4 Mixed strategies in strategic games

4.1 Mixed strategies and mixed-strategy Nash equilibria

Example 4.1 (Matching pennies). Alice and Bob both turn a
H T
penny with either heads or tails. If they match, Alice takes them,
H 1, -1 -1, 1
otherwise Bob takes them. The strategic form is on the right; as
T -1, 1 1, -1
usual, Alice is the row player and Bob is the column player.
Game Theory cannot possibly advise Alice to play H or T. For instance, if the correct move for
Alice is H, then Bob can play T and win. A more natural course of action is to play randomly:
for instance, Alice can choose to play H with probability ↵ and T with probability 1 ↵, and
Bob can choose H with probability and T with probability 1 . We can then look which
values of ↵, lead to desirable outcomes.
Note that if Alice chooses to play H with ↵ = 0, then that is the same as choosing the pure
strategy of playing T. Thus our expanded notion of strategies by including randomness still
includes pure strategies. The formal definitions are as follows.

Definition 4.2. Consider a game (S1 , . . . , Sn , u1 , . . . , un ). Assume that all strategy sets
S1 , . . . , Sn are finite (so each player has a finite number of actions to choose from).
P
• A mixed strategy for player i is a function p : Si ! [0, 1] such that si 2Si p(si ) = 1.
The real number 0  p(si )  1 should be interpreted as the probability that player i plays
the pure strategy si .
• Given mixed strategies p1 , . . . , pn for players 1, . . . , n, the expected utility or expected
payo↵ ui (p1 , . . . , pn ) for player i when these mixed strategies are played is defined by the
equation
X
ui (p1 , . . . , pn ) = p1 (s1 ) . . . pn (sn )ui (s1 , . . . , sn ).
(s1 ,...,sn )2S1 ⇥···⇥Sn

• The support of a mixed strategy p : Si ! [0, 1] for player i, denoted by Supp(p), is the
set of all set of pure strategies which occur with positive probability. That is,

Supp(p) = {s 2 Si | p(s) > 0}.

Remark 4.3. A mixed strategy for Player i is nothing more than a probability distribution on
the strategy set Si . Given mixed strategies p1 , . . . , pn for Player 1, . . . , Player n respectively, we
then have the product probability distribution on the strategy profile set S1 ⇥ · · · ⇥ Sn where
the strategy profile (s1 , ⇥, sn ) 2 S1 ⇥ · · · ⇥ Sn is assigned the probability p1 (s1 ) · · · pn (sn ). We
are making use of the independence of actions of di↵erent players here. The expected utility is
then nothing more than the expectation of the random variable ui : S1 ⇥ · · · ⇥ Sn ! R.

So that there is no ambiguity, we formally record the following running hypothesis:

13
 Assumption: (The Von-Neumann, Morgenstein expected utility hy-
pothesis.) Given the choice between two uncertain outcomes, players
will choose the one with highest expected utility.

In preceding chapters, utility encoded rankings and only the preference mattered. Now the
actual values of the utility function matter! Note that we have transformed our original game
to a game where probability distributions become strategies and expected utility as the utility
function, and we can transfer the discussion from Chapter 2 to this setting.

Definition 4.4. Mixed strategies p1 , . . . , pn for players 1, . . . , n are a mixed strategy Nash
equilibrium of the game if for all 1  i  n and any mixed strategy p0i for player i we have

0
ui (p1 , . . . , pi 1 , pi , pi+1 . . . pn ) ui (p1 , . . . , pi 1 , pi , pi+1 . . . pn ).

In other words, each pi gives a response with highest expected value to the other mixed strate-
gies.

Example 4.5 (Matching pennies again). Suppose Alice plays p where p(H) = ↵, p(T ) = 1 ↵,
and Bob plays q where q(H) = , q(T ) = 1 . The expected utility u1 (p, q) of Alice is then

u1 (p, q) = p(H)q(H)u1 (H, H) + p(H)q(T )u1 (H, T ) + p(T )q(H)u1 (T, H) + p(T )q(T )u1 (T, T )
=↵ ↵(1 ) (1 ↵) + (1 ↵)(1 )
=1 2↵ 2 + 4↵ = (1 2↵)(1 2 ).

Likewise, we find that the expected utility for Bob is given by

u2 (p, q) = 1 + 2↵ + 2 4↵ = (1 2↵)(1 2 )= u1 (p, q).

