A Model for Social Network Formation:

Efficiency, Stability and Dynamics

L. Elisa Celis, École Polytechnique Fédérale de Lausanne
Aida Sadat Mousavifar, University of Tehran

We introduce a simple network formation model for social networks. Agents are nodes, connecting to another agent by
building a directed edge (or accepting a connection from another agent) has a cost, and reaching (or being reached by) other
agents via short directed paths has a benefit; in effect, an agent wants to reach others quickly, but without the cost of directly
connecting each and every one. We prove that asynchronous edge dynamics always converge to a stable network; in fact, for
nontrivial ranges of parameters this convergence is fast. Moreover, the set of fixed points of the dynamics form a nontrivial
arXiv:1510.09025v2 [cs.SI] 21 Apr 2016

class of networks. For the static game, we give classes of efficient networks for nontrivial parameter ranges and further study
their stability. We close several problems, and leave many interesting ones open.

1. INTRODUCTION
Online social networks such as Facebook and Twitter are now an ubiquitous part of modern life.
Moreover, given the prevalence of economic situations in which the network of relationships be-
tween agents play an important role in outcomes, it is essential to rigorously understand how net-
works form and what network structures are likely to emerge. Large interdisciplinary subfields that
combine economics, sociology, mathematics and computer science in the study of social networks
are emerging (see [Carrington et al. 2005] for a survey). While many models for social network
exist, most are either stochastic (i.e., probabilistic models) or are learned models (i.e., constructed
by fitting a set of parameters). The game theoretic approaches to network formation that exist are
largely motivated by games where network infrastructure is being built and costs are shared amongst
agents (see, e.g., [Nisan et al. 2007] Chapter 19), and do not necessarily capture natural properties
of online social networks. We introduce a simple directed network model that has a natural interpre-
tation with respect to many online social networks. In this model the agents are nodes in the network
and the model is defined by three key parameters:
(1) the cost cs of directly connecting to another agent (i.e., making a friend request),
(2) the cost c` of accepting a connection another agent (i.e., confirming a friend request), and
(3) the distance k (i.e., the maximum path length) that suffices for gaining utility from an indirect
connection to another agent.
Agents trade off decisions between the cost of maintaining edges against the rewards (in terms of
connectivity) from doing so. Allowing cs , c` > 0 captures many online social networks such as
Facebook and LinkedIn in which one agent initiates a connection request and the other choses to
accept or decline. When c` = 0, the model captures other online social networks such as Twitter in
which a connection can be made unilaterally. The distance k captures the maximum path distance
that suffices for deriving utility from (indirect) connections; a generalization of this model can fur-
ther allow target sets T(v) which defines the set of agents that v would like to reach within distance
k.1
In particular, we study natural dynamics in which agents periodically make asynchronous deci-
sions on whether to add or sever edges. We show that irrespective of the initial network, the dynamic
process converges; in fact when cs , c` > 0 the convergence is fast. Furthermore, the fixed points
of this process form a nontrivial class of networks.2 We further study the static game, and prove
that for a nontrivial range of parameters a flower graph is efficient and stable, and a Kautz graph
is symmetric-efficient. We leave open the technically challenging question of whether symmetric
networks that are also stable exist for nontrivial parameter ranges. Finally, we study the extremal

1 See Section 7 for a discussion – in this paper we only consider T (v) = V for all agents v.
2 This is in stark contrast to related settings in which the only fixedpoints are cycles and empty graphs [Bala and Goyal 2000].
properties of stable networks and show that the price of anarchy can be arbitrarily bad, although the
price of stability is 1.
Lastly, this model allows us to define a multi-dimensional generalized clustering coefficient which
relates to the stability of the network; this could be of independent interest in the study of social
networks [Wasserman and Faust 1994].
2. PRELIMINARIES
Let V be a set of agents with n = |V |.
D EFINITION 1 (B IDIRECTED N ETWORK). A bidirected network G = (V, Es , E` ) is a graph
with vertex set V with two types of directed edges; speaking edges suv ∈ Es ,3 and listening edges
`vu ∈ E` .4
Note that suv exists independently of `uv ; both, one, or none may be present in the network. When
clear from context, with some abuse of notation, we drop the s/` demarkations and simply refer to
edges uv ∈ E. We let ds+ G (v) = |{w|svw ∈ Es }| denote the number of outgoing speaking edges
−
of v, and let dsG (v) = |{w|swv ∈ Es }| denote the number of incoming speaking edges to v. The
−
analogous definition is used for listening out-degree (d`+G (v)) and in-degree (d`G (v)).
D EFINITION 2 (S PEAKING AND L ISTENING R EACHABILITY). We say there is a speaking
path of length k from v to u if there exists a set of directed edges svv1 , sv1 v2 , . . . , svk u ∈ Es and
edges `uvk , `vk vk−1 , . . . , `v1 v ∈ E` . We say that a vertex u that has a speaking path of length at
s
most k from v is k-speaking-reachable from v, and let RG,k (v) ⊆ V be the set of all such vertices.
`
Listening paths, listening reachability, and the set of listening-reachable vertices RG,k (v) ⊆ V are
defined in an analogous manner.
With some abuse of notation, when k is clear from context we drop it from the notation above. Note
that if u is speaking-reachable from v then v is listening-reachable from u.
2.1. Our Model
Each agent v ∈ V has a strategy Sv = (Svs , Sv` ), which consists of subsets of agents Svs , Sv` ⊆ V .
Thinking of agents as vertices, Svs (respectively Sv` ) corresponds to the set of vertices that v connects
to by building speaking (respectively listening) edges svu (respectively `vu ). Thus, the strategy
vector S = (S1 , . . . , Sn ) defines a bidirected graph G = (V, Es , E` ) where Es = {vu|u ∈ Svs }
and E` = {vu|u ∈ Sv` }. With some abuse of notation, we often refer to G as the set of strategies
and will use G and S interchangeably.
The utility of v is given by
s `
UG (v) = UG,k (v) + UG,k (v)
where
s
UG,k s
(v) = |RG,k (v)| − cs · ds+
G (v)| and
`
UG,k `
(v) = |RG,k (v)| − c` · d`+
G (v)
are the utilities derived from speaking and listening respectively. The costs cs and c` capture the
cost of maintaining speaking and listening edges respectively.
A natural special case is that in which one of the costs is 0 (without loss of generality c` = 0). For
such a model, an agent can always set Sv` = V without loss to her utility. Hence, the strategy space
boils down to Svs . Moreover, we can consider only the speaking portion of the utility UG,k (v) =
s
UG,k (v) without loss of generality. In such cases we drop all s/` demarkations and simply think of
a directed graph G = (V, E) = (V, Es ). This special case, when we further assume that k = ∞, is
equivalent to the network model in [Bala and Goyal 2000].

3 An edge suv can be thought of as u initiating contact with v.

4A listening edge `vu can be thought of as v accepting contact from u.
Fig. 1. Example network depicting addable edges (green), removable edges (red) and existing edges that are not removable
(black) for k = 2, c = cs = 1.5 and c` = 0; note that we only depict speaking edges (one can assume that all listening
edges are present).

