Bayesian Learning in Social Networks∗

Daron Acemoglu,†Munther A. Dahleh,‡Ilan Lobel,§and Asuman Ozdaglar¶

January 14, 2010

Abstract

We study the (perfect Bayesian) equilibrium of a model of learning over a general

social network. Each individual receives a signal about the underlying state of the world,
observes the past actions of a stochastically-generated neighborhood of individuals, and
chooses one of two possible actions. The stochastic process generating the neighbor-
hoods defines the network topology (social network). We characterize pure-strategy
equilibria for arbitrary stochastic and deterministic social networks and characterize the
conditions under which there will be asymptotic learning—convergence (in probability)
to the right action as the social network becomes large. We show that when private be-
liefs are unbounded (meaning that the implied likelihood ratios are unbounded), there
will be asymptotic learning as long as there is some minimal amount of “expansion in
observations”. We also characterize conditions under which there will be asymptotic
learning when private beliefs are bounded.
Keywords: information aggregation, learning, social networks, herding, information
cascades.
JEL Classification: C72, D83.

∗
We thank three anonymous referees, and Andrea Prat, Lones Smith and Peter Sorensen for useful
comments and suggestions. We gratefully acknowledge financial support from the AFOSR and the NSF.
†
Department of Economics, MIT, Cambridge MA, 02139 (e-mail: [email protected])
‡
Laboratory for Information and Decision Systems, Electrical Engineering and Computer Science
Department, MIT, Cambridge MA, 02139 (e-mail: [email protected])
§
Microsoft Research New England Lab, 1 Memorial Drive, Cambridge MA, 02142 (e-mail: ilo-
[email protected])
¶
Laboratory for Information and Decision Systems, Electrical Engineering and Computer Science
Department, MIT, Cambridge MA, 02139 (e-mail: [email protected])
1 Introduction
How is dispersed and decentralized information held by a large number of individuals ag-
gregated? Imagine a situation in which each of a large number of individuals has a noisy
signal about an underlying state of the world. This state of the world might concern,
among other things, earning opportunities in a certain occupation, the quality of a new
product, the suitability of a particular political candidate for office or payoff-relevant ac-
tions taken by the government. If signals are unbiased, the combination—aggregation—
of the information of the individuals will be sufficient for the society to “learn” the true
underlying state. The above question can be formulated as the investigation of what
types of behaviors and communication structures will lead to this type of information
aggregation.
Condorcet’s Jury Theorem provides a natural benchmark, where sincere (truthful)
reporting of their information by each individual is sufficient for aggregation of informa-
tion by a law of large numbers argument (Condorcet, 1788). Against this background,
a number of papers, most notably Bikchandani, Hirshleifer and Welch (1992), Banerjee
(1992) and Smith and Sorensen (2000), show how this type of aggregation might fail in
the context of the (perfect) Bayesian equilibrium of a dynamic game: when individuals
act sequentially and observe the actions of all previous individuals (agents), there may
be “herding” on the wrong action, preventing efficient aggregation of information.
An important modeling assumption in these papers is that each individual observes
all past actions. In practice, individuals are situated in complex social networks, which
provide their main source of information.1 In this paper, we address how the structure
of social networks, which determines the information that individuals receive, affects
equilibrium information aggregation.
As a motivating example, consider a group of consumers deciding which one of two
possible new smartphones to switch to. Each consumer makes this decision when her
existing service contract expires, so that there is an exogenous sequence of actions de-
termined by contract expiration dates. Each consumer observes the choices of some of
her friends, neighbors and coworkers, and the approximate timing of these choices (as
she sees when they start using the new phone). She does not, however, observe who else
these friends, neighbors and coworkers have themselves observed. Additional informa-
tion from direct communication is limited, since consumers cannot easily identify and
communicate their valuations soon after a new purchase. A related example would be
the choice of a firm between two new technologies after the (exogenous) breakdown of
its current technology. The firm observes the nature and timing of the choices of some of
the nearby firms, but not what other information these firms had at the time they made
their decisions. In this case also, the main source of information would be observation
of past actions rather than direct communication.
Similar to these examples, in our model a large number of agents sequentially choose
1
Granovetter (1973), Montgomery (1991), Munshi (2003) and Iaonnides and Loury (2004) document
the importance of information obtained from the social network of an individual for employment out-
comes. Besley and Case (1994), Foster and Rosenzweig (1995), Munshi (2004), and Udry and Conley
(2001) show the importance of the information obtained from social networks for technology adoption.
Jackson (2006, 2007) provides excellent surveys.

1
between two actions. An underlying state determines the payoffs of these two actions.
Each agent receives a signal on which of these two actions yields a higher payoff. Prefer-
ences of all agents are aligned in the sense that, given the underlying state of the world,
they all prefer the same action. In addition to his own signal, each agent observes the
choices of others in a stochastically generated neighborhood. In line with our social net-
work interpretation and the examples above, each individual knows the identity of the
agents in his neighborhood. But he does not observe their private signal or necessarily
know what information these agents had access to when making their own decisions.
More formally, this dynamic game of incomplete information is characterized by two
features: (i) the signal structure, which determines how informative the signals received
by the individuals are; (ii) the (social) network topology. We represent the network
topology by a sequence of probability distributions (one for each agent) over subsets of
past actions. The environment most commonly studied in the previous literature, the
full observation network topology, is the special case where all past actions are observed.
Another deterministic special case is the network topology where each agent observes
the actions of the most recent M ≥ 1 individuals. Other relevant social networks include
stochastic topologies in which each agent observes a random subset of past actions, as
well as those in which, with a high probability, each agent observes the actions of some
“influential” group of agents, who may be thought of as “informational leaders” or the
media. In addition to these examples, our representation of social networks is sufficiently
general to nest several commonly studied models of stochastic networks, including mod-
els of preferential attachment, small-world models, and observation structures in which
each individual sees the actions of one or several agents randomly drawn from the entire
past or the recent past. Our representation also does not impose any restriction on the
degree distribution (cardinality) of the agents’ neighborhoods or the degree of clustering
in the network.
We provide a systematic characterization of the conditions under which there will
be asymptotic learning in this model. We say that there is asymptotic learning if as
the size of the social network becomes arbitrarily large, equilibrium actions converge (in
probability) to the action that yields the higher payoff. Conversely, asymptotic learning
fails if, as the social network becomes large, the correct action is not chosen (or more
formally, the lim inf of the probability that the right action is chosen is strictly less than
1).
Two concepts turn out to be crucial in the study of information aggregation in social
networks. The first is whether the likelihood ratio implied by individual signals is always
finite and bounded away from 0.2 Smith and Sorensen (2000) refer to beliefs that satisfy
this property as bounded (private) beliefs. With bounded beliefs, there is a maximum
amount of information in any individual signal. In contrast, when there exist signals
with arbitrarily high and low likelihood ratios, (private) beliefs are unbounded. Whether
bounded or unbounded beliefs provide a better approximation to reality is partly an
interpretational and partly an empirical question. Smith and Sorensen’s main result is
that when each individual observes all past actions and private beliefs are unbounded,
2
The likelihood ratio is the ratio of the probabilities or the densities of a signal in one state relative
to the other.

2
information will be aggregated and the correct action will be chosen asymptotically. In
contrast, the results in Bikchandani, Hirshleifer and Welch (1992), Banerjee (1992) and
Smith and Sorensen (2000) indicate that with bounded beliefs, there will not be asymp-
totic learning (or information aggregation). Instead, as emphasized by Bikchandani,
Hirshleifer and Welch (1992) and Banerjee (1992), there will be “herding” or “infor-
mational cascades,” where individuals copy past actions and/or completely ignore their
own signals.
The second key concept is that of a network topology with expanding observations.
To describe this concept, let us first introduce another notion: a finite group of agents is
excessively influential if there exists an infinite number of agents who, with probability
uniformly bounded away from 0, observe only the actions of a subset of this group. For
example, a group is excessively influential if it is the source of all information (except
individual signals) for an infinitely large component of the social network. If there exists
an excessively influential group of individuals, then the social network has nonexpanding
observations, and conversely, if there exists no excessively influential group, the network
has expanding observations. This definition implies that most reasonable social networks
have expanding observations, and in particular, a minimum amount of “arrival of new
information ” in the social network is sufficient for the expanding observations property.3
For example, the environment studied in most of the previous work in this area, where
all past actions are observed, has expanding observations. Similarly, a social network
in which each individual observes one uniformly drawn individual from those who have
taken decisions in the past or a network in which each individual observes his immedi-
ate neighbor all feature expanding observations. Note also that a social network with
expanding observations need not be connected. For example, the network in which even-
numbered [odd-numbered] individuals only observe the past actions of even-numbered
[odd-numbered] individuals has expanding observations, but is not connected. A simple,
but typical, example of a network with nonexpanding observations is the one in which
all future individuals only observe the actions of the first K < ∞ agents.
Our main results in this paper are presented in four theorems. In particular, The-
orems 2 and 4 are the most substantive contributions of this paper. Theorem 1 shows
that there is no asymptotic learning in networks with nonexpanding observations. This
result is not surprising, since information aggregation is not possible when the set of ob-
servations on which (an infinite subset of) individuals can build their decisions remains
limited forever.
Theorem 2, shows that when (private) beliefs are unbounded and the network topol-
ogy is expanding, there will be asymptotic learning. This is a strong result (particularly
if we consider unbounded beliefs to be a better approximation to reality than bounded
beliefs), since almost all reasonable social networks have the expanding observations
property. This theorem, for example, implies that when some individuals, such as “in-
formational leaders,” are overrepresented in the neighborhoods of future agents (and are
thus “influential,” though not excessively so), learning may slow down, but asymptotic
3
Here, “arrival of new information” refers to the property that the probability of each individual
observing the action of some individual from the recent past converges to one as the social network
becomes arbitrarily large.

3
learning will still obtain as long as private beliefs are unbounded.4
Theorem 3 presents a partial converse to Theorem 2. It shows that for many common
deterministic and stochastic networks, bounded private beliefs are incompatible with
asymptotic learning. It therefore generalizes existing results on asymptotic learning,
for example, those in Bikchandani, Hirshleifer and Welch (1992), Banerjee (1992), and
Smith and Sorensen (2000) to general networks.
Our final main result, Theorem 4, establishes that there is asymptotic learning with
bounded private beliefs for a large class of stochastic network topologies. Suppose there
exists a subset S of the agents such that agents in S have a probability ϵ > 0 of
observing the entire history of actions and that an infinite subset of the agents in S
makes decisions (partially) based on their private signals because they have neighbors
whose actions are not completely informative. The remaining agents in the society have
expanding observations with respect to S (in the sense that they are likely to observe
some recent actions from S). Then, Theorem 4 shows that asymptotic learning occurs
for any distribution of private signals.5 This result is particularly important, since it
shows how moving away from simple network structures, which have been the focus of
prior work, has major implications for equilibrium learning dynamics.
Our paper is related to the large and growing literature on social learning. Bikchan-
dani, Hirshleifer and Welch (1992) and Banerjee (1992) started the literature on learning
in situations in which individuals are Bayesian and observe past actions. Smith and
Sorensen (2000) provide the most comprehensive and complete analysis of this environ-
ment. Their results and the importance of the concepts of bounded and unbounded
beliefs, which they introduced, have already been discussed in the introduction and will
play an important role in our analysis in the rest of the paper. Other important contri-
butions in this area include, among others, Welch (1992), Lee (1993), Chamley and Gale
(1994), and Vives (1997). An excellent general discussion is contained in Bikchandani,
Hirshleifer and Welch (1998). These papers typically focus on the special case of full
observation network topology in terms of our general model.
The two papers most closely related to ours are Banerjee and Fudenberg (2004) and
Smith and Sorensen (1998). Both of these papers study social learning with sampling of
past actions. In Banerjee and Fudenberg, there is a continuum of agents and the focus
is on proportional sampling (whereby individuals observe a “representative” sample of
the overall population). They establish that asymptotic learning is achieved under mild
4
The idea of the proof of Theorem 2 is as follows. We first establish a strong improvement prin-
ciple under unbounded beliefs, whereby in a network where each individual has a single agent in his
neighborhood, he can receive a strictly higher payoff than this agent and this improvement remains
bounded away from zero as long as asymptotic learning has not been achieved. We then show that
the same insight applies when individuals stochastically observe one or multiple agents (in particular,
with multiple agents, the improvement is no less than the case in which the individual observes a single
agent from the past). Finally, the property that the network topology has expanding observations is
sufficient for these improvements to accumulate to asymptotic learning.
5
The idea of the proof is as follows. We first show using a martingale convergence argument that the
beliefs of the agents in S must converge, and the limiting belief must be correct since infinitely many
agents in S are acting based, at least in part, on their private signals. We then use an improvement
argument and expanding observations with respect to S to establish asymptotic learning for the entire
society.

4
assumptions as long as the sample size is no smaller than two. The existence of a
continuum of agents is important for this result since it ensures that the fraction of
individuals with different posteriors evolves deterministically. Smith and Sorensen, on
the other hand, consider a related model with a countable number of agents. In their
model, as in ours, the evolution of beliefs is stochastic. Smith and Sorensen provide
conditions under which asymptotic learning takes place.
A crucial difference between Banerjee and Fudenberg and Smith and Sorensen, on the
one hand, and our work, on the other, is the information structure. These papers assume
that “samples are unordered” in the sense that individuals do not know the identity of
the agents they have observed. In contrast, as mentioned above, our setup is motivated
by a social network and assumes that individuals have stochastic neighborhoods, but
know the identity of the agents in their realized neighborhood. We view this as a better
approximation to learning in social networks. While situations in which an individual
observes the actions of “strangers” would naturally correspond to “unordered samples,”
in most social networks situations individuals have some idea about where others are
situated in the network. For example, an individual would have some idea about which
of their friends and coworkers are likely to have observed the choices of many others,
which in the context of our model corresponds to “ordered samples”. In addition to its
descriptive realism, this assumption leads to a sharper characterization of the conditions
under which asymptotic learning occurs. For example, in Smith and Sorensen’s environ-
ment, asymptotic learning fails whenever an individual is “oversampled,” in the sense of
being overrepresented in the samples of future agents. In contrast, in our environment,
asymptotic learning occurs when the network topology features expanding observations
(and private beliefs are unbounded). Expanding observations is a much weaker require-
ment than “non-oversampling.” For example, when each individual observes agent 1 and
a randomly chosen agent from his predecessors, the network topology satisfies expanding
observations, but there is oversampling.6
Other recent work on social learning includes Celen and Kariv (2004) who study
Bayesian learning when each individual observes his immediate predecessor, Callander
and Horner (2006), who show that it may be optimal to follow the actions of agents
that deviate from past average behavior and Gale and Kariv (2003) who generalize
the payoff equalization result of Bala and Goyal (1998) in connected social networks
(discussed below) to Bayesian learning.7
The second branch of the literature focuses on non-Bayesian learning, typically with
agents using some reasonable rules of thumb. This literature considers both learning from
past actions and from payoffs (or directly from beliefs). Early papers in this literature
include Ellison and Fudenberg (1993, 1995), which show how rule-of-thumb learning can
converge to the true underlying state in some simple environments. The papers most
closely related to our work in this genre are Bala and Goyal (1998, 2001), DeMarzo,
6
This also implies that, in the terminology of Bala and Goyal, a “royal family” precludes learning
in Smith and Sorensen’s model, but not in ours, see below.
7
Gale and Kariv show that the behavior of agents in a connected social network who repeatedly act
according to their posterior beliefs eventually converge. This does not constitute asymptotic learning
according to our definition since behavior does not converge to the optimal action given the available
information.

5
Vayanos and Zwiebel (2003) and Golub and Jackson (2007). These papers study non-
Bayesian learning over an arbitrary, connected social network. Bala and Goyal (1998)
establish the important and intuitive payoff equalization result that, asymptotically,
each individual must receive a payoff equal to that of an arbitrary individual in his
“social network,” since otherwise he could copy the behavior of this other individual.
Our paper can be viewed as extending Bala and Goyal’s results to a situation with
Bayesian learning. A similar “imitation” intuition plays an important role in our proof
of asymptotic learning with unbounded beliefs and unbounded observations. A key
difference between our results and those in Bala and Goyal, DeMarzo, Vayanos and
Zwiebel, and Golub and Jackson concerns the effects of “influential” groups on learning.
In these non-Bayesian papers, an influential group or what Bala and Goyal refer to as a
royal family, which is highly connected to the rest of the network, prevents aggregation of
information. In contrast, in our model Bayesian updating ensures that such individuals
do not receive disproportionate weight and their presence does not preclude efficient
aggregation of information. Only when there is an excessively influential group, that is,
a group that is the sole source of information for an infinite subset of individuals, that
asymptotic learning breaks down.8
The rest of the paper is organized as follows. Section 2 introduces our model. Sec-
tion 3 characterizes the (pure-strategy) perfect Bayesian equilibria and introduces the
concepts of bounded and unbounded beliefs. Section 4 presents our main results, The-
orems 1-4 and discusses some of their implications (as well as presenting a number of
corollaries to facilitate interpretation). Section 5 concludes. Appendix A contains the
main proofs, while the not-for-publication Appendix B contains additional results and
omitted proofs.

2 Model
A countably infinite number of agents (individuals), indexed by n ∈ N, sequentially
make a single decision each. The payoff of agent n depends on an underlying state of
the world θ and his decision. To simplify the notation and the exposition, we assume
that both the underlying state and decisions are binary. In particular, the decision of
agent n is denoted by xn ∈ {0, 1} and the underlying state is θ ∈ {0, 1}. The payoff of
agent n is {
1 if xn = θ
un (xn , θ) =
0 if xn ̸= θ.
Again to simplify notation, we assume that both values of the underlying state are
equally likely, so that P(θ = 0) = P(θ = 1) = 1/2.
The state θ is unknown. Each agent n ∈ N forms beliefs about this state from a
private signal sn ∈ S (where S is a metric space or simply a Euclidean space) and from
8
There is also a literature in engineering, which studies related problems, especially motivated by
aggregation of information collected by decentralized sensors. These include Cover (1969), Papastavrou
and Athans (1990), Lorenz, Marciniszyn and, Steger (2007), and Tay, Tsitsiklis and Win (2007). The
work by Papastavrou and Athans contains a result that is equivalent to the characterization of asymp-
totic learning with the observation of the immediate neighbor.

