1: \documentclass[twocolumn,showpacs,preprintnumbers,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: \usepackage{epsfig}
7: \usepackage{psfrag}
8: \usepackage{subfigure}
9:
10: %\nofiles
11:
12:
13: \newcommand{\insfig}[2]
14: {
15: \begin{center}
16: \mbox{\epsfig{file=./#1.eps, width=#2\textwidth}}
17: \end{center}
18: }
19:
20:
21: \begin{document}
22:
23: %\preprint{APS/123-QED}
24:
25: \title{Layered Complex Networks}
26:
27: \author{Maciej Kurant}
28: %\email{Maciej.Kurant@epfl.ch}
29: \author{Patrick Thiran}
30: \affiliation{EPFL, Switzerland}
31:
32:
33: \begin{abstract}
34: Many complex networks are only a part of larger systems, where a number of coexisting topologies interact and depend on
35: each other. We introduce a layered model to facilitate the description and analysis of such systems. As an example of
36: its application we study the load distribution in three real-life transportation systems, where the lower layer is the
37: physical infrastructure and the upper layer represents the traffic flows. This layered view allows us to capture the
38: fundamental differences between the real load and commonly used load estimators, which explains why these estimators
39: fail to approximate the real load.
40: \end{abstract}
41:
42: %89.75.Hc Networks and genealogical trees
43: %89.75.Fb Structures and organization in complex systems
44: %89.75.-k Complex systems
45: %89.40.Bb Land transportation
46: %89.65.Lm Urban planning and construction
47: %89.20.Hh World Wide Web, Internet
48: %05.50.+q Lattice theory and statistics (Ising, Potts, etc.) (see also 64.60.Cn Order-disorder transformations and
49: %statistical mechanics of model systems and 75.10.Hk Classical spin models)
50:
51:
52: \pacs{89.75.Hc, 89.75.Fb, 89.40.Bb, 89.20.Hh}
53: %\keywords{Suggested keywords}%Use showkeys class option if keyword
54: %display desired
55: \maketitle
56:
57: The topologies of the Internet at the IP layer~\cite{Faloutsos99}, of the World Wide Web~\cite{Albert99} or the
58: networks formed by Peer To Peer (P2P) applications~\cite{Adamic01} have recently drawn a lot of attention. These graphs
59: have been studied separately, as distinct objects. However, they are closely related: each WWW or P2P link virtually
60: connects two IP nodes. These two IP nodes are usually distant in the underlying IP topology and the virtual connection
61: is realized as a path found by IP routers. In other words, the graph formed by an application is mapped on the
62: underlying IP network. Although the topologies at both layers might share a number of statistical properties (such as a
63: heavy-tailed degree distribution), they are very different.\\
64: There exist layers also under the IP layer; even in a simplified view of the Internet we must distinguish at least one
65: - the physical layer. It consists of a mesh of optical fibers that are usually put in the ground along roads, rails, or
66: power-lines. This results in topologies very different from those observed at the IP layer. A mapping of the IP graph
67: onto the physical layer must satisfy a number of constraints (see e.g.,~\cite{Kurant05}).
68:
69: Another important class of real-life systems is transportation networks. The graphs based on the physical
70: infrastructure of such networks were analyzed on the examples of a power grid \cite{Watts98,AlbertUSAPowerGrid},
71: railway network \cite{Indian03}, road networks \cite{Gastner04,Rosvall05}, or urban mass transportation systems
72: \cite{Latora01,Stations04,Gastner04b,Sienkiewicz05}. Although this approach often gives a valuable insight into the
73: studied topology, it ignores the real-life traffic pattern and hence captures only a part of the full picture.
74: Interestingly, the networks of traffic flows were studied separately, for instance the flows of people within a
75: city~\cite{Chowell03}, and commuting traffic flows between different cities~\cite{Montis05}. These studies, in turn,
76: neglect the underlying physical topology. A comprehensive view of the system requires to analyze both layers (physical
77: and traffic) together. Of course, a partial knowledge of the traffic pattern could be introduced into the physical
78: graph by assigning weights reflecting the amount of carried traffic to the physical edges. This describes well a
79: specific type of transportation network, where all traffic flows are one-hop long and where the two layers actually
80: coincide, such as airport networks~\cite{Barrat04,Guimera05}. However, in the presence of longer (than one hop) traffic
81: flows, the weighted physical graph is not sufficient. For instance, the failure of a single physical node/edge should
82: affect (delete or cause to reroute) all traffic flows using this edge/node, which requires the knowledge of the traffic
83: graph and of the actual routes of these flows in the physical graph.
