Game Theory Lecture Math486
Game Theory Lecture Math486
Notes
Version 1.1.1
Christopher Griffin
« 2010-2012
Licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 United States License
With Major Contributions By:
James Fan
George Kesidis
List of Figures v
Chapter 1. Preface and an Introduction to Game Theory xi
1. Using These Notes xi
2. An Overview of Game Theory xi
Chapter 2. Probability Theory and Games Against the House 1
1. Probability 1
2. Random Variables and Expected Values 6
3. Conditional Probability 7
4. Bayes Rule 12
Chapter 3. Utility Theory 15
1. Decision Making Under Certainty 15
2. Advanced Decision Making under Uncertainty 22
Chapter 4. Game Trees, Extensive Form, Normal Form and Strategic Form 25
1. Graphs and Trees 25
2. Game Trees with Complete Information and No Chance 28
3. Game Trees with Incomplete Information 32
4. Games of Chance 34
5. Pay-off Functions and Equilibria 36
Chapter 5. Normal and Strategic Form Games and Matrices 47
1. Normal and Strategic Form 47
2. Strategic Form Games 48
3. Review of Basic Matrix Properties 50
4. Special Matrices and Vectors 52
5. Strategy Vectors and Matrix Games 53
Chapter 6. Saddle Points, Mixed Strategies and the Minimax Theorem 55
1. Saddle Points 55
2. Zero-Sum Games without Saddle Points 58
3. Mixed Strategies 60
4. Mixed Strategies in Matrix Games 63
5. Dominated Strategies and Nash Equilibria 64
6. The Minimax Theorem 69
7. Finding Nash Equilibria in Simple Games 74
8. A Note on Nash Equilibria in General 76
Chapter 7. An Introduction to Optimization and the Karush-Kuhn-Tucker Conditions 79
iii
iv CONTENTS
1.1 There are several sub-disciplines within Game Theory. Each one has its own
unique sets of problems and applications. We will study Classical Game Theory,
which focuses on questions like, “What is my best decision in a given economic
scenario, where a reward function provides a way for me to understand how my
decision will impact my result.” We may also investigate Combinatorial Game
Theory, which is interested in games like Chess or Go. If there’s time, we’ll
study Evolutionary Game Theory, which is interesting in its own right. xiii
2.1 The Monty Hall Problem is a multi-stage decision problem whose solution
relies on conditional probability. The stages of decision making are shown in
the diagram. We assume that the prizes are randomly assigned to the doors.
We can’t see this step–so we’ve adorned this decision with a square box. We’ll
discuss these boxes more when we talk about game trees. You the player must
first choose a door. Lastly, you must decide whether or not to switch doors
having been shown a door that is incorrect. 10
4.1 Digraphs on 3 Vertices: There are 64 = 26 distinct graphs on three vertices. The
increased number of edges graphs is caused by the fact that the edges are now
directed. 26
4.2 Two Paths: We illustrate two paths in a digraph on three vertices. 26
4.3 Directed Tree: We illustrate a directed tree. Every directed tree has a unique
vertex called the root. The root is connected by a directed path to every other
vertex in the directed tree. 27
4.4 Sub Tree: We illustrate a sub-tree. This tree is the collection of all nodes that
are descended from a vertex u. 28
4.5 Rock-Paper-Scissors with Perfect Information: Player 1 moves first and holds up
a symbol for either rock, paper or scissors. This is illustrated by the three edges
leaving the root node, which is assigned to Player 1. Player 2 then holds up a
symbol for either rock, paper or scissors. Payoffs are assigned to Player 1 and 2
at terminal nodes. The index of the payoff vector corresponds to the players. 29
4.6 New Guinea is located in the south pacific and was a major region of contention
during World War II. The northern half was controlled by Japan through 1943,
while the southern half was controlled by the Allies. (Image created from
Wikipedia (http://en.wikipedia.org/wiki/File:LocationNewGuinea.svg),
originally sourced from http://commons.wikimedia.org/wiki/File:
LocationPapuaNewGuinea.svg. 30
v
vi LIST OF FIGURES
4.7 The game tree for the Battle of the Bismark Sea. The Japanese could choose to
sail either north or south of New Britain. The Americans (Allies) could choose
to concentrate their search efforts on either the northern or southern routes.
Given this game tree, the Americans would always choose to search the North
if they knew the Japanese had chosen to sail on the north side of New Britain;
alternatively, they would search the south route, if they knew the Japanese had
taken that. Assuming the Americans have perfect intelligence, the Japanese
would always choose to sail the northern route as in this instance they would
expose themselves to only 2 days of bombing as opposed to 3 with the southern
route. 31
4.8 Simple tic-tac-toe: Players in this case try to get two in a row. 32
4.9 The game tree for the Battle of the Bismark Sea with incomplete information.
Obviously Kenney could not have known a priori which path the Japanese
would choose to sail. He could have reasoned (as they might) that there best
plan was to sail north, but he wouldn’t really know. We can capture this fact by
showing that when Kenney chooses his move, he cannot distinguish between the
two intermediate nodes that belong to the Allies. 34
4.10 Poker: The root node of the game tree is controlled by Nature. At this node, a
single random card is dealt to Player 1. Player 1 can then decide whether to end
the game by folding (and thus receiving a payoff or not) or continuing the game
by raising. At this point, Player 2 can then decide whether to call or fold, thus
potentially receiving a payoff. 36
4.11 Reduced Red Black Poker: We are told that Player 1 receives a red card. The
resulting game tree is substantially simpler. Because the information set on
Player 2 controlled nodes indicated a lack of knowledge of Player 1’s card, we
can see that this sub-game is now a complete information game. 37
4.12 A unique path through the game tree of the Battle of the Bismark Sea.
Since each player determines a priori the unique edge he/she will select when
confronted with a specific information set, a path through the tree can be
determined from these selections. 38
4.13 The probability space constructed from fixed player strategies in a game of
chance. The strategy space is constructed from the unique choices determined
by the strategy of the players and the independent random events that are
determined by the chance moves. 40
4.14 The probability space constructed from fixed player strategies in a game of
chance. The strategy space is constructed from the unique choices determined
by the strategy of the players and the independent random events that are
determined by the chance moves. Note in this example that constructing the
probabilities of the various events requires multiplying the probabilities of the
chance moves in each path. 41
4.15 Game tree paths derived from the Simple Poker Game as a result of the strategy
(Fold, Fold). The probability of each of these paths is 1/2. 42
4.16 The game tree for the Battle of the Bismark Sea. If the Japanese sail north, the
best move for the Allies is to search north. If the Japanese sail south, then the
LIST OF FIGURES vii
best move for the Allies is to search south. The Japanese, observing the payoffs,
note that given these best strategies for the Allies, there best course of action is
to sail North. 45
5.1 In Chicken, two cars drive toward one another. The player who swerves first
loses 1 point, the other player wins 1 point. If both players swerve, then each
receives 0 points. If neither player swerves, a very bad crash occurs and both
players lose 10 points. 49
5.2 A three dimensional array is like a matrix with an extra dimension. They are
difficult to capture on a page. The elements of the array for Player i store the
various payoffs for Player i under different strategy combinations of the different
players. If there are three players, then there will be three different arrays. 50
6.1 The minimax analysis of the game of competing networks. The row player knows
that Player 2 (the column player) is trying to maximize her [Player 2’s] payoff.
Thus, Player 1 asks: “What is the worst possible outcome I could see if I played
a strategy corresponding to this row?” Having obtained these worst possible
scenarios he chooses the row with the highest value. Player 2 does something
similar in columns. 56
6.2 In August 1944, the allies broke out of their beachhead at Avranches and started
heading in toward the mainland of France. At this time, General Bradley was in
command of the Allied forces. He faced General von Kluge of the German ninth
army. Each commander faced several troop movement choices. These choices
can be modeled as a game. 59
6.3 At the battle of Avranches General Bradley and General von Kluge faced off
over the advancing Allied Army. Each had decisions to make. This game matrix
shows that this game has no saddle point solution. There is no position in the
matrix where an element is simultaneously the maximum value in its column
and the minimum value in its row. 59
6.4 When von Kluge chooses to retreat, Bradley can benefit by playing a strategy
different from his maximin strategy and he moves east. When Bradley does this,
von Kluge realizes he could benefit by attacking and not playing his maximin
strategy. Bradley realizes this and realizes he should play his maximin strategy
and wait. This causes von Kluge to realize that he should retreat, causing this
cycle to repeat. 60
6.5 The payoff matrix for Player P1 in Rock-Paper-Scissors. This payoff matrix can
be derived from Figure 4.5. 61
6.6 In three dimensional space ∆3 is the face of a tetrahedron. In four dimensional
space, it would be a tetrahedron, which would itself be the face of a four
dimensional object. 62
6.7 To show that Confess dominates over Don’t Confess in Prisoner’s dilemma for
Bonnie, we can compute e1 T Az and e2 Az for any arbitrary mixed strategy z for
Clyde. The resulting payoff to Bonnie is 5z − 5 when she confesses and 9z − 10
when she doesn’t confess. Here z is the probability that Clyde will not confess.
viii LIST OF FIGURES
7.1 Goat pen with unknown side lengths. The objective is to identify the values of
x and y that maximize the area of the pen (and thus the number of goats that
can be kept). 79
7.2 Plot with Level Sets Projected on the Graph of z. The level sets existing in R2
while the graph of z existing R3 . The level sets have been projected onto their
appropriate heights on the graph. 83
7.3 Contour Plot of z = x2 + y 2 . The circles in R2 are the level sets of the function.
The lighter the circle hue, the higher the value of c that defines the level set. 83
7.4 A Line Function: The points in the graph shown in this figure are in the set
produced using the expression x0 + vt where x0 = (2, 1) and let v = (2, 2). 84
7.5 A Level Curve Plot with Gradient Vector: We’ve scaled the gradient vector
in this case to make the picture understandable. Note that the gradient is
perpendicular to the level set curve at the point (1, 1), where the gradient was
evaluated. You can also note that the gradient is pointing in the direction of
steepest ascent of z(x, y). 86
7.6 Level Curves and Feasible Region: At optimality the level curve of the objective
function is tangent to the binding constraints. 87
7.7 Gradients of the Binding Constraint and Objective: At optimality the gradient
of the binding constraints and the objective function are scaled versions of each
other. 87
7.8 Examples of Convex Sets: The set on the left (an ellipse and its interior) is
a convex set; every pair of points inside the ellipse can be connected by a line
contained entirely in the ellipse. The set on the right is clearly not convex as
we’ve illustrated two points whose connecting line is not contained inside the
set. 89
7.9 A convex function: A convex function satisfies the expression f (λx1 +(1−λ)x2 ) ≤
λf (x1 ) + (1 − λ)f (x2 ) for all x1 and x2 and λ ∈ [0, 1]. 89
8.1 Feasible Region and Level Curves of the Objective Function: The shaded region
in the plot is the feasible region and represents the intersection of the five
LIST OF FIGURES ix
10.1 The three plots shown the competitive payoff region, cooperative payoff region
and and overlay of the regions for the Battle of the Sexes game. Note that the
cooperative payoff region completely contains the competitive payoff region. 129
10.2 The Pareto Optimal, Nash Bargaining Solution, to the Battle of the Sexes is for
each player to do what makes them happiest 50% of the time. This seems like
the basis for a fairly happy marriage, and it yields a Pareto optimal solution,
shown by the green dot. 139
10.3 Matlab input for solving Nash’s bargaining problem with the Battle of the Sexes
problem. Note that we are solving a maximization problem, but Matlab solve
mnimization problems by default. Thus we change the sign on the objective
matrices. 140
CHAPTER 1
a car, your decisions and the decisions of those around you will affect the quality of the end
result for everyone.
Game Theory is a broad discipline within Applied Mathematics that influences and is
itself influenced by Operations Research, Economics, Control Theory, Computer Science,
Psychology, Biology and Sociology (to name a few disciplines). If you want to start a fight
in bar with a Game Theorist (or an Economist) you might say that Game Theory can be
broadly classified into four main sub-categories of study:
(1) Classical Game Theory: Focuses on optimal play in situations where one or more
people must make a decision and the impact of that decision and the decisions of
those involved is known. Decisions may be made by use of a randomizing device (like
flipping a coin). Classical Game Theory has helped people understand everything
from the commanders in military engagements to the behavior of the car salesman
during negotiations. See [vNM04, LR89, Mor94, Mye01, Dre81, PR71] and
Chapter 1 of [Wei97] or [Bra04] for extensive details on this sub-discipline of Game
Theory.
(2) Combinatorial Game Theory: Focuses on optimal play in two-player games in which
each player takes turns changing in pre-defined ways. Combinatorial Game Theory
does not consider games with chance (no randomness). Combinatorial Game Theory
is used to investigate games like Chess, Checkers or Go. Of all branches, Combina-
torial Game Theory is the least directly related to real life scenarios. See[Con76]
and [BCG01a, BCG01b, BCG01c, BCG01d], which are widely regarded as the
bible of Combinatorial Game Theory.
(3) Dynamic Game Theory: Focuses on the analysis of games in which players must
make decisions over time and in which those decisions will affect the outcome at the
next moment in time. Dynamic Game Theory often relies on differential equations to
model the behavior of players over time. Dynamic Game Theory can help optimize
the behavior of unmanned vehicles or it can help you capture your baby sister who
has escaped from her playpen. See [DJLS00, BO82] for a survey on dynamic
games. The latter reference is extremely technical.
(4) Other Topics in Game Theory: Game Theory, as noted, is broad. This category
captures those topics that are derivative from the three other branches. Examples
include, but are not limited to: (i) Evolutionary Game Theory, which attempts to
model evolution as competition between species, (ii) Dual games in which players
may choose from an infinite number of strategies, but time is not a factor, (iii) Ex-
perimental Game Theory, in which people are studied to determine how accurately
classical game theoretic models truly explain their behavior. See [Wei97, Bra04]
for examples.
Figure 1.1 summarizes the various types of Game Theory.
In these notes, we focus primarily on Classical Game Theory. This work is relatively
young (under 70 years old) and was initiated by Von Neumann and Morgenstern. Major
contributors to this field include Nash (of A Beautiful Mind fame), and several other Nobel
Laureates.
2. AN OVERVIEW OF GAME THEORY xiii
GAME
THEORY
Games with finite Games with time. Games with no Evolutionary Game
numbers of strategies. chance. Theory
Games with motion or
Games with probability a dynamic component. Generally two player Experimental /
(either induced by the strategic games Behavioral Game
player or the game). Examples: played on boards. Theory
Optimal play in a dog
Games with coalitions. fight. Chasing your Moves change the Examples:
brother across a room. structure of a game Evolutionary dynamics
Examples: board. in closed populations,
Poker, Strategic Determining why
Military Decision Examples: altruism is present in
Making, Negotiations. Chess, Checkers, Go, human society.
Nim.
Figure 1.1. There are several sub-disciplines within Game Theory. Each one has its
own unique sets of problems and applications. We will study Classical Game Theory,
which focuses on questions like, “What is my best decision in a given economic
scenario, where a reward function provides a way for me to understand how my
decision will impact my result.” We may also investigate Combinatorial Game
Theory, which is interested in games like Chess or Go. If there’s time, we’ll study
Evolutionary Game Theory, which is interesting in its own right.
CHAPTER 2
1. Probability
Our study of Game Theory starts with a characterization of optimal decision making for
an individual in the absence of any other players. The games we often see on television fall
into this category. TV Game Shows (that do not pit players against each other in knowledge
tests) often require a single player (who is, in a sense, playing against The House) to make
a decision that will affect only his life.
Example 2.1. Congratulations! You have made it to the very final stage of Deal or No
Deal. Two suitcases with money remain in play, one contains $0.01 while the other contains
$1, 000, 000. The banker has offered you a payoff of $499, 999. Do you accept the banker’s
safe offer or do you risk it all to try for $1, 000, 000. Suppose the banker offers you $100, 000
what about $500, 000 or $10, 000?
Example 2.1 may seem contrived, but it has real world implications and most of the
components needed for a serious discussion of decision making under risk. In order to study
these concepts formally, we will need a grounding in probability. Unfortunately, a formal
study of probability requires a heavy dose of Measure Theory, which is well beyond the scope
of an introductory course on Game Theory. Therefore, the following definitions are meant
to be intuitive rather than mathematically rigorous.
Let Ω be a finite set of elements describing the outcome of a chance event (a coin toss,
a roll of the dice etc.). We will call Ω the Sample Space. Each element of Ω is called an
outcome.
Example 2.2. In the case of Example 2.1, the world as we care about it is purely the
position of the $1, 000, 000 and $0.01 within the suitcases. In this case Ω consists of two
possible outcomes: $1, 000, 000 is in suitcase number 1 (while $0.01 is in suitcase number 2)
or $1, 000, 000 is in suitcase number 2 (while $0.01 is in suitcase number 1).
Formally, let us refer to the first outcome as A and the second outcome as B. Then
Ω = {A, B}.
Definition 2.3 (Event). If Ω is a sample space, then an event is any subset of Ω.
Example 2.4. Clearly, the sample space in Example 2.1 consists of precisely four events:
∅ (the empty event), {A}, {B} and {A, B} = Ω. These four sets represent all possible subsets
of the set Ω = {A, B}.
Definition 2.5 (Union). If E, F ⊆ Ω are both events, then E ∪ F is the union of the
sets E and F and consists of all outcomes in either E or F . Event E ∪ F occurs if either
event E or event F occurs.1
1Thanks to Arlan Stutler for finding a typo in this definition. Fixed in versions ≥ 1.1.2.
1
2 2. PROBABILITY THEORY AND GAMES AGAINST THE HOUSE
Example 2.6. Consider the role of a fair six sided dice. The outcomes are 1,. . . ,6. If
E = {1, 3} and F = {2, 4}, then E ∪ F = {1, 2, 3, 4} and will occur as long as we don’t roll
a 5 or 6.
Definition 2.7 (Intersection). If E, F ⊆ Ω are both events, then E∩F is the intersection
of the sets E and F and consists of all outcomes in both E and F . Event E ∩ F occurs if
both even E or event F occur.
Example 2.8. Again, consider the role of a fair six sided dice. The outcomes are 1,. . . ,6.
If E = {1, 2} and F = {2, 4}, then E ∩ F = {2} and will occur only if we roll a 2.
Definition 2.9 (Mutual Exclusivity). Two events E, F ⊆ Ω are said to be mutually
exclusive if and only if E ∩ F = ∅.
Definition 2.10 (Discrete Probability Distribution (Function)). Given discrete sample
space Ω, let F be the set of all events on Ω. A discrete probability function is a mapping
from P : F → [0, 1] with the properties:
(1) P (Ω) = 1
(2) If E, F ∈ F and E ∩ F = ∅, then P (E ∪ F ) = P (E) + P (F )
Remark 2.11 (Power Set). In this definition, we talked about the set F as the set of all
events over a set of outcomes Ω. This is an example of the power set: the set of all subsets
of a set. We sometimes denote this set as 2Ω . Thus, if Ω is a set, then 2Ω is the power set of
Ω or the set of all subsets of Ω.
Definition 2.10 is surprisingly technical and probably does not conform to your ordinary
sense of what probability is. It’s best not to think of probability in this very formal way and
instead to think that a probability function assigns a number to an outcome (or event) that
tells you the chances of it occurring. Put more simply, suppose we could run an experiment
where the result of that experiment will be an outcome in Ω. The the function P simply
tells us the proportion of times we will observe an event E ⊂ Ω if we ran this experiment an
exceedingly large number of times.
Example 2.12. Suppose we could play the Deal or No Deal example over and over again
and observe where the money ends up. A smart game show would mix the money up so that
approximately one-half of the time we observe $1, 000, 000 in suitcase 1 and the other half
the time we observe this money in suitcase 2.
A probability distribution formalizes this notion and might assign 1/2 to event {A} and
1/2 to event {B}. However to obtain a true probability distribution, we must also assign
probabilities to ∅ and {A, B}. In the former case, we know that something must happen!
Therefore, we can assign 0 to event ∅. In the latter case, we know that for certain that either
outcome A or B must occur and so in this case we assign a value of 1.
Example 2.13. In a fair six sided dice, the probability of rolling any value is 1/6.
Formally, Ω = {1, 2, . . . , 6} any role yields is an event with only one element: {ω} where
ω is some value in Ω. If we consider the event E = {1, 2, 3} then P (E) gives us the
probability that we will roll a 1, 2 or 3. Since {1}, {2} and {3} are disjoint sets and
{1, 2, 3} = {1} ∪ {2} ∪ {3}, we know that:
1 1 1 1
P (E) = + + =
6 6 6 2
1. PROBABILITY 3
Definition 2.14 (Discrete Probability Space). The triple (Ω, F, P ) is called a discrete
probability space over Ω.
Lemma 2.15. Let (Ω, F, P ) be a discrete probability space. Then P (∅) = 0.
Proof. The set Ω ∈ F and ∅ ∈ F are disjoint (i.e., Ω ∩ ∅ = ∅). Thus:
P (Ω ∪ ∅) = P (Ω) + P (∅)
We know that Ω ∪ ∅ = Ω. Thus we have:
P (Ω) = P (Ω) + P (∅) =⇒ 1 = 1 + P (∅) =⇒ 0 = P (∅)
Lemma 2.16. Let (Ω, F, P ) be a discrete probability space and let E, F ∈ F. Then:
(2.1) P (E ∪ F ) = P (E) + P (F ) − P (E ∩ F )
Proof. If E ∩ F = ∅ then by definition P (E ∪ F ) = P (E) + P (F ) but P (∅) = 0, so
P (E ∪ F ) = P (E) + P (F ) − P (E ∩ F ).
Suppose E ∩ F 6= ∅. Then let:
E 0 = {ω ∈ E|ω 6∈ F }
F 0 = {ω ∈ F |ω 6∈ E}
Then we know:
(1) E 0 ∩ F 0 = ∅,
(2) E 0 ∩ (E ∩ F ) = ∅,
(3) F 0 ∩ (E ∩ F ) = ∅,
(4) E = E 0 ∪ (E ∩ F ) and
(5) F = F 0 ∪ (E ∩ F ).
Thus, (by inductive extension of the definition of discrete probability function) we know:
(2.2) P (E ∪ F ) = P (E 0 ∪ F 0 ∪ (E ∩ F )) = P (E 0 ) + P (F 0 ) + P (E ∩ F )
We also know that:
(2.3) P (E) = P (E 0 ) + P (E ∩ F ) =⇒ P (E 0 ) = P (E) − P (E ∩ F )
and
(2.4) P (F ) = P (F 0 ) + P (E ∩ F ) =⇒ P (F 0 ) = P (F ) − P (E ∩ F )
Combing these three equations yields:
(2.5) P (E ∪ F ) = P (E) − P (E ∩ F ) + P (F ) − P (E ∩ F ) + P (E ∩ F ) =
P (E ∪ F ) = P (E) + P (F ) − P (E ∩ F )
This completes the proof.
Exercise 1. A fair 4 sided die is rolled. Assume the sample space of interest is the
number appearing on the die and the numbers run from 1 to 4. Identify the space Ω
precisely and all the possible outcomes and events within the space. What is the (logical)
fair probability distribution in this case. [Hint: See Example 2.13.]
4 2. PROBABILITY THEORY AND GAMES AGAINST THE HOUSE
Exercise 2. Prove the following: Let E ⊆ Ω and define E c to be the set of elements of
Ω not in E (this is called the complement of E). Suppose (Ω, F, P ) is a discrete probability
space. Show that P (E c ) = 1 − P (E).
Lemma 2.17. Let (Ω, F, P ) be a discrete probability space and let E, F ∈ F. Then:
(2.6) P (E) = P (E ∩ F ) + P (E ∩ F c )
Exercise 3. Prove Lemma 2.17. [Hint: Show that E ∩ F and E ∩ F c are mutually
exclusive events. Then show that E = (E ∩ F ) ∪ (E ∩ F c ).]
The following lemma is provided without proof. The exercise to prove it is somewhat
challenging.
Lemma 2.18. Let (Ω, F, P ) be a probability space and suppose that E, F1 , . . . , Fn are
subsets of Ω. Then:
[n n
[
(2.7) E∩ Fi = (E ∩ Fi )
i=1 i=1
n
[
(2.8) Ω= Fi
i=1
and Fi ∩ Fj = ∅ if i 6= j. Then:
n
X
(2.9) P (E) = P (E ∩ Fi )
i=1
Exercise 5. Suppose that I change the definition of Fi to read: value i appears on either
die, while keeping the definition of event E the same. Do we still have:
6
X
P (E) = P (E ∩ Fi )
i=1
Example 2.27. Let’s play a die rolling game. You put up your own money. Even
numbers lose $10 times the number rolled, while odd numbers win $12 times the number
rolled. What is the expected amount of money you’ll win in this game?
Let Ω = {1, . . . , 6}. Then D = {12, −20, 36, −40, 60, −60}: these are the dollar values
you will win for various rolls of the dice. Then the expected value of X is:
1 1 1 1 1 1
(2.16) E(X) = 12 + (−20) + 36 + (−40) + 60 + (−60) = −2
6 6 6 6 6 6
Would you still want to play this game considering the expected payoff is −$2?
3. CONDITIONAL PROBABILITY 7
3. Conditional Probability
Suppose we are given a discrete probability space (Ω, F, P ) and we are told that an
event E has occurred. We now wish to compute the probability that some other event F
has occurred. This value is called the conditional probability of event F given event E and
is written P (F |E).
Example 2.28. Suppose we roll a fair 6 sided die twice. The sample space in this case is
the set Ω = {(x, y)|x = 1, . . . , 6, y = 1, . . . , 6}. Suppose I roll a 2 on the first try. I want to
know what the probability of rolling a combined score of 8 is. That is, given that I’ve rolled
a 2, I wish to determine the conditional probability of rolling a 6.
Since the die is fair, the probability of rolling any pair of values (x, y) ∈ Ω is equally
likely. There are 36 elements in Ω and so each is assigned a probability of 1/36. That is,
(Ω, F, P ) is defined so that P ((x, y)) = 1/36 for each (x, y) ∈ Ω.
Let E be the event that we roll a 2 on the first try. We wish to assign a new set of
probabilities to the elements of Ω to reflect this information. We know that our final outcome
must have the form (2, y) where y ∈ {1, . . . , 6}. In essence, E becomes our new sample
space. Further, we know that each of these outcomes is equally likely because the die is fair.
Thus, we may assign P ((2, y)|E) = 1/6 for each y ∈ {1, . . . , 6} and P ((x, y)|E) = 0 just in
case x 6= 2, so (x, y) 6∈ E. This last definition occurs because we know that we’ve already
observed a 2 on the first roll, so it’s impossible to see another first number not equal to 2.
At last, we can answer the question we originally posed. The only way to obtain a sum
equal to 8 is to roll a six on the second attempt. Thus, the probability of rolling a combined
score of 8 given a 2 on the first roll is 1/6.
Lemma 2.29. Let (Ω, F, P ) be a discrete probability space and suppose that event E ⊆ Ω.
Then (E, FE , PE ) is a discrete probability space when:
P (F )
(2.17) PE (F ) =
P (E)
for all F ⊆ E and PE (ω) = 0 for any ω 6∈ E.
Proof. Our objective is to construct a new probability space (E, FE , PE ).
If ω 6∈ E, then we can assign PE (ω) = 0. Suppose that ω ∈ E. For (E, FE , PE ) to be a
discrete probability space, we must have: PE (E) = 1 or:
X
(2.18) PE (E) = PE (ω) = 1
ω∈E
Thus, if we assign PE (ω) = P (ω)/P (E) for all ω ∈ E, then Equation 2.18 will be satisfied
automatically. Since for any F ⊆ E we know that:
X
P (F ) = P (ω)
ω∈F
Remark 2.30. The previous lemma gives us a direct way to construct P (F |E) for arbi-
trary F ⊆ Ω. Clearly if F ⊆ E, then
P (F )
P (F |E) = PE (F ) =
P (E)
Now suppose that F is not a subset of E but that F ∩ E 6= ∅. Then clearly, the only possible
events that can occur in F , given that E has occurred are the ones that are also in E. Thus,
PE (F ) = PE (E ∩ F ). More to the point, we have:
P (F ∩ E)
(2.19) P (F |E) = PE (F ∩ E) =
P (E)
Definition 2.31 (Conditional Probability). Given a discrete probability space (Ω, F, P )
and an event E ∈ F, the conditional probability of event F ∈ F given event E is:
P (F ∩ E)
(2.20) P (F |E) =
P (E)
Example 2.32 (Simple Blackjack). Blackjack is a game in which decisions can be made
entirely based on conditional probabilities. The chances of a card appearing are based
entirely on whether or not you have seen that card already since cards are discarded as the
dealer works her way through the deck.
Consider a simple game of Blackjack played with only the cards A, 2, 3, 4, 5, 6, 7, 8, 9,
10, J, Q, K. In this game, the dealer deals two cards to the player and two to herself. The
objective is to obtain a score as close to 21 as possible without going over. Face cards are
worth 10, A is worth 1 or 11 all other cards are worth their face value. We’ll assume that
the dealer must hit (take a new card) on 16 and below and will stand on 17 and above.
