Advanced Topics in Microeconomics

Choice Under Uncertainty

Duozhe Li
Department of Economics
Chinese University of Hong Kong

Spring 2023
1 Outcomes and Lotteries

• Many economic decisions contain some element of

uncertainty (more precisely, risk). That is, the deci-
sion maker (DM) is not certain about the outcome
of the choice that is made, but the DM may know
the probabilities of different possible outcomes.

• Consider a finite set of outcomes

C = {1 2   } 

Each  is a deterministic (no uncertainty) outcome,
which may refer to a consumption bundle, a certain
amount of money, or anything at all.

• Definition Fix C, a simple lottery is a probability

distribution over C :

 = (1 2   )

P
with  ≥ 0 for all  and =1  = 1, where 
is the probability that outcome  occurs.
• Example Suppose that you make the following bet
with a friend: if the toss of a fair coin comes up
heads, she pays you 10 dollars, and you pay her 10
dollars if it comes up tails. Then, C = {−10 10},
and the bet is a simple lottery  = (12 12),
which can be depicted by a tree
10
1/2
Lottery 1
1/2 -10

• Definition A lottery that has a simple lottery as a

possible “outcome” is called a compound lottery.

• Example Lottery 2 below is compound lottery: with

probability 23, the DM receives Lottery 1 as de-
scribed above; with probability 1/3, the DM receives
 nothing. Now the outcome set is C = {−10 10 0}.
Lottery 1
2/3
Lottery 2
1/3 0

• A lottery can have many levels of compounding, that

is, the “outcome” of a compound lottery can also
be a compound lottery. For simplicity, we rule out
infinitely layered compound lotteries, that is, a com-
pound lottery must result in an outcome in C after
finitely many randomizations.

• A compound lottery can be reduced to a simple lot-

tery by computing the probability of each outcome
resulting from the sequential randomizations.
1/3 10

1/3
Lottery 2 -10
1/3
0
• More precisely, fix C, given  simple lotteries
³ ´
  
 = 1  2    
with  = 1 2  , and probabilities  ≥ 0 with
X
 = 1
=1 
the compound lottery, written as
(11 22    )
can be reduced to a simple lottery
∗ = (∗1 ∗2  ∗ )
with
∗ = 11 + 22 +  +  


• If the compound lottery contains multiple levels of

compounding, we can calculate the corresponding
reduced simple lottery similarly. (Exercise)

• The DM is assumed to be a consequentialist, i.e.,

he views a compound lottery and its reduced simple
lottery as equivalent.
2 Expected Utility Theory

• The DM has a preference, %, over the set L of all

simple lotteries:

—  % 0: the DM prefers  to 0.

—  Â 0 ( % 0 and 0 ´ ): the DM strictly

prefers  to 0.

—  ∼ 0 ( % 0 and 0 % ): the DM is

indifferent between  and 0.

• Can the DM’s preference be represented by a utility

function  : L → R, that is,
³ ´
%  ⇔  () ≥  0
0

for all  0 ∈ L?
• The most well-known theory is due to von Neumann
and Morgenstern.

• Definition A preference % over L is said to be rep-

resented by a von Neumann-Morgenstern (vNM)
utility function  : L → R iff for all  0 ∈ L

 % 0 ⇔

X 
X ³ ´
 () ≡  () ≥  () ≡  0 
0
=1 =1
where  : C → R gives the DM’s utility from deter-
ministic outcomes.

• Proposition A utility function  : L → R is a vNM

utility function iff it is linear, that is, iff it satisfies
µX ¶ X

   =
=1  
  ( )
=1 
(∗)

for any  lotteries  ∈ L,  = 1   and prob-

P
abilities (1   ) ≥ 0, =1  = 1.
Proof. Suppose that  (·) satisfies
³ condition (∗).
´
1
Write  as a compound lottery 1     , 

where  refers to the degenerate simple lottery

which gives  with probability 1. Then by (∗),
µX ¶
 
 () =  =1
 
X X
= 
  ( ) =   () 
=1  =1 
Thus,  (·) is a vNM utility function.