84:
85: \begin{figure*}[!t]
86: \psfrag{GP}[][]{$G^\phi$}
87: \psfrag{GL}[][]{$G^\lambda$}
88: % \psfrag{GP}[][]{{\large$G^\phi$}}
89: % \psfrag{GL}[][]{{\large$G^\lambda$}}
90: \psfrag{Gpb}[][]{{\large$G^\phi$}}
91: \psfrag{Glb}[][]{{\large$G^\lambda$}}
92: \psfrag{M}[][]{$\;M(\!E^\lambda\!)$}
93: \psfrag{A}[][]{a)}
94: \psfrag{B}[][]{b)}
95: \psfrag{El1}[][]{\footnotesize{$e^\lambda_1$}}
96: \psfrag{V1}[][]{\footnotesize{$v^\phi_1$}}
97: \psfrag{V2}[][]{\footnotesize{$v^\phi_2$}}
98: \psfrag{V3}[][]{\footnotesize{$v^\phi_3$}}
99: \psfrag{M1}[][]{\footnotesize{$M(\!e^\lambda_1\!)$}}
100: \includegraphics[width=1\textwidth]{figures/ExampleNEW3.eps}
101: \\[-0.25cm]
102: \caption{(a) Illustration of the two-layered model. The logical graph~$G^\lambda$ is mapped onto the physical graph~
103: $G^\phi$ by a mapping~$M(E^\lambda)$. In this example the logical edge $e^\lambda_1$ is mapped on $G^\phi$ as the path
104: $M(e^\lambda_1)=(v^\phi_1, v^\phi_2, v^\phi_3)$. Assuming that~$G^\lambda$ is unweighted, the loads of the three
105: indicated nodes are $l(v^\phi_1)\!=\!3$, $l(v^\phi_2)\!=\!2$ and $l(v^\phi_3)\!=\!4$. \; (b) A part of the logical and
106: physical graphs of the EU dataset. Here, the traffic intensities (weights) are indicated by multiedges in the logical
107: graph.\\[-0.7cm]} \label{fig:Example}
108: \end{figure*}
109:
110:
111: Coexisting and dependent graphs can also be observed in social networks~\cite{Wasserman94}, where the same set of nodes
112: may be connected in various ways, depending on the type of relationship chosen to be represented by edges. These graphs
113: are related to each other. It is common, for instance, to establish a new link in a business relationship graph (e.g.,
114: to find a job) by performing a search in our acquaintanceship network (i.e., by asking our friends who ask their
115: friends, etc)~\cite{Granovetter73}. This new direct business link translates into a path in the acquaintanceship
116: network.
117:
118: The above examples call for the introduction of additional layers to the description of some complex systems. Therefore
119: we propose a general \emph{multi-layered model}. We explain it on the example of two layers; all the definitions
120: naturally extend to any number of layers. In the two-layered model, the lower-layer topology is called \emph{physical
121: graph} $G^\phi=(V^\phi, E^\phi)$, and the upper-layer topology is called \emph{logical graph} $G^\lambda=(V^\lambda,
122: E^\lambda)$. We assume that the sets of nodes at both layers are identical, i.e., $V^\phi\equiv V^\lambda$, but as a
123: general rule we keep the indexes $\phi$ and $\lambda$ to make the description unambiguous. Let $N$ be the number of
124: nodes, $N\!\!=\!\!|V^\phi|\!\!=\!\!|V^\lambda|$. The physical and logical graphs can be directed or undirected,
125: depending on the application. The nodes and edges can have \emph{weights} assigned to them and denoted by $w(\cdot)$,
126: with $w=1$ for unweighted graphs. Every logical edge $e^\lambda=(u^\lambda,v^\lambda)$ is mapped on the physical graph
127: as a path $M(e^\lambda)\subset G^\phi$ connecting the nodes~$u^\phi$ and~$v^\phi$, corresponding to $u^\lambda$ and
128: $v^\lambda$. (A path is defined by the sequence of nodes it traverses.) The set of paths corresponding to all logical
129: edges is called \emph{mapping} $M(E^\lambda)$ of the logical topology on the physical topology. Now, the \emph{load}
130: $l$ of a node
131: $v^\phi$ is the sum of the weights of all logical edges whose paths traverse~$v^\phi$:\\[-0.3cm]
132: \begin{equation}\label{eq:node_load}
133: l(v^\phi)=\sum_{e^\lambda\!:\;v^\phi\in M(e^\lambda)} w(e^\lambda)\\[-0.1cm]
134: \end{equation}
135: %(Replace $v^\phi$ with $e^\phi$ in (\ref{eq:node_load}) to obtain the definition of load of an edge $e^\phi$.)