The complete sample space in this case is very complex; it consists of all possible valid
hands that could be dealt over the course of a standard play of the game. We can however
consider a simplified sample space of hands after the initial deal. In this case, the sample
space has the form:
Ω = {(hx, yi, hs, ti)}
Here x, y, s, t are cards without repeats. The total size of the sample space is
13 × 12 × 11 × 10 = 17, 160
This can be seen by noting that the player can receive any of the 13 cards as first card and
any of the remaining 12 cards for the second card. The dealer then receives 1 of the 11
remaining cards and then 1 of the 11 remaining cards.
Let’s suppose that the player is dealt 10 and 6 for a score of 16 while the dealer receives
a 4 and 5 for a total of 9. If we suppose that the player decides to hit, then the large sample
space (Ω) becomes:
Ω = {(hx, y, zi, hs, ti)}
which has size:
13 × 12 × 11 × 10 × 9 = 154, 440
while the event is:
E = {(h10, 6, zi, h4, 5i)}
3. CONDITIONAL PROBABILITY 9
There are 9 possible values for z and thus P (E) = 9/154, 440.
Let us now consider the probability of busting on our first hit. This is event F and is
given as:
F = {(hx, y, zi, hs, ti) : x + y + z > 21}
(Here we take some liberty by assuming that we can add card values like digits.)
The set F is very complex, but we can see immediately that:
E ∩ F = {(h10, 6, zi, h4, 5i) : z ∈ {7, 8, 9, J, Q, K}}
because these are the hands that will cause us to bust. Thus we can easily compute:
P (E ∩ F ) 6/154, 440 6 2
(2.21) P (F |E) = = = =
P (E) 9/154, 440 9 3
Thus the probability of not busting given the hand we have drawn must be 1/3. We can see
at once that our odds when taking a hit are not very good. Depending on the probabilities
associated with the dealer busting, it may be smarter for us to not take a hit and see what
happens to the dealer, however in order to be sure we’d have to work out the chances of the
dealer busting (since we know she will continue to hit until she busts or exceeds our value
of 16).
Unfortunately, this computation is quite tedious and we will not include it here.
Remark 2.33. The complexity associated with blackjack makes knowing exact probabil-
ities difficult, if not impossible. Thus most card counting strategies use heuristics to attempt
to understand approximately what the probabilities are for winning given the history of ob-
served hands. To do this, simple numeric values are assigned to cards, generally a +1 to
cards with low values (2,3, 4 etc.) a 0 to cards with mid-range values (7, 8, 9) and negative
values for face cards (10, J, Q, K). As the count gets high there are more face cards in the
deck and thus the chances of the dealer busting or the player drawing blackjack increase. If
the count is low, there are fewer face cards in the deck and the chance of the dealer drawing
a sufficient number of cards without busting is higher. Thus, players favor tables with high
counts.
The chief roadblock to card counters is knowing the count before sitting at the table. The
MIT card counting team (featured in the movie 21 ) used a big player team strategy. In this
strategy, card counters would sit at a table and make safe bets winning or losing very little
over the course of time. They would keep the card count and signal big players from their
team who would arrive at the table and make large bets when the count was high (in their
favor). The big players would leave once signaled that the count had dropped. Using this
strategy, the MIT players cleared millions from the casinos using basic probability theory.
Exercise 6. Use Definition 2.31 to compute the probability of obtaining a sum of 8 in
two rolls of a die given that in the first roll a 1 or 2 appears. [Hint: The space of outcomes
is still Ω = {(x, y)|x = 1, . . . , 6, y = 1, . . . , 6}. First identify the event E within this space.
How many elements within this set will enable you to obtain an 8 in two rolls? This is the
set E ∩ F What is the probability of E ∩ F ? What is the probability of E? Use the formula
in Defintion 2.31. It might help to write out the space Ω.]
Example 2.34 (The Monty Hall Problem). Congratulations! You are a contestant on
Let’s Make a Deal and you are playing for The Big Deal of the Day! You must choose
between Door Number 1, Door Number 2 and Door Number 3. Behind one of these doors is
10 2. PROBABILITY THEORY AND GAMES AGAINST THE HOUSE
a fabulous prize! Behind the other two doors, are goats. Once you choose your door, Monty
Hall (or Wayne Brady, you pick) will reveal a door that did not have the big deal. At this
point you can decide if you want to keep the original door you chose or switch doors. When
the time comes, what do you do?
It is tempting at first to suppose that it doesn’t matter whether you switch or not. You
have a 1/3 chance of choosing the correct door on your first try, so why would that change
after you are given information about an incorrect door? It turns out–it does matter.
To solve this problem, it helps to understand the set of potential outcomes. There are
really three possible pieces of information that determine an outcome:
(1) Which door the producer chooses for the big deal,
(2) Which door you choose first, and
(3) Whether you switch or not.
For the first decision, there are three possibilities (three doors). For the second decision,
there are again three possibilities (again three doors). For the third decision there are two
possibilities (either you switch, or not). Thus, there are 3 × 3 × 2 = 18 possible outcomes.
These outcomes can be visualized in the order in which the decisions are made (more or less)
this is shown in Figure 2.1. The first step (where the producers choose a door to hide the
prize) is not observable by the contestant, so we adorn this part of the diagram with a box.
We’ll get into what this box means when we discuss game trees.
Prize is Behind: 1 2 3
Choose Door: 1 2 3 1 2 3 1 2 3
Switch: Y N Y N Y N Y N Y N Y N Y N Y N Y N
Win/Lose: L W W L W L W L L W W L W L W L L W
Figure 2.1. The Monty Hall Problem is a multi-stage decision problem whose
solution relies on conditional probability. The stages of decision making are shown
in the diagram. We assume that the prizes are randomly assigned to the doors. We
can’t see this step–so we’ve adorned this decision with a square box. We’ll discuss
these boxes more when we talk about game trees. You the player must first choose
a door. Lastly, you must decide whether or not to switch doors having been shown
a door that is incorrect.
The next to the last row (labeled “Switch”) of Figure 2.1 illustrates the 18 elements of the
probability space. We could assume that they are all equally likely (i.e., that you randomly
choose a door and that you randomly decide to switch and that the producers of the show
randomly choose a door for hiding the prize). In this case, the probability of any outcome
is 1/18. Now, let’s focus exclusively on the outcomes in which we decide to switch. In the
figure, these appear with bold, colored borders. This is our event set E. Suppose event set
F consists of those outcomes in which the contestant wins. (This is shown in the bottom
3. CONDITIONAL PROBABILITY 11
row of the diagram with a W .) We are now interesting in P (F |E). That is, what are our
chances of winning, given we actively choose to switch?
Within E, there are precisely 6 outcomes in which we win. If each of these mutually
exclusive outcomes has probability 1/18:
1 1
P (E ∩ F ) = 6 =
18 3
Obviously, we switch in 9 of the possible 18 outcomes, so:
1 1
P (E) = 9 =
18 2
Thus we can compute:
P (E ∩ F ) 1/3 2
P (F |E) = = =
P (E) 1/2 3
Thus if we switch, there is a 2/3 chance we will win the prize. If we don’t switch, there is
only a 1/3 chance we win the prize. Thus, switching is better than not switching.
If this reasoning doesn’t appeal to you, there’s another way to see that the chance of
winning given switching is 2/3: In the case of switching we’re making a conscious decision;
there is no probabilistic voodoo that is affecting this part of the outcome. So just consider
the outcomes in which we switch. Notice there are 9 outcomes in which we switch from our
original door to a door we did not pick first. In 6 of these 9 we win the prize, while in 3 we
fail to win the prize. Thus, the chances of winning the prize when we switch is 6/9 or 2/3.
Exercise 7. Show (in anyway you like) that the probability of winning given that you
do not switch doors is 1/3.
Exercise 8. In the little known Lost Episodes of Let’s Make a Deal, Monty (or Wayne)
introduces a fourth door. Suppose that you choose a door and then are shown two incorrect
doors and given the chance to switch. Should you switch? Why? [Hint: Build a figure like
Figure 2.1. It will be a bit large. Use the same reasoning we used to compute the probability
of successfully winning the prize in the previous example.
Remark 2.35. The Monty Hall Problem first appeared in 1975 in the American Statis-
tician (if you believe Wikipedia–http://en.wikipedia.org/wiki/Monty_Hall_problem).
It’s one of those great problems that seems so obvious until you start drawing diagrams with
probability spaces. Speaking of Wikipedia, the referenced article is accessible, but contains
more advanced material. We’ll cover some of it later. On a related note, this example takes
us into our first real topic in game theory, Optimal Decision Making Under Uncertainty. As
we remarked in the example, the choice of whether to switch is really not a probabilistic
thing; it’s a decision that you must make in order to improve your happiness. This, at the
core, is what decision science, optimization theory and game theory is all about. Making a
good decision given all the information (stochastic or not) to improve your happiness.
Definition 2.36 (Independence). Let (Ω, F, P ) be a discrete probability space. Two
events E, F ∈ F are called independent if P (E|F ) = P (E) and P (F |E) = P (F ).
Theorem 2.37. Let (Ω, F, P ) be a discrete probability space. If E, F ∈ F are independent
events, then P (E ∩ F ) = P (E)P (F ).
12 2. PROBABILITY THEORY AND GAMES AGAINST THE HOUSE
Theorem 2.41 (Bayes Theorem 2). Let (Ω, F, P ) be a discrete probability space and
suppose that E, F1 , . . . , Fn ∈ F with F1 , . . . , Fn being pairwise disjoint and
n
[
Ω= Fi
i=1
Then:
P (E|Fi )P (Fi )
(2.23) P (Fi |E) = Pn
j=1 P (E|Fj )P (Fj )
Proof. Consider:
n n n
P (E ∩ Fj )
X X X
P (E|Fj )P (Fj ) = P (Fj ) = P (E ∩ Fj ) = P (E)
j=1 j=1
P (F j ) j=1
Exercise 11. In the previous example, for what probability of having the disease is
there a 1 in 100 chance of having the disease given that you’ve tested positive? [Hint: I’m
asking for what value of P (F ) is the value of P (F |E) 1 in 100. Draw a graph of P (F |E)
and use your calculator.
Exercise 12. Use Bayes Rule to show that the probability of winning in the Monty Hall
Problem is 2/3.
CHAPTER 3
Utility Theory
second digit in the price of the model car. What is the probability you choose the correct
value?
Remark 3.6. In a lottery (of this type) we do not assume that we will determine the
probability distribution P as a result of repeated exposure. (This is not like The Pennsylvania
Lottery.) Instead, the probability is given ab initio and is constant.
Definition 3.7 (Preference). Let L1 and L2 be lotteries. We write L1 L2 to indicate
that an individual prefers lottery L1 to lottery L2 . If both L1 L2 and L2 L1 , then
L1 ∼ L2 and L1 and L2 are considered equivalent to the individual.
Remark 3.8. The axiomatic treatment of utility theory rests on certain assumptions
about an individual’s behavior when they are confronted with a choice of two or more
lotteries. We have already seen this type of scenario in Example 3.5. We assume these
choices are governed by preference. Preference can vary from individual to individual.
Remark 3.9. For the remainder of this section we will assume that every lottery
consists of prizes A1 , . . . , An and that there prizes are preferred in order:
(3.1) A1 A2 · · · An
Assumption 1. Let L1 , L2 and L3 be lotteries:
(1) Either L1 L2 or L2 L1 or L1 ∼ L2 .
(2) If L1 L2 and L2 L3 , then L1 L3 .
(3) If L1 ∼ L2 and L2 ∼ L3 , then L1 ∼ L3 .
(4) If L1 L2 and L2 L1 , then L1 ∼ L2 .
Remark 3.10. Item 1 of Assumption 1 states that the ordering is a total ordering on
the set of all lotteries with which an individual may be presented. That is, we can compare
any two lotteries two each other and always be able to decide which one is preferred or
whether they are equivalent.
Item 2 of Assumption 1 states that this ordering is transitive.
It should be noted that these assumptions rarely work out in real-life. The idea that
everyone has in their mind a total ranking of all possible lotteries (or could construct one)
is difficult to believe. Ignoring that however, problems often arise more often with the
assumption of transitivity.
Remark 3.11. Assumption 1 asserts that preference is transitive over the set of all
lotteries. Since it is clear that preference should be reflexive (i.e., L1 ∼ L1 for all lotteries
L1 ) and symmetric (L1 ∼ L2 if and only if L2 ∼ L1 for all lotteries L1 and L2 ) preferential
equivalence is an equivalence relation over the set of all lotteries.
Example 3.12 (Problem with transitivity). For this example, you must use your imag-
ination and think like a pre-schooler (probably a boy pre-schooler).
Suppose I present a pre-schooler with the following choices (lotteries with only one item):
a ball, a stick and a crayon (and paper). If I present the choice of the stick and crayon, the
child may choose the crayon (crayons are fun to use when you have lot’s of imagination).
In presenting the stick and the ball, the child may choose the stick (a stick can be made
into anything using imagination). On the other hand, suppose I present the crayon and
the ball. If the child chooses the ball, then transitivity is violated. Why might the child
choose the ball? Suppose that the ball is not a ball but the ultimate key to the galaxy’s last
1. DECISION MAKING UNDER CERTAINTY 17
energy source! The child’s preferences will change depending upon the current requirements
of his/her imagination. Thus leading to a simple example of an intransitive ordering on the
items he is presented. This is evident only when presenting the items in pairs.
Definition 3.13 (Compound Lottery). Let L1 , . . . , Ln be a set of lotteries and suppose
that the probability of being presented with lottery i (i = 1, . . . , n) is qi . A lottery Q =
h(L1 , q1 ), . . . , (Ln , qn )i is called a compound lottery.
Example 3.14. Two contestants are playing a new game called Flip of a Coin! in which
“Your life can change on the flip of a coin!” The contestants first enter a round in which they
choose heads or tails. A coin is flipped and the winner is offered a choice of a sure $1, 000
or a 10% chance of winning a car. The loser is presented with a lottery in which they can
leave with nothing (and stay dry) or choose a lottery in which there is a 10% chance they
will win $1,000 and 90% they will fall into a tank of water dyed blue.
The coin flip stage is a compound lottery composed of the lotteries the contestants will
be offered later in the show.
Assumption 2. Let L1 , . . . , Ln be a compound lottery with probabilities q1 , . . . , qn and
suppose each Li is composed of prizes A1 , . . . , Am with probabilities pij (j = 1, . . . , m).
Then this compound lottery is equivalent to a simply lottery in which the probability
of prize Aj is:
rj = q1 p1j + q2 p2j + · · · + qn pnj
Remark 3.15. All Assumption 2 is saying is that compound lotteries can be transformed
into equivalent simple lotteries. Note further that the probability of prize j (Aj ) is actually:
n
X
(3.2) P (Aj ) = P (Aj |Lj )P (Lj )
i=1
This statement should be very clear from Theorem 2.19, when we define our probability
space in the right way.
Assumption 3. For each prize (or lottery) Ai there is a number ui ∈ [0, 1] so that
the prize Ai (or lottery Li ) is preferentially equivalent to the lottery in which you win
prize A1 with probability ui and An with probability 1 − ui and all other prizes with
probability 0. This lottery will be denoted Ãi .
Remark 3.16. Assumption 3 is a strange assumption often called the continuity assump-
tion. It assumes that for any ordered set of prizes (A1 , . . . , An ) that a person would view
winning any specific prize (Ai ) as equivalent to playing a game of chance in which either the
worst or best prize could be obtained.
This assumption is clearly not valid in all cases. Suppose that the best prize was a new
car, while the worst prize is spending 10 years in jail. If the prize in question (Ai ) is that
you receive $100, is there a game of chance you would play involving a new car or 10 years
in jail that would be equal to receiving $100?
Assumption 4. If L = h(A1 , p1 ), . . . , (Ai , pi ), . . . , (An , pn )i is a lottery, then L is pref-
erentially equivalent to the lottery h(A1 , p1 ), . . . , (Ãi , pi ), . . . , (An , pn )i
18 3. UTILITY THEORY
Remark 3.17. Assumption 4 only asserts that we can substitute any equivalent lottery
in for a prize and not change the individuals preferential ordering. It is up to you to evaluate
the veracity of this claim in real life.
Assumption 5. A lottery L in which A1 is obtained with probability p and An is
obtained with probability (1 − p) is always preferred or equivalent to a lottery in which
A1 is obtained with probability p0 and An is obtained with probability (1 − p0 ) if and
only if p ≥ p0 .
Remark 3.18. Our last assumption, Assumption 5 states that we would prefer (or be
indifferent) to win A1 with a higher probability and An with lower probability. This assump-
tion is reasonable when we have the case A1 An , however as [LR89] point out, there are
psychological reasons why this assumption may be violated.
At last we’ve reached the fundamental theorem in our study of utility.
Theorem 3.19 (Expected Utility Theorem). Let be a preference relation satisfying
Assumptions 1 - 5 over the set of all lotteries L defined over prizes A1 , . . . , An . Furthermore,
assume that:
A1 A2 · · · An
Then there is a function u : L → [0, 1] with the property that:
(3.3) u(L1 ) ≥ u(L2 ) ⇐⇒ L1 L2
Proof. The trick to this proof is to define the utility function and then show the if and
only if statement. We will define the utility function as follows:
(1) Define u(A1 ) = 1. Recall that A1 is not only prize A1 but also the lottery in which we
receive A1 with probability 1. That is the lottery in which p1 = 1 and p2 . . . , pn = 0.
(2) Define u(An ) = 0. Again, recall that An is also the lottery in which we receive An
with probability 1.
(3) By Assumption 3, for lottery Ai (i 6= 1 and i 6= n) there is a ui so that Ai is
equivalent to Ãi : the lottery in which you win prize A1 with probability ui and An
with probability 1 − ui and all other prizes with probability 0. Define u(Ai ) = ui .
(4) Let L ∈ L be a lottery in which we win prize Ai with probability pi . Then
(3.4) u(L) = p1 u1 + p2 u2 + · · · + pn un
Here u1 ≡ 1 and un ≡ 0.
We now show that this utility function satisfies Expression 3.3.
(⇐) Let L1 , L2 ∈ L and suppose that L1 L2 . Suppose:
L1 = h(A1 , p1 ), (A2 , p2 ), . . . , (An , pn )i
L2 = h(A1 , q1 ), (A2 , q2 ), . . . , (An , qn )i
By Assimption 3, for each Ai , (i 6= 1, i 6= n), we know that Ai ∼ Ãi with Ãi ≡ h(A1 , ui ), (An , 1 − ui )i.
Then by Assumption 4 we know:
L1 ∼ h(A1 , p1 ), (Ã2 , p2 ), . . . , (Ãn−1 , pn−1 ), (An , pn )i
L2 ∼ h(A1 , q1 ), (Ã2 , q2 ), . . . , (Ãn−1 , qn−1 ), (An , qn )i
1. DECISION MAKING UNDER CERTAINTY 19
Remark 3.20. This theorem is called the Expected Utility Theorem because the utility
for any lottery is really the expected utility from any of the prizes. That is, let U be the
random variable that takes value ui if prize Ai is received. Then:
n
X
(3.12) E(U ) = ui p(Ai ) = u1 p1 + u2 p2 + · · · + un pn
i=1
This is just the utility of the lottery in which prize i is received with probability pi .
Example 3.21. Congratulations! You’re on Let’s Make a Deal. The following prizes are
up for grabs:
(1) A1 : A new car (worth $15, 000)
(2) A2 : A gift card (worth $1, 000) to Best Buy
(3) A3 : A new iPad (worth $800)
(4) A4 : A Donkey (technically worth $500, but somewhat challenging)
We’ll assume that you prefer these prizes in the order in which they appear. Wayne Brady
offers you the following deal you can compete in either of the following games (lotteries):
(1) L1 = h(A1 , 0.25), (A2 , 0.25), (A3 , 0.25), (A4 , 0.25)i
(2) L2 = h(A1 , 0.15), (A2 , 0.4), (A3 , 0.4), (A4 , 0.05)i
Which games should you choose to make you the most happy? The problem here is actually
valuing the prizes. Maybe you really really need a new car (or you just bought a new car).
The car may be worth more than it’s dollar value. Alternatively, suppose you actually want
a donkey? Suppose you know that donkeys are expensive to own and the “retail” $450 value
is false. Maybe there’s not a Best Buy near you and it would be hard to use the gift card.
For the sake of argument, let’s suppose that you determine that the donkey is worth
nothing to you. You might say that:
(1) A2 ∼ h(A1 , 0.1), (A4 , 0.9)i
(2) A3 ∼ h(A1 , 0.05), (A4 , 0.95)i
The numbers really don’t make any difference, you can supply any values you want for 0.1
and 0.05 as long as the other numbers enforce Assumption 3. Then we can write:
(1) L1 ∼ h(A1 , 0.25), (h(A1 , 0.1), (A4 , 0.9)i, 0.25), (h(A1 , 0.05), (A4 , 0.95)i, 0.25), (A4 , 0.25)i
(2) L2 ∼ h(A1 , 0.15), (h(A1 , 0.1), (A4 , 0.9)i, 0.4), (h(A1 , 0.05), (A4 , 0.95)i, 0.4), (A4 , 0.05)i
We can now simplify this by expanding these compound lotteries into simple lotteries in
terms of A1 and A4 :
To see how we do this, let’s consider just Lottery 1: Lottery 1 is a compound lottery that
contains the following sub-lotteries:
(1) S1 : A1 with probability 0.25
(2) S2 : h(A1 , 0.1), (A4 , 0.9)i with probability 0.25
(3) S3 : h(A1 , 0.05), (A4 , 0.95)i with probability 0.25
(4) S4 : A4 with probability 0.25
To convert this lottery into a simpler lottery, we apply Assumption 2. The probability of
winning prize A1 is just the probability of winning prize A1 in one of the lotteries that make
up the compound lottery multiplied by the probability of playing in that lottery. Or:
P (A1 ) = P (A1 |S1 )P (S1 ) + P (A1 |S2 )P (S2 ) + P (A1 |S3 )P (S3 ) + P (A1 |S4 )P (S4 )
1. DECISION MAKING UNDER CERTAINTY 21
Exercise 14. Make up an example of a game with four prizes and perform the same
calculation that we did in Example 3.21. Explain what happens to your computation if you
replace the “donkey prize” with something more severe like being imprisoned for 10 years.
Does a penalty that is difficult to compare to prizes make it difficult to believe that the ui
values actually exist in all cases?
Definition 3.22 (Linear Utility Function). We say that a utility function u : L → R is
linear if given any lotteries L1 , L2 ∈ L and some q ∈ [0, 1], then:
(3.15) u (h(L1 , q), (L2 , (1 − q))i) = qu(L1 ) + (1 − q)u(L2 )
Here h(L1 , q), (L2 , (1 − q))i is the compound lottery made up of lotteries L1 and L2 each
having probabilities q and (1 − q) respectively.
Lemma 3.23. Let L be the collection of lotteries defined over prizes A1 , . . . , An with
A1 A2 · · · An . Let u : L → [0, 1] be the utility function defined in Theorem 3.19.
Then L1 ∼ L2 if and only if u(L1 ) = u(L2 ).
Exercise 15. Prove Lemma 3.23. [Hint: We know L1 L2 and L2 L1 if and only if
L1 ∼ L2 . We also know L1 L2 if and only if u(L1 ) ≥ u(L2 ). What, then do we know is
true about u(L1 ) and u(L2 ) when L2 L1 ? Use this, along with the rules of ordering in the
real numbers to prove the lemma.]
Theorem 3.24. The utility function u : L → [0, 1] in Theorem 3.19 is linear.
Proof. Let:
L1 = h(A1 , p1 ), (A2 , p2 ), . . . , (An , pn )i
L2 = h(A1 , r1 ), (A2 , r2 ), . . . , (An , rn )i
Thus we know that:
n
X
u(L1 ) = pi u i
i=1
Xn
u(L2 ) = ri ui
i=1
22 3. UTILITY THEORY
Choose q ∈ [0, 1]. The lottery L = h(L1 , q), (L2 , (1 − q))i is equivalent to a lottery in
which prize Ai is obtained with probability:
Pr(Ai ) = qpi + (1 − q)ri
Thus, applying Assumption 2 we have:
L̃ = h(A1 , [qp1 + (1 − q)r1 ]), . . . , (An , [qp1 + (1 − q)r1 ])i ∼ L
Applying Lemma 3.23, we can compute:
n
X n
X n
X
(3.16) u(L) = u(L̃) = [qpi + (1 − q)ri ] ui = qpi ui + (1 − q)ri ui =
i=1 i=1 i=1
n
! n
!
X X
q pi ui + (1 − q) ri ui = qu(L1 ) = (1 − q)u(L2 )
i=1 i=1
Remark 3.29. In this situation, we assume that the lotteries can change depending upon
the state of the world, which is provided by an event in S ⊆ Ω.
Example 3.30. In this world, suppose that the set of outcomes is the days of the week.
A game show might go something like this: on Tuesday and Thursday a contestant has a
50% chance of winning a car and a 50% chance of winning a donkey. On Monday, Wednesday
and Friday, there is a 20% chance of winning $100 and an 80% chance of winning $2, 000.
On Saturday and Sunday there is a 100% chance of winning nothing (because the game show
does not tape on the weekend).
Under these conditions, the Assumptions 1 through 5 must be modified to deal with the
state of the world. This is done by making the preference relation dependent on any given
subset S ⊆ Ω. Thus we end up with a collection of orderings S for any given subset S ⊆ Ω
(S 6= ∅).
The transformation of our assumptions into assumptions for the more complex case is
beyond the scope of our course. If you are interested, see Myerson’s Book (Chapter 1) for a
complete discussion. For those students interesting in studying graduate economics, this is a
worthwhile activity. The proof of the Generalized Expected Utility Theorem is substantially
more complex than our proof here. It is worth the effort to work through if you are properly
motivated.
CHAPTER 4
The purpose of this chapter is to create a formal and visual representation for a certain
class of games. This representation will be called extensive form, which we will define
formally as we proceed. We will proceed with our study of games under the following
assumptions:
(1) There are a finite set of Players: P = {P1 , . . . , PN }
(2) Each player has a knowledge of the rules of the game (the rules under which the
game state evolves) and the rules are fixed.
(3) At any time t ∈ R+ during game play, the player has a finite set of moves or choices
to make. These choices will affect the evolution of the game. The set of all available
moves will be denoted S.
(4) The game ends after some finite number of moves.
(5) At the end of the game, each player receives a prize. (Using the results in the
previous section, we assume that these prizes can be ordered according to preference
and that a utility function exists to assign numerical values to these prizes.)
In addition to these assumptions, some games may incorporate two other components:
(1) At certain points, there may be chance moves which advance the game in a non-
deterministic way. This only occurs in games of chance. (This occurs, e.g., in poker
when the cards are dealt.)
(2) In some games the players will know the entire history of moves that have been
made at all times. (This occurs, e.g., in Tic-Tac-Toe and Chess, but not e.g., in
Poker.)
v0 v0
�v0 , v1 , v2 � �v0 , v1 , v2 , v0 �
Path Path (Cycle)
Figure 4.2. Two Paths: We illustrate two paths in a digraph on three vertices.
Root
Terminal Vertices
Figure 4.3. Directed Tree: We illustrate a directed tree. Every directed tree has
a unique vertex called the root. The root is connected by a directed path to every
other vertex in the directed tree.
Thus,
Eo (v) = {(v, u) ∈ V : (v, u) ∈ E}
Definition 4.9 (Terminating Vertex). If T = (V, E) is a directed tree and v ∈ V so that
v has no descendants, then v is called a terminal vertex. All vertices that are not terminal
are non-terminal or intermediate.
Definition 4.10 (Tree Height). Let T = (V, E) be a tree. The height of the tree is the
length of the longest path in T .
Example 4.11. The height of the tree shown in Figure 4.3 is 4. There are three paths
of length 4 in the tree that start at the root of the tree and lead to three of the four terminal
vertices.
Lemma 4.12. Let T = (V, E) be a directed tree. If v is a vertex of v and u is a descendent
of v, then there is no path from u to v.
Proof. Let r be the root of the tree. Clearly if v = r, then the theorem is proved.
Suppose not. Let hw0 , w1 , . . . , wn i be a path from u to v with w0 = u and wn = v. Let
hx0 , x1 , . . . , xm i be the path from the root of the tree to the node v (thus x0 = r and xm = v).
Let hy0 , y1 , . . . , yk i be the path leading from the r to u (thus y0 = r and yk = u. Then we
can construct a new path:
hr = y0 , y1 , . . . , yk = u = w0 , w1 , . . . , wn = vi
from r (the root) to the vertex v. Thus there are two paths leading from the root to vertex
v, contradicting our assertion that T was a tree.
Theorem 4.13. Let T = (V, E) be a tree. Suppose u ∈ V is a vertex and let:
V (u) = {v ∈ V : v = u or v is a descendent of u}
Let E(u) be the set of all edges defined in paths connecting u to a vertex in V (u). Then the
graph Tu = (V (u), E(u)) is a tree with root u and is called the sub-tree of T descended from
u.
28 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
Root
Sub-tree
Figure 4.4. Sub Tree: We illustrate a sub-tree. This tree is the collection of all
nodes that are descended from a vertex u.
