Lecture Notes on

DECISION THEORY
Brian Weatherson

2015
 Contents

1 Introduction 1
1.1 Decisions and Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Previews . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Example: Newcomb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Example: Sleeping Beauty . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Simple Reasoning Strategies 6

2.1 Dominance Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 States and Choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Maximin and Maximax . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Ordinal and Cardinal Utilities . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Regret . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Uncertainty 13
3.1 Likely Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Do What’s Likely to Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Probability and Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Measures 18
4.1 Probability Defined . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3 Normalised Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 Formalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.5 Possibility Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5 Truth Tables 24
5.1 Compound Sentences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2 Equivalence, Entailment, Inconsistency, and Logical Truth . . . . . . . . . 27
5.3 Two Important Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

6 Axioms for Probability 30

6.1 Axioms of Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6.2 Truth Tables and Possibilities . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.3 Propositions and Possibilities . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
7 Conditional Probability 38
7.1 Conditional Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.2 Bayes Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.3 Conditionalisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

8 About Conditional Probability 44

8.1 Conglomerability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.2 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.3 Kinds of Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.4 Gamblers’ Fallacy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

9 Expected Utility 49
9.1 Expected Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
9.2 Maximise Expected Utility Rule . . . . . . . . . . . . . . . . . . . . . . . . 50
9.3 Structural Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

10 Sure Thing Principle 54

10.1 Generalising Dominance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
10.2 Sure Thing Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
10.3 Allais Paradox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
10.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

11 Understanding Probability 61
11.1 Kinds of Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
11.2 Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
11.3 Degrees of Belief . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

12 Objective Probabilities 66
12.1 Credences and Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
12.2 Evidential Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
12.3 Objective Chances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
12.4 The Principal Principle and Direct Inference . . . . . . . . . . . . . . . . . 69

13 Understanding Utility 71
13.1 Utility and Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
13.2 Experiences and Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
13.3 Objective List Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

14 Subjective Utility 76
14.1 Preference Based Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
14.2 Interpersonal Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . 77
14.3 Which Desires Count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
15 Declining Marginal Utilities 80
15.1 Money and Utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
15.2 Insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
15.3 Diversification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
15.4 Selling Insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

16 Newcomb’s Problem 85
16.1 The Puzzle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
16.2 Two Principles of Decision Theory . . . . . . . . . . . . . . . . . . . . . . 86
16.3 Bringing Two Principles Together . . . . . . . . . . . . . . . . . . . . . . . 87
16.4 Well Meaning Friends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

17 Realistic Newcomb Problems 90

17.1 Real Life Newcomb Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
17.2 Tickle Defence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

18 Causal Decision Theory 95

18.1 Causal and Evidential Decision Theory . . . . . . . . . . . . . . . . . . . . 95
18.2 Right and Wrong Tabulations . . . . . . . . . . . . . . . . . . . . . . . . . 95
18.3 Why Ain’Cha Rich . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
18.4 Dilemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
18.5 Weak Newcomb Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

19 Introduction to Games 100

19.1 Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
19.2 Zero-Sum Games and Backwards Induction . . . . . . . . . . . . . . . . . 102
19.3 Zero-Sum Games and Nash Equilibrium . . . . . . . . . . . . . . . . . . . 103

20 Zero-Sum Games 105

20.1 Mixed Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
20.2 Surprising Mixed Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . 106
20.3 Calculating Mixed Strategy Nash Equilibrium . . . . . . . . . . . . . . . . 108

21 Nash Equilibrium 110

21.1 Illustrating Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . 110
21.2 Why Play Equilibrium Moves? . . . . . . . . . . . . . . . . . . . . . . . . . 111
21.3 Causal Decision Theory and Game Theory . . . . . . . . . . . . . . . . . . 113

22 Many Move Games 115

22.1 Games with Multiple Moves . . . . . . . . . . . . . . . . . . . . . . . . . . 115
22.2 Extensive and Normal Form . . . . . . . . . . . . . . . . . . . . . . . . . . 115
22.3 Two Types of Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
22.4 Normative Significance of Subgame Perfect Equilibrium . . . . . . . . . . . 117
22.5 Cooperative Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
22.6 Pareto Efficient Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
22.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
23 Backwards Induction 121
23.1 Puzzles About Backwards Induction . . . . . . . . . . . . . . . . . . . . . 121
23.2 Pettit and Sugden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

24 Group Decisions 125

24.1 Making a Decision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
24.2 Desiderata for Preference Aggregation Mechanisms . . . . . . . . . . . . . 128
24.3 Assessing Plurality Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

25 Arrow’s Theorem 130

25.1 Ranking Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
25.2 Cyclic Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
25.3 Proofs of Arrow’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 133

26 Voting Systems 135

26.1 Plurality voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
26.2 Runoff Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
26.3 Instant Runoff Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

27 More Voting Systems 140

27.1 Borda Count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
27.2 Approval Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
27.3 Range Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
27.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
 Chapter 1

Introduction
1.1 Decisions and Games
This course is an introduction to decision theory. We’re interested in what to do when the
outcomes of your actions depend on some external facts about which you are uncertain.
The simplest such decision has the following structure.

State 1 State 2
Choice 1 a b
Choice 2 c d

The choices are the options you can take. The states are the ways the world can be that affect
how good an outcome you’ll get. And the variables, a, b, c and d are numbers measuring how
good those outcomes are. For now we’ll simply have higher numbers representing better
outcomes, though eventually we’ll want the numbers to reflect how good various outcomes
are.
Let’s illustrate this with a simple example. It’s a Sunday afternoon, and you have the
choice between watching a football game and finishing a paper due on Monday. It will be
a little painful to do the paper after the football, but not impossible. It will be fun to watch
football, at least if your team wins. But if they lose you’ll have spent the afternoon watching
them lose, and still have the paper to write. On the other hand, you’ll feel bad if you skip
the game and they win. So we might have the following decision table.

Your Team Wins Your Team Loses

Watch Football 4 1
Work on Paper 2 3

The numbers of course could be different if you have different preferences. Perhaps your
desire for your team to win is stronger than your desire to avoid regretting missing the
game. In that case the table might look like this.

1
 Your Team Wins Your Team Loses
Watch Football 4 1
Work on Paper 3 2

Either way, what turns out to be for the best depends on what the state of the world is. These
are the kinds of decisions with which we’ll be interested.
Sometimes the relevant state of the world is the action of someone who is, in some
loose sense, interacting with you. For instance, imagine you are playing a game of rock-
paper-scissors. We can represent that game using the following table, with the rows for
your choices and the columns for the other person’s choices.

Rock Paper Scissors

Rock 0 -1 1
Paper 1 0 -1
Scissors -1 1 0

Not all games are competitive like this. Some games involve coordination. For instance,
imagine you and a friend are trying to meet up somewhere in New York City. You want
to go to a movie, and your friend wants to go to a play, but neither of you wants to go to
something on their own. Sadly, your cell phone is dead, so you’ll just have to go to either the
movie theater or the playhouse, and hope your friend goes to the same location. We might
represent the game you and your friend are playing this way.

Movie Theater Playhouse

Movie Theater (2, 1) (0, 0)
Playhouse (0, 0) (1, 2)

In each cell now there are two numbers, representing first how good the outcome is for you,
and second how good it is for your friend. So if you both go to the movies, that’s the best
outcome for you, and the second-best for your friend. But if you go to different things, that’s
the worst result for both of you. We’ll look a bit at games like this where the party’s interests
are neither strictly allied nor strictly competitive.
Traditionally there is a large division between decision theory, where the outcome de-
pends just on your choice and the impersonal world, and game theory, where the outcome
depends on the choices made by multiple interacting agents. We’ll follow this tradition here,
focussing on decision theory for the first two-thirds of the course, and then shifting our at-
tention to game theory. But it’s worth noting that this division is fairly arbitrary. Some
decisions depend for their outcome on the choices of entities that are borderline agents,
such as animals or very young children. And some decisions depend for their outcome on
choices of agents that are only minimally interacting with you. For these reasons, among
others, we should be suspicious of theories that draw a sharp line between decision theory
and game theory.

2
1.2 Previews
Just thinking intuitively about decisions like whether to watch football, it seems clear that
how likely the various states of the world are is highly relevant to what you should do. If
you’re more or less certain that your team will win, and you’ll enjoy watching the win, then
you should watch the game. But if you’re more or less certain that your team will lose, then
it’s better to start working on the term paper. That intuition, that how likely the various
states are affects what the right decision is, is central to modern decision theory.
The best way we have to formally regiment likelihoods is probability theory. So we’ll
spend quite a bit of time in this course looking at probability, because it is central to good
decision making. In particular, we’ll be looking at four things.
First, we’ll spend some time going over the basics of probability theory itself. Many
people, most people in fact, make simple errors when trying to reason probabilistically. This
is especially true when trying to reason with so-called conditional probabilities. We’ll look
at a few common errors, and look at ways to avoid them.
Second, we’ll look at some questions that come up when we try to extend probability
theory to cases where there are infinitely many ways the world could be. Some issues that
come up in these cases affect how we understand probability, and in any case the issues are
philosophically interesting in their own right.
Third, we’ll look at some arguments as to why we should use probability theory, rather
than some other theory of uncertainty, in our reasoning. Outside of philosophy it is some-
times taken for granted that we should mathematically represent uncertainties as proba-
bilities, but this is in fact quite a striking and, if true, profound result. So we’ll pay some
attention to arguments in favour of using probabilities. Some of these arguments will also
be relevant to questions about whether we should represent the value of outcomes with
numbers.
Finally, we’ll look a little at where probabilities come from. The focus here will largely be
negative. We’ll look at reasons why some simple identifications of probabilities either with
numbers of options or with frequencies are unhelpful at best.
In the middle of the course, we’ll look at a few modern puzzles that have been the focus
of attention in decision theory. Later today we’ll go over a couple of examples that illustrate
what we’ll be covering in this section.
The final part of the course will be on game theory. We’ll be looking at some of the
famous examples of two person games. (We’ve already seen a version of one, the movie and
play game, above.) And we’ll be looking at the use of equilibrium concepts in analysing
various kinds of games.
We’ll end with a point that we mentioned above, the connection between decision theory
and game theory. Some parts of the standard treatment of game theory seem not to be
consistent with the best form of decision theory that we’ll look at. So we’ll want to see how
much revision is needed to accommodate our decision theoretic results.

3
1.3 Example: Newcomb
In front of you are two boxes, call them A and B. You call see that in box B there is $1000,
but you cannot see what is in box A. You have a choice, but not perhaps the one you were
expecting. Your first option is to take just box A, whose contents you do not know. Your
other option is to take both box A and box B, with the extra $1000.
There is, as you may have guessed, a catch. A demon has predicted whether you will
take just one box or take two boxes. The demon is very good at predicting these things –
in the past she has made many similar predictions and been right every time. If the demon
predicts that you will take both boxes, then she’s put nothing in box A. If the demon predicts
you will take just one box, she has put $1,000,000 in box A. So the table looks like this.

Predicts 1 box Predicts 2 boxes

Take 1 box $1,000,000 $0
Take 2 boxes $1,001,000 $1,000

There are interesting arguments for each of the two options here.
The argument for taking just one box is easy. The way the story has been set up, lots
of people have taken this challenge before you. Those that have taken 1 box have walked
away with a million dollars. Those that have taken both have walked away with a thousand
dollars. You’d prefer to be in the first group to being in the second group, so you should take
just one box.
The argument for taking both boxes is also easy. Either the demon has put the million
in box A or she hasn’t. If she has, you’re better off taking both boxes. That way you’ll get
$1,001,000 rather than $1,000,000. If she has not, you’re better off taking both boxes. That
way you’ll get $1,000 rather than $0. Either way, you’re better off taking both boxes, so you
should do that.
Both arguments seem quite strong. The problem is that they lead to incompatible con-
clusions. So which is correct?

1.4 Example: Sleeping Beauty

Sleeping Beauty is about to undergo a slightly complicated experiment. It is now Sunday
night, and a fair coin is about to be tossed, though Sleeping Beauty won’t see how it lands.
Then she will be asked a question, and then she’ll go to sleep. She’ll be woken up on Monday,
asked the same question, and then she’ll go back to sleep, and her memory of being woken
on Monday will be wiped. Then, if (and only if) the coin landed tails, she’ll be woken on
Tuesday, and asked the same question, and then she’ll go back to sleep. Finally, she’ll wake
on Wednesday.
The question she’ll be asked is How probable do you think it is that the coin landed
heads? What answers should she give

1. When she is asked on Sunday?

2. When she is asked on Monday?
3. If she is asked on Tuesday?

4
 It seems plausible to suggest that the answers to questions 2 and 3 should be the same.
After all, given that Sleeping Beauty will have forgotten about the Monday waking if she
wakes on Tuesday, then she won’t be able to tell the difference between the Monday and
Tuesday waking. So she should give the same answers on Monday and Tuesday. We’ll as-
sume that in what follows.
First, there seems to be a very good argument for answering 1/2 to question 1. It’s a fair
coin, so it has a probability of 1/2 of landing heads. And it has just been tossed, and there
hasn’t been any ‘funny business’. So that should be the answer.
Second, there seems to be a good, if a little complicated, argument for answering 1/3 to
questions 2 and 3. Assume that questions 2 and 3 are in some sense the same question. And
assume that Sleeping Beauty undergoes this experiment many times. Then she’ll be asked
the question twice as often when the coin lands tails as when it lands heads. That’s because
when it lands tails, she’ll be asked that question twice, but only once when it lands heads.
So only 1/3 of the time when she’s asked this question, will it be true that the coin landed
heads. And plausibly, if you’re going to be repeatedly asked How probable is it that such-
and-such happened, and 1/3 of the time when you’re asked that question, such-and-such will
have happened, then you should answer 1/3 each time.
Finally, there seems to be a good argument for answering questions 1 and 2 the same way.
After all, Sleeping Beauty doesn’t learn anything new between the two questions. She wakes
up, but she knew she was going to wake up. And she’s asked the question, but she knew
she was going to be asked the question. And it seems like a decent principle that if nothing
happens between Sunday and Monday to give you new evidence about a proposition, the
probability that you think it did happen shouldn’t change.
But of course, these three arguments can’t all be correct. So we have to decide which one
is incorrect.

Upcoming
These are just two of the puzzles we’ll be looking at as the course proceeds. Some of these
will be decision puzzles, like Newcomb’s Problem. Some of them will be probability puzzles
that are related to decision theory, like Sleeping Beauty. And some will be game puzzles. I
hope the puzzles are somewhat interesting. I hope even more that we learn something from
them.

5
 Chapter 2

Simple Reasoning Strategies

2.1 Dominance Reasoning
The simplest rule we can use for decision making is never choose dominated options. There
is a stronger and a weaker version of this rule.
An option A strongly dominated another option B if in every state, A leads to better
outcomes than B. A weakly dominates B if in every state, A leads to at least as good an
outcome as B, and in some states it leads to better outcomes.
We can use each of these as decision principles. The dominance principle we’ll be pri-
marily interested in says that if A strongly dominates B, then A should be preferred to B. We
get a slightly stronger principle if we use weak dominance. That is, we get a slightly stronger
principle if we say that whenever A weakly dominates B, A should be chosen over B. It’s a
stronger principle because it applies in more cases — that is, whenever A strongly dominates
B, it also weakly dominates B.
Dominance principles seem very intuitive when applied to everyday decision cases.
Consider, for example, a revised version of our case about choosing whether to watch foot-
ball or work on a term paper. Imagine that you’ll do very badly on the term paper if you
leave it to the last minute. And imagine that the term paper is vitally important for some-
thing that matters to your future. Then we might set up the decision table as follows.

Your team wins Your team loses

Watch football 2 1
Work on paper 4 3

If your team wins, you are better off working on the paper, since 4 > 2. And if your team
loses, you are better off working on the paper, since 3 > 1. So either way you are better off
working on the paper. So you should work on the paper.

6
2.2 States and Choices
Here is an example from Jim Joyce that suggests that dominance might not be as straight-
forward a rule as we suggested above.

Suupose you have just parked in a seedy neighborhood when a man approaches
and offers to “protect” your car from harm for $10. You recognize this as ex-
tortion and have heard that people who refuse “protection” invariably return
to find their windshields smashed. Those who pay find their cars intact. You
cannot park anywhere else because you are late for an important meeting. It
costs $400 to replace a windshield. Should you buy “protection”? Dominance
says that you should not. Since you would rather have the extra $10 both in
the even that your windshield is smashed and in the event that it is not, Domi-
nance tells you not to pay. (Joyce, The Foundations of Causal Decision Theory,
pp 115-6.)

We can put this in a table to make the dominance argument that Joyce suggests clearer.

Broken Windshield Unbroken Windshield

Pay extortion -$410 -$10
Don’t pay -$400 0

In each column, the number in the ‘Don’t pay’ row is higher than the number in the ‘Pay
extortion’ row. So it looks just like the case above where we said dominance gives a clear
answer about what to do. But the conclusion is crazy. Here is how Joyce explains what goes
wrong in the dominance argument.

Of course, this is absurd. Your choice has a direct influence on the state of the
world; refusing to pay makes it likly that your windshield will be smashed while
paying makes this unlikely. The extortionist is a despicable person, but he has
you over a barrel and investing a mere $10 now saves $400 down the line. You
should pay now (and alert the police later).

This seems like a general principle we should endorse. We should define states as being,
intuitively, independent of choices. The idea behind the tables we’ve been using is that the
outcome should depend on two factors - what you do and what the world does. If the ‘states’
are dependent on what choice you make, then we won’t have successfully ‘factorised’ the
dependence of outcomes into these two components.
We’ve used a very intuitive notion of ‘independence’ here, and we’ll have a lot more
to say about that in later sections. It turns out that there are a lot of ways to think about
independence, and they yield different recommendations about what to do. For now, we’ll
try to use ‘states’ that are clearly independent of the choices we make.

7
2.3 Maximin and Maximax
Dominance is a (relatively) uncontroversial rule, but it doesn’t cover a lot of cases. We’ll
start now lookintg at rules that are more or less comprehensive. To start off, let’s consider
rules that we might consider rules for optimists and pessimists respectively.
The Maximax rule says that you should maximise the maximum outcome you can get.
Basically, consider the best possible outcome, consider what you’d have to do to bring that
about, and do it. In general, this isn’t a very plausible rule. It recommends taking any kind
of gamble that you are offered. If you took this rule to Wall St, it would recommend buying
the riskiest derivatives you could find, because they might turn out to have the best results.
Perhaps needless to say, I don’t recommend that strategy.
The Maximin rule says that you should maximise the minimum outcome you can get.
So for every choice, you look at the worst-case scenario for that choice. You then pick the
option that has the least bad worst case scenario. Consider the following list of preferences
from our watch football/work on paper example.

Your team wins Your team loses

Watch football 4 1
Work on paper 3 2

So you’d prefer your team to win, and you’d prefer to watch if they win, and work if they lose.
So the worst case scenario if you watch the game is that they lose - the worst case scenario
of all in the game. But the worst case scenario if you don’t watch is also that they lose. Still
that wasn’t as bad as watching the game and seeing them lose. So you should work on the
paper.
We can change the example a little without changing the recommendation.

Your team wins Your team loses

Watch football 4 1
Work on paper 2 3

In this example, your regret at missing the game overrides your desire for your team to win.
So if you don’t watch, you’d prefer that they lose. Still the worst case scenario is you don’t
watch is 2, and the worst case scenario if you do watch is 1. So, according to maximin, you
should not watch.
Note in this case that the worst case scenario is a different state for different choices.
Maximin doesn’t require that you pick some ‘absolute’ worst-case scenario and decide on
the assumption it is going to happen. Rather, you look at different worst case scenarios for
different choices, and compare them.

8
2.4 Ordinal and Cardinal Utilities
All of the rules we’ve looked at so far depend only on the ranking of various options. They
don’t depend on how much we prefer one option over another. They just depend on which
order we rank goods is.
To use the technical language, so far we’ve just looked at rules that just rely on ordinal
utilities. The term ordinal here means that we only look at the order of the options. The
rules that we’ll look at rely on cardinal utilities. Whenever we’re associating outcomes with
numbers in a way that the magnitudes of the differences between the numbers matters, we’re
using cardinal utilities.
It is rather intuitive that something more than the ordering of outcomes should matter
to what decisions we make. Imagine that two agents, Chris and Robin, each have to make
a decision between two airlines to fly them from New York to San Francisco. One airline is
more expensive, the other is more reliable. To oversimplify things, let’s say the unreliable
airline runs well in good weather, but in bad weather, things go wrong. And Chris and Robin
have no way of finding out what the weather along the way will be. They would prefer to
save money, but they’d certainly not prefer for things to go badly wrong. So they face the
following decision table.

Good weather Bad weather

Fly cheap airline 4 1
Fly good airline 3 2

If we’re just looking at the ordering of outcomes, that is the decision problem facing both
Chris and Robin.
But now let’s fill in some more details about the cheap airlines they could fly. The cheap
airline that Chris might fly has a problem with luggage. If the weather is bad, their passen-
gers’ luggage will be a day late getting to San Francisco. The cheap airline that Robin might
fly has a problem with staying in the air. If the weather is bad, their plane will crash.
Those seem like very different decision problems. It might be worth risking one’s luggage
being a day late in order to get a cheap plane ticket. It’s not worth risking, seriously risking,
a plane crash. (Of course, we all take some risk of being in a plane crash, unless we only ever
fly the most reliable airline that we possibly could.) That’s to say, Chris and Robin are facing
very different decision problems, even though the ranking of the four possible outcomes
is the same in each of their cases. So it seems like some decision rules should be sensitive
to magnitudes of differences between options. The first kind of rule we’ll look at uses the
notion of regret.

9
2.5 Regret
Whenever you are faced with a decision problem without a dominating option, there is a
chance that you’ll end up taking an option that turns out to be sub-optimal. If that happens
there is a chance that you’ll regret the choice you take. That isn’t always the case. Sometimes
you decide that you’re happy with the choice you made after all. Sometimes you’re in no
position to regret what you chose because the combination of your choice and the world
leaves you dead.
Despite these complications, we’ll define the regret of a choice to be the difference be-
tween the value of the best choice given that state, and the value of the choice in question. So
imagine that you have a choice between going to the movies, going on a picnic or going to
a baseball game. And the world might produce a sunny day, a light rain day, or a thunder-
storm. We might imagine that your values for the nine possible choice-world combinations
are as follows.

Sunny Light rain Thunderstorm

Picnic 20 5 0
Baseball 15 2 6
Movies 8 10 9

Then the amount of regret associated with each choice, in each state, is as follows

Sunny Light rain Thunderstorm

Picnic 0 5 9
Baseball 5 8 3
Movies 12 0 0

Look at the middle cell in the table, the 8 in the baseball row and light rain column. The
reason that’s a 8 is that in that possibility, you get utility 2. But you could have got utility 10
from going to the movies. So the regret level is 10 - 2, that is, 8.
There are a few rules that we can describe using the notion of regret. The most com-
monly discussed one is called Minimax regret. The idea behind this rule is that you look at
what the maximum possible regret is for each option. So in the above example, the picnic
could end up with a regret of 9, the baseball with a regret of 8, and the movies with a regret
of 12. Then you pick the option with the lowest maximum possible regret. In this case, that’s
the baseball.
The minimax regret rule leads to plausible outcomes in a lot of cases. But it has one odd
structural property. In this case it recommends choosing the baseball over the movies and
picnic. Indeed, it thinks going to the movies is the worst option of all. But now imagine that
the picnic is ruled out as an option. (Perhaps we find out that we don’t have any way to get
picnic food.) Then we have the following table.

10
 Sunny Light rain Thunderstorm
Baseball 15 2 6
Movies 8 10 9

And now the amount of regret associated with each option is as follows.

Sunny Light rain Thunderstorm

Baseball 0 8 3
Movies 7 0 0

Now the maximum regret associated with going to the baseball is 8. And the maximum
regret associated with going to the movies is 7. So minimax regret recommends going to
the movies.
Something very odd just happened. We had settled on a decision: going to the baseball.
Then an option that we’d decided against, a seemingly irrelevant option, was ruled out. And
because of that we made a new decision: going to the movies. It seems that this is an odd
result. It violates what decision theorists call the Irrelevance of Independence Alternatives.
Formally, this principle says that if option C is chosen from some set S of options, then C
should be chosen from any set of options that (a) includes C and (b) only includes choices
in S. The minimax regret rule violates this principle, and that seems like an unattractive
feature of the rule.

11
2.6 Exercises
2.6.1 St Crispin’s Day Speech
In his play Henry V, Shakespeare gives the title character the following little speech. The
context is that the English are about to go to battle with the French at Agincourt, and they
are heavily outnumbered. The king’s cousin Westmoreland has said that he wishes they had
more troops, and Henry strongly disagrees.

What’s he that wishes so?

My cousin Westmoreland? No, my fair cousin;
If we are marked to die, we are enough
To do our country loss; and if to live,
The fewer men, the greater share of honor.
God’s will! I pray thee, wish not one man more.

Is the decision principle Henry is using here (a) dominance, (b) maximin, (c) maximax or
(d) minimax regret? Is his argument persuasive?

2.6.2 Dominance and Regret

Assume that in a decision problem, choice C is a dominating option. Will the minimax
regret rule recommend choosing C? Defend your answer; i.e. say why C must be chosen, it
must not be chosen, or why there isn’t enough information to tell whether C will be chosen.

2.6.3 Irrelevance of Independent Alternatives

Sam always chooses by the maximax rule. Will Sam’s choices satisfy the irrelevance of in-
dependent alternatives condition? Say why or why not.

2.6.4 Applying the Rules

For each of the following decision tables, say which decision would be preferred by (a) the
maximin rule, (b) the maximax rule and (c) the minimax regret rule. Also say whether the
minimax regret rule would lead to a different choice if a non-chosen option were elminated.
(You just have to give answers here, not show your workings.)

S1 S2 S3
C1 9 5 1
C2 8 6 3
C3 7 2 4

S1 S2 S3
C1 15 2 1
C2 9 9 9
C3 4 4 16

12
 Chapter 3

Uncertainty
3.1 Likely Outcomes
Earlier we considered the a decision problem, basically deciding what to do with a Sunday
afternoon, that had the following table.

Sunny Light rain Thunderstorm

Picnic 20 5 0
Baseball 15 2 6
Movies 8 10 9

We looked at how a few different decision rules would treat this decision. The maximin
rule would recommend going to the movies, the maximax rule going to the picnic, and the
minimax regret rule going to the baseball.
But if we were faced with that kind of decision in real life, we wouldn’t sit down to
start thinking about which of those three rules were correct, and using the answer to that
philosophical question to determine what to do. Rather, we’d consult a weather forecast. If
it looked like it was going to be sunny, we’d go on a picnic. If it looked like it was going to
rain, we’d go to the movie. What’s relevant is how likely each of the three states of the world
are. That’s something none of our decision rules to date have considered, and it seems like
a large omission.
In general, how likely various states are plays a major role in deciding what to do. Con-
sider the following broad kind of decision problem. There is a particular disease that, if you
catch it and don’t have any drugs to treat it with, is likely fatal. Buying the drugs in question
will cost $500. Do you buy the drugs?
Well, that probably depends on how likely it is that you’ll catch the disease in the first
place. The case isn’t entirely hypothetical. You or I could, at this moment, be stockpiling
drugs that treat anthrax poisoning, or avian flu. I’m not buying drugs to defend against
either thing. If it looked more likely that there would be more terrorist attacks using anthrax,
or an avian flu epidemic, then it would be sensible to spend $500, and perhaps a lot more,
defending against them. As it stands, that doesn’t seem particularly sensible. (I have no
idea exactly how much buying the relevant drugs would cost; the $500 figure was somewhat
made up. I suspect it would be a rolling cost because the drugs would go ‘stale’.)

13
 We’ll start off today looking at various decision rules that might be employed taking
account of the likelihood of various outcomes. Then we’ll look at what we might mean by
likelihoods. This will start us down the track to discussions of probability, a subject that
we’ll be interested in for most of the rest of the course.

3.2 Do What’s Likely to Work

The following decision rule doesn’t have a catchy name, but I’ll call it Do What’s Likely to
Work. The idea is that we should look at the various states that could come about, and decide
which of them is most likely to actually happen. This is more or less what we would do in
the decision above about what to do with a Sunday afternoon. The rule says then we should
make the choice that will result in the best outcome in that most likely of states.
The rule has two nice advantages. First, it doesn’t require a very sophisticated theory of
likelihoods. It just requires us to be able to rank the various states in terms of how likely
they are. Using some language from the previous section, we rely on a ordinarl rather than a
cardinal theory of likelihoods. Second, it matches up well enough with a lot of our everyday
decisions. In real life cases like the above example, we really do decide what state is likely to
be actual (i.e. decide what the weather is likely to be) then decide what would be best to do
in that circumstance.
But the rule also leads to implausible recommendations in other real life cases. Indeed,
in some cases it is so implausible that it seems that it must at some level be deeply mistaken.
Here is a simple example of such a case.
You have been exposed to a deadly virus. About 1/3 of people who are exposed to the
virus are infected by it, and all those infected by it die unless they receive a vaccine. By
the time any symptoms of the virus show up, it is too late for the vaccine to work. You are
offered a vaccine for $500. Do you take it or not?
Well, the most likely state of the world is that you don’t have the virus. After all, only 1/3
of people who are exposed catch the virus. The other 2/3 do not, and the odds are that you
are in that group. And if you don’t have the virus, it isn’t worth paying $500 for a vaccine
against a virus you haven’t caught. So by “Do What’s Likely to Work”, you should decline
the vaccine.
But that’s crazy! It seems as clear as anything that you should pay for the vaccine. You’re
in serious danger of dying here, and getting rid of that risk for $500 seems like a good deal.
So “Do What’s Likely to Work” gives you the wrong result. There’s a reason for this. You
stand to lose a lot if you die. And while $500 is a lot of money, it’s a lot less of a loss than
dying. Whenever the downside is very different depending on which choice you make,
sometimes you should avoid the bigger loss, rather than doing the thing that is most likely
to lead to the right result.
Indeed, sometimes the sensible decision is one that leads to the best outcome in no
possible states at all. Consider the following situation. You’ve caught a nasty virus, which
will be fatal unless treated. Happily, there is a treatment for the virus, and it only costs $100.
Unhappily, there are two strands of the virus, call them A and B. And each strand requires
a different treatment. If you have the A strand, and only get the treatment for the B virus,
you’ll die. Happily, you can have each of the two treatments; they don’t interact with each
other in nasty ways. So here are your options.

14
 Have strand A Have strand B
Get treatment A only Pay $100 + live Pay $100 + die
Get treatment B only Pay $100 + die Pay $100 + live
Get both treatments Pay $200 + live Pay $200 + live

Now the sensible thing to do is to get both treatments. But if you have strand A, the best
thing to do is to get treatment A only. And if you have strand B, the best thing to do is to
get treatment B only. There is no state whatsoever in which getting both treatments leads
to the best outcome. Note that “Do What’s Likely to Work” only ever recommends options
that are the best in some state or other. So it’s a real problem that sometimes the thing to do
does not produce the best outcome in any situation.

3.3 Probability and Uncertainty

As I mentioned above, none of the rules we’d looked at before today took into account the
likelihood of the various states of the world. Some authors have been tempted to see this as
a feature not a bug. To see why, we need to look at a common three-fold distinction between
states.
There are lots of things we know, even that we’re certain about. If we are certain which
state of the world will be actual, call the decision we face a decision under certainty.
Some times we don’t know which state will be actual. But we can state precisely what the
probability is that each of the states in question will be actual. For instance, if we’re trying to
decide whether to bet on a roulette wheel, then the relevant states will be the 37 or 38 slots
in which the ball can land. We can’t know which of those will happen, but we do know the
probability of each possible state. In cases where we can state the relevant probabilities, call
the decision we face a decision under risk.
In other cases, we can’t even state any probabilities. Imagine the following (not entirely
unrealistic) case. You have an option to invest in a speculative mining venture. The people
doing the projections for the investment say that it will be a profitable investment, over its
lifetime, if private cars running primarily on fossil fuel products are still the dominant form
of transportation in 20 years time. Maybe that will happen, maybe it won’t. It depends a
lot on how non fossil-fuel energy projects go, and I gather that that’s very hard to predict.
Call such a decision, one where we can’t even assign probabilities to states, a decision under
uncertainty.
It is sometimes proposed that rules like maximin, and minimax regret, while they are
clearly bad rules to use for decisions under risk, might be good rules for decisions under
uncertainty. I suspect that isn’t correct, largely because I suspect the distinction between
decisions under risk and decisions under uncertainty is not as sharp as the above tripartite
distinction suggests. Here is a famous passage from John Maynard Keynes, written in 1937,
describing the distinction between risk and uncertainty.

By “uncertain” knowledge, let me explain, I do not mean merely to distinguish

what is known for certain from what is only probable. The game of roulette is
not subject, in this sense, to uncertainty; nor is the prospect of a Victory bond
being drawn. Or, again, the expectation of life is only slightly uncertain. Even
the weather is only moderately uncertain. The sense in which I am using the

15
 term is that in which the prospect of a European war is uncertain, or the price of
copper and the rate of interest twenty years hence, or the obsolescence of a new
invention, or the position of private wealth owners in the social system in 1970.
About these matters there is no scientific basis on which to form any calculable
probability whatever. We simply do not know. Nevertheless, the necessity for
action and for decision compels us as practical men to do our best to overlook
this awkward fact and to behave exactly as we should if we had behind us a good
Benthamite calculation of a series of prospective advantages and disadvantages,
each multiplied by its appropriate probability, waiting to be summed.

There’s something very important about how Keynes sets up the distinction between risk
and uncertainty here. He says that it is a matter of degree. Some things are very uncertain,
such as the position of wealth holders in the social system a generation hence. Some things
are a little uncertain, such as the weather in a week’s time. We need a way of thinking about
risk and uncertainty that allows that in many cases, we can’t say exactly what the relevant
probabilities are, but we can say something about the comparative likelihoods.
Let’s look more closely at the case of the weather. In particular, think about decisions
that you have to make which turn on what the weather will be like in 7 to 10 days time.
These are a particularly tricky range of cases to think about.
If your decision turns on what the weather will be like in the distant future, you can
look at historical data. That data might not tell you much about the particular day you’re
interested in, but it will be reasonably helpful in setting probabilities. For instance, if it has
historically rained on 17% of August days in your hometown, then it isn’t utterly crazy to
think the probability it will rain on August 19 in 3 years time is about 0.17.
If your decision turns on what the weather will be like in the near future, such as the next
few hours or days, you have a lot of information ready to hand on which to base a decision.
Looking out the window is a decent guide to what the weather will be like for the next hour,
and looking up a professional weather service is a decent guide to what it will be for days
after that.
But in between those two it is hard. What do you think if historically it rarely rains at
this time of year, but the forecasters think there’s a chance a storm is brewing out west that
could arrive in 7 to 10 days? It’s hard even to assign probabilities to whether it will rain.
But this doesn’t mean that we should throw out all information we have about relative
likelihoods. I don’t know what the weather will be like in 10 days time, and I can’t even
sensibly assign probabilities to outcomes, but I’m not in a state of complete uncertainty.
I have a little information, and that information is useful in making decisions. Imagine
that I’m faced with the following decision table. The numbers at the top refer to what the
temperature will be, to the nearest 10 degrees Fahrenheit, 8 days from now, here in New
York in late summer.

60 70 80 90
Have picnic 0 4 5 6
Watch baseball 2 3 4 5

16
Both the maximin rule, and the minimax regret rule say that I should watch baseball rather
than having a picnic. (Exercise: prove this.) But this seems wrong. I don’t know exactly
how probable the various outcomes are, but I know that 60 degree days in late summer are
pretty rare, and nothing much in the long range forecast suggests that 8 days time will be
unseasonally mild.
The point is, even when we can’t say exactly how probable the various states are, we
still might be able to say something inexact. We might be able to say that some state is fairly
likely, or that another is just about certain not to happen. And that can be useful information
for decision making purposes. Rules like minimax regret throw out that information, and
that seems to make them bad rules.
We won’t get to it in these notes, but it’s important to be able to be able to think about
these cases where we have some information, but not complete information, about the
salient probabilities. The orthodox treatment in decision theory is to say that these cases
are rather like cases of decision making when you know the probabilities. That is, ortho-
doxy doesn’t distinguish decision making under risk and decision making under uncer-
tainty. We’re going to mostly assume here that orthodoxy is right. That’s in part because
it’s important to know what the standard views (in philosophy, economics, political science
and so on) are. And in part it’s because the orthodox views are close to being correct. Sadly,
getting clearer than that will be a subject for a much longer set of lecture notes.

