Elements of Decision Under Uncertainty: Table 1.1

Download as pdf or txt
Download as pdf or txt
You are on page 1of 28

1

Elements of Decision under Uncertainty

We introduce a model for decision making under uncertainty that will be


our workhorse throughout the book. Uncertainty is modeled with a set of
states of nature, one of which will occur. The decision maker or individual
has a probability distribution over the states of nature that represents his (or
her) subjective beliefs about the likelihood of different states of nature. This
individual chooses actions and actions have consequences. The consequence
for the individual depends on the state of nature and his choice of action.
The states of nature are represented in a way that the probabilities of states
are unaffected by the individual’s actions. The individual’s preferences over
consequences are captured by a utility function. The probability distribution
over states of nature and the utility function over consequences, both of
which are subjective,1 are combined by the expected-utility rule to induce
an expected utility over actions.
An individual must choose among acts – or synonymously, he or she
must make decisions, or select among actions, options, or moves. And, where
there is uncertainty, nature may be said to “choose” the state of the world
(or state, for short). You decide whether or not to carry an umbrella; nature
“decides” on rain or sunshine. Table 1.1 pictures an especially simple 2 × 2
situation. Your alternative acts x = (1, 2) are shown along the left margin,
and nature’s alternative states s = (1, 2) across the top. The body of the
table shows the consequences cxs resulting from your choice of act x and
nature’s choice of state s.

1
Subjective in the sense that another individual, faced with the same decision problem, may
have a different probability distribution and a different utility function: beliefs and tastes
may differ from person to person.

7
8 Elements of Decision under Uncertainty

Table 1.1. Consequences of alternative acts and states

States
s=1 s=2

x=1 c11 c12


Acts
x=2 c21 c22

More generally, the individual under uncertainty will, according to this


analysis, specify the following elements of his decision problem:

(1) a set of acts (l, . . . , x, . . . , X) available to him;


(2) a set of states (l, . . . , s, . . . , S) available to nature;
(3) a consequence function cxs showing outcomes under all combinations
of acts and states.

And, in addition:

(4) a probability function π(s) expressing his beliefs (as to the likelihood
of nature choosing each and every state);
(5) an elementary-utility function v(c) measuring the desirability of the
different possible consequences to him.

We will explain below how the “expected-utility rule” integrates all these
elements so as to enable the individual to decide upon the most advanta-
geous action. Put another way, we will show how the economic agent can
derive a personal preference ordering of his possible acts from his given
preference scaling over consequences.
comment: The approach here does not allow for the psychological sensa-
tions of vagueness or confusion that people often suffer in facing situations
with uncertain (risky) outcomes. In our model, the individual is neither
vague nor confused. While recognizing that his knowledge is imperfect, so
that he cannot be sure which state of the world will occur, he nevertheless can
assign exact numerical probabilities representing his degree of belief as to
the likelihood of each possible state. Our excuse for not picturing vagueness
or confusion is that we are trying to model economics, not psychology. Even
the very simplest models in economic textbooks, for example, indifference-
curve diagrams, implicitly postulate a degree of precise self-knowledge that
is descriptively unrealistic. The ultimate justification, for indifference-curve
diagrams or for theories of decision under uncertainty, is the ability of such
models to help us understand and predict behavior.
1.2 The Probability Distribution 9

1.1 The Menu of Acts


There are two main classes of individual actions: terminal moves versus
informational moves. Here, in Part I of the book, we consider a simplified
world where only terminal acts are available, so that the individual is limited
to making the best of his or her existing combination of knowledge and
ignorance. An example of terminal action under uncertainty is the statistical
problem of coming to a decision on the basis of sample evidence now in
hand: for instance, when a regulatory agency has to decide whether or not
to approve a new drug on the basis of experimental test results. We will be
considering terminal actions of this type, and especially the risk-involved
decisions of individuals in markets: whether or not to purchase insurance, to
buy or sell stocks and bonds, to participate in a partnership, etc. Anticipating
a bit, a key theme of our analysis will be that markets allow decision makers
to share risks and returns in ways that accord with the particular preferences
and opportunities of the different transactors.
Part II of the book will be covering informational actions – decisions
concerning whether and how to improve upon one’s state of knowledge
before making a terminal move. In the class of informational actions would
fall statistical choices such as how much additional evidence to collect before
coming to a terminal decision, what sampling technique to employ, etc. Once
again, our emphasis will be on ways of acquiring new information through
markets. Knowledge can be acquired by direct market purchase – by buying
newspapers for weather and stock market reports, by undergoing a course
of training to gain “know how” in a trade, or by employing an expert for
private advice. Rather less obviously, markets open up an indirect means
of acquiring information: for example, a person can observe the market
choices of better-informed traders, or might draw inferences from people’s
reputations acquired in the course of their previous market dealings. Or,
a producing firm might imitate other commercially successful firms. But
these interesting phenomena involving information-involved actions will
have to be set aside until Part II.

1.2 The Probability Distribution


We assume that each person is able to represent his beliefs as to the likelihood
of the different states of the world (e.g., as to whether nature will choose
rain or shine) by a “subjective” probability distribution (Savage, 1954).
Assuming discrete states of the world, the individual is supposed to be able
to assign to each state s a degree of belief, in the form of numerical weights
10 Elements of Decision under Uncertainty

π s lying between zero and one inclusive, and summing to unity: s π s = 1.


In the extreme case, if the person were certain that some particular state s
would be occurring, the full probabilistic weight of unity would be assigned
to that state. Then π s = 1, so that zero probability is attached to every other
state in the set 1, . . . , s, . . . , S. More generally, a high degree of subjective
assurance will be reflected by a relatively “tight” probability distribution
over the range of possible states; a high degree of doubt would be reflected
by a wide dispersion.
At times, we shall find it will be more convenient to assume that the
variable or variables defining the state of the world vary continuously (rather
than discretely) so that the number of distinct states is uncountably infinite.
Here the probability of any exact single state coming about is regarded as
zero (“infinitesimal”), although the event is not impossible. Making use of a
continuous state-defining variable s, where s can be any real number between
0 and S, the individual’s subjective probability beliefs  S would be represented
by a probability density function π(s) such that 0 π (s) ds = 1.