We can now find Alice’s best response to Bob paying the mixed strategy q (that is, Bob plays
H with probability ). For this, we need to maximise u1 for a given . If < 1/2, then
u1 (p, q) = (1 2↵)(1 2 ) is maximised at ↵ = 0. If > 1/2 on the other hand, then we
need to take ↵ = 1. Finally, if = 1/2, then Alice can take any value of ↵ between 0 and 1.
Similarly, u2 is maximised by = 0 if ↵ > 1/2, by = 1 if ↵ < 1/2, and any when
↵ = 1/2. We can now conclude that the only mixed strategy Nash equilibrium occurs with
↵ = = 1/2. This fits pretty well with our intuition!

Example 4.6 (The battle of the sexes). Alice and Bob want to meet up for a night out. Alice
suggests going to the pub and Bob suggests watching a film, and both would prefer meeting up
to missing out. Unfortunately, the MAS348 lecture is over before they can agree and they rush
to di↵erent lectures. Realising they don’t have a way to contact each other before their date,
they have to decide independently where to go. How should they strategise?
Let us model Alice and Bob’s problem by the following game (where, as usual, Alice is the
row player and Bob is the column player).

14
 Pub Film
Pub 3,1 0,0
Film 0,0 1,3

This game has two Nash equilibria in pure strategies, namely (P ub, P ub) and (F ilm, F ilm).
Are there any other Nash equilibria?
For 0   1, let [ ] denote the mixed strategy of going to the pub with probability and
opting for cinema with probability 1 .
Suppose Alice plays [p] and Bob mixes with [q]. Then, Alice’s expected utility is 3pq + (1
p)(1 q) and Bob’s expected utility is pq + 3(1 p)(1 q). To find Alice’s best responses to
Bob playing [q], Alice will maximise her payo↵

3pq + (1 p)(1 q) = 1 p q + 4pq = 1 q + p(4q 1)

as a function of p. There are three cases to distinguish depending on the sign of 4q 1 and we
obtain the following:
• if q > 1/4, then Alice’s expected utility is maximised when p = 1;
• if q < 1/4, then Alice’s expected utility is maximised when p = 0;
• if q = 1/4, then Alice’s expected utility is 3/4, regardless of p.
Similarly, Bob’s best responses to Alice playing p are given by q = 0 when p < 3/4, by q = 1
when p > 3/4, and any choice of q when p = 3/4 with expected utility 3/4.
Now let ([p], [q]) be a Nash equilibrium. If p < 3/4, then q = 0, and consequently p = 0. If
p > 3/4, then q = 1, and so p = 1. The two Nash equilibria ([0], [0]) and ([1], [1]) are the ones
that come from pure strategies.
We now consider p = 3/4. If q 6= 1/4, then a similar analysis leads to p = 0 or p = 1.
Hence q = 1/4, and so ([3/4], [1/4]) is a Nash equilibrium. Note that in equilibrium each player
has an expected utility independent of the strategy adopted by the other player. This is not a
coincidence.

The Indi↵erence Principle. The preceding method of calculating best responses is unfeasible
even for slightly more complex games (for example, 2-player games with 3 by 3 payo↵ matrix).
A more e↵ective process is to use the Indi↵erence Principle, which we now describe.
Consider an n-player game (S1 , . . . , Sn ; u1 , . . . , un ) with finite strategy sets.

Notation 4.7.
• Let s be an action/pure strategy of a player in a game. We will denote the mixed strategy
which plays s with probability 1 by sb. If it is clear from context and there is cause
confusion, then we will continue to use s instead of sb.
• Let p = (p1 , . . . , pn ) is a mixed strategy profile. Then we write p i to denote the mixed
strategies of other players di↵erent from player i.

15
 Key Observation. Suppose we are given actions a, b 2 Si of player i such that

a, p i ) > ui (bb, p i ).
ui (b

In other words, player i does strictly better by playing a instead of b when faced with p i .
Then b cannot be in the support of any mixed best response to p i (that is, b cannot
occur with non-zero probability in any best response). This is because player i can always
get a better return by replacing every occurence of b with a.

Theorem 4.8 (The Indi↵erence Principle). Suppose p is a best response from player i given
their belief that other players have played p i = {pj |j 6= i}. We must then have ui (b s, p i ) =
ui (p, p i ) for any pure strategy in the support Supp(p) (that is, any s 2 Si with p(s) > 0).