Suppose that  (·) is a vNM. Recall that the com-

pound lottery (11 22 ³   ) can ´be re-
duced to a simple lottery ∗ = ∗1 ∗2  ∗ with
P
 = =1  . Then,
∗

µX ¶  µX
X ¶
    ( )
  
=1  
=  
=1   
=1
X 
X
= 
=1 
 ()
=1
X
=   ( ) 
=1 
Thus, condition (∗) is satisfied.
• The vNM representation has two crucial elements:

1.  : L → R represents % in the ordinal sense.

2. The representation has a cardinal property as

it takes the linear form. The DM acts as if he
maximizes the expected value of  (·).

• Next, we explore the conditions on the preference

under which the vNM representation is possible.

• We first assume that the DM has a rational prefer-

ence, i.e., Axiom 1 and Axiom 2.

• Axiom 1: Completeness. For any two lotteries, 

and 0 in L, either  % 0, or 0 % .

• Axiom 2: Transitivity. For any three lotteries, ,

0 00 in L, if  % 0 and 0 % 00, then  % 00.
• By Axioms 1 and 2, the finitely many elements in C
can be indexed such that 1 % 2 %  %  

• Axiom 3: Continuity. For any three lotteries , 0

00 in L with  % 0 % 00, there exists  ∈ [0 1]
such that 0 ∼  + (1 − ) 00.

• The implication of continuity is that the DM’s pref-

erence cannot exhibit “jump”.

• Example Let C = {$100, $1, death}. For a typical

DM: $100 Â $1 Âdeath, but
$1 Â  ≡ (1 − ) ◦ $100  ◦ death
for all   0.

— The DM is not willing to take any risk of his life

for the extra $99, which seems reasonable.

— However, as the preference exhibits a “jump” at

 = 0, it is not continuous.
• Axiom 4: Independence. For all  0 00 ∈ L
and  ∈ (0 1),  % 0 if and only if (iff)
 + (1 − ) 00 % 0 + (1 − ) 00

• The independence axiom is at the heart of the the-

ory of choice under uncertainty. It has some useful
implications:

1. For any  0 00 000 ∈ L with 00 ∼ 000 and

any  ∈ (0 1],
 % 0 ⇔ +(1 − ) 00 % 0+(1 − ) 000

2. For any  0 ∈ L with  Â 0 and any   ∈

[0 1] with   ,
 + (1 − ) 0 Â  + (1 − ) 0

3. For any  ∈ L,
1 %  %  
(Recall that the finitely many outcomes in C can
be indexed such that 1 % 2 %  %  )
• The independence axiom imposes a structure on the
DM’s indifference sets on L:

1. The indifference sets are straight lines (i.e. hy-

perplanes).

2. The indifference sets are parallel to each other.

• Consider the case with three outcomes. Recall that

L is the 2-dimensional simplex taking the form of an
equilateral triangle.
Direction of
Increasing
Desirability

Indifference Sets ( = 3)
• Suppose that the indifference set is not a straight
line:
Direction of
Increasing
Desirability

0.5L+0.5L’

L L’

0 ¡ 0 ¢
Then we have  ∼  but 05 + 05 Â ,
violating the independence axiom.

• Suppose that the indifference sets are not parallel:

L’’
0.4L+0.6L’’ 0.4L’+0.6L’’

Direction of
Increasing
Desirability

L L’

¡ ¢
Then we have  ∼ 0 but 040 + 0600 Â
¡ ¢
04 + 0600 , violating the independence axiom.
• Theorem A preference % over L can be represented
by a vNM utility function  : L → R if it satisfies
Axioms 1, 2, 3 and 4.

Proof. The outcomes in C can be indexed such that

1 % 2 %  %  (Axioms 1 & 2). For all all
 ∈ L, we have 1 %  %  (Axiom 4).

If 1 ∼  , the DM is indifferent between every-

thing, and thus  () = 1 for all  is a vNM repre-
sentation of %.

Now, assume 1 Â  . By Axioms 3 & 4, for all

 ∈ L, there exists a unique  ∈ [0 1] such that

 ∼  ◦ 1 + (1 − ) ◦  

Let  () = . Next we show that  (·) is indeed

a vNM representation of %.

