Economics 500, Fall 2021

Problem Set 6, Due Tuesday, October 14

Question 1 We start with some issues in risk attitudes.

1.1 Consider a Qsimple form of the Cobb-Douglas utility function, given

by U (x) = L α
1 xℓ . The simplification here is that the exponent α is
common across good. Derive the indirect utility function. Identify
for what values of α the function is concave in w, and for which it is
convex in w. Us this to characterize when this person is risk averse,
neutral or seeking in income.
Solution As we already calculated many times the demand function
is
α w
xi (p, w) = ,
Lα pi
Y w
v(p, w) = ( )α .
Lpi
i

When α > 1 the indirect utility function is convex in w and the person
is risk seeking in income. When α < 1 the indirect utility function is
concave in w and the person is risk averse in income. When α = 1 the
person is risk neutral.

1.2 Give the definition of risk aversion and risk seeking in income (or
wealth) in terms of the comparison of the expected utility of a lottery
and the utility of its expected value. Offer a similar definition in terms
of risk aversion over prices. Use your indirect utility function from [1.1]
to identify when this person is risk averse, neutral or seeking in prices.
Solution We can define that a person is risk seeking/averse/neutral if
the the indirect utility function is convex/concave/linear in p. From
[1.1] we know

∂v v
= −α ,
∂pi pi
∂2v v ∂2v v
2 = α(α + 1) 2 , = α2 .
∂pi pi ∂pi pj pi pj
 We prove v is convex in p by showing that the Hessian H is positive
semi-definite. For this, we that that for any real vector (y1 , ..., yn ) ∈
Rn , we have yHy ≥ 0, which is
X vyi2 X 2 vyi yj
α(α + 1) + α ,
i
p2i pi pj
i̸=j
 y2 X yi yj 
i
= αv +α ,
p2i ij
pi pj
 y2 X yi 
= αv i
+ α( )2 ≥ 0.
p2i pi
i

Thus under the utility function of [1.1] the person is always risk seeking
in price p.

1.3 Consider the expenditure minimization problem for a fixed utility level
u. Give the definition of risk aversion and risk seeking in prices in
terms of the comparison of the expected expenditure of a lottery and
the expenditure of its expected value. Establish whether or when the
expenditure minimizer will be risk averse neutral or seeking in prices.
Solution The lottery l in this problem is a lottery over price. Notice
here the consumer is minimizing the cost. The lower the cost is, the
better the consumer feels. Thus a consumer is risk seeking in price if
X X
l(pi )e(pi , u) ≤ e( l(pi )pi , u).
i i

This is equivalent to the concavity of the e in p. As we know this is

always satisfied.

1.4 Consider an expected utility maximizer. As is customary, assume there

is a finite set of alternatives X and a Bernoulli utility function u : X →
R. Now introduce a finite state space Θ, where the probability of state
θ is p(θ), and associate with each state a lottery over X. Now think
about how the expected utility maximizer evaluates these lotteries.
Give the definition of risk aversion and risk seeking in probabilities
in terms of the comparison of the expected utility of a lottery over
probabilities and the expected utility of the expected probabilities.
Establish whether or when the expected utility maximizer will be risk

2
 averse neutral or seeking in probabilities. If it helps, it is fine to keep
things simple by assuming there are only two states.
Solution When we typically talk about risk attitudes, we examine lot-
teries over X, say with q(x) denoting the probability
P of alternative X,
with lottery q evaluated by the expected utility x q(x)u(x). Now we
are pushing this back one step and considering a lottery over lotteries.
There are different ways of doing this, but one that helps keep things
straight is to introduce the state space Θ, with each state θ ∈ Θ ap-
pearing with probability p(θ), and then to associate a lottery over X
with each state theta. To pursue this, we first define what is a lottery.
X
L = {l : Θ × X → [0, 1] with {x : l(θ, x) > 0} finite l(θ, x) = 1 ∀θ}
x

For an expected utility maximizer, the utility of a lottery is

X X
U (l) = p(θ) l(θ, x)u(x),
θ∈Θ x∈X
X X 
= p(θ)l(θ, x) u(x)
x∈X θ∈Θ

We can define a consumer to be risk averse if

X X  XX 
U (l) = p(θ)l(θ, x) u(x) ≤ u p(θ)l(θ, x)x ∀l.
x∈X θ∈Θ x∈X θ∈Θ
P P
Denote λx = θ∈Θ p(θ)l(θ, x), we know x λx = 1 and λ has finite
support. The above inequality is equivalent to
X X
λx u(x) ≤ u( λx x), ∀λ ∈ ∆′ X,
x x

where we use ∆′ X to denote all the finitely supported distribution in

X. The above definition is equivalent to the concavity of u. Similarly
we define the a consumer to be risk seeking if the inequality is reverse.
What we see here is that probabilities over probabilities can be reduced
to ordinary probabilities, at which point our usual analysis of risk
aversion applies.