6
his observation of the actions of other agents. Conditional on the state of the world θ,
the signals are independently generated according to a probability measure Fθ . We refer
to the pair of measures (F0 , F1 ) as the signal structure of the model. We assume that F0
and F1 are absolutely continuous with respect to each other, which immediately implies
that no signal is fully revealing about the underlying state. We also assume that F0 and
F1 are not identical, so that some signals are informative. These two assumptions on
the signal structure are maintained throughout the paper and will not be stated in the
theorems explicitly.
In contrast to much of the literature on social learning, we assume that agents do
not necessarily observe all previous actions. Instead, they observe the actions of other
agents according to the structure of the social network. To introduce the notion of a social
network, let us first define a neighborhood. Each agent n observes the decisions of the
agents in his (stochastically-generated) neighborhood, denoted by B(n).9 Since agents
can only observe actions taken previously, B(n) ⊆ {1, 2, ..., n − 1}. Each neighborhood
B(n) is generated according to an arbitrary probability distribution Qn over the set of
all subsets of {1, 2, ..., n − 1}. We impose no special assumptions on the sequence of
distributions {Qn }n∈N except that the draws from each Qn are independent from each
other for all n and from the realizations of private signals. The sequence {Qn }n∈N is the
network topology of the social network formed by the agents. The network topology is
common knowledge, whereas the realized neighborhood B(n) and the private signal sn
are the private information of agent n. We say that {Qn }n∈N is a deterministic network
topology if the probability distribution Qn is a degenerate (Dirac) distribution for all n.
Otherwise, that is, if {Qn } for some n is nondegenerate, {Qn }n∈N is a stochastic network
topology.
A social network consists of a network topology {Qn }n∈N and a signal structure
(F0 , F1 ).
Example 1 Here are some examples of network topologies.
1. If {Qn }n∈N assigns probability 1 to neighborhood {1, 2..., n − 1} for each n ∈ N ,
then the network topology is identical to the canonical one studied in the previous
literature where each agent observes all previous actions (e.g., Banerjee (1992),
Bikchandani, Hirshleifer and Welch (1992), Smith and Sorensen (2000)).
2. If {Qn }n∈N assigns probability 1/(n − 1) to each one of the subsets of size 1 of
{1, 2..., n − 1} for each n ∈ N , then we have a network topology of random sam-
pling of one agent from the past.
3. If {Qn }n∈N assigns probability 1 to neighborhood {n − 1} for each n ∈ N , then
we have a network topology where each individual only observes his immediate
neighbor.
4. If {Qn }n∈N assigns probability 1 to neighborhoods that are subsets of {1, 2, ..., K}
for each n ∈ N for some K ∈ N . In this case, all agents observe the actions of at
most K agents.
9
If n′ ∈ B(n), then agent n not only observes the action of n′ , but also knows the identity of this
agent. Crucially, however, n does not observe B (n′ ) or the actions of the agents in B (n′ ).

7
 2 5

7
1 6
3

STATE

Figure 1: The figure illustrates the world from the perspective of agent 7. Agent 7
knows her private signal s7 , her realized neighborhood, B(7) = {4, 6} and the decisions
of agents 4 and 6, x4 and x6 . She also knows the probabilistic model {Qn }n<7 for
neighborhoods of all agents n < 7.

5. Figure 1 depicts an arbitrary stochastic topology until agent 7. The thickness of

the lines represents the probability with which a particular agent will observe the
action of the corresponding preceding agent.

Notice that our framework is general enough to nest the majority of social net-
work models studied in the literature, including the popular preferential attachment
and small-world network structures. For example, the preferential attachment model
can be nested by choosing a stochastic network topology {Qn }n∈N with a collection of
subsets S1 ,...,Sk of agents such that agents in S1 have a very high probability of being in
each B (n), those in S2 also have a high, but lower than the corresponding probability
for those in S2 , of being in each B (n), and so on. The small-world network structure
can be nested by choosing a partition {Sj } of N such that for each n ∈ Sj , the prob-
ability that for m < n that is also in Sj is in B (n) is very high, while the probability
that m is not in Sj is low but positive. More generally, any network structure can be
represented by a judicious choice of {Qn }n∈N provided that we keep the assumption that
the realizations of {Qn }n∈N are independent, which is adopted to simplify the analysis.10
The independence assumption on the neighborhoods does not impose a restriction on
the degree distribution (cardinality) of the agents or on the degree of clustering of the
10
This independence assumption rules out situations in which the fact that a particular individual is
observed more makes it even more likely that it will be observed more in the future. Although this is
related to the preferential attachment models and in particular to Jackson and Rogers’s (2007) gener-
ative model that leads to preferential attachment, as noted in this paragraph, preferential attachment
models can also be cast as special cases of our model without abandoning the independence assumption.

8
agents. To observe this, note that any given deterministic network topology satisfies the
independence assumption and it can be selected to have an arbitrary degree distribution
or level of clustering.

3 Equilibrium Strategies
In this section, we introduce the definitions of equilibrium and asymptotic learning,
and we provide a characterization of equilibrium strategies. In particular, we show
that equilibrium decision rules of individuals can be decomposed into two parts, one
that only depends on an individual’s private signal, and the other that is a function of
the observations of past actions. We also show why a full characterization of individual
decisions is nontrivial and motivate an alternative proof technique, relying on developing
bounds on improvements in the probability of the correct decisions, that will be used in
the rest of our analysis.

3.1 Perfect Bayesian Equilibrium and the Asymptotic Learn-

ing
Given the description above, it is evident that the information set In of agent n is given
by her signal sn , her neighborhood B(n), and all decisions of agents in B(n), that is,

In = {sn , B(n), xk for all k ∈ B(n)}. (1)

The set of all possible information sets of agent n is denoted by In . A strategy for
individual n is a mapping σn : In → {0, 1} that selects a decision for each possible
information set. A strategy profile is a sequence of strategies σ = {σn }n∈N . We use the
standard notation σ−n = {σ1 , . . . , σn−1 , σn+1 , . . .} to denote the strategies of all agents
other than n and also (σn , σ−n ) for any n to denote the strategy profile σ. Given a
strategy profile σ, the sequence of decisions {xn }n∈N is a stochastic process and we
denote the measure generated by this stochastic process by Pσ .

Definition 1 A strategy profile σ ∗ is a pure-strategy Perfect Bayesian Equilibrium

of this game of social learning if for each n ∈ N, σn∗ maximizes the expected payoff of
∗
agent n given the strategies of other agents σ−n .

In the rest of the paper, we focus on pure-strategy Perfect Bayesian Equilibria, and
simply refer to this as “equilibrium” (without the pure-strategy and the Perfect Bayesian
qualifiers).
Given a strategy profile σ, the expected payoff of agent n from action xn = σn (In ) is
simply Pσ (xn = θ | In ). Therefore, for any equilibrium σ ∗ , we have

σn∗ (In ) ∈ argmax P(y,σ−n

∗ ) (y = θ | In ). (2)
y∈{0,1}

We denote the set of equilibria (pure-strategy Perfect Bayesian Equilibria) of the game
∗
by Σ∗ . It is clear that Σ∗ is nonempty. Given the sequence of strategies {σ1∗ , . . . , σn−1 },

9
the maximization problem in (2) has a solution for each agent n and each In ∈ In .
Proceeding inductively, and choosing either one of the actions in case of indifference
determines an equilibrium. We note the existence of equilibrium here.

Proposition 1 There exists a pure-strategy Perfect Bayesian Equilibrium.

Our main focus is whether equilibrium behavior will lead to information aggregation.
This is captured by the notion of asymptotic learning, which is introduced next.

Definition 2 Given a signal structure (F0 , F1 ) and a network topology {Qn }n∈N , we say
that asymptotic learning occurs in equilibrium σ if xn converges to θ in probability
(according to measure Pσ ), that is,

lim Pσ (xn = θ) = 1.
n→∞

Notice that asymptotic learning requires that the probability of taking the correct
action converges to 1.11 Therefore, asymptotic learning will fail when, as the network
becomes large, the limit inferior of the probability of all individuals taking the correct
action is strictly less than 1.
Our goal in this paper is to characterize conditions on social networks—on signal
structures and network topologies—that ensure asymptotic learning.

3.2 Characterization of Individual Decisions

Our first result shows that individual decisions can be characterized as a function of the
sum of two posteriors. These posteriors play an important role in our analysis. We will
refer to these posteriors as the individual’s private belief and the social belief.

Proposition 2 Let σ ∈ Σ∗ be an equilibrium of the game. Let In ∈ In be an information

set of agent n. Then, the decision of agent n, xn = σ(In ), satisfies
{ ( )
1, if Pσ (θ = 1 | sn ) + Pσ (θ = 1 | B(n), xk , k ∈ B(n)) > 1,
xn =
0, if Pσ (θ = 1 | sn ) + Pσ θ = 1 | B(n), xk , k ∈ B(n) < 1,

and xn ∈ {0, 1} otherwise.

Proof. See Appendix A.

This proposition establishes an additive decomposition in the equilibrium decision
rule between the information obtained from the private signal of the individual and from
the observations of others’ actions (in his neighborhood). The next definition formally
distinguishes between the two components of an individual’s information.
11
It is also clear that asymptotic learning is equivalent to the posterior beliefs converging to a distri-
bution putting probability 1 on the true state.

10
Definition 3 We refer to the probability Pσ (θ = 1 | sn ) as the private belief of agent
n, and the probability
( )
Pσ θ = 1 B(n), xk for all k ∈ B(n) ,
as the social belief of agent n.
Proposition 2 and Definition 3 imply that the equilibrium decision rule for agent
n ∈ N is equivalent to choosing xn = 1 when the sum of his private and social beliefs
is greater than 1. Consequently, the properties of private and social beliefs will shape
equilibrium learning behavior.
Notice that the social belief depends on n since it is a function of the (realized) neigh-
borhood of agent n. In most learning models, social beliefs have natural monotonicity
properties. For example, a greater fraction of individuals choosing action x = 1 in the
information set of agent n would increase qn . It is straightforward to construct exam-
ples, where such monotonicity properties do not hold under general social networks (see
Appendix B). For this reason, we will use a different line of attack, based on developing
lower bounds on the probability of taking the correct action, for establishing our main
results, which are presented in the next section.

3.3 Bounded and Unbounded Private Beliefs

The private belief of an individual is a function of his private signal s ∈ S and is not
a function of the strategy profile σ since it does not depend on the decisions of other
agents. We represent probabilities that do not depend on the strategy profile by P. We
use the notation pn to represent the private belief of agent n, i.e.,
pn = P(θ = 1 | sn ).
A straightforward application of Bayes’ Rule implies that for any n and any signal
sn ∈ S, the private belief pn of agent n is given by
( )−1
dF0
pn = 1 + (sn ) . (3)
dF1
We next define the support of a private belief. In our subsequent analysis, we will
see that properties of the support of private beliefs play a key role in asymptotic learn-
ing behavior. Since the pn are identically distributed for all n (which follows by the
assumption that the private signals sn are identically distributed), in the following, we
will use agent 1’s private belief p1 to define the support and the conditional distributions
of private beliefs.
Definition 4 The support of the private beliefs is the interval [β, β], where the end
points of the interval are given by
β = inf {r ∈ [0, 1] | P(p1 ≤ r) > 0}, and β = sup {r ∈ [0, 1] | P(p1 ≤ r) < 1}.

The signal structure has bounded private beliefs if β > 0 and β < 1 and unbounded
private beliefs if β = 1 − β = 1.

11
 When private beliefs are bounded, there is a maximum informativeness to any signal.
When they are unbounded, agents may receive arbitrarily strong signals favoring either
state (this follows from the assumption that (F0 , F1 ) are absolutely continuous with
respect to each other).
The conditional distribution of private beliefs given the underlying state j ∈ {0, 1}
can be directly computed as

Gj (r) = P(p1 ≤ r | θ = j). (4)

The signal structure (F0 , F1 ) can be equivalently represented by the corresponding pri-
vate belief distributions (G0 , G1 ), and in what follows, it will typically be more conve-
nient to work with (G0 , G1 ) rather than (F0 , F1 ). It is straightforward to verify that
G0 (r)/G1 (r) is nonincreasing in r and G0 (r)/G1 (r) > 1 for all r ∈ (β, β) (see Lemma 1
in Appendix A).

4 Main Results
In this section, we present our main results on asymptotic learning and provide the main
intuition for the proofs.

4.1 Expanding Observations

We start by introducing the key properties of network topologies and signal structures
that impact asymptotic learning. Intuitively, for asymptotic learning to occur, the infor-
mation that each agent receives from other agents should not be confined to a bounded
subset of agents. This property is established in the following definition. For this defi-
nition and throughout the paper, if the set B(n) is empty, we set maxb∈B(n) b = 0.

Definition 5 The network topology has expanding observations if for all K ∈ N,

we have ( )
lim Qn max b < K = 0.
n→∞ b∈B(n)

If the network topology does not satisfy this property, then we say it has nonexpanding
observations.

Recall that the neighborhood of agent n is a random variable B(n) (with values in the
set of subsets of {1, 2, ..., n−1}) and distributed according to Qn . Therefore, maxb∈B(n) b
is a random variable that takes values in {0, 1, ..., n − 1}. The expanding observations
condition can be restated as the sequence of random variables {maxb∈B(n) b}n∈N converg-
ing to infinity in probability. Similarly, it follows from the preceding definition that the
network topology has nonexpanding observations if and only if there exists some K ∈ N
and some scalar ϵ > 0 such that
( )
lim sup Qn max b < K ≥ ϵ.
n→∞ b∈B(n)

12
An alternative restatement of this definition might clarify its meaning. Let us refer to
a finite set of individuals C as excessively influential if there exists a subsequence of
agents who, with probability uniformly bounded away from zero, observe the actions of
a subset of C. Then, the network topology has nonexpanding observations if and only
if there exists an excessively influential group of agents. Note also that if there is a
minimum amount of arrival of new information in the network, so that the probability
of an individual observing some other individual from the recent past goes to one as the
network becomes large, then the network topology will feature expanding observations.
This discussion therefore highlights that the requirement that a network topology has
expanding observations is quite mild and most social networks, including all of those
discussed above, satisfy this requirement.
When the topology has nonexpanding observations, there is a subsequence of agents
that draws information from the first K decisions with positive probability (uniformly
bounded away from 0). It is then intuitive that network topologies with nonexpanding
observations will preclude asymptotic learning. Our first theorem states and proofs this
result.

Theorem 1 Assume that the network topology {Qn }n∈N has nonexpanding observations.
Then, there exists no equilibrium σ ∈ Σ∗ with asymptotic learning.

Proof. See Appendix A.

This theorem states the intuitive result that with nonexpanding observations, asymp-
totic learning will fail. This result is not surprising, since asymptotic learning requires
the aggregation of the information of different individuals. But a network topology with
nonexpanding observations does not allow such aggregation. Intuitively, nonexpanding
observations, or equivalently the existence of an excessively influential group of agents,
imply that infinitely many individuals will observe finitely many actions with positive
probability and this will not enable them to aggregate the dispersed information collec-
tively held by the entire social network.

4.2 Asymptotic Learning with Unbounded Private Beliefs

A central question is then whether, once we exclude network topologies with nonex-
panding observations, what other conditions need to be imposed to ensure asymptotic
learning. The following theorem is one of the main results of the paper and shows that
for general network topologies, unbounded private beliefs play a key role. In particu-
lar, unbounded private beliefs and expanding observations are sufficient for asymptotic
learning in all equilibria.

Theorem 2 Assume that the signal structure (F0 , F1 ) has unbounded private beliefs and
the network topology {Qn }n∈N has expanding observations. Then, asymptotic learning
occurs in every equilibrium σ ∈ Σ∗ .

Proof. See Appendix A.

13
 Theorem 2 is quite a striking result. It implies that unbounded private beliefs are
sufficient for asymptotic learning for most (but not all) network topologies. In particu-
lar, the condition that the network topology has expanding observations is fairly mild
and only requires a minimum amount of arrival of recent information to the network.
Social networks in which each individual observes all past actions, those in which each
observes just his neighbor, and those in which each individual observes M ≥ 1 agents
independently and uniformly drawn from his predecessors are all examples of network
topologies with expanding observations. Theorem 2 therefore implies that unbounded
private beliefs are sufficient to guarantee asymptotic learning in social networks with
these properties and many others.
The proof of this theorem is presented in Appendix A. Here we provide a roadmap
and the general intuition. As noted in the previous section, there is no monotonicity
result linking the behavior of an agent to the fraction of actions he or she observes. Nev-
ertheless, we prove Theorem 2 by makinge use of an informational monotonicity, related
to the (expected) welfare improvement principle in Banerjee and Fudenberg (2004) and
in Smith and Sorensen (1998), and the imitation principle in Bala and Goyal (1998) and
Gale and Kariv (2003). We first consider a special case in which each individual only
observes one other from the past (i.e., B (n) is a singleton for each n). We then estab-
lish the following strong improvement principle: with unbounded private beliefs, there
exists a strict lower bound on the increase in the ex ante probability that an individual
will make a correct decision over his neighbor’s probability (recall that for now there
is a single agent in each individual’s neighborhood, thus each individual has a single
“neighbor”). Intuitively, each individual can copy the behavior of their neighbor unless
they have a very strong signal that points in a different direction. We show that this
results in a strict improvement in the probability of taking the right action whenever
this probability is not equal to 1.
We then prove a generalized strong improvement principle for arbitrary social net-
works, by showing that each individual can obtain such an improvement even if they
ignore all but one of the individuals in their information set. The overall improvement
in the probability that each individual will take the right action is a priori greater than
this lower bound.
Finally, the proof of Theorem 2 follows by using this generalized strong improvement
principle to construct a subsequence of informational improvements at each point, and
showing that there is a strict improvement at each step of the subsequence. This com-
bined with the expanding observations assumption on the network topology establishes
the asymptotic learning result in the proof of the theorem.
The following corollary to Theorems 1 and 2 shows that for an interesting class
of stochastic network topologies, there is a critical topology at which there is a phase
transition—that is, for all network topologies with greater expansion of observations
than this critical topology, there will be asymptotic learning and for all topologies with
less expansion, asymptotic learning will fail.