2.2. Dynamics
In this paper we only consider dynamics that are asynchronous (i.e., one agent updates at a time)
and stochastic (i.e., agents update in a random order). A shorthand notation for the network obtained
by adding (alternatively, deleting) the edge vw from an existing network G is G + vw (alternatively,
G − vw). Similarly, we let G + Sv be the network obtained by adding all vu edges where u ∈ Sv
to G. The following definition is convenient for a variety of our definitions and results.
D EFINITION 3 (A DDABLE AND R EMOVABLE E DGES). We say an edge uv ∈ G is removable
if UG−uv (u) > UG (u). Similarly, we say an edge uv 6∈ G is addable if UG+uv (u) > UG (u).
See Figure 2.2 for an example of a network where the addable and removable edges are depicted.
We can then define the edge dynamics as follows:
D EFINITION 4 (E DGE DYNAMICS). In each round, one potential (speaking or listening) edge
vw is selected at random. Without loss of generality, assume it is a speaking edge. If svw ∈ Es then
the edge is deleted if and only if it is removable. Alternatively, if svw ∈
/ Es then the edge is added if
and only if it is addable. The analogous definition is used for listening edges.
The bulk of our convergence results concern edge dynamics. In some cases they also apply to a
natural vertex dynamics (see Definition 25).
R EMARK 5. Note that we do not consider best-response dynamics in which, in each time step,
a vertex v is selected at random and she updates update her strategy Sv in order to maximize her
utility (potentially by changing multiple edges simultaneously) with respect to the current strategies
S−v of the other agents. Indeed, our results do not extend naturally to this setting. For some special
cases of our model, if the dynamics are performed synchronously, the results from [Bala and Goyal
2000] apply.
2.3. Stability and Efficiency
Denote by S−v by the (n − 1)-dimensional vector of the strategies played by all agents other than
v. With some abuse of notation we use (Sv , S−v ) and S and G interchangeably as is convenient.
D EFINITION 6 (S TABILITY). A strategy vector S is said to be stable if for all agents v and each
potential strategy Sv0 ⊆ V , we have that
USv ,S−v (v) ≥ USv0 ,S−v (v).
This is equivalent to saying that S is a Nash equilibrium.
In other words, no agent v has any incentive to change her strategy from Sv to Sv0 , assuming that
all other agents stick to their current strategies. Observe that such a solution is self-enforcing in the
sense that once the agents are playing such a solution, no one has any incentive to deviate. In fact,
for our model, something stronger holds:
P ROPOSITION 7. A strategy vector is stable if and only if no edge is addable or removable.
Pairwise stability is a common strengthening of the notion of stability. It is natural in social networks
where, effectively, a link between two agents is formed only if both endpoints are in agreement, but
either can unilaterally delete an edge. In our model, an agent’s utility is never decreased by an
incoming edge, hence there is no difference between stability and pairwise stability. However, in
the bidirected case (when cs , c` > 0), an extended notion of pairwise stability where a speaking
edge and its corresponding listening edge are consider in conjunction is natural.
D EFINITION 8 (B IDIRECTED PAIRWISE S TABILITY). A strategy vector S is said to be bi-
pairwise stable if for all pairs of agents u, v
US−suv (u) ≤ US (u), US−`uv (u) ≤ US (u),
and if
US+suv +`vu (u) > US (u) then US+suv +`vu (v) < US (v).
For the remainder of this paper we refer to this notion as pairwise stability.
Often, notion of fairness or global optimality are important considerations. The utilitarian objec-
tive welfare of a P
set of strategies is the collective utility of all of the agents; i.e., for the strategy set
G, it is ϕ(G) = v UG (v).
D EFINITION 9 (E FFICIENCY AND S YMMETRY). We say a set of strategies S, and the network
G it defines, is efficient if it maximizes ϕ(G). It is symmetric if UG (v) = UG (u) for all u, v ∈ V , and
is asymmetric otherwise. It is symmetric-efficient if it maximizes ϕ(G) over to the set of symmetric
networks.
Note that, a priori, efficiency, symmetry, and stability need not be satisfied simultaneously. One aim
of this work is to explore these relationships for our model.
2.4. Preliminary Observations
Before stating our main results, we make a few preliminary observations which will become useful
in later proofs.
L EMMA 10. If G is bi-pairwise stable then G is stable.
P ROOF. For sake of contradiction, assume that G is pairwise stable but is not stable. Since G
is pairwise stable so for any speaking edge svw ∈ Es we have UG (v) ≥ UG−svw (v) and for any
listening edge `vw ∈ E` we have UG (v) ≥ UG−`vw (v). Hence, G has no removable edge. And
because we assume that G is not stable so there must exist a speaking edge svw 6∈ Es where
UG (v) < UG+svw or a listening edge `vw 6∈ E` where UG (v) < UG+`vw (v) which contradicts the
definition of pairwise stability.
D EFINITION 11 (C OMPLETE E DGES). With some abuse of notation, we say a speaking edge
svw is complete if
svw ∈ Es ⇒ `wv ∈ E`
We say a listening edge `vw is complete if
`vw ∈ E` ⇒ swv ∈ Es .
In any stable or efficient network, if cs , c` > 0, all edges are complete. Hence, despite allowing
unilateral actions, agreement naturally emerges.
L EMMA 12. If cs > 0, c` > 0, then for any stable, pairwise stable, or efficient network G all
speaking and listening edges are complete.
P ROOF. We will prove the contrapositive. Assume that there exist vertices v, w ∈ V such that
s
svw ∈ Es and `wv ∈ / E` . Since `wv ∈ / E` , if for any node z we have z ∈ RG,k (v), then, by the
definition of reachability, there exists a path from v to z that does not use edge svw . Thus there is no
 s s s s
node z such that z ∈ RG,k (v) and z 6∈ RG−s vw ,k
(v). In other words |RG−s vw ,k
(v)| = |RG,k (v)|,
+ +
and the positive component of the utility is the same for v. However, dsG−svw (v) = dsG (v) − 1, so
the negative component of the utility is reduced by cs in G − svw . Thus, UG−svw (v) = UG (v) + cs
and, since cs > 0, we conclude that UG−svw (v) > UG (v). Hence, G is could not be stable or
pairwise stable. Moreover, deleting svw does not affect on the utility of other vertices so ϕ(G −
svw ) > ϕ(G). Hence G is not efficient. The proof follows analogously for `vw .