1.2.1 Risk versus Uncertainty


A number of economists have attempted to distinguish between risk and
uncertainty, as originally proposed by Frank H. Knight (1921, pp. 20, 226).
(1) “Risk,” Knight said, refers to situations where an individual is able to
calculate probabilities on the basis of an objective classification of instances.
For example, in tossing a fair die the chance of any single one of the six
faces showing is exactly one-sixth. (2) “Uncertainty,” he contended, refers
to situations where no objective classification is possible, for example, in
estimating whether or not a cure for cancer will be discovered in the next
decade.
In this book, we disregard Knight’s distinction. For our purposes, risk and
uncertainty mean the same thing. It does not matter, we contend, whether
an “objective” classification is or is not possible. For we will be dealing
throughout with a “subjective” probability concept (as developed especially
by Savage, 1954): probability is simply degree of belief. In fact, even in
cases like the toss of a die where assigning “objective” probabilities appears
possible, such an appearance is really illusory. That the chance of any single
face turning up is one-sixth is a valid inference only if the die is a fair one – a
condition about which no one could ever be “objectively” certain. Decision
makers are therefore never in Knight’s world of risk but instead always in
his world of uncertainty. That this approach, assigning probabilities on the
1.2 The Probability Distribution 11

basis of subjective degree of belief, is a workable and fruitful procedure will


be shown constructively throughout the book.2

1.2.2 “Hard” versus “Soft” Probabilities


While we do not distinguish between what Knight called risk and uncer-
tainty, he was getting at – though imperfectly expressing – an important
and valid point. In his discussion, Knight suggested that a person’s actions
may well depend upon his “estimate of the chance that his estimates are
correct,” or, we shall say, upon his confidence in his beliefs. This brings us to
a distinction between “hard” versus “soft” probability estimates.
Suppose that for purposes of an immediate bet you had to estimate the
probability of heads coming up on the next toss of coin A – the coin having
been previously tested many times by you and found to have historically
come up heads and tails with just about equal frequency. If you are a reason-
able person, you would assign a degree of belief (subjective probability) of
about 0.5 to heads, and you would be rather confident about that number.
In contrast, imagine instead that you are dealing with coin B, about which
you know absolutely nothing. You have not even been able to inspect it to
verify whether it is possibly two tailed or two headed. Nevertheless, if you
had to pick some single number you would be compelled again to assign
0.5 probability to heads coming up on the next toss, since as a reasonable
person you lack any basis for a greater or lesser degree of belief in heads
than tails. But, your confidence in the 0.5 figure for coin B would surely be
much less.
It is not the psychological sensation of confidence or doubt that inter-
ests us, but the possible implications for decisions. If the same probability
assignment of 0.5 will be made either way, as has just been argued, is there
any action-relevant difference between the two cases? For our purpose, the
answer is no, if you are committed to terminal action.3 If you must bet
now on the basis of your current information, 0.5 is the relevant probability
for guiding your choice of heads or tails. In either situation, you have no
grounds for thinking heads more likely or tails more likely. But the answer
is yes, there is indeed a difference between the two situations if you have the

2
See Schmeidler (1989) for the foundations of an alternative approach that explicitly models
Knightian uncertainty and individuals’ attitudes to it.
3
Later in this chapter, we describe the Ellsberg paradox, which is an experiment indicating
that individuals may react differently to hard and soft probabilities in a setting with
terminal actions.
12 Elements of Decision under Uncertainty

option of informational action. When this option is available, you should


be more willing to invest money or effort to obtain additional information
about coin B than about coin A. In short, greater prior doubt (lesser degree
of confidence) makes it more important to acquire additional evidence
before making a terminal move. So, we see that a person’s informational
actions, though not his terminal actions, do depend upon his confidence
in his beliefs – in Knight’s language, upon his “estimate of the chance that
his estimates are correct.” Confidence will be an important topic in Part II
of the book, where we cover the economics of information, but will not be
involved in our more elementary treatment of the economics of uncertainty
in Part I.

Exercises and Excursions 1.2


1 Consistency of Probability Beliefs
An individual believes that credible information will soon arrive in the form
of news about the probability of rain. He believes there is a 50% chance that
the news will be “rain certain,” a 30% chance that the news will be “no
rain,” and a 20% chance that the news will be “rain with probability 0.5.”
Is this consistent with his currently believing that the odds in favor of rain
are 2:1?

2 Information and Confidence


In terms of the chances of a coin coming up heads, suppose there are three
states of the world regarded as possible:

State 1: chance of heads is 100% [coin is two headed]


State 2: chance of heads is 50% [coin is fair]
State 3: chance of heads is 0% [coin is two tailed].

An individual initially assigns equal probabilities (π 1 , π 2 , π 3 ) = ( 13 , 13 , 13 )


to all three states.

(A) For an immediate bet (terminal action), what is his best estimate for
the probability p of heads on the next toss?
(B) Suppose new information were now to change his probability vector
to (π 1 , π 2 , π 3 ) = (0, 1, 0). What can you now say about his best
estimate for p? What has happened to his confidence in that estimate?
(C) Same question if, instead, the new information changed his proba-
bility vector to ( 12 , 0, 12 ).
1.4 The Expected-Utility Rule 13

1.3 The Utility Function


As shown in Table 1.1, each consequence is the outcome of an economic
agent’s choice of action combined with nature’s “choice” of the state of the
world. In principle, the consequence is a full description of all aspects of the
individual’s environment resulting from such an interaction. For example,
if someone decides not to carry an umbrella and nature chooses rain, the
consequences might include getting wet, being late for work, and a variety
of other discomforts. But we shall mainly be concerned with consequences
describable in terms of alternative baskets of consumption goods that enter
into individuals’ utility functions. Very frequently we shall deal with an even
simpler picture in which consequences take the form of entitlements to a
single summary variable like monetary income.
Consequences might be quantities that are certain, or might themselves
be probabilistic – depending upon how states of the world are described.
If the states are defined deterministically, as in “Coin shows heads,” and
supposing the action chosen was “Bet $1 at even money on heads,” then
the consequence would be “Win one dollar.” But states of the world can
sometimes be defined as probabilistic processes. The relevant states might
be “Coin has 50% chance of coming up heads” versus “Coin is biased to
have 75% chance of coming up heads.” Here the act “Bet on heads” will be
reflected, in either state of the world, by an uncertain consequence taking
the form of a specified chance of winning the dollar.
We shall use the notation v(c) to represent a person’s utility function (or
elementary-utility function) over the consequences c.

1.4 The Expected-Utility Rule


Utility attaches directly to consequences, and only derivatively to actions.
To emphasize this distinction, we shall use the notation U(x) for a person’s
derived preference ordering over actions x. The expected-utility rule is used
to derive U(x) from the utility function v(c), as explained below.
A CRUCIAL DISTINCTION
v(c) is a utility function defined over consequences
U(x) is the expected-utility function defined over actions

The analytical problem is to explain and justify this derivation, that is, to
show how, given his direct preferences over consequences, the individual can
order the desirability of the actions available to him.
14 Elements of Decision under Uncertainty

To choose an act is to choose one of the rows of a consequence matrix like


Table 1.1. As the individual is also supposed to have attached a probability
(degree of belief) to the occurrence of every state, each such row can be
regarded as a probability distribution. We may therefore think of a person
as choosing among probability distributions or “prospects.” A convenient
notation for the “prospect” associated with an act x, whose uncertain con-
sequences cx = (cx1 , cx2 , . . . , cxS ) are to be received with respective state
probabilities π = (π 1 , π 2 , . . . , π S ) – the probabilities summing, of course,
to unity – is:

X ≡ (cx1 , cx2 , . . . , cxS ; π1 , π2 , . . . , πS )

The crucial step is to connect the v(c) function for consequences with
the utility ordering U(x) of acts. We can take this step using the famous
“expected-utility rule” of John von Neumann and Oskar Morgenstern (1944,
pp. 15–31):
EXPECTED-UTILITY RULE

