MATH4512 Fundamentals of Mathematical Finance

Topic 1 Mean variance portfolio theory

1.1 Mean and variance of portfolio return
1.2 Markowitz mean-variance formulation
1.3 Two-fund Theorem
1.4 Inclusion of the risk free asset: One-fund Theorem
1.5 Addition of risk tolerance factor
Appendix: Degenerate case of singular covariance matrix
1
1.1 Mean and variance of portfolio return
Single-period investment model Asset return
Suppose that you purchase an asset at time zero, and 1 year later
you sell the asset. The total return on your investment is dened
to be
total return =
amount received
amount invested
.
If X
0
and X
1
are, respectively, the amounts of money invested and
received and R is the total return, then
R =
X
1
X
0
.
The rate of return is dened by
r =
amount received amount invested
amount invested
=
X
1
X
0
X
0
.
It is clear that
R = 1 +r and X
1
= (1 +r)X
0
.
2
Amount received = dividend received during the investment
period + terminal asset value,
both quantities are uncertain.
We treat r as a random variable, characterized by its probability
distribution. For example, in the discrete case, we have
rate of return r
1
r
2
r
n
probability of occurrence p
1
p
2
p
n
Two important statistics (discrete random variable)
mean = r =
n

i=1
r
i
p
i
variance =
2
(r) =
n

i=1
(r
i
r)
2
p
i
.
3
Statement of the problem
A portfolio is dened by allocating fractions of initial wealth to
individual assets. The fractions (or weights) must sum to one
(some of these weights may be negative, corresponding to short
selling).
Return is quantied by portfolios expected rate of return;
Risk is quantied by variance of portfolios rate of return.
Goal: Maximize return for a given level of risk;
or minimize risk for a given level of return.
(i) How do we determine the optimal portfolio allocation?
(ii) The characterization of the set of optimal portfolios (mini-
mum variance funds and ecient funds).
4
Limitations in the mean variance portfolio theory
Only the mean and variance of rates of returns are taken into
consideration in the portfolio analysis. The higher order mo-
ments (like the skewness) of the probability distribution of the
rates of return are irrelevant in the formulation.
Indeed, only the Gaussian (normal) distribution is fully specied
by its mean and variance. Unfortunately, the rates of return of
asset investment are not Gaussian in general.
Calibration of parameters in the model is always challenging.
Sample mean: r =
1
n

n
t=1
r
t
.
Sample variance
2
=
1
n1

n
t=1
(r
t
r)
2
,
where r
t
is the historical rate of return observed at time t, t =
1, 2, , n.
5
Short sales
It is possible to sell an asset that you do not own through the
process of short selling, or shorting, the asset. You then sell
the borrowed asset to someone else, receiving an amount X
0
.
At a later date, you repay your loan by purchasing the asset
for, say, X
1
and return the asset to your lender. Short selling is
protable if the asset price declines.
When short selling a stock, you are essentially duplicating the
role of the issuing corporation. You sell the stock to raise im-
mediate capital. If the stock pays dividends during the period
that you have borrowed it, you too must pay that same dividend
to the person from whom you borrowed the stock.
6
Return associated with short selling
We receive X
0
initially and pay X
1
later, so the outlay
()
is X
0
and the nal receipt
()
is X
1
, and hence the total return
is
R =
X
1
X
0
=
X
1
X
0
.
The minus signs cancel out, so we obtain the same expression as
that for purchasing the asset. The return value R applies alge-
braically to both purchases and short sales.
We can write
X
1
= X
0
R = X
0
(1 +r)
to show how the nal receipt is related to the initial outlay.
7
Example of short selling transaction
Suppose I short 100 shares of stock in company CBA. This stock
is currently selling for $10 per share. I borrow 100 shares from my
broker and sell these in the stock market, receiving $1, 000. At the
end of 1 year the price of CBA has dropped to $9 per share. I buy
back 100 shares for $900 and give these shares to my broker to
repay the original loan. Because the stock price fell, this has been
a favorable transaction for me. I made a prot of $100.
The rate of return is clearly negative as r = 10%.
Shorting converts a negative rate of return into a prot because the
original investment is also negative.
8
Portfolio weights
Suppose now that n dierent assets are available. We form a port-
folio of these n assets. Suppose that this is done by apportion-
ing an amount X
0
among the n assets. We then select amounts
X
0i
, i = 1, 2, , n, such that
n

i=1
X
0i
= X
0
, where X
0i
represents the
amount invested in the i
th
asset. If we are allowed to sell an asset
short, then some of the X
0i
s can be negative.
We write
X
0i
= w
i
X
0
, i = 1, 2, , n
where w
i
is the weight of asset i in the portfolio. Clearly,
n

i=1
w
i
= 1
and some w
i
s may be negative if short selling is allowed.
9
Portfolio return
Let R
i
denote the total return of asset i. Then the amount of money
generated at the end of the period by the i
th
asset is R
i
X
0i
= R
i
w
i
X
0
.
The total amount received by this portfolio at the end of the period
is therefore
n

i=1
R
i
w
i
X
0
. The overall total return of the portfolio is
R
P
=

n
i=1
R
i
w
i
X
0
X
0
=
n

i=1
w
i
R
i
.
Since
n

i=1
w
i
= 1, we have
r
P
= R
P
1 =
n

i=1
w
i
(R
i
1) =
n

i=1
w
i
r
i
.
10
Covariance of a pair of random variables
When considering two or more random variables, their mutual de-
pendence can be summarized by their covariance.
Let x
1
and x
2
be a pair random variables with expected values x
1
and x
2
, respectively. The covariance of this pair of random variables
is dened to be the expectation of the product of deviations from
the respective mean of x
1
and x
2
:
cov(x
1
, x
2
) = E[(x
1
x
1
)(x
2
x
2
)].
The covariance of two random variables x and y is denoted by
xy
.
We write cov(x
1
, x
2
) =
12
. By symmetry,
12
=
21
, where

12
= E[x
1
x
2
x
1
x
2
x
1
x
2
+x
1
x
2
] = E[x
1
x
2
] x
1
x
2
.
11
Correlation
If the two random variables x
1
and x
2
have the property that

12
= 0, then they are said to be uncorrelated.
If the two random variables are independent, then they are un-
correlated. When x
1
and x
2
are independent, E[x
1
x
2
] = x
1
x
2
so
that cov(x
1
, x
2
) = 0.
If
12
> 0, then the two variables are said to be positively
correlated. In this case, if one variable is above its mean, the
other is likely to be above its mean as well.
On the other hand, if
12
< 0, the two variables are said to be
negatively correlated.
12
When x
1
and x
2
are positively correlated, a positive deviation from
mean of one random variable has a higher tendency to have a pos-
itive deviation from mean of the other random variable.
13
The correlation coecient of a pair of random variables is dened
as

12
=

12

2
.
It can be shown that |
12
| 1.
This would imply that the covariance of two random variables sat-
ises
|
12
|
1

2
.
If
12
=
1

2
, the variables are perfectly correlated. In this situa-
tion, the covariance is as large as possible for the given variance. If
one random variable were a xed positive multiple of the other, the
two would be perfectly correlated.
Conversely, if
12
=
1

2
, the two variables exhibit perfect neg-
ative correlation.
14
Mean return of a portfolio
Suppose that there are N assets with (random) rates of return
r
1
, r
2
, , r
N
, and their expected values E[r
1
] = r
1
, E[r
2
] = r
2
, ,
E[r
N
] = r
N
. The rate of return of the portfolio in terms of the rate
of return of the individual assets is
r
P
= w
1
r
1
+w
2
r
2
+ +w
n
r
N
,
so that
E[r
P
] = w
1
E[r
1
] +w
2
E[r
2
] + +w
n
E[r
N
].
The portfolios rate of return is simply the weighted average of the
rates of return of the assets.
15
Variance of portfolios rate of return
We denote the variance of the return of asset i by
2
i
, the variance
of the return of the portfolio by
2
P
, and the covariance of the return
of asset i with that of asset j by
ij
. Portfolio variance is given by

