Chapter 4

Mean-Variance Portfolio Analysis

As a first example of a portfolio problem, we consider standard mean-variance optimiza-

tion. Historically this was one of the earliest problems considered. It is important because
mean-variance analysis provides a basis for the derivation of the equilibrium model known
variously as the capital asset pricing model (CAPM), Sharpe-Lintner model, Black model,
and two- factor model.
Mean-variance analysis is fully consistent with expected utility maximization only under
special circumstances. The required assumptions will be discussed later.

4.1 The Standard Mean-Variance Portfolio Problem

We assume for now that the investors’ preferences can be represented by a (derived) utility
function defined over the mean and the variance of a portfolio’s return, V (Ẑ, σ 2 ). Using
the notation in Chapter 2, we obtain that the expected return and variance on a portfolio are
N
X
′
Z̄ = w z̄ + w0 R = wi z̄i , (4.1a)
i=0

N X
X N
σ 2 = w′ Σw = wi wj σij . (4.1b)
i=1 j=1

The standard assumption is that preferences induce the favoring of higher means V1 > 0
and smaller variances V2 < 0. Under this assumption the class of potentially optimal port-
folios for such investors are therefore those with the greatest expected return for a given
level of variance and, simultaneously, the smallest variance for a given expected return.
(If short sales are unrestricted, the first condition alone is an adequate description.) Such
portfolios are termed mean-variance efficient. Here we shall work with the larger class
of minimum-variance portfolios. This is the set that includes the single portfolio with the
smallest variance at every level of expected return. All mean-variance efficient portfolios
are also minimum-variance portfolios, but the converse is not true. These sets are illus-
trated in Figures 4.1 and 4.2. (Not only is the latter class easier to characterize and work
with analytically, but also under certain asymmetric distributions V1 is ambiguous in sign
even for investors who strictly prefer more to less. Thus, in general, all minimum-variance
portfolios must be included in the optimal set. Whenever mean-variance analysis is con-
sistent with expected utility maximization by a risk-averse investor, however, V2 must be
4.1 The Standard Mean-Variance Portfolio Problem 53

negative.)
Z
Expected
Mean-Variance
Return
Efficient Locus

Feasible Portfolios

σ
Standard Deviation

Figure 4.1 Mean-Variance Efficient Portfolios

We first find the minimum-variance portfolios in the absence of a riskless asset. The
minimum-variance portfolio with expected return µ. is the solution w(µ) to
1 ′
Min 2
w Σw (4.2a)

Subject to 1′ w = 1, (4.2b)

z̄′ w = µ. (4.2c)

We have not imposed any positivity constraints of the form wi > 0, so unrestricted short
sales are permitted.

Z
Expected
Mean-Variance
Return
Locus

Feasible Portfolios

σ
Standard Deviation

Figure 4.2 Mean-Variance Portfolios

Forming the Lagrangian L ≡ w′ Σw/2 + λ(1 − 1′ w) + γ(µ − z̄′ w) and differentiating

give the first-order conditions
∂L
0= = Σw − λ1 − γz̄ (4.3)
∂w
54 Mean-Variance Portfolio Analysis

and (2b) and (2c). The solution set is

w∗ = λΣ−1 1 + γΣ−1 z̄,
N N
X X (4.4)
wi∗ =λ vij + γ j = 1vij z̄j .
j=1

λ and γ can be determined from the constraints (2b) and (2c):

C − µB µA − B
λ= , γ= ,
∆ ∆
A ≡ 1′ Σ−1 1 > 0, B ≡ 1′ Σ−1 z̄, (4.5)

C ≡ z̄′ Σ−1 z̄ > 0, ∆ ≡ AC − B 2 > 0.

(Note: ∆ > 0 by the Cauchy-Schwarz inequality since we have assumed that Σ is non-
singular and all assets do not have the same mean, z̄ 6= k 1. If all means were the same,
then ∆ = 0, and this problem has no solution except when µ = k .) The equation of the
minimum-variance set is, from (4), (2b), (2c), and (5),

σ 2 = w′ Σw = w′ Σ λΣ−1 1 + γΣ−1 z̄
¡ ¢

= λw′ 1 + γ w′ z̄ = λ + γµ
Aµ2 − 2Bµ + C
= . (4.6)
∆
This is the equation of a parabola. In mean-standard deviation space the curve is a hyper-
bola, as illustrated in Figure 4.3.

Z
Z = B A + σ (∆ A)1/ 2
Expected
Return

portfolio d

B A global minimum
variance portfolio g
σ
Standard Deviation

Figure 4.3 Risky-Asset-Only Minimum-Variance Set

The global minimum variance portfolio is located by

dσ 2 2Aµ − 2B
0= = . (4.7)
dµ ∆
That is, it has an expected return of B/A and, from (6), a variance of 1/A. Substituting
for µ in (5) gives λ = 1/A and γ = 0, so the global minimum- variance portfolio is
Σ−1 1 Σ−1
wg = = ′ −1 . (4.8)
A 1Σ
4.1 The Standard Mean-Variance Portfolio Problem 55

This can also be determined by setting γ , the Lagrange multiplier associated with the mean
constraint, to zero in (4). With γ = 0 the constraint is not binding, and the global minimum-
variance portfolio is found.
The slopes of the asymptotes can be computed from dµ/dσ as µ approaches ±∞.
√
dµ dµ dσ 2 ∆ ∆ p 2
= 2 = 2σ = Aµ − 2Bµ + C,
dσ dσ dσ 2Aµ − 2B Aµ − B
r (4.9)
dµ ∆
lim =± .
µ→±∞ dσ A
The second and third equalities follow from (7) and (6), respectively.
From (4) we can see that all minimum-variance portfolios are portfolio combinations of
only two distinct portfolios. Since the global minimum-variance portfolio corresponds to
the first term in (4), we choose, for simplicity, the portfolio corresponding to the second
term as the second diversified portfolio. Assuming B 6= 0, we define wd ≡ Σ−1 z̄/B =
Σ−1 z̄/1′ Σ−1 z̄, and we can write (4) as
w∗ = (λA)wg + (γB)wd . (4.10)
Since λA + γB = 1 [see (5)], we have verified this proposition.
As indicated, portfolio g is the global minimum-variance portfolio at the apex of the
hyperbola in Figure 4.3. We can locate the other as follows. The difference in expected
returns is
C B ∆
Z̄d − Z̄g = − = . (4.11)
B A AB
As noted previously, ∆ > 0. A is also positive since it is a quadratic form. In principal,
B can have either sign; however, if the expected return on the global minimum-variance
portfolio is positive, then B > 0. We take this as the typical case, so wd will be on the
upper limb of the hyperbola. (Note: Even limited liability of the assets is no guarantee of a
positive B because the global minimum-variance portfolio may be short in some assets.)
The result we have derived is an example of a separation or mutual fund theorem because
all investors who choose portfolios by examining only mean and variance can be satisfied
by holding different combinations of only a few, in this case two, mutual funds regardless of
their preferences. All of the original assets, therefore, can be purchased by just two mutual
funds, and the investors can then just buy these. Separation theorems are very important in
finance. We shall see many further examples.
Any two distinct minimum-variance portfolios will serve in place of wg and wd . For
example, if wa and wb are two minimum-variance portfolios, then, from (10), wa = (1 −
a)wg + awd and similarly for portfolio b. Thus
λA + b − 1 1 − a − λA
w∗ = wa + wb . (4.12)
b−a b−a
These coefficients also sum to unity so the proposition is proved.
The portfolio weight of any asset is linear in µ along the minimum-variance frontier.
Substituting (5) into (10) gives
(Aµ − B)B(wid − wig )
wi∗ (µ) = wig + . (4.13)
∆
In the typical case, B > 0, assets represented more (less) heavily in wd than wg will be held
in greater amounts in the minimum-variance portfolios with high (low) expected returns.
For each asset there is one minimum-variance portfolio in which it has a weight of zero. In
all portfolios below (above) this one the asset is sold short.
56 Mean-Variance Portfolio Analysis

