arxiv: v1 [cs.si] 19 Dec 2013

Transcription

1 Vulnerability and power on networks arxiv: v1 [cs.si] 19 Dec 2013 Enrico Bozzo Department of Mathematics and Computer Science University of Udine Massimo Franceschet Department of Mathematics and Computer Science University of Udine Franca Rinaldi Department of Mathematics and Computer Science University of Udine December 20, 2013 Abstract Inspired by socio-political scenarios, like dictatorships, in which a minority of people exercise control over a majority of weakly interconnected individuals, we propose vulnerability and power measures defined on groups of actors of networks. We establish an unexpected connection between the defined measures with the concepts of regularizability and robustness of networks. We provide effective methods to compute vulnerability and power and prove the corresponding computational complexity. Furthermore, we use the Shapley value of coalition games to introduce fresh notions of vulnerability and power at node level. Finally, we experimentally investigate the probability of being vulnerable in the network universe. Keywords: centrality measures; algebraic connectivity; robustness; regularizable graphs; supermodular functions; coalitional games; Shapley value. 1 Introduction Our investigation moves from the observation that there exists a recurrent topology in socio-political scenarios characterized by a minority of people (that, with a bit of emphasis, we call executioners) that exercise control over a majority of 1

2 individuals (that we name victims). This topology consists of a large number of victims that have few or no connections among them and are controlled by (are connected to) few other executioners. It can be portrayed as a restricted core of executioners linked to a sparse periphery of victims, a sort of generalization of a star network (see Figure 1). This topology is typically adopted in dictatorships to gain control over people. Meetings and associations among people(the victims) are prohibited. Links of victims to external sources of information are hampered. This is accomplished, for instance, by imposing limitations to the use of Internet and popular social networking services. On the other hand, communication necessarily flows only between the dictator or a group of few individuals (the executioners) and the isolated victims. The crucial role of Internet and in particular of social networking services (Twitter in particular) during the uprisings of the Arab Spring has been largely acknowledged. These media have been used by insurgents to break isolation with the external world as well as to organize the internal revolution. These communication links decreased the vulnerability of victims with respect to the executioners. Further instances of a similar topological exploitation are described in [14]; we quote a couple of them in the following: Plantation owners in Hawaii a century ago expressly hired workers who spoke different native languages to ensure that communication among them would be limited, thus discouraging labor action. And the extraordinary longevity of the Ottoman Empire ( ) and its remarkable integration and taxation of diverse ethnic and religious communities was based on a network structure that made peripheral elites dependent on the center, communicating only with the center rather than with one another. In this paper we conduct a formal investigation of the described topology in the context of network science. We define a vulnerability measure that quantifies the tendency ofaset ofactorsof the networkto be the potential victims of some group of executioners, as well as a symmetric power measure that assesses the capacity of a set of actors to play the executioners of some pool of victims. We first reveal an unexpected connection between the defined notion of vulnerability and that of graph regularizability [1], a seasoned concept in Graph Theory. Then, we investigate the computational complexity of deciding whether a network as a whole is vulnerable or not, as well as the complexity of the harder problem of finding an actual vulnerability score for a given network. To this end, we show that both the vulnerability and the power measures are supermodular functions which allows for suitable polynomial solution algorithms. In particular, this implies that the value of the whole is greater than or equal to the sum of the values of its parts, proving that both measures define emergent properties of the underlying complex system modeled by the network [9]. We also make a detour through Game Theory [17], culminating in new game-theoretic measures of vulnerability and power defined at node level in terms of the Shapley value of coalition games. Moreover, we formally and experimentally investigate the relationship between vulnerability and robustness of networks defined in terms of 2

3 Figure 1: The archetypal vulnerable node set: white nodes, the victims, are quite a few and share no communication edge among them. The only communication link is with the black node, the unique executioner. All white nodes form a vulnerable node set with respect to the executioner. The topology of the network is that of a star graph where center is the executioner and periphery is the isolated victims (although we used a circle layout in the picture). algebraic connectivity [16]. Finally, we experimentally estimate the probability of being a vulnerable network in the class of random graphs. The rest of the paper is organized as follows. In Section 2 we give a gentle account on the concepts of vulnerability, power and regularizability of networks. Section 3 does the formal work, defining and investigating vulnerability and power from various angles. The outcomes of the experimental investigation are described in Section 4. We review the related literature in Section 5 and draw our conclusions in Section 6. 2 An informal account In this section we give an informal account on the notions of vulnerability, power, and regularizability of networks, which are fundamental in the present work. Vulnerability and power are symmetric notions defined on independent sets of nodes of a network. A set of nodes is said independent if no pairs of nodes of the set are connected by an edge. In a social network, for instance, an independent set of nodes corresponds to a group of actors that do not directly communicate with each other; communication can only occur trough another actor outside the independent set. An independent set of nodes is just the opposite of a clique, which is a set of nodes such that each pair of nodes in the set is connected by an edge. Informally, a set of nodes V is vulnerable if V is independent, large, and totally controllable by a smaller set E of other nodes. For convenience, we 3