Example 4.14. A sub-tree of the tree shown in Example 4.6 is shown in Figure 4.4.
Sub-trees can be useful in analyzing decisions in games.
Proof. If u is the root of T , then the statement is clear. There is a unique path from
u (the root) to every vertex in T , by definition. Thus, Tu is the whole tree.
Suppose that u is not the root of T . The set V (u) consists of all descendants of u and u
itself. Thus between u and each v ∈ V (u) there is a path p = hv0 , v1 , . . . , vn i where v0 = u
and vn = v. To see this path must be unique, suppose that it is not, then there is at least
one other distinct path hw0 , w1 , . . . , wm i with w0 = u and wm = v. But if that’s so, we know
there is a unique path hx0 , . . . , xk i with x0 being the root of T and xk = u. It follows that
there are two paths:
hx0 , . . . , xk = v0 = u, v1 , . . . , vn = vi
hx0 , . . . , xk = w0 = u, w1 , . . . , wm = vi
between the root x0 and the vertex v. This is a contradiction of our assumption that T was
a directed tree.
To see that there is no path leading from any element in V (u) back to u, we apply Lemma
4.12. Since, by definition, every edge in the paths connecting u with its descendants are in
E(u) it follows that Tu is a directed tree and u is the root since there is a unique path from
u to each element of V (u) and there is no path leading from any element of V (u) back to u.
This completes the proof.
R P S
P2 P2 P2
R P S
R P S R P S
defined, because it looks like Player 2 has an advantage in knowing Player 1’s move before
making his own move. Irrespective of feelings, this is a valid game tree.
30 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
Figure 4.6. New Guinea is located in the south pacific and was a major
region of contention during World War II. The northern half was controlled
by Japan through 1943, while the southern half was controlled by the Al-
lies. (Image created from Wikipedia (http://en.wikipedia.org/wiki/File:
LocationNewGuinea.svg), originally sourced from http://commons.wikimedia.
org/wiki/File:LocationPapuaNewGuinea.svg.
massing troops to reinforce their army on New Guinea in an attempt to control the entire
island. These troops had to be delivered by naval convoy. The Japanese had a choice of
sailing either north of New Britain, where rain and poor visibility was expected or south of
New Britain, where the weather was expected to be good. Either route required the same
amount of sailing time.
General Kenney, the Allied Forces Commander in the Southwest Pacific had been ordered
to do as much damage to the Japanese convoy fleet as possible. He had reconnaissance
aircraft to detect the Japanese fleet, but had to determine whether to concentrate his search
planes on the northern or southern route.
The following game tree summarizes the choice for the Japanese (J) and American (A)
commanders (players), with payoffs given as the number of days available for bombing of the
Japanese fleet. (Since the Japanese cannot benefit, there payoff is reported as the negative
2. GAME TREES WITH COMPLETE INFORMATION AND NO CHANCE 31
of these values.) The moves for each player are sail north or sail south for the Japanese and
search north or search south for the Americans.
J
N S
A A
N S N S
Figure 4.7. The game tree for the Battle of the Bismark Sea. The Japanese
could choose to sail either north or south of New Britain. The Americans (Allies)
could choose to concentrate their search efforts on either the northern or southern
routes. Given this game tree, the Americans would always choose to search the
North if they knew the Japanese had chosen to sail on the north side of New Britain;
alternatively, they would search the south route, if they knew the Japanese had taken
that. Assuming the Americans have perfect intelligence, the Japanese would always
choose to sail the northern route as in this instance they would expose themselves
to only 2 days of bombing as opposed to 3 with the southern route.
This example illustrates the importance of intelligence in warfare. In this game tree,
we assume perfect information. Thus, the Americans know (through backchannels) which
route the Japanese will sail. In knowing this, they can make an optimal choice for each
contingency. If the Japanese sail north, then the Americans search north and will be able
to bomb the Japanese fleet for 2 days. Similarly, if the Japanese sail south, the Americans
will search south and be able to bomb the Japanese fleet for 3 days.
The Japanese, however, also have access to this game tree and reasoning that the Amer-
icans are payoff maximizers, will chose a path to minimize their exposure to attack. They
must choose to go north and accept 2 days of bombing. If they choose to go south, then
they know they will be exposed to 3 days of bombing. Thus, their optimal strategy is to sail
north.
Naturally, the Allies did not know which route the Japanese would take and there was
no backchannel intelligence. We will come back to this case later. However, this example
serves to show how important intelligence is in warfare since it can help commanders make
optimal decisions.
Exercise 18. Using the approach from Example 4.23 derive a strategy for Player 2 in
the Rock-Paper-Scissors game (Example 4.20) assuming she will attempt to maximize her
payoff. Similarly, show that it doesn’t matter whether Player 1 chooses Rock, Paper or
Scissors in this game and thus any strategy for Player 1 is equally good (or bad).
Remark 4.24. The complexity of a game (especially one with perfect information and no
chance moves) can often be measured by how many nodes are in its game tree. A computer
that wishes to play a game often attempts to explore the game tree in order to make its
moves. Certain games, like Chess and Go, have huge game trees. Another measure of
complexity is the length of the longest path in the game tree.
32 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
In our odd version Rock-Paper-Scissors, the length of the longest path in the game tree
is 3 nodes. This reflects the fact that there are only two moves in the game: first Player 1
moves and then Player 2 moves.
Exercise 19. Consider a simplified game of tic-tac-toe where the objective is to fill in
a board shown in Figure 4.8
XX
O
Game Board X Wins!
Figure 4.8. Simple tic-tac-toe: Players in this case try to get two in a row.
Assuming that X goes first. Construct the game tree for this game by assuming that
the winner receives +1 while the loser receives −1 and draws result in 0 for both players.
Compute the depth of the longest path in the game tree. Show that there is a strategy so
that the first player always wins. [Hint: You will need to consider each position in the board
as one of the moves that can be made.]
Exercise 20. In a standard 3 × 3 tic-tac-toe board, compute the length of the longest
path in the game tree. [Hint: Assume you draw in this game.]
3. Game Trees with Incomplete Information
Remark 4.25 (Power Set and Partitions). Recall from Remark 2.11 that, if X is a set,
then 2X is the power set of X or the set of all subsets of X. Any parition of X is a set
I ⊆ 2X so that: For all x ∈ X there is exactly one element I ∈ I so that x ∈ I.
(Remember, I is a subset of X and as such, I ∈ I ⊆ 2X .
Definition 4.26 (Information Sets). If T = (V, E) is a tree and D ⊂ V are the in-
termediate (decision) nodes of the tree, ν is a player assignment function and µ is a move
assignment, then information sets are a set I ⊂ 2D , satisfying the following:
(1) For all v ∈ D there is exactly one set Iv ∈ I so that v ∈ Iv . This is the information
set of the vertex v.
(2) If v1 , v2 ∈ Iv , then ν(v1 ) = ν(v2 ).
(3) If (v1 , v) ∈ E and µ(v1 , v) = m, and v2 ∈ Iv1 (that is, v1 and v2 are in the same
information set), then there is some w ∈ V so that (v2 , w) ∈ E and µ(v2 , w) = m
Thus I is a partition of D.
Remark 4.27. Definition 4.26 says that every vertex in a game tree is assigned a single
information set. It also says that if two vertices are in the same information set, then they
must both be controlled by the same player. Finally, the definition says that two vertices can
be in the same information set only if the moves from these vertices are indistinguishable.
An information set is used to capture the notion that a player doesn’t know what vertex
of the game tree he is at; i.e., that he cannot distinguish between two nodes in the game tree.
All that is known is that the same moves are available at all vertices in a given information
set.
3. GAME TREES WITH INCOMPLETE INFORMATION 33
In a case like this, it is possible that the player doesn’t know which vertex in the game
tree will come next as a result of choosing a move, but he can certainly limit the possible
vertices.
Remark 4.28. We can also think of the information set as being a mapping ξ : V → I
where I is a finite set of information labels and the labels satisfy requirements like those in
Definition 4.26. This is the approach that Myerson [Mye01] takes.
Exercise 21. Consider the information sets a set of labels I and let ξ : V → I. Write
down the constraints that ξ must satisfy so that this definition of information set is analogous
to Definition 4.26.
Definition 4.29 (Game Tree with Incomplete Information and No Chance Moves). A
game tree with incomplete information and no chance is a tuple G = (T, P, S, ν, µ, π, I) such
that T is a directed tree, ν is a player vertex assignment on intermediate vertices of T , µ is
a move assignment on the edges of T and π is a payoff function on T and I are information
sets.
Definition 4.30 (Strategy–Imperfect Information). Let G = (T, P, S, ν, µ, π, I) be a
game tree with incomplete information and no chance moves, with T = (V, E). Let Ii be
the information sets controlled by Player i. A pure strategy for Player Pi is a mapping
σi : Ii → S with the property that if I ∈ Ii and σi (I) = s, then for every v ∈ I there is
some edge (v, w) ∈ E so that µ(v, w) = s.
Proposition 4.31. If G = (T, P, S, ν, µ, π, I) and I consists of only singleton sets, then
G is equivalent to a game with complete information.
Proof. The information sets are used only in defining strategies. Since each I ∈ I is
a singleton, we know that for each I ∈ I we have I = {v} where v ∈ D. (Here D is the
set of decision nodes in V with T = (V, E).) Thus any strategy σi : Ii → E can easily be
converted into σi : Vi → E by stating that σi (v) = σi ({v}) for all v ∈ Vi . This completes the
proof.
Example 4.32 (The Battle of the Bismark Sea (Part 2)). Obviously, General Kenney
did not know a priori which route the Japanese would take. This can be modeled using
information sets. In this game, the two nodes that are owned by the Allies in the game tree
are in the same information set. General Kenney doesn’t know whether the Japanese will
sail north or south. He could (in theory) have reasoned that they should sail north, but he
doesn’t know. The information set for the Japanese is likewise shown in the diagram.
In determining a strategy, the Allies and Japanese must think a little differently. The
Japanese could choose to go south. If the Allies are lucky and choose to search south, the
Japanese will be in for three days worth of attacks. If the allies are unlucky and choose to go
north, the Japanese will still face two days of bombing. On the other hand, if the Japanese
choose to go north, then they may be unlucky and the Allies will choose to search north in
which case they will again take 2 days of bombing. If however, the allies are unlucky, the
japanese will face only 1 day of bombing.
From the perspective of the Japanese, since the routes will take the same amount of
time, the northern route is more favorable. To see this note Table 1: If the Japanese sail
north, then the worst they will suffer is 2 days of bombing and the best they will suffer
is one day of bombing. If the Japanese sail south, the worse they will suffer is 3 days of
34 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
Japanese Information Set
N S
A A
N S N S
Figure 4.9. The game tree for the Battle of the Bismark Sea with incomplete
information. Obviously Kenney could not have known a priori which path the
Japanese would choose to sail. He could have reasoned (as they might) that there
best plan was to sail north, but he wouldn’t really know. We can capture this fact
by showing that when Kenney chooses his move, he cannot distinguish between the
two intermediate nodes that belong to the Allies.
bombing and the best they will suffer is 2 days of bombing. Thus, the northern route should
be preferable as the cost to taking it is never worse than taking the southern route. We say
that the northern route strategy dominates the southern route strategy. If General Kenney
could reason this, then he might choose to commit his reconnaissance forces to searching the
north, even without being able to determine whether the Japanese sailed north or south.
Exercise 22. Identify the information sets for Rock-Paper-Scissors and draw the game
tree to illustrate the incomplete information. Do not worry about trying to identify an
optimal strategy for either player.
4. Games of Chance
In games of chance, there is always a point in the game where a chance move is made. In
card games, the initial deal is one of these points. To accommodate chance moves, we assume
the existence of a Player 0 who is sometimes called Nature. When dealing with games of
chance, we assume that the player vertex assignment function assigns some vertices the label
P0 .
Definition 4.33 (Moves of Player 0). Let T = (V, E) and let ν be a player vertex
assignment function. For all v ∈ D such that ν(v) = P0 here is a probability assignment
4. GAMES OF CHANCE 35
Remark 4.34. The probability function(s) pv in Definition 4.33 essentially defines an roll
of the dice. When game play reaches a vertex owned by P0 , Nature (or Player 0 or Chance)
probabilistically advances the game by moving along an randomly chosen edge. The fact
that Equation 4.1 holds simply asserts that the chance moves of Nature form a probability
space at that point, whose outcomes are all the possible chance moves.
Definition 4.35 (Game Tree). Let T = (V, E) be a directed tree, let F ⊆ V be the
terminal vertices and let D = V \ F be the intermediate (or decision) vertices. Let P =
{P0 , P1 , . . . , Pn } be a set of players including P0 the chance player. Let S be a set of moves
for the players. Let ν : D → P be a player vertex assignment function and µ : E → S be a
move assignment function. Let
P = {pv : ν(v) = P0 and pv is the moves of Player 0}
Let π : F → Rn be a payoff function. Let I ⊆ 2D be the set of information sets.
A game tree is a tuple G = (T, P, S, ν, µ, π, I, P). In this form, the game defined by the
game tree G is said to be in extensive form.
Remark 4.36. A strategy for Player i in a game tree like the one in Definition 4.35 is
the same as that in Definition 4.30
Example 4.37 (Red-Black Poker). This example is taken from Chapter 2 of [Mye01].
At the beginning of this game, each player antes up $1 into a common pot. Player 1 takes
a card from a randomized (shuffled) deck. After looking at the card, Player 1 will decide
whether to raise or fold.
(1) If Player 1 folds, he shows the card to Player 2: If the card is red, then Player 1
wins the pot and Player 2 loses the pot. If the card is black, then Player 1 loses the
pot and Player 2 wins the pot.
(2) If Player 1 raises, then Player 1 adds another dollar to the pot and Player 2 must
decide whether to call or fold.
(a) If Player 2 folds, then the game ends and Player 1 takes the money irrespective
of his card.
(b) If Player 2 calls, then he adds $1 to the pot. Player 1 shows his card. If his
card is red, then he wins the pot ($2) and Player 2 loses the pot. If Player 1’s
card is black, then he loses the pot and Player 2 wins the pot ($2).
The game tree for this game is shown in Figure 4.10 The root node of the game tree is
controlled by Nature (Player 0). This corresponds to the initial draw of Player 1, which is
random and will result in a red card 50% of the time and a black card 50% of the time.
Notice that the nodes controlled by P2 are in the same information set. This is because
it is impossible for Player 2 to know whether or not Player 1 has a red card or a black card.
The payoffs shown on the terminal nodes are determined by how much each player will
win or loose.
36 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
P0
Red (0.5) Black (0.5)
P1 P1
Raise Raise
Fold Fold
P2 P2
Figure 4.10. Poker: The root node of the game tree is controlled by Nature. At
this node, a single random card is dealt to Player 1. Player 1 can then decide whether
to end the game by folding (and thus receiving a payoff or not) or continuing the
game by raising. At this point, Player 2 can then decide whether to call or fold,
thus potentially receiving a payoff.
Exercise 23. Draw a game tree for the following game: At the beginning of this game,
each player antes up $1 into a common pot. Player 1 takes a card from a randomized
(shuffled) deck. After looking at the card, Player 1 will decide whether to raise or fold.
(1) If Player 1 folds, he shows the card to Player 2: If the card is red, then Player 1
wins the pot and Player 2 loses the pot. If the card is black, then Player 1 loses the
pot and Player 2 wins the pot.
(2) If Player 1 raises, then Player 1 adds another dollar to the pot and Player 2 picks
a card and must decide whether to call or fold.
(a) If Player 2 folds, then the game ends and Player 1 takes the money irrespective
of any cards drawn.
(b) If Player 2 calls, then he adds $1 to the pot. Both players show their cards.
If both cards of the same suit, then Player 1 wins the pot ($2) and Player 2
loses the pot. If the cards are of opposite suits, then Player 2 wins the pot and
Player 1 loses.
P1
Raise
Fold
P2
Call Fold
Figure 4.11. Reduced Red Black Poker: We are told that Player 1 receives a red
card. The resulting game tree is substantially simpler. Because the information set
on Player 2 controlled nodes indicated a lack of knowledge of Player 1’s card, we
can see that this sub-game is now a complete information game.
noting that when we restrict our attention to this sub-tree, a game that was originally an
incomplete information game becomes a complete information game. That is, each vertex is
now the sole member in its information set. Additionally, we have removed chance from the
game.
Exercise 24. Continuing from Exercise 23 draw the game tree when we know that
Player 1 is dealt a red card. Illustrate in your drawing how it is a sub-tree of the tree you
drew in Exercise 23. Determine whether this game is still (i) a game of chance and (ii)
whether it is a complete information game or not.
Theorem 4.40. Let G = (T, P, S, ν, µ, π, I) be a game with no chance. Let σ1 , . . . , σN
be set of strategies for Players 1 through n. Then these strategies determine a unique path
through the game tree.
Proof. To see this, suppose we begin at the root node r. If this node is controlled by
Player i, then node r exists in information set Ir ∈ Ii . Then σi (Ir ) = s ∈ S and there is
38 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
some edge (r, u) ∈ E so that µ(r, u) = s. The next vertex determined by the strategy σi is
u. In either case, we have a two vertex path (r, u).
Consider the game tree G 0 constructed from sub-tree Tu and determined as in Theorem
4.38. This game tree has root u. We can apply the same argument to construct a two
vertex path (u, u0 ), which when joined with the initial path forms the three node path
(r, u, u0 ). Repeating this argument inductively will yield a path through the game tree that
is determined by the strategy functions of the players. Since the number of vertices in the
tree is finite, this process must stop, producing the desired path. Uniqueness of the path
is ensured by the fact that at the strategies are functions and thus at any information set,
exactly one move will be chosen by the player in control.
Example 4.41. In the Battle of the Bismark Sea, the strategy we defined in Example
4.23 clearly defines a unique path through the tree: Since each player determines a priori
J
N S
A A
N S N S
Figure 4.12. A unique path through the game tree of the Battle of the Bismark
Sea. Since each player determines a priori the unique edge he/she will select when
confronted with a specific information set, a path through the tree can be determined
from these selections.
the unique edge he/she will select when confronted with a specific information set, a path
through the tree can be determined from these selections. This is illustrated in Figure 4.12.
Exercise 25. Define a strategy for Rock-Paper-Scissors and show the unique path
through the tree in Figure 4.5 determined by this strategy. Do the same for the game
tree describing the Battle of the Bismark Sea with incomplete information.
Theorem 4.42. Let G = (T, P, S, ν, µ, π, I, P). Let σ1 , . . . , σN be a collection of strate-
gies for Players 1 through n. Then these strategies determine a discrete probability space
(Ω, F, P ) where Ω is a set of paths leading from the root of the tree to a subset of the ter-
minal nodes and if ω ∈ Ω, then P (ω) is the product of the probabilities of the chance moves
defined by the path ω.
Proof. We will proceed inductively on the height of the tree T . Suppose the tree T
has a height of 1. Then there is only one decision vertex (the root). If that decision vertex
is controlled by a player other than chance, then applying Theorem 4.40 we know that the
strategies σ1 , . . . , σN defined a unique path through the tree. The only paths in a tree of
height 1 have the form hr, ui where r is the root of T and u is a terminal vertex. Thus, Ω is
the singleton consisting of only the path hr, ui determined by the strategies and it is assigned
a probability of 1.
5. PAY-OFF FUNCTIONS AND EQUILIBRIA 39
ensures that we can define the probability space (Ω, F, P ). Thus we have shown that the
theorem is true for game trees of height 1.
Suppose the statement is true for game trees with height up to k ≥ 1. We will show that
the theorem is true for game trees of height k + 1. Let r be the root of tree T and consider
the set of children of U = {u ∈ V : (r, u) ∈ E}. For each u ∈ U , we can define a game
tree of height k with tree Tu by Theorem 4.38. The fact that this tree has height k implies
that we can define a probability space (Ωu , Fu , Pu ) with Ωu composed of paths from u to the
terminal vertices of Tu .
Suppose that vertex r is controlled by Player Pj (j 6= 0). Then the strategy σj determines
a unique move that will be made by Player j at vertex r. Suppose that move m is determined
by σj at vertex r and µ(r, u) = m for edge (r, u) ∈ E with u ∈ U (that is edge (r, u) is labeld
m). We can define the new event set Ω of paths in the tree T from root r to a terminal
vertex. The probability function on paths can then be defined as:
(
Pu (hv1 , . . . , vk i) hv1 , . . . , vk i ∈ Ωu
P (hr, v1 , . . . , vk i) =
0 else
The fact that Pu is a properly defined probability function over Ωu implies that P is a
properly defined probability function over Ω and thus (Ω, F, P ) is a probability space over
the paths in T .
Now suppose that chance (Player P0 ) controls r in the game tree. Again, Ω is the set of
paths leading from r to a terminal vertex of T . The probability function on paths can then
be defined as:
P (hr, v1 , . . . , vk i) = pr (r, v1 )Pv1 (hr, v1 , . . . , vk i)
Here v1 ∈ U and hr, v1 , . . . , vk i ∈ Ωv1 , the set of paths leading from v1 to a terminal vertex
in tree Tv1 and p(r, v1 ) is the probability chance assigns to edge (r, v1 ) ∈ E.
To see that this is a properly defined probability function, suppose that ω ∈ Ωu that
is, ω is a path in tree Tu leading from u to a terminal vertex of Tu . Then a path in Ω is
constructed by joining the path that leads from vertex r to vertex u and then following a
path ω ∈ Ωu . Let hr, ωi denote such a path. Then we know:
XX XX
(4.2) P (hr, ωi) = p(r, u)Pu (ω) =
u∈U ω∈Ωu u∈U ω∈Ωu
!
X X X
p(r, u) Pu (ω) = p(r, u) = 1
u∈U u∈Ωu u∈U
40 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
P
This is because ω∈Ωu Pu (ω) = 1. Since clearly P (hr, ωi) ∈ [0, 1] and the paths through the
game tree are independent, it follows that (Ω, F, P ) is a properly defined probability space.
Thus the theorem follows by induction. This completes the proof.
Example 4.43. Consider the simple game of poker we defined in Example 4.37. Suppose
we fix strategies in which Player 1 always raises and Player 2 always calls. Then the resulting
probability distribution defined as in Theorem 4.42 contains two paths (one when a red card
is dealt and another when a black card is dealt. This is shown in Figure 4.13. The sample
P0
Red (0.5) Black (0.5)
P1 P1
Raise Raise
Fold Fold
P2 P2
{ → 50%
{
P0 Red (0.5) P1 Raise P2 Call
Ω=
P0 Black (0.5) P1 Raise P2 Call → 50%
Figure 4.13. The probability space constructed from fixed player strategies in a
game of chance. The strategy space is constructed from the unique choices deter-
mined by the strategy of the players and the independent random events that are
determined by the chance moves.
space consists of the possible paths through the game tree. Notice that as in Theorem 4.40
the paths through the game tree are completely specified (and therefore unique) when the
non-chance players are determining the moves. The only time probabilistic moves occur is
when chance is causes the game to progress..
Example 4.44. Suppose we play a game in which Players 1 and 2 ante $1 each. One
card each is dealt to Player 1 and Player 2. Player 1 can choose to raise (and add a $1 to
the pot) or fold (and lose the pot). Player 2 can then choose to call (adding $1) or fold (and
loose the pot). Player 1 wins if both cards are black. Player 2 wins if both cards are red.
The pot is split if the cards have opposite color. Suppose that Player 1 always chooses to
raise and Player 2 always chooses to call. Then the game tree and strategies are shown in
Figure 4.14. The sample space in this case consists of 4 distinct paths each with probability
1/4, assuming that the cards are dealt with equal probability. Note in this example that
constructing the probabilities of the various events requires multiplying the probabilities of
the chance moves in each path. This is illustrated in the theorem when we write:
P (hr, v1 , . . . , vk i) = pr (r, v1 )Pv1 (hr, v1 , . . . , vk i)
Exercise 26. Suppose that players always raise and call in the game defined in Exercise
23. Compute the probability space defined by these strategies in the game tree you developed.
5. PAY-OFF FUNCTIONS AND EQUILIBRIA 41
P0
Red (0.5) Card to Player 1 Black (0.5)
P0 P0
Red (0.5) Card to Player 2 Black (0.5) Red (0.5) Card to Player 2 Black (0.5)
P1 P1 P1 P1
P2 P2 P2 P2
(-1, 1) (-2, 2) (1,-1) (0,0) (1,-1) (-1,1) (-1,1) (0,0) (1,-1) (2,-2) (1,-1) (-1,1)
{ {
P0 Red (0.5) P0 Red (0.5) P1 Raise P2 Call → 25%
P0 Red (0.5) P0 Black (0.5) P1 Raise P2 Call → 25%
Ω=
P0 Black (0.5) P0 Red (0.5) P1 Raise P2 Call → 25%
P0 Black (0.5) P0 Black (0.5) P1 Raise P2 Call → 25%
Figure 4.14. The probability space constructed from fixed player strategies in a
game of chance. The strategy space is constructed from the unique choices deter-
mined by the strategy of the players and the independent random events that are
determined by the chance moves. Note in this example that constructing the prob-
abilities of the various events requires multiplying the probabilities of the chance
moves in each path.
Definition 4.45 (Strategy Space). Let Σi be the set of all strategies for Player i in a
game tree G. Then the entire strategy space is Σ = Σ1 × Σ2 × · · · × Σn .
Definition 4.46 (Strategy Payoff Function). Let G be a game tree with no chance
moves. The strategy payoff function is a mapping π : Σ → Rn . If σ1 , . . . , σN are strategies for
Players 1 through n, then π(σ1 , . . . , σN ) is the vector of payoffs assigned to the terminal node
of the path determined by the strategies σ1 , . . . , σN in game tree G. For each i = 1, . . . , N
πi (σ1 , . . . , σN ) is the payoff to Player i in πi (σ1 , . . . , σN ).
Example 4.47. Consider the Battle of the Bismark Sea game from Example 4.32. Then
there are four distinct strategies in Σ with the following payoffs:
π (Sail North, Search North) = (−2, 2)
π (Sail South, Search North) = (−2, 2)
π (Sail North, Search South) = (−1, 1)
π (Sail South, Search South) = (−3, 3)
Definition 4.48 (Expected Strategy Payoff Function). Let G be a game tree with chance
moves. The expected strategy payoff function is a mapping π : Σ → Rn defined as follows: If
σ1 , . . . , σN are strategies for Players 1 through n, then let (Ω, F, P ) be the probability space
over the paths constructed by these strategies as given in Theorem 4.42. Let Πi be a random
42 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
variable that maps ω ∈ Ω to the payoff for Player i at the terminal node in path ω. Let:
πi (σ1 , . . . , σN ) = E(Πi )
Then:
π(σ1 , . . . , σN ) = hπ1 (σ1 , . . . , σN ), . . . , πN (σ1 , . . . , σN )i
As before, πi (σ1 , . . . , σN ) is the expected payoff to Player i in π(σ1 , . . . , σN ).
Example 4.49. Consider Example 4.37. There are 4 distinct strategies in Σ:
(Fold, Call)
(Fold, Fold)
(Raise, Call)
(Raise, Fold)
Let’s focus on the strategy (Fold, Call). Then the resulting paths in the graph defined by
these strategies are shown in Figure 4.15. There are two paths and we note that the decision
Figure 4.15. Game tree paths derived from the Simple Poker Game as a result of
the strategy (Fold, Fold). The probability of each of these paths is 1/2.
made by Player 2 makes no difference in this case because Player 1 folds. Each path has
probability 1/2. Our random variable Π1 will map the top path (in Figure 4.15) to a $1
payoff for Player 1 and will map the bottom path (in Figure 4.15) to a payoff of −$1 for
Player 1. Thus we can compute:
1 1
π1 (Fold, Fold) = (1) + (−1) = 0
2 2
Likewise,
1 1
π2 (Fold, Fold) = (−1) + (1) = 0
2 2
Thus we compute:
π (Fold, Fold) = (0, 0)
Using this approach, we can compute the expected payoff function to be:
π (Fold, Call) = (0, 0)
π (Fold, Fold) = (0, 0)
π (Raise, Call) = (0, 0)
π (Raise, Fold) = (1, −1)
Exercise 27. Explicitly show that the expected payoff function for Simple Poker is the
one given in the previous example.
5. PAY-OFF FUNCTIONS AND EQUILIBRIA 43
∗
Definition 4.50 (Equilibrium). A strategy (σ1∗ , . . . , σN ) ∈ Σ is an equilibrium if for all
i.
∗
πi (σ1∗ , . . . , σi∗ , . . . , σN ∗
) ≥ πi (σ1∗ , . . . , σi , . . . , σN )
where σi ∈ Σi .
Example 4.51. Consider the Battle of the Bismark Sea. We can show that (Sail North, Search North)
is an equilibrium strategy. Recall that:
π (Sail North, Search North) = (−2, 2)
Now, suppose that the Japanese deviate from this strategy and decide to sail south. Then
the new payoff is:
π (Sail South, Search North) = (−2, 2)
Thus:
π1 (Sail North, Search North) ≥ π1 (Sail South, Search North)
Now suppose that the Allies deviate from the strategy and decide to search south. Then
the new payoff is:
π (Sail North, Search South) = (−1, 1)
Thus:
π2 (Sail North, Search North) > π2 (Sail North, Search South)
Exercise 28. Show that the strategy (Raise, Call) is an equilibrium strategy in Simple
Poker.
