Internet Mathematics Vol.

10: 222–262

Axioms for Centrality

Paolo Boldi and Sebastiano Vigna

Abstract. Given a social network, which of its nodes are more central? This question
has been asked many times in sociology, psychology, and computer science, and a whole
plethora of centrality measures (a.k.a. centrality indices, or rankings) were proposed to
account for the importance of the nodes of a network. In this study, we try to provide a
mathematically sound survey of the most important classic centrality measures known
from the literature and propose an axiomatic approach to establish whether they are
actually doing what they have been designed to do. Our axioms suggest some simple,
basic properties that a centrality measure should exhibit.
Surprisingly, only a new simple measure based on distances, harmonic centrality,
turns out to satisfy all axioms; essentially, harmonic centrality is a correction to Bave-
las’s classic closeness centrality [Bavelas 50] designed to take unreachable nodes into
account in a natural way.
As a sanity check, we examine in turn each measure under the lens of informa-
tion retrieval, leveraging state-of-the-art knowledge in the discipline to measure the
effectiveness of the various indices in locating webpages that are relevant to a query.
Although there are some examples of such comparisons in the literature, here, for the
first time, we also take into consideration centrality measures based on distances, such
as closeness, in an information-retrieval setting. The results closely match the data we
gathered using our axiomatic approach.
Our results suggest that centrality measures based on distances, which in recent years
have been neglected in information retrieval in favor of spectral centrality measures,
do provide high-quality signals; moreover, harmonic centrality pops up as an excellent
general-purpose centrality index for arbitrary directed graphs.


C Taylor & Francis Group, LLC
222 ISSN: 1542-7951 print
Boldi and Vigna: Axioms for Centrality 223

1. Introduction
In recent years, there has been an ever-increasing research activity in the study
of real-world complex networks [Wasserman and Faust 94] (the world-wide web,
the autonomous-systems graph within the Internet, coauthorship graphs, phone-
call graphs, e-mail graphs and biological networks, to cite but a few). These
networks, typically generated directly or indirectly by human activity and inter-
action (and, therefore, hereafter dubbed “social”), appear in a large variety of
contexts and often exhibit a surprisingly similar structure. One of the most im-
portant notions that researchers have been trying to capture in such networks is
“node centrality”: ideally, every node (often representing an individual) has some
degree of influence or importance within the social domain under consideration,
and one expects such importance to surface in the structure of the social net-
work; centrality is a quantitative measure that aims at revealing the importance
of a node.
Among the types of centrality that have been considered in the literature
(see [Borgatti 05] for a good survey), many have to do with distances between
nodes.1 Take, for example, a node in an undirected connected network: if the sum
of distances to all other nodes is large, the node under consideration is peripheral ;
this is the starting point to define Bavelas’s closeness centrality [Bavelas 50],
which is the reciprocal of peripherality (i.e., the reciprocal of the sum of distances
to all other nodes).
The role played by shortest paths is justified by one of the most well-known
features of complex networks, the so-called small-world phenomenon. A small-
world network [Cohen and Havlin 10] is a graph where the average distance
between nodes is logarithmic in the size of the network, whereas the clustering
coefficient is larger (that is, neighborhoods tend to be denser) than in a random
Erdős-Rényi graph with the same size and average distance.2 The fact that social
networks (whether electronically mediated or not) exhibit the small-world prop-
erty has been known at least since Milgram’s famous experiment [Milgram 67]
and is arguably the most popular of all features of complex networks. For in-
stance, the average distance of the Facebook graph was recently established to
be just 4.74 [Backstrom et al. 12].

1 Here and in the following, by “distance” we mean the length of a shortest path between
two nodes.
2 The reader might find this definition a bit vague, and some variants are often spotted in the
literature: this is a general well-known problem, also highlighted recently, for example, in [Li
et al. 05].
224 Internet Mathematics

The purpose of this article is to pave the way for a formal, well-grounded as-
sessment of centrality measures based on some simple guiding principles; we seek
notions of centrality that are at the same time robust (they should be applicable
to arbitrary directed graphs, possibly nonconnected, without modifications) and
understandable (they should have a clear combinatorial interpretation).
With these principles in mind, we shall present and compare the most popular
and well-known centrality measures proposed in the last decades. The comparison
will be based on a set of axioms, each trying to capture a specific trait.
In the last part of the article, as a sanity check, we compare the measures we
discuss in an information-retrieval setting, extracting from the classic GOV2 web
collection, documents satisfying a query and ranking by centrality the subgraph
of retrieved documents.
The results are somehow surprising, and suggest that simple measures based
on distances, and in particular harmonic centrality (which we introduce formally
in this study) can give better results than some of the most sophisticated indices
used in the literature. These unexpected outcomes are the main contribution of
this work, together with the set of axioms we propose, which provide a conceptual
framework for understanding centrality measures in a formal way. We also try
to give an orderly account of centrality in social and network sciences, gathering
scattered results and folklore knowledge in a systematic way.

2. An Historical Account
In this section, we sketch the historical development of centrality, focusing on
ten classical centrality measures that we decided to include in this study: the
overall growth of the field is, of course, much more complex, and the literature
contains a myriad of alternative proposals that will not be discussed here.
Centrality is a fundamental tool in the study of social networks: the first ef-
forts to define formally centrality indices were put forth in the late 1940s by the
Group Networks Laboratory at MIT directed by Alex Bavelas [Bavelas 50], in the
framework of communication patterns and group collaboration [Leavitt 51, Bave-
las and Barrett 51]; those pioneering experiments concluded that centrality was
related to group efficiency in problem-solving, and agreed with the subjects’ per-
ceptions of leadership. In the following decades, various measures of centrality
were employed in a multitude of contexts (to understand political integration in
Indian social life [Cohn and Marriott 58], to examine the consequences of central-
ity in communication paths for urban development [Pitts 65], to analyze their
implications to the efficient design of organizations [Beauchamp 65, Macken-
zie 66], or even to explain the wealth of the Medici family based on their central
Boldi and Vigna: Axioms for Centrality 225

position with respect to marriages and financial transactions in 15th-century

Florence [Padget and Ansell 93]). We can certainly say that the problem of sin-
gling out influential individuals in a social group is a holy grail that sociologists
have been trying to capture for at least sixty years.
Although all researchers agree that centrality is an important structural at-
tribute of social networks and that it is directly related to other important group
properties and processes, there is no consensus on what centrality is exactly or
on its conceptual foundations, and there is very little agreement on the proper
procedures for its measurement [Burgess 69, Freeman 79]. It is true that often
different centrality indices are designed to capture different properties of the net-
work under observation, but as Freeman observed, “several measures are often
only vaguely related to the intuitive ideas they purport to index, and many are
so complex that it is difficult or impossible to discover what, if anything, they
are measuring” [Freeman 79].
Freeman acutely remarks that the implicit starting point of all centrality mea-
sures is the same: the central node of a star should be deemed more important
than the other vertices; paradoxically, it is precisely the unanimous agreement
on this requirement that may have produced quite different approaches to the
problem. In fact, the center of a star is at the same time

1. the node with largest degree;

2. the node that is closest to the other nodes (e.g., that has the smallest
average distance to other nodes);
3. the node through which most shortest paths pass;
4. the node with the largest number of incoming paths of length k, for every
k;
5. the node that maximizes the dominant eigenvector of the graph matrix;
6. the node with the highest probability in the stationary distribution of the
natural random walk on the graph.

These observations lead to corresponding (competing) views of centrality. Degree

is probably the oldest measure of importance ever used, being equivalent to
majority voting in elections (where x → y is interpreted as “x voted for y”).
The most classical notion of closeness, instead, was introduced by [Bavelas 50]
for undirected, connected networks as the reciprocal of the sum of distances
from a given node. Closeness was originally aimed at establishing how much a
vertex can communicate without relying on third parties for his messages to
226 Internet Mathematics

be delivered.3 In the seventies, Lin proposed to adjust the definition of close-

ness to make it usable on directed networks that are not necessarily strongly
connected [Lin 76].
Centrality indices based on the count of shortest paths were formally developed
independently by [Anthonisse 71] and [Freeman 77], who introduced betweenness
as a measure of the probability that a random shortest path passes through a
given node or edge.
Katz’s index [Katz 53] is based instead on a weighted count of all paths coming
into a node: more precisely, the weight of a path of length t is β t , for some
attenuation factor β, and the score of x is the sum of the weights of all paths
coming into x. Of course, β must be chosen so that all the summations converge.
Whereas the above notions of centrality are combinatorial in nature and based
on the discrete structure of the underlying graph, another line of research studies
spectral techniques (in the sense of linear algebra) to define centrality [Vigna 09].
The earliest known proposal of this kind that of [Seeley 49], who normalized
the rows of an adjacency matrix representing the “I like him” relations among a
group of children and assigned a centrality score using the resulting left dominant
eigenvector. This approach is equivalent to studying the stationary distribution
of the Markov chain defined by the natural random walk on the graph. A few
years later, [Wei 52] proposed to rank sport teams using the right dominant
eigenvector of a tournament matrix, which contains 1 or 0 depending on whether
a team defeated another team. Wei’s work was then popularized by [Kendall 55],
and the technique is known in the literature about ranking of sport teams as
“Kendall–Wei ranking.” Building on Wei’s approach, [Berge 58] generalized the
usage of dominant eigenvectors to arbitrary directed graphs, and in particular
to sociograms, in which an arc represents a relationship of influence between two
individuals.4
Curiously enough, the most famous among spectral centrality scores is also
one of the most recent, [Page et al. 98]: PageRank was a centrality measure
specifically geared toward web graphs, and it was introduced precisely with the
aim of implementing it in a search engine (specifically, Google, that the inventors
of PageRank founded in 1997).

3 The notion can also be generalized to a weighted summation of node contributions mul-
tiplied by some discount functions applied to their distance to a given node [Cohen and Ka-
plan 07].
4 Dominant eigenvectors were rediscovered as a generic way of computing centralities on
graphs by [Bonacich 72].
Boldi and Vigna: Axioms for Centrality 227

In the same span of years, Jon Kleinberg defined another centrality measure
called hyperlink-induced topic search (HITS) [Kleinberg 99]. The idea5 is that
every node of a graph is associated with two importance indices: one (called
“authority score”) measures how reliable (important, authoritative. . . ) a node
is, and another (called “hub score”) measures how good the node is in pointing
to authoritative nodes, with the two scores mutually reinforcing each other. The
result is again the dominant eigenvector of a suitable matrix. Stochastic approach
for link-structure analysis (SALSA) [Lempel and Moran 01] is a more recent and
strictly related score based on the same idea, with the difference that it applies
some normalization to the matrix.
We conclude this brief historical account mentioning that there were, in the
past, some attempts to axiomatize the notion of centrality: we postpone a dis-
cussion on these attempts to Section 4.4.

