dynamic network formation with incomplete information€¦ · that a much wider collection of...
TRANSCRIPT
Dynamic Network Formation with Incomplete
Information∗
Yangbo Song† Mihaela van der Schaar‡
November 20, 2013
Abstract
How do networks form and what is their ultimate topology? Most of the literature that addresses
these questions assumes complete information: agents know in advance the value of linking to other
agents, even with agents they have never met and with whom they have had no previous interaction
(direct or indirect). This paper addresses the same questions under what seems to us to be the
much more natural assumption of incomplete information: agents do not know in advance – but
must learn – the value of linking to agents they have never met. We show that the assumption
of incomplete information has profound implications for the process of network formation and the
topology of networks that ultimately form. Under complete information, the networks that form
and are stable typically have a star, wheel or core-periphery form, with high-value agents in the core.
Under incomplete information, the presence of positive externalities (the value of indirect links) implies
that a much wider collection of network topologies can emerge and be stable. Moreover, even when
the topologies that emerge are the same, the locations of agents can be very different. For instance,
when information is incomplete, it is possible for a hub-and-spokes network with a low-value agent
in the center to form and endure permanently: an agent can achieve a central position purely as the
result of chance rather than as the result of merit. Perhaps even more strikingly: when information
is incomplete, a connected network could form and persist even if, when information were complete,
no links would ever form, so that the final form would be a totally disconnected network. All of this
can occur even in settings where agents eventually learn everything so that information, although
initially incomplete, eventually becomes complete.
Keywords: Network Formation, Incomplete Information, Dynamic Network Formation, Link
Formation, Formation History, Externalities
JEL Classification: A14, C72, D62, D83, D85
∗We are grateful to William Zame, Ichiro Obara, Moritz Meyer-ter-Vehn, Sanjeev Goyal, Luca Canzian,
Simpson Zhang and a number of seminar audiences for suggestions which have significantly improved the paper.†Department of Economics, UCLA. Email: [email protected].‡Department of Electrical Engineering, UCLA, and Director of UCLA Center for Engineering Economics,
Learning, and Networks. Email: [email protected].
1
1 Introduction
How do social and economic networks form and what is their ultimate shape (topology)? The lit-
erature, which includes work by Jackson and Wolinsky (1996)[20], Barabasi and Albert (1999)[4]
Bala and Goyal (2000)[2], Jackson and Watts (2002)[19] and Ballester et al. (2006)[3], has ad-
dressed these questions in both static and dynamic contexts. The central conclusion of this
work is that special shapes of networks can occur and persist. However, this literature makes
the strong assumption of complete information: agents know in advance the value of linking
to other agents - even agents they have never met and with whom they have had no previous
interaction (direct or indirect). The present paper addresses the same questions under what
seems to us to be the much more natural assumption of incomplete information: agents do not
know in advance – but must learn – the values of linking to agents they have never met1. As
is usual in environments of incomplete information, agents begin only with beliefs about the
values of linking to other agents, make choices on the basis of their beliefs, and update their
beliefs (learn the true values) on the basis of their experience (history).
We show that the assumption of incomplete information has profound implications for both
the process of network formation and the topology of networks that ultimately form. When
information is complete, the networks that form and persist typically have a star or core-
periphery form, with high-value agents in the core. By contrast, when information is incomplete,
a much larger variety of networks and network shapes can form and persist. Indeed, the set of
networks that can form and persist when information is incomplete is a superset (typically a
strict superset) of the set of networks that can form and persist when information is complete.
Moreover, even when the network shapes that form are the same or similar, the locations of
agents within the network can be very different. For instance, when information is incomplete,
it is possible for a star network with a low-value agent in the center to form and persist
indefinitely; thus, an agent can achieve a central position purely as the result of chance rather
than as the result of merit. Perhaps even more strikingly, when information is incomplete,
a connected network can form and persist even if, when information were complete, no links
would ever form so that the final form would be a totally disconnected network.
However, the most important consequence of incomplete information is not that a larger
variety of network shapes (topologies) might emerge, but that the particular shape that does
emerge depends on the history of link formation and of link formation opportunities. For
instance, when information is incomplete, agents i, j might choose to form a link because each
expects the value of the link to exceed the cost of forming it. Having formed the link, the agents
may learn that their expectations were wrong and so might wish to sever it. However, before
the agents have the opportunity to sever the link, each of them may have formed other links,
so that the indirect value of the link between i and j - the value of the connection to other
1We emphasize that we study learning during the network formation process, rather than learning in an
exogenously given and fixed network structure. For a study on the latter, see for example Acemoglu et al.
(2008)[1]
2
agents - may be sufficiently large that they prefer to maintain the link between them after all.2
However, whether these other links have formed will depend not only on the values of those
links but also on the random opportunities presented to form or not form them.
We stress that all of this can occur even in settings where agents eventually learn everything
so that information, although initially incomplete, eventually becomes complete. Incomplete-
ness of information may eventually disappear but its influence may persist forever.
2 Literature Review
The literature on network formation studies the characteristics of networks emerging as a
result of the strategic interaction of self-interested agents. This literature includes renowned
early papers in this area, including Jackson and Wolinsky (1996)[20], Bala and Goyal (2000)[2],
Watts (2001)[32], etc. These papers and subsequent works building on them (for instance
Johnson and Giles (2000)[21]) assume that agents are homogeneous, i.e. the value that each
agent in the group provides to and receives from another agent is the same. This is the strongest
form of complete information, in the sense that agents do not only know their exact payoffs
from linking to others, but are also aware that the payoffs are solely determined by the network
topology, and that the agents’ identities play no role in affecting payoff characteristics. As a
result, a prominent feature in network topologies that emerge and/or become stable is that they
are either empty or connected. By contrast, in our paper, agents are of different types and thus
connecting to them results in different (heterogeneous) payoffs; moreover, there is incomplete
information: an agent does not know the types of agents that he has never connected with,
but he is able to form beliefs based on which he chooses the optimal action. Networks which
result under this, more realistic, assumption are strikingly different from those obtained in the
model assuming homogeneous agents and complete information. On one hand, connectedness
is no longer a key property of stable networks - the formation process converges to connected
networks in some range of parameters and to multiple components in other ranges. On the
other hand, even if the network stays empty forever under complete information, a non-empty,
even connected network may emerge and be stable with positive probability under incomplete
information.
The literature studying how agent heterogeneity affects network formation defines ”hetero-
geneity” in various manners: as differentiated failure probabilities for different links (Haller
and Sarangi (2003)[15]), as differences in values obtained from links and costs for forming links
across agents (Galeotti (2006)[11], Galeotti et al. (2006)[13]), as different amounts of valuable
information produced endogenously (Galeotti and Goyal (2010)[12], Zhang and van der Schaar
(2012)[34]), as different endogenous effort levels (Koenig et al. (2012)[22]), or as agent-specific
resource holding amounts (Zhang and van der Schaar (2013)[35]). Despite such diverse views
on heterogeneity, complete information is still a common assumption, and the predictions in
2Indirect linking provides a positive externality when information is complete as well, but the effect is
completely different: i, j will never wish to form a link at some point in time and sever it later.
3
the above papers are often restricted to a few, specific, types of network topologies such as
stars, wheels or core-periphery networks. Moreover, in networks with asymmetric connectivity
degree distribution, for instance stars and core-periphery networks, agents with high values
or low costs always stay at the center or core, whereas agents with low values or high costs
remain in the periphery. We differ from these works in four aspects. First, as mentioned above,
agents have no precise knowledge about their exact payoffs due to incomplete information;
instead, they choose an optimal action according to their beliefs about the payoffs that they
will obtain from connecting to others. Secondly, in most of the above papers an agent’s payoff
is affected by a single factor, either his own characteristics or other agents’ characteristics;
our model allows both factors to jointly determine the payoffs, for instance, a type a agent’s
payoff from linking to another type a agent may differ from his payoff from linking to a type
b agent, and vice versa. Thirdly, we show that the interaction of incomplete information and
agent heterogeneity produces a much wider range of network topologies, which includes stars,
wheels, core-periphery networks, etc. Finally, the topology that emerges and becomes stable
strongly depends on the formation history : an agent may exhibit a high degree of connectivity
in equilibrium not necessarily because he is of a special (high) type but also because initially
he was fortunate to obtain sufficiently many links by chance, which in turn attracted others to
form and maintain links with him due to the large indirect benefits that he can offer. Therefore,
unlike most existing literature that only emphasizes what topologies can be formed, we argue
that how a certain topology comes into being is equally important.
Among the empirical literature on network formation games, works such as Falk and Kosfeld
(2003)[8], Corbae and Duffy (2008)[5], Goeree et al. (2009)[14] and Rong and Houser (2012)[30]
have conducted experimental studies on the types of emerging topologies. In particular, Goeree
et al. (2009)[14] raises similar concerns to ours regarding the assumption of complete infor-
mation being unrealistic in practice, and introduces incomplete information about the private
types of agents in their experiments. The experimental results indicate that (1) typical equilib-
rium network topologies predicted by the existing theoretical analysis are not always consistent
with the empirical observations; especially, stars are formed only in a proportion of the total
number of experiments conducted([5],[30]), and such proportions, under some treatments such
as a two-way flow of payoffs, are rather low ([8]); (2) even in the experiments where equilibrium
network topologies do emerge with high frequency, such topologies are developed rather than
born([14]), which we believe suggests a dynamic network formation process of a sufficiently
long duration as a more appropriate environment for stable networks to emerge, compared
with a static one. Moreover, the study of networks which actually get formed in large social
communities, as presented for instance by Mele (2010)[25] and Leung (2013)[23], shows that in
environments where agents are heterogeneous and withhold certain private information in their
payoffs from links, numerous phenomena which are not predicted by the existing theoretical
literature can happen. For instance, networks can have multiple components, agents with dif-
ferent attributes can cluster over time, agents can remain singletons even in the long run, etc.
Our model does not explain all these phenomena; however, we do believe that incorporating
4
incomplete information in the dynamic network formation game represents a first and impor-
tant step towards understanding why several previously seemingly irregular network topologies
can emerge and remain stable in practice.
One assumption in our model that does stay in conformity with the above theoretical works
is agent myopia, i.e. agents only care about their current period payoffs when making decisions.