3. RELATED WORK
Due to the vast range of applications; from sociology to commerce, biology and physics, with
drastically different underlying properties, many models have been developed and studied in depth
(see [Newman 2003] for a survey). Starting with G(n, p) [Bollobás 2001; Erdos and Rényi 1960],
stochastic models have often taken a forefront. Depending on the observed graph properties, differ-
ent models take the forefront, such as preferential attachment models [Barabási and Albert 1999]
for specific degree distributions, or small-worlds models [Watts and Strogatz 1998] for capturing
social networks. An alternate approach, is to take an existing network and fit a model using tech-
niques from machine learning. For example, a the authors of [Karandikar et al. 2008] attempt to
understand the Twitter network by fitting a stochastic model. However, while stochastic and learned
models can explain on a macro level what is occurring in a network, on a micro level, i.e., looking at
individual nodes and its edges, they remain uninformative; the motivation as to why a node would
maintain an edge is abstracted away. We instead consider game theoretic models of a network in
which each node is a selfish agent and decided if and whom to connect to based on her utility.
There has been a lot of very interesting work on network formation (see [Jackson 2005] and
[Nisan et al. 2007] Chapter 19 for nice surveys coming out of the Economics and Algorithmic
Game Theory literature respectively). Myerson was the first to consider such models (see, e.g.,
[Aumann and Myerson 1988; Myerson 1977]). However, formulated the problem as a cooperative
game where agents worked together towards a common goal; in our setting we assume agents have
individual, or selfish, goals. In another line of work, global connection games (see [Nisan et al.
2007]) are studied in which agents are not nodes in the network, rather vested parties in (individual)
global connectivity properties of the game.
More closely related models consider selfish agents that are nodes in undirected networks. [Jack-
son and Wolinsky 1996] introduced a model for the study of the (static) stability of undirected net-
works. Also known as the local connection game (see [Nisan et al. 2007]), nodes have discounted
(based on path length) rewards for being connected to another agent, and cost for making a link.
Their goal was to understand the relationship between stability and efficiency, which led to further
results in this direction (see, e.g., [Dutta and Mutuswami 1997; Jackson 2003]). The authors of
[Watts 2003] consider edge dynamics for this undirected model. This is further studied by [Jackson
and Watts 2002] where it is shown that the stochastic best response dynamics may not converge;
this is in contrast to our model which will always converge to a stable network.
Directed networks allow one individual to connect to another without the consent of the sec-
ond individual, and thus applications are to settings such as Twitter following, while undirected
network capture social networks where links are reciprocal, such as Facebook friendship. The dif-
ference between directed and undirected graphs is not just a technicality when it comes to modeling.
In undirected networks, edges are implicitly reciprocal, hence consent is required from both end-
points; thus, undirected models are suitable for many forms of economic and social relationships.
For directed networks, however, a vertex can link directly to another without reciprocally; thus,
directed models are more suitable for capturing interactions that are passive in one direction, as
with the consumption of public content. Not only are the applications and modeling considerations
distinct, but they can lead to dramatic differences in the resulting graphs, their properties, and their
efficiency.
The closest related work to our setting, by Bala and Goyal [Bala and Goyal 2000], studies the
stability, efficiency, and dynamics of directed networks. In fact, their can be viewed as a special
 case of ours when c` = 0 and k = ∞. However, their dynamics differ significantly from ours; they
use lazy simultaneous best-response dynamics while we consider asynchronous stochastic edge
dynamics. Due to the nature of their update, their process always converges to either a cycle or the
empty network. Not only are our dynamics more natural because coordination is not required and
connections are evaluated on an individual basis, but the class of networks that are fixed points of
our dynamics is nontrivial.

4. DYNAMICS
In studying the dynamics, different approaches are required for the special case where c` = 0
(i.e., the graph is directed) and more general bidirected setting (where cs , c` > 0). In the latter, an
agent’s hand is often forced; as reachability can only occur via paths of complete edges, u never has
incentive to add an edge suv if `vu is not present (a layman’s interpretation is to say that one cannot
accept a connection that is not initiated). This, in effect, speeds convergence and makes our work
easier. We consider both settings in Sections 4.1 and 4.2 respectively.

4.1. The Directed Setting (c` = 0)

In this section we consider the setting where c` = 0. Recall that in this case we can assume S` = V
and the model reduces to that for a directed graph with only speaking edges. Moreover, for the
directed setting, our results only hold for k = ∞. Hence, for ease of notation we drop all s, ` and k
s
demarkations, e.g., G = (V, E) = (V, Es ), c = cs , UG = UG,k , and uv = suv .
T HEOREM 13. The edge dynamics for k = ∞ converge to a stable graph.
It is a priori not clear that any dynamics should converge, or that any interesting structure could
emerge at a fixed point of the dynamics.5 On the other hand, as is evident in our proof, a nontrivial
class of networks are fixedpoints of the edge dynamics. The entire proof of convergence is rather
technical and requires an in-depth case analysis which we have split out into individual lemmas for
the reader’s clarity.
P ROOF OF T HEOREM 13. In order to prove convergence, it suffices to show that for any initial
graph G1 , there exists a finite sequence of graphs G1 , . . . , Gs such that Gi 6= Gj for any i 6= j and
Gi and Gi+1 differ by a single edge that is either addable or removable in Gi such that Gs is stable.
In other words, there is a path to convergence from any state which only uses addable or removable
edges. If this path is finite, there is some (albeit tiny) probability, which is bounded away from 0,
that this sequence of edges is selected. Hence, it suffices to construct such a sequence. We construct
it as follows:
(1) Recursively delete removable edges one at a time until no removable edge remains.
(2) If there is no addable edge then we have reached a stable graph.
(3) Otherwise, partition the graph into maximal strongly connected components; the component-
level graph must be a directed acyclic graph (Lemma 16). We call a component a root if it
has no incoming edge, and a leaf if it has no outgoing edge. It is often convenient for us to
restrict to the components of size > c, which we call large-components; roots and leaves are then
defined within the large-component graph. Note that there must be at least one large component,
otherwise we contradict the fact that an edge is addable (Lemma 17).
(4) For each root Ti of the large-component graph, designate a special vertex ri . Note that for any
(large or small) component C such that Ti 6⊆ RG (C), the edge xri is addable for any x ∈ C
(Lemma 18); hence, the edges specified in steps 5-7 must be addable.
(5) If some leaf Li 6= Ti exists in the large-component graph rooted at Ti , add the edge `i ri for some
`i ∈ Li and go back to step 1.