Theorem 4.52. Let G = (T, P, S, ν, µ, π, I, P) be a game tree with complete information.
∗
Then there is an equilibrium strategy (σ1∗ , . . . , σN ) ∈ Σ.
Proof. We will apply induction on the height of the game tree T = (V, E). Before
proceeding to the proof, recall that a game with complete information is one in which if
v ∈ V and Iv ∈ I is the information set of vertex v, then Iv = {v}. Thus we can think of a
strategy σi for player Pi as being as being a mapping from V to S as in Definition 4.21. We
now proceed to the proof.
Suppose the height of the tree is 1. Then the tree consists of a root node r and a collection
of terminal nodes F so that if u ∈ F then (r, u) ∈ E. If chance controls r, then there is no
strategy for any of the players, they are randomly assigned a payoff. Thus we can think of
the empty strategy as the equilibrium strategy. On the other hand, if player Pi controls r,
then we let σi (r) = m ∈ S so that if µ(r, u) = m for some u ∈ F then πi (u) ≥ πi (v) for all
other v ∈ U . That is, the vertex reached by making move m has a payoff for Player i that
is greater than or equal to any other payoff Player i might receive at another vertex. All
other players are assigned empty strategies (as they never make a move). Thus it is easy to
see that this is an equilibrium strategy since no player can improve their payoff by changing
strategies. Thus we have proved that there is an equilibrium strategy in this case.
Now suppose that the theorem is true for game trees G with complete information of
height some k ≥ 1. We will show that the statement holds for a game tree of height k + 1.
Let r be the root of the tree and let U = {u ∈ V : (r, u) ∈ E} be the set of children of r in
T . If r is controlled by chance, then the first move of the game is controlled by chance. For
each u ∈ U , we can construct a game tree with tree Tu by Theorem 4.38. By the induction
44 4. GAME TREES, EXTENSIVE FORM, NORMAL FORM AND STRATEGIC FORM
∗ ∗ ∗
hypothesis, we know there is some equilibrium strategy (σ1u , . . . , σN u
). Let πiu be the payoff
associated with using this strategy for Player Pi . Now consider any alternative strategy
∗ u∗ u∗ u∗
(σ1u , . . . , σi−1 , σiu , σi+1 . . . , σN ). Let πiu be the payoff to Player Pi that results from using
this new strategy in the game with game tree Tu . It must be that
∗
(4.3) πiu ≥ πiu ∀i ∈ {1, . . . , N }, u ∈ U
Thus we construct a new strategy for Player Pi so that if chance causes the game to transition
∗
to vertex u in the first step, then Player Pi will use strategy σiu . Equation 4.3 ensures that
Player i will never have a motivation to deviate from this strategy as the assumption of
complete information assures us that Player i will know for certain to which u ∈ U the game
has transitioned.
∗
Alternatively, suppose that the root is controlled by Player Pj . Let U and πiu be as
above. Then let σj (r) = m ∈ S so that if µ(r, u) = m then:
∗ ∗
(4.4) πju ≥ πjv
for all v ∈ U . That is, Player Pj chooses a move that will yield a new game tree Tu that has
∗ u∗
the greatest terminal payoff using the equilibrium strategy (σ1u , . . . , σN ) in that game tree.
We can now define a new strategy:
(1) At vertex r, σj (r) = m.
∗ u∗
(2) Every move in tree Tu is governed by (σ1u , . . . , σN )
(3) If v 6= r and v 6∈ Tu and ν(v) = i, then σi (v) may be chosen at random from S
(because this vertex will never be reached during game play).
We can show that this is an equilibrium strategy. To see this, consider any other strategy.
If Player i 6= j deviates, then we know that this player will receive payoff πiu (as above)
because Player j will force the game into the tree Tu after the first move. We know further
∗
that πiu ≥ πiu . Thus, there is no incentive for Player Pi to deviate from the given strategy.
∗ u∗
He must play (σ1u , . . . , σN ) in Tu . If Player j deviates at some vertex in Tu , then we know
∗
Player j will receive payoff πju ≤ πju . Thus, once game play takes place inside tree Tu there
is no reason to deviate from the given strategy. If Player j deviates on the first move and
chooses a move m0 so that µ(r, v) = m0 , then there are two possibilities:
∗ ∗
(1) πjv = πju
∗ ∗
(2) πjv < πju
In the first case, we can construct a strategy as before in which Player Pj will still receive
the same payoff as if he played the strategy in which σj (r) = m (instead of σj (r) = m0 ). In
∗ ∗
the second case, the best payoff Player Pj can obtain is πjv < πju , so there is certainly no
reason for Player Pj to deviate and chose to define σj (r) = m0 . Thus, we have shown that
this new strategy is an equilibrium. Thus there is an equilibrium strategy for this tree of
height k + 1 and the proof follows by induction.
Example 4.53. We can illustrate the construction in the theorem with the Battle of
the Bismark Sea. In fact, you have already seen this construction once. Consider the game
tree in Figure 4.12: We construct the equilibrium solution from the bottom of the tree up.
Consider the vertex controlled by the Allies in which the Japanese sail north. In the sub-tree
below this node, the best move for the Allies is to search north (they receive the highest
payoff). This is highlighted in blue. Now consider the vertex controlled by the Allies where
the Japanese sail south. The best move for the Allies is to search south. Now, consider the
5. PAY-OFF FUNCTIONS AND EQUILIBRIA 45
N S
A A
N S N S
Figure 4.16. The game tree for the Battle of the Bismark Sea. If the Japanese sail
north, the best move for the Allies is to search north. If the Japanese sail south, then
the best move for the Allies is to search south. The Japanese, observing the payoffs,
note that given these best strategies for the Allies, there best course of action is to
sail North.
root node controlled by the Japanese. The Japanese can examine the two sub-trees below
this node and determine that the payoffs resulting from the equilibrium solutions in these
trees are −2 (from sailing north) and −3 (from sailing south). Naturally, the Japanese will
choose to so make the move of sailing north as this is the highest payoff they can achieve.
Thus the equilibrium strategy is shown in red and blue in the tree in Figure 4.16.
Exercise 29. Show that in Rock-Paper-Scissors with perfect information, there are three
equilibrium strategies.
Corollary 4.54 (Zermelo’s Theorem). Let G = (T, P, S, ν, µ, π) be a two-player game
with complete information and no chance. Assume that the payoff is such that:
(1) The only payoffs are +1 (win), −1 (lose).
(2) Player 1 wins +1 if and only if Player 2 wins −1.
(3) Player 2 wins +1 if and only if Player 1 wins −1.
Finally, assume that the players alternate turns. Then one of the two players must have a
strategy to obtain +1.
Exercise 30. Prove Zermelo’s Theorem. Can you illustrate a game of this type?[Hint:
Use Theorems 4.52 and 4.40. There are many games of this type.]
CHAPTER 5
Figure 5.1. In Chicken, two cars drive toward one another. The player who swerves
first loses 1 point, the other player wins 1 point. If both players swerve, then each
receives 0 points. If neither player swerves, a very bad crash occurs and both players
lose 10 points.
Player 2 are in the columns, then the two matrices for the players are:
Swerve Don’t Swerve Swerve Don’t Swerve
Swerve 0 -1 Swerve 0 1
Don’t Swerve 1 -10 Don’t Swerve -1 -10
From this we can see the matrices are:
0 −1
A1 =
1 −10
0 1
A2 =
−1 −10
Note that the Game of Chicken is not a zero-sum game, i.e. it is a general sum game.
Exercise 32. Construct payoff matrices for Rock-Paper-Scissors. Also construct the
normal form of the game.
Remark 5.11. Definition 5.7 can be extended to N player games. However, we no
longer have matrices with payoff values for various strategies. Instead, we construct N N -
dimensional arrays (or tensors). So a game with 3 players yields 3 arrays with dimension 3.
This is illustrated in Figure 5.2 Multidimensional arrays are easy to represent in computers,
but hard to represent on the page. They have multiple indices, instead of just 1 index like
a vector or 2 indices like a matrix. The elements of the array for Player i store the various
payoffs for Player i under different strategy combinations of the different players. If there
are three players, then there will be three different arrays, one for each player.
Remark 5.12. The normal form of a (two-player) game is essentially the recipe for
transforming a game in extensive form into a game in strategic form. Any game in extensive
form can be transformed in this way and the strategic form can be analyzed. Reasons for
doing this include the fact that the strategic form is substantially more compact. However,
it can be complex to compute if the size of the game tree in extensive form is very large.
Exercise 33. Compute the strategic form of the two-player Simple Poker game using
the expected payoff function defined in Example 4.49
50 5. NORMAL AND STRATEGIC FORM GAMES AND MATRICES
r 2
ye
Pla
Payoff Values
Figure 5.2. A three dimensional array is like a matrix with an extra dimension.
They are difficult to capture on a page. The elements of the array for Player i store
the various payoffs for Player i under different strategy combinations of the different
players. If there are three players, then there will be three different arrays.
Remark 5.14. We can apply Definition 5.13 to the case when x and y are column or
row vectors in the obvious way.
Definition 5.15 (Matrix Addition). If A and B are both in Rm×n , then C = A + B is
the matrix sum of A and B and
(5.5) Cij = Aij + Bij for i = 1, . . . , m and j = 1, . . . , n
Example 5.16.
1 2 5 6 1+5 2+6 6 8
(5.6) + = =
3 4 7 8 3+7 4+8 10 12
Definition 5.17 (Row/Column Vector). A 1 × n matrix is called a row vector, and a
m × 1 matrix is called a column vector. For the remainder of these notes, every vector will
be thought of column vector unless otherwise noted.
It should be clear that any row of matrix A could be considered a row vector in Rn and
any column of A could be considered a column vector in Rm .
Definition 5.18 (Matrix Multiplication). If A ∈ Rm×n and B ∈ Rn×p , then C = AB
is the matrix product of A and B and
(5.7) Cij = Ai· · B·j
Note, Ai· ∈ R1×n (an n-dimensional vector) and B·j ∈ Rn×1 (another n-dimensional vector),
thus making the dot product meaningful.
3. REVIEW OF BASIC MATRIX PROPERTIES 51
Example 5.19.
1 2 5 6 1(5) + 2(7) 1(6) + 2(8) 19 22
(5.8) = =
3 4 7 8 3(5) + 4(7) 3(6) + 4(8) 43 50
Definition 5.20 (Matrix Transpose). If A ∈ Rm×n is a m×n matrix, then the transpose
of A dented AT is an m × n matrix defined as:
(5.9) ATij = Aji
Example 5.21.
T
1 2 1 3
(5.10) =
3 4 2 4
The matrix transpose is a particularly useful operation and makes it easy to transform
column vectors into row vectors, which enables multiplication. For example, suppose x is
an n × 1 column vector (i.e., x is a vector in Rn ) and suppose y is an n × 1 column vector.
Then:
(5.11) x · y = xT y
Exercise 34. Let A, B ∈ Rm×n . Use the definitions of matrix addition and transpose
to prove that:
(5.12) (A + B)T = AT + BT
[Hint: If C = A + B, then Cij = Aij + Bij , the element in the (i, j) position of matrix C.
This element moves to the (j, i) position in the transpose. The (j, i) position of AT + BT is
ATji + BTji , but ATji = Aij . Reason from this point.]
Exercise 35. Let A, B ∈ Rm×n . Prove by example that AB 6= BA; that is, matrix
multiplication is not commutative. [Hint: Almost any pair of matrices you pick (that can be
multiplied) will not commute.]
Exercise 36. Let A ∈ Rm×n and let, B ∈ Rn×p . Use the definitions of matrix multipli-
cation and transpose to prove that:
(5.13) (AB)T = BT AT
[Hint: Use similar reasoning to the hint in Exercise 34. But this time, note that Cij = Ai· ·B·j ,
which moves to the (j, i) position. Now figure out what is in the (j, i) position of BT AT .]
Let A and B be two matrices with the same number of rows (so A ∈ Rm×n and B ∈
m×p
R ). Then the augmented matrix [A|B] is:
a11 a12 . . . a1n b11 b12 . . . b1p
a21 a22 . . . a2n b21 b22 . . . b2p
(5.14) ... ... .. .. ... ..
. . .
am1 am2 . . . amn bm1 bm2 . . . bmp
Thus, [A|B] is a matrix in Rm×(n+p) .
Example 5.22. Consider the following matrices:
1 2 7
A= , b=
3 4 8
52 5. NORMAL AND STRATEGIC FORM GAMES AND MATRICES
1 0 ... 0
0 1 ... 0
(5.15) In = ... . . ..
. .
0 0 ... 1
When it is clear from context, we may simply write I and omit the subscript n.
Exercise 38. Let A ∈ Rn×n . Show that AIn = In A = A. Hence, I is an identify for
the matrix multiplication operation on square matrices. [Hint: Do the multiplication out
long hand.]
Definition 5.24 (Standard Basis Vector). The standard basis vector ei ∈ Rn is:
ei = 0, 0, . . ., 1, 0, . . . , 0
| {z } | {z }
i−1 n−i−1
Note, this definition is only valid for n ≥ i. Further the standard basis vector ei is also the
ith row or column of In .
Definition 5.25 (Unit and Zero Vectors). The vector e ∈ Rn is the one vector e =
(1, 1, . . . , 1). Similarly, the zero vector 0 = (0, 0, . . . , 0) ∈ Rn . We assume that the length of
e and 0 will be determined from context.
Proof. For any matrix A ∈ Rm×n , Aec returns column c of matrix A, that is, A·c .
Likewise eTr A·c is the rth element of this vector. Thus, eTr Aec is the (r, c)th element of the
matrix A. By definition, this must be the payoff for the strategy pair (σr1 , σc2 ) for Player P1 .
A similar argument follows for Player P2 and matrix B.
Remark 5.28. What Proposition 5.27 says is that for two-player matrix games, we can
relate any choice of strategy that Player Pi makes with a unit vector. Thus, we can actually
define the payoff function in terms of vector and matrix multiplication. We will see that
this can be generalized to cases when the strategies of the players are not represented by
standard basis vectors.
Example 5.29. Consider the game of Chicken. Suppose Player P1 decides to swerve,
while Player P2 decides not to swerve. Then we can represent the strategy of Player P1 by
the vector:
1
e1 =
0
while the strategy of Player P2 is represented by the vector:
0
e2 =
1
Recall the payoff matrices for this game:
0 −1
A=
1 −10
0 1
B=
−1 −10
54 5. NORMAL AND STRATEGIC FORM GAMES AND MATRICES
Let us return to the notion of an equilibrium point for a two-player zero sum game. For
the remainder of this section, we will assume that Σ = Σ1 × Σ2 and Σ1 = {σ11 , . . . , σm
1
} and
2 2
Σ2 = {σ1 , . . . , σn }. Then any two-player zero-sum game in strategic form will be a tuple:
G = (P, Σ, A) with A ∈ Rm×n .
1. Saddle Points
Theorem 6.1. Let G = (P, Σ, A) be a zero-sum two player game. A strategy pair (ei , ej )
is an equilibrium strategy if and only if:
(6.1) eTi Aej = max min Akl = min max Akl
k∈{1,...,m} l∈{1,...,n} l∈{1,...,n} k∈{1,...,m}
Example 6.2. Before we prove Theorem 6.1, let’s first consider an example. This
example comes from [WV02] (Chapter 12). Two network corporations believe there are
100, 000, 000 viewers to be had during Thursday night, prime-time (8pm - 9pm). The corpo-
rations must decide which type of programming to run: Science Fiction, Drama or Comedy.
If the two networks initially split the 100, 000, 000 viewers evenly, we can think of the pay-
off matrix as determining how many excess viewers the networks’ strategies will yield over
50, 000, 000: The payoff matrix (in millions) for Network 1 is shown in Expression 6.2:
−15 −35 10
(6.2) A = −5 8 0
−12 −36 20
The expression:
min max Akl
l∈{1,...,n} k∈{1,...,m}
asks us to compute the maximum value in each column to create the set:
Cmax = {c∗l = max{Akl : k ∈ {1, . . . , m}} : l ∈ {1, . . . , n}}
and then choose the smallest value in this case. If we look at this matrix, the column
maximums are:
−5 8 20
We then choose the minimum value in this case and it is −5. This value occurs at position
(2, 1).
The expression
max min Akl
k∈{1,...,m} l∈{1,...,n}
55
56 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
asks us to compute the minimum value in each row to create the set:
Rmin = {rk∗ = min{Akl : l ∈ {1, . . . , n}} : k ∈ {1, . . . , m}}
and then choose the largest value in this case. Again, if we look at the matrix in Expression
6.2 we see that the minimum values in the rows are:
−35
−5
−36
The largest value in this case is −5. Again, this value occurs at position (2, 1).
Putting this all together, we get Figure 6.1:
Figure 6.1. The minimax analysis of the game of competing networks. The row
player knows that Player 2 (the column player) is trying to maximize her [Player
2’s] payoff. Thus, Player 1 asks: “What is the worst possible outcome I could see if I
played a strategy corresponding to this row?” Having obtained these worst possible
scenarios he chooses the row with the highest value. Player 2 does something similar
in columns.
Let’s try and understand why we would do this. The row player (Player 1) knows that
Player 2 (the column player) is trying to maximize her [Player 2’s] payoff. Since this is a
zero-sum game, any increase to Player 2’s payoff will come at the expense of Player 1. So
Player 1 looks at each row independently (since his strategy comes down to choosing a row)
and asks, “What is the worst possible outcome I could see if I played a strategy corresponding
to this row?” Having obtained these worst possible scenarios he chooses the row with the
highest value.
Player 2 faces a similar problem. She knows that Player 1 wishes to maximize his payoff
and that any gain will come at her expense. So Player 2 looks across each column of matrix
A and asks what is the best possible score Player 1 can achieve if I [Player 2] choose to play
the strategy corresponding to the given column. Remember, the negation of this value will
be Player 2’s payoff in this case. Having done that, Player 2 then chooses the column that
minimizes this value and thus maximizes her payoff.
If these two values are equal, then the theorem claims that the resulting strategy pair is
an equilibrium.
Exercise 40. Show that the strategy (e2 , e1 ) is an equilibrium for the game in Example
6.2. That is, show that the strategy (Drama, Science Fiction) is an equilibrium strategy for
the networks.
Exercise 41. Show that (Sail North, Search North) is an equilibrium solution for the
Battle of the Bismark Sea using the approach from Example 6.2 and Theorem 6.1.
1. SADDLE POINTS 57
Proof of Theorem 6.1. (⇒) Suppose that (ei , ej ) is an equilibrium solution. Then
we know that:
eTi Aej ≥ eTk Aej
eTi (−A)ej ≥ eTi (−A)el
for all k ∈ {1, . . . , m} and l ∈ {1, . . . , n}. We can obviously write this as:
(6.3) eTi Aej ≥ eTk Aej
and
(6.4) eTi Aej ≤ eTi Ael
We know that eTi Aej = Aij and that Equation 6.3 holds if and only if:
(6.5) Aij ≥ Akj
for all k ∈ {1, . . . , m}. From this we deduce that element i must be a maximal element in
column A·j . Based on this, we know that for each row k ∈ {1, . . . , m}:
(6.6) Aij ≥ min{Akl : l ∈ {1, . . . , n}}
To see this, note that for a fixed row k ∈ {1, . . . , m}:
Akj ≥ min{Akl : l ∈ {1, . . . , n}}
This means that if we compute the minimum value in a row k, then the value in column j,
Akj must be at least as large as that minimal value. But, Expression 6.6 implies that:
(6.7) eTi Aej = Aij = max min Akl
k∈{1,...,m} l∈{1,...,n}
Consider:
eTk Aej = Akj
58 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
implies that Aij ≥ Akj for any k ∈ {1, . . . , m}. To see this remember:
(6.11) Cmax = {c∗l = max{Akl : k ∈ {1, . . . , m}} : l ∈ {1, . . . , n}}
and Aij ∈ Cmax by construction. Thus it follows that:
eTi Aej ≥ eTk Aej
for any k ∈ {1, . . . , m}. By a similar argument we know that:
Aij = min max Akl
l∈{1,...,m} k∈{1,...,n}
implies that Aij ≤ Ail for any l ∈ {1, . . . , n}. To see this remember:
Rmin = {rk∗ = min{Akl : l ∈ {1, . . . , n}} : k ∈ {1, . . . , m}}
and Aij ∈ Rmin by construction. Thus it follows that:
eTi Aej ≤ eTi Ael
for any l ∈ {1, . . . , n}. Thus (ei , ej ) is an equilibrium solution. This completes the proof.
Theorem 6.3. Suppose that G = (P, Σ, A) be a zero-sum two player game. Let (ei , ej )
be an equilibrium strategy pair for this game. Show that if (ek , el ) is a second equilibrium
strategy pair, then
Aij = Akl = Ail = Akj
Exercise 42. Prove Theorem 6.3. [Hint: This proof is in Morris, Page 36.]
Definition 6.4 (Saddle Point). Let G = (P, Σ, A) be a zero-sum two player game. If
(ei , ej ) is an equilibrium, then it is called a saddle point.
Avranches
Figure 6.2. In August 1944, the allies broke out of their beachhead at Avranches
and started heading in toward the mainland of France. At this time, General Bradley
was in command of the Allied forces. He faced General von Kluge of the German
ninth army. Each commander faced several troop movement choices. These choices
can be modeled as a game.
We can see that the player set can be written as P = {Bradley, von Kluge}. The strategy
sets are:
Σ1 = {Reinforce the gap, Send forces east, Wait}
Σ2 = {Attack, Retreat}
In real life, there were no pay-off values (as there were in the Battle of the Bismark Sea),
however General Bradley’s diary indicates the scenarios he preferred in order. There are six
possible scenarios; i.e., there are six elements in Σ = Σ1 × Σ2 . Bradley ordered them from
most to least preferable and using this ranking, we can construct the game matrix shown in
Figure 6.3. Notice that the maximin value of the rows is not equal to the minimax value of
Figure 6.3. At the battle of Avranches General Bradley and General von Kluge
faced off over the advancing Allied Army. Each had decisions to make. This game
matrix shows that this game has no saddle point solution. There is no position in
the matrix where an element is simultaneously the maximum value in its column
and the minimum value in its row.
60 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
the columns. This is indicative of the fact that there is not a pair of strategies that form an
equilibrium for this game.
To see this, suppose that von Kluge plays his minimax strategy to retreat then Bradley
would do better not play his maximin strategy (wait) and instead move east, cutting of von
Kluge’s retreat, thus obtaining a payoff of (5, −5). But von Kluge would realize this and
deduce that he should attack, which would yield a payoff of (1, −1). However, Bradley could
deduce this as well and would know to play his maximin strategy (wait), which yields payoff
(6, −6). However, von Kluge would realize that this would occur in which case he would
decide to retreat yielding a payoff of (4, −4). The cycle then repeats. This is illustrated in
Figure 6.4.
Figure 6.4. When von Kluge chooses to retreat, Bradley can benefit by playing a
strategy different from his maximin strategy and he moves east. When Bradley does
this, von Kluge realizes he could benefit by attacking and not playing his maximin
strategy. Bradley realizes this and realizes he should play his maximin strategy and
wait. This causes von Kluge to realize that he should retreat, causing this cycle to
repeat.
Definition 6.7 (Game Value). Let G = (P, Σ, A) be a zero-sum game. If there exists a
strategy pair (ei , ej ) so that:
max min Akl = min max Akl
k∈{1,...,m} l∈{1,...,n} l∈{1,...,n} k∈{1,...,m}
then:
(6.12) VG = eTi Aej
is the value of the game.
Remark 6.8. We will see that we can define the value of a zero-sum game even when
there is no equilibrium point in strategies in Σ. Using Theorem 6.3 we can see that this value
is unique, that is any equilibrium pair for a game will yield the same value for a zero-sum
game. This is not the case in a general-sum game.
Exercise 43. Show that Rock-Paper-Scissors does not have a saddle-point strategy.
3. Mixed Strategies
Heretofore we have assumed that Player Pi will deterministically choose a strategy in Σi .
It’s possible, however, that Player Pi might choose a strategy at random. In this case, we
assign probability to each strategy in Σi .
3. MIXED STRATEGIES 61
Definition 6.9 (Mixed Strategy). Let G = (P, Σ, π) be a game in normal form with
P = {P1 , . . . , PN }. A mixed strategy for Player Pi ∈ P is a discrete probability distribution
function ρi defined over the sample space Σ. That is, we can define a discrete probability
space (Σi , FΣi , ρi ) where Σi is the discrete sample space, FΣi is the power set of Σi and ρi is
the discrete probability function that assigns probabilities to events in FΣi .
Remark 6.10. We assume that players choose their mixed strategies independently.
Thus we can compute the probability of a strategy element (σ 1 , . . . , σ N ) ∈ Σ as:
(6.13) ρ(σ 1 , . . . , σ N ) = ρ1 (σ 1 )ρ2 (σ 2 ) · · · ρN (σ n )
Using this, we can define a discrete probability distribution over the sample space Σ as:
(Σ, FΣ , ρ). Define Πi as a random variable that maps Σ into R so that Πi returns the payoff
to Player Pi as a result of the random outcome (σ 1 , . . . , σ N ). Therefore, the expected payoff
for Player Pi for a given mixed strategy (ρ1 , . . . , ρN ) is given as:
X X X
E(Πi ) = ··· πi (σ 1 , . . . , σ n )ρ1 (σ 1 )ρ2 (σ 2 ) · · · ρN (σ N )
σ 1 ∈Σ1 σ 2 ∈Σ2 σ N ∈ΣN
Example 6.11. Consider the Rock-Paper-Scissors Game. The payoff matrix for Player
1 is given in Figure 6.5: Suppose that each strategy is chosen with probability 31 by each
Figure 6.5. The payoff matrix for Player P1 in Rock-Paper-Scissors. This payoff
matrix can be derived from Figure 4.5.
player. Then the expected payoff to Player P1 with this strategy is:
1 1 1 1
E(π1 ) = π1 (Rock, Rock) + π1 (Rock, Paper)+
3 3 3 3
1 1 1 1
π1 (Rock, Scissors) + π1 (Paper, Rock)+
3 3 3 3
1 1 1 1
π1 (Paper, Paper) + π1 (Paper, Scissors)+
3 3 3 3
1 1 1 1
π1 (Scissors, Rock) + π1 (Scissors, Paper)+
3 3 3 3
1 1
π1 (Scissors, Scissors) = 0
3 3
We can likewise compute the same value for E(π2 ) for Player P2 .
3.1. Mixed Strategy Vectors.
Definition 6.12 (Mixed Strategy Vector). Let G = (P, Σ, π) be a game in normal form
with P = {P1 , . . . , PN }. Let Σi = {σ1i , . . . , σni i }. To any mixed strategy for Player Pi we
may associate a vector xi = [xi1 , . . . , xini ]T provided that it satisfies the properties:
62 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
x3
1
∆3 = Face of a tetrahedron
1 1
x1 x2
Figure 6.6. In three dimensional space ∆3 is the face of a tetrahedron. In four
dimensional space, it would be a tetrahedron, which would itself be the face of a
four dimensional object.
Definition 6.15 (Pure Strategy). Let Σi be the strategy set for Player Pi in a game. If
Σi = {σ1i , . . . , σni i }, then ej ∈ ∆ni (for j = 1, . . . , ni ). These standard basis vectors are the
pure strategies in ∆ni and ej corresponds to a pure strategy choice σji ∈ Σi .
Definition 6.16 (Mixed Strategy Space). Let G = (P, Σ, π) be a game in normal form
with P = {P1 , . . . , PN }. Let Σi = {σ1i , . . . , σni i }. Then the mixed strategy space for the game
G is the set:
(6.15) ∆ = ∆n1 × ∆n2 × · · · × ∆nN
4. MIXED STRATEGIES IN MATRIX GAMES 63
Here xji
is the ith element of vector xj . The function ui : ∆ → R defined in Equation 6.16
is the mixed strategy payoff function for Player Pi . (Note: This notation is adapted from
[Wei97].)
Example 6.18. For Rock-Paper-Scissors, since each player has 3 strategies, n = 3 and
∆3 consists of those vectors [x1 , x2 , x3 ]T so that x1 , x2 , x3 ≥ 0 and x1 + x2 + x3 = 1. For
example, the vectors:
1
3
x = y = 13
1
3
are mixed strategies for Players 1 and 2 respectively that instruct the players to play rock
1/3 of the time, paper 1/3 of the time and scissors 1/3 of the time.
Definition 6.19 (Nash Equilibrium). Let G = (P, Σ, π) be a game in normal form with
P = {P1 , . . . , PN }. Let Σi = {σ1i , . . . , σni i }. A Nash equilibrium is a tuple of mixed strategies
∗ ∗
(x1 , . . . , xN ) ∈ ∆ so that for all i = 1, . . . , N :
∗ ∗ ∗ ∗ ∗
(6.17) ui (x1 , . . . , xi , . . . , xN ) ≥ ui (x1 , . . . , xi , . . . , xN )
for all xi ∈ ∆ni
∗ ∗
Remark 6.20. What Definition 6.19 says is that a tuple of mixed strategies (x1 , . . . , xN )
is a Nash equilibrium if no player has any reason to deviate unilaterally from her mixed
strategy.