1.5 Let Θ be a finite set of states and X a set of alternatives. The function
u : Θ × X → R is Bernoulli a utility function with u(θ, x) giving the

3
 utility of alternative x in state θ. The probability of state θ is p(θ).
Explain how the Bernoulli utility functions here and in [1.4] differ.
Give an example of a case in which each is applicable.
Solution The utility here depends on both the consumption bundle
and the state, while the utility in [1.4] depends only on the consump-
tion bundle. It is most common to model utility as depending only on
the consumption bundle, but there is no conceptual difficulty is having
it also depend on the state.
For an example of [1.4], imagine you invest all your stipend in Bitcoin
(which you probably shouldn’t). The state is unknown so the value of a
Bitcoin might double or halve in a month. Your utility of consumption
next month per se does not depends on whether Bitcoin goes up or
down. However your consumption does depend on how much money
you have and thus indirectly affected by the value of Bitcoin.
For an example of [1.5], imagine you are going to a outdoor barbecue.
Then your utility does not only depend on the quality of the food (the
consumption bundle) but also on weather (the state).

1.6 Continuing with the setting in [1.5], suppose that an alternative must
be chosen before the state is known, and write the attendant utility
maximization problem. Then suppose that an alternative x can be
chosen after the state is known, and write the utility maximization
problem. Show that the expected utility in the latter case is always
larger than in the former. Construct an example in which the differ-
ence in expected utilities is strict. Explain how these results can be
interpreted as showing that “more information is always better.”
Solution The first maximization problem is
X
max p(θ)u(θ, x).
x∈X
θ∈Θ

The second maximization problem is

X
p(θ) max u(θ, x).
x∈X
θ∈Θ

The expected utility in the latter case is always larger because denote
X
x∗ = argmax p(θ)u(θ, x).
x∈X θ∈Θ

4
 Then
X X X
p(θ) max u(θ, x) ≥ p(θ)u(θ, x∗ ) = max p(θ)u(θ, x).
x∈X x∈X
θ∈Θ θ∈Θ θ∈Θ

Consider the case where Θ = {1, 2} and X = {1, 2}, p(1) = p(2) = 0.5,
and

u(i, j) = 1i=j .

Then
X 1 1 1
max p(θ)u(θ, x) = max{ , } = .
x∈X 2 2 2
θ∈Θ
X 1 1
p(θ) max u(θ, x) = + = 1.
x∈X 2 2
θ∈Θ

1.7 Repeat your answer to [1.4] for the utility function introduced in [1.5]
and the two cases considered in [1.6]. Explain when the person is risk
averse, neutral and seeking.
Solution We start with the case where the consumer can choose an
alternative after the state is known. Then we can define a consumer
to be risk averse if
X X
λx u(θ, x) ≤ u(θ, λx x), ∀λ ∈ ∆′ X, ∀θ.
x∈X x∈X

Again ∆′ X represents the set of all finitely supported distribution.

This is equivalent to u being concave in x for any θ. Similarly we
define the a consumer to be risk seeking if the inequality is reverse.
When the consumer must choose alternative before the state is known.
We define a consumer to be risk averse if
X X X
λx,θ u(θ, x) ≤ u( λx,θ θ, λx,θ x), ∀λ ∈ ∆′ (X × Θ).
(x,θ)∈X×Θ x,θ x,θ

This is equivalent to u being concave in x and θ jointly.

1.8 Identify the principle that explains your answer to the preceding ques-
tions.

5
 These questions all consider settings in which a person faces a lot-
tery, and evaluates the lottery by taking the expectation of a function
of lottery outcomes. The lottery can be over different objects (such
as consumption bundles or probabilities). and function can be over
different objects (such as just consumption bundles or consumption
bundles and states both; or over the utility of a consumption bundle
chosen before the state is observed or the optimal consumption bun-
dle chosen after the state is observed), and the function itself can take
different shapes (most notably, concave or convex). In each case, we
say that a person is risk averse (with risk seeking being the reverse) if
the person prefers the expected outcome of the lottery to the lottery.
This invariably comes down to a question of whether the evaluation
function is concave.

Question 2. Now we focus on risk aversion.