5 E.g., recall that the authors of [Bala and Goyal 2000] find that synchronous best-response dynamics must converge to either
the cycle or the empty graph under the same conditions.
(6) Otherwise, all large-components have no edges between them in the large-component graph. If
there are more than one large components, let T1 , T2 be any two large-components such that
T1 6⊆ RG (T2 ). Add the edge r2 r1 . If T2 6⊆ RG+r2 r1 (T1 ), also add the edge r1 r2 . Go back to
step 1.
(7) Otherwise, there is exactly one large-component T1 . If there exists a small component Sj that is
a leaf, add the edge sj r1 for some sj ∈ Sj and go back to step 1.
(8) Again, there is exactly one large-component T1 , and this set T1 is reachable from every compo-
nent (otherwise we would be in step 7). Moreover, there must be at least one small component
Sk that is a root and in which |RG (Sk ) \ T1 | > c (Lemma 22). Let tk rk be an edge that reaches
T1 on a path from Sk , i.e., rk ∈ T1 and tk 6∈ T1 , and tk ∈ RG (Sk ). Add the edge rk sk for some
sk ∈ Sk and go back to step 1.
We prove in Lemma 15 that the only addable edges are between two different components; hence
only inter-component edges must be addable; the above steps cover all possible types of addable
edges in sequence until none remain.
After step 5+1 (alternatively 6+1), the number of large components that existed before step 5
(alternatively 6) is reduced by at least one (see Lemma 19 and Lemma 20 respectively). Hence, we
will reach step 7 in finitely many rounds.
In step 7 there is exactly one large component (see also Lemma 17), and the number components
without a direct edge to T1 is reduced by one. Moreover, after step 7 there are no removable edges,
so the number of large components and the size of the largest component does not change in step
7+1 (Lemma 21). Thus, we never go back to steps 5 or 6, and after finitely many rounds, we will
move on to step 8.
In step 8 again there is exactly one large component. In step 8+1 the the number of large com-
ponents does not increase, thus, we never go back to step 5 or 6. While we may go back to step 7,
however, every time we complete step 8 we have removed at least one edge from being addable at
any point in the future (Lemma 14). Hence, we can remain in steps 7 and 8 for finitely many rounds,
and the process will terminate in step 2.
L EMMA 14. After step 8, we removed at least one edge from being addable at any point in the
future. Moreover, the number of large components does not increase.
P ROOF. By the same argument as in the proof of Lemma 19, we know that step 8 + 1 cannot
increase the number of large components.
We prove that the edge rk sk cannot be added again in the future. Let T1 be the largest component
in G (before step 8) and let T100 be the largest component in G00 (after step 8+1). We reserve T10 and
G0 for an intermediary graphs.
It now suffices to argue that T100 ∈ RG00 (sk ); hence either sk ∈ T100 , or rk sk was removed in step 1
(after we added it in step 8), in which case |RG00 (sk ) \ T100 | < c. Since steps 7 and 8 cannot increase
RG00 (sk ) for any future G00 , the edge rk sk will never become addable again.
For sake of contradiction, assume T100 6∈ RG00 (sk ). Let xy be the first removable edge such
that T10 6⊆ RG0 −xy (sk ); hence, xy must have been on every path from sk to T10 , and in particular
x ∈ RG0 (sk ) and T1 6⊆ RG0 −xy (x). However, |T1 | > c, hence xy is not removable which gives a
contradiction.
L EMMA 15. If vw is an addable edge in G, then v and w belong to different strongly connected
components of G.
P ROOF. An edge vw is addable if UG+vw (v) > UG (v), in other words |RG+vw (v) − RG (v)| ≥
c. Assume that v and w belong to same strongly connected component. Thus RG (v) = RG (w), and
hence RG+vw (v) = RG (v). Therefore the positive component of the utility for v does not increase,
while the negative component would decrease. Hence no such vw edge is addable.
L EMMA 16. In step 3, the component-level graph is a directed acyclic graph.
 P ROOF. Assume for sake of contradiction, that some set of two or more components form a
cycle in the component-level graph. Then, for every x, y in the cycle, there must be some path from
x to y as there is a path between any two vertices within every component, and a path between
any two components via the cycle. Hence, the cycle in fact forms a strongly connected component
contradicting the maximality of the partition.
L EMMA 17. Any leaf must either be an isolated vertex (with no parent) or is of size at least c;
in particular a leaf that is not a root must be of size at least c. Moreover, in step 3, there exists a
component of size at least c.
P ROOF. If there is an edge ab such that |RG (b)| < c, then by removing ab, the positive compo-
nent of the utility decreases by less than c while the negative component decreases by c. Hence, ab
is removable, and would have been eliminated in step 1. Thus, any edge ab must have |RG (b)| ≥ c.
Consider a component B, and consider any edge ab such that b ∈ B. If some such edge exists, then
RG (b) = RG (B), and hence |RG (B)| ≥ c as shown above. If no such edge exists, then B must be
an isolated vertex.
In step 3, there exists some edge ab that is addable. From Lemma 15, we know that a and b must be
in different components A, B. By assumption, we know that |B| < c, however, |RG (b) \ RG (a)| ≥
c. Thus, there must be at least one other component that is reachable from B. Thus, the component
graph has at least one leaf, and, from above, that leaf must be of size at least c. Thus, there exists at
least one component of size at least c.
L EMMA 18. In step 4, given any component C such that Ti 6⊆ RG (C), for any x ∈ C the edge
xri is addable.
P ROOF. Let Ti be a root of the large-component graph; trivially |Ti | ≥ c. If Ti is not reachable
from a component C, then, for any x ∈ C we know |RG+xri (x) \ RG (x)| ≥ |RG (ri ) \ RG (x)| ≥
|Ti | ≥ c. Thus, xri is addable.
L EMMA 19. After adding an edge as in step 5 and then completing step 1, the number of large
components at the beginning of step 5 has reduced by at least 1.
P ROOF. Clearly, adding edges cannot increase the number of components. Moreover, since Li is
a leaf of Ti , and Li and Ti there is clearly a path from every ri ∈ Ti to every xi ∈ Li in the original
graph. Using the new edge `i ri , there is now also a path from every xi ∈ Li to every ri ∈ Ti . Hence
this reduces the number of large components by at least one.
Let G0 be the current graph (after adding `i ri , and removing an unknown number of edges). We
must now show that, after deleting an edge, the number of large components has not increased.
Let xy be a removable edge in G0 and let X be the component that contains x. If y 6∈ X, then
removing xy does not increase the number of components of G0 . Assume y ∈ X. Let S be the set
of vertices that are reachable from x only through xy. Thus, S ∩ X and X \ S will form two new
strongly connected components after the removal of xy. Because xy is removable, we know that
|RG0 −xy (x) \ RG0 (x)| < c. Note that we have exactly defined S = RG0 −xy (x) \ RG0 (x). Hence,
|S ∩ X| ≤ |S| < c. Thus we have not increased the number of large connected components in G0 .
Thus, overall, the number of large components reduces by at least one.
L EMMA 20. After adding an edge as in step 6 and then completing step 1, the number of large
components at the beginning of step 6 reduces by at least one.
P ROOF. Clearly, adding edges cannot increase the number of components. Moreover, after
adding r2 r1 (and if necessary r1 r2 ), the two large components T1 , T2 are in the same large compo-
nent in the new graph. Hence, the number of large components has reduced by at least one.
By the same argument as in the proof of Lemma 19, we see that after deleting edges, the number
of large components does not increase, and hence, overall, the number of large components has
reduced by at least one.
 L EMMA 21. After adding an edge as in step 7 and then completing step 1, the number of large
components at the beginning of step 7 does not increase and the number of components without a
direct edge to T1 decreases by one.
P ROOF. Note that, at the beginning of step 7 there is a single large component (otherwise we
would be in step 5 or 6). Clearly, adding the edge reduces the number of components without a
direct edge to T1 , and adding edges cannot increase the number of components or create more roots.
We now show that no edge is removable when we go to step 1, everything that holds after step 7
must also hold after step 7+1.
Let G be the graph at the beginning of step 7. For sake of contradiction, assume that xy is
a removable edge; thus, |RG+sj r1 −xy (x) \ RG+sj r1 (x)| < c. Since sj r1 was addable in G, we
know that xy 6= sj r1 . Additionally, since sj was an isolated vertex in G we know that it has no
other edge and hence x 6= sj . Moreover, since sj was an isolated vertex in G, we know it has no
incoming edge, and hence x 6∈ RG+sjr1 (z) for any z 6= x. Thus, RG+sj r1 −xy (x) = RG−xy (x) and
RG+sj r1 (x) = RG (x), and this implies that |RG−xy (x)\RG (x)| < c, and hence xy was removable
in G. This gives a contradiction, as no such xy exists at the beginning of step 7 (as it would have
been removed in the previous step 1).
L EMMA 22. There is exactly one large-component T1 , this set T1 is reachable from every
component, and there must be at least one small component Sk with no incoming edge such that
|RG (Sk ) \ T1 | > c.
P ROOF. Exactly one large-component T1 , since at least one large component exists (Lemma 17)
and if two or more exists we would be in step 5 or 6. Moreover, T1 is reachable from every compo-
nent, otherwise we would be in step 7. Lastly, we know that there must be some small component
Sk such that |RG (Sk ) \ T1 | > c, otherwise no edge addable. In particular, any rsk edge with r ∈ T1
and sk ∈ Sk is addable.
One important question that remains open is the time until convergence; in particular, we con-
jecture that the convergence time is fast. In effect, this is equivalent to showing that there are many
short paths to convergence. Proving this, however, remains a challenging technical open problem.
4.2. The Bidirected Setting (cs , c` > 0)
We now go back to the bidirected setting. As observed above, the convergence is less surprising;
as edges that are not complete are easily deleted. This is formalized in the following lemma, which
will be useful for our result:
L EMMA 23. If cs , c` > 0 and there exist vertices v, w ∈ V such that svw ∈
/ Es , `wv ∈
/ E` ,
then the edge and vertex dynamics will not add either edge back.
P ROOF. We will prove the contrapositive. Assume that some such svw is added back. We know
that `wv ∈ / E` . Thus, by the definition of reachability, there is no path from v to any arbitrary
s k
node z which passes through svw . Hence, z ∈ RG,k (v) if and only if z ∈ SG+s vw
(v). Therefore,
k k + +
|SG (v)| = |SG+s vw
(v)|. However, ds G+svw (v) = dsG (v) + 1 so UG+svw (v) = U G (v) − cs and
since cs > 0 thus UG+svw (v) < UG (v). Therefore agent v does not gain by adding svw and wo not
add it for ever. The similar proof used for listening edges.
Using this lemma, we give a very different proof than the one for the directed setting. In addition
to working for arbitrary k, this proof also allows us to conclude that the convergence is fast.
T HEOREM 24. If cs , c` > 0, then the edge dynamics, vertex dynamics (see Definition 25) con-
verge to a stable graph in (expected) polynomial time in the number of nodes.
P ROOF. We will prove the theorem for edge dynamics; the analogous proof for vertex dynamics
follows from Lemma 26. Let pG be the number of complete edges (see Definition 11) and let mG
be the number of edges in network G. We will prove the theorem by induction on pG + mG .
 Base Case:. Suppose that pG + mG = 0, thus pG = 0, mG = 0 so G is empty and by Lemma
34, the empty graph is stable.