4 call V the set of victims and E the set of executioners. More specifically, the following factors contribute to the vulnerability of a set of potential victims: 1. Number of victims: the bigger the number of victims that can be controlled by executioners, the higher the vulnerability of the set of victims. 2. Number of executioners: the lower the number of executioners that can control the victims, the higher the vulnerability of the set of victims. 3. Independence of victims: victims must be independent, that is, they should not share any link among each other. 4. Controllability of victims by executioners: victims must be totally controllable by executioners, meaning that victims are connected in the network only with executioners(not with other victims, and not with actors outside the set of executioners). The archetypal vulnerable set of actors of the network is a large independent set of victims that can be controlled by a unique executioner (see Figure 1). Vulnerability of such a set is high since: (a) potential victims are numerous; (b) all victims are controllable by a single executioner; (c) victims cannot exchange information among them. In such a case, the (potentially malicious) executioner might easily take control of a multitude of isolated victims, exercising power over a large portion of the network. One might argue that a large loosely-interconnected set of victims that is controlled by a smaller set of executioners is also vulnerable, even if the victims are not independent. In general, any link among victims decreases the vulnerability of the set, since if victims can internally communicate, they are no more at the mercy of executioners. However, if these communication bonds are rare, only few victims can share information and most of them are in fact isolated, and this might be a potential vulnerability. We will also investigate this more general notion of vulnerability, which is defined on generic (not necessarily independent) sets of nodes. Power is a symmetric concept with respect to vulnerability. Informally, a set of executioner nodes E is powerful if E is small and has total control over a larger independent set V of victim nodes. Executioners have total control on victims if victims are linked in the network only to executioners. Hence, a vulnerable set V of victims induces a powerful set E of executioners and vice versa. The archetypal powerful set of actors of the network is the center of a star network (see Figure 1). We say that a network is vulnerable if it contains a group of nodes that is vulnerable. Hence, vulnerable networks are those that include at least one vulnerable fragment. On the other hand, a network is not vulnerable if any subset of nodes is not vulnerable. Notice that being vulnerable is a local property of a network (there exists at least one vulnerable subset), while being not vulnerable is a global property of a network (all subsets are not vulnerable). 4

5 Pazzi 2 Ginori 2 2 Lamberteschi Albizzi 0 0 Guadagni 0 0 Tornabuoni 2 0 Bischeri Salviati Medici 0 Ridolfi 0 Strozzi Peruzzi Acciaiuoli 2 0 Barbadori 2 Castellani Figure 2: The network of intermarriage relations between the 15th century Florentine families. This is an example of quasi-regularizable but not regularizable network. Edges are labelled with the (possibly null) strengths that make the corresponding weighted network regular (with degree 2). We will explore the equivalence of vulnerability (or the symmetric concept of power) with a seemingly unrelated topological property of a network called regularizability. Let assume that the binary relationships of a network are weighted with a positive number, called weight of the relationship, indicating the strength of the link among the involved two actors. The weighted degree of a node in a weighted network is the sum of weights of the edges incident in the node. A network, either weighted or unweighted, is regular if all nodes have the same (weighted) degree. 1 Such regular networks are rare in the real world; they are mainly the object of study of mathematicians and graph theorists. Typically, nodes of real networks have different degrees, and the node degree distribution displays a long tail, with many nodes with low degree and few (but not rare) ones with a remarkably high degree. An unweighted network is said regularizable (respectively, quasi-regularizable) if we can assign positive weights (respectively, null or positive weights) to the edges of the networks such that the resulting weighted network is regular. Hence, regularizable and quasi-regularizable networks are those networks that can be made regular by properly weighting their edges. Of course, a regular network is regularizable, and a regularizable network is quasi-regularizable. The class of regular networks is properly contained in that of regularizable networks, which is a proper subset of that of quasi-regularizable network. An example of network 1 It is worth noticing that in both weighted and unweighted regular networks all nodes have the same degree centrality as well as the same eigenvector centrality. However, other centrality measures, like closeness and betweenness, might differ among nodes. 5

6 Tornabuoni Medici Ridolfi Guadagni 4 Barbadori 2 Castellani 4 2 Bischeri 1 Strozzi Peruzzi Figure 3: The (unique) biconnected component of the intermarriage network depicted in Figure 2. This is a regularizable network: edges are labelled with the positive strengths that make the corresponding weighted network regular (with degree 6). The unweighted network is not regular. that is not quasi-regularizable (and hence is not regularizable) is that provided in Figure 1 - it is easy to see that it is not possible to assign weights to the edges of the star graph so that the center has the same degree of all peripheral nodes. A real example of a quasi-regularizable network that is not regularizable is depicted in Figure 2. The network depicts the intermarriage relations between the 15th century Florentine families [18]. Its unique biconnected component, depicted in Figure 3, is instead regularizable, but, when unweighted, it is not regular. Finally, a regular unweighted graph is given in Figure 4. Finally, we will explore an unexpected connection between vulnerability and robustness of networks. Robustness is the property of networks that reflects how easily a network can be divided: it is small for networks that can be easily partitionedintwogroupsofnodes, anditislargefornetworksthatcanbehardly partitioned in two fragments. Notice that being robust is a global property of a network (no two subsets of nodes are easily separable), while being not robust is a local property of a network (there exist at least two subsets of nodes that are easily separable). 3 Vulnerability and power on networks We start by formally defining the notion of vulnerability. Let G = (V,E) be an undirected graph without isolated nodes. For every subset T V, we denote by N(T) = {j V : there is i T such that ij E} the set of the neighbors of the nodes in T and by S(G) the collection of the independent sets of G, i.e., 6