136: %This definition is very intuitive.
137: In a transportation network $l(v^\phi)$ is the total amount of traffic that flows
138: through the node $v^\phi$; if the logical graph is unweighted, $l(v^\phi)$ counts the number of logical edges that are
139: mapped on $v^\phi$.
140:
141:
142: Here, we apply this two-layered framework to study transportation networks. The undirected, unweighted physical graph
143: $G^\phi$ will henceforth capture the physical infrastructure of a transportation network, and the logical graph
144: $G^\lambda$ will reflect the undirected traffic flows. All data studied in this paper is extracted from timetables of
145: public transportation systems. First, we take a list of all of trains, metros and buses departing in the system within
146: one weekday (time-span of 24 hours). A timetable gives the exact route of each vehicle, which translates directly into
147: a logical edge $e^\lambda$ (connecting the first and the last station) and its mapping $M(e^\lambda)$. The number of
148: vehicles following the same path in both possible directions defines the flow intensity - the weight $w(e^\lambda)$ of
149: the logical link.
150: % \footnote{Note that we interpret every vehicle as one traffic unit, regardless of its size or the number
151: %of people using it. Although this gives us a picture at a low level of granularity, we believe it reflects well the
152: %general direction and intensity of travels.}.
153: (In this context, the logical graph is equivalent to a traffic matrix in transportation
154: science~\cite{TransportationBook}.) We describe the algorithm to extract the two layers and the mapping from timetables
155: in~\cite{KurantRailwayAlgorithm}.
156:
157: \begin{figure}[!b]
158: ${}$\\[-0.5cm]
159: \psfrag{x1}[][]{\small{$k^\phi$}} \psfrag{y1}[][]{{\small $P(k^\phi)$}} \psfrag{x2}[][]{{\small$k^\lambda$}}
160: \psfrag{y2}[][]{{\small $P(k^\lambda)$}} \psfrag{x3}[][]{{\small$w^\phi$}} \psfrag{y3}[][]{{\small$P(w^\phi)$}}
161: \psfrag{x4}[][]{{\small$h$}} \psfrag{y4}[][]{{\small$P(h)$}} \psfrag{A}[][]{a)} \psfrag{B}[][]{b)} \psfrag{C}[][]{c)}
162: \psfrag{D}[][]{d)} \psfrag{real}[][]{{\footnotesize {}\;real}} \psfrag{betweenness}[][]{{\footnotesize all-to-all}}
163: \includegraphics[width=0.45\textwidth]{figures/Flows.eps}\\[-0.2cm]
164: \caption{EU network (WA and CH yield similar results). Node degree distribution in the physical graph~(a), and in the
165: logical graph~(b); edge weight distribution in the logical graph~(c); and the distribution of the lengths of traffic
166: flows~(d), counted in a number of hops $h$ in the physical graph.} \label{fig:Flows}
167: \end{figure}
168:
169: \begin{table}[!htb]
170: \begin{tabular}{|l|c|c|c|c|c|c|c|}
171: \hline
172: % Dataset & &\multicolumn{2}{c|}{Physical graph}&\multicolumn{2}{c|}{Logical graph}\\\cline{2-6}
173: Dataset & $N$ & $|E^\phi|$&$d^\phi$&$|E^\lambda|$&\# vehicles\\
174: \hline
175: WA - Warsaw & 1529 & 1827 & 90 &324& 26075\\
176: CH - Switzerland& 1679 & 1750 & 142 & 539&7482\\
177: EU - Europe &6276& 7273 & 181 & 6623& 54073\\
178: \hline
179: \end{tabular}
180: \caption{The studied datasets. $N$ is the number of nodes, $|E^\phi|$ (respectively, $|E^\lambda|$) is the number of
181: edges in the physical (respectively, logical) graph, and $d^\phi$ denotes the diameter of the physical graph.
182: %$\langle w^\lambda\rangle$ - average logical edge weight (flow intensity).