4.2 Covariance Properties of the Minimum-Variance Portfolios

As we shall see, the covariance properties of the portfolios in the minimum-variance set are
of particular importance. We will examine some of them.
The global minimum-variance portfolio possesses the peculiar property that its covari-
ance with any asset or portfolio is always 1/A.
³ ´ 1′ Σ−1 Σwp 1
Cov Z̃g , Z̃p = w′g Σwp = = . (4.14)
A A
For every other minimum-variance portfolio the range of possible covariances is (−∞, ∞).
Let wa ≡ (1−a)wg +awd denote a minimum-variance portfolio. For two such portfolios
³ ´
Cov Z̃a , Z̃b = (1 − a)(1 − b)σg2 + abσd2 + [a(1 − b) + b(1 − a)]σdg
(1 − a)(1 − b) abC a + b − 2ab
= + 2 +
A B A
1 ab∆
= + .
A AB 2
(4.15)

For a fixed portfolio a on the upper limb of the hyperbola, a > 0, this covariance increases
from −∞ to ∞ as b is increased; that is, portfolio b is moved up the hyperbola in the
standard case. The portfolio uncorrelated with a is on the lower limb. For a fixed portfolio
a on the lower limb, the opposite is true. This will be proved later.
Again for a fixed portfolio a on the upper limb, the correlation coefficient increases from
−a/(B 2 /∆+a2 )1/2 to 1 as b increases from −∞ to a. It then falls off to a/(B 2 /∆+a2 )1/2
as b goes to ∞.

4.3 The Mean-Variance Problem with a Riskless Asset

When a riskless asset is available, the portfolio problem is changed in two ways. First, there
is no budget constraint on the risky assets corresponding to (2b). A position in the riskless
asset makes up the residual, or, if negative, finances the risky asset purchases. Second, the
constraint on the mean return must be expressed in excess return form

(z̄ − R1)′ w = µ − R. (2c′ )

The resulting maximization yields a solution

w∗ = γΣ−1 (z̄ − R1), w0∗ = 1 − 1′ w∗ . (4.16)

γ can be obtained from (2c’) and (16) as the solution to

µ − R = γ(z̄ − R1)′ Σ−1 (z̄ − R1) = γ[C − 2RB + R2 A]. (4.17)

The equation of the minimum-variance set is, from (16), (2c’), and (17),

σ 2 = (µ − R)2 [C − 2RB + R2 A]−1 . (4.18)

Again, this is a parabola; however, this time in mean-standard deviation space the locus is
a pair of rays with common intercepts at R and slopes of ±(C − 2RB + R2 A)1/2 . This is
4.3 The Mean-Variance Problem with a Riskless Asset 57

illustrated in Figure 4.4.

1/ 2
Z Z = R + σ ( C − 2 RB + R 2 A )
Expected
Return

tangency portfolio t
global minimum
variance portfolio g
R

σ
Standard Deviation

Figure 4.4 Minimum-Variance Set

As before, all minimum-variance portfolios are combinations of only two distinct port-
folios. Again, any two minimum-variance portfolios will span the set. In this case, however,
there is a natural choice of funds - namely, the riskless asset and that portfolio with none of
the riskless asset, the “tangency” portfolio

Σ−1 (z̄ − R1)

w0t = 0, wt = . (4.19)
B − AR
The mean and variance of this portfolio are
C − BR
Z̄t = z̄′ wt = ,
B − AR
(4.20)
C − 2RB + R2 A
σt2 = w′t Σwt = .
(B − AR)2
Portfolio t is a member of the original risky-asset-only minimum-variance set. This can
be easily verified by showing that Z̄t , and σt2 satisfy Equation (6); however, it is also obvious
from the geometry in Figure 4.4. Portfolio t is at the tangency point of the hyperbola and a
line passing throughR. This can be proved by calculating the slope of the hyperbola at this
point, but again Figure 4.4 should be sufficient. The case illustrated shows this tangency
portfolio on the upper limb of the hyperbola; however, a lower tangency is possible instead.
The upper (lower) tangency occurs if the riskless return is less (greater) than the expected
return on the (risky-asset-only) global minimum-variance portfolio, That
is, either Z̄t > B/A > R or the reverse ordering is correct. To prove this proposition,
we compute µ ¶µ ¶ · ¸
B B C − BR B B − AR
Z̃t − −R = −
A A B − AR A A
2 2
(4.21)
C − BR B BR CA − B ∆
= − 2+ = = > 0.
A A A A A
Therefore, both differences must be of the same sign. (Note: In the singular case B/A = R
there is no finite point of tangency, and the minimum-variance locus coincides with the
asymptotes of the risky-asset-only hyperbola.)
58 Mean-Variance Portfolio Analysis

4.4 Expected Returns Relations

The expected excess return on any asset is proportional to its covariance with the return on
Zt . Let σ t denote the vector of these covariances. Then since the covariance of a random
variable with a weighted sum of random variables is the weighted sum of the individual
covariances, we have, from (19),

z̄ − R1
σ t = Σwt = . (4.22)
B − AR
Premultiplying (22) by wt gives
Z̄t − R
σt2 = w′t σ t = . (4.23)
B − AR
Combining (22) and (23) gives
Z̄t − R
z̄ − R1 = σ t ≡ β t (Z̄t − R). (4.24)
σt2
Because all other portfolios in the minimum variance set are perfectly correlated with
the tangency portfolio, exactly the same result holds for them.
All expected returns can also be written in terms of portfolio wd . If σ d is the vector of
covariances with this portfolio, then, from the definition of wd ,
µ ¶
z̄ σd
σ d = Σwd = , z̄ = Z̄d ≡ β d Z̄d . (4.25)
B σd2
This relation holds only for portfolio d and not for any of the other hyperbola portfolios.
For other hyperbola portfolios we can express re turns by using a second portfolio. Let
wa ≡ (1 − a)wg + awd (a 6= 0, a 6= 1) be a risky-asset-only minimum-variance portfolio.
Then µ ¶
1−a ³a´
σ a ≡ Σwa = 1+ z̃. (4.26)
A B
Now let wp be any other risky-asset-only portfolio. Then we can derive from (26) the
system of equations
1 − a Z̄a a 1 − a Z̄p a
σa2 = + , σpa = + , (4.27)
A B A B
which can be solved for (1 − a)/A and a/B . Substituting these into (26) and rearranging
terms give
Z̄p σa2 − Z̄a σap Z̄a − Z̄p
z̄ = 2
1+ 2 σa. (4.28)
σa − σap σa − σap
If we choose as portfolio p a portfolio Zz which has a zero beta with (is uncorrelated with)
portfolio a, then (28) simplifies to
¡ ¢
z̄ = Z̄z 1 + Z̄a − Z̄z β a . (4.29)

Although (29) is generally true, it is particularly useful when no riskless asset is avail-
able. Comparing (29) to (24), we can see that all of the portfolios that are uncorrelated
with a particular minimum-variance portfolio have an expected return equal to the intercept
on the Z̄ axis of a line drawn tangent to the (risky-asset-only) efficient hyperbola at the
particular minimum-variance portfolio. From Figure 4.4 or (21), the zero beta portfolio of
4.5 Equilibrium: The Capital Asset Pricing Model 59

a minimum-variance portfolio on the upper (lower) limb of the hyperbola must be on the
lower (upper) limb.
This linearity property of expected returns and beta holds only if the bench-mark port-
folio is in the minimum-variance set. To prove this converse proposition, assume that

z̄ = α1 + γσ p = α1 + γΣwp (4.30)

for some portfolio p. Then

1 −1 α
wp = Σ z̄ − Σ−1 1 = δ wd + η wg . (4.31)
γ γ
Furthermore, since 1′ wp = 1, δ + η = 1, so portfolio p is a portfolio combination of
portfolios g and d and must lie in the minimum-variance set.
The relation between the expected return on any portfolio and its variance can now be
expressed as follows. For any portfolio wp let wm be the minimum-variance portfolio with
the same expected return. Then wp is

wp = wg + (wm − wg ) + (wp − wm )
= wg + ω s + ω d ,
(4.32)

where ω s and ω d are the arbitrage portfolios defined as ω s ≡ wm − wg and ω d ≡ wp − wm .