17
 Chapter 4

Measures
4.1 Probability Defined
We talk informally about probabilities all the time. We might say that it is more probable
than not that such-and-such team will make the playoffs. Or we might say that it’s very prob-
able that a particular defendant will be convicted at his trial. Or that it isn’t very probable
that the next card will be the one we need to complete this royal flush.
We also talk formally about probability in mathematical contexts. Formally, a probabil-
ity function is a normalised measure over a possibility space. Below we’ll be saying a fair bit
about what each of those terms mean. We’ll start with measure, then say what a normalised
measure is, and finally (over the next two days) say something about possibility spaces.
There is a very important philosophical question about the connection between our in-
formal talk and our formal talk. In particular, it is a very deep question whether this par-
ticular kind of formal model is the right model to represent our informal, intuitive concept.
The vast majority of philosophers, statisticians, economists and others who work on these
topics think it is, though as always there are dissenters. We’ll be spending a fair bit of time
later in this course on this philosophical question. But before we can even answer that ques-
tion we need to understand what the mathematicians are talking about when they talk about
probabilities. And that requires starting with the notion of a measure.

4.2 Measures
A measure is a function from ‘regions’ of some space to non-negative numbers with the fol-
lowing property. If A is a region that divides exactly into regions B and C, then the measure
of A is the sum of the measures of B and C. And more generally, if A divides exactly into
regions B1 , B2 , ..., Bn , then the measure of A will be the sum of the measures of B1 , B2 , ...
and Bn .
Here’s a simple example of a measure: the function that takes as input any part of New
York City, and returns as output the population of that part. Assume that the following
numbers are the populations of New York’s five boroughs. (These numbers are far from
accurate.)

18
 Borough Population
Brooklyn 2,500,000
Queens 2,000,000
Manhattan 1,500,000
The Bronx 1,000,000
Staten Island 500,000

We can already think of this as a function, with the left hand column giving the inputs, and
the right hand column the values. Now if this function is a measure, it should be additive
in the sense described above. So consider the part of New York City that’s on Long Island.
That’s just Brooklyn plus Queens. If the population function is a measure, the value of that
function, as applied to the Long Island part of New York, should be 2,500,000 plus 2,000,000,
i.e. 4,500,000. And that makes sense: the population of Brooklyn plus Queens just is the
population of Brooklyn plus the population of Queens.
Not every function from regions to numbers is a measure. Consider the function that
takes a region of New York City as input, and returns as output the proportion of people
in that region who are New York Mets fans. We can imagine that this function has the
following values.

Borough Mets Proportion

Brooklyn 0.6
Queens 0.75
Manhattan 0.5
The Bronx 0.25
Staten Island 0.5

Now think again about the part of New York we discussed above: the Brooklyn plus Queens
part. What proportion of people in that part of the city are Mets fans? We certainly can’t
figure that out by just looking at the Brooklyn number from the above table, 0.6, and the
Queens number, 0.75, and adding them together. That would yield the absurd result that
the proportion of people in that part of the city who are Mets fans is 1.35.
That’s to say, the function from a region to the proportion of people in that region who
are Mets fans is not a measure. Measures are functions that are always additive over sub-
regions. The value of the function applied to a whole region is the sum of the values the
function takes when applied to the parts. ‘Counting’ functions, like population, have this
property.
The measure function we looked at above takes real regions, parts of New York City, as
inputs. But measures can also be defined over things that are suitably analogous to regions.
Imagine a family of four children, named below, who eat the following amounts of meat at
dinner.

19
 Child Meat Consumption (g)
Alice 400
Bruce 300
Chuck 200
Daria 100

We can imagine a function that takes a group of children (possibly including just one child,
or even no children) as inputs, and has as output how many grams of meat those children
ate. This function will be a measure. If the ‘groups’ contain just the one child, the values of
the function will be given by the above table. If the group contains two children, the values
will be given by the addition rule. So for the group consisting of Alice and Chuck, the value
of the function will be 600. That’s because the amount of meat eaten by Alice and Chuck just
is the amount of meat eaten by Alice, plus the amount of meat eaten by Chuck. Whenever
the value of a function, as applied to a group, is the sum of the values of the function as
applied to the members, we have a measure function.

4.3 Normalised Measures

A measure function is defined over some regions. Usually one of those regions will be the
‘universe’ of the function; that is, the region made up of all those regions the function is
defined over. In the case where the regions are regions of physical space, as in our New
York example, that will just be the physical space consisting of all the smaller regions that
are inputs to the function. In our New York example, the universe is just New York City.
In cases where the regions are somewhat more metaphorical, as in the case of the children’s
meat-eating, the universe will also be defined somewhat more metaphorically. In that case,
it is just the group consisting of the four children.
However the universe is defined, a normalised measure is simply a measure function
where the value the function gives to the universe is 1. So for every sub-region of the uni-
verse, its measure can be understood as a proportion of the universe.
We can ‘normalise’ any measure by simply dividing each value through by the value of
the universe. If we wanted to normalise our New York City population measure, we would
simply divide all values by 7,500,000. The values we would then end up with are as follows.

Borough Population
Brooklyn 1/3
Queens 4/15
Manhattan 1/5
The Bronx 2/15
Staten Island 1/3

Some measures may not have a well-defined universe, and in those cases we cannot nor-
malise the measure. But generally normalisation is a simple matter of dividing everything
by the value the function takes when applied to the whole universe. And the benefit of doing
this is that it gives us a simple way of representing proportions.

20
4.4 Formalities
So far I’ve given a fairly informal description of what measures are, and what normalised
measures are. In this section we’re going to go over the details more formally. If you under-
stand the concepts well enough already, or if you aren’t familiar enough with set theory to
follow this section entirely, you should feel free to skip forward to the next section. Note that
this is a slightly simplified, and hence slightly inaccurate, presentation; we aren’t focussing
on issues to do with infinity.
A measure is a function m satisfying the following conditions.

1. The domain D is a set of sets.

2. The domain is closed under union, intersection and complementation with respect
to the relevant universe U. That is, if A ∈ D and B ∈ D, then (A ∪ B) ∈ D and
(A ∪ B) ∈ D and U \ A ∈ D
3. The range is a set of non-negative real numbers
4. The function is additive in the following sense: If A ∩ B = ∅, then m(A ∪ B) =
m(A) + m(B)

We can prove some important general results about measures using just these properties.
Note that we the following results follow more or less immediately from additivity.

1. m(A) = m(A ∩ B) + m(A ∩ (U \ B))

2. m(B) = m(A ∩ B) + m(B ∩ (U \ A))
3. m(A ∪ B) = m(A ∩ B) + m(A ∩ (U \ B)) + m(B ∩ (U \ A))

The first says that the measure of A is the measure of A’s intersection with B, plus the
measure of A’s intersection with the complement of B. The first says that the measure of
B is the measure of A’s intersection with B, plus the measure of B’s intersection with the
complement of A. In each case the point is that a set is just made up of its intersection with
some other set, plus its intersection with the complement of that set. The final line relies
on the fact that the union of A and B is made up of (i) their intersection, (ii) the part of A
that overlaps B’s complement and (iii) the part of B that overlaps A’s complement. So the
measure of A ∪ B should be the sum of the measure of those three sets.
Note that if we add up the LHS and RHS of lines 1 and 2 above, we get

m(A) + m(B) = m(A ∩ B) + m(A ∩ (U \ B)) + m(A ∩ B) + m(A ∩ (U \ B))

And subtracting m(A ∩ B) from each side, we get

m(A) + m(B) – m(A ∩ B) = m(A ∩ B) + m(A ∩ (U \ B)) + m(A ∩ (U \ B))

But that equation, plus line 3 above, entails that

m(A) + m(B) – m(A ∩ B) = m(A ∪ B)

And that identity holds whether or not A ∩ B is empty. If A ∩ B is empty, the result is just
equivalent to the addition postulate, but in general it is a stronger result, and one we’ll be
using a fair bit in what follows.

21
4.5 Possibility Space
Imagine you’re watching a baseball game. There are lots of ways we could get to the final
result, but there are just two ways the game could end. The home team could win, call this
possibility H, or the away team could win, call this possibility A.
Let’s complicate the example somewhat. Imagine that you’re watching one game while
keeping track of what’s going on in another game. Now there are four ways that the games
could end. Both home teams could win. The home team could win at your game while the
away team wins the other game. The away team could win at your game while the home
team wins the other game. Or both away teams could win. This is a little easier to represent
on a chart.

Your game Other game

H H
H A
A H
A A

Here H stands for home team winning, and A stands for away team winning. If we start to
consider a third game, there are now 8 possibilities. We started with 4 possibilities, but now
each of these divides in 2: one where the home team wins the third game, and one where
the away team wins. It’s just about impossible to represent these verbally, so we’ll just use a
chart.

Game 1 Game 2 Game 3

H H H
H H A
H A H
H A A
A H H
A H A
A A H
A A A

Of course, in general we’re interested in more things than just the results of baseball games.
But the same structure can be applied to many more cases.
Say that there are three propositions, p, q and r that we’re interested in. And assume
that all we’re interested in is whether each of these propositions is true or false. Then there
are eight possible ways things could turn out, relative to what we’re interested in. In the
following table, each row is a possibility. T means the proposition at the head of that column
is true, F means that it is false.

22
 p q r
T T T
T T F
T F T
T F F
F T T
F T F
F F T
F F F

These eight possibilities are the foundation of the possibility space we’ll use to build a prob-
ability function.
A measure is an additive function. So once you’ve set the values of the smallest parts,
you’ve fixed the values of the whole. That’s because for any larger part, you can work out its
value by summing the values of its smaller parts. We can see this in the above example. Once
you’ve fixed how much meat each child has eaten, you’ve fixed how much meat each group
of children have eaten. The same goes for probability functions. In the cases we’re interested
in, once you’ve fixed the measure, i.e. the probability of each of the eight basic possibilities
represented by the above eight rows, you’ve fixed the probability of all propositions that
we’re interested in.
For concreteness, let’s say the probability of each row is given as follows.

p q r
T T T 0.0008
T T F 0.008
T F T 0.08
T F F 0.8
F T T 0.0002
F T F 0.001
F F T 0.01
F F F 0.1

So the probability of the fourth row, where p is true while q and r are false, is 0.8. (Don’t
worry for now about where these numbers come from; we’ll spend much more time on
that in what follows.) Note that these numbers sum to 1. This is required; probabilities are
normalised measures, so they must sum to 1.
Then the probability of any proposition is simply the sum of the probabilities of each
row on which it is true. For instance, the probability of p is the sum of the probabilities of
the first four rows. That is, it is 0.0008 + 0.008 + 0.08 + 0.8, which is 0.8888.
In the next class we’ll look at how we tell which propositions are true on which rows.
Once we’ve done that, we’ll have a fairly large portion of the formalities needed to look at
many decision-theoretic puzzles.

23
 Chapter 5

Truth Tables
5.1 Compound Sentences
Some sentences have other sentences as parts. We’re going to be especially interested in
sentences that have the following structures, where A and B are themselves sentences.

• A and B; which we’ll write as A ∧ B

• A or B; which we’ll write as A ∨ B
• It is not the case that A; which we’ll write as ¬A

What’s special about these three compound formations is that the truth value of the
whole sentence is fixed by the truth value of the parts. In fact, we can present the relationship
between the truth value of the whole and the truth value of the parts using the truth tables
discussed in the previous chapter. Here are the tables for the three connectives. First for
and,

A B A∧B
T T T
T F F
F T F
F F F

Then for or. (Note that this is so-called inclusive disjunction. The whole sentence is true if
both disjuncts are true.)

A B A∨B
T T T
T F T
F T T
F F F

Finally for not.

24
 A ¬A
T F
F T

The important thing about this way of thinking about compound sentences is that it is re-
cursive. I said above that some sentences have other sentences as parts. The easiest cases of
this to think about are cases where A and B are atomic sentences, i.e. sentences that don’t
themselves have other sentences as parts. But nothing in the definitions we gave, or in the
truth tables, requires that. A and B themselves could also be compound. And when they
are, we can use truth tables to figure out how the truth value of the whole sentence relates
to the truth value of its smallest constituents.
It will be easiest to see this if we work through an example. So let’s spend some time
considering the following sentence.

(p ∧ q) ∨ ¬r
The sentence has the form A ∨ B. But in this case A is the compound sentence p ∧ q,
and B is the compound sentence ¬r. If we’re looking at the possible truth values of the three
sentences p, q and r, we saw in the previous chapter that there are 23 , i.e. 8 possibilities. And
they can be represented as follows.

p q r
T T T
T T F
T F T
T F F
F T T
F T F
F F T
F F F

It isn’t too hard, given what we said above, to see what the truth values of p ∧ q, and of ¬r
will be in each of those possibilities. The first of these, p ∧ q, is true at a possibility just in
case there’s a T in the first column (i.e. p is true) and a T in the second column (i.e. q is
true). The second sentence, ¬r is true just in case there’s an F in the third column (i.e. r is
false). So let’s represent all that on the table.

p q r p∧q ¬r
T T T T F
T T F T T
T F T F F
T F F F T
F T T F F
F T F F T
F F T F F
F F F F T

25
Now the whole sentence is a disjunction, i.e. an or sentence, with the fourth and fifth
columns representing the two disjuncts. So the whole sentence is true just in case either
there’s a T in the fourth column, i.e. p ∧ q is true, or a T in the fifth column, i.e. ¬r is true.
We can represent that on the table as well.

p q r p∧q ¬r (p ∧ q) ∨ ¬r
T T T T F T
T T F T T T
T F T F F F
T F F F T T
F T T F F F
F T F F T T
F F T F F F
F F F F T T

And this gives us the full range of dependencies of the truth value of our whole sentence on
the truth value of its parts.
This is relevant to probability because, as we’ve been stressing, probability is a measure
over possibility space. So if you want to work out the probability of a sentence like (p∧q)∨¬r,
one way is to work out the probability of each of the eight basic possibilities here, then work
out at which of those possibilities (p ∧ q) ∨ ¬r is true, then sum the probabilities of those
possibilities at which it is true. To illustrate this, let’s again use the table of probabilities from
the previous chapter.

p q r
T T T 0.0008
T T F 0.008
T F T 0.08
T F F 0.8
F T T 0.0002
F T F 0.001
F F T 0.01
F F F 0.1

If those are the probabilities of each basic possibility, then the probability of (p ∧ q) ∨ ¬r is
the sum of the values on the lines on which it is true. That is, it is the sum of the values on
lines 1, 2, 4, 6 and 8. That is, it is 0.0008 + 0.008 + 0.8 + 0.001 + 0.1, which is 0.9098.

26
5.2 Equivalence, Entailment, Inconsistency, and Logical Truth
To a first approximation, we can define logical equivalence and logical entailment within
the truth-table framework. The accounts we’ll give here aren’t quite accurate, and we’ll make
them a bit more precise in the next section. But they are on the right track, and they suggest
some results that are, as it turns out, true in the more accurate structure.
If two sentences have the same pattern of Ts and Fs in their truth table, they are logically
equivalent. Consider, for example, the sentences ¬A ∨ ¬B and ¬(A ∧ B). Their truth tables
are given in the fifth and seventh columns of this table.

A B ¬A ¬B ¬A ∨ ¬B A∧B ¬(A ∧ B)
T T F F F T F
T F F T T F T
F T T F T F T
F F T T T F T

Note that those two columns are the same. That means that the two sentences are logically
equivalent.
Now something important follows from the fact that the sentences are true in the same
rows. For each sentence, the probability of the sentence is the sum of the probabilities of
the rows in which it is true. But if the sentences are true in the same row, those are the same
sums in each case. So the probability of the two sentences is the same. This leads to an
important result.

• Logically equivalent sentences have the same probability

Note that we haven’t quite proven this yet, because our account of logical equivalence is not
quite accurate. But the result will turn out to hold when we fix that inaccuracy.
One of the notions that logicians care most about is validity. An argument with premises
A1 , A2 , ..., An and conclusion B is valid if it is impossible for the premises to be true and the
conclusion false. Slightly more colloquially, if the premises are true, then the conclusion
has to be true. Again, we can approximate this notion using truth tables. An argument is
invalid if there is a line where the premises are true and the conclusion false. An argument
is valid if there is no such line. That is, it is valid if in all possibilities where all the premises
are true, the conclusion is also true.
When the argument that has A as its only premise, and B as its conclusion, is valid, we
say that A entails B. If every line on the truth table where A is true is also a line where B is
true, then A entails B.
Again, this has consequences for probability. The probability of a sentence is the sum
of the probability of the possibilities in which it is true. If A entails B, then the possibili-
ties where B is true will include all the possibilities where A is true, and may include some
more. So the probability of B can’t be lower than the probability of A. That’s because each
of these probabilities are sums of non-negative numbers, and each of the summands in the
probability of A is also a summand in the probability of B.

• If A entails B, then the probability of B is at least as great as the probability of A

27
 The argument we’ve given for this is a little rough, because we’re working with an ap-
proximation of the definition of entailment, but it will turn out that the result goes through
even when we tidy up the details.
Two sentences are inconsistent if they cannot be true together. Roughly, that means
there is no line on the truth table where they are both true. Assume that A and B are incon-
sistent. So A is true at lines L1 , L2 , ..., Ln , and B is true at lines Ln+1 , ..., Lm , where these do
not overlap. So A ∨ B is true at lines L1 , L2 , ..., Ln , Ln+1 , ..., Lm . So the probability of A is the
probability of L1 plus the probability of L2 plus ... plus the probability of Ln . And the prob-
ability of B is the probability of Ln+1 plus ... plus the probability of Lm . And the probability
of A ∨ B is the probability of L1 plus the probability of L2 plus ... plus the probability of Ln
plus Ln+1 plus ... plus the probability of Lm . That’s to say

• If A and B are inconsistent, then the probability of A ∨ B equals the probability of

A plus the probability of B

This is just the addition rule for measures transposed to probabilities. And it is a crucial
rule, one that we will use all the time. (Indeed, it is sometimes taken to be the characteristic
axiom of probability theory. We will look at axiomatic approaches to probability in the next
chapter.)
Finally, a logical truth is something that is true in virtue of logic alone. It is true in all
possibilities, since what logic is does not change. A logical truth is entailed by any sentence.
And a logical truth only entails other sentences.
Any sentence that is true in all possibilities must have probability 1. That’s because prob-
ability is a normalised measure, and in a normalised measure, the measure of the universe
is 1. And a logical truth is true at every point in the ‘universe’ of logical space.

• Any logical truth has probability 1

5.3 Two Important Results

None of the three connectives is particularly hard to process, but the rule for negation may
well be the easiest of the lot. The truth value of ¬A is just the opposite of the truth value of A.
So if A is true at a line, then ¬A is false. And if A is false at a line, then ¬A is true. So exactly
one of A and ¬A is true at each line. So the sum of the probabilities of those propositions
must be 1.
We can get to this result another way. It is easy to see that A ∨ ¬A is a logical truth by
simply looking at its truth table.

A ¬A A ∨ ¬A
T F T
F T T

28
The sentence A ∨ ¬A is true on each line, so it is a logical truth. And logical truths have
probability 1. Now A and ¬A are clearly inconsistent. So the probability of their disjunction
equals the sum of their probabilities. That’s to say, Pr(A ∨ ¬A) = Pr(A) + Pr(¬A). But
Pr(A ∨ ¬A) = 1. So,

Pr(A) + Pr(¬A) = 1
One important consequence of this is that the probabilities of A and ¬A can’t vary in-
dependently. Knowing how probable A is settles how probable ¬A is.
The next result is slightly more complicated, but only a little. Consider the following
table of truth values and probabilities.

Pr A B A∧B A∨B
x1 T T T T
x2 T F F T
x3 F T F T
x4 F F F F

The variables in the first column represent the probability of each row. We can see from the
table that the following results all hold.

1. Pr(A) = x1 + x2 , since A is true on the first and second lines

2. Pr(B) = x1 + x3 , since B is true on the first and third lines
3. Pr(A ∧ B) = x1 , since A ∧ B is true on the first line
4. Pr(A ∨ B) = x1 + x2 + x3 , since A ∨ B is true on the first, second and third lines

Adding the first and second lines together, we get

Pr(A) + Pr(B) = x1 + x2 + x1 + x3

And adding the third and fourth lines together, we get

Pr(A ∧ B) + Pr(A ∨ B) = x1 + x1 + x2 + x3

And simply rearranging the variables a little reveals that

Pr(A) + Pr(B) = Pr(A ∧ B) + Pr(A ∨ B)

Again, this is a result that we will use a lot in what follows.

29
 Chapter 6

Axioms for Probability

6.1 Axioms of Probability
We’ve introduced probability so far through the truth tables. If you are concerned with some
finite number, say n of sentences, you can make up a truth table with 2n rows representing
all the possible combinations of truth values for those sentences. And then a probability
function is simply a measure defined over sets of those rows, i.e. sets of possibilities.
But we can also introduce probability more directly. A probability function is a function
that takes sentences as inputs, has outputs in [0, 1], and satisfies the following constraints.

• If A is a logical truth, then Pr(A) = 1

• If A and B are logically equivalent, then Pr(A) = Pr(B)
• If A and B are logically disjoint, i.e. ¬(A ∧ B) is a logical truth, then Pr(A) + Pr(B) =
Pr(A ∨ B)

To get a feel for how these axioms operate, I’ll run through a few proofs using the axioms.
The results we prove will be familiar from the previous chapter, but the interest here is in
seeing how the axioms interact with the definitions of logical truth, logical equivalence and
logical disjointedness to derive familiar results.

• Pr(A) + Pr(¬A) = 1

Proof: It is a logical truth that A ∨ ¬A. This can be easily seen on a truth table. So by axiom
1, Pr(A ∨ ¬A) = 1. The truth tables can also be used to show that ¬(A ∧ A) is a logical truth,
so A and ¬A are disjoint. So Pr(A) + Pr(¬A) = Pr(A ∨ ¬A). But since Pr(A ∨ ¬A) = 1, it
follows that Pr(A) + Pr(¬A) = 1.

• If A is a logical falsehood, i.e. ¬A is a logical truth, then Pr(A) = 0

Proof: If ¬A is a logical truth, then by axiom 1, Pr(¬A) = 1. We just proved that Pr(A) +
Pr(¬A) = 1. From this it follows that Pr(A) = 0.

• Pr(A) + Pr(B) = Pr(A ∨ B) + Pr(A ∧ B)

Proof: First, note that A is logically equivalent to (A ∧ B) ∨ (A ∧ ¬B), and that (A ∧ B) and
(A ∧ ¬B) are logically disjoint. We can see both these facts in the following truth table.

30
 A B ¬B (A ∧ B) (A ∧ ¬B) (A ∧ B) ∨ (A ∧ ¬B)
T T F T F T
T F T F T T
F T F F F F
F F T F F F

The first and sixth columns are identical, so A and (A ∧ B) ∨ (A ∧ ¬B). By axiom 2, that
means that Pr(A) = Pr((A ∧ B) ∨ (A ∧ ¬B)).
The fourth and fifth column never have a T on the same row, so (A ∧ B) and (A ∧ ¬B)
are disjoint. That means that Pr((A ∧ B) ∨ (A ∧ ¬B) = Pr((A ∧ B) + Pr(A ∧ ¬B). Putting
the two results together, we get that Pr(A) = Pr((A ∧ B) + Pr(A ∧ ¬B).
The next truth table is designed to get us two results. First, that A ∨ B is equivalent to
B ∨ (A ∧ ¬B). And second that B and (A ∧ ¬B) are disjoint.

A B A∨B ¬B A ∧ ¬B B ∨ (A ∧ ¬B)
T T T F F T
T F T T T T
F T T F F T
F F F T F F

Note that the third column, A ∨ B, and the sixth column, B ∨ (A ∧ ¬B), are identical. So
those two propositions are equivalent. So Pr(A ∨ B) = Pr(B ∨ (A ∧ ¬B)).
Note also that the second column, B and the fifth column, A∧¬B, have no Ts in common.
So they are disjoint. So Pr(B ∨ (A ∧ ¬B)) = Pr(B) + Pr(A ∧ ¬B). Putting the last two results
together, we get that Pr(A ∨ B) = Pr(B) + Pr(A ∧ ¬B).
If we add Pr(A ∧ B) to both sides of that last equation, we get Pr(A ∨ B) + Pr(A ∧ B) =
Pr(B)+Pr(A∧¬B)+Pr(A∧B). But note that we already proved that Pr(A∧¬B)+Pr(A∧B) =
Pr(A). So we can rewrite Pr(A ∨ B) + Pr(A ∧ B) = Pr(B) + Pr(A ∧ ¬B) + Pr(A ∧ B) as
Pr(A ∨ B) + Pr(A ∧ B) = Pr(B) + Pr(A). And simply rearranging terms around gives us
Pr(A) + Pr(B) = Pr(A ∨ B) + Pr(A ∧ B), which is what we set out to prove.

6.2 Truth Tables and Possibilities

So far we’ve been assuming that whenever we are interested in n sentences, there are 2n
possibilities. But this isn’t always the case. Sometimes a combination of truth values doesn’t
express a real possibility. Consider, for example, the case where A = Many people enjoyed
the play, and B = Some people enjoyed the play. Now we might start trying to draw up a truth
table as follows.

A B
T T
T F
F T
F F

31
But there’s something deeply wrong with this table. The second line doesn’t represent a
real possibility. It isn’t possible that it’s true that many people enjoyed the play, but false
that some people enjoyed the play. In fact there are only three real possibilities here. First,
many people (and hence some people) enjoyed the play. Second, some people, but not many
people, enjoyed the play. Third, no one enjoyed the play. That’s all the possibilities that there
are. There isn’t a fourth possibility.
In this case, A entails B, which is why there is no possibility where A is true and B is
false. In other cases there might be more complicated interrelations between sentences that
account for some of the lines not representing real possibilities. Consider, for instance, the
following case.

• A = Alice is taller than Betty

• B = Betty is taller than Carla
• C = Carla is taller than Alice

Again, we might try and have a regular, 8 line, truth table for these, as below.

A B C
T T T
T T F
T F T
T F F
F T T
F T F
F F T
F F F

But here the first line is not a genuine possibility. If Alice is taller than Betty, and Betty is
taller than Carla, then Carla can’t be taller than Alice. So there are, at most, 7 real possibilities
here. (We’ll leave the question of whether there are fewer than 7 possibilities as an exercise.)
Again, one of the apparent possibilities is not real.
The chance that there are lines on the truth tables that don’t represent real possibilities
means that we have to modify several of the definitions we offered above. More carefully,
we should say.

• Two sentences are A and B are logically equivalent if (and only if) they have the same
truth value at every line on the truth table that represents a real possibility.
• Some sentences A1 , ..., An entail a sentence B if (and only if) at every line which (a)
represents a real possibility and (b) each of A1 , ..., An is true, B is true. Another way
of putting this is that the argument from A1 , ..., An to B is valid.
• Two sentences A and B are logically disjoint if (and only if) there is no line which (a)
represents a real possibility and (b) they are both true at that line

32
 Surprisingly perhaps, we don’t have to change the definition of a probability function
all that much. We started off by saying that you got a probability function, defined over
A1 , ..., An by starting with the truth table for those sentences, all 2n rows of it, and assigning
numbers to each row in a way that they added up to 1. The probability of any sentence was
then the sum of the numbers assigned to each row at which it is true.
This needs to be changed a little. If something does not represent a real possibility, then
its negation is a logical truth. And all logical truths have to get probability 1. So we have to
assign 0 to every row that does not represent a real possibility.
But that’s the only change we have to make. Still, any way of assigning numbers to rows
such that the numbers sum to 1, and any row that does not represent a real possibility is
assigned 0, will be a probability function. And, as long as we are only interested in sentences
with A1 , An as parts, any probability function can be generated this way.
So in fact all of the proofs in the previous chapter of the notes will still go through.
There we generated a lot of results from the assumption that any probability function is a
measure over the possibility space generated by a truth table. And that assumption is, strictly
speaking, true. Any probability function is a measure over the possibility space generated
by a truth table. It’s true that some such measures are not probability functions because they
assign positive values to lines that don’t represent real possibilities. But that doesn’t matter
for the proofs we were making there.
The upshot is that we can, for the purposes of decision theory, continue to think about
probability functions using truth tables. Occasionally we will have to be a little more careful,
but for the most part, just assigning numbers to rows gives us all the basic probability theory
we will need.

6.3 Propositions and Possibilities

There are many things we can be uncertain about. Some of these concern matters of fact,
especially facts about the future. We can be uncertain about horseraces, or elections, or
the weather. And some of them concern matters to do with mathematics or logic. We
might be uncertain about whether two propositions are logically equivalent. Or we might
be uncertain whether a particular mathematical conjecture is true or false.
Sometimes our uncertainty about a subject matter relates to both things. I’m writing this
in the middle of hurricane season, and we’re frequently uncertain about what the hurricanes
will do. There are computer models to predict them, but the models are very complicated,
and take hours to produce results even once all the data is in. So we might also be uncertain
about a purely mathematical fact, namely what this model will predict given these inputs.
One of the consequences of the axioms for probability theory we gave above is that any
logical truth, and for current purposes at least mathematical truths count as logical truths,
get probability 1. This might seem counterintuitive. Surely we can sensibly say that such
and such a mathematical claim is likely to be true, or probable to be true. Or we can say that
someone’s logical conjecture is probably false. How could it be that the axioms of probability
say otherwise?
Well, the important thing to remember here is that what we’re developing is a formal,
mathematical notion. It remains an open question, indeed a deep philosophical question,
whether that mathematical notion is useful in making sense of our intuitive, informal notion
of what’s more or less likely, or more or less probable. It is natural to think at this point that

33
probability theory, the mathematical version, will not be of much help in modelling our
uncertainty about logic or mathematics.
At one level this should not be too surprising. In order to use a logical/mathematical
model, we have to use logic and mathematics. And to use logic and mathematics, we have
to presuppose that they are given and available to use. But that’s very close already to pre-
supposing that they aren’t at all uncertain. Now this little argument isn’t very formal, and it
certainly isn’t meant to be a conclusive proof that there couldn’t be a mathematical model
of uncertainty about mathematics. But it’s a reason to think that such a model would have
to solve some tricky conceptual questions that a model of uncertainty about the facts does
not have to solve.
And not only should this not be surprising, it should not necessarily be too worrying.
In decision theory, what we’re usually concerned with is uncertainty about the facts. It’s
possible that probability theory can be the foundation for an excellent model for uncertainty
about the facts even if such a model is a terrible tool for understanding uncertainty about
mathematics. In most areas of science, we don’t expect every model to solve every problem.
I mentioned above that at this time of year, we spend a lot of time looking at computer
models of hurricane behaviour. Those models are not particularly useful guides to, say,
snowfall over winter. (Let alone guides to who will win the next election.) But that doesn’t
make them bad hurricane models.
The same thing is going to happen here. We’re going to try to develop a mathematical
model for uncertainty about matters of fact. That model will be extremely useful, when
applied to its intended questions. If you apply the model to uncertainty about mathematics,
you’ll get the crazy result that no mathematical question could ever be uncertain, because
every mathematical truth gets probability 1, and every falsehood probability 0. That’s not a
sign the model is failing; it is a sign that it is being misapplied. (Caveat: Given that the model
has limits, we might worry about whether its limits are being breached in some applications.
This is a serious question about some applications of decision theory to the Sleeping Beauty
puzzle, for example.)
To end, I want to note a connection between this section and two large philosophical
debates. The first is about the relationship between mathematics and logic. The second is
about the nature of propositions. I’ll spend one all-too-brief paragraph on each.
I’ve freely moved between talk of logical truths and mathematical truths in the above.
Whether this is appropriate turns out to be a tricky philosophical question. One view about
the nature of mathematics, called logicisim, holds that mathematics is, in some sense, part
of logic. If that’s right, then mathematical truths are logical truths, and everything I’ve said is
fine. But logicism is very controversial, to put it mildly. So we shouldn’t simply assume that
mathematical truths are logical truths. But we can safely assume the following disjunction
is true. Either (a) simple arithmetical truths (which is all we’ve been relying on) are part of
logic, or (b) the definition of a probability function needs to be clarified so all logical and
(simple) mathematical truths get probability 1. With that assumption, everything I’ve said
here will go through.
I’ve taken probability functions to be defined over sentences. But it is more common in
mathematics, and perhaps more elegant, to define probability functions over sets of possi-
bilities. Now some philosophers, most notably Robert Stalnaker, have argued that sets of
possibilities also have a central philosophical role. They’ve argued that propositions, the

34
things we believe, assert, are uncertain about etc, just are sets of possibilities. If that’s right,
there’s a nice connection between the mathematical models of probability, and the psy-
chological notion of uncertainty we’re interested in. But this view is controversial. Many
philosophers think that, especially in logic and mathematics, there are many distinct propo-
sitions that are true in the same possibilities. (One can be uncertain about one mathematical
truth while being certain that another is true, they think.) In any case, one of the upshots
of the discussion above is that we’re going to write as if Stalnaker was right, i.e. as if sets
of possibilities are the things that we are certain/uncertain about. We’ll leave the tricky
philosophical questions about whether he’s actually right for another day.

35
6.4 Exercises
6.4.1 Truth Tables and Probabilities
Consider this table of possibilities and probabilities, that we’ve used before.

p q r
T T T 0.0008
T T F 0.008
T F T 0.08
T F F 0.8
F T T 0.0002
F T F 0.001
F F T 0.01
F F F 0.1

If those numbers on each row express the probability that the row is actual, what is the
probability of each of the following sentences?

1. q
2. ¬r
3. p∧q
4. q ∨ ¬r
5. p ∧ (q ∨ ¬r)
6. (¬p ∧ r) ∨ (r ∧ ¬q)

6.4.2 Tables and Proofs

There’s just one question here, but I want you to answer it twice. Make the following as-
sumptions.

• Pr(p ∨ q) = 0.84
• Pr(¬p ∨ q) = 0.77
• Pr(p ∨ ¬q) = 0.59

What I want you to figure out is, what is Pr(p). But I want you to show the workings out
for this twice.
First, I want you to use the information given to work out what the probability of each
row of the truth table is, and use that to work out Pr(p).
Second, I want an argument directly from the axioms for probability (plus facts about
logical relations, as necessary) that ends up with the right value for Pr(p).

36
6.4.3 Possibiilities
We discussed above the following example.

• A = Alice is taller than Betty

• B = Betty is taller than Carla
• C = Carla is taller than Alice

And we noted that one of the eight lines on the truth table, the top one, does not rep-
resent a real possibility. How many other lines on the truth table do not represent real
possibilities?

37
 Chapter 7

Conditional Probability
7.1 Conditional Probability
So far we’ve talked simply about the probability of various propositions. But sometimes
we’re not interested in the absolute probability of a proposition, we’re interested in its con-
ditional probability. That is, we’re interested in the probability of the proposition assuming
or conditional on some other proposition obtaining.
For example, imagine we’re trying to decide whether to go to a party. At first glance,
we might think that one of the factors that is relevant to our decision is the probability that
it will be a successful party. But on second thought that isn’t particularly relevant at all. If
the party is going to be unpleasant if we are there (because we’ll annoy the host) but quite
successful if we aren’t there, then it might be quite probable that it will be a successful party,
but that will be no reason at all for us to go. What matters is the probabiilty of it being a
good, happy party conditional on our being there.
It isn’t too hard to visualise how conditional probability works if we think of measures
over lines on the truth table. If we assume that something , call it B is true, then we should
‘zero out’, i.e. assign probability 0, to all the possibilities where B doesn’t obtain. We’re now
left with a measure over only the B-possibilities. The problem is that it isn’t a normalised
measure. The values will only sum to Pr(B), not to 1. We need to renormalise. So we divide
by Pr(B) and we get a probability back. In a formula, we’re left with

Pr(A ∧ B)
Pr(A|B) =
Pr(B)
We can work through an example of this using a table that we’ve seen once or twice in
the past.

p q r
T T T 0.0008
T T F 0.008
T F T 0.08
T F F 0.8
F T T 0.0002
F T F 0.001
F F T 0.01
F F F 0.1

38
Assume now that we’re trying to find the conditional probability of p given q. We could do
this in two different ways.
First, we could set the probability of any line where q is false to 0. So we will get the
following table.

p q r
T T T 0.0008
T T F 0.008
T F T 0
T F F 0
F T T 0.0002
F T F 0.001
F F T 0
F F F 0

The numbers don’t sum to 1 any more. They sum to 0.01. So we need to divide everything by
0.01. It’s sometimes easier to conceptualise this as multiplying by 1/Pr(q), i.e. by multiplying
by 100. Then we’ll end up with:

p q r
T T T 0.08
T T F 0.8
T F T 0
T F F 0
F T T 0.02
F T F 0.1
F F T 0
F F F 0

And since p is true on the top two lines, the ‘new’ probability of p is 0.88. That is, the
conditional probability of p given q is 0.88. As we were writing things above, Pr(p|q) = 0.88.
Alternatively we could just use the formula given above. Just adding up rows gives us
the following numbers.