U (x) ≡ π1 v(cx1 ) + π2 v(cx2 ) + · · · + πS v(cxS )


S
≡ πs v(cxs ) (1.4.1)
s=1

This says that the expected utility U(x) of act x is calculable in an espe-
cially simple way: to wit, as the mathematical expectation (the probability-
weighted average) of the elementary utilities v(cxs ) of the associated con-
sequences. Note that Equation (1.4.1) is simply additive over states of the
world, which means that the consequence cxs realized in any state s in no
way affects the preference scaling v(cxs0 ) of consequences in any other state
s 0 . Equation (1.4.1) is also linear in the probabilities, another very specific
and special functional form. As the von Neumann–Morgenstern expected-
utility rule is absolutely crucial for our theory of decision under uncertainty,
we shall be devoting considerable space to it.
It turns out that the expected-utility rule is applicable if and only if the
v(c) function has been determined in a particular way that has been termed
the assignment of “cardinal” utilities to consequences. More specifically, the
proposition that we will attempt to explain and justify (though not rigor-
ously prove) can be stated as follows:

Given certain “postulates of rational choice,” there is a way of assigning a cardinal


utility function v(c) over consequences such that the expected-utility rule determines
the individual’s preference ranking U(x) over actions.
1.4 The Expected-Utility Rule 15

A “cardinal” variable is one that can be measured quantitatively, like


altitude, time, or temperature. While different measuring scales might be
employed, such scales can diverge only in zero-point and unit-interval.
Temperature, for example, can be measured according to the Celsius or
the Fahrenheit scales; 32° Fahrenheit is 0° Celsius, and each degree up
or down of Celsius is 1.8° up or down of Fahrenheit. Similarly, altitude
could be measured from sea level or from the center of the Earth (shift of
zero-point) and in feet or meters (shift of unit-interval). Cardinal variables
have the following property: regardless of shift of zero-point and unit-
interval, the relative magnitudes of differences remains unchanged. The
altitude difference between the base and crest of Mount Everest exceeds the
difference between the foundation and roof of even the tallest man-made
building – whether we measure in feet above sea level or in meters from
the center of the Earth.
In dealing with certainty choices, standard economic theory treats utility
(intensity of preference) as an ordinal rather than a cardinal variable. The
individual, it is postulated, can say “I prefer basket A to basket B.” He is not
required to quantify how much he prefers A to B. Put another way, if any given
utility function in the form of an assignment of cardinal numbers to conse-
quences (consumption baskets) correctly describes choices under certainty,
so will any ordinal (positive monotonic) transformation of that function.
Suppose that, for choices not involving risks, some scale u of cardinal num-
bers was attached as preference labels to consequences – where, of course,
higher u indicates greater level of satisfaction. Then any positive monotonic
transformation of those numbers would lead to the same decisions. For
example, suppose an individual always prefers more consumption income c
to less. Then we might say, “He is trying to maximize the function u = c.”
But the income level that maximizes u also maximizes log u or eu , both of
which are positive monotonic transformations of u. So u = e c or u = log
c could equally well have served to indicate the preference scaling. More
formally, if u is a satisfactory function for choices under certainty, then so
is û  F(u), provided only that the first derivative is positive: F  (u) > 0.
In contrast, when it comes to choices under uncertainty, the expected-
utility rule is applicable only if the utility function v(c) has been constructed
in a particular way that provides fewer degrees of freedom. In fact, as will
shortly be seen, given any initially satisfactory v(c) function, only the car-
dinal (positive linear, rather than positive monotonic) transformations of
v(c) will leave preference rankings unchanged. Formally, if v(c) satisfacto-
rily describes the individual’s choices under uncertainty, then so does ν̂ =
α + βν, where α is any constant and β is any positive constant.
16 Elements of Decision under Uncertainty

Why are all the positive monotonic transformations of the utility function
permissible in the riskless case, while only the positive linear transformations
are allowed when it comes to risky choices? In the absence of uncertainty,
deciding upon an action is immediately equivalent to selecting a single defi-
nite consequence. It follows that if someone can rank consequences in terms
of preferences he has already determined the preference ordering of his
actions – which is all that is needed for purposes of decision. But in dealing
with risky choices it is not immediately evident how a ranking of conse-
quences leads to an ordering of actions, since each action will in general
imply a probabilistic mix of possible consequences. The great contribu-
tion of von Neumann and Morgenstern was to show that, given plausible
assumptions about individual preferences, it is possible to construct a v(c)
function – “cardinal” in that only positive linear transformations thereof
are permissible – whose joint use with the expected-utility rule (1.4.1) will
lead to the correct ordering of actions.

1.4.1 An Informal Presentation


To formally justify the joint use of a cardinal utility function and the
expected-utility rule, for dealing with choices among risky prospects,
involves a somewhat higher order of technical difficulty. What follows here
is an informal presentation (based mainly upon Schlaifer, 1959) illustrat-
ing, by direct construction, how the required type of utility function can be
developed.
For the purpose of this discussion, assume that the consequences c are
simply amounts of income a person might receive. Let m represent the
worst possible consequence (the smallest amount of income) that can occur
with positive probability, and M the best possible consequence (the largest
amount of income). More income is preferred to less – so the individual
already has, to begin with, an ordinal utility scale. The problem is to “cardi-
nalize” this scale, that is, to show that there is a way of assigning numerical
values (arbitrary only with respect to zero-point and unit-interval) to the
degrees of preference associated with all levels of income. These values must
be rising with income, else they would not be consistent with the given ordi-
nal preference (“more is preferred to less”). But the chosen scale must also
lead to correct answers when used in conjunction with the expected-utility
rule. The method we shall employ to establish such a cardinal scale is called
“the reference-lottery technique.”
Consider any level of income c ∗ between m and M. Imagine that the
individual is faced with the choice between c ∗ and some “reference lottery”
1.4 The Expected-Utility Rule 17

having the form (M, m; π, 1 − π) in prospect notation. That is, he has


a choice between c ∗ for certain versus a gamble yielding the best possible
outcome M with probability π and the worst possible outcome m with
probability 1 − π. We shall suppose that the individual can say to himself:
“When π becomes very close to unity, I surely will prefer the gamble;
for lotteries with π very close to zero, I surely prefer the certainty of c ∗ .
Consequently, in between there must be some intermediate probability π ∗
of success in the reference lottery, such that I am exactly indifferent between
the certain income c∗ and the prospect (M, m; π ∗ , 1 − π ∗ ).” After due
introspection, we assume, the individual can in fact specify this π ∗ . We may
set v(m) = 0 and v(M) = 1. Then the π ∗ so derived is a cardinal measure
of the utility of income level c∗ for him. That is: v(c∗ ) = π ∗ .4 Or, more
elaborately:

v(c ∗ ) ≡ U (M, m; π ∗ , 1 − π ∗ ) ≡ π ∗ (1.4.2)