183: The total number of vehicles taken into account for every dataset is given by ``\# vehicles''. Note
184: that $|E^\lambda|\ll$~\# vehicles, because many vehicles follow the same route.\\[-0.5cm]}
185: \label{tab:Data}
186: \end{table}
187: \begin{figure*}[!htb]
188: % \psfrag{b}[][]{\small{$b^\phi$}}
189: % \psfrag{k}[][]{\small{$k^\phi$}}
190: % \psfrag{l}[][]{\small{$l$}}
191: % \psfrag{A}[][]{\normalsize{\hspace{0.12\linewidth}a) Node degree $k^\phi$}}
192: % \psfrag{B}[][]{\normalsize{\hspace{0.12\linewidth}b) Betweenness $b^\phi$}}
193: % \psfrag{C}[][]{\normalsize{\hspace{0.12\linewidth}c) Real load $l$}}
194: % \psfrag{D}[][]{\normalsize{\hspace{0.03\linewidth}d)}}
195: \includegraphics[width=1\linewidth]{figures/LoadComparison.eps}
196: \\[-0.25cm]
197: \caption{EU dataset (WA and CH yield similar results).
198: % (In computing betweenness, restricted betweenness and real load, the end nodes of flows are always taken into account.
199: %However, excluding them yielded very similar results.)
200: The first three figures present the physical layout of the node degree $k^\phi$~(a), betweenness $b^\phi$~(b) and real
201: load $l$~(c). The size of a node is proportional to the measured value. In (d) we present the scatter-plots of the node
202: degree $k^\phi$ (top) and betweenness $b^\phi$ (bottom) versus the real load~$l$.
203: In the top left corner of every plot we give the value of the corresponding Pearson's correlation coefficient.\\[-0.8cm]}
204: \label{fig:LoadComparison}
205: \end{figure*}
206: We study three examples of transportation networks, with sizes ranging from city to continent. As an example of a city,
207: we take the mass transportation system (buses, trams and metros) of Warsaw (WA), Poland. At a country level, we study
208: the railway network of Switzerland (CH). Finally we investigate the railway network formed by major trains and stations
209: in most countries of central Europe (EU). The basic characteristics of these networks can be found in
210: Table~\ref{tab:Data} and in Fig.~\ref{fig:Flows}. All physical topologies are connected, planar or close to planar,
211: with the diameter $d^\phi$ in the order of $O(\sqrt{N})$ (the diameter of a graph is the length of the longest of all
212: possible shortest paths), and node degree distributions decaying exponentially (the degree of a node is the number of
213: edges incident on this node). These features are common to many physical transportation graphs, such as a road network
214: or a railway system. The logical graphs are strikingly different. They are sparse and have multiple components, among
215: which many isolated nodes. The degree distributions of the logical graphs are \emph{right-skewed}, meaning that there
216: is a small number of nodes with very high degree. This is confirmed by the almost linear shape of the distribution in
217: the log-log scale plot shown in Fig.~\ref{fig:Flows}b; a fully linear shape would indicate a power-law (a heavy-tailed
218: distribution).
219: % cover two or more decades, and have the standard deviation several times larger than the mean.
220: %Stretched exponential or log-normal function
221: %. They cover two or more decades, and have the standard deviation several times larger than the mean.
222: %The best fit is given by a log-normal function, which indicates a fat-tail.
223: Similar right-skewed distributions are observed for the weights of edges in the logical graphs (Fig.~\ref{fig:Flows}c).
224: In Fig.~\ref{fig:Flows}d, we compare the length distribution of real traffic flows with the length distribution of
225: all-to-all shortest paths. The former is very much left-skewed, which means that the real flows tend to be local.
226:
227: Knowing the topologies and the mapping of both layers, we can easily compute the load of a node with
228: formula~(\ref{eq:node_load}). For comparison purposes, we present below two \emph{load estimators} based exclusively on
229: the physical graph~$G^\phi$. For load estimators we take two metrics known in social networks as centrality measures;
230: they are used to assess the importance of nodes. Our first metric is \emph{node degree}~$k^\phi$. It seems natural that
231: the nodes with a high degree carry more traffic than the less connected nodes. Our second metric is
232: \emph{betweenness}~$b^\phi$~\cite{Freeman77}. The betweenness of a vertex $v$ is the fraction of shortest paths between
233: all pairs of vertices in a network, that pass through $v$. If there are more than one shortest path between a given
234: pair of vertices, then all such paths are taken into account with equal weights summing to one. Betweenness aims at
235: capturing the amount of information passing through a vertex. Indeed, many authors take betweenness as a measure of
236: load either directly~\cite{Goh01,Holme02,Motter02,Szabo02,Bollobas04,Zhao04}, or with slight
237: modifications~\cite{AlbertUSAPowerGrid,Moreno04,Tadic04}
238: %Sometimes a partial knowledge
239: %of the real traffic pattern is given. For instance in \cite{AlbertUSAPowerGrid}, the authors took advantage of the
240: %knowledge of location of all traffic sources and destinations in the US power grid. Therefore they naturally restricted
241: %the definition of betweenness by considering all source-destination pairs only, instead of all existing pairs in the
242: %graph (they generated shortest paths from every generator node to all non-generator nodes). We will refer to this
243: %method as \emph{restricted betweenness} $r$.