Note that the covariances between these parts are all zero:

σgs ≡ w′g Σω s = w′g Σ(wm − wg ) = 0, (4.33a)

σgd ≡ w′g Σω d = w′g Σ(wp − wm ) = 0, (4.33b)

σsd ≡ ω ′s Σω d = w′m Σ(wp − wm ) − w′g Σ(wp − wm ) = 0, (4.33c)
Equations (33a) and (33b) are zero since they are each the difference of the covariances
of two portfolios with the global minimum-variance portfolio, and by (14) the covariance
of every portfolio with wg is equal to 1/A. Equation (33c) is also zero. The second term
is zero by (33b). The first term is zero by (27). It is the difference of the covariance of
portfolios wp and wm with portfolio wm . Since wp and wm have the same expected returns,
their covariances with wm , a minimum-variance portfolio, must be the same.
Thus, the variance of portfolio wp , can be separated into three parts:

σp2 = σg2 + σs2 + σd2 . (4.34)

σg2 measures unavoidable risk, σs2 measures the systematic risk associated with a level of
expected return of Z̄p , and σd2 measures diversifiable risk. Only the first two portions of the
variance contribute to expected return.

4.5 Equilibrium: The Capital Asset Pricing Model

The pricing results just given cannot be used to explain the expected returns observed in the
market in equilibrium. The minimum-variance portfolios (with the exception of the global
minimum-variance portfolio) cannot be identified until the vector of expected returns is
known. Furthermore, even if investors do not choose portfolios through mean- variance
analysis (so that there need be no optimal properties of the minimum-variance portfolios),
60 Mean-Variance Portfolio Analysis

the minimum-variance portfolios exist provided only that there are no riskless arbitrage
opportunities (and means and variances are defined), and the pricing results of the previ-
ous section obtain. Thus, these results have no equilibrium content until some minimum-
variance portfolio can be identified by another route. This identification is precisely what
the capital asset pricing model accomplishes.
To derive the CAPM we make the following assumptions: (i) Each investor chooses a
portfolio with the objective of maximizing a derived utility function of the form V (Z̄, σ 2 )
with V2 < 0, V1 > 0, and V concave. (ii) All investors have a common time horizon and
homogenous beliefs about z̄ and Σ. (iii) Each asset is infinitely divisible (iv) The riskless
asset can be bought or sold in unlimited amounts.
These conditions are sufficient to show that each investor holds a minimum-variance
portfolio. Intuitively, no other portfolio could be optimal, because the minimum-variance
portfolio with the same expected return would be preferred since V2 < 0. Furthermore,
since all minimum-variance portfolios are combinations of any two minimum-variance
portfolios, a separation or mutual fund theorem obtains, as previously explained.
Formally, the problem is

Max V (R + w′ (z̄ − R1), w′ Σw) , (4.35)

where we have substituted for w0 using the budget constraint. Solving (35) gives
∂V
0= = V1 (·)(z̄ − R1) + 2V2 (·)Σw,
∂w
(4.36)
−V1 (·) −1
w∗ = Σ (z̄ − R1).
2V2 (·)
The second-order condition is met since
∂ 2V
= 2V2 Σ + V11 (z̄ − R1)(z̄ − R1)′
∂ w∂ w′
+ 2V12 [Σw(z̄ − R1)′ + (z̄ − R1)w′ Σ] + 4V22 Σww′ Σ
(4.37)

is negative definite. (V2 (·) < 0, and Σ is positive definite. V11 < 0, V22 < 0, and
V11 + 4V12 + 4V22 < 0 by concavity, and the other matrices are positive semidefinite with
that multiplying V12 positively dominated.)
Comparing (36) to (19), we see that the optimal investment in the risky assets is propor-
tional to the tangency portfolio. The remainder of wealth is invested in the riskless asset.
Since all investors hold combinations of the tangency portfolio and the riskless asset, the
aggregate demand for each risky asset must be in proportion to its representation in the
tangency portfolio. In equilibrium, demand and supply are equal; therefore, the supply of
each risky asset in the market portfolio must also be in proportion to the tangency portfolio:
wM ∝ wt . If the riskless asset is a financial security in zero net supply, then wt is the
market portfolio wM . In other cases the market portfolio will be somewhere to the left on
the tangency line. Using (24) we can therefore write the Sharpe-Lintner CAPM pricing
equation:
¡ ¢
z̄ − R1 = βM Z̄M − R . (4.38)
In the absence of a riskless asset, or if there are limits on its sale or purchase, a similar
result can still be derived provided that (iv’) there are no restrictions on the sale of risky
4.5 Equilibrium: The Capital Asset Pricing Model 61

assets. (The assumption that V1 > 0 is not required.) The same intuition is valid. The
formal problem is

Max V (w′ z̄, w′ Σw) + λ(1 − 1′ w). (4.39)

The solution to (39) satisfies

∂L
0 = = V1 (·)z̄ + 2V2 (·)Σw − λ1,
∂w
0 = 1 − 1′ w.
(4.40)

The optimal portfolio is

−V1 (·) −1 λ
w = Σ z̄ + Σ−1 1
2V2 (·) 2V2 (·)
µ ¶
BV1 BV1
= − wd + 1 + wg
2V2 2V2
(4.41)

when λ is substituted out by using the budget constraint, Optimal portfolios are combina-
tions of wd and wg . Thus, aggregate demand is a combination of wd and wg . In equilibrium,
supply equals demand; therefore, the market portfolio must be a minimum-variance portfo-
lio, and (29) describes expected returns:
¡ ¢
z̄ = Z̄z 1 + Z̄M − Z̄z β M , (4.42)

where z denotes a zero beta portfolio uncorrelated with the market. This is the Black CAPM
or the “two-factor” model.
In the standard case when the expected return on the global minimum- variance portfolio
is positive (B > 0), the expected return (and variance) on portfolio d exceeds that on
portfolio g . Investors who are more “risk averse” (−V1 /V2 is smaller) hold less of portfolio
d, as expected.
The primary difference in the assumptions leading to these results is that short sales
need not be permitted for the risky assets when unlimited riskless borrowing and lending
is available, provided that V1 > 0. In this case all investors will want to hold portfolios on
the upper portion of the frontier. If an upper (lower) tangency exists, all investors will have
a positive (negative) demand for the tangency portfolio. Since this is the market portfolio
in equilibrium, the net demand for it must be positive. The first conclusion therefore is that
an upper tangency must exist and Z̃M > R. Second, since each investor’s demand for the
market portfolio is positive and it is long in each asset, every investor’s desired holding of
every risky asset is positive and a no-short-sales constraint would not be binding. Without
a riskless asset some short sales must be allowed in order to fashion every portfolio on the
upper limb of the minimum variance hyperbola. Since the portfolios on the lower limb can
then also be formed, restricting our attention to utilities with V1 > 0 is not necessary. We
demonstrate in Appendix B that V1 < 0 is not consistent with expected utility maximization
62 Mean-Variance Portfolio Analysis

when a riskless asset is available, but it can be in the absence of one.