Corollary 1 Assume that the signal structure (F0 , F1 ) has unbounded private beliefs.

14
Assume also that the network topology is given by {Qn }n∈N such that
a
Qn (m ∈ B(n)) = for all n and all m < n,
(n − 1)c

where, given n, the draws for m, m′ < n are independent and a and c are positive
constants. If c < 1 then asymptotic learning occurs in all equilibria. If c ≥ 1, then
asymptotic learning does not occur in any equilibrium.

Proof. See Appendix A.

Given the class of network topologies in this corollary, c < 1 implies that as the
network becomes large, there will be sufficient expansion of observations. In contrast,
for c ≥ 1, stochastic process Qn does not place enough probability on observing recent
actions and the network topology is nonexpanding. Consequently, Theorem 1 applies
and there is no asymptotic learning.
To highlight the implications of Theorems 1 and 2 for deterministic network topolo-
gies, let us introduce the following definition.

Definition 6 Assume that the network topology is deterministic. Then, we say a finite
sequence of agents π is an information path of agent n if for each i, πi ∈ B(πi+1 ) and
the last element of π is n. Let π(n) be an information path of agent n that has maximal
length. Then, we let L(n) denote the number of elements in π(n) and call it agent n’s
information depth.

Intuitively, the concepts of information path and information depth capture the
intuitive notion of how long the “trail” of the information in the neighborhood of an
individual is. For example, if each individual observes only his immediate neighbor
(i.e., B(n) = {n − 1} with probability one), each will have a small neighborhood, but
the information depth of a high-indexed individual will be high (or the “trail” will be
long), because the immediate neighbor’s action will contain information about the signals
of all previous individuals. The next corollary shows that with deterministic network
topologies, asymptotic learning will occur if and only if the information depth (or the
trail of the information) increases without bound as the network becomes larger.

Corollary 2 Assume that the signal structure (F0 , F1 ) has unbounded private beliefs.
Assume that the network topology is deterministic. Then, asymptotic learning occurs
for all equilibria if the sequence of information depths {L(n)}n∈N goes to infinity. If the
sequence {L(n)}n∈N does not go to infinity, then asymptotic learning does not occur in
any equilibrium.

Proof. See Appendix A.

4.3 No Learning under Bounded Private Beliefs

In the full observation network topology, bounded beliefs imply lack of asymptotic learn-
ing. One might thus expect a converse to Theorem 2, whereby asymptotic learning fails

15
whenever signals are bounded. Under general network topologies, learning dynamics
turn out to be more interesting and richer. The next theorem provides a partial con-
verse to Theorem 2 and shows that for a wide range of deterministic and stochastic
network topologies, bounded beliefs imply no asymptotic learning. However, somewhat
surprisingly, Theorem 4 will show that the same is not true with more general stochastic
network topologies.

Theorem 3 Assume that the signal structure (F0 , F1 ) has bounded private beliefs. If
the network topology {Qn }n∈N satisfies one of the following conditions,

(a) B(n) = {1, . . . , n − 1} for all n,

(b) |B(n)| ≤ 1 for all n, or
(c) there exists some constant M such that |B(n)| ≤ M for all n and

lim max b = ∞ with probability 1,

n→∞ b∈B(n)

then, asymptotic learning does not occur in any equilibrium σ ∈ Σ∗ .

Proof. See Appendix B.

This theorem implies that in many common deterministic and stochastic network
topologies, bounded private beliefs imply lack of asymptotic learning. Part (a) of this
theorem is already proved by Smith and Sorensen (2000), we provide an alternative
proof in Appendix B that highlights the importance of the concepts emphasized here.
Intuitively, in all cases we can provide an upper bound on the amount of learning. We
establish this upper bound by showing that either individuals start following potentially
incorrect social beliefs after a certain stage of learning has been reached, or they rely on
their own private signal, taking the incorrect action with positive probability.
Part (c) of the theorem is the more substantive contribution. The idea behind the
proof of Part (c) is as follows: suppose, to obtain a contradiction, that asymptotic
learning occurs. We show that this implies that the social belief of agents who only
observe neighbors selecting action 1 converges to 1. The probability that the first K
agents select action 1 is positive (for any K). In case of such an event and K large,
eventually all agents will have a social belief close to 1. But since with bounded private
beliefs an agent with social belief close to 1 will ignore his signal, after a while decisions
will contain no new information, contradicting asymptotic learning.
The following corollary illustrates the implications of Theorem 3. It shows that, when
private beliefs are bounded, there will be no asymptotic learning (in any equilibrium) in
stochastic networks with random sampling.

Corollary 3 Assume that the signal structure (F0 , F1 ) has bounded private beliefs.
Assume that each agent n samples M agents uniformly and independently among
{1, ..., n − 1}, for some M ≥ 1. Then, asymptotic learning does not occur in any
equilibrium σ ∈ Σ∗ .

Proof. See Appendix B.

16
4.4 Asymptotic Learning with Bounded Private Beliefs
While in the full observation network topology studied by Smith and Sorensen (2000) and
in other instances investigated in the previous literature, bounded private beliefs preclude
asymptotic learning, this is no longer true under general stochastic social networks.
Characterizing the set of stochastic network topologies under which learning occurs for
all distributions of private beliefs is the next major contribution of our paper. To present
these results, we first introduce the notion of a nonpersuasive neighborhood.

Definition 7 A finite set B ⊂ N is a nonpersuasive neighborhood in equilibrium

σ ∈ Σ∗ if ( )
Pσ (θ = 1|xk = yk for all k ∈ B) ∈ β, β
for any set of values yk ∈ {0, 1} for each k. We denote the set of all nonpersuasive
neighborhoods by Uσ .

A neighborhood B is nonpersuasive in equilibrium σ ∈ Σ∗ if for any set of decisions

that agent n observes, his behavior may still depend on his private signal. A nonpersua-
sive neighborhood is defined with respect to a particular equilibrium. However, we will
provide below several different sufficient conditions for a collection of neighborhoods to
be nonpersuasive in any equilibrium.
Our main theorem for learning with bounded beliefs provides a broad class of stochas-
tic social networks where asymptotic learning takes place for any signal structure. It
relies on the society comprising two subsets, one learning from agents with nonpersua-
sive neighborhoods and the other learning from the former subset. This necessitates
strengthening the expanding observations condition to ensure that the latter subset re-
ceives new information from the former subset.

Definition 8 Given a subset S ⊆ N, the network topology {Qn }n∈N has expanding
observations with respect to S if for all K,
( )
lim Qn max b < K = 0.
n→∞ b∈B(n)∩S

Theorem 4 Let (F0 , F1 ) be an arbitrary signal structure and let S ⊆ N. Assume the
network topology {Qn }n∈N has expanding observations with respect to S and has a lower
bound on the probability of observing the entire history of actions along S, i.e., there
exists some ϵ > 0 such that

Qn (B(n) = {1, ..., n − 1}) ≥ ϵ for all n ∈ S.

Assume further that for some positive integer M and nonpersuasive sets C1 , ..., CM , i.e.,
Ci ∈ Uσ for all i = 1, . . . , M , we have

∑∑
M
Qn (B(n) = Ci ) = ∞.
n∈S i=1

Then, asymptotic learning occurs in equilibrium σ.

17
Proof. See Appendix A.
This is a rather surprising result, particularly in view of existing results in the liter-
ature, which generate herds and information cascades (and no learning) with bounded
beliefs. This theorem indicates that learning dynamics become significantly richer when
we consider general social networks.
Theorem 4 highlights a class of network structures that lead to learning. In these
structures, there is an infinite subset S that forms the core of the network. Within this
core S, there are two important subgroups of agents. There are infinitely many agents
who act based (partially) on their signals since they have nonpersuasive neighborhoods.
The existence of these agents does not preclude learning since the probability that an
agent n has a nonpersuasive neighborhood can go to 0 as n goes to infinity. By acting
based on their private signals, these agents play the essential function of bringing new
information into the society. The second key subgroup within the core of agents in S is
the set of all agents who observe the entire history of play. These agents play the role
of collecting the information brought in by the agents who have nonpersuasive neigh-
borhoods. From the martingale convergence theorem, the social belief of these agents
converges with probability 1, and because there is sufficient of information generated
by the agents with nonpersuasive neighborhoods, the social belief must converge to the
correct state of the world. The expanding observations with respect to S guarantees
that information spreads to all other agents. In particular, agents who do not observe
the entire history of actions can simply copy the action of their highest neighbors within
S. This strategy provides a lower bound on the payoff to these agents and in fact guar-
antees asymptotic learning. Therefore, these agents also asymptotically learn the state
in equilibrium.12
We conclude this subsection by providing sufficient conditions for a neighborhood to
be nonpersuasive in any equilibrium. It is straightforward to see that B = ∅, i.e., the
empty neighborhood, is nonpersuasive in any equilibrium. The following two proposi-
tions present two other classes of nonpersuasive neighborhoods.

Proposition 3 Let (G0 , G1 ) be private belief distributions. Assume that the first K
agents have empty neighborhoods, i.e., B(n) = ∅ for all n ≤ K, and K satisfies
 ( ) ( ) 
 log β β
log 1−β 
1−β
K < min ( ), ( ) . (5)
 log G0 (1/2) log 1−G0 (1/2) 
G1 (1/2) 1−G1 (1/2)

Then any subset B ⊆ {1, 2, ..., K} is a nonpersuasive neighborhood.

12
It is important to emphasize the difference between this result and that in Sgroi (2002), which
shows that a social planner can ensure some degree of information aggregation by forcing a subsequence
of agents to make decisions without observing past actions. With the same reasoning, one might
conjecture that asymptotic learning may occur if a particular subsequence of agents, such as that
indexed by prime numbers, has empty neighborhoods. However, there will not be asymptotic learning
in this deterministic topology since lim inf n→∞ Pσ (xn = θ) < 1. For the result that there is asymptotic
learning (i.e., lim inf n→∞ Pσ (xn = θ) = 1) in Theorem 4, the feature that the network topology is
stochastic is essential.

18
Proof. See Appendix A.
To obtain the intuition for (5), consider the equivalent pair of inequalities
( )−1 ( )−1
G0 (1/2)K (1 − G0 (1/2))K
1+ > 1 − β and 1+ < 1 − β, (6)
G1 (1/2)K (1 − G1 (1/2))K

The factor of (1 + (G0 (1/2)/G1 (1/2))K )−1 represents the probability that the state θ
is equal to 1 conditional on K agents independently selecting action 0. For such a
neighborhood of K agents acting independently to be nonpersuasive, there must exist
a signal strong enough in favor of state 1 such that an agent would select action 1 after
observing K independent agents choosing 0. From the equilibrium decision rule (cf.
Proposition 2), it follows that this holds if (1 + (G0 (1/2)/G1 (1/2))K )−1 + β > 1. We can
repeat the same argument for action 1 to obtain the second inequality in Eq. (6).
The condition in Eq. (5) has a natural interpretation: for a neighborhood B to be
nonpersuasive, any |B| decisions have to be less informative than a single very infor-
mative signal. In the case where an agent k ∈ B has an empty neighborhood, the
informativeness of her decision xk = 0 is given by G0 (1/2)/G1 (1/2) and can interpreted
as the informativeness of an “average” signal in favor of state 0 since any signal mildly
in favor of state 0 would lead to action 0. Therefore, the ratio between the informative
value of an average signal and the informative value of an extreme signal determines the
size |B| of the largest nonpersuasive neighborhood.

Proposition 4 Let (G0 , G1 ) be private belief distributions. Assume that the first K
agents observe the full history of past actions, i.e., B(n) = {1, ..., n − 1} for all n ≤ K,
and K satisfies  ( ) ( ) 
 log β log
β
1−β

1−β
K < min ( ), ( ) . (7)
 log G0 (1/2) log 1−G0 (1/2) 
G1 (1/2) 1−G1 (1/2)

Assume also that the private belief distributions (G0 , G1 ) satisfy the following mono-
tonicity conditions: the functions
( ) ( )
1 G0 (1 − q) 1 1 − G0 (1 − q)
−1 and −1 (8)
q G1 (1 − q) q 1 − G1 (1 − q)
are both nonincreasing in q. Then any subset B ⊆ {1, 2, ..., K} is a nonpersuasive
neighborhood.

Proof. See Appendix A.

The monotonicity condition of Eq. (8) guarantees the following: let 0 ≤ q ′ ≤ q ′′ ≤ 1;
then Pσ (θ = 1|qn = q ′ , xn = 0) ≤ Pσ (θ = 1|qn = q ′′ , xn = 0) and Pσ (θ = 1|qn = q ′ , xn =
1) ≤ Pσ (θ = 1|qn = q ′′ , xn = 1). That is, an agent n with a given social belief qn = q ′
selecting action 0 (or action 1) represents a stronger signal in favor of state θ = 0 than
an agent with social belief qn = q ′′ ≥ q ′ choosing action 0 (or, respectively, action 1).
When the monotonicity condition of Eq. (8) holds, we obtain a partial ordering on
the informativeness of sets of decisions. In particular, it determines that the most in-
formative sequence of actions (x1 , x2 , ..., xK ) of the first K agents in favor of state 0

19
is the sequence (0, 0, ..., 0). Without the monotonicity condition, this seemingly mild
claim is not generally true. We can bound the informativeness of the sequence of de-
cisions (x1 , ..., xK ) = (0, ..., 0) using the fact (cf. Lemma 1(c) from Appendix A) that
G0 (q1 )/G1 (q1 ) ≥ G0 (q2 )/G1 (q2 ) ≥ ... ≥ G0 (qK )/G1 (qK ), where qk is the social belief
of agent K. In this situation, all decisions are less informative than agent 1’s decision
yielding
( )K ( )K
Pσ (xk = 0, k ≤ K|θ = 0) G0 (q1 ) G0 (1/2)
≤ = .
Pσ (xk = 0, k ≤ K|θ = 1) G1 (q1 ) G1 (1/2)
Using the same argument as in Proposition 3, we obtain that the condition in Eq. (7) is
sufficient to guarantee the neighborhood B is unpersuasive.

5 Conclusion
In this paper, we studied the problem of Bayesian (equilibrium) learning over a general
social network. A large social learning literature, pioneered by Bikhchandani, Hirshleifer
and Welch (1992), Banerjee (1992) and Smith and Sorensen (2000), has studied equilibria
of sequential-move games, where each individual observes all past actions. The focus has
been on whether equilibria lead to aggregation of information (and thus to asymptotic
learning).
In many relevant situations, individuals obtain their information not by observing
all past actions, but from their “social network”. This raises the question of how the
structures of social networks in which individuals are situated affects learning behavior.
To address these questions, we formulated a sequential-move equilibrium learning model
over a general social network.
In our model, each individual receives a signal about the underlying state of the
world and observes the past actions of a stochastically-generated neighborhood of in-
dividuals. The stochastic process generating the neighborhoods defines the network
topology. The signal structure determines the conditional distributions of the signals
received by each individual as a function of the underlying state. The social network
consists of the network topology and the signal structure. Each individual then chooses
one of two possible actions depending on his posterior beliefs given his signal and the
realized neighborhood. We characterized pure-strategy (perfect Bayesian) equilibria for
arbitrary stochastic and deterministic social networks, and characterized the conditions
under which there is asymptotic learning. Asymptotic learning corresponds to individual
decisions converging (in probability) to the right action as the social network becomes
large.
Two concepts turn out to be crucial in determining whether there will be asymp-
totic learning. The first is common with the previous literature. Following Smith and
Sorensen (2000), we say that private beliefs are bounded if the likelihood ratio implied
by individual signals is bounded and there is a maximum amount of information that
can be generated from these signals. Conversely, private beliefs are unbounded if the
corresponding likelihood ratio is unbounded. The second important concept is that of
expanding or nonexpanding observations. A network topology has nonexpanding ob-
servations if there exists infinitely many agents observing the actions of only a finite

20
subset of (excessively influential) agents. Most network topologies feature expanding
observations.
Nonexpanding observations do not allow asymptotic learning, since there exists in-
finitely many agents who do not receive sufficiently many observations to be able to
aggregate information.
Our first main theorem, Theorem 2, shows that expanding observations and un-
bounded private signals are sufficient to ensure asymptotic learning. Since expanding
observations is a relatively mild restriction, to the extent that unbounded private beliefs
constitute a good approximation to the informativeness of individual signals, this result
implies that all equilibria feature asymptotic learning applies in a wide variety of set-
tings. Another implication is that asymptotic learning is possible even when there are
“influential agents” or “information leaders”, that is, individuals who are observed by
many, most or even all agents (while others may be observed not at all or much less fre-
quently). It is only when individuals are excessively influential—loosely speaking when
they act as the sole source of information for infinitely many agents—that asymptotic
learning ceases to apply.
We also provide a partial converse to this result, showing that under the most com-
mon deterministic or stochastic network topologies, bounded beliefs imply no asymp-
totic learning. However, our second main theorem, Theorem 4, characterizes the class
of network structures where asymptotic learning occurs with bounded beliefs. This re-
sult shows the importance of general (stochastic) network topologies in the study of
Bayesian learning, since asymptotic learning with bounded beliefs is impossible in the
deterministic network topologies studied in the previous literature.
Beyond the specific results presented in this paper, we believe that the framework
developed here opens the way for a more general analysis of the structure of social
networks on learning dynamics. Among the questions that can be studied in future
work using this framework are the following: (1) the effect of network structure on the
speed (rate of convergence) of Bayesian learning; (2) equilibrium learning when there
are heterogeneous preferences; (3) equilibrium learning when the underlying state is
changing dynamically; (4) the influence of a subset of a social network (for example,
the media or interested parties) in influencing the views of the rest as a function of the
network structure.