An individual proceeding to assign cardinal preference values to income in


this way will generate a v(c) function over the range m  c  M, which
can be employed with the expected-utility rule (1.4.1) to order his choices
among actions.
Figure 1.1 illustrates a hypothetical individual situation. Let m = 0 and
M = 1,000 (in dollars, we can suppose) be the extremes of income that
need be considered. For the specific income c ∗ = 250, the person’s success-
in-equivalent-reference-lottery probability is assumed to be π ∗ = 0.5 –
meaning that he finds himself indifferent between a sure income of $250 and
a 50% chance of winning in an income lottery whose alternative outcomes
are $1,000 or nothing. Then the utility assigned to the sure consequence
$250 is just 12 – that is, v(250) = 0.5, determining the location of point Q
on the v(c) curve. Repeating this process, the reference-lottery technique
generates the entire v(c) curve between m = 0 and M = 1,000.
A full justification, showing why this particular procedure works to derive
a suitable cardinal scale, requires a more formal analysis (to be touched on
in Section 1.4.2). But we can give a geometric intuition here. The essential
point is that the v(c) measure obtained via the reference-lottery technique is
in the form of a probability, so that the expected-utility rule (1.4.1) becomes
equivalent to the standard formula for compounding probabilities.

4
Because shifts of zero-point and unit-interval are permissible for cardinal scaling, more
generally we can write v(c ∗ ) = α + βπ ∗ , for arbitrary α and β > 0. This is equivalent
to assuming v(m) = α and v(M) = α + β. We will henceforth ignore this uninteresting
generalization.
18 Elements of Decision under Uncertainty

Figure 1.1. Utility function.

A Geometric Interpretation
The expected-utility rule is equivalent to the assumption that indifference
curves over lotteries are parallel straight lines. To see this, consider lotteries
over three possible consequences m, c ∗ , and M. Any lottery

x = (m, c ∗ , M; π̂1 , π̂2 , π̂3 ) = (m, c ∗ , M; π̂1 , 1 − π̂1 − π̂3 , π̂3 )

may be represented as a point in (π1 , π3 ) space; see Figure 1.2. The triangle
ABC is the set of all possible lotteries with outcomes m, c ∗ , and M. Point
A corresponds to getting c ∗ for sure, point B is M for sure, and point C
is m for sure. In the lottery x, the probabilities π̂1 and π̂3 (of outcomes m
and M, respectively) are the coordinates of the point x. The probability of
outcome c∗ in this lottery, π̂2 , is the horizontal (or equivalently vertical)
distance from point x in Figure 1.2 to the hypotenuse BC of the triangle.
The expected utility of x is:

U (x) = π̂1 v(m) + (1 − π̂1 − π̂3 )v(c ∗ ) + π̂3 v(M )


= v(c ∗ ) − π̂1 v(c ∗ ) + π̂3 (1 − v(c ∗ ))

where we substitute v(m) = 0 and v(M ) = 1.


1.4 The Expected-Utility Rule 19

π3

1 B

D: (1 − π*, π*)

•x= (πˆ1 , πˆ3 )

Increasing preference

A C π1
0 1

Figure 1.2. Indifference curves under expected-utility rule.

Let y = (m, c ∗ , M; π1 , 1 − π1 − π3 , π3 ) be any other lottery. A similar


calculation shows that U (y) = v(c ∗ ) − π1 v(c ∗ ) − π3 (1 − v(c ∗ )). If lottery
y yields the same expected utility as lottery x, then:

U (y) = v(c ∗ ) − π1 v(c ∗ ) + π3 (1 − v(c ∗ ))


= v(c ∗ ) − π̂1 v(c ∗ ) + π̂3 (1 − v(c ∗ )) = U (x)

Re-arranging this we have:

U (x) − v(c ∗ ) v(c ∗ )


π3 = + π
1 − v(c ∗ ) 1 − v(c ∗ ) 1

This is the equation of the (straight line) indifference curve through x.


Observe that the slope of the indifference curve, v(c ∗ )/(1 − v(c ∗ )), does not
depend on the lottery x. Thus, all indifference curves are parallel straight
lines (shown as broken lines in the Figure 1.2). The indifference lines have
positive slope because (i) any rightward movement from x leads to a less
desirable lottery as it corresponds to increasing π1 at the expense π2 and
(ii) any upward movement from x leads to a more desirable lottery as it
corresponds to increasing π3 at the expense π2 . The direction of increasing
preference is northwest, as indicated by the arrow in Figure 1.2.
20 Elements of Decision under Uncertainty

Point D on the line segment BC in Figure 1.2 corresponds to the lottery


(m, c ∗ , M; 1 − π ∗ , 0, π ∗ ). This lottery has expected utility π ∗ . As v(c ∗ ) =
π ∗ , point D is on the same indifference line as point A.

Example 1.1: Imagine that an individual finds that his reference-lottery


utilities over the range 0  c  1,000 satisfy the specific utility func-
tion v (c) = (c /1,000)1/2 . (This formula is consistent with the previously
obtained point v(250) = 0.5 in Figure 1.1.) Suppose he is now offered a
choice between option A, representing $250 for certain once again, and
option E taking the form of a three-way prospect: E = (810, 360, 160; 0.1,
0.5, 0.4). Which should he choose?5
We already know that v(250) = 0.5: option A is equivalent to a reference
lottery with 50% chance of success. For the elements of option E, we can
readily compute: v(810) = 0.9, v(360) = 0.6, and v(160) = 0.4. That is,
in option E the high possible payoff of $810 is equivalent in preference to
a reference lottery with 90% chance of success, the middling payoff $360
is equivalent to a 60% chance of success, and the poor payoff $160 to a
40% chance of success. Now we ask ourselves: What is the overall equivalent
probability of success associated with option E? We can simply compute it
by using the rule for compounding probabilities:
0.1(0.9) + 0.5(0.6) + 0.4(0.4) = 0.55
So prospect E offers, overall, the equivalent of a 0.55 chance of success in
the reference lottery whereas option A was equivalent only to a 0.5 chance
of success. Evidently, option E is better. The key point is that the equation
leading to the 0.55 number, which we presented as the familiar formula for
compounding probabilities, is also an instance of applying the expected-
utility rule (1.4.1). 