3. Definitions and Conventions

In this study, we consider directed graphs defined by a set N of n nodes and a set
A ⊆ N × N of arcs; we write x → y when x, y ∈ A and call x and y the source
and target of the arc, respectively. An arc with the same source and target is
called a loop.
The transpose of a graph is obtained by reversing all arc directions (i.e., it has
an arc y → x for every arc x → y of the original graph). A symmetric graph is a
graph such that x → y whenever y → x; such a graph is fixed by transposition
and can be identified with an undirected graph, that is, a graph whose arcs
(usually called edges) are a subset of unordered pairs of nodes. A successor of x
is a node y such that x → y, and a predecessor of x is a node y such that y → x.
The outdegree d+ (x) of a node x is the number of its successors, and the indegree
d− (x) is the number of its predecessors.
A path (of length k) is a sequence x0 , x1 , . . . , xk −1 , where xj → xj +1 ,
0 ≤ j < k. A walk (of length k) is a sequence x0 , x1 , . . . , xk −1 , where xj → xj +1
or xj +1 → xj , 0 ≤ j < k. A connected (strongly connected, respectively) compo-
nent of a graph is a maximal subset in which every pair of nodes is connected
by a walk (path, respectively). Components form a partition of the nodes of a

5 To be precise, Kleinberg’s algorithm works in two phases; in the first phase, one selects
a subgraph of the starting web graph based on the pages that match the given query; in the
second phase, the centrality score is computed on the subgraph. Because in this paper we
are looking at HITS simply as a centrality index, we will simply apply it to the graph under
examination.
228 Internet Mathematics

graph. A graph is (strongly) connected if there is a single (strongly) connected

component, that is, for every choice of x and y there is a walk (path) from x to
y. A strongly connected component is terminal if its nodes have no arc toward
other components.
The distance d(x, y) from x to y is the length of the shortest path from x
to y, or ∞ if no such path exists. The nodes reachable from x are the nodes y
such that d(x, y) < ∞. The nodes coreachable from x are the nodes y such that
d(y, x) < ∞. A node has a trivial (co)reachable set if the latter contains only the
node itself.
The notation Ā, where A is a nonnegative matrix, will be used throughout
the article to denote the matrix obtained by 1 -normalizing the rows of A, that
is, dividing each element of a row by the sum of the row (null rows are left un-
changed). If there are no null rows, Ā is (row-)stochastic, that is, it is nonnegative
and the row sums are all equal to one.
We use Iverson’s notation: if P is a predicate, [P ] has value 0 if P is false and 1

if P is true [Knuth 92]; we denote with Hi the ith harmonic number 1≤k ≤i 1/k;
finally, χi (j) = [j = i] is the characteristic vector of i. We number graph nodes
and corresponding vector coordinates starting from zero.

3.1. Geometric Measures

We call geometric those measures assuming that importance is a function of
distances; more precisely, a geometric centrality depends only on how many
nodes exist at every distance. These are some of the oldest measures defined in
the literature.

3.1.1. Indegree. Indegree, the number of incoming arcs d− (x), can be considered
a geometric measure: it is simply the number of nodes at distance one.6 It is
probably the oldest measure of importance ever used, because it is equivalent
to majority voting in elections (where x → y if x voted for y). Indegree has
a number of obvious shortcomings (e.g., it is easy to spam), but it is a good
baseline, and in some cases it turned out to provide better results than more
sophisticated methods (see, e.g., [Craswell et al. 03]).

6 Most centrality measures proposed in the literature were actually described only for undi-
rected, connected graphs. Because the study of web graphs and online social networks has
posed the problem of extending centrality concepts to networks that are directed, and possi-
bly not strongly connected, in the rest of this study we consider measures depending on the
incoming arcs of a node (e.g., incoming paths, left dominant eigenvectors, distances from all
nodes to a fixed node). If necessary, these measures can be called “negative,” opposed to the
“positive” versions obtained by considering outgoing paths, or (equivalently) by transposing
the graph.
Boldi and Vigna: Axioms for Centrality 229

3.1.2. Closeness. Bavelas introduced closeness in the late forties [Bavelas 50]; the
closeness of x is defined by
1
 . (3.1)
y d(y, x)
The intuition behind closeness is that nodes that are more central have smaller
distances, and thus a smaller denominator, resulting in a larger centrality. We
remark that for this definition to make sense, the graph must be strongly con-
nected. Lacking that condition, some of the denominators will be ∞, resulting
in a null score for all nodes that cannot coreach the whole graph.
It was probably not Bavelas’s intent to apply the measure to directed graphs,
and even less so to graphs with infinite distances; nonetheless, closeness is some-
times “patched” by simply not including unreachable nodes, that is,
1
 ,
d(y ,x)< ∞ d(y, x)
and assuming that nodes with an empty coreachable set have centrality 0 by
definition: this is actually the definition we shall use in the rest of the article.
These apparently innocuous adjustments, however, introduce a strong bias to-
ward nodes with a small coreachable set.

3.1.3. Lin’s Index. [Lin 76] tried to repair the definition of closeness for graphs
with infinite distances by weighting closeness using the square of the number of
coreachable nodes; his definition for the centrality of a node x with a nonempty
coreachable set is
 
{y | d(y, x) < ∞}2
 .
d(y ,x)< ∞ d(y, x)

The rationale behind this definition is as follows: first, we consider closeness not
the inverse of a sum of distances, but rather the inverse of the average distance,
which entails a first multiplication by the number of coreachable nodes. This
change normalizes closeness across the graph. Now, however, we want nodes with
a larger coreachable set to be more important, given that the average distance is
the same, so we multiply again by the number of coreachable nodes. Nodes with
an empty coreachable set have centrality 1 by definition.
Lin’s index was (somewhat surprisingly) ignored in the following literature.
Nonetheless, it seems to provide a reasonable solution for the problems caused
by the original definition of closeness.

3.1.4. Harmonic Centrality. As we noticed, the main problem of closeness lies in the
presence of pairs of unreachable nodes. We thus get inspiration from [Marchiori
230 Internet Mathematics

and Latora 00]: faced with the problem of providing a sensible notion of “average
shortest path” for a generic directed network, they propose to replace the average
distance with the harmonic mean of all distances (i.e., the n(n − 1) distances
between every pair of distinct nodes). Indeed, in case a large number of pairs
of nodes are not reachable, the average of finite distances can be misleading:
a graph might have a very low average distance while it is almost completely
disconnected (e.g., a perfect matching has average distance 1/2). The harmonic
mean has the useful property of handling ∞ cleanly (assuming, of course, that
∞−1 = 0). For example, the harmonic mean of distances of a perfect matching
is n − 1: in fact, for every node there is exactly another node at a noninfinite
distance, and its distance is 1; so the sum of the inverse of all distances is n,
making the harmonic average equal to n(n − 1)/n = n − 1.
In general, for each graph-theoretical notion based on arithmetic averaging
or maximization, there is an equivalent notion based on the harmonic mean. If
we consider closeness the reciprocal of a denormalized average of distances, it
is natural to consider also the reciprocal of a denormalized harmonic mean of
distances. We thus define the harmonic centrality of x as7
 1  1
= . (3.2)
d(y, x) d(y, x)
y =x d(y ,x)< ∞,y =x

The difference with (3.1) might seem minor, but actually it is a radical change.
Harmonic centrality is strongly correlated to closeness centrality in simple net-
works, but naturally also accounts for nodes y that cannot reach x. Thus, it can
be fruitfully applied to graphs that are not strongly connected.

3.2. Spectral Measures

Spectral measures compute the left dominant eigenvector of some matrix derived
from the graph, and depending on how the matrix is modified before the compu-
tation, we can obtain a number of different measures. Existence and uniqueness of
such measures is usually derivable by the theory of nonnegative matrices started
by Perron and Frobenius [Berman and Plemmons 94]; we will not, however,

7 We remark that Yannick Rochat, in a talk at Application of Social Network Analysys

(ASNA 2009) observed that in an undirected graph with several disconnected components, the
inverse of the harmonic mean of distances offers a better notion of centrality than closeness,
because it weights less those elements that belong to smaller components. Tore Opsahl made
the same observation in a March 2010 blog posting. Raj Kumar Pan and Jari Saramäki deviated
from the classical definition of closeness in [Pan and Saramäki 11], using, in practice, harmonic
centrality, with the motivation of better handling disconnected nodes (albeit with no reference
to the harmonic mean). Edith Cohen and Haim Kaplan defined in [Cohen and Kaplan 07] the
notion of a spatially decaying aggregate, of which harmonic centrality is a particular instance.
Boldi and Vigna: Axioms for Centrality 231

discuss such issues, because there is a large body of established literature about
the topic. All observations in this section are true for strongly connected graphs;
the modifications for graphs that are not strongly connected can be found in
[Berman and Plemmons 94, Vigna 09].

3.2.1. The Left Dominant Eigenvector. The first and most obvious spectral measure is the
left dominant eigenvector of the plain adjacency matrix. Indeed, the dominant
eigenvector can be thought of as the fixed point of an iterated computation in
which every node starts with the same score and then replaces its score with
the sum of the scores of its predecessors. The vector is then normalized, and the
process is repeated until convergence.
Dominant eigenvectors fail to behave as expected on graphs that are not
strongly connected. Depending on the dominant eigenvalue of the strongly con-
nected components, the dominant eigenvector might or might not be nonzero on
nonterminal components (a detailed characterization can be found in [Berman
and Plemmons 94]).

3.2.2. Seeley’s Index. The dominant eigenvector rationale can be slightly amended
with the observation that the update rule we described can be thought of as if
each node gives away its score to its successors; or even that each node has a
reputation and is giving its reputation to its successors so that they can build
their own.
Once we take this viewpoint, it is clear that it is not very sensible to give away
our own amount of reputation to everybody; it is more reasonable to divide
equally our reputation among our successors. From a linear-algebra viewpoint,
this corresponds to normalizing each row of the adjacency matrix using the 1
norm.
Seeley advocated this approach [Seeley 49] for computing the popularity among
groups of children, given a graph representing whether each child liked another
one. The matrix resulting from the 1 -normalization process is stochastic, so
the score can be interpreted as the stationary state of a Markov chain. In par-
ticular, if the underlying graph is symmetric, then Seeley’s index collapses to
a degree (modulo normalization) because of the well-known characterization of
the stationary distribution of the natural random walk on a symmetric graph.
Also, Seeley’s index does not react very well to the lack of strong connectivity:
the only nodes with a nonzero score are those belonging to terminal components
that are not formed by a single node of outdegree zero.

3.2.3. Katz’s Index. Katz introduced his celebrated index [Katz 53] using a sum-
mation over all paths coming into a node, but weighting each path so that the
232 Internet Mathematics

summation would be finite. Katz’s index can be expressed as

∞

k=1 β i Ai ,
i=0

because of the interplay between the powers of the adjacency matrix and the
number of paths connecting two nodes. For the summation above to be finite,
the attenuation factor β must be smaller than 1/λ, where λ is the dominant
eigenvalue of A.
Katz immediately noted that the index was expressible using linear algebra
operations:
k = 1(1 − βA)−1 .
It took some more time to realize that, due to Brauer’s theorem on the displace-
ment of eigenvalues [Brauer 52], Katz’s index is the left dominant eigenvector of
a perturbed matrix
βλA + (1 − βλ)eT 1, (3.3)
where e is a right dominant eigenvector of A such that 1eT = λ [Vigna 09].
An easy generalization (actually suggested by [Hubbell 65]) replaces the vec-
tor 1 with some preference vector v so that paths are also weighted differently
depending on their starting node.8
The normalized limit of Katz’s index when β → 1/λ is a dominant eigenvector;
if the dominant eigenvector is not unique, the limit depends on the preference
vector v [Vigna 09].