Pr(p ∧ q) = 0.0008 + 0.008 = 0.0088

Pr(q) = 0.0008 + 0.008 + 0.0002 + 0.001 = 0.01

Then we can apply the formula.

Pr(p ∧ q)
Pr(p|q) =
Pr(q)
0.0088
=
0.01
= 0.88

39
7.2 Bayes Theorem
It is often easier to calculate conditional probabilities in the ‘inverse’ direction to what we
are interested in. That is, if we want to know Pr(A|B), it might be much easier to discover
Pr(B|A). In these cases, we use Bayes Theorem to get the right result. I’ll state Bayes Theorem
in two distinct ways, then show that the two ways are ultimately equivalent.

Pr(B|A)Pr(A)
Pr(A|B) =
Pr(B)
Pr(B|A)Pr(A)
=
Pr(B|A)Pr(A) + Pr(B|¬A)Pr(¬A)

These are equivalent because Pr(B) = Pr(B|A)Pr(A) + Pr(B|¬A)Pr(¬A). Since this is an

independently interesting result, it’s worth going through the proof of it. First note that

Pr(A ∧ B)
Pr(B|A)Pr(A) = Pr(A)
Pr(A)
= Pr(A ∧ B)

Pr(¬A ∧ B)
Pr(B|¬A)Pr(¬A) = Pr(¬A)
Pr¬(A)
= Pr(¬A ∧ B)

Adding those two together we get

Pr(B|A)Pr(A) + Pr(B|¬A)Pr(¬A) = Pr(A ∧ B) + Pr(¬A ∧ B)

= Pr((A ∧ B) ∨ (¬A ∧ B))
= Pr(B)

The second line uses the fact that A ∧ B and ¬A ∧ B are inconsistent, which can be verified
using the truth tables. And the third line uses the fact that (A ∧ B) ∨ (¬A ∧ B) is equivalent
to A, which can also be verified using truth tables. So we get a nice result, one that we’ll have
occasion to use a bit in what follows.

Pr(B) = Pr(B|A)Pr(A) + Pr(B|¬A)Pr(¬A)

So the two forms of Bayes Theorem are the same. We’ll often find ourselves in a position to
use the second form.
One kind of case where we have occasion to use Bayes Theorem is when we want to know
how significant a test finding is. So imagine we’re trying to decide whether the patient has
disease D, and we’re interested in how probable it is that the patient has the disease condi-
tional on them returning a test that’s positive for the disease. We also know the following
background facts.

• In the relevant demographic group, 5% of patients have the disease.

• When a patient has the disease, the test returns a position result 80% of the time

40
 • When a patient does not have the disease, the test returns a negative result 90% of the
time

So in some sense, the test is fairly reliable. It usually returns a positive result when applied
to disease carriers. And it usually returns a negative result when applied to non-carriers.
But as we’ll see when we apply Bayes Theorem, it is very unreliable in another sense. So let
A be that the patient has the disease, and B be that the patient returns a positive test. We
can use the above data to generate some ‘prior’ probabilities, i.e. probabilities that we use
prior to getting information about the test.
• Pr(A) = 0.05, and hence Pr(¬A) = 0.95
• Pr(B|A) = 0.8
• Pr(B|¬A) = 0.1

Now we can apply Bayes theorem in its second form.

Pr(B|A)Pr(A)
Pr(A|B) =
Pr(B|A)Pr(A) + Pr(B|¬A)Pr(¬A)
0.8 × 0.05
=
0.08 × 0.05 + 0.1 × 0.95
0.04
=
0.04 + 0.095
0.04
=
0.135
≈ 0.296

So in fact the probability of having the disease, conditional on having a positive test, is less
than 0.3. So in that sense the test is quite unreliable.
This is actually a quite important point. The fact that the probability of B given A is
quite high does not mean that the probability of A given B is equally high. By tweaking the
percentages in the example I gave you, you can come up with cases where the probability of
B given A is arbitrarily high, even 1, while the probability of A given B is arbitrarily low.
Confusing these two conditional probabilities is sometimes referred to as the prosecutors’
fallacy, though it’s not clear how many actual prosecutors are guilty of it! The thought is
that some prosecutors start with the premise that the probability of the defendant’s blood
(or DNA or whatever) matching the blood at the crime scene, conditional on the defendant
being innocent, is 1 in a billion (or whatever it exactly is). They conclude that the probability
of the defendant being innocent, conditional on their blood matching the crime scene, is
about 1 in a billion. Because of derivations like the one we just saw, that is a clearly invalid
move.

7.3 Conditionalisation
The following two concepts seem fairly closely related.

• The probability of some hypothesis H given evidence E

• The new probability of hypothesis H when evidence E comes in

41
In fact these are distinct concepts, though there are interesting philosophical questions
about how intimately they are connected.
The first one is a static concept. It says, at one particular time, what the probability of
H is given E. It doesn’t say anything about whether or not E actually obtains. It doesn’t
say anything about changing your views, or your probabilities. It just tells us something
about our current probabilities, i.e. our current measure on possibility space. And what it
tells us is what proportion of the space where E obtains is occupied by possibilities where
H obtains. (The talk of ‘proportion’ here is potentially misleading, since there’s no physical
space to measure. What we care about is the measure of the E ∧ H space as a proportion of
the measure of the E space.)
The second one is a dynamic concept. It says what we do when evidence E actually comes
in. Once this happens, old probabilities go out the window, because we have to adjust to
the new evidence that we have to hand. If E indicates H, then the probability of H should
presumably go up, for instance.
Because these are two distinct concepts, we’ll have two different symbols for them. We’ll
use Pr(H|E) for the static concept, and PrE (H) for the dynamic concept. So Pr(H|E) is what
the current probability of H given E is, and PrE (H) is what the probability of H will be when
we get evidence E.
Many philosophers think that these two should go together. More precisely, they think
that a rational agent always updates by conditionalisation. That’s just to say that for any ra-
tional agent, Pr(H|E) = PrE (H). When we get evidence E, we always replace the probability
of H with the probability of H given E.
The conditionalisation thesis occupies a quirky place in contemporary philosophy. On
the one hand it is almost universally accepted, and an extremely interesting set of theoret-
ical results have been built up using the assumption it is true. (Pretty much everything in
Bayesian philosophy of science relies in one way or another on the assumption that con-
ditionalisation is correct. And since Bayesian philosophy of science is a thriving research
program, this is a non-trivial fact.) On the other hand, there are remarkably few direct, and
plausible, arguments in favor of conditionalisation. In the absence of a direct argument we
can say two things.
First, the fact that a lot of philosophers (and statisticians and economists etc) accept
conditionalisation, and have derived many important results using it, is a reason to take it
seriously. The research programs that are based around conditionalisation do not seem to be
degenerating, or failing to produce new insights. Second, in a lot of everyday applications,
conditionalisation seems to yield sensible results. The simplest cases here are cases involving
card games or roulette wheels where we can specify the probabilities of various outcomes in
advance.
Let’s work through a very simple example to see this. A deck of cards has 52 cards, of
which 13 are hearts. Imagine we’re about to draw 2 cards, without replacement, from that
deck, which has been well-shuffled. The probability that the first is a heart is 13/52, or, more
simply, 1/4. If we assume that a heart has been taken out, e.g. if we draw a heart with the first
card, the probability that we’ll draw another heart if 12/51. That is, conditional on the first
card we draw being a heart, the probability that the second is a heart if 12/51.
Now imagine that we do actually draw the first card, and it’s a heart. What should the
probability be that the next card will be a heart? It seems like it should be 12/51. Indeed, it is

42
hard to see what else it could be. If A is The first card drawn is a heart and B is The second
card drawn is a heart, then it seems both Pr(A|B) and PrB (A) should be 12/51. And examples
like this could be multiplied endlessly.
The support here for conditionalisation is not just that we ended up with the same result.
It’s that we seem to be making the same calculations both times. In cases like this, when
we’re trying to figure out Pr(A|B), we pretend we’re trying to work out PrB (A), and then stop
pretending when we’ve worked out the calculation. If that’s always the right way to work out
Pr(A|B), then Pr(A|B) should always turn out to be equal to PrB (A). Now this argument goes
by fairly quickly obviously, and we might want to look over more details before deriving very
heavy duty results from the idea that updating is always by conditionalisation, but it’s easy
to see we might take conditionalisation to be a plausible model for updating probabilities.

43
 Chapter 8

About Conditional Probability

8.1 Conglomerability
Here is a feature that we’d like an updating rule to have. If getting some evidence E will make
a hypothesis H more probable, then not getting E will not also make H more probable.
Indeed, in standard cases, not getting evidence that would have made H more probable
should make H less probable. It would be very surprising if we could know, before running
a test, that however it turns out some hypothesis H will be more probable at the end of the
test than at the beginning of it. We might have to qualify this in odd cases where H is, e.g.,
that the test is completed. But in standard cases if H will be likely whether some evidence
comes in or doesn’t come in, then H should be already likely.
We’ll say that an update rule is conglomerable if it has this feature, and non-conglomerable
otherwise. That is, it is non-conglomerable iff there are H and E such that,

PrE (H) > Pr(H)andPr¬E (H) > Pr(H)

Now a happy result for conditionalisation, the rule that says PE (H) = Pr(H|E), is that it
is conglomerable. This result is worth going over in some detail. Assume that Pr(H|E) >
Pr(H)andPr¬E (H) > Pr(H). Then we can derive a contradicton as follows

Pr(H) = Pr((H ∧ E) ∨ (H ∧ ¬E)) since H = (H ∧ E) ∨ (H ∧ ¬E)

= Pr(H ∧ E) + Pr(H ∧ ¬E) since (H ∧ E) and (H ∧ ¬E) are disjoint
= Pr(H|E)Pr(E) + Pr(H|¬E)Pr(¬E) since Pr(H|E)Pr(E) = Pr(H ∧ E)
> Pr(H)Pr(E) + Pr(H)Pr(¬E) since by assumption Pr(H|E) > Pr(H) and Pr(H|¬E) > Pr(H)
= Pr(H)(Pr(E) + Pr(¬E))
= Pr(H)Pr(E ∨ ¬E) since E and ¬E are disjoint
= Pr(H) since Pr(E ∨ ¬E) = 1

Conglomerability is related to dominance. The dominance rule of decision making says

(among other things) that if C1 is preferable to C2 given E, and C1 is preferable to C2 given
¬E, then C1 is simply preferable to C2 . Conglomerability says (among other things) that if
Pr(H) is greater than x given E, and it is greater than x given ¬E, then it is simply greater
than x.

44
 Contemporary decision theory makes deep and essential use of principles of this form,
i.e. that if something holds given E, and given ¬E, then it simply holds. And one of the run-
ning themes of these notes will be sorting out just which such principles hold, and which do
not hold. The above proof shows that we get one nice result relating conditional probability
and simple probability which we can rely on.

8.2 Independence
The probability of some propositions depends on other propositions. The probability that
I’ll be happy on Monday morning is not independent of whether I win the lottery on the
weekend. On the other hand, the probability that I win the lottery on the weekend is in-
dependent of whether it rains in Seattle next weekend. Formally, we define probabilistic
indepdendence as follows.

• Propositions A and B are independent iff Pr(A|B) = Pr(A).

There is something odd about this definition. We purported to define a relationship that
holds between pairs of propositions. It looked like it should be a symmetric relation: A is
independent from B iff B is independent from A. But the definition looks asymmetric: A
and B play very different roles on the right-hand side of the definition. Happily, this is just
an appearance. Assuming that A and B both have positive probability, we can show that
Pr(A|B) = Pr(A) is equivalent to Pr(B|A) = Pr(B).

Pr(A|B) = Pr(A)
Pr(A ∧ B)
⇔ = Pr(A)
Pr(B)
⇔ Pr(A ∧ B) = Pr(A) × Pr(B)
Pr(A ∧ B)
⇔ = Pr(B)
Pr(A)
⇔ Pr(B|A) = Pr(B)

We’ve multiplied and divided by Pr(A) and Pr(B), so these equivalences don’t hold if Pr(A)
or Pr(B) is 0. But in other cases, it turns out that Pr(A|B) = Pr(A) is equivalent to Pr(B|A) =
Pr(B). And each of these is equivalent to the claim that Pr(A ∧ B) = Pr(A)Pr(B). This is an
important result, and one that we’ll refer to a bit.

• For independent propositions, the probability of their conjunction is the product of

their probabilities.
• That is, if A and B are independent, then Pr(A ∧ B) = Pr(A)Pr(B)

This rule doesn’t apply in cases where A and B are dependent. To take an extreme case,
when A is equivalent to B, then A ∧ B is equivalent to A. In that case, Pr(A ∧ B) = Pr(A),
not Pr(A)2 . So we have to be careful applying this multiplication rule. But it is a powerful
rule in those cases where it works.

45
8.3 Kinds of Independence
The formula Pr(A|B) = Pr(A) is, by definition, what probabilistic independence amounts
to. It’s important to note that probabilistic dependence is very different from causal depen-
dence, and so we’ll spend a bit of time going over the differences.
The phrase ‘causal dependence’ is a little ambiguous, but one natural way to use it is that
A causally depends on B just in case B causes A. If we use it that way, it is an asymmetric
relation. If B causes A, then A doesn’t cause B. But probabilistic dependence is symmetric.
That’s what we proved in the previous section.
Indeed, there will typically be a quite strong probabilistic dependence between effects
and their causes. So not only is the probability that I’ll be happy on Monday dependent on
whether I win the lottery, the probability that I’ll win the lottery is dependent on whether
I’ll be happy on Monday. It isn’t causally dependent; my moods don’t cause lottery results.
But the probability of my winning (or, perhaps better, having won) is higher conditional on
my being happy on Monday than on my not being happy.
One other frequent way in which we get probabilistic dependence without causal de-
pendence is when we have common effects of a cause. So imagine that Fred and I jointly
purchased some lottery tickets. If one of those tickets wins, that will cause each of us to be
happy. So if I’m happy, that is some evidence that I won the lottery, which is some evidence
that Fred is happy. So there is a probabilistic connection between my being happy and Fred’s
being happy. This point is easier to appreciate if we work through an example numerically.
Make each of the following assumptions.

• We have a 10% chance of winning the lottery, and hence a 90% chance of losing.
• If we win, it is certain that we’ll be happy. The probability of either of us not being
happy after winning is 0.
• If we lose, the probability that we’ll be unhappy is 0.5.
• Moreover, if we lose, our happiness is completely independent of one another, so con-
ditional on losing, the proposition that I’m happy is independent of the proposition
that Fred’s happy

So conditional on losing, each of the four possible outcomes have the same probability.
Since these probabilities have to sum to 0.9, they’re each equal to 0.225. So we can list the
possible outcomes in a table. In this table A is winning the lottery, B is my being happy and
C is Fred’s being happy.

A B C Pr
T T T 0.1
T T F 0
T F T 0
T F F 0
F T T 0.225
F T F 0.225
F F T 0.225
F F F 0.225

46
Adding up the various rows tells us that each of the following are true.
• Pr(B) = 0.1 + 0.225 + 0.225 = 0.55
• Pr(C) = 0.1 + 0.225 + 0.225 = 0.55
• Pr(B ∧ C) = 0.1 + 0.225 = 0.325

From that it follows that Pr(B|C) = 0.325/0.55 ≈ 0.59. So Pr(B|C) > Pr(B). So B and C
are not independent. Conditionalising on C raises the probability of B because it raises the
probability of one of the possible causes of C, and that cause is also a possible cause of B.
Often we know a lot more about probabilistic dependence than we know about causal
connections and we have work to do to figure out the causal connections. It’s very hard,
especially in for example public health settings, to figure out what is a cause-effect pair, and
what is the result of a common cause. One of the most important research programs in
modern statistics is developing methods for solving just this problem. The details of those
methods won’t concern us here, but we’ll just note that there’s a big gap between probabilistic
dependence and causal dependence.
On the other hand, it is usually safe to infer probabilistic dependence from causal depen-
dence. If E is one of the (possible) causes of H, then usually E will change the probabilities
of H. We can perhaps dimly imagine exceptions to this rule.
So imagine that a quarterback is trying to decide whether to run or pass on the final
play of a football game. He decides to pass, and the pass is successful, and his team wins.
Now as it happens, had he decided to run, the team would have had just as good a chance
of winning, since their run game was exactly as likely to score as their pass game. It’s not
crazy to think in those circumstances that the decision to pass was among the causes of the
win, but the win was probabilistically independent of the decision to pass. In general we can
imagine cases where some event moves a process down one of two possible paths to success,
and where the other path had just as good a chance of success. (Imagine a doctor deciding
to operate in a certain way, a politician campaigning in one area rather than another, a storm
moving a battle from one piece of land to another, or any number of such cases.) In these
cases we might have causal dependence (though whether we do is a contentious issue in the
metaphysics of causation) without probabilistic dependence.
But such cases are rare at best. It is a completely commonplace occurrence to have prob-
abilistic dependence without clear lines of causal dependence. We have to have very deli-
cately balanced states of the world in order to have causal dependence without probabilistic
dependence, and in every day cases we can safely assume that such a situation is impossible
without probabilistic connections.

8.4 Gamblers’ Fallacy

If some events are independent, then the probability of one is independent of the proba-
bility of the others. So knowing the results of one event gives you no guidance, not even
probabilistic guidance, into whether the other will happen.
These points may seem completely banal, but in fact they are very hard to fully incor-
porate into our daily lives. In particular, they are very hard to completely incorporate in
cases where we are dealing with successive outcomes of a particular chance process, such as
a dice roll or a coin flip. In those cases we know that the individual events are independent

47
of one another. But it’s very hard not to think that, after a long run of heads say, that the
coin landing tails is ‘due’.
This feeling is what is known as the Gamblers’ Fallacy. It is the fallacy of thinking that,
when events A and B are independent, that what happens in A can be a guide of some kind
to event B.
One way of noting how hard a grip the Gamblers’ Fallacy has over our thoughts is to try
to simulate a random device such as a coin flip. As an exercise, imagine that you’re writing
down the results of a series of 100 coin flips. Don’t actually flip the coin, just write down a
sequence of 100 Hs (for Heads) and Ts (for Tails) that look like what you think a random
series of coin flips will look like. I suspect that it won’t look a lot like what an actual sequence
does look like, in part because it is hard to avoid the Gamblers’ Fallacy.
Occasionally people will talk about the Inverse Gamblers’ Fallacy, but this is a much less
clear notion. The worry would be someone inferring from the fact that the coin has landed
heads a lot that it will probably land heads next time. Now sometimes, if we know that it
is a fair coin for example, this will be just as fallacious as the Gamblers’ Fallacy itself. But it
isn’t always a fallacy. Sometimes the fact that the coin lands heads a few times in a row is
evidence that it isn’t really a fair coin.
It’s important to remember the gap between causal and probabilistic dependence here.
In normal coin-tossing situations, it is a mistake to think that the earlier throws have a causal
impact on the later throws. But there are many ways in which we can have probabilistic
dependence without causal dependence. And in cases where the coin has been landing
heads a suspiciously large number of times, it might be reasonable to think that there is a
common cause of it landing heads in the past and in the future - namely that it’s a biased
coin! And when there’s a common cause of two causally independent events, they may be
probabilistically dependent. That’s to say, the first event might change the probabilities of
the second event. In those cases, it doesn’t seem fallacious to think that various patterns will
continue.
This does all depend on just how plausible it is that there is such a causal mechanism.
It’s one thing to think, because the coin has landed heads ten times in a row, that it might
be biased. There are many causal mechanisms that could explain that. It’s another thing
to think, because the coin has alternated heads and tails for the last ten tosses that it will
continue to do so in the future. It’s very hard, in normal circumstances, to see what could
explain that. And thinking that patterns for which there’s no natural causal explanation will
continue is probably a mistake.

48
 Chapter 9

Expected Utility
9.1 Expected Values
A random variable is simply a variable that takes different numerical values in different
states. In other words, it is a function from possibilities to numbers. Typically, random
variables are denoted by capital letters. So we might have a random variable X whose value
is the age of the next President of the United States, and his or her inauguration. Or we
might have a random variable that is the number of children you will have in your lifetime.
Basically any mapping from possibilities to numbers can be a random variable.
It will be easier to work with a specific example, so let’s imagine the following case.
You’ve asked each of your friends who will win the big football game this weekend, and 9
said the home team will win, while 5 said the away team will win. (Let’s assume draws are
impossible to make the equations easier.) Then we can let X be a random variable measuring
the number of your friends who correctly predicted the result of the game. The value X takes
is {
9, if the home team wins,
X=
5, if the away team wins.
Given a random variable X and a probability function Pr, we can work out the expected
value of that random variable with respect to that probability function. Intuitively, the ex-
pected value of X is a weighted average of the possible values of X, where the weights are
given by the probability (according to Pr) of each value coming about. More formally, we
work out the expected value of X this way. For each case, we multiply the value of X in that
case by the probability of the case obtaining. Then we sum the numbers we’ve got, and the
result is the expected value of X. We’ll write the expected value of X as Exp(X). So if the
probability that the home wins is 0.8, and the probability that the away team wins is 0.2,
then

Exp(X) = 9 × 0.8 + 5 × 0.2

= 7.2 + 1
= 8.2

There are a couple of things to note about this result. First, the expected value of X isn’t in
any sense the value that we expect X to take. Indeed, the expected value of X is not even
a value that X could take. So we shouldn’t think that “expected value” is a phrase we can

49
understand by simply understanding the notion of expectation and of value. Rather, we
should think of the expected value as a kind of average.
Indeed, thinking of the expected value as an average lets us relate it back to the common
notion of expectation. If you repeated the situation here – where there’s an 0.8 chance that
9 of your friends will be correct, and an 0.2 chance that 5 of your friends will be correct
– very often, then you would expect that in the long run the number of friends who were
correct on each occasion would average about 8.2. That is, the expected value of a random
variable X is what you’d expect the average value of X to be if (perhaps per impossible) the
underlying situation was repeated many many times.

9.2 Maximise Expected Utility Rule

The orthodox view in modern decision theory is that the right decision is the one that max-
imises the expected utility of your choice. Let’s work through a few examples to see how this
might work. Consider again the decision about whether to take a cheap airline or a more
reliable airline, where the cheap airline is cheaper, but it performs badly in bad weather. In
cases where the probability is that the plane won’t run into difficulties, and you have much
to gain by taking the cheaper ticket, and even if something goes wrong it won’t go badly
wrong, it seems that you should take the cheaper plane. Let’s set up that situation in a table.

Good weather Bad weather

Pr = 0.8 Pr = 0.2
Cheap Airline 10 0
Reliable Airline 6 5

We can work out the expected utility of each action fairly easily.
Exp(Cheap Airline) = 0.8 × 10 + 0.2 × 0
=8+0
=8
Exp(Reliable Airline) = 0.8 × 6 + 0.2 × 5
= 4.8 + 1
= 5.8
So the cheap airline has an expected utility of 8, the reliable airline has an expected utility
of 5.8. The cheap airline has a higher expected utility, so it is what you should take.
We’ll now look at three changes to the example. Each change should intuitively change
the correct decision, and we’ll see that the maximise expected utility rule does change in
each case. First, change the downside of getting the cheap airline so it is now more of a risk
to take it.

Good weather Bad weather

Pr = 0.8 Pr = 0.2
Cheap Airline 10 -20
Reliable Airline 6 5

50
Here are the new expected utility considerations.

Exp(Cheap Airline) = 0.8 × 10 + 0.2 × –20

= 8 + (–4)
=4
Exp(Reliable Airline) = 0.8 × 6 + 0.2 × 5
= 4.8 + 1
= 5.8

Now the expected utility of catching the reliable airline is higher than the expected utility of
catching the cheap airline. So it is better to catch the reliable airline.
Alternatively, we could lower the price of the reliable airline, so it is closer to the cheap
airline, even if it isn’t quite as cheap.

Good weather Bad weather

Pr = 0.8 Pr = 0.2
Cheap Airline 10 0
Reliable Airline 9 8

Here are the revised expected utility considerations.

Exp(Cheap Airline) = 0.8 × 10 + 0.2 × 0

=8+0
=8
Exp(Reliable Airline) = 0.8 × 9 + 0.2 × 8
= 7.2 + 1.6
= 8.8

And again this is enough to make the reliable airline the better choice.
Finally, we can go back to the original utility tables and simply increase the probability
of bad weather.

Good weather Bad weather

Pr = 0.3 Pr = 0.7
Cheap Airline 10 0
Reliable Airline 6 5

51
We can work out the expected utility of each action fairly easily.

Exp(Cheap Airline) = 0.3 × 10 + 0.7 × 0

=3+0
=3
Exp(Reliable Airline) = 0.3 × 6 + 0.7 × 5
= 1.8 + 3.5
= 5.3

We’ve looked at four versions of the same case. In each case the ordering of the outcomes,
from best to worst, was:
1. Cheap airline and good weather
2. Reliable airline and good weather
3. Reliable airline and bad weather
4. Cheap airline and bad weather
As we originally set up the case, the cheap airline was the better choice. But there were three
ways to change this. First, we increased the possible loss from taking the cheap airline. (That
is, we increased the gap between the third and fourth options.) Second, we decreased the
gain from taking the cheap airline. (That is, we decreased the gap between the first and
second options.) Finally, we increased the risk of things going wrong, i.e. we increased the
probability of the bad weather state. Any of these on their own was sufficient to change the
recommendation that “Maximise Expected Utility” makes. And that’s all to the good, since
any of these things does seem like it should be sufficient to change what’s best to do.

9.3 Structural Features

When using the “Maximise Expected Utility” rule we assign a number to each choice, and
then pick the option with the highest number. Moreover, the number we assign is inde-
pendent of the other options that are available. The number we assign to a choice depends
on the utility of that choice in each state and the probability of the states. Any decision rule
that works this way is guaranteed to have a number of interesting properties.
First, it is guaranteed to be transitive. That is, if it recommends A over B, and B over
C, then it recommends A over C. To see this, let’s write the expected utility of a choice A
as Exp(U(A)). If A is chosen over B, then Exp(U(A)) > Exp(U(B)). And if B is chosen over
C, then Exp(U(B)) > Exp(U(C)). Now >, defined over numbers, is transitive. That is, if
Exp(U(A)) > Exp(U(B)) and Exp(U(B)) > Exp(U(C)), then Exp(U(A)) > Exp(U(C)). So
the rule will recommend A over B.
Second, it satisfies the independence of irrelevant alternatives. Assume A is chosen over
B and C. That is, Exp(U(A)) > Exp(U(B)) and Exp(U(A)) > Exp(U(C)). Then A will be
chosen when the only options are A and B, since Exp(U(A)) > Exp(U(B)). And A will
be chosen when the only options are A and C, since Exp(U(A)) > Exp(U(C)). These two
features are intuitively pleasing features of a decision rule.
Numbers are totally ordered by >. That is, for any two numbers x and y, either x > y
or y > x or x = y. So if each choice is associated with a number, a similar relation holds

52
among choices. That is, either A is preferable to B, or B is preferable to A, or they are equally
preferable.
Expected utility maximisation never recommends choosing dominated options. As-
sume that A dominates B. For each state Si , write utility of A in Si as U(A|Si ). Then domi-
nance means that for all i, U(A|Si ) > U(B|Si ). Now Exp(U(A)) and Exp(U(B)) are given by
the following formulae. (In what follows n is the number of possible states.)

Exp(A) = Pr(S1 )U(A|S1 ) + Pr(S2 )U(A|S2 ) + ... + Pr(Sn )U(A|Sn )

Exp(B) = Pr(S1 )U(B|S1 ) + Pr(S2 )U(B|S2 ) + ... + Pr(Sn )U(B|Sn )

Note that the two values are each the sum of n terms. Note also that, given dominance,
each term on the top row is at least as great as than the term immediately below it on the
second row. (This follows from the fact that U(A|Si ) > U(B|Si ) and the fact that Pr(Si ) ≥ 0.)
Moreover, at least one of the terms on the top row is greater than the term immediately
below it. (This follows from the fact that U(A|Si ) > U(B|Si ) and the fact that for at least one
i, Pr(Si ) > 0. That in turn has to be true because if Pr(Si ) = 0 for each i, then Pr(S1 ∨ S2 ∨
... ∨ Sn ) = 0. But S1 ∨ S2 ∨ ... ∨ Sn has to be true.) So Exp(A) has to be greater than Exp(B).
So if A dominates B, it has a higher expected utility.

53
 Chapter 10

Sure Thing Principle

10.1 Generalising Dominance
The maximise expected utility rule also supports a more general version of dominance. We’ll
state the version of dominance using an example, then spend some time going over how we
know maximise expected utility satisfies that version.
The original dominance principle said that if A is better than B in every state, then A is
simply better than B simply. But we don’t have to just compare choices in individual states,
we can also compare them across any number of states. So imagine that we have to choose
between A and B and we know that one of four states obtains. The utility of each choice in
each state is given as follows.

S1 S2 S3 S4
A 10 9 9 0
B 8 3 3 3

And imagine we’re using the maximin rule. Then the rule says that A does better than B in
S1 , while B does better than A in S4 . The rule also says that B does better than A overall,
since it’s worst case scenario is 3, while A’s worst case scenario is 0. But we can also compare
A and B with respect to pairs of states. So conditional on us just being in S1 or S2 , then A is
better. Because between those two states, its worst case is 9, while B’s worst case is 3.
Now imagine we’ve given up on maximin, and are applying a new rule we’ll call maxi-
average. The maxiaverage rule tells us make the choice that has the highest (or maximum)
average of best case and worst case scenarios. The rule says that B is better overall, since it
has a best case of 8 and a worst case of 3 for an average of 5.5, while A has a best case of 10
and a worst case of 0, for an average of 5.
But if we just know we’re in S1 or S2 , then the rule recommends A over B. That’s because
among those two states, A has a maximum of 10 and a minimum of 9, for an average of 9.5,
while B has a maximum of 8 and a minimum of 3 for an average of 5.5.
And if we just know we’re in S3 or S4 , then the rule also recommends A over B. That’s
because among those two states, A has a maximum of 9 and a minimum of 0, for an average
of 4.5, while B has a maximum of 3 and a minimum of 3 for an average of 3.
This is a fairly odd result. We know that either we’re in one of S1 or S2 , or that we’re in
one of S3 or S4 . And the rule tells us that if we find out which, i.e. if we find out we’re in S1

54
or S2 , or we find out we’re in S3 or S4 , either way we should choose A. But before we find
this out, we should choose B.
Here then is a more general version of dominance. Assume our initial states are {S1 , S2 , ..., Sn }.
Call this set S. A binary partition of S is a pair of sets of states, call them T1 and T2 , such
that every state in S is in exactly one of T1 and T2 . (We’re simplifying a little here - generally
a partition is any way of dividing a collection up into parts such that every member of the
original collection is in one of the ‘parts’. But we’ll only be interested in cases where we di-
vide the original states in two, i.e., into a binary partition.) Then the generalised version of
dominance says that if A is better than B among the states in T1 , and it is better than B among
the states in T2 , where T1 and T2 provide a partition of S, then it is better than B among the
states in S. That’s the principle that maxiaverage violates. A is better than B among the states
{S1 , S2 }. And it is better than B among the states {S3 , S4 }. But it isn’t better than B among
the states {S1 , S2 , S3 , S4 }. That is, it isn’t better than B among the states generally.
We’ll be interested in this principle of dominance because, unlike perhaps dominance
itself, there are some cases where it leads to slightly counterintuitive results. For this reason
some theorists have been interested in theories which, although they satisfy dominance, do
not satisfy this general version of dominance.
On the other hand, maximise expected utility does respect this principle. In fact, it
respects an even stronger principle, one that we’ll state using the notion of conditional
expected utility. Recall that as well as probabilities, we defined conditional probabilities
above. Well conditional expected utilities are just the expectations of the utility function
with respect to a conditional probability. More formally, if there are states S1 , S2 , ..., Sn , then
the expected utility of A conditional on E, which we’ll write Exp(U(A|E), is

Exp(U(A|E)) = Pr(S1 |E)U(S1 |A) + Pr(S2 |E)U(S2 |A) + ... + Pr(Sn |E)U(Sn |A)

That is, we just replace the probabilities in the definition of expected utility with conditional
probabilities. (You might wonder why we didn’t also replace the utilities with conditional
utilities. That’s because we’re assuming that states are defined so that given an action, the
state has a fixed utility. If we didn’t make this simplifying assumption, we’d have to be more
careful here.) Now we can prove the following theorem.

• If Exp(U(A|E)) > Exp(U(B|E)), and Exp(U(B|¬E)) > Exp(U(B|¬E)), then Exp(U(A)) >
Exp(U(B)).

We’ll prove this by proving something else that will be useful in many contexts.

• Exp(U(A)) = Exp(U(A|E))Pr(E) + Exp(U(A|¬E))Pr(¬E)

To see this, note the following

Pr(Si ) = Pr((Si ∧ E) ∨ (Si ∧ ¬E))

= Pr(Si ∧ E) + Pr(Si ∧ ¬E)
= Pr(Si |E)Pr(E) + Pr(Si |¬E)Pr(¬E)

55
And now we’ll use this when we’re expanding Exp(U(A|E))Pr(E).

Exp(U(A|E))Pr(E) = Pr(E)[Pr(S1 |E)U(S1 |A) + Pr(S2 |E)U(S2 |A) + ... + Pr(Sn |E)U(Sn |A)]
= Pr(E)Pr(S1 |E)U(S1 |A) + Pr(E)Pr(S2 |E)U(S2 |A) + ... + Pr(E)Pr(Sn |E)U(Sn |A)
Exp(U(A|¬E))Pr(¬E) = Pr(¬E)[Pr(S1 |¬E)U(S1 |A) + Pr(S2 |¬E)U(S2 |A) + ... + Pr(Sn |¬E)U(Sn |A)]
= Pr(¬E)Pr(S1 |¬E)U(S1 |A) + Pr(¬E)Pr(S2 ¬|E)U(S2 |A) + ... + Pr(¬E)Pr(Sn |¬E)U(Sn |A)

Putting those two together, we get

Exp(U(A|E))Pr(E) + Exp(U(A|¬E))Pr(¬E)
= Pr(E)Pr(S1 |E)U(S1 |A) + ... + Pr(E)Pr(Sn |E)U(Sn |A)+
Pr(¬E)Pr(S1 |¬E)U(S1 |A) + ... + Pr(¬E)Pr(Sn |¬E)U(Sn |A)
= (Pr(E)Pr(S1 |E) + Pr(¬E)Pr(S1 |¬E))U(S1 |A) + ... + (Pr(E)Pr(Sn |E) + Pr(¬E)Pr(Sn |¬E))U(Sn |A)
= Pr(S1 )U(S1 |A) + Pr(S2 )U(S2 |A) + ...Pr(Sn )U(Sn |A)
= Exp(U(A))

Now if Exp(U(A|E)) > Exp(U(B|E)), and Exp(U(B|¬E)) > Exp(U(B|¬E)), then the follow-
ing two inequalities hold.