Remark 6.21 (Notational Remark). In many texts, it becomes cumbersome in N player
games to denote the mixed strategy tuple (x1 , . . . , xN ) especially when (as in Definition 6.19)
you are only interested in one player (Player Pi ). To deal with this, textbooks sometimes
adopt the notation (xi , x−i ). Here xi is the mixed strategy for Player Pi ) while x−i denotes
the mixed strategy tuple for the other Players (who are not Player Pi ). When expressed this
way, Equation 6.17 is written as:
∗ ∗ ∗
ui (xi , x−i ) ≥ ui (xi , x−i )
for all i = 1, . . . , N . While notationally convenient, we will restrict our attention to two
player games, so this will generally not be necessary.
4. Mixed Strategies in Matrix Games
Proposition 6.22. Let G = (P, Σ, A, B) be a two-player matrix game. Let Σ = Σ1 × Σ2
where Σ1 = {σ11 , . . . , σm
1
} and Σ2 = {σ12 , . . . , σn2 }. Let x ∈ ∆m and y ∈ ∆n be mixed strategies
for Players 1 and 2 respectively. Then:
(6.18) u1 (x, y) = xT Ay
(6.19) u2 (x, y) = xT By
64 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
That is, xT A is a row vector whose j th element is xT A·j . For fixed j we have:
m
X
T
x A·j = x1 A1j + x2 A2j + · · · + xm Amj = π1 (σi1 , σj2 )xi
i=1
yn
This simplifies to:
(6.20) xT A·1 y1 + · · · + xT A·n yn =
(x1 A11 + x2 A21 + · · · + xm Am1 ) y1 + · · · + (x1 A1n + x2 A2n + · · · + xm Amn ) ym
Distributing multiplication through, we can simplify Equation 6.20 as:
Xm X n m X
X n
T
(6.21) x Ay = Aij xi yj = π1 (σi1 , σj2 )xi yj = u1 (x, y)
i=1 j=1 i=1 j=1
A similar argument shows that u2 (x, y) = xT By. This completes the proof.
Exercise 44. Show explicitly that u2 (x, y) = xT By as we did in the previous proof.
Remark 6.26. In a two player matrix game G = (P, Σ, A, B) with A, B ∈ Rm×n a mixed
strategy x ∈ ∆m for Player 1 weakly dominates a strategy y ∈ ∆m if for all z ∈ ∆n (mixed
strategies for Player 2) we have:
(6.24) xT Az ≥ yT Az
and the inequality is strict for at least one z ∈ ∆n . If x strictly dominates y then we have:
(6.25) xT Az > yT Az
for all z ∈ ∆n .
Exercise 45. For a two player matrix game, write what it means for a strategy y ∈ ∆n
for Player 2 to weakly dominate a strategy x. Also write what it means if y strictly dominates
x. [Hint: Remember, Player 2 multiplies on the right hand side of the payoff matrix. Also,
you’ll need to use B.]
Example 6.27 (Prisoner’s Dilemma). The following example is called Prisoner’s Dilemma
and is a classic example in Game Theory. Two prisoner’s Bonnie and Clyde commit a bank
robbery. They stash the cash and are driving around wondering what to do next when they
are pulled over and arrested for a weapons violation. The police suspect Bonnie and Clyde
of the bank robbery, but do not have any hard evidence. They separate the prisoners and
offer them the following options to Bonnie:
(1) If neither Bonnie nor Clyde confess, they will go to prison for 1 year on the weapons
violation.
(2) If Bonnie confesses, but Clyde does not, then Bonnie can go free while Clyde will
go to jail for 10 years.
(3) If Clyde confesses and Bonnie does not, then Bonnie will go to jail for 10 years while
Clyde will go free.
(4) If both Bonnie and Clyde confess, then they will go to jail for 5 years.
A similar offer is made to Clyde. The following two-player matrix game describes the sce-
nario: P = {Bonnie, Clyde}; Σ1 = Σ2 = {Don’t Confess, Confess}. The matrices for this
game are given below:
−1 −10
A=
0 −5
−1 0
B=
−10 −5
Here payoffs are given in negative years (for years lost to prison). Bonnie’s matrix is A and
Clyde’s matrix is B. The rows (columns) correspond to the strategies Don’t Confess and
Confess. Thus, we see that if Bonnie does not confess and Clyde does (row 1, column 2),
then Bonnie loses 10 years and Clyde loses 0 years.
We can show that the strategy Confess dominates Don’t Confess for Bonnie. Pure strate-
gies correspond to standard basis vectors. Thus we’re claiming that e2 strictly dominates e1
for Bonnie. We can use remark 6.26 to see that we must show:
(6.26) eT2 Az > eT1 Az
We know that z is a mixed strategy. That means that:
z
z= 1
z2
66 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
5z − 5
9z − 10
Figure 6.7. To show that Confess dominates over Don’t Confess in Prisoner’s
dilemma for Bonnie, we can compute e1 T Az and e2 Az for any arbitrary mixed
strategy z for Clyde. The resulting payoff to Bonnie is 5z − 5 when she confesses
and 9z − 10 when she doesn’t confess. Here z is the probability that Clyde will
not confess. The fact that 5z − 5 is greater than 9z − 10 at every point in the
domain z ∈ [0, 1] demonstrates that Confess dominates Don’t Confess for Bonnie.
Exercise 46. Show that Confess strictly dominates Don’t Confess for Clyde in Example
6.27.
5. DOMINATED STRATEGIES AND NASH EQUILIBRIA 67
Remark 6.28. Strict dominance can be extremely useful for identifying pure Nash equi-
libria. This is especially true in matrix games. This is summarized in the following two
theorems.
Theorem 6.29. Let G = (P, Σ, A, B) be a two player matrix game with A, B ∈ Rm×n .
If
(6.27) eTi Aek > eTj Aek
for k = 1, . . . , n, then ei strictly dominates ej for Player 1.
Remark 6.30. We know that eTi A is the ith row of A. Theorem 6.29 says: if every
element in Ai· (the ith row of A) is greater than its corresponding element in Aj· , (the j th
row of A), then Player 1’s ith strategy strictly dominates Player 1’s j th strategy.
Proof. For all k = 1, . . . , n we know that:
eTi Aek > eTj Aek
Suppose that z1 , . . . , zn ∈ [0, 1] with z1 + · · · + zn = 1. Then for each zk we know that:
eTi Aek zk > eTj Aek zk
for k = 1, . . . , n. This implies that:
eTi Ae1 z1 + · · · + eTi Aen zn > eTj Ae1 z1 + · · · + eTj Aen zn
Factoring we have:
eTi A (z1 e1 + · · · + zn en ) > eTj A (z1 e1 + · · · + zn en )
Define:
z1
z = z1 e1 + · · · + zn en = ...
zn
Since the original z1 , . . . , zn where chosen arbitrarily from [0, 1] so that z1 + . . . zn = 1, we
know that:
eTi Az > eTj Az
for all z ∈ ∆n . Thus ei strictly dominates ej by Definition 6.24.
Remark 6.31. There is an analogous theorem for Player 2 which states that if each
element of a column B·i is greater than the corresponding element in column B·j , then ei
strictly dominates strategy ej for Player 2.
Exercise 47. Using Theorem 6.29, state and prove an analogous theorem for Player 2.
Remark 6.32. Theorem 6.29 can be generalized to N players. Unfortunately, the no-
tation becomes complex and is outside the scope of this set of notes. It is worth knowing,
however, that this is the case.
Theorem 6.33. Let G = (P, Σ, A, B) be a two player matrix game. Suppose pure strategy
ej ∈ ∆m for Player 1 is strictly dominated by pure strategy ei ∈ ∆m . If (x∗ , y∗ ) is a Nash
equilibrium, then x∗j = 0. Similarly, if pure strategy ej ∈ ∆n for Player 2 is strictly dominated
by pure strategy ei ∈ ∆n , then yj∗ = 0
68 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
Proof. We will prove the theorem for Player 1; the proof for Player 2 is completely
analogous. We will proceed by contradiction. Suppose that x∗j > 0. We know:
eTi Ay∗ > e∗j Ay∗
because ei strictly dominates ej . We can express:
x∗ T Ay = x∗1 eT1 + · · · + x∗i eTi + · · · + x∗j eTj + · · · + x∗m eTm Ay∗
(6.28)
Here x∗i is the ith element of vector x∗ . Since x∗j > 0 we know that:
x∗j eTi Ay∗ > x∗j e∗j Ay∗
Thus we can conclude that:
x∗1 eT1 + · · · + x∗i eTi + · · · + x∗j eTi + · · · + x∗m eTm Ay∗ >
(6.29)
x∗1 eT1 + · · · + x∗i eTi + · · · + x∗j eTj + · · · + x∗m eTm Ay∗
If we define z ∈ ∆m so that:
∗ ∗
xi + xj k = i
(6.30) zk = 0 k=j
x
k else
Then Equation 6.29 implies:
(6.31) zT Ay∗ > x∗ T Ay∗
Thus, (x∗ , y∗ ) could not have been a Nash equilibrium. This completes the proof.
Example 6.34. We can use the two previous theorems to our advantage. Consider the
Prisoner’s Dilemma (Example 6.27). The payoff matrices (again) are:
−1 −10
A=
0 −5
−1 0
B=
−10 −5
For Bonnie Row (Strategy) 1 is strictly dominated by Row (Strategy) 2. Thus Bonnie will
never play Strategy 1 (Don’t Confess) in a Nash equilibrium. That is:
A1· < A2· ≡ −1 −10 < 0 −5
Thus, we can consider a new game in which we remove this strategy for Bonnie (since Bonnie
will never play this strategy). The new game has P = {Bonnie, Clyde}, Σ1 = {Confess},
Σ2 = {Don’t Confess, Confess}. The new game matrices are:
A0 = 0 −5
B0 = −10 −5
In this new game, we note that for Clyde (Player 2) Column (Strategy) 2 strictly dominates
Column (Strategy 1). That is:
B0·1 < B0·2 ≡ −10 < −5
6. THE MINIMAX THEOREM 69
Clyde will never play Strategy 1 (Don’t Confess) in a Nash equilibrium. We can construct a
new game with P = {Bonnie, Clyde}, Σ1 = {Confess}, Σ2 = {Confess} and (trivial) payoff
matrices:
A00 = −5
B00 = −5
In this game, there is only one Nash equilibrium in which both players confess. And this
equilibrium is the Nash equilibrium of the original game.
Remark 6.35 (Iterative Dominance). A game whose Nash equilibrium is computed using
the method from Example 6.34 in which strictly dominated are iteratively eliminated for the
two players is said to be solved by iterative dominance. A game that can be analyzed in this
way is said to be strictly dominance solvable.
Exercise 48. Consider the game matrix (matrices) 6.2. Show that this game is strictly
dominance solvable. Recall that the game matrix is:
−15 −35 10
A = −5 8 0
−12 −36 20
[Hint: Start with Player 2 (the Column Player) instead of Player 1. Note that Column 3
is strictly dominated by Column 1, so you can remove Column 3. Go from there. You can
eliminate two rows (or columns) at a time if you want.]
That is, given x ∈ ∆m , we choose a vector y that minimizes xT Ay. This value is the best
possible result Player 1 can expect if he announces to Player 2 that he will play strategy x.
Player 1 then faces the problem that he would like to maximize this value by choosing x
appropriately. That is, Player 1 hopes to solve the problem:
(6.33) max v1 (x)
x∈∆m
70 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
Thus we have:
(6.34) max v1 (x) = max min xT Ay
x∈∆m x y
That is, given y ∈ ∆n , we choose a vector x that maximizes xT Ay. This value is the
best possible result that Player 2 can expect if she announces to Player 1 that she will play
strategy y. Player 2 then faces the problem that she would like to minimize this value by
choosing y appropriately. That is, Player 2 hopes to solve the problem:
(6.36) min v2 (y)
y∈∆n
Thus we have:
(6.37) min v2 (y) = min max xT Ay
y∈∆n y x
Note that this is the precise analogy in mixed strategies to the concept of a saddle point. The
functions v1 and v2 are called the value functions for Player 1 and 2 respectively. The main
problem we must tackle now is to determine whether these maximization and minimization
problems can be solved.
Lemma 6.38. Let G = (P, Σ, A) be a zero-sum game with A ∈ Rm×n . Then:
(6.38) max v1 (x) ≤ min v2 (y)
x∈∆m y∈∆n
Exercise 49. Prove Lemma 6.38. [Hint: Argue that for all x ∈ ∆m and for all y ∈ ∆n
we know that v1 (x) ≤ v2 (y) by showing that v2 (y) ≥ xT Ay ≥ v1 (x). From this conclude
that miny v2 (y) ≥ maxx v1 (x).]
Theorem 6.39. Let G = (P, Σ, A) be a zero-sum game with A ∈ Rm×n . Then the
following are equivalent:
(1) There is a Nash equilibrium (x∗ , y∗ ) for G
(2) The following equation holds:
(6.39) v1 = max min xT Ay = min max xT Ay = v2
x y y x
implies that:
x∗ T Ay∗ = max xT Ay∗
x
Thus we have:
v2 = min max xT Ay ≤ max xT Ay∗ = x∗ T Ay∗
y x x
Thus:
v2 = min max xT Ay ≤ max xT Ay∗ = x∗ T Ay∗ = min x∗ T Ay
y x x y
when we let v1 = maxx miny x∗ T Ay. By Lemma 6.38 we know that v1 ≤ v2 . Thus we have
v2 ≤ v1 and v1 ≤ v2 so v1 = v2 as required.
(2 =⇒ 3): Let v = v1 = v2 and let x∗ be the vector that solves maxx v1 (x) and y∗ be
the vector that solves miny v2 (y). For fixed j we know:
X
Aij x∗i = x∗ T Aej
i
y = ...
yn
we can conclude that:
(6.42) x∗ T Ay ≥ v
for any y ∈ ∆n . By a similar argument we know that:
(6.43) xT Ay∗ ≤ v
for all x ∈ ∆m . From Equation 6.43 we conclude that:
(6.44) x∗ T Ay∗ ≤ v
and from Equation 6.42 we conclude that:
(6.45) x∗ T Ay∗ ≥ v
Thus v = x∗ T Ay∗ and we know for all x and y:
x∗ T Ay∗ ≥ xT Ay∗
x∗ T Ay∗ ≤ x∗ T Ay
Thus (x∗ , y∗ ) is a Nash equilibrium. This completes the proof.
Remark 6.40. Theorem 6.39 does not assert the existence of a Nash equilibrium, it just
provides insight into what happens if one exists. In particular, we know that the game has
a unique value:
(6.46) v = max min xT Ay = min max xT Ay
x y y x
Proving the existence of a Nash equilibrium can be accomplished in several ways, the oldest
of which uses a topological argument, which we present next. We can also use a linear
programming based argument, which we will explore in the next chapter.
6. THE MINIMAX THEOREM 73
Lemma 6.41 (Brouwer Fixed Point Theorem). Let ∆ be the mixed strategy space of a
two-player zero sum game. If T : ∆ → ∆ is continuous, then there exists a pair of strategies
(x∗ , y∗ ) so that T (x∗ , y∗ ) = (x∗ , y∗ ). That is (x∗ , y∗ ) is a fixed point of the mapping T .
Remark 6.42. The proof of Brouwer’s Fixed Point Theorem is well outside the scope of
these notes. It is a deep theorem in topology. The interested reader should consult [Mun00]
(Page 351 - 353).
Theorem 6.43 (Minimax Theorem). Let G = (P, Σ, A) be a zero-sum game with A ∈
Rm×n . Then there is a Nash equilibrium (x∗ , y∗ ).
Nash’s Proof. (A version of this proof is given in [LR89], Appendix 2.) Let (x, y) ∈ ∆
be mixed strategies for Players 1 and 2. Define the following:
(
eTi Ay − xT Ay if this quantity is positive
(6.47) ci (x, y) =
0 else
(
xT Ay − xT Aej if this quantity is positive
(6.48) dj (x, y) =
0 else
Let T : ∆ → ∆ where T (x, y) = (x0 , y0 ) so that for i = 1, . . . , m we have:
xi + ci (x, y)
(6.49) x0i =
1+ m
P
k=1 ck (x, y)
and for j = 1, . . . , n we have:
yj + dj (x, y)
(6.50) yj0 =
1 + nk=1 dk (x, y)
P
P
Since i xi = 1 we know that:
x1 + · · · + xm + m
P
k=1 ck (x, y)
(6.51) x01 + ··· + x0m = Pm =1
1 + k=1 ck (x, y)
It is also clear that since xi ≥ 0 for P i= 1, . . . , m we have x0i ≥ 0. A similar argument shows
0 0
that yj ≥ 0 for j = 1, . . . , n and j yj = 1. Thus T is a proper map from ∆ to ∆. The fact
that T is continuous follows from the continuity of the payoff function (See Exercise 51). We
now show that (x, y) is a Nash equilibrium if and only if it is a fixed point of T .
To see this note that ci (x, y) measures the amount that the pure strategy ei is better
than x as a response to y. That is, if Player 2 decides to play strategy y then ci (x, y)
tells us if and how much playing pure strategy ei is better than playing x ∈ ∆m . Similarly,
dj (x, y) measures how much better ej is as a response to Player 1’s strategy x than strategy
y for Player 2. Suppose that (x, y) is a Nash equilibrim. Then ci (x, y) = 0 = dj (x, y) for
i = 1, . . . , m and j = 1, . . . , n by the definition of equilibrium. Thus x0i = xi for i = 1, . . . , m
and yj0 = yj for j = 1, . . . , n and thus (x, y) is a fixed point of T .
To show the converse, suppose that (x, y) is a fixed point of T . It suffices to show that
there is at least one i so that xi > 0 and ci (x, y) = 0. Clearly there is at least one i for
which xi > 0. Note that:
Xm
T
x Ay = xi eTi Ay
i=1
74 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
Thus, xT Ay < eTi Ay cannot hold for all i = 1, . . . , m with xi > 0 (otherwise the previous
equation would not hold). Thus for at least one i with xi > 0 we must have ci (x, y) = 0.
But for this i, the fact that (x, y) is a fixed point implies that:
x
(6.52) xi = Pm i
1 + k=1 ck (x, y)
Pm
This implies that k=1 ck (x, y) = 0. The fact that ck (x, y) ≥ 0 for all k = 1, . . . , m
implies that ck (x, y) = 0. A similar argument can be shown for y. Thus we know that
ci (x, y) = 0 = dj (x, y) for i = 1, . . . , m and j = 1, . . . , n and thus x is at least as good a
strategy for Player 1 responding to y as any ei ∈ ∆m ; likewise y is at least as good a strategy
for Player 2 responding to x as any ej ∈ ∆n . This fact implies that (x, y) is an equilibrium
(see Exercise 50) for details).
Applying Lemma 6.41 (Brouwer’s Fixed Point Theorem) we see that T must have a fixed
point and thus every two player zero sum game has a Nash equilibrium. This completes the
proof.
Exercise 50. Prove the following: G = (P, Σ, A) be a zero-sum game with A ∈ Rm×n .
Let x∗ ∈ ∆m and y∗ ∈ ∆n . If:
x∗ T Ay∗ ≥ eTi Ay∗
for all i = 1, . . . , m and
x∗ T Ay∗ ≤ x∗ T Aej
for all j = 1, . . . , n, then (x∗ , y∗ ) is an equilibrium.
Exercise 51. Verify that the function T in Theorem 6.43 is continuous.
7. Finding Nash Equilibria in Simple Games
It is relatively straightforward to find a Nash equilibrium in 2 × 2 zero-sum games,
assuming that a saddle-point cannot be identified using the approach from Example 6.2. We
illustrate the approach using The Battle of Avranches.
Example 6.44. Consider the Battle of Avranches (Example 6.6). The payoff matrix is:
2 3
A = 1 5
6 4
Note first that Row 1 (Bradley’ first strategy) is strictly dominated by Row 3 (Bradley’s
third strategy) and thus we can reduce the payoff matrix to:
1 5
A=
6 4
Let’s suppose that Bradley chooses a strategy:
x
x=
1−x
with x ∈ [0, 1]. If Von Kluge chooses to Attack (Column 1), then Bradley’s expected payoff
will be:
T
1 5 1
x Ae1 = x 1 − x = x + 6(1 − x) = −5x + 6
6 4 0
7. FINDING NASH EQUILIBRIA IN SIMPLE GAMES 75
A similar argument shows that if Von Kluge chooses to Retreat (Column 2), then Bradley’s
expected payoff will be:
xT Ae2 = 5x + 4(1 − x) = x + 4
We can visualize these strategies by plotting them (see Figure 6.8, left)). Plotting the
y
−5x + 6
x+4
1
x=
3
1
y=
6
−2y − 4
4y − 5
Figure 6.8. Plotting the expected payoff to Bradley by playing a mixed strategy
[x (1 − x)]T when Von Kluge plays pure strategies shows which strategy Von Kluge
should pick. When x ≤ 1/3, Von Kluge does better if he retreats because x + 4 is
below −5x + 6. On the other hand, if x ≥ 1/3, then Von Kluge does better if he
attacks because −5x + 6 is below x + 4. Remember, Von Kluge wants to minimize
the payoff to Bradley. The point at which Bradley does best (i.e., maximizes his
expected payoff) comes at x = 1/3. By a similar argument, when y ≤ 1/6, Bradley
does better if he choose Row 1 (Move East) while when y ≥ 1/6, Bradley does best
when he waits. Remember, Bradley is minimizing Von Kluge’s payoff (since we are
working with −A).
expected payoff to Bradley by playing a mixed strategy [x (1 − x)]T when Von Kluge plays
pure strategies shows which strategy Von Kluge should pick. When x ≤ 1/3, Von Kluge
does better if he retreats because x + 4 is below −5x + 6. That is, the best Bradley can hope
to get is −5x + 6 if he announced to Von Kluge that he was playing x ≤ 1/3.
On the other hand, if x ≥ 1/3, then Von Kluge does better if he attacks because −5x + 6
is below x + 4. That is, the best Bradley can hope to get is x + 4 if he tells Von Kluge that
he is playing x ≥ 1/3. Remember, Von Kluge wants to minimize the payoff to Bradley. The
point at which Bradley does best (i.e., maximizes his expected payoff) comes at x = 1/3.
By a similar argument, we can compute the expected payoff to Von Kluge when he plays
mixed strategy [y (1 − y)]T and Bradley plays pure strategies. The expected payoff to Von
Kluge when Bradley plays Row 1, is:
eT1 (−A)y = −y − 5(1 − y) = 4y − 5
When Bradley plays Row 2, the expected payoff to Von Kluge is:
eT2 (−A)y = −6y − 4(1 − y) = −2y − 4
We can plot these expressions (see Figure 6.8, right). When y ≤ 1/6, Bradley does better
if he choose Row 1 (Move East) while when y ≥ 1/6, Bradley does best when he waits.
76 6. SADDLE POINTS, MIXED STRATEGIES AND THE MINIMAX THEOREM
Remember, Bradley is minimizing Von Kluge’s payoff (since we are working with −A). We
know that Bradley cannot do any better than when he plays x∗ = [1/3 2/3]T . Similarly,
Von Kluge cannot do any better than when he plays y∗ = [1/6 5/6]T . The pair (x∗ , y∗ ) is
the Nash equilibrium for this problem.
Often, any Nash equilibrium for a zero-sum game is called a saddle-point. To see why we
called these points saddle points, consider Figure 6.9. This figure shows the payoff function
for Player 1 as a function of x and y (from the example). This function is:
1 5 y
(6.53) x 1−x = −6yx + 2y + x + 4
6 4 1−y
The figure is a hyperbolic saddle. In 3D space, it looks like a twisted combination of an
upside down parabola (like the plot of y = −x2 from high school algebra) and a right-side
up parabola (like y = x2 from high school algebra). Note that the maximum of one parabola
and minimum of another parabola occur precisely at the point (x, y) = (1/3, 1/5), the point
in 2D space corresponding to this Nash equilibrium.
Figure 6.9. The payoff function for Player 1 as a function of x and y. Notice that
the Nash equilibrium does in fact occur at a saddle point.
Exercise 52. Consider the football game in Example 5.4. Ignoring the Blitz option
for the defense, compute the Nash equilibrium strategy in terms of Running Plays, Passing
Plays, Running Defense and Passing Defense.
Remark 6.45. The techniques discussed in Example 6.44 can be extended to cases when
one player has 2 strategies and another player has more than 2 strategies, but these methods
are not efficient for finding Nash equilibria in general. In the next chapter we will show how
to find Nash equilibria for games by finding solving a specific simple optimization problem.
This technique will work for general two player zero-sum games. We will also discuss the
problem of finding Nash equilibria in two player general sum matrix games.
produce the values in these functions are called best replies and are used in proving the
existence of Nash equilibria for general sum N player games.
Definition 6.47 (Player Best Response). Let G = (P, Σ, π) be an N player game in
normal form with Σi = {σ1i , . . . , σni i } and let ∆ be the mixed strategy space for this game.
If y ∈ ∆ is a mixed strategy for all players, then the best reply for Player Pi is the set:
(6.54) Bi (y) = xi ∈ ∆ni : ui (xi , y−i ) ≥ ui (zi , y−i ) ∀zi ∈ ∆ni
∗ ∗
Thus xi ∈ Bi (xi ). Since this holds for each i = 1, . . . , N it follows that x∗ ∈ B(x∗ ).
To prove the converse, suppose that x∗ ∈ B(x∗ ). Then for all i = 1, . . . , N :
∗
xi ∈ xi ∈ ∆ni : ui (xi , x−i ) ≥ ui (z, y−i ) ∀z ∈ ∆ni
Remark 6.53. The proof based on Kakutani’s Fixed Point Theorem is neither useful
nor satisfying. Nash realized this and constructed an alternate proof using Brouwer’s Fixed
Point theorem following the same steps we used to prove Theorem 6.43. We can generalize
the proof of Theorem 6.43 by defining:
(6.56) Jki (x) = max 0, ui (ek , x−i ) − ui (xi , x−i )
The function Jki (x) measures the benefit of changing to the pure strategy ek for Player Pi
when all other players hold their strategy fixed at x−i .
We can now define:
0 xij + Jji (x)
(6.57) xij =
1 + nk=1
P i i
Jk (x)
Using this equation, we can construct a mapping T : ∆ → ∆ and show that every fixed
point of T is a Nash Equilibrium. Using the Brouwer fixed point theorem, it then follows
that a Nash equilibrium exists. Unfortunately, this is still not a very useful way to construct
a Nash equilibrium.
In the next chapter we will explore this problem in depth for two player zero-sum games
and then go on to explore the problem for two player general sum-games. The story of
computing Nash equilibria takes on a life of its own and is an important study within com-
putational game theory that has had a substantial impact on the literature in mathematical
programming (optimization), computer science, and economics.
CHAPTER 7
In this chapter we’re going to take a detour into optimization theory. We’ll need many
of these results and definitions later when we tackle methods for solving two player zero and
general sum games. Optimization is an exciting sub-discipline within applied mathematics!
Optimization is all about making things better; this could mean helping a company make
better decisions to maximize profit; helping a factory make products with less environmental
impact; or helping a zoologist improve the diet of an animal. When we talk about optimiza-
tion, we often use terms like better or improvement. It’s important to remember that words
like better can mean more of something (as in the case of profit) or less of something as in
the case of waste. As we study linear programming, we’ll quantify these terms in a mathe-
matically precise way. For the time being, let’s agree that when we optimize something we
are trying to make some decisions that will make it better.
Example 7.1. Let’s recall a simple optimization problem from differential calculus (Math
140): Goats are an environmentally friendly and inexpensive way to control a lawn when
there are lots of rocks or lots of hills. (Seriously, both Google and some U.S. Navy bases use
goats on rocky hills instead of paying lawn mowers!)
Suppose I wish to build a pen to keep some goats. I have 100 meters of fencing and I
wish to build the pen in a rectangle with the largest possible area. How long should the sides
of the rectangle be? In this case, making the pen better means making it have the largest
possible area.
The problem is illustrated in Figure 7.1. Clearly, we know that:
Goat Pen y
Figure 7.1. Goat pen with unknown side lengths. The objective is to identify the
values of x and y that maximize the area of the pen (and thus the number of goats
that can be kept).
(7.1) 2x + 2y = 100
79
80 7. AN INTRODUCTION TO OPTIMIZATION AND THE KARUSH-KUHN-TUCKER CONDITIONS
because 2x + 2y is the perimeter of the pen and I have 100 meters of fencing to build my
pen. The area of the pen is A(x, y) = xy. We can use Equation 7.1 to solve for x in terms
of y. Thus we have:
(7.2) y = 50 − x
and A(x) = x(50 − x). To maximize A(x), recall we take the first derivative of A(x) with
respect to x, set this derivative to zero and solve for x:
dA
(7.3) = 50 − 2x = 0;
dx
Thus, x = 25 and y = 50 − x = 25. We further recall from basic calculus how to confirm
that this is a maximum; note:
d2 A
(7.4) = −2 < 0
dx2
x=25
Which implies that x = 25 is a local maximum for this function. Another way of seeing this
is to note that A(x) = 50x − x2 is an “upside-down” parabola. As we could have guessed, a
square will maximize the area available for holding goats.