3.2.4. PageRank. PageRank is one of the most discussed and quoted spectral indices
in use today, mainly because of its alleged use in Google’s ranking algorithm.9
By definition, PageRank is the unique vector p satisfying
p = αpĀ + (1 − α)v, (3.4)
where Ā is again the 1 -normalized adjacency matrix of the graph, α ∈ [0 . . 1) is
a damping factor, and v is a preference vector (which must be a distribution).

8 We
∞
β i Ai =
∞ must note that the original definition of Katz’s index is 1A
∞ i i i= 0
1/β i+ 1 i+ 1
i= 0
β A = (1/β) i= 0
β A − 1/β. This additional multiplication by A is some-
what common in the literature, even for PageRank; clearly, it alters the order induced by the
scores only when there is a nonuniform preference vector. Our discussion can be easily adapted
for this version.
9 The reader should be aware, however, that the literature about the actual effectiveness of
PageRank in information retrieval is rather scarce, and comprises mainly negative results such
as those found in [Najork et al. 07a] and [Craswell et al. 03].
Boldi and Vigna: Axioms for Centrality 233

This is the definition appearing in Brin and Page’s paper about Google [Brin and
Page 98]; the authors claim that the PageRank score is a probability distribution
over webpages, that is, it has unit 1 norm; but this is not necessarily true if A has
null rows. The following scientific literature thus proposed several ways to patch
the matrix Ā to make it stochastic, which would guarantee p 1 = 1. A common
solution is to replace every null row with the preference vector v itself, but
other solutions have been proposed (e.g., adding a loop to all nodes of outdegree
zero), which have led to different scores. This issue is definitely not academic;
in typical web snapshots a significant fraction of the nodes have outdegree zero
(the so-called dangling nodes).
It is interesting to note, however, that in the preprint written in collaboration
with Motwani and Winograd and commonly quoted as defining PageRank [Page
et al. 98], Brin and Page themselves propose a different but essentially equivalent
linear recurrence in the style of Hubbell’s index [Hubbell 65] and acknowledge
that Ā can have null rows, in which case the dominant eigenvalue of Ā could be
smaller than one, and the solution might need to be normalized to have unit 1
norm.
(3.4) is of course solvable even without any patching, giving
 −1
p = (1 − α)v 1 − αĀ , (3.5)

and finally
∞

p = (1 − α)v αi Āi , (3.6)
i=0

which shows immediately that Katz’s index and PageRank differ only by a con-
stant factor and by the 1 normalization applied to A, similarly to the difference
between the dominant eigenvector and Seeley’s index.
If A has no null rows, or Ā has been patched to be stochastic, PageRank
can be equivalently defined as the stationary distribution (i.e., the left dominant
eigenvector) of the Markov chain with transition matrix

αĀ + (1 − α)1T v, (3.7)

which is of course analogous to (3.3). [Del Corso et al. 06] have shown that
the resulting scores (which have always unit 1 norm) differ from the PageRank
vector defined by (3.4) only by a normalization factor, provided that in (3.7)
null rows have been replaced with v. If A had no null rows, the scores of course
would be identical as (3.5) can be easily derived from (3.7).
Both definitions have been used in the literature: the linear-recurrence def-
inition (3.4) is particularly useful when one needs linear dependence on v
234 Internet Mathematics

[Boldi et al. 05]. The Markov-chain definition (3.7) is nonetheless more com-
mon, albeit it raises the issue of patching null rows.
Analogously to what happens with Katz’s index, if the dominant eigenvalue of
Ā is one, then the limit of PageRank when α goes to 1 is a dominant eigenvector
of Ā, that is, Seeley’s index, and if the dominant eigenvector is not unique, the
limit depends on the preference vector v [Boldi et al. 09].
In the rest of this article, except when explicitly stated, we shall define PageR-
ank as in (3.5), without any patching or normalization, and use a uniform pref-
erence vector.

3.2.5. HITS. Kleinberg introduced his celebrated HITS algorithm [Kleinberg 99]10
using the web metaphor of “mutual reinforcement”: a page is authoritative if it
is pointed to by many good hubs—pages that contain good lists of authoritative
pages— and a hub is good if it points to authoritative pages. This suggests an
iterative process that computes at the same time an authoritativeness score ai
and a “hubbiness” score hi starting with a0 = 1, and then applying the update
rule

hi+1 = ai AT
ai+1 = hi+1 A.

This process converges to the left dominant eigenvector of the matrix AT A, which
gives the final authoritativeness score, and which we label with HITS throughout
the paper.11
Inverting the process, and considering the left dominant eigenvector of the
matrix AAT , gives the final hubbiness score. The two vectors are the left and
right singular vectors associated with the largest singular value in the singular-
value decomposition of A. Note also that hubbiness is the positive version of
authoritativeness.

3.2.6. SALSA. Finally, we consider SALSA, a measure introduced by [Lempel and

Moran 01] always using the metaphor of mutual reinforcement between author-
itativeness and hubbiness, but 1 -normalizing the matrices A and AT . We start

1 0 A few years before, [Bonacich 91] had introduced the same method for computing the
centrality of individuals and groups related by a rectangular incidence matrix. Kleinberg’s
approach is slightly more general, as it does not assume that the two scores are assigned to
different types of entities.
1 1 As discussed in [Farahat et al. 06], the dominant eigenvector may not be unique; equiva-
lently, the limit of the recursive definition given above may depend on the way the authority
and hub scores are initialized. Here, we consider the result of the iterative process starting with
a0 = 1.
Boldi and Vigna: Axioms for Centrality 235

with a0 = 1 and proceed with

hi+1 = ai AT
ai+1 = hi+1 Ā.

We remark that this normalization process is analogous to the one that brings
us from the dominant eigenvector to Seeley’s index or from Katz’s index to
PageRank.
Similarly to what happens with Seeley’s index on symmetric graphs, SALSA
does not need such an iterative process to be computed.12 First, one computes
the connected components of the symmetric graph induced by the matrix AT A;
in this graph, x and y are adjacent if x and y have some common predecessor
in the original graph. Then, the SALSA score of a node is the ratio between its
indegree and the sum of the indegrees of nodes in the same component, multiplied
by the ratio between the component size and n. Thus, contrarily to HITS, a single
linear scan of the graph is sufficient to compute SALSA, albeit the computation

of the intersection graph requires time proportional to x d+ (x)2 .

3.3. Path-Based Measures

Path-based measures exploit not only the existence of shortest paths but actually
examine all shortest paths (or all paths) coming into a node. We remark that
indegree can be considered a path-based measure, because it is the equivalent to
the number of incoming paths of length one.

3.3.1. Betweenness. Betweenness centrality was introduced by [Anthonisse 71] for

edges, and then rephrased by Freeman for nodes [Freeman 77]. The idea is to
measure the probability that a random shortest path passes through a given
node: if σy z is the number of shortest paths going from y to z, and σy z (x) is the
number of such paths that pass through x, we define the betweenness of x as
 σy z (x)
.
σy z
y ,z =x,σ y z =0

The intuition behind betweenness is that if a large fraction of shortest paths

passes through x, then x is an important junction point of the network. In-
deed, removing nodes in betweenness order causes a very quick disruption of the
network [Boldi et al. 13].

1 2 This property, which appears to be little known, is proved in Proposition 2 of the original
paper [Lempel and Moran 01].
236 Internet Mathematics

3.3.2. Spectral Measures as Path-Based Measures. It is a general observation that all spectral
measures can be interpreted as path-based measures because they depend on
taking the limit of some summations of powers of A, or on the limit of powers
of A, and in both cases we can express these algebraic operations in terms of
suitable paths.
For instance, the left dominant eigenvector of a nonnegative matrix can be
computed with the power method by taking the limit of 1Ak / 1Ak for k → ∞.
Since, however, 1Ak is a vector associating with each node the number of paths
of length k coming into the node, we can see that the dominant eigenvector
expresses the relative growth of the number of paths coming into each node as
their length increases.
Analogously, Seeley’s index can be computed (modulo a normalization factor)
by taking the limit of 1Āk (in this case, the 1 norm cannot grow, so we do
not need to renormalize at each iteration). The vector 1Āk has the following
combinatorial interpretation: it assigns to each x the sums of the weights of the
paths coming into x, where the weight of a path x0 , x1 , . . . , xt is

t−1
1
. (3.8)
i=0
d+ (xi )
When we switch to attenuated versions of the previous indices (that is, Katz’s
index and PageRank), we switch from limits to infinite summations and at the
same time multiply the weight of paths of length t by β t or αt . Indeed, the Katz
index of x was originally defined as the summation over all t of the number of
paths of length t coming into x multiplied by β t , and PageRank is the summation
over all paths coming into x of the weight (3.8) multiplied by αt .
The reader can easily work out similar definitions for HITS and SALSA,
which depend on a suitable definition of alternate “back-and-forth path” (see,
e.g., [Borodin et al. 05]).

4. Axioms for Centrality

The comparative evaluation of centrality measures is a challenging, difficult,
arduous task, for many different reasons. The datasets that are classically used in
social sciences are very small (typically, some tens of nodes), and it is difficult to
draw conclusions out of them. Nonetheless, some attempts in this direction were
put forth (e.g., [Stephenson and Zelen 89]); sometimes, the attitude was actually
to provide evidence that different measures highlight different kinds of centralities
and are, therefore, equally incomparably interesting [Friedkin 91]. Whether the
latter consideration is the only sensible conclusion or not is debatable. Although
Boldi and Vigna: Axioms for Centrality 237

it is clear that the notion of centrality, in its vagueness, can be interpreted

differently, giving rise to many good but incompatible measures, we will provide
evidence that some measures tend to reward nodes that are in no way central.
If results obtained on small corpora may be misleading, a comparison on larger
corpora is much more difficult to deal with, because of the lack of ground truth
and the unavailability of implementations of efficient algorithms to compute the
measures under consideration (at least in cases where efficient, possibly approx-
imate, algorithms do exist). Among the few attempts that try a comparison on
large networks we cite [Upstill et al. 03] and [Najork et al. 07b], which never-
theless focus only on web graphs and on a very limited number of centrality
indices.
In this study, we propose to understand (part of) the behavior of a centrality
measure using a set of axioms. Of course, it is not sensible to prescribe a set of
axioms that define what centrality should be (in the vein of Shannon’s defini-
tion of entropy [Shannon 48] or Altman and Tennenholtz axiomatic definition
of Seeley’s index [Altman and Tennenholtz 05]13 ), however, as different indices
serve different purposes, it is reasonable to set up some necessary axioms that
an index should satisfy to behave predictably and follow our intuition.
The other interesting aspect of defining axioms is that, even if one does not
believe they are really so discriminative or necessary, they provide a very specific,
formal, provable piece of information about a centrality measure that is much
more precise than folklore intuitions such as “this centrality is really correlated
to indegree” or “this centrality is really fooled by cliques.” We believe that a
theory of centrality should provide exactly this kind of compact, meaningful,
reusable information (in the sense that it can be used to prove other properties).
This is indeed what happens, for example, in topology, where the information
that a space is T0 , rather than T1 , is a compact way to provide a lot of knowledge
about the structure of the space.
Defining such axioms is a delicate matter. First of all, the semantics of the
axioms must be very clear. Second, the axioms must be evaluable in an exact
way on the most common centrality measures. Third, they should be formulated
to avoid the trap of small, finite (counter)examples, on which many centrality
measures collapse (e.g., using an asymptotic definition as suggested by Lempel
and Moran for rank similarity and stability [Lempel and Moran 05]). We assume
from the beginning that the centrality measures under examination are invariant