2.1 We defined the Arrow-Pratt measure of risk aversion. This is often
referred to as a measure of absolute risk aversion. Write the cor-
responding measure of relative risk aversion. Explain how these two
measures are related—if one is constant in income or wealth, how does
the other behave?
Solution The coefficient of absolute risk aversion is
u′′ (x)
ra (x) = − .
u′ (x)
The coefficient of relative risk aversion is
xu′′ (x)
rr (x) = − .
u′ (x)
If ra is constant in x then ra (x) will go to infinity as does x, while if
rr stays constant then ra (x) will converge to 0.
1
2.2 Consider the Bernoulli utility function u(x) = 1−θ x1−θ . What are rea-
sonable values of θ, and for what values of θ does this utility function
exhibit risk averse, risk neutral, and risk-seeking preferences? Calcu-
late the coefficient of relative risk aversion and explain how it depends
on θ, and indicate why this is called a constant-relative-risk-aversion
utility function. Explain how you can work with the case of θ = 1.
Solution
u′ (x) = x−θ , u′′ (x) = −θx−θ−1 .

6
 We would like utility to be increasing in x, and so we would like θ ̸= 0.
Interestingly, we can accommodate both θ > 0 and θ < 0.
If we are interested in risk aversion, then we want the utility function
to be concave, which is the case if θ > 0.
x(−θx−θ−1 )
rr (x) = − = θ.
xθ
To address the case of θ = 1 we first do a strictly monotone transfor-
mation
1
u(x) = (x1−θ − 1).
1−θ
Now observe that for when θ → 1, both numerator and denominator
converge to 0. We can thus apply L’Hospital’s law:
− log x
lim u(x, θ) = = log x.
θ→1 −1
It is easy to check that, in this case the relative risk aversion is a
constant 1.
−θx
2.3 Consider the Bernoulli utility function u(x) = 1−eθ . What are rea-
sonable values of θ, and for what values of θ does this utility function
exhibit risk averse, risk neutral, and risk-seeking preferences? Calcu-
late the coefficient of absolute risk aversion and explain how it depends
on θ, and indicate why this is called a constant-absolute-risk-aversion
utility function. Explain how you can work with the case of θ = 0.
Solution
u′ (x) = e−θx , u′′ (x) = −θe−θx .
Again, for monotonicity, the case we need to exclude is θ = 0. We
want the utility to be concave (risk averse) in x thus it is reasonable
to assume θ ≥ 0.
−θe−θx
ra (x) = − = θ.
e−θx
Observe that when θ → 0, both the numerator and the denominator
converges to 0, thus we can apply L’Hospital’s law:
x
lim u(x, θ) = = x.
θ→0 1
It is easy to see that in this case ra (x) = 0.

7
2.4 It seems intuitive that for a risk aversion person, it is good for a lottery
over income or wealth to have a higher mean and a low variance.
Let’s investigate a setting in which we can make this precise. The
Bernoulli utility function is given by u(x) = −e−θx . Relate this utility
function to those considered in [2.2] and [2.3] and interpret θ. Now
consider a lottery over x given by a normal distribution with mean µ
and variance σ 2 . Then show that the expected utility of this lottery
can be written as a linear function of the mean µ and variance σ 2 , and
explain how this mean-variance representation captures differences in
risk aversion. (To do this, first write the expression for expected utility.
This will give you an integral over an expression whose form is e raised
to some rather unwieldy looking expression. Separate this expression
into those terms that do not contain x, and hence can be taken our of
the integral, and those that contain x and hence must remain. Then
argue the remaining integral can be shown to equal one, and then
simplify what’s left.)
Solution This is exactly the constant-absolute-risk-aversion utility
function that we just consider in [2.3] (up to a strictly monotone trans-
formation). θ is the coefficient of the absolute risk aversion.
Z +∞ 2
1 1 (x−µ)
Eu(X) = − e−θx √ e− 2 σ2 dx
∞ 2πσ
Z +∞ 2 2 2
1 1 x −2(µ−θσ )x+µ
=− √ e− 2 σ2 dx
∞ 2πσ
Z +∞ 2 2 2 2
1 1 x −2(µ−θσ )x+(µ−θσ ) µ2 −(µ−θσ 2 )2
=− √ e− 2 σ2 e− 2σ 2 dx
∞ 2πσ
(µ−θσ 2 )2 −µ2 θ 2 σ 4 −2µθσ 2 2 /2)
= −e 2σ 2 = −e 2σ 2 = −e−θ(µ−θσ

Hence equivalently we can say that the consumer is trying to maximize

µ − θσ 2 /2. Observe the expected utility is increasing in the mean
µ while decreasing in the variance σ 2 . Also, the variance term σ 2
gains higher relative weight as θ increases. This is consistent with the
interpretation that higher θ represents higher risk aversion.