In contrast to this assumption, there is a literature including Page Jr. et al. (2003)[29],
Deroian (2003)[6] and Dutta et al. (2005)[7], that considers farsighted agents, in the sense
that agents take the effect that their current actions have on future payoffs into account. The
major objective of these papers is to construct equilibrium network topologies that are strongly
efficient. We would like to point out that in our setting, determining the topology of a strongly
efficient network would be by itself a challenging task since such topologies may not be unique
(to be discussed in more detail later). We leave a thorough analysis on this topic as a promising
direction for future research.
Finally, regarding learning in social networks, the usual approach as in Gale and Kariv
(2003)[10] and Acemoglu et al. (2008)[1] is to assume a fixed network topology or an exogenous
process determining the topology, and study the evolution of actions and beliefs as information
gets updated over time. Essentially, the learning process in our paper goes in line with the
above literature in the sense that agents are learning about a payoff relevant state which is
the type vector; however, in our model the network formation process is endogenous, which
is a key difference to the works existing in this area. In other words, in our environment two
processes evolve simultaneously : the network formation process, which pins down the network
topology, and the belief formation process, which determines the information each agent learns
about others’ types. On one hand, network formation serves as the medium of belief formation
since agents acquire knowledge on others’ types via connection. On the other hand, belief
formation provides the basis for network formation because agents choose actions according to
their current beliefs.
The rest of the paper is organized as follows. Section 3 introduces the model. Section 4
analyzes the model in detail and interprets the results. Section 5 analyzes a special case in
the context of large networks. Section 6 discusses some extensions of the model. Section 7
concludes and introduces relevant future research topics.
3 Model
Our notation mainly follows Jackson and Wolinsky (1996)[20], Bala and Goyal (2000)[2] and
Watts (2001)[32].
3.1 Networks with Incomplete Information
Networks and the Agents’ Types
Let I = {1, 2, ..., N} denote a group of N agents. Each agent can either be of type a or
5
type b. Most of our analysis in a two-type environment carries over to one with more than two
types, yielding similar conclusions with slightly modified conditions. Let ki denote agent i’s
type, and let na and nb denote the number of type a and type b agents. Let κ = {ki}Ni=1 denote
the type vector of the agents. The realization of this type vector is drawn i.i.d. for each agent,
with Prob(a) = p and Prob(b) = 1− p.A network is denoted by g ⊂ {ij : i, j ∈ I, i 6= j}, and a sub-network of g on I ′ ⊂ I, denoted
gsub(I′), is defined as a subset of g such that ij ∈ gsub(I
′) if and only if i, j ∈ I ′ and ij ∈ g.
ij is called a link between agents i and j. We assume throughout that links are undirected, in
the sense that we do not specify whether link ij points from i to j or vice versa. A network g
is empty if g = ∅.
We say that agents i and j are connected, denoted ig↔ j, if there exist j1, j2, ..., jn for some
n such that ij1, j1j2, ..., jnj ∈ g. Let dij denote the distance, or the smallest number of links
between i and j. If i and j are not connected, define dij := ∞. An agent i in a network is a
singleton if ij /∈ g for any j 6= i.
Let N(g) = {i|∃j s.t. ij ∈ g}. A component of network g is a maximal connected sub-
network, i.e. a set C ⊂ g such that for all i ∈ N(C) and j ∈ N(C), i 6= j, we have iC↔ j, and
for any i ∈ N(C) and j ∈ N(g), ij ∈ g implies that ij ∈ C. Let Ci denote the component that
contains link ij for some j 6= i. Unless otherwise specified, in the remaining parts of the paper
we use the word ”component” to refer to any non-empty component.
A network g is said to be empty if g = ∅, and connected if g has only one component which
is itself. g is minimal if for any component C ⊂ g and any link ij ∈ C, C − ij is no longer a
component. g is minimally connected if it is minimal and connected.
Payoff Structure
Following the assumption in literature on networks that explicitly model non-local external-
ities3, such as Jackson and Wolinsky (1996)[20], Watts (2001)[32] and Galeotti et al. (2006)[13],
we assume that once agents i and j form a link, i not only obtains payoffs from his immediate
neighbor j, but also from the agents that he is indirectly connected to via that particular link.
As stated before, the payoff of an agent also depends on the type vector. Specifically, an agent
i’s payoff from a network g is given by
ui(ki, k−i,g) = ui(ki, k−i, Ci) :=∑jCi↔i
δdij−1f(ki, kj)−∑j:ij∈Ci
c
where f : {a, b} × {a, b} → R++ is the payoff function for an agent from a link with another
agent. Therefore, f(ki, kj) > 0 denotes the payoff to an agent i of type ki by linking to an
agent j of type kj. c > 0 is the cost of maintaining a link, which is assumed to be bilateral and
homogeneous across agents. δ ∈ [0, 1] denotes a common decay factor4, such that the payoff of
i from j with a distance of dij is δdij−1f(ki, kj). Note that c is assumed to be independent of
3As opposed to papers that assume pure local externalities, i.e. agents only obtain payoffs from their
immediate neighbors. See for example Galeotti and Goyal (2010)[12].4In this paper, unless otherwise specified, we assume that δ ∈ (0, 1).
6
the agents’ types5. Without loss of generality, we assume that f(a, a) ≥ f(b, b), but we impose
no additional assumptions on the relative magnitude among f(a, a), f(a, b), f(b, a) and f(b, b).
In particular, the payoffs may be asymmetric, i.e. f(ki, kj) may be different from f(kj, ki)
when ki 6= kj. As we will see in the analysis, the relative magnitudes of these parameters have
decisive effects on the resulting convergence of the formation process.
Note that, given the considered payoff structure, there are increasing returns to link for-
mation: linking with an agent is more valuable if that particular agent connects to a larger
number of agents. Moreover, when agents are heterogeneous, not only does the (discounted)
number of connections matter, but also the types of agents connected.
From the formulation above, two forces are at work in determining whether a certain link
is valuable for an agent. On one hand, the payoff an agent gets from a link not only depends
on his own type, but also the type of the counterparty. As a result, the private information an
agent possesses–his type–determines his willingness to link to others as well as his preference
on what types of agents to link to. On the other hand, the belief of an agent over the types
of others plays an equally crucial role. In contrast to works in the literature which assume
complete information on payoffs, the same realization of a type vector may result in rather
distinct network formation processes given different beliefs. We provide the following notations
to emphasize the above two points, which will be discussed throughout the analysis.
We define the self effect by fself = |f(a, a)− f(b, a)|+ |f(a, b)− f(b, b)|, and the peer effect
by fpeer = |f(a, a)−f(a, b)|+ |f(b, a)−f(b, b)|.6 As the names indicate, the self effect measures
how much an agent’s payoff depends on his own type, and the peer effect measures how much
his payoff depends on the types of agents he connects to. For example, a strong self effect may
be found in a network among workers collaborating on a job, where agents need to exchange
homogeneous messages from time to time but some agents are introverted/extroverted and
thus have different benefits from social interaction, regardless of the type of the neighbor. In
this case, the payoff of an agent would only depend on his own characteristics. Strong peer
effects are often present in networks dedicated to information and resource sharing (e.g. peer
to peer networks), such as data, knowledge, news as well as computational and processing
capabilities, since in such networks the resource acquirer’s characteristics would not affect the
quality of resources it obtains from others. The two effects can both be strong in social networks
– not only do different appearances or personalities leave different impressions, but different
people may have different preferences even towards the same person. Note that when both
effects are zero, the model essentially becomes one with homogeneous agents and no incomplete
information, since there is no payoff heterogeneity whatsoever. In the analysis, we will show
that the magnitude of these two effects will greatly influence the formation process and will
5We make this assumption to economize on notations. In the case where costs are also heterogeneous, our
analysis and results hold with slightly modified conditions.6The effect-relevant results we derive are mostly about pure self effect or pure peer effect, i.e. one of the
effects is zero. Thus we tacitly assume the particular forms of the two terms as above, and even if we defined
the effects alternatively by putting positive coefficients before the absolute values, for instance a weighted sum,
the results would not be affected.
7
lead to the emergence of very different network structures.
Let fa = pf(a, a) + (1 − p)f(a, b) denote a type a agent’s expected payoff from linking to
a singleton agent, and let fb = pf(b, a) + (1− p)f(b, b) denote that of a type b agent. Because
information is incomplete, the expected payoff and the actual payoff are different, which can lead
to a very different formation process from the one under complete information. For example,
consider a group consisting of only type b agents where f(b, b) < c. Under complete information,
no link will ever form since no agent has the incentive to form the first link. However, under
incomplete information, if fb is large enough, agents would be incentivized to make initial links,
and it is possible that a number of agents would be connected after a few periods. Due to the
assumption that externalities are not purely local, i.e. agents get payoffs from other indirectly
connected agents, the positive externality may become large enough to maintain a non-empty,
even connected network. We will derive general properties and provide more examples in the
analysis.
3.2 Dynamic Network Formation Game
The Game
We model the dynamic game in a similar fashion to Watts (2001)[32] and Jackson and Watts
(2002)[19]. Time is discrete and the horizon is infinite: t = 0, 1, 2, .... The game is played as
follows: agents start with an empty network ∅ in period 0. In each following period, a pair of
agents (i, j) is randomly selected to update the link between them. For simplicity, we assume
that the matching probability is uniform, so (i, j) is selected with probability 2N(N−1) . The two
agents then play a simultaneous move game, where each can choose to sever the link between
them if there is one, and if there is not, whether to agree to form a link with the other agent.
Let aij = 1 denote the strategy that i agrees to form a link with j (if there is no existing link)
or not to sever the link (if there is an existing one), and aij = 0 otherwise. A link is formed or
maintained after bilateral consent (i.e. aij = aji = 1). Let γ(t) := {(i, j)t′}tt′=1 (t ≥ 1) denote
a selection path up to time t, or the set of selected pairs of agents, ordered from 1 to t. The
agents are assumed to be myopic, i.e. they only care about their current period payoffs.
In addition, we make the following assumptions throughout the paper:
Assumption 1. (Component observation) When agents i and j are selected, i can observe Cj,
and vice versa.
Assumption 2. (Linking is valuable in expectation) The expected payoff fk ≥ c ∀k ∈ {a, b}.
