Algorithmic Game Theory, Summer 2019 Lecture 9 (5 pages)

Inefficiencies in Cost and Market Sharing Games

Thomas Kesselheim Last Update: May 6, 2019

Today, we will wrap up our discussion of the inefficiency due to selfish behavior and show
two further results. The first one is about a cost-minimization game, for which we will consider
a less pessimistic perspective than the price of anarchy. The second one is about a very similar
game but now with a utility-maximization objective.

1 Cost-Sharing Games
We first turn to fair cost-sharing games, which are congestion games with delays dr (x) = cr /x
for constant cr for every resource r ∈ R. That is, we have a set N of n players and a set R of
m resources. Player i allocates some resources, i.e., his strategy set is Σi ⊆ 2R . Each resource
r ∈ R has fixed cost cr ≥ 0. The cost cr is assigned in equal shares to the players allocating r
(if any).
Social cost turns out to be the sum of costs of resources allocated by at least one player:
X XX X X
SC(S) = ci (S) = dr (nr (S)) = nr (S) · cr /nr (S) = cr . (1)
i∈N i∈N r∈Si r∈R r∈R
nr (S)≥1 nr (S)≥1

The price of anarchy for pure Nash equilibria can be as big as the number of players n, even
in a symmetric game. For  > 0, consider the example
1+

s t
n

Edge labels indicate the cost value cr for this resource. It is a pure Nash equilibrium if all
players use the bottom edge, whereas the social optimum would be that all users use the top
edge.
Although this is a very stylized example, there are indeed examples of such bad equilibria
occurring in reality. A prime example are mediocre technologies, which win against better
ones just because they are in the market early and get their share. This way, they are widely
supported. Maybe another example are social networks and messaging apps. Many people
would prefer not to use, say, Facebook but they cannot switch to an alternative platform unless
their friends do.

1.1 Price of Stability

The price-on-anarchy viewpoint is still a pessimistic one because we make statements about
the worst equilibria. This is different in price of stability. For an equilibrium concept Eq, it is
defined as
minp∈Eq SC(p)
P oSEq = .
mins∈S SC(s)
As the set of equilibria gets larger, the minimum gets smaller and smaller. Therefore, if the
respective equilibria exist, we now have

1 ≤ P oSCCE ≤ P oSCE ≤ P oSMNE ≤ P oSPNE ≤ P oAPNE ≤ . . . .

Algorithmic Game Theory, Summer 2019 Lecture 9 (page 2 of 5)

Theorem 9.1. In a symmetric cost-sharing game, the price of stability for pure Nash equilibria
is 1.

Proof. We explicitly construct a pure Nash equilibrium as follows. It is a symmetric equilibrium,

meaning that all playersP use the same strategy. Consider player 1, and set S1 to the strategy
from Σ1 that minimizes r∈S1 cr . Set S2 = . . . = Sn = S1 . This certainly minimizes social cost
according to Equation (1).
It is also an equilibrium because for each i strategy Si is a best response against S−i . To
see this, consider some alterantive Si0 ∈ Σi . We have
X cr X 1 X 1 X
ci (Si0 , S−i ) = + cr ≥ cr ≥ cr = ci (S) ,
n n n
r∈Si0 ∩Si 0 0
r∈Si \Si r∈Si r∈Si
P
where we used that Si was chosen to minimize r∈Si cr .

The core insight of the previous proof is that in a symmetric game every social optimum is
a pure Nash equilibrium. For general, asymmetric games, the social optimum is not necessarily
a pure Nash equilibrium. Consider the following game with n players. Each player i has source
node si and destination node t.

1 1 1
1 2 3 n

s1 s2 s3 ··· sn 1+

0 0 0 0

A player two possible strategies: Either take the direct edge or take the detour via v. The
social optimum lets all players choose the indirect path, ending up with social cost 1 + . This,
however, is no Nash equilibrium. Player n would opt out and take the direct edge. Therefore,
the only pure Nash
Pnequilibrium lets all players choose their direct edge, yielding social cost of
Hn . Here, Hn = i=1 i = 1 + 21 + 13 + . . . + n1 denotes the n-th harmonic number. We have
1

Hn = Θ(log n).

Theorem 9.2. The Price of Stability for pure Nash equilibria in fair cost sharing games is at
most Hn .

Proof. Rosenthal’s potential function for cost-sharing delays is

nr  
XX X 1 1 1
Φ(S) = cr /i = cr · 1 + + + . . . +
2 3 nr
r∈R i=1 r∈R
nr ≥1
X
≤ cr · Hn
r∈R
nr ≥1
= SC(S) · Hn .
Algorithmic Game Theory, Summer 2019 Lecture 9 (page 3 of 5)

In Φ(S) we account for each player allocating resource r a contribution of cr /i for some
i = 1, . . . , nr , whereas in his cost ci (S) we account only cr /nr . Hence, for every state S of a
cost-sharing game we have
SC(S) ≤ Φ(S) ≤ SC(S) · Hn .
Now suppose we start at the optimum state S ∗ and iteratively perform improvement steps for
single players. This eventually leads to a pure Nash equilibrium. Every such move decreases
the potential function. For the resulting Nash equilibrium S we thus have Φ(S) ≤ Φ(S ∗ ) and

SC(S) ≤ Φ(S) ≤ Φ(S ∗ ) ≤ SC(S ∗ ) · Hn .

This proves that there is a pure Nash equilibrium that is only a factor of Hn more costly than
S∗.

