \end{small}
290: We now evaluate this expression explicitly for the first few values of $k' = 0,1,2,3$ to generate the degree distribution for
291: the nodes with degrees $1\rightarrow4$:
292: \begin{tiny}
293: \begin{eqnarray}
294: c_1(t)&=&\frac{1}{t^2}\int_0^t G_1(t_1)\textrm{d}t_1, \nonumber\\
295: c_2(t)&=&\frac{1}{t^3}\Bigg\{ \int_0^t\int_0^{t_1} G_1(t_2)\textrm{d}t_2 \textrm{d}t_1
296: + \int_0^t G_2(t_1)\textrm{d}t_1
297: \Bigg\}, \nonumber\\
298: c_3(t)&=&\frac{1}{t^4}\Bigg\{\int_0^t\int_0^{t_1}\int_0^{t_2} 2G_1(t_3)\textrm{d}t_3 \textrm{d}t_2 \textrm{d}t_1 \nonumber\\
299: {}&{}& + \int_0^t\int_0^{t_1} 2G_2(t_2)\textrm{d}t_2 \textrm{d}t_1 \nonumber\\
300: {}&{}& + \int_0^t G_3(t_1)\textrm{d}t_1 \Bigg\},\nonumber\\
301: c_4(t)&=&\frac{1}{t^5}\Bigg\{ \int_0^t\int_0^{t_1}\int_0^{t_2}\int_0^{t_3} 3\times 2 G_1(t_4)\textrm{d}t_4 \textrm{d}t_3
302: \textrm{d}t_2 \textrm{d}t_1\nonumber\\
303: {}&{}& + \int_0^t\int_0^{t_1}\int_0^{t_2} 2\times3G_2(t_3)\textrm{d}t_3 \textrm{d}t_2 \textrm{d}t_1 \nonumber\\
304: {}&{}& + \int_0^t\int_0^{t_1} 3G_3(t_2)\textrm{d}t_2 \textrm{d}t_1 \nonumber\\
305: {}&{}& + \int_0^{t} G_4(t_1)\textrm{d}t_1 \Bigg\}.
306: \end{eqnarray}
307: %\begin{eqnarray}
308: %c_5(t)&=&\frac{1}{t^6}\Bigg\{ \int_0^t\int_0^{t_5}\int_0^{t_4}\int_0^{t_3}\int_0^{t_2} 2\times3\times4 G_1(t_1)\textrm{d}t_1
309: %\textrm{d}t_2 \textrm{d}t_3\textrm{d}t_4 \textrm{d}t_5 \nonumber\\
310: %{}&{}&+ \int_0^t\int_0^{t_4}\int_0^{t_3}\int_0^{t_2} 2\times3\times4 G_2(t_1)\textrm{d}t_1 \textrm{d}t_2 \textrm{d}t_3 \textrm{d}t_4
311: %\nonumber\\
312: %{}&{}&
313: %+ \int_0^t\int_0^{t_3}\int_0^{t_2} 3\times4G_3(t_1)\textrm{d}t_1 \textrm{d}t_2 \textrm{d}t_3 \nonumber\\
314: %{}&{}&
315: %+ \int_0^t\int_0^{t_2} 4G_4(t_1)\textrm{d}t_1 \textrm{d}t_2 \nonumber\\
316: %{}&{}&
317: %+ \int_0^t G_5(t_1)\textrm{d}t_1
318: %\Bigg\}.\nonumber
319: %\end{eqnarray}
320: \end{tiny}
321: We tabulate the coefficients and number of repeated integrals for the contribution of each $G_x(t)$ to each degree distribution
322: value $c_k$ in Table~\ref{tab:1} using the notation $(f,g)$ such that $f$ represents the coefficient and $g$ the number of
323: repeated definite integrations undergone.