1 3 The authors claim to formalize PageRank [Page et al. 98], but they do not consider the
damping factor (equivalently, they are setting α = 1), so they are actually formalizing Seeley’s
venerable index [Seeley 49].
238 Internet Mathematics

by isomorphism, that is, they depend just on the structure of the graph, and not
on particular labeling chosen for each node.
To meet these constraints, we propose to study the reaction of centrality mea-
sures to change of size, to (local) change of density and to arc additions. We
expect that nodes belonging to larger groups, when every other parameter is
fixed, should be more important, and that nodes with a denser neighborhood
(i.e., having more friends), when every other parameter is fixed, should also be
more important. We also expect that adding an arc should increase the impor-
tance of the target.
How can we actually determine if this happens in an exact way, and possibly
in an asymptotic setting? To do so, we need to try something entirely new—
evaluating exactly (i.e., in algebraic closed form) all measures of interest on all
nodes of some representative classes of networks.
A good approach to reduce the amount of computation is to use strongly con-
nected vertex-transitive 14 graphs as basic building blocks: these graphs exhibit
a high degree of symmetry, which should entail a simplification of our computa-
tions. Finally, since we want to compare density, a natural choice is to pick the
densest strongly connected vertex-transitive graph, the clique, and the sparsest
strongly connected, the directed cycle. Choosing two graphs at the extreme of
the density spectrum should better highlight the reaction of centrality measures
to density. Moreover, k-cliques and directed p-cycles exist for every k and p (this
might not happen for more complicated structures, e.g., a cubic graph).

4.1. The Size Axiom

Let us consider a graph made by a k-clique and a p-cycle (see the figure in
Table 1).15 Because of invariance by isomorphism, all nodes of the clique have
equal score, and all nodes of the cycle have equal score, too; but which nodes are
more important? Probably everybody would answer that if p = k the elements
on the clique are more important, and indeed this axiom is so trivial that is
satisfied by almost any measure of which we are aware, but we are interested in
assessing the sensitivity to size,16 and thus we state our first axiom as follows:

1 4 A graph is vertex-transitive if for every nodes x and y there is an automorphism exchanging

x and y.
1 5 The graph is of course disconnected. It is a common theme of this work that centrality
measures should work also on graphs that are not strongly connected, for the very simple
reason that we meet this kind of graphs in the real world, the web being a common example.
1 6 [Lempel and Moran 01] while designing their own centrality measure observe that “the
size of the community should be considered when evaluating the quality of the top pages in
that community”.
Boldi and Vigna: Axioms for Centrality 239

Definition 4.1. (Size Axiom). Consider the graph Sk ,p made by a k-clique and a directed
p-cycle (Table 1). A centrality measure satisfies the size axiom if for every k there
is a Pk such that for all p ≥ Pk in Sk ,p the centrality of a node of the p-cycle
is strictly larger than the centrality of a node of the k-clique, and if for every
p there is a Kp such that for all k ≥ Kp in Sk ,p the centrality of a node of the
k-clique is strictly larger than the centrality of a node of the p-cycle.

Intuitively, when p = k we do expect nodes of the cycle to be less important

than nodes of the clique. The rationale behind the case k → ∞ is rather clear:
the denser community is also getting larger, and thus its members are expected
to become ultimately even more important.
However, if the cycle becomes very large (more precisely, when its size goes
to infinity), its nodes are still part of a very large (albeit badly connected) com-
munity, and we expect them to achieve ultimately greater importance than the
nodes of a fixed-size community, no matter how dense it can be.
Because one might devise some centrality measures that satisfy the size axiom
for p and not for k, stating both properties in Definition 4.1. gives us a finer
granularity and avoids pathological cases.

4.2. The Density Axiom

Designing an axiom for density is a more delicate issue, because we must be able
to define an increase of density “with all other parameters fixed,” including size.
Let us start, ideally, from a graph made by a directed k-cycle and a directed
p-cycle, and connect a node x of the k-cycle with a node y of the p-cycle through
a bidirectional arc, the bridge. If k = p, the vertices x and y are symmetric,
and thus must necessarily have the same score. Now, we increase the density
of the k-cycle as much as possible, turning it into a k-clique (see the figure in
Table 2). Note that this change of density is local to x, as the degree of y has
not changed. We are thus strictly increasing the local density around x, leaving
all other parameters fixed, and in these circumstances we expect that the score
of x increases.

Definition 4.2. (Density axiom).

Consider the graph Dk ,p made by a k-clique and a p-
cycle (p, k ≥ 3) connected by a bidirectional bridge x ↔ y, where x is a node of
the clique and y is a node of the cycle (Table 2). A centrality measure satisfies the
density axiom if for k = p the centrality of x is strictly larger than the centrality
of y.
240 Internet Mathematics

Note that our axiom does not specify any constraint when k = p. While study-
ing the behavior of the graph Dk ,p of the previous definition when k = p sheds
some light on the inner behavior of centrality measures, it is essential, in an
axiom asserting the sensitivity to density, that size is not involved.
In our proofs for the density axiom, we actually let k and p be independent
parameters (even if the axiom is stated for k = p) to compute the watershed,
that is, the value of k (expressed as a function of p) at which the centrality of
x becomes larger than the centrality of y (if any). The watershed can give some
insight as to how badly a measure can miss satisfying the density axiom.

4.3. The Score-Monotonicity Axiom

Finally, we propose an axiom that specifies strictly monotonic behavior on the
addition of an arc:

Definition 4.3. (Score–Monotonicity Axiom).

A centrality measure satisfies the score-
monotonicity axiom if for every graph G and every pair of nodes x, y such
that x → y, when we add x → y to G the centrality of y increases.

In some sense, this axiom is trivial: it is satisfied by essentially all centrality

measures we consider on strongly connected graphs. Thus, it is an excellent test
to verify that a measure is able to correctly handle partially disconnected graphs.
The reader might be tempted to define a weak monotonicity axiom, which
just requires the rank of y to be nondecreasing. However, the centrality measure
associating a constant value to every node of every network would satisfy such
an axiom, which makes it not very interesting for our goals if used in isolation.
Nonetheless, a form of rank monotonicity would be an essential counterpoint
to score monotonicity, because it would detect pathological centralities in which
the addition of an arc toward y does increase the score of y, but also increases,
in a counterintuitive way, the score of other nodes, resulting in a rank decrease
of y. We leave the statement and the study of such an axiom for future work.

4.4. Previous Attempts at Axiomatization

The idea of formalizing through axioms the behavior of centrality indices is
not new: Sabidussi, in his much-quoted paper on centrality [Sabidussi 66],17
describes in a section named “Axioms for Centrality” a set of axioms that should

1 7 This paper contributed so much to the popularization of Bavelas’s closeness centrality

that the latter is often called “Sabidussi’s centrality” in the literature.
Boldi and Vigna: Axioms for Centrality 241

be satisfied by a sensible centrality on an undirected graph. Indeed, one of his

conditions is identical to our score-monotonicity axiom (see Section 4.3).
A few years later, [Nieminen 73] attempted a similar formalization for directed
graphs. His axioms are meaningful only for geometric centralities because they
are stated in terms of the number of nodes at a given distance. Consider the
negative neighborhood function of x in a graph G defined by
 
NG− (x, t) = {y | d(y, x) ≤ t }.
Nieminen essentially states that if the neighborhood function of x dominates
strictly that of y (i.e., NG− (x, t) ≥ NG− (y, t) for all t and NG− (x, t) > NG− (y, t) for at
least one t) then the centrality of x must be strictly larger than that of y. Because
the new neighborhood function of y after the addition of an arc toward y always
dominates strictly the old one, Nieminen’s axiom implies score monotonicity. He
notes, indeed, that Bavelas’s centrality does not satisfy some of his axioms.
The work of [Chien et al. 04], albeit targeted exclusively at PageRank, contains
a definition of score monotonicity equivalent to ours (and Sabidussi’s) and a
definition of rank monotonicity that we think captures the essence of the problem:
when we add an arc toward y, nodes with a score smaller than y must continue
to have a score smaller than that of y, whereas nodes with a score equal to y
must get a score that is smaller than or equal to that of y.
Sabidussi’s work tries to capture rank monotonicity, too, but with much weaker
requirements: he requires that if y has a maximum score in the network, then it
should have a maximum score also after the addition of an arc toward y.
Very recently, [Brandes et al. 12] proposed to call radial 18 a centrality in which
the addition of an arc x → y does not decrease the rank of y, in the sense that
when we add an arc toward y, nodes with a score smaller than or equal to y
continue to have this property. This is actually a weaker condition than the one
used in [Chien et al. 04] to state rank monotonicity; in fact, it makes a form of
non-monotonicity of ranks possible, because nodes with a score smaller than y
might end up having a score equal to y when we add an arc toward y. The authors
of [Brandes et al. 12] propose also a different axiom for undirected networks.

5. Proofs and Counterexamples

We have finally reached the core of this work: because we are considering 11
centralities and 3 axioms, we have to verify 33 statements. For the size and

1 8 There are actually two notions of radiality, which correspond to our notions of “positive”
and “negative” centralities.
242 Internet Mathematics

density axioms, we compute in closed form the values of all measures, from which
we can derive the desired results, whereas, for the score-monotonicity axiom we
provide direct proofs or counterexamples.
We remark that in all our tables we use the proportionality symbol ∝ to mark
values that have been rescaled by a common factor to make them more readable.

5.1. Size
Table 1 provides scores for the graph Sp,k , from which we can check whether the
size axiom is satisfied. The scores are immediately computable from the basic

Centrality k-clique p-cycle

Degree k−1 1
Harmonic k−1 Hp −1
1 2
Closeness
k−1 p(p − 1)
k2 2p
Lin
k−1 p−1
(p − 1)(p − 2)
Betweenness 0
2
Dominant ∝ 1 0
Seeley ∝ 1 1
1 1
Katz
1 − (k − 1)β 1−β
PageRank ∝ 1 1
HITS ∝ 1 0
SALSA ∝ 1 1

Table 1. Centrality scores for the graph Sk , p . Hi denotes the ith harmonic num-
ber. The parameter β is Katz’s attenuation factor.
Boldi and Vigna: Axioms for Centrality 243

definitions; as we noticed, Sk ,p is highly symmetrical, and so there are only two

actual scores—the score of a node of the clique and the score of a node of the
cycle. In the case of some spectral centrality measures, there are several possible
solutions, in which case we use the one returned by the power method starting
from the uniform vector.