2 Market-Sharing Games
Most of our examples so far in this course were cost-minimization games. For the basic defini-
tions there is no real difference when one turns to utility-maximization games instead. However,
for the price of anarchy, the story is different, as we will seen in the following example.
Let us consider the following market sharing game. There are n firms, which are our players
N , and m markets M . Each firm can decide to invest in one of these markets. Therefore, for
player i ∈ N , the strategy set Si is a subset of M .
Each market j ∈ M has a total demand vj . If k firms invest in the same market, then the
v
market’s demand is shared equally. So every firm gets a utility of kj .
This way, the utility of player i ∈ N in state s ∈ S is
vsi
ui (s) = , where nj (s) = |{i ∈ N | si = j}| .
nsi (s)

The social welfare of a state s is defined as the sum of player utilities, or equivalently, as
the sum of demands that are fulfilled
X X X
SW (s) = ui (s) = vj = vj .
i∈N j∈M j∈{s1 ,...,sn }
nj (s)≥1

Example 9.3. There are n markets 1, . . . , n; each player can invest in every market. For some
 > 0, the demands are v1 = n + , v2 = . . . = vn = 1.
The social welfare is maximized by each player investing in a different market. In this case,
SW (s) = 2n − 1 + . However, the only pure Nash equilibrium is that all players invest in
market 1. Here, SW (s) = n + .

So far, these games look a lot like the cost-sharing games. And, indeed, they are. We can
even interpret them as congestion games: Set R = M , so the markets become the resources,
v
and set dj (k) = − kj for all j ∈ M and all k. Now the players’ cost functions in the congestion
game are exactly the negative utility functions of the market sharing game: ci (s) = −ui (s).

Observation 9.4. Every Market Sharing Game has a pure Nash equilibrium.

2.1 Price of Anarchy

Interestingly, despite the similarity to cost-sharing games, the price of anarchy is a lot differ-
ent. Let us first define the price of anarchy for utility-maximization games. The definition is
analogous to the one for cost-minimization game.
Algorithmic Game Theory, Summer 2019 Lecture 9 (page 4 of 5)

Definition 9.5. Given a utility-maximization game, let Eq be a set of probability distributions

over
P the set of states S. For some probability distribution p, let SW (p) = Es∼p [SW (s)] =
s∈S p(s)SW (s) be the expected social welfare. The price of anarchy for Eq is defined as

maxs∈S SW (s)
P oAEq = .
minp∈Eq SW (p)

So, we swap minima and maxima and the social optimum is now in the numerator and the
equilibrium in the denominator. This way, the PoA is still greater than 1. You will also find it
defined as the reciprocal. In any case, values closer to 1 are better.
We can also adapt the smoothness definition as follows.

Definition 9.6. A utility-maximization game is called (λ, µ)-smooth for λ > 0 and µ ≥ 0 if,
for every pair of states s, s∗ ∈ S, we have
X
ui (s∗i , s−i ) ≥ λ · SW (s∗ ) − µ · SW (s) .
i∈N

There is again an analogous theorem that smoothness implies a bound on the price of
anarchy.

Theorem 9.7. In a (λ, µ)-smooth utility-maximization game, the PoA for coarse correlated
equilibria is at most 1+µ
λ .

Proof. For simplicity, we prove the theorem only for pure Nash equilibria. The generalization
to coarse correlated equilibria works exactly as in the case of cost-minimization games.
Let s be a pure Nash equilibrium, s∗ be a social optimum. Then we have
X
SW (s) = ui (s)
i∈N
X
≥ ui (s∗i , s−i )
i∈N
≥ λ · SW (s∗ ) − µ · SW (s) .

So (1 + µ)SW (s) ≥ λSW (s∗ ).

So, it only remains to give a smoothness proof.

Theorem 9.8. The market-sharing game is (1, 1)-smooth. So P oACCE ≤ 2.

Proof. Observe that

vs∗i
ui (s∗i , s−i ) =
ns∗ (s∗i , s−i )
i
vs∗i if s∗i 6∈ {s1 , . . . , si−1 , si+1 , . . . , sn }
≥
0 otherwise
vs∗i if s∗i 6∈ {s1 , . . . , sn , s∗1 , . . . , s∗i−1 }

≥ . (2)
0 otherwise

Denote by T = {s1 , . . . , sn } all markets invested in in s and by T ∗ = {s∗1 , . . . , s∗n } all markets
invested in in s∗ . We can now write
X X X X
ui (s∗i , s−i ) ≥ vj = vj − vj .
i∈N j∈T ∗ \T j∈T ∗ ∪T j∈T
Algorithmic Game Theory, Summer 2019 Lecture 9 (page 5 of 5)

This is because if we take the sum over the terms (2) then every element of T ∗ exactly appears
once unless it is in T .
This gives us
X X X X X
ui (s∗i , s−i ) ≥ vj − vj ≥ vj − vj = SW (s∗ ) − SW (s) .
i∈N j∈T ∗ ∪T j∈T j∈T ∗ j∈T

This is exactly the requirement for smoothness.

Further Reading
• Chapter 19.3 in the AGT book. (PoS bound)

• Tim Roughgarden’s lecture notes http://theory.stanford.edu/~tim/f13/l/l15.pdf

and lecture video https://youtu.be/VjCKNl-GENI

• A. Vetta, Nash Equilibria in Competitive Societies, with Applications to Facility Location,

Traffic Routing and Auctions, FOCS 2002 (Generalization of the result for market sharing
game)