21
Appendix A: Main Proofs
Proof of Proposition 2
We prove that if
( )
Pσ (θ = 1 | sn ) + Pσ θ = 1 | B(n), xk for all k ∈ B(n) > 1, (A1)

then xn = 1. The proofs of the remaining statements follow the same line of argument.
We first show that Eq. (A1) holds if and only if

Pσ (θ = 1 | In ) > 1/2, (A2)

therefore implying that xn = 1 by the equilibrium condition [cf. Eq. (2)]. By Bayes’
Rule, Eq. (A2) is equivalent to

dPσ (In | θ = 1)Pσ (θ = 1) dPσ (In | θ = 1)

Pσ (θ = 1 | In ) = ∑1 = ∑1 > 1/2,
j=0 dPσ (In | θ = j)Pσ (θ = j) j=0 dPσ (In | θ = j)

where the second equality follows from the assumption that states 0 and 1 are equally
likely. Hence, Eq. (A2) holds if and only if

dPσ (In | θ = 1) > dPσ (In | θ = 0).

Conditional on state θ, the private signals and the observed decisions are independent,
i.e.,
dPσ (In | θ = j) = dPσ (sn | θ = j)Pσ (B(n), xk , k ∈ B(n)|θ = j).
Combining the preceding two relations, it follows that Eq. (A2) is equivalent to

Pσ (B(n), xk , k ∈ B(n) | θ = 1) dPσ (sn | θ = 0)

∑1 > ∑1 .
j=0 Pσ (B(n), xk , k ∈ B(n) | θ = j) j=0 dPσ (sn | θ = j)

Since both states are equally likely, this can be rewritten as

Pσ (B(n), xk , k ∈ B(n) | θ = 1)Pσ (θ = 1) dPσ (sn | θ = 0)Pσ (θ = 0)

∑1 > ∑1 .
j=0 Pσ (B(n), xk , k ∈ B(n) | θ = j)Pσ (θ = j) j=0 dPσ (sn | θ = j)Pσ (θ = j)

Applying Bayes’ Rule on both sides of the inequality above, we see that the preceding
relation is identical to

Pσ (θ = 1 | B(n), xk , k ∈ B(n)) > Pσ (θ = 0 | sn ) = 1 − Pσ (θ = 1 | sn ),

completing the proof. 

22
Proof of Theorem 1
Suppose that the network has nonexpanding observations. This implies that there exists
some K ∈ N, ϵ > 0, and a subsequence of agents N such that for all n ∈ N ,
( )
Qn max b < K ≥ ϵ. (A3)
b∈B(n)

For any such agent n ∈ N , we have

( ) ( )
Pσ (xn = θ) = Pσ xn = θ max b < K Qn max b < K
b∈B(n) b∈B(n)
( ) ( )
+ Pσ xn = θ max b ≥ K Qn max b ≥ K
b∈B(n) b∈B(n)
( ) ( ) ( )
≤ Pσ xn = θ max b < K Qn max b < K + Qn max b ≥ K
b∈B(n) b∈B(n) b∈B(n)
( )
≤ 1 − ϵ + ϵPσ xn = θ max b < K , (A4)
b∈B(n)

where the second inequality follows from Eq. (A3).

Given some equilibrium σ ∈ Σ∗ and agent n, we define zn as the decision that
maximizes the conditional probability of making a correct decision given the private
signals and neighborhoods of the first K − 1 agents and agent n, i.e.,

zn = argmax Py,σ−n (y = θ | si , B(i), for i = 1, . . . , K − 1, n). (A5)

y∈{0,1}

We denote a particular realization of private signal si by si and a realization of neigh-

borhood B(i) by B(i) for all i. Given the equilibrium σ and the realization s1 , ..., sK−1
and B(1), ..., B(K − 1), all decisions x1 , . . . , xK−1 are recursively defined [i.e., they are
non-stochastic; see the definition of the information set in Eq. (1)]. Therefore, for any
B(n) that satisfies maxb∈B(n) b < K, the decision xn is also defined. By the definition of
zn [cf. Eq. (A5)], this implies that

Pσ (xn = θ | si = si , B(i) = B(i), for i = 1, . . . , K − 1, n)

≤ Pσ (zn = θ | si = si , B(i) = B(i), for i = 1, . . . , K − 1, n).

By integrating over all possible s1 , ..., sK−1 , sn , B(1), ..., B(K − 1), this yields

Pσ (xn = θ | B(n) = B(n)) ≤ Pσ (zn = θ | B(n) = B(n)),

for any B(n) that satisfies maxb∈B(n) b < K. By integrating over all B(n) that satisfy
this condition, we obtain
( ) ( )
Pσ xn = θ max b < K ≤ Pσ zn = θ max b < K . (A6)
b∈B(n) b∈B(n)

23
 Moreover, since the sequence of neighborhoods {B(i)}i∈N is independent of θ and the
sequence of private signals {si }i∈N , it follows from Eq. (A5) that the decision zn is given
by
zn = argmax Pσ (y = θ | s1 , ..., sK−1 , sn ). (A7)
y∈{0,1}

Therefore, zn is also independent of the sequence of neighborhoods {B(i)}i∈N and we

have ( )
Pσ zn = θ max b < K = Pσ (zn = θ) .
b∈B(n)

Since the private signals have the same distribution, it follows from Eq. (A7) that for
any n, m ≥ K, the random variables zn and zm have identical probability distributions.
Hence, for any n ≥ K, Eq. (A7) implies that
Pσ (zn = θ) = Pσ (zK = θ) .
Combining the preceding two relations with Eq. (A6), we have for any n ≥ K,
( ) ( )
Pσ xn = θ max b < K ≤ Pσ zn = θ max b < K = Pσ (zn = θ) = Pσ (zK = θ) .
b∈B(n) b∈B(n)

Substituting this relation in Eq. (A4), we obtain for any n ∈ N , n ≥ K,

Pσ (xn = θ) ≤ 1 − ϵ + ϵPσ (zK = θ) .
Therefore,
lim inf Pσ (xn = θ) ≤ 1 − ϵ + ϵPσ (zK = θ). (A8)
n→∞
We finally show that in view of the assumption that F0 and F1 are absolutely contin-
uous with respect to each other [which implies Pσ (x1 = θ) < 1], we have Pσ (zK = θ) < 1
for any given K. If Pσ (x1 = θ) < 1 holds, then we have either Pσ (x1 = θ | θ = 1) < 1 or
Pσ (x1 = θ | θ = 0) < 1. Assume without loss of generality that we have
Pσ (x1 = θ | θ = 1) < 1. (A9)
Let S denote the set of all private signals such that if s1 ∈ S σ , then x1 = 0 in equilibrium
σ. Since the first agent’s decision is a function of s1 , then Eq. (A9) is equivalent to
Pσ (s1 ∈ S σ | θ = 1) > 0.
Since the private signals are conditionally independent given θ, this implies that
Pσ (si ∈ S σ for all i ≤ K | θ = 1) = Pσ (s1 ∈ S σ | θ = 1)K > 0. (A10)
We next show that if si ∈ S σ for all i ≤ K, then zK = 0. Using Bayes’ Rule, we have
[ ]−1
Pσ (si ∈ S σ for all i ≤ K | θ = 1)
Pσ (θ = 0 | si ∈ S σ for all i ≤ K) = 1 +
Pσ (si ∈ S σ for all i ≤ K | θ = 0)
[ ∏K ]−1
i=1 Pσ (si ∈ S σ | θ = 1)
= 1 + ∏K
i=1 Pσ (si ∈ S σ | θ = 0)
[ ( )K ]−1
Pσ (s1 ∈ S σ | θ = 1)
= 1+ , (A11)
Pσ (s1 ∈ S σ | θ = 0)

24
where the second equality follows from the conditional independence of the private sig-
nals and the third equality holds since private signals are identically distributed. Ap-
plying Bayes’ Rule on the second term in parentheses in Eq. (A11), this implies that
[ ( )K ]−1
1
Pσ (θ = 0 | si ∈ S σ for all i ≤ K) = 1+ −1 (. A12)
Pσ (θ = 0 | s1 ∈ S σ )

Since s1 ∈ S σ induces x1 = 0, we have Pσ (θ = 0 | s1 ∈ S σ ) ≥ 1/2. Because the function

on the right-handside of Eq. (A12) is nondecreasing in Pσ (θ = 0 | s1 ∈ S σ ) for any value
in [1/2,1], we obtain
1
Pσ (θ = 0 | si ∈ S σ for all i ≤ K) ≥ .
2
By the definition of zK , this implies that if si ∈ S σ for all i ≤ K, then zK = 0
(we can let zn be equal to 0 whenever both states are equally likely given the private
signals). Combined with the fact that the event {si ∈ S σ for all i ≤ K} has positive
conditional probability given θ = 1 under measure Pσ [cf. Eq. (A10)], this implies that
Pσ (zK = θ) < 1. Substituting this relation in Eq. (A8), we have

lim inf Pσ (xn = θ) < 1,

n→∞

thus showing that asymptotic learning does not occur. 

Proof of Theorem 2
The proof follows by combining several lemmas, which we next present.
As a first step, we characterize certain important properties of private belief distri-
butions.

Lemma 1 For any private belief distributions (G0 , G1 ), the following relations hold.

(a) For all r ∈ (0, 1), we have

dG0 1−r
(r) = .
dG1 r
(b) We have
( )
1−r r−z
G0 (r) ≥ G1 (r) + G1 (z) for all 0 < z < r < 1,
r 2
( )
r w−r
1 − G1 (r) ≥ (1 − G0 (r)) + (1 − G1 (w)) for all 0 < r < w < 1.
1−r 2

(c) The ratio G0 (r)/G1 (r) is nonincreasing in r and G0 (r)/G1 (r) > 1 for all r ∈ (β, β).

25
Proof. See Appendix B.
We next show that the ex ante probability of an agent making the correct decision
(and thus his expected payoff) is no less than the probability of any of the agents in his
realized neighborhood making the correct decision.

Lemma 2 (Information Monotonicity) Let σ ∈ Σ∗ be an equilibrium. For any agent n

and neighborhood B, we have

Pσ (xn = θ | B(n) = B) ≥ max Pσ (xb = θ).

b∈B

Proof. See Appendix B.

Let us next focus on a specific network topology where each agent observes the
decision of a single agent. This will enable us to establish a preliminary version of
the strong improvement principle, which provides a lower bound on the increase in the
ex ante probability that an individual will make a correct decision over his neighbor’s
probability. We will then generalize this result arbitrate social networks.
For each n and strategy profile σ, let us define Ynσ and Nnσ as the probabilities of
agent n making the correct decision conditional on state θ. More formally, these are
defined as
Ynσ = Pσ (xn = 1 | θ = 1), Nnσ = Pσ (xn = 0 | θ = 0).
The unconditional probability of a correct decision is then
1 σ
(Y + Nnσ ) = Pσ (xn = θ). (A13)
2 n
We also define the thresholds Lσn and Unσ in terms of these probabilities:

1 − Nnσ Nnσ
Lσn = , Unσ = . (A14)
1 − Nnσ + Ynσ Nnσ + 1 − Ynσ

The next lemma shows that the equilibrium decisions are fully characterized in terms of
these thresholds.

Lemma 3 Let B(n) = {b} for some agent n. Let σ ∈ Σ∗ be an equilibrium, and let Lσb
and Ubσ be given by Eq. (A14). Then, agent n’s decision xn in equilibrium σ satisfies

 0, if pn < Lσb
xn = xb , if pn ∈ (Lσb , Ubσ )

1, if pn > Ubσ .

The proof is omitted since it is an immediate application of Proposition 2 [use Bayes’

Rule to determine Pσ (θ = 1|xb = j) for each j ∈ {0, 1}].
Note that the sequence {(Un , Ln )} only depends on {(Yn , Nn )}, and is thus determin-
istic. This reflects the fact that each individual recognizes the amount of information
that will be contained in the action of the previous agent, which determines his own

26
Figure 2: The equilibrium decision rule when observing a single agent, illustrated on the
private belief space.

decision thresholds. Individual actions are still stochastic since they are determined by
whether the individual’s private belief is below Lb , above Ub , or in between (see Figure
3).
Using the structure of the equilibrium decision rule, the next lemma provides an
expression for the probability of agent n making the correct decision conditional on his
observing agent b < n, in terms of the private belief distributions and the thresholds Lσb
and Ubσ .

Lemma 4 Let B(n) = {b} for some agent n. Let σ ∈ Σ∗ be an equilibrium, and let Lσb
and Ubσ be given by Eq. (A14). Then,

Pσ (xn = θ | B(n) = {b})

1[ ( ) ( ) ]
= G0 (Lσb ) + G0 (Ubσ ) − G0 (Lσb ) Nbσ + (1 − G1 (Ubσ )) + G1 (Ubσ ) − G1 (Lσb ) Ybσ .
2
Proof. By definition, agent n receives the same expected utility from all his possible
equilibrium choices. We can thus compute the expected utility by supposing that the
agent will choose xn = 0 when indifferent. Then, the expected utility of agent n (the
probability of the correct decision) can be written as

Pσ (xn = θ | B(n) = {b})

= Pσ (pn ≤ Lσb | θ = 0)P(θ = 0) + Pσ (pn ∈ (Lσb , Ubσ ], xb = 0 | θ = 0)P(θ = 0)
+Pσ (pn > Ubσ | θ = 1)P(θ = 1) + Pσ (pn ∈ (Lσb , Ubσ ], xb = 1 | θ = 1)P(θ = 1).

The result then follows using the fact that pn and xb are conditionally independent given
θ and the notation for the private belief distributions [cf. Eq. (4)].

27
 Using the previous lemma, we next provide a lower bound on the amount of improve-
ment in the ex ante probability of making the correct decision between an agent and his
neighbor.

Lemma 5 Let B(n) = {b} for some agent n. Let σ ∈ Σ∗ be an equilibrium, and let Lσb
and Ubσ be given by Eq. (A14). Then,
( σ)
(1 − Nbσ )Lσb Lb
Pσ (xn = θ | B(n) = {b}) ≥ Pσ (xb = θ) + G1
8 2
[ ( )]
(1 − Yb )(1 − Ub )
σ σ
1 + Ubσ
+ 1 − G0 .
8 2

Proof. In Lemma 1(b), let r = Lσb , z = Lσb /2, so that we obtain

( σ)
(1 − Nbσ )Lσb Lb
(1 − Nb )G0 (Lb ) ≥ Yb G1 (Lb ) +
σ σ σ σ
G1 .
4 2

Next, again using Lemma 1(b) and letting r = Ubσ and w = (1 + Ubσ )/2, we have
[ ( )]
(1 − Ybσ )(1 − Ubσ ) 1 + Ubσ
(1 − Yb )[1 − G1 (Ub )] ≥ Nb [1 − G0 (b )] +
σ σ σ σ
1 − G0 .
4 2

Combining the preceding two relations with Lemma 4 and using the fact that Ybσ +Nbσ =
2 Pσ (xb = θ) [cf. Eq. (A13)], the desired result follows.
The next lemma establishes that the lower bound on the amount of improvement
in the ex ante probability is uniformly bounded away from zero for unbounded private
beliefs and when Pσ (xb = θ) < 1, i.e., when asymptotic learning is not achieved.

Lemma 6 Let B(n) = {b} for some n. Let σ ∈ Σ∗ be an equilibrium, and denote
α = Pσ (xb = θ). Then,
{ ( ) ( )}
(1 − α)2 1−α 1+α
Pσ (xn = θ | B(n) = {b}) ≥ α + min G1 , 1 − G0 .
8 2 2

Proof. We consider two cases separately.

Case 1: Nbσ ≤ α. From the definition of Lσb and the fact that Ybσ = 2α − Nbσ [cf. Eq.
(A13)], we have
σ 1 − Nbσ
Lb = .
1 − 2Nbσ + 2α
Since σ is an equilibrium, we have α ≥ 1/2, and thus the right hand-side of the preceding
inequality is a nonincreasing function of Nbσ . Since Nbσ ≤ α, this relation therefore
implies that Lσb ≥ 1 − α. Combining the relations 1 − Nbσ ≥ 1 − α and Lσb ≥ 1 − α, we
obtain ( σ) ( )
(1 − Nbσ )Lσb Lb (1 − α)2 1−α
G1 ≥ G1 . (A15)
8 2 8 2

28
Case 2: Nbσ ≥ α. Since Ybσ + Nbσ = 2α, this implies that Ybσ ≤ α. Using the definition
of Ubσ and a similar argument as the one above, we obtain
[ ( )] [ ( )]
(1 − Ybσ )(1 − Ubσ ) 1 + Ubσ (1 − α)2 1+α
1 − G0 ≥ 1 − G0 . (A16)
8 2 8 2

Combining Eqs. (A15) and (A16), we obtain

( σ) [ ( )]
(1 − Nbσ )Lσb Lb (1 − Ybσ )(1 − Ubσ ) 1 + Ubσ
G1 + 1 − G0
8 2 8 2
{ ( ) ( )}
(1 − α) 2
1−α 1+α
≥ min G1 , 1 − G0 ,
8 2 2
where we also used the fact that each term on the left hand-side of the preceding in-
equality is nonnegative. Substituting this into Lemma 5, the desired result follows.

The preceding lemma characterizes the improvements in the probability of making

the correct decision between an agent and his neighbor. To study the limiting behavior
of these improvements, we introduce the function Z : [1/2, 1] → [1/2, 1] defined by
{ ( ) ( )}
(1 − α)2 1−α 1+α
Z(α) = α + min G1 , 1 − G0 . (A17)
8 2 2
Lemma 6 establishes that for n, which has B(n) = {b}, we have
Pσ (xn = θ|B(n) = {b}) ≥ Z (Pσ (xb = θ)) , (A18)
i.e., the function Z acts as an improvement function for the evolution of the probability
of making the correct decision. The function Z(·) has several important properties,
which are formally stated in the next lemma.

Lemma 7 The function Z : [1/2, 1] → [1/2, 1] given in (A17) satisfy the following
properties:

(a) The function Z has no upwards jumps. That is, for any α ∈ [1/2, 1],
Z(α) = lim Z(r) ≥ lim Z(r).
r↑α r↓α

(b) For any α ∈ [1/2, 1], Z(α) ≥ α.

