Games and Economic Behavior 79 (2013) 129

Contents lists available at SciVerse ScienceDirect

Games and Economic Behavior
www.elsevier.com/locate/geb
Dynamics in tree formation games
E. Arcaute
a,1
, K. Dyagilev
b
, R. Johari
c
, S. Mannor
b,
a
Adchemy, Inc., 1001 E. Hillsdale Blvd., Foster City, CA 94404, USA
b
Faculty of Electrical Engineering, Technion, Haifa 32000, Israel
c
Department of Management Science and Engineering, Stanford University, Stanford, CA 94305-4121, USA
a r t i c l e i n f o a b s t r a c t
Article history:
Received 20 March 2012
Available online 11 January 2013
JEL classication:
D85
Keywords:
Network formation games
Network dynamics
Strong stability
Tree formation games
Network formation games capture two conicting objectives of selsh nodes in a network:
such nodes wish to form a well-connected network and, at the same time, to minimize
their cost of participation. We consider three families of such models where nodes avoid
forming edges beyond those necessary for connectivity, thus forming tree networks. We
focus on two local two-stage best-response dynamics in these models, where nodes can
only form links with others in a restricted neighborhood. Despite this locality, both our
dynamics converge to ecient outcomes in two of the considered families of models.
In the third family of models, both our dynamics guarantee at most constant eciency
loss. This is in contrast with the standard best-response dynamics whose eciency loss is
unbounded in all three families of models. Thus we present a globally constrained network
formation game where local dynamics naturally select desirable outcomes.
2013 Elsevier Inc. All rights reserved.
1. Introduction
Network formation games (Jackson, 2008) typically embody a tension between two countervailing forces. On one hand,
individuals benet from connections to each other because of the network reachability such connections afford; for example,
in a transportation network, additional links allow a node to shorten the distance to other nodes in the network. On the
other hand, link creation and maintenance is typically costly. Thus individuals are incentivized to balance the number of
links with the benets they provide.
In this paper, we study this particular tension in the context of a network formation game model inspired by data com-
munication and transportation networks. Our cost structure has two key features. First, in these applications, connectivity is
typically a minimal requirement. To account for the desire for connectivity, we assume that nodes experience a very large
cost when the network is disconnected. Second, we assume that in a connected network, nodes experience a cost consisting
of two terms: (1) a link maintenance cost, 0, per link the node participates in; and (2) a trac cost, per unit of trac
sent by the node. We assume that nodes have an ex ante demand to send trac to each other and form links with each
other with the goal of minimizing the total cost of both link maintenance and trac handling.
In our model the entire cost of sending trac is borne by the sender of the trac itself. This is the case, for example, in
logistics networks, where transportation costs are borne by the sender. It is also the case in certain types of data transport
contracts for Internet service. In our paper we consider several forms for the trac cost; special cases of this cost recover
*
Corresponding author.
E-mail addresses: arcaute@alumni.stanford.edu (E. Arcaute), kirilld@tx.technion.ac.il (K. Dyagilev), ramesh.johari@stanford.edu (R. Johari),
shie@ee.technion.ac.il (S. Mannor).
1
This research was conducted when the author was at the Institute for Computational and Mathematical Engineering at Stanford University, Stanford,
CA, USA.
0899-8256/$ see front matter 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.geb.2013.01.002
2 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
both the model of Corbo and Parkes (2005) and the connections model of Jackson and Wolinsky (1996). (Indeed, the model
where the sender bears the cost of trac is one special case of an allocation rule, as dened by Jackson, 2008.)
Our model has two orthogonal dimensions that can be explored. We can x the value of the link maintenance cost and
study the static game and dynamics for a range of suitable trac-related cost functions. Alternatively, we can x the trac-
related cost function and vary the link maintenance cost . We select the former approach and x to be large enough so
that the cost of redundant links is prohibitive. This large value of together with the requirement of connectivity of the
network implies that the set of pairwise Nash networks is a set of trees.
Our work is primarily concerned with dynamics: in particular, are there plausible dynamics that lead to ecient or near-
ecient equilibria? We begin by analyzing a natural starting point, myopic best-response dynamics. First, we characterize
equilibrium outcomes of the static game and show that even the strongly stable outcomes may be inecient. In fact, the
corresponding price of anarchy can grow linearly in the size of the network. Second, we observe that any static equilibria
is a xed point of best-response dynamics. Hence, even with respect to strong stability, the best-response dynamics may
select a grossly inecient network.
This ineciency problem is addressed by considering two particular dynamic processes inspired by our earlier work
(Arcaute et al., 2007, 2008, 2009), that we call NodeDynamics and LinkDynamics. In NodeDynamics, at the beginning of
a round, an active node u is exogenously selected by a stochastic process. Node u is allowed to minimize the cost it
experiences at the end of the round by undertaking either a unilateral deviation or a bilateral deviation with any node w
in the active nodes vicinity. In case of unilateral deviation, the active node is allowed to erase several of its adjacent edges.
A bilateral deviation yields erasure of several links adjacent to either u or w and creation of a new link between these
two nodes. This deviation proceeds in two stages. At the rst stage, the active node u erases some of its adjacent links and
suggests node w to create a link between them. At the second stage, node w erases some of its adjacent links and decides
whether to accept the link proposal from node u. We assume that link proposal is accepted if the cost experienced at the
end of the round by node w is smaller than the one at the end of the rst stage.
In contrast, in LinkDynamics, at the beginning of a round, a pair of nodes is exogenously selected by a stochastic process
and one of the nodes is designated to be an active node. We let u denote the active node and let w denote the other node.
We assume that the dynamics are local, i.e., the selected nodes u and w are necessarily close to each other in the current
network. Node u is allowed to minimize the cost it experiences at the end of the round by undertaking either a unilateral
deviation or a bilateral deviation with the selected node w. Unilateral and bilateral deviation are performed in the identical
manner to these in NodeDynamics. Essentially, the activation process in LinkDynamics selects a potential link (a pair of
nodes) to activate as opposed to NodeDynamics in which a single node is activated.
We analyze NodeDynamics and LinkDynamics for three specic subclasses of the distance-based trac cost models. We
show that our dynamics converge almost surely in all three subclasses. We then show that in two subclasses of models,
these dynamics converge to the most ecient network. In the third subclass of models, the ratio of eciency of the selected
outcome to the most ecient network is shown to be bounded by a constant that is independent of the size of the network.
We show that the expected time until convergence of our dynamics is polynomial for two subclasses of models and, in the
third subclass, it is at most exponential. Any xed point of NodeDynamics and LinkDynamics is necessarily a 2-stable
equilibrium, therefore it is fair to compare the eciency of networks selected by these dynamics to the eciency of the
standard best-response dynamics with respect to 2-stability. Recall that the price of anarchy associated with strong stability
is of O(|V |). This implies that the price of anarchy of the standard best response dynamics with respect to 2-stability is at
least of O(|V |). In all three subclasses, it can be seen that the outcomes selected by our dynamics are signicantly more
ecient than the worst case 2-stable equilibrium.
Our analysis of NodeDynamics and LinkDynamics relies on an auxiliary tree formation game. This game does not have a
value of its own, however, it helps us to streamline the analysis of NodeDynamics and LinkDynamics in the original game.
Tree formation game is a version of Myerson link announcement game where:
all nodes incur an unbounded cost in any outcome other than a tree;
the link maintenance cost is removed and the trac-related cost retained; and
any deviation involving one or two nodes is allowed.
We dene two constrained versions of best-response dynamics of the tree formation game, dubbed TreeNode and TreeLink.
We show that TreeNode is equivalent to NodeDynamics in the original game and that TreeLink is equivalent to LinkDy-
namics. We derive convergence results for TreeNode and TreeLink, as they are easier to analyze. Due to the equivalence of
the dynamics, these results automatically apply to NodeDynamics and LinkDynamics.
The rest of the paper is organized as follows. We begin with a survey of relevant literature in Section 2. Section 3 presents
the formal denition of our link announcement game in its most general form. This section also introduces the correspond-
ing notions of equilibrium and eciency and the results of our analysis of the static game. In Section 4 we formally dene
NodeDynamics and LinkDynamics and analyze some of their basic properties. In Section 5 we introduce distance-based cost
models, and we constrain our analysis to these cost models for the remainder of the paper. In Section 6 we introduce the
tree formation game and the two versions of the associated best-response dynamics TreeNode and TreeLink. We show the
equivalence of the latter to NodeDynamics and LinkDynamics, respectively. In Sections 7 and 8 we address three specic
subclasses of the distance-based cost model. For each class, we rene our characterization of stable outcomes of the link
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 3
announcement game and analyze convergence properties of both our dynamics. We conclude this paper with a summary of
central results and suggestions for future research in Section 9.
2. Related work
Our work touches on several related threads of the literature. Most closely related is the work on network formation
games in economics, see Jackson (2003), Goyal (2007), Vega-Redondo (2007) for a survey. In the context of communication
networks, Johari et al. (2006) considered a static network formation game related to the model in this paper; by contrast,
the focus of the current paper is on the dynamics of such formation processes. The work of Jackson and Watts also considers
dynamics for network formation games (Jackson and Watts, 2001, 2002), but for a utility model that is unrelated to ours;
while in their dynamics any unilateral or bilateral deviation may occur in a single stage, our dynamics are designed so that
bilateral deviation proceeds in two stages. This latter property allows our dynamics to select desirable equilibria. Despite
the dissimilarities in cost structure and dynamics, random activation of nodes in the dynamics is needed for obtaining
convergence results both in our setting and in that of Jackson and Watts (2002).
In Arcaute et al. (2008), we considered dynamics in which each step is composed of two stages. During the rst stage,
the active node is allowed to unilaterally deviate by erasing some of the adjacent edges. During the second stage, the active
node is allowed to undertake either a unilateral deviation or a bilateral deviation with a node in its close vicinity. It can
be seen that these dynamics are essentially identical to NodeDynamics. However, Arcaute et al. (2008) considered a utility
model that is unrelated to ours.
In contrast to our approach, which is inspired by the literature on network formation games, in recent years a large
body of literature has considered eciency loss due to self-interested behavior in routing-related applications; see, e.g.,
Roughgarden (2005). A related thread of literature has considered the eciency of equilibria in network design games; see,
e.g., Chen et al. (2008), Anshelevich et al. (2003), Fiat et al. (2006). In this line of work, a network is formed based on
unilateral decisions of the players, and the cost of the resulting network is shared among the players. The goal of that body
of research is to design mechanisms that provide agents with incentives to use the network in an ecient manner.
Our work is also related to the literature on learning in games; see, e.g., Fudenberg and Levine (1998), Cesa-Bianchi
and Lugosi (2004), Mannor and Shamma (2007) for surveys. In this literature the emphasis is on studying classes of dy-
namic methods that converge into the set of equilibria (e.g., correlated or Nash equilibria), without regard to eciency. Our
approach departs signicantly from this literature, as we are interested in convergence to desirable equilibria.
Finally, there is an extensive body of research in the application of game theory to networks; see, e.g., Altman et al.
(2006) for a survey, and Falkner et al. (2000), Briscoe et al. (2003) for a discussion of pricing in networks. In the application
domain, our work is related to papers on topology formation in ad hoc networks; e.g., Komali and MacKenzie (2006),
Eidenbenz et al. (2003), Fabrikant et al. (2003). However, these works all consider Nash equilibria, whereas our focus is on
pairwise interactions between nodes.
3. Preliminaries
In this section we introduce the general game theoretic model we study throughout the paper. We begin by presenting
in Section 3.1 some of the notations used in this paper, which we adopt from Arcaute et al. (2007, 2008), Jackson and
Wolinsky (1996). In Section 3.2 we describe a general version of the cost model we employ. In this model, nodes face a
cost both to maintain connectivity to other nodes, as well as a cost based on the quality of connectivity to other nodes.
We further make an assumption that will prove to be central in our analysis: the link maintenance cost is suciently high
to deter nodes from keeping redundant links in the network. In Section 3.3 we describe the game that we study, which is
a variant of the Myerson link announcement game. Section 3.4 introduces a notion of eciency of a network. Finally, in
Section 3.5 we analyze networks that correspond to its equilibrium outcomes and discuss their eciency.
3.1. Notation
Let V be a set of n strategic agents alternatively referred to as nodes. We denote a network connecting these nodes by
an undirected graph G = (V , E), where V is the set of nodes and E is the set of edges (or links). We write uv for the
undirected edge between distinct nodes u and v. Abusing notation, we will typically write uv G if uv E.
Given a network G =(V , E) and two nodes u = v, we let G +uv be the network over V with edge set E {uv}. Similarly,
we let G uv be the network over V with edge set E \ {uv}. For two networks G
1
= (V , E
1
) and G
2
= (V , E
2
) over the
same set V of nodes, we denote by G
1
\ G
2
the set of all edges uv G
1
such that uv / G
2
. Finally, we let (u; G) denote
the number of links adjacent to node u in G (the degree of node u) and let d(u, v; G) denote the shortest path distance (in
number of edges) between nodes u and v in G. We let d(u, v; G) = if u and v are not connected in G.
3.2. Cost model
In modern communication and transportation networks, it is expected that any pair of users should be able to commu-
nicate. We assume that any node in a disconnected network is penalized by a suciently large cost so we could focus on
4 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
connected networks alone. In a connected network, nodes experience a cost consisting of two terms: (1) a link maintenance
cost incurred per link the node participates in; and (2) a connectivity cost, that quanties the quality of connectivity a node
obtains in a given network. Informally, the network formation game we consider involves a tradeoff: forming more links will
typically decrease the connectivity cost due to higher quality of connectivity, but it will also increase the link maintenance
cost.
Formally, let G be a connected network and x a node u. We assume that the node incurs a link maintenance cost
per link, and a connectivity cost h(u; G). Thus we dene the cost C(u; G) to node u in G as:
C(u; G) =(u; G) +h(u; G). (1)
If G is not connected, we assume the cost C(u; G) to any node u V to be an arbitrarily large constant. In particular, for
the remainder of the paper, we restrict attention to studying eciency and stability of connected networks.
We make a second assumption that plays a key role in our analysis. Denote by h
min
the minimum value taken by h(u; G)
over all possible nodes u and connected networks G, and by h
max
its maximum value. Then,
Assumption 1. The link maintenance cost satises > (1/2)|V |(h
max
h
min
), where |V | is the cardinality of the set V .
This assumption remains in force throughout the paper.
Informally, Assumption 1 ensures that there is no value to redundant links in the network: provided that the link uv
can be erased from network G without compromising its connectivity, both node u and v are incentivized to remove uv.
Moreover, it means that no improvement of connectivity can compensate for the creation of an additional link. These
observations are made formal in the following corollary:
Corollary 1 (Costliness of redundant links).
1. Let G and G

be two connected networks over the same set of nodes so that (u; G

) > (u; G). Then, C(u; G

) > C(u; G).

