Network Games Under Incomplete Information on

Graph Realizations

Promit K. Chaudhuri∗ Sudipta Sarangi† Hector Tzavellas‡

December 2022

Abstract

We consider an incomplete information game in which agents have beliefs about an

ex ante distribution over all unweighted and undirected networks on a given vertex set.
Agents have private information which restricted only to the identity of their neigh-
bors. Given these beliefs, agents play a linear-quadratic effort game a la Ballester et
al. (2006) to maximize interim payoffs. We establish the existence and uniqueness of
Bayesian-Nash equilibria in pure strategies and characterize agent behavior. We show
that when agents are endowed with beliefs about network topology, local connectivity
provides information about indirect connectivity which is used for equilibrium play. In
particular, equilibrium actions are computable by the expected number of walks indi-
viduals have in the network.

JEL Classifications: C72, D81, D85

Keywords: Incomplete information, Network Games, Centrality, Local comple-
mentarities

∗
Department of Economics, Virginia Tech, Blacksburg, USA, Email: [email protected]
†
Department of Economics, Virginia Tech, Blacksburg, USA, Email: [email protected]
‡
Department of Economics, Virginia Tech, Blacksburg, USA, Email: [email protected]

1
1 Introduction

The majority of the work on network games with local complementarities assumes that
agents have complete knowledge of the network architecture they are embedded in. This
assumption is crucial for such games, as equilibrium behavior of agents is driven by metrics
defined on the architecture of the entire network (Ballester et al. (2006)). In most real
world social and economic networks, however, agents rarely have access to this information.
In fact, as demostrated by Breza, Chandrasekhar, and Salehi (2018), agents are mostly
aware of the identity of their immediate neighbors. Despite this, as well as the fact that
network interactions are important drivers of behavior, there has been little work on trying
to understand how this particular type of information drives equilibrium outcomes.

Within the confines of a game of incomplete information, the first question that arises is
the following: as to what object is the information regarding agents neighbors informative
of? In other words, if we presume that agent behavior within a network system is the
result of equilibrium behavior in a game of incomplete information, over what object do
agents have ex-ante beliefs? Given that agents can identify the number of agents they are
connected to, the prevailing approach has been to assume that these ex-ante beliefs are
representative of degree distributions. That is, existing models endow agents with beliefs
about the number of connections that each individual may have in the network. An important
network characteristic that this approach abstracts away from, however, is agent identity.

In this paper, we are motivated by the subset of complete information network games in which
common knowledge of the entire network is assumed, and where the identity of individuals
play a role in equilibrium behavior. We, therefore, take a different approach and assume the
aforementioned ex-ante beliefs are representative of a distribution over networks themselves.
In particular, we study an incomplete information variant of the popular quadratic game of
Ballester et al. (2006) where ex-ante beliefs are prescribed by a probability mass function
over the set of simple graphs on n vertices, and in which agent’s information is restricted to
the identity of their neighbors only.

Our game proceeds as follows. Nature moves first and chooses an unweighted and undirected
network on n vertices from an ex-ante distribution over all such graphs. Agent i’s type (in the
Harsanyi (1967) sense) corresponds to the ith row of the adjacency representation of nature’s
chosen network. Agents are thus classified by the identity of their neighbors and are hence
able to identify the agents from whom they will directly extract network complememtarities
from. However, they are unaware of the types of these adjacent agents. Given their realized

2
type, and using Bayes rule, agents update their beliefs regarding these types and, therefore,
their beliefs about the true topology of the network. They then proceed to simultaneously
exert actions to maximize interim linear quadratic payoffs.

We establish existence and uniqueness of pure strategy Bayesian Nash equilibria (BNE)
for arbitrary ex-ante distributions. These properties hold for a bound on the modularity
parameter of cross activity that is identical to the complete information variant of the model.

Turning to the properties of the BNE, we show that agents can over or under exert actions
when compared to the complete information Nash equilibrium induced over nature’s chosen
network. The extent to which they do so depends on the number of agents they are con-
nected to (i.e. their degree), as well as the ex-ante distribution of networks. We show that
in equilibrium, agents will use the information regarding their direct connections to make
inferences about the complementarity strength of their actions with those of other agents.
The streghth of this complementarity is computed by their ex-post expectation regarding the
number of walks they have in the network. In this sense, the BNE results from a calculation
similar to the one performed by agents in the complete information case, where the Nash
equilibrium is proportional to the actual number of walks that agents have in the network
(Balletser et al. (2006)) i.e. KB centrality.

It is important to note, however, that this expected sum of walks does not correspond to
the agents ex-ante expected Katz-Bonacich (KB) centrality. This point raises an interesting
question with regard to how network uncertainty should be gauged from an applied perspec-
tive. Motivated by results on complete information games and centrality driven outcomes,
simulation based, and empirical approaches have been employed to deal with situations in
which the network is not observed. The empirical approach has proposed estimators of net-
work effects in environments in which researchers cannot observe the network (De Paula,
Rasul, and Souza (2018) and Lewbel, Qu, and Tang (2021)). In the simulation based ap-
proaches, researchers will often employ random network models in an attempt to generate
the distribution of the centrality metric that drives behavior in the network system of inter-
est (Crucitti et al. (2006), Latora and Marchiori (2007)). There have also been analytical
attempts towards this goal. Dasaratha (2020), for instance, characterizes the distributions
of KB and eigenvalue centrality on large networks.

Our result, however, suggests that whenever network participants lack information about
the true architecture of the network, their behavior is not consistent with complete infor-
mation behavior, nor with expectations over complete information equilibrium outcomes.
The caveat with such applied approaches is, therefore, that they do not internalize the fact

3
that the subjects themselves may not know the network. When this is the case, centrality
distributions, or estimators of complete information network effects, may not be informative
as to the true behavior of the network system of interest.

Next, we restrict attention to the role of degrees and impose the ex-ante distribution to be
uniform. In this case, these expected walks, and consequently BNE actions, are directly
proportional to agents realized degrees. In other words, the higher an agent’s number of
connections, the higher is their exerted action is in equilibrium. This result is closely related
to the finding of Galeotti et al. (2010). They show that in an environment in which agents
are endowed with information about their degree, but not the identity of the agents they are
connected to, equilibrium actions are also driven by degrees. They do this by endowing agents
with ex-ante beliefs on the degree distribution of the true networks. Degree distributions,
however, are anonymous implying that agents do not employ local information towards
equilibrium play. In contrast, we show that by endowing agents with beliefs about network
topology itself, agents will internalize local information in order to form beliefs about the
number of walks they have in the network.