Assumption 1 is common in the literature on network formation. It means that there is
no incomplete information on the network topology regarding the link to be updated, so our
analysis focuses on the effect of incomplete information about types. Assumption 2 works as a
participation incentive: since agents are myopic and the network is empty to start with, we let
their expected payoff from a first link be non-negative, so that they are willing to participate
in this network formation game from the beginning. Note that even if fk ≥ c, it could still be
8
the case that f(k, k′) < c for some type k′. For example, suppose that p = 0.5, f(a, a) = 10,
f(a, b) = 2, c = 5. Then fa = 6 > c but we still have f(a, b) < c.
Updating Rule
With incomplete information, agents maximize their expected payoffs, rather than actual
payoffs, when deciding the optimal action. Therefore, the belief of an agent on the types of
the other agents plays a crucial role in shaping his behavioral patterns. For any i, we let
Bi ∈ ∆{a, b}N denote agent i’s belief on the type vector of the group, and B = (B1, ..., BN)
define a belief vector for the group of agents. Note that for any i, Bi must put zero probability
on any type vector where i’s type differs from the true type. Moreover, we assume the following
simple updating rule for the agents:
1. If two agents are ever connected, they know each other’s type.
2. Otherwise, their belief on each other’s type remains at the prior.
In other words, agents can only observe their own formation history: they do not observe
the actions by agents in other components and thus make no relevant inferences on their
types. A similar assumption appears in McBride (2006)[24], which is denoted as imperfect
monitoring and describes agents’ inability to observe all other agents’ strategies in a static
network formation game. A plausible alternative updating rule is to allow agents to observe
each other’s past actions and perform Bayesian updating accordingly, which results in very
complicated belief formation. We will discuss this in a later section. Throughout the paper,
we assume that the updating rule is common knowledge among the agents.
4 Analysis
In this section, we analyze the nature of the network formation process and show a clear contrast
between results under complete information and incomplete information. We begin by defining
the solution concept for the two-player game each period, and the notion of a stable network.
4.1 Stable Optimistic Equilibrium and Stable Network
Stable Optimistic Equilibrium
Let C denote the set of all possible components (including the empty component), i.e. when
agents i and j are selected, for agent i, C is the set of all possible observations of Cj. Let B denote
the set of all possible beliefs for an agent. Denote K = {a, b}×C×B as the set of information7
for an agent who is selected to update a link. The interpretation of knowledge is the following:
the first argument is the agent’s own type, the second argument is the component containing
7Technically speaking, agent i’s information also includes Ci, but omitting such information does not affect
the existence and any property of the equilibrium we define below.
9
his counterparty, i.e. the other agent selected, and the third argument is his current belief on
the type vector.
Now we define an agent’s (pure) strategy in the two-player game after a pair of agents is
selected, followed by our proposed solution concept, which we call a stable optimistic equilibrium
(SOE):
Definition 1. Agent i’s (pure) strategy towards j is sij : K → {0, 1}, a function from agent i’s
set of information to his set of actions.
Definition 2. A strategy profile s = (sij, sji) is a stable optimistic equilibrium (SOE) if: (1) it
is mutual best response; (2) sij = 1 if
E[ui(ki, k−i, (Ci ∪ Cj) + ij)|Bi] ≥ ui(ki, k−i, Ci)
In words, a SOE is an equilibrium where any agent would choose to agree to form a link (if
there is no existing link) or choose not to sever the link (if there is an existing one) as long as
the expected payoff from the link is non-negative. It is stable because the prescribed strategy is
robust to small probabilistic changes in the counterparty’s strategy when the above inequality
is strict. More specifically, if some agent j other than i changes sji with a sufficiently small
probability ε, i’s best response would not change. It is optimistic in the sense that it excludes
the “pessimistic”, or null equilibrium in which no agent agrees to form any link. In other words,
agents choose to link when they are indifferent. If two selected agents know each other’s type,
then they simply play a weakly dominant strategy; otherwise, by Assumption 2 the expected
payoff from linking is non-negative, so it is a mutual best response for each agent to agree.
This solution concept can be viewed as an extension of pairwise stability introduced by
Jackson and Wolinsky (1996)[20], with the new element of incomplete information and belief
formation. In their definition, a link between two agents is pairwise stable if it brings non-
negative payoffs to both, and there is no need to distinguish between payoffs before and after
link formation because information is complete. In our case, we define the equilibrium in
conformity with the idea that a link is formed, or is not severed, if it brings bilateral non-
negative payoffs; however, we highlight the clear difference between expected payoff, i.e. the
payoff before link formation and realized payoff, i.e. the payoff after.
We would like to emphasize the connection between the existence of SOE and Assumption
2. By Assumption 2, the expected payoff from a link to some unknown agent is non-negative;
when two agents who have never been connected before, say i and j, are selected, even though i
may not know j’s belief about the types of other agents in Ci (and vice versa), i does know that
j would play the equilibrium strategy of agreeing to form a link, no matter what j’s type is. As
a result, also agreeing would be i’s best response. Assumption 2 ensures that an agent makes
no inference on the counterparty’s type from equilibrium strategies, and thus the prescribed
strategy in an SOE indeed constitutes an equilibrium.
Lemma 1. In every period, for any pair of selected agents, any network topology and any belief
vector according to the simple updating rule, a SOE exists and is unique.
10
Lemma 1 together with the above robustness property of a SOE ensures that the outcome
of the formation process is unique and robust to small perturbation.
Stable Network
Let g(γ(t)) denote the unique network formed after period t, following a selection path of
γ(t). Let B(γ(t)) denote the associated belief vector after period t. By Lemma 1 we know
that both g(γ(t)) and B(γ(t)) are well-defined. Let σ(t) := (γ(t), {g(γ(t′))}tt′=1) denote a
formation history up to time t. We say that:
1. A network together with the associated beliefs (g, B) can emerge if there exists a selection
path γ(t) for some t, such that g = g(γ(t)), B = B(γ(t)).
2. (g, B) is a stable network if no link is formed or broken given any subsequent selection path.
3. The formation process can converge to g if there exists a selection path γ(t) for some
t, such that g = g(γ(t)) and (g(γ(t)), B(γ(t))) is stable. Mathematically, we denote it as
limt→∞ g(γ(t)) = g for some fixed γ(∞).
4.2 Contrast with Complete Information
Before discussing the differences between the network topologies that can emerge and be stable
under complete information and incomplete information, we first inspect how long incomplete
information can persist. We say that information is complete when every agent’s belief is the
degenerate belief on the true type vector κ, i.e. Prob(κ|Bi) = 1 for any i, and that information
is incomplete otherwise.
Proposition 1. In the formation process, for every κ:
1. Information becomes complete within finitely many periods almost surely.
2. Information is complete in any stable network.
In essence, everyone eventually learns the true type vector with probability 1 over time.
Indeed, as agents are willing to form links with others of unknown type, after sufficiently many
periods the probability of unconnected agents still existing in the group would be arbitrarily
small. In other words, the effect of incomplete information and thus the difference with complete
information only occurs in an early stage of the formation process. Moreover, we can now
suppress the notation for beliefs in (g;B) and only use g when referring to a stable network.
Nevertheless, even though Proposition 1 may leave the impression that incomplete infor-
mation is not so important as it only takes effect in the short run, we will emphasize in the
following analysis that such short-term influence is actually persistent over time.
Let GC(κ) = {g : g = g(γ(t)) for some γ(t)} denote the set of networks that can emerge
under complete information given κ, and GIC(κ) that under incomplete information. Similarly,
let G∗C(κ) = {g ∈ GC(κ) : g = limt→∞ g(γ(t)) for some γ(∞)} denote the set of networks
11
that can emerge and be stable under complete information given κ, and G∗IC(κ) that under
incomplete information.
Theorem 1. For every κ, GC(κ) ⊂ GIC(κ).
Theorem 1 states that if some network can emerge under complete information, then it can
also emerge under incomplete information. Intuitively, if a link could be formed under complete
information, then given Assumption 2 and the simple updating rule, it can also be formed under
incomplete information, whether or not the relevant agents know each other’s type. Note that
the reverse is not necessarily true: even if a link could be formed under incomplete information,
it may not form under complete information because the expected payoffs are sufficiently higher
than the realized payoffs, i.e. in the incomplete information setting, were the agents to know
each other’s type beforehand, the link may never be formed.
Our next result shows a similar relation between complete and incomplete information when
the formation process converges.
Theorem 2. For every κ, G∗C(κ) ⊂ G∗IC(κ).
Theorem 2 states that if the formation process can converge to some network under complete
information, then it can converge to the same network under incomplete information as well.
Again, the reverse may not be true: under incomplete information, high expected payoffs can
initialize link formation such that even though agents would “regret” the links they form after
knowing each other’s type, more links have formed before they are selected again to update
their initial links. Then due to increasing returns to link formation, the positive externalities
would in turn ensure that the initial links are maintained. The following example illustrates
this point.
Example 1. Assume the following parameter values: N = 5, ki = b ∀i = 1, ..., 5, and the other
parameters are such that f(b, b) < c, fb ≥ c, (1 + δ − δ2 − δ3)f(b, b) ≥ c.
Consider the following selection path: (1, 2), (2, 3), (3, 4), (4, 5), (1, 5). Under complete
information, it is clear that the network remains empty regardless of the selection path, as in
Figure 1(A). Under incomplete information, the SOE can be explicitly computed in each period.
For example, in period 1, agent 1’s expected payoff from the link with agent 2 is fb ≥ c and vice
versa, and thus the link is formed. The formation process is shown in Figure 1(B).
According to the assumptions on parameters, one can then easily show that the network
formed in period 5 is stable.
From the above example, we can see the stark contrast between the two environments:
even if no agent would make any single link under complete information, a connected network
can emerge under incomplete information, and finally be stable due to the sufficient positive
externalities created. With the same setting as Example 1, if we consider a larger group
of agents and an appropriate selection path, the formation process converges to the network
shown in Figure 2(B)8.
8For instance, under the following selection path: (1, 2), (1, 3), (2, 6), (3, 7), (6, , 7), (1, 5), (5, 9), (6, 9), (1, 4),
(4, 8), (7, 8), (8, 9).
12
Figure 1: Empty under Complete Info. vs. Connected under Incomplete Info.
Figure 2: Empty under Complete Info. vs. ”Hub-and-Spokes” under Incomplete Info.
