Artem Tsvetkov & Lammertjan Dam

FTM

Week 4: Luenberger, Chapter 8 and 9

Luenberger 8.5 (Are more data helpful? ⋄) Suppose a stock’s rate of return has annual
mean and variance of r and σ2 . To estimate these quantities, we divide 1 year into n equal
periods and record the return for each period. Let r n and σn2 be the mean and the variance
for the rate of return for each period. Specifically, assume that r n = r/n and σn2 = σ2 /n.
If r̂ n and σ̂n2 are the estimates of these, then r̂ = nr̂ n and σ̂2 = nσ̂n2 . Let σ (r̂ ) and σ(σ̂2 ) be
the standard deviations of these estimates.
(a) Show that σ(r̂ ) is independent of n.
(b) Show how σ(σ̂2 ) depends on n. (Assume the returns are normal random variables.)
Answer the question posed as the title to this exercise.

Solution: We will use the variances of the estimates, r̂ and σ̂2 , for convenience.
(a) Since r̂ = nr̂ n , we have

var(r̂ ) = n2 var(r̂ n ), or for standard deviation

σ(r̂ ) = nσ(r̂ n ). (1)

The variance of return estimate over a single perios is

var(rn ) = σn2 , (2)

Because the mean rate of return is averaged over n intervals, the variance of mean
rate of return is reduced by n
!
1 n σn2
n i∑
var(r̂ n ) = var r n ( i ) = , (3)
=1
n

where rn (i ) is an observed return over the i-th interval. In turn, the variance over a
shorter period is related to the longer period variance as

σ2
σn2 =
n
we have
σn2 σ2
var(r̂ ) = n2 var(r̂ n ) = n2 = n2 2 = σ 2 ,
n n
independent of n.

1
(b)

var(σ̂2 ) = var(nσ̂n2 ) = n2 var(σ̂n2 ) or for standard deviation

σ (σ̂2 ) = nσ (σ̂n2 ) (4)

For normally distributed returns

2σn4 2σ4
var(σ̂n2 ) = = 2 .
n−1 n ( n − 1)

Then
2 2σ4
var(σ̂ ) =
n−1
and
r
2
σ (σ̂2 ) = · σ2 . (5)
n−1
Thus under the assumptions, more frequent sampling reduces the standard deviation
of the variance estimator.
■

2
Luenberger 8.6 (A record) A record of annual percentage rates of return of the stock S
is shown in Table.
(a) Estimate the arithmetic mean rate of return, expressed in percent per year.
(b) Estimate the arithmetic standard deviation of these returns, again as percent per year.
(c) Estimate the accuracy of the estimates found in parts (a) and (b).
(d) How do you think the answers to (c) would change if you had 2 years of weekly data
instead of monthly data? (See Exercise 5.)

Record of rates of Return

Month Percentage rate of return Month Percentage rate of return

1 1.0 13 4.2
2 0.5 14 4.5
3 4.2 15 -2.5
4 -2.7 16 2.1
5 -2.0 17 -1.7
6 3.5 18 3.7
7 -3.1 19 3.2
8 4.1 20 -2.4
9 1.7 21 2.7
10 0.1 22 2.9
11 -2.4 23 -1.9
12 3.2 24 1.1

Solution:
(a) An arithmetic mean of monthly returns is

1 24
24 i∑
r̂ M = ri = 1.0%
=1

Yearly mean return is

r̂Y = 12 r̂ M = 12.0%

(b) Arithmetic monthly standard deviations is

v
24
u
u 1
24 − 1 i∑
σ̂M = t (ri − r̂ M )2 = 2.68%.
=1

Annual standard deviation

√
σ̂Y = 12 σ̂M = 9.29%.