244: %The "betweenness approach" is also used in a dynamic setting, that is
245: %where, in time, the communicating pairs (origin-destination) are chosen at random at a given constant rate and the
246: %units of traffic travel through the network (\cite{Lawniczak03,Moreno04,Tadic04}). Although in some of
247: %these cases the underlying routing algorithm is randomized, the main assumption of "all-to-all at the same rate"
248: %traffic pattern is common with the definition of betweenness.
249:
250: In Fig.~\ref{fig:LoadComparison} we compare the distribution of the real load with its two estimators: node degree and
251: betweenness. The geographical patterns formed by the three metrics differ substantially (see
252: Fig.~\ref{fig:LoadComparison}abc). To quantify these differences, in Fig.~\ref{fig:LoadComparison}d we present the
253: scatter plots of these two estimators versus the real load $l$. The correlations between them are very low, which is
254: confirmed by low values of the corresponding Pearson's coefficients (top left corner of every plot). For instance, for
255: the value of load $l\simeq 10^2$, the corresponding values of betweenness $b^\phi$ cover more than two orders of
256: magnitude. Surprisingly, contrary to the commonly admitted view, the node degree approximates the real load better than
257: betweenness.
258: %As another example consider 100 most loaded (by real load) nodes in the Europe network. Only 10 of them can be found
259: %among 100 top-betweenness, and 12 among 100 top-degree nodes.
260:
261:
262: Why do load estimators fail to mimic the real load pattern? Are there some fundamental reasons behind this? The layered
263: view of the system is very helpful in answering these questions. First, observe that the ways we compute node degree,
264: betweenness and real load, can be unified by recasting the first two in the two-layered setting. Indeed, both the node
265: degree and the betweenness can be computed as the node load~(\ref{eq:node_load}) in two-layered systems with specific
266: logical topologies mapped on the physical graph~$G^\phi$ using shortest paths. We denote these specific logical graphs
267: by $G^\lambda_{\!k\!^\phi}$ and $G^\lambda_{\!b^\phi}$, for the node degree~$k^\phi$ and the
268: betweenness~$b^\phi$, respectively. They are defined as follows.\\
269: In the case of the node degree, pick~$G^\lambda_{\!k\!^\phi}=G^\phi$: the logical graph is identical to the physical
270: graph~$G^\phi$. Hence the mapping of $G^\lambda_{\!k\!^\phi}$ on $G^\phi$ reduces trivially to single hop traffic
271: flows, and~(\ref{eq:node_load}) boils down to $l(v^\phi)=k^\phi(v^\phi)$.\\
272: %Now, since we ``map'' $G^\phi$ on $G^\phi$, all the resulting traffic flows are one-hop long, and the
273: %formula (\ref{eq:node_load}) boils down to $l(v^\phi)=k^\phi(v^\phi)$.\\
274: For the betweenness, $G^\lambda_{\!b^\phi}$ is an unweighted and complete (fully connected) graph. Indeed, the
275: definition of betweenness requires to find shortest paths between every possible pair of vertices. Note that the
276: mapping defined by betweenness splits the path (and its weight) if there are more than one shortest path, whereas the
277: shortest-path mapping simply returns one of them. However, in large graphs this difference is negligible, especially if
278: the shortest-path algorithm picks one of the possible paths at random.\\
279: The same two-layered methodology can therefore be used to compute node degree, betweenness and real load. Moreover, in
280: all three cases we use the same physical graph $G^\phi$ and a mapping that follows the shortest path~\footnote{
281: %(THE NUMERICAL SIMULATIONS HAVE CONFIRMED ALMOST NO DIFFERENCE IN THE RESULTING LOAD - modify Newmans betweenness algorithm for that).