Other than the large literature on linear-quadratic network games of complete information,
there have been two other attempts that introduce incomplete information into the model.
De Marti and Zenou (2015) study a linear quadratic game of incomplete information in which
agents lack information regarding model parameters other than the network itself. These
include the link complementarity strength, and the return to own action exertion. Unlike
their work, we are interested in incomplete information on the network.

Closer to our model is Breza, Chandrasekhar, and Salehi (2018) who also employ a linear
quadratic game in which agents lack complete information regarding the network itself.
One of their crucial assumptions, however, is that the information set of any agent (i.e. the
identity of their neighbors) doesn’t provide any information about their indirect connections.
In other words, their expectations regarding the existence of links between their neighbors
and other agents is independent of the information they are endowed with. As a result, the
equilibrium gets mapped to their ex-ante beliefs about the network. In contrast, we find
that as long as agent are endowed beliefs about network topology itself, local connectivity
provides information regarding indirect connectivity and agents will make use of it towards
equilibrium play. This local information being different for each player in turn, implies that
the equilibrium is no longer mapped to their ex-ante beliefs about the network.

The rest of the paper is structured as follows. Section 2 contains tool from network theory
that will be throughout the paper and sets up the game. In section 3, we derive the Bayesian

4
Nash equilibrium and study its properties. Section 4 discusses welfare. Section 5 concludes.
All proofs are relegated to the the appendix.

2 Model

2.1 Preliminaries

Let N = {1, 2, ..., n} denote the set of players. Letting i ∼ j denote a link between players
i and j, a network (graph) g is the collection of all pairwise links that exist between the
players. The links are undirected such that i ∼ j ∈ g implies j ∼ i ∈ g. The network can be
represented by its adjacency matrix which, with some abuse of notation, is also denoted as
g = [gij ], where gij = 1 if a link exists between players i and j, and gij = 0 otherwise. There
are no self-loops and thus gii = 0 for all i ∈ N . The fact that links are undirected implies
g = gT . We denote by Gn the set of all unweighted and undirected networks on n vertices.
n(n−1)
We consider simple graphs, so the cardinality of Gn is 2 2 .

Given the adjacency representation of a network g ∈ Gn we let gi denote its ith row. That is,
gi = (gi1 , gi2 , ...., gin ) ∈ {0, 1}n , where it is understood that gii = 0. In the following section,
it will be convenient to represent any network g by the rows of its adjacency matrix:

g = (g1 , g2 , .., gn ) (1)

The neighborhood of player i is the set of players to whom i is linked and is denoted by:
N(gi ) = {j : gij = 1}. The size of this set is i’s degree which counts the agent’s direct
connections: d(gi ) ≡ |N(gi )|. A network is called k-regular if all players have the same
degree k (which is also the degree of the network); otherwise, the network is irregular.

A walk of length p from a node i to a node jp is a sequence of links in the network i ∼ j1 ,

j1 ∼ j2 ,..., jp−1 ∼ jp . It is denoted by ij1 j2 ..., jp . Given two nodes i and jp there may exist
more than one such walk. Using the adjacency representation, the number of walks of length
p from node i to node jp can be computed by the ijp element of the matrix gp .

Finally, let g0 = I, then for a sufficiently small λ > 0, the following influence matrix
M(g,λ) = [mij (g)] is well-defined and non-negative:
∞
X
M(g, λ) ≡ [I−λg]−1 = λs gs
s=0

5
Each element mij (g) measure the total number of walks of all lengths from agent i to agent
j. Given M(g,λ), the Katz-Bonacich (KB) centrality of player i, bi (g), is the ith -component
of the vector b (g) = M(g, λ)1. It measures the total number of discounted walks of all
lengths originating from player i to all other players in g where longer paths are discounted
more.

2.2 The Game

We study a variant of the simultaneous move local complementarities game of Ballester et

al. (2006), in which agents have incomplete information about the full architecture of the
network.

We follow Harsanyi’s (1967) approach to games of incomplete information by introducing

Nature as a non-strategic player who chooses a network out of set of all possible graphs on
the number of vertices equal to the number of agents. The network is chosen from an ex-ante
distribution that is common knowledge among all agents. Following Nature’s draw, players
realize their direct connections (they can see the agents with who they are linked with) but
do not know the network’s architecture beyond that. In other words they do not observe the
links of their neighbors. Using the information on their direct connections, agents proceed to
update their beliefs about the graph chosen by Nature according to Bayes’ rule. Given these
updates beliefs, agents will simultaneously exert actions to maximize their interim payoffs.

We proceed to describe the game formally.

Agents and Types

N is the set of players (nodes), with |N | = n. For each i ∈ N , we let Gi denote the
player’s type set. We want agent types to be representative of their corresponding row in the
adjacency representation of the network over which the game will be played. To this end,
each player’s types set is taken to be as follows:

Gi = {(gi1 , gi2 , ...., gin )i ∈ {0, 1}n : gii = 0}

where we will think of gij = 1 if player i is connected to j and 0 otherwise.1 The cardinality
of each agent’s type set is:
|Gi | ≡ γ = 2n−1
1
Note that the subscript in (gi1 , gi2 , ...., gin )i is imposed to differentiate between agents whose types
consist of the same sequence of 0’s and 1’s. For instance, if n = 3 it differentiates between the type (0, 1, 0)1
for agent 1 and (0, 1, 0)3 for agent 3.