5.2. Density
Table 2 provides scores for the graph Dp,k . Because the graph is strongly con-
nected, there is no uniqueness issue. Although the computation of geometric and
path-based centrality measures is a tedious but rather straightforward exercise
(it is just a matter of finite summations), spectral indices require some more care.
In the rest of this section, we shall first detail the steps leading to the explicit
formulae of Table 2. Then we will prove that the density axiom holds.

5.2.1. Explicit Formulae for Spectral Indices. We will write the parametric equations ex-
pressing the matrix computation that defines each spectral index and solve them.
As noted before, even if the axiom requires k = p, we prefer to perform the com-
putation with two independent parameters k and p to compute the watershed.
In all cases, we can always use the bounds imposed by symmetry to write
down just a small number of variables: c for the centrality of an element of the
clique,  for the clique bridge (“left”), r for the cycle bridge (“right”), and some
function t(d) of the distance from the cycle bridge for the nodes of the cycle
(with 0 ≤ d < p), with the condition t(0) = r.
• The left dominant eigenvector. In this case, the equations are given by
the standard eigenvalue problem of the adjacency matrix:

λ = r + (k − 1)c
λc =  + (k − 2)c
r
λr =  + p−1 ,
λ
subject to the condition that we choose λ real and positive with maximum abso-
lute value. Note that in the case of the last equation we “unrolled” the equations
about the elements of the cycle, λt(d + 1) = t(d). Solving the system and choos-
ing c = 1/(λ − k + 1) gives the solutions found in Table 2.
Since for nonnegative matrices the dominant eigenvalue is monotone in the
matrix entries, λ ≥ k − 1, because the k-clique has dominant eigenvalue equal
to k − 1. Otherwise, λ ≤ k by the row-sum bound. As the eigenvalue equations
have no solution for λ = k − 1, we conclude that k − 1 < λ ≤ k.
 244

Centrality Clique Clique Bridge Cycle Bridge Cycle (d > 0 from the Bridge) Watershed

Degree k −1 k 2 1 —

k −1 1 k −1
Harm onic k − 2 + Hp + 1 k − 1 + Hp 1+ + H p −1 + + H p −1 —
2 d+1 d+2
1 1 1 1
Closeness k ≤p
k − 1 + 2p + p(p − 1)/2 k − 1 + p + p(p − 1)/2 2k − 1 + p(p − 1)/2 k (d + 2) − 1 + p(p − 1)/2
(p − 1)(p − 2) (p − 1)(p − 2) p2 + p + 2
Betweenness 0 2p(k − 1) 2k (p − 1) + 2k (p − 2) + k≤
2 2 4
1 1 1+λ
Dom inant 1+ 1+λ —
λ −k + 1 λ −k + 1 λd
Seeley ∝ k −1 k 2 1 —

1 + β 1 β 1 βd+ 1
Katz ∝  +  +  —
1 − β (k − 2) 1 −β 1 − βp 1 −β 1 − βp
(k − 1)(k − α k + α ) α − k α − k
PageRank ∝  2+2 1 + αd —
k (k − 1 − α (k − 2)) k (2 − α p ) k (2 − α p )

HITS ∝ μ 2 − μ(k + 1) + k − 1 (k − 1)(k − 2)(μ − 1) μ 3 − (k 2 − 2k + 4)μ 2 [d = 1](k − 1)(k − 2) —

+ (3k 2 − 7k + 6)μ − (k − 1) 2
SALSA ∝ (k − 1)(k + 2) k (k + 2) 2(k + 2) k + 2 + [d = 1](k 2 − 2k + 2) —

Table 2. Centrality scores for the graph Dk , p . The parameter β is Katz’s attenuation factor, α is PageRank’s damping
Internet Mathematics

factor, λ is the dominant eigenvalue of the adjacency matrix A and μ is the dominant eigenvalue of the matrix AT A. Lin’s
centrality is omitted because it is proportional to closeness (the graph being strongly connected).
Boldi and Vigna: Axioms for Centrality 245

• Katz’s Index. In this case, the equations can be obtained by the standard
technique of “taking one summand out,” that is, writing
∞
 ∞
  ∞
k=1 i i
β A =1+1 i i
β A =1+ 1 β i Ai βA = 1 + kβA.
i=0 i=1 i=0

The equations are then

 = 1 + βr + β(k − 1)c
c = 1 + β + β(k − 2)c

1 − β p−1
r = 1 + β + β + β p−1 r ,
1−β

where again we “unrolled” the equations about the elements of the cycle, as we
would have just t(d + 1) = 1 + βt(d), so

1 − βd
t(d) = + β d r.
1−β
The explicit values of the solutions are quite ugly, so we present them in Table 2
as a function of the centrality of the clique bridge .
• PageRank. To simplify the computation, we use 1, rather than 1/(k + p), as
preference vector (by linearity, the result obtained just needs to be rescaled). We
use the same technique employed in the computation of Katz’s index, leading to
1
 = 1 − α + αr + αc
2
α k−2
c=1−α+ +α c
k k−1

α 1
r = 1 − α +  + α 1 − αp−1 + αp−1 r ,
k 2

noting once again that unrolling the equation of the cycle t(1) = 1 − α + αr/2
and t(d + 1) = 1 − α + αt(d) for d > 1 we get
1
t(d) = 1 − αd + αd r.
2
The explicit values for PageRank are even uglier than those of Katz’s index, so
again we present them in Table 2 as a function of the centrality of the clique
bridge .
• Seeley’s Index. This is a freebie; we can just compute PageRank’s limit when
α → 1.
246 Internet Mathematics

• HITS. In this case, we write down the eigenvalue problem for AT A. Writing
t for t(1), we have
μc = (k − 1)c + (k − 2)2 c + (k − 2) + r
μ = k + (k − 1)(k − 2)c + t
μr = 2r + (k − 1)c
μt = t + .
By normalizing the result so that c = μ2 − μ(k + 1) + k − 1, we obtain the com-
plex but somewhat readable values shown in Table 2. Note that p has no role in
the solution, because AT A can be decomposed into two independent blocks, one
of which is an identity matrix corresponding to all elements of the cycle except
for the first two.
• SALSA. It is easy to check that the components of the intersection graph
of predecessors are given by the clique together with the cycle bridge and its
successor and then by one component for each node of the cycle. The computation
of the scores is then trivial using the noniterative method.

5.2.2. Proofs. Armed with our explicit formulation of spectral scores, we have now
to prove whether the density axiom holds, that is, whether  > r when k = p.
Note that in Table 2 we report no watershed for all spectral centrality measures,
which means even more:  > r even when k = p, provided that k, p ≥ 3. The
proofs in this section cover this stronger statement.

Theorem 5.1. HITS satisfies the density axiom.

Proof. As we have seen, we can normalize the solution to the HITS equations so
that
 = (k − 1)(k − 2)(μ − 1)
r = μ3 − (k 2 − 2k + 4)μ2 + (3k 2 − 7k + 6)μ − (k − 1)2 .
Moreover, the characteristic polynomial can be computed explicitly from the set
of equations and by observing that the vectors χk +i and χ0 − χi , 0 < i < p, are
linearly independent eigenvectors for the eigenvalue 1:

p(μ) = μ4 − (k 2 − 2k + 6)μ3 + (5k 2 − 12k + 15)μ2

− (6k 2 − 16k + 14)μ + k 2 − 2k + 1 (μ − 1)k +p−4 .
The largest eigenvalue μ0 satisfies the inequality (k − 1)2 ≤ μ0 ≤ k 2 − 2k + 5/4
for every k ≥ 9 as shown following (the statement of the theorem can be verified
in the remaining cases by explicit computation; it does not depend on p). Using
Boldi and Vigna: Axioms for Centrality 247

the stated upper and lower bounds on μ0 , we can say that

 − r = (k − 1)(k − 2)(μ − 1) − (μ3 − (k 2 − 2k + 4)μ2

+ (3k 2 − 7k + 6)μ − (k − 1)2 )
= −μ3 + (k 2 − 2k + 4)μ2 − (2k 2 − 4k + 4)μ + k − 1
 3
5
≥ − k 2 − 2k + + (k 2 − 2k + 4)(k − 1)4
4

5
− (2k − 4k + 4) k 2 − 2k +
2
+k−1
4
1 19 43 253
= k4 − k3 − k2 + k − ,
4 16 8 64
which is positive for k > 4.
We are left to prove the bounds on μ0 . The lower bound can be easily ob-
tained by monotonicity of the dominant eigenvalue in the matrix entries, be-
cause the dominant eigenvalue of a k-clique is k − 1. For the upper bound,
first we observe that μ0 can be computed explicitly (because it is the solu-
tion of a quartic equation) and, using its expression in closed form, it is pos-
sible to show that limk →∞ μ0 − (k − 1)2 = 0. This guarantees that the bound
μ0 ≤ k 2 − 2k + 5/4 is true ultimately. To obtain an explicit value of k after
which the bound holds true, observe that k 2 − 2k + 5/4 = μ0 implies q(k) = 0,
where q(k) = p(k 2 − 2k + 5/4). Computing the Sturm sequence associated with
q(k), one can prove that q(k) has no zeroes for k ≥ 9, hence our lower bound
on k.

Theorem 5.2. The dominant eigenvector satisfies the density axiom.

Proof. From the proof of Theorem 5.1. we know that λ2 ≤ k2 − 2k + 5/4 for every
√
k ≥ 9, because μ is the spectral norm of A and thus dominates its spectral
radius λ, that is, λ2 ≤ μ. We conclude that

1 −λ2 + (k − 1)λ + 1
−r =1+ − (1 + λ) =
λ−k+1 λ−k+1
−(k 2 − 2k + 5/4) + (k − 1)2 + 1 3
> = > 0.
λ−k+1 4(λ − k + 1)

The remaining cases (k < 9) are verifiable by explicit computation when p ≤ k.

To prove that there is no watershed, we first note that  − r is decreasing in
λ. Then one can use the characteristic equation defining λ and implicit differ-
entiation to write down the values of the derivative of λ as a function of p for
each k < 9. It is then easy to verify that in the range p > k, k − 1 < λ ≤ k, the
248 Internet Mathematics

derivative is always negative, that is, λ is decreasing in p, which completes the

proof.

Theorem 5.3. Katz’s index satisfies the density axiom when β ∈ (0 . . 1/λ).

Proof. Recall that the equations for Katz’s index are

 = 1 + βr + β(k − 1)c
c = 1 + β + β(k − 2)c

1 − β p−1
r = 1 + β + β + β p−1 r .
1−β
First, we remark that as β → 1/λ Katz’s index tends to the dominant eigenvec-
tor, so  > r for β is close enough to 1/λ. Thus, by continuity, we just need to
show that  = r never happens in the range of our parameters. If we solve the
previous equations for c, , and r and impose  = r, we obtain
β2 + k − 2
ln
p= k−1 .
ln β
Now observe that
β2 + k − 2
β≤
k−1
is always true for β ≤ 1 and k ≥ 3. This implies that under the same conditions
p ≤ 1, which concludes the proof.

Theorem 5.4. PageRank with constant preference vector satisfies the density axiom
when α ∈ (0 . . 1).