Z
Expected
borrowing portfolio
Return
lending portfolio

RB

RL
σ
Standard Deviation

Figure 4.5 M-V Problem with Restricted Borrowing and Lending

When there are restrictions on the riskless asset, such as only lending permitted or there
is a higher borrowing than lending rate, but no restrictions on the other assets, then the zero
beta version of the CAPM is still valid. These two cases are illustrated in Figure 4.5. The
most risk-averse investors will hold tangency portfolio L combined with lending. Less risk-
averse investors will hold long positions in both L and B . The least risk-averse investors
will be overly long in B financed by borrowing or a short position in L if borrowing is not
allowed. In any case each investor’s demand for the risky assets will be some combina-
tion of portfolios L and B . Since these are both minimum-variance portfolios, aggregate
demand will be minimum variance as well. In equilibrium, therefore, the market portfolio
will be in the minimum-variance class.
To summarize, it is the mean-variance analysis that generates the separation theorem and
the pricing result that relates expected returns to covariances with minimum-variance port-
folios. The equilibrium analysis simply leads to the identification of the market portfolio as
being minimum variance.

4.6 Consistency of Mean-Variance Analysis and Expected Utility Max-

imization

Thus far we have simply assumed that investors’ optimal policies consist of selecting among
mean-variance efficient portfolios. In this section we explore the conditions under which
this behavior maximizes expected utility.
The first case examined is quadratic utility. Suppose that an investor has a quadratic
utility function defined over end-of-period wealth. We can choose the units of wealth so
that the investor’s initial endowment is unity, so the argument of the utility function is
return. Also since utility is defined only up to an increasing linear transformation, we may
write
bZ 2
u(Z) = Z − (4.43)
2
with complete generality.
4.6 Consistency of Mean-Variance Analysis and Expected Utility Maximization 63

For the utility function in (43), expected utility is

b¡ 2
V (Z̄, σ 2 ) ≡ Eu(Z̃) = Z̄ − Z̄ + σ 2 .
¢
(4.44)
2
Thus, for quadratic utility of wealth functions, expected utility depends solely on the mean
and variance of return.
Furthermore, for risk-averse utility functions (b > 0), V2 < 0, so only minimum-
variance portfolios are held. If outcomes are confined to the range of increasing utility
(Z < 1/b), then V1 > 0, and only mean-variance efficient portfolios can be optimal.
From (44) we can see that the indifference curves in mean-standard deviation space are
semicircles centered at Z̄ = 1/b, σ = 0. For the relevant portion (Z̄ < 1/b), they are
convex so a tangency with the minimum-variance frontier exists.
Assuming quadratic utility has two drawbacks. First, all concave quadratic functions are
decreasing after a certain point, so some care must be employed to assure that the outcomes
remain in the lower, relevant range of utility. Second, quadratic utility functions display
increasing absolute risk aversion
u′′ (Z) b
A(Z) ≡ − ′ = ,
u (Z) 1 − bZ
µ ¶2 (4.45)
′ b
A (Z) = > 0.
1 − bZ
The more commonly adopted justification for mean-variance analysis is multivariate
normally distributed returns. Since all weighted sums of multivariate normal random vari-
ables are normally distributed, and since a normal distribution is completely characterized
by its mean and variance, all portfolios formed from such assets are completely described
by means and variances.
We can also demonstrate that only mean-variance efficient portfolios are optimal. Let
f (Z; Z̄, σ 2 ) denote the normal density function for the return on a portfolio. Then
Z ∞
2
u(Z)f (Z; Z̄, σ 2 )dZ
¡ ¢
V Z̄, σ ≡
Z−∞
∞
= u(Z̄ + σε)n(ε)dε,
−∞
(4.46)

where ε ≡ (Z − Z̄)/σ is a standard normal deviate and n(·) is its density function. The
partial derivative with respect to mean return is
Z ∞
V1 = u′ (Z̄ + σε)n(ε)dε > 0 (4.47)
−∞

since u′ (·) is uniformly positive. The partial derivative with respect to variance is
Z ∞
1 ∂V 1
V2 = = u′ (Z̄ + σε)εn(ε)dε. (4.48)
2σ ∂σ 2σ −∞
The negativity of V2 can now be deduced from the symmetry R of n(·) and the positivity of
u′ . Alternatively, integrating by parts by using the property εn(ε)dε = −n(ε) gives
Z ∞
1
V2 = u′′ (Z̄ + σε)σn(ε)dε < 0 (4.49)
2σ −∞
64 Mean-Variance Portfolio Analysis

since u′′ (·) is uniformly negative.

The indifference curves in mean standard deviation space are again convex since

∂2V
Z
2
= u′′ (·)n(ε)dε < 0, (4.50a)
∂ Z̄

∂ 2V
Z
= u′′ (·)ε2 n(ε)dε < 0, (4.50b)
∂σ 2
¶2 ¸2
∂2V ∂2V ∂ 2V
µ Z Z ·Z
′′ 2 ′′ ′′
− = u (·)n(ε)dε ε u (·)n(ε)dε − εu (·)n(ε)dε
∂ Z̄ 2 ∂σ 2 ∂ Z̄∂σ
"Z Z µZ ¶2 #
= (−1)2 g(ε)dε ε2 g(ε)dε − εg(ε)dε .
(4.50c)
′′
The first two inequalities follow from the negativity of u . The last expression is positive
by the Cauchy-Schwarz inequality since g(ε) = −u′′ (·)n(ε) is a positive measure.

4.7 Solving A Specific Problem

′
Consider an investor with utility function U (W ) = −e−a W facing an investment envi-
ronment with a riskless asset and n risky assets distributed multivariate normally N (z̄, Σ).
′
Without loss of generality we can assume U (Z) = −e−a Z , where a = a′ W0 . Substituting
for the budget constraint, we see that the investor’s problem is
n ′
o
Max E −e−a[w (z̄−R1)+R] . (4.51)

Since w′ z̃ is normally distributed with mean w′ z̄ and variance w′ Σw, we can write the
expectation in (51) as −φ(ia), where φ is the characteristic function of a normal distribution
with mean and variance equal to that of the portfolio. Thus, the problem can be restated as

a2 w′ Σw
½ µ ¶¾
′
Max − exp −a [w (z̄ − R1) + R] + . (4.52)
2

The utility function to be maximized in (52) is an ordinal function of mean and variance.
Since it is ordinal, we can operate on it with the increasing function θ(x) = −(1/a)[log(−x)−
R] to give the problem

aw′ Σw
Max w′ (z̄ − R1) − . (52′ )
2
The solution to (52’) satisfies

0 = z̄ − R1 − aΣw or w∗ = a−1 Σ−1 (z̄ − R1).

(4.53)
w0∗ = 1 − 1′ w∗ = 1 − a−1 (B − AR).

Comparing this to (19), we see that the risky assets are held in the same proportions as in
the tangency portfolio in (19). The more risk averse is the investor (the higher is a), the
fewer risky assets will be held.
4.8 The State Prices Under Mean-Variance Analysis 65

4.8 The State Prices Under Mean-Variance Analysis

Starting with the expected returns relation (38), we can compute state prices consistent with
the CAPM as follows:
Z̄M − R
z̄i = R + Cov(z̃i , Z̃M ) 2
≡ R + λ Cov(z̃i , Z̃M ),
σM
Z Z
£ ¤
π(s)zi (s)ds = R + λ π(s)zi (s) ZM (s) − Z̄M ds,
Z
−1
£ ¡ ¢¤
1 = R π(s) 1 − λ ZM (s) − Z̄M zi (s)ds,

p(s) = R−1 π(s) 1 − λ ZM (s) − Z̄M .