(c) If the private beliefs are unbounded, then for any α ∈ [1/2, 1), Z(α) > α.

Proof. Since G0 and G1 are cumulative distribution functions, they cannot have down-
wards jumps, i.e., for each j ∈ {0, 1}, limr↑α Gj (r) ≤ limr↓α Gj (r) for any α ∈ [1/2, 1],
establishing Part (a). Part (b) follows from the fact that cumulative distribution func-
tions take values in [0, 1]. For Part (c), suppose that for some α ∈ [1/2, 1), Z(α) = α.
This implies that { ( ) ( )}
1−α 1+α
min G1 , 1 − G0 = 0. (A19)
2 2

29
However, from the assumption on the private beliefs, we have that for all α ∈ (0, 1) and
any j ∈ {0, 1}, Gj (α) ∈ (0, 1), contradicting Eq. (A19).
The properties of the Z function will be used in the analysis of asymptotic learning
in general networks in the last step of the proof of Theorem 2. The analysis of asymp-
totic learning requires the relevant improvement function to be both continuous and
monotone. However, Z does not necessarily satisfy these properties. We next construct
a related function Z : [1/2, 1] → [1/2, 1] that satisfies these properties and can be used
as the improvement function in the asymptotic analysis. Let Z be defined as:
( )
1
Z(α) = α + sup Z(r) . (A20)
2 r∈[1/2,α]

This function shares the same “improvement” properties as Z, but is also nondecreasing
and continuous. The properties of the function Z(·) stated in the following lemma.

Lemma 8 The function Z : [1/2, 1] → [1/2, 1] given in (A20) satisfy the following
properties:

(a) For any α ∈ [1/2, 1], Z(α) ≥ α.

(b) If the private beliefs are unbounded, then for any α ∈ [1/2, 1), Z(α) > α.

(c) The function Z is increasing and continuous.

Proof. Parts (a) and (b) follow immediately from Lemma 7, parts (b) and (c) respec-
tively. The function supr∈[1/2,α] Z(r) is nondecreasing and the function α is increasing,
therefore the average of these two functions, which is Z, is an increasing function, es-
tablishing the first part of part (c).
We finally show that Z is a continuous function. We first show Z(α) is continuous
for all α ∈ [1/2, 1). To obtain a contradiction, assume that Z is discontinuous at
some α∗ ∈ [1/2, 1). This implies that supr∈[1/2,α] Z(r) is discontinuous at α∗ . Since
supr∈[1/2,α] Z(r) is a nondecreasing function, we have

lim sup Z(r) > sup Z(r),

α↓α∗ r∈[1/2,α] r∈[1/2,α∗ ]

from which it follows that there exists some ϵ > 0 such that for any δ > 0

sup Z(r) > Z(α) + ϵ for all α ∈ [1/2, α∗ ).

r∈[1/2,α∗ +δ]

This contradicts the fact that the function Z does not have an upward jump [cf. Lemma
7(a)], and establishes the continuity of Z(α) for all α ∈ [1/2, 1). The continuity of the
function Z(α) at α = 1 follows from part (a).
The next lemma shows that the function Z is also a (strong) improvement function
for the evolution of the probability of making the correct decision.

30
Lemma 9 (Strong Improvement Principle) Let B(n) = {b} for some n. Let σ ∈
Σ∗ igma∗ be an equilibrium. Then, we have
Pσ (xn = θ | B(n) = {b}) ≥ Z (Pσ (xb = θ)) . (A21)

Proof. Let α denote Pσ (xb = θ). If Z (α) = α, then the result is immediate. Suppose
next that Z (α) > α. This implies that Z(α) < supr∈[1/2,α] Z(r). Therefore, there exists
some α ∈ [1/2, α] such that
Z(α) > Z(α). (A22)
We next show that Pσ (xn = θ|B(n) = b) ≥ Z(α). Agent n can always (privately)
make the information from his observation of xb coarser (i.e., not observe xb according
to some probability). Let the observation thus generated by agent n be denoted by x̃b ,
and suppose that it is given by

 xb , with probability (2α − 1)/(2α − 1)
x̃b = 0, with probability (α − α)/(2α − 1)

1, with probability (α − α)/(2α − 1),
where the realizations of x̃b are independent from agent n’s information set. Next observe
that Pσ (x̃b = θ) = α. Then, Lemma 6 implies that Pσ (xn = θ|B(n) = b) ≥ Z(α). Since
Z(α) > Z(α) [cf. Eq. (A22)], the desired result follows.
We next generalize the results presented so far to an arbitrary network topology. We
first present an information monotonicity result with the amount of improvement given
by the improvement function Z defined in Eq. (A21). Even though a full characterization
of equilibrium decisions in general network topologies is a nontractable problem, it is
possible to establish an analogue of Lemma 9, that is, a generalized strong improvement
principle, which provides a lower bound on the amount of increase in the probabilities
of making the correct decision.

Lemma 10 (Generalized Strong Improvement Principle) For any n ∈ N, any set B ⊆

{1, ..., n − 1} and any equilibrium σ ∈ Σ∗ , we have
( )
Pσ (xn = θ | B(n) = B) ≥ Z max Pσ (xb = θ) .
b∈B

Proof. Given an equilibrium σ ∈ Σ∗ and agent n , let hσ be a function that maps

any subset of {1, . . . , n − 1} to an element of {1, . . . , n − 1} such that for any B ⊂
{1, . . . , n − 1}, we have
hσ (B) ∈ argmax Pσ (xb = θ). (A23)
b∈B

We define wn as the decision that maximizes the conditional probability of making the
correct decision given the private signal sn and the decision of the agent hσ (B(n)), i.e.,
( )
wn ∈ argmax Pσ y = θ sn , xhσ (B(n)) .
y∈{0,1}

The equilibrium decision xn of agent n satisfies

Pσ (xn = θ | sn , B(n), xk , k ∈ B(n)) ≥ Pσ (wn = θ | sn , B(n), xk , k ∈ B(n)),

31
[cf. the characterization of the equilibrium decision rule in Eq. (2)]. Integrating over all
possible private signals and decisions of neighbors, we obtain for any B ⊂ {1, . . . , n −1},

Pσ (xn = θ | B(n) = B) ≥ Pσ (wn = θ | B(n) = B) . (A24)

Because wn is an optimal choice given a single observation, Eq. (A21) holds and yields
( )
Pσ (wn = θ | B(n) = B) ≥ Z Pσ (xhσ (B) = θ) . (A25)

Combining Eqs. (A23), (A24) and (A25) we obtain the desired result.
This lemma is the key step in our proof. It shows that, under unbounded private
beliefs, there are improvements in payoffs (probabilities of making correct decisions)
that are bounded away from zero. We will next use this generalized strong improvement
principle to prove Theorem 2. The proof involves showing that under the expanding
observations and the unbounded private beliefs assumptions, the amount of improve-
ment in the probabilities of making the correct decision given by Z accumulates until
asymptotic learning is reached.

Proof of Theorem 2: The proof consists of two parts. In the first part of the
proof, we construct two sequences {αk } and {ϕk } such that for all k ≥ 0, there holds

Pσ (xn = θ) ≥ ϕk for all n ≥ αk . (A26)

The second part of the proof shows that ϕk converges to 1, thus establishing the result.
Given some integer K > 0 and scalar ϵ > 0, let N (K, ϵ) > 0 be an integer such that
for all n ≥ N (K, ϵ), ( )
Qn max b < K < ϵ,
b∈B(n)

(such an integer exists in view of the fact that, by hypothesis, the network topology
features expanding observations). We let α1 = 1 and ϕ1 = 1/2 and define the sequences
{αk } and ϕk recursively by
( [ ])
1 ϕk ϕk + Z(ϕk )
αk+1 = N αk , 1− , ϕk+1 = .
2 Z(ϕk ) 2

Using the fact that the range of the function Z is [1/2, 1], it can be seen that ϕk ∈ [1/2, 1]
for all k, therefore the preceding sequences are well-defined.
We use induction on the index k to prove relation (A26). Since σ is an equilibrium,
we have
1
Pσ (xn = θ) ≥ for all n ≥ 1,
2
which together with α1 = 1 and ϕ1 = 1/2 shows relation (A26) for k = 1. Assume that
the relation (A26) holds for an arbitrary k, i.e.,

Pσ (xj = θ) ≥ ϕk for all j ≥ αk . (A27)

32
Consider some agent n with n ≥ αk+1 . By integrating the relation from Lemma 10 over
all possible neighborhoods B(n), we obtain
[ ( )]
Pσ (xn = θ) ≥ EB(n) Z max Pσ (xb = θ) ,
b∈B(n)

where EB(n) denotes the expectation with respect to the neighborhood B(n) (i.e., the
weighted sum over all possible neighborhoods B(n)). We can rewrite the preceding as
[ ( ) ] ( )
Pσ (xn = θ) ≥ EB(n) Z max Pσ (xb = θ) max b ≥ αk Qn max b ≥ αk
b∈B(n) b∈B(n) b∈B(n)
[ ( ) ] ( )
+ EB(n) Z max Pσ (xb = θ) max b < αk Qn max b < αk .
b∈B(n) b∈B(n) b∈B(n)

Since the terms on the right hand-side of the preceding relation are nonnegative, this
implies that
[ ( ) ] ( )
Pσ (xn = θ) ≥ EB(n) Z max Pσ (xb = θ) max b ≥ αk Qn max b ≥ αk .
b∈B(n) b∈B(n) b∈B(n)

Since maxb∈B(n) b ≥ αk , Eq. (A27) implies that

max Pσ (xb = θ) ≥ ϕk .
b∈B(n)

Since the function Z is nondecreasing [cf. Lemma 8(c)], combining the preceding two
relations, we obtain
[ ] ( )
Pσ (xn = θ) ≥ EB(n) Z (ϕk ) max b ≥ αk Qn max b ≥ αk
b∈B(n) b∈B(n)
( )
= Z(ϕk )Qn max b ≥ αk ,
b∈B(n)

where the equality follows since the sequence {ϕk } is deterministic. Using the definition
of αk , this implies that
[ ]
1 ϕk
Pσ (xn = θ) ≥ Z (ϕk ) 1+ = ϕk+1 ,
2 Z(ϕk )
thus completing the induction.
We finally prove that ϕk → 1 as k → ∞. Since Z(α) ≥ α for all α ∈ [1/2, 1]
[cf. Lemma 8(a)], it follows from the definition of ϕk that {ϕk }k∈N is a nondecreasing
sequence. It is also bounded and therefore it converges to some ϕ∗ . Taking the limit in
the definition of ϕk , we obtain
[ ]
2ϕ = 2 lim ϕk = lim ϕk + Z(ϕk ) = ϕ∗ + Z(ϕ∗ ),
∗
k→∞ k→∞

where the third equality follows since Z is a continuous function [cf. Lemma 8(c)]. This
shows that ϕ∗ = Z(ϕ∗ ), i.e., ϕ∗ is a fixed point of Z. Since the private beliefs are
unbounded, the unique fixed point of Z is 1, showing that ϕk → 1 as k → ∞ and
completing the proof. 

33
Proofs of Corollaries 1 and 2
Proof of Corollary 1. We first show that if c ≥ 1, then the network topology has
nonexpanding observations. To show this, we set K = 1 in Definition 5 and show
that the probability of infinitely many agents having empty neighborhoods is uniformly
bounded away from 0. We first consider the case c > 1. Then, the probability that the
neighborhood of agent n + 1 is the empty set is given by
( a )n
Qn+1 (B(n + 1) = ∅) = 1 − c ,
n
which converges to 1 as n goes to infinity. If c = 1, then
( a )n
lim Qn+1 (B(n + 1) = ∅) = lim 1 − = e−a .
n→∞ n→∞ n
Therefore, for infinitely many agents, Qn+1 (B(n + 1) = ∅) ≥ e−a /2. The preceding show
that the network topology has nonexpanding observations for c ≥ 1, hence the result
follows from Theorem 1.
We next assume that c < 1. For any K and all n ≥ K, we have
( ) (
a )n−K
Qn+1 max b ≤ K = 1 − c ,
b∈B(n+1) n

which converges to 0 as n goes to infinity. Hence the network topology is expanding in

observations and the result follows from Theorem 2. 
Proof of Corollary 2. We show that {L(n)}n∈N goes to infinity if and only if the de-
terministic sequence {maxb∈B(n) b}n∈N goes to infinity. Suppose first {L(n)}n∈N diverges
to infinity. Then, for every K there exists N such that for all n ≥ N , L(n) ≥ K. Note
that
L(n) ≤ 1 + max b
b∈B(n)

because the longest information path must be a subset of the sequence

(1, 2, ..., maxb∈B(n) b, n). So, for n ≥ N , if L(n) ≥ K, then maxb∈B(n) b > K, thus
proving the first part of the lemma. Suppose next that {maxb∈B(n) b}n∈N goes to infinity
as n goes to infinity. We show by induction that for each d ∈ N, there exists some integer
cd such that L(n) ≥ d for all n ≥ cd . Since L(n) ≥ 1 for all n, then c1 = 1. Assume such
cd exists for some d. Then, we show that such a cd+1 also exists. Since {maxb∈B(n) b}n∈N
goes to infinity, there exists some Nd such that for all n ≥ Nd ,

max b ≥ cd .
b∈B(n)

Now, for any n ≥ Nd , there exists a path with size d up to some k ≥ cd and then another
observation from k to n, therefore L(n) ≥ d + 1. Hence, cd+1 = Nd . 

34
Proof of Theorem 4
This proof consists of two parts. In the first part, we show that agents who observe
the full history of actions are able to learn the state of the world. This part of the
proof is centered on the martingale convergence of beliefs for agents who observe the
entire history. The second part of the proof uses the improvement principle to extend
the asymptotic learning result to all agents. For simplicity, we assume in this proof (as
well as in the proof of the next two propositions) that all agents break ties in favor of
action 0. Relaxing this assumption does not change our results, but would necessitate
additional notation.
Note that the conditions of the theorem do not hold if the set S is finite. We,
therefore, assume that S has infinite cardinality without loss of generality.
Part 1. For each n, let xn = (x1 , . . . , xn ) represent the sequence of decisions up
to and including xn . Let q ∗ (xn ) denote the “social belief” when xn is observed under
equilibrium σ, i.e.,
q ∗ (xn ) = Pσ (θ = 1 | xn ).
The social belief q ∗ (xn ) is a martingale and, by the martingale convergence theorem,
converges with probability 1 to some random variable q̂. Conditional on θ = 1, the
likelihood ratio
1 − q ∗ (xn ) Pσ (θ = 0 | xn )
=
q ∗ (xn ) Pσ (θ = 1 | xn )
is also a martingale [see Doob, 1953, Eq. (7.12)]. Therefore, conditional on θ = 1, the
ratio (1−q ∗ (xn ))/q ∗ (xn ) converges with probability 1 to some random variable (1−q̂1 )/q̂1 .
In particular, we have [ ]
1 − q̂1
Eσ < ∞,
q̂1
[see Breiman, Theorem 5.14], and therefore q̂1 > 0 with probability 1. Similarly,
q ∗ (xn )/(1 − q ∗ (xn )) is a martingale conditional on θ = 0 and converges with proba-
bility 1 to some random variable q̂0 /(1 − q̂0 ), where q̂0 < 1 with probability 1. Therefore,
Pσ (q̂ > 0 | θ = 1) = 1 and Pσ (q̂ < 1 | θ = 0) = 1. (A28)
The key element of part 1 of the proof is to show that the support of q̂ is contained in
the set {0, 1}. This fact combined with Eq. (A28) guarantees that q̂ = θ (i.e., the agents
that observe the entire history eventually know what the state of the world θ is).
To show that the support of q̂ is contained in {0, 1}, we study the evolution dynamics
of q ∗ (xn ). Suppose xn+1 = 0. Using Bayes’ Rule twice, we have
[ ]−1
∗ Pσ (xn+1 = 0, xn | θ = 1) Pσ (xn+1 = 0, xn | θ = 0)
q ((x , 0)) = ∑1
n
= 1+
j=0 Pσ (xn+1 = 0, x | θ = j)
n Pσ (xn+1 = 0, xn | θ = 1)
[ ( ) ]−1
1 Pσ (xn+1 = 0 | θ = 0, xn )
= 1+ −1 .
q ∗ (xn ) Pσ (xn+1 = 0 | θ = 1, xn )
To simplify notation, let
Pσ (xn+1 = 0 | θ = 0, xn )
fn (xn ) = , (A29)
Pσ (xn+1 = 0 | θ = 1, xn )

35
so that
[ ( ) ]−1
∗ 1
n
q ((x , 0)) = 1 + − 1 fn (x )
n
. (A30)
q ∗ (xn )

We next show that if q ∗ (xn ) ≥ 1/2 for some n + 1 ∈ S, we have

fn (xn ) ≥ 1 + δ for some δ > 0, (A31)

which will allow us to establish a bound on the difference between q ∗ (xn ) and q ∗ ((xn , 0)).
By conditioning on the neighborhood B(n + 1), we have for any j ∈ {0, 1},
∑
Pσ (xn+1 = 0 | θ = j, xn ) = Qn+1 (B(n+1) = B)Pσ (xn+1 = 0 | θ = j, xn , B(n+1) = B).
B
(A32)
Define the auxiliary measure Q̃n+1 over the subsets of {1, ..., n} to have the following
value for any B ⊆ {1, ..., n},
{
Qn+1 (B(n + 1) = B), if B ̸= {1, ..., n};
Q̃n+1 (B) =
Qn+1 (B(n + 1) = B) − ϵ, if B = {1, ..., n},

where ϵ > 0 is given by the lower bound on the probability of observing the history of
actions for n + 1 ∈ S. Note that by the definition of ϵ, Q̃n+1 is a non-negative measure.
Rewriting Eq. (A32) in terms of Q̃n+1 , we obtain

Pσ (xn+1 = 0 | θ = j, xn ) = ϵPσ (xn+1 = 0 | θ = j, xn , B(n + 1) = {1, ..., n}) +(A33)

∑
Q̃n+1 (B)Pσ (xn+1 = 0 | θ = j, xn , B(n + 1) = B).
B

Let q̃(B, xn ) denote the “social belief” when the subset B of actions of xn is observed
under equilibrium σ, i.e.,

q̃(B, xn ) = Pσ (θ = 1 | xk for all k ∈ B).