Exercise 53. A canning company is producing canned corn for the holidays. They
have determined that each family prefers to purchase their corn in units of 12 fluid ounces.
Assuming that metal costs 1 cent per square inch and 1 fluid ounce is about 1.8 cubic inches,
compute the ideal height and radius for a can of corn assuming that cost is to be minimized.
[Hint: Suppose that our can has radius r and height h. The formula for the surface area of
a can is 2πrh + 2πr2 . Since metal is priced by the square inch, the cost is a function of the
surface area. The volume of the can is πr2 h and is constrained. Use the same trick we did
in the example to find the values of r and h that minimize cost.
problems, we can have many constraints. The set of all points in Rn for which the constraints
are true is called the feasible set (or feasible region). Our problem is to decide the best values
of x and y to maximize the area A(x, y). The variables x and y are called decision variables.
Let z : D ⊆ Rn → R; for i = 1, . . . , m, gi : D ⊆ Rn → R; and for j = 1, . . . , l
hj : D ⊆ Rn → R be functions. Then the general maximization problem with objec-
tive function z(x1 , . . . , xn ) and inequality constraints gi (x1 , . . . , xn ) ≤ bi (i = 1, . . . , m) and
equality constraints hj (x1 , . . . , xn ) = rj (j = 1, . . . , s)is written as:
max z(x1 , . . . , xn )
s.t. g1 (x1 , . . . , xn ) ≤ b1
..
.
(7.5) gm (x1 , . . . , xn ) ≤ bm
h1 (x1 , . . . , xn ) = r1
..
.
hl (x1 , . . . , xn ) = rl
Expression 7.5 is also called a mathematical programming problem. Naturally when con-
straints are involved we define the global and local maxima for the objective function
z(x1 , . . . , xn ) in terms of the feasible region instead of the entire domain of z, since we
are only concerned with values of x1 , . . . , xn that satisfy our constraints.
Example 7.3 (Continuation of Example 7.1). We can re-write the problem in Example
7.1:
max A(x, y) = xy
s.t. 2x + 2y = 100
(7.6)
x≥0
y≥0
Note we’ve added two inequality constraints x ≥ 0 and y ≥ 0 because it doesn’t really make
any sense to have negative lengths. We can re-write these constraints as −x ≤ 0 and −y ≤ 0
where g1 (x, y) = −x and g2 (x, y) = −y to make Expression 7.6 look like Expression 7.5.
We have formulated the general maximization problem in Proble 7.5. Suppose that we
are interested in finding a value that minimizes an objective function z(x1 , . . . , xn ) subject
to certain constraints. Then we can write Problem 7.5 replacing max with min.
Exercise 55. Write the problem from Exercise 53 as a general minimization problem.
Add any appropriate non-negativity constraints. [Hint: You must change max to min.]
An alternative way of dealing with minimization is to transform a minimization prob-
lem into a maximization problem. If we want to minimize z(x1 , . . . , xn ), we can maximize
−z(x1 , . . . , xn ). In maximizing the negation of the objective function, we are actually finding
a value that minimizes z(x1 , . . . , xn ).
Exercise 56. Prove the following statement: Consider Problem 7.5 with the objective
function z(x1 , . . . , xn ) replaced by −z(x1 , . . . , xn ). Then the solution to this new problem
minimizes z(x1 , . . . , xn ) subject to the constraints of Problem 7.5.[Hint: Use the definition of
82 7. AN INTRODUCTION TO OPTIMIZATION AND THE KARUSH-KUHN-TUCKER CONDITIONS
global maximum and a multiplication by −1. Be careful with the direction of the inequality
when you multiply by −1.]
When z : D ⊆ R → R, the graph is precisely what you’d expect. It’s the set of pairs
(x, y) ∈ R2 so that y = z(x). This is the graph that you learned about back in Algebra 1.
Definition 7.6 (Level Set). Let z : Rn → R be a function and let c ∈ R. Then the level
set of value c for function z is the set:
(7.10) {x = (x1 , . . . , xn ) ∈ Rn |z(x) = c} ⊆ Rn
2. SOME GEOMETRY FOR OPTIMIZATION 83
Example 7.7. Consider the function z = x2 + y 2 . The level set of z at 4 is the set of
points (x, y) ∈ R2 such that:
(7.11) x2 + y 2 = 4
You will recognize this as the equation for a circle with radius 4. We illustrate this in the
following two figures. Figure 7.2 shows the level sets of z as they sit on the 3D plot of the
function, while Figure 7.3 shows the level sets of z in R2 . The plot in Figure 7.3 is called a
contour plot.
Level Set
Figure 7.2. Plot with Level Sets Projected on the Graph of z. The level sets
existing in R2 while the graph of z existing R3 . The level sets have been projected
onto their appropriate heights on the graph.
Level Set
Figure 7.3. Contour Plot of z = x2 + y 2 . The circles in R2 are the level sets of the
function. The lighter the circle hue, the higher the value of c that defines the level
set.
Definition 7.8. (Line) Let x0 , v ∈ Rn . Then the line defined by vectors x0 and v is
the function l(t) = x0 + tv. Clearly l : R → Rn . The vector v is called the direction of the
line.
Example 7.9. Let x0 = (2, 1) and let v = (2, 2). Then the line defined by x0 and v
is shown in Figure 7.4. The set of points on this line is the set L = {(x, y) ∈ R2 : x =
2 + 2t, y = 1 + 2t, t ∈ R}.
Definition 7.10 (Directional Derivative). Let z : Rn → R and let v ∈ Rn be a vector
(direction) in n-dimensional space. Then the directional derivative of z at point x0 ∈ Rn in
84 7. AN INTRODUCTION TO OPTIMIZATION AND THE KARUSH-KUHN-TUCKER CONDITIONS
Figure 7.4. A Line Function: The points in the graph shown in this figure are in
the set produced using the expression x0 + vt where x0 = (2, 1) and let v = (2, 2).
the direction of v is
d
(7.12) z(x0 + tv)
dt t=0
when this derivative exists.
Proposition 7.11. The directional derivative of z at x0 in the direction v is equal to:
z(x0 + hv) − z(x0 )
(7.13) lim
h→0 h
Exercise 58. Prove Proposition 7.11. [Hint: Use the definition of derivative for a
univariate function and apply it to the definition of directional derivative and evaluate t = 0.]
Definition 7.12 (Gradient). Let z : Rn → R be function and let x0 ∈ Rn . Then the
gradient of z at x0 is the vector in Rn given by:
∂z ∂z
(7.14) ∇z(x0 ) = (x0 ), . . . , (x0 )
∂x1 ∂xn
Gradients are extremely important concepts in optimization (and vector calculus in gen-
eral). Gradients have many useful properties that can be exploited. The relationship between
the directional derivative and the gradient is of critical importance.
Theorem 7.13. If z : Rn → R is differentiable, then all directional derivatives exist.
Furthermore, the directional derivative of z at x0 in the direction of v is given by:
(7.15) ∇z(x0 ) · v
where · denotes the dot product of two vectors.
Proof. Let l(t) = x0 + vt. Then l(t) = (l1 (t), . . . , ln (t)); that is, l(t) is a vector function
whose ith component is given by li (t) = x0i + vi t.
Apply the chain rule:
dz(l(t)) ∂z dl1 ∂z dln
(7.16) = + ··· +
dt ∂l1 dt ∂ln dt
2. SOME GEOMETRY FOR OPTIMIZATION 85
Thus:
d dl
(7.17) z(l(t)) = ∇z ·
dt dt
Clearly dl/dt = v. We have l(0) = x0 . Thus:
d
(7.18) z(x0 + tv) = ∇z(x0 ) · v
dt t=0
We now come to the two most important results about gradients, (i) the fact that they
always point in the direction of steepest ascent with respect to the level curves of a function
and (ii) that they are perpendicular (normal) to the level curves of a function. We can
exploit this fact as we seek to maximize (or minimize) functions.
Theorem 7.14. Let z : Rn → R be differentiable and let x0 ∈ Rn . If ∇z(x0 ) 6= 0, then
∇z(x0 ) points in the direction in which z is increasing fastest.
Proof. Recall ∇z(x0 ) · n is the directional derivative of z in direction n at x0 . Assume
that n is a unit vector. We know that:
(7.19) ∇z(x0 ) · n = ||∇z(x0 )|| cos θ
where θ is the angle between the vectors ∇z(x0 ) and n. The function cos θ is largest when
θ = 0, that is when n and ∇z(x0 ) are parallel vectors. (If ∇z(x0 ) = 0, then the directional
derivative is zero in all directions.)
Theorem 7.15. Let z : Rn → R be differentiable and let x0 lie in the level set S defined
by z(x) = k for fixed k ∈ R. Then ∇z(x0 ) is normal to the set S in the sense that if v
is a tangent vector at t = 0 of a path c(t) contained entirely in S with c(0) = x0 , then
∇z(x0 ) · v = 0.
Before giving the proof, we illustrate this theorem in Figure 7.5. The function is z(x, y) =
x + y 2 + 2xy and x0 = (1, 1). At this point ∇z(x0 ) = (6, 4).
4
Proof. As stated, let c(t) be a curve in S. Then c : R → Rn and z(c(t)) = k for all
t ∈ R. Let v be the tangent vector to c at t = 0; that is:
dc(t)
(7.20) =v
dt t=0
Differentiating z(c(t)) with respect to t using the chain rule and evaluating at t = 0 yields:
d
(7.21) z(c(t)) = ∇z(c(0)) · v = ∇z(x0 ) · v = 0
dt t=0
Thus ∇z(x0 ) is perpendicular to v and thus normal to the set S as required.
Exercise 59. In this exercise you will use elementary calculus (and a little bit of vector
algebra) to show that the gradient of a simple function is perpendicular to its level sets:
(a): Plot the level sets of z(x, y) = x2 + y 2 . Draw the gradient at the point (x, y) =
(2, 0). Convince yourself that it is normal to the level set x2 + y 2 = 4.
(b): Now, choose any level set x2 + y 2 = k. Use implicit differentiation to find dy/dx.
This is the slope of a tangent line to the circle x2 + y 2 = k. Let (x0 , y0 ) be a point
on this circle.
86 7. AN INTRODUCTION TO OPTIMIZATION AND THE KARUSH-KUHN-TUCKER CONDITIONS
Figure 7.5. A Level Curve Plot with Gradient Vector: We’ve scaled the gradient
vector in this case to make the picture understandable. Note that the gradient
is perpendicular to the level set curve at the point (1, 1), where the gradient was
evaluated. You can also note that the gradient is pointing in the direction of steepest
ascent of z(x, y).
(c): Find an expression for a vector parallel to the tangent line at (x0 , y0 ) [Hint: you
can use the slope you just found.]
(d): Compute the gradient of z at (x0 , y0 ) and use it and the vector expression you just
computed to show that two vectors are perpendicular. [Hint: use the dot product.]
Figure 7.6. Level Curves and Feasible Region: At optimality the level curve of the
objective function is tangent to the binding constraints.
which is just a scaled version of the gradient of the objective function. Thus the gradient
of the objective function is just a dilation of gradient of the binding constraint. This is
illustrated in Figure 7.7.
Figure 7.7. Gradients of the Binding Constraint and Objective: At optimality the
gradient of the binding constraints and the objective function are scaled versions of
each other.
The elements illustrated in the previous example are true in general. You may have
discussed a simple example of these when you talked about Lagrange Multipliers in Vector
Calculus (Math 230/231). We’ll revisit these concepts when we discuss 7.31.
Exercise 60. Plot the level sets of the objective function and the feasible region in
Exercise 53. At the point of optimality you identified, show that the gradient of the objective
function is a scaled version of the gradient (linear combination) of the binding constraints.
88 7. AN INTRODUCTION TO OPTIMIZATION AND THE KARUSH-KUHN-TUCKER CONDITIONS
then
m
X
(7.24) x= λi xi
i=1
x1
x1 x2 x2
X X
Figure 7.8. Examples of Convex Sets: The set on the left (an ellipse and its
interior) is a convex set; every pair of points inside the ellipse can be connected by
a line contained entirely in the ellipse. The set on the right is clearly not convex as
we’ve illustrated two points whose connecting line is not contained inside the set.
be the set formed from the intersection of these sets. Choose x1 , x2 ∈ C and λ ∈ [0, 1].
Consider x = λx1 + (1 − λ)x2 . We know that x1 , x2 ∈ C1 , . . . , Cn by definition of C. By
convexity, we know that x ∈ C1 , . . . , Cn by convexity of each set. Therefore, x ∈ C. Thus
C is a convex set.
f (λx1 + (1 − λ)x2 )
Figure 7.9. A convex function: A convex function satisfies the expression f (λx1 +
(1 − λ)x2 ) ≤ λf (x1 ) + (1 − λ)f (x2 ) for all x1 and x2 and λ ∈ [0, 1].
can be shown to be equivalent to the definition you learned in Calculus I (Math 140) using
first and second derivatives.
Definition 7.26 (Concave Function). A function f : Rn → R is a convex function if it
satisfies:
(7.27) f (λx1 + (1 − λ)x2 ) ≥ λf (x1 ) + (1 − λ)f (x2 )
for all x1 , x2 ∈ Rn and for all λ ∈ [0, 1].
90 7. AN INTRODUCTION TO OPTIMIZATION AND THE KARUSH-KUHN-TUCKER CONDITIONS
To visualize this definition, simply flip Figure 7.9 upside down. The following theorem
is a powerful tool that can be used to show sets are convex. It’s proof is outside the scope
of the class, but relatively easy.
(7.28) z(x1 , . . . , xn ) = c1 x1 + · · · + cn xn
Example 7.29. We have had experience with many linear functions already. The left-
hand-side of the constraint 2x + 2y ≤ 100 is a linear function. That is the function z(x, y) =
2x + 2y is a linear function of x and y.
Exercise 62. Prove that every affine function is both convex and concave.
6. Karush-Kuhn-Tucker Conditions
It turns out there is a very powerful theorem that discusses when a point x∗ ∈ Rn will
maximize a function. The following is the Kuhn-Karush-Tucker theorem, which we will
state, but not prove.
Remark 7.35. This theorem holds as a necessary condition even if z(x) is not concave
or the functions gi (x) (i = 1, . . . , m) are not convex or the functions hj (x) (j = 1, . . . , l) are
not linear. In this case though, the fact that a triple: (x, λ, µ) ∈ Rn × Rm × Rl does not
ensure that this is an optimal solution for Problem P .
Remark 7.36. Looking more closely at the dual feasibility conditions, we see something
interesting. Suppose that there are no equality constraints (i.e., not constraints of the form
hj (x) = 0). Then the statements:
m
X l
X
∇z(x∗ ) − λi ∇gi (x∗ ) − µj ∇hj (x∗ ) = 0
i=1 j=1
λi ≥ 0 for i = 1, . . . , m
imply that:
m
X
∇z(x∗ ) = λi ∇gi (x∗ )
i=1
λi ≥ 0 for i = 1, . . . , m
Specifically, this says that the gradient of z at x∗ is a positive combination of the gradients
of the constraints at x∗ . But more importantly, since we also have complementary slackness,
we know that if gi (x∗ ) 6= 0, then λi = 0 because λi gi (x∗ ) = 0 for i = 1, . . . , m. Thus, what
dual feasibility is really saying is that gradient of z at x∗ is a positive combination of the
gradients of the binding constraints at x∗ . Remember, a constraint is binding if gi (x∗ ) = 0,
in which case λi ≥ 0.
Remark 7.37. Continuing from the previous remark, in the general case when we have
some equality constraints, then dual feasibility says:
m
X l
X
∗ ∗
∇z(x ) = λi ∇gi (x ) + µj ∇hj (x∗ )
i=1 j=1
λi ≥ 0 for i = 1, . . . , m
µj ∈ R for j = 1, . . . , l
Since equality constraints are always binding this says that the gradient of z at x∗ is a linear
combination of the gradients of the binding constraints at x∗ .
Example 7.38. We’ll finish the example we started with Example 7.1. Let’s rephrase
this optimization problem in the form we saw in the theorem: We’ll have:
max A(x, y) = xy
s.t. 2x + 2y − 100 = 0
(7.29)
−x≤0
−y ≤0
Note that the greater-than inequalities x ≥ 0 and y ≥ 0 in Expression 7.6 have been changes
to less-than inequalities by multiplying by −1. The constraints 2x + 2y = 100 has simply
been transformed to 2x + 2y − 100 = 0. Thus, if h(x, y) = 2x + 2y − 100, we can see
h(x, y) = 0 is our constraint. We can let g1 (x, y) = −x and g2 (x, y) = −y. Then we have
g1 (x, y) ≤ 0 and g2 (x, y) ≤ 0 as our inequality constraints. We already know that x = y = 25
7. RELATING BACK TO GAME THEORY 93
is our optimal solution. Thus we know that there must be Lagrange multipliers µ, λ1 and
λ2 corresponding to the constraints h(x, y) =, g1 (x, y) ≤ 0 and g2 (x, y) ≤ 0 that satisfy the
KKT conditions.
Let’s investigate the three components of the KKT conditions.
Primal Feasibility: If x = y = 25, then h(x, y) = 2x + 2y − 100 and clearly
h(25, 25) = 0. Further g1 (x, y) = −x and g2 (x, y) = −y then g1 (25, 25) = −25 ≤ 0
and g2 (25, 25) = −25 ≤ 0. So primal feasibility is satisfied.
Complementary Slackness: We know that g1 (x, y) = g2 (x, y) = −25. Since neither
of these functions is 0, we know that λ1 = λ2 = 0. This will force complementary
slackness, namely:
λ1 g1 (25, 25) = 0
λ2 g2 (25, 25) = 0
Dual Feasibility: We already know that λ1 = λ2 = 0. That means we need to find
µ ∈ R so that:
∇A(25, 25) − µ∇h(25, 25) = 0
We know that:
y
∇A(x, y) = ∇xy =
x
2
∇h(x, y) = ∇(2x + 2y − 100) =
2
Evaluating ∇A(25, 25) yields:
25 2 0
−µ =
25 2 0
Thus setting µ = 25/2 will accomplish our goal.
Exercise 63. Find the values of the dual variables for the optimal point in Exercise 53.
Show that the KKT conditions hold for the values you found.
This is a mathematical programming problem, provided that ui (xi , x−i ) is known. However,
it assumes that all other players are holding their strategy constant e.g., playing x−i . The
interesting part (and the part that makes Game Theory hard) is that each player is solving
this problem simultaneously. Thus an equilibrium solution is a simultaneous solution to:
i −i
max ui (x , x )
(7.31) ∀i : s.t. xi1 + · · · + xini = 1
xij ≥ 0 j = 1, . . . , ni
1. Linear Programs
When both the objective and all the constraints in Expression 7.5 are linear functions,
then the optimization problem is called a linear programming problem. This has the general
form:
max z(x1 , . . . , xn ) = c1 x1 + · · · + cn xn
s.t. a11 x1 + · · · + a1n xn ≤ b1
..
.
(8.1) am1 x1 + · · · + amn xn ≤ bm
h11 x1 + · · · + hn1 xn = r1
..
.
hl1 x1 + · · · + hln xn = rl
Example 8.1. Consider the problem of a toy company that produces toy planes and toy
boats. The toy company can sell its planes for $10 and its boats for $8 dollars. It costs $3
in raw materials to make a plane and $2 in raw materials to make a boat. A plane requires
3 hours to make and 1 hour to finish while a boat requires 1 hour to make and 2 hours to
finish. The toy company knows it will not sell anymore than 35 planes per week. Further,
given the number of workers, the company cannot spend anymore than 160 hours per week
finishing toys and 120 hours per week making toys. The company wishes to maximize the
profit it makes by choosing how much of each toy to produce.
We can represent the profit maximization problem of the company as a linear program-
ming problem. Let x1 be the number of planes the company will produce and let x2 be
the number of boats the company will produce. The profit for each plane is $10 − $3 = $7
per plane and the profit for each boat is $8 − $2 = $6 per boat. Thus the total profit the
company will make is:
(8.2) z(x1 , x2 ) = 7x1 + 6x2
The company can spend no more than 120 hours per week making toys and since a plane
takes 3 hours to make and a boat takes 1 hour to make we have:
(8.3) 3x1 + x2 ≤ 120
Likewise, the company can spend no more than 160 hours per week finishing toys and since
it takes 1 hour to finish a plane and 2 hour to finish a boat we have:
(8.4) x1 + 2x2 ≤ 160
95
96 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
Finally, we know that x1 ≤ 35, since the company will make no more than 35 planes per
week. Thus the complete linear programming problem is given as:
max z(x1 , x2 ) = 7x1 + 6x2
s.t. 3x1 + x2 ≤ 120
x1 + 2x2 ≤ 160
(8.5)
x1 ≤ 35
x1 ≥ 0
x2 ≥ 0
Remark 8.2. Strictly speaking, the linear programming problem in Example 8.1 is not a
true linear programming problem because we don’t want to manufacture a fractional number
of boats or planes and therefore x1 and x2 must really be drawn from the integers and not
the real numbers (a requirement for a linear programming problem). This type of problem
is generally called an integer programming problem. However, we will ignore this fact and
assume that we can indeed manufacture a fractional number of boats and planes. If you’re
interested in this distinction, you might consider taking Math 484, where we discuss this
issue in depth.
Exercise 64. A chemical manufacturer produces three chemicals: A, B and C. These
chemical are produced by two processes: 1 and 2. Running process 1 for 1 hour costs $4 and
yields 3 units of chemical A, 1 unit of chemical B and 1 unit of chemical C. Running process 2
for 1 hour costs $1 and produces 1 units of chemical A, and 1 unit of chemical B (but none of
Chemical C). To meet customer demand, at least 10 units of chemical A, 5 units of chemical
B and 3 units of chemical C must be produced daily. Assume that the chemical manufacturer
wants to minimize the cost of production. Develop a linear programming problem describing
the constraints and objectives of the chemical manufacturer. [Hint: Let x1 be the amount
of time Process 1 is executed and let x2 be amount of time Process 2 is executed. Use the
coefficients above to express the cost of running Process 1 for x1 time and Process 2 for x2
time. Do the same to compute the amount of chemicals A, B, and C that are produced.]
3x1 + x2 = 120
x1 + 2x2 = 160
Figure 8.1. Feasible Region and Level Curves of the Objective Function: The
shaded region in the plot is the feasible region and represents the intersection of
the five inequalities constraining the values of x1 and x2 . On the right, we see the
optimal solution is the “last” point in the feasible region that intersects a level set
as we move in the direction of increasing profit.
This is a set of parallel lines with slope −7/6 and intercept c/6 where c can be varied as
needed. The level curves for various values of c are parallel lines. In Figure 8.1 they are
shown in colors ranging from red to yellow depending upon the value of c. Larger values of
c are more yellow.
To solve the linear programming problem, follow the level sets along the gradient (shown
as the black arrow) until the last level set (line) intersects the feasible region. If you are
doing this by hand, you can draw a single line of the form 7x1 + 6x2 = c and then simply
draw parallel lines in the direction of the gradient (7, 6). At some point, these lines will fail
to intersect the feasible region. The last line to intersect the feasible region will do so at a
point that maximizes the profit. In this case, the point that maximizes z(x1 , x2 ) = 7x1 +6x2 ,
subject to the constraints given, is (x∗1 , x∗2 ) = (16, 72).
Note the point of optimality (x∗1 , x∗2 ) = (16, 72) is at a corner of the feasible region. This
corner is formed by the intersection of the two lines: 3x1 + x2 = 120 and x1 + 2x2 = 160. In
this case, the constraints
3x1 + x2 ≤ 120
x1 + 2x2 ≤ 160
are both binding, while the other constraints are non-binding. In general, we will see that
when an optimal solution to a linear programming problem exists, it will always be at the
intersection of several binding constraints; that is, it will occur at a corner of a higher-
dimensional polyhedron.
2.1. KKT Conditions for Linear Programs. As with any mathematical program-
ming problem, we can derive the Karush-Kuhn-Tucker conditions for the a linear program-
ming problem. We’ll illustrate this by deriving the KKT conditions for Example 8.1. Note
98 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
since linear (affine) functions are both convex and concave functions, we know that finding
a Lagrange multipliers satisfying the KKT conditions is necessary and sufficient for proving
that a point is an optimal point.
Example 8.4. Let z(x1 , x2 ) = 7x1 + 6x2 , the objective function in Problem 8.5. We have
argued that the point of optimality is (x∗1 , x∗2 ) = (16, 72). The KKT conditions for Problem
8.5 are:
Primal Feasibility:
Lagrange Multiplier
g1 (x∗1 , x∗2 ) = 3x∗1 + x∗2 − 120 ≤ 0 (λ1 )
∗ ∗ ∗ ∗
g2 (x , x ) = x + 2x − 160 ≤ 0
(λ2 )
1 2 1 2
(8.6)
g3 (x∗1 , x∗2 ) = x∗1 − 35 ≤ 0 (λ3 )
∗ ∗ ∗
g4 (x1 , x2 ) = −x1 ≤ 0 (λ4 )
∗ ∗ ∗
g5 (x1 , x2 ) = −x2 ≤ 0 (λ5 )
Dual Feasibility:
5
∇z(x∗ , x∗ ) −
X
∗ ∗ 0
λi ∇gi (x1 , x2 ) =
1 2 0
(8.7) i=1
λi ≥ 0 i = 1, . . . , 5
Complementary Slackness:
(8.8) {λi gi (x∗1 , x∗2 ) = 0 i = 1, . . . , 5
We have [0 0]T in our dual feasible conditions because the gradients of our functions will all
be two-dimensional vectors (there are two variables). Specifically, we can compute
(1) ∇z(x∗1 , x∗2 ) = [7 6]T
(2) ∇g1 (x∗1 , x∗2 ) = [3 1]T
(3) ∇g2 (x∗1 , x∗2 ) = [1 2]T
(4) ∇g3 (x∗1 , x∗2 ) = [1 0]T
(5) ∇g4 (x∗1 , x∗2 ) = [−1 0]T
(6) ∇g5 (x∗1 , x∗2 ) = [0 − 1]T
Notice that g3 (16, 72) = 16 − 35 = −17 6= 0. This means that for complementary
slackness to be satisfied we must have λ2 = 0. The the same reasoning, λ4 = 0 because
g4 (16, 72) = −16 6= 0 and λ5 = 0 because g5 (16, 72) = −72 6= 0. Thus, dual feasibility can
be simplified to:
7 −λ 3 −λ 1 = 0
1 2
(8.9) 6 1 2 0
λi ≥ 0 i = 1, . . . , 5
This is just a set of linear equations (with some non-negativity constraints, which we’ll
ignore). We have:
(8.10) 7 − 3λ1 − λ2 = 0 =⇒ 3λ1 + λ2 = 7
(8.11) 6 − λ1 − 2λ2 = 0 =⇒ λ1 + 2λ2 = 6
2. INTUITION ON THE SOLUTION OF LINEAR PROGRAMS 99
We can solve these linear equations (and hope that the solution is positive). Doing so yields:
8
(8.12) λ1 =
5
11
(8.13) λ2 =
5
Thus we have found a KKT point:
x∗1 = 16
x∗2 = 72
8
λ1 =
5
(8.14) 11
λ2 =
5
λ3 = 0
λ4 = 0
λ5 = 0
This proves (via Theorem 7.31) that the point we found graphically is in fact the optimal
solution to the Problem 8.5.
2.2. Problems with an Infinite Number of Solutions. We’ll study a specific lin-
ear programming problem with an infinite number of solutions by modifying the objective
function in Example 8.1.
Example 8.5. Suppose the toy maker in Example 8.1 finds that it can sell planes for a
profit of $18 each instead of $7 each. The new linear programming problem becomes:
max z(x1 , x2 ) = 18x1 + 6x2
s.t. 3x1 + x2 ≤ 120
x1 + 2x2 ≤ 160
(8.15)
x1 ≤ 35
x1 ≥ 0
x2 ≥ 0
Applying our graphical method for finding optimal solutions to linear programming problems
yields the plot shown in Figure 8.2. The level curves for the function z(x1 , x2 ) = 18x1 + 6x2
are parallel to one face of the polygon boundary of the feasible region. Hence, as we move
further up and to the right in the direction of the gradient (corresponding to larger and
larger values of z(x1 , x2 )) we see that there is not one point on the boundary of the feasible
region that intersects that level set with greatest value, but instead a side of the polygon
boundary described by the line 3x1 + x2 = 120 where x1 ∈ [16, 35]. Let:
S = {(x1 , x2 |3x1 + x2 ≤ 120, x1 + 2x2 ≤ 160, x1 ≤ 35, x1 , x2 ≥ 0}
that is, S is the feasible region of the problem. Then for any value of x∗1 ∈ [16, 35] and any
value x∗2 so that 3x∗1 + x∗2 = 120, we will have z(x∗1 , x∗2 ) ≥ z(x1 , x2 ) for all (x1 , x2 ) ∈ S. Since
there are infinitely many values that x1 and x2 may take on, we see this problem has an
infinite number of alternative optimal solutions.