Proof. From the Key Observation, we see that all actions in the Supp(p) must all have the
same expected payo↵ against p i . If we set ui (t, p i ) = c for all t 2 Supp(p), then
X X X
ui (p, p i ) = p(t)ui (b
t, p i ) = p(t)ui (b
t, p i ) = c p(t) = c.
t2Si t2Supp(p) t2Supp(p)

Observation, following on from Key Observation and Indi↵erence Principle. Player i’s
best responses to p i = {pj |j 6= i} consist of those mixed strategies supported only on those
actions/pure strategies of player i that give maximum return against p i . ⇤

Example 4.9 (Rock-Paper-Scissors). Let us find Nash equilibrium mixed strategy profiles for
the following obscure game:

Rock Paper Scissors

Rock 0, 0 1, 1 1, 1
Paper 1, 1 0, 0 1, 1
Scissors 1, 1 1, 1 0, 0

We first check that there is no Nash equilibrium in which one of the players plays a pure
strategy. For definiteness, let’s suppose player 1 (row player) plays Rock. Then player 2’s
best response must be Paper, which in turn implies that player 1 must have played Scissors—a
contradiction.
We now rule out the case when one of players mixes only two actions. Again, for definiteness,
let us suppose that player 1 mixes Rock and Paper non-trivially in the Nash equilibrium strategy
profile. Since Paper always does better than Rock against a mixed Rock and Paper strategy,
player 2’s response cannot involve Rock. Hence player 2’s response must mix Paper and Scissors
non-trivially (as we already have ruled out pure strategies being part of a Nash equilibrium).

16
But then, by the same reasoning, player 1 must have mixed Scissors and Rock, leading to a
contradiction.
We conclude that Nash equilibrium strategy profiles must involve strategies that mix all three
actions (non-trivially). Suppose now player 1 plays Rock with probability p1 , Paper with proba-
bility p2 and Scissors with probability p3 . Since player 2’s best response mixes all three actions,
they must all do equally well against (p1 , p2 , p3 ). This is the Indi↵erence Principle! Now
calculate player 2’s expected payo↵ for each pure strategy against p = (p1 , p2 , p3 ): we have

u2 (p, Rock) = p 2 + p3 , u2 (p, Paper) = p1 p3 , u2 (p, Scissors) = p1 + p2 .

As these returns are equal, we obtain p2 + p3 = p1 p3 = p1 + p2 , which together with

p1 + p2 + p3 = 1 yields the mixed strategy (1/3, 1/3, 1/3). Similarly, player 2’s strategy is
(1/3, 1/3, 1/3).

Example 4.10. Alice and Bob are playing tennis. Bob is at the net and Alice needs to decide
to hit the ball to Bob’s right or left. Bob needs to anticipate Alice’s passing shot and has to
commit to jumping towards his right or left; Bob does not enough time to react after Alice has
played her shot. We assume Alice and Bob are playing the following game

left right
left 0.5, 0.5 0.8, 0.2
right 0.9, 0.1 0.2, 0.8

where the entries denote the probabilities of Alice and Bob winning the game from their chosen
actions. Notice that there are no dominant strategies, nor are there Nash equilibria in pure
strategies.
When is ((p, 1 p), (q, 1 q)) a Nash equilibrium? From the indi↵erence principle, we
obtain 0.5p + 0.1(1 p) = 0.2p + 0.8(1 p), which gives p = 0.7. Likewise, we must have
0.5q + 0.8(1 q) = 0.9q + 0.2(1 q), which gives q = 0.6.

Example 4.11 (The attrition game). Consider the game on the

Stay Leave
right. There are two pure-strategy Nash equilibria: (Stay, Leave)
Stay 2, 2 1, 0
and (Leave, Stay). There is also a mixed-strategy Nash equilibrium:
Leave 0, 1 0, 0
((1/3, 2/3), (1/3, 2/3)).
What happens when we change the payo↵s? We consider the following versions where x>1
in game (A) and 0 < y < 1 in game (B).