The network exhibits a certain “law of the few”, in the sense that only agent 1 in the
center has the highest connectivity degree. However, unlike in Galeotti and Goyal[12] where
the few agents with high connectivity degree stand out because they are of a more valuable
type (i.e. they are the only ones producing valuable resources), in the above network agents are
identical in type, and some agent has a higher connectivity degree than others merely because
the resulting convergence is history dependent. In other words, even if agents turn out to be
homogeneous, it may well be the case that some agent gets ”lucky” in the selection process
and becomes linked with a considerable number of other agents. In subsequent periods, this
agent attracts even more links to be formed because linking to him grants access to large
indirect benefits. Though occurring by pure chance initially, such a topology in which agents
are homogeneous in type but heterogeneous in connectivity may finally be stable, due to the
existence of increasing returns to link formation.
In fact, such history dependence may generate an even stronger result: even if a type is more
valuable or preferable than the other, under incomplete information it is not necessary that an
agent of that type ends up with a higher connectivity degree. Consider the following example:
assume the same values for fb, f(b, b), c and δ as in Example 1, and in addition that fa ≥ c,
13
f(a, b) ≥ c, (1− δ)f(a, b) < c, and consider a group of agents consisting of 1 type a agent and 8
type b agents. Under complete information, the formation process will only converge to a star
network with the type a agent as the center (Figure 3(A)); under incomplete information, the
formation process may again converge to a ”hub-and-spokes” network (Figure 3(B)).
Figure 3: Different Connectivity Degree Distributions
Under complete information, the type a agent has to be the center of the star network
because the only possible type of link that agents would be willing to form is one between a
type a agent and a type b agent. By contrast, under incomplete information, it first becomes
possible for two type b agents to form a link; and then, as it turns out in this particular topology,
each type b agent’s distance with the type a agent is sufficiently small. Even though the type
a agent has a low connectivity degree, the other agents do not find a new link with the type a
agent attractive, because it does not offer sufficient indirect benefits. Hence, the one agent with
the more valuable type – type a – ends up with the lowest connectivity degree in the network.
This is in stark contrast with the existing results in the literature, which often show that a
more valuable agent is better connected.
Furthermore, under incomplete information, the probability that the formation process con-
verges to a different network topology from those under complete information can be significant.
The following figure shows simulation results9 under an environment with a pure peer effect
(i.e. f(a, a) = f(b, a), f(b, b) = f(a, b)). We make this assumption only to economize on pa-
rameters. Under complete information, no type b agent would ever get linked; under incomplete
information, the probability that some type b agent remains linked by the end rises to more
than 0.8 for a range of the group size N .
The above examples also indicate that, unlike in most existing literature that specifies only
a few types of network topologies as the only possible ones in equilibrium, under incomplete
information the formation process can converge to potentially more types of networks, as there
9We assume for the simulation that δ = 0.8 and p = 0.5. For each value of N , we run 500 simulations and
average the results. Each simulation consists of a fixed long run of 5N(N − 1) periods, so that each link is
selected for 10 times on average.
14
Figure 4: Simulation: Significant Probability of Difference
may exist some network g which can emerge and be stable under incomplete information, but
not under complete information. We define some typical network structures in the literature
and the empirical works below:
1. g is complete if ij ∈ g ∀i, j such that i 6= j.
2. g is a star network if there exists i ∈ I such that ij ∈ g ∀j 6= i, j ∈ I and i′j /∈ g ∀i′, j 6= i.
3. g is a core-periphery network if there exists non-empty I ′ ( I, such that ij ∈ g ∀i, j ∈ I ′, i 6=j, and that ∀j′ ∈ I \ I ′, ij′ ∈ g for some i ∈ I ′ and jj′ /∈ g ∀j 6= i. Note that a star network is
a special case of a core-periphery network.
4. g is a tree network if there exists a partition of I, I1, ..., In, such that (1)#(I1) = 1;
(2)∀n′ = 2, ..., n, each agent in In′ has one and only one link with some agent in In′−1; (3)no
other link exists.10
5. g is a wheel network if there exists a bijection π : I → I such that g = {π−1(1)π−1(2),
π−1(2)π−1(3), ..., π−1(N − 1)π−1(N)}.
Theorem 3. Given κ, if a connected network g is stable and belongs to one of the following
categories:
1. Empty network;
2. Minimally connected network (i.e. tree network, including star network);
10Essentially, a tree network is equivalent to a minimally connected network, and a star network is a special
case of a tree network.
15
Figure 5: Some typical networks
3. Fully connected network;
4. Core-periphery network;
5. Wheel network.
then g ∈ G∗IC.
Theorem 3 explicitly characterizes types of connected networks that can emerge and be
stable under incomplete information, and most typical networks in both the literature and
empirical studies are included. However, note that there may be some stable networks that
cannot emerge under complete or incomplete information – e.g. networks with links such that
(1) the benefit from any one link cannot cover the maintenance cost without the existence of
the other links and (2) the network would still be connected if these links were severed. Such
network topologies may never be formed since only one pair of agents is selected in each period,
and the agents are myopic.
We end this section with an observation when δ = 1, i.e. the value from connection does
not decay with distance.
Proposition 2. Given κ and δ = 1, every g ∈ GIC(κ) is minimal.
Indeed, when there is no decay of indirect benefits, it is never optimal for agents to form
any “redundant” links. Note that according to Theorems 1 and 2, the above result applies to
GC(κ), G∗C(κ) and G∗IC(κ) as well. It also applies to situations where link formation and the
associated costs are directed (one-sided), as long as the benefit flow is undirected (two-sided).
See for instance Zhang and van der Schaar (2013)[35] for such a model which also predicts
minimal equilibrium topologies.
16
4.3 Long-Term Phenomena
We have shown that even though incomplete information only exists for finitely many periods,
its impact on the formation process can be persistent. Therefore, the long-term structure of a
network is determined by two factors: the influence of incomplete information accounts for the
formation, and the payoff characteristics indicate properties of the network once information
becomes complete. In this section, we focus on teasing apart the influences of these factors.
Our findings are in conformity with the empirical works by Newman (2002, 2003)[26][27],
Holme et al. (2004)[16], Hu and Wang (2009)[17] and Fricke et al. (2013)[9], who analyze data
on a variety of social, biological, technological and business networks and show that assortative
mixing (i.e. agents with high connectivity degree tend to link to each other) prevails in some
cases and neutral or disassortative mixing in others. In one of our applications where agents
are divided into subgroups, we propose particular network formation processes that essentially
exhibit disassortative mixing locally and assortative mixing across subgroups. However, we
diverge from the above studies in arguing that besides connectivity degree, assortativity can
also exist in types, which may in turn alter the network topology considerably.
Proposition 3. If types k, k′ (k and k′ may be the same type) are such that min{f(k, k′),
f(k′, k)} ≥ c, then
1. Every type k agent and every type k′ agent are connected for infinitely many periods almost
surely.
2. In every network g ∈ G∗IC(κ), every type k agent is connected to every type k′ agent.
Proposition 3 reveals a pattern of connection over time: as long as two agents both have an
incentive to link to each other when they are not connected, they would be connected (though
may not be linked) repeatedly starting from any period. Moreover, they are connected in every
stable network.
Note that even if min{f(k, k′), f(k′, k)} ≥ c, an existing link between a type k agent and
a type k′ agent may be severed. For instance: suppose fa, fb ≥ c f(a, b), f(b, a) ≥ c, (1 −δ)f(a, a) ≥ c, (1− δ)f(b, a) < c. Consider agents 1, 2, 3 such that k1 = b, k2 = k3 = a and the
selection path (1, 2), (1, 3), (2, 3), (1, 2). By computing the SOE for each period, it is clear that
every link is formed in periods 1− 3. In period 4, the payoff of agent 1 from the link between
agents 1 and 2 is (1− δ)f(b, a) < c, and therefore the link would be severed.
Corollary 1. If different types k, k′ are such that min{f(k, k′), f(k′, k)} ≥ c and there is at
least one agent of each type in I, then every g ∈ G∗IC(κ) is a connected network.
The implication of the above corollary is straightforward: if both types are willing to link
with each other, then connectedness is guaranteed once the network topology becomes stable.
Next, we analyze how self effect and peer effect, and a mixture of the two, influence the
formation process and the resulting network topologies.
Proposition 4. The following are true:
17
1. With pure self effect, given κ and γ(∞), the formation process is the same under complete
information and incomplete information.
2. With pure peer effect, every g ∈ G∗IC(κ) has at most one component.
3. If f(b, b) ≥ c > max{f(a, b), f(b, a)} and there are at least two agents of each different
type, then there exists δ > 0 such that for any δ ∈ [0, δ], every g ∈ G∗IC(κ) has exactly two
components, each one consisting of all the agents of one type.
Different effects impose different influences on the network structure. When an agent’s
payoff depends solely on his own type, information does not play a role and agents know their
exact payoffs as long as they know the pattern of connection; on the other extreme, when agents
only care about others’ types, which can be interpreted as the acquisition of others’ resources,
either there is always incentive to connect between any components, or the whole network
would stay empty in the long run; when both effects are non-trivial, clustering according to
types could emerge and be sustained.
Furthermore, the relative significance of self and peer effects determines the influence of
incomplete information as well. As shown by Proposition 4, the more significant the self
effect is, the smaller the difference between network topologies that result from complete and
incomplete information. At the extreme, where there is pure self effect, the formation processes
under complete and incomplete information, given a fixed selection path, become identical. On
the other hand, the more significant the peer effect is, the more likely that such differences are
large, as illustrated in Example 1.
4.4 Long-Term Connectedness of Large Networks
In this section we discuss asymptotic properties of the formation process, i.e. what types of
networks can emerge as a result of convergence when the group size goes to infinity. We focus
on the ex ante probability, which takes the initial choice of nature on the type vector κ into
account.
Proposition 5. If for some type k, k′ 6= k, min{f(k, k), f(k, k′)} ≥ c and f(k′, k) > (1− δ)c,then
limN→∞
Prob(every g ∈ G∗IC(κ) is connected) = 1
Intuitively, since when f(k, k) > c all the type k agents must be connected in any stable
network, as the group size grows the sub-group of type k agents could be sufficiently large to
“attract” any other agent. In other words, the increasing returns to link formation create incen-
tives for link formation between type k and k′ agents. Comparing this result with Proposition
4, we see that even if clustering occurs almost surely given almost every realization of the type
vector in a group of fixed size, clustering may eventually disappear and a connected network
can be sustained if the group size grows.