7 Figure 4: An unweighted regular network with degree 4. those subsets S V such that N(S) S =. We introduce a vulnerability function v : 2 V Z defined by v(t) = T N(T) T V. (1) Since for every set T V each node in T N(T) gives a null contribution to v(t), the vulnerability function v(t) can be equivalently expressed as v(t) = I(T) N(T)\T (2) where I(T) = T \N(T) denotes the independent set containing all the nodes of T that have no neighbor in T. The vulnerability ν(g) of the network G is the maximum vulnerability of an independent set of nodes in G: ν(g) = max v(s). (3) =S S(G) We say that G is vulnerable if ν(g) > 0, i.e., there exists an independent set S such that S > N(S). On the contrary, in non-vulnerable networks, S N(S) for every independent set. A weaker notion of vulnerability can be defined by maximizing the function v(t) over all the subsets of V, not only the independent ones, that is by setting ν(g) = maxv(t). (4) T V Clearly ν(g) ν(g). Moreover, the following property holds. 7

8 Lemma 1. If ν(g) > 0 then ν(g) = ν(g). (5) Proof. Let T be a subset of V that maximizes v(t). Then v( T) = I( T) N( T)\ T > 0 which implies that the independent set I( T) is not empty. From N(I( T)) N( T)\ T it follows v(i( T)) v( T) that proves the statement. Note that the hypothesis that G has no isolated nodes implies v(v) = 0 and thus ν(g) 0. As a consequence, the above result implies that ν(g) ν(g) if and only if ν(g) < Determining if a network is vulnerable As a first aspect, we consider the problem of determining if a network G is vulnerable or not. In graph theory the networks G with ν(g) 0 and ν(g) < 0 have been characterized from several perspectives. A first characterization arises from the study of quasi-regularizable and regularizable graphs. We recall that a graph G is quasi-regularizable if it is possible to assign non-negative integer weightsto the edges ofthe graphsin such a waythat the sum ofthe weightsover the edges incident in any node is the same non-null value. The graph is called regularizable if these weights can be chosen strictly positive. An alternative characterization, useful from a computational point of view, involves the notion of 2-matching. A 2-matching is an assignment of weights 0, 1 or 2 to the edges of the graph with the property that the sum of weights of the edges incident in any node is at most 2. If this sum is exactly 2 for each node, the 2-matching is called perfect. We summarize the main relations between the above concepts and the properties ν(g) 0 and ν(g) < 0 in the following two theorems [15]. Theorem 1. Let G = (V,E) be a connected undirected graph. Then the following conditions are equivalent: 1. S N(S) for every independent set S V, i.e., ν(g) 0; 2. G is quasi-regularizable [2]; 3. G admits a perfect 2-matching [22]. Theorem 2. Let G = (V,E) be a connected undirected graph. Then the following conditions are equivalent: 1. S < N(S) for every independent set S V, i.e., ν(g) < 0; 2. G is a regularizable graph that is not elementary bipartite 2 [1]; 3. G is a 2-bicritical graph, i.e., for each node v V the graph G(V \{v}) admits a perfect 2-matching [19]. 2 A bipartite graph is elementary if every edge is contained in a perfect matching. 8

9 Algorithm 1 Procedure B for the maximum independent set problem Input: an undirected graph G = (V,E); Procedure: S := A(G); W := V \( S N( S)); if W = then S := S; else S W := B(G(W)); S := S S W ; return S. Since the problem of determining if a graph admits a perfect 2-matching is polynomial [21], Theorems 1 and 2 imply that one can determine in polynomial time if ν(g) < 0, ν(g) = 0 or ν(g) > Computing the vulnerability of a network Let us now address the problem of computing the vulnerability of a network, that is problem (3). As a first result, we show that this is an NP-hard problem. Theorem 3. Let G = (V,E) be an undirected network. The problem of computing ν(g) is NP-hard. Proof. We show that if a polynomial algorithm A to compute ν(g) would exist then we could solve in polynomial time the Maximum Independent Set Problem, i.e., the problem of finding an independent set of G of maximum cardinality, which is an NP-hard problem[6]. To this aim consider the recursive procedure B reported in Algorithm 1. The procedure applied to the input graph G first calls for algorithm A to find an independent set S of G of maximum vulnerability. If S N( S) = V, the procedure returns S = S. Otherwise, denoting by W = V \( S N( S)) the subset of nodes that are neither in S nor in N( S), the procedure B is recursively applied to the subgraph G(W) of G induced by W. In this case the final output is the set S = S S W where S W is the set returned by B when applied to G(W). We first prove by induction on V that the set S returned by B is a maximum independent set of G. If V = 1 the procedure clearly returns V and the thesis holds. Assuming V > 1, let S = A(G) and k = v( S) = v(g). Then N( S) + k = S and V can be partitioned as V = S N( S) W where W = V \( S N( S)). If W, being W < V, the independent set S W returned by the procedure B applied to the graph G(W) is a maximum independent set of G(W) by induction. Moreover, since S W N( S) =, the subset S = S S W is an independent set of G. It remains to show that S S for each independent set S of G. To this aim partition S as S = (S S) (S N( S)) (S W). Since S W S W by the 9