2
P
= E[(r
P
r
P
)
2
]
= E
_

_
_
_
N

i=1
w
i
r
i

i=1
w
i
r
i
_
_
2
_

_
= E
_
_
_
_
N

i=1
w
i
(r
i
r
i
)
_
_
_
_
N

j=1
w
j
(r
j
r
j
)
_
_
_
_
= E
_
_
N

i=1
N

j=1
w
i
w
j
(r
i
r
i
)(r
j
r
j
)
_
_
=
N

i=1
N

j=1
w
i
w
j

ij
.
16
Zero correlation
Suppose that a portfolio is constructed by taking equal portions of
N of these assets; that is, w
i
=
1
N
for each i. The overall rate of
return of this portfolio is
r
P
=
1
N
N

i=1
r
i
.
Let
2
i
be the variance of the rate of return of asset i. When the
rates of return are uncorrelated, the corresponding variance is
var(r
P
) =
1
N
2
N

i=1

2
i
=

2
aver
N
.
The variance decreases rapidly as N increases.
17
Uncorrelated assets
When the rates of return of assets are uncorrelated, the variance of
a portfolio can be made very small.
18
Non-zero correlation
We form a portfolio by taking equal portions of w
i
=
1
N
of these
assets. In this case,
var(r
P
) = E
_
_
N

i=1
1
N
(r
i
r)
_
_
2
=
1
N
2
E
_
_
_
_
_
N

i=1
(r
i
r)
_
_
_
_
N

j=1
(r
j
r)
_
_
_
_
_
=
1
N
2

i,j

ij
=
1
N
2
_
_
_

i=j

ij
+

i=j

ij
_
_
_
=
1
N
2
{N(
2
i
)
aver
+(N
2
N)(
ij
)
aver
}
=
1
N
_
(
2
i
)
aver
(
ij
)
aver
_
+(
ij
)
aver
.
The covariance terms remain even when we take N .
Also, var(r
P
) may be decreased by choosing assets that are nega-
tively correlated.
19
Correlated assets
If returns of assets are correlated, there is likely to be a lower limit
to the portfolio variance that can be achieved.
20
1.2 Markowitz mean-variance formulation
We consider a single-period investment model. Suppose there are N
risky assets, whose rates of return are given by the random variables
r
1
, , r
N
, where
r
n
=
S
n
(1) S
n
(0)
S
n
(0)
, n = 1, 2, , N.
Here S
n
(0) is known while S
n
(1) is random, n = 1, 2, , N. Let
w = (w
1
w
N
)
T
, w
n
denotes the proportion of wealth invested in
asset n, with
N

n=1
w
n
= 1. The rate of return of the portfolio r
P
is
r
P
=
N

n=1
w
n
r
n
.
21
Assumption
The two vectors = (r
1
r
2
r
N
) and 1= (1 1 1) are linearly
independent. If otherwise, the mean rates of return are equal and so
the portfolio return can only be the common mean rate of return.
Under this degenerate case, the portfolio choice problem becomes
a simpler minimization problem.
The rst two moments of r
P
are

P
= E[r
P
] =
N

n=1
E[w
n
r
n
] =
N

n=1
w
n

n
, where
n
= r
n
,
and

2
P
= var(r
P
) =
N

i=1
N

j=1
w
i
w
j
cov(r
i
, r
j
) =
N

i=1
N

j=1
w
i
w
j

ij
.
22
Covariance matrix
Let denote the covariance matrix so that

2
P
= w
T
w,
where is symmetric and ()
ij
=
ij
= cov(r
i
, r
j
). For example,
when n = 2, we have
(w
1
w
2
)
_

11

12

21

22
__
w
1
w
2
_
= w
2
1

2
1
+w
1
w
2
(
12
+
21
) +w
2
2

2
2
.
Since portfolio variance
2
P
must be non-negative, so the covari-
ance matrix must be symmetric and semi-positive denite. The
eigenvalues are all real non-negative.
In our later discussion, we always assume to be symmetric and
positive denite (avoiding the unlikely event where one of the
eigenvalues is zero) so that
1
always exists. The degenerate
case of singular covariance matrix is discussed in the Appendix.
23
Sensitivity of
2
P
with respect to w
k
By the product rule in dierentiation

2
P
w
k
=
N

j=1
N

i=1
w
i
w
k
w
j

ij
+
N

i=1
N

j=1
w
i
w
j
w
k

ij
=
N

j=1
w
j

kj
+
N

i=1
w
i

ik
.
Since
kj
=
jk
, we obtain

2
P
w
k
= 2
N

j=1
w
j

kj
= 2(w)
k
,
where (w)
k
is the k
th
component of the vector w. Alternatively,
we may write

2
P
= 2w,
where is the gradient operator. This partial derivative gives the
sensitivity of the portfolio variance with respect to the weight of a
particular asset.
24
Remark
1. The portfolio risk of return is quantied by
2
P
. In the mean-
variance analysis, only the rst two moments are considered in
the portfolio investment model. Earlier investment theory prior
to Markowitz only considered the maximization of
P
without

P
.
2. The measure of risk by variance would place equal weight on the
upside and downside deviations. In reality, positive deviations
should be more welcomed.
3. The assets are characterized by their random rates of return,
r
i
, i = 1, , N. In the mean-variance model, it is assumed that
their rst and second order moments:
i
,
i
and
ij
are all known.
In the Markowitz mean-variance formulation, we would like to
determine the choice variables: w
1
, , w
N
such that
2
P
is min-
imized for a given preset value of
P
.
25
Two-asset portfolio
Consider a portfolio of two assets with known means r
1
and r
2
,
variances
2
1
and
2
2
, of the rates of return r
1
and r
2
, together with
the correlation coecient , where cov(r
1
, r
2
) =
1

2
.
Let 1 and be the weights of assets 1 and 2 in this two-asset
portfolio.
Portfolio mean: r
P
= (1 )r
1
+r
2
,
Portfolio variance:
2
P
= (1 )
2

2
1
+2(1 )
1

2
+
2

2
2
.
26
assets mean and variance
Asset A Asset B
Mean return (%) 10 20
Variance (%) 10 15
Portfolio mean
a
and variance
b
for weights and asset correlations
weight = 1 = 0.5 = 0.5 = 1
w
A
w
B
= 1 w
A
Mean Variance Mean Variance Mean Variance Mean Variance
1.0 0.0 10.0 10.00 10.0 10.00 10.0 10.00 10.0 10.00
0.8 0.2 12.0 3.08 12.0 5.04 12.0 8.96 12.0 10.92
0.5 0.5 15.0 0.13 15.0 3.19 15.0 9.31 15.0 12.37
0.2 0.8 18.0 6.08 18.0 8.04 18.0 11.96 18.00 13.92
0.0 1.0 20.0 15.00 20.0 15.00 20.0 15.00 20.0 15.00
a
The mean is calculated as E(R) = w
A
10 +(1 w
A
)20.
b
The variance is calculated as
2
P
= w
2
A
10+(1w
A
)
2
15+2w
A
(1w
A
)

10

5
where is the assumed correlation coecient and

10 and

5 are standard
deviations of the returns of the two assets, respectively.
Observation: A lower variance is achieved for a given mean when
the correlation of the pair of assets returns becomes more negative.
27
We represent the two assets in a mean-standard deviation diagram
As varies, (
P
, r
P
) traces out a conic curve in the -r plane. With
= 1, it is possible to have
P
= 0 for some suitable choice of
weight .
28
Consider the special case where = 1,

P
(; = 1) =

(1 )
2

2
1
+2(1 )
1

2
+
2

2
2
= (1 )
1
+
2
.
Since r
P
and
P
are linear in , and if we choose 0 1, then
the portfolios are represented by the straight line joining P
1
(
1
, r
1
)
and P
2
(
2
, r
2
).
When = 1, we have

P
(; = 1) =

[(1 )
1

2
]
2
= |(1 )
1

2
|.
When is small (close to zero), the corresponding point is close to
P
1
(
1
, r
1
). The line AP
1
corresponds to

P
(; = 1) = (1 )
1

2
.
The point A corresponds to =

1

1
+
2
. It is a point on the vertical
axis which has zero value of
P
.
29
The quantity (1 )
1

2
remains positive until =

1

1
+
2
.
When >

1

1
+
2
, the locus traces out the upper line AP
2
.
Suppose 1 < < 1, the minimum variance point on the curve that
represents various portfolio combinations is determined by

2
P

= 2(1 )
2
1
+2
2
2
+2(1 2)
1

2
= 0

set
giving
=

2
1

2
1
2
1

2
+
2
2
.
30
Mean-standard deviation diagram
31
Mathematical formulation of Markowitzs mean-variance analysis
minimize
1
2
N

i=1
N

j=1
w
i
w
j

ij
subject to
N

i=1
w
i
r
i
=
P
and
N

i=1
w
i
= 1. Given the target expected
rate of return of portfolio
P
, we nd the optimal portfolio strategy
that minimizes
2
P
. The constraint:
N

i=1
w
i
= 1 refers to the strategy
of putting all wealth into investment of risky assets.
Solution
We form the Lagrangian
L =
1
2
N

i=1
N

j=1
w
i
w
j

ij

1
_
_
N

i=1
w
i
1
_
_

2
_
_
N

i=1
w
i
r
i

P
_
_
,
where
1
and
2
are the Lagrangian multipliers.
32
We dierentiate L with respect to w
i
and the Lagrangian multipliers,
then set all the derivatives be zero.
L
w
i
=
N

j=1

ij
w
j

2
r
i
= 0, i = 1, 2, , N; (1)
L

1
=
N

i=1
w
i
1 = 0; (2)
L

2
=
N

i=1
w
i
r
i

P
= 0. (3)
From Eq. (1), we deduce that the optimal portfolio vector weight
w

admits solution of the form

w

=
1
1+
2
or w

=
1
(
1
1+
2
)
where 1= (1 1 1)
T
and = (r
1
r
2
r
N
)
T
.
33
Degenerate case
Consider the case where all assets have the same expected rate
of return, that is, = h1 for some constant h. In this case,
the solution to Eqs. (2) and (3) gives
P
= h. The assets are
represented by points that all lie on the horizontal line: r = h.
In this case, the expected portfolio return cannot be arbitrarily pre-
scribed. Actually, we have to take
P
= h, so the constraint on the
expected portfolio return becomes irrelevant.
34
Solution procedure
To determine
1
and
2
, we apply the two constraints:
1 = 1
T