18
Our next result proves the existence of an upper bound on the diameter of a network, which
is defined to be the longest distance between any two agents in the network, as the group size
enlarges.
Proposition 6. If for some type k, k′ 6= k, f(k, k) > c and f(k, k′) ≥ c, then there exists
d∗ ∈ N+11 such that
limN→∞
Prob(every g ∈ G∗IC(κ) has a diameter of at most d∗) = 1
In other words, as long as agents of some type tend to cluster ”densely”, i.e. within a
bounded distance, as the group size grows large, agents of the other type would be incentivized
to stay close enough to this clustered sub-group. This “small-world” property has been con-
firmed in both the folk wisdom such as the “six degrees of separation” between any two persons
on the planet, and renowned research papers such as Watts and Strogatz (1998)[33], Newman
and Watts (1999)[28] and Jackson and Rogers (2007)[18].
We would like to emphasize that there can still be significant differences between network
structures under complete information and incomplete information, even in the long run and
asymptotically. Consider for example a case similar to Example 1, where if information were
complete no agent would be willing to form a first link with a type b agent. Even asymptotically,
the type b agents would stay singletons. Yet under incomplete information, if sufficiently many
agents are connected to a type b agent because of the initial incentive of link formation, then
the positive externalities thus created may ensure a type b agent stays connected even in the
long run and in a group of arbitrarily large size.
In the following section, we discuss some of network structures where the diameter of the
network does not increase with N , which enables an agent’s payoff to go up unboundedly as
the group size is enlarged. More concise conditions and more precise assertions can be obtained
as a result.
5 A Special Case: Preferential Attachment
So far we have assumed that opportunities for link formation are completely and uniformly
random. However, it is easy to imagine situations in which link formation is not entirely
random but is at least partially directed. In this section, we construct special selection processes
to illustrate how some particular types of networks could form asymptotically, and how to
determine the corresponding probabilities in the limit. In the selection processes we consider
below, an agent is always paired with the agent with the highest connectivity degree (i.e. the
number of links an agent has) globally or locally. Formally, if a network (or a sub-network) is
empty, then any pair of agents can be selected with equal probability; if there is one link ij,
then either ij is selected, or i (or j) and some agent other than i, j are selected.
11In fact, we have a closed form characterization of this upper bound. Please see the proof of this proposition
in the Appendix.
19
This characterization reflects the idea of preferential attachment in Barabasi and Albert
(1999)[4] and Vasquez (2003)[31], where agents are assumed to prefer to connect with other
well-connected agents. The importance of a “key player” with high centrality is also discussed
in Ballester et al. (2006)[3].
5.1 Centralized Preferential Attachment
We define centralized preferential attachment as a selection process in which any agent can only
be selected together with the agent (or one of the agents) who has the highest connectivity
degree (we call such an agent the center of the group) in the network. The following result
shows that for large N , there is a complete dichotomy.
Proposition 7. Under centralized preferential attachment, define the following events:
A. The formation process converges to a network with a star sub-network. A type a agent is
the center and all type k agents such that f(a, k) ≥ c are periphery agents.
B. The formation process converges to a network with a star sub-network. A type b agent is the
center; all type k agents such that f(b, k) ≥ c are periphery agents.
We have limN→∞ Prob(A) = p, limN→∞ Prob(B) = 1− p.
Now we will further explore the difference between complete and incomplete information in
large networks. Consider the following example.
Example 2. Let the parameters be such that fa ≥ c, f(a, a) ≥ c, f(a, b) < c, fb ≥ c, f(b, a) ≥c, f(b, b) < c. Consider centralized preferential attachment as the selection process. Under
complete information, only a type a agent can be the center (Figure 6(A)), since a type b agent
can never form a single link. Under incomplete information, however, as shown in Proposition
7, the probability of a type a agent being the center and that of a type b agent being the center
go to p and 1− p as N goes to infinity (Figure 6(B)).
Figure 6: Contrast between Complete and Incomplete Info in Large Networks
20
In the above example, even though a star forms almost surely in both environments, it is only
under incomplete information that a ”bad” agent–an agent of type b–may end up as the center,
due to the difference between expected payoff and realized payoff. Even more interestingly, this
example exhibits a certain inverse relation between type and connectivity degree: in any stable
network among a sufficiently large group of agents, the only possible place for a non-singleton
“bad” agent is the center of a star – even though the agent himself provides less value to others
than the periphery agents he enjoys the best-connected position in the whole network, and no
agent is willing to unilaterally sever their link with him!
5.2 Regional Preferential Attachment
Assume that the group of agents can be partitioned into M subgroups, with each group con-
taining n agents. One may regard such an environment as one with hierarchy in communica-
tion: preferential attachment only occurs locally such that periphery agents can only link with
the agent with the highest connectivity degree in their corresponding subgroup. We define
regional preferential attachment as a selection process in which (1) only the agent(s) with the
highest discounted connectivity degree in his own subgroup (we call such an agent the center of
the corresponding subgroup) can be selected together with an agent from another subgroup; (2)
any non-center agent in a subgroup can only be selected with the corresponding center agent.
As before, we fully characterize the resulting network.
Proposition 8. Under regional preferential attachment, define the following event for integer
m ∈ [0,M ]:
A(m). The formation process converges to a network with a core-periphery sub-network. m
type a agents and M −m type b agents are in the core, and in any subgroup with a type a (b)
center, all type k agents such that f(a, k) ≥ c (f(b, k) ≥ c) are periphery agents.
We have limn→∞ Prob(A(m)) =(Mm
)pm(1− p)M−m.
Furthermore, as we have emphasized before, network formation in our model exhibits strong
history dependence, which can be typically represented by different selection rules. Consider
the following example.
Example 3. Consider the following environment: fa, fb, f(a, a), f(b, b) ≥ c, f(b, a)+f(b, b) nδ1−nδ , f(a, b)+
f(a, a) nδ1−nδ < c for some positive integer n > 1. We will show by the following example that
under one selection rule, there is a positive probability that some type a and type b agents are
connected in a large stable network, whereas under the other selection rule, such an event is
impossible no matter how large the network is.
Consider the following two selection rules:
1. Regional preferential attachment: By Proposition 8, we know that the formation process
converges to a core-periphery sub-network almost surely. According to the payoff characteristics,
in each subgroup the periphery agents are of the same type as the center agent. Since the
21
probability that there are both type a and type b agents as centers is positive, we know that with
positive probability some type a and type b agents (the number of such agents goes to infinity
as the group size goes to infinity) are connected. Moreover, some agents of each type are left
as singletons with positive probability.
2. Anti-preferential attachment: Assume that no agent can be selected with any agent with more
than n− 1 links (i.e. with connectivity degree of more than n− 1). By an argument similar to
the proof of Proposition 4, we can show that in any stable network of any size no type a agent
can be connected to any type b agent. Moreover, no agent would be left as singleton almost
surely, as long as there are at least two agents of each type.
6 Discussions
6.1 Alternative Updating Rules
The results we have derived so far are based on the simple updating rule, which assumes that
every agent’s posterior belief on another agent’s type is binary: either it is the degenerate
belief on the true type, or the prior. Note again that such an updating rule implicitly assumes
that agents can only observe their own formation history. If the agents adopt a different
updating rule, which reflects either more or less available information, the formation process
can exhibit a much different pattern. We discuss one such alternative in detail, which we call
Bayesian learning by formation history.
We assume that agents can observe the entire formation history, i.e. the pair of agents
selected and the resulting network structure each period, in addition to knowing the types
of agents connected to themselves. They then apply Bayesian updating in forming posterior
beliefs. Formally, let p′ij(σ(t)) denote the posterior probability in i’s belief after σ(t) that j is
of type a. The following result highlights the key difference between the simple learning rule
and this alternative.
Proposition 9. Under Bayesian learning by formation history, assume that f(b, b) < c and
that there is a pure peer effect12. Then for any type a agent i, there is a sufficiently large nb
and a formation history σ(t) for some t such that p′ji(σ(t))f(kj, a) + (1− p′ji(σ(t)))f(kj, b) < c
for any other agent j and any kj ∈ {a, b}.
A major implication here is that Bayesian learning by formation history makes it possible
for the posterior probability of an agent being of high type to fall close to 0. As a result, even if
making a link with some agent i is incentivized with the simple learning rule, it may no longer
be the case under Bayesian learning by formation history, given some particular selection path
that would drag posterior beliefs towards i being of type b. The following example illustrates
this difference.
12It is easy to check that, with pure peer effect, the SOE still exists and is unique under Bayesian learning
by formation history.
22
Example 4. Assume the following parameter values in an environment with pure peer effect
(i.e. f(a, a) = f(b, a), f(a, b) = f(b, b)): N = 5, k1 = k2 = a, k3 = k4 = k5 = b, and the
other parameters are such that fa = fb ≥ c, f(b, b) = f(a, b) < c, (1 + δ − δ2 − δ3)f(b, b) ≥ c,p(1−p)1−p2 f(a, a) + p(1−p)+(1−p)2
1−p2 f(a, b) < c. Consider the following selection path in period 1-9:
(1, 3), (1, 3), (2, 4), (2, 4), (1, 2), (3, 4), (4, 5), (2, 3), (1, 5). Note that in this environment with
pure peer effect, the existence of a SOE (and thus the uniqueness) is still valid under Bayesian
learning by formation history, as any agent’s expected payoff from any link is not affected by
his own type.
Under the simple learning rule, the formation process is shown in Figure 7(A). The network
formed in period 9 is stable; under Bayesian learning by formation history, the formation process
is shown in Figure 7(B). The network formed in period 4 is stable.
Figure 7: Connected under Simple Updating Rule vs. Empty under Bayesian Learning by
Formation History
Here with the simple learning rule, agents hold the prior belief each time they are selected
with another agent with an unknown type, and thus the given selection path induces a connected
network at last. Yet with Bayesian learning by formation history, each agent updates from their
observation to conclude that others are of type b with a sufficiently large probability, and thus
are unwilling to make any link.
One implication of the above proposition and example is that more learning can sometimes
be “bad”, i.e. it may lead to inefficient outcomes. Despite the specific differences brought about
by an alternative updating rule, our general results still hold under a range of parameters. For
instance, it can be shown that in an environment with pure peer effect, if any typical network
as depicted in Theorem 3 can emerge and be stable under complete information, then it can
under incomplete information and Bayesian learning by formation history as well.