Stay Leave Stay Leave

(A) Stay 2, 2 x, 0 (B) Stay 2, 2 1, 0
Leave 0, 1 0, 0 Leave y, 1 y, 0

Pure-strategy Nash equilibria are preserved in both games. In game (A), the mixed-strategy
Nash equilibrium is now: ((1/3, 2/3), (x/(2 + x), 2/(2 + x))). In game (B), there is a mixed

17
strategy Nash equilibrium when player 1 stays with probability 1/3 and player 2 stays with
probability (1 y)/3. For varying y, player 1’s mixed strategy remains the same but player 2’s
mixed strategy changes (probability of player 2 choosing to stay decreases as y increases).

Remark 4.12 (Dominating mixed strategies). Even when no action strictly dominates an-
other, a mixed strategy might, as the following example shows.

Example 4.13 (A game with a dominating mixed strategy). Consider the game

L R
U 3, 1 0, 2
M 1, 1 1, 0
D 0, 0 3, 1

None of the actions is dominated. Now suppose player 1 plays U and D with probability 1/2
each. The expected payo↵ when player 2 chooses L or R is 3/2 in both cases. Therefore the
mixed strategy of playing U and D with probability 1/2 each dominates M. After eliminating
M, Bob’s R dominates L, and we obtain a solution (D, R) for this game.

The indi↵erence principle (Theorem 4.8) generalizes to any number of players: in a Nash
equilibrium, all actions played by a player with positive probability must do equally well against
the strategies played by the other players.

Example 4.14 (The indi↵erence principle for n > 2 players). In a remote village with n > 1
inhabitants there is nothing to do in the evening but to meditate on the top of a nearby mountain.
Meditation yields a utility of 10 units, but only one person can comfortably meditate at the top
of the mountain: if two or more people climb the mountain, they get no utility. Climbing the
mountain is difficult and it costs 3 units of utility. Staying at home yields no utility. Find a
symmetric mixed-strategy Nash equilibrium of this game, i.e., find a mixed strategy which, if
played by all players, yields a Nash equilibrium.
Consider a mixed strategy consisting of climbing the mountain with probability p. By the
principle of indi↵erence, everyone following this strategy gives a Nash equilibrium precisely when
each player is indi↵erent between climbing and not climbing, given that everyone else climbs with
probability p. In this situation, the expected utility of climbing is 7(1 p)n 1 3(1 (1 p)n 1 )
p
and the expected utility of not climbing is 0. An equality between these yields p = 1 n 1 3/10.

Existence of Nash equilibria. We have already seen examples of games without any Nash
equilibrium in pure strategy (e.g., matching pennies, rock-paper-scissors). Here are couple more
examples of games without any Nash equilibria.
(1) Two players choose an integer. The player with the biggest integer gets 1, the other 0,
and in case of a tie both get 0.
(2) n > 1 players choose a number in the open interval (0, 1) and the mean µ of these
numbers is computed. The players closest to µ/2 win, the others lose.

18
 The fundamental result at the heart of Game Theory, due to John Nash and for which Nash
got the Nobel Prize in Economics, is the following

Theorem 4.15 (Nash’s Theorem). Any game (S1 , . . . , Sn , u1 , . . . , un ) with finite strategy sets
has at least one Nash equilibrium in mixed strategies.

The proof of this theorem is quite intricate! (See, for example, Chapter 8 in Ken Binmore’s
Playing For Real.)

4.2 Nash Bargaining

In this section we do something new: rather than finding equlibria of games, we look for equitable
outcomes. A standard example to keep in mind is pay negotiation: The employer might secretly
admit that their employees are worth paying 5,000 pounds more; employees want at least 2,000
pounds. How much should employees ask for? What would be a fair outcome?

Convex sets. We begin by discussing some geometric notions in Euclidean spaces. To under-
stand the definitions, just work in the plane R2 or usual 3-dimensional space R3 .

Definition 4.16 (Convex sets). A subset S of Rd is convex if for any v, w 2 S, the vectors
tv + (1 t)w 2 S for all 0  t  1.

Remark 4.17. Given vectors v, w 2 Rd , the set of vectors tv + (1 t)w for 0  t  1 is nothing
more than the line segment from w to v. So a convex set is one that contains all line segments
with end points in the given set.

Convex sets in R2 and R3 can be visualised read. A solid sphere is convex; a doughnut is not
convex. Other examples: vector-subspaces of Rd ; cones (non-negative linear combinations) in
Rd . In general, it is simplest to use the definition to check convexity. For example:

Proposition 4.18. The intersection of convex sets (in a fixed Euclidean space) is also convex.