In short, the prescribed way of determining a cardinal v(c) function for use
with the expected-utility rule makes it possible to interpret each v(c) value as
a probability – to wit, the equivalent chance of success in a standardized ref-
erence lottery – and therefore to use the laws of compounding probabilities
for determining the desirability of more complicated prospects.
A few additional comments:
1. We have been assuming here that consequences take the form of simple
quantities of income. More generally, each consequence c might be a
5
The prospect E cannot be represented in Figure 1.2. Only prospects that yield $0, $250, or
$1000 are depicted in Figure 1.2.
1.4 The Expected-Utility Rule 21

basket (vector) of consumption goods. The same technique can be


employed so long as the individual has an ordinal preference scaling of
baskets (an indifference map) to begin with.
2. We have also assumed that the same v(c) scale is applicable in each
and every state of the world. But, if the states are defined as “rain
versus shine,” or “healthy versus sick,” it might appear that attitudes
toward income and income risks, as reflected in the v(c) function,
could differ from state to state. We shall see in Chapter 2, under
the heading of “state-dependent utilities,” how this difficulty can be
handled.
3. Some people find it disturbing that the additive form of the expected-
utility rule (1.4.1) excludes any “complementarities,” positive or neg-
ative, between consequences in different states. For example, if conse-
quences are simple incomes, a higher or lower income in any state sº
is supposed in no way to affect the v(c) number assigned to income
received in any other state s∗ . The reason is simple: incomes in the
distinct states sº and s∗ can never be received in combination but only
as mutually exclusive alternatives. There can be no complementarity
where no possibility of jointness exists.
4. There can be confusion over whether or not the von Neumann–
Morgenstern analysis proves that utility is “really” cardinal rather than
ordinal. Some of the difficulty stems from a mix-up between the v(c)
and the U(x) functions. The cardinality restriction applies to the v(c)
function – the preference scaling over consequences. But we are ulti-
mately interested in the utility rankings of alternative actions, and when
it comes to actions any ordinal transformation of an acceptable utility
measure will always serve equally well. Suppose, for example, that use
of the reference-lottery technique provides the needed utility function
v(c) such that an individual’s actions (prospects) are correctly ordered
by the expected-utility formula U(x)=  S π S v(cS ). Then any posi-
tive monotonic transformation of U(x), such as Û (x) = eU (x) , would
provide an equally correct ordering of the actions. Observe that if
U (x) > U (y) then Û (x) = eU (x) > eU (y) = Û (y).
5. We have emphasized that the von Neumann–Morgenstern analysis
justifies this particular method of constructing a cardinal v(c) scale only
when jointly used with the expected-utility rule. Correspondingly, the
expected-utility rule has not been “proved” to be true. All that has been
shown is that there exists a way of constructing a v(c) function that
makes the expected-utility rule valid as a way of deriving preferences
as to actions from given preferences as to consequences.
22 Elements of Decision under Uncertainty

1.4.2 The Independence Axiom


We are not providing here a formal proof of the expected-utility rule.
Instead, our objective is to clarify the crucial element in the proof, the prin-
ciple of non-complementarity of incomes in different states (see Comment
3 above). The formal postulate expressing this principle, the Independence
Axiom, is also known as the substitution axiom or the axiom of complex
gambles.
Independence Axiom: Suppose an individual is indifferent between two
actions or prospects x and y. Then, for any other prospect z and any fixed
probability p, he will be indifferent between a first complex lottery in which
he receives x with probability p and z otherwise, versus a second com-
plex lottery yielding y with probability p and z otherwise. Moreover, if he
strictly prefers x over y, he will strictly prefer the first complex lottery. Thus,
using the symbol  to indicate indifference and the symbol  for strict
preference:

If x  y, then (x, z; p, 1 − p)  (y, z; p, 1 − p)


If x  y, then (x, z; p, 1 − p)  (y, z; p, 1 − p)

This axiom would be violated if, in a complex prospect, the presence of z


differentially affected the attractiveness of x relative to y – i.e., if there were
any complementarity effect. It might seem this could happen if, say, x and
y were amounts of ordinary commodities like bread and margarine and z
were a commodity like butter (since butter is a consumption complement
for bread but a substitute for margarine). However, in the complex prospects
or lotteries dealt with here, positive or negative complementarity has no role.
The rationale behind this is that the occurrence of x in the one case or of y
in the other rules out z. An individual can never simultaneously enjoy both
x and z together, or both y and z.
An immediate implication of this axiom is that, for two lotteries x and
y such that x  y, we can substitute one for the other in any prospect in
which either appears, without changing the relative preference ordering of
prospects.
In the reference-lottery process, the v(c) associated with any income level
c was determined by finding the probability of success in the reference
lottery equally preferred to that income, i.e.:

If c  (M, m; π, 1 − π), then, because v(m) = 0 and v(M ) = 1, we have


v(c) = π
1.4 The Expected-Utility Rule 23

Figure 1.3. Tree diagram of compound lottery.

In what follows, it will be helpful to introduce the notation l ∗ (π) to


represent a reference lottery in which M is the outcome with probability π
and m is the outcome with probability 1 − π:
l ∗ (π ) = (M, m; π, 1 − π )
Thus, if c ∼ l ∗ (π ) then v(c) = π.
Consider now two levels of income c1 and c2 and their equivalent reference
lotteries l ∗ (π1 ) and l ∗ (π2 ). Then v(c1 ) = π 1 and v(c2 ) = π 2 . Suppose we
wanted to find the preference equivalent of a lottery (c1 , c2 ; p, 1 − p)
involving consequences c1 and c2 with respective probabilities p and 1 − p.
Using the ability to substitute preference-equivalent prospects:
c1 ∼ l ∗ (π1 ) ⇒ (c1 , c2 ; p, 1 − p) ∼ (l ∗ (π1 ), c2 ; p, 1 − p)
Moreover:
c2 ∼ l ∗ (π2 ) ⇒ (l ∗ (π1 ), c2 ; p, 1 − p) ∼ (l ∗ (π1 ), l ∗ (π2 ); p, 1 − p)
Combining these implications:
(c1 , c2 ; p, 1 − p) ∼ (l ∗ (π1 ), l ∗ (π2 ); p, 1 − p) (1.4.3)
The lottery on the right-hand side of (1.4.3) is depicted as a “tree diagram”
in Figure 1.3. Each box or “node” represents a point at which nature makes
a move. Outcomes are indicated at the end of each branch of the tree.
At the initial node, nature “chooses” probabilistically between the two
reference lotteries. Then, depending on this choice, one of the reference
lotteries is played. Note that there are only two outcomes of this compound
lottery, M and m. Adding probabilities, outcome M is reached with proba-
bility pπ 1 + (1 − p)π 2 . Then the compound lottery is itself equivalent to a
reference lottery:
(l ∗ (π1 ), l ∗ (π2 ); p, 1 − p) = l ∗ (pπ1 + (1 − p)π2 ) (1.4.4)
24 Elements of Decision under Uncertainty

Combining (1.4.3) and (1.4.4) it follows that the individual is indifferent


between (c1 , c2 ; p, 1 − p) and a reference lottery in which the probability of
success is pπ 1 + (1 − p)π 2 . As π 1  v(c1 ) and π 2  v(c2 ), it follows that:

U (c1 , c2 ; p, 1 − p) = pπ1 + (1 − p)π2


= pv(c1 ) + (1 − p)v(c2 )

Thus, the independence axiom, which formalizes the principle of non-


complementarity of income over states of the world, leads directly to the
von Neumann–Morgenstern expected-utility rule.

Exercises and Excursions 1.4


1 Transformation of Preferences
An individual claims to be maximizing:

U = (1 + c1 )π1 (1 + c2 )π2

where (c1 , c2 ; π 1 , π 2 ) is a two-state prospect (which means that π 1 + π 2 =


1). Is he a von Neumann–Morgenstern expected-utility (EU) maximizer?
Would all his decisions be consistent with those of an EU maximizer?