2.5 Let Θ be a finite set of states and p a prior distribution over θ. Consider
a person whose Bernoulli utility function u : R → R exhibits risk
aversion. This person’s current allocation is given by x(θ). Suppose
this person can purchase, at zero cost, any lottery z(θ) that has zero
expected value. Characterize the utility-maximizing lottery. In light

8
 of your result, assess the following statement: “Risk aversion people
in competitive markets will fully ensure.”
Solution The person will solve the following maximization problem:
X
max p(θ)u(x(θ) + z(θ)),
z()
θ
X
s.t. p(θ)z(θ) = 0.
θ

Denote y(θ) = x(θ) + z(θ) the problem is equivalent to

X
max p(θ)u(y(θ)),
y()
θ
X X
s.t. p(θ)y(θ) = p(θ)x(θ).
θ θ

Since the consumer is risk averse we know u is concave and thus

X X
p(θ)u(y(θ)) ≤ u( p(θ)y(θ))
θ θ
X
= u( p(θ)x(θ)).
θ

Also notice that the constant consumption

X
y ∗ (θ) = p(θ)x(θ)
θ

is feasible. Thus it is optimal to choose y ∗ (θ) or equivalently choose

z ∗ (θ) = y ∗ (θ) − x(θ).
The statement relies on the assumption that consumer “can purchase,
at zero cost, any lottery z(θ) that has zero expected value”. As we
will learn in the general equilibrium or macro economic class, this
condition holds when there is only idiosyncratic risk. When there is
also systematic risk (risk about the whole economy) this is typically
not satisfied.

Question 3 Here, we look at some of the challenges to expected utility

theory.

9
3.1 In response to the Allais paradox, it is often suggested that axioms A1-
A3 should be weakened. The typical target is the independence axiom.
One suggested axiom is the betweenness axiom: If p and q are lotteries
with p ∼ q and α ∈ [0, 1], then αp + (1 − α)q ∼ p. Give your best
argument for why you expect betweenness to hold. Present an example
in which you think independence is problematic but betweenness is
reasonable. Show that independence implies betweenness, but that
the converse can fail. Show that the choices associated with the Allais
paradox are consistent with betweenness.
Solution

Argument Suppose p ∼ q so you are indifferent between two lotter-

ies. Now imagine there is a compound lottery αp + (1 − α)q that gives
you the lottery p with probability α and the lottery q with probability
1 − α. When facing with this lottery, you should feel equally happy
about either lottery it offers. Thus you should be indifferent in the
value of α.

Example As we will show later, Allais paradox is exactly an example

where independence is problematic while betweenness is reasonable.
For another example, imagine you are planing for your thanksgiving
holiday. You can either play video games in your apartment alone
for the whole holiday (p), or go back home to spend the holiday with
your family (q). Suppose you are indifferent between these two plans.
Then it seems reasonable that you are equally happy with a random
plan which send you back home with probability α. Now let us in-
troduce another event (r). Suppose there is 10% probability that you
will be murdered at the start of the holiday by a serial killer. If you
survive, the killer will be caught and you can choose to go back home
to spend thanksgiving with your family (0.9q) or play video games in
your apartment alone for the whole holiday (0.9p). Then it is very
reasonable that people will strictly prefer 0.1r + 0.9q to 0.1r + 0.9p.
What goes wrong here? A common argument for the independence
axiom is as follow. There is only one outcome p, q, or r that actually
happens in the end, and when the plan (0.1r + 0.9p) actually differs
from (0.1r + 0.9q), r does not happen anyway. Thus you preference
should remain unchanged. However, in this example even if r does not
happens, the existence of the possibility of being killed and the fact

10
that we survive changes our preference for later decisions. We simply
want to share the happiness of surviving with family in thanksgiving
holiday.

Relationship with independence Next we prove independence

implies betweenness

p ∼ q =⇒ p ⪰ q & q ⪰ p.

Let p be the r in the axiom of independence we have

αp + (1 − α)p ⪰ αp + (1 − α)q,
αp + (1 − α)q ⪰ αp + (1 − α)p.

Thus

αp + (1 − α)q ∼ αp + (1 − α)p = p.