Exp(U(A|E))Pr(E) ≥ Exp(U(B|E))Pr(E)
Exp(U(A|¬E))Pr(¬E) ≥ Exp(U(B|¬E))Pr(¬E)

In each case we have equality only if the probability in question (Pr(E) in the first line,
Pr(¬E) in the second) is zero. Since not both Pr(E) and Pr(¬E) are zero, one of those is a
strict inequality. (That is, the left hand side is greater than, not merely greater than or equal
to, the right hand side.) So adding up the two lines, and using the fact that in one case we
have a strict inequality, we get

Exp(U(A|E))Pr(E) + Exp(U(A|¬E))Pr(¬E) ≥ Exp(U(B|E))Pr(E) + Exp(U(B|¬E))Pr(¬E)

i.e. Exp(U(A)) > Exp(U(B))

That is, if A is better than B conditional on E, and it is better than B conditional on ¬E, then
it is simply better than B.

10.2 Sure Thing Principle

The result we just proved is very similar to a famous principle of decision theory, the Sure
Thing Principle. The Sure Thing Principle is usually stated in terms of one option being at
least as good as another, rather than one option being better than another, as follows.

Sure Thing Principle If AE ⪰ BE and A¬E ⪰ B¬E, then A ⪰ B.

The terminology there could use some spelling out. By A ≻ B we mean that A is preferred
to B. By A ⪰ B we mean that A is regarded as at least as good as B. The relation between ≻
and ⪰ is like the relation between > and ≥. In each case the line at the bottom means that
we’re allowing equality between the values on either side.

56
 The odd thing here is using AE ⪰ BE rather than something that’s explicitly conditional.
We should read the terms on each side of the inequality sign as conjunctions. It means that A
and E is regarded as at least as good an outcome as B and E. But that sounds like something
that’s true just in case the agent prefers A to B conditional on E obtaining. So we can use
preferences over conjunctions like AE as proxy for conditional preferences.
So we can read the Sure Thing Principle as saying that if A is at least as good as B con-
ditional on E, and conditional on ¬E, then it really is at least as good as B. Again, this looks
fairly plausible in the abstract, though we’ll soon see some reasons to worry about it.
Expected Utility maximisation satisfies the Sure Thing Principle. I won’t go over the
proof here because it’s really just the same as the proof from the previous section with >
replaced by ≥ in a lot of places. But if we regard the Sure Thing Principle as a plausible
principle of decision making, then it is a good feature of Expected Utility maximisation that
it satisfies it.
It is tempting to think of the Sure Thing Principle as a generalisation of a principle of
logical implication we all learned in propositional logic. The principle in question said that
from X → Z, and Y → Z, and X ∨ Y, we can infer C. If we let Z be that A is better than
B, let X be E, and Y be ¬E, it looks like we have all the premises, and the reasoning looks
intuitively right. But this analogy is misleading for two reasons.
First, for technical reasons we can’t get into in depth here, preferring A to B conditional
on E isn’t the same as it being true that if E is true you prefer A to B. To see some problems
with this, think about cases where you don’t know E is true, and A is something quite hor-
rible that mitigates the effects of the unpleasant E. In this case you do prefer AE to BE, and
E is true, but you don’t prefer A to B. But we’ll set this question, which is largely a logical
question about the nature of conditionals, to one side.
The bigger problem is that the analogy with logic would suggest that the following gen-
eralisation of the Sure Thing Principle will hold.
Disjunction Principle If AE1 ⪰ BE1 and AE2 ⪰ BE2 , and Pr(E1 ∨ E2 ) = 1 then A ⪰ B.

But this “Disjunction Principle” seems no good in cases like the following. I’m going to toss
two coins. Let p be the proposition that they will land differently, i.e. one heads and one
tails. I offer you a bet that pays you $2 if p, and costs you $3 if ¬p. This looks like a bad
bet, since Pr(p) = 0.5, and losing $3 is worse than gaining $2. But consider the following
argument.
Let E1 be that at least one of the coins landing heads. It isn’t too hard to show that
Pr(p|E1 ) = 2/3. So conditional on E1 , the expected return of the bet is 2/3 × 2 – 1/3 × 3 =
4/3 – 1 = 1/3. That’s a positive return. So if we let A be taking the bet, and B be declining the
bet, then conditional on E1 , A is better than B, because the expected return is positive.
Let E2 be that at least one of the coins landing tails. It isn’t too hard to show that
Pr(p|E1 ) = 2/3. So conditional on E2 , the expected return of the bet is 2/3 × 2 – 1/3 × 3 =
4/3 – 1 = 1/3. That’s a positive return. So if we let A be taking the bet, and B be declining the
bet, then conditional on E2 , A is better than B, because the expected return is positive.
Now if E1 fails, then both of the coins lands tails. That means that at least one of the
coins lands tails. That means that E2 is true. So if E1 fails E2 is true. So one of E1 and E2

57
has to be true, i.e. Pr(E1 ∨ E2 ) = 1. And AE1 ⪰ BE1 and AE2 ⪰ BE2 . Indeed AE1 ≻ BE1
and AE2 ≻ BE2 . But B ≻ A. So the disjunction principle isn’t in general true.
It’s a deep philosophical question how seriously we should worry about this. If the Sure
Thing Principle isn’t any more plausible intuitively than the Disjunction Principle, and the
Disjunction Principle seems false, does that mean we should be sceptical of the Sure Thing
Principle? As I said, that’s a very hard question, and it’s one we’ll return to a few times in
what follows.

10.3 Allais Paradox

The Sure Thing Principle is one of the more controversial principles in decision theory be-
cause there seem to be cases where it gives the wrong answer. The most famous of these is
the Allais paradox, first discovered by the French economist (and Nobel Laureate) Maurice
Allais. In this paradox, the subject is first offered the following choice between A and B.
The results of their choice will depend on the drawing of a coloured ball from an urn. The
urn contains 10 white balls, 1 yellow ball, and 89 black balls, and assume the balls are all
randomly distributed so the probability of drawing each is identical.

White Yellow Black

A $1,000,000 $1,000,000 $0
B $5,000,000 $0 $0

That is, they are offered a choice between an 11% shot at $1,000,000, and a 10% shot at
$5,000,000. Second, the subjects are offered the following choice between C and D, which
are dependent on drawings from a similarly constructed urn.

White Yellow Black

C $1,000,000 $1,000,000 $1,000,000
D $5,000,000 $0 $1,000,000

That is, they are offered a choice between $1,000,000 for sure, and a complex bet that gives
them a 10% shot at $5,000,000, an 89% shot at $1,000,000, and a 1% chance of striking out
and getting nothing.
Now if we were trying to maximise expected dollars, then we’d have to choose both B
and D. But, and this is an important point that we’ll come back to, dollars aren’t utilities.
Getting $2,000,000 isn’t twice as good as getting $1,000,000. Pretty clearly if you were offered
a million dollars or a 50% chance at two million dollars you would, and should, take the
million for sure. That’s because the two million isn’t twice as useful to you as the million.
Without a way of figuring out the utility of $1,000,000 versus the utility of $5,000,000, we
can’t say whether A is better than B. But we can say one thing. You can’t consistently hold
the following three views.

• B≻A
• C≻D
• The Sure Thing Principle holds

58
This is relevant because a lot of people think B ≻ A and C ≻ D. Let’s work through the
proof of this to finish with.
Let E be that either a white or yellow ball is drawn. So ¬E is that a black ball is drawn.
Now note that A¬E is identical to B¬E. In either case you get nothing. So A¬E ⪰ B¬E. So
if AE ⪰ BE then, by Sure Thing, A ⪰ B. Equivalently, if B ≻ A, then BE ≻ AE. Since we’ve
assumed B ≻ A, then BE ≻ AE.
Also note that C¬E is identical to D¬E. In either case you get a million dollars. So
D¬E ⪰ C¬E. So if DE ⪰ CE then, by Sure Thing, D ⪰ C. Equivalently, if C ≻ D, then
CE ≻ DE. Since we’ve assumed C ≻ D, then CE ≻ DE.
But now we have a problem, since BE = DE, and AE = CE. Given E, then choice between
A and B just is the choice between C and D. So holding simultaneously that BE ≻ AE and
CE ≻ DE is incoherent.
It’s hard to say for sure just what’s going on here. Part of what’s going on is that we have
a ‘certainty premium’. We prefer options like C that guarantee a positive result. Now having
a certainly good result is a kind of holistic property of C. The Sure Thing Principle in effect
rules out assigning value to holistic properties like that. The value of the whole need not be
identical to the value of the parts, but any comparisons between the values of the parts has
to be reflected in the value of the whole. Some theorists have thought that a lesson of the
Allais paradox is that this is a mistake.
We won’t be looking in this course at theories which violate the Sure Thing Principle,
but we will be looking at justifications of the Sure Thing Principle, so it is worth thinking
about reasons you might have for rejecting it.

59
10.4 Exercises
10.4.1 Calculate Expected Utilities
In the following example Pr(S1 ) = 0.4, Pr(S2 ) = 0.3, Pr(S3 ) = 0.2 and Pr(S4 ) = 0.1. The table
gives the utility of each of the possible actions (A, B, C, D and E) in each state. What is the
expected utility of each action?

S1 S2 S3 S4
A 0 2 10 2
B 6 2 1 7
C 1 8 9 7
D 3 1 8 6
E 4 7 1 4

10.4.2 Conditional Choices

In the previous example, C is the best thing to do conditional on S2 . It has expected utility 8
in that case, and all the others are lower. It is also the best thing to do conditional on S2 ∨ S3 .
It has expected utility 8.4 if we conditionalise on S2 ∨ S3 , and again all the others are lower.
For each of the actions A, B, C, D and E, find a proposition such that conditional on that
proposition, the action in question has the highest expected utility.

10.4.3 Generalised Dominance

Does the maximax decision rule satisfy the generalised dominance principle we discussed
in the text? That principle says that if the initial range of states is S, and T1 and T2 form a
partition of S, and if A is a better choice than B conditional on being in T1 , and A is also
a better choice than B conditional on being in T2 , then A is simply a better choice than B.
Does this principle hold for the maximax decision rule?

10.4.4 Sure Thing Principle

Assume we’re using the ‘Maximise Expected Utility’ rule. And assume that B is not the best
choice out of our available choices conditional on E. Assume also that B is not the best choice
out of our available choices conditional on ¬E. Does it follow that B is not the best available
choice? If so, provide an argument that this is the case. If not, provide a counterexample,
i.e. a case where B is not the best choice conditional on E, not the best choice conditional
on ¬E, but the best choice overall.

60
 Chapter 11

Understanding Probability
11.1 Kinds of Probability
As might be clear from the discussion of what probability functions are, there are a lot of
probability functions. For instance, the following is a probability function for any (logically
independent) p and q.

p q Pr
T T 0.97
T F 0.01
F T 0.01
F F 0.01

But if p actually is that the moon is made of green cheese, and q is that there are little green
men on Mars, you probably won’t want to use this probability function in decision making.
That would commit you to making some bets that are intuitively quite crazy.
So we have to put some constraints on the kinds of probability we use if the “Maximise
Expected Utility” rule is likely to make sense. As it is sometimes put, we need to have an in-
terpretation of the Pr in the expected utility rule. We’ll look at three possible interpretations
that might be used.

11.2 Frequency
Historically probabilities were often identified with frequencies. If we say that the probabil-
ity that this F is a G is, say, 23 , that means that the proportion of F’s that are G’s is 23 .
Such an approach is plausible in a lot of cases. If we want to know what the probability
is that a particular student will catch influenza this winter, a good first step would be to find
1
out the proportion of students who will catch influenza this winter. Let’s say this is 10 . Then,
to a first approximation, if we need to feed into our expected utility calculator the probability
1
that this student will catch influenza this winter, using 10 is not a bad first step. Indeed, the
insurance industry does not a bad job using frequencies as guides to probabilities in just this
way.
But that can hardly be the end of the story. If we know that this particular student has
not had an influenza shot, and that their boyfriend and their roommate have both caught
influenza, then the probability of them catching influenza would now be much higher. With

61
that new information, you wouldn’t want to take a bet that paid $1 if they didn’t catch in-
fluenza, but lost you $8 if they did catch influenza. The odds now look like that’s a bad
bet.
Perhaps the thing to say is that the relevant group is not all students. Perhaps the relevant
group is students who haven’t had influenza shots and whose roommates and boyfriends
have also caught influenza. And if, say, 23 of such students have caught influenza, then per-
haps the probability that this student will catch influenza is 32 .
You might be able to see where this story is going by now. We can always imagine more
details that will make that number look inappropriate as well. Perhaps the student in ques-
tion is spending most of the winter doing field work in South America, so they have little
chance to catch influenza from their infected friends. And now the probability should be
lower. Or perhaps we can imagine that they have a genetic predisposition to catch influenza,
so the probability should be higher. There is always more information that could be relevant.
The problem for using frequencies as probabilities then is that there could always be
more precise information that is relevant to the probability. Every time we find that the
person in question isn’t merely an F (a student, say), but is a particular kind of F (a student
who hasn’t had an influenza shot, whose close contacts are infected, who has a genetic pre-
disposition to influenza), we want to know the proportion not of F’s who are G’s, but the
proportion of the more narrowly defined class who are G’s. But eventually this will leave us
with no useful probabilities at all, because we’ll have found a way of describing the student
in question such that they are the only person in history who satisfies this description.
This is hardly a merely theoretical concern. If we are interested in the probability that a
particular bank will go bankrupt, or that a particular Presidential candidate will win elec-
tion, it isn’t too hard to come up with a list of characteristics of the bank or candidate in
question in such a way that they are the only one in history to meet that description. So the
frequency that such banks will go bankrupt is either 1 (1 out of 1 go bankrupt) or 0 (0 out
of 1 do). But those aren’t particularly useful probabilities. So we should look elsewhere for
an interpretation of the Pr that goes into our definition of expected utility.
In the literature there are two objections to using frequencies as probabilities that seem
related to the argument we’re looking at here.
One of these is the Reference Class Problem. This is the problem that if we’re interested
in the probability that a particular person is G, then the frequency of G-hood amongst the
different classes the person is in might differ.
The other is the Single Case Problem. This is the problem that we’re often interested in
one-off events, like bank failures, elections, wars etc, that don’t naturally fit into any natural
broader category.
I think the reflections here support the idea that these are two sides of a serious problem
for the view that probabilities are frequencies. In general, there actually is a natural solution
to the Reference Class Problem. We look to the most narrowly drawn reference class we
have available. So if we’re interested in whether a particular person will survive for 30 years,
and we know they are a 52 year old man who smokes, we want to look not to the survival
frequencies of people in general, or men in general, or 52 year old men in general, but 52
year old male smokers.
Perhaps by looking at cases like this, we can convince ourselves that there is a natural
solution to the Reference Class Problem. But the solution makes the Single Case Problem

62
come about. Pretty much anything that we care about is distinct in some way or another.
That’s to say, if we look closely we’ll find that the most natural reference class for it just
contains that one thing. That’s to say, it’s a single case in some respect. And one-off events
don’t have interesting frequencies. So frequencies aren’t what we should be looking to as
probabilities.

11.3 Degrees of Belief

In response to these worries, a lot of philosophers and statisticians started thinking of prob-
ability in purely subjective terms. The probability of a proposition p is just how confident
the agent is that p will obtain. This level of confidence is the agent’s degree of belief that p
will obtain.
Now it isn’t altogether to measure degrees of belief. I might be fairly confident that my
baseball team will win tonight, and more confident that they’ll win at least one of the next
three games, and less confident that they’ll win all of their next three games, but how could
we measure numerically each of those strengths. Remember that probabilities are numbers.
So if we’re going to identify probabilities with degrees of belief, we have to have a way to
convert strengths of confidence to numbers.
The core idea about how to do this uses the very decision theory that we’re looking for
input to. I’ll run through a rough version of how the measurement works; we’ll be refining
this quite a bit as the course goes on. Imagine you have a chance to buy a ticket that pays
$1 if p is true. How much, in dollars, is the most would you pay for this? Well, it seems that
how much you should pay for this is the probability of p. Let’s see why this is true. (Assume
in what follows that the utility of each action is given by how many dollars you get from the
action; this is the simplifying assumption we’re making.) If you pay $Pr(p) for the ticket,
then you’ve performed some action (call it A) that has the following payout structure.
{
1 – Pr(p) if p,
U(A) =
–Pr(p) if ¬p.
So the expected value of U(A) is

Exp(U(A)) = Pr(p)U(Ap) + Pr(¬p)U(A¬p)

= Pr(p)(1 – Pr(p)) + Pr(¬p)U(A¬p)
= Pr(p)(1 – Pr(p)) + (1 – Pr(p))(–Pr(p))
= Pr(p)(1 – Pr(p)) – (1 – Pr(p))(Pr(p))
=0

So if you pay $Pr(p) for the bet, your expected return is exactly 0. Obviously if you pay
more, you’re worse off, and if you pay less, you’re better off. $Pr(p) is the break even point,
so that’s the fair price for the bet.
And that’s how we measure degrees of belief. We look at the agent’s ‘fair price’ for a bet
that returns $1 if p. (Alternatively, we look at the maximum they’ll pay for such a bet.) And
that’s they’re degree of belief that p. If we’re taking probabilities to be degrees of belief, if we

63
are (as it is sometimes put) interpreting probability subjectively, then that’s the probability
of p.
This might look suspiciously circular. The expected utility rule was meant to give us
guidance as to how we should make decisions. But the rule needed a probability as an input.
And now we’re taking that probability to not only be a subjective state of the agent, but a
subjective state that is revealed in virtue of the agent’s own decisions. Something seems odd
here.
Perhaps we can make it look even odder. Let p be some proposition that might be true
and might be false, and assume that the agent’s choice is to take or decline a bet on p that has
some chance of winning and some chance of losing. Then if the agent takes the bet, that’s
a sign that their degree of belief in p was higher than the odds of the bet on p, so therefore
they are increasing their expected utility by taking the bet, so they are doing the right thing.
On the other hand, if they decline the bet, that’s a sign that their degree of belief in p was
lower than the odds of the bet on p, so therefore they are increasing their expected utility
by taking the bet, so they are doing the right thing. So either way, they do the right thing.
But a rule that says they did the right thing whatever they do isn’t much of a rule.
There are two important responses to this, which are related to one another. The first
is that although the rule does (more or less) put no restrictions at all on what you do when
faced with a single choice, it can put quite firm constraints on your sets of choices when
you have to make multiple decisions. The second is that the rule should be thought of as a
procedural rather than substantive rule of rationality. We’ll look at these more closely.
If we take probabilities to be subjective probabilities, i.e. degrees of belief, then the max-
imise expected utility rule turns out to be something like a consistency constraint. Compare
it to a rule like Have Consistent Beliefs. As long as we’re talking about logically contingent
matters, this doesn’t put any constraint at all on what you do when faced with a single ques-
tion of whether to believe p or ¬p. But it does put constraints on what further beliefs you
can have once you believe p. For instance, you can’t now believe ¬p.
The maximise expected utility rule is like this. Indeed we already saw this in the Allais
paradox. The rule, far from being empty, rules out the pair of choices that many people
intuitively think is best. So if the objection is that the rule has no teeth, that objection can’t
hold up.
We can see this too in simpler cases. Let’s say I offer the agent a ticket that pays $1 if p,
and she pays 60c for it. So her degree of belief in p must be at least 0.6. Then I offer her a
ticket that pays $1 if ¬p, and she pays 60c for it too. So her degree of belief in ¬p must be at
least 0.6. But, and here’s the constraint, we think degrees of belief have to be probabilities.
And if Pr(p) > 0.6, then Pr(¬p) < 0.4. So if Pr(¬p) > 0.6, we have an inconsistency. That’s
bad, and it’s the kind of badness it is the job of the theory to rule out.
One way to think about the expected utility rule is to compare it to norms of means-
end rationality. At times when we’re thinking about what someone should do, we really
focus on what the best means is to their preferred end. So we might say If you want to go to
Harlem, you should take the A train, without it even being a relevant question whether they
should, in the circumstances, want to go to Harlem.
The point being made here is quite striking when we consider people with manifestly
crazy beliefs. If we’re just focussing on means to an end, then we might look at someone
who, say, wants to crawl from the southern tip of Broadway to its northern tip. And we’ll

64
say “You should get some kneepads so you don’t scrape your knees, and you should take lots
of water, and you should catch the 1 train down to near to where Broadway starts, etc.” But
if we’re not just offering procedural advice, but are taking a more substantive look at their
position, we’ll say “You should come up with a better idea about what to do, because that’s
an absolutely crazy thing to want.”
As we’ll see, the combination of the maximise expected utility rule with the use of de-
grees of belief as probabilities leads to a similar set of judgments. On the one hand, it is a
very good guide to procedural questions. But it leaves some substantive questions worry-
ingly unanswered. Next time we’ll come back to this distinction, and see if there’s a better
way to think about probability.

65
 Chapter 12

Objective Probabilities
12.1 Credences and Norms
We ended last time with looking at the idea that the probabilities in expected utility calcu-
lations should be subjective. As it is sometimes put, they should be degrees of belief. Or, as
it is also sometimes put, they should be credences. We noted that under this interpretation,
the maximise expected utility rule doesn’t put any constraints on certain simple decisions.
That’s because we use the rule to calculate what credences are, and then use the very same
credences to say what the rule requires. But the rule isn’t useless. It puts constraints, often
sharp constraints, on sets of decisions. In this respect it is more like the rule Have Consistent
Beliefs than like the rule Believe What’s True, or Believe What Your Evidence Supports. And
we compared it to procedural, as opposed to substantive norms.
What’s left from all that are two large questions.

• Do we get the right procedural/consistency constraints from the expected utility rule?
In particular (a) should credences be probabilities, and (b) should we make complex
decisions by the expected utility rule? We’ll look a bit in what follows at each of these
questions.
• Is a purely procedural constraint all we’re looking for in a decision theory?

And intuitively the answer to the second question is No. Let’s consider a particular case.
Alex is very confident that the Kansas City Royals will win baseball’s World Series next year.
In fact, Alex’s credence in this is 0.9, very close to 1. Unfortunately, there is little reason for
this confidence. Kansas City has been one of the worst teams in baseball for many years,
the players they have next year will be largely the same as the players they had when doing
poorly this year, and many other teams have players who have performed much much better.
Even if Kansas City were a good team, there are 30 teams in baseball, and relatively random
events play a big role in baseball, making it unwise to be too confident that any one team
will win.
Now, Alex is offered a bet that leads to a $1 win if Kansas City win the World Series, and
a $1 loss if they do not. The expected return of that bet, given Alex’s credences, is +80c. So
should Alex make the bet?
Intuitively, Alex should not. It’s true that given Alex’s credences, the bet is a good one.
But it’s also true that Alex has crazy credences. Given more sensible credences, the bet has
a negative expected return. So Alex should not make the bet.

66
 It’s worth stepping away from probabilities, expected values and the like to think about
this in a simpler context. Imagine a person has some crazy beliefs about what is an effective
way to get some good end. And assume they, quite properly, want that good end. In fact,
however, acting on their crazy beliefs will be counterproductive; it will just make things
worse for everyone. And their evidence supports this. Should they act on their beliefs?
Intuitively not. To be sure, if they didn’t act on their beliefs, there would be some inconsis-
tency between their beliefs and their actions. But inconsistency isn’t the worst thing in the
world. They should, instead, have different beliefs.
Similarly Alex should have different credences in the case in question. The question,
what should Alex do given these credences, seems less interesting than the question, what
should Alex do? And that’s what we’ll look at.

12.2 Evidential Probability

We get a better sense of what an agent should do if we look not to what credences they have,
but to what credences they should have. Let’s try to formalise this as the credences they
would have if they were perfectly rational.
Remember credences are still being measured by betting behaviour, but now it is betting
behaviour under the assumption of perfect rationality. So the probability of p is the highest
price the agent would pay for a bet that pays $1 if p, if they were perfectly rational. The thing
that should be done then is the thing that has the highest expected utility, relative to this
probability function. In the simple case where the choice is between taking and declining a
bet, this becomes a relatively boring theory - you should take the bet if you would take the
bet if you were perfectly rational. In the case of more complicated decisions, it becomes a
much more substantive theory. (We’ll see examples of this in later weeks.)
But actually we’ve said enough to give us two philosophical puzzles.
The first concerns whether there determinately is a thing that you would do if you were
perfectly rational. Consider a case where you have quite a bit of evidence for and against
p. Different rational people will evaluate the evidence in different ways. Some people will
evaluate p as being more likely than not, and so take a bet at 50/50 odds on p. Others will
consider the evidence against p to be stronger, and hence decline a bet at 50/50 odds. It
seems possible that both sides in such a dispute could be perfectly rational.
The danger here is that if we define rational credences as the credences a perfectly ratio-
nal person would have, we might not have a precise definition. There may be many different
credences that a perfectly rational person would have. That’s bad news for a purported def-
inition of rational credence.
The other concerns cases where p is about your own rationality. Let’s say p is the propo-
sition that you are perfectly rational. Then if you were perfectly rational, your credence in
this would probably be quite high. But that’s not the rational credence for you to have right
now in p. You should be highly confident that you, like every other human being on the
planet, are susceptible to all kinds of failures of rationality. So it seems like a mistake in
general to set your credences to what they would be were you perfectly rational.
What seems better in general is to proportion your credences to the evidence. The ra-
tional credences are the ones that best reflect the evidence you have in favour of various
propositions. The idea here to to generate what’s usually called an evidential probability.

67
The probability of each proposition is a measure of how strongly it is supported by the evi-
dence.
That’s different from what a rational person would believe in two respects. For one thing,
there is a fact about how strongly the evidence supports p, even if different people might
disagree about just how strongly that is. For another thing, it isn’t true that the evidence
supports that you are perfectly rational, even though you would believe that if you were
perfectly rational. So the two objections we just mentioned are not an issue here.
From now on then, when we talk about probability in the context of expected utility, we’ll
talk about evidential probabilities. There’s an issue, one we’ll return to later, about whether
we can numerically measure strengths of evidence. That is, there’s an issue about whether
strengths of evidence are the right kind of thing to be put on a numerical scale. Even if they
are, there’s a tricky issue about how we can even guess what they are. I’m going to cheat a
little here. Despite the arguments above that evidential probabilities can’t be identified with
betting odds of perfectly rational agents, I’m going to assume that, unless we have reason to
the contrary, those betting odds will be our first approximation. So when we have to guess
what the evidential probability of p is, we’ll start with what odds a perfectly rational agent
(with your evidence) would look for before betting on p.

12.3 Objective Chances

There is another kind of probability that theorists are often interested in, one that plays a
particularly important role in modern physics. Classical physics was, or at least was thought
to be, deterministic. Once the setup of the universe at a time t was set, the laws of nature
determined what would happen after t. Modern physics is not deterministic. The laws don’t
determine, say, how long it will take for an unstable particle to decay. Rather, all the laws
say is that the particle has such-and-such a chance of decaying in a certain time period. You
might have heard references to the half-life of different radioactive particles; this is the time
in which the particle has a 21 probabiilty of decaying.
What are these probabilities that the scientists are talking about? Let’s call them ‘chances’
to give them a name. So the question is, what is the status of chances. We know chances
aren’t evidential probabilities. We know this for three reasons.
One is that it is a tricky empirical question whether any event has any chance other than
0 or 1. It is now something of a scientific consensus that some events are indeed chancy.
But this relies on some careful scientific investigation. It isn’t something we can tell from
our armchairs. But we can tell from just thinking about decisions under uncertainty that
the evidential probability of some outcomes is between 0 and 1.
Another is that, as chances are often conceived, events taking place in the past do not,
right now, have chances other than 0 or 1. There might have been, at a point in the past,
some intermediate chance of a particle decaying. But if we’re now asking about whether a
particle did decay or not in the last hour, then either it did decay, and its chance is 0, or it
did not decay, and its chance is 1. (I should note that not everyone thinks about chances
in quite this way, but it is a common way to think about them.) There are many events
that took place in the past, however, whose evidential probability is between 0 and 1. For
instance, if we’re trying to meet up a friend, and hence trying to figure out where the friend
might have gone to, we’ll think about, and assign evidential probabilities to, various paths

68
the friend might have taken in the past. These thoughts won’t be thoughts about chances in
the physicists’ sense; they’ll be about evidential probabilities.
Finally, chances are objective. The evidential probability that p is true might be different
for me than for you. For instance, the evidence she has might make it quite likely for the
juror that the suspect is guilty, even if he is not. But the evidence the suspect has makes it
extremely likely that he is innocent. Evidential probabilities differ between different people.
Chances do not. Someone might not know what the chance of a particular outcome is, but
what they are ignorant of is a matter of objective fact.
The upshot seems to be that chances are quite different things from evidential probabil-
ities, and the best thing to do is simply to take them to be distinct basic concepts.

12.4 The Principal Principle and Direct Inference

Although chances and evidential probabilities are distinct, it seems they stand in some close
relation. If a trustworthy physicist tells you that a particle has an 0.8 chance of decaying
in the next hour, then it seems your credences should be brought into line with what the
physicists say. This idea has been dubbed the Principal Principle, because it is the main
principle linking chances and credences. If we use Pr for evidential probabilities, and Ch
for objective chances in the physicists’ sense, then the idea behind the principle is this.

Principal Principle Pr(p|Ch(p) = x) = x

That is, the probability of p, conditional on the chance of p being x, is x.

The Principal Principle may need to be qualified. If your evidence also includes that p,
then even if the chance of p is 0.8, perhaps your credence in p should be 1. After all, p is
literally evident to you. But perhaps it is impossible for p to be part of your evidence while
its chance is less than 1. The examples given in the literature of how this could come about
are literally spectacular. Perhaps God tells you that p is true. Or perhaps a fortune teller
with a crystal ball sees that it is true. Or something equally bizarre happens. Any suggested
exceptions to the principle have been really outlandish. So whether the principle is true for
all possible people in all possible worlds, it seems to hold for us around here.
Chances, as the physicists think of them, are not frequencies. It might be possible to
compute the theoretical chance of a rare kind of particle not decaying over the course of
an hour, even though the particle is so rare, and so unstable, that no such particle has ever
survived an hour. In that case the frequency of survival (i.e. the proportion of all such
particles that do actually survive an hour) is 0, but physical theory might tell us that the
chance is greater than 0. Nevertheless chances are like frequencies in some respects.
One such respect is that chances are objective. Just as the chance of a particle decay is an
objective fact, one that we might or might not be aware of, the frequency of particle decay
is also an objective fact that we might or might not be aware of. Neither of these facts are in
any way relative to the evidence of a particular agent, the way that evidential probabilities
are.
And just like chances, frequencies might seem to put a constraint on credences. Con-
sider a case where the only thing you know about a is that it is G. And you know that the
frequency of F-hood among Gs is x. For instance, let a be a person you’ve never met, G be
the property of being a 74 year old male smoker, and F the property of surviving 10 more

69
years. Then you might imagine knowing the survival statistics, but knowing nothing else
about the person. In that case, it’s very tempting to think the probability that a is F is x. In
our example, we’d be identifying the probability of this person surviving with the frequency
of survival among people of the same type.
This inference from frequencies to probabilities is sometimes called “Direct Inference”. It
is, at least on the surface, a lot like the Principal Principle. But it is a fair bit more contentious.
We’ll say a bit more about this once we’ve looked about probabilities of events with infinite
possibility spaces. But for now just note that it is really rather rare that all we know about
an individual can be summed up in one statistic like this. Even if the direct inference can
be philosophically justified (and I’m a little unsure that it can be) it will rarely be applicable.
So it is less important than the Principal Principle.
We’ll often invoke the Principal Principle tacitly in setting up problems. That is, when I
want to set up a problem where the probabilities of the various outcomes are given, I’ll often
use objective chances to fix the probabilities of various states. We’ll use the direct inference
more sparingly, because it isn’t as clearly useful.

70
 Chapter 13

Understanding Utility
13.1 Utility and Welfare
So far we’ve frequently talked about the utility of various outcomes. What we haven’t said a
lot about is just what it is that we’re measuring when we measure the utility of an outcomes.
The intuitive idea is that utility is a measure of welfare - having outcomes with higher utility
if a matter of having a higher level of welfare. But this doesn’t necessarily move the idea
forward, because we’d like to know a bit more about what it is to have more welfare. There
are a number of ways we can frame the same question. We can talk about ‘well-being’ instead
of welfare, or we can talk about having a good life instead, or having a life that goes well. But
the underlying philosophical question, what makes it the case that a life has these features,
remains more or less the same.
There are three primary kinds of theories of welfare in contemporary philosophy. These
are
• Experience Based theories
• Objective List theories
• Preference Based theories

In decision theory, and indeed in economics, people usually focus on preference based the-
ories. Indeed, the term ‘utility’ is sometimes used in such way that A has more utility than
B just means that the agent prefers A to B. Indeed, I’ve sometimes earlier moved back and
forth previously between saying A has higher utility and saying A is preferred. And the
focus here (and in the next set of notes) will be on why people have moved to preference
based accounts, and technical challenges within those accounts. But we’ll start with the
non-preference based accounts of welfare.

13.2 Experiences and Welfare

One tradition, tracing back at least to Jeremy Bentham, is to identify welfare with having
good experiences. A person’s welfare is high if they have lots of pleasures, and few pains.
More generally, a person’s welfare is high if they have good experiences.
Of course it is possible that a person might be increasing their welfare by having bad
experiences at any one time. They might be at work earning the money they need to finance
activities that lead to good experiences later, or they might just be looking for money to
stave off bad experiences (starvation, lack of shelter) later. Or perhaps the bad experiences,

71
such as in strenuous exercise, are needed in order to be capable of later doing the things,
e.g. engaging in sporting activities, that produce good experiences. Either way, the point
has to be that a person’s welfare is not simply measured by what their experiences are like
right now, but by what their experiences have been, are, and will be over the course of their
lives.
There is one well known objection to any such account - what Robert Nozick called the
“experience machine”. Imagine that a person is, in their sleep, kidnapped and wired up to
a machine that produces in their brain the experiences as of a fairly good life. The person
still seems to be having good days filled with enjoyable experiences. And they aren’t merely
raw pleasurable sensations - the person is having experiences as of having rich fulfilling
relationships with the friends and family they have known and loved for years. But in fact
the person is not in any contact with those people, and for all the friends and family know,
the person was kidnapped and killed. This continues for decades, until the person has a
peaceful death at an advanced age.
Has this person had a good life or a bad life? Many people think intuitively that they
have had a bad life. Their entire world has been based on an illusion. They haven’t really
had fulfilling relationships, travelled to exciting places, and so on. Instead they have been
systematically deceived about the world. But on an experience based view of welfare, they
have had all of the goods you could want in life. Their experiences are just the experiences
that a person having a good life would have. So the experience based theorist is forced to
say that they have had a good life, and this seems mistaken.
Many philosophers find this a compelling objection to the experience based view of
welfare. But many people are not persuaded. So it’s worth thinking a little through some
other puzzles for purely experience based views of welfare.
It’s easy enough to think about paradigmatic pains, or bad experiences. It isn’t too hard
to come up with paradigmatic good experiences, though perhaps there would be more dis-
agreement about what experiences are paradigms of the good than are paradigms of the bad.
But many experiences are less easy to classify. Even simple experiences like tickles might be
good experiences for some, and bad experiences for others.
When we get to more complicated experiences, things are even more awkward for the
experience based theorist. Some people like listening to heavily distorted music, or watching
horror movies, or drinking pineapple schnapps. Other people, indeed most people, do not
enjoy these things. The experience theory has a couple of choices here. Either we can say
that one group is wrong, and these things either do, or do not, raise one’s welfare. But this
seems implausible for all experiences. Perhaps at the fringes there are experiences people
seek that nevertheless decrease their welfare, but it seems strange to argue that the same
experiences are good for everyone.
The other option is to say that there are really two experiences going on when you, say,
listen to a kind of music that some, but not all, people like. There is a ‘first-order’ experience
of hearing the music. And there is a ‘second-order’ experience, an experience of enjoying
the experience of hearing the music. Perhaps this is right in some cases. (Perhaps for horror
movies, fans both feel horrified and have a pleasant reaction to being horrified, at least some
of the time.) But it seems wrong in general. If there is a food that I like and you dislike, that
won’t usually be because I’ll have a positive second-order experience, and you won’t have
such a thing. Intuitively, the experience of, say, drinking a good beer, isn’t like that, because

72
it just isn’t that complicated. Rather, I just have a certain kind of experience, and I like it,
and you, perhaps, do not.
A similar problem arises when considering the choices people make about how to dis-
tribute pleasures over their lifetime. Some people are prepared to undergo quite unpleasant
experiences, e.g. working in painful conditions, in exchange for pleasant experiences later
(e.g. early retirement, higher pay, shorter hours). Other people are not. Perhaps in some
cases people are making a bad choice, and their welfare would be higher if they made dif-
ferent trade-offs. But this doesn’t seem to be universally true - it just isn’t clear that there’s
such a thing as the universally correct answer to how to trade off current unpleasantness for
future pleasantness.
Note that this intertemporal trade-off question actually conceals two distinct questions
we have to answer. One is how much we want to ‘discount’ the future. Economists think,
with some empirical support, that people mentally discount future goods. People value a
dollar now more than they value a dollar ten years hence, or even an inflation adjusted
dollar ten years hence. The same is true for experiences: people value good experiences
now more than good experiences in the future. But it isn’t clear how much discount, if any,
is consistent with maximising welfare. The other question is how much we value high ‘peaks’
of experience versus avoiding low ‘troughs’. Some people are prepared to put up with the
bad to get the good, others are not. And the worry for the experience based theorist is that
neither need be making a mistake. Perhaps what is best for a person isn’t just a function of
their experiences over time, but on how much they value the kind of experiences that they
get.
So we’ve ended up with three major kinds of objections to experience based accounts of
welfare.
• The experience machine does not increase our welfare
• Different people get welfare from different experiences
• Different people get different amounts of welfare from the same sequences of expe-
riences over time, even if they agree about the welfare of each of the moment-to-
moment experiences.