6
and we denote its elements by giti ∈ Gi .2 Given each player’s type set, we can write down
the type space of the game:
G = "i∈N Gi

Observe that if we invoke network representation (1), an element g ∈ G may correspond to

an adjacency matrix of an undirected and unweighted network. However, not all elements of
G are valid representations. As an example, consider the case with 3 players, N = {1, 2, 3}.
The type set of each player is given by:

G1 = {(0, 0, 0)1 , (0, 1, 0)1 , (0, 0, 1)1 , (0, 1, 1)1 }

G2 = {(0, 0, 0)2 , (1, 0, 0)2 , (0, 0, 1)2 , (1, 0, 1)2 }
G3 = {(0, 0, 0)3 , (1, 0, 0)3 , (0, 1, 0)3 , (1, 1, 0)3 }

with corresponding type space G = G1 × G2 × G3 . One element of G is ((0, 1, 0)1 , (0, 0, 1)2 ,
(0, 1, 0)3 ), the vectors of which do not correspond to rows of the adjacency matrix of any
undirected and un-weighted network. This is because agent 1 is connected to agent 2 while
agent 2 is not connected to agent 1, therefore, the corresponding adjacency matrix is not
symmetric.

In our model, we want to restrict attention to elements of G that have valid network rep-
resentations so that Nature’s choice is reflective of a network. In what follows, we do so
through the information structure.

Ex-Ante Beliefs

We denote by, p ∈ ∆(G) the probability distribution over the type space, with ∆(G) denoting
set of all probability distributions on G. In our game, Nature moves first and chooses an
element of the type space g ∈ G. As noted above, we want to restrict Nature’s choice to
those elements in G that have valid network representations. Towards this, we define the
following set of admissible distributions and impose the assumption that Nature draws a
network according to a distribution in this set.

Definition 1. We say that the probability distribution p ∈ ∆(G) is admissible if it satisfies:

p(g) = 0 ∀ g ∈ G s.t g ̸= gT

and denote set of all admissible distributions by ∆A (G)

2
Whenever the context is clear and we need not enumerate the elements of each type set we suppress
the superscript.

7
Assumption 1 : p ∈ ∆A (G) and this is common knowledge.

Observe that the imposition of assumption 1 implies that p(g) > 0 if and only if g ∈ Gn .
Consequently, Nature will choose an unweighted and undirected network with certainty, and
the fact that the agents are part of one is common knowledge.

In the following section we will make use of the uniform admissible distribution which is
defined as follows:

Definition 2. The probability distribution p ∈ ∆A (G) is uniform if it satisfies:


1
 n(n−1) if g ∈ Gn
p(g) = 2 2
0 if g ̸= gT

In the 3-player case, for instance, we have that |G3 | = 8 and thus Nature will choose any
network with probability p(g) = 81 .

Belief Updating

Given a assumption 1, agents know that Nature draws a network and proceed to update
their beliefs regarding its true topology according to Bayes’ Rule. These ex-post updated
beliefs can be written as:
q q
P
p(gi , gj ) q p(G ).I{gi , gj ∈ G }
p(gj |gi ) = = P k k
∀ i, j ∈ N (2)
p(gi ) k p(G ).I{gi ∈ G }

where, Gq , Gk ∈ G. Intuitively, equation (2) states that agent i who is of type gi ∈ Gi will
assign a probability to agent j being of type gj ∈ Gj according to (i) the number of states in
the state space that contain both of these types, and (ii) their ex-ante probabilities. Given
assumption 1, since agent types correspond to rows of an adjacency matrix, the probability
the agent i (whose row is gi ) will assign to an agent having a row gj will depend on the
number of networks that themselves contain these rows and the probability that nature
chooses them.

As an example, consider the 3-player case and suppose that after Nature’s draw, agent 2
is of type (1, 0, 0)2 . In other words, agent 2 learns that it is connected to agent 1 but is
not connected to agent 3. Since players can only observe the neighbors, agent 2 does not
know if agents 1 and 3 are connected between themselves and will thus have to form beliefs
about the existence of a link between them. This is demonstrated in the left-most network

8
in figure 1. However, the state space only contains 2 elements in which agent 2’s type is
admissible with a valid network representation and these are ((0, 1, 0)1 , (1, 0, 0)2 , (0, 0, 0)3 )
and ((0, 1, 1)1 , (1, 0, 0)2 , (1, 0, 0)3 ). In other words, there are only two graphs on 3 vertices
that contain the link 1 ∼ 2 and do not contain the link 2 ∼ 3. If we we took the ex-ante
distribution to uniform, then agent 2 would assign a probability of 21 that nature chose either
of these.

Figure 1: Graphs of the types (1, 0, 0)2 , (1, 1, 0)3 , (0, 0, 1)1 and (0, 0, 0)2

Finally, note that assumption 1 implies implies that beliefs are consistent, in the sense that
agents will give zero probability to others being of types that do not match the adjacency
pattern induced by their own type. This is expressed formally in the proceeding remark.
Remark 1. For all gi ∈ Gi , p(gj |gi ) = 0 ∀gj ∈ Gj for which gij ̸= gji , ∀i, j ∈ N

State Game and Equilibrium

Given the above, conditional on a state g ∈ G being realized, agents play the state game:

sg = (N, A, (ui (ai , a−i ))i∈N )

where the action set is the same for each agent and is equal to the positive real numbers
A ≡ R+ . Interim utilities assume a linear-quadratic form:
n
1 X X
ui (ai (giti ); ai (gi−ti ), a−i ) = ai (giti ) − ai (giti )2 + λai (giti ) gijti p(gj |giti )aj (gj ) (3)
2 j=1 g ∈G
j j

where ai (gi−ti ) = (ai (gi1 ), ..., ai (giti −1 ), ai (giti +1 ), ..., ai (giγ )), a−i = (a1 , ..., ai−1 , ai+1 , ..., an )
and aj = (aj (gj1 ), ..., aj (gjγ )). As in Ballester et al. (2006) the first two term in the utility
specification capture the cost and direct benefit to agent i from exerting its own action. The
third term captures local complementarities with the agents that the player is connected
to, with λ measuring the strength of this complementarity. Unlike Ballester et al. (2006)

9
agents need to form beliefs about the actions of their neighbors. Agents simultaneously exert
actions to maximize (3). For each agent i, a pure strategy σi maps each possible type to an
action. That is,
σi = (ai (gi1 ), ..., ai (giγ ))

This is a simultaneous move game of incomplete information so we use the Bayesian-Nash

equilibrium notion.
∗
Definition 3. The pure strategy profile σ ∗ = (σi∗ , σ−i ) is a Bayesian-Nash equilibrium
(BNE) if:

a∗i (giti ) = arg max

t
ui (ai (giti ), a∗i (gi−ti ), σ−i
∗
) ∀ i ∈ N, ∀ giti ∈ Gi
ai (gi i )

The above game can be summarized according to the tuple:

Γ = ⟨N, (Gi )i∈N , p, (sg )g∈G ⟩ (4)

3 Bayesian Nash Equilibrium

We proceed to characterize the BNE of Γ starting with best responses.