Proof. Consider two points u and v in the intersection. Then u and v are in each of the given
sets. As these sets are convex, the line segment {tu + (1 t)v | 0  t  1} is contained in
each of them. Therefore {tu + (1 t)v | 0  t  1} is in the intersection as well.

Proposition 4.19. Given n points v1 , . . . , vn 2 Rd , let C(v1 , . . . , vn ) be the set of convex

combinations of a1 , . . . , an given by
( )
def 0  1 , . . . , n  1 satisfies
C(v1 , . . . , vn ) = 1 v1 + · · · + n vn .
1 + ··· + n = 1

Then C(v1 , . . . , vn ) is a convex set. Furthermore, any convex set in Rd containing v1 , . . . , vn

contains the set C(v1 , . . . , vn ).

19
Proof. Write C for C(v1 , . . . , vn ). To prove that C is convex, consider two points u, v 2 C
given by
8
) >
u = p1 v 1 + · · · + pn v n , < 0  p1 , . . . , pn , q1 , . . . , qn  1,
where p1 + · · · + pn = 1,
v = q1 v 1 + · · · + qn v n >
:
q1 + · · · + qn = 1.

Take 0  t  1 and consider

tu + (1 t)v = (tp1 + (1 t)q1 )v1 + · · · + (tpn + (1 t)qn )vn .

Now each coefficient tpi + (1 t)qi is between 0 and 1, and

(tp1 + (1 t)q1 )v1 + · · · + (tpn + (1 t)qn ) = t(p1 + · · · + pn ) + (1 t)(q1 + · · · + qn )

= t + (1 t) = 1.

Thus tu + (1 t)v 2 C, and therefore C is convex.

Now let A be a convex set containing v1 , . . . , vn , and we want to show C ⇢ A. Consider a
convex combination v = p1 v1 + · · · + pn vn 2 C. If p2 = · · · = pn = 0, then v 2 A. So assume
that p2 + · · · + pn is a positive real number. We then have
⇣ p2 pn ⌘
v = p1 v1 + (p2 + · · · + pn ) v2 + · · · + vn .
p 2 + · · · + pn p2 + · · · + p n
p2 pn
We can inductively assume p2 +···+pn v2 + ··· + p2 +···+pn vn 2 A, and then use convexity of A
to conclude v 2 A.

Definition 4.20. The convex hull of A ✓ Rd is the smallest convex set containing A.

In other words, the convex hull is a convex set containing A and contained in every convex
set containing A. Existence follows from Proposition 4.18: take the intersection of all convex
sets containing A. (Note that there is at least one convex set containing A: the whole space
Rd .)
A constructive description of the convex hull is given by the following proposition.

Proposition 4.21. The convex hull of a set A ✓ Rd is the subset of Rd consisting of all convex
combinations of points in A: that is, points of the form 1 a1 + · · · + k ak where k 1,
1 , . . . , k are non-negative real numbers, 1 + · · · + k = 1 and a1 , . . . , ak 2 A.

The above proposition follows easily from Proposition 4.19; details are left to the reader.

Remark 4.22. Here’s one way of visualising convex hulls in the plane. Let A ⇢ R2 , and assume
that you can wrap a piece of string around A (i.e., assume that it is bounded). Now tighten
the string until no slack is left to obtain the convex hull.

20
Feasible expected utilities of a game. We now transfer the preceding geometric notions
to the setting of mixed strategies. We will restrict our discussion to 2-player games with finite
actions; the general n-player case is not that much di↵erent (but requires some advanced analysis
in higher dimensional spaces).

Definition 4.23 (Cooperative payo↵ region). Let G = (S, T, u1 , u2 ) be a game with finite
action sets. The cooperative payo↵ region of G is the convex hull of {(u1 (s, t), u2 (s, t)) | s 2
S, t 2 T } ✓ R2 . The elements of the cooperative payo↵ region are the feasible outcomes.

Remark 4.24. Informally, the players agree to play out strategy profiles (s, t) 2 S ⇥T randomly
with preassigned probabilities. The expected payo↵s then are represented by points in the convex
hull of {(u1 (s, t), u2 (s, t)) | s 2 S, t 2 T } ✓ R2 . In other words, points in the cooperative payo↵
region represent expected utilities that the players can obtain by agreeing on a deal (involving
choosing random numbers).