3
(c)
σ̂M
σ(r̂ M ) = √ = 0.55%, see Eq. 2 in L 8.5,
24
σ(r̂Y ) = 12 σ(r̂ M ) = 6.6%, see Eq. 1 in L 8.5,
r
2 2
σ (σ̂M )= σ̂2 = 0.00021, see Eq. 5 in L 8.5,
24 − 1 M
σ (σ̂Y2 ) = σ(12 σ̂M2 2
) = 12 σ(σ̂M ) = 0.0025, see Eq. 4 in L 8.5.

(d) Under the assumptions of Exercise 8.5 the estimate for σ (r̂Y ) does not depend on n.
The accuracy cannot be improved by more frequent sampling. However, the accu-
racy of estimate for variance can be improved

2
var(σ̂Y2 ) = × σ̂Y4
m−1
var(σ̂Y2 )W 23
2
=
var(σ̂Y ) M 103
r
2 23
σ(12 σ̂M )W = × 0.0025 = 0.0012.
103
■

4
Luenberger 8.8 (General tilting ⋄) A general model for information about expected re-
turns can be expressed in vector–matrix form as

p = P r + e.

In the model P is an m × n matrix, r is an n-dimensional vector, and p and e are m-

dimensional vectors. The vector p is a set of observation values and e is a vector of errors
having zero mean. The error vector has a covariance matrix Q. The best (minimum–
variance) estimate of r is
r̂ = (PT Q−1 P)−1 PT Q−1 p. (8.12)
(a) Suppose there is a single asset and just one measurement of the form p = r + e. Show
that according to (8.12), we have r̂ = p.
(b) Suppose there are two uncorrelated measurements with values p1 and p2 , having
variances σ12 and σ22 . Show that
! ! −1
p1 p 1 1
r̂ = 2
+ 22 2
+ 2 .
σ1 σ2 σ1 σ2

(c) Consider Example 8.5. There are measurements of the form

r1 = p 1 + e1
r2 = p 2 + e2
r3 = p 3 + e3
r4 = p 4 + e4
r1 = r f + β1 f M
r2 = r f + β 2 f M
r3 = r f + β 3 f M
r4 = r f + β 4 f M

where the ei ’s are uncorrelated, but where cov(ei , f M ) = 0.25σi2 . Using the data of the
example, and assuming the β’s are known exactly, find the best estimates of the ri ’s.
[Note: You should only need to invert 2 × 2 matrices.]
Solution:
(a) Since p = r + e, P and Q are scalar, where P = PT = 1, and Q = σ2 . Using the
formula, we get
r̂ = (PT Q−1 P)−1 PT Q−1 p = QQ−1 p = p.

5
(b) The corresponding matrices can be written as

σ1−2 0
     2   
1 p1 σ1 0 −1
P= , p= , r = r, Q = , Q = .
1 p2 0 σ22 0 σ2−2

Substituting into the formula we get

! ! −1
p1 p 1 1
r̂ = 2
+ 22 2
+ 2 .
σ1 σ2 σ1 σ2

(c) Example 8.5 contains the following data

Table 1: Data from example 8.5

Stock 1 Stock 2 Stock 3 Stock 4 Market Risk-free
r̂ 15.00 14.34 10.9 15.09 13.83 5.84
var(r ) 90.28 107.24 162.19 68.27 72.12
cov(r ) 65.08 73.62 100.78 48.99 72.12
β 0.90 1.02 1.40 0.68 1.00
r̂ c 13.05 14.00 17.01 11.27

For each stock i in Example 8.5 there are two estimates: historical and CAPM.
Historical average gives (see Example 8.1)
10
1
pih ≡ r̂ ih =
10 ∑ ri ( j) = ri + eih , i = 1, 2, 3, 4,
j =1

where ri ( j) is the return if stock i in year j. If σi is the standard

√ deviation of return ri ,
h h h
the standard deviation of estimate r̂ i is σi = σ (ei ) = σi / 10, since the estimate is
the average over 10 years.
The CAPM estimate is according to the model

pic ≡ r̂ ic = r f + β i bf M = r f + β i (r̂ M − r f ) = r f + β i (r M − r f ) + β i (r̂ M − r M )