1
w

=
1
1
T

1
1+
2
1
T

P
=
T

1
w

=
1

1
1+
2

1
.
Writing a = 1
T

1
1, b = 1
T

1
and c =
T

1
, we have two
equations for
1
and
2
:
1 =
1
a +
2
b and
P
=
1
b +
2
c.
Solving for
1
and
2
:

1
=
c b
P

and
2
=
a
P
b

,
where = ac b
2
. Provided that = h1 for some scalar h, we
then have = 0.
35
Solution to the minimum portfolio variance
Both
1
and
2
have dependence on
P
, where
P
is the target
mean prescribed in the variance minimization problem.
The minimum portfolio variance for a given value of
P
is given
by

2
P
= w

T
w

= w

T
(
1
1+
2
)
=
1
+
2

P
=
a
2
P
2b
P
+c

.

2
P
= w
T
w 0, for all w, so is guaranteed to be semi-
positive denite. In our subsequent analysis, we assume to
be positive denite. Given that is positive denite, we have
a > 0, c > 0 and
1
exists. By virtue of the Cauchy-Schwarz
inequality, > 0. Since a and are both positive, the quantity
a
2
P
2b
P
+c is guaranteed to be positive.
36
The set of minimum variance portfolios is represented by a parabolic
curve in the
2
P

P
plane. The parabolic curve is generated by
varying the value of the parameter
P
. Note that
1
a
> 0 while
b
a
may become negative under some extreme adverse cases of negative
mean rates of return.
Non-optimal portfolios are represented by points which must fall on
the right side of the parabolic curve.
37
Global minimum variance portfolio
Given
P
, we obtain
1
=
c b
P

and
2
=
a
P
b

, and the optimal

weight w

=
1
(
1
1+
2
) =
c b
P

1
1+
a
P
b

1
.
To nd the global minimum variance portfolio, we set
d
2
P
d
P
=
2a
P
2b

= 0
so that
P
= b/a and
2
P
= 1/a. Correspondingly,
1
= 1/a and

2
= 0. The weight vector that gives the global minimum variance
portfolio is found to be
w
g
=
1

1
1=

1
1
a
=

1
1
1
T

1
1
.
Note that w
g
is independent of . Obviously, w
T
g
1 = 1 due to the
normalization factor 1
T

1
1 in the denominator. As a check, we
have
g
=
T
w
g
=
b
a
and
2
g
= w
T
g
w
g
=
1
a
.
38
Example
Given the variance matrix
=
_
_
_
2 0.5 0
0.5 3 0.5
0 0.5 2
_
_
_,
nd w
g
. This can be obtained eectively by solving
2v
1
+0.5v
2
= 1
0.5v
1
+3v
2
+0.5v
3
= 1
0.5v
2
+2v
3
= 1.
Observing the symmetry between asset 1 and asset 3 since
2
1
=
2
3
and
12
=
32
, etc.
39
We expect v
1
= v
3
so that the above system reduces to
2v
1
+0.5v
2
= 1
v
1
+3v
2
= 1
giving v
1
= v
3
=
5
11
and v
2
=
2
11
. Lastly, by normalization to sum
of weights equals 1, we obtain
w
g
=
_
5
12
2
12
5
12
_
T
.
40
Two-parameter (
1

2
) family of minimum variance portfolios
Recall w

=
1

1
1+
2

1
, so the minimum variance portfolios
(frontier funds) are seen to be generated by a linear combination of

1
1+
1
.
It is not surprising to see that
2
= 0 corresponds to w

g
since the
constraint on the target mean vanishes when
2
is taken to be zero.
In this case, we minimize risk while paying no regard to the target
mean, thus the global minimum variance portfolio is resulted.
Suppose we normalize
1
by b and dene
w
d
=

1

b
=

1

1
T

.
Obviously, w
d
also lies on the frontier.
41
The corresponding expected rate of return
d
and
2
d
are given by

d
=
T
w
d
=
c
b

2
d
=
_

_
T

_
b
2
=

T

b
2
=
c
b
2
.
Since
1
1 = aw
g
and
1
= bw
d
, the weight of any frontier
fund (minimum variance fund) can be represented by
w

= (
1
a)w
g
+(
2
b)w
d
=
c b
P

aw
g
+
a
P
b

bw
d
.
This provides the motivation of the Two-Fund Theorem. The above
representation indicates that the optimal portfolio weight w

de-
pends on
P
set by the investor.
Any minimum variance fund can be generated by an appropriate
combination of the two funds corresponding to w
g
and w
d
(see
Sec. 1.3: Two-fund Theorem).
42
Feasible set
Given N risky assets, we can form various portfolios from these N
assets. We plot the point (
P
, r
P
) that represents a particular port-
folio in the r diagram. The collection of these points constitutes
the feasible set or feasible region.
43
Argument to show that the collection of the points representing
(
P
, r
P
) of a 3-asset portfolio generates a solid region in the -r
plane
Consider a 3-asset portfolio, the various combinations of assets
2 and 3 sweep out a curve between them (the particular curve
taken depends on the correlation coecient
23
).
A combination of assets 2 and 3 (labelled 4) can be combined
with asset 1 to form a curve joining 1 and 4. As 4 moves
between 2 and 3, the family of curves joining 1 and 4 sweep out
a solid region.
44
Properties of the feasible regions
1. For a portfolio with at least 3 risky assets (not perfectly cor-
related and with dierent means), the feasible set is a solid
two-dimensional region.
2. The feasible region is convex to the left. That is, given any
two points in the region, the straight line connecting them does
not cross the left boundary of the feasible region. This property
must be observed since any combination of two portfolios also
lies in the feasible region. Indeed, the left boundary of a feasible
region is a hyperbola (as solved by the Markowitz constrained
minimization model).
45
Locate the ecient and inecient investment strategies
Since investors prefer the lowest variance for the same expected
return, they will focus on the set of portfolios with the small-
est variance for a given mean, or the mean-variance frontier
(collection of minimum variance portfolios).
The mean-variance frontier can be divided into two parts: an
ecient frontier and an inecient frontier.
The ecient part includes the portfolios with the highest mean
for a given variance.
46
Minimum variance set and ecient funds
The left boundary of a feasible region is called the minimum variance
set. The most left point on the minimum variance set is called the
global minimum variance point. The portfolios in the minimum
variance set are called the frontier funds.
For a given level of risk, only those portfolios on the upper half of
the ecient frontier with a higher return are desired by investors.
They are called the ecient funds.
A portfolio w

is said to be mean-variance ecient if there exists no

portfolio w with
P

P
and
2
P

2
P
, except itself. That is, you
cannot nd a portfolio that has a higher return and lower risk than
those of an ecient portfolio. The funds on the inecient frontier
do not exhibit the above properties.
47
Example Uncorrelated assets with short sales constraint
Suppose there are three uncorrelated assets. Each has variance 1,
and the mean values are 1, 2 and 3 (in percentage points), respec-
tively. We have
2
1
=
2
2
=
2
3
= 1 and
12
=
23
=
13
= 0.
The rst order conditions give w =
1
1+
2
,
T
w =
P
and
1
T
w = 1, so we obtain
w
1

1
= 0
w
2
2
2

1
= 0
w
3
3
2

1
= 0
w
1
+2w
2
+3w
3
=
P
w
1
+w
2
+w
3
= 1.
By eliminating w
1
, w
2
, w
3
, we obtain two equations for
1
and
2
14
2
+6
1
=
P
6
2
+3
1
= 1.
48
These two equations can be solved to yield
2
=

P
2
1 and
1
=
2
1
3

P
. The portfolio weights are expressed in terms of
P
:
w
1
=
4
3

P
2
w
2
=
1
3
w
3
=

P
2

2
3
.
The standard deviation of r
P
at the solution is

w
2
1
+w
2
2
+w
2
3
,
which by direct substitution gives

P
=

7
3
2
P
+

2
P
2
.
The minimum-variance point is, by symmetry, at
P
= 2, with

P
=

3/3 = 0.58. When

P
= 2, we obtain
w
1
= w
2
= w
3
=
1
3
.
49
Short sales not allowed (adding the constraints: w
i
0, i = 1, 2, 3)
Unlike the unrestricted case of allowing short sales, we now impose
w
i
0, i = 1, 2, 3. As a result,
P
can only lie between 1
P
3
[recall
P
= w
1
+2w
2
+3w
3
]. The lower bound is easily seen since