23
6.2 Robustness to Link Failure
Consider the following variation of the model: in each period, before a pair of agents is selected
to update the (potential) link between them, each existing link may fail independently with
probability ε ∈ (0, 1). One could think of this scenario as introducing an exogenous environ-
mental error, such as a possible system breakdown for a social network website or software.
One particular influence of possible link failure on the formation process is that some stable
network is not “robust”, in the sense that if some link fails by accident, the formation process
will never converge to the previous network again. The following example illustrates this point.
Example 5. Assume the same parameter values and the same selection path as in Example
1. We know that if there were no link failures, the connected network formed in period 5 is
stable. If there is a positive probability of link failure, and in some period two adjacent links
(for instance the link between agents 1 and 2, and that between 2 and 3) of the five fail, then
since agents know each other’s types already, the same connected network can never be formed
thereafter. Moreover, given that f(b, b) < c, any network with some agent with only one link
is never stable because the link would be severed once selected for update. As a result, the
formation process would converge to an empty network almost surely.
The intuition behind this phenomenon is simple. Should there be no incomplete information,
no link would ever form between any two agents; the connected network emerges due to high
expected payoffs from potential links, and without link failure it is stable because each link carries
sufficient indirect benefits. Now if two adjacent links fail, which leaves one agent a singleton, it
is clear that no agent would be willing to link to the singleton agent again. Moreover, there is
no network structure among the remaining four type b agents that could sustain enough indirect
benefit to maintain any one link.
From the above example it is rather obvious that a stronger concept for a stable net-
work is needed in the presence of link failure. One feasible candidate is what we denote as
a robust stable network: a network g is a robust stable network if (1) g is a stable network
without link failure; (2) with link failure, given any number of failing links in any period, g
will emerge again in some future period almost surely. The second condition guarantees that
no matter how serious the link failure is, g would be formed again in the future, and hence the
name “robust stable network”. Formally, let G∗∗C (κ) and G∗∗IC(κ) denote the set of robust stable
networks that can emerge and be stable under incomplete information given the type vector κ.
Proposition 10. Given any κ, G∗∗IC(κ) = G∗∗C (κ) = G∗C(κ).
We see from the above result that once robustness is imposed as an additional constraint,
there is no essential difference between complete and incomplete information regarding the
resulting convergence under possible link failure. Note that condition (2) in the definition of
a robust stable network is a strong condition implying that a robust stable network does not
always exist, and it remains a future research direction to characterize the network formation
process under possible link failure in detail.
24
7 Conclusion and Future Research
In this paper we analyzed the network formation process under agent heterogeneity and in-
complete information. Our results are in stark contrast with the existing literature: instead
of restricting the equilibrium network topologies to fall into one or two specific categories, our
model generates a great variety of network types. Besides what networks can emerge as a result
of convergence, we argue that it is also important to understand how a network gets formed,
since we want to know, for instance, why some agents become central and others do not. While
link formation and belief formation are usually treated as two independent processes to be
studied separately, we combine them in our model and show that belief formation is in fact a
key factor that could facilitate or deter link formation. Even if incomplete information vanishes
in the long run, its impact on shaping the network topology is persistent.
Several future research topics can be built up on the basis of our model. One natural
extension of the model is to assume that there are more than two, or even a continuum of types
of agents. As long as the initial incentive to form links is guaranteed for each type, the contrast
between complete and incomplete information, as reflected by Proposition 1 and Theorems 1-3,
would remain. If the agents adopted a different updating rule such as Bayesian learning by
formation history, then more types would induce a more complicated updating process and
results would also be different from those in the two type environment. In terms of the network
structure, adding more types creates more possibilities for the realized payoffs, and thus adds to
the variety of emerging topologies. For example, a stable network may have more components,
or the pattern of clustering across types may be more complicated.
Another future research topic is to pin down the structure of an efficient network and
implement it in a game-theoretic setting. The usual definition on efficiency in networks adopted
in the literature is strong efficiency, i.e. a network is strongly efficient if it maximizes the sum
of agents’ payoffs. In general, we know that a strongly efficient network must exist (though
not necessarily be unique) because the set of possible network structures is finite. By Theorem
2, we also know that more networks can emerge and be stable under incomplete information,
so maximal social welfare is at least as large as that under complete information. However,
since payoffs are heterogeneous across agents according to the type vector, the exact topology
of an efficient network becomes difficult to characterize; moreover, the efficient network may
not be unique because in an agent-heterogeneous environment there could be multiple ways of
generating the same level of social welfare.
Most importantly, we have assumed throughout, as does most of the literature, that agents
are myopic rather than forward-looking. If it is assumed otherwise that agents are foresighted
and are concerned about both their current and future welfare, then the aim of analysis es-
sentially becomes solving an agent’s dynamic optimization problem in the presence of other
similarly foresighted agents. One can then easily anticipate a very different evolution pattern
of network topologies as well as very different stable network topologies in the limit, for now
link formation does not only serve as an action of maximizing the current expected payoff, but
also as a way of acquiring information for potential future benefit. It is our conjecture that
25
link formation would be more frequent at least in an early stage of the formation process, since
on top of the incentives already discussed in this paper, agents may also be willing to endure
current payoff losses in return for valuable information about the type vector.
26
Appendix
Proof of Lemma 1. Assume that agents i and j are selected in some period. Based on the
simple updating rule, either i and j know each other’s type, or their belief on each other’s type
remains at the prior. If they know each other’s type, then since actions are binary for each
agent, the strategy described in a SOE exists and is unique for each agent. It is also clear that
the strategy is a best response for each agent since it is a weakly dominant strategy.
If the agents do not know each other’s type, and hence their belief remains at the prior,
then by Assumption 2 i’s expected payoff from the link with j is non-negative regardless of
i’s type, and vice versa. Therefore, the strategy profile described in a SOE forms mutual best
responses for i and j and thus is indeed an equilibrium. Finally, it is clear that such a profile
is unique.
Proof of Proposition 1. 1: It suffices to show that any two agents are connected for at least
one period within finitely many periods almost surely. By Assumption 2, it further suffices to
show that any two agents are selected at least once within finitely many periods almost surely.
Consider any two agents i and j; the probability of the event that they are not selected in one
period is 1 − 2N(N−1) < 1, and thus the probability of this event occurring for infinitely many
periods is 0.
2: If g is connected, then clearly there is complete information. If g is unconnected and
information is not complete, then there must exist two unconnected agents such that their
beliefs on each other’s type remain at the prior. When they are selected, by Assumption 2 they
would form a link, which implies that g, B is not stable, a contradiction.
Proof of Theorem 1. For any g ∈ GC(κ):
If g is empty: since g ∈ GC(κ), there must exist two agents i, j such that f(ki, kj) < c or
f(kj, ki) < c. Consider the selection path γ(2) = ((i, j)1, (i, j)2) under incomplete information.
It is clear that a link would form between i and j in period 1, but the link would then be
severed in period 2, and thus g = g(γ(2)), which implies that g ∈ GIC(κ).
If g is non-empty: consider any selection path γC(t) such that g emerges for the first time in
period t. By the definition of GC(κ), we know that such γC(t) exists. Let γIC(t′) be a selection
path constructed from γC(t) such that the pairs of agents in γC(t) between whom there is no
existing link, but a new link is not formed either, are deleted.
Consider the formation process under incomplete information given γIC(t′). By Assumption
2 and the simple updating rule, we know that given the same network structure, if a link is
formed between i and j under complete information, it will also be formed under incomplete
information regardless of whether i and j know each other’s type. Also, it is clear that the
decision of severing a link by any agent is the same under complete information and incomplete
information, since such a decision is based on the realized payoff. Therefore, the formation
process yields the same link formation and severance results under complete information given
γC(t) and under incomplete information given γIC(t′). Hence given γIC(t′), g emerges for the
27
first time in period t′ under incomplete information, which implies that g ∈ GIC(κ). Therefore
GC(κ) ∈ GIC(κ).
Proof of Theorem 2. By Theorem 1, we already know that any network that can emerge
under complete information can also emerge under incomplete information. Thus it suffices to
show that, for any network g that can emerge and be stable under complete information, there
exists a subsequent selection path under incomplete information after g’s first appearance that
would make g stable. We prove the result by construction.
If g is empty: consider the selection path γ(N(N−1)2
) such that every pair of agents is selected
exactly twice consecutively. Since by assumption g ∈ G∗C(κ), we know that for every pair of
agents a link would first form and then be severed in the next period. In period N(N−1)2
+ 1,
information is complete and the empty network becomes stable.
If g is non-empty: denoting the number of components in g as q(g), under incomplete in-
formation let the subsequent selection path after g’s first appearance be such that in the first
q(g)(q(g)− 1) periods, two agents from different components are selected exactly twice consec-
utively and every two components are involved. Since by assumption g ∈ G∗C(κ), which means
that g is stable under complete information, when a pair of agents from different components is
selected for the second time, either there is no existing link between them and no link would be
formed, or an existing link would be severed. In either case, the agents know each other’s type
as well as the types of agents in the counterparty’s component. Therefore, after q(g)(q(g)− 1)
periods, every agent knows κ and essentially there is no incomplete information. Again by the
assumption that g ∈ G∗C(κ), we can conclude that g is such that no link would be formed or
severed in any later period given any selection path. Thus g ∈ G∗IC(κ), which completes the
proof.
Proof of Theorem 3. By the assumption that g is connected and stable, it suffices to show
that when g belongs to any of the categories there exists a selection path such that g can
emerge in the formation process. We discuss case by case and prove by construction below.
1: See the proof of Theorem 2.
2: let L be the total number of links in g. Let the selection path be such that the pair of
agents for each link in g is selected once and only once in the first L periods. By Assumption
2, we know that each link will be formed, and thus g emerges in period L.
3: Since g is fully connected and stable, we know that for any two agents i and j, δf(kikj) ≥c. Therefore regardless of the selection path g would emerge.
4: Let the selection path be such that: first each periphery agent is selected once and only
once with their corresponding core agent, then every two core agents are selected once and only
once before any other pair of agents is selected. By Assumption 2 and the assumption that g
is stable, we know that each link will be formed, and thus g emerges after the last pair of core
agents is selected.