= r + β f + β (bf − f ) = r + β (bf − f ) = r + ec , i = 1, 2, 3, 4.
f i M i M M i i M M i i

Here, r f is the risk-free rate, β i is the beta of asset i according to CAPM, r̂ ic , r̂ M , and
r̂ M are the estimates of return ri , market factor f M , and market return r M , respec-
tively. See Example 8.2. Besides, if σM is the standard
√ deviation of market return, the
c c c
standard deviation of r̂ i is σi = σ (ei ) = βσM / 10.

6
 2
We are given that the eih ’s are uncorrelated, but cov(eih , bf M ) = 0.25σih . This gives the
covariance between eih and eic equal to
2
cov(eih , eic ) = 0.25β i σih .

Finally, for each return ri we have a system of two equations

pih = ri + eih
pic = ri + eic .

The corresponding matrices for Eq. 8.12 are

2 2
" #
pih σh i 0.25β i σih
   
1
p= c , P= , r = ri , Q= 2 .
pi 1 0.25β i σih β2 σM2 /10

Inverse matrix Q−1 can be computed as

  −1  
−1 q q 1 q22 −q12
Q = 11 12 = .
q21 q22 (q11 q22 − q12 q21 ) −q21 q11

For example, for the first asset we have

   
15.00 90.28/10 0.25 · 0.9 · 90.28/10
p= , Q= .
13.05 0.25 · 0.9 · 90.28/10 0.92 · 72.12/10

Substitution of the corresponding values for the four assets into Eq. 8.12 gives

r̂1 = 13.73%, r̂2 = 14.13%, r̂3 = 14.29%, r̂4 = 12.33%.

The results are not much different from the “tilt” values in Table 8.5 of Example 8.5
where zero correlations between errors were considered.
The following question may arise. Shouldn’t the correlations between the returns of
asset i and asset j, i.e. between ric and r cj as well as between rih and r hj be taken into
account? The answer is no. These correlations will not affect the value of the esti-
mates. They will affect the estimates’ error, though. One can see the former from the
following considerations. Suppose we have two estimates for two different returns,
which do not have common factors, i.e. P is diagonal, but have correlated errors:

p i = r i + ei
pj = r j + ej, i ̸= j
E(ei e j ) ̸= 0.

7
The corresponding matrices for Eq. 8.12 are
     
pi 1 0 r
p= , P= = 1, r= i ,
pj 0 1 rj

and Q is some covariance matrix with non-zero off-diagonal elements. The solution
for r is

r = (PT Q−1 P)−1 PT Q−1 p = (1 · Q−1 · 1)−1 · 1 · Q−1 p = QQ−1 p = p,

i.e. independent of Q and equal to the original estimate.

■

8
Luenberger 9.2 (Wealth independence) Suppose an investor has exponential utility
function U ( x ) = −e−ax and an initial wealth level of W. The investor is faced with an
opportunity to invest an amount w < W and obtain a random payoff x. Show that his
evaluation of this incremental investment is independent of W.
Solution: If we invest, we spend w and receive x as the result of the investment. Our
payoff at the end of the investment period is W − w + x. If we decided to not invest, we
keep our initial wealth W. It makes sense to invest if the expected utility of the investment
exceeds the utility of keeping money:

E[U (W − w + x )] > E[U (W )]

E[−e−a(W −w+ x) ] > E[−e−aW ]
−e−aW E[e−a(−w+x) ] > −e−aW
E[e−a(−w+ x) ] < 1

One can see that the initial wealth, W, is excluded from the last equation and does not
affect the decision to invest or not to invest. ■