These seem like enough reasons to move to other theories of welfare.

13.3 Objective List Theories

One response to these problems with experience based accounts is to move to a theory based
around desire satisfaction. Since that’s the theory that’s most commonly used in decision
theory, we’ll look at it last. Before that, we’ll look briefly at so called objective list theories
of welfare. These theories hold that there isn’t necessarily any one thing that makes your
life better. Welfare isn’t all about good experiences, or about having preferences that are
satisfied. Rather, there are many ways in which your welfare can be improved. The list of
things that make your life better may include:
• Knowledge
• Engaging in rational activity
• Good health, adequate shelter, and more generally good physical well-being
• Being in loving relationships, and in sustained friendships

73
 • Being virtuous
• Experiencing beauty
• Desiring the things that make life better, i.e. the things on this list
Some objective list theorists hold that the things that should go on the list do have something
in common, but this isn’t an essential part of the theory.
The main attraction of the objective list approach is negative. We’ve already seen some of
the problems with experience based theories of welfare. We’ll see later some of the problems
with desire based theories. A natural response to this is to think that welfare is heteroege-
nous, and that no simple theory of welfare can capture all that makes human lives go well.
That’s the response of the objective list theorist.
The first thing to note about these theories is that the lists in question always seem open
to considerable debate. If there was a clearer principle about what’s going on the lists and
what is not, this would not be such a big deal. But in the absence of a clear (or easy to apply)
principle, there is a sense of arbitrariness about the process.
Indeed, the lists that are produced by Western academics seem notably aligned with the
desires and values of Western academics. It’s notable that the lists produced tend to give
very little role to the family, to religion, to community and to tradition. Of course all these
things can come in indirectly. If being in loving relationships is a good, and families promote
loving relationships, then families are an indirect good. And the same thing can be said
religion, and community, and traditional practices. But still, many people might hold those
things to be valuable in their own right, not just because of the goods that they produce. Or
they might hold some things on the canonical lists, such as education and knowledge to be
instrumental goods, rather than making them primary goods as philosophers often do.
This can’t be an objection to objective list theories of welfare as such. Nothing in the
theory rules out extending the list to include families, or traditions, in the mix, for instance.
(Indeed, these kinds of goods are included in some versions of the theory.) But it is perhaps
revealing that the lists hew so closely to the Western academic’s idea of the good life. (Indeed
the list I’ve got here is more universal than several proposed lists, since I’ve included health
and shelter, which is left off some.) It might well be thought that there isn’t one list of goods
that make life good for any person in any community at any time. There might well be a list
of what makes for a good life in a community like ours, and maybe even lists like the one
above capture it, but claims to universality should be treated sceptically.
A more complicated question is how to generate comparative welfare judgments from
the list. Utilities are meant to be represented numerically, so we need to be able to say which
of two outcomes is better, or that the outcomes are exactly as good as one another. (Perhaps
we need something more, some way of saying how much better one life is than another. But
we’ll set that question aside for now.) We already saw one hard aspect of this question above
- how do we turn facts about the welfare of a person at different times of their life into an
overall welfare judgment? That question is just as hard for the objective list theorist as for
the experience theorist. (And again, part of why it is so hard is that it is far from clear that
there is a unique correct answer.)
But the objective list theorist has a challenge that the experience theorist does not have:
how do we weigh up the various goods involved? Let’s think about a very simple list - say
the only things on the list are friendship and beauty. Now in some cases, saying which of

74
two outcomes is better will be easy. If outcome A will produce improve your friendship, and
let you experience beautiful things, more than outcome B will, then A is better than B. But
not all choices are like that. What if you are faced with a choice between seeing a beautiful
art exhibit, that is closing today, or keeping a promise to meet your friend for lunch? Which
choice will maximise your welfare? The art gallery will do better from a beauty standpoint,
while the lunch will do better from a friendship standpoint. We need to know something
more to know how this tradeoff will be made.
There are actually three related objections here. One is that the theory is incomplete
unless there is some way to weigh up the various things on the list, and the list itself does
not produce the means to do the weighting. A second is that it isn’t obvious that there is
a unique way to weigh up the things on the list. Perhaps one person is made better off by
focussing on friendship and the expense of beauty, and for another person it goes the other
way. So perhaps there is no natural weighing consistent with the spirit behind the objective
list theories that works in all contexts. Finally, it isn’t obvious that there is a fact of the matter
in many cases, leaving us with many choices where there is no fact of the matter about which
will produce more utility. But that will be a problem for creating a numerical measure of
value that can be plugged into expected utility calculations.
Let’s sum up. There are really two core worries about objective list theories. These are:
• Different things are good for different people
• There’s no natural way to produce a utility measure out of the goodness of each ‘com-
ponent’ of welfare

Next time we’ll look at desire based theories of utility, which are the standard in decision
theory and in economics.

75
 Chapter 14

Subjective Utility
14.1 Preference Based Theories
So far we’ve looked at two big theories of the nature of preferences. Both of them have
thought that in some sense people don’t get a say in what’s good for them. There is an
impersonal fact about what is best for a person, and that is good for you whether you like it
or not. The experience theory says that it is the having of good experiences, and the objective
list theory says that it includes a larger number of features. Preference-based, or ‘subjective’
theories of welfare start with the idea that what’s good for different people might be radically
different. It also takes the idea that people often are the best judge of what’s best for them
very seriously.
What we end up with is the theory that A is better for an agent than B if and only if
the agent prefers A to B. We’ll look at some complications to this, but for now we’ll work
with the simple picture that welfare is a matter of preference satisfaction. This theory has a
number of advantages over the more objective theories.
First, it easily deals with the idea that different things might be good for different people.
That’s accommodated by the simple fact that people have very different desires, so different
things increase their welfare.
Second, it also deals easily with the issues about comparing bundles of goods, either
bundles of different goods, or bundles of goods at different times. An agent need not only
have preferences about whether they, for instance, prefer time with their family to material
possessions. They also have more fine-grained preferences about various trade offs between
different goods, and trade offs about sequences of goods across time. So if one person has a
strong preference for getting goods now, and another person is prepared to wait for greater
goods later, the theory can accommodate that difference. Or if one person is prepared to
put up with unpleasant events in order to have greater goods at other times, the theory
can accommodate that, as well as the person who prefers a more steady life. If they are
both doing what they want, then even though they are doing different things, they are both
maximising their welfare.
But there are several serious problems concerning this approach to welfare. We’ll start
with the intuitive idea that people sometimes don’t know what is good for them.
We probably all can think about things in everyday life where we, or a friend of ours,
has done things that quite clearly are not in their own best interests. In many such cases,
it won’t be that the person is doing what they don’t want to do. Indeed, part of the reason
that people acting against their own best interests is such a problem is that the actions in

76
question are ones they very much want to perform. Or so we might think antecedently. If a
person’s interests are just measured by their desires, then it is impossible to want what’s bad
for you. That seems very odd.
It is particularly odd when you think about the effect of advertising and other forms of
persuasion. The point of advertising is to change your preferences, and presumably it works
frequently enough to be worth spending a lot of money on. But it is hard to believe that
the effect of advertising is to change how good for you various products are. Yet if your
welfare is measured by how many of your desires are satisfied, then anything that changes
your desires changes what is good for you.
Note that sometimes we even have internalised the fact that we desire the wrong things.
Sometimes we desire something, while desiring that we don’t desire it. So we can say things
like “I wish I didn’t want to smoke so much”. In that case it seems that what would, on a
strict subjective standpoint, have our best outcome be smoking and wanting not to smoke,
since then both our ‘first-order’ desire to smoke and our ‘second-order’ desire not to want
to smoke would be satisfied. But that sounds crazy.
Perhaps the best thing to do here would be to modify the subjective theory of welfare.
Perhaps we could say that our welfare is maximised by the satisfaction of those desires we
wish we had. Or perhaps we could say that it is maximised by the satisfaction of our ‘un-
defeated’ desires, i.e. desires that we don’t wish we didn’t have. There are various options
here for keeping the spirit of a subjective approach to welfare, while allowing that people
sometimes desire the bad.

14.2 Interpersonal Comparisons

I mentioned above that the subjective approach does better than the other approaches at
converting the welfare someone gets from the different parts of their life into a coherent
whole. That’s because agent’s don’t only have preferences over how the parts of their lives
go, they also have preferences over different distributions of welfare over the different parts
of their lives, and preferences over bundles of goods they may receive. The downside of
this is that a kind of comparison that the objective theory might do well at, interpersonal
comparisons, are very hard for the subjective theorist to make.
Intuitively there are cases where the welfare of a group is improved or decreased by a
change in events. But this is hard, in general, to capture on a subjective theory of welfare.
There is one kind of group comparison that we can make. If some individuals prefer A to B,
and none prefer B to A, then A is said to be a Pareto-improvement over B. (The name comes
from the Italian economist Wilfredo Pareto.) An outcome is Pareto-optimal if no outcome
is a Pareto-improvement over it.
But Pareto-improvements, and even Pareto-inefficiency, are rare. If I’m trying to decide
who to give $1000 to, then pretty much whatever choice I make will be Pareto-optimal.
Assume I give the money to x. Then any other choice will involve x not getting $1000, and
hence not preferring that outcome. So not everyone will prefer the alternative.
But intuitively, there are cases which are not Pareto-improvements which make a group
better off. Consider again the fact that the marginal utility of money is declining. That
suggests that if we took $1,000,000 from Bill Gates, and gave $10,000 each to 100 people on
the borderline of losing their houses, then we’d have increased the net welfare. It might not
be just to simply take money from Gates in this way, so many people will think it would be

77
wrong to do even if it wouldn’t increase welfare. But it would be odd to say that this didn’t
increase welfare. It might be odder still to say, as the subjective theory seems forced to say,
that there’s no way to tell whether it increased welfare, or perhaps that there is no fact of
the matter about whether it increased net welfare, because welfare comparisons only make
sense for something that has desires, e.g. an agent, not something that does not, e.g. a group.
There have been various attempts to get around this problem. Most of them start with
the idea that we can put everyone’s preferences on a scale with some fixed points. Perhaps for
each person we can say that utility of 0 is where they have none of their desires satisfied, and
utility of 1 is where they have all of their desires satisfied. The difficulty with this approach
is that it suggests that one way to become very very well off is to have few desires. The easily
satisfied do just as well as the super wealthy on such a model. So this doesn’t look like a
promising way forward.
Since we’re only looking at decisions made by a single individual here, the difficulties
that subjective theories of welfare have with interpersonal comparisons might not be the
biggest concern in the world. But it is an issue that comes up whenever we try to apply
subjective theories broadly.

14.3 Which Desires Count

There is another technical problem about using preferences as a foundation for utilities.
Sometimes I’ll choose A over B, not because A really will produce more welfare for me than
B, but because I think that A will produce more utility. In particular, if A is a gamble, then I
might take the gamble even though the actual result of A will be be worse, by anyone’s lights,
including my own, than B.
Now the subjectivist about welfare does want to use preferences over gambles in the
theory. In particular, it is important for figuring out how much an agent prefers A to B
to look at the agent’s preferences over gambles. In particular, if the agent thinks that one
gamble has a 50% chance of generating A, and a 50% chance of generating C, and the agent
is indifferent between that gamble and B, then the utility of B is exactly half-way between
A’s utility and C’s utility. That’s a very useful thing to be able to say. But it doesn’t help with
the original problem - how much do we value actual outcomes, not gambles over outcomes.
What we want is a way of separating instrumental from non-instrumental desires. Most
of our desires are, at least to some extent, instrumental. But that’s a problem for using them
in generating welfare functions. If I have an instrumental desire for A, that means I regard
A as a gamble that will, under conditions I give a high probability of obtaining, lead to some
result C that I want. What we really want to do is to specify these non-instrumental desires.
A tempting thing to say here is to look at our desires under conditions of full knowledge.
If I know that the train and the car will take equally long to get to a destination I desire, and
I still want to take the train, that’s a sign that I have a genuine preference for catching the
train. In normal circumstances, I might catch the train rather than take the car not because
I have such a preference, but because I could be stuck in arbitrarily long traffic jams when
driving, and I’d rather not take that risk.
But focussing on conditions of full knowledge won’t get us quite the results that we want.
For one thing, there are many things where full knowledge changes the relevant preferences.
Right now I might like to watch a football game, even though this is something of a gamble.
I’d rather do other things conditional on my team losing, but I’d rather watch conditional

78
on them winning. But if I knew the result of the game, I wouldn’t watch - it’s a little boring
to watch games where you know the result. The same goes of course for books, movies etc.
And if I had full knowledge I wouldn’t want to learn so much, but I do prefer learning to
not learning.
A better option is to look at desires over fully specific options. A fully specific option is
an option where, no matter how the further details are filled out, it doesn’t change how much
you’d prefer it. So if we were making choices over complete possible worlds, we’d be making
choices over fully specific options. But even less detailed options might be fully specific in
this sense. Whether it rains in an uninhabited planet on the other side of the universe on a
given day doesn’t affect how much I like the world, for instance.
The nice thing about fully specific options is that preferences for one rather than the
other can’t be just instrumental. In the fully specific options, all the possible consequences
are played out, so preferences for one rather than another must be non-instrumental. The
problem is that this is psychologically very unrealistic. We simply don’t have that fine-
grained a preference set. In some cases we have sufficient dispositions to say that we do
prefer one fully specific option to another, even if we hadn’t thought of them under those
descriptions. But it isn’t clear that this will always be the case.
To the extent that the subjective theory of welfare requires us to have preferences over
options that are more complex than we have the capacity to consider, it is something of an
idealisation. It isn’t clear that this is necessarily a bad thing, but it is worth noting that the
theory is in this sense a little unrealistic.

79
 Chapter 15

Declining Marginal Utilities

15.1 Money and Utility
In simple puzzles involving money, it is easy to think of the dollar amounts involved as
being proxy for the utility of each outcome. In a lot of cases, that’s a very misleading way of
thinking about things though. In general, a certain amount of money will be less useful to
you if you have more money. So $1000 will be more useful to a person who earns $20,000
per year than a person who earns $100,000 per year. And $1,000,000 will be more useful to
either of them than it will be to, say, Bill Gates.
This matters for decision making. It matters because it implies that in an important
sense, $2x is generally not twice as valuable to you as $x. That’s because $2x is like getting
$x, and then getting $x again. (A lot like it really!) And when we’re thinking about the utility
of the second $x, we have to think about its utility not to you, but to the person you’ll be
once you’ve already got the first $x. And that person might not value the second $x that
much.
To put this in perspective, consider having a choice between $1,000,000 for certain, and
a 50% chance at $2,000,000. Almost everyone would take the sure million. And that would
be rational, because it has a higher utility. It’s a tricky question to think about just what is
the smallest x for which you’d prefer a 50% chance at $x to $1,000,000. It might be many
many times more than a million.
The way economists’ put this is that money (like most goods) has a declining marginal
utility. The marginal utility of a good is, roughly, the utility of an extra unit of the good. For
a good like money that comes in (more or less) continuous quantities, the marginal utility
is the slope of the utility graph, as below.
You should read the x-axis there are measuring possible incomes in thousands of dollars per
1
year, and the y-axis as measuring utility. The curve there is y = x 2 . That isn’t necessarily
a plausible account of how much utility each income might give you, but it’s close enough
for our purposes. Note that although more income gives you more utility, the amount of
extra utility you get from each extra bit of income goes down as you get more income. More
precisely, the slope of the income-utility graph keeps getting shallower and shallower as
your income/utility rises. (More precisely yet, a little calculus shows that the slope of the
graph at any point is 2y1 , which is obviously always positive, but gets less and less as your
income/utility gets higher and higher.)
The fact that there is a declining marginal utility of money explains certain features of
economic life. We’ll look at models of two simple economic decisions, buying insurance

80
 320

240

160

80

0 10 20 30 40 50 60 70 80 90

Figure 15.1: Declining Marginal Utility of Money

and diversifying an investment portfolio. We’ll then use what we said about diversified in-
vestments to explain some features of the actual insurance markets that we find.

15.2 Insurance
1
Imagine the utility an agent gets from an income of x dollars is x 2 . And imagine that right
now their income is $90,000. But there is a 5% chance that something catastrophic will
happen, and their income will be just $14,400. So their expected income is 0.95 × 90, 000 +
0.05 × 14, 400 = 86220. But their expected utility is just 0.95 × 300 + 0.05 × 120 = 291, or
the utility they would have with an income of $84,861.
Now imagine this person is offered insurance against the catastrophic scenario. They can
pay, say, $4,736, and the insurance company will restore the $75,600 that they will lose if the
catastrophic event takes place. Their income is now sure to be $85,264 (after the insurance
is taken out), so they have a utility of 292. That’s higher than what their utility was, so this
is a good deal for them.
But note that it might also be a good deal for the insurance company. They receive in
premiums $4,736. And they have a 5% chance of paying out $75,600. So the expected outlay,
in dollars, for them, is $3,780. So they turn an expected profit on the deal. If they repeat
this deal often enough, the probability that they will make a profit goes very close to 1.
The point of the example is that people are trying to maximise expected utility, while
insurance companies are trying to maximise expected profits. Since there are cases where
lowering your expected income can raise your expected utility, there is a chance for a win-
win trade. And this possibility, that expected income can go down while expected utility can
go up, is explained in virtue of the fact that there is a declining marginal utility of money.

15.3 Diversification
Imagine that an agent has a starting wealth of 1, and the utility the agent gets from wealth
1
x is x 2 . (We won’t specify 2 what, but take this to be some kind of substantial unit.) The

81
agent has an opportunity to make an investment that has a 50% chance of success and a 50%
chance of failure. If the agent invests y in the scheme, the returns r will be
{
4y, if success,
r=
0, if failure.

The expected profit, in money, is y. That’s because there is a 50% chance of the profit being
3y, and a 50% chance of it being –y. But in utility, the expected return of investing 1 unit
is 0. The agent has a 50% chance of ending with a wealth of 4, i.e. a utility of 2, and a 50%
chance of ending with a wealth of 0, i.e. a utility of 0.
So making the investment doesn’t seem like a good idea. But now imagine that the agent
could, instead of putting all their money into this one venture, split the investment between
two ventures that (a) have the same probability of returns as this one, and (b) their success
of failure is probabilistically independent. So the agent invests 12 in each deal. The agent’s
return will be 

4, if both succeed,
r = 2, if one succeeds and the other fails,


0, if both fail.
The probability that both will succeed is 14 . The probability that one will succeed and the
other fail is 21 . (Exercise: why is this number greater?) The probability that both will fail is
4 . So the agent’s expected profit, in wealth, is 1. That is, it is 4 × 4 + 2 × 2 + 0 × 4 , i.e. 2,
1 1 1 1

minus the 1 that is invested, so it is 2 minus 1, i.e. 1. So it’s the same as before. Indeed, the
expected profit on each investment is 12 . And the expected profits on a pair of investments
is just the sum of the expected profits on each of the investments.
But the expected utility of the ‘portfolio’ of two investments is considerably better than
other portfolios with the same expected profit. One such portfolio is investing all of the
1 1
starting wealth in one 50/50 scheme. The expected utility of the portfolio is 4 2 × 41 + 2 2 ×
2 + 0 × 4 , which is about 1.21. So it’s a much more valuable portfolio to the agent than the
1 1

portfolio which had just a single investment. Indeed, the diversified investment is worth
making, while the single investment was not worth making.
This is the general reason why it is good to have a diversified portfolio of investments. It
isn’t because the expected profits, measured in dollars, are higher this way. Indeed, diver-
sification couldn’t possibly produce a higher expected profit. That’s because the expected
profit of a portfolio is just the sum of the expected profits of each investment in the portfo-
lio. What diversification can do is increase the expected utility of that return. Very roughly,
the way it does this is by decreasing the probability of the worst case scenarios, and of the
best case scenarios. Because the worst case scenario is more relevant to the expected utility
calculation than the best case scenario, because in general it will be further from the median
outcome, the effect is to increase the expected utility overall.
One way of seeing how important diversification is is to consider what happens if the
agent again makes two investments like this, but the two investments are probabilistically
linked. So if one investment succeeds, the other has an 80% chance of success. Now the
probability that both will succeed is 0.4, the probability that both will fail is 0.4, and the

82
probability that one will succeed and the other fail is 0.2. The expected profit of the invest-
ments is still 1. (Each investment still has an expected profit of 21 , and expected profits are
1 1
additive.) But the expected utility of the portfolio is just 4 2 × 0.4 + 2 2 × 0.2 + 0 × 0.4, which
is about 1.08. The return on investment, in utility terms, has dropped by more than half.
The lesson is that for agents with declining marginal utilities for money, a diversified
portfolio of investments can be more valuable to them than any member of the portfolio
on its own could be. But this fact turns on the investments being probabilistically separated
from one another.

15.4 Selling Insurance

In the toy example about insurance, we assumed that the marginal utility of money for the
insurance company was flat. That isn’t really true. The insurance company is owned by peo-
ple, and the utility of return to those people is diminishing as the returns get higher. There
is also the complication that the insurance company faces very different kinds of returns
when it is above and below the solvency line.
Nevertheless, the assumption that the marginal utility of money is constant for the in-
surance company is constant is a useful fiction. And the reason that it is a useful fiction is
that if the insurance company is well enough run, then the assumption is close to being true.
By ’well enough run’, I simply mean that their insurance portfolio is highly diversified.
We won’t even try to prove this here, but there are various results in probability theory
that suggest that as long as there are a lot of different, and probabilistically independent,
investments in a portfolio, then with a very high probability, the actual returns will be close
to the expected returns. In particular, if the expected returns are positive, and the portfolio
is large and diverse enough, then with a very high probability the actual returns will be
positive. So, at least in optimal cases, it isn’t a terrible simplification to treat the insurance
company as if it was sure that it would actually get its expected profits. And if that’s the case,
the changing marginal utility of money is simply indifferent.
The mathematical results that are relevant here are what are sometimes called the ”Law
of Large Numbers”. The law says that if you sample independent and identically distributed
random variables repeatedly, then for any positive number e, the probability that the average
output is within e of the expected output goes to 1 as the number of samples goes to infinity.
The approach can be quite quick in some cases. The following table lists the probability that
the number of heads on n flips of a random coin will be (strictly) between 0.4n and 0.6n for
various values of n.

Number of flips Probabiilty of between 0.4n and 0.6n heads

1 0
10 0.246
20 0.497
50 0.797
100 0.943
200 0.994
500 > 0.99

83
This depends crucially on independence. If the coin flips were all perfectly dependent, then
the probabilities would not converge at all.
Note we’ve made two large assumptions about insurance companies. One is that the
insurance company is large, the other is that it is diversified. Arguably both of these as-
sumptions are true of most real-world insurance companies. There tend to be very few
insurance companies in most economies. More importantly, those companies tend to be
fairly diversified. You can see this in a couple of different features of modern insurance
companies.
One is that they work across multiple sectors. Most car insurance companies will also
offer home insurance. Compare this to other industries. It isn’t common for car sales agents
to also be house sales agents. And it isn’t common for car builders to also be house builders.
The insurance industry tends to be special here. And that’s because it’s very attractive for the
insurance companies to have somewhat independent business wings, such as car insurance
and house insurance.
Another is that the products that are offered tend to be insurance on events that are
somewhat probabilistically independent. If I get in a car accident, this barely makes a dif-
ference to the probability that you’ll be in a car accident. So offering car insurance is an
attractive line of business. Other areas of life are a little trickier to insure. If I lose my home
to a hurricane, that does increase, perhaps substantially, the probability of you losing your
house to a hurricane. That’s because the probability of their being a hurricane, conditional
on my losing my house to a hurricane, is 1. And conditional on their being a hurricane, the
probability of you losing your house to a hurricane rises substantially. So offering hurricane
insurance isn’t as attractive a line of business as car insurance. Finally, if I lose my home to
an invading army, the probability that the same will happen to you is very high indeed. In
part for that reason, very few companies ever offer ‘invasion insurance’.
It is very hard to say with certainty at this stage whether this is true, but it seems that
a large part of the financial crisis that is now ongoing is related to a similar problem. A lot
of the financial institutions that failed were selling, either explicitly or effectively, mortgage
insurance. That is, they were insuring various banks against the possibility of default. One
problem with this is that mortgage defaults are not probabilistically independent. If I default
on my mortgage, that could be because I lost my job, or it could be because my house price
collapsed and I have no interest in sustaining my mortgage. Either way, the probability that
you will also default goes up. (It goes up dramatically if I defaulted for the second reason.)
What may have been sensible insurance policies to write on their own turned into massive
losses because the insurers underestimated the probability of having to pay out on many
policies all at once.

84
 Chapter 16

Newcomb’s Problem
16.1 The Puzzle
In front of you are two boxes, call them A and B. You call see that in box B there is $1000,
but you cannot see what is in box A. You have a choice, but not perhaps the one you were
expecting. Your first option is to take just box A, whose contents you do not know. Your
other option is to take both box A and box B, with the extra $1000.
There is, as you may have guessed, a catch. A demon has predicted whether you will
take just one box or take two boxes. The demon is very good at predicting these things –
in the past she has made many similar predictions and been right every time. If the demon
predicts that you will take both boxes, then she’s put nothing in box A. If the demon predicts
you will take just one box, she has put $1,000,000 in box A. So the table looks like this.

Predicts 1 box Predicts 2 boxes

Take 1 box $1,000,000 $0
Take 2 boxes $1,001,000 $1,000

There are interesting arguments for each of the two options here.
The argument for taking just one box is easy. The way the story has been set up, lots
of people have taken this challenge before you. Those that have taken 1 box have walked
away with a million dollars. Those that have taken both have walked away with a thousand
dollars. You’d prefer to be in the first group to being in the second group, so you should take
just one box.
The argument for taking both boxes is also easy. Either the demon has put the million
in the opaque or she hasn’t. If she has, you’re better off taking both boxes. That way you’ll
get $1,001,000 rather than $1,000,000. If she has not, you’re better off taking both boxes.
That way you’ll get $1,000 rather than $0. Either way, you’re better off taking both boxes, so
you should do that.
Both arguments seem quite strong. The problem is that they lead to incompatible con-
clusions. So which is correct?

85
16.2 Two Principles of Decision Theory
The puzzle was first introduced to philosophers by Robert Nozick. And he suggested that
the puzzle posed a challenge for the compatibility of two decision theoretic rules. These
rules are

• Never choose dominated options

• Maximise expected utility

Nozick argued that if we never chose dominated options, we would choose both boxes.
The reason for this is clear enough. If the demon has put $1,000,000 in the opaque box, then
it is better to take both boxes, since getting $1,001,000 is better than getting $1,000,000. And
if the demon put nothing in the opaque box, then your choices are $1,000 if you take both
boxes, or $0 if you take just the empty box. Either way, you’re better off taking both boxes.
This is obviously just the standard argument for taking both boxes. But note that however
plausible it is as an argument for taking both boxes, it is compelling as an argument that
taking both boxes is a dominating option.
To see why Nozick thought that maximising expected utility leads to taking one box, we
need to see how he is thinking of the expected utility formula. That formula takes as an input
the probability of each state. Nozick’s way of approaching things, which was the standard at
the time, was to take the expected utility of an action A to be given by the following sum

Exp(U(A)) = Pr(S1 |A)U(AS1 ) + ... + Pr(Sn |A)U(ASn )

Note in particular that we put into this formula the probability of each state given that A
is chosen. We don’t take the unconditional probability of being in that state. These numbers
can come quite dramatically apart.
In Newcomb’s problem, it is actually quite hard to say what the probability of each state
is. (The states here, of course, are just that there is either $1,000,000 in the opaque box or that
there is nothing in it.) But what’s easy to say is the probability of each state given the choices
you make. If you choose both boxes, the probability that there is nothing in the opaque box
is very high, and the probability that there is $1,000,000 in it is very low. Conversely, if you
choose just the one box, the probability that there is $1,000,000 in it is very high, and the
probability that there is nothing in it is very low. Simplifying just a little, we’ll say that this
high probability is 1, and the low probabiilty is 0. The expected utility of each choice then is

86
Exp(U(Take both boxes))
= Pr(Million in opaque box|Take both boxes)U(Take both boxes and million in opaque box)
+ Pr(Nothing in opaque box|Take both boxes)U(Take both boxes and nothing in opaque box)
= 0 × 1, 001, 000 + 1 × 1, 000
= 1, 000
Exp(U(Take one box))
= Pr(Million in opaque box|Take one box)U(Take one box and million in opaque box)
+ Pr(Nothing in opaque box|Take one box)U(Take one box and nothing in opaque box)
= 1 × 1, 000, 000 + 0 × 0
= 1, 000, 000

I’ve assumed here that the marginal utility of money is constant, so we can measure
utility by the size of the numerical prize. That’s an idealisation, but hopefully a harmless
enough one.

16.3 Bringing Two Principles Together

In earlier chapters we argued that the expected utility rule never led to a conflict with the
dominance principle. But here it has led to a conflict. Something seems to have gone badly
wrong.
The problem was that we’ve used two distinct definitions of expected utility in the two
arguments. In the version we had used in previous chapters, we presupposed that the prob-
ability of the states was independent of the choices that were made. So we didn’t talk about
Pr(S1 |A) or Pr(S1 |B) or whatever. We simply talked about Pr(S1 ).
If you make that assumption, expected utility maximisation does indeed imply domi-
nance. We won’t rerun the entire proof here, but let’s see how it works in this particular
case. Let’s say that the probability that there is $1,000,000 in the opaque box is x. It won’t
matter at all what x is. And assume that the expected utility of a choice A is given by this
formula, where we use the unconditional probability of states as inputs.

Exp(U(A)) = Pr(S1 )U(AS1 ) + ... + Pr(Sn |A)U(ASn )

Applied to our particular case, that would give us the following calculations.

87
Exp(U(Take both boxes))
= Pr(Million in opaque box)U(Take both boxes and million in opaque box)
+ Pr(Nothing in opaque box)U(Take both boxes and nothing in opaque box)
= x × 1, 001, 000 + (1 – x) × 1, 000
= 1, 000 + 1, 000, 000x
Exp(U(Take one box))
= Pr(Million in opaque box)U(Take one box and million in opaque box)
+ Pr(Nothing in opaque box)U(Take one box and nothing in opaque box)
= x × 1, 000, 000 + (1 – x) × 0
= 1, 000, 000x

And clearly the expected value of taking both boxes is 1,000 higher than the expected
utility of taking just one box. So as long as we don’t conditionalise on the act we are per-
forming, there isn’t a conflict between the dominance principle and expected utility max-
imisation.
While that does resolve the mathematical puzzle, it hardly resolves the underlying philo-
sophical problem. Why, we might ask, shouldn’t we conditionalise on the actions we are
performing? In general, it’s a bad idea to throw away information, and the choice that we’re
about to make is a piece of information. So we might think it should make a difference to
the probabilities that we are using.
The best response to this argument, I think, is that it leads to the wrong results in New-
comb’s problem, and related problems. But this is a somewhat controversial clam. After
all, some people think that taking one box is the right result in Newcomb’s problem. And
as we saw above, if we conditionalise on our action, then the expected utility of taking one
box is higher than the expected utility of taking both. So such theorists will not think that
it gives the wrong answer at all. To address this worry, we need to look more closely back at
Newcomb’s original problem, and its variants.

16.4 Well Meaning Friends

The next few sections are going to involve looking at arguments that we should take both
boxes in Newcomb’s problem, or to rejecting arguments that we should only take one box.
The simplest argument is just a dramatisation of the dominance argument. But still, it
is a way to see the force of that argument. Imagine that you have a friend who can see into
the opaque box. Perhaps the box is clear from behind, and your friend is standing behind
the box. Or perhaps your friend has super-powers that let them see into opaque boxes. If
your friend was able to give you advice, and has your best interests at heart, they’ll tell you
to take both boxes. That’s true whether or not there is a million dollars in the opaque box.
Either way, they’ll know that you’re better off taking both boxes.
Of course, there are lots of cases where a friend with more knowledge than you and your
interests at heart will give you advice that is different to what you might intuitively think is
correct. Imagine that I have just tossed a biased coin that has an 80% chance of landing

88
heads. The coin has landed, but neither of us can see how it has landed. I offer you a choice
between a bet that pays $1 if it landed heads, and a bet that pays $1 if it landed tails. Since
heads is more likely, it seems you should take the bet on heads. But if the coin has landed
tails, then a well meaning and well informed friend will tell you that you should bet on tails.
But that case is somewhat different to the friend in Newcomb’s problem. The point here
is that you know what the friend will tell you. And plausibly, whenever you know what
advice a friend will give you, you should follow that advice. Even in the coin-flip case, if you
knew that your friend would tell you to bet on tails, it would be smart to bet on tails. After
all, knowing that your friend would give you that advice would be equivalent to knowing
that the coin landed tails. And if you knew the coin landed tails, then whatever arguments
you could come up with concerning chances of landing tails would be irrelevant. It did land
tails, so that’s what you should bet on.
There is another way to dramatise the dominance argument. Imagine that after the boxes
are opened, i.e. after you know which state you are in, you are given a chance to revise your
choice if you pay $500. If you take just one box, then whatever is in the opaque box, this will
be a worthwhile switch to make. It will either take you from $0 to $500, or from $1,000,000
to $1,000,500. And once the box is open, there isn’t even an intuition that you should worry
about how the box got filled. So you should make the switch.
But it seems plausible in general that if right now you’ve got a chance to do X, and you
know that if you don’t do X now you’ll certainly pay good money to do X later, and you
know that when you do that you’ll be acting perfectly rationally, then you should simply do
X. After all, you’ll get the same result whether you do X now or later, you’ll simply not have
to pay the ‘late fee’ for taking X any later. More relevantly to our case, if you would switch to
X once the facts were known, even if doing so required paying a fee, then it seems plausible
that you should simply do X now. It doesn’t seem that including the option of switching
after the boxes are revealed changes anything about what you should do before the boxes
are revealed, after all.
Ultimately, I’m not sure that either of the arguments I gave here, either the well meaning
friend argument or the switching argument, are any more powerful than the dominance ar-
gument. Both of them are just ways of dramatising the dominance argument. And someone
who thinks that you should take just one box is, by definition, someone who isn’t moved by
the dominance argument. In the next set of notes we’ll look at other arguments for taking
both boxes.

89
 Chapter 17

Realistic Newcomb Problems

17.1 Real Life Newcomb Cases
In the previous notes we ended up saying that there are two quite different ways to think
about utility expectations. We can use the unconditional probability of each state, or, for
each choice, we can use the probabilities of each state conditional on the choice the agent
makes. That is, we can take the expected utility of a choice A to be given by one or other of
the following formulae.