2. Let uv be a link in the network G so that G uv is connected. Then, C(u; G uv) < C(u; G) and C(v; G uv) < C(v; G).
Proof. Consider the difference in cost the node u incurs in networks G and G

:
C(u; G) C
_
u; G

_
=
_
(u; G)
_
u; G

__
+
_
h(u; G) h
_
u; G

__
< +(h
max
h
min
)
< (h
max
h
min
)
_
1 1/2 |V |
_
< 0.
Here, the rst equality follows by a straightforward substitution. The rst inequality stems from the denition of h
max
and h
min
, and the observation that (u; G

) (u; G) + 1. The third and forth inequalities hold true to Assumption 1 and
the observation that the set V contains at least two nodes (otherwise, (u; G

) =(u; G) =0). Statement 2 follows trivially

from Statement 1. 2
3.3. The game
We consider a network formation game introduced by Myerson (1977) and dubbed the link announcement game. In this
game, each node u simultaneously announces the set S
u
of nodes it wishes to connect to. Given the composite strategy
vector S = (S
u
: u V ), the graph G(S) is created as follows: for all pairs of nodes u and v, the link uv exists in G if
and only if u S
v
and v S
u
. The costs experienced by nodes in network G are described by function C(u; G) dened in
Section 3.2.
An important property of a game is its Nash equilibria strategies, namely, composite strategy vectors under which no
node can strictly prot by unilaterally changing its action. We refer to graphs G(S) that correspond to a Nash equilibrium
strategy S as a Nash equilibrium network.
Due to Assumption 1, we have the following simple observation: any connected Nash equilibrium network is a tree.
Lemma 2. If a Nash equilibrium network G is connected, then it is a tree.
Proof. Assume by contradiction that G is not a tree. Since G is connected, there must exist a link uv G so that G uv
is connected. Node u can decrease its degree (u; G) by unilaterally erasing the link uv. By Statement 2 of Corollary 1, this
deviation strictly prots node u, which is a contradiction to G being a Nash equilibrium network. 2
3.4. Eciency and the price of anarchy
We measure the eciency of a connected network G via the social cost (G), which is the total cost experienced by all
nodes in the network:
(G) =

u
C(u; G).
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 5
We say that a connected network G is ecient if and only if G argmin
G
(G

), where the minimum is over all possible

connected networks G

. The following corollary shows that all ecient networks are trees:
Corollary 3. Let G be an ecient network. Then G is a tree.
Proof. Assume by contradiction that G is not a tree. Then, there exists a link uv G so that G uv is connected. Consider
the difference of social costs of networks G and G uv:
(G uv) (G) =2 +

wV
_
C(w; G uv) C(w; G)
_
<2 +|V |(h
max
h
min
) < 0.
Here, the rst equality is obtained by straightforward substitution. The rst inequality follows from denition of h
max
and h
min
. The last inequality stems from Assumption 1. We thus obtained that (G uv) (G) < 0 which is in contra-
diction with the denition of G as an ecient network. 2
In this paper we provide insights into the eciency of outcomes achieved through a variety of stability concepts (both
static and dynamic), all of which rene the Nash equilibrium concept. We observe that in our model, both ecient networks
(see Corollary 3) and stable networks (see Lemma 2) are always trees. For any tree, the sum of link maintenance costs for
all nodes is 2(n1). For this reason, throughout the paper when comparing the eciency of two outcomes, we do so only
by comparing their total connectivity costs. Formally, given a tree G, we dene the social connectivity cost as:
(G) =

u
h(u; G). (2)
While occasionally stable outcomes will be fully ecient, it is also possible that there may be ineciencies in the
obtained networks. In such a case, we measure the relative ineciency via the price of anarchy (Papadimitriou, 2001), dened
in terms of . Formally, let T denote the set of all trees. For a given subset G T of trees, we dene:
PoA(G) =
max
G

G
(G

)
min
GT
(G)
as the eciency ratio of the least ecient network in G to that of an ecient tree. We refer to PoA(G) as the price of
anarchy of G.
3.5. Static analysis: k-stability
This section is devoted to static analysis of the link announcement game, via the concept of k-stability. Several solution
concepts have been considered when analyzing the link announcement game. Since bilateral agreement is necessary to add
a link, it is natural to consider a form of pairwise stability as our solution concept (Jackson, 2008; Ma, 1995, 2002). We will
more generally consider k-stability (k > 1) as our solution concept. Informally, we say that G is k-stable if no coalition of k
or fewer nodes can protably deviate from G.
Formally, let S be a non-empty subset of nodes. We say that the network G

is obtainable from the network G through

deviations by S if, for all nodes u = v:
1. uv G

\ G implies u, v S, and
2. uv G \ G

implies either u S or v S.
The rst condition states that any new link can only be made between players in S. The second condition states that all
removed links must have at least one endpoint in S. We refer the reader to Jackson (2008) for a more thorough exposition
of the above denition. We say that G

S-defeats G if the following conditions hold: (1) G

is obtainable from G through

deviations by S; and (2) for all u S, it holds that C(u; G) C(u; G

) with a strict inequality for at least one node in S. We

can now formally dene k-stability.
Denition 1 (k-Stability). Given k > 1, the network G is k-stable if for every subset S of nodes so that |S| k there is no
network G

that S-defeats G.
In the rest of this section we analyze k-stable networks in our link announcement game and their eciency. We begin
by showing that the set of 3-stable networks and the set of tree networks are identical. We further focus on tree networks
of two specic topologies. Their importance will become evident later in the section. The rst network topology we consider
is a star, which is a tree with all-but-one nodes connected to a single hub node. We show that all stars are strongly stable,
namely, are |V |-stable. The second network topology we consider is a path, where all nodes lie on a single path. We show
that the set of all path networks is a sink equilibria with respect to strong stability. Namely, the only network that can
6 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
V -defeat a path network is another path network. We refer the reader to Goemans et al. (2005) for the thorough discussion
of the concept of sink equilibrium. The latter result has strong implication on network dynamics. For instance, consider the
best-response dynamics in which at every time step a different coalition S of nodes is allowed to deviate from the current
strategy. No node in coalition S will agree to a deviation that increases its cost, hence they can only deviate to a network
that S-defeats the current network. Combining with the previous observation, we conclude that once the best-response
dynamics arrives to the set of path networks it will never leave it. We conclude this section with a specic example of
the connectivity cost function. We show that for this function both the most and the least ecient connected networks are
strongly stable. We thus conclude that strong stability may fail to provide an insight to network eciency.
Observations above are made formal by the following proposition:
Proposition 4.
1. The set of 3-stable networks is the set of all trees over V .
2. All stars
2
are strongly stable.
3. The set of all path networks
3
is a sink equilibrium under strong stability.
The proof of Proposition 4 builds on the following auxiliary lemma:
Lemma 5. Let G be a tree network and let G

be a network that S-defeats G for some subset of nodes S. Then,

1. Network G

is a tree.
2. The degrees of all nodes are identical in both networks, i.e., (u; G) =(u; G

) for all u V .
3. For any link uv such that uv G and uv / G

it holds that u, v S.
Proof. Statement 1. We note that network G

has to be connected otherwise it could not defeat a connected network G. Let

us assume by contradiction that G

is not a tree. Then, the total number of links in G

is larger than (|V | 1), hence

uV

_
u; G

_
> 2
_
|V | 1
_
=

uV

_
u; G

_
.
This implies that there exists a node v so that (v; G

) > (v; G), which in turn implies that there exists a link vw G \ G

.
By denition, network G

is obtainable from G through deviations of nodes in S. Furthermore, new links can only emerge
between nodes in S, we therefore conclude that v S. Statement 1 of Corollary 1 yields that C(v; G

) > C(v; G), namely,

there exists a node v S that pays a higher cost in G

than in G. This is in contradiction with denition of G

as a network
that S-defeats G.
Statement 2. We begin by noting that the sum of degrees of all nodes is identical in both networks:

vV
(v; G) =

vV

_
v; G

_
=2|V | 2. (3)
In what follows, we show that (u; G

) (u; G) for all u V . Combining with (3) it yields that (u; G) = (u; G

) for all
u V , which concludes the proof of Statement 2. Note that for all w / S, it holds that (w; G

) (w; G), as links cannot be

added to nodes not in S. It remains to show that (w; G

) (w; G) for all w S. Assume by contradiction that there exists

w S such that (w; G

) > (w; G). Statement 1 of Corollary 1 yields that C(v; G

) > C(v; G), which is in contradiction with

denition of G

as a network that S-defeats G.

Statement 3. By Statement 2 it holds that (u; G

) = (u; G), hence, there should be a new link uw G

\ G that com-
pensates for the erasure of the link uv. However, since G

is S obtainable from G, it implies that u S. The exact same

argument holds for v, hence v S. 2
We can now prove Proposition 4:
Proof of Proposition 4. Statement 1. It can be readily seen that any 3-stable network is also a Nash equilibrium network,
therefore, by Lemma 2 it is a tree. It remains to show that every tree G is a 3-stable network. Indeed, assume by con-
tradiction that there exists a network G

that S-defeats G for some subset S of at most 3 nodes. Since G = G

and both
networks contain the same number of edges, there exists a link uv G \ G

. By Statement 3 of Lemma 5 it holds that

u, v S. Furthermore, by Statement 2 of Lemma 5 it holds that (u; G

) =(u; G) and (v; G

) =(v; G), which implies that

nodes u and v each generated at least one new link during the deviation. Let ua G

\ G and ub G

\ G to denote these
links. By denition of G

, it holds that a, b S, hence, combining with conditions |S| 3 and u, v S, we conclude that
2
Stars are trees in which all but one node are leaves.
3
Path network is a tree in which exactly two nodes are leaves and all other nodes participate in exactly two edges.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 7
a and b are the same node. Again, by Statement 2 of Lemma 5 it holds that (a, G

) = (a, G). However, this means that

node a erased at least two links in its deviation from G to G

. Let us denote two of these links by ax and ay. We note that

x, y / {u, v, a} S. This is in contradiction with Statement 3 of Lemma 5.
Proof of Statement 2. Let G be a star network with center u V , i.e., (v; G) = 1 for all v = u. In other words, all nodes
but u are leaves. Assume by contradiction that there exists a network G

that V -defeats network in G. By Statements 1

and 2 of Lemma 5 it holds G

is a tree and that the degrees of all nodes in G

are identical to these G. However, the only

tree with all nodes but u V being leaves is the star centered at u. Thus, G

= G in contradiction to the denition of G

.
Proof of Statement 3. Let G be a path network. It is evident that for every node v V it holds that (v; G) 2. Assume
by contradiction that there exists a non-path network G

that V -defeats network G. Since network G

is a tree, but not a

path, there exists a node v V so that (v; G

) > 2. However, this is in contradiction with Statement 2 of Lemma 5 that

states that the degrees of all nodes in G

remain identical to these in G. 2

We conclude this section by considering a specic form of connectivity cost presented in the following example:
Example 1 (Distance-based connectivity cost). Assume that G is connected. The distance-based trac cost to an agent u in G is
dened by
h(u; G) =

v=u
d(u, v; G).
Such a cost might arise when modeling units of trac being sent through the network, with each unit of trac experiencing
a unit delay per node traversed. This cost model has received a signicant amount of attention in the recent literature
(Fabrikant et al., 2003; Corbo and Parkes, 2005; Demaine et al., 2007).
Under this cost, both star and path topologies are strongly stable, however, their social connectivity costs differs in a
signicant way (by a factor of (|V |)). Moreover, their social connectivity costs are the lowest (star) and the highest (path)
among all connected networks. We therefore conclude that strong stability does not provide any insight into the eciency
of networks we can expect to arise due to self-interested interaction of the nodes. This observation is made formal in the
following proposition:
Proposition 6.
1. Any star network G
S
is strongly stable under distance-based connectivity cost. Its social connectivity cost is the lowest among all
connected networks and is given by
(G
S
) =2
_
|V | 1
_
2
.
2. Any path network G
P
is strongly stable under distance-based connectivity cost. Its social connectivity cost is the highest among
all connected networks and is bounded from below in the following way
(G
P
)
_
|V | 1
_
3
/6.
Proof. The strong stability of G
S
and G
P
follows from Theorem 20 in Section 8.1. The rest of the results are evident. 2
4. Dynamics
As shown in Section 3.5, even strong stability can fail to provide denite conclusions on the eciency of stable net-
works of the link announcement game. In this section we consider an alternate approach: rather than characterizing stable
networks via a static notion, such as k-stability, we consider stable networks that might arise as limit points of dynamic
strategic adjustment processes.
A natural starting point might be myopic best-response dynamics, with respect to k-stabilityi.e., if G is the current
network, the next network in the dynamic process is a network G

that S-defeats G, for some coalition of at most k nodes.