3.1 Best Responses

Given the payoff structure, the best response of the ith player whose is of type giti is given
by:
n
X X
ai (giti ) = 1 + λ gijti p(gj |giti )aj (gj )
j=1 gj ∈Gj

The system of best responses can be written in vector notation as follows:

a = 1nγ + λBa (5)

γ
where 1nγ is the nγ-dimesnional column vector of 1’s, a = [ai ]ni=1 , ai = ai (giti ) ti =1 , γ = 2n−1


is the total number of types of each player, and B is a block matrix that assumes the following

10
form  
0 G1∼2 . . . G1∼n
 
 G2∼1 0 . . . G2∼n 
B=
 ...

 ... ... ...  
Gn∼1 Gn∼2 ... 0 nγ×nγ

with
t t
[Gi∼j ]ti tj = gijti p(gjj |giti ) ∀ tj , ti = 1, .., γ and ∀ gjj ∈ Gj , giti ∈ Gi

It can be verified that if the ex-ante distribution satisfies p(g) = 1 for some g ∈ Gn and
p(g′ ) = 0 for all g′ ̸=g, then ai (gi ) = 0 for all i ∈ N for which gi ∈ / g. In this case,
the system of best responses would reduce to the complete information Nash equilibrium
(Ballester et al. (2006)):
ac = 1n + λgac (6)

where ac = (a1 (g1 ), .., an (gn )) with g = (g1 , .., gn ). In other words, in the complete informa-
tion case, the matrix B would reduce to the actual network over which the game is played,
and agents would best respond to the actions of their adjacent agents. In the incomplete
information case, however, agents do not know the types of their neighbors and best respond
to the updated beliefs regarding their neighbors actions. This is captured by the elements
within the blocks of B. For instance, consider agent i and the block [Gi∼j ]ti tj . Its elements
t
are of the form gijti p(gjj |giti ), which state that if agent i whose type giti is such that it is
t
connected to agent j, it will assign the probability to this agent being of type gjj equal to
t
p(gjj |giti ). Observe that this updating takes place as long as agent i is connected to agent j,
implying that agents form beliefs about other as long as a link exists between them. In this
way, the matrix B may be interpreted in a similar way to the complete information case, but
instead of adjacency over agents, it provides the adjacency pattern over network admissible
types. In turn, this gives rise to a network between types themselves.

For example, suppose n = 3 and let the underlying distribution be uniform on G3 . This
implies that updated beliefs are given by p(gj |gi ) = 12 , ∀gj ∈ Gj . Setting λ = 0.1, numerical

11
computation gives following characterization:
     
a1 ((0, 0, 0)1 ) 0 0 0 0 0 0 0 0 0 0 0 0 1.
     
a1 ((0, 1, 0)1 )  0 0 0 0 0 1 0 1 0 0 0 0  1.11765
     
a ((0, 0, 1) )  0 0 0 0 0 0 0 0 0 1 0 1  1.11765
 1 1     
     
a1 ((0, 1, 1)1 )  0 0 0 0 0 1 0 1 0 1 0 1  1.23529
     
a ((0, 0, 0) )  0 0 0 0 0 0 0 0 0 0 0 0   1. 
 2 2     
     
a2 ((1, 0, 0)2 ) 1 0 1 0 1 0 0 0 0 0 0 0 0  ∗
1.11765
a=  B=   a = 
a ((0, 0, 1) )
 2 2  2
 0 0 0 0 0 0 0 0 0 0 1 1 

1.11765
 
     
a2 ((1, 0, 1)2 )  0 1 0 1 0 0 0 0 0 0 1 1  1.23529
     
a ((0, 0, 0) )  0 0 0 0 0 0 0 0 0 0 0 0   1. 
 3 3     
     
a3 ((1, 0, 0)3 )  0 0 1 1 0 0 0 0 0 0 0 0  1.11765
     
a ((0, 1, 0) )  0 0 0 0 0 0 1 1 0 0 0 0  1.11765
 3 3     
a3 ((1, 1, 0)3 ) 0 0 1 1 0 0 1 1 0 0 0 0 1.23529

If we think of B as an adjacency matrix whose entries are representative of links between

network admissible types, it may then be visualized by the figure 2.

Consider player 2 and suppose it has realized the type (1, 0, 0)2 (as visualized in the left-most
network of figure 1). The player knows that it connected to player 1, as g21 = 1, and that
it is not connected to 3, as g23 = 0. Therefore, agent 2 will form beliefs over agent 1’s
types. Since the only types of agent 1 that are network admissible with the type (1, 0, 0)2
are (0, 1, 1)1 and (0, 1, 0)1 , then there exists a link between the types (1, 0, 0)2 and (0, 1, 1)1
as well as (1, 0, 0)2 and (0, 1, 0)1 . A similar argument holds for all other agents and all of
their possible types. This in turn produces figure 2.

Before we proceed, we note that closest to our best response characterization is the interaction
structure considered in Golub and Morris (2020). Although the signal realizations of each
agent in their model can be thought of as arising from a more general information structure
(which could potentially allow for network signals themselves), the network realization itself
is nonetheless common knowledge. In their general theory of networks and information,
agent behavior is driven by an endowed interaction structure similar to our matrix B. In our
model, however, this is generated endogenously as a result of optimizing behavior.

12
 Figure 2: Network between types of each players

3.2 Existence-Uniqueness

According to definition 3, the BNE is characterized by the fixed point of the system of
equations in (5). We have the following classification:
1


Proposition 1. There exists an unique pure strategy BNE for all λ ∈ 0, n−1

Observe that the bound on the local complementarity parameter λ which guarantees the

13
existence and uniqueness of an equilibrium is identical to the complete information bound.3
Formally, this holds because the elements in each row of B sum to at most n − 1. This can be
seen from the fact that the non-zero rows of its blocks [Gi∼j ]kl sum to 1, as they correspond
to conditional probability distributions between network admissible types.