First,  % 0 ⇔

 () ◦ 1 + (1 −  ()) ◦  %
³ ´ ³ ³ ´´
  ◦ 1 + 1 −  0 ◦ 
0
 ¡ 0¢
⇔  () ≥   . (Axiom 4)

It remains to show that  (·) is linear, that is

³ ´ ³ ´
  + (1 − )  =  ()+(1 − )  0 
0

By definition
 ∼  () ◦  + (1³ −  ())
³ ´ 1
◦
³ ´´ 
0 ∼  0 ◦ 1 + 1 −  0 ◦  
By Axiom 4:
 + (1 − ) 0 ∼h  ³[ ()
´
◦ 1 +
³
(1 −  ()) ◦ i]
³ ´´
+ (1 − )  0 ◦ 1 + 1 −  0 ◦ 
h ´i³
0
∼  () + (1 − )   ◦ 1
h ³ ´i
+ 1 −  () − (1 − )  0 ◦  
and then by the definition of  (·), we get
³ ´ ³ ´
  + (1 − )  =  ()+(1 − )  0 
0

To be complete, write  () as  (). Then, for

any simple lottery  = (1 2   ), we have
X X
 () =   () =
=1 
  () .
=1 
• Next theorem says that the vNM representation is
unique up to positive affine transformations.

• Theorem Suppose that  : L → R is a vNM rep-

resentation of %, then ̃ : L → R is also a vNM
representation of % iff there are some   0 and
 ∈ R such that ̃ () =  () +  for every
 ∈ L.

Proof. The “If” part is obvious (Do it by yourself).

“Only If”: For every  ∈ L, define  ∈ [0 1] by

 () =  (1) + (1 − )  ( )

 () −  ( )
thus,  = 
 (1) −  ( )
As  is a vNM representation of %,

 ( ◦ 1 + (1 − ) ◦  ) =
 (1) + (1 − )  ( )
and thus,

 ∼  ◦ 1 + (1 − ) ◦  
Then, since ̃ is also a vNM representation of %,

̃ () = ̃ ( ◦ 1 + (1 − ) ◦  )

= ̃ (1) + (1 − ) ̃ ( )
³ ´
=  ̃ (1) − ̃ ( ) + ̃ ( )
Plug in
 () −  ( )
 =
 (1) −  ( )
and rearrange the terms, we get

̃ () =  () + 
where
̃ (1) − ̃ ( )
 =
 (1) −  ( )
̃ (1) − ̃ ( )
 = ̃ ( ) −  ( ) 
 (1) −  ( )
• The expected utility theory (most importantly, its
central assumption, the Independence Axiom) is not
always convincing. The following example is perhaps
the most famous challenge to the theory.

• The Allais Paradox

This is a thought experiment. There are three out-

comes (monetary prizes)
First Prize Second Prize Third Prize
$2 500 000 $500 000 $0
The DM is subject to two choice tests.

The first consists of a choice between lotteries 1

and 01 :

1 = (0 1 0) and 01 = (010 089 001) 

The second consists of a choice between 2 and 02 :

2 = (0 011 089) and 02 = (010 0 090) 

• It is common for a DM to express the preferences

1 = (0 1 0) Â 01 = (010 089 001)

and

02 = (010 0 090) Â 2 = (0 011 089) 

• These choices are inconsistent with expected utility.

To see this, suppose there is a vNM representation of

such preference, and let 1 2 and 3 be the utility
values of the three outcomes.

Then by 1 Â 01, we have

2  0101 + 0892 + 0013

Adding 0893 − 0892 to both sides, we get

0112 + 0893  0101 + 0903

which means that the DM with a vNM utility func-
tion must have 2 Â 02.
 [Link] i
3 Risk Aversion
ti
3.1 Theory

• Why do people buy insurance?

• Why do people invest on risky assets?

• More specifically, if a DM is given the choice between

receiving a monetary lottery and receiving with cer-
tainty the expected amount of money of the same
lottery, it seems that he is often willing to choose
the latter; in other words, the DM displays aversion
to risks. This is relevant for many economic set-
tings. We now formalize the notion of risk aversion
and introduce its measurements and properties.
• We focus on lotteries over monetary outcomes. For-
mally speaking, the set of outcome is R, and a lot-
tery can be represented by a cumulative distribution
function (CDF),

[Link]
 : R → [0 1]  (
where  () is the probability that the monetary out-
come is less than or equal to . We denote by  the
density of  whenever it exists.
1
F (a) 0 F (b)
⼆
: ,

• We can still consider lotteries over a discrete set of

outcomes. For example, the following lottery

 = ( ◦ $10 (1 − ) ◦ $100)
assigns probability  and (1 − ) on two outcomes:
$10 and $100. Then, we have the cumulative dis-
tribution function:
⎧
⎪
⎨ 0 if   10
 () =  if  ∈ [10 100) 
⎪
⎩ 1 if  ≥ 100
 • Denote by L the set of all monetary lotteries. Con-
sider a DM with a rational preference % over L.