Of course, the xed points of such a process are the k-stable networks, and so based on our results in the preceding section
we gain no additional insight into eciency by studying such dynamics.
Instead, we consider the following approach: are there dynamics that are almost best-response dynamics, that can
yield greater insight into eciency? In particular, we focus on two dynamics that are based on deviations consistent with
2-stability and consider the possibility of granting one of the nodes a right of rst move. At rst glance, this shift seems to
allow a single node to act more selshly, by creating an intermediate state that is favorable towards an outcome achieved
in two steps. In fact, we show in Corollary 8 that xed points of these two dynamics are necessarily 2-stable. We compare
the eciency of worst-case outcomes selected by our dynamics to the worst-case 2-stable network. Specically, we show
that given our cost structure, our dynamics choose more ecient networks.
8 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
In this paper we consider two dynamicsNodeDynamics and LinkDynamics.
NodeDynamics. Let > 1 be a given integer. Let k be the number of the current round and let G
(k)
be the network in the
beginning of round k. We assume that G
(k)
is connected.
4
At the beginning of round k, a node u
(k)
is exogenously marked
as the active node. Node u
(k)
seeks to minimize its cost by the end of the round by undertaking either a unilateral or a two-
stage bilateral deviation. Unilateral deviation allows the active node to erase some of its adjacent edges. Bilateral deviation
allows a pair of nodes u
(k)
and w to perform any deviation consistent with 2-stability, namely, each node is allowed to erase
several of its adjacent link and at most one new edge (u
(k)
w) can be created. We assume that our dynamics are -local, in
the sense that node u
(k)
can deviate only with nodes w in its -neighborhood (i.e., d(u
(k)
, w; G
(k)
) ). A bilateral deviation
with node w proceeds in the following two stages:
Stage 1: the active node u
(k)
removes a (possibly empty) subset of its adjacent edges and suggest creating a link with
node w. Let G
(k)
1
be the state of the network at the end of the rst stage.
Stage 2: node w removes a (possibly empty) subset of its adjacent edges and decides whether to accept a suggestion to
create the link u
(k)
w.
These dynamics are myopic in the following sense. The active node u
(k)
selects its actions in order to minimize the cost
it experiences in the resulting network G
(k+1)
. If a bilateral deviation with some node w is chosen, node w accepts the link
suggestion from node u
(k)
if the cost node w experiences in G
(k+1)
is smaller than the cost it experiences in G
(k)
1
. Node w
can further minimize its cost in G
(k+1)
by erasing some of its adjacent edges.
LinkDynamics. Let , k and G
(k)
be as above. Similarly to NodeDynamics, at the beginning of round k, a node u
(k)
is
exogenously marked as the active node. An auxiliary node w
(k)
is then selected uniformly at random in the -neighborhood
of u
(k)
, i.e., d(u
(k)
, w
(k)
; G
(k)
) . Node u
(k)
is then allowed to minimize its cost at the end of the round by undertaking
either a unilateral deviation or a two-stage bilateral deviation with node w
(k)
. The two-stage bilateral deviation proceeds in
the exact same fashion as in NodeDynamics. Similarly, in case of bilateral deviation, node w
(k)
accepts the link suggestion if
its cost in G
(k+1)
is smaller than in G
(k)
1
. Node w
(k)
can further minimize its cost in G
(k+1)
by erasing some of its adjacent
edges.
It can be seen that essentially LinkDynamics activates a potential link (a pairs of nodes) in contrast to NodeDynamics
that activates a single node.
We assume that both NodeDynamics and LinkDynamics are lazy in the sense that the deviation is undertaken only if
it is strictly protable for the active node.
In the remainder of this section we turn our attention to two preliminary results on NodeDynamics and LinkDynamics.
We rst show that starting from any connected network, in expectation, after O(|V | log(|V |)) rounds, the state of the
network is a tree. We also prove that the set of trees is an attractor for our dynamics, namely, if the network in the current
round is a tree than it will remain a tree in all the following rounds. The second result states that when we consider our
dynamics on the set of trees, we can signicantly simplify the deviations considered by the active node.
Proposition 7. Let G
(0)
be a connected network. The following result holds both for NodeDynamics and for LinkDynamics:
In expectation, after k = O(|V | log(|V |)) rounds, the network G
(k)
is a tree. Furthermore, for all subsequent q >k, G
(q)
is a tree.
Proof. In what follows, we prove Proposition 7 for NodeDynamics. An identical argument also proves this proposition for
LinkDynamics.
We proceed in four steps. First, we prove that at the beginning of round k 0, network G
(k)
is connected. We then
prove that, at the end of the round, for all v V such that u
(k)
v G
(k+1)
, it holds that G
(k+1)
u
(k)
v is disconnected.
Next, we show that if the network topology is a tree at the beginning of a round, it remains so in all subsequent rounds.
We conclude by showing that it suces for all nodes to be activated at least once for the network topology to be a tree.
Hence, under the uniform activation process, the expected number of rounds until NodeDynamics converges to a tree is of
O(|V | log(|V |)).
Step 1. First, we prove by contradiction that the network topology is connected at the beginning of each round. Let k 0
be such that G
(k+1)
is disconnected, and that for all q k, G
(q)
is connected. By our assumption on the cost to a node
in a disconnected network, we conclude that C(u
(k)
, G
(k)
) < C(u
(k)
, G
(k+1)
). Recall that the active node u
(k)
can decide not
to deviate during round k, thus the previous inequality contradicts the assumption that u
(k)
selects its actions in order to
minimize its cost at the end of round k.
4
It is straightforward to show that, if G
(0)
is connected, then the network topology remains connected at the beginning of each round.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 9
Step 2. Assume by contradiction that there exists v V such that u
(k)
v G
(k+1)
and G
(k+1)
u
(k)
v is connected. In what fol-
lows, we suggest a unilateral deviation of node u
(k)
that is strictly more protable than deviation to G
(k+1)
. This contradicts
the denition of NodeDynamics by which G
(k+1)
is the most protable the deviation for node u
(k)
.
Let G

k
be the graph created from G
(k)
by removing all edges adjacent to node u
(k)
. This graph consists of connected
component {u
(k)
} and several additional connected components {C
i
}
L
i=1
for L > 0. Since the network G
(k)
is connected, there
exist w
i
C
i
for all i = 1, . . . , L such that u
(k)
w
i
G
(k)
. Let G

k
be the network created by addition of edges {u
(k)
w
i
}
L
i=1
to G

(k). It can be seen that this network can be obtained by unilateral deviation of node u
(k)
from G
(k)
, e.g., by uni-
lateral removal of all edges adjacent to u
(k)
except for {u
(k)
w
i
}
L
i=1
. We note that network G

k
is connected and it holds
that (u
(k)
; G

k
) = L. We proceed to show that (u
(k)
; G

k
) < (u
(k)
; G
(k+1)
), hence C(u
(k)
; G

k
) < C(u
(k)
; G
(k+1)
) by Corol-
lary 1.
Let G

k+1
be the graph created from G
(k+1)
by removing all edges adjacent to node u
(k)
. This graph consists of connected
component {u
(k)
} and several additional connected components {C

i
}
L

i=1
for L

> 0. By denition of NodeDynamics, only

edges adjacent to u
(k)
could have been added during the deviation from G
(k)
to G
(k+1)
, hence L

L. We further recall that

G
(k+1)
u
(k)
v is connected hence (u
(k)
; G
(k+1)
) 1 L

. Combining these observations yields (u

(k)
; G

k
) < (u
(k)
; G
(k+1)
)
which concludes the proof of this step.
Step 3. Let us show that the property discussed in Step 2 is preserved in subsequent rounds. Namely, for any m k + 1
and v V such that u
(k)
v G
(m)
the network G
(m)
u
(k)
v is disconnected. We again proceed by contradiction. Let m be
the rst round so that the property does not holds. Namely, there exists a cycle C = (v
0
= u
(k)
, v
1
, . . . , v
p
, v
p+1
= u
(k)
) in
G
(m)
containing u
(k)
. Since C is not contained in G
(m1)
, we conclude that there exists at least one 0 i p such that
v
i
v
i+1
/ G
(m1)
. Note that during round m1, only edges adjacent to u
(m1)
can be added, hence u
(m1)
{v
i
; v
i+1
}, and
thus u
(m1)
is in a cycle in G
(m)
. This is a contradiction to Step 2 in which we showed that all links u
(m1)
w G
(m)
being
such that G
(m)
u
(m1)
w are disconnected.
Step 4. We now show that the network topology eventually becomes a tree, and remains so thereafter. Let k be a round such
that all nodes in V have been activated at least once. It follows from Step 2 that, for all uv G
(k)
, G
(k)
uv is disconnected.
Since G
(k)
is connected, we conclude that G
(k)
is a tree. Furthermore, by Step 3 this remains true for all subsequent rounds.
Hence, after activating all nodes at least once, the network topology is always a tree at the beginning of each round.
To prove the bound on the number of rounds, note that when using the uniform activation process, bounding the
expected number of rounds needed to activate all nodes at least once is exactly the coupon collector problem
5
(Flajolet and
Sedgewick, 2009). Thus we conclude that, in expectation, after O(|V | log(|V |)) rounds, the network topology is a tree and
remains so. 2
An immediate consequence of Proposition 7 is that all xed points of the dynamics NodeDynamics and LinkDynamics
are 2-stable:
Corollary 8. Any xed point of NodeDynamics or LinkDynamics is a 2-stable network.
Proof. By Proposition 7, it holds that any xed point of NodeDynamics is necessarily a tree. By Statement 1 of Proposition 4,
any tree is 3-stable and therefore is also 2-stable. This implies any xed point of NodeDynamics is a 2-stable network. The
same argument also applies to xed points of LinkDynamics. 2
Another consequence of Proposition 7 is that we can assume without loss of generality that G
(0)
is a tree. We further
know that, for all k > 0, G
(k)
is a tree as well. The following result states that for both NodeDynamics and LinkDynamics
if the network topology changes in a given round k, then a bilateral deviation is undertaken. At the rst stage of this de-
viation a single link u
(k)
v is erased and at the second stage a single new link u
(k)
w is created. (By denition, w = w
(k)
under LinkDynamics.) For NodeDynamics, we characterize the set of possible node pairs (v, w). We show that selection
of a specic pair (v, w) is dictated only by the future total connectivity cost h(u
(k)
; G
(k+1)
) of the active node and is
independent of costs C(w; G
(k+1)
) and h(w; G
(k+1)
) incurred to the node w. There might be more than one bilateral de-
viation that maximize the value of h(u
(k)
; G
(k+1)
). In order to break such ties, we index all nodes and give preference to
deviations in which node v has the smallest index. If several deviations use node v with the minimal index, we give
preference to the deviation with the smallest index of node w. For LinkDynamics, we show that the bilateral devia-
tion is taken if h(u
(k)
; G
(k+1)
) < h(u
(k)
; G
(k)
), independently of costs C(w
(k)
; G
(k+1)
) and h(w
(k)
; G
(k+1)
) incurred to the
node w.
5
The coupon collector problem is the following well-known problem in analytical combinatorics. Suppose there are coupons of n different types. Each
time step, a person collects a coupon drawn uniformly at random from one of the types. Let N be the minimal number of coupons this person needs to
obtain in order to have at least one coupon of each type. Then, the expected value of N is of O(nlogn). We refer the reader to Example II.11 on pp. 116118
in Flajolet and Sedgewick (2009) for complete details.
10 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Proposition 9. Assume G
(0)
is a tree. Let k 0 be such that G
(k)
= G
(k+1)
. Then the following statements hold:
1. The active node performs a two-stage bilateral deviation under both NodeDynamics and LinkDynamics.
2. Let N(u
(k)
) be the set of all nodes x so that u
(k)
x G
(k)
. For any node v N(u
(k)
), let W(v) denote the set of all nodes y so that
nodes v and y are in the same connected component of the graph (G
(k)
u
(k)
v) and d(u
(k)
, y; G
(k)
) . Then,
For NodeDynamics, there exist v N(u
(k)
) and w W(v) such that G
(k+1)
= G
(k)
u
(k)
v +u
(k)
w.
For LinkDynamics, there exist v N(u
(k)
) such that w
(k)
W(v) and G
(k+1)
= G
(k)
u
(k)
v +u
(k)
w
(k)
.
3. Under both dynamics, the degree of the active node remains unchanged in the deviation, i.e., (u
(k)
; G
(k)
) =(u
(k)
; G
(k+1)
).
4. Under both dynamics, node w (w
(k)
for LinkDynamics) always accepts link suggestion from node u
(k)
. Moreover, this implies that
the specic selection of v (and w for NodeDynamics) is made by the active node with a goal of minimizing its total connectivity
cost h(u
(k)
; G
(k+1)
) (see Eq. (1) for denition) regardless of the values of C(w; G
(k+1)
) and h(w; G
(k+1)
) (or C(w
(k)
; G
(k+1)
) and
h(w
(k)
; G
(k+1)
) for LinkDynamics).
Proof. Statement 1. It suces to show that no unilateral deviation is possible. Consider NodeDynamics or LinkDynamics and
assume by contradiction that there exists a strictly protable unilateral deviation of node u
(k)
. This deviation will remove
at least one edge adjacent to u
(k)
thus creating a disconnected network. By the denition of the game, the cost incurred
to node u
(k)
will be larger than in the current network, which is a contradiction to the denitions of NodeDynamics and
LinkDynamics.
Statement 2. Consider NodeDynamics. Since at most one edge can be added during the second stage and since network
G
(k+1)
is connected, we conclude that exactly one edge can be removed during the rst stage. The removed edge is adjacent
to u
(k)
, as otherwise the deviation would increase the degree of u
(k)
, which is unprotable by Statement 1 of Corollary 1.
Therefore, the removed link is of the form u
(k)
v where v N(u
(k)
). Furthermore, the new link u
(k)
w restores the connectiv-
ity of the network (G
(k+1)
u
(k)
v) if and only if w W(v). Repeating the same argument for LinkDynamics, we conclude
that G
(k+1)
= G
(k)
u
(k)
v +u
(k)
w
(k)
, where v N(u
(k)
). Since G
(k+1)
is connected, it holds w
(k)
W(v).
Statement 3. This statement follows trivially by Statement 2.
Statement 4. Consider NodeDynamics. We begin by showing that the link u
(k)
w will be accepted by node w regardless of
values of C(w; G
(k+1)
) and h(w; G
(k+1)
). Indeed, note that the network G
(k)
1
= G
(k)
u
(k)
v is disconnected and G
(k)
1
+u
(k)
w =
G
(k+1)
is connected, thus C(w; G
(k)
1
) > C(w; G
(k+1)
). Hence, node w is compelled to accept the edge u
(k)
w. This observation
implies that the only restriction on the bilateral deviation is the minimization of the cost incurred to the active node.
Now consider LinkDynamics. Following the same argument, we obtain that w
(k)
is compelled to accept the edge u
(k)
w
(k)
since C(w; G
(k)
1
) > C(w; G
(k+1)
) for corresponding networks G
(k)
1
and G
(k+1)
. Again, this observation implies that the only
restriction on the bilateral deviation is the minimization of the cost incurred to the active node.
We conclude the proof of this statement by showing that the minimization of the cost incurred to the active node is
equivalent to the minimization of the connectivity cost alone. Indeed, by Statement 3, it holds that both under NodeDy-
namics and LinkDynamics, the degree (u
(k)
; ) of the active node remains unchanged. Combining this observation with the
denition of the cost given in Eq. (1) concludes the proof. 2
5. Distance-based cost models
In the remainder of the paper, we focus on a specic class of connectivity cost functions h relevant to transportation and
communication networks. In this section we describe this family of connectivity costs, that we refer to as distance-based
cost models. Informally, these models capture the fact that when trac is routed through a network, the cost incurred is
related to the distance traversed by that trac. The model we consider generalizes this intuition.
We consider a setting where each node u in a connected graph G wishes to send one unit of trac to another node v.
We assume this trac traverses a (simple) path P(u, v; G). (For a given edge xy G, we write xy P(u, v; G) if the edge is
traversed in the path P(u, v; G).) Let (u, v; G) be the cost associated with successfully sending a unit of trac from u to
v in G.
6
We assume that is a sum of costs over the edges traversed in P(u, v; G); formally, we have:
(u, v; G) =

eP(u,v;G)
(e), (4)
where is a positive weight function on the set of edges. We assume the following properties for the weight function
throughout the paper.
Assumption 2. Let be a weight function. For all three distinct nodes u, v and z in V ,
1. The edge cost is undirected: (u, v) =(v, u); and
2. The strict triangle inequality holds: (u, v) <(u, z) +(z, v).
6
We are considering the case where G is connected. When u and v are not in the same connected component, we assume that (u, v; G) is arbitrarily
large.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 11
An immediate consequence of Assumption 2 is the following proposition:
Proposition 10. Let G be a tree.
7
Let u, v and w be three nodes in G such that w is located on the (unique) simple path P(u, v; G)
from u to v in G. Then,
1. (u, v; G) = (v, u; G);
2. (u, v; G) = (u, w; G) + (w, v; G); and
3. if uv / G, then (u, v; G) >(u, v).
In the cost model we consider, the sender bears all the cost of sending all units of trac it sends throughout the
network. More generally, we assume that in a connected network G, the connectivity cost h(u; G) decomposes into the sum
of |V | 1 terms:
h(u; G) =