P
= (w
1
+w
2
+w
3
) +(w
2
+2w
3
) = 1 +w
2
+2w
3
1 since w
2
0
and w
3
0. Also,
P
cannot go above 3 as the maximum value
of
P
can only be achieved by choosing w
3
= 1, w
1
= w
2
= 0.
For certain range of
P
, some of the optimal portfolio weights may
become negative when there is no short sales constraint.
50
It is instructive to consider seperately, the following 3 intervals for

P
:
_
1,
4
3
_
,
_
4
3
,
8
3
_
and
_
8
3
, 3
_
.
1
P

4
3
4
3

P

8
3
8
3

P
3
w
1
= 2
P
4
3

P
2
0
w
2
=
P
1
1
3
3
P
w
3
= 0

P
2

2
3

P
2

P
=

2
2
P
6
P
+5

7
3
2
P
+

2
P
2

2
2
P
10
P
+13.
51
From w
3
=

P
2

2
3
, we deduce that when 1
P
<
4
3
, w
3
becomes
negative in the minimum variance portfolio when short sales are
allowed. This is truly an inferior investment choice as the in-
vestor sets
P
to be too low while asset 3 has the highest mean
rate of return. When short sales are not allowed, we expect to
have w
3
= 0 in the minimum variance portfolio. The prob-
lem reduces to two-asset portfolio model and the corresponding
optimal weights w
1
and w
2
can be easily obtained by solving
w
1
+2w
2
=
P
w
1
+w
2
= 1.
Similarly, when
8
3

P
3, it is optimal to choose w
1
= 0. In
this case, the investor sets
P
to be too high while asset 1 has
the lowest mean rate of return.
When
4
3

P

8
3
, we have the same solution as the case without
the short sales constraint. This is because the solutions to
the weights happen to be non-negative under the unconstrained
case. The short sales constraint becomes redundant.
52
1.3 Two-fund Theorem
Take any two frontier funds (portfolios), then any combination of
these two frontier funds remains to be a frontier fund. Indeed, any
frontier portfolio can be duplicated, in terms of mean and variance,
as a combination of these two frontier funds. In other words, all
investors seeking frontier portfolios need only invest in various com-
binations of these two funds.
This property can be extended to a combination of ecient
funds (frontier funds that lie on the upper portion of the ecient
frontier)?
53
Proof of the Two-fund Theorem
Let w
1
= (w
1
1
w
1
n
),
1
1
,
1
2
and w
2
= (w
2
1
w
2
n
)
T
,
2
1
,
2
2
be two
known solutions to the minimum variance formulation with expected
rates of return
1
P
and
2
P
, respectively. By setting
P
equal
1
P
and

2
P
successively, both solutions satisfy
n

j=1

ij
w
j

2
r
i
= 0, i = 1, 2, , n (1)
n

i=1
w
i
r
i
=
P
(2)
n

i=1
w
i
= 1. (3)
We would like to show that w
1
+(1)w
2
is a solution corresponds
to the expected rate of return
1
P
+(1 )
2
P
.
54
1. The new weight vector w
1
+(1)w
2
is a legitimate portfolio
with weights that sum to one.
2. Check the condition on the expected rate of return
n

i=1
_
w
1
i
+(1 )w
2
i
_
r
i
=
n

i=1
w
1
i
r
i
+(1 )
n

i=1
w
2
i
r
i
=
1
P
+(1 )
2
P
.
3. Eq. (1) is satised by w
1
+ (1 )w
2
since the system of
equations is linear. The corresponding
1
and
2
are given by

1
=
1
1
+(1 )
2
1
and
2
=
1
2
+(1 )
2
2
.
4. Given
P
, the appropriate portion is determined by

P
=
1
P
+(1 )
2
P
.
55
Global minimum variance portfolio w
g
and the counterpart w
d
For convenience, we choose the two frontier funds to be w
g
and
w
d
. To obtain the optimal weight w

for a given
P
, we solve for
using
g
+(1)
d
=
P
and w

is then given by w

g
+(1)w

d
.
Recall
g
= b/a and
d
= c/b, so =
(c b
P
)a

.
Proposition
Any minimum variance portfolio with the target mean
P
can be
uniquely decomposed into the sum of two portfolios
w

P
= w
g
+(1 )w
d
where =
c b
P

a.
56
Indeed, any two minimum variance portfolios w
u
and w
v
can be used
to substitute for w
g
and w
d
. Suppose
w
u
= (1 u)w
g
+uw
d
w
v
= (1 v)w
g
+vw
d
we then solve for w
g
and w
d
in terms of w
u
and w
v
. Recall
w

P
=
1

1
1+
2

so that
w

P
=
1
aw
g
+(1
1
a)w
d
=

1
a +v 1
v u
w
u
+
1 u
1
a
v u
w
v
,
whose sum of coecients remains to be 1 and
1
=
c b
P

.
57
Convex combination of ecient portfolios
Any convex combination (that is, weights are non-negative) of e-
cient portfolios is an ecient portfolio.
Proof
Let w
i
0 be the weight of the ecient fund i whose random rate
of return is r
i
e
. Recall that
b
a
is the expected rate of return of the
global minimum variance portfolio.
It suces to show that such convex combination has an expected
rate of return greater than
b
a
in order that the combination of funds
remains to be ecient.
Since E
_
r
i
e
_

b
a
for all i as all these funds are ecient and w
i
0,
i = 1, 2, . . . , n, we have
n

i=1
w
i
E
_
r
i
e
_

i=1
w
i
b
a
=
b
a
.
58
Example
Mean, variances, and covariances of the rates of return of 5 risky
assets are listed:
Security covariance,
ij
mean, r
i
1 2.30 0.93 0.62 0.74 0.23 15.1
2 0.93 1.40 0.22 0.56 0.26 12.5
3 0.62 0.22 1.80 0.78 0.27 14.7
4 0.74 0.56 0.78 3.40 0.56 9.02
5 0.23 0.26 0.27 0.56 2.60 17.68
Recall that w

has the following closed form solution

w

=
c b
P

1
1+
a
P
b

= w
g
+(1 )w
d
,
where = (c b
P
)
a

. Here, satises

P
=
g
+(1 )
d
=
_
b
a
_
+(1 )
c
b
.
59
We compute w

g
and w

d
through nding
1
1 and
1
, then
normalize by enforcing the condition that their weights are summed
to one.
1. To nd v
1
=
1
1, we solve the system of equations
5

j=1

ij
v
1
j
= 1, i = 1, 2, , 5.
Normalize the component v
1
i
s so that they sum to one
w
1
i
=
v
1
i

5
j=1
v
1
j
.
After normalization, this gives the solution to w
g
. Why?
60
We rst solve for v
1
=
1
1 and later divide v
1
by some constant
k such that 1
T
v
1
/k = 1. We see that k must be equal to a, where
a = 1
T

1
1. Actually, a = 1
T

1
1=
N

j=1
v
1
j
.
2. To nd v
2
=
1
, we solve the system of equations:
5

j=1

ij
v
2
j
= r
i
, i = 1, 2, , 5.
Normalize v
2
i
s to obtain w
2
i
. After normalization, this gives the
solution to w
d
. Also, b = 1
T

1
=
N

j=1
v
2
j
and c =
T

1
=
N

j=1
r
j
v
2
j
.
61
security v
1
v
2
w
g
w
d
1 0.141 3.652 0.088 0.158
2 0.401 3.583 0.251 0.155
3 0.452 7.284 0.282 0.314
4 0.166 0.874 0.104 0.038
5 0.440 7.706 0.275 0.334
mean 14.413 15.202
variance 0.625 0.659
standard deviation 0.791 0.812
Recall v
1
=
1
1 and v
2
=
1
so that
sum of components in v
1
= 1
T

1
1= a
sum of components in v
2
= 1
T

1
= b.
Note that w
g
= v
1
/a and w
d
= v
2
/b.
62
Relation between w
g
and w
d
Both w
g
and w
d
are frontier funds with

g
=

T

1
1
a
=
b
a
and
d
=

T

b
=
c
b
.
Their variances are

2
g
= w
T
g

w
g
=
(
1
1)
T
(
1
1)
a
2
=
1
a
,

2
d
= w
T
d
w
d
=
(
1
)
T
(
1
)
b
2
=
c
b
2
.
Dierence in expected returns =
d

g
=
c
b

b
a
=

ab
. Note that

d
>
g
if and only if b > 0.
Also, dierence in variances =
2
d

2
g
=
c
b
2

1
a
=

ab
2
> 0.
63
Covariance of the portfolio returns for any two minimum vari-
ance portfolios
The random rates of return of u-portfolio and v-portfolio are given
by
r
u
P
= w
T
u
r and r
v
P
= w
T
v
r
where r = (r
1
r
N
)
T
is the random rate of return vector. First, for
the two special frontier funds, w
g
and w
d
, their covariance is given
by

gd
= cov(r
g
P
, r
d
P
) = cov
_
_
N

i=1
w
g
i
r
i
,
N

j=1
w
d
j
r
j
_
_
=
N

i=1
N

j=1
w
g
i
w
d
j
cov(r
i
, r
j
) (bilinear property of covariance)
= w
T
g
w
d
=
_
_