9
Exam 3.2 (Utility and state prices) Suppose an investor exhibits constant relative risk
aversion (CRRA), that is, U (c) = (1/(1 − γ))c1−γ , where U (·) is utility, c is his consump-
tion, and γ = 2 (> 0).
There are two time periods, today (t = 0) and tomorrow (t = 1). Two things can happen
tomorrow (at t = 1), either the economy is in a good state (G), or in a bad state (B),
that both occur with probability 1/2. The investor makes choices at t = 0 to maximize
(terminal) expected utility at time t = 1, i.e.

max E0 [U (c1 )],

where, as usual, his future consumption is subject to portfolio and wealth constraints.
The realization of the random consumption c1 is denoted by c1,G in the good state and
c1,B in the bad state. It turns out that when the investor solves this problem, the optimal
allocation implies that consumption in the good state is equal to c1,G ∗ = 2, and in the
∗
bad state consumption is equal to c1,B = 1/2. Moreover, there exists a risk free asset,
that yields a gross return of 18/17 ≈ 1.059. (i.e. its price is PRF = 1 and its payoff
d RF = 18/17).
(a) What is the price of asset A, denoted PA , that gives a payoff d A,G = −1 in the good
state and d A,B = 1 in the bad state? (4 points)
(b) What is the price of asset B, denoted PB , that gives a payoff d B,G = 1 in the good state
and payoff d B,B = −1 in the bad state? (4 points)
(c) Asset A and B have the same mean and variance. Does this imply that their prices
should actually be the same? Give economic intuition for why or why not. (4 points)
(d) Calculate the elementary state price (or contingent claim price) of the good state, ψG ,
and the elementary state price of the bad state, ψB . (4 points)
(e) Calculate the risk-neutral probabilities of the good state, qG , and the bad state, q B . (3
points)
Solution: (1) A security can be priced as (see Luenberger p. 244)

E [U ′ ( c ∗ ) d X ]
PX = .
RE [U ′ (c∗ )]

In our case ′
c 1− γ

′
U (c) = = c−γ , γ = 2,
1−γ

10
so we have
pG · d A,G · U ′ (cG ) + p B · d A,B · U ′ (c B )
PA =
R ( pG · U ′ (cG ) + p B · U ′ (c B ))
  −2
1 − 2 1
+ 2 · 1 · 12
2 · (−1) · 2 5
=    −2  = 6 ,
R 12 · 2−2 + 12 · 12

where pG = 1/2 and p B = 1/2 are probabilities to reach Good and Bad states, d A,G/B is a
payoff in Good and Bad states, and cG/B is the value of the optimal consumption in these
two states.
(2) Since payoffs A and B differ just in sign, d A = −d B , and using the linearity of expecta-
tion, we get
5
PB = − PA = − .
6
(3) No. Assets that pay well when consumption is low are valued higher. Put differently,
the covariance with the optimal portfolio determines the price, not the variance, similar
to the CAPM logic.
(4) Elementary state prices are the prices of the following payoffs
eG =< 1, 0 >
e B =< 0, 1 > .
Following the same logic as in (1) we get
pG · dG,G · U ′ (cG ) + p B · (dG,B = 0) · U ′ (c B ) 1
ψG = G ′ G B ′ B
=
R ( p · U (c ) + p · U (c )) 18
pG · (d B,G = 0) · U ′ (cG ) + p B · d B,B · U ′ (c B ) 16
ψB = =
R ( pG · U ′ (cG ) + p B · U ′ (c B )) 18
Risk-free asset gives payoff 18/17 in both states. The price of this payoff is 1, in agreement
with the value of risk-free asset.
(5) Risk-neutral probabilities are normalized state prices
ψG 1
qG = =
ψG + ψ B 17
ψB 16
qB = =
ψG + ψ B 17
The obviously add up to one. We can use them to price our assets, like for example
Ê[d] −1 · 1/17 + 16/17 5
PA = = = .
R 18/17 6
■