Intuitively, n − 1 represents the maximal number of agents that each individual can extract
direct complementarities from. Since the complementarity strength arising from a single link
is λ, the maximal direct complementarity that may be extracted by a single agent is λ(n−1).
Moreover, agents are embedded in a network, so they can also extract complementarities from
their indirect connections. In the complete information case, the maximal complementarity
that can be extracted by a single agent due to their pth order indirect connections is λp (n −
1)p .4 Therefore, summing over all p ∈ N+ gives the maximal complementarity that any agent
can extract from any network, which in turn produces a bound on the strength λ for actions
to be bounded.

In the incomplete information case, a similar argument holds, but the bound on the maximal
complementarity that may be extracted from the network is attained by decomposing it
across states rather than links. For example, consider an agent i who is of type gi , and
who is connected to agent j. Given updated beliefs, agent i assigns a probability p(gj |gi )
to agent j being of type gj . This in turn induces a complementarity strength of λp(gj |gi )
between the action of agent i and that of an agent j who is of the particular type gj . Since
P
gj ∈Gj p(gj |gi ) = 1, then the maximal complementarity that can be extracted from a single
neighbor is λ. A similar argument holds for indirect connections.

In other words, the complementarity an agent i extracts from another j, is spread out across
all of j ′ s types that are admissible with the realized type of agent i, In this sense, the
model generates network externalities on the agent-state specific level rather than the agent
specific level. This has important consequences for the properties of the BNE. We turn to
these properties next.

3
Note that a more general bound for the complete information case is λρ(g) < 1 where ρ(g) is the
1
spectral radius of the adjacency matrix g. The bound λ < n−1 is the tightest possible, since that maximal
spectral radius for any graph g ∈ Gn is n − 1 and this corresponds to the complete network. We use the
1
bound λ < n−1 so that comparisons between complete and incomplete information is possible throughout
the paper without having to consider heterogeneous λ for the complete and incomplete information cases.
4
Since any agent is connected to at most n − 1 others so that p links aways from any node are at most
(n − 1)p other nodes from which a complementarity strength of λp is extracted.

14
3.3 Walk Characterization

We now turn to the properties of the BNE. In what follows, we denote by a∗i (giti ) the equi-
librium action of agent i who has realized the type giti ∈ Gi , and by a∗ (g) the vector of
equilibrium actions given that nature have drawn g = (g1t1 , ..., gntn ) ∈ Gn . The following
proposition characterizes the BNE for any ex-ante distribution and any realized network.

Proposition 2. For any p ∈ N+ let j1 , j2 , .., jp denote an arbitrary collection of p indices. For
any admissible probability distribution, and for any realized network g ∈ G, the equilibrium
actions of agents are given by:
∞
X (p)
a∗i (giti ) = λp βi,ti ∀ i ∈ N, ∀ giti ∈ Gi
p=0

where giti = (gijti )j∈N is the realized type of agent i, and where:

n γ
(p)
X X t tj tj tj tj tj t
βi,ti = gijti1 gj1j1j2 ...gjp−1
p−1 p−1 p−2 p−2 p−3 j1 ti
jp p(gjp−1 |gjp−2 )p(gjp−2 |gjp−3 )...p(gj1 |gi )
j1 ,j2 ,..,jp =1 tj1 ,tj2 ,..,tjp−1 =1

Proposition 2 is best understood when compared to the complete information Nash equilib-
rium over the same network:
 
∞ n ∞
c ti
X
p
X
ti tj1 tjp−1 X (p)
ai (gi ) = λ gij1 gj1 j2 ...gjp−1 jp ≡
 λp di
p=0 j1 ,j2 ,..,jp =1 p=1

(p)
For each p ∈ N+ , di measures the total number of walks of length p originating from player
i to all others (including i itself). In the complete information scenario, each agent has
knowledge of the full architecture of the network and can thus compute these walks for all
lengths p. Intuitively, each of these walks ij1 j2 , .., jp , captures the complementarity of agent
i′ s action with that of agent jp due the existence of a particular sequence of intermediate
links i ∼ j1 , j1 ∼ j2 ,.., jp−1 ∼ jp connecting them. Thus, each agent will take into account
all of these complementarities and exert an action equal to their total strength. In turn this
produces a Nash equilibrium equal to the KB centrality vector.

In the incomplete information case, knowledge of these walks is limited to those that are of
first order, as agents can only identify their neighbors. Even though information is limited,

15
agents are, nonetheless, aware of the fact that they participate in a network, and hence, in-
ternalize the fact that walks of arbitrary orders may exist between them and all other agents.
Since these walks capture complementarity strengths, that in turn dictate the magnitude of
actions, agents will need to form expectations as to what their actual strength is. In the
(p)
statement of proposition 2, each term βi,ti captures this expected measure for all walks of
the particular order p.

To describe this expected measure in more detail, consider the case p = 3. With a slight
rearranging of terms we can write
n γ
t t t
(3)
X X tk
βi,ti = gijti gjkj gkl p(gktk |gjj )p(gjj |giti )
j,k,l=1 tj ,tk =1
n X
n γ γ
t t t
X X X tk
= gijti gjkj p(gjj |giti ) gkl p(gktk |gjj )
j,k=1 l=1 tj =1 tk =1

As per the timing of events in the game, agent i gets to know its type giti and hence has full
knowledge of the links gijti . The player is thus aware of the agents through which it can form
a walk of length three. To fix ideas, suppose that player i wants to form of a belief about
the complementarity strength of its action with that of agent l due to the particular walk
ijkl. As for the reasons mentioned earlier, agent i has complete information about gijti , i.e.
the link between it and agent j. However, it does not have complete information about j ′ s
type, nor about j ′ s neighbors’ neighbors’ type which in turn may or may not include a link
with agent k through which the walk of interest ijkl reaches agent l.