On the other direction, we provide the following counterexample. Con-

sider the lottery for three outcome: {(p1 , p2 , 1−p1 , p2 )|pi ≥ 0, p1 +p2 ≤
1}. Consider the following preference

(x, 0, 1 − x) ∼ (0, x2 , 1 − x2 ) ∼ (λx, (1 − λ)x2 , 1 − λx − (1 − λ)x2 ),

∀x ∈ [0, 1], λ ∈ [0, 1];
(0, x, 1 − x) ≻ (0, y, 1 − y), ∀x > y.

This preference satisfies between axiom by construction. The indif-

ference curve is a straight segment that connects (x, 0, 1 − x) with
(0, x2 , 1 − x2 ). Importantly, the slope of these straight indifference
curve is different. On the other hand, if this preference the axiom of
independence, we must have

(1, 0, 0) ∼ (0, 1, 0) =⇒ (0.5, 0, 0) + (0, 0, 0.5) ∼ (0, 0.5, 0) + (0, 0, 0.5).

But by construction we know

(0.5, 0, 0.5) ∼ (0, 0.25, 0.75) ≺ (0, 0.5, 0.5).

This is a contradiction.

11
 Allais paradox The profile in Allasis paradox is summarised in the
table . We want to rationalized the preference A ≻ B and D ≻ C. In
fact it can be done by the preference we just constructed above. We
have

C = (0.11, 0, 0.89) ∼ (0, 0.112 , 1 − 0.112 ) ≺ D,

A = (1, 0, 0) ∼ (0.9, 0.1, 0) ≻ B.

The last strict preference in both lines comes from the fact that the
constructed preference is strictly increasing in the first two coordinates.

1 5 0
A 1 0 0
B .89 0.1 .01
C .11 0 .89
D 0 .1 .9

Table 1: Allais Paradox

3.2 Suppose an urn contains 100 balls, each of which is either red or green,
but you have no information as to the proportion of red or green balls.
One ball is to be drawn from the urn. You must choose one of the
following three lotteries: (a) receive a payoff of 1 (add some zeros to
make it interesting if you would like) with probability .49, no matter
what color ball is drawn; (b) you receive a payoff of 1 if a red ball is
drawn; (c) you receive a payoff of 1 if a green ball is drawn. Many
people are indifference between lotteries b and c, but prefer lottery a
to both b and c. Now suppose you encounter the spirit of Howard
Raiffa, who argues, “It is irrational to prefer a to b or c. If you could
announce you choice of b or c until after a ball was drawn but before its
color was revealed, you could flip a coin between b and c and ensure
a payoff of 1 with probability half, no matter what the color of the
ball, giving an outcome that first-order stochastically dominates a and
hence which you would surely prefer to a. But then it cannot matter
whether you flip the coin before or after the ball is drawn, and so
flipping a coin and then choosing b or c as appropriate must be better
than a.” Explain why, since there are comparing lotteries with only
two outcomes, risk attitudes are irrelevant here. Then explain why
you do or do not find this reasoning convincing.

12
Solution Denote the utility of receiving 1 as u(1) and the utility of
receiving nothing as u(0). Suppose the consumer has a prior ρ on the
number of red balls in the urn:

ρ(i) = P(there is i red balls in the urn).

Then the expected utility of a profile (pa , pb , pc ), which means choos-

ing a with probability pa , choosing b with probability pb and c with
probability pc is
X i 100 − i

U (pa , pb , pc ) = pb ρ(i)( u(1) + u(0))
100 100
i
X i 100 − i

+pa (0.49u(1) + 0.51u(0)) + pc ρ(i)( u(0) + u(1)) .
100 100
i

Notice that any u such that u(1) > u(0) will induce a same ranking
over all profiles (pa , pb , pc ) thus the risk attitude does not matter. In
particular,
1
U (0, 0.5, 0.5) = (u(0) + u(1)) > 0.49u(1) + 0.51u(1) = U (1, 0, 0).
2

Some people find this argument convincing, others do not. This argu-
ment is convincing if you believe what matters for the decision of the
consumer is the probability of the eventual outcome. Under this view,
a consumer shall make the same decision under the following to two
prior ρ1 and ρ2 :

ρ1 (50) = 1, ρ1 (i) = 0, ∀i ̸= 50;

ρ2 (i) = 1/100 ∀i.

Under both prior, the probability that a picked ball is red is just
1/2. However, one can argue that there is much more uncertainty or
ambiguity about the whole system when the consumer have prior ρ2 .
Put it differently, the consumer might also dislike the extra randomness
introduced by the lottery of lotteries.

13