10 maximality of S W, it is sufficient to show that S (S S) (S N( S)). We can write S = N( S) +k = N( S) (V \S) + N( S) S +k and note that the set N( S) (V \S) contains N(S S). Then the maximality of v( S) = v(g) implies N( S) (V \S) N(S S) S S k that leads to S S S k + N( S) S +k = S S + N( S) S as required. Finally we note that since W strictly decreases at each step of the recursion, the complexity of the procedure B is O( V H) where H denotes the time complexity of the procedure A. As a consequence, if A is polynomial then B is polynomial as well. We now prove that, despite of the previous negative result, the problem of computing ν(g) can be solved in polynomial time whenever ν(g) 0. Notice that, as shown in Section 3.1, graphs with ν(g) = 0, that is, quasi-regularizable but not regularizable graphs, can be recognized in polynomial time. So we may restrictto the case ν(g) > 0, that is, to the classofnetworksthat arevulnerable or, equivalently, to the class of networks that are not quasi-regularizable. By Lemma 1, if ν(g) > 0 then ν(g) = ν(g) = max T V v(t). Now the problem of maximizing the vulnerability function over all the subsets of V can be solved by exploiting the fact that v(t) is a supermodular function. Definition 1. Let V be a finite set. A function f defined on the collection 2 V of all the subsets of V is called supermodular if f(s T)+f(S T) f(s)+f(t) for every S,T V. Moreover, f is called submodular if g = f is supermodular and modular if f is both supermodular and modular. Theorem 4. The vulnerability function v(t) is supermodular. Proof. Since T is a modular function it is sufficient to show that N(T) is a submodular function. This immediately follows from the fact that for each pair of subsets S,T V it holds N(S T) = N(S) + N(T) N(S) N(T) and N(S T) N(S) N(T). Submodular (and supermodular) functions have been introduced as an interesting generalization of the rank functions of matroids, where matroids are those combinatorial structures on which the greedy algorithm yields an optimal solution. Likely, several results of the matroid theory can be extended to the theory of submodular functions (see [21] vol.b for an extensive survey on the subject). In particular, [8] showed that the problem of minimizing an integervalued submodular function, i.e. to find a subset T V of minimum value f(t), can be solved in strongly polynomial time if f is given by a value giving oracle and the function is bounded. The proposed algorithm makes use of the ellipsoid method and is not efficient in practice. Combinatorial strongly polynomial algorithms to minimize a submodular function have been later proposed 10

11 Figure 5: A 3-regular graph with an idependent set of maximum vulnerability (black nodes) which is not the independent set of maximum cardinality. in [12, 20, 5, 13]; a survey on the topic is [11]. Clearly, the above algorithms can be used, with straightforward modifications, to maximize a supermodular function and in particular, in our case, to compute the vulnerability of a vulnerable network in strongly polynomial time. The algorithm provided by [13], for instance, runs (for integer valued functions) in O( V 4 E +n 5 )log V M) where M is the maximum value of f and E is the time for the function evaluation. Theorem 5. If ν(g) 0, the problem of computing ν(g) can be solved in strongly polynomial time. Proof. We can determine in polynomial time if either ν(g) = 0 or ν(g) > 0 by solving a sequence of O( V ) 2-matching problems (see Theorems 1 and 2). If ν(g) > 0 then ν(g) = ν(g) by Lemma 1. Being the vulnerability function bounded, computable in polynomial time and supermodular by Theorem 4, we can determine in strongly polynomial time a subset T V with maximum value v( T) = ν(g). If T is not independent, the subset I( T) = T \N( T) is an independent set of G with v( T) = v(i( T)). Corollary 1. If G = (V 1 V 2,E) is a bipartite graph, then computing v(g) is a polynomial problem. Proof. The statement follows from Theorem 5 and the fact that, beeing V 1 and V 2 twoindependentsetsofgwithn(v 1 ) = V 2,N(V 2 ) = V 1 andmax{ V 1, V 2 } min{ V 1, V 2 }, it holds that v(g) 0. When ν(g) < 0, the equivalence among the problems of maximizing the vulnerability function v over S(G) and 2 V does not hold anymore. In this case a simple lower bound on ν(g) is given by LB(G) = 1 δ(g), where δ(g) denotes the minimum degree of a node of G. However there are cases where this bound is not tight, as it happens for the graph G shown in Fig. 5. The graph is a cubic (regular) graph, thus ν(g) < 0 and LB(G) = 2. The black 11

12 nodes form an independent set S with S = 4, N(S) = 5 and this implies ν(g) = v(s) = 1 > LB(G). Note that, since each node of V \(S N(S)) is not adjacent to any node of S, S is not a maximum independent set of G. When ν(g) < 0 one can address the problem of computing ν(g) by integer linear programming. A 0-1 linear programming model of problem (3) can be defined by introducing for each i V two binary variables x i and y i with the meaning that x i = 1 if i S, 0 otherwise, and y i = 1 if i N(S), 0 otherwise. The model is max (x i y i ) i V x i +x j 1 ij E (6) y j x i ij E (7) y i x j ij E (8) x i,y i {0,1} i V. (9) Constraints(6)assurethatthe sets ofthe nodesiwithx i = 1isanindependent set, while constraints (7) and (8) force to 1 all the variables y j associated with nodes in N(S). Note that the integrality constraints on the y variables can be relaxed since they are anyway satisfied in every optimal solution. It is worth noticing that when G is a bipartite graphthe matrix A associated with constraints (6)-(8) is a totally unimodular matrix. This implies that all the vertices of the polyhedron defined by constraints (6)-(8) are integer vectors and thus the integer linear programming model can be solved by simply solving its continuous relaxation, i.e., the problem obtained by substituting constraints (9) with 0 x i,y i 1 for each i V. 3.3 A symmetric perspective: power Assuming a symmetric perspective, in this section we introduce two power functions that measure the capacity of a set of nodes T V to completely control a set of victim nodes. To this aim we denote by B(T) = {i V : N({i}) T} the subset of nodes whose neighbors are contained in T. By definition, the subset S(T) = B(T)\T is an independent set, thus the maximum independent set in B(T) (V \T). We define two power functions p, p : 2 V Z by setting for each T V and p(t) = B(T) T (10) p(t) = S(T) T (11) and we associate to the graph the two corresponding values π(g) = max T V p(t) 12