• Assume that the preference can be represented by

a vNM utility function  (·), which is differentiable
and 0 ()  0 for all .
※ …

• The expected utility of a lottery  is given by

Z 以
所
 ( ) ≡  () ≡  ()  () 
where  is the expectation operator under  .

• Meanwhile, the expected value of the monetary out-

come under  is

[Link]
Z
 () ≡  () 

• By comparing  () to  ( ()), one can learn

about the DM’s attitude towards risk.
TH TTH T_T
2 4 8
Sfopwter c. .

学⼼
gtcnad 主 厽卡 字 …

名
[Link] yeonhpn 。
 Zfulx ) 5
[Link] $
⼆

[Link] ) 。

Edxkhoro ,

⼆ 55

• Definition Let  (·) be a DM’s vNM utility function

for monetary lotteries. The DM is said to be ul EFL非 后 50

= 71

1. risk averse if  () ≤  ( ()) for all  ,

2. risk neutral if  () =  ( ()) for all  ,

3. risk loving if  () ≥  ( ()) for all  .

• The DM is said to be strictly risk averse or risk loving

if the inequality is always strict for nondegenerate
lotteries (i.e., lotteries that assign positive probability
to at least two distinct outcomes).

R → R is concave iff
Z
 ()  () ≤ 
µZ dgen [Link] fiton [Link]
• Jensen’s inequality: a one-variable function  :

 ()

for all  (·). In the discrete case, it becomes

¶
[Link]

[Link]
X µX ¶

  () ≤ 
=1 
 
=1  

u
"

(e) ⼆ 0
[Link]
nlioredwa
 • Thus, a DM is risk averse iff his utility function  (·)
is concave, and he is risk neutral iff  (·) is linear.

• Note that if  (·) is concave and twice differentiable,

then its second order derivative 00 (·) ≤ 0.

• Definition Given a vNM utility function  (·) and a

lottery  (·), we define the following:

1. the certainty equivalent, denoted by  ( | ),

is an amount of money, offered with certainty,
such that
Z
 () =  ()  () =  ( ( )) 
⼀、

2. the risk premium  is an amount of money such

that  () =  ( () −  ). Clearly,
lheckcaveis eàetef rhaeu .
,

 ≡  () −  ( | ) 
→

ofhnctoe

[Link]
• Thus, a DM is risk averse iff  ( | ) ≤  ()
for all  .

Erluk 竑 .tt
$ 1 000 0

。
wlx ) = dxdiso
以
⼆ 55、
以⼋ ⼆
(d) (d) ,
uk ECFHK 5 5、 rìskdh
d El →

sfrìhyriddke
30251
CECF ⼼ 巧 ⼦
Mu
⼆ ˋ

→
5050-3025 2015
p
=
=
,
 • It is sometimes useful to quantify the degree of risk
aversion. We introduce two important measures of

[Link]-ist [Link]
risk aversion.

• Definition The Arrow-Pratt measure of absolute

risk aversion is given by  ( | ) ≡ −00 () 0 () 

• Definition A We say that DM1 is more risk averse

than DM2 if  ( | 1) ≥  ( | 2) for every .

• Discussion Why not just use 00 (·) to measure risk

aversion? (Positive affine transformation on .)

[Link] .li
位
ulx ) (於 → ulx )
⼆ =
=

vcxii
• Definition B DM1 is more risk averse than DM2 if
there exists an increasing concave function  (·) such
that 1 () =  (2 ()) for all .

• Definition C DM1 is more risk averse than DM2 if

 ( | 1) ≤  ( | 2) for all  .

-,,- ,
( 1
1
1 1
1
1
 Stictlyincuusigshauìnuufuion
• Theorem Definitions A, B and C are equivalent.