Proof. The proof is similar to that of Theorem 5.3. Recall that the equations for
PageRank are
1
 = 1 − α + αr + αc
2
α α(k − 2)
c=1−α+ + c
k k−1

α 1
r = 1 − α +  + α 1 − αp−1 + αp−1 r .
k 2
First, we remark that as α → 1 PageRank tends to Seeley’s index, so  > r for α
close enough to 1. By continuity, we just need to show that  = r never happens
in our range of parameters. If we solve the equations above for c, , and r and
Boldi and Vigna: Axioms for Centrality 249

impose  = r, we obtain

2α2 − (k 2 − 4k + 6)α + k 2 − 3k + 2
ln − 2
(k − 3k + 2)α2 − (2k 2 − 3k)α + k 2 − k
p=1+ .
ln α
Now observe that 2α2 − (k 2 − 4k + 6)α + k 2 − 3k + 2 ≥ 0 for k ≥ 3. Thus, a so-
lution for p exists only when the denominator is negative. However, in that region

2α2 − (k 2 − 4k + 6)α + k 2 − 3k + 2
− ≥ 1.
(k 2 − 3k + 2)α2 − (2k 2 − 3k)α + k 2 − k

This implies that under the same conditions p ≤ 1, which concludes the proof.

5.3. Score Monotonicity

In this section, we briefly discuss the only nontrivial cases.
Harmonic. If we add an arc x → y, the harmonic centrality of y can only in-
crease, because this addition can only decrease distances (possibly even turning
some of them from infinite to finite), so it will increase their reciprocals (strictly
increasing the one from x).
Closeness. If we consider a one-arc graph z → y and add an arc x → y, the
closeness of y decreases from 1 to 1/2.
Lin. Consider the graph in Figure 1: the Lin centrality of y is (k + 1)2 /k. After
adding an arc x → y, the centrality becomes (k + 5)2 /(k + 9), which is smaller
than the previous value when k > 3.
Betweenness. If we consider a graph made of two isolated nodes x and y, the
addition of the arc x → y leaves the betweenness of x and y unchanged.

2
y x

k
Figure 1. A counterexample showing that Lin’s index fails to satisfy the score-
monotonicity axiom.
250 Internet Mathematics

x z

Figure 2. A counterexample showing that SALSA fails to satisfy the score-

monotonicity axiom.

Katz. The score of y after adding x → y can only increase because the set of
paths coming into y now contains new elements.19 If the constant vector 1 is
replaced by a preference vector v in the definition, it is necessary that x have a
nonzero score before the addition in order for score monotonicity to hold.
Dominant eigenvector, Seeley’s index, HITS. If we consider a clique and
two isolated nodes x, y, the score given by the dominant eigenvector, Seeley’s
index and HITS to x and y is zero, and it remains unchanged when we add the
arc x → y.
SALSA. Consider the graph in Figure 2: the indegree of y is 1, and its component
in the intersection graph of predecessors is trivial, so its SALSA centrality is
(1/1) · (1/6) = 1/6. After adding an arc x → y, the indegree of y becomes 2, but
now its component is { y, z }; so the sum of indegrees within the component is
2 + 3 = 5, hence the centrality of y becomes (2/5) · (2/6) = 2/15 < 1/6.
PageRank. Score monotonicity of PageRank was proved by Chien, and col-
leagues [Chien et al. 04]. Their proof works for a generic regular Markov chain:
in the case of PageRank this condition is true, for instance, if the preference
vector is strictly positive or if the graph is strongly connected. Score monotonic-
ity under the same hypotheses is also a consequence of the analysis made by
Avrachenkov and Litvak [Avrachenkov and Litvak 06] of the behavior of PageR-
ank when multiple new links are added.
Their result can be extended to a much more general setting. Suppose that
we are adding the arc x → y, with the proviso that the PageRank of x before
adding the arc was strictly positive. We will show that, under this condition,
the score of y will increase for arbitrary graphs and preference vectors. The same
argument shows also that if the score of x is  zero, the
−1score of y does not change.
For this proof, we define PageRank as v 1 − αĀ (i.e., without the normal-
izing factor 1 − α), so to simplify our calculations. By linearity, the result for the
standard definition follows immediately.

1 9 It should be noted, however, that this is true only for the values of the parameter β that
still make sense after the addition.
Boldi and Vigna: Axioms for Centrality 251

Consider two nodes x and y of a graph G such that there is no arc from x to
y, and let d be the outdegree of x. Given the normalized matrix Ā of G, and the
normalized matrix Ā of the graph G obtained by adding to G the arc x → y,
we have
Ā − Ā = χTx δ,
where δ is the difference between the rows corresponding to x in Ā and Ā , which
contains 1/d(d + 1) in the positions corresponding to the successors of x in G,
and −1/(d + 1) in the position corresponding to y (note that if d = 0, we have
just the latter entry).
We now use the Sherman–Morrison formula to write down the inverse of 1 −
αĀ as a function of 1 − αĀ. More precisely,
 −1   −1  −1
1 − αĀ = 1 − α Ā − χTx δ = 1 − αĀ + αχTx δ
 −1  −1
 −1 1 − αĀ αχTx δ 1 − αĀ
= 1 − αĀ −  −1 .
1 + αδ 1 − αĀ χTx
We now multiply by the preference vector v, obtaining the explicit PageRank
correction:
 −1  −1
 −1  −1 1 − αĀ αχTx δ 1 − αĀ
v 1 − αĀ = v 1 − αĀ −v  −1
1 + αδ 1 − αĀ χTx
 −1  −1
αrχTx δ 1 − αĀ αrx δ 1 − αĀ
=r−  −1 =r−  −1 .
1 + αδ 1 − αĀ χTx 1 + αδ 1 − αĀ χTx
 −1
Now remember that rx > 0, and note that 1 − αĀ χTx is the vector of positive
contributions to the PageRank of x, modulo the normalization factor 1 − α. As
such, it is made of positive values adding up to, at most, 1/(1 − α). When the
vector is multiplied by δ, in the worst case (d = 0), we obtain 1/(1 − α), so
given the conditions on α, it is easy to see that the denominator is positive. This
implies that we can gather all constants in a single positive constant c and just
write
 −1   −1
v 1 − αĀ = v − cδ 1 − αĀ .
The above equation rewrites the rank one correction due to the addition of the
arc x → y as a formal correction of the preference vector. We are interested in
the difference
  −1  −1  −1
v − cδ 1 − αĀ − v 1 − αĀ = −cδ 1 − αĀ ,
because we can conclude our proof by just showing that its yth coordinate is
strictly positive.
252 Internet Mathematics

 
We now note that being 1 − αĀ strictly diagonally dominant, the (nonnega-
 −1
tive) inverse B = 1 − αĀ has the property that the entries bii on the diagonal
are strictly larger than off-diagonal entries bk i on the same column [McDonald
et al. 95, Remark 3.3], and in particular they are nonzero. Thus, if d = 0
 −1 c
−cδ 1 − αĀ y
= by y > 0,
d+1
and if d = 0
 −1 c  c c
−cδ 1 − αĀ y
= by y − bz y > by y
d+1 x→z
d(d + 1) d + 1
 c
− by y = 0.
x→z
d(d + 1)

We remark that the above discussion applies to PageRank as defined by (3.5):

if the scores are forced to be 1 -normalized, the score-monotonicity axiom may
fail to hold even under the assumption that the score of x is positive. If we take,
for example, the graph with adjacency matrix
 
00
A=
10
and the preference vector v = (0, 1), we have p = (α(1 − α), 1 − α). Adding one
arc from the first node to the second yields p = (α/(1 + α), 1/(1 + α)); the score
of the second node increases (1/(1 + α) is always larger than 1 − α) as stated in
the theorem. However, the normalized PageRank score of the second node does
not change (it is equal, in both cases, to 1/(1 + α)).

6. Roundup
All our results are summarized in Table 3, where we distilled them into sim-
ple yes/no answers to the question: does a given centrality measure satisfy the
axioms?
It was surprising for us to discover that only harmonic centrality satisfies all ax-
ioms.20 All spectral centrality measures are sensitive to density. Row-normalized
spectral centrality measures (Seeley’s index, PageRank and SALSA) are insen-
sitive to size, whereas the remaining ones are sensitive only to the increase of k
(or p in the case of betweenness). All nonattenuated spectral measures are also

2 0 It is interesting to note that it is actually the only centrality satisfying the size axiom—in
fact, one needs a cycle of ≈ ek nodes to beat a k-clique.
Boldi and Vigna: Axioms for Centrality 253

Centrality Size Density Score-Monotonicity

Degree only k yes yes

Harmonic yes yes yes
Closeness no no no
Lin only k no no
Betweenness only p no no
Dominant only k yes no
Seeley no yes no
Katz only k yes yes
PageRank no yes yes
HITS only k yes no
SALSA no yes no

Table 3. For each centrality and each axiom, we report whether it is satisfied.

non-monotone. Both Lin’s index and closeness centrality fail density tests.21
Closeness has, indeed, the worst possible behavior, failing to satisfy all our ax-
ioms. Although this result might seem counterintuitive, it is actually a conse-
quence of the known tendency of very far nodes to dominate the score, hid-
ing the contribution of closer nodes whose presence is more correlated to local
density.
All centralities satisfying the density axiom have no watershed: the axiom is
satisfied for all p, k ≥ 3. The watershed for closeness (and Lin’s index) is k ≤ p,
meaning that they just miss it, whereas the watershed for betweenness is a
quite pathological condition (k ≤ (p2 + p + 2)/4): one needs a clique whose size
is quadratic in the size of the cycle before the node of the clique on the bridge
becomes more important than the one on the cycle (compare this with closeness,
where k = p + 1 is sufficient).
We remark that our results on geometric indices do not change if we replace
the directed cycle with a symmetric (i.e., undirected) cycle, with the additional
condition that k > 3. It is possible that the same is true also of spectral central-
ities, but the geometry of the paths of the undirected cycle makes it extremely
difficult to carry on the analogous computations in that case.