13 and π(g) = max T V :S(T) p(t). While p(t) measures the relative ability of T to completely control any node in the graph, p(t)takesinconsiderationonlynodesoutside T. Since S(T) B(T) for each T V, it holds π(g) π(g). The next proposition points out the strong relationship between π(g) and π(g) and the vulnerability notions ν(g) and ν(g) introduced in the previous section. Proposition 1. For every network G it holds that: 1. p(t) = v(v \T) for each T V; 2. ν(g) = π(g) and ν(g) = π(g). Proof. Item 1 follows from the fact that for each T V it holds that B(T) = V \N(V \T) and thus p(t) = V\N(V\T) T = V N(V\T) T = V\T N(V\T) = v(v\t). We now show item 2 of the proposition. The first identity immediately follows from item 1. For every independent set S it holds that S B(N(S))\ N(S), thus S S(N(S)). So we obtain v(s) = S N(S) S(N(S)) N(S) = π(n(s)) that implies v(g) π(g). On the other hand for each T with S(T) it holds N(S(T)) T and this implies v(s(t)) p(t). As a consequence we obtain ν(g) π(g). As a consequence of the above result, if S is an independent set of maximum vulnerability, then its neighbor set N(S) is a subset of maximum power π(g). In particular, the problems of computing ν(g) and π(g) and, on the other hand, ν(g) and π(g) are equivalent. Moreover, by Lemma 1, if G is vulnerable π(g) = ν(g) = ν(g) = π(g) > 0 and a set of maximum power can be found in polynomial time. Also, item 1 of Proposition 1 implies that the power function p(t), as the vulnerability function v(t), is supermodular. 3.4 A game-theoretic interpretation Both the power and the vulnerability functions introduced above associate values with subset of nodes, and not with single nodes as it is common for the centrality measures proposed in network theory. In this section we show how to define vulnerability and power at node level using a game-theoretic interpretation. This can be done by using the power and vulnerability functions to define suitable coalitional games on node sets of the network and by considering a classical game solution, the Shapley value. For the game theory notions in this section the reader is referred, among others, to [17]. 13

14 We caninterpretthe powerfunction p(t)definedin(10)asthecharacteristic function of a coalitional game with player set V. In game theory, a characteristic function is commonly used to assign to each coalition of players a value corresponding to the power of the coalition, i.e., how much these players can globally get if they decide to play together, independently on the other players actions. In our case p(t), as a measure of power, returns the difference between the number of the potential victims of T and the cardinality of T. A common task in game theory is that of deriving, on the base of the characteristic function, an assignment ofscoresto the playersas an index ofthe power of the single players in the game. Probably the most popular and used solution proposed for coalitional games is the Shapley value. This solution associates with each game G = (N,w), where N is the set of players and w : 2 N R the characteristic function, a vector φ R N whose components are given by φ i = T V :i/ T T!( N T 1)! (w(t {i}) w(t)). (12) N! According to this definition, the score assigned to each player i is the average over all the orders L of the player set N of the contribution that player i gives when reaches the coalition T formed by the players that precede i in L. It has been shown that the Shapley value is the only solution that satisfies a suitable i N φ i = w(n) (group set of axioms. In particular these axioms require: rationality axiom) and φ(i) = φ(j) if w(t {i}) = w(t {j}) for each coalition T not containing i, j (symmetry axiom). For the game G 1 = (V,p) defined by the power function p(t), p(v) = 0 and the difference p(t {i}) p(t), T V and i / T, has the form p(t,i) = p(t {i}) p(t)= B(T {i})\b(t) 1= N(i) B(T {i}) 1. (13) We now compute the Shapley value for two simple examples. Example 1 As a first example consider a star S n with n+1 nodes and denote by c its center. Then T if T n 1, c / T, 1 n if T = n, c / T p(t) = (14) n T if T n, c T 0 if T = V Let us compute the Shapley value of the game G 1. Since all the nodes of V \{c} are symmetric players, they get the same value ˆφ. Moreover by the group rationality axiom, φ c = 0 nˆφ. Computing φ c is an easy task. Indeed p(t {c}) = n 1 for each coalition T V \{c} and this implies that the average contribution of c is φ c = n 1. As a consequence ˆφ = n 1 n = 1 n 1. Example 2 In order to capture how the Shapley value of the power game depends on the topology of the graph we analyze the very simple graph G with 14

15 Figure 6: The network of Example 2. 5 nodes and edge-set E = {15,25,34,35,45} shown in Figure 6. The graph has a positive value π(g) since, for instance, p({4,5}) = 1. The power function takes values { 1 if i 5, 2 if i,j 5, p({i}) = p({i,j}) = 0 if i = 5, j {1,2} (15) 1 if i = 5 1 if i = 5, j {3,4} 3 if i,j,k 5, 1 if i = 5, j = 1, k = 2 p({i,j,k}) = (16) 0 if i = 5, j {1,2}, k {3,4} 1 if i = 5, j = 3, k = 4 and finally 3 if i = 5, p(v \{i}) = 1 if i = 3,4 (17) 0 if i = 1,2 Because of the symmetric property and the group rationality axioms of the Shapley value it must be φ 1 = φ 2, φ 3 = φ 4 and φ 5 = 4 i=1 φ i. Then a direct computation of the Shapley value gives φ 1 = φ 2 = 3 4, φ 3 = φ 4 = 1 4 φ 5 = 2. By adopting a symmetric approach, we can consider the coalitional game G 2 (V,v) defined by the vulnerability function v(t) that measures how the nodes of T (seen as victims) are exposed to the control of their neighbors in N(T) (seen as executioners). In this case the components of the Shapley value can be interpreted as a measure of the contribution of the single nodes to the vulnerability of the graph. For each subset T V and node i V \T the difference v(s {i}) v(s) is given by v(t,i) = v(t {i}) v(t) = 1 N(T {i})\n(t) = 1 N({i})\N(T). (18) The following proposition shows how the symmetry between the vulnerability and power functions is reflected in the symmetry of the Shapley values of the 15