2 Indifference Curves in Consequence Space


1
(A) If the utility function is v (c) = c 2 , where c is income, suppose a
person’s preference ordering over actions or prospects in a two-state
world is given by:

1 1
U (c1 , c2 ; π1 , π2 ) = π1 (c1 ) 2 + π2 (c2 ) 2

Depict the indifference curves in a diagram with c1 on the horizontal


axis and c2 on the vertical axis (probabilities held constant). Show that
each indifference curve touches the axes and is everywhere bowed
toward the origin.
(B) If U = 21 πs v(cs ) and v(.) is a strictly concave function,
  show that if
the individual is indifferent between (c1 , c2 ) and c1 , c2 he will strictly
prefer the convex combination (λc1 + (1 − λ)c1 , λc2 + (1 − λ)c2 ).
Hence draw a conclusion about the shape of the indifference curves
in the (c1 , c2 ) plane.
1.4 The Expected-Utility Rule 25

3 The Expected-Utility Rule


Let v(c) be the utility functions for certain outcomes. Then, for lotteries of
the form (c1 , c2 ; π 1 , π 2 ), we have seen that:

2
U (c1 , c2 ; π1 , π2 ) = πs v(cs )
s=1

In this exercise, you are asked to generalize this result to lotteries with three
outcomes. An inductive argument can then be used to show that for any
lottery (c1 , c2 , . . . , cs ; π 1 , π 2 , . . . , π s ):

S
U (c1 , . . . , cs ; π1 , . . . , πs ) = πs v(cs )
s=1

(A) Consider the lottery:


 
ˆl ≡ c , c ; π1 π2
1 2 ,
π1 + π2 π1 + π2

Explain why lˆ ∼ l ∗ (v̄) where:


π1 π2
v̄ ≡ v(c ) + v(c )
π1 + π2 1 π 1 + π2 2
(B) Appeal to the independence axiom to establish that:
ˆ c ; 1 − π , π ) ∼ (l ∗ (v̄), c ; 1 − π , π )
(l, 3 3 3 3 3 3

and

(l ∗ (v̄), c3 ; 1 − π3 , π3 ) ∼ (l ∗ (v̄), l ∗ (c3 ); 1 − π3 , π3 )


ˆ c ; 1 − π , π ) and (l ∗ (v̄), l ∗ (v(c )); 1 −
(C) Depict the two lotteries (l, 3 3 3 3
π3 , π3 ) in tree diagrams.
(D) Confirm that the first is equivalent to the lottery (c1 , c2 , c3 ; π1 ,
π2 , π3 ). Confirm that thesecond is equivalent to the reference lottery
with success probability 3s=1 πs v(cs ).
(E) Suppose the expected-utility rule is true for prospects with S out-
comes. (We have seen that it is true for S = 2 and 3.) Show that
the above argument can, with only slight modifications, be used to
establish that the expected-utility rule must be true for prospects
with S + 1 outcomes.
26 Elements of Decision under Uncertainty

1.5 Risk Aversion


In Figure 1.1 the individual pictured was indifferent between a certainty
income of $250 and a prospect yielding equal chances of $1,000 or nothing.
Such a person is termed risk averse. More generally:
definition: A person is risk averse (displays risk aversion) if he strictly prefers a
certainty consequence to any risky prospect whose mathematical expectation of
consequences equals that certainty. If his preferences go the other way he is a risk
preferrer or loving (displays risk preference); if he is indifferent between the certainty
consequence and such a risky prospect he is risk neutral (displays risk neutrality).

The risky prospect described above, equal chances of $1,000 or nothing,


has a mathematical expectation of $500 of income. Since our individual
was indifferent between the prospect and a mere $250 certain, for him
$500 certain is surely preferable to the risky prospect, which verifies that he
is indeed risk averse.
The term “fair gamble” is used to describe an uncertain prospect whose
mathematical expectation is zero. (A gamble with negative expectation is
called “unfavorable”; one with positive expectation is called “favorable.”)
For example, odds of 5:1 on a roll of a fair die represent a fair gamble: since
you lose (say) a dollar if the face you name does not come up, andwin  five
dollars if it does come up, the expectation of gain is (−1) 56 + 5 16 = 0.
Then a risk-averse person would refuse a fair gamble; a risk preferrer would
accept a fair gamble; and a risk-neutral person would be indifferent.6
Figure 1.4 displays three possible utility functions: v1 (c) would apply to
a risk-averse individual, v2 (c) to someone who is risk neutral, and v3 (c) to
a risk preferrer. Consider the fair prospect or gamble G = (750, 250; 12 , 12 )
whose mathematical expectation is $500. For the first or risk-averse individ-
ual the utility of $500 certain, v1 (500), is indicated by the height of point T
along the v1 (c) curve. The utility he attaches to the risky prospect, choosing
the gamble G, is indicated by point L – whose height is the probability-
weighted average of the heights of points J and K. This is, of course,
the geometrical equivalent of the expected-utility rule, which tells us that
U1 (G) = 12 v1 (750) + 12 v1 (250). Evidently, whenever the utility function
has the “concave” shape of v1 (c), points associated with a certainty income
(like T in the diagram) will be higher than points (like L) representing a fair
gamble with the same expectation of income. By an analogous argument,
6
However, as we shall see below, a risk-averse individual would accept a fair gamble if it
offset other risks to which he was exposed. To purchase insurance, for example, is to accept
an offsetting (risk-reducing) gamble.
1.5 Risk Aversion 27

Figure 1.4. Attitudes toward risk.

for the risk-preferring individual, v3 (500) at point Y will be less than at


point W; such a person would choose the gamble G rather than receive its
mathematical expectation of income, $500, as a certainty. Finally, the v2 (c)
curve indicates that the risk-neutral person would be indifferent between
the gamble G and the certainty of $500.
We will often have occasion to make use of Jensen’s inequality: If c̃ is a
random variable (taking on at least two values with non-zero probability)
and v(c) is a twice-differentiable function:

If v (c) < 0, then E[v(c)] < v[E(c)]


If v (c) = 0, then E[v(c)] = v[E(c)]
If v (c) > 0, then E[v(c)] > v[E(c)]

Evidently, these conditions correspond immediately to the risk-averse, risk-


neutral, and risk-preferring cases of Figure 1.4.
It is useful to consider how attitude toward risk is reflected in the triangle
diagram introduced in Section 1.4.1. Figure 1.5 below shows the set of
lotteries over three income levels, $0, $500, $1,000. Thus, π1 is the probability
of getting $0 and π3 is the probability of $1,000. The origin corresponds to
getting $500 for sure.
Point L in Figure 1.5 is the lottery that gives either $1,000 or nothing
with equal probability. A risk-neutral individual is indifferent between this
prospect and $500 for sure. Hence, the solid line joining L to the origin is
the indifference line through L for a risk-neutral person. As indifference
28 Elements of Decision under Uncertainty

π3

D: (1 − π*, π*)

L: (0.5, 0.5)