282: The real-life flows almost always coincide with shortest paths connecting their end-nodes.}. Consequently, all the
283: differences between the three metrics are completely captured by the logical graphs $G^\lambda_{\!k\!^\phi}$,
284: $G^\lambda_{\!b^\phi}$ and $G^\lambda$. We compare them in Table~\ref{tab:LogGraphs}. The
285: graph~$G^\lambda_{\!k\!^\phi}$ is moderately dense, planar, unweighted, with the degree distribution decaying
286: exponentially. The edge length, counted in the number of hops in the mapping of this edge, is equal to one for all
287: edges of $G^\lambda_{\!k\!^\phi}$. In contrast, the graph $G^\lambda_{\!b^\phi}$ is an unweighted and complete graph,
288: which means it is very dense, with every node of degree equal to $N\!-\!1$. In $G^\lambda_{\!b^\phi}$ we find both
289: short and long edges; their distribution is bell-shaped, as shown by the ``all-to-all'' curve in Fig.~\ref{fig:Flows}d.
290: Finally, the real-life logical graph $G^\lambda$ is sparse, weighted and has rather local edges (see the ``real'' curve
291: in Fig.~\ref{fig:Flows}d). Moreover, the node degree and edge weight distributions of $G^\lambda$ are both very much
292: right-skewed.
293: %THINK ABOUT OTHER CENTRALITY MEASURES - MAYBE TWO-LAYERED APPROACH CAN COVER THEM ALL?
294:
295: \begin{table}[!t]
296: \begin{small}
297: %\begin{tabular}{|c|p{0.1\textwidth}|p{0.1\textwidth}|p{0.1\textwidth}|}
298: \begin{tabular}{|c|c|c|c|}
299: \hline
300: Property &
301: \raisebox{0.5cm}{\normalsize $G^\lambda_{\!k\!^\phi}$}\hspace{-0.6cm}
302: \begin{minipage}[c]{1.9cm} \includegraphics[width=0.85\textwidth]{figures/planar.eps}\end{minipage}&
303: \raisebox{0.5cm}{\normalsize $G^\lambda_{\!b^\phi}$}\hspace{-0.6cm}
304: \begin{minipage}[c]{1.9cm} \includegraphics[width=0.9\textwidth]{figures/complete.eps}\end{minipage}&
305: \raisebox{0.5cm}{\normalsize $G^\lambda$}\hspace{-0.6cm}
306: \begin{minipage}[c]{1.9cm} \includegraphics[width=0.9\textwidth]{figures/scalefree.eps}\end{minipage}\\
307: \hline
308: $\mid E^\lambda\mid$ & $=\mid E^\phi\mid$ & $=N(N-1)/2$ & $<\mid E^\phi\mid$\\
309: Planar & Yes & No &No\\
310: Weights $w(e^\lambda)$& = 1& = 1 & Right-skewed\\
311: Degrees $k(v^\lambda)$& Exponential & $= N-1$ & Right-skewed\\
312: Edge lengths& = 1 & Bell-shaped & Exponential\\
313: \hline
314: \end{tabular}
315: \caption{The properties of the logical graphs $G^\lambda_{\!k\!^\phi}$, $G^\lambda_{\!b^\phi}$ and $G^\lambda$.
316: %Right-skewed
317: %Fat-tail indicates a log-normal distribution, light-tail indicates a distribution with exponential decay.
318: ``Edge length'' is the number of hops in the mapping of the edge on the physical graph.\\[-0.7cm]} \label{tab:LogGraphs}
319: \end{small}
320: \end{table}
321:
322:
323: \pagebreak[4] There are thus a number of fundamental differences between the three logical
324: graphs~$G^\lambda_{\!k\!^\phi}$, $G^\lambda_{\!b^\phi}$ and $G^\lambda$. They explain why the node degree and
325: betweenness fail to mimic the real load distribution. We expect to observe similar differences in other fields. For
326: instance, the logical graph representing the traffic in the Internet shares many properties with the logical graphs of
327: transportation systems studied here. In particular, in the Internet, the distribution of intensity of traffic flows
328: (which is, in this paper, equivalent to the edge weights in the logical graph) was shown to be
329: heavy-tailed~\cite{Elephants,Meiss05}.
330: %,Lakhina04,
331: This is known in the field as ``the elephants and mice phenomenon''~\cite{Elephants}, where a small fraction of flows
332: is responsible for carrying most of the traffic. Moreover, the number of flows originating from a given node (which is,
333: in this paper, equivalent to the node degree in the logical graph), was also shown to follow a power-law
334: distribution~\cite{Meiss05}.
335:
336: To summarize, we have introduced a framework for studying complex systems in which we distinguish graphs on two or more
337: layers. We have shown on the example of transportation networks how the layered view can facilitate the description,
338: comparison and analysis of such systems. Our work represents only a fraction of the possibilities in this area. For
339: example, the layered perspective can completely change our view of the error and attack tolerance of considered
340: systems. It would be also interesting to study how the properties of the topologies at different layers affect the
341: interactions between the layers.