£ ¡ ¢¤

(4.54)

Although this pricing “vector” is clearly a set of supporting prices, it need not be the positive
state pricing vector we seek. In particular, if the market portfolio is normally distributed,
then Z̃M can be arbitrarily larger than its mean. For those states in which λ(ZM (s)−Z̄M ) >
1, the state pricing function given in (54) will be negative.
Since the CAPM equation above is an equilibrium relation, there must be an alternative
way to describe state prices in which they are all positive, for a market in equilibrium cannot
contain any arbitrage opportunities.
If the CAPM equilibrium arises because all investors have quadratic utility, then po-
tential arbitrage opportunities are not a concern. Investors with quadratic utility will not
take unlimited arbitrage positions, for this would increase their wealth past the point where
marginal utility becomes negative, and the guaranteed profit of an arbitrage would only
hurt.
With multivariate normal returns supporting the CAPM, however, there can be no arbi-
trage opportunities. This is true for the equilibrium reason stated earlier. In addition, there
can be no possible arbitrage opportunities among assets with normally distributed returns
simply because no portfolio that has a completely nonnegative return can be constructed
from the risky assets.
Valid positive state prices for the case of normally distributed asset returns can be con-
structed as follows. Consider the problem stated in Equation (51) of the previous section.
The first-order conditions for this (unmodified) problem are
h ∗
i
E e−aZ̃ (z̃i − R) = 0. (4.55)

From (53) an investor with risk aversion a ≡ aM ≡ B − AR > 0 will hold just the
tangency (market) portfolio. For this investor the first-order condition can be written as
Z Z
π(s) exp[−aM ZM (s)]zi (s)ds = R π(s) exp[−aM ZM (s)]ds ≡ K −1 . (4.56)

Thus
p(s) = Kπ(s) exp[−aM ZM (s)] > 0. (4.57)

4.9 Portfolio Analysis Using Higher Moments

Apart from tractability and tradition there are no reasons not to consider third and higher
central moments in portfolio problems. indeed, a close examination of these problems can
66 Mean-Variance Portfolio Analysis

provide understanding even for the mean-variance problem, which is sometimes otherwise
missed through too much familiarity. We outline the equilibrium by using mean-variance-
skewness analysis. (Note: Skewness usually means the third central moment divided by the
cube of the standard deviation; here skewness is the unnormalized
h ithird central moment.)
3 3
To compute the skewness of a portfolio, m ≡ E (Z̃ − Z̄) , we need to know not
only the assets’ skewnesses but also the coskewnesses

mijk ≡ E [(z̃i − z̄i )(z̃j − z̄j )(z̃k − z̄k )] . (4.58)

Then
N X
X N X
N
m3p = wi wj wk mijk (4.59)
1 1 1

Now consider an investor with the derived utility function V (Z̄, σ 2 , m3 ) with optimal
portfolio characterized by the moments Z̄o , σo2 , m3o . The portfolio combination of asset i
and the optimal portfolio is characterized by the moments

Z̄ = (1 − w)Z̄o + wz̄i ,
σ 2 = (1 − w)2 σo2 + 2w(1 − w)σio + w2 σi2 ,
m3 = (1 − w)3 m3o + 3(1 − w)2 wmooi + 3(1 − w)w2 moii + w3 m3i .
(4.60)

Since at the optimum, differential changes in the optimal portfolio leave utility unaffected,
dσ 2 dm3 ¯¯
¯
dZ̄
0 = V1 + V2 + V3
dw dw dw ¯w=0
= V1 (z̄i − Z̄o ) + 2V2 (σio − σo2 ) + 3V3 (mioo − m3o ).
(4.61)

Rearranging (61) gives the expected return on asset i as a function of its covariance and
coskewness with portfolio o and two utility measures
2V2 3V3
z̄i = Z̄o − (σio − σo2 ) − (mioo − m3o ). (4.62)
V1 V1
Equation (62) also holds for the riskless asset if one is available, and since σ0o = m0oo =
0,
2V2 2 3V3 3
R = Z̄o + σ + m. (4.63)
V1 o V1 o
Solving (63) for 2V2 /V1 and substituting into (62) yield
¢ 3V3 Mo3 o
z̄i = R + βio Z̄o − R + (βi − γio ),
¡
(4.64)
V1
σio mioo
βio ≡ 2 , γio ≡ .
σo m3o
Now if some investor’s optimal portfolio is the market (see Chapter 5 and 6), then the first
two terms in (64) duplicate the Sharpe-Lintner CAPM.
The deviations from that model can be either positive or negative; however, the contri-
bution of the skewness term
∂ z̄i 3V3
=− (4.65)
∂mioo V1
4.9 Portfolio Analysis Using Higher Moments 67

is easily signed. Assuming higher skewness is preferred, we see that a comparative static
decrease in coskewness requires an increase in expected return to induce the same holding
of the asset at the margin. The intuition is the same as in the mean-variance model. If
both mean and skewness are desirable, then the investor can remain at the margin after the
comparative static change only if mo decreases when Z̄o increases, and vice versa. Again,
as in the mean-variance model, it is only the co-moment that matters. Skewness and miio
are irrelevant at the margin, just as individual asset variances are. The intuition is that they
provided no marginal tradeoff opportunities, since, as seen in (60), ∂m3 /∂w = mioo − m3o
at w = 0. The co-moment mioo , however, contributes fully since skewness is the weighted
average of this coskewnesses measure
X
m3o = wi∗ mioo (4.66)

just as σo2 = wi∗ σioh. A second explanation 3

P
i of mioo ’s contribution to mo comes from its
definition: mioo ≡ E (z̃i − z̄i )(Z̃o − Z̄o )2 . If this is positive, then z̃i tends to be above
(below) its mean when the return on the remainder of Z̃o is very far from (close to) its
mean. This tends to shift the density function of Z̃o to the right (left) for very high or low
(intermediate) values. That is, it imparts a positive skewness.
The second utility term in the pricing relation can be eliminated by introducing a second
portfolio. For example, let the subscript z denote the minimum variance portfolio uncorre-
lated with Z̃o . Then, since (62) holds for Z̄z and βzo = 0,
¢ γo − βo ¡
z̄i = R + βio Z̄o − R + i o i Z̄z − R .
¡ ¢
(4.67)
γz

Any uncorrelated portfolio could be used in place of Z̃z ; however, unlike in the mean-
variance model each may have a different expected return. Different values of the parameter
γzo will correct for this.
Appendix A

The Budget Constraint

Throughout this chapter a strict budget constraint 1′ w = 1 was imposed, indicating that all
wealth must be invested. In some cases the appropriate budget constraint should actually be
1′ w 6 1; that is, no more than the investor’s entire wealth can be invested. It might appear
that the more stringent requirement is perfectly general since it would seem foolish ever to
throw away wealth. Actually, in the standard mean-variance problem this can be done as a
part of an optimal portfolio.
Consider the problem posed earlier with multivariate normally distributed asset returns
and an investor with exponential utility but without a riskless asset. As in (52) the strictly
constrained problem can be stated now as a Lagrangian:
aw′ Σw
L(w, λ) ≡ w′ z̄ − − λ(1′ w − 1). (A.1)
2
The first-order conditions are
z̄ − aΣw − λ1 = 0, (A.2a)
′
1 w − 1 = 0. (A.2b)
Solving (A2a) gives
1 £ −1
w∗ = λΣ 1 + Σ−1 z̄ .
¤
(A.3)
a
Premultiplying (A3) by 1′ and imposing the budget constraint (A2b) yield an equation in λ
which can be solved, giving
a − 1′ Σ−1 z̄ a−B
λ= ′ −1
= . (A.4)
1Σ 1 A
Thus, the final solution for the optimal portfolio is
1 − B/a 1
w∗ = Σ−1 1 + Σ−1 z̄ . (A.5)
A a
The investor holds a combination of portfolios wg and wd The greater is the risk aversion,
a, the more the investor holds of the former.
For the second problem the weak budget constraint can be restated as 1′ w = b 6 1,
where b is a control variable. For this problem the Lagrangian is
aw′ Σw
L(w, λ, b) ≡ w′ z̄ − − λ(1′ w − b) (A.6)
2
with the constraint b 6 1. Using the Kuhn-Tucker methodology, we obtain that the first-
order conditions are (A2a) and
∂L
= −1′ w + b = 0, (A.7a)
∂λ
 69