The expectation regarding the strength of this complementarity is formed in three steps.
First, the agent will condition on the fact that it has a link with agent j. This occurs with
probability p(giti |giti ) = 1 (since we are assuming that gijti = 1) and thus, we may think
of gijti as the expected number of ways that agent i can reach agent j. Second, the agent
t
internalizes its own type through p(gjj |giti ), to compute expectations over the links between
its neighbor j and its neighbors’ neighbor k. Using this, the agent counts the expected
number of ways it can reach k through j, given the link gijti exists i.e. given it is of type giti .
t t
This is given by γtj =1 gjkj p(gjj |giti ). Third, the agent internalizes the information about the
P
t
possible types of its neighbor j through p(gktk |gjj ), to compute the expectations over the links
between its neighbors’ neighbor k and agent l, (who is its neighbors’ neighbors’ neighbor).
Using this, the agent counts the expected number of possible ways it can reach l through k,
t t
conditional on the existence of the link gjkj i.e. given that j is of type gjj . This is given by
Pγ tk tk tj
tk =1 gkl p(gk |gj ).

16
 Figure 3: Expected walk of length 3

Given the above, the expected total number of ways player i can reach player l via a walk of
length three is given by the product of (i) the actual link between it and j, (ii) the number
of ways it can reach k from j given the previous link gijti exists and (iii) the number of ways
t
it can reach l from k given gjkj exists. Repeating this process for all possible walks of length
three which start from agent i, and summing over all possible values of j, k and l, gives the
expected complementarity strength of agent i′ s action with all other agents due to walks of
(3)
length three βi,ti .

There are a couple of remarks that we make with regard to the nature  of the preceding
(p) (p)
expected complementarity calculation. First, we note that, βi,ti ̸= E di . That is, the
expected complementarity arising from walks of length p does not equal to the ex-ante expec-
tation of these walks. This to be expected as the equilibrium is an interim notion, allowing for
belief-updating. An important consequence of this, is that the BNE equilibrium of this game
does not equal the ex-ante expectation of KB centrality. Motivated by complete information
equilibrium notions, applied work has tried to estimate network effects in environments in
which researchers cannot observe the network. These approaches implicitly presume that
although the researcher does not have information about the network, the agents themselves
have this information. The proposed estimators are reflective of this, as they correspond
to ex-ante expectations of complete information outcomes. As demonstrated by Breza et
al. (2018), however, the assumption that a researcher is unaware of the network while the
subjects are aware of it, may in some cases be inconsistent. If so, and as Proposition 2

17
suggests, agent behavior under such information settings would not correspond to ex-ante
expectations over complete information outcomes.
 
(p) (p)
Next, we also note that, βi,ti ̸= E di | giti . To shed more light into this, consider the case
of p = 3.

n X
n
!
  X
(3)
E di | giti = E gij gjk gkl | giti
j,k=1 l=1
n X
X n
E gij gjk gkl | giti


=
j,k=1 l=1
n X n Xγ γ  
t tk t tk
X X
= gijti gjkj gkl p gjkj gkl | giti
j,k=1 l=1 tj =1 tk =1

This hypothetical ex-post expectation calculation, fails to capture the dynamics by which the
agent i internalizes the possible types of its neighbors, its neighbors’ neighbors and so on. In
other words, only conditioning on their own type giti , makes it a more restricted measure. On
(p)
the other hand, βi,ti gives us the process by which agent i internalizes the possible types of
its neighbors, its neighbors’ neighbors and so on. The interim pay-off structure being linear
quadratic, implies that the payoff to agent i is dependent on their neighbors’ actions. As a
result, the equilibrium action strategy given by ∞ p (p)
P
p=0 λ βi,ti , enables us to understand the
dynamics of how the agents internalizes the information about the possible types of others,
and how they are employed towards equilibrium play.

3.4 The Uniform Case

Before we proceed to welfare, we discuss the special case in which the ex-ante distribution
is uniform over the set of all networks. We have the following characterization.

Proposition 3. When the underlying probability distribution is uniform, the equilibrium

actions of the agents for any realized graph are given by:

λd
a∗ (d) = 1 + , ∀d ∈ {0, 1, 2, . . . , (n − 1)}
1 − nλ
2

where d is the degree of an agent in the realized network.

Proposition 3 is similar to the finding of Galleotti et al. (2010) (Proposition 2). They study a
reduced form game of incomplete information in which agents are endowed with beliefs about

18
degree distributions, establishing the existence of a BNE in which the actions of agents are
monotonically increasing in their degrees. The result is as consequence of two assumptions.
The first is anonymity, which states that while agents are aware of the number of agents
they are connected to, they are unaware of the identity of these adjacent agents. The second
is independence, which states that the degree of any agent who is connected to another is
independent of the latter’s degree.

While our result provides the same quantitative insight as their finding, it holds due to a
different belief structure. First, we do not assume anonymity. As seen from Proposition
2, our BNE is the result of an expected walk calculation. These walks are computed for
all possible sequences of nodes, which in turn require agent identity to be accounted for.
Clearly, these expected walks would be different for different ex-ante distributions that place
higher probability mass on specific sets of networks which contain specific sets of walks.
While anonymity is not imposed in our model, Proposition 3 provides a condition for it to
appear as if it arises in equilibrium. This appearance is due to the fact that equilibrium
actions are completely characterized by agent degrees, which in turn may be thought of as
abstracting away from the computation of possible walks. However, this is a consequence of
the uniform assumption and the corresponding expected walks it induces. As the following
lemma suggests, the uniform case exhibits the special property that the ex-post expectation
of any agent regarding any other’s (who is either connected or non connected to the former)
degree is the same for all agents.

Lemma 1. For any player i denote by Ei (dj1 ), Ei (dj2 ), .. the agent’s ex-post expectations of
any of its neighbor’s degree, any of its possible neighbor’s neighbor’s degree and so on. When
the underlying probability distribution is uniform, we have:

n
Ei [dj1 ] = Ei [dj2 ] = . . . . . . =
2

Consequently,  n p
(p)
βi,ti =
2

Uniformity in ex-ante beliefs provides the least amount of information with respect to the
identification of which walks are present in the network, inducing the trivial belief that each
other agent has a degree equal to n2 . Consequently, each agent expects the complementarity

p
strength of its action with any other agent due to a walk of length p to be equal to n2 .

Lemma 1 also speaks to the second assumption in Galleotti et al. (2010) regarding degree
independence. In their set up, independence of degrees implies that the belief that a player

19
who is of degree d, and that of another who is of degree d + 1, regarding the degrees of
each of their neighbors are the same. In our case, the same property holds, but it arises
endogenously as a result of ex-ante uniformity.