vV \{u}
g
_
(u, v; G)
_
, (5)
where g is an increasing, non-negative function. The connectivity cost function in Example 1 naturally ts into this setting:
it suces to set (u, v) =1 for all u, v, and g(x) = x. This yields (u, v; G) =d(u, v; G), and h(u; G) =

v
d(u, v; G) as in
the example.
In the following section we analyze behavior of NodeDynamics and LinkDynamics in games with distance-based cost
models. In particular, we focus on two subfamilies of these models:
(1) Homogeneous distance model. In this model it holds that (u, v) = 1 for all u, v. It follows that the connectivity cost
function h does not depend on the identity of the node but only on its position in the network.
(2) General linear model. In this model it holds that g(x) = x for all x, namely, the connectivity cost is linear in the weights
of the edges.
These subfamilies are not disjoint. For instance, the connectivity cost function in Example 1 belongs to both subfamilies.
6. Tree formation games
Section 4 made the observation that NodeDynamics and LinkDynamics converge to trees. Further, recall that in Propo-
sition 9, once these two dynamics converge to the set of trees, they exhibit a relatively simple dynamic behavior: either
no deviation is undertaken or one edge is removed and one edge is added. In this section, we exploit this property to
demonstrate an equivalence between NodeDynamics and LinkDynamics in the original game, and two versions of the best-
response dynamics in an auxiliary tree formation games. Because these new dynamics and the tree formation game are easier
to analyze, this equivalence proves instrumental in the results we derive in the remainder of the paper.
Formally, a tree formation game is a network formation game with the same game structure as in Section 3.3, but with
the following cost model: Given a graph G and a node u, the cost to u is arbitrarily large if G is not a tree, and otherwise equal
to h(u; G). In other words, in this game, there is a global constraint that the outcome must be a tree, and there is no link
formation cost; nodes are only subject to the connectivity cost. We consider 2-stability as our solution concept for this class
of games (cf. Section 3.5).
In particular, we consider the following two versions of the -local best-response dynamics for tree formation games
with respect to 2-stability. The rst dynamics we consider is called TreeNode. Let G
(k)
denote the network topology at the
beginning of round k. In TreeNode, we activate a node u
(k)
at round k (as in NodeDynamics). The node u
(k)
then seeks
to minimize its connectivity cost undertaking either a unilateral deviation or a bilateral deviation with some node w in
the -neighborhood of u
(k)
. We allow bilateral deviation only of the following specic form: node u
(k)
removes a single
adjacent edge and proposes creation of a single edge with node w. Naturally, node w accepts the deviation only if it does
not increase its connectivity cost. The same process then repeats indenitely. Similar to NodeDynamics, we assume that
the deviation is undertaken only if it is strictly protable for the active node u
(k)
. We further use an identical tie-breaking
mechanism.
We refer to the second dynamics as TreeLink and dene it in the following way. At the beginning of each round we
activate a pair of -close nodes and designate one of them to be the active node. We let u
(k)
denote the active node and
let w
(k)
denote the other node in the pair. The node u
(k)
then seeks to minimize its connectivity cost undertaking either
a unilateral deviation or a bilateral deviation with node v
(k)
. During bilateral deviation node u
(k)
is allowed to erase one
adjacent edge and to add the link u
(k)
w
(k)
. Node w
(k)
accepts this link suggestion if it does not increase its connectivity
cost. As usual, we assume that these dynamics are lazy, i.e., the deviation is undertaken only if it is strictly protable for
the active node u
(k)
.
We begin by introducing several properties of TreeNode and TreeLink:
7
We state Proposition 10 only in trees as in a tree there is a unique simple path between any two given nodes.
12 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Proposition 11.
1. Let tree network G be the current network and let u be the active node. Under both TreeNode and TreeLink, node u either does
nothing or it undertakes a bilateral deviation.
2. Let tree network G be the current network, let u be the active node, and let G

be the network after a single round of TreeNode or

TreeLink. Then (u; G) =(u; G

).
Proof. Statement 1. The proof of this statement is identical to the proof of Statement 1 in Proposition 9.
Statement 2. This statement follows trivially from Statement 1. 2
The following proposition is the main result of this section. It establishes the equivalence of TreeNode and NodeDy-
namics and of TreeLink and LinkDynamics, when the cost model is a distance-based cost model (as dened in the previous
section). In the remainder of the paper, we exploit this equivalence to prove convergence of NodeDynamics and LinkDynam-
ics under various assumptions on and/or g (cf. Section 5), and to provide insight into the xed points of the dynamics.
Proposition 12 (Equivalence of dynamics). Let > 1 be given. Assume G is a tree, and x a node u. Let the connectivity cost h(; ) be
given by (5).
1. Let G
1
be the network topology after one round of NodeDynamics, where u is the active node, and G is the initial state of the
network. Similarly, let G
2
be the network topology after one round of TreeNode, where u is the active node and G is the initial
state of the network. Then G
1
= G
2
.
2. Fix a node w in the -neighborhood of node u. Let G
3
be the network topology after one round of LinkDynamics, where u is the
active node, w is node us deviation partner and G is the initial state of the network. Similarly, let G
4
be the network topology
after one round of TreeLink, where u is the active node, w is node us deviation partner and G is the initial state of the network.
Then G
3
= G
4
.
The proof of the proposition above relies on the following lemma:
Lemma 13. Let G be a tree, and let u = v be two nodes such that uv / G. Assume the connectivity cost h(;) is given by (5). Let
w P(u, v; G) and uw G. Then h(v; G) >h(v; G uw +uv).
Proof. For notational simplicity, let G

= G uw + uv. First, note that v = w, hence (u, v; G) > (u, v) by Statement 3

of Proposition 10. Next, since uv G

, (u, v; G

) = (u, v), and thus (u, v; G) > (u, v; G

). Let T
u
be the subtree of G
rooted at u obtained by removing uw. Similarly dene T
w
. It follows that v T
w
. We can now write
h(v; G) h
_
v; G

_
=

zT
u
_
g
_
(v, z; G)
_
g
_

_
v, z; G

___
+

zT
w
\{v}
_
g
_
(v, z; G)
_
g
_

_
v, z; G

___
.
But, for all z T
w
, P(v, z; G) = P(v, z; G

), hence the second term of the above sum is zero. Next, note that for all
z T
u
, (v, z; G) = (v, u; G) + (u, z; G) and (v, z; G

) = (v, u; G

) + (u, z; G

). Now, since vu / G and vu G

,
by Proposition 10, we conclude (v, u; G) > (v, u; G

). Since P(u, z; G) = P(u, z; G

) for all z T
u
, we conclude that
(v, z; G) > (v, z; G

). Since g is strictly increasing, we conclude that

g
_
(v, z; G)
_
g
_

_
v, z; G

__
> 0,
which proves h(v; G) >h(v; G uw +uv), as claimed. 2
We proceed to prove Proposition 12.
Proof of Proposition 12. Statement 1. Let N(u) be the set of all nodes x so that ux G. For any node v N(u), let W(v)
denote the set of all nodes y so that nodes v and y are in the same connected component of the graph (G uv) and
d(u, y; G) . By Statement 3 of Proposition 11 it holds that the active node u under TreeNode either does nothing or
undertakes a deviation G
2
= G uv +uw for some pair of nodes v N(u) and w W(v). We note that the latter deviation
is protable to u only if w and v belong to the same connected component in (G uv), namely, w W(v). Otherwise, the
network G
2
would be disconnected thus incurring a very high cost to node u. We further note that w W(v) implies that
v P(u, w; G), thus by Lemma 13 it holds that h(w; G) h(w; G uv +uw). This means the link uw will be accepted by
node w for any selection of v N(u) and w W(v), hence specic selection of v and w is made with a goal of minimizing
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 13
h(u; ) regardless of h(w; ). Comparing the observations above with Statements 2 and 4 of Proposition 9, we conclude that
both dynamics perform the same deviation.
Statement 2. The proof of Statement 2 follows an argument similar to the proof of Statement 1. By Statement 3 of
Proposition 11 it holds that the active node u under TreeNode either does nothing or undertakes a deviation G
2
= G
uv +uw for some v N(u) such that w W(v). Again, the observation that w W(v) implies that v P(u, w; G), thus
by Lemma 13 it holds that h(w; G) h(w; G uv +uw). This means the link uw will necessarily be accepted by node w.
Hence specic selection of v is made with a goal of minimizing h(u; ) regardless of h(w; ). Comparing the observations
above with Statements 2 and 4 of Proposition 9, we conclude that dynamics LinkDynamics and TreeLink perform the same
deviation. 2
The following result follows trivially from Proposition 12.
Corollary 14. TreeNode and NodeDynamics have the same xed points. Moreover, given that the initial network is a tree, these
dynamics have the same convergence rate. Similarly, TreeLink and LinkDynamics have the same xed points and, given that the
initial network is a tree, have the same convergence rate.
The rest of the paper studies the convergence of our dynamics under specic assumptions on the weight function
and/or the function g. We establish convergence results for TreeNode and TreeLink, which automatically apply to NodeDy-
namics and LinkDynamics, respectively.
7. General linear model
In this section we consider connectivity cost models of the form (5), where g(x) = x, with general weight functions .
We prove that TreeNode and TreeLink (and thus NodeDynamics and LinkDynamics) converge almost surely, and that in
expectation, convergence time is polynomial in the number of nodes. Finally, we prove that the price of anarchy of both
dynamics is bounded from above by an expression that depends on the edge weights. For a certain edge weight functions,
this expression has (small) value that is independent of graph cardinality.
We begin by introducing several auxiliary notations. Let
max
and
min
denote the minimal and the maximal edge
weights respectively. Namely,

max
max
u,v
(u, v) and
min
min
u,v
(u, v).
Recall that edge weights function satises strict triangular inequality. The next two expressions we introduce reect how
close this inequality to be an equality.
max
u,v,wV
(v, u) +(u, w) (v, w)
(v, w) +(w, u) (v, u)
<, (6)
min
u,v,wV
_
(v, u) +(u, w) (v, w)
_
> 0. (7)
Here, the inequalities are implied by Assumption 2.
7.1. Convergence analysis
In this subsection we prove that for all neighborhood widths > 1, the -local dynamics TreeNode and TreeLink con-
verges almost surely. Our main result is the following theorem.
Theorem 15 (Convergence in the linear model). Let > 1 be given. Assume the activation process is i.i.d. with full support. Finally,
assume that T
(0)
is a tree. Then the -local dynamics TreeNode and TreeLink converge almost surely.
Proof. We rst prove that if the total connectivity cost function (see denition in (2)) is an ordinal potential function,
then the dynamics converge almost surely. Formally, assume that for all k 0, if T
(k)
= T
(k+1)
, then (T
(k)
) > (T
(k+1)
).
Note that for all k 0, T
(k)
is a tree. Thus the number of possible states of the network is nite. Hence, by a standard
ordinal potential function argument (Jackson, 2008), the dynamics converge.
Assume T
(k)
is not a xed point of the dynamics. For TreeNode, this implies that there exists u V such that, if u
is activated at round k, then T
(k)
= T
(k+1)
. We assumed the activation process to be i.i.d. with full support, hence u is
activated with positive probability at round k; the dynamics converge almost surely. For TreeLink, this implies that there
exists a pair of -close nodes u, v such that if u is activated at round k and node v is selected as his deviation partner,
then T
(k)
= T
(k+1)
. Again, we assumed that the activation process to be i.i.d. with full support and that the second node for
deviation is selected uniformly at random from the -neighborhood of the active node. Hence, the pair u, v can be chosen
with positive probability at round k; the dynamics converge almost surely.
14 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
We now prove that the total connectivity cost function is an ordinal potential function. Let k 0 be a given round,
with u
(k)
the active node, such that T
(k)
= T
(k+1)
. Then there exist two distinct nodes v = u
(k)
and w = u
(k)
such that
T
(k+1)
= T
(k)
u
(k)
v +u
(k)
w. (It holds that w = w
(k)
for TreeLink.) Also, since u
(k)
is the active node,
h
_
u
(k)
; T
(k)
_
>h
_
u
(k)
; T
(k+1)
_
.
Now recall that the total connectivity cost function can be written as follows
(G) =

xV
h(x; G) =

xV
_

y=x
(x, y; G)
_
.
For T = T
(k)
or T = T
(k+1)
, we can further decompose this expression into four terms:
(T ) =

(x,y)T
2
u
(x, y; T ) +

(x,y)T
u
T
v
(x, y; T ) +

(x,y)T
v
T
u
(x, y; T ) +

x,yT
2
v
(x, y; T ),
where T
u
is the subtree of T
(k)
rooted at u
(k)
obtained after removing u
(k)
v from T
(k)
, and T
v
is similarly the subtree
rooted at v. First, note that by Assumption 2 the second and third terms are equal. Now, note that for all (x, y) T
u
T
u
,
P(x, y; T
(k)
) = P(x, y; T
(k+1)
) = P(x, y; T
u
). Hence (x, y; T
(k)
) = (x, y; T
(k+1)
). Similarly, we can prove that, for all (x, y)
T
v
T
v
, (x, y; T
(k)
) = (x, y; T
(k+1)
). We can now write the difference in total connectivity cost as follows