Inductive step:. For sake of contradiction, assume that the dynamics do not converge to a stable
graph. Thus, since the number of possible networks is finite, there must exist a sequence of networks
that we cycle over. If there exists v, w ∈ V such that svw ∈ Es and `wv ∈ / E` , then with probability
1 1 1
at least · · , the edge svw is chosen. By Lemma 26, this edge is removable for v. Therefore,
2 n n−1
the number of edges is reduced, so pG−svw + mG−svw = pG + mG − 1. Thus by the induction
hypothesis, G − svw converges to a stable network. Similarly, if svw ∈ / Es and `wv ∈ E` an
analogous proof follows.
Otherwise, for all v, w ∈ V , if svw ∈ Es then `wv ∈ E` and if `vw ∈ E` then swv ∈ Es . Since G
is not stable, there must exist an edge which addable or removable. By Lemma 23 we know that no
removed edge will be added again. Hence, the only possible change is to remove an edge. Suppose
that svw is an existing removable edge for v. The probability of selecting svw and `wv consecutively
 2
1 1 1
is at least · · . Moreover, we know that if svw and `wv are selected in sequence then
2 n n−1
both will be removed as svw is removable, and then `wv will be removable by Lemma 26. Thus, for
new graph we have that pG−svw −`vw + mG−svw −`vw = pG + mG − 3.
The convergence happens in expected polynomial time because the probabilities, as computed
above, shows that after O(n4 ) moves in expectation, the value of pG + mG will decrease. We know
that pG + mG = O(n2 ). Thus, after at most O(n6 ) moves in expectation we will reach a stable
network.

D EFINITION 25 (V ERTEX DYNAMICS). In each round, a vertex v is selected at random, and

two options are selected at random 1) speaking or listening and 2) adding or deleting. Without
loss of generality, assume speaking and deleting was selected. Then, all removable svw ∈ Es are
deleted. Similarly, if speaking and adding is selected, then all addable svw ∈
/ Es are added. The
analogous definitions are used for listening edges.

L EMMA 26. Suppose that cs > 0. If there exists a vertex w ∈ V such that svw ∈ Es , `wv ∈
/ E` ,
then if v is selected to remove a speaking edge, she will select some such svw to remove. The
analogous statement holds for listening edges.

P ROOF. When agent v decides to delete a speaking edge, it chooses an edge which has the least
gain for it since by removing it, the lost be minimal. We know that the profit of existence a speaking
edge is made by providing speaking reachability. Moreover since `wv ∈ / E` and by the definition
of reachability, there is no path from v to arbitrary node z that passes through svw so the gain of it
zero. Therefore removing svw has the least lost for v and is the best strategy among removing all
possible speaking edges.

5. STABILITY AND EFFICIENCY

In this section we construct classes of efficient and stable networks. We first consider nontrivial
ranges of parameters k > 1 and cs , c` ≤ n − 1 (see Section 5.1 for a discussion of extremal
parameter ranges and networks). For this range of parameters, we see interesting networks emerge
in the study of efficient and stable networks.

R EMARK 27. For ease of notation, in the first part of this section we consider only the directed
version of our model (i.e., without loss of generality, c` = 0). However, all results and proofs
follow immediately for the bidirectional case by replacing the directed graphs in question by the
analogous bidirected graph where each directed edge is replaced by a complete bidirected edge. In
the statement of the theorems, constraints on c apply to both cs and c` .
 Fig. 2. A balanced flower graph with n = 26 vertices and k = 10. This is efficient and stable for c < 5.

√
The first network we consider, called balanced flower graph,
 is defined for k ≤ 2 n and is
 of length /2 + 1 . Select one node from this set
constructed as follows: Make a directedcycle k

and call it the center. As long as at least /2 nodes remain, select

k
  them, and, along with the center
node, form another directed cycle. Repeat until fewer than k/2 nodes remain; then remove one
non-central node  from
 each petal (severing its edges and connecting its predecessor and successor)
until you have k/2 − 1 nodes and form them into the final petal. We denote by q the number of
√
petals. Note at most
 one node is removed from each petal in balancing since since k ≤ 2 n, and
hence q ≥ k/2 . Note that the balanced flower graph has diameter at most k. See Figure 2 for an
example.
√
T HEOREM 28. For any 4 ≤ k ≤ 2 n and 3 ≤ n, the social welfare of the balanced flower
graph (see Figure 2) is
& '
n−1
n(n − 1) − c  k  − c(n − 1)
2
 
Moreover, if 1 ≤ c < k/2 − 1 the balanced flower graph is efficient and pairwise stable.
Before we prove the theorem, in order to show pairwise stability, a useful lemma is stated:
L EMMA 29. If G be stable and strongly connected, then G is pairwise stable.
P ROOF. Since G is stable, it contains no removable edge. As G is strongly connected, for any
vertex v, the positive component of UG (v) is 2(n − 1) and cannot be increased. Hence, there is no
addable pair in G. Thus, G is pairwise stable.
P ROOF OF T HEOREM 28. The social welfare of the balanced flower graph is
& '
n−1
n(n − 1) − c  k  − c(n − 1)
2
 
as the utility of the center node is (n − 1) − cq = (n − 1) − c n−1
k
 , and the utility of all other
2
nodes is (n − 1) − c. Hence, the social welfare is
& ' & '
n−1 n−1
(n − 1) · ((n − 1) − c) + (n − 1) − c  k  = n(n − 1) − c  k  − c(n − 1)
2 2
 Let us now prove that the flower graph G is stable. By Lemma 29, this will give pairwise stability.
We must show that for all agents v, and alternate strategy s0v ⊆ V , we have
USv ,S−v (v) ≥ USv0 ,S−v (v).
First, consider a vertex v that is not the center; since all vertices are reachable from v by paths of
length at most k, the positive component of the utility cannot be increased. Hence, utility can only
be improved by lowering the cost; however, |Sv | = 1, and if |Sv0 | = ∅, then v 0 s utility is 0, which is
strictly less than UG (v). Thus, v does not have an alternate strategy that can improve her utility.
Now consider the center vertex u. Again, all vertices are reachable from u by paths of length at
most k, so the positive component of the utility cannot be increased, and the negative component can
only be decreased by decreasing the number of outgoing edges from u. For sake of contradiction,
assume there is some |Su0 | < q that is u’s best response.
 Note that Su0 must disconnect
  u from at
least one petal. Since each petal has at least /2 − 1 non-center nodes and c < k/2 − 1, adding
k