342: %We hope that researchers will feel encouraged to investigate more this interesting class of systems.
343:
344: \begin{thebibliography}{10}
345:
346: \bibitem{Faloutsos99}
347: M.~Faloutsos, P.~Faloutsos, and C.~Faloutsos.
348: \newblock On power-law relationships of the internet topology.
349: \newblock {\em Proc. of ACM SIGCOMM}, 1999.
350:
351: \bibitem{Albert99}
352: R.~Albert, H.~Jeong, and A.-L. Barabási.
353: \newblock Diameter of the world wide web.
354: \newblock {\em Nature}, 401:130--131, 1999.
355:
356: \bibitem{Adamic01}
357: Lada~A. Adamic, Rajan~M. Lukose, Amit~R. Puniyani, and Bernardo~A. Huberman.
358: \newblock Search in power-law networks.
359: \newblock {\em Phys. Rev. E}, 64:046135, 2001.
360:
361: \bibitem{Kurant05}
362: M.~Kurant and P.~Thiran.
363: \newblock On \uppercase{S}urvivable \uppercase{R}outing of \uppercase{M}esh
364: \uppercase{T}opologies in \uppercase{IP}-over-\uppercase{WDM}
365: \uppercase{N}etworks.
366: \newblock {\em Proc. of Infocom'05}, 2005.
367:
368: \bibitem{Watts98}
369: D.~J. Watts and S.~H. Strogatz.
370: \newblock Collective dynamics of ``small-world'' networks.
371: \newblock {\em Nature}, 393:440--442, 1998.
372:
373: \bibitem{AlbertUSAPowerGrid}
374: Reka Albert, Istvan Albert, and Gary~L. Nakarado.
375: \newblock Structural vulnerability of the north american power grid.
376: \newblock {\em Phys. Rev. E}, 69:025103(R), 2004.
377:
378: \bibitem{Indian03}
379: Parongama Sen, Subinay Dasgupta, Arnab Chatterjee, P.~A. Sreeram, G.~Mukherjee,
380: and S.~S. Manna.
381: \newblock Small-world properties of the \uppercase{I}ndian railway network.
382: \newblock {\em Phys. Rev. E}, 67:036106, 2003.
383:
384: \bibitem{Gastner04}
385: Michael~T. Gastner and M.~E.~J. Newman.
386: \newblock The spatial structure of networks.
387: \newblock {\em cond-mat/0407680}, 2004.
388:
389: \bibitem{Rosvall05}
390: M.~Rosvall, A.~Trusina, P.~Minnhagen, and K.~Sneppen.
391: \newblock Networks and cities: An information perspective.
392: \newblock {\em Phys. Rev. Lett.}, 94:028701, 2005.
393:
394: \bibitem{Latora01}
395: V.~Latora and M.~Marchiori.
396: \newblock Efficient behavior of small-world networks.
397: \newblock {\em Phys. Rev. Lett.}, 87:198701, 2001.
398:
399: \bibitem{Stations04}
400: Katherine~A. Seaton and Lisa~M. Hackett.
401: \newblock Stations, trains and small-world networks.
402: \newblock {\em Physica A}, 339:635, 2004.
403:
404: \bibitem{Gastner04b}
405: Michael~T. Gastner and M.~E.~J. Newman.
406: \newblock Shape and efficiency in spatial distribution networks.
407: \newblock {\em cond-mat/0409702}, 2004.
408:
409: \bibitem{Sienkiewicz05}
410: J.~Sienkiewicz and J.~A. Ho{\l}yst.
411: \newblock Statistical analysis of 22 public transport networks in poland.
412: \newblock {\em physics/0506074}, 2005.
413:
414: \bibitem{Chowell03}
415: G.~Chowell, J.~M. Hyman, S.~Eubank, and C.~Castillo-Chavez.
416: \newblock Scaling laws for the movement of people between locations in a large
417: city.
418: \newblock {\em Phys. Rev. E}, 68:066102, 2003.
419:
420: \bibitem{Montis05}
421: Andrea~De Montis, Marc Barth\'elemy, Alessandro Chessa, and Alessandro
422: Vespignani.
423: \newblock The structure of inter-urban traffic: A weighted network analysis.
424: \newblock {\em physics/0507106}, 2005.
425:
426: \bibitem{Barrat04}
427: A.~Barrat, M.~Barth\'elemy, R.~Pastor-Satorras, and A.~Vespignani.