Pr(S1 )U(S1 A) + ... + Pr(Sn )U(Sn A)

Pr(S1 |A)U(S1 A) + ... + Pr(Sn |A)U(Sn A)

It would be nice to know which of these is the right formula, since the two formulae
disagree about cases like Newcomb’s problem. Since we have a case where they disagree, a
simple methodology suggests itself. Figure out what we should do in Newcomb’s problem,
and then select the formula which agrees with the correct answer. But this method has two
flaws.
First, intuitions about Newcomb’s puzzle are themselves all over the place. If we try to
adjust our theory to match our judgments in Newcomb’s problem, then different people will
have different theories.
Second, Newcomb’s problem is itself quite fantastic. This is part of why different people
have such divergent intuitions on the example. But it also might make us think that the
problem is not particularly urgent. If the two equations only come apart in fantastic cases
like this, perhaps we can ignore the puzzles.
So it would be useful to come up with more realistic examples where the two equations
come apart. It turns out that what is driving the divergence between the equations is that
there is a common cause of the world being in a certain state and you making the choice
that you make. Any time there is something in the world that tracks your decision making
processes, we’ll have a Newcomb like problem.
For example, imagine that we are in a Prisoners’ Dilemma situation where we know that
the other prisoner uses very similar decision making procedures to what we use. Here is the
table for a Prisoners’ Dilemma.

90
 Other Other
Cooperates Defects
You Cooperate (3,3) (0, 5)
You Defect (5, 0) (1, 1)

In this table the notation (x, y) means that you get x utils and the other person gets y utils.
Remember that utils are meant to be an overall measure of what you value, so it includes
your altruistic care for the other person.
Let’s see why this resembles a Newcomb problem. Assume that conditional on your
performing an action A, the probability that the other person will do the same action is 0.9.
Then, if we are taking probabilities to be conditional on choices, the expected utility of the
two choices is

Exp(U(Coop)) = 0.9 × 3 + 0.1 × 0

= 2.7
Exp(U(Defect)) = 0.1 × 5 + 0.9 × 1
= 1.4

So if we use probabilities conditional on choices, we end up with the result that you should
cooperate. But note that cooperation is dominated by defection. If the other person defects,
then your choice is to get 1 (by defecting) or 0 (by cooperating). You’re better off cooper-
ating. If the other person cooperates, then your choice is to get 5 (by defecting) or 0 (by
cooperating). So whatever probability we give to the possible actions of the other person,
provided we don’t conditionalise on our choice, we’ll end up deciding to defect.
Prisoners Dilemma cases are much less fantastic than Newcomb problems. Even Pris-
oners Dilemma cases where we have some confidence that the other party sufficiently re-
sembles us that they will likely (not certainly) make the same choice as us are fairly realistic.
So they are somewhat better than Newcomb’s original problem for detecting intuitions. But
the problem of divergent intuitions still remains. Many people are unsure about what the
right thing to do in a Prisoners Dilemma problem is. (We’ll come back to this point when
we look at game theory.)
So it is worth looking at some cases without that layer of complication. Real life cases
are tricky to come by, but for a while some people suggested that the following might be
a case. We’ve known for a long time that smoking causes various cancers. We’ve known
for even longer than that that smoking is correlated with various cancers. For a while there
was a hypothesis that smoking did not cause cancer, but was correlated with cancer because
there was a common cause. Something, presumably genetic, caused people to (a) have a
disposition to smoke, and (b) develop cancer. Crucially, this hypothesis went, smoking did
not raise the risk of cancer; whether you got cancer or not was largely due to the genes that
led to a desire for smoking.
We now know, by means of various tests, that this isn’t true. (For one thing, the reduction
in cancer rates among people who give up smoking is truly impressive, and hard to explain
on the model that these cancers are all genetic.) But at least at some point in history it was a
not entirely crazy hypothesis. Let’s assume this hypothesis is actually true (contrary to fact).

91
And let’s assume that you (a) want to smoke, other things being equal, and (b) really don’t
want to get cancer. You don’t know whether you have the desire for smoking/disposition to
get cancer gene or not? What should you do?
Plausibly, you should smoke. You either have the gene or you don’t. If you do, you’ll
probably get cancer, but you can either get cancer while smoking, or get cancer while not
smoking, and since you enjoy smoking, you should smoke. If you don’t, you won’t get cancer
whether you smoke or not, so you should indulge your preference for smoking.
It isn’t just philosophers who think this way. At some points (after the smoking/cancer
correlation was discovered but before the causal connection was established) various to-
bacco companies were trying very hard to get evidence for this ‘common cause’ hypothesis.
Presumably the reason they were doing this was because they thought that if it were true, it
would be rational for people to smoke more, and hence people would smoke more.
But note that this presumption is true if and only if we use the ‘unconditional’ version
of expected utility theory. To see this, we’ll use the following table for the various outcomes.

Get Cancer Don’t get Cancer

Smoke 1 6
Don’t Smoke 0 5

The assumption is that not getting cancer is worth 5 to you, while smoking is worth 1 to
you. Now we know that smoking is evidence that you have the cancer gene, and this raises
dramatically the chance of you getting cancer. So the (evidential) probability of getting
cancer conditional on smoking is, we’ll assume, 0.8, while the (evidential) probability of
getting cancer conditional on not smoking is, we’ll assume, 0.2. And remember this isn’t
because cancer causes smoking in our example, but rather that there is a common cause of
the two. Still, this is enough to make the expected utilities work out as follows.

Exp(U(Smoke)) = 0.8 × 1 + 0.2 × 6

=2
Exp(U(Don’t Smoke)) = 0.2 × 0 + 0.8 × 5
=4

And the recommendation is not to smoke, even though smoking dominates. This seems
very odd. As it is sometimes put, the recommendation here seems to be a matter of manag-
ing the ‘news’, not managing the outcome. What’s bad about smoking is that if you smoke
you get some evidence that something bad is going to happen to you. In particular, you get
evidence that you have this cancer gene, and that’s really bad news to get because dramati-
cally raises the probability of getting cancer. But not smoking doesn’t mean that you don’t
have the gene, it just means that you don’t find out that you have the gene. Not smoking
looks like a policy of denying yourself good outcomes because you don’t want to get bad
news. And this doesn’t look rational.
So this case has convinced a lot of decision theorists that we shouldn’t use conditional
probabilities of states when working out the utility of various outcomes. Using conditional

92
probabilities will be good if we want to learn the ‘news value’ of some choices, but not if we
want to learn how useful those choices will be to us.

17.2 Tickle Defence

Not everyone has been convinced by these ‘real-life’ examples. The counter-argument is that
in any realistic case, the gene that leads to smoking has to work by changing our dispositions.
So there isn’t just a direct causal connection between some genetic material and smoking.
Rather, the gene causes a desire to smoke, and the desire to smoke cause the smoking. As
it is sometimes put, between the gene and the smoking there has to be something mental, a
‘tickle’ that leads to the smoking.
Now this is important because we might think that rational agents know their own men-
tal states. Let’s assume that for now. So if an agent has the smoking desire they know it,
perhaps because this desire has a distinctive phenomenology, a tickle of sorts. And if the
agent knows this, then they won’t get any extra evidence that they have a desire to smoke
from their actual smoking. So the probability of getting cancer given smoking is not higher
than the probability of getting cancer given not smoking.
In the case we have in mind, the bad news is probably already here. Once the agent
realises that their values are given by the table above, they’ve already got the bad news.
Someone who didn’t have the gene wouldn’t value smoking more than not smoking. Once
the person conditionalises on the fact that that is their value table, the evidence that they
actually smoke is no more evidence. Either way, they are (say) 80% likely to get cancer. So
the calculations are really something like this

Exp(U(Smoke)) = 0.8 × 1 + 0.2 × 6

=2
Exp(U(Don’t Smoke)) = 0.8 × 0 + 0.2 × 5
=1

And we get the correct answer that in this situation we should smoke. So this isn’t a case
where the two different equations we’ve used give different answers. And hence it isn’t a
reason for using unconditional probabilities rather than conditional probabilities.
There are two common responses to this argument. The first is that it isn’t clear that
there is always a ‘tickle’. The second is that it isn’t a requirement of rationality that we know
what tickles we have. Let’s look at these in turn.
First, it was crucial to this defence that the gene (or whatever) that causes both smoking
and cancer causes smoking by causing some particular mental state first. But this isn’t a
necessary feature of the story. It might be that, say, everyone has the ‘tickle’ that goes along
with wanting to smoke. (Perhaps this desire has some evolutionary advantage. Or, more
likely, it might be a result of something that genuinely had evolutionary advantage.) Perhaps
what the gene does is to affect how much willpower we have, and hence how likely we are
to overcome the desire.
Second, it was also crucial to the defence that it is a requirement of rationality that people
know what ‘tickles’ they have. If this isn’t supposed, we can just imagine that our agent is a
rational person who is ignorant of their own desires. But this supposition is quite strong. It

93
is generally not a requirement of rationality that we know things about the external world.
Some things are just hidden from us, and it isn’t a requirement of rationality that we be
able to see what is hidden. Similarly, it seems at least possible that some things in our own
mind should be hidden. Whether or not you believe in things like subconscious desires, the
possibility of them doesn’t seem to systematically undermine human rationality.
Note that these two responses dovetail nicely. If we think that the gene works not by
producing individual desires, but by modifying quite general standing dispositions like how
much willpower we have, it is even more plausible to think that this is not something a
rational person will always know about. It is a little odd to think of a person who desires to
smoke but doesn’t realise that they desire to smoke. It isn’t anywhere near as odd to think
about a person who has very little willpower but, perhaps because their willpower is rarely
tested, doesn’t realise that they have low willpower. Unless they are systematically ignoring
evidence that they lack willpower, they aren’t being clearly irrational.
So it seems there are possible, somewhat realistic, cases where one choice is evidence,
to a rational agent, that something bad is likely to happen, even though the choice does
not bring about the bad outcome. In such a case using conditional probabilities will lead
to avoiding the bad news, rather than producing the best outcomes. And that seems to be
irrational.

94
 Chapter 18

Causal Decision Theory

18.1 Causal and Evidential Decision Theory
Over the last two chapters we’ve looked at two ways of thinking about the expected utility
of an action A. These are

Pr(S1 )U(S1 A) + ... + Pr(Sn )U(Sn A)

Pr(S1 |A)U(S1 A) + ... + Pr(Sn |A)U(Sn A)

It will be convenient to have names for these two approaches. So let’s say that the first of
these, which uses unconditional probabilities, is causal expected value, and the second of
these, which uses conditional probabilities is the evidential expected value. The reason for
the names should be clear enough. The causal expected value measures what you can expect
to bring about by your action. The evidential expected value measures what kind of result
your action is evidence that you’ll get.
Causal Decision Theory then is the theory that rational agents aim to maximise causal
expected utility.
Evidential Decision Theory is the theory that rational agents aim to maximise evidential
expected utility.
Over the past two chapters we’ve been looking at reasons why we should be causal de-
cision theorists rather than evidential decision theorists. We’ll close out this section by
looking at various puzzles for causal decision theory, and then looking at one reason why
we might want some kind of hybrid approach.

18.2 Right and Wrong Tabulations

If we use the causal approach, it is very important how we divide up the states. We can see
this by thinking again about an example from Jim Joyce that we discussed a while ago.

Suupose you have just parked in a seedy neighborhood when a man ap-
proaches and offers to “protect” your car from harm for $10. You recognize
this as extortion and have heard that people who refuse “protection” invari-
ably return to find their windshields smashed. Those who pay find their cars
intact. You cannot park anywhere else because you are late for an important
meeting. It costs $400 to replace a windshield. Should you buy “protection”?
Dominance says that you should not. Since you would rather have the extra $10

95
 both in the even that your windshield is smashed and in the event that it is not,
Dominance tells you not to pay. (Joyce, The Foundations of Causal Decision
Theory, pp 115-6.)

If we set this up as a table, we get the following possible states and outcomes.

Broken Windshield Unbroken Windshield

Pay extortion -$410 -$10
Don’t pay -$400 0

Now if you look at the causal expected value of each action, the expected value of not paying
will be higher. And this will be so whatever probabilities you assign to broken windshield
and unbroken windshield. Say that the probability of the first is x and of the second is 1 – x.
Then we’ll have the following (assuming dollars equal utils)

Exp(U(Pay extortion)) = –410x – 10(1 – x)

= –400x – 10
Exp(U(Don’t pay) = –400x – 0(1 – x)
= –400x

Whatever x is, the causal expected value of not paying is higher by 10. That’s obviously
a bad result. Is it a problem for causal decision theory though? No. As the name ‘causal’
suggests, it is crucial to causal decision theory that we separate out what we have causal
power over from what we don’t have causal power over. The states of the world represent
what we can’t control. If something can be causally affected by our actions, it can’t be a
background state.
So this is a complication in applying causal decision theory. Note that it is not a problem
for evidential decision theory. We can even use the very table that we have there. Let’s
assume that the probability of broken windshield given paying is 0, and the probability of
unbroken windshield given paying is 0. Then the expected utilities will work out as follows

Exp(U(Pay extortion)) = –410 × 0 – 10 × 1

= –10
Exp(U(Don’t pay) = –400 × 1 – 10 × 0
= –400

So we get the right result that we should pay up. It is a nice feature of evidential decision
theory that we don’t have to be so careful about what states are and aren’t under our control.
Of course, if the only reason we don’t have to worry about what is and isn’t under our control
is that the theory systematically ignores such facts, even though they are intuitively relevant
to decision theory, this isn’t perhaps the best advertisement for evidential decision theory.

96
18.3 Why Ain’Cha Rich
There is one other argument for evidential decision theory that we haven’t yet addressed.
Causal decision theory recommends taking two boxes in Newcomb’s problem; evidential
decision theory recommends only taking one. People who take both boxes tend, as a rule,
to end up poorer than people who take just the one box. Since the aim here is to get the best
outcome, this might be thought to be embarrassing for causal decision theorists.
Causal decision theorists have a response to this argument. They say that Newcomb’s
problem is a situation where there is someone who is quite smart, and quite determined to
reward irrationality. In such a case, they say, it isn’t too surprising that irrational people, i.e.
evidential decision theorists, get rewarded. Moreover, if a rational person like them were to
have taken just one box, they would have ended up with even less money, i.e., they would
have ended up with nothing.
One way that causal decision theorists would have liked to make this objection stronger
would be to show that there is a universal problem for decision theories - whenever there is
someone whose aim is to reward people who don’t follow the dictates of their theory, then
the followers of their theory will end up poorer than the non-followers. That’s what happens
to causal decision theorists in Newcomb’s problem. It turns out it is hard, however, to play
such a trick on evidential decision theorists.
Of course we could have someone go around and just give money to people who have
done irrational things. That wouldn’t be any sign that the theory is wrong however. What’s
distinctive about Newcomb’s problem is that we know this person is out there, rewarding
non-followers of causal decision theory, and yet the causal decision theorist does not change
their recommendation. In this respect they differ from evidential decision theorists.
It turns out to be very hard, perhaps impossible, to construct a problem of this sort for
evidential decision theorists. That is, it turns out to be hard to construct a problem where
(a) an agent aims to enrich all and only those who don’t follow evidential decision theory,
(b) other agents know what the devious agent is doing, but (c) evidential decision theory
still ends up recommending that you side with those who end up getting less money. If the
devious agent rewards doing X, then evidential decision theory will (other things equal)
recommend doing X. The devious agent will make such a large evidential difference that
evidential decision theory will recommend doing the thing the devious agent is rewarding.
So there’s no simple response to the “Why Ain’Cha Rich” rhetorical question. The causal
decision theorist says it is because there is a devious agent rewarding irrationality. The evi-
dential decision theorist says that a theory should not allow the existence of such an agent.
This seems to be a standoff.

18.4 Dilemmas
Consider the following story, told by Allan Gibbard and William Harper in their paper set-
ting out causal decision theory.
Consider the story of the man who met Death in Damascus. Death looked
surprised, but then recovered his ghastly composure and said, ‘I am coming
for you tomorrow’. The terrified man that night bought a camel and rode to
Aleppo. The next day, Death knocked on the door of the room where he was
hiding, and said ‘I have come for you’.

97
 ‘But I thought you would be looking for me in Damascus’, said the man.
‘Not at all’, said Death ‘that is why I was surprised to see you yesterday.
I knew that today I was to find you in Aleppo’.
Now suppose the man knows the following. Death works from an appointment
book which states time and place; a person dies if and only if the book correctly
states in what city he will be at the stated time. The book is made up weeks in
advance on the basis of highly reliable predictions. An appointment on the next
day has been inscribed for him. Suppose, on this basis, the man would take his
being in Damascus the next day as strong evidence that his appointment with
Death is in Damascus, and would take his being in Aleppo the next day as
strong evidence that his appointment is in Aleppo...
If... he decides to go to Aleppo, he then has strong grounds for expecting that
Aleppo is where Death already expects him to be, and hence it is rational for
him to prefer staying in Damascus. Similarly, deciding to stay in Damascus
would give him strong grounds for thinking that he ought to go to Aleppo.
In cases like this, the agent is in a real dilemma. Whatever he does, it seems that it will be
the wrong thing. If he goes to Aleppo, then Death will probably be there. And if he stays in
Damascus, then Death will probably be there as well. So it seems like he is stuck.
Of course in one sense, there is clearly a right thing to do, namely go wherever Death
isn’t. But that isn’t the sense of right decision we’re typically using in decision theory. Is
there something that he can do that maximises expected utility. In a sense the answer is
“No”. Whatever he does, doing that will be some evidence that Death is elsewhere. And
what he should do is go wherever his evidence suggests Death isn’t. This turns out to be
impossible, so the agent is bound not to do the rational thing.
Is this a problem for causal decision theory? It is if you think that we should always have
a rational option available to us. If you think that ‘rational’ here is a kind of ‘ought’, and
you think ‘ought’ implies ‘can’, then you might think we have a problem, because in this case
there’s a sense in which the man can’t do the right thing. (Though this is a bit unclear; in the
actual story, there’s a perfectly good sense in which he could have stayed in Aleppo, and the
right thing to do, given his evidence, would have been to stay in Aleppo. So in one sense
he could have done the right thing.) But both the premises of the little argument here are
somewhat contentious. It isn’t clear that we should say you ought, in any sense, maximise
expected utility. And the principle that ought implies can is rather controversial. So perhaps
this isn’t a clear counterexample to causal decision theory.

18.5 Weak Newcomb Problems

Imagine a small change to the original Newcomb problem. Instead of there being $1000 in
the clear box, there is $800,000. Still, evidential decision theory recommends taking one
box. The evidential expected value of taking both boxes is now roughly $800,000, while the
evidential expected value of taking just the one box is $1,000,000. Causal decision theory
recommends taking both boxes, as before.

98
 So neither theory changes its recommendations when we increase the amount in the
clear box. But I think many people find the case for taking just the one box to be less com-
pelling in this variant. Does that suggest we need a third theory, other than just causal or
evidential decision theory?
It turns out that we can come up with hybrid theories that recommend taking one box
in the original case, but two boxes in the original case. Remember that in principle any-
thing can have a probability, including theories of decision. So let’s pretend that given the
(philosophical) evidence on the table, the probability of causal decision theory is, say, 0.8,
while the probability of evidential decision theory is 0.2. (I’m not saying these numbers
are right, this is just a possibility to float.) And let’s say that we should do the thing that
has the highest expected expected utility, where we work out expected expected utilities by
summing over the expectation of the action on different theories, times the probability of
each theory. (Again, I’m not endorsing this, just floating it.)
Now in the original Newcomb problem, evidential decision theory says taking one boxes
is $999,000 better, while causal decision theory say staking both boxes is $1,000 better. So
the expected expected utility of taking one box rather than both boxes is 0.2 × 999, 000 –
0.8 × 1, 000, which is 199,000. So taking one box is ‘better’ by 199,000
In the modified Newcomb problem, evidential decision theory says taking one boxes is
$200,000 better, while causal decision theory says taking both boxes is $800,000 better. So
the expected expected utility of taking one box rather than both boxes is 0.2 × 200, 000 –
0.8 × 800, 000, i.e., -600,000. So taking both boxes is ‘better’ by 600,000.
If you think that changing the amount in the clear box can change your decision in
Newcomb’s problem, then possibly you want a hybrid theory, perhaps like the one floated
here.

99
 Chapter 19

Introduction to Games
19.1 Games
A game is any decision problem where the outcome turns on the actions of two or more
individuals. We’ll entirely be concerned here with games where the outcome turns on the
actions of just two agents, though that’s largely because the larger cases are more mathemat-
ically complicated.
Given a definition that broad, pretty much any human interaction can be described as a
game. And indeed game theory, the study of games in this sense, is one of the most thriving
areas of modern decision theory. Game theory is routinely used in thinking about conflicts,
such as warfare or elections. It is also used in thinking about all sorts of economic inter-
actions. Game theorists have played crucial (and lucrative) roles in recent years designing
high-profile auctions, for example. The philosopher and economist Ken Binmore, for exam-
ple, led the team that used insights from modern game theory to design the auction of the
3G wireless spectrum in Britain. That auction yielded the government billions of pounds
more than was anticipated.
When we think of the ordinary term ‘game’, we naturally think of games like football or
chess, where there are two players with conflicting goals. But these games are really quite
special cases. What’s distinctive of football and chess is that, to a first approximation, the
players’ goals are completely in conflict. Whatever is good for the interests of one player is
bad for the interests of the other player. This isn’t what’s true of most human interaction.
Most human interaction is not, as we will put it here, zero sum. When we say that an inter-
action is zero sum, what we mean (roughly) that the net outcome for the players is constant.
(Such games may better be called ‘constant-sum’.)
We’ll generally represent games using tables like the following. Each row represents a
possible move (or strategy) for a player called Row, and each column represents a possible
move (or strategy) for a player called Column. Each cell represents the payoffs for the two
players. The first number is the utility that Row receives for that outcome, and the second
number is the utility that Column receives for that outcome. Here is an example of a game.
(It’s a version of a game called the Stag Hunt.)

Team Solo
Team (4, 4) (1, 3)
Solo (3, 1) (3, 3)

100
Each player has a choice between two strategies, one called ‘Team’ and the other called ‘Solo’.
(The model here is whether players choose to hunt alone or as a team. A team produces
better results for everyone; if it is large enough.) Whoever plays Solo is guaranteed to get an
outcome of 3. If someone plays Team, they get 4 if the other player plays Team as well, and
1 if the other player plays solo.
A zero sum game is where the outcomes all sum to a constant. (For simplicity, we usually
make this constant zero.) So here is a representation of (a single game of) Rock-Paper-
Scissors.

Rock Paper Scissors

Rock (0, 0) (-1, 1) (1, -1)
Paper (1, -1) (0, 0) (-1, 1)
Scissors (-1, 1) (1, -1) (0, 0)

Sometimes we will specify that the game is a zero sum game and simply report the payoffs
for Row. In that case we’d represent Rock-Paper-Scissors in the following way.

Rock Paper Scissors

Rock 0 -1 1
Paper 1 0 -1
Scissors -1 1 0

The games we’ve discussed so far are symmetric, but that need not be the case. Consider a
situation where two people are trying to meet up and don’t have any way of getting in touch
with each other. Row would prefer to meet at the Cinema, Column would prefer to meet at
the Opera. But they would both prefer to meet up than to not meet up. We might represent
the game as follows.

Cinema Opera
Cinema (3, 2) (1, 1)
Opera (0, 0) (2, 3)

We will make the following assumptions about all games we discuss. Not all game the-
orists make these assumptions, but we’re just trying to get started here. First, we’ll assume
that the players have no means of communicating, and hence no means of negotiating. Sec-
ond, we’ll assume that all players know everything about the game table. That is, they know
exactly how much each outcome is worth to each player.
Finally, we’ll assume that all the payoffs are in ‘utils’. We won’t assume that the payoffs
are fully determinate. The payoff might be a probability distribution over outcomes. For
example, in the game above, consider the top left outcome, where we say Row’s payoff is 3.
It might be that Row doesn’t know if the movie will be any good, and thinks there is a 50%
chance of a good movie, with utility 5, and a 50% chance of a bad movie, with utility 1. In
that case Row’s expected utility will be 3, so that’s what we put in the table. (Note that this

101
makes the assumption that the players know the full payoff structure quite unrealistic, since
players typically don’t know the probabilities that other players assign to states of the world.
So this is an assumption that we might like to drop in more careful work.)
For the next few handouts, we’ll assume that the interaction between the players is ended
when they make their, simultaneous, moves. So these are very simple one-move games.
We’ll get to games that involve series of moves in later handouts. But for now we just want
to simplify by thinking of cases where Row and Column move simultaneously, and that ends
the game/interaction.

19.2 Zero-Sum Games and Backwards Induction

Zero-sum games are the simplest to theorise about, so we’ll start with them. They are also
quite familiar as ‘games’, though as we said above, most human interaction is not zero-sum.
Zero-sum games are sometimes called ‘strictly competitive’ games, and we’ll use that ter-
minology as well sometimes, just to avoid repetition. For all of this section we’ll represent
zero-sum games by the ‘one-number’ method mentioned above, where the aim of Row is to
maximise that number, and the aim of Column is to minimise it.
Zero-sum games can’t have pairs strictly dominating options for each player. That’s be-
cause what is good for Row is bad for Column. But they can have outcomes that are ended
up at by a process of removing something like dominated outcomes. Consider, for instance,
the following game.

A B C
A 5 6 7
B 3 7 8
C 4 1 9

Column pretty clearly isn’t going to want to play C, because that is the worst possible out-
come whatever Row plays. Now C could have been a good play for Row, it could have ended
up with the 9 in the bottom-right corner. But that isn’t a live option any more. Column isn’t
going to play C, so really Row is faced with something like this game table.

A B
A 5 6
B 3 7
C 4 1

And in that table, C is a dominated outcome. Row is better off playing A than C, whatever
Column plays. Now Column can figure this out too. So Column knows that Row won’t play
C, so really Column is faced with this choice.

A B
A 5 6
B 3 7

102
And whatever Row plays now, Column is better off playing A. Note that this really requires
the prior inference that Row won’t play C. If C was a live option for Row, then B might be the
best option for Column. But that isn’t really a possibility. So Column will play A. And given
that’s what Column will do, the best thing for Row to do is to play A. So just eliminating
dominated options repeatedly in this way gets us to the solution that both players will play
A.
So something like repeated dominance reasoning can sometimes get us to the solution
of a game. It’s worth spending a bit of time reflecting on the assumptions that went into
the arguments we’ve used here. We had to assume that Row could figure out that Column
will play A. And that required Column figuring out that Row will not play C. And Column
could only figure that out if they could figure out that Row would figure out that they, i.e.
Column, would not play C. So Column has to make some strong assumptions about not
only the rationality of the other player, but also about how much the other player can know
about their own rationality. In games with more than three outcomes, the players may have
to use more complicated assumptions, e.g. assumptions about how rational the other player
knows that they know that that other player is, or about whether the other player knows
they are in a position to make such assumptions, and so on.
This is all to say that even a relatively simple argument like this, and it was fairly sim-
ple as game theoretic arguments go, has some heavy duty assumptions about the players’
knowledge and assumptions built into it. This will be a theme we’ll return to a few times.

19.3 Zero-Sum Games and Nash Equilibrium

Not all games can be solved by the method described in the previous section. Sometimes
there are no dominated options for either player, so we can’t get started on this strategy.
And sometimes the method described there won’t get to a result at one point or another.
Consider, for example, the following game.

A B C
A 5 6 7
B 3 7 2
C 4 2 9

No option is dominated for either player, so we can’t use the ‘eliminate dominated options’
method. But there is still something special about the (A, A) outcome. That is, if either
player plays A, the other player can’t do better than by playing A. That’s to say, the outcome
(A, A) is a Nash equilibrium.

• A pair of moves (xi , yi ) by Row and Column respectively is a Nash equilibrium if (a)
Row can’t do any better than playing xi given that Column is playing yi , and Column
can’t do any better than playing yi , given that Row is playing xi .

103
 Assume that each player knows everything the other player knows. And assume that the
players are equally, and perfectly, rational. Then you might conclude that each player will be
able to figure out the strategy of the other. Now assume that the players pick (between them)
a pair of moves that do not form a Nash equilibrium. Since the players know everything
about the other player, they know what the other will do. But if the moves picked do not
form a Nash equilibrium, then one or other player could do better, given what the other
does. Since each player knows what the other will do, that means that they could do better,
given what they know. And that isn’t rational.
The argument from the previous paragraph goes by fairly fast, and it isn’t obviously wa-
tertight, but it suggests that there is a reason to think that players should end up playing parts
of Nash equilibrium strategies. So identifying Nash equilibria, like (A, A) in this game, is a
useful way to figure out what they each should do.
Some games have more than one Nash equilibria. Consider, for instance, the following
game.

A B C D
A 5 6 5 6
B 5 7 5 7
C 4 8 3 8
D 3 8 4 9

In this game, both (A, A) and (B, C) are Nash equilibria. Note two things about the game.
First, the ‘cross-strategies’, where Row plays one half of one Nash equilibrium, and Columns
plays the other half of a different Nash equilibrium, are also Nash equilibria. So (A, C) and
(B, A) are both Nash equilibria. Second, all four of these Nash equilibria have the same
value. In one of the exercises later on, you will be asked to prove both of these facts.

104
 Chapter 20

Zero-Sum Games
20.1 Mixed Strategies
In a zero-sum game, there is a simple way to tell that an outcome is a Nash equilibrium
outcome. It has to be the smallest value in the row it is (else Column could do better going
elsewhere) and the highest value in the column it is in (else Row could do better by going
elsewhere). But once we see this, we can see that several games do not have any simple Nash
equilibrium. Consider again Rock-Paper-Scissors.

Rock Paper Scissors

Rock 0 -1 1
Paper 1 0 -1
Scissors -1 1 0

There is no number that’s both the lowest number in the row that it is in, and the highest
number in the row that it is in. And this shouldn’t be too surprising. Let’s think about what
Nash equilibrium means. It means that a move is the best each player can do even if the
other player plays their part of the equilibrium strategy. That is, it is a move such that if
one player announced their move, the other player wouldn’t want to change. And there’s no
such move in Rock-Paper-Scissors. The whole point is to try to trick the other player about
what your move will be.
So in one sense there is no Nash equilibrium to the game. But in another sense there is
an equilibrium to the game. Let’s expand the scope of possible moves. As well as picking
one particular play, a player can pick a mixed strategy.

• A mixed strategy is where the player doesn’t decide which move they will make, but
decides merely the probability with which they will make certain moves.
• Intuitively, picking a mixed strategy is deciding to let a randomising device choose
what move you’ll make; the player’s strategy is limited to adjusting the settings on the
randomising device.

105
 We will represent mixed strategies in the following way. <0.6 Rock; 0.4 Scissors> is the
strategy of playing Rock with probability 0.6, and Scissors with probability 0.4. Now this
isn’t a great strategy to announce. The other player can do well enough by responding Rock,
which has an expected return of 0.4 (Proof: if the other player plays Rock, they have an 0.6
chance of getting a return of 0, and an 0.4 chance of getting a return of 1. So their expected
return is 0.6 × 0 + 0.4 × 1 = 0.4.) But this is already a little better than any ‘pure’ strategy. A
pure strategy is just any strategy that’s not a mixed strategy. For any pure strategy that you
announce, the other player can get an expected return of 1.
Now consider the strategy < 13 Rock, 13 Paper, 13 Scissors>. Whatever pure strategy the
other player chooses, it has an expected return of 0. That’s because it has a 13 chance of a
return of 1, a 13 chance of a return of 0, and a 13 chance of a return of -1. As a consequence
of that, whatever mixed strategy they choose has an expected return of 0. That’s because the
expected return of a mixed strategy can be calculated by taking the expected return of each
pure strategy that goes into the mixed strategy, multiplying each number by the probability
of that pure strategy being played, and summing the numbers.
The consequence is that if both players play < 13 Rock, 13 Paper, 13 Scissors>, then each
has an expected return of 0. Moreover, if each player plays this strategy, the other player’s
expected return is 0 no matter what they play. That’s to say, playing < 13 Rock, 13 Paper,
1 1 1 1
3 Scissors> does as well as anything they can do. So the ‘outcome’ (< 3 Rock, 3 Paper, 3
1 1 1
Scissors>, < 3 Rock, 3 Paper, 3 Scissors>), i.e. the outcome where both players simply choose
at random which move to make, is a Nash equilibrium. In fact it is the only Nash equilibrium
for this game, though we won’t prove this.
It turns out that every zero-sum game has at least one Nash equilibrium if we allow the
players to use mixed strategies. (In fact every game has at least one Nash equilibrium if
we allow mixed strategies, though we won’t get to this general result for a while.) So the
instruction play your half of Nash equilibrium strategies is a strategy that you can follow.

20.2 Surprising Mixed Strategies

Consider the following zero-sum game. Row and Column each have to pick either Side or
Center. If they pick differently, then Row wins, which we’ll represent as a return of 5. If they
both pick Center, then Column wins, which we’ll represent as a return of 0. If they both
pick Side, then Row wins with probability 0.6. In that case Row’s expected return is 3. So
we can represent the game as follows.

Side Center
Side 3 5
Center 5 0

There is no pure Nash equilibrium here. But you might think that Row is best off concen-
trating their attention on Side possibilities, since it lets them have more chance of winning.
You’d be right, but only to an extent. The Nash equilibrium solution is (< 75 Side, 72 Center>,
< 57 Side, 27 Center>). (Exercise: Verify that this is a Nash equilibrium solution.) So even
though the outcomes look a lot better for Row if they play Side, they should play Center

106
with some probability. And conversely, although Column’s best outcome comes with Cen-
ter, Column should in fact play Side quite a bit.
Let’s expand the game a little bit. Imagine that each player doesn’t get to just pick Side,
but split this into Left and Right. Again, Row wins if they don’t pick the same way. So the
game is now more generous to Row. And the table looks like this.

Left Center Right

Left 3 5 5
Center 5 0 5
Right 5 5 3

5
It is a little harder to see, but the Nash solution to this game is (< 12 Left, 16 Center, 12
5
Right>,
5 1 5
< 12 Left, 6 Center, 12 Right>). That is, even though Row could keep Column on their toes
simply by randomly choosing between Left and Right, they do a little better sometimes
playing Center. I’ll leave confirming this as an exercise for you, but if Row played <0.5 Left,
0.5 Right>, then Column could play the same, and Row’s expected return would be 4. But
in this solution, Row’s expected return is a little higher, it is 4 16 .
The above game is based on a study of penalty kicks in soccer that Stephen Levitt (of
Freakonomics fame) did with some colleagues. In a soccer penalty kick, a player, call them
Kicker, stands 12 yards in front of the goal and tries to kick it into the goal. The goalkeeper
stands in the middle of the goal and tries to stop them. At professional level, the ball moves
too quickly for the goalkeeper to see where the ball is going and then react and move to
stop it. Rather, the goalkeeper has to move simultaneously with Kicker. Simplifying a little,
Kicker can aim left, or right, or straight ahead. Simplifying even more, if the goalkeeper
does not guess Kicker’s move, a goal will be scored with high probability. (We’ve made this
probability 1 in the game.) If Kicker aims left or right, and goalkeeper guesses this, there is
still a very good chance a goal will be scored, but the goalkeeper has some chance of stopping
it. And if Kicker aims straight at center, and goalkeeper simply stands centrally, rather than
diving to one side or the other, the ball will certainly not go in.
One of the nice results Levitt’s team found was that, even when we put in more realistic
numbers for the goal-probability than I have used, the Nash equilibrium solution of the
game has Kicker having some probability of kicking straight at Center. And it has some
probability for goalkeeper standing centrally. So there is some probability that the Kicker
will kick the ball straight where the goalkeeper is standing, and the goalkeeper will gratefully
stand there and catch the ball.
This might seem like a crazy result in fact; who would play soccer that way? Well, they
discovered that professional players do just this. Players do really kick straight at the goal-
keeper some of the time, and occasionally the goalkeeper doesn’t dive to the side. And very
occasionally, both of those things happen. (It turns out that when you are more careful with
the numbers, goalkeepers should dive almost all the time, while players should kick straight
reasonably often, and that’s just what happens.) So in at least one high profile game, players
do make the Nash equilibrium play.