100 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
Exercise 65. Modify the linear programming problem from Exercise 64 to obtain a
linear programming problem with an infinite number of alternative optimal solutions. Solve
the new problem and obtain a description for the set of alternative optimal solutions. [Hint:
Just as in the example, x1 will be bound between two value corresponding to a side of the
polygon. Find those values and the constraint that is binding. This will provide you with a
description of the form for any x∗1 ∈ [a, b] and x∗2 is chosen so that cx∗1 + dx∗2 = v, the point
(x∗1 , x∗2 ) is an alternative optimal solution to the problem. Now you fill in values for a, b, c,
d and v.]
2.3. Other Possibilities. In addition to the two scenarios above in which a linear
programming problem has a unique solution or an infinite number of alternative optimal
solutions, it is also possible that a linear programming problem can have:
(1) No solution, which occurs when the feasible region is empty,
(2) An unbounded solution, which can occur if the feasible region is an unbounded set.
Fortunately, we will not encounter either of those situations in our study of zero-sum games
and so we blissfully ignore these possibilities.
∗ ∗
Pexists a∗ real number v and x ∈ ∆m and y ∈ ∆n so that:
(3) There
(a) Pi Aij xi ≥ v for j = 1, . . . , n and
(b) j Aij yj∗ ≤ v for i = 1, . . . , m
The fact that x∗ ∈ ∆m implies that:
(8.17) x∗1 + · · · + x∗ m = 1
In this set of constraints we have m + 1 variables: x∗1 , . . . , x∗m and v, the value of the game.
We know that Player 1 (the row player) is a value maximizer, therefore Player 1 is interested
in solving the linear programming problem:
max v
s.t. A11 x1 + · · · + Am1 xm − v ≥ 0
A12 x1 + · · · + Am2 xm − v ≥ 0
(8.19) ..
.
A1n x1 + · · · + Amn xm − v ≥ 0
x1 + · · · + xm = 1
xi ≥ 0 i = 1, . . . , m
We know that Player 2 (the column player) is a value minimizer, therefore Player 2 is
interested in solving the linear programming problem:
min v
s.t. A11 y1 + · · · + A1n yn − v ≤ 0
A21 y1 + · · · + A2n yn − v ≤ 0
(8.21) ..
.
Am1 y1 + · · · + Amn yn − v ≤ 0
y1 + · · · + yn = 1
yi ≥ 0 i = 1, . . . , n
Example 8.6. Consider the game from Example 6.2. The payoff matrix for Player 1 is
given as:
−15 −35 10
A = −5 8 0
−12 −36 20
This is a zero sum game, so the payoff matrix for Player 2 is simply the negation of this
matrix. The linear programming problem for Player 1 is:
max v
s.t. − 15x1 − 5x2 − 12x3 − v ≥ 0
− 35x1 + 8x2 − 36x3 − v ≥ 0
(8.22)
10x1 + 20x3 − v ≥ 0
x1 + x2 + x3 = 1
x1 , x2 , x3 ≥ 0
Notice, we simply work our way down each column of the matrix A in forming the constraints
of the linear programming problem. To form the problem for Player 2, we work our way
across the rows of A and obtain:
min v
s.t. − 15y1 − 35y2 + 10y3 − v ≤ 0
− 5y1 + 8y2 − v ≤ 0
(8.23)
− 12y1 − 36y2 + 20y3 − v ≤ 0
y1 + y2 + y3 = 1
y1 , y2 , y3 ≥ 0
Exercise 66. Construct the two linear programming problems for Bradley and von
Kluge in the Battle of Avranches.
4. MATRIX NOTATION, SLACK AND SURPLUS VARIABLES FOR LINEAR PROGRAMMING 103
Example 8.7. Consider a zero-sum game with payoff matrix A ∈ Rm×n . We can write
the problem that arises for Player 1 in matrix notation. The decision variables are x ∈ Rm×1
and v ∈ R. We can write these decision variables as a single vector z:
x
z=
v
Let:
0
0
.
..
c=
0
1
Then our objective function is cT z = v. Our inequality constraints have the form:
T
A |−e z≥0
Here e = [1, 1, . . . , 1]T is a column vector of ones with n elements to make the augmented
matrix meaningful. Our equality constraints are x1 + · · · + xm = 1. This can be written as:
T
e |0 z = 1
104 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
Again, e is an appropriately sized vector of ones (this time with m elements). The resulting
linear program is then:
max cT z
s.t. AT | − e z ≥ 0
T
e |0 z = 1
eTi z ≥ 0 i = 1, . . . , m
The last constraint simply says that xi ≥ 0 and since v is the m + 1st variable, we do not
constraint v to be positive.
Exercise 67. Construct the matrix form of the linear program for Player 2 in a zero-sum
game.
4.1. Standard Form, Slack and Surplus Variables.
Definition 8.8 (Standard Form). A linear programming problem is in standard form if
it is written as:
T
max z(x) =c x
(8.28) s.t. Ax = b
x≥0
Remark 8.9. It is relatively easy to convert any inequality constraint into an equality
constraint. Consider the inequality constraint:
(8.29) ai1 x1 + ai2 x2 + · · · + ain xn ≤ bi
We can add a new slack variable si to this constraint to obtain:
ai1 x1 + ai2 x2 + · · · + ain xn + si = bi
Obviously this slack variable si ≥ 0. The slack variable then becomes just another variable
whose value we must discover as we solve the linear program for which Expression 8.29 is a
constraint.
We can deal with constraints of the form:
(8.30) ai1 x1 + ai2 x2 + · · · + ain xn ≥ bi
in a similar way. In this case we subtract a surplus variable si to obtain:
ai1 x1 + ai2 x2 + · · · + ain xn − si = bi
Again, we must have si ≥ 0.
Example 8.10. Consider the linear programming problem:
max z(x1 , x2 ) = 2x1 − x2
s.t. x − x ≤ 1
1 2
2x1 + x2 ≥ 6
x1 , x2 ≥ 0
5. SOLVING LINEAR PROGRAMS BY COMPUTER 105
This linear programming problem can be put into standard form by using both a slack and
surplus variable.
Here, l is a vector of lower bounds for the vector x and u is a vector of upper bounds for the
vector x.
In Matlab, almost all input is in the form of matrices. Thus we enter the vector for the
objective function c, the matrix A and vector b for constraints of the form Ax ≤ b, a matrix
H and vector r for constraints of the form Hx = r and finally the two vectors l and u for
constraints of the form x ≥ l and x ≤ u. If a variable in unconstrained, then we can use the
value inf to indicate an infinite bound. We can solve the Battle of the Networks problem
for Players 1 and 2 using Matlab and confirm our saddle point solution from Example 6.2.
Recall the game matrix for Battle of the Networks is:
−15 −35 10
G = −5 8 0
−12 −36 20
106 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
We’ll use G so that we can reserve A for the inequality matrix for Matlab. Using Equations
8.22 and 8.23, we’ll have the linear programming problem for Player 1:
max v = 0x1 + 0x2 + 0x3 + v
s.t. − 15x1 − 5x2 − 12x3 − v ≥ 0
− 35x1 + 8x2 − 36x3 − v ≥ 0
10x1 + 0x2 + 20x3 − v ≥ 0
x1 + x2 + x3 + 0v = 1
x1 , x2 , x3 ≥ 0
This problem is not in a format Matlab likes, we must convert the greater-than (≥) con-
straints to less-than (≤) constraints. We must also convert this to a minimization problem.
We can do this by multiplying the objective by −1 and each ≥ constraint by −1 to obtain:
min − v = 0x1 + 0x2 + 0x3 − v
s.t. 15x1 + 5x2 + 12x3 + v ≤ 0
35x1 − 8x2 + 36x3 + v ≤ 0
− 10x1 + 0x2 − 20x3 + v ≤ 0
x1 + x2 + x3 + 0v = 1
x1 , x2 , x3 ≥ 0
We can read the matrices and vectors for Player 1 as:
15 5 12 1 0
A = 35
−8 36 1 b = 0
−10 0 −20 1 0
H= 1 1 1 0 r= 1
0 0 +∞
0 0
u = +∞
c= 0
l= 0 +∞
−1 −∞ +∞
Note our lower bound for v is −∞ and our upper bound for all variables is +∞. Though
we should note that since x1 + x2 + x3 = 1, these values will automatically be less than 1.
The Matlab solution is shown in Figure 8.3 (Player 1). We can also construct the Matlab
problem for Player 2. Player 2’s problem will be
min 0y1 + 0y2 + 0y3 + v
s.t. − 15y1 − 35y2 + 10y3 − v ≤ 0
− 5y1 + 8y2 − 0y3 − v ≤ 0
− 12y1 − 36y2 + 20y3 − v ≤ 0
y1 + y2 + y3 + 0v = 1
y1 , y2 , y3 ≥ 0
We can construct the matrices and vectors for this problem just as we did before and use
Matlab to find the optimal solution. This is shown in Figure 8.3 (Player 2). Notice that it’s
a lot easier to solve for Player 2’s strategy because it’s already in a Matlab approved form.
5. SOLVING LINEAR PROGRAMS BY COMPUTER 107
(a) Player 1
(b) Player 2
Figure 8.3. We solve for the strategy for Player 1 in the Battle of the Networks.
Player 1 maximizes v subject to the constraints given in Problem 8.19. The result is
Player 1 should play strategy 2 all the time. We also solve for the strategy for Player
2 in the Battle of the Networks. Player 2 minimizes v subject to the constraints
given in Problem 8.21. The result is Player 2 should play strategy 1 all of the time.
This agrees with our saddle-point solution.
108 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
That is, Player 1 should always play pure strategy 2, while Player 2 should play pure strategy
1. This agrees exactly with our observation of the minimax value in Figure 6.1 from Ex-
ample 6.2 in which we concluded that the minimax and maximin values of the game matrix
corresponded precisely to when Player 1 played pure strategy 2 and Player 2 played pure
strategy 1 (element (2, 1) in the matrix G).
5.2. Closing Remarks. In a perfect world, there would be time to teach you everything
you want to know about the Simplex Algorithm (or any other method) for solving linear
programs. If you’re interested in these types of problems, you should consider taking Math
484 (Linear Programming) or getting a good book on the subject.
Proof. We’ll begin by showing the statements that make up Primal Feasibility must
hold. Clearly v is unrestricted and xi ≥ 0 for i = 1, . . . , m. The fact that x1 + · · · + xm = 1
is also clear from the problem. We can rewrite each constraint of the form:
(8.32) A1j x1 + · · · + Amj xm − v ≥ 0
where j = 1, . . . , n as:
(8.33) A1j x1 + · · · + Amj xm − v + sj = 0
where sj ≥ 0. Each variable sj is a surplus variable. Thus it’s clear that if x1 , . . . , xm is a
feasible solution, then at least variables s1 , . . . , sn ≥ 0 exist and Primal Feasibility holds.
Let us re-write the constraints of the form in Expression 8.32 as:
(8.34) −A1j x1 − · · · − Amj xm + v ≤ 0 j = 1, . . . , n
and each non-negativity constraint as:
(8.35) −xi ≤ 0 i = 1, . . . , m
We know that each affine function is both concave and convex and therefore, by Theorem 7.31
(the Karush-Kuhn-Tucker theorem), there are Lagrange multipliers y1 , . . . , yn corresponding
to the constraints of the form in Expression 8.34 and Lagrange multipliers ρ1 , . . . , ρm cor-
responding to the constraints of the form in Expression 8.35. Lastly, there is a Lagrange
multiplier ν corresponding to the constraint:
(8.36) x1 + x2 + · · · + xm − 1 = 0
We know from Theorem 7.31 that:
yj ≥ 0 j = 1, . . . , n
ρi ≥ 0 i = 1, . . . , m
ν unrestricted
Before showing that
Xn
(8.37) Aij yj − ν + ρi = 0 i = 1, . . . , m
j=1
Xn
(8.38) yj = 1
j=1
110 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
holds, we show that Complementary Slackness holds. To see this, note that by Theorem
7.31, we know that:
yj (−A1j x1 − · · · − Amj xm + v) = 0 j = 1, . . . , n
ρi (−xi ) = 0 i = 1, . . . , m
1
1
. (m+1)×1
(8.47) .. ∈ R
∇h(x1 , . . . , xm , v) =
1
0
and
−A1j
−A2j
. (m+1)×1
(8.49) .. ∈ R
∇gj (x1 , . . . , xm , v) =
−A
mj
1
Before proceeding, note that in computing ∇fi (x1 , . . . , xm , v), (i = 1, . . . , m), we will have
−e1 , . . . , em ∈ R(m+1)×1 . Thus, we will never see the vector:
0
0
. (m+1)×1
−em+1 = .. ∈ R
0
−1
because there is no function fm+1 (x1 , . . . , xm , v). We can now rewrite Expression 8.45 as:
0 −A1j 1
0 X −A2j ! 1
. n m
. X .
.. −
(8.50) .. −
yj .. = 0
ρi (−ei ) − ν
0 j=1 −A i=1 1
mj
1 1 0
Consider element i the first m terms of these vectors. Adding term-by-term we have:
n
X
(8.51) 0+ Aij yj + ρi − ν = 0
j=1
This is the ith row of vector that results from adding the terms on the left-hand-side of
Expression 8.50. Now consider row m + 1. We have:
n
X
(8.52) 1− yj + 0 + 0 = 0
j=1
Thus, we have shown that Dual Feasibility holds. Necessity and sufficiency of the statement
follows at once from Theorem 7.31. This completes the proof.
112 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
Theorem 8.12. Let G = (P, Σ, A) be a zero-sum two player game with A ∈ Rm×n . Then
the linear program for Player 2:
min ν
s.t. A11 y1 + · · · + A1n yn − ν ≤ 0
A21 y1 + · · · + A2n yn − ν ≤ 0
..
.
Am1 y1 + · · · + Amn yn − ν ≤ 0
y1 + · · · + yn − 1 = 0
yi ≥ 0 i = 1, . . . , m
has optimal solution (y1 , . . . , yn ) if and only if there exists Lagrange multipliers: x1 , . . . , xm ,
s1 , . . . , sn and v and slack variables ρ1 , . . . , ρm such that:
n
X
Aij yj − ν + ρi = 0 i = 1, . . . , m
j=1
X n
yj = 1
Primal Feasibility :
j=1
yj ≥ 0 j = 1, . . . , n
ρi ≥ 0 i = 1, . . . , m
ν unrestricted
m
X
Aij xi − v − sj = 0 j = 1, . . . , n
i=1
m
X
xi = 1
Dual Feasibility :
i=1
xi ≥ 0 for i = 1, . . . , m
sj ≥ 0 for j = 1, . . . , n
v unrestricted
yj sj = 0 j = 1, . . . , n
Complementary Slackness :
ρi xi = 0 i = 1, . . . , m
Proposition 8.15. The linear programming problem for Player 1 is the dual problem of
the linear programming problem for Player 2 in a zero-sum two player game G = (P, Σ, A)
with A ∈ Rm×n .
There is a very deep theorem about dual linear programming problems, which is beyond
the scope of this course. (We prove it in Math 484.) We state it and make use of it to prove
the minimax theorem in a totally new way.
Theorem 8.16 (Strong Duality Theorem). Let P and D be dual linear programming
problems (like the linear programming problems of Players 1 and 2 in a zero-sum game).
Then either:
(1) Both P and D have a solution and at optimality, the objective function value for
Problem P is identical to the objective function value for Problem D.
(2) Problem P has no solution because it is unbounded and Problem D has no solution
because it is infeasible.
(3) Problem D has no solution because it is unbounded and Problem P has no solution
because it is infeasible.
(4) Both Problem P and Problem D are infeasible.
Theorem 8.17 (Minimax Theorem (redux)). Let G = (P, Σ, A) be a zero-sum two player
game with A ∈ Rm×n , then there exists a Nash equilibrium (x∗ , y∗ ) ∈ ∆. Furthermore, for
every Nash equilibrium pair (x∗ , y∗ ) ∈ ∆ there is one value v ∗ = x∗ T Ay∗ .
Sketch of Proof. Let Problem P1 and Problem P2 be the linear programming prob-
lems for Player 1 and 2 respectively that arise from G. That is:
max v
s.t. A11 x1 + · · · + Am1 xm − v ≥ 0
A12 x1 + · · · + Am2 xm − v ≥ 0
P1 ..
.
A1n x1 + · · · + Amn xm − v ≥ 0
x1 + · · · + xm − 1 = 0
xi ≥ 0 i = 1, . . . , m
min ν
s.t. A11 y1 + · · · + A1n yn − ν ≤ 0
A21 y1 + · · · + A2n yn − ν ≤ 0
P2 ..
.
Am1 y1 + · · · + Amn yn − ν ≤ 0
y1 + · · · + yn − 1 = 0
yi ≥ 0 i = 1, . . . , m
These linear programming problems are dual and therefore if Problem P1 has a solution,
then so does problem P2 . More importantly, at these optimal solutions (x∗ , v ∗ ), (y∗ , ν ∗ ) we
know that v ∗ = ν ∗ as the objective function values must be equal by Theorem 8.16.
Consider Problem P1 : we know that (x1 , . . . , xm ) ∈ ∆m and therefore, this space is
bounded. The value v clearly cannot exceed maxij Aij as a result of the constraints and the
114 8. ZERO-SUM MATRIX GAMES WITH LINEAR PROGRAMMING
fact that xi ∈ [0, 1] for i = 1, . . . , m. Obviously, v can be made as small as we like, but this
won’t happen since this is a maximization problem. The fact that v is bounded from above
and (x1 , . . . , xm ) ∈ ∆m and P1 is a maximization problem (on v) implies that there is at least
one solution (x∗ , v ∗ ) to Problem P1 . In this case, there is a solution (y ∗ , ν ∗ ) to Problem P2
and v ∗ = ν ∗ . Since the constraints for Problem P1 and Problem P2 were taken from Theorem
6.39, we know that (x∗ , y∗ ) is a Nash equilibrium and therefore such an equilibrium must
exist.
Furthermore, while we have not proved this explicitly, one can prove that if (x∗ , y∗ ) is a
Nash equilibrium, then it must be a part of solutions (x∗ , v ∗ ), (y∗ , ν ∗ ) to Problems P1 and
P2 . Thus, any two equilibrium solutions are simply alternative optimal solutions to P1 and
P2 respectively. Thus, for any Nash equilibrium pair we have:
(8.55) ν ∗ = v ∗ = x∗ T Ay∗
This completes the proof sketch.
Remark 8.18 (A remark on Complementary Slackness). Consider the KKT conditions
for Players 1 and 2 (Theorems 8.11 and 8.12). Suppose (for the sake of argument) that in
an optimal solution of the problem for Player 1, sj > 0. Then, it follows that yj = 0 by
complementary slackness. We can understand this from a game theoretic perspective. The
expression:
A1j x1 + · · · + Amj xm
is the expected payoff to Player 1 if Player 2 plays column j. If sj > 0, then:
A1j x1 + · · · + Amj xm > v
But that means that if Player 2 ever played column j, then Player 1 could do better than
the equilibrium value of the game, thus Player 2 has no incentive to ever play this strategy
and the result is that yj = 0 (as required by complementary slackness).
Exercise 69. Use the logic from the preceding remark to argue that xi = 0 when ρi > 0
for Player 2.
Remark 8.19. The connection between zero-sum games and linear programming is sub-
stantially deeper than the previous theorem suggests. Luce and Raiffa [LR89] show the
equivalence between Linear Programming and Zero-Sum games by demonstrating (as we
have done) that for each zero-sum game there is a linear programming problem whose so-
lution yields an equilibrium and for each linear programming problem there is a zero-sum
game whose equilibrium solution yields an optimal solution.
In the next chapter, we’ll continue our discussion of the equivalence of games and opti-
mization problems by investigating general sum two-player games.
CHAPTER 9
x≥0
y≥0
Obviously, we can put this problem in precisely the format given in Expression 9.1, if so
desired.
Remark 9.3. Quadratic programs are just a special instance of nonlinear (or mathe-
matical) programming problems. There are many applications for quadratic programs that
are beyond the scope of these notes. There are also many solution techniques for quadratic
programs, which are also beyond the scope of these notes. Interested readers should consult
[BSS06] for details.
115
116 9. QUADRATIC PROGRAMS AND GENERAL SUM GAMES
The user will supply the matrices and vectors Q, c, A, b, H, r, l and u. The function for
solving quadratic programs in Matlab is quadprog.
If we were to solve the problem from Example 9.2 we would have to multiply the objective
function by −1 to transform the problem from a maximization problem to a minimization
problem:
0 −1/2 x x
min x y − 0 0
−1/2 0 y y
x
s.t. 2 2 = 100
y
x 0
≥
y 0
Figure 9.1. Solving quadratic programs is relatively easy with Matlab. We sim-
ply provide the necessary matrix inputs remembering that we have the objective
(1/2)xT Qx + cT x.
there is at least one Nash equilibrium (x∗ , y∗ ). If either Player were to play his/her Nash
equilibrium, then the optimization problems for the players would be:
xT Ay∗
max
P1 s.t. 1Tm x = 1
x≥0
x∗ T By
max
P2 s.t. 1Tn y = 1
y≥0
Individually, these are linear programs. The problem is, we don’t know the values of (x∗ , y∗ )
a priori. However, we can draw insight from these problems.
Lemma 9.4. Let G = (P, Σ, A, B) be a general sum two-player matrix game with A, B ∈
R . A point (x∗ , y ∗ ) ∈ ∆ is a Nash equilibrium if and only if there exists scalar values α
m×n
118 9. QUADRATIC PROGRAMS AND GENERAL SUM GAMES
Proof. Assume that x∗ = [x∗1 , . . . , x∗m ]T and y∗ = [y1∗ , . . . , yn∗ ]T . Consider the KKT
conditions for the linear programming problem for P1 . The objective function is:
z(x1 , . . . , xn ) = xT Ay∗ = cT x
here c ∈ Rn×1 and
ci = Ai· y∗ = ai1 y1∗ + ai2 y2∗ + · · · + ain yn∗
The vector x∗ is an optimal solution for this problem if and only if there exists multipliers
λ1 , . . . , λm (corresponding to constraints x ≥ 0) and α (corresponding to the constraint
1Tm x = 1 so that:
x∗1 + · · · + x∗m = 1
Primal Feasibility :
x∗i ≥ 0 i = 1, . . . , m
X m
∗
∇z(x ) − λi (−ei ) − α1m = 0
i=1
Dual Feasibility :
λi ≥ 0 for i = 1, . . . , m
α unrestricted
∗
Complementary Slackness : λi xi = 0 i = 1, . . . , m
We observe first that ∇z(x∗ ) = Ay∗ . Therefore, we can write the first equation in the Dual
Feasibility condition as:
m
X
∗
(9.4) Ay − α1m = − λi ei
i=1
Since λi ≥ 0 and ei is just the ith standard basis vector, we know that λi ei ≥ 0 and thus:
(9.5) Ay∗ − α1m ≤ 0
Now, again consider the first equation in Dual Feasibility written as:
m
X
∗
Ay + λi ei − α1m = 0
i=1
3. GENERAL SUM GAMES AND QUADRATIC PROGRAMMING 119
Theorem 9.5. Let G = (P, Σ, A, B) be a general sum two-player matrix game with
A, B ∈ Rm×n . A point (x∗ , y ∗ ) ∈ ∆ is a Nash equilibrium if and only if there are reals α∗
and β ∗ so that (x∗ , y∗ , α∗ , β ∗ ), is a global maximizer for the quadratic programming problem:
max xT (A + B)y − α − β
s.t. Ay − α1m ≤ 0
xT B − β1Tn ≤ 0
(9.11) 1Tm x − 1 = 0
1Tn y − 1 = 0
x≥0
y≥0
Proof. First observe that:
(9.12) Ay − α1m ≤ 0 =⇒ xT Ay − αxT 1m ≤ xT 0 =⇒ xT Ay − α ≤ 0
Similarly,
(9.13) xT B − β1Tn ≤ 0 =⇒ xT By − β1Tn y ≤ 0y =⇒ xT By − β ≤ 0
Combining these inequalities we see that z(x, y, α, β) = xT (A + B)y − α − β ≤ 0. Thus any
set of variables (x∗ , y∗ , α∗ , β ∗ ) so that z(x∗ , y∗ , α∗ , β ∗ ) = 0 is a global maximum.
(⇐) We now show that at a global optimal solution, the KKT conditions for the qua-
dratic program are identical to the conditions given in Lemma 9.4. At an optimal point
(x∗ , y∗ , α∗ , β ∗ ), there are multipliers
(1) λ1 , . . . , λm (corresponding to the constraints Ay − α1m ≤ 0)
(2) µ1 , . . . , µn (corresponding to the constraints xT B − β1Tn ≤ 0),
(3) ν1 (corresponding to the constraint 1Tm x − 1),
(4) ν2 (corresponding to the constraint 1Tn y − 1 = 0),
(5) φ1 , . . . , φm (corresponding to the constraints x ≥ 0) and
(6) θ1 , . . . , θn (corresponding to the constraints y ≥ 0).
We can compute the gradients of the various constraints and objective as (remembering that
we will write x ≥ 0 as −x ≤ 0 and y ≥ 0 as −y ≤ 0. Additionally we note that each
gradient has m + n + 2 components (one for each variable in x, y and α and β. The vector
0 will vary in size to ensure that all vectors have the correct size:
(1)
(A + B)y
(A + B)T x
∇z(x, y, α, β) =
−1
−1
(2)
0
ATi·
∇ (Ay − α1m ) =
−1
0
3. GENERAL SUM GAMES AND QUADRATIC PROGRAMMING 121
(3)
B·j
0
∇ BxT − β1n =
0
−1
(4)
1m
0
∇(1Tm x − 1) =
0
0
(5)
0
1n
∇(1Tn y − 1) =
0
0
(6)
−ei
0
∇(−xi ) =
0
0
(7)
0
−ej
∇(−yj ) =
0
0
In the final gradients, ei ∈ Rm×1 and ej ∈ Rn×1 so that the standard basis vectors agree
with the dimensionality of x and y respectively. The Dual Feasibility constraints of the KKT
conditions for the quadratic program assert that
(1) λ1 , . . . , λn ≥ 0
(2) µ1 , . . . , µ m ≥ 0
(3) φ1 , . . . , φm ≥ 0,
(4) θ1 , . . . , θn ≥ 0,
(5) ν1 ∈ R, and
(6) ν2 ∈ R
122 9. QUADRATIC PROGRAMS AND GENERAL SUM GAMES
We can similarly analyze the component corresponding to α and see that dual feasibility
implies that:
m
X m
X
(9.16) −1 − λi = 0 =⇒ λi = 1
i=1 i=1
Thus dual feasibility shows that (λ1 , . . . , λm ) ∈ ∆m and (µ1 , . . . µn ) ∈ ∆n . Let us now
analyze the component corresponding to variable yj . Dual feasibility implies:
m
X m
X
T T
(9.17) x (A·j + B·j ) − λi Aij − ν2 + θj = 0 =⇒ x (A·j + B·j ) − λi Aij − ν2 ≤ 0
i=1 i=1
We can similarly analyze the component corresponding to variable xi . Dual feasibility implies
that:
n
X n
X
(9.18) (Ai· + Bi· )y − µj Bij − ν1 + φi = 0 =⇒ (Ai· + Bi· )y − µj Bij − ν1 ≤ 0
j=1 j=1
There is now a trick required to complete to proof. Suppose we choose Lagrange multi-
pliers so that xi = λi (i = 1, . . . , m) and yj = µj (j = 1, . . . , n). We are allowed to do so
because of the constraints on the λi and µj . Furthermore, suppose we choose ν1 = α and
ν2 = β. Then if x∗ , y∗ , α∗ , β ∗ is an optimal solution, then Equations 9.17 and 9.18 become:
x∗ T (A + B) − x∗ T A − β ∗ 1Tn ≤ 0 =⇒ x∗ T B − β ∗ 1Tn ≤ 0
(A + B)y∗ − By∗ − α∗ 1m ≤ 0 =⇒ Ay∗ − α∗ 1m ≤ 0
We also know that:
(1) 1Tm x∗ = 1,
(2) 1Tn y∗ = 1,
(3) x ≥ 0, and
(4) y≥0
3. GENERAL SUM GAMES AND QUADRATIC PROGRAMMING 123
Lastly, complementary slackness for the quadratic programming problem implies that:
(9.19) λi (Ai· y − α) = 0 i = 1, . . . , m
xT B·j − β µj = 0 j = 1, . . . , n
(9.20)
m
X m
X m
X
(9.21) x∗i ∗ ∗
(Ai· y − α ) = 0 =⇒ x∗i Ai· y∗ − α∗ x∗i = 0 =⇒ x∗ T Ay∗ − α∗ = 0
i=1 i=1 i=1
Xn Xn n
X
∗T ∗T
∗
B·j yj∗ β ∗ yj∗ = 0 =⇒ x∗ T By∗ − β ∗ = 0
(9.22) x B·j − β µj = 0 =⇒ x −
j=1 j=1 j=1
From this we conclude that any tuple (x∗ , y∗ , α∗ , β ∗ ) satisfying these KKT conditions must
be a global maximizer because adding these final two equations yields:
(9.23) x∗ T (A + B)y∗ − α∗ − β ∗ = 0
x∗ T Ay∗ − α∗ = 0
x∗ T By∗ − β ∗ = 0
x∗ T (A + B)y∗ − α∗ − β ∗ = 0
holds and thus (x∗ , y∗ , α∗ , β ∗ ) must be a global maximizer for the quadratic program because
the objective function achieves its upper bound. This completes the proof.