_
T
(k)
_

_
T
(k+1)
_
=2

(x,y)T
u
T
v
_

_
x, y; T
(k)
_

_
x, y; T
(k+1)
__
=2

(x,y)T
u
T
v
_

_
x, u
(k)
; T
(k)
_
+
_
u
(k)
, y; T
(k)
_

_
x, u
(k)
; T
(k+1)
_

_
u
(k)
, y; T
(k+1)
__
=2

xT
u

yT
v
_

_
u
(k)
, y; T
(k)
_

_
u
(k)
, y; T
(k+1)
__
. (8)
We observe that

yT
v
_

_
u
(k)
, y; T
(k)
_

_
u
(k)
, y; T
(k+1)
__
=

yT
v

_
u
(k)
, y; T
(k)
_

yT
v

_
u
(k)
, y; T
(k+1)
_
+

xT
u

_
u
(k)
, x; T
(k)
_

xT
u

_
u
(k)
, x; T
(k+1)
_
=h
_
u
(k)
; T
(k)
_
h
_
u
(k)
; T
(k+1)
_
> 0, (9)
where the second equality stems from observation that P(u
(k)
, x; T
(k)
) = P(u
(k)
, x; T
(k+1)
) for all x T
u
. Combining Eqs. (8)
and (9) we obtain that (T
(k)
) (T
(k+1)
) > 0, which concludes the proof. 2
Note that the proof of Theorem 15 provides an upper bound for the expected number of steps needed for the dynamics
to converge. At each round where the network topology changes, the value of is strictly decreased. We exploit this
property to derive the following bound on the convergence time of NodeDynamics and LinkDynamics.
Lemma 16 (Expected convergence time in the linear model). Assume that takes only integer values and that the activation process
is uniform. Starting from any connected graph G
(0)
, the expected number of networks visited by NodeDynamics is O(n
4
), which is
polynomial in the size n of the graph. The expected number of networks visited by LinkDynamics in this scenario is O(n
5
).
Proof. From Proposition 7, in expectation the network is a tree after O(nlog(n)) rounds. Now note that for any tree T on
n nodes, (T ) < n
3

max
= O(n
3
), where
max
< by assumption. Thus the dynamics can visit at most O(n
3
) distinct
networks, since the value of the total cost must decrease each time the network topology changes.
Consider NodeDynamics. As long the current network is not a xed point of NodeDynamics, there exists at least one
node whose activation will yield a topology change. Assuming that the activation process is uniform, the expected time
between these activation is of O(n). Thus NodeDynamics converge in expectation in O(nlog(n)) + O(n
4
) = O(n
4
) rounds.
Now consider LinkDynamics. As long the current network is not a xed point of LinkDynamics, there exists at least one
pair of nodes whose activation will yield a topology change. Assuming that the activation process is uniform, the expected
time between these activation is of O(n
2
), hence the convergence time is O(n
5
). 2
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 15
7.2. Price of anarchy
We now turn our attention to the price of anarchy of the xed points of NodeDynamics and LinkDynamics (or equiv-
alently, TreeNode and TreeLink, respectively) in this model. We begin this section with a lower and upper bound on the
total connectivity cost of any network. Recall the denition of
max
and
min
in the beginning of this section.
Lemma 17. For any connected network G on n nodes, n(n1)
min
(G) n(n1) diam(G)
max
, where diam(G) is the diameter
of G.
Proof. Let G be a connected network, and let diam(G) be its diameter. Let u = v be two given nodes in V . First, note that
P(u, v; G) traverses at least one edge and at most diam(G) edges. Hence
min
(u, v; G) diam(G)
max
. Next, recall the
denition of the total connectivity cost:
(G) =

uV
h(u; G) =

uV
_

v=u
(u, v; G)
_
.
Thus, the total connectivity cost is the sum of n(n 1) terms, which proves the result. 2
An immediate corollary of Lemma 17 is that for the tree formation game, the total connectivity cost (T
n
) of a sequence
of trees T
n
with constant diameter is within a constant of the optimal total connectivity cost achievable.
Corollary 18. Let T be a tree with n nodes and let
OPT
be the total connectivity cost of the ecient outcome of the tree formation
game. Then,

OPT
(T )

max

min
diam(T )
OPT
. (10)
Proof. By denition of
OPT
it holds that
OPT
(T ). By Lemma 17, it holds that
OPT
n(n 1)
min
and (T )
n(n 1)
max
diam(T ), hence

max

min
diam(T )
OPT
n(n 1)
max
diam(T ) (T ).
We thus proved inequality (10). 2
From the preceding corollary, we observe that if the xed points of NodeDynamics and LinkDynamics have constant
diameter independently on graph size, then their price of anarchy is constant as well. In the next theorem we establish this
result.
Theorem19 (Constant price of anarchy of xed points). Let > 1 be given. Assume T is a xed point of TreeNode or TreeLink with n
nodes, and let be its diameter. Then 1 +2, where is dened in Eq. (6). Thus the price of anarchy of the set of xed points of
TreeNode and TreeLink (and of NodeDynamics and LinkDynamics, respectively) is bounded fromabove by the following expression:
PoA

max

min
(1 +2).
Proof. For the duration of the proof, we suppress the dependence on the number of nodes n for notational simplicity. In
what follows we prove the theorem for dynamics TreeNode. The theorems correctness for TreeLink follows by an identical
argument.
Assume T is a xed point of dynamics TreeNode, and let be its diameter. For the case of = 2 (star network) the
theorem follows trivially, therefore in what follows we focus on the case 3.
This proof proceeds through the following steps. Consider a path P of length in the tree T , which exists by denition
of the diameter of the graph. We partition nodes in the tree T into disjoint subsets in the following way. The rst two
subsets contain one leaf each from path P. Each one of the additional (2) disjoint sets contains at least a 1/(1 +2)-
fraction of all n nodes in the network. This readily implies an upper bound on , which in combination with Corollary 18
proves the claim.
This idea is made formal in the following way. Let us assume the following property, which is proved later on.
Property (): Let u be an internal node (i.e., not a leaf) and let v and w be such that vu T , wu T and v = w.
We denote by T
u
the subtree of (T vu wu) rooted at u and let V (T
u
) be the set of nodes in this subtree. Then,
|V (T
u
)| n/(1 +2).
Assume Property () holds. Let P = v
0
v
1
. . . v
1
v

be a path in T of length . First, note that the longest path in the

tree graph lies between two leafs, hence (v
0
; T ) =(v

; T ) =1, and that v

i
are internal nodes for all i with 1 i 1.
16 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Fig. 1. Illustration for the proof of Theorem 19. Proof of constant diameter. Here, bold points stand for nodes and lines stand for edges.
Fig. 2. Illustration for the proof of Theorem 19. Protable joint deviation by v and w. Here, bold points stand for nodes and lines stand for edges.
For each i such that 1 i 1, dene T
i
to be the subtree of T v
i1
v
i
v
i+1
v
i
rooted at v
i
(see Fig. 1). Then, by
assumption, |V (T
i
)| n/(1 +2).
Since both v
0
and v

are leaves,
V ={v
0
, v

}
_
1
_
i=1
V (T
i
)
_
.
Further, from the denition of the T
i
s, we can conclude that, for 1 i < j 1, V (T
i
) V (T
j
) =. Hence
|V | =n =2 +

i=1

V (T
i
)

2 + n/(1 +2),
which proves that (1 +2)(n 2)/n (1 +2). Let F be the set of all xed points of TreeNode. Then,
PoA(F ) =
max
T F
(T )

OPT

max

min
max
T F
diam(T )

max

min
(1 +2).
Here, the rst equality follows by the denition of price of anarchy and of
OPT
. The rst inequality follows by Corollary 10
and the last inequality follows from previously obtained bound on .
We now prove Property (). Assume u is an internal node, and let v and w be two of its neighbors. Dene T
u
, T
v
and
T
w
as in Fig. 2.
Consider the deviation in Fig. 2. Let T

= T uv + vw. From Lemma 13, we know that h(w; T ) >h(w; T

). Since T is a
xed point of TreeNode than h(v; T ) h(v; T

).
Let z
u
T
u
, z
v
T
v
and z
w
T
w
and calculate h(v; T ) h(v; T

):
h(v; T ) h
_
v; T

_
=

z
u
V (T
u
)
_
(v, z
u
; T )
_
v, z
u
; T

__
+

z
v
V (T
v
)
_
(v, z
v
; T )
_
v, z
v
; T

__
+

z
w
V (T
w
)
_
(v, z
w
; T )
_
v, z
w
; T

__
=

z
u
V (T
u
)
_
(v, z
u
; T )
_
v, z
u
; T

__
+

z
w
V (T
w
)
_
(v, z
w
; T )
_
v, z
w
; T

__
=

z
u
V (T
u
)
_
(v, u; T ) + (u, z
u
; T )
_
v, u; T

_
u, z
u
; T

__
+

z
w
V (T
w
)
_
(v, w; T ) + (w, z
w
; T )
_
v, w; T

_
w, z
w
; T

__
=

z
u
V (T
u
)
_
(v, u; T )
_
v, u; T

_
+ (u, z
u
; T )
_
u, z
u
; T

__
+

z
w
V (T
w
)
_
(v, w; T )
_
v, w; T

_
+ (w, z
w
; T )
_
w, z
w
; T

__
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 17
=

z
u
V (T
u
)
_
(v, u)
_
(v, w) +(w, u)
__
+

z
w
V (T
w
)
_
(v, u) +(u, w) (v, w)
_
=

V (T
u
)

_
(v, u)
_
(v, w) +(w, u)
__
+

V (T
w
)

_
(v, u) +(u, w) (v, w)
_
.
Here, the rst equality is obtained by substitution and rearranging the terms. The second equality stems from observa-
tion that (v, z
v
; T ) = (v, z
v
; T

) for all z
v
V (T
v
). Observations that (w, z
w
; T ) = (w, z
w
; T

) and (u, z
u
; T ) =
(u, z
u
; T

) for all z
w
V (T
w
) and z
u
V (T
u
) yield the third inequality. The last three equalities are algebraic.
Rearranging the terms and recalling that h(v; T ) h(v; T

) yields that

V (T
u
)

V (T
w
)

.
Following the same argument we obtain that |V (T
u
)| |V (T
v
)|. Summing the last two inequalities with the trivial in-
equality |V (T
u
)| |V (T
u
)| yields
(1 +2)

V (T
u
)

V (T
v
)

V (T
u
)

V (T
w
)

_
=n,
which concludes the proof. 2
Recall the cost model introduced in Example 1 and note that it can be seen as a general linear model for 1. For
this case it holds that
max
= 1,
min
= 1 and = 1. Therefore, the price of anarchy for all xed points of NodeDynamics
and LinkDynamics is bounded from above by 3. Recall also that by Proposition 6 the price of anarchy of strongly stable
outcomes in this model is (n). We therefore conclude for this class of models NodeDynamics and LinkDynamics allow us
to select a much more ecient outcomes than strong stability.
8. Homogeneous distance model
In Section 7 we considered connectivity cost models of the form (5) where we xed g(x) = x and considered general
weight functions . In this section we consider connectivity cost models where the weight function is uniform over all
pairs of nodes, and focus on two general classes of functions g.
We rst prove that both the least (path) and the most (star) ecient connected networks are strongly stable. Therefore,
strong stability does not provide any guarantees on the eciency of networks. We then consider the set of xed points
of TreeNode and TreeLink (and hence of NodeDynamics and LinkDynamics, respectively) for two classes of functions g.
For one of the two classes of functions g considered, the xed points are ecient (for all > 1). For the other class of
functions g, eciency is only guaranteed for the global version of dynamics TreeNode and TreeLink (i.e. when n 1).
Without loss of generality, we assume = 1. From Eq. (4) it follows that (u, v; G) = d(u, v; G) for all nodes u = v
in the same connected component of a graph G. It follows that, for all connected networks G and all nodes u V , the
connectivity cost function h is given by
h(u; G) =

vV
g
_
d(u, v; G)
_
. (11)
We assume that the function g is strictly increasing and that g(0) =0.
This model incorporates two important connectivity cost functions that have been extensively studied, as explained in
the examples below.
Example 2 (Linear connectivity cost). Assume that the connectivity cost function g is the identity function g(m) =m. Then
the connectivity cost to a node is proportional to its average distance to any other node in the network. This setting has
been studied by Fabrikant et al. (2003), as well as Corbo and Parkes (2005) and Demaine et al. (2007).
Example 3 (Exponential connectivity cost). Let > 0 be a given absolute constant. Assume that the connectivity cost function
is g(m) =

m1
i=0

i
. For such a connectivity cost function, the interpretation is that the marginal cost of each extra edge
needed to connect to another node grows/decays ( > 1 and < 1 respectively) exponentially.
Then it is easy to see that, up to a rescaling and shifting of the cost, the connectivity cost function is that of the
connections game rst studied by Jackson and Wolinsky (1996).
8.1. Static analysis
In this section we present the following result on network stability for the general homogeneous distance model:
Theorem 20. Let the connectivity cost function h satisfy Eq. (11). Then the following statements hold:
1. All star networks are strongly stable in the link announcement game.
18 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
2. All line networks are strongly stable in the link announcement game.
3. All star networks are ecient with respect to .
4. Line networks are the least ecient among tree networks. Namely, let L be a line network and let T be any tree network respec-
tively. Then, (L) (T ).
Proof. Statement 1 is identical to Statement 2 of Proposition 4, hence was already proved.
Statement 2. Let L be a line network. Assume by contradiction that there exists a set S V of nodes and a network G
such that G S-defeats L. By Statement 3 of Proposition 4 it holds that G is also a line network.
The proof of this statement proceeds through the following steps. We begin by dening an auxiliary notion of the nodes
rank in a line network, which is the distance between this node and the closest leaf. We show that cost incurred by a node
is fully dened by its rank. Furthermore, the higher the rank (the closer the node to the middle of the line) the smaller the
corresponding cost. We then focus on the deviation from network L to G and show that there exists a node v S whose
rank in G is smaller than in L. This implies that his cost in G is greater than in L, which contradict the assumption that G
S-defeats L, hence concludes the proof.
Let v
1
and v
n
be leaves in network L, namely, (v
1
; L) = (v
n
; L) = 1. By Statement 2 of Lemma 5, their degrees in
network G are identical to these in L, hence (v
1
; G) =(v
n
; G) =1 and nodes v
1
and v
n
are leaves in G. We further note
that any internal node u in network L is also an internal node in G. Let v
i
for i =2, 3, . . . , n 1 denote the unique node in
network L so that d(v
1
, v
i
) =i 1.
Dene the rank function (over the set of nodes) in both L and G as follows:
r(u; G) =min
_
d(u, v
1
; G), d(u, v
n
; G)
_
and r(u; L) =min
_
d(u, v
1
; L), d(u, v
n
; L)
_
.
We introduce the following two properties that will be established later on:
Property 1. For any node v / {v
1
, v
n
}, it holds that
r(v; G) >r(v; L) h(v; G) <h(v; L)
or equivalently, C(u; G) < C(v; L).
Property 2. For any three nodes x, y, z so that d(x, v
1
; L) <d(y, v
1
; L) <d(z, v
1
; L) it holds that
r(y; L) =min
_
r(x; L) +d(x, y; L), r(z; L) +d(z, y; L)
_
.
Similarly, for all x, y, z such that d(x, v
1
; G) <d(y, v
1
; G) <d(z, v
1
; G) it holds that
r(y; G) =min
_
r(x; G) +d(x, y; G), r(z; G) +d(z, y; G)
_
.
We now proceed to prove Statement 2. Note that since G S-defeats L, there exists a node v S so that C(v; L) > C(v; G).
By Statement 2 of Lemma 5 it implies that h(v; L) > h(v; G), hence r(v; L) < r(v; G) by Property 1. Furthermore, both
networks contain exactly two nodes of rank 0, two nodes of rank 1 and so on. Hence, by pigeonhole principle, there exist
a node v
l
V so that r(v
l
; L) >r(v
l
; G). It is evident that v
l
/ S since h(v
l
; L) <h(v
l
; G) by Property 1. Moreover, l > 1 and
l <n since the rank of nodes v
1
and v
n
equals 0 in both networks. Consider a node v
k
such that
k =max{1 i l 1 s.t. v
i
S}.
We note that such k exists as otherwise the path between nodes v
1
and v
l
would contain the same nodes both in L and
in G. This would imply that d(v
1
, v
l
; L) =d(v
1
, v
l
; G), hence r(v
l
; L) =r(v
l
; G) in contradiction with the denition of v
l
.
We note that v
i
/ S for all k < i l, hence the path between nodes v
k
and v
l
contains the same nodes both in G
and in L. This implies that d(v
k
, v
l
; G) = d(v
k
, v
l
; L). Moreover, since v
k
S it holds that C(v
k
; G) C(v
k
; L), therefore
r(v
k
; G) r(v
k
; L) by Property 1. Combining the last observations we obtain
r(v
k
; G) +d(v
k
, v
l
; G) r(v
k
; L) +d(v
k
, v
l
; L). (12)
Following a similar argument there exists a node v
m
such that
m=min{l +1 i n 1 s.t. v
i
S}.
As above, we obtain that d(v
m
, v
l
; G) =d(v
m
, v
l
; L) and
r(v
m
; G) +d(v
m
, v
l
; G) r(v
m
; L) +d(v
m
, v
l
; L). (13)
Combining Property 2 and Eqs. (12) and (13) yields:
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 19
r(v
l
; G) =min
_
r(v
k
; G) +d(v
k
, v
l
; G), r(v
m
; G) +d(v
m
, v
l
; G)
_
min
_
r(v
k
; L) +d(v
k
, v
l
; L), r(v
m
; L) +d(v
m
, v
l
; L)
_
=r(v
l
; L).
Thus, r(v
l
; G) r(v
l
; L) in contradiction with the denition of v
l
.
We conclude the proof of Statement 2 by establishing Properties 1 and 2. Property 1 follows from the facts that g is
strictly increasing, (u; G) =(u; L) for any u V , and that for any u / {v
1
, v
n
} it holds that
h(u; G) =2
r(u;G)

i=1
g(i) +
nr(u;G)1

i=r(u;G)+1
g(i) and h(u; L) =2
r(u;L)

i=1
g(i) +
nr(u;L)1

i=r(u;L)+1
g(i).
Now consider network L. Property 2 is established by consideration of the following four cases:
Case 1d(x, v
1
; L) d(x, v
n
; L), d(y, v
1
; L) d(y, v
n
; L) and d(z, v
1
; L) d(z, v
n
; L). Namely, all three nodes x, y and z
are closer to node v
1
than to node v
n
. Recalling that d(v
1
, x; L) < d(y, v
1
; L), we obtain that node x is also closer to node
v
1
than v
n
. By denition of the rank it holds that r(x; L) =d(x, v
1
; L), r(y; L) =d(y, v
1
; L) and r(z; L) =d(z, v
1
; L). Further,
the relative location of nodes x, y, and z implies that
d(y, v
1
; L) =d(v
1
, x; L) +d(x, y; L) =d(v
1
, z; L) d(z, y; L).
Combining with observations above yields that r(y; L) =r(x; L) +d(x, y; L) and
r(z; L) +d(z, y; L) =r(y; L) +2d(z, y; L) >r(y; L),
establishing Property 2 for this case.
Case 2d(x, v
1
; L) d(x, v
n
; L), d(y, v
1
; L) d(y, v
n
; L) and d(z, v
1
; L) >d(z, v
n
; L). Namely, nodes x and y are closer to
v
1
than to v
n
and node z is closer to node v
n
than to v
1
. Using an argument similar to the previous case we show that
r(y; L) =r(x; L) +d(x, y; L) and that
r(z; L) +d(z, y; L) =d(y, v
n
; L) d(y, v
1
; L) =r(y; L).
Case 3d(x, v
1
; L) d(x, v
n
; L), d(y, v
1
; L) >d(y, v
n
; L) and d(z, v
1
; L) >d(z, v
n
; L). Namely, node x is closer to v
1
than
to v
n
and nodes y and z are closer to node v
n
than to v
1
. Property 2 for this case is established following an argument
analogous to Case 2 above.
Case 4d(x, v
1
; L) > d(x, v
n
; L), d(y, v
1
; L) > d(y, v
n
; L) and d(z, v
1
; L) > d(z, v
n
; L). Namely, all three nodes x, y and x
are closer to v
n
than to v
1
. Property 2 for this case is established following an argument analogous to Case 1 above.
Statement 3. First note that for |V | < 4 the result is trivially true as all trees over less than four nodes are stars. Let T
be a non-star tree. Then, there exists at least two nodes x, y V such that (x; T ) > 1 and (y; T ) > 1. Hence there exist a
node l
x
in the neighborhood of x that is not in the shortest path from x to y in T . Formally, let l
x
V be such that l
x
x T
and l
x
/ P(x, y; T ). Similarly dene l
y
. Then P(l
x
, l
y
; T ) ={l
x
} P(x, y; T ) {l
y
}, and thus d(l
x
, l
y
; T ) 3.
Now note that we can write (T ) in the following way:
(T ) =

(u,v)V V
g
_
d(u, v; T )
_
;
and note that, for all (u, v) V V , d(u, v; T ) = 1 if and only if uv T . Since T is a tree, it has exactly (n 1) edges,
where n =|V |. Thus,
(T ) =2(n 1)g(1) +

(u,v)V V |uv / T
g
_
d(u, v; T )
_
.
Since d(l
x
, l
y
; T ) 3, we conclude that
(T ) 2(n 1)g(1) +2g(3) +
_
n(n 1) 2(n 1) 2
_
g(2),
where the second term comes from the bound on d(l
x
, l
y
; T ), and the third one from the following facts: (1) the term g(2)
is a lower bound on g(d(u, v; T )) for (u, v) such that uv / T ; and (2) the term [n(n 1) 2(n 1) 2] is the number of
pairs (u, v) such that u = v, uv / T and u, v / {l
x
, l
y
}.
Let S be a star. It follows that (S) =2(n1)g(1) +[n(n1) 2(n1) 2]g(2), and thus (T ) (S) 2(g(3) g(2))
> 0 as claimed.
Statement 4. Note that, for n |V | < 4, up to isomorphism of graphs, there is a unique tree. For n =4, there are exactly
two non-isomorphic trees: the star tree S
4
and the line tree L
4
. For the base case n = 4, we prove that (L
4
) > (S
4
).
A straightforward calculation yields
(L
4
) =6g(1) +6g(2),
(S
4
) =6g(1) +4g(2) +2g(3),
hence (S
4
) (L
4
) =2(g(3) g(2)) > 0, which proves the base case.
20 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Assume that, for all non-line trees T of size n 1 4, (T ) < (L
n1
), where L
k
is a line network over k nodes.
Let T

be a tree over n nodes, non-isomorphic to L

n
. Since T

is a tree over more than two nodes, there exist a leaf l

u
in
T

and an internal node u in T

such that l
u
u T

. Let F = T

l
u
u be the forest over V obtained by removing the link l
u
u
from T . Then F has two connected components: {l
u
} and T
u
, the subtree of T

rooted at u not containing l

u
. We conclude
that T
u
has n 1 nodes. By the induction hypothesis, (T
u
) < (L
n1
).
Let l
v
L
n
be a leaf of a line network over n nodes. By a similar argument, there exists v L
n
such that l
v
v L
n
and
L
n
l
v
v is a forest composed of a line network of size n 1, and {l
v
}.
Now note that (T

) = (T
u
) +2h(l
u
; T

) and (L
n
) = (L
n1
) +2h(l
v
; L
n
). Thus it suces to prove h(l
v
; L
n
) h(l
u
; T

).
Let x be the node that is located furthers from node l
u
in T

and let d
1
= d(l
u
, x; T

) be the distance between them.

Then,
h
_
l
u
; T

_
=

zP(l
u
,x;T

)
g
_
d
_
l
u
, z; T

__
+

zV \P(l
u
,x;T

)
g
_
d
_
l
u
, z; T

__

d
1

i=1
g(i) +
n1

i=d
1
+1
g(i) h(l
v
; L
n
).
Here, the rst equality is obtained by substitution and rearranging the terms. The rst inequality follows from the fact that
d(l
u
, z; T

) i for all z = x and i d

1
. 2
An important implication of theorem above is that the price of anarchy of the set of strongly stable networks is the same
as the price of anarchy of the set of trees:
Corollary 21 (Price of anarchy and strong stability). Assume the connectivity cost function is given by Eq. (11). Under Assumption 1,
the price of anarchy of the link announcement game under strong stability is identical to that obtained under 3-stability or just of the
set of trees.
Proof. Recall, that by Statement 1 of Proposition 4 the set of tree networks and the set of all 3-stable networks are identical.
Then, Corollary 21 follows from Theorem 20. 2
8.2. Convergence analysis
In this section we focus on two classes of functions g. For each class, we prove that TreeNode and TreeLink converges
almost surely, bound its converge rate and analyze the eciency of the selected outcomes.
8.2.1. Assumption on g-bounded discrete derivative
The rst class of functions g is the following.
Property 1 (Bounded discrete derivatives). We say that g has bounded discrete derivatives if g is such that
i 3, g(i) g(i 1) g(3) g(2) g(2) g(1) > 0. (14)
Note that g satisfying the condition from Property 1 implies that g is strictly increasing. Further, note that all strictly
increasing, convex functions satisfy condition (14).
We begin with some useful results about trees. We rst introduce the notion of a star-like node, then we state that all
non-star trees have at least two such nodes. Finally, we conclude that all non-star trees have at least one star-like node of
degree at most n/2.
Denition 2 (Star-like nodes). Let T be a tree, and assume n > 2. We say that u V is a star-like node of T if
(i) u is an internal node (i.e. (u; T ) > 1); and
(ii) at least (u; T ) 1 of its neighbors are leaves, i.e.,

_
w V : (w; T ) =1 and uw T
_

(u; T ) 1.
We denote the set of all star-like nodes in T by V
SL
(T ).
Proposition 22 (Number of star-like nodes). Let T be a tree, its diameter and assume n > 2. If = 2 (i.e., T is a star), then there
exists exactly one star-like node, i.e., |V
SL
(T )| =1. If > 2, then |V
SL
(T )| > 1.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 21
Fig. 3. Illustration for the proof of Proposition 22. Left: the tree T . Right: the tree T assuming v
1
/ V
SL
(T ). Here, discs stand for nodes and lines for edges.
Proof. Let be the diameter of T . Since n > 2, it follows that 2. If = 2, then T is a star, and thus its center is a
star-like node. Hence we assume that > 2. Let u
1
, u
2
V be two nodes such that d(u
1
, u
2
; T ) = . Then it follows that
u
1
and u
2
are leaves in T .
Let P(u
1
, u
2
; T ) =u
1
, v
1
, . . . , v
2
, u
2
be a simple path from u
1
to u
2
in T . Since > 2, it follows that v
1
= v
2
, and thus
there is a non-trivial path P(v
1
, v
2
; T ) from v
1
to v
2
. Let w be the node following v
1
in P(v
1
, v
2
; T ). Note that w in an
internal node in T (w might be v
2
), i.e., (w; T ) > 1. See Fig. 3 for an illustration.
Assume v
1
is not a star-like node. Then, since v
1
w V and (w; T ) > 1, there exist z = w such that (z; T ) > 1 and
v
1
z T . Recall w P(v
1
, v
2
; T ), then z / P(u
1
, u
2
; T ). Since (z; T ) > 1, let y be such that yz T and y = v
1
. Consider the
path P

= y, z, v
1
, . . . , v
2
, u
2
. From the denition of y and z, P

is a simple path in T . Since all simple paths in a tree are

shortest paths, it follows that
d(y, u
2
; T ) =2 +d(v
1
, u
2
; T ) =1 +d(u
1
, u
2
; T ) =1 +
which is a contradiction. Hence v
1
V
SL
(T ). Repeating the same argument, we obtain that also v
2
V
SL
(T ), therefore
|V
SL
(T )| > 1. 2
Proposition 22 reveals that only stars have exactly one star-like node, and all other trees have at least two star-like
nodes. We continue with a useful corollary.
Corollary 23 (Minimum degree of star-like nodes). Let T be a tree and its diameter. If > 2, then there exist u V
SL
(T ) such that
(u; T ) n/2.
Corollary 23 is an immediate consequence of Proposition 22, and of the fact that two nodes in a tree can share at most
one neighbor.
In this section we use the following notation. For any given tree T over V , let (T ) be the number of leaves in T :
(T ) =

_
z V : (z; T ) =1
_

.
Before we state and prove the main theorem of this subsection, we prove the following lemma.
Lemma 24 (Leaves and potential function). Let > 1 be given, and assume the activation process has full support. Assume T
(0)
is a tree
and let k 0 be given. Then is an ordinal potential function with respect to -local dynamics TreeNode and TreeLink. Formally,
for any realization of the activation sequence,

_
T
(k)
_

_
T
(k+1)
_
.
Proof. It suces to show that for any tree T and any active node v, v would never nd it benecial to exchange a
connection to an internal node, say w, for a connection to a leaf, say l
u
T . Formally, let w be the immediate neighbor of v
in vs path to l
u
. Let T
v
be the subtree of T rooted at v obtained by removing the edge vw. Let u be the parent node of l
u
(see Fig. 4). Note that u = v. Dene T
u
= T vw +uv and T
l
u
= T vw +l
u
v. We now prove that h(v; T
l
u
) h(v; T
u
) > 0.
Then we write
h(v; T
l
u
) h(v; T
u
) =

z / T
v
_
g
_
d(v, z; T
l
u
)
_
g
_
d(v, z; T
u
)
__
=

z / T
v
{l
u
,u}
_
g
_
d(v, z; T
l
u
)
_
g
_
d(v, z; T
u
)
__
=

z / T
v
{l
u
,u}
_
g
_
d(v, z; T
u
) +1
_
g
_
d(v, z; T
u
)
__
> 0.
This proves that if a node v is activated, it would never propose a bilateral deviation to a leaf l
u
. Thus is an ordinal
potential function. 2
22 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Fig. 4. Illustration for the proof of Lemma 24. Unprotable deviation for v. Here, discs stand for nodes, solid lines for edges and dotted lines for paths.
Our rst result establishes that, for all > 1, dynamics TreeNode and TreeLink converge almost surely when the con-
nectivity cost function satises Property 1, and that the selected outcomes are ecient. Our proof technique also yields an
exponential upper bound on the expected convergence time. To do so, we just prove that, if the current network is not
ecient, then, with at least exponentially small positive probability, the potential function decreases.
Theorem 25 (Local convergence, exponential bound). Let > 1 be given. Assume that the activation process has full support, and
that the initial graph T
(0)
is a tree. Suppose that the connectivity cost function g has bounded discrete derivatives (i.e., it satises
Property 1). Then, the following statements hold:
1. Dynamics TreeNode, hence NodeDynamics, converge almost surely.
2. Dynamics TreeLink, hence LinkDynamics, converge almost surely.
3. The selected outcomes of all dynamics are star networks, hence are ecient.
4. If the activation process is uniform, then the expected convergence time is at most exponential in n. Specically, the expected
convergence time of NodeDynamics is O(n
n/2
) and the expected convergence time LinkDynamics is O(n
n
).
Proof. Let T
(k)
be the current network topology and assume T
(k)
is not a star. For TreeNode, we provide a sequence of
nodes (or pairs of nodes) such that, if the activation process activates all these nodes before activating a node (a pair
of nodes) not in the sequence, then the potential function is increased. Similarly, for TreeLink, we provide a sequence of
pairs of nodes such that their consequent activation increases the value of .
We begin by observing that by Corollary 23 there exists a node u V
SL
(T
(k)
) such that (u; T
(k)
) n/2. Let

=
(u; T
(k)
) 1, and denote by v
i
, for 1 i

, the leaves connected to u.

Assume the following property that we will prove later on:
Property (

). Let T be a tree network and let u V

SL
(T ) such that (u; T ) n/2. Let l
u
be a leaf connected to node u,
i.e., (l
u
; T ) = 1 and l
u
u T . Then, node l
u
will deviate under TreeNode if l
u
is activated. For TreeLink, there exists a pair
-close nodes l
u
and x =u that will deviate if node l
u
is activated and x is selected as l
u
s deviation partner.
Consider TreeNode. By Property (

), if node v
1
is activated then (u; T
(k+1)
) =(u; T
(k)
) 1. Thus, at round k +1, either
(u; T
(k+1)
) =1, which implies that u is now a leaf, or u V
SL
(T
(k+1)
) and (u; T
(k+1)
) n/2.
One can see that, by induction on i, if at round k 1 + i we activate v
i
, all the v
i
s will connect to a node different
from u, and thus, after

rounds, u becomes a leaf. We therefore conclude that the potential function is increased. Since
the probability of such activation sequence is bounded from below, we conclude that TreeNode converges almost surely to
a star network.
Note that, if the activation process is uniform, the probability of such event can be O((1/n)
n/2
) and thus we can only
conclude that expected convergence time is at most exponential in n.
A similar argument applies to TreeLink. By Property (

), if node v
1
is activated and the corresponding node x is selected
as its deviation partner, then (u; T
(k+1)
) =(u; T
(k)
)1. Following the same line of reasoning as for TreeNode, we conclude
that TreeLink converges almost surely to a star network. Moreover, if the activation process is uniform, the probability of
such event can be O((1/n
2
)
n/2
), thus the expected convergence time is at most exponential in n.
The proof of the theorem is concluded by showing Property (

). Let T , u and l
u
be as dened above. We note that there
exists a unique node w T such that uw T and (w; T ) > 1. Consider the tree T

= T l
u
u +l
u
w as in Fig. 5.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 23
Fig. 5. Illustration for the proof of Theorem 25. Protable deviation. Here, discs stand for nodes and lines stand for edges.
We now prove that h(l
u
; T ) > h(l
u
; T

). First, note that l

u
is a leaf, and l
u
u T . Next, since uw T

and (u; T

)
n/2 1, it follows that the number of leaves connected to u in T

is strictly smaller than the number of nodes in T

w
minus one, where T
w
is the subtree of T rooted at w obtained by removing the link uw (similarly we dene T
u
as the
subtree of T rooted at u). Formally:

_
z V
_
T

zu T

and
_
z; T

_
=1
_

V (T
u
) \ {u}

<

V (T
w
)

1 =

V (T
w
) \ {w}

,
where V (T

) is the set of nodes in subtree T

. Equivalently, |V (T
u
)| <|V (T
w
)|. We now calculate h(l
u
; T ) h(l
u
; T

).
h(l
u
; T ) h
_
l
u
; T

_
=

zV (T
u
)
_
g
_
d(l
u
, z; T )
_
g
_
d
_
l
u
, z; T

___
+

zV (T
w
)
_
g
_
d(l
u
, z; T )
_
g
_
d
_
l
u
, z; T

___
=

zV (T
u
)\{u}
_
g
_
d(l
u
, z; T )
_
g
_
d
_
l
u
, z; T

___
+

zV (T
w
)\{w}
_
g
_
d(l
u
, z; T )
_
g
_
d
_
l
u
, z; T

___
,
since g(d(l
u
, u; T )) g(d(l
u
, u; T

)) =[g(d(l
u
, w; T )) g(d(l
u
, w; T

))].
Now note that, for all z V (T
u
) \ {u}, d(l
u
, z; T

)) =3 =1 +d(l
u
, z; T ), hence
h(l
u
; T ) h
_
l
u
; T

_
=

zV (T
u
)\{u}
_
g(2) g(3)
_
+

zV (T
w
)\{w}
_
g
_
d(l
u
, z; T )
_
g
_
d
_
l
u
, z; T

___
=
_

V (T
u
)

1
__
g(2) g(3)
_
+

zV (T
w
)\{w}
_
g
_
d(l
u
, z; T )
_
g
_
d(l
u
, z; T ) 1
__

V (T
u
)

1
__
g(2) g(3)
_
+
_

V (T
w
)

1
__
g(3) g(2)
_
=
_

V (T
w
)

V (T
u
)

__
g(3) g(2)
_
>
_
g(3) g(2)
_
> 0.
Next, notice that
h(w; T ) h
_
w; T

_
= g
_
d(w, l
u
; T )
_
g
_
d
_
w, l
u
; T

__
= g(2) g(1) > 0,
thus both l
u
and w strictly prot by deviating from T to T

. Therefore, under TreeNode the node l

u
deviates if activated.
Under TreeLink, nodes l
u
and w deviate if node l
u
is activated and node w is selected to be its deviation partner. 2
It is trivial to check that if the connectivity cost function is non-negative, convex, and strictly increasing, then all the
conditions from Theorem 25 are satised. This proves that for linear connectivity cost (Example 2), and for exponential
connectivity cost (Example 3), for all > 1, dynamics TreeNode and TreeLink converge almost surely to ecient networks.
8.2.2. Assumption on hLeaf Choices Property
We now turn our attention the second class of functions g we consider. Unlike Property 1, we now consider an aggregated
property that can best be described as a property on h and not on g.
Property 2 (Leaf Choices). Let V = V

{u}. We say that h satises the Leaf Choices Property if, for all trees T

over V

,
v, w V

, h
_
v; T

_
>h
_
w; T

_
h(u; F +uv) >h(u; F +uw),
where F is the forest
8
over V consisting of T

and {u}.
Informally, if the cost to v in T

is strictly greater than the cost of w in T

, then the cost to a new node u is strictly

greater if it connects to v than if it connects to w. Of course, this new node u would be a leaf when added to the existing
tree T

hence the name leaf choices.

We now show that both the linear connectivity function, as dened in Example 2, and the exponential connectivity
function, as dened in Example 3, satisfy the Leaf Choices Property. To do so, we use a characterization that encompasses
both connectivity functions, and then prove that all such connectivity functions satisfy Property 2.
8
We use a standard denition of forest from graph theory (see Cormen et al., 2001), i.e., disjoint set of trees.
24 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Proposition 26. Let > 0, = 1 be given. Assume the connectivity cost function g is given either by g(m) = m or by g(m) =

m1
i=0

i
. Then h satises Property 2.
Proof. Suppose there exists a constant > 0 such that, for all m > 0,
g(m+1) =1 +g(m). (15)
Note that for g linear =1, and for g exponential =.
Assume V = V

{u}, and let T

be a tree over V

, and let F be the forest over V containing T

and {u}. We calculate

h(u; F +uv) h(u; F +uw):
h(u; F +uv) h(u; F +uw) =

zV

_
g
_
d(u, z; F +uv)
_
g
_
d(u, z; F +uw)
__
=

zV

_
g
_
d(v, z; F +uv) +1
_
g
_
d(w, z; F +uw) +1
__
=

zV

_
g
_
d(v, z; F +uv)
_
g
_
d(w, z; F +uw)
__
=
_
h
_
v; T

_
h
_
w; T

__
.
Here, the third equality is obtained using 15. The result above implies that h(u; F + uv) > h(u; F + uw) if and only if
h(v; T

) >h(w; T

) since > 0. 2
The following theorem shows that if the function h satises Property 2, then TreeNode and TreeLink converge almost
surely. Moreover, if the dynamics are global then the limiting networks are ecient.
Theorem 27 (Global convergence). Let > 1 be given and let the initial graph T
(0)
be a tree. Assume that the activation process has
full support, and that the connectivity cost function satises the Leaf Choices Property (i.e. Property 2). Then, the following statements
hold:
1. TreeNode, and hence NodeDynamics, converge almost surely.
2. TreeLink, and hence LinkDynamics, converge almost surely.
3. If the dynamics are global (i.e., n 1), then the selected outcomes of all four dynamics are star networks, hence ecient.
The proof of the theorem above relies the following technical lemma:
Lemma 28. Let G =(V , E) be a tree network and let v, w V be two nodes. Then, the following two statements hold:
1. It holds that
h(v; G) h(w; G)

zV
_
g
_
d(v, z; G)
_
g
_
d(w, z; G)
__
> 0

zV
_
g
_
d(v, z; G) +1
_
g
_
d(w, z; G) +1
__
> 0.
2. For all i 1,
h(v; G) h(w; G)

zV
_
g
_
d(v, z; G)
_
g
_
d(w, z; G)
__
> 0

zV
_
g
_
d(v, z; G) +i
_
g
_
d(w, z; G) +i
__
> 0.
Proof. Statement 1. Let u / V be a new node and let G

(0) =(V {u}, E) be a forest consisting of G and {u}. See Fig. 6 for
an illustration. Note that
h
_
u; G

(0) +uv
_
h
_
u; G

(0) +uw
_

zV
g
_
d
_
u, z; G

(0) +uv
__

zV
g
_
d
_
u, z; G

(0) +uw
__
=

zV
_
g
_
d(v, z; G) +1
_
g
_
d(w, z; G) +1
__
.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 25
Fig. 6. Auxiliary graph for the proof of Statement 1 of Lemma 28. Here, discs stand for nodes.
Fig. 7. Illustration for the proof of Statement 2 of Lemma 28. Here, discs stand for nodes.
Since h satises Property 2, we conclude that
h(v; G) h(w; G) > 0 h
_
u; G

(0) +uv
_
h
_
u; G

(0) +uw
_
> 0,
which proves the result.
Statement 2. Let us now assume that h(v; G) h(w; G) > 0 and show that for every i 1 it holds that

zV
_
g
_
d(v, z; G) +i
_
g
_
d(w, z; G) +i
__
> 0. (16)
We prove the latter by induction. Let i = 1 be the basis of the induction and note that (16) holds by Statement 1. Let us
now assume that inequality (16) holds for i =k 1, where k is some integer greater than 1. We proceed to show that (16)
also holds for i =k.
Our argument uses two auxiliary graphs G

(k) and G

(k) dened in the following way. Let G

(k) be a network over the

set V

(k) of nodes and the set E

(k) of edges, where

V

(k) = V {v
i
}
k1
i=1
{w
i
}
k1
i=1
,
E

(k) = E {v
i1
v
i
}
k1
i=1
{w
i1
w
i
}
k1
i=1
.
For brevity of exposition, we alternatively refer to nodes v and w in expressions above as v
0
and w
0
respectively. Network
G

(k) is dened to be a network over the set V

(k) = V

(k) {u} of nodes and the set E

(k) = E

(k) of edges. We illustrate

these graphs in Fig. 7.
Let us consider the connectivity cost of node w
k1
in network G

(k):
h
_
w
k1
; G

(k)
_
=
k1

j=1
g
_
d
_
w
k1
, w
j
; G

(k)
__
+

zV
g
_
d
_
w
k1
, z; G

(k)
__
+
k1

j=1
g
_
d
_
w
k1
, v
j
; G

(k)
__
=
k2

j=0
g( j) +

zV
g
_
d
_
w, z; G

(k)
_
+(k 1)
_
+
k1

j=1
g
_
d
_
w, v; G

(k)
_
+(k 1) + j
_
.
Here, the second equality stems from the observations that d(w
k1
, w
j
; G

(k)) = j, that d(w

k1
, z; G

(k)) = d(w
k1
, w;
G

(k)) +d(w, w; G

(k)) and that d(w

k1
, v
j
; G

(k)) =d(w
k1
, w; G

(k)) +d(w, v; G

(k)) +d(v, v
j
; G

(k)). Similarly,
h
_
v
k1
; G

(k)
_
=
k2

j=0
g( j) +

zV
g
_
d
_
v, z; G

(k)
_
+(k 1)
_
+
k1

j=1
g
_
d
_
w, v; G

(k)
_
+(k 1) + j
_
,
26 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Fig. 8. Auxiliary graph for the proof of Statement 1 of Theorem 27. Here, discs stand for nodes and a line stands for an edge.
therefore,
h
_
v
k1
; G

(k)
_
h
_
w
k1
; G

(k)
_
=

zV
_
g
_
d(v, z; G) +(k 1)
_
g
_
d(w, z; G) +(k 1)
__
> 0,
where the last inequality is justied by the induction and by the fact that d(x, y; G

(k)) =d(x, y; G) for all x, y V .

Now consider node u. Since function h satises Property 2 and h(v
k1
; G

(k)) h(w
k1
; G

(k)) > 0, it holds that

h(u; G

(k) + uv
k1
) h(u; G

(k) + uw
k1
) > 0. Using straightforward substitution and rearranging the terms we obtain
that

zV
_
g
_
d
_
v, z; G

(k)
_
+k
_
g
_
d
_
w, z; G

(k)
_
+k
__
=h
_
u; G

(k) +uv
k1
_
h
_
u; G

(k) +uw
k1
_
> 0,
which concludes the proof. 2
We now proceed to prove Theorem 27.
Proof. Statements 1 and 2. It suces to show that there exist an ordinal potential function for the dynamics considered
(Jackson, 2008). We prove that the social connectivity cost is an ordinal potential function for TreeNode and TreeLink.
First, recall that a function F is an ordinal potential function (Jackson, 2008) if, for all realizations of the activation sequence
and for all k 0,
T
(k)
= T
(k+1)
F
_
T
(k)
_
> F
_
T
(k+1)
_
.
Let such k 0 be given. Then, there exist three distinct nodes u, v and w satisfying u
(k)
=u and T
(k+1)
= T
(k)
uv +uw.
(For TreeLink, it also holds that w = w
(k)
.) In other words, u is the active node during round k, and it successfully deviates
with w by removing uv and adding uw.
We know that the active node proposes a given deviation only if its cost strictly decreases. We conclude that h(u; T
(k)
) >
h(u; T
(k+1)
). Let T
v
be the subtree of T
(k)
rooted at v and not containing uv; note that w must be in T
v
. We let V (T

)
denote the set of nodes in a subtree T

. See Fig. 8 for illustration.