428: \newblock The architecture of complex weighted networks.
429: \newblock {\em Proc. Nat. Acad. Sci. USA}, 101(11):3747, 2004.
430:
431: \bibitem{Guimera05}
432: R.~Guimer\`a, S.~Mossa, A.~Turtschi, and L.A.N. Amaral.
433: \newblock The worldwide air transportation network: Anomalous centrality,
434: community structure, and cities' global roles.
435: \newblock {\em Proc. Natl. Acad. Sci. USA}, 102:7794, 2005.
436:
437: \bibitem{Wasserman94}
438: Stanley Wasserman and Katherine Faust.
439: \newblock {\em Social Network Analysis}.
440: \newblock Cambridge University Press, Cambridge, 1994.
441:
442: \bibitem{Granovetter73}
443: M.~Granovetter.
444: \newblock The strength of weak ties.
445: \newblock {\em American Journal of Sociology}, 78(6):1360--1380, 1973.
446:
447: \bibitem{TransportationBook}
448: William~R. Black.
449: \newblock {\em Transportation. A Geographical Analysis}.
450: \newblock The Guilford Press, New York, London, 2003.
451:
452: \bibitem{KurantRailwayAlgorithm}
453: M.~Kurant and P.~Thiran.
454: \newblock Extracting the traffic flows and the physical graphs from timetables.
455: \newblock {\em physics/0510151}, 2005.
456:
457: \bibitem{Freeman77}
458: L.~C. Freeman.
459: \newblock A set of measures of centrality based on betweenness.
460: \newblock {\em Sociometry}, 40:35, 1977.
461:
462: \bibitem{Goh01}
463: K.-I. Goh, B.~Kahng, and D.~Kim.
464: \newblock Universal behavior of load distribution in scale-free networks.
465: \newblock {\em Phys. Rev. Lett.}, 87(27):278701, December 2001.
466:
467: \bibitem{Motter02}
468: Adilson~E. Motter and Ying-Cheng Lai.
469: \newblock Cascade-based attacks on complex networks.
470: \newblock {\em Phys. Rev. E}, 66:065102(R), 2002.
471:
472: \bibitem{Szabo02}
473: G.~Szab\'o, M.~Alava, and J.~Kert\'osz.
474: \newblock Shortest paths and load scaling in scale-free trees.
475: \newblock {\em Phys. Rev. E}, 66:026101, 2002.
476:
477: \bibitem{Bollobas04}
478: B.~{Bollob{\' a}s} and O.~{Riordan}.
479: \newblock {Shortest paths and load scaling in scale-free trees}.
480: \newblock {\em Phys. Rev. E}, 69(3):036114, March 2004.
481:
482: \bibitem{Zhao04}
483: L.~Zhao, K.~Park, and Y.-C. Lai.
484: \newblock {Attack vulnerability of scale-free networks due to cascading
485: breakdown}.
486: \newblock {\em Phys. Rev. E}, 70:035101(R), 2004.
487:
488: \bibitem{Holme02}
489: P.~Holme and Beom~Jun Kim.
490: \newblock Vertex overload breakdown in evolving networks.
491: \newblock {\em Phys. Rev. E}, 65:066109, 2002.
492:
493: \bibitem{Moreno04}
494: P.~Echenique, J.~G. Gardenes, and Y.Moreno.
495: \newblock {Improved Routing Strategies for Internet Traffic Delivery}.
496: \newblock {\em Phys. Rev. E}, 70:056105, 2004.
497:
498: \bibitem{Tadic04}
499: B.~{Tadi{\' c}}, S.~{Thurner}, and G.~J. {Rodgers}.
500: \newblock {Traffic on complex networks: Towards understanding global
501: statistical properties from microscopic density fluctuations}.
502: \newblock {\em Phys. Rev. E}, 69(3):036102, March 2004.
503:
504: \bibitem{Meiss05}
505: M.~Meiss, F.~Menczer, and A.~Vespignani.
506: \newblock On the lack of typical behavior in the global web traffic network.
507: \newblock {\em Proc. of WWW2005}, 2005.
508:
509: \bibitem{Elephants}
510: D.~Papagiannaki, N.~Taft, S.~Bhattacharyya, P.~Thiran, K.~Salamatian, and
511: C.~Diot.
512: \newblock A pragmatic definition of elephants in internet backbone traffic.
513: \newblock {\em Proc. of Internet Measurement Workshop}, 2002.
514:
515: \end{thebibliography}
516:
517:
518: \end{document}
519: