Game Theory

Lecture #6

Focus of Lecture:

• Cost Sharing Games

• Core

• Minimum Spanning Tree Games

1 The Cost Sharing Problem

The previous lectures focused on two problems pertaining to the analysis and design of
sociotechnical systems. The first problem we discussed was social choice, where the goal was
to aggregate the preferences of many individuals while satisfying five desirable axioms. Here,
we presented a negative result by Kenneth Arrow that demonstrated that there does not exist
a social choice mechanism that satisfies this objective. The second problem we discussed was
the stable matching problem, where our goal was to produce an efficient matching between
two distinct groups. Here, we presented a series of positive results demonstrating that
(i) a stable matching always exists and (ii) the Gale-Shapley algorithm will always find
a stable matching. The sharp contrast between these two sets of results highlights that
a system designer should not take anything for granted when working with sociotechnical
systems. Rather, a system designer should approach the analysis and design of such systems
in axiomatic way, rigorously arguing about the resulting conclusions from a first principles
perspective.
This lecture will present on a third problem, namely cost sharing. In any cost sharing
problem there is a collection of individuals that engage in a joint venture accompanied by
an associated cost. How should the costs associated with this joint venture be distributed to
the individuals? Note that this choice will directly impact the behavior of individuals who
seek to meet their demands at the least possible cost. The goal here is to establish a cost
sharing protocol that promotes the most efficient structures to emerge, where efficiency is
measured by the cumulative cost incurred to meet the demands of all. (Note that the above
discussion also could have been expressed in terms of revenue or profit sharing.)
The following examples shed some light on the challenges associated with cost sharing.

Example 1.1 (A Cost Sharing Problem Among Two Towns) Suppose two nearby towns,
which we will refer to as A and B, respectively, decide to construct a joint facility to serve the
needs of both towns, e.g., a community center, a water distribution center, or a recreational
facility. Each town is required to provide this facility to their community and must decide
whether to proceed independently or jointly. The costs associated with the different scenarios
are as follows:
 15

B’s
Payment

7
The core

11 15 A’s
Payment

Figure 1: Illustration of the core in the two town example.

• Town A can build its own facility for $11 million

• Town B can build its own facility for $ 7 million

• A joint facility serving Town A and B costs $15 million

Clearly, having the towns construct a joint facility is the fiscally advantageous option—$15
million jointly vs $18 million individually. However, pursuing this option requires that the
towns find an equitable division of the $15 million costs.
What division of costs would incentivize each of the towns to engage in this joint venture?
One possible division of costs is to divide the costs equally, i.e., each Town A and B pays
$7.5 million. However, Town B would not engage in such a joint venture as Town B can
build its own facility for only $7 million. Hence, any division of costs that incentive this
joint venture must ensure that each town pays less than the alternative cost of going alone.
Figure 1 illustrates all costs divisions that incentive both Town A and Town B to pursue a
joint venture. In the above example, such a division of costs must satisfy the following three
conditions: (i) the costs to Town A and B equals $15 million; (ii) the cost to Town A is less
than $11 million; and (iii) the cost to Town B is less than $7 million. Here, the strategic
behavior, or self-interest, of the towns imposes the constraints (ii) and (iii).

In the general setting of cost sharing games, the set of all cost divisions that incentivize the
towns to pursue the fiscally advantageous joint venture is known as the core.
The following example illustrates this concept in a slightly more general setting, namely, one
with three towns.

Example 1.2 (A Cost Sharing Problem Among Three Towns) Now add a third Town
C to Example 1.1. Once again, each town is required to provide this facility to their commu-
nity and can proceed alone, or all three towns can enter a partnership. What is new in this
setting is that there is a new option for two towns to form their own partnership. The costs
associated with the various scenarios are as follows:

• Town A can build its own facility for $11 million

• Town B can build its own facility for $7 million

• Town C can build its own facility for $8 million

• Towns A and B can build a joint facility for $15 million

• Towns A and C can build a joint facility for $14 million

• Towns B and C can build a joint facility for $13 million

• Towns A, B, and C can build a joint facility for $20 million

While not as obvious than in the previous example, having all three towns partner to form
a joint facility ($20 million) is the most fiscally responsible option. Verifying this statement
would require comparing the total cost of a three way partnership with all other combinations,
i.e., the total cost of Towns A, B, and C each building their own facility; Towns A and B
forming a partnership, with C building its own; etc. Still, a three way partnership requires
that the three towns find an equitable division of the $20 million cost.
What division of costs would incentivize each of the towns to engage in this joint venture?
One possible division of costs is to divide the costs equally, i.e., each town pays $20/3 million.
At first glance, it seems appealing since the distributed cost of $20/3 is less than each town’s
solo cost, e.g., Town B prefers to pay $20/3 million over its solo cost of $7 million. However,
we must now examine also the alternatives where only two of the three towns pursue a joint
venture, e.g., Towns B and C engaging in a joint venture. Note that if costs are divided
equally in the three-way partnership, Towns B and C would together pay a combined $40/3
million. However, Towns B and C could instead pursue their own joint venture, resulting
in a total cost of $13 million, which is less than $40/3 million (i.e, $20/3 × 2). Hence, an
equal division of costs in the three-way partnership is not in the core.
A division of costs that incentivizes the three towns to pursue a joint ventures, i.e., is in the
core, must satisfy the following:

• The costs for Towns A, B, and C totals $20 million

• The cost for Town A ≤ $11 million

• The cost for Town B ≤ $7 million

• The cost for Town C ≤ $8 million

• The total costs for Towns A and B ≤ $15 million

• The total costs for Towns A and C ≤ $14 million

 A pays 20

A+C=14
B=7
A = 11

The Core
B+C=13
Equal share

C=8 A+B=15
C pays 20 B pays 20

Figure 2: Illustration of the core in the three town example.

• The total costs for Towns B and C ≤ $13 million

Figure 2 illustrates these constraints and the resulting core for this three town cost sharing
problem. To understand this plot, first consider the line connecting the A and C vertices.
This line identifies all possible cost shares where Town B pays $0. All parallel lines to the
(A, C) line represents divisions where B pays a constant amount, and the amount B pays
increases as we get closer to the B vertex.

2 Model
A cost sharing problem is defined by (i) a player set N = {1, 2, . . . , |N |} and (ii) a cost
function1 c : 2N \∅ → R that defines the collective cost for any subset of players S ∈ 2N .
For the three town example given above, we have N = {A, B, C} and c({A}) = 11, c({B}) =
7, c({C}) = 8, c({A, B}) = 15, c({A, C}) = 13, c({B, C}) = 10, and c({A, B, C}) = 20.
A standing assumption is that the cost of the entire set N is less than the sum of the costs
of any partition2 of N , i.e.,
XK
c(N ) ≤ c(Si ) (1)
i=1
for all partitions {S1 , S2 , ..., SK } of N . The interpretation is that it is fiscally advantageous
for all players to be part of a joint venture.
1
Recall that the notation 2N denotes the power set of N , i.e., the set of all subsets of N . For example, if
N = {A, B, C} then 2N = {∅, {A}, {B}, {C}, {A, B}, {A, C}, {B, C}, {A, B, C}}.
2 N
AT S {S1 , S2 , ..., SK } such that for all i, j = 1, 2, ..., K: (i) Si ∈ 2 ;
partition of N is a collection ofSsubsets
(ii) Si Sj = ∅ for i 6= j; and (iii) S1 ... SK = N .
Definition 2.1 (Core) The core of a cost sharing game is the set of vectors v ∈ R|N | such
that (i) v1 + ... + v|N | = c(N ) and (ii) for any subset S ⊆ N
X
vi ≤ c(S)
i∈S

It may be the case for a specific cost sharing game that no vector v ∈ R|N | satisfies these
conditions, in which case we will say that the core is empty.
The vector v represents the player by player allocation of costs in a joint partnership, c(N ).
This interpretation is captured in condition (i) of the definition of the core. Condition (ii)
reflects that sum of costs of a subset is less than the total cost of that subset. To interpret
this statement, suppose that for some subset S ∈ 2N ,
X
vi > c(S).
i∈S