2 1 We note that since D

k , p is strongly connected, closeness and Lin’s centrality differ just
by a multiplicative constant.
254 Internet Mathematics

7. Sanity Check via Information Retrieval

Information retrieval has developed in the last fifty years a large body of research
about extracting knowledge from data. In this section, we want to leverage the
work done in that field to check that our axioms actually describe interesting
features of centrality measures. We are, in this sense, following the same line of
thought as in [Najork et al. 07a]: in that study, the authors tried to establish in a
methodologically sound way which of degree, HITS, and PageRank works better
as a feature in web retrieval. Here we ask the same question, but we include for
the first time geometric indices also, which had never been considered before in
the literature about information retrieval, most likely because it was not possible
to compute them efficiently on large networks.22
The community working on information retrieval developed a number of stan-
dard datasets with associated queries and ground truth about which documents
are relevant for every query; those collections are typically used to compare the
(de)merits of new retrieval methods. Because many of those collections are made
of hyperlinked documents, it is possible to use them to assess centrality measures,
too.
In this work we consider the somewhat classical TREC GOV2 collection (about
25 million web documents) and the 149 associated queries. For each query (topic,
in TREC parlance), we have solved the corresponding Boolean conjunction of
terms, obtaining a subset of matching webpages. Each subset induces a graph
(whose nodes are the pages satisfying the conjunctive query), which can then be
ranked using any centrality measure. Finally, the pages in the graph are listed
in score order as results of the query, and standard relevance measures can be
applied to see how much they correspond to the available ground truth about
the assessed relevance of pages to queries.
There are a few methodological remarks that are necessary before discussing
the results:
r The results we present are for GOV2; there are other publicly available
collections with queries and relevant documents that can be used to this
purpose.
r As observed in earlier works [Najork et al. 07a], centrality scores in isolation
have a very poor performance when compared with text-based ranking func-
tions but can improve the results of the latter. We purposely avoid measur-
ing performance in conjunction with text-based ranking because this would

2 2 It is actually now possible to approximate them efficiently [Boldi and Vigna 13].
Boldi and Vigna: Axioms for Centrality 255

introduce further parameters. Moreover, our idea is to use information-

retrieval techniques to judge centrality measures, not to improve retrieval
performance per se (albeit, of course, a better centrality measure could be
used to improve the relevance of retrieved documents).
r Some methods are claimed to work better if nepotistic links (that is, links
between pages of the same host) are excluded from the graph. Therefore,
we report also results on the procedure applied to GOV2 with inter-host
links only.
r There are several ways to build a graph associated with a query. Here, we
choose a very straightforward approach—we solve the query in conjunctive
form and build the induced subgraph. Variants may include enlarging the
resulting graph with successors/predecessors, possibly by sampling [Najork
et al. 09].
r There are many measures of effectiveness that are used in information re-
trieval; among those, we focus here on the Precision at 10 (P@10, i.e.,
fraction of relevant documents retrieved among the first ten) and on the
NDCG@10 [Järvelin and Kekäläinen 02].
r Because of the poor performance, even for the best documents, about half
of the queries have null score. Thus, the data we report must be taken
with a grain of salt—confidence intervals for our measures of effectiveness
would be largely overlapping (i.e., our experiments have limited statistical
significance).

Our results are presented in Table 5: even if obtained in a completely different

way, they confirm the information we have been gathering with our axioms.
Harmonic centrality has the best overall scores. When we eliminate nepotistic
links, the landscape changes drastically—SALSA and PageRank now lead the
results—but the best performances are worse than those obtained using the
whole structure of the web. Note that, again, consistently with the information
gathered up to now, closeness performs very badly and betweenness performs
essentially like using no ranking at all (i.e., showing the documents in some
arbitrary order).
There are four new centrality measures appearing in Table 5 that deserve an
explanation. When we first computed these tables, we were very puzzled: HITS is
supposed to work very badly on disconnected graphs (it fails score monotonicity),
whereas it provided the second best ranking after harmonic centrality. Also, when
one eliminates nepotistic links, the graphs become highly disconnected and all
rankings tend to correlate with one another simply because most nodes obtain
256 Internet Mathematics

Indegree Negative β-measure

Number of coreachable nodes Indegree← β-measure←

Number of weakly reachable nodes Indegree↔ β-measure↔

Table 4. The names and definition of the four naive centrality measures used in
Table 5. Each centrality is obtained by multiplying the values described by its
row and column labels.

a null score. How is it possible that PageRank and SALSA work so well (albeit
less than harmonic centrality on the whole graph) with so little information?
Our suspect was that these measures were actually picking up some much
more elementary signal than their definition could make one think. In a highly
disconnected graph, the values assigned by such measures depend mainly on the
indegree and on some additional ranking provided by coreachable (or weakly
reachable) nodes.
We thus devised four “naive” centrality measures around our axioms. These
measures depend on density, and on size. We use two kinds of scores based on
density: the indegree and the negative β-measure [van den Brink and Gilles 94],23
that is,
 1
.
y →x
d+ (y)

The negative β-measure is a kind of “Markovian indegree” inspired by the 1 nor-

malization typical of Seeley’s index, PageRank, and SALSA. Indeed, indegree is
expressible in matrix form as 1A, whereas the negative β-measure is expressible
as 1Ā. We also use two kinds of scores based on size: the number of coreach-
able nodes, and the number of weakly reachable nodes. By multiplying a score
based on density with a score based on size, we obtain four centrality measures,
displayed in Table 4, which satisfy all our axioms.
As it is evident from Table 5, such simple measures outperform in this test
most of the very sophisticated alternatives proposed in the literature: this shows,
on one hand, that it is possible to extract information from the graph underly-
ing a query in very simple ways that do not involve any spectral or geometric
technique and, on the other hand, that designing centralities around our axioms

2 3 Note that the β-measure originally defined by van den Brink and Gilles in [van den Brink
and Gilles 94] is the positive version, that is, the negative β-measure can be obtained by
applying the β-measure defined in [van den Brink and Gilles 94] to the transposed graph.
Boldi and Vigna: Axioms for Centrality 257

All links Inter-host links only

NDCG@10 P@10 NDCG@10 P@10
BM25 0.5842 0.5644 BM25 0.5842 0.5644
Harmonic 0.1438 0.1416 β-measure↔ 0.1417 0.1349
Indegree← 0.1373 0.1356 SALSA 0.1384 0.1282
HITS 0.1364 0.1349 PageRank 1/4 0.1347 0.1295
Indegree↔ 0.1357 0.1349 β-measure← 0.1328 0.1275
Lin 0.1307 0.1289 Indegree↔ 0.1318 0.1255
β-measure↔ 0.1302 0.1322 PageRank 1/2 0.1315 0.1268
β-measure← 0.1275 0.1248 PageRank 3/4 0.1313 0.1255
Katz 3/4λ 0.1228 0.1242 Katz 1/2λ 0.1297 0.1262
Katz 1/2λ 0.1222 0.1228 Indegree← 0.1295 0.1262
Indegree 0.1222 0.1208 Harmonic 0.1293 0.1262
Katz 1/4λ 0.1204 0.1181 Katz 1/4λ 0.1289 0.1255
SALSA 0.1194 0.1221 Lin 0.1286 0.1248
Closeness 0.1093 0.1114 Indegree 0.1283 0.1248
PageRank 1/2 0.1091 0.1094 Katz 3/4λ 0.1278 0.1242
PageRank 1/4 0.1085 0.1107 HITS 0.1179 0.1107
Dominant 0.1061 0.1027 Closeness 0.1168 0.1121
PageRank 3/4 0.1060 0.1094 Dominant 0.1131 0.1067
Betweenness 0.0595 0.0584 Betweenness 0.0588 0.0577
— 0.0588 0.0577 — 0.0588 0.0577

Table 5. Normalized discounted cumulative gain (NDCG) and precision at 10

retrieved documents (P@10) for the GOV2 collection using all links and using
only inter-host links. The tables include, for reference, the results obtained using a
state-of-the-art text ranking function, BM25, and a final line obtained by applying
no ranking function at all (documents are sorted by the document identifier).

actually pays off. We consider this fact as a further confirmation that the traits
of centrality represented by our axioms are important.

8. Conclusions and Future Work

We have presented a set of axioms that try to capture part of the intended be-
havior of centrality measures. We have proved or disproved all our axioms for
ten classical centrality measures and for harmonic centrality, a variant to Bave-
las’s closeness that we have introduced formally in this article. The results are
surprising and confirmed by some information-retrieval experiments: harmonic
centrality is a very simple measure providing a good notion of centrality. It is
258 Internet Mathematics

almost identical to closeness centrality on undirected, connected networks but

provides a sensible centrality notion for arbitrary directed graphs.
There is, of course, a large measure of arbitrariness in the formulation of
our axioms: we believe that this is actually a feature—building an ecosystem of
interesting axioms is just a healthy way of understanding centrality better and
less anecdotally. Promoting the growth of such an ecosystem is one of the goals
of this work.
As a final note, the experiments on information retrieval that we have re-
ported are just the beginning. Testing with different collections (and possibly
with different ways of generating the graph associated with a query) may lead to
different results. Nonetheless, we believe we have made the important point that
geometric measures are relevant not only to social networks, but also to informa-
tion retrieval. In the literature comparing exogenous (i.e., link-based) rankings,
one can find different instances of spectral measures and indegree, but up to now
the venerable measures based on distances have been neglected. We suggest that
it is time to change this attitude.

Acknowledgments. We thank David Gleich for useful pointers leading to the proof of the
score-monotonicity of PageRank in the general case, and Edith Cohen for useful discus-
sions on the behavior of centrality indices. Marco Rosa participated in the first phases
of the development of this paper.

Funding. The authors have been supported by the EU-FET grant NADINE (GA 288956).

References
[Anthonisse 71] Jac M. Anthonisse. “The Rush in a Graph.” Technical Report, Ams-
terdam: University of Amsterdam Mathematical Centre, 1971.
[Altman and Tennenholtz 05] Alon Altman and Moshe Tennenholtz. “Ranking Sys-
tems: The PageRank Axioms.” In Proceedings of the 6th ACM Conference on Elec-
tronic Commerce, pp. 1–8. New York, NY: ACM, 2005.
[Avrachenkov and Litvak 06] Konstantin Avrachenkov and Nelly Litvak. “The Effect
of New Links on Google PageRank.” Stochastic Models 22:2 (2006), 319–331.
[Backstrom et al. 12] Lars Backstrom, Paolo Boldi, Marco Rosa, Johan Ugander, and
Sebastiano Vigna. “Four Degrees of separation.” In ACM Web Science 2012: Con-
ference Proceedings, pp. 45–54. ACM Press, 2012.
[Bavelas 50] Alex Bavelas. “Communication Patterns in Task-Oriented Groups.” Jour-
nal of Acoustical Socciety of America 22:6 (1950), 725–730.
[Bavelas and Barrett 51] A. Bavelas, D. Barrett, and American Management Associ-
ation. An Experimental Approach to Organizational Communication. Publications
Boldi and Vigna: Axioms for Centrality 259

(Massachusetts Institute of Technology. Dept. of Economics and Social Science):