∂L
= λ ≥ 0, (A.7b)
∂b
(1 − b)λ = 0. (A.7c)
If a > B , then the same solution applies. This is true, because from (A4) λ is positive, so
from (A7c) b = 1 and (A7a) is identical to (A2b). For smaller values of a, λ must be set to
zero to satisfy both (A7b) and (A7c). In this case the solution to (A2a) is w∗ = Σ−1 z̄/a.
The entire solution can be written as
· ¸
−1 1 − B/a 1
∗
w = Σ 1 max 0, + Σ−1 z̄ . (A.8)
A a
As before, the investor holds some combination of portfolios wg and wd . If B > 0,
investors who are sufficiently close to risk neutral are short in the first portfolio and long
in the second. At higher levels of risk aversion, wealth is shifted from wd into wg . For
a = 1′ Σ−1 z̄ just wd is held, but at this point the shifting into wg stops. For even higher
risk-aversion levels, wealth is discarded rather than being invested in wg .
The opportunity to freely discard wealth can be viewed as a riskless asset with a zero
gross return to those investors who are sufficiently risk averse to desire this opportunity.
Less risk-averse investors cannot, however, borrow at this rate, −100%. This solution is
illustrated in Figure 43, The lower portion of the dashed line, which was used to locate
portfolio d, can now be interpreted as the borrowing-lending line.
If B < 0 so that portfolio wd is on the lower limb of the hyperbola, the solution appears
even stranger. Now for risk-averse investors a must exceed B , so the solution for every
investor is to short portfolio wd and discard all wealth, including the proceeds from this
short sale.
The explanation for these counterintuitive results lies in the unlimited liability embodied
in the normal distribution and the interpretation of negative wealth implicit in exponential
utility. Since the exponential function is decreasing over the entire real line, negative ex-
ponential utility views negative wealth as a defined concept and attributes less utility to
larger bankruptcies. Were this not the case - that is, were, for example, W = −100 and
W = −1000 to be considered equally undesirable - then the optimal solution would not
have this characteristic. Wealth would not be discarded, Alternatively, if some asset had
limited liability, its return was bounded below by −1, then it would always be preferable to
invest in this asset rather than to discard wealth, or if consumption in the first period were
modeled explicitly, the investor would just consume more and invest all remaining wealth.
Appendix B

The Elliptical Distributions

Multivariate normality is the usual, but not the only, distribution of returns for which mean-
variance (or generally median-dispersion) analysis is fully consistent with expected utility
maximization. Clearly, if there are only two assets and they have different expected re-
turns, each possible portfolio is completely characterized by its mean return. A fortiori,
mean-variance analysis is valid. For n > 2, Tobin originally conjectured that any family
of two-parameter distributions would be consistent with mean-variance analysis. but this
class is too broad. The important missing requirement was that any portfolio (i.e., linear
combination) of the random returns also had to have a distribution in the same family.
The distributions which permit mean-variance analysis are the elliptical distributions, so
called because the isoprobability contours are ellipsoidal. A vector x̃ of n random variables
is said to be elliptically distributed if its density function can be expressed as

f (x) = |Ω|−1/2 g (x − µ)′ Ω−1 (x − µ); n .

£ ¤
(B.1)

Ω is the positive definite dispersion matrix, and µ is the vector of medians. If variances
exist, then the covariances matrix is proportional to Ω. If means exist, µ is also the vector
of means. Note that g(·) is a univariate function with parameter n. The functions g(·; n)
in the same class but with different parameter values can be quite different in appearance.
Membership in a particular class can be verified through the characteristic function, which
does not depend on n.
The characteristic function of elliptical random variables has the form
h ′ i ′
φn t ≡ E eit x̃ = eit µΨ(t′ Ωt) (B.2)

for some function Ψ which does not depend on n.

Multivariate normal random variables are the most well-known ellipticals. For these,
g(s; n) ∝ exp(−s/2) and Ψ(T ) = exp(−T /2). Many other types are also possible. For
example, the multivariate student t distribution with v degrees of freedom is characterized
by
g(s; n) ∝ (v + s)−(n+v)/2 . (B.3)
(The multivariate Cauchy distribution, v = 1, is a special case of an elliptical distribution in
which neither the means nor variances are defined.) In fact, any nonnegative function, when
properly normalized, can be used as the basis of an elliptical distribution. Note that the g(s)
can be identically zero for sufficiently large s, so that the distribution of asset returns could
be bounded above and below. Thus, the criticism of the CAPM - that it requires assets to
have unlimited liability - although true for the normal case, is not generally valid.
 71

The marginal distribution of any component of x is also elliptical of the same class. If
x1 is a subvector of x of length v , then its density function is

f (x1 ) = |Ω11 |−1/2 g (x1 − µ1 )′ Ω−1

£ ¤
11 (x1 − µ1 ); v . (B.4)

This can be easily demonstrated by evaluating φv (t) = φn (t, 0) by using (B2). More
generally, weighted sums of any elliptical random variables have an identical joint elliptical
form that depends only on their mean vector and dispersion matrix.
Let v ≡ Tx be a set of weighted sums for any matrix T. Then the mean (median) vector
for v is m = Tµ, and the dispersion matrix is proportional to the covariance matrix

∆ ≡ kE [(v − m)(v − m)′ ] = k TE [(x − µ)(x − µ)′ ] T′ = TΩT′ .

The characteristic function of v is

h i h i
it′ v it′ Tx
φv (t) ≡ E e =E e = φx (T′ t)
′ ′
= eit TµΨ(t′ TΩT′ t) = eit m Ψ(t′ ∆t).
(B.5)

Since the characteristic function of v has the same form as the characteristic function of any
subvector of x with the same number of elements, this proves the point.
To prove that “mean-variance” analysis is valid for any elliptical distribution for returns,
we need only show that the distribution of any portfolio is completely specified by its mean
and dispersion and that the latter is disliked, so that investors hold minimum dispersion
portfolios. Of course when variances exist, any portfolio’s variance is proportional to its
dispersion, so optimal portfolios are minimum variance as well. The intuition behind these
steps is identical to that for the multivariate normal case.
Any portfolio of risky assets Z̃p = w′ z̃ is a weighted sum, so if the assets have an
elliptical distribution, the distribution of the portfolio is the appropriate marginal with char-
acteristic function
φp (t) = φz (tw) = eitµ Ψ(t2 ω 2 ), (B.6)
where µ = w′ z̄ and ω 2 ≡ w′ Ωw are the portfolio’s mean (or median) return and dispersion.
Now the mth central moment of Z̃p will be
m
¯
d
i−m m Ψ(t2 ω 2 )¯¯ .
¯
(B.7)
dt t=0

It is clear that all odd moments will be zero, and even moments will be proportional to ω m .
Therefore, the distribution of any portfolio of risky assets is characterized completely by µ
and ω .
If the riskless asset is included, then Z̃p − R = w′ (Z̃ − R1), which is a weighted sum
of ellipticals. The latter is characterized by ω 2 ≡ w′ Ωw and µ − R ≡ w′ (Z̄ − R1) as
just shown, but ω 2 is the portfolio’s variance, which, as before, determines all the higher
moments.
We now show that dispersion is disliked. For the density function of the portfolio’s
return f (Z̃p ), expected utility is
Z ∞ Z ∞
V̂ (µ, ω) ≡ Eu(Z̃p ) = u(Zp )f (Zp )dZp = u(µ + ωξ)g(ξ 2 ; 1)dξ. (B.8)
−∞ −∞
72 The Elliptical Distributions