the edge from u to the petal to Su0 would increase the utility; this contradicts the fact that |Su0 | < q
can be a best response. Hence, the balanced flower graph is stable.
Now we prove that the balanced flower graph is efficient. First, consider the graph that would arise
if we simply connected the remainder nodes into a small petal without balancing. This is what is
known as simply a flower graph. It is known that, for any n and 2 ≤ k ≤ n, the flower graph attains
the fewest number of edges of any connected graph on n nodes with diameter k (see Theorems 1
and 2 in [Hirschberg and Wong 1979]). Note that our balanced flower graph has the same number
of edges as the flower graph; hence it also attains the fewest number of edges of any connected
graph on n nodes with diameter k. Therefore, the balanced flower graph is optimal amongst the set
of strongly connected graphs.
Hence, it suffices to show that a graph that is not strongly connected cannot be efficient. We
prove the contrapositive. Assume there is a graph that is not strongly connected; we show that we
can combine any two maximal strongly connected components into a single strongly connected
component that is at least as efficient. Take two strongly connected of size a ≥ 1 and n − a ≥ 1.
Clearly, from above, the social welfare of each component can only improve by making it a balanced
flower, and this does not affect the social welfare of the remaining graph. Since it is possible that
one component is connected to the other (but not vice versa), without loss of generality the social
welfare of this graph is at most
& ' & '
a−1 n−a−1
a(a − 1) − c   − c(a − 1) + (n − a)(n − 1) − c   − c(n − a − 1)
k/2 k/2
& ' !
n−1
≤ n(n − 1) − a(n − a) − c   − 2 − cn
k/2
& '
n−1
= n(n − 1) − c   − c(n − 1) + c − a(n − a)
k/2
& '
n−1 √
≤ n(n − 1) − c   − c(n − 1) + n − (n − 1),
k/2

  √  √
where the last inequality follows as c < k/2 − 1 ≤ n − 1 ≤ n and 1 ≤ a ≤ n − 1. For
n ≥ 3, this is less than the social welfare attained by the balanced flower.
R EMARK 30. In fact, for any 2 ≤ k, the (unbalanced) flower graph which leaves the last petal √
at it’s size without rebalancing is also efficient; the constraint is due to the fact that for k > 2 n
balancing as described above may not be possible. We could, instead, define a recursive balancing
process that continues to steal vertices (possibly several from the same petal) until a balanced flower
Fig. 3. Kautz graph with 24 vertices, outdegree 2 and diameter 4. This is a symmetric-efficient graph for n = 24, k = 4,
and c < 10, and is stable for c < 1.

is reached for some k 0 < k; efficiency follows directly, as does stability for a reduced c which is a
function of n and k.
While the above graphs are efficient, they are highly asymmetric with a single node taking on
most of the cost. Hence, we now turn our attention towards symmetric graphs, and consider a second
class of graphs, known as Kautz graphs [Elspas et al. 1968; Imase and Itoh 1983] (see Figure 3).
The Kautz graph KdD is a directed graph with (d + 1)dD−1 vertices. The vertices are labeled by
strings x0 . . . xD−1 of length D with xi ∈ {0, . . . , d} with xi 6= xi+1 . The set of edges is defined
by
{(x0 , x1 . . . xD , x1 . . . xD xD+1 )| xi ∈ {0, . . . , d} and xi 6= xi+1 }.
Clearly, the graph has outdegree d, (d + 1)dD edges, and diameter D + 1.
Kautz graphs arose in the study of the following question: Given a graph with n nodes and m
edges, what is the smallest possible diameter k? Through a series of works, it was shown that Kautz
graphs are asymptotically optimal with respect to this question (see [Miller and Širán 2005] for a
survey).
In our case, we can rephrase the question as follows: Given a graph with n nodes and diameter
k, what is the smallest possible number of edges m? Clearly, such a graph would maximize social
welfare restricted to the set of strongly connected graphs; we can extend this result to all graphs.
k−1
T HEOREM 31. For any k ≥ 4, n ≥ 16, and c < n/logk (n) the Kautz graph Klog (see
k (n)
6
Figure 3) is asymptotically symmetric-efficient and its social welfare is
n((n − 1) − clogk (n)).
Moreover, for c ≤ 1 the Kautz graph is pairwise stable.
P ROOF OF T HEOREM 31. The Kautz graph is strongly connected and all vertices have the same
degree; hence it is symmetric. Each vertex is connected to all others and each has degree dn =
logk (n), hence each vertex contributes (n − 1) − c logk (n) to the social welfare.
As discussed in the main body of the text, the Kautz graph is known to asymptotically have the
fewest number of edges necessary in order to have a graph of size n with maximum diameter k
(see [Miller and Širán 2005] for a survey), and hence it is asymptotically optimal amongst the set

6 Generally, we assume that logk (n) ∈ Z, and then can make a statement about asymptotic as n → ∞ for a fixed k on the
set of well-defined n.
of strongly connected graphs. Analogously to the Proof of Theorem 28 (see the last paragraph), we
can bootstrap this result in order to prove efficiency amongst the set of all graphs by showing that
the social welfare of a graph that is not strongly connected is (asymptotically) at most the social
welfare of the corresponding Kautz graph for this range of parameters.
We prove the contrapositive. Assume there is a graph that is not strongly connected; take two
strongly connected components of size a ≥ 1 and b = n − a; we show that the Kautz graph of size
n has at least as much social welfare as this graph. Clearly, from above, the social welfare of each
component can only improve by making it a Kautz graph, and this does not affect the social welfare
of the remaining graph. Then, (since the two components may be connected) the social welfare is at
most
a(a − 1) − cada + (n − a)(n − 1) − c(n − a)dn−a
 
n−a
= n(n − 1) − a(n − a) − c logk (aa (n − a) )
 
n−1
≤ n(n − 1) − (n − 1) − c logk ((n − 1) )
!
n nn
≤ n(n − 1) − c logk (n ) + c logk n−1 − (n − 1)
(n − 1)
≤ n(n − 1) − c logk (nn )
where the last inequality follows for any constant c and k and large enough n. This is the social
welfare of the Kautz graph; hence the Kautz graph is asymptotically symmetric-efficient.
We now prove stability. By Lemma 29, this also gives pairwise stability. Clearly, no strategy Sx0
such that |Sx0 | ≥ |Sx | can improve the utility of x. Hence we first consider strategies Sx0 ⊆ Sx .
Let x = x0 , . . . , xD and y = y0 , . . . , yD be two nodes such that there is there is no i such that
xi = yD−i−1 , . . . , xD = y0 , . . . , yi . Hence, the distance from x to y must be exactly k in the Kautz
graph. Consider any edge xz such that z0 6= y0 . It must again hold that there is no i such that
zi = yD−i−1 , . . . , zD = y0 , . . . , yi ; otherwise there would have been some such i for x. Hence,
the distance from x to y is k + 1 if we remove edge xx0 where x0 = x1 , . . . , xD , y0 . Hence, the
strategy Sx \ xx0 does not improve x’s utility since the cost reduces by at most 1 while at least one
node becomes unreachable. Note that, for every allowable y0 , some such vertex y exists. Hence, no
strategy Sx0 ⊆ Sx improves x’s utility. Lastly, for sake of contradiction, assume there is an optimal
strategy Sx0 such that |Sx0 | < |Sx |. Then, there must be some allowable y0 to which x does not build
an edge. However, as we just observed, Sx0 + xy0 improves the utility of x. This contradicts the
optimality of Sx0 . Hence, this concludes the proof.
R EMARK 32. We leave open the question of whether Kautz graphs are stable for any c > 1,
and more generally, the question of characterizing optimal symmetric stable graphs.