Using the assumption that agents break ties in favor of action 0, Proposition 2 implies
that for any B ⊆ {1, ..., n},

Pσ (xn+1 = 0 | θ = j, xn , B(n + 1) = B) = Gj (1 − q̃(B, xn )).

Hence, Eq. (A33) can be rewritten as

∑
Pσ (xn+1 = 0 | θ = j, xn ) = ϵGj (1 − q ∗ (xn )) + Q̃n+1 (B)Gj (1 − q̃(B, xn )).
B

Returning to Eq. (A29), we have that

∑
ϵG0 (1 − q ∗ (xn )) + B Q̃n+1 (B)G0 (1 − q̃(B, xn ))
n
fn (x ) = ∑ .
ϵG1 (1 − q ∗ (xn )) + B Q̃n+1 (B)G1 (1 − q̃(B, xn ))

36
From Lemma 1(c), we know that G0 (r) ≥ G1 (r) for any r and, therefore,
∑
ϵG0 (1 − q ∗ (xn )) + B Q̃n+1 (B)G1 (1 − q̃(B, xn ))
fn (x ) ≥
n
∑ .
ϵG1 (1 − q ∗ (xn )) + B Q̃n+1 (B)G1 (1 − q̃(B, xn ))

We also obtain from Lemma 1(c) that

G0 (1/2)
G0 (1 − q ∗ (xn )) ≥ G1 (1 − q ∗ (xn )),
G1 (1/2)
since q ∗ (xn ) is assumed to be greater than or equal to 1/2. Therefore,
∑
ϵG0 (1/2)
G (1 − q ∗ (xn )) + B Q̃n+1 (B)G1 (1 − q̃(B, xn ))
G1 (1/2) 1
fn (x ) ≥
n
∑ .
ϵG1 (1 − q ∗ (xn )) + B Q̃n+1 (B)G1 (1 − q̃(B, xn ))
G0 (1/2)
Note that since > 1, the right-hand side of the equation above is increasing in
G1 (1/2)
∑
G1 (1 − q (x )) and decreasing in B Q̃n+1 (B)G1 (1 − q̃(B, xn )). Since G1 (1 − q ∗ (xn )) ≤
∗ n
∑ ∑
G1 (1/2) and B Q̃n+1 (B)G1 (1 − q̃(B, xn )) ≥ B Q̃n+1 (B) = 1 − ϵ, we obtain

ϵG0 (1/2) + 1 − ϵ
fn (xn ) ≥ =1+δ for some δ > 0 if n + 1 ∈ S,
ϵG1 (1/2) + 1 − ϵ
thus proving Eq. (A31).
We now show that the support of the limiting social belief q̂ does not include the
interval (1/2, 1). Combining Eq. (A31) with Eq. (A30) yields
[ ( ) ]−1
∗ 1
q ((x , 0)) ≤ 1 +
n
− 1 (1 + δ) if q ∗ (xn ) ≥ 1/2 and n + 1 ∈ S. (A34)
q ∗ (xn )

Suppose now q ∗ (xn ) ∈ [1/2, 1 − ϵ] for some ϵ > 0. We show that there exists some
constant K(δ, ϵ) > 0 such that

q ∗ (xn ) − q ∗ ((xn , 0)) ≥ K(δ, ϵ) if n + 1 ∈ S. (A35)

Define g : [1/2, 1 − ϵ] → [0, 1] as

[ ( ) ]−1
1
g(q) = q − 1 + − 1 (1 + δ) .
q

It can be seen that g(q) is a concave function over q ∈ [1/2, 1 − ϵ]. Let K(δ, ϵ) be
{ }
δ ϵδ(1 − ϵ)
K(δ, ϵ) = inf g(q) = min{g(1/2), g(1 − ϵ)} = min , > 0.
q∈[1/2,1−ϵ] 2(2 + δ) 1 + ϵδ

From Eq. (A34), it follows that

[ ( ) ]−1
∗ ∗ ∗ 1
q (x ) − q ((x , 0)) ≥ q (x ) − 1 +
n n n
− 1 (1 + δ) ≥ g (q ∗ (xn )) ≥ K(δ, ϵ),
q ∗ (xn )

37
thus proving Eq. (A35).
Recall that q ∗ (xn ) converges to q̂ with probability 1. We show that for any ϵ > 0, the
support of q̂ does not contain (1/2, 1 − ϵ). Assume, to arrive at a contradiction, that it
does. Consider a sample path that converges to a value in the interval (1/2, 1 − ϵ). For
this sample path, there exists some N such that for all n ≥ N , q ∗ (xn ) ∈ [1/2, 1 − ϵ]. By
the Borel-Cantelli Lemma, there are infinitely many agents n within S that observe a
nonpersuasive
∑ ∑Mneighborhood Ci for some i because the neighborhoods are independent
and n∈S i=1 ri (n) = ∞. Since these are nonpersuasive neighborhoods, an infinite
subset will choose action 0. Therefore, for infinitely many n that satisfy n ≥ N and
n ∈ S, by Eq. (A35),
q ∗ (xn+1 ) ≤ q ∗ (xn ) − K(δ, ϵ).
But this implies the sequence q ∗ (xn ) is not Cauchy and contradicts the fact that q ∗ (xn )
is a convergent sequence. Hence, we conclude that the support of q̂ does not contain
(1/2, 1 − ϵ). Since this argument holds for any ϵ > 0, the support of q̂ cannot contain
(1/2, 1). A similar argument leads to the conclusion that the support of q̂ does not
include (0, 1/2]. Therefore, the support of q̂ is a subset of the set {0, 1}. By Eq. (A28),
this implies that q̂ = θ with probability 1. Combining the result above with Proposition
2,

lim Pσ (xn = θ|B(n) = {1, ..., n − 1}) = lim Pσ (⌊pn + qn ⌋ = θ|B(n) = {1, ..., n − 1})
n→∞ n→∞
= lim Pσ (⌊pn + q ∗ (xn−1 )⌋ = θ)
n→∞
= lim Pσ (⌊pn + θ⌋ = θ) = 1,
n→∞

where pn and qn are respectively the private and social beliefs of agent n. We thus
conclude that the subsequence of agents that observe the full history of actions
eventually learn the state of the world θ.

Part 2. In this part of the proof, we show that asymptotic learning occurs. We
first show that agents in S asymptotically learn the state θ. In the previous part, we
obtained that Pσ (xn = θ|B(n) = {1, ..., n − 1}) converges to 1 as n goes to infinity.
Define {ψn }n∈S to be a nondecreasing sequence of numbers such that limn→∞ ψn = 1
and for all n ∈ S,
Pσ (xn = θ|B(n) = {1, ..., n − 1}) ≥ ψn .
Define also a sequence of binary random variables {Cn }n∈S such that Cn = 1 implies
that B(n) = {1, ..., n − 1} and that, conditionally on B(n), Cn is independent of all
other events. Furthermore, set Qn (Cn = 1) = ϵ, where ϵ is given by the lower bound on
the probability of observing the history of actions for S. Conditioning the event xn = θ
on Cn we obtain that for any n ∈ S,

Pσ (xn = θ) = Pσ (xn = θ|Cn = 0)(1 − ϵ) + Pσ (xn = θ|Cn = 1)ϵ

≥ Pσ (xn = θ|Cn = 0)(1 − ϵ) + ψn ϵ. (A36)

Note that expanding observations with respect to S holds conditionally on Cn = 0 since

38
for any K,
( ) ( )
0 = lim Qn max b < K ≥ lim Qn max b < K Cn = 0 (1 − ϵ) ≥ 0.
n→∞ b∈B(n)∩S n→∞ b∈B(n)∩S

We can thus define a function N (K, δ) such that for all K ∈ N, δ > 0 and n ≥ N (K, δ),
( )
Qn max b < K Cn = 0 < δ. (A37)
b∈B(n)∩S

We now show inductively that agents in S asymptotically learn the state θ. We construct
two sequences {ϕK }K∈N and {αK }K∈N such that for all n ∈ S and all K ∈ N,

Pσ (xn = θ) ≥ ϕK for all n ≥ αK . (A38)

Let α1 = 1 and ϕ1 = 1/2 and the equation above holds for K = 1. Assume Eq. (A38)
holds for some K, and let
ϵ
ϕK+1 = ϕK + (1 − ϕK ). (A39)
2
We now show that there exists some αK+1 such that for all n ∈ S and n ≥ αK+1 ,
Pσ (xn = θ) ≥ ϕK+1 . From Eq. (A36), we have that for any n ∈ S,

Pσ (xn = θ) ≥ Pσ (xn = θ|Cn = 0)(1 − ϵ) + ψn ϵ

[ ]
≥ Eσ max Pσ (xb = θ) Cn = 0 (1 − ϵ) + ψn ϵ,
b∈B(n)∩S

where the second inequality follows from the fact that agent n can select a better action
(ex ante) than anyone in his neighborhood (cf. Lemma 2). From Eqs. (A37) and (A38),
we obtain that for any δ > 0, n ∈ S and n ≥ N (αK , δ),

Pσ (xn = θ) ≥ ϕK (1 − δ)(1 − ϵ) + ψn ϵ.

By selecting a large n large (in order to have ψn close to 1) and a small δ, we can
have ϕK (1 − δ)(1 − ϵ) + ψn ϵ be arbitrarily close to ϕK + ϵ(1 − ϕK ). Therefore, there
exists some αK+1 that satisfies Eq. (A39). Therefore, Eq. (A38) holds. To complete
the argument that agents in S asymptotically learn the state note that limK→∞ ϕK = 1
from the recursion in Eq. (A39) and

lim Pσ (xn = θ) ≥ lim ϕK = 1.

n→∞,n∈S K→∞

We now show that all agents asymptotically learn the state θ. From Lemma 2 we
obtain that
[ ]
Pσ (xn = θ) ≥ Eσ max Pσ (xb = θ)
b∈B(n)∩S
[ ] ( )
≥ Eσ max Pσ (xb = θ) max b ≥ K Qn max b ≥ K ,
b∈B(n)∩S b∈B(n)∩S b∈B(n)∩S

39
for any K ∈ N. Since agents in S learn the state, for any K,
[ ]
lim Eσ max Pσ (xb = θ) max b ≥ K = 1.
n→∞ b∈B(n)∩S b∈B(n)∩S

From the expanding observations condition with respect to S we obtain that for all K,
( )
lim Qn max b ≥ K = 1.
n→∞ b∈B(n)∩S

Therefore, limn→∞ Pσ (xn = θ) = 1, completing the proof. 

Proofs of Propositions 3 and 4

Proof of Proposition 3. Using Bayes’ Rule and the assumption of equal priors on the
state θ, we have that the social belief given by observing neighborhood B, with actions
xk = yk for each k ∈ B in equilibrium σ is given by
( )−1
Pσ (xk = yk for all k ∈ B|θ = 0)
Pσ (θ = 1|xk = yk for all k ∈ B) = 1 + . (A40)
Pσ (xk = yk for all k ∈ B|θ = 1)
Since the neighborhoods of the agents in B are all empty, their actions are independent
conditional on the state θ. Hence,
( ∏ )−1
k∈B Pσ (xk = yk |θ = 0)
Pσ (θ = 1|xk = yk for all k ∈ B) = 1+ ∏ .
k∈B Pσ (xk = yk |θ = 1)

Given the assumption that all agents break ties in favor of action 0, we can obtain that
for each agent n with an empty neighborhood, and each j ∈ {0, 1}, Pσ (x1 = 0|θ = j) =
Gj (1/2) and Pσ (x1 = 1|θ = j) = 1 − Gj (1/2). Therefore, the social belief of Eq. (A40)
can be rewritten as
( ∑ ∑ )−1
(1 − G0 (1/2)) k∈B yk G0 (1/2)|B|− k∈B yk
Pσ (θ = 1|xk = yk for all k ∈ B) = 1 + ∑ ∑ .
(1 − G1 (1/2)) k∈B yk G1 (1/2)|B|− k∈B yk
(A41)
From Lemma 1(c), we have that
G0 (1/2) 1 − G0 (1/2)
>1> ,
G1 (1/2) 1 − G1 (1/2)
which combined with Eq. (A41) implies
min Pσ (θ = 1|xk = yk for all k ∈ B)
B⊆{1,...,K},{yk }k≤K ∈{0,1}K
( )−1
G0 (1/2)K
= Pσ (θ = 1|xk = 0 for all k ∈ {1, ..., K}) = 1+ and
G1 (1/2)K
max Pσ (θ = 1|xk = yk for all k ∈ B)
B⊆{1,...,K},{yk }k≤K ∈{0,1}K
( )−1
(1 − G0 (1/2))K
= Pσ (θ = 1|xk = 1 for all k ∈ {1, ..., K}) = 1+ .
(1 − G1 (1/2))K

40
Therefore, any such neighborhood B ⊆ {1, ..., K} is nonpersuasive if it satisfies the
conditions
( )−1 ( )−1
G0 (1/2)K (1 − G0 (1/2))K
1+ > 1 − β and 1+ < 1 − β, (A42)
G1 (1/2)K (1 − G1 (1/2))K
which together yield Eq. (5).

Proof of Proposition 4. We start by showing inductively that all networks of the

form {1, ..., n} for any n ≤ K are nonpersuasive. In particular, we first show that for all
n ≤ K, and any set of values yk ∈ {0, 1} for each k ≤ n,
( ( )n )−1
G0 (1/2)
Pσ (θ = 1|xk = yk for all k ≤ n) ≥ 1 + . (A43)
G1 (1/2)
The upper bound on Pσ (θ = 1|xk = yk for all k ≤ n) is symmetric and we do not prove
it to avoid repetition.
Note that Eq. (A43) is trivially true for n = 0, i.e., Pσ (θ = 1) = 1/2. Assume that
Eq. (A43) is true for n − 1 as an inductive hypothesis. We now show it is true for n.
Let the social belief given the neighborhood {1, ..., n} and decisions y n = (y1 , ..., yn ) be
given by q ∗ (y n ) = Pσ (θ = 1|xk = yk for all k ≤ n). Then, using Bayes’ Rule, we obtain
that
( )
1 1 Pσ (xn = yn |θ = 0, xk = yk for all k ≤ n − 1)
− 1 = − 1 .
q ∗ (y n ) q ∗ (y n−1 ) Pσ (xn = yn |θ = 1, xk = yk for all k ≤ n − 1)
Since agent n observes the full history of past actions and breaks ties in favor of action
0, we have
 ( )

 1
− 1 1−G
∗ n−1 ))
0 (1−q (y
, if yn = 1;
1  ( q ∗ (y n−1 )
) 1−G1 (1−q ∗ (y n−1 ))
−1= 1
−1 G
∗
0 (1−q (y
n−1 ))
, if yn = 0.
q ∗ (y n ) 
 q ∗ (y n−1 ) G1 (1−q ∗ (y n−1 ))


If yn = 1, then q ∗ (y n ) ≥ q ∗ (y n−1 ) because

1 − G0 (1 − r))
≤ 1 for any r ≥ β,
1 − G1 (1 − r))
by Lemma 1(c). Using the inductive hypothesis,
( ( )n−1 )−1 ( ( )n )−1
∗ n ∗ n−1 G0 (1/2) G0 (1/2)
q (y ) ≥ q (y ) ≥ 1 + ≥ 1+ ,
G1 (1/2) G1 (1/2)

thus proving Eq. (A43) for yn = 1. If yn = 0, then we need to consider two separate
cases. Suppose first q ∗ (y n−1 ) < 1/2. Then, by Lemma 1(c), we have that
G0 (1 − q ∗ (y n−1 )) G0 (1/2)
≤ .
G1 (1 − q (y ))
∗ n−1 G1 (1/2)

41
Combining the relation above with the inductive hypothesis, we obtain
( ) ( ) ( )n
1 G0 (1 − q ∗ (y n−1 )) 1 G0 (1/2) G0 (1/2)
−1 ≤ −1 ≤ ,
q ∗ (y n−1 ) G1 (1 − q ∗ (y n−1 )) q ∗ (y n−1 ) G1 (1/2) G1 (1/2)

thus proving that Eq. (A43) holds in this case as well. Finally, consider the case where
q ∗ (y n−1 ) ≥ 1/2. Here, we use the monotonicity assumption from Eq. (8) to show that
( ) ( ) ( )n
1 1 G0 (1 − q ∗ (y n−1 )) 1 G0 (1/2) G0 (1/2)
−1 = −1 ≤ −1 ≤ ,
q ∗ (y n ) q ∗ (y n−1 ) G1 (1 − q ∗ (y n−1 )) 1/2 G1 (1/2) G1 (1/2)

thus completing the proof of Eq. (A43). We thus conclude that any neighborhood
{1, ..., n} is nonpersuasive if n ≤ K and K satisfies
( ( )K )−1
G0 (1/2)
1+ > 1 − β,
G1 (1/2)

as well as the symmetric (upper bound)

( ( )K )−1
1 − G0 (1/2)
1+ < 1 − β.
1 − G1 (1/2)

Both conditions on K combined yield the condition of Eq. (7).

We have proved that the set {1, ..., K} is nonpersuasive given the conditions of the
proposition. We now show that any subset of B ⊆ {1, ..., K} is also nonpersuasive. To
prove this result, we bound by the worst possible event in terms of the actions of agents
not B with respect to the social belief, i.e.,

Pσ (θ = 1|xk = yk for all k ∈ B)

∑
= Pσ (θ = 1|xk = yk for k ∈ B, xk = zk for k ∈ {1, ..., K}\B)
z n ∈{0,1}{1,...,K}\B

Pσ (xk = zk for all k ∈ {1, ..., K}\B|xk = yk for all k ∈ B)

≥ min Pσ (θ = 1|xk = yk for k ∈ B, xk = zk for k ∈ {1, ..., K}\B).
z n ∈{0,1}{1,...,K}\B

Use a symmetric bound to show the maximum value of the social belief. The proof is,
therefore, complete since we have previously proven that for the neighborhood {1, ..., K}
is nonpersuasive. 