11
Luenberger 9.5 (Equivalence) A young woman uses the first procedure described in
Section 9.4 to deduce her utility function U ( x ) over the range A ≤ x ≤ B. She uses the
normalization U ( A) = A, U ( B) = B. To check her result, she repeats the whole procedure
over the range A′ ≤ x ≤ B′ , where A < A′ < B′ < B. The result is a utility function V ( x ).
with V ( A′ ) = A′ , V ( B′ ) = B′ . If the results are consistent, U and V should be equivalent:
that is, V ( x ) = aU ( x ) + b for some a > 0 and b. Find a and b.
Solution: The resulting functions should be equivalent, which means that

V ( x ) = aU ( x ) + b, ∀x : A′ ≤ x ≤ B′ .

If we invest in an asset that pays A′ without any risk, we should have utility function
V ( A′ ) = A′ and similar V ( B′ ) = B′ . Respectively,
(
V ( A′ ) = A′ = aU ( A′ ) + b
V ( B′ ) = B′ = aU ( B′ ) + b

The solution of the system of linear equations is

B′ − A′
a=
U ( B′ ) − U ( A′ )
U ( B′ ) A′ − U ( A′ ) B′
b = A′ − aU ( A′ ) = B′ − aU ( B′ ) = .
U ( B′ ) − U ( A′ )
■

12
Luenberger 9.11 (Money-back guarantee) The promoter of the film venture offers a new
investment designed to attract reluctant investors. One unit of this new investment has a
payoff of $3,000 if the venture is highly successful, and it refunds the original investment
otherwise. Assuming that the other three investment alternatives described in Example
9.6 are also available, what is the price of this money-back guaranteed investment?
Solution: There are at least three solutions, a textbook solution from the first principles,
a textbook solution based on log-optimal pricing, and a replication solution.
(a) Textbook solution (from the first principles):
Let us denote the unknown price of the guarantee by P.e An investment of 1$ gives
RG = $3, 000/ Pe return in case of High outcome and 1$ otherwise. Thus, the market
contains 4 assets with the following payoffs.

Outcome Film Risk-free Residual Guarantee Probability, p(s)

High, di ( H ) 3.0 1.2 6 RG p( H ) = 0.3
Moderate, di ( M ) 1.0 1.2 0 1 p( M ) = 0.4
Failure, di ( F ) 0.0 1.2 0 1 p( F ) = 0.3
Price, Pi 1 1 1 1

From the portfolio pricing equation 9.5 in the book we have

E[U ′ ( x ∗ )di ] = λPi , i = 1, 2, 3, 4. (9.5)
Since U ( x ) = ln x and U ′ ( x ) = 1/x, we can multiply both sides of Eq. 9.5 by the
corresponding weight in the optimal portfolio θi and sum over all i to get
4 4
∑ θ i E [U ′ ( x ∗ ) d i ] = ∑ λθi Pi
i =1 i =1
" #
4 4
E U ′ ( x ∗ ) ∑ di θi = λ ∑ θi Pi
i =1 i =1
′ ∗ ∗
E U ( x ) x = λW
 
 
1 ∗
E ∗ x = 1 = λW
x
λ = W −1 .

Using the notations from the table above and the fact the utility is a logarithm func-
tion, U ( x ) = ln x, the equation can be written as
E [U ′ ( x ∗ ) d i ] = ∑ p(s) U ′ ( x ∗ (s)) di (s)
s∈{ H,M,F }
1 1 1
= p( H ) d i ( H ) + p ( M ) d i ( M ) + p ( F ) d ( F)
x∗ ( H ) x∗ ( M) x∗ ( F) i
= W −1 , i = 1, 2, 3, 4.