Proof. (A = B) Since both 1 and 2 are increas-

ing, there exists an increasing function  such that
1 () =  (2 ()). Then,

01 () =  0 (2 ()) 02 () (∗)

h i2
1 () =  (2 ()) 2 () +  0 (2 ()) 002 () 
00 00 0

器 :
tn 㦛啦 機 …

Dividing both sides by 01 () and plug in (∗),

器
[ ⼼ 72
台 ⼀
⼮川
'

( [Link])
"

g
=

shnanui  00 (2 ()) 0

加

⼀些 …" 
 ( | 1) =  ( | 2) − 0 2 () 
 ( ())
背
器 2
ui ⼼ ⼼ "

⼋)

吼⼼⼼ ⼼⼼ Thus,
⼼ -5
 ( | 1) ≥  ( | 2) iff  00 (2 ()) ≤ 0
哭 …
⼼…

( (·) is concave) everywhere.

吼⼼ ⼩ 吣⼼
深 uiso ,

(B = C) Again, there is an increasing function  (·)

such that 1 () =  (2 ()). By definition,

1 ( ( | 1)) =  (1) =  ( (2)) 

and
2 ( ( | 2)) =  (2)
 then
1 ( ( | 2)) =  (2 ( ( | 2)))
=  ( (2)) 
As 1 (·) is increasing, we have
 ( | 1) ≤  ( | 2)
iff
 ( (2)) ≤  ( (2)) 
that is,  (·) is concave by Jensen’s inequality.

• In economic analysis we sometimes assume constant

absolute risk aversion (CARA). A CARA utility
function takes the simple form of
"
 () = −− hlxlidē 。

[Link] inlxtijlxi .tn

⼼ (x) ⼆

where  is the coefficient of absolute risk aversion.

• One may find it more plausible that absolute risk

aversion is decreasing with wealth level (DARA),
that is, richer people are willing take higher risks.

Mul ⼆点
"
⼆点
• The absolute risk aversion measure evaluates a DM’s
attitude towards risky projects whose outcomes are
absolute gains or losses from the initial wealth. Some-
times we are interested in evaluate a DM’s attitude
towards risky projects whose outcomes are percent-
age gains or losses of initial wealth. This leads to
the concept of relative risk aversion.

• Definition The coefficient of relative risk aversion

is given by  ( | ) ≡ −00 () 0 ().

• The constant relative risk aversion (CRRA) utility

function takes the form of

() =
−1−

u (x ) × =

1
-

品
1− u ( ) Ox
"
=
x

where  is the coefficient of constant relative risk

aversion. When  = 1, () = log()

• Another common assumption in economic analysis

is nonincreasing or decreasing relative risk aversion
(DRRA). Note that DRRA implies DARA, but the
converse is not necessarily true.
3.2 Applications

[Link]
-

3.2.1 Insurance

• Consider a strictly risk-averse DM who has initial

wealth of  and may lose $ with probability . He
can buy an insurance. One unit of insurance costs 
and pays $1 if the loss occurs.
500 dneo

• Thus, the DM’s utility maximization problem is

max (1 − )  ( − ) +  ( −  −  + ) 
≥0
If ∗ is an optimum, it must satisfy the first-order
condition:

−  (1 − ) 0 ( − ∗) +
 (1 − ) 0 ( −  + ∗ (1 − )) ≤ 0

← with equality if ∗  0.
[Link]

[Link]
I optintuìnsne
I
7
proftl -ldp-caldp-akdlp-kop-a-s [Link]
Suppose that the insurance price is actuarially fair,
i.e.,  = . Then, the FOC becomes
0 ( −  + ∗ (1 − )) − 0 ( − ∗) ≤ 0
with equality if ∗  0. If ∗ = 0, it becomes
isclecucig
0 ( − ) − 0 () ≤ 0
E
which is impossible when  (·) is strictly concave.
Thus, it must be that ∗  0, and
0 ( −  + ∗ (1 − )) = 0 ( − ∗)
 −  + ∗ (1 − ) =  − ∗
∗ = 
caidohinsurad
If insurance is actuarially fair, the DM insures com-
pletely, that is, his final wealth is  − , whether
or not the loss occurs.
ˊ

• Note that the actuarially fair price corresponds to

the competitive price on the insurance market.