First, let us consider the difference between h(u; T
(k)
) and h(u; T
(k+1)
) h(u; T
(k)
uv +uw):
h
_
u; T
(k)
_
h
_
u; T
(k+1)
_
=

zV (T
u
)
_
g
_
d
_
u, z; T
(k)
__
g
_
d
_
u, z; T
(k+1)
___
+

zV (T
v
)
_
g
_
d
_
u, z; T
(k)
__
g
_
d
_
u, z; T
(k+1)
___
=

zV (T
v
)
_
g
_
d(v, z; T
v
) +1
_
g
_
d(w, z; T
v
) +1
__
.
By Statement 1 of Lemma 28 the inequality

zV (T
v
)
[g(d(v, z; T
v
) +1) g(d(w, z; T
v
) +1)] > 0 implies
h(v; T
v
) >h(w; T
v
). (17)
Now let us consider the difference between h(x; T
(k)
) and h(x; T
(k+1)
) h(x; T
(k)
uv +uw) for some x T
u
.
h
_
x; T
(k)
_
h
_
x; T
(k+1)
_
=

zV (T
u
)
_
g
_
d
_
x, z; T
(k)
__
g
_
d
_
x, z; T
(k+1)
___
+

zV (T
v
)
_
g
_
d
_
x, z; T
(k)
__
g
_
d
_
x, z; T
(k+1)
___
=

zV (T
v
)
_
g
_
d(x, u; T
u
) +1 +d(v, z; T
v
)
_
g
_
d(x, u; T
u
) +1 +d(w, z; T
v
)
__
> 0,
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 27
where the last inequality follows by Statement 2 of Lemma 28 for G T
v
. We thus showed that
h
_
x; T
(k)
_
h
_
x; T
(k+1)
_
> 0 for all x T
u
. (18)
Finally, consider the following expression

yG
2
[h(y; T
(k)
) h(y; T
(k+1)
)]. We note that for any y T
u
it holds that
h
_
y; T
(k)
_
h
_
y; T
(k+1)
_
=

zT
v
_
g
_
d
_
y, z; T
(k)
__
g(d
_
y, z; T
(k+1)
__
,
hence

yT
v
_
h
_
y; T
(k)
_
h
_
y; T
(k+1)
__
=

yT
v
_

zT
u
_
g
_
d
_
y, z; T
(k)
__
g
_
d
_
y, z; T
(k+1)
___
_
=

zT
u
_

yT
v
_
g
_
d
_
y, z; T
(k)
__
g
_
d
_
y, z; T
(k+1)
___
_
=

zT
u
_

yT
v
_
g
_
d
_
z, y; T
(k)
__
g
_
d
_
z, y; T
(k+1)
___
_
=

zT
u
_
h
_
z; T
(k)
_
h
_
z; T
(k+1)
__
> 0,
where the last inequality is due to (18).
We conclude the proof by considering the difference between the total connectivity cost of two networks:

_
T
(k)
_

_
T
(k+1)
_
=

xT
u
_
h
_
x; T
(k)
_
h
_
x; T
(k+1)
__
+

yT
y
_
h
_
y; T
(k)
_
h
_
y; T
(k+1)
__
> 0,
where the inequality follows from (17) and (18).
Statement 3. Consider dynamics TreeNode and TreeLink. For TreeNode, it suces to show that in any non-star network T
there exists a node x that will necessary deviate if activated. For TreeLink, it suces to show that in any non-star network T
there exists a pair of nodes x and y that will necessary deviate if x is activated and y is selected as his deviation partner.
Recall that by Proposition 22, there exist two distinct nodes u and v in V
SL
(T ). Since u V
SL
(T ), it follows that there
exists a leaf l
u
such that l
u
u T . The same holds for another leaf l
v
such that l
v
v T .
Let T

be the tree obtained from T by removing l

u
and l
v
from the vertex set (and by removing edges adjacent to them
as well). Assume, without loss of generality, that h(u; T

) h(v; T

). In what follows, we show that the connectivity cost

of both node l
v
and node u strictly decreases in deviation from T to (T l
v
v +l
v
u). Since the dynamics are global (i.e.,
l n 1), node u is in l-neighborhood of l
v
. Hence, node l
v
will necessarily deviate under TreeNode if activated. The same
argument holds for TreeLink if node l
v
is activated and node u is selected as his deviation partner.
We begin by observing that
h(u; T ) h(u; T l
v
v +l
v
u) = g
_
d(u, l
v
; T )
_
g
_
d(u, l
v
; T l
v
v +l
v
u)
_
g(2) g(1) > 0.
It thus remains to show that h(l
v
; T ) h(l
v
; T l
v
v +l
v
u) > 0.
Note that
h(v; T l
v
v) =h
_
v; T

+l
u
u
_
=h
_
v; T

_
+ g
_
d
_
v, l
u
; T

+l
u
u
__
h
_
v; T

_
+ g(2); and
h(u; T l
v
v) =h
_
u; T

+l
u
u
_
=h
_
u; T

_
+ g
_
d
_
u, l
u
; T

+l
u
u
__
=h
_
v; T

_
+ g(1),
hence h(v; T l
v
v) >h(u; T l
v
v). By Property 2, we conclude that h(l
u
; T ) >h(l
u
; T l
v
v +l
v
u).
So far we proved that a non-star network cannot be a xed point neither of TreeNode nor of TreeLink. By Statements 1
and 2 we know that these dynamics converges almost surely, hence it converges to a star network. We further note that by
Statement 3 of Theorem 20 these outcomes are ecient. 2
As a corollary, we show that if the connectivity cost function is linear and the activation process is uniform, then the
expected convergence time is polynomial in n.
Corollary 29 (Linear connectivity cost convergence time). If the connectivity cost function is linear, and the activation process is
uniform, then, for all > 1, TreeNode and TreeLink converge almost surely to ecient networks. The expected convergence time of
TreeNode is O(n
4
) and of TreeLink is O(n
5
).
28 E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129
Proof. We have already shown that TreeNode and TreeLink converge almost surely. In order to give a polynomial upper
bound on the expected convergence time, note that is a positive integer valued function. Further, note that the maximum
value can attain is upper bounded by n n
2
= n
3
. Thus the potential function can change at most O(n
3
) times. For
TreeNode, there is at least one node willing to deviate at any non-convergent state, the expected number of rounds between
deviations is at most O(n). We conclude that the dynamics converge in O(n
4
) rounds in expectation. For TreeLink, there is
at least one pair of nodes willing to deviate at any non-convergent state, the expected number of rounds between deviations
is at most O(n
2
). We conclude that the dynamics converge in O(n
5
) rounds in expectation. 2
9. Conclusion
In this work we considered a network formation game with high maintenance cost for the links and analyzed its static
equilibria and ecient outcomes. We showed that even the notion of strong stability may provide a very limited insight
to the network eciency. In particular, the corresponding price of anarchy might grow linearly in the size of the network.
This ineciency is addressed by introducing two natural and local dynamics dubbed NodeDynamics and LinkDynamics. We
focused on three subclasses of distance-based cost models and show that NodeDynamics and LinkDynamics converge almost
surely in all three subclasses of models. In two classes of models these dynamics converge to the most ecient network and
in the third class the price of anarchy of the selected outcome is shown to be bounded by a constant independent on the size
of the network. Outcomes selected by our dynamics are signicantly more ecient than strongly stable equilibria. The price
of anarchy of networks selected by our dynamics is either 1 or bounded by a constant, while the price of anarchy of strongly
stable networks may grow linearly with the size of the network. Finally, we analyzed the expected time until convergence
of our dynamics. We showed that it is polynomial for two subclasses of models and that it is at most exponential in the
third class.
The possible future research of this topic can be divided into the following three directions: First, we plan to extend
our cost model to incorporate considerations of network reliability (see, e.g., Bachrach et al., 2011) in the presence of link
failures. Namely, we assume that links maintenance cost allows for redundant links, however, each link has a probability
of failure. Such link failures may break the connectivity between some nodes, thus decreasing their expected utilities. Each
node faces tension between a desire to boost the network reliability by adding redundant links and unwillingness to pay
for unnecessary links. Second, we are interested in dynamics that allow deviation of coalitions of several nodes rather than
of a single node. These dynamics may provide faster of convergence rate and lower price of anarchy for selected outcomes.
Finally, the third research direction is to consider tree formation game in the presence of an underlying infrastructure
network, which constrains the set of possible links in the network or determining their cost.
Acknowledgments
This research was partially supported by the National Science Foundation; the Defense Advanced Research Projects
Agency under the ITMANET program; and the United StatesIsrael Binational Science Foundation (BSF), Jerusalem, Israel.
References
Altman, E., Boulogne, T., El-Azouzi, R., Jimnez, T., Wynter, L., 2006. A survey on networking games in telecommunications. Comput. Oper. Res. 33 (2),
286311.
Anshelevich, E., Dasgupta, A., Tardos, ., Wexler, T., 2003. Near-optimal network design with selsh agents. In: Proceedings of ACM Symposium on the
Theory of Computing, pp. 511520.
Arcaute, E., Johari, R., Mannor, S., 2007. Network formation: Bilateral contracting and myopic dynamics. In: Proceedings of the Third International Workshop
on Internet and Network Economics, WINE, San Diego, CA, USA, December 1214, 2007, pp. 191207.
Arcaute, E., Johari, R., Mannor, S., 2008. Local two-stage myopic dynamics for network formation games. In: Proceedings of the Fourth International Work-
shop on Internet and Network Economics, WINE, Shanghai, China, December 1720, 2008, pp. 263277.
Arcaute, E., Johari, R., Mannor, S., 2009. Network formation: Bilateral contracting and myopic dynamics. IEEE Trans. Autom. Control 54 (8), 17651778.
http://dx.doi.org/10.1109/TAC.2009.2024564.
Bachrach, Y., Meir, R., Feldman, M., Tennenholtz, M., 2011. Solving cooperative reliability games. In: Proceedings of the 27th Conference on Uncertainty in
Articial Intelligence.
Briscoe, B., Darlagiannis, V., Heckman, O., Oliver, H., Siris, V., Songhurst, D., et al., 2003. A market managed multiservice Internet (M3I). Comput. Commun. 26
(4), 404414.
Cesa-Bianchi, N., Lugosi, G., 2004. Prediction, Learning, and Games. Cambridge University Press, Cambridge, UK.
Chen, H.-L., Roughgarden, T., Valiant, G., 2008. Designing networks with good equilibria. In: Proceedings of the Nineteenth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA, 2007.
Corbo, J. Parkes, D., 2005. The price of selsh behavior in bilateral network formation, in: Proceedings of the Twenty-Fourth Annual ACM Symposium on
Principles of Distributed Computing, PODC, Las Vegas, NV, USA, July 1720, 2005, 2005, pp. 99107.
Cormen, T., Leiserson, C., Rivest, R., Stein, C., 2001. Introduction to Algorithms. MIT Press.
Demaine, E.D., Hajiaghayi, M., Mahini, H., Zadimoghaddam, M., 2007. The price of anarchy in network creation games. In: Proceedings of the Twenty-Sixth
Annual ACM Symposium on Principles of Distributed Computing, PODC 2007, Portland, OR, USA, August 1215, 2007, pp. 292298.
Eidenbenz, S., Kumar, V., Zust, S., 2003. Equilibria in topology control games for ad hoc networks. In: Proceedings of International Conference on Mobile
Computing and Networking, pp. 211.
Fabrikant, A., Luthra, A., Maneva, E.N., Papadimitriou, C.H., Shenker, S., 2003. On a network creation game. In: Proceedings of the Twenty-Second ACM
Symposium on Principles of Distributed Computing, PODC 2003, Boston, MA, USA, pp. 347351.
Falkner, M., Devetsikiotis, M., Lambadaris, I., 2000. An overview of pricing concepts for broadband IP networks. IEEE Commun. Surv. Tutor. 3 (2), 213.
E. Arcaute et al. / Games and Economic Behavior 79 (2013) 129 29
Fiat, A., Kaplan, H., Levy, M., Olonetsky, S., Shabo, R., 2006. On the price of stability for designing undirected networks with fair cost allocations. In: ICALP,
vol. 1, pp. 608618.
Flajolet, P., Sedgewick, R., 2009. Analytic Combinatorics. Cambridge University Press.
Fudenberg, D., Levine, D.K., 1998. The Theory of Learning in Games. MIT Press, Cambridge, MA.
Goemans, M., Mirrokni, V., Vetta, A., 2005. Sink equilibria and convergence. In: 46th Annual IEEE Symposium Foundations of Computer Science, FOCS 2005.
IEEE, pp. 142151.
Goyal, S., 2007. Connections: An Introduction to the Economics of Networks. Princeton University Press.
Jackson, M.O., 2003. A survey of models of network formation: Stability and eciency. Division of the Humanities and Social Sciences. Working paper 1161,
California Institute of Technology. Available at http://ideas.repec.org/p/clt/sswopa/1161.html.
Jackson, M.O., 2008. Social and Economic Networks. Princeton University Press, Princeton, NJ, USA, ISBN 978-0691134406.
Jackson, M.O., Watts, A., 2001. The existence of pairwise stable networks. Seoul J. Econ. 14 (3), 299321.
Jackson, M.O., Watts, A., 2002. The evolution of social and economic networks. J. Econ. Theory 106 (2), 265295.
Jackson, M.O., Wolinsky, A., 1996. A strategic model of social and economic networks. J. Econ. Theory 71 (1), 4474. Available at http://ideas.repec.org/a/
eee/jetheo/v71y1996i1p44-74.html.
Johari, R., Mannor, S., Tsitsiklis, J.N., 2006. A contract-based model for directed network formation. Games Econ. Behav. 56 (2), 201224. Available at
http://ideas.repec.org/a/eee/gamebe/v56y2006i2p201-224.html.
Komali, R., MacKenzie, A., 2006. Distributed topology control in ad hoc networks: A game theoretic perspective. In: Proceedings of IEEE CCNC, pp. 563568.
Ma, J., 1995. Stable matchings and rematching-proof equilibria in a two-sided matching market. J. Econ. Theory 66 (2), 352369.
Ma, J., 2002. Stable matchings and the small core in Nash equilibrium in the college admissions problem. Rev. Econ. Design 7 (2), 117134.
Mannor, S., Shamma, J., 2007. Multi-agent learning for engineers. In: Special Issue on Foundations of Multi-Agent Learning. Articial Intelligence, 417422.
Myerson, R.B., 1977. Graphs and cooperation in games. Math. Oper. Res. 2 (3), 225229.
Papadimitriou, C.H., 2001. Algorithms, games, and the internet. In: Proceedings on 33rd Annual ACM Symposium on Theory of Computing, Heraklion, Crete,
Greece, July 68, 2001, pp. 749753.
Roughgarden, T., 2005. Selsh Routing and the Price of Anarchy. MIT Press.
Vega-Redondo, F., 2007. Complex Social Networks. Cambridge University Press, New York.