4 Welfare

Let W ∗ (g), W I (g) and W C (g) respectively denote ex-post, interim, and complete informa-
tion welfare, given that nature has chosen some g ∈ Gn . These are given by:

1 X
W ∗ (g) = a∗i (gi ) − a∗i (gi )2 + λa∗i (gi ) gij a∗j (gj )
2 j
n
I 1X ∗
W (g) = a (gi )2
2 i=1 i
n
C 1X c
W (g) = a (gi )2
2 i=1 i

where a∗i (gi ) is the BNE action of agent i who is of type gi and aci (gi ) is complete information
Nash equilibrium action of agent i whose row is also gi . Note that the difference between
ex-post and interim payoffs lies in the manner by which network externalities are accounted
for. Interim welfare considers these externalities in expectation, while ex-post computes the
actual level of externalities produced in equilibrium given that nature has chosen a particular
network.

Proposition 4. If the underlying probability distribution is uniform, then for a d−regular

graph with d ≤ n2 we have
W C (g) ≤ W ∗ (g)

Since the underlying distribution is assumed to uniform, Lemma 1 suggests that each agent

p
expects the complementarity of its action due to walks of length p to be n2 , regardless of
the realization of their type. In other words, even though the degree induced by their realized

p
type is less n2 , agents still expect that they will have n2 walks of length p which is less than
what they actually have in the realized network. Consequently, this induces over-exertion of
actions relative to the complete information Nash equilibrium leading to larger welfare.

20
5 Conclusion

We study a linear quadratic network game of incomplete information in which agents infor-
mation is restricted to the identity of their neighbors only. We characterize Bayesian-Nash
equilibria, demonstrating that agents will make use of local information in order to form
beliefs about the number of walks they have in the network and consequently the comple-
mentarity strength of their action with all other agents.

21
References

[1] Acemoglu, D., Ozdaglar,A. and Tahbaz-Selehi, A. (2016). Networks, shocks and sys-
temic risk. The Handbook of the Economics of Networks

[2] Calvo-Armengol, and Y. Zenou (2006). Who’s who in networks. wanted: The key player.
Econometrica 74(5), 1403–1417.

[3] Bravard, C., Durieu, J., Sarangi, S., & Semirat, S. (2022). False information from near
and far. Games and Economic Behavior.

[4] Bonacich, P. (1987). Power and centrality: A family of measures. American journal of
sociology, 92(5), 1170-1182.

[5] Banerjee, A., Chandrasekhar, A. G., Duflo, E and Jackson, .M. O (2013). The diffusion
of microfinance. Science 341, 1236498.

[6] Banerjee, A., Chandrasekhar, A. G., Duflo, E and Jackson, .M. O (2014). Gossip:
Identifying central individuals in a social network. NBER Working PaperNo. w20422

[7] Breza, Emily, Arun G. Chandrasekhar, and Alireza Tahbaz-Salehi (2018). Seeing the
forest for the trees? An investigation of network knowledge. No. w24359. National
Bureau of Economic Research

[8] Calvo-Armengol, A., E. Patacchini, and Y. Zenou (2009). Peer effects and social net-
works in education, Review of Economic Studies, 76, 1239–1267.

[9] Charness, G., F. Feri, M. A. Melendez-Jimenez, and M. Sutter (2014). Experimental

games on networks: Underpinnings of behavior and equilibrium selection. Econometrica
82(5), 1615– 1670.

[10] Crucitti, P., V. Latora, and S. Porta (2006). Centrality measures in spatial networks of
urban streets. Phys. Rev. E 73, 036125.

[11] Dasaratha, K. (2020). Distributions of centrality on networks. Games and Economic

Behavior 122, 1–27.

[12] De Marti, J. and Y. Zenou (2015). Network games with incomplete information. Journal
of Mathematical Economics 61, 221–240.

[13] De Paula, A., Rasul, I., and Souza, P. (2018). Recovering social networks from panel
data: identification, simulations and an application.

22
[14] Galeotti, A., S. Goyal, M. O. Jackson, F. Vega-Redondo, and L. Yariv (2010). Network
games. The review of economic studies 77(1), 218–244.

[15] Golub, B., and Morris, S. (2020). Expectations, networks, and conventions. arXiv
preprint arXiv:2009.13802.

[16] Goyal, S., & Joshi, S. (2003). Networks of collaboration in oligopoly. Games and Eco-
nomic behavior, 43(1), 57-85.

[17] Goyal, S. and J. L. Moraga-Gonzalez (2001): R&D networks. The RAND Journal of
Economics, 32, 686–707.

[18] Harsanyi, J. C. (1967). Games with incomplete information played by “bayesian” players,
i–iii part i. the basic model. Management science 14(3), 159–182.

[19] Hernandez, P., M. Munoz-Herrera, and A. Sanchez (2013). Heterogeneous network

games: Conflicting preferences. Games and Economic Behavior 79, 56–66.

[20] Kets, W. (2008). Beliefs in network games.

[21] Latora, V. and M. Marchiori (2007). A measure of centrality based on network efficiency.
New Journal of Physics 9(6), 188–188.

[22] Lewbel, A., Qu, X., and Tang, X. (2021). Social networks with unobserved links. Boston
College Working Papers in Economics 1004.

[23] Ruiz-Palazuelos, S. (2021). Clustering in network games. Economics Letters 205, 109922.

[24] Song, Y. and M. van der Schaar (2015). Dynamic network formation with incomplete
information. Economic Theory 59(2), 301–331.

23
Appendix

Proof of Proposition 1

Define a mapping P : Rγ −→ Rγ , such that

P (a) = 1 + λ(n).B(n)a

where, 
0 if i=j
γ×γ
[B(n)]ij =
G
i∼j if i ̸= j
and

[Gi∼j ]kl = gijk p(gjl |gik ) ∀l, k = 1(1)γ and ∀gjl ∈ Gj , gik ∈ Gi

Let (Rγ , ∥.∥∞ ) be a metric space with ∥.∥∞ being the sup-norm defined on Rγ , i.e.