107
20.3 Calculating Mixed Strategy Nash Equilibrium
Here is a completely general version of a two-player zero-sum game with just two moves
available for each player.

C1 C2
R1 a b
R2 c d

If one player has a dominating strategy, then they will play that, and the Nash equilibrium
will be the pair consisting of that dominating move and the best move the other player can
make, assuming the first player makes the dominating move. If that doesn’t happen, we can
use the following method to construct a Nash equilbrium. What we’re going to do is to find
a pair of mixed strategies such that for each mixed strategy, if it is made, any strategy the
other player follows has equal probability.
So let’s say that Row plays <pR1 , 1 – pR2 > and Column plays <qC1 and1 – qC2 >. We want
to find values of p and q such that the other player’s expected utility is invariant over their
possible choices. We’ll do this first for Column. Row’s expected return is

Pr(R1 C1 )U(R1 C1 ) + Pr(R1 C2 )U(R1 C2 ) + Pr(R2 C1 )U(R2 C1 ) + Pr(R2 C2 )U(R2 C2 )

=pq × a + p(1 – q) × b + (1 – p)q × c + (1 – p)(1 – q) × d
=pqa + pb – pqb + qc – pqc + d – pd – qd + pqd
=p(qa + b – qb – qc – d + qd) + qc + d – qd

Now our aim is to make that value a constant when p varies. So we have to make qa +
b – qb – qc – d + qd equal 0, and then Row’s expected return will be exactly qc + d – qd. So
we have the following series of equations.

qa + b – qb – qc – d + qd = 0
qa + qd – qb – qc = d – b
q(a + d – (b + c)) = d – b
d–b
q=
a + d – (b + c)
Let’s do the same thing for Row. Again, we’re assuming that there is no pure Nash equi-
librium, and we’re trying to find a mixed equilibrium. And in such a state, whatever Column
does, it won’t change her expected return. Now Column’s expected return is the negation
of Row’s return. So her return is

108
 Pr(R1 C1 )U(R1 C1 ) + Pr(R1 C2 )U(R1 C2 ) + Pr(R2 C1 )U(R2 C1 ) + Pr(R2 C2 )U(R2 C2 )
=pq × –a + p(1 – q) × –b + (1 – p)q × –c + (1 – p)(1 – q) × –d
= – pqa – pb + pqb – qc + pqc – d + pd + qd – pqd
=q(–pa + pb – c + pc + d – pd) – pb – c + d

Again, our aim is to make that value a constant when p varies. So we have to make
–pa+pb–c+pc+d–pd equal 0, and then Column’s expected return will be exactly qc+d–qd.
So we have the following series of equations.

–pa + pb – c + pc + d – pd = 0
–pa + pb + pc – pd = c – d
p(b + c – (a + d)) = c – d
c–d
p=
b + c – (a + d)
c–d b–a
So if Row plays < b+c–(a+d) R1 , b+c–(a+d) R2 >, Column’s expected return is the same what-
d–b a–c
ever she plays. And if Column plays < a+d–(b+c) C1 and a+d–(b+c) C2 >, Row’s expected return is
the same whatever she plays. So that pair of plays forms a Nash equilibrium.

109
 Chapter 21

Nash Equilibrium
21.1 Illustrating Nash Equilibrium
In the previous notes, we worked out what the Nash equilibrium was for a general 2 × 2
zero-sum game with these payoffs.

C1 C2
R1 a b
R2 c d

And we worked out that the Nash equilibrium is where Row and Column play the following
strategies.

c–d b–a
Row plays < R1 , R2 >
b + c – (a + d) b + c – (a + d)
d–b a–c
Column plays < C1 , C2 >
a + d – (b + c) a + d – (b + c)
Let’s see how this works with a particular example. Our task is to find the Nash equilib-
rium for the following game.

C1 C2
R1 1 6
R2 3 2

There is no Nash equilibrium here. Basically Column aims to play the same as what Row
plays, though just how the payouts go depends on just what they select.
3–2 6–1
Row’s part of the Nash equilibrium, according to the formula above, is < 6+3–(2+1) R1 , 6+3–(2+1) R2 >.
1 5
That is, it is < 6 R1 , 6 R2 >. Row’s part of the Nash equilibrium then is to usually play R2 , and
occasionally play R1 , just to stop Column from being sure what Row is playing.
2–6 1–3
Column’s part of the Nash equilibrium, according to the formula above, is < 2+1–(6+3) C1 , 2+1–(6+3) C2 >.
2 1
That is, it is < 3 C1 , 3 C2 >. Column’s part of the Nash equilibrium then is to frequently play
C1 , but often play C1 .

110
 The following example is more complicated. To find the Nash equilibrium, we first elim-
inate dominated options, then apply our formulae for finding mixed strategy Nash equilib-
rium.

C1 C2 C3
R1 1 5 2
R2 3 2 4
R3 0 4 6

Column is trying to minimise the relevant number. So whatever Row plays, it is better for
Column to play C1 than C3 . Equivalently, C1 dominates C3 . So Column won’t play C3 . So
effectively, we’re faced with the following game.

C1 C2
R1 1 5
R2 3 2
R3 0 4

In this game, R1 dominates R3 for Row. Whatever Column plays, Row gets a better return
playing R1 than R3 . So Row won’t play R3 . Effectively, then, we’re faced with this game.

C1 C2
R1 1 5
R2 3 2

And now we can apply the above formulae. When we do, we see that the Nash equilibrium
for this game is with Row playing < 15 R1 , 45 R2 >, and Column playing < 53 C1 , 25 C2 >.

21.2 Why Play Equilibrium Moves?

We’ve spent a lot of time so far on the mathematics of equilibrium solutions to games, but
we haven’t said a lot about the normative significance of these equilibrium solutions. We’ve
occasionally talked as if playing your part of a Nash equilibrium is what you should do. Yet
this is far from obvious.
One reason it isn’t obvious is that often the only equilibrium solution to a game is a mixed
strategy equilibrium. So if you should only play equilibrium solutions, then sometimes you
have to play a mixed strategy. So sometimes, the only rational thing to do is to randomise
your choices. This seems odd. In regular decision problems, we didn’t have any situation
where it was better to play a mixed strategy than any pure strategy.
Indeed, it is hard to conceptualise how a mixed strategy is better than any pure strategy.
The expected return of a mixed strategy is presumably a weighted average of the expected
returns of the pure strategies of which it is made up. That is, if you’re playing a mixed strategy
of the form < 0.6A, 0.4B >, then the expected utility of that strategy looks like it should be

111
0.6 × U(A) + 0.4 × U(B). And that can’t possibly be higher than both U(A) and U(B). So
what’s going on?
We can build up to an argument for playing Nash equilibrium by considering two cases
where it seems to really be the rational thing to do. These cases are

• Repeated plays of a zero-sum game

• When the other person can figure out your strategy

Let’s take these in turn. Consider again Rock-Paper-Scissors. It might be unclear why, in
a one-shot game, it is better to play the mixed strategy < 13 Rock, 13 Paper, 13 Scissors> than to
play any pure strategy, such as say Rock. But it is clear why the mixed strategy will be better
over the long run than the pure strategy Rock. If you just play Rock all the time, then the
other player will eventually figure this out, and play Paper every time and win every time.
In short, if you are playing repeatedly, then it is important to be unpredictable. And
mixed strategies are ideal for being unpredictable. In real-life, this is an excellent reason for
using mixed strategies in zero-sum games. (The penalty kicks study we referred to above
is a study of one such game.) Indeed, we’ve often referred to mixed strategies in ways that
only make sense in long run cases. So we would talk about Row as usually, or frequently, or
occasionally, playing R1 , and we’ve talked about how doing this avoids detection of Row’s
strategy by Column. In a repeated game, that talk makes sense. But Nash equilibrium is also
meant to be relevant to one-off games. So we need another reason to take mixed strategy
equilibrium solutions seriously.
Another case where it seems to make sense to play a mixed strategy is where you have
reason to believe that the other player will figure out your strategy. Perhaps the other player
has spies in your camp, spies who will figure out what strategy you’ll play. If that’s so, then
often a mixed strategy will be best. That’s because, in effect, the other player’s move is not
independent of what strategy you’ll pick. Crucially, it is neither evidentially nor causally
independent of what you do. If that’s so, then the mixed strategy could possibly produce
different results to either mixed strategy, because it will change the probability of the other
player’s move.
Put more formally, the Nash equilibrium move is the best move you can make condi-
tional on the assumption that the other player will know your move before you make their
move. Consider a simple game of ’matching pennies’, where each player puts down a coin,
and Row wins if they are facing the same way (either both Heads or both Tails), and Column
wins if they are facing opposite ways. The game table is

Heads Tails
Heads 1 -1
Tails -1 1

The equilibrium solution to this game is for each player to play < 0.5 Heads, 0.5 Tails >. In
other words, the equilibrium thing to do with your coin is to flip it. And if the other player
knows what you’ll do with your coin, that’s clearly the right thing to do. If Row plays Heads,
Column will play Tails and win. If Row plays Tails, Column will play Heads and win. But if
Row flips their coin, Column can’t guarantee a win.

112
 Now in reality, most times you are playing a game, there isn’t any such spy around. But
the other player may not need a spy. They might simply be able to guess, or predict, what
you’ll do. So if you play a pure strategy, there is reason to suspect that the other player will
figure out that you’ll play that strategy. And if you play a mixed strategy, the other player
will figure out this as well. Again, assuming the other player will make the optimal move in
response to your strategy, the mixed strategy may well be best.
Here’s why this is relevant to actual games. We typically assume in game theory that each
player is rational, that each player knows the other player is rational, and so on. So the other
player can perfectly simulate what you do. That’s because they, as a rational person, knows
how a rational person thinks. So if it is rational you to pick strategy S, the other player will
predict that you’ll pick strategy S. And you’ll pick strategy S if and only if it is rational to do
so. Putting those last two conditionals together, we get the conclusion that the other player
will predict whatever strategy you play.
And with that comes the justification for playing Nash equilibrium moves. Given our
assumptions about rationality, we should assume that the other player will predict our strat-
egy. And conditional on the other player playing the best response to our strategy, whatever
it is, the Nash equilibrium play has the highest expected utility. So we should make Nash
equilibrium plays.

21.3 Causal Decision Theory and Game Theory

In the last section we gave what is essentially the orthodox argument for playing equilib-
rium plays. The other player is as rational as you, so the other player can figure out the
rational play, i.e. what you’ll play. So you should make the play that returns the highest
result conditional on the other player figuring out that you’ll play it.
This kind of reasoning might be familiar. It is the reasoning that leads to taking one-box
in Newcomb’s problem. If we think that we are perfectly rational players facing Newcomb’s
problem, then we should think that the demon can predict what we’ll do by simply applying
her own rationality. So the demon will predict our play. So we should make the move that
has the highest expected utility conditional on it being predicted by the demon. And that’s
to take one-box. Conditional on the demon figuring out what we’ll do, taking one-box leads
to $1,000,00 reward, and taking both leads to a $1,000 reward. So we should take one-box.
But not everyone agrees with this conclusion. Some people are causal decision theorists,
not evidential decision theorists. They think that if the demon is merely predicting what
we will do, then it is wrong to conditionalise on the assumption that the demon will be
correct. That’s because our actions could at best be evidence for what the demon predicted;
they couldn’t cause what the demon predicted. So the demon’s predictions are effectively
states of the world; they are causally independent of our choices. And then applying causal
decision theory recommends taking both boxes.
The causal decision theorist will think that the argument from the previous section con-
tained an important, but illegitimate, move. The story we told about the case where there
was a spy in our ranks made sense. If there is a spy, then what we do causes the moves of the
other player. So the other player’s move isn’t an antecedently obtaining state of the world
in the relevant sense. But when we drop the spy, and assume that the other player is merely
predicting what we will do, then their choice really is a causally independent state of the
world. So our selection of a pure strategy doesn’t cause the other person’s moves to change,

113
though they may well be evidence that the other person’s moves will be different to what we
thought they would be.
The core idea behind causal decision theory is that it is illegitimate to conditionalise on
our actual choice when working out the probability of various states of the world. We should
work out the probability of the different states, and take those as inputs to our expected
utility calculations. But to give a high probability to the hypothesis that our choice will be
predicted, whatever it is, is to not use one probability for each possible state of the world.
And that’s what both the expected utility theorist does, and what the game theorist who
offers the above defence of equilibrium plays does.
There’s an interesting theoretical point here. The use of equilibrium reasoning is en-
demic in game theory. But the standard justification of equiilbrium strategies relies on one
of the two big theories of decision making, namely evidential decision theory. And that’s
not even the more popular of the two models of decision making. We’ll come back to this
point a little as we go along.
In practice this is a little different to in theory. Most games in real-life are repeat games,
and in repeat games the difference between causal and evidential decision theory is less than
in one-shot games. If you were to play Newcomb’s Problem many times, you may well be best
off picking one-box on the early plays to get the demon to think you are a one-box player.
But to think through cases like this one more seriously we need to look at the distinctive
features of games involving more than one move, and that’s what we’ll do next.

114
 Chapter 22

Many Move Games

22.1 Games with Multiple Moves
Most real life games have more than one move in them. The players in chess, for instance,
do not just make one simultaneous move and then stop. In fact, games like chess differ from
the simple games we’ve been studying in two respects. First, the players make more than
one move. Second, the players do not move simultaneously.
To a first approximation, those differences might be different than they first appear. We
can imagine two super-duper-computers playing chess as follows. Each of them announces,
simultaneously, their strategies for the complete game. A strategy here is a decision about
what to do at any stage the game might come to. The ‘play’ of the game would then consist in
moving the various pieces around in accord with the various strategies the computers laid
out.
Of course, this is completely impractical. Even the best of modern computers can’t deal
with all the possible positions that might come up on a chess board. What they have to do,
like what we do, is to look at the positions that actually arise and deal with them when they
come up. But if we’re abstracting away from computational costs (as we are throughout)
the difference between chess as it actually is (with turn-by-turn moves) and ‘strategy chess’
looks a little smaller.

22.2 Extensive and Normal Form

What we’ve noted so far is that there are two ways to ‘play’ a many move game. We can
wait and watch the moves get played. Or we can have each player announce their strategy
at the start of the game. Somewhat reflecting these two ways of thinking about games, there
are two ways of representing many move games. First, we can represent them in extensive
form. The following is an extensive form representation of a zero-sum game.
Each node in the chart represents a move that a player makes. The nodes are marked
with the name of the player who moves at that point. So in this game, Row moves first,
then Column moves, then the game is done. The numbers at the end represent the payoffs.
Eventually there will be two numbers there, but for now we’re still in the realm of zero-sum
games, so we’re just using a single number.
In this game, Row plays first and has to choose between L and R. Then Column plays,
and the choices Column has depend on what move Row made. If Row played L, then Col-
umn could choose between a and b. (And presumably would choose b, since Column is
trying to minimise the number.) If Row played R, then Column could choose between c

115
 Rowb
 HH
L R
  Column HHH
r
p p p p p p p p p p p p p p p p p p pH
p r
a @ b c @ d
@ @
r @r r @r
4 3 2 1

Figure 22.1: An extensive game

and d. (And presumably would choose d, since Column is trying to minimise the num-
ber.) If we assume Column will make the rational play, then Row’s choice is really between
getting 3, if she plays L, and 1, if she plays R, so she should play L.
As well as this extensive form representation, we can also represent the game in normal
form. A normal form representation is where we set out each player’s possible strategies for
the game. As above, a strategy is a decision about what to do in every possibility that may
arise. Since Row only has one move in this game, her strategy is just a matter of that first
move. But Column’s strategy has to specify two things: what to do if Row plays L, and what
to do if Row plays R. We’ll represent a strategy for Column with two letters, e.g., ac. That’s
the strategy of playing a if Row plays L and c if Row plays R. The normal form of this game
is then

ac ad bc bd
L 4 4 3 3
R 2 1 2 1

22.3 Two Types of Equilibrium

In the game we’ve been looking at above, there are two Nash equilibrium outcomes. The first
is < L, bc >, and the second is < L, bd >. Both of these end up with a payoff of 3. But there is
something odd about the first equilibrium. In that equilibrium, Column has a strategy that
embeds some odd dispositions. If Row (foolishly) plays R, then Column’s strategy says to
(equally foolishly) play c. But clearly the best play for Column in this circumstance is d, not
c.
So in a sense, < L, bc > is not an equilibrium strategy. True, it is as good as any strat-
egy that Column can follow given Row’s other choices. But it isn’t an optimal strategy for
Column to follow with respect to every decision that Column has to make.
We’ll say a subgame perfect equilibrium is a pair of strategies for Row and Column
such that for any given node in the game, from that node on, neither can do better given the
other’s strategy. A Nash equilibrium satisfies this condition for the ‘initial’ node; subgame
perfect equilibrium requires that it be satisfied for all nodes.

116
22.4 Normative Significance of Subgame Perfect Equilibrium
Subgame perfect equilibrium is a very significant concept in modern game theory. Some
writers take it to be an important restriction on rational action that players play strategies
which are part of subgame perfect equilibria. But it is a rather odd concept for a few reasons.
We’ll say more about this after we stop restricting our attention to zero-sum games, but for
now, consider the game in Figure 22.2. (I’ve used R and C for Row and Column to save
space. Again, it’s a zero-sum game. And the initial node in these games is always the open
circle; the closed circles are nodes that we may or may not get to.)

Rb r Cr r Rr r r1

d d d
r r r
3 2 4

Figure 22.2: Illustrating subgame perfect equilibrium

Note that Row’s strategy has to include two choices: what to do at the first node, and
what to do at the third node. But Column has (at most) one choice. Note also that the game
ends as soon as any player plays d. The game continues as long as players are playing r, until
there are 3 plays of r.
We can work out the subgame perfect equilibrium by backwards induction from the
terminal nodes of the game. At the final node, the dominating option for Row is d, so Row
should play d. Given that Row is going to play d at that final choice-point, and hence end
the game with 4, Column is better off playing d at her one and only choice, and ending
the game with 2 rather than the 4 it would end with if Row was allowed to play last. And
given that that’s what Column is planning to do, Row is better off ending the game straight
away by playing d at the very first opportunity, and ending with 3. So the subgame perfect
equilibrium is < dd, d >.
There are three oddities about this game.
First, if Row plays d straight away, then the game is over and it doesn’t matter what the
rest of the strategies are. So there are many Nash equilibria for this game. That implies that
there are Nash equilibria that are not subgame perfect equilibria. For instance, < dr, d > is a
Nash equilibria, but isn’t subgame perfect. That’s not, however, something we haven’t seen
before.
Second, the reason that < dr, d > is not an equilibrium is that it is an irrational thing for
Row to play if the game were to get to the third node. But if Row plays that very strategy,
then the game won’t get to that third node. Oddly, Row is being criticised here for playing
a strategy that could, in principle, have a bad outcome, but will only have a bad outcome if
she doesn’t play that very strategy. So it isn’t clear that her strategy is so bad.
Finally, let’s think again about Column’s option at the middle node. We worked out what
Column should do by working backwards. But the game is played forwards. And if we reach
that second node, where Column is playing, then Column knows that Row is not playing
an equilibrium strategy. Given that Column knows this, perhaps it isn’t altogether obvious

117
that Column should hold onto the assumption that Row is perfectly rational. But without
the assumption that Row is perfectly rational, then it isn’t obvious that Column should play
d. After all, that’s only the best move on the assumption that Row is rational.
The philosophical points here are rather tricky, and we’ll come back to them when we’ve
looked more closely at non zero sum games.

22.5 Cooperative Games

As we’ve stressed several times, most human interactions are not zero sum. Most of the
time, there is some opportunity for the players’ interests to be aligned. This is so even when
we look at games involving one (simultaneous) move.
We won’t prove this, but it turns out that even when we drop the restriction to zero-
sum games, every game has a Nash equilibrium. Sometimes this will be a mixed strategy
equilibrium, but often it will be a pure strategy. What is surprising about non-zero sum
games is that it is possible for there to be multiple Nash equilibria that are not equal in their
outcomes. For instance, consider the following game.

C1 C2
R1 (4, 1) (0, 0)
R2 (0, 0) (2, 2)

Both (R1 , C1 ) and (R2 , C2 ) are Nash equilibria. I won’t prove this, but there is also a mixed
strategy Nash equilibria, (< 23 R1 , 13 R2 >, < 13 C1 , 23 C2 >). This is an incredibly inefficient
Nash equilibrium, since the players end up with the (0, 0) outcome most of the time. But
given that that’s what the other player is playing, they can’t do better.
The players are not indifferent over these three equilibria. Row would prefer the (R1 , C1 )
equilibrium, and Column would prefer the (R2 , C2 ) equilibrium. The mixed equilibrium is
the worst outcome for both of them. Unlike in the zero-sum case, it does matter which
equilibrium we end up at. Unfortunately, in the absence the possibility for negotiation, it
isn’t clear what advice game theory can give about cases like this one, apart from saying that
the players should play their part in some equilibrium play or other.

22.6 Pareto Efficient Outcomes

In game theory, and in economics generally, we say that one outcome O1 is Pareto superior
to another O2 if at least one person is better off in O1 than in O2 , and no one is worse off
in O1 than O2 . O2 is Pareto inferior to O1 iff O1 is Pareto superior to O2 . An outcome is
Pareto inefficient if there is some outcome that is Pareto superior to it. And an outcome is
Pareto efficient otherwise.
Some games have multiple equilibria where one equilibrium outcome is Pareto superior
to another. We’ve already seen one example of this with the previous game. In that game,
there was a mixed strategy equilibrium that was worse for both players than either pure
strategy equilibrium. But there are considerably simpler cases of the same phenomenon.

118
 C1 C2
R1 (2, 2) (0, 0)
R2 (0, 0) (1, 1)

In this case, the (R1 , C1 ) outcome is clearly superior to the (R2 , C2 ) outcome. (There’s also a
mixed strategy equilibrium that is worse again for both players.) And it would be surprising
if the players ended up with anything other than the (R1 , C1 ) outcome.
It might be tempting at this point to add an extra rule to the Only choose equilibrium
strategies rule, namely Never choose an equilibrium that is Pareto inefficient. Unfortunately,
that won’t always work. In one famous game, the Prisoners Dilemma, the only equilibrium
is Pareto inefficient. Here is a version of the Prisoners Dilemma.

C1 C2
R1 (3, 3) (5, 0)
R2 (0, 5) (1, 1)

The (R2 , C2 ) outcome is Pareto inferior to the (R1 , C1 ) outcome. But the (R2 , C2 ) outcome
is the only equilibrium. Indeed, (R2 , C2 ) is the outcome we get to if both players simply
eliminate dominated options. Whatever the other player does, each player is better off play-
ing their half of (R2 , C2 ). So equilibrium seeking not only fails to avoid Pareto inefficient
options; sometimes it actively seeks out Pareto inefficiencies.

22.7 Exercises
22.7.1 Nash Equilibrium
Find the Nash equilibrium in each of the following zero-sum games.

C1 C2
R1 4 6
R2 3 7

C1 C2
R1 4 3
R2 3 7

C1 C2 C3
R1 3 2 4
R2 1 5 3
R3 0 1 6

119
22.7.2 Subgame Perfect Equilibrium
In the following game, which pairs of strategies form a Nash equilibrium? Which pairs
form a subgame perfect equilibrium? In each node, the first number represents R’s payoff,
the second represents C’s payoff. Remember that a strategy for each player has to specify
what they would do at each node they could possibly come to.

Row b
 HH
L  HHR
  HH

r
p p p p p p p p p p Column
p p p p p p p p p p p p p pHp rH
a HH c d  e HH
f
  b H H   H
r r Hr r r Hr
(25, 25) (20, 30) (15, 35) (30, 0) (25, 5) (20, 10)

Figure 22.3: Extensive Game for question 2

22.7.3 Equality of Nash Equilibria

In a particular zero-sum game, (R1 , C1 ) and (R2 , C2 ) are Nash equilibria. Prove that (a) both
(R1 , C2 ) and (R2 , C1 ) are Nash equilibria, and (b) (R1 , C1 ) and (R2 , C2 ) have the same payoffs.

120
 Chapter 23

Backwards Induction
23.1 Puzzles About Backwards Induction
In the previous notes, we showed that one way to work out the subgame perfect equilibrium
for a strategic game is by backwards induction. The idea is that we find the Nash equilib-
rium for the terminal nodes, then we work out the best move at the ‘penultimate’ nodes by
working out the best plays for each player assuming a Nash equilibrium play will be made
at the terminal nodes. Then we work out the best play at the third-last node by working out
the best thing to do assuming players will make the rational play at the last two nodes, and
so on until we get back to the start of the game.
The method, which we’ll call backwards induction, is easy enough in practice to im-
plement. And the rational seems sound at first glance. It is reasonable to assume that the
players will make rational moves at the end of the game, and that earlier moves should be
made predicated on our best guesses of later moves. So it seems sensible enough to use
backwards induction.
But it leads to crazy results in a few cases. Consider, for example, the centipede game.
I’ve done a small version of it here, where each player has (up to) 7 moves. You should be
able to see the pattern, and imagine a version of the game where each player has 50, or for
that matter, 50,000 possible moves.

Rb r Cr r Rr r Cr r Rr r Cr r Rr r Cr r Rr r Cr r Rr r Cr r Rr r Cr r r (8,6)
d d d d d d d d d d d d d d
r r r r r r r r r r r r r r
(1,0) (0,2) (2,1) (1,3) (3,2) (2,4) (4,3) (3,5) (5,4) (4,6) (6,5) (5,7) (7,6) (6,8)

Figure 23.1: Centipede Game

At each node, players have a choice between playing d, which will end the game, and playing
r, which will (usually) continue it. At the last node, the game will end whatever Column
plays. The longer the game goes, the larger the ‘pot’, i.e. the combined payouts to the two
players. But whoever plays d and ends the game gets a slightly larger than average share of
the pot. Let’s see how that works out in practice.
If the game gets to the terminal node, Column will have a choice between 8 (if she plays
d) and 6 (if she plays r). Since she prefers 8 to 6, she should play d and get the 8. If we

121
assume that Column will play d at the terminal node, then at the penultimate node, Row
has a choice between playing d, and ending the game with 7, or playing r, and hence, after
Column plays d, ending the game with 6. Since she prefers getting 7 to 6, she should play
d at this point. If we assume Row will play d at the second last node, leaving Column with
6, then Column is better off playing d at the third last node and getting 7. And so on. At
every node, if you assume the other player will play d at the next node, if it arrives, then the
player who is moving has a reason to play d now and end the game. So working backwards
from the end of the game, we conclude that Row should play d at the first position, and end
the game.
A similar situation arises in a repeated Prisoners Dilemma. Here is a basic version of a
Prisoners Dilemma game.

Coop Rat
Coop (3, 3) (5, 0)
Rat (0, 5) (1, 1)

Imagine that Row and Column have to play this game 100 times in a row. There might be
some incentive here to play Coop in the early rounds, if it will encourage the other player to
play Coop in later rounds. Of course, neither player wants to be a sucker, but it seems plau-
sible to think that there might be some benefit to playing ‘Tit-For-Tat’. This is the strategy
of playing Coop on the first round, then on subsequent rounds playing whatever the other
player had played on previous rounds.
There is some empirical evidence that this is the rational thing to do. In the late 1970s a
political scientist, Robert Axelrod, set up just this game, and asked people to send in com-
puter programs with strategies for how to play each round. Some people wrote quite sophis-
ticated programs that were designed to trigger general cooperation, but also occasionally
exploit the other player by playing Rat occasionally. Axelrod had all of the strategies sent
in play ‘against’ each other, and added up the points each got. Despite the sophistication
of some of the submitted strategies, it turned out that the most successful one was simply
Tit-For-Tat. After writing up the results of this experiment, Axelrod ran the experiment
again, this time with more players because of the greater prominence he’d received from the
first experiment. And Tit-For-Tat won again. (There was one other difference in the second
version of the game that was important to us, and which we’ll get to below.)
But backwards induction suggests that the best thing to do is always to Rat. The rational
thing for each player to do in the final game is Rat. That’s true whatever the players have
done in earlier games, and whatever signals have been sent, tacit agreements formed etc.
The players, we are imagining, can’t communicate except through their moves, so there is
no chance of an explicit agreement forming. But by playing cooperatively, they might in
effect form a pact. But that can have no relevance to their choices on the final game, where
they should both Rat.
And if they should both play Rat on the final game, then there can’t be a strategic benefit
from playing Coop on the second last game, since whatever they do, they will both play Rat
on the last game. And whenever there is no strategic benefit from playing Coop, the rational
thing to do is to play Rat, so they will both play Rat on the second last game.

122
 And if they should both play Rat on the second last game, whatever has happened before,
then similar reasoning shows that they should both play Rat on the third last game, and
hence on the fourth last game, and so on. So they should both play Rat on every game.
This is, to say the least, extremely counterintuitive. It isn’t just that playing Coop in
earlier rounds is Pareto superior to playing Rat. After all, each playing Coop on the final
round is Pareto superior to playing Rat. It is that it is very plausible that each player has more
to gain by trying to set up tacit agreements to cooperate than they have to lose by playing
Coop on a particular round.
It would be useful to have Axelrod’s experiments to back this up, but they aren’t quite as
good evidence as we might like. His first experiment was exactly of this form, and Tit-For-
Tat did win (with always Rat finishing in last place). But the more thorough experiment,
with more players, was not quite of this form. So as to avoid complications about the back-
wards induction argument, Axelrod made the second game a repeated Prisoners Dilemma
with a randomised end point. Without common knowledge of the end point, the backwards
induction argument doesn’t get off the ground.
Still, it seems highly implausible that rationality requires us to play Rat at every stage, or
to play d at every stage in the Centipede game. In the next section we’ll look at an argument
by Philip Pettit and Robert Sugden that suggests this is not a requirement of rationality.

23.2 Pettit and Sugden

The backwards induction reasoning is, as the name suggests, from the back of the game
to the front of it. But games are played from front to back, and we should check how the
reasoning looks from this perspective. For simplicity, we’ll work with Centipede, though
what we say should carry over to finitely iterated Prisoners Dilemma as well.
First, imagine that one of the players does not play the subgame perfect equilibrium
solution to the game. So imagine that Row plays r at the first move. Now Column has a
choice to make at the second node. Column knows that if Row plays the subgame perfect
equilibrium at the third move, then the best thing for her to do now is to play d. And
Column presupposed, at the start of the game, that Row was rational. And we’re supposing,
so far, that rational players play subgame perfect equilibrium plays. So Column should play
d right?
Not necessarily. At the start of the game, Column assumed that Row was a rational
player. But Row has given her evidence, irrefutable evidence given our assumption that
rationality equals making subgame perfect equilibrium plays, that she is not rational. And
it isn’t at all clear that the right thing to do when playing with a less than rational player is
to play d. If there is even a small chance that they will keep playing r, then it is probably
worthwhile to give them the chance to do so.
That’s all to say, given the assumptions that we made, if Row plays r, Column might well
reciprocate by playing r. But if that’s true, then there is no reason for Row to play d at the
start. The argument that Row should play d turned on the assumption that if she played
r, Column would play d. And given various assumptions we made at the start of the game,
Column would have played d. But, and here is the crucial point, if Column were in a position
to make a move at all, those assumptions would no longer still be operational. So perhaps
it is rational for Row to play r.

123
 None of this is to say that Row should play r on her last move. After all, whatever Column
thinks about Row’s rationality, Column will play d on the last move, so Row should play d
if it gets to her last move. It isn’t even clear that it gives Row or Column a reason to play r
on their second last moves, since even then it isn’t clear there is a strategic benefit to be had.
But it might given them a reason to play r on earlier moves, as was intuitively plausible.
There is something that might seem odd about this whole line of reasoning. We started
off saying that the uniquely rational option was to play d everywhere. We then said that if
Row played r, Column wouldn’t think that Row was rational, so all bets were off with respect
to backwards induction reasoning. So it might be sensible for Row to play r. Now you might
worry that if all that’s true, then when Row plays r, that won’t be a sign that Row is irrational.
Indeed, it will be a sign that Row is completely rational! So how can Pettit and Sugden argue
that Column won’t play d at the second node?
Well, if their reasoning is right that r is a rational move at the initial node, then it is also
good reasoning that Column can play r at the second node. Either playing r early in the game
is rational or it isn’t. If it is, then both players can play r for a while as a rational resolution
of the game. If it isn’t, then Row can play r as a way of signaling that she is irrational, and
hence Column has some reason to play r. Either way, the players can keep on playing r.
The upshot of this is that backwards induction reasoning is less impressive than it looked
at first.

124
 Chapter 24

Group Decisions
So far, we’ve been looking at the way that an individual may make a decision. In practice, we
are just as often concerned with group decisions as with individual decisions. These range
from relatively trivial concerns (e.g. Which movie shall we see tonight?) to some of the
most important decisions we collectively make (e.g. Who shall be the next President?). So
methods for grouping individual judgments into a group decision seem important.
Unfortunately, it turns out that there are several challenges facing any attempt to merge
preferences into a single decision. In this chapter, we’ll look at various approaches that
different groups take to form decisions, and how these different methods may lead to dif-
ferent results. The different methods have different strengths and, importantly, different
weaknesses. We might hope that there would be a method with none of these weaknesses.
Unfortunately, this turns out to be impossible.
One of the most important results in modern decision theory is the Arrow Impossi-
bility Theorem, named after the economist Kenneth Arrow who discovered it. The Arrow
Impossibility Theorem says that there is no method for making group decisions that satisfies
a certain, relatively small, list of desiderata. The next chapter will set out the theorem, and
explore a little what those constraints are.
Finally, we’ll look a bit at real world voting systems, and their different strengths and
weaknesses. Different democracies use quite different voting systems to determine the win-
ner of an election. (Indeed, within the United States there is an interesting range of systems
used.) And some theorists have promoted the use of yet other systems than are currently
used. Choosing a voting system is not quite like choosing a method for making a group
decision. For the next two chapters, when we’re looking at ways to aggregate individual
preferences into a group decision, we’ll assume that we have clear access to the preferences
of individual agents. A voting system is not meant to tally preferences into a decision, it is
meant to tally votes. And voters may have reasons (some induced by the system itself) for
voting in ways other than their preferences. For instance, many voters in American presi-
dential elections vote for their preferred candidate of the two major candidates, rather than
‘waste’ their vote on a third party candidate.
For now we’ll put those problems to one side, and assume that members of the group
express themselves honestly when voting. Still, it turns out there are complications that arise
for even relatively simple decisions.

125
24.1 Making a Decision
Seven friends, who we’ll imaginatively name F1 , F2 , ..., F7 are trying to decide which restau-
rant to go to. They have four options, which we’ll also imaginatively name R1 , R2 , R3 , R4 . The
first thing they do is ask which restaurant each person prefers. The results are as follows.

• F1 , F2 and F3 all vote for R1 , so it gets 3 votes

• F4 and F5 both vote for R2 , so it gets 2 votes
• F6 votes for R3 , so it gets 1 vote
• F7 votes for R4 , so it gets 1 vote

It looks like R1 should be the choice then. It, after all, has the most votes. It has a ‘plu-
rality’ of the votes - that is, it has the most votes. In most American elections, the candidate
with a plurality wins. This is sometimes known as plurality voting, or (for unclear reasons)
first-past-the-post or winner-take-all. The obvious advantage of such a system is that it is
easy enough to implement.
But it isn’t clear that it is the ideal system to use. Only 3 of the 7 friends wanted to go to
R1 . Possibly the other friends are all strongly opposed to this particular restaurant. It seems
unhappy to choose a restaurant that a majority is strongly opposed to, especially if this is
avoidable.
So the second thing the friends do is hold a ‘runoff ’ election. This is the method used for
voting in some U.S. states (most prominently in Georgia and Louisiana) and many European
countries. The idea is that if no candidate (or in this case no restaurant) gets a majority of
the vote, then there is a second vote, held just between the top two vote getters. (Such a
runoff election is scheduled for December 3 in Georgia to determine the next United States
Senator.) Since R1 and R2 were the top vote getters, the choice will just be between those
two. When this vote is held the results are as follows.