Example 4.25. Suppose Alice and Bob play the following game in tabular form

L M R
U 2, 1 2, 1 1, 1
D 1, 2 0, 2 1, 2

The cooperative payo↵ region is the convex hull of (2, 1), ( 2, 1), (1, 1), ( 1, 2), (0, 2), (1, 2)
in R2
To see how this might arise, imagine that we as arbitrators could force each outcome of the
game to occur with a given probability: we choose real numbers 0  p1 , p2 , p3 , p4 , p5 , p6  1
such that p1 + p2 + p3 + p4 + p5 + p6 = 1, and impose that scenarios UL, UM, UR, DL, DM,
DR happen with probabilites p1 , p2 , p3 , p4 , p5 , p6 , respectively. By agreeing to this arbitrations,
Alice and Bob’s expected payo↵ becomes

p1 (2, 1) + p2 ( 2, 1) + p3 (1, 1) + p4 ( 1, 2) + p5 (0, 2) + p6 (1, 2).

Whether Alice and Bob will agree to this arrangement will depend on how willing they are to
compromise. More pertinently, they might have red lines. Can the arbitrators now come up
with a better arrangement, still within rules that apply to all disputes?

Axiomatic set up Nash Bargaining. As indicated above, the arbitrator’s task is to suggest
a feasible outcome acceptable to both players every time a game is brought to arbitration. We
begin with the following definition

Definition 4.26 (Nash bargaining problem). A Nash bargaining problem is a pair (P, (a, b))
where P ⇢ R2 is a convex polygon and (a, b) 2 R2 is in P (the disagreement point).

Remark 4.27. A polygon includes its interior points as well as the points on the boundary
edges.

21
Remark 4.28. In the bargaining problem (P, (a, b)), the convex polygon corresponds to the
cooperative payo↵ region of a game. The distinguished feasible outcome or disagreement point
(a, b) would be one which both players come up with initially and can settle on or walk away if
nothing ‘better’ can be found.

The arbitration procedure must now generate a feasible outcome to every bargaining problem:
Given a bargaining problem (P, (a, b)), associate a feasible outcome (P, (a, b)) = (a⇤ , b⇤ ) 2 P
satisfying the following axioms.
(A1) Individual Rationality: a⇤ a and b⇤ b.
(A2) Pareto Optimality: If (a0 , b0 ) 2 P satisfies a0 a ⇤ , b0 b⇤ simultaneously, then (a0 , b0 ) =
(a⇤ , b⇤ ).
(A3) Independence of Irrelevant Alternatives: If P 0 ✓ P is a convex polygon, and contains
both (a, b) and (a⇤ , b⇤ ), then ((P 0 , (a, b))) = (a⇤ , b⇤ ).
(A4) Invariance Under Affine Linear Transformations: If P 0 = {(↵x + , y + | (x, y) 2 P }
for some real constants ↵, , , with ↵ > 0, > 0, then (P 0 , (↵a + , b + )) =
(↵a⇤ + , b⇤ + ).
(A5) Symmetry: If P is symmetric (i.e., (x, y) = P ) (y, x) 2 P ) and a = b, then a⇤ = b⇤ .

We then have the following theorem.

Theorem 4.29 (Nash). There exists a unique arbitration procedure which satisfies all five
conditions above.
The value (a⇤ , b⇤ ) = (P, (a, b)) is often referred to as a Nash Bargain.

The proof will proceed in two parts: we will first construct an arbitration procedure explicitly,
and then show that there is only one solution.

Note 4.30. Let (P, (a, b)) be a bargaining problem with solution (a⇤ , b⇤ ). Then individual ra-
tionality and Pareto optimality (axioms (A1), (A2)) show that (a⇤ , b⇤ ) must lie on the boundary
of the polygon P and satisfies a⇤ a, b⇤ b.

We begin our construction of a candidate Nash bargain with the following lemma.

Lemma 4.31. Let P ⇢ R2 be a convex polygon with a distinguished fixed point (a, b) 2 P ,
and let K = {(x, y) 2 P | x a, y b}. Consider

f :K!R defined by f (x, y) = (x a)(y b).

The function f then attains its maximum on the boundary of K. Moreover, the maximum is
unique if K is a non-degenerate polygon (which is equivalent to the condition that there exists
(a0 , b0 ) 2 K satisfying a0 > a, b0 > b).