13
Since we use the returns of assets on a 1$ investment as payoff di , the price is Pi = 1
for all i, and the initial wealth is W = ∑4i=1 θi Pi = ∑4i=1 θ1 . The return of the optimal
portfolio in state s is x ∗ (s) = ∑4i=1 θi di (s). With the values from the table it becomes

3θ1 + 1.2θ2 + 6θ3 + R p θ4 for s = H

∗
x (s) = 1θ1 + 1.2θ2 + 0θ3 + 1θ4 for s = M

0θ1 + 1.2θ2 + 0θ3 + 1θ4 for s = F


Substitution of x ∗ in the pricing equation gives

 0.3 · 3 0.4
 + = W −1 (Film)
3θ1 + 1.2θ2 + 6θ3 + RG θ4 θ1 + 1.2θ2 + θ4




0.3 · 1.2 0.4 · 1.2 0.3 · 1.2


= W −1 (Risk-free)


 + +
3θ1 + 1.2θ2 + 6θ3 + RG θ4 θ1 + 1.2θ2 + θ4 1.2θ2 + θ4

0.3 · 6
= W −1 (Residual)






 3θ1 + 1.2θ2 + 6θ3 + RG θ4
 0.3RG 0.4 0.3
= W −1 (Guarantee)

+ +


3θ1 + 1.2θ2 + 6θ3 + R P θ4 θ1 + 1.2θ2 + θ4 1.2θ2 + θ4
The system contains 4 equations and 5 variables: four θi and RG , and seems to be
underspecified, i.e. it seems to have infinite number of solutions. This is not the case
as the variables θi enter the equations only in three combinations, namely x ∗ ( H ),
x ∗ ( M ), and x ∗ ( F ).

0.3 · 3 · x ∗ ( H )−1 + 0.4 · x ∗ ( M)−1 = W −1 (Film)





 0.3 · 1.2 · x ∗ ( H )−1 + 0.4 · 1.2 · x ∗ ( M )−1 + 0.3 · 1.2 · x ∗ ( F )−1 = W −1 (Risk-free)




 0.3 · 6 · x ∗ ( H )−1 = W −1 (Residual)
0.3 · RG · x ∗ ( H )−1 + 0.4 · x ∗ ( M )−1 + 0.3 · x ∗ ( F )−1 = W −1 (Guarantee)



Thus in fact, there are only four unknowns. We can find them in the following way.
From the Residual equation we have

1.8W = x ∗ ( H ). (A)

Multiplying the Film equation by 2 and subtracting the Residual equation we have

0.8W = x ∗ ( M )

Multiplying the Risk-free equation by 3/1.2, adding the Residual equation, and sub-
tracting the Film equation multiplied by 3 gives

1.8W = x ∗ ( F )

14
 Finally, subtracting from the Guarantee equation the Risk-free equation multiplied
by 1/1.2, we get
1.8( RG − 1)W = x ∗ ( H ).
Together with Eq. A, this gives
RG = 2.
The price of the investment is Pe = $3, 000/RG = $1, 500.

(b) Log-optimal pricing:

In case of logarithmic utility function. the log-optimal pricing can be used. Eq. 9.9 in
the book defines the price of a security as
 
d
P=E .
R∗
Here, d is the payoff of the security and R∗ = x ∗ is the value of the log-optimal
portfolio. The weights θi of the log-optimal portfolio are computed in Example 9.6,
θ1 = −1.0W, θ2 = 1.5W, θ3 = 0.5W. The return on the log-optimal portfolio is
thus R∗ ( H ) = −1.0 · 3.0 + 1.5 · 1.2 + 0.5 · 6 = 1.8, R∗ ( M ) = 0.8, and R∗ ( F ) = 1.8
(see Example 9.7). The payoff of Guarantee is d1 = $3, 000, d2 = P, e and d3 = P. e
Substituting these values in Eq. 9.9, we get
 
d $3, 000 Pe Pe
P=E
e = 0.3 + 0.4 + 0.3
R∗ 1.8 0.8 1.8
 
0.4 0.3 e $3, 000
1− − P = 0.3
0.8 1.8 1.8
1e 1
P = $3, 000
3 6
P = $1, 500.
e

(c) Replication solution:

There are three states in the model (H, M, and F), and three securities (Film venture,
Risk free, and Residual rights). The new alternative can be replicated by a combina-
tion of these three securities. Suppose we buy a, b, c numbers of these three securi-
ties, respectively. The value of such investment is a + b + c. Since we want that this
replicating portfolio reproduce the outcome of the new Guarantee security in every
possible state, we should have the following system of equations
3.0a + 1.2b + 6.0c = 3, 000 for s = H
1.0a + 1.2b = Pe for s = M
1.2b = Pe for s = F
a + b + c = Pe initial investment

15
Solving the system of equations gives the price Pe = $1.500.
Alternatively and specifically to this problem, we can see that we can reproduce the
payoff of the guarantee by the risk-free asset and by the residual.

Outcome Probability Film Risk-free Residual Guarantee

High 0.3 3.0 1.2 6 RG
Moderate 0.4 1.0 1.2 0 1
Failure 0.3 0.0 1.2 0 1

Construct portfolio from these two assets, respectively, x f and xr

x f + xr = 1.

From the second and third outcome we see that

1
xf = .
1.2
Respectively,
1 1
xr = 1 − = .
1.2 6
The payoff in the high outcome is then

RG = 1.2 ∗ x f + 6 ∗ xr = 2.

Pe = $3, 000/RG = $1, 500.

16
Luenberger 9.12 (General positive state prices result) The following is a general result
from matrix theory: Let A be an m × n matrix. Suppose that the equation Ax = p can
achieve no p ≥ 0 except p = 0. Then there is a vector y > 0 with AT y = 0. Use this
result to show that if there is no arbitrage, there are positive state prices; that is, prove the
positive state price theorem in Section 9.9. [Hint: If there are S states and N securities, let
A be an appropriate (S + 1) × N matrix.]
Solution: Construct matrix A,
 
d11 d12 · · · d1N
 d21 d22
 · · · d2N 
A =  ... .. .. ..  ,

. . . 
 
 dS1 dS2 · · · dSN 
− p1 − p2 · · · − pN

where dij is dividend of the security j in state i, and p j are the prices of these securities.
Construct a portfolio of these securities with weights θ = ⟨θ1 θ2 . . . θ N ⟩. The payoff of this
portfolio in the states is
    
d11 d12 · · · d1N θ1 D1
 d21 d22 · · · d2N   θ2   D2 
    
 .. .
.. . .. . .
..   ..  =  ..  .
Aθ =  . =T
   
    
 dS1 dS2 · · · dSN  θ N −1   DS 
− p1 − p2 · · · − p N θN −P

Here Di is the payoff in state i and P is the price of the portfolio.

If P < 0 and Di ≥ 0, ∀i = 1, S, there is an arbitrage.
If P = 0 and Di ≥ 0, ∀i = 1, S and ∃ k : Dk > 0, we again have arbitrage.
This means that the only achievable T is T = 0. This means that matrix A satisfy the
conditions from the matrix theory and ∃ y > 0 s.t.
 
  y1
d11 d21 · · · dS1 − p1 
  y2 

 d12 d22 · · · d − p 2
S2 . 
AT y =  .. ..   ..  = 0.
  
.. . . ..
 . . . . .  
 yS 
d1N d2N · · · dSN − p N
y S +1

As yi > 0, we normalize all yi by yS+1 :

 
  ψ1
d11 d21 · · · dS1 − p1  
 d12 d22 · · · dS2 − p2   ψ2 
  .. 
AT ψ =  .. ..   .  = 0,

.. .. ..
 . . . . .  
ψS 
d1N d2N · · · dSN − pN
1

17
where ψj > 0, ∀ j = 1, S. This means that we have

S
pi = ∑ d ji ψj
j =1

and we have constructed a positive state prices ψj for every state and we can price any
security by the formula above. ■

18
Luenberger 9.15 (General risk-neutral pricing) We can transform the log-optimal pric-
ing formula into a risk-neutral pricing equation. From the log-optimal pricing equation
we have  
d
P=E
R∗
where R∗ is the return on the log-optimal portfolio. We can then define a new expectation
operation E
b by
 
Rx
E( x ) = E
b .
R∗
This can be regarded as the expectation of an artificial probability. Note that the usual
rules of expectation hold. Namely:
(a) If x is certain, then E
b ( x ) = x. This is because E(1/R∗ ) = 1/R.
(b) For any random variables x and y, there holds E b ( ax + by) = aE
b (x) + E
b ( y ).