• F1 , F2 and F3 all vote for R1 , so it gets 3 votes

• F4 , F5 , F6 and F7 all vote for R2 , so it gets 4 votes

This is sometimes called ‘runoff ’ voting, for the natural reason that there is a runoff.
Now we’ve at least arrived at a result that the majority may not have as their first choice, but
which a majority are at least happy to vote for.
But both of these voting systems seem to put a lot of weight on the various friends’ first
preferences, and less weight on how they rank options that aren’t optimal for them. There
are a couple of notable systems that allow for these later preferences to count. For instance,
here is how the polls in American college sports work. A number of voters rank the best
teams from 1 to n, for some salient n in the relevant sport. Each team then gets a number
of points per ballot, depending on where it is ranked, with n points for being ranked first,
n – 1 points for being ranked second, n – 2 points for being ranked third, and so on down
to 1 point for being ranked n’th. The teams’ overall ranking is then determined by who has
the most points.
In the college sport polls, the voters don’t rank every team, only the top n, but we can
imagine doing just that. So let’s have each of our friends rank the restaurants in order, and
we’ll give 4 points to each restaurant that is ranked first, 3 to each second place, etc. The
points that each friend awards are given by the following table.

126
 F1 F2 F3 F4 F5 F6 F7 Total
R1 4 4 4 1 1 1 1 16
R2 1 3 3 4 4 2 2 19
R3 3 2 2 3 3 4 3 20
R4 2 1 1 2 2 3 4 15

Now we have yet a different choice. By this method, R3 comes out as the best option. This
voting method is sometimes called the Borda count. The nice advantage of it is that it lets
all preferences, not just first preferences, count. Note that previously we didn’t look at all
at the preferences of the first three friends, beside noting that R1 is their first choice. Note
also that R3 is no one’s least favourite option, and is many people’s second best choice. These
seem to make it a decent choice for the group, and it is these facts that the Borda count is
picking up on.
But there is something odd about the Borda count. Sometimes when we prefer one
restaurant to another, we prefer it by just a little. Other times, the first is exactly what we
want, and the second is, by our lights, terrible. The Borda count tries to approximately
measure this - if X strongly prefers A to B, then often there will be many choices between A
and B, so A will get many more points on X’s ballot. But this is not necessary. It is possible
to have a strong preference for A over B without there being any live option that is ‘between’
them. In any case, why try to come up with some proxy for strength of preference when we
can measure it directly?
That’s what happens if we use ‘range voting’. Under this method, we get each voter to give
each option a score, say a number between 0 and 10, and then add up all the scores. This
is, approximately, what’s used in various sporting competitions that involve judges, such
as gymnastics or diving. In those sports there is often some provision for eliminating the
extreme scores, but we won’t be borrowing that feature of the system. Instead, we’ll just get
each friend to give each restaurant a score out of 10, and add up the scores. Here is how the
numbers fall out.

F1 F2 F3 F4 F5 F6 F7 Total
R1 10 10 10 5 5 5 0 45
R2 7 9 9 10 10 7 1 53
R3 9 8 8 9 9 10 2 55
R4 8 7 7 8 8 9 10 57

Now R4 is the choice! But note that the friends’ individual preferences have not changed
throughout. The way each friend would have voted in the previous ‘elections’ is entirely
determined by their scores as given in this table. But using four different methods for ag-
gregating preferences, we ended up with four different decisions for where to go for dinner.
I’ve been assuming so far that the friends are accurately expressing their opinions. If the
votes came in just like this though, some of them might wonder whether this is really the
case. After all, F7 seems to have had an outsized effect on the overall result here. We’ll come
back to this when looking at options for voting systems.

127
24.2 Desiderata for Preference Aggregation Mechanisms
None of the four methods we used so far are obviously crazy. But they lead to four different
results. Which of these, if any, is the correct result? Put another way, what is the ideal
method for aggregating preferences? One natural way to answer this question is to think
about some desirable features of aggregation methods. We’ll then look at which systems
have the most such features, or ideally have all of them.
One feature we’d like is that each option has a chance of being chosen. It would be a very
bad preference aggregation method that didn’t give any possibility to, say, R3 being chosen.
More strongly, it would be bad if the aggregation method chose an option X when there
was another option Y that everyone preferred to X. Using some terminology from the game
theory notes, we can express this constraint by saying our method should never choose a
Pareto inferior option. Call this the Pareto condition.
We might try for an even stronger constraint. Some of the time, not always but some
of the time, there will be an option C such than a majority of voters prefers C to X, for
every alternative X. That is, in a two-way match-up between C and any other option X, C
will get more votes. Such an option is sometimes called a Condorcet option, after Marie
Jean Antoine Nicolas Caritat, the Marquis de Condorcet, who discussed such options. The
Condorcet condition on aggregation methods is that a Condorcet option always comes first,
if such an option exists.
Moving away from these comparative norms, we might also want our preference ag-
gregation system to be fair to everyone. A method that said F2 is the dictator, and F2 ’s
preferences are the group’s preferences, would deliver a clear answer, but does not seem to
be particularly fair to the group. There should be no dictators; for any person, it is possible
that the group’s decision does not match up with their preference.
More generally than that, we might restrict attention to preference aggregation systems
that don’t pay attention to who has various preferences, just to what preferences people have.
Here’s one way of stating this formally. Assume that two members of the group, v1 and v2 ,
swap preferences, so v1 ’s new preference ordering is v2 ’s old preference ordering and vice
versa. This shouldn’t change what the group’s decision is, since from a group level, nothing
has changed. Call this the symmetry condition.
Finally, we might want to impose a condition that we said is a condition we imposed
on independent agents: the irrelevance of independent alternatives. If the group would
choose A when the options are A and B, then they wouldn’t choose B out of any larger set of
options that also include A. More generally, adding options can change the group’s choice,
but only to one of the new options.

24.3 Assessing Plurality Voting

It is perhaps a little disturbing to think how few of those conditions are met by plurality
voting, which is how Presidents of the USA are elected. Plurality voting clearly satisfies
the Pareto condition. If everyone prefers A to B, then B will get no votes, and so won’t
win. So far so good. And since any one person might be the only person who votes for
their preferred candidate, and since other candidates might get more than one vote, no one
person can dictate who wins. So it satisfies no dictators. Finally, since the system only looks
at votes, and not at who cast them, it satisfies symmetry.

128
 But it does not satisfy the Condorcet condition. Consider an election with three candi-
dates. A gets 40% of the vote, B gets 35% of the vote, and C gets 25% of the vote. A wins, and
C doesn’t even finish second. But assume also that everyone who didn’t vote for C has her as
their second preference after either A or B. Something like this may happen if, for instance,
C is an independent moderate, and A and B are doctrinaire candidates from the major par-
ties. Then 60% prefer C to A, and 65% prefer C to B. So C is a Condorcet candidate, yet is
not elected.
A similar example shows that the system does not satisfy the irrelevance of independent
alternatives condition. If B was not running, then presumably A would still have 40% of
the vote, while C would have 60% of the vote, and would win. One thing you might want to
think about is how many elections in recent times would have had the outcome changed by
eliminating (or adding) unsuccessful candidates in this way.

129
 Chapter 25

Arrow’s Theorem
25.1 Ranking Functions
The purpose of this chapter is to set out Arrow’s Theorem, and its implications for the con-
struction of group preferences from individual preferences. We’ll also say a little about the
implications of the theorem for the design of voting systems, though we’ll leave most of that
to the next chapter.
The theorem is a mathematical result, and needs careful setup. We’ll assume that each
agent has a complete and transitive preference ordering over the options. If we say A >V B
means that V prefers A to B, that A =V B means that V is indifferent between A and B, and
that A ≥V B means that A >V B ∨ A =V B, then these constraints can be expressed as
follows.

Completeness For any voter V and options A, B, either A ≥V B or B ≥V A

Transitivity For any voter V and options A, B, the following three conditions hold:
• If A >V B and B >V C then A >V C
• If A =V B and B =V C then A =V C
• If A ≥V B and B ≥V C then A ≥V C

More generally, we assume the substitutivity of indifferent options. That is, if A =V B, then
whatever is true of the agent’s attitude towards A is also true of the agent’s attitude towards
B. In particular, whatever comparison holds in the agent’s mind between A and C holds
between B and C. (The last two bullet points under transitivity follow from this principle
about indifference and the earlier bullet point.)
The effect of these assumptions is that we can represent the agent’s preferences by lining
up the options from best to worst, with the possibility that we’ll have to put two options in
one ‘spot’ to represent the fact that the agent values each of them equally.
A ranking function is a function from the preference orderings of the agent to a new
preference ordering, which we’ll call the preference ordering of the group. We’ll use the
subscript G to note that it is the group’s ordering we are designing. We’ll also assume that
the group’s preference ordering is complete and transitive.
There are any number ranking functions that don’t look at all like the group’s preferences
in any way. For instance, if the function is meant to work out the results of an election, we
could consider the function that takes any input whatsoever, and returns a ranking that
simply lists by age, with the oldest first, the second oldest second, etc. This doesn’t seem

130
like it is the group’s preferences in any way. Whatever any member of the group thinks, the
oldest candidate wins. What Arrow called the citizen sovereignty condition is that for any
possible ranking, it should be possible to have the group end up with that ranking.
The citizen sovereignty follows from another constraint we might put on ranking func-
tions. If everyone in the group prefers A to B, then A >G B, i.e. the group prefers A to B.
We’ll call this the Pareto constraint. It is sometimes called the unanimity constraint, but
we’ll call it the Pareto condition.
One way to satisfy the Pareto constraint is to pick a particular person, and make them
dictator. That is, the function ‘selects’ a person V, and says that A >G B if and only if A >V B.
If everyone prefers A to B, then V will, so this is consistent with the Pareto constraint. But
it also doesn’t seem like a way of constructing the group’s preferences. So let’s say that we’d
like a non-dictatorial ranking function.
The last constraint is one we discussed in the previous chapter: the independence of
irrelevant alternatives. Formally, this means that whether A >G B is true depends only on
how the voters rank A and B. So changing how the voters rank, say B and C, doesn’t change
what the group says about the A, B comparison.
It’s sometimes thought that it would be a very good thing if the voting system respected
this constraint. Let’s say that you believe that if Ralph Nader had not been a candidate in
the 2000 U.S. Presidential election, then Al Gore, not George Bush, would have won the
election. Then you might think it is a little odd that whether Gore or Bush wins depends on
who else is in the election, and not on the voters’ preferences between Gore and Bush. This
is a special case of the independence of irrelevant alternatives - you think that the voting
system should end up with the result that it would have come up with had there been just
those two candidates. If we generalise this motivation a lot, we get the conclusion that third
possibilities should be irrelevant.
Unfortunately, we’ve now got ourselves into an impossible situation. Arrow’s theorem
says that any ranking function that satisfies the Pareto and independence of irrelevant al-
ternatives constraints, has a dictator in any case where the number of alternatives is greater
than 2. When there are only 2 choices, majority rule satisfies all the constraints. But nothing,
other than dictatorship, works in the general case.

25.2 Cyclic Preferences

We can see why three option cases are a problem by considering one very simple example.
Say there are three voters, V1 , V2 , V3 and three choices A, B, C. The agent’s rankings are
given in the table below. (The column under each voter lists the choices from their first
preference, on top, to their least favourite option, on the bottom.)

V1 V2 V3
A B C
B C A
C A B

131
If we just look at the A/B comparison, A looks pretty good. After all, 2 out of 3 voters prefer
A to B. But if we look at the B/C comparison, B looks pretty good. After all, 2 out of 3 voters
prefer B to C. So perhaps we should say A is best, B second best and C worst. But wait! If
we just look at the C/A comparison, C looks pretty good. After all, 2 out of 3 voters prefer
C to A.
It might seem like one natural response here is to say that the three options should be
tied. The group preference ranking should just be that A =G B =G = C. But note what hap-
pens if we say that and accept independence of irrelevant alternatives. If we eliminate option
C, then we shouldn’t change the group’s ranking of A and B. That’s what independence of
irrelevant alternatives says. So now we’ll be left with the following rankings.

V1 V2 V3
A B A
B A B

By independence of irrelevant alternatives, we should still have A =G B. But 2 out of 3

voters wanted A over B. The one voter who preferred B to A is making it that the group
ranks them equally. That’s a long way from making them a dictator, but it’s our first sign
that our constraints give excessive power to one voter. One other thing the case shows is
that we can’t have the following three conditions on our ranking function.

• If there are just two choices, then the majority choice is preferred by the group.
• If there are three choices, and they are symmetrically arranged, as in the table above,
then all choices are equally preferred.
• The ranking function satisfies independence of irrelevant alternatives.

I noted after the example that V2 has quite a lot of power. Their preference makes it that
the group doesn’t prefer A to B. We might try to generalise this power. Maybe we could try
for a ranking function that worked strictly by consensus. The idea would be that if everyone
prefers A to B, then A >G B, but if there is no consensus, then A =G B. Since how the group
ranks A and B only depends on how individuals rank A and B, this method easily satisfies
independence of irrelevant alternatives. And there are no dictators, and the method satisfies
the Pareto condition. So what’s the problem?
Unfortunately, the consensus method described here violates transitivity, so doesn’t even
produce a group preference ordering in the formal sense we’re interested in. Consider the
following distribution of preferences.

V1 V2 V3
A A B
B C A
C B C

132
Everyone prefers A to C, so by unanimity, A >G C. But there is no consensus over the A/B
comparison. Two people prefer A to B, but one person prefers B to A. And there is no
consensus over the B/C comparison. Two people prefer B to C, but one person prefers C to
B. So if we’re saying the group is indifferent between any two options over which there is
no consensus, then we have to say that A =G B, and B =G C. By transitivity, it follows that
A =G C, contradicting our earlier conclusion that A >G C.
This isn’t going to be a formal argument, but we might already be able to see a difficulty
here. Just thinking about our first case, where the preferences form a cycle suggests that the
only way to have a fair ranking consistent with independence of irrelevant alternatives is to
say that the group only prefers options when there is a consensus in favour of that option.
But the second case shows that consensus based methods do not in general produce rankings
of the options. So we have a problem. Arrow’s Theorem shows how deep that problem goes.

25.3 Proofs of Arrow’s Theorem

The proofs of Arrow’s Theorem, though not particularly long, are a little tricky to follow.
So we won’t go through them in any detail at all. But I’ll sketch one proof due to John
Geanakopolos of the Cowles Foundation at Yale.1 Geanakopolos assumes that we have a
ranking function that satisfies Pareto and independence of irrelevant alternatives, and aims
to show that in this function there must be a dictator.
The first thing he proves is a rather nice lemma. Assume that every voter puts some
option B on either the top or the bottom of their preference ranking. Don’t assume they all
agree: some people hold that B is the very best option, and the rest hold that it is the worst.
Geanakopolos shows that in this case the ranking function must put B either at the very top
or the very bottom.
To see this, assume that it isn’t true. So there are some options A and C such that A ≥G B
and B ≥G C. Now imagine changing each voter’s preferences so that C is moved above A
while B stays where it is - either on the top or the bottom of that particular voter’s prefer-
ences. By Pareto, we’ll now have C >G A, since everyone prefers C to A. But we haven’t
changed how any person thinks about any comparison involving B. So by independence of
irrelevant alternatives, A ≥G B and B ≥G C must still be true. By transitivity, it follows that
A ≥G C, contradicting our conclusion that C >G A.
This is a rather odd conclusion I think. Imagine that we have four voters with the fol-
lowing preferences.

V1 V2 V3 V4
B B A C
A C C A
C A B B

1 The proof is available at http://ideas.repec.org/p/cwl/cwldpp/1123r3.html.

133
By what we’ve proven so far, B has to come out either best or worst in the group’s rankings.
But which should it be? Since half the people love B, and half hate it, it seems it should get
a middling ranking. One lesson of this is that independence of irrelevant alternatives is a
very strong condition, one that we might want to question.
The next stage of Geanakopolos’s proof is to consider a situation where at the start every-
one thinks B is the very worst option out of some long list of options. One by one the voters
change their mind, with each voter in turn coming to think that B is the best option. By
the result we proved above, at every stage of the process, B must be either the worst option
according to the group, or the best option. B starts off as the worst option, and by Pareto B
must end up as the best option. So at one point, when one voter changes their mind, B must
go from being the worst option on the group’s ranking to being the best option, simply in
virtue of that person changing their mind.
We won’t go through the rest, but the proof continues by showing that that person has
to be a dictator. Informally, the idea is to prove two things about that person, both of which
are derived by repeated applications of independence of irrelevant alternatives. First, this
person has to retain their power to move B from worst to first whatever the other people
think of A and C. Second, since they can make B jump all options by changing their mind
about B, if they move B ‘halfway’, say they come to have the view A >V B >V C, then B will
jump (in the group’s ranking) over all options that it jumps over in this voter’s rankings. But
that’s possible (it turns out) only if the group’s ranking of A and C is dependent entirely on
this voter’s rankings of A and C. So the voter is a dictator with respect to this pair. A further
argument shows that the voter is a dictator with respect to every pair, which shows there
must be a dictator.

134
 Chapter 26

Voting Systems
The Arrow Impossibility Theorem shows that we can’t have everything that we want in a
voting system. In particular, we can’t have a voting system that takes as inputs the prefer-
ences of each voter, and outputs a preference ordering of the group that satisfies these three
constraints.

1. Unanimity: If everyone prefers A to B, then the group prefers A to B.

2. Independence of Irrelevant Alternatives: If nobody changes their mind about the
relative ordering of A and B, then the group can’t change its mind about the relative
ordering of A and B.
3. No Dictators: For each voter, it is possible that the group’s ranking will be different
to their ranking

Any voting system either won’t be a function in the sense that we’re interested in for
Arrow’s Theorem, or will violate some of those constraints. (Or both.) But still there could
be better or worse voting systems. Indeed, there are many voting systems in use around the
world, and serious debate about which is best. In these notes we’ll look at the pros and cons
of a few different voting systems.
The discussion here will be restricted in two respects. First, we’re only interested in
systems for making political decisions, indeed, in systems for electing representatives to
political positions. We’re not interested in, for instance, the systems that a group of friends
might use to choose which movie to see, or that an academic department might use to hire
new faculty. Some of the constraints we’ll be looking at are characteristic of elections in
particular, not of choices in general.
Second, we’ll be looking only at elections to fill a single position. This is a fairly sub-
stantial constraint. Many elections are to fill multiple positions. The way a lot of electoral
systems work is that many candidates are elected at once, with the number of representa-
tives each party gets being (roughly) in proportion to the number of people who vote for that
party. This is how the parliament is elected in many countries around the world (includ-
ing, for instance, Mexico, Germany and Spain). Perhaps more importantly, it is basically
the norm for new parliaments to have such kind of multi-member constituencies. But the
mathematical issues get a little complicated when we look at the mechanisms for select-
ing multiple candidates, and we’ll restrict ourselves to looking at mechanisms for electing a
single candidate.

135
26.1 Plurality voting
By far the most common method used in America, and throughout much of the rest of the
world, is plurality voting. Every voter selects one of the candidates, and the candidates with
the most votes wins. As we’ve already noted, this is called plurality, or first-past-the-post,
voting.
Plurality voting clearly does not satisfy the independence of irrelevant alternatives con-
dition. We can see this if we imagine that the voting distribution starts off with the table
on the left, and ends with the table on the right. (The three candidates are A, B and C, with
the numbers at the top of each column representing the percentage of voters who have the
preference ordering listed below it.)

40% 35% 25% 40% 35% 25%

A B C A B B
B A B B A C
C C A C C A

All that happens as we go from left-to-right is that some people who previously favoured C
over B, come to favour B over C. Yet this change, which is completely independent of how
anyone feels about A, is sufficient for B to go from losing the election 40-35 to winning the
election 60-40.
This is how we show that a system does not satisfy independent of irrelevant alternatives
- coming up with a pair of situations where no voter’s opinion about the relative merits of
two choices (in this case A and B) changes, but the group’s ranking of those two choices
changes.
One odd effect of this is that whether B wins the election depends not just on how voters
compare A and B, but on how voters compare B and C. One of the consequences of Arrow’s
Theorem might be taken to be that this kind of thing is unavoidable, but it is worth stopping
to reflect on just how pernicious this is to the democratic system.
Imagine that we are in the left-hand situation, and you are one of the 25% of voters who
like C best, then B then A. It seems that there is a reason for you to not vote the way your
preferences go; you’ll have a better chance of electing a candidate you prefer if you vote,
against your preferences, for B. So the voting system might encourage voters to not express
their preferences adequately. This can have a snowball effect - if in one election a number
of people who prefer C vote for B, at future elections other people who might have voted for
C will also vote for B because they don’t think enough other people share their preferences
for C to make such a vote worthwhile.
Indeed, if the candidate C themselves strongly prefers B to A, but thinks a lot of people
will vote for them if they run, then C might even be discouraged from running because it
will lead to a worse election result. This doesn’t seem like a democratically ideal situation.
Some of these consequences are inevitable consequences of a system that doesn’t satisfy
independence of irrelevant alternatives. And the Arrow Theorem shows that it is hard to
avoid independence of irrelevant alternatives. But some of them seem like serious demo-
cratic shortcomings, the effects of which can be seen in American democracy, and especially
in the extreme power the two major parties have. (Though, to be fair, a number of other

136
electoral systems that use plurality voting do not have such strong major parties. Indeed,
Canada seems to have very strong third parties despite using this system.)
One clear advantage of plurality voting should be stressed: it is quick and easy. There is
little chance that voters will not understand what they have to do in order to express their
preferences. (Although as Palm Beach county keeps showing us, this can happen.) And
voting is, or at least should be, relatively quick. The voter just has to make one mark on
a piece of paper, or press a single button, to vote. When the voter is expected to vote for
dozens of offices, as is usual in America (though not elsewhere) this is a serious benefit. In
the recent U.S. elections we saw queues hours long of people waiting to vote. Were voting
any slower than it actually is, these queues might have been worse.
Relatedly, it is easy to count the votes in a plurality system. You just sort all the votes
into different bundles and count the size of each bundle. Some of the other systems we’ll be
looking at are much harder to count the votes in. I’m writing this a month after the 2008
U.S. elections, and some of the votes still haven’t been counted in some elections. If the U.S.
didn’t use plurality voting, this would likely be a much worse problem.

26.2 Runoff Voting

One solution to some of the problems with plurality voting is runoff voting, which is used in
parts of America (notably Georgia and Louisiana) and is very common throughout Europe
and South America. The idea is that there are, in general, two elections. At the first election,
if one candidate has majority support, then they win. But otherwise the top two candidates
go into a runoff. In the runoff, voters get to vote for one of those two candidates, and the
candidate with the most votes wins.
This doesn’t entirely deal with the problem of a spoiler candidate having an outsized ef-
fect on the election, but it makes such cases a little harder to produce. For instance, imagine
that there are four candidates, and the arrangement of votes is as follows.

35% 30% 20% 15%

A B C D
B D D C
C C B B
D A A A

In a plurality election, A will win with only 35% of the vote.2 In a runoff election, the runoff
will be between A and B, and presumably B will win, since 65% of the voters prefer B to A.
But look what happens if D drops out of the election, or all of D’s supporters decide to vote
more strategically.
2 This isn’t actually that unusual in the overall scope of American elections. John McCain won several crucial

Republican primary elections, especially in Florida and Missouri, with under 35% of the vote. Without those wins,
the Republican primary contest would have been much closer.

137
 35% 30% 20% 15%
A B C C
B C B B
C A A A

Now the runoff is between C and A, and C will win. D being a candidate means that the
candidate most like D, namely C, loses a race they could have won.
In one respect this is much like what happens with plurality voting. On the other hand,
it is somewhat harder to find real life cases that show this pattern of votes. That’s in part
because it is hard to find cases where there are (a) four serious candidates, and (b) the third
and fourth candidates are so close ideologically that they eat into each other’s votes and (c)
the top two candidates are so close that these third and fourth candidates combined could
leapfrog over each of them. Theoretically, the problem about spoiler candidates might look
as severe, but it is much less of a problem in practice.
The downside of runoff voting of course is that it requires people to go and vote twice.
This can be a major imposition on the time and energy of the voters. More seriously from
a democratic perspective, it can lead to an unrepresentative electorate. In American runoff
elections, the runoff typically has a much lower turnout than the initial election, so the
election comes down to the true party loyalists. In Europe, the first round often has a very
low turnout, which has led on occasion to fringe candidates with a small but loyal supporter
base making the final round.

26.3 Instant Runoff Voting

One approach to this problem is to do, in effect, the initial election and the runoff at the
same time. In instant runoff voting, every voter lists their preference ordering over their
desired candidates. In practice, that means marking ‘1’ beside their first choice candidate,
‘2’ beside their second choice and so on through the candidates.
When the votes are being counted, the first thing that is done is to count how many
first-place votes each candidate gets. If any candidate has a majority of votes, they win. If
not, the candidate with the lowest number of votes is eliminated. The vote counter then
distributes each ballot for that eliminated candidate to whichever candidate receives the ‘2’
vote on that ballot. If that leads to a candidate having a majority, that candidate wins. If not,
the candidate with the lowest number of votes at this stage is eliminated, and their votes
are distributed, each voter’s vote going to their most preferred candidate of the remaining
candidates. This continues until a candidate gets a majority of the votes.
This avoids the particular problem we discussed about runoff voting. In that case, D
would have been eliminated at the first round, and D’s votes would all have flowed to C.
That would have moved C about B, eliminating B. Then with B’s preferences, C would have
won the election comfortably. But it doesn’t remove all problems. In particular, it leads to
an odd kind of strategic voting possibility. The following situation does arise, though rarely.
Imagine the voters are split the following way.

138
 45% 28% 27%
A B C
B A B
C C A

As things stand, C will be eliminated. And when C is eliminated, all of C’s votes will be
transferred to B, leading to B winning. Now imagine that a few of A’s voters change the way
they vote, voting for C instead of their preferred candidate A, so now the votes look like this.

43% 28% 27% 2%

A B C C
B A B A
C C A B

Now C has more votes than B, so B will be eliminated. But B’s voters have A as their second
choice, so now A will get all the new votes, and A will easily win. Some theorists think that
this possibility for strategic voting is a sign that instant runoff voting is flawed.
Perhaps a more serious worry is that the voting and counting system is more compli-
cated. This slows down voting itself, though this is a problem can be partially dealt with
by having more resources dedicated to making it possible to vote. The vote count is also
somewhat slower. A worse consequence is that because the voter has more to do, there is
more chance for the voter to make a mistake. In some jurisdictions, if the voter does not put
a number down for each candidate, their vote is invalid, even if it is clear which candidate
they wish to vote for. It also requires the voter to have opinions about all the candidates
running, and this may include a number of frivolous candidates. But it isn’t clear that this is
a major problem if it does seem worthwhile to avoid the problems with plurality and runoff
voting.

139
 Chapter 27

More Voting Systems

In the previous chapter we looked at a number of voting systems that are in widespread
use in various democracies. Here we look at three voting systems that are not used for mass
elections anywhere around the world, though all of them have been used in various purposes
for combining the views of groups. (For instance, they have been used for elections in small
groups.)

27.1 Borda Count

In a Borda Count election, each voter ranks each of the candidates, as in Instant Runoff
Voting. Each candidate then receives n points for each first place vote they receive (where
n is the number of candidates), n – 1 points for each second place vote, and so on through
the last place candidate getting 1 point. The candidate with the most points wins.
One nice advantage of the Borda Count is that it eliminates the chance for the kind of
strategic voting that exists in Instant Runoff Voting, or for that matter any kind of Runoff
Voting. It can never make it more likely that A will win by someone changing their vote
away from A. Indeed, this could only lead to A having fewer votes. This certainly seems to
be reasonable.
Another advantage is that many preferences beyond first place votes count. A candidate
who is every single voter’s second best choice will not do very well under any voting system
that gives a special weight to first preferences. But such a candidate may well be in a certain
sense the best representative of the whole community.
And a third advantage is that the Borda Count includes a rough approximation of voter’s
strength of preference. If one voter ranks A a little above B, and another votes B many places
above A, that’s arguably a sign that B is a better representative of the two of them than A.
Although only one of the two prefers B, one voter will be a little disappointed that B wins,
while the other would be very disappointed if B lost.
These are not trivial advantages. But there are also many disadvantages which explain
why no major electoral system has adopted Borda Count voting yet, despite its strong sup-
port from some theorists.
First, Borda Count is particularly complicated to implement. It is just as difficult for the
voter to as in Instant Runoff Voting; in each case they have to express a complete preference
ordering. But it is much harder to count, because the vote counter has to detect quite a bit
of information from each ballot. Getting this information from millions of ballots is not a
trivial exercise.

140
 Second, Borda Count has a serious problem with ‘clone candidates’. In plurality voting,
a candidate suffers if there is another candidate much like them on the ballot. In Borda
Count, a candidate can seriously gain if such a candidate is added. Consider the following
situation. In a certain electorate, of say 100,000 voters, 60% of the voters are Republicans,
and 40% are Democrats. But there is only one Republican, call them R, on the ballot, and
there are 2 Democrats, D1 and D2 on the ballot. Moreover, D2 is clearly a worse candidate
than D1, but the Democrats still prefer the Democrat to the Republican. Since the district is
overwhelmingly Republican, intuitively the Republican should win. But let’s work through
what happens if 60,000 Republicans vote for R, then D1, then D2, and the 40,000 Democrats
vote D1 then D2 then R. In that case, R will get 60, 000×3+40, 000×1 = 220, 000 points, D1
will get 60, 000×2+40, 000×3 = 240, 000 points, and D2 will get 60, 000×1+40, 000×2 =
140, 000 points, and D1 will win. Having a ‘clone’ on the ticket was enough to push D1 over
the top.
On the one hand, this may look a lot like the mirror image of the ‘spoiler’ problem for
plurality voting. But in another respect it is much worse. It is hard to get someone who
is a lot ideologically like your opponent to run in order to improve your electoral chances.
It is much easier to convince someone who already wants you to win to add their name to
the ballot in order to improve your chances. In practice, this would either lead to an arms
race between the two parties, each trying to get the most names onto the ballot, or very
restrictive (and hence undemocratic) rules about who was even allowed to be on the ballot,
or, most likely, both.
The third problem comes from thinking through the previous problem from the point
of view of a Republican voter. If the Republican voters realise what is up, they might vote
tactically for D2 over D1, putting R back on top. In a case where the electorate is as partisan
as in this case, this might just work. But this means that Borda Count is just as susceptible
to tactical voting as other systems; it is just that the tactical voting often occurs downticket.
(There are more complicated problems, that we won’t work through, about what happens
if the voters mistakenly judge what is likely to happen in the election, and tactical voting
backfires.)
Finally, it’s worth thinking about whether the supposed major virtue of Borda Count,
the fact that it considers all preferences and not just first choices, is a real gain. The core idea
behind Borda Count is that all preferences should count equally. So the difference between
first place and second place in a voter’s affections counts just as much as the difference be-
tween third and fourth. But for many elections, this isn’t how the voters themselves feel.
I suspect many people reading this have strong feelings about who was the best candidate
in the past Presidential election. I suspect very few people had strong feelings about who
was the third best versus fourth best candidate. This is hardly a coincidence; people identify
with a party that is their first choice. They say, “I’m a Democrat” or “I’m a Green” or “I’m
a Republican”. They don’t identify with their third versus fourth preference. Perhaps vot-
ing systems that give primary weight to first place preferences are genuinely reflecting the
desires of the voters.

27.2 Approval Voting

In plurality voting, every voter gets to vote for one candidate, and the candidate with the
most votes wins. Approval voting is similar, except that each voter is allowed to vote for as

141
many candidates as they like. The votes are then added up, and the candidate with the most
votes wins. Of course, the voter has an interest in not voting for too many candidates. If
they vote for all of the candidates, this won’t advantage any candidate; they may as well have
voted for no candidates at all.
The voters who are best served by approval voting, at least compared to plurality voting,
are those voters who wish to vote for a non-major candidate, but who also have a preference
between the two major candidates. Under approval voting, they can vote for the minor
candidate that they most favour, and also vote for the the major candidate who they hope
will win. Of course, runoff voting (and Instant Runoff Voting) also allow these voters to
express a similar preference. Indeed, the runoff systems allow the voters to express not only
two preferences, but express the order in which they hold those preferences. Under approval
voting, the voter only gets to vote for more than one candidate, they don’t get to express any
ranking of those candidates.
But arguably approval voting is easier on the voter. The voter can use a ballot that looks
just like the ballot used in plurality voting. And they don’t have to learn about preference
flows, or Borda Counts, to understand what is going on in the voting. Currently there are
many voters who vote for, or at least appear to try to vote for, multiple candidates. This is
presumably inadvertent, but approval voting would let these votes be counted, which would
refranchise a number of voters. Approval voting has never been used as a mass electoral
tool, so it is hard to know how quick it would be to count, but presumably it would not be
incredibly difficult.
One striking thing about approval voting is that it is not a function from voter prefer-
ences to group preferences. Hence it is not subject to the Arrow Impossibility Theorem. It
isn’t such a function because the voters have to not only rank the candidates, they have to
decide where on their ranking they will ‘draw the line’ between candidates that they will vote
for, and candidates that they will not vote for. Consider the following two sets of voters. In
each case candidates are listed from first preference to last preference, with stars indicating
which candidates the voters vote for.

40% 35% 25% 40% 35% 25%

∗A ∗B ∗C ∗A ∗B ∗C
B A B B A ∗B
C C A C C A

In the election on the left-hand-side, no voter takes advantage of approval voting to vote
for more than one candidate. So A wins with 40% of the vote. In the election on the right-
hand-side, no one’s preferences change. But the 25% who prefer C also decide to vote for B.
So now B has 60% of the voters voting for them, as compared to 40% for A and 25% for C,
so B wins.
This means that the voting system is not a function from voter preferences to group
preferences. If it were a function, fixing the group preferences would fix who wins. But
in this case, without a single voter changing their preference ordering of the candidates, a
different candidate won. Since the Arrow Impossibility Theorem only applies to functions
from voter preferences to group preferences, it does not apply to Approval Voting.

142
27.3 Range Voting
In Range Voting, every voter gives each candidate a score. Let’s say that score is from 0 to
10. The name ‘Range’ comes from the range of options the voter has. In the vote count, the
score that each candidate receives from each voter is added up, and the candidate with the
most points wins.
In principle, this is a way for voters to express very detailed opinions about each of the
candidates. They don’t merely rank the candidates, they measure how much better each
candidate is than all the other candidates. And this information is then used to form an
overall ranking of the various candidates.
In practice, it isn’t so clear this would be effective. Imagine that a voter V thinks that
candidate A would be reasonably good, and candidate B would be merely OK, and that
no other candidates have a serious chance of winning. If V was genuinely expressing their
opinions, they might think that A deserves an 8 out of 10, and B deserves a 5 out of 10. But
V wants A to win, since V thinks A is the better candidate. And V knows that what will
make the biggest improvement in A’s chances is if they score A a 10 out of 10, and B a 0 out
of 10. That will give A a 10 point advantage, whereas they may only get a 3 point advantage
if the voter voted sincerely.
It isn’t unusual for a voter to find themselves in V’s position. So we might suspect that
although Range Voting will give the voters quite a lot of flexibility, and give them the chance
to express detailed opinions, it isn’t clear how often it would be in a voter’s interests to use
these options.
And Range Voting is quite complex, both from the perspective of the voter and of the
vote counter. There is a lot of information to be gleaned from each ballot in Range Voting.
This means the voter has to go to a lot of work to fill out the ballot, and the vote counter has
to do a lot of work to process all that information. This means that Range Voting might be
very slow, both in terms of voting and counting. And if voters have a tactical reason for not
wanting to fill in detailed ballots, this might mean it’s a lot of effort for not a lot of reward,
and that we should stick to somewhat simpler vote counting methods.

27.4 Exercises
For each of the following voting systems, say (a) whether they are functions from expressed
preferences of voters to a preference ordering by the group, and, if so, (b) which of the
Arrow constraints (unanimity, no dictators, independence of irrelevant alternatives) they
fail to satisfy.

1. Runoff Voting
2. Instant Runoff Voting
3. Borda Count
4. Range Voting

For each case where you say the voting system is not a function, or say that a constraint
is not satisfied, you should give a pair of examples (like the pairs on pages 122 and 127) to
demonstrate this.

143