• If the price is not actuarially fair, i.e.,   , the

DM will not insure completely, i.e., ∗  .
 E ( ))
1
4
,

3.2.2 Investment on Risky Assets

• There is a risky asset that yields $ for every dollar

R
invested. Let  ∼  () and  ()  1.

• Consider a DM with initial wealth of . The utility

maximization problem is
Z
max  ( +  − )  () s.t. 0 ≤  ≤ 

where  is the DM’s investment on the risky asset.

The optimum ∗ is determined by the FOC:

Z
0 [ + ∗ ( − 1)] ( − 1)  () = 0
Note that ∗  0 because at ∗ = 0, the FOC
>R >

leads to 0 () (  () − 1)  0 and thus, the

expected utility will increase if  is increased from 0
'( ) ( ] ↑
to some   0. d ^
:

n x > 0

bi
∴ a x

∴ wt -

dig .

• The DM will always invest a positive amount on the

R
risky asset as long as  ()  1.
• The following statements can be rigorously proved.

1. A more risk averse DM invests less on the risky

asset.

2a For a DM with DARA, the optimal investment

on the risky asset ∗ is increasing with the initial
wealth .

2b For a DM with CARA, the optimal investment

on the risky asset ∗ is independent of the initial
wealth .
( ìuatatánpeuntgej ,

3a For a DM with DRRA, the optimal investment

level  ∗ as a percentage of the initial wealth 
is increasing in .

3b For a DM with CRRA, the optimal investment

level  ∗ as a percentage of the initial wealth 
is independent of .
4 Stochastic Dominance

• Suppose that we know a DM is an expected utility

maximizer, but do not know his utility function. Can
we say anything about his preference between two
different monetary lotteries?

• To compare lotteries (or, payoff distributions), we

need the notion of stochastic dominance.

• We explore two ideas, roughly speaking, one is to

identify which lottery gives unambiguously higher re-
turn, and the other is to see which lottery is unam-
biguously less risky, namely, first-order and second-
order stochastic dominance.

• Consider monetary lotteries over an interval [ ],

i.e.,  () = 0 and  () = 1.

• Let  : R → R be the DM’s vNM utility function.

We only assume that  is nondecreasing.
retumone
implies
exeoled cant ominance
,

E= (x) =
EG ( x) F 年G

4.1 First-Order Stochastic Dominance

• Definition For any lotteries  (·) and  (·),  (·)

first-order stochastically dominates  (·) iff
Z Z
() () ≥ () ()

for every nondecreasing  : R → R.

ulx ) ( F ( x ) .
G (x 1 " 三 Julxldx
SF x - Gx
(
)

• As long as the DM prefers having more wealth to

less (i.e.  is increasing), he would prefer to have
the payoff scheme distributed by  (·), rather than
that by  (·).

• Theorem For any lotteries  (·) and  (·),  (·)

first-order stochastically dominates  (·) iff

 () ≤  () for any .

 《
• Graph

1
G(x)
州
i
i
0.3 ˋ
ˋ
x

1
↑ F(x)
1

1
i
0 . 1 -

i ix
b
'
a a
 (·) first-order stochastically dominates  (·)

• First-order stochastic dominance can be obtained by

transferring probability weights upwards. F4 ) ⼆ 0
EGM

[Link]
⼆年
tèxpededoetnn ≠ ↑ 1)

沾炸
• Example A discrete case G 年 0 年 F (2)
1点

Outcomes $1 $2 $3 $4 $5
Dist. F 0 1/4 1/4 1/4 1/4
Dist. G 1/2 0 0 1/2 0

ae-Oistochasticallydominat.
-[Link] [Link]
•

i. Ffistorder
⼆
年 ⼀
.

i 出
 iv) ⼆号
[Link]
"" '

Ǐ
n
EGC Ea (v) 法

thiilgifhxs2_zfuwuytfbul
niyf. 1fCxIdFlDfntGG1xHHMEO-HCDHHre.
@[Link]
以
A

Proof. Let  () =  () −  (). Suppose that

 ()  0 for some . Define a utility function as:
(
HXE 2
1 for   
q
0
 () = 
0 for  ≤ 
Z 
(c)
1st  ()  () = − ()  0
heedmoe 
Surùre and so the “only if” part of the theorem follows.
.