This inequality implies that it would be advantageous for all members of S to defect from
the joint partnership and form a partnership amongst themselves. In particular, there exists
an alternative allocation of costs, v 0 ∈ R|N | , such that (i) i∈S vi0 = c(S) and (ii) vi0 < vi for
P
all i ∈ S. Condition (ii) of the core definition removes the incentive for such defections.
In the two player example, there were 2 constraints on the allocation needed for the core. In
the three player example, there were 6 constraints on the allocation needed for the core. In
both of these specific cases, the core was not empty.
In general, there are 2|N | − 2 constraints, which is exponential in the number of individu-
als |N |. Exhaustively checking all of these constraints is computationally prohibitive and
intractable, hence more general arguments are needed to be developed for ensuring the exis-
tence (or lack thereof) with regards to the core. Clearly, any results along this direction are
going to hinge on the structure of the cost function, c(·).

3 Minimum Spanning Tree Games

In general, the core could be either empty or non-empty, and there are advanced results that
can be employed to make this determination (e.g., the Bondareva-Shapley theorem and the
concept of balanced games). Rather than introducing these concepts here, we will study
a specific setting for which the core is always non-empty, namely minimum spanning tree
games. These games model the problem of distributing the infrastructure costs incurred by
connecting a set of customers to a common resource. This problem appears in a variety of
contexts, such as distributing costs in multicast transmissions.
A minimum spanning tree game consists of a finite set of customers N that need to be
connected, either directly or indirectly, to a single supplier, which we denote by ∅. We
define an interconnection graph with nodes N ∗ = N ∪ {∅} and directed edges E = N ∗ × N ∗ ,
where each edge (i, j) ∈ E is associated with a given cost cij ≥ 0. Here, we adopt the
convention that the edge (i, j) is a directed edge pointing from i to j. A spanning tree of the
above interconnection graph is defined as a collection of edges EN ⊆ E such that there is a
unique path from every node i ∈ N to the source {∅} in the edge set EN . More formally,
from every node i ∈ N , there is a unique sequence of nodes i = i0 , i1 , ..., ik = {∅} such that
i0 , . . . , ik−1 ∈ N and (ix , ix+1 ) ∈ EN for all x ∈ {0, . . . , k − 1}. Accordingly, any spanning
tree EN will have exactly |N | edges with a unique edge leaving each node N .
We will call a spanning tree, EN , a minimum spanning tree if there does notPexist another
0
collection of edges EN that form a spanning tree with a lower total cost, i.e., (i,j)∈EN cij ≤
0
P
c
(i,j)∈E 0 ij for any spanning tree EN .
N

Likewise, we also can define a spanning tree and minimum spanning tree associated with any
coalition S ⊆ N in the same fashion, where we restrict attention to nodes in S and edges in
(S ∪ {∅}) × (S ∪ {∅}).
With these definitions in hand, we can now define the cost function c : 2N \∅ → R as follows.
For any set S ∈ 2N , let ES∗ denote a minimum spanning tree restricted to the nodes in S,
and define X
c(S) = cij
∗
(i,j)∈ES

An example of a minimum spanning tree is presented below. Here, there are four individuals
N = {1, 2, 3, 4}, and the relevant (undirected) edge costs cij are shown in the figure on the
left, e.g., c30 = c03 = 4 and c24 = c42 = 3. The edges that are not explicitly highlighted can
be thought of as having infinite cost, e.g., c23 = c32 = ∞.
Clearly, in any minimum spanning tree game, the minimum spanning tree for the joint
venture of all individuals N is the graph that meets all the individual demands at the lowest
total cost. Accordingly, the standing assumption in (1) is satisfied.
Are there cost allocations that incentivize this outcome? In other words, is the core guaran-
teed to be non-empty for any minimum spanning tree game? The following theorem provides
the answer to this question.