Proof. If we fix y b, then f (x, y) = (x a)(y b) is increasing for x a. From this, it follows
that the maximal values of f , if it exists, must occur on the boundary of K. But the boundary

22
 Figure 1: Nash bargaining problem (P, (a, b)). The Nash
bargain will be located in the blue boundary.

is made up of straight line segments, and f is a quadratic function in a single parameter in each
segment. Thus f will have a maximum in each boundary segment, and because there are only
finitely many segments, we can take the largest of these maximums. This proves the first part
of the lemma.
We now prove uniqueness. By a simple change of coordinates, we can assume that (a, b) =
(0, 0). Thus f (x, y) = xy. We are assuming that K = {(x, y) 2 P | x 0, y 0} contains a
point with (strictly) positive coordinates. Thus the maximum value of xy on K, call it M , will
be positive real number.
Now suppose there are distinct points (a1 , b1 ), (a2 , b2 ) on which the maximum is attained.
Let (a3 , b3 ) = (1/2)(a1 , b1 ) + (1/2)(a2 , b2 ) be the mid-point of the line segment joining (a1 , b1 )
and (a2 , b2 ). This midpoint is in K since K is convex. Using the identity

(a1 + a2 )(b1 + b2 ) + (a1 a2 )(b1 b2 ) = 2a1 b1 + 2a2 b2

together with a1 b1 = a2 b2 = M , we obtain

f (a3 , b3 ) + 14 (a1 a2 )(b1 b2 ) = M.

But then a1 b1 = a2 b2 = M > 0 implies that either (a1 > a2 and b1 < b2 ) or (a1 < a2 and
b1 > b2 ).In either case (a1 a2 )(b1 b2 ) < 0, and we obtain f (a3 , b3 ) > M contradicting the
maximality of M .

Proof of Nash’s Theorem: existence. We construct explicitly as follows. Let (P, (a, b))
be given bargaining problem, set K = P \ {(x, y) 2 R2 | x a, y b}, and let f (x, y) =
(x a)(y b).
Case 1. K has non-empty interior: that is, we can find (a0 , b0 ) 2 K satisfying a0 > a, b0 > b.
In this case, we use Lemma 4.31 to find unique (a⇤ , b⇤ ) 2 K on which f attains its maximum,

23
and declare (P, (a, b)) = (a⇤ , b⇤ ).
Case 2a. K is a vertical line segment: that is, (a, y) 2 P for some y > b. Let b⇤ be the
largest b for which this occurs, and declare (P, (a0 , b0 )) = (a0 , b⇤ ).
Case 2b. K is a horizontal line segment: that is, (x, b) 2 P for some x > a. Let a⇤ be the
largest a for which this occurs, and declare (P, (a0 , b0 )) = (a⇤ , b0 ).
Case 3. K = {(a, b)}. Declare (P, (a0 , b0 )) = (a0 , b0 ).
The verification that as constructed has the desired properties is quite straightforward and
follows from the construction itself. (But do check that the axioms hold!)

Proof of Nash’s Theorem: uniqueness. To prove uniqueness assume that there is a di↵erent
procedure b with the same properties, and choose a bargaining problem P with disagreement
point (a, b) 2 P such that (P, (a, b)) 6= b (P, (a, b)). We can take (a, b) = (0, 0) (by using
using the linear transformation (x, y) ! (x a, y b) and axiom (A4), if necessary).
If K = {(x, y) 2 P |x 0, y 0} has empty interior, then K is either a horizontal or
vertical line segment or singleton. If K is a horizontal line, then (P, (0, 0)) = (a⇤ , 0) where
a⇤ is thex-coordinate of the right end K. Individual rationality and Pareto optimality then
implies b (P, (0, 0)) = (a⇤ , 0). The case when K is a vertical line is similar, and the case when
K = {(0, 0)} is trivial.
So from here one, let us assume that K has non-empty interior, and let (P, (0, 0)) = (a⇤ , b⇤ ).
Now, both a⇤ , b⇤ are positive real numbers by construction. By applying the linear function
L(x, y) = (x/a⇤ , y/b⇤ ), and relabelling L(P ) as P if needed, we can use the invariance under
affine transformations axiom to assume (P, (0, 0)) = (1, 1) and b (P, (0, 0)) 6= (1, 1).
We claim that any point (u, v) 2 P satisfies u + v  2. Suppose otherwise, and define
h(t) = f (t(u, v) + (1 t)(1, 1)) = (tu + (1 t))(tv + (1 t)). We have h(0) = 1 and
dh/dt(0) = u + v 2 > 0 by assumption. So h is increasing around t = 0 and we can find a
small t > 0 such that f (t(u, v) + (1 t)(1, 1)) > 1 = f (1, 1) and t(u, v) + (1 t)(1, 1) having
both coordinates positive (so in K), a contradiction.
Let Pb be the symmetric convex hull of P : that is, Pb is the convex hull of P [ {(v, u) | (u, v) 2
P }. We also have u + v  2 for (u, v) 2 Pb. Hence if (a, a) 2 Pb, then a  1, and we deduce
that b (Pb , (0, 0)) = (1, 1). Finally, the independence of irrelevant alternatives axiom the implies
that (P 0 , (0, 0)) = (1, 1), a contradiction.