5.1. Price of Stability and Price of Anarchy

We can further observe some properties of the extremal ranges of the parameters where the stable
and efficient graph structures are not as interesting. These results also lead to the following obser-
vation.
C OROLLARY 33. The price of anarchy is 0 and the price of stability is 1.
This follows from the definition of the price of anarchy and price of stability by combining Propo-
sitions 34 and 35 for the former and Propositions 36 and 35 for the latter.
P ROPOSITION 34. If cs ≥ 1 or c` ≥ 1 then the empty graph is pairwise stable. Moreover, if
cs > n − 1 or c` > n − 1 then the only stable graph (and hence the only pairwise stable graph) is
the empty graph.
 P ROOF OF P ROPOSITION 34. Assume without loss of generality that cs ≥ 1. We will prove the
contrapositive. Assume that G is not pairwise stable. Since the graph is empty, there are no possible
s s
edges to remove. Thus, there must exists a pair of vertices v, w such that UG+s vw +`wv
(v) > UG (v)
` ` s s s
and UG+svw +`wv (w) ≥ UG (w). This implies that UG+svw +`vw (v) = UG (v) + 1 − cs > UG (v),
and thus cs < 1.
Now, suppose that cs > n − 1. Note that in a stable network, the utility for any vertex v is at
least 0 since this can always by attained by taking the strategy Sv = ∅. Assume that a graph G is
nonempty. Thus, there must exist at least one vertex v such that ds+ (v) > 0. Let v be such a vertex.
s s
Then UG (v) = |RG (v)| − cs · ds+ (v) ≤ (n − 1) − cs < 0. Thus, G is not stable.
P ROPOSITION 35. If k = ∞, and 0 < cs ≤ n − 1, 0 < c` ≤ n − 1, then the cycle is the only
efficient network.
P ROOF. For ease of notation, we drop k = ∞ from the subscripts. For an arbitrary node v, if
ds+G (v) = 0 then there does not exist a speaking path from v to rest of the network, and hence
UGs
(v) = |RGs
(v)| − cs · ds+ +
G (v) = 0. Otherwise, if dsG (v) ≥ 1, then the rest of network may be
s
reachable from v. Hence, |RG (v)| ≤ n − 1 where equality occurs if and only if all other nodes are
reachable from v. Therefore
s
UG s
(v) = |RG (v)| − cs · ds+
G (v) ≤ (n − 1) − cs ,
and we know that cs < (n − 1). Hence, in the case of equality, the utility is improved over the
setting where ds+
G (v) = 0. Similarly, it follows that
`
UG (v) ≤ (n − 1) − c` .
Thus we can conclude that
s `
UG (v) = UG (v) + UG (v) ≤ (n − 1) − (cs + c` )
so
X
ϕ(G) = UG (v) ≤ n · ((n − 1) − (cs + c` ))
v
By the definition of efficiency, a network is efficient if it maximizes ϕ thus if we be able to find a
network say G such that ϕ(G) = n · ((n − 1) − (cs + c` )) then we could conclude that G is efficient.
Now we consider cycle C to compute ϕ(C). A cycle C is a closed path v1 , v2 , . . . vn , vn+1 such
that vn+1 = v1 , for all 1 ≤ i ≤ n svi vi+1 ∈ Es , `vi+1 vi ∈ E` . Thus by the definition of reachability
s ` s `
for any v, w ∈ C, v ∈ RG,∞ (w), v ∈ RG,∞ (w) so |RG (v)| = |RG,∞ (v)| = n − 1. Moreover for
+ +
any node v ∈ C we have dsC (v) = d`C (v) = 1. Therefore
X X
ϕ(G) = UG (v) = (n − 1) − (cs + c` )) = n · ((n − 1) − (cs + c` )).
v v
Now we prove that cycle is the unique efficient network. By contradiction assume that there
exists another efficient network G 6= C where ϕ(G) = n · ((n − 1) − (cs + c` )). According to
the first part of proof, the only condition for holding this equality is that for all v ∈ V it holds that
ds+ + s `
G (v) = d`G (v) = 1, |RG (v)| = |RG (v)| = n − 1, thus G is connected in other words, for all
s `
v, w ∈ V , we have that v ∈ RG (w) and v ∈ RG (w) = 1. Moreover since the speaking outdegree
of each node is one, so if we only consider speaking edges, we will have one cycle on the nodes.
Now we want to show that listening edges are exactly in opposite direction of speaking edges. If
not, there must exist pair (v, w) such that svw ∈ Es , `wv ∈ / E` thus by the same proof used in
proposition 12, node v could improve her utility by removing svw and this action dos not change
other nodes’ utility. This improvement in social welfare contradicts with efficiency of G.
P ROPOSITION 36. For k = ∞, if 0 ≤ cs ≤ n − 1 and 0 ≤ c` ≤ n − 1, then the cycle C is
pairwise stable.
 P ROOF. We first prove stability, and then use Lemma 29 to show that pairwise stability must
also hold. Using the same proof presented in Proposition 35, we know that for all v ∈ V , we have
|RGs
(v)| = |RG `
(v)| = n − 1 and ds+ +
C (v) = d`C (v) = 1. To prove the stability of C we must show
that for all agents v and any alternate strategy Sv0 ⊆ V , we have
USv ,S−v (v) ≥ USv0 ,S−v (v).
There are only four possible moves for node vi .
(1) Adding svi w such that w 6= vi+1
(2) Removing svi vi+1
(3) Adding `vi w such that w 6= vi+1
(4) Removing `vi vi+1 .
Now we show that after any such change, the utility of vi will decrease. Without loss of generality,
s `
consider a change to a speaking edge. We know that |RC (v)| = |RC (v)| = n − 1 which is the
maximum possible, so by adding a new speaking edge we could not gain; however we will loose
cs thus UC+svi w (vi ) = UC (vi ) − cs and since cs > 0 so UC+svi w (vi ) < UC (vi ). In the case
of deleting, if we remove a speaking edge, then ds+ C (vi ) = 0 thus for all w ∈ V, w 6= vi we have
s s
w ∈ / RC (vi ) so |RC (vi )| = 0. Therefore, UC−svi w (vi ) = UC (vi ) + cs − (n − 1), and since
cs < (n − 1) so UC−svi w (vi ) < UC (vi ). The same reasoning used for listening edges. Hence the
cycle is stable. By Lemma 29, it is also pairwise stable.
We conclude this section with a characterization of all efficient and stable networks for k = 1.
The proof follows directly from the definitions.
P ROPOSITION 37. Assume k = 1. If cs , c` > 1 then the empty graph is efficient and is the only
stable graph. If cs , c` ≤ 1 then the complete graph is efficient; if cs < 1 or cs < 1 then the complete
graph is the only only stable graph. If cs = c` = 1, then all graphs are efficient and stable.
6. THE GENERALIZED CLUSTERING COEFFICIENT
There is a wide literature on networks and their social and economic properties (see [Easley and
Kleinberg 2010]). A clustering coefficient is a measure of the degree to which nodes in a graph
tend to be highly interconnected in small clusters. While clustering is high in many real-world
networks, evidence suggests that social networks in particular are highly clustered [Holland and
Leinhardt 1971; Watts and Strogatz 1998]. Many (1-dimensional) measures of clustering have been
defined [Luce and Perry 1949; Opsahl 2013; Opsahl and Panzarasa 2009; Watts and Strogatz 1998].
These clustering coefficients capture statistics about local connections; i.e., vertices that are apart
by at most two edges. Our k-dimensional clustering coefficient generalizes these approaches by
considering clustering between vertices that are at most k + 1 edges apart.
Inspired by our model, we define a generalized clustering coefficient below. We let f (G, c, k) ∈
[0, 1] be the percentage of removable edges in G for the model specified by c ∈ R+ , k ∈
{2, 3, . . . , ∞}. We can define an analogous local version f (v, c, k) ∈ [0, 1] as the percentage of
removable edges going out from v. These functions implicitly define clustering coefficients as a
function of our model parameters. In particular, we define two specific such clustering coefficients
which we believe are of general interest.
Definition 6.1. The generalized (global) clustering coefficient of dimension k of a graph G to
be the k-dimensional vector g where
gi (G) = f (G, 1, i + 1, T (·) = V ).
Similarly, the generalized (local) clustering coefficient of dimension k of a vertex v to be the k-
dimensional vector h where
hi (G) = f (v, 1, i + 1, V ).
 For any c ≤ 1, an edge uv is removable if and only if there is a path of length at most k from
u to v that does not use uv. Hence, the 1-dimensional generalized clustering coefficient makes a
statement about the number of edges uv that participate in at least one closed triangle.
R EMARK 38. Our definition of clustering is closely related to the class of clustering coefficients
which defined to be the percentage of (potential) triangles that are closed by an edge [Watts and
Strogatz 1998]. We instead measure the percentage of edges that close at least one triangle. Sim-
ilarly, we can generalize the original definition to higher k: it would measure the percentage of
(potential) cycles of length at most k that are closed, while we measure the percentage of edges
completing at least one cycle of length at most k.
7. DISCUSSION & FUTURE WORK
For the directed setting (c` = 0), we show that for k = ∞, asynchronous stochastic edge dynamics
converge to a stable network; moreover these fixed points can have non-trivial network structure.
Our proof does not generalize to the case of k < ∞, and we leave open the question of whether these
dynamics converge in that setting. Proving bounds on the time to convergence and understanding
the regions of attraction would also be of interest as they inform the distribution of networks we
would expect to see from a generative version of this model.
For the bidirected setting, we show that for arbitrary k the edge dynamics exhibit fast converge.
Moreover, vertex dynamics (in which multiple addable or removable edges are modified) are also
fast to converge. However, developing an understanding of asynchronous best-response dynamics
remains open as an interesting line for future work; our results do not generalize as there is no reason
why a best response would be limited to only changing addable and removable edges (given two
addable edges, it is not necessarily optimal to add both). √
With respect to the static game, we give classes of efficient networks (for k . n) and
symmetric-efficient networks (for 4 ≤ k). While the former are stable for any cs , c` . k/2, the
latter are only stable for cs , c` ≤ 1; for the latter we rely on a long line of work from combinatorics,
and determining whether any symmetric stable network exists appears to be a deep and technically
challenging open problem.
Lastly, a natural generalization of our model is to define target sets T s (v), T ` (v) ⊆ V and con-
sider utilities
s
UG,k s
(v) = |RG,k (v) ∩ T s (v)| − cs · ds+
G (v) and `
UG,k `
(v) = |RG,k (v) ∩ T ` (v)| − c` · d`+
G (v).
Such a model captures agents who participate in a larger network, but are only interested in connect-
ing to some subset of nodes. Understanding the stability and efficiency of networks under certain
classes of target sets would be of interest.