324: \begin{table*}
325: \begin{center}
326: \begin{tabular}{c c c c c c c}
327: {}& $G_1(t)$ & $G_2(t)$ &$G_3(t)$ &$G_4(t)$&$G_5(t)$ &$G_6(t)$\\
328: $c_1(t)$& $(1,1)$ & $ - $& $ - $& $ - $ &$ - $&$ - $\\
329: $c_2(t)$& $(1,2)$ & $(1,1)$ & $- $& $- $ &$ - $&$ - $\\
330: $c_3(t)$& $(1\times2,3)$ & $(2,2)$ &$(1,1)$ & $- $ &$-$&$ - $\\
331: $c_4(t)$& $(1\times2\times3,4)$ & $(2\times3,3)$ &$(3,2)$ & $(1,1) $ &$ - $&$ - $\\
332: $c_5(t)$& $(1\times2\times3\times4,5)$ & $(2\times3\times4,4)$ &$(3\times4,3)$ & $(4,2) $ &$ (1,1) $&$ -$\\
333: $c_6(t)$& $(1\times2\times3\times4\times5,6)$ & $(2\times3\times4\times5,5)$ &$(3\times4\times5,5)$ & $(4\times5,3) $ &$
334: (5,2) $&$ (1,1) $\\
335: \end{tabular}
336: \caption{\label{tab:1} The contribution of the functions $G_x(t)$ to the values of the degree distribution $c_k(t)$. The notation
337: $(f,g)$ is such that $f$ is the coefficient and $g$ is the number of repeated integrals.}
338: \end{center}
339: \end{table*}
340: By inspection from Table~\ref{tab:1}, we can write the solution of the time-dependent degree distribution as
341: \begin{small}
342: \begin{eqnarray}\label{eqn:appfull0}
343: c_k(t)&=&\frac{1}{t^{k+1}}\sum_{x=1}^k\frac{(k-1)!}{(x-1)!}
344: \underbrace{\int_0^t \cdots \int_0^{t_{k-x}}}_{k-x+1}G_x(t_{k-x+1})\textrm{d}t_{k-x+1} \cdots \textrm{d}t_{1}.\nonumber\\
345: {}&{}&{}
346: \end{eqnarray}
347: \end{small}
348: Each integral, when evaluated at the lower limit, is always zero as $G_x( 0)=0$ for all $x$. The upper limits are all equal in value. This allows further streamlining of the expression for the degree distribution as the following identity holds for the repeated integral~\cite{RepeatedIntegral}:
349: \begin{small}
350: \begin{eqnarray}
351: \underbrace{\int_0^t \cdots \int_0^{t_{k-x}}}_{k-x+1}G_x(t_{k-x+1})\textrm{d}t_{k-x+1} \cdots \textrm{d}t_{1}
352: &=& \int_0^t \frac{G_x(y) (t-y)^{k-x}}{(k-x)!}\textrm{d}y.\nonumber\\
353: {}&{}&{}
354: \end{eqnarray}
355: \end{small}
356: Substituting back in the expression for $G_x(t)$, the distribution of Eq.~(\ref{eqn:appfull0}) can now be written as in Eq.~(\ref{eqn:caccfinal}).
357: %\begin{small}
358: %\begin{eqnarray}\label{eqn:appcaccfinal}
359: %c_k(t)&=&\frac{1}{t^{k+1}}\sum_{x=1}^{k}\frac{(k-1)!}{x!(x-1)!(k-x)!}\int_0^t (t-y)^{k-x}(ay^2)^x e^{-ay} \textrm{d}y \nonumber\\
360: %{}&=&\sum_{x=1}^k \frac{(k-1)!(2x)!(at)^x}{x!(x-1)!(k+x+1)!} {_1F_1}[1+2x,2+k+x,-at],\nonumber
361: %\end{eqnarray}
362: %\end{small}
363: %where $_1F_1[\kappa,\rho,\phi]$ is the confluent hypergeometric function. Comparison of the analytic solution Eq.~(\ref{eqn:appcaccfinal}) is made %to a simulated network in Fig.~\ref{fig:prefrandattach}.
364:
365: \begin{thebibliography}{99}
366: \bibitem{GM} M.J. Gagen and J.S. Mattick, Phys. Rev. E {\bf 72}, 016123 (2005).
367:
368: \bibitem{AcceleratingNetworks} D.M.D. Smith, J.P. Onnela and N.F. Johnson, New Journal of Physics, {\bf 9}, 181 (2007).
369:
370: \bibitem{Zhang:2007} Z. Zhang, F. Fang, S. Zhou and J. Guan, Physica A {\bf 388}, 225-232 (2009).
371:
372: \bibitem{Sen} P. Sen, Phys. Rev. E {\bf 69}, 046107 (2004).
373:
374: \bibitem{internet} S.N. Dorogovtsev and J.F.F Mendes, Phys. Rev. E {\bf 63}, 025101(R) (2001).
375:
376: \bibitem{Internet2} M. Faloutsos, P. Faloutsos, C. Faloutsos, ACM SIGCOMM Computer Communications {\bf 29}, 251--262 (1999).
377:
378: \bibitem{Vazquez2} A. V\'azquez, R. Pastor-Satorras and A. Vespignani, Phys. Rev. E {\bf 65}, 066130 (2002).
379:
380: \bibitem{WWW} A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajaopalan, R. Stata, A. Tomkins and J. Wiener, Computer Networks {\bf 33}, 309--320 (2000).
381:
382: \bibitem{collaboration} A.-L. Barab\'asi, H. Jeong, Z. N\'eda, E. Ravasz, A. Schubert and T. Vicsek, Physica A, {\bf 311}, 590--614 (2002).
383:
384: \bibitem{NewmanReview} M.E.J. Newman, SIAM Rev. {\bf 45}, 167--256 (2003).
385:
386: \bibitem{Dorogovtsev} S.N. Dorogovtsev and J.F.F. Mendes, {\it `Evolution of Networks: From Biological Nets to the Internet and WWW}, Oxford University Press, Oxford (2003).