∥a∥∞ = max{|aj | : j ∈ N }

Hence, we can write:

∥P (x) − P (y)∥∞ = |λ(n)| ∥B(n).(x − y)∥∞

≤ λ(n)(n − 1) ∥x − y∥∞
= r ∥x − y∥∞

where the first inequality results from the fact that ∥B(n).a∥∞ ≤ (n − 1) ∥a∥∞ for B(n) ∈
{0, 1}nγ . Thus, we get
∥P (x) − P (y)∥∞ ≤ r ∥x − y∥∞

so that P is a contraction mapping on Rγ as long as r ∈ [0, 1). This holds if:

1
0 ≤ λ(n).(n − 1) ≤ 1 ⇒ 0 ≤ λ(n) <
n−1
1
Hence, for 0 ≤ λ(n) < n−1 , P is a contraction mapping on Rγ and (Rγ , ∥.∥∞ ) is a complete
metric space. Therefore, by the Banach fixed point theorem, there exists an unique a∗ ∈ Rγ ,
such that
P (a∗ ) = a∗ ⇒ a∗ = 1 + λ(n).B(n)a∗

24
Consequently, there exists an unique pure strategy BNE for the game Γ whenever λ(n) ∈
1


0, n−1 .

Proof of Proposotion 2

The equilibrium actions can be written from (5) as,

a∗ = (I − λ(n)B(n))−1 · 1
= 1 + λB · 1 + λ2 B2 · 1 + . . . . . .

For an agent i ∈ N of type giti ,

a∗ = 1 + λ [B · 1]i,ti + λ2 B2 · 1 i,ti + . . . . . .
 

Then,

(1)
βi,ti = [B · 1]i,ti
Xn X
= [Gi∼j ]ti ,tj
j=1 tj ∈Gj
n X
t
X
= gijti p(gjj |giti )
j=1 tj ∈Gj
n
t
X X
= gijti p(gjj |giti )
j=1 tj ∈Gj
n
X
= gijti
j=1

25
Similarly,

(2)
βi,ti = B2 · 1 i,ti
 

= [B · (B · 1)]i,ti
Xn X
(1)
= [Gi∼j ]ti ,tj βj,tj
j=1 tj ∈Gj
n X n
t t
X X
= gijti p(gjj |giti ) gjkj
j=1 tj ∈Gj k=1
n n
t t
X X X
= gijti gjkj p(gjj |giti )
j=1 tj ∈Gj k=1

(m)
This calculation generalizes for βi,ti .

Proof of Lemma 1

Fix an agent i ∈ N , then, for any j a uniform ex-ante distribution implies

n−2

(n−1)(n−2) −(n−2) n−2


d−1
2 2
d−1
p(dj1 = d) = (n−1)(n−2)
=
2 2 2n−2

Similarly, we can write

n−2

(n−1)(n−2) −(n−2) n−2


d−1
2 2
d−1
p(dj2 = d) = (n−1)(n−2)
=
2 2 2n−2

and so on. Therefore,

n−1
X
Ei [dj2 ] = p(dj1 = d) d
d=1
n−1 n−2


X d−1
= d
d=1
2n−2
n
=
2

Similarly, we can show the rest.

26
Proof of Proposition 3

To prove proposition 3, if we characterise the actions of each agents with respect to degrees,
then first order condition can be re-written as:
n−1 n−2


X d−1
a(d) = 1 + λd a(d) (7)
d=1
2n−2

    " n−1   # 
n−2 n−2 λ X n−2 n−2
∴ a(d) = + n−2 a(d) d
d−1 d−1 2 d=1
d−1 d−1
n−1   n−1   " n−1   # n−1 
X n − 2
X n−2 X n−2 λ X n−2
⇒ a(d) = + n−2 a(d) d
d=1
d−1 d=1
d−1 2 d=1
d−1 d=1
d−1
n−1   " n−1   # n−1 
X n−2 λ X n − 2 X n − 2
n−2
⇒ a(d) = 2 + n−2 a(d) d
d=1
d−1 2 d=1
d−1 d=1
d−1

We have
n−1   n−1   Xn−1  
X n−2 X n−2 n−2
d= (d − 1) +
d=1
d−1 d=1
d−1 d=1
d−1
n−1
X (n − 2)!
= + 2n−2
d=1
(d − 2)!(n − 2 − d − 1)!
n−1  
X n−3
= (n − 2) + 2n−2
d=2
d − 2
= (n − 2)2n−3 + 2n−2

so we can write,

n−1   "
n−1   #
X n−2 λ X n − 2
a(d) = 2n−2 + n−2 a(d) (n − 2)2n−3 + 2n−2


d=1
d−1 2 d=1
d−1
2n−2
=
1 − nλ
2

Thus from equation (8),

λd
a(d) = 1 + , ∀d ∈ {0, 1, 2, . . . , (n − 1)} (8)
1 − nλ
2

27
■

Proof of Proposition 4

For the complete information case, the equilibrium actions of each agent in a realized d-
regular graph g ∈ G is given by:

λd
aci (g) = 1 + ∀i ∈ N
1 − λd

Since d ≤ n2 , we get that

n
1 − λd ≥ 1 − λ
2
λd λd
=⇒ 1 + ≤1+
1 − λd 1 − nλ
2

=⇒ aci (g) ≤ a∗i (g), ∀i ∈ N

Hence the ex-post utilities for each players,

n
1 c 2 1 ∗ 2 X
uci (g) − u∗i (g) ∗ ∗
= ai (g) − ai (g) + ai (g) − λai (g) gij a∗j (g)
2 2 j=1
1 1
= aci (g)2 − a∗i (g) + a∗i (g)2 − λda∗i (g)2
2 2
1 c 2 1
≤ ai (g) − aci (g) − (λd − )aci (g)2
2 2
c 2 c
= (1 − λd)ai (g) − ai (g)
= ((1 − λd)aci (g) − 1)aci (g)
=0

Where the second equality is due to the fact that for a d-regular graph realization, a∗i (g) =
a∗j (g), ∀i, j and the third inequality is due to the fact that aci (g) ≤ a∗i (g), ∀i ∈ N . Thus we
have

uci (g) − u∗i (g) ≤ 0, ∀i ∈ N

Summing over all agents gives the result.

28