π1
0 1
Figure 1.5. Risk-averse and risk-neutral indifference lines.

lines under the expected-utility rule are parallel straight lines, the solid lines
inside the triangle are indifference lines under risk neutrality. That is, all
lotteries on a solid line have the same expected value and the same expected
utility for a risk-neutral person.
Any risk-averse individual strictly prefers $500 for sure to the prospect
L. Thus, since the direction of increasing preference is to the northwest,
any risk-averse indifference line through the origin must intersect the
hypotenuse of the triangle at a point D to the northwest of L (π ∗ > 0.5).
Hence, indifference lines for a risk-averse person (the broken lines in the
triangle) are steeper than the indifference lines for a risk-neutral person (the
solid lines in the triangle). Similarly, the indifference lines for a risk-neutral
person are steeper than indifference lines for a risk-preferring person.
We now consider what observation of the world tells us about the actual
v(c) curves entering into people’s decisions. First of all, we have already
postulated that more income is preferred to less, justified by the observation
that only rarely do people throw away income. This implies a rising v(c)
function, with positive first derivative v  (c), that is, positive marginal utility
of income. The question of risk aversion versus risk preference concerns the
second derivative v  (c) – whether marginal utility of income falls or rises
with income.
Risk aversion – “concave” curves like v1 (c) displaying diminishing
marginal utility – is considered to be the normal case, based upon the
1.5 Risk Aversion 29

observation that individuals typically hold diversified portfolios. Suppose


someone were merely risk neutral, so that for him v  (c) = 0. Then he
would ignore the riskiness or variance of different investment options or
assets (gambles), and take account only of the mathematical expectation
of income associated with each. Such a person would plunge all his wealth
into that single asset that, regardless of its riskiness, offered the highest
mathematical expectation of income. But we scarcely ever see this behavior
pattern, and more commonly observe individuals holding a variety of assets.
Since the risks associated with different assets are generally partially offset-
ting, diversification reduces the chance of ending up with an extremely low
level of income. This safety feature is achieved, however, only by accepting a
lower overall mathematical expectation of income; some expected income
has been sacrificed in order to reduce risk.7
What of the seemingly contrary evidence that “unfavorable” (negative
mathematical expectation) gambles are cheerfully accepted by bettors at Las
Vegas and elsewhere? Even more puzzling, why is it that the same person
might behave quite conservatively (insure his house, diversify his asset
holdings) in some circumstances, and in other circumstances accept fair
or even unfavorable gambles? There have been attempts to construct utility
functions v(c) that would be consistent with avoiding gambles (insuring)
over certain ranges of income and with seeking gambles over other ranges
(Friedman and Savage, 1948; Markowitz, 1952). We will briefly discuss the
Friedman-Savage version.
Consider the doubly inflected utility function in Figure 1.6. The v(c)
curve is concave, reflecting normal risk aversion, in the region OK and once
again in the region LN. But it is convex, reflecting risk preference, in the
middle region KL. With this sort of v(c) function, risk-taking behavior will
vary with wealth. For those whose endowments fall in the first concave
segment, the tendency is to insure against relatively small risks but to accept
fair (or even mildly adverse) long-shot big-payoff gambles, offering a chance
of landing somewhere toward the upper end of the curve. It can be verified
that this pattern will particularly apply for those with incomes toward the
upper edge of the bottom segment – the less indigent poor, and perhaps the
lower-middle class. The very poor, in contrast, would be much less inclined
to gamble. Looking now toward the top of the scale, those with incomes near
the lower edge of the upper concave segment – the rich but not super-rich,

7
An individual characterized by risk-preference might also plunge all of his wealth into
a single asset, but this need not be the asset with the highest mathematical expectation
of income. He might choose an asset with greater riskiness over the asset with high-
est income yield (that is, he would sacrifice some expected income in order to enlarge
his risk).
30 Elements of Decision under Uncertainty

Figure 1.6. Gambling and insuring – doubly inflected utility function.

and perhaps the upper middle class – would seem to have a taste for risks
likely to have a favorable payoff but offering a long-shot chance of a really
large loss. (But the super-rich, like the super-poor, are very disinclined to
gamble at all.) The central group, finally, would be happy to accept almost
any fair or not-too-unfavorable gamble.
The doubly inflected utility function of Figure 1.6 does then explain
why a person might gamble in some circumstances and insure in others, or
accept some fair gambles while rejecting other ones. But it also implies other
behavior that is quite inconsistent with common observation. It is hard to
believe that people of middling incomes are always great gamblers. If the
picture in Figure 1.6 were correct, the middle group in the convex KL seg-
ment would be so anxious to gamble as to seek out enormous riches-or-ruin
bets. These middle ranges of income would then rapidly be depopulated,
which is surely not what is observed. And that the really solid risk avoiders
in our society are only the very poor and the super-rich is equally difficult
to credit.
A more acceptable explanation, of why people simultaneously gamble and
insure, is that most of us engage in gambling as a recreational rather than
an income-determining activity. Put another way, gambling is normally
more like a consumption good than an investment good. As it happens,
it is quite possible operationally to distinguish recreational or pleasure-
oriented from serious wealth-oriented gambling. The latter, if efficiently
conducted, would take the form of once-and-for-all wagers at enormous
stakes. Pleasure-oriented gambling, in contrast, being designed to yield
enjoyment over some period of time, will be characterized by repetitive
1.5 Risk Aversion 31

minuscule bets practically guaranteed not to change one’s wealth status


in any drastic way. What is observed at Las Vegas is very much more the
repetitive small-stake than the riches-or-ruin huge-stake betting pattern.
Nevertheless, in exceptional situations, risk-preferring behavior does
indeed surely occur. Consider the following. As bank cashier you have
dipped into the till to the extent of $30,000. The bank examiners are arriv-
ing tomorrow, so you have time to replace the missing funds, but you have
only $10,000 left on hand. Suppose you value the consumption benefit of
spending the remaining $10,000 today far less than you value avoiding the
shame and pain of exposure as an embezzler. Then you surely would be
willing to risk the $10,000 on a fair gamble today – say, with a 13 chance
of winning $20,000. You would probably even take quite an adverse bet if
necessary, so long as the possible payoff sufficed to cover the $20,000 of
additional funds you need.
What is involved here is a “threshold” phenomenon, a critical level of
income where a little bit more can make a big difference. Put another
way, there is a range of increasing marginal utility – in the extreme, a
single discrete step to a higher utility level. Threshold phenomena are quite
common in nature. In many species, animals must take risks in accumulating
resources or engaging in combat in order to achieve nutritional viability or
win the privilege of mating. These phenomena have evident analogs for
humans living in primitive societies. To what extent they may explain risk-
taking behavior under modern conditions may be left an open question.8
This discussion may possibly suggest, contrary to a point made earlier,
that it is after all true that utility must “really” be cardinal. A viability
threshold, for example, might seem to be a cardinal feature of preference that
would apply to riskless as well as to risky decision making. Nevertheless, our
original point remains valid. For certainty choices, only ordinal comparisons
of consequences are needed. For decisions under uncertainty we can derive,
by the reference-lottery technique, a v(c) function that may have convex
or concave or mixed curvature, as the case may be. But the shape of this
function for any individual is an inseparable blend of two elements: (i) the
individual’s valuations of the consequences, and (ii) his attitudes toward