16 corresponding games. It implies, in particular, that it is sufficient to compute any one of the two values. Proposition 2. For every network G = (V,E), the Shapley values φ 1 and φ 2 of the games G 1 (V,p) and G 2 (V,v) are symmetric, i.e., φ 2 = φ 1. Proof. By item 1 of Proposition 1, for each T V and i / T v(t,i) = p(v \(T {i})) p(v \T) = p(v \(T {i}),i). Since v(t,i) and p(v \(T {i})) have the same coefficient in the expression of the Shapley value (12) the statement holds. 4 Experimental evaluation In this section we report the outcomes of a pair of experiments that we conducted on random graphs. The goal of the first experiment is to assess the relationship among vulnerability and robustness of a graph: are robust graphs less vulnerable? Do fragile networks have high vulnerability? The aim of the second experiment is to estimate the probability of being a regularizable or quasiregularizable graph: how many graphs are regularizable? How many graphs are quasi-regularizable? Notice that, because of Theorem 1, a network is vulnerable if and only if it is not quasi-regularizable, hence the probability of finding a vulnerable network is the complement to 1 of the probability of finding a quasi-regularizable network. We performed both experiments in the computing environment R, taking advantage of the network analysis package igraph. For the first experiment we generate randoms graphs according to the following two graph models: Barabási-Albert graphs (BA graphs, for short), also known as scale-free graphs, and Erdős-Rényi graphs (ER graphs, for short). We first generate a sample of random BA graphs, varying the edge density. In particular, we vary the number of edges to add in each step of the preferential attachment process from 1 to n/2, where n is the number of graph nodes. Hence, both sparse and dense graphs are generated. Next, we generate a sample of the size of random ER graphs with the same edge densities of the BA graphs previously sampled. On each graph of the sample, we compute the vulnerability and the algebraic connectivity. The algebraic connectivity of a graph is the second-smallest eigenvalue of the Laplacian matrix of the graph. This eigenvalue is greater than 0 if and only if the graph is connected. The magnitude of this value reflects how easily a network can be divided: it is small for networks that can be easily partitioned in two groups of nodes, that is, the network divides by removing few edges from it, and it is large for networks that can be hardly partitioned in two fragments, that is, to divide the network a large number of edges must be removed. Algebraic connectivity is hence a measure of the robustness of networks [16]. As shown in Figure 7, for both BA and ER graphs, vulnerability and algebraic connectivity are negatively correlated as soon as vulnerability is lower 16

17 than or equal to the watershed score of 0 (recall that the same score of vulnerability determines if the network is regularizable or not). This means that, regularizable networks with low vulnerability have high algebraic connectivity, and hence are robust graphs. On the other hand, for graphs with positive vulnerability, that is, networks that are not regularizable, there is no association between vulnerability and algebraic connectivity. Given these experimental outcomes, we conjecture a partial mathematical relationship between vulnerability and algebraic connectivity of networks. 3 A first step towards a precise formalization of this relationship is the following. Let G = (V,E), with V = n and let S V. The set of the edges connecting S with the rest of the graph makes up the boundary of S, that we denote with (S). Formally (S) = {uv E : S {u,v} = 1}. Clearly, in the case where S is an independent set then (S) = v S ({v}). Actually, for every S V it turns out that (S) S ( λ 2 1 S n where λ 2 is the second-smallest eigenvalue of the graph Laplacian, that is, the graph algebraic connectivity [7]. If S is an independent set, then (S) v S = ({v}) S S is the mean degree of the nodes of S. For any node set S, we have that N(S) is always greater than or equal to the maximum degree of the nodes in S, and hence, it is also greater than or equal to the mean degree of the nodes in S. Summing up, if S is an independent set, we have ( λ 2 1 S ) n (S) S ), N(S). This inequality is weak and makes sense only for λ 2 > 1; however it partially explains the results of the experiments: if algebraic connectivity (λ 2 ) is high, then, any independent set S has a large set of neighbors N(S), and hence the vulnerability of the graph cannot be large (see Figure 7). Another simple observation helps us complementing the explanation of the experimental results. If a graph G has two nodes of degree 1 connected to a third node (of arbitrary degree), then 1 is an eigenvalue of the Laplacian matrix 3 This intuition is corroborated by the known result that expanders (see Section 5) are graphs with large algebraic connectivity. 17