1
1
a
_
_
T

b
_
=
1
T

ab
=
1
a
since b = 1
T

1
.
64
In general, consider the two portfolios parametrized by u and v:
w
u
= (1 u)w
g
+uw
d
and w
v
= (1 v)w
g
+vw
d
so that the covariance of their portfolio returns is given by
cov(r
u
P
, r
v
P
) = (1 u)(1 v)
2
g
+uv
2
d
+[u(1 v) +v(1 u)]
gd
=
(1 u)(1 v)
a
+
uvc
b
2
+
u +v 2uv
a
=
1
a
+
uv
ab
2
.
For any portfolio w
P
, we always have
cov(r
g
, r
P
) = w
T
g
w
P
=
1
T

1
w
P
a
=
1
a
= var(r
g
).
65
Minimum variance portfolio and its uncorrelated counterpart
For any frontier portfolio u, we can nd another frontier portfolio v
such that these two portfolios are uncorrelated. This can be done
by setting
1
a
+
uv
ab
2
= 0,
and solve for v, provided that u = 0. Portfolio v is the uncorrelated
counterpart of portfolio u.
The case u = 0 corresponds to w
g
. We cannot solve for v when
u = 0, indicating that the uncorrelated counterpart of the global
minimum variance portfolio does not exist. This observation is con-
sistent with the result that cov(r
g
, r
P
) = var(r
g
) = 1/a = 0.
66
1.4 Inclusion of the risk free asset: One-fund Theorem
Consider a portfolio with weight for the risk free asset and 1
for a risky asset. The risk free asset has the deterministic rate of
return r
f
. The expected rate of portfolio return is
r
P
= r
f
+(1 )r
j
(note that r
f
= r
f
).
The covariance
fj
between the risk free asset and any risky asset
j is zero since
E[(r
j
r
j
) (r
f
r
f
)
. .
zero
] = 0.
Therefore, the variance of portfolio return
2
P
is

2
P
=
2

2
f
..
zero
+(1 )
2

2
j
+2(1 )
fj
..
zero
so that

P
= |1 |
j
.
67
Since both r
P
and
P
are linear functions of , so (
P
, r
P
) lies on a
pair of line segments in the -r diagram.
1. For 0 < < 1, the points representing (
P
, r
P
) for varying values
of lie on the straight line segment joining (0, r
f
) and (
j
, r
j
).
68
2. If borrowing of the risk free asset is allowed, then can be
negative. In this case, the line extends beyond the right side of
(
j
, r
j
) (possibly up to innity).
3. When > 1, this corresponds to short selling of the risky asset.
In this case, the portfolios are represented by a line with slope
negative to that of the line segment joining (0, r
f
) and (
j
, r
j
)
(see the lower dotted-dashed line).
The lower dotted-dashed line can be seen as the mirror image
with respect to the vertical r-axis of the upper solid line segment
that would have been extended beyond the left side of (0, r
f
).
This is due to the swapping in sign in |1 |
j
when > 1.
The holder bears the same risk, like long holding of the risky
asset, while
P
falls below r
f
. This is highly undesirable for the
investor.
69
Consider a portfolio that starts with N risky assets originally, what
is the impact of the inclusion of a risk free asset on the feasible
region?
Lending and borrowing of the risk free asset is allowed
For each portfolio formed using the N risky assets, the new combi-
nations with the inclusion of the risk free asset trace out the pair of
symmetric half-lines originating from the risk free point and passing
through the point representing the original portfolio.
The totality of these lines forms an innite triangular feasible region
bounded by a pair of symmetric half-lines through the risk free point,
one line is tangent to the original feasible region while the other line
is the mirror image about the horizontal line: r = r
f
. The innite
triangular wedge contains the original feasible region.
70
We consider the more realistic case where r
f
<
g
(a risky portfolio
should demand an expected rate of return high than r
f
). For r
f
<
b
a
,
the upper line of the symmetric double line pair touches the original
feasible region.
The new ecient set is the single straight line on the top of the
new triangular feasible region. This tangent line touches the original
feasible region at a point F, where F lies on the ecient frontier of
the original feasible set.
71
No shorting of the risk free asset (r
f
<
g
)
The line originating from the risk free point cannot be extended
beyond the points in the original feasible region (otherwise entails
borrowing of the risk free asset). The upper half line is extended up
to the tangency point only while the lower half line can be extended
to innity.
72
One-fund Theorem
Any ecient portfolio (represented by a point on the upper tangent
line) can be expressed as a combination of the risk free asset and
the portfolio (or fund) represented by M.
There is a single fund M of risky assets such that any ecient
portfolio can be constructed as a combination of the fund M and
the risk free asset.
The One-fund Theorem is based on the assumptions that
every investor is a mean-variance optimizer
they all agree on the probabilistic structure of asset returns
a unique risk free asset exists.
Then everyone purchases a single fund, which is then called the
market portfolio.
73
The proportion of wealth invested in the risk free asset is 1
N

i=1
w
i
.
Write r as the constant rate of return of the risk free asset.
Modied Lagrangian formulation
minimize

2
P
2
=
1
2
w
T
w
subject to
T
w +(1 1
T
w)r =
P
.
Dene the Lagrangian: L =
1
2
w
T
w +[
P
r ( r1)
T
w]
L
w
i
=
N

j=1

ij
w
j
(
i
r) = 0, i = 1, 2, , N (1)
L

= 0 giving ( r1)
T
w =
P
r. (2)
( r1)
T
w is interpreted as the weighted sum of the expected
excess rate of return above the risk free rate r.
74
Remark
In the earlier mean-variance model without the risk free asset, we
have
N

j=1
w
j
r
j
=
P
.
However, with the inclusion of the risk free asset, the corresponding
relation is modied to become
N

j=1
w
j
(r
j
r) =
P
r.
In the new formulation, we now consider r
j
r, which is the excess
expected rate of return of asset j above the riskfree rate of return r.
This is more convenient since the contribution of the riskfree asset
to this excess expected rate of return is zero so that the weight of
the riskfree asset becomes immaterial in the new formulation.
75
Solution to the constrained optimization model
Comparing to the earlier Markowitz model without the riskfree asset,
the new formulation considers the expected rate of return above the
riskfree rate of the risky assets. There is no constraint on sum of
weights equals one. We have
(1): w

= ( r1) and (2): w

T
( r1) =
P
r.
Solving (1): w

=
1
( r1). Substituting into (2)

P
r = ( r1)
T

1
( r1) = (c 2br +ar
2
).
By eliminating , the relation between
P
and
P
is given by the
following pair of half lines ending at the risk free asset point (0, r):

2
P
= w

T
w

= (w

T
rw

T
1)
= (
P
r) = (
P
r)
2
/(c 2br +ar
2
).
76
With the inclusion of the risk free asset, the set of minimum variance
portfolios are represented by portfolios on the two half lines
L
up
:
P
r =
P

ar
2
2br +c (3a)
L
low
:
P
r =
P

ar
2
2br +c. (3b)
Recall that ar
2
2br +c > 0 for all values of r since = ac b
2
> 0.
The pair of half lines give the frontier boundary of the feasible region
of the risky assets plus the risk free asset?
The minimum variance portfolios without the risk free asset lie on
the hyperbola

2
P
=
a
2
P
2b
P
+c

.
77
When r <
g
=
b
a
, one can show geometrically that the upper half
line is a tangent to the hyperbola. The tangency portfolio is the
tangent point to the ecient frontier (upper part of the hyperbolic
curve) through the point (0, r).
78
Though we normally expect r <
g
=
b
a
, what happen when r >
b
a
?
The lower half line touches the feasible region with risky assets only.
Any portfolio on the upper half line involves short selling of the
tangency portfolio and investing the proceeds in the risk free
asset. It makes good sense to short sell the tangency portfolio
since it has an expected rate of return that is lower than the
risk free asset.
79
Solution of the tangency portfolio when r <
g
The tangency portfolio M is represented by the point (
P,M
,
M
P
),
and the solution to
P,M
and
M
P
are obtained by solving simultane-
ously

2
P
=
a
2
P
2b
P
+c

P
= r +
P

ar
2
2br +c.
From the rst order conditions that are obtained by dierentiating
the Lagrangian by the control variables w, we obtain
w

=
1
( r1), (a)
where is then determined by the constraint condition:

P
r = ( r1)
T
w. (b)
80
Recall that the tangency portfolio lies in the feasible region that
corresponds to the absence of the riskfree asset, so 1
T
w
M
= 1.
Note that w
M
should satisfy eq. (a) while pays no reference to eq.
(b) since
M
p
is not yet known (to be determined as part of the
solution). We then deduce that
1 =
M
[1
T

1
r1
T

1
1]
so that
M
=
1
bar
(note that r <
b
a
has been assumed). The
corresponding
M
p
and
2
P,M
can be determined as follows:

M
p
=
T
w

M
=
1
b ar
(
T

1
r
T

1
1) =
c br
b ar
,

2
P,M
= w

T
M
w

M
=
1
(b ar)
2
( r1)
T

1
( r1)
=
ar
2
2br +c
(b ar)
2
.
81
Recall
g
=
b
a
. When r <
b
a
, we can establish
M
P
>
g
as follows:
_

M
P

b
a
__
b
a
r
_
=
_
c br
b ar

b
a
_
b ar
a
=
c br
a

b
2
a
2
+
br
a
=
ac b
2
a
2
=

a
2
> 0,
so we deduce that
M
P
>
b
a
> r.
Similarly, when r >
b
a
, we have
M
p
<
b
a
< r.
Also, we can deduce that
P,M
>
g
as expected. This is because
both Portfolio M and Portfolio g are portfolios generated by the
universe of risky assets (with no inclusion of the riskfree asset), and
g is the global minimum variance portfolio.
82
Example (5 risky assets and one riskfree asset)
Data of the 5 risky assets are given in the earlier example, and
r = 10%.
The system of linear equations to be solved is
5

j=1

ij
v
j
= r
i
r = 1 r
i
r 1, i = 1, 2, , 5.
Recall that v
1
and v
2
in the earlier example are solutions to
5

j=1

ij
v
1
j
= 1 and
5

j=1

ij
v
2
j
= r
i
, respectively, i = 1, 2, . . . , 5.
Hence, v
j
= v
2
j
rv
1
j
, j = 1, 2, , 5 (numerically, we take r = 10%).
In matrix representation, we have
v
1
=
1
1 and v
2
=
1
.
83
Now, we have obtained v where
v =
1
( r1) = v
1
rv
2
Note that the optimal weight vector for the 5 risky assets satises
w = v for some scalar .
We determine by enforcing ( r1)
T
w =
P
r, or equivalently,
( r1)
T
v = (c 2br +ar
2
) =
P
r,
where
P
is the target rate of return of the portfolio.
Recall a = 1
T

1
1=
5

j=1
v
1
j
, b = 1
T

1
=
5

j=1
v
2
j
, and
c =
T

1
=
5

j=1
r
j
v
2
j
.
The weight of the risk free asset is then given by 1
5

j=1
w
j
.
84
Properties of the minimum variance portfolios for r < b/a
1. Ecient portfolios
Any portfolio on the upper half line

P
= r +
P

ar
2
2br +c
within the segment FM joining the two points F(0, r) and M
involves long holding of the market portfolio M and the risk free
asset F, while those outside FM involves short selling of the risk
free asset and long holding of the market portfolio.
2. Any portfolio on the lower half line

P
= r
P

ar
2
2br +c
involves short selling of the market portfolio and investing the
proceeds in the risk free asset. This represents a non-optimal
investment strategy since the investor faces risk but gains no
extra expected return above r.
85
Furthermore, note that
M
P
r =
ar
2
2br+c
bar
and
P,M
=

ar
2
2br+c
|bar|
.
One can show that
(i) when r <
b
a
, we obtain

M
P
r =
P,M

ar
2
2br c (equation of L
up
);
(ii) when r >
b
a
, we obtain

M
P
r =
P,M

ar
2
2br +c (equation of L
low
).
Interestingly, the ip of sign in b ar with respect to r <
g
or
r >
g
would indicate whether (
P,M
,
M
P
) lies in the upper or lower
half line, respectively.
86
Degenerate case occurs when
g
=
b
a
= r
What happens when r = b/a? The pair of half lines become

P
= r
P

c 2
_
b
a
_
b +
b
2
a
= r
P

a
,
which correspond to the asymptotes of the hyperbolic left bound-
ary of the feasible region with risky assets only. The tangency
portfolio does not exist, consistent with the mathematical result
that
M
=
1
b ar
is not dened when r =
b
a
.
Under the scenario: r =
b
a
, ecient funds still lie on the upper
half line, though the tangency portfolio does not exist. Recall
that
w

=
1
( r1) so that
1
T
w

= (1
T

1
r1
T

1
1) = (b ra).
87
When r = b/a, sum of weights of risky assets = 1
T
w

= 0 as is
nite. Any minimum variance portfolio involves investing everything
in the riskfree asset and holding a portfolio of risky assets whose
weights are summed to zero.
The optimal weight vector w

equals
1
( r1). Suppose
we specify
P
to be the target expected rate of return of the
ecient portfolio, then the multiplier is determined by (see
p.75)
=

P
r
c 2br +ar
2

r=b/a
=

P
r
c 2
_
b
a
_
b +
b
2
a
=
a(
P
r)

.
88
Financial interpretation
Given the target expected rate of portfolio return
P
, the corre-
sponding optimal portfolio is to hold 100% on the riskfree asset
and w
j
on the j
th
risky asset, j = 1, 2, , N, where w
j
is given by
the j
th
component of
a(
P
r)

1
( r1).
One should check whether the expected rate of return of the whole
portfolio equals
P
.
The expected rate of return from all the risky assets is
a(
P
r)

T
[
1
( r1)] =
a(
P
r)
ac b
2
_
c
b
2
a
_
=
P
r.
The overall expected rate of return of the portfolio is
w
0
r +
N

j=1
w
j
r
j
= r +(
P
r) =
P
, where w
0
= 1.
89
One-fund Theorem under r =
g
= b/a
In this degenerate case, r = b/a, the tangency fund does not exist.
At its global minimum variance portfolio g, the universe of risky
assets just provide an expected rate of return that is the same as
the riskfree return r.
The optimal portfolio is to invest 100% on the riskfree asset and a
scalar multiple of the fund z whose weight vector is
w
z
=
1
( r1).
The scalar is determined by the investors target expected rate of
return
P
. The sum of weights in this portfolio z is zero.
Since the global minimum variance portfolio of risky assets g has the
same expected rate of return as that of the riskfree asset, a sensible
investor would place 100% weight on the riskfree asset to generate
the level of expected rate of return equals r. This is consistent with
the observation that = 0 when
P
= r.
90
Nature of the portfolio z: w
z
=
1
( r1), where r = b/a
1. Recall w
d
=
1
/b and w
g
=
1
1/a, so
w
z
= bw
d

b
a
(aw
g
) = b(w
d
w
g
).
Its sum of weights is seen to be zero since it longs b units of w
d
and short the same number of units of w
g
.
2. Location of the portfolio z in the mean-variance plot

2
z
= w
T
z
w
z
= b
2
(w
d
w
g
)
T
(w
d
w
g
)
= b
2
(
2
d
2
g
d
+
2
g
) = b
2

ab
2
=

a
;

z
=
T
w
z
= b(
d

g
) = b
_
c
b

b
a
_
=

a
.
We have
z
=

a
and
z
=

a
; so any scalar multiple of z lies
on the line: =

a
.
91
3. Location of ecient portfolios in the mean-variance diagram
Suppose the investor species her target rate of return to be

P
, then the scalar is determined by setting

P
= r +
z
= r +

a
giving =
a(
P
r)

.
Also, the standard deviation of the optimal portfolios return is

P
=
z
=

a
. The ecient portfolio lies on the line:

P
r =

P
.
The target expected rate of portfolio return above r is produced
by appropriate long and short positions on the risky assets with
sum of weights equals zero.
92

p
o
) , 0 ( r
|
|
.
|

\
|
A A
a a
,
P p
a
r o
A
=
P P
a
o
A
=
The upper line represents the set of frontier funds. The point
_
_

a
,

a
_
_
on the lower line represents the z-fund, w
z
.
93
Market Portfolio is a portfolio consisting of a weighted sum of
every asset in the market, with weights in the proportions that they
exist in the market (under the assumption that these assets are
innitely divisible). This is also called a value-weighted portfolio.
In the market portfolio, HSBC has a larger weight than Bank of
East Asia due to her larger market capitalization.
The Hang Seng index may be considered as a proxy of the market
portfolio of the Hong Kong stock market.
94
Tangency portfolio under One-fund Theorem and market port-
folio
The One-fund Theorem states that everyone purchases a single
fund (tangency portfolio) of risky assets and borrow or lend at
the riskfree rate.
If everyone purchases the same tangency portfolio (same weights
of all risky assets in the market), what must that fund be? This
fund is simply the market portfolio. In other words, if everyone
buys just one fund, and their purchases add up to the market,
then the proportional weights in the tangency fund must be the
same as those of the market portfolio.
In the situation where everyone follows the mean-variance method-
ology with the same estimates of parameters, the ecient fund
of risky assets will be the market portfolio.
95
How can this happen? The answer is based on the equilibrium
argument.
If everyone else (or at least a large number of people) solves the
problem, we do not need to. The return on an asset depends
on both its initial price and its nal price. The other investors
solve the mean-variance portfolio problem using their common
estimates, and they place orders in the market to acquire their
portfolios.
If orders placed do not match with what is available, the prices
must change. The prices of the assets under heavy demand
will increase while the prices of the assets under light demand
will decrease. These price changes aect the estimates of asset
returns directly, and hence investors will recalculate their optimal
portfolio.
96
This process continues until demand exactly matches supply,
that is, it continues until an equilibrium prevails.
Summary
In the idealized world, where every investor is a mean-variance
investor and all have the same estimates, everyone buys the
same portfolio and that must be equal to the market portfolio.
Prices adjust to drive the market to eciency. Then after other
people have made the adjustments, we can be sure that the
single ecient portfolio is the market portfolio.
97
1.5 Addition of a risk tolerance factor
The universe of assets include N risky assets but no riskfree asset.
Maximize
P

2
P
2
, with 0, where is the risk tolerance.
Optimization problem: max
wR
N

2
P
2
subject to 1
T
w = 1.
Instead of only minimizing risk as in the mean variance mod-
els, the new objective function represents the tradeo between
return and risk with weighted factor 2. When is high, the
investor is more interested in expected return and has a high
tolerance on risk.
Note that
P
in the objective function is not preset. The tol-
erance factor is chosen by the investor and will be xed in
the formulation. The choice variables are the portfolio weights
w
i
, i = 1, 2, , N.
98
Quadratic optimization problem
max
wR
N
_

T
w
w
T
w
2
_
subject to w
T
1= 1.
The Lagrangian formulation becomes
L(w; ) =
T
w
w
T
w
2
+(w
T
11).
The rst order conditions are
_
w

+1= 0
w
T
1= 1
.
When is taken to be zero, the problem reduces to the minimization
of portfolio return variance without regard to expected portfolio
return. This gives the global minimum variance portfolio w

g
.
99
One can show that the optimal solution w

can be expressed as
w
g
+z

, 0.
1. When = 0, the two rst order conditions become
w =
0
1 and 1
T
w
g
= 1.
Solving
w
g
=
0

1
1 and 1 = 1
T
w
g
=
0
1
T

1
1
hence
w
g
=

1
1
1
T

1
1
=

1
1
a
(independent of ).
The formulation does not depend on when is taken to be
zero.
100
2. When > 0, we obtain w

=
1
+
1
1. To determine
, we apply
1 = 1
T
w = 1
T

1
+1
T

1
1 so that =
1 b
a
.
w

=
1
+
1 b
a

1
1
=
_

1

b
a

1
1
_
+w
g
.
We obtain w

= w
g
+z

, where
z

=
1

b
a

1
1= b(w
d
w
g
) and 1
T
z

= 0.
Write r
w
g
as the random rate of return of portfolio g, and a
similar notation for z

. Observe that cov(r

wg
, r
z
) = z
T
w
g
=
0,

z
=
T
z

= c
b
2
a
=

a
> 0 and
2
z
=

a
> 0.
101
Financial interpretation
The zero tolerance solution w
g
leads to the global minimum risk
position. This position is modied by investing in the portfolio z

[note that 1
T
z

= 0] so as to maximize
T
w
w
T
w
2
.
Set of optimal portfolios
For a given value of , we have solved for w

(with dependence
on ). We then compute
P
and
2
P
corresponding to the optimal
weight w

P
=
T
(w
g
+z

) =
g
+
z

2
P
=
2
g
+2 cov(r
w
g
, r
z
)
. .
z

T
w
g
=0
+
2

2
z
.
By eliminating , we obtain

2
P
=
2
g
+
_

P

g

_
2

2
z
=
2
g
+
_

P

g

_
2
,
z
=
2
z
=

a
.
This is an equation of a hyperbola in the
P
-
P
diagram.
102

P
=
b
a
+

a

2
P
=
1
a
+

a

2
The points representing these optimal portfolios in the
P
-
P
dia-
gram lie on the upper half of the hyperbola. We expect that for a
higher value of chosen by the investor, the optimal portfolio has
higher
P
and
P
.
103
How to reconcile

the mean-variance model and risk-tolerance
model?
Recall that the left boundary of the feasible region of the risky assets
is given by

2
P
=
a
2
P
2b
P
+c

, = ac b
2
. (1)
The parabolic curve that traces all optimal portfolios of the risk-
tolerance model in the
2
P
-
P
diagram is

P
=
b
a
+

a
and
2
P
=
1
a
+

a

2
. (2)
It can be shown that the solutions to
P
and
2
P
in Eq. (2) satisfy
the parabolic equation (1) since
a
_
b
a
+

a

_
2
2b
_
b
a
+

a

_
+c

=
1
a
+

a

2
.
104
With a given tolerance , the objective function line

2
P
2
= constant
is pushed upwards until it becomes the tangent line to the parabolic
curve.
s
2
p
m
P
s
2
p
=
a 2b + c m - m
P P
2
D
tm -
P
s
P
2
2
= constant
105
The objective function line:
P

2
P
2
= constant in the
2
P
-
P
diagram is pushed upwards as much as possible in the maximiza-
tion procedure.
However, the optimal portfolio must lie in the feasible region of
risky assets. Recall that the feasible region is bounded on the
left by the parabolic curve:
2
P
=
a
2
P
2b
P
+c

. The objective
function
P

2
P
2
is maximized when the objective function line
touches the left boundary of the feasible region.
In the degenerate case where = 0, the slope of the line:

P


2
P
2
= constant becomes innite. When we maximize

2
P
2
= constant by pushing the vertical line to the far left, the
corresponding optimal portfolio obtained is Portfolio g.
106
Another version of the Two-fund Theorem
Given
P
, the ecient fund under the mean-variance model is given
by
w

= w
g
+
a

P

b
a
_
z

,
P
>
g
=
b
a
,
where w
g
=

1
1
a
, z

=
1

b
a

1
1.
This implies that any ecient fund can be generated by the two
funds: global minimum variance fund w
g
and the fund z

.
107
Proof
1. Note that w

is of the form
1

1
1+
2

1
.
2. Consider the expected portfolio return:

T
w

=
g
+
a

P

b
a
_

z
=
b
a
+
a

P

b
a
_

a
=
P
.
3. Consider the sum of weights:
1
T
w

= 1
T
w
g
+
a

P

b
a
_
1
T
z

= 1.
108
Comparing the rst order conditions of the mean-variance model
and risk-tolerance model
1. w

=
1
1+
2
2. w

= 1+
1
T
w

= 1 1
T
w

= 1
1
T
=
P
We observe that

1
= =
c b
P

2
=
a
P
b

is simply .
The specication of the risk tolerance factor is somewhat equiv-
alent to the specication of
P
.
109
Summary
1. The objective function
T
w
w
T
w
2
represents a balance of
maximizing return
T
w against risk
w
T
w
2
.
2. The optimal solution takes the form
w

= w
g
+z

where w
g
is the portfolio weight of the global minimum variance
portfolio and the weights in z

are summed to zero.

110
Note that
z

=
1

b
a

1
1= b(w
d
w
g
).
and and
P
are related by
=
a

P

b
a
_
.
Putting the results together, we have
w

= w
g
+
ab

P

b
a
_
(w
d
w
g
)
=
ab

_
c
b

P
_
w
g
+
ab

P

b
a
_
w
d
.
111
3. The additional variance above
2
g
is given by

2
z
=
2

a
, = ac b
2
.
Also, cov(r
wg
, r
z
) = 0, that is, r
w
g
and r
z
are uncorrelated
4. The ecient frontier of the mean-variance model coincides with
the set of optimal portfolios of the risk-tolerance model. The
risk tolerance and expected portfolio return
P
are related by

P

g

=

P

g

2
z

= .
5. Another version of the Two-fund Theorem can be established
where any ecient fund can be generated by the two funds: w
g
and z

.
112
Asset-liability model
Liabilities of a pension fund = future benets future contributions
Market value can hardly be determined since liabilities are not read-
ily marketable, unlike tradeable assets. Assume that some specic
accounting rules are used to calculate an initial value L
0
. If the
same rule is applied one period later, a value for L
1
results. Note
that L
1
is random.
Rate of growth of the liabilities = r
L
=
L
1
L
0
L
0
, where r
L
is ex-
pected to depend on the changes of interest rate structure, mortality
and other stochastic factors.
Let A
0
be the initial value of assets. The investment strategy of
the pension fund is characterized by the portfolio choice w. Let r
w
denote the rate of growth of the asset portfolio.
113
Surplus optimization
Depending on the portfolio choice w, the surplus gain after one
period
S
1
S
0
= [A
0
(1 +r
w
) L
0
(1 +r
L
)] (A
0
L
0
) = A
0
r
w
L
0
r
L
.
The rate of return on the surplus relative to asset value is dened
by
r
S
=
S
1
S
0
A
0
= r
w