Example 4.32. Suppose Alice and Bob bargain over the following game in tabular form

L M R
U 2, 1 2, 1 1, 1
D 1, 2 0, 2 1, 2

and if they fail to reach agreement they both get 0. The corresponding bargaining problem is
(P, (0, 0)) where P is the convex hull of (2, 1), ( 2, 1), (1, 1), ( 1, 2), (0, 2), (1, 2) in R2 .

24
 Figure 2: Example 4.32: The Nash bargain for (P, (0, 0))
will be located in the red boundary.

The Nash bargain will be somewhere on the red boundary above. This red boundary is the
union of two line segments given parametrically by t(0, 2) + (1 t)(1, 1) = (1 t, t + 1)
and t(1, 1) + (1 t)(3/2, 0) = ( t/2 + 3/2, t) with 0  t  1. Now f (1 t, t + 1) =
(1 t)(t + 1) attains its maximum value 1 at t = 0, and this corresponds to the point (1, 1);
f ( t/2 + 3/2, t) = (3 t)t/2 restricted to 0  t  1 attains its maximum value 1 at t = 1,
and this also corresponds to the point (1, 1). We conclude that the Nash bargain results in
expected utility 1 for both Alice and Bob.
Let us now consider what happens when we modify the example so that failure to agree results
in a payo↵ ( 2/5, 1). The new bargaining problem is (P, ( 2/5, 1). We then need to maximise
f (x, y) = (x + 2/5)(y 1) for (x, y) 2 P satisfying x 2/5, y 1. The possible bargain
must lie on the line segment joining ( 2/5, 2) to (0, 2), or the line segment joining (0, 2) to
(1, 1); see the red boundary in Figure 3.
Points on the line joining ( 2/5, 2) to (0, 2) are given by (x, 2) with 2/5  x  0. So
f (x, y) = x + 2/5 is maximised at the point (0, 2) with maximum 2/5 on this segment.
On the line joining (0, 2) to (1, 1), we need to maximise

f (1 t, t + 1) = (1 t + 2/5)(t + 1 1) = (7/5 t)t

for 0  t  1. We find that (7/5 t)t attains its maximum value of 49/100 at t = 7/10.
Finally, since 0.49 > 2/5, we conclude that the maximum value of f (x, y) we are looking for
is attained at (3/10, 17/10). Hence the Nash bargain results in expected utility 3/10 for Alice
and 17/10 for Bob.

25
 Figure 3: Example 4.32: The Nash bargain for
(P, ( 2/5, 1)) will be located in the red boundary.

5 Sequential games

5.1 Games with perfect information

In a sequential game, players take turns to make a move (e.g., chess, bidding at an open auction,
etc.). The game has perfect information if players know the history of the game at any point.
Equivalently, every end point of the game can only be reached by one specific sequence of
moves.
What this translates to is that sequential games with perfect information can be represented
by directed rooted trees (often referred to as extensive form of the game). Recall that a directed
rooted tree is a graph consisting of vertices (nodes) and directed edges (arrows) joining vertices
with the following properties:
• There is a unique distinguished vertex, called the root, which has only edges going out
and no edges going in.
• Every other vertex can by reached from the root in precisely one way by following a
sequence of directed edges.
Vertices with no arrows going out are called leaves or terminal nodes.

Root

Leaf

(A directed rooted tree) (Not a directed rooted tree)

26