18 Algebraic connectivity Algebraic connectivity Vulnerability Vulnerability Figure 7: Scatter plots comparing vulnerability and algebraic connectivity over Barabási-Albert graphs (left plot) and Erdős-Rényi graphs (right plot). [7], so that λ 2 1. But at the same time the graph vulnerability ν(g) 1, and, if the nodes of degree one connected to the same node are k, then ν(g) k 1. This suggests that when algebraic connectivity is small (λ 2 1) we cannot expect any relationship between vulnerability and algebraic connectivity (see again Figure 7). The second experiment we concoct aims to estimate how frequent are regularizable and quasi-regularizable graphs among random networks. To this end, we generate a sample of Erdős-Rényi graphs, increasing the average node degree from 1 to 10. We use the model G(n,p) of ER graphs, where n is the number of nodes and p is the probability of edges between vertices. The mean degree of a node in a G(n,p) graph is k = p(n 1). We fix the number of nodes n = 100 and increase p so that we obtain the mean degree sequence from 1 to 10. For each graph in the sample, we check whether the graph is regularizable and, if not, whether it is quasi-regularizable. As it is clear from Figure 8, the frequency of quasi-regularizable and that of regularizable increase as the mean node degree k grows. More precisely, when k is low, both frequencies are negligible. As soon as k is sufficiently large, both frequencies start growing very rapidly. By way of example, when n = 100, the frequency of quasi-regularizable graphs is negligible up as soon as k 3, it is significantly above 0 (14%) when k = 4, when k = 5 almost half (48%) of the graphs in the sample are quasi-regularizable, and as soon as k = 6 more than three-quarters (77%) of the sampled random networks are quasi-regularizable. For higher values of 18

19 Frequency regularizable quasi regularizable Mean node degree Figure 8: Frequency of Erdős-Rényi graphs that are regularizable and quasiregularizable by increasing the mean node degree. the mean node degree, the frequency of quasi-regularizable graphs is close to 100%. As for regularizability, the frequency is negligible as soon as k 5. Graphs with k = 6 have 21% probability of being regularizable, those with k = 7 have 50% chance of being regularizable, while networks with k 9 are almost certainly regularizable. We notice, however, that these frequencies tend to become lower as soon as the number of nodes increases. We conjecture that there exists a transition phase of regularizability of networks that depends predominantly on the mean degree of the network. 4 This seems reasonable with the benefit of hindsight. Recall that regularizability is the process of assigning weights to edges to that the resulting graph is regular. When the mean node degree is low, nodes have few incident edges, hence the process of regularizability is hampered. However, as soon as node degrees grow, there are many more possibilities of assigning weights to edges, significantly increasing the probability of success of the regularizability process. Finally, when node degrees are sufficiently large, there are so many possible weight assignments that the graph is almost certainly regularizable. 4 A similar transition phase has been noticed for the giant component of networks: as soon as the mean degree of a node is higher than 1, a giant connected component including the majority of the graph nodes emerges [16]. 19

20 5 Related literature The notion of vulnerability we have proposed is related to the ubiquitous notion of expander graph [10]. Informally, an expander graph is an undirected unweighted graph that is both sparse and robustly connected. Sparsity is achieved by constraining all nodes of the graph to have the same small degree k, which is constant with respect to the number of nodes n (hence expander graphs are k-regular graphs). Robustness holds since every not too large subset of nodes of an expander graph has a relatively large boundary, where the boundary S of a node set S is defined as the set of edges emanating from S to its complement (or sometimes as the set of neighbors of nodes in S belonging to its complement). The expansion parameter for a regular graph G is defined as S h(g) = min S: S n/2 S and a regular graph is a good expander if its expansion parameter is well above 0. One reason that expanders are ubiquitous is that they can be defined and investigated in different languages including graph theory, geometry, probability and algebra. In graph theory, expanders are graphs that are both sparse (hence economical) and robust (to failure or attacks): to disconnect a large part of the graph, one has to remove many edges. Using the geometric notion of isoperimetry, every set of vertices of an expander graph has a relatively large boundary. From the probabilistic perspective, expanders are graphs for which a natural random walk on the graph converges to its limiting distribution very rapidly. Algebraically, expanders are graphs with a large eigengap between the largest and second-largest eigenvalues of the adjacent matrix of the graph(this property is related to the convergence speed of the above mentioned random walk on the graph). Equivalently, expanders are graphs with a large second-smallest eigenvalue of the Laplacian matrix of the graph. We recall that the second-smallest eigenvalue of the graph Laplacian is known as the algebraic connectivity of the graph and it gives an indication of how easily the graph can be divided in two sets of nodes, hence it represents a measure of network robustness. The property of being an expander seems significant in many computational, mathematical and physical contexts. Expanders are useful in the design of economical robust communication networks, in the theory of error correcting codes, in the theory of pseudo-randomness, in computational complexity theory as well as in cryptography. In mathematics, expanders have a role in the study of metric embeddings. Expansion is closely related to the convergence rates of Markov chains, and so they are involved in the study of Monte-Carlo algorithms in statistical mechanics. Recall that we defined vulnerability of an arbitrary graph as v(g) = max S S(G) S N(S). Our definition diverges from that of expander graph for the following reasons: 20