1
f
0
r
L
where f
0
= A
0
/L
0
is the initial funding ratio.
Here, r
S
is the dierence of the two random variables: r
w
and
1
f
0
r
L
,
one of them is dependent on the choice variable w while the other
is not.
114
Maximization formulation
max
wR
N
_
E
_
r
w

1
f
0
r
L
_

1
2
var
_
r
w

1
f
0
r
L
__
subject to
N

i=1
w
i
= 1. Since E
_
1
f
0
r
L
_
and var(r
L
) are independent
of w so that they can be omitted from the objective function.
We rewrite the quadratic maximization formulation as
max
wR
N
_
E[r
w
]
var(r
w
)
2
+
1
f
0
cov(r
w
, r
L
)
_
subject to
N

i=1
w
i
= 1. Recall that
cov(r
w
, r
L
) = cov
_
_
N

i=1
w
i
r
i
, r
L
_
_
=
N

i=1
w
i
cov(r
i
, r
L
).
115
Modied maximization formulation
max
wR
N
_

T
w +
T
w
w
T
w
2
_
subject to 1
T
w = 1,
where
T
= (
1

N
) with
i
=
1
f
0
cov(r
i
, r
L
),

T
= (
1

N
) with
i
= E[r
i
],
ij
= cov(r
i
, r
j
).
1. The additional term
T
w in the objective function arises from
the correlation cov(r
i
, r
L
) multiplied by the factor L
0
/A
0
.
2. Compared to the earlier risk tolerance model, we just need to
replace by +
1

. The ecient portfolios are of the form

w

= w
g
+z
L
+z

, 0,
where z
L
=
1

1
T

1
T

1
1

1
1 with
N

i=1
z
L
i
= 0.
116
Appendix: Mathematical properties of covariance matrix
1. All eigenvalues of are non-negative
If otherwise, suppose is a negative eigenvalue of and x is
the corresponding eigenvector. We have
x = x, x = 0,
so that
x
T
x = x
T
x < 0,
a contradiction to the semi-positive denite property of .
2. is non-singular (
1
exists) if and only if all eigenvalues are
positive
Recall
det = product of eigenvalues
and
is non-singular det = 0.
117
3. Decomposition of and representation of
1
when is non-
singular
Let
i
, i = 1, 2, . . . , n, be the eigenvalues of (allowing multi-
plicities) and x
i
be the corresponding eigenvector of eigenvalue

i
. Since is symmetric, it has a full set of eigenvectors. We
then have
S = S,
where is the diagonal matrix whose entries are the eigenvalues
of and S is the matrix whose columns are the eigenvectors
of (arranged in the corresponding sequential order). The
eigenvectors are orthogonal to each other since is symmetric
and we can always normalize the eigenvectors to be unit length.
That is, S can be constructed to be an orthonormal matrix so
that S
1
= S
T
. We then have
= SS
1
= SS
T
.
118
Provided that all eigenvalues of are non-zero so that
1
exists, we then have

1
= (SS
T
)
1
= (S
T
)
1

1
S
1
= S
1
S
T
.
4. = ac b
2
> 0, where = h1
Note that
a = 1
T

1
1= (1
T
S
1/2
)(
1/2
S
T
1)
b =
T

1
1= (
T
S
1/2
)(
1/2
S
T
1)
c =
T

1
= (
T
S
1/2
)(
1/2
S
T
).
We write x =
1/2
S
T
1 and y =
1/2
S
T
so that a = x
T
x,
b = y
T
x and c = y
T
y.
The Cauchy-Schwarz inequality gives
|y
T
x|
2
(x
T
x)(y
T
y).
For = h1, x and y are then linearly independent, we have
= ac b
2
> 0.
119
Appendix: Degenerate case of singular covariance matrix
All the formulas derived earlier are based on the assumption that the
covariance matrix is non-singular. In this Appendix, we investigate
how to determine the feasible region and the corresponding ecient
funds when
1
does not exist. Recall
is singular the set of eigenvalues of contains zero.
That is, there exists non-zero vector w
0
such that
w
0
= 0.
We seek w
0
such that w
T
0
1 = 1. Write r
w
0
as the random rate of
return of the portfolio with the weight vector w
0
. Note that
var(r
w
0
) = w
T
0
w
0
= 0.
120
Since var(r
w
0
) = 0, the corresponding portfolio can be considered
as a proxy for the riskfree asset (zero variance), whose riskfree
rate of return r
0
is given by
r
0
= w
T
0
.
Example
Consider the two-asset portfolio with = 1, the corresponding
covariance matrix is
=
_

2
1

1

2

2
2
_
.
Obviously, is singular since the columns are dependent. Accord-
ingly, we obtain
w
0
=
_

2

1
+
2

1
+
2
_
,
where w
0
= 0 and w
T
0
1= 1. See p. 28 for the pictorial represen-
tation of the zero-variance portfolio in the mean-variance diagram.
121
Let E
0
denote the eigenspace corresponding to the eigenvalue 0. It
is seen that E
0
is simply the null space of .
First, we assume that there exists unique (up to a scalar multiple)
eigenvector w
0
corresponding to the eigenvalue 0. We have
rank() = dim(column space of ) = N 1,
where is a N N matrix. The global minimum variance portfolio
is the proxy riskfree asset generated by w
0
, whose riskfree rate of
return is r
0
= w
T
0
.
The feasible region and ecient frontier would be similar to the
portfolio model with the inclusion of the riskfree asset. Since any
combination of any portfolio and w
0
lies on the line joining the
portfolio and this proxy riskfree asset, the feasible region is an
innite triangular wedge with its tip at (0, r
0
) in the
P
r
P
diagram.
122
We would like to nd the equation of the upper half line that bounds
the triangular wedge region. Similar to the portfolio model with the
inclusion of the riskfree asset, we impose the constraint:

P
r
0
=
N

j=1
w
j
(r
j
r
0
).
Note that

N
j=1
w
j
= 1 in general. The weight vector w of an
ecient fund is obtained by solving
w = ( r
0
1) = ( 1w
T
0
) = (I 1w
T
0
), (1)
where the matrix 1w
T
0
is obtained by taking the outer product of
1 and w
0
. Similarly, the parameter is determined by specifying
the investors target expected rate of return of the portfolio.
123
Several mathematical issues:
1. Since
1
does not exist, does solution to eq. (1) always exist?
2. Suppose solution to eq. (1) exists, it is not unique since adding
any scalar multiple of w
0
is also a solution (since w
0
= 0).
3. What is the interpretation of in eq. (1) and how to determine
its value?
Though is singular, recall that solution w exists in eq. (1) provided
that the column vector (I 1w
T
0
) lies in the column space of .
124
Mathematical proof that solution to eq. (1) always exists
Since w
0
= 0, the eigenvector w
0
is orthogonal to all the rows in .
Since is symmetric, so the column space of is the orthogonal
complement of the eigenspace E
0
. To establish the existence of
solution, it suces to show that r
0
1 lies in the orthogonal
complement of E
0
(which is the column space of ). This is easily
seen since
w
T
0
( r
0
1) = w
T
0
r
0
= 0.
125
Example
Given the singular matrix
=
_

2
1

1

2

2
2
_
and w
0
=
1

1
+
2
_

1
_
in E
0
, we have w
0
= 0. The proxy
riskfree rate of return is given by
r
0
= w
T
0
=
1

1
+
2
_

2

1
_
_
r
1
r
2
_
=

2
r
1
+
1
r
2

1
+
2
.
As a check, it is readily seen that solution exists for
w =
__
r
1
r
2
_

2
r
1
+
1
r
2

1
+
2
_
1
1
__
=
_
_

1
( r
1
r
2
)

1
+
2

2
( r
2
r
1
)

1
+
2
_
_
=
( r
1
r
2
)

1
+
2
_

1

2
_
since the column vector
(r
1
r
2
)

1
+
2
_

1

2
_
T
lies in the column
space of .
126
Other mathematical issues
1. What would happen when dim(E
0
) > 1? In this case, there are
multiple proxy riskfree assets. However, such scenario is ruled
out; otherwise, this leads to arbitrage opportunities where the
riskfree rate of return of one portfolio is higher than the riskfree
rate of return of another portfolio.
2. Suppose happens to lie in E
0
, the projection of onto the
orthogonal complement of E
0
is the zero vector. Now, solution
to eq. (1) is w
0
since the right-hand side vector is the zero
vector. In this case, the investor puts 100% of her wealth into
the proxy riskfree asset.
127