For the “if” part, assume  (·) is differentiable. Then,

Z  '" "派 ⼈ 。

 ()  ()
 yZ
?
( igtinkgpau=
 _
 − 0 ()  () 
[ ()  ()]

Z 
= − 0 ()  () 

If 0 () ≥ 0 and  () ≤ 0 for all , then
Z  Z 
 ()  () = − 0 ()  ()  ≥ 0
 
and
Z Z
() () ≥ () () 
4.2 Second-Order Stochastic Dominance

• Now we assume that lotteries  (·) and  (·) have

the same mean, i.e.,
Z Z
 () =  () 
and then we try to see which lottery is less risky. If
two lotteries have the same mean, a risk averse DM
prefers the one with less risk.

• Definition For any lotteries  (·) and  (·),  (·)

second-order stochastically dominates  (·) iff
Z Z
() () ≥ () ()
for every nondecreasing concave .
 $1 0 (i) (i)
[Link]
0 -100
F
㵆
G $1 (i) -1g [Link]
Qoota
00 …
pi
-

1.9.60-rc0
-

然 羹
• Definition  (·) is a mean-preserving spread of唸
上
⼀

 (·) if we have  ∼  ,  ∼ , and  =  +  for

some , such that  (|) = 0 for all .

• Example Lottery  (·) assigns probability 12 to

two outcomes: $2 and $3. Construct lottery  (·)
from  (·) by spreading the $2 ($3 resp.) outcome
between $1 and $3 ($2 and $4 resp.) with equal
probability. Hence,  (·) assigns probability 14 to
four outcomes: $1, $2, $3 and $4.
p

3/4 F(x)

1/2
G(x)

1/4

x
1 2 3 4

 (·) is a mean-preserving spread of  (·).

 •  (·) is generated from  (·) by adding a noise (extra
risk) at every realized outcome, hence it is riskier
than  (·).

• Definition For any lotteries  (·) and  (·),  (·) is

an elementary increase in risk from  (·) if  (·)
is generated from  (·) by taking
£ 0 00¤
all the mass that
 (·) assigns to an interval    and transferring
it to the endpoints 0 and 00 in such a manner that
the mean is preserved.
ˋ
Fiiiu
p
Area (A) = Area (B) GG )
※[Link]立
1
G :
、

F(x) gànaea
[Link]
Ex G(x)
[Link] B
→ weūsk

[Link]
A

(幻
)
xcrx)
[Link]
⼆
-

x
[Link] ) 我 ⼼比 x’
沿
x’’
、

 (·) is an elementray increase in risk from  (·).

• Regions A and B must have the same area. Why?

• Integrating by parts yields

Z 
( () − ()) 

Z 
= [ () − ()] −  ( () − ())

= 0
that is, the areas below  (·) and  (·) are always
the same. (Note that the last equality holds because
 (·) and  (·) have the same mean.)

• This leads to another way to capture the second-

order stochastic dominance idea:
Z  Z 
() ≥  () for all .
 
 i

[Link]
-_- →

• Example

1
of G(x) D

1st - 2nd .
F(x) C
uis

hebwt samewen dqualmea-sidewical. t

1st →
B
Area (A) >= Area (B)

A Area (A+C) = Area (B+D)

x

 (·) second-order stochastically dominates  (·)

• Theorem For any lotteries  (·) and  (·) with the

same mean, the following are equivalent:
%( F 的 Glxlldx

[Link]/-alwwiIcakICb7=[Link]=CFk)-G(x)ldxg!wuntn(
4
ix)
ix)
⼆

une
1.  (·) second-order stochastically dominates  (·).
幻 比⼼ 唸 悩 以 比 灿 ) dxifiuhhlxyin 㟮 fnn
2.  (·) is a mean-preserving spread of  (·). fikio
正
R R
[Link]
G) I

3. For any  ∈ [ ],  () ≥   ().

[Link]
1 fffa.ci1
$11 州
$ $ 3 4
0 c
上 -

F i 0 0 B ⼼
1

[Link]
G 予 1
0 1
0

jp
[Link]
-1 i
:P 五 frxsz Exfxǜ ) 0 + _ _

Chae G Cholei Edu) 望 送㘭 污 垮

:
,
 0 5 lo

Onebasket al 09
[Link] 0.81 ← addanoiuat 5 ,

※
+5 -5 。

[Link] dqgitonebaskt 、