The second line follows from the substitutions ξ = (Zp − µ)/ω , dZp = ωdξ . Then
Z ∞
V̂2 = u′ (µ + ωξ)ξg(ξ 2 ; 1)dξ
Z−∞∞
= ξg(ξ 2 ; 1) [u′ (µ + ωξ) − u′ (µ − ωξ)] dξ < 0.
0
(B.9)
The inequality follows from the concavity of u, which guarantees that the term in brackets
is negative. Since variance is proportional to dispersion, ω , it is likewise disliked.
These two steps complete the proof that mean-variance analysis is valid for elliptical
distributions, and the CAPM continues to obtain. Note that the CAPM would even be valid
if investors disagreed about the distribution of returns, that is, about g(·), provided they had
identical assessments about µ and Ω. Although we shall not prove so here, for more than
two assets this condition is also necessary for the validity of mean-variance analysis (but
not for the CAPM; see Chapters 6 and 9) when a riskless asset is available.
When there is no riskless asset, the class of distributions permitting mean-variance anal-
ysis to be consistent with expected utility maximization is enlarged somewhat to
z̃ = ε̃1 1 + ε̃2 a + x̃, (B.10)
where the vector x̃ is elliptically distributed around zero conditional on ε̃. That is, the form
of g may depend upon their realizations, but the dispersion matrix Ω remains the same.
We shall not analyze this case in detail, but the intuition is straightforward. Since
′
w 1 = 1, any portfolio formed from these assets includes all the risk associated with ε̃1 .
Consequently, investors “ignore” it and minimize the remaining risk. (ε̃1 may affect the
choice of which minimum-variance portfolio is chosen.) Also, for any portfolio,
Z̄p = ε̄1 + ε̄2 w′ a, (B.11)
so the mean return (together with the exogenous quantities ε̃1 and ε̃2 , which are the same
for all portfolios) is a sufficient statistic for w′ a, which is the amount of ε̃2 risk included in
the portfolio. Finally, since x̃ is conditionally elliptical around zero given ε̃, E(x̃|ε) = 0,
implying Cov(x̃, ε1 ) = Cov(x̃, ε˜2 ) = 0. Thus
Var(Z̃p ) = Var(ε̃1 ) + 2w′ a Cov(ε̃1 , ε̃2 ) + (w′ a)2 Var(ε̃2 ) + Var(w′ x̃). (B.12)
From (B12) it is clear that the portfolio’s mean (which determines w′ a) and variance (again
together with the three exogenous variance numbers) are sufficient statistics for variance
of the elliptical random variable affecting the portfolio’s return. But we already know that
the variance of a particular zero mean elliptical random variable is a sufficient statistic for
its entire distribution. Thus, the portfolio mean and variance completely determine the
distribution of the portfolio’s return.
Again with more than two assets this condition is also necessary for the validity of
mean-variance analysis in the absence of a riskless asset. We shall refer to this class of
distributions as the augmented elliptical class.

B.1 Some Examples of Elliptical Variables

As stated previously, the number of different types of elliptical distributions is literally
uncountable. To illustrate the properties of elliptical variables, we use the following simple
example.
B.1 Some Examples of Elliptical Variables 73

Let (x1 , x2 ) be uniformly distributed on the unit circle. That is, the density function is
for x21 + x22 6 1,
½ −1
π ,
f (x) = (B.13)
0, elsewhere.
p
The marginal density of xi can be obtained by integrating out xj . Since |xj | 6 1 − x22 ≡
L, Z L
2¡ ¢1/2
f (xi ) = π −1 dxj = 1 − x2i (B.14)
−L π
for x2i ≤ 1 and zero elsewhere.
Note that as suggested earlier the marginal distribution can be quite different in appear-
ance from the joint distribution. Also, the conditional distribution is obviously uniform;
specifically
f (x) (1 − x2j )−1/2
f (xi |xj ) = = for x2i ≤ 1 − x2j .
f (xj ) 2
Thus, we see that the marginal and conditional densities can be different in appearance as
well. This latter fact indicates that the random variables are not independent even though
each is conditionally independent of the other.
Now let v ≡ ax1 + bx2 . The density function for this sum is
Z µ ¶
v − ax1
f (v) = f x1 , dx1 , (B.15)
b
where the integral is over the region where x2i + b−2 (v − ax1 )2 ≤ 1, giving
¸1/2
v2
·
2
f (v) = 1− 2 , (B.16)
π a + b2
for v 2 ≤ a2 + b2 , and zero elsewhere. Comparing (B16) to (B12), we see that all linear
combinations of x1 and x2 have the same distributional form as either xi - namely,
¶1/2
v2
µ
2
f (v) = 1− 2 , (B.17)
π ω
where ω is the semirange, a measure of dispersion.
This example demonstrates the properties of elliptical distributions and illustrates that
mean-variance analysis (and hence the CAPM) can be valid even with substantially nonnor-
mal returns. It is not, however, a particularly apt description of asset returns. One general
and useful category of augmented elliptical random variables is the subordinated normal
class.
A random vector of returns on the risky assets is said to be jointly subordinated normal if
its distribution is conditionally normal. That is, it is distributed as N (p̃θ, q̃Υ), where p̃ and
q̃ are scalar random variables independent of the normal deviates. q must be nonnegative
with probability 1.
p̃ and q̃ are called the directing variables. We will show that this class of distributions
is augmented elliptical by demonstrating that all portfolios can be described by their means
and variances alone.
Assume that the directing variables have the bivariate density h(p, q) and characteristic
function θh (t1 , t2 ) ≡ E [exp(it1 p̃ + it2 q̃)]. Then the characteristic function of the un-
conditional distribution can also be derived. For a given portfolio of risky assets, w, the
74 The Elliptical Distributions

conditional distribution is normal with conditional mean and conditional variance of

pθ = pw′ θ, qv 2 = q w′ Υw, (B.18)

respectively. The characteristic function of the unconditional portfolio distribution is

Z ∞ ·Z Z µ ¶ ¸
itZ Z − pθ
φ(t) ≡ e n √ h(p, q)dpdg dZ
−∞ v q
Z Z ·Z ∞ µ ¶ ¸
itZ Z − pθ
= e n √ dZ h(p, q)dpdq
−∞ v q
t2 v 2 g
Z Z µ ¶
= exp itθp − h(p, q)dpdq
2
it2 v 2
µ ¶
h
= φ θt, .
2
(B.19)

In the second line we have assumed that it is permissible to change the order of the
integration. In that case the innermost integral is the characteristic function of the normal
distribution. The fourth line follows from the definition of φh . It is clear from (B19) that the
distribution of returns on any portfolio is completely characterized by θ and v since these
two variables completely determine its characteristic function (together with the distribution
of p̃ and q̃ , which is, of course, the same for all portfolios). We next show the the portfolio’s
mean and variance are sufficient statistics.
Evaluating the derivatives