5: Let the selection path be such that the pair of agents for each link in g is selected once
and only once in the first N −1 periods. By Assumption 2 and the assumption that g is stable,
28
we know that each link will be formed, and thus g emerges in period N − 1.
Proof of Proposition 2. For any non-minimal network g, consider agents i and j such that
ij ∈ g and ig−ij↔ j. For g to emerge, either ij is formed before some i′j′ belonging to the path
connecting i and j in g − ij, or such i′j′ is formed before ij. In the first case, since δ = 1,
the formation of i′j′ would only bring additional costs to i′ and j′, but no additional benefits,
and thus i′j′ would never be formed. A similar argument would show that ij would never be
formed in the second case. Therefore, we can conclude that g /∈ GIC(κ), which completes the
proof.
Proof of Proposition 3. Without loss of generality we assume that types k and k′ are dif-
ferent. The case where they are the same can be proved analogously.
1: Consider any type k agent i and any type k′ agent j. By Assumption 2, when i and j are
selected for the first time a link would be formed between them. After that, whenever i and j
are selected and they are not connected, by the assumption that min{f(k, k′), f(k′, k)} ≥ c, a
link would be formed between them. Then we know that for any m,n ∈ N+, n > m:
Prob(i and j are connected for m periods before period n, and unconnected thereafter)
≤Prob(i and j are unconnected after period n)
≤Prob(i and j are never selected after period n)
=(N(N − 1)− 2
N(N − 1))∞ = 0
which implies that
Prob(i and j are connected for m periods before period n, and unconnected thereafter) = 0
Therefore, we have
Prob(i and j are connected for finitely many periods)
≤∞∑m=0
∞∑n=1
Prob(i and j are connected for m periods before period n, and unconnected thereafter)
=∞∑m=0
∞∑n=1
0 = 0
We can then conclude that the probability that i and j are connected for finitely many periods
is 0, so i and j are connected for infinitely many periods almost surely.
2: From the above argument, whenever a type k agent and a type k′ agent are selected, they
would form a link. Therefore a network with unconnected type k and type k′ agents is never
stable, which implies that the formation process would never converge to such a network.
Proof of Corollary 1. Consider any two unconnected agents i and j in any unconnected
network g. If i and j are of different types, then according to Proposition 3 we know that they
have an incentive to form a link once selected, and thus g /∈ G∗IC(κ). If they are of the same
29
type, without loss of generality assumed to be type k, then since there is at least one agent of
each type, by Proposition 3 we know that for g to be stable there exists some agent i′ of type
k′ who is connected to either i or j. Without loss of generality, assume that i′g↔ i. But then
if i′ and j are selected, again by Proposition 3 they have an incentive to form a link, which
implies that g /∈ G∗IC(κ). This completes the proof.
Proof of Proposition 4. 1: Clear, since with pure self effect the realized payoff coincides
with the expected payoff.
For 2 and 3, by Lemma 2 we can restrict our attention to the case of complete information
without loss of generality.
2: We prove by contradiction that a network with more than one component is never stable,
which is a sufficient condition for 2. Suppose that a network with two components C1, C2 is
stable. If there is a type a agent in C1 and a type a agent in C2, since fa ≥ c by Assumption
2 and f(a, a) ≥ c, when they are selected they would form a link regardless of whether they
know each other’s type. Thus without loss of generality we only need to discuss the case where
there are only type b agents in C1. For any agent i ∈ C1, the assumption that some other agent
j ∈ C1 is willing to link to i in a stable network together with f(b, a) = f(a, a) ≥ c implies
that any agent l outside C1 is willing to link to i whether or not he knows i’s type, as both
l’s expected payoff and realized payoff from a link with i are strictly larger than j’s realized
payoff from the link with i. The same argument would establish that any agent outside C2 is
willing to link to any type b agent in C2. Finally, it is clear that any agent outside C2 is willing
to link to any type a agent in C2, since both the expected payoff and the realized payoff are
non-negative. Therefore, there is always one agent in C1 and one agent in C2 who are willing
to form a link, and thus a network with more than one component is never stable.
3: Let δ = c−max{f(a,b),f(b,a)}(N−2)f(a,a) > 0. We first prove by contradiction that a network with at
most one component is never stable. By Assumption 2 and the assumption that f(a, a) ≥f(b, b) ≥ c, it is clear that an empty network is never stable, and thus we only need to discuss
the case of a network with only one component. Since there are at least two agents of each
type, the largest possible payoff that a type a agent gets by linking to a type b agent is
(1 + δ)f(a, b) + (N − 3)δf(a, a). By the above construction, we know that this payoff is strictly
less than c. A similar argument would show that the largest possible payoff that a type b agent
gets by linking to a type a agent is also strictly less than c. Therefore, suppose that a network
with only one component is stable, then the component can only consist of one type of agent.
Without loss of generality assume that it consists of only type a agents. Since fb ≥ c and
f(b, b) ≥ c, we know that two type b agents would form a link, regardless of beliefs. Therefore
a network with only one component, which consists of only type a agents, cannot be stable,
which is a contradiction.
Now assume that the network formation process converges. We know that in any stable
network there cannot be a link between any type a agent and type b agent; also, since there
are at least two agents of each type, there cannot be any singleton agent. Therefore any stable
network is partitioned by components consisting only of type a or type b agents. Suppose
30
that there are two components consisting of the same type of agent; we know from the above
construction that any agent in one component would be willing to link with any agent in the
other, regardless of beliefs. Therefore such a network cannot be stable, which implies that in
any stable network there is exactly one component consisting of all the type a agents, and
another component consisting of all the type b agents.
Proof of Proposition 5. Without loss of generality we only consider the case of complete
information. First we find N ′ such that∑N ′
n=1 δn−1f(k′, k) ≥ c. By the assumption that
f(k′, k) > (1− δ)c we know that such N ′ exists. Then for any ε > 0, we find N∗ such that for
all N ≥ N∗, the ex ante probability that there are at least N ′ type k agents in the group is
at least 1 − ε. By the assumptions that p ∈ (0, 1) and agents’ types are i.i.d., we know that
such N∗ exists. We assume in the remaining part of the proof that there are at least N ′ type
k agents in the group.
Assume that the network formation process converges. The construction of N ′ is such that
the payoff of a type k′ agent from connecting to a subgroup of N ′ connected type k agents is
non-negative. Since f(k, k) ≥ c, in any stable network all type k agents must be connected.
Then since there are at least N ′ connected type k agents in the group, we know that each type
k′ agent must be connected to the subgroup of type k agents. Therefore, any stable network
must be connected, which completes the proof.
Proof of Proposition 6. We prove the result in the following steps. Without loss of gener-
ality we only consider the case of complete information.
Step 1: Define d′ := min{d : (1− δd−1)f(k, k) ≥ c} − 1. By the construction of d′, we know
that if two type k agents with a distance of at least d′ + 1 are selected, they both have the
incentive to form a link. It immediately follows that for any κ, in any g ∈ G∗IC(κ), the distance
between any two type k agents is at most d′.
Step 2: We find N ′ such that (1− δd′)f(k′, k) + (N ′ − 1)(δd′ − δd′+1)f(k′, k) ≥ c. Then for
any ε > 0, we find N∗ such that for all N ≥ N∗, the ex ante probability that there are at least
N ′ type k agents in the group is at least 1− ε. By the assumptions that p ∈ (0, 1) and agents’
types are i.i.d., we know that such N∗ exists. We assume in the remaining part of the proof
that there are at least N ′ type k agents in the group.
Consider a sub-group of N ′ type k agents, among which the distance between any two agents
is at most d′, and consider any type k′ agent i. Let j be the type k agent in the sub-group
whose distance with i is the shortest, and suppose that dij = d′ + 2. If i and j are selected, for
j it is clear that he has an incentive to form a link since f(k, k′) ≥ c; for i, a link with j would
generate an additional benefit of (1 − δd′)f(k′, k) from j, and an additional benefit of at least
(δd′ − δd′+1)f(k′, k) from every other type k agent in the sub-group. By the construction of N ′,
we know that i also has an incentive to form a link. According to the conclusion in step 1, we
know that in any g ∈ G∗IC(κ), for any type k′ agent, the shortest distance between himself and
any type k′ agent is at most d′ + 1. It then follows that the distance between any type k agent
and any type k′ agent is at most 2d′ + 1.
31
Step 3: Consider any two type k′ agents i′ and j′, and let i′′ and j′′ denote the type k
agents with the shortest distance to i′ and j′ respectively. From the above two steps, we
know that in any g ∈ G∗IC(κ), di′i′′ ≤ d′ + 1, dj′j′′ ≤ d′ + 1 and di′′j′′ ≤ d′, which imply
di′j′ ≤ di′i′′ + di′′j′′ + dj′j′′ = 3d′ + 2. Therefore d∗ = 3d′ + 2 is one of the desired upper bounds
on the network diameter. This completes the proof.
Proof of Proposition 7. We first assume that two agents of the same type are selected in
the first period. The case where two agents of different types are selected can be analyzed anal-
ogously. As N →∞, in the second period almost surely an agent among the two (denote this
agent as i) and some singleton agent are selected. LetN∗ = max{ c−min{f(a,a),f(b,b),f(a,b),f(b,a)}δmin{f(a,a),f(b,b),f(a,b),f(b,a)} }+1.
As N →∞, the probability that apart from i there are at least N∗+1 agents of type ki goes to
1. Hereby we prove that the probability that at least N∗ + 1 agents of type ki would be linked
to i before any of the existing links between i and type ki agents is selected to be updated goes
to 1 as N → ∞. For notational convenience, let P (n) denote the probability that n agents of
type ki are linked to i, and let P ′(n+ 1|n) denote the probability that given n agents of type ki
are linked to i, another agent of type ki is linked to i before any of the existing n links is ever
selected to be updated.
We know that limN→∞ P (1) = 1 since fki ≥ c and p ∈ (0, 1). Given that n agents of type
ki are linked to i and none of the existing n links is ever selected to be updated, we know that
almost surely i and another agent of type ki would be selected in some period in the future;
given this event, the ratio of the probability that a singleton type ki agent is selected with i,
and the probability that a non-singleton type ki agent (i.e. one that has already been linked to
i) is selected with i, ismax{nki
−n−1}n
. The value of this ratio goes to∞ almost surely as N →∞,
and thus limN→∞ P′(n+ 1|n) = 1 ∀n > 0. Now we discuss the following two cases for k′ 6= ki:
(1) If min{f(ki, ki), f(ki, k′)} ≥ c: by the construction of N∗, once i is linked with at least
N∗ + 1 agents, then every new link would be formed and no existing link would be severed.