References
R. Aumann and R. Myerson. 1988. An application of the Shapley value. In Endogenous formation
of links between players and coalitions. Cambridge Univ. Press, Cambridge, UK.
V. Bala and S. Goyal. 2000. Self-organization in communication networks. Econometrica (2000).
Albert-László Barabási and Réka Albert. 1999. Emergence of scaling in random networks. science
286, 5439 (1999), 509–512.
Bélla Bollobás. 2001. Random Graphs. Cambridge University Press: Cambridge, UK.
Peter J Carrington, John Scott, and Stanley Wasserman. 2005. Models and methods in social net-
work analysis. Vol. 28. Cambridge university press.
Bhaskar Dutta and Suresh Mutuswami. 1997. Stable networks. Journal of Economic Theory 76, 2
(1997), 322–344.
David Easley and Jon Kleinberg. 2010. Networks, crowds, and markets: Reasoning about a highly
connected world. Cambridge University Press.
Bernard Elspas, William H Kautz, and James Turner. 1968. THEORY OF CELLULAR LOGIC
NETWORKS AND MACHINES. Technical Report. DTIC Document.
Paul Erdos and Alfréd Rényi. 1960. On the evolution of random graphs. Bulletin of the International
Statistical Institute 38, 4 (1960), 343–347.
D. S. Hirschberg and C. K. Wong. 1979. Upper and Lower Bounds for Graph-Diameter Problems
with Application to Record Allocation. Journal of Combinatorial Theory (1979).
Paul W Holland and Samuel Leinhardt. 1971. Transitivity in structural models of small groups.
Comparative Group Studies (1971).
Makoto Imase and Masaki Itoh. 1983. A design for directed graphs with minimum diameter. IEEE
Trans. Comput. 8 (1983), 782–784.
Matthew O Jackson. 2003. The stability and efficiency of economic and social networks. In Net-
works and Groups. Springer, 99–140.
Matthew O Jackson. 2005. A survey of network formation models: stability and efficiency. Group
Formation in Economics: Networks, Clubs, and Coalitions (2005), 11–49.
Matthew O Jackson and Alison Watts. 2002. The evolution of social and economic networks.
Journal of Economic Theory 106, 2 (2002), 265–295.
Matthew O Jackson and Asher Wolinsky. 1996. A strategic model of social and economic networks.
Journal of economic theory 71, 1 (1996), 44–74.
Amit Karandikar, Akshay Java, Anupam Joshi, Tim Finin, Yaacov Yesha, and Yelena Yesha. 2008.
Second Space: A Generative Model for the Blogosphere.. In ICWSM.
R Duncan Luce and Albert D Perry. 1949. A method of matrix analysis of group structure. Psy-
chometrika 14, 2 (1949), 95–116.
Mirka Miller and Jozef Širán. 2005. Moore graphs and beyond: A survey of the degree/diameter
problem. Electronic Journal of Combinatorics 61 (2005), 1–63.
Roger B. Myerson. 1977. Graphs and cooperation in games. Mathematics of Operations Research
(1977).
Mark EJ Newman. 2003. The structure and function of complex networks. SIAM review 45, 2
(2003), 167–256.
Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V. Vazirani. 2007. Algorithmic Game The-
ory. Cambridge University Press. 411–443 pages.
Tore Opsahl. 2013. Triadic closure in two-mode networks: Redefining the global and local clustering
coefficients. Social Networks 35, 2 (2013), 159–167.
Tore Opsahl and Pietro Panzarasa. 2009. Clustering in weighted networks. Social networks 31, 2
(2009), 155–163.
Stanley Wasserman and Katherine Faust. 1994. Social network analysis: Methods and applications.
Vol. 8. Cambridge university press.
Alison Watts. 2003. A dynamic model of network formation. In Networks and Groups. Springer,
337–345.
Duncan J Watts and Steven H Strogatz. 1998. Collective dynamics of ‘small-world’networks. nature
393, 6684 (1998), 440–442.