42
References
[1] Bala V. and Goyal S., “Learning from Neighbours,” The Review of Economic Stud-
ies, vol. 65, no. 3, pp. 595-621, 1998.

[2] Bala V. and Goyal S., “Conformism and Diversity under Social Learning,” Economic
Theory, vol. 17, pp. 101-120, 2001.

[3] Banerjee A., “A Simple Model of Herd Behavior,” The Quarterly Journal of Eco-
nomics, vol. 107, pp. 797-817, 1992.

[4] Banerjee A. and Fudenberg D., “Word-of-mouth Learning,” Games and Economic
Behavior, vol. 46, pp. 1-22, 2004.

[5] Besley T. and Case A., “Diffusion as a Learning Process: Evidence from HYV
Cotton,” Working Papers 228, Princeton University, Woodrow Wilson School of
Public and International Affairs, Research Program in Development Studies, 1994.

[6] Bikhchandani S., Hirshleifer D., and Welch I., “A Theory of Fads, Fashion, Custom,
and Cultural Change as Information Cascades,” The Journal of Political Economy,
vol. 100, pp. 992-1026, 1992.

[7] Bikhchandani S., Hirshleifer D., and Welch I., “Learning from the Behavior of
Others: Conformity, Fats, and Informational Cascades,” The Journal of Economic
Perspectives, vol. 12, pp. 151-170, 1998.

[8] Breiman L., Probability Addison-Wellesley, 1968.

[9] Callander S. and Horner J., “The Wisdom of the Minority,” Northwestern mimeo,
2006.

[10] Celen B. and Kariv S., “Observational Learning Under Imperfect Information,”
Games and Economic Behavior, vol. 47, no. 1, pp. 72-86, 2004.

[11] Celen B. and Kariv S., “Distinguishing Informational Cascades from Herd Behavior
in the Laboratory,” The American Economic Review, vol. 94, no. 3, pp. 484-498,
2004.

[12] Chamley C. and Gale D., “Information Revelation and Strategic Delay in a Model
of Investment,” Econometrica, vol. 62, pp. 1065-1086, 1994.

[13] Choi S., Celen B., and Kariv S., “Learning in Networks: An Experimental Study,”
UCLA Technical Report, Dec. 2005.

[14] Condorcet N.C. , “Essai sur l’Application de l’Analyze à la Probabilité des Décisions
Rendues à la Pluralité des Voix,” Imprimérie Royale, Paris, France, 1785.

[15] Cover T., “Hypothesis Testing with Finite Statistics,” The Annals of Mathematical
Statistics, vol. 40, no. 3, pp. 828-835, 1969.

43
[16] DeMarzo P.M., Vayanos D., and Zwiebel J., “Persuasion Bias, Social Influence, and
Unidimensional Opinions,” The Quarterly Journal of Economics, vol. 118, no. 3,
pp. 909-968, 2003.

[17] Doob J.L., Stochastic Processes, John Wiley and Sons Inc., 1955.

[18] Ellison G. and Fudenberg D., “Rules of Thumb for Social Learning,” The Journal
of Political Economy, vol. 101, no. 4, pp. 612-643, 1993.

[19] Ellison G. and Fudenberg D., “Word-of-mouth communication and social learning,”
The Quarterly Journal of Economics, vol. 110, pp. 93-126, 1995.

[20] Foster A. and Rosenzweig M., “Learning by Doing and Learning from Others: Hu-
man Capital and Technical Change in Agriculture,” The Journal of Political Econ-
omy, vol. 103, no. 6, pp. 1176-1209, 1995.

[21] Gale D. and Kariv S., “Bayesian Learning in Social Networks,” Games and Eco-
nomic Behavior, vol. 45, no. 2, pp. 329-346, 2003.

[22] Golub B. and Jackson M.O., “Naive Learning in Social Networks: Convergence,
Influence, and the Wisdom of Crowds,” unpublished manuscript, 2007.

[23] Granovetter M., “The Strength of Weak Ties,” The American Journal of Sociology,
vol. 78, no. 6, pp. 1360-1380, 1973.

[24] Ioannides M. and Loury L., “Job Information Networks, Neighborhood Effects, and
Inequality,” The Journal of Economic Literature, vol. 42, no. 2, pp. 1056-1093, 2004.

[25] Jackson M.O., “The Economics of Social Networks,” chapter in “Volume I of Ad-
vances in Economics and Econometrics, Theory and Applications: Ninth World
Congress of the Econometric Society”, Richard Blundell, Whitney Newey, and
Torsten Persson, Cambridge University Press, 2006.

[26] Jackson M.O., “The Study of Social Networks In Economics,” chapter in “The
Missing Links: Formation and Decay of Economic Networks”, edited by James E.
Rauch; Russell Sage Foundation, 2007.

[27] Jackson M.O. and Wolinsky A., “A Strategic Model of Social and Economic Net-
works,” The Journal of Economic Theory, vol. 71, pp. 44-74, 1996.

[28] Jackson M.O. and Watts A., “The Evolution of Social and Economic Networks,”
The Journal of Economic Theory, vol. 106, pp. 265-295, 2002.

[29] Lee I., “On the Convergence of Informational Cascades,” The Journal of Economic
Theory, vol. 61, pp. 395-411, 1993.

[30] Lorenz J., Marciniszyn M., and Steger A., “Observational Learning in Random
Networks,” chapter in “Learning Theory”, Springer Berlin / Heidelberg, 2007.

44
[31] Montgomery J.D., “Social Networks and Labor-Market Outcomes: Toward an Eco-
nomic Analysis,” The American Economic Review, vol. 81 no. 5, pp. 1408-1418,
1991.

[32] Munshi K., “Networks in the Modern Economy: Mexican Migrants in the U.S.
Labor Market,” The Quarterly Journal of Economics, vol. 118 no. 2, pp. 549-597,
2003.

[33] Munshi K., “Social Learning in a Heterogeneous Population: Technology Diffusion

in the Indian Green Revolution,” The Journal of Development Economics, vol. 73
no. 1, pp. 185-215, 2004.

[34] Myerson R.B., “Extended Poisson Games and the Condorcet Jury Theorem,”
Games and Economic Behavior, vol. 25, no. 1, pp. 111-131, 1998.

[35] Myerson R.B., “Large Poisson Games,” The Journal of Economic Theory, vol. 94
no. 1, pp. 7-45, 2000.

[36] Papastavrou J.D. and Athans M., “Distributed Detection by a Large Team of Sen-
sors in Tandem,” Proceedings of IEEE CDC, 1990.

[37] Sgroi D., “Optimizing Information in the Herd: Guinea Pigs, Profits, and Welfare,”
Games and Economic Behavior, vol. 39, pp. 137-166, 2002.

[38] Smith L. and Sorensen P., “Rational Social Learning with Random Sampling,”
unpublished manuscript, 1998.

[39] Smith L. and Sorensen P., “Pathological Outcomes of Observational Learning,”

Econometrica, vol. 68, no. 2, pp. 371-398, 2000.

[40] Tay W.P., Tsitsiklis J.N., and Win M.Z., “On the Sub-Exponential Decay of De-
tection Error Probabilities in Long Tandems,” Proceedings of ICASSP, 2007.

[41] Udry C. and Conley T., “Social Learning through Networks: the Adoption of New
Agricultural Technologies in Ghana,” American Journal of Agricultural Economics,
vol. 83, pp. 668-673, 2001.

[42] Vives X., “Learning from Others: A Welfare Analysis,” Games and Economic Be-
havior, vol. 20, pp. 177-200, 1997.

[43] Welch I., “Sequential Sales, Learning and Cascades,” The Journal of Finance, vol.
47, pp. 695-732, 1992.

[44] Young H.P., “Condorcet’s Theory of Voting,” The American Political Science Re-
view, vol. 82, pp. 1231–1244, 1988.

45
Appendix B: Omitted Proofs from “Bayesian Learn-
ing in Social Networks” (Not for Publication)
Proof of Lemma 1
(a) By the definition of a private belief, we have for any pn ∈ (0, 1),
P(θ = 1|sn ) = P(θ = 1|pn ).
Using Bayes’ Rule, it follows that
dP(pn |θ = 1)P(θ = 1) dP(pn |θ = 1) dG1 (pn )
pn = Pσ (θ = 1|pn ) = ∑1 = ∑1 = ∑1 .
j=0 dP(pn |θ = j)P(θ = j) j=0 dP(pn |θ = j) j=0 dGj (pn )

Because of the assumption that no signal is completely informative, i.e., pn ∈

/ {0, 1}, we
can rewrite this equation as
dG0 1 − pn
(pn ) = ,
dG1 pn
completing the proof.
(b) For any p ∈ (0, 1),
∫ p ∫ ( ) ∫ ( )
p
1−r 1−p p
1 1
G0 (p) = dG0 (r) = dG1 (r) = G1 (p) + − dG1 (r),
r=0 r=0 r p r=0 r p
where the second equality follows from part (a) of this lemma. We can provide a lower
bound on the last integral as
∫ p ( ) ∫ z ( ) ∫ z ( )
1 1 1 1 1 2 p−z
− dG1 (r) ≥ − dG1 (r) ≥ − dG1 (r) ≥ G1 (z) ,
r=0 r p r=0 r p r=0 z z+p 2
for any z ∈ (0, p). Equivalently, the second relation is obtained by
∫ 1 ∫ 1
r
1 − G1 (p) = dG1 (r) = dG0 (r)
r=p r=p 1 − r
( ) ∫ 1 ( )
p r p
= (1 − G0 (p)) + − dG0 (r),
1−p r=p 1−r 1−p
where the following bound is valid for any p < w < 1,
∫ 1 ( ) ∫ 1 ( )
r p r p
− dG0 (r) ≥ − dG0 (r)
r=p 1−r 1−p r=w 1−r 1−p
∫ 1 ( )
w p+w w−p
≥ − dG0 (r) ≥ (1 − G0 (w)).
r=w 1−w 2−p−w 2

(c) From part (a), we have for any r ∈ (0, 1),

∫ r ∫ r ( ) ∫ r ( ) ( )
1−x 1−r 1−r
G0 (r) = dG0 (x) = dG1 (x) ≥ dG1 (x) = G1 (r).
x=0 x=0 x x=0 r r
(B1)

1
Using part (a) again,
( )
G0 (r) dG0 (r)G1 (r) − G0 (r)dG1 (r)
d =
G1 (r) (G (r))2
[( 1 ) ]
dG1 (r) 1−r
= G1 (r) − G0 (r) .
(G1 (r))2 r
Since G1 (r) > 0 for r > β, dG1 (r) ≥ 0 and the term in brackets above is non-positive
by Eq. (B1), we have
( )
G0 (r)
d ≤ 0,
G1 (r)
thus proving the ratio G0 (r)/G1 (r) is non-increasing.
We now show that
G0 (r) ≥ G1 (r) for all r ∈ [0, 1]. (B2)
From Eq. (B1), we obtain that Eq. (B2) is true for r ≤ 1/2. For r > 1/2,
∫ 1 ∫ 1 ( ) ∫ 1
1−x
1 − G0 (r) = dG0 (x) = dG1 (x) ≤ dG1 (x) = 1 − G1 (r),
x=r x=r x x=r

thus proving Eq. (B2).

We proceed to prove the second part of the lemma. Suppose G0 (r)/G1 (r) = 1 for
some r < β. Suppose first r ∈ (1/2, β). Then,
∫ 1
G0 (1) = G0 (r) + dG0 (x)
x=r
∫ 1
= G1 (r) + dG0 (x)
∫ 1 (
x=r
)
1−x
= G1 (r) + dG1 (x)
x=r x
( )∫ 1
1−r
≥ G1 (r) + dG1 (x)
r
( ) x=r
1−r
≥ G1 (r) + [1 − G1 (r)] ,
r

which yields a contradiction unless G1 (r) = 1. However, G1 (r) = 1 implies r ≥ β – also a

contradiction. Now, suppose r ∈ (β, 1/2]. Since the ratio G0 (r)/G1 (r) is non-increasing,
this implies that for all x ∈ (r, 1], G0 (x)/G1 (x) ≤ 1. Combined with Eq. (B2), this
yields G0 (x)/G1 (x) = 1 for all x ∈ (r, 1], which yields a contradiction for x ∈ (1/2, β).


Nonmonotonicity of Social Beliefs

In this subsection, we illustrate the difficulties involved in determining equilibrium learn-
ing in general social networks. In particular, we show that social beliefs, as defined in

2
 2

4
1 8
5

Figure 3: The figure illustrates a deterministic topology in which the social beliefs are
nonmonotone.

Definition 3, may be nonmonotone, in the sense that additional observations of xn = 1

in the neighborhood of an individual may reduce the social belief (i.e., the posterior
derived from past observations that xn = 1 is the correct action).
The following example establishes this point. Suppose the private signals are such
that G0 (r) = 2r − r2 and G1 (r) = r2 , which is a pair of private belief distributions
(G0 , G1 ). Suppose the network topology is deterministic and for the first eight agents,
it has the following structure: B(1) = ∅, B(2) = ... = B(7) = {1} and B(8) = {1, ..., 7}
(see Figure 2).
For this social network, agent 1 has 3/4 probability of making a correct decision in
either state of the world. If agent 1 chooses the action that yields a higher payoff (i.e.,
the correct decision), then agents 2 to 7 each have 15/16 probability of choosing the
correct decision. However, if agent 1 fails to choose the correct decision, then agents 2
to 7 have a 7/16 probability of choosing the correct decision. Now suppose agents 1 to
4 choose action xn = 0, while agents 5 to 7 choose xn = 1. The probability of this event
happening in each state of the world is:
( )3 ( )3
3 15 1 10125
Pσ (x1 = ... = x4 = 0, x5 = x6 = x7 = 1|θ = 0) = = 26 ,
4 16 16 2
( )3 ( )3
1 9 7 250047
Pσ (x1 = ... = x4 = 0, x5 = x6 = x7 = 1|θ = 1) = = .
4 16 16 226
Using Bayes’ Rule, the social belief of agent 8 is given by
[ ]−1
10125
1+ w 0.961.
250047

3
 Now, consider a change in x1 from 0 to 1, while keeping all decisions as they are.
Then,
( )3 ( )3
1 7 9 250047
Pσ (x1 = 1, x2 = x3 = x4 = 0, x5 = x6 = x7 = 1|θ = 0) = = ,
4 16 16 226
( )3 ( )3
1 1 16 10125
Pσ (x1 = 1, x2 = x3 = x4 = 0, x5 = x6 = x7 = 1|θ = 1) = = 26 .
4 16 16 2
This leads to a social belief of agent 8 given by
[ ]−1
250047
1+ w 0.039.
10125
Therefore, this example has established that when x1 changes from 0 to 1, agent 8’s
social belief declines from 0.961 to 0.039. That is, while the agent strongly believes the
state is 1 when x1 = 0, he equally strongly believes the state is 0 when x1 = 1. This
happens because when half of the agents in {2, . . . , 7} choose action 0 and the other half
choose action 1, agent n places a high probability to the event that x1 ̸= θ. This leads
to a nonmonotonicity in social beliefs.

Proof of Lemma 2. Let h : {(n, B(n)) : n ∈ N, B(n) ⊆ {1, 2, ..., n − 1}} → N

be an arbitrary function that maps an agent and a neighborhood of the agent into an
element of the neighborhood, i.e., h(n, B(n)) ∈ B(n). In view of the characterization
of the equilibrium decision xn [cf. Eq. (2)], it follows that for any private signal sn ,
neighborhood B(n) ⊆ {1, 2, ..., n − 1}, and decisions xk , k ∈ B(n), we have

Pσ (xn = θ | sn , B(n), xk , k ∈ B(n)) ≥ Pσ (xh(n,B(n)) = θ | sn , B(n), xk , k ∈ B(n)).

By integrating over all possible private signals and decisions of agents in the neighbor-
hood, we obtain that for any n and any B(n) = B,

Pσ (xn = θ | B(n) = B) ≥ Pσ (xh(n,B(n)) = θ | B(n) = B) = Pσ (xh(n,B) = θ),

where the equality follows by the assumption that each neighborhood is generated inde-
pendently from all other neighborhoods. By taking the maximum over all functions h,
we obtain
Pσ (xn = θ | B(n) = B) ≥ max Pσ (xb = θ),
b∈B

showing the desired relation. 

Proof of Theorem 3
Proof of part (a): The proof consists of two steps. We first show that the lower and
upper supports of the social belief qn = Pσ (θ = 1|x1 , ..., xn−1 ) are bounded away from 0
and 1. We next show that this implies that xn does not converge to θ in probability.
Let xn−1 = (x1 , ..., xn−1 ) denote the sequence of decisions up to and including n − 1.
Let φσ,xn−1 (qn , xn ) represent the social belief qn+1 given the social belief qn and the

4
decision xn , for a given strategy σ and decisions xn−1 . We use Bayes’ Rule to determine
the dynamics of the social belief. For any xn−1 compatible with qn , and xn = x with
x ∈ {0, 1}, we have
φσ,xn−1 (qn , x) = Pσ (θ = 1 | xn = x, qn , xn−1 )
[ ]−1
Pσ (xn = x, qn , xn−1 , θ = 0)
= 1+
Pσ (xn = x, qn , xn−1 , θ = 1)
[ ]−1
Pσ (qn , xn−1 | θ = 0) Pσ (xn = x | qn , xn−1 , θ = 0)
= 1+
Pσ (qn , xn−1 | θ = 1) Pσ (xn = x | qn , xn−1 , θ = 1)
[ ( ) ]−1
1 Pσ (xn = x | qn , xn−1 , θ = 0)
= 1+ −1 . (B3)
qn Pσ (xn = x | qn , xn−1 , θ = 1)
Let ασ,xn−1 denote the probability that agent n chooses x = 0 in equilibrium σ when he
observes history xn−1 and is indifferent between the two actions. Let
G−
j (r) = lim Gj (s),
s↑r

for any r ∈ [0, 1] and any j ∈ {0, 1}. Then, for any j ∈ {0, 1},
Pσ (xn = 0 | qn , xn−1 , θ = j) = Pσ (pn < 1 − qn | qn , θ = j)
+ ασ,xn−1 Pσ (pn = 1 − qn | qn , θ = j)
[ ]
= Gj (1 − qn ) + ασ,xn−1 Gj (1 − qn ) − G−
−
j (1 − qn ) .