From the above analysis, the probability that at least N∗ + 1 agents would be linked to i is
bounded below by P (1)×N∗i=1P′(i+1|i), which goes to 1 as N →∞. Thus the formation process
would converge to a star network.
(2) If min{f(ki, ki), f(ki, k′)} < c: without loss of generality we only discuss the case where
f(ki, ki) < f(ki, k′). From the assumption fki ≥ c we know that f(ki, ki) < c and f(ki, k
′) > c.
From above we know that with probability going to 1 as N → ∞, at least N∗ + 1 agents of
type ki would be linked to i before any of the existing links between i and type ki agents is
selected to be updated. By the same argument, with probability going to 1 as N →∞, at least
N∗ + 1 agents of type k′ 6= ki would be selected before the existing links between i and type ki
agents is selected to be updated, and a link would be formed upon each selection. Afterwards,
by the construction of N∗, every link between i and a type k′ 6= ki agent would be formed,
and since f(ki, ki) < c every link between i and a type ki agent would be severed. Thus the
formation process would converge to a star sub-network on the set containing i and all type
k′ 6= ki agents, and all other type ki agents would remain singletons.
Finally, note that by the Glivenko-Cantelli Theorem, the empirical distribution of the agents’
32
types converges uniformly to the prior almost surely as N →∞. Then we know that as N →∞,
(1) a pair of type a agents is selected in the first period with probability converging to p2, (2) a
pair of type b agents is selected in the first period with probability converging to (1− p)2, (3) a
type a agent and a type b agent are selected in the first period with probability converging to
2p(1− p), and (4) given that a type a agent and a type b agent are selected in the first period,
each is selected in the next period with equal probability.
It is clear that asN →∞, in the second period a new agent other than the two selected in the
first period would be selected almost surely. Then with probability converging to p2+ 2p(1−p)2
= p
((1− p)2 + 2p(1−p)2
= 1− p), a star with a high (low) type agent as the center and two periphery
agents would be formed. Then we can apply the above argument to show the result.
Proof of Proposition 8. Let im denote the center agent of subgroup m. By Proposition 7 we
know that as n→∞, with probability converging to 1 a star sub-network would form in each
subgroup. Given that, when there are sufficiently many periphery agents in subgroup m (which
happens with probability converging to 1 as n→∞), we know that with probability converging
to 1 as n→∞: (1) all other center agents would form a link with im if selected; (2) a star sub-
network would emerge in each group; (3) any two unlinked center agents would form a link once
selected; (4) any type k′ agent in group m′ such that f(kim′ , k′) < c would ultimately remain a
singleton. Finally, the limiting probabilities are derived using Glivenko-Cantelli Theorem.
Proof of Proposition 9. We prove the result by construction. Consider the selection path
that i is selected twice consecutively with a type b agent, then selected twice consecutively
with another type b agent, and so on. We know that initially a link forms and then breaks
each time i is selected with a different type b agent. Let p′m be the posterior probability for
some other type a agent j that i is a type a after the link between i and the mth type b
agent breaks. We know that by Bayesian updating, p′m+1 = p′m(1−p)1−p′mp
with the initial condition
p′0 = p. Thereforep′m+1
p′m= 1−p
1−p′mp≤ 1−p
1−p2 < 1, and thus there exists nb such that p′nbsatisfies
p′nbf(k, a) + (1− p′nb
)f(k, b) < c ∀k ∈ {a, b}.
Proof of Proposition 10. We first prove that G∗∗IC(κ) = G∗∗C (κ). Given any κ, it is first clear
that G∗∗C (κ) ⊂ G∗∗IC(κ), since by Proposition 1 we know that information must be complete in
any stable network under incomplete information.
Consider any network g ∈ G∗∗IC(κ). Again by Proposition 1 we know that information must
be complete in g. By the definition of a robust stable network, g can still emerge even if each
link in g fails, which implies that g can emerge from an empty network. Therefore g ∈ G∗∗C (κ),
and thus G∗∗IC(κ) ⊂ G∗∗C (κ). Therefore, G∗∗IC(κ) = G∗∗C (κ).
Now we prove that G∗∗C (κ) = G∗C(κ). By the definition of G∗∗C (κ), it is clear that G∗∗C (κ) ⊂G∗C(κ). For any g ∈ G∗C(κ), fix a large enough number of periods T such that there is a
positive probability PT that starting with an empty network, g emerges at least once within
T periods. Starting with some period that any link(s) fails in g, we know that almost surely
the network will become empty in some future period; starting from then, there is a positive
probability PT that g emerges within T periods; if that does not happen, which is an event of
33
probability 1− PT , again almost surely the network will become empty in some future period,
and an identical argument follows. Therefore, the probability of g emerging again in some
future period is bounded below by∑∞
i=0 PT (1 − PT )i = 1. Thus by the definition of G∗∗C (κ),
g ∈ G∗∗C (κ), which implies that G∗C(κ) ⊂ G∗∗C (κ). This completes the proof.
34
References
[1] Acemoglu, D., Dahleh, M. A., Lobel, I., and Ozdaglar, A. Bayesian learning
in social networks. NBER Working Papers (2008), 14040.
[2] Bala, V., and Goyal, S. A noncoorperative model of network formation. Econometrica
68, 5 (2000), 1181–1229.
[3] Ballester, C., Calvo-Armengol, A., and Zenou, Y. Who’s who in networks.
wanted: The key player. Econometrica 74, 5 (2006), 1403–1417.
[4] Barabasi, A.-L., and Albert, R. Emergence of scaling in random networks. Science
286 (1999), 509–512.
[5] Corbae, D., and Duffy, J. Experiments with network formation. Games and Economic
Behavior 64, 1 (2008), 81–120.
[6] Deroian, F. Farsighted strategies in the formation of a communication network. Eco-
nomic Letters 80 (2003), 343–349.
[7] Dutta, B., Ghosal, S., and Ray, D. Farsighted network formation. Journal of
Economic Theory 122 (2005), 143–164.
[8] Falk, A., and Kosfeld, M. It’s all about connections: Evidence on network formation.
CEPR Discussion Papers, 3970 (2003).
[9] Fricke, D., Finger, K., and Lux, T. On assortative and dissasortative mixing in
scale-free networks: The case of interbank credit networks. Kiel Working Paper (2013),
1830.
[10] Gale, D., and Kariv, S. Bayesian learning in social networks. Games and Economic
Behavior 45 (2003), 329–346.
[11] Galeotti, A. One-way flow networks: The role of heterogeneity. Economic Theory 29,
1 (2006), 163–179.
[12] Galeotti, A., and Goyal, S. The law of the few. The American Economic Review
100, 4 (2010), 1468–1492.
[13] Galeotti, A., Goyal, S., and Kamphorst, J. Network formation with heterogeneous
players. Games and Economic Behavior 54 (2006), 353–372.
[14] Goeree, J. K., Riedl, A., and Ule, A. In search of stars: Network formation among
heterogeneous agents. Games and Economic Behavior 67, 2 (2009), 455–466.
[15] Haller, H., and Sarangi, S. Nash networks with heterogeneous agents. Discussion
Papers of DIW Berlin (2003), 337.
35
[16] Holme, P., Edling, C. R., and Liljeros, F. Structure and time evolution of an
internet dating community. Social Networks 26 (2004), 155–174.
[17] Hu, H., and Wang, X. Disassortative mixing in online social networks. European
Physics Letters (2009), 18003.
[18] Jackson, M. O., and Rogers, B. W. Meeting strangers and friends of friends: How
random are social networks? The American Economic Review 97, 3 (2007), 890–915.
[19] Jackson, M. O., and Watts, A. The evolution of social and economic networks.
Journal of Economic Theory 106 (2002), 265–295.
[20] Jackson, M. O., and Wolinsky, A. A strategic model of social and economic networks.
Journal of Economic Theory 71 (1996), 44–74.
[21] Johnson, C., and Gilles, R. P. Spatial social networks. Review of Economic Design
5, 3 (2000), 273–299.
[22] Koenig, M. D., Tessone, C. J., and Zenou, Y. Nestedness in networks: A theoretical
model and some applications. Stanford Institute for Economic Policy Research Discussion
Papers (2012), 11–005.
[23] Leung, M. Two-step estimation of network-formation models with incomplete informa-
tion. Working Paper (2013).
[24] McBride, M. Imperfect monitoring in communication networks. Journal of Economic
Theory 126 (2006), 97–119.
[25] Mele, A. A structural model of segregation in social networks. NET Institute Working
Papers, 10-16 (2010).
[26] Newman, M. E. J. Assortive mixing in networks. Phisics Review Letters 89 (2002),
208701.
[27] Newman, M. E. J. Mixing patterns in networks. Phisics Review E 67 (2003), 026126.
[28] Newman, M. E. J., and Watts, D. J. Scaling and percolation in the small-world
network model. Physical Review E 60, 6 (1999), 7332–7342.
[29] Page Jr., F. H., Wooders, M. H., and Kamat, S. Networks and farsighted stability.
Warwick Ecomomic Research Papers, 689 (2003).
[30] Rong, R., and Houser, D. Emergent star networks with ex ante homogeneous agents.
GMU Working Paper in Economics, 12-33 (2012).
[31] Vazquez, A. Growing network with local rules: Preferential attachment, clustering hier-
archy, and degree correlations. Physical Review E 67 (2003), 056104.
36
[32] Watts, A. A dynamic model of network formation. Games and Economic Behavior 34
(2001), 331–341.
[33] Watts, D. J., and Strogatz, S. H. Collective dynamics of ’small world’ networks.
Nature 393 (1998), 440–442.
[34] Zhang, Y., and van der Schaar, M. Information production and link formation in
social computing systems. IEEE Journal on Selected Areas in Communications 30, 11
(2012), 2136–2145.
[35] Zhang, Y., and van der Schaar, M. Strategic networks: Information dissemina-
tion and link formation among self-interested agents. IEEE Journal on Selected Areas in
Communications 31, 6 (2013), 1115–1123.
37