(c) For any non-negative random variable x, there holds E

b ( x ) ≥ 0.
Using this new expectation operation, with the implied artificial probabilities, show that
the price of any security d is
Eb (d)
P=
R
This is risk neutral pricing.
Solution: From log-optimal pricing
       
d R d R b d = 1E
P=E ∗
=E ∗ = E ∗x = E
b [x] = E b [d] ,
R R R R R R

where we introduced a new variable x = d/R in the intermediate steps. ■

19
Luenberger 9.10 (Portfolio Optimization ⋄) Suppose an investor has utility function U.
There are n risky assets with rates of return ri , i = 1, n, and one risk-free asset with rate of
return r f . The investor has initial wealth W0 . Suppose that the optimal portfolio for this
investor has (random) payoff x ∗ . Show that

E U ′ ( x ∗ )(ri − r f ) = 0
 

for i = 1, n.
Solution: From to Luerberger Eq. 9.4 for the risk-free asset

λ = E U ′ ( x ∗ ) R = E U ′ ( x ∗ ) (1 + r f ).
   

For asset i with return Ri

λ = E U ′ ( x ∗ ) Ri = E U ′ ( x ∗ )(1 + ri ) .
   

Subtracting first equation from the second we get

E U ′ ( x ∗ )(1 + ri ) − E U ′ ( x ∗ ) (1 + r f )
   

=E U ′ ( x ∗ )(1 + ri ) − E U ′ ( x ∗ )(1 + r f )
   

=E U ′ ( x ∗ )(1 + ri ) − U ′ ( x ∗ )(1 + r f )
 

=E U ′ ( x ∗ )(ri − r f ) = 0
 

20
Luenberger 9.13 (Quadratic pricing ⋄) Suppose an investor uses the quadratic utility
function U ( x ) = x − 12 cx2 . Suppose there are n risky assets and one risk-free asset with
total return R. Let R M , be the total return on the optimal portfolio of risky assets. Show
that the expected return of any asset i is given by the formula

Ri − R = β i ( R M − R )
2 . [Hint: Use Exercise 9.10. Apply the result to R itself.]
where β i = cov( R M , Ri )/σM M

Solution: From Exercise 9.10

E U ′ ( x ∗ )(ri − r f ) = 0.
 

As U ( x ) is quadratic, we have U ( x )′ = 1 − cx. If R M is the total return of the optimal

portfolio, Ri = 1 + ri , R = 1 + r f are total returns of assets i and a risk-free asset, and W
is the initial wealth, we have

0 = E U ′ (WR M )(ri − r f )
 

= E U ′ (WR M )( Ri − R)
 

= E [(1 − cWR M )( Ri − R)] .

Respectively,

Ri − R = cW E[ R M Ri ] − R M R

= cW E[( R M − R M + R M )( Ri − Ri + Ri )] − R M R




= cW cov( R M , Ri ) + R M ( Ri − R) .

Further

Ri − R − cWR M ( Ri − R) = cWcov( R M , Ri )
Ri − R = γcov( R M , Ri )
cW
γ=
1 − cWR M
We apply this formula to the optimal portfolio

R M − R = γcov( R M , R M ) = γvar( R M )

Substituting this γ into the equation for Ri we get

cov( R M , Ri )
Ri − R = ( R M − R ) = β i ( R M − R ).
var( R M )
■

21