387:
388: \bibitem{ABReview} R. Albert and A.-L. Barab\'{a}si Rev. Mod. Phys. {\bf 74}, 47--97 (2002).
389:
390: \bibitem{mattick:2005} J.S. Mattick and M.J. Gagen, Science {\bf 307}, 856--858 (2005).
391:
392: \bibitem{Nimwegen} E. van Nimwegen, Trends in Genetics {\bf 19}, 479-–484 (2003).
393:
394: \bibitem{Croftetal} L.J. Croft, M.J. Lercher, M.J. Gagen and J.S. Mattick, q-bio.MN/0311021 (2003).
395:
396: \bibitem{Failed} M.J. Gagen and J.S. Mattick, q-bio.MN/0312022 (2004).
397:
398: \bibitem{SizeConstraints} M.J. Gagen and J.S. Mattick, q-bio.MN/0312021 (2004).
399:
400: \bibitem{Ahnert} S.E. Ahnert, T.M.A. Fink and A. Zinovyev, Journal of Theoretical Biology {\bf 252}, 587--592 (2007).
401:
402: \bibitem{BarabasiDeriv2} A.-L. Barab\'asi, R. Albert and H. Jeong, Physica A {\bf 272}, 173--187 (1999).
403:
404: \bibitem{Barabasi} A.-L. Barab\'asi and R. Albert, Science {\bf 286}, 509--512 (1999).
405:
406: \bibitem{topology} X. Yu, Z. Li, D. Zhang, F. Liang, X. Wang and X. Wu, J. Phys. A {\bf 39}, 14343–-14351 (2006).
407:
408: \bibitem{tunabledegree} W.-J. Yuan, X-S Luo, J.-F. Zhou and B.-H. Wang, Physica A {\bf 387}, 5311-5316 (2008).
409:
410: \bibitem{KrapivskyDeriv} P.L. Krapivsky, S. Redner and F. Leyvraz, Phys. Rev. Lett. {\bf 85}, 4629--4632 (2000).
411:
412: \bibitem{DorogovtsevDeriv} S.N. Dorogovtsev, J.F.F. Mendes and A.N. Samukhin, Phys Rev. Lett. {\bf 85}, 4633--4636 (2000).
413:
414: \bibitem{Jeon} Y.-P. Jeon and B.J. McCoy, Eur. Phys. J. B {\bf 60}, 521–-528 (2007).
415:
416: \bibitem{LinkSpaceNetworkAnalysis} D.M.D. Smith, C.F. Lee, J.P. Onnela and N.F. Johnson, Phys. Rev. E. {\bf 77} 036112 (2008).
417:
418: \bibitem{history} M. Mitzenmacher, Internet Mathematics {\bf 1}, 226--251 (2003).
419:
420: \bibitem{Boll} B. Bollob\'as, O. Riordan, J. Spencer and G. Tusn\'ady, Random Structures and Algorithms {\bf18}, 279-–290 (2001).
421:
422: \bibitem{ER} E. Erd\H{o}s and A. R\'enyi, Publicationes Mathematicae Debrecen {\bf 6}, 290--297 (1959).
423:
424: \bibitem{Boll2} B. Bollob\'as, Discrete Mathematics {\bf 33}, 1--19 (1981).
425:
426: \bibitem{Evans} T.S. Evans and A.D.K. Plato, Phys. Rev. E {\bf 75}, 056101 (2007).
427:
428: \bibitem{Abramowitz} M. Abramowitz and I.A. Stegun, {\it Handbook of Mathematical Functions with Formulas, Graphs, and
429: Mathematical Tables}, Wiley, New York (1972).
430:
431: \bibitem{Approximations} K.E. Muller, Numerische Mathematik {\bf 90}, 179--196 (2001).
432:
433: \bibitem{RepeatedIntegral} \url{http://mathworld.wolfram.com/RepeatedIntegral.html}.
434:
435: \end{thebibliography}
436:
437:
438: \end{document}
439: %\begin{small}
440: %\begin{eqnarray}\label{eqn:full1}
441: %c_k(t)&=&\frac{1}{t^{k+1}}\sum_{x=1}^k\frac{(k-1)!}{(x-1)!}
442: %\underbrace{\int_0^t \cdots \int_0^{t_2}}_{k-x+1}G_x(t_1)\textrm{d}t_1 \cdots \textrm{d}t_{k-x+1},\nonumber\\
443: %{}&{}&{}
444: %\end{eqnarray}
445: %\end{small}
446: %where
447: %\begin{eqnarray}
448: %G_x(t) &=&\frac{e^{-a t}(a t^2)^x}{x!}.\nonumber
449: %\end{eqnarray}
450: %Each integral in Eq.~\ref{eqn:full1}, when evaluated at the lower limit, is always zero. This allows further streamlining of the degree distribution (see Appendix~\ref{sec:derivation} for details):