φ′ (t) = θφh1 + itv 2 φh2 ,

φ(t) = θ2 φh11 + 2iθv 2 tφh22 − v 4 t2 φh22 + iv 2 φh2
(B.20)

gives the moment as

1 1
E[Z̃] = φ′ (0) = θφhi (0, 0) = θE[p̃], (B.21a)
i i
E[Z̃ 2 ] = −φ′′ (0) = −θ2 φh11 (0, 0) − iv 2 φh2 (0, 0) = θ2 E[p̃2 ] + v 2 E[q̃], (B.21b)
Var[Z̃] = θ2 Var[p̃2 ] + v 2 E[q̃]. (B.21c)
Thus, Z̄ is a sufficient statistic for the conditional mean pθ, and Z̄ and Var(Z̃) jointly de-
termine the conditional variance qv 2 . This concludes the proof; however, since (B19) holds
for all portfolios, we can additionally use it and (B6) to write the multivariate characteristic
function as
it2 w′ Υw
h ′ i µ ¶
itw Z̃ h ′
φ(t) ≡ E e = φ tw θ, , (B.22a)
2
it′ Υt
µ ¶
h ′
φz (t) = φ t θ, , (B.22b)
2
where t = tw.
Whenever p̃, the random variable affecting the means, is a constant, the subordinated
normals are elliptical, and a riskless asset can also be added to the opportunity set. In some
cases when p̃ is not degenerate, there is an alternative choice of directing variables p̃ and
q̃ with the former constant which yields the same distribution of returns. This will be true,
B.2 Solving a Specific Problem 75

for example, whenever p̃ and q̃ are independent and p̃ has a symmetric distribution. Again
a riskless asset can be added. In other cases the unconditional distribution for the assets’
returns are only in the augmented elliptical class, and no riskless asset is permitted.

B.2 Solving a Specific Problem

As a specific example take p̃ to be degenerate, that is p̃ ≡ 1, and q̃ to have an exponential
distribution
h(p, q) = δ(p − 1)λe−λq , (B.23a)
µ ¶−1
h is it
φ (s, t) = e 1− . (B.23b)
λ
′
From (B22) and (B23b) the characteristic function of the risky assets is eit θ (1+t′ Υt/2λ)−1 .
The corresponding distribution
µ ¶1/2
λ
exp −[2λ(z − θ)′ Υ−1 (z − θ)]1/2
¡ ¢
f (z) = (B.24)
2|Υ|
is known as the Laplace or double exponential distribution. Its mean vector is z̄ = θ, and its
covariance matrix is Σ = Υ/λ. The marginal distributions are symmetric, but the kurtosis
is 6 (“excess” is 3), so they are “fat tailed.”
An exponential utility investor, u(Z) = −e−aZ , facing this opportunity set would
solve the problem stated in (50). Using the characteristic function, we get that this is
max[−e−aW R φz (iaw)] or
¶−1
w′ Υw
µ
−a(R+w′ (θ−R1))
Min e 1− . (B.25)
2λ
The first-order condition is
aΥw
0 = θ − R1 − ,
(1 −a2 w′ Υw/2λ)λ
(B.26)
1 − a2 w′ Σw/2 −1
w∗ = Σ (z̄ − R1).
a
Recall that z̄ = θ and Υ = λΣ. Comparing (B26) to (53) we see that the investor’s
holdings in the risky assets are proportional in these two cases. We know this must be true
because in each case the same two-fund separation obtains. He will, however, change the
leverage of his portfolio.
Under the first investment environment the investor holds the fraction 1′ w∗ = (B −
AR)/a in the risky assets. In the present economy the fraction [1−a2 w′ Σw/2](B−AR)/a
is placed at risk. To compute the expression in brackets in terms of exogenous quantities
alone, define

v ≡ a2 w′ Σw,
³ v ´2
v = 1− (z̄ − R1)′ Σ−1 ΣΣ−1 (z̄ − R1)
2
³ v ´2 2
³ v ´2 2
= 1− (C − 2RB + R A) ≡ 1 − b.
2 2
(B.27)
76 The Elliptical Distributions

The second line follows from (B26). From (18) or Figure 4.4, b is the slope of the efficient
frontier. (B27) is a quadratic equation in v .
Solving for v and substituting into (B27) give a total risky holding of
√
′ ∗ 1 + 2b2 − 1 B − AR
1w = . (B.28)
b2 a
√
Since 1 + 2b2 − 1 < b2 , the investor places fewer funds at risk in this economy. This
is just the result expected because portfolio returns are now leptokurtotic, and the investor
dislikes kurtosis. (u(4) (Z) < 0.)

B.3 Preference Over Mean Return

We stated previously that whenever return distributions are such that mean-variance anal-
ysis is consistent with expected utility maximization and a riskless asset is available, then
investors who prefer more to less will always prefer the portfolio with the highest mean
among portfolios with the same variance. That is V1 > 0, and investors’ optimal portfolios
are on the upper limb of the locus of minimum variance portfolios. On the other hand,
the preference of higher mean is not true for some of the additional distributions permitted
when no riskless asset is available.
To prove the first part of this statement, recall that mean-variance analysis is valid with a
riskless asset only if all risky assets have an elliptical distribution. In this case the expected
utility of any portfolio can be expressed via the derived utility function defined over mean
and variance given in (B8). Evaluating the derivative gives
Z ∞
V1 = V̂1 = u′ (µ + ωξ)g(ξ 2 ; 1)dξ > 0. (B.29)
−∞

The inequality follows from the assumption of positive marginal utility.

In the absence of a riskless asset the returns must only be elliptical conditional upon an
arbitrary random vector ε̃. The expected utility of a portfolio can still be written in terms
of the mean and variance of the elliptical distribution
Z Z
V̂ (µ, ω) ≡ Eu(Zp ) = u(µ + ωξ) g(ξ 2 , ε; 1)f (ε)dεdξ. (B.30)

As before, Z Z
′
V̂1 = u (·) g(·)f (·)εdξ > 0; (B.31)

however, now an increase in µ, holding ω constant, is not the same as an increase in Z̄p ,
holding σp constant, because σp , may depend on µ through interactions via the distribution
of ε. If increasing Z̄p increases both µ and ω , then a higher mean may not be preferred.
A specific illustration of a situation in which some investor, with positive marginal
utility, holds a minimum-variance portfolio with mean return less than that on the global
minimum-variance portfolio is the following simple example. Assume that there are two
assets with returns z1 = {1, 3} and z2 = {0, 6}. The two outcomes are each equally likely,
and the returns are independent. Then ẑ1 = 2, ẑ2 = 3, σ12 = 1, σ22 = 9, and σ12 = 0.
Because there are only two assets, all portfolios are minimum variance, and mean-variance
analysis is obviously valid.
The global minimum-variance portfolio is w′g = (0.9, 0.1), with Ẑg = 2.1 and σg2 = 0.9.
Since Ẑg > ẑ1 , asset 1 is on the lower, “inefficient” portion of the minimum-variance
B.3 Preference Over Mean Return 77

frontier. Nevertheless, asset 1 and the portfolios between asset 1 and the global minimum-
variance portfolio are each preferred to the corresponding portfolio with the same variance
on the upper limb by some investors. (They are each even the optimal portfolio for some
investor.)
For example, the portfolio with the same variance as asset 1 is w′ = (0.8, 0.2), with
Z̄p = 2.2 and the four equally likely returns {0.8, 2.0, 2.4, 3.6}. For the utility function
−e−4Z , the expected utility of this portfolio is approximately -.010, whereas for asset 1
expected utility is about -.009. Thus, asset 1 is preferred to the portfolio even though the
former is mean-variance dominated.
The reason asset 1 is preferred to the portfolio despite being mean-variance dominated
is that its remaining moments are “better.” Both distributions are symmetric, but all the
even central moments for asset one are less than those for the portfolio. Since all even
moments are “disliked” by exponential utility functions, there is a tradeoff between mean
and these higher moments when comparing asset 1 and the portfolio. For exponential utility
−exp(−aZ), the larger a is, the more important are the higher moments.
If all assets are joint elliptically distributed, then such tradeoffs cannot exist because
all portfolios with the same variance have identical central moments of all orders. In the
augmented elliptical class, however, these tradeoffs do arise through the variable ε̃2 , which
links mean returns with the other moments.