Industrial Relations. American Management Association, 1951.
[Beauchamp 65] Murray A. Beauchamp. “An Improved Index of Centrality.” Behavioral
Science 10:2 (1965), 161–163.
[Berge 58] Claude Berge. Théorie des Graphes et ses Applications. Paris, France:Dunod,
1958.
[Berman and Plemmons 94] Abraham Berman and Robert J. Plemmons. Nonnegative
Matrices in the Mathematical Sciences. Classics in Applied Mathematics. SIAM,
1994.
[Boldi et al.13] Paolo Boldi, Marco Rosa, and Sebastiano Vigna. “Robustness of Social
and Web Graphs to Node Removal.” Social Network Analysis and Mining 3:4 (2013),
829–842.
[Boldi et al. 05] Paolo Boldi, Massimo Santini, and Sebastiano Vigna. “PageRank as
a Function of the Damping factor.” In Proc. of the Fourteenth International World
Wide Web Conference (WWW 2005), pp. 557–566. Chiba, Japan, 2005. ACM Press,
2005.
[Boldi et al. 09] Paolo Boldi, Massimo Santini, and Sebastiano Vigna. “PageRank:
Functional Dependencies.” ACM Trans. Inf. Sys. 27:4 (2009), 1–23.
[Boldi and Vigna 13] Paolo Boldi and Sebastiano Vigna. “In-Core Computation of
Geometric Centralities with HyperBall: A Hundred Billion Nodes and Beyond.”
In Proc. of 2013 IEEE 13th International Conference on Data Mining Workshops
(ICDMW 2013). IEEE, 2013.
[Bonacich 72] Phillip Bonacich. “Factoring and Weighting Approaches to Status
Scores and Clique Identification.” Journal of Mathematical Sociology 2:1 (1972),
113–120.
[Bonacich 91] Phillip Bonacich. “Simultaneous Group and Individual Centralities.” So-
cial Networks 13:2 (1991), 155–168.
[Borgatti 05] Stephen P. Borgatti. “Centrality and Network Flow.” Social Networks
27:1 (2005), 55–71.
[Borodin et al. 05] Allan Borodin, Gareth O. Roberts, Jeffrey S. Rosenthal, and Panayi-
otis Tsaparas. “Link analysis Ranking: Algorithms, Theory, and Experiments.” ACM
Transactions on Internet Technology (TOIT) 5:1 (2005), 231–297.
[Brandes et al. 12] Ulrik Brandes, Sven Kosub, and Bobo Nick. “Was messen Zen-
tralitätsindizes?” In Die Integration von Theorie und Methode in der Netzwerk-
forschung, edited by Marina Hennig and Christian Stegbauer, pp. 33–52. VS Verlag
für Sozialwissenschaften, 2012.
[Brauer 52] Alfred Brauer. “Limits for the Characteristic Roots of a Matrix. IV: Ap-
plications to Stochastic Matrices.” Duke Math. J. 19:1 (1952), 75–91.
[Brin and Page 98] Sergey Brin and Lawrence Page. “The Anatomy of a Large-Scale
Hypertextual Web Search Engine.” Computer Networks and ISDN Systems 30:1
(1998), 107–117.
[Burgess 69] Robert L. Burgess. “Communication Networks and Behavioral Conse-
quences.” Human Relations 22:2 (1969), 137–159.
260 Internet Mathematics

[Chien et al. 04] Steve Chien, Cynthia Dwork, Ravi Kumar, Daniel R. Simon, and
D. Sivakumar. “Link Evolution: Analysis and Algorithms.” Internet Math. 1:3 (2004),
277–304.
[Cohen and Kaplan 07] Edith Cohen and Haim Kaplan. “Spatially-Decaying Aggre-
gation over a Network.” Journal of Computer and System Sciences 73:3 (2007),
265–288.
[Cohen and Havlin 10] Reuven Cohen and Shlomo Havlin. Complex Networks: Struc-
ture, Robustness and Function. New York, NY: Cambridge University Press, 2010.
[Cohn and Marriott 58] B.S. Cohn and M. Marriott. “Networks and Centres of Inte-
gration in Indian Civilization.” Journal of Social Research 1:1 (1958), 1–9.
[Craswell et al. 03] Nick Craswell, David Hawking, and Trystan Upstill. “Predicting
Fame and Fortune: PageRank or indegree?” In Proceedings of the Australasian Doc-
ument Computing Symposium, ADCS2003, pp. 31–40, 2003.
[Del Corso et al. 06] Gianna Del Corso, Antonio Gullı̀, and Francesco Romani. “Fast
PageRank Computation via a Sparse Linear System.” Internet Math. 2:3 (2006),
251–273.
[Farahat et al. 06] Ayman Farahat, Thomas Lofaro, Joel C. Miller, Gregory Rae, and
Lesley A. Ward. “Authority Rankings from HITS, PageRank, and SALSA: Existence,
Uniqueness, and Effect of Initialization.” SIAM Journal on Scientific Computing 27:4
(2006), 1181–1201.
[Freeman 77] Linton C. Freeman. “A Set of Measures of Centrality Based on Between-
ness.” Sociometry 40:1 (1977), 35–41.
[Freeman 79] L. Freeman. “Centrality in Social Networks: Conceptual Clarification.”
Social Networks 1:3 (1979), 215–239.
[Friedkin 91] N.E. Friedkin. “Theoretical Foundations for Centrality Measures.” The
American Journal of Sociology 96:6 (1991), 1478–1504.
[Hubbell 65] Charles H. Hubbell. “An Input-Output Approach to Clique Identifica-
tion.” Sociometry 28:4 (1965), 377–399.
[Järvelin and Kekäläinen 02] Kalervo Järvelin and Jaana Kekäläinen. “Cumulated
Gain-Based Evaluation of IR Techniques.” ACM Trans. Inf. Syst. 20:4 (2002),
422–446.
[Katz 53] Leo Katz. “A New Status Index Derived from Sociometric Analysis.” Psy-
chometrika 18:1 (1953), 39–43.
[Kendall 55] Maurice G. Kendall. “Further Contributions to the Theory of Paired Com-
parisons.” Biometrics 11:1 (1955), 43–62.
[Kleinberg 99] Jon M. Kleinberg. “Authoritative Sources in a Hyperlinked Environ-
ment.” Journal of the ACM 46: (1999), 604–632.
[Knuth 92] Donald E. Knuth. “Two Notes on Notation.” American Mathematical
Monthly 99:5 (1992), 403–422.
[Leavitt 51] H. J. Leavitt. “Some Effects of Certain Communication Patterns on Group
Performance.” J. Abnorm. Psychol. 46:1 (1951), 38–50.
Boldi and Vigna: Axioms for Centrality 261

[Lempel and Moran 01] Ronny Lempel and Shlomo Moran. “SALSA: The Stochastic
Approach for Link-Structure Analysis.” ACM Trans. Inf. Syst. 19:2 (2001), 131–160.
[Lempel and Moran 05] Ronny Lempel and Shlomo Moran. “Rank-Stability and Rank-
Similarity of Link-Based Web Ranking Algorithms in Authority-Connected Graphs.”
Information Retrieval 8:2 (2005), 245–264.
[Li et al. 05] Lun Li, David L. Alderson, John Doyle, and Walter Willinger. “Towards
a Theory of Scale-Free Graphs: Definition, Properties, and Implications.” Internet
Math. 2:4 (2005), 431–523.
[Lin 76] Nan Lin. Foundations of Social Research. McGraw-Hill, New York, 1976.
[Mackenzie 66] Kenneth Mackenzie. “Structural Centrality in Communications Net-
works.” Psychometrika 31:1 (1966), 17–25.
[Marchiori and Latora 00] Massimo Marchiori and Vito Latora. “Harmony in the
Small-World.” Physica A: Statistical Mechanics and its Applications 285:3–4 (2000),
539–546.
[McDonald et al. 95] J.J. McDonald, M. Neumann, H. Schneider, and M.J. Tsat-
someros. “Inverse m-Matrix Inequalities and Generalized Ultrametric Matrices.”
Linear Algebra and its Applications 220 (1995), 321–341.
[Milgram 67] Stanley Milgram. “The Small World Problem.” Psychology Today 2:1
(1967), 60–67.
[Najork et al. 09] Marc Najork, Sreenivas Gollapudi, and Rina Panigrahy. “Less is
More: Sampling the Neighborhood Graph Makes SALSA Better and Faster.” In
Proceedings of the Second ACM International Conference on Web Search and Data
Mining, pp. 242–251. New York, NY: ACM, 2009.
[Najork et al. 07a] Marc Najork, Hugo Zaragoza, and Michael J. Taylor. “HITS on the
Web: How Does it Compare?” In Wessel Kraaij, Arjen P. de Vries, Charles L. A.
Clarke, Norbert Fuhr, and Noriko Kando, pp. 471–478. SIGIR 2007: Proceedings of
the 30th Annual International ACM SIGIR Conference on Research and Develop-
ment in Information Retrieval. Amsterdam, The Netherlands: New York, NY: ACM,
2007.
[Najork et al. 07b] Marc A. Najork, Hugo Zaragoza, and Michael J. Taylor. HITS on
the web: How does it compare? In Proceedings of the 30th annual international ACM
SIGIR conference on Research and development in information retrieval, SIGIR ’07,
pages 471–478. ACM, 2007.
[Nieminen 73] U.J. Nieminen. “On the Centrality in a Directed Graph.” Social Science
Research 2:4 (1973), 371–378.
[Padgett and Ansell 93] John F. Padgett and Christopher K. Ansell. “Robust Action
and the Rise of the Medici, 1400–1434”. The American Journal of Sociology 98:6
(1993), 1259–1319.
[Page et al. 98] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd.
“The PageRank Citation Ranking: Bringing Order to the Web.” Technical Report
SIDL-WP-1999-0120, Stanford Digital Library Technologies Project, Stanford Uni-
versity, 1998.
262 Internet Mathematics

[Pan and Saramäki 11] Raj Kumar Pan and Jari Saramäki. “Path Lengths, Correla-
tions, and Centrality in Temporal Networks.” Phys. Rev. E 84:1(2011),016105.
[Pitts 65] Forrest R. Pitts. “A Graph Theoretic Approach to Historical Geography.”
The Professional Geographer 17:5 (1965), 15–20.
[Sabidussi 66] G. Sabidussi. “The Centrality Index of a Graph.” Psychometrika 31:4
(1966), 581–603.
[Seeley 49] John R. Seeley. “The Net of Reciprocal Influence: A Problem in Treating
Sociometric Data.” Canadian Journal of Psychology 3:4 (1949), 234–240.
[Shannon 148] Claude E. Shannon. “A Mathematical Theory of Communication.” Bell
Syst. Tech. J. 27:3 (1948), 379–423, 623–656.
[Stephenson and Zelen 89] Karen Stephenson and Marvin Zelen. “Rethinking Central-
ity: Methods and Examples.” Social Networks 11:1 (1989), 1–37.
[Upstill et al. 03] Trystan Upstill, Nick Craswell, and David Hawking. “Query-
Independent Evidence in Home Page Finding.” ACM Trans. Inf. Syst. 21:3 (2003),
286–313.
[van den Brink and Gilles 94] René van den Brink and Robert P. Gilles. “A Social
Power Index for Hierarchically Structured Populations of Economic Agents.” In Im-
perfections and Behavior in Economic Organizations, vol. 11 edited by Robert P.
Gilles and Pieter H. M. Ruys, pp. 279–318. Netherlands: Springer, 1994.
[Vigna 09] Sebastiano Vigna. “Spectral Ranking.” Preprint, 2009. (http://arxiv.org/
abs/0912.0238).
[Wasserman and Faust 94] Stanley Wasserman and Katherine Faust. Social Network
Analysis: Methods and Applications. New York, Cambridge: Cambridge Univ Press,
1994.
[Wei 52] T.-H. Wei. The Algebraic Foundations of Ranking Theory. PhD thesis, Uni-
versity of Cambridge, 1952.

Paolo Boldi, Dipartimento di informatica, Università degli Studi di Milano, via

Comelico 39, 20135 Milano, Italy ([email protected])

Sebastiano Vigna, Dipartimento di informatica, Università degli Studi di Milano, via

Comelico 39, 20135 Milano, Italy ([email protected])