From Lemma 1(a), dG0 /dG1 (r) = (1 − r)/r for all r ∈ [0, 1]. Therefore,
1−r G0 (r) G−
0 (r)
1−β
≤ , and − ≤ .
r G1 (r) G1 (r) β
Hence, for any ασ,xn−1 ,
[ ]
Pσ (xn = 0 | qn , xn−1 , θ = 0) G− −
0 (1 − qn ) + ασ,xn−1 G0 (1 − qn ) − G0 (1 − qn )
= [ ]
Pσ (xn = 0 | qn , xn−1 , θ = 1) G− −
1 (1 − qn ) + ασ,xn−1 G1 (1 − qn ) − G0 (1 − qn )
[ ]
qn 1−β
∈ , .
1 − qn β
Combining this with Eq. (B3), we obtain
[( ( )( ))−1 ( ( )( ))−1 ]
1 1−β 1 qn
φσ,xn−1 (qn , 0) ∈ 1+ −1 , 1+ −1
qn β qn 1 − qn
[ ]
βqn 1
= ,
1 − β − qn + 2βqn 2
βqn
Note that 1−β−qn +2βqn
is an increasing function of qn and if qn ∈ [1 − β, 1 − β], then this
function is minimized at 1 − β. This implies that
[ ] [ ]
β(1 − β) 1 1
φσ,xn−1 (qn , 0) ∈ , = ∆, ,
−β + β + 2β(1 − β) 2 2

5
where ∆ is a constant strictly greater than 0. An analogous argument for xn = 1
establishes that there exists some ∆ < 1 such that if qn ∈ [1 − β, 1 − β], then
[ ]
1
φσ,xn−1 (qn , 1) ∈ , ∆ .
2
We next show that qn ∈ [∆, ∆] for all n. Suppose this is not true. Let N be the first
agent such that
qN ∈ [0, ∆) ∪ (∆, 1] (B4)
in some equilibrium and some realized history. Then, qN −1 ∈ [0, 1 − β) ∪ (1 − β, 1]
because otherwise, the dynamics of qn implies a violation of Eq. (B4) for any xN −1 . But
note that if qN −1 < 1 − β, then by Proposition 2 agent N − 1 chooses action xN −1 = 0
and, thus by Eq. (B3),
[ ( ) ]−1
1 Pσ (xN −1 = 0 | qN −1 , xN −2 , θ = 0)
qN = 1 + −1
qN −1 Pσ (xN −1 = 0 | qN −1 , xN −2 , θ = 1)
[ ( ) ]−1
1 1
= 1+ −1
qN −1 1
= qN −1 .
By the same argument, if qN −1 > 1 − β, we have that qN = qN −1 . Therefore, qN = qN −1 ,
which contradicts the fact that N is the first agent that satisfies Eq. (B4).
We next show that qn ∈ [∆, ∆] for all n implies that xn does not converge in proba-
bility to θ. Let yk denote a realization of xk . Then, for any n and any sequence of yk ’s,
we have
Pσ (θ = 1, xk = yk for all k ≤ n) ≤ ∆Pσ (xk = yk for all k ≤ n),
Pσ (θ = 0, xk = yk for all k ≤ n) ≤ (1 − ∆)Pσ (xk = yk for all k ≤ n).
By summing the preceding relations over all yk for k < n, we obtain
Pσ (θ = 1, xn = 1) ≤ ∆Pσ (xn = 1) and Pσ (θ = 0, xn = 0) ≤ (1 − ∆)Pσ (xn = 0).
Therefore, for any n, we have
Pσ (xn = θ) ≤ ∆Pσ (xn = 1) + (1 − ∆)Pσ (xn = 0) ≤ max{∆, 1 − ∆} < 1,
which completes the proof. 

Proof of part (b): The first step is the following lemma.

Lemma 11 Let B(n) = {b} for some n. We define
{ }
β 3 1
f (β, β) = max 1 − , − , (B5)
2(1 − β) 2 2β

where β and β are the lower and upper supports of the private beliefs (cf. Definition 4).
Let σ be an equilibrium. Assume that Pσ (xb = θ) ≤ f (β, β). Then, we have

Pσ (xn = θ | B(n) = {b}) ≤ f (β, β).

6
Proof. We first assume that Pσ (xb = θ) = f (β, β) and show that this implies

Ubσ ≥ β, and Lσb ≤ β, (B6)

where Ubσ and Lσb are defined in Eq. (A14). We can rewrite Ubσ as
Nbσ Nbσ
Ubσ = = .
1 − 2Pσ (xb = θ) + 2Nbσ 1 − 2f (β, β) + 2Nbσ

This is a decreasing function of Nbσ and, therefore,

1
Ubσ ≥ .
1 − 2f (β, β) + 2

Using f (β, β) ≥ 32 − 2β
1
, the preceding relation implies Ubσ ≥ β. An analogous argument
shows that Lσb ≤ β.
Since the support of the private beliefs is [β, β], using Lemma 3 and Eq. (B6), there
exists an equilibrium σ ′ = (σn′ , σ−n ) such that xn = xb with probability one (with respect
to measure Pσ′ ). Since this gives an expected payoff Pσ′ (xn = θ | B(n) = b) = Pσ (xb = θ),
it follows that, Pσ (xn = θ | B(n) = b) = Pσ (xb = θ). This establishes the claim that

if Pσ (xb = θ) = f (β, β), then Pσ (xn = θ | B(n) = {b}) = f (β, β). (B7)

We next assume that Pσ (xb = θ) < f (β, β). To arrive at a contradiction, suppose
that
Pσ (xn = θ | B(n) = {b}) > f (β, β). (B8)
Now consider a hypothetical situation where agent n observes a private signal generated
with conditional probabilities (F0 , F1 ) and a coarser version of the observation xb , i.e.,
the random variable x̃b distributed according to
[ ] [ ]
1 − f (β, β) 1 − f (β, β)
P(x̃b = 1 | θ = 1) = 1−Ybσ and P(x̃b = 0 | θ = 0) = 1−Nbσ .
Pσ (xb = θ) Pσ (xb = θ)

It follows from the preceding conditional probabilities that P(x̃b = θ) = f (β, β). We
assume that agent n uses the equilibrium strategy σn . Using a similar argument as in
the proof of Eq. (B7), this implies that

P(xn = θ | B(n) = {b}) = f (β, β). (B9)

Let z be a binary random variable with values {0, 1} and is generated independent of θ
with probabilities
2Ybσ 2Nbσ
P(z = 1) = 1 − and P(z = 0) = 1 − .
Pσ (xb = θ) Pσ (xb = θ)
This implies
[ that P(z = j ]| θ = j) = P(z = j) for j ∈ {0, 1}. Using x̃b with probability
b −1)Pσ (xb =θ)
(Y σ
1
1+f (β,β)
2 + Yσ
and z otherwise generates the original observation (random
b

7
variable) xb . Therefore, from Eq. (B8), P(xn = θ|B(n) = {b}) > f (β, β), which contra-
dicts Eq. (B9), and completes the proof.
Let f be defined in Eq. (B5). We show by induction that

Pσ (xn = θ) ≤ f (β, β) for all n. (B10)

Suppose that for all agents up to n − 1 the preceding inequality holds. Since |B(n)| ≤ 1,
we have

Pσ (xn = θ) = Pσ (xn = θ | B(n) = ∅)Qn (B(n) = ∅) (B11)

∑
n−1
+ Pσ (xn = θ | B(n) = b)Qn (B(n) = {b})
b=1
∑
n−1
≤ Pσ (xn = θ | B(n) = ∅)Qn (B(n) = ∅) + f (β, β)Qn (B(n) = {b}),
b=1

where the inequality follows from the induction hypothesis and Lemma 11. Note that
having B(n) = ∅ is equivalent to observing a decision b such that Pσ (xb = θ) = 1/2.
Since 1/2 ≤ f (β, β), Lemma 11 implies that Pσ (xn = θ|B(n) = ∅) ≤ f (β, β). Combined
with Eq. (B11), this completes the induction.
Since the private beliefs are bounded, i.e., β > 0 and β < 1, we have f (β, β) < 1 [cf.
Eq. (B5)]. Combined with Eq. (B10), this establishes that lim inf n→∞ Pσ (xn = θ) < 1,
showing that asymptotic learning does not occur at any equilibrium σ. 
Proof of part (c): We start with the following lemma, which will be used subse-
quently in the proof.
Proof.
Lemma 12 Assume that asymptotic learning occurs in some equilibrium σ, i.e., we
have limn→∞ Pσ (xn = θ) = 1. For some constant K, let D be the set of all subsets of
{1, ..., K}. Then,
lim min Pσ (xn = θ | xk = 1, k ∈ D) = 1.
n→∞ D∈D

Proof. First note that since the event xk = 1 for all k ≤ K is the intersection of events
xk = 1 for each k ≤ K,

min Pσ (xk = 1, k ∈ D) = Pσ (xk = 1, k ≤ K).

D∈D

Let ∆ = Pσ (xk = 1, k ≤ K). Fix some D̃ ∈ D. Then,

Pσ (xk = 1, k ∈ D̃) ≥ ∆ > 0,

where the second inequality follows from the fact that there is a positive probability of
the first K agents choosing xk = 1. Let A = {0, 1}|D̃| , i.e., A is the set of all possible
actions for the set of agents D̃. Then,
∑
Pσ (xn = θ) = Pσ (xn = θ | xk = ak , k ∈ D̃)Pσ (xk = ak , k ∈ D̃).
ak ∈A

8
Since Pσ (xn = θ) converges to 1 and all elements in the set Pσ (xk = 1, k ∈ D̃)}D̃∈A are
greater than or equal to ∆ > 0, it follows that the sequence {Pσ (xn = θ | xk = 1, k ∈ D̃)
also converges to 1. Hence, for each ϵ > 0, there exists some Nϵ (D̃) such that for all
n ≥ Nϵ (D̃),
Pσ (xn = θ | xk = 1, k ∈ D̃) ≥ 1 − ϵ.
Therefore, for any ϵ > 0,

min Pσ (xn = θ | xk = 1, k ∈ D) ≥ 1 − ϵ for all n ≥ max Nϵ (D),

D∈D D∈D

thus completing the proof.

We prove the claim in this part by contradiction. To arrive at a contradiction, we
assume that in some equilibrium σ ∈ Σ∗ , limn→∞ Pσ (xn = θ) = 1. The key part of the
proof is to show that this implies

lim Pσ (θ = 1 | xk = 1, k ∈ B(n)) = 1. (B12)

n→∞

To prove this claim, we show that for any ϵ > 0, there exists some K̃(ϵ) such that for
any neighborhood B with |B| ≤ M and maxb∈B b ≥ K̃(ϵ) we have

Pσ (θ = 1 | xk = 1, k ∈ B) ≥ 1 − ϵ. (B13)

In view of the assumption that maxb∈B(n) b converges to infinity with probability 1, this
implies the desired claim (B12).
For a fixed ϵ > 0, we define K̃(ϵ) as follows: We recursively construct M thresholds
K0 < ... < KM −1 and let K̃(ϵ) = KM −1 . We consider an arbitrary neighborhood B with
|B| ≤ M and maxb∈B b ≥ KM −1 , and for each d ∈ {0, ..., M − 1}, define the sets

Bd = {b ∈ B : b ≥ Kd } and Cd = {b ∈ B : b < Kd−1 },

where C0 = ∅. With this construction, it follows that there exists at least one d ∈
{0, ..., M − 1} such that B = Bd ∪ Cd , in which case we say B is of type d. We show
below that for any B of type d, we have

Pσ (θ = 1 | xk = 1, k ∈ Bd ∪ Cd ) ≥ 1 − ϵ, (B14)

which implies the relation in (B13).

We first define K0 and show that for any B of type 0, relation (B14) holds. Since
limn→∞ Pσ (xn = θ) = 1 by assumption, there exists some N0 such that for all n ≥ N0 ,
ϵ
Pσ (xn = θ) ≥ 1 − .
2M
Let K0 = N0 . Let B be a neighborhood of type 0, implying that B = B0 and all
elements b ∈ B0 satisfy b ≥ K0 . By using a union bound, the preceding inequality
implies ∑ ϵ
Pσ (xk = θ, k ∈ B0 ) ≥ 1 − Pσ (xk ̸= θ) ≥ 1 − .
k∈B
2
0

9
Hence, we have
1 1 ϵ
Pσ (xk = θ, k ∈ B0 | θ = 1) + Pσ (xk = θ, k ∈ B0 | θ = 0) ≥ 1 − ,
2 2 2
and for any j ∈ {0, 1},
Pσ (xk = θ, k ∈ B0 | θ = j) ≥ 1 − ϵ. (B15)
Therefore, for any such B0 ,
[ ]−1
Pσ (xk = 1, k ∈ B0 | θ = 0)Pσ (θ = 0)
Pσ (θ = 1 | xk = 1, k ∈ B0 ) = 1 +
Pσ (xk = 1, k ∈ B0 | θ = 1)Pσ (θ = 1)
[ ]−1
ϵ
≥ 1+ = 1 − ϵ,
1−ϵ
showing that relation (B14) holds for any B of type 0.
We proceed recursively, i.e., given Kd−1 we define Kd and show that relation (B14)
holds for any neighborhood B of type d. Lemma 12 implies that
lim min Pσ (xn = θ | xk = 1, k ∈ D) = 1.
n→∞ D⊆{1,...,Kd−1 −1}

Therefore, for any δ > 0, there exists some Kd such that for all n ≥ Kd ,
min Pσ (xn = θ | xk = 1, k ∈ D) ≥ 1 − δϵ.
D⊆{1,...,Kd−1 −1}

From the equation above and definition of Cd it follows that for any Cd ,
Pσ (xn = θ | xk = 1, k ∈ Cd ) ≥ 1 − δϵ.
By a union bound,
∑
Pσ (xk = θ, k ∈ Bd | xk = 1, k ∈ Cd ) ≥ 1 − Pσ (xk ̸= θ | xk = 1, k ∈ Cd )
k∈Bd
≥ 1 − (M − d)δϵ.
Repeating the argument from Eq. (B15), for any j ∈ {0, 1},
(M − d)δϵ
Pσ (xk = θ, k ∈ Bd | θ = j, xk = 1, k ∈ Cd ) ≥ 1 − .
Pσ (θ = j | xk = 1, k ∈ Cd )
Hence, for any such Bd ,
Pσ (θ = 1 | xk = 1, k ∈ Bd ∪ Cd )
[ ]−1
Pσ (xk = 1, k ∈ Bd | θ = 0, xk = 1, k ∈ Cd )Pσ (θ = 0, xk = 1, k ∈ Cd )
= 1+
Pσ (xk = 1, k ∈ Bd | θ = 1, xk = 1, k ∈ Cd )Pσ (θ = 1, xk = 1, k ∈ Cd )
 −1
(M −d)δϵ
P (θ=0 | xk =1, k∈Cd )
P σ (θ = 0, x k = 1, k ∈ Cd )
≥ 1 + ( σ ) 
(M −d)δϵ
1 − Pσ (θ=1 | xk =1, k∈Cd ) Pσ (θ = 1, xk = 1, k ∈ Cd )
(M − d)δϵ
= 1− .
Pσ (θ = 1 | xk = 1, k ∈ Cd )

10
Choosing ( )
1
δ= min Pσ (θ = 1 | xk = 1, k ∈ D),
M − d D⊆{1,...,Kd−1 −1}
we obtain that for any neighborhood B of type d
Pσ (θ = 1 | xk = 1, k ∈ Bd ∪ Cd ) ≥ 1 − ϵ.
This proves that Eq. (B14) holds for any neighborhood B of type d, and completing the
proof of Eq. (B13) and therefore of Eq. (B12).
Since the private beliefs are bounded, we have β > 0. By Eq. (B12), there exists
some N such that
β
Pσ (θ = 1|xk = 1, k ∈ B(n)) ≥ 1 − for all n ≥ N .
2
Suppose the first N agents choose 1, i.e., xk = 1 for all k ≤ N , which is an event with
positive probability for any state of the world θ. We now prove inductively that this
event implies that xn = 1 for all n ∈ N. Suppose it holds for some n ≥ N . Then, by
Eq. (B13),
β
Pσ (θ = 1 | sn+1 ) + Pσ (θ = 1 | xk = 1, ∈ B(n + 1)) ≥ β + 1 − > 1.
2
By Proposition 2, this implies that xn+1 = 1. Hence, we conclude there is a positive
probability xn = 1 for all n ∈ N in any state of the world, contradicting limn→∞ Pσ (xn =
θ) = 1, and completing the proof. 

Proof of Corollary 3
For M = 1, the result follows from Theorem 3 part (b). For M ≥ 2, we show that, under
the assumption on the network topology, maxb∈B(n) b goes to infinity with probability
one. To arrive at a contradiction, suppose this is not true. Then, there exists some
K ∈ N and scalar ϵ > 0 such that
( )
Q max b ≤ K for infinitely many n ≥ ϵ.
b∈B(n)

By the Borel-Cantelli Lemma (see, e.g., Breiman, Lemma 3.14, p. 41), this implies that
∑∞ ( )
Qn max b ≤ K = ∞.
b∈B(n)
n=1
Since the samples are all uniformly drawn and independent, for all n ≥ 2,
( ) ( )M
min{K, n − 1}
Qn max b ≤ K = .
b∈B(n) n−1
Therefore,
∑∞ ( ) ∑∞ ( )M ∑ ∞ ( )M
min{K, n − 1} K
Qn max b ≤ K = 1 + ≤1+ < ∞,
n=1
b∈B(n)
n=1
n n=1
n
where the last inequality holds since M ≥ 2. Hence, we obtain a contradiction. The
result follows by using Theorem 3 part (c). 

11