21 1. vulnerability is defined on arbitrary graphs, while an expander is a k- regular graph with small k; 2. the graph vulnerability considers any independent subset of nodes of the graph, while the expansion parameter counts arbitrary node sets that are not too large (those with cardinality not higher than half of the nodes); 3. finally, in the context of network science, graph expanders have been studied with the goal of designing future communication networks with good topological properties, while we propose graph vulnerability with the aim of analyzing existing real networks. Vulnerability is also weakly related to the sociological theory of structural holes suggested in [3]. The author argues very convincingly that opinion and behavior are more homogeneous within than between groups, so people connected across groups are more familiar with alternative ways of thinking and behaving. Brokerage across the structural holes between groups provides a vision of options otherwise unseen, which is the mechanism by which brokerage becomes social capital. [...] Compensation, positive performance evaluations, promotions, and good ideas are disproportionately in the hands of people whose networks span structural holes. In short, these social brokers see bridges where others see holes. A quantitative measure of the mentioned local betweenness centrality is the local clustering coefficient [23, 16]. For a given node i, the local clustering coefficient is the ratio of the number of pairs of neighbors of i that are connected and the number of pairs of neighbors of i. This coefficient is low if there are many structural holes among the neighbors of node i, making the subgraph induced by the neighborhood of i loosely connected. In such a case the broker i has power over information flow between those friends that are not directly connected. The coefficient is high if the neighbors of i are instead tightly connected, and information between these friends can flow directly without passing through i, lowering the power of i. In fact, the inverse of the local clustering coefficient might be regarded as a centrality measure of local betweenness [16]. Now consider a highly powerful set of executioners E and the set V of its potential victims. Since the victim set V is independent, and the victims are connected only to executioners in E, there are many structural holes among the victims in V, and executioners in E have a low local clustering, and hence a high local betweenness. However, a node i with high local betweenness is not necessarily a powerful executioner, since the set of neighbors of i need neither to be independent nor to be totally controllable by i. Finally, standard node centrality measures, like degree, closeness and betweenness, have been extended to sets of nodes in [4]. In particular they define group degree centrality as the relative number of non-group nodes that are connected to group members, that is, for a node set S in a graph with nodes in V, group degree centrality is δ(s) = N(S)\S. V \S 21

22 Thecoefficientrunsfrom0to1and, assumingaconnectedgraph, itismaximum for dominating sets S such that every node not in S is adjacent to at least one memberofs. Tobe effective, it wouldbedesirableforthegroups tobeassmall as possible without sacrificing centrality [4]. Therefore, the authors propose to search for the smallest set S with the maximum degree centrality, that is, the smallest dominating set. In Graph Theory, the cardinality of the smallest dominating set is known as domination number of the graph, and finding the domination number of an arbitrary graph is a classical NP-complete problem in Computational Complexity Theory. Therefore it is believed that there is no efficient algorithm that finds a smallest dominating set for a given graph. The problem of finding the smallest dominating set bears some analogy with that of finding the set ofmaximum power in our setting. However,there arealso significant differences: while the former problem searches for a small set with a large neighbor set, the latter seeks for a small set with a large independent set that is totally controllable by it. 6 Conclusion Classical centrality measures on networks, like closeness and betweenness, are defined at node level. In this contribution, on the other hand, we define measuresatthe levelofnode sets; thesemeasuresquantifythe inclination foraset of nodes to play the role of victims controlled by executioners (vulnerability) or to interpret the symmetric character of executioners that control victims (power). We thoroughly investigate the provided measures from different perspectives: classical graph theory, algorithmic optimization, computational complexity theory, as well as game theory. We end up with original game-theoretic definitions of vulnerability and power at node level. Node-level power assesses the average marginal contribution to power carried by a node when it is added to an arbitrary group of nodes, and similarly for vulnerability. It is worth highlighting that game-theoretic node-level measures are intrinsically defined in terms of corresponding group-level measures, marking a significant departure from classical centrality measures. References [1] C. Berge. Regularizable graphs I. Discrete Mathemtics, 23:85 89, [2] C. Berge. Some common properties for regularizable graphs, edge-critical graphs and B-graphs. In N. Saito and T. Nishizeki, editors, Graph Theory and Algorithms, volume 108 of Lecture Notes in Computer Science, pages Springer, [3] R. S. Burt. Structural holes and good ideas. American Journal of Sociology, 110(2): ,

23 [4] M. G. Everett and S. P. Borgatti. The centrality of groups and classes. Journal of Mathematical Sociology, 23(3): , [5] L. Fleischer and S. Iwata. A push-relabel framework for submodular function minimization and applications to parametric optimization. Discrete Applied Mathematics, 131: , [6] M.R. Garey and D.S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman, San Francisco, [7] C. Godsil and G. F. Royle. Algebraic Graph Theory. Springer, [8] M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial Optimization. Springer, Heidelberg, [9] J. H. Holland. Emergence from Chaos to Order. Oxford University Press, [10] S. Hoory, N. Linial, and A. Wigderson. Expander graphs and their applications. Bulletin of the American Mathematical Society, 43: , [11] S. Iwata. Submodular function minimization. Mathematical Programming, Ser. B, 112:45 64, [12] S. Iwata, L. Fleischer, and S. Fujishige. A combinatorial strongly polynomial algorithm for minimizing submodular functions. Journal of the ACM, 48: , [13] S. Iwata and J.B. Orlin. A simple combinatorial algorithm for submodular function minimization. In Proceedings of the ACM-SIAM Symposium on Discrete Algorithms, pages , [14] W. Kets, G. Iyengar, R. Sethi, and S. Bowles. Inequality and network structure. Games and Economic Behavior, 73(1): , [15] L. Lovász and M.D. Plummer. Matching Theory, volume 29 of Annals of Discrete Mathematics. North Holland, [16] M. E. J. Newman. Networks: An introduction. Oxford University Press, [17] M. J. Osborne and A. Rubinstein. A Course in Game Theory. Annals of Discrete Mathematics. MIT Press, [18] J. F. Padgett and C. K. Ansell. Robust action and the rise of the Medici. American Journal of Sociology, 98: , [19] W. R. Pulleyblank. Minimum node covers and 2-bicritical graphs. Mathematical Programming, 17:91 103,