8
See Rubin and Paul (1979). These authors suggest that the propensity of young males to
engage in highly risky activities – as evidenced, for example, by their high automobile
accident rates – may be the result of natural selection for risk-taking. The evolutionary
history of the human species may have instilled risk-preferring attitudes among individuals
in age and sex groups liable to encounter viability or mating thresholds. (Note that the
threshold argument is also consistent with the observation that risk-taking behavior will
be observed predominantly among the poor.)
32 Elements of Decision under Uncertainty

risk. We may therefore interpret a concave v(c) function as reflecting either


risk aversion (attitude toward risk) or diminishing marginal utility (attitude
toward income); similarly, a convex v(c) function can be said to reflect
either risk preference or increasing marginal utility. Both terminologies are
somewhat misleading, since what the curvature of v(c) really represents is
the interaction of the two factors working together.
Finally, another category of seeming risk-taking behavior may be explain-
able in terms of state-dependent utility functions. An example: Suppose it is
very important to me, as the psychological equivalent of having a large sum
of money, that the home team wins the big game. Then I might plausibly
bet against the home team, at fair or even adverse odds! (How this works
out in detail will be left for the chapter following.)

Exercises and Excursions 1.5


1 Risk Aversion, Risk Preference, Risk Neutrality
(A) Identify each of the following “cardinal” utility functions with risk-
averse, risk-preferring, or risk-neutral behavior:
(i) v = ln c (ii) v = ac – bc 2 (a, b positive constants)
1
(iii) v = c 2 (iv) v = c 2
(v) v = 100 + 6c (vi) v = 1 – e −c
(B) The quadratic form (ii) above has an unsatisfactory feature for
c > a/2b. Explain.

2 Diversification
Three individuals have respective utility functions v1 = c (risk neutral),
v2 = c 0.5 (risk averse), and v3 = c 2 (risk preferrer). They each have the
option of investing in any one of the three following prospects or gambles,
with mathematical expectations of income as shown:
G1 = (480, 480; 0.5, 0.5) E[G1] = 480
G2 = (850, 200; 0.5, 0.5) E[G2] = 525
G3 = (1,000, 0; 0.5, 0.5) E[G3] = 500
Notice that, comparing the first two gambles, higher risk is associated with
greater mathematical expectation of income. The third gamble has highest
risk of all, but intermediate mathematical expectation.
(A) Show that risk-neutral individual 1 will prefer gamble G2 with the
highest expectation, while risk-averse individual 2 will prefer gamble
1.5 Risk Aversion 33

G1 with the lowest risk. Show that the risk-preferring individual 3 is


willing to sacrifice some expectation to increase his risk, by choosing
G3.
(B) If the individuals could “diversify” by choosing any desired mixture
of these gambles, which of them would diversify? (Assume that the
payoffs of gambles G2 and G3 are perfectly correlated.)

3 Doubly Inflected Utility Function


In the doubly inflected v(c) curve shown in Figure 1.6, suppose that the
borders of the segments (inflection points) occur at c = 250 and at c = 750.
(A) Illustrate geometrically that an individual with initial income of
$240 would be likely to accept a (fair) gamble offering a one-sixth
chance of a $600 gain and a five-sixth chance of a $120 loss. Show
that someone with initial income of $120 would be much less likely
to accept the same gamble.
(B) Show that someone with initial endowed income of $760 would be
likely to accept a fair gamble which is the reverse of the above: a five-
sixth chance of a $120 gain and a one-sixth chance of a $600 loss.
What about a person with initial wealth of $880?
(C) Show that someone with endowed wealth of exactly $500 would
surely accept any fair gamble with 50: 50 odds – at least up to a scale
of $250 gain and $250 loss. He might even accept much larger fair
gambles of this type; indicate geometrically the limits of what he
would accept.

4 Linear Risk Tolerance


Risk aversion is characterized by the condition v  (c) < 0. For some pur-
poses, as we shall see below, the ratio −v  /v  is a useful measure of risk
aversion. The reciprocal of this ratio, −v /v , is known as the risk tolerance.
An interesting class of v(c) functions is defined by the condition of linear
risk tolerance: −v /v = α + βc.
(A) Show that, for arbitrary constants M, N with N > 0:
(i) β = 0 implies v = M − Ne−c/α
(ii) α = 0, β  1 implies v = M + Nc 1−γ /(1−γ )) where γ = 1/β
(iii) α = 0, β = 1 implies v = M + N ln c
(iv) α > 0, β = −1 implies v = M − N (α-c)2
(B) Some of the above functions are valid only in restricted ranges of c.
Indicate the restrictions, if any, that apply in each case. Also explain
why N must be positive if v is to be a well-behaved utility function.
34 Elements of Decision under Uncertainty

5 The Bank Examiner Is Coming


You have stolen $30,000 from the bank but have the opportunity to replace
it by winning a fair gamble. You have at your disposal just $10,000. Your
utility function is such that v(c) = −B, where B is a very big number, when
c < 0 (i.e., should you not replace all the missing funds), and otherwise
1
v(c) = c 2 . Assuming fair gambles are available at any terms you desire,
solve geometrically for your optimal fair gamble. Will you surely stake all
your $10,000? Will you look only for a $20,000 payoff, or would you prefer
a bet with a smaller chance of a bigger payoff?

6 Utility Functions with Multiple Goods


The argument in the text above, developing a cardinal utility function v(c)
for use with the expected-utility rule, ran in terms of a single desired good
or commodity c. Extend the argument to cardinal utility functions of two
goods, in the form v(a, b). Show that, starting with an ordinal preference
function defined over combinations of a and b (that is, starting with an
ordinary indifference map on a, b axes), the reference-lottery technique can
be used to generate a cardinal scaling that amounts to giving a numerical
utility value to each indifference curve.

7 Risk Aversion with Multiple Goods 1 1


An individual has a utility functionv (a, b) = a 2 b 4 . He has income I avail-
able for spending on a and b, and faces fixed prices Pa = Pb = 1.
(A) Show that he would strictly prefer the certain income of 50 to an
equal chance of his income rising or falling by 49 before he makes
his consumption choices.
(B) Obtain an expression for the individual’s “indirect” utility function.
(That is, the maximized level of v given income I and prices Pa and
Pb .) Hence show that this individual exhibits aversion to income
risks.
(C) Suppose I = 50 and Pb = 16. Would the individual prefer to face a
certain Pa = 64 or a stochastically varying Pa that might equal 1 or
81 with equal chances? Does your answer cast doubt upon whether
the individual is really risk averse? Explain.

8 Jensen’s Inequality (I)


(A) If the utility function v(c) is twice continuously differentiable with
v (c)  0, show that for any random variable c̃:
E[v(c̃)] ≤ v(E[c̃])

You might also like