Example 9.6. We can find a third Nash equilibrium for the Chicken game using this
approach. Recall we have:
0 −1
A=
1 −10
0 1
B=
−1 −10
124 9. QUADRATIC PROGRAMS AND GENERAL SUM GAMES
0 0 y1
max x1 x2 −α−β
0 −20 y2
0 −1 y1
α 0
− ≤
s.t.
1 −10 y2
α 0
0 1
x 1 x 2 − β β ≤ 0 0
−1 −10
x1
(9.24) 1 1 =1
x2
y1
1 1 =1
y2
x1 0
≥
x2 0
y1 0
≥
y2 0
max − 20x2 y2 − α − β
− y2 − α ≤ 0
s.t.
y1 − 10y2 − α ≤ 0
− x2 − β ≤ 0
(9.25)
x1 − 10x2 − β ≤ 0
x1 + x2 = 1
y1 + y2 = 1
x1 , x2 , y1 , y2 ≥ 0
An optimal solution to this problem is x1 = 0.9, x2 = 0.1, y1 = 0.9, y2 = 0.1. This is a third
Nash equilibrium in mixed strategies for this instance of Chicken. Identifying this third Nash
equilibrium in Matlab is shown in Figure 9.2. In order to correctly input this problem into
Matlab, we need to first write the problem as a proper quadratic program. This is done by
letting the vector of decision variables be:
x1
x2
y
z = 1
y2
α
β
3. GENERAL SUM GAMES AND QUADRATIC PROGRAMMING 125
Figure 9.2. We can use the power of Matlab to find a third Nash equilibrium in
mixed strategies for the game of Chicken by solving the Problem 9.26. Note, we
have to change this problem to a minimization problem by multiplying the objective
by −1.
CHAPTER 10
Heretofore we have considered games in which the players were unable to communicate
before play began or in which players has no way of trusting each other with certainty
(remember Prisoner’s dilemma). In this chapter, we remove this restriction and consider
those games in which players may put in place a pre-play agreement on their play in an
attempt to identify a solution with which both players can live happily.
1. Payoff Regions in Two Player Games
Definition 10.1 (Cooperative Mixed Strategy). Let G = (P, Σ, A, B) be a two-player
matrix game with A, B ∈ Rn×m . Then a cooperative strategy is a collection of probabilities
xij (i = 1, . . . , m, j = 1, . . . , n) so that:
Xm X n
xij = 1
i=1 j=1
xij ≥ 0 i = 1, . . . , m, j = 1, . . . , n
To any cooperative strategy, we can associate a vector x ∈ ∆mn .
Remark 10.2. For any cooperative strategy xij (i = 1, . . . , m, j = 1, . . . , n), xij gives
the probability that Player 1 plays row i while Player 2 players column j. Note, x could be
thought of as a matrix, but for the sake of notational consistency, it is easier to think of it
as a vector with an strange indexing scheme.
Definition 10.3 (Cooperative Expected Payoff). Let G = (P, Σ, A, B) be a two-player
matrix game with A, B ∈ Rn×m and let xij (i = 1, . . . , m, j = 1, . . . , n) be a cooperative
strategy for Player 1 and 2. Then:
Xm X n
(10.1) u1 (x) = Aij xij
i=1 j=1
(c) Overlap
Figure 10.1. The three plots shown the competitive payoff region, cooperative
payoff region and and overlay of the regions for the Battle of the Sexes game. Note
that the cooperative payoff region completely contains the competitive payoff region.
This set is defined by equalities associated with linear functions (which are both convex and
concave). We can rewrite this as:
m X n
X
Aij xij − u1 ≤ 0
i=1 j=1
X m X n
− Aij xij + u1 ≤ 0
i=1 j=1
m X n
X
Bij xij − u2 ≤ 0
i=1 j=1
X m X n
− Bij xij + u2 ≤ 0
i=1 j=1
Xm X n
xij ≤ 1
i=1 j=1
m X n
X
− xij = −1
i=1 j=1
−x ≤ 0 i = 1, . . . , m, j = 1, . . . , n
ij
130 10. NASH’S BARGAINING PROBLEM AND COOPERATIVE GAMES
Thus, since linear functions are convex, the set of tuples (u1 , u2 , x) that satisfy these con-
straints is a convex set by Theorems 7.24 and 7.27. Suppose that (u11 , u12 , x1 ) and (u21 , u22 , x2 )
are two tuples satisfying these constraints. Then clearly, (u11 , u12 ), (u21 , u22 ) ∈ P (A, B). Since
the set of tuples (u1 , u2 , x) that satisfy these constraints form a convex set we know that for
all λ ∈ [0, 1] we have:
(10.8) λ(u11 , u12 , x1 ) + (1 − λ)(u21 , u22 , x2 ) = (u1 , u2 , x)
and (u1 , u2 , x) satisfies the constraints. But then, (u1 , u2 ) ∈ P (A, B) and therefore
(10.9) λ(u11 , u12 ) + (1 − λ)(u21 , u22 ) ∈ P (A, B)
for all λ. It follows that P (A, B) is convex.
Remark 10.10. The next theorem assumes that the reader knows the definition of a
closed set in Euclidean space. There are many consistent definitions for a closed set in Rn ,
however we will take the definition to be that the set is defined by a collection of equalities
and non-strict (i.e., ≤) inequalities.
Theorem 10.11. Let G = (P, Σ, A, B) be a two-player matrix game with A, B ∈ Rn×m .
The cooperative payoff region P (A, B) is a bounded and closed set.
Proof. Again, consider the defining equalities:
m n
XX
Aij xij − u1 = 0
i=1 j=1
X m X n
Bij xij − u2 = 0
i=1 j=1
m X
n
X
xij = 1
i=1 j=1
xij ≥ 0 i = 1, . . . , m, j = 1, . . . , n
This set must be bounded because xij ∈ [0, 1] for i = 1, . . . , m and j = 1, . . . , n. As a result
of this, the value of u1 is bounded above and below by the largest and smallest values in A
while the value of u2 is bounded above and below by the largest and smallest values in B.
Closure of the set is ensured by the fact that the set is defined by non-strict inequalities and
equalities.
Remark 10.12. What we’ve actually proved in these theorems (and more importantly)
is that the set of tuples (u1 , u2 , x) defined by the system of equations and inequalities:
m n
XX
Aij xij − u1 = 0
i=1 j=1
Xm X n
Bij xij − u2 = 0
i=1 j=1
m X
n
X
xij = 1
i=1 j=1
xij ≥ 0 i = 1, . . . , m, j = 1, . . . , n
2. COLLABORATION AND MULTI-CRITERIA OPTIMIZATION 131
is closed, bounded and convex. We will actually use this result, rather than the generic
statements on P (A, B).
Remark 10.13. You note that the objective function has now been replaced with a
vector of objective functions. Multi-criteria optimization problems can be challenging to
solve because (e.g.) making z1 (x1 , . . . , xn ) larger may make z2 (x1 , . . . , xn ) smaller and vice
versa.
Example 10.14 (The Green Toy Maker). For the sake of argument, consider the Toy
Maker problem from Example 8.1. We had the linear programming problem:
max z(x1 , x2 ) = 7x1 + 6x2
s.t. 3x1 + x2 ≤ 120
x1 + 2x2 ≤ 160
x1 ≤ 35
x1 ≥ 0
x2 ≥ 0
Suppose a certain amount of pollution is created each time a toy is manufactured. Sup-
pose each plane generates 3 units of pollution, while manufacturing a boat generates only 2
132 10. NASH’S BARGAINING PROBLEM AND COOPERATIVE GAMES
units of pollution. Since x1 was the number of planes produced and x2 was the number of
boats produced, we could create a multi-criteria optimization problem in which we simul-
taneously attempt to maximize profit 7x1 + 6x2 and minimize pollution 3x1 + 2x2 . Since
every minimization problem can be transformed into a maximization problem by negating
the objective we would have the problem:
max 7x 1 + 6x 2 , −3x 1 − 2x 2
s.t. 3x1 + x2 ≤ 120
x1 + 2x2 ≤ 160
x1 ≤ 35
x1 ≥ 0
x2 ≥ 0
Remark 10.15. For n > 1, we can choose many different ways to order elements in Rn .
For example, in the plane there are many ways to decide that a point (x1 , y1 ) is greater than
or less than or equivalent to another point (x2 , y2 ). We can think of these as the various
ways of assigning a preference relation to points in the plane (or more generally points in
Rn ). Among other things, we could:
(1) Order them based on their standard euclidean distance to the origin (as points);
i.e.,
q q
(x1 , y1 ) (x2 , y2 ) ⇐⇒ x1 + y1 > x22 + y22
2 2
(2) We could alphabetize them by comparing the first component and then the second
component. (This is called the lexicographic ordering.)
(3) We could specify a parameter λ ∈ R and declare:
(x1 , y1 ) (x2 , y2 ) ⇐⇒ x1 + λy1 > x2 + λy1 2
For this reason, a multi-criteria optimization problem may have many equally good solutions.
There is a substantial amount of information on solving these types of problems, which arise
frequently in the real world. The interested reader might consider [Coh03].
Definition 10.16 (Pareto Optimality). Let gi : Rn → R (i = 1, . . . , m) and hj : Rn → R
and zk : Rn → R (k = 1, . . . , s). Consider the mult-criteria optimization problem:
max z1 (x 1 , . . . , x n ) z2 (x 1 , . . . , x n ) · · · zs (x 1 , . . . , x n )
s.t. g1 (x1 , . . . , xn ) ≤ 0
..
.
gm (x1 , . . . , xn ) ≤ 0
h1 (x1 , . . . , xn ) = 0
..
.
h (x , . . . , x ) = 0
l 1 n
A payoff vector z(x ) dominates another payoff vector z(x) (for two feasible points x, x∗ ) if:
∗
are the cooperative expected payoff functions and u01 and u02 are status quo payoff values–
usually assumed to be a Nash equilibrium payoff value for the two players.
Remark 10.24. Assumption 6 simply states that if (u1 , u2 ) ∈ P (A, B) (for u1 , u2 ∈ R),
then (u2 , u1 ) ∈ P (A, B) also. Thus, P (A, B) is symmetric in R2 about the line y = x.
Inspection of Figure 10.1 reveals this is (in fact) true.
Remark 10.25. Our goal is to now show that there is an arbitration procedure x∗ ∈ ∆nm
that satisfies these assumptions and that the resulting pair (u1 (x∗ , u2 (x∗ )) ∈ P (A, B) is
unique. This is Nash’s Bargaining Theorem.
has at least one global optimal solution (u∗1 , u∗2 , x∗ ). Furthermore if (u01 , u02 , x0 ) is an alterna-
tive optimal solution, then u∗1 = u01 and u∗2 = u02 .
Proof. By the same argument as in the proof of Theorem 10.11 the feasible region of
this problem is a closed bounded and convex set. Moreover, since (u01 , u02 ) ∈ P (A, B) we
know that there is some x0 satisfying the constraints given in Expression 10.7 and that the
tuple (u01 , u02 , x0 ) is feasible to this problem. Thus, the feasible region is non-empty. Thus
applying Lemma 10.26 we know that there is at least one (global optimal) solution to this
problem.
To see the uniqueness of (u∗1 , u∗2 ), suppose that M = (u∗1 − u01 )(u∗2 − u02 ) and we have
a second solution (u01 , u02 , x0 ) so that (without loss of generality) u01 > u∗1 and u02 < u∗2 but
M = (u01 − u01 )(u02 − u02 ). We showed that P (A, B) is convex (see Theorem 10.9). Then there
is some feasible (u001 , u002 , x00 ) so that:
1 1
(10.15) u00i = u∗i + u0i
2 2
for i = 1, 2. Evaluating the objective function at this point yields:
00 0 00 0 1 ∗ 1 0 0 1 ∗ 1 0 0
(10.16) (u1 − u1 )(u2 − u2 ) = u + u − u1 u + u − u2
2 1 2 1 2 2 2 2
Expanding yields:
1 ∗ 1 0 1 0 1 0 1 ∗ 1 0 1 0 1 0
(10.17) u − u + u − u u − u + u − u =
2 1 2 1 2 1 2 1 2 2 2 2 2 2 2 2
1
(u∗1 − u01 ) + (u01 − u01 ) (u∗2 − u02 ) + (u02 − u02 )
4
∗ ∗ 0 0 0 0
Let Hi = (ui − ui ) and Hi = (ui − ui ) for i = 1, 2. Then our expression reduces to:
1 1
(10.18) (H1∗ + H10 )(H2∗ + H20 ) = (H1∗ H2∗ + H10 H2∗ + H1∗ H20 + H10 H20 )
4 4
We have the following:
(1) H1∗ H2∗ = M (by definition).
136 10. NASH’S BARGAINING PROBLEM AND COOPERATIVE GAMES
(10.19) H10 H2∗ + H1∗ H20 = u01 u∗2 − u01 u02 − u∗2 u01 + u01 u02 + u∗1 u02 − u∗1 u02 − u02 u01 + u01 u02
We can write:
(10.20) H1∗ H2∗ + H10 H2∗ + H1∗ H20 + H10 H20 = 2M + H1∗ H20 + H10 H20 =
4M + H1∗ H20 + H10 H20 − 2M = 4M + H1∗ H20 + H10 H20 − H1∗ H2∗ − H10 H20
(10.21) H1∗ H20 + H10 H20 − H1∗ H2∗ − H10 H20 = u01 u∗2 − u01 u02 − u∗2 u01 + u01 u02 +
u∗1 u02 − u∗1 u02 − u02 u01 + u01 u02 − u∗1 u∗2 − u∗1 u02 − u∗2 u01 + u01 u02 −
(10.22) H1∗ H20 + H10 H20 − H1∗ H2∗ − H10 H20 = u01 u∗2 + u∗1 u02 − u∗1 u∗2 − u01 u02 =
(u∗1 − u01 )(u02 − u∗2 )
Thus:
1
(10.23) (H ∗ H ∗ + H10 H2∗ + H1∗ H20 + H10 H20 = 2M + H1∗ H20 + H10 H20 ) =
4 1 2
1
(4M + H1∗ H20 + H10 H20 − H1∗ H2∗ − H10 H20 ) =
4
1 1
M + (H1∗ H20 + H10 H20 − H1∗ H2∗ − H10 H20 ) = M + (u∗1 − u01 )(u02 − u∗2 ) > M
4 4
because (u∗1 − u01 )(u02 − u∗2 ) > 0 by our assumption that u01 > u∗1 and u02 < u∗2 . But since we
assumed that M was the maximum value the objective function attained, we know that we
must have u∗1 = u01 and u∗2 = u02 . This completes the proof.
It suffices to show that the solution of this quadratic program provides an arbitration proce-
dure x satisfying Nash’s assumptions. Uniqueness follows immediately from Lemma 10.27.
Denote the feasible region of this problem by F (A, B). That is F (A, B) is the set of all tuples
(u1 , u2 , x) satisfying the constraints of Problem 10.24. Clearly u1 = u1 (x) and u2 = u2 (x).
Before proceeding, recall that Q(A, B), the payoff region for the competitive game G is
contained in P (A, B). Clearly if u01 , u02 is chosen as an equilibrium for the competitive game,
we know that (u01 , u02 ) ∈ P (A, B). Thus there is a x0 so that (u01 , u02 , x0 ) ∈ F (A, B) and it
follows that 0 is a lower bound for the maximal value of the objective function.
Assumption 1: By construction of this problem, we know that u1 (x∗ ) ≥ u01 and
u2 (x∗ ) ≥ u02 .
Assumption 2: By Lemma 10.27 any solution (u∗1 , u∗2 , x∗ ) has unique u∗1 and u∗2 . Thus,
any other feasible solution (u1 , u2 , x) must have the property that either u1 < u∗1 or u2 < u∗2 .
Therefore, the (u∗1 , u∗2 ) must be Pareto optimal.
Assumption 3: Since the constraints of Problem 10.24 properly contain the constraints
in Expression 10.7, the assumption of feasibility is ensured.
Assumption 4: Suppose that P 0 ⊆ P (A, B). Then there is a subset F 0 ⊆ F (A, B)
corresponding to P 0 . If (u∗1 , u∗2 ) ∈ P 0 and (u01 , u02 ) ∈ P 0 , it follows that (u∗1 , u∗2 , x∗ ) ∈ F 0 and
(u01 , u02 , x0 ) ∈ F 0 . Then we can define the new optimization problem:
0 0
max (u1 − u1 )(u2 − u2 )
(10.25) s.t. (u1 , u2 , x) ∈ F
(u1 , u2 , x) ∈ F 0
The constraints of the problem will not be changed since we assume that α1 , α2 ≥ 0. To see
this note that linear transformation of the payoff values implies the new constraints:
m X
X n
(10.28) (α1 Aij + β1 )xij − (α1 u1 + β1 ) = 0
i=1 j=1
m X
X n m X
X n
⇐⇒ α1 Aij xij + β1 xij − (α1 u1 + β1 ) = 0 ⇐⇒
i=1 j=1 i=1 j=1
Xm X
n m X
X n
α1 Aij xij + β1 − α1 u1 − β1 = 0 ⇐⇒ Aij xij − u1 = 0
i=1 j=1 i=1 j=1
m X
X n
(10.29) (α2 Bij + β2 )xij − (α2 u2 + β2 ) = 0
i=1 j=1
m X
X n m X
X n
⇐⇒ α2 Bij xij + β2 xij − (α2 u2 + β2 ) = 0 ⇐⇒
i=1 j=1 i=1 j=1
Xm X
n m X
X n
α2 Bij xij + β2 − α2 u2 − β2 = 0 ⇐⇒ Bij xij − u2 = 0
i=1 j=1 i=1 j=1
Figure 10.2. The Pareto Optimal, Nash Bargaining Solution, to the Battle of the
Sexes is for each player to do what makes them happiest 50% of the time. This seems
like the basis for a fairly happy marriage, and it yields a Pareto optimal solution,
shown by the green dot.
The following Matlab code will solve the Nash bargaining problem associated with the
Battle of the Sexes game. Note that we are solving a maximization problem, but Matlab
solve mnimization problems by default. Thus we change the sign on the objective matrices.
As before, calling quadprog will solve the maximization problem associated with Battle of
the Sexes. We must compute the appropriate matrices and vectors for this problem. In order
to see that this is the correct problem, note we can read the H matrix and r vector directly
from the equality constraints of Problem 10.33. There are no inequality constraints (that
are not bounds) thus A = b = [], the empty matrix. The matrix and vector that make up
the objective functions can be found by noting that if we let our vector of decision variables
140 10. NASH’S BARGAINING PROBLEM AND COOPERATIVE GAMES
[x obj] = quadprog(-Q,-c,A,b,H,r,lb,ub);
Figure 10.3. Matlab input for solving Nash’s bargaining problem with the Battle
of the Sexes problem. Note that we are solving a maximization problem, but Matlab
solve mnimization problems by default. Thus we change the sign on the objective
matrices.
Lemma 11.4. For any Two-Coalition Game GS , there is a Nash equilibrium strategy for
both the coalition S and its dual S c .
Exercise 75. Prove the previous lemma. [Hint: Use Nash’s theorem.]
141
142 11. A SHORT INTRODUCTION TO N -PLAYER COOPERATIVE GAMES
by superadditivity. Now:
X
v(S c ) ≥ v({i})
i∈S c
Thus:
N
X
v(P) > v({i})
i=1
Remark 11.16. The first criterion for a tuple (x1 , . . . , xN ) to be an imputation says that
each player must do better in the grand coalition then they would on their own (against the
world). The second criterion says that the total allotment of payoff to the players cannot
exceed the payoff received by the grand coalition itself. Essentially, this second criterion
asserts that the coalition Pcannot go into debt to maintain its members. It is also worth
noting that the condition i∈P xi = v(P) is equivalent to a statement on Pareto optimality
in so far as players all together can’t expect to do any better than the net payoff accorded
to the grand coalition.
Definition 11.17 (Dominance). Let (P, v) be a coalition game. Suppose x = (x1 , . . . , xN )
and y = (y1 , . . . , yN ) are two imputations. Then x dominates y over some coalition S ⊂ P
if
Pi > yi for all i ∈ S and
(1) x
(2) i∈S xi ≤ v(S)
Remark 11.18. The previous definition states that Players in coalition S prefer the
payoffs they receive under x to the payoffs they receive under y. Furthermore, these same
players can threaten to leave the grand coalition P because they may actually improve their
payoff by playing coalition S.
Definition 11.19 (Stable Set). A stable set X ⊆ Rn of imputations is a set satisfying:
(1) No payoff vector x ∈ X is dominated in any coalition by another coalition y ∈ X
and
(2) All payoff vectors y 6∈ X are dominated by at least one vector x ∈ X.
Remark 11.20. Stable sets are (in some way) very good sets of imputations in so far as
they represent imputations that will make players want remain in the grand coalition.
4. The Core
Definition 11.21 (Core). Given a coalition game (P, v), the core is:
( N
!)
X X
C(v) = x ∈ Rn : xi = v(P) and ∀S ⊆ P xi ≥ v(S)
i=1 i∈S
Theorem 11.24. Let (P, v) be a coalition game. Consider the linear programming prob-
lem:
min x1 + · · · + xN
(11.2)
X
s.t. xi ≥ v(S) ∀S ⊆ P
i∈S
PN
If there is no solution x∗ so that i=1 xi = v(P ), then C(v) = ∅.
Exercise 78. Prove the preceding theorem. [Hint: Note that the constraints enforce
the requirement:
!
X
∀S ⊆ P xi ≥ v(S)
i∈S
PN
while the objective function yields i=1 xi .]
Corollary 11.25. The core of a coalition game (P, v) may be empty.
Exercise 79. Find the core of the previous three player cooperative game.
Theorem 11.26 (Bondarvera-Shapley Theorem). Let (P, v) be a coalition game with
|P| = N . The core C(v) is non-empty if and only if there exists y1 , . . . , y2N where each yi
corresponds to a set Si ⊆ P so that:
2N
X
v(P) = yi v(Si )
i=1
X
yi = 1 ∀j ∈ P
Si ⊇{j}
yi ≥ 0 ∀Si ⊆ P
Proof. The dual linear programming problem (See Chapter 8.6) for Problem 11.2 is:
2N
X
max yi v(Si )
i=1
(11.3) X
s.t. yi = 1 ∀j ∈ P
Si ⊇{j}
yi ≥ 0 ∀Si ⊆ P
To see this, we note that there are 2N constraints in Problem 11.2 and N variables and
thus there will be N constraints in the dual problem, but 2N variables and the resulting
dual problem is Problem 11.3. By Theorem 8.16 (the Strong Duality Theorem), Problem
11.3 has a solution if and only if Problem 11.2 does and moreover the objective functions at
optimality coincide.
Exercise 80. Prove that Problems 11.2 and 11.3 are in fact dual linear programming
problems by showing that they have the same KKT conditions.
Corollary 11.27. A non-empty core is not necessarily a singleton.
Exercise 81. Prove the preceding corollary. [Hint: Think about alternative optimal
solutions.]
146 11. A SHORT INTRODUCTION TO N -PLAYER COOPERATIVE GAMES
Exercise 82. Show that computing the core is an exponential problem even though
solving a linear programming problem is known to be polynomial in the size of the problem.
Remark 11.28. The core can be thought of as the possible “equilibrium” imputations
that smart players will agree to and that cause the grand coalition to hold together; i.e., no
players or coalition have any motivation to leave the coalition. Unfortunately, the fact that
the core may be empty is not helpful.
5. Shapley Values
Definition 11.29 (Shapley Values). Let (P, v) be a coalition game with N players.
Then the Shapley value for Player i is:
X |S|!(N − |S| − 1)!
(11.4) xi = φi (v) = (v (S ∪ {i}) − v(S))
N!
S⊆P\{i}
Remark 11.30. The Shapley value is the average extra value Player i contributes to
each possible coalition that might form. Imagine forming the grand coalition one player at
a time. There are N ! ways to do this. Hence, in an average, N ! is in the denominator of the
Shapley value.
Now, if we’ve formed coalition S (on our way to forming P), then there are |S|! ways
we could have done this. Each of these ways yields v(S) in value because the characteristic
function does not value how a coalition is formed, only the members of the coalition.
Once we add i to the coalition S, the new value is v (S ∪ {i}) and the value player i
added was v (S ∪ {i}) − v(S). We then add the other N − |S| − 1 players to achieve the
grand coalition. There are (N − |S| − 1)! ways of doing this.
Thus, the extra value Player i adds in each case is v (S ∪ {i}) − v(S) multiplied by
|S|!(N − |S| − 1)! for each of the possible ways this exact scenario occurs. Summing over
all possible subsets S and dividing by N !, as noted, yields the average excess value Player i
brings to a coalition.
Remark 11.31. We state, but do not prove, the following theorem. The proof rests on
the linear properties of averages. That is, we note that is a linear expression in v(S) and
v (S ∪ {i}).
Theorem 11.32. For any coalition game (P, v) with N players, then:
φi (v) ≥ v({i})
(1) P
(2) i∈P φi (v) = v(P)
(3) From (1) and (2) we conclude that (φ1 (v), . . . , φN (v)) is an imputation.
(4) If for all S ⊆ P, v (S ∪ {i}) = v (S ∪ {j}) with i, j 6∈ S, then φi (v) = φj (v).
(5) If v and w are two characteristic functions in coalition games (P, v) and (P, w),
then φi (v + w) = φi (v) + φi (w) for all i ∈ P.
(6) If v (S ∪ {i}) = v(S) for all S ⊆ P with i 6∈ S then φi (v) = 0 because Player i
contributes nothing to the grand coalition.
Exercise 83. Prove the previous theorem.
Exercise 84. Find the Shapley values for each player in the previous three player game.
Remark 11.33. There is substantially more information on coalition games and economists
have spent a large quantity of time investigating the various properties of these games. The
5. SHAPLEY VALUES 147
interested reader should consider [LR89] and [Mye01] for more detailed information. Ad-
ditionally, for general game theoretic research the journals, The International Journal of
Game Theory, Games and Economic Behavior and IEEE Trans. Automatic Control have a
substantial number of articles on game theory, including coalition games.
Bibliography
[BCG01a] E. R. Berlekamp, J. H. Conway, and R. K. Guy, Winning ways for your mathematical plays,
vol. 1, A. K. Peters, 2001.
[BCG01b] , Winning ways for your mathematical plays, vol. 2, A. K. Peters, 2001.
[BCG01c] , Winning ways for your mathematical plays, vol. 3, A. K. Peters, 2001.
[BCG01d] , Winning ways for your mathematical plays, vol. 4, A. K. Peters, 2001.
[BJS04] Mokhtar S. Bazaraa, John J. Jarvis, and Hanif D. Sherali, Linear programming and network flows,
Wiley-Interscience, 2004.
[BO82] Tamer Başar and Geert Jan Olsder, Dynamic noncooperative game theory, Academic Press, 1982.
[Bra04] S. J. Brams, Game theory and politics, Dover Press, 2004.
[BSS06] M. S. Bazaraa, H. D. Sherali, and C. M. Shetty, Nonlinear programming: Theory and algorithms,
John Wiley and Sons, 2006.
[Coh03] J. L. Cohen, Multiobjective progamming and planning, Dover, 2003.
[Con76] J. H. Conway, On numbers and games, Academic Press, 1976.
[DJLS00] E. J. Dockner, S. Jørgensen, N. V. Long, and G. Sorger, Differential games in economics and
management science, Cambridge University Press, 2000.
[Dre81] M. Dresher, The mathematics of games of strategy, Dover Press, 1981.
[LH61] C. E. Lemke and J. T. Howson, Equilibrum points of bimatrix games, J. Soc. Indust. Appl. Math.
12 (1961), no. 2, 413–423.
[LR89] R. D. Luce and H. Raiffa, Games and decisions: Introduction and critical survey, Dover Press,
1989.
[Mor94] P. Morris, Introduction to Game Theory, Springer, 1994.
[MS64] O. L. Mangasarian and H. Stone, Two-Person Nonzero-Sum Games and Quadratic Programming,
J. Math. Analysis and Applications 9 (1964), 348–355.
[MT03] J. E. Marsden and A. Tromba, Vector calculus, 5 ed., W. H. Freeman, 2003.
[Mun00] J. Munkres, Topology, Prentice Hall, 2000.
[Mye01] R. B. Myerson, Game theory: Analysis of conflict, Harvard University Press, 2001.
[PR71] T. Parthasarathy and T. E. S. Raghavan, Some topics in two-person games, Elsevier Science
LTD, 1971.
[vNM04] J. von Neumann and O. Morgenstern, The Theory of Games and Economic Behavior, 60th an-
niversary edition ed., Princeton University Press, 2004.
[Wei97] J. W. Weibull, Evolutionary game theory, MIT Press, 1997.
[WV02] W. L. Winston and M. Venkataramanan, Introduction to mathematical programming: Applica-
tions and algorithms, vol. 1, Duxbury Press, 2002.
149