0 0

5 6

1 2 1 2
5
4 2 2
3 3 3 3

3 4 3 4
1 1

Minimum Spanning Tree Game Minimum Spanning Tree

Theorem 3.1 The core of any minimum spanning tree game is non-empty.
The proof of this theorem is straightforward and amounts to showing that a particular
allocation structure is always in the core, and hence the core must be non-empty. We
construct these cost shares using the structure of the minimum spanning tree for the set
∗
N , which we denote by EN . For each individual i ∈ N there exists a single directed edge
∗
(i, j) ∈ EN where j ∈ N ∪ {0}. Accordingly, we define the cost share of each individual
i ∈ N as the cost of this outgoing edge (i, j), i.e., vi = cij . For the above minimum spanning
tree game we have:
v1 = 3, v2 = 3, v3 = 1, v4 = 2.
Note that by definition of c(N ), v1 + ... + v|N | = c(N ), and so condition (i) of Definition 2.1
is satisfied.
The following arguments verify that these cost shares are in the core. Proving this statement
requires showing that the minimum spanning tree for any coalition S ⊆ N , denoted by ES∗ ,
must satisfy (see condition (ii) in Definition 2.1)
X X
cij ≤ cij = c(S), (2)
∗ :i∈S
(i,j)∈EN ∗ :i∈S
(i,j)∈ES

meaning that it is the best interest of coalition S to stay in the joint venture N . We will
establish this inequality by contradiction. Suppose that
X X
cij > cij , (3)
∗ :i∈S
(i,j)∈EN ∗
(i,j)∈ES

and consider the collection of edges

0 ∗ ∗
EN = (EN − {(i, j) ∈ EN : i ∈ S}) ∪ ES∗
∗
which entails removing the outgoing edges of each individual i ∈ S in the graph EN and
∗
replacing them with the edges in the minimum spanning tree ES . Given (3), we know that
0 ∗
c(EN ) < c(EN ).
0
We will now complete the proof by arguing that the graph EN is in fact a spanning tree
∗
for the set N , which provides our contradiction since EN was a minimum spanning tree.
0
Showing that the graph EN is a spanning tree involves arguing that each node i ∈ N has
0
a path to the source {∅} in the tree EN . First, note that if i ∈ S then this conclusion is
immediate because there is a path to the source {∅} in the set ES∗ ⊆ EN 0
. Next, if i ∈
/ S then
∗
there are two options: (i) the path from i to the source in EN did not contain any nodes in
∗
S or (ii) the path from i to the source in EN did contain nodes in S. If (i), then there is still
0 ∗
a path from i to {0} in the graph EN as this is the same path as in EN since none of those
edges were removed. Alternatively, if (ii) then there exists a path to some node in k ∈ S.
However, since ES∗ is a minimum spanning tree for the nodes S, this means that there exists
0 0
a path from k to {∅} in EN . Hence, there is also a path from i to {∅} in EN . Accordingly,
0
EN is a spanning tree for the set N , which completes the proof.
4 Conclusions
In this chapter we introduced the problem of cost sharing and the associated notion of the
core, and we presented examples where the core is non-empty. The next lecture will focus
on identifying mechanisms for generating cost shares that are in the core for certain problem
instances.

5 Exercises
1. Consider the following cost sharing problem:
• Player set: N = {1, 2, 3}
• Costs: c : 2N \∅ → R
c({1}) = 9, c({2}) = 8, c({3}) = 9
c({1, 2}) = 14, c({1, 3}) = 15, c({2, 3}) = 13
c({1, 2, 3}) = 21
(a) Identify the core graphically.
(b) Is the core nonempty? If so, provide an allocation in the core.
2. Consider the example of the minimum spanning tree game considered in this lecture.
For each coalition S ⊆ {1, 2, 3, 4} determine the minimum spanning tree and resulting
cost. Furthermore, define the proposed allocation (i.e., cost shares) for each coalition
S. Note that for each coalition S the accompanying minimum spanning tree will have
exactly |S| edges.
3. A scientist has been invited for consultation at three distant cities. In addition to her
consultation fees, she expects travel compensation. But since these three cities are
relatively close together, travel expenses can be greatly reduced if she accommodates
them all in one trip. The problem is how to decide how the travel expenses should be
split among her hosts in the three cities. The one-way travel expenses among these
three cities, A, B, and C, and her home base, H, are given as follows:
• Between H and A, cost = 7
• Between H and B, cost = 9
• Between H and C, cost = 6
• Between A and B, cost = 2
• Between A and C, cost = 4
• Between B and C, cost = 4
Assume that the value of the visit is the same for each of the hosts, say 20 units each.
Set the problem up as a three-player cost-